4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
30 * DVA-based Adjustable Replacement Cache
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal ARC algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
96 * Each ARC state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an ARC list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
102 * It as also possible to register a callback which is run when the
103 * arc_meta_limit is reached and no buffers can be safely evicted. In
104 * this case the arc user should drop a reference on some arc buffers so
105 * they can be reclaimed and the arc_meta_limit honored. For example,
106 * when using the ZPL each dentry holds a references on a znode. These
107 * dentries must be pruned before the arc buffer holding the znode can
110 * Note that the majority of the performance stats are manipulated
111 * with atomic operations.
113 * The L2ARC uses the l2ad_mtx on each vdev for the following:
115 * - L2ARC buflist creation
116 * - L2ARC buflist eviction
117 * - L2ARC write completion, which walks L2ARC buflists
118 * - ARC header destruction, as it removes from L2ARC buflists
119 * - ARC header release, as it removes from L2ARC buflists
125 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
126 * This structure can point either to a block that is still in the cache or to
127 * one that is only accessible in an L2 ARC device, or it can provide
128 * information about a block that was recently evicted. If a block is
129 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
130 * information to retrieve it from the L2ARC device. This information is
131 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
132 * that is in this state cannot access the data directly.
134 * Blocks that are actively being referenced or have not been evicted
135 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
136 * the arc_buf_hdr_t that will point to the data block in memory. A block can
137 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
138 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
139 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
141 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
142 * ability to store the physical data (b_pabd) associated with the DVA of the
143 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
144 * it will match its on-disk compression characteristics. This behavior can be
145 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
146 * compressed ARC functionality is disabled, the b_pabd will point to an
147 * uncompressed version of the on-disk data.
149 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
150 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
151 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
152 * consumer. The ARC will provide references to this data and will keep it
153 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
154 * data block and will evict any arc_buf_t that is no longer referenced. The
155 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
156 * "overhead_size" kstat.
158 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
159 * compressed form. The typical case is that consumers will want uncompressed
160 * data, and when that happens a new data buffer is allocated where the data is
161 * decompressed for them to use. Currently the only consumer who wants
162 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
163 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
164 * with the arc_buf_hdr_t.
166 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
167 * first one is owned by a compressed send consumer (and therefore references
168 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
169 * used by any other consumer (and has its own uncompressed copy of the data
184 * | b_buf +------------>+-----------+ arc_buf_t
185 * | b_pabd +-+ |b_next +---->+-----------+
186 * +-----------+ | |-----------| |b_next +-->NULL
187 * | |b_comp = T | +-----------+
188 * | |b_data +-+ |b_comp = F |
189 * | +-----------+ | |b_data +-+
190 * +->+------+ | +-----------+ |
192 * data | |<--------------+ | uncompressed
193 * +------+ compressed, | data
194 * shared +-->+------+
199 * When a consumer reads a block, the ARC must first look to see if the
200 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
201 * arc_buf_t and either copies uncompressed data into a new data buffer from an
202 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
203 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
204 * hdr is compressed and the desired compression characteristics of the
205 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
206 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
207 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
208 * be anywhere in the hdr's list.
210 * The diagram below shows an example of an uncompressed ARC hdr that is
211 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
212 * the last element in the buf list):
224 * | | arc_buf_t (shared)
225 * | b_buf +------------>+---------+ arc_buf_t
226 * | | |b_next +---->+---------+
227 * | b_pabd +-+ |---------| |b_next +-->NULL
228 * +-----------+ | | | +---------+
230 * | +---------+ | |b_data +-+
231 * +->+------+ | +---------+ |
233 * uncompressed | | | |
236 * | uncompressed | | |
239 * +---------------------------------+
241 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
242 * since the physical block is about to be rewritten. The new data contents
243 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
244 * it may compress the data before writing it to disk. The ARC will be called
245 * with the transformed data and will bcopy the transformed on-disk block into
246 * a newly allocated b_pabd. Writes are always done into buffers which have
247 * either been loaned (and hence are new and don't have other readers) or
248 * buffers which have been released (and hence have their own hdr, if there
249 * were originally other readers of the buf's original hdr). This ensures that
250 * the ARC only needs to update a single buf and its hdr after a write occurs.
252 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
253 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
254 * that when compressed ARC is enabled that the L2ARC blocks are identical
255 * to the on-disk block in the main data pool. This provides a significant
256 * advantage since the ARC can leverage the bp's checksum when reading from the
257 * L2ARC to determine if the contents are valid. However, if the compressed
258 * ARC is disabled, then the L2ARC's block must be transformed to look
259 * like the physical block in the main data pool before comparing the
260 * checksum and determining its validity.
262 * The L1ARC has a slightly different system for storing encrypted data.
263 * Raw (encrypted + possibly compressed) data has a few subtle differences from
264 * data that is just compressed. The biggest difference is that it is not
265 * possible to decrypt encrypted data (or visa versa) if the keys aren't loaded.
266 * The other difference is that encryption cannot be treated as a suggestion.
267 * If a caller would prefer compressed data, but they actually wind up with
268 * uncompressed data the worst thing that could happen is there might be a
269 * performance hit. If the caller requests encrypted data, however, we must be
270 * sure they actually get it or else secret information could be leaked. Raw
271 * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore,
272 * may have both an encrypted version and a decrypted version of its data at
273 * once. When a caller needs a raw arc_buf_t, it is allocated and the data is
274 * copied out of this header. To avoid complications with b_pabd, raw buffers
280 #include <sys/spa_impl.h>
281 #include <sys/zio_compress.h>
282 #include <sys/zio_checksum.h>
283 #include <sys/zfs_context.h>
285 #include <sys/refcount.h>
286 #include <sys/vdev.h>
287 #include <sys/vdev_impl.h>
288 #include <sys/dsl_pool.h>
289 #include <sys/zio_checksum.h>
290 #include <sys/multilist.h>
293 #include <sys/fm/fs/zfs.h>
295 #include <sys/vmsystm.h>
297 #include <sys/fs/swapnode.h>
299 #include <linux/mm_compat.h>
301 #include <sys/callb.h>
302 #include <sys/kstat.h>
303 #include <sys/dmu_tx.h>
304 #include <zfs_fletcher.h>
305 #include <sys/arc_impl.h>
306 #include <sys/trace_arc.h>
309 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
310 boolean_t arc_watch = B_FALSE;
313 static kmutex_t arc_reclaim_lock;
314 static kcondvar_t arc_reclaim_thread_cv;
315 static boolean_t arc_reclaim_thread_exit;
316 static kcondvar_t arc_reclaim_waiters_cv;
319 * The number of headers to evict in arc_evict_state_impl() before
320 * dropping the sublist lock and evicting from another sublist. A lower
321 * value means we're more likely to evict the "correct" header (i.e. the
322 * oldest header in the arc state), but comes with higher overhead
323 * (i.e. more invocations of arc_evict_state_impl()).
325 int zfs_arc_evict_batch_limit = 10;
327 /* number of seconds before growing cache again */
328 static int arc_grow_retry = 5;
330 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
331 int zfs_arc_overflow_shift = 8;
333 /* shift of arc_c for calculating both min and max arc_p */
334 static int arc_p_min_shift = 4;
336 /* log2(fraction of arc to reclaim) */
337 static int arc_shrink_shift = 7;
339 /* percent of pagecache to reclaim arc to */
341 static uint_t zfs_arc_pc_percent = 0;
345 * log2(fraction of ARC which must be free to allow growing).
346 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
347 * when reading a new block into the ARC, we will evict an equal-sized block
350 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
351 * we will still not allow it to grow.
353 int arc_no_grow_shift = 5;
357 * minimum lifespan of a prefetch block in clock ticks
358 * (initialized in arc_init())
360 static int arc_min_prefetch_ms;
361 static int arc_min_prescient_prefetch_ms;
364 * If this percent of memory is free, don't throttle.
366 int arc_lotsfree_percent = 10;
371 * The arc has filled available memory and has now warmed up.
373 static boolean_t arc_warm;
376 * log2 fraction of the zio arena to keep free.
378 int arc_zio_arena_free_shift = 2;
381 * These tunables are for performance analysis.
383 unsigned long zfs_arc_max = 0;
384 unsigned long zfs_arc_min = 0;
385 unsigned long zfs_arc_meta_limit = 0;
386 unsigned long zfs_arc_meta_min = 0;
387 unsigned long zfs_arc_dnode_limit = 0;
388 unsigned long zfs_arc_dnode_reduce_percent = 10;
389 int zfs_arc_grow_retry = 0;
390 int zfs_arc_shrink_shift = 0;
391 int zfs_arc_p_min_shift = 0;
392 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
394 int zfs_compressed_arc_enabled = B_TRUE;
397 * ARC will evict meta buffers that exceed arc_meta_limit. This
398 * tunable make arc_meta_limit adjustable for different workloads.
400 unsigned long zfs_arc_meta_limit_percent = 75;
403 * Percentage that can be consumed by dnodes of ARC meta buffers.
405 unsigned long zfs_arc_dnode_limit_percent = 10;
408 * These tunables are Linux specific
410 unsigned long zfs_arc_sys_free = 0;
411 int zfs_arc_min_prefetch_ms = 0;
412 int zfs_arc_min_prescient_prefetch_ms = 0;
413 int zfs_arc_p_dampener_disable = 1;
414 int zfs_arc_meta_prune = 10000;
415 int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED;
416 int zfs_arc_meta_adjust_restarts = 4096;
417 int zfs_arc_lotsfree_percent = 10;
420 static arc_state_t ARC_anon;
421 static arc_state_t ARC_mru;
422 static arc_state_t ARC_mru_ghost;
423 static arc_state_t ARC_mfu;
424 static arc_state_t ARC_mfu_ghost;
425 static arc_state_t ARC_l2c_only;
427 typedef struct arc_stats {
428 kstat_named_t arcstat_hits;
429 kstat_named_t arcstat_misses;
430 kstat_named_t arcstat_demand_data_hits;
431 kstat_named_t arcstat_demand_data_misses;
432 kstat_named_t arcstat_demand_metadata_hits;
433 kstat_named_t arcstat_demand_metadata_misses;
434 kstat_named_t arcstat_prefetch_data_hits;
435 kstat_named_t arcstat_prefetch_data_misses;
436 kstat_named_t arcstat_prefetch_metadata_hits;
437 kstat_named_t arcstat_prefetch_metadata_misses;
438 kstat_named_t arcstat_mru_hits;
439 kstat_named_t arcstat_mru_ghost_hits;
440 kstat_named_t arcstat_mfu_hits;
441 kstat_named_t arcstat_mfu_ghost_hits;
442 kstat_named_t arcstat_deleted;
444 * Number of buffers that could not be evicted because the hash lock
445 * was held by another thread. The lock may not necessarily be held
446 * by something using the same buffer, since hash locks are shared
447 * by multiple buffers.
449 kstat_named_t arcstat_mutex_miss;
451 * Number of buffers skipped when updating the access state due to the
452 * header having already been released after acquiring the hash lock.
454 kstat_named_t arcstat_access_skip;
456 * Number of buffers skipped because they have I/O in progress, are
457 * indirect prefetch buffers that have not lived long enough, or are
458 * not from the spa we're trying to evict from.
460 kstat_named_t arcstat_evict_skip;
462 * Number of times arc_evict_state() was unable to evict enough
463 * buffers to reach its target amount.
465 kstat_named_t arcstat_evict_not_enough;
466 kstat_named_t arcstat_evict_l2_cached;
467 kstat_named_t arcstat_evict_l2_eligible;
468 kstat_named_t arcstat_evict_l2_ineligible;
469 kstat_named_t arcstat_evict_l2_skip;
470 kstat_named_t arcstat_hash_elements;
471 kstat_named_t arcstat_hash_elements_max;
472 kstat_named_t arcstat_hash_collisions;
473 kstat_named_t arcstat_hash_chains;
474 kstat_named_t arcstat_hash_chain_max;
475 kstat_named_t arcstat_p;
476 kstat_named_t arcstat_c;
477 kstat_named_t arcstat_c_min;
478 kstat_named_t arcstat_c_max;
479 kstat_named_t arcstat_size;
481 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
482 * Note that the compressed bytes may match the uncompressed bytes
483 * if the block is either not compressed or compressed arc is disabled.
485 kstat_named_t arcstat_compressed_size;
487 * Uncompressed size of the data stored in b_pabd. If compressed
488 * arc is disabled then this value will be identical to the stat
491 kstat_named_t arcstat_uncompressed_size;
493 * Number of bytes stored in all the arc_buf_t's. This is classified
494 * as "overhead" since this data is typically short-lived and will
495 * be evicted from the arc when it becomes unreferenced unless the
496 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
497 * values have been set (see comment in dbuf.c for more information).
499 kstat_named_t arcstat_overhead_size;
501 * Number of bytes consumed by internal ARC structures necessary
502 * for tracking purposes; these structures are not actually
503 * backed by ARC buffers. This includes arc_buf_hdr_t structures
504 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
505 * caches), and arc_buf_t structures (allocated via arc_buf_t
508 kstat_named_t arcstat_hdr_size;
510 * Number of bytes consumed by ARC buffers of type equal to
511 * ARC_BUFC_DATA. This is generally consumed by buffers backing
512 * on disk user data (e.g. plain file contents).
514 kstat_named_t arcstat_data_size;
516 * Number of bytes consumed by ARC buffers of type equal to
517 * ARC_BUFC_METADATA. This is generally consumed by buffers
518 * backing on disk data that is used for internal ZFS
519 * structures (e.g. ZAP, dnode, indirect blocks, etc).
521 kstat_named_t arcstat_metadata_size;
523 * Number of bytes consumed by dmu_buf_impl_t objects.
525 kstat_named_t arcstat_dbuf_size;
527 * Number of bytes consumed by dnode_t objects.
529 kstat_named_t arcstat_dnode_size;
531 * Number of bytes consumed by bonus buffers.
533 kstat_named_t arcstat_bonus_size;
535 * Total number of bytes consumed by ARC buffers residing in the
536 * arc_anon state. This includes *all* buffers in the arc_anon
537 * state; e.g. data, metadata, evictable, and unevictable buffers
538 * are all included in this value.
540 kstat_named_t arcstat_anon_size;
542 * Number of bytes consumed by ARC buffers that meet the
543 * following criteria: backing buffers of type ARC_BUFC_DATA,
544 * residing in the arc_anon state, and are eligible for eviction
545 * (e.g. have no outstanding holds on the buffer).
547 kstat_named_t arcstat_anon_evictable_data;
549 * Number of bytes consumed by ARC buffers that meet the
550 * following criteria: backing buffers of type ARC_BUFC_METADATA,
551 * residing in the arc_anon state, and are eligible for eviction
552 * (e.g. have no outstanding holds on the buffer).
554 kstat_named_t arcstat_anon_evictable_metadata;
556 * Total number of bytes consumed by ARC buffers residing in the
557 * arc_mru state. This includes *all* buffers in the arc_mru
558 * state; e.g. data, metadata, evictable, and unevictable buffers
559 * are all included in this value.
561 kstat_named_t arcstat_mru_size;
563 * Number of bytes consumed by ARC buffers that meet the
564 * following criteria: backing buffers of type ARC_BUFC_DATA,
565 * residing in the arc_mru state, and are eligible for eviction
566 * (e.g. have no outstanding holds on the buffer).
568 kstat_named_t arcstat_mru_evictable_data;
570 * Number of bytes consumed by ARC buffers that meet the
571 * following criteria: backing buffers of type ARC_BUFC_METADATA,
572 * residing in the arc_mru state, and are eligible for eviction
573 * (e.g. have no outstanding holds on the buffer).
575 kstat_named_t arcstat_mru_evictable_metadata;
577 * Total number of bytes that *would have been* consumed by ARC
578 * buffers in the arc_mru_ghost state. The key thing to note
579 * here, is the fact that this size doesn't actually indicate
580 * RAM consumption. The ghost lists only consist of headers and
581 * don't actually have ARC buffers linked off of these headers.
582 * Thus, *if* the headers had associated ARC buffers, these
583 * buffers *would have* consumed this number of bytes.
585 kstat_named_t arcstat_mru_ghost_size;
587 * Number of bytes that *would have been* consumed by ARC
588 * buffers that are eligible for eviction, of type
589 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
591 kstat_named_t arcstat_mru_ghost_evictable_data;
593 * Number of bytes that *would have been* consumed by ARC
594 * buffers that are eligible for eviction, of type
595 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
597 kstat_named_t arcstat_mru_ghost_evictable_metadata;
599 * Total number of bytes consumed by ARC buffers residing in the
600 * arc_mfu state. This includes *all* buffers in the arc_mfu
601 * state; e.g. data, metadata, evictable, and unevictable buffers
602 * are all included in this value.
604 kstat_named_t arcstat_mfu_size;
606 * Number of bytes consumed by ARC buffers that are eligible for
607 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
610 kstat_named_t arcstat_mfu_evictable_data;
612 * Number of bytes consumed by ARC buffers that are eligible for
613 * eviction, of type ARC_BUFC_METADATA, and reside in the
616 kstat_named_t arcstat_mfu_evictable_metadata;
618 * Total number of bytes that *would have been* consumed by ARC
619 * buffers in the arc_mfu_ghost state. See the comment above
620 * arcstat_mru_ghost_size for more details.
622 kstat_named_t arcstat_mfu_ghost_size;
624 * Number of bytes that *would have been* consumed by ARC
625 * buffers that are eligible for eviction, of type
626 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
628 kstat_named_t arcstat_mfu_ghost_evictable_data;
630 * Number of bytes that *would have been* consumed by ARC
631 * buffers that are eligible for eviction, of type
632 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
634 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
635 kstat_named_t arcstat_l2_hits;
636 kstat_named_t arcstat_l2_misses;
637 kstat_named_t arcstat_l2_feeds;
638 kstat_named_t arcstat_l2_rw_clash;
639 kstat_named_t arcstat_l2_read_bytes;
640 kstat_named_t arcstat_l2_write_bytes;
641 kstat_named_t arcstat_l2_writes_sent;
642 kstat_named_t arcstat_l2_writes_done;
643 kstat_named_t arcstat_l2_writes_error;
644 kstat_named_t arcstat_l2_writes_lock_retry;
645 kstat_named_t arcstat_l2_evict_lock_retry;
646 kstat_named_t arcstat_l2_evict_reading;
647 kstat_named_t arcstat_l2_evict_l1cached;
648 kstat_named_t arcstat_l2_free_on_write;
649 kstat_named_t arcstat_l2_abort_lowmem;
650 kstat_named_t arcstat_l2_cksum_bad;
651 kstat_named_t arcstat_l2_io_error;
652 kstat_named_t arcstat_l2_lsize;
653 kstat_named_t arcstat_l2_psize;
654 kstat_named_t arcstat_l2_hdr_size;
655 kstat_named_t arcstat_memory_throttle_count;
656 kstat_named_t arcstat_memory_direct_count;
657 kstat_named_t arcstat_memory_indirect_count;
658 kstat_named_t arcstat_memory_all_bytes;
659 kstat_named_t arcstat_memory_free_bytes;
660 kstat_named_t arcstat_memory_available_bytes;
661 kstat_named_t arcstat_no_grow;
662 kstat_named_t arcstat_tempreserve;
663 kstat_named_t arcstat_loaned_bytes;
664 kstat_named_t arcstat_prune;
665 kstat_named_t arcstat_meta_used;
666 kstat_named_t arcstat_meta_limit;
667 kstat_named_t arcstat_dnode_limit;
668 kstat_named_t arcstat_meta_max;
669 kstat_named_t arcstat_meta_min;
670 kstat_named_t arcstat_async_upgrade_sync;
671 kstat_named_t arcstat_demand_hit_predictive_prefetch;
672 kstat_named_t arcstat_demand_hit_prescient_prefetch;
673 kstat_named_t arcstat_need_free;
674 kstat_named_t arcstat_sys_free;
675 kstat_named_t arcstat_raw_size;
678 static arc_stats_t arc_stats = {
679 { "hits", KSTAT_DATA_UINT64 },
680 { "misses", KSTAT_DATA_UINT64 },
681 { "demand_data_hits", KSTAT_DATA_UINT64 },
682 { "demand_data_misses", KSTAT_DATA_UINT64 },
683 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
684 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
685 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
686 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
687 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
688 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
689 { "mru_hits", KSTAT_DATA_UINT64 },
690 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
691 { "mfu_hits", KSTAT_DATA_UINT64 },
692 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
693 { "deleted", KSTAT_DATA_UINT64 },
694 { "mutex_miss", KSTAT_DATA_UINT64 },
695 { "access_skip", KSTAT_DATA_UINT64 },
696 { "evict_skip", KSTAT_DATA_UINT64 },
697 { "evict_not_enough", KSTAT_DATA_UINT64 },
698 { "evict_l2_cached", KSTAT_DATA_UINT64 },
699 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
700 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
701 { "evict_l2_skip", KSTAT_DATA_UINT64 },
702 { "hash_elements", KSTAT_DATA_UINT64 },
703 { "hash_elements_max", KSTAT_DATA_UINT64 },
704 { "hash_collisions", KSTAT_DATA_UINT64 },
705 { "hash_chains", KSTAT_DATA_UINT64 },
706 { "hash_chain_max", KSTAT_DATA_UINT64 },
707 { "p", KSTAT_DATA_UINT64 },
708 { "c", KSTAT_DATA_UINT64 },
709 { "c_min", KSTAT_DATA_UINT64 },
710 { "c_max", KSTAT_DATA_UINT64 },
711 { "size", KSTAT_DATA_UINT64 },
712 { "compressed_size", KSTAT_DATA_UINT64 },
713 { "uncompressed_size", KSTAT_DATA_UINT64 },
714 { "overhead_size", KSTAT_DATA_UINT64 },
715 { "hdr_size", KSTAT_DATA_UINT64 },
716 { "data_size", KSTAT_DATA_UINT64 },
717 { "metadata_size", KSTAT_DATA_UINT64 },
718 { "dbuf_size", KSTAT_DATA_UINT64 },
719 { "dnode_size", KSTAT_DATA_UINT64 },
720 { "bonus_size", KSTAT_DATA_UINT64 },
721 { "anon_size", KSTAT_DATA_UINT64 },
722 { "anon_evictable_data", KSTAT_DATA_UINT64 },
723 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
724 { "mru_size", KSTAT_DATA_UINT64 },
725 { "mru_evictable_data", KSTAT_DATA_UINT64 },
726 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
727 { "mru_ghost_size", KSTAT_DATA_UINT64 },
728 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
729 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
730 { "mfu_size", KSTAT_DATA_UINT64 },
731 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
732 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
733 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
734 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
735 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
736 { "l2_hits", KSTAT_DATA_UINT64 },
737 { "l2_misses", KSTAT_DATA_UINT64 },
738 { "l2_feeds", KSTAT_DATA_UINT64 },
739 { "l2_rw_clash", KSTAT_DATA_UINT64 },
740 { "l2_read_bytes", KSTAT_DATA_UINT64 },
741 { "l2_write_bytes", KSTAT_DATA_UINT64 },
742 { "l2_writes_sent", KSTAT_DATA_UINT64 },
743 { "l2_writes_done", KSTAT_DATA_UINT64 },
744 { "l2_writes_error", KSTAT_DATA_UINT64 },
745 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
746 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
747 { "l2_evict_reading", KSTAT_DATA_UINT64 },
748 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
749 { "l2_free_on_write", KSTAT_DATA_UINT64 },
750 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
751 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
752 { "l2_io_error", KSTAT_DATA_UINT64 },
753 { "l2_size", KSTAT_DATA_UINT64 },
754 { "l2_asize", KSTAT_DATA_UINT64 },
755 { "l2_hdr_size", KSTAT_DATA_UINT64 },
756 { "memory_throttle_count", KSTAT_DATA_UINT64 },
757 { "memory_direct_count", KSTAT_DATA_UINT64 },
758 { "memory_indirect_count", KSTAT_DATA_UINT64 },
759 { "memory_all_bytes", KSTAT_DATA_UINT64 },
760 { "memory_free_bytes", KSTAT_DATA_UINT64 },
761 { "memory_available_bytes", KSTAT_DATA_INT64 },
762 { "arc_no_grow", KSTAT_DATA_UINT64 },
763 { "arc_tempreserve", KSTAT_DATA_UINT64 },
764 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
765 { "arc_prune", KSTAT_DATA_UINT64 },
766 { "arc_meta_used", KSTAT_DATA_UINT64 },
767 { "arc_meta_limit", KSTAT_DATA_UINT64 },
768 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
769 { "arc_meta_max", KSTAT_DATA_UINT64 },
770 { "arc_meta_min", KSTAT_DATA_UINT64 },
771 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
772 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
773 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
774 { "arc_need_free", KSTAT_DATA_UINT64 },
775 { "arc_sys_free", KSTAT_DATA_UINT64 },
776 { "arc_raw_size", KSTAT_DATA_UINT64 }
779 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
781 #define ARCSTAT_INCR(stat, val) \
782 atomic_add_64(&arc_stats.stat.value.ui64, (val))
784 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
785 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
787 #define ARCSTAT_MAX(stat, val) { \
789 while ((val) > (m = arc_stats.stat.value.ui64) && \
790 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
794 #define ARCSTAT_MAXSTAT(stat) \
795 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
798 * We define a macro to allow ARC hits/misses to be easily broken down by
799 * two separate conditions, giving a total of four different subtypes for
800 * each of hits and misses (so eight statistics total).
802 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
805 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
807 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
811 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
813 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
818 static arc_state_t *arc_anon;
819 static arc_state_t *arc_mru;
820 static arc_state_t *arc_mru_ghost;
821 static arc_state_t *arc_mfu;
822 static arc_state_t *arc_mfu_ghost;
823 static arc_state_t *arc_l2c_only;
826 * There are several ARC variables that are critical to export as kstats --
827 * but we don't want to have to grovel around in the kstat whenever we wish to
828 * manipulate them. For these variables, we therefore define them to be in
829 * terms of the statistic variable. This assures that we are not introducing
830 * the possibility of inconsistency by having shadow copies of the variables,
831 * while still allowing the code to be readable.
833 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
834 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
835 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
836 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
837 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
838 #define arc_no_grow ARCSTAT(arcstat_no_grow) /* do not grow cache size */
839 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
840 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
841 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
842 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
843 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
844 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
845 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
846 #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
847 #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */
848 #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
849 #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */
850 #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */
852 /* size of all b_rabd's in entire arc */
853 #define arc_raw_size ARCSTAT(arcstat_raw_size)
854 /* compressed size of entire arc */
855 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
856 /* uncompressed size of entire arc */
857 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
858 /* number of bytes in the arc from arc_buf_t's */
859 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
861 static list_t arc_prune_list;
862 static kmutex_t arc_prune_mtx;
863 static taskq_t *arc_prune_taskq;
865 #define GHOST_STATE(state) \
866 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
867 (state) == arc_l2c_only)
869 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
870 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
871 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
872 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
873 #define HDR_PRESCIENT_PREFETCH(hdr) \
874 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
875 #define HDR_COMPRESSION_ENABLED(hdr) \
876 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
878 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
879 #define HDR_L2_READING(hdr) \
880 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
881 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
882 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
883 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
884 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
885 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
886 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
887 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
889 #define HDR_ISTYPE_METADATA(hdr) \
890 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
891 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
893 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
894 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
895 #define HDR_HAS_RABD(hdr) \
896 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
897 (hdr)->b_crypt_hdr.b_rabd != NULL)
898 #define HDR_ENCRYPTED(hdr) \
899 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
900 #define HDR_AUTHENTICATED(hdr) \
901 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
903 /* For storing compression mode in b_flags */
904 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
906 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
907 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
908 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
909 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
911 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
912 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
913 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
914 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
920 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
921 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
922 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
925 * Hash table routines
928 #define HT_LOCK_ALIGN 64
929 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
934 unsigned char pad[HT_LOCK_PAD];
938 #define BUF_LOCKS 8192
939 typedef struct buf_hash_table {
941 arc_buf_hdr_t **ht_table;
942 struct ht_lock ht_locks[BUF_LOCKS];
945 static buf_hash_table_t buf_hash_table;
947 #define BUF_HASH_INDEX(spa, dva, birth) \
948 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
949 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
950 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
951 #define HDR_LOCK(hdr) \
952 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
954 uint64_t zfs_crc64_table[256];
960 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
961 #define L2ARC_HEADROOM 2 /* num of writes */
964 * If we discover during ARC scan any buffers to be compressed, we boost
965 * our headroom for the next scanning cycle by this percentage multiple.
967 #define L2ARC_HEADROOM_BOOST 200
968 #define L2ARC_FEED_SECS 1 /* caching interval secs */
969 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
972 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
973 * and each of the state has two types: data and metadata.
