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.
262 * The L1ARC has a slightly different system for storing encrypted data.
263 * Raw (encrypted + possibly compressed) data has a few subtle differences from
264 * data that is just compressed. The biggest difference is that it is not
265 * possible to decrypt encrypted data (or visa versa) if the keys aren't loaded.
266 * The other difference is that encryption cannot be treated as a suggestion.
267 * If a caller would prefer compressed data, but they actually wind up with
268 * uncompressed data the worst thing that could happen is there might be a
269 * performance hit. If the caller requests encrypted data, however, we must be
270 * sure they actually get it or else secret information could be leaked. Raw
271 * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore,
272 * may have both an encrypted version and a decrypted version of its data at
273 * once. When a caller needs a raw arc_buf_t, it is allocated and the data is
274 * copied out of this header. To avoid complications with b_pabd, raw buffers
280 #include <sys/spa_impl.h>
281 #include <sys/zio_compress.h>
282 #include <sys/zio_checksum.h>
283 #include <sys/zfs_context.h>
285 #include <sys/refcount.h>
286 #include <sys/vdev.h>
287 #include <sys/vdev_impl.h>
288 #include <sys/dsl_pool.h>
289 #include <sys/zio_checksum.h>
290 #include <sys/multilist.h>
293 #include <sys/fm/fs/zfs.h>
295 #include <sys/vmsystm.h>
297 #include <sys/fs/swapnode.h>
299 #include <linux/mm_compat.h>
301 #include <sys/callb.h>
302 #include <sys/kstat.h>
303 #include <sys/dmu_tx.h>
304 #include <zfs_fletcher.h>
305 #include <sys/arc_impl.h>
306 #include <sys/trace_arc.h>
309 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
310 boolean_t arc_watch
= B_FALSE
;
313 static kmutex_t arc_reclaim_lock
;
314 static kcondvar_t arc_reclaim_thread_cv
;
315 static boolean_t arc_reclaim_thread_exit
;
316 static kcondvar_t arc_reclaim_waiters_cv
;
319 * The number of headers to evict in arc_evict_state_impl() before
320 * dropping the sublist lock and evicting from another sublist. A lower
321 * value means we're more likely to evict the "correct" header (i.e. the
322 * oldest header in the arc state), but comes with higher overhead
323 * (i.e. more invocations of arc_evict_state_impl()).
325 int zfs_arc_evict_batch_limit
= 10;
327 /* number of seconds before growing cache again */
328 static int arc_grow_retry
= 5;
330 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
331 int zfs_arc_overflow_shift
= 8;
333 /* shift of arc_c for calculating both min and max arc_p */
334 static int arc_p_min_shift
= 4;
336 /* log2(fraction of arc to reclaim) */
337 static int arc_shrink_shift
= 7;
339 /* percent of pagecache to reclaim arc to */
341 static uint_t zfs_arc_pc_percent
= 0;
345 * log2(fraction of ARC which must be free to allow growing).
346 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
347 * when reading a new block into the ARC, we will evict an equal-sized block
350 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
351 * we will still not allow it to grow.
353 int arc_no_grow_shift
= 5;
357 * minimum lifespan of a prefetch block in clock ticks
358 * (initialized in arc_init())
360 static int arc_min_prefetch_ms
;
361 static int arc_min_prescient_prefetch_ms
;
364 * If this percent of memory is free, don't throttle.
366 int arc_lotsfree_percent
= 10;
371 * The arc has filled available memory and has now warmed up.
373 static boolean_t arc_warm
;
376 * log2 fraction of the zio arena to keep free.
378 int arc_zio_arena_free_shift
= 2;
381 * These tunables are for performance analysis.
383 unsigned long zfs_arc_max
= 0;
384 unsigned long zfs_arc_min
= 0;
385 unsigned long zfs_arc_meta_limit
= 0;
386 unsigned long zfs_arc_meta_min
= 0;
387 unsigned long zfs_arc_dnode_limit
= 0;
388 unsigned long zfs_arc_dnode_reduce_percent
= 10;
389 int zfs_arc_grow_retry
= 0;
390 int zfs_arc_shrink_shift
= 0;
391 int zfs_arc_p_min_shift
= 0;
392 int zfs_arc_average_blocksize
= 8 * 1024; /* 8KB */
394 int zfs_compressed_arc_enabled
= B_TRUE
;
397 * ARC will evict meta buffers that exceed arc_meta_limit. This
398 * tunable make arc_meta_limit adjustable for different workloads.
400 unsigned long zfs_arc_meta_limit_percent
= 75;
403 * Percentage that can be consumed by dnodes of ARC meta buffers.
405 unsigned long zfs_arc_dnode_limit_percent
= 10;
408 * These tunables are Linux specific
410 unsigned long zfs_arc_sys_free
= 0;
411 int zfs_arc_min_prefetch_ms
= 0;
412 int zfs_arc_min_prescient_prefetch_ms
= 0;
413 int zfs_arc_p_aggressive_disable
= 1;
414 int zfs_arc_p_dampener_disable
= 1;
415 int zfs_arc_meta_prune
= 10000;
416 int zfs_arc_meta_strategy
= ARC_STRATEGY_META_BALANCED
;
417 int zfs_arc_meta_adjust_restarts
= 4096;
418 int zfs_arc_lotsfree_percent
= 10;
421 static arc_state_t ARC_anon
;
422 static arc_state_t ARC_mru
;
423 static arc_state_t ARC_mru_ghost
;
424 static arc_state_t ARC_mfu
;
425 static arc_state_t ARC_mfu_ghost
;
426 static arc_state_t ARC_l2c_only
;
428 typedef struct arc_stats
{
429 kstat_named_t arcstat_hits
;
430 kstat_named_t arcstat_misses
;
431 kstat_named_t arcstat_demand_data_hits
;
432 kstat_named_t arcstat_demand_data_misses
;
433 kstat_named_t arcstat_demand_metadata_hits
;
434 kstat_named_t arcstat_demand_metadata_misses
;
435 kstat_named_t arcstat_prefetch_data_hits
;
436 kstat_named_t arcstat_prefetch_data_misses
;
437 kstat_named_t arcstat_prefetch_metadata_hits
;
438 kstat_named_t arcstat_prefetch_metadata_misses
;
439 kstat_named_t arcstat_mru_hits
;
440 kstat_named_t arcstat_mru_ghost_hits
;
441 kstat_named_t arcstat_mfu_hits
;
442 kstat_named_t arcstat_mfu_ghost_hits
;
443 kstat_named_t arcstat_deleted
;
445 * Number of buffers that could not be evicted because the hash lock
446 * was held by another thread. The lock may not necessarily be held
447 * by something using the same buffer, since hash locks are shared
448 * by multiple buffers.
450 kstat_named_t arcstat_mutex_miss
;
452 * Number of buffers skipped because they have I/O in progress, are
453 * indrect prefetch buffers that have not lived long enough, or are
454 * not from the spa we're trying to evict from.
456 kstat_named_t arcstat_evict_skip
;
458 * Number of times arc_evict_state() was unable to evict enough
459 * buffers to reach its target amount.
461 kstat_named_t arcstat_evict_not_enough
;
462 kstat_named_t arcstat_evict_l2_cached
;
463 kstat_named_t arcstat_evict_l2_eligible
;
464 kstat_named_t arcstat_evict_l2_ineligible
;
465 kstat_named_t arcstat_evict_l2_skip
;
466 kstat_named_t arcstat_hash_elements
;
467 kstat_named_t arcstat_hash_elements_max
;
468 kstat_named_t arcstat_hash_collisions
;
469 kstat_named_t arcstat_hash_chains
;
470 kstat_named_t arcstat_hash_chain_max
;
471 kstat_named_t arcstat_p
;
472 kstat_named_t arcstat_c
;
473 kstat_named_t arcstat_c_min
;
474 kstat_named_t arcstat_c_max
;
475 kstat_named_t arcstat_size
;
477 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
478 * Note that the compressed bytes may match the uncompressed bytes
479 * if the block is either not compressed or compressed arc is disabled.
481 kstat_named_t arcstat_compressed_size
;
483 * Uncompressed size of the data stored in b_pabd. If compressed
484 * arc is disabled then this value will be identical to the stat
487 kstat_named_t arcstat_uncompressed_size
;
489 * Number of bytes stored in all the arc_buf_t's. This is classified
490 * as "overhead" since this data is typically short-lived and will
491 * be evicted from the arc when it becomes unreferenced unless the
492 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
493 * values have been set (see comment in dbuf.c for more information).
495 kstat_named_t arcstat_overhead_size
;
497 * Number of bytes consumed by internal ARC structures necessary
498 * for tracking purposes; these structures are not actually
499 * backed by ARC buffers. This includes arc_buf_hdr_t structures
500 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
501 * caches), and arc_buf_t structures (allocated via arc_buf_t
504 kstat_named_t arcstat_hdr_size
;
506 * Number of bytes consumed by ARC buffers of type equal to
507 * ARC_BUFC_DATA. This is generally consumed by buffers backing
508 * on disk user data (e.g. plain file contents).
510 kstat_named_t arcstat_data_size
;
512 * Number of bytes consumed by ARC buffers of type equal to
513 * ARC_BUFC_METADATA. This is generally consumed by buffers
514 * backing on disk data that is used for internal ZFS
515 * structures (e.g. ZAP, dnode, indirect blocks, etc).
517 kstat_named_t arcstat_metadata_size
;
519 * Number of bytes consumed by dmu_buf_impl_t objects.
521 kstat_named_t arcstat_dbuf_size
;
523 * Number of bytes consumed by dnode_t objects.
525 kstat_named_t arcstat_dnode_size
;
527 * Number of bytes consumed by bonus buffers.
529 kstat_named_t arcstat_bonus_size
;
531 * Total number of bytes consumed by ARC buffers residing in the
532 * arc_anon state. This includes *all* buffers in the arc_anon
533 * state; e.g. data, metadata, evictable, and unevictable buffers
534 * are all included in this value.
536 kstat_named_t arcstat_anon_size
;
538 * Number of bytes consumed by ARC buffers that meet the
539 * following criteria: backing buffers of type ARC_BUFC_DATA,
540 * residing in the arc_anon state, and are eligible for eviction
541 * (e.g. have no outstanding holds on the buffer).
543 kstat_named_t arcstat_anon_evictable_data
;
545 * Number of bytes consumed by ARC buffers that meet the
546 * following criteria: backing buffers of type ARC_BUFC_METADATA,
547 * residing in the arc_anon state, and are eligible for eviction
548 * (e.g. have no outstanding holds on the buffer).
550 kstat_named_t arcstat_anon_evictable_metadata
;
552 * Total number of bytes consumed by ARC buffers residing in the
553 * arc_mru state. This includes *all* buffers in the arc_mru
554 * state; e.g. data, metadata, evictable, and unevictable buffers
555 * are all included in this value.
557 kstat_named_t arcstat_mru_size
;
559 * Number of bytes consumed by ARC buffers that meet the
560 * following criteria: backing buffers of type ARC_BUFC_DATA,
561 * residing in the arc_mru state, and are eligible for eviction
562 * (e.g. have no outstanding holds on the buffer).
564 kstat_named_t arcstat_mru_evictable_data
;
566 * Number of bytes consumed by ARC buffers that meet the
567 * following criteria: backing buffers of type ARC_BUFC_METADATA,
568 * residing in the arc_mru state, and are eligible for eviction
569 * (e.g. have no outstanding holds on the buffer).
571 kstat_named_t arcstat_mru_evictable_metadata
;
573 * Total number of bytes that *would have been* consumed by ARC
574 * buffers in the arc_mru_ghost state. The key thing to note
575 * here, is the fact that this size doesn't actually indicate
576 * RAM consumption. The ghost lists only consist of headers and
577 * don't actually have ARC buffers linked off of these headers.
578 * Thus, *if* the headers had associated ARC buffers, these
579 * buffers *would have* consumed this number of bytes.
581 kstat_named_t arcstat_mru_ghost_size
;
583 * Number of bytes that *would have been* consumed by ARC
584 * buffers that are eligible for eviction, of type
585 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
587 kstat_named_t arcstat_mru_ghost_evictable_data
;
589 * Number of bytes that *would have been* consumed by ARC
590 * buffers that are eligible for eviction, of type
591 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
593 kstat_named_t arcstat_mru_ghost_evictable_metadata
;
595 * Total number of bytes consumed by ARC buffers residing in the
596 * arc_mfu state. This includes *all* buffers in the arc_mfu
597 * state; e.g. data, metadata, evictable, and unevictable buffers
598 * are all included in this value.
600 kstat_named_t arcstat_mfu_size
;
602 * Number of bytes consumed by ARC buffers that are eligible for
603 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
606 kstat_named_t arcstat_mfu_evictable_data
;
608 * Number of bytes consumed by ARC buffers that are eligible for
609 * eviction, of type ARC_BUFC_METADATA, and reside in the
612 kstat_named_t arcstat_mfu_evictable_metadata
;
614 * Total number of bytes that *would have been* consumed by ARC
615 * buffers in the arc_mfu_ghost state. See the comment above
616 * arcstat_mru_ghost_size for more details.
618 kstat_named_t arcstat_mfu_ghost_size
;
620 * Number of bytes that *would have been* consumed by ARC
621 * buffers that are eligible for eviction, of type
622 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
624 kstat_named_t arcstat_mfu_ghost_evictable_data
;
626 * Number of bytes that *would have been* consumed by ARC
627 * buffers that are eligible for eviction, of type
628 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
630 kstat_named_t arcstat_mfu_ghost_evictable_metadata
;
631 kstat_named_t arcstat_l2_hits
;
632 kstat_named_t arcstat_l2_misses
;
633 kstat_named_t arcstat_l2_feeds
;
634 kstat_named_t arcstat_l2_rw_clash
;
635 kstat_named_t arcstat_l2_read_bytes
;
636 kstat_named_t arcstat_l2_write_bytes
;
637 kstat_named_t arcstat_l2_writes_sent
;
638 kstat_named_t arcstat_l2_writes_done
;
639 kstat_named_t arcstat_l2_writes_error
;
640 kstat_named_t arcstat_l2_writes_lock_retry
;
641 kstat_named_t arcstat_l2_evict_lock_retry
;
642 kstat_named_t arcstat_l2_evict_reading
;
643 kstat_named_t arcstat_l2_evict_l1cached
;
644 kstat_named_t arcstat_l2_free_on_write
;
645 kstat_named_t arcstat_l2_abort_lowmem
;
646 kstat_named_t arcstat_l2_cksum_bad
;
647 kstat_named_t arcstat_l2_io_error
;
648 kstat_named_t arcstat_l2_lsize
;
649 kstat_named_t arcstat_l2_psize
;
650 kstat_named_t arcstat_l2_hdr_size
;
651 kstat_named_t arcstat_memory_throttle_count
;
652 kstat_named_t arcstat_memory_direct_count
;
653 kstat_named_t arcstat_memory_indirect_count
;
654 kstat_named_t arcstat_memory_all_bytes
;
655 kstat_named_t arcstat_memory_free_bytes
;
656 kstat_named_t arcstat_memory_available_bytes
;
657 kstat_named_t arcstat_no_grow
;
658 kstat_named_t arcstat_tempreserve
;
659 kstat_named_t arcstat_loaned_bytes
;
660 kstat_named_t arcstat_prune
;
661 kstat_named_t arcstat_meta_used
;
662 kstat_named_t arcstat_meta_limit
;
663 kstat_named_t arcstat_dnode_limit
;
664 kstat_named_t arcstat_meta_max
;
665 kstat_named_t arcstat_meta_min
;
666 kstat_named_t arcstat_sync_wait_for_async
;
667 kstat_named_t arcstat_demand_hit_predictive_prefetch
;
668 kstat_named_t arcstat_demand_hit_prescient_prefetch
;
669 kstat_named_t arcstat_need_free
;
670 kstat_named_t arcstat_sys_free
;
671 kstat_named_t arcstat_raw_size
;
674 static arc_stats_t arc_stats
= {
675 { "hits", KSTAT_DATA_UINT64
},
676 { "misses", KSTAT_DATA_UINT64
},
677 { "demand_data_hits", KSTAT_DATA_UINT64
},
678 { "demand_data_misses", KSTAT_DATA_UINT64
},
679 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
680 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
681 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
682 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
683 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
684 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
685 { "mru_hits", KSTAT_DATA_UINT64
},
686 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
687 { "mfu_hits", KSTAT_DATA_UINT64
},
688 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
689 { "deleted", KSTAT_DATA_UINT64
},
690 { "mutex_miss", KSTAT_DATA_UINT64
},
691 { "evict_skip", KSTAT_DATA_UINT64
},
692 { "evict_not_enough", KSTAT_DATA_UINT64
},
693 { "evict_l2_cached", KSTAT_DATA_UINT64
},
694 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
695 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
696 { "evict_l2_skip", KSTAT_DATA_UINT64
},
697 { "hash_elements", KSTAT_DATA_UINT64
},
698 { "hash_elements_max", KSTAT_DATA_UINT64
},
699 { "hash_collisions", KSTAT_DATA_UINT64
},
700 { "hash_chains", KSTAT_DATA_UINT64
},
701 { "hash_chain_max", KSTAT_DATA_UINT64
},
702 { "p", KSTAT_DATA_UINT64
},
703 { "c", KSTAT_DATA_UINT64
},
704 { "c_min", KSTAT_DATA_UINT64
},
705 { "c_max", KSTAT_DATA_UINT64
},
706 { "size", KSTAT_DATA_UINT64
},
707 { "compressed_size", KSTAT_DATA_UINT64
},
708 { "uncompressed_size", KSTAT_DATA_UINT64
},
709 { "overhead_size", KSTAT_DATA_UINT64
},
710 { "hdr_size", KSTAT_DATA_UINT64
},
711 { "data_size", KSTAT_DATA_UINT64
},
712 { "metadata_size", KSTAT_DATA_UINT64
},
713 { "dbuf_size", KSTAT_DATA_UINT64
},
714 { "dnode_size", KSTAT_DATA_UINT64
},
715 { "bonus_size", KSTAT_DATA_UINT64
},
716 { "anon_size", KSTAT_DATA_UINT64
},
717 { "anon_evictable_data", KSTAT_DATA_UINT64
},
718 { "anon_evictable_metadata", KSTAT_DATA_UINT64
},
719 { "mru_size", KSTAT_DATA_UINT64
},
720 { "mru_evictable_data", KSTAT_DATA_UINT64
},
721 { "mru_evictable_metadata", KSTAT_DATA_UINT64
},
722 { "mru_ghost_size", KSTAT_DATA_UINT64
},
723 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64
},
724 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
725 { "mfu_size", KSTAT_DATA_UINT64
},
726 { "mfu_evictable_data", KSTAT_DATA_UINT64
},
727 { "mfu_evictable_metadata", KSTAT_DATA_UINT64
},
728 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
729 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64
},
730 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
731 { "l2_hits", KSTAT_DATA_UINT64
},
732 { "l2_misses", KSTAT_DATA_UINT64
},
733 { "l2_feeds", KSTAT_DATA_UINT64
},
734 { "l2_rw_clash", KSTAT_DATA_UINT64
},
735 { "l2_read_bytes", KSTAT_DATA_UINT64
},
736 { "l2_write_bytes", KSTAT_DATA_UINT64
},
737 { "l2_writes_sent", KSTAT_DATA_UINT64
},
738 { "l2_writes_done", KSTAT_DATA_UINT64
},
739 { "l2_writes_error", KSTAT_DATA_UINT64
},
740 { "l2_writes_lock_retry", KSTAT_DATA_UINT64
},
741 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
742 { "l2_evict_reading", KSTAT_DATA_UINT64
},
743 { "l2_evict_l1cached", KSTAT_DATA_UINT64
},
744 { "l2_free_on_write", KSTAT_DATA_UINT64
},
745 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
746 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
747 { "l2_io_error", KSTAT_DATA_UINT64
},
748 { "l2_size", KSTAT_DATA_UINT64
},
749 { "l2_asize", KSTAT_DATA_UINT64
},
750 { "l2_hdr_size", KSTAT_DATA_UINT64
},
751 { "memory_throttle_count", KSTAT_DATA_UINT64
},
752 { "memory_direct_count", KSTAT_DATA_UINT64
},
753 { "memory_indirect_count", KSTAT_DATA_UINT64
},
754 { "memory_all_bytes", KSTAT_DATA_UINT64
},
755 { "memory_free_bytes", KSTAT_DATA_UINT64
},
756 { "memory_available_bytes", KSTAT_DATA_INT64
},
757 { "arc_no_grow", KSTAT_DATA_UINT64
},
758 { "arc_tempreserve", KSTAT_DATA_UINT64
},
759 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
760 { "arc_prune", KSTAT_DATA_UINT64
},
761 { "arc_meta_used", KSTAT_DATA_UINT64
},
762 { "arc_meta_limit", KSTAT_DATA_UINT64
},
763 { "arc_dnode_limit", KSTAT_DATA_UINT64
},
764 { "arc_meta_max", KSTAT_DATA_UINT64
},
765 { "arc_meta_min", KSTAT_DATA_UINT64
},
766 { "sync_wait_for_async", KSTAT_DATA_UINT64
},
767 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64
},
768 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64
},
769 { "arc_need_free", KSTAT_DATA_UINT64
},
770 { "arc_sys_free", KSTAT_DATA_UINT64
},
771 { "arc_raw_size", KSTAT_DATA_UINT64
}
774 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
776 #define ARCSTAT_INCR(stat, val) \
777 atomic_add_64(&arc_stats.stat.value.ui64, (val))
779 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
780 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
782 #define ARCSTAT_MAX(stat, val) { \
784 while ((val) > (m = arc_stats.stat.value.ui64) && \
785 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
789 #define ARCSTAT_MAXSTAT(stat) \
790 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
793 * We define a macro to allow ARC hits/misses to be easily broken down by
794 * two separate conditions, giving a total of four different subtypes for
795 * each of hits and misses (so eight statistics total).
797 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
800 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
802 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
806 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
808 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
813 static arc_state_t
*arc_anon
;
814 static arc_state_t
*arc_mru
;
815 static arc_state_t
*arc_mru_ghost
;
816 static arc_state_t
*arc_mfu
;
817 static arc_state_t
*arc_mfu_ghost
;
818 static arc_state_t
*arc_l2c_only
;
821 * There are several ARC variables that are critical to export as kstats --
822 * but we don't want to have to grovel around in the kstat whenever we wish to
823 * manipulate them. For these variables, we therefore define them to be in
824 * terms of the statistic variable. This assures that we are not introducing
825 * the possibility of inconsistency by having shadow copies of the variables,
826 * while still allowing the code to be readable.
828 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
829 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
830 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
831 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
832 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
833 #define arc_no_grow ARCSTAT(arcstat_no_grow) /* do not grow cache size */
834 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
835 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
836 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
837 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
838 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
839 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
840 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
841 #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
842 #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */
843 #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
844 #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */
845 #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */
847 /* size of all b_rabd's in entire arc */
848 #define arc_raw_size ARCSTAT(arcstat_raw_size)
849 /* compressed size of entire arc */
850 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
851 /* uncompressed size of entire arc */
852 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
853 /* number of bytes in the arc from arc_buf_t's */
854 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
856 static list_t arc_prune_list
;
857 static kmutex_t arc_prune_mtx
;
858 static taskq_t
*arc_prune_taskq
;
860 #define GHOST_STATE(state) \
861 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
862 (state) == arc_l2c_only)
864 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
865 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
866 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
867 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
868 #define HDR_PRESCIENT_PREFETCH(hdr) \
869 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
870 #define HDR_COMPRESSION_ENABLED(hdr) \
871 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
873 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
874 #define HDR_L2_READING(hdr) \
875 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
876 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
877 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
878 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
879 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
880 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
881 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
882 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
884 #define HDR_ISTYPE_METADATA(hdr) \
885 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
886 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
888 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
889 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
890 #define HDR_HAS_RABD(hdr) \
891 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
892 (hdr)->b_crypt_hdr.b_rabd != NULL)
893 #define HDR_ENCRYPTED(hdr) \
894 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
895 #define HDR_AUTHENTICATED(hdr) \
896 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
898 /* For storing compression mode in b_flags */
899 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
901 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
902 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
903 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
904 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
906 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
907 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
908 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
909 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
915 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
916 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
917 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
920 * Hash table routines
923 #define HT_LOCK_ALIGN 64
924 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
929 unsigned char pad
[HT_LOCK_PAD
];
933 #define BUF_LOCKS 8192
934 typedef struct buf_hash_table
{
936 arc_buf_hdr_t
**ht_table
;
937 struct ht_lock ht_locks
[BUF_LOCKS
];
940 static buf_hash_table_t buf_hash_table
;
942 #define BUF_HASH_INDEX(spa, dva, birth) \
943 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
944 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
945 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
946 #define HDR_LOCK(hdr) \
947 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
949 uint64_t zfs_crc64_table
[256];
955 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
956 #define L2ARC_HEADROOM 2 /* num of writes */
959 * If we discover during ARC scan any buffers to be compressed, we boost
960 * our headroom for the next scanning cycle by this percentage multiple.
962 #define L2ARC_HEADROOM_BOOST 200
963 #define L2ARC_FEED_SECS 1 /* caching interval secs */
964 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
967 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
968 * and each of the state has two types: data and metadata.
970 #define L2ARC_FEED_TYPES 4
972 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
973 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
975 /* L2ARC Performance Tunables */
976 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
977 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
978 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
979 unsigned long l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
980 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
981 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
982 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
983 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
984 int l2arc_norw
= B_FALSE
; /* no reads during writes */
989 static list_t L2ARC_dev_list
; /* device list */
990 static list_t
*l2arc_dev_list
; /* device list pointer */
991 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
992 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
993 static list_t L2ARC_free_on_write
; /* free after write buf list */
994 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
995 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
996 static uint64_t l2arc_ndev
; /* number of devices */
998 typedef struct l2arc_read_callback
{
999 arc_buf_hdr_t
*l2rcb_hdr
; /* read header */
1000 blkptr_t l2rcb_bp
; /* original blkptr */
1001 zbookmark_phys_t l2rcb_zb
; /* original bookmark */
1002 int l2rcb_flags
; /* original flags */
1003 abd_t
*l2rcb_abd
; /* temporary buffer */
1004 } l2arc_read_callback_t
;
1006 typedef struct l2arc_data_free
{
1007 /* protected by l2arc_free_on_write_mtx */
1010 arc_buf_contents_t l2df_type
;
1011 list_node_t l2df_list_node
;
1012 } l2arc_data_free_t
;
1014 typedef enum arc_fill_flags
{
1015 ARC_FILL_LOCKED
= 1 << 0, /* hdr lock is held */
1016 ARC_FILL_COMPRESSED
= 1 << 1, /* fill with compressed data */
1017 ARC_FILL_ENCRYPTED
= 1 << 2, /* fill with encrypted data */
1018 ARC_FILL_NOAUTH
= 1 << 3, /* don't attempt to authenticate */
1019 ARC_FILL_IN_PLACE
= 1 << 4 /* fill in place (special case) */
1022 static kmutex_t l2arc_feed_thr_lock
;
1023 static kcondvar_t l2arc_feed_thr_cv
;
1024 static uint8_t l2arc_thread_exit
;
1026 static abd_t
*arc_get_data_abd(arc_buf_hdr_t
*, uint64_t, void *);
1027 static void *arc_get_data_buf(arc_buf_hdr_t
*, uint64_t, void *);
1028 static void arc_get_data_impl(arc_buf_hdr_t
*, uint64_t, void *);
1029 static void arc_free_data_abd(arc_buf_hdr_t
*, abd_t
*, uint64_t, void *);
1030 static void arc_free_data_buf(arc_buf_hdr_t
*, void *, uint64_t, void *);
1031 static void arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
);
1032 static void arc_hdr_free_abd(arc_buf_hdr_t
*, boolean_t
);
1033 static void arc_hdr_alloc_abd(arc_buf_hdr_t
*, boolean_t
);
1034 static void arc_access(arc_buf_hdr_t
*, kmutex_t
*);
1035 static boolean_t
arc_is_overflowing(void);
1036 static void arc_buf_watch(arc_buf_t
*);
1037 static void arc_tuning_update(void);
1038 static void arc_prune_async(int64_t);
1039 static uint64_t arc_all_memory(void);
1041 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t
*);
1042 static uint32_t arc_bufc_to_flags(arc_buf_contents_t
);
1043 static inline void arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
1044 static inline void arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
1046 static boolean_t
l2arc_write_eligible(uint64_t, arc_buf_hdr_t
*);
1047 static void l2arc_read_done(zio_t
*);
1050 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
1052 uint8_t *vdva
= (uint8_t *)dva
;
1053 uint64_t crc
= -1ULL;
1056 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
1058 for (i
= 0; i
< sizeof (dva_t
); i
++)
1059 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
1061 crc
^= (spa
>>8) ^ birth
;
1066 #define HDR_EMPTY(hdr) \
1067 ((hdr)->b_dva.dva_word[0] == 0 && \
1068 (hdr)->b_dva.dva_word[1] == 0)
1070 #define HDR_EQUAL(spa, dva, birth, hdr) \
1071 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1072 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1073 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1076 buf_discard_identity(arc_buf_hdr_t
*hdr
)
1078 hdr
->b_dva
.dva_word
[0] = 0;
1079 hdr
->b_dva
.dva_word
[1] = 0;
1083 static arc_buf_hdr_t
*
1084 buf_hash_find(uint64_t spa
, const blkptr_t
*bp
, kmutex_t
**lockp
)
1086 const dva_t
*dva
= BP_IDENTITY(bp
);
1087 uint64_t birth
= BP_PHYSICAL_BIRTH(bp
);
1088 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
1089 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1092 mutex_enter(hash_lock
);
1093 for (hdr
= buf_hash_table
.ht_table
[idx
]; hdr
!= NULL
;
1094 hdr
= hdr
->b_hash_next
) {
1095 if (HDR_EQUAL(spa
, dva
, birth
, hdr
)) {
1100 mutex_exit(hash_lock
);
1106 * Insert an entry into the hash table. If there is already an element
1107 * equal to elem in the hash table, then the already existing element
1108 * will be returned and the new element will not be inserted.
1109 * Otherwise returns NULL.
1110 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1112 static arc_buf_hdr_t
*
1113 buf_hash_insert(arc_buf_hdr_t
*hdr
, kmutex_t
**lockp
)
1115 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1116 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1117 arc_buf_hdr_t
*fhdr
;
1120 ASSERT(!DVA_IS_EMPTY(&hdr
->b_dva
));
1121 ASSERT(hdr
->b_birth
!= 0);
1122 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
1124 if (lockp
!= NULL
) {
1126 mutex_enter(hash_lock
);
1128 ASSERT(MUTEX_HELD(hash_lock
));
1131 for (fhdr
= buf_hash_table
.ht_table
[idx
], i
= 0; fhdr
!= NULL
;
1132 fhdr
= fhdr
->b_hash_next
, i
++) {
1133 if (HDR_EQUAL(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
, fhdr
))
1137 hdr
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
1138 buf_hash_table
.ht_table
[idx
] = hdr
;
1139 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1141 /* collect some hash table performance data */
1143 ARCSTAT_BUMP(arcstat_hash_collisions
);
1145 ARCSTAT_BUMP(arcstat_hash_chains
);
1147 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
1150 ARCSTAT_BUMP(arcstat_hash_elements
);
1151 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
1157 buf_hash_remove(arc_buf_hdr_t
*hdr
)
1159 arc_buf_hdr_t
*fhdr
, **hdrp
;
1160 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1162 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
1163 ASSERT(HDR_IN_HASH_TABLE(hdr
));
1165 hdrp
= &buf_hash_table
.ht_table
[idx
];
1166 while ((fhdr
= *hdrp
) != hdr
) {
1167 ASSERT3P(fhdr
, !=, NULL
);
1168 hdrp
= &fhdr
->b_hash_next
;
1170 *hdrp
= hdr
->b_hash_next
;
1171 hdr
->b_hash_next
= NULL
;
1172 arc_hdr_clear_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1174 /* collect some hash table performance data */
1175 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
1177 if (buf_hash_table
.ht_table
[idx
] &&
1178 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
1179 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
1183 * Global data structures and functions for the buf kmem cache.
1186 static kmem_cache_t
*hdr_full_cache
;
1187 static kmem_cache_t
*hdr_full_crypt_cache
;
1188 static kmem_cache_t
*hdr_l2only_cache
;
1189 static kmem_cache_t
*buf_cache
;
1196 #if defined(_KERNEL) && defined(HAVE_SPL)
1198 * Large allocations which do not require contiguous pages
1199 * should be using vmem_free() in the linux kernel\
1201 vmem_free(buf_hash_table
.ht_table
,
1202 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1204 kmem_free(buf_hash_table
.ht_table
,
1205 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1207 for (i
= 0; i
< BUF_LOCKS
; i
++)
1208 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
1209 kmem_cache_destroy(hdr_full_cache
);
1210 kmem_cache_destroy(hdr_full_crypt_cache
);
1211 kmem_cache_destroy(hdr_l2only_cache
);
1212 kmem_cache_destroy(buf_cache
);
1216 * Constructor callback - called when the cache is empty
1217 * and a new buf is requested.
1221 hdr_full_cons(void *vbuf
, void *unused
, int kmflag
)
1223 arc_buf_hdr_t
*hdr
= vbuf
;
1225 bzero(hdr
, HDR_FULL_SIZE
);
1226 cv_init(&hdr
->b_l1hdr
.b_cv
, NULL
, CV_DEFAULT
, NULL
);
1227 refcount_create(&hdr
->b_l1hdr
.b_refcnt
);
1228 mutex_init(&hdr
->b_l1hdr
.b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1229 list_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1230 list_link_init(&hdr
->b_l2hdr
.b_l2node
);
1231 multilist_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1232 arc_space_consume(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1239 hdr_full_crypt_cons(void *vbuf
, void *unused
, int kmflag
)
1241 arc_buf_hdr_t
*hdr
= vbuf
;
1243 hdr_full_cons(vbuf
, unused
, kmflag
);
1244 bzero(&hdr
->b_crypt_hdr
, sizeof (hdr
->b_crypt_hdr
));
1245 arc_space_consume(sizeof (hdr
->b_crypt_hdr
), ARC_SPACE_HDRS
);
1252 hdr_l2only_cons(void *vbuf
, void *unused
, int kmflag
)
1254 arc_buf_hdr_t
*hdr
= vbuf
;
1256 bzero(hdr
, HDR_L2ONLY_SIZE
);
1257 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1264 buf_cons(void *vbuf
, void *unused
, int kmflag
)
1266 arc_buf_t
*buf
= vbuf
;
1268 bzero(buf
, sizeof (arc_buf_t
));
1269 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1270 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1276 * Destructor callback - called when a cached buf is
1277 * no longer required.
1281 hdr_full_dest(void *vbuf
, void *unused
)
1283 arc_buf_hdr_t
*hdr
= vbuf
;
1285 ASSERT(HDR_EMPTY(hdr
));
1286 cv_destroy(&hdr
->b_l1hdr
.b_cv
);
1287 refcount_destroy(&hdr
->b_l1hdr
.b_refcnt
);
1288 mutex_destroy(&hdr
->b_l1hdr
.b_freeze_lock
);
1289 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1290 arc_space_return(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1295 hdr_full_crypt_dest(void *vbuf
, void *unused
)
1297 arc_buf_hdr_t
*hdr
= vbuf
;
1299 hdr_full_dest(vbuf
, unused
);
1300 arc_space_return(sizeof (hdr
->b_crypt_hdr
), ARC_SPACE_HDRS
);
1305 hdr_l2only_dest(void *vbuf
, void *unused
)
1307 ASSERTV(arc_buf_hdr_t
*hdr
= vbuf
);
1309 ASSERT(HDR_EMPTY(hdr
));
1310 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1315 buf_dest(void *vbuf
, void *unused
)
1317 arc_buf_t
*buf
= vbuf
;
1319 mutex_destroy(&buf
->b_evict_lock
);
1320 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1324 * Reclaim callback -- invoked when memory is low.
1328 hdr_recl(void *unused
)
1330 dprintf("hdr_recl called\n");
1332 * umem calls the reclaim func when we destroy the buf cache,
1333 * which is after we do arc_fini().
