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) 2018, Joyent, Inc.
24 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2017 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/shrinker.h>
296 #include <sys/vmsystm.h>
298 #include <linux/page_compat.h>
300 #include <sys/callb.h>
301 #include <sys/kstat.h>
302 #include <sys/zthr.h>
303 #include <zfs_fletcher.h>
304 #include <sys/arc_impl.h>
305 #include <sys/trace_arc.h>
306 #include <sys/aggsum.h>
307 #include <sys/cityhash.h>
310 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
311 boolean_t arc_watch = B_FALSE;
315 * This thread's job is to keep enough free memory in the system, by
316 * calling arc_kmem_reap_soon() plus arc_reduce_target_size(), which improves
317 * arc_available_memory().
319 static zthr_t *arc_reap_zthr;
322 * This thread's job is to keep arc_size under arc_c, by calling
323 * arc_adjust(), which improves arc_is_overflowing().
325 static zthr_t *arc_adjust_zthr;
327 static kmutex_t arc_adjust_lock;
328 static kcondvar_t arc_adjust_waiters_cv;
329 static boolean_t arc_adjust_needed = B_FALSE;
332 * The number of headers to evict in arc_evict_state_impl() before
333 * dropping the sublist lock and evicting from another sublist. A lower
334 * value means we're more likely to evict the "correct" header (i.e. the
335 * oldest header in the arc state), but comes with higher overhead
336 * (i.e. more invocations of arc_evict_state_impl()).
338 int zfs_arc_evict_batch_limit = 10;
340 /* number of seconds before growing cache again */
341 static int arc_grow_retry = 5;
344 * Minimum time between calls to arc_kmem_reap_soon().
346 int arc_kmem_cache_reap_retry_ms = 1000;
348 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
349 int zfs_arc_overflow_shift = 8;
351 /* shift of arc_c for calculating both min and max arc_p */
352 int arc_p_min_shift = 4;
354 /* log2(fraction of arc to reclaim) */
355 static int arc_shrink_shift = 7;
357 /* percent of pagecache to reclaim arc to */
359 static uint_t zfs_arc_pc_percent = 0;
363 * log2(fraction of ARC which must be free to allow growing).
364 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
365 * when reading a new block into the ARC, we will evict an equal-sized block
368 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
369 * we will still not allow it to grow.
371 int arc_no_grow_shift = 5;
375 * minimum lifespan of a prefetch block in clock ticks
376 * (initialized in arc_init())
378 static int arc_min_prefetch_ms;
379 static int arc_min_prescient_prefetch_ms;
382 * If this percent of memory is free, don't throttle.
384 int arc_lotsfree_percent = 10;
387 * hdr_recl() uses this to determine if the arc is up and running.
389 static boolean_t arc_initialized;
392 * The arc has filled available memory and has now warmed up.
394 static boolean_t arc_warm;
397 * log2 fraction of the zio arena to keep free.
399 int arc_zio_arena_free_shift = 2;
402 * These tunables are for performance analysis.
404 unsigned long zfs_arc_max = 0;
405 unsigned long zfs_arc_min = 0;
406 unsigned long zfs_arc_meta_limit = 0;
407 unsigned long zfs_arc_meta_min = 0;
408 unsigned long zfs_arc_dnode_limit = 0;
409 unsigned long zfs_arc_dnode_reduce_percent = 10;
410 int zfs_arc_grow_retry = 0;
411 int zfs_arc_shrink_shift = 0;
412 int zfs_arc_p_min_shift = 0;
413 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
416 * ARC dirty data constraints for arc_tempreserve_space() throttle.
418 unsigned long zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */
419 unsigned long zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */
420 unsigned long zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */
423 * Enable or disable compressed arc buffers.
425 int zfs_compressed_arc_enabled = B_TRUE;
428 * ARC will evict meta buffers that exceed arc_meta_limit. This
429 * tunable make arc_meta_limit adjustable for different workloads.
431 unsigned long zfs_arc_meta_limit_percent = 75;
434 * Percentage that can be consumed by dnodes of ARC meta buffers.