975 #define L2ARC_FEED_TYPES 4
977 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
978 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
980 /* L2ARC Performance Tunables */
981 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
982 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
983 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
984 unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
985 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
986 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
987 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
988 int l2arc_feed_again = B_TRUE; /* turbo warmup */
989 int l2arc_norw = B_FALSE; /* no reads during writes */
994 static list_t L2ARC_dev_list; /* device list */
995 static list_t *l2arc_dev_list; /* device list pointer */
996 static kmutex_t l2arc_dev_mtx; /* device list mutex */
997 static l2arc_dev_t *l2arc_dev_last; /* last device used */
998 static list_t L2ARC_free_on_write; /* free after write buf list */
999 static list_t *l2arc_free_on_write; /* free after write list ptr */
1000 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1001 static uint64_t l2arc_ndev; /* number of devices */
1003 typedef struct l2arc_read_callback {
1004 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1005 blkptr_t l2rcb_bp; /* original blkptr */
1006 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1007 int l2rcb_flags; /* original flags */
1008 abd_t *l2rcb_abd; /* temporary buffer */
1009 } l2arc_read_callback_t;
1011 typedef struct l2arc_data_free {
1012 /* protected by l2arc_free_on_write_mtx */
1015 arc_buf_contents_t l2df_type;
1016 list_node_t l2df_list_node;
1017 } l2arc_data_free_t;
1019 typedef enum arc_fill_flags {
1020 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
1021 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
1022 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
1023 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
1024 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
1027 static kmutex_t l2arc_feed_thr_lock;
1028 static kcondvar_t l2arc_feed_thr_cv;
1029 static uint8_t l2arc_thread_exit;
1031 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1032 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1033 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1034 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1035 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1036 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1037 static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t);
1038 static void arc_hdr_alloc_abd(arc_buf_hdr_t *, boolean_t);
1039 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1040 static boolean_t arc_is_overflowing(void);
1041 static void arc_buf_watch(arc_buf_t *);
1042 static void arc_tuning_update(void);
1043 static void arc_prune_async(int64_t);
1044 static uint64_t arc_all_memory(void);
1046 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1047 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1048 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1049 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1051 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1052 static void l2arc_read_done(zio_t *);
1055 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1057 uint8_t *vdva = (uint8_t *)dva;
1058 uint64_t crc = -1ULL;
1061 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1063 for (i = 0; i < sizeof (dva_t); i++)
1064 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1066 crc ^= (spa>>8) ^ birth;
1071 #define HDR_EMPTY(hdr) \
1072 ((hdr)->b_dva.dva_word[0] == 0 && \
1073 (hdr)->b_dva.dva_word[1] == 0)
1075 #define HDR_EQUAL(spa, dva, birth, hdr) \
1076 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1077 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1078 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1081 buf_discard_identity(arc_buf_hdr_t *hdr)
1083 hdr->b_dva.dva_word[0] = 0;
1084 hdr->b_dva.dva_word[1] = 0;
1088 static arc_buf_hdr_t *
1089 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1091 const dva_t *dva = BP_IDENTITY(bp);
1092 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1093 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1094 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1097 mutex_enter(hash_lock);
1098 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1099 hdr = hdr->b_hash_next) {
1100 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1105 mutex_exit(hash_lock);
1111 * Insert an entry into the hash table. If there is already an element
1112 * equal to elem in the hash table, then the already existing element
1113 * will be returned and the new element will not be inserted.
1114 * Otherwise returns NULL.
1115 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1117 static arc_buf_hdr_t *
1118 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1120 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1121 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1122 arc_buf_hdr_t *fhdr;
1125 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1126 ASSERT(hdr->b_birth != 0);
1127 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1129 if (lockp != NULL) {
1131 mutex_enter(hash_lock);
1133 ASSERT(MUTEX_HELD(hash_lock));
1136 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1137 fhdr = fhdr->b_hash_next, i++) {
1138 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1142 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1143 buf_hash_table.ht_table[idx] = hdr;
1144 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1146 /* collect some hash table performance data */
1148 ARCSTAT_BUMP(arcstat_hash_collisions);
1150 ARCSTAT_BUMP(arcstat_hash_chains);
1152 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1155 ARCSTAT_BUMP(arcstat_hash_elements);
1156 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1162 buf_hash_remove(arc_buf_hdr_t *hdr)
1164 arc_buf_hdr_t *fhdr, **hdrp;
1165 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1167 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1168 ASSERT(HDR_IN_HASH_TABLE(hdr));
1170 hdrp = &buf_hash_table.ht_table[idx];
1171 while ((fhdr = *hdrp) != hdr) {
1172 ASSERT3P(fhdr, !=, NULL);
1173 hdrp = &fhdr->b_hash_next;
1175 *hdrp = hdr->b_hash_next;
1176 hdr->b_hash_next = NULL;
1177 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1179 /* collect some hash table performance data */
1180 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1182 if (buf_hash_table.ht_table[idx] &&
1183 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1184 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1188 * Global data structures and functions for the buf kmem cache.
1191 static kmem_cache_t *hdr_full_cache;
1192 static kmem_cache_t *hdr_full_crypt_cache;
1193 static kmem_cache_t *hdr_l2only_cache;
1194 static kmem_cache_t *buf_cache;
1201 #if defined(_KERNEL) && defined(HAVE_SPL)
1203 * Large allocations which do not require contiguous pages
1204 * should be using vmem_free() in the linux kernel\
1206 vmem_free(buf_hash_table.ht_table,
1207 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1209 kmem_free(buf_hash_table.ht_table,
1210 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1212 for (i = 0; i < BUF_LOCKS; i++)
1213 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1214 kmem_cache_destroy(hdr_full_cache);
1215 kmem_cache_destroy(hdr_full_crypt_cache);
1216 kmem_cache_destroy(hdr_l2only_cache);
1217 kmem_cache_destroy(buf_cache);
1221 * Constructor callback - called when the cache is empty
1222 * and a new buf is requested.
1226 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1228 arc_buf_hdr_t *hdr = vbuf;
1230 bzero(hdr, HDR_FULL_SIZE);
1231 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1232 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1233 refcount_create(&hdr->b_l1hdr.b_refcnt);
1234 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1235 list_link_init(&hdr->b_l1hdr.b_arc_node);
1236 list_link_init(&hdr->b_l2hdr.b_l2node);
1237 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1238 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1245 hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
1247 arc_buf_hdr_t *hdr = vbuf;
1249 hdr_full_cons(vbuf, unused, kmflag);
1250 bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr));
1251 arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1258 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1260 arc_buf_hdr_t *hdr = vbuf;
1262 bzero(hdr, HDR_L2ONLY_SIZE);
1263 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1270 buf_cons(void *vbuf, void *unused, int kmflag)
1272 arc_buf_t *buf = vbuf;
1274 bzero(buf, sizeof (arc_buf_t));
1275 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1276 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1282 * Destructor callback - called when a cached buf is
1283 * no longer required.
1287 hdr_full_dest(void *vbuf, void *unused)
1289 arc_buf_hdr_t *hdr = vbuf;
1291 ASSERT(HDR_EMPTY(hdr));
1292 cv_destroy(&hdr->b_l1hdr.b_cv);
1293 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1294 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1295 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1296 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1301 hdr_full_crypt_dest(void *vbuf, void *unused)
1303 arc_buf_hdr_t *hdr = vbuf;
1305 hdr_full_dest(vbuf, unused);
1306 arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1311 hdr_l2only_dest(void *vbuf, void *unused)
1313 ASSERTV(arc_buf_hdr_t *hdr = vbuf);
1315 ASSERT(HDR_EMPTY(hdr));
1316 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1321 buf_dest(void *vbuf, void *unused)
1323 arc_buf_t *buf = vbuf;
1325 mutex_destroy(&buf->b_evict_lock);
1326 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1330 * Reclaim callback -- invoked when memory is low.
1334 hdr_recl(void *unused)
1336 dprintf("hdr_recl called\n");
1338 * umem calls the reclaim func when we destroy the buf cache,
1339 * which is after we do arc_fini().
1342 cv_signal(&arc_reclaim_thread_cv);
1348 uint64_t *ct = NULL;
1349 uint64_t hsize = 1ULL << 12;
1353 * The hash table is big enough to fill all of physical memory
1354 * with an average block size of zfs_arc_average_blocksize (default 8K).
1355 * By default, the table will take up
1356 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1358 while (hsize * zfs_arc_average_blocksize < arc_all_memory())
1361 buf_hash_table.ht_mask = hsize - 1;
1362 #if defined(_KERNEL) && defined(HAVE_SPL)
1364 * Large allocations which do not require contiguous pages
1365 * should be using vmem_alloc() in the linux kernel
1367 buf_hash_table.ht_table =
1368 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
1370 buf_hash_table.ht_table =
1371 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1373 if (buf_hash_table.ht_table == NULL) {
1374 ASSERT(hsize > (1ULL << 8));
1379 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1380 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1381 hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
1382 HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
1383 hdr_recl, NULL, NULL, 0);
1384 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1385 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1387 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1388 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1390 for (i = 0; i < 256; i++)
1391 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1392 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1394 for (i = 0; i < BUF_LOCKS; i++) {
1395 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1396 NULL, MUTEX_DEFAULT, NULL);
1400 #define ARC_MINTIME (hz>>4) /* 62 ms */
1403 * This is the size that the buf occupies in memory. If the buf is compressed,
1404 * it will correspond to the compressed size. You should use this method of
1405 * getting the buf size unless you explicitly need the logical size.
1408 arc_buf_size(arc_buf_t *buf)
1410 return (ARC_BUF_COMPRESSED(buf) ?
1411 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1415 arc_buf_lsize(arc_buf_t *buf)
1417 return (HDR_GET_LSIZE(buf->b_hdr));
1421 * This function will return B_TRUE if the buffer is encrypted in memory.
1422 * This buffer can be decrypted by calling arc_untransform().
1425 arc_is_encrypted(arc_buf_t *buf)
1427 return (ARC_BUF_ENCRYPTED(buf) != 0);
1431 * Returns B_TRUE if the buffer represents data that has not had its MAC
1435 arc_is_unauthenticated(arc_buf_t *buf)
1437 return (HDR_NOAUTH(buf->b_hdr) != 0);
1441 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1442 uint8_t *iv, uint8_t *mac)
1444 arc_buf_hdr_t *hdr = buf->b_hdr;
1446 ASSERT(HDR_PROTECTED(hdr));
1448 bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
1449 bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
1450 bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
1451 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1452 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1456 * Indicates how this buffer is compressed in memory. If it is not compressed
1457 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1458 * arc_untransform() as long as it is also unencrypted.
1461 arc_get_compression(arc_buf_t *buf)
1463 return (ARC_BUF_COMPRESSED(buf) ?
1464 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1468 * Return the compression algorithm used to store this data in the ARC. If ARC
1469 * compression is enabled or this is an encrypted block, this will be the same
1470 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1472 static inline enum zio_compress
1473 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1475 return (HDR_COMPRESSION_ENABLED(hdr) ?
1476 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1479 static inline boolean_t
1480 arc_buf_is_shared(arc_buf_t *buf)
1482 boolean_t shared = (buf->b_data != NULL &&
1483 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1484 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1485 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1486 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1487 IMPLY(shared, ARC_BUF_SHARED(buf));
1488 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1491 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1492 * already being shared" requirement prevents us from doing that.
1499 * Free the checksum associated with this header. If there is no checksum, this
1503 arc_cksum_free(arc_buf_hdr_t *hdr)
1505 ASSERT(HDR_HAS_L1HDR(hdr));
1507 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1508 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1509 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1510 hdr->b_l1hdr.b_freeze_cksum = NULL;
1512 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1516 * Return true iff at least one of the bufs on hdr is not compressed.
1517 * Encrypted buffers count as compressed.
1520 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1522 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1523 if (!ARC_BUF_COMPRESSED(b)) {
1532 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1533 * matches the checksum that is stored in the hdr. If there is no checksum,
1534 * or if the buf is compressed, this is a no-op.
1537 arc_cksum_verify(arc_buf_t *buf)
1539 arc_buf_hdr_t *hdr = buf->b_hdr;
1542 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1545 if (ARC_BUF_COMPRESSED(buf)) {
1546 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1547 arc_hdr_has_uncompressed_buf(hdr));
1551 ASSERT(HDR_HAS_L1HDR(hdr));
1553 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1554 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1555 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1559 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1560 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1561 panic("buffer modified while frozen!");
1562 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1566 * This function makes the assumption that data stored in the L2ARC
1567 * will be transformed exactly as it is in the main pool. Because of
1568 * this we can verify the checksum against the reading process's bp.
1571 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1573 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1574 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1577 * Block pointers always store the checksum for the logical data.
1578 * If the block pointer has the gang bit set, then the checksum
1579 * it represents is for the reconstituted data and not for an
1580 * individual gang member. The zio pipeline, however, must be able to
1581 * determine the checksum of each of the gang constituents so it
1582 * treats the checksum comparison differently than what we need
1583 * for l2arc blocks. This prevents us from using the
1584 * zio_checksum_error() interface directly. Instead we must call the
1585 * zio_checksum_error_impl() so that we can ensure the checksum is
1586 * generated using the correct checksum algorithm and accounts for the
1587 * logical I/O size and not just a gang fragment.
1589 return (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1590 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1591 zio->io_offset, NULL) == 0);
1595 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1596 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1597 * isn't modified later on. If buf is compressed or there is already a checksum
1598 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1601 arc_cksum_compute(arc_buf_t *buf)
1603 arc_buf_hdr_t *hdr = buf->b_hdr;
1605 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1608 ASSERT(HDR_HAS_L1HDR(hdr));
1610 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1611 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1612 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1613 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1615 } else if (ARC_BUF_COMPRESSED(buf)) {
1616 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1620 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1621 ASSERT(!ARC_BUF_COMPRESSED(buf));
1622 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1624 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1625 hdr->b_l1hdr.b_freeze_cksum);
1626 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1632 arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
1634 panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr);
1640 arc_buf_unwatch(arc_buf_t *buf)
1644 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1645 PROT_READ | PROT_WRITE));
1652 arc_buf_watch(arc_buf_t *buf)
1656 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1661 static arc_buf_contents_t
1662 arc_buf_type(arc_buf_hdr_t *hdr)
1664 arc_buf_contents_t type;
1665 if (HDR_ISTYPE_METADATA(hdr)) {
1666 type = ARC_BUFC_METADATA;
1668 type = ARC_BUFC_DATA;
1670 VERIFY3U(hdr->b_type, ==, type);
1675 arc_is_metadata(arc_buf_t *buf)
1677 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1681 arc_bufc_to_flags(arc_buf_contents_t type)
1685 /* metadata field is 0 if buffer contains normal data */
1687 case ARC_BUFC_METADATA:
1688 return (ARC_FLAG_BUFC_METADATA);
1692 panic("undefined ARC buffer type!");
1693 return ((uint32_t)-1);
1697 arc_buf_thaw(arc_buf_t *buf)
1699 arc_buf_hdr_t *hdr = buf->b_hdr;
1701 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1702 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1704 arc_cksum_verify(buf);
1707 * Compressed buffers do not manipulate the b_freeze_cksum or
1708 * allocate b_thawed.
1710 if (ARC_BUF_COMPRESSED(buf)) {
1711 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1712 arc_hdr_has_uncompressed_buf(hdr));
1716 ASSERT(HDR_HAS_L1HDR(hdr));
1717 arc_cksum_free(hdr);
1718 arc_buf_unwatch(buf);
1722 arc_buf_freeze(arc_buf_t *buf)
1724 arc_buf_hdr_t *hdr = buf->b_hdr;
1725 kmutex_t *hash_lock;
1727 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1730 if (ARC_BUF_COMPRESSED(buf)) {
1731 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1732 arc_hdr_has_uncompressed_buf(hdr));
1736 hash_lock = HDR_LOCK(hdr);
1737 mutex_enter(hash_lock);
1739 ASSERT(HDR_HAS_L1HDR(hdr));
1740 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
1741 hdr->b_l1hdr.b_state == arc_anon);
1742 arc_cksum_compute(buf);
1743 mutex_exit(hash_lock);
1747 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1748 * the following functions should be used to ensure that the flags are
1749 * updated in a thread-safe way. When manipulating the flags either
1750 * the hash_lock must be held or the hdr must be undiscoverable. This
1751 * ensures that we're not racing with any other threads when updating
1755 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1757 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1758 hdr->b_flags |= flags;
1762 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1764 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1765 hdr->b_flags &= ~flags;
1769 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1770 * done in a special way since we have to clear and set bits
1771 * at the same time. Consumers that wish to set the compression bits
1772 * must use this function to ensure that the flags are updated in
1773 * thread-safe manner.
1776 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1778 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1781 * Holes and embedded blocks will always have a psize = 0 so
1782 * we ignore the compression of the blkptr and set the
1783 * want to uncompress them. Mark them as uncompressed.
1785 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1786 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1787 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1789 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1790 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1793 HDR_SET_COMPRESS(hdr, cmp);
1794 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1798 * Looks for another buf on the same hdr which has the data decompressed, copies
1799 * from it, and returns true. If no such buf exists, returns false.
1802 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1804 arc_buf_hdr_t *hdr = buf->b_hdr;
1805 boolean_t copied = B_FALSE;
1807 ASSERT(HDR_HAS_L1HDR(hdr));
1808 ASSERT3P(buf->b_data, !=, NULL);
1809 ASSERT(!ARC_BUF_COMPRESSED(buf));
1811 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1812 from = from->b_next) {
1813 /* can't use our own data buffer */
1818 if (!ARC_BUF_COMPRESSED(from)) {
1819 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1826 * There were no decompressed bufs, so there should not be a
1827 * checksum on the hdr either.
1829 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1835 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1838 arc_hdr_size(arc_buf_hdr_t *hdr)
1842 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1843 HDR_GET_PSIZE(hdr) > 0) {
1844 size = HDR_GET_PSIZE(hdr);
1846 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1847 size = HDR_GET_LSIZE(hdr);
1853 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1857 uint64_t lsize = HDR_GET_LSIZE(hdr);
1858 uint64_t psize = HDR_GET_PSIZE(hdr);
1859 void *tmpbuf = NULL;
1860 abd_t *abd = hdr->b_l1hdr.b_pabd;
1862 ASSERT(HDR_LOCK(hdr) == NULL || MUTEX_HELD(HDR_LOCK(hdr)));
1863 ASSERT(HDR_AUTHENTICATED(hdr));
1864 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1867 * The MAC is calculated on the compressed data that is stored on disk.
1868 * However, if compressed arc is disabled we will only have the
1869 * decompressed data available to us now. Compress it into a temporary
1870 * abd so we can verify the MAC. The performance overhead of this will
1871 * be relatively low, since most objects in an encrypted objset will
1872 * be encrypted (instead of authenticated) anyway.
1874 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1875 !HDR_COMPRESSION_ENABLED(hdr)) {
1876 tmpbuf = zio_buf_alloc(lsize);
1877 abd = abd_get_from_buf(tmpbuf, lsize);
1878 abd_take_ownership_of_buf(abd, B_TRUE);
1880 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1881 hdr->b_l1hdr.b_pabd, tmpbuf, lsize);
1882 ASSERT3U(csize, <=, psize);
1883 abd_zero_off(abd, csize, psize - csize);
1887 * Authentication is best effort. We authenticate whenever the key is
1888 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1890 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1891 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1892 ASSERT3U(lsize, ==, psize);
1893 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1894 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1896 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1897 hdr->b_crypt_hdr.b_mac);
1901 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1902 else if (ret != ENOENT)
1918 * This function will take a header that only has raw encrypted data in
1919 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1920 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1921 * also decompress the data.
1924 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1927 dsl_crypto_key_t *dck = NULL;
1930 boolean_t no_crypt = B_FALSE;
1931 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1933 ASSERT(HDR_LOCK(hdr) == NULL || MUTEX_HELD(HDR_LOCK(hdr)));
1934 ASSERT(HDR_ENCRYPTED(hdr));
1936 arc_hdr_alloc_abd(hdr, B_FALSE);
1939 * We must be careful to use the passed-in dsobj value here and
1940 * not the value in b_dsobj. b_dsobj is meant to be a best guess for
1941 * the L2ARC, which has the luxury of being able to fail without real
1942 * consequences (the data simply won't make it to the L2ARC). In
1943 * reality, the dsobj stored in the header may belong to a dataset
1944 * that has been unmounted or otherwise disowned, meaning the key
1945 * won't be accessible via that dsobj anymore.
1947 ret = spa_keystore_lookup_key(spa, dsobj, FTAG, &dck);
1949 ret = SET_ERROR(EACCES);
1953 ret = zio_do_crypt_abd(B_FALSE, &dck->dck_key,
1954 hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_ot,
1955 hdr->b_crypt_hdr.b_iv, hdr->b_crypt_hdr.b_mac,
1956 HDR_GET_PSIZE(hdr), bswap, hdr->b_l1hdr.b_pabd,
1957 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1962 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1963 HDR_GET_PSIZE(hdr));
1967 * If this header has disabled arc compression but the b_pabd is
1968 * compressed after decrypting it, we need to decompress the newly
1971 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1972 !HDR_COMPRESSION_ENABLED(hdr)) {
1974 * We want to make sure that we are correctly honoring the
1975 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1976 * and then loan a buffer from it, rather than allocating a
1977 * linear buffer and wrapping it in an abd later.
1979 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
1980 tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
1982 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1983 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
1984 HDR_GET_LSIZE(hdr));
1986 abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
1990 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
1991 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1992 arc_hdr_size(hdr), hdr);
1993 hdr->b_l1hdr.b_pabd = cabd;
1996 spa_keystore_dsl_key_rele(spa, dck, FTAG);
2001 arc_hdr_free_abd(hdr, B_FALSE);
2003 spa_keystore_dsl_key_rele(spa, dck, FTAG);
2005 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
2011 * This function is called during arc_buf_fill() to prepare the header's
2012 * abd plaintext pointer for use. This involves authenticated protected
2013 * data and decrypting encrypted data into the plaintext abd.
2016 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
2017 uint64_t dsobj, boolean_t noauth)
2021 ASSERT(HDR_PROTECTED(hdr));
2023 if (hash_lock != NULL)
2024 mutex_enter(hash_lock);
2026 if (HDR_NOAUTH(hdr) && !noauth) {
2028 * The caller requested authenticated data but our data has
2029 * not been authenticated yet. Verify the MAC now if we can.
2031 ret = arc_hdr_authenticate(hdr, spa, dsobj);
2034 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
2036 * If we only have the encrypted version of the data, but the
2037 * unencrypted version was requested we take this opportunity
2038 * to store the decrypted version in the header for future use.
2040 ret = arc_hdr_decrypt(hdr, spa, dsobj);
2045 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2047 if (hash_lock != NULL)
2048 mutex_exit(hash_lock);
2053 if (hash_lock != NULL)
2054 mutex_exit(hash_lock);
2060 * This function is used by the dbuf code to decrypt bonus buffers in place.
2061 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
2062 * block, so we use the hash lock here to protect against concurrent calls to
2066 arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock)
2068 arc_buf_hdr_t *hdr = buf->b_hdr;
2070 ASSERT(HDR_ENCRYPTED(hdr));
2071 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2072 ASSERT(HDR_LOCK(hdr) == NULL || MUTEX_HELD(HDR_LOCK(hdr)));
2073 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2075 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
2077 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
2078 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2079 hdr->b_crypt_hdr.b_ebufcnt -= 1;
2083 * Given a buf that has a data buffer attached to it, this function will
2084 * efficiently fill the buf with data of the specified compression setting from
2085 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2086 * are already sharing a data buf, no copy is performed.
2088 * If the buf is marked as compressed but uncompressed data was requested, this
2089 * will allocate a new data buffer for the buf, remove that flag, and fill the
2090 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2091 * uncompressed data, and (since we haven't added support for it yet) if you
2092 * want compressed data your buf must already be marked as compressed and have
2093 * the correct-sized data buffer.
2096 arc_buf_fill(arc_buf_t *buf, spa_t *spa, uint64_t dsobj, arc_fill_flags_t flags)
2099 arc_buf_hdr_t *hdr = buf->b_hdr;
2100 boolean_t hdr_compressed =
2101 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
2102 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
2103 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
2104 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2105 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
2107 ASSERT3P(buf->b_data, !=, NULL);
2108 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
2109 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2110 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
2111 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
2112 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
2113 IMPLY(encrypted, !ARC_BUF_SHARED(buf));
2116 * If the caller wanted encrypted data we just need to copy it from
2117 * b_rabd and potentially byteswap it. We won't be able to do any
2118 * further transforms on it.
2121 ASSERT(HDR_HAS_RABD(hdr));
2122 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
2123 HDR_GET_PSIZE(hdr));
2128 * Adjust encrypted and authenticated headers to accomodate the
2129 * request if needed.
2131 if (HDR_PROTECTED(hdr)) {
2132 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
2133 dsobj, !!(flags & ARC_FILL_NOAUTH));
2139 * There is a special case here for dnode blocks which are
2140 * decrypting their bonus buffers. These blocks may request to
2141 * be decrypted in-place. This is necessary because there may
2142 * be many dnodes pointing into this buffer and there is
2143 * currently no method to synchronize replacing the backing
2144 * b_data buffer and updating all of the pointers. Here we use
2145 * the hash lock to ensure there are no races. If the need
2146 * arises for other types to be decrypted in-place, they must
2147 * add handling here as well.
2149 if ((flags & ARC_FILL_IN_PLACE) != 0) {
2150 ASSERT(!hdr_compressed);
2151 ASSERT(!compressed);
2154 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
2155 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2157 if (hash_lock != NULL)
2158 mutex_enter(hash_lock);
2159 arc_buf_untransform_in_place(buf, hash_lock);
2160 if (hash_lock != NULL)
2161 mutex_exit(hash_lock);
2163 /* Compute the hdr's checksum if necessary */
2164 arc_cksum_compute(buf);
2170 if (hdr_compressed == compressed) {
2171 if (!arc_buf_is_shared(buf)) {
2172 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2176 ASSERT(hdr_compressed);
2177 ASSERT(!compressed);
2178 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2181 * If the buf is sharing its data with the hdr, unlink it and
2182 * allocate a new data buffer for the buf.
2184 if (arc_buf_is_shared(buf)) {
2185 ASSERT(ARC_BUF_COMPRESSED(buf));
2187 /* We need to give the buf it's own b_data */
2188 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2190 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2191 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2193 /* Previously overhead was 0; just add new overhead */
2194 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2195 } else if (ARC_BUF_COMPRESSED(buf)) {
2196 /* We need to reallocate the buf's b_data */
2197 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2200 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2202 /* We increased the size of b_data; update overhead */
2203 ARCSTAT_INCR(arcstat_overhead_size,
2204 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2208 * Regardless of the buf's previous compression settings, it
2209 * should not be compressed at the end of this function.
2211 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2214 * Try copying the data from another buf which already has a
2215 * decompressed version. If that's not possible, it's time to
2216 * bite the bullet and decompress the data from the hdr.
2218 if (arc_buf_try_copy_decompressed_data(buf)) {
2219 /* Skip byteswapping and checksumming (already done) */
2220 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2223 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2224 hdr->b_l1hdr.b_pabd, buf->b_data,
2225 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2228 * Absent hardware errors or software bugs, this should
2229 * be impossible, but log it anyway so we can debug it.
2233 "hdr %p, compress %d, psize %d, lsize %d",
2234 hdr, arc_hdr_get_compress(hdr),
2235 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2236 return (SET_ERROR(EIO));
2242 /* Byteswap the buf's data if necessary */
2243 if (bswap != DMU_BSWAP_NUMFUNCS) {
2244 ASSERT(!HDR_SHARED_DATA(hdr));
2245 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2246 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2249 /* Compute the hdr's checksum if necessary */
2250 arc_cksum_compute(buf);
2256 * If this function is being called to decrypt an encrypted buffer or verify an
2257 * authenticated one, the key must be loaded and a mapping must be made
2258 * available in the keystore via spa_keystore_create_mapping() or one of its
2262 arc_untransform(arc_buf_t *buf, spa_t *spa, uint64_t dsobj, boolean_t in_place)
2264 arc_fill_flags_t flags = 0;
2267 flags |= ARC_FILL_IN_PLACE;
2269 return (arc_buf_fill(buf, spa, dsobj, flags));
2273 * Increment the amount of evictable space in the arc_state_t's refcount.
2274 * We account for the space used by the hdr and the arc buf individually
2275 * so that we can add and remove them from the refcount individually.
2278 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2280 arc_buf_contents_t type = arc_buf_type(hdr);
2282 ASSERT(HDR_HAS_L1HDR(hdr));
2284 if (GHOST_STATE(state)) {
2285 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2286 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2287 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2288 ASSERT(!HDR_HAS_RABD(hdr));
2289 (void) refcount_add_many(&state->arcs_esize[type],
2290 HDR_GET_LSIZE(hdr), hdr);
2294 ASSERT(!GHOST_STATE(state));
2295 if (hdr->b_l1hdr.b_pabd != NULL) {
2296 (void) refcount_add_many(&state->arcs_esize[type],
2297 arc_hdr_size(hdr), hdr);
2299 if (HDR_HAS_RABD(hdr)) {
2300 (void) refcount_add_many(&state->arcs_esize[type],
2301 HDR_GET_PSIZE(hdr), hdr);
2304 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2305 buf = buf->b_next) {
2306 if (arc_buf_is_shared(buf))
2308 (void) refcount_add_many(&state->arcs_esize[type],
2309 arc_buf_size(buf), buf);
2314 * Decrement the amount of evictable space in the arc_state_t's refcount.
2315 * We account for the space used by the hdr and the arc buf individually
2316 * so that we can add and remove them from the refcount individually.
2319 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2321 arc_buf_contents_t type = arc_buf_type(hdr);
2323 ASSERT(HDR_HAS_L1HDR(hdr));
2325 if (GHOST_STATE(state)) {
2326 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2327 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2328 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2329 ASSERT(!HDR_HAS_RABD(hdr));
2330 (void) refcount_remove_many(&state->arcs_esize[type],
2331 HDR_GET_LSIZE(hdr), hdr);
2335 ASSERT(!GHOST_STATE(state));
2336 if (hdr->b_l1hdr.b_pabd != NULL) {
2337 (void) refcount_remove_many(&state->arcs_esize[type],
2338 arc_hdr_size(hdr), hdr);
2340 if (HDR_HAS_RABD(hdr)) {
2341 (void) refcount_remove_many(&state->arcs_esize[type],
2342 HDR_GET_PSIZE(hdr), hdr);
2345 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2346 buf = buf->b_next) {
2347 if (arc_buf_is_shared(buf))
2349 (void) refcount_remove_many(&state->arcs_esize[type],
2350 arc_buf_size(buf), buf);
2355 * Add a reference to this hdr indicating that someone is actively
2356 * referencing that memory. When the refcount transitions from 0 to 1,
2357 * we remove it from the respective arc_state_t list to indicate that
2358 * it is not evictable.
2361 add_reference(arc_buf_hdr_t *hdr, void *tag)
2365 ASSERT(HDR_HAS_L1HDR(hdr));
2366 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2367 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2368 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2369 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2372 state = hdr->b_l1hdr.b_state;
2374 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2375 (state != arc_anon)) {
2376 /* We don't use the L2-only state list. */
2377 if (state != arc_l2c_only) {
2378 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2380 arc_evictable_space_decrement(hdr, state);
2382 /* remove the prefetch flag if we get a reference */
2383 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2388 * Remove a reference from this hdr. When the reference transitions from
2389 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2390 * list making it eligible for eviction.
2393 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2396 arc_state_t *state = hdr->b_l1hdr.b_state;
2398 ASSERT(HDR_HAS_L1HDR(hdr));
2399 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2400 ASSERT(!GHOST_STATE(state));
2403 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2404 * check to prevent usage of the arc_l2c_only list.