1336 cv_signal(&arc_reclaim_thread_cv
);
1342 uint64_t *ct
= NULL
;
1343 uint64_t hsize
= 1ULL << 12;
1347 * The hash table is big enough to fill all of physical memory
1348 * with an average block size of zfs_arc_average_blocksize (default 8K).
1349 * By default, the table will take up
1350 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1352 while (hsize
* zfs_arc_average_blocksize
< arc_all_memory())
1355 buf_hash_table
.ht_mask
= hsize
- 1;
1356 #if defined(_KERNEL) && defined(HAVE_SPL)
1358 * Large allocations which do not require contiguous pages
1359 * should be using vmem_alloc() in the linux kernel
1361 buf_hash_table
.ht_table
=
1362 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
1364 buf_hash_table
.ht_table
=
1365 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
1367 if (buf_hash_table
.ht_table
== NULL
) {
1368 ASSERT(hsize
> (1ULL << 8));
1373 hdr_full_cache
= kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE
,
1374 0, hdr_full_cons
, hdr_full_dest
, hdr_recl
, NULL
, NULL
, 0);
1375 hdr_full_crypt_cache
= kmem_cache_create("arc_buf_hdr_t_full_crypt",
1376 HDR_FULL_CRYPT_SIZE
, 0, hdr_full_crypt_cons
, hdr_full_crypt_dest
,
1377 hdr_recl
, NULL
, NULL
, 0);
1378 hdr_l2only_cache
= kmem_cache_create("arc_buf_hdr_t_l2only",
1379 HDR_L2ONLY_SIZE
, 0, hdr_l2only_cons
, hdr_l2only_dest
, hdr_recl
,
1381 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
1382 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
1384 for (i
= 0; i
< 256; i
++)
1385 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
1386 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
1388 for (i
= 0; i
< BUF_LOCKS
; i
++) {
1389 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
1390 NULL
, MUTEX_DEFAULT
, NULL
);
1394 #define ARC_MINTIME (hz>>4) /* 62 ms */
1397 * This is the size that the buf occupies in memory. If the buf is compressed,
1398 * it will correspond to the compressed size. You should use this method of
1399 * getting the buf size unless you explicitly need the logical size.
1402 arc_buf_size(arc_buf_t
*buf
)
1404 return (ARC_BUF_COMPRESSED(buf
) ?
1405 HDR_GET_PSIZE(buf
->b_hdr
) : HDR_GET_LSIZE(buf
->b_hdr
));
1409 arc_buf_lsize(arc_buf_t
*buf
)
1411 return (HDR_GET_LSIZE(buf
->b_hdr
));
1415 * This function will return B_TRUE if the buffer is encrypted in memory.
1416 * This buffer can be decrypted by calling arc_untransform().
1419 arc_is_encrypted(arc_buf_t
*buf
)
1421 return (ARC_BUF_ENCRYPTED(buf
) != 0);
1425 * Returns B_TRUE if the buffer represents data that has not had its MAC
1429 arc_is_unauthenticated(arc_buf_t
*buf
)
1431 return (HDR_NOAUTH(buf
->b_hdr
) != 0);
1435 arc_get_raw_params(arc_buf_t
*buf
, boolean_t
*byteorder
, uint8_t *salt
,
1436 uint8_t *iv
, uint8_t *mac
)
1438 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1440 ASSERT(HDR_PROTECTED(hdr
));
1442 bcopy(hdr
->b_crypt_hdr
.b_salt
, salt
, ZIO_DATA_SALT_LEN
);
1443 bcopy(hdr
->b_crypt_hdr
.b_iv
, iv
, ZIO_DATA_IV_LEN
);
1444 bcopy(hdr
->b_crypt_hdr
.b_mac
, mac
, ZIO_DATA_MAC_LEN
);
1445 *byteorder
= (hdr
->b_l1hdr
.b_byteswap
== DMU_BSWAP_NUMFUNCS
) ?
1446 ZFS_HOST_BYTEORDER
: !ZFS_HOST_BYTEORDER
;
1450 * Indicates how this buffer is compressed in memory. If it is not compressed
1451 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1452 * arc_untransform() as long as it is also unencrypted.
1455 arc_get_compression(arc_buf_t
*buf
)
1457 return (ARC_BUF_COMPRESSED(buf
) ?
1458 HDR_GET_COMPRESS(buf
->b_hdr
) : ZIO_COMPRESS_OFF
);
1462 * Return the compression algorithm used to store this data in the ARC. If ARC
1463 * compression is enabled or this is an encrypted block, this will be the same
1464 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1466 static inline enum zio_compress
1467 arc_hdr_get_compress(arc_buf_hdr_t
*hdr
)
1469 return (HDR_COMPRESSION_ENABLED(hdr
) ?
1470 HDR_GET_COMPRESS(hdr
) : ZIO_COMPRESS_OFF
);
1473 static inline boolean_t
1474 arc_buf_is_shared(arc_buf_t
*buf
)
1476 boolean_t shared
= (buf
->b_data
!= NULL
&&
1477 buf
->b_hdr
->b_l1hdr
.b_pabd
!= NULL
&&
1478 abd_is_linear(buf
->b_hdr
->b_l1hdr
.b_pabd
) &&
1479 buf
->b_data
== abd_to_buf(buf
->b_hdr
->b_l1hdr
.b_pabd
));
1480 IMPLY(shared
, HDR_SHARED_DATA(buf
->b_hdr
));
1481 IMPLY(shared
, ARC_BUF_SHARED(buf
));
1482 IMPLY(shared
, ARC_BUF_COMPRESSED(buf
) || ARC_BUF_LAST(buf
));
1485 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1486 * already being shared" requirement prevents us from doing that.
1493 * Free the checksum associated with this header. If there is no checksum, this
1497 arc_cksum_free(arc_buf_hdr_t
*hdr
)
1499 ASSERT(HDR_HAS_L1HDR(hdr
));
1501 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1502 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1503 kmem_free(hdr
->b_l1hdr
.b_freeze_cksum
, sizeof (zio_cksum_t
));
1504 hdr
->b_l1hdr
.b_freeze_cksum
= NULL
;
1506 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1510 * Return true iff at least one of the bufs on hdr is not compressed.
1511 * Encrypted buffers count as compressed.
1514 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t
*hdr
)
1516 for (arc_buf_t
*b
= hdr
->b_l1hdr
.b_buf
; b
!= NULL
; b
= b
->b_next
) {
1517 if (!ARC_BUF_COMPRESSED(b
)) {
1526 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1527 * matches the checksum that is stored in the hdr. If there is no checksum,
1528 * or if the buf is compressed, this is a no-op.
1531 arc_cksum_verify(arc_buf_t
*buf
)
1533 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1536 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1539 if (ARC_BUF_COMPRESSED(buf
)) {
1540 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1541 arc_hdr_has_uncompressed_buf(hdr
));
1545 ASSERT(HDR_HAS_L1HDR(hdr
));
1547 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1548 if (hdr
->b_l1hdr
.b_freeze_cksum
== NULL
|| HDR_IO_ERROR(hdr
)) {
1549 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1553 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
, &zc
);
1554 if (!ZIO_CHECKSUM_EQUAL(*hdr
->b_l1hdr
.b_freeze_cksum
, zc
))
1555 panic("buffer modified while frozen!");
1556 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1560 * This function makes the assumption that data stored in the L2ARC
1561 * will be transformed exactly as it is in the main pool. Because of
1562 * this we can verify the checksum against the reading process's bp.
1565 arc_cksum_is_equal(arc_buf_hdr_t
*hdr
, zio_t
*zio
)
1567 ASSERT(!BP_IS_EMBEDDED(zio
->io_bp
));
1568 VERIFY3U(BP_GET_PSIZE(zio
->io_bp
), ==, HDR_GET_PSIZE(hdr
));
1571 * Block pointers always store the checksum for the logical data.
1572 * If the block pointer has the gang bit set, then the checksum
1573 * it represents is for the reconstituted data and not for an
1574 * individual gang member. The zio pipeline, however, must be able to
1575 * determine the checksum of each of the gang constituents so it
1576 * treats the checksum comparison differently than what we need
1577 * for l2arc blocks. This prevents us from using the
1578 * zio_checksum_error() interface directly. Instead we must call the
1579 * zio_checksum_error_impl() so that we can ensure the checksum is
1580 * generated using the correct checksum algorithm and accounts for the
1581 * logical I/O size and not just a gang fragment.
1583 return (zio_checksum_error_impl(zio
->io_spa
, zio
->io_bp
,
1584 BP_GET_CHECKSUM(zio
->io_bp
), zio
->io_abd
, zio
->io_size
,
1585 zio
->io_offset
, NULL
) == 0);
1589 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1590 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1591 * isn't modified later on. If buf is compressed or there is already a checksum
1592 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1595 arc_cksum_compute(arc_buf_t
*buf
)
1597 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1599 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1602 ASSERT(HDR_HAS_L1HDR(hdr
));
1604 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1605 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1606 ASSERT(arc_hdr_has_uncompressed_buf(hdr
));
1607 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1609 } else if (ARC_BUF_COMPRESSED(buf
)) {
1610 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1614 ASSERT(!ARC_BUF_ENCRYPTED(buf
));
1615 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1616 hdr
->b_l1hdr
.b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
),
1618 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
,
1619 hdr
->b_l1hdr
.b_freeze_cksum
);
1620 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1626 arc_buf_sigsegv(int sig
, siginfo_t
*si
, void *unused
)
1628 panic("Got SIGSEGV at address: 0x%lx\n", (long)si
->si_addr
);
1634 arc_buf_unwatch(arc_buf_t
*buf
)
1638 ASSERT0(mprotect(buf
->b_data
, arc_buf_size(buf
),
1639 PROT_READ
| PROT_WRITE
));
1646 arc_buf_watch(arc_buf_t
*buf
)
1650 ASSERT0(mprotect(buf
->b_data
, arc_buf_size(buf
),
1655 static arc_buf_contents_t
1656 arc_buf_type(arc_buf_hdr_t
*hdr
)
1658 arc_buf_contents_t type
;
1659 if (HDR_ISTYPE_METADATA(hdr
)) {
1660 type
= ARC_BUFC_METADATA
;
1662 type
= ARC_BUFC_DATA
;
1664 VERIFY3U(hdr
->b_type
, ==, type
);
1669 arc_is_metadata(arc_buf_t
*buf
)
1671 return (HDR_ISTYPE_METADATA(buf
->b_hdr
) != 0);
1675 arc_bufc_to_flags(arc_buf_contents_t type
)
1679 /* metadata field is 0 if buffer contains normal data */
1681 case ARC_BUFC_METADATA
:
1682 return (ARC_FLAG_BUFC_METADATA
);
1686 panic("undefined ARC buffer type!");
1687 return ((uint32_t)-1);
1691 arc_buf_thaw(arc_buf_t
*buf
)
1693 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1695 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
1696 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
1698 arc_cksum_verify(buf
);
1701 * Compressed buffers do not manipulate the b_freeze_cksum or
1702 * allocate b_thawed.
1704 if (ARC_BUF_COMPRESSED(buf
)) {
1705 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1706 arc_hdr_has_uncompressed_buf(hdr
));
1710 ASSERT(HDR_HAS_L1HDR(hdr
));
1711 arc_cksum_free(hdr
);
1712 arc_buf_unwatch(buf
);
1716 arc_buf_freeze(arc_buf_t
*buf
)
1718 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1719 kmutex_t
*hash_lock
;
1721 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1724 if (ARC_BUF_COMPRESSED(buf
)) {
1725 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1726 arc_hdr_has_uncompressed_buf(hdr
));
1730 hash_lock
= HDR_LOCK(hdr
);
1731 mutex_enter(hash_lock
);
1733 ASSERT(HDR_HAS_L1HDR(hdr
));
1734 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
||
1735 hdr
->b_l1hdr
.b_state
== arc_anon
);
1736 arc_cksum_compute(buf
);
1737 mutex_exit(hash_lock
);
1741 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1742 * the following functions should be used to ensure that the flags are
1743 * updated in a thread-safe way. When manipulating the flags either
1744 * the hash_lock must be held or the hdr must be undiscoverable. This
1745 * ensures that we're not racing with any other threads when updating
1749 arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1751 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1752 hdr
->b_flags
|= flags
;
1756 arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1758 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1759 hdr
->b_flags
&= ~flags
;
1763 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1764 * done in a special way since we have to clear and set bits
1765 * at the same time. Consumers that wish to set the compression bits
1766 * must use this function to ensure that the flags are updated in
1767 * thread-safe manner.
1770 arc_hdr_set_compress(arc_buf_hdr_t
*hdr
, enum zio_compress cmp
)
1772 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1775 * Holes and embedded blocks will always have a psize = 0 so
1776 * we ignore the compression of the blkptr and set the
1777 * want to uncompress them. Mark them as uncompressed.
1779 if (!zfs_compressed_arc_enabled
|| HDR_GET_PSIZE(hdr
) == 0) {
1780 arc_hdr_clear_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1781 ASSERT(!HDR_COMPRESSION_ENABLED(hdr
));
1783 arc_hdr_set_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1784 ASSERT(HDR_COMPRESSION_ENABLED(hdr
));
1787 HDR_SET_COMPRESS(hdr
, cmp
);
1788 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, cmp
);
1792 * Looks for another buf on the same hdr which has the data decompressed, copies
1793 * from it, and returns true. If no such buf exists, returns false.
1796 arc_buf_try_copy_decompressed_data(arc_buf_t
*buf
)
1798 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1799 boolean_t copied
= B_FALSE
;
1801 ASSERT(HDR_HAS_L1HDR(hdr
));
1802 ASSERT3P(buf
->b_data
, !=, NULL
);
1803 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1805 for (arc_buf_t
*from
= hdr
->b_l1hdr
.b_buf
; from
!= NULL
;
1806 from
= from
->b_next
) {
1807 /* can't use our own data buffer */
1812 if (!ARC_BUF_COMPRESSED(from
)) {
1813 bcopy(from
->b_data
, buf
->b_data
, arc_buf_size(buf
));
1820 * There were no decompressed bufs, so there should not be a
1821 * checksum on the hdr either.
1823 EQUIV(!copied
, hdr
->b_l1hdr
.b_freeze_cksum
== NULL
);
1829 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1832 arc_hdr_size(arc_buf_hdr_t
*hdr
)
1836 if (arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
&&
1837 HDR_GET_PSIZE(hdr
) > 0) {
1838 size
= HDR_GET_PSIZE(hdr
);
1840 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, 0);
1841 size
= HDR_GET_LSIZE(hdr
);
1847 arc_hdr_authenticate(arc_buf_hdr_t
*hdr
, spa_t
*spa
, uint64_t dsobj
)
1851 uint64_t lsize
= HDR_GET_LSIZE(hdr
);
1852 uint64_t psize
= HDR_GET_PSIZE(hdr
);
1853 void *tmpbuf
= NULL
;
1854 abd_t
*abd
= hdr
->b_l1hdr
.b_pabd
;
1856 ASSERT(HDR_LOCK(hdr
) == NULL
|| MUTEX_HELD(HDR_LOCK(hdr
)));
1857 ASSERT(HDR_AUTHENTICATED(hdr
));
1858 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
1861 * The MAC is calculated on the compressed data that is stored on disk.
1862 * However, if compressed arc is disabled we will only have the
1863 * decompressed data available to us now. Compress it into a temporary
1864 * abd so we can verify the MAC. The performance overhead of this will
1865 * be relatively low, since most objects in an encrypted objset will
1866 * be encrypted (instead of authenticated) anyway.
1868 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
1869 !HDR_COMPRESSION_ENABLED(hdr
)) {
1870 tmpbuf
= zio_buf_alloc(lsize
);
1871 abd
= abd_get_from_buf(tmpbuf
, lsize
);
1872 abd_take_ownership_of_buf(abd
, B_TRUE
);
1874 csize
= zio_compress_data(HDR_GET_COMPRESS(hdr
),
1875 hdr
->b_l1hdr
.b_pabd
, tmpbuf
, lsize
);
1876 ASSERT3U(csize
, <=, psize
);
1877 abd_zero_off(abd
, csize
, psize
- csize
);
1881 * Authentication is best effort. We authenticate whenever the key is
1882 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1884 if (hdr
->b_crypt_hdr
.b_ot
== DMU_OT_OBJSET
) {
1885 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1886 ASSERT3U(lsize
, ==, psize
);
1887 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, spa
, dsobj
, abd
,
1888 psize
, hdr
->b_l1hdr
.b_byteswap
!= DMU_BSWAP_NUMFUNCS
);
1890 ret
= spa_do_crypt_mac_abd(B_FALSE
, spa
, dsobj
, abd
, psize
,
1891 hdr
->b_crypt_hdr
.b_mac
);
1895 arc_hdr_clear_flags(hdr
, ARC_FLAG_NOAUTH
);
1896 else if (ret
!= ENOENT
)
1912 * This function will take a header that only has raw encrypted data in
1913 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1914 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1915 * also decompress the data.
1918 arc_hdr_decrypt(arc_buf_hdr_t
*hdr
, spa_t
*spa
, uint64_t dsobj
)
1921 dsl_crypto_key_t
*dck
= NULL
;
1924 boolean_t no_crypt
= B_FALSE
;
1925 boolean_t bswap
= (hdr
->b_l1hdr
.b_byteswap
!= DMU_BSWAP_NUMFUNCS
);
1927 ASSERT(HDR_LOCK(hdr
) == NULL
|| MUTEX_HELD(HDR_LOCK(hdr
)));
1928 ASSERT(HDR_ENCRYPTED(hdr
));
1930 arc_hdr_alloc_abd(hdr
, B_FALSE
);
1933 * We must be careful to use the passed-in dsobj value here and
1934 * not the value in b_dsobj. b_dsobj is meant to be a best guess for
1935 * the L2ARC, which has the luxury of being able to fail without real
1936 * consequences (the data simply won't make it to the L2ARC). In
1937 * reality, the dsobj stored in the header may belong to a dataset
1938 * that has been unmounted or otherwise disowned, meaning the key
1939 * won't be accessible via that dsobj anymore.
1941 ret
= spa_keystore_lookup_key(spa
, dsobj
, FTAG
, &dck
);
1943 ret
= SET_ERROR(EACCES
);
1947 ret
= zio_do_crypt_abd(B_FALSE
, &dck
->dck_key
,
1948 hdr
->b_crypt_hdr
.b_salt
, hdr
->b_crypt_hdr
.b_ot
,
1949 hdr
->b_crypt_hdr
.b_iv
, hdr
->b_crypt_hdr
.b_mac
,
1950 HDR_GET_PSIZE(hdr
), bswap
, hdr
->b_l1hdr
.b_pabd
,
1951 hdr
->b_crypt_hdr
.b_rabd
, &no_crypt
);
1956 abd_copy(hdr
->b_l1hdr
.b_pabd
, hdr
->b_crypt_hdr
.b_rabd
,
1957 HDR_GET_PSIZE(hdr
));
1961 * If this header has disabled arc compression but the b_pabd is
1962 * compressed after decrypting it, we need to decompress the newly
1965 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
1966 !HDR_COMPRESSION_ENABLED(hdr
)) {
1968 * We want to make sure that we are correctly honoring the
1969 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1970 * and then loan a buffer from it, rather than allocating a
1971 * linear buffer and wrapping it in an abd later.
1973 cabd
= arc_get_data_abd(hdr
, arc_hdr_size(hdr
), hdr
);
1974 tmp
= abd_borrow_buf(cabd
, arc_hdr_size(hdr
));
1976 ret
= zio_decompress_data(HDR_GET_COMPRESS(hdr
),
1977 hdr
->b_l1hdr
.b_pabd
, tmp
, HDR_GET_PSIZE(hdr
),
1978 HDR_GET_LSIZE(hdr
));
1980 abd_return_buf(cabd
, tmp
, arc_hdr_size(hdr
));
1984 abd_return_buf_copy(cabd
, tmp
, arc_hdr_size(hdr
));
1985 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
,
1986 arc_hdr_size(hdr
), hdr
);
1987 hdr
->b_l1hdr
.b_pabd
= cabd
;
1990 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
1995 arc_hdr_free_abd(hdr
, B_FALSE
);
1997 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
1999 arc_free_data_buf(hdr
, cabd
, arc_hdr_size(hdr
), hdr
);
2005 * This function is called during arc_buf_fill() to prepare the header's
2006 * abd plaintext pointer for use. This involves authenticated protected
2007 * data and decrypting encrypted data into the plaintext abd.
2010 arc_fill_hdr_crypt(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, spa_t
*spa
,
2011 uint64_t dsobj
, boolean_t noauth
)
2015 ASSERT(HDR_PROTECTED(hdr
));
2017 if (hash_lock
!= NULL
)
2018 mutex_enter(hash_lock
);
2020 if (HDR_NOAUTH(hdr
) && !noauth
) {
2022 * The caller requested authenticated data but our data has
2023 * not been authenticated yet. Verify the MAC now if we can.
2025 ret
= arc_hdr_authenticate(hdr
, spa
, dsobj
);
2028 } else if (HDR_HAS_RABD(hdr
) && hdr
->b_l1hdr
.b_pabd
== NULL
) {
2030 * If we only have the encrypted version of the data, but the
2031 * unencrypted version was requested we take this opportunity
2032 * to store the decrypted version in the header for future use.
2034 ret
= arc_hdr_decrypt(hdr
, spa
, dsobj
);
2039 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2041 if (hash_lock
!= NULL
)
2042 mutex_exit(hash_lock
);
2047 if (hash_lock
!= NULL
)
2048 mutex_exit(hash_lock
);
2054 * This function is used by the dbuf code to decrypt bonus buffers in place.
2055 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
2056 * block, so we use the hash lock here to protect against concurrent calls to
2060 arc_buf_untransform_in_place(arc_buf_t
*buf
, kmutex_t
*hash_lock
)
2062 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2064 ASSERT(HDR_ENCRYPTED(hdr
));
2065 ASSERT3U(hdr
->b_crypt_hdr
.b_ot
, ==, DMU_OT_DNODE
);
2066 ASSERT(HDR_LOCK(hdr
) == NULL
|| MUTEX_HELD(HDR_LOCK(hdr
)));
2067 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2069 zio_crypt_copy_dnode_bonus(hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
2071 buf
->b_flags
&= ~ARC_BUF_FLAG_ENCRYPTED
;
2072 buf
->b_flags
&= ~ARC_BUF_FLAG_COMPRESSED
;
2073 hdr
->b_crypt_hdr
.b_ebufcnt
-= 1;
2077 * Given a buf that has a data buffer attached to it, this function will
2078 * efficiently fill the buf with data of the specified compression setting from
2079 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2080 * are already sharing a data buf, no copy is performed.
2082 * If the buf is marked as compressed but uncompressed data was requested, this
2083 * will allocate a new data buffer for the buf, remove that flag, and fill the
2084 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2085 * uncompressed data, and (since we haven't added support for it yet) if you
2086 * want compressed data your buf must already be marked as compressed and have
2087 * the correct-sized data buffer.
2090 arc_buf_fill(arc_buf_t
*buf
, spa_t
*spa
, uint64_t dsobj
, arc_fill_flags_t flags
)
2093 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2094 boolean_t hdr_compressed
=
2095 (arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
);
2096 boolean_t compressed
= (flags
& ARC_FILL_COMPRESSED
) != 0;
2097 boolean_t encrypted
= (flags
& ARC_FILL_ENCRYPTED
) != 0;
2098 dmu_object_byteswap_t bswap
= hdr
->b_l1hdr
.b_byteswap
;
2099 kmutex_t
*hash_lock
= (flags
& ARC_FILL_LOCKED
) ? NULL
: HDR_LOCK(hdr
);
2101 ASSERT3P(buf
->b_data
, !=, NULL
);
2102 IMPLY(compressed
, hdr_compressed
|| ARC_BUF_ENCRYPTED(buf
));
2103 IMPLY(compressed
, ARC_BUF_COMPRESSED(buf
));
2104 IMPLY(encrypted
, HDR_ENCRYPTED(hdr
));
2105 IMPLY(encrypted
, ARC_BUF_ENCRYPTED(buf
));
2106 IMPLY(encrypted
, ARC_BUF_COMPRESSED(buf
));
2107 IMPLY(encrypted
, !ARC_BUF_SHARED(buf
));
2110 * If the caller wanted encrypted data we just need to copy it from
2111 * b_rabd and potentially byteswap it. We won't be able to do any
2112 * further transforms on it.
2115 ASSERT(HDR_HAS_RABD(hdr
));
2116 abd_copy_to_buf(buf
->b_data
, hdr
->b_crypt_hdr
.b_rabd
,
2117 HDR_GET_PSIZE(hdr
));
2122 * Adjust encrypted and authenticated headers to accomodate the
2123 * request if needed.
2125 if (HDR_PROTECTED(hdr
)) {
2126 error
= arc_fill_hdr_crypt(hdr
, hash_lock
, spa
,
2127 dsobj
, !!(flags
& ARC_FILL_NOAUTH
));
2133 * There is a special case here for dnode blocks which are
2134 * decrypting their bonus buffers. These blocks may request to
2135 * be decrypted in-place. This is necessary because there may
2136 * be many dnodes pointing into this buffer and there is
2137 * currently no method to synchronize replacing the backing
2138 * b_data buffer and updating all of the pointers. Here we use
2139 * the hash lock to ensure there are no races. If the need
2140 * arises for other types to be decrypted in-place, they must
2141 * add handling here as well.
2143 if ((flags
& ARC_FILL_IN_PLACE
) != 0) {
2144 ASSERT(!hdr_compressed
);
2145 ASSERT(!compressed
);
2148 if (HDR_ENCRYPTED(hdr
) && ARC_BUF_ENCRYPTED(buf
)) {
2149 ASSERT3U(hdr
->b_crypt_hdr
.b_ot
, ==, DMU_OT_DNODE
);
2151 if (hash_lock
!= NULL
)
2152 mutex_enter(hash_lock
);
2153 arc_buf_untransform_in_place(buf
, hash_lock
);
2154 if (hash_lock
!= NULL
)
2155 mutex_exit(hash_lock
);
2157 /* Compute the hdr's checksum if necessary */
2158 arc_cksum_compute(buf
);
2164 if (hdr_compressed
== compressed
) {
2165 if (!arc_buf_is_shared(buf
)) {
2166 abd_copy_to_buf(buf
->b_data
, hdr
->b_l1hdr
.b_pabd
,
2170 ASSERT(hdr_compressed
);
2171 ASSERT(!compressed
);
2172 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, HDR_GET_PSIZE(hdr
));
2175 * If the buf is sharing its data with the hdr, unlink it and
2176 * allocate a new data buffer for the buf.
2178 if (arc_buf_is_shared(buf
)) {
2179 ASSERT(ARC_BUF_COMPRESSED(buf
));
2181 /* We need to give the buf it's own b_data */
2182 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
2184 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
2185 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2187 /* Previously overhead was 0; just add new overhead */
2188 ARCSTAT_INCR(arcstat_overhead_size
, HDR_GET_LSIZE(hdr
));
2189 } else if (ARC_BUF_COMPRESSED(buf
)) {
2190 /* We need to reallocate the buf's b_data */
2191 arc_free_data_buf(hdr
, buf
->b_data
, HDR_GET_PSIZE(hdr
),
2194 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
2196 /* We increased the size of b_data; update overhead */
2197 ARCSTAT_INCR(arcstat_overhead_size
,
2198 HDR_GET_LSIZE(hdr
) - HDR_GET_PSIZE(hdr
));
2202 * Regardless of the buf's previous compression settings, it
2203 * should not be compressed at the end of this function.
2205 buf
->b_flags
&= ~ARC_BUF_FLAG_COMPRESSED
;
2208 * Try copying the data from another buf which already has a
2209 * decompressed version. If that's not possible, it's time to
2210 * bite the bullet and decompress the data from the hdr.
2212 if (arc_buf_try_copy_decompressed_data(buf
)) {
2213 /* Skip byteswapping and checksumming (already done) */
2214 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, !=, NULL
);
2217 error
= zio_decompress_data(HDR_GET_COMPRESS(hdr
),
2218 hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
2219 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
2222 * Absent hardware errors or software bugs, this should
2223 * be impossible, but log it anyway so we can debug it.
2227 "hdr %p, compress %d, psize %d, lsize %d",
2228 hdr
, arc_hdr_get_compress(hdr
),
2229 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
2230 return (SET_ERROR(EIO
));
2236 /* Byteswap the buf's data if necessary */
2237 if (bswap
!= DMU_BSWAP_NUMFUNCS
) {
2238 ASSERT(!HDR_SHARED_DATA(hdr
));
2239 ASSERT3U(bswap
, <, DMU_BSWAP_NUMFUNCS
);
2240 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, HDR_GET_LSIZE(hdr
));
2243 /* Compute the hdr's checksum if necessary */
2244 arc_cksum_compute(buf
);
2250 * If this function is being called to decrypt an encrypted buffer or verify an
2251 * authenticated one, the key must be loaded and a mapping must be made
2252 * available in the keystore via spa_keystore_create_mapping() or one of its
2256 arc_untransform(arc_buf_t
*buf
, spa_t
*spa
, uint64_t dsobj
, boolean_t in_place
)
2258 arc_fill_flags_t flags
= 0;
2261 flags
|= ARC_FILL_IN_PLACE
;
2263 return (arc_buf_fill(buf
, spa
, dsobj
, flags
));
2267 * Increment the amount of evictable space in the arc_state_t's refcount.
2268 * We account for the space used by the hdr and the arc buf individually
2269 * so that we can add and remove them from the refcount individually.
2272 arc_evictable_space_increment(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
2274 arc_buf_contents_t type
= arc_buf_type(hdr
);
2276 ASSERT(HDR_HAS_L1HDR(hdr
));
2278 if (GHOST_STATE(state
)) {
2279 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
2280 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2281 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2282 ASSERT(!HDR_HAS_RABD(hdr
));
2283 (void) refcount_add_many(&state
->arcs_esize
[type
],
2284 HDR_GET_LSIZE(hdr
), hdr
);
2288 ASSERT(!GHOST_STATE(state
));
2289 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2290 (void) refcount_add_many(&state
->arcs_esize
[type
],
2291 arc_hdr_size(hdr
), hdr
);
2293 if (HDR_HAS_RABD(hdr
)) {
2294 (void) refcount_add_many(&state
->arcs_esize
[type
],
2295 HDR_GET_PSIZE(hdr
), hdr
);
2298 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2299 buf
= buf
->b_next
) {
2300 if (arc_buf_is_shared(buf
))
2302 (void) refcount_add_many(&state
->arcs_esize
[type
],
2303 arc_buf_size(buf
), buf
);
2308 * Decrement the amount of evictable space in the arc_state_t's refcount.
2309 * We account for the space used by the hdr and the arc buf individually
2310 * so that we can add and remove them from the refcount individually.
2313 arc_evictable_space_decrement(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
2315 arc_buf_contents_t type
= arc_buf_type(hdr
);
2317 ASSERT(HDR_HAS_L1HDR(hdr
));
2319 if (GHOST_STATE(state
)) {
2320 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
2321 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2322 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2323 ASSERT(!HDR_HAS_RABD(hdr
));
2324 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2325 HDR_GET_LSIZE(hdr
), hdr
);
2329 ASSERT(!GHOST_STATE(state
));
2330 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2331 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2332 arc_hdr_size(hdr
), hdr
);
2334 if (HDR_HAS_RABD(hdr
)) {
2335 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2336 HDR_GET_PSIZE(hdr
), hdr
);
2339 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2340 buf
= buf
->b_next
) {
2341 if (arc_buf_is_shared(buf
))
2343 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2344 arc_buf_size(buf
), buf
);
2349 * Add a reference to this hdr indicating that someone is actively
2350 * referencing that memory. When the refcount transitions from 0 to 1,
2351 * we remove it from the respective arc_state_t list to indicate that
2352 * it is not evictable.
2355 add_reference(arc_buf_hdr_t
*hdr
, void *tag
)
2359 ASSERT(HDR_HAS_L1HDR(hdr
));
2360 if (!MUTEX_HELD(HDR_LOCK(hdr
))) {
2361 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
2362 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2363 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2366 state
= hdr
->b_l1hdr
.b_state
;
2368 if ((refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
) == 1) &&
2369 (state
!= arc_anon
)) {
2370 /* We don't use the L2-only state list. */
2371 if (state
!= arc_l2c_only
) {
2372 multilist_remove(state
->arcs_list
[arc_buf_type(hdr
)],
2374 arc_evictable_space_decrement(hdr
, state
);
2376 /* remove the prefetch flag if we get a reference */
2377 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
2382 * Remove a reference from this hdr. When the reference transitions from
2383 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2384 * list making it eligible for eviction.
2387 remove_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
2390 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2392 ASSERT(HDR_HAS_L1HDR(hdr
));
2393 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
2394 ASSERT(!GHOST_STATE(state
));
2397 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2398 * check to prevent usage of the arc_l2c_only list.
2400 if (((cnt
= refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
)) == 0) &&
2401 (state
!= arc_anon
)) {
2402 multilist_insert(state
->arcs_list
[arc_buf_type(hdr
)], hdr
);
2403 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
2404 arc_evictable_space_increment(hdr
, state
);
2410 * Returns detailed information about a specific arc buffer. When the
2411 * state_index argument is set the function will calculate the arc header
2412 * list position for its arc state. Since this requires a linear traversal
2413 * callers are strongly encourage not to do this. However, it can be helpful
2414 * for targeted analysis so the functionality is provided.
2417 arc_buf_info(arc_buf_t
*ab
, arc_buf_info_t
*abi
, int state_index
)
2419 arc_buf_hdr_t
*hdr
= ab
->b_hdr
;
2420 l1arc_buf_hdr_t
*l1hdr
= NULL
;
2421 l2arc_buf_hdr_t
*l2hdr
= NULL
;
2422 arc_state_t
*state
= NULL
;
2424 memset(abi
, 0, sizeof (arc_buf_info_t
));
2429 abi
->abi_flags
= hdr
->b_flags
;
2431 if (HDR_HAS_L1HDR(hdr
)) {
2432 l1hdr
= &hdr
->b_l1hdr
;
2433 state
= l1hdr
->b_state
;
2435 if (HDR_HAS_L2HDR(hdr
))
2436 l2hdr
= &hdr
->b_l2hdr
;
2439 abi
->abi_bufcnt
= l1hdr
->b_bufcnt
;
2440 abi
->abi_access
= l1hdr
->b_arc_access
;
2441 abi
->abi_mru_hits
= l1hdr
->b_mru_hits
;
2442 abi
->abi_mru_ghost_hits
= l1hdr
->b_mru_ghost_hits
;
2443 abi
->abi_mfu_hits
= l1hdr
->b_mfu_hits
;
2444 abi
->abi_mfu_ghost_hits
= l1hdr
->b_mfu_ghost_hits
;
2445 abi
->abi_holds
= refcount_count(&l1hdr
->b_refcnt
);
2449 abi
->abi_l2arc_dattr
= l2hdr
->b_daddr
;
2450 abi
->abi_l2arc_hits
= l2hdr
->b_hits
;
2453 abi
->abi_state_type
= state
? state
->arcs_state
: ARC_STATE_ANON
;
2454 abi
->abi_state_contents
= arc_buf_type(hdr
);
2455 abi
->abi_size
= arc_hdr_size(hdr
);
2459 * Move the supplied buffer to the indicated state. The hash lock
2460 * for the buffer must be held by the caller.
2463 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*hdr
,
2464 kmutex_t
*hash_lock
)
2466 arc_state_t
*old_state
;
2469 boolean_t update_old
, update_new
;
2470 arc_buf_contents_t buftype
= arc_buf_type(hdr
);
2473 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2474 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2475 * L1 hdr doesn't always exist when we change state to arc_anon before
2476 * destroying a header, in which case reallocating to add the L1 hdr is
2479 if (HDR_HAS_L1HDR(hdr
)) {
2480 old_state
= hdr
->b_l1hdr
.b_state
;
2481 refcnt
= refcount_count(&hdr
->b_l1hdr
.b_refcnt
);
2482 bufcnt
= hdr
->b_l1hdr
.b_bufcnt
;
2483 update_old
= (bufcnt
> 0 || hdr
->b_l1hdr
.b_pabd
!= NULL
||
2486 old_state
= arc_l2c_only
;
2489 update_old
= B_FALSE
;
2491 update_new
= update_old
;
2493 ASSERT(MUTEX_HELD(hash_lock
));
2494 ASSERT3P(new_state
, !=, old_state
);
2495 ASSERT(!GHOST_STATE(new_state
) || bufcnt
== 0);
2496 ASSERT(old_state
!= arc_anon
|| bufcnt
<= 1);
2499 * If this buffer is evictable, transfer it from the
2500 * old state list to the new state list.
