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_lifespan
;
363 * If this percent of memory is free, don't throttle.
365 int arc_lotsfree_percent
= 10;
370 * The arc has filled available memory and has now warmed up.
372 static boolean_t arc_warm
;
375 * log2 fraction of the zio arena to keep free.
377 int arc_zio_arena_free_shift
= 2;
380 * These tunables are for performance analysis.
382 unsigned long zfs_arc_max
= 0;
383 unsigned long zfs_arc_min
= 0;
384 unsigned long zfs_arc_meta_limit
= 0;
385 unsigned long zfs_arc_meta_min
= 0;
386 unsigned long zfs_arc_dnode_limit
= 0;
387 unsigned long zfs_arc_dnode_reduce_percent
= 10;
388 int zfs_arc_grow_retry
= 0;
389 int zfs_arc_shrink_shift
= 0;
390 int zfs_arc_p_min_shift
= 0;
391 int zfs_arc_average_blocksize
= 8 * 1024; /* 8KB */
393 int zfs_compressed_arc_enabled
= B_TRUE
;
396 * ARC will evict meta buffers that exceed arc_meta_limit. This
397 * tunable make arc_meta_limit adjustable for different workloads.
399 unsigned long zfs_arc_meta_limit_percent
= 75;
402 * Percentage that can be consumed by dnodes of ARC meta buffers.
404 unsigned long zfs_arc_dnode_limit_percent
= 10;
407 * These tunables are Linux specific
409 unsigned long zfs_arc_sys_free
= 0;
410 int zfs_arc_min_prefetch_lifespan
= 0;
411 int zfs_arc_p_aggressive_disable
= 1;
412 int zfs_arc_p_dampener_disable
= 1;
413 int zfs_arc_meta_prune
= 10000;
414 int zfs_arc_meta_strategy
= ARC_STRATEGY_META_BALANCED
;
415 int zfs_arc_meta_adjust_restarts
= 4096;
416 int zfs_arc_lotsfree_percent
= 10;
419 static arc_state_t ARC_anon
;
420 static arc_state_t ARC_mru
;
421 static arc_state_t ARC_mru_ghost
;
422 static arc_state_t ARC_mfu
;
423 static arc_state_t ARC_mfu_ghost
;
424 static arc_state_t ARC_l2c_only
;
426 typedef struct arc_stats
{
427 kstat_named_t arcstat_hits
;
428 kstat_named_t arcstat_misses
;
429 kstat_named_t arcstat_demand_data_hits
;
430 kstat_named_t arcstat_demand_data_misses
;
431 kstat_named_t arcstat_demand_metadata_hits
;
432 kstat_named_t arcstat_demand_metadata_misses
;
433 kstat_named_t arcstat_prefetch_data_hits
;
434 kstat_named_t arcstat_prefetch_data_misses
;
435 kstat_named_t arcstat_prefetch_metadata_hits
;
436 kstat_named_t arcstat_prefetch_metadata_misses
;
437 kstat_named_t arcstat_mru_hits
;
438 kstat_named_t arcstat_mru_ghost_hits
;
439 kstat_named_t arcstat_mfu_hits
;
440 kstat_named_t arcstat_mfu_ghost_hits
;
441 kstat_named_t arcstat_deleted
;
443 * Number of buffers that could not be evicted because the hash lock
444 * was held by another thread. The lock may not necessarily be held
445 * by something using the same buffer, since hash locks are shared
446 * by multiple buffers.
448 kstat_named_t arcstat_mutex_miss
;
450 * Number of buffers skipped because they have I/O in progress, are
451 * indrect prefetch buffers that have not lived long enough, or are
452 * not from the spa we're trying to evict from.
454 kstat_named_t arcstat_evict_skip
;
456 * Number of times arc_evict_state() was unable to evict enough
457 * buffers to reach its target amount.
459 kstat_named_t arcstat_evict_not_enough
;
460 kstat_named_t arcstat_evict_l2_cached
;
461 kstat_named_t arcstat_evict_l2_eligible
;
462 kstat_named_t arcstat_evict_l2_ineligible
;
463 kstat_named_t arcstat_evict_l2_skip
;
464 kstat_named_t arcstat_hash_elements
;
465 kstat_named_t arcstat_hash_elements_max
;
466 kstat_named_t arcstat_hash_collisions
;
467 kstat_named_t arcstat_hash_chains
;
468 kstat_named_t arcstat_hash_chain_max
;
469 kstat_named_t arcstat_p
;
470 kstat_named_t arcstat_c
;
471 kstat_named_t arcstat_c_min
;
472 kstat_named_t arcstat_c_max
;
473 kstat_named_t arcstat_size
;
475 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
476 * Note that the compressed bytes may match the uncompressed bytes
477 * if the block is either not compressed or compressed arc is disabled.
479 kstat_named_t arcstat_compressed_size
;
481 * Uncompressed size of the data stored in b_pabd. If compressed
482 * arc is disabled then this value will be identical to the stat
485 kstat_named_t arcstat_uncompressed_size
;
487 * Number of bytes stored in all the arc_buf_t's. This is classified
488 * as "overhead" since this data is typically short-lived and will
489 * be evicted from the arc when it becomes unreferenced unless the
490 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
491 * values have been set (see comment in dbuf.c for more information).
493 kstat_named_t arcstat_overhead_size
;
495 * Number of bytes consumed by internal ARC structures necessary
496 * for tracking purposes; these structures are not actually
497 * backed by ARC buffers. This includes arc_buf_hdr_t structures
498 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
499 * caches), and arc_buf_t structures (allocated via arc_buf_t
502 kstat_named_t arcstat_hdr_size
;
504 * Number of bytes consumed by ARC buffers of type equal to
505 * ARC_BUFC_DATA. This is generally consumed by buffers backing
506 * on disk user data (e.g. plain file contents).
508 kstat_named_t arcstat_data_size
;
510 * Number of bytes consumed by ARC buffers of type equal to
511 * ARC_BUFC_METADATA. This is generally consumed by buffers
512 * backing on disk data that is used for internal ZFS
513 * structures (e.g. ZAP, dnode, indirect blocks, etc).
515 kstat_named_t arcstat_metadata_size
;
517 * Number of bytes consumed by dmu_buf_impl_t objects.
519 kstat_named_t arcstat_dbuf_size
;
521 * Number of bytes consumed by dnode_t objects.
523 kstat_named_t arcstat_dnode_size
;
525 * Number of bytes consumed by bonus buffers.
527 kstat_named_t arcstat_bonus_size
;
529 * Total number of bytes consumed by ARC buffers residing in the
530 * arc_anon state. This includes *all* buffers in the arc_anon
531 * state; e.g. data, metadata, evictable, and unevictable buffers
532 * are all included in this value.
534 kstat_named_t arcstat_anon_size
;
536 * Number of bytes consumed by ARC buffers that meet the
537 * following criteria: backing buffers of type ARC_BUFC_DATA,
538 * residing in the arc_anon state, and are eligible for eviction
539 * (e.g. have no outstanding holds on the buffer).
541 kstat_named_t arcstat_anon_evictable_data
;
543 * Number of bytes consumed by ARC buffers that meet the
544 * following criteria: backing buffers of type ARC_BUFC_METADATA,
545 * residing in the arc_anon state, and are eligible for eviction
546 * (e.g. have no outstanding holds on the buffer).
548 kstat_named_t arcstat_anon_evictable_metadata
;
550 * Total number of bytes consumed by ARC buffers residing in the
551 * arc_mru state. This includes *all* buffers in the arc_mru
552 * state; e.g. data, metadata, evictable, and unevictable buffers
553 * are all included in this value.
555 kstat_named_t arcstat_mru_size
;
557 * Number of bytes consumed by ARC buffers that meet the
558 * following criteria: backing buffers of type ARC_BUFC_DATA,
559 * residing in the arc_mru state, and are eligible for eviction
560 * (e.g. have no outstanding holds on the buffer).
562 kstat_named_t arcstat_mru_evictable_data
;
564 * Number of bytes consumed by ARC buffers that meet the
565 * following criteria: backing buffers of type ARC_BUFC_METADATA,
566 * residing in the arc_mru state, and are eligible for eviction
567 * (e.g. have no outstanding holds on the buffer).
569 kstat_named_t arcstat_mru_evictable_metadata
;
571 * Total number of bytes that *would have been* consumed by ARC
572 * buffers in the arc_mru_ghost state. The key thing to note
573 * here, is the fact that this size doesn't actually indicate
574 * RAM consumption. The ghost lists only consist of headers and
575 * don't actually have ARC buffers linked off of these headers.
576 * Thus, *if* the headers had associated ARC buffers, these
577 * buffers *would have* consumed this number of bytes.
579 kstat_named_t arcstat_mru_ghost_size
;
581 * Number of bytes that *would have been* consumed by ARC
582 * buffers that are eligible for eviction, of type
583 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
585 kstat_named_t arcstat_mru_ghost_evictable_data
;
587 * Number of bytes that *would have been* consumed by ARC
588 * buffers that are eligible for eviction, of type
589 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
591 kstat_named_t arcstat_mru_ghost_evictable_metadata
;
593 * Total number of bytes consumed by ARC buffers residing in the
594 * arc_mfu state. This includes *all* buffers in the arc_mfu
595 * state; e.g. data, metadata, evictable, and unevictable buffers
596 * are all included in this value.
598 kstat_named_t arcstat_mfu_size
;
600 * Number of bytes consumed by ARC buffers that are eligible for
601 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
604 kstat_named_t arcstat_mfu_evictable_data
;
606 * Number of bytes consumed by ARC buffers that are eligible for
607 * eviction, of type ARC_BUFC_METADATA, and reside in the
610 kstat_named_t arcstat_mfu_evictable_metadata
;
612 * Total number of bytes that *would have been* consumed by ARC
613 * buffers in the arc_mfu_ghost state. See the comment above
614 * arcstat_mru_ghost_size for more details.
616 kstat_named_t arcstat_mfu_ghost_size
;
618 * Number of bytes that *would have been* consumed by ARC
619 * buffers that are eligible for eviction, of type
620 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
622 kstat_named_t arcstat_mfu_ghost_evictable_data
;
624 * Number of bytes that *would have been* consumed by ARC
625 * buffers that are eligible for eviction, of type
626 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
628 kstat_named_t arcstat_mfu_ghost_evictable_metadata
;
629 kstat_named_t arcstat_l2_hits
;
630 kstat_named_t arcstat_l2_misses
;
631 kstat_named_t arcstat_l2_feeds
;
632 kstat_named_t arcstat_l2_rw_clash
;
633 kstat_named_t arcstat_l2_read_bytes
;
634 kstat_named_t arcstat_l2_write_bytes
;
635 kstat_named_t arcstat_l2_writes_sent
;
636 kstat_named_t arcstat_l2_writes_done
;
637 kstat_named_t arcstat_l2_writes_error
;
638 kstat_named_t arcstat_l2_writes_lock_retry
;
639 kstat_named_t arcstat_l2_evict_lock_retry
;
640 kstat_named_t arcstat_l2_evict_reading
;
641 kstat_named_t arcstat_l2_evict_l1cached
;
642 kstat_named_t arcstat_l2_free_on_write
;
643 kstat_named_t arcstat_l2_abort_lowmem
;
644 kstat_named_t arcstat_l2_cksum_bad
;
645 kstat_named_t arcstat_l2_io_error
;
646 kstat_named_t arcstat_l2_lsize
;
647 kstat_named_t arcstat_l2_psize
;
648 kstat_named_t arcstat_l2_hdr_size
;
649 kstat_named_t arcstat_memory_throttle_count
;
650 kstat_named_t arcstat_memory_direct_count
;
651 kstat_named_t arcstat_memory_indirect_count
;
652 kstat_named_t arcstat_memory_all_bytes
;
653 kstat_named_t arcstat_memory_free_bytes
;
654 kstat_named_t arcstat_memory_available_bytes
;
655 kstat_named_t arcstat_no_grow
;
656 kstat_named_t arcstat_tempreserve
;
657 kstat_named_t arcstat_loaned_bytes
;
658 kstat_named_t arcstat_prune
;
659 kstat_named_t arcstat_meta_used
;
660 kstat_named_t arcstat_meta_limit
;
661 kstat_named_t arcstat_dnode_limit
;
662 kstat_named_t arcstat_meta_max
;
663 kstat_named_t arcstat_meta_min
;
664 kstat_named_t arcstat_sync_wait_for_async
;
665 kstat_named_t arcstat_demand_hit_predictive_prefetch
;
666 kstat_named_t arcstat_need_free
;
667 kstat_named_t arcstat_sys_free
;
668 kstat_named_t arcstat_raw_size
;
671 static arc_stats_t arc_stats
= {
672 { "hits", KSTAT_DATA_UINT64
},
673 { "misses", KSTAT_DATA_UINT64
},
674 { "demand_data_hits", KSTAT_DATA_UINT64
},
675 { "demand_data_misses", KSTAT_DATA_UINT64
},
676 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
677 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
678 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
679 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
680 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
681 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
682 { "mru_hits", KSTAT_DATA_UINT64
},
683 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
684 { "mfu_hits", KSTAT_DATA_UINT64
},
685 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
686 { "deleted", KSTAT_DATA_UINT64
},
687 { "mutex_miss", KSTAT_DATA_UINT64
},
688 { "evict_skip", KSTAT_DATA_UINT64
},
689 { "evict_not_enough", KSTAT_DATA_UINT64
},
690 { "evict_l2_cached", KSTAT_DATA_UINT64
},
691 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
692 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
693 { "evict_l2_skip", KSTAT_DATA_UINT64
},
694 { "hash_elements", KSTAT_DATA_UINT64
},
695 { "hash_elements_max", KSTAT_DATA_UINT64
},
696 { "hash_collisions", KSTAT_DATA_UINT64
},
697 { "hash_chains", KSTAT_DATA_UINT64
},
698 { "hash_chain_max", KSTAT_DATA_UINT64
},
699 { "p", KSTAT_DATA_UINT64
},
700 { "c", KSTAT_DATA_UINT64
},
701 { "c_min", KSTAT_DATA_UINT64
},
702 { "c_max", KSTAT_DATA_UINT64
},
703 { "size", KSTAT_DATA_UINT64
},
704 { "compressed_size", KSTAT_DATA_UINT64
},
705 { "uncompressed_size", KSTAT_DATA_UINT64
},
706 { "overhead_size", KSTAT_DATA_UINT64
},
707 { "hdr_size", KSTAT_DATA_UINT64
},
708 { "data_size", KSTAT_DATA_UINT64
},
709 { "metadata_size", KSTAT_DATA_UINT64
},
710 { "dbuf_size", KSTAT_DATA_UINT64
},
711 { "dnode_size", KSTAT_DATA_UINT64
},
712 { "bonus_size", KSTAT_DATA_UINT64
},
713 { "anon_size", KSTAT_DATA_UINT64
},
714 { "anon_evictable_data", KSTAT_DATA_UINT64
},
715 { "anon_evictable_metadata", KSTAT_DATA_UINT64
},
716 { "mru_size", KSTAT_DATA_UINT64
},
717 { "mru_evictable_data", KSTAT_DATA_UINT64
},
718 { "mru_evictable_metadata", KSTAT_DATA_UINT64
},
719 { "mru_ghost_size", KSTAT_DATA_UINT64
},
720 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64
},
721 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
722 { "mfu_size", KSTAT_DATA_UINT64
},
723 { "mfu_evictable_data", KSTAT_DATA_UINT64
},
724 { "mfu_evictable_metadata", KSTAT_DATA_UINT64
},
725 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
726 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64
},
727 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
728 { "l2_hits", KSTAT_DATA_UINT64
},
729 { "l2_misses", KSTAT_DATA_UINT64
},
730 { "l2_feeds", KSTAT_DATA_UINT64
},
731 { "l2_rw_clash", KSTAT_DATA_UINT64
},
732 { "l2_read_bytes", KSTAT_DATA_UINT64
},
733 { "l2_write_bytes", KSTAT_DATA_UINT64
},
734 { "l2_writes_sent", KSTAT_DATA_UINT64
},
735 { "l2_writes_done", KSTAT_DATA_UINT64
},
736 { "l2_writes_error", KSTAT_DATA_UINT64
},
737 { "l2_writes_lock_retry", KSTAT_DATA_UINT64
},
738 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
739 { "l2_evict_reading", KSTAT_DATA_UINT64
},
740 { "l2_evict_l1cached", KSTAT_DATA_UINT64
},
741 { "l2_free_on_write", KSTAT_DATA_UINT64
},
742 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
743 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
744 { "l2_io_error", KSTAT_DATA_UINT64
},
745 { "l2_size", KSTAT_DATA_UINT64
},
746 { "l2_asize", KSTAT_DATA_UINT64
},
747 { "l2_hdr_size", KSTAT_DATA_UINT64
},
748 { "memory_throttle_count", KSTAT_DATA_UINT64
},
749 { "memory_direct_count", KSTAT_DATA_UINT64
},
750 { "memory_indirect_count", KSTAT_DATA_UINT64
},
751 { "memory_all_bytes", KSTAT_DATA_UINT64
},
752 { "memory_free_bytes", KSTAT_DATA_UINT64
},
753 { "memory_available_bytes", KSTAT_DATA_INT64
},
754 { "arc_no_grow", KSTAT_DATA_UINT64
},
755 { "arc_tempreserve", KSTAT_DATA_UINT64
},
756 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
757 { "arc_prune", KSTAT_DATA_UINT64
},
758 { "arc_meta_used", KSTAT_DATA_UINT64
},
759 { "arc_meta_limit", KSTAT_DATA_UINT64
},
760 { "arc_dnode_limit", KSTAT_DATA_UINT64
},
761 { "arc_meta_max", KSTAT_DATA_UINT64
},
762 { "arc_meta_min", KSTAT_DATA_UINT64
},
763 { "sync_wait_for_async", KSTAT_DATA_UINT64
},
764 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64
},
765 { "arc_need_free", KSTAT_DATA_UINT64
},
766 { "arc_sys_free", KSTAT_DATA_UINT64
},
767 { "arc_raw_size", KSTAT_DATA_UINT64
}
770 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
772 #define ARCSTAT_INCR(stat, val) \
773 atomic_add_64(&arc_stats.stat.value.ui64, (val))
775 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
776 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
778 #define ARCSTAT_MAX(stat, val) { \
780 while ((val) > (m = arc_stats.stat.value.ui64) && \
781 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
785 #define ARCSTAT_MAXSTAT(stat) \
786 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
789 * We define a macro to allow ARC hits/misses to be easily broken down by
790 * two separate conditions, giving a total of four different subtypes for
791 * each of hits and misses (so eight statistics total).
793 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
796 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
798 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
802 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
804 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
809 static arc_state_t
*arc_anon
;
810 static arc_state_t
*arc_mru
;
811 static arc_state_t
*arc_mru_ghost
;
812 static arc_state_t
*arc_mfu
;
813 static arc_state_t
*arc_mfu_ghost
;
814 static arc_state_t
*arc_l2c_only
;
817 * There are several ARC variables that are critical to export as kstats --
818 * but we don't want to have to grovel around in the kstat whenever we wish to
819 * manipulate them. For these variables, we therefore define them to be in
820 * terms of the statistic variable. This assures that we are not introducing
821 * the possibility of inconsistency by having shadow copies of the variables,
822 * while still allowing the code to be readable.
824 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
825 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
826 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
827 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
828 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
829 #define arc_no_grow ARCSTAT(arcstat_no_grow) /* do not grow cache size */
830 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
831 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
832 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
833 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
834 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
835 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
836 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
837 #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
838 #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */
839 #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
840 #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */
841 #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */
843 /* size of all b_rabd's in entire arc */
844 #define arc_raw_size ARCSTAT(arcstat_raw_size)
845 /* compressed size of entire arc */
846 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
847 /* uncompressed size of entire arc */
848 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
849 /* number of bytes in the arc from arc_buf_t's */
850 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
852 static list_t arc_prune_list
;
853 static kmutex_t arc_prune_mtx
;
854 static taskq_t
*arc_prune_taskq
;
856 #define GHOST_STATE(state) \
857 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
858 (state) == arc_l2c_only)
860 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
861 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
862 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
863 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
864 #define HDR_COMPRESSION_ENABLED(hdr) \
865 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
867 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
868 #define HDR_L2_READING(hdr) \
869 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
870 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
871 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
872 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
873 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
874 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
875 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
876 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
878 #define HDR_ISTYPE_METADATA(hdr) \
879 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
880 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
882 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
883 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
884 #define HDR_HAS_RABD(hdr) \
885 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
886 (hdr)->b_crypt_hdr.b_rabd != NULL)
887 #define HDR_ENCRYPTED(hdr) \
888 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
889 #define HDR_AUTHENTICATED(hdr) \
890 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
892 /* For storing compression mode in b_flags */
893 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
895 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
896 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
897 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
898 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
900 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
901 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
902 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
903 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
909 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
910 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
911 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
914 * Hash table routines
917 #define HT_LOCK_ALIGN 64
918 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
923 unsigned char pad
[HT_LOCK_PAD
];
927 #define BUF_LOCKS 8192
928 typedef struct buf_hash_table
{
930 arc_buf_hdr_t
**ht_table
;
931 struct ht_lock ht_locks
[BUF_LOCKS
];
934 static buf_hash_table_t buf_hash_table
;
936 #define BUF_HASH_INDEX(spa, dva, birth) \
937 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
938 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
939 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
940 #define HDR_LOCK(hdr) \
941 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
943 uint64_t zfs_crc64_table
[256];
949 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
950 #define L2ARC_HEADROOM 2 /* num of writes */
953 * If we discover during ARC scan any buffers to be compressed, we boost
954 * our headroom for the next scanning cycle by this percentage multiple.
956 #define L2ARC_HEADROOM_BOOST 200
957 #define L2ARC_FEED_SECS 1 /* caching interval secs */
958 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
961 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
962 * and each of the state has two types: data and metadata.
964 #define L2ARC_FEED_TYPES 4
966 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
967 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
969 /* L2ARC Performance Tunables */
970 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
971 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
972 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
973 unsigned long l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
974 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
975 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
976 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
977 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
978 int l2arc_norw
= B_FALSE
; /* no reads during writes */
983 static list_t L2ARC_dev_list
; /* device list */
984 static list_t
*l2arc_dev_list
; /* device list pointer */
985 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
986 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
987 static list_t L2ARC_free_on_write
; /* free after write buf list */
988 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
989 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
990 static uint64_t l2arc_ndev
; /* number of devices */
992 typedef struct l2arc_read_callback
{
993 arc_buf_hdr_t
*l2rcb_hdr
; /* read header */
994 blkptr_t l2rcb_bp
; /* original blkptr */
995 zbookmark_phys_t l2rcb_zb
; /* original bookmark */
996 int l2rcb_flags
; /* original flags */
997 abd_t
*l2rcb_abd
; /* temporary buffer */
998 } l2arc_read_callback_t
;
1000 typedef struct l2arc_data_free
{
1001 /* protected by l2arc_free_on_write_mtx */
1004 arc_buf_contents_t l2df_type
;
1005 list_node_t l2df_list_node
;
1006 } l2arc_data_free_t
;
1008 typedef enum arc_fill_flags
{
1009 ARC_FILL_LOCKED
= 1 << 0, /* hdr lock is held */
1010 ARC_FILL_COMPRESSED
= 1 << 1, /* fill with compressed data */
1011 ARC_FILL_ENCRYPTED
= 1 << 2, /* fill with encrypted data */
1012 ARC_FILL_NOAUTH
= 1 << 3, /* don't attempt to authenticate */
1013 ARC_FILL_IN_PLACE
= 1 << 4 /* fill in place (special case) */
1016 static kmutex_t l2arc_feed_thr_lock
;
1017 static kcondvar_t l2arc_feed_thr_cv
;
1018 static uint8_t l2arc_thread_exit
;
1020 static abd_t
*arc_get_data_abd(arc_buf_hdr_t
*, uint64_t, void *);
1021 static void *arc_get_data_buf(arc_buf_hdr_t
*, uint64_t, void *);
1022 static void arc_get_data_impl(arc_buf_hdr_t
*, uint64_t, void *);
1023 static void arc_free_data_abd(arc_buf_hdr_t
*, abd_t
*, uint64_t, void *);
1024 static void arc_free_data_buf(arc_buf_hdr_t
*, void *, uint64_t, void *);
1025 static void arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
);
1026 static void arc_hdr_free_abd(arc_buf_hdr_t
*, boolean_t
);
1027 static void arc_hdr_alloc_abd(arc_buf_hdr_t
*, boolean_t
);
1028 static void arc_access(arc_buf_hdr_t
*, kmutex_t
*);
1029 static boolean_t
arc_is_overflowing(void);
1030 static void arc_buf_watch(arc_buf_t
*);
1031 static void arc_tuning_update(void);
1032 static void arc_prune_async(int64_t);
1033 static uint64_t arc_all_memory(void);
1035 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t
*);
1036 static uint32_t arc_bufc_to_flags(arc_buf_contents_t
);
1037 static inline void arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
1038 static inline void arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
1040 static boolean_t
l2arc_write_eligible(uint64_t, arc_buf_hdr_t
*);
1041 static void l2arc_read_done(zio_t
*);
1044 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
1046 uint8_t *vdva
= (uint8_t *)dva
;
1047 uint64_t crc
= -1ULL;
1050 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
1052 for (i
= 0; i
< sizeof (dva_t
); i
++)
1053 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
1055 crc
^= (spa
>>8) ^ birth
;
1060 #define HDR_EMPTY(hdr) \
1061 ((hdr)->b_dva.dva_word[0] == 0 && \
1062 (hdr)->b_dva.dva_word[1] == 0)
1064 #define HDR_EQUAL(spa, dva, birth, hdr) \
1065 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1066 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1067 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1070 buf_discard_identity(arc_buf_hdr_t
*hdr
)
1072 hdr
->b_dva
.dva_word
[0] = 0;
1073 hdr
->b_dva
.dva_word
[1] = 0;
1077 static arc_buf_hdr_t
*
1078 buf_hash_find(uint64_t spa
, const blkptr_t
*bp
, kmutex_t
**lockp
)
1080 const dva_t
*dva
= BP_IDENTITY(bp
);
1081 uint64_t birth
= BP_PHYSICAL_BIRTH(bp
);
1082 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
1083 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1086 mutex_enter(hash_lock
);
1087 for (hdr
= buf_hash_table
.ht_table
[idx
]; hdr
!= NULL
;
1088 hdr
= hdr
->b_hash_next
) {
1089 if (HDR_EQUAL(spa
, dva
, birth
, hdr
)) {
1094 mutex_exit(hash_lock
);
1100 * Insert an entry into the hash table. If there is already an element
1101 * equal to elem in the hash table, then the already existing element
1102 * will be returned and the new element will not be inserted.
1103 * Otherwise returns NULL.
1104 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1106 static arc_buf_hdr_t
*
1107 buf_hash_insert(arc_buf_hdr_t
*hdr
, kmutex_t
**lockp
)
1109 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1110 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1111 arc_buf_hdr_t
*fhdr
;
1114 ASSERT(!DVA_IS_EMPTY(&hdr
->b_dva
));
1115 ASSERT(hdr
->b_birth
!= 0);
1116 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
1118 if (lockp
!= NULL
) {
1120 mutex_enter(hash_lock
);
1122 ASSERT(MUTEX_HELD(hash_lock
));
1125 for (fhdr
= buf_hash_table
.ht_table
[idx
], i
= 0; fhdr
!= NULL
;
1126 fhdr
= fhdr
->b_hash_next
, i
++) {
1127 if (HDR_EQUAL(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
, fhdr
))
1131 hdr
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
1132 buf_hash_table
.ht_table
[idx
] = hdr
;
1133 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1135 /* collect some hash table performance data */
1137 ARCSTAT_BUMP(arcstat_hash_collisions
);
1139 ARCSTAT_BUMP(arcstat_hash_chains
);
1141 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
1144 ARCSTAT_BUMP(arcstat_hash_elements
);
1145 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
1151 buf_hash_remove(arc_buf_hdr_t
*hdr
)
1153 arc_buf_hdr_t
*fhdr
, **hdrp
;
1154 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1156 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
1157 ASSERT(HDR_IN_HASH_TABLE(hdr
));
1159 hdrp
= &buf_hash_table
.ht_table
[idx
];
1160 while ((fhdr
= *hdrp
) != hdr
) {
1161 ASSERT3P(fhdr
, !=, NULL
);
1162 hdrp
= &fhdr
->b_hash_next
;
1164 *hdrp
= hdr
->b_hash_next
;
1165 hdr
->b_hash_next
= NULL
;
1166 arc_hdr_clear_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1168 /* collect some hash table performance data */
1169 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
1171 if (buf_hash_table
.ht_table
[idx
] &&
1172 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
1173 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
1177 * Global data structures and functions for the buf kmem cache.
1180 static kmem_cache_t
*hdr_full_cache
;
1181 static kmem_cache_t
*hdr_full_crypt_cache
;
1182 static kmem_cache_t
*hdr_l2only_cache
;
1183 static kmem_cache_t
*buf_cache
;
1190 #if defined(_KERNEL) && defined(HAVE_SPL)
1192 * Large allocations which do not require contiguous pages
1193 * should be using vmem_free() in the linux kernel\
1195 vmem_free(buf_hash_table
.ht_table
,
1196 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1198 kmem_free(buf_hash_table
.ht_table
,
1199 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1201 for (i
= 0; i
< BUF_LOCKS
; i
++)
1202 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
1203 kmem_cache_destroy(hdr_full_cache
);
1204 kmem_cache_destroy(hdr_full_crypt_cache
);
1205 kmem_cache_destroy(hdr_l2only_cache
);
1206 kmem_cache_destroy(buf_cache
);
1210 * Constructor callback - called when the cache is empty
1211 * and a new buf is requested.
1215 hdr_full_cons(void *vbuf
, void *unused
, int kmflag
)
1217 arc_buf_hdr_t
*hdr
= vbuf
;
1219 bzero(hdr
, HDR_FULL_SIZE
);
1220 cv_init(&hdr
->b_l1hdr
.b_cv
, NULL
, CV_DEFAULT
, NULL
);
1221 refcount_create(&hdr
->b_l1hdr
.b_refcnt
);
1222 mutex_init(&hdr
->b_l1hdr
.b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1223 list_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1224 list_link_init(&hdr
->b_l2hdr
.b_l2node
);
1225 multilist_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1226 arc_space_consume(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1233 hdr_full_crypt_cons(void *vbuf
, void *unused
, int kmflag
)
1235 arc_buf_hdr_t
*hdr
= vbuf
;
1237 hdr_full_cons(vbuf
, unused
, kmflag
);
1238 bzero(&hdr
->b_crypt_hdr
, sizeof (hdr
->b_crypt_hdr
));
1239 arc_space_consume(sizeof (hdr
->b_crypt_hdr
), ARC_SPACE_HDRS
);
1246 hdr_l2only_cons(void *vbuf
, void *unused
, int kmflag
)
1248 arc_buf_hdr_t
*hdr
= vbuf
;
1250 bzero(hdr
, HDR_L2ONLY_SIZE
);
1251 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1258 buf_cons(void *vbuf
, void *unused
, int kmflag
)
1260 arc_buf_t
*buf
= vbuf
;
1262 bzero(buf
, sizeof (arc_buf_t
));
1263 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1264 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1270 * Destructor callback - called when a cached buf is
1271 * no longer required.
1275 hdr_full_dest(void *vbuf
, void *unused
)
1277 arc_buf_hdr_t
*hdr
= vbuf
;
1279 ASSERT(HDR_EMPTY(hdr
));
1280 cv_destroy(&hdr
->b_l1hdr
.b_cv
);
1281 refcount_destroy(&hdr
->b_l1hdr
.b_refcnt
);
1282 mutex_destroy(&hdr
->b_l1hdr
.b_freeze_lock
);
1283 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1284 arc_space_return(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1289 hdr_full_crypt_dest(void *vbuf
, void *unused
)
1291 arc_buf_hdr_t
*hdr
= vbuf
;
1293 hdr_full_dest(vbuf
, unused
);
1294 arc_space_return(sizeof (hdr
->b_crypt_hdr
), ARC_SPACE_HDRS
);
1299 hdr_l2only_dest(void *vbuf
, void *unused
)
1301 ASSERTV(arc_buf_hdr_t
*hdr
= vbuf
);
1303 ASSERT(HDR_EMPTY(hdr
));
1304 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1309 buf_dest(void *vbuf
, void *unused
)
1311 arc_buf_t
*buf
= vbuf
;
1313 mutex_destroy(&buf
->b_evict_lock
);
1314 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1318 * Reclaim callback -- invoked when memory is low.
1322 hdr_recl(void *unused
)
1324 dprintf("hdr_recl called\n");
1326 * umem calls the reclaim func when we destroy the buf cache,
1327 * which is after we do arc_fini().
1330 cv_signal(&arc_reclaim_thread_cv
);
1336 uint64_t *ct
= NULL
;
1337 uint64_t hsize
= 1ULL << 12;
1341 * The hash table is big enough to fill all of physical memory
1342 * with an average block size of zfs_arc_average_blocksize (default 8K).
1343 * By default, the table will take up
1344 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1346 while (hsize
* zfs_arc_average_blocksize
< arc_all_memory())
1349 buf_hash_table
.ht_mask
= hsize
- 1;
1350 #if defined(_KERNEL) && defined(HAVE_SPL)
1352 * Large allocations which do not require contiguous pages
1353 * should be using vmem_alloc() in the linux kernel
1355 buf_hash_table
.ht_table
=
1356 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
1358 buf_hash_table
.ht_table
=
1359 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
1361 if (buf_hash_table
.ht_table
== NULL
) {
1362 ASSERT(hsize
> (1ULL << 8));
1367 hdr_full_cache
= kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE
,
1368 0, hdr_full_cons
, hdr_full_dest
, hdr_recl
, NULL
, NULL
, 0);
1369 hdr_full_crypt_cache
= kmem_cache_create("arc_buf_hdr_t_full_crypt",
1370 HDR_FULL_CRYPT_SIZE
, 0, hdr_full_crypt_cons
, hdr_full_crypt_dest
,
1371 hdr_recl
, NULL
, NULL
, 0);
1372 hdr_l2only_cache
= kmem_cache_create("arc_buf_hdr_t_l2only",
1373 HDR_L2ONLY_SIZE
, 0, hdr_l2only_cons
, hdr_l2only_dest
, hdr_recl
,
1375 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
1376 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
1378 for (i
= 0; i
< 256; i
++)
1379 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
1380 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
1382 for (i
= 0; i
< BUF_LOCKS
; i
++) {
1383 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
1384 NULL
, MUTEX_DEFAULT
, NULL
);
1388 #define ARC_MINTIME (hz>>4) /* 62 ms */
1391 * This is the size that the buf occupies in memory. If the buf is compressed,
1392 * it will correspond to the compressed size. You should use this method of
1393 * getting the buf size unless you explicitly need the logical size.
1396 arc_buf_size(arc_buf_t
*buf
)
1398 return (ARC_BUF_COMPRESSED(buf
) ?
1399 HDR_GET_PSIZE(buf
->b_hdr
) : HDR_GET_LSIZE(buf
->b_hdr
));
1403 arc_buf_lsize(arc_buf_t
*buf
)
1405 return (HDR_GET_LSIZE(buf
->b_hdr
));
1409 * This function will return B_TRUE if the buffer is encrypted in memory.
1410 * This buffer can be decrypted by calling arc_untransform().