2406 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2407 (state != arc_anon)) {
2408 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2409 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2410 arc_evictable_space_increment(hdr, state);
2416 * Returns detailed information about a specific arc buffer. When the
2417 * state_index argument is set the function will calculate the arc header
2418 * list position for its arc state. Since this requires a linear traversal
2419 * callers are strongly encourage not to do this. However, it can be helpful
2420 * for targeted analysis so the functionality is provided.
2423 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2425 arc_buf_hdr_t *hdr = ab->b_hdr;
2426 l1arc_buf_hdr_t *l1hdr = NULL;
2427 l2arc_buf_hdr_t *l2hdr = NULL;
2428 arc_state_t *state = NULL;
2430 memset(abi, 0, sizeof (arc_buf_info_t));
2435 abi->abi_flags = hdr->b_flags;
2437 if (HDR_HAS_L1HDR(hdr)) {
2438 l1hdr = &hdr->b_l1hdr;
2439 state = l1hdr->b_state;
2441 if (HDR_HAS_L2HDR(hdr))
2442 l2hdr = &hdr->b_l2hdr;
2445 abi->abi_bufcnt = l1hdr->b_bufcnt;
2446 abi->abi_access = l1hdr->b_arc_access;
2447 abi->abi_mru_hits = l1hdr->b_mru_hits;
2448 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2449 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2450 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2451 abi->abi_holds = refcount_count(&l1hdr->b_refcnt);
2455 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2456 abi->abi_l2arc_hits = l2hdr->b_hits;
2459 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2460 abi->abi_state_contents = arc_buf_type(hdr);
2461 abi->abi_size = arc_hdr_size(hdr);
2465 * Move the supplied buffer to the indicated state. The hash lock
2466 * for the buffer must be held by the caller.
2469 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2470 kmutex_t *hash_lock)
2472 arc_state_t *old_state;
2475 boolean_t update_old, update_new;
2476 arc_buf_contents_t buftype = arc_buf_type(hdr);
2479 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2480 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2481 * L1 hdr doesn't always exist when we change state to arc_anon before
2482 * destroying a header, in which case reallocating to add the L1 hdr is
2485 if (HDR_HAS_L1HDR(hdr)) {
2486 old_state = hdr->b_l1hdr.b_state;
2487 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2488 bufcnt = hdr->b_l1hdr.b_bufcnt;
2489 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
2492 old_state = arc_l2c_only;
2495 update_old = B_FALSE;
2497 update_new = update_old;
2499 ASSERT(MUTEX_HELD(hash_lock));
2500 ASSERT3P(new_state, !=, old_state);
2501 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2502 ASSERT(old_state != arc_anon || bufcnt <= 1);
2505 * If this buffer is evictable, transfer it from the
2506 * old state list to the new state list.
2509 if (old_state != arc_anon && old_state != arc_l2c_only) {
2510 ASSERT(HDR_HAS_L1HDR(hdr));
2511 multilist_remove(old_state->arcs_list[buftype], hdr);
2513 if (GHOST_STATE(old_state)) {
2515 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2516 update_old = B_TRUE;
2518 arc_evictable_space_decrement(hdr, old_state);
2520 if (new_state != arc_anon && new_state != arc_l2c_only) {
2522 * An L1 header always exists here, since if we're
2523 * moving to some L1-cached state (i.e. not l2c_only or
2524 * anonymous), we realloc the header to add an L1hdr
2527 ASSERT(HDR_HAS_L1HDR(hdr));
2528 multilist_insert(new_state->arcs_list[buftype], hdr);
2530 if (GHOST_STATE(new_state)) {
2532 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2533 update_new = B_TRUE;
2535 arc_evictable_space_increment(hdr, new_state);
2539 ASSERT(!HDR_EMPTY(hdr));
2540 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2541 buf_hash_remove(hdr);
2543 /* adjust state sizes (ignore arc_l2c_only) */
2545 if (update_new && new_state != arc_l2c_only) {
2546 ASSERT(HDR_HAS_L1HDR(hdr));
2547 if (GHOST_STATE(new_state)) {
2551 * When moving a header to a ghost state, we first
2552 * remove all arc buffers. Thus, we'll have a
2553 * bufcnt of zero, and no arc buffer to use for
2554 * the reference. As a result, we use the arc
2555 * header pointer for the reference.
2557 (void) refcount_add_many(&new_state->arcs_size,
2558 HDR_GET_LSIZE(hdr), hdr);
2559 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2560 ASSERT(!HDR_HAS_RABD(hdr));
2562 uint32_t buffers = 0;
2565 * Each individual buffer holds a unique reference,
2566 * thus we must remove each of these references one
2569 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2570 buf = buf->b_next) {
2571 ASSERT3U(bufcnt, !=, 0);
2575 * When the arc_buf_t is sharing the data
2576 * block with the hdr, the owner of the
2577 * reference belongs to the hdr. Only
2578 * add to the refcount if the arc_buf_t is
2581 if (arc_buf_is_shared(buf))
2584 (void) refcount_add_many(&new_state->arcs_size,
2585 arc_buf_size(buf), buf);
2587 ASSERT3U(bufcnt, ==, buffers);
2589 if (hdr->b_l1hdr.b_pabd != NULL) {
2590 (void) refcount_add_many(&new_state->arcs_size,
2591 arc_hdr_size(hdr), hdr);
2594 if (HDR_HAS_RABD(hdr)) {
2595 (void) refcount_add_many(&new_state->arcs_size,
2596 HDR_GET_PSIZE(hdr), hdr);
2601 if (update_old && old_state != arc_l2c_only) {
2602 ASSERT(HDR_HAS_L1HDR(hdr));
2603 if (GHOST_STATE(old_state)) {
2605 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2606 ASSERT(!HDR_HAS_RABD(hdr));
2609 * When moving a header off of a ghost state,
2610 * the header will not contain any arc buffers.
2611 * We use the arc header pointer for the reference
2612 * which is exactly what we did when we put the
2613 * header on the ghost state.
2616 (void) refcount_remove_many(&old_state->arcs_size,
2617 HDR_GET_LSIZE(hdr), hdr);
2619 uint32_t buffers = 0;
2622 * Each individual buffer holds a unique reference,
2623 * thus we must remove each of these references one
2626 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2627 buf = buf->b_next) {
2628 ASSERT3U(bufcnt, !=, 0);
2632 * When the arc_buf_t is sharing the data
2633 * block with the hdr, the owner of the
2634 * reference belongs to the hdr. Only
2635 * add to the refcount if the arc_buf_t is
2638 if (arc_buf_is_shared(buf))
2641 (void) refcount_remove_many(
2642 &old_state->arcs_size, arc_buf_size(buf),
2645 ASSERT3U(bufcnt, ==, buffers);
2646 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2649 if (hdr->b_l1hdr.b_pabd != NULL) {
2650 (void) refcount_remove_many(
2651 &old_state->arcs_size, arc_hdr_size(hdr),
2655 if (HDR_HAS_RABD(hdr)) {
2656 (void) refcount_remove_many(
2657 &old_state->arcs_size, HDR_GET_PSIZE(hdr),
2663 if (HDR_HAS_L1HDR(hdr))
2664 hdr->b_l1hdr.b_state = new_state;
2667 * L2 headers should never be on the L2 state list since they don't
2668 * have L1 headers allocated.
2670 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2671 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2675 arc_space_consume(uint64_t space, arc_space_type_t type)
2677 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2682 case ARC_SPACE_DATA:
2683 ARCSTAT_INCR(arcstat_data_size, space);
2685 case ARC_SPACE_META:
2686 ARCSTAT_INCR(arcstat_metadata_size, space);
2688 case ARC_SPACE_BONUS:
2689 ARCSTAT_INCR(arcstat_bonus_size, space);
2691 case ARC_SPACE_DNODE:
2692 ARCSTAT_INCR(arcstat_dnode_size, space);
2694 case ARC_SPACE_DBUF:
2695 ARCSTAT_INCR(arcstat_dbuf_size, space);
2697 case ARC_SPACE_HDRS:
2698 ARCSTAT_INCR(arcstat_hdr_size, space);
2700 case ARC_SPACE_L2HDRS:
2701 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
2705 if (type != ARC_SPACE_DATA)
2706 ARCSTAT_INCR(arcstat_meta_used, space);
2708 atomic_add_64(&arc_size, space);
2712 arc_space_return(uint64_t space, arc_space_type_t type)
2714 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2719 case ARC_SPACE_DATA:
2720 ARCSTAT_INCR(arcstat_data_size, -space);
2722 case ARC_SPACE_META:
2723 ARCSTAT_INCR(arcstat_metadata_size, -space);
2725 case ARC_SPACE_BONUS:
2726 ARCSTAT_INCR(arcstat_bonus_size, -space);
2728 case ARC_SPACE_DNODE:
2729 ARCSTAT_INCR(arcstat_dnode_size, -space);
2731 case ARC_SPACE_DBUF:
2732 ARCSTAT_INCR(arcstat_dbuf_size, -space);
2734 case ARC_SPACE_HDRS:
2735 ARCSTAT_INCR(arcstat_hdr_size, -space);
2737 case ARC_SPACE_L2HDRS:
2738 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
2742 if (type != ARC_SPACE_DATA) {
2743 ASSERT(arc_meta_used >= space);
2744 if (arc_meta_max < arc_meta_used)
2745 arc_meta_max = arc_meta_used;
2746 ARCSTAT_INCR(arcstat_meta_used, -space);
2749 ASSERT(arc_size >= space);
2750 atomic_add_64(&arc_size, -space);
2754 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2755 * with the hdr's b_pabd.
2758 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2761 * The criteria for sharing a hdr's data are:
2762 * 1. the buffer is not encrypted
2763 * 2. the hdr's compression matches the buf's compression
2764 * 3. the hdr doesn't need to be byteswapped
2765 * 4. the hdr isn't already being shared
2766 * 5. the buf is either compressed or it is the last buf in the hdr list
2768 * Criterion #5 maintains the invariant that shared uncompressed
2769 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2770 * might ask, "if a compressed buf is allocated first, won't that be the
2771 * last thing in the list?", but in that case it's impossible to create
2772 * a shared uncompressed buf anyway (because the hdr must be compressed
2773 * to have the compressed buf). You might also think that #3 is
2774 * sufficient to make this guarantee, however it's possible
2775 * (specifically in the rare L2ARC write race mentioned in
2776 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2777 * is sharable, but wasn't at the time of its allocation. Rather than
2778 * allow a new shared uncompressed buf to be created and then shuffle
2779 * the list around to make it the last element, this simply disallows
2780 * sharing if the new buf isn't the first to be added.
2782 ASSERT3P(buf->b_hdr, ==, hdr);
2783 boolean_t hdr_compressed =
2784 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF;
2785 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2786 return (!ARC_BUF_ENCRYPTED(buf) &&
2787 buf_compressed == hdr_compressed &&
2788 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2789 !HDR_SHARED_DATA(hdr) &&
2790 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2794 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2795 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2796 * copy was made successfully, or an error code otherwise.
2799 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj, void *tag,
2800 boolean_t encrypted, boolean_t compressed, boolean_t noauth,
2801 boolean_t fill, arc_buf_t **ret)
2804 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2806 ASSERT(HDR_HAS_L1HDR(hdr));
2807 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2808 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2809 hdr->b_type == ARC_BUFC_METADATA);
2810 ASSERT3P(ret, !=, NULL);
2811 ASSERT3P(*ret, ==, NULL);
2812 IMPLY(encrypted, compressed);
2814 hdr->b_l1hdr.b_mru_hits = 0;
2815 hdr->b_l1hdr.b_mru_ghost_hits = 0;
2816 hdr->b_l1hdr.b_mfu_hits = 0;
2817 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
2818 hdr->b_l1hdr.b_l2_hits = 0;
2820 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2823 buf->b_next = hdr->b_l1hdr.b_buf;
2826 add_reference(hdr, tag);
2829 * We're about to change the hdr's b_flags. We must either
2830 * hold the hash_lock or be undiscoverable.
2832 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2835 * Only honor requests for compressed bufs if the hdr is actually
2836 * compressed. This must be overriden if the buffer is encrypted since
2837 * encrypted buffers cannot be decompressed.
2840 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2841 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2842 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2843 } else if (compressed &&
2844 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2845 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2846 flags |= ARC_FILL_COMPRESSED;
2851 flags |= ARC_FILL_NOAUTH;
2855 * If the hdr's data can be shared then we share the data buffer and
2856 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2857 * allocate a new buffer to store the buf's data.
2859 * There are two additional restrictions here because we're sharing
2860 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2861 * actively involved in an L2ARC write, because if this buf is used by
2862 * an arc_write() then the hdr's data buffer will be released when the
2863 * write completes, even though the L2ARC write might still be using it.
2864 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2865 * need to be ABD-aware.
2867 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2868 hdr->b_l1hdr.b_pabd != NULL && abd_is_linear(hdr->b_l1hdr.b_pabd);
2870 /* Set up b_data and sharing */
2872 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2873 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2874 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2877 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2878 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2880 VERIFY3P(buf->b_data, !=, NULL);
2882 hdr->b_l1hdr.b_buf = buf;
2883 hdr->b_l1hdr.b_bufcnt += 1;
2885 hdr->b_crypt_hdr.b_ebufcnt += 1;
2888 * If the user wants the data from the hdr, we need to either copy or
2889 * decompress the data.
2892 return (arc_buf_fill(buf, spa, dsobj, flags));
2898 static char *arc_onloan_tag = "onloan";
2901 arc_loaned_bytes_update(int64_t delta)
2903 atomic_add_64(&arc_loaned_bytes, delta);
2905 /* assert that it did not wrap around */
2906 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2910 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2911 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2912 * buffers must be returned to the arc before they can be used by the DMU or
2916 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2918 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2919 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2921 arc_loaned_bytes_update(size);
2927 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2928 enum zio_compress compression_type)
2930 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2931 psize, lsize, compression_type);
2933 arc_loaned_bytes_update(psize);
2939 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2940 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2941 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2942 enum zio_compress compression_type)
2944 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2945 byteorder, salt, iv, mac, ot, psize, lsize, compression_type);
2947 atomic_add_64(&arc_loaned_bytes, psize);
2953 * Return a loaned arc buffer to the arc.
2956 arc_return_buf(arc_buf_t *buf, void *tag)
2958 arc_buf_hdr_t *hdr = buf->b_hdr;
2960 ASSERT3P(buf->b_data, !=, NULL);
2961 ASSERT(HDR_HAS_L1HDR(hdr));
2962 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2963 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2965 arc_loaned_bytes_update(-arc_buf_size(buf));
2968 /* Detach an arc_buf from a dbuf (tag) */
2970 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2972 arc_buf_hdr_t *hdr = buf->b_hdr;
2974 ASSERT3P(buf->b_data, !=, NULL);
2975 ASSERT(HDR_HAS_L1HDR(hdr));
2976 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2977 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2979 arc_loaned_bytes_update(arc_buf_size(buf));
2983 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2985 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2988 df->l2df_size = size;
2989 df->l2df_type = type;
2990 mutex_enter(&l2arc_free_on_write_mtx);
2991 list_insert_head(l2arc_free_on_write, df);
2992 mutex_exit(&l2arc_free_on_write_mtx);
2996 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
2998 arc_state_t *state = hdr->b_l1hdr.b_state;
2999 arc_buf_contents_t type = arc_buf_type(hdr);
3000 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3002 /* protected by hash lock, if in the hash table */
3003 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3004 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3005 ASSERT(state != arc_anon && state != arc_l2c_only);
3007 (void) refcount_remove_many(&state->arcs_esize[type],
3010 (void) refcount_remove_many(&state->arcs_size, size, hdr);
3011 if (type == ARC_BUFC_METADATA) {
3012 arc_space_return(size, ARC_SPACE_META);
3014 ASSERT(type == ARC_BUFC_DATA);
3015 arc_space_return(size, ARC_SPACE_DATA);
3019 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
3021 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
3026 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
3027 * data buffer, we transfer the refcount ownership to the hdr and update
3028 * the appropriate kstats.
3031 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3033 ASSERT(arc_can_share(hdr, buf));
3034 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3035 ASSERT(!ARC_BUF_ENCRYPTED(buf));
3036 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3039 * Start sharing the data buffer. We transfer the
3040 * refcount ownership to the hdr since it always owns
3041 * the refcount whenever an arc_buf_t is shared.
3043 refcount_transfer_ownership(&hdr->b_l1hdr.b_state->arcs_size, buf, hdr);
3044 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
3045 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
3046 HDR_ISTYPE_METADATA(hdr));
3047 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3048 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3051 * Since we've transferred ownership to the hdr we need
3052 * to increment its compressed and uncompressed kstats and
3053 * decrement the overhead size.
3055 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3056 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3057 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3061 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3063 ASSERT(arc_buf_is_shared(buf));
3064 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3065 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3068 * We are no longer sharing this buffer so we need
3069 * to transfer its ownership to the rightful owner.
3071 refcount_transfer_ownership(&hdr->b_l1hdr.b_state->arcs_size, hdr, buf);
3072 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3073 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3074 abd_put(hdr->b_l1hdr.b_pabd);
3075 hdr->b_l1hdr.b_pabd = NULL;
3076 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3079 * Since the buffer is no longer shared between
3080 * the arc buf and the hdr, count it as overhead.
3082 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3083 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3084 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3088 * Remove an arc_buf_t from the hdr's buf list and return the last
3089 * arc_buf_t on the list. If no buffers remain on the list then return
3093 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3095 ASSERT(HDR_HAS_L1HDR(hdr));
3096 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3098 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3099 arc_buf_t *lastbuf = NULL;
3102 * Remove the buf from the hdr list and locate the last
3103 * remaining buffer on the list.
3105 while (*bufp != NULL) {
3107 *bufp = buf->b_next;
3110 * If we've removed a buffer in the middle of
3111 * the list then update the lastbuf and update
3114 if (*bufp != NULL) {
3116 bufp = &(*bufp)->b_next;
3120 ASSERT3P(lastbuf, !=, buf);
3121 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3122 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3123 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3129 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3133 arc_buf_destroy_impl(arc_buf_t *buf)
3135 arc_buf_hdr_t *hdr = buf->b_hdr;
3138 * Free up the data associated with the buf but only if we're not
3139 * sharing this with the hdr. If we are sharing it with the hdr, the
3140 * hdr is responsible for doing the free.
3142 if (buf->b_data != NULL) {
3144 * We're about to change the hdr's b_flags. We must either
3145 * hold the hash_lock or be undiscoverable.
3147 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3149 arc_cksum_verify(buf);
3150 arc_buf_unwatch(buf);
3152 if (arc_buf_is_shared(buf)) {
3153 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3155 uint64_t size = arc_buf_size(buf);
3156 arc_free_data_buf(hdr, buf->b_data, size, buf);
3157 ARCSTAT_INCR(arcstat_overhead_size, -size);
3161 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3162 hdr->b_l1hdr.b_bufcnt -= 1;
3164 if (ARC_BUF_ENCRYPTED(buf)) {
3165 hdr->b_crypt_hdr.b_ebufcnt -= 1;
3168 * If we have no more encrypted buffers and we've
3169 * already gotten a copy of the decrypted data we can
3170 * free b_rabd to save some space.
3172 if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
3173 HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
3174 !HDR_IO_IN_PROGRESS(hdr)) {
3175 arc_hdr_free_abd(hdr, B_TRUE);
3180 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3182 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3184 * If the current arc_buf_t is sharing its data buffer with the
3185 * hdr, then reassign the hdr's b_pabd to share it with the new
3186 * buffer at the end of the list. The shared buffer is always
3187 * the last one on the hdr's buffer list.
3189 * There is an equivalent case for compressed bufs, but since
3190 * they aren't guaranteed to be the last buf in the list and
3191 * that is an exceedingly rare case, we just allow that space be
3192 * wasted temporarily. We must also be careful not to share
3193 * encrypted buffers, since they cannot be shared.
3195 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
3196 /* Only one buf can be shared at once */
3197 VERIFY(!arc_buf_is_shared(lastbuf));
3198 /* hdr is uncompressed so can't have compressed buf */
3199 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3201 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3202 arc_hdr_free_abd(hdr, B_FALSE);
3205 * We must setup a new shared block between the
3206 * last buffer and the hdr. The data would have
3207 * been allocated by the arc buf so we need to transfer
3208 * ownership to the hdr since it's now being shared.
3210 arc_share_buf(hdr, lastbuf);
3212 } else if (HDR_SHARED_DATA(hdr)) {
3214 * Uncompressed shared buffers are always at the end
3215 * of the list. Compressed buffers don't have the
3216 * same requirements. This makes it hard to
3217 * simply assert that the lastbuf is shared so
3218 * we rely on the hdr's compression flags to determine
3219 * if we have a compressed, shared buffer.
3221 ASSERT3P(lastbuf, !=, NULL);
3222 ASSERT(arc_buf_is_shared(lastbuf) ||
3223 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3227 * Free the checksum if we're removing the last uncompressed buf from
3230 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3231 arc_cksum_free(hdr);
3234 /* clean up the buf */
3236 kmem_cache_free(buf_cache, buf);
3240 arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, boolean_t alloc_rdata)
3244 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3245 ASSERT(HDR_HAS_L1HDR(hdr));
3246 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3247 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3250 size = HDR_GET_PSIZE(hdr);
3251 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3252 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr);
3253 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3254 ARCSTAT_INCR(arcstat_raw_size, size);
3256 size = arc_hdr_size(hdr);
3257 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3258 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr);
3259 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3262 ARCSTAT_INCR(arcstat_compressed_size, size);
3263 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3267 arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3269 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3271 ASSERT(HDR_HAS_L1HDR(hdr));
3272 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3273 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3276 * If the hdr is currently being written to the l2arc then
3277 * we defer freeing the data by adding it to the l2arc_free_on_write
3278 * list. The l2arc will free the data once it's finished
3279 * writing it to the l2arc device.
3281 if (HDR_L2_WRITING(hdr)) {
3282 arc_hdr_free_on_write(hdr, free_rdata);
3283 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3284 } else if (free_rdata) {
3285 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3287 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr);
3291 hdr->b_crypt_hdr.b_rabd = NULL;
3292 ARCSTAT_INCR(arcstat_raw_size, -size);
3294 hdr->b_l1hdr.b_pabd = NULL;
3297 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3298 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3300 ARCSTAT_INCR(arcstat_compressed_size, -size);
3301 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3304 static arc_buf_hdr_t *
3305 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3306 boolean_t protected, enum zio_compress compression_type,
3307 arc_buf_contents_t type, boolean_t alloc_rdata)
3311 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3313 hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
3315 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3318 ASSERT(HDR_EMPTY(hdr));
3319 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3320 HDR_SET_PSIZE(hdr, psize);
3321 HDR_SET_LSIZE(hdr, lsize);
3325 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3326 arc_hdr_set_compress(hdr, compression_type);
3328 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3330 hdr->b_l1hdr.b_state = arc_anon;
3331 hdr->b_l1hdr.b_arc_access = 0;
3332 hdr->b_l1hdr.b_bufcnt = 0;
3333 hdr->b_l1hdr.b_buf = NULL;
3336 * Allocate the hdr's buffer. This will contain either
3337 * the compressed or uncompressed data depending on the block
3338 * it references and compressed arc enablement.
3340 arc_hdr_alloc_abd(hdr, alloc_rdata);
3341 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3347 * Transition between the two allocation states for the arc_buf_hdr struct.
3348 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3349 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3350 * version is used when a cache buffer is only in the L2ARC in order to reduce
3353 static arc_buf_hdr_t *
3354 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3356 ASSERT(HDR_HAS_L2HDR(hdr));
3358 arc_buf_hdr_t *nhdr;
3359 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3361 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3362 (old == hdr_l2only_cache && new == hdr_full_cache));
3365 * if the caller wanted a new full header and the header is to be
3366 * encrypted we will actually allocate the header from the full crypt
3367 * cache instead. The same applies to freeing from the old cache.
3369 if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
3370 new = hdr_full_crypt_cache;
3371 if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
3372 old = hdr_full_crypt_cache;
3374 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3376 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3377 buf_hash_remove(hdr);
3379 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3381 if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
3382 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3384 * arc_access and arc_change_state need to be aware that a
3385 * header has just come out of L2ARC, so we set its state to
3386 * l2c_only even though it's about to change.
3388 nhdr->b_l1hdr.b_state = arc_l2c_only;
3390 /* Verify previous threads set to NULL before freeing */
3391 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3392 ASSERT(!HDR_HAS_RABD(hdr));
3394 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3395 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3396 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3399 * If we've reached here, We must have been called from
3400 * arc_evict_hdr(), as such we should have already been
3401 * removed from any ghost list we were previously on
3402 * (which protects us from racing with arc_evict_state),
3403 * thus no locking is needed during this check.
3405 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3408 * A buffer must not be moved into the arc_l2c_only
3409 * state if it's not finished being written out to the
3410 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3411 * might try to be accessed, even though it was removed.
3413 VERIFY(!HDR_L2_WRITING(hdr));
3414 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3415 ASSERT(!HDR_HAS_RABD(hdr));
3417 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3420 * The header has been reallocated so we need to re-insert it into any
3423 (void) buf_hash_insert(nhdr, NULL);
3425 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3427 mutex_enter(&dev->l2ad_mtx);
3430 * We must place the realloc'ed header back into the list at
3431 * the same spot. Otherwise, if it's placed earlier in the list,
3432 * l2arc_write_buffers() could find it during the function's
3433 * write phase, and try to write it out to the l2arc.
3435 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3436 list_remove(&dev->l2ad_buflist, hdr);
3438 mutex_exit(&dev->l2ad_mtx);
3441 * Since we're using the pointer address as the tag when
3442 * incrementing and decrementing the l2ad_alloc refcount, we
3443 * must remove the old pointer (that we're about to destroy) and
3444 * add the new pointer to the refcount. Otherwise we'd remove
3445 * the wrong pointer address when calling arc_hdr_destroy() later.
3448 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3449 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3451 buf_discard_identity(hdr);
3452 kmem_cache_free(old, hdr);
3458 * This function allows an L1 header to be reallocated as a crypt
3459 * header and vice versa. If we are going to a crypt header, the
3460 * new fields will be zeroed out.
3462 static arc_buf_hdr_t *
3463 arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
3465 arc_buf_hdr_t *nhdr;
3467 kmem_cache_t *ncache, *ocache;
3469 ASSERT(HDR_HAS_L1HDR(hdr));
3470 ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
3471 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3472 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3475 ncache = hdr_full_crypt_cache;
3476 ocache = hdr_full_cache;
3478 ncache = hdr_full_cache;
3479 ocache = hdr_full_crypt_cache;
3482 nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);
3483 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3484 nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
3485 nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
3486 nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
3487 nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
3488 nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
3489 nhdr->b_l1hdr.b_mru_hits = hdr->b_l1hdr.b_mru_hits;
3490 nhdr->b_l1hdr.b_mru_ghost_hits = hdr->b_l1hdr.b_mru_ghost_hits;
3491 nhdr->b_l1hdr.b_mfu_hits = hdr->b_l1hdr.b_mfu_hits;
3492 nhdr->b_l1hdr.b_mfu_ghost_hits = hdr->b_l1hdr.b_mfu_ghost_hits;
3493 nhdr->b_l1hdr.b_l2_hits = hdr->b_l1hdr.b_l2_hits;
3494 nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
3495 nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
3496 nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
3499 * This refcount_add() exists only to ensure that the individual
3500 * arc buffers always point to a header that is referenced, avoiding
3501 * a small race condition that could trigger ASSERTs.
3503 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
3505 for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) {
3506 mutex_enter(&buf->b_evict_lock);
3508 mutex_exit(&buf->b_evict_lock);
3511 refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
3512 (void) refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
3515 arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
3517 arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
3520 buf_discard_identity(hdr);
3521 kmem_cache_free(ocache, hdr);
3527 * This function is used by the send / receive code to convert a newly
3528 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3529 * is also used to allow the root objset block to be uupdated without altering
3530 * its embedded MACs. Both block types will always be uncompressed so we do not
3531 * have to worry about compression type or psize.
3534 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3535 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3538 arc_buf_hdr_t *hdr = buf->b_hdr;
3540 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3541 ASSERT(HDR_HAS_L1HDR(hdr));
3542 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3544 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3545 if (!HDR_PROTECTED(hdr))
3546 hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
3547 hdr->b_crypt_hdr.b_dsobj = dsobj;
3548 hdr->b_crypt_hdr.b_ot = ot;
3549 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3550 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3551 if (!arc_hdr_has_uncompressed_buf(hdr))
3552 arc_cksum_free(hdr);
3555 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3557 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3559 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3563 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3564 * The buf is returned thawed since we expect the consumer to modify it.
3567 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3569 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3570 B_FALSE, ZIO_COMPRESS_OFF, type, B_FALSE);
3571 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3573 arc_buf_t *buf = NULL;
3574 VERIFY0(arc_buf_alloc_impl(hdr, spa, 0, tag, B_FALSE, B_FALSE,
3575 B_FALSE, B_FALSE, &buf));
3582 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3583 * for bufs containing metadata.
3586 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3587 enum zio_compress compression_type)
3589 ASSERT3U(lsize, >, 0);
3590 ASSERT3U(lsize, >=, psize);
3591 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3592 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3594 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3595 B_FALSE, compression_type, ARC_BUFC_DATA, B_FALSE);
3596 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3598 arc_buf_t *buf = NULL;
3599 VERIFY0(arc_buf_alloc_impl(hdr, spa, 0, tag, B_FALSE,
3600 B_TRUE, B_FALSE, B_FALSE, &buf));
3602 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3604 if (!arc_buf_is_shared(buf)) {
3606 * To ensure that the hdr has the correct data in it if we call
3607 * arc_untransform() on this buf before it's been written to
3608 * disk, it's easiest if we just set up sharing between the
3611 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3612 arc_hdr_free_abd(hdr, B_FALSE);
3613 arc_share_buf(hdr, buf);
3620 arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder,
3621 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
3622 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3623 enum zio_compress compression_type)
3627 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3628 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3630 ASSERT3U(lsize, >, 0);
3631 ASSERT3U(lsize, >=, psize);
3632 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3633 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3635 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3636 compression_type, type, B_TRUE);
3637 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3639 hdr->b_crypt_hdr.b_dsobj = dsobj;
3640 hdr->b_crypt_hdr.b_ot = ot;
3641 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3642 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3643 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3644 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3645 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3648 * This buffer will be considered encrypted even if the ot is not an
3649 * encrypted type. It will become authenticated instead in
3650 * arc_write_ready().