2503 if (old_state
!= arc_anon
&& old_state
!= arc_l2c_only
) {
2504 ASSERT(HDR_HAS_L1HDR(hdr
));
2505 multilist_remove(old_state
->arcs_list
[buftype
], hdr
);
2507 if (GHOST_STATE(old_state
)) {
2509 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2510 update_old
= B_TRUE
;
2512 arc_evictable_space_decrement(hdr
, old_state
);
2514 if (new_state
!= arc_anon
&& new_state
!= arc_l2c_only
) {
2516 * An L1 header always exists here, since if we're
2517 * moving to some L1-cached state (i.e. not l2c_only or
2518 * anonymous), we realloc the header to add an L1hdr
2521 ASSERT(HDR_HAS_L1HDR(hdr
));
2522 multilist_insert(new_state
->arcs_list
[buftype
], hdr
);
2524 if (GHOST_STATE(new_state
)) {
2526 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2527 update_new
= B_TRUE
;
2529 arc_evictable_space_increment(hdr
, new_state
);
2533 ASSERT(!HDR_EMPTY(hdr
));
2534 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(hdr
))
2535 buf_hash_remove(hdr
);
2537 /* adjust state sizes (ignore arc_l2c_only) */
2539 if (update_new
&& new_state
!= arc_l2c_only
) {
2540 ASSERT(HDR_HAS_L1HDR(hdr
));
2541 if (GHOST_STATE(new_state
)) {
2545 * When moving a header to a ghost state, we first
2546 * remove all arc buffers. Thus, we'll have a
2547 * bufcnt of zero, and no arc buffer to use for
2548 * the reference. As a result, we use the arc
2549 * header pointer for the reference.
2551 (void) refcount_add_many(&new_state
->arcs_size
,
2552 HDR_GET_LSIZE(hdr
), hdr
);
2553 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2554 ASSERT(!HDR_HAS_RABD(hdr
));
2556 uint32_t buffers
= 0;
2559 * Each individual buffer holds a unique reference,
2560 * thus we must remove each of these references one
2563 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2564 buf
= buf
->b_next
) {
2565 ASSERT3U(bufcnt
, !=, 0);
2569 * When the arc_buf_t is sharing the data
2570 * block with the hdr, the owner of the
2571 * reference belongs to the hdr. Only
2572 * add to the refcount if the arc_buf_t is
2575 if (arc_buf_is_shared(buf
))
2578 (void) refcount_add_many(&new_state
->arcs_size
,
2579 arc_buf_size(buf
), buf
);
2581 ASSERT3U(bufcnt
, ==, buffers
);
2583 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2584 (void) refcount_add_many(&new_state
->arcs_size
,
2585 arc_hdr_size(hdr
), hdr
);
2588 if (HDR_HAS_RABD(hdr
)) {
2589 (void) refcount_add_many(&new_state
->arcs_size
,
2590 HDR_GET_PSIZE(hdr
), hdr
);
2595 if (update_old
&& old_state
!= arc_l2c_only
) {
2596 ASSERT(HDR_HAS_L1HDR(hdr
));
2597 if (GHOST_STATE(old_state
)) {
2599 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2600 ASSERT(!HDR_HAS_RABD(hdr
));
2603 * When moving a header off of a ghost state,
2604 * the header will not contain any arc buffers.
2605 * We use the arc header pointer for the reference
2606 * which is exactly what we did when we put the
2607 * header on the ghost state.
2610 (void) refcount_remove_many(&old_state
->arcs_size
,
2611 HDR_GET_LSIZE(hdr
), hdr
);
2613 uint32_t buffers
= 0;
2616 * Each individual buffer holds a unique reference,
2617 * thus we must remove each of these references one
2620 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2621 buf
= buf
->b_next
) {
2622 ASSERT3U(bufcnt
, !=, 0);
2626 * When the arc_buf_t is sharing the data
2627 * block with the hdr, the owner of the
2628 * reference belongs to the hdr. Only
2629 * add to the refcount if the arc_buf_t is
2632 if (arc_buf_is_shared(buf
))
2635 (void) refcount_remove_many(
2636 &old_state
->arcs_size
, arc_buf_size(buf
),
2639 ASSERT3U(bufcnt
, ==, buffers
);
2640 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
||
2643 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2644 (void) refcount_remove_many(
2645 &old_state
->arcs_size
, arc_hdr_size(hdr
),
2649 if (HDR_HAS_RABD(hdr
)) {
2650 (void) refcount_remove_many(
2651 &old_state
->arcs_size
, HDR_GET_PSIZE(hdr
),
2657 if (HDR_HAS_L1HDR(hdr
))
2658 hdr
->b_l1hdr
.b_state
= new_state
;
2661 * L2 headers should never be on the L2 state list since they don't
2662 * have L1 headers allocated.
2664 ASSERT(multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]) &&
2665 multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]));
2669 arc_space_consume(uint64_t space
, arc_space_type_t type
)
2671 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2676 case ARC_SPACE_DATA
:
2677 ARCSTAT_INCR(arcstat_data_size
, space
);
2679 case ARC_SPACE_META
:
2680 ARCSTAT_INCR(arcstat_metadata_size
, space
);
2682 case ARC_SPACE_BONUS
:
2683 ARCSTAT_INCR(arcstat_bonus_size
, space
);
2685 case ARC_SPACE_DNODE
:
2686 ARCSTAT_INCR(arcstat_dnode_size
, space
);
2688 case ARC_SPACE_DBUF
:
2689 ARCSTAT_INCR(arcstat_dbuf_size
, space
);
2691 case ARC_SPACE_HDRS
:
2692 ARCSTAT_INCR(arcstat_hdr_size
, space
);
2694 case ARC_SPACE_L2HDRS
:
2695 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
2699 if (type
!= ARC_SPACE_DATA
)
2700 ARCSTAT_INCR(arcstat_meta_used
, space
);
2702 atomic_add_64(&arc_size
, space
);
2706 arc_space_return(uint64_t space
, arc_space_type_t type
)
2708 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2713 case ARC_SPACE_DATA
:
2714 ARCSTAT_INCR(arcstat_data_size
, -space
);
2716 case ARC_SPACE_META
:
2717 ARCSTAT_INCR(arcstat_metadata_size
, -space
);
2719 case ARC_SPACE_BONUS
:
2720 ARCSTAT_INCR(arcstat_bonus_size
, -space
);
2722 case ARC_SPACE_DNODE
:
2723 ARCSTAT_INCR(arcstat_dnode_size
, -space
);
2725 case ARC_SPACE_DBUF
:
2726 ARCSTAT_INCR(arcstat_dbuf_size
, -space
);
2728 case ARC_SPACE_HDRS
:
2729 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
2731 case ARC_SPACE_L2HDRS
:
2732 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
2736 if (type
!= ARC_SPACE_DATA
) {
2737 ASSERT(arc_meta_used
>= space
);
2738 if (arc_meta_max
< arc_meta_used
)
2739 arc_meta_max
= arc_meta_used
;
2740 ARCSTAT_INCR(arcstat_meta_used
, -space
);
2743 ASSERT(arc_size
>= space
);
2744 atomic_add_64(&arc_size
, -space
);
2748 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2749 * with the hdr's b_pabd.
2752 arc_can_share(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2755 * The criteria for sharing a hdr's data are:
2756 * 1. the buffer is not encrypted
2757 * 2. the hdr's compression matches the buf's compression
2758 * 3. the hdr doesn't need to be byteswapped
2759 * 4. the hdr isn't already being shared
2760 * 5. the buf is either compressed or it is the last buf in the hdr list
2762 * Criterion #5 maintains the invariant that shared uncompressed
2763 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2764 * might ask, "if a compressed buf is allocated first, won't that be the
2765 * last thing in the list?", but in that case it's impossible to create
2766 * a shared uncompressed buf anyway (because the hdr must be compressed
2767 * to have the compressed buf). You might also think that #3 is
2768 * sufficient to make this guarantee, however it's possible
2769 * (specifically in the rare L2ARC write race mentioned in
2770 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2771 * is sharable, but wasn't at the time of its allocation. Rather than
2772 * allow a new shared uncompressed buf to be created and then shuffle
2773 * the list around to make it the last element, this simply disallows
2774 * sharing if the new buf isn't the first to be added.
2776 ASSERT3P(buf
->b_hdr
, ==, hdr
);
2777 boolean_t hdr_compressed
=
2778 arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
;
2779 boolean_t buf_compressed
= ARC_BUF_COMPRESSED(buf
) != 0;
2780 return (!ARC_BUF_ENCRYPTED(buf
) &&
2781 buf_compressed
== hdr_compressed
&&
2782 hdr
->b_l1hdr
.b_byteswap
== DMU_BSWAP_NUMFUNCS
&&
2783 !HDR_SHARED_DATA(hdr
) &&
2784 (ARC_BUF_LAST(buf
) || ARC_BUF_COMPRESSED(buf
)));
2788 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2789 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2790 * copy was made successfully, or an error code otherwise.
2793 arc_buf_alloc_impl(arc_buf_hdr_t
*hdr
, spa_t
*spa
, uint64_t dsobj
, void *tag
,
2794 boolean_t encrypted
, boolean_t compressed
, boolean_t noauth
,
2795 boolean_t fill
, arc_buf_t
**ret
)
2798 arc_fill_flags_t flags
= ARC_FILL_LOCKED
;
2800 ASSERT(HDR_HAS_L1HDR(hdr
));
2801 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2802 VERIFY(hdr
->b_type
== ARC_BUFC_DATA
||
2803 hdr
->b_type
== ARC_BUFC_METADATA
);
2804 ASSERT3P(ret
, !=, NULL
);
2805 ASSERT3P(*ret
, ==, NULL
);
2806 IMPLY(encrypted
, compressed
);
2808 hdr
->b_l1hdr
.b_mru_hits
= 0;
2809 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
2810 hdr
->b_l1hdr
.b_mfu_hits
= 0;
2811 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
2812 hdr
->b_l1hdr
.b_l2_hits
= 0;
2814 buf
= *ret
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
2817 buf
->b_next
= hdr
->b_l1hdr
.b_buf
;
2820 add_reference(hdr
, tag
);
2823 * We're about to change the hdr's b_flags. We must either
2824 * hold the hash_lock or be undiscoverable.
2826 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2829 * Only honor requests for compressed bufs if the hdr is actually
2830 * compressed. This must be overriden if the buffer is encrypted since
2831 * encrypted buffers cannot be decompressed.
2834 buf
->b_flags
|= ARC_BUF_FLAG_COMPRESSED
;
2835 buf
->b_flags
|= ARC_BUF_FLAG_ENCRYPTED
;
2836 flags
|= ARC_FILL_COMPRESSED
| ARC_FILL_ENCRYPTED
;
2837 } else if (compressed
&&
2838 arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
) {
2839 buf
->b_flags
|= ARC_BUF_FLAG_COMPRESSED
;
2840 flags
|= ARC_FILL_COMPRESSED
;
2845 flags
|= ARC_FILL_NOAUTH
;
2849 * If the hdr's data can be shared then we share the data buffer and
2850 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2851 * allocate a new buffer to store the buf's data.
2853 * There are two additional restrictions here because we're sharing
2854 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2855 * actively involved in an L2ARC write, because if this buf is used by
2856 * an arc_write() then the hdr's data buffer will be released when the
2857 * write completes, even though the L2ARC write might still be using it.
2858 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2859 * need to be ABD-aware.
2861 boolean_t can_share
= arc_can_share(hdr
, buf
) && !HDR_L2_WRITING(hdr
) &&
2862 hdr
->b_l1hdr
.b_pabd
!= NULL
&& abd_is_linear(hdr
->b_l1hdr
.b_pabd
);
2864 /* Set up b_data and sharing */
2866 buf
->b_data
= abd_to_buf(hdr
->b_l1hdr
.b_pabd
);
2867 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2868 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2871 arc_get_data_buf(hdr
, arc_buf_size(buf
), buf
);
2872 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2874 VERIFY3P(buf
->b_data
, !=, NULL
);
2876 hdr
->b_l1hdr
.b_buf
= buf
;
2877 hdr
->b_l1hdr
.b_bufcnt
+= 1;
2879 hdr
->b_crypt_hdr
.b_ebufcnt
+= 1;
2882 * If the user wants the data from the hdr, we need to either copy or
2883 * decompress the data.
2886 return (arc_buf_fill(buf
, spa
, dsobj
, flags
));
2892 static char *arc_onloan_tag
= "onloan";
2895 arc_loaned_bytes_update(int64_t delta
)
2897 atomic_add_64(&arc_loaned_bytes
, delta
);
2899 /* assert that it did not wrap around */
2900 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
2904 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2905 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2906 * buffers must be returned to the arc before they can be used by the DMU or
2910 arc_loan_buf(spa_t
*spa
, boolean_t is_metadata
, int size
)
2912 arc_buf_t
*buf
= arc_alloc_buf(spa
, arc_onloan_tag
,
2913 is_metadata
? ARC_BUFC_METADATA
: ARC_BUFC_DATA
, size
);
2915 arc_loaned_bytes_update(size
);
2921 arc_loan_compressed_buf(spa_t
*spa
, uint64_t psize
, uint64_t lsize
,
2922 enum zio_compress compression_type
)
2924 arc_buf_t
*buf
= arc_alloc_compressed_buf(spa
, arc_onloan_tag
,
2925 psize
, lsize
, compression_type
);
2927 arc_loaned_bytes_update(psize
);
2933 arc_loan_raw_buf(spa_t
*spa
, uint64_t dsobj
, boolean_t byteorder
,
2934 const uint8_t *salt
, const uint8_t *iv
, const uint8_t *mac
,
2935 dmu_object_type_t ot
, uint64_t psize
, uint64_t lsize
,
2936 enum zio_compress compression_type
)
2938 arc_buf_t
*buf
= arc_alloc_raw_buf(spa
, arc_onloan_tag
, dsobj
,
2939 byteorder
, salt
, iv
, mac
, ot
, psize
, lsize
, compression_type
);
2941 atomic_add_64(&arc_loaned_bytes
, psize
);
2947 * Return a loaned arc buffer to the arc.
2950 arc_return_buf(arc_buf_t
*buf
, void *tag
)
2952 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2954 ASSERT3P(buf
->b_data
, !=, NULL
);
2955 ASSERT(HDR_HAS_L1HDR(hdr
));
2956 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2957 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2959 arc_loaned_bytes_update(-arc_buf_size(buf
));
2962 /* Detach an arc_buf from a dbuf (tag) */
2964 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
2966 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2968 ASSERT3P(buf
->b_data
, !=, NULL
);
2969 ASSERT(HDR_HAS_L1HDR(hdr
));
2970 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2971 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2973 arc_loaned_bytes_update(arc_buf_size(buf
));
2977 l2arc_free_abd_on_write(abd_t
*abd
, size_t size
, arc_buf_contents_t type
)
2979 l2arc_data_free_t
*df
= kmem_alloc(sizeof (*df
), KM_SLEEP
);
2982 df
->l2df_size
= size
;
2983 df
->l2df_type
= type
;
2984 mutex_enter(&l2arc_free_on_write_mtx
);
2985 list_insert_head(l2arc_free_on_write
, df
);
2986 mutex_exit(&l2arc_free_on_write_mtx
);
2990 arc_hdr_free_on_write(arc_buf_hdr_t
*hdr
, boolean_t free_rdata
)
2992 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2993 arc_buf_contents_t type
= arc_buf_type(hdr
);
2994 uint64_t size
= (free_rdata
) ? HDR_GET_PSIZE(hdr
) : arc_hdr_size(hdr
);
2996 /* protected by hash lock, if in the hash table */
2997 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
2998 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2999 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
3001 (void) refcount_remove_many(&state
->arcs_esize
[type
],
3004 (void) refcount_remove_many(&state
->arcs_size
, size
, hdr
);
3005 if (type
== ARC_BUFC_METADATA
) {
3006 arc_space_return(size
, ARC_SPACE_META
);
3008 ASSERT(type
== ARC_BUFC_DATA
);
3009 arc_space_return(size
, ARC_SPACE_DATA
);
3013 l2arc_free_abd_on_write(hdr
->b_crypt_hdr
.b_rabd
, size
, type
);
3015 l2arc_free_abd_on_write(hdr
->b_l1hdr
.b_pabd
, size
, type
);
3020 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
3021 * data buffer, we transfer the refcount ownership to the hdr and update
3022 * the appropriate kstats.
3025 arc_share_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
3027 ASSERT(arc_can_share(hdr
, buf
));
3028 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
3029 ASSERT(!ARC_BUF_ENCRYPTED(buf
));
3030 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
3033 * Start sharing the data buffer. We transfer the
3034 * refcount ownership to the hdr since it always owns
3035 * the refcount whenever an arc_buf_t is shared.
3037 refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, buf
, hdr
);
3038 hdr
->b_l1hdr
.b_pabd
= abd_get_from_buf(buf
->b_data
, arc_buf_size(buf
));
3039 abd_take_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
,
3040 HDR_ISTYPE_METADATA(hdr
));
3041 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
3042 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
3045 * Since we've transferred ownership to the hdr we need
3046 * to increment its compressed and uncompressed kstats and
3047 * decrement the overhead size.
3049 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
3050 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
3051 ARCSTAT_INCR(arcstat_overhead_size
, -arc_buf_size(buf
));
3055 arc_unshare_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
3057 ASSERT(arc_buf_is_shared(buf
));
3058 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
3059 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
3062 * We are no longer sharing this buffer so we need
3063 * to transfer its ownership to the rightful owner.
3065 refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, hdr
, buf
);
3066 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
3067 abd_release_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
);
3068 abd_put(hdr
->b_l1hdr
.b_pabd
);
3069 hdr
->b_l1hdr
.b_pabd
= NULL
;
3070 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
3073 * Since the buffer is no longer shared between
3074 * the arc buf and the hdr, count it as overhead.
3076 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
3077 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
3078 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
3082 * Remove an arc_buf_t from the hdr's buf list and return the last
3083 * arc_buf_t on the list. If no buffers remain on the list then return
3087 arc_buf_remove(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
3089 ASSERT(HDR_HAS_L1HDR(hdr
));
3090 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
3092 arc_buf_t
**bufp
= &hdr
->b_l1hdr
.b_buf
;
3093 arc_buf_t
*lastbuf
= NULL
;
3096 * Remove the buf from the hdr list and locate the last
3097 * remaining buffer on the list.
3099 while (*bufp
!= NULL
) {
3101 *bufp
= buf
->b_next
;
3104 * If we've removed a buffer in the middle of
3105 * the list then update the lastbuf and update
3108 if (*bufp
!= NULL
) {
3110 bufp
= &(*bufp
)->b_next
;
3114 ASSERT3P(lastbuf
, !=, buf
);
3115 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, lastbuf
!= NULL
);
3116 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, hdr
->b_l1hdr
.b_buf
!= NULL
);
3117 IMPLY(lastbuf
!= NULL
, ARC_BUF_LAST(lastbuf
));
3123 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3127 arc_buf_destroy_impl(arc_buf_t
*buf
)
3129 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3132 * Free up the data associated with the buf but only if we're not
3133 * sharing this with the hdr. If we are sharing it with the hdr, the
3134 * hdr is responsible for doing the free.
3136 if (buf
->b_data
!= NULL
) {
3138 * We're about to change the hdr's b_flags. We must either
3139 * hold the hash_lock or be undiscoverable.
3141 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
3143 arc_cksum_verify(buf
);
3144 arc_buf_unwatch(buf
);
3146 if (arc_buf_is_shared(buf
)) {
3147 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
3149 uint64_t size
= arc_buf_size(buf
);
3150 arc_free_data_buf(hdr
, buf
->b_data
, size
, buf
);
3151 ARCSTAT_INCR(arcstat_overhead_size
, -size
);
3155 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
3156 hdr
->b_l1hdr
.b_bufcnt
-= 1;
3158 if (ARC_BUF_ENCRYPTED(buf
))
3159 hdr
->b_crypt_hdr
.b_ebufcnt
-= 1;
3162 * If we have no more encrypted buffers and we've already
3163 * gotten a copy of the decrypted data we can free b_rabd to
3166 if (hdr
->b_crypt_hdr
.b_ebufcnt
== 0 && HDR_HAS_RABD(hdr
) &&
3167 hdr
->b_l1hdr
.b_pabd
!= NULL
&& !HDR_IO_IN_PROGRESS(hdr
)) {
3168 arc_hdr_free_abd(hdr
, B_TRUE
);
3172 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
3174 if (ARC_BUF_SHARED(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
3176 * If the current arc_buf_t is sharing its data buffer with the
3177 * hdr, then reassign the hdr's b_pabd to share it with the new
3178 * buffer at the end of the list. The shared buffer is always
3179 * the last one on the hdr's buffer list.
3181 * There is an equivalent case for compressed bufs, but since
3182 * they aren't guaranteed to be the last buf in the list and
3183 * that is an exceedingly rare case, we just allow that space be
3184 * wasted temporarily. We must also be careful not to share
3185 * encrypted buffers, since they cannot be shared.
3187 if (lastbuf
!= NULL
&& !ARC_BUF_ENCRYPTED(lastbuf
)) {
3188 /* Only one buf can be shared at once */
3189 VERIFY(!arc_buf_is_shared(lastbuf
));
3190 /* hdr is uncompressed so can't have compressed buf */
3191 VERIFY(!ARC_BUF_COMPRESSED(lastbuf
));
3193 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
3194 arc_hdr_free_abd(hdr
, B_FALSE
);
3197 * We must setup a new shared block between the
3198 * last buffer and the hdr. The data would have
3199 * been allocated by the arc buf so we need to transfer
3200 * ownership to the hdr since it's now being shared.
3202 arc_share_buf(hdr
, lastbuf
);
3204 } else if (HDR_SHARED_DATA(hdr
)) {
3206 * Uncompressed shared buffers are always at the end
3207 * of the list. Compressed buffers don't have the
3208 * same requirements. This makes it hard to
3209 * simply assert that the lastbuf is shared so
3210 * we rely on the hdr's compression flags to determine
3211 * if we have a compressed, shared buffer.
3213 ASSERT3P(lastbuf
, !=, NULL
);
3214 ASSERT(arc_buf_is_shared(lastbuf
) ||
3215 arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
);
3219 * Free the checksum if we're removing the last uncompressed buf from
3222 if (!arc_hdr_has_uncompressed_buf(hdr
)) {
3223 arc_cksum_free(hdr
);
3226 /* clean up the buf */
3228 kmem_cache_free(buf_cache
, buf
);
3232 arc_hdr_alloc_abd(arc_buf_hdr_t
*hdr
, boolean_t alloc_rdata
)
3236 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
3237 ASSERT(HDR_HAS_L1HDR(hdr
));
3238 ASSERT(!HDR_SHARED_DATA(hdr
) || alloc_rdata
);
3239 IMPLY(alloc_rdata
, HDR_PROTECTED(hdr
));
3241 if (hdr
->b_l1hdr
.b_pabd
== NULL
&& !HDR_HAS_RABD(hdr
))
3242 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
3245 size
= HDR_GET_PSIZE(hdr
);
3246 ASSERT3P(hdr
->b_crypt_hdr
.b_rabd
, ==, NULL
);
3247 hdr
->b_crypt_hdr
.b_rabd
= arc_get_data_abd(hdr
, size
, hdr
);
3248 ASSERT3P(hdr
->b_crypt_hdr
.b_rabd
, !=, NULL
);
3249 ARCSTAT_INCR(arcstat_raw_size
, size
);
3251 size
= arc_hdr_size(hdr
);
3252 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
3253 hdr
->b_l1hdr
.b_pabd
= arc_get_data_abd(hdr
, size
, hdr
);
3254 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
3257 ARCSTAT_INCR(arcstat_compressed_size
, size
);
3258 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
3262 arc_hdr_free_abd(arc_buf_hdr_t
*hdr
, boolean_t free_rdata
)
3264 uint64_t size
= (free_rdata
) ? HDR_GET_PSIZE(hdr
) : arc_hdr_size(hdr
);
3266 ASSERT(HDR_HAS_L1HDR(hdr
));
3267 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
|| HDR_HAS_RABD(hdr
));
3268 IMPLY(free_rdata
, HDR_HAS_RABD(hdr
));
3271 * If the hdr is currently being written to the l2arc then
3272 * we defer freeing the data by adding it to the l2arc_free_on_write
3273 * list. The l2arc will free the data once it's finished
3274 * writing it to the l2arc device.
3276 if (HDR_L2_WRITING(hdr
)) {
3277 arc_hdr_free_on_write(hdr
, free_rdata
);
3278 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
3279 } else if (free_rdata
) {
3280 arc_free_data_abd(hdr
, hdr
->b_crypt_hdr
.b_rabd
, size
, hdr
);
3282 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
, size
, hdr
);
3286 hdr
->b_crypt_hdr
.b_rabd
= NULL
;
3287 ARCSTAT_INCR(arcstat_raw_size
, -size
);
3289 hdr
->b_l1hdr
.b_pabd
= NULL
;
3292 if (hdr
->b_l1hdr
.b_pabd
== NULL
&& !HDR_HAS_RABD(hdr
))
3293 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
3295 ARCSTAT_INCR(arcstat_compressed_size
, -size
);
3296 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
3299 static arc_buf_hdr_t
*
3300 arc_hdr_alloc(uint64_t spa
, int32_t psize
, int32_t lsize
,
3301 boolean_t
protected, enum zio_compress compression_type
,
3302 arc_buf_contents_t type
, boolean_t alloc_rdata
)
3306 VERIFY(type
== ARC_BUFC_DATA
|| type
== ARC_BUFC_METADATA
);
3308 hdr
= kmem_cache_alloc(hdr_full_crypt_cache
, KM_PUSHPAGE
);
3310 hdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
3313 ASSERT(HDR_EMPTY(hdr
));
3314 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
3315 HDR_SET_PSIZE(hdr
, psize
);
3316 HDR_SET_LSIZE(hdr
, lsize
);
3320 arc_hdr_set_flags(hdr
, arc_bufc_to_flags(type
) | ARC_FLAG_HAS_L1HDR
);
3321 arc_hdr_set_compress(hdr
, compression_type
);
3323 arc_hdr_set_flags(hdr
, ARC_FLAG_PROTECTED
);
3325 hdr
->b_l1hdr
.b_state
= arc_anon
;
3326 hdr
->b_l1hdr
.b_arc_access
= 0;
3327 hdr
->b_l1hdr
.b_bufcnt
= 0;
3328 hdr
->b_l1hdr
.b_buf
= NULL
;
3331 * Allocate the hdr's buffer. This will contain either
3332 * the compressed or uncompressed data depending on the block
3333 * it references and compressed arc enablement.
3335 arc_hdr_alloc_abd(hdr
, alloc_rdata
);
3336 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3342 * Transition between the two allocation states for the arc_buf_hdr struct.
3343 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3344 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3345 * version is used when a cache buffer is only in the L2ARC in order to reduce
3348 static arc_buf_hdr_t
*
3349 arc_hdr_realloc(arc_buf_hdr_t
*hdr
, kmem_cache_t
*old
, kmem_cache_t
*new)
3351 ASSERT(HDR_HAS_L2HDR(hdr
));
3353 arc_buf_hdr_t
*nhdr
;
3354 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
3356 ASSERT((old
== hdr_full_cache
&& new == hdr_l2only_cache
) ||
3357 (old
== hdr_l2only_cache
&& new == hdr_full_cache
));
3360 * if the caller wanted a new full header and the header is to be
3361 * encrypted we will actually allocate the header from the full crypt
3362 * cache instead. The same applies to freeing from the old cache.
3364 if (HDR_PROTECTED(hdr
) && new == hdr_full_cache
)
3365 new = hdr_full_crypt_cache
;
3366 if (HDR_PROTECTED(hdr
) && old
== hdr_full_cache
)
3367 old
= hdr_full_crypt_cache
;
3369 nhdr
= kmem_cache_alloc(new, KM_PUSHPAGE
);
3371 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
3372 buf_hash_remove(hdr
);
3374 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
3376 if (new == hdr_full_cache
|| new == hdr_full_crypt_cache
) {
3377 arc_hdr_set_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
3379 * arc_access and arc_change_state need to be aware that a
3380 * header has just come out of L2ARC, so we set its state to
3381 * l2c_only even though it's about to change.
3383 nhdr
->b_l1hdr
.b_state
= arc_l2c_only
;
3385 /* Verify previous threads set to NULL before freeing */
3386 ASSERT3P(nhdr
->b_l1hdr
.b_pabd
, ==, NULL
);
3387 ASSERT(!HDR_HAS_RABD(hdr
));
3389 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
3390 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
3391 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
3394 * If we've reached here, We must have been called from
3395 * arc_evict_hdr(), as such we should have already been
3396 * removed from any ghost list we were previously on
3397 * (which protects us from racing with arc_evict_state),
3398 * thus no locking is needed during this check.
3400 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3403 * A buffer must not be moved into the arc_l2c_only
3404 * state if it's not finished being written out to the
3405 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3406 * might try to be accessed, even though it was removed.
3408 VERIFY(!HDR_L2_WRITING(hdr
));
3409 VERIFY3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
3410 ASSERT(!HDR_HAS_RABD(hdr
));
3412 arc_hdr_clear_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
3415 * The header has been reallocated so we need to re-insert it into any
3418 (void) buf_hash_insert(nhdr
, NULL
);
3420 ASSERT(list_link_active(&hdr
->b_l2hdr
.b_l2node
));
3422 mutex_enter(&dev
->l2ad_mtx
);
3425 * We must place the realloc'ed header back into the list at
3426 * the same spot. Otherwise, if it's placed earlier in the list,
3427 * l2arc_write_buffers() could find it during the function's
3428 * write phase, and try to write it out to the l2arc.
3430 list_insert_after(&dev
->l2ad_buflist
, hdr
, nhdr
);
3431 list_remove(&dev
->l2ad_buflist
, hdr
);
3433 mutex_exit(&dev
->l2ad_mtx
);
3436 * Since we're using the pointer address as the tag when
3437 * incrementing and decrementing the l2ad_alloc refcount, we
3438 * must remove the old pointer (that we're about to destroy) and
3439 * add the new pointer to the refcount. Otherwise we'd remove
3440 * the wrong pointer address when calling arc_hdr_destroy() later.
3443 (void) refcount_remove_many(&dev
->l2ad_alloc
, arc_hdr_size(hdr
), hdr
);
3444 (void) refcount_add_many(&dev
->l2ad_alloc
, arc_hdr_size(nhdr
), nhdr
);
3446 buf_discard_identity(hdr
);
3447 kmem_cache_free(old
, hdr
);
3453 * This function allows an L1 header to be reallocated as a crypt
3454 * header and vice versa. If we are going to a crypt header, the
3455 * new fields will be zeroed out.
3457 static arc_buf_hdr_t
*
3458 arc_hdr_realloc_crypt(arc_buf_hdr_t
*hdr
, boolean_t need_crypt
)
3460 arc_buf_hdr_t
*nhdr
;
3462 kmem_cache_t
*ncache
, *ocache
;
3464 ASSERT(HDR_HAS_L1HDR(hdr
));
3465 ASSERT3U(!!HDR_PROTECTED(hdr
), !=, need_crypt
);
3466 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
3467 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3470 ncache
= hdr_full_crypt_cache
;
3471 ocache
= hdr_full_cache
;
3473 ncache
= hdr_full_cache
;
3474 ocache
= hdr_full_crypt_cache
;
3477 nhdr
= kmem_cache_alloc(ncache
, KM_PUSHPAGE
);
3478 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
3479 nhdr
->b_l1hdr
.b_freeze_cksum
= hdr
->b_l1hdr
.b_freeze_cksum
;
3480 nhdr
->b_l1hdr
.b_bufcnt
= hdr
->b_l1hdr
.b_bufcnt
;
3481 nhdr
->b_l1hdr
.b_byteswap
= hdr
->b_l1hdr
.b_byteswap
;
3482 nhdr
->b_l1hdr
.b_state
= hdr
->b_l1hdr
.b_state
;
3483 nhdr
->b_l1hdr
.b_arc_access
= hdr
->b_l1hdr
.b_arc_access
;
3484 nhdr
->b_l1hdr
.b_mru_hits
= hdr
->b_l1hdr
.b_mru_hits
;
3485 nhdr
->b_l1hdr
.b_mru_ghost_hits
= hdr
->b_l1hdr
.b_mru_ghost_hits
;
3486 nhdr
->b_l1hdr
.b_mfu_hits
= hdr
->b_l1hdr
.b_mfu_hits
;
3487 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= hdr
->b_l1hdr
.b_mfu_ghost_hits
;
3488 nhdr
->b_l1hdr
.b_l2_hits
= hdr
->b_l1hdr
.b_l2_hits
;
3489 nhdr
->b_l1hdr
.b_acb
= hdr
->b_l1hdr
.b_acb
;
3490 nhdr
->b_l1hdr
.b_pabd
= hdr
->b_l1hdr
.b_pabd
;
3491 nhdr
->b_l1hdr
.b_buf
= hdr
->b_l1hdr
.b_buf
;
3494 * This refcount_add() exists only to ensure that the individual
3495 * arc buffers always point to a header that is referenced, avoiding
3496 * a small race condition that could trigger ASSERTs.
3498 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, FTAG
);
3500 for (buf
= nhdr
->b_l1hdr
.b_buf
; buf
!= NULL
; buf
= buf
->b_next
) {
3501 mutex_enter(&buf
->b_evict_lock
);
3503 mutex_exit(&buf
->b_evict_lock
);
3506 refcount_transfer(&nhdr
->b_l1hdr
.b_refcnt
, &hdr
->b_l1hdr
.b_refcnt
);
3507 (void) refcount_remove(&nhdr
->b_l1hdr
.b_refcnt
, FTAG
);
3510 arc_hdr_set_flags(nhdr
, ARC_FLAG_PROTECTED
);
3512 arc_hdr_clear_flags(nhdr
, ARC_FLAG_PROTECTED
);
3515 buf_discard_identity(hdr
);
3516 kmem_cache_free(ocache
, hdr
);
3522 * This function is used by the send / receive code to convert a newly
3523 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3524 * is also used to allow the root objset block to be uupdated without altering
3525 * its embedded MACs. Both block types will always be uncompressed so we do not
3526 * have to worry about compression type or psize.
3529 arc_convert_to_raw(arc_buf_t
*buf
, uint64_t dsobj
, boolean_t byteorder
,
3530 dmu_object_type_t ot
, const uint8_t *salt
, const uint8_t *iv
,
3533 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3535 ASSERT(ot
== DMU_OT_DNODE
|| ot
== DMU_OT_OBJSET
);
3536 ASSERT(HDR_HAS_L1HDR(hdr
));
3537 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
3539 buf
->b_flags
|= (ARC_BUF_FLAG_COMPRESSED
| ARC_BUF_FLAG_ENCRYPTED
);
3540 if (!HDR_PROTECTED(hdr
))
3541 hdr
= arc_hdr_realloc_crypt(hdr
, B_TRUE
);
3542 hdr
->b_crypt_hdr
.b_dsobj
= dsobj
;
3543 hdr
->b_crypt_hdr
.b_ot
= ot
;
3544 hdr
->b_l1hdr
.b_byteswap
= (byteorder
== ZFS_HOST_BYTEORDER
) ?
3545 DMU_BSWAP_NUMFUNCS
: DMU_OT_BYTESWAP(ot
);
3546 if (!arc_hdr_has_uncompressed_buf(hdr
))
3547 arc_cksum_free(hdr
);
3550 bcopy(salt
, hdr
->b_crypt_hdr
.b_salt
, ZIO_DATA_SALT_LEN
);
3552 bcopy(iv
, hdr
->b_crypt_hdr
.b_iv
, ZIO_DATA_IV_LEN
);
3554 bcopy(mac
, hdr
->b_crypt_hdr
.b_mac
, ZIO_DATA_MAC_LEN
);
3558 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3559 * The buf is returned thawed since we expect the consumer to modify it.