1413 arc_is_encrypted(arc_buf_t
*buf
)
1415 return (ARC_BUF_ENCRYPTED(buf
) != 0);
1419 * Returns B_TRUE if the buffer represents data that has not had its MAC
1423 arc_is_unauthenticated(arc_buf_t
*buf
)
1425 return (HDR_NOAUTH(buf
->b_hdr
) != 0);
1429 arc_get_raw_params(arc_buf_t
*buf
, boolean_t
*byteorder
, uint8_t *salt
,
1430 uint8_t *iv
, uint8_t *mac
)
1432 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1434 ASSERT(HDR_PROTECTED(hdr
));
1436 bcopy(hdr
->b_crypt_hdr
.b_salt
, salt
, ZIO_DATA_SALT_LEN
);
1437 bcopy(hdr
->b_crypt_hdr
.b_iv
, iv
, ZIO_DATA_IV_LEN
);
1438 bcopy(hdr
->b_crypt_hdr
.b_mac
, mac
, ZIO_DATA_MAC_LEN
);
1439 *byteorder
= (hdr
->b_l1hdr
.b_byteswap
== DMU_BSWAP_NUMFUNCS
) ?
1440 ZFS_HOST_BYTEORDER
: !ZFS_HOST_BYTEORDER
;
1444 * Indicates how this buffer is compressed in memory. If it is not compressed
1445 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1446 * arc_untransform() as long as it is also unencrypted.
1449 arc_get_compression(arc_buf_t
*buf
)
1451 return (ARC_BUF_COMPRESSED(buf
) ?
1452 HDR_GET_COMPRESS(buf
->b_hdr
) : ZIO_COMPRESS_OFF
);
1456 * Return the compression algorithm used to store this data in the ARC. If ARC
1457 * compression is enabled or this is an encrypted block, this will be the same
1458 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1460 static inline enum zio_compress
1461 arc_hdr_get_compress(arc_buf_hdr_t
*hdr
)
1463 return (HDR_COMPRESSION_ENABLED(hdr
) ?
1464 HDR_GET_COMPRESS(hdr
) : ZIO_COMPRESS_OFF
);
1467 static inline boolean_t
1468 arc_buf_is_shared(arc_buf_t
*buf
)
1470 boolean_t shared
= (buf
->b_data
!= NULL
&&
1471 buf
->b_hdr
->b_l1hdr
.b_pabd
!= NULL
&&
1472 abd_is_linear(buf
->b_hdr
->b_l1hdr
.b_pabd
) &&
1473 buf
->b_data
== abd_to_buf(buf
->b_hdr
->b_l1hdr
.b_pabd
));
1474 IMPLY(shared
, HDR_SHARED_DATA(buf
->b_hdr
));
1475 IMPLY(shared
, ARC_BUF_SHARED(buf
));
1476 IMPLY(shared
, ARC_BUF_COMPRESSED(buf
) || ARC_BUF_LAST(buf
));
1479 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1480 * already being shared" requirement prevents us from doing that.
1487 * Free the checksum associated with this header. If there is no checksum, this
1491 arc_cksum_free(arc_buf_hdr_t
*hdr
)
1493 ASSERT(HDR_HAS_L1HDR(hdr
));
1495 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1496 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1497 kmem_free(hdr
->b_l1hdr
.b_freeze_cksum
, sizeof (zio_cksum_t
));
1498 hdr
->b_l1hdr
.b_freeze_cksum
= NULL
;
1500 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1504 * Return true iff at least one of the bufs on hdr is not compressed.
1505 * Encrypted buffers count as compressed.
1508 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t
*hdr
)
1510 for (arc_buf_t
*b
= hdr
->b_l1hdr
.b_buf
; b
!= NULL
; b
= b
->b_next
) {
1511 if (!ARC_BUF_COMPRESSED(b
)) {
1520 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1521 * matches the checksum that is stored in the hdr. If there is no checksum,
1522 * or if the buf is compressed, this is a no-op.
1525 arc_cksum_verify(arc_buf_t
*buf
)
1527 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1530 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1533 if (ARC_BUF_COMPRESSED(buf
)) {
1534 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1535 arc_hdr_has_uncompressed_buf(hdr
));
1539 ASSERT(HDR_HAS_L1HDR(hdr
));
1541 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1542 if (hdr
->b_l1hdr
.b_freeze_cksum
== NULL
|| HDR_IO_ERROR(hdr
)) {
1543 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1547 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
, &zc
);
1548 if (!ZIO_CHECKSUM_EQUAL(*hdr
->b_l1hdr
.b_freeze_cksum
, zc
))
1549 panic("buffer modified while frozen!");
1550 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1554 * This function makes the assumption that data stored in the L2ARC
1555 * will be transformed exactly as it is in the main pool. Because of
1556 * this we can verify the checksum against the reading process's bp.
1559 arc_cksum_is_equal(arc_buf_hdr_t
*hdr
, zio_t
*zio
)
1561 ASSERT(!BP_IS_EMBEDDED(zio
->io_bp
));
1562 VERIFY3U(BP_GET_PSIZE(zio
->io_bp
), ==, HDR_GET_PSIZE(hdr
));
1565 * Block pointers always store the checksum for the logical data.
1566 * If the block pointer has the gang bit set, then the checksum
1567 * it represents is for the reconstituted data and not for an
1568 * individual gang member. The zio pipeline, however, must be able to
1569 * determine the checksum of each of the gang constituents so it
1570 * treats the checksum comparison differently than what we need
1571 * for l2arc blocks. This prevents us from using the
1572 * zio_checksum_error() interface directly. Instead we must call the
1573 * zio_checksum_error_impl() so that we can ensure the checksum is
1574 * generated using the correct checksum algorithm and accounts for the
1575 * logical I/O size and not just a gang fragment.
1577 return (zio_checksum_error_impl(zio
->io_spa
, zio
->io_bp
,
1578 BP_GET_CHECKSUM(zio
->io_bp
), zio
->io_abd
, zio
->io_size
,
1579 zio
->io_offset
, NULL
) == 0);
1583 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1584 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1585 * isn't modified later on. If buf is compressed or there is already a checksum
1586 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1589 arc_cksum_compute(arc_buf_t
*buf
)
1591 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1593 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1596 ASSERT(HDR_HAS_L1HDR(hdr
));
1598 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1599 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1600 ASSERT(arc_hdr_has_uncompressed_buf(hdr
));
1601 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1603 } else if (ARC_BUF_COMPRESSED(buf
)) {
1604 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1608 ASSERT(!ARC_BUF_ENCRYPTED(buf
));
1609 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1610 hdr
->b_l1hdr
.b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
),
1612 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
,
1613 hdr
->b_l1hdr
.b_freeze_cksum
);
1614 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1620 arc_buf_sigsegv(int sig
, siginfo_t
*si
, void *unused
)
1622 panic("Got SIGSEGV at address: 0x%lx\n", (long)si
->si_addr
);
1628 arc_buf_unwatch(arc_buf_t
*buf
)
1632 ASSERT0(mprotect(buf
->b_data
, arc_buf_size(buf
),
1633 PROT_READ
| PROT_WRITE
));
1640 arc_buf_watch(arc_buf_t
*buf
)
1644 ASSERT0(mprotect(buf
->b_data
, arc_buf_size(buf
),
1649 static arc_buf_contents_t
1650 arc_buf_type(arc_buf_hdr_t
*hdr
)
1652 arc_buf_contents_t type
;
1653 if (HDR_ISTYPE_METADATA(hdr
)) {
1654 type
= ARC_BUFC_METADATA
;
1656 type
= ARC_BUFC_DATA
;
1658 VERIFY3U(hdr
->b_type
, ==, type
);
1663 arc_is_metadata(arc_buf_t
*buf
)
1665 return (HDR_ISTYPE_METADATA(buf
->b_hdr
) != 0);
1669 arc_bufc_to_flags(arc_buf_contents_t type
)
1673 /* metadata field is 0 if buffer contains normal data */
1675 case ARC_BUFC_METADATA
:
1676 return (ARC_FLAG_BUFC_METADATA
);
1680 panic("undefined ARC buffer type!");
1681 return ((uint32_t)-1);
1685 arc_buf_thaw(arc_buf_t
*buf
)
1687 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1689 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
1690 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
1692 arc_cksum_verify(buf
);
1695 * Compressed buffers do not manipulate the b_freeze_cksum or
1696 * allocate b_thawed.
1698 if (ARC_BUF_COMPRESSED(buf
)) {
1699 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1700 arc_hdr_has_uncompressed_buf(hdr
));
1704 ASSERT(HDR_HAS_L1HDR(hdr
));
1705 arc_cksum_free(hdr
);
1706 arc_buf_unwatch(buf
);
1710 arc_buf_freeze(arc_buf_t
*buf
)
1712 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1713 kmutex_t
*hash_lock
;
1715 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1718 if (ARC_BUF_COMPRESSED(buf
)) {
1719 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1720 arc_hdr_has_uncompressed_buf(hdr
));
1724 hash_lock
= HDR_LOCK(hdr
);
1725 mutex_enter(hash_lock
);
1727 ASSERT(HDR_HAS_L1HDR(hdr
));
1728 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
||
1729 hdr
->b_l1hdr
.b_state
== arc_anon
);
1730 arc_cksum_compute(buf
);
1731 mutex_exit(hash_lock
);
1735 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1736 * the following functions should be used to ensure that the flags are
1737 * updated in a thread-safe way. When manipulating the flags either
1738 * the hash_lock must be held or the hdr must be undiscoverable. This
1739 * ensures that we're not racing with any other threads when updating
1743 arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1745 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1746 hdr
->b_flags
|= flags
;
1750 arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1752 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1753 hdr
->b_flags
&= ~flags
;
1757 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1758 * done in a special way since we have to clear and set bits
1759 * at the same time. Consumers that wish to set the compression bits
1760 * must use this function to ensure that the flags are updated in
1761 * thread-safe manner.
1764 arc_hdr_set_compress(arc_buf_hdr_t
*hdr
, enum zio_compress cmp
)
1766 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1769 * Holes and embedded blocks will always have a psize = 0 so
1770 * we ignore the compression of the blkptr and set the
1771 * want to uncompress them. Mark them as uncompressed.
1773 if (!zfs_compressed_arc_enabled
|| HDR_GET_PSIZE(hdr
) == 0) {
1774 arc_hdr_clear_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1775 ASSERT(!HDR_COMPRESSION_ENABLED(hdr
));
1777 arc_hdr_set_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1778 ASSERT(HDR_COMPRESSION_ENABLED(hdr
));
1781 HDR_SET_COMPRESS(hdr
, cmp
);
1782 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, cmp
);
1786 * Looks for another buf on the same hdr which has the data decompressed, copies
1787 * from it, and returns true. If no such buf exists, returns false.
1790 arc_buf_try_copy_decompressed_data(arc_buf_t
*buf
)
1792 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1793 boolean_t copied
= B_FALSE
;
1795 ASSERT(HDR_HAS_L1HDR(hdr
));
1796 ASSERT3P(buf
->b_data
, !=, NULL
);
1797 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1799 for (arc_buf_t
*from
= hdr
->b_l1hdr
.b_buf
; from
!= NULL
;
1800 from
= from
->b_next
) {
1801 /* can't use our own data buffer */
1806 if (!ARC_BUF_COMPRESSED(from
)) {
1807 bcopy(from
->b_data
, buf
->b_data
, arc_buf_size(buf
));
1814 * There were no decompressed bufs, so there should not be a
1815 * checksum on the hdr either.
1817 EQUIV(!copied
, hdr
->b_l1hdr
.b_freeze_cksum
== NULL
);
1823 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1826 arc_hdr_size(arc_buf_hdr_t
*hdr
)
1830 if (arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
&&
1831 HDR_GET_PSIZE(hdr
) > 0) {
1832 size
= HDR_GET_PSIZE(hdr
);
1834 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, 0);
1835 size
= HDR_GET_LSIZE(hdr
);
1841 arc_hdr_authenticate(arc_buf_hdr_t
*hdr
, spa_t
*spa
, uint64_t dsobj
)
1845 uint64_t lsize
= HDR_GET_LSIZE(hdr
);
1846 uint64_t psize
= HDR_GET_PSIZE(hdr
);
1847 void *tmpbuf
= NULL
;
1848 abd_t
*abd
= hdr
->b_l1hdr
.b_pabd
;
1850 ASSERT(HDR_LOCK(hdr
) == NULL
|| MUTEX_HELD(HDR_LOCK(hdr
)));
1851 ASSERT(HDR_AUTHENTICATED(hdr
));
1852 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
1855 * The MAC is calculated on the compressed data that is stored on disk.
1856 * However, if compressed arc is disabled we will only have the
1857 * decompressed data available to us now. Compress it into a temporary
1858 * abd so we can verify the MAC. The performance overhead of this will
1859 * be relatively low, since most objects in an encrypted objset will
1860 * be encrypted (instead of authenticated) anyway.
1862 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
1863 !HDR_COMPRESSION_ENABLED(hdr
)) {
1864 tmpbuf
= zio_buf_alloc(lsize
);
1865 abd
= abd_get_from_buf(tmpbuf
, lsize
);
1866 abd_take_ownership_of_buf(abd
, B_TRUE
);
1868 csize
= zio_compress_data(HDR_GET_COMPRESS(hdr
),
1869 hdr
->b_l1hdr
.b_pabd
, tmpbuf
, lsize
);
1870 ASSERT3U(csize
, <=, psize
);
1871 abd_zero_off(abd
, csize
, psize
- csize
);
1875 * Authentication is best effort. We authenticate whenever the key is
1876 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1878 if (hdr
->b_crypt_hdr
.b_ot
== DMU_OT_OBJSET
) {
1879 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1880 ASSERT3U(lsize
, ==, psize
);
1881 ret
= spa_do_crypt_objset_mac_abd(B_FALSE
, spa
, dsobj
, abd
,
1882 psize
, hdr
->b_l1hdr
.b_byteswap
!= DMU_BSWAP_NUMFUNCS
);
1884 ret
= spa_do_crypt_mac_abd(B_FALSE
, spa
, dsobj
, abd
, psize
,
1885 hdr
->b_crypt_hdr
.b_mac
);
1889 arc_hdr_clear_flags(hdr
, ARC_FLAG_NOAUTH
);
1890 else if (ret
!= ENOENT
)
1906 * This function will take a header that only has raw encrypted data in
1907 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1908 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1909 * also decompress the data.
1912 arc_hdr_decrypt(arc_buf_hdr_t
*hdr
, spa_t
*spa
, uint64_t dsobj
)
1915 dsl_crypto_key_t
*dck
= NULL
;
1918 boolean_t no_crypt
= B_FALSE
;
1919 boolean_t bswap
= (hdr
->b_l1hdr
.b_byteswap
!= DMU_BSWAP_NUMFUNCS
);
1921 ASSERT(HDR_LOCK(hdr
) == NULL
|| MUTEX_HELD(HDR_LOCK(hdr
)));
1922 ASSERT(HDR_ENCRYPTED(hdr
));
1924 arc_hdr_alloc_abd(hdr
, B_FALSE
);
1927 * We must be careful to use the passed-in dsobj value here and
1928 * not the value in b_dsobj. b_dsobj is meant to be a best guess for
1929 * the L2ARC, which has the luxury of being able to fail without real
1930 * consequences (the data simply won't make it to the L2ARC). In
1931 * reality, the dsobj stored in the header may belong to a dataset
1932 * that has been unmounted or otherwise disowned, meaning the key
1933 * won't be accessible via that dsobj anymore.
1935 ret
= spa_keystore_lookup_key(spa
, dsobj
, FTAG
, &dck
);
1937 ret
= SET_ERROR(EACCES
);
1941 ret
= zio_do_crypt_abd(B_FALSE
, &dck
->dck_key
,
1942 hdr
->b_crypt_hdr
.b_salt
, hdr
->b_crypt_hdr
.b_ot
,
1943 hdr
->b_crypt_hdr
.b_iv
, hdr
->b_crypt_hdr
.b_mac
,
1944 HDR_GET_PSIZE(hdr
), bswap
, hdr
->b_l1hdr
.b_pabd
,
1945 hdr
->b_crypt_hdr
.b_rabd
, &no_crypt
);
1950 abd_copy(hdr
->b_l1hdr
.b_pabd
, hdr
->b_crypt_hdr
.b_rabd
,
1951 HDR_GET_PSIZE(hdr
));
1955 * If this header has disabled arc compression but the b_pabd is
1956 * compressed after decrypting it, we need to decompress the newly
1959 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
1960 !HDR_COMPRESSION_ENABLED(hdr
)) {
1962 * We want to make sure that we are correctly honoring the
1963 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1964 * and then loan a buffer from it, rather than allocating a
1965 * linear buffer and wrapping it in an abd later.
1967 cabd
= arc_get_data_abd(hdr
, arc_hdr_size(hdr
), hdr
);
1968 tmp
= abd_borrow_buf(cabd
, arc_hdr_size(hdr
));
1970 ret
= zio_decompress_data(HDR_GET_COMPRESS(hdr
),
1971 hdr
->b_l1hdr
.b_pabd
, tmp
, HDR_GET_PSIZE(hdr
),
1972 HDR_GET_LSIZE(hdr
));
1974 abd_return_buf(cabd
, tmp
, arc_hdr_size(hdr
));
1978 abd_return_buf_copy(cabd
, tmp
, arc_hdr_size(hdr
));
1979 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
,
1980 arc_hdr_size(hdr
), hdr
);
1981 hdr
->b_l1hdr
.b_pabd
= cabd
;
1984 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
1989 arc_hdr_free_abd(hdr
, B_FALSE
);
1991 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
1993 arc_free_data_buf(hdr
, cabd
, arc_hdr_size(hdr
), hdr
);
1999 * This function is called during arc_buf_fill() to prepare the header's
2000 * abd plaintext pointer for use. This involves authenticated protected
2001 * data and decrypting encrypted data into the plaintext abd.
2004 arc_fill_hdr_crypt(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, spa_t
*spa
,
2005 uint64_t dsobj
, boolean_t noauth
)
2009 ASSERT(HDR_PROTECTED(hdr
));
2011 if (hash_lock
!= NULL
)
2012 mutex_enter(hash_lock
);
2014 if (HDR_NOAUTH(hdr
) && !noauth
) {
2016 * The caller requested authenticated data but our data has
2017 * not been authenticated yet. Verify the MAC now if we can.
2019 ret
= arc_hdr_authenticate(hdr
, spa
, dsobj
);
2022 } else if (HDR_HAS_RABD(hdr
) && hdr
->b_l1hdr
.b_pabd
== NULL
) {
2024 * If we only have the encrypted version of the data, but the
2025 * unencrypted version was requested we take this opportunity
2026 * to store the decrypted version in the header for future use.
2028 ret
= arc_hdr_decrypt(hdr
, spa
, dsobj
);
2033 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2035 if (hash_lock
!= NULL
)
2036 mutex_exit(hash_lock
);
2041 if (hash_lock
!= NULL
)
2042 mutex_exit(hash_lock
);
2048 * This function is used by the dbuf code to decrypt bonus buffers in place.
2049 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
2050 * block, so we use the hash lock here to protect against concurrent calls to
2054 arc_buf_untransform_in_place(arc_buf_t
*buf
, kmutex_t
*hash_lock
)
2056 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2058 ASSERT(HDR_ENCRYPTED(hdr
));
2059 ASSERT3U(hdr
->b_crypt_hdr
.b_ot
, ==, DMU_OT_DNODE
);
2060 ASSERT(HDR_LOCK(hdr
) == NULL
|| MUTEX_HELD(HDR_LOCK(hdr
)));
2061 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2063 zio_crypt_copy_dnode_bonus(hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
2065 buf
->b_flags
&= ~ARC_BUF_FLAG_ENCRYPTED
;
2066 buf
->b_flags
&= ~ARC_BUF_FLAG_COMPRESSED
;
2067 hdr
->b_crypt_hdr
.b_ebufcnt
-= 1;
2071 * Given a buf that has a data buffer attached to it, this function will
2072 * efficiently fill the buf with data of the specified compression setting from
2073 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2074 * are already sharing a data buf, no copy is performed.
2076 * If the buf is marked as compressed but uncompressed data was requested, this
2077 * will allocate a new data buffer for the buf, remove that flag, and fill the
2078 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2079 * uncompressed data, and (since we haven't added support for it yet) if you
2080 * want compressed data your buf must already be marked as compressed and have
2081 * the correct-sized data buffer.
2084 arc_buf_fill(arc_buf_t
*buf
, spa_t
*spa
, uint64_t dsobj
, arc_fill_flags_t flags
)
2087 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2088 boolean_t hdr_compressed
=
2089 (arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
);
2090 boolean_t compressed
= (flags
& ARC_FILL_COMPRESSED
) != 0;
2091 boolean_t encrypted
= (flags
& ARC_FILL_ENCRYPTED
) != 0;
2092 dmu_object_byteswap_t bswap
= hdr
->b_l1hdr
.b_byteswap
;
2093 kmutex_t
*hash_lock
= (flags
& ARC_FILL_LOCKED
) ? NULL
: HDR_LOCK(hdr
);
2095 ASSERT3P(buf
->b_data
, !=, NULL
);
2096 IMPLY(compressed
, hdr_compressed
|| ARC_BUF_ENCRYPTED(buf
));
2097 IMPLY(compressed
, ARC_BUF_COMPRESSED(buf
));
2098 IMPLY(encrypted
, HDR_ENCRYPTED(hdr
));
2099 IMPLY(encrypted
, ARC_BUF_ENCRYPTED(buf
));
2100 IMPLY(encrypted
, ARC_BUF_COMPRESSED(buf
));
2101 IMPLY(encrypted
, !ARC_BUF_SHARED(buf
));
2104 * If the caller wanted encrypted data we just need to copy it from
2105 * b_rabd and potentially byteswap it. We won't be able to do any
2106 * further transforms on it.
2109 ASSERT(HDR_HAS_RABD(hdr
));
2110 abd_copy_to_buf(buf
->b_data
, hdr
->b_crypt_hdr
.b_rabd
,
2111 HDR_GET_PSIZE(hdr
));
2116 * Adjust encrypted and authenticated headers to accomodate the
2117 * request if needed.
2119 if (HDR_PROTECTED(hdr
)) {
2120 error
= arc_fill_hdr_crypt(hdr
, hash_lock
, spa
,
2121 dsobj
, !!(flags
& ARC_FILL_NOAUTH
));
2127 * There is a special case here for dnode blocks which are
2128 * decrypting their bonus buffers. These blocks may request to
2129 * be decrypted in-place. This is necessary because there may
2130 * be many dnodes pointing into this buffer and there is
2131 * currently no method to synchronize replacing the backing
2132 * b_data buffer and updating all of the pointers. Here we use
2133 * the hash lock to ensure there are no races. If the need
2134 * arises for other types to be decrypted in-place, they must
2135 * add handling here as well.
2137 if ((flags
& ARC_FILL_IN_PLACE
) != 0) {
2138 ASSERT(!hdr_compressed
);
2139 ASSERT(!compressed
);
2142 if (HDR_ENCRYPTED(hdr
) && ARC_BUF_ENCRYPTED(buf
)) {
2143 ASSERT3U(hdr
->b_crypt_hdr
.b_ot
, ==, DMU_OT_DNODE
);
2145 if (hash_lock
!= NULL
)
2146 mutex_enter(hash_lock
);
2147 arc_buf_untransform_in_place(buf
, hash_lock
);
2148 if (hash_lock
!= NULL
)
2149 mutex_exit(hash_lock
);
2151 /* Compute the hdr's checksum if necessary */
2152 arc_cksum_compute(buf
);
2158 if (hdr_compressed
== compressed
) {
2159 if (!arc_buf_is_shared(buf
)) {
2160 abd_copy_to_buf(buf
->b_data
, hdr
->b_l1hdr
.b_pabd
,
2164 ASSERT(hdr_compressed
);
2165 ASSERT(!compressed
);
2166 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, HDR_GET_PSIZE(hdr
));
2169 * If the buf is sharing its data with the hdr, unlink it and
2170 * allocate a new data buffer for the buf.
2172 if (arc_buf_is_shared(buf
)) {
2173 ASSERT(ARC_BUF_COMPRESSED(buf
));
2175 /* We need to give the buf it's own b_data */
2176 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
2178 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
2179 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2181 /* Previously overhead was 0; just add new overhead */
2182 ARCSTAT_INCR(arcstat_overhead_size
, HDR_GET_LSIZE(hdr
));
2183 } else if (ARC_BUF_COMPRESSED(buf
)) {
2184 /* We need to reallocate the buf's b_data */
2185 arc_free_data_buf(hdr
, buf
->b_data
, HDR_GET_PSIZE(hdr
),
2188 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
2190 /* We increased the size of b_data; update overhead */
2191 ARCSTAT_INCR(arcstat_overhead_size
,
2192 HDR_GET_LSIZE(hdr
) - HDR_GET_PSIZE(hdr
));
2196 * Regardless of the buf's previous compression settings, it
2197 * should not be compressed at the end of this function.
2199 buf
->b_flags
&= ~ARC_BUF_FLAG_COMPRESSED
;
2202 * Try copying the data from another buf which already has a
2203 * decompressed version. If that's not possible, it's time to
2204 * bite the bullet and decompress the data from the hdr.
2206 if (arc_buf_try_copy_decompressed_data(buf
)) {
2207 /* Skip byteswapping and checksumming (already done) */
2208 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, !=, NULL
);
2211 error
= zio_decompress_data(HDR_GET_COMPRESS(hdr
),
2212 hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
2213 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
2216 * Absent hardware errors or software bugs, this should
2217 * be impossible, but log it anyway so we can debug it.
2221 "hdr %p, compress %d, psize %d, lsize %d",
2222 hdr
, arc_hdr_get_compress(hdr
),
2223 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
2224 return (SET_ERROR(EIO
));
2230 /* Byteswap the buf's data if necessary */
2231 if (bswap
!= DMU_BSWAP_NUMFUNCS
) {
2232 ASSERT(!HDR_SHARED_DATA(hdr
));
2233 ASSERT3U(bswap
, <, DMU_BSWAP_NUMFUNCS
);
2234 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, HDR_GET_LSIZE(hdr
));
2237 /* Compute the hdr's checksum if necessary */
2238 arc_cksum_compute(buf
);
2244 * If this function is being called to decrypt an encrypted buffer or verify an
2245 * authenticated one, the key must be loaded and a mapping must be made
2246 * available in the keystore via spa_keystore_create_mapping() or one of its
2250 arc_untransform(arc_buf_t
*buf
, spa_t
*spa
, uint64_t dsobj
, boolean_t in_place
)
2252 arc_fill_flags_t flags
= 0;
2255 flags
|= ARC_FILL_IN_PLACE
;
2257 return (arc_buf_fill(buf
, spa
, dsobj
, flags
));
2261 * Increment the amount of evictable space in the arc_state_t's refcount.
2262 * We account for the space used by the hdr and the arc buf individually
2263 * so that we can add and remove them from the refcount individually.
2266 arc_evictable_space_increment(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
2268 arc_buf_contents_t type
= arc_buf_type(hdr
);
2270 ASSERT(HDR_HAS_L1HDR(hdr
));
2272 if (GHOST_STATE(state
)) {
2273 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
2274 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2275 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2276 ASSERT(!HDR_HAS_RABD(hdr
));
2277 (void) refcount_add_many(&state
->arcs_esize
[type
],
2278 HDR_GET_LSIZE(hdr
), hdr
);
2282 ASSERT(!GHOST_STATE(state
));
2283 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2284 (void) refcount_add_many(&state
->arcs_esize
[type
],
2285 arc_hdr_size(hdr
), hdr
);
2287 if (HDR_HAS_RABD(hdr
)) {
2288 (void) refcount_add_many(&state
->arcs_esize
[type
],
2289 HDR_GET_PSIZE(hdr
), hdr
);
2292 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2293 buf
= buf
->b_next
) {
2294 if (arc_buf_is_shared(buf
))
2296 (void) refcount_add_many(&state
->arcs_esize
[type
],
2297 arc_buf_size(buf
), buf
);
2302 * Decrement the amount of evictable space in the arc_state_t's refcount.
2303 * We account for the space used by the hdr and the arc buf individually
2304 * so that we can add and remove them from the refcount individually.
2307 arc_evictable_space_decrement(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
2309 arc_buf_contents_t type
= arc_buf_type(hdr
);
2311 ASSERT(HDR_HAS_L1HDR(hdr
));
2313 if (GHOST_STATE(state
)) {
2314 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
2315 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2316 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2317 ASSERT(!HDR_HAS_RABD(hdr
));
2318 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2319 HDR_GET_LSIZE(hdr
), hdr
);
2323 ASSERT(!GHOST_STATE(state
));
2324 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2325 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2326 arc_hdr_size(hdr
), hdr
);
2328 if (HDR_HAS_RABD(hdr
)) {
2329 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2330 HDR_GET_PSIZE(hdr
), hdr
);
2333 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2334 buf
= buf
->b_next
) {
2335 if (arc_buf_is_shared(buf
))
2337 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2338 arc_buf_size(buf
), buf
);
2343 * Add a reference to this hdr indicating that someone is actively
2344 * referencing that memory. When the refcount transitions from 0 to 1,
2345 * we remove it from the respective arc_state_t list to indicate that
2346 * it is not evictable.
2349 add_reference(arc_buf_hdr_t
*hdr
, void *tag
)
2353 ASSERT(HDR_HAS_L1HDR(hdr
));
2354 if (!MUTEX_HELD(HDR_LOCK(hdr
))) {
2355 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
2356 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2357 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2360 state
= hdr
->b_l1hdr
.b_state
;
2362 if ((refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
) == 1) &&
2363 (state
!= arc_anon
)) {
2364 /* We don't use the L2-only state list. */
2365 if (state
!= arc_l2c_only
) {
2366 multilist_remove(state
->arcs_list
[arc_buf_type(hdr
)],
2368 arc_evictable_space_decrement(hdr
, state
);
2370 /* remove the prefetch flag if we get a reference */
2371 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
2376 * Remove a reference from this hdr. When the reference transitions from
2377 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2378 * list making it eligible for eviction.
2381 remove_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
2384 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2386 ASSERT(HDR_HAS_L1HDR(hdr
));
2387 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
2388 ASSERT(!GHOST_STATE(state
));
2391 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2392 * check to prevent usage of the arc_l2c_only list.
2394 if (((cnt
= refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
)) == 0) &&
2395 (state
!= arc_anon
)) {
2396 multilist_insert(state
->arcs_list
[arc_buf_type(hdr
)], hdr
);
2397 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
2398 arc_evictable_space_increment(hdr
, state
);
2404 * Returns detailed information about a specific arc buffer. When the
2405 * state_index argument is set the function will calculate the arc header
2406 * list position for its arc state. Since this requires a linear traversal
2407 * callers are strongly encourage not to do this. However, it can be helpful
2408 * for targeted analysis so the functionality is provided.
2411 arc_buf_info(arc_buf_t
*ab
, arc_buf_info_t
*abi
, int state_index
)
2413 arc_buf_hdr_t
*hdr
= ab
->b_hdr
;
2414 l1arc_buf_hdr_t
*l1hdr
= NULL
;
2415 l2arc_buf_hdr_t
*l2hdr
= NULL
;
2416 arc_state_t
*state
= NULL
;
2418 memset(abi
, 0, sizeof (arc_buf_info_t
));
2423 abi
->abi_flags
= hdr
->b_flags
;
2425 if (HDR_HAS_L1HDR(hdr
)) {
2426 l1hdr
= &hdr
->b_l1hdr
;
2427 state
= l1hdr
->b_state
;
2429 if (HDR_HAS_L2HDR(hdr
))
2430 l2hdr
= &hdr
->b_l2hdr
;
2433 abi
->abi_bufcnt
= l1hdr
->b_bufcnt
;
2434 abi
->abi_access
= l1hdr
->b_arc_access
;
2435 abi
->abi_mru_hits
= l1hdr
->b_mru_hits
;
2436 abi
->abi_mru_ghost_hits
= l1hdr
->b_mru_ghost_hits
;
2437 abi
->abi_mfu_hits
= l1hdr
->b_mfu_hits
;
2438 abi
->abi_mfu_ghost_hits
= l1hdr
->b_mfu_ghost_hits
;
2439 abi
->abi_holds
= refcount_count(&l1hdr
->b_refcnt
);
2443 abi
->abi_l2arc_dattr
= l2hdr
->b_daddr
;
2444 abi
->abi_l2arc_hits
= l2hdr
->b_hits
;
2447 abi
->abi_state_type
= state
? state
->arcs_state
: ARC_STATE_ANON
;
2448 abi
->abi_state_contents
= arc_buf_type(hdr
);
2449 abi
->abi_size
= arc_hdr_size(hdr
);
2453 * Move the supplied buffer to the indicated state. The hash lock
2454 * for the buffer must be held by the caller.
2457 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*hdr
,
2458 kmutex_t
*hash_lock
)
2460 arc_state_t
*old_state
;
2463 boolean_t update_old
, update_new
;
2464 arc_buf_contents_t buftype
= arc_buf_type(hdr
);
2467 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2468 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2469 * L1 hdr doesn't always exist when we change state to arc_anon before
2470 * destroying a header, in which case reallocating to add the L1 hdr is
2473 if (HDR_HAS_L1HDR(hdr
)) {
2474 old_state
= hdr
->b_l1hdr
.b_state
;
2475 refcnt
= refcount_count(&hdr
->b_l1hdr
.b_refcnt
);
2476 bufcnt
= hdr
->b_l1hdr
.b_bufcnt
;
2477 update_old
= (bufcnt
> 0 || hdr
->b_l1hdr
.b_pabd
!= NULL
||
2480 old_state
= arc_l2c_only
;
2483 update_old
= B_FALSE
;
2485 update_new
= update_old
;
2487 ASSERT(MUTEX_HELD(hash_lock
));
2488 ASSERT3P(new_state
, !=, old_state
);
2489 ASSERT(!GHOST_STATE(new_state
) || bufcnt
== 0);
2490 ASSERT(old_state
!= arc_anon
|| bufcnt
<= 1);
2493 * If this buffer is evictable, transfer it from the
2494 * old state list to the new state list.
2497 if (old_state
!= arc_anon
&& old_state
!= arc_l2c_only
) {
2498 ASSERT(HDR_HAS_L1HDR(hdr
));
2499 multilist_remove(old_state
->arcs_list
[buftype
], hdr
);
2501 if (GHOST_STATE(old_state
)) {
2503 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2504 update_old
= B_TRUE
;
2506 arc_evictable_space_decrement(hdr
, old_state
);
2508 if (new_state
!= arc_anon
&& new_state
!= arc_l2c_only
) {
2510 * An L1 header always exists here, since if we're
2511 * moving to some L1-cached state (i.e. not l2c_only or
2512 * anonymous), we realloc the header to add an L1hdr
2515 ASSERT(HDR_HAS_L1HDR(hdr
));
2516 multilist_insert(new_state
->arcs_list
[buftype
], hdr
);
2518 if (GHOST_STATE(new_state
)) {
2520 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2521 update_new
= B_TRUE
;
2523 arc_evictable_space_increment(hdr
, new_state
);
2527 ASSERT(!HDR_EMPTY(hdr
));
2528 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(hdr
))
2529 buf_hash_remove(hdr
);
2531 /* adjust state sizes (ignore arc_l2c_only) */
2533 if (update_new
&& new_state
!= arc_l2c_only
) {
2534 ASSERT(HDR_HAS_L1HDR(hdr
));
2535 if (GHOST_STATE(new_state
)) {
2539 * When moving a header to a ghost state, we first
2540 * remove all arc buffers. Thus, we'll have a
2541 * bufcnt of zero, and no arc buffer to use for
2542 * the reference. As a result, we use the arc
2543 * header pointer for the reference.