3653 VERIFY0(arc_buf_alloc_impl(hdr, spa, dsobj, tag, B_TRUE, B_TRUE,
3654 B_FALSE, B_FALSE, &buf));
3656 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3662 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3664 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3665 l2arc_dev_t *dev = l2hdr->b_dev;
3666 uint64_t psize = arc_hdr_size(hdr);
3668 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3669 ASSERT(HDR_HAS_L2HDR(hdr));
3671 list_remove(&dev->l2ad_buflist, hdr);
3673 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3674 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3676 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3678 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3679 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3683 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3685 if (HDR_HAS_L1HDR(hdr)) {
3686 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3687 hdr->b_l1hdr.b_bufcnt > 0);
3688 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3689 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3691 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3692 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3694 if (!HDR_EMPTY(hdr))
3695 buf_discard_identity(hdr);
3697 if (HDR_HAS_L2HDR(hdr)) {
3698 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3699 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3702 mutex_enter(&dev->l2ad_mtx);
3705 * Even though we checked this conditional above, we
3706 * need to check this again now that we have the
3707 * l2ad_mtx. This is because we could be racing with
3708 * another thread calling l2arc_evict() which might have
3709 * destroyed this header's L2 portion as we were waiting
3710 * to acquire the l2ad_mtx. If that happens, we don't
3711 * want to re-destroy the header's L2 portion.
3713 if (HDR_HAS_L2HDR(hdr))
3714 arc_hdr_l2hdr_destroy(hdr);
3717 mutex_exit(&dev->l2ad_mtx);
3720 if (HDR_HAS_L1HDR(hdr)) {
3721 arc_cksum_free(hdr);
3723 while (hdr->b_l1hdr.b_buf != NULL)
3724 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3726 if (hdr->b_l1hdr.b_pabd != NULL) {
3727 arc_hdr_free_abd(hdr, B_FALSE);
3730 if (HDR_HAS_RABD(hdr))
3731 arc_hdr_free_abd(hdr, B_TRUE);
3734 ASSERT3P(hdr->b_hash_next, ==, NULL);
3735 if (HDR_HAS_L1HDR(hdr)) {
3736 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3737 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3739 if (!HDR_PROTECTED(hdr)) {
3740 kmem_cache_free(hdr_full_cache, hdr);
3742 kmem_cache_free(hdr_full_crypt_cache, hdr);
3745 kmem_cache_free(hdr_l2only_cache, hdr);
3750 arc_buf_destroy(arc_buf_t *buf, void* tag)
3752 arc_buf_hdr_t *hdr = buf->b_hdr;
3753 kmutex_t *hash_lock = HDR_LOCK(hdr);
3755 if (hdr->b_l1hdr.b_state == arc_anon) {
3756 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3757 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3758 VERIFY0(remove_reference(hdr, NULL, tag));
3759 arc_hdr_destroy(hdr);
3763 mutex_enter(hash_lock);
3764 ASSERT3P(hdr, ==, buf->b_hdr);
3765 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3766 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3767 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3768 ASSERT3P(buf->b_data, !=, NULL);
3770 (void) remove_reference(hdr, hash_lock, tag);
3771 arc_buf_destroy_impl(buf);
3772 mutex_exit(hash_lock);
3776 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3777 * state of the header is dependent on its state prior to entering this
3778 * function. The following transitions are possible:
3780 * - arc_mru -> arc_mru_ghost
3781 * - arc_mfu -> arc_mfu_ghost
3782 * - arc_mru_ghost -> arc_l2c_only
3783 * - arc_mru_ghost -> deleted
3784 * - arc_mfu_ghost -> arc_l2c_only
3785 * - arc_mfu_ghost -> deleted
3788 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3790 arc_state_t *evicted_state, *state;
3791 int64_t bytes_evicted = 0;
3792 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3793 arc_min_prescient_prefetch_ms : arc_min_prefetch_ms;
3795 ASSERT(MUTEX_HELD(hash_lock));
3796 ASSERT(HDR_HAS_L1HDR(hdr));
3798 state = hdr->b_l1hdr.b_state;
3799 if (GHOST_STATE(state)) {
3800 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3801 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3804 * l2arc_write_buffers() relies on a header's L1 portion
3805 * (i.e. its b_pabd field) during it's write phase.
3806 * Thus, we cannot push a header onto the arc_l2c_only
3807 * state (removing its L1 piece) until the header is
3808 * done being written to the l2arc.
3810 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3811 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3812 return (bytes_evicted);
3815 ARCSTAT_BUMP(arcstat_deleted);
3816 bytes_evicted += HDR_GET_LSIZE(hdr);
3818 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3820 if (HDR_HAS_L2HDR(hdr)) {
3821 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3822 ASSERT(!HDR_HAS_RABD(hdr));
3824 * This buffer is cached on the 2nd Level ARC;
3825 * don't destroy the header.
3827 arc_change_state(arc_l2c_only, hdr, hash_lock);
3829 * dropping from L1+L2 cached to L2-only,
3830 * realloc to remove the L1 header.
3832 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3835 arc_change_state(arc_anon, hdr, hash_lock);
3836 arc_hdr_destroy(hdr);
3838 return (bytes_evicted);
3841 ASSERT(state == arc_mru || state == arc_mfu);
3842 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3844 /* prefetch buffers have a minimum lifespan */
3845 if (HDR_IO_IN_PROGRESS(hdr) ||
3846 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3847 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3848 MSEC_TO_TICK(min_lifetime))) {
3849 ARCSTAT_BUMP(arcstat_evict_skip);
3850 return (bytes_evicted);
3853 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3854 while (hdr->b_l1hdr.b_buf) {
3855 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3856 if (!mutex_tryenter(&buf->b_evict_lock)) {
3857 ARCSTAT_BUMP(arcstat_mutex_miss);
3860 if (buf->b_data != NULL)
3861 bytes_evicted += HDR_GET_LSIZE(hdr);
3862 mutex_exit(&buf->b_evict_lock);
3863 arc_buf_destroy_impl(buf);
3866 if (HDR_HAS_L2HDR(hdr)) {
3867 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3869 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3870 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3871 HDR_GET_LSIZE(hdr));
3873 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3874 HDR_GET_LSIZE(hdr));
3878 if (hdr->b_l1hdr.b_bufcnt == 0) {
3879 arc_cksum_free(hdr);
3881 bytes_evicted += arc_hdr_size(hdr);
3884 * If this hdr is being evicted and has a compressed
3885 * buffer then we discard it here before we change states.
3886 * This ensures that the accounting is updated correctly
3887 * in arc_free_data_impl().
3889 if (hdr->b_l1hdr.b_pabd != NULL)
3890 arc_hdr_free_abd(hdr, B_FALSE);
3892 if (HDR_HAS_RABD(hdr))
3893 arc_hdr_free_abd(hdr, B_TRUE);
3895 arc_change_state(evicted_state, hdr, hash_lock);
3896 ASSERT(HDR_IN_HASH_TABLE(hdr));
3897 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3898 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3901 return (bytes_evicted);
3905 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3906 uint64_t spa, int64_t bytes)
3908 multilist_sublist_t *mls;
3909 uint64_t bytes_evicted = 0;
3911 kmutex_t *hash_lock;
3912 int evict_count = 0;
3914 ASSERT3P(marker, !=, NULL);
3915 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3917 mls = multilist_sublist_lock(ml, idx);
3919 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3920 hdr = multilist_sublist_prev(mls, marker)) {
3921 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3922 (evict_count >= zfs_arc_evict_batch_limit))
3926 * To keep our iteration location, move the marker
3927 * forward. Since we're not holding hdr's hash lock, we
3928 * must be very careful and not remove 'hdr' from the
3929 * sublist. Otherwise, other consumers might mistake the
3930 * 'hdr' as not being on a sublist when they call the
3931 * multilist_link_active() function (they all rely on
3932 * the hash lock protecting concurrent insertions and
3933 * removals). multilist_sublist_move_forward() was
3934 * specifically implemented to ensure this is the case
3935 * (only 'marker' will be removed and re-inserted).
3937 multilist_sublist_move_forward(mls, marker);
3940 * The only case where the b_spa field should ever be
3941 * zero, is the marker headers inserted by
3942 * arc_evict_state(). It's possible for multiple threads
3943 * to be calling arc_evict_state() concurrently (e.g.
3944 * dsl_pool_close() and zio_inject_fault()), so we must
3945 * skip any markers we see from these other threads.
3947 if (hdr->b_spa == 0)
3950 /* we're only interested in evicting buffers of a certain spa */
3951 if (spa != 0 && hdr->b_spa != spa) {
3952 ARCSTAT_BUMP(arcstat_evict_skip);
3956 hash_lock = HDR_LOCK(hdr);
3959 * We aren't calling this function from any code path
3960 * that would already be holding a hash lock, so we're
3961 * asserting on this assumption to be defensive in case
3962 * this ever changes. Without this check, it would be
3963 * possible to incorrectly increment arcstat_mutex_miss
3964 * below (e.g. if the code changed such that we called
3965 * this function with a hash lock held).
3967 ASSERT(!MUTEX_HELD(hash_lock));
3969 if (mutex_tryenter(hash_lock)) {
3970 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3971 mutex_exit(hash_lock);
3973 bytes_evicted += evicted;
3976 * If evicted is zero, arc_evict_hdr() must have
3977 * decided to skip this header, don't increment
3978 * evict_count in this case.
3984 * If arc_size isn't overflowing, signal any
3985 * threads that might happen to be waiting.
3987 * For each header evicted, we wake up a single
3988 * thread. If we used cv_broadcast, we could
3989 * wake up "too many" threads causing arc_size
3990 * to significantly overflow arc_c; since
3991 * arc_get_data_impl() doesn't check for overflow
3992 * when it's woken up (it doesn't because it's
3993 * possible for the ARC to be overflowing while
3994 * full of un-evictable buffers, and the
3995 * function should proceed in this case).
3997 * If threads are left sleeping, due to not
3998 * using cv_broadcast, they will be woken up
3999 * just before arc_reclaim_thread() sleeps.
4001 mutex_enter(&arc_reclaim_lock);
4002 if (!arc_is_overflowing())
4003 cv_signal(&arc_reclaim_waiters_cv);
4004 mutex_exit(&arc_reclaim_lock);
4006 ARCSTAT_BUMP(arcstat_mutex_miss);
4010 multilist_sublist_unlock(mls);
4012 return (bytes_evicted);
4016 * Evict buffers from the given arc state, until we've removed the
4017 * specified number of bytes. Move the removed buffers to the
4018 * appropriate evict state.
4020 * This function makes a "best effort". It skips over any buffers
4021 * it can't get a hash_lock on, and so, may not catch all candidates.
4022 * It may also return without evicting as much space as requested.
4024 * If bytes is specified using the special value ARC_EVICT_ALL, this
4025 * will evict all available (i.e. unlocked and evictable) buffers from
4026 * the given arc state; which is used by arc_flush().
4029 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
4030 arc_buf_contents_t type)
4032 uint64_t total_evicted = 0;
4033 multilist_t *ml = state->arcs_list[type];
4035 arc_buf_hdr_t **markers;
4037 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
4039 num_sublists = multilist_get_num_sublists(ml);
4042 * If we've tried to evict from each sublist, made some
4043 * progress, but still have not hit the target number of bytes
4044 * to evict, we want to keep trying. The markers allow us to
4045 * pick up where we left off for each individual sublist, rather
4046 * than starting from the tail each time.
4048 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
4049 for (int i = 0; i < num_sublists; i++) {
4050 multilist_sublist_t *mls;
4052 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
4055 * A b_spa of 0 is used to indicate that this header is
4056 * a marker. This fact is used in arc_adjust_type() and
4057 * arc_evict_state_impl().
4059 markers[i]->b_spa = 0;
4061 mls = multilist_sublist_lock(ml, i);
4062 multilist_sublist_insert_tail(mls, markers[i]);
4063 multilist_sublist_unlock(mls);
4067 * While we haven't hit our target number of bytes to evict, or
4068 * we're evicting all available buffers.
4070 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
4071 int sublist_idx = multilist_get_random_index(ml);
4072 uint64_t scan_evicted = 0;
4075 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
4076 * Request that 10% of the LRUs be scanned by the superblock
4079 if (type == ARC_BUFC_DATA && arc_dnode_size > arc_dnode_limit)
4080 arc_prune_async((arc_dnode_size - arc_dnode_limit) /
4081 sizeof (dnode_t) / zfs_arc_dnode_reduce_percent);
4084 * Start eviction using a randomly selected sublist,
4085 * this is to try and evenly balance eviction across all
4086 * sublists. Always starting at the same sublist
4087 * (e.g. index 0) would cause evictions to favor certain
4088 * sublists over others.
4090 for (int i = 0; i < num_sublists; i++) {
4091 uint64_t bytes_remaining;
4092 uint64_t bytes_evicted;
4094 if (bytes == ARC_EVICT_ALL)
4095 bytes_remaining = ARC_EVICT_ALL;
4096 else if (total_evicted < bytes)
4097 bytes_remaining = bytes - total_evicted;
4101 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4102 markers[sublist_idx], spa, bytes_remaining);
4104 scan_evicted += bytes_evicted;
4105 total_evicted += bytes_evicted;
4107 /* we've reached the end, wrap to the beginning */
4108 if (++sublist_idx >= num_sublists)
4113 * If we didn't evict anything during this scan, we have
4114 * no reason to believe we'll evict more during another
4115 * scan, so break the loop.
4117 if (scan_evicted == 0) {
4118 /* This isn't possible, let's make that obvious */
4119 ASSERT3S(bytes, !=, 0);
4122 * When bytes is ARC_EVICT_ALL, the only way to
4123 * break the loop is when scan_evicted is zero.
4124 * In that case, we actually have evicted enough,
4125 * so we don't want to increment the kstat.
4127 if (bytes != ARC_EVICT_ALL) {
4128 ASSERT3S(total_evicted, <, bytes);
4129 ARCSTAT_BUMP(arcstat_evict_not_enough);
4136 for (int i = 0; i < num_sublists; i++) {
4137 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4138 multilist_sublist_remove(mls, markers[i]);
4139 multilist_sublist_unlock(mls);
4141 kmem_cache_free(hdr_full_cache, markers[i]);
4143 kmem_free(markers, sizeof (*markers) * num_sublists);
4145 return (total_evicted);
4149 * Flush all "evictable" data of the given type from the arc state
4150 * specified. This will not evict any "active" buffers (i.e. referenced).
4152 * When 'retry' is set to B_FALSE, the function will make a single pass
4153 * over the state and evict any buffers that it can. Since it doesn't
4154 * continually retry the eviction, it might end up leaving some buffers
4155 * in the ARC due to lock misses.
4157 * When 'retry' is set to B_TRUE, the function will continually retry the
4158 * eviction until *all* evictable buffers have been removed from the
4159 * state. As a result, if concurrent insertions into the state are
4160 * allowed (e.g. if the ARC isn't shutting down), this function might
4161 * wind up in an infinite loop, continually trying to evict buffers.
4164 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4167 uint64_t evicted = 0;
4169 while (refcount_count(&state->arcs_esize[type]) != 0) {
4170 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4180 * Helper function for arc_prune_async() it is responsible for safely
4181 * handling the execution of a registered arc_prune_func_t.
4184 arc_prune_task(void *ptr)
4186 arc_prune_t *ap = (arc_prune_t *)ptr;
4187 arc_prune_func_t *func = ap->p_pfunc;
4190 func(ap->p_adjust, ap->p_private);
4192 refcount_remove(&ap->p_refcnt, func);
4196 * Notify registered consumers they must drop holds on a portion of the ARC
4197 * buffered they reference. This provides a mechanism to ensure the ARC can
4198 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
4199 * is analogous to dnlc_reduce_cache() but more generic.
4201 * This operation is performed asynchronously so it may be safely called
4202 * in the context of the arc_reclaim_thread(). A reference is taken here
4203 * for each registered arc_prune_t and the arc_prune_task() is responsible
4204 * for releasing it once the registered arc_prune_func_t has completed.
4207 arc_prune_async(int64_t adjust)
4211 mutex_enter(&arc_prune_mtx);
4212 for (ap = list_head(&arc_prune_list); ap != NULL;
4213 ap = list_next(&arc_prune_list, ap)) {
4215 if (refcount_count(&ap->p_refcnt) >= 2)
4218 refcount_add(&ap->p_refcnt, ap->p_pfunc);
4219 ap->p_adjust = adjust;
4220 if (taskq_dispatch(arc_prune_taskq, arc_prune_task,
4221 ap, TQ_SLEEP) == TASKQID_INVALID) {
4222 refcount_remove(&ap->p_refcnt, ap->p_pfunc);
4225 ARCSTAT_BUMP(arcstat_prune);
4227 mutex_exit(&arc_prune_mtx);
4231 * Evict the specified number of bytes from the state specified,
4232 * restricting eviction to the spa and type given. This function
4233 * prevents us from trying to evict more from a state's list than
4234 * is "evictable", and to skip evicting altogether when passed a
4235 * negative value for "bytes". In contrast, arc_evict_state() will
4236 * evict everything it can, when passed a negative value for "bytes".
4239 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4240 arc_buf_contents_t type)
4244 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
4245 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
4246 return (arc_evict_state(state, spa, delta, type));
4253 * The goal of this function is to evict enough meta data buffers from the
4254 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4255 * more complicated than it appears because it is common for data buffers
4256 * to have holds on meta data buffers. In addition, dnode meta data buffers
4257 * will be held by the dnodes in the block preventing them from being freed.
4258 * This means we can't simply traverse the ARC and expect to always find
4259 * enough unheld meta data buffer to release.
4261 * Therefore, this function has been updated to make alternating passes
4262 * over the ARC releasing data buffers and then newly unheld meta data
4263 * buffers. This ensures forward progress is maintained and arc_meta_used
4264 * will decrease. Normally this is sufficient, but if required the ARC
4265 * will call the registered prune callbacks causing dentry and inodes to
4266 * be dropped from the VFS cache. This will make dnode meta data buffers
4267 * available for reclaim.
4270 arc_adjust_meta_balanced(void)
4272 int64_t delta, prune = 0, adjustmnt;
4273 uint64_t total_evicted = 0;
4274 arc_buf_contents_t type = ARC_BUFC_DATA;
4275 int restarts = MAX(zfs_arc_meta_adjust_restarts, 0);
4279 * This slightly differs than the way we evict from the mru in
4280 * arc_adjust because we don't have a "target" value (i.e. no
4281 * "meta" arc_p). As a result, I think we can completely
4282 * cannibalize the metadata in the MRU before we evict the
4283 * metadata from the MFU. I think we probably need to implement a
4284 * "metadata arc_p" value to do this properly.
4286 adjustmnt = arc_meta_used - arc_meta_limit;
4288 if (adjustmnt > 0 && refcount_count(&arc_mru->arcs_esize[type]) > 0) {
4289 delta = MIN(refcount_count(&arc_mru->arcs_esize[type]),
4291 total_evicted += arc_adjust_impl(arc_mru, 0, delta, type);
4296 * We can't afford to recalculate adjustmnt here. If we do,
4297 * new metadata buffers can sneak into the MRU or ANON lists,
4298 * thus penalize the MFU metadata. Although the fudge factor is
4299 * small, it has been empirically shown to be significant for
4300 * certain workloads (e.g. creating many empty directories). As
4301 * such, we use the original calculation for adjustmnt, and
4302 * simply decrement the amount of data evicted from the MRU.
4305 if (adjustmnt > 0 && refcount_count(&arc_mfu->arcs_esize[type]) > 0) {
4306 delta = MIN(refcount_count(&arc_mfu->arcs_esize[type]),
4308 total_evicted += arc_adjust_impl(arc_mfu, 0, delta, type);
4311 adjustmnt = arc_meta_used - arc_meta_limit;
4313 if (adjustmnt > 0 &&
4314 refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) {
4315 delta = MIN(adjustmnt,
4316 refcount_count(&arc_mru_ghost->arcs_esize[type]));
4317 total_evicted += arc_adjust_impl(arc_mru_ghost, 0, delta, type);
4321 if (adjustmnt > 0 &&
4322 refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) {
4323 delta = MIN(adjustmnt,
4324 refcount_count(&arc_mfu_ghost->arcs_esize[type]));
4325 total_evicted += arc_adjust_impl(arc_mfu_ghost, 0, delta, type);
4329 * If after attempting to make the requested adjustment to the ARC
4330 * the meta limit is still being exceeded then request that the
4331 * higher layers drop some cached objects which have holds on ARC
4332 * meta buffers. Requests to the upper layers will be made with
4333 * increasingly large scan sizes until the ARC is below the limit.
4335 if (arc_meta_used > arc_meta_limit) {
4336 if (type == ARC_BUFC_DATA) {
4337 type = ARC_BUFC_METADATA;
4339 type = ARC_BUFC_DATA;
4341 if (zfs_arc_meta_prune) {
4342 prune += zfs_arc_meta_prune;
4343 arc_prune_async(prune);
4352 return (total_evicted);
4356 * Evict metadata buffers from the cache, such that arc_meta_used is
4357 * capped by the arc_meta_limit tunable.
4360 arc_adjust_meta_only(void)
4362 uint64_t total_evicted = 0;
4366 * If we're over the meta limit, we want to evict enough
4367 * metadata to get back under the meta limit. We don't want to
4368 * evict so much that we drop the MRU below arc_p, though. If
4369 * we're over the meta limit more than we're over arc_p, we
4370 * evict some from the MRU here, and some from the MFU below.
4372 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
4373 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4374 refcount_count(&arc_mru->arcs_size) - arc_p));
4376 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4379 * Similar to the above, we want to evict enough bytes to get us
4380 * below the meta limit, but not so much as to drop us below the
4381 * space allotted to the MFU (which is defined as arc_c - arc_p).
4383 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
4384 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
4386 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4388 return (total_evicted);
4392 arc_adjust_meta(void)
4394 if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY)
4395 return (arc_adjust_meta_only());
4397 return (arc_adjust_meta_balanced());
4401 * Return the type of the oldest buffer in the given arc state
4403 * This function will select a random sublist of type ARC_BUFC_DATA and
4404 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4405 * is compared, and the type which contains the "older" buffer will be
4408 static arc_buf_contents_t
4409 arc_adjust_type(arc_state_t *state)
4411 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4412 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4413 int data_idx = multilist_get_random_index(data_ml);
4414 int meta_idx = multilist_get_random_index(meta_ml);
4415 multilist_sublist_t *data_mls;
4416 multilist_sublist_t *meta_mls;
4417 arc_buf_contents_t type;
4418 arc_buf_hdr_t *data_hdr;
4419 arc_buf_hdr_t *meta_hdr;
4422 * We keep the sublist lock until we're finished, to prevent
4423 * the headers from being destroyed via arc_evict_state().
4425 data_mls = multilist_sublist_lock(data_ml, data_idx);
4426 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4429 * These two loops are to ensure we skip any markers that
4430 * might be at the tail of the lists due to arc_evict_state().
4433 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4434 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4435 if (data_hdr->b_spa != 0)
4439 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4440 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4441 if (meta_hdr->b_spa != 0)
4445 if (data_hdr == NULL && meta_hdr == NULL) {
4446 type = ARC_BUFC_DATA;
4447 } else if (data_hdr == NULL) {
4448 ASSERT3P(meta_hdr, !=, NULL);
4449 type = ARC_BUFC_METADATA;
4450 } else if (meta_hdr == NULL) {
4451 ASSERT3P(data_hdr, !=, NULL);
4452 type = ARC_BUFC_DATA;
4454 ASSERT3P(data_hdr, !=, NULL);
4455 ASSERT3P(meta_hdr, !=, NULL);
4457 /* The headers can't be on the sublist without an L1 header */
4458 ASSERT(HDR_HAS_L1HDR(data_hdr));
4459 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4461 if (data_hdr->b_l1hdr.b_arc_access <
4462 meta_hdr->b_l1hdr.b_arc_access) {
4463 type = ARC_BUFC_DATA;
4465 type = ARC_BUFC_METADATA;
4469 multilist_sublist_unlock(meta_mls);
4470 multilist_sublist_unlock(data_mls);
4476 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4481 uint64_t total_evicted = 0;
4486 * If we're over arc_meta_limit, we want to correct that before
4487 * potentially evicting data buffers below.
4489 total_evicted += arc_adjust_meta();
4494 * If we're over the target cache size, we want to evict enough
4495 * from the list to get back to our target size. We don't want
4496 * to evict too much from the MRU, such that it drops below
4497 * arc_p. So, if we're over our target cache size more than
4498 * the MRU is over arc_p, we'll evict enough to get back to
4499 * arc_p here, and then evict more from the MFU below.
4501 target = MIN((int64_t)(arc_size - arc_c),
4502 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4503 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
4506 * If we're below arc_meta_min, always prefer to evict data.
4507 * Otherwise, try to satisfy the requested number of bytes to
4508 * evict from the type which contains older buffers; in an
4509 * effort to keep newer buffers in the cache regardless of their
4510 * type. If we cannot satisfy the number of bytes from this
4511 * type, spill over into the next type.
4513 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4514 arc_meta_used > arc_meta_min) {
4515 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4516 total_evicted += bytes;
4519 * If we couldn't evict our target number of bytes from
4520 * metadata, we try to get the rest from data.
4525 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4527 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4528 total_evicted += bytes;
4531 * If we couldn't evict our target number of bytes from
4532 * data, we try to get the rest from metadata.
4537 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4543 * Now that we've tried to evict enough from the MRU to get its
4544 * size back to arc_p, if we're still above the target cache
4545 * size, we evict the rest from the MFU.
4547 target = arc_size - arc_c;
4549 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4550 arc_meta_used > arc_meta_min) {
4551 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4552 total_evicted += bytes;
4555 * If we couldn't evict our target number of bytes from
4556 * metadata, we try to get the rest from data.
4561 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4563 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4564 total_evicted += bytes;
4567 * If we couldn't evict our target number of bytes from
4568 * data, we try to get the rest from data.
4573 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4577 * Adjust ghost lists
4579 * In addition to the above, the ARC also defines target values
4580 * for the ghost lists. The sum of the mru list and mru ghost
4581 * list should never exceed the target size of the cache, and
4582 * the sum of the mru list, mfu list, mru ghost list, and mfu
4583 * ghost list should never exceed twice the target size of the
4584 * cache. The following logic enforces these limits on the ghost
4585 * caches, and evicts from them as needed.
4587 target = refcount_count(&arc_mru->arcs_size) +
4588 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4590 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4591 total_evicted += bytes;
4596 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4599 * We assume the sum of the mru list and mfu list is less than
4600 * or equal to arc_c (we enforced this above), which means we
4601 * can use the simpler of the two equations below:
4603 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4604 * mru ghost + mfu ghost <= arc_c
4606 target = refcount_count(&arc_mru_ghost->arcs_size) +
4607 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4609 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4610 total_evicted += bytes;
4615 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4617 return (total_evicted);
4621 arc_flush(spa_t *spa, boolean_t retry)
4626 * If retry is B_TRUE, a spa must not be specified since we have
4627 * no good way to determine if all of a spa's buffers have been
4628 * evicted from an arc state.
4630 ASSERT(!retry || spa == 0);
4633 guid = spa_load_guid(spa);
4635 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4636 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4638 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4639 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4641 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4642 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4644 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4645 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4649 arc_shrink(int64_t to_free)
4653 if (c > to_free && c - to_free > arc_c_min) {
4654 arc_c = c - to_free;
4655 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4656 if (arc_c > arc_size)
4657 arc_c = MAX(arc_size, arc_c_min);
4659 arc_p = (arc_c >> 1);
4660 ASSERT(arc_c >= arc_c_min);
4661 ASSERT((int64_t)arc_p >= 0);
4666 if (arc_size > arc_c)
4667 (void) arc_adjust();
4671 * Return maximum amount of memory that we could possibly use. Reduced
4672 * to half of all memory in user space which is primarily used for testing.
4675 arc_all_memory(void)
4678 #ifdef CONFIG_HIGHMEM
4679 return (ptob(totalram_pages - totalhigh_pages));
4681 return (ptob(totalram_pages));
4682 #endif /* CONFIG_HIGHMEM */
4684 return (ptob(physmem) / 2);
4685 #endif /* _KERNEL */
4689 * Return the amount of memory that is considered free. In user space
4690 * which is primarily used for testing we pretend that free memory ranges
4691 * from 0-20% of all memory.
4694 arc_free_memory(void)
4697 #ifdef CONFIG_HIGHMEM
4700 return (ptob(si.freeram - si.freehigh));
4702 #ifdef ZFS_GLOBAL_NODE_PAGE_STATE
4703 return (ptob(nr_free_pages() +
4704 global_node_page_state(NR_INACTIVE_FILE) +
4705 global_node_page_state(NR_INACTIVE_ANON) +
4706 global_node_page_state(NR_SLAB_RECLAIMABLE)));
4708 return (ptob(nr_free_pages() +
4709 global_page_state(NR_INACTIVE_FILE) +
4710 global_page_state(NR_INACTIVE_ANON) +
4711 global_page_state(NR_SLAB_RECLAIMABLE)));
4712 #endif /* ZFS_GLOBAL_NODE_PAGE_STATE */
4713 #endif /* CONFIG_HIGHMEM */
4715 return (spa_get_random(arc_all_memory() * 20 / 100));
4716 #endif /* _KERNEL */
4719 typedef enum free_memory_reason_t {
4724 FMR_PAGES_PP_MAXIMUM,
4727 } free_memory_reason_t;
4729 int64_t last_free_memory;
4730 free_memory_reason_t last_free_reason;
4734 * Additional reserve of pages for pp_reserve.
4736 int64_t arc_pages_pp_reserve = 64;
4739 * Additional reserve of pages for swapfs.
4741 int64_t arc_swapfs_reserve = 64;
4742 #endif /* _KERNEL */
4745 * Return the amount of memory that can be consumed before reclaim will be
4746 * needed. Positive if there is sufficient free memory, negative indicates
4747 * the amount of memory that needs to be freed up.
