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, 2017 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;
310 /* number of seconds before growing cache again */
311 static int arc_grow_retry
= 5;
313 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
314 int zfs_arc_overflow_shift
= 8;
316 /* shift of arc_c for calculating both min and max arc_p */
317 static int arc_p_min_shift
= 4;
319 /* log2(fraction of arc to reclaim) */
320 static int arc_shrink_shift
= 7;
323 * log2(fraction of ARC which must be free to allow growing).
324 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
325 * when reading a new block into the ARC, we will evict an equal-sized block
328 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
329 * we will still not allow it to grow.
331 int arc_no_grow_shift
= 5;
335 * minimum lifespan of a prefetch block in clock ticks
336 * (initialized in arc_init())
338 static int arc_min_prefetch_lifespan
;
341 * If this percent of memory is free, don't throttle.
343 int arc_lotsfree_percent
= 10;
348 * The arc has filled available memory and has now warmed up.
350 static boolean_t arc_warm
;
353 * log2 fraction of the zio arena to keep free.
355 int arc_zio_arena_free_shift
= 2;
358 * These tunables are for performance analysis.
360 unsigned long zfs_arc_max
= 0;
361 unsigned long zfs_arc_min
= 0;
362 unsigned long zfs_arc_meta_limit
= 0;
363 unsigned long zfs_arc_meta_min
= 0;
364 unsigned long zfs_arc_dnode_limit
= 0;
365 unsigned long zfs_arc_dnode_reduce_percent
= 10;
366 int zfs_arc_grow_retry
= 0;
367 int zfs_arc_shrink_shift
= 0;
368 int zfs_arc_p_min_shift
= 0;
369 int zfs_arc_average_blocksize
= 8 * 1024; /* 8KB */
371 int zfs_compressed_arc_enabled
= B_TRUE
;
374 * ARC will evict meta buffers that exceed arc_meta_limit. This
375 * tunable make arc_meta_limit adjustable for different workloads.
377 unsigned long zfs_arc_meta_limit_percent
= 75;
380 * Percentage that can be consumed by dnodes of ARC meta buffers.
382 unsigned long zfs_arc_dnode_limit_percent
= 10;
385 * These tunables are Linux specific
387 unsigned long zfs_arc_sys_free
= 0;
388 int zfs_arc_min_prefetch_lifespan
= 0;
389 int zfs_arc_p_aggressive_disable
= 1;
390 int zfs_arc_p_dampener_disable
= 1;
391 int zfs_arc_meta_prune
= 10000;
392 int zfs_arc_meta_strategy
= ARC_STRATEGY_META_BALANCED
;
393 int zfs_arc_meta_adjust_restarts
= 4096;
394 int zfs_arc_lotsfree_percent
= 10;
397 static arc_state_t ARC_anon
;
398 static arc_state_t ARC_mru
;
399 static arc_state_t ARC_mru_ghost
;
400 static arc_state_t ARC_mfu
;
401 static arc_state_t ARC_mfu_ghost
;
402 static arc_state_t ARC_l2c_only
;
404 typedef struct arc_stats
{
405 kstat_named_t arcstat_hits
;
406 kstat_named_t arcstat_misses
;
407 kstat_named_t arcstat_demand_data_hits
;
408 kstat_named_t arcstat_demand_data_misses
;
409 kstat_named_t arcstat_demand_metadata_hits
;
410 kstat_named_t arcstat_demand_metadata_misses
;
411 kstat_named_t arcstat_prefetch_data_hits
;
412 kstat_named_t arcstat_prefetch_data_misses
;
413 kstat_named_t arcstat_prefetch_metadata_hits
;
414 kstat_named_t arcstat_prefetch_metadata_misses
;
415 kstat_named_t arcstat_mru_hits
;
416 kstat_named_t arcstat_mru_ghost_hits
;
417 kstat_named_t arcstat_mfu_hits
;
418 kstat_named_t arcstat_mfu_ghost_hits
;
419 kstat_named_t arcstat_deleted
;
421 * Number of buffers that could not be evicted because the hash lock
422 * was held by another thread. The lock may not necessarily be held
423 * by something using the same buffer, since hash locks are shared
424 * by multiple buffers.
426 kstat_named_t arcstat_mutex_miss
;
428 * Number of buffers skipped because they have I/O in progress, are
429 * indrect prefetch buffers that have not lived long enough, or are
430 * not from the spa we're trying to evict from.
432 kstat_named_t arcstat_evict_skip
;
434 * Number of times arc_evict_state() was unable to evict enough
435 * buffers to reach its target amount.
437 kstat_named_t arcstat_evict_not_enough
;
438 kstat_named_t arcstat_evict_l2_cached
;
439 kstat_named_t arcstat_evict_l2_eligible
;
440 kstat_named_t arcstat_evict_l2_ineligible
;
441 kstat_named_t arcstat_evict_l2_skip
;
442 kstat_named_t arcstat_hash_elements
;
443 kstat_named_t arcstat_hash_elements_max
;
444 kstat_named_t arcstat_hash_collisions
;
445 kstat_named_t arcstat_hash_chains
;
446 kstat_named_t arcstat_hash_chain_max
;
447 kstat_named_t arcstat_p
;
448 kstat_named_t arcstat_c
;
449 kstat_named_t arcstat_c_min
;
450 kstat_named_t arcstat_c_max
;
451 kstat_named_t arcstat_size
;
453 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
454 * Note that the compressed bytes may match the uncompressed bytes
455 * if the block is either not compressed or compressed arc is disabled.
457 kstat_named_t arcstat_compressed_size
;
459 * Uncompressed size of the data stored in b_pabd. If compressed
460 * arc is disabled then this value will be identical to the stat
463 kstat_named_t arcstat_uncompressed_size
;
465 * Number of bytes stored in all the arc_buf_t's. This is classified
466 * as "overhead" since this data is typically short-lived and will
467 * be evicted from the arc when it becomes unreferenced unless the
468 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
469 * values have been set (see comment in dbuf.c for more information).
471 kstat_named_t arcstat_overhead_size
;
473 * Number of bytes consumed by internal ARC structures necessary
474 * for tracking purposes; these structures are not actually
475 * backed by ARC buffers. This includes arc_buf_hdr_t structures
476 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
477 * caches), and arc_buf_t structures (allocated via arc_buf_t
480 kstat_named_t arcstat_hdr_size
;
482 * Number of bytes consumed by ARC buffers of type equal to
483 * ARC_BUFC_DATA. This is generally consumed by buffers backing
484 * on disk user data (e.g. plain file contents).
486 kstat_named_t arcstat_data_size
;
488 * Number of bytes consumed by ARC buffers of type equal to
489 * ARC_BUFC_METADATA. This is generally consumed by buffers
490 * backing on disk data that is used for internal ZFS
491 * structures (e.g. ZAP, dnode, indirect blocks, etc).
493 kstat_named_t arcstat_metadata_size
;
495 * Number of bytes consumed by dmu_buf_impl_t objects.
497 kstat_named_t arcstat_dbuf_size
;
499 * Number of bytes consumed by dnode_t objects.
501 kstat_named_t arcstat_dnode_size
;
503 * Number of bytes consumed by bonus buffers.
505 kstat_named_t arcstat_bonus_size
;
507 * Total number of bytes consumed by ARC buffers residing in the
508 * arc_anon state. This includes *all* buffers in the arc_anon
509 * state; e.g. data, metadata, evictable, and unevictable buffers
510 * are all included in this value.
512 kstat_named_t arcstat_anon_size
;
514 * Number of bytes consumed by ARC buffers that meet the
515 * following criteria: backing buffers of type ARC_BUFC_DATA,
516 * residing in the arc_anon state, and are eligible for eviction
517 * (e.g. have no outstanding holds on the buffer).
519 kstat_named_t arcstat_anon_evictable_data
;
521 * Number of bytes consumed by ARC buffers that meet the
522 * following criteria: backing buffers of type ARC_BUFC_METADATA,
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_metadata
;
528 * Total number of bytes consumed by ARC buffers residing in the
529 * arc_mru state. This includes *all* buffers in the arc_mru
530 * state; e.g. data, metadata, evictable, and unevictable buffers
531 * are all included in this value.
533 kstat_named_t arcstat_mru_size
;
535 * Number of bytes consumed by ARC buffers that meet the
536 * following criteria: backing buffers of type ARC_BUFC_DATA,
537 * residing in the arc_mru state, and are eligible for eviction
538 * (e.g. have no outstanding holds on the buffer).
540 kstat_named_t arcstat_mru_evictable_data
;
542 * Number of bytes consumed by ARC buffers that meet the
543 * following criteria: backing buffers of type ARC_BUFC_METADATA,
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_metadata
;
549 * Total number of bytes that *would have been* consumed by ARC
550 * buffers in the arc_mru_ghost state. The key thing to note
551 * here, is the fact that this size doesn't actually indicate
552 * RAM consumption. The ghost lists only consist of headers and
553 * don't actually have ARC buffers linked off of these headers.
554 * Thus, *if* the headers had associated ARC buffers, these
555 * buffers *would have* consumed this number of bytes.
557 kstat_named_t arcstat_mru_ghost_size
;
559 * Number of bytes that *would have been* consumed by ARC
560 * buffers that are eligible for eviction, of type
561 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
563 kstat_named_t arcstat_mru_ghost_evictable_data
;
565 * Number of bytes that *would have been* consumed by ARC
566 * buffers that are eligible for eviction, of type
567 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
569 kstat_named_t arcstat_mru_ghost_evictable_metadata
;
571 * Total number of bytes consumed by ARC buffers residing in the
572 * arc_mfu state. This includes *all* buffers in the arc_mfu
573 * state; e.g. data, metadata, evictable, and unevictable buffers
574 * are all included in this value.
576 kstat_named_t arcstat_mfu_size
;
578 * Number of bytes consumed by ARC buffers that are eligible for
579 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
582 kstat_named_t arcstat_mfu_evictable_data
;
584 * Number of bytes consumed by ARC buffers that are eligible for
585 * eviction, of type ARC_BUFC_METADATA, and reside in the
588 kstat_named_t arcstat_mfu_evictable_metadata
;
590 * Total number of bytes that *would have been* consumed by ARC
591 * buffers in the arc_mfu_ghost state. See the comment above
592 * arcstat_mru_ghost_size for more details.
594 kstat_named_t arcstat_mfu_ghost_size
;
596 * Number of bytes that *would have been* consumed by ARC
597 * buffers that are eligible for eviction, of type
598 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
600 kstat_named_t arcstat_mfu_ghost_evictable_data
;
602 * Number of bytes that *would have been* consumed by ARC
603 * buffers that are eligible for eviction, of type
604 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
606 kstat_named_t arcstat_mfu_ghost_evictable_metadata
;
607 kstat_named_t arcstat_l2_hits
;
608 kstat_named_t arcstat_l2_misses
;
609 kstat_named_t arcstat_l2_feeds
;
610 kstat_named_t arcstat_l2_rw_clash
;
611 kstat_named_t arcstat_l2_read_bytes
;
612 kstat_named_t arcstat_l2_write_bytes
;
613 kstat_named_t arcstat_l2_writes_sent
;
614 kstat_named_t arcstat_l2_writes_done
;
615 kstat_named_t arcstat_l2_writes_error
;
616 kstat_named_t arcstat_l2_writes_lock_retry
;
617 kstat_named_t arcstat_l2_evict_lock_retry
;
618 kstat_named_t arcstat_l2_evict_reading
;
619 kstat_named_t arcstat_l2_evict_l1cached
;
620 kstat_named_t arcstat_l2_free_on_write
;
621 kstat_named_t arcstat_l2_abort_lowmem
;
622 kstat_named_t arcstat_l2_cksum_bad
;
623 kstat_named_t arcstat_l2_io_error
;
624 kstat_named_t arcstat_l2_size
;
625 kstat_named_t arcstat_l2_asize
;
626 kstat_named_t arcstat_l2_hdr_size
;
627 kstat_named_t arcstat_memory_throttle_count
;
628 kstat_named_t arcstat_memory_direct_count
;
629 kstat_named_t arcstat_memory_indirect_count
;
630 kstat_named_t arcstat_no_grow
;
631 kstat_named_t arcstat_tempreserve
;
632 kstat_named_t arcstat_loaned_bytes
;
633 kstat_named_t arcstat_prune
;
634 kstat_named_t arcstat_meta_used
;
635 kstat_named_t arcstat_meta_limit
;
636 kstat_named_t arcstat_dnode_limit
;
637 kstat_named_t arcstat_meta_max
;
638 kstat_named_t arcstat_meta_min
;
639 kstat_named_t arcstat_sync_wait_for_async
;
640 kstat_named_t arcstat_demand_hit_predictive_prefetch
;
641 kstat_named_t arcstat_need_free
;
642 kstat_named_t arcstat_sys_free
;
645 static arc_stats_t arc_stats
= {
646 { "hits", KSTAT_DATA_UINT64
},
647 { "misses", KSTAT_DATA_UINT64
},
648 { "demand_data_hits", KSTAT_DATA_UINT64
},
649 { "demand_data_misses", KSTAT_DATA_UINT64
},
650 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
651 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
652 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
653 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
654 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
655 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
656 { "mru_hits", KSTAT_DATA_UINT64
},
657 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
658 { "mfu_hits", KSTAT_DATA_UINT64
},
659 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
660 { "deleted", KSTAT_DATA_UINT64
},
661 { "mutex_miss", KSTAT_DATA_UINT64
},
662 { "evict_skip", KSTAT_DATA_UINT64
},
663 { "evict_not_enough", KSTAT_DATA_UINT64
},
664 { "evict_l2_cached", KSTAT_DATA_UINT64
},
665 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
666 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
667 { "evict_l2_skip", KSTAT_DATA_UINT64
},
668 { "hash_elements", KSTAT_DATA_UINT64
},
669 { "hash_elements_max", KSTAT_DATA_UINT64
},
670 { "hash_collisions", KSTAT_DATA_UINT64
},
671 { "hash_chains", KSTAT_DATA_UINT64
},
672 { "hash_chain_max", KSTAT_DATA_UINT64
},
673 { "p", KSTAT_DATA_UINT64
},
674 { "c", KSTAT_DATA_UINT64
},
675 { "c_min", KSTAT_DATA_UINT64
},
676 { "c_max", KSTAT_DATA_UINT64
},
677 { "size", KSTAT_DATA_UINT64
},
678 { "compressed_size", KSTAT_DATA_UINT64
},
679 { "uncompressed_size", KSTAT_DATA_UINT64
},
680 { "overhead_size", KSTAT_DATA_UINT64
},
681 { "hdr_size", KSTAT_DATA_UINT64
},
682 { "data_size", KSTAT_DATA_UINT64
},
683 { "metadata_size", KSTAT_DATA_UINT64
},
684 { "dbuf_size", KSTAT_DATA_UINT64
},
685 { "dnode_size", KSTAT_DATA_UINT64
},
686 { "bonus_size", KSTAT_DATA_UINT64
},
687 { "anon_size", KSTAT_DATA_UINT64
},
688 { "anon_evictable_data", KSTAT_DATA_UINT64
},
689 { "anon_evictable_metadata", KSTAT_DATA_UINT64
},
690 { "mru_size", KSTAT_DATA_UINT64
},
691 { "mru_evictable_data", KSTAT_DATA_UINT64
},
692 { "mru_evictable_metadata", KSTAT_DATA_UINT64
},
693 { "mru_ghost_size", KSTAT_DATA_UINT64
},
694 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64
},
695 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
696 { "mfu_size", KSTAT_DATA_UINT64
},
697 { "mfu_evictable_data", KSTAT_DATA_UINT64
},
698 { "mfu_evictable_metadata", KSTAT_DATA_UINT64
},
699 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
700 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64
},
701 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
702 { "l2_hits", KSTAT_DATA_UINT64
},
703 { "l2_misses", KSTAT_DATA_UINT64
},
704 { "l2_feeds", KSTAT_DATA_UINT64
},
705 { "l2_rw_clash", KSTAT_DATA_UINT64
},
706 { "l2_read_bytes", KSTAT_DATA_UINT64
},
707 { "l2_write_bytes", KSTAT_DATA_UINT64
},
708 { "l2_writes_sent", KSTAT_DATA_UINT64
},
709 { "l2_writes_done", KSTAT_DATA_UINT64
},
710 { "l2_writes_error", KSTAT_DATA_UINT64
},
711 { "l2_writes_lock_retry", KSTAT_DATA_UINT64
},
712 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
713 { "l2_evict_reading", KSTAT_DATA_UINT64
},
714 { "l2_evict_l1cached", KSTAT_DATA_UINT64
},
715 { "l2_free_on_write", KSTAT_DATA_UINT64
},
716 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
717 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
718 { "l2_io_error", KSTAT_DATA_UINT64
},
719 { "l2_size", KSTAT_DATA_UINT64
},
720 { "l2_asize", KSTAT_DATA_UINT64
},
721 { "l2_hdr_size", KSTAT_DATA_UINT64
},
722 { "memory_throttle_count", KSTAT_DATA_UINT64
},
723 { "memory_direct_count", KSTAT_DATA_UINT64
},
724 { "memory_indirect_count", KSTAT_DATA_UINT64
},
725 { "arc_no_grow", KSTAT_DATA_UINT64
},
726 { "arc_tempreserve", KSTAT_DATA_UINT64
},
727 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
728 { "arc_prune", KSTAT_DATA_UINT64
},
729 { "arc_meta_used", KSTAT_DATA_UINT64
},
730 { "arc_meta_limit", KSTAT_DATA_UINT64
},
731 { "arc_dnode_limit", KSTAT_DATA_UINT64
},
732 { "arc_meta_max", KSTAT_DATA_UINT64
},
733 { "arc_meta_min", KSTAT_DATA_UINT64
},
734 { "sync_wait_for_async", KSTAT_DATA_UINT64
},
735 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64
},
736 { "arc_need_free", KSTAT_DATA_UINT64
},
737 { "arc_sys_free", KSTAT_DATA_UINT64
}
740 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
742 #define ARCSTAT_INCR(stat, val) \
743 atomic_add_64(&arc_stats.stat.value.ui64, (val))
745 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
746 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
748 #define ARCSTAT_MAX(stat, val) { \
750 while ((val) > (m = arc_stats.stat.value.ui64) && \
751 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
755 #define ARCSTAT_MAXSTAT(stat) \
756 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
759 * We define a macro to allow ARC hits/misses to be easily broken down by
760 * two separate conditions, giving a total of four different subtypes for
761 * each of hits and misses (so eight statistics total).
763 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
766 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
768 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
772 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
774 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
779 static arc_state_t
*arc_anon
;
780 static arc_state_t
*arc_mru
;
781 static arc_state_t
*arc_mru_ghost
;
782 static arc_state_t
*arc_mfu
;
783 static arc_state_t
*arc_mfu_ghost
;
784 static arc_state_t
*arc_l2c_only
;
787 * There are several ARC variables that are critical to export as kstats --
788 * but we don't want to have to grovel around in the kstat whenever we wish to
789 * manipulate them. For these variables, we therefore define them to be in
790 * terms of the statistic variable. This assures that we are not introducing
791 * the possibility of inconsistency by having shadow copies of the variables,
792 * while still allowing the code to be readable.
794 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
795 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
796 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
797 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
798 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
799 #define arc_no_grow ARCSTAT(arcstat_no_grow) /* do not grow cache size */
800 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
801 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
802 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
803 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
804 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
805 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
806 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
807 #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
808 #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */
809 #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
810 #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */
811 #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */
813 /* compressed size of entire arc */
814 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
815 /* uncompressed size of entire arc */
816 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
817 /* number of bytes in the arc from arc_buf_t's */
818 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
820 static list_t arc_prune_list
;
821 static kmutex_t arc_prune_mtx
;
822 static taskq_t
*arc_prune_taskq
;
824 #define GHOST_STATE(state) \
825 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
826 (state) == arc_l2c_only)
828 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
829 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
830 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
831 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
832 #define HDR_COMPRESSION_ENABLED(hdr) \
833 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
835 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
836 #define HDR_L2_READING(hdr) \
837 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
838 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
839 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
840 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
841 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
842 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
844 #define HDR_ISTYPE_METADATA(hdr) \
845 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
846 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
848 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
849 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
851 /* For storing compression mode in b_flags */
852 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
854 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
855 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
856 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
857 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
859 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
860 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
861 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
867 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
868 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
871 * Hash table routines
874 #define HT_LOCK_ALIGN 64
875 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
880 unsigned char pad
[HT_LOCK_PAD
];
884 #define BUF_LOCKS 8192
885 typedef struct buf_hash_table
{
887 arc_buf_hdr_t
**ht_table
;
888 struct ht_lock ht_locks
[BUF_LOCKS
];
891 static buf_hash_table_t buf_hash_table
;
893 #define BUF_HASH_INDEX(spa, dva, birth) \
894 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
895 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
896 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
897 #define HDR_LOCK(hdr) \
898 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
900 uint64_t zfs_crc64_table
[256];
906 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
907 #define L2ARC_HEADROOM 2 /* num of writes */
910 * If we discover during ARC scan any buffers to be compressed, we boost
911 * our headroom for the next scanning cycle by this percentage multiple.
913 #define L2ARC_HEADROOM_BOOST 200
914 #define L2ARC_FEED_SECS 1 /* caching interval secs */
915 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
918 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
919 * and each of the state has two types: data and metadata.
921 #define L2ARC_FEED_TYPES 4
923 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
924 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
926 /* L2ARC Performance Tunables */
927 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
928 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
929 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
930 unsigned long l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
931 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
932 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
933 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
934 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
935 int l2arc_norw
= B_FALSE
; /* no reads during writes */
940 static list_t L2ARC_dev_list
; /* device list */
941 static list_t
*l2arc_dev_list
; /* device list pointer */
942 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
943 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
944 static list_t L2ARC_free_on_write
; /* free after write buf list */
945 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
946 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
947 static uint64_t l2arc_ndev
; /* number of devices */
949 typedef struct l2arc_read_callback
{
950 arc_buf_hdr_t
*l2rcb_hdr
; /* read header */
951 blkptr_t l2rcb_bp
; /* original blkptr */
952 zbookmark_phys_t l2rcb_zb
; /* original bookmark */
953 int l2rcb_flags
; /* original flags */
954 } l2arc_read_callback_t
;
956 typedef struct l2arc_data_free
{
957 /* protected by l2arc_free_on_write_mtx */
960 arc_buf_contents_t l2df_type
;
961 list_node_t l2df_list_node
;
964 static kmutex_t l2arc_feed_thr_lock
;
965 static kcondvar_t l2arc_feed_thr_cv
;
966 static uint8_t l2arc_thread_exit
;
968 static abd_t
*arc_get_data_abd(arc_buf_hdr_t
*, uint64_t, void *);
969 static void *arc_get_data_buf(arc_buf_hdr_t
*, uint64_t, void *);
970 static void arc_get_data_impl(arc_buf_hdr_t
*, uint64_t, void *);
971 static void arc_free_data_abd(arc_buf_hdr_t
*, abd_t
*, uint64_t, void *);
972 static void arc_free_data_buf(arc_buf_hdr_t
*, void *, uint64_t, void *);
973 static void arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
);
974 static void arc_hdr_free_pabd(arc_buf_hdr_t
*);
975 static void arc_hdr_alloc_pabd(arc_buf_hdr_t
*);
976 static void arc_access(arc_buf_hdr_t
*, kmutex_t
*);
977 static boolean_t
arc_is_overflowing(void);
978 static void arc_buf_watch(arc_buf_t
*);
979 static void arc_tuning_update(void);
980 static void arc_prune_async(int64_t);
981 static uint64_t arc_all_memory(void);
983 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t
*);
984 static uint32_t arc_bufc_to_flags(arc_buf_contents_t
);
985 static inline void arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
986 static inline void arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
988 static boolean_t
l2arc_write_eligible(uint64_t, arc_buf_hdr_t
*);
989 static void l2arc_read_done(zio_t
*);
992 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
994 uint8_t *vdva
= (uint8_t *)dva
;
995 uint64_t crc
= -1ULL;
998 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
1000 for (i
= 0; i
< sizeof (dva_t
); i
++)
1001 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
1003 crc
^= (spa
>>8) ^ birth
;
1008 #define HDR_EMPTY(hdr) \
1009 ((hdr)->b_dva.dva_word[0] == 0 && \
1010 (hdr)->b_dva.dva_word[1] == 0)
1012 #define HDR_EQUAL(spa, dva, birth, hdr) \
1013 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1014 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1015 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1018 buf_discard_identity(arc_buf_hdr_t
*hdr
)
1020 hdr
->b_dva
.dva_word
[0] = 0;
1021 hdr
->b_dva
.dva_word
[1] = 0;
1025 static arc_buf_hdr_t
*
1026 buf_hash_find(uint64_t spa
, const blkptr_t
*bp
, kmutex_t
**lockp
)
1028 const dva_t
*dva
= BP_IDENTITY(bp
);
1029 uint64_t birth
= BP_PHYSICAL_BIRTH(bp
);
1030 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
1031 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1034 mutex_enter(hash_lock
);
1035 for (hdr
= buf_hash_table
.ht_table
[idx
]; hdr
!= NULL
;
1036 hdr
= hdr
->b_hash_next
) {
1037 if (HDR_EQUAL(spa
, dva
, birth
, hdr
)) {
1042 mutex_exit(hash_lock
);
1048 * Insert an entry into the hash table. If there is already an element
1049 * equal to elem in the hash table, then the already existing element
1050 * will be returned and the new element will not be inserted.
1051 * Otherwise returns NULL.
1052 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1054 static arc_buf_hdr_t
*
1055 buf_hash_insert(arc_buf_hdr_t
*hdr
, kmutex_t
**lockp
)
1057 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1058 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1059 arc_buf_hdr_t
*fhdr
;
1062 ASSERT(!DVA_IS_EMPTY(&hdr
->b_dva
));
1063 ASSERT(hdr
->b_birth
!= 0);
1064 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
1066 if (lockp
!= NULL
) {
1068 mutex_enter(hash_lock
);
1070 ASSERT(MUTEX_HELD(hash_lock
));
1073 for (fhdr
= buf_hash_table
.ht_table
[idx
], i
= 0; fhdr
!= NULL
;
1074 fhdr
= fhdr
->b_hash_next
, i
++) {
1075 if (HDR_EQUAL(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
, fhdr
))
1079 hdr
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
1080 buf_hash_table
.ht_table
[idx
] = hdr
;
1081 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1083 /* collect some hash table performance data */
1085 ARCSTAT_BUMP(arcstat_hash_collisions
);
1087 ARCSTAT_BUMP(arcstat_hash_chains
);
1089 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
1092 ARCSTAT_BUMP(arcstat_hash_elements
);
1093 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
1099 buf_hash_remove(arc_buf_hdr_t
*hdr
)
1101 arc_buf_hdr_t
*fhdr
, **hdrp
;
1102 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1104 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
1105 ASSERT(HDR_IN_HASH_TABLE(hdr
));
1107 hdrp
= &buf_hash_table
.ht_table
[idx
];
1108 while ((fhdr
= *hdrp
) != hdr
) {
1109 ASSERT3P(fhdr
, !=, NULL
);
1110 hdrp
= &fhdr
->b_hash_next
;
1112 *hdrp
= hdr
->b_hash_next
;
1113 hdr
->b_hash_next
= NULL
;
1114 arc_hdr_clear_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1116 /* collect some hash table performance data */
1117 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
1119 if (buf_hash_table
.ht_table
[idx
] &&
1120 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
1121 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
1125 * Global data structures and functions for the buf kmem cache.
1127 static kmem_cache_t
*hdr_full_cache
;
1128 static kmem_cache_t
*hdr_l2only_cache
;
1129 static kmem_cache_t
*buf_cache
;
1136 #if defined(_KERNEL) && defined(HAVE_SPL)
1138 * Large allocations which do not require contiguous pages
1139 * should be using vmem_free() in the linux kernel\
1141 vmem_free(buf_hash_table
.ht_table
,
1142 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1144 kmem_free(buf_hash_table
.ht_table
,
1145 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1147 for (i
= 0; i
< BUF_LOCKS
; i
++)
1148 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
1149 kmem_cache_destroy(hdr_full_cache
);
1150 kmem_cache_destroy(hdr_l2only_cache
);
1151 kmem_cache_destroy(buf_cache
);
1155 * Constructor callback - called when the cache is empty
1156 * and a new buf is requested.
1160 hdr_full_cons(void *vbuf
, void *unused
, int kmflag
)
1162 arc_buf_hdr_t
*hdr
= vbuf
;
1164 bzero(hdr
, HDR_FULL_SIZE
);
1165 cv_init(&hdr
->b_l1hdr
.b_cv
, NULL
, CV_DEFAULT
, NULL
);
1166 refcount_create(&hdr
->b_l1hdr
.b_refcnt
);
1167 mutex_init(&hdr
->b_l1hdr
.b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1168 list_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1169 list_link_init(&hdr
->b_l2hdr
.b_l2node
);
1170 multilist_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1171 arc_space_consume(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1178 hdr_l2only_cons(void *vbuf
, void *unused
, int kmflag
)
1180 arc_buf_hdr_t
*hdr
= vbuf
;
1182 bzero(hdr
, HDR_L2ONLY_SIZE
);
1183 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1190 buf_cons(void *vbuf
, void *unused
, int kmflag
)
1192 arc_buf_t
*buf
= vbuf
;
1194 bzero(buf
, sizeof (arc_buf_t
));
1195 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1196 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1202 * Destructor callback - called when a cached buf is
1203 * no longer required.