2545 (void) refcount_add_many(&new_state
->arcs_size
,
2546 HDR_GET_LSIZE(hdr
), hdr
);
2547 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2548 ASSERT(!HDR_HAS_RABD(hdr
));
2550 uint32_t buffers
= 0;
2553 * Each individual buffer holds a unique reference,
2554 * thus we must remove each of these references one
2557 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2558 buf
= buf
->b_next
) {
2559 ASSERT3U(bufcnt
, !=, 0);
2563 * When the arc_buf_t is sharing the data
2564 * block with the hdr, the owner of the
2565 * reference belongs to the hdr. Only
2566 * add to the refcount if the arc_buf_t is
2569 if (arc_buf_is_shared(buf
))
2572 (void) refcount_add_many(&new_state
->arcs_size
,
2573 arc_buf_size(buf
), buf
);
2575 ASSERT3U(bufcnt
, ==, buffers
);
2577 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2578 (void) refcount_add_many(&new_state
->arcs_size
,
2579 arc_hdr_size(hdr
), hdr
);
2582 if (HDR_HAS_RABD(hdr
)) {
2583 (void) refcount_add_many(&new_state
->arcs_size
,
2584 HDR_GET_PSIZE(hdr
), hdr
);
2589 if (update_old
&& old_state
!= arc_l2c_only
) {
2590 ASSERT(HDR_HAS_L1HDR(hdr
));
2591 if (GHOST_STATE(old_state
)) {
2593 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2594 ASSERT(!HDR_HAS_RABD(hdr
));
2597 * When moving a header off of a ghost state,
2598 * the header will not contain any arc buffers.
2599 * We use the arc header pointer for the reference
2600 * which is exactly what we did when we put the
2601 * header on the ghost state.
2604 (void) refcount_remove_many(&old_state
->arcs_size
,
2605 HDR_GET_LSIZE(hdr
), hdr
);
2607 uint32_t buffers
= 0;
2610 * Each individual buffer holds a unique reference,
2611 * thus we must remove each of these references one
2614 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2615 buf
= buf
->b_next
) {
2616 ASSERT3U(bufcnt
, !=, 0);
2620 * When the arc_buf_t is sharing the data
2621 * block with the hdr, the owner of the
2622 * reference belongs to the hdr. Only
2623 * add to the refcount if the arc_buf_t is
2626 if (arc_buf_is_shared(buf
))
2629 (void) refcount_remove_many(
2630 &old_state
->arcs_size
, arc_buf_size(buf
),
2633 ASSERT3U(bufcnt
, ==, buffers
);
2634 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
||
2637 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2638 (void) refcount_remove_many(
2639 &old_state
->arcs_size
, arc_hdr_size(hdr
),
2643 if (HDR_HAS_RABD(hdr
)) {
2644 (void) refcount_remove_many(
2645 &old_state
->arcs_size
, HDR_GET_PSIZE(hdr
),
2651 if (HDR_HAS_L1HDR(hdr
))
2652 hdr
->b_l1hdr
.b_state
= new_state
;
2655 * L2 headers should never be on the L2 state list since they don't
2656 * have L1 headers allocated.
2658 ASSERT(multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]) &&
2659 multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]));
2663 arc_space_consume(uint64_t space
, arc_space_type_t type
)
2665 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2670 case ARC_SPACE_DATA
:
2671 ARCSTAT_INCR(arcstat_data_size
, space
);
2673 case ARC_SPACE_META
:
2674 ARCSTAT_INCR(arcstat_metadata_size
, space
);
2676 case ARC_SPACE_BONUS
:
2677 ARCSTAT_INCR(arcstat_bonus_size
, space
);
2679 case ARC_SPACE_DNODE
:
2680 ARCSTAT_INCR(arcstat_dnode_size
, space
);
2682 case ARC_SPACE_DBUF
:
2683 ARCSTAT_INCR(arcstat_dbuf_size
, space
);
2685 case ARC_SPACE_HDRS
:
2686 ARCSTAT_INCR(arcstat_hdr_size
, space
);
2688 case ARC_SPACE_L2HDRS
:
2689 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
2693 if (type
!= ARC_SPACE_DATA
)
2694 ARCSTAT_INCR(arcstat_meta_used
, space
);
2696 atomic_add_64(&arc_size
, space
);
2700 arc_space_return(uint64_t space
, arc_space_type_t type
)
2702 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2707 case ARC_SPACE_DATA
:
2708 ARCSTAT_INCR(arcstat_data_size
, -space
);
2710 case ARC_SPACE_META
:
2711 ARCSTAT_INCR(arcstat_metadata_size
, -space
);
2713 case ARC_SPACE_BONUS
:
2714 ARCSTAT_INCR(arcstat_bonus_size
, -space
);
2716 case ARC_SPACE_DNODE
:
2717 ARCSTAT_INCR(arcstat_dnode_size
, -space
);
2719 case ARC_SPACE_DBUF
:
2720 ARCSTAT_INCR(arcstat_dbuf_size
, -space
);
2722 case ARC_SPACE_HDRS
:
2723 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
2725 case ARC_SPACE_L2HDRS
:
2726 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
2730 if (type
!= ARC_SPACE_DATA
) {
2731 ASSERT(arc_meta_used
>= space
);
2732 if (arc_meta_max
< arc_meta_used
)
2733 arc_meta_max
= arc_meta_used
;
2734 ARCSTAT_INCR(arcstat_meta_used
, -space
);
2737 ASSERT(arc_size
>= space
);
2738 atomic_add_64(&arc_size
, -space
);
2742 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2743 * with the hdr's b_pabd.
2746 arc_can_share(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2749 * The criteria for sharing a hdr's data are:
2750 * 1. the buffer is not encrypted
2751 * 2. the hdr's compression matches the buf's compression
2752 * 3. the hdr doesn't need to be byteswapped
2753 * 4. the hdr isn't already being shared
2754 * 5. the buf is either compressed or it is the last buf in the hdr list
2756 * Criterion #5 maintains the invariant that shared uncompressed
2757 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2758 * might ask, "if a compressed buf is allocated first, won't that be the
2759 * last thing in the list?", but in that case it's impossible to create
2760 * a shared uncompressed buf anyway (because the hdr must be compressed
2761 * to have the compressed buf). You might also think that #3 is
2762 * sufficient to make this guarantee, however it's possible
2763 * (specifically in the rare L2ARC write race mentioned in
2764 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2765 * is sharable, but wasn't at the time of its allocation. Rather than
2766 * allow a new shared uncompressed buf to be created and then shuffle
2767 * the list around to make it the last element, this simply disallows
2768 * sharing if the new buf isn't the first to be added.
2770 ASSERT3P(buf
->b_hdr
, ==, hdr
);
2771 boolean_t hdr_compressed
=
2772 arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
;
2773 boolean_t buf_compressed
= ARC_BUF_COMPRESSED(buf
) != 0;
2774 return (!ARC_BUF_ENCRYPTED(buf
) &&
2775 buf_compressed
== hdr_compressed
&&
2776 hdr
->b_l1hdr
.b_byteswap
== DMU_BSWAP_NUMFUNCS
&&
2777 !HDR_SHARED_DATA(hdr
) &&
2778 (ARC_BUF_LAST(buf
) || ARC_BUF_COMPRESSED(buf
)));
2782 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2783 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2784 * copy was made successfully, or an error code otherwise.
2787 arc_buf_alloc_impl(arc_buf_hdr_t
*hdr
, spa_t
*spa
, uint64_t dsobj
, void *tag
,
2788 boolean_t encrypted
, boolean_t compressed
, boolean_t noauth
,
2789 boolean_t fill
, arc_buf_t
**ret
)
2792 arc_fill_flags_t flags
= ARC_FILL_LOCKED
;
2794 ASSERT(HDR_HAS_L1HDR(hdr
));
2795 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2796 VERIFY(hdr
->b_type
== ARC_BUFC_DATA
||
2797 hdr
->b_type
== ARC_BUFC_METADATA
);
2798 ASSERT3P(ret
, !=, NULL
);
2799 ASSERT3P(*ret
, ==, NULL
);
2800 IMPLY(encrypted
, compressed
);
2802 hdr
->b_l1hdr
.b_mru_hits
= 0;
2803 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
2804 hdr
->b_l1hdr
.b_mfu_hits
= 0;
2805 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
2806 hdr
->b_l1hdr
.b_l2_hits
= 0;
2808 buf
= *ret
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
2811 buf
->b_next
= hdr
->b_l1hdr
.b_buf
;
2814 add_reference(hdr
, tag
);
2817 * We're about to change the hdr's b_flags. We must either
2818 * hold the hash_lock or be undiscoverable.
2820 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2823 * Only honor requests for compressed bufs if the hdr is actually
2824 * compressed. This must be overriden if the buffer is encrypted since
2825 * encrypted buffers cannot be decompressed.
2828 buf
->b_flags
|= ARC_BUF_FLAG_COMPRESSED
;
2829 buf
->b_flags
|= ARC_BUF_FLAG_ENCRYPTED
;
2830 flags
|= ARC_FILL_COMPRESSED
| ARC_FILL_ENCRYPTED
;
2831 } else if (compressed
&&
2832 arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
) {
2833 buf
->b_flags
|= ARC_BUF_FLAG_COMPRESSED
;
2834 flags
|= ARC_FILL_COMPRESSED
;
2839 flags
|= ARC_FILL_NOAUTH
;
2843 * If the hdr's data can be shared then we share the data buffer and
2844 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2845 * allocate a new buffer to store the buf's data.
2847 * There are two additional restrictions here because we're sharing
2848 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2849 * actively involved in an L2ARC write, because if this buf is used by
2850 * an arc_write() then the hdr's data buffer will be released when the
2851 * write completes, even though the L2ARC write might still be using it.
2852 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2853 * need to be ABD-aware.
2855 boolean_t can_share
= arc_can_share(hdr
, buf
) && !HDR_L2_WRITING(hdr
) &&
2856 hdr
->b_l1hdr
.b_pabd
!= NULL
&& abd_is_linear(hdr
->b_l1hdr
.b_pabd
);
2858 /* Set up b_data and sharing */
2860 buf
->b_data
= abd_to_buf(hdr
->b_l1hdr
.b_pabd
);
2861 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2862 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2865 arc_get_data_buf(hdr
, arc_buf_size(buf
), buf
);
2866 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2868 VERIFY3P(buf
->b_data
, !=, NULL
);
2870 hdr
->b_l1hdr
.b_buf
= buf
;
2871 hdr
->b_l1hdr
.b_bufcnt
+= 1;
2873 hdr
->b_crypt_hdr
.b_ebufcnt
+= 1;
2876 * If the user wants the data from the hdr, we need to either copy or
2877 * decompress the data.
2880 return (arc_buf_fill(buf
, spa
, dsobj
, flags
));
2886 static char *arc_onloan_tag
= "onloan";
2889 arc_loaned_bytes_update(int64_t delta
)
2891 atomic_add_64(&arc_loaned_bytes
, delta
);
2893 /* assert that it did not wrap around */
2894 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
2898 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2899 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2900 * buffers must be returned to the arc before they can be used by the DMU or
2904 arc_loan_buf(spa_t
*spa
, boolean_t is_metadata
, int size
)
2906 arc_buf_t
*buf
= arc_alloc_buf(spa
, arc_onloan_tag
,
2907 is_metadata
? ARC_BUFC_METADATA
: ARC_BUFC_DATA
, size
);
2909 arc_loaned_bytes_update(size
);
2915 arc_loan_compressed_buf(spa_t
*spa
, uint64_t psize
, uint64_t lsize
,
2916 enum zio_compress compression_type
)
2918 arc_buf_t
*buf
= arc_alloc_compressed_buf(spa
, arc_onloan_tag
,
2919 psize
, lsize
, compression_type
);
2921 arc_loaned_bytes_update(psize
);
2927 arc_loan_raw_buf(spa_t
*spa
, uint64_t dsobj
, boolean_t byteorder
,
2928 const uint8_t *salt
, const uint8_t *iv
, const uint8_t *mac
,
2929 dmu_object_type_t ot
, uint64_t psize
, uint64_t lsize
,
2930 enum zio_compress compression_type
)
2932 arc_buf_t
*buf
= arc_alloc_raw_buf(spa
, arc_onloan_tag
, dsobj
,
2933 byteorder
, salt
, iv
, mac
, ot
, psize
, lsize
, compression_type
);
2935 atomic_add_64(&arc_loaned_bytes
, psize
);
2941 * Return a loaned arc buffer to the arc.
2944 arc_return_buf(arc_buf_t
*buf
, void *tag
)
2946 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2948 ASSERT3P(buf
->b_data
, !=, NULL
);
2949 ASSERT(HDR_HAS_L1HDR(hdr
));
2950 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2951 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2953 arc_loaned_bytes_update(-arc_buf_size(buf
));
2956 /* Detach an arc_buf from a dbuf (tag) */
2958 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
2960 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2962 ASSERT3P(buf
->b_data
, !=, NULL
);
2963 ASSERT(HDR_HAS_L1HDR(hdr
));
2964 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2965 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2967 arc_loaned_bytes_update(arc_buf_size(buf
));
2971 l2arc_free_abd_on_write(abd_t
*abd
, size_t size
, arc_buf_contents_t type
)
2973 l2arc_data_free_t
*df
= kmem_alloc(sizeof (*df
), KM_SLEEP
);
2976 df
->l2df_size
= size
;
2977 df
->l2df_type
= type
;
2978 mutex_enter(&l2arc_free_on_write_mtx
);
2979 list_insert_head(l2arc_free_on_write
, df
);
2980 mutex_exit(&l2arc_free_on_write_mtx
);
2984 arc_hdr_free_on_write(arc_buf_hdr_t
*hdr
, boolean_t free_rdata
)
2986 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2987 arc_buf_contents_t type
= arc_buf_type(hdr
);
2988 uint64_t size
= (free_rdata
) ? HDR_GET_PSIZE(hdr
) : arc_hdr_size(hdr
);
2990 /* protected by hash lock, if in the hash table */
2991 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
2992 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2993 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
2995 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2998 (void) refcount_remove_many(&state
->arcs_size
, size
, hdr
);
2999 if (type
== ARC_BUFC_METADATA
) {
3000 arc_space_return(size
, ARC_SPACE_META
);
3002 ASSERT(type
== ARC_BUFC_DATA
);
3003 arc_space_return(size
, ARC_SPACE_DATA
);
3007 l2arc_free_abd_on_write(hdr
->b_crypt_hdr
.b_rabd
, size
, type
);
3009 l2arc_free_abd_on_write(hdr
->b_l1hdr
.b_pabd
, size
, type
);
3014 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
3015 * data buffer, we transfer the refcount ownership to the hdr and update
3016 * the appropriate kstats.
3019 arc_share_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
3021 ASSERT(arc_can_share(hdr
, buf
));
3022 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
3023 ASSERT(!ARC_BUF_ENCRYPTED(buf
));
3024 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
3027 * Start sharing the data buffer. We transfer the
3028 * refcount ownership to the hdr since it always owns
3029 * the refcount whenever an arc_buf_t is shared.
3031 refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, buf
, hdr
);
3032 hdr
->b_l1hdr
.b_pabd
= abd_get_from_buf(buf
->b_data
, arc_buf_size(buf
));
3033 abd_take_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
,
3034 HDR_ISTYPE_METADATA(hdr
));
3035 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
3036 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
3039 * Since we've transferred ownership to the hdr we need
3040 * to increment its compressed and uncompressed kstats and
3041 * decrement the overhead size.
3043 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
3044 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
3045 ARCSTAT_INCR(arcstat_overhead_size
, -arc_buf_size(buf
));
3049 arc_unshare_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
3051 ASSERT(arc_buf_is_shared(buf
));
3052 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
3053 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
3056 * We are no longer sharing this buffer so we need
3057 * to transfer its ownership to the rightful owner.
3059 refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, hdr
, buf
);
3060 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
3061 abd_release_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
);
3062 abd_put(hdr
->b_l1hdr
.b_pabd
);
3063 hdr
->b_l1hdr
.b_pabd
= NULL
;
3064 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
3067 * Since the buffer is no longer shared between
3068 * the arc buf and the hdr, count it as overhead.
3070 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
3071 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
3072 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
3076 * Remove an arc_buf_t from the hdr's buf list and return the last
3077 * arc_buf_t on the list. If no buffers remain on the list then return
3081 arc_buf_remove(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
3083 ASSERT(HDR_HAS_L1HDR(hdr
));
3084 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
3086 arc_buf_t
**bufp
= &hdr
->b_l1hdr
.b_buf
;
3087 arc_buf_t
*lastbuf
= NULL
;
3090 * Remove the buf from the hdr list and locate the last
3091 * remaining buffer on the list.
3093 while (*bufp
!= NULL
) {
3095 *bufp
= buf
->b_next
;
3098 * If we've removed a buffer in the middle of
3099 * the list then update the lastbuf and update
3102 if (*bufp
!= NULL
) {
3104 bufp
= &(*bufp
)->b_next
;
3108 ASSERT3P(lastbuf
, !=, buf
);
3109 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, lastbuf
!= NULL
);
3110 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, hdr
->b_l1hdr
.b_buf
!= NULL
);
3111 IMPLY(lastbuf
!= NULL
, ARC_BUF_LAST(lastbuf
));
3117 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3121 arc_buf_destroy_impl(arc_buf_t
*buf
)
3123 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3126 * Free up the data associated with the buf but only if we're not
3127 * sharing this with the hdr. If we are sharing it with the hdr, the
3128 * hdr is responsible for doing the free.
3130 if (buf
->b_data
!= NULL
) {
3132 * We're about to change the hdr's b_flags. We must either
3133 * hold the hash_lock or be undiscoverable.
3135 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
3137 arc_cksum_verify(buf
);
3138 arc_buf_unwatch(buf
);
3140 if (arc_buf_is_shared(buf
)) {
3141 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
3143 uint64_t size
= arc_buf_size(buf
);
3144 arc_free_data_buf(hdr
, buf
->b_data
, size
, buf
);
3145 ARCSTAT_INCR(arcstat_overhead_size
, -size
);
3149 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
3150 hdr
->b_l1hdr
.b_bufcnt
-= 1;
3152 if (ARC_BUF_ENCRYPTED(buf
))
3153 hdr
->b_crypt_hdr
.b_ebufcnt
-= 1;
3156 * If we have no more encrypted buffers and we've already
3157 * gotten a copy of the decrypted data we can free b_rabd to
3160 if (hdr
->b_crypt_hdr
.b_ebufcnt
== 0 && HDR_HAS_RABD(hdr
) &&
3161 hdr
->b_l1hdr
.b_pabd
!= NULL
&& !HDR_IO_IN_PROGRESS(hdr
)) {
3162 arc_hdr_free_abd(hdr
, B_TRUE
);
3166 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
3168 if (ARC_BUF_SHARED(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
3170 * If the current arc_buf_t is sharing its data buffer with the
3171 * hdr, then reassign the hdr's b_pabd to share it with the new
3172 * buffer at the end of the list. The shared buffer is always
3173 * the last one on the hdr's buffer list.
3175 * There is an equivalent case for compressed bufs, but since
3176 * they aren't guaranteed to be the last buf in the list and
3177 * that is an exceedingly rare case, we just allow that space be
3178 * wasted temporarily. We must also be careful not to share
3179 * encrypted buffers, since they cannot be shared.
3181 if (lastbuf
!= NULL
&& !ARC_BUF_ENCRYPTED(lastbuf
)) {
3182 /* Only one buf can be shared at once */
3183 VERIFY(!arc_buf_is_shared(lastbuf
));
3184 /* hdr is uncompressed so can't have compressed buf */
3185 VERIFY(!ARC_BUF_COMPRESSED(lastbuf
));
3187 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
3188 arc_hdr_free_abd(hdr
, B_FALSE
);
3191 * We must setup a new shared block between the
3192 * last buffer and the hdr. The data would have
3193 * been allocated by the arc buf so we need to transfer
3194 * ownership to the hdr since it's now being shared.
3196 arc_share_buf(hdr
, lastbuf
);
3198 } else if (HDR_SHARED_DATA(hdr
)) {
3200 * Uncompressed shared buffers are always at the end
3201 * of the list. Compressed buffers don't have the
3202 * same requirements. This makes it hard to
3203 * simply assert that the lastbuf is shared so
3204 * we rely on the hdr's compression flags to determine
3205 * if we have a compressed, shared buffer.
3207 ASSERT3P(lastbuf
, !=, NULL
);
3208 ASSERT(arc_buf_is_shared(lastbuf
) ||
3209 arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
);
3213 * Free the checksum if we're removing the last uncompressed buf from
3216 if (!arc_hdr_has_uncompressed_buf(hdr
)) {
3217 arc_cksum_free(hdr
);
3220 /* clean up the buf */
3222 kmem_cache_free(buf_cache
, buf
);
3226 arc_hdr_alloc_abd(arc_buf_hdr_t
*hdr
, boolean_t alloc_rdata
)
3230 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
3231 ASSERT(HDR_HAS_L1HDR(hdr
));
3232 ASSERT(!HDR_SHARED_DATA(hdr
) || alloc_rdata
);
3233 IMPLY(alloc_rdata
, HDR_PROTECTED(hdr
));
3235 if (hdr
->b_l1hdr
.b_pabd
== NULL
&& !HDR_HAS_RABD(hdr
))
3236 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
3239 size
= HDR_GET_PSIZE(hdr
);
3240 ASSERT3P(hdr
->b_crypt_hdr
.b_rabd
, ==, NULL
);
3241 hdr
->b_crypt_hdr
.b_rabd
= arc_get_data_abd(hdr
, size
, hdr
);
3242 ASSERT3P(hdr
->b_crypt_hdr
.b_rabd
, !=, NULL
);
3243 ARCSTAT_INCR(arcstat_raw_size
, size
);
3245 size
= arc_hdr_size(hdr
);
3246 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
3247 hdr
->b_l1hdr
.b_pabd
= arc_get_data_abd(hdr
, size
, hdr
);
3248 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
3251 ARCSTAT_INCR(arcstat_compressed_size
, size
);
3252 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
3256 arc_hdr_free_abd(arc_buf_hdr_t
*hdr
, boolean_t free_rdata
)
3258 uint64_t size
= (free_rdata
) ? HDR_GET_PSIZE(hdr
) : arc_hdr_size(hdr
);
3260 ASSERT(HDR_HAS_L1HDR(hdr
));
3261 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
|| HDR_HAS_RABD(hdr
));
3262 IMPLY(free_rdata
, HDR_HAS_RABD(hdr
));
3265 * If the hdr is currently being written to the l2arc then
3266 * we defer freeing the data by adding it to the l2arc_free_on_write
3267 * list. The l2arc will free the data once it's finished
3268 * writing it to the l2arc device.
3270 if (HDR_L2_WRITING(hdr
)) {
3271 arc_hdr_free_on_write(hdr
, free_rdata
);
3272 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
3273 } else if (free_rdata
) {
3274 arc_free_data_abd(hdr
, hdr
->b_crypt_hdr
.b_rabd
, size
, hdr
);
3276 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
, size
, hdr
);
3280 hdr
->b_crypt_hdr
.b_rabd
= NULL
;
3281 ARCSTAT_INCR(arcstat_raw_size
, -size
);
3283 hdr
->b_l1hdr
.b_pabd
= NULL
;
3286 if (hdr
->b_l1hdr
.b_pabd
== NULL
&& !HDR_HAS_RABD(hdr
))
3287 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
3289 ARCSTAT_INCR(arcstat_compressed_size
, -size
);
3290 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
3293 static arc_buf_hdr_t
*
3294 arc_hdr_alloc(uint64_t spa
, int32_t psize
, int32_t lsize
,
3295 boolean_t
protected, enum zio_compress compression_type
,
3296 arc_buf_contents_t type
, boolean_t alloc_rdata
)
3300 VERIFY(type
== ARC_BUFC_DATA
|| type
== ARC_BUFC_METADATA
);
3302 hdr
= kmem_cache_alloc(hdr_full_crypt_cache
, KM_PUSHPAGE
);
3304 hdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
3307 ASSERT(HDR_EMPTY(hdr
));
3308 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
3309 HDR_SET_PSIZE(hdr
, psize
);
3310 HDR_SET_LSIZE(hdr
, lsize
);
3314 arc_hdr_set_flags(hdr
, arc_bufc_to_flags(type
) | ARC_FLAG_HAS_L1HDR
);
3315 arc_hdr_set_compress(hdr
, compression_type
);
3317 arc_hdr_set_flags(hdr
, ARC_FLAG_PROTECTED
);
3319 hdr
->b_l1hdr
.b_state
= arc_anon
;
3320 hdr
->b_l1hdr
.b_arc_access
= 0;
3321 hdr
->b_l1hdr
.b_bufcnt
= 0;
3322 hdr
->b_l1hdr
.b_buf
= NULL
;
3325 * Allocate the hdr's buffer. This will contain either
3326 * the compressed or uncompressed data depending on the block
3327 * it references and compressed arc enablement.
3329 arc_hdr_alloc_abd(hdr
, alloc_rdata
);
3330 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3336 * Transition between the two allocation states for the arc_buf_hdr struct.
3337 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3338 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3339 * version is used when a cache buffer is only in the L2ARC in order to reduce
3342 static arc_buf_hdr_t
*
3343 arc_hdr_realloc(arc_buf_hdr_t
*hdr
, kmem_cache_t
*old
, kmem_cache_t
*new)
3345 ASSERT(HDR_HAS_L2HDR(hdr
));
3347 arc_buf_hdr_t
*nhdr
;
3348 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
3350 ASSERT((old
== hdr_full_cache
&& new == hdr_l2only_cache
) ||
3351 (old
== hdr_l2only_cache
&& new == hdr_full_cache
));
3354 * if the caller wanted a new full header and the header is to be
3355 * encrypted we will actually allocate the header from the full crypt
3356 * cache instead. The same applies to freeing from the old cache.
3358 if (HDR_PROTECTED(hdr
) && new == hdr_full_cache
)
3359 new = hdr_full_crypt_cache
;
3360 if (HDR_PROTECTED(hdr
) && old
== hdr_full_cache
)
3361 old
= hdr_full_crypt_cache
;
3363 nhdr
= kmem_cache_alloc(new, KM_PUSHPAGE
);
3365 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
3366 buf_hash_remove(hdr
);
3368 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
3370 if (new == hdr_full_cache
|| new == hdr_full_crypt_cache
) {
3371 arc_hdr_set_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
3373 * arc_access and arc_change_state need to be aware that a
3374 * header has just come out of L2ARC, so we set its state to
3375 * l2c_only even though it's about to change.
3377 nhdr
->b_l1hdr
.b_state
= arc_l2c_only
;
3379 /* Verify previous threads set to NULL before freeing */
3380 ASSERT3P(nhdr
->b_l1hdr
.b_pabd
, ==, NULL
);
3381 ASSERT(!HDR_HAS_RABD(hdr
));
3383 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
3384 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
3385 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
3388 * If we've reached here, We must have been called from
3389 * arc_evict_hdr(), as such we should have already been
3390 * removed from any ghost list we were previously on
3391 * (which protects us from racing with arc_evict_state),
3392 * thus no locking is needed during this check.
3394 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3397 * A buffer must not be moved into the arc_l2c_only
3398 * state if it's not finished being written out to the
3399 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3400 * might try to be accessed, even though it was removed.
3402 VERIFY(!HDR_L2_WRITING(hdr
));
3403 VERIFY3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
3404 ASSERT(!HDR_HAS_RABD(hdr
));
3406 arc_hdr_clear_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
3409 * The header has been reallocated so we need to re-insert it into any
3412 (void) buf_hash_insert(nhdr
, NULL
);
3414 ASSERT(list_link_active(&hdr
->b_l2hdr
.b_l2node
));
3416 mutex_enter(&dev
->l2ad_mtx
);
3419 * We must place the realloc'ed header back into the list at
3420 * the same spot. Otherwise, if it's placed earlier in the list,
3421 * l2arc_write_buffers() could find it during the function's
3422 * write phase, and try to write it out to the l2arc.
3424 list_insert_after(&dev
->l2ad_buflist
, hdr
, nhdr
);
3425 list_remove(&dev
->l2ad_buflist
, hdr
);
3427 mutex_exit(&dev
->l2ad_mtx
);
3430 * Since we're using the pointer address as the tag when
3431 * incrementing and decrementing the l2ad_alloc refcount, we
3432 * must remove the old pointer (that we're about to destroy) and
3433 * add the new pointer to the refcount. Otherwise we'd remove
3434 * the wrong pointer address when calling arc_hdr_destroy() later.
3437 (void) refcount_remove_many(&dev
->l2ad_alloc
, arc_hdr_size(hdr
), hdr
);
3438 (void) refcount_add_many(&dev
->l2ad_alloc
, arc_hdr_size(nhdr
), nhdr
);
3440 buf_discard_identity(hdr
);
3441 kmem_cache_free(old
, hdr
);
3447 * This function allows an L1 header to be reallocated as a crypt
3448 * header and vice versa. If we are going to a crypt header, the
3449 * new fields will be zeroed out.
3451 static arc_buf_hdr_t
*
3452 arc_hdr_realloc_crypt(arc_buf_hdr_t
*hdr
, boolean_t need_crypt
)
3454 arc_buf_hdr_t
*nhdr
;
3456 kmem_cache_t
*ncache
, *ocache
;
3458 ASSERT(HDR_HAS_L1HDR(hdr
));
3459 ASSERT3U(!!HDR_PROTECTED(hdr
), !=, need_crypt
);
3460 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
3461 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3464 ncache
= hdr_full_crypt_cache
;
3465 ocache
= hdr_full_cache
;
3467 ncache
= hdr_full_cache
;
3468 ocache
= hdr_full_crypt_cache
;
3471 nhdr
= kmem_cache_alloc(ncache
, KM_PUSHPAGE
);
3472 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
3473 nhdr
->b_l1hdr
.b_freeze_cksum
= hdr
->b_l1hdr
.b_freeze_cksum
;
3474 nhdr
->b_l1hdr
.b_bufcnt
= hdr
->b_l1hdr
.b_bufcnt
;
3475 nhdr
->b_l1hdr
.b_byteswap
= hdr
->b_l1hdr
.b_byteswap
;
3476 nhdr
->b_l1hdr
.b_state
= hdr
->b_l1hdr
.b_state
;
3477 nhdr
->b_l1hdr
.b_arc_access
= hdr
->b_l1hdr
.b_arc_access
;
3478 nhdr
->b_l1hdr
.b_mru_hits
= hdr
->b_l1hdr
.b_mru_hits
;
3479 nhdr
->b_l1hdr
.b_mru_ghost_hits
= hdr
->b_l1hdr
.b_mru_ghost_hits
;
3480 nhdr
->b_l1hdr
.b_mfu_hits
= hdr
->b_l1hdr
.b_mfu_hits
;
3481 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= hdr
->b_l1hdr
.b_mfu_ghost_hits
;
3482 nhdr
->b_l1hdr
.b_l2_hits
= hdr
->b_l1hdr
.b_l2_hits
;
3483 nhdr
->b_l1hdr
.b_acb
= hdr
->b_l1hdr
.b_acb
;
3484 nhdr
->b_l1hdr
.b_pabd
= hdr
->b_l1hdr
.b_pabd
;
3485 nhdr
->b_l1hdr
.b_buf
= hdr
->b_l1hdr
.b_buf
;
3488 * This refcount_add() exists only to ensure that the individual
3489 * arc buffers always point to a header that is referenced, avoiding
3490 * a small race condition that could trigger ASSERTs.
3492 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, FTAG
);
3494 for (buf
= nhdr
->b_l1hdr
.b_buf
; buf
!= NULL
; buf
= buf
->b_next
) {
3495 mutex_enter(&buf
->b_evict_lock
);
3497 mutex_exit(&buf
->b_evict_lock
);
3500 refcount_transfer(&nhdr
->b_l1hdr
.b_refcnt
, &hdr
->b_l1hdr
.b_refcnt
);
3501 (void) refcount_remove(&nhdr
->b_l1hdr
.b_refcnt
, FTAG
);
3504 arc_hdr_set_flags(nhdr
, ARC_FLAG_PROTECTED
);
3506 arc_hdr_clear_flags(nhdr
, ARC_FLAG_PROTECTED
);
3509 buf_discard_identity(hdr
);
3510 kmem_cache_free(ocache
, hdr
);
3516 * This function is used by the send / receive code to convert a newly
3517 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3518 * is also used to allow the root objset block to be uupdated without altering
3519 * its embedded MACs. Both block types will always be uncompressed so we do not
3520 * have to worry about compression type or psize.
3523 arc_convert_to_raw(arc_buf_t
*buf
, uint64_t dsobj
, boolean_t byteorder
,
3524 dmu_object_type_t ot
, const uint8_t *salt
, const uint8_t *iv
,
3527 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3529 ASSERT(ot
== DMU_OT_DNODE
|| ot
== DMU_OT_OBJSET
);
3530 ASSERT(HDR_HAS_L1HDR(hdr
));
3531 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
3533 buf
->b_flags
|= (ARC_BUF_FLAG_COMPRESSED
| ARC_BUF_FLAG_ENCRYPTED
);
3534 if (!HDR_PROTECTED(hdr
))
3535 hdr
= arc_hdr_realloc_crypt(hdr
, B_TRUE
);
3536 hdr
->b_crypt_hdr
.b_dsobj
= dsobj
;
3537 hdr
->b_crypt_hdr
.b_ot
= ot
;
3538 hdr
->b_l1hdr
.b_byteswap
= (byteorder
== ZFS_HOST_BYTEORDER
) ?
3539 DMU_BSWAP_NUMFUNCS
: DMU_OT_BYTESWAP(ot
);
3540 if (!arc_hdr_has_uncompressed_buf(hdr
))
3541 arc_cksum_free(hdr
);
3544 bcopy(salt
, hdr
->b_crypt_hdr
.b_salt
, ZIO_DATA_SALT_LEN
);
3546 bcopy(iv
, hdr
->b_crypt_hdr
.b_iv
, ZIO_DATA_IV_LEN
);
3548 bcopy(mac
, hdr
->b_crypt_hdr
.b_mac
, ZIO_DATA_MAC_LEN
);
3552 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3553 * The buf is returned thawed since we expect the consumer to modify it.
3556 arc_alloc_buf(spa_t
*spa
, void *tag
, arc_buf_contents_t type
, int32_t size
)
3558 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), size
, size
,
3559 B_FALSE
, ZIO_COMPRESS_OFF
, type
, B_FALSE
);
3560 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
3562 arc_buf_t
*buf
= NULL
;
3563 VERIFY0(arc_buf_alloc_impl(hdr
, spa
, 0, tag
, B_FALSE
, B_FALSE
,
3564 B_FALSE
, B_FALSE
, &buf
));
3571 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3572 * for bufs containing metadata.
3575 arc_alloc_compressed_buf(spa_t
*spa
, void *tag
, uint64_t psize
, uint64_t lsize
,
3576 enum zio_compress compression_type
)
3578 ASSERT3U(lsize
, >, 0);
3579 ASSERT3U(lsize
, >=, psize
);
3580 ASSERT3U(compression_type
, >, ZIO_COMPRESS_OFF
);
3581 ASSERT3U(compression_type
, <, ZIO_COMPRESS_FUNCTIONS
);
3583 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
3584 B_FALSE
, compression_type
, ARC_BUFC_DATA
, B_FALSE
);
3585 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
3587 arc_buf_t
*buf
= NULL
;
3588 VERIFY0(arc_buf_alloc_impl(hdr
, spa
, 0, tag
, B_FALSE
,
3589 B_TRUE
, B_FALSE
, B_FALSE
, &buf
));
3591 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
3593 if (!arc_buf_is_shared(buf
)) {
3595 * To ensure that the hdr has the correct data in it if we call
3596 * arc_untransform() on this buf before it's been written to
3597 * disk, it's easiest if we just set up sharing between the
3600 ASSERT(!abd_is_linear(hdr
->b_l1hdr
.b_pabd
));
3601 arc_hdr_free_abd(hdr
, B_FALSE
);
3602 arc_share_buf(hdr
, buf
);
3609 arc_alloc_raw_buf(spa_t
*spa
, void *tag
, uint64_t dsobj
, boolean_t byteorder
,
3610 const uint8_t *salt
, const uint8_t *iv
, const uint8_t *mac
,
3611 dmu_object_type_t ot
, uint64_t psize
, uint64_t lsize
,
3612 enum zio_compress compression_type
)
3616 arc_buf_contents_t type
= DMU_OT_IS_METADATA(ot
) ?