436 unsigned long zfs_arc_dnode_limit_percent = 10;
439 * These tunables are Linux specific
441 unsigned long zfs_arc_sys_free = 0;
442 int zfs_arc_min_prefetch_ms = 0;
443 int zfs_arc_min_prescient_prefetch_ms = 0;
444 int zfs_arc_p_dampener_disable = 1;
445 int zfs_arc_meta_prune = 10000;
446 int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED;
447 int zfs_arc_meta_adjust_restarts = 4096;
448 int zfs_arc_lotsfree_percent = 10;
451 static arc_state_t ARC_anon;
452 static arc_state_t ARC_mru;
453 static arc_state_t ARC_mru_ghost;
454 static arc_state_t ARC_mfu;
455 static arc_state_t ARC_mfu_ghost;
456 static arc_state_t ARC_l2c_only;
458 typedef struct arc_stats {
459 kstat_named_t arcstat_hits;
460 kstat_named_t arcstat_misses;
461 kstat_named_t arcstat_demand_data_hits;
462 kstat_named_t arcstat_demand_data_misses;
463 kstat_named_t arcstat_demand_metadata_hits;
464 kstat_named_t arcstat_demand_metadata_misses;
465 kstat_named_t arcstat_prefetch_data_hits;
466 kstat_named_t arcstat_prefetch_data_misses;
467 kstat_named_t arcstat_prefetch_metadata_hits;
468 kstat_named_t arcstat_prefetch_metadata_misses;
469 kstat_named_t arcstat_mru_hits;
470 kstat_named_t arcstat_mru_ghost_hits;
471 kstat_named_t arcstat_mfu_hits;
472 kstat_named_t arcstat_mfu_ghost_hits;
473 kstat_named_t arcstat_deleted;
475 * Number of buffers that could not be evicted because the hash lock
476 * was held by another thread. The lock may not necessarily be held
477 * by something using the same buffer, since hash locks are shared
478 * by multiple buffers.
480 kstat_named_t arcstat_mutex_miss;
482 * Number of buffers skipped when updating the access state due to the
483 * header having already been released after acquiring the hash lock.
485 kstat_named_t arcstat_access_skip;
487 * Number of buffers skipped because they have I/O in progress, are
488 * indirect prefetch buffers that have not lived long enough, or are
489 * not from the spa we're trying to evict from.
491 kstat_named_t arcstat_evict_skip;
493 * Number of times arc_evict_state() was unable to evict enough
494 * buffers to reach its target amount.
496 kstat_named_t arcstat_evict_not_enough;
497 kstat_named_t arcstat_evict_l2_cached;
498 kstat_named_t arcstat_evict_l2_eligible;
499 kstat_named_t arcstat_evict_l2_ineligible;
500 kstat_named_t arcstat_evict_l2_skip;
501 kstat_named_t arcstat_hash_elements;
502 kstat_named_t arcstat_hash_elements_max;
503 kstat_named_t arcstat_hash_collisions;
504 kstat_named_t arcstat_hash_chains;
505 kstat_named_t arcstat_hash_chain_max;
506 kstat_named_t arcstat_p;
507 kstat_named_t arcstat_c;
508 kstat_named_t arcstat_c_min;
509 kstat_named_t arcstat_c_max;
510 /* Not updated directly; only synced in arc_kstat_update. */
511 kstat_named_t arcstat_size;
513 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
514 * Note that the compressed bytes may match the uncompressed bytes
515 * if the block is either not compressed or compressed arc is disabled.
517 kstat_named_t arcstat_compressed_size;
519 * Uncompressed size of the data stored in b_pabd. If compressed
520 * arc is disabled then this value will be identical to the stat
523 kstat_named_t arcstat_uncompressed_size;
525 * Number of bytes stored in all the arc_buf_t's. This is classified
526 * as "overhead" since this data is typically short-lived and will
527 * be evicted from the arc when it becomes unreferenced unless the
528 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
529 * values have been set (see comment in dbuf.c for more information).
531 kstat_named_t arcstat_overhead_size;
533 * Number of bytes consumed by internal ARC structures necessary
534 * for tracking purposes; these structures are not actually
535 * backed by ARC buffers. This includes arc_buf_hdr_t structures
536 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
537 * caches), and arc_buf_t structures (allocated via arc_buf_t
539 * Not updated directly; only synced in arc_kstat_update.