1207 hdr_full_dest(void *vbuf
, void *unused
)
1209 arc_buf_hdr_t
*hdr
= vbuf
;
1211 ASSERT(HDR_EMPTY(hdr
));
1212 cv_destroy(&hdr
->b_l1hdr
.b_cv
);
1213 refcount_destroy(&hdr
->b_l1hdr
.b_refcnt
);
1214 mutex_destroy(&hdr
->b_l1hdr
.b_freeze_lock
);
1215 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1216 arc_space_return(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1221 hdr_l2only_dest(void *vbuf
, void *unused
)
1223 ASSERTV(arc_buf_hdr_t
*hdr
= vbuf
);
1225 ASSERT(HDR_EMPTY(hdr
));
1226 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1231 buf_dest(void *vbuf
, void *unused
)
1233 arc_buf_t
*buf
= vbuf
;
1235 mutex_destroy(&buf
->b_evict_lock
);
1236 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1240 * Reclaim callback -- invoked when memory is low.
1244 hdr_recl(void *unused
)
1246 dprintf("hdr_recl called\n");
1248 * umem calls the reclaim func when we destroy the buf cache,
1249 * which is after we do arc_fini().
1252 cv_signal(&arc_reclaim_thread_cv
);
1258 uint64_t *ct
= NULL
;
1259 uint64_t hsize
= 1ULL << 12;
1263 * The hash table is big enough to fill all of physical memory
1264 * with an average block size of zfs_arc_average_blocksize (default 8K).
1265 * By default, the table will take up
1266 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1268 while (hsize
* zfs_arc_average_blocksize
< arc_all_memory())
1271 buf_hash_table
.ht_mask
= hsize
- 1;
1272 #if defined(_KERNEL) && defined(HAVE_SPL)
1274 * Large allocations which do not require contiguous pages
1275 * should be using vmem_alloc() in the linux kernel
1277 buf_hash_table
.ht_table
=
1278 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
1280 buf_hash_table
.ht_table
=
1281 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
1283 if (buf_hash_table
.ht_table
== NULL
) {
1284 ASSERT(hsize
> (1ULL << 8));
1289 hdr_full_cache
= kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE
,
1290 0, hdr_full_cons
, hdr_full_dest
, hdr_recl
, NULL
, NULL
, 0);
1291 hdr_l2only_cache
= kmem_cache_create("arc_buf_hdr_t_l2only",
1292 HDR_L2ONLY_SIZE
, 0, hdr_l2only_cons
, hdr_l2only_dest
, hdr_recl
,
1294 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
1295 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
1297 for (i
= 0; i
< 256; i
++)
1298 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
1299 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
1301 for (i
= 0; i
< BUF_LOCKS
; i
++) {
1302 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
1303 NULL
, MUTEX_DEFAULT
, NULL
);
1307 #define ARC_MINTIME (hz>>4) /* 62 ms */
1310 * This is the size that the buf occupies in memory. If the buf is compressed,
1311 * it will correspond to the compressed size. You should use this method of
1312 * getting the buf size unless you explicitly need the logical size.
1315 arc_buf_size(arc_buf_t
*buf
)
1317 return (ARC_BUF_COMPRESSED(buf
) ?
1318 HDR_GET_PSIZE(buf
->b_hdr
) : HDR_GET_LSIZE(buf
->b_hdr
));
1322 arc_buf_lsize(arc_buf_t
*buf
)
1324 return (HDR_GET_LSIZE(buf
->b_hdr
));
1328 arc_get_compression(arc_buf_t
*buf
)
1330 return (ARC_BUF_COMPRESSED(buf
) ?
1331 HDR_GET_COMPRESS(buf
->b_hdr
) : ZIO_COMPRESS_OFF
);
1334 static inline boolean_t
1335 arc_buf_is_shared(arc_buf_t
*buf
)
1337 boolean_t shared
= (buf
->b_data
!= NULL
&&
1338 buf
->b_hdr
->b_l1hdr
.b_pabd
!= NULL
&&
1339 abd_is_linear(buf
->b_hdr
->b_l1hdr
.b_pabd
) &&
1340 buf
->b_data
== abd_to_buf(buf
->b_hdr
->b_l1hdr
.b_pabd
));
1341 IMPLY(shared
, HDR_SHARED_DATA(buf
->b_hdr
));
1342 IMPLY(shared
, ARC_BUF_SHARED(buf
));
1343 IMPLY(shared
, ARC_BUF_COMPRESSED(buf
) || ARC_BUF_LAST(buf
));
1346 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1347 * already being shared" requirement prevents us from doing that.
1354 * Free the checksum associated with this header. If there is no checksum, this
1358 arc_cksum_free(arc_buf_hdr_t
*hdr
)
1360 ASSERT(HDR_HAS_L1HDR(hdr
));
1361 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1362 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1363 kmem_free(hdr
->b_l1hdr
.b_freeze_cksum
, sizeof (zio_cksum_t
));
1364 hdr
->b_l1hdr
.b_freeze_cksum
= NULL
;
1366 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1370 * Return true iff at least one of the bufs on hdr is not compressed.
1373 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t
*hdr
)
1375 for (arc_buf_t
*b
= hdr
->b_l1hdr
.b_buf
; b
!= NULL
; b
= b
->b_next
) {
1376 if (!ARC_BUF_COMPRESSED(b
)) {
1385 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1386 * matches the checksum that is stored in the hdr. If there is no checksum,
1387 * or if the buf is compressed, this is a no-op.
1390 arc_cksum_verify(arc_buf_t
*buf
)
1392 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1395 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1398 if (ARC_BUF_COMPRESSED(buf
)) {
1399 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1400 arc_hdr_has_uncompressed_buf(hdr
));
1404 ASSERT(HDR_HAS_L1HDR(hdr
));
1406 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1407 if (hdr
->b_l1hdr
.b_freeze_cksum
== NULL
|| HDR_IO_ERROR(hdr
)) {
1408 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1412 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
, &zc
);
1413 if (!ZIO_CHECKSUM_EQUAL(*hdr
->b_l1hdr
.b_freeze_cksum
, zc
))
1414 panic("buffer modified while frozen!");
1415 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1419 arc_cksum_is_equal(arc_buf_hdr_t
*hdr
, zio_t
*zio
)
1421 enum zio_compress compress
= BP_GET_COMPRESS(zio
->io_bp
);
1422 boolean_t valid_cksum
;
1424 ASSERT(!BP_IS_EMBEDDED(zio
->io_bp
));
1425 VERIFY3U(BP_GET_PSIZE(zio
->io_bp
), ==, HDR_GET_PSIZE(hdr
));
1428 * We rely on the blkptr's checksum to determine if the block
1429 * is valid or not. When compressed arc is enabled, the l2arc
1430 * writes the block to the l2arc just as it appears in the pool.
1431 * This allows us to use the blkptr's checksum to validate the
1432 * data that we just read off of the l2arc without having to store
1433 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1434 * arc is disabled, then the data written to the l2arc is always
1435 * uncompressed and won't match the block as it exists in the main
1436 * pool. When this is the case, we must first compress it if it is
1437 * compressed on the main pool before we can validate the checksum.
1439 if (!HDR_COMPRESSION_ENABLED(hdr
) && compress
!= ZIO_COMPRESS_OFF
) {
1443 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1445 cbuf
= zio_buf_alloc(HDR_GET_PSIZE(hdr
));
1446 lsize
= HDR_GET_LSIZE(hdr
);
1447 csize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1449 ASSERT3U(csize
, <=, HDR_GET_PSIZE(hdr
));
1450 if (csize
< HDR_GET_PSIZE(hdr
)) {
1452 * Compressed blocks are always a multiple of the
1453 * smallest ashift in the pool. Ideally, we would
1454 * like to round up the csize to the next
1455 * spa_min_ashift but that value may have changed
1456 * since the block was last written. Instead,
1457 * we rely on the fact that the hdr's psize
1458 * was set to the psize of the block when it was
1459 * last written. We set the csize to that value
1460 * and zero out any part that should not contain
1463 bzero((char *)cbuf
+ csize
, HDR_GET_PSIZE(hdr
) - csize
);
1464 csize
= HDR_GET_PSIZE(hdr
);
1466 zio_push_transform(zio
, cbuf
, csize
, HDR_GET_PSIZE(hdr
), NULL
);
1470 * Block pointers always store the checksum for the logical data.
1471 * If the block pointer has the gang bit set, then the checksum
1472 * it represents is for the reconstituted data and not for an
1473 * individual gang member. The zio pipeline, however, must be able to
1474 * determine the checksum of each of the gang constituents so it
1475 * treats the checksum comparison differently than what we need
1476 * for l2arc blocks. This prevents us from using the
1477 * zio_checksum_error() interface directly. Instead we must call the
1478 * zio_checksum_error_impl() so that we can ensure the checksum is
1479 * generated using the correct checksum algorithm and accounts for the
1480 * logical I/O size and not just a gang fragment.
1482 valid_cksum
= (zio_checksum_error_impl(zio
->io_spa
, zio
->io_bp
,
1483 BP_GET_CHECKSUM(zio
->io_bp
), zio
->io_abd
, zio
->io_size
,
1484 zio
->io_offset
, NULL
) == 0);
1485 zio_pop_transforms(zio
);
1486 return (valid_cksum
);
1490 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1491 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1492 * isn't modified later on. If buf is compressed or there is already a checksum
1493 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1496 arc_cksum_compute(arc_buf_t
*buf
)
1498 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1500 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1503 ASSERT(HDR_HAS_L1HDR(hdr
));
1505 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1506 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1507 ASSERT(arc_hdr_has_uncompressed_buf(hdr
));
1508 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1510 } else if (ARC_BUF_COMPRESSED(buf
)) {
1511 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1515 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1516 hdr
->b_l1hdr
.b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
),
1518 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
,
1519 hdr
->b_l1hdr
.b_freeze_cksum
);
1520 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1526 arc_buf_sigsegv(int sig
, siginfo_t
*si
, void *unused
)
1528 panic("Got SIGSEGV at address: 0x%lx\n", (long)si
->si_addr
);
1534 arc_buf_unwatch(arc_buf_t
*buf
)
1538 ASSERT0(mprotect(buf
->b_data
, arc_buf_size(buf
),
1539 PROT_READ
| PROT_WRITE
));
1546 arc_buf_watch(arc_buf_t
*buf
)
1550 ASSERT0(mprotect(buf
->b_data
, arc_buf_size(buf
),
1555 static arc_buf_contents_t
1556 arc_buf_type(arc_buf_hdr_t
*hdr
)
1558 arc_buf_contents_t type
;
1559 if (HDR_ISTYPE_METADATA(hdr
)) {
1560 type
= ARC_BUFC_METADATA
;
1562 type
= ARC_BUFC_DATA
;
1564 VERIFY3U(hdr
->b_type
, ==, type
);
1569 arc_is_metadata(arc_buf_t
*buf
)
1571 return (HDR_ISTYPE_METADATA(buf
->b_hdr
) != 0);
1575 arc_bufc_to_flags(arc_buf_contents_t type
)
1579 /* metadata field is 0 if buffer contains normal data */
1581 case ARC_BUFC_METADATA
:
1582 return (ARC_FLAG_BUFC_METADATA
);
1586 panic("undefined ARC buffer type!");
1587 return ((uint32_t)-1);
1591 arc_buf_thaw(arc_buf_t
*buf
)
1593 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1595 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
1596 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
1598 arc_cksum_verify(buf
);
1601 * Compressed buffers do not manipulate the b_freeze_cksum or
1602 * allocate b_thawed.
1604 if (ARC_BUF_COMPRESSED(buf
)) {
1605 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1606 arc_hdr_has_uncompressed_buf(hdr
));
1610 ASSERT(HDR_HAS_L1HDR(hdr
));
1611 arc_cksum_free(hdr
);
1612 arc_buf_unwatch(buf
);
1616 arc_buf_freeze(arc_buf_t
*buf
)
1618 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1619 kmutex_t
*hash_lock
;
1621 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1624 if (ARC_BUF_COMPRESSED(buf
)) {
1625 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1626 arc_hdr_has_uncompressed_buf(hdr
));
1630 hash_lock
= HDR_LOCK(hdr
);
1631 mutex_enter(hash_lock
);
1633 ASSERT(HDR_HAS_L1HDR(hdr
));
1634 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
||
1635 hdr
->b_l1hdr
.b_state
== arc_anon
);
1636 arc_cksum_compute(buf
);
1637 mutex_exit(hash_lock
);
1641 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1642 * the following functions should be used to ensure that the flags are
1643 * updated in a thread-safe way. When manipulating the flags either
1644 * the hash_lock must be held or the hdr must be undiscoverable. This
1645 * ensures that we're not racing with any other threads when updating
1649 arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1651 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1652 hdr
->b_flags
|= flags
;
1656 arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1658 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1659 hdr
->b_flags
&= ~flags
;
1663 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1664 * done in a special way since we have to clear and set bits
1665 * at the same time. Consumers that wish to set the compression bits
1666 * must use this function to ensure that the flags are updated in
1667 * thread-safe manner.
1670 arc_hdr_set_compress(arc_buf_hdr_t
*hdr
, enum zio_compress cmp
)
1672 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1675 * Holes and embedded blocks will always have a psize = 0 so
1676 * we ignore the compression of the blkptr and set the
1677 * want to uncompress them. Mark them as uncompressed.
1679 if (!zfs_compressed_arc_enabled
|| HDR_GET_PSIZE(hdr
) == 0) {
1680 arc_hdr_clear_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1681 HDR_SET_COMPRESS(hdr
, ZIO_COMPRESS_OFF
);
1682 ASSERT(!HDR_COMPRESSION_ENABLED(hdr
));
1683 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1685 arc_hdr_set_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1686 HDR_SET_COMPRESS(hdr
, cmp
);
1687 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, cmp
);
1688 ASSERT(HDR_COMPRESSION_ENABLED(hdr
));
1693 * Looks for another buf on the same hdr which has the data decompressed, copies
1694 * from it, and returns true. If no such buf exists, returns false.
1697 arc_buf_try_copy_decompressed_data(arc_buf_t
*buf
)
1699 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1700 boolean_t copied
= B_FALSE
;
1702 ASSERT(HDR_HAS_L1HDR(hdr
));
1703 ASSERT3P(buf
->b_data
, !=, NULL
);
1704 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1706 for (arc_buf_t
*from
= hdr
->b_l1hdr
.b_buf
; from
!= NULL
;
1707 from
= from
->b_next
) {
1708 /* can't use our own data buffer */
1713 if (!ARC_BUF_COMPRESSED(from
)) {
1714 bcopy(from
->b_data
, buf
->b_data
, arc_buf_size(buf
));
1721 * There were no decompressed bufs, so there should not be a
1722 * checksum on the hdr either.
1724 EQUIV(!copied
, hdr
->b_l1hdr
.b_freeze_cksum
== NULL
);
1730 * Given a buf that has a data buffer attached to it, this function will
1731 * efficiently fill the buf with data of the specified compression setting from
1732 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1733 * are already sharing a data buf, no copy is performed.
1735 * If the buf is marked as compressed but uncompressed data was requested, this
1736 * will allocate a new data buffer for the buf, remove that flag, and fill the
1737 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1738 * uncompressed data, and (since we haven't added support for it yet) if you
1739 * want compressed data your buf must already be marked as compressed and have
1740 * the correct-sized data buffer.
1743 arc_buf_fill(arc_buf_t
*buf
, boolean_t compressed
)
1745 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1746 boolean_t hdr_compressed
= (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
1747 dmu_object_byteswap_t bswap
= hdr
->b_l1hdr
.b_byteswap
;
1749 ASSERT3P(buf
->b_data
, !=, NULL
);
1750 IMPLY(compressed
, hdr_compressed
);
1751 IMPLY(compressed
, ARC_BUF_COMPRESSED(buf
));
1753 if (hdr_compressed
== compressed
) {
1754 if (!arc_buf_is_shared(buf
)) {
1755 abd_copy_to_buf(buf
->b_data
, hdr
->b_l1hdr
.b_pabd
,
1759 ASSERT(hdr_compressed
);
1760 ASSERT(!compressed
);
1761 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, HDR_GET_PSIZE(hdr
));
1764 * If the buf is sharing its data with the hdr, unlink it and
1765 * allocate a new data buffer for the buf.
1767 if (arc_buf_is_shared(buf
)) {
1768 ASSERT(ARC_BUF_COMPRESSED(buf
));
1770 /* We need to give the buf it's own b_data */
1771 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
1773 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
1774 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
1776 /* Previously overhead was 0; just add new overhead */
1777 ARCSTAT_INCR(arcstat_overhead_size
, HDR_GET_LSIZE(hdr
));
1778 } else if (ARC_BUF_COMPRESSED(buf
)) {
1779 /* We need to reallocate the buf's b_data */
1780 arc_free_data_buf(hdr
, buf
->b_data
, HDR_GET_PSIZE(hdr
),
1783 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
1785 /* We increased the size of b_data; update overhead */
1786 ARCSTAT_INCR(arcstat_overhead_size
,
1787 HDR_GET_LSIZE(hdr
) - HDR_GET_PSIZE(hdr
));
1791 * Regardless of the buf's previous compression settings, it
1792 * should not be compressed at the end of this function.
1794 buf
->b_flags
&= ~ARC_BUF_FLAG_COMPRESSED
;
1797 * Try copying the data from another buf which already has a
1798 * decompressed version. If that's not possible, it's time to
1799 * bite the bullet and decompress the data from the hdr.
1801 if (arc_buf_try_copy_decompressed_data(buf
)) {
1802 /* Skip byteswapping and checksumming (already done) */
1803 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, !=, NULL
);
1806 int error
= zio_decompress_data(HDR_GET_COMPRESS(hdr
),
1807 hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
1808 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
1811 * Absent hardware errors or software bugs, this should
1812 * be impossible, but log it anyway so we can debug it.
1816 "hdr %p, compress %d, psize %d, lsize %d",
1817 hdr
, HDR_GET_COMPRESS(hdr
),
1818 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
1819 return (SET_ERROR(EIO
));
1824 /* Byteswap the buf's data if necessary */
1825 if (bswap
!= DMU_BSWAP_NUMFUNCS
) {
1826 ASSERT(!HDR_SHARED_DATA(hdr
));
1827 ASSERT3U(bswap
, <, DMU_BSWAP_NUMFUNCS
);
1828 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, HDR_GET_LSIZE(hdr
));
1831 /* Compute the hdr's checksum if necessary */
1832 arc_cksum_compute(buf
);
1838 arc_decompress(arc_buf_t
*buf
)
1840 return (arc_buf_fill(buf
, B_FALSE
));
1844 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1847 arc_hdr_size(arc_buf_hdr_t
*hdr
)
1851 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
1852 HDR_GET_PSIZE(hdr
) > 0) {
1853 size
= HDR_GET_PSIZE(hdr
);
1855 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, 0);
1856 size
= HDR_GET_LSIZE(hdr
);
1862 * Increment the amount of evictable space in the arc_state_t's refcount.
1863 * We account for the space used by the hdr and the arc buf individually
1864 * so that we can add and remove them from the refcount individually.
1867 arc_evictable_space_increment(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
1869 arc_buf_contents_t type
= arc_buf_type(hdr
);
1872 ASSERT(HDR_HAS_L1HDR(hdr
));
1874 if (GHOST_STATE(state
)) {
1875 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
1876 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1877 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
1878 (void) refcount_add_many(&state
->arcs_esize
[type
],
1879 HDR_GET_LSIZE(hdr
), hdr
);
1883 ASSERT(!GHOST_STATE(state
));
1884 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
1885 (void) refcount_add_many(&state
->arcs_esize
[type
],
1886 arc_hdr_size(hdr
), hdr
);
1888 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
; buf
= buf
->b_next
) {
1889 if (arc_buf_is_shared(buf
))
1891 (void) refcount_add_many(&state
->arcs_esize
[type
],
1892 arc_buf_size(buf
), buf
);
1897 * Decrement the amount of evictable space in the arc_state_t's refcount.
1898 * We account for the space used by the hdr and the arc buf individually
1899 * so that we can add and remove them from the refcount individually.
1902 arc_evictable_space_decrement(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
1904 arc_buf_contents_t type
= arc_buf_type(hdr
);
1907 ASSERT(HDR_HAS_L1HDR(hdr
));
1909 if (GHOST_STATE(state
)) {
1910 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
1911 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1912 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
1913 (void) refcount_remove_many(&state
->arcs_esize
[type
],
1914 HDR_GET_LSIZE(hdr
), hdr
);
1918 ASSERT(!GHOST_STATE(state
));
1919 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
1920 (void) refcount_remove_many(&state
->arcs_esize
[type
],
1921 arc_hdr_size(hdr
), hdr
);
1923 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
; buf
= buf
->b_next
) {
1924 if (arc_buf_is_shared(buf
))
1926 (void) refcount_remove_many(&state
->arcs_esize
[type
],
1927 arc_buf_size(buf
), buf
);
1932 * Add a reference to this hdr indicating that someone is actively
1933 * referencing that memory. When the refcount transitions from 0 to 1,
1934 * we remove it from the respective arc_state_t list to indicate that
1935 * it is not evictable.
1938 add_reference(arc_buf_hdr_t
*hdr
, void *tag
)
1942 ASSERT(HDR_HAS_L1HDR(hdr
));
1943 if (!MUTEX_HELD(HDR_LOCK(hdr
))) {
1944 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
1945 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
1946 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1949 state
= hdr
->b_l1hdr
.b_state
;
1951 if ((refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
) == 1) &&
1952 (state
!= arc_anon
)) {
1953 /* We don't use the L2-only state list. */
1954 if (state
!= arc_l2c_only
) {
1955 multilist_remove(state
->arcs_list
[arc_buf_type(hdr
)],
1957 arc_evictable_space_decrement(hdr
, state
);
1959 /* remove the prefetch flag if we get a reference */
1960 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
1965 * Remove a reference from this hdr. When the reference transitions from
1966 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
1967 * list making it eligible for eviction.
1970 remove_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1973 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
1975 ASSERT(HDR_HAS_L1HDR(hdr
));
1976 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
1977 ASSERT(!GHOST_STATE(state
));
1980 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1981 * check to prevent usage of the arc_l2c_only list.
1983 if (((cnt
= refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
)) == 0) &&
1984 (state
!= arc_anon
)) {
1985 multilist_insert(state
->arcs_list
[arc_buf_type(hdr
)], hdr
);
1986 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
1987 arc_evictable_space_increment(hdr
, state
);
1993 * Returns detailed information about a specific arc buffer. When the
1994 * state_index argument is set the function will calculate the arc header
1995 * list position for its arc state. Since this requires a linear traversal
1996 * callers are strongly encourage not to do this. However, it can be helpful
1997 * for targeted analysis so the functionality is provided.
2000 arc_buf_info(arc_buf_t
*ab
, arc_buf_info_t
*abi
, int state_index
)
2002 arc_buf_hdr_t
*hdr
= ab
->b_hdr
;
2003 l1arc_buf_hdr_t
*l1hdr
= NULL
;
2004 l2arc_buf_hdr_t
*l2hdr
= NULL
;
2005 arc_state_t
*state
= NULL
;
2007 memset(abi
, 0, sizeof (arc_buf_info_t
));
2012 abi
->abi_flags
= hdr
->b_flags
;
2014 if (HDR_HAS_L1HDR(hdr
)) {
2015 l1hdr
= &hdr
->b_l1hdr
;
2016 state
= l1hdr
->b_state
;
2018 if (HDR_HAS_L2HDR(hdr
))
2019 l2hdr
= &hdr
->b_l2hdr
;
2022 abi
->abi_bufcnt
= l1hdr
->b_bufcnt
;
2023 abi
->abi_access
= l1hdr
->b_arc_access
;
2024 abi
->abi_mru_hits
= l1hdr
->b_mru_hits
;
2025 abi
->abi_mru_ghost_hits
= l1hdr
->b_mru_ghost_hits
;
2026 abi
->abi_mfu_hits
= l1hdr
->b_mfu_hits
;
2027 abi
->abi_mfu_ghost_hits
= l1hdr
->b_mfu_ghost_hits
;
2028 abi
->abi_holds
= refcount_count(&l1hdr
->b_refcnt
);
2032 abi
->abi_l2arc_dattr
= l2hdr
->b_daddr
;
2033 abi
->abi_l2arc_hits
= l2hdr
->b_hits
;
2036 abi
->abi_state_type
= state
? state
->arcs_state
: ARC_STATE_ANON
;
2037 abi
->abi_state_contents
= arc_buf_type(hdr
);
2038 abi
->abi_size
= arc_hdr_size(hdr
);
2042 * Move the supplied buffer to the indicated state. The hash lock
2043 * for the buffer must be held by the caller.
2046 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*hdr
,
2047 kmutex_t
*hash_lock
)
2049 arc_state_t
*old_state
;
2052 boolean_t update_old
, update_new
;
2053 arc_buf_contents_t buftype
= arc_buf_type(hdr
);
2056 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2057 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2058 * L1 hdr doesn't always exist when we change state to arc_anon before
2059 * destroying a header, in which case reallocating to add the L1 hdr is
2062 if (HDR_HAS_L1HDR(hdr
)) {
2063 old_state
= hdr
->b_l1hdr
.b_state
;
2064 refcnt
= refcount_count(&hdr
->b_l1hdr
.b_refcnt
);
2065 bufcnt
= hdr
->b_l1hdr
.b_bufcnt
;
2066 update_old
= (bufcnt
> 0 || hdr
->b_l1hdr
.b_pabd
!= NULL
);
2068 old_state
= arc_l2c_only
;
2071 update_old
= B_FALSE
;
2073 update_new
= update_old
;
2075 ASSERT(MUTEX_HELD(hash_lock
));
2076 ASSERT3P(new_state
, !=, old_state
);
2077 ASSERT(!GHOST_STATE(new_state
) || bufcnt
== 0);
2078 ASSERT(old_state
!= arc_anon
|| bufcnt
<= 1);
2081 * If this buffer is evictable, transfer it from the
2082 * old state list to the new state list.
2085 if (old_state
!= arc_anon
&& old_state
!= arc_l2c_only
) {
2086 ASSERT(HDR_HAS_L1HDR(hdr
));
2087 multilist_remove(old_state
->arcs_list
[buftype
], hdr
);
2089 if (GHOST_STATE(old_state
)) {
2091 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2092 update_old
= B_TRUE
;
2094 arc_evictable_space_decrement(hdr
, old_state
);
2096 if (new_state
!= arc_anon
&& new_state
!= arc_l2c_only
) {
2098 * An L1 header always exists here, since if we're
2099 * moving to some L1-cached state (i.e. not l2c_only or
2100 * anonymous), we realloc the header to add an L1hdr
2103 ASSERT(HDR_HAS_L1HDR(hdr
));
2104 multilist_insert(new_state
->arcs_list
[buftype
], hdr
);
2106 if (GHOST_STATE(new_state
)) {
2108 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2109 update_new
= B_TRUE
;
2111 arc_evictable_space_increment(hdr
, new_state
);
2115 ASSERT(!HDR_EMPTY(hdr
));
2116 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(hdr
))
2117 buf_hash_remove(hdr
);
2119 /* adjust state sizes (ignore arc_l2c_only) */
2121 if (update_new
&& new_state
!= arc_l2c_only
) {
2122 ASSERT(HDR_HAS_L1HDR(hdr
));
2123 if (GHOST_STATE(new_state
)) {
2127 * When moving a header to a ghost state, we first
2128 * remove all arc buffers. Thus, we'll have a
2129 * bufcnt of zero, and no arc buffer to use for
2130 * the reference. As a result, we use the arc
2131 * header pointer for the reference.
2133 (void) refcount_add_many(&new_state
->arcs_size
,
2134 HDR_GET_LSIZE(hdr
), hdr
);
2135 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2138 uint32_t buffers
= 0;
2141 * Each individual buffer holds a unique reference,
2142 * thus we must remove each of these references one
2145 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2146 buf
= buf
->b_next
) {
2147 ASSERT3U(bufcnt
, !=, 0);
2151 * When the arc_buf_t is sharing the data
2152 * block with the hdr, the owner of the
2153 * reference belongs to the hdr. Only
2154 * add to the refcount if the arc_buf_t is
2157 if (arc_buf_is_shared(buf
))
2160 (void) refcount_add_many(&new_state
->arcs_size
,
2161 arc_buf_size(buf
), buf
);
2163 ASSERT3U(bufcnt
, ==, buffers
);
2165 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2166 (void) refcount_add_many(&new_state
->arcs_size
,
2167 arc_hdr_size(hdr
), hdr
);
2169 ASSERT(GHOST_STATE(old_state
));
2174 if (update_old
&& old_state
!= arc_l2c_only
) {
2175 ASSERT(HDR_HAS_L1HDR(hdr
));
2176 if (GHOST_STATE(old_state
)) {
2178 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2181 * When moving a header off of a ghost state,
2182 * the header will not contain any arc buffers.