3617 ARC_BUFC_METADATA
: ARC_BUFC_DATA
;
3619 ASSERT3U(lsize
, >, 0);
3620 ASSERT3U(lsize
, >=, psize
);
3621 ASSERT3U(compression_type
, >=, ZIO_COMPRESS_OFF
);
3622 ASSERT3U(compression_type
, <, ZIO_COMPRESS_FUNCTIONS
);
3624 hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
, B_TRUE
,
3625 compression_type
, type
, B_TRUE
);
3626 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
3628 hdr
->b_crypt_hdr
.b_dsobj
= dsobj
;
3629 hdr
->b_crypt_hdr
.b_ot
= ot
;
3630 hdr
->b_l1hdr
.b_byteswap
= (byteorder
== ZFS_HOST_BYTEORDER
) ?
3631 DMU_BSWAP_NUMFUNCS
: DMU_OT_BYTESWAP(ot
);
3632 bcopy(salt
, hdr
->b_crypt_hdr
.b_salt
, ZIO_DATA_SALT_LEN
);
3633 bcopy(iv
, hdr
->b_crypt_hdr
.b_iv
, ZIO_DATA_IV_LEN
);
3634 bcopy(mac
, hdr
->b_crypt_hdr
.b_mac
, ZIO_DATA_MAC_LEN
);
3637 * This buffer will be considered encrypted even if the ot is not an
3638 * encrypted type. It will become authenticated instead in
3639 * arc_write_ready().
3642 VERIFY0(arc_buf_alloc_impl(hdr
, spa
, dsobj
, tag
, B_TRUE
, B_TRUE
,
3643 B_FALSE
, B_FALSE
, &buf
));
3645 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
3651 arc_hdr_l2hdr_destroy(arc_buf_hdr_t
*hdr
)
3653 l2arc_buf_hdr_t
*l2hdr
= &hdr
->b_l2hdr
;
3654 l2arc_dev_t
*dev
= l2hdr
->b_dev
;
3655 uint64_t psize
= arc_hdr_size(hdr
);
3657 ASSERT(MUTEX_HELD(&dev
->l2ad_mtx
));
3658 ASSERT(HDR_HAS_L2HDR(hdr
));
3660 list_remove(&dev
->l2ad_buflist
, hdr
);
3662 ARCSTAT_INCR(arcstat_l2_psize
, -psize
);
3663 ARCSTAT_INCR(arcstat_l2_lsize
, -HDR_GET_LSIZE(hdr
));
3665 vdev_space_update(dev
->l2ad_vdev
, -psize
, 0, 0);
3667 (void) refcount_remove_many(&dev
->l2ad_alloc
, psize
, hdr
);
3668 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
3672 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
3674 if (HDR_HAS_L1HDR(hdr
)) {
3675 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
||
3676 hdr
->b_l1hdr
.b_bufcnt
> 0);
3677 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3678 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
3680 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3681 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
3683 if (!HDR_EMPTY(hdr
))
3684 buf_discard_identity(hdr
);
3686 if (HDR_HAS_L2HDR(hdr
)) {
3687 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
3688 boolean_t buflist_held
= MUTEX_HELD(&dev
->l2ad_mtx
);
3691 mutex_enter(&dev
->l2ad_mtx
);
3694 * Even though we checked this conditional above, we
3695 * need to check this again now that we have the
3696 * l2ad_mtx. This is because we could be racing with
3697 * another thread calling l2arc_evict() which might have
3698 * destroyed this header's L2 portion as we were waiting
3699 * to acquire the l2ad_mtx. If that happens, we don't
3700 * want to re-destroy the header's L2 portion.
3702 if (HDR_HAS_L2HDR(hdr
))
3703 arc_hdr_l2hdr_destroy(hdr
);
3706 mutex_exit(&dev
->l2ad_mtx
);
3709 if (HDR_HAS_L1HDR(hdr
)) {
3710 arc_cksum_free(hdr
);
3712 while (hdr
->b_l1hdr
.b_buf
!= NULL
)
3713 arc_buf_destroy_impl(hdr
->b_l1hdr
.b_buf
);
3715 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
3716 arc_hdr_free_abd(hdr
, B_FALSE
);
3719 if (HDR_HAS_RABD(hdr
))
3720 arc_hdr_free_abd(hdr
, B_TRUE
);
3723 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
3724 if (HDR_HAS_L1HDR(hdr
)) {
3725 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3726 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
3728 if (!HDR_PROTECTED(hdr
)) {
3729 kmem_cache_free(hdr_full_cache
, hdr
);
3731 kmem_cache_free(hdr_full_crypt_cache
, hdr
);
3734 kmem_cache_free(hdr_l2only_cache
, hdr
);
3739 arc_buf_destroy(arc_buf_t
*buf
, void* tag
)
3741 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3742 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
3744 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
3745 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
3746 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3747 VERIFY0(remove_reference(hdr
, NULL
, tag
));
3748 arc_hdr_destroy(hdr
);
3752 mutex_enter(hash_lock
);
3753 ASSERT3P(hdr
, ==, buf
->b_hdr
);
3754 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
3755 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3756 ASSERT3P(hdr
->b_l1hdr
.b_state
, !=, arc_anon
);
3757 ASSERT3P(buf
->b_data
, !=, NULL
);
3759 (void) remove_reference(hdr
, hash_lock
, tag
);
3760 arc_buf_destroy_impl(buf
);
3761 mutex_exit(hash_lock
);
3765 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3766 * state of the header is dependent on its state prior to entering this
3767 * function. The following transitions are possible:
3769 * - arc_mru -> arc_mru_ghost
3770 * - arc_mfu -> arc_mfu_ghost
3771 * - arc_mru_ghost -> arc_l2c_only
3772 * - arc_mru_ghost -> deleted
3773 * - arc_mfu_ghost -> arc_l2c_only
3774 * - arc_mfu_ghost -> deleted
3777 arc_evict_hdr(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
3779 arc_state_t
*evicted_state
, *state
;
3780 int64_t bytes_evicted
= 0;
3782 ASSERT(MUTEX_HELD(hash_lock
));
3783 ASSERT(HDR_HAS_L1HDR(hdr
));
3785 state
= hdr
->b_l1hdr
.b_state
;
3786 if (GHOST_STATE(state
)) {
3787 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3788 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
3791 * l2arc_write_buffers() relies on a header's L1 portion
3792 * (i.e. its b_pabd field) during it's write phase.
3793 * Thus, we cannot push a header onto the arc_l2c_only
3794 * state (removing its L1 piece) until the header is
3795 * done being written to the l2arc.
3797 if (HDR_HAS_L2HDR(hdr
) && HDR_L2_WRITING(hdr
)) {
3798 ARCSTAT_BUMP(arcstat_evict_l2_skip
);
3799 return (bytes_evicted
);
3802 ARCSTAT_BUMP(arcstat_deleted
);
3803 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3805 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, hdr
);
3807 if (HDR_HAS_L2HDR(hdr
)) {
3808 ASSERT(hdr
->b_l1hdr
.b_pabd
== NULL
);
3809 ASSERT(!HDR_HAS_RABD(hdr
));
3811 * This buffer is cached on the 2nd Level ARC;
3812 * don't destroy the header.
3814 arc_change_state(arc_l2c_only
, hdr
, hash_lock
);
3816 * dropping from L1+L2 cached to L2-only,
3817 * realloc to remove the L1 header.
3819 hdr
= arc_hdr_realloc(hdr
, hdr_full_cache
,
3822 arc_change_state(arc_anon
, hdr
, hash_lock
);
3823 arc_hdr_destroy(hdr
);
3825 return (bytes_evicted
);
3828 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
3829 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
3831 /* prefetch buffers have a minimum lifespan */
3832 if (HDR_IO_IN_PROGRESS(hdr
) ||
3833 ((hdr
->b_flags
& (ARC_FLAG_PREFETCH
| ARC_FLAG_INDIRECT
)) &&
3834 ddi_get_lbolt() - hdr
->b_l1hdr
.b_arc_access
<
3835 arc_min_prefetch_lifespan
)) {
3836 ARCSTAT_BUMP(arcstat_evict_skip
);
3837 return (bytes_evicted
);
3840 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
3841 while (hdr
->b_l1hdr
.b_buf
) {
3842 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
3843 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
3844 ARCSTAT_BUMP(arcstat_mutex_miss
);
3847 if (buf
->b_data
!= NULL
)
3848 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3849 mutex_exit(&buf
->b_evict_lock
);
3850 arc_buf_destroy_impl(buf
);
3853 if (HDR_HAS_L2HDR(hdr
)) {
3854 ARCSTAT_INCR(arcstat_evict_l2_cached
, HDR_GET_LSIZE(hdr
));
3856 if (l2arc_write_eligible(hdr
->b_spa
, hdr
)) {
3857 ARCSTAT_INCR(arcstat_evict_l2_eligible
,
3858 HDR_GET_LSIZE(hdr
));
3860 ARCSTAT_INCR(arcstat_evict_l2_ineligible
,
3861 HDR_GET_LSIZE(hdr
));
3865 if (hdr
->b_l1hdr
.b_bufcnt
== 0) {
3866 arc_cksum_free(hdr
);
3868 bytes_evicted
+= arc_hdr_size(hdr
);
3871 * If this hdr is being evicted and has a compressed
3872 * buffer then we discard it here before we change states.
3873 * This ensures that the accounting is updated correctly
3874 * in arc_free_data_impl().
3876 if (hdr
->b_l1hdr
.b_pabd
!= NULL
)
3877 arc_hdr_free_abd(hdr
, B_FALSE
);
3879 if (HDR_HAS_RABD(hdr
))
3880 arc_hdr_free_abd(hdr
, B_TRUE
);
3882 arc_change_state(evicted_state
, hdr
, hash_lock
);
3883 ASSERT(HDR_IN_HASH_TABLE(hdr
));
3884 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
3885 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, hdr
);
3888 return (bytes_evicted
);
3892 arc_evict_state_impl(multilist_t
*ml
, int idx
, arc_buf_hdr_t
*marker
,
3893 uint64_t spa
, int64_t bytes
)
3895 multilist_sublist_t
*mls
;
3896 uint64_t bytes_evicted
= 0;
3898 kmutex_t
*hash_lock
;
3899 int evict_count
= 0;
3901 ASSERT3P(marker
, !=, NULL
);
3902 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3904 mls
= multilist_sublist_lock(ml
, idx
);
3906 for (hdr
= multilist_sublist_prev(mls
, marker
); hdr
!= NULL
;
3907 hdr
= multilist_sublist_prev(mls
, marker
)) {
3908 if ((bytes
!= ARC_EVICT_ALL
&& bytes_evicted
>= bytes
) ||
3909 (evict_count
>= zfs_arc_evict_batch_limit
))
3913 * To keep our iteration location, move the marker
3914 * forward. Since we're not holding hdr's hash lock, we
3915 * must be very careful and not remove 'hdr' from the
3916 * sublist. Otherwise, other consumers might mistake the
3917 * 'hdr' as not being on a sublist when they call the
3918 * multilist_link_active() function (they all rely on
3919 * the hash lock protecting concurrent insertions and
3920 * removals). multilist_sublist_move_forward() was
3921 * specifically implemented to ensure this is the case
3922 * (only 'marker' will be removed and re-inserted).
3924 multilist_sublist_move_forward(mls
, marker
);
3927 * The only case where the b_spa field should ever be
3928 * zero, is the marker headers inserted by
3929 * arc_evict_state(). It's possible for multiple threads
3930 * to be calling arc_evict_state() concurrently (e.g.
3931 * dsl_pool_close() and zio_inject_fault()), so we must
3932 * skip any markers we see from these other threads.
3934 if (hdr
->b_spa
== 0)
3937 /* we're only interested in evicting buffers of a certain spa */
3938 if (spa
!= 0 && hdr
->b_spa
!= spa
) {
3939 ARCSTAT_BUMP(arcstat_evict_skip
);
3943 hash_lock
= HDR_LOCK(hdr
);
3946 * We aren't calling this function from any code path
3947 * that would already be holding a hash lock, so we're
3948 * asserting on this assumption to be defensive in case
3949 * this ever changes. Without this check, it would be
3950 * possible to incorrectly increment arcstat_mutex_miss
3951 * below (e.g. if the code changed such that we called
3952 * this function with a hash lock held).
3954 ASSERT(!MUTEX_HELD(hash_lock
));
3956 if (mutex_tryenter(hash_lock
)) {
3957 uint64_t evicted
= arc_evict_hdr(hdr
, hash_lock
);
3958 mutex_exit(hash_lock
);
3960 bytes_evicted
+= evicted
;
3963 * If evicted is zero, arc_evict_hdr() must have
3964 * decided to skip this header, don't increment
3965 * evict_count in this case.
3971 * If arc_size isn't overflowing, signal any
3972 * threads that might happen to be waiting.
3974 * For each header evicted, we wake up a single
3975 * thread. If we used cv_broadcast, we could
3976 * wake up "too many" threads causing arc_size
3977 * to significantly overflow arc_c; since
3978 * arc_get_data_impl() doesn't check for overflow
3979 * when it's woken up (it doesn't because it's
3980 * possible for the ARC to be overflowing while
3981 * full of un-evictable buffers, and the
3982 * function should proceed in this case).
3984 * If threads are left sleeping, due to not
3985 * using cv_broadcast, they will be woken up
3986 * just before arc_reclaim_thread() sleeps.
3988 mutex_enter(&arc_reclaim_lock
);
3989 if (!arc_is_overflowing())
3990 cv_signal(&arc_reclaim_waiters_cv
);
3991 mutex_exit(&arc_reclaim_lock
);
3993 ARCSTAT_BUMP(arcstat_mutex_miss
);
3997 multilist_sublist_unlock(mls
);
3999 return (bytes_evicted
);
4003 * Evict buffers from the given arc state, until we've removed the
4004 * specified number of bytes. Move the removed buffers to the
4005 * appropriate evict state.
4007 * This function makes a "best effort". It skips over any buffers
4008 * it can't get a hash_lock on, and so, may not catch all candidates.
4009 * It may also return without evicting as much space as requested.
4011 * If bytes is specified using the special value ARC_EVICT_ALL, this
4012 * will evict all available (i.e. unlocked and evictable) buffers from
4013 * the given arc state; which is used by arc_flush().
4016 arc_evict_state(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
4017 arc_buf_contents_t type
)
4019 uint64_t total_evicted
= 0;
4020 multilist_t
*ml
= state
->arcs_list
[type
];
4022 arc_buf_hdr_t
**markers
;
4024 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
4026 num_sublists
= multilist_get_num_sublists(ml
);
4029 * If we've tried to evict from each sublist, made some
4030 * progress, but still have not hit the target number of bytes
4031 * to evict, we want to keep trying. The markers allow us to
4032 * pick up where we left off for each individual sublist, rather
4033 * than starting from the tail each time.
4035 markers
= kmem_zalloc(sizeof (*markers
) * num_sublists
, KM_SLEEP
);
4036 for (int i
= 0; i
< num_sublists
; i
++) {
4037 multilist_sublist_t
*mls
;
4039 markers
[i
] = kmem_cache_alloc(hdr_full_cache
, KM_SLEEP
);
4042 * A b_spa of 0 is used to indicate that this header is
4043 * a marker. This fact is used in arc_adjust_type() and
4044 * arc_evict_state_impl().
4046 markers
[i
]->b_spa
= 0;
4048 mls
= multilist_sublist_lock(ml
, i
);
4049 multilist_sublist_insert_tail(mls
, markers
[i
]);
4050 multilist_sublist_unlock(mls
);
4054 * While we haven't hit our target number of bytes to evict, or
4055 * we're evicting all available buffers.
4057 while (total_evicted
< bytes
|| bytes
== ARC_EVICT_ALL
) {
4058 int sublist_idx
= multilist_get_random_index(ml
);
4059 uint64_t scan_evicted
= 0;
4062 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
4063 * Request that 10% of the LRUs be scanned by the superblock
4066 if (type
== ARC_BUFC_DATA
&& arc_dnode_size
> arc_dnode_limit
)
4067 arc_prune_async((arc_dnode_size
- arc_dnode_limit
) /
4068 sizeof (dnode_t
) / zfs_arc_dnode_reduce_percent
);
4071 * Start eviction using a randomly selected sublist,
4072 * this is to try and evenly balance eviction across all
4073 * sublists. Always starting at the same sublist
4074 * (e.g. index 0) would cause evictions to favor certain
4075 * sublists over others.
4077 for (int i
= 0; i
< num_sublists
; i
++) {
4078 uint64_t bytes_remaining
;
4079 uint64_t bytes_evicted
;
4081 if (bytes
== ARC_EVICT_ALL
)
4082 bytes_remaining
= ARC_EVICT_ALL
;
4083 else if (total_evicted
< bytes
)
4084 bytes_remaining
= bytes
- total_evicted
;
4088 bytes_evicted
= arc_evict_state_impl(ml
, sublist_idx
,
4089 markers
[sublist_idx
], spa
, bytes_remaining
);
4091 scan_evicted
+= bytes_evicted
;
4092 total_evicted
+= bytes_evicted
;
4094 /* we've reached the end, wrap to the beginning */
4095 if (++sublist_idx
>= num_sublists
)
4100 * If we didn't evict anything during this scan, we have
4101 * no reason to believe we'll evict more during another
4102 * scan, so break the loop.
4104 if (scan_evicted
== 0) {
4105 /* This isn't possible, let's make that obvious */
4106 ASSERT3S(bytes
, !=, 0);
4109 * When bytes is ARC_EVICT_ALL, the only way to
4110 * break the loop is when scan_evicted is zero.
4111 * In that case, we actually have evicted enough,
4112 * so we don't want to increment the kstat.
4114 if (bytes
!= ARC_EVICT_ALL
) {
4115 ASSERT3S(total_evicted
, <, bytes
);
4116 ARCSTAT_BUMP(arcstat_evict_not_enough
);
4123 for (int i
= 0; i
< num_sublists
; i
++) {
4124 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
4125 multilist_sublist_remove(mls
, markers
[i
]);
4126 multilist_sublist_unlock(mls
);
4128 kmem_cache_free(hdr_full_cache
, markers
[i
]);
4130 kmem_free(markers
, sizeof (*markers
) * num_sublists
);
4132 return (total_evicted
);
4136 * Flush all "evictable" data of the given type from the arc state
4137 * specified. This will not evict any "active" buffers (i.e. referenced).
4139 * When 'retry' is set to B_FALSE, the function will make a single pass
4140 * over the state and evict any buffers that it can. Since it doesn't
4141 * continually retry the eviction, it might end up leaving some buffers
4142 * in the ARC due to lock misses.
4144 * When 'retry' is set to B_TRUE, the function will continually retry the
4145 * eviction until *all* evictable buffers have been removed from the
4146 * state. As a result, if concurrent insertions into the state are
4147 * allowed (e.g. if the ARC isn't shutting down), this function might
4148 * wind up in an infinite loop, continually trying to evict buffers.
4151 arc_flush_state(arc_state_t
*state
, uint64_t spa
, arc_buf_contents_t type
,
4154 uint64_t evicted
= 0;
4156 while (refcount_count(&state
->arcs_esize
[type
]) != 0) {
4157 evicted
+= arc_evict_state(state
, spa
, ARC_EVICT_ALL
, type
);
4167 * Helper function for arc_prune_async() it is responsible for safely
4168 * handling the execution of a registered arc_prune_func_t.
4171 arc_prune_task(void *ptr
)
4173 arc_prune_t
*ap
= (arc_prune_t
*)ptr
;
4174 arc_prune_func_t
*func
= ap
->p_pfunc
;
4177 func(ap
->p_adjust
, ap
->p_private
);
4179 refcount_remove(&ap
->p_refcnt
, func
);
4183 * Notify registered consumers they must drop holds on a portion of the ARC
4184 * buffered they reference. This provides a mechanism to ensure the ARC can
4185 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
4186 * is analogous to dnlc_reduce_cache() but more generic.
4188 * This operation is performed asynchronously so it may be safely called
4189 * in the context of the arc_reclaim_thread(). A reference is taken here
4190 * for each registered arc_prune_t and the arc_prune_task() is responsible
4191 * for releasing it once the registered arc_prune_func_t has completed.
4194 arc_prune_async(int64_t adjust
)
4198 mutex_enter(&arc_prune_mtx
);
4199 for (ap
= list_head(&arc_prune_list
); ap
!= NULL
;
4200 ap
= list_next(&arc_prune_list
, ap
)) {
4202 if (refcount_count(&ap
->p_refcnt
) >= 2)
4205 refcount_add(&ap
->p_refcnt
, ap
->p_pfunc
);
4206 ap
->p_adjust
= adjust
;
4207 if (taskq_dispatch(arc_prune_taskq
, arc_prune_task
,
4208 ap
, TQ_SLEEP
) == TASKQID_INVALID
) {
4209 refcount_remove(&ap
->p_refcnt
, ap
->p_pfunc
);
4212 ARCSTAT_BUMP(arcstat_prune
);
4214 mutex_exit(&arc_prune_mtx
);
4218 * Evict the specified number of bytes from the state specified,
4219 * restricting eviction to the spa and type given. This function
4220 * prevents us from trying to evict more from a state's list than
4221 * is "evictable", and to skip evicting altogether when passed a
4222 * negative value for "bytes". In contrast, arc_evict_state() will
4223 * evict everything it can, when passed a negative value for "bytes".
4226 arc_adjust_impl(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
4227 arc_buf_contents_t type
)
4231 if (bytes
> 0 && refcount_count(&state
->arcs_esize
[type
]) > 0) {
4232 delta
= MIN(refcount_count(&state
->arcs_esize
[type
]), bytes
);
4233 return (arc_evict_state(state
, spa
, delta
, type
));
4240 * The goal of this function is to evict enough meta data buffers from the
4241 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4242 * more complicated than it appears because it is common for data buffers
4243 * to have holds on meta data buffers. In addition, dnode meta data buffers
4244 * will be held by the dnodes in the block preventing them from being freed.
4245 * This means we can't simply traverse the ARC and expect to always find
4246 * enough unheld meta data buffer to release.
4248 * Therefore, this function has been updated to make alternating passes
4249 * over the ARC releasing data buffers and then newly unheld meta data
4250 * buffers. This ensures forward progress is maintained and arc_meta_used
4251 * will decrease. Normally this is sufficient, but if required the ARC
4252 * will call the registered prune callbacks causing dentry and inodes to
4253 * be dropped from the VFS cache. This will make dnode meta data buffers
4254 * available for reclaim.
4257 arc_adjust_meta_balanced(void)
4259 int64_t delta
, prune
= 0, adjustmnt
;
4260 uint64_t total_evicted
= 0;
4261 arc_buf_contents_t type
= ARC_BUFC_DATA
;
4262 int restarts
= MAX(zfs_arc_meta_adjust_restarts
, 0);
4266 * This slightly differs than the way we evict from the mru in
4267 * arc_adjust because we don't have a "target" value (i.e. no
4268 * "meta" arc_p). As a result, I think we can completely
4269 * cannibalize the metadata in the MRU before we evict the
4270 * metadata from the MFU. I think we probably need to implement a
4271 * "metadata arc_p" value to do this properly.
4273 adjustmnt
= arc_meta_used
- arc_meta_limit
;
4275 if (adjustmnt
> 0 && refcount_count(&arc_mru
->arcs_esize
[type
]) > 0) {
4276 delta
= MIN(refcount_count(&arc_mru
->arcs_esize
[type
]),
4278 total_evicted
+= arc_adjust_impl(arc_mru
, 0, delta
, type
);
4283 * We can't afford to recalculate adjustmnt here. If we do,
4284 * new metadata buffers can sneak into the MRU or ANON lists,
4285 * thus penalize the MFU metadata. Although the fudge factor is
4286 * small, it has been empirically shown to be significant for
4287 * certain workloads (e.g. creating many empty directories). As
4288 * such, we use the original calculation for adjustmnt, and
4289 * simply decrement the amount of data evicted from the MRU.
4292 if (adjustmnt
> 0 && refcount_count(&arc_mfu
->arcs_esize
[type
]) > 0) {
4293 delta
= MIN(refcount_count(&arc_mfu
->arcs_esize
[type
]),
4295 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, delta
, type
);
4298 adjustmnt
= arc_meta_used
- arc_meta_limit
;
4300 if (adjustmnt
> 0 &&
4301 refcount_count(&arc_mru_ghost
->arcs_esize
[type
]) > 0) {
4302 delta
= MIN(adjustmnt
,
4303 refcount_count(&arc_mru_ghost
->arcs_esize
[type
]));
4304 total_evicted
+= arc_adjust_impl(arc_mru_ghost
, 0, delta
, type
);
4308 if (adjustmnt
> 0 &&
4309 refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]) > 0) {
4310 delta
= MIN(adjustmnt
,
4311 refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]));
4312 total_evicted
+= arc_adjust_impl(arc_mfu_ghost
, 0, delta
, type
);
4316 * If after attempting to make the requested adjustment to the ARC
4317 * the meta limit is still being exceeded then request that the
4318 * higher layers drop some cached objects which have holds on ARC
4319 * meta buffers. Requests to the upper layers will be made with
4320 * increasingly large scan sizes until the ARC is below the limit.
4322 if (arc_meta_used
> arc_meta_limit
) {
4323 if (type
== ARC_BUFC_DATA
) {
4324 type
= ARC_BUFC_METADATA
;
4326 type
= ARC_BUFC_DATA
;
4328 if (zfs_arc_meta_prune
) {
4329 prune
+= zfs_arc_meta_prune
;
4330 arc_prune_async(prune
);
4339 return (total_evicted
);
4343 * Evict metadata buffers from the cache, such that arc_meta_used is
4344 * capped by the arc_meta_limit tunable.
4347 arc_adjust_meta_only(void)
4349 uint64_t total_evicted
= 0;
4353 * If we're over the meta limit, we want to evict enough
4354 * metadata to get back under the meta limit. We don't want to
4355 * evict so much that we drop the MRU below arc_p, though. If
4356 * we're over the meta limit more than we're over arc_p, we
4357 * evict some from the MRU here, and some from the MFU below.
4359 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
4360 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
4361 refcount_count(&arc_mru
->arcs_size
) - arc_p
));
4363 total_evicted
+= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
4366 * Similar to the above, we want to evict enough bytes to get us
4367 * below the meta limit, but not so much as to drop us below the
4368 * space allotted to the MFU (which is defined as arc_c - arc_p).
4370 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
4371 (int64_t)(refcount_count(&arc_mfu
->arcs_size
) - (arc_c
- arc_p
)));
4373 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
4375 return (total_evicted
);
4379 arc_adjust_meta(void)
4381 if (zfs_arc_meta_strategy
== ARC_STRATEGY_META_ONLY
)
4382 return (arc_adjust_meta_only());
4384 return (arc_adjust_meta_balanced());
4388 * Return the type of the oldest buffer in the given arc state
4390 * This function will select a random sublist of type ARC_BUFC_DATA and
4391 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4392 * is compared, and the type which contains the "older" buffer will be
4395 static arc_buf_contents_t
4396 arc_adjust_type(arc_state_t
*state
)
4398 multilist_t
*data_ml
= state
->arcs_list
[ARC_BUFC_DATA
];
4399 multilist_t
*meta_ml
= state
->arcs_list
[ARC_BUFC_METADATA
];
4400 int data_idx
= multilist_get_random_index(data_ml
);
4401 int meta_idx
= multilist_get_random_index(meta_ml
);
4402 multilist_sublist_t
*data_mls
;
4403 multilist_sublist_t
*meta_mls
;
4404 arc_buf_contents_t type
;
4405 arc_buf_hdr_t
*data_hdr
;
4406 arc_buf_hdr_t
*meta_hdr
;
4409 * We keep the sublist lock until we're finished, to prevent
4410 * the headers from being destroyed via arc_evict_state().
4412 data_mls
= multilist_sublist_lock(data_ml
, data_idx
);
4413 meta_mls
= multilist_sublist_lock(meta_ml
, meta_idx
);
4416 * These two loops are to ensure we skip any markers that
4417 * might be at the tail of the lists due to arc_evict_state().
4420 for (data_hdr
= multilist_sublist_tail(data_mls
); data_hdr
!= NULL
;
4421 data_hdr
= multilist_sublist_prev(data_mls
, data_hdr
)) {
4422 if (data_hdr
->b_spa
!= 0)
4426 for (meta_hdr
= multilist_sublist_tail(meta_mls
); meta_hdr
!= NULL
;
4427 meta_hdr
= multilist_sublist_prev(meta_mls
, meta_hdr
)) {
4428 if (meta_hdr
->b_spa
!= 0)
4432 if (data_hdr
== NULL
&& meta_hdr
== NULL
) {
4433 type
= ARC_BUFC_DATA
;
4434 } else if (data_hdr
== NULL
) {
4435 ASSERT3P(meta_hdr
, !=, NULL
);
4436 type
= ARC_BUFC_METADATA
;
4437 } else if (meta_hdr
== NULL
) {
4438 ASSERT3P(data_hdr
, !=, NULL
);
4439 type
= ARC_BUFC_DATA
;
4441 ASSERT3P(data_hdr
, !=, NULL
);
4442 ASSERT3P(meta_hdr
, !=, NULL
);
4444 /* The headers can't be on the sublist without an L1 header */
4445 ASSERT(HDR_HAS_L1HDR(data_hdr
));
4446 ASSERT(HDR_HAS_L1HDR(meta_hdr
));
4448 if (data_hdr
->b_l1hdr
.b_arc_access
<
4449 meta_hdr
->b_l1hdr
.b_arc_access
) {
4450 type
= ARC_BUFC_DATA
;
4452 type
= ARC_BUFC_METADATA
;
4456 multilist_sublist_unlock(meta_mls
);
4457 multilist_sublist_unlock(data_mls
);
4463 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4468 uint64_t total_evicted
= 0;
4473 * If we're over arc_meta_limit, we want to correct that before
4474 * potentially evicting data buffers below.
4476 total_evicted
+= arc_adjust_meta();
4481 * If we're over the target cache size, we want to evict enough
4482 * from the list to get back to our target size. We don't want
4483 * to evict too much from the MRU, such that it drops below
4484 * arc_p. So, if we're over our target cache size more than
4485 * the MRU is over arc_p, we'll evict enough to get back to
4486 * arc_p here, and then evict more from the MFU below.
4488 target
= MIN((int64_t)(arc_size
- arc_c
),
4489 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
4490 refcount_count(&arc_mru
->arcs_size
) + arc_meta_used
- arc_p
));
4493 * If we're below arc_meta_min, always prefer to evict data.
4494 * Otherwise, try to satisfy the requested number of bytes to
4495 * evict from the type which contains older buffers; in an
4496 * effort to keep newer buffers in the cache regardless of their
4497 * type. If we cannot satisfy the number of bytes from this
4498 * type, spill over into the next type.
4500 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
4501 arc_meta_used
> arc_meta_min
) {
4502 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
4503 total_evicted
+= bytes
;
4506 * If we couldn't evict our target number of bytes from
4507 * metadata, we try to get the rest from data.
4512 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
4514 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
4515 total_evicted
+= bytes
;
4518 * If we couldn't evict our target number of bytes from
4519 * data, we try to get the rest from metadata.
4524 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
4530 * Now that we've tried to evict enough from the MRU to get its
4531 * size back to arc_p, if we're still above the target cache
4532 * size, we evict the rest from the MFU.
4534 target
= arc_size
- arc_c
;
4536 if (arc_adjust_type(arc_mfu
) == ARC_BUFC_METADATA
&&
4537 arc_meta_used
> arc_meta_min
) {
4538 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
4539 total_evicted
+= bytes
;
4542 * If we couldn't evict our target number of bytes from
4543 * metadata, we try to get the rest from data.
4548 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
4550 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
4551 total_evicted
+= bytes
;
4554 * If we couldn't evict our target number of bytes from
4555 * data, we try to get the rest from data.
4560 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
4564 * Adjust ghost lists
4566 * In addition to the above, the ARC also defines target values
4567 * for the ghost lists. The sum of the mru list and mru ghost
4568 * list should never exceed the target size of the cache, and
4569 * the sum of the mru list, mfu list, mru ghost list, and mfu
4570 * ghost list should never exceed twice the target size of the
4571 * cache. The following logic enforces these limits on the ghost
4572 * caches, and evicts from them as needed.
4574 target
= refcount_count(&arc_mru
->arcs_size
) +
4575 refcount_count(&arc_mru_ghost
->arcs_size
) - arc_c
;
4577 bytes
= arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_DATA
);
4578 total_evicted
+= bytes
;
4583 arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_METADATA
);
4586 * We assume the sum of the mru list and mfu list is less than
4587 * or equal to arc_c (we enforced this above), which means we
4588 * can use the simpler of the two equations below:
4590 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4591 * mru ghost + mfu ghost <= arc_c
4593 target
= refcount_count(&arc_mru_ghost
->arcs_size
) +
4594 refcount_count(&arc_mfu_ghost
->arcs_size
) - arc_c
;
4596 bytes
= arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_DATA
);
4597 total_evicted
+= bytes
;
4602 arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_METADATA
);
4604 return (total_evicted
);
4608 arc_flush(spa_t
*spa
, boolean_t retry
)
4613 * If retry is B_TRUE, a spa must not be specified since we have
4614 * no good way to determine if all of a spa's buffers have been
4615 * evicted from an arc state.
4617 ASSERT(!retry
|| spa
== 0);
4620 guid
= spa_load_guid(spa
);
4622 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_DATA
, retry
);
4623 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_METADATA
, retry
);
4625 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_DATA
, retry
);
4626 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_METADATA
, retry
);
4628 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_DATA
, retry
);
4629 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
4631 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_DATA
, retry
);
4632 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
4636 arc_shrink(int64_t to_free
)
4640 if (c
> to_free
&& c
- to_free
> arc_c_min
) {
4641 arc_c
= c
- to_free
;
4642 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
4643 if (arc_c
> arc_size
)
4644 arc_c
= MAX(arc_size
, arc_c_min
);
4646 arc_p
= (arc_c
>> 1);
4647 ASSERT(arc_c
>= arc_c_min
);
4648 ASSERT((int64_t)arc_p
>= 0);
4653 if (arc_size
> arc_c
)
4654 (void) arc_adjust();
4658 * Return maximum amount of memory that we could possibly use. Reduced
4659 * to half of all memory in user space which is primarily used for testing.
4662 arc_all_memory(void)
4665 #ifdef CONFIG_HIGHMEM
4666 return (ptob(totalram_pages
- totalhigh_pages
));
4668 return (ptob(totalram_pages
));
4669 #endif /* CONFIG_HIGHMEM */
4671 return (ptob(physmem
) / 2);
4672 #endif /* _KERNEL */
4676 * Return the amount of memory that is considered free. In user space
4677 * which is primarily used for testing we pretend that free memory ranges
4678 * from 0-20% of all memory.