541 kstat_named_t arcstat_hdr_size;
543 * Number of bytes consumed by ARC buffers of type equal to
544 * ARC_BUFC_DATA. This is generally consumed by buffers backing
545 * on disk user data (e.g. plain file contents).
546 * Not updated directly; only synced in arc_kstat_update.
548 kstat_named_t arcstat_data_size;
550 * Number of bytes consumed by ARC buffers of type equal to
551 * ARC_BUFC_METADATA. This is generally consumed by buffers
552 * backing on disk data that is used for internal ZFS
553 * structures (e.g. ZAP, dnode, indirect blocks, etc).
554 * Not updated directly; only synced in arc_kstat_update.
556 kstat_named_t arcstat_metadata_size;
558 * Number of bytes consumed by dmu_buf_impl_t objects.
559 * Not updated directly; only synced in arc_kstat_update.
561 kstat_named_t arcstat_dbuf_size;
563 * Number of bytes consumed by dnode_t objects.
564 * Not updated directly; only synced in arc_kstat_update.
566 kstat_named_t arcstat_dnode_size;
568 * Number of bytes consumed by bonus buffers.
569 * Not updated directly; only synced in arc_kstat_update.
571 kstat_named_t arcstat_bonus_size;
573 * Total number of bytes consumed by ARC buffers residing in the
574 * arc_anon state. This includes *all* buffers in the arc_anon
575 * state; e.g. data, metadata, evictable, and unevictable buffers
576 * are all included in this value.
577 * Not updated directly; only synced in arc_kstat_update.
579 kstat_named_t arcstat_anon_size;
581 * Number of bytes consumed by ARC buffers that meet the
582 * following criteria: backing buffers of type ARC_BUFC_DATA,
583 * residing in the arc_anon state, and are eligible for eviction
584 * (e.g. have no outstanding holds on the buffer).
585 * Not updated directly; only synced in arc_kstat_update.
587 kstat_named_t arcstat_anon_evictable_data;
589 * Number of bytes consumed by ARC buffers that meet the
590 * following criteria: backing buffers of type ARC_BUFC_METADATA,
591 * residing in the arc_anon state, and are eligible for eviction
592 * (e.g. have no outstanding holds on the buffer).
593 * Not updated directly; only synced in arc_kstat_update.
595 kstat_named_t arcstat_anon_evictable_metadata;
597 * Total number of bytes consumed by ARC buffers residing in the
598 * arc_mru state. This includes *all* buffers in the arc_mru
599 * state; e.g. data, metadata, evictable, and unevictable buffers
600 * are all included in this value.
601 * Not updated directly; only synced in arc_kstat_update.
603 kstat_named_t arcstat_mru_size;
605 * Number of bytes consumed by ARC buffers that meet the
606 * following criteria: backing buffers of type ARC_BUFC_DATA,
607 * residing in the arc_mru state, and are eligible for eviction
608 * (e.g. have no outstanding holds on the buffer).
609 * Not updated directly; only synced in arc_kstat_update.
611 kstat_named_t arcstat_mru_evictable_data;
613 * Number of bytes consumed by ARC buffers that meet the
614 * following criteria: backing buffers of type ARC_BUFC_METADATA,
615 * residing in the arc_mru state, and are eligible for eviction
616 * (e.g. have no outstanding holds on the buffer).
617 * Not updated directly; only synced in arc_kstat_update.
619 kstat_named_t arcstat_mru_evictable_metadata;
621 * Total number of bytes that *would have been* consumed by ARC
622 * buffers in the arc_mru_ghost state. The key thing to note
623 * here, is the fact that this size doesn't actually indicate
624 * RAM consumption. The ghost lists only consist of headers and
625 * don't actually have ARC buffers linked off of these headers.
626 * Thus, *if* the headers had associated ARC buffers, these
627 * buffers *would have* consumed this number of bytes.
628 * Not updated directly; only synced in arc_kstat_update.
630 kstat_named_t arcstat_mru_ghost_size;
632 * Number of bytes that *would have been* consumed by ARC
633 * buffers that are eligible for eviction, of type
634 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
635 * Not updated directly; only synced in arc_kstat_update.