2183 * We use the arc header pointer for the reference
2184 * which is exactly what we did when we put the
2185 * header on the ghost state.
2188 (void) refcount_remove_many(&old_state
->arcs_size
,
2189 HDR_GET_LSIZE(hdr
), hdr
);
2192 uint32_t buffers
= 0;
2195 * Each individual buffer holds a unique reference,
2196 * thus we must remove each of these references one
2199 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2200 buf
= buf
->b_next
) {
2201 ASSERT3U(bufcnt
, !=, 0);
2205 * When the arc_buf_t is sharing the data
2206 * block with the hdr, the owner of the
2207 * reference belongs to the hdr. Only
2208 * add to the refcount if the arc_buf_t is
2211 if (arc_buf_is_shared(buf
))
2214 (void) refcount_remove_many(
2215 &old_state
->arcs_size
, arc_buf_size(buf
),
2218 ASSERT3U(bufcnt
, ==, buffers
);
2219 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2220 (void) refcount_remove_many(
2221 &old_state
->arcs_size
, arc_hdr_size(hdr
), hdr
);
2225 if (HDR_HAS_L1HDR(hdr
))
2226 hdr
->b_l1hdr
.b_state
= new_state
;
2229 * L2 headers should never be on the L2 state list since they don't
2230 * have L1 headers allocated.
2232 ASSERT(multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]) &&
2233 multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]));
2237 arc_space_consume(uint64_t space
, arc_space_type_t type
)
2239 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2244 case ARC_SPACE_DATA
:
2245 ARCSTAT_INCR(arcstat_data_size
, space
);
2247 case ARC_SPACE_META
:
2248 ARCSTAT_INCR(arcstat_metadata_size
, space
);
2250 case ARC_SPACE_BONUS
:
2251 ARCSTAT_INCR(arcstat_bonus_size
, space
);
2253 case ARC_SPACE_DNODE
:
2254 ARCSTAT_INCR(arcstat_dnode_size
, space
);
2256 case ARC_SPACE_DBUF
:
2257 ARCSTAT_INCR(arcstat_dbuf_size
, space
);
2259 case ARC_SPACE_HDRS
:
2260 ARCSTAT_INCR(arcstat_hdr_size
, space
);
2262 case ARC_SPACE_L2HDRS
:
2263 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
2267 if (type
!= ARC_SPACE_DATA
)
2268 ARCSTAT_INCR(arcstat_meta_used
, space
);
2270 atomic_add_64(&arc_size
, space
);
2274 arc_space_return(uint64_t space
, arc_space_type_t type
)
2276 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2281 case ARC_SPACE_DATA
:
2282 ARCSTAT_INCR(arcstat_data_size
, -space
);
2284 case ARC_SPACE_META
:
2285 ARCSTAT_INCR(arcstat_metadata_size
, -space
);
2287 case ARC_SPACE_BONUS
:
2288 ARCSTAT_INCR(arcstat_bonus_size
, -space
);
2290 case ARC_SPACE_DNODE
:
2291 ARCSTAT_INCR(arcstat_dnode_size
, -space
);
2293 case ARC_SPACE_DBUF
:
2294 ARCSTAT_INCR(arcstat_dbuf_size
, -space
);
2296 case ARC_SPACE_HDRS
:
2297 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
2299 case ARC_SPACE_L2HDRS
:
2300 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
2304 if (type
!= ARC_SPACE_DATA
) {
2305 ASSERT(arc_meta_used
>= space
);
2306 if (arc_meta_max
< arc_meta_used
)
2307 arc_meta_max
= arc_meta_used
;
2308 ARCSTAT_INCR(arcstat_meta_used
, -space
);
2311 ASSERT(arc_size
>= space
);
2312 atomic_add_64(&arc_size
, -space
);
2316 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2317 * with the hdr's b_pabd.
2320 arc_can_share(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2323 * The criteria for sharing a hdr's data are:
2324 * 1. the hdr's compression matches the buf's compression
2325 * 2. the hdr doesn't need to be byteswapped
2326 * 3. the hdr isn't already being shared
2327 * 4. the buf is either compressed or it is the last buf in the hdr list
2329 * Criterion #4 maintains the invariant that shared uncompressed
2330 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2331 * might ask, "if a compressed buf is allocated first, won't that be the
2332 * last thing in the list?", but in that case it's impossible to create
2333 * a shared uncompressed buf anyway (because the hdr must be compressed
2334 * to have the compressed buf). You might also think that #3 is
2335 * sufficient to make this guarantee, however it's possible
2336 * (specifically in the rare L2ARC write race mentioned in
2337 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2338 * is sharable, but wasn't at the time of its allocation. Rather than
2339 * allow a new shared uncompressed buf to be created and then shuffle
2340 * the list around to make it the last element, this simply disallows
2341 * sharing if the new buf isn't the first to be added.
2343 ASSERT3P(buf
->b_hdr
, ==, hdr
);
2344 boolean_t hdr_compressed
= HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
;
2345 boolean_t buf_compressed
= ARC_BUF_COMPRESSED(buf
) != 0;
2346 return (buf_compressed
== hdr_compressed
&&
2347 hdr
->b_l1hdr
.b_byteswap
== DMU_BSWAP_NUMFUNCS
&&
2348 !HDR_SHARED_DATA(hdr
) &&
2349 (ARC_BUF_LAST(buf
) || ARC_BUF_COMPRESSED(buf
)));
2353 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2354 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2355 * copy was made successfully, or an error code otherwise.
2358 arc_buf_alloc_impl(arc_buf_hdr_t
*hdr
, void *tag
, boolean_t compressed
,
2359 boolean_t fill
, arc_buf_t
**ret
)
2363 ASSERT(HDR_HAS_L1HDR(hdr
));
2364 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2365 VERIFY(hdr
->b_type
== ARC_BUFC_DATA
||
2366 hdr
->b_type
== ARC_BUFC_METADATA
);
2367 ASSERT3P(ret
, !=, NULL
);
2368 ASSERT3P(*ret
, ==, NULL
);
2370 hdr
->b_l1hdr
.b_mru_hits
= 0;
2371 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
2372 hdr
->b_l1hdr
.b_mfu_hits
= 0;
2373 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
2374 hdr
->b_l1hdr
.b_l2_hits
= 0;
2376 buf
= *ret
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
2379 buf
->b_next
= hdr
->b_l1hdr
.b_buf
;
2382 add_reference(hdr
, tag
);
2385 * We're about to change the hdr's b_flags. We must either
2386 * hold the hash_lock or be undiscoverable.
2388 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2391 * Only honor requests for compressed bufs if the hdr is actually
2394 if (compressed
&& HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
)
2395 buf
->b_flags
|= ARC_BUF_FLAG_COMPRESSED
;
2398 * If the hdr's data can be shared then we share the data buffer and
2399 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2400 * allocate a new buffer to store the buf's data.
2402 * There are two additional restrictions here because we're sharing
2403 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2404 * actively involved in an L2ARC write, because if this buf is used by
2405 * an arc_write() then the hdr's data buffer will be released when the
2406 * write completes, even though the L2ARC write might still be using it.
2407 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2408 * need to be ABD-aware.
2410 boolean_t can_share
= arc_can_share(hdr
, buf
) && !HDR_L2_WRITING(hdr
) &&
2411 abd_is_linear(hdr
->b_l1hdr
.b_pabd
);
2413 /* Set up b_data and sharing */
2415 buf
->b_data
= abd_to_buf(hdr
->b_l1hdr
.b_pabd
);
2416 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2417 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2420 arc_get_data_buf(hdr
, arc_buf_size(buf
), buf
);
2421 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2423 VERIFY3P(buf
->b_data
, !=, NULL
);
2425 hdr
->b_l1hdr
.b_buf
= buf
;
2426 hdr
->b_l1hdr
.b_bufcnt
+= 1;
2429 * If the user wants the data from the hdr, we need to either copy or
2430 * decompress the data.
2433 return (arc_buf_fill(buf
, ARC_BUF_COMPRESSED(buf
) != 0));
2439 static char *arc_onloan_tag
= "onloan";
2442 arc_loaned_bytes_update(int64_t delta
)
2444 atomic_add_64(&arc_loaned_bytes
, delta
);
2446 /* assert that it did not wrap around */
2447 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
2451 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2452 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2453 * buffers must be returned to the arc before they can be used by the DMU or
2457 arc_loan_buf(spa_t
*spa
, boolean_t is_metadata
, int size
)
2459 arc_buf_t
*buf
= arc_alloc_buf(spa
, arc_onloan_tag
,
2460 is_metadata
? ARC_BUFC_METADATA
: ARC_BUFC_DATA
, size
);
2462 arc_loaned_bytes_update(size
);
2468 arc_loan_compressed_buf(spa_t
*spa
, uint64_t psize
, uint64_t lsize
,
2469 enum zio_compress compression_type
)
2471 arc_buf_t
*buf
= arc_alloc_compressed_buf(spa
, arc_onloan_tag
,
2472 psize
, lsize
, compression_type
);
2474 arc_loaned_bytes_update(psize
);
2481 * Return a loaned arc buffer to the arc.
2484 arc_return_buf(arc_buf_t
*buf
, void *tag
)
2486 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2488 ASSERT3P(buf
->b_data
, !=, NULL
);
2489 ASSERT(HDR_HAS_L1HDR(hdr
));
2490 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2491 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2493 arc_loaned_bytes_update(-arc_buf_size(buf
));
2496 /* Detach an arc_buf from a dbuf (tag) */
2498 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
2500 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2502 ASSERT3P(buf
->b_data
, !=, NULL
);
2503 ASSERT(HDR_HAS_L1HDR(hdr
));
2504 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2505 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2507 arc_loaned_bytes_update(arc_buf_size(buf
));
2511 l2arc_free_abd_on_write(abd_t
*abd
, size_t size
, arc_buf_contents_t type
)
2513 l2arc_data_free_t
*df
= kmem_alloc(sizeof (*df
), KM_SLEEP
);
2516 df
->l2df_size
= size
;
2517 df
->l2df_type
= type
;
2518 mutex_enter(&l2arc_free_on_write_mtx
);
2519 list_insert_head(l2arc_free_on_write
, df
);
2520 mutex_exit(&l2arc_free_on_write_mtx
);
2524 arc_hdr_free_on_write(arc_buf_hdr_t
*hdr
)
2526 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2527 arc_buf_contents_t type
= arc_buf_type(hdr
);
2528 uint64_t size
= arc_hdr_size(hdr
);
2530 /* protected by hash lock, if in the hash table */
2531 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
2532 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2533 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
2535 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2538 (void) refcount_remove_many(&state
->arcs_size
, size
, hdr
);
2539 if (type
== ARC_BUFC_METADATA
) {
2540 arc_space_return(size
, ARC_SPACE_META
);
2542 ASSERT(type
== ARC_BUFC_DATA
);
2543 arc_space_return(size
, ARC_SPACE_DATA
);
2546 l2arc_free_abd_on_write(hdr
->b_l1hdr
.b_pabd
, size
, type
);
2550 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2551 * data buffer, we transfer the refcount ownership to the hdr and update
2552 * the appropriate kstats.
2555 arc_share_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2557 ASSERT(arc_can_share(hdr
, buf
));
2558 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2559 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2562 * Start sharing the data buffer. We transfer the
2563 * refcount ownership to the hdr since it always owns
2564 * the refcount whenever an arc_buf_t is shared.
2566 refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, buf
, hdr
);
2567 hdr
->b_l1hdr
.b_pabd
= abd_get_from_buf(buf
->b_data
, arc_buf_size(buf
));
2568 abd_take_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
,
2569 HDR_ISTYPE_METADATA(hdr
));
2570 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2571 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2574 * Since we've transferred ownership to the hdr we need
2575 * to increment its compressed and uncompressed kstats and
2576 * decrement the overhead size.
2578 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
2579 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
2580 ARCSTAT_INCR(arcstat_overhead_size
, -arc_buf_size(buf
));
2584 arc_unshare_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2586 ASSERT(arc_buf_is_shared(buf
));
2587 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2588 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2591 * We are no longer sharing this buffer so we need
2592 * to transfer its ownership to the rightful owner.
2594 refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, hdr
, buf
);
2595 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2596 abd_release_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
);
2597 abd_put(hdr
->b_l1hdr
.b_pabd
);
2598 hdr
->b_l1hdr
.b_pabd
= NULL
;
2599 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
2602 * Since the buffer is no longer shared between
2603 * the arc buf and the hdr, count it as overhead.
2605 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
2606 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
2607 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2611 * Remove an arc_buf_t from the hdr's buf list and return the last
2612 * arc_buf_t on the list. If no buffers remain on the list then return
2616 arc_buf_remove(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2618 ASSERT(HDR_HAS_L1HDR(hdr
));
2619 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2621 arc_buf_t
**bufp
= &hdr
->b_l1hdr
.b_buf
;
2622 arc_buf_t
*lastbuf
= NULL
;
2625 * Remove the buf from the hdr list and locate the last
2626 * remaining buffer on the list.
2628 while (*bufp
!= NULL
) {
2630 *bufp
= buf
->b_next
;
2633 * If we've removed a buffer in the middle of
2634 * the list then update the lastbuf and update
2637 if (*bufp
!= NULL
) {
2639 bufp
= &(*bufp
)->b_next
;
2643 ASSERT3P(lastbuf
, !=, buf
);
2644 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, lastbuf
!= NULL
);
2645 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, hdr
->b_l1hdr
.b_buf
!= NULL
);
2646 IMPLY(lastbuf
!= NULL
, ARC_BUF_LAST(lastbuf
));
2652 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
2656 arc_buf_destroy_impl(arc_buf_t
*buf
)
2658 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2661 * Free up the data associated with the buf but only if we're not
2662 * sharing this with the hdr. If we are sharing it with the hdr, the
2663 * hdr is responsible for doing the free.
2665 if (buf
->b_data
!= NULL
) {
2667 * We're about to change the hdr's b_flags. We must either
2668 * hold the hash_lock or be undiscoverable.
2670 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2672 arc_cksum_verify(buf
);
2673 arc_buf_unwatch(buf
);
2675 if (arc_buf_is_shared(buf
)) {
2676 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2678 uint64_t size
= arc_buf_size(buf
);
2679 arc_free_data_buf(hdr
, buf
->b_data
, size
, buf
);
2680 ARCSTAT_INCR(arcstat_overhead_size
, -size
);
2684 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
2685 hdr
->b_l1hdr
.b_bufcnt
-= 1;
2688 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
2690 if (ARC_BUF_SHARED(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
2692 * If the current arc_buf_t is sharing its data buffer with the
2693 * hdr, then reassign the hdr's b_pabd to share it with the new
2694 * buffer at the end of the list. The shared buffer is always
2695 * the last one on the hdr's buffer list.
2697 * There is an equivalent case for compressed bufs, but since
2698 * they aren't guaranteed to be the last buf in the list and
2699 * that is an exceedingly rare case, we just allow that space be
2700 * wasted temporarily.
2702 if (lastbuf
!= NULL
) {
2703 /* Only one buf can be shared at once */
2704 VERIFY(!arc_buf_is_shared(lastbuf
));
2705 /* hdr is uncompressed so can't have compressed buf */
2706 VERIFY(!ARC_BUF_COMPRESSED(lastbuf
));
2708 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2709 arc_hdr_free_pabd(hdr
);
2712 * We must setup a new shared block between the
2713 * last buffer and the hdr. The data would have
2714 * been allocated by the arc buf so we need to transfer
2715 * ownership to the hdr since it's now being shared.
2717 arc_share_buf(hdr
, lastbuf
);
2719 } else if (HDR_SHARED_DATA(hdr
)) {
2721 * Uncompressed shared buffers are always at the end
2722 * of the list. Compressed buffers don't have the
2723 * same requirements. This makes it hard to
2724 * simply assert that the lastbuf is shared so
2725 * we rely on the hdr's compression flags to determine
2726 * if we have a compressed, shared buffer.
2728 ASSERT3P(lastbuf
, !=, NULL
);
2729 ASSERT(arc_buf_is_shared(lastbuf
) ||
2730 HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
2734 * Free the checksum if we're removing the last uncompressed buf from
2737 if (!arc_hdr_has_uncompressed_buf(hdr
)) {
2738 arc_cksum_free(hdr
);
2741 /* clean up the buf */
2743 kmem_cache_free(buf_cache
, buf
);
2747 arc_hdr_alloc_pabd(arc_buf_hdr_t
*hdr
)
2749 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2750 ASSERT(HDR_HAS_L1HDR(hdr
));
2751 ASSERT(!HDR_SHARED_DATA(hdr
));
2753 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2754 hdr
->b_l1hdr
.b_pabd
= arc_get_data_abd(hdr
, arc_hdr_size(hdr
), hdr
);
2755 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
2756 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2758 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
2759 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
2763 arc_hdr_free_pabd(arc_buf_hdr_t
*hdr
)
2765 ASSERT(HDR_HAS_L1HDR(hdr
));
2766 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2769 * If the hdr is currently being written to the l2arc then
2770 * we defer freeing the data by adding it to the l2arc_free_on_write
2771 * list. The l2arc will free the data once it's finished
2772 * writing it to the l2arc device.
2774 if (HDR_L2_WRITING(hdr
)) {
2775 arc_hdr_free_on_write(hdr
);
2776 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
2778 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
,
2779 arc_hdr_size(hdr
), hdr
);
2781 hdr
->b_l1hdr
.b_pabd
= NULL
;
2782 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
2784 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
2785 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
2788 static arc_buf_hdr_t
*
2789 arc_hdr_alloc(uint64_t spa
, int32_t psize
, int32_t lsize
,
2790 enum zio_compress compression_type
, arc_buf_contents_t type
)
2794 VERIFY(type
== ARC_BUFC_DATA
|| type
== ARC_BUFC_METADATA
);
2796 hdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
2797 ASSERT(HDR_EMPTY(hdr
));
2798 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2799 HDR_SET_PSIZE(hdr
, psize
);
2800 HDR_SET_LSIZE(hdr
, lsize
);
2804 arc_hdr_set_flags(hdr
, arc_bufc_to_flags(type
) | ARC_FLAG_HAS_L1HDR
);
2805 arc_hdr_set_compress(hdr
, compression_type
);
2807 hdr
->b_l1hdr
.b_state
= arc_anon
;
2808 hdr
->b_l1hdr
.b_arc_access
= 0;
2809 hdr
->b_l1hdr
.b_bufcnt
= 0;
2810 hdr
->b_l1hdr
.b_buf
= NULL
;
2813 * Allocate the hdr's buffer. This will contain either
2814 * the compressed or uncompressed data depending on the block
2815 * it references and compressed arc enablement.
2817 arc_hdr_alloc_pabd(hdr
);
2818 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2824 * Transition between the two allocation states for the arc_buf_hdr struct.
2825 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
2826 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
2827 * version is used when a cache buffer is only in the L2ARC in order to reduce
2830 static arc_buf_hdr_t
*
2831 arc_hdr_realloc(arc_buf_hdr_t
*hdr
, kmem_cache_t
*old
, kmem_cache_t
*new)
2833 arc_buf_hdr_t
*nhdr
;
2834 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
2836 ASSERT(HDR_HAS_L2HDR(hdr
));
2837 ASSERT((old
== hdr_full_cache
&& new == hdr_l2only_cache
) ||
2838 (old
== hdr_l2only_cache
&& new == hdr_full_cache
));
2840 nhdr
= kmem_cache_alloc(new, KM_PUSHPAGE
);
2842 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
2843 buf_hash_remove(hdr
);
2845 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
2847 if (new == hdr_full_cache
) {
2848 arc_hdr_set_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
2850 * arc_access and arc_change_state need to be aware that a
2851 * header has just come out of L2ARC, so we set its state to
2852 * l2c_only even though it's about to change.
2854 nhdr
->b_l1hdr
.b_state
= arc_l2c_only
;
2856 /* Verify previous threads set to NULL before freeing */
2857 ASSERT3P(nhdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2859 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2860 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
2861 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2864 * If we've reached here, We must have been called from
2865 * arc_evict_hdr(), as such we should have already been
2866 * removed from any ghost list we were previously on
2867 * (which protects us from racing with arc_evict_state),
2868 * thus no locking is needed during this check.
2870 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
2873 * A buffer must not be moved into the arc_l2c_only
2874 * state if it's not finished being written out to the
2875 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
2876 * might try to be accessed, even though it was removed.
2878 VERIFY(!HDR_L2_WRITING(hdr
));
2879 VERIFY3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2881 arc_hdr_clear_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
2884 * The header has been reallocated so we need to re-insert it into any
2887 (void) buf_hash_insert(nhdr
, NULL
);
2889 ASSERT(list_link_active(&hdr
->b_l2hdr
.b_l2node
));
2891 mutex_enter(&dev
->l2ad_mtx
);
2894 * We must place the realloc'ed header back into the list at
2895 * the same spot. Otherwise, if it's placed earlier in the list,
2896 * l2arc_write_buffers() could find it during the function's
2897 * write phase, and try to write it out to the l2arc.
2899 list_insert_after(&dev
->l2ad_buflist
, hdr
, nhdr
);
2900 list_remove(&dev
->l2ad_buflist
, hdr
);
2902 mutex_exit(&dev
->l2ad_mtx
);
2905 * Since we're using the pointer address as the tag when
2906 * incrementing and decrementing the l2ad_alloc refcount, we
2907 * must remove the old pointer (that we're about to destroy) and
2908 * add the new pointer to the refcount. Otherwise we'd remove
2909 * the wrong pointer address when calling arc_hdr_destroy() later.
2912 (void) refcount_remove_many(&dev
->l2ad_alloc
, arc_hdr_size(hdr
), hdr
);
2913 (void) refcount_add_many(&dev
->l2ad_alloc
, arc_hdr_size(nhdr
), nhdr
);
2915 buf_discard_identity(hdr
);
2916 kmem_cache_free(old
, hdr
);
2922 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
2923 * The buf is returned thawed since we expect the consumer to modify it.
2926 arc_alloc_buf(spa_t
*spa
, void *tag
, arc_buf_contents_t type
, int32_t size
)
2928 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), size
, size
,
2929 ZIO_COMPRESS_OFF
, type
);
2930 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
2932 arc_buf_t
*buf
= NULL
;
2933 VERIFY0(arc_buf_alloc_impl(hdr
, tag
, B_FALSE
, B_FALSE
, &buf
));
2940 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
2941 * for bufs containing metadata.
2944 arc_alloc_compressed_buf(spa_t
*spa
, void *tag
, uint64_t psize
, uint64_t lsize
,
2945 enum zio_compress compression_type
)
2947 ASSERT3U(lsize
, >, 0);
2948 ASSERT3U(lsize
, >=, psize
);
2949 ASSERT(compression_type
> ZIO_COMPRESS_OFF
);
2950 ASSERT(compression_type
< ZIO_COMPRESS_FUNCTIONS
);
2952 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
2953 compression_type
, ARC_BUFC_DATA
);
2954 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
2956 arc_buf_t
*buf
= NULL
;
2957 VERIFY0(arc_buf_alloc_impl(hdr
, tag
, B_TRUE
, B_FALSE
, &buf
));
2959 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2961 if (!arc_buf_is_shared(buf
)) {
2963 * To ensure that the hdr has the correct data in it if we call
2964 * arc_decompress() on this buf before it's been written to
2965 * disk, it's easiest if we just set up sharing between the
2968 ASSERT(!abd_is_linear(hdr
->b_l1hdr
.b_pabd
));
2969 arc_hdr_free_pabd(hdr
);
2970 arc_share_buf(hdr
, buf
);
2977 arc_hdr_l2hdr_destroy(arc_buf_hdr_t
*hdr
)
2979 l2arc_buf_hdr_t
*l2hdr
= &hdr
->b_l2hdr
;
2980 l2arc_dev_t
*dev
= l2hdr
->b_dev
;
2981 uint64_t asize
= arc_hdr_size(hdr
);
2983 ASSERT(MUTEX_HELD(&dev
->l2ad_mtx
));
2984 ASSERT(HDR_HAS_L2HDR(hdr
));
2986 list_remove(&dev
->l2ad_buflist
, hdr
);
2988 ARCSTAT_INCR(arcstat_l2_asize
, -asize
);
2989 ARCSTAT_INCR(arcstat_l2_size
, -HDR_GET_LSIZE(hdr
));
2991 vdev_space_update(dev
->l2ad_vdev
, -asize
, 0, 0);
2993 (void) refcount_remove_many(&dev
->l2ad_alloc
, asize
, hdr
);
2994 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
2998 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
3000 if (HDR_HAS_L1HDR(hdr
)) {
3001 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
||
3002 hdr
->b_l1hdr
.b_bufcnt
> 0);
3003 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3004 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
3006 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3007 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
3009 if (!HDR_EMPTY(hdr
))
3010 buf_discard_identity(hdr
);
3012 if (HDR_HAS_L2HDR(hdr
)) {
3013 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
3014 boolean_t buflist_held
= MUTEX_HELD(&dev
->l2ad_mtx
);
3017 mutex_enter(&dev
->l2ad_mtx
);
3020 * Even though we checked this conditional above, we
3021 * need to check this again now that we have the
3022 * l2ad_mtx. This is because we could be racing with
3023 * another thread calling l2arc_evict() which might have
3024 * destroyed this header's L2 portion as we were waiting
3025 * to acquire the l2ad_mtx. If that happens, we don't
3026 * want to re-destroy the header's L2 portion.
3028 if (HDR_HAS_L2HDR(hdr
))
3029 arc_hdr_l2hdr_destroy(hdr
);
3032 mutex_exit(&dev
->l2ad_mtx
);
3035 if (HDR_HAS_L1HDR(hdr
)) {
3036 arc_cksum_free(hdr
);
3038 while (hdr
->b_l1hdr
.b_buf
!= NULL
)
3039 arc_buf_destroy_impl(hdr
->b_l1hdr
.b_buf
);
3041 if (hdr
->b_l1hdr
.b_pabd
!= NULL
)
3042 arc_hdr_free_pabd(hdr
);
3045 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
3046 if (HDR_HAS_L1HDR(hdr
)) {
3047 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3048 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
3049 kmem_cache_free(hdr_full_cache
, hdr
);
3051 kmem_cache_free(hdr_l2only_cache
, hdr
);
3056 arc_buf_destroy(arc_buf_t
*buf
, void* tag
)
3058 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3059 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
3061 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
3062 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
3063 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3064 VERIFY0(remove_reference(hdr
, NULL
, tag
));
3065 arc_hdr_destroy(hdr
);
3069 mutex_enter(hash_lock
);
3070 ASSERT3P(hdr
, ==, buf
->b_hdr
);
3071 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
3072 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3073 ASSERT3P(hdr
->b_l1hdr
.b_state
, !=, arc_anon
);
3074 ASSERT3P(buf
->b_data
, !=, NULL
);
3076 (void) remove_reference(hdr
, hash_lock
, tag
);
3077 arc_buf_destroy_impl(buf
);
3078 mutex_exit(hash_lock
);
3082 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3083 * state of the header is dependent on its state prior to entering this
3084 * function. The following transitions are possible:
3086 * - arc_mru -> arc_mru_ghost
3087 * - arc_mfu -> arc_mfu_ghost
3088 * - arc_mru_ghost -> arc_l2c_only
3089 * - arc_mru_ghost -> deleted
3090 * - arc_mfu_ghost -> arc_l2c_only
3091 * - arc_mfu_ghost -> deleted
3094 arc_evict_hdr(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
3096 arc_state_t
*evicted_state
, *state
;
3097 int64_t bytes_evicted
= 0;
3099 ASSERT(MUTEX_HELD(hash_lock
));
3100 ASSERT(HDR_HAS_L1HDR(hdr
));
3102 state
= hdr
->b_l1hdr
.b_state
;
3103 if (GHOST_STATE(state
)) {
3104 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3105 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
3108 * l2arc_write_buffers() relies on a header's L1 portion
3109 * (i.e. its b_pabd field) during it's write phase.
3110 * Thus, we cannot push a header onto the arc_l2c_only
3111 * state (removing its L1 piece) until the header is
3112 * done being written to the l2arc.
3114 if (HDR_HAS_L2HDR(hdr
) && HDR_L2_WRITING(hdr
)) {
3115 ARCSTAT_BUMP(arcstat_evict_l2_skip
);
3116 return (bytes_evicted
);
3119 ARCSTAT_BUMP(arcstat_deleted
);
3120 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3122 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, hdr
);
3124 if (HDR_HAS_L2HDR(hdr
)) {
3125 ASSERT(hdr
->b_l1hdr
.b_pabd
== NULL
);
3127 * This buffer is cached on the 2nd Level ARC;
3128 * don't destroy the header.
3130 arc_change_state(arc_l2c_only
, hdr
, hash_lock
);
3132 * dropping from L1+L2 cached to L2-only,
3133 * realloc to remove the L1 header.