3562 arc_alloc_buf(spa_t
*spa
, void *tag
, arc_buf_contents_t type
, int32_t size
)
3564 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), size
, size
,
3565 B_FALSE
, ZIO_COMPRESS_OFF
, type
, B_FALSE
);
3566 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
3568 arc_buf_t
*buf
= NULL
;
3569 VERIFY0(arc_buf_alloc_impl(hdr
, spa
, 0, tag
, B_FALSE
, B_FALSE
,
3570 B_FALSE
, B_FALSE
, &buf
));
3577 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3578 * for bufs containing metadata.
3581 arc_alloc_compressed_buf(spa_t
*spa
, void *tag
, uint64_t psize
, uint64_t lsize
,
3582 enum zio_compress compression_type
)
3584 ASSERT3U(lsize
, >, 0);
3585 ASSERT3U(lsize
, >=, psize
);
3586 ASSERT3U(compression_type
, >, ZIO_COMPRESS_OFF
);
3587 ASSERT3U(compression_type
, <, ZIO_COMPRESS_FUNCTIONS
);
3589 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
3590 B_FALSE
, compression_type
, ARC_BUFC_DATA
, B_FALSE
);
3591 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
3593 arc_buf_t
*buf
= NULL
;
3594 VERIFY0(arc_buf_alloc_impl(hdr
, spa
, 0, tag
, B_FALSE
,
3595 B_TRUE
, B_FALSE
, B_FALSE
, &buf
));
3597 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
3599 if (!arc_buf_is_shared(buf
)) {
3601 * To ensure that the hdr has the correct data in it if we call
3602 * arc_untransform() on this buf before it's been written to
3603 * disk, it's easiest if we just set up sharing between the
3606 ASSERT(!abd_is_linear(hdr
->b_l1hdr
.b_pabd
));
3607 arc_hdr_free_abd(hdr
, B_FALSE
);
3608 arc_share_buf(hdr
, buf
);
3615 arc_alloc_raw_buf(spa_t
*spa
, void *tag
, uint64_t dsobj
, boolean_t byteorder
,
3616 const uint8_t *salt
, const uint8_t *iv
, const uint8_t *mac
,
3617 dmu_object_type_t ot
, uint64_t psize
, uint64_t lsize
,
3618 enum zio_compress compression_type
)
3622 arc_buf_contents_t type
= DMU_OT_IS_METADATA(ot
) ?
3623 ARC_BUFC_METADATA
: ARC_BUFC_DATA
;
3625 ASSERT3U(lsize
, >, 0);
3626 ASSERT3U(lsize
, >=, psize
);
3627 ASSERT3U(compression_type
, >=, ZIO_COMPRESS_OFF
);
3628 ASSERT3U(compression_type
, <, ZIO_COMPRESS_FUNCTIONS
);
3630 hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
, B_TRUE
,
3631 compression_type
, type
, B_TRUE
);
3632 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
3634 hdr
->b_crypt_hdr
.b_dsobj
= dsobj
;
3635 hdr
->b_crypt_hdr
.b_ot
= ot
;
3636 hdr
->b_l1hdr
.b_byteswap
= (byteorder
== ZFS_HOST_BYTEORDER
) ?
3637 DMU_BSWAP_NUMFUNCS
: DMU_OT_BYTESWAP(ot
);
3638 bcopy(salt
, hdr
->b_crypt_hdr
.b_salt
, ZIO_DATA_SALT_LEN
);
3639 bcopy(iv
, hdr
->b_crypt_hdr
.b_iv
, ZIO_DATA_IV_LEN
);
3640 bcopy(mac
, hdr
->b_crypt_hdr
.b_mac
, ZIO_DATA_MAC_LEN
);
3643 * This buffer will be considered encrypted even if the ot is not an
3644 * encrypted type. It will become authenticated instead in
3645 * arc_write_ready().
3648 VERIFY0(arc_buf_alloc_impl(hdr
, spa
, dsobj
, tag
, B_TRUE
, B_TRUE
,
3649 B_FALSE
, B_FALSE
, &buf
));
3651 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
3657 arc_hdr_l2hdr_destroy(arc_buf_hdr_t
*hdr
)
3659 l2arc_buf_hdr_t
*l2hdr
= &hdr
->b_l2hdr
;
3660 l2arc_dev_t
*dev
= l2hdr
->b_dev
;
3661 uint64_t psize
= arc_hdr_size(hdr
);
3663 ASSERT(MUTEX_HELD(&dev
->l2ad_mtx
));
3664 ASSERT(HDR_HAS_L2HDR(hdr
));
3666 list_remove(&dev
->l2ad_buflist
, hdr
);
3668 ARCSTAT_INCR(arcstat_l2_psize
, -psize
);
3669 ARCSTAT_INCR(arcstat_l2_lsize
, -HDR_GET_LSIZE(hdr
));
3671 vdev_space_update(dev
->l2ad_vdev
, -psize
, 0, 0);
3673 (void) refcount_remove_many(&dev
->l2ad_alloc
, psize
, hdr
);
3674 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
3678 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
3680 if (HDR_HAS_L1HDR(hdr
)) {
3681 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
||
3682 hdr
->b_l1hdr
.b_bufcnt
> 0);
3683 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3684 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
3686 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3687 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
3689 if (!HDR_EMPTY(hdr
))
3690 buf_discard_identity(hdr
);
3692 if (HDR_HAS_L2HDR(hdr
)) {
3693 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
3694 boolean_t buflist_held
= MUTEX_HELD(&dev
->l2ad_mtx
);
3697 mutex_enter(&dev
->l2ad_mtx
);
3700 * Even though we checked this conditional above, we
3701 * need to check this again now that we have the
3702 * l2ad_mtx. This is because we could be racing with
3703 * another thread calling l2arc_evict() which might have
3704 * destroyed this header's L2 portion as we were waiting
3705 * to acquire the l2ad_mtx. If that happens, we don't
3706 * want to re-destroy the header's L2 portion.
3708 if (HDR_HAS_L2HDR(hdr
))
3709 arc_hdr_l2hdr_destroy(hdr
);
3712 mutex_exit(&dev
->l2ad_mtx
);
3715 if (HDR_HAS_L1HDR(hdr
)) {
3716 arc_cksum_free(hdr
);
3718 while (hdr
->b_l1hdr
.b_buf
!= NULL
)
3719 arc_buf_destroy_impl(hdr
->b_l1hdr
.b_buf
);
3721 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
3722 arc_hdr_free_abd(hdr
, B_FALSE
);
3725 if (HDR_HAS_RABD(hdr
))
3726 arc_hdr_free_abd(hdr
, B_TRUE
);
3729 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
3730 if (HDR_HAS_L1HDR(hdr
)) {
3731 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3732 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
3734 if (!HDR_PROTECTED(hdr
)) {
3735 kmem_cache_free(hdr_full_cache
, hdr
);
3737 kmem_cache_free(hdr_full_crypt_cache
, hdr
);
3740 kmem_cache_free(hdr_l2only_cache
, hdr
);
3745 arc_buf_destroy(arc_buf_t
*buf
, void* tag
)
3747 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3748 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
3750 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
3751 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
3752 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3753 VERIFY0(remove_reference(hdr
, NULL
, tag
));
3754 arc_hdr_destroy(hdr
);
3758 mutex_enter(hash_lock
);
3759 ASSERT3P(hdr
, ==, buf
->b_hdr
);
3760 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
3761 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3762 ASSERT3P(hdr
->b_l1hdr
.b_state
, !=, arc_anon
);
3763 ASSERT3P(buf
->b_data
, !=, NULL
);
3765 (void) remove_reference(hdr
, hash_lock
, tag
);
3766 arc_buf_destroy_impl(buf
);
3767 mutex_exit(hash_lock
);
3771 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3772 * state of the header is dependent on its state prior to entering this
3773 * function. The following transitions are possible:
3775 * - arc_mru -> arc_mru_ghost
3776 * - arc_mfu -> arc_mfu_ghost
3777 * - arc_mru_ghost -> arc_l2c_only
3778 * - arc_mru_ghost -> deleted
3779 * - arc_mfu_ghost -> arc_l2c_only
3780 * - arc_mfu_ghost -> deleted
3783 arc_evict_hdr(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
3785 arc_state_t
*evicted_state
, *state
;
3786 int64_t bytes_evicted
= 0;
3787 int min_lifetime
= HDR_PRESCIENT_PREFETCH(hdr
) ?
3788 arc_min_prescient_prefetch_ms
: arc_min_prefetch_ms
;
3790 ASSERT(MUTEX_HELD(hash_lock
));
3791 ASSERT(HDR_HAS_L1HDR(hdr
));
3793 state
= hdr
->b_l1hdr
.b_state
;
3794 if (GHOST_STATE(state
)) {
3795 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3796 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
3799 * l2arc_write_buffers() relies on a header's L1 portion
3800 * (i.e. its b_pabd field) during it's write phase.
3801 * Thus, we cannot push a header onto the arc_l2c_only
3802 * state (removing its L1 piece) until the header is
3803 * done being written to the l2arc.
3805 if (HDR_HAS_L2HDR(hdr
) && HDR_L2_WRITING(hdr
)) {
3806 ARCSTAT_BUMP(arcstat_evict_l2_skip
);
3807 return (bytes_evicted
);
3810 ARCSTAT_BUMP(arcstat_deleted
);
3811 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3813 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, hdr
);
3815 if (HDR_HAS_L2HDR(hdr
)) {
3816 ASSERT(hdr
->b_l1hdr
.b_pabd
== NULL
);
3817 ASSERT(!HDR_HAS_RABD(hdr
));
3819 * This buffer is cached on the 2nd Level ARC;
3820 * don't destroy the header.
3822 arc_change_state(arc_l2c_only
, hdr
, hash_lock
);
3824 * dropping from L1+L2 cached to L2-only,
3825 * realloc to remove the L1 header.
3827 hdr
= arc_hdr_realloc(hdr
, hdr_full_cache
,
3830 arc_change_state(arc_anon
, hdr
, hash_lock
);
3831 arc_hdr_destroy(hdr
);
3833 return (bytes_evicted
);
3836 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
3837 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
3839 /* prefetch buffers have a minimum lifespan */
3840 if (HDR_IO_IN_PROGRESS(hdr
) ||
3841 ((hdr
->b_flags
& (ARC_FLAG_PREFETCH
| ARC_FLAG_INDIRECT
)) &&
3842 ddi_get_lbolt() - hdr
->b_l1hdr
.b_arc_access
< min_lifetime
* hz
)) {
3843 ARCSTAT_BUMP(arcstat_evict_skip
);
3844 return (bytes_evicted
);
3847 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
3848 while (hdr
->b_l1hdr
.b_buf
) {
3849 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
3850 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
3851 ARCSTAT_BUMP(arcstat_mutex_miss
);
3854 if (buf
->b_data
!= NULL
)
3855 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3856 mutex_exit(&buf
->b_evict_lock
);
3857 arc_buf_destroy_impl(buf
);
3860 if (HDR_HAS_L2HDR(hdr
)) {
3861 ARCSTAT_INCR(arcstat_evict_l2_cached
, HDR_GET_LSIZE(hdr
));
3863 if (l2arc_write_eligible(hdr
->b_spa
, hdr
)) {
3864 ARCSTAT_INCR(arcstat_evict_l2_eligible
,
3865 HDR_GET_LSIZE(hdr
));
3867 ARCSTAT_INCR(arcstat_evict_l2_ineligible
,
3868 HDR_GET_LSIZE(hdr
));
3872 if (hdr
->b_l1hdr
.b_bufcnt
== 0) {
3873 arc_cksum_free(hdr
);
3875 bytes_evicted
+= arc_hdr_size(hdr
);
3878 * If this hdr is being evicted and has a compressed
3879 * buffer then we discard it here before we change states.
3880 * This ensures that the accounting is updated correctly
3881 * in arc_free_data_impl().
3883 if (hdr
->b_l1hdr
.b_pabd
!= NULL
)
3884 arc_hdr_free_abd(hdr
, B_FALSE
);
3886 if (HDR_HAS_RABD(hdr
))
3887 arc_hdr_free_abd(hdr
, B_TRUE
);
3889 arc_change_state(evicted_state
, hdr
, hash_lock
);
3890 ASSERT(HDR_IN_HASH_TABLE(hdr
));
3891 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
3892 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, hdr
);
3895 return (bytes_evicted
);
3899 arc_evict_state_impl(multilist_t
*ml
, int idx
, arc_buf_hdr_t
*marker
,
3900 uint64_t spa
, int64_t bytes
)
3902 multilist_sublist_t
*mls
;
3903 uint64_t bytes_evicted
= 0;
3905 kmutex_t
*hash_lock
;
3906 int evict_count
= 0;
3908 ASSERT3P(marker
, !=, NULL
);
3909 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3911 mls
= multilist_sublist_lock(ml
, idx
);
3913 for (hdr
= multilist_sublist_prev(mls
, marker
); hdr
!= NULL
;
3914 hdr
= multilist_sublist_prev(mls
, marker
)) {
3915 if ((bytes
!= ARC_EVICT_ALL
&& bytes_evicted
>= bytes
) ||
3916 (evict_count
>= zfs_arc_evict_batch_limit
))
3920 * To keep our iteration location, move the marker
3921 * forward. Since we're not holding hdr's hash lock, we
3922 * must be very careful and not remove 'hdr' from the
3923 * sublist. Otherwise, other consumers might mistake the
3924 * 'hdr' as not being on a sublist when they call the
3925 * multilist_link_active() function (they all rely on
3926 * the hash lock protecting concurrent insertions and
3927 * removals). multilist_sublist_move_forward() was
3928 * specifically implemented to ensure this is the case
3929 * (only 'marker' will be removed and re-inserted).
3931 multilist_sublist_move_forward(mls
, marker
);
3934 * The only case where the b_spa field should ever be
3935 * zero, is the marker headers inserted by
3936 * arc_evict_state(). It's possible for multiple threads
3937 * to be calling arc_evict_state() concurrently (e.g.
3938 * dsl_pool_close() and zio_inject_fault()), so we must
3939 * skip any markers we see from these other threads.
3941 if (hdr
->b_spa
== 0)
3944 /* we're only interested in evicting buffers of a certain spa */
3945 if (spa
!= 0 && hdr
->b_spa
!= spa
) {
3946 ARCSTAT_BUMP(arcstat_evict_skip
);
3950 hash_lock
= HDR_LOCK(hdr
);
3953 * We aren't calling this function from any code path
3954 * that would already be holding a hash lock, so we're
3955 * asserting on this assumption to be defensive in case
3956 * this ever changes. Without this check, it would be
3957 * possible to incorrectly increment arcstat_mutex_miss
3958 * below (e.g. if the code changed such that we called
3959 * this function with a hash lock held).
3961 ASSERT(!MUTEX_HELD(hash_lock
));
3963 if (mutex_tryenter(hash_lock
)) {
3964 uint64_t evicted
= arc_evict_hdr(hdr
, hash_lock
);
3965 mutex_exit(hash_lock
);
3967 bytes_evicted
+= evicted
;
3970 * If evicted is zero, arc_evict_hdr() must have
3971 * decided to skip this header, don't increment
3972 * evict_count in this case.
3978 * If arc_size isn't overflowing, signal any
3979 * threads that might happen to be waiting.
3981 * For each header evicted, we wake up a single
3982 * thread. If we used cv_broadcast, we could
3983 * wake up "too many" threads causing arc_size
3984 * to significantly overflow arc_c; since
3985 * arc_get_data_impl() doesn't check for overflow
3986 * when it's woken up (it doesn't because it's
3987 * possible for the ARC to be overflowing while
3988 * full of un-evictable buffers, and the
3989 * function should proceed in this case).
3991 * If threads are left sleeping, due to not
3992 * using cv_broadcast, they will be woken up
3993 * just before arc_reclaim_thread() sleeps.
3995 mutex_enter(&arc_reclaim_lock
);
3996 if (!arc_is_overflowing())
3997 cv_signal(&arc_reclaim_waiters_cv
);
3998 mutex_exit(&arc_reclaim_lock
);
4000 ARCSTAT_BUMP(arcstat_mutex_miss
);
4004 multilist_sublist_unlock(mls
);
4006 return (bytes_evicted
);
4010 * Evict buffers from the given arc state, until we've removed the
4011 * specified number of bytes. Move the removed buffers to the
4012 * appropriate evict state.
4014 * This function makes a "best effort". It skips over any buffers
4015 * it can't get a hash_lock on, and so, may not catch all candidates.
4016 * It may also return without evicting as much space as requested.
4018 * If bytes is specified using the special value ARC_EVICT_ALL, this
4019 * will evict all available (i.e. unlocked and evictable) buffers from
4020 * the given arc state; which is used by arc_flush().
4023 arc_evict_state(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
4024 arc_buf_contents_t type
)
4026 uint64_t total_evicted
= 0;
4027 multilist_t
*ml
= state
->arcs_list
[type
];
4029 arc_buf_hdr_t
**markers
;
4031 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
4033 num_sublists
= multilist_get_num_sublists(ml
);
4036 * If we've tried to evict from each sublist, made some
4037 * progress, but still have not hit the target number of bytes
4038 * to evict, we want to keep trying. The markers allow us to
4039 * pick up where we left off for each individual sublist, rather
4040 * than starting from the tail each time.
4042 markers
= kmem_zalloc(sizeof (*markers
) * num_sublists
, KM_SLEEP
);
4043 for (int i
= 0; i
< num_sublists
; i
++) {
4044 multilist_sublist_t
*mls
;
4046 markers
[i
] = kmem_cache_alloc(hdr_full_cache
, KM_SLEEP
);
4049 * A b_spa of 0 is used to indicate that this header is
4050 * a marker. This fact is used in arc_adjust_type() and
4051 * arc_evict_state_impl().
4053 markers
[i
]->b_spa
= 0;
4055 mls
= multilist_sublist_lock(ml
, i
);
4056 multilist_sublist_insert_tail(mls
, markers
[i
]);
4057 multilist_sublist_unlock(mls
);
4061 * While we haven't hit our target number of bytes to evict, or
4062 * we're evicting all available buffers.
4064 while (total_evicted
< bytes
|| bytes
== ARC_EVICT_ALL
) {
4065 int sublist_idx
= multilist_get_random_index(ml
);
4066 uint64_t scan_evicted
= 0;
4069 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
4070 * Request that 10% of the LRUs be scanned by the superblock
4073 if (type
== ARC_BUFC_DATA
&& arc_dnode_size
> arc_dnode_limit
)
4074 arc_prune_async((arc_dnode_size
- arc_dnode_limit
) /
4075 sizeof (dnode_t
) / zfs_arc_dnode_reduce_percent
);
4078 * Start eviction using a randomly selected sublist,
4079 * this is to try and evenly balance eviction across all
4080 * sublists. Always starting at the same sublist
4081 * (e.g. index 0) would cause evictions to favor certain
4082 * sublists over others.
4084 for (int i
= 0; i
< num_sublists
; i
++) {
4085 uint64_t bytes_remaining
;
4086 uint64_t bytes_evicted
;
4088 if (bytes
== ARC_EVICT_ALL
)
4089 bytes_remaining
= ARC_EVICT_ALL
;
4090 else if (total_evicted
< bytes
)
4091 bytes_remaining
= bytes
- total_evicted
;
4095 bytes_evicted
= arc_evict_state_impl(ml
, sublist_idx
,
4096 markers
[sublist_idx
], spa
, bytes_remaining
);
4098 scan_evicted
+= bytes_evicted
;
4099 total_evicted
+= bytes_evicted
;
4101 /* we've reached the end, wrap to the beginning */
4102 if (++sublist_idx
>= num_sublists
)
4107 * If we didn't evict anything during this scan, we have
4108 * no reason to believe we'll evict more during another
4109 * scan, so break the loop.
4111 if (scan_evicted
== 0) {
4112 /* This isn't possible, let's make that obvious */
4113 ASSERT3S(bytes
, !=, 0);
4116 * When bytes is ARC_EVICT_ALL, the only way to
4117 * break the loop is when scan_evicted is zero.
4118 * In that case, we actually have evicted enough,
4119 * so we don't want to increment the kstat.
4121 if (bytes
!= ARC_EVICT_ALL
) {
4122 ASSERT3S(total_evicted
, <, bytes
);
4123 ARCSTAT_BUMP(arcstat_evict_not_enough
);
4130 for (int i
= 0; i
< num_sublists
; i
++) {
4131 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
4132 multilist_sublist_remove(mls
, markers
[i
]);
4133 multilist_sublist_unlock(mls
);
4135 kmem_cache_free(hdr_full_cache
, markers
[i
]);
4137 kmem_free(markers
, sizeof (*markers
) * num_sublists
);
4139 return (total_evicted
);
4143 * Flush all "evictable" data of the given type from the arc state
4144 * specified. This will not evict any "active" buffers (i.e. referenced).
4146 * When 'retry' is set to B_FALSE, the function will make a single pass
4147 * over the state and evict any buffers that it can. Since it doesn't
4148 * continually retry the eviction, it might end up leaving some buffers
4149 * in the ARC due to lock misses.
4151 * When 'retry' is set to B_TRUE, the function will continually retry the
4152 * eviction until *all* evictable buffers have been removed from the
4153 * state. As a result, if concurrent insertions into the state are
4154 * allowed (e.g. if the ARC isn't shutting down), this function might
4155 * wind up in an infinite loop, continually trying to evict buffers.
4158 arc_flush_state(arc_state_t
*state
, uint64_t spa
, arc_buf_contents_t type
,
4161 uint64_t evicted
= 0;
4163 while (refcount_count(&state
->arcs_esize
[type
]) != 0) {
4164 evicted
+= arc_evict_state(state
, spa
, ARC_EVICT_ALL
, type
);
4174 * Helper function for arc_prune_async() it is responsible for safely
4175 * handling the execution of a registered arc_prune_func_t.
4178 arc_prune_task(void *ptr
)
4180 arc_prune_t
*ap
= (arc_prune_t
*)ptr
;
4181 arc_prune_func_t
*func
= ap
->p_pfunc
;
4184 func(ap
->p_adjust
, ap
->p_private
);
4186 refcount_remove(&ap
->p_refcnt
, func
);
4190 * Notify registered consumers they must drop holds on a portion of the ARC
4191 * buffered they reference. This provides a mechanism to ensure the ARC can
4192 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
4193 * is analogous to dnlc_reduce_cache() but more generic.
4195 * This operation is performed asynchronously so it may be safely called
4196 * in the context of the arc_reclaim_thread(). A reference is taken here
4197 * for each registered arc_prune_t and the arc_prune_task() is responsible
4198 * for releasing it once the registered arc_prune_func_t has completed.
4201 arc_prune_async(int64_t adjust
)
4205 mutex_enter(&arc_prune_mtx
);
4206 for (ap
= list_head(&arc_prune_list
); ap
!= NULL
;
4207 ap
= list_next(&arc_prune_list
, ap
)) {
4209 if (refcount_count(&ap
->p_refcnt
) >= 2)
4212 refcount_add(&ap
->p_refcnt
, ap
->p_pfunc
);
4213 ap
->p_adjust
= adjust
;
4214 if (taskq_dispatch(arc_prune_taskq
, arc_prune_task
,
4215 ap
, TQ_SLEEP
) == TASKQID_INVALID
) {
4216 refcount_remove(&ap
->p_refcnt
, ap
->p_pfunc
);
4219 ARCSTAT_BUMP(arcstat_prune
);
4221 mutex_exit(&arc_prune_mtx
);
4225 * Evict the specified number of bytes from the state specified,
4226 * restricting eviction to the spa and type given. This function
4227 * prevents us from trying to evict more from a state's list than
4228 * is "evictable", and to skip evicting altogether when passed a
4229 * negative value for "bytes". In contrast, arc_evict_state() will
4230 * evict everything it can, when passed a negative value for "bytes".
4233 arc_adjust_impl(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
4234 arc_buf_contents_t type
)
4238 if (bytes
> 0 && refcount_count(&state
->arcs_esize
[type
]) > 0) {
4239 delta
= MIN(refcount_count(&state
->arcs_esize
[type
]), bytes
);
4240 return (arc_evict_state(state
, spa
, delta
, type
));
4247 * The goal of this function is to evict enough meta data buffers from the
4248 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4249 * more complicated than it appears because it is common for data buffers
4250 * to have holds on meta data buffers. In addition, dnode meta data buffers
4251 * will be held by the dnodes in the block preventing them from being freed.
4252 * This means we can't simply traverse the ARC and expect to always find
4253 * enough unheld meta data buffer to release.
4255 * Therefore, this function has been updated to make alternating passes
4256 * over the ARC releasing data buffers and then newly unheld meta data
4257 * buffers. This ensures forward progress is maintained and arc_meta_used
4258 * will decrease. Normally this is sufficient, but if required the ARC
4259 * will call the registered prune callbacks causing dentry and inodes to
4260 * be dropped from the VFS cache. This will make dnode meta data buffers
4261 * available for reclaim.
4264 arc_adjust_meta_balanced(void)
4266 int64_t delta
, prune
= 0, adjustmnt
;
4267 uint64_t total_evicted
= 0;
4268 arc_buf_contents_t type
= ARC_BUFC_DATA
;
4269 int restarts
= MAX(zfs_arc_meta_adjust_restarts
, 0);
4273 * This slightly differs than the way we evict from the mru in
4274 * arc_adjust because we don't have a "target" value (i.e. no
4275 * "meta" arc_p). As a result, I think we can completely
4276 * cannibalize the metadata in the MRU before we evict the
4277 * metadata from the MFU. I think we probably need to implement a
4278 * "metadata arc_p" value to do this properly.
4280 adjustmnt
= arc_meta_used
- arc_meta_limit
;
4282 if (adjustmnt
> 0 && refcount_count(&arc_mru
->arcs_esize
[type
]) > 0) {
4283 delta
= MIN(refcount_count(&arc_mru
->arcs_esize
[type
]),
4285 total_evicted
+= arc_adjust_impl(arc_mru
, 0, delta
, type
);
4290 * We can't afford to recalculate adjustmnt here. If we do,
4291 * new metadata buffers can sneak into the MRU or ANON lists,
4292 * thus penalize the MFU metadata. Although the fudge factor is
4293 * small, it has been empirically shown to be significant for
4294 * certain workloads (e.g. creating many empty directories). As
4295 * such, we use the original calculation for adjustmnt, and
4296 * simply decrement the amount of data evicted from the MRU.
4299 if (adjustmnt
> 0 && refcount_count(&arc_mfu
->arcs_esize
[type
]) > 0) {
4300 delta
= MIN(refcount_count(&arc_mfu
->arcs_esize
[type
]),
4302 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, delta
, type
);
4305 adjustmnt
= arc_meta_used
- arc_meta_limit
;
4307 if (adjustmnt
> 0 &&
4308 refcount_count(&arc_mru_ghost
->arcs_esize
[type
]) > 0) {
4309 delta
= MIN(adjustmnt
,
4310 refcount_count(&arc_mru_ghost
->arcs_esize
[type
]));
4311 total_evicted
+= arc_adjust_impl(arc_mru_ghost
, 0, delta
, type
);
4315 if (adjustmnt
> 0 &&
4316 refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]) > 0) {
4317 delta
= MIN(adjustmnt
,
4318 refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]));
4319 total_evicted
+= arc_adjust_impl(arc_mfu_ghost
, 0, delta
, type
);
4323 * If after attempting to make the requested adjustment to the ARC
4324 * the meta limit is still being exceeded then request that the
4325 * higher layers drop some cached objects which have holds on ARC
4326 * meta buffers. Requests to the upper layers will be made with
4327 * increasingly large scan sizes until the ARC is below the limit.
4329 if (arc_meta_used
> arc_meta_limit
) {
4330 if (type
== ARC_BUFC_DATA
) {
4331 type
= ARC_BUFC_METADATA
;
4333 type
= ARC_BUFC_DATA
;
4335 if (zfs_arc_meta_prune
) {
4336 prune
+= zfs_arc_meta_prune
;
4337 arc_prune_async(prune
);
4346 return (total_evicted
);
4350 * Evict metadata buffers from the cache, such that arc_meta_used is
4351 * capped by the arc_meta_limit tunable.
4354 arc_adjust_meta_only(void)
4356 uint64_t total_evicted
= 0;
4360 * If we're over the meta limit, we want to evict enough
4361 * metadata to get back under the meta limit. We don't want to
4362 * evict so much that we drop the MRU below arc_p, though. If
4363 * we're over the meta limit more than we're over arc_p, we
4364 * evict some from the MRU here, and some from the MFU below.
4366 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
4367 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
4368 refcount_count(&arc_mru
->arcs_size
) - arc_p
));
4370 total_evicted
+= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
4373 * Similar to the above, we want to evict enough bytes to get us
4374 * below the meta limit, but not so much as to drop us below the
4375 * space allotted to the MFU (which is defined as arc_c - arc_p).
4377 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
4378 (int64_t)(refcount_count(&arc_mfu
->arcs_size
) - (arc_c
- arc_p
)));
4380 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
4382 return (total_evicted
);
4386 arc_adjust_meta(void)
4388 if (zfs_arc_meta_strategy
== ARC_STRATEGY_META_ONLY
)
4389 return (arc_adjust_meta_only());
4391 return (arc_adjust_meta_balanced());
4395 * Return the type of the oldest buffer in the given arc state
4397 * This function will select a random sublist of type ARC_BUFC_DATA and
4398 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4399 * is compared, and the type which contains the "older" buffer will be
4402 static arc_buf_contents_t
4403 arc_adjust_type(arc_state_t
*state
)
4405 multilist_t
*data_ml
= state
->arcs_list
[ARC_BUFC_DATA
];
4406 multilist_t
*meta_ml
= state
->arcs_list
[ARC_BUFC_METADATA
];
4407 int data_idx
= multilist_get_random_index(data_ml
);
4408 int meta_idx
= multilist_get_random_index(meta_ml
);
4409 multilist_sublist_t
*data_mls
;
4410 multilist_sublist_t
*meta_mls
;
4411 arc_buf_contents_t type
;
4412 arc_buf_hdr_t
*data_hdr
;
4413 arc_buf_hdr_t
*meta_hdr
;
4416 * We keep the sublist lock until we're finished, to prevent
4417 * the headers from being destroyed via arc_evict_state().
4419 data_mls
= multilist_sublist_lock(data_ml
, data_idx
);
4420 meta_mls
= multilist_sublist_lock(meta_ml
, meta_idx
);
4423 * These two loops are to ensure we skip any markers that
4424 * might be at the tail of the lists due to arc_evict_state().
4427 for (data_hdr
= multilist_sublist_tail(data_mls
); data_hdr
!= NULL
;
4428 data_hdr
= multilist_sublist_prev(data_mls
, data_hdr
)) {
4429 if (data_hdr
->b_spa
!= 0)
4433 for (meta_hdr
= multilist_sublist_tail(meta_mls
); meta_hdr
!= NULL
;
4434 meta_hdr
= multilist_sublist_prev(meta_mls
, meta_hdr
)) {
4435 if (meta_hdr
->b_spa
!= 0)
4439 if (data_hdr
== NULL
&& meta_hdr
== NULL
) {
4440 type
= ARC_BUFC_DATA
;
4441 } else if (data_hdr
== NULL
) {
4442 ASSERT3P(meta_hdr
, !=, NULL
);
4443 type
= ARC_BUFC_METADATA
;
4444 } else if (meta_hdr
== NULL
) {
4445 ASSERT3P(data_hdr
, !=, NULL
);
4446 type
= ARC_BUFC_DATA
;
4448 ASSERT3P(data_hdr
, !=, NULL
);
4449 ASSERT3P(meta_hdr
, !=, NULL
);
4451 /* The headers can't be on the sublist without an L1 header */
4452 ASSERT(HDR_HAS_L1HDR(data_hdr
));
4453 ASSERT(HDR_HAS_L1HDR(meta_hdr
));
4455 if (data_hdr
->b_l1hdr
.b_arc_access
<
4456 meta_hdr
->b_l1hdr
.b_arc_access
) {
4457 type
= ARC_BUFC_DATA
;
4459 type
= ARC_BUFC_METADATA
;
4463 multilist_sublist_unlock(meta_mls
);
4464 multilist_sublist_unlock(data_mls
);
4470 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4475 uint64_t total_evicted
= 0;
4480 * If we're over arc_meta_limit, we want to correct that before
4481 * potentially evicting data buffers below.
4483 total_evicted
+= arc_adjust_meta();
4488 * If we're over the target cache size, we want to evict enough
4489 * from the list to get back to our target size. We don't want
4490 * to evict too much from the MRU, such that it drops below
4491 * arc_p. So, if we're over our target cache size more than
4492 * the MRU is over arc_p, we'll evict enough to get back to
4493 * arc_p here, and then evict more from the MFU below.
4495 target
= MIN((int64_t)(arc_size
- arc_c
),
4496 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
4497 refcount_count(&arc_mru
->arcs_size
) + arc_meta_used
- arc_p
));
4500 * If we're below arc_meta_min, always prefer to evict data.
4501 * Otherwise, try to satisfy the requested number of bytes to
4502 * evict from the type which contains older buffers; in an
4503 * effort to keep newer buffers in the cache regardless of their
4504 * type. If we cannot satisfy the number of bytes from this
4505 * type, spill over into the next type.
4507 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
4508 arc_meta_used
> arc_meta_min
) {
4509 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
4510 total_evicted
+= bytes
;
4513 * If we couldn't evict our target number of bytes from
4514 * metadata, we try to get the rest from data.
4519 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
4521 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
4522 total_evicted
+= bytes
;
4525 * If we couldn't evict our target number of bytes from
4526 * data, we try to get the rest from metadata.
4531 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
4537 * Now that we've tried to evict enough from the MRU to get its
4538 * size back to arc_p, if we're still above the target cache
4539 * size, we evict the rest from the MFU.
4541 target
= arc_size
- arc_c
;
4543 if (arc_adjust_type(arc_mfu
) == ARC_BUFC_METADATA
&&
4544 arc_meta_used
> arc_meta_min
) {
4545 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
4546 total_evicted
+= bytes
;
4549 * If we couldn't evict our target number of bytes from
4550 * metadata, we try to get the rest from data.
4555 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
4557 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
4558 total_evicted
+= bytes
;
4561 * If we couldn't evict our target number of bytes from
4562 * data, we try to get the rest from data.
4567 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
4571 * Adjust ghost lists
4573 * In addition to the above, the ARC also defines target values
4574 * for the ghost lists. The sum of the mru list and mru ghost
4575 * list should never exceed the target size of the cache, and
4576 * the sum of the mru list, mfu list, mru ghost list, and mfu
4577 * ghost list should never exceed twice the target size of the
4578 * cache. The following logic enforces these limits on the ghost
4579 * caches, and evicts from them as needed.
4581 target
= refcount_count(&arc_mru
->arcs_size
) +
4582 refcount_count(&arc_mru_ghost
->arcs_size
) - arc_c
;
4584 bytes
= arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_DATA
);
4585 total_evicted
+= bytes
;
4590 arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_METADATA
);
4593 * We assume the sum of the mru list and mfu list is less than
4594 * or equal to arc_c (we enforced this above), which means we
4595 * can use the simpler of the two equations below:
4597 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4598 * mru ghost + mfu ghost <= arc_c
4600 target
= refcount_count(&arc_mru_ghost
->arcs_size
) +
4601 refcount_count(&arc_mfu_ghost
->arcs_size
) - arc_c
;
4603 bytes
= arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_DATA
);
4604 total_evicted
+= bytes
;
4609 arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_METADATA
);
4611 return (total_evicted
);
4615 arc_flush(spa_t
*spa
, boolean_t retry
)
4620 * If retry is B_TRUE, a spa must not be specified since we have
4621 * no good way to determine if all of a spa's buffers have been
4622 * evicted from an arc state.
4624 ASSERT(!retry
|| spa
== 0);
4627 guid
= spa_load_guid(spa
);
4629 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_DATA
, retry
);
4630 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_METADATA
, retry
);
4632 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_DATA
, retry
);
4633 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_METADATA
, retry
);
4635 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_DATA
, retry
);
4636 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
4638 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_DATA
, retry
);
4639 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
4643 arc_shrink(int64_t to_free
)
4647 if (c
> to_free
&& c
- to_free
> arc_c_min
) {
4648 arc_c
= c
- to_free
;
4649 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
4650 if (arc_c
> arc_size
)
4651 arc_c
= MAX(arc_size
, arc_c_min
);
4653 arc_p
= (arc_c
>> 1);
4654 ASSERT(arc_c
>= arc_c_min
);
4655 ASSERT((int64_t)arc_p
>= 0);
4660 if (arc_size
> arc_c
)
4661 (void) arc_adjust();
4665 * Return maximum amount of memory that we could possibly use. Reduced
4666 * to half of all memory in user space which is primarily used for testing.