4750 arc_available_memory(void)
4752 int64_t lowest = INT64_MAX;
4753 free_memory_reason_t r = FMR_UNKNOWN;
4760 pgcnt_t needfree = btop(arc_need_free);
4761 pgcnt_t lotsfree = btop(arc_sys_free);
4762 pgcnt_t desfree = 0;
4763 pgcnt_t freemem = btop(arc_free_memory());
4767 n = PAGESIZE * (-needfree);
4775 * check that we're out of range of the pageout scanner. It starts to
4776 * schedule paging if freemem is less than lotsfree and needfree.
4777 * lotsfree is the high-water mark for pageout, and needfree is the
4778 * number of needed free pages. We add extra pages here to make sure
4779 * the scanner doesn't start up while we're freeing memory.
4781 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4789 * check to make sure that swapfs has enough space so that anon
4790 * reservations can still succeed. anon_resvmem() checks that the
4791 * availrmem is greater than swapfs_minfree, and the number of reserved
4792 * swap pages. We also add a bit of extra here just to prevent
4793 * circumstances from getting really dire.
4795 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4796 desfree - arc_swapfs_reserve);
4799 r = FMR_SWAPFS_MINFREE;
4803 * Check that we have enough availrmem that memory locking (e.g., via
4804 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4805 * stores the number of pages that cannot be locked; when availrmem
4806 * drops below pages_pp_maximum, page locking mechanisms such as
4807 * page_pp_lock() will fail.)
4809 n = PAGESIZE * (availrmem - pages_pp_maximum -
4810 arc_pages_pp_reserve);
4813 r = FMR_PAGES_PP_MAXIMUM;
4819 * If we're on a 32-bit platform, it's possible that we'll exhaust the
4820 * kernel heap space before we ever run out of available physical
4821 * memory. Most checks of the size of the heap_area compare against
4822 * tune.t_minarmem, which is the minimum available real memory that we
4823 * can have in the system. However, this is generally fixed at 25 pages
4824 * which is so low that it's useless. In this comparison, we seek to
4825 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4826 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4829 n = vmem_size(heap_arena, VMEM_FREE) -
4830 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
4838 * If zio data pages are being allocated out of a separate heap segment,
4839 * then enforce that the size of available vmem for this arena remains
4840 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4842 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4843 * memory (in the zio_arena) free, which can avoid memory
4844 * fragmentation issues.
4846 if (zio_arena != NULL) {
4847 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4848 (vmem_size(zio_arena, VMEM_ALLOC) >>
4849 arc_zio_arena_free_shift);
4856 /* Every 100 calls, free a small amount */
4857 if (spa_get_random(100) == 0)
4859 #endif /* _KERNEL */
4861 last_free_memory = lowest;
4862 last_free_reason = r;
4868 * Determine if the system is under memory pressure and is asking
4869 * to reclaim memory. A return value of B_TRUE indicates that the system
4870 * is under memory pressure and that the arc should adjust accordingly.
4873 arc_reclaim_needed(void)
4875 return (arc_available_memory() < 0);
4879 arc_kmem_reap_now(void)
4882 kmem_cache_t *prev_cache = NULL;
4883 kmem_cache_t *prev_data_cache = NULL;
4884 extern kmem_cache_t *zio_buf_cache[];
4885 extern kmem_cache_t *zio_data_buf_cache[];
4886 extern kmem_cache_t *range_seg_cache;
4889 if ((arc_meta_used >= arc_meta_limit) && zfs_arc_meta_prune) {
4891 * We are exceeding our meta-data cache limit.
4892 * Prune some entries to release holds on meta-data.
4894 arc_prune_async(zfs_arc_meta_prune);
4898 * Reclaim unused memory from all kmem caches.
4904 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4906 /* reach upper limit of cache size on 32-bit */
4907 if (zio_buf_cache[i] == NULL)
4910 if (zio_buf_cache[i] != prev_cache) {
4911 prev_cache = zio_buf_cache[i];
4912 kmem_cache_reap_now(zio_buf_cache[i]);
4914 if (zio_data_buf_cache[i] != prev_data_cache) {
4915 prev_data_cache = zio_data_buf_cache[i];
4916 kmem_cache_reap_now(zio_data_buf_cache[i]);
4919 kmem_cache_reap_now(buf_cache);
4920 kmem_cache_reap_now(hdr_full_cache);
4921 kmem_cache_reap_now(hdr_l2only_cache);
4922 kmem_cache_reap_now(range_seg_cache);
4924 if (zio_arena != NULL) {
4926 * Ask the vmem arena to reclaim unused memory from its
4929 vmem_qcache_reap(zio_arena);
4934 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4935 * enough data and signal them to proceed. When this happens, the threads in
4936 * arc_get_data_impl() are sleeping while holding the hash lock for their
4937 * particular arc header. Thus, we must be careful to never sleep on a
4938 * hash lock in this thread. This is to prevent the following deadlock:
4940 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4941 * waiting for the reclaim thread to signal it.
4943 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4944 * fails, and goes to sleep forever.
4946 * This possible deadlock is avoided by always acquiring a hash lock
4947 * using mutex_tryenter() from arc_reclaim_thread().
4951 arc_reclaim_thread(void *unused)
4953 fstrans_cookie_t cookie = spl_fstrans_mark();
4954 hrtime_t growtime = 0;
4957 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4959 mutex_enter(&arc_reclaim_lock);
4960 while (!arc_reclaim_thread_exit) {
4961 uint64_t evicted = 0;
4962 uint64_t need_free = arc_need_free;
4963 arc_tuning_update();
4966 * This is necessary in order for the mdb ::arc dcmd to
4967 * show up to date information. Since the ::arc command
4968 * does not call the kstat's update function, without
4969 * this call, the command may show stale stats for the
4970 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4971 * with this change, the data might be up to 1 second
4972 * out of date; but that should suffice. The arc_state_t
4973 * structures can be queried directly if more accurate
4974 * information is needed.
4977 if (arc_ksp != NULL)
4978 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4980 mutex_exit(&arc_reclaim_lock);
4983 * We call arc_adjust() before (possibly) calling
4984 * arc_kmem_reap_now(), so that we can wake up
4985 * arc_get_data_buf() sooner.
4987 evicted = arc_adjust();
4989 int64_t free_memory = arc_available_memory();
4990 if (free_memory < 0) {
4992 arc_no_grow = B_TRUE;
4996 * Wait at least zfs_grow_retry (default 5) seconds
4997 * before considering growing.
4999 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
5001 arc_kmem_reap_now();
5004 * If we are still low on memory, shrink the ARC
5005 * so that we have arc_shrink_min free space.
5007 free_memory = arc_available_memory();
5010 (arc_c >> arc_shrink_shift) - free_memory;
5013 to_free = MAX(to_free, need_free);
5015 arc_shrink(to_free);
5017 } else if (free_memory < arc_c >> arc_no_grow_shift) {
5018 arc_no_grow = B_TRUE;
5019 } else if (gethrtime() >= growtime) {
5020 arc_no_grow = B_FALSE;
5023 mutex_enter(&arc_reclaim_lock);
5026 * If evicted is zero, we couldn't evict anything via
5027 * arc_adjust(). This could be due to hash lock
5028 * collisions, but more likely due to the majority of
5029 * arc buffers being unevictable. Therefore, even if
5030 * arc_size is above arc_c, another pass is unlikely to
5031 * be helpful and could potentially cause us to enter an
5034 if (arc_size <= arc_c || evicted == 0) {
5036 * We're either no longer overflowing, or we
5037 * can't evict anything more, so we should wake
5038 * up any threads before we go to sleep and remove
5039 * the bytes we were working on from arc_need_free
5040 * since nothing more will be done here.
5042 cv_broadcast(&arc_reclaim_waiters_cv);
5043 ARCSTAT_INCR(arcstat_need_free, -need_free);
5046 * Block until signaled, or after one second (we
5047 * might need to perform arc_kmem_reap_now()
5048 * even if we aren't being signalled)
5050 CALLB_CPR_SAFE_BEGIN(&cpr);
5051 (void) cv_timedwait_sig_hires(&arc_reclaim_thread_cv,
5052 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
5053 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
5057 arc_reclaim_thread_exit = B_FALSE;
5058 cv_broadcast(&arc_reclaim_thread_cv);
5059 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
5060 spl_fstrans_unmark(cookie);
5066 * Determine the amount of memory eligible for eviction contained in the
5067 * ARC. All clean data reported by the ghost lists can always be safely
5068 * evicted. Due to arc_c_min, the same does not hold for all clean data
5069 * contained by the regular mru and mfu lists.
5071 * In the case of the regular mru and mfu lists, we need to report as
5072 * much clean data as possible, such that evicting that same reported
5073 * data will not bring arc_size below arc_c_min. Thus, in certain
5074 * circumstances, the total amount of clean data in the mru and mfu
5075 * lists might not actually be evictable.
5077 * The following two distinct cases are accounted for:
5079 * 1. The sum of the amount of dirty data contained by both the mru and
5080 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5081 * is greater than or equal to arc_c_min.
5082 * (i.e. amount of dirty data >= arc_c_min)
5084 * This is the easy case; all clean data contained by the mru and mfu
5085 * lists is evictable. Evicting all clean data can only drop arc_size
5086 * to the amount of dirty data, which is greater than arc_c_min.
5088 * 2. The sum of the amount of dirty data contained by both the mru and
5089 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5090 * is less than arc_c_min.
5091 * (i.e. arc_c_min > amount of dirty data)
5093 * 2.1. arc_size is greater than or equal arc_c_min.
5094 * (i.e. arc_size >= arc_c_min > amount of dirty data)
5096 * In this case, not all clean data from the regular mru and mfu
5097 * lists is actually evictable; we must leave enough clean data
5098 * to keep arc_size above arc_c_min. Thus, the maximum amount of
5099 * evictable data from the two lists combined, is exactly the
5100 * difference between arc_size and arc_c_min.
5102 * 2.2. arc_size is less than arc_c_min
5103 * (i.e. arc_c_min > arc_size > amount of dirty data)
5105 * In this case, none of the data contained in the mru and mfu
5106 * lists is evictable, even if it's clean. Since arc_size is
5107 * already below arc_c_min, evicting any more would only
5108 * increase this negative difference.
5111 arc_evictable_memory(void)
5113 uint64_t arc_clean =
5114 refcount_count(&arc_mru->arcs_esize[ARC_BUFC_DATA]) +
5115 refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA]) +
5116 refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_DATA]) +
5117 refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
5118 uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0);
5121 * Scale reported evictable memory in proportion to page cache, cap
5122 * at specified min/max.
5124 #ifdef ZFS_GLOBAL_NODE_PAGE_STATE
5125 uint64_t min = (ptob(global_node_page_state(NR_FILE_PAGES)) / 100) *
5128 uint64_t min = (ptob(global_page_state(NR_FILE_PAGES)) / 100) *
5131 min = MAX(arc_c_min, MIN(arc_c_max, min));
5133 if (arc_dirty >= min)
5136 return (MAX((int64_t)arc_size - (int64_t)min, 0));
5140 * If sc->nr_to_scan is zero, the caller is requesting a query of the
5141 * number of objects which can potentially be freed. If it is nonzero,
5142 * the request is to free that many objects.
5144 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
5145 * in struct shrinker and also require the shrinker to return the number
5148 * Older kernels require the shrinker to return the number of freeable
5149 * objects following the freeing of nr_to_free.
5151 static spl_shrinker_t
5152 __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc)
5156 /* The arc is considered warm once reclaim has occurred */
5157 if (unlikely(arc_warm == B_FALSE))
5160 /* Return the potential number of reclaimable pages */
5161 pages = btop((int64_t)arc_evictable_memory());
5162 if (sc->nr_to_scan == 0)
5165 /* Not allowed to perform filesystem reclaim */
5166 if (!(sc->gfp_mask & __GFP_FS))
5167 return (SHRINK_STOP);
5169 /* Reclaim in progress */
5170 if (mutex_tryenter(&arc_reclaim_lock) == 0) {
5171 ARCSTAT_INCR(arcstat_need_free, ptob(sc->nr_to_scan));
5175 mutex_exit(&arc_reclaim_lock);
5178 * Evict the requested number of pages by shrinking arc_c the
5182 arc_shrink(ptob(sc->nr_to_scan));
5183 if (current_is_kswapd())
5184 arc_kmem_reap_now();
5185 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
5186 pages = MAX((int64_t)pages -
5187 (int64_t)btop(arc_evictable_memory()), 0);
5189 pages = btop(arc_evictable_memory());
5192 * We've shrunk what we can, wake up threads.
5194 cv_broadcast(&arc_reclaim_waiters_cv);
5196 pages = SHRINK_STOP;
5199 * When direct reclaim is observed it usually indicates a rapid
5200 * increase in memory pressure. This occurs because the kswapd
5201 * threads were unable to asynchronously keep enough free memory
5202 * available. In this case set arc_no_grow to briefly pause arc
5203 * growth to avoid compounding the memory pressure.
5205 if (current_is_kswapd()) {
5206 ARCSTAT_BUMP(arcstat_memory_indirect_count);
5208 arc_no_grow = B_TRUE;
5209 arc_kmem_reap_now();
5210 ARCSTAT_BUMP(arcstat_memory_direct_count);
5215 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func);
5217 SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS);
5218 #endif /* _KERNEL */
5221 * Adapt arc info given the number of bytes we are trying to add and
5222 * the state that we are coming from. This function is only called
5223 * when we are adding new content to the cache.
5226 arc_adapt(int bytes, arc_state_t *state)
5229 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
5230 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
5231 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
5233 if (state == arc_l2c_only)
5238 * Adapt the target size of the MRU list:
5239 * - if we just hit in the MRU ghost list, then increase
5240 * the target size of the MRU list.
5241 * - if we just hit in the MFU ghost list, then increase
5242 * the target size of the MFU list by decreasing the
5243 * target size of the MRU list.
5245 if (state == arc_mru_ghost) {
5246 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
5247 if (!zfs_arc_p_dampener_disable)
5248 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
5250 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
5251 } else if (state == arc_mfu_ghost) {
5254 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
5255 if (!zfs_arc_p_dampener_disable)
5256 mult = MIN(mult, 10);
5258 delta = MIN(bytes * mult, arc_p);
5259 arc_p = MAX(arc_p_min, arc_p - delta);
5261 ASSERT((int64_t)arc_p >= 0);
5263 if (arc_reclaim_needed()) {
5264 cv_signal(&arc_reclaim_thread_cv);
5271 if (arc_c >= arc_c_max)
5275 * If we're within (2 * maxblocksize) bytes of the target
5276 * cache size, increment the target cache size
5278 ASSERT3U(arc_c, >=, 2ULL << SPA_MAXBLOCKSHIFT);
5279 if (arc_size >= arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
5280 atomic_add_64(&arc_c, (int64_t)bytes);
5281 if (arc_c > arc_c_max)
5283 else if (state == arc_anon)
5284 atomic_add_64(&arc_p, (int64_t)bytes);
5288 ASSERT((int64_t)arc_p >= 0);
5292 * Check if arc_size has grown past our upper threshold, determined by
5293 * zfs_arc_overflow_shift.
5296 arc_is_overflowing(void)
5298 /* Always allow at least one block of overflow */
5299 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5300 arc_c >> zfs_arc_overflow_shift);
5302 return (arc_size >= arc_c + overflow);
5306 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5308 arc_buf_contents_t type = arc_buf_type(hdr);
5310 arc_get_data_impl(hdr, size, tag);
5311 if (type == ARC_BUFC_METADATA) {
5312 return (abd_alloc(size, B_TRUE));
5314 ASSERT(type == ARC_BUFC_DATA);
5315 return (abd_alloc(size, B_FALSE));
5320 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5322 arc_buf_contents_t type = arc_buf_type(hdr);
5324 arc_get_data_impl(hdr, size, tag);
5325 if (type == ARC_BUFC_METADATA) {
5326 return (zio_buf_alloc(size));
5328 ASSERT(type == ARC_BUFC_DATA);
5329 return (zio_data_buf_alloc(size));
5334 * Allocate a block and return it to the caller. If we are hitting the
5335 * hard limit for the cache size, we must sleep, waiting for the eviction
5336 * thread to catch up. If we're past the target size but below the hard
5337 * limit, we'll only signal the reclaim thread and continue on.
5340 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5342 arc_state_t *state = hdr->b_l1hdr.b_state;
5343 arc_buf_contents_t type = arc_buf_type(hdr);
5345 arc_adapt(size, state);
5348 * If arc_size is currently overflowing, and has grown past our
5349 * upper limit, we must be adding data faster than the evict
5350 * thread can evict. Thus, to ensure we don't compound the
5351 * problem by adding more data and forcing arc_size to grow even
5352 * further past it's target size, we halt and wait for the
5353 * eviction thread to catch up.
5355 * It's also possible that the reclaim thread is unable to evict
5356 * enough buffers to get arc_size below the overflow limit (e.g.
5357 * due to buffers being un-evictable, or hash lock collisions).
5358 * In this case, we want to proceed regardless if we're
5359 * overflowing; thus we don't use a while loop here.
5361 if (arc_is_overflowing()) {
5362 mutex_enter(&arc_reclaim_lock);
5365 * Now that we've acquired the lock, we may no longer be
5366 * over the overflow limit, lets check.
5368 * We're ignoring the case of spurious wake ups. If that
5369 * were to happen, it'd let this thread consume an ARC
5370 * buffer before it should have (i.e. before we're under
5371 * the overflow limit and were signalled by the reclaim
5372 * thread). As long as that is a rare occurrence, it
5373 * shouldn't cause any harm.
5375 if (arc_is_overflowing()) {
5376 cv_signal(&arc_reclaim_thread_cv);
5377 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5380 mutex_exit(&arc_reclaim_lock);
5383 VERIFY3U(hdr->b_type, ==, type);
5384 if (type == ARC_BUFC_METADATA) {
5385 arc_space_consume(size, ARC_SPACE_META);
5387 arc_space_consume(size, ARC_SPACE_DATA);
5391 * Update the state size. Note that ghost states have a
5392 * "ghost size" and so don't need to be updated.
5394 if (!GHOST_STATE(state)) {
5396 (void) refcount_add_many(&state->arcs_size, size, tag);
5399 * If this is reached via arc_read, the link is
5400 * protected by the hash lock. If reached via
5401 * arc_buf_alloc, the header should not be accessed by
5402 * any other thread. And, if reached via arc_read_done,
5403 * the hash lock will protect it if it's found in the
5404 * hash table; otherwise no other thread should be
5405 * trying to [add|remove]_reference it.
5407 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5408 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5409 (void) refcount_add_many(&state->arcs_esize[type],
5414 * If we are growing the cache, and we are adding anonymous
5415 * data, and we have outgrown arc_p, update arc_p
5417 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
5418 (refcount_count(&arc_anon->arcs_size) +
5419 refcount_count(&arc_mru->arcs_size) > arc_p))
5420 arc_p = MIN(arc_c, arc_p + size);
5425 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5427 arc_free_data_impl(hdr, size, tag);
5432 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5434 arc_buf_contents_t type = arc_buf_type(hdr);
5436 arc_free_data_impl(hdr, size, tag);
5437 if (type == ARC_BUFC_METADATA) {
5438 zio_buf_free(buf, size);
5440 ASSERT(type == ARC_BUFC_DATA);
5441 zio_data_buf_free(buf, size);
5446 * Free the arc data buffer.
5449 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5451 arc_state_t *state = hdr->b_l1hdr.b_state;
5452 arc_buf_contents_t type = arc_buf_type(hdr);
5454 /* protected by hash lock, if in the hash table */
5455 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5456 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5457 ASSERT(state != arc_anon && state != arc_l2c_only);
5459 (void) refcount_remove_many(&state->arcs_esize[type],
5462 (void) refcount_remove_many(&state->arcs_size, size, tag);
5464 VERIFY3U(hdr->b_type, ==, type);
5465 if (type == ARC_BUFC_METADATA) {
5466 arc_space_return(size, ARC_SPACE_META);
5468 ASSERT(type == ARC_BUFC_DATA);
5469 arc_space_return(size, ARC_SPACE_DATA);
5474 * This routine is called whenever a buffer is accessed.
5475 * NOTE: the hash lock is dropped in this function.
5478 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5482 ASSERT(MUTEX_HELD(hash_lock));
5483 ASSERT(HDR_HAS_L1HDR(hdr));
5485 if (hdr->b_l1hdr.b_state == arc_anon) {
5487 * This buffer is not in the cache, and does not
5488 * appear in our "ghost" list. Add the new buffer
5492 ASSERT0(hdr->b_l1hdr.b_arc_access);
5493 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5494 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5495 arc_change_state(arc_mru, hdr, hash_lock);
5497 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5498 now = ddi_get_lbolt();
5501 * If this buffer is here because of a prefetch, then either:
5502 * - clear the flag if this is a "referencing" read
5503 * (any subsequent access will bump this into the MFU state).
5505 * - move the buffer to the head of the list if this is
5506 * another prefetch (to make it less likely to be evicted).
5508 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5509 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5510 /* link protected by hash lock */
5511 ASSERT(multilist_link_active(
5512 &hdr->b_l1hdr.b_arc_node));
5514 arc_hdr_clear_flags(hdr,
5516 ARC_FLAG_PRESCIENT_PREFETCH);
5517 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5518 ARCSTAT_BUMP(arcstat_mru_hits);
5520 hdr->b_l1hdr.b_arc_access = now;
5525 * This buffer has been "accessed" only once so far,
5526 * but it is still in the cache. Move it to the MFU
5529 if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access +
5532 * More than 125ms have passed since we
5533 * instantiated this buffer. Move it to the
5534 * most frequently used state.
5536 hdr->b_l1hdr.b_arc_access = now;
5537 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5538 arc_change_state(arc_mfu, hdr, hash_lock);
5540 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5541 ARCSTAT_BUMP(arcstat_mru_hits);
5542 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5543 arc_state_t *new_state;
5545 * This buffer has been "accessed" recently, but
5546 * was evicted from the cache. Move it to the
5550 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5551 new_state = arc_mru;
5552 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5553 arc_hdr_clear_flags(hdr,
5555 ARC_FLAG_PRESCIENT_PREFETCH);
5557 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5559 new_state = arc_mfu;
5560 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5563 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5564 arc_change_state(new_state, hdr, hash_lock);
5566 atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits);
5567 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5568 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5570 * This buffer has been accessed more than once and is
5571 * still in the cache. Keep it in the MFU state.
5573 * NOTE: an add_reference() that occurred when we did
5574 * the arc_read() will have kicked this off the list.
5575 * If it was a prefetch, we will explicitly move it to
5576 * the head of the list now.
5579 atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits);
5580 ARCSTAT_BUMP(arcstat_mfu_hits);
5581 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5582 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5583 arc_state_t *new_state = arc_mfu;
5585 * This buffer has been accessed more than once but has
5586 * been evicted from the cache. Move it back to the
5590 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5592 * This is a prefetch access...
5593 * move this block back to the MRU state.
5595 new_state = arc_mru;
5598 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5599 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5600 arc_change_state(new_state, hdr, hash_lock);
5602 atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits);
5603 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5604 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5606 * This buffer is on the 2nd Level ARC.
5609 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5610 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5611 arc_change_state(arc_mfu, hdr, hash_lock);
5613 cmn_err(CE_PANIC, "invalid arc state 0x%p",
5614 hdr->b_l1hdr.b_state);
5619 * This routine is called by dbuf_hold() to update the arc_access() state
5620 * which otherwise would be skipped for entries in the dbuf cache.
5623 arc_buf_access(arc_buf_t *buf)
5625 mutex_enter(&buf->b_evict_lock);
5626 arc_buf_hdr_t *hdr = buf->b_hdr;
5629 * Avoid taking the hash_lock when possible as an optimization.
5630 * The header must be checked again under the hash_lock in order
5631 * to handle the case where it is concurrently being released.
5633 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5634 mutex_exit(&buf->b_evict_lock);
5638 kmutex_t *hash_lock = HDR_LOCK(hdr);
5639 mutex_enter(hash_lock);
5641 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5642 mutex_exit(hash_lock);
5643 mutex_exit(&buf->b_evict_lock);
5644 ARCSTAT_BUMP(arcstat_access_skip);
5648 mutex_exit(&buf->b_evict_lock);
5650 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5651 hdr->b_l1hdr.b_state == arc_mfu);
5653 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5654 arc_access(hdr, hash_lock);
5655 mutex_exit(hash_lock);
5657 ARCSTAT_BUMP(arcstat_hits);
5658 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr) && !HDR_PRESCIENT_PREFETCH(hdr),
5659 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5662 /* a generic arc_read_done_func_t which you can use */
5665 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5666 arc_buf_t *buf, void *arg)
5671 bcopy(buf->b_data, arg, arc_buf_size(buf));
5672 arc_buf_destroy(buf, arg);
5675 /* a generic arc_read_done_func_t */
5678 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5679 arc_buf_t *buf, void *arg)
5681 arc_buf_t **bufp = arg;
5687 ASSERT(buf->b_data);
5692 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5694 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5695 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5696 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5698 if (HDR_COMPRESSION_ENABLED(hdr)) {
5699 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5700 BP_GET_COMPRESS(bp));
5702 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5703 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5704 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5709 arc_read_done(zio_t *zio)
5711 blkptr_t *bp = zio->io_bp;
5712 arc_buf_hdr_t *hdr = zio->io_private;
5713 kmutex_t *hash_lock = NULL;
5714 arc_callback_t *callback_list;
5715 arc_callback_t *acb;
5716 boolean_t freeable = B_FALSE;
5719 * The hdr was inserted into hash-table and removed from lists
5720 * prior to starting I/O. We should find this header, since
5721 * it's in the hash table, and it should be legit since it's
5722 * not possible to evict it during the I/O. The only possible
5723 * reason for it not to be found is if we were freed during the
5726 if (HDR_IN_HASH_TABLE(hdr)) {
5727 arc_buf_hdr_t *found;
5729 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5730 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5731 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5732 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5733 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5735 found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock);
5737 ASSERT((found == hdr &&
5738 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5739 (found == hdr && HDR_L2_READING(hdr)));
5740 ASSERT3P(hash_lock, !=, NULL);
5743 if (BP_IS_PROTECTED(bp)) {
5744 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5745 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5746 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5747 hdr->b_crypt_hdr.b_iv);
5749 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5752 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5753 sizeof (zil_chain_t));
5754 zio_crypt_decode_mac_zil(tmpbuf,
5755 hdr->b_crypt_hdr.b_mac);
5756 abd_return_buf(zio->io_abd, tmpbuf,
5757 sizeof (zil_chain_t));
5759 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
5763 if (zio->io_error == 0) {
5764 /* byteswap if necessary */
5765 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5766 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5767 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5769 hdr->b_l1hdr.b_byteswap =
5770 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5773 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5777 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5778 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5779 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5781 callback_list = hdr->b_l1hdr.b_acb;
5782 ASSERT3P(callback_list, !=, NULL);
5784 if (hash_lock && zio->io_error == 0 &&
5785 hdr->b_l1hdr.b_state == arc_anon) {
5787 * Only call arc_access on anonymous buffers. This is because
5788 * if we've issued an I/O for an evicted buffer, we've already
5789 * called arc_access (to prevent any simultaneous readers from
5790 * getting confused).
5792 arc_access(hdr, hash_lock);
5796 * If a read request has a callback (i.e. acb_done is not NULL), then we
5797 * make a buf containing the data according to the parameters which were
5798 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5799 * aren't needlessly decompressing the data multiple times.
5801 int callback_cnt = 0;
5802 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5808 if (zio->io_error != 0)
5811 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5812 acb->acb_dsobj, acb->acb_private, acb->acb_encrypted,
5813 acb->acb_compressed, acb->acb_noauth, B_TRUE,
5816 arc_buf_destroy(acb->acb_buf, acb->acb_private);
5817 acb->acb_buf = NULL;
5821 * Assert non-speculative zios didn't fail because an
5822 * encryption key wasn't loaded
5824 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) || error == 0);
5827 * If we failed to decrypt, report an error now (as the zio
5828 * layer would have done if it had done the transforms).
5830 if (error == ECKSUM) {
5831 ASSERT(BP_IS_PROTECTED(bp));
5832 error = SET_ERROR(EIO);
5833 spa_log_error(zio->io_spa, &zio->io_bookmark);
5834 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5835 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
5836 zio->io_spa, NULL, &zio->io_bookmark, zio,
5841 if (zio->io_error == 0)
5842 zio->io_error = error;
5844 hdr->b_l1hdr.b_acb = NULL;
5845 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5846 if (callback_cnt == 0)
5847 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
5849 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5850 callback_list != NULL);
5852 if (zio->io_error == 0) {
5853 arc_hdr_verify(hdr, zio->io_bp);
5855 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5856 if (hdr->b_l1hdr.b_state != arc_anon)
5857 arc_change_state(arc_anon, hdr, hash_lock);
5858 if (HDR_IN_HASH_TABLE(hdr))
5859 buf_hash_remove(hdr);
5860 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5864 * Broadcast before we drop the hash_lock to avoid the possibility
5865 * that the hdr (and hence the cv) might be freed before we get to
5866 * the cv_broadcast().
5868 cv_broadcast(&hdr->b_l1hdr.b_cv);
5870 if (hash_lock != NULL) {
5871 mutex_exit(hash_lock);
5874 * This block was freed while we waited for the read to
5875 * complete. It has been removed from the hash table and
5876 * moved to the anonymous state (so that it won't show up
5879 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5880 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5883 /* execute each callback and free its structure */
5884 while ((acb = callback_list) != NULL) {
5885 if (acb->acb_done) {
5886 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5887 acb->acb_buf, acb->acb_private);
5890 if (acb->acb_zio_dummy != NULL) {
5891 acb->acb_zio_dummy->io_error = zio->io_error;
5892 zio_nowait(acb->acb_zio_dummy);
5895 callback_list = acb->acb_next;
5896 kmem_free(acb, sizeof (arc_callback_t));
5900 arc_hdr_destroy(hdr);
5904 * "Read" the block at the specified DVA (in bp) via the
5905 * cache. If the block is found in the cache, invoke the provided
5906 * callback immediately and return. Note that the `zio' parameter
5907 * in the callback will be NULL in this case, since no IO was
5908 * required. If the block is not in the cache pass the read request
5909 * on to the spa with a substitute callback function, so that the
5910 * requested block will be added to the cache.