3135 hdr
= arc_hdr_realloc(hdr
, hdr_full_cache
,
3138 arc_change_state(arc_anon
, hdr
, hash_lock
);
3139 arc_hdr_destroy(hdr
);
3141 return (bytes_evicted
);
3144 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
3145 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
3147 /* prefetch buffers have a minimum lifespan */
3148 if (HDR_IO_IN_PROGRESS(hdr
) ||
3149 ((hdr
->b_flags
& (ARC_FLAG_PREFETCH
| ARC_FLAG_INDIRECT
)) &&
3150 ddi_get_lbolt() - hdr
->b_l1hdr
.b_arc_access
<
3151 arc_min_prefetch_lifespan
)) {
3152 ARCSTAT_BUMP(arcstat_evict_skip
);
3153 return (bytes_evicted
);
3156 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
3157 while (hdr
->b_l1hdr
.b_buf
) {
3158 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
3159 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
3160 ARCSTAT_BUMP(arcstat_mutex_miss
);
3163 if (buf
->b_data
!= NULL
)
3164 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3165 mutex_exit(&buf
->b_evict_lock
);
3166 arc_buf_destroy_impl(buf
);
3169 if (HDR_HAS_L2HDR(hdr
)) {
3170 ARCSTAT_INCR(arcstat_evict_l2_cached
, HDR_GET_LSIZE(hdr
));
3172 if (l2arc_write_eligible(hdr
->b_spa
, hdr
)) {
3173 ARCSTAT_INCR(arcstat_evict_l2_eligible
,
3174 HDR_GET_LSIZE(hdr
));
3176 ARCSTAT_INCR(arcstat_evict_l2_ineligible
,
3177 HDR_GET_LSIZE(hdr
));
3181 if (hdr
->b_l1hdr
.b_bufcnt
== 0) {
3182 arc_cksum_free(hdr
);
3184 bytes_evicted
+= arc_hdr_size(hdr
);
3187 * If this hdr is being evicted and has a compressed
3188 * buffer then we discard it here before we change states.
3189 * This ensures that the accounting is updated correctly
3190 * in arc_free_data_impl().
3192 arc_hdr_free_pabd(hdr
);
3194 arc_change_state(evicted_state
, hdr
, hash_lock
);
3195 ASSERT(HDR_IN_HASH_TABLE(hdr
));
3196 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
3197 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, hdr
);
3200 return (bytes_evicted
);
3204 arc_evict_state_impl(multilist_t
*ml
, int idx
, arc_buf_hdr_t
*marker
,
3205 uint64_t spa
, int64_t bytes
)
3207 multilist_sublist_t
*mls
;
3208 uint64_t bytes_evicted
= 0;
3210 kmutex_t
*hash_lock
;
3211 int evict_count
= 0;
3213 ASSERT3P(marker
, !=, NULL
);
3214 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3216 mls
= multilist_sublist_lock(ml
, idx
);
3218 for (hdr
= multilist_sublist_prev(mls
, marker
); hdr
!= NULL
;
3219 hdr
= multilist_sublist_prev(mls
, marker
)) {
3220 if ((bytes
!= ARC_EVICT_ALL
&& bytes_evicted
>= bytes
) ||
3221 (evict_count
>= zfs_arc_evict_batch_limit
))
3225 * To keep our iteration location, move the marker
3226 * forward. Since we're not holding hdr's hash lock, we
3227 * must be very careful and not remove 'hdr' from the
3228 * sublist. Otherwise, other consumers might mistake the
3229 * 'hdr' as not being on a sublist when they call the
3230 * multilist_link_active() function (they all rely on
3231 * the hash lock protecting concurrent insertions and
3232 * removals). multilist_sublist_move_forward() was
3233 * specifically implemented to ensure this is the case
3234 * (only 'marker' will be removed and re-inserted).
3236 multilist_sublist_move_forward(mls
, marker
);
3239 * The only case where the b_spa field should ever be
3240 * zero, is the marker headers inserted by
3241 * arc_evict_state(). It's possible for multiple threads
3242 * to be calling arc_evict_state() concurrently (e.g.
3243 * dsl_pool_close() and zio_inject_fault()), so we must
3244 * skip any markers we see from these other threads.
3246 if (hdr
->b_spa
== 0)
3249 /* we're only interested in evicting buffers of a certain spa */
3250 if (spa
!= 0 && hdr
->b_spa
!= spa
) {
3251 ARCSTAT_BUMP(arcstat_evict_skip
);
3255 hash_lock
= HDR_LOCK(hdr
);
3258 * We aren't calling this function from any code path
3259 * that would already be holding a hash lock, so we're
3260 * asserting on this assumption to be defensive in case
3261 * this ever changes. Without this check, it would be
3262 * possible to incorrectly increment arcstat_mutex_miss
3263 * below (e.g. if the code changed such that we called
3264 * this function with a hash lock held).
3266 ASSERT(!MUTEX_HELD(hash_lock
));
3268 if (mutex_tryenter(hash_lock
)) {
3269 uint64_t evicted
= arc_evict_hdr(hdr
, hash_lock
);
3270 mutex_exit(hash_lock
);
3272 bytes_evicted
+= evicted
;
3275 * If evicted is zero, arc_evict_hdr() must have
3276 * decided to skip this header, don't increment
3277 * evict_count in this case.
3283 * If arc_size isn't overflowing, signal any
3284 * threads that might happen to be waiting.
3286 * For each header evicted, we wake up a single
3287 * thread. If we used cv_broadcast, we could
3288 * wake up "too many" threads causing arc_size
3289 * to significantly overflow arc_c; since
3290 * arc_get_data_impl() doesn't check for overflow
3291 * when it's woken up (it doesn't because it's
3292 * possible for the ARC to be overflowing while
3293 * full of un-evictable buffers, and the
3294 * function should proceed in this case).
3296 * If threads are left sleeping, due to not
3297 * using cv_broadcast, they will be woken up
3298 * just before arc_reclaim_thread() sleeps.
3300 mutex_enter(&arc_reclaim_lock
);
3301 if (!arc_is_overflowing())
3302 cv_signal(&arc_reclaim_waiters_cv
);
3303 mutex_exit(&arc_reclaim_lock
);
3305 ARCSTAT_BUMP(arcstat_mutex_miss
);
3309 multilist_sublist_unlock(mls
);
3311 return (bytes_evicted
);
3315 * Evict buffers from the given arc state, until we've removed the
3316 * specified number of bytes. Move the removed buffers to the
3317 * appropriate evict state.
3319 * This function makes a "best effort". It skips over any buffers
3320 * it can't get a hash_lock on, and so, may not catch all candidates.
3321 * It may also return without evicting as much space as requested.
3323 * If bytes is specified using the special value ARC_EVICT_ALL, this
3324 * will evict all available (i.e. unlocked and evictable) buffers from
3325 * the given arc state; which is used by arc_flush().
3328 arc_evict_state(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
3329 arc_buf_contents_t type
)
3331 uint64_t total_evicted
= 0;
3332 multilist_t
*ml
= state
->arcs_list
[type
];
3334 arc_buf_hdr_t
**markers
;
3337 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3339 num_sublists
= multilist_get_num_sublists(ml
);
3342 * If we've tried to evict from each sublist, made some
3343 * progress, but still have not hit the target number of bytes
3344 * to evict, we want to keep trying. The markers allow us to
3345 * pick up where we left off for each individual sublist, rather
3346 * than starting from the tail each time.
3348 markers
= kmem_zalloc(sizeof (*markers
) * num_sublists
, KM_SLEEP
);
3349 for (i
= 0; i
< num_sublists
; i
++) {
3350 multilist_sublist_t
*mls
;
3352 markers
[i
] = kmem_cache_alloc(hdr_full_cache
, KM_SLEEP
);
3355 * A b_spa of 0 is used to indicate that this header is
3356 * a marker. This fact is used in arc_adjust_type() and
3357 * arc_evict_state_impl().
3359 markers
[i
]->b_spa
= 0;
3361 mls
= multilist_sublist_lock(ml
, i
);
3362 multilist_sublist_insert_tail(mls
, markers
[i
]);
3363 multilist_sublist_unlock(mls
);
3367 * While we haven't hit our target number of bytes to evict, or
3368 * we're evicting all available buffers.
3370 while (total_evicted
< bytes
|| bytes
== ARC_EVICT_ALL
) {
3371 int sublist_idx
= multilist_get_random_index(ml
);
3372 uint64_t scan_evicted
= 0;
3375 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
3376 * Request that 10% of the LRUs be scanned by the superblock
3379 if (type
== ARC_BUFC_DATA
&& arc_dnode_size
> arc_dnode_limit
)
3380 arc_prune_async((arc_dnode_size
- arc_dnode_limit
) /
3381 sizeof (dnode_t
) / zfs_arc_dnode_reduce_percent
);
3384 * Start eviction using a randomly selected sublist,
3385 * this is to try and evenly balance eviction across all
3386 * sublists. Always starting at the same sublist
3387 * (e.g. index 0) would cause evictions to favor certain
3388 * sublists over others.
3390 for (i
= 0; i
< num_sublists
; i
++) {
3391 uint64_t bytes_remaining
;
3392 uint64_t bytes_evicted
;
3394 if (bytes
== ARC_EVICT_ALL
)
3395 bytes_remaining
= ARC_EVICT_ALL
;
3396 else if (total_evicted
< bytes
)
3397 bytes_remaining
= bytes
- total_evicted
;
3401 bytes_evicted
= arc_evict_state_impl(ml
, sublist_idx
,
3402 markers
[sublist_idx
], spa
, bytes_remaining
);
3404 scan_evicted
+= bytes_evicted
;
3405 total_evicted
+= bytes_evicted
;
3407 /* we've reached the end, wrap to the beginning */
3408 if (++sublist_idx
>= num_sublists
)
3413 * If we didn't evict anything during this scan, we have
3414 * no reason to believe we'll evict more during another
3415 * scan, so break the loop.
3417 if (scan_evicted
== 0) {
3418 /* This isn't possible, let's make that obvious */
3419 ASSERT3S(bytes
, !=, 0);
3422 * When bytes is ARC_EVICT_ALL, the only way to
3423 * break the loop is when scan_evicted is zero.
3424 * In that case, we actually have evicted enough,
3425 * so we don't want to increment the kstat.
3427 if (bytes
!= ARC_EVICT_ALL
) {
3428 ASSERT3S(total_evicted
, <, bytes
);
3429 ARCSTAT_BUMP(arcstat_evict_not_enough
);
3436 for (i
= 0; i
< num_sublists
; i
++) {
3437 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
3438 multilist_sublist_remove(mls
, markers
[i
]);
3439 multilist_sublist_unlock(mls
);
3441 kmem_cache_free(hdr_full_cache
, markers
[i
]);
3443 kmem_free(markers
, sizeof (*markers
) * num_sublists
);
3445 return (total_evicted
);
3449 * Flush all "evictable" data of the given type from the arc state
3450 * specified. This will not evict any "active" buffers (i.e. referenced).
3452 * When 'retry' is set to B_FALSE, the function will make a single pass
3453 * over the state and evict any buffers that it can. Since it doesn't
3454 * continually retry the eviction, it might end up leaving some buffers
3455 * in the ARC due to lock misses.
3457 * When 'retry' is set to B_TRUE, the function will continually retry the
3458 * eviction until *all* evictable buffers have been removed from the
3459 * state. As a result, if concurrent insertions into the state are
3460 * allowed (e.g. if the ARC isn't shutting down), this function might
3461 * wind up in an infinite loop, continually trying to evict buffers.
3464 arc_flush_state(arc_state_t
*state
, uint64_t spa
, arc_buf_contents_t type
,
3467 uint64_t evicted
= 0;
3469 while (refcount_count(&state
->arcs_esize
[type
]) != 0) {
3470 evicted
+= arc_evict_state(state
, spa
, ARC_EVICT_ALL
, type
);
3480 * Helper function for arc_prune_async() it is responsible for safely
3481 * handling the execution of a registered arc_prune_func_t.
3484 arc_prune_task(void *ptr
)
3486 arc_prune_t
*ap
= (arc_prune_t
*)ptr
;
3487 arc_prune_func_t
*func
= ap
->p_pfunc
;
3490 func(ap
->p_adjust
, ap
->p_private
);
3492 refcount_remove(&ap
->p_refcnt
, func
);
3496 * Notify registered consumers they must drop holds on a portion of the ARC
3497 * buffered they reference. This provides a mechanism to ensure the ARC can
3498 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
3499 * is analogous to dnlc_reduce_cache() but more generic.
3501 * This operation is performed asynchronously so it may be safely called
3502 * in the context of the arc_reclaim_thread(). A reference is taken here
3503 * for each registered arc_prune_t and the arc_prune_task() is responsible
3504 * for releasing it once the registered arc_prune_func_t has completed.
3507 arc_prune_async(int64_t adjust
)
3511 mutex_enter(&arc_prune_mtx
);
3512 for (ap
= list_head(&arc_prune_list
); ap
!= NULL
;
3513 ap
= list_next(&arc_prune_list
, ap
)) {
3515 if (refcount_count(&ap
->p_refcnt
) >= 2)
3518 refcount_add(&ap
->p_refcnt
, ap
->p_pfunc
);
3519 ap
->p_adjust
= adjust
;
3520 if (taskq_dispatch(arc_prune_taskq
, arc_prune_task
,
3521 ap
, TQ_SLEEP
) == TASKQID_INVALID
) {
3522 refcount_remove(&ap
->p_refcnt
, ap
->p_pfunc
);
3525 ARCSTAT_BUMP(arcstat_prune
);
3527 mutex_exit(&arc_prune_mtx
);
3531 * Evict the specified number of bytes from the state specified,
3532 * restricting eviction to the spa and type given. This function
3533 * prevents us from trying to evict more from a state's list than
3534 * is "evictable", and to skip evicting altogether when passed a
3535 * negative value for "bytes". In contrast, arc_evict_state() will
3536 * evict everything it can, when passed a negative value for "bytes".
3539 arc_adjust_impl(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
3540 arc_buf_contents_t type
)
3544 if (bytes
> 0 && refcount_count(&state
->arcs_esize
[type
]) > 0) {
3545 delta
= MIN(refcount_count(&state
->arcs_esize
[type
]), bytes
);
3546 return (arc_evict_state(state
, spa
, delta
, type
));
3553 * The goal of this function is to evict enough meta data buffers from the
3554 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
3555 * more complicated than it appears because it is common for data buffers
3556 * to have holds on meta data buffers. In addition, dnode meta data buffers
3557 * will be held by the dnodes in the block preventing them from being freed.
3558 * This means we can't simply traverse the ARC and expect to always find
3559 * enough unheld meta data buffer to release.
3561 * Therefore, this function has been updated to make alternating passes
3562 * over the ARC releasing data buffers and then newly unheld meta data
3563 * buffers. This ensures forward progress is maintained and arc_meta_used
3564 * will decrease. Normally this is sufficient, but if required the ARC
3565 * will call the registered prune callbacks causing dentry and inodes to
3566 * be dropped from the VFS cache. This will make dnode meta data buffers
3567 * available for reclaim.
3570 arc_adjust_meta_balanced(void)
3572 int64_t delta
, prune
= 0, adjustmnt
;
3573 uint64_t total_evicted
= 0;
3574 arc_buf_contents_t type
= ARC_BUFC_DATA
;
3575 int restarts
= MAX(zfs_arc_meta_adjust_restarts
, 0);
3579 * This slightly differs than the way we evict from the mru in
3580 * arc_adjust because we don't have a "target" value (i.e. no
3581 * "meta" arc_p). As a result, I think we can completely
3582 * cannibalize the metadata in the MRU before we evict the
3583 * metadata from the MFU. I think we probably need to implement a
3584 * "metadata arc_p" value to do this properly.
3586 adjustmnt
= arc_meta_used
- arc_meta_limit
;
3588 if (adjustmnt
> 0 && refcount_count(&arc_mru
->arcs_esize
[type
]) > 0) {
3589 delta
= MIN(refcount_count(&arc_mru
->arcs_esize
[type
]),
3591 total_evicted
+= arc_adjust_impl(arc_mru
, 0, delta
, type
);
3596 * We can't afford to recalculate adjustmnt here. If we do,
3597 * new metadata buffers can sneak into the MRU or ANON lists,
3598 * thus penalize the MFU metadata. Although the fudge factor is
3599 * small, it has been empirically shown to be significant for
3600 * certain workloads (e.g. creating many empty directories). As
3601 * such, we use the original calculation for adjustmnt, and
3602 * simply decrement the amount of data evicted from the MRU.
3605 if (adjustmnt
> 0 && refcount_count(&arc_mfu
->arcs_esize
[type
]) > 0) {
3606 delta
= MIN(refcount_count(&arc_mfu
->arcs_esize
[type
]),
3608 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, delta
, type
);
3611 adjustmnt
= arc_meta_used
- arc_meta_limit
;
3613 if (adjustmnt
> 0 &&
3614 refcount_count(&arc_mru_ghost
->arcs_esize
[type
]) > 0) {
3615 delta
= MIN(adjustmnt
,
3616 refcount_count(&arc_mru_ghost
->arcs_esize
[type
]));
3617 total_evicted
+= arc_adjust_impl(arc_mru_ghost
, 0, delta
, type
);
3621 if (adjustmnt
> 0 &&
3622 refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]) > 0) {
3623 delta
= MIN(adjustmnt
,
3624 refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]));
3625 total_evicted
+= arc_adjust_impl(arc_mfu_ghost
, 0, delta
, type
);
3629 * If after attempting to make the requested adjustment to the ARC
3630 * the meta limit is still being exceeded then request that the
3631 * higher layers drop some cached objects which have holds on ARC
3632 * meta buffers. Requests to the upper layers will be made with
3633 * increasingly large scan sizes until the ARC is below the limit.
3635 if (arc_meta_used
> arc_meta_limit
) {
3636 if (type
== ARC_BUFC_DATA
) {
3637 type
= ARC_BUFC_METADATA
;
3639 type
= ARC_BUFC_DATA
;
3641 if (zfs_arc_meta_prune
) {
3642 prune
+= zfs_arc_meta_prune
;
3643 arc_prune_async(prune
);
3652 return (total_evicted
);
3656 * Evict metadata buffers from the cache, such that arc_meta_used is
3657 * capped by the arc_meta_limit tunable.
3660 arc_adjust_meta_only(void)
3662 uint64_t total_evicted
= 0;
3666 * If we're over the meta limit, we want to evict enough
3667 * metadata to get back under the meta limit. We don't want to
3668 * evict so much that we drop the MRU below arc_p, though. If
3669 * we're over the meta limit more than we're over arc_p, we
3670 * evict some from the MRU here, and some from the MFU below.
3672 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
3673 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
3674 refcount_count(&arc_mru
->arcs_size
) - arc_p
));
3676 total_evicted
+= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3679 * Similar to the above, we want to evict enough bytes to get us
3680 * below the meta limit, but not so much as to drop us below the
3681 * space allotted to the MFU (which is defined as arc_c - arc_p).
3683 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
3684 (int64_t)(refcount_count(&arc_mfu
->arcs_size
) - (arc_c
- arc_p
)));
3686 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3688 return (total_evicted
);
3692 arc_adjust_meta(void)
3694 if (zfs_arc_meta_strategy
== ARC_STRATEGY_META_ONLY
)
3695 return (arc_adjust_meta_only());
3697 return (arc_adjust_meta_balanced());
3701 * Return the type of the oldest buffer in the given arc state
3703 * This function will select a random sublist of type ARC_BUFC_DATA and
3704 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3705 * is compared, and the type which contains the "older" buffer will be
3708 static arc_buf_contents_t
3709 arc_adjust_type(arc_state_t
*state
)
3711 multilist_t
*data_ml
= state
->arcs_list
[ARC_BUFC_DATA
];
3712 multilist_t
*meta_ml
= state
->arcs_list
[ARC_BUFC_METADATA
];
3713 int data_idx
= multilist_get_random_index(data_ml
);
3714 int meta_idx
= multilist_get_random_index(meta_ml
);
3715 multilist_sublist_t
*data_mls
;
3716 multilist_sublist_t
*meta_mls
;
3717 arc_buf_contents_t type
;
3718 arc_buf_hdr_t
*data_hdr
;
3719 arc_buf_hdr_t
*meta_hdr
;
3722 * We keep the sublist lock until we're finished, to prevent
3723 * the headers from being destroyed via arc_evict_state().
3725 data_mls
= multilist_sublist_lock(data_ml
, data_idx
);
3726 meta_mls
= multilist_sublist_lock(meta_ml
, meta_idx
);
3729 * These two loops are to ensure we skip any markers that
3730 * might be at the tail of the lists due to arc_evict_state().
3733 for (data_hdr
= multilist_sublist_tail(data_mls
); data_hdr
!= NULL
;
3734 data_hdr
= multilist_sublist_prev(data_mls
, data_hdr
)) {
3735 if (data_hdr
->b_spa
!= 0)
3739 for (meta_hdr
= multilist_sublist_tail(meta_mls
); meta_hdr
!= NULL
;
3740 meta_hdr
= multilist_sublist_prev(meta_mls
, meta_hdr
)) {
3741 if (meta_hdr
->b_spa
!= 0)
3745 if (data_hdr
== NULL
&& meta_hdr
== NULL
) {
3746 type
= ARC_BUFC_DATA
;
3747 } else if (data_hdr
== NULL
) {
3748 ASSERT3P(meta_hdr
, !=, NULL
);
3749 type
= ARC_BUFC_METADATA
;
3750 } else if (meta_hdr
== NULL
) {
3751 ASSERT3P(data_hdr
, !=, NULL
);
3752 type
= ARC_BUFC_DATA
;
3754 ASSERT3P(data_hdr
, !=, NULL
);
3755 ASSERT3P(meta_hdr
, !=, NULL
);
3757 /* The headers can't be on the sublist without an L1 header */
3758 ASSERT(HDR_HAS_L1HDR(data_hdr
));
3759 ASSERT(HDR_HAS_L1HDR(meta_hdr
));
3761 if (data_hdr
->b_l1hdr
.b_arc_access
<
3762 meta_hdr
->b_l1hdr
.b_arc_access
) {
3763 type
= ARC_BUFC_DATA
;
3765 type
= ARC_BUFC_METADATA
;
3769 multilist_sublist_unlock(meta_mls
);
3770 multilist_sublist_unlock(data_mls
);
3776 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3781 uint64_t total_evicted
= 0;
3786 * If we're over arc_meta_limit, we want to correct that before
3787 * potentially evicting data buffers below.
3789 total_evicted
+= arc_adjust_meta();
3794 * If we're over the target cache size, we want to evict enough
3795 * from the list to get back to our target size. We don't want
3796 * to evict too much from the MRU, such that it drops below
3797 * arc_p. So, if we're over our target cache size more than
3798 * the MRU is over arc_p, we'll evict enough to get back to
3799 * arc_p here, and then evict more from the MFU below.
3801 target
= MIN((int64_t)(arc_size
- arc_c
),
3802 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
3803 refcount_count(&arc_mru
->arcs_size
) + arc_meta_used
- arc_p
));
3806 * If we're below arc_meta_min, always prefer to evict data.
3807 * Otherwise, try to satisfy the requested number of bytes to
3808 * evict from the type which contains older buffers; in an
3809 * effort to keep newer buffers in the cache regardless of their
3810 * type. If we cannot satisfy the number of bytes from this
3811 * type, spill over into the next type.
3813 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
3814 arc_meta_used
> arc_meta_min
) {
3815 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3816 total_evicted
+= bytes
;
3819 * If we couldn't evict our target number of bytes from
3820 * metadata, we try to get the rest from data.
3825 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3827 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3828 total_evicted
+= bytes
;
3831 * If we couldn't evict our target number of bytes from
3832 * data, we try to get the rest from metadata.
3837 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3843 * Now that we've tried to evict enough from the MRU to get its
3844 * size back to arc_p, if we're still above the target cache
3845 * size, we evict the rest from the MFU.
3847 target
= arc_size
- arc_c
;
3849 if (arc_adjust_type(arc_mfu
) == ARC_BUFC_METADATA
&&
3850 arc_meta_used
> arc_meta_min
) {
3851 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3852 total_evicted
+= bytes
;
3855 * If we couldn't evict our target number of bytes from
3856 * metadata, we try to get the rest from data.
3861 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3863 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3864 total_evicted
+= bytes
;
3867 * If we couldn't evict our target number of bytes from
3868 * data, we try to get the rest from data.
3873 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3877 * Adjust ghost lists
3879 * In addition to the above, the ARC also defines target values
3880 * for the ghost lists. The sum of the mru list and mru ghost
3881 * list should never exceed the target size of the cache, and
3882 * the sum of the mru list, mfu list, mru ghost list, and mfu
3883 * ghost list should never exceed twice the target size of the
3884 * cache. The following logic enforces these limits on the ghost
3885 * caches, and evicts from them as needed.
3887 target
= refcount_count(&arc_mru
->arcs_size
) +
3888 refcount_count(&arc_mru_ghost
->arcs_size
) - arc_c
;
3890 bytes
= arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_DATA
);
3891 total_evicted
+= bytes
;
3896 arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_METADATA
);
3899 * We assume the sum of the mru list and mfu list is less than
3900 * or equal to arc_c (we enforced this above), which means we
3901 * can use the simpler of the two equations below:
3903 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3904 * mru ghost + mfu ghost <= arc_c
3906 target
= refcount_count(&arc_mru_ghost
->arcs_size
) +
3907 refcount_count(&arc_mfu_ghost
->arcs_size
) - arc_c
;
3909 bytes
= arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_DATA
);
3910 total_evicted
+= bytes
;
3915 arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_METADATA
);
3917 return (total_evicted
);
3921 arc_flush(spa_t
*spa
, boolean_t retry
)
3926 * If retry is B_TRUE, a spa must not be specified since we have
3927 * no good way to determine if all of a spa's buffers have been
3928 * evicted from an arc state.
3930 ASSERT(!retry
|| spa
== 0);
3933 guid
= spa_load_guid(spa
);
3935 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_DATA
, retry
);
3936 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_METADATA
, retry
);
3938 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_DATA
, retry
);
3939 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_METADATA
, retry
);
3941 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3942 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3944 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3945 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3949 arc_shrink(int64_t to_free
)
3953 if (c
> to_free
&& c
- to_free
> arc_c_min
) {
3954 arc_c
= c
- to_free
;
3955 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
3956 if (arc_c
> arc_size
)
3957 arc_c
= MAX(arc_size
, arc_c_min
);
3959 arc_p
= (arc_c
>> 1);
3960 ASSERT(arc_c
>= arc_c_min
);
3961 ASSERT((int64_t)arc_p
>= 0);
3966 if (arc_size
> arc_c
)
3967 (void) arc_adjust();
3971 * Return maximum amount of memory that we could possibly use. Reduced
3972 * to half of all memory in user space which is primarily used for testing.
3975 arc_all_memory(void)
3978 return (MIN(ptob(physmem
),
3979 vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
)));
3981 return (ptob(physmem
) / 2);
3985 typedef enum free_memory_reason_t
{
3990 FMR_PAGES_PP_MAXIMUM
,
3993 } free_memory_reason_t
;
3995 int64_t last_free_memory
;
3996 free_memory_reason_t last_free_reason
;
4000 * Additional reserve of pages for pp_reserve.
4002 int64_t arc_pages_pp_reserve
= 64;
4005 * Additional reserve of pages for swapfs.
4007 int64_t arc_swapfs_reserve
= 64;
4008 #endif /* _KERNEL */
4011 * Return the amount of memory that can be consumed before reclaim will be
4012 * needed. Positive if there is sufficient free memory, negative indicates
4013 * the amount of memory that needs to be freed up.
4016 arc_available_memory(void)
4018 int64_t lowest
= INT64_MAX
;
4019 free_memory_reason_t r
= FMR_UNKNOWN
;
4021 uint64_t available_memory
= ptob(freemem
);
4024 pgcnt_t needfree
= btop(arc_need_free
);
4025 pgcnt_t lotsfree
= btop(arc_sys_free
);
4026 pgcnt_t desfree
= 0;
4031 MIN(available_memory
, vmem_size(heap_arena
, VMEM_FREE
));
4035 n
= PAGESIZE
* (-needfree
);
4043 * check that we're out of range of the pageout scanner. It starts to
4044 * schedule paging if freemem is less than lotsfree and needfree.
4045 * lotsfree is the high-water mark for pageout, and needfree is the
4046 * number of needed free pages. We add extra pages here to make sure
4047 * the scanner doesn't start up while we're freeing memory.
4049 n
= PAGESIZE
* (btop(available_memory
) - lotsfree
- needfree
- desfree
);
4057 * check to make sure that swapfs has enough space so that anon
4058 * reservations can still succeed. anon_resvmem() checks that the
4059 * availrmem is greater than swapfs_minfree, and the number of reserved
4060 * swap pages. We also add a bit of extra here just to prevent
4061 * circumstances from getting really dire.
4063 n
= PAGESIZE
* (availrmem
- swapfs_minfree
- swapfs_reserve
-
4064 desfree
- arc_swapfs_reserve
);
4067 r
= FMR_SWAPFS_MINFREE
;
4072 * Check that we have enough availrmem that memory locking (e.g., via
4073 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4074 * stores the number of pages that cannot be locked; when availrmem
4075 * drops below pages_pp_maximum, page locking mechanisms such as
4076 * page_pp_lock() will fail.)