4669 arc_all_memory(void)
4672 #ifdef CONFIG_HIGHMEM
4673 return (ptob(totalram_pages
- totalhigh_pages
));
4675 return (ptob(totalram_pages
));
4676 #endif /* CONFIG_HIGHMEM */
4678 return (ptob(physmem
) / 2);
4679 #endif /* _KERNEL */
4683 * Return the amount of memory that is considered free. In user space
4684 * which is primarily used for testing we pretend that free memory ranges
4685 * from 0-20% of all memory.
4688 arc_free_memory(void)
4691 #ifdef CONFIG_HIGHMEM
4694 return (ptob(si
.freeram
- si
.freehigh
));
4696 #ifdef ZFS_GLOBAL_NODE_PAGE_STATE
4697 return (ptob(nr_free_pages() +
4698 global_node_page_state(NR_INACTIVE_FILE
) +
4699 global_node_page_state(NR_INACTIVE_ANON
) +
4700 global_node_page_state(NR_SLAB_RECLAIMABLE
)));
4702 return (ptob(nr_free_pages() +
4703 global_page_state(NR_INACTIVE_FILE
) +
4704 global_page_state(NR_INACTIVE_ANON
) +
4705 global_page_state(NR_SLAB_RECLAIMABLE
)));
4706 #endif /* ZFS_GLOBAL_NODE_PAGE_STATE */
4707 #endif /* CONFIG_HIGHMEM */
4709 return (spa_get_random(arc_all_memory() * 20 / 100));
4710 #endif /* _KERNEL */
4713 typedef enum free_memory_reason_t
{
4718 FMR_PAGES_PP_MAXIMUM
,
4721 } free_memory_reason_t
;
4723 int64_t last_free_memory
;
4724 free_memory_reason_t last_free_reason
;
4728 * Additional reserve of pages for pp_reserve.
4730 int64_t arc_pages_pp_reserve
= 64;
4733 * Additional reserve of pages for swapfs.
4735 int64_t arc_swapfs_reserve
= 64;
4736 #endif /* _KERNEL */
4739 * Return the amount of memory that can be consumed before reclaim will be
4740 * needed. Positive if there is sufficient free memory, negative indicates
4741 * the amount of memory that needs to be freed up.
4744 arc_available_memory(void)
4746 int64_t lowest
= INT64_MAX
;
4747 free_memory_reason_t r
= FMR_UNKNOWN
;
4754 pgcnt_t needfree
= btop(arc_need_free
);
4755 pgcnt_t lotsfree
= btop(arc_sys_free
);
4756 pgcnt_t desfree
= 0;
4757 pgcnt_t freemem
= btop(arc_free_memory());
4761 n
= PAGESIZE
* (-needfree
);
4769 * check that we're out of range of the pageout scanner. It starts to
4770 * schedule paging if freemem is less than lotsfree and needfree.
4771 * lotsfree is the high-water mark for pageout, and needfree is the
4772 * number of needed free pages. We add extra pages here to make sure
4773 * the scanner doesn't start up while we're freeing memory.
4775 n
= PAGESIZE
* (freemem
- lotsfree
- needfree
- desfree
);
4783 * check to make sure that swapfs has enough space so that anon
4784 * reservations can still succeed. anon_resvmem() checks that the
4785 * availrmem is greater than swapfs_minfree, and the number of reserved
4786 * swap pages. We also add a bit of extra here just to prevent
4787 * circumstances from getting really dire.
4789 n
= PAGESIZE
* (availrmem
- swapfs_minfree
- swapfs_reserve
-
4790 desfree
- arc_swapfs_reserve
);
4793 r
= FMR_SWAPFS_MINFREE
;
4797 * Check that we have enough availrmem that memory locking (e.g., via
4798 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4799 * stores the number of pages that cannot be locked; when availrmem
4800 * drops below pages_pp_maximum, page locking mechanisms such as
4801 * page_pp_lock() will fail.)
4803 n
= PAGESIZE
* (availrmem
- pages_pp_maximum
-
4804 arc_pages_pp_reserve
);
4807 r
= FMR_PAGES_PP_MAXIMUM
;
4813 * If we're on a 32-bit platform, it's possible that we'll exhaust the
4814 * kernel heap space before we ever run out of available physical
4815 * memory. Most checks of the size of the heap_area compare against
4816 * tune.t_minarmem, which is the minimum available real memory that we
4817 * can have in the system. However, this is generally fixed at 25 pages
4818 * which is so low that it's useless. In this comparison, we seek to
4819 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4820 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4823 n
= vmem_size(heap_arena
, VMEM_FREE
) -
4824 (vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
) >> 2);
4832 * If zio data pages are being allocated out of a separate heap segment,
4833 * then enforce that the size of available vmem for this arena remains
4834 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4836 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4837 * memory (in the zio_arena) free, which can avoid memory
4838 * fragmentation issues.
4840 if (zio_arena
!= NULL
) {
4841 n
= (int64_t)vmem_size(zio_arena
, VMEM_FREE
) -
4842 (vmem_size(zio_arena
, VMEM_ALLOC
) >>
4843 arc_zio_arena_free_shift
);
4850 /* Every 100 calls, free a small amount */
4851 if (spa_get_random(100) == 0)
4853 #endif /* _KERNEL */
4855 last_free_memory
= lowest
;
4856 last_free_reason
= r
;
4862 * Determine if the system is under memory pressure and is asking
4863 * to reclaim memory. A return value of B_TRUE indicates that the system
4864 * is under memory pressure and that the arc should adjust accordingly.
4867 arc_reclaim_needed(void)
4869 return (arc_available_memory() < 0);
4873 arc_kmem_reap_now(void)
4876 kmem_cache_t
*prev_cache
= NULL
;
4877 kmem_cache_t
*prev_data_cache
= NULL
;
4878 extern kmem_cache_t
*zio_buf_cache
[];
4879 extern kmem_cache_t
*zio_data_buf_cache
[];
4880 extern kmem_cache_t
*range_seg_cache
;
4883 if ((arc_meta_used
>= arc_meta_limit
) && zfs_arc_meta_prune
) {
4885 * We are exceeding our meta-data cache limit.
4886 * Prune some entries to release holds on meta-data.
4888 arc_prune_async(zfs_arc_meta_prune
);
4892 * Reclaim unused memory from all kmem caches.
4898 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
4900 /* reach upper limit of cache size on 32-bit */
4901 if (zio_buf_cache
[i
] == NULL
)
4904 if (zio_buf_cache
[i
] != prev_cache
) {
4905 prev_cache
= zio_buf_cache
[i
];
4906 kmem_cache_reap_now(zio_buf_cache
[i
]);
4908 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
4909 prev_data_cache
= zio_data_buf_cache
[i
];
4910 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
4913 kmem_cache_reap_now(buf_cache
);
4914 kmem_cache_reap_now(hdr_full_cache
);
4915 kmem_cache_reap_now(hdr_l2only_cache
);
4916 kmem_cache_reap_now(range_seg_cache
);
4918 if (zio_arena
!= NULL
) {
4920 * Ask the vmem arena to reclaim unused memory from its
4923 vmem_qcache_reap(zio_arena
);
4928 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4929 * enough data and signal them to proceed. When this happens, the threads in
4930 * arc_get_data_impl() are sleeping while holding the hash lock for their
4931 * particular arc header. Thus, we must be careful to never sleep on a
4932 * hash lock in this thread. This is to prevent the following deadlock:
4934 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4935 * waiting for the reclaim thread to signal it.
4937 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4938 * fails, and goes to sleep forever.
4940 * This possible deadlock is avoided by always acquiring a hash lock
4941 * using mutex_tryenter() from arc_reclaim_thread().
4945 arc_reclaim_thread(void *unused
)
4947 fstrans_cookie_t cookie
= spl_fstrans_mark();
4948 hrtime_t growtime
= 0;
4951 CALLB_CPR_INIT(&cpr
, &arc_reclaim_lock
, callb_generic_cpr
, FTAG
);
4953 mutex_enter(&arc_reclaim_lock
);
4954 while (!arc_reclaim_thread_exit
) {
4955 uint64_t evicted
= 0;
4956 uint64_t need_free
= arc_need_free
;
4957 arc_tuning_update();
4960 * This is necessary in order for the mdb ::arc dcmd to
4961 * show up to date information. Since the ::arc command
4962 * does not call the kstat's update function, without
4963 * this call, the command may show stale stats for the
4964 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4965 * with this change, the data might be up to 1 second
4966 * out of date; but that should suffice. The arc_state_t
4967 * structures can be queried directly if more accurate
4968 * information is needed.
4971 if (arc_ksp
!= NULL
)
4972 arc_ksp
->ks_update(arc_ksp
, KSTAT_READ
);
4974 mutex_exit(&arc_reclaim_lock
);
4977 * We call arc_adjust() before (possibly) calling
4978 * arc_kmem_reap_now(), so that we can wake up
4979 * arc_get_data_buf() sooner.
4981 evicted
= arc_adjust();
4983 int64_t free_memory
= arc_available_memory();
4984 if (free_memory
< 0) {
4986 arc_no_grow
= B_TRUE
;
4990 * Wait at least zfs_grow_retry (default 5) seconds
4991 * before considering growing.
4993 growtime
= gethrtime() + SEC2NSEC(arc_grow_retry
);
4995 arc_kmem_reap_now();
4998 * If we are still low on memory, shrink the ARC
4999 * so that we have arc_shrink_min free space.
5001 free_memory
= arc_available_memory();
5004 (arc_c
>> arc_shrink_shift
) - free_memory
;
5007 to_free
= MAX(to_free
, need_free
);
5009 arc_shrink(to_free
);
5011 } else if (free_memory
< arc_c
>> arc_no_grow_shift
) {
5012 arc_no_grow
= B_TRUE
;
5013 } else if (gethrtime() >= growtime
) {
5014 arc_no_grow
= B_FALSE
;
5017 mutex_enter(&arc_reclaim_lock
);
5020 * If evicted is zero, we couldn't evict anything via
5021 * arc_adjust(). This could be due to hash lock
5022 * collisions, but more likely due to the majority of
5023 * arc buffers being unevictable. Therefore, even if
5024 * arc_size is above arc_c, another pass is unlikely to
5025 * be helpful and could potentially cause us to enter an
5028 if (arc_size
<= arc_c
|| evicted
== 0) {
5030 * We're either no longer overflowing, or we
5031 * can't evict anything more, so we should wake
5032 * up any threads before we go to sleep and remove
5033 * the bytes we were working on from arc_need_free
5034 * since nothing more will be done here.
5036 cv_broadcast(&arc_reclaim_waiters_cv
);
5037 ARCSTAT_INCR(arcstat_need_free
, -need_free
);
5040 * Block until signaled, or after one second (we
5041 * might need to perform arc_kmem_reap_now()
5042 * even if we aren't being signalled)
5044 CALLB_CPR_SAFE_BEGIN(&cpr
);
5045 (void) cv_timedwait_sig_hires(&arc_reclaim_thread_cv
,
5046 &arc_reclaim_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
5047 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_lock
);
5051 arc_reclaim_thread_exit
= B_FALSE
;
5052 cv_broadcast(&arc_reclaim_thread_cv
);
5053 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_lock */
5054 spl_fstrans_unmark(cookie
);
5060 * Determine the amount of memory eligible for eviction contained in the
5061 * ARC. All clean data reported by the ghost lists can always be safely
5062 * evicted. Due to arc_c_min, the same does not hold for all clean data
5063 * contained by the regular mru and mfu lists.
5065 * In the case of the regular mru and mfu lists, we need to report as
5066 * much clean data as possible, such that evicting that same reported
5067 * data will not bring arc_size below arc_c_min. Thus, in certain
5068 * circumstances, the total amount of clean data in the mru and mfu
5069 * lists might not actually be evictable.
5071 * The following two distinct cases are accounted for:
5073 * 1. The sum of the amount of dirty data contained by both the mru and
5074 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5075 * is greater than or equal to arc_c_min.
5076 * (i.e. amount of dirty data >= arc_c_min)
5078 * This is the easy case; all clean data contained by the mru and mfu
5079 * lists is evictable. Evicting all clean data can only drop arc_size
5080 * to the amount of dirty data, which is greater than arc_c_min.
5082 * 2. The sum of the amount of dirty data contained by both the mru and
5083 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5084 * is less than arc_c_min.
5085 * (i.e. arc_c_min > amount of dirty data)
5087 * 2.1. arc_size is greater than or equal arc_c_min.
5088 * (i.e. arc_size >= arc_c_min > amount of dirty data)
5090 * In this case, not all clean data from the regular mru and mfu
5091 * lists is actually evictable; we must leave enough clean data
5092 * to keep arc_size above arc_c_min. Thus, the maximum amount of
5093 * evictable data from the two lists combined, is exactly the
5094 * difference between arc_size and arc_c_min.
5096 * 2.2. arc_size is less than arc_c_min
5097 * (i.e. arc_c_min > arc_size > amount of dirty data)
5099 * In this case, none of the data contained in the mru and mfu
5100 * lists is evictable, even if it's clean. Since arc_size is
5101 * already below arc_c_min, evicting any more would only
5102 * increase this negative difference.
5105 arc_evictable_memory(void)
5107 uint64_t arc_clean
=
5108 refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]) +
5109 refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]) +
5110 refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]) +
5111 refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
5112 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
5115 * Scale reported evictable memory in proportion to page cache, cap
5116 * at specified min/max.
5118 #ifdef ZFS_GLOBAL_NODE_PAGE_STATE
5119 uint64_t min
= (ptob(global_node_page_state(NR_FILE_PAGES
)) / 100) *
5122 uint64_t min
= (ptob(global_page_state(NR_FILE_PAGES
)) / 100) *
5125 min
= MAX(arc_c_min
, MIN(arc_c_max
, min
));
5127 if (arc_dirty
>= min
)
5130 return (MAX((int64_t)arc_size
- (int64_t)min
, 0));
5134 * If sc->nr_to_scan is zero, the caller is requesting a query of the
5135 * number of objects which can potentially be freed. If it is nonzero,
5136 * the request is to free that many objects.
5138 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
5139 * in struct shrinker and also require the shrinker to return the number
5142 * Older kernels require the shrinker to return the number of freeable
5143 * objects following the freeing of nr_to_free.
5145 static spl_shrinker_t
5146 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
5150 /* The arc is considered warm once reclaim has occurred */
5151 if (unlikely(arc_warm
== B_FALSE
))
5154 /* Return the potential number of reclaimable pages */
5155 pages
= btop((int64_t)arc_evictable_memory());
5156 if (sc
->nr_to_scan
== 0)
5159 /* Not allowed to perform filesystem reclaim */
5160 if (!(sc
->gfp_mask
& __GFP_FS
))
5161 return (SHRINK_STOP
);
5163 /* Reclaim in progress */
5164 if (mutex_tryenter(&arc_reclaim_lock
) == 0) {
5165 ARCSTAT_INCR(arcstat_need_free
, ptob(sc
->nr_to_scan
));
5169 mutex_exit(&arc_reclaim_lock
);
5172 * Evict the requested number of pages by shrinking arc_c the
5176 arc_shrink(ptob(sc
->nr_to_scan
));
5177 if (current_is_kswapd())
5178 arc_kmem_reap_now();
5179 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
5180 pages
= MAX((int64_t)pages
-
5181 (int64_t)btop(arc_evictable_memory()), 0);
5183 pages
= btop(arc_evictable_memory());
5186 * We've shrunk what we can, wake up threads.
5188 cv_broadcast(&arc_reclaim_waiters_cv
);
5190 pages
= SHRINK_STOP
;
5193 * When direct reclaim is observed it usually indicates a rapid
5194 * increase in memory pressure. This occurs because the kswapd
5195 * threads were unable to asynchronously keep enough free memory
5196 * available. In this case set arc_no_grow to briefly pause arc
5197 * growth to avoid compounding the memory pressure.
5199 if (current_is_kswapd()) {
5200 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
5202 arc_no_grow
= B_TRUE
;
5203 arc_kmem_reap_now();
5204 ARCSTAT_BUMP(arcstat_memory_direct_count
);
5209 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
5211 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
5212 #endif /* _KERNEL */
5215 * Adapt arc info given the number of bytes we are trying to add and
5216 * the state that we are coming from. This function is only called
5217 * when we are adding new content to the cache.
5220 arc_adapt(int bytes
, arc_state_t
*state
)
5223 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
5224 int64_t mrug_size
= refcount_count(&arc_mru_ghost
->arcs_size
);
5225 int64_t mfug_size
= refcount_count(&arc_mfu_ghost
->arcs_size
);
5227 if (state
== arc_l2c_only
)
5232 * Adapt the target size of the MRU list:
5233 * - if we just hit in the MRU ghost list, then increase
5234 * the target size of the MRU list.
5235 * - if we just hit in the MFU ghost list, then increase
5236 * the target size of the MFU list by decreasing the
5237 * target size of the MRU list.
5239 if (state
== arc_mru_ghost
) {
5240 mult
= (mrug_size
>= mfug_size
) ? 1 : (mfug_size
/ mrug_size
);
5241 if (!zfs_arc_p_dampener_disable
)
5242 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
5244 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
5245 } else if (state
== arc_mfu_ghost
) {
5248 mult
= (mfug_size
>= mrug_size
) ? 1 : (mrug_size
/ mfug_size
);
5249 if (!zfs_arc_p_dampener_disable
)
5250 mult
= MIN(mult
, 10);
5252 delta
= MIN(bytes
* mult
, arc_p
);
5253 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
5255 ASSERT((int64_t)arc_p
>= 0);
5257 if (arc_reclaim_needed()) {
5258 cv_signal(&arc_reclaim_thread_cv
);
5265 if (arc_c
>= arc_c_max
)
5269 * If we're within (2 * maxblocksize) bytes of the target
5270 * cache size, increment the target cache size
5272 ASSERT3U(arc_c
, >=, 2ULL << SPA_MAXBLOCKSHIFT
);
5273 if (arc_size
>= arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
5274 atomic_add_64(&arc_c
, (int64_t)bytes
);
5275 if (arc_c
> arc_c_max
)
5277 else if (state
== arc_anon
)
5278 atomic_add_64(&arc_p
, (int64_t)bytes
);
5282 ASSERT((int64_t)arc_p
>= 0);
5286 * Check if arc_size has grown past our upper threshold, determined by
5287 * zfs_arc_overflow_shift.
5290 arc_is_overflowing(void)
5292 /* Always allow at least one block of overflow */
5293 uint64_t overflow
= MAX(SPA_MAXBLOCKSIZE
,
5294 arc_c
>> zfs_arc_overflow_shift
);
5296 return (arc_size
>= arc_c
+ overflow
);
5300 arc_get_data_abd(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
5302 arc_buf_contents_t type
= arc_buf_type(hdr
);
5304 arc_get_data_impl(hdr
, size
, tag
);
5305 if (type
== ARC_BUFC_METADATA
) {
5306 return (abd_alloc(size
, B_TRUE
));
5308 ASSERT(type
== ARC_BUFC_DATA
);
5309 return (abd_alloc(size
, B_FALSE
));
5314 arc_get_data_buf(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
5316 arc_buf_contents_t type
= arc_buf_type(hdr
);
5318 arc_get_data_impl(hdr
, size
, tag
);
5319 if (type
== ARC_BUFC_METADATA
) {
5320 return (zio_buf_alloc(size
));
5322 ASSERT(type
== ARC_BUFC_DATA
);
5323 return (zio_data_buf_alloc(size
));
5328 * Allocate a block and return it to the caller. If we are hitting the
5329 * hard limit for the cache size, we must sleep, waiting for the eviction
5330 * thread to catch up. If we're past the target size but below the hard
5331 * limit, we'll only signal the reclaim thread and continue on.
5334 arc_get_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
5336 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
5337 arc_buf_contents_t type
= arc_buf_type(hdr
);
5339 arc_adapt(size
, state
);
5342 * If arc_size is currently overflowing, and has grown past our
5343 * upper limit, we must be adding data faster than the evict
5344 * thread can evict. Thus, to ensure we don't compound the
5345 * problem by adding more data and forcing arc_size to grow even
5346 * further past it's target size, we halt and wait for the
5347 * eviction thread to catch up.
5349 * It's also possible that the reclaim thread is unable to evict
5350 * enough buffers to get arc_size below the overflow limit (e.g.
5351 * due to buffers being un-evictable, or hash lock collisions).
5352 * In this case, we want to proceed regardless if we're
5353 * overflowing; thus we don't use a while loop here.
5355 if (arc_is_overflowing()) {
5356 mutex_enter(&arc_reclaim_lock
);
5359 * Now that we've acquired the lock, we may no longer be
5360 * over the overflow limit, lets check.
5362 * We're ignoring the case of spurious wake ups. If that
5363 * were to happen, it'd let this thread consume an ARC
5364 * buffer before it should have (i.e. before we're under
5365 * the overflow limit and were signalled by the reclaim
5366 * thread). As long as that is a rare occurrence, it
5367 * shouldn't cause any harm.
5369 if (arc_is_overflowing()) {
5370 cv_signal(&arc_reclaim_thread_cv
);
5371 cv_wait(&arc_reclaim_waiters_cv
, &arc_reclaim_lock
);
5374 mutex_exit(&arc_reclaim_lock
);
5377 VERIFY3U(hdr
->b_type
, ==, type
);
5378 if (type
== ARC_BUFC_METADATA
) {
5379 arc_space_consume(size
, ARC_SPACE_META
);
5381 arc_space_consume(size
, ARC_SPACE_DATA
);
5385 * Update the state size. Note that ghost states have a
5386 * "ghost size" and so don't need to be updated.
5388 if (!GHOST_STATE(state
)) {
5390 (void) refcount_add_many(&state
->arcs_size
, size
, tag
);
5393 * If this is reached via arc_read, the link is
5394 * protected by the hash lock. If reached via
5395 * arc_buf_alloc, the header should not be accessed by
5396 * any other thread. And, if reached via arc_read_done,
5397 * the hash lock will protect it if it's found in the
5398 * hash table; otherwise no other thread should be
5399 * trying to [add|remove]_reference it.
5401 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
5402 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5403 (void) refcount_add_many(&state
->arcs_esize
[type
],
5408 * If we are growing the cache, and we are adding anonymous
5409 * data, and we have outgrown arc_p, update arc_p
5411 if (arc_size
< arc_c
&& hdr
->b_l1hdr
.b_state
== arc_anon
&&
5412 (refcount_count(&arc_anon
->arcs_size
) +
5413 refcount_count(&arc_mru
->arcs_size
) > arc_p
))
5414 arc_p
= MIN(arc_c
, arc_p
+ size
);
5419 arc_free_data_abd(arc_buf_hdr_t
*hdr
, abd_t
*abd
, uint64_t size
, void *tag
)
5421 arc_free_data_impl(hdr
, size
, tag
);
5426 arc_free_data_buf(arc_buf_hdr_t
*hdr
, void *buf
, uint64_t size
, void *tag
)
5428 arc_buf_contents_t type
= arc_buf_type(hdr
);
5430 arc_free_data_impl(hdr
, size
, tag
);
5431 if (type
== ARC_BUFC_METADATA
) {
5432 zio_buf_free(buf
, size
);
5434 ASSERT(type
== ARC_BUFC_DATA
);
5435 zio_data_buf_free(buf
, size
);
5440 * Free the arc data buffer.
5443 arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
5445 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
5446 arc_buf_contents_t type
= arc_buf_type(hdr
);
5448 /* protected by hash lock, if in the hash table */
5449 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
5450 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5451 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
5453 (void) refcount_remove_many(&state
->arcs_esize
[type
],
5456 (void) refcount_remove_many(&state
->arcs_size
, size
, tag
);
5458 VERIFY3U(hdr
->b_type
, ==, type
);
5459 if (type
== ARC_BUFC_METADATA
) {
5460 arc_space_return(size
, ARC_SPACE_META
);
5462 ASSERT(type
== ARC_BUFC_DATA
);
5463 arc_space_return(size
, ARC_SPACE_DATA
);
5468 * This routine is called whenever a buffer is accessed.
5469 * NOTE: the hash lock is dropped in this function.
5472 arc_access(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
5476 ASSERT(MUTEX_HELD(hash_lock
));
5477 ASSERT(HDR_HAS_L1HDR(hdr
));
5479 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
5481 * This buffer is not in the cache, and does not
5482 * appear in our "ghost" list. Add the new buffer
5486 ASSERT0(hdr
->b_l1hdr
.b_arc_access
);
5487 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
5488 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
5489 arc_change_state(arc_mru
, hdr
, hash_lock
);
5491 } else if (hdr
->b_l1hdr
.b_state
== arc_mru
) {
5492 now
= ddi_get_lbolt();
5495 * If this buffer is here because of a prefetch, then either:
5496 * - clear the flag if this is a "referencing" read
5497 * (any subsequent access will bump this into the MFU state).
5499 * - move the buffer to the head of the list if this is
5500 * another prefetch (to make it less likely to be evicted).
5502 if (HDR_PREFETCH(hdr
) || HDR_PRESCIENT_PREFETCH(hdr
)) {
5503 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
5504 /* link protected by hash lock */
5505 ASSERT(multilist_link_active(
5506 &hdr
->b_l1hdr
.b_arc_node
));
5508 arc_hdr_clear_flags(hdr
,
5510 ARC_FLAG_PRESCIENT_PREFETCH
);
5511 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
5512 ARCSTAT_BUMP(arcstat_mru_hits
);
5514 hdr
->b_l1hdr
.b_arc_access
= now
;
5519 * This buffer has been "accessed" only once so far,
5520 * but it is still in the cache. Move it to the MFU
5523 if (ddi_time_after(now
, hdr
->b_l1hdr
.b_arc_access
+
5526 * More than 125ms have passed since we
5527 * instantiated this buffer. Move it to the
5528 * most frequently used state.
5530 hdr
->b_l1hdr
.b_arc_access
= now
;
5531 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
5532 arc_change_state(arc_mfu
, hdr
, hash_lock
);
5534 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
5535 ARCSTAT_BUMP(arcstat_mru_hits
);
5536 } else if (hdr
->b_l1hdr
.b_state
== arc_mru_ghost
) {
5537 arc_state_t
*new_state
;
5539 * This buffer has been "accessed" recently, but
5540 * was evicted from the cache. Move it to the
5544 if (HDR_PREFETCH(hdr
) || HDR_PRESCIENT_PREFETCH(hdr
)) {
5545 new_state
= arc_mru
;
5546 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0) {
5547 arc_hdr_clear_flags(hdr
,
5549 ARC_FLAG_PRESCIENT_PREFETCH
);
5551 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
5553 new_state
= arc_mfu
;
5554 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
5557 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
5558 arc_change_state(new_state
, hdr
, hash_lock
);
5560 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_ghost_hits
);
5561 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
5562 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu
) {
5564 * This buffer has been accessed more than once and is
5565 * still in the cache. Keep it in the MFU state.
5567 * NOTE: an add_reference() that occurred when we did
5568 * the arc_read() will have kicked this off the list.
5569 * If it was a prefetch, we will explicitly move it to
5570 * the head of the list now.
5573 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_hits
);
5574 ARCSTAT_BUMP(arcstat_mfu_hits
);
5575 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
5576 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu_ghost
) {
5577 arc_state_t
*new_state
= arc_mfu
;
5579 * This buffer has been accessed more than once but has
5580 * been evicted from the cache. Move it back to the
5584 if (HDR_PREFETCH(hdr
) || HDR_PRESCIENT_PREFETCH(hdr
)) {
5586 * This is a prefetch access...
5587 * move this block back to the MRU state.
5589 new_state
= arc_mru
;
5592 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
5593 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
5594 arc_change_state(new_state
, hdr
, hash_lock
);
5596 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_ghost_hits
);
5597 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
5598 } else if (hdr
->b_l1hdr
.b_state
== arc_l2c_only
) {
5600 * This buffer is on the 2nd Level ARC.
5603 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
5604 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
5605 arc_change_state(arc_mfu
, hdr
, hash_lock
);
5607 cmn_err(CE_PANIC
, "invalid arc state 0x%p",
5608 hdr
->b_l1hdr
.b_state
);
5612 /* a generic arc_read_done_func_t which you can use */
5615 arc_bcopy_func(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
5616 arc_buf_t
*buf
, void *arg
)
5621 bcopy(buf
->b_data
, arg
, arc_buf_size(buf
));
5622 arc_buf_destroy(buf
, arg
);
5625 /* a generic arc_read_done_func_t */
5628 arc_getbuf_func(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
5629 arc_buf_t
*buf
, void *arg
)
5631 arc_buf_t
**bufp
= arg
;
5637 ASSERT(buf
->b_data
);
5642 arc_hdr_verify(arc_buf_hdr_t
*hdr
, blkptr_t
*bp
)
5644 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
)) {
5645 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, 0);
5646 ASSERT3U(arc_hdr_get_compress(hdr
), ==, ZIO_COMPRESS_OFF
);
5648 if (HDR_COMPRESSION_ENABLED(hdr
)) {
5649 ASSERT3U(arc_hdr_get_compress(hdr
), ==,
5650 BP_GET_COMPRESS(bp
));
5652 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(bp
));
5653 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, BP_GET_PSIZE(bp
));
5654 ASSERT3U(!!HDR_PROTECTED(hdr
), ==, BP_IS_PROTECTED(bp
));
5659 arc_read_done(zio_t
*zio
)
5661 blkptr_t
*bp
= zio
->io_bp
;
5662 arc_buf_hdr_t
*hdr
= zio
->io_private
;
5663 kmutex_t
*hash_lock
= NULL
;
5664 arc_callback_t
*callback_list
;
5665 arc_callback_t
*acb
;
5666 boolean_t freeable
= B_FALSE
;
5669 * The hdr was inserted into hash-table and removed from lists
5670 * prior to starting I/O. We should find this header, since
5671 * it's in the hash table, and it should be legit since it's
5672 * not possible to evict it during the I/O. The only possible
5673 * reason for it not to be found is if we were freed during the
5676 if (HDR_IN_HASH_TABLE(hdr
)) {
5677 arc_buf_hdr_t
*found
;
5679 ASSERT3U(hdr
->b_birth
, ==, BP_PHYSICAL_BIRTH(zio
->io_bp
));
5680 ASSERT3U(hdr
->b_dva
.dva_word
[0], ==,
5681 BP_IDENTITY(zio
->io_bp
)->dva_word
[0]);
5682 ASSERT3U(hdr
->b_dva
.dva_word
[1], ==,
5683 BP_IDENTITY(zio
->io_bp
)->dva_word
[1]);
5685 found
= buf_hash_find(hdr
->b_spa
, zio
->io_bp
, &hash_lock
);
5687 ASSERT((found
== hdr
&&
5688 DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
5689 (found
== hdr
&& HDR_L2_READING(hdr
)));
5690 ASSERT3P(hash_lock
, !=, NULL
);
5693 if (BP_IS_PROTECTED(bp
)) {
5694 hdr
->b_crypt_hdr
.b_ot
= BP_GET_TYPE(bp
);
5695 hdr
->b_crypt_hdr
.b_dsobj
= zio
->io_bookmark
.zb_objset
;
5696 zio_crypt_decode_params_bp(bp
, hdr
->b_crypt_hdr
.b_salt
,
5697 hdr
->b_crypt_hdr
.b_iv
);
5699 if (BP_GET_TYPE(bp
) == DMU_OT_INTENT_LOG
) {
5702 tmpbuf
= abd_borrow_buf_copy(zio
->io_abd
,
5703 sizeof (zil_chain_t
));
5704 zio_crypt_decode_mac_zil(tmpbuf
,
5705 hdr
->b_crypt_hdr
.b_mac
);
5706 abd_return_buf(zio
->io_abd
, tmpbuf
,
5707 sizeof (zil_chain_t
));
5709 zio_crypt_decode_mac_bp(bp
, hdr
->b_crypt_hdr
.b_mac
);
5713 if (zio
->io_error
== 0) {
5714 /* byteswap if necessary */
5715 if (BP_SHOULD_BYTESWAP(zio
->io_bp
)) {
5716 if (BP_GET_LEVEL(zio
->io_bp
) > 0) {
5717 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_UINT64
;
5719 hdr
->b_l1hdr
.b_byteswap
=
5720 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
5723 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
5727 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_EVICTED
);
5728 if (l2arc_noprefetch
&& HDR_PREFETCH(hdr
))
5729 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2CACHE
);
5731 callback_list
= hdr
->b_l1hdr
.b_acb
;
5732 ASSERT3P(callback_list
, !=, NULL
);
5734 if (hash_lock
&& zio
->io_error
== 0 &&
5735 hdr
->b_l1hdr
.b_state
== arc_anon
) {
5737 * Only call arc_access on anonymous buffers. This is because
5738 * if we've issued an I/O for an evicted buffer, we've already
5739 * called arc_access (to prevent any simultaneous readers from
5740 * getting confused).
5742 arc_access(hdr
, hash_lock
);
5746 * If a read request has a callback (i.e. acb_done is not NULL), then we
5747 * make a buf containing the data according to the parameters which were
5748 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5749 * aren't needlessly decompressing the data multiple times.
5751 int callback_cnt
= 0;
5752 for (acb
= callback_list
; acb
!= NULL
; acb
= acb
->acb_next
) {
5758 if (zio
->io_error
!= 0)
5761 int error
= arc_buf_alloc_impl(hdr
, zio
->io_spa
,
5762 acb
->acb_dsobj
, acb
->acb_private
, acb
->acb_encrypted
,
5763 acb
->acb_compressed
, acb
->acb_noauth
, B_TRUE
,
5766 arc_buf_destroy(acb
->acb_buf
, acb
->acb_private
);
5767 acb
->acb_buf
= NULL
;
5771 * Assert non-speculative zios didn't fail because an
5772 * encryption key wasn't loaded
5774 ASSERT((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) || error
== 0);
5777 * If we failed to decrypt, report an error now (as the zio
5778 * layer would have done if it had done the transforms).
5780 if (error
== ECKSUM
) {
5781 ASSERT(BP_IS_PROTECTED(bp
));
5782 error
= SET_ERROR(EIO
);
5783 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
);
5784 if ((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) == 0) {
5785 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION
,
5786 zio
->io_spa
, NULL
, &zio
->io_bookmark
, zio
,
5791 if (zio
->io_error
== 0)
5792 zio
->io_error
= error
;
5794 hdr
->b_l1hdr
.b_acb
= NULL
;
5795 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5796 if (callback_cnt
== 0)
5797 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
|| HDR_HAS_RABD(hdr
));
5799 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
) ||
5800 callback_list
!= NULL
);
5802 if (zio
->io_error
== 0) {
5803 arc_hdr_verify(hdr
, zio
->io_bp
);
5805 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_ERROR
);
5806 if (hdr
->b_l1hdr
.b_state
!= arc_anon
)
5807 arc_change_state(arc_anon
, hdr
, hash_lock
);
5808 if (HDR_IN_HASH_TABLE(hdr
))
5809 buf_hash_remove(hdr
);
5810 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
5814 * Broadcast before we drop the hash_lock to avoid the possibility
5815 * that the hdr (and hence the cv) might be freed before we get to
5816 * the cv_broadcast().
5818 cv_broadcast(&hdr
->b_l1hdr
.b_cv
);
5820 if (hash_lock
!= NULL
) {
5821 mutex_exit(hash_lock
);
5824 * This block was freed while we waited for the read to
5825 * complete. It has been removed from the hash table and
5826 * moved to the anonymous state (so that it won't show up
5829 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
5830 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
5833 /* execute each callback and free its structure */
5834 while ((acb
= callback_list
) != NULL
) {
5835 if (acb
->acb_done
) {
5836 acb
->acb_done(zio
, &zio
->io_bookmark
, zio
->io_bp
,
5837 acb
->acb_buf
, acb
->acb_private
);
5840 if (acb
->acb_zio_dummy
!= NULL
) {
5841 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
5842 zio_nowait(acb
->acb_zio_dummy
);
5845 callback_list
= acb
->acb_next
;
5846 kmem_free(acb
, sizeof (arc_callback_t
));
5850 arc_hdr_destroy(hdr
);
5854 * "Read" the block at the specified DVA (in bp) via the
5855 * cache. If the block is found in the cache, invoke the provided
5856 * callback immediately and return. Note that the `zio' parameter
5857 * in the callback will be NULL in this case, since no IO was
5858 * required. If the block is not in the cache pass the read request
5859 * on to the spa with a substitute callback function, so that the
5860 * requested block will be added to the cache.