5912 * If a read request arrives for a block that has a read in-progress,
5913 * either wait for the in-progress read to complete (and return the
5914 * results); or, if this is a read with a "done" func, add a record
5915 * to the read to invoke the "done" func when the read completes,
5916 * and return; or just return.
5918 * arc_read_done() will invoke all the requested "done" functions
5919 * for readers of this block.
5922 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
5923 arc_read_done_func_t *done, void *private, zio_priority_t priority,
5924 int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5926 arc_buf_hdr_t *hdr = NULL;
5927 kmutex_t *hash_lock = NULL;
5929 uint64_t guid = spa_load_guid(spa);
5930 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
5931 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
5932 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5933 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
5934 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5937 ASSERT(!BP_IS_EMBEDDED(bp) ||
5938 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5941 if (!BP_IS_EMBEDDED(bp)) {
5943 * Embedded BP's have no DVA and require no I/O to "read".
5944 * Create an anonymous arc buf to back it.
5946 hdr = buf_hash_find(guid, bp, &hash_lock);
5950 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5951 * we maintain encrypted data seperately from compressed / uncompressed
5952 * data. If the user is requesting raw encrypted data and we don't have
5953 * that in the header we will read from disk to guarantee that we can
5954 * get it even if the encryption keys aren't loaded.
5956 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
5957 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
5958 arc_buf_t *buf = NULL;
5959 *arc_flags |= ARC_FLAG_CACHED;
5961 if (HDR_IO_IN_PROGRESS(hdr)) {
5962 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5964 ASSERT3P(head_zio, !=, NULL);
5965 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5966 priority == ZIO_PRIORITY_SYNC_READ) {
5968 * This is a sync read that needs to wait for
5969 * an in-flight async read. Request that the
5970 * zio have its priority upgraded.
5972 zio_change_priority(head_zio, priority);
5973 DTRACE_PROBE1(arc__async__upgrade__sync,
5974 arc_buf_hdr_t *, hdr);
5975 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5977 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5978 arc_hdr_clear_flags(hdr,
5979 ARC_FLAG_PREDICTIVE_PREFETCH);
5982 if (*arc_flags & ARC_FLAG_WAIT) {
5983 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5984 mutex_exit(hash_lock);
5987 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5990 arc_callback_t *acb = NULL;
5992 acb = kmem_zalloc(sizeof (arc_callback_t),
5994 acb->acb_done = done;
5995 acb->acb_private = private;
5996 acb->acb_compressed = compressed_read;
5997 acb->acb_encrypted = encrypted_read;
5998 acb->acb_noauth = noauth_read;
5999 acb->acb_dsobj = zb->zb_objset;
6001 acb->acb_zio_dummy = zio_null(pio,
6002 spa, NULL, NULL, NULL, zio_flags);
6004 ASSERT3P(acb->acb_done, !=, NULL);
6005 acb->acb_zio_head = head_zio;
6006 acb->acb_next = hdr->b_l1hdr.b_acb;
6007 hdr->b_l1hdr.b_acb = acb;
6008 mutex_exit(hash_lock);
6011 mutex_exit(hash_lock);
6015 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
6016 hdr->b_l1hdr.b_state == arc_mfu);
6019 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
6021 * This is a demand read which does not have to
6022 * wait for i/o because we did a predictive
6023 * prefetch i/o for it, which has completed.
6026 arc__demand__hit__predictive__prefetch,
6027 arc_buf_hdr_t *, hdr);
6029 arcstat_demand_hit_predictive_prefetch);
6030 arc_hdr_clear_flags(hdr,
6031 ARC_FLAG_PREDICTIVE_PREFETCH);
6034 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
6036 arcstat_demand_hit_prescient_prefetch);
6037 arc_hdr_clear_flags(hdr,
6038 ARC_FLAG_PRESCIENT_PREFETCH);
6041 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
6043 /* Get a buf with the desired data in it. */
6044 rc = arc_buf_alloc_impl(hdr, spa, zb->zb_objset,
6045 private, encrypted_read, compressed_read,
6046 noauth_read, B_TRUE, &buf);
6048 arc_buf_destroy(buf, private);
6052 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) || rc == 0);
6053 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
6054 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
6055 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6057 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
6058 arc_access(hdr, hash_lock);
6059 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6060 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6061 if (*arc_flags & ARC_FLAG_L2CACHE)
6062 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6063 mutex_exit(hash_lock);
6064 ARCSTAT_BUMP(arcstat_hits);
6065 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6066 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
6067 data, metadata, hits);
6070 done(NULL, zb, bp, buf, private);
6072 uint64_t lsize = BP_GET_LSIZE(bp);
6073 uint64_t psize = BP_GET_PSIZE(bp);
6074 arc_callback_t *acb;
6077 boolean_t devw = B_FALSE;
6082 * Gracefully handle a damaged logical block size as a
6085 if (lsize > spa_maxblocksize(spa)) {
6086 rc = SET_ERROR(ECKSUM);
6091 /* this block is not in the cache */
6092 arc_buf_hdr_t *exists = NULL;
6093 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
6094 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
6095 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), type,
6098 if (!BP_IS_EMBEDDED(bp)) {
6099 hdr->b_dva = *BP_IDENTITY(bp);
6100 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
6101 exists = buf_hash_insert(hdr, &hash_lock);
6103 if (exists != NULL) {
6104 /* somebody beat us to the hash insert */
6105 mutex_exit(hash_lock);
6106 buf_discard_identity(hdr);
6107 arc_hdr_destroy(hdr);
6108 goto top; /* restart the IO request */
6112 * This block is in the ghost cache or encrypted data
6113 * was requested and we didn't have it. If it was
6114 * L2-only (and thus didn't have an L1 hdr),
6115 * we realloc the header to add an L1 hdr.
6117 if (!HDR_HAS_L1HDR(hdr)) {
6118 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
6122 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
6123 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6124 ASSERT(!HDR_HAS_RABD(hdr));
6125 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6126 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
6127 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
6128 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
6129 } else if (HDR_IO_IN_PROGRESS(hdr)) {
6131 * If this header already had an IO in progress
6132 * and we are performing another IO to fetch
6133 * encrypted data we must wait until the first
6134 * IO completes so as not to confuse
6135 * arc_read_done(). This should be very rare
6136 * and so the performance impact shouldn't
6139 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6140 mutex_exit(hash_lock);
6145 * This is a delicate dance that we play here.
6146 * This hdr might be in the ghost list so we access
6147 * it to move it out of the ghost list before we
6148 * initiate the read. If it's a prefetch then
6149 * it won't have a callback so we'll remove the
6150 * reference that arc_buf_alloc_impl() created. We
6151 * do this after we've called arc_access() to
6152 * avoid hitting an assert in remove_reference().
6154 arc_access(hdr, hash_lock);
6155 arc_hdr_alloc_abd(hdr, encrypted_read);
6158 if (encrypted_read) {
6159 ASSERT(HDR_HAS_RABD(hdr));
6160 size = HDR_GET_PSIZE(hdr);
6161 hdr_abd = hdr->b_crypt_hdr.b_rabd;
6162 zio_flags |= ZIO_FLAG_RAW;
6164 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6165 size = arc_hdr_size(hdr);
6166 hdr_abd = hdr->b_l1hdr.b_pabd;
6168 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
6169 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6173 * For authenticated bp's, we do not ask the ZIO layer
6174 * to authenticate them since this will cause the entire
6175 * IO to fail if the key isn't loaded. Instead, we
6176 * defer authentication until arc_buf_fill(), which will
6177 * verify the data when the key is available.
6179 if (BP_IS_AUTHENTICATED(bp))
6180 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
6183 if (*arc_flags & ARC_FLAG_PREFETCH &&
6184 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))
6185 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6186 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6187 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6188 if (*arc_flags & ARC_FLAG_L2CACHE)
6189 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6190 if (BP_IS_AUTHENTICATED(bp))
6191 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6192 if (BP_GET_LEVEL(bp) > 0)
6193 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
6194 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
6195 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
6196 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
6198 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
6199 acb->acb_done = done;
6200 acb->acb_private = private;
6201 acb->acb_compressed = compressed_read;
6202 acb->acb_encrypted = encrypted_read;
6203 acb->acb_noauth = noauth_read;
6204 acb->acb_dsobj = zb->zb_objset;
6206 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6207 hdr->b_l1hdr.b_acb = acb;
6208 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6210 if (HDR_HAS_L2HDR(hdr) &&
6211 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
6212 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
6213 addr = hdr->b_l2hdr.b_daddr;
6215 * Lock out device removal.
6217 if (vdev_is_dead(vd) ||
6218 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
6223 * We count both async reads and scrub IOs as asynchronous so
6224 * that both can be upgraded in the event of a cache hit while
6225 * the read IO is still in-flight.
6227 if (priority == ZIO_PRIORITY_ASYNC_READ ||
6228 priority == ZIO_PRIORITY_SCRUB)
6229 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6231 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6234 * At this point, we have a level 1 cache miss. Try again in
6235 * L2ARC if possible.
6237 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
6239 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
6240 uint64_t, lsize, zbookmark_phys_t *, zb);
6241 ARCSTAT_BUMP(arcstat_misses);
6242 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6243 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
6244 data, metadata, misses);
6246 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
6248 * Read from the L2ARC if the following are true:
6249 * 1. The L2ARC vdev was previously cached.
6250 * 2. This buffer still has L2ARC metadata.
6251 * 3. This buffer isn't currently writing to the L2ARC.
6252 * 4. The L2ARC entry wasn't evicted, which may
6253 * also have invalidated the vdev.
6254 * 5. This isn't prefetch and l2arc_noprefetch is set.
6256 if (HDR_HAS_L2HDR(hdr) &&
6257 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
6258 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
6259 l2arc_read_callback_t *cb;
6263 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
6264 ARCSTAT_BUMP(arcstat_l2_hits);
6265 atomic_inc_32(&hdr->b_l2hdr.b_hits);
6267 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
6269 cb->l2rcb_hdr = hdr;
6272 cb->l2rcb_flags = zio_flags;
6274 asize = vdev_psize_to_asize(vd, size);
6275 if (asize != size) {
6276 abd = abd_alloc_for_io(asize,
6277 HDR_ISTYPE_METADATA(hdr));
6278 cb->l2rcb_abd = abd;
6283 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6284 addr + asize <= vd->vdev_psize -
6285 VDEV_LABEL_END_SIZE);
6288 * l2arc read. The SCL_L2ARC lock will be
6289 * released by l2arc_read_done().
6290 * Issue a null zio if the underlying buffer
6291 * was squashed to zero size by compression.
6293 ASSERT3U(arc_hdr_get_compress(hdr), !=,
6294 ZIO_COMPRESS_EMPTY);
6295 rzio = zio_read_phys(pio, vd, addr,
6298 l2arc_read_done, cb, priority,
6299 zio_flags | ZIO_FLAG_DONT_CACHE |
6301 ZIO_FLAG_DONT_PROPAGATE |
6302 ZIO_FLAG_DONT_RETRY, B_FALSE);
6303 acb->acb_zio_head = rzio;
6305 if (hash_lock != NULL)
6306 mutex_exit(hash_lock);
6308 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6310 ARCSTAT_INCR(arcstat_l2_read_bytes,
6311 HDR_GET_PSIZE(hdr));
6313 if (*arc_flags & ARC_FLAG_NOWAIT) {
6318 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6319 if (zio_wait(rzio) == 0)
6322 /* l2arc read error; goto zio_read() */
6323 if (hash_lock != NULL)
6324 mutex_enter(hash_lock);
6326 DTRACE_PROBE1(l2arc__miss,
6327 arc_buf_hdr_t *, hdr);
6328 ARCSTAT_BUMP(arcstat_l2_misses);
6329 if (HDR_L2_WRITING(hdr))
6330 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6331 spa_config_exit(spa, SCL_L2ARC, vd);
6335 spa_config_exit(spa, SCL_L2ARC, vd);
6336 if (l2arc_ndev != 0) {
6337 DTRACE_PROBE1(l2arc__miss,
6338 arc_buf_hdr_t *, hdr);
6339 ARCSTAT_BUMP(arcstat_l2_misses);
6343 rzio = zio_read(pio, spa, bp, hdr_abd, size,
6344 arc_read_done, hdr, priority, zio_flags, zb);
6345 acb->acb_zio_head = rzio;
6347 if (hash_lock != NULL)
6348 mutex_exit(hash_lock);
6350 if (*arc_flags & ARC_FLAG_WAIT) {
6351 rc = zio_wait(rzio);
6355 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6360 spa_read_history_add(spa, zb, *arc_flags);
6365 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6369 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6371 p->p_private = private;
6372 list_link_init(&p->p_node);
6373 refcount_create(&p->p_refcnt);
6375 mutex_enter(&arc_prune_mtx);
6376 refcount_add(&p->p_refcnt, &arc_prune_list);
6377 list_insert_head(&arc_prune_list, p);
6378 mutex_exit(&arc_prune_mtx);
6384 arc_remove_prune_callback(arc_prune_t *p)
6386 boolean_t wait = B_FALSE;
6387 mutex_enter(&arc_prune_mtx);
6388 list_remove(&arc_prune_list, p);
6389 if (refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6391 mutex_exit(&arc_prune_mtx);
6393 /* wait for arc_prune_task to finish */
6395 taskq_wait_outstanding(arc_prune_taskq, 0);
6396 ASSERT0(refcount_count(&p->p_refcnt));
6397 refcount_destroy(&p->p_refcnt);
6398 kmem_free(p, sizeof (*p));
6402 * Notify the arc that a block was freed, and thus will never be used again.
6405 arc_freed(spa_t *spa, const blkptr_t *bp)
6408 kmutex_t *hash_lock;
6409 uint64_t guid = spa_load_guid(spa);
6411 ASSERT(!BP_IS_EMBEDDED(bp));
6413 hdr = buf_hash_find(guid, bp, &hash_lock);
6418 * We might be trying to free a block that is still doing I/O
6419 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6420 * dmu_sync-ed block). If this block is being prefetched, then it
6421 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6422 * until the I/O completes. A block may also have a reference if it is
6423 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6424 * have written the new block to its final resting place on disk but
6425 * without the dedup flag set. This would have left the hdr in the MRU
6426 * state and discoverable. When the txg finally syncs it detects that
6427 * the block was overridden in open context and issues an override I/O.
6428 * Since this is a dedup block, the override I/O will determine if the
6429 * block is already in the DDT. If so, then it will replace the io_bp
6430 * with the bp from the DDT and allow the I/O to finish. When the I/O
6431 * reaches the done callback, dbuf_write_override_done, it will
6432 * check to see if the io_bp and io_bp_override are identical.
6433 * If they are not, then it indicates that the bp was replaced with
6434 * the bp in the DDT and the override bp is freed. This allows
6435 * us to arrive here with a reference on a block that is being
6436 * freed. So if we have an I/O in progress, or a reference to
6437 * this hdr, then we don't destroy the hdr.
6439 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6440 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6441 arc_change_state(arc_anon, hdr, hash_lock);
6442 arc_hdr_destroy(hdr);
6443 mutex_exit(hash_lock);
6445 mutex_exit(hash_lock);
6451 * Release this buffer from the cache, making it an anonymous buffer. This
6452 * must be done after a read and prior to modifying the buffer contents.
6453 * If the buffer has more than one reference, we must make
6454 * a new hdr for the buffer.
6457 arc_release(arc_buf_t *buf, void *tag)
6459 arc_buf_hdr_t *hdr = buf->b_hdr;
6462 * It would be nice to assert that if its DMU metadata (level >
6463 * 0 || it's the dnode file), then it must be syncing context.
6464 * But we don't know that information at this level.
6467 mutex_enter(&buf->b_evict_lock);
6469 ASSERT(HDR_HAS_L1HDR(hdr));
6472 * We don't grab the hash lock prior to this check, because if
6473 * the buffer's header is in the arc_anon state, it won't be
6474 * linked into the hash table.
6476 if (hdr->b_l1hdr.b_state == arc_anon) {
6477 mutex_exit(&buf->b_evict_lock);
6478 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6479 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6480 ASSERT(!HDR_HAS_L2HDR(hdr));
6481 ASSERT(HDR_EMPTY(hdr));
6483 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6484 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6485 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6487 hdr->b_l1hdr.b_arc_access = 0;
6490 * If the buf is being overridden then it may already
6491 * have a hdr that is not empty.
6493 buf_discard_identity(hdr);
6499 kmutex_t *hash_lock = HDR_LOCK(hdr);
6500 mutex_enter(hash_lock);
6503 * This assignment is only valid as long as the hash_lock is
6504 * held, we must be careful not to reference state or the
6505 * b_state field after dropping the lock.
6507 arc_state_t *state = hdr->b_l1hdr.b_state;
6508 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6509 ASSERT3P(state, !=, arc_anon);
6511 /* this buffer is not on any list */
6512 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6514 if (HDR_HAS_L2HDR(hdr)) {
6515 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6518 * We have to recheck this conditional again now that
6519 * we're holding the l2ad_mtx to prevent a race with
6520 * another thread which might be concurrently calling
6521 * l2arc_evict(). In that case, l2arc_evict() might have
6522 * destroyed the header's L2 portion as we were waiting
6523 * to acquire the l2ad_mtx.
6525 if (HDR_HAS_L2HDR(hdr))
6526 arc_hdr_l2hdr_destroy(hdr);
6528 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6532 * Do we have more than one buf?
6534 if (hdr->b_l1hdr.b_bufcnt > 1) {
6535 arc_buf_hdr_t *nhdr;
6536 uint64_t spa = hdr->b_spa;
6537 uint64_t psize = HDR_GET_PSIZE(hdr);
6538 uint64_t lsize = HDR_GET_LSIZE(hdr);
6539 boolean_t protected = HDR_PROTECTED(hdr);
6540 enum zio_compress compress = arc_hdr_get_compress(hdr);
6541 arc_buf_contents_t type = arc_buf_type(hdr);
6542 VERIFY3U(hdr->b_type, ==, type);
6544 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6545 (void) remove_reference(hdr, hash_lock, tag);
6547 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6548 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6549 ASSERT(ARC_BUF_LAST(buf));
6553 * Pull the data off of this hdr and attach it to
6554 * a new anonymous hdr. Also find the last buffer
6555 * in the hdr's buffer list.
6557 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6558 ASSERT3P(lastbuf, !=, NULL);
6561 * If the current arc_buf_t and the hdr are sharing their data
6562 * buffer, then we must stop sharing that block.
6564 if (arc_buf_is_shared(buf)) {
6565 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6566 VERIFY(!arc_buf_is_shared(lastbuf));
6569 * First, sever the block sharing relationship between
6570 * buf and the arc_buf_hdr_t.
6572 arc_unshare_buf(hdr, buf);
6575 * Now we need to recreate the hdr's b_pabd. Since we
6576 * have lastbuf handy, we try to share with it, but if
6577 * we can't then we allocate a new b_pabd and copy the
6578 * data from buf into it.
6580 if (arc_can_share(hdr, lastbuf)) {
6581 arc_share_buf(hdr, lastbuf);
6583 arc_hdr_alloc_abd(hdr, B_FALSE);
6584 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6585 buf->b_data, psize);
6587 VERIFY3P(lastbuf->b_data, !=, NULL);
6588 } else if (HDR_SHARED_DATA(hdr)) {
6590 * Uncompressed shared buffers are always at the end
6591 * of the list. Compressed buffers don't have the
6592 * same requirements. This makes it hard to
6593 * simply assert that the lastbuf is shared so
6594 * we rely on the hdr's compression flags to determine
6595 * if we have a compressed, shared buffer.
6597 ASSERT(arc_buf_is_shared(lastbuf) ||
6598 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6599 ASSERT(!ARC_BUF_SHARED(buf));
6602 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6603 ASSERT3P(state, !=, arc_l2c_only);
6605 (void) refcount_remove_many(&state->arcs_size,
6606 arc_buf_size(buf), buf);
6608 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6609 ASSERT3P(state, !=, arc_l2c_only);
6610 (void) refcount_remove_many(&state->arcs_esize[type],
6611 arc_buf_size(buf), buf);
6614 hdr->b_l1hdr.b_bufcnt -= 1;
6615 if (ARC_BUF_ENCRYPTED(buf))
6616 hdr->b_crypt_hdr.b_ebufcnt -= 1;
6618 arc_cksum_verify(buf);
6619 arc_buf_unwatch(buf);
6621 /* if this is the last uncompressed buf free the checksum */
6622 if (!arc_hdr_has_uncompressed_buf(hdr))
6623 arc_cksum_free(hdr);
6625 mutex_exit(hash_lock);
6628 * Allocate a new hdr. The new hdr will contain a b_pabd
6629 * buffer which will be freed in arc_write().
6631 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6632 compress, type, HDR_HAS_RABD(hdr));
6633 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6634 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6635 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
6636 VERIFY3U(nhdr->b_type, ==, type);
6637 ASSERT(!HDR_SHARED_DATA(nhdr));
6639 nhdr->b_l1hdr.b_buf = buf;
6640 nhdr->b_l1hdr.b_bufcnt = 1;
6641 if (ARC_BUF_ENCRYPTED(buf))
6642 nhdr->b_crypt_hdr.b_ebufcnt = 1;
6643 nhdr->b_l1hdr.b_mru_hits = 0;
6644 nhdr->b_l1hdr.b_mru_ghost_hits = 0;
6645 nhdr->b_l1hdr.b_mfu_hits = 0;
6646 nhdr->b_l1hdr.b_mfu_ghost_hits = 0;
6647 nhdr->b_l1hdr.b_l2_hits = 0;
6648 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6651 mutex_exit(&buf->b_evict_lock);
6652 (void) refcount_add_many(&arc_anon->arcs_size,
6653 HDR_GET_LSIZE(nhdr), buf);
6655 mutex_exit(&buf->b_evict_lock);
6656 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6657 /* protected by hash lock, or hdr is on arc_anon */
6658 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6659 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6660 hdr->b_l1hdr.b_mru_hits = 0;
6661 hdr->b_l1hdr.b_mru_ghost_hits = 0;
6662 hdr->b_l1hdr.b_mfu_hits = 0;
6663 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
6664 hdr->b_l1hdr.b_l2_hits = 0;
6665 arc_change_state(arc_anon, hdr, hash_lock);
6666 hdr->b_l1hdr.b_arc_access = 0;
6668 mutex_exit(hash_lock);
6669 buf_discard_identity(hdr);
6675 arc_released(arc_buf_t *buf)
6679 mutex_enter(&buf->b_evict_lock);
6680 released = (buf->b_data != NULL &&
6681 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6682 mutex_exit(&buf->b_evict_lock);
6688 arc_referenced(arc_buf_t *buf)
6692 mutex_enter(&buf->b_evict_lock);
6693 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6694 mutex_exit(&buf->b_evict_lock);
6695 return (referenced);
6700 arc_write_ready(zio_t *zio)
6702 arc_write_callback_t *callback = zio->io_private;
6703 arc_buf_t *buf = callback->awcb_buf;
6704 arc_buf_hdr_t *hdr = buf->b_hdr;
6705 blkptr_t *bp = zio->io_bp;
6706 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6707 fstrans_cookie_t cookie = spl_fstrans_mark();
6709 ASSERT(HDR_HAS_L1HDR(hdr));
6710 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6711 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6714 * If we're reexecuting this zio because the pool suspended, then
6715 * cleanup any state that was previously set the first time the
6716 * callback was invoked.
6718 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6719 arc_cksum_free(hdr);
6720 arc_buf_unwatch(buf);
6721 if (hdr->b_l1hdr.b_pabd != NULL) {
6722 if (arc_buf_is_shared(buf)) {
6723 arc_unshare_buf(hdr, buf);
6725 arc_hdr_free_abd(hdr, B_FALSE);
6729 if (HDR_HAS_RABD(hdr))
6730 arc_hdr_free_abd(hdr, B_TRUE);
6732 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6733 ASSERT(!HDR_HAS_RABD(hdr));
6734 ASSERT(!HDR_SHARED_DATA(hdr));
6735 ASSERT(!arc_buf_is_shared(buf));
6737 callback->awcb_ready(zio, buf, callback->awcb_private);
6739 if (HDR_IO_IN_PROGRESS(hdr))
6740 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6742 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6744 if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr))
6745 hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp));
6747 if (BP_IS_PROTECTED(bp)) {
6748 /* ZIL blocks are written through zio_rewrite */
6749 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6750 ASSERT(HDR_PROTECTED(hdr));
6752 if (BP_SHOULD_BYTESWAP(bp)) {
6753 if (BP_GET_LEVEL(bp) > 0) {
6754 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6756 hdr->b_l1hdr.b_byteswap =
6757 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6760 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6763 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6764 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6765 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6766 hdr->b_crypt_hdr.b_iv);
6767 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6771 * If this block was written for raw encryption but the zio layer
6772 * ended up only authenticating it, adjust the buffer flags now.
6774 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6775 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6776 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6777 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6778 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6779 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6780 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6781 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6784 /* this must be done after the buffer flags are adjusted */
6785 arc_cksum_compute(buf);
6787 enum zio_compress compress;
6788 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6789 compress = ZIO_COMPRESS_OFF;
6791 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6792 compress = BP_GET_COMPRESS(bp);
6794 HDR_SET_PSIZE(hdr, psize);
6795 arc_hdr_set_compress(hdr, compress);
6797 if (zio->io_error != 0 || psize == 0)
6801 * Fill the hdr with data. If the buffer is encrypted we have no choice
6802 * but to copy the data into b_radb. If the hdr is compressed, the data
6803 * we want is available from the zio, otherwise we can take it from
6806 * We might be able to share the buf's data with the hdr here. However,
6807 * doing so would cause the ARC to be full of linear ABDs if we write a
6808 * lot of shareable data. As a compromise, we check whether scattered
6809 * ABDs are allowed, and assume that if they are then the user wants
6810 * the ARC to be primarily filled with them regardless of the data being
6811 * written. Therefore, if they're allowed then we allocate one and copy
6812 * the data into it; otherwise, we share the data directly if we can.
6814 if (ARC_BUF_ENCRYPTED(buf)) {
6815 ASSERT3U(psize, >, 0);
6816 ASSERT(ARC_BUF_COMPRESSED(buf));
6817 arc_hdr_alloc_abd(hdr, B_TRUE);
6818 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6819 } else if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6821 * Ideally, we would always copy the io_abd into b_pabd, but the
6822 * user may have disabled compressed ARC, thus we must check the
6823 * hdr's compression setting rather than the io_bp's.
6825 if (BP_IS_ENCRYPTED(bp)) {
6826 ASSERT3U(psize, >, 0);
6827 arc_hdr_alloc_abd(hdr, B_TRUE);
6828 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6829 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
6830 !ARC_BUF_COMPRESSED(buf)) {
6831 ASSERT3U(psize, >, 0);
6832 arc_hdr_alloc_abd(hdr, B_FALSE);
6833 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6835 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6836 arc_hdr_alloc_abd(hdr, B_FALSE);
6837 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6841 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6842 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6843 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6845 arc_share_buf(hdr, buf);
6849 arc_hdr_verify(hdr, bp);
6850 spl_fstrans_unmark(cookie);
6854 arc_write_children_ready(zio_t *zio)
6856 arc_write_callback_t *callback = zio->io_private;
6857 arc_buf_t *buf = callback->awcb_buf;
6859 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6863 * The SPA calls this callback for each physical write that happens on behalf
6864 * of a logical write. See the comment in dbuf_write_physdone() for details.
6867 arc_write_physdone(zio_t *zio)
6869 arc_write_callback_t *cb = zio->io_private;
6870 if (cb->awcb_physdone != NULL)
6871 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6875 arc_write_done(zio_t *zio)
6877 arc_write_callback_t *callback = zio->io_private;
6878 arc_buf_t *buf = callback->awcb_buf;
6879 arc_buf_hdr_t *hdr = buf->b_hdr;
6881 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6883 if (zio->io_error == 0) {
6884 arc_hdr_verify(hdr, zio->io_bp);
6886 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6887 buf_discard_identity(hdr);
6889 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6890 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6893 ASSERT(HDR_EMPTY(hdr));
6897 * If the block to be written was all-zero or compressed enough to be
6898 * embedded in the BP, no write was performed so there will be no
6899 * dva/birth/checksum. The buffer must therefore remain anonymous
6902 if (!HDR_EMPTY(hdr)) {
6903 arc_buf_hdr_t *exists;
6904 kmutex_t *hash_lock;
6906 ASSERT3U(zio->io_error, ==, 0);
6908 arc_cksum_verify(buf);
6910 exists = buf_hash_insert(hdr, &hash_lock);
6911 if (exists != NULL) {
6913 * This can only happen if we overwrite for
6914 * sync-to-convergence, because we remove
6915 * buffers from the hash table when we arc_free().