4078 n
= PAGESIZE
* (availrmem
- pages_pp_maximum
-
4079 arc_pages_pp_reserve
);
4082 r
= FMR_PAGES_PP_MAXIMUM
;
4088 * If we're on an i386 platform, it's possible that we'll exhaust the
4089 * kernel heap space before we ever run out of available physical
4090 * memory. Most checks of the size of the heap_area compare against
4091 * tune.t_minarmem, which is the minimum available real memory that we
4092 * can have in the system. However, this is generally fixed at 25 pages
4093 * which is so low that it's useless. In this comparison, we seek to
4094 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4095 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4098 n
= vmem_size(heap_arena
, VMEM_FREE
) -
4099 (vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
) >> 2);
4107 * If zio data pages are being allocated out of a separate heap segment,
4108 * then enforce that the size of available vmem for this arena remains
4109 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4111 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4112 * memory (in the zio_arena) free, which can avoid memory
4113 * fragmentation issues.
4115 if (zio_arena
!= NULL
) {
4116 n
= (int64_t)vmem_size(zio_arena
, VMEM_FREE
) -
4117 (vmem_size(zio_arena
, VMEM_ALLOC
) >>
4118 arc_zio_arena_free_shift
);
4125 /* Every 100 calls, free a small amount */
4126 if (spa_get_random(100) == 0)
4128 #endif /* _KERNEL */
4130 last_free_memory
= lowest
;
4131 last_free_reason
= r
;
4137 * Determine if the system is under memory pressure and is asking
4138 * to reclaim memory. A return value of B_TRUE indicates that the system
4139 * is under memory pressure and that the arc should adjust accordingly.
4142 arc_reclaim_needed(void)
4144 return (arc_available_memory() < 0);
4148 arc_kmem_reap_now(void)
4151 kmem_cache_t
*prev_cache
= NULL
;
4152 kmem_cache_t
*prev_data_cache
= NULL
;
4153 extern kmem_cache_t
*zio_buf_cache
[];
4154 extern kmem_cache_t
*zio_data_buf_cache
[];
4155 extern kmem_cache_t
*range_seg_cache
;
4157 if ((arc_meta_used
>= arc_meta_limit
) && zfs_arc_meta_prune
) {
4159 * We are exceeding our meta-data cache limit.
4160 * Prune some entries to release holds on meta-data.
4162 arc_prune_async(zfs_arc_meta_prune
);
4165 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
4167 /* reach upper limit of cache size on 32-bit */
4168 if (zio_buf_cache
[i
] == NULL
)
4171 if (zio_buf_cache
[i
] != prev_cache
) {
4172 prev_cache
= zio_buf_cache
[i
];
4173 kmem_cache_reap_now(zio_buf_cache
[i
]);
4175 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
4176 prev_data_cache
= zio_data_buf_cache
[i
];
4177 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
4180 kmem_cache_reap_now(buf_cache
);
4181 kmem_cache_reap_now(hdr_full_cache
);
4182 kmem_cache_reap_now(hdr_l2only_cache
);
4183 kmem_cache_reap_now(range_seg_cache
);
4185 if (zio_arena
!= NULL
) {
4187 * Ask the vmem arena to reclaim unused memory from its
4190 vmem_qcache_reap(zio_arena
);
4195 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4196 * enough data and signal them to proceed. When this happens, the threads in
4197 * arc_get_data_impl() are sleeping while holding the hash lock for their
4198 * particular arc header. Thus, we must be careful to never sleep on a
4199 * hash lock in this thread. This is to prevent the following deadlock:
4201 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4202 * waiting for the reclaim thread to signal it.
4204 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4205 * fails, and goes to sleep forever.
4207 * This possible deadlock is avoided by always acquiring a hash lock
4208 * using mutex_tryenter() from arc_reclaim_thread().
4211 arc_reclaim_thread(void)
4213 fstrans_cookie_t cookie
= spl_fstrans_mark();
4214 hrtime_t growtime
= 0;
4217 CALLB_CPR_INIT(&cpr
, &arc_reclaim_lock
, callb_generic_cpr
, FTAG
);
4219 mutex_enter(&arc_reclaim_lock
);
4220 while (!arc_reclaim_thread_exit
) {
4222 uint64_t evicted
= 0;
4223 uint64_t need_free
= arc_need_free
;
4224 arc_tuning_update();
4227 * This is necessary in order for the mdb ::arc dcmd to
4228 * show up to date information. Since the ::arc command
4229 * does not call the kstat's update function, without
4230 * this call, the command may show stale stats for the
4231 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4232 * with this change, the data might be up to 1 second
4233 * out of date; but that should suffice. The arc_state_t
4234 * structures can be queried directly if more accurate
4235 * information is needed.
4238 if (arc_ksp
!= NULL
)
4239 arc_ksp
->ks_update(arc_ksp
, KSTAT_READ
);
4241 mutex_exit(&arc_reclaim_lock
);
4244 * We call arc_adjust() before (possibly) calling
4245 * arc_kmem_reap_now(), so that we can wake up
4246 * arc_get_data_buf() sooner.
4248 evicted
= arc_adjust();
4250 int64_t free_memory
= arc_available_memory();
4251 if (free_memory
< 0) {
4253 arc_no_grow
= B_TRUE
;
4257 * Wait at least zfs_grow_retry (default 5) seconds
4258 * before considering growing.
4260 growtime
= gethrtime() + SEC2NSEC(arc_grow_retry
);
4262 arc_kmem_reap_now();
4265 * If we are still low on memory, shrink the ARC
4266 * so that we have arc_shrink_min free space.
4268 free_memory
= arc_available_memory();
4270 to_free
= (arc_c
>> arc_shrink_shift
) - free_memory
;
4273 to_free
= MAX(to_free
, need_free
);
4275 arc_shrink(to_free
);
4277 } else if (free_memory
< arc_c
>> arc_no_grow_shift
) {
4278 arc_no_grow
= B_TRUE
;
4279 } else if (gethrtime() >= growtime
) {
4280 arc_no_grow
= B_FALSE
;
4283 mutex_enter(&arc_reclaim_lock
);
4286 * If evicted is zero, we couldn't evict anything via
4287 * arc_adjust(). This could be due to hash lock
4288 * collisions, but more likely due to the majority of
4289 * arc buffers being unevictable. Therefore, even if
4290 * arc_size is above arc_c, another pass is unlikely to
4291 * be helpful and could potentially cause us to enter an
4294 if (arc_size
<= arc_c
|| evicted
== 0) {
4296 * We're either no longer overflowing, or we
4297 * can't evict anything more, so we should wake
4298 * up any threads before we go to sleep and remove
4299 * the bytes we were working on from arc_need_free
4300 * since nothing more will be done here.
4302 cv_broadcast(&arc_reclaim_waiters_cv
);
4303 ARCSTAT_INCR(arcstat_need_free
, -need_free
);
4306 * Block until signaled, or after one second (we
4307 * might need to perform arc_kmem_reap_now()
4308 * even if we aren't being signalled)
4310 CALLB_CPR_SAFE_BEGIN(&cpr
);
4311 (void) cv_timedwait_sig_hires(&arc_reclaim_thread_cv
,
4312 &arc_reclaim_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
4313 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_lock
);
4317 arc_reclaim_thread_exit
= B_FALSE
;
4318 cv_broadcast(&arc_reclaim_thread_cv
);
4319 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_lock */
4320 spl_fstrans_unmark(cookie
);
4326 * Determine the amount of memory eligible for eviction contained in the
4327 * ARC. All clean data reported by the ghost lists can always be safely
4328 * evicted. Due to arc_c_min, the same does not hold for all clean data
4329 * contained by the regular mru and mfu lists.
4331 * In the case of the regular mru and mfu lists, we need to report as
4332 * much clean data as possible, such that evicting that same reported
4333 * data will not bring arc_size below arc_c_min. Thus, in certain
4334 * circumstances, the total amount of clean data in the mru and mfu
4335 * lists might not actually be evictable.
4337 * The following two distinct cases are accounted for:
4339 * 1. The sum of the amount of dirty data contained by both the mru and
4340 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4341 * is greater than or equal to arc_c_min.
4342 * (i.e. amount of dirty data >= arc_c_min)
4344 * This is the easy case; all clean data contained by the mru and mfu
4345 * lists is evictable. Evicting all clean data can only drop arc_size
4346 * to the amount of dirty data, which is greater than arc_c_min.
4348 * 2. The sum of the amount of dirty data contained by both the mru and
4349 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4350 * is less than arc_c_min.
4351 * (i.e. arc_c_min > amount of dirty data)
4353 * 2.1. arc_size is greater than or equal arc_c_min.
4354 * (i.e. arc_size >= arc_c_min > amount of dirty data)
4356 * In this case, not all clean data from the regular mru and mfu
4357 * lists is actually evictable; we must leave enough clean data
4358 * to keep arc_size above arc_c_min. Thus, the maximum amount of
4359 * evictable data from the two lists combined, is exactly the
4360 * difference between arc_size and arc_c_min.
4362 * 2.2. arc_size is less than arc_c_min
4363 * (i.e. arc_c_min > arc_size > amount of dirty data)
4365 * In this case, none of the data contained in the mru and mfu
4366 * lists is evictable, even if it's clean. Since arc_size is
4367 * already below arc_c_min, evicting any more would only
4368 * increase this negative difference.
4371 arc_evictable_memory(void)
4373 uint64_t arc_clean
=
4374 refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]) +
4375 refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]) +
4376 refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]) +
4377 refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
4378 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
4380 if (arc_dirty
>= arc_c_min
)
4383 return (MAX((int64_t)arc_size
- (int64_t)arc_c_min
, 0));
4387 * If sc->nr_to_scan is zero, the caller is requesting a query of the
4388 * number of objects which can potentially be freed. If it is nonzero,
4389 * the request is to free that many objects.
4391 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
4392 * in struct shrinker and also require the shrinker to return the number
4395 * Older kernels require the shrinker to return the number of freeable
4396 * objects following the freeing of nr_to_free.
4398 static spl_shrinker_t
4399 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
4403 /* The arc is considered warm once reclaim has occurred */
4404 if (unlikely(arc_warm
== B_FALSE
))
4407 /* Return the potential number of reclaimable pages */
4408 pages
= btop((int64_t)arc_evictable_memory());
4409 if (sc
->nr_to_scan
== 0)
4412 /* Not allowed to perform filesystem reclaim */
4413 if (!(sc
->gfp_mask
& __GFP_FS
))
4414 return (SHRINK_STOP
);
4416 /* Reclaim in progress */
4417 if (mutex_tryenter(&arc_reclaim_lock
) == 0) {
4418 ARCSTAT_INCR(arcstat_need_free
, ptob(sc
->nr_to_scan
));
4422 mutex_exit(&arc_reclaim_lock
);
4425 * Evict the requested number of pages by shrinking arc_c the
4426 * requested amount. If there is nothing left to evict just
4427 * reap whatever we can from the various arc slabs.
4430 arc_shrink(ptob(sc
->nr_to_scan
));
4431 arc_kmem_reap_now();
4432 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
4433 pages
= MAX(pages
- btop(arc_evictable_memory()), 0);
4435 pages
= btop(arc_evictable_memory());
4438 * We've shrunk what we can, wake up threads.
4440 cv_broadcast(&arc_reclaim_waiters_cv
);
4443 arc_kmem_reap_now();
4444 pages
= SHRINK_STOP
;
4448 * When direct reclaim is observed it usually indicates a rapid
4449 * increase in memory pressure. This occurs because the kswapd
4450 * threads were unable to asynchronously keep enough free memory
4451 * available. In this case set arc_no_grow to briefly pause arc
4452 * growth to avoid compounding the memory pressure.
4454 if (current_is_kswapd()) {
4455 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
4457 arc_no_grow
= B_TRUE
;
4458 ARCSTAT_BUMP(arcstat_memory_direct_count
);
4463 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
4465 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
4466 #endif /* _KERNEL */
4469 * Adapt arc info given the number of bytes we are trying to add and
4470 * the state that we are coming from. This function is only called
4471 * when we are adding new content to the cache.
4474 arc_adapt(int bytes
, arc_state_t
*state
)
4477 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
4478 int64_t mrug_size
= refcount_count(&arc_mru_ghost
->arcs_size
);
4479 int64_t mfug_size
= refcount_count(&arc_mfu_ghost
->arcs_size
);
4481 if (state
== arc_l2c_only
)
4486 * Adapt the target size of the MRU list:
4487 * - if we just hit in the MRU ghost list, then increase
4488 * the target size of the MRU list.
4489 * - if we just hit in the MFU ghost list, then increase
4490 * the target size of the MFU list by decreasing the
4491 * target size of the MRU list.
4493 if (state
== arc_mru_ghost
) {
4494 mult
= (mrug_size
>= mfug_size
) ? 1 : (mfug_size
/ mrug_size
);
4495 if (!zfs_arc_p_dampener_disable
)
4496 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
4498 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
4499 } else if (state
== arc_mfu_ghost
) {
4502 mult
= (mfug_size
>= mrug_size
) ? 1 : (mrug_size
/ mfug_size
);
4503 if (!zfs_arc_p_dampener_disable
)
4504 mult
= MIN(mult
, 10);
4506 delta
= MIN(bytes
* mult
, arc_p
);
4507 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
4509 ASSERT((int64_t)arc_p
>= 0);
4511 if (arc_reclaim_needed()) {
4512 cv_signal(&arc_reclaim_thread_cv
);
4519 if (arc_c
>= arc_c_max
)
4523 * If we're within (2 * maxblocksize) bytes of the target
4524 * cache size, increment the target cache size
4526 ASSERT3U(arc_c
, >=, 2ULL << SPA_MAXBLOCKSHIFT
);
4527 if (arc_size
>= arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
4528 atomic_add_64(&arc_c
, (int64_t)bytes
);
4529 if (arc_c
> arc_c_max
)
4531 else if (state
== arc_anon
)
4532 atomic_add_64(&arc_p
, (int64_t)bytes
);
4536 ASSERT((int64_t)arc_p
>= 0);
4540 * Check if arc_size has grown past our upper threshold, determined by
4541 * zfs_arc_overflow_shift.
4544 arc_is_overflowing(void)
4546 /* Always allow at least one block of overflow */
4547 uint64_t overflow
= MAX(SPA_MAXBLOCKSIZE
,
4548 arc_c
>> zfs_arc_overflow_shift
);
4550 return (arc_size
>= arc_c
+ overflow
);
4554 arc_get_data_abd(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4556 arc_buf_contents_t type
= arc_buf_type(hdr
);
4558 arc_get_data_impl(hdr
, size
, tag
);
4559 if (type
== ARC_BUFC_METADATA
) {
4560 return (abd_alloc(size
, B_TRUE
));
4562 ASSERT(type
== ARC_BUFC_DATA
);
4563 return (abd_alloc(size
, B_FALSE
));
4568 arc_get_data_buf(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4570 arc_buf_contents_t type
= arc_buf_type(hdr
);
4572 arc_get_data_impl(hdr
, size
, tag
);
4573 if (type
== ARC_BUFC_METADATA
) {
4574 return (zio_buf_alloc(size
));
4576 ASSERT(type
== ARC_BUFC_DATA
);
4577 return (zio_data_buf_alloc(size
));
4582 * Allocate a block and return it to the caller. If we are hitting the
4583 * hard limit for the cache size, we must sleep, waiting for the eviction
4584 * thread to catch up. If we're past the target size but below the hard
4585 * limit, we'll only signal the reclaim thread and continue on.
4588 arc_get_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4590 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
4591 arc_buf_contents_t type
= arc_buf_type(hdr
);
4593 arc_adapt(size
, state
);
4596 * If arc_size is currently overflowing, and has grown past our
4597 * upper limit, we must be adding data faster than the evict
4598 * thread can evict. Thus, to ensure we don't compound the
4599 * problem by adding more data and forcing arc_size to grow even
4600 * further past it's target size, we halt and wait for the
4601 * eviction thread to catch up.
4603 * It's also possible that the reclaim thread is unable to evict
4604 * enough buffers to get arc_size below the overflow limit (e.g.
4605 * due to buffers being un-evictable, or hash lock collisions).
4606 * In this case, we want to proceed regardless if we're
4607 * overflowing; thus we don't use a while loop here.
4609 if (arc_is_overflowing()) {
4610 mutex_enter(&arc_reclaim_lock
);
4613 * Now that we've acquired the lock, we may no longer be
4614 * over the overflow limit, lets check.
4616 * We're ignoring the case of spurious wake ups. If that
4617 * were to happen, it'd let this thread consume an ARC
4618 * buffer before it should have (i.e. before we're under
4619 * the overflow limit and were signalled by the reclaim
4620 * thread). As long as that is a rare occurrence, it
4621 * shouldn't cause any harm.
4623 if (arc_is_overflowing()) {
4624 cv_signal(&arc_reclaim_thread_cv
);
4625 cv_wait(&arc_reclaim_waiters_cv
, &arc_reclaim_lock
);
4628 mutex_exit(&arc_reclaim_lock
);
4631 VERIFY3U(hdr
->b_type
, ==, type
);
4632 if (type
== ARC_BUFC_METADATA
) {
4633 arc_space_consume(size
, ARC_SPACE_META
);
4635 arc_space_consume(size
, ARC_SPACE_DATA
);
4639 * Update the state size. Note that ghost states have a
4640 * "ghost size" and so don't need to be updated.
4642 if (!GHOST_STATE(state
)) {
4644 (void) refcount_add_many(&state
->arcs_size
, size
, tag
);
4647 * If this is reached via arc_read, the link is
4648 * protected by the hash lock. If reached via
4649 * arc_buf_alloc, the header should not be accessed by
4650 * any other thread. And, if reached via arc_read_done,
4651 * the hash lock will protect it if it's found in the
4652 * hash table; otherwise no other thread should be
4653 * trying to [add|remove]_reference it.
4655 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
4656 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4657 (void) refcount_add_many(&state
->arcs_esize
[type
],
4662 * If we are growing the cache, and we are adding anonymous
4663 * data, and we have outgrown arc_p, update arc_p
4665 if (arc_size
< arc_c
&& hdr
->b_l1hdr
.b_state
== arc_anon
&&
4666 (refcount_count(&arc_anon
->arcs_size
) +
4667 refcount_count(&arc_mru
->arcs_size
) > arc_p
))
4668 arc_p
= MIN(arc_c
, arc_p
+ size
);
4673 arc_free_data_abd(arc_buf_hdr_t
*hdr
, abd_t
*abd
, uint64_t size
, void *tag
)
4675 arc_free_data_impl(hdr
, size
, tag
);
4680 arc_free_data_buf(arc_buf_hdr_t
*hdr
, void *buf
, uint64_t size
, void *tag
)
4682 arc_buf_contents_t type
= arc_buf_type(hdr
);
4684 arc_free_data_impl(hdr
, size
, tag
);
4685 if (type
== ARC_BUFC_METADATA
) {
4686 zio_buf_free(buf
, size
);
4688 ASSERT(type
== ARC_BUFC_DATA
);
4689 zio_data_buf_free(buf
, size
);
4694 * Free the arc data buffer.
4697 arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4699 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
4700 arc_buf_contents_t type
= arc_buf_type(hdr
);
4702 /* protected by hash lock, if in the hash table */
4703 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
4704 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4705 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
4707 (void) refcount_remove_many(&state
->arcs_esize
[type
],
4710 (void) refcount_remove_many(&state
->arcs_size
, size
, tag
);
4712 VERIFY3U(hdr
->b_type
, ==, type
);
4713 if (type
== ARC_BUFC_METADATA
) {
4714 arc_space_return(size
, ARC_SPACE_META
);
4716 ASSERT(type
== ARC_BUFC_DATA
);
4717 arc_space_return(size
, ARC_SPACE_DATA
);
4722 * This routine is called whenever a buffer is accessed.
4723 * NOTE: the hash lock is dropped in this function.
4726 arc_access(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
4730 ASSERT(MUTEX_HELD(hash_lock
));
4731 ASSERT(HDR_HAS_L1HDR(hdr
));
4733 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
4735 * This buffer is not in the cache, and does not
4736 * appear in our "ghost" list. Add the new buffer
4740 ASSERT0(hdr
->b_l1hdr
.b_arc_access
);
4741 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4742 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
4743 arc_change_state(arc_mru
, hdr
, hash_lock
);
4745 } else if (hdr
->b_l1hdr
.b_state
== arc_mru
) {
4746 now
= ddi_get_lbolt();
4749 * If this buffer is here because of a prefetch, then either:
4750 * - clear the flag if this is a "referencing" read
4751 * (any subsequent access will bump this into the MFU state).
4753 * - move the buffer to the head of the list if this is
4754 * another prefetch (to make it less likely to be evicted).
4756 if (HDR_PREFETCH(hdr
)) {
4757 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
4758 /* link protected by hash lock */
4759 ASSERT(multilist_link_active(
4760 &hdr
->b_l1hdr
.b_arc_node
));
4762 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
4763 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
4764 ARCSTAT_BUMP(arcstat_mru_hits
);
4766 hdr
->b_l1hdr
.b_arc_access
= now
;
4771 * This buffer has been "accessed" only once so far,
4772 * but it is still in the cache. Move it to the MFU
4775 if (ddi_time_after(now
, hdr
->b_l1hdr
.b_arc_access
+
4778 * More than 125ms have passed since we
4779 * instantiated this buffer. Move it to the
4780 * most frequently used state.
4782 hdr
->b_l1hdr
.b_arc_access
= now
;
4783 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4784 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4786 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
4787 ARCSTAT_BUMP(arcstat_mru_hits
);
4788 } else if (hdr
->b_l1hdr
.b_state
== arc_mru_ghost
) {
4789 arc_state_t
*new_state
;
4791 * This buffer has been "accessed" recently, but
4792 * was evicted from the cache. Move it to the
4796 if (HDR_PREFETCH(hdr
)) {
4797 new_state
= arc_mru
;
4798 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0)
4799 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
4800 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
4802 new_state
= arc_mfu
;
4803 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4806 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4807 arc_change_state(new_state
, hdr
, hash_lock
);
4809 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_ghost_hits
);
4810 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
4811 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu
) {
4813 * This buffer has been accessed more than once and is
4814 * still in the cache. Keep it in the MFU state.
4816 * NOTE: an add_reference() that occurred when we did
4817 * the arc_read() will have kicked this off the list.
4818 * If it was a prefetch, we will explicitly move it to
4819 * the head of the list now.
4821 if ((HDR_PREFETCH(hdr
)) != 0) {
4822 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4823 /* link protected by hash_lock */
4824 ASSERT(multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4826 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_hits
);
4827 ARCSTAT_BUMP(arcstat_mfu_hits
);
4828 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4829 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu_ghost
) {
4830 arc_state_t
*new_state
= arc_mfu
;
4832 * This buffer has been accessed more than once but has
4833 * been evicted from the cache. Move it back to the
4837 if (HDR_PREFETCH(hdr
)) {
4839 * This is a prefetch access...
4840 * move this block back to the MRU state.
4842 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
4843 new_state
= arc_mru
;
4846 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4847 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4848 arc_change_state(new_state
, hdr
, hash_lock
);
4850 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_ghost_hits
);
4851 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
4852 } else if (hdr
->b_l1hdr
.b_state
== arc_l2c_only
) {
4854 * This buffer is on the 2nd Level ARC.
4857 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4858 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4859 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4861 cmn_err(CE_PANIC
, "invalid arc state 0x%p",
4862 hdr
->b_l1hdr
.b_state
);
4866 /* a generic arc_done_func_t which you can use */
4869 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4871 if (zio
== NULL
|| zio
->io_error
== 0)
4872 bcopy(buf
->b_data
, arg
, arc_buf_size(buf
));
4873 arc_buf_destroy(buf
, arg
);
4876 /* a generic arc_done_func_t */
4878 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4880 arc_buf_t
**bufp
= arg
;
4881 if (zio
&& zio
->io_error
) {
4882 arc_buf_destroy(buf
, arg
);
4886 ASSERT(buf
->b_data
);
4891 arc_hdr_verify(arc_buf_hdr_t
*hdr
, blkptr_t
*bp
)
4893 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
)) {
4894 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, 0);
4895 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
4897 if (HDR_COMPRESSION_ENABLED(hdr
)) {
4898 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==,
4899 BP_GET_COMPRESS(bp
));
4901 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(bp
));
4902 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, BP_GET_PSIZE(bp
));
4907 arc_read_done(zio_t
*zio
)
4909 arc_buf_hdr_t
*hdr
= zio
->io_private
;
4910 kmutex_t
*hash_lock
= NULL
;
4911 arc_callback_t
*callback_list
;
4912 arc_callback_t
*acb
;
4913 boolean_t freeable
= B_FALSE
;
4914 boolean_t no_zio_error
= (zio
->io_error
== 0);
4917 * The hdr was inserted into hash-table and removed from lists
4918 * prior to starting I/O. We should find this header, since
4919 * it's in the hash table, and it should be legit since it's
4920 * not possible to evict it during the I/O. The only possible
4921 * reason for it not to be found is if we were freed during the
4924 if (HDR_IN_HASH_TABLE(hdr
)) {
4925 arc_buf_hdr_t
*found
;
4927 ASSERT3U(hdr
->b_birth
, ==, BP_PHYSICAL_BIRTH(zio
->io_bp
));
4928 ASSERT3U(hdr
->b_dva
.dva_word
[0], ==,
4929 BP_IDENTITY(zio
->io_bp
)->dva_word
[0]);
4930 ASSERT3U(hdr
->b_dva
.dva_word
[1], ==,
4931 BP_IDENTITY(zio
->io_bp
)->dva_word
[1]);
4933 found
= buf_hash_find(hdr
->b_spa
, zio
->io_bp
, &hash_lock
);
4935 ASSERT((found
== hdr
&&
4936 DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
4937 (found
== hdr
&& HDR_L2_READING(hdr
)));
4938 ASSERT3P(hash_lock
, !=, NULL
);
4942 /* byteswap if necessary */
4943 if (BP_SHOULD_BYTESWAP(zio
->io_bp
)) {
4944 if (BP_GET_LEVEL(zio
->io_bp
) > 0) {
4945 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_UINT64
;
4947 hdr
->b_l1hdr
.b_byteswap
=
4948 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
4951 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
4955 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_EVICTED
);
4956 if (l2arc_noprefetch
&& HDR_PREFETCH(hdr
))
4957 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2CACHE
);
4959 callback_list
= hdr
->b_l1hdr
.b_acb
;
4960 ASSERT3P(callback_list
, !=, NULL
);
4962 if (hash_lock
&& no_zio_error
&& hdr
->b_l1hdr
.b_state
== arc_anon
) {
4964 * Only call arc_access on anonymous buffers. This is because
4965 * if we've issued an I/O for an evicted buffer, we've already
4966 * called arc_access (to prevent any simultaneous readers from
4967 * getting confused).
4969 arc_access(hdr
, hash_lock
);
4973 * If a read request has a callback (i.e. acb_done is not NULL), then we
4974 * make a buf containing the data according to the parameters which were
4975 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
4976 * aren't needlessly decompressing the data multiple times.
4978 int callback_cnt
= 0;
4979 for (acb
= callback_list
; acb
!= NULL
; acb
= acb
->acb_next
) {
4983 /* This is a demand read since prefetches don't use callbacks */
4986 int error
= arc_buf_alloc_impl(hdr
, acb
->acb_private
,
4987 acb
->acb_compressed
, no_zio_error
, &acb
->acb_buf
);
4989 zio
->io_error
= error
;
4992 hdr
->b_l1hdr
.b_acb
= NULL
;
4993 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
4994 if (callback_cnt
== 0) {
4995 ASSERT(HDR_PREFETCH(hdr
));
4996 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
4997 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5000 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
) ||
5001 callback_list
!= NULL
);
5004 arc_hdr_verify(hdr
, zio
->io_bp
);
5006 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_ERROR
);
5007 if (hdr
->b_l1hdr
.b_state
!= arc_anon
)
5008 arc_change_state(arc_anon
, hdr
, hash_lock
);
5009 if (HDR_IN_HASH_TABLE(hdr
))
5010 buf_hash_remove(hdr
);
5011 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
5015 * Broadcast before we drop the hash_lock to avoid the possibility
5016 * that the hdr (and hence the cv) might be freed before we get to
5017 * the cv_broadcast().
5019 cv_broadcast(&hdr
->b_l1hdr
.b_cv
);
5021 if (hash_lock
!= NULL
) {
5022 mutex_exit(hash_lock
);
5025 * This block was freed while we waited for the read to
5026 * complete. It has been removed from the hash table and
5027 * moved to the anonymous state (so that it won't show up
5030 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
5031 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
5034 /* execute each callback and free its structure */
5035 while ((acb
= callback_list
) != NULL
) {
5037 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
5039 if (acb
->acb_zio_dummy
!= NULL
) {
5040 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
5041 zio_nowait(acb
->acb_zio_dummy
);
5044 callback_list
= acb
->acb_next
;
5045 kmem_free(acb
, sizeof (arc_callback_t
));
5049 arc_hdr_destroy(hdr
);
5053 * "Read" the block at the specified DVA (in bp) via the
5054 * cache. If the block is found in the cache, invoke the provided
5055 * callback immediately and return. Note that the `zio' parameter
5056 * in the callback will be NULL in this case, since no IO was
5057 * required. If the block is not in the cache pass the read request
5058 * on to the spa with a substitute callback function, so that the
5059 * requested block will be added to the cache.