5862 * If a read request arrives for a block that has a read in-progress,
5863 * either wait for the in-progress read to complete (and return the
5864 * results); or, if this is a read with a "done" func, add a record
5865 * to the read to invoke the "done" func when the read completes,
5866 * and return; or just return.
5868 * arc_read_done() will invoke all the requested "done" functions
5869 * for readers of this block.
5872 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
5873 arc_read_done_func_t
*done
, void *private, zio_priority_t priority
,
5874 int zio_flags
, arc_flags_t
*arc_flags
, const zbookmark_phys_t
*zb
)
5876 arc_buf_hdr_t
*hdr
= NULL
;
5877 kmutex_t
*hash_lock
= NULL
;
5879 uint64_t guid
= spa_load_guid(spa
);
5880 boolean_t compressed_read
= (zio_flags
& ZIO_FLAG_RAW_COMPRESS
) != 0;
5881 boolean_t encrypted_read
= BP_IS_ENCRYPTED(bp
) &&
5882 (zio_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0;
5883 boolean_t noauth_read
= BP_IS_AUTHENTICATED(bp
) &&
5884 (zio_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0;
5887 ASSERT(!BP_IS_EMBEDDED(bp
) ||
5888 BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
);
5891 if (!BP_IS_EMBEDDED(bp
)) {
5893 * Embedded BP's have no DVA and require no I/O to "read".
5894 * Create an anonymous arc buf to back it.
5896 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
5900 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5901 * we maintain encrypted data seperately from compressed / uncompressed
5902 * data. If the user is requesting raw encrypted data and we don't have
5903 * that in the header we will read from disk to guarantee that we can
5904 * get it even if the encryption keys aren't loaded.
5906 if (hdr
!= NULL
&& HDR_HAS_L1HDR(hdr
) && (HDR_HAS_RABD(hdr
) ||
5907 (hdr
->b_l1hdr
.b_pabd
!= NULL
&& !encrypted_read
))) {
5908 arc_buf_t
*buf
= NULL
;
5909 *arc_flags
|= ARC_FLAG_CACHED
;
5911 if (HDR_IO_IN_PROGRESS(hdr
)) {
5913 if ((hdr
->b_flags
& ARC_FLAG_PRIO_ASYNC_READ
) &&
5914 priority
== ZIO_PRIORITY_SYNC_READ
) {
5916 * This sync read must wait for an
5917 * in-progress async read (e.g. a predictive
5918 * prefetch). Async reads are queued
5919 * separately at the vdev_queue layer, so
5920 * this is a form of priority inversion.
5921 * Ideally, we would "inherit" the demand
5922 * i/o's priority by moving the i/o from
5923 * the async queue to the synchronous queue,
5924 * but there is currently no mechanism to do
5925 * so. Track this so that we can evaluate
5926 * the magnitude of this potential performance
5929 * Note that if the prefetch i/o is already
5930 * active (has been issued to the device),
5931 * the prefetch improved performance, because
5932 * we issued it sooner than we would have
5933 * without the prefetch.
5935 DTRACE_PROBE1(arc__sync__wait__for__async
,
5936 arc_buf_hdr_t
*, hdr
);
5937 ARCSTAT_BUMP(arcstat_sync_wait_for_async
);
5939 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5940 arc_hdr_clear_flags(hdr
,
5941 ARC_FLAG_PREDICTIVE_PREFETCH
);
5944 if (*arc_flags
& ARC_FLAG_WAIT
) {
5945 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
5946 mutex_exit(hash_lock
);
5949 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5952 arc_callback_t
*acb
= NULL
;
5954 acb
= kmem_zalloc(sizeof (arc_callback_t
),
5956 acb
->acb_done
= done
;
5957 acb
->acb_private
= private;
5958 acb
->acb_compressed
= compressed_read
;
5959 acb
->acb_encrypted
= encrypted_read
;
5960 acb
->acb_noauth
= noauth_read
;
5961 acb
->acb_dsobj
= zb
->zb_objset
;
5963 acb
->acb_zio_dummy
= zio_null(pio
,
5964 spa
, NULL
, NULL
, NULL
, zio_flags
);
5966 ASSERT3P(acb
->acb_done
, !=, NULL
);
5967 acb
->acb_next
= hdr
->b_l1hdr
.b_acb
;
5968 hdr
->b_l1hdr
.b_acb
= acb
;
5969 mutex_exit(hash_lock
);
5972 mutex_exit(hash_lock
);
5976 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
5977 hdr
->b_l1hdr
.b_state
== arc_mfu
);
5980 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5982 * This is a demand read which does not have to
5983 * wait for i/o because we did a predictive
5984 * prefetch i/o for it, which has completed.
5987 arc__demand__hit__predictive__prefetch
,
5988 arc_buf_hdr_t
*, hdr
);
5990 arcstat_demand_hit_predictive_prefetch
);
5991 arc_hdr_clear_flags(hdr
,
5992 ARC_FLAG_PREDICTIVE_PREFETCH
);
5995 if (hdr
->b_flags
& ARC_FLAG_PRESCIENT_PREFETCH
) {
5997 arcstat_demand_hit_prescient_prefetch
);
5998 arc_hdr_clear_flags(hdr
,
5999 ARC_FLAG_PRESCIENT_PREFETCH
);
6002 ASSERT(!BP_IS_EMBEDDED(bp
) || !BP_IS_HOLE(bp
));
6004 /* Get a buf with the desired data in it. */
6005 rc
= arc_buf_alloc_impl(hdr
, spa
, zb
->zb_objset
,
6006 private, encrypted_read
, compressed_read
,
6007 noauth_read
, B_TRUE
, &buf
);
6009 arc_buf_destroy(buf
, private);
6013 ASSERT((zio_flags
& ZIO_FLAG_SPECULATIVE
) || rc
== 0);
6014 } else if (*arc_flags
& ARC_FLAG_PREFETCH
&&
6015 refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
6016 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
6018 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
6019 arc_access(hdr
, hash_lock
);
6020 if (*arc_flags
& ARC_FLAG_PRESCIENT_PREFETCH
)
6021 arc_hdr_set_flags(hdr
, ARC_FLAG_PRESCIENT_PREFETCH
);
6022 if (*arc_flags
& ARC_FLAG_L2CACHE
)
6023 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
6024 mutex_exit(hash_lock
);
6025 ARCSTAT_BUMP(arcstat_hits
);
6026 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
6027 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
6028 data
, metadata
, hits
);
6031 done(NULL
, zb
, bp
, buf
, private);
6033 uint64_t lsize
= BP_GET_LSIZE(bp
);
6034 uint64_t psize
= BP_GET_PSIZE(bp
);
6035 arc_callback_t
*acb
;
6038 boolean_t devw
= B_FALSE
;
6043 * Gracefully handle a damaged logical block size as a
6046 if (lsize
> spa_maxblocksize(spa
)) {
6047 rc
= SET_ERROR(ECKSUM
);
6052 /* this block is not in the cache */
6053 arc_buf_hdr_t
*exists
= NULL
;
6054 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
6055 hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
6056 BP_IS_PROTECTED(bp
), BP_GET_COMPRESS(bp
), type
,
6059 if (!BP_IS_EMBEDDED(bp
)) {
6060 hdr
->b_dva
= *BP_IDENTITY(bp
);
6061 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
6062 exists
= buf_hash_insert(hdr
, &hash_lock
);
6064 if (exists
!= NULL
) {
6065 /* somebody beat us to the hash insert */
6066 mutex_exit(hash_lock
);
6067 buf_discard_identity(hdr
);
6068 arc_hdr_destroy(hdr
);
6069 goto top
; /* restart the IO request */
6073 * This block is in the ghost cache or encrypted data
6074 * was requested and we didn't have it. If it was
6075 * L2-only (and thus didn't have an L1 hdr),
6076 * we realloc the header to add an L1 hdr.
6078 if (!HDR_HAS_L1HDR(hdr
)) {
6079 hdr
= arc_hdr_realloc(hdr
, hdr_l2only_cache
,
6083 if (GHOST_STATE(hdr
->b_l1hdr
.b_state
)) {
6084 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
6085 ASSERT(!HDR_HAS_RABD(hdr
));
6086 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
6087 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
6088 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
6089 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
6090 } else if (HDR_IO_IN_PROGRESS(hdr
)) {
6092 * If this header already had an IO in progress
6093 * and we are performing another IO to fetch
6094 * encrypted data we must wait until the first
6095 * IO completes so as not to confuse
6096 * arc_read_done(). This should be very rare
6097 * and so the performance impact shouldn't
6100 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
6101 mutex_exit(hash_lock
);
6106 * This is a delicate dance that we play here.
6107 * This hdr might be in the ghost list so we access
6108 * it to move it out of the ghost list before we
6109 * initiate the read. If it's a prefetch then
6110 * it won't have a callback so we'll remove the
6111 * reference that arc_buf_alloc_impl() created. We
6112 * do this after we've called arc_access() to
6113 * avoid hitting an assert in remove_reference().
6115 arc_access(hdr
, hash_lock
);
6116 arc_hdr_alloc_abd(hdr
, encrypted_read
);
6119 if (encrypted_read
) {
6120 ASSERT(HDR_HAS_RABD(hdr
));
6121 size
= HDR_GET_PSIZE(hdr
);
6122 hdr_abd
= hdr
->b_crypt_hdr
.b_rabd
;
6123 zio_flags
|= ZIO_FLAG_RAW
;
6125 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
6126 size
= arc_hdr_size(hdr
);
6127 hdr_abd
= hdr
->b_l1hdr
.b_pabd
;
6129 if (arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
) {
6130 zio_flags
|= ZIO_FLAG_RAW_COMPRESS
;
6134 * For authenticated bp's, we do not ask the ZIO layer
6135 * to authenticate them since this will cause the entire
6136 * IO to fail if the key isn't loaded. Instead, we
6137 * defer authentication until arc_buf_fill(), which will
6138 * verify the data when the key is available.
6140 if (BP_IS_AUTHENTICATED(bp
))
6141 zio_flags
|= ZIO_FLAG_RAW_ENCRYPT
;
6144 if (*arc_flags
& ARC_FLAG_PREFETCH
&&
6145 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
))
6146 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
6147 if (*arc_flags
& ARC_FLAG_PRESCIENT_PREFETCH
)
6148 arc_hdr_set_flags(hdr
, ARC_FLAG_PRESCIENT_PREFETCH
);
6149 if (*arc_flags
& ARC_FLAG_L2CACHE
)
6150 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
6151 if (BP_IS_AUTHENTICATED(bp
))
6152 arc_hdr_set_flags(hdr
, ARC_FLAG_NOAUTH
);
6153 if (BP_GET_LEVEL(bp
) > 0)
6154 arc_hdr_set_flags(hdr
, ARC_FLAG_INDIRECT
);
6155 if (*arc_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
)
6156 arc_hdr_set_flags(hdr
, ARC_FLAG_PREDICTIVE_PREFETCH
);
6157 ASSERT(!GHOST_STATE(hdr
->b_l1hdr
.b_state
));
6159 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
6160 acb
->acb_done
= done
;
6161 acb
->acb_private
= private;
6162 acb
->acb_compressed
= compressed_read
;
6163 acb
->acb_encrypted
= encrypted_read
;
6164 acb
->acb_noauth
= noauth_read
;
6165 acb
->acb_dsobj
= zb
->zb_objset
;
6167 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
6168 hdr
->b_l1hdr
.b_acb
= acb
;
6169 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6171 if (HDR_HAS_L2HDR(hdr
) &&
6172 (vd
= hdr
->b_l2hdr
.b_dev
->l2ad_vdev
) != NULL
) {
6173 devw
= hdr
->b_l2hdr
.b_dev
->l2ad_writing
;
6174 addr
= hdr
->b_l2hdr
.b_daddr
;
6176 * Lock out device removal.
6178 if (vdev_is_dead(vd
) ||
6179 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
6183 if (priority
== ZIO_PRIORITY_ASYNC_READ
)
6184 arc_hdr_set_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
6186 arc_hdr_clear_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
6188 if (hash_lock
!= NULL
)
6189 mutex_exit(hash_lock
);
6192 * At this point, we have a level 1 cache miss. Try again in
6193 * L2ARC if possible.
6195 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, lsize
);
6197 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
6198 uint64_t, lsize
, zbookmark_phys_t
*, zb
);
6199 ARCSTAT_BUMP(arcstat_misses
);
6200 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
6201 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
6202 data
, metadata
, misses
);
6204 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
6206 * Read from the L2ARC if the following are true:
6207 * 1. The L2ARC vdev was previously cached.
6208 * 2. This buffer still has L2ARC metadata.
6209 * 3. This buffer isn't currently writing to the L2ARC.
6210 * 4. The L2ARC entry wasn't evicted, which may
6211 * also have invalidated the vdev.
6212 * 5. This isn't prefetch and l2arc_noprefetch is set.
6214 if (HDR_HAS_L2HDR(hdr
) &&
6215 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
6216 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
6217 l2arc_read_callback_t
*cb
;
6221 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
6222 ARCSTAT_BUMP(arcstat_l2_hits
);
6223 atomic_inc_32(&hdr
->b_l2hdr
.b_hits
);
6225 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
6227 cb
->l2rcb_hdr
= hdr
;
6230 cb
->l2rcb_flags
= zio_flags
;
6232 asize
= vdev_psize_to_asize(vd
, size
);
6233 if (asize
!= size
) {
6234 abd
= abd_alloc_for_io(asize
,
6235 HDR_ISTYPE_METADATA(hdr
));
6236 cb
->l2rcb_abd
= abd
;
6241 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
6242 addr
+ asize
<= vd
->vdev_psize
-
6243 VDEV_LABEL_END_SIZE
);
6246 * l2arc read. The SCL_L2ARC lock will be
6247 * released by l2arc_read_done().
6248 * Issue a null zio if the underlying buffer
6249 * was squashed to zero size by compression.
6251 ASSERT3U(arc_hdr_get_compress(hdr
), !=,
6252 ZIO_COMPRESS_EMPTY
);
6253 rzio
= zio_read_phys(pio
, vd
, addr
,
6256 l2arc_read_done
, cb
, priority
,
6257 zio_flags
| ZIO_FLAG_DONT_CACHE
|
6259 ZIO_FLAG_DONT_PROPAGATE
|
6260 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
6262 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
6264 ARCSTAT_INCR(arcstat_l2_read_bytes
,
6265 HDR_GET_PSIZE(hdr
));
6267 if (*arc_flags
& ARC_FLAG_NOWAIT
) {
6272 ASSERT(*arc_flags
& ARC_FLAG_WAIT
);
6273 if (zio_wait(rzio
) == 0)
6276 /* l2arc read error; goto zio_read() */
6278 DTRACE_PROBE1(l2arc__miss
,
6279 arc_buf_hdr_t
*, hdr
);
6280 ARCSTAT_BUMP(arcstat_l2_misses
);
6281 if (HDR_L2_WRITING(hdr
))
6282 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
6283 spa_config_exit(spa
, SCL_L2ARC
, vd
);
6287 spa_config_exit(spa
, SCL_L2ARC
, vd
);
6288 if (l2arc_ndev
!= 0) {
6289 DTRACE_PROBE1(l2arc__miss
,
6290 arc_buf_hdr_t
*, hdr
);
6291 ARCSTAT_BUMP(arcstat_l2_misses
);
6295 rzio
= zio_read(pio
, spa
, bp
, hdr_abd
, size
,
6296 arc_read_done
, hdr
, priority
, zio_flags
, zb
);
6298 if (*arc_flags
& ARC_FLAG_WAIT
) {
6299 rc
= zio_wait(rzio
);
6303 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
6308 spa_read_history_add(spa
, zb
, *arc_flags
);
6313 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
6317 p
= kmem_alloc(sizeof (*p
), KM_SLEEP
);
6319 p
->p_private
= private;
6320 list_link_init(&p
->p_node
);
6321 refcount_create(&p
->p_refcnt
);
6323 mutex_enter(&arc_prune_mtx
);
6324 refcount_add(&p
->p_refcnt
, &arc_prune_list
);
6325 list_insert_head(&arc_prune_list
, p
);
6326 mutex_exit(&arc_prune_mtx
);
6332 arc_remove_prune_callback(arc_prune_t
*p
)
6334 boolean_t wait
= B_FALSE
;
6335 mutex_enter(&arc_prune_mtx
);
6336 list_remove(&arc_prune_list
, p
);
6337 if (refcount_remove(&p
->p_refcnt
, &arc_prune_list
) > 0)
6339 mutex_exit(&arc_prune_mtx
);
6341 /* wait for arc_prune_task to finish */
6343 taskq_wait_outstanding(arc_prune_taskq
, 0);
6344 ASSERT0(refcount_count(&p
->p_refcnt
));
6345 refcount_destroy(&p
->p_refcnt
);
6346 kmem_free(p
, sizeof (*p
));
6350 * Notify the arc that a block was freed, and thus will never be used again.
6353 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
6356 kmutex_t
*hash_lock
;
6357 uint64_t guid
= spa_load_guid(spa
);
6359 ASSERT(!BP_IS_EMBEDDED(bp
));
6361 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
6366 * We might be trying to free a block that is still doing I/O
6367 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6368 * dmu_sync-ed block). If this block is being prefetched, then it
6369 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6370 * until the I/O completes. A block may also have a reference if it is
6371 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6372 * have written the new block to its final resting place on disk but
6373 * without the dedup flag set. This would have left the hdr in the MRU
6374 * state and discoverable. When the txg finally syncs it detects that
6375 * the block was overridden in open context and issues an override I/O.
6376 * Since this is a dedup block, the override I/O will determine if the
6377 * block is already in the DDT. If so, then it will replace the io_bp
6378 * with the bp from the DDT and allow the I/O to finish. When the I/O
6379 * reaches the done callback, dbuf_write_override_done, it will
6380 * check to see if the io_bp and io_bp_override are identical.
6381 * If they are not, then it indicates that the bp was replaced with
6382 * the bp in the DDT and the override bp is freed. This allows
6383 * us to arrive here with a reference on a block that is being
6384 * freed. So if we have an I/O in progress, or a reference to
6385 * this hdr, then we don't destroy the hdr.
6387 if (!HDR_HAS_L1HDR(hdr
) || (!HDR_IO_IN_PROGRESS(hdr
) &&
6388 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
))) {
6389 arc_change_state(arc_anon
, hdr
, hash_lock
);
6390 arc_hdr_destroy(hdr
);
6391 mutex_exit(hash_lock
);
6393 mutex_exit(hash_lock
);
6399 * Release this buffer from the cache, making it an anonymous buffer. This
6400 * must be done after a read and prior to modifying the buffer contents.
6401 * If the buffer has more than one reference, we must make
6402 * a new hdr for the buffer.
6405 arc_release(arc_buf_t
*buf
, void *tag
)
6407 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
6410 * It would be nice to assert that if its DMU metadata (level >
6411 * 0 || it's the dnode file), then it must be syncing context.
6412 * But we don't know that information at this level.
6415 mutex_enter(&buf
->b_evict_lock
);
6417 ASSERT(HDR_HAS_L1HDR(hdr
));
6420 * We don't grab the hash lock prior to this check, because if
6421 * the buffer's header is in the arc_anon state, it won't be
6422 * linked into the hash table.
6424 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
6425 mutex_exit(&buf
->b_evict_lock
);
6426 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
6427 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
6428 ASSERT(!HDR_HAS_L2HDR(hdr
));
6429 ASSERT(HDR_EMPTY(hdr
));
6431 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
6432 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), ==, 1);
6433 ASSERT(!list_link_active(&hdr
->b_l1hdr
.b_arc_node
));
6435 hdr
->b_l1hdr
.b_arc_access
= 0;
6438 * If the buf is being overridden then it may already
6439 * have a hdr that is not empty.
6441 buf_discard_identity(hdr
);
6447 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
6448 mutex_enter(hash_lock
);
6451 * This assignment is only valid as long as the hash_lock is
6452 * held, we must be careful not to reference state or the
6453 * b_state field after dropping the lock.
6455 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
6456 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
6457 ASSERT3P(state
, !=, arc_anon
);
6459 /* this buffer is not on any list */
6460 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), >, 0);
6462 if (HDR_HAS_L2HDR(hdr
)) {
6463 mutex_enter(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
6466 * We have to recheck this conditional again now that
6467 * we're holding the l2ad_mtx to prevent a race with
6468 * another thread which might be concurrently calling
6469 * l2arc_evict(). In that case, l2arc_evict() might have
6470 * destroyed the header's L2 portion as we were waiting
6471 * to acquire the l2ad_mtx.
6473 if (HDR_HAS_L2HDR(hdr
))
6474 arc_hdr_l2hdr_destroy(hdr
);
6476 mutex_exit(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
6480 * Do we have more than one buf?
6482 if (hdr
->b_l1hdr
.b_bufcnt
> 1) {
6483 arc_buf_hdr_t
*nhdr
;
6484 uint64_t spa
= hdr
->b_spa
;
6485 uint64_t psize
= HDR_GET_PSIZE(hdr
);
6486 uint64_t lsize
= HDR_GET_LSIZE(hdr
);
6487 boolean_t
protected = HDR_PROTECTED(hdr
);
6488 enum zio_compress compress
= arc_hdr_get_compress(hdr
);
6489 arc_buf_contents_t type
= arc_buf_type(hdr
);
6490 VERIFY3U(hdr
->b_type
, ==, type
);
6492 ASSERT(hdr
->b_l1hdr
.b_buf
!= buf
|| buf
->b_next
!= NULL
);
6493 (void) remove_reference(hdr
, hash_lock
, tag
);
6495 if (arc_buf_is_shared(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
6496 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
6497 ASSERT(ARC_BUF_LAST(buf
));
6501 * Pull the data off of this hdr and attach it to
6502 * a new anonymous hdr. Also find the last buffer
6503 * in the hdr's buffer list.
6505 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
6506 ASSERT3P(lastbuf
, !=, NULL
);
6509 * If the current arc_buf_t and the hdr are sharing their data
6510 * buffer, then we must stop sharing that block.
6512 if (arc_buf_is_shared(buf
)) {
6513 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
6514 VERIFY(!arc_buf_is_shared(lastbuf
));
6517 * First, sever the block sharing relationship between
6518 * buf and the arc_buf_hdr_t.
6520 arc_unshare_buf(hdr
, buf
);
6523 * Now we need to recreate the hdr's b_pabd. Since we
6524 * have lastbuf handy, we try to share with it, but if
6525 * we can't then we allocate a new b_pabd and copy the
6526 * data from buf into it.
6528 if (arc_can_share(hdr
, lastbuf
)) {
6529 arc_share_buf(hdr
, lastbuf
);
6531 arc_hdr_alloc_abd(hdr
, B_FALSE
);
6532 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
,
6533 buf
->b_data
, psize
);
6535 VERIFY3P(lastbuf
->b_data
, !=, NULL
);
6536 } else if (HDR_SHARED_DATA(hdr
)) {
6538 * Uncompressed shared buffers are always at the end
6539 * of the list. Compressed buffers don't have the
6540 * same requirements. This makes it hard to
6541 * simply assert that the lastbuf is shared so
6542 * we rely on the hdr's compression flags to determine
6543 * if we have a compressed, shared buffer.
6545 ASSERT(arc_buf_is_shared(lastbuf
) ||
6546 arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
);
6547 ASSERT(!ARC_BUF_SHARED(buf
));
6550 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
|| HDR_HAS_RABD(hdr
));
6551 ASSERT3P(state
, !=, arc_l2c_only
);
6553 (void) refcount_remove_many(&state
->arcs_size
,
6554 arc_buf_size(buf
), buf
);
6556 if (refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
)) {
6557 ASSERT3P(state
, !=, arc_l2c_only
);
6558 (void) refcount_remove_many(&state
->arcs_esize
[type
],
6559 arc_buf_size(buf
), buf
);
6562 hdr
->b_l1hdr
.b_bufcnt
-= 1;
6563 if (ARC_BUF_ENCRYPTED(buf
))
6564 hdr
->b_crypt_hdr
.b_ebufcnt
-= 1;
6566 arc_cksum_verify(buf
);
6567 arc_buf_unwatch(buf
);
6569 /* if this is the last uncompressed buf free the checksum */
6570 if (!arc_hdr_has_uncompressed_buf(hdr
))
6571 arc_cksum_free(hdr
);
6573 mutex_exit(hash_lock
);
6576 * Allocate a new hdr. The new hdr will contain a b_pabd
6577 * buffer which will be freed in arc_write().
6579 nhdr
= arc_hdr_alloc(spa
, psize
, lsize
, protected,
6580 compress
, type
, HDR_HAS_RABD(hdr
));
6581 ASSERT3P(nhdr
->b_l1hdr
.b_buf
, ==, NULL
);
6582 ASSERT0(nhdr
->b_l1hdr
.b_bufcnt
);
6583 ASSERT0(refcount_count(&nhdr
->b_l1hdr
.b_refcnt
));
6584 VERIFY3U(nhdr
->b_type
, ==, type
);
6585 ASSERT(!HDR_SHARED_DATA(nhdr
));
6587 nhdr
->b_l1hdr
.b_buf
= buf
;
6588 nhdr
->b_l1hdr
.b_bufcnt
= 1;
6589 if (ARC_BUF_ENCRYPTED(buf
))
6590 nhdr
->b_crypt_hdr
.b_ebufcnt
= 1;
6591 nhdr
->b_l1hdr
.b_mru_hits
= 0;
6592 nhdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
6593 nhdr
->b_l1hdr
.b_mfu_hits
= 0;
6594 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
6595 nhdr
->b_l1hdr
.b_l2_hits
= 0;
6596 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, tag
);
6599 mutex_exit(&buf
->b_evict_lock
);
6600 (void) refcount_add_many(&arc_anon
->arcs_size
,
6601 HDR_GET_LSIZE(nhdr
), buf
);
6603 mutex_exit(&buf
->b_evict_lock
);
6604 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 1);
6605 /* protected by hash lock, or hdr is on arc_anon */
6606 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
6607 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
6608 hdr
->b_l1hdr
.b_mru_hits
= 0;
6609 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
6610 hdr
->b_l1hdr
.b_mfu_hits
= 0;
6611 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
6612 hdr
->b_l1hdr
.b_l2_hits
= 0;
6613 arc_change_state(arc_anon
, hdr
, hash_lock
);
6614 hdr
->b_l1hdr
.b_arc_access
= 0;
6616 mutex_exit(hash_lock
);
6617 buf_discard_identity(hdr
);
6623 arc_released(arc_buf_t
*buf
)
6627 mutex_enter(&buf
->b_evict_lock
);
6628 released
= (buf
->b_data
!= NULL
&&
6629 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
6630 mutex_exit(&buf
->b_evict_lock
);
6636 arc_referenced(arc_buf_t
*buf
)
6640 mutex_enter(&buf
->b_evict_lock
);
6641 referenced
= (refcount_count(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
6642 mutex_exit(&buf
->b_evict_lock
);
6643 return (referenced
);
6648 arc_write_ready(zio_t
*zio
)
6650 arc_write_callback_t
*callback
= zio
->io_private
;
6651 arc_buf_t
*buf
= callback
->awcb_buf
;
6652 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
6653 blkptr_t
*bp
= zio
->io_bp
;
6654 uint64_t psize
= BP_IS_HOLE(bp
) ? 0 : BP_GET_PSIZE(bp
);
6655 fstrans_cookie_t cookie
= spl_fstrans_mark();
6657 ASSERT(HDR_HAS_L1HDR(hdr
));
6658 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
6659 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
6662 * If we're reexecuting this zio because the pool suspended, then
6663 * cleanup any state that was previously set the first time the
6664 * callback was invoked.
6666 if (zio
->io_flags
& ZIO_FLAG_REEXECUTED
) {
6667 arc_cksum_free(hdr
);
6668 arc_buf_unwatch(buf
);
6669 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
6670 if (arc_buf_is_shared(buf
)) {
6671 arc_unshare_buf(hdr
, buf
);
6673 arc_hdr_free_abd(hdr
, B_FALSE
);
6677 if (HDR_HAS_RABD(hdr
))
6678 arc_hdr_free_abd(hdr
, B_TRUE
);
6680 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
6681 ASSERT(!HDR_HAS_RABD(hdr
));
6682 ASSERT(!HDR_SHARED_DATA(hdr
));
6683 ASSERT(!arc_buf_is_shared(buf
));
6685 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
6687 if (HDR_IO_IN_PROGRESS(hdr
))
6688 ASSERT(zio
->io_flags
& ZIO_FLAG_REEXECUTED
);
6690 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6692 if (BP_IS_PROTECTED(bp
) != !!HDR_PROTECTED(hdr
))
6693 hdr
= arc_hdr_realloc_crypt(hdr
, BP_IS_PROTECTED(bp
));
6695 if (BP_IS_PROTECTED(bp
)) {
6696 /* ZIL blocks are written through zio_rewrite */
6697 ASSERT3U(BP_GET_TYPE(bp
), !=, DMU_OT_INTENT_LOG
);
6698 ASSERT(HDR_PROTECTED(hdr
));
6700 hdr
->b_crypt_hdr
.b_ot
= BP_GET_TYPE(bp
);
6701 hdr
->b_crypt_hdr
.b_dsobj
= zio
->io_bookmark
.zb_objset
;
6702 zio_crypt_decode_params_bp(bp
, hdr
->b_crypt_hdr
.b_salt
,
6703 hdr
->b_crypt_hdr
.b_iv
);
6704 zio_crypt_decode_mac_bp(bp
, hdr
->b_crypt_hdr
.b_mac
);
6708 * If this block was written for raw encryption but the zio layer
6709 * ended up only authenticating it, adjust the buffer flags now.
6711 if (BP_IS_AUTHENTICATED(bp
) && ARC_BUF_ENCRYPTED(buf
)) {
6712 arc_hdr_set_flags(hdr
, ARC_FLAG_NOAUTH
);
6713 buf
->b_flags
&= ~ARC_BUF_FLAG_ENCRYPTED
;
6714 if (BP_GET_COMPRESS(bp
) == ZIO_COMPRESS_OFF
)
6715 buf
->b_flags
&= ~ARC_BUF_FLAG_COMPRESSED
;
6718 /* this must be done after the buffer flags are adjusted */
6719 arc_cksum_compute(buf
);
6721 enum zio_compress compress
;
6722 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
)) {
6723 compress
= ZIO_COMPRESS_OFF
;
6725 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(bp
));
6726 compress
= BP_GET_COMPRESS(bp
);
6728 HDR_SET_PSIZE(hdr
, psize
);
6729 arc_hdr_set_compress(hdr
, compress
);
6731 if (zio
->io_error
!= 0 || psize
== 0)
6735 * Fill the hdr with data. If the buffer is encrypted we have no choice
6736 * but to copy the data into b_radb. If the hdr is compressed, the data
6737 * we want is available from the zio, otherwise we can take it from
6740 * We might be able to share the buf's data with the hdr here. However,
6741 * doing so would cause the ARC to be full of linear ABDs if we write a
6742 * lot of shareable data. As a compromise, we check whether scattered
6743 * ABDs are allowed, and assume that if they are then the user wants
6744 * the ARC to be primarily filled with them regardless of the data being
6745 * written. Therefore, if they're allowed then we allocate one and copy
6746 * the data into it; otherwise, we share the data directly if we can.
6748 if (ARC_BUF_ENCRYPTED(buf
)) {
6749 ASSERT3U(psize
, >, 0);
6750 ASSERT(ARC_BUF_COMPRESSED(buf
));
6751 arc_hdr_alloc_abd(hdr
, B_TRUE
);
6752 abd_copy(hdr
->b_crypt_hdr
.b_rabd
, zio
->io_abd
, psize
);
6753 } else if (zfs_abd_scatter_enabled
|| !arc_can_share(hdr
, buf
)) {
6755 * Ideally, we would always copy the io_abd into b_pabd, but the
6756 * user may have disabled compressed ARC, thus we must check the
6757 * hdr's compression setting rather than the io_bp's.
6759 if (BP_IS_ENCRYPTED(bp
)) {
6760 ASSERT3U(psize
, >, 0);
6761 arc_hdr_alloc_abd(hdr
, B_TRUE
);
6762 abd_copy(hdr
->b_crypt_hdr
.b_rabd
, zio
->io_abd
, psize
);
6763 } else if (arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
&&
6764 !ARC_BUF_COMPRESSED(buf
)) {
6765 ASSERT3U(psize
, >, 0);
6766 arc_hdr_alloc_abd(hdr
, B_FALSE
);
6767 abd_copy(hdr
->b_l1hdr
.b_pabd
, zio
->io_abd
, psize
);
6769 ASSERT3U(zio
->io_orig_size
, ==, arc_hdr_size(hdr
));
6770 arc_hdr_alloc_abd(hdr
, B_FALSE
);
6771 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
6775 ASSERT3P(buf
->b_data
, ==, abd_to_buf(zio
->io_orig_abd
));
6776 ASSERT3U(zio
->io_orig_size
, ==, arc_buf_size(buf
));
6777 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
6779 arc_share_buf(hdr
, buf
);
6783 arc_hdr_verify(hdr
, bp
);
6784 spl_fstrans_unmark(cookie
);
6788 arc_write_children_ready(zio_t
*zio
)
6790 arc_write_callback_t
*callback
= zio
->io_private
;
6791 arc_buf_t
*buf
= callback
->awcb_buf
;
6793 callback
->awcb_children_ready(zio
, buf
, callback
->awcb_private
);
6797 * The SPA calls this callback for each physical write that happens on behalf
6798 * of a logical write. See the comment in dbuf_write_physdone() for details.
6801 arc_write_physdone(zio_t
*zio
)
6803 arc_write_callback_t
*cb
= zio
->io_private
;
6804 if (cb
->awcb_physdone
!= NULL
)
6805 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
6809 arc_write_done(zio_t
*zio
)
6811 arc_write_callback_t
*callback
= zio
->io_private
;
6812 arc_buf_t
*buf
= callback
->awcb_buf
;
6813 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
6815 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
6817 if (zio
->io_error
== 0) {
6818 arc_hdr_verify(hdr
, zio
->io_bp
);
6820 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
6821 buf_discard_identity(hdr
);
6823 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
6824 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
6827 ASSERT(HDR_EMPTY(hdr
));
6831 * If the block to be written was all-zero or compressed enough to be
6832 * embedded in the BP, no write was performed so there will be no
6833 * dva/birth/checksum. The buffer must therefore remain anonymous
6836 if (!HDR_EMPTY(hdr
)) {
6837 arc_buf_hdr_t
*exists
;
6838 kmutex_t
*hash_lock
;
6840 ASSERT3U(zio
->io_error
, ==, 0);
6842 arc_cksum_verify(buf
);
6844 exists
= buf_hash_insert(hdr
, &hash_lock
);
6845 if (exists
!= NULL
) {
6847 * This can only happen if we overwrite for
6848 * sync-to-convergence, because we remove
6849 * buffers from the hash table when we arc_free().