6917 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6918 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6919 panic("bad overwrite, hdr=%p exists=%p",
6920 (void *)hdr, (void *)exists);
6921 ASSERT(refcount_is_zero(
6922 &exists->b_l1hdr.b_refcnt));
6923 arc_change_state(arc_anon, exists, hash_lock);
6924 mutex_exit(hash_lock);
6925 arc_hdr_destroy(exists);
6926 exists = buf_hash_insert(hdr, &hash_lock);
6927 ASSERT3P(exists, ==, NULL);
6928 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6930 ASSERT(zio->io_prop.zp_nopwrite);
6931 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6932 panic("bad nopwrite, hdr=%p exists=%p",
6933 (void *)hdr, (void *)exists);
6936 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6937 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6938 ASSERT(BP_GET_DEDUP(zio->io_bp));
6939 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6942 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6943 /* if it's not anon, we are doing a scrub */
6944 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6945 arc_access(hdr, hash_lock);
6946 mutex_exit(hash_lock);
6948 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6951 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6952 callback->awcb_done(zio, buf, callback->awcb_private);
6954 abd_put(zio->io_abd);
6955 kmem_free(callback, sizeof (arc_write_callback_t));
6959 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
6960 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc,
6961 const zio_prop_t *zp, arc_write_done_func_t *ready,
6962 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
6963 arc_write_done_func_t *done, void *private, zio_priority_t priority,
6964 int zio_flags, const zbookmark_phys_t *zb)
6966 arc_buf_hdr_t *hdr = buf->b_hdr;
6967 arc_write_callback_t *callback;
6969 zio_prop_t localprop = *zp;
6971 ASSERT3P(ready, !=, NULL);
6972 ASSERT3P(done, !=, NULL);
6973 ASSERT(!HDR_IO_ERROR(hdr));
6974 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6975 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6976 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6978 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6980 if (ARC_BUF_ENCRYPTED(buf)) {
6981 ASSERT(ARC_BUF_COMPRESSED(buf));
6982 localprop.zp_encrypt = B_TRUE;
6983 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6984 localprop.zp_byteorder =
6985 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
6986 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
6987 bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt,
6989 bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv,
6991 bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac,
6993 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
6994 localprop.zp_nopwrite = B_FALSE;
6995 localprop.zp_copies =
6996 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
6998 zio_flags |= ZIO_FLAG_RAW;
6999 } else if (ARC_BUF_COMPRESSED(buf)) {
7000 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
7001 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
7002 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
7004 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
7005 callback->awcb_ready = ready;
7006 callback->awcb_children_ready = children_ready;
7007 callback->awcb_physdone = physdone;
7008 callback->awcb_done = done;
7009 callback->awcb_private = private;
7010 callback->awcb_buf = buf;
7013 * The hdr's b_pabd is now stale, free it now. A new data block
7014 * will be allocated when the zio pipeline calls arc_write_ready().
7016 if (hdr->b_l1hdr.b_pabd != NULL) {
7018 * If the buf is currently sharing the data block with
7019 * the hdr then we need to break that relationship here.
7020 * The hdr will remain with a NULL data pointer and the
7021 * buf will take sole ownership of the block.
7023 if (arc_buf_is_shared(buf)) {
7024 arc_unshare_buf(hdr, buf);
7026 arc_hdr_free_abd(hdr, B_FALSE);
7028 VERIFY3P(buf->b_data, !=, NULL);
7031 if (HDR_HAS_RABD(hdr))
7032 arc_hdr_free_abd(hdr, B_TRUE);
7034 if (!(zio_flags & ZIO_FLAG_RAW))
7035 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
7037 ASSERT(!arc_buf_is_shared(buf));
7038 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
7040 zio = zio_write(pio, spa, txg, bp,
7041 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
7042 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
7043 (children_ready != NULL) ? arc_write_children_ready : NULL,
7044 arc_write_physdone, arc_write_done, callback,
7045 priority, zio_flags, zb);
7051 arc_memory_throttle(uint64_t reserve, uint64_t txg)
7054 uint64_t available_memory = arc_free_memory();
7055 static uint64_t page_load = 0;
7056 static uint64_t last_txg = 0;
7060 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
7063 if (available_memory > arc_all_memory() * arc_lotsfree_percent / 100)
7066 if (txg > last_txg) {
7071 * If we are in pageout, we know that memory is already tight,
7072 * the arc is already going to be evicting, so we just want to
7073 * continue to let page writes occur as quickly as possible.
7075 if (current_is_kswapd()) {
7076 if (page_load > MAX(arc_sys_free / 4, available_memory) / 4) {
7077 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
7078 return (SET_ERROR(ERESTART));
7080 /* Note: reserve is inflated, so we deflate */
7081 page_load += reserve / 8;
7083 } else if (page_load > 0 && arc_reclaim_needed()) {
7084 /* memory is low, delay before restarting */
7085 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
7086 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
7087 return (SET_ERROR(EAGAIN));
7095 arc_tempreserve_clear(uint64_t reserve)
7097 atomic_add_64(&arc_tempreserve, -reserve);
7098 ASSERT((int64_t)arc_tempreserve >= 0);
7102 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
7108 reserve > arc_c/4 &&
7109 reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT))
7110 arc_c = MIN(arc_c_max, reserve * 4);
7113 * Throttle when the calculated memory footprint for the TXG
7114 * exceeds the target ARC size.
7116 if (reserve > arc_c) {
7117 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
7118 return (SET_ERROR(ERESTART));
7122 * Don't count loaned bufs as in flight dirty data to prevent long
7123 * network delays from blocking transactions that are ready to be
7124 * assigned to a txg.
7127 /* assert that it has not wrapped around */
7128 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
7130 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
7131 arc_loaned_bytes), 0);
7134 * Writes will, almost always, require additional memory allocations
7135 * in order to compress/encrypt/etc the data. We therefore need to
7136 * make sure that there is sufficient available memory for this.
7138 error = arc_memory_throttle(reserve, txg);
7143 * Throttle writes when the amount of dirty data in the cache
7144 * gets too large. We try to keep the cache less than half full
7145 * of dirty blocks so that our sync times don't grow too large.
7146 * Note: if two requests come in concurrently, we might let them
7147 * both succeed, when one of them should fail. Not a huge deal.
7150 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
7151 anon_size > arc_c / 4) {
7152 uint64_t meta_esize =
7153 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7154 uint64_t data_esize =
7155 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7156 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
7157 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
7158 arc_tempreserve >> 10, meta_esize >> 10,
7159 data_esize >> 10, reserve >> 10, arc_c >> 10);
7160 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
7161 return (SET_ERROR(ERESTART));
7163 atomic_add_64(&arc_tempreserve, reserve);
7168 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
7169 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
7171 size->value.ui64 = refcount_count(&state->arcs_size);
7172 evict_data->value.ui64 =
7173 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
7174 evict_metadata->value.ui64 =
7175 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
7179 arc_kstat_update(kstat_t *ksp, int rw)
7181 arc_stats_t *as = ksp->ks_data;
7183 if (rw == KSTAT_WRITE) {
7184 return (SET_ERROR(EACCES));
7186 arc_kstat_update_state(arc_anon,
7187 &as->arcstat_anon_size,
7188 &as->arcstat_anon_evictable_data,
7189 &as->arcstat_anon_evictable_metadata);
7190 arc_kstat_update_state(arc_mru,
7191 &as->arcstat_mru_size,
7192 &as->arcstat_mru_evictable_data,
7193 &as->arcstat_mru_evictable_metadata);
7194 arc_kstat_update_state(arc_mru_ghost,
7195 &as->arcstat_mru_ghost_size,
7196 &as->arcstat_mru_ghost_evictable_data,
7197 &as->arcstat_mru_ghost_evictable_metadata);
7198 arc_kstat_update_state(arc_mfu,
7199 &as->arcstat_mfu_size,
7200 &as->arcstat_mfu_evictable_data,
7201 &as->arcstat_mfu_evictable_metadata);
7202 arc_kstat_update_state(arc_mfu_ghost,
7203 &as->arcstat_mfu_ghost_size,
7204 &as->arcstat_mfu_ghost_evictable_data,
7205 &as->arcstat_mfu_ghost_evictable_metadata);
7207 as->arcstat_memory_all_bytes.value.ui64 =
7209 as->arcstat_memory_free_bytes.value.ui64 =
7211 as->arcstat_memory_available_bytes.value.i64 =
7212 arc_available_memory();
7219 * This function *must* return indices evenly distributed between all
7220 * sublists of the multilist. This is needed due to how the ARC eviction
7221 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7222 * distributed between all sublists and uses this assumption when
7223 * deciding which sublist to evict from and how much to evict from it.
7226 arc_state_multilist_index_func(multilist_t *ml, void *obj)
7228 arc_buf_hdr_t *hdr = obj;
7231 * We rely on b_dva to generate evenly distributed index
7232 * numbers using buf_hash below. So, as an added precaution,
7233 * let's make sure we never add empty buffers to the arc lists.
7235 ASSERT(!HDR_EMPTY(hdr));
7238 * The assumption here, is the hash value for a given
7239 * arc_buf_hdr_t will remain constant throughout its lifetime
7240 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7241 * Thus, we don't need to store the header's sublist index
7242 * on insertion, as this index can be recalculated on removal.
7244 * Also, the low order bits of the hash value are thought to be
7245 * distributed evenly. Otherwise, in the case that the multilist
7246 * has a power of two number of sublists, each sublists' usage
7247 * would not be evenly distributed.
7249 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
7250 multilist_get_num_sublists(ml));
7254 * Called during module initialization and periodically thereafter to
7255 * apply reasonable changes to the exposed performance tunings. Non-zero
7256 * zfs_* values which differ from the currently set values will be applied.
7259 arc_tuning_update(void)
7261 uint64_t allmem = arc_all_memory();
7262 unsigned long limit;
7264 /* Valid range: 64M - <all physical memory> */
7265 if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) &&
7266 (zfs_arc_max > 64 << 20) && (zfs_arc_max < allmem) &&
7267 (zfs_arc_max > arc_c_min)) {
7268 arc_c_max = zfs_arc_max;
7270 arc_p = (arc_c >> 1);
7271 if (arc_meta_limit > arc_c_max)
7272 arc_meta_limit = arc_c_max;
7273 if (arc_dnode_limit > arc_meta_limit)
7274 arc_dnode_limit = arc_meta_limit;
7277 /* Valid range: 32M - <arc_c_max> */
7278 if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) &&
7279 (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) &&
7280 (zfs_arc_min <= arc_c_max)) {
7281 arc_c_min = zfs_arc_min;
7282 arc_c = MAX(arc_c, arc_c_min);
7285 /* Valid range: 16M - <arc_c_max> */
7286 if ((zfs_arc_meta_min) && (zfs_arc_meta_min != arc_meta_min) &&
7287 (zfs_arc_meta_min >= 1ULL << SPA_MAXBLOCKSHIFT) &&
7288 (zfs_arc_meta_min <= arc_c_max)) {
7289 arc_meta_min = zfs_arc_meta_min;
7290 if (arc_meta_limit < arc_meta_min)
7291 arc_meta_limit = arc_meta_min;
7292 if (arc_dnode_limit < arc_meta_min)
7293 arc_dnode_limit = arc_meta_min;
7296 /* Valid range: <arc_meta_min> - <arc_c_max> */
7297 limit = zfs_arc_meta_limit ? zfs_arc_meta_limit :
7298 MIN(zfs_arc_meta_limit_percent, 100) * arc_c_max / 100;
7299 if ((limit != arc_meta_limit) &&
7300 (limit >= arc_meta_min) &&
7301 (limit <= arc_c_max))
7302 arc_meta_limit = limit;
7304 /* Valid range: <arc_meta_min> - <arc_meta_limit> */
7305 limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit :
7306 MIN(zfs_arc_dnode_limit_percent, 100) * arc_meta_limit / 100;
7307 if ((limit != arc_dnode_limit) &&
7308 (limit >= arc_meta_min) &&
7309 (limit <= arc_meta_limit))
7310 arc_dnode_limit = limit;
7312 /* Valid range: 1 - N */
7313 if (zfs_arc_grow_retry)
7314 arc_grow_retry = zfs_arc_grow_retry;
7316 /* Valid range: 1 - N */
7317 if (zfs_arc_shrink_shift) {
7318 arc_shrink_shift = zfs_arc_shrink_shift;
7319 arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1);
7322 /* Valid range: 1 - N */
7323 if (zfs_arc_p_min_shift)
7324 arc_p_min_shift = zfs_arc_p_min_shift;
7326 /* Valid range: 1 - N ms */
7327 if (zfs_arc_min_prefetch_ms)
7328 arc_min_prefetch_ms = zfs_arc_min_prefetch_ms;
7330 /* Valid range: 1 - N ms */
7331 if (zfs_arc_min_prescient_prefetch_ms) {
7332 arc_min_prescient_prefetch_ms =
7333 zfs_arc_min_prescient_prefetch_ms;
7336 /* Valid range: 0 - 100 */
7337 if ((zfs_arc_lotsfree_percent >= 0) &&
7338 (zfs_arc_lotsfree_percent <= 100))
7339 arc_lotsfree_percent = zfs_arc_lotsfree_percent;
7341 /* Valid range: 0 - <all physical memory> */
7342 if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free))
7343 arc_sys_free = MIN(MAX(zfs_arc_sys_free, 0), allmem);
7348 arc_state_init(void)
7350 arc_anon = &ARC_anon;
7352 arc_mru_ghost = &ARC_mru_ghost;
7354 arc_mfu_ghost = &ARC_mfu_ghost;
7355 arc_l2c_only = &ARC_l2c_only;
7357 arc_mru->arcs_list[ARC_BUFC_METADATA] =
7358 multilist_create(sizeof (arc_buf_hdr_t),
7359 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7360 arc_state_multilist_index_func);
7361 arc_mru->arcs_list[ARC_BUFC_DATA] =
7362 multilist_create(sizeof (arc_buf_hdr_t),
7363 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7364 arc_state_multilist_index_func);
7365 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
7366 multilist_create(sizeof (arc_buf_hdr_t),
7367 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7368 arc_state_multilist_index_func);
7369 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
7370 multilist_create(sizeof (arc_buf_hdr_t),
7371 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7372 arc_state_multilist_index_func);
7373 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
7374 multilist_create(sizeof (arc_buf_hdr_t),
7375 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7376 arc_state_multilist_index_func);
7377 arc_mfu->arcs_list[ARC_BUFC_DATA] =
7378 multilist_create(sizeof (arc_buf_hdr_t),
7379 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7380 arc_state_multilist_index_func);
7381 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
7382 multilist_create(sizeof (arc_buf_hdr_t),
7383 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7384 arc_state_multilist_index_func);
7385 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
7386 multilist_create(sizeof (arc_buf_hdr_t),
7387 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7388 arc_state_multilist_index_func);
7389 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
7390 multilist_create(sizeof (arc_buf_hdr_t),
7391 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7392 arc_state_multilist_index_func);
7393 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
7394 multilist_create(sizeof (arc_buf_hdr_t),
7395 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7396 arc_state_multilist_index_func);
7398 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7399 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7400 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7401 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7402 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7403 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7404 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7405 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7406 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7407 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7408 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7409 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7411 refcount_create(&arc_anon->arcs_size);
7412 refcount_create(&arc_mru->arcs_size);
7413 refcount_create(&arc_mru_ghost->arcs_size);
7414 refcount_create(&arc_mfu->arcs_size);
7415 refcount_create(&arc_mfu_ghost->arcs_size);
7416 refcount_create(&arc_l2c_only->arcs_size);
7418 arc_anon->arcs_state = ARC_STATE_ANON;
7419 arc_mru->arcs_state = ARC_STATE_MRU;
7420 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7421 arc_mfu->arcs_state = ARC_STATE_MFU;
7422 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7423 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7427 arc_state_fini(void)
7429 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7430 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7431 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7432 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7433 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7434 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7435 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7436 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7437 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7438 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7439 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7440 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7442 refcount_destroy(&arc_anon->arcs_size);
7443 refcount_destroy(&arc_mru->arcs_size);
7444 refcount_destroy(&arc_mru_ghost->arcs_size);
7445 refcount_destroy(&arc_mfu->arcs_size);
7446 refcount_destroy(&arc_mfu_ghost->arcs_size);
7447 refcount_destroy(&arc_l2c_only->arcs_size);
7449 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
7450 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7451 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7452 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7453 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
7454 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7455 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
7456 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7457 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7458 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7462 arc_target_bytes(void)
7470 uint64_t percent, allmem = arc_all_memory();
7472 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
7473 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
7474 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
7476 arc_min_prefetch_ms = 1000;
7477 arc_min_prescient_prefetch_ms = 6000;
7481 * Register a shrinker to support synchronous (direct) memory
7482 * reclaim from the arc. This is done to prevent kswapd from
7483 * swapping out pages when it is preferable to shrink the arc.
7485 spl_register_shrinker(&arc_shrinker);
7487 /* Set to 1/64 of all memory or a minimum of 512K */
7488 arc_sys_free = MAX(allmem / 64, (512 * 1024));
7492 /* Set max to 1/2 of all memory */
7493 arc_c_max = allmem / 2;
7496 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more */
7497 arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT);
7500 * In userland, there's only the memory pressure that we artificially
7501 * create (see arc_available_memory()). Don't let arc_c get too
7502 * small, because it can cause transactions to be larger than
7503 * arc_c, causing arc_tempreserve_space() to fail.
7505 arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT);
7509 arc_p = (arc_c >> 1);
7512 /* Set min to 1/2 of arc_c_min */
7513 arc_meta_min = 1ULL << SPA_MAXBLOCKSHIFT;
7514 /* Initialize maximum observed usage to zero */
7517 * Set arc_meta_limit to a percent of arc_c_max with a floor of
7518 * arc_meta_min, and a ceiling of arc_c_max.
7520 percent = MIN(zfs_arc_meta_limit_percent, 100);
7521 arc_meta_limit = MAX(arc_meta_min, (percent * arc_c_max) / 100);
7522 percent = MIN(zfs_arc_dnode_limit_percent, 100);
7523 arc_dnode_limit = (percent * arc_meta_limit) / 100;
7525 /* Apply user specified tunings */
7526 arc_tuning_update();
7528 /* if kmem_flags are set, lets try to use less memory */
7529 if (kmem_debugging())
7531 if (arc_c < arc_c_min)
7537 list_create(&arc_prune_list, sizeof (arc_prune_t),
7538 offsetof(arc_prune_t, p_node));
7539 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7541 arc_prune_taskq = taskq_create("arc_prune", max_ncpus, defclsyspri,
7542 max_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
7544 arc_reclaim_thread_exit = B_FALSE;
7546 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7547 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7549 if (arc_ksp != NULL) {
7550 arc_ksp->ks_data = &arc_stats;
7551 arc_ksp->ks_update = arc_kstat_update;
7552 kstat_install(arc_ksp);
7555 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
7556 TS_RUN, defclsyspri);
7562 * Calculate maximum amount of dirty data per pool.
7564 * If it has been set by a module parameter, take that.
7565 * Otherwise, use a percentage of physical memory defined by
7566 * zfs_dirty_data_max_percent (default 10%) with a cap at
7567 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
7569 if (zfs_dirty_data_max_max == 0)
7570 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
7571 allmem * zfs_dirty_data_max_max_percent / 100);
7573 if (zfs_dirty_data_max == 0) {
7574 zfs_dirty_data_max = allmem *
7575 zfs_dirty_data_max_percent / 100;
7576 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7577 zfs_dirty_data_max_max);
7587 spl_unregister_shrinker(&arc_shrinker);
7588 #endif /* _KERNEL */
7590 mutex_enter(&arc_reclaim_lock);
7591 arc_reclaim_thread_exit = B_TRUE;
7593 * The reclaim thread will set arc_reclaim_thread_exit back to
7594 * B_FALSE when it is finished exiting; we're waiting for that.
7596 while (arc_reclaim_thread_exit) {
7597 cv_signal(&arc_reclaim_thread_cv);
7598 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
7600 mutex_exit(&arc_reclaim_lock);
7602 /* Use B_TRUE to ensure *all* buffers are evicted */
7603 arc_flush(NULL, B_TRUE);
7607 if (arc_ksp != NULL) {
7608 kstat_delete(arc_ksp);
7612 taskq_wait(arc_prune_taskq);
7613 taskq_destroy(arc_prune_taskq);
7615 mutex_enter(&arc_prune_mtx);
7616 while ((p = list_head(&arc_prune_list)) != NULL) {
7617 list_remove(&arc_prune_list, p);
7618 refcount_remove(&p->p_refcnt, &arc_prune_list);
7619 refcount_destroy(&p->p_refcnt);
7620 kmem_free(p, sizeof (*p));
7622 mutex_exit(&arc_prune_mtx);
7624 list_destroy(&arc_prune_list);
7625 mutex_destroy(&arc_prune_mtx);
7626 mutex_destroy(&arc_reclaim_lock);
7627 cv_destroy(&arc_reclaim_thread_cv);
7628 cv_destroy(&arc_reclaim_waiters_cv);
7633 ASSERT0(arc_loaned_bytes);
7639 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7640 * It uses dedicated storage devices to hold cached data, which are populated
7641 * using large infrequent writes. The main role of this cache is to boost
7642 * the performance of random read workloads. The intended L2ARC devices
7643 * include short-stroked disks, solid state disks, and other media with
7644 * substantially faster read latency than disk.
7646 * +-----------------------+
7648 * +-----------------------+
7651 * l2arc_feed_thread() arc_read()
7655 * +---------------+ |
7657 * +---------------+ |
7662 * +-------+ +-------+
7664 * | cache | | cache |
7665 * +-------+ +-------+
7666 * +=========+ .-----.
7667 * : L2ARC : |-_____-|
7668 * : devices : | Disks |
7669 * +=========+ `-_____-'
7671 * Read requests are satisfied from the following sources, in order:
7674 * 2) vdev cache of L2ARC devices
7676 * 4) vdev cache of disks
7679 * Some L2ARC device types exhibit extremely slow write performance.
7680 * To accommodate for this there are some significant differences between
7681 * the L2ARC and traditional cache design:
7683 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7684 * the ARC behave as usual, freeing buffers and placing headers on ghost
7685 * lists. The ARC does not send buffers to the L2ARC during eviction as
7686 * this would add inflated write latencies for all ARC memory pressure.
7688 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7689 * It does this by periodically scanning buffers from the eviction-end of
7690 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7691 * not already there. It scans until a headroom of buffers is satisfied,
7692 * which itself is a buffer for ARC eviction. If a compressible buffer is
7693 * found during scanning and selected for writing to an L2ARC device, we
7694 * temporarily boost scanning headroom during the next scan cycle to make
7695 * sure we adapt to compression effects (which might significantly reduce
7696 * the data volume we write to L2ARC). The thread that does this is
7697 * l2arc_feed_thread(), illustrated below; example sizes are included to
7698 * provide a better sense of ratio than this diagram:
7701 * +---------------------+----------+
7702 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7703 * +---------------------+----------+ | o L2ARC eligible
7704 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7705 * +---------------------+----------+ |
7706 * 15.9 Gbytes ^ 32 Mbytes |
7708 * l2arc_feed_thread()
7710 * l2arc write hand <--[oooo]--'
7714 * +==============================+
7715 * L2ARC dev |####|#|###|###| |####| ... |
7716 * +==============================+
7719 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7720 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7721 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7722 * safe to say that this is an uncommon case, since buffers at the end of
7723 * the ARC lists have moved there due to inactivity.
7725 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7726 * then the L2ARC simply misses copying some buffers. This serves as a
7727 * pressure valve to prevent heavy read workloads from both stalling the ARC
7728 * with waits and clogging the L2ARC with writes. This also helps prevent
7729 * the potential for the L2ARC to churn if it attempts to cache content too
7730 * quickly, such as during backups of the entire pool.
7732 * 5. After system boot and before the ARC has filled main memory, there are
7733 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7734 * lists can remain mostly static. Instead of searching from tail of these
7735 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7736 * for eligible buffers, greatly increasing its chance of finding them.
7738 * The L2ARC device write speed is also boosted during this time so that
7739 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7740 * there are no L2ARC reads, and no fear of degrading read performance
7741 * through increased writes.
7743 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7744 * the vdev queue can aggregate them into larger and fewer writes. Each
7745 * device is written to in a rotor fashion, sweeping writes through
7746 * available space then repeating.
7748 * 7. The L2ARC does not store dirty content. It never needs to flush
7749 * write buffers back to disk based storage.
7751 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7752 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7754 * The performance of the L2ARC can be tweaked by a number of tunables, which
7755 * may be necessary for different workloads:
7757 * l2arc_write_max max write bytes per interval
7758 * l2arc_write_boost extra write bytes during device warmup
7759 * l2arc_noprefetch skip caching prefetched buffers
7760 * l2arc_headroom number of max device writes to precache
7761 * l2arc_headroom_boost when we find compressed buffers during ARC
7762 * scanning, we multiply headroom by this
7763 * percentage factor for the next scan cycle,
7764 * since more compressed buffers are likely to
7766 * l2arc_feed_secs seconds between L2ARC writing
7768 * Tunables may be removed or added as future performance improvements are
7769 * integrated, and also may become zpool properties.
7771 * There are three key functions that control how the L2ARC warms up:
7773 * l2arc_write_eligible() check if a buffer is eligible to cache
7774 * l2arc_write_size() calculate how much to write
7775 * l2arc_write_interval() calculate sleep delay between writes
7777 * These three functions determine what to write, how much, and how quickly
7782 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
7785 * A buffer is *not* eligible for the L2ARC if it:
7786 * 1. belongs to a different spa.
7787 * 2. is already cached on the L2ARC.
7788 * 3. has an I/O in progress (it may be an incomplete read).
7789 * 4. is flagged not eligible (zfs property).
7791 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
7792 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
7799 l2arc_write_size(void)
7804 * Make sure our globals have meaningful values in case the user
7807 size = l2arc_write_max;
7809 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
7810 "be greater than zero, resetting it to the default (%d)",
7812 size = l2arc_write_max = L2ARC_WRITE_SIZE;
7815 if (arc_warm == B_FALSE)
7816 size += l2arc_write_boost;
7823 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
7825 clock_t interval, next, now;
7828 * If the ARC lists are busy, increase our write rate; if the
7829 * lists are stale, idle back. This is achieved by checking
7830 * how much we previously wrote - if it was more than half of
7831 * what we wanted, schedule the next write much sooner.
7833 if (l2arc_feed_again && wrote > (wanted / 2))
7834 interval = (hz * l2arc_feed_min_ms) / 1000;
7836 interval = hz * l2arc_feed_secs;
7838 now = ddi_get_lbolt();
7839 next = MAX(now, MIN(now + interval, began + interval));
7845 * Cycle through L2ARC devices. This is how L2ARC load balances.
7846 * If a device is returned, this also returns holding the spa config lock.
7848 static l2arc_dev_t *
7849 l2arc_dev_get_next(void)
7851 l2arc_dev_t *first, *next = NULL;
7854 * Lock out the removal of spas (spa_namespace_lock), then removal
7855 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7856 * both locks will be dropped and a spa config lock held instead.
7858 mutex_enter(&spa_namespace_lock);
7859 mutex_enter(&l2arc_dev_mtx);
7861 /* if there are no vdevs, there is nothing to do */
7862 if (l2arc_ndev == 0)
7866 next = l2arc_dev_last;
7868 /* loop around the list looking for a non-faulted vdev */
7870 next = list_head(l2arc_dev_list);
7872 next = list_next(l2arc_dev_list, next);
7874 next = list_head(l2arc_dev_list);
7877 /* if we have come back to the start, bail out */
7880 else if (next == first)
7883 } while (vdev_is_dead(next->l2ad_vdev));
7885 /* if we were unable to find any usable vdevs, return NULL */
7886 if (vdev_is_dead(next->l2ad_vdev))
7889 l2arc_dev_last = next;
7892 mutex_exit(&l2arc_dev_mtx);
7895 * Grab the config lock to prevent the 'next' device from being
7896 * removed while we are writing to it.
7899 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7900 mutex_exit(&spa_namespace_lock);
7906 * Free buffers that were tagged for destruction.
7909 l2arc_do_free_on_write(void)
7912 l2arc_data_free_t *df, *df_prev;
7914 mutex_enter(&l2arc_free_on_write_mtx);
7915 buflist = l2arc_free_on_write;
7917 for (df = list_tail(buflist); df; df = df_prev) {
7918 df_prev = list_prev(buflist, df);
7919 ASSERT3P(df->l2df_abd, !=, NULL);
7920 abd_free(df->l2df_abd);
7921 list_remove(buflist, df);
7922 kmem_free(df, sizeof (l2arc_data_free_t));
7925 mutex_exit(&l2arc_free_on_write_mtx);
7929 * A write to a cache device has completed. Update all headers to allow
7930 * reads from these buffers to begin.
7933 l2arc_write_done(zio_t *zio)
7935 l2arc_write_callback_t *cb;
7938 arc_buf_hdr_t *head, *hdr, *hdr_prev;
7939 kmutex_t *hash_lock;
7940 int64_t bytes_dropped = 0;
7942 cb = zio->io_private;
7943 ASSERT3P(cb, !=, NULL);
7944 dev = cb->l2wcb_dev;
7945 ASSERT3P(dev, !=, NULL);
7946 head = cb->l2wcb_head;
7947 ASSERT3P(head, !=, NULL);
7948 buflist = &dev->l2ad_buflist;
7949 ASSERT3P(buflist, !=, NULL);
7950 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7951 l2arc_write_callback_t *, cb);
7953 if (zio->io_error != 0)
7954 ARCSTAT_BUMP(arcstat_l2_writes_error);
7957 * All writes completed, or an error was hit.
7960 mutex_enter(&dev->l2ad_mtx);
7961 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7962 hdr_prev = list_prev(buflist, hdr);
7964 hash_lock = HDR_LOCK(hdr);
7967 * We cannot use mutex_enter or else we can deadlock
7968 * with l2arc_write_buffers (due to swapping the order
7969 * the hash lock and l2ad_mtx are taken).
7971 if (!mutex_tryenter(hash_lock)) {
7973 * Missed the hash lock. We must retry so we
7974 * don't leave the ARC_FLAG_L2_WRITING bit set.
7976 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7979 * We don't want to rescan the headers we've
7980 * already marked as having been written out, so
7981 * we reinsert the head node so we can pick up
7982 * where we left off.
7984 list_remove(buflist, head);
7985 list_insert_after(buflist, hdr, head);
7987 mutex_exit(&dev->l2ad_mtx);
7990 * We wait for the hash lock to become available
7991 * to try and prevent busy waiting, and increase
7992 * the chance we'll be able to acquire the lock
7993 * the next time around.
7995 mutex_enter(hash_lock);
7996 mutex_exit(hash_lock);
8001 * We could not have been moved into the arc_l2c_only
8002 * state while in-flight due to our ARC_FLAG_L2_WRITING
8003 * bit being set. Let's just ensure that's being enforced.
8005 ASSERT(HDR_HAS_L1HDR(hdr));
8008 * Skipped - drop L2ARC entry and mark the header as no
8009 * longer L2 eligibile.
8011 if (zio->io_error != 0) {
8013 * Error - drop L2ARC entry.