5061 * If a read request arrives for a block that has a read in-progress,
5062 * either wait for the in-progress read to complete (and return the
5063 * results); or, if this is a read with a "done" func, add a record
5064 * to the read to invoke the "done" func when the read completes,
5065 * and return; or just return.
5067 * arc_read_done() will invoke all the requested "done" functions
5068 * for readers of this block.
5071 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_done_func_t
*done
,
5072 void *private, zio_priority_t priority
, int zio_flags
,
5073 arc_flags_t
*arc_flags
, const zbookmark_phys_t
*zb
)
5075 arc_buf_hdr_t
*hdr
= NULL
;
5076 kmutex_t
*hash_lock
= NULL
;
5078 uint64_t guid
= spa_load_guid(spa
);
5079 boolean_t compressed_read
= (zio_flags
& ZIO_FLAG_RAW
) != 0;
5082 ASSERT(!BP_IS_EMBEDDED(bp
) ||
5083 BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
);
5086 if (!BP_IS_EMBEDDED(bp
)) {
5088 * Embedded BP's have no DVA and require no I/O to "read".
5089 * Create an anonymous arc buf to back it.
5091 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
5094 if (hdr
!= NULL
&& HDR_HAS_L1HDR(hdr
) && hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5095 arc_buf_t
*buf
= NULL
;
5096 *arc_flags
|= ARC_FLAG_CACHED
;
5098 if (HDR_IO_IN_PROGRESS(hdr
)) {
5100 if ((hdr
->b_flags
& ARC_FLAG_PRIO_ASYNC_READ
) &&
5101 priority
== ZIO_PRIORITY_SYNC_READ
) {
5103 * This sync read must wait for an
5104 * in-progress async read (e.g. a predictive
5105 * prefetch). Async reads are queued
5106 * separately at the vdev_queue layer, so
5107 * this is a form of priority inversion.
5108 * Ideally, we would "inherit" the demand
5109 * i/o's priority by moving the i/o from
5110 * the async queue to the synchronous queue,
5111 * but there is currently no mechanism to do
5112 * so. Track this so that we can evaluate
5113 * the magnitude of this potential performance
5116 * Note that if the prefetch i/o is already
5117 * active (has been issued to the device),
5118 * the prefetch improved performance, because
5119 * we issued it sooner than we would have
5120 * without the prefetch.
5122 DTRACE_PROBE1(arc__sync__wait__for__async
,
5123 arc_buf_hdr_t
*, hdr
);
5124 ARCSTAT_BUMP(arcstat_sync_wait_for_async
);
5126 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5127 arc_hdr_clear_flags(hdr
,
5128 ARC_FLAG_PREDICTIVE_PREFETCH
);
5131 if (*arc_flags
& ARC_FLAG_WAIT
) {
5132 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
5133 mutex_exit(hash_lock
);
5136 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5139 arc_callback_t
*acb
= NULL
;
5141 acb
= kmem_zalloc(sizeof (arc_callback_t
),
5143 acb
->acb_done
= done
;
5144 acb
->acb_private
= private;
5145 acb
->acb_compressed
= compressed_read
;
5147 acb
->acb_zio_dummy
= zio_null(pio
,
5148 spa
, NULL
, NULL
, NULL
, zio_flags
);
5150 ASSERT3P(acb
->acb_done
, !=, NULL
);
5151 acb
->acb_next
= hdr
->b_l1hdr
.b_acb
;
5152 hdr
->b_l1hdr
.b_acb
= acb
;
5153 mutex_exit(hash_lock
);
5156 mutex_exit(hash_lock
);
5160 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
5161 hdr
->b_l1hdr
.b_state
== arc_mfu
);
5164 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5166 * This is a demand read which does not have to
5167 * wait for i/o because we did a predictive
5168 * prefetch i/o for it, which has completed.
5171 arc__demand__hit__predictive__prefetch
,
5172 arc_buf_hdr_t
*, hdr
);
5174 arcstat_demand_hit_predictive_prefetch
);
5175 arc_hdr_clear_flags(hdr
,
5176 ARC_FLAG_PREDICTIVE_PREFETCH
);
5178 ASSERT(!BP_IS_EMBEDDED(bp
) || !BP_IS_HOLE(bp
));
5180 /* Get a buf with the desired data in it. */
5181 VERIFY0(arc_buf_alloc_impl(hdr
, private,
5182 compressed_read
, B_TRUE
, &buf
));
5183 } else if (*arc_flags
& ARC_FLAG_PREFETCH
&&
5184 refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
5185 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
5187 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
5188 arc_access(hdr
, hash_lock
);
5189 if (*arc_flags
& ARC_FLAG_L2CACHE
)
5190 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5191 mutex_exit(hash_lock
);
5192 ARCSTAT_BUMP(arcstat_hits
);
5193 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
5194 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
5195 data
, metadata
, hits
);
5198 done(NULL
, buf
, private);
5200 uint64_t lsize
= BP_GET_LSIZE(bp
);
5201 uint64_t psize
= BP_GET_PSIZE(bp
);
5202 arc_callback_t
*acb
;
5205 boolean_t devw
= B_FALSE
;
5209 * Gracefully handle a damaged logical block size as a
5212 if (lsize
> spa_maxblocksize(spa
)) {
5213 rc
= SET_ERROR(ECKSUM
);
5218 /* this block is not in the cache */
5219 arc_buf_hdr_t
*exists
= NULL
;
5220 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
5221 hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
5222 BP_GET_COMPRESS(bp
), type
);
5224 if (!BP_IS_EMBEDDED(bp
)) {
5225 hdr
->b_dva
= *BP_IDENTITY(bp
);
5226 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
5227 exists
= buf_hash_insert(hdr
, &hash_lock
);
5229 if (exists
!= NULL
) {
5230 /* somebody beat us to the hash insert */
5231 mutex_exit(hash_lock
);
5232 buf_discard_identity(hdr
);
5233 arc_hdr_destroy(hdr
);
5234 goto top
; /* restart the IO request */
5238 * This block is in the ghost cache. If it was L2-only
5239 * (and thus didn't have an L1 hdr), we realloc the
5240 * header to add an L1 hdr.
5242 if (!HDR_HAS_L1HDR(hdr
)) {
5243 hdr
= arc_hdr_realloc(hdr
, hdr_l2only_cache
,
5247 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5248 ASSERT(GHOST_STATE(hdr
->b_l1hdr
.b_state
));
5249 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5250 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5251 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
5252 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
5255 * This is a delicate dance that we play here.
5256 * This hdr is in the ghost list so we access it
5257 * to move it out of the ghost list before we
5258 * initiate the read. If it's a prefetch then
5259 * it won't have a callback so we'll remove the
5260 * reference that arc_buf_alloc_impl() created. We
5261 * do this after we've called arc_access() to
5262 * avoid hitting an assert in remove_reference().
5264 arc_access(hdr
, hash_lock
);
5265 arc_hdr_alloc_pabd(hdr
);
5267 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5268 size
= arc_hdr_size(hdr
);
5271 * If compression is enabled on the hdr, then will do
5272 * RAW I/O and will store the compressed data in the hdr's
5273 * data block. Otherwise, the hdr's data block will contain
5274 * the uncompressed data.
5276 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
) {
5277 zio_flags
|= ZIO_FLAG_RAW
;
5280 if (*arc_flags
& ARC_FLAG_PREFETCH
)
5281 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
5282 if (*arc_flags
& ARC_FLAG_L2CACHE
)
5283 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5284 if (BP_GET_LEVEL(bp
) > 0)
5285 arc_hdr_set_flags(hdr
, ARC_FLAG_INDIRECT
);
5286 if (*arc_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
)
5287 arc_hdr_set_flags(hdr
, ARC_FLAG_PREDICTIVE_PREFETCH
);
5288 ASSERT(!GHOST_STATE(hdr
->b_l1hdr
.b_state
));
5290 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
5291 acb
->acb_done
= done
;
5292 acb
->acb_private
= private;
5293 acb
->acb_compressed
= compressed_read
;
5295 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5296 hdr
->b_l1hdr
.b_acb
= acb
;
5297 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5299 if (HDR_HAS_L2HDR(hdr
) &&
5300 (vd
= hdr
->b_l2hdr
.b_dev
->l2ad_vdev
) != NULL
) {
5301 devw
= hdr
->b_l2hdr
.b_dev
->l2ad_writing
;
5302 addr
= hdr
->b_l2hdr
.b_daddr
;
5304 * Lock out device removal.
5306 if (vdev_is_dead(vd
) ||
5307 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
5311 if (priority
== ZIO_PRIORITY_ASYNC_READ
)
5312 arc_hdr_set_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
5314 arc_hdr_clear_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
5316 if (hash_lock
!= NULL
)
5317 mutex_exit(hash_lock
);
5320 * At this point, we have a level 1 cache miss. Try again in
5321 * L2ARC if possible.
5323 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, lsize
);
5325 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
5326 uint64_t, lsize
, zbookmark_phys_t
*, zb
);
5327 ARCSTAT_BUMP(arcstat_misses
);
5328 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
5329 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
5330 data
, metadata
, misses
);
5332 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
5334 * Read from the L2ARC if the following are true:
5335 * 1. The L2ARC vdev was previously cached.
5336 * 2. This buffer still has L2ARC metadata.
5337 * 3. This buffer isn't currently writing to the L2ARC.
5338 * 4. The L2ARC entry wasn't evicted, which may
5339 * also have invalidated the vdev.
5340 * 5. This isn't prefetch and l2arc_noprefetch is set.
5342 if (HDR_HAS_L2HDR(hdr
) &&
5343 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
5344 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
5345 l2arc_read_callback_t
*cb
;
5347 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
5348 ARCSTAT_BUMP(arcstat_l2_hits
);
5349 atomic_inc_32(&hdr
->b_l2hdr
.b_hits
);
5351 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
5353 cb
->l2rcb_hdr
= hdr
;
5356 cb
->l2rcb_flags
= zio_flags
;
5358 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
5359 addr
+ lsize
< vd
->vdev_psize
-
5360 VDEV_LABEL_END_SIZE
);
5363 * l2arc read. The SCL_L2ARC lock will be
5364 * released by l2arc_read_done().
5365 * Issue a null zio if the underlying buffer
5366 * was squashed to zero size by compression.
5368 ASSERT3U(HDR_GET_COMPRESS(hdr
), !=,
5369 ZIO_COMPRESS_EMPTY
);
5370 rzio
= zio_read_phys(pio
, vd
, addr
,
5371 size
, hdr
->b_l1hdr
.b_pabd
,
5373 l2arc_read_done
, cb
, priority
,
5374 zio_flags
| ZIO_FLAG_DONT_CACHE
|
5376 ZIO_FLAG_DONT_PROPAGATE
|
5377 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
5379 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
5381 ARCSTAT_INCR(arcstat_l2_read_bytes
, size
);
5383 if (*arc_flags
& ARC_FLAG_NOWAIT
) {
5388 ASSERT(*arc_flags
& ARC_FLAG_WAIT
);
5389 if (zio_wait(rzio
) == 0)
5392 /* l2arc read error; goto zio_read() */
5394 DTRACE_PROBE1(l2arc__miss
,
5395 arc_buf_hdr_t
*, hdr
);
5396 ARCSTAT_BUMP(arcstat_l2_misses
);
5397 if (HDR_L2_WRITING(hdr
))
5398 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
5399 spa_config_exit(spa
, SCL_L2ARC
, vd
);
5403 spa_config_exit(spa
, SCL_L2ARC
, vd
);
5404 if (l2arc_ndev
!= 0) {
5405 DTRACE_PROBE1(l2arc__miss
,
5406 arc_buf_hdr_t
*, hdr
);
5407 ARCSTAT_BUMP(arcstat_l2_misses
);
5411 rzio
= zio_read(pio
, spa
, bp
, hdr
->b_l1hdr
.b_pabd
, size
,
5412 arc_read_done
, hdr
, priority
, zio_flags
, zb
);
5414 if (*arc_flags
& ARC_FLAG_WAIT
) {
5415 rc
= zio_wait(rzio
);
5419 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5424 spa_read_history_add(spa
, zb
, *arc_flags
);
5429 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
5433 p
= kmem_alloc(sizeof (*p
), KM_SLEEP
);
5435 p
->p_private
= private;
5436 list_link_init(&p
->p_node
);
5437 refcount_create(&p
->p_refcnt
);
5439 mutex_enter(&arc_prune_mtx
);
5440 refcount_add(&p
->p_refcnt
, &arc_prune_list
);
5441 list_insert_head(&arc_prune_list
, p
);
5442 mutex_exit(&arc_prune_mtx
);
5448 arc_remove_prune_callback(arc_prune_t
*p
)
5450 boolean_t wait
= B_FALSE
;
5451 mutex_enter(&arc_prune_mtx
);
5452 list_remove(&arc_prune_list
, p
);
5453 if (refcount_remove(&p
->p_refcnt
, &arc_prune_list
) > 0)
5455 mutex_exit(&arc_prune_mtx
);
5457 /* wait for arc_prune_task to finish */
5459 taskq_wait_outstanding(arc_prune_taskq
, 0);
5460 ASSERT0(refcount_count(&p
->p_refcnt
));
5461 refcount_destroy(&p
->p_refcnt
);
5462 kmem_free(p
, sizeof (*p
));
5466 * Notify the arc that a block was freed, and thus will never be used again.
5469 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
5472 kmutex_t
*hash_lock
;
5473 uint64_t guid
= spa_load_guid(spa
);
5475 ASSERT(!BP_IS_EMBEDDED(bp
));
5477 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
5482 * We might be trying to free a block that is still doing I/O
5483 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5484 * dmu_sync-ed block). If this block is being prefetched, then it
5485 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5486 * until the I/O completes. A block may also have a reference if it is
5487 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5488 * have written the new block to its final resting place on disk but
5489 * without the dedup flag set. This would have left the hdr in the MRU
5490 * state and discoverable. When the txg finally syncs it detects that
5491 * the block was overridden in open context and issues an override I/O.
5492 * Since this is a dedup block, the override I/O will determine if the
5493 * block is already in the DDT. If so, then it will replace the io_bp
5494 * with the bp from the DDT and allow the I/O to finish. When the I/O
5495 * reaches the done callback, dbuf_write_override_done, it will
5496 * check to see if the io_bp and io_bp_override are identical.
5497 * If they are not, then it indicates that the bp was replaced with
5498 * the bp in the DDT and the override bp is freed. This allows
5499 * us to arrive here with a reference on a block that is being
5500 * freed. So if we have an I/O in progress, or a reference to
5501 * this hdr, then we don't destroy the hdr.
5503 if (!HDR_HAS_L1HDR(hdr
) || (!HDR_IO_IN_PROGRESS(hdr
) &&
5504 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
))) {
5505 arc_change_state(arc_anon
, hdr
, hash_lock
);
5506 arc_hdr_destroy(hdr
);
5507 mutex_exit(hash_lock
);
5509 mutex_exit(hash_lock
);
5515 * Release this buffer from the cache, making it an anonymous buffer. This
5516 * must be done after a read and prior to modifying the buffer contents.
5517 * If the buffer has more than one reference, we must make
5518 * a new hdr for the buffer.
5521 arc_release(arc_buf_t
*buf
, void *tag
)
5523 kmutex_t
*hash_lock
;
5525 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5528 * It would be nice to assert that if its DMU metadata (level >
5529 * 0 || it's the dnode file), then it must be syncing context.
5530 * But we don't know that information at this level.
5533 mutex_enter(&buf
->b_evict_lock
);
5535 ASSERT(HDR_HAS_L1HDR(hdr
));
5538 * We don't grab the hash lock prior to this check, because if
5539 * the buffer's header is in the arc_anon state, it won't be
5540 * linked into the hash table.
5542 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
5543 mutex_exit(&buf
->b_evict_lock
);
5544 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5545 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
5546 ASSERT(!HDR_HAS_L2HDR(hdr
));
5547 ASSERT(HDR_EMPTY(hdr
));
5549 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
5550 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), ==, 1);
5551 ASSERT(!list_link_active(&hdr
->b_l1hdr
.b_arc_node
));
5553 hdr
->b_l1hdr
.b_arc_access
= 0;
5556 * If the buf is being overridden then it may already
5557 * have a hdr that is not empty.
5559 buf_discard_identity(hdr
);
5565 hash_lock
= HDR_LOCK(hdr
);
5566 mutex_enter(hash_lock
);
5569 * This assignment is only valid as long as the hash_lock is
5570 * held, we must be careful not to reference state or the
5571 * b_state field after dropping the lock.
5573 state
= hdr
->b_l1hdr
.b_state
;
5574 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
5575 ASSERT3P(state
, !=, arc_anon
);
5577 /* this buffer is not on any list */
5578 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), >, 0);
5580 if (HDR_HAS_L2HDR(hdr
)) {
5581 mutex_enter(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
5584 * We have to recheck this conditional again now that
5585 * we're holding the l2ad_mtx to prevent a race with
5586 * another thread which might be concurrently calling
5587 * l2arc_evict(). In that case, l2arc_evict() might have
5588 * destroyed the header's L2 portion as we were waiting
5589 * to acquire the l2ad_mtx.
5591 if (HDR_HAS_L2HDR(hdr
))
5592 arc_hdr_l2hdr_destroy(hdr
);
5594 mutex_exit(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
5598 * Do we have more than one buf?
5600 if (hdr
->b_l1hdr
.b_bufcnt
> 1) {
5601 arc_buf_hdr_t
*nhdr
;
5602 uint64_t spa
= hdr
->b_spa
;
5603 uint64_t psize
= HDR_GET_PSIZE(hdr
);
5604 uint64_t lsize
= HDR_GET_LSIZE(hdr
);
5605 enum zio_compress compress
= HDR_GET_COMPRESS(hdr
);
5606 arc_buf_contents_t type
= arc_buf_type(hdr
);
5607 VERIFY3U(hdr
->b_type
, ==, type
);
5609 ASSERT(hdr
->b_l1hdr
.b_buf
!= buf
|| buf
->b_next
!= NULL
);
5610 (void) remove_reference(hdr
, hash_lock
, tag
);
5612 if (arc_buf_is_shared(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
5613 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
5614 ASSERT(ARC_BUF_LAST(buf
));
5618 * Pull the data off of this hdr and attach it to
5619 * a new anonymous hdr. Also find the last buffer
5620 * in the hdr's buffer list.
5622 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
5623 ASSERT3P(lastbuf
, !=, NULL
);
5626 * If the current arc_buf_t and the hdr are sharing their data
5627 * buffer, then we must stop sharing that block.
5629 if (arc_buf_is_shared(buf
)) {
5630 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
5631 VERIFY(!arc_buf_is_shared(lastbuf
));
5634 * First, sever the block sharing relationship between
5635 * buf and the arc_buf_hdr_t.
5637 arc_unshare_buf(hdr
, buf
);
5640 * Now we need to recreate the hdr's b_pabd. Since we
5641 * have lastbuf handy, we try to share with it, but if
5642 * we can't then we allocate a new b_pabd and copy the
5643 * data from buf into it.
5645 if (arc_can_share(hdr
, lastbuf
)) {
5646 arc_share_buf(hdr
, lastbuf
);
5648 arc_hdr_alloc_pabd(hdr
);
5649 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
,
5650 buf
->b_data
, psize
);
5652 VERIFY3P(lastbuf
->b_data
, !=, NULL
);
5653 } else if (HDR_SHARED_DATA(hdr
)) {
5655 * Uncompressed shared buffers are always at the end
5656 * of the list. Compressed buffers don't have the
5657 * same requirements. This makes it hard to
5658 * simply assert that the lastbuf is shared so
5659 * we rely on the hdr's compression flags to determine
5660 * if we have a compressed, shared buffer.
5662 ASSERT(arc_buf_is_shared(lastbuf
) ||
5663 HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
5664 ASSERT(!ARC_BUF_SHARED(buf
));
5666 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5667 ASSERT3P(state
, !=, arc_l2c_only
);
5669 (void) refcount_remove_many(&state
->arcs_size
,
5670 arc_buf_size(buf
), buf
);
5672 if (refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
)) {
5673 ASSERT3P(state
, !=, arc_l2c_only
);
5674 (void) refcount_remove_many(&state
->arcs_esize
[type
],
5675 arc_buf_size(buf
), buf
);
5678 hdr
->b_l1hdr
.b_bufcnt
-= 1;
5679 arc_cksum_verify(buf
);
5680 arc_buf_unwatch(buf
);
5682 mutex_exit(hash_lock
);
5685 * Allocate a new hdr. The new hdr will contain a b_pabd
5686 * buffer which will be freed in arc_write().
5688 nhdr
= arc_hdr_alloc(spa
, psize
, lsize
, compress
, type
);
5689 ASSERT3P(nhdr
->b_l1hdr
.b_buf
, ==, NULL
);
5690 ASSERT0(nhdr
->b_l1hdr
.b_bufcnt
);
5691 ASSERT0(refcount_count(&nhdr
->b_l1hdr
.b_refcnt
));
5692 VERIFY3U(nhdr
->b_type
, ==, type
);
5693 ASSERT(!HDR_SHARED_DATA(nhdr
));
5695 nhdr
->b_l1hdr
.b_buf
= buf
;
5696 nhdr
->b_l1hdr
.b_bufcnt
= 1;
5697 nhdr
->b_l1hdr
.b_mru_hits
= 0;
5698 nhdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
5699 nhdr
->b_l1hdr
.b_mfu_hits
= 0;
5700 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
5701 nhdr
->b_l1hdr
.b_l2_hits
= 0;
5702 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, tag
);
5705 mutex_exit(&buf
->b_evict_lock
);
5706 (void) refcount_add_many(&arc_anon
->arcs_size
,
5707 HDR_GET_LSIZE(nhdr
), buf
);
5709 mutex_exit(&buf
->b_evict_lock
);
5710 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 1);
5711 /* protected by hash lock, or hdr is on arc_anon */
5712 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
5713 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5714 hdr
->b_l1hdr
.b_mru_hits
= 0;
5715 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
5716 hdr
->b_l1hdr
.b_mfu_hits
= 0;
5717 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
5718 hdr
->b_l1hdr
.b_l2_hits
= 0;
5719 arc_change_state(arc_anon
, hdr
, hash_lock
);
5720 hdr
->b_l1hdr
.b_arc_access
= 0;
5721 mutex_exit(hash_lock
);
5723 buf_discard_identity(hdr
);
5729 arc_released(arc_buf_t
*buf
)
5733 mutex_enter(&buf
->b_evict_lock
);
5734 released
= (buf
->b_data
!= NULL
&&
5735 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
5736 mutex_exit(&buf
->b_evict_lock
);
5742 arc_referenced(arc_buf_t
*buf
)
5746 mutex_enter(&buf
->b_evict_lock
);
5747 referenced
= (refcount_count(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
5748 mutex_exit(&buf
->b_evict_lock
);
5749 return (referenced
);
5754 arc_write_ready(zio_t
*zio
)
5756 arc_write_callback_t
*callback
= zio
->io_private
;
5757 arc_buf_t
*buf
= callback
->awcb_buf
;
5758 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5759 uint64_t psize
= BP_IS_HOLE(zio
->io_bp
) ? 0 : BP_GET_PSIZE(zio
->io_bp
);
5760 enum zio_compress compress
;
5761 fstrans_cookie_t cookie
= spl_fstrans_mark();
5763 ASSERT(HDR_HAS_L1HDR(hdr
));
5764 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
5765 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
5768 * If we're reexecuting this zio because the pool suspended, then
5769 * cleanup any state that was previously set the first time the
5770 * callback was invoked.
5772 if (zio
->io_flags
& ZIO_FLAG_REEXECUTED
) {
5773 arc_cksum_free(hdr
);
5774 arc_buf_unwatch(buf
);
5775 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5776 if (arc_buf_is_shared(buf
)) {
5777 arc_unshare_buf(hdr
, buf
);
5779 arc_hdr_free_pabd(hdr
);
5783 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5784 ASSERT(!HDR_SHARED_DATA(hdr
));
5785 ASSERT(!arc_buf_is_shared(buf
));
5787 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
5789 if (HDR_IO_IN_PROGRESS(hdr
))
5790 ASSERT(zio
->io_flags
& ZIO_FLAG_REEXECUTED
);
5792 arc_cksum_compute(buf
);
5793 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5795 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
5796 compress
= ZIO_COMPRESS_OFF
;
5798 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(zio
->io_bp
));
5799 compress
= BP_GET_COMPRESS(zio
->io_bp
);
5801 HDR_SET_PSIZE(hdr
, psize
);
5802 arc_hdr_set_compress(hdr
, compress
);
5805 * Fill the hdr with data. If the hdr is compressed, the data we want
5806 * is available from the zio, otherwise we can take it from the buf.
5808 * We might be able to share the buf's data with the hdr here. However,
5809 * doing so would cause the ARC to be full of linear ABDs if we write a
5810 * lot of shareable data. As a compromise, we check whether scattered
5811 * ABDs are allowed, and assume that if they are then the user wants
5812 * the ARC to be primarily filled with them regardless of the data being
5813 * written. Therefore, if they're allowed then we allocate one and copy
5814 * the data into it; otherwise, we share the data directly if we can.
5816 if (zfs_abd_scatter_enabled
|| !arc_can_share(hdr
, buf
)) {
5817 arc_hdr_alloc_pabd(hdr
);
5820 * Ideally, we would always copy the io_abd into b_pabd, but the
5821 * user may have disabled compressed ARC, thus we must check the
5822 * hdr's compression setting rather than the io_bp's.
5824 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
) {
5825 ASSERT3U(BP_GET_COMPRESS(zio
->io_bp
), !=,
5827 ASSERT3U(psize
, >, 0);
5829 abd_copy(hdr
->b_l1hdr
.b_pabd
, zio
->io_abd
, psize
);
5831 ASSERT3U(zio
->io_orig_size
, ==, arc_hdr_size(hdr
));
5833 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
5837 ASSERT3P(buf
->b_data
, ==, abd_to_buf(zio
->io_orig_abd
));
5838 ASSERT3U(zio
->io_orig_size
, ==, arc_buf_size(buf
));
5839 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
5841 arc_share_buf(hdr
, buf
);
5844 arc_hdr_verify(hdr
, zio
->io_bp
);
5845 spl_fstrans_unmark(cookie
);
5849 arc_write_children_ready(zio_t
*zio
)
5851 arc_write_callback_t
*callback
= zio
->io_private
;
5852 arc_buf_t
*buf
= callback
->awcb_buf
;
5854 callback
->awcb_children_ready(zio
, buf
, callback
->awcb_private
);
5858 * The SPA calls this callback for each physical write that happens on behalf
5859 * of a logical write. See the comment in dbuf_write_physdone() for details.
5862 arc_write_physdone(zio_t
*zio
)
5864 arc_write_callback_t
*cb
= zio
->io_private
;
5865 if (cb
->awcb_physdone
!= NULL
)
5866 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
5870 arc_write_done(zio_t
*zio
)
5872 arc_write_callback_t
*callback
= zio
->io_private
;
5873 arc_buf_t
*buf
= callback
->awcb_buf
;
5874 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5876 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5878 if (zio
->io_error
== 0) {
5879 arc_hdr_verify(hdr
, zio
->io_bp
);
5881 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
5882 buf_discard_identity(hdr
);
5884 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
5885 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
5888 ASSERT(HDR_EMPTY(hdr
));
5892 * If the block to be written was all-zero or compressed enough to be
5893 * embedded in the BP, no write was performed so there will be no
5894 * dva/birth/checksum. The buffer must therefore remain anonymous
5897 if (!HDR_EMPTY(hdr
)) {
5898 arc_buf_hdr_t
*exists
;
5899 kmutex_t
*hash_lock
;
5901 ASSERT3U(zio
->io_error
, ==, 0);
5903 arc_cksum_verify(buf
);
5905 exists
= buf_hash_insert(hdr
, &hash_lock
);
5906 if (exists
!= NULL
) {
5908 * This can only happen if we overwrite for
5909 * sync-to-convergence, because we remove
5910 * buffers from the hash table when we arc_free().