4681 arc_free_memory(void)
4684 #ifdef CONFIG_HIGHMEM
4687 return (ptob(si
.freeram
- si
.freehigh
));
4689 #ifdef ZFS_GLOBAL_NODE_PAGE_STATE
4690 return (ptob(nr_free_pages() +
4691 global_node_page_state(NR_INACTIVE_FILE
) +
4692 global_node_page_state(NR_INACTIVE_ANON
) +
4693 global_node_page_state(NR_SLAB_RECLAIMABLE
)));
4695 return (ptob(nr_free_pages() +
4696 global_page_state(NR_INACTIVE_FILE
) +
4697 global_page_state(NR_INACTIVE_ANON
) +
4698 global_page_state(NR_SLAB_RECLAIMABLE
)));
4699 #endif /* ZFS_GLOBAL_NODE_PAGE_STATE */
4700 #endif /* CONFIG_HIGHMEM */
4702 return (spa_get_random(arc_all_memory() * 20 / 100));
4703 #endif /* _KERNEL */
4706 typedef enum free_memory_reason_t
{
4711 FMR_PAGES_PP_MAXIMUM
,
4714 } free_memory_reason_t
;
4716 int64_t last_free_memory
;
4717 free_memory_reason_t last_free_reason
;
4721 * Additional reserve of pages for pp_reserve.
4723 int64_t arc_pages_pp_reserve
= 64;
4726 * Additional reserve of pages for swapfs.
4728 int64_t arc_swapfs_reserve
= 64;
4729 #endif /* _KERNEL */
4732 * Return the amount of memory that can be consumed before reclaim will be
4733 * needed. Positive if there is sufficient free memory, negative indicates
4734 * the amount of memory that needs to be freed up.
4737 arc_available_memory(void)
4739 int64_t lowest
= INT64_MAX
;
4740 free_memory_reason_t r
= FMR_UNKNOWN
;
4747 pgcnt_t needfree
= btop(arc_need_free
);
4748 pgcnt_t lotsfree
= btop(arc_sys_free
);
4749 pgcnt_t desfree
= 0;
4750 pgcnt_t freemem
= btop(arc_free_memory());
4754 n
= PAGESIZE
* (-needfree
);
4762 * check that we're out of range of the pageout scanner. It starts to
4763 * schedule paging if freemem is less than lotsfree and needfree.
4764 * lotsfree is the high-water mark for pageout, and needfree is the
4765 * number of needed free pages. We add extra pages here to make sure
4766 * the scanner doesn't start up while we're freeing memory.
4768 n
= PAGESIZE
* (freemem
- lotsfree
- needfree
- desfree
);
4776 * check to make sure that swapfs has enough space so that anon
4777 * reservations can still succeed. anon_resvmem() checks that the
4778 * availrmem is greater than swapfs_minfree, and the number of reserved
4779 * swap pages. We also add a bit of extra here just to prevent
4780 * circumstances from getting really dire.
4782 n
= PAGESIZE
* (availrmem
- swapfs_minfree
- swapfs_reserve
-
4783 desfree
- arc_swapfs_reserve
);
4786 r
= FMR_SWAPFS_MINFREE
;
4790 * Check that we have enough availrmem that memory locking (e.g., via
4791 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4792 * stores the number of pages that cannot be locked; when availrmem
4793 * drops below pages_pp_maximum, page locking mechanisms such as
4794 * page_pp_lock() will fail.)
4796 n
= PAGESIZE
* (availrmem
- pages_pp_maximum
-
4797 arc_pages_pp_reserve
);
4800 r
= FMR_PAGES_PP_MAXIMUM
;
4806 * If we're on a 32-bit platform, it's possible that we'll exhaust the
4807 * kernel heap space before we ever run out of available physical
4808 * memory. Most checks of the size of the heap_area compare against
4809 * tune.t_minarmem, which is the minimum available real memory that we
4810 * can have in the system. However, this is generally fixed at 25 pages
4811 * which is so low that it's useless. In this comparison, we seek to
4812 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4813 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4816 n
= vmem_size(heap_arena
, VMEM_FREE
) -
4817 (vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
) >> 2);
4825 * If zio data pages are being allocated out of a separate heap segment,
4826 * then enforce that the size of available vmem for this arena remains
4827 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4829 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4830 * memory (in the zio_arena) free, which can avoid memory
4831 * fragmentation issues.
4833 if (zio_arena
!= NULL
) {
4834 n
= (int64_t)vmem_size(zio_arena
, VMEM_FREE
) -
4835 (vmem_size(zio_arena
, VMEM_ALLOC
) >>
4836 arc_zio_arena_free_shift
);
4843 /* Every 100 calls, free a small amount */
4844 if (spa_get_random(100) == 0)
4846 #endif /* _KERNEL */
4848 last_free_memory
= lowest
;
4849 last_free_reason
= r
;
4855 * Determine if the system is under memory pressure and is asking
4856 * to reclaim memory. A return value of B_TRUE indicates that the system
4857 * is under memory pressure and that the arc should adjust accordingly.
4860 arc_reclaim_needed(void)
4862 return (arc_available_memory() < 0);
4866 arc_kmem_reap_now(void)
4869 kmem_cache_t
*prev_cache
= NULL
;
4870 kmem_cache_t
*prev_data_cache
= NULL
;
4871 extern kmem_cache_t
*zio_buf_cache
[];
4872 extern kmem_cache_t
*zio_data_buf_cache
[];
4873 extern kmem_cache_t
*range_seg_cache
;
4876 if ((arc_meta_used
>= arc_meta_limit
) && zfs_arc_meta_prune
) {
4878 * We are exceeding our meta-data cache limit.
4879 * Prune some entries to release holds on meta-data.
4881 arc_prune_async(zfs_arc_meta_prune
);
4885 * Reclaim unused memory from all kmem caches.
4891 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
4893 /* reach upper limit of cache size on 32-bit */
4894 if (zio_buf_cache
[i
] == NULL
)
4897 if (zio_buf_cache
[i
] != prev_cache
) {
4898 prev_cache
= zio_buf_cache
[i
];
4899 kmem_cache_reap_now(zio_buf_cache
[i
]);
4901 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
4902 prev_data_cache
= zio_data_buf_cache
[i
];
4903 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
4906 kmem_cache_reap_now(buf_cache
);
4907 kmem_cache_reap_now(hdr_full_cache
);
4908 kmem_cache_reap_now(hdr_l2only_cache
);
4909 kmem_cache_reap_now(range_seg_cache
);
4911 if (zio_arena
!= NULL
) {
4913 * Ask the vmem arena to reclaim unused memory from its
4916 vmem_qcache_reap(zio_arena
);
4921 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4922 * enough data and signal them to proceed. When this happens, the threads in
4923 * arc_get_data_impl() are sleeping while holding the hash lock for their
4924 * particular arc header. Thus, we must be careful to never sleep on a
4925 * hash lock in this thread. This is to prevent the following deadlock:
4927 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4928 * waiting for the reclaim thread to signal it.
4930 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4931 * fails, and goes to sleep forever.
4933 * This possible deadlock is avoided by always acquiring a hash lock
4934 * using mutex_tryenter() from arc_reclaim_thread().
4938 arc_reclaim_thread(void *unused
)
4940 fstrans_cookie_t cookie
= spl_fstrans_mark();
4941 hrtime_t growtime
= 0;
4944 CALLB_CPR_INIT(&cpr
, &arc_reclaim_lock
, callb_generic_cpr
, FTAG
);
4946 mutex_enter(&arc_reclaim_lock
);
4947 while (!arc_reclaim_thread_exit
) {
4948 uint64_t evicted
= 0;
4949 uint64_t need_free
= arc_need_free
;
4950 arc_tuning_update();
4953 * This is necessary in order for the mdb ::arc dcmd to
4954 * show up to date information. Since the ::arc command
4955 * does not call the kstat's update function, without
4956 * this call, the command may show stale stats for the
4957 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4958 * with this change, the data might be up to 1 second
4959 * out of date; but that should suffice. The arc_state_t
4960 * structures can be queried directly if more accurate
4961 * information is needed.
4964 if (arc_ksp
!= NULL
)
4965 arc_ksp
->ks_update(arc_ksp
, KSTAT_READ
);
4967 mutex_exit(&arc_reclaim_lock
);
4970 * We call arc_adjust() before (possibly) calling
4971 * arc_kmem_reap_now(), so that we can wake up
4972 * arc_get_data_buf() sooner.
4974 evicted
= arc_adjust();
4976 int64_t free_memory
= arc_available_memory();
4977 if (free_memory
< 0) {
4979 arc_no_grow
= B_TRUE
;
4983 * Wait at least zfs_grow_retry (default 5) seconds
4984 * before considering growing.
4986 growtime
= gethrtime() + SEC2NSEC(arc_grow_retry
);
4988 arc_kmem_reap_now();
4991 * If we are still low on memory, shrink the ARC
4992 * so that we have arc_shrink_min free space.
4994 free_memory
= arc_available_memory();
4997 (arc_c
>> arc_shrink_shift
) - free_memory
;
5000 to_free
= MAX(to_free
, need_free
);
5002 arc_shrink(to_free
);
5004 } else if (free_memory
< arc_c
>> arc_no_grow_shift
) {
5005 arc_no_grow
= B_TRUE
;
5006 } else if (gethrtime() >= growtime
) {
5007 arc_no_grow
= B_FALSE
;
5010 mutex_enter(&arc_reclaim_lock
);
5013 * If evicted is zero, we couldn't evict anything via
5014 * arc_adjust(). This could be due to hash lock
5015 * collisions, but more likely due to the majority of
5016 * arc buffers being unevictable. Therefore, even if
5017 * arc_size is above arc_c, another pass is unlikely to
5018 * be helpful and could potentially cause us to enter an
5021 if (arc_size
<= arc_c
|| evicted
== 0) {
5023 * We're either no longer overflowing, or we
5024 * can't evict anything more, so we should wake
5025 * up any threads before we go to sleep and remove
5026 * the bytes we were working on from arc_need_free
5027 * since nothing more will be done here.
5029 cv_broadcast(&arc_reclaim_waiters_cv
);
5030 ARCSTAT_INCR(arcstat_need_free
, -need_free
);
5033 * Block until signaled, or after one second (we
5034 * might need to perform arc_kmem_reap_now()
5035 * even if we aren't being signalled)
5037 CALLB_CPR_SAFE_BEGIN(&cpr
);
5038 (void) cv_timedwait_sig_hires(&arc_reclaim_thread_cv
,
5039 &arc_reclaim_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
5040 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_lock
);
5044 arc_reclaim_thread_exit
= B_FALSE
;
5045 cv_broadcast(&arc_reclaim_thread_cv
);
5046 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_lock */
5047 spl_fstrans_unmark(cookie
);
5053 * Determine the amount of memory eligible for eviction contained in the
5054 * ARC. All clean data reported by the ghost lists can always be safely
5055 * evicted. Due to arc_c_min, the same does not hold for all clean data
5056 * contained by the regular mru and mfu lists.
5058 * In the case of the regular mru and mfu lists, we need to report as
5059 * much clean data as possible, such that evicting that same reported
5060 * data will not bring arc_size below arc_c_min. Thus, in certain
5061 * circumstances, the total amount of clean data in the mru and mfu
5062 * lists might not actually be evictable.
5064 * The following two distinct cases are accounted for:
5066 * 1. The sum of the amount of dirty data contained by both the mru and
5067 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5068 * is greater than or equal to arc_c_min.
5069 * (i.e. amount of dirty data >= arc_c_min)
5071 * This is the easy case; all clean data contained by the mru and mfu
5072 * lists is evictable. Evicting all clean data can only drop arc_size
5073 * to the amount of dirty data, which is greater than arc_c_min.
5075 * 2. The sum of the amount of dirty data contained by both the mru and
5076 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5077 * is less than arc_c_min.
5078 * (i.e. arc_c_min > amount of dirty data)
5080 * 2.1. arc_size is greater than or equal arc_c_min.
5081 * (i.e. arc_size >= arc_c_min > amount of dirty data)
5083 * In this case, not all clean data from the regular mru and mfu
5084 * lists is actually evictable; we must leave enough clean data
5085 * to keep arc_size above arc_c_min. Thus, the maximum amount of
5086 * evictable data from the two lists combined, is exactly the
5087 * difference between arc_size and arc_c_min.
5089 * 2.2. arc_size is less than arc_c_min
5090 * (i.e. arc_c_min > arc_size > amount of dirty data)
5092 * In this case, none of the data contained in the mru and mfu
5093 * lists is evictable, even if it's clean. Since arc_size is
5094 * already below arc_c_min, evicting any more would only
5095 * increase this negative difference.
5098 arc_evictable_memory(void)
5100 uint64_t arc_clean
=
5101 refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]) +
5102 refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]) +
5103 refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]) +
5104 refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
5105 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
5108 * Scale reported evictable memory in proportion to page cache, cap
5109 * at specified min/max.
5111 #ifdef ZFS_GLOBAL_NODE_PAGE_STATE
5112 uint64_t min
= (ptob(global_node_page_state(NR_FILE_PAGES
)) / 100) *
5115 uint64_t min
= (ptob(global_page_state(NR_FILE_PAGES
)) / 100) *
5118 min
= MAX(arc_c_min
, MIN(arc_c_max
, min
));
5120 if (arc_dirty
>= min
)
5123 return (MAX((int64_t)arc_size
- (int64_t)min
, 0));
5127 * If sc->nr_to_scan is zero, the caller is requesting a query of the
5128 * number of objects which can potentially be freed. If it is nonzero,
5129 * the request is to free that many objects.
5131 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
5132 * in struct shrinker and also require the shrinker to return the number
5135 * Older kernels require the shrinker to return the number of freeable
5136 * objects following the freeing of nr_to_free.
5138 static spl_shrinker_t
5139 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
5143 /* The arc is considered warm once reclaim has occurred */
5144 if (unlikely(arc_warm
== B_FALSE
))
5147 /* Return the potential number of reclaimable pages */
5148 pages
= btop((int64_t)arc_evictable_memory());
5149 if (sc
->nr_to_scan
== 0)
5152 /* Not allowed to perform filesystem reclaim */
5153 if (!(sc
->gfp_mask
& __GFP_FS
))
5154 return (SHRINK_STOP
);
5156 /* Reclaim in progress */
5157 if (mutex_tryenter(&arc_reclaim_lock
) == 0) {
5158 ARCSTAT_INCR(arcstat_need_free
, ptob(sc
->nr_to_scan
));
5162 mutex_exit(&arc_reclaim_lock
);
5165 * Evict the requested number of pages by shrinking arc_c the
5169 arc_shrink(ptob(sc
->nr_to_scan
));
5170 if (current_is_kswapd())
5171 arc_kmem_reap_now();
5172 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
5173 pages
= MAX((int64_t)pages
-
5174 (int64_t)btop(arc_evictable_memory()), 0);
5176 pages
= btop(arc_evictable_memory());
5179 * We've shrunk what we can, wake up threads.
5181 cv_broadcast(&arc_reclaim_waiters_cv
);
5183 pages
= SHRINK_STOP
;
5186 * When direct reclaim is observed it usually indicates a rapid
5187 * increase in memory pressure. This occurs because the kswapd
5188 * threads were unable to asynchronously keep enough free memory
5189 * available. In this case set arc_no_grow to briefly pause arc
5190 * growth to avoid compounding the memory pressure.
5192 if (current_is_kswapd()) {
5193 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
5195 arc_no_grow
= B_TRUE
;
5196 arc_kmem_reap_now();
5197 ARCSTAT_BUMP(arcstat_memory_direct_count
);
5202 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
5204 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
5205 #endif /* _KERNEL */
5208 * Adapt arc info given the number of bytes we are trying to add and
5209 * the state that we are coming from. This function is only called
5210 * when we are adding new content to the cache.
5213 arc_adapt(int bytes
, arc_state_t
*state
)
5216 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
5217 int64_t mrug_size
= refcount_count(&arc_mru_ghost
->arcs_size
);
5218 int64_t mfug_size
= refcount_count(&arc_mfu_ghost
->arcs_size
);
5220 if (state
== arc_l2c_only
)
5225 * Adapt the target size of the MRU list:
5226 * - if we just hit in the MRU ghost list, then increase
5227 * the target size of the MRU list.
5228 * - if we just hit in the MFU ghost list, then increase
5229 * the target size of the MFU list by decreasing the
5230 * target size of the MRU list.
5232 if (state
== arc_mru_ghost
) {
5233 mult
= (mrug_size
>= mfug_size
) ? 1 : (mfug_size
/ mrug_size
);
5234 if (!zfs_arc_p_dampener_disable
)
5235 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
5237 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
5238 } else if (state
== arc_mfu_ghost
) {
5241 mult
= (mfug_size
>= mrug_size
) ? 1 : (mrug_size
/ mfug_size
);
5242 if (!zfs_arc_p_dampener_disable
)
5243 mult
= MIN(mult
, 10);
5245 delta
= MIN(bytes
* mult
, arc_p
);
5246 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
5248 ASSERT((int64_t)arc_p
>= 0);
5250 if (arc_reclaim_needed()) {
5251 cv_signal(&arc_reclaim_thread_cv
);
5258 if (arc_c
>= arc_c_max
)
5262 * If we're within (2 * maxblocksize) bytes of the target
5263 * cache size, increment the target cache size
5265 ASSERT3U(arc_c
, >=, 2ULL << SPA_MAXBLOCKSHIFT
);
5266 if (arc_size
>= arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
5267 atomic_add_64(&arc_c
, (int64_t)bytes
);
5268 if (arc_c
> arc_c_max
)
5270 else if (state
== arc_anon
)
5271 atomic_add_64(&arc_p
, (int64_t)bytes
);
5275 ASSERT((int64_t)arc_p
>= 0);
5279 * Check if arc_size has grown past our upper threshold, determined by
5280 * zfs_arc_overflow_shift.
5283 arc_is_overflowing(void)
5285 /* Always allow at least one block of overflow */
5286 uint64_t overflow
= MAX(SPA_MAXBLOCKSIZE
,
5287 arc_c
>> zfs_arc_overflow_shift
);
5289 return (arc_size
>= arc_c
+ overflow
);
5293 arc_get_data_abd(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
5295 arc_buf_contents_t type
= arc_buf_type(hdr
);
5297 arc_get_data_impl(hdr
, size
, tag
);
5298 if (type
== ARC_BUFC_METADATA
) {
5299 return (abd_alloc(size
, B_TRUE
));
5301 ASSERT(type
== ARC_BUFC_DATA
);
5302 return (abd_alloc(size
, B_FALSE
));
5307 arc_get_data_buf(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
5309 arc_buf_contents_t type
= arc_buf_type(hdr
);
5311 arc_get_data_impl(hdr
, size
, tag
);
5312 if (type
== ARC_BUFC_METADATA
) {
5313 return (zio_buf_alloc(size
));
5315 ASSERT(type
== ARC_BUFC_DATA
);
5316 return (zio_data_buf_alloc(size
));
5321 * Allocate a block and return it to the caller. If we are hitting the
5322 * hard limit for the cache size, we must sleep, waiting for the eviction
5323 * thread to catch up. If we're past the target size but below the hard
5324 * limit, we'll only signal the reclaim thread and continue on.
5327 arc_get_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
5329 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
5330 arc_buf_contents_t type
= arc_buf_type(hdr
);
5332 arc_adapt(size
, state
);
5335 * If arc_size is currently overflowing, and has grown past our
5336 * upper limit, we must be adding data faster than the evict
5337 * thread can evict. Thus, to ensure we don't compound the
5338 * problem by adding more data and forcing arc_size to grow even
5339 * further past it's target size, we halt and wait for the
5340 * eviction thread to catch up.
5342 * It's also possible that the reclaim thread is unable to evict
5343 * enough buffers to get arc_size below the overflow limit (e.g.
5344 * due to buffers being un-evictable, or hash lock collisions).
5345 * In this case, we want to proceed regardless if we're
5346 * overflowing; thus we don't use a while loop here.
5348 if (arc_is_overflowing()) {
5349 mutex_enter(&arc_reclaim_lock
);
5352 * Now that we've acquired the lock, we may no longer be
5353 * over the overflow limit, lets check.
5355 * We're ignoring the case of spurious wake ups. If that
5356 * were to happen, it'd let this thread consume an ARC
5357 * buffer before it should have (i.e. before we're under
5358 * the overflow limit and were signalled by the reclaim
5359 * thread). As long as that is a rare occurrence, it
5360 * shouldn't cause any harm.
5362 if (arc_is_overflowing()) {
5363 cv_signal(&arc_reclaim_thread_cv
);
5364 cv_wait(&arc_reclaim_waiters_cv
, &arc_reclaim_lock
);
5367 mutex_exit(&arc_reclaim_lock
);
5370 VERIFY3U(hdr
->b_type
, ==, type
);
5371 if (type
== ARC_BUFC_METADATA
) {
5372 arc_space_consume(size
, ARC_SPACE_META
);
5374 arc_space_consume(size
, ARC_SPACE_DATA
);
5378 * Update the state size. Note that ghost states have a
5379 * "ghost size" and so don't need to be updated.
5381 if (!GHOST_STATE(state
)) {
5383 (void) refcount_add_many(&state
->arcs_size
, size
, tag
);
5386 * If this is reached via arc_read, the link is
5387 * protected by the hash lock. If reached via
5388 * arc_buf_alloc, the header should not be accessed by
5389 * any other thread. And, if reached via arc_read_done,
5390 * the hash lock will protect it if it's found in the
5391 * hash table; otherwise no other thread should be
5392 * trying to [add|remove]_reference it.
5394 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
5395 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5396 (void) refcount_add_many(&state
->arcs_esize
[type
],
5401 * If we are growing the cache, and we are adding anonymous
5402 * data, and we have outgrown arc_p, update arc_p
5404 if (arc_size
< arc_c
&& hdr
->b_l1hdr
.b_state
== arc_anon
&&
5405 (refcount_count(&arc_anon
->arcs_size
) +
5406 refcount_count(&arc_mru
->arcs_size
) > arc_p
))
5407 arc_p
= MIN(arc_c
, arc_p
+ size
);
5412 arc_free_data_abd(arc_buf_hdr_t
*hdr
, abd_t
*abd
, uint64_t size
, void *tag
)
5414 arc_free_data_impl(hdr
, size
, tag
);
5419 arc_free_data_buf(arc_buf_hdr_t
*hdr
, void *buf
, uint64_t size
, void *tag
)
5421 arc_buf_contents_t type
= arc_buf_type(hdr
);
5423 arc_free_data_impl(hdr
, size
, tag
);
5424 if (type
== ARC_BUFC_METADATA
) {
5425 zio_buf_free(buf
, size
);
5427 ASSERT(type
== ARC_BUFC_DATA
);
5428 zio_data_buf_free(buf
, size
);
5433 * Free the arc data buffer.
5436 arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
5438 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
5439 arc_buf_contents_t type
= arc_buf_type(hdr
);
5441 /* protected by hash lock, if in the hash table */
5442 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
5443 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5444 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
5446 (void) refcount_remove_many(&state
->arcs_esize
[type
],
5449 (void) refcount_remove_many(&state
->arcs_size
, size
, tag
);
5451 VERIFY3U(hdr
->b_type
, ==, type
);
5452 if (type
== ARC_BUFC_METADATA
) {
5453 arc_space_return(size
, ARC_SPACE_META
);
5455 ASSERT(type
== ARC_BUFC_DATA
);
5456 arc_space_return(size
, ARC_SPACE_DATA
);
5461 * This routine is called whenever a buffer is accessed.
5462 * NOTE: the hash lock is dropped in this function.
5465 arc_access(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
5469 ASSERT(MUTEX_HELD(hash_lock
));
5470 ASSERT(HDR_HAS_L1HDR(hdr
));
5472 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
5474 * This buffer is not in the cache, and does not
5475 * appear in our "ghost" list. Add the new buffer
5479 ASSERT0(hdr
->b_l1hdr
.b_arc_access
);
5480 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
5481 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
5482 arc_change_state(arc_mru
, hdr
, hash_lock
);
5484 } else if (hdr
->b_l1hdr
.b_state
== arc_mru
) {
5485 now
= ddi_get_lbolt();
5488 * If this buffer is here because of a prefetch, then either:
5489 * - clear the flag if this is a "referencing" read
5490 * (any subsequent access will bump this into the MFU state).
5492 * - move the buffer to the head of the list if this is
5493 * another prefetch (to make it less likely to be evicted).
5495 if (HDR_PREFETCH(hdr
)) {
5496 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
5497 /* link protected by hash lock */
5498 ASSERT(multilist_link_active(
5499 &hdr
->b_l1hdr
.b_arc_node
));
5501 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
5502 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
5503 ARCSTAT_BUMP(arcstat_mru_hits
);
5505 hdr
->b_l1hdr
.b_arc_access
= now
;
5510 * This buffer has been "accessed" only once so far,
5511 * but it is still in the cache. Move it to the MFU
5514 if (ddi_time_after(now
, hdr
->b_l1hdr
.b_arc_access
+
5517 * More than 125ms have passed since we
5518 * instantiated this buffer. Move it to the
5519 * most frequently used state.
5521 hdr
->b_l1hdr
.b_arc_access
= now
;
5522 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
5523 arc_change_state(arc_mfu
, hdr
, hash_lock
);
5525 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
5526 ARCSTAT_BUMP(arcstat_mru_hits
);
5527 } else if (hdr
->b_l1hdr
.b_state
== arc_mru_ghost
) {
5528 arc_state_t
*new_state
;
5530 * This buffer has been "accessed" recently, but
5531 * was evicted from the cache. Move it to the
5535 if (HDR_PREFETCH(hdr
)) {
5536 new_state
= arc_mru
;
5537 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0)
5538 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
5539 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
5541 new_state
= arc_mfu
;
5542 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
5545 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
5546 arc_change_state(new_state
, hdr
, hash_lock
);
5548 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_ghost_hits
);
5549 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
5550 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu
) {
5552 * This buffer has been accessed more than once and is
5553 * still in the cache. Keep it in the MFU state.
5555 * NOTE: an add_reference() that occurred when we did
5556 * the arc_read() will have kicked this off the list.
5557 * If it was a prefetch, we will explicitly move it to
5558 * the head of the list now.
5560 if ((HDR_PREFETCH(hdr
)) != 0) {
5561 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5562 /* link protected by hash_lock */
5563 ASSERT(multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
5565 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_hits
);
5566 ARCSTAT_BUMP(arcstat_mfu_hits
);
5567 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
5568 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu_ghost
) {
5569 arc_state_t
*new_state
= arc_mfu
;
5571 * This buffer has been accessed more than once but has
5572 * been evicted from the cache. Move it back to the
5576 if (HDR_PREFETCH(hdr
)) {
5578 * This is a prefetch access...
5579 * move this block back to the MRU state.
5581 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
5582 new_state
= arc_mru
;
5585 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
5586 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
5587 arc_change_state(new_state
, hdr
, hash_lock
);
5589 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_ghost_hits
);
5590 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
5591 } else if (hdr
->b_l1hdr
.b_state
== arc_l2c_only
) {
5593 * This buffer is on the 2nd Level ARC.
5596 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
5597 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
5598 arc_change_state(arc_mfu
, hdr
, hash_lock
);
5600 cmn_err(CE_PANIC
, "invalid arc state 0x%p",
5601 hdr
->b_l1hdr
.b_state
);
5605 /* a generic arc_read_done_func_t which you can use */
5608 arc_bcopy_func(zio_t
*zio
, int error
, arc_buf_t
*buf
, void *arg
)
5611 bcopy(buf
->b_data
, arg
, arc_buf_size(buf
));
5612 arc_buf_destroy(buf
, arg
);
5615 /* a generic arc_read_done_func_t */
5617 arc_getbuf_func(zio_t
*zio
, int error
, arc_buf_t
*buf
, void *arg
)
5619 arc_buf_t
**bufp
= arg
;
5621 arc_buf_destroy(buf
, arg
);
5625 ASSERT(buf
->b_data
);
5630 arc_hdr_verify(arc_buf_hdr_t
*hdr
, blkptr_t
*bp
)
5632 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
)) {
5633 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, 0);
5634 ASSERT3U(arc_hdr_get_compress(hdr
), ==, ZIO_COMPRESS_OFF
);
5636 if (HDR_COMPRESSION_ENABLED(hdr
)) {
5637 ASSERT3U(arc_hdr_get_compress(hdr
), ==,
5638 BP_GET_COMPRESS(bp
));
5640 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(bp
));
5641 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, BP_GET_PSIZE(bp
));
5642 ASSERT3U(!!HDR_PROTECTED(hdr
), ==, BP_IS_PROTECTED(bp
));
5647 arc_read_done(zio_t
*zio
)
5649 blkptr_t
*bp
= zio
->io_bp
;
5650 arc_buf_hdr_t
*hdr
= zio
->io_private
;
5651 kmutex_t
*hash_lock
= NULL
;
5652 arc_callback_t
*callback_list
;
5653 arc_callback_t
*acb
;
5654 boolean_t freeable
= B_FALSE
;
5655 boolean_t no_zio_error
= (zio
->io_error
== 0);
5658 * The hdr was inserted into hash-table and removed from lists
5659 * prior to starting I/O. We should find this header, since
5660 * it's in the hash table, and it should be legit since it's
5661 * not possible to evict it during the I/O. The only possible
5662 * reason for it not to be found is if we were freed during the
5665 if (HDR_IN_HASH_TABLE(hdr
)) {
5666 arc_buf_hdr_t
*found
;
5668 ASSERT3U(hdr
->b_birth
, ==, BP_PHYSICAL_BIRTH(zio
->io_bp
));
5669 ASSERT3U(hdr
->b_dva
.dva_word
[0], ==,
5670 BP_IDENTITY(zio
->io_bp
)->dva_word
[0]);
5671 ASSERT3U(hdr
->b_dva
.dva_word
[1], ==,
5672 BP_IDENTITY(zio
->io_bp
)->dva_word
[1]);
5674 found
= buf_hash_find(hdr
->b_spa
, zio
->io_bp
, &hash_lock
);
5676 ASSERT((found
== hdr
&&
5677 DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
5678 (found
== hdr
&& HDR_L2_READING(hdr
)));
5679 ASSERT3P(hash_lock
, !=, NULL
);
5682 if (BP_IS_PROTECTED(bp
)) {
5683 hdr
->b_crypt_hdr
.b_ot
= BP_GET_TYPE(bp
);
5684 hdr
->b_crypt_hdr
.b_dsobj
= zio
->io_bookmark
.zb_objset
;
5685 zio_crypt_decode_params_bp(bp
, hdr
->b_crypt_hdr
.b_salt
,
5686 hdr
->b_crypt_hdr
.b_iv
);
5688 if (BP_GET_TYPE(bp
) == DMU_OT_INTENT_LOG
) {
5691 tmpbuf
= abd_borrow_buf_copy(zio
->io_abd
,
5692 sizeof (zil_chain_t
));
5693 zio_crypt_decode_mac_zil(tmpbuf
,
5694 hdr
->b_crypt_hdr
.b_mac
);
5695 abd_return_buf(zio
->io_abd
, tmpbuf
,
5696 sizeof (zil_chain_t
));
5698 zio_crypt_decode_mac_bp(bp
, hdr
->b_crypt_hdr
.b_mac
);
5703 /* byteswap if necessary */
5704 if (BP_SHOULD_BYTESWAP(zio
->io_bp
)) {
5705 if (BP_GET_LEVEL(zio
->io_bp
) > 0) {
5706 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_UINT64
;
5708 hdr
->b_l1hdr
.b_byteswap
=
5709 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
5712 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
5716 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_EVICTED
);
5717 if (l2arc_noprefetch
&& HDR_PREFETCH(hdr
))
5718 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2CACHE
);
5720 callback_list
= hdr
->b_l1hdr
.b_acb
;
5721 ASSERT3P(callback_list
, !=, NULL
);
5723 if (hash_lock
&& no_zio_error
&& hdr
->b_l1hdr
.b_state
== arc_anon
) {
5725 * Only call arc_access on anonymous buffers. This is because
5726 * if we've issued an I/O for an evicted buffer, we've already
5727 * called arc_access (to prevent any simultaneous readers from
5728 * getting confused).
5730 arc_access(hdr
, hash_lock
);
5734 * If a read request has a callback (i.e. acb_done is not NULL), then we
5735 * make a buf containing the data according to the parameters which were
5736 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5737 * aren't needlessly decompressing the data multiple times.
5739 int callback_cnt
= 0;
5740 for (acb
= callback_list
; acb
!= NULL
; acb
= acb
->acb_next
) {
5744 /* This is a demand read since prefetches don't use callbacks */
5747 int error
= arc_buf_alloc_impl(hdr
, zio
->io_spa
,
5748 acb
->acb_dsobj
, acb
->acb_private
, acb
->acb_encrypted
,
5749 acb
->acb_compressed
, acb
->acb_noauth
, no_zio_error
,
5753 * Assert non-speculative zios didn't fail because an
5754 * encryption key wasn't loaded
5756 ASSERT((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) || error
== 0);
5759 * If we failed to decrypt, report an error now (as the zio
5760 * layer would have done if it had done the transforms).
5762 if (error
== ECKSUM
) {
5763 ASSERT(BP_IS_PROTECTED(bp
));
5764 error
= SET_ERROR(EIO
);
5765 spa_log_error(zio
->io_spa
, &zio
->io_bookmark
);
5766 if ((zio
->io_flags
& ZIO_FLAG_SPECULATIVE
) == 0) {
5767 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION
,
5768 zio
->io_spa
, NULL
, &zio
->io_bookmark
, zio
,
5774 zio
->io_error
= error
;
5777 hdr
->b_l1hdr
.b_acb
= NULL
;
5778 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5779 if (callback_cnt
== 0)
5780 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
|| HDR_HAS_RABD(hdr
));
5782 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
) ||
5783 callback_list
!= NULL
);
5786 arc_hdr_verify(hdr
, zio
->io_bp
);
5788 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_ERROR
);
5789 if (hdr
->b_l1hdr
.b_state
!= arc_anon
)
5790 arc_change_state(arc_anon
, hdr
, hash_lock
);
5791 if (HDR_IN_HASH_TABLE(hdr
))
5792 buf_hash_remove(hdr
);
5793 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
5797 * Broadcast before we drop the hash_lock to avoid the possibility
5798 * that the hdr (and hence the cv) might be freed before we get to
5799 * the cv_broadcast().
5801 cv_broadcast(&hdr
->b_l1hdr
.b_cv
);
5803 if (hash_lock
!= NULL
) {
5804 mutex_exit(hash_lock
);
5807 * This block was freed while we waited for the read to
5808 * complete. It has been removed from the hash table and
5809 * moved to the anonymous state (so that it won't show up
5812 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
5813 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
5816 /* execute each callback and free its structure */
5817 while ((acb
= callback_list
) != NULL
) {
5818 if (acb
->acb_done
) {
5819 acb
->acb_done(zio
, zio
->io_error
, acb
->acb_buf
,
5823 if (acb
->acb_zio_dummy
!= NULL
) {
5824 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
5825 zio_nowait(acb
->acb_zio_dummy
);
5828 callback_list
= acb
->acb_next
;
5829 kmem_free(acb
, sizeof (arc_callback_t
));
5833 arc_hdr_destroy(hdr
);
5837 * "Read" the block at the specified DVA (in bp) via the
5838 * cache. If the block is found in the cache, invoke the provided
5839 * callback immediately and return. Note that the `zio' parameter
5840 * in the callback will be NULL in this case, since no IO was
5841 * required. If the block is not in the cache pass the read request
5842 * on to the spa with a substitute callback function, so that the
5843 * requested block will be added to the cache.
5845 * If a read request arrives for a block that has a read in-progress,
5846 * either wait for the in-progress read to complete (and return the
5847 * results); or, if this is a read with a "done" func, add a record
5848 * to the read to invoke the "done" func when the read completes,
5849 * and return; or just return.
5851 * arc_read_done() will invoke all the requested "done" functions
5852 * for readers of this block.