6851 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
6852 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
6853 panic("bad overwrite, hdr=%p exists=%p",
6854 (void *)hdr
, (void *)exists
);
6855 ASSERT(refcount_is_zero(
6856 &exists
->b_l1hdr
.b_refcnt
));
6857 arc_change_state(arc_anon
, exists
, hash_lock
);
6858 mutex_exit(hash_lock
);
6859 arc_hdr_destroy(exists
);
6860 exists
= buf_hash_insert(hdr
, &hash_lock
);
6861 ASSERT3P(exists
, ==, NULL
);
6862 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
6864 ASSERT(zio
->io_prop
.zp_nopwrite
);
6865 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
6866 panic("bad nopwrite, hdr=%p exists=%p",
6867 (void *)hdr
, (void *)exists
);
6870 ASSERT(hdr
->b_l1hdr
.b_bufcnt
== 1);
6871 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
6872 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
6873 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
6876 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6877 /* if it's not anon, we are doing a scrub */
6878 if (exists
== NULL
&& hdr
->b_l1hdr
.b_state
== arc_anon
)
6879 arc_access(hdr
, hash_lock
);
6880 mutex_exit(hash_lock
);
6882 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6885 ASSERT(!refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
6886 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
6888 abd_put(zio
->io_abd
);
6889 kmem_free(callback
, sizeof (arc_write_callback_t
));
6893 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
6894 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
,
6895 const zio_prop_t
*zp
, arc_write_done_func_t
*ready
,
6896 arc_write_done_func_t
*children_ready
, arc_write_done_func_t
*physdone
,
6897 arc_write_done_func_t
*done
, void *private, zio_priority_t priority
,
6898 int zio_flags
, const zbookmark_phys_t
*zb
)
6900 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
6901 arc_write_callback_t
*callback
;
6903 zio_prop_t localprop
= *zp
;
6905 ASSERT3P(ready
, !=, NULL
);
6906 ASSERT3P(done
, !=, NULL
);
6907 ASSERT(!HDR_IO_ERROR(hdr
));
6908 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
6909 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
6910 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
6912 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
6914 if (ARC_BUF_ENCRYPTED(buf
)) {
6915 ASSERT(ARC_BUF_COMPRESSED(buf
));
6916 localprop
.zp_encrypt
= B_TRUE
;
6917 localprop
.zp_compress
= HDR_GET_COMPRESS(hdr
);
6918 localprop
.zp_byteorder
=
6919 (hdr
->b_l1hdr
.b_byteswap
== DMU_BSWAP_NUMFUNCS
) ?
6920 ZFS_HOST_BYTEORDER
: !ZFS_HOST_BYTEORDER
;
6921 bcopy(hdr
->b_crypt_hdr
.b_salt
, localprop
.zp_salt
,
6923 bcopy(hdr
->b_crypt_hdr
.b_iv
, localprop
.zp_iv
,
6925 bcopy(hdr
->b_crypt_hdr
.b_mac
, localprop
.zp_mac
,
6927 if (DMU_OT_IS_ENCRYPTED(localprop
.zp_type
)) {
6928 localprop
.zp_nopwrite
= B_FALSE
;
6929 localprop
.zp_copies
=
6930 MIN(localprop
.zp_copies
, SPA_DVAS_PER_BP
- 1);
6932 zio_flags
|= ZIO_FLAG_RAW
;
6933 } else if (ARC_BUF_COMPRESSED(buf
)) {
6934 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, arc_buf_size(buf
));
6935 localprop
.zp_compress
= HDR_GET_COMPRESS(hdr
);
6936 zio_flags
|= ZIO_FLAG_RAW_COMPRESS
;
6938 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
6939 callback
->awcb_ready
= ready
;
6940 callback
->awcb_children_ready
= children_ready
;
6941 callback
->awcb_physdone
= physdone
;
6942 callback
->awcb_done
= done
;
6943 callback
->awcb_private
= private;
6944 callback
->awcb_buf
= buf
;
6947 * The hdr's b_pabd is now stale, free it now. A new data block
6948 * will be allocated when the zio pipeline calls arc_write_ready().
6950 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
6952 * If the buf is currently sharing the data block with
6953 * the hdr then we need to break that relationship here.
6954 * The hdr will remain with a NULL data pointer and the
6955 * buf will take sole ownership of the block.
6957 if (arc_buf_is_shared(buf
)) {
6958 arc_unshare_buf(hdr
, buf
);
6960 arc_hdr_free_abd(hdr
, B_FALSE
);
6962 VERIFY3P(buf
->b_data
, !=, NULL
);
6965 if (HDR_HAS_RABD(hdr
))
6966 arc_hdr_free_abd(hdr
, B_TRUE
);
6968 if (!(zio_flags
& ZIO_FLAG_RAW
))
6969 arc_hdr_set_compress(hdr
, ZIO_COMPRESS_OFF
);
6971 ASSERT(!arc_buf_is_shared(buf
));
6972 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
6974 zio
= zio_write(pio
, spa
, txg
, bp
,
6975 abd_get_from_buf(buf
->b_data
, HDR_GET_LSIZE(hdr
)),
6976 HDR_GET_LSIZE(hdr
), arc_buf_size(buf
), &localprop
, arc_write_ready
,
6977 (children_ready
!= NULL
) ? arc_write_children_ready
: NULL
,
6978 arc_write_physdone
, arc_write_done
, callback
,
6979 priority
, zio_flags
, zb
);
6985 arc_memory_throttle(uint64_t reserve
, uint64_t txg
)
6988 uint64_t available_memory
= arc_free_memory();
6989 static uint64_t page_load
= 0;
6990 static uint64_t last_txg
= 0;
6994 MIN(available_memory
, vmem_size(heap_arena
, VMEM_FREE
));
6997 if (available_memory
> arc_all_memory() * arc_lotsfree_percent
/ 100)
7000 if (txg
> last_txg
) {
7005 * If we are in pageout, we know that memory is already tight,
7006 * the arc is already going to be evicting, so we just want to
7007 * continue to let page writes occur as quickly as possible.
7009 if (current_is_kswapd()) {
7010 if (page_load
> MAX(arc_sys_free
/ 4, available_memory
) / 4) {
7011 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
7012 return (SET_ERROR(ERESTART
));
7014 /* Note: reserve is inflated, so we deflate */
7015 page_load
+= reserve
/ 8;
7017 } else if (page_load
> 0 && arc_reclaim_needed()) {
7018 /* memory is low, delay before restarting */
7019 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
7020 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
7021 return (SET_ERROR(EAGAIN
));
7029 arc_tempreserve_clear(uint64_t reserve
)
7031 atomic_add_64(&arc_tempreserve
, -reserve
);
7032 ASSERT((int64_t)arc_tempreserve
>= 0);
7036 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
7042 reserve
> arc_c
/4 &&
7043 reserve
* 4 > (2ULL << SPA_MAXBLOCKSHIFT
))
7044 arc_c
= MIN(arc_c_max
, reserve
* 4);
7047 * Throttle when the calculated memory footprint for the TXG
7048 * exceeds the target ARC size.
7050 if (reserve
> arc_c
) {
7051 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
7052 return (SET_ERROR(ERESTART
));
7056 * Don't count loaned bufs as in flight dirty data to prevent long
7057 * network delays from blocking transactions that are ready to be
7058 * assigned to a txg.
7061 /* assert that it has not wrapped around */
7062 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
7064 anon_size
= MAX((int64_t)(refcount_count(&arc_anon
->arcs_size
) -
7065 arc_loaned_bytes
), 0);
7068 * Writes will, almost always, require additional memory allocations
7069 * in order to compress/encrypt/etc the data. We therefore need to
7070 * make sure that there is sufficient available memory for this.
7072 error
= arc_memory_throttle(reserve
, txg
);
7077 * Throttle writes when the amount of dirty data in the cache
7078 * gets too large. We try to keep the cache less than half full
7079 * of dirty blocks so that our sync times don't grow too large.
7080 * Note: if two requests come in concurrently, we might let them
7081 * both succeed, when one of them should fail. Not a huge deal.
7084 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
7085 anon_size
> arc_c
/ 4) {
7086 uint64_t meta_esize
=
7087 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
7088 uint64_t data_esize
=
7089 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
7090 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
7091 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
7092 arc_tempreserve
>> 10, meta_esize
>> 10,
7093 data_esize
>> 10, reserve
>> 10, arc_c
>> 10);
7094 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
7095 return (SET_ERROR(ERESTART
));
7097 atomic_add_64(&arc_tempreserve
, reserve
);
7102 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
7103 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
7105 size
->value
.ui64
= refcount_count(&state
->arcs_size
);
7106 evict_data
->value
.ui64
=
7107 refcount_count(&state
->arcs_esize
[ARC_BUFC_DATA
]);
7108 evict_metadata
->value
.ui64
=
7109 refcount_count(&state
->arcs_esize
[ARC_BUFC_METADATA
]);
7113 arc_kstat_update(kstat_t
*ksp
, int rw
)
7115 arc_stats_t
*as
= ksp
->ks_data
;
7117 if (rw
== KSTAT_WRITE
) {
7118 return (SET_ERROR(EACCES
));
7120 arc_kstat_update_state(arc_anon
,
7121 &as
->arcstat_anon_size
,
7122 &as
->arcstat_anon_evictable_data
,
7123 &as
->arcstat_anon_evictable_metadata
);
7124 arc_kstat_update_state(arc_mru
,
7125 &as
->arcstat_mru_size
,
7126 &as
->arcstat_mru_evictable_data
,
7127 &as
->arcstat_mru_evictable_metadata
);
7128 arc_kstat_update_state(arc_mru_ghost
,
7129 &as
->arcstat_mru_ghost_size
,
7130 &as
->arcstat_mru_ghost_evictable_data
,
7131 &as
->arcstat_mru_ghost_evictable_metadata
);
7132 arc_kstat_update_state(arc_mfu
,
7133 &as
->arcstat_mfu_size
,
7134 &as
->arcstat_mfu_evictable_data
,
7135 &as
->arcstat_mfu_evictable_metadata
);
7136 arc_kstat_update_state(arc_mfu_ghost
,
7137 &as
->arcstat_mfu_ghost_size
,
7138 &as
->arcstat_mfu_ghost_evictable_data
,
7139 &as
->arcstat_mfu_ghost_evictable_metadata
);
7141 as
->arcstat_memory_all_bytes
.value
.ui64
=
7143 as
->arcstat_memory_free_bytes
.value
.ui64
=
7145 as
->arcstat_memory_available_bytes
.value
.i64
=
7146 arc_available_memory();
7153 * This function *must* return indices evenly distributed between all
7154 * sublists of the multilist. This is needed due to how the ARC eviction
7155 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7156 * distributed between all sublists and uses this assumption when
7157 * deciding which sublist to evict from and how much to evict from it.
7160 arc_state_multilist_index_func(multilist_t
*ml
, void *obj
)
7162 arc_buf_hdr_t
*hdr
= obj
;
7165 * We rely on b_dva to generate evenly distributed index
7166 * numbers using buf_hash below. So, as an added precaution,
7167 * let's make sure we never add empty buffers to the arc lists.
7169 ASSERT(!HDR_EMPTY(hdr
));
7172 * The assumption here, is the hash value for a given
7173 * arc_buf_hdr_t will remain constant throughout its lifetime
7174 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7175 * Thus, we don't need to store the header's sublist index
7176 * on insertion, as this index can be recalculated on removal.
7178 * Also, the low order bits of the hash value are thought to be
7179 * distributed evenly. Otherwise, in the case that the multilist
7180 * has a power of two number of sublists, each sublists' usage
7181 * would not be evenly distributed.
7183 return (buf_hash(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
) %
7184 multilist_get_num_sublists(ml
));
7188 * Called during module initialization and periodically thereafter to
7189 * apply reasonable changes to the exposed performance tunings. Non-zero
7190 * zfs_* values which differ from the currently set values will be applied.
7193 arc_tuning_update(void)
7195 uint64_t allmem
= arc_all_memory();
7196 unsigned long limit
;
7198 /* Valid range: 64M - <all physical memory> */
7199 if ((zfs_arc_max
) && (zfs_arc_max
!= arc_c_max
) &&
7200 (zfs_arc_max
> 64 << 20) && (zfs_arc_max
< allmem
) &&
7201 (zfs_arc_max
> arc_c_min
)) {
7202 arc_c_max
= zfs_arc_max
;
7204 arc_p
= (arc_c
>> 1);
7205 if (arc_meta_limit
> arc_c_max
)
7206 arc_meta_limit
= arc_c_max
;
7207 if (arc_dnode_limit
> arc_meta_limit
)
7208 arc_dnode_limit
= arc_meta_limit
;
7211 /* Valid range: 32M - <arc_c_max> */
7212 if ((zfs_arc_min
) && (zfs_arc_min
!= arc_c_min
) &&
7213 (zfs_arc_min
>= 2ULL << SPA_MAXBLOCKSHIFT
) &&
7214 (zfs_arc_min
<= arc_c_max
)) {
7215 arc_c_min
= zfs_arc_min
;
7216 arc_c
= MAX(arc_c
, arc_c_min
);
7219 /* Valid range: 16M - <arc_c_max> */
7220 if ((zfs_arc_meta_min
) && (zfs_arc_meta_min
!= arc_meta_min
) &&
7221 (zfs_arc_meta_min
>= 1ULL << SPA_MAXBLOCKSHIFT
) &&
7222 (zfs_arc_meta_min
<= arc_c_max
)) {
7223 arc_meta_min
= zfs_arc_meta_min
;
7224 if (arc_meta_limit
< arc_meta_min
)
7225 arc_meta_limit
= arc_meta_min
;
7226 if (arc_dnode_limit
< arc_meta_min
)
7227 arc_dnode_limit
= arc_meta_min
;
7230 /* Valid range: <arc_meta_min> - <arc_c_max> */
7231 limit
= zfs_arc_meta_limit
? zfs_arc_meta_limit
:
7232 MIN(zfs_arc_meta_limit_percent
, 100) * arc_c_max
/ 100;
7233 if ((limit
!= arc_meta_limit
) &&
7234 (limit
>= arc_meta_min
) &&
7235 (limit
<= arc_c_max
))
7236 arc_meta_limit
= limit
;
7238 /* Valid range: <arc_meta_min> - <arc_meta_limit> */
7239 limit
= zfs_arc_dnode_limit
? zfs_arc_dnode_limit
:
7240 MIN(zfs_arc_dnode_limit_percent
, 100) * arc_meta_limit
/ 100;
7241 if ((limit
!= arc_dnode_limit
) &&
7242 (limit
>= arc_meta_min
) &&
7243 (limit
<= arc_meta_limit
))
7244 arc_dnode_limit
= limit
;
7246 /* Valid range: 1 - N */
7247 if (zfs_arc_grow_retry
)
7248 arc_grow_retry
= zfs_arc_grow_retry
;
7250 /* Valid range: 1 - N */
7251 if (zfs_arc_shrink_shift
) {
7252 arc_shrink_shift
= zfs_arc_shrink_shift
;
7253 arc_no_grow_shift
= MIN(arc_no_grow_shift
, arc_shrink_shift
-1);
7256 /* Valid range: 1 - N */
7257 if (zfs_arc_p_min_shift
)
7258 arc_p_min_shift
= zfs_arc_p_min_shift
;
7260 /* Valid range: 1 - N ms */
7261 if (zfs_arc_min_prefetch_ms
)
7262 arc_min_prefetch_ms
= zfs_arc_min_prefetch_ms
;
7264 /* Valid range: 1 - N ms */
7265 if (zfs_arc_min_prescient_prefetch_ms
) {
7266 arc_min_prescient_prefetch_ms
=
7267 zfs_arc_min_prescient_prefetch_ms
;
7270 /* Valid range: 0 - 100 */
7271 if ((zfs_arc_lotsfree_percent
>= 0) &&
7272 (zfs_arc_lotsfree_percent
<= 100))
7273 arc_lotsfree_percent
= zfs_arc_lotsfree_percent
;
7275 /* Valid range: 0 - <all physical memory> */
7276 if ((zfs_arc_sys_free
) && (zfs_arc_sys_free
!= arc_sys_free
))
7277 arc_sys_free
= MIN(MAX(zfs_arc_sys_free
, 0), allmem
);
7282 arc_state_init(void)
7284 arc_anon
= &ARC_anon
;
7286 arc_mru_ghost
= &ARC_mru_ghost
;
7288 arc_mfu_ghost
= &ARC_mfu_ghost
;
7289 arc_l2c_only
= &ARC_l2c_only
;
7291 arc_mru
->arcs_list
[ARC_BUFC_METADATA
] =
7292 multilist_create(sizeof (arc_buf_hdr_t
),
7293 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7294 arc_state_multilist_index_func
);
7295 arc_mru
->arcs_list
[ARC_BUFC_DATA
] =
7296 multilist_create(sizeof (arc_buf_hdr_t
),
7297 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7298 arc_state_multilist_index_func
);
7299 arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
7300 multilist_create(sizeof (arc_buf_hdr_t
),
7301 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7302 arc_state_multilist_index_func
);
7303 arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
] =
7304 multilist_create(sizeof (arc_buf_hdr_t
),
7305 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7306 arc_state_multilist_index_func
);
7307 arc_mfu
->arcs_list
[ARC_BUFC_METADATA
] =
7308 multilist_create(sizeof (arc_buf_hdr_t
),
7309 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7310 arc_state_multilist_index_func
);
7311 arc_mfu
->arcs_list
[ARC_BUFC_DATA
] =
7312 multilist_create(sizeof (arc_buf_hdr_t
),
7313 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7314 arc_state_multilist_index_func
);
7315 arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
7316 multilist_create(sizeof (arc_buf_hdr_t
),
7317 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7318 arc_state_multilist_index_func
);
7319 arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
] =
7320 multilist_create(sizeof (arc_buf_hdr_t
),
7321 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7322 arc_state_multilist_index_func
);
7323 arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
] =
7324 multilist_create(sizeof (arc_buf_hdr_t
),
7325 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7326 arc_state_multilist_index_func
);
7327 arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
] =
7328 multilist_create(sizeof (arc_buf_hdr_t
),
7329 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7330 arc_state_multilist_index_func
);
7332 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
7333 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
7334 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
7335 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
7336 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
7337 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
7338 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
7339 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
7340 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
7341 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
7342 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
7343 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
7345 refcount_create(&arc_anon
->arcs_size
);
7346 refcount_create(&arc_mru
->arcs_size
);
7347 refcount_create(&arc_mru_ghost
->arcs_size
);
7348 refcount_create(&arc_mfu
->arcs_size
);
7349 refcount_create(&arc_mfu_ghost
->arcs_size
);
7350 refcount_create(&arc_l2c_only
->arcs_size
);
7352 arc_anon
->arcs_state
= ARC_STATE_ANON
;
7353 arc_mru
->arcs_state
= ARC_STATE_MRU
;
7354 arc_mru_ghost
->arcs_state
= ARC_STATE_MRU_GHOST
;
7355 arc_mfu
->arcs_state
= ARC_STATE_MFU
;
7356 arc_mfu_ghost
->arcs_state
= ARC_STATE_MFU_GHOST
;
7357 arc_l2c_only
->arcs_state
= ARC_STATE_L2C_ONLY
;
7361 arc_state_fini(void)
7363 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
7364 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
7365 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
7366 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
7367 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
7368 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
7369 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
7370 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
7371 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
7372 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
7373 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
7374 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
7376 refcount_destroy(&arc_anon
->arcs_size
);
7377 refcount_destroy(&arc_mru
->arcs_size
);
7378 refcount_destroy(&arc_mru_ghost
->arcs_size
);
7379 refcount_destroy(&arc_mfu
->arcs_size
);
7380 refcount_destroy(&arc_mfu_ghost
->arcs_size
);
7381 refcount_destroy(&arc_l2c_only
->arcs_size
);
7383 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
7384 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
7385 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
7386 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
7387 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
7388 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
7389 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
7390 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
7391 multilist_destroy(arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]);
7392 multilist_destroy(arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]);
7396 arc_target_bytes(void)
7404 uint64_t percent
, allmem
= arc_all_memory();
7406 mutex_init(&arc_reclaim_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
7407 cv_init(&arc_reclaim_thread_cv
, NULL
, CV_DEFAULT
, NULL
);
7408 cv_init(&arc_reclaim_waiters_cv
, NULL
, CV_DEFAULT
, NULL
);
7410 /* Convert seconds to clock ticks */
7411 arc_min_prefetch_ms
= 1;
7412 arc_min_prescient_prefetch_ms
= 6;
7416 * Register a shrinker to support synchronous (direct) memory
7417 * reclaim from the arc. This is done to prevent kswapd from
7418 * swapping out pages when it is preferable to shrink the arc.
7420 spl_register_shrinker(&arc_shrinker
);
7422 /* Set to 1/64 of all memory or a minimum of 512K */
7423 arc_sys_free
= MAX(allmem
/ 64, (512 * 1024));
7427 /* Set max to 1/2 of all memory */
7428 arc_c_max
= allmem
/ 2;
7431 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more */
7432 arc_c_min
= MAX(allmem
/ 32, 2ULL << SPA_MAXBLOCKSHIFT
);
7435 * In userland, there's only the memory pressure that we artificially
7436 * create (see arc_available_memory()). Don't let arc_c get too
7437 * small, because it can cause transactions to be larger than
7438 * arc_c, causing arc_tempreserve_space() to fail.
7440 arc_c_min
= MAX(arc_c_max
/ 2, 2ULL << SPA_MAXBLOCKSHIFT
);
7444 arc_p
= (arc_c
>> 1);
7447 /* Set min to 1/2 of arc_c_min */
7448 arc_meta_min
= 1ULL << SPA_MAXBLOCKSHIFT
;
7449 /* Initialize maximum observed usage to zero */
7452 * Set arc_meta_limit to a percent of arc_c_max with a floor of
7453 * arc_meta_min, and a ceiling of arc_c_max.
7455 percent
= MIN(zfs_arc_meta_limit_percent
, 100);
7456 arc_meta_limit
= MAX(arc_meta_min
, (percent
* arc_c_max
) / 100);
7457 percent
= MIN(zfs_arc_dnode_limit_percent
, 100);
7458 arc_dnode_limit
= (percent
* arc_meta_limit
) / 100;
7460 /* Apply user specified tunings */
7461 arc_tuning_update();
7463 /* if kmem_flags are set, lets try to use less memory */
7464 if (kmem_debugging())
7466 if (arc_c
< arc_c_min
)
7472 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
7473 offsetof(arc_prune_t
, p_node
));
7474 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7476 arc_prune_taskq
= taskq_create("arc_prune", max_ncpus
, defclsyspri
,
7477 max_ncpus
, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
7479 arc_reclaim_thread_exit
= B_FALSE
;
7481 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
7482 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
7484 if (arc_ksp
!= NULL
) {
7485 arc_ksp
->ks_data
= &arc_stats
;
7486 arc_ksp
->ks_update
= arc_kstat_update
;
7487 kstat_install(arc_ksp
);
7490 (void) thread_create(NULL
, 0, arc_reclaim_thread
, NULL
, 0, &p0
,
7491 TS_RUN
, defclsyspri
);
7497 * Calculate maximum amount of dirty data per pool.
7499 * If it has been set by a module parameter, take that.
7500 * Otherwise, use a percentage of physical memory defined by
7501 * zfs_dirty_data_max_percent (default 10%) with a cap at
7502 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
7504 if (zfs_dirty_data_max_max
== 0)
7505 zfs_dirty_data_max_max
= MIN(4ULL * 1024 * 1024 * 1024,
7506 allmem
* zfs_dirty_data_max_max_percent
/ 100);
7508 if (zfs_dirty_data_max
== 0) {
7509 zfs_dirty_data_max
= allmem
*
7510 zfs_dirty_data_max_percent
/ 100;
7511 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
7512 zfs_dirty_data_max_max
);
7522 spl_unregister_shrinker(&arc_shrinker
);
7523 #endif /* _KERNEL */
7525 mutex_enter(&arc_reclaim_lock
);
7526 arc_reclaim_thread_exit
= B_TRUE
;
7528 * The reclaim thread will set arc_reclaim_thread_exit back to
7529 * B_FALSE when it is finished exiting; we're waiting for that.
7531 while (arc_reclaim_thread_exit
) {
7532 cv_signal(&arc_reclaim_thread_cv
);
7533 cv_wait(&arc_reclaim_thread_cv
, &arc_reclaim_lock
);
7535 mutex_exit(&arc_reclaim_lock
);
7537 /* Use B_TRUE to ensure *all* buffers are evicted */
7538 arc_flush(NULL
, B_TRUE
);
7542 if (arc_ksp
!= NULL
) {
7543 kstat_delete(arc_ksp
);
7547 taskq_wait(arc_prune_taskq
);
7548 taskq_destroy(arc_prune_taskq
);
7550 mutex_enter(&arc_prune_mtx
);
7551 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
7552 list_remove(&arc_prune_list
, p
);
7553 refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
7554 refcount_destroy(&p
->p_refcnt
);
7555 kmem_free(p
, sizeof (*p
));
7557 mutex_exit(&arc_prune_mtx
);
7559 list_destroy(&arc_prune_list
);
7560 mutex_destroy(&arc_prune_mtx
);
7561 mutex_destroy(&arc_reclaim_lock
);
7562 cv_destroy(&arc_reclaim_thread_cv
);
7563 cv_destroy(&arc_reclaim_waiters_cv
);
7568 ASSERT0(arc_loaned_bytes
);
7574 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7575 * It uses dedicated storage devices to hold cached data, which are populated
7576 * using large infrequent writes. The main role of this cache is to boost
7577 * the performance of random read workloads. The intended L2ARC devices
7578 * include short-stroked disks, solid state disks, and other media with
7579 * substantially faster read latency than disk.
7581 * +-----------------------+
7583 * +-----------------------+
7586 * l2arc_feed_thread() arc_read()
7590 * +---------------+ |
7592 * +---------------+ |
7597 * +-------+ +-------+
7599 * | cache | | cache |
7600 * +-------+ +-------+
7601 * +=========+ .-----.
7602 * : L2ARC : |-_____-|
7603 * : devices : | Disks |
7604 * +=========+ `-_____-'
7606 * Read requests are satisfied from the following sources, in order:
7609 * 2) vdev cache of L2ARC devices
7611 * 4) vdev cache of disks
7614 * Some L2ARC device types exhibit extremely slow write performance.
7615 * To accommodate for this there are some significant differences between
7616 * the L2ARC and traditional cache design:
7618 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7619 * the ARC behave as usual, freeing buffers and placing headers on ghost
7620 * lists. The ARC does not send buffers to the L2ARC during eviction as
7621 * this would add inflated write latencies for all ARC memory pressure.
7623 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7624 * It does this by periodically scanning buffers from the eviction-end of
7625 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7626 * not already there. It scans until a headroom of buffers is satisfied,
7627 * which itself is a buffer for ARC eviction. If a compressible buffer is
7628 * found during scanning and selected for writing to an L2ARC device, we
7629 * temporarily boost scanning headroom during the next scan cycle to make
7630 * sure we adapt to compression effects (which might significantly reduce
7631 * the data volume we write to L2ARC). The thread that does this is
7632 * l2arc_feed_thread(), illustrated below; example sizes are included to
7633 * provide a better sense of ratio than this diagram:
7636 * +---------------------+----------+
7637 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7638 * +---------------------+----------+ | o L2ARC eligible
7639 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7640 * +---------------------+----------+ |
7641 * 15.9 Gbytes ^ 32 Mbytes |
7643 * l2arc_feed_thread()
7645 * l2arc write hand <--[oooo]--'
7649 * +==============================+
7650 * L2ARC dev |####|#|###|###| |####| ... |
7651 * +==============================+
7654 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7655 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7656 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7657 * safe to say that this is an uncommon case, since buffers at the end of
7658 * the ARC lists have moved there due to inactivity.
7660 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7661 * then the L2ARC simply misses copying some buffers. This serves as a
7662 * pressure valve to prevent heavy read workloads from both stalling the ARC
7663 * with waits and clogging the L2ARC with writes. This also helps prevent
7664 * the potential for the L2ARC to churn if it attempts to cache content too
7665 * quickly, such as during backups of the entire pool.
7667 * 5. After system boot and before the ARC has filled main memory, there are
7668 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7669 * lists can remain mostly static. Instead of searching from tail of these
7670 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7671 * for eligible buffers, greatly increasing its chance of finding them.
7673 * The L2ARC device write speed is also boosted during this time so that
7674 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7675 * there are no L2ARC reads, and no fear of degrading read performance
7676 * through increased writes.
7678 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7679 * the vdev queue can aggregate them into larger and fewer writes. Each
7680 * device is written to in a rotor fashion, sweeping writes through
7681 * available space then repeating.
7683 * 7. The L2ARC does not store dirty content. It never needs to flush
7684 * write buffers back to disk based storage.
7686 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7687 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7689 * The performance of the L2ARC can be tweaked by a number of tunables, which
7690 * may be necessary for different workloads:
7692 * l2arc_write_max max write bytes per interval
7693 * l2arc_write_boost extra write bytes during device warmup
7694 * l2arc_noprefetch skip caching prefetched buffers
7695 * l2arc_headroom number of max device writes to precache
7696 * l2arc_headroom_boost when we find compressed buffers during ARC
7697 * scanning, we multiply headroom by this
7698 * percentage factor for the next scan cycle,
7699 * since more compressed buffers are likely to
7701 * l2arc_feed_secs seconds between L2ARC writing
7703 * Tunables may be removed or added as future performance improvements are
7704 * integrated, and also may become zpool properties.
7706 * There are three key functions that control how the L2ARC warms up:
7708 * l2arc_write_eligible() check if a buffer is eligible to cache
7709 * l2arc_write_size() calculate how much to write
7710 * l2arc_write_interval() calculate sleep delay between writes
7712 * These three functions determine what to write, how much, and how quickly
7717 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*hdr
)
7720 * A buffer is *not* eligible for the L2ARC if it:
7721 * 1. belongs to a different spa.
7722 * 2. is already cached on the L2ARC.
7723 * 3. has an I/O in progress (it may be an incomplete read).
7724 * 4. is flagged not eligible (zfs property).
7726 if (hdr
->b_spa
!= spa_guid
|| HDR_HAS_L2HDR(hdr
) ||
7727 HDR_IO_IN_PROGRESS(hdr
) || !HDR_L2CACHE(hdr
))
7734 l2arc_write_size(void)
7739 * Make sure our globals have meaningful values in case the user
7742 size
= l2arc_write_max
;
7744 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
7745 "be greater than zero, resetting it to the default (%d)",
7747 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
7750 if (arc_warm
== B_FALSE
)
7751 size
+= l2arc_write_boost
;
7758 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
7760 clock_t interval
, next
, now
;
7763 * If the ARC lists are busy, increase our write rate; if the
7764 * lists are stale, idle back. This is achieved by checking
7765 * how much we previously wrote - if it was more than half of
7766 * what we wanted, schedule the next write much sooner.
7768 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
7769 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
7771 interval
= hz
* l2arc_feed_secs
;
7773 now
= ddi_get_lbolt();
7774 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
7780 * Cycle through L2ARC devices. This is how L2ARC load balances.
7781 * If a device is returned, this also returns holding the spa config lock.
7783 static l2arc_dev_t
*
7784 l2arc_dev_get_next(void)
7786 l2arc_dev_t
*first
, *next
= NULL
;
7789 * Lock out the removal of spas (spa_namespace_lock), then removal
7790 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7791 * both locks will be dropped and a spa config lock held instead.
7793 mutex_enter(&spa_namespace_lock
);
7794 mutex_enter(&l2arc_dev_mtx
);
7796 /* if there are no vdevs, there is nothing to do */
7797 if (l2arc_ndev
== 0)
7801 next
= l2arc_dev_last
;
7803 /* loop around the list looking for a non-faulted vdev */
7805 next
= list_head(l2arc_dev_list
);
7807 next
= list_next(l2arc_dev_list
, next
);
7809 next
= list_head(l2arc_dev_list
);
7812 /* if we have come back to the start, bail out */
7815 else if (next
== first
)
7818 } while (vdev_is_dead(next
->l2ad_vdev
));
7820 /* if we were unable to find any usable vdevs, return NULL */
7821 if (vdev_is_dead(next
->l2ad_vdev
))
7824 l2arc_dev_last
= next
;
7827 mutex_exit(&l2arc_dev_mtx
);
7830 * Grab the config lock to prevent the 'next' device from being
7831 * removed while we are writing to it.
7834 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
7835 mutex_exit(&spa_namespace_lock
);
7841 * Free buffers that were tagged for destruction.
7844 l2arc_do_free_on_write(void)
7847 l2arc_data_free_t
*df
, *df_prev
;
7849 mutex_enter(&l2arc_free_on_write_mtx
);
7850 buflist
= l2arc_free_on_write
;
7852 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
7853 df_prev
= list_prev(buflist
, df
);
7854 ASSERT3P(df
->l2df_abd
, !=, NULL
);
7855 abd_free(df
->l2df_abd
);
7856 list_remove(buflist
, df
);
7857 kmem_free(df
, sizeof (l2arc_data_free_t
));
7860 mutex_exit(&l2arc_free_on_write_mtx
);
7864 * A write to a cache device has completed. Update all headers to allow
7865 * reads from these buffers to begin.
7868 l2arc_write_done(zio_t
*zio
)
7870 l2arc_write_callback_t
*cb
;
7873 arc_buf_hdr_t
*head
, *hdr
, *hdr_prev
;
7874 kmutex_t
*hash_lock
;
7875 int64_t bytes_dropped
= 0;
7877 cb
= zio
->io_private
;
7878 ASSERT3P(cb
, !=, NULL
);
7879 dev
= cb
->l2wcb_dev
;
7880 ASSERT3P(dev
, !=, NULL
);
7881 head
= cb
->l2wcb_head
;
7882 ASSERT3P(head
, !=, NULL
);
7883 buflist
= &dev
->l2ad_buflist
;
7884 ASSERT3P(buflist
, !=, NULL
);
7885 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
7886 l2arc_write_callback_t
*, cb
);
7888 if (zio
->io_error
!= 0)
7889 ARCSTAT_BUMP(arcstat_l2_writes_error
);
7892 * All writes completed, or an error was hit.
7895 mutex_enter(&dev
->l2ad_mtx
);
7896 for (hdr
= list_prev(buflist
, head
); hdr
; hdr
= hdr_prev
) {
7897 hdr_prev
= list_prev(buflist
, hdr
);
7899 hash_lock
= HDR_LOCK(hdr
);
7902 * We cannot use mutex_enter or else we can deadlock
7903 * with l2arc_write_buffers (due to swapping the order
7904 * the hash lock and l2ad_mtx are taken).
7906 if (!mutex_tryenter(hash_lock
)) {
7908 * Missed the hash lock. We must retry so we
7909 * don't leave the ARC_FLAG_L2_WRITING bit set.
7911 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry
);
7914 * We don't want to rescan the headers we've
7915 * already marked as having been written out, so
7916 * we reinsert the head node so we can pick up
7917 * where we left off.
7919 list_remove(buflist
, head
);
7920 list_insert_after(buflist
, hdr
, head
);
7922 mutex_exit(&dev
->l2ad_mtx
);
7925 * We wait for the hash lock to become available
7926 * to try and prevent busy waiting, and increase
7927 * the chance we'll be able to acquire the lock
7928 * the next time around.
7930 mutex_enter(hash_lock
);
7931 mutex_exit(hash_lock
);
7936 * We could not have been moved into the arc_l2c_only
7937 * state while in-flight due to our ARC_FLAG_L2_WRITING
7938 * bit being set. Let's just ensure that's being enforced.
7940 ASSERT(HDR_HAS_L1HDR(hdr
));
7943 * Skipped - drop L2ARC entry and mark the header as no
7944 * longer L2 eligibile.
7946 if (zio
->io_error
!= 0) {
7948 * Error - drop L2ARC entry.
7950 list_remove(buflist
, hdr
);
7951 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
7953 ARCSTAT_INCR(arcstat_l2_psize
, -arc_hdr_size(hdr
));
7954 ARCSTAT_INCR(arcstat_l2_lsize
, -HDR_GET_LSIZE(hdr
));
7956 bytes_dropped
+= arc_hdr_size(hdr
);
7957 (void) refcount_remove_many(&dev
->l2ad_alloc
,
7958 arc_hdr_size(hdr
), hdr
);
7962 * Allow ARC to begin reads and ghost list evictions to
7965 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_WRITING
);
7967 mutex_exit(hash_lock
);
7970 atomic_inc_64(&l2arc_writes_done
);
7971 list_remove(buflist
, head
);
7972 ASSERT(!HDR_HAS_L1HDR(head
));
7973 kmem_cache_free(hdr_l2only_cache
, head
);
7974 mutex_exit(&dev
->l2ad_mtx
);
7976 vdev_space_update(dev
->l2ad_vdev
, -bytes_dropped
, 0, 0);
7978 l2arc_do_free_on_write();
7980 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
7984 l2arc_untransform(zio_t
*zio
, l2arc_read_callback_t
*cb
)
7987 spa_t
*spa
= zio
->io_spa
;
7988 arc_buf_hdr_t
*hdr
= cb
->l2rcb_hdr
;
7989 blkptr_t
*bp
= zio
->io_bp
;
7990 dsl_crypto_key_t
*dck
= NULL
;
7991 uint8_t salt
[ZIO_DATA_SALT_LEN
];
7992 uint8_t iv
[ZIO_DATA_IV_LEN
];
7993 uint8_t mac
[ZIO_DATA_MAC_LEN
];
7994 boolean_t no_crypt
= B_FALSE
;
7997 * ZIL data is never be written to the L2ARC, so we don't need
7998 * special handling for its unique MAC storage.