5912 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
5913 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
5914 panic("bad overwrite, hdr=%p exists=%p",
5915 (void *)hdr
, (void *)exists
);
5916 ASSERT(refcount_is_zero(
5917 &exists
->b_l1hdr
.b_refcnt
));
5918 arc_change_state(arc_anon
, exists
, hash_lock
);
5919 mutex_exit(hash_lock
);
5920 arc_hdr_destroy(exists
);
5921 exists
= buf_hash_insert(hdr
, &hash_lock
);
5922 ASSERT3P(exists
, ==, NULL
);
5923 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
5925 ASSERT(zio
->io_prop
.zp_nopwrite
);
5926 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
5927 panic("bad nopwrite, hdr=%p exists=%p",
5928 (void *)hdr
, (void *)exists
);
5931 ASSERT(hdr
->b_l1hdr
.b_bufcnt
== 1);
5932 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
5933 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
5934 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
5937 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5938 /* if it's not anon, we are doing a scrub */
5939 if (exists
== NULL
&& hdr
->b_l1hdr
.b_state
== arc_anon
)
5940 arc_access(hdr
, hash_lock
);
5941 mutex_exit(hash_lock
);
5943 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5946 ASSERT(!refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5947 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
5949 abd_put(zio
->io_abd
);
5950 kmem_free(callback
, sizeof (arc_write_callback_t
));
5954 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
5955 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
,
5956 const zio_prop_t
*zp
, arc_done_func_t
*ready
,
5957 arc_done_func_t
*children_ready
, arc_done_func_t
*physdone
,
5958 arc_done_func_t
*done
, void *private, zio_priority_t priority
,
5959 int zio_flags
, const zbookmark_phys_t
*zb
)
5961 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5962 arc_write_callback_t
*callback
;
5965 ASSERT3P(ready
, !=, NULL
);
5966 ASSERT3P(done
, !=, NULL
);
5967 ASSERT(!HDR_IO_ERROR(hdr
));
5968 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5969 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5970 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
5972 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5973 if (ARC_BUF_COMPRESSED(buf
)) {
5974 ASSERT3U(zp
->zp_compress
, !=, ZIO_COMPRESS_OFF
);
5975 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, arc_buf_size(buf
));
5976 zio_flags
|= ZIO_FLAG_RAW
;
5978 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
5979 callback
->awcb_ready
= ready
;
5980 callback
->awcb_children_ready
= children_ready
;
5981 callback
->awcb_physdone
= physdone
;
5982 callback
->awcb_done
= done
;
5983 callback
->awcb_private
= private;
5984 callback
->awcb_buf
= buf
;
5987 * The hdr's b_pabd is now stale, free it now. A new data block
5988 * will be allocated when the zio pipeline calls arc_write_ready().
5990 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5992 * If the buf is currently sharing the data block with
5993 * the hdr then we need to break that relationship here.
5994 * The hdr will remain with a NULL data pointer and the
5995 * buf will take sole ownership of the block.
5997 if (arc_buf_is_shared(buf
)) {
5998 arc_unshare_buf(hdr
, buf
);
6000 arc_hdr_free_pabd(hdr
);
6002 VERIFY3P(buf
->b_data
, !=, NULL
);
6003 arc_hdr_set_compress(hdr
, ZIO_COMPRESS_OFF
);
6005 ASSERT(!arc_buf_is_shared(buf
));
6006 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
6008 zio
= zio_write(pio
, spa
, txg
, bp
,
6009 abd_get_from_buf(buf
->b_data
, HDR_GET_LSIZE(hdr
)),
6010 HDR_GET_LSIZE(hdr
), arc_buf_size(buf
), zp
,
6012 (children_ready
!= NULL
) ? arc_write_children_ready
: NULL
,
6013 arc_write_physdone
, arc_write_done
, callback
,
6014 priority
, zio_flags
, zb
);
6020 arc_memory_throttle(uint64_t reserve
, uint64_t txg
)
6023 uint64_t available_memory
= ptob(freemem
);
6024 static uint64_t page_load
= 0;
6025 static uint64_t last_txg
= 0;
6027 pgcnt_t minfree
= btop(arc_sys_free
/ 4);
6032 MIN(available_memory
, vmem_size(heap_arena
, VMEM_FREE
));
6035 if (available_memory
> arc_all_memory() * arc_lotsfree_percent
/ 100)
6038 if (txg
> last_txg
) {
6043 * If we are in pageout, we know that memory is already tight,
6044 * the arc is already going to be evicting, so we just want to
6045 * continue to let page writes occur as quickly as possible.
6047 if (current_is_kswapd()) {
6048 if (page_load
> MAX(ptob(minfree
), available_memory
) / 4) {
6049 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
6050 return (SET_ERROR(ERESTART
));
6052 /* Note: reserve is inflated, so we deflate */
6053 page_load
+= reserve
/ 8;
6055 } else if (page_load
> 0 && arc_reclaim_needed()) {
6056 /* memory is low, delay before restarting */
6057 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
6058 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
6059 return (SET_ERROR(EAGAIN
));
6067 arc_tempreserve_clear(uint64_t reserve
)
6069 atomic_add_64(&arc_tempreserve
, -reserve
);
6070 ASSERT((int64_t)arc_tempreserve
>= 0);
6074 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
6080 reserve
> arc_c
/4 &&
6081 reserve
* 4 > (2ULL << SPA_MAXBLOCKSHIFT
))
6082 arc_c
= MIN(arc_c_max
, reserve
* 4);
6085 * Throttle when the calculated memory footprint for the TXG
6086 * exceeds the target ARC size.
6088 if (reserve
> arc_c
) {
6089 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
6090 return (SET_ERROR(ERESTART
));
6094 * Don't count loaned bufs as in flight dirty data to prevent long
6095 * network delays from blocking transactions that are ready to be
6096 * assigned to a txg.
6099 /* assert that it has not wrapped around */
6100 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
6102 anon_size
= MAX((int64_t)(refcount_count(&arc_anon
->arcs_size
) -
6103 arc_loaned_bytes
), 0);
6106 * Writes will, almost always, require additional memory allocations
6107 * in order to compress/encrypt/etc the data. We therefore need to
6108 * make sure that there is sufficient available memory for this.
6110 error
= arc_memory_throttle(reserve
, txg
);
6115 * Throttle writes when the amount of dirty data in the cache
6116 * gets too large. We try to keep the cache less than half full
6117 * of dirty blocks so that our sync times don't grow too large.
6118 * Note: if two requests come in concurrently, we might let them
6119 * both succeed, when one of them should fail. Not a huge deal.
6122 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
6123 anon_size
> arc_c
/ 4) {
6124 uint64_t meta_esize
=
6125 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6126 uint64_t data_esize
=
6127 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6128 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6129 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6130 arc_tempreserve
>> 10, meta_esize
>> 10,
6131 data_esize
>> 10, reserve
>> 10, arc_c
>> 10);
6132 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
6133 return (SET_ERROR(ERESTART
));
6135 atomic_add_64(&arc_tempreserve
, reserve
);
6140 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
6141 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
6143 size
->value
.ui64
= refcount_count(&state
->arcs_size
);
6144 evict_data
->value
.ui64
=
6145 refcount_count(&state
->arcs_esize
[ARC_BUFC_DATA
]);
6146 evict_metadata
->value
.ui64
=
6147 refcount_count(&state
->arcs_esize
[ARC_BUFC_METADATA
]);
6151 arc_kstat_update(kstat_t
*ksp
, int rw
)
6153 arc_stats_t
*as
= ksp
->ks_data
;
6155 if (rw
== KSTAT_WRITE
) {
6158 arc_kstat_update_state(arc_anon
,
6159 &as
->arcstat_anon_size
,
6160 &as
->arcstat_anon_evictable_data
,
6161 &as
->arcstat_anon_evictable_metadata
);
6162 arc_kstat_update_state(arc_mru
,
6163 &as
->arcstat_mru_size
,
6164 &as
->arcstat_mru_evictable_data
,
6165 &as
->arcstat_mru_evictable_metadata
);
6166 arc_kstat_update_state(arc_mru_ghost
,
6167 &as
->arcstat_mru_ghost_size
,
6168 &as
->arcstat_mru_ghost_evictable_data
,
6169 &as
->arcstat_mru_ghost_evictable_metadata
);
6170 arc_kstat_update_state(arc_mfu
,
6171 &as
->arcstat_mfu_size
,
6172 &as
->arcstat_mfu_evictable_data
,
6173 &as
->arcstat_mfu_evictable_metadata
);
6174 arc_kstat_update_state(arc_mfu_ghost
,
6175 &as
->arcstat_mfu_ghost_size
,
6176 &as
->arcstat_mfu_ghost_evictable_data
,
6177 &as
->arcstat_mfu_ghost_evictable_metadata
);
6184 * This function *must* return indices evenly distributed between all
6185 * sublists of the multilist. This is needed due to how the ARC eviction
6186 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6187 * distributed between all sublists and uses this assumption when
6188 * deciding which sublist to evict from and how much to evict from it.
6191 arc_state_multilist_index_func(multilist_t
*ml
, void *obj
)
6193 arc_buf_hdr_t
*hdr
= obj
;
6196 * We rely on b_dva to generate evenly distributed index
6197 * numbers using buf_hash below. So, as an added precaution,
6198 * let's make sure we never add empty buffers to the arc lists.
6200 ASSERT(!HDR_EMPTY(hdr
));
6203 * The assumption here, is the hash value for a given
6204 * arc_buf_hdr_t will remain constant throughout its lifetime
6205 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
6206 * Thus, we don't need to store the header's sublist index
6207 * on insertion, as this index can be recalculated on removal.
6209 * Also, the low order bits of the hash value are thought to be
6210 * distributed evenly. Otherwise, in the case that the multilist
6211 * has a power of two number of sublists, each sublists' usage
6212 * would not be evenly distributed.
6214 return (buf_hash(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
) %
6215 multilist_get_num_sublists(ml
));
6219 * Called during module initialization and periodically thereafter to
6220 * apply reasonable changes to the exposed performance tunings. Non-zero
6221 * zfs_* values which differ from the currently set values will be applied.
6224 arc_tuning_update(void)
6226 uint64_t percent
, allmem
= arc_all_memory();
6228 /* Valid range: 64M - <all physical memory> */
6229 if ((zfs_arc_max
) && (zfs_arc_max
!= arc_c_max
) &&
6230 (zfs_arc_max
> 64 << 20) && (zfs_arc_max
< allmem
) &&
6231 (zfs_arc_max
> arc_c_min
)) {
6232 arc_c_max
= zfs_arc_max
;
6234 arc_p
= (arc_c
>> 1);
6235 /* Valid range of arc_meta_limit: arc_meta_min - arc_c_max */
6236 percent
= MIN(zfs_arc_meta_limit_percent
, 100);
6237 arc_meta_limit
= MAX(arc_meta_min
, (percent
* arc_c_max
) / 100);
6238 percent
= MIN(zfs_arc_dnode_limit_percent
, 100);
6239 arc_dnode_limit
= (percent
* arc_meta_limit
) / 100;
6242 /* Valid range: 32M - <arc_c_max> */
6243 if ((zfs_arc_min
) && (zfs_arc_min
!= arc_c_min
) &&
6244 (zfs_arc_min
>= 2ULL << SPA_MAXBLOCKSHIFT
) &&
6245 (zfs_arc_min
<= arc_c_max
)) {
6246 arc_c_min
= zfs_arc_min
;
6247 arc_c
= MAX(arc_c
, arc_c_min
);
6250 /* Valid range: 16M - <arc_c_max> */
6251 if ((zfs_arc_meta_min
) && (zfs_arc_meta_min
!= arc_meta_min
) &&
6252 (zfs_arc_meta_min
>= 1ULL << SPA_MAXBLOCKSHIFT
) &&
6253 (zfs_arc_meta_min
<= arc_c_max
)) {
6254 arc_meta_min
= zfs_arc_meta_min
;
6255 arc_meta_limit
= MAX(arc_meta_limit
, arc_meta_min
);
6256 arc_dnode_limit
= arc_meta_limit
/ 10;
6259 /* Valid range: <arc_meta_min> - <arc_c_max> */
6260 if ((zfs_arc_meta_limit
) && (zfs_arc_meta_limit
!= arc_meta_limit
) &&
6261 (zfs_arc_meta_limit
>= zfs_arc_meta_min
) &&
6262 (zfs_arc_meta_limit
<= arc_c_max
))
6263 arc_meta_limit
= zfs_arc_meta_limit
;
6265 /* Valid range: <arc_meta_min> - <arc_c_max> */
6266 if ((zfs_arc_dnode_limit
) && (zfs_arc_dnode_limit
!= arc_dnode_limit
) &&
6267 (zfs_arc_dnode_limit
>= zfs_arc_meta_min
) &&
6268 (zfs_arc_dnode_limit
<= arc_c_max
))
6269 arc_dnode_limit
= zfs_arc_dnode_limit
;
6271 /* Valid range: 1 - N */
6272 if (zfs_arc_grow_retry
)
6273 arc_grow_retry
= zfs_arc_grow_retry
;
6275 /* Valid range: 1 - N */
6276 if (zfs_arc_shrink_shift
) {
6277 arc_shrink_shift
= zfs_arc_shrink_shift
;
6278 arc_no_grow_shift
= MIN(arc_no_grow_shift
, arc_shrink_shift
-1);
6281 /* Valid range: 1 - N */
6282 if (zfs_arc_p_min_shift
)
6283 arc_p_min_shift
= zfs_arc_p_min_shift
;
6285 /* Valid range: 1 - N ticks */
6286 if (zfs_arc_min_prefetch_lifespan
)
6287 arc_min_prefetch_lifespan
= zfs_arc_min_prefetch_lifespan
;
6289 /* Valid range: 0 - 100 */
6290 if ((zfs_arc_lotsfree_percent
>= 0) &&
6291 (zfs_arc_lotsfree_percent
<= 100))
6292 arc_lotsfree_percent
= zfs_arc_lotsfree_percent
;
6294 /* Valid range: 0 - <all physical memory> */
6295 if ((zfs_arc_sys_free
) && (zfs_arc_sys_free
!= arc_sys_free
))
6296 arc_sys_free
= MIN(MAX(zfs_arc_sys_free
, 0), allmem
);
6301 arc_state_init(void)
6303 arc_anon
= &ARC_anon
;
6305 arc_mru_ghost
= &ARC_mru_ghost
;
6307 arc_mfu_ghost
= &ARC_mfu_ghost
;
6308 arc_l2c_only
= &ARC_l2c_only
;
6310 arc_mru
->arcs_list
[ARC_BUFC_METADATA
] =
6311 multilist_create(sizeof (arc_buf_hdr_t
),
6312 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6313 arc_state_multilist_index_func
);
6314 arc_mru
->arcs_list
[ARC_BUFC_DATA
] =
6315 multilist_create(sizeof (arc_buf_hdr_t
),
6316 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6317 arc_state_multilist_index_func
);
6318 arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
6319 multilist_create(sizeof (arc_buf_hdr_t
),
6320 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6321 arc_state_multilist_index_func
);
6322 arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
] =
6323 multilist_create(sizeof (arc_buf_hdr_t
),
6324 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6325 arc_state_multilist_index_func
);
6326 arc_mfu
->arcs_list
[ARC_BUFC_METADATA
] =
6327 multilist_create(sizeof (arc_buf_hdr_t
),
6328 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6329 arc_state_multilist_index_func
);
6330 arc_mfu
->arcs_list
[ARC_BUFC_DATA
] =
6331 multilist_create(sizeof (arc_buf_hdr_t
),
6332 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6333 arc_state_multilist_index_func
);
6334 arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
6335 multilist_create(sizeof (arc_buf_hdr_t
),
6336 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6337 arc_state_multilist_index_func
);
6338 arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
] =
6339 multilist_create(sizeof (arc_buf_hdr_t
),
6340 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6341 arc_state_multilist_index_func
);
6342 arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
] =
6343 multilist_create(sizeof (arc_buf_hdr_t
),
6344 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6345 arc_state_multilist_index_func
);
6346 arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
] =
6347 multilist_create(sizeof (arc_buf_hdr_t
),
6348 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6349 arc_state_multilist_index_func
);
6351 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6352 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6353 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
6354 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
6355 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6356 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6357 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
6358 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
6359 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6360 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6361 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
6362 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
6364 refcount_create(&arc_anon
->arcs_size
);
6365 refcount_create(&arc_mru
->arcs_size
);
6366 refcount_create(&arc_mru_ghost
->arcs_size
);
6367 refcount_create(&arc_mfu
->arcs_size
);
6368 refcount_create(&arc_mfu_ghost
->arcs_size
);
6369 refcount_create(&arc_l2c_only
->arcs_size
);
6371 arc_anon
->arcs_state
= ARC_STATE_ANON
;
6372 arc_mru
->arcs_state
= ARC_STATE_MRU
;
6373 arc_mru_ghost
->arcs_state
= ARC_STATE_MRU_GHOST
;
6374 arc_mfu
->arcs_state
= ARC_STATE_MFU
;
6375 arc_mfu_ghost
->arcs_state
= ARC_STATE_MFU_GHOST
;
6376 arc_l2c_only
->arcs_state
= ARC_STATE_L2C_ONLY
;
6380 arc_state_fini(void)
6382 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6383 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6384 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
6385 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
6386 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6387 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6388 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
6389 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
6390 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6391 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6392 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
6393 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
6395 refcount_destroy(&arc_anon
->arcs_size
);
6396 refcount_destroy(&arc_mru
->arcs_size
);
6397 refcount_destroy(&arc_mru_ghost
->arcs_size
);
6398 refcount_destroy(&arc_mfu
->arcs_size
);
6399 refcount_destroy(&arc_mfu_ghost
->arcs_size
);
6400 refcount_destroy(&arc_l2c_only
->arcs_size
);
6402 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
6403 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
6404 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
6405 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
6406 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
6407 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
6408 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
6409 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
6410 multilist_destroy(arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]);
6411 multilist_destroy(arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]);
6423 uint64_t percent
, allmem
= arc_all_memory();
6425 mutex_init(&arc_reclaim_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
6426 cv_init(&arc_reclaim_thread_cv
, NULL
, CV_DEFAULT
, NULL
);
6427 cv_init(&arc_reclaim_waiters_cv
, NULL
, CV_DEFAULT
, NULL
);
6429 /* Convert seconds to clock ticks */
6430 arc_min_prefetch_lifespan
= 1 * hz
;
6434 * Register a shrinker to support synchronous (direct) memory
6435 * reclaim from the arc. This is done to prevent kswapd from
6436 * swapping out pages when it is preferable to shrink the arc.
6438 spl_register_shrinker(&arc_shrinker
);
6440 /* Set to 1/64 of all memory or a minimum of 512K */
6441 arc_sys_free
= MAX(allmem
/ 64, (512 * 1024));
6445 /* Set max to 1/2 of all memory */
6446 arc_c_max
= allmem
/ 2;
6449 * In userland, there's only the memory pressure that we artificially
6450 * create (see arc_available_memory()). Don't let arc_c get too
6451 * small, because it can cause transactions to be larger than
6452 * arc_c, causing arc_tempreserve_space() to fail.
6455 arc_c_min
= MAX(arc_c_max
/ 2, 2ULL << SPA_MAXBLOCKSHIFT
);
6457 arc_c_min
= 2ULL << SPA_MAXBLOCKSHIFT
;
6461 arc_p
= (arc_c
>> 1);
6464 /* Set min to 1/2 of arc_c_min */
6465 arc_meta_min
= 1ULL << SPA_MAXBLOCKSHIFT
;
6466 /* Initialize maximum observed usage to zero */
6469 * Set arc_meta_limit to a percent of arc_c_max with a floor of
6470 * arc_meta_min, and a ceiling of arc_c_max.
6472 percent
= MIN(zfs_arc_meta_limit_percent
, 100);
6473 arc_meta_limit
= MAX(arc_meta_min
, (percent
* arc_c_max
) / 100);
6474 percent
= MIN(zfs_arc_dnode_limit_percent
, 100);
6475 arc_dnode_limit
= (percent
* arc_meta_limit
) / 100;
6477 /* Apply user specified tunings */
6478 arc_tuning_update();
6480 /* if kmem_flags are set, lets try to use less memory */
6481 if (kmem_debugging())
6483 if (arc_c
< arc_c_min
)
6489 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
6490 offsetof(arc_prune_t
, p_node
));
6491 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6493 arc_prune_taskq
= taskq_create("arc_prune", max_ncpus
, defclsyspri
,
6494 max_ncpus
, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
6496 arc_reclaim_thread_exit
= B_FALSE
;
6498 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
6499 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
6501 if (arc_ksp
!= NULL
) {
6502 arc_ksp
->ks_data
= &arc_stats
;
6503 arc_ksp
->ks_update
= arc_kstat_update
;
6504 kstat_install(arc_ksp
);
6507 (void) thread_create(NULL
, 0, arc_reclaim_thread
, NULL
, 0, &p0
,
6508 TS_RUN
, defclsyspri
);
6514 * Calculate maximum amount of dirty data per pool.
6516 * If it has been set by a module parameter, take that.
6517 * Otherwise, use a percentage of physical memory defined by
6518 * zfs_dirty_data_max_percent (default 10%) with a cap at
6519 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
6521 if (zfs_dirty_data_max_max
== 0)
6522 zfs_dirty_data_max_max
= MIN(4ULL * 1024 * 1024 * 1024,
6523 allmem
* zfs_dirty_data_max_max_percent
/ 100);
6525 if (zfs_dirty_data_max
== 0) {
6526 zfs_dirty_data_max
= allmem
*
6527 zfs_dirty_data_max_percent
/ 100;
6528 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
6529 zfs_dirty_data_max_max
);
6539 spl_unregister_shrinker(&arc_shrinker
);
6540 #endif /* _KERNEL */
6542 mutex_enter(&arc_reclaim_lock
);
6543 arc_reclaim_thread_exit
= B_TRUE
;
6545 * The reclaim thread will set arc_reclaim_thread_exit back to
6546 * B_FALSE when it is finished exiting; we're waiting for that.
6548 while (arc_reclaim_thread_exit
) {
6549 cv_signal(&arc_reclaim_thread_cv
);
6550 cv_wait(&arc_reclaim_thread_cv
, &arc_reclaim_lock
);
6552 mutex_exit(&arc_reclaim_lock
);
6554 /* Use B_TRUE to ensure *all* buffers are evicted */
6555 arc_flush(NULL
, B_TRUE
);
6559 if (arc_ksp
!= NULL
) {
6560 kstat_delete(arc_ksp
);
6564 taskq_wait(arc_prune_taskq
);
6565 taskq_destroy(arc_prune_taskq
);
6567 mutex_enter(&arc_prune_mtx
);
6568 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
6569 list_remove(&arc_prune_list
, p
);
6570 refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
6571 refcount_destroy(&p
->p_refcnt
);
6572 kmem_free(p
, sizeof (*p
));
6574 mutex_exit(&arc_prune_mtx
);
6576 list_destroy(&arc_prune_list
);
6577 mutex_destroy(&arc_prune_mtx
);
6578 mutex_destroy(&arc_reclaim_lock
);
6579 cv_destroy(&arc_reclaim_thread_cv
);
6580 cv_destroy(&arc_reclaim_waiters_cv
);
6585 ASSERT0(arc_loaned_bytes
);
6591 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6592 * It uses dedicated storage devices to hold cached data, which are populated
6593 * using large infrequent writes. The main role of this cache is to boost
6594 * the performance of random read workloads. The intended L2ARC devices
6595 * include short-stroked disks, solid state disks, and other media with
6596 * substantially faster read latency than disk.
6598 * +-----------------------+
6600 * +-----------------------+
6603 * l2arc_feed_thread() arc_read()
6607 * +---------------+ |
6609 * +---------------+ |
6614 * +-------+ +-------+
6616 * | cache | | cache |
6617 * +-------+ +-------+
6618 * +=========+ .-----.
6619 * : L2ARC : |-_____-|
6620 * : devices : | Disks |
6621 * +=========+ `-_____-'
6623 * Read requests are satisfied from the following sources, in order:
6626 * 2) vdev cache of L2ARC devices
6628 * 4) vdev cache of disks
6631 * Some L2ARC device types exhibit extremely slow write performance.
6632 * To accommodate for this there are some significant differences between
6633 * the L2ARC and traditional cache design:
6635 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6636 * the ARC behave as usual, freeing buffers and placing headers on ghost
6637 * lists. The ARC does not send buffers to the L2ARC during eviction as
6638 * this would add inflated write latencies for all ARC memory pressure.
6640 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6641 * It does this by periodically scanning buffers from the eviction-end of
6642 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6643 * not already there. It scans until a headroom of buffers is satisfied,
6644 * which itself is a buffer for ARC eviction. If a compressible buffer is
6645 * found during scanning and selected for writing to an L2ARC device, we
6646 * temporarily boost scanning headroom during the next scan cycle to make
6647 * sure we adapt to compression effects (which might significantly reduce
6648 * the data volume we write to L2ARC). The thread that does this is
6649 * l2arc_feed_thread(), illustrated below; example sizes are included to
6650 * provide a better sense of ratio than this diagram:
6653 * +---------------------+----------+
6654 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6655 * +---------------------+----------+ | o L2ARC eligible
6656 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6657 * +---------------------+----------+ |
6658 * 15.9 Gbytes ^ 32 Mbytes |
6660 * l2arc_feed_thread()
6662 * l2arc write hand <--[oooo]--'
6666 * +==============================+
6667 * L2ARC dev |####|#|###|###| |####| ... |
6668 * +==============================+
6671 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6672 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6673 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6674 * safe to say that this is an uncommon case, since buffers at the end of
6675 * the ARC lists have moved there due to inactivity.
6677 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6678 * then the L2ARC simply misses copying some buffers. This serves as a
6679 * pressure valve to prevent heavy read workloads from both stalling the ARC
6680 * with waits and clogging the L2ARC with writes. This also helps prevent
6681 * the potential for the L2ARC to churn if it attempts to cache content too
6682 * quickly, such as during backups of the entire pool.
6684 * 5. After system boot and before the ARC has filled main memory, there are
6685 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6686 * lists can remain mostly static. Instead of searching from tail of these
6687 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6688 * for eligible buffers, greatly increasing its chance of finding them.
6690 * The L2ARC device write speed is also boosted during this time so that
6691 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6692 * there are no L2ARC reads, and no fear of degrading read performance
6693 * through increased writes.
6695 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6696 * the vdev queue can aggregate them into larger and fewer writes. Each
6697 * device is written to in a rotor fashion, sweeping writes through
6698 * available space then repeating.
6700 * 7. The L2ARC does not store dirty content. It never needs to flush
6701 * write buffers back to disk based storage.
6703 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6704 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6706 * The performance of the L2ARC can be tweaked by a number of tunables, which
6707 * may be necessary for different workloads:
6709 * l2arc_write_max max write bytes per interval
6710 * l2arc_write_boost extra write bytes during device warmup
6711 * l2arc_noprefetch skip caching prefetched buffers
6712 * l2arc_headroom number of max device writes to precache
6713 * l2arc_headroom_boost when we find compressed buffers during ARC
6714 * scanning, we multiply headroom by this
6715 * percentage factor for the next scan cycle,
6716 * since more compressed buffers are likely to
6718 * l2arc_feed_secs seconds between L2ARC writing
6720 * Tunables may be removed or added as future performance improvements are
6721 * integrated, and also may become zpool properties.
6723 * There are three key functions that control how the L2ARC warms up:
6725 * l2arc_write_eligible() check if a buffer is eligible to cache
6726 * l2arc_write_size() calculate how much to write
6727 * l2arc_write_interval() calculate sleep delay between writes
6729 * These three functions determine what to write, how much, and how quickly
6734 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*hdr
)
6737 * A buffer is *not* eligible for the L2ARC if it:
6738 * 1. belongs to a different spa.
6739 * 2. is already cached on the L2ARC.
6740 * 3. has an I/O in progress (it may be an incomplete read).
6741 * 4. is flagged not eligible (zfs property).
6743 if (hdr
->b_spa
!= spa_guid
|| HDR_HAS_L2HDR(hdr
) ||
6744 HDR_IO_IN_PROGRESS(hdr
) || !HDR_L2CACHE(hdr
))
6751 l2arc_write_size(void)
6756 * Make sure our globals have meaningful values in case the user
6759 size
= l2arc_write_max
;
6761 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
6762 "be greater than zero, resetting it to the default (%d)",
6764 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
6767 if (arc_warm
== B_FALSE
)
6768 size
+= l2arc_write_boost
;
6775 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
6777 clock_t interval
, next
, now
;
6780 * If the ARC lists are busy, increase our write rate; if the
6781 * lists are stale, idle back. This is achieved by checking
6782 * how much we previously wrote - if it was more than half of
6783 * what we wanted, schedule the next write much sooner.
6785 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
6786 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
6788 interval
= hz
* l2arc_feed_secs
;
6790 now
= ddi_get_lbolt();
6791 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
6797 * Cycle through L2ARC devices. This is how L2ARC load balances.
6798 * If a device is returned, this also returns holding the spa config lock.
6800 static l2arc_dev_t
*
6801 l2arc_dev_get_next(void)
6803 l2arc_dev_t
*first
, *next
= NULL
;
6806 * Lock out the removal of spas (spa_namespace_lock), then removal
6807 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6808 * both locks will be dropped and a spa config lock held instead.
6810 mutex_enter(&spa_namespace_lock
);
6811 mutex_enter(&l2arc_dev_mtx
);
6813 /* if there are no vdevs, there is nothing to do */
6814 if (l2arc_ndev
== 0)
6818 next
= l2arc_dev_last
;
6820 /* loop around the list looking for a non-faulted vdev */
6822 next
= list_head(l2arc_dev_list
);
6824 next
= list_next(l2arc_dev_list
, next
);
6826 next
= list_head(l2arc_dev_list
);
6829 /* if we have come back to the start, bail out */
6832 else if (next
== first
)
6835 } while (vdev_is_dead(next
->l2ad_vdev
));
6837 /* if we were unable to find any usable vdevs, return NULL */
6838 if (vdev_is_dead(next
->l2ad_vdev
))
6841 l2arc_dev_last
= next
;
6844 mutex_exit(&l2arc_dev_mtx
);
6847 * Grab the config lock to prevent the 'next' device from being
6848 * removed while we are writing to it.
6851 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
6852 mutex_exit(&spa_namespace_lock
);
6858 * Free buffers that were tagged for destruction.