5855 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
5856 arc_read_done_func_t
*done
, void *private, zio_priority_t priority
,
5857 int zio_flags
, arc_flags_t
*arc_flags
, const zbookmark_phys_t
*zb
)
5859 arc_buf_hdr_t
*hdr
= NULL
;
5860 kmutex_t
*hash_lock
= NULL
;
5862 uint64_t guid
= spa_load_guid(spa
);
5863 boolean_t compressed_read
= (zio_flags
& ZIO_FLAG_RAW_COMPRESS
) != 0;
5864 boolean_t encrypted_read
= BP_IS_ENCRYPTED(bp
) &&
5865 (zio_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0;
5866 boolean_t noauth_read
= BP_IS_AUTHENTICATED(bp
) &&
5867 (zio_flags
& ZIO_FLAG_RAW_ENCRYPT
) != 0;
5870 ASSERT(!BP_IS_EMBEDDED(bp
) ||
5871 BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
);
5874 if (!BP_IS_EMBEDDED(bp
)) {
5876 * Embedded BP's have no DVA and require no I/O to "read".
5877 * Create an anonymous arc buf to back it.
5879 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
5883 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5884 * we maintain encrypted data seperately from compressed / uncompressed
5885 * data. If the user is requesting raw encrypted data and we don't have
5886 * that in the header we will read from disk to guarantee that we can
5887 * get it even if the encryption keys aren't loaded.
5889 if (hdr
!= NULL
&& HDR_HAS_L1HDR(hdr
) && (HDR_HAS_RABD(hdr
) ||
5890 (hdr
->b_l1hdr
.b_pabd
!= NULL
&& !encrypted_read
))) {
5891 arc_buf_t
*buf
= NULL
;
5892 *arc_flags
|= ARC_FLAG_CACHED
;
5894 if (HDR_IO_IN_PROGRESS(hdr
)) {
5896 if ((hdr
->b_flags
& ARC_FLAG_PRIO_ASYNC_READ
) &&
5897 priority
== ZIO_PRIORITY_SYNC_READ
) {
5899 * This sync read must wait for an
5900 * in-progress async read (e.g. a predictive
5901 * prefetch). Async reads are queued
5902 * separately at the vdev_queue layer, so
5903 * this is a form of priority inversion.
5904 * Ideally, we would "inherit" the demand
5905 * i/o's priority by moving the i/o from
5906 * the async queue to the synchronous queue,
5907 * but there is currently no mechanism to do
5908 * so. Track this so that we can evaluate
5909 * the magnitude of this potential performance
5912 * Note that if the prefetch i/o is already
5913 * active (has been issued to the device),
5914 * the prefetch improved performance, because
5915 * we issued it sooner than we would have
5916 * without the prefetch.
5918 DTRACE_PROBE1(arc__sync__wait__for__async
,
5919 arc_buf_hdr_t
*, hdr
);
5920 ARCSTAT_BUMP(arcstat_sync_wait_for_async
);
5922 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5923 arc_hdr_clear_flags(hdr
,
5924 ARC_FLAG_PREDICTIVE_PREFETCH
);
5927 if (*arc_flags
& ARC_FLAG_WAIT
) {
5928 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
5929 mutex_exit(hash_lock
);
5932 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5935 arc_callback_t
*acb
= NULL
;
5937 acb
= kmem_zalloc(sizeof (arc_callback_t
),
5939 acb
->acb_done
= done
;
5940 acb
->acb_private
= private;
5941 acb
->acb_compressed
= compressed_read
;
5942 acb
->acb_encrypted
= encrypted_read
;
5943 acb
->acb_noauth
= noauth_read
;
5944 acb
->acb_dsobj
= zb
->zb_objset
;
5946 acb
->acb_zio_dummy
= zio_null(pio
,
5947 spa
, NULL
, NULL
, NULL
, zio_flags
);
5949 ASSERT3P(acb
->acb_done
, !=, NULL
);
5950 acb
->acb_next
= hdr
->b_l1hdr
.b_acb
;
5951 hdr
->b_l1hdr
.b_acb
= acb
;
5952 mutex_exit(hash_lock
);
5955 mutex_exit(hash_lock
);
5959 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
5960 hdr
->b_l1hdr
.b_state
== arc_mfu
);
5963 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5965 * This is a demand read which does not have to
5966 * wait for i/o because we did a predictive
5967 * prefetch i/o for it, which has completed.
5970 arc__demand__hit__predictive__prefetch
,
5971 arc_buf_hdr_t
*, hdr
);
5973 arcstat_demand_hit_predictive_prefetch
);
5974 arc_hdr_clear_flags(hdr
,
5975 ARC_FLAG_PREDICTIVE_PREFETCH
);
5977 ASSERT(!BP_IS_EMBEDDED(bp
) || !BP_IS_HOLE(bp
));
5979 /* Get a buf with the desired data in it. */
5980 rc
= arc_buf_alloc_impl(hdr
, spa
, zb
->zb_objset
,
5981 private, encrypted_read
, compressed_read
,
5982 noauth_read
, B_TRUE
, &buf
);
5983 ASSERT((zio_flags
& ZIO_FLAG_SPECULATIVE
) || rc
== 0);
5984 } else if (*arc_flags
& ARC_FLAG_PREFETCH
&&
5985 refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
5986 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
5988 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
5989 arc_access(hdr
, hash_lock
);
5990 if (*arc_flags
& ARC_FLAG_L2CACHE
)
5991 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5992 mutex_exit(hash_lock
);
5993 ARCSTAT_BUMP(arcstat_hits
);
5994 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
5995 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
5996 data
, metadata
, hits
);
5999 done(NULL
, rc
, buf
, private);
6001 uint64_t lsize
= BP_GET_LSIZE(bp
);
6002 uint64_t psize
= BP_GET_PSIZE(bp
);
6003 arc_callback_t
*acb
;
6006 boolean_t devw
= B_FALSE
;
6011 * Gracefully handle a damaged logical block size as a
6014 if (lsize
> spa_maxblocksize(spa
)) {
6015 rc
= SET_ERROR(ECKSUM
);
6020 /* this block is not in the cache */
6021 arc_buf_hdr_t
*exists
= NULL
;
6022 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
6023 hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
6024 BP_IS_PROTECTED(bp
), BP_GET_COMPRESS(bp
), type
,
6027 if (!BP_IS_EMBEDDED(bp
)) {
6028 hdr
->b_dva
= *BP_IDENTITY(bp
);
6029 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
6030 exists
= buf_hash_insert(hdr
, &hash_lock
);
6032 if (exists
!= NULL
) {
6033 /* somebody beat us to the hash insert */
6034 mutex_exit(hash_lock
);
6035 buf_discard_identity(hdr
);
6036 arc_hdr_destroy(hdr
);
6037 goto top
; /* restart the IO request */
6041 * This block is in the ghost cache or encrypted data
6042 * was requested and we didn't have it. If it was
6043 * L2-only (and thus didn't have an L1 hdr),
6044 * we realloc the header to add an L1 hdr.
6046 if (!HDR_HAS_L1HDR(hdr
)) {
6047 hdr
= arc_hdr_realloc(hdr
, hdr_l2only_cache
,
6051 if (GHOST_STATE(hdr
->b_l1hdr
.b_state
)) {
6052 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
6053 ASSERT(!HDR_HAS_RABD(hdr
));
6054 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
6055 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
6056 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
6057 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
6058 } else if (HDR_IO_IN_PROGRESS(hdr
)) {
6060 * If this header already had an IO in progress
6061 * and we are performing another IO to fetch
6062 * encrypted data we must wait until the first
6063 * IO completes so as not to confuse
6064 * arc_read_done(). This should be very rare
6065 * and so the performance impact shouldn't
6068 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
6069 mutex_exit(hash_lock
);
6074 * This is a delicate dance that we play here.
6075 * This hdr might be in the ghost list so we access
6076 * it to move it out of the ghost list before we
6077 * initiate the read. If it's a prefetch then
6078 * it won't have a callback so we'll remove the
6079 * reference that arc_buf_alloc_impl() created. We
6080 * do this after we've called arc_access() to
6081 * avoid hitting an assert in remove_reference().
6083 arc_access(hdr
, hash_lock
);
6084 arc_hdr_alloc_abd(hdr
, encrypted_read
);
6087 if (encrypted_read
) {
6088 ASSERT(HDR_HAS_RABD(hdr
));
6089 size
= HDR_GET_PSIZE(hdr
);
6090 hdr_abd
= hdr
->b_crypt_hdr
.b_rabd
;
6091 zio_flags
|= ZIO_FLAG_RAW
;
6093 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
6094 size
= arc_hdr_size(hdr
);
6095 hdr_abd
= hdr
->b_l1hdr
.b_pabd
;
6097 if (arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
) {
6098 zio_flags
|= ZIO_FLAG_RAW_COMPRESS
;
6102 * For authenticated bp's, we do not ask the ZIO layer
6103 * to authenticate them since this will cause the entire
6104 * IO to fail if the key isn't loaded. Instead, we
6105 * defer authentication until arc_buf_fill(), which will
6106 * verify the data when the key is available.
6108 if (BP_IS_AUTHENTICATED(bp
))
6109 zio_flags
|= ZIO_FLAG_RAW_ENCRYPT
;
6112 if (*arc_flags
& ARC_FLAG_PREFETCH
&&
6113 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
))
6114 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
6115 if (*arc_flags
& ARC_FLAG_L2CACHE
)
6116 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
6117 if (BP_IS_AUTHENTICATED(bp
))
6118 arc_hdr_set_flags(hdr
, ARC_FLAG_NOAUTH
);
6119 if (BP_GET_LEVEL(bp
) > 0)
6120 arc_hdr_set_flags(hdr
, ARC_FLAG_INDIRECT
);
6121 if (*arc_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
)
6122 arc_hdr_set_flags(hdr
, ARC_FLAG_PREDICTIVE_PREFETCH
);
6123 ASSERT(!GHOST_STATE(hdr
->b_l1hdr
.b_state
));
6125 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
6126 acb
->acb_done
= done
;
6127 acb
->acb_private
= private;
6128 acb
->acb_compressed
= compressed_read
;
6129 acb
->acb_encrypted
= encrypted_read
;
6130 acb
->acb_noauth
= noauth_read
;
6131 acb
->acb_dsobj
= zb
->zb_objset
;
6133 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
6134 hdr
->b_l1hdr
.b_acb
= acb
;
6135 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6137 if (HDR_HAS_L2HDR(hdr
) &&
6138 (vd
= hdr
->b_l2hdr
.b_dev
->l2ad_vdev
) != NULL
) {
6139 devw
= hdr
->b_l2hdr
.b_dev
->l2ad_writing
;
6140 addr
= hdr
->b_l2hdr
.b_daddr
;
6142 * Lock out device removal.
6144 if (vdev_is_dead(vd
) ||
6145 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
6149 if (priority
== ZIO_PRIORITY_ASYNC_READ
)
6150 arc_hdr_set_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
6152 arc_hdr_clear_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
6154 if (hash_lock
!= NULL
)
6155 mutex_exit(hash_lock
);
6158 * At this point, we have a level 1 cache miss. Try again in
6159 * L2ARC if possible.
6161 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, lsize
);
6163 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
6164 uint64_t, lsize
, zbookmark_phys_t
*, zb
);
6165 ARCSTAT_BUMP(arcstat_misses
);
6166 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
6167 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
6168 data
, metadata
, misses
);
6170 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
6172 * Read from the L2ARC if the following are true:
6173 * 1. The L2ARC vdev was previously cached.
6174 * 2. This buffer still has L2ARC metadata.
6175 * 3. This buffer isn't currently writing to the L2ARC.
6176 * 4. The L2ARC entry wasn't evicted, which may
6177 * also have invalidated the vdev.
6178 * 5. This isn't prefetch and l2arc_noprefetch is set.
6180 if (HDR_HAS_L2HDR(hdr
) &&
6181 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
6182 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
6183 l2arc_read_callback_t
*cb
;
6187 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
6188 ARCSTAT_BUMP(arcstat_l2_hits
);
6189 atomic_inc_32(&hdr
->b_l2hdr
.b_hits
);
6191 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
6193 cb
->l2rcb_hdr
= hdr
;
6196 cb
->l2rcb_flags
= zio_flags
;
6198 asize
= vdev_psize_to_asize(vd
, size
);
6199 if (asize
!= size
) {
6200 abd
= abd_alloc_for_io(asize
,
6201 HDR_ISTYPE_METADATA(hdr
));
6202 cb
->l2rcb_abd
= abd
;
6207 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
6208 addr
+ asize
<= vd
->vdev_psize
-
6209 VDEV_LABEL_END_SIZE
);
6212 * l2arc read. The SCL_L2ARC lock will be
6213 * released by l2arc_read_done().
6214 * Issue a null zio if the underlying buffer
6215 * was squashed to zero size by compression.
6217 ASSERT3U(arc_hdr_get_compress(hdr
), !=,
6218 ZIO_COMPRESS_EMPTY
);
6219 rzio
= zio_read_phys(pio
, vd
, addr
,
6222 l2arc_read_done
, cb
, priority
,
6223 zio_flags
| ZIO_FLAG_DONT_CACHE
|
6225 ZIO_FLAG_DONT_PROPAGATE
|
6226 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
6228 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
6230 ARCSTAT_INCR(arcstat_l2_read_bytes
,
6231 HDR_GET_PSIZE(hdr
));
6233 if (*arc_flags
& ARC_FLAG_NOWAIT
) {
6238 ASSERT(*arc_flags
& ARC_FLAG_WAIT
);
6239 if (zio_wait(rzio
) == 0)
6242 /* l2arc read error; goto zio_read() */
6244 DTRACE_PROBE1(l2arc__miss
,
6245 arc_buf_hdr_t
*, hdr
);
6246 ARCSTAT_BUMP(arcstat_l2_misses
);
6247 if (HDR_L2_WRITING(hdr
))
6248 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
6249 spa_config_exit(spa
, SCL_L2ARC
, vd
);
6253 spa_config_exit(spa
, SCL_L2ARC
, vd
);
6254 if (l2arc_ndev
!= 0) {
6255 DTRACE_PROBE1(l2arc__miss
,
6256 arc_buf_hdr_t
*, hdr
);
6257 ARCSTAT_BUMP(arcstat_l2_misses
);
6261 rzio
= zio_read(pio
, spa
, bp
, hdr_abd
, size
,
6262 arc_read_done
, hdr
, priority
, zio_flags
, zb
);
6264 if (*arc_flags
& ARC_FLAG_WAIT
) {
6265 rc
= zio_wait(rzio
);
6269 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
6274 spa_read_history_add(spa
, zb
, *arc_flags
);
6279 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
6283 p
= kmem_alloc(sizeof (*p
), KM_SLEEP
);
6285 p
->p_private
= private;
6286 list_link_init(&p
->p_node
);
6287 refcount_create(&p
->p_refcnt
);
6289 mutex_enter(&arc_prune_mtx
);
6290 refcount_add(&p
->p_refcnt
, &arc_prune_list
);
6291 list_insert_head(&arc_prune_list
, p
);
6292 mutex_exit(&arc_prune_mtx
);
6298 arc_remove_prune_callback(arc_prune_t
*p
)
6300 boolean_t wait
= B_FALSE
;
6301 mutex_enter(&arc_prune_mtx
);
6302 list_remove(&arc_prune_list
, p
);
6303 if (refcount_remove(&p
->p_refcnt
, &arc_prune_list
) > 0)
6305 mutex_exit(&arc_prune_mtx
);
6307 /* wait for arc_prune_task to finish */
6309 taskq_wait_outstanding(arc_prune_taskq
, 0);
6310 ASSERT0(refcount_count(&p
->p_refcnt
));
6311 refcount_destroy(&p
->p_refcnt
);
6312 kmem_free(p
, sizeof (*p
));
6316 * Notify the arc that a block was freed, and thus will never be used again.
6319 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
6322 kmutex_t
*hash_lock
;
6323 uint64_t guid
= spa_load_guid(spa
);
6325 ASSERT(!BP_IS_EMBEDDED(bp
));
6327 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
6332 * We might be trying to free a block that is still doing I/O
6333 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6334 * dmu_sync-ed block). If this block is being prefetched, then it
6335 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6336 * until the I/O completes. A block may also have a reference if it is
6337 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6338 * have written the new block to its final resting place on disk but
6339 * without the dedup flag set. This would have left the hdr in the MRU
6340 * state and discoverable. When the txg finally syncs it detects that
6341 * the block was overridden in open context and issues an override I/O.
6342 * Since this is a dedup block, the override I/O will determine if the
6343 * block is already in the DDT. If so, then it will replace the io_bp
6344 * with the bp from the DDT and allow the I/O to finish. When the I/O
6345 * reaches the done callback, dbuf_write_override_done, it will
6346 * check to see if the io_bp and io_bp_override are identical.
6347 * If they are not, then it indicates that the bp was replaced with
6348 * the bp in the DDT and the override bp is freed. This allows
6349 * us to arrive here with a reference on a block that is being
6350 * freed. So if we have an I/O in progress, or a reference to
6351 * this hdr, then we don't destroy the hdr.
6353 if (!HDR_HAS_L1HDR(hdr
) || (!HDR_IO_IN_PROGRESS(hdr
) &&
6354 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
))) {
6355 arc_change_state(arc_anon
, hdr
, hash_lock
);
6356 arc_hdr_destroy(hdr
);
6357 mutex_exit(hash_lock
);
6359 mutex_exit(hash_lock
);
6365 * Release this buffer from the cache, making it an anonymous buffer. This
6366 * must be done after a read and prior to modifying the buffer contents.
6367 * If the buffer has more than one reference, we must make
6368 * a new hdr for the buffer.
6371 arc_release(arc_buf_t
*buf
, void *tag
)
6373 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
6376 * It would be nice to assert that if its DMU metadata (level >
6377 * 0 || it's the dnode file), then it must be syncing context.
6378 * But we don't know that information at this level.
6381 mutex_enter(&buf
->b_evict_lock
);
6383 ASSERT(HDR_HAS_L1HDR(hdr
));
6386 * We don't grab the hash lock prior to this check, because if
6387 * the buffer's header is in the arc_anon state, it won't be
6388 * linked into the hash table.
6390 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
6391 mutex_exit(&buf
->b_evict_lock
);
6392 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
6393 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
6394 ASSERT(!HDR_HAS_L2HDR(hdr
));
6395 ASSERT(HDR_EMPTY(hdr
));
6397 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
6398 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), ==, 1);
6399 ASSERT(!list_link_active(&hdr
->b_l1hdr
.b_arc_node
));
6401 hdr
->b_l1hdr
.b_arc_access
= 0;
6404 * If the buf is being overridden then it may already
6405 * have a hdr that is not empty.
6407 buf_discard_identity(hdr
);
6413 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
6414 mutex_enter(hash_lock
);
6417 * This assignment is only valid as long as the hash_lock is
6418 * held, we must be careful not to reference state or the
6419 * b_state field after dropping the lock.
6421 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
6422 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
6423 ASSERT3P(state
, !=, arc_anon
);
6425 /* this buffer is not on any list */
6426 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), >, 0);
6428 if (HDR_HAS_L2HDR(hdr
)) {
6429 mutex_enter(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
6432 * We have to recheck this conditional again now that
6433 * we're holding the l2ad_mtx to prevent a race with
6434 * another thread which might be concurrently calling
6435 * l2arc_evict(). In that case, l2arc_evict() might have
6436 * destroyed the header's L2 portion as we were waiting
6437 * to acquire the l2ad_mtx.
6439 if (HDR_HAS_L2HDR(hdr
))
6440 arc_hdr_l2hdr_destroy(hdr
);
6442 mutex_exit(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
6446 * Do we have more than one buf?
6448 if (hdr
->b_l1hdr
.b_bufcnt
> 1) {
6449 arc_buf_hdr_t
*nhdr
;
6450 uint64_t spa
= hdr
->b_spa
;
6451 uint64_t psize
= HDR_GET_PSIZE(hdr
);
6452 uint64_t lsize
= HDR_GET_LSIZE(hdr
);
6453 boolean_t
protected = HDR_PROTECTED(hdr
);
6454 enum zio_compress compress
= arc_hdr_get_compress(hdr
);
6455 arc_buf_contents_t type
= arc_buf_type(hdr
);
6456 VERIFY3U(hdr
->b_type
, ==, type
);
6458 ASSERT(hdr
->b_l1hdr
.b_buf
!= buf
|| buf
->b_next
!= NULL
);
6459 (void) remove_reference(hdr
, hash_lock
, tag
);
6461 if (arc_buf_is_shared(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
6462 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
6463 ASSERT(ARC_BUF_LAST(buf
));
6467 * Pull the data off of this hdr and attach it to
6468 * a new anonymous hdr. Also find the last buffer
6469 * in the hdr's buffer list.
6471 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
6472 ASSERT3P(lastbuf
, !=, NULL
);
6475 * If the current arc_buf_t and the hdr are sharing their data
6476 * buffer, then we must stop sharing that block.
6478 if (arc_buf_is_shared(buf
)) {
6479 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
6480 VERIFY(!arc_buf_is_shared(lastbuf
));
6483 * First, sever the block sharing relationship between
6484 * buf and the arc_buf_hdr_t.
6486 arc_unshare_buf(hdr
, buf
);
6489 * Now we need to recreate the hdr's b_pabd. Since we
6490 * have lastbuf handy, we try to share with it, but if
6491 * we can't then we allocate a new b_pabd and copy the
6492 * data from buf into it.
6494 if (arc_can_share(hdr
, lastbuf
)) {
6495 arc_share_buf(hdr
, lastbuf
);
6497 arc_hdr_alloc_abd(hdr
, B_FALSE
);
6498 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
,
6499 buf
->b_data
, psize
);
6501 VERIFY3P(lastbuf
->b_data
, !=, NULL
);
6502 } else if (HDR_SHARED_DATA(hdr
)) {
6504 * Uncompressed shared buffers are always at the end
6505 * of the list. Compressed buffers don't have the
6506 * same requirements. This makes it hard to
6507 * simply assert that the lastbuf is shared so
6508 * we rely on the hdr's compression flags to determine
6509 * if we have a compressed, shared buffer.
6511 ASSERT(arc_buf_is_shared(lastbuf
) ||
6512 arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
);
6513 ASSERT(!ARC_BUF_SHARED(buf
));
6516 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
|| HDR_HAS_RABD(hdr
));
6517 ASSERT3P(state
, !=, arc_l2c_only
);
6519 (void) refcount_remove_many(&state
->arcs_size
,
6520 arc_buf_size(buf
), buf
);
6522 if (refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
)) {
6523 ASSERT3P(state
, !=, arc_l2c_only
);
6524 (void) refcount_remove_many(&state
->arcs_esize
[type
],
6525 arc_buf_size(buf
), buf
);
6528 hdr
->b_l1hdr
.b_bufcnt
-= 1;
6529 if (ARC_BUF_ENCRYPTED(buf
))
6530 hdr
->b_crypt_hdr
.b_ebufcnt
-= 1;
6532 arc_cksum_verify(buf
);
6533 arc_buf_unwatch(buf
);
6535 /* if this is the last uncompressed buf free the checksum */
6536 if (!arc_hdr_has_uncompressed_buf(hdr
))
6537 arc_cksum_free(hdr
);
6539 mutex_exit(hash_lock
);
6542 * Allocate a new hdr. The new hdr will contain a b_pabd
6543 * buffer which will be freed in arc_write().
6545 nhdr
= arc_hdr_alloc(spa
, psize
, lsize
, protected,
6546 compress
, type
, HDR_HAS_RABD(hdr
));
6547 ASSERT3P(nhdr
->b_l1hdr
.b_buf
, ==, NULL
);
6548 ASSERT0(nhdr
->b_l1hdr
.b_bufcnt
);
6549 ASSERT0(refcount_count(&nhdr
->b_l1hdr
.b_refcnt
));
6550 VERIFY3U(nhdr
->b_type
, ==, type
);
6551 ASSERT(!HDR_SHARED_DATA(nhdr
));
6553 nhdr
->b_l1hdr
.b_buf
= buf
;
6554 nhdr
->b_l1hdr
.b_bufcnt
= 1;
6555 if (ARC_BUF_ENCRYPTED(buf
))
6556 nhdr
->b_crypt_hdr
.b_ebufcnt
= 1;
6557 nhdr
->b_l1hdr
.b_mru_hits
= 0;
6558 nhdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
6559 nhdr
->b_l1hdr
.b_mfu_hits
= 0;
6560 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
6561 nhdr
->b_l1hdr
.b_l2_hits
= 0;
6562 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, tag
);
6565 mutex_exit(&buf
->b_evict_lock
);
6566 (void) refcount_add_many(&arc_anon
->arcs_size
,
6567 HDR_GET_LSIZE(nhdr
), buf
);
6569 mutex_exit(&buf
->b_evict_lock
);
6570 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 1);
6571 /* protected by hash lock, or hdr is on arc_anon */
6572 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
6573 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
6574 hdr
->b_l1hdr
.b_mru_hits
= 0;
6575 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
6576 hdr
->b_l1hdr
.b_mfu_hits
= 0;
6577 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
6578 hdr
->b_l1hdr
.b_l2_hits
= 0;
6579 arc_change_state(arc_anon
, hdr
, hash_lock
);
6580 hdr
->b_l1hdr
.b_arc_access
= 0;
6582 mutex_exit(hash_lock
);
6583 buf_discard_identity(hdr
);
6589 arc_released(arc_buf_t
*buf
)
6593 mutex_enter(&buf
->b_evict_lock
);
6594 released
= (buf
->b_data
!= NULL
&&
6595 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
6596 mutex_exit(&buf
->b_evict_lock
);
6602 arc_referenced(arc_buf_t
*buf
)
6606 mutex_enter(&buf
->b_evict_lock
);
6607 referenced
= (refcount_count(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
6608 mutex_exit(&buf
->b_evict_lock
);
6609 return (referenced
);
6614 arc_write_ready(zio_t
*zio
)
6616 arc_write_callback_t
*callback
= zio
->io_private
;
6617 arc_buf_t
*buf
= callback
->awcb_buf
;
6618 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
6619 blkptr_t
*bp
= zio
->io_bp
;
6620 uint64_t psize
= BP_IS_HOLE(bp
) ? 0 : BP_GET_PSIZE(bp
);
6621 fstrans_cookie_t cookie
= spl_fstrans_mark();
6623 ASSERT(HDR_HAS_L1HDR(hdr
));
6624 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
6625 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
6628 * If we're reexecuting this zio because the pool suspended, then
6629 * cleanup any state that was previously set the first time the
6630 * callback was invoked.
6632 if (zio
->io_flags
& ZIO_FLAG_REEXECUTED
) {
6633 arc_cksum_free(hdr
);
6634 arc_buf_unwatch(buf
);
6635 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
6636 if (arc_buf_is_shared(buf
)) {
6637 arc_unshare_buf(hdr
, buf
);
6639 arc_hdr_free_abd(hdr
, B_FALSE
);
6643 if (HDR_HAS_RABD(hdr
))
6644 arc_hdr_free_abd(hdr
, B_TRUE
);
6646 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
6647 ASSERT(!HDR_HAS_RABD(hdr
));
6648 ASSERT(!HDR_SHARED_DATA(hdr
));
6649 ASSERT(!arc_buf_is_shared(buf
));
6651 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
6653 if (HDR_IO_IN_PROGRESS(hdr
))
6654 ASSERT(zio
->io_flags
& ZIO_FLAG_REEXECUTED
);
6656 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6658 if (BP_IS_PROTECTED(bp
) != !!HDR_PROTECTED(hdr
))
6659 hdr
= arc_hdr_realloc_crypt(hdr
, BP_IS_PROTECTED(bp
));
6661 if (BP_IS_PROTECTED(bp
)) {
6662 /* ZIL blocks are written through zio_rewrite */
6663 ASSERT3U(BP_GET_TYPE(bp
), !=, DMU_OT_INTENT_LOG
);
6664 ASSERT(HDR_PROTECTED(hdr
));
6666 hdr
->b_crypt_hdr
.b_ot
= BP_GET_TYPE(bp
);
6667 hdr
->b_crypt_hdr
.b_dsobj
= zio
->io_bookmark
.zb_objset
;
6668 zio_crypt_decode_params_bp(bp
, hdr
->b_crypt_hdr
.b_salt
,
6669 hdr
->b_crypt_hdr
.b_iv
);
6670 zio_crypt_decode_mac_bp(bp
, hdr
->b_crypt_hdr
.b_mac
);
6674 * If this block was written for raw encryption but the zio layer
6675 * ended up only authenticating it, adjust the buffer flags now.
6677 if (BP_IS_AUTHENTICATED(bp
) && ARC_BUF_ENCRYPTED(buf
)) {
6678 arc_hdr_set_flags(hdr
, ARC_FLAG_NOAUTH
);
6679 buf
->b_flags
&= ~ARC_BUF_FLAG_ENCRYPTED
;
6680 if (BP_GET_COMPRESS(bp
) == ZIO_COMPRESS_OFF
)
6681 buf
->b_flags
&= ~ARC_BUF_FLAG_COMPRESSED
;
6684 /* this must be done after the buffer flags are adjusted */
6685 arc_cksum_compute(buf
);
6687 enum zio_compress compress
;
6688 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
)) {
6689 compress
= ZIO_COMPRESS_OFF
;
6691 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(bp
));
6692 compress
= BP_GET_COMPRESS(bp
);
6694 HDR_SET_PSIZE(hdr
, psize
);
6695 arc_hdr_set_compress(hdr
, compress
);
6697 if (zio
->io_error
!= 0 || psize
== 0)
6701 * Fill the hdr with data. If the buffer is encrypted we have no choice
6702 * but to copy the data into b_radb. If the hdr is compressed, the data
6703 * we want is available from the zio, otherwise we can take it from
6706 * We might be able to share the buf's data with the hdr here. However,
6707 * doing so would cause the ARC to be full of linear ABDs if we write a
6708 * lot of shareable data. As a compromise, we check whether scattered
6709 * ABDs are allowed, and assume that if they are then the user wants
6710 * the ARC to be primarily filled with them regardless of the data being
6711 * written. Therefore, if they're allowed then we allocate one and copy
6712 * the data into it; otherwise, we share the data directly if we can.
6714 if (ARC_BUF_ENCRYPTED(buf
)) {
6715 ASSERT3U(psize
, >, 0);
6716 ASSERT(ARC_BUF_COMPRESSED(buf
));
6717 arc_hdr_alloc_abd(hdr
, B_TRUE
);
6718 abd_copy(hdr
->b_crypt_hdr
.b_rabd
, zio
->io_abd
, psize
);
6719 } else if (zfs_abd_scatter_enabled
|| !arc_can_share(hdr
, buf
)) {
6721 * Ideally, we would always copy the io_abd into b_pabd, but the
6722 * user may have disabled compressed ARC, thus we must check the
6723 * hdr's compression setting rather than the io_bp's.
6725 if (BP_IS_ENCRYPTED(bp
)) {
6726 ASSERT3U(psize
, >, 0);
6727 arc_hdr_alloc_abd(hdr
, B_TRUE
);
6728 abd_copy(hdr
->b_crypt_hdr
.b_rabd
, zio
->io_abd
, psize
);
6729 } else if (arc_hdr_get_compress(hdr
) != ZIO_COMPRESS_OFF
&&
6730 !ARC_BUF_COMPRESSED(buf
)) {
6731 ASSERT3U(psize
, >, 0);
6732 arc_hdr_alloc_abd(hdr
, B_FALSE
);
6733 abd_copy(hdr
->b_l1hdr
.b_pabd
, zio
->io_abd
, psize
);
6735 ASSERT3U(zio
->io_orig_size
, ==, arc_hdr_size(hdr
));
6736 arc_hdr_alloc_abd(hdr
, B_FALSE
);
6737 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
6741 ASSERT3P(buf
->b_data
, ==, abd_to_buf(zio
->io_orig_abd
));
6742 ASSERT3U(zio
->io_orig_size
, ==, arc_buf_size(buf
));
6743 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
6745 arc_share_buf(hdr
, buf
);
6749 arc_hdr_verify(hdr
, bp
);
6750 spl_fstrans_unmark(cookie
);
6754 arc_write_children_ready(zio_t
*zio
)
6756 arc_write_callback_t
*callback
= zio
->io_private
;
6757 arc_buf_t
*buf
= callback
->awcb_buf
;
6759 callback
->awcb_children_ready(zio
, buf
, callback
->awcb_private
);
6763 * The SPA calls this callback for each physical write that happens on behalf
6764 * of a logical write. See the comment in dbuf_write_physdone() for details.
6767 arc_write_physdone(zio_t
*zio
)
6769 arc_write_callback_t
*cb
= zio
->io_private
;
6770 if (cb
->awcb_physdone
!= NULL
)
6771 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
6775 arc_write_done(zio_t
*zio
)
6777 arc_write_callback_t
*callback
= zio
->io_private
;
6778 arc_buf_t
*buf
= callback
->awcb_buf
;
6779 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
6781 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
6783 if (zio
->io_error
== 0) {
6784 arc_hdr_verify(hdr
, zio
->io_bp
);
6786 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
6787 buf_discard_identity(hdr
);
6789 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
6790 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
6793 ASSERT(HDR_EMPTY(hdr
));
6797 * If the block to be written was all-zero or compressed enough to be
6798 * embedded in the BP, no write was performed so there will be no
6799 * dva/birth/checksum. The buffer must therefore remain anonymous
6802 if (!HDR_EMPTY(hdr
)) {
6803 arc_buf_hdr_t
*exists
;
6804 kmutex_t
*hash_lock
;
6806 ASSERT3U(zio
->io_error
, ==, 0);
6808 arc_cksum_verify(buf
);
6810 exists
= buf_hash_insert(hdr
, &hash_lock
);
6811 if (exists
!= NULL
) {
6813 * This can only happen if we overwrite for
6814 * sync-to-convergence, because we remove
6815 * buffers from the hash table when we arc_free().
6817 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
6818 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
6819 panic("bad overwrite, hdr=%p exists=%p",
6820 (void *)hdr
, (void *)exists
);
6821 ASSERT(refcount_is_zero(
6822 &exists
->b_l1hdr
.b_refcnt
));
6823 arc_change_state(arc_anon
, exists
, hash_lock
);
6824 mutex_exit(hash_lock
);
6825 arc_hdr_destroy(exists
);
6826 exists
= buf_hash_insert(hdr
, &hash_lock
);
6827 ASSERT3P(exists
, ==, NULL
);
6828 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
6830 ASSERT(zio
->io_prop
.zp_nopwrite
);
6831 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
6832 panic("bad nopwrite, hdr=%p exists=%p",
6833 (void *)hdr
, (void *)exists
);
6836 ASSERT(hdr
->b_l1hdr
.b_bufcnt
== 1);
6837 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
6838 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
6839 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
6842 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6843 /* if it's not anon, we are doing a scrub */
6844 if (exists
== NULL
&& hdr
->b_l1hdr
.b_state
== arc_anon
)
6845 arc_access(hdr
, hash_lock
);
6846 mutex_exit(hash_lock
);
6848 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6851 ASSERT(!refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
6852 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
6854 abd_put(zio
->io_abd
);
6855 kmem_free(callback
, sizeof (arc_write_callback_t
));
6859 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
6860 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
,
6861 const zio_prop_t
*zp
, arc_write_done_func_t
*ready
,
6862 arc_write_done_func_t
*children_ready
, arc_write_done_func_t
*physdone
,
6863 arc_write_done_func_t
*done
, void *private, zio_priority_t priority
,
6864 int zio_flags
, const zbookmark_phys_t
*zb
)
6866 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
6867 arc_write_callback_t
*callback
;
6869 zio_prop_t localprop
= *zp
;
6871 ASSERT3P(ready
, !=, NULL
);
6872 ASSERT3P(done
, !=, NULL
);
6873 ASSERT(!HDR_IO_ERROR(hdr
));
6874 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
6875 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
6876 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
6878 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
6880 if (ARC_BUF_ENCRYPTED(buf
)) {
6881 ASSERT(ARC_BUF_COMPRESSED(buf
));
6882 localprop
.zp_encrypt
= B_TRUE
;
6883 localprop
.zp_compress
= HDR_GET_COMPRESS(hdr
);
6884 localprop
.zp_byteorder
=
6885 (hdr
->b_l1hdr
.b_byteswap
== DMU_BSWAP_NUMFUNCS
) ?