8000 ASSERT3U(BP_GET_TYPE(bp
), !=, DMU_OT_INTENT_LOG
);
8001 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
8002 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
8005 * If the data was encrypted, decrypt it now. Note that
8006 * we must check the bp here and not the hdr, since the
8007 * hdr does not have its encryption parameters updated
8008 * until arc_read_done().
8010 if (BP_IS_ENCRYPTED(bp
)) {
8011 abd_t
*eabd
= arc_get_data_abd(hdr
,
8012 arc_hdr_size(hdr
), hdr
);
8014 zio_crypt_decode_params_bp(bp
, salt
, iv
);
8015 zio_crypt_decode_mac_bp(bp
, mac
);
8017 ret
= spa_keystore_lookup_key(spa
,
8018 cb
->l2rcb_zb
.zb_objset
, FTAG
, &dck
);
8020 arc_free_data_abd(hdr
, eabd
, arc_hdr_size(hdr
), hdr
);
8024 ret
= zio_do_crypt_abd(B_FALSE
, &dck
->dck_key
,
8025 salt
, BP_GET_TYPE(bp
), iv
, mac
, HDR_GET_PSIZE(hdr
),
8026 BP_SHOULD_BYTESWAP(bp
), eabd
, hdr
->b_l1hdr
.b_pabd
,
8029 arc_free_data_abd(hdr
, eabd
, arc_hdr_size(hdr
), hdr
);
8030 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
8034 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
8037 * If we actually performed decryption, replace b_pabd
8038 * with the decrypted data. Otherwise we can just throw
8039 * our decryption buffer away.
8042 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
,
8043 arc_hdr_size(hdr
), hdr
);
8044 hdr
->b_l1hdr
.b_pabd
= eabd
;
8047 arc_free_data_abd(hdr
, eabd
, arc_hdr_size(hdr
), hdr
);
8052 * If the L2ARC block was compressed, but ARC compression
8053 * is disabled we decompress the data into a new buffer and
8054 * replace the existing data.
8056 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
8057 !HDR_COMPRESSION_ENABLED(hdr
)) {
8058 abd_t
*cabd
= arc_get_data_abd(hdr
, arc_hdr_size(hdr
), hdr
);
8059 void *tmp
= abd_borrow_buf(cabd
, arc_hdr_size(hdr
));
8061 ret
= zio_decompress_data(HDR_GET_COMPRESS(hdr
),
8062 hdr
->b_l1hdr
.b_pabd
, tmp
, HDR_GET_PSIZE(hdr
),
8063 HDR_GET_LSIZE(hdr
));
8065 abd_return_buf_copy(cabd
, tmp
, arc_hdr_size(hdr
));
8066 arc_free_data_abd(hdr
, cabd
, arc_hdr_size(hdr
), hdr
);
8070 abd_return_buf_copy(cabd
, tmp
, arc_hdr_size(hdr
));
8071 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
,
8072 arc_hdr_size(hdr
), hdr
);
8073 hdr
->b_l1hdr
.b_pabd
= cabd
;
8075 zio
->io_size
= HDR_GET_LSIZE(hdr
);
8086 * A read to a cache device completed. Validate buffer contents before
8087 * handing over to the regular ARC routines.
8090 l2arc_read_done(zio_t
*zio
)
8093 l2arc_read_callback_t
*cb
;
8095 kmutex_t
*hash_lock
;
8096 boolean_t valid_cksum
, using_rdata
;
8098 ASSERT3P(zio
->io_vd
, !=, NULL
);
8099 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
8101 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
8103 cb
= zio
->io_private
;
8104 ASSERT3P(cb
, !=, NULL
);
8105 hdr
= cb
->l2rcb_hdr
;
8106 ASSERT3P(hdr
, !=, NULL
);
8108 hash_lock
= HDR_LOCK(hdr
);
8109 mutex_enter(hash_lock
);
8110 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
8113 * If the data was read into a temporary buffer,
8114 * move it and free the buffer.
8116 if (cb
->l2rcb_abd
!= NULL
) {
8117 ASSERT3U(arc_hdr_size(hdr
), <, zio
->io_size
);
8118 if (zio
->io_error
== 0) {
8119 abd_copy(hdr
->b_l1hdr
.b_pabd
, cb
->l2rcb_abd
,
8124 * The following must be done regardless of whether
8125 * there was an error:
8126 * - free the temporary buffer
8127 * - point zio to the real ARC buffer
8128 * - set zio size accordingly
8129 * These are required because zio is either re-used for
8130 * an I/O of the block in the case of the error
8131 * or the zio is passed to arc_read_done() and it
8134 abd_free(cb
->l2rcb_abd
);
8135 zio
->io_size
= zio
->io_orig_size
= arc_hdr_size(hdr
);
8137 if (BP_IS_ENCRYPTED(&cb
->l2rcb_bp
) &&
8138 (cb
->l2rcb_flags
& ZIO_FLAG_RAW_ENCRYPT
)) {
8139 ASSERT(HDR_HAS_RABD(hdr
));
8140 zio
->io_abd
= zio
->io_orig_abd
=
8141 hdr
->b_crypt_hdr
.b_rabd
;
8143 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
8144 zio
->io_abd
= zio
->io_orig_abd
= hdr
->b_l1hdr
.b_pabd
;
8148 ASSERT3P(zio
->io_abd
, !=, NULL
);
8151 * Check this survived the L2ARC journey.
8153 ASSERT(zio
->io_abd
== hdr
->b_l1hdr
.b_pabd
||
8154 (HDR_HAS_RABD(hdr
) && zio
->io_abd
== hdr
->b_crypt_hdr
.b_rabd
));
8155 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
8156 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
8158 valid_cksum
= arc_cksum_is_equal(hdr
, zio
);
8159 using_rdata
= (HDR_HAS_RABD(hdr
) &&
8160 zio
->io_abd
== hdr
->b_crypt_hdr
.b_rabd
);
8163 * b_rabd will always match the data as it exists on disk if it is
8164 * being used. Therefore if we are reading into b_rabd we do not
8165 * attempt to untransform the data.
8167 if (valid_cksum
&& !using_rdata
)
8168 tfm_error
= l2arc_untransform(zio
, cb
);
8170 if (valid_cksum
&& tfm_error
== 0 && zio
->io_error
== 0 &&
8171 !HDR_L2_EVICTED(hdr
)) {
8172 mutex_exit(hash_lock
);
8173 zio
->io_private
= hdr
;
8176 mutex_exit(hash_lock
);
8178 * Buffer didn't survive caching. Increment stats and
8179 * reissue to the original storage device.
8181 if (zio
->io_error
!= 0) {
8182 ARCSTAT_BUMP(arcstat_l2_io_error
);
8184 zio
->io_error
= SET_ERROR(EIO
);
8186 if (!valid_cksum
|| tfm_error
!= 0)
8187 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
8190 * If there's no waiter, issue an async i/o to the primary
8191 * storage now. If there *is* a waiter, the caller must
8192 * issue the i/o in a context where it's OK to block.
8194 if (zio
->io_waiter
== NULL
) {
8195 zio_t
*pio
= zio_unique_parent(zio
);
8196 void *abd
= (using_rdata
) ?
8197 hdr
->b_crypt_hdr
.b_rabd
: hdr
->b_l1hdr
.b_pabd
;
8199 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
8201 zio_nowait(zio_read(pio
, zio
->io_spa
, zio
->io_bp
,
8202 abd
, zio
->io_size
, arc_read_done
,
8203 hdr
, zio
->io_priority
, cb
->l2rcb_flags
,
8208 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
8212 * This is the list priority from which the L2ARC will search for pages to
8213 * cache. This is used within loops (0..3) to cycle through lists in the
8214 * desired order. This order can have a significant effect on cache
8217 * Currently the metadata lists are hit first, MFU then MRU, followed by
8218 * the data lists. This function returns a locked list, and also returns
8221 static multilist_sublist_t
*
8222 l2arc_sublist_lock(int list_num
)
8224 multilist_t
*ml
= NULL
;
8227 ASSERT(list_num
>= 0 && list_num
< L2ARC_FEED_TYPES
);
8231 ml
= arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
8234 ml
= arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
8237 ml
= arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
8240 ml
= arc_mru
->arcs_list
[ARC_BUFC_DATA
];
8247 * Return a randomly-selected sublist. This is acceptable
8248 * because the caller feeds only a little bit of data for each
8249 * call (8MB). Subsequent calls will result in different
8250 * sublists being selected.
8252 idx
= multilist_get_random_index(ml
);
8253 return (multilist_sublist_lock(ml
, idx
));
8257 * Evict buffers from the device write hand to the distance specified in
8258 * bytes. This distance may span populated buffers, it may span nothing.
8259 * This is clearing a region on the L2ARC device ready for writing.
8260 * If the 'all' boolean is set, every buffer is evicted.
8263 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
8266 arc_buf_hdr_t
*hdr
, *hdr_prev
;
8267 kmutex_t
*hash_lock
;
8270 buflist
= &dev
->l2ad_buflist
;
8272 if (!all
&& dev
->l2ad_first
) {
8274 * This is the first sweep through the device. There is
8280 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
8282 * When nearing the end of the device, evict to the end
8283 * before the device write hand jumps to the start.
8285 taddr
= dev
->l2ad_end
;
8287 taddr
= dev
->l2ad_hand
+ distance
;
8289 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
8290 uint64_t, taddr
, boolean_t
, all
);
8293 mutex_enter(&dev
->l2ad_mtx
);
8294 for (hdr
= list_tail(buflist
); hdr
; hdr
= hdr_prev
) {
8295 hdr_prev
= list_prev(buflist
, hdr
);
8297 hash_lock
= HDR_LOCK(hdr
);
8300 * We cannot use mutex_enter or else we can deadlock
8301 * with l2arc_write_buffers (due to swapping the order
8302 * the hash lock and l2ad_mtx are taken).
8304 if (!mutex_tryenter(hash_lock
)) {
8306 * Missed the hash lock. Retry.
8308 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
8309 mutex_exit(&dev
->l2ad_mtx
);
8310 mutex_enter(hash_lock
);
8311 mutex_exit(hash_lock
);
8316 * A header can't be on this list if it doesn't have L2 header.
8318 ASSERT(HDR_HAS_L2HDR(hdr
));
8320 /* Ensure this header has finished being written. */
8321 ASSERT(!HDR_L2_WRITING(hdr
));
8322 ASSERT(!HDR_L2_WRITE_HEAD(hdr
));
8324 if (!all
&& (hdr
->b_l2hdr
.b_daddr
>= taddr
||
8325 hdr
->b_l2hdr
.b_daddr
< dev
->l2ad_hand
)) {
8327 * We've evicted to the target address,
8328 * or the end of the device.
8330 mutex_exit(hash_lock
);
8334 if (!HDR_HAS_L1HDR(hdr
)) {
8335 ASSERT(!HDR_L2_READING(hdr
));
8337 * This doesn't exist in the ARC. Destroy.
8338 * arc_hdr_destroy() will call list_remove()
8339 * and decrement arcstat_l2_lsize.
8341 arc_change_state(arc_anon
, hdr
, hash_lock
);
8342 arc_hdr_destroy(hdr
);
8344 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_l2c_only
);
8345 ARCSTAT_BUMP(arcstat_l2_evict_l1cached
);
8347 * Invalidate issued or about to be issued
8348 * reads, since we may be about to write
8349 * over this location.
8351 if (HDR_L2_READING(hdr
)) {
8352 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
8353 arc_hdr_set_flags(hdr
, ARC_FLAG_L2_EVICTED
);
8356 arc_hdr_l2hdr_destroy(hdr
);
8358 mutex_exit(hash_lock
);
8360 mutex_exit(&dev
->l2ad_mtx
);
8364 * Handle any abd transforms that might be required for writing to the L2ARC.
8365 * If successful, this function will always return an abd with the data
8366 * transformed as it is on disk in a new abd of asize bytes.
8369 l2arc_apply_transforms(spa_t
*spa
, arc_buf_hdr_t
*hdr
, uint64_t asize
,
8374 abd_t
*cabd
= NULL
, *eabd
= NULL
, *to_write
= hdr
->b_l1hdr
.b_pabd
;
8375 enum zio_compress compress
= HDR_GET_COMPRESS(hdr
);
8376 uint64_t psize
= HDR_GET_PSIZE(hdr
);
8377 uint64_t size
= arc_hdr_size(hdr
);
8378 boolean_t ismd
= HDR_ISTYPE_METADATA(hdr
);
8379 boolean_t bswap
= (hdr
->b_l1hdr
.b_byteswap
!= DMU_BSWAP_NUMFUNCS
);
8380 dsl_crypto_key_t
*dck
= NULL
;
8381 uint8_t mac
[ZIO_DATA_MAC_LEN
] = { 0 };
8382 boolean_t no_crypt
= B_FALSE
;
8384 ASSERT((HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
8385 !HDR_COMPRESSION_ENABLED(hdr
)) ||
8386 HDR_ENCRYPTED(hdr
) || HDR_SHARED_DATA(hdr
) || psize
!= asize
);
8387 ASSERT3U(psize
, <=, asize
);
8390 * If this data simply needs its own buffer, we simply allocate it
8391 * and copy the data. This may be done to elimiate a depedency on a
8392 * shared buffer or to reallocate the buffer to match asize.
8394 if (HDR_HAS_RABD(hdr
) && asize
!= psize
) {
8395 ASSERT3U(size
, ==, psize
);
8396 to_write
= abd_alloc_for_io(asize
, ismd
);
8397 abd_copy(to_write
, hdr
->b_crypt_hdr
.b_rabd
, size
);
8399 abd_zero_off(to_write
, size
, asize
- size
);
8403 if ((compress
== ZIO_COMPRESS_OFF
|| HDR_COMPRESSION_ENABLED(hdr
)) &&
8404 !HDR_ENCRYPTED(hdr
)) {
8405 ASSERT3U(size
, ==, psize
);
8406 to_write
= abd_alloc_for_io(asize
, ismd
);
8407 abd_copy(to_write
, hdr
->b_l1hdr
.b_pabd
, size
);
8409 abd_zero_off(to_write
, size
, asize
- size
);
8413 if (compress
!= ZIO_COMPRESS_OFF
&& !HDR_COMPRESSION_ENABLED(hdr
)) {
8414 cabd
= abd_alloc_for_io(asize
, ismd
);
8415 tmp
= abd_borrow_buf(cabd
, asize
);
8417 psize
= zio_compress_data(compress
, to_write
, tmp
, size
);
8418 ASSERT3U(psize
, <=, HDR_GET_PSIZE(hdr
));
8420 bzero((char *)tmp
+ psize
, asize
- psize
);
8421 psize
= HDR_GET_PSIZE(hdr
);
8422 abd_return_buf_copy(cabd
, tmp
, asize
);
8426 if (HDR_ENCRYPTED(hdr
)) {
8427 eabd
= abd_alloc_for_io(asize
, ismd
);
8430 * If the dataset was disowned before the buffer
8431 * made it to this point, the key to re-encrypt
8432 * it won't be available. In this case we simply
8433 * won't write the buffer to the L2ARC.
8435 ret
= spa_keystore_lookup_key(spa
, hdr
->b_crypt_hdr
.b_dsobj
,
8440 ret
= zio_do_crypt_abd(B_TRUE
, &dck
->dck_key
,
8441 hdr
->b_crypt_hdr
.b_salt
, hdr
->b_crypt_hdr
.b_ot
,
8442 hdr
->b_crypt_hdr
.b_iv
, mac
, psize
, bswap
, to_write
,
8448 abd_copy(eabd
, to_write
, psize
);
8451 abd_zero_off(eabd
, psize
, asize
- psize
);
8453 /* assert that the MAC we got here matches the one we saved */
8454 ASSERT0(bcmp(mac
, hdr
->b_crypt_hdr
.b_mac
, ZIO_DATA_MAC_LEN
));
8455 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
8457 if (to_write
== cabd
)
8464 ASSERT3P(to_write
, !=, hdr
->b_l1hdr
.b_pabd
);
8465 *abd_out
= to_write
;
8470 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
8481 * Find and write ARC buffers to the L2ARC device.
8483 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
8484 * for reading until they have completed writing.
8485 * The headroom_boost is an in-out parameter used to maintain headroom boost
8486 * state between calls to this function.
8488 * Returns the number of bytes actually written (which may be smaller than
8489 * the delta by which the device hand has changed due to alignment).
8492 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
)
8494 arc_buf_hdr_t
*hdr
, *hdr_prev
, *head
;
8495 uint64_t write_asize
, write_psize
, write_lsize
, headroom
;
8497 l2arc_write_callback_t
*cb
;
8499 uint64_t guid
= spa_load_guid(spa
);
8501 ASSERT3P(dev
->l2ad_vdev
, !=, NULL
);
8504 write_lsize
= write_asize
= write_psize
= 0;
8506 head
= kmem_cache_alloc(hdr_l2only_cache
, KM_PUSHPAGE
);
8507 arc_hdr_set_flags(head
, ARC_FLAG_L2_WRITE_HEAD
| ARC_FLAG_HAS_L2HDR
);
8510 * Copy buffers for L2ARC writing.
8512 for (int try = 0; try < L2ARC_FEED_TYPES
; try++) {
8513 multilist_sublist_t
*mls
= l2arc_sublist_lock(try);
8514 uint64_t passed_sz
= 0;
8516 VERIFY3P(mls
, !=, NULL
);
8519 * L2ARC fast warmup.
8521 * Until the ARC is warm and starts to evict, read from the
8522 * head of the ARC lists rather than the tail.
8524 if (arc_warm
== B_FALSE
)
8525 hdr
= multilist_sublist_head(mls
);
8527 hdr
= multilist_sublist_tail(mls
);
8529 headroom
= target_sz
* l2arc_headroom
;
8530 if (zfs_compressed_arc_enabled
)
8531 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
8533 for (; hdr
; hdr
= hdr_prev
) {
8534 kmutex_t
*hash_lock
;
8535 abd_t
*to_write
= NULL
;
8537 if (arc_warm
== B_FALSE
)
8538 hdr_prev
= multilist_sublist_next(mls
, hdr
);
8540 hdr_prev
= multilist_sublist_prev(mls
, hdr
);
8542 hash_lock
= HDR_LOCK(hdr
);
8543 if (!mutex_tryenter(hash_lock
)) {
8545 * Skip this buffer rather than waiting.
8550 passed_sz
+= HDR_GET_LSIZE(hdr
);
8551 if (passed_sz
> headroom
) {
8555 mutex_exit(hash_lock
);
8559 if (!l2arc_write_eligible(guid
, hdr
)) {
8560 mutex_exit(hash_lock
);
8565 * We rely on the L1 portion of the header below, so
8566 * it's invalid for this header to have been evicted out
8567 * of the ghost cache, prior to being written out. The
8568 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8570 ASSERT(HDR_HAS_L1HDR(hdr
));
8572 ASSERT3U(HDR_GET_PSIZE(hdr
), >, 0);
8573 ASSERT3U(arc_hdr_size(hdr
), >, 0);
8574 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
||
8576 uint64_t psize
= HDR_GET_PSIZE(hdr
);
8577 uint64_t asize
= vdev_psize_to_asize(dev
->l2ad_vdev
,
8580 if ((write_asize
+ asize
) > target_sz
) {
8582 mutex_exit(hash_lock
);
8587 * We rely on the L1 portion of the header below, so
8588 * it's invalid for this header to have been evicted out
8589 * of the ghost cache, prior to being written out. The
8590 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8592 arc_hdr_set_flags(hdr
, ARC_FLAG_L2_WRITING
);
8593 ASSERT(HDR_HAS_L1HDR(hdr
));
8595 ASSERT3U(HDR_GET_PSIZE(hdr
), >, 0);
8596 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
||
8598 ASSERT3U(arc_hdr_size(hdr
), >, 0);
8601 * If this header has b_rabd, we can use this since it
8602 * must always match the data exactly as it exists on
8603 * disk. Otherwise, the L2ARC can normally use the
8604 * hdr's data, but if we're sharing data between the
8605 * hdr and one of its bufs, L2ARC needs its own copy of
8606 * the data so that the ZIO below can't race with the
8607 * buf consumer. To ensure that this copy will be
8608 * available for the lifetime of the ZIO and be cleaned
8609 * up afterwards, we add it to the l2arc_free_on_write
8610 * queue. If we need to apply any transforms to the
8611 * data (compression, encryption) we will also need the
8614 if (HDR_HAS_RABD(hdr
) && psize
== asize
) {
8615 to_write
= hdr
->b_crypt_hdr
.b_rabd
;
8616 } else if ((HDR_COMPRESSION_ENABLED(hdr
) ||
8617 HDR_GET_COMPRESS(hdr
) == ZIO_COMPRESS_OFF
) &&
8618 !HDR_ENCRYPTED(hdr
) && !HDR_SHARED_DATA(hdr
) &&
8620 to_write
= hdr
->b_l1hdr
.b_pabd
;
8623 arc_buf_contents_t type
= arc_buf_type(hdr
);
8625 ret
= l2arc_apply_transforms(spa
, hdr
, asize
,
8628 arc_hdr_clear_flags(hdr
,
8629 ARC_FLAG_L2_WRITING
);
8630 mutex_exit(hash_lock
);
8634 l2arc_free_abd_on_write(to_write
, asize
, type
);
8639 * Insert a dummy header on the buflist so
8640 * l2arc_write_done() can find where the
8641 * write buffers begin without searching.
8643 mutex_enter(&dev
->l2ad_mtx
);
8644 list_insert_head(&dev
->l2ad_buflist
, head
);
8645 mutex_exit(&dev
->l2ad_mtx
);
8648 sizeof (l2arc_write_callback_t
), KM_SLEEP
);
8649 cb
->l2wcb_dev
= dev
;
8650 cb
->l2wcb_head
= head
;
8651 pio
= zio_root(spa
, l2arc_write_done
, cb
,
8655 hdr
->b_l2hdr
.b_dev
= dev
;
8656 hdr
->b_l2hdr
.b_hits
= 0;
8658 hdr
->b_l2hdr
.b_daddr
= dev
->l2ad_hand
;
8659 arc_hdr_set_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
8661 mutex_enter(&dev
->l2ad_mtx
);
8662 list_insert_head(&dev
->l2ad_buflist
, hdr
);
8663 mutex_exit(&dev
->l2ad_mtx
);
8665 (void) refcount_add_many(&dev
->l2ad_alloc
,
8666 arc_hdr_size(hdr
), hdr
);
8668 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
8669 hdr
->b_l2hdr
.b_daddr
, asize
, to_write
,
8670 ZIO_CHECKSUM_OFF
, NULL
, hdr
,
8671 ZIO_PRIORITY_ASYNC_WRITE
,
8672 ZIO_FLAG_CANFAIL
, B_FALSE
);
8674 write_lsize
+= HDR_GET_LSIZE(hdr
);
8675 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
8678 write_psize
+= psize
;
8679 write_asize
+= asize
;
8680 dev
->l2ad_hand
+= asize
;
8682 mutex_exit(hash_lock
);
8684 (void) zio_nowait(wzio
);
8687 multilist_sublist_unlock(mls
);
8693 /* No buffers selected for writing? */
8695 ASSERT0(write_lsize
);
8696 ASSERT(!HDR_HAS_L1HDR(head
));
8697 kmem_cache_free(hdr_l2only_cache
, head
);
8701 ASSERT3U(write_asize
, <=, target_sz
);
8702 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
8703 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_psize
);
8704 ARCSTAT_INCR(arcstat_l2_lsize
, write_lsize
);
8705 ARCSTAT_INCR(arcstat_l2_psize
, write_psize
);
8706 vdev_space_update(dev
->l2ad_vdev
, write_psize
, 0, 0);
8709 * Bump device hand to the device start if it is approaching the end.
8710 * l2arc_evict() will already have evicted ahead for this case.
8712 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
8713 dev
->l2ad_hand
= dev
->l2ad_start
;
8714 dev
->l2ad_first
= B_FALSE
;
8717 dev
->l2ad_writing
= B_TRUE
;
8718 (void) zio_wait(pio
);
8719 dev
->l2ad_writing
= B_FALSE
;
8721 return (write_asize
);
8725 * This thread feeds the L2ARC at regular intervals. This is the beating
8726 * heart of the L2ARC.
8730 l2arc_feed_thread(void *unused
)
8735 uint64_t size
, wrote
;
8736 clock_t begin
, next
= ddi_get_lbolt();
8737 fstrans_cookie_t cookie
;
8739 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
8741 mutex_enter(&l2arc_feed_thr_lock
);
8743 cookie
= spl_fstrans_mark();
8744 while (l2arc_thread_exit
== 0) {
8745 CALLB_CPR_SAFE_BEGIN(&cpr
);
8746 (void) cv_timedwait_sig(&l2arc_feed_thr_cv
,
8747 &l2arc_feed_thr_lock
, next
);
8748 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
8749 next
= ddi_get_lbolt() + hz
;
8752 * Quick check for L2ARC devices.
8754 mutex_enter(&l2arc_dev_mtx
);
8755 if (l2arc_ndev
== 0) {
8756 mutex_exit(&l2arc_dev_mtx
);
8759 mutex_exit(&l2arc_dev_mtx
);
8760 begin
= ddi_get_lbolt();
8763 * This selects the next l2arc device to write to, and in
8764 * doing so the next spa to feed from: dev->l2ad_spa. This
8765 * will return NULL if there are now no l2arc devices or if
8766 * they are all faulted.
8768 * If a device is returned, its spa's config lock is also
8769 * held to prevent device removal. l2arc_dev_get_next()
8770 * will grab and release l2arc_dev_mtx.
8772 if ((dev
= l2arc_dev_get_next()) == NULL
)
8775 spa
= dev
->l2ad_spa
;
8776 ASSERT3P(spa
, !=, NULL
);
8779 * If the pool is read-only then force the feed thread to
8780 * sleep a little longer.
8782 if (!spa_writeable(spa
)) {
8783 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
8784 spa_config_exit(spa
, SCL_L2ARC
, dev
);
8789 * Avoid contributing to memory pressure.
8791 if (arc_reclaim_needed()) {
8792 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
8793 spa_config_exit(spa
, SCL_L2ARC
, dev
);
8797 ARCSTAT_BUMP(arcstat_l2_feeds
);
8799 size
= l2arc_write_size();
8802 * Evict L2ARC buffers that will be overwritten.
8804 l2arc_evict(dev
, size
, B_FALSE
);
8807 * Write ARC buffers.
8809 wrote
= l2arc_write_buffers(spa
, dev
, size
);
8812 * Calculate interval between writes.
8814 next
= l2arc_write_interval(begin
, size
, wrote
);
8815 spa_config_exit(spa
, SCL_L2ARC
, dev
);
8817 spl_fstrans_unmark(cookie
);
8819 l2arc_thread_exit
= 0;
8820 cv_broadcast(&l2arc_feed_thr_cv
);
8821 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
8826 l2arc_vdev_present(vdev_t
*vd
)
8830 mutex_enter(&l2arc_dev_mtx
);
8831 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
8832 dev
= list_next(l2arc_dev_list
, dev
)) {
8833 if (dev
->l2ad_vdev
== vd
)
8836 mutex_exit(&l2arc_dev_mtx
);
8838 return (dev
!= NULL
);
8842 * Add a vdev for use by the L2ARC. By this point the spa has already
8843 * validated the vdev and opened it.
8846 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
8848 l2arc_dev_t
*adddev
;
8850 ASSERT(!l2arc_vdev_present(vd
));
8853 * Create a new l2arc device entry.
8855 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
8856 adddev
->l2ad_spa
= spa
;
8857 adddev
->l2ad_vdev
= vd
;
8858 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
8859 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
8860 adddev
->l2ad_hand
= adddev
->l2ad_start
;
8861 adddev
->l2ad_first
= B_TRUE
;
8862 adddev
->l2ad_writing
= B_FALSE
;
8863 list_link_init(&adddev
->l2ad_node
);
8865 mutex_init(&adddev
->l2ad_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
8867 * This is a list of all ARC buffers that are still valid on the
8870 list_create(&adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
8871 offsetof(arc_buf_hdr_t
, b_l2hdr
.b_l2node
));
8873 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
8874 refcount_create(&adddev
->l2ad_alloc
);
8877 * Add device to global list
8879 mutex_enter(&l2arc_dev_mtx
);
8880 list_insert_head(l2arc_dev_list
, adddev
);
8881 atomic_inc_64(&l2arc_ndev
);
8882 mutex_exit(&l2arc_dev_mtx
);
8886 * Remove a vdev from the L2ARC.
8889 l2arc_remove_vdev(vdev_t
*vd
)
8891 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
8894 * Find the device by vdev
8896 mutex_enter(&l2arc_dev_mtx
);
8897 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
8898 nextdev
= list_next(l2arc_dev_list
, dev
);
8899 if (vd
== dev
->l2ad_vdev
) {
8904 ASSERT3P(remdev
, !=, NULL
);
8907 * Remove device from global list
8909 list_remove(l2arc_dev_list
, remdev
);
8910 l2arc_dev_last
= NULL
; /* may have been invalidated */
8911 atomic_dec_64(&l2arc_ndev
);
8912 mutex_exit(&l2arc_dev_mtx
);
8915 * Clear all buflists and ARC references. L2ARC device flush.
8917 l2arc_evict(remdev
, 0, B_TRUE
);
8918 list_destroy(&remdev
->l2ad_buflist
);
8919 mutex_destroy(&remdev
->l2ad_mtx
);
8920 refcount_destroy(&remdev
->l2ad_alloc
);
8921 kmem_free(remdev
, sizeof (l2arc_dev_t
));
8927 l2arc_thread_exit
= 0;
8929 l2arc_writes_sent
= 0;
8930 l2arc_writes_done
= 0;
8932 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
8933 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
8934 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
8935 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
8937 l2arc_dev_list
= &L2ARC_dev_list
;
8938 l2arc_free_on_write
= &L2ARC_free_on_write
;
8939 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
8940 offsetof(l2arc_dev_t
, l2ad_node
));
8941 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
8942 offsetof(l2arc_data_free_t
, l2df_list_node
));
8949 * This is called from dmu_fini(), which is called from spa_fini();
8950 * Because of this, we can assume that all l2arc devices have
8951 * already been removed when the pools themselves were removed.
8954 l2arc_do_free_on_write();
8956 mutex_destroy(&l2arc_feed_thr_lock
);
8957 cv_destroy(&l2arc_feed_thr_cv
);
8958 mutex_destroy(&l2arc_dev_mtx
);
8959 mutex_destroy(&l2arc_free_on_write_mtx
);
8961 list_destroy(l2arc_dev_list
);
8962 list_destroy(l2arc_free_on_write
);
8968 if (!(spa_mode_global
& FWRITE
))
8971 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
8972 TS_RUN
, defclsyspri
);
8978 if (!(spa_mode_global
& FWRITE
))
8981 mutex_enter(&l2arc_feed_thr_lock
);
8982 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
8983 l2arc_thread_exit
= 1;
8984 while (l2arc_thread_exit
!= 0)
8985 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
8986 mutex_exit(&l2arc_feed_thr_lock
);
8989 #if defined(_KERNEL) && defined(HAVE_SPL)
8990 EXPORT_SYMBOL(arc_buf_size
);
8991 EXPORT_SYMBOL(arc_write
);
8992 EXPORT_SYMBOL(arc_read
);
8993 EXPORT_SYMBOL(arc_buf_info
);
8994 EXPORT_SYMBOL(arc_getbuf_func
);
8995 EXPORT_SYMBOL(arc_add_prune_callback
);
8996 EXPORT_SYMBOL(arc_remove_prune_callback
);
8999 module_param(zfs_arc_min
, ulong
, 0644);
9000 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
9002 module_param(zfs_arc_max
, ulong
, 0644);
9003 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
9005 module_param(zfs_arc_meta_limit
, ulong
, 0644);
9006 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
9008 module_param(zfs_arc_meta_limit_percent
, ulong
, 0644);
9009 MODULE_PARM_DESC(zfs_arc_meta_limit_percent
,
9010 "Percent of arc size for arc meta limit");
9012 module_param(zfs_arc_meta_min
, ulong
, 0644);
9013 MODULE_PARM_DESC(zfs_arc_meta_min
, "Min arc metadata");
9015 module_param(zfs_arc_meta_prune
, int, 0644);
9016 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Meta objects to scan for prune");
9018 module_param(zfs_arc_meta_adjust_restarts
, int, 0644);
9019 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts
,
9020 "Limit number of restarts in arc_adjust_meta");
9022 module_param(zfs_arc_meta_strategy
, int, 0644);
9023 MODULE_PARM_DESC(zfs_arc_meta_strategy
, "Meta reclaim strategy");
9025 module_param(zfs_arc_grow_retry
, int, 0644);
9026 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
9028 module_param(zfs_arc_p_aggressive_disable
, int, 0644);
9029 MODULE_PARM_DESC(zfs_arc_p_aggressive_disable
, "disable aggressive arc_p grow");
9031 module_param(zfs_arc_p_dampener_disable
, int, 0644);
9032 MODULE_PARM_DESC(zfs_arc_p_dampener_disable
, "disable arc_p adapt dampener");
9034 module_param(zfs_arc_shrink_shift
, int, 0644);
9035 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
9037 module_param(zfs_arc_pc_percent
, uint
, 0644);
9038 MODULE_PARM_DESC(zfs_arc_pc_percent
,
9039 "Percent of pagecache to reclaim arc to");
9041 module_param(zfs_arc_p_min_shift
, int, 0644);
9042 MODULE_PARM_DESC(zfs_arc_p_min_shift
, "arc_c shift to calc min/max arc_p");
9044 module_param(zfs_arc_average_blocksize
, int, 0444);
9045 MODULE_PARM_DESC(zfs_arc_average_blocksize
, "Target average block size");
9047 module_param(zfs_compressed_arc_enabled
, int, 0644);
9048 MODULE_PARM_DESC(zfs_compressed_arc_enabled
, "Disable compressed arc buffers");
9050 module_param(zfs_arc_min_prefetch_ms
, int, 0644);
9051 MODULE_PARM_DESC(zfs_arc_min_prefetch_ms
, "Min life of prefetch block in ms");
9053 module_param(zfs_arc_min_prescient_prefetch_ms
, int, 0644);
9054 MODULE_PARM_DESC(zfs_arc_min_prescient_prefetch_ms
,
9055 "Min life of prescient prefetched block in ms");
9057 module_param(l2arc_write_max
, ulong
, 0644);
9058 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
9060 module_param(l2arc_write_boost
, ulong
, 0644);
9061 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
9063 module_param(l2arc_headroom
, ulong
, 0644);
9064 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
9066 module_param(l2arc_headroom_boost
, ulong
, 0644);
9067 MODULE_PARM_DESC(l2arc_headroom_boost
, "Compressed l2arc_headroom multiplier");
9069 module_param(l2arc_feed_secs
, ulong
, 0644);
9070 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
9072 module_param(l2arc_feed_min_ms
, ulong
, 0644);
9073 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
9075 module_param(l2arc_noprefetch
, int, 0644);
9076 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
9078 module_param(l2arc_feed_again
, int, 0644);
9079 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
9081 module_param(l2arc_norw
, int, 0644);
9082 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");
9084 module_param(zfs_arc_lotsfree_percent
, int, 0644);
9085 MODULE_PARM_DESC(zfs_arc_lotsfree_percent
,
9086 "System free memory I/O throttle in bytes");
9088 module_param(zfs_arc_sys_free
, ulong
, 0644);
9089 MODULE_PARM_DESC(zfs_arc_sys_free
, "System free memory target size in bytes");
9091 module_param(zfs_arc_dnode_limit
, ulong
, 0644);
9092 MODULE_PARM_DESC(zfs_arc_dnode_limit
, "Minimum bytes of dnodes in arc");
9094 module_param(zfs_arc_dnode_limit_percent
, ulong
, 0644);
9095 MODULE_PARM_DESC(zfs_arc_dnode_limit_percent
,
9096 "Percent of ARC meta buffers for dnodes");
9098 module_param(zfs_arc_dnode_reduce_percent
, ulong
, 0644);
9099 MODULE_PARM_DESC(zfs_arc_dnode_reduce_percent
,
9100 "Percentage of excess dnodes to try to unpin");