6861 l2arc_do_free_on_write(void)
6864 l2arc_data_free_t
*df
, *df_prev
;
6866 mutex_enter(&l2arc_free_on_write_mtx
);
6867 buflist
= l2arc_free_on_write
;
6869 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
6870 df_prev
= list_prev(buflist
, df
);
6871 ASSERT3P(df
->l2df_abd
, !=, NULL
);
6872 abd_free(df
->l2df_abd
);
6873 list_remove(buflist
, df
);
6874 kmem_free(df
, sizeof (l2arc_data_free_t
));
6877 mutex_exit(&l2arc_free_on_write_mtx
);
6881 * A write to a cache device has completed. Update all headers to allow
6882 * reads from these buffers to begin.
6885 l2arc_write_done(zio_t
*zio
)
6887 l2arc_write_callback_t
*cb
;
6890 arc_buf_hdr_t
*head
, *hdr
, *hdr_prev
;
6891 kmutex_t
*hash_lock
;
6892 int64_t bytes_dropped
= 0;
6894 cb
= zio
->io_private
;
6895 ASSERT3P(cb
, !=, NULL
);
6896 dev
= cb
->l2wcb_dev
;
6897 ASSERT3P(dev
, !=, NULL
);
6898 head
= cb
->l2wcb_head
;
6899 ASSERT3P(head
, !=, NULL
);
6900 buflist
= &dev
->l2ad_buflist
;
6901 ASSERT3P(buflist
, !=, NULL
);
6902 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
6903 l2arc_write_callback_t
*, cb
);
6905 if (zio
->io_error
!= 0)
6906 ARCSTAT_BUMP(arcstat_l2_writes_error
);
6909 * All writes completed, or an error was hit.
6912 mutex_enter(&dev
->l2ad_mtx
);
6913 for (hdr
= list_prev(buflist
, head
); hdr
; hdr
= hdr_prev
) {
6914 hdr_prev
= list_prev(buflist
, hdr
);
6916 hash_lock
= HDR_LOCK(hdr
);
6919 * We cannot use mutex_enter or else we can deadlock
6920 * with l2arc_write_buffers (due to swapping the order
6921 * the hash lock and l2ad_mtx are taken).
6923 if (!mutex_tryenter(hash_lock
)) {
6925 * Missed the hash lock. We must retry so we
6926 * don't leave the ARC_FLAG_L2_WRITING bit set.
6928 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry
);
6931 * We don't want to rescan the headers we've
6932 * already marked as having been written out, so
6933 * we reinsert the head node so we can pick up
6934 * where we left off.
6936 list_remove(buflist
, head
);
6937 list_insert_after(buflist
, hdr
, head
);
6939 mutex_exit(&dev
->l2ad_mtx
);
6942 * We wait for the hash lock to become available
6943 * to try and prevent busy waiting, and increase
6944 * the chance we'll be able to acquire the lock
6945 * the next time around.
6947 mutex_enter(hash_lock
);
6948 mutex_exit(hash_lock
);
6953 * We could not have been moved into the arc_l2c_only
6954 * state while in-flight due to our ARC_FLAG_L2_WRITING
6955 * bit being set. Let's just ensure that's being enforced.
6957 ASSERT(HDR_HAS_L1HDR(hdr
));
6960 * Skipped - drop L2ARC entry and mark the header as no
6961 * longer L2 eligibile.
6963 if (zio
->io_error
!= 0) {
6965 * Error - drop L2ARC entry.
6967 list_remove(buflist
, hdr
);
6968 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
6970 ARCSTAT_INCR(arcstat_l2_asize
, -arc_hdr_size(hdr
));
6971 ARCSTAT_INCR(arcstat_l2_size
, -HDR_GET_LSIZE(hdr
));
6973 bytes_dropped
+= arc_hdr_size(hdr
);
6974 (void) refcount_remove_many(&dev
->l2ad_alloc
,
6975 arc_hdr_size(hdr
), hdr
);
6979 * Allow ARC to begin reads and ghost list evictions to
6982 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_WRITING
);
6984 mutex_exit(hash_lock
);
6987 atomic_inc_64(&l2arc_writes_done
);
6988 list_remove(buflist
, head
);
6989 ASSERT(!HDR_HAS_L1HDR(head
));
6990 kmem_cache_free(hdr_l2only_cache
, head
);
6991 mutex_exit(&dev
->l2ad_mtx
);
6993 vdev_space_update(dev
->l2ad_vdev
, -bytes_dropped
, 0, 0);
6995 l2arc_do_free_on_write();
6997 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
7001 * A read to a cache device completed. Validate buffer contents before
7002 * handing over to the regular ARC routines.
7005 l2arc_read_done(zio_t
*zio
)
7007 l2arc_read_callback_t
*cb
;
7009 kmutex_t
*hash_lock
;
7010 boolean_t valid_cksum
;
7012 ASSERT3P(zio
->io_vd
, !=, NULL
);
7013 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
7015 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
7017 cb
= zio
->io_private
;
7018 ASSERT3P(cb
, !=, NULL
);
7019 hdr
= cb
->l2rcb_hdr
;
7020 ASSERT3P(hdr
, !=, NULL
);
7022 hash_lock
= HDR_LOCK(hdr
);
7023 mutex_enter(hash_lock
);
7024 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
7026 ASSERT3P(zio
->io_abd
, !=, NULL
);
7029 * Check this survived the L2ARC journey.
7031 ASSERT3P(zio
->io_abd
, ==, hdr
->b_l1hdr
.b_pabd
);
7032 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
7033 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
7035 valid_cksum
= arc_cksum_is_equal(hdr
, zio
);
7036 if (valid_cksum
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
7037 mutex_exit(hash_lock
);
7038 zio
->io_private
= hdr
;
7041 mutex_exit(hash_lock
);
7043 * Buffer didn't survive caching. Increment stats and
7044 * reissue to the original storage device.
7046 if (zio
->io_error
!= 0) {
7047 ARCSTAT_BUMP(arcstat_l2_io_error
);
7049 zio
->io_error
= SET_ERROR(EIO
);
7052 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
7055 * If there's no waiter, issue an async i/o to the primary
7056 * storage now. If there *is* a waiter, the caller must
7057 * issue the i/o in a context where it's OK to block.
7059 if (zio
->io_waiter
== NULL
) {
7060 zio_t
*pio
= zio_unique_parent(zio
);
7062 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
7064 zio_nowait(zio_read(pio
, zio
->io_spa
, zio
->io_bp
,
7065 hdr
->b_l1hdr
.b_pabd
, zio
->io_size
, arc_read_done
,
7066 hdr
, zio
->io_priority
, cb
->l2rcb_flags
,
7071 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
7075 * This is the list priority from which the L2ARC will search for pages to
7076 * cache. This is used within loops (0..3) to cycle through lists in the
7077 * desired order. This order can have a significant effect on cache
7080 * Currently the metadata lists are hit first, MFU then MRU, followed by
7081 * the data lists. This function returns a locked list, and also returns
7084 static multilist_sublist_t
*
7085 l2arc_sublist_lock(int list_num
)
7087 multilist_t
*ml
= NULL
;
7090 ASSERT(list_num
>= 0 && list_num
< L2ARC_FEED_TYPES
);
7094 ml
= arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
7097 ml
= arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
7100 ml
= arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
7103 ml
= arc_mru
->arcs_list
[ARC_BUFC_DATA
];
7110 * Return a randomly-selected sublist. This is acceptable
7111 * because the caller feeds only a little bit of data for each
7112 * call (8MB). Subsequent calls will result in different
7113 * sublists being selected.
7115 idx
= multilist_get_random_index(ml
);
7116 return (multilist_sublist_lock(ml
, idx
));
7120 * Evict buffers from the device write hand to the distance specified in
7121 * bytes. This distance may span populated buffers, it may span nothing.
7122 * This is clearing a region on the L2ARC device ready for writing.
7123 * If the 'all' boolean is set, every buffer is evicted.
7126 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
7129 arc_buf_hdr_t
*hdr
, *hdr_prev
;
7130 kmutex_t
*hash_lock
;
7133 buflist
= &dev
->l2ad_buflist
;
7135 if (!all
&& dev
->l2ad_first
) {
7137 * This is the first sweep through the device. There is
7143 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
7145 * When nearing the end of the device, evict to the end
7146 * before the device write hand jumps to the start.
7148 taddr
= dev
->l2ad_end
;
7150 taddr
= dev
->l2ad_hand
+ distance
;
7152 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
7153 uint64_t, taddr
, boolean_t
, all
);
7156 mutex_enter(&dev
->l2ad_mtx
);
7157 for (hdr
= list_tail(buflist
); hdr
; hdr
= hdr_prev
) {
7158 hdr_prev
= list_prev(buflist
, hdr
);
7160 hash_lock
= HDR_LOCK(hdr
);
7163 * We cannot use mutex_enter or else we can deadlock
7164 * with l2arc_write_buffers (due to swapping the order
7165 * the hash lock and l2ad_mtx are taken).
7167 if (!mutex_tryenter(hash_lock
)) {
7169 * Missed the hash lock. Retry.
7171 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
7172 mutex_exit(&dev
->l2ad_mtx
);
7173 mutex_enter(hash_lock
);
7174 mutex_exit(hash_lock
);
7178 if (HDR_L2_WRITE_HEAD(hdr
)) {
7180 * We hit a write head node. Leave it for
7181 * l2arc_write_done().
7183 list_remove(buflist
, hdr
);
7184 mutex_exit(hash_lock
);
7188 if (!all
&& HDR_HAS_L2HDR(hdr
) &&
7189 (hdr
->b_l2hdr
.b_daddr
> taddr
||
7190 hdr
->b_l2hdr
.b_daddr
< dev
->l2ad_hand
)) {
7192 * We've evicted to the target address,
7193 * or the end of the device.
7195 mutex_exit(hash_lock
);
7199 ASSERT(HDR_HAS_L2HDR(hdr
));
7200 if (!HDR_HAS_L1HDR(hdr
)) {
7201 ASSERT(!HDR_L2_READING(hdr
));
7203 * This doesn't exist in the ARC. Destroy.
7204 * arc_hdr_destroy() will call list_remove()
7205 * and decrement arcstat_l2_size.
7207 arc_change_state(arc_anon
, hdr
, hash_lock
);
7208 arc_hdr_destroy(hdr
);
7210 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_l2c_only
);
7211 ARCSTAT_BUMP(arcstat_l2_evict_l1cached
);
7213 * Invalidate issued or about to be issued
7214 * reads, since we may be about to write
7215 * over this location.
7217 if (HDR_L2_READING(hdr
)) {
7218 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
7219 arc_hdr_set_flags(hdr
, ARC_FLAG_L2_EVICTED
);
7222 /* Ensure this header has finished being written */
7223 ASSERT(!HDR_L2_WRITING(hdr
));
7225 arc_hdr_l2hdr_destroy(hdr
);
7227 mutex_exit(hash_lock
);
7229 mutex_exit(&dev
->l2ad_mtx
);
7233 * Find and write ARC buffers to the L2ARC device.
7235 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7236 * for reading until they have completed writing.
7237 * The headroom_boost is an in-out parameter used to maintain headroom boost
7238 * state between calls to this function.
7240 * Returns the number of bytes actually written (which may be smaller than
7241 * the delta by which the device hand has changed due to alignment).
7244 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
)
7246 arc_buf_hdr_t
*hdr
, *hdr_prev
, *head
;
7247 uint64_t write_asize
, write_psize
, write_sz
, headroom
;
7249 l2arc_write_callback_t
*cb
;
7251 uint64_t guid
= spa_load_guid(spa
);
7254 ASSERT3P(dev
->l2ad_vdev
, !=, NULL
);
7257 write_sz
= write_asize
= write_psize
= 0;
7259 head
= kmem_cache_alloc(hdr_l2only_cache
, KM_PUSHPAGE
);
7260 arc_hdr_set_flags(head
, ARC_FLAG_L2_WRITE_HEAD
| ARC_FLAG_HAS_L2HDR
);
7263 * Copy buffers for L2ARC writing.
7265 for (try = 0; try < L2ARC_FEED_TYPES
; try++) {
7266 multilist_sublist_t
*mls
= l2arc_sublist_lock(try);
7267 uint64_t passed_sz
= 0;
7269 VERIFY3P(mls
, !=, NULL
);
7272 * L2ARC fast warmup.
7274 * Until the ARC is warm and starts to evict, read from the
7275 * head of the ARC lists rather than the tail.
7277 if (arc_warm
== B_FALSE
)
7278 hdr
= multilist_sublist_head(mls
);
7280 hdr
= multilist_sublist_tail(mls
);
7282 headroom
= target_sz
* l2arc_headroom
;
7283 if (zfs_compressed_arc_enabled
)
7284 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
7286 for (; hdr
; hdr
= hdr_prev
) {
7287 kmutex_t
*hash_lock
;
7288 uint64_t asize
, size
;
7291 if (arc_warm
== B_FALSE
)
7292 hdr_prev
= multilist_sublist_next(mls
, hdr
);
7294 hdr_prev
= multilist_sublist_prev(mls
, hdr
);
7296 hash_lock
= HDR_LOCK(hdr
);
7297 if (!mutex_tryenter(hash_lock
)) {
7299 * Skip this buffer rather than waiting.
7304 passed_sz
+= HDR_GET_LSIZE(hdr
);
7305 if (passed_sz
> headroom
) {
7309 mutex_exit(hash_lock
);
7313 if (!l2arc_write_eligible(guid
, hdr
)) {
7314 mutex_exit(hash_lock
);
7318 if ((write_asize
+ HDR_GET_LSIZE(hdr
)) > target_sz
) {
7320 mutex_exit(hash_lock
);
7326 * Insert a dummy header on the buflist so
7327 * l2arc_write_done() can find where the
7328 * write buffers begin without searching.
7330 mutex_enter(&dev
->l2ad_mtx
);
7331 list_insert_head(&dev
->l2ad_buflist
, head
);
7332 mutex_exit(&dev
->l2ad_mtx
);
7335 sizeof (l2arc_write_callback_t
), KM_SLEEP
);
7336 cb
->l2wcb_dev
= dev
;
7337 cb
->l2wcb_head
= head
;
7338 pio
= zio_root(spa
, l2arc_write_done
, cb
,
7342 hdr
->b_l2hdr
.b_dev
= dev
;
7343 hdr
->b_l2hdr
.b_hits
= 0;
7345 hdr
->b_l2hdr
.b_daddr
= dev
->l2ad_hand
;
7346 arc_hdr_set_flags(hdr
,
7347 ARC_FLAG_L2_WRITING
| ARC_FLAG_HAS_L2HDR
);
7349 mutex_enter(&dev
->l2ad_mtx
);
7350 list_insert_head(&dev
->l2ad_buflist
, hdr
);
7351 mutex_exit(&dev
->l2ad_mtx
);
7354 * We rely on the L1 portion of the header below, so
7355 * it's invalid for this header to have been evicted out
7356 * of the ghost cache, prior to being written out. The
7357 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7359 ASSERT(HDR_HAS_L1HDR(hdr
));
7361 ASSERT3U(HDR_GET_PSIZE(hdr
), >, 0);
7362 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
7363 ASSERT3U(arc_hdr_size(hdr
), >, 0);
7364 size
= arc_hdr_size(hdr
);
7366 (void) refcount_add_many(&dev
->l2ad_alloc
, size
, hdr
);
7369 * Normally the L2ARC can use the hdr's data, but if
7370 * we're sharing data between the hdr and one of its
7371 * bufs, L2ARC needs its own copy of the data so that
7372 * the ZIO below can't race with the buf consumer. To
7373 * ensure that this copy will be available for the
7374 * lifetime of the ZIO and be cleaned up afterwards, we
7375 * add it to the l2arc_free_on_write queue.
7377 if (!HDR_SHARED_DATA(hdr
)) {
7378 to_write
= hdr
->b_l1hdr
.b_pabd
;
7380 to_write
= abd_alloc_for_io(size
,
7381 HDR_ISTYPE_METADATA(hdr
));
7382 abd_copy(to_write
, hdr
->b_l1hdr
.b_pabd
, size
);
7383 l2arc_free_abd_on_write(to_write
, size
,
7386 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
7387 hdr
->b_l2hdr
.b_daddr
, size
, to_write
,
7388 ZIO_CHECKSUM_OFF
, NULL
, hdr
,
7389 ZIO_PRIORITY_ASYNC_WRITE
,
7390 ZIO_FLAG_CANFAIL
, B_FALSE
);
7392 write_sz
+= HDR_GET_LSIZE(hdr
);
7393 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
7396 write_asize
+= size
;
7398 * Keep the clock hand suitably device-aligned.
7400 asize
= vdev_psize_to_asize(dev
->l2ad_vdev
, size
);
7401 write_psize
+= asize
;
7402 dev
->l2ad_hand
+= asize
;
7404 mutex_exit(hash_lock
);
7406 (void) zio_nowait(wzio
);
7409 multilist_sublist_unlock(mls
);
7415 /* No buffers selected for writing? */
7418 ASSERT(!HDR_HAS_L1HDR(head
));
7419 kmem_cache_free(hdr_l2only_cache
, head
);
7423 ASSERT3U(write_asize
, <=, target_sz
);
7424 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
7425 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_asize
);
7426 ARCSTAT_INCR(arcstat_l2_size
, write_sz
);
7427 ARCSTAT_INCR(arcstat_l2_asize
, write_asize
);
7428 vdev_space_update(dev
->l2ad_vdev
, write_asize
, 0, 0);
7431 * Bump device hand to the device start if it is approaching the end.
7432 * l2arc_evict() will already have evicted ahead for this case.
7434 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
7435 dev
->l2ad_hand
= dev
->l2ad_start
;
7436 dev
->l2ad_first
= B_FALSE
;
7439 dev
->l2ad_writing
= B_TRUE
;
7440 (void) zio_wait(pio
);
7441 dev
->l2ad_writing
= B_FALSE
;
7443 return (write_asize
);
7447 * This thread feeds the L2ARC at regular intervals. This is the beating
7448 * heart of the L2ARC.
7451 l2arc_feed_thread(void)
7456 uint64_t size
, wrote
;
7457 clock_t begin
, next
= ddi_get_lbolt();
7458 fstrans_cookie_t cookie
;
7460 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
7462 mutex_enter(&l2arc_feed_thr_lock
);
7464 cookie
= spl_fstrans_mark();
7465 while (l2arc_thread_exit
== 0) {
7466 CALLB_CPR_SAFE_BEGIN(&cpr
);
7467 (void) cv_timedwait_sig(&l2arc_feed_thr_cv
,
7468 &l2arc_feed_thr_lock
, next
);
7469 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
7470 next
= ddi_get_lbolt() + hz
;
7473 * Quick check for L2ARC devices.
7475 mutex_enter(&l2arc_dev_mtx
);
7476 if (l2arc_ndev
== 0) {
7477 mutex_exit(&l2arc_dev_mtx
);
7480 mutex_exit(&l2arc_dev_mtx
);
7481 begin
= ddi_get_lbolt();
7484 * This selects the next l2arc device to write to, and in
7485 * doing so the next spa to feed from: dev->l2ad_spa. This
7486 * will return NULL if there are now no l2arc devices or if
7487 * they are all faulted.
7489 * If a device is returned, its spa's config lock is also
7490 * held to prevent device removal. l2arc_dev_get_next()
7491 * will grab and release l2arc_dev_mtx.
7493 if ((dev
= l2arc_dev_get_next()) == NULL
)
7496 spa
= dev
->l2ad_spa
;
7497 ASSERT3P(spa
, !=, NULL
);
7500 * If the pool is read-only then force the feed thread to
7501 * sleep a little longer.
7503 if (!spa_writeable(spa
)) {
7504 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
7505 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7510 * Avoid contributing to memory pressure.
7512 if (arc_reclaim_needed()) {
7513 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
7514 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7518 ARCSTAT_BUMP(arcstat_l2_feeds
);
7520 size
= l2arc_write_size();
7523 * Evict L2ARC buffers that will be overwritten.
7525 l2arc_evict(dev
, size
, B_FALSE
);
7528 * Write ARC buffers.
7530 wrote
= l2arc_write_buffers(spa
, dev
, size
);
7533 * Calculate interval between writes.
7535 next
= l2arc_write_interval(begin
, size
, wrote
);
7536 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7538 spl_fstrans_unmark(cookie
);
7540 l2arc_thread_exit
= 0;
7541 cv_broadcast(&l2arc_feed_thr_cv
);
7542 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
7547 l2arc_vdev_present(vdev_t
*vd
)
7551 mutex_enter(&l2arc_dev_mtx
);
7552 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
7553 dev
= list_next(l2arc_dev_list
, dev
)) {
7554 if (dev
->l2ad_vdev
== vd
)
7557 mutex_exit(&l2arc_dev_mtx
);
7559 return (dev
!= NULL
);
7563 * Add a vdev for use by the L2ARC. By this point the spa has already
7564 * validated the vdev and opened it.
7567 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
7569 l2arc_dev_t
*adddev
;
7571 ASSERT(!l2arc_vdev_present(vd
));
7574 * Create a new l2arc device entry.
7576 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
7577 adddev
->l2ad_spa
= spa
;
7578 adddev
->l2ad_vdev
= vd
;
7579 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
7580 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
7581 adddev
->l2ad_hand
= adddev
->l2ad_start
;
7582 adddev
->l2ad_first
= B_TRUE
;
7583 adddev
->l2ad_writing
= B_FALSE
;
7584 list_link_init(&adddev
->l2ad_node
);
7586 mutex_init(&adddev
->l2ad_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7588 * This is a list of all ARC buffers that are still valid on the
7591 list_create(&adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
7592 offsetof(arc_buf_hdr_t
, b_l2hdr
.b_l2node
));
7594 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
7595 refcount_create(&adddev
->l2ad_alloc
);
7598 * Add device to global list
7600 mutex_enter(&l2arc_dev_mtx
);
7601 list_insert_head(l2arc_dev_list
, adddev
);
7602 atomic_inc_64(&l2arc_ndev
);
7603 mutex_exit(&l2arc_dev_mtx
);
7607 * Remove a vdev from the L2ARC.
7610 l2arc_remove_vdev(vdev_t
*vd
)
7612 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
7615 * Find the device by vdev
7617 mutex_enter(&l2arc_dev_mtx
);
7618 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
7619 nextdev
= list_next(l2arc_dev_list
, dev
);
7620 if (vd
== dev
->l2ad_vdev
) {
7625 ASSERT3P(remdev
, !=, NULL
);
7628 * Remove device from global list
7630 list_remove(l2arc_dev_list
, remdev
);
7631 l2arc_dev_last
= NULL
; /* may have been invalidated */
7632 atomic_dec_64(&l2arc_ndev
);
7633 mutex_exit(&l2arc_dev_mtx
);
7636 * Clear all buflists and ARC references. L2ARC device flush.
7638 l2arc_evict(remdev
, 0, B_TRUE
);
7639 list_destroy(&remdev
->l2ad_buflist
);
7640 mutex_destroy(&remdev
->l2ad_mtx
);
7641 refcount_destroy(&remdev
->l2ad_alloc
);
7642 kmem_free(remdev
, sizeof (l2arc_dev_t
));
7648 l2arc_thread_exit
= 0;
7650 l2arc_writes_sent
= 0;
7651 l2arc_writes_done
= 0;
7653 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
7654 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
7655 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7656 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7658 l2arc_dev_list
= &L2ARC_dev_list
;
7659 l2arc_free_on_write
= &L2ARC_free_on_write
;
7660 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
7661 offsetof(l2arc_dev_t
, l2ad_node
));
7662 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
7663 offsetof(l2arc_data_free_t
, l2df_list_node
));
7670 * This is called from dmu_fini(), which is called from spa_fini();
7671 * Because of this, we can assume that all l2arc devices have
7672 * already been removed when the pools themselves were removed.
7675 l2arc_do_free_on_write();
7677 mutex_destroy(&l2arc_feed_thr_lock
);
7678 cv_destroy(&l2arc_feed_thr_cv
);
7679 mutex_destroy(&l2arc_dev_mtx
);
7680 mutex_destroy(&l2arc_free_on_write_mtx
);
7682 list_destroy(l2arc_dev_list
);
7683 list_destroy(l2arc_free_on_write
);
7689 if (!(spa_mode_global
& FWRITE
))
7692 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
7693 TS_RUN
, defclsyspri
);
7699 if (!(spa_mode_global
& FWRITE
))
7702 mutex_enter(&l2arc_feed_thr_lock
);
7703 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
7704 l2arc_thread_exit
= 1;
7705 while (l2arc_thread_exit
!= 0)
7706 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
7707 mutex_exit(&l2arc_feed_thr_lock
);
7710 #if defined(_KERNEL) && defined(HAVE_SPL)
7711 EXPORT_SYMBOL(arc_buf_size
);
7712 EXPORT_SYMBOL(arc_write
);
7713 EXPORT_SYMBOL(arc_read
);
7714 EXPORT_SYMBOL(arc_buf_info
);
7715 EXPORT_SYMBOL(arc_getbuf_func
);
7716 EXPORT_SYMBOL(arc_add_prune_callback
);
7717 EXPORT_SYMBOL(arc_remove_prune_callback
);
7720 module_param(zfs_arc_min
, ulong
, 0644);
7721 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
7723 module_param(zfs_arc_max
, ulong
, 0644);
7724 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
7726 module_param(zfs_arc_meta_limit
, ulong
, 0644);
7727 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
7729 module_param(zfs_arc_meta_limit_percent
, ulong
, 0644);
7730 MODULE_PARM_DESC(zfs_arc_meta_limit_percent
,
7731 "Percent of arc size for arc meta limit");
7733 module_param(zfs_arc_meta_min
, ulong
, 0644);
7734 MODULE_PARM_DESC(zfs_arc_meta_min
, "Min arc metadata");
7736 module_param(zfs_arc_meta_prune
, int, 0644);
7737 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Meta objects to scan for prune");
7739 module_param(zfs_arc_meta_adjust_restarts
, int, 0644);
7740 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts
,
7741 "Limit number of restarts in arc_adjust_meta");
7743 module_param(zfs_arc_meta_strategy
, int, 0644);
7744 MODULE_PARM_DESC(zfs_arc_meta_strategy
, "Meta reclaim strategy");
7746 module_param(zfs_arc_grow_retry
, int, 0644);
7747 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
7749 module_param(zfs_arc_p_aggressive_disable
, int, 0644);
7750 MODULE_PARM_DESC(zfs_arc_p_aggressive_disable
, "disable aggressive arc_p grow");
7752 module_param(zfs_arc_p_dampener_disable
, int, 0644);
7753 MODULE_PARM_DESC(zfs_arc_p_dampener_disable
, "disable arc_p adapt dampener");
7755 module_param(zfs_arc_shrink_shift
, int, 0644);
7756 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
7758 module_param(zfs_arc_p_min_shift
, int, 0644);
7759 MODULE_PARM_DESC(zfs_arc_p_min_shift
, "arc_c shift to calc min/max arc_p");
7761 module_param(zfs_arc_average_blocksize
, int, 0444);
7762 MODULE_PARM_DESC(zfs_arc_average_blocksize
, "Target average block size");
7764 module_param(zfs_compressed_arc_enabled
, int, 0644);
7765 MODULE_PARM_DESC(zfs_compressed_arc_enabled
, "Disable compressed arc buffers");
7767 module_param(zfs_arc_min_prefetch_lifespan
, int, 0644);
7768 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan
, "Min life of prefetch block");
7770 module_param(l2arc_write_max
, ulong
, 0644);
7771 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
7773 module_param(l2arc_write_boost
, ulong
, 0644);
7774 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
7776 module_param(l2arc_headroom
, ulong
, 0644);
7777 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
7779 module_param(l2arc_headroom_boost
, ulong
, 0644);
7780 MODULE_PARM_DESC(l2arc_headroom_boost
, "Compressed l2arc_headroom multiplier");
7782 module_param(l2arc_feed_secs
, ulong
, 0644);
7783 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
7785 module_param(l2arc_feed_min_ms
, ulong
, 0644);
7786 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
7788 module_param(l2arc_noprefetch
, int, 0644);
7789 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
7791 module_param(l2arc_feed_again
, int, 0644);
7792 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
7794 module_param(l2arc_norw
, int, 0644);
7795 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");
7797 module_param(zfs_arc_lotsfree_percent
, int, 0644);
7798 MODULE_PARM_DESC(zfs_arc_lotsfree_percent
,
7799 "System free memory I/O throttle in bytes");
7801 module_param(zfs_arc_sys_free
, ulong
, 0644);
7802 MODULE_PARM_DESC(zfs_arc_sys_free
, "System free memory target size in bytes");
7804 module_param(zfs_arc_dnode_limit
, ulong
, 0644);
7805 MODULE_PARM_DESC(zfs_arc_dnode_limit
, "Minimum bytes of dnodes in arc");
7807 module_param(zfs_arc_dnode_limit_percent
, ulong
, 0644);
7808 MODULE_PARM_DESC(zfs_arc_dnode_limit_percent
,
7809 "Percent of ARC meta buffers for dnodes");
7811 module_param(zfs_arc_dnode_reduce_percent
, ulong
, 0644);
7812 MODULE_PARM_DESC(zfs_arc_dnode_reduce_percent
,
7813 "Percentage of excess dnodes to try to unpin");