6886 ZFS_HOST_BYTEORDER
: !ZFS_HOST_BYTEORDER
;
6887 bcopy(hdr
->b_crypt_hdr
.b_salt
, localprop
.zp_salt
,
6889 bcopy(hdr
->b_crypt_hdr
.b_iv
, localprop
.zp_iv
,
6891 bcopy(hdr
->b_crypt_hdr
.b_mac
, localprop
.zp_mac
,
6893 if (DMU_OT_IS_ENCRYPTED(localprop
.zp_type
)) {
6894 localprop
.zp_nopwrite
= B_FALSE
;
6895 localprop
.zp_copies
=
6896 MIN(localprop
.zp_copies
, SPA_DVAS_PER_BP
- 1);
6898 zio_flags
|= ZIO_FLAG_RAW
;
6899 } else if (ARC_BUF_COMPRESSED(buf
)) {
6900 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, arc_buf_size(buf
));
6901 localprop
.zp_compress
= HDR_GET_COMPRESS(hdr
);
6902 zio_flags
|= ZIO_FLAG_RAW_COMPRESS
;
6904 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
6905 callback
->awcb_ready
= ready
;
6906 callback
->awcb_children_ready
= children_ready
;
6907 callback
->awcb_physdone
= physdone
;
6908 callback
->awcb_done
= done
;
6909 callback
->awcb_private
= private;
6910 callback
->awcb_buf
= buf
;
6913 * The hdr's b_pabd is now stale, free it now. A new data block
6914 * will be allocated when the zio pipeline calls arc_write_ready().
6916 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
6918 * If the buf is currently sharing the data block with
6919 * the hdr then we need to break that relationship here.
6920 * The hdr will remain with a NULL data pointer and the
6921 * buf will take sole ownership of the block.
6923 if (arc_buf_is_shared(buf
)) {
6924 arc_unshare_buf(hdr
, buf
);
6926 arc_hdr_free_abd(hdr
, B_FALSE
);
6928 VERIFY3P(buf
->b_data
, !=, NULL
);
6931 if (HDR_HAS_RABD(hdr
))
6932 arc_hdr_free_abd(hdr
, B_TRUE
);
6934 if (!(zio_flags
& ZIO_FLAG_RAW
))
6935 arc_hdr_set_compress(hdr
, ZIO_COMPRESS_OFF
);
6937 ASSERT(!arc_buf_is_shared(buf
));
6938 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
6940 zio
= zio_write(pio
, spa
, txg
, bp
,
6941 abd_get_from_buf(buf
->b_data
, HDR_GET_LSIZE(hdr
)),
6942 HDR_GET_LSIZE(hdr
), arc_buf_size(buf
), &localprop
, arc_write_ready
,
6943 (children_ready
!= NULL
) ? arc_write_children_ready
: NULL
,
6944 arc_write_physdone
, arc_write_done
, callback
,
6945 priority
, zio_flags
, zb
);
6951 arc_memory_throttle(uint64_t reserve
, uint64_t txg
)
6954 uint64_t available_memory
= arc_free_memory();
6955 static uint64_t page_load
= 0;
6956 static uint64_t last_txg
= 0;
6960 MIN(available_memory
, vmem_size(heap_arena
, VMEM_FREE
));
6963 if (available_memory
> arc_all_memory() * arc_lotsfree_percent
/ 100)
6966 if (txg
> last_txg
) {
6971 * If we are in pageout, we know that memory is already tight,
6972 * the arc is already going to be evicting, so we just want to
6973 * continue to let page writes occur as quickly as possible.
6975 if (current_is_kswapd()) {
6976 if (page_load
> MAX(arc_sys_free
/ 4, available_memory
) / 4) {
6977 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
6978 return (SET_ERROR(ERESTART
));
6980 /* Note: reserve is inflated, so we deflate */
6981 page_load
+= reserve
/ 8;
6983 } else if (page_load
> 0 && arc_reclaim_needed()) {
6984 /* memory is low, delay before restarting */
6985 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
6986 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
6987 return (SET_ERROR(EAGAIN
));
6995 arc_tempreserve_clear(uint64_t reserve
)
6997 atomic_add_64(&arc_tempreserve
, -reserve
);
6998 ASSERT((int64_t)arc_tempreserve
>= 0);
7002 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
7008 reserve
> arc_c
/4 &&
7009 reserve
* 4 > (2ULL << SPA_MAXBLOCKSHIFT
))
7010 arc_c
= MIN(arc_c_max
, reserve
* 4);
7013 * Throttle when the calculated memory footprint for the TXG
7014 * exceeds the target ARC size.
7016 if (reserve
> arc_c
) {
7017 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
7018 return (SET_ERROR(ERESTART
));
7022 * Don't count loaned bufs as in flight dirty data to prevent long
7023 * network delays from blocking transactions that are ready to be
7024 * assigned to a txg.
7027 /* assert that it has not wrapped around */
7028 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
7030 anon_size
= MAX((int64_t)(refcount_count(&arc_anon
->arcs_size
) -
7031 arc_loaned_bytes
), 0);
7034 * Writes will, almost always, require additional memory allocations
7035 * in order to compress/encrypt/etc the data. We therefore need to
7036 * make sure that there is sufficient available memory for this.
7038 error
= arc_memory_throttle(reserve
, txg
);
7043 * Throttle writes when the amount of dirty data in the cache
7044 * gets too large. We try to keep the cache less than half full
7045 * of dirty blocks so that our sync times don't grow too large.
7046 * Note: if two requests come in concurrently, we might let them
7047 * both succeed, when one of them should fail. Not a huge deal.
7050 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
7051 anon_size
> arc_c
/ 4) {
7052 uint64_t meta_esize
=
7053 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
7054 uint64_t data_esize
=
7055 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
7056 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
7057 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
7058 arc_tempreserve
>> 10, meta_esize
>> 10,
7059 data_esize
>> 10, reserve
>> 10, arc_c
>> 10);
7060 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
7061 return (SET_ERROR(ERESTART
));
7063 atomic_add_64(&arc_tempreserve
, reserve
);
7068 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
7069 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
7071 size
->value
.ui64
= refcount_count(&state
->arcs_size
);
7072 evict_data
->value
.ui64
=
7073 refcount_count(&state
->arcs_esize
[ARC_BUFC_DATA
]);
7074 evict_metadata
->value
.ui64
=
7075 refcount_count(&state
->arcs_esize
[ARC_BUFC_METADATA
]);
7079 arc_kstat_update(kstat_t
*ksp
, int rw
)
7081 arc_stats_t
*as
= ksp
->ks_data
;
7083 if (rw
== KSTAT_WRITE
) {
7084 return (SET_ERROR(EACCES
));
7086 arc_kstat_update_state(arc_anon
,
7087 &as
->arcstat_anon_size
,
7088 &as
->arcstat_anon_evictable_data
,
7089 &as
->arcstat_anon_evictable_metadata
);
7090 arc_kstat_update_state(arc_mru
,
7091 &as
->arcstat_mru_size
,
7092 &as
->arcstat_mru_evictable_data
,
7093 &as
->arcstat_mru_evictable_metadata
);
7094 arc_kstat_update_state(arc_mru_ghost
,
7095 &as
->arcstat_mru_ghost_size
,
7096 &as
->arcstat_mru_ghost_evictable_data
,
7097 &as
->arcstat_mru_ghost_evictable_metadata
);
7098 arc_kstat_update_state(arc_mfu
,
7099 &as
->arcstat_mfu_size
,
7100 &as
->arcstat_mfu_evictable_data
,
7101 &as
->arcstat_mfu_evictable_metadata
);
7102 arc_kstat_update_state(arc_mfu_ghost
,
7103 &as
->arcstat_mfu_ghost_size
,
7104 &as
->arcstat_mfu_ghost_evictable_data
,
7105 &as
->arcstat_mfu_ghost_evictable_metadata
);
7107 as
->arcstat_memory_all_bytes
.value
.ui64
=
7109 as
->arcstat_memory_free_bytes
.value
.ui64
=
7111 as
->arcstat_memory_available_bytes
.value
.i64
=
7112 arc_available_memory();
7119 * This function *must* return indices evenly distributed between all
7120 * sublists of the multilist. This is needed due to how the ARC eviction
7121 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7122 * distributed between all sublists and uses this assumption when
7123 * deciding which sublist to evict from and how much to evict from it.
7126 arc_state_multilist_index_func(multilist_t
*ml
, void *obj
)
7128 arc_buf_hdr_t
*hdr
= obj
;
7131 * We rely on b_dva to generate evenly distributed index
7132 * numbers using buf_hash below. So, as an added precaution,
7133 * let's make sure we never add empty buffers to the arc lists.
7135 ASSERT(!HDR_EMPTY(hdr
));
7138 * The assumption here, is the hash value for a given
7139 * arc_buf_hdr_t will remain constant throughout its lifetime
7140 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7141 * Thus, we don't need to store the header's sublist index
7142 * on insertion, as this index can be recalculated on removal.
7144 * Also, the low order bits of the hash value are thought to be
7145 * distributed evenly. Otherwise, in the case that the multilist
7146 * has a power of two number of sublists, each sublists' usage
7147 * would not be evenly distributed.
7149 return (buf_hash(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
) %
7150 multilist_get_num_sublists(ml
));
7154 * Called during module initialization and periodically thereafter to
7155 * apply reasonable changes to the exposed performance tunings. Non-zero
7156 * zfs_* values which differ from the currently set values will be applied.
7159 arc_tuning_update(void)
7161 uint64_t allmem
= arc_all_memory();
7162 unsigned long limit
;
7164 /* Valid range: 64M - <all physical memory> */
7165 if ((zfs_arc_max
) && (zfs_arc_max
!= arc_c_max
) &&
7166 (zfs_arc_max
> 64 << 20) && (zfs_arc_max
< allmem
) &&
7167 (zfs_arc_max
> arc_c_min
)) {
7168 arc_c_max
= zfs_arc_max
;
7170 arc_p
= (arc_c
>> 1);
7171 if (arc_meta_limit
> arc_c_max
)
7172 arc_meta_limit
= arc_c_max
;
7173 if (arc_dnode_limit
> arc_meta_limit
)
7174 arc_dnode_limit
= arc_meta_limit
;
7177 /* Valid range: 32M - <arc_c_max> */
7178 if ((zfs_arc_min
) && (zfs_arc_min
!= arc_c_min
) &&
7179 (zfs_arc_min
>= 2ULL << SPA_MAXBLOCKSHIFT
) &&
7180 (zfs_arc_min
<= arc_c_max
)) {
7181 arc_c_min
= zfs_arc_min
;
7182 arc_c
= MAX(arc_c
, arc_c_min
);
7185 /* Valid range: 16M - <arc_c_max> */
7186 if ((zfs_arc_meta_min
) && (zfs_arc_meta_min
!= arc_meta_min
) &&
7187 (zfs_arc_meta_min
>= 1ULL << SPA_MAXBLOCKSHIFT
) &&
7188 (zfs_arc_meta_min
<= arc_c_max
)) {
7189 arc_meta_min
= zfs_arc_meta_min
;
7190 if (arc_meta_limit
< arc_meta_min
)
7191 arc_meta_limit
= arc_meta_min
;
7192 if (arc_dnode_limit
< arc_meta_min
)
7193 arc_dnode_limit
= arc_meta_min
;
7196 /* Valid range: <arc_meta_min> - <arc_c_max> */
7197 limit
= zfs_arc_meta_limit
? zfs_arc_meta_limit
:
7198 MIN(zfs_arc_meta_limit_percent
, 100) * arc_c_max
/ 100;
7199 if ((limit
!= arc_meta_limit
) &&
7200 (limit
>= arc_meta_min
) &&
7201 (limit
<= arc_c_max
))
7202 arc_meta_limit
= limit
;
7204 /* Valid range: <arc_meta_min> - <arc_meta_limit> */
7205 limit
= zfs_arc_dnode_limit
? zfs_arc_dnode_limit
:
7206 MIN(zfs_arc_dnode_limit_percent
, 100) * arc_meta_limit
/ 100;
7207 if ((limit
!= arc_dnode_limit
) &&
7208 (limit
>= arc_meta_min
) &&
7209 (limit
<= arc_meta_limit
))
7210 arc_dnode_limit
= limit
;
7212 /* Valid range: 1 - N */
7213 if (zfs_arc_grow_retry
)
7214 arc_grow_retry
= zfs_arc_grow_retry
;
7216 /* Valid range: 1 - N */
7217 if (zfs_arc_shrink_shift
) {
7218 arc_shrink_shift
= zfs_arc_shrink_shift
;
7219 arc_no_grow_shift
= MIN(arc_no_grow_shift
, arc_shrink_shift
-1);
7222 /* Valid range: 1 - N */
7223 if (zfs_arc_p_min_shift
)
7224 arc_p_min_shift
= zfs_arc_p_min_shift
;
7226 /* Valid range: 1 - N ticks */
7227 if (zfs_arc_min_prefetch_lifespan
)
7228 arc_min_prefetch_lifespan
= zfs_arc_min_prefetch_lifespan
;
7230 /* Valid range: 0 - 100 */
7231 if ((zfs_arc_lotsfree_percent
>= 0) &&
7232 (zfs_arc_lotsfree_percent
<= 100))
7233 arc_lotsfree_percent
= zfs_arc_lotsfree_percent
;
7235 /* Valid range: 0 - <all physical memory> */
7236 if ((zfs_arc_sys_free
) && (zfs_arc_sys_free
!= arc_sys_free
))
7237 arc_sys_free
= MIN(MAX(zfs_arc_sys_free
, 0), allmem
);
7242 arc_state_init(void)
7244 arc_anon
= &ARC_anon
;
7246 arc_mru_ghost
= &ARC_mru_ghost
;
7248 arc_mfu_ghost
= &ARC_mfu_ghost
;
7249 arc_l2c_only
= &ARC_l2c_only
;
7251 arc_mru
->arcs_list
[ARC_BUFC_METADATA
] =
7252 multilist_create(sizeof (arc_buf_hdr_t
),
7253 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7254 arc_state_multilist_index_func
);
7255 arc_mru
->arcs_list
[ARC_BUFC_DATA
] =
7256 multilist_create(sizeof (arc_buf_hdr_t
),
7257 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7258 arc_state_multilist_index_func
);
7259 arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
7260 multilist_create(sizeof (arc_buf_hdr_t
),
7261 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7262 arc_state_multilist_index_func
);
7263 arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
] =
7264 multilist_create(sizeof (arc_buf_hdr_t
),
7265 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7266 arc_state_multilist_index_func
);
7267 arc_mfu
->arcs_list
[ARC_BUFC_METADATA
] =
7268 multilist_create(sizeof (arc_buf_hdr_t
),
7269 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7270 arc_state_multilist_index_func
);
7271 arc_mfu
->arcs_list
[ARC_BUFC_DATA
] =
7272 multilist_create(sizeof (arc_buf_hdr_t
),
7273 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7274 arc_state_multilist_index_func
);
7275 arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
7276 multilist_create(sizeof (arc_buf_hdr_t
),
7277 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7278 arc_state_multilist_index_func
);
7279 arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
] =
7280 multilist_create(sizeof (arc_buf_hdr_t
),
7281 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7282 arc_state_multilist_index_func
);
7283 arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
] =
7284 multilist_create(sizeof (arc_buf_hdr_t
),
7285 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7286 arc_state_multilist_index_func
);
7287 arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
] =
7288 multilist_create(sizeof (arc_buf_hdr_t
),
7289 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
7290 arc_state_multilist_index_func
);
7292 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
7293 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
7294 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
7295 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
7296 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
7297 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
7298 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
7299 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
7300 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
7301 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
7302 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
7303 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
7305 refcount_create(&arc_anon
->arcs_size
);
7306 refcount_create(&arc_mru
->arcs_size
);
7307 refcount_create(&arc_mru_ghost
->arcs_size
);
7308 refcount_create(&arc_mfu
->arcs_size
);
7309 refcount_create(&arc_mfu_ghost
->arcs_size
);
7310 refcount_create(&arc_l2c_only
->arcs_size
);
7312 arc_anon
->arcs_state
= ARC_STATE_ANON
;
7313 arc_mru
->arcs_state
= ARC_STATE_MRU
;
7314 arc_mru_ghost
->arcs_state
= ARC_STATE_MRU_GHOST
;
7315 arc_mfu
->arcs_state
= ARC_STATE_MFU
;
7316 arc_mfu_ghost
->arcs_state
= ARC_STATE_MFU_GHOST
;
7317 arc_l2c_only
->arcs_state
= ARC_STATE_L2C_ONLY
;
7321 arc_state_fini(void)
7323 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
7324 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
7325 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
7326 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
7327 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
7328 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
7329 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
7330 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
7331 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
7332 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
7333 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
7334 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
7336 refcount_destroy(&arc_anon
->arcs_size
);
7337 refcount_destroy(&arc_mru
->arcs_size
);
7338 refcount_destroy(&arc_mru_ghost
->arcs_size
);
7339 refcount_destroy(&arc_mfu
->arcs_size
);
7340 refcount_destroy(&arc_mfu_ghost
->arcs_size
);
7341 refcount_destroy(&arc_l2c_only
->arcs_size
);
7343 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
7344 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
7345 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
7346 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
7347 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
7348 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
7349 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
7350 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
7351 multilist_destroy(arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]);
7352 multilist_destroy(arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]);
7356 arc_target_bytes(void)
7364 uint64_t percent
, allmem
= arc_all_memory();
7366 mutex_init(&arc_reclaim_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
7367 cv_init(&arc_reclaim_thread_cv
, NULL
, CV_DEFAULT
, NULL
);
7368 cv_init(&arc_reclaim_waiters_cv
, NULL
, CV_DEFAULT
, NULL
);
7370 /* Convert seconds to clock ticks */
7371 arc_min_prefetch_lifespan
= 1 * hz
;
7375 * Register a shrinker to support synchronous (direct) memory
7376 * reclaim from the arc. This is done to prevent kswapd from
7377 * swapping out pages when it is preferable to shrink the arc.
7379 spl_register_shrinker(&arc_shrinker
);
7381 /* Set to 1/64 of all memory or a minimum of 512K */
7382 arc_sys_free
= MAX(allmem
/ 64, (512 * 1024));
7386 /* Set max to 1/2 of all memory */
7387 arc_c_max
= allmem
/ 2;
7390 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more */
7391 arc_c_min
= MAX(allmem
/ 32, 2ULL << SPA_MAXBLOCKSHIFT
);
7394 * In userland, there's only the memory pressure that we artificially
7395 * create (see arc_available_memory()). Don't let arc_c get too
7396 * small, because it can cause transactions to be larger than
7397 * arc_c, causing arc_tempreserve_space() to fail.
7399 arc_c_min
= MAX(arc_c_max
/ 2, 2ULL << SPA_MAXBLOCKSHIFT
);
7403 arc_p
= (arc_c
>> 1);
7406 /* Set min to 1/2 of arc_c_min */
7407 arc_meta_min
= 1ULL << SPA_MAXBLOCKSHIFT
;
7408 /* Initialize maximum observed usage to zero */
7411 * Set arc_meta_limit to a percent of arc_c_max with a floor of
7412 * arc_meta_min, and a ceiling of arc_c_max.
7414 percent
= MIN(zfs_arc_meta_limit_percent
, 100);
7415 arc_meta_limit
= MAX(arc_meta_min
, (percent
* arc_c_max
) / 100);
7416 percent
= MIN(zfs_arc_dnode_limit_percent
, 100);
7417 arc_dnode_limit
= (percent
* arc_meta_limit
) / 100;
7419 /* Apply user specified tunings */
7420 arc_tuning_update();
7422 /* if kmem_flags are set, lets try to use less memory */
7423 if (kmem_debugging())
7425 if (arc_c
< arc_c_min
)
7431 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
7432 offsetof(arc_prune_t
, p_node
));
7433 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7435 arc_prune_taskq
= taskq_create("arc_prune", max_ncpus
, defclsyspri
,
7436 max_ncpus
, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
7438 arc_reclaim_thread_exit
= B_FALSE
;
7440 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
7441 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
7443 if (arc_ksp
!= NULL
) {
7444 arc_ksp
->ks_data
= &arc_stats
;
7445 arc_ksp
->ks_update
= arc_kstat_update
;
7446 kstat_install(arc_ksp
);
7449 (void) thread_create(NULL
, 0, arc_reclaim_thread
, NULL
, 0, &p0
,
7450 TS_RUN
, defclsyspri
);
7456 * Calculate maximum amount of dirty data per pool.
7458 * If it has been set by a module parameter, take that.
7459 * Otherwise, use a percentage of physical memory defined by
7460 * zfs_dirty_data_max_percent (default 10%) with a cap at
7461 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
7463 if (zfs_dirty_data_max_max
== 0)
7464 zfs_dirty_data_max_max
= MIN(4ULL * 1024 * 1024 * 1024,
7465 allmem
* zfs_dirty_data_max_max_percent
/ 100);
7467 if (zfs_dirty_data_max
== 0) {
7468 zfs_dirty_data_max
= allmem
*
7469 zfs_dirty_data_max_percent
/ 100;
7470 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
7471 zfs_dirty_data_max_max
);
7481 spl_unregister_shrinker(&arc_shrinker
);
7482 #endif /* _KERNEL */
7484 mutex_enter(&arc_reclaim_lock
);
7485 arc_reclaim_thread_exit
= B_TRUE
;
7487 * The reclaim thread will set arc_reclaim_thread_exit back to
7488 * B_FALSE when it is finished exiting; we're waiting for that.
7490 while (arc_reclaim_thread_exit
) {
7491 cv_signal(&arc_reclaim_thread_cv
);
7492 cv_wait(&arc_reclaim_thread_cv
, &arc_reclaim_lock
);
7494 mutex_exit(&arc_reclaim_lock
);
7496 /* Use B_TRUE to ensure *all* buffers are evicted */
7497 arc_flush(NULL
, B_TRUE
);
7501 if (arc_ksp
!= NULL
) {
7502 kstat_delete(arc_ksp
);
7506 taskq_wait(arc_prune_taskq
);
7507 taskq_destroy(arc_prune_taskq
);
7509 mutex_enter(&arc_prune_mtx
);
7510 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
7511 list_remove(&arc_prune_list
, p
);
7512 refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
7513 refcount_destroy(&p
->p_refcnt
);
7514 kmem_free(p
, sizeof (*p
));
7516 mutex_exit(&arc_prune_mtx
);
7518 list_destroy(&arc_prune_list
);
7519 mutex_destroy(&arc_prune_mtx
);
7520 mutex_destroy(&arc_reclaim_lock
);
7521 cv_destroy(&arc_reclaim_thread_cv
);
7522 cv_destroy(&arc_reclaim_waiters_cv
);
7527 ASSERT0(arc_loaned_bytes
);
7533 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7534 * It uses dedicated storage devices to hold cached data, which are populated
7535 * using large infrequent writes. The main role of this cache is to boost
7536 * the performance of random read workloads. The intended L2ARC devices
7537 * include short-stroked disks, solid state disks, and other media with
7538 * substantially faster read latency than disk.
7540 * +-----------------------+
7542 * +-----------------------+
7545 * l2arc_feed_thread() arc_read()
7549 * +---------------+ |
7551 * +---------------+ |
7556 * +-------+ +-------+
7558 * | cache | | cache |
7559 * +-------+ +-------+
7560 * +=========+ .-----.
7561 * : L2ARC : |-_____-|
7562 * : devices : | Disks |
7563 * +=========+ `-_____-'
7565 * Read requests are satisfied from the following sources, in order:
7568 * 2) vdev cache of L2ARC devices
7570 * 4) vdev cache of disks
7573 * Some L2ARC device types exhibit extremely slow write performance.
7574 * To accommodate for this there are some significant differences between
7575 * the L2ARC and traditional cache design:
7577 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7578 * the ARC behave as usual, freeing buffers and placing headers on ghost
7579 * lists. The ARC does not send buffers to the L2ARC during eviction as
7580 * this would add inflated write latencies for all ARC memory pressure.
7582 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7583 * It does this by periodically scanning buffers from the eviction-end of
7584 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7585 * not already there. It scans until a headroom of buffers is satisfied,
7586 * which itself is a buffer for ARC eviction. If a compressible buffer is
7587 * found during scanning and selected for writing to an L2ARC device, we
7588 * temporarily boost scanning headroom during the next scan cycle to make
7589 * sure we adapt to compression effects (which might significantly reduce
7590 * the data volume we write to L2ARC). The thread that does this is
7591 * l2arc_feed_thread(), illustrated below; example sizes are included to
7592 * provide a better sense of ratio than this diagram:
7595 * +---------------------+----------+
7596 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7597 * +---------------------+----------+ | o L2ARC eligible
7598 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7599 * +---------------------+----------+ |
7600 * 15.9 Gbytes ^ 32 Mbytes |
7602 * l2arc_feed_thread()
7604 * l2arc write hand <--[oooo]--'
7608 * +==============================+
7609 * L2ARC dev |####|#|###|###| |####| ... |
7610 * +==============================+
7613 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7614 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7615 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7616 * safe to say that this is an uncommon case, since buffers at the end of
7617 * the ARC lists have moved there due to inactivity.
7619 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7620 * then the L2ARC simply misses copying some buffers. This serves as a
7621 * pressure valve to prevent heavy read workloads from both stalling the ARC
7622 * with waits and clogging the L2ARC with writes. This also helps prevent
7623 * the potential for the L2ARC to churn if it attempts to cache content too
7624 * quickly, such as during backups of the entire pool.
7626 * 5. After system boot and before the ARC has filled main memory, there are
7627 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7628 * lists can remain mostly static. Instead of searching from tail of these
7629 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7630 * for eligible buffers, greatly increasing its chance of finding them.
7632 * The L2ARC device write speed is also boosted during this time so that
7633 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7634 * there are no L2ARC reads, and no fear of degrading read performance
7635 * through increased writes.
7637 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7638 * the vdev queue can aggregate them into larger and fewer writes. Each
7639 * device is written to in a rotor fashion, sweeping writes through
7640 * available space then repeating.
7642 * 7. The L2ARC does not store dirty content. It never needs to flush
7643 * write buffers back to disk based storage.
7645 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7646 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7648 * The performance of the L2ARC can be tweaked by a number of tunables, which
7649 * may be necessary for different workloads:
7651 * l2arc_write_max max write bytes per interval
7652 * l2arc_write_boost extra write bytes during device warmup
7653 * l2arc_noprefetch skip caching prefetched buffers
7654 * l2arc_headroom number of max device writes to precache
7655 * l2arc_headroom_boost when we find compressed buffers during ARC
7656 * scanning, we multiply headroom by this
7657 * percentage factor for the next scan cycle,
7658 * since more compressed buffers are likely to
7660 * l2arc_feed_secs seconds between L2ARC writing
7662 * Tunables may be removed or added as future performance improvements are
7663 * integrated, and also may become zpool properties.
7665 * There are three key functions that control how the L2ARC warms up:
7667 * l2arc_write_eligible() check if a buffer is eligible to cache
7668 * l2arc_write_size() calculate how much to write
7669 * l2arc_write_interval() calculate sleep delay between writes
7671 * These three functions determine what to write, how much, and how quickly
7676 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*hdr
)
7679 * A buffer is *not* eligible for the L2ARC if it:
7680 * 1. belongs to a different spa.
7681 * 2. is already cached on the L2ARC.
7682 * 3. has an I/O in progress (it may be an incomplete read).
7683 * 4. is flagged not eligible (zfs property).
7685 if (hdr
->b_spa
!= spa_guid
|| HDR_HAS_L2HDR(hdr
) ||
7686 HDR_IO_IN_PROGRESS(hdr
) || !HDR_L2CACHE(hdr
))
7693 l2arc_write_size(void)
7698 * Make sure our globals have meaningful values in case the user
7701 size
= l2arc_write_max
;
7703 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
7704 "be greater than zero, resetting it to the default (%d)",
7706 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
7709 if (arc_warm
== B_FALSE
)
7710 size
+= l2arc_write_boost
;
7717 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
7719 clock_t interval
, next
, now
;
7722 * If the ARC lists are busy, increase our write rate; if the
7723 * lists are stale, idle back. This is achieved by checking
7724 * how much we previously wrote - if it was more than half of
7725 * what we wanted, schedule the next write much sooner.
7727 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
7728 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
7730 interval
= hz
* l2arc_feed_secs
;
7732 now
= ddi_get_lbolt();
7733 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
7739 * Cycle through L2ARC devices. This is how L2ARC load balances.
7740 * If a device is returned, this also returns holding the spa config lock.
7742 static l2arc_dev_t
*
7743 l2arc_dev_get_next(void)
7745 l2arc_dev_t
*first
, *next
= NULL
;
7748 * Lock out the removal of spas (spa_namespace_lock), then removal
7749 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7750 * both locks will be dropped and a spa config lock held instead.
7752 mutex_enter(&spa_namespace_lock
);
7753 mutex_enter(&l2arc_dev_mtx
);
7755 /* if there are no vdevs, there is nothing to do */
7756 if (l2arc_ndev
== 0)
7760 next
= l2arc_dev_last
;
7762 /* loop around the list looking for a non-faulted vdev */
7764 next
= list_head(l2arc_dev_list
);
7766 next
= list_next(l2arc_dev_list
, next
);
7768 next
= list_head(l2arc_dev_list
);
7771 /* if we have come back to the start, bail out */
7774 else if (next
== first
)
7777 } while (vdev_is_dead(next
->l2ad_vdev
));
7779 /* if we were unable to find any usable vdevs, return NULL */
7780 if (vdev_is_dead(next
->l2ad_vdev
))
7783 l2arc_dev_last
= next
;
7786 mutex_exit(&l2arc_dev_mtx
);
7789 * Grab the config lock to prevent the 'next' device from being
7790 * removed while we are writing to it.
7793 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
7794 mutex_exit(&spa_namespace_lock
);
7800 * Free buffers that were tagged for destruction.
7803 l2arc_do_free_on_write(void)
7806 l2arc_data_free_t
*df
, *df_prev
;
7808 mutex_enter(&l2arc_free_on_write_mtx
);
7809 buflist
= l2arc_free_on_write
;
7811 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
7812 df_prev
= list_prev(buflist
, df
);
7813 ASSERT3P(df
->l2df_abd
, !=, NULL
);
7814 abd_free(df
->l2df_abd
);
7815 list_remove(buflist
, df
);
7816 kmem_free(df
, sizeof (l2arc_data_free_t
));
7819 mutex_exit(&l2arc_free_on_write_mtx
);
7823 * A write to a cache device has completed. Update all headers to allow
7824 * reads from these buffers to begin.
7827 l2arc_write_done(zio_t
*zio
)
7829 l2arc_write_callback_t
*cb
;
7832 arc_buf_hdr_t
*head
, *hdr
, *hdr_prev
;
7833 kmutex_t
*hash_lock
;
7834 int64_t bytes_dropped
= 0;
7836 cb
= zio
->io_private
;
7837 ASSERT3P(cb
, !=, NULL
);
7838 dev
= cb
->l2wcb_dev
;
7839 ASSERT3P(dev
, !=, NULL
);
7840 head
= cb
->l2wcb_head
;
7841 ASSERT3P(head
, !=, NULL
);
7842 buflist
= &dev
->l2ad_buflist
;
7843 ASSERT3P(buflist
, !=, NULL
);
7844 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
7845 l2arc_write_callback_t
*, cb
);
7847 if (zio
->io_error
!= 0)
7848 ARCSTAT_BUMP(arcstat_l2_writes_error
);
7851 * All writes completed, or an error was hit.
7854 mutex_enter(&dev
->l2ad_mtx
);
7855 for (hdr
= list_prev(buflist
, head
); hdr
; hdr
= hdr_prev
) {
7856 hdr_prev
= list_prev(buflist
, hdr
);
7858 hash_lock
= HDR_LOCK(hdr
);
7861 * We cannot use mutex_enter or else we can deadlock
7862 * with l2arc_write_buffers (due to swapping the order
7863 * the hash lock and l2ad_mtx are taken).
7865 if (!mutex_tryenter(hash_lock
)) {
7867 * Missed the hash lock. We must retry so we
7868 * don't leave the ARC_FLAG_L2_WRITING bit set.
7870 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry
);
7873 * We don't want to rescan the headers we've
7874 * already marked as having been written out, so
7875 * we reinsert the head node so we can pick up
7876 * where we left off.
7878 list_remove(buflist
, head
);
7879 list_insert_after(buflist
, hdr
, head
);
7881 mutex_exit(&dev
->l2ad_mtx
);
7884 * We wait for the hash lock to become available
7885 * to try and prevent busy waiting, and increase
7886 * the chance we'll be able to acquire the lock
7887 * the next time around.
7889 mutex_enter(hash_lock
);
7890 mutex_exit(hash_lock
);
7895 * We could not have been moved into the arc_l2c_only
7896 * state while in-flight due to our ARC_FLAG_L2_WRITING
7897 * bit being set. Let's just ensure that's being enforced.
7899 ASSERT(HDR_HAS_L1HDR(hdr
));
7902 * Skipped - drop L2ARC entry and mark the header as no
7903 * longer L2 eligibile.
7905 if (zio
->io_error
!= 0) {
7907 * Error - drop L2ARC entry.
7909 list_remove(buflist
, hdr
);
7910 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
7912 ARCSTAT_INCR(arcstat_l2_psize
, -arc_hdr_size(hdr
));
7913 ARCSTAT_INCR(arcstat_l2_lsize
, -HDR_GET_LSIZE(hdr
));
7915 bytes_dropped
+= arc_hdr_size(hdr
);
7916 (void) refcount_remove_many(&dev
->l2ad_alloc
,
7917 arc_hdr_size(hdr
), hdr
);
7921 * Allow ARC to begin reads and ghost list evictions to
7924 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_WRITING
);
7926 mutex_exit(hash_lock
);
7929 atomic_inc_64(&l2arc_writes_done
);
7930 list_remove(buflist
, head
);
7931 ASSERT(!HDR_HAS_L1HDR(head
));
7932 kmem_cache_free(hdr_l2only_cache
, head
);
7933 mutex_exit(&dev
->l2ad_mtx
);
7935 vdev_space_update(dev
->l2ad_vdev
, -bytes_dropped
, 0, 0);
7937 l2arc_do_free_on_write();
7939 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
7943 l2arc_untransform(zio_t
*zio
, l2arc_read_callback_t
*cb
)
7946 spa_t
*spa
= zio
->io_spa
;
7947 arc_buf_hdr_t
*hdr
= cb
->l2rcb_hdr
;
7948 blkptr_t
*bp
= zio
->io_bp
;
7949 dsl_crypto_key_t
*dck
= NULL
;
7950 uint8_t salt
[ZIO_DATA_SALT_LEN
];
7951 uint8_t iv
[ZIO_DATA_IV_LEN
];
7952 uint8_t mac
[ZIO_DATA_MAC_LEN
];
7953 boolean_t no_crypt
= B_FALSE
;
7956 * ZIL data is never be written to the L2ARC, so we don't need
7957 * special handling for its unique MAC storage.
7959 ASSERT3U(BP_GET_TYPE(bp
), !=, DMU_OT_INTENT_LOG
);
7960 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
7961 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
7964 * If the data was encrypted, decrypt it now. Note that
7965 * we must check the bp here and not the hdr, since the
7966 * hdr does not have its encryption parameters updated
7967 * until arc_read_done().