8015 list_remove(buflist, hdr);
8016 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
8018 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
8019 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
8021 bytes_dropped += arc_hdr_size(hdr);
8022 (void) refcount_remove_many(&dev->l2ad_alloc,
8023 arc_hdr_size(hdr), hdr);
8027 * Allow ARC to begin reads and ghost list evictions to
8030 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
8032 mutex_exit(hash_lock);
8035 atomic_inc_64(&l2arc_writes_done);
8036 list_remove(buflist, head);
8037 ASSERT(!HDR_HAS_L1HDR(head));
8038 kmem_cache_free(hdr_l2only_cache, head);
8039 mutex_exit(&dev->l2ad_mtx);
8041 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
8043 l2arc_do_free_on_write();
8045 kmem_free(cb, sizeof (l2arc_write_callback_t));
8049 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
8052 spa_t *spa = zio->io_spa;
8053 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
8054 blkptr_t *bp = zio->io_bp;
8055 dsl_crypto_key_t *dck = NULL;
8056 uint8_t salt[ZIO_DATA_SALT_LEN];
8057 uint8_t iv[ZIO_DATA_IV_LEN];
8058 uint8_t mac[ZIO_DATA_MAC_LEN];
8059 boolean_t no_crypt = B_FALSE;
8062 * ZIL data is never be written to the L2ARC, so we don't need
8063 * special handling for its unique MAC storage.
8065 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
8066 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
8067 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8070 * If the data was encrypted, decrypt it now. Note that
8071 * we must check the bp here and not the hdr, since the
8072 * hdr does not have its encryption parameters updated
8073 * until arc_read_done().
8075 if (BP_IS_ENCRYPTED(bp)) {
8076 abd_t *eabd = arc_get_data_abd(hdr,
8077 arc_hdr_size(hdr), hdr);
8079 zio_crypt_decode_params_bp(bp, salt, iv);
8080 zio_crypt_decode_mac_bp(bp, mac);
8082 ret = spa_keystore_lookup_key(spa,
8083 cb->l2rcb_zb.zb_objset, FTAG, &dck);
8085 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8089 ret = zio_do_crypt_abd(B_FALSE, &dck->dck_key,
8090 salt, BP_GET_TYPE(bp), iv, mac, HDR_GET_PSIZE(hdr),
8091 BP_SHOULD_BYTESWAP(bp), eabd, hdr->b_l1hdr.b_pabd,
8094 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8095 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8099 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8102 * If we actually performed decryption, replace b_pabd
8103 * with the decrypted data. Otherwise we can just throw
8104 * our decryption buffer away.
8107 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8108 arc_hdr_size(hdr), hdr);
8109 hdr->b_l1hdr.b_pabd = eabd;
8112 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8117 * If the L2ARC block was compressed, but ARC compression
8118 * is disabled we decompress the data into a new buffer and
8119 * replace the existing data.
8121 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8122 !HDR_COMPRESSION_ENABLED(hdr)) {
8123 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
8124 void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
8126 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
8127 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
8128 HDR_GET_LSIZE(hdr));
8130 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8131 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
8135 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8136 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8137 arc_hdr_size(hdr), hdr);
8138 hdr->b_l1hdr.b_pabd = cabd;
8140 zio->io_size = HDR_GET_LSIZE(hdr);
8151 * A read to a cache device completed. Validate buffer contents before
8152 * handing over to the regular ARC routines.
8155 l2arc_read_done(zio_t *zio)
8158 l2arc_read_callback_t *cb;
8160 kmutex_t *hash_lock;
8161 boolean_t valid_cksum, using_rdata;
8163 ASSERT3P(zio->io_vd, !=, NULL);
8164 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
8166 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
8168 cb = zio->io_private;
8169 ASSERT3P(cb, !=, NULL);
8170 hdr = cb->l2rcb_hdr;
8171 ASSERT3P(hdr, !=, NULL);
8173 hash_lock = HDR_LOCK(hdr);
8174 mutex_enter(hash_lock);
8175 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
8178 * If the data was read into a temporary buffer,
8179 * move it and free the buffer.
8181 if (cb->l2rcb_abd != NULL) {
8182 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
8183 if (zio->io_error == 0) {
8184 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
8189 * The following must be done regardless of whether
8190 * there was an error:
8191 * - free the temporary buffer
8192 * - point zio to the real ARC buffer
8193 * - set zio size accordingly
8194 * These are required because zio is either re-used for
8195 * an I/O of the block in the case of the error
8196 * or the zio is passed to arc_read_done() and it
8199 abd_free(cb->l2rcb_abd);
8200 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
8202 if (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
8203 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT)) {
8204 ASSERT(HDR_HAS_RABD(hdr));
8205 zio->io_abd = zio->io_orig_abd =
8206 hdr->b_crypt_hdr.b_rabd;
8208 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8209 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8213 ASSERT3P(zio->io_abd, !=, NULL);
8216 * Check this survived the L2ARC journey.
8218 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8219 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8220 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
8221 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
8223 valid_cksum = arc_cksum_is_equal(hdr, zio);
8224 using_rdata = (HDR_HAS_RABD(hdr) &&
8225 zio->io_abd == hdr->b_crypt_hdr.b_rabd);
8228 * b_rabd will always match the data as it exists on disk if it is
8229 * being used. Therefore if we are reading into b_rabd we do not
8230 * attempt to untransform the data.
8232 if (valid_cksum && !using_rdata)
8233 tfm_error = l2arc_untransform(zio, cb);
8235 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8236 !HDR_L2_EVICTED(hdr)) {
8237 mutex_exit(hash_lock);
8238 zio->io_private = hdr;
8241 mutex_exit(hash_lock);
8243 * Buffer didn't survive caching. Increment stats and
8244 * reissue to the original storage device.
8246 if (zio->io_error != 0) {
8247 ARCSTAT_BUMP(arcstat_l2_io_error);
8249 zio->io_error = SET_ERROR(EIO);
8251 if (!valid_cksum || tfm_error != 0)
8252 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8255 * If there's no waiter, issue an async i/o to the primary
8256 * storage now. If there *is* a waiter, the caller must
8257 * issue the i/o in a context where it's OK to block.
8259 if (zio->io_waiter == NULL) {
8260 zio_t *pio = zio_unique_parent(zio);
8261 void *abd = (using_rdata) ?
8262 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8264 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8266 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
8267 abd, zio->io_size, arc_read_done,
8268 hdr, zio->io_priority, cb->l2rcb_flags,
8273 kmem_free(cb, sizeof (l2arc_read_callback_t));
8277 * This is the list priority from which the L2ARC will search for pages to
8278 * cache. This is used within loops (0..3) to cycle through lists in the
8279 * desired order. This order can have a significant effect on cache
8282 * Currently the metadata lists are hit first, MFU then MRU, followed by
8283 * the data lists. This function returns a locked list, and also returns
8286 static multilist_sublist_t *
8287 l2arc_sublist_lock(int list_num)
8289 multilist_t *ml = NULL;
8292 ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
8296 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
8299 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
8302 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
8305 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
8312 * Return a randomly-selected sublist. This is acceptable
8313 * because the caller feeds only a little bit of data for each
8314 * call (8MB). Subsequent calls will result in different
8315 * sublists being selected.
8317 idx = multilist_get_random_index(ml);
8318 return (multilist_sublist_lock(ml, idx));
8322 * Evict buffers from the device write hand to the distance specified in
8323 * bytes. This distance may span populated buffers, it may span nothing.
8324 * This is clearing a region on the L2ARC device ready for writing.
8325 * If the 'all' boolean is set, every buffer is evicted.
8328 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
8331 arc_buf_hdr_t *hdr, *hdr_prev;
8332 kmutex_t *hash_lock;
8335 buflist = &dev->l2ad_buflist;
8337 if (!all && dev->l2ad_first) {
8339 * This is the first sweep through the device. There is
8345 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
8347 * When nearing the end of the device, evict to the end
8348 * before the device write hand jumps to the start.
8350 taddr = dev->l2ad_end;
8352 taddr = dev->l2ad_hand + distance;
8354 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
8355 uint64_t, taddr, boolean_t, all);
8358 mutex_enter(&dev->l2ad_mtx);
8359 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
8360 hdr_prev = list_prev(buflist, hdr);
8362 hash_lock = HDR_LOCK(hdr);
8365 * We cannot use mutex_enter or else we can deadlock
8366 * with l2arc_write_buffers (due to swapping the order
8367 * the hash lock and l2ad_mtx are taken).
8369 if (!mutex_tryenter(hash_lock)) {
8371 * Missed the hash lock. Retry.
8373 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
8374 mutex_exit(&dev->l2ad_mtx);
8375 mutex_enter(hash_lock);
8376 mutex_exit(hash_lock);
8381 * A header can't be on this list if it doesn't have L2 header.
8383 ASSERT(HDR_HAS_L2HDR(hdr));
8385 /* Ensure this header has finished being written. */
8386 ASSERT(!HDR_L2_WRITING(hdr));
8387 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
8389 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
8390 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
8392 * We've evicted to the target address,
8393 * or the end of the device.
8395 mutex_exit(hash_lock);
8399 if (!HDR_HAS_L1HDR(hdr)) {
8400 ASSERT(!HDR_L2_READING(hdr));
8402 * This doesn't exist in the ARC. Destroy.
8403 * arc_hdr_destroy() will call list_remove()
8404 * and decrement arcstat_l2_lsize.
8406 arc_change_state(arc_anon, hdr, hash_lock);
8407 arc_hdr_destroy(hdr);
8409 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
8410 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
8412 * Invalidate issued or about to be issued
8413 * reads, since we may be about to write
8414 * over this location.
8416 if (HDR_L2_READING(hdr)) {
8417 ARCSTAT_BUMP(arcstat_l2_evict_reading);
8418 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
8421 arc_hdr_l2hdr_destroy(hdr);
8423 mutex_exit(hash_lock);
8425 mutex_exit(&dev->l2ad_mtx);
8429 * Handle any abd transforms that might be required for writing to the L2ARC.
8430 * If successful, this function will always return an abd with the data
8431 * transformed as it is on disk in a new abd of asize bytes.
8434 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
8439 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
8440 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
8441 uint64_t psize = HDR_GET_PSIZE(hdr);
8442 uint64_t size = arc_hdr_size(hdr);
8443 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
8444 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
8445 dsl_crypto_key_t *dck = NULL;
8446 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
8447 boolean_t no_crypt = B_FALSE;
8449 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8450 !HDR_COMPRESSION_ENABLED(hdr)) ||
8451 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
8452 ASSERT3U(psize, <=, asize);
8455 * If this data simply needs its own buffer, we simply allocate it
8456 * and copy the data. This may be done to elimiate a depedency on a
8457 * shared buffer or to reallocate the buffer to match asize.
8459 if (HDR_HAS_RABD(hdr) && asize != psize) {
8460 ASSERT3U(size, ==, psize);
8461 to_write = abd_alloc_for_io(asize, ismd);
8462 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, size);
8464 abd_zero_off(to_write, size, asize - size);
8468 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
8469 !HDR_ENCRYPTED(hdr)) {
8470 ASSERT3U(size, ==, psize);
8471 to_write = abd_alloc_for_io(asize, ismd);
8472 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
8474 abd_zero_off(to_write, size, asize - size);
8478 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
8479 cabd = abd_alloc_for_io(asize, ismd);
8480 tmp = abd_borrow_buf(cabd, asize);
8482 psize = zio_compress_data(compress, to_write, tmp, size);
8483 ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
8485 bzero((char *)tmp + psize, asize - psize);
8486 psize = HDR_GET_PSIZE(hdr);
8487 abd_return_buf_copy(cabd, tmp, asize);
8491 if (HDR_ENCRYPTED(hdr)) {
8492 eabd = abd_alloc_for_io(asize, ismd);
8495 * If the dataset was disowned before the buffer
8496 * made it to this point, the key to re-encrypt
8497 * it won't be available. In this case we simply
8498 * won't write the buffer to the L2ARC.
8500 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
8505 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
8506 hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_ot,
8507 hdr->b_crypt_hdr.b_iv, mac, psize, bswap, to_write,
8513 abd_copy(eabd, to_write, psize);
8516 abd_zero_off(eabd, psize, asize - psize);
8518 /* assert that the MAC we got here matches the one we saved */
8519 ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
8520 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8522 if (to_write == cabd)
8529 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
8530 *abd_out = to_write;
8535 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8546 * Find and write ARC buffers to the L2ARC device.
8548 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
8549 * for reading until they have completed writing.
8550 * The headroom_boost is an in-out parameter used to maintain headroom boost
8551 * state between calls to this function.
8553 * Returns the number of bytes actually written (which may be smaller than
8554 * the delta by which the device hand has changed due to alignment).
8557 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
8559 arc_buf_hdr_t *hdr, *hdr_prev, *head;
8560 uint64_t write_asize, write_psize, write_lsize, headroom;
8562 l2arc_write_callback_t *cb;
8564 uint64_t guid = spa_load_guid(spa);
8566 ASSERT3P(dev->l2ad_vdev, !=, NULL);
8569 write_lsize = write_asize = write_psize = 0;
8571 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
8572 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
8575 * Copy buffers for L2ARC writing.
8577 for (int try = 0; try < L2ARC_FEED_TYPES; try++) {
8578 multilist_sublist_t *mls = l2arc_sublist_lock(try);
8579 uint64_t passed_sz = 0;
8581 VERIFY3P(mls, !=, NULL);
8584 * L2ARC fast warmup.
8586 * Until the ARC is warm and starts to evict, read from the
8587 * head of the ARC lists rather than the tail.
8589 if (arc_warm == B_FALSE)
8590 hdr = multilist_sublist_head(mls);
8592 hdr = multilist_sublist_tail(mls);
8594 headroom = target_sz * l2arc_headroom;
8595 if (zfs_compressed_arc_enabled)
8596 headroom = (headroom * l2arc_headroom_boost) / 100;
8598 for (; hdr; hdr = hdr_prev) {
8599 kmutex_t *hash_lock;
8600 abd_t *to_write = NULL;
8602 if (arc_warm == B_FALSE)
8603 hdr_prev = multilist_sublist_next(mls, hdr);
8605 hdr_prev = multilist_sublist_prev(mls, hdr);
8607 hash_lock = HDR_LOCK(hdr);
8608 if (!mutex_tryenter(hash_lock)) {
8610 * Skip this buffer rather than waiting.
8615 passed_sz += HDR_GET_LSIZE(hdr);
8616 if (passed_sz > headroom) {
8620 mutex_exit(hash_lock);
8624 if (!l2arc_write_eligible(guid, hdr)) {
8625 mutex_exit(hash_lock);
8630 * We rely on the L1 portion of the header below, so
8631 * it's invalid for this header to have been evicted out
8632 * of the ghost cache, prior to being written out. The
8633 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8635 ASSERT(HDR_HAS_L1HDR(hdr));
8637 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8638 ASSERT3U(arc_hdr_size(hdr), >, 0);
8639 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
8641 uint64_t psize = HDR_GET_PSIZE(hdr);
8642 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
8645 if ((write_asize + asize) > target_sz) {
8647 mutex_exit(hash_lock);
8652 * We rely on the L1 portion of the header below, so
8653 * it's invalid for this header to have been evicted out
8654 * of the ghost cache, prior to being written out. The
8655 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8657 arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
8658 ASSERT(HDR_HAS_L1HDR(hdr));
8660 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8661 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
8663 ASSERT3U(arc_hdr_size(hdr), >, 0);
8666 * If this header has b_rabd, we can use this since it
8667 * must always match the data exactly as it exists on
8668 * disk. Otherwise, the L2ARC can normally use the
8669 * hdr's data, but if we're sharing data between the
8670 * hdr and one of its bufs, L2ARC needs its own copy of
8671 * the data so that the ZIO below can't race with the
8672 * buf consumer. To ensure that this copy will be
8673 * available for the lifetime of the ZIO and be cleaned
8674 * up afterwards, we add it to the l2arc_free_on_write
8675 * queue. If we need to apply any transforms to the
8676 * data (compression, encryption) we will also need the
8679 if (HDR_HAS_RABD(hdr) && psize == asize) {
8680 to_write = hdr->b_crypt_hdr.b_rabd;
8681 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
8682 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
8683 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
8685 to_write = hdr->b_l1hdr.b_pabd;
8688 arc_buf_contents_t type = arc_buf_type(hdr);
8690 ret = l2arc_apply_transforms(spa, hdr, asize,
8693 arc_hdr_clear_flags(hdr,
8694 ARC_FLAG_L2_WRITING);
8695 mutex_exit(hash_lock);
8699 l2arc_free_abd_on_write(to_write, asize, type);
8704 * Insert a dummy header on the buflist so
8705 * l2arc_write_done() can find where the
8706 * write buffers begin without searching.
8708 mutex_enter(&dev->l2ad_mtx);
8709 list_insert_head(&dev->l2ad_buflist, head);
8710 mutex_exit(&dev->l2ad_mtx);
8713 sizeof (l2arc_write_callback_t), KM_SLEEP);
8714 cb->l2wcb_dev = dev;
8715 cb->l2wcb_head = head;
8716 pio = zio_root(spa, l2arc_write_done, cb,
8720 hdr->b_l2hdr.b_dev = dev;
8721 hdr->b_l2hdr.b_hits = 0;
8723 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
8724 arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR);
8726 mutex_enter(&dev->l2ad_mtx);
8727 list_insert_head(&dev->l2ad_buflist, hdr);
8728 mutex_exit(&dev->l2ad_mtx);
8730 (void) refcount_add_many(&dev->l2ad_alloc,
8731 arc_hdr_size(hdr), hdr);
8733 wzio = zio_write_phys(pio, dev->l2ad_vdev,
8734 hdr->b_l2hdr.b_daddr, asize, to_write,
8735 ZIO_CHECKSUM_OFF, NULL, hdr,
8736 ZIO_PRIORITY_ASYNC_WRITE,
8737 ZIO_FLAG_CANFAIL, B_FALSE);
8739 write_lsize += HDR_GET_LSIZE(hdr);
8740 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
8743 write_psize += psize;
8744 write_asize += asize;
8745 dev->l2ad_hand += asize;
8747 mutex_exit(hash_lock);
8749 (void) zio_nowait(wzio);
8752 multilist_sublist_unlock(mls);
8758 /* No buffers selected for writing? */
8760 ASSERT0(write_lsize);
8761 ASSERT(!HDR_HAS_L1HDR(head));
8762 kmem_cache_free(hdr_l2only_cache, head);
8766 ASSERT3U(write_asize, <=, target_sz);
8767 ARCSTAT_BUMP(arcstat_l2_writes_sent);
8768 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
8769 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
8770 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
8771 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
8774 * Bump device hand to the device start if it is approaching the end.
8775 * l2arc_evict() will already have evicted ahead for this case.
8777 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
8778 dev->l2ad_hand = dev->l2ad_start;
8779 dev->l2ad_first = B_FALSE;
8782 dev->l2ad_writing = B_TRUE;
8783 (void) zio_wait(pio);
8784 dev->l2ad_writing = B_FALSE;
8786 return (write_asize);
8790 * This thread feeds the L2ARC at regular intervals. This is the beating
8791 * heart of the L2ARC.
8795 l2arc_feed_thread(void *unused)
8800 uint64_t size, wrote;
8801 clock_t begin, next = ddi_get_lbolt();
8802 fstrans_cookie_t cookie;
8804 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
8806 mutex_enter(&l2arc_feed_thr_lock);
8808 cookie = spl_fstrans_mark();
8809 while (l2arc_thread_exit == 0) {
8810 CALLB_CPR_SAFE_BEGIN(&cpr);
8811 (void) cv_timedwait_sig(&l2arc_feed_thr_cv,
8812 &l2arc_feed_thr_lock, next);
8813 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
8814 next = ddi_get_lbolt() + hz;
8817 * Quick check for L2ARC devices.
8819 mutex_enter(&l2arc_dev_mtx);
8820 if (l2arc_ndev == 0) {
8821 mutex_exit(&l2arc_dev_mtx);
8824 mutex_exit(&l2arc_dev_mtx);
8825 begin = ddi_get_lbolt();
8828 * This selects the next l2arc device to write to, and in
8829 * doing so the next spa to feed from: dev->l2ad_spa. This
8830 * will return NULL if there are now no l2arc devices or if
8831 * they are all faulted.
8833 * If a device is returned, its spa's config lock is also
8834 * held to prevent device removal. l2arc_dev_get_next()
8835 * will grab and release l2arc_dev_mtx.
8837 if ((dev = l2arc_dev_get_next()) == NULL)
8840 spa = dev->l2ad_spa;
8841 ASSERT3P(spa, !=, NULL);
8844 * If the pool is read-only then force the feed thread to
8845 * sleep a little longer.
8847 if (!spa_writeable(spa)) {
8848 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
8849 spa_config_exit(spa, SCL_L2ARC, dev);
8854 * Avoid contributing to memory pressure.
8856 if (arc_reclaim_needed()) {
8857 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
8858 spa_config_exit(spa, SCL_L2ARC, dev);
8862 ARCSTAT_BUMP(arcstat_l2_feeds);
8864 size = l2arc_write_size();
8867 * Evict L2ARC buffers that will be overwritten.
8869 l2arc_evict(dev, size, B_FALSE);
8872 * Write ARC buffers.
8874 wrote = l2arc_write_buffers(spa, dev, size);
8877 * Calculate interval between writes.
8879 next = l2arc_write_interval(begin, size, wrote);
8880 spa_config_exit(spa, SCL_L2ARC, dev);
8882 spl_fstrans_unmark(cookie);
8884 l2arc_thread_exit = 0;
8885 cv_broadcast(&l2arc_feed_thr_cv);
8886 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
8891 l2arc_vdev_present(vdev_t *vd)
8895 mutex_enter(&l2arc_dev_mtx);
8896 for (dev = list_head(l2arc_dev_list); dev != NULL;
8897 dev = list_next(l2arc_dev_list, dev)) {
8898 if (dev->l2ad_vdev == vd)
8901 mutex_exit(&l2arc_dev_mtx);
8903 return (dev != NULL);
8907 * Add a vdev for use by the L2ARC. By this point the spa has already
8908 * validated the vdev and opened it.
8911 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
8913 l2arc_dev_t *adddev;
8915 ASSERT(!l2arc_vdev_present(vd));
8918 * Create a new l2arc device entry.
8920 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
8921 adddev->l2ad_spa = spa;
8922 adddev->l2ad_vdev = vd;
8923 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
8924 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
8925 adddev->l2ad_hand = adddev->l2ad_start;
8926 adddev->l2ad_first = B_TRUE;
8927 adddev->l2ad_writing = B_FALSE;
8928 list_link_init(&adddev->l2ad_node);
8930 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
8932 * This is a list of all ARC buffers that are still valid on the
8935 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
8936 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
8938 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
8939 refcount_create(&adddev->l2ad_alloc);
8942 * Add device to global list
8944 mutex_enter(&l2arc_dev_mtx);
8945 list_insert_head(l2arc_dev_list, adddev);
8946 atomic_inc_64(&l2arc_ndev);
8947 mutex_exit(&l2arc_dev_mtx);
8951 * Remove a vdev from the L2ARC.
8954 l2arc_remove_vdev(vdev_t *vd)
8956 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
8959 * Find the device by vdev
8961 mutex_enter(&l2arc_dev_mtx);
8962 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
8963 nextdev = list_next(l2arc_dev_list, dev);
8964 if (vd == dev->l2ad_vdev) {
8969 ASSERT3P(remdev, !=, NULL);
8972 * Remove device from global list
8974 list_remove(l2arc_dev_list, remdev);
8975 l2arc_dev_last = NULL; /* may have been invalidated */
8976 atomic_dec_64(&l2arc_ndev);
8977 mutex_exit(&l2arc_dev_mtx);
8980 * Clear all buflists and ARC references. L2ARC device flush.
8982 l2arc_evict(remdev, 0, B_TRUE);
8983 list_destroy(&remdev->l2ad_buflist);
8984 mutex_destroy(&remdev->l2ad_mtx);
8985 refcount_destroy(&remdev->l2ad_alloc);
8986 kmem_free(remdev, sizeof (l2arc_dev_t));
8992 l2arc_thread_exit = 0;
8994 l2arc_writes_sent = 0;
8995 l2arc_writes_done = 0;
8997 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
8998 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
8999 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
9000 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
9002 l2arc_dev_list = &L2ARC_dev_list;
9003 l2arc_free_on_write = &L2ARC_free_on_write;
9004 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
9005 offsetof(l2arc_dev_t, l2ad_node));
9006 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
9007 offsetof(l2arc_data_free_t, l2df_list_node));
9014 * This is called from dmu_fini(), which is called from spa_fini();
9015 * Because of this, we can assume that all l2arc devices have
9016 * already been removed when the pools themselves were removed.
9019 l2arc_do_free_on_write();
9021 mutex_destroy(&l2arc_feed_thr_lock);
9022 cv_destroy(&l2arc_feed_thr_cv);
9023 mutex_destroy(&l2arc_dev_mtx);
9024 mutex_destroy(&l2arc_free_on_write_mtx);
9026 list_destroy(l2arc_dev_list);
9027 list_destroy(l2arc_free_on_write);
9033 if (!(spa_mode_global & FWRITE))
9036 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
9037 TS_RUN, defclsyspri);
9043 if (!(spa_mode_global & FWRITE))
9046 mutex_enter(&l2arc_feed_thr_lock);
9047 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
9048 l2arc_thread_exit = 1;
9049 while (l2arc_thread_exit != 0)
9050 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
9051 mutex_exit(&l2arc_feed_thr_lock);
9054 #if defined(_KERNEL) && defined(HAVE_SPL)
9055 EXPORT_SYMBOL(arc_buf_size);
9056 EXPORT_SYMBOL(arc_write);
9057 EXPORT_SYMBOL(arc_read);
9058 EXPORT_SYMBOL(arc_buf_info);
9059 EXPORT_SYMBOL(arc_getbuf_func);
9060 EXPORT_SYMBOL(arc_add_prune_callback);
9061 EXPORT_SYMBOL(arc_remove_prune_callback);
9064 module_param(zfs_arc_min, ulong, 0644);
9065 MODULE_PARM_DESC(zfs_arc_min, "Min arc size");
9067 module_param(zfs_arc_max, ulong, 0644);
9068 MODULE_PARM_DESC(zfs_arc_max, "Max arc size");
9070 module_param(zfs_arc_meta_limit, ulong, 0644);
9071 MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size");
9073 module_param(zfs_arc_meta_limit_percent, ulong, 0644);
9074 MODULE_PARM_DESC(zfs_arc_meta_limit_percent,
9075 "Percent of arc size for arc meta limit");
9077 module_param(zfs_arc_meta_min, ulong, 0644);
9078 MODULE_PARM_DESC(zfs_arc_meta_min, "Min arc metadata");
9080 module_param(zfs_arc_meta_prune, int, 0644);
9081 MODULE_PARM_DESC(zfs_arc_meta_prune, "Meta objects to scan for prune");
9083 module_param(zfs_arc_meta_adjust_restarts, int, 0644);
9084 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts,
9085 "Limit number of restarts in arc_adjust_meta");
9087 module_param(zfs_arc_meta_strategy, int, 0644);
9088 MODULE_PARM_DESC(zfs_arc_meta_strategy, "Meta reclaim strategy");
9090 module_param(zfs_arc_grow_retry, int, 0644);
9091 MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size");
9093 module_param(zfs_arc_p_dampener_disable, int, 0644);
9094 MODULE_PARM_DESC(zfs_arc_p_dampener_disable, "disable arc_p adapt dampener");
9096 module_param(zfs_arc_shrink_shift, int, 0644);
9097 MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)");
9099 module_param(zfs_arc_pc_percent, uint, 0644);
9100 MODULE_PARM_DESC(zfs_arc_pc_percent,
9101 "Percent of pagecache to reclaim arc to");
9103 module_param(zfs_arc_p_min_shift, int, 0644);
9104 MODULE_PARM_DESC(zfs_arc_p_min_shift, "arc_c shift to calc min/max arc_p");
9106 module_param(zfs_arc_average_blocksize, int, 0444);
9107 MODULE_PARM_DESC(zfs_arc_average_blocksize, "Target average block size");
9109 module_param(zfs_compressed_arc_enabled, int, 0644);
9110 MODULE_PARM_DESC(zfs_compressed_arc_enabled, "Disable compressed arc buffers");
9112 module_param(zfs_arc_min_prefetch_ms, int, 0644);
9113 MODULE_PARM_DESC(zfs_arc_min_prefetch_ms, "Min life of prefetch block in ms");
9115 module_param(zfs_arc_min_prescient_prefetch_ms, int, 0644);
9116 MODULE_PARM_DESC(zfs_arc_min_prescient_prefetch_ms,
9117 "Min life of prescient prefetched block in ms");
9119 module_param(l2arc_write_max, ulong, 0644);
9120 MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval");
9122 module_param(l2arc_write_boost, ulong, 0644);
9123 MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup");
9125 module_param(l2arc_headroom, ulong, 0644);
9126 MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache");
9128 module_param(l2arc_headroom_boost, ulong, 0644);
9129 MODULE_PARM_DESC(l2arc_headroom_boost, "Compressed l2arc_headroom multiplier");
9131 module_param(l2arc_feed_secs, ulong, 0644);
9132 MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing");
9134 module_param(l2arc_feed_min_ms, ulong, 0644);
9135 MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds");
9137 module_param(l2arc_noprefetch, int, 0644);
9138 MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers");
9140 module_param(l2arc_feed_again, int, 0644);
9141 MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup");
9143 module_param(l2arc_norw, int, 0644);
9144 MODULE_PARM_DESC(l2arc_norw, "No reads during writes");
9146 module_param(zfs_arc_lotsfree_percent, int, 0644);
9147 MODULE_PARM_DESC(zfs_arc_lotsfree_percent,
9148 "System free memory I/O throttle in bytes");
9150 module_param(zfs_arc_sys_free, ulong, 0644);
9151 MODULE_PARM_DESC(zfs_arc_sys_free, "System free memory target size in bytes");
9153 module_param(zfs_arc_dnode_limit, ulong, 0644);
9154 MODULE_PARM_DESC(zfs_arc_dnode_limit, "Minimum bytes of dnodes in arc");
9156 module_param(zfs_arc_dnode_limit_percent, ulong, 0644);
9157 MODULE_PARM_DESC(zfs_arc_dnode_limit_percent,
9158 "Percent of ARC meta buffers for dnodes");
9160 module_param(zfs_arc_dnode_reduce_percent, ulong, 0644);
9161 MODULE_PARM_DESC(zfs_arc_dnode_reduce_percent,
9162 "Percentage of excess dnodes to try to unpin");