7969 if (BP_IS_ENCRYPTED(bp
)) {
7970 abd_t
*eabd
= arc_get_data_abd(hdr
,
7971 arc_hdr_size(hdr
), hdr
);
7973 zio_crypt_decode_params_bp(bp
, salt
, iv
);
7974 zio_crypt_decode_mac_bp(bp
, mac
);
7976 ret
= spa_keystore_lookup_key(spa
,
7977 cb
->l2rcb_zb
.zb_objset
, FTAG
, &dck
);
7979 arc_free_data_abd(hdr
, eabd
, arc_hdr_size(hdr
), hdr
);
7983 ret
= zio_do_crypt_abd(B_FALSE
, &dck
->dck_key
,
7984 salt
, BP_GET_TYPE(bp
), iv
, mac
, HDR_GET_PSIZE(hdr
),
7985 BP_SHOULD_BYTESWAP(bp
), eabd
, hdr
->b_l1hdr
.b_pabd
,
7988 arc_free_data_abd(hdr
, eabd
, arc_hdr_size(hdr
), hdr
);
7989 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
7993 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
7996 * If we actually performed decryption, replace b_pabd
7997 * with the decrypted data. Otherwise we can just throw
7998 * our decryption buffer away.
8001 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
,
8002 arc_hdr_size(hdr
), hdr
);
8003 hdr
->b_l1hdr
.b_pabd
= eabd
;
8006 arc_free_data_abd(hdr
, eabd
, arc_hdr_size(hdr
), hdr
);
8011 * If the L2ARC block was compressed, but ARC compression
8012 * is disabled we decompress the data into a new buffer and
8013 * replace the existing data.
8015 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
8016 !HDR_COMPRESSION_ENABLED(hdr
)) {
8017 abd_t
*cabd
= arc_get_data_abd(hdr
, arc_hdr_size(hdr
), hdr
);
8018 void *tmp
= abd_borrow_buf(cabd
, arc_hdr_size(hdr
));
8020 ret
= zio_decompress_data(HDR_GET_COMPRESS(hdr
),
8021 hdr
->b_l1hdr
.b_pabd
, tmp
, HDR_GET_PSIZE(hdr
),
8022 HDR_GET_LSIZE(hdr
));
8024 abd_return_buf_copy(cabd
, tmp
, arc_hdr_size(hdr
));
8025 arc_free_data_abd(hdr
, cabd
, arc_hdr_size(hdr
), hdr
);
8029 abd_return_buf_copy(cabd
, tmp
, arc_hdr_size(hdr
));
8030 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
,
8031 arc_hdr_size(hdr
), hdr
);
8032 hdr
->b_l1hdr
.b_pabd
= cabd
;
8034 zio
->io_size
= HDR_GET_LSIZE(hdr
);
8045 * A read to a cache device completed. Validate buffer contents before
8046 * handing over to the regular ARC routines.
8049 l2arc_read_done(zio_t
*zio
)
8052 l2arc_read_callback_t
*cb
;
8054 kmutex_t
*hash_lock
;
8055 boolean_t valid_cksum
, using_rdata
;
8057 ASSERT3P(zio
->io_vd
, !=, NULL
);
8058 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
8060 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
8062 cb
= zio
->io_private
;
8063 ASSERT3P(cb
, !=, NULL
);
8064 hdr
= cb
->l2rcb_hdr
;
8065 ASSERT3P(hdr
, !=, NULL
);
8067 hash_lock
= HDR_LOCK(hdr
);
8068 mutex_enter(hash_lock
);
8069 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
8072 * If the data was read into a temporary buffer,
8073 * move it and free the buffer.
8075 if (cb
->l2rcb_abd
!= NULL
) {
8076 ASSERT3U(arc_hdr_size(hdr
), <, zio
->io_size
);
8077 if (zio
->io_error
== 0) {
8078 abd_copy(hdr
->b_l1hdr
.b_pabd
, cb
->l2rcb_abd
,
8083 * The following must be done regardless of whether
8084 * there was an error:
8085 * - free the temporary buffer
8086 * - point zio to the real ARC buffer
8087 * - set zio size accordingly
8088 * These are required because zio is either re-used for
8089 * an I/O of the block in the case of the error
8090 * or the zio is passed to arc_read_done() and it
8093 abd_free(cb
->l2rcb_abd
);
8094 zio
->io_size
= zio
->io_orig_size
= arc_hdr_size(hdr
);
8096 if (BP_IS_ENCRYPTED(&cb
->l2rcb_bp
) &&
8097 (cb
->l2rcb_flags
& ZIO_FLAG_RAW_ENCRYPT
)) {
8098 ASSERT(HDR_HAS_RABD(hdr
));
8099 zio
->io_abd
= zio
->io_orig_abd
=
8100 hdr
->b_crypt_hdr
.b_rabd
;
8102 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
8103 zio
->io_abd
= zio
->io_orig_abd
= hdr
->b_l1hdr
.b_pabd
;
8107 ASSERT3P(zio
->io_abd
, !=, NULL
);
8110 * Check this survived the L2ARC journey.
8112 ASSERT(zio
->io_abd
== hdr
->b_l1hdr
.b_pabd
||
8113 (HDR_HAS_RABD(hdr
) && zio
->io_abd
== hdr
->b_crypt_hdr
.b_rabd
));
8114 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
8115 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
8117 valid_cksum
= arc_cksum_is_equal(hdr
, zio
);
8118 using_rdata
= (HDR_HAS_RABD(hdr
) &&
8119 zio
->io_abd
== hdr
->b_crypt_hdr
.b_rabd
);
8122 * b_rabd will always match the data as it exists on disk if it is
8123 * being used. Therefore if we are reading into b_rabd we do not
8124 * attempt to untransform the data.
8126 if (valid_cksum
&& !using_rdata
)
8127 tfm_error
= l2arc_untransform(zio
, cb
);
8129 if (valid_cksum
&& tfm_error
== 0 && zio
->io_error
== 0 &&
8130 !HDR_L2_EVICTED(hdr
)) {
8131 mutex_exit(hash_lock
);
8132 zio
->io_private
= hdr
;
8135 mutex_exit(hash_lock
);
8137 * Buffer didn't survive caching. Increment stats and
8138 * reissue to the original storage device.
8140 if (zio
->io_error
!= 0) {
8141 ARCSTAT_BUMP(arcstat_l2_io_error
);
8143 zio
->io_error
= SET_ERROR(EIO
);
8145 if (!valid_cksum
|| tfm_error
!= 0)
8146 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
8149 * If there's no waiter, issue an async i/o to the primary
8150 * storage now. If there *is* a waiter, the caller must
8151 * issue the i/o in a context where it's OK to block.
8153 if (zio
->io_waiter
== NULL
) {
8154 zio_t
*pio
= zio_unique_parent(zio
);
8155 void *abd
= (using_rdata
) ?
8156 hdr
->b_crypt_hdr
.b_rabd
: hdr
->b_l1hdr
.b_pabd
;
8158 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
8160 zio_nowait(zio_read(pio
, zio
->io_spa
, zio
->io_bp
,
8161 abd
, zio
->io_size
, arc_read_done
,
8162 hdr
, zio
->io_priority
, cb
->l2rcb_flags
,
8167 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
8171 * This is the list priority from which the L2ARC will search for pages to
8172 * cache. This is used within loops (0..3) to cycle through lists in the
8173 * desired order. This order can have a significant effect on cache
8176 * Currently the metadata lists are hit first, MFU then MRU, followed by
8177 * the data lists. This function returns a locked list, and also returns
8180 static multilist_sublist_t
*
8181 l2arc_sublist_lock(int list_num
)
8183 multilist_t
*ml
= NULL
;
8186 ASSERT(list_num
>= 0 && list_num
< L2ARC_FEED_TYPES
);
8190 ml
= arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
8193 ml
= arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
8196 ml
= arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
8199 ml
= arc_mru
->arcs_list
[ARC_BUFC_DATA
];
8206 * Return a randomly-selected sublist. This is acceptable
8207 * because the caller feeds only a little bit of data for each
8208 * call (8MB). Subsequent calls will result in different
8209 * sublists being selected.
8211 idx
= multilist_get_random_index(ml
);
8212 return (multilist_sublist_lock(ml
, idx
));
8216 * Evict buffers from the device write hand to the distance specified in
8217 * bytes. This distance may span populated buffers, it may span nothing.
8218 * This is clearing a region on the L2ARC device ready for writing.
8219 * If the 'all' boolean is set, every buffer is evicted.
8222 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
8225 arc_buf_hdr_t
*hdr
, *hdr_prev
;
8226 kmutex_t
*hash_lock
;
8229 buflist
= &dev
->l2ad_buflist
;
8231 if (!all
&& dev
->l2ad_first
) {
8233 * This is the first sweep through the device. There is
8239 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
8241 * When nearing the end of the device, evict to the end
8242 * before the device write hand jumps to the start.
8244 taddr
= dev
->l2ad_end
;
8246 taddr
= dev
->l2ad_hand
+ distance
;
8248 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
8249 uint64_t, taddr
, boolean_t
, all
);
8252 mutex_enter(&dev
->l2ad_mtx
);
8253 for (hdr
= list_tail(buflist
); hdr
; hdr
= hdr_prev
) {
8254 hdr_prev
= list_prev(buflist
, hdr
);
8256 hash_lock
= HDR_LOCK(hdr
);
8259 * We cannot use mutex_enter or else we can deadlock
8260 * with l2arc_write_buffers (due to swapping the order
8261 * the hash lock and l2ad_mtx are taken).
8263 if (!mutex_tryenter(hash_lock
)) {
8265 * Missed the hash lock. Retry.
8267 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
8268 mutex_exit(&dev
->l2ad_mtx
);
8269 mutex_enter(hash_lock
);
8270 mutex_exit(hash_lock
);
8275 * A header can't be on this list if it doesn't have L2 header.
8277 ASSERT(HDR_HAS_L2HDR(hdr
));
8279 /* Ensure this header has finished being written. */
8280 ASSERT(!HDR_L2_WRITING(hdr
));
8281 ASSERT(!HDR_L2_WRITE_HEAD(hdr
));
8283 if (!all
&& (hdr
->b_l2hdr
.b_daddr
>= taddr
||
8284 hdr
->b_l2hdr
.b_daddr
< dev
->l2ad_hand
)) {
8286 * We've evicted to the target address,
8287 * or the end of the device.
8289 mutex_exit(hash_lock
);
8293 if (!HDR_HAS_L1HDR(hdr
)) {
8294 ASSERT(!HDR_L2_READING(hdr
));
8296 * This doesn't exist in the ARC. Destroy.
8297 * arc_hdr_destroy() will call list_remove()
8298 * and decrement arcstat_l2_lsize.
8300 arc_change_state(arc_anon
, hdr
, hash_lock
);
8301 arc_hdr_destroy(hdr
);
8303 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_l2c_only
);
8304 ARCSTAT_BUMP(arcstat_l2_evict_l1cached
);
8306 * Invalidate issued or about to be issued
8307 * reads, since we may be about to write
8308 * over this location.
8310 if (HDR_L2_READING(hdr
)) {
8311 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
8312 arc_hdr_set_flags(hdr
, ARC_FLAG_L2_EVICTED
);
8315 arc_hdr_l2hdr_destroy(hdr
);
8317 mutex_exit(hash_lock
);
8319 mutex_exit(&dev
->l2ad_mtx
);
8323 * Handle any abd transforms that might be required for writing to the L2ARC.
8324 * If successful, this function will always return an abd with the data
8325 * transformed as it is on disk in a new abd of asize bytes.
8328 l2arc_apply_transforms(spa_t
*spa
, arc_buf_hdr_t
*hdr
, uint64_t asize
,
8333 abd_t
*cabd
= NULL
, *eabd
= NULL
, *to_write
= hdr
->b_l1hdr
.b_pabd
;
8334 enum zio_compress compress
= HDR_GET_COMPRESS(hdr
);
8335 uint64_t psize
= HDR_GET_PSIZE(hdr
);
8336 uint64_t size
= arc_hdr_size(hdr
);
8337 boolean_t ismd
= HDR_ISTYPE_METADATA(hdr
);
8338 boolean_t bswap
= (hdr
->b_l1hdr
.b_byteswap
!= DMU_BSWAP_NUMFUNCS
);
8339 dsl_crypto_key_t
*dck
= NULL
;
8340 uint8_t mac
[ZIO_DATA_MAC_LEN
] = { 0 };
8341 boolean_t no_crypt
= B_FALSE
;
8343 ASSERT((HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
8344 !HDR_COMPRESSION_ENABLED(hdr
)) ||
8345 HDR_ENCRYPTED(hdr
) || HDR_SHARED_DATA(hdr
) || psize
!= asize
);
8346 ASSERT3U(psize
, <=, asize
);
8349 * If this data simply needs its own buffer, we simply allocate it
8350 * and copy the data. This may be done to elimiate a depedency on a
8351 * shared buffer or to reallocate the buffer to match asize.
8353 if (HDR_HAS_RABD(hdr
) && asize
!= psize
) {
8354 ASSERT3U(size
, ==, psize
);
8355 to_write
= abd_alloc_for_io(asize
, ismd
);
8356 abd_copy(to_write
, hdr
->b_crypt_hdr
.b_rabd
, size
);
8358 abd_zero_off(to_write
, size
, asize
- size
);
8362 if ((compress
== ZIO_COMPRESS_OFF
|| HDR_COMPRESSION_ENABLED(hdr
)) &&
8363 !HDR_ENCRYPTED(hdr
)) {
8364 ASSERT3U(size
, ==, psize
);
8365 to_write
= abd_alloc_for_io(asize
, ismd
);
8366 abd_copy(to_write
, hdr
->b_l1hdr
.b_pabd
, size
);
8368 abd_zero_off(to_write
, size
, asize
- size
);
8372 if (compress
!= ZIO_COMPRESS_OFF
&& !HDR_COMPRESSION_ENABLED(hdr
)) {
8373 cabd
= abd_alloc_for_io(asize
, ismd
);
8374 tmp
= abd_borrow_buf(cabd
, asize
);
8376 psize
= zio_compress_data(compress
, to_write
, tmp
, size
);
8377 ASSERT3U(psize
, <=, HDR_GET_PSIZE(hdr
));
8379 bzero((char *)tmp
+ psize
, asize
- psize
);
8380 psize
= HDR_GET_PSIZE(hdr
);
8381 abd_return_buf_copy(cabd
, tmp
, asize
);
8385 if (HDR_ENCRYPTED(hdr
)) {
8386 eabd
= abd_alloc_for_io(asize
, ismd
);
8389 * If the dataset was disowned before the buffer
8390 * made it to this point, the key to re-encrypt
8391 * it won't be available. In this case we simply
8392 * won't write the buffer to the L2ARC.
8394 ret
= spa_keystore_lookup_key(spa
, hdr
->b_crypt_hdr
.b_dsobj
,
8399 ret
= zio_do_crypt_abd(B_TRUE
, &dck
->dck_key
,
8400 hdr
->b_crypt_hdr
.b_salt
, hdr
->b_crypt_hdr
.b_ot
,
8401 hdr
->b_crypt_hdr
.b_iv
, mac
, psize
, bswap
, to_write
,
8407 abd_copy(eabd
, to_write
, psize
);
8410 abd_zero_off(eabd
, psize
, asize
- psize
);
8412 /* assert that the MAC we got here matches the one we saved */
8413 ASSERT0(bcmp(mac
, hdr
->b_crypt_hdr
.b_mac
, ZIO_DATA_MAC_LEN
));
8414 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
8416 if (to_write
== cabd
)
8423 ASSERT3P(to_write
, !=, hdr
->b_l1hdr
.b_pabd
);
8424 *abd_out
= to_write
;
8429 spa_keystore_dsl_key_rele(spa
, dck
, FTAG
);
8440 * Find and write ARC buffers to the L2ARC device.
8442 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
8443 * for reading until they have completed writing.
8444 * The headroom_boost is an in-out parameter used to maintain headroom boost
8445 * state between calls to this function.
8447 * Returns the number of bytes actually written (which may be smaller than
8448 * the delta by which the device hand has changed due to alignment).
8451 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
)
8453 arc_buf_hdr_t
*hdr
, *hdr_prev
, *head
;
8454 uint64_t write_asize
, write_psize
, write_lsize
, headroom
;
8456 l2arc_write_callback_t
*cb
;
8458 uint64_t guid
= spa_load_guid(spa
);
8460 ASSERT3P(dev
->l2ad_vdev
, !=, NULL
);
8463 write_lsize
= write_asize
= write_psize
= 0;
8465 head
= kmem_cache_alloc(hdr_l2only_cache
, KM_PUSHPAGE
);
8466 arc_hdr_set_flags(head
, ARC_FLAG_L2_WRITE_HEAD
| ARC_FLAG_HAS_L2HDR
);
8469 * Copy buffers for L2ARC writing.
8471 for (int try = 0; try < L2ARC_FEED_TYPES
; try++) {
8472 multilist_sublist_t
*mls
= l2arc_sublist_lock(try);
8473 uint64_t passed_sz
= 0;
8475 VERIFY3P(mls
, !=, NULL
);
8478 * L2ARC fast warmup.
8480 * Until the ARC is warm and starts to evict, read from the
8481 * head of the ARC lists rather than the tail.
8483 if (arc_warm
== B_FALSE
)
8484 hdr
= multilist_sublist_head(mls
);
8486 hdr
= multilist_sublist_tail(mls
);
8488 headroom
= target_sz
* l2arc_headroom
;
8489 if (zfs_compressed_arc_enabled
)
8490 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
8492 for (; hdr
; hdr
= hdr_prev
) {
8493 kmutex_t
*hash_lock
;
8494 abd_t
*to_write
= NULL
;
8496 if (arc_warm
== B_FALSE
)
8497 hdr_prev
= multilist_sublist_next(mls
, hdr
);
8499 hdr_prev
= multilist_sublist_prev(mls
, hdr
);
8501 hash_lock
= HDR_LOCK(hdr
);
8502 if (!mutex_tryenter(hash_lock
)) {
8504 * Skip this buffer rather than waiting.
8509 passed_sz
+= HDR_GET_LSIZE(hdr
);
8510 if (passed_sz
> headroom
) {
8514 mutex_exit(hash_lock
);
8518 if (!l2arc_write_eligible(guid
, hdr
)) {
8519 mutex_exit(hash_lock
);
8524 * We rely on the L1 portion of the header below, so
8525 * it's invalid for this header to have been evicted out
8526 * of the ghost cache, prior to being written out. The
8527 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8529 ASSERT(HDR_HAS_L1HDR(hdr
));
8531 ASSERT3U(HDR_GET_PSIZE(hdr
), >, 0);
8532 ASSERT3U(arc_hdr_size(hdr
), >, 0);
8533 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
||
8535 uint64_t psize
= HDR_GET_PSIZE(hdr
);
8536 uint64_t asize
= vdev_psize_to_asize(dev
->l2ad_vdev
,
8539 if ((write_asize
+ asize
) > target_sz
) {
8541 mutex_exit(hash_lock
);
8546 * We rely on the L1 portion of the header below, so
8547 * it's invalid for this header to have been evicted out
8548 * of the ghost cache, prior to being written out. The
8549 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8551 arc_hdr_set_flags(hdr
, ARC_FLAG_L2_WRITING
);
8552 ASSERT(HDR_HAS_L1HDR(hdr
));
8554 ASSERT3U(HDR_GET_PSIZE(hdr
), >, 0);
8555 ASSERT(hdr
->b_l1hdr
.b_pabd
!= NULL
||
8557 ASSERT3U(arc_hdr_size(hdr
), >, 0);
8560 * If this header has b_rabd, we can use this since it
8561 * must always match the data exactly as it exists on
8562 * disk. Otherwise, the L2ARC can normally use the
8563 * hdr's data, but if we're sharing data between the
8564 * hdr and one of its bufs, L2ARC needs its own copy of
8565 * the data so that the ZIO below can't race with the
8566 * buf consumer. To ensure that this copy will be
8567 * available for the lifetime of the ZIO and be cleaned
8568 * up afterwards, we add it to the l2arc_free_on_write
8569 * queue. If we need to apply any transforms to the
8570 * data (compression, encryption) we will also need the
8573 if (HDR_HAS_RABD(hdr
) && psize
== asize
) {
8574 to_write
= hdr
->b_crypt_hdr
.b_rabd
;
8575 } else if ((HDR_COMPRESSION_ENABLED(hdr
) ||
8576 HDR_GET_COMPRESS(hdr
) == ZIO_COMPRESS_OFF
) &&
8577 !HDR_ENCRYPTED(hdr
) && !HDR_SHARED_DATA(hdr
) &&
8579 to_write
= hdr
->b_l1hdr
.b_pabd
;
8582 arc_buf_contents_t type
= arc_buf_type(hdr
);
8584 ret
= l2arc_apply_transforms(spa
, hdr
, asize
,
8587 arc_hdr_clear_flags(hdr
,
8588 ARC_FLAG_L2_WRITING
);
8589 mutex_exit(hash_lock
);
8593 l2arc_free_abd_on_write(to_write
, asize
, type
);
8598 * Insert a dummy header on the buflist so
8599 * l2arc_write_done() can find where the
8600 * write buffers begin without searching.
8602 mutex_enter(&dev
->l2ad_mtx
);
8603 list_insert_head(&dev
->l2ad_buflist
, head
);
8604 mutex_exit(&dev
->l2ad_mtx
);
8607 sizeof (l2arc_write_callback_t
), KM_SLEEP
);
8608 cb
->l2wcb_dev
= dev
;
8609 cb
->l2wcb_head
= head
;
8610 pio
= zio_root(spa
, l2arc_write_done
, cb
,
8614 hdr
->b_l2hdr
.b_dev
= dev
;
8615 hdr
->b_l2hdr
.b_hits
= 0;
8617 hdr
->b_l2hdr
.b_daddr
= dev
->l2ad_hand
;
8618 arc_hdr_set_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
8620 mutex_enter(&dev
->l2ad_mtx
);
8621 list_insert_head(&dev
->l2ad_buflist
, hdr
);
8622 mutex_exit(&dev
->l2ad_mtx
);
8624 (void) refcount_add_many(&dev
->l2ad_alloc
,
8625 arc_hdr_size(hdr
), hdr
);
8627 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
8628 hdr
->b_l2hdr
.b_daddr
, asize
, to_write
,
8629 ZIO_CHECKSUM_OFF
, NULL
, hdr
,
8630 ZIO_PRIORITY_ASYNC_WRITE
,
8631 ZIO_FLAG_CANFAIL
, B_FALSE
);
8633 write_lsize
+= HDR_GET_LSIZE(hdr
);
8634 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
8637 write_psize
+= psize
;
8638 write_asize
+= asize
;
8639 dev
->l2ad_hand
+= asize
;
8641 mutex_exit(hash_lock
);
8643 (void) zio_nowait(wzio
);
8646 multilist_sublist_unlock(mls
);
8652 /* No buffers selected for writing? */
8654 ASSERT0(write_lsize
);
8655 ASSERT(!HDR_HAS_L1HDR(head
));
8656 kmem_cache_free(hdr_l2only_cache
, head
);
8660 ASSERT3U(write_asize
, <=, target_sz
);
8661 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
8662 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_psize
);
8663 ARCSTAT_INCR(arcstat_l2_lsize
, write_lsize
);
8664 ARCSTAT_INCR(arcstat_l2_psize
, write_psize
);
8665 vdev_space_update(dev
->l2ad_vdev
, write_psize
, 0, 0);
8668 * Bump device hand to the device start if it is approaching the end.
8669 * l2arc_evict() will already have evicted ahead for this case.
8671 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
8672 dev
->l2ad_hand
= dev
->l2ad_start
;
8673 dev
->l2ad_first
= B_FALSE
;
8676 dev
->l2ad_writing
= B_TRUE
;
8677 (void) zio_wait(pio
);
8678 dev
->l2ad_writing
= B_FALSE
;
8680 return (write_asize
);
8684 * This thread feeds the L2ARC at regular intervals. This is the beating
8685 * heart of the L2ARC.
8689 l2arc_feed_thread(void *unused
)
8694 uint64_t size
, wrote
;
8695 clock_t begin
, next
= ddi_get_lbolt();
8696 fstrans_cookie_t cookie
;
8698 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
8700 mutex_enter(&l2arc_feed_thr_lock
);
8702 cookie
= spl_fstrans_mark();
8703 while (l2arc_thread_exit
== 0) {
8704 CALLB_CPR_SAFE_BEGIN(&cpr
);
8705 (void) cv_timedwait_sig(&l2arc_feed_thr_cv
,
8706 &l2arc_feed_thr_lock
, next
);
8707 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
8708 next
= ddi_get_lbolt() + hz
;
8711 * Quick check for L2ARC devices.
8713 mutex_enter(&l2arc_dev_mtx
);
8714 if (l2arc_ndev
== 0) {
8715 mutex_exit(&l2arc_dev_mtx
);
8718 mutex_exit(&l2arc_dev_mtx
);
8719 begin
= ddi_get_lbolt();
8722 * This selects the next l2arc device to write to, and in
8723 * doing so the next spa to feed from: dev->l2ad_spa. This
8724 * will return NULL if there are now no l2arc devices or if
8725 * they are all faulted.
8727 * If a device is returned, its spa's config lock is also
8728 * held to prevent device removal. l2arc_dev_get_next()
8729 * will grab and release l2arc_dev_mtx.
8731 if ((dev
= l2arc_dev_get_next()) == NULL
)
8734 spa
= dev
->l2ad_spa
;
8735 ASSERT3P(spa
, !=, NULL
);
8738 * If the pool is read-only then force the feed thread to
8739 * sleep a little longer.
8741 if (!spa_writeable(spa
)) {
8742 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
8743 spa_config_exit(spa
, SCL_L2ARC
, dev
);
8748 * Avoid contributing to memory pressure.
8750 if (arc_reclaim_needed()) {
8751 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
8752 spa_config_exit(spa
, SCL_L2ARC
, dev
);
8756 ARCSTAT_BUMP(arcstat_l2_feeds
);
8758 size
= l2arc_write_size();
8761 * Evict L2ARC buffers that will be overwritten.
8763 l2arc_evict(dev
, size
, B_FALSE
);
8766 * Write ARC buffers.
8768 wrote
= l2arc_write_buffers(spa
, dev
, size
);
8771 * Calculate interval between writes.
8773 next
= l2arc_write_interval(begin
, size
, wrote
);
8774 spa_config_exit(spa
, SCL_L2ARC
, dev
);
8776 spl_fstrans_unmark(cookie
);
8778 l2arc_thread_exit
= 0;
8779 cv_broadcast(&l2arc_feed_thr_cv
);
8780 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
8785 l2arc_vdev_present(vdev_t
*vd
)
8789 mutex_enter(&l2arc_dev_mtx
);
8790 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
8791 dev
= list_next(l2arc_dev_list
, dev
)) {
8792 if (dev
->l2ad_vdev
== vd
)
8795 mutex_exit(&l2arc_dev_mtx
);
8797 return (dev
!= NULL
);
8801 * Add a vdev for use by the L2ARC. By this point the spa has already
8802 * validated the vdev and opened it.
8805 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
8807 l2arc_dev_t
*adddev
;
8809 ASSERT(!l2arc_vdev_present(vd
));
8812 * Create a new l2arc device entry.
8814 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
8815 adddev
->l2ad_spa
= spa
;
8816 adddev
->l2ad_vdev
= vd
;
8817 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
8818 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
8819 adddev
->l2ad_hand
= adddev
->l2ad_start
;
8820 adddev
->l2ad_first
= B_TRUE
;
8821 adddev
->l2ad_writing
= B_FALSE
;
8822 list_link_init(&adddev
->l2ad_node
);
8824 mutex_init(&adddev
->l2ad_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
8826 * This is a list of all ARC buffers that are still valid on the
8829 list_create(&adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
8830 offsetof(arc_buf_hdr_t
, b_l2hdr
.b_l2node
));
8832 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
8833 refcount_create(&adddev
->l2ad_alloc
);
8836 * Add device to global list
8838 mutex_enter(&l2arc_dev_mtx
);
8839 list_insert_head(l2arc_dev_list
, adddev
);
8840 atomic_inc_64(&l2arc_ndev
);
8841 mutex_exit(&l2arc_dev_mtx
);
8845 * Remove a vdev from the L2ARC.
8848 l2arc_remove_vdev(vdev_t
*vd
)
8850 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
8853 * Find the device by vdev
8855 mutex_enter(&l2arc_dev_mtx
);
8856 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
8857 nextdev
= list_next(l2arc_dev_list
, dev
);
8858 if (vd
== dev
->l2ad_vdev
) {
8863 ASSERT3P(remdev
, !=, NULL
);
8866 * Remove device from global list
8868 list_remove(l2arc_dev_list
, remdev
);
8869 l2arc_dev_last
= NULL
; /* may have been invalidated */
8870 atomic_dec_64(&l2arc_ndev
);
8871 mutex_exit(&l2arc_dev_mtx
);
8874 * Clear all buflists and ARC references. L2ARC device flush.
8876 l2arc_evict(remdev
, 0, B_TRUE
);
8877 list_destroy(&remdev
->l2ad_buflist
);
8878 mutex_destroy(&remdev
->l2ad_mtx
);
8879 refcount_destroy(&remdev
->l2ad_alloc
);
8880 kmem_free(remdev
, sizeof (l2arc_dev_t
));
8886 l2arc_thread_exit
= 0;
8888 l2arc_writes_sent
= 0;
8889 l2arc_writes_done
= 0;
8891 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
8892 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
8893 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
8894 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
8896 l2arc_dev_list
= &L2ARC_dev_list
;
8897 l2arc_free_on_write
= &L2ARC_free_on_write
;
8898 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
8899 offsetof(l2arc_dev_t
, l2ad_node
));
8900 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
8901 offsetof(l2arc_data_free_t
, l2df_list_node
));
8908 * This is called from dmu_fini(), which is called from spa_fini();
8909 * Because of this, we can assume that all l2arc devices have
8910 * already been removed when the pools themselves were removed.
8913 l2arc_do_free_on_write();
8915 mutex_destroy(&l2arc_feed_thr_lock
);
8916 cv_destroy(&l2arc_feed_thr_cv
);
8917 mutex_destroy(&l2arc_dev_mtx
);
8918 mutex_destroy(&l2arc_free_on_write_mtx
);
8920 list_destroy(l2arc_dev_list
);
8921 list_destroy(l2arc_free_on_write
);
8927 if (!(spa_mode_global
& FWRITE
))
8930 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
8931 TS_RUN
, defclsyspri
);
8937 if (!(spa_mode_global
& FWRITE
))
8940 mutex_enter(&l2arc_feed_thr_lock
);
8941 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
8942 l2arc_thread_exit
= 1;
8943 while (l2arc_thread_exit
!= 0)
8944 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
8945 mutex_exit(&l2arc_feed_thr_lock
);
8948 #if defined(_KERNEL) && defined(HAVE_SPL)
8949 EXPORT_SYMBOL(arc_buf_size
);
8950 EXPORT_SYMBOL(arc_write
);
8951 EXPORT_SYMBOL(arc_read
);
8952 EXPORT_SYMBOL(arc_buf_info
);
8953 EXPORT_SYMBOL(arc_getbuf_func
);
8954 EXPORT_SYMBOL(arc_add_prune_callback
);
8955 EXPORT_SYMBOL(arc_remove_prune_callback
);
8958 module_param(zfs_arc_min
, ulong
, 0644);
8959 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
8961 module_param(zfs_arc_max
, ulong
, 0644);
8962 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
8964 module_param(zfs_arc_meta_limit
, ulong
, 0644);
8965 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
8967 module_param(zfs_arc_meta_limit_percent
, ulong
, 0644);
8968 MODULE_PARM_DESC(zfs_arc_meta_limit_percent
,
8969 "Percent of arc size for arc meta limit");
8971 module_param(zfs_arc_meta_min
, ulong
, 0644);
8972 MODULE_PARM_DESC(zfs_arc_meta_min
, "Min arc metadata");
8974 module_param(zfs_arc_meta_prune
, int, 0644);
8975 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Meta objects to scan for prune");
8977 module_param(zfs_arc_meta_adjust_restarts
, int, 0644);
8978 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts
,
8979 "Limit number of restarts in arc_adjust_meta");
8981 module_param(zfs_arc_meta_strategy
, int, 0644);
8982 MODULE_PARM_DESC(zfs_arc_meta_strategy
, "Meta reclaim strategy");
8984 module_param(zfs_arc_grow_retry
, int, 0644);
8985 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
8987 module_param(zfs_arc_p_aggressive_disable
, int, 0644);
8988 MODULE_PARM_DESC(zfs_arc_p_aggressive_disable
, "disable aggressive arc_p grow");
8990 module_param(zfs_arc_p_dampener_disable
, int, 0644);
8991 MODULE_PARM_DESC(zfs_arc_p_dampener_disable
, "disable arc_p adapt dampener");
8993 module_param(zfs_arc_shrink_shift
, int, 0644);
8994 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
8996 module_param(zfs_arc_pc_percent
, uint
, 0644);
8997 MODULE_PARM_DESC(zfs_arc_pc_percent
,
8998 "Percent of pagecache to reclaim arc to");
9000 module_param(zfs_arc_p_min_shift
, int, 0644);
9001 MODULE_PARM_DESC(zfs_arc_p_min_shift
, "arc_c shift to calc min/max arc_p");
9003 module_param(zfs_arc_average_blocksize
, int, 0444);
9004 MODULE_PARM_DESC(zfs_arc_average_blocksize
, "Target average block size");
9006 module_param(zfs_compressed_arc_enabled
, int, 0644);
9007 MODULE_PARM_DESC(zfs_compressed_arc_enabled
, "Disable compressed arc buffers");
9009 module_param(zfs_arc_min_prefetch_lifespan
, int, 0644);
9010 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan
, "Min life of prefetch block");
9012 module_param(l2arc_write_max
, ulong
, 0644);
9013 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
9015 module_param(l2arc_write_boost
, ulong
, 0644);
9016 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
9018 module_param(l2arc_headroom
, ulong
, 0644);
9019 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
9021 module_param(l2arc_headroom_boost
, ulong
, 0644);
9022 MODULE_PARM_DESC(l2arc_headroom_boost
, "Compressed l2arc_headroom multiplier");
9024 module_param(l2arc_feed_secs
, ulong
, 0644);
9025 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
9027 module_param(l2arc_feed_min_ms
, ulong
, 0644);
9028 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
9030 module_param(l2arc_noprefetch
, int, 0644);
9031 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
9033 module_param(l2arc_feed_again
, int, 0644);
9034 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
9036 module_param(l2arc_norw
, int, 0644);
9037 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");
9039 module_param(zfs_arc_lotsfree_percent
, int, 0644);
9040 MODULE_PARM_DESC(zfs_arc_lotsfree_percent
,
9041 "System free memory I/O throttle in bytes");
9043 module_param(zfs_arc_sys_free
, ulong
, 0644);
9044 MODULE_PARM_DESC(zfs_arc_sys_free
, "System free memory target size in bytes");
9046 module_param(zfs_arc_dnode_limit
, ulong
, 0644);
9047 MODULE_PARM_DESC(zfs_arc_dnode_limit
, "Minimum bytes of dnodes in arc");
9049 module_param(zfs_arc_dnode_limit_percent
, ulong
, 0644);
9050 MODULE_PARM_DESC(zfs_arc_dnode_limit_percent
,
9051 "Percent of ARC meta buffers for dnodes");
9053 module_param(zfs_arc_dnode_reduce_percent
, ulong
, 0644);
9054 MODULE_PARM_DESC(zfs_arc_dnode_reduce_percent
,
9055 "Percentage of excess dnodes to try to unpin");