637 kstat_named_t arcstat_mru_ghost_evictable_data;
639 * Number of bytes that *would have been* consumed by ARC
640 * buffers that are eligible for eviction, of type
641 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
642 * Not updated directly; only synced in arc_kstat_update.
644 kstat_named_t arcstat_mru_ghost_evictable_metadata;
646 * Total number of bytes consumed by ARC buffers residing in the
647 * arc_mfu state. This includes *all* buffers in the arc_mfu
648 * state; e.g. data, metadata, evictable, and unevictable buffers
649 * are all included in this value.
650 * Not updated directly; only synced in arc_kstat_update.
652 kstat_named_t arcstat_mfu_size;
654 * Number of bytes consumed by ARC buffers that are eligible for
655 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
657 * Not updated directly; only synced in arc_kstat_update.
659 kstat_named_t arcstat_mfu_evictable_data;
661 * Number of bytes consumed by ARC buffers that are eligible for
662 * eviction, of type ARC_BUFC_METADATA, and reside in the
664 * Not updated directly; only synced in arc_kstat_update.
666 kstat_named_t arcstat_mfu_evictable_metadata;
668 * Total number of bytes that *would have been* consumed by ARC
669 * buffers in the arc_mfu_ghost state. See the comment above
670 * arcstat_mru_ghost_size for more details.
671 * Not updated directly; only synced in arc_kstat_update.
673 kstat_named_t arcstat_mfu_ghost_size;
675 * Number of bytes that *would have been* consumed by ARC
676 * buffers that are eligible for eviction, of type
677 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
678 * Not updated directly; only synced in arc_kstat_update.
680 kstat_named_t arcstat_mfu_ghost_evictable_data;
682 * Number of bytes that *would have been* consumed by ARC
683 * buffers that are eligible for eviction, of type
684 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
685 * Not updated directly; only synced in arc_kstat_update.
687 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
688 kstat_named_t arcstat_l2_hits;
689 kstat_named_t arcstat_l2_misses;
690 kstat_named_t arcstat_l2_feeds;
691 kstat_named_t arcstat_l2_rw_clash;
692 kstat_named_t arcstat_l2_read_bytes;
693 kstat_named_t arcstat_l2_write_bytes;
694 kstat_named_t arcstat_l2_writes_sent;
695 kstat_named_t arcstat_l2_writes_done;
696 kstat_named_t arcstat_l2_writes_error;
697 kstat_named_t arcstat_l2_writes_lock_retry;
698 kstat_named_t arcstat_l2_evict_lock_retry;
699 kstat_named_t arcstat_l2_evict_reading;
700 kstat_named_t arcstat_l2_evict_l1cached;
701 kstat_named_t arcstat_l2_free_on_write;
702 kstat_named_t arcstat_l2_abort_lowmem;
703 kstat_named_t arcstat_l2_cksum_bad;
704 kstat_named_t arcstat_l2_io_error;
705 kstat_named_t arcstat_l2_lsize;
706 kstat_named_t arcstat_l2_psize;
707 /* Not updated directly; only synced in arc_kstat_update. */
708 kstat_named_t arcstat_l2_hdr_size;
709 kstat_named_t arcstat_memory_throttle_count;
710 kstat_named_t arcstat_memory_direct_count;
711 kstat_named_t arcstat_memory_indirect_count;
712 kstat_named_t arcstat_memory_all_bytes;
713 kstat_named_t arcstat_memory_free_bytes;
714 kstat_named_t arcstat_memory_available_bytes;
715 kstat_named_t arcstat_no_grow;
716 kstat_named_t arcstat_tempreserve;
717 kstat_named_t arcstat_loaned_bytes;
718 kstat_named_t arcstat_prune;
719 /* Not updated directly; only synced in arc_kstat_update. */
720 kstat_named_t arcstat_meta_used;
721 kstat_named_t arcstat_meta_limit;
722 kstat_named_t arcstat_dnode_limit;
723 kstat_named_t arcstat_meta_max;
724 kstat_named_t arcstat_meta_min;
725 kstat_named_t arcstat_async_upgrade_sync;
726 kstat_named_t arcstat_demand_hit_predictive_prefetch;
727 kstat_named_t arcstat_demand_hit_prescient_prefetch;
728 kstat_named_t arcstat_need_free;
729 kstat_named_t arcstat_sys_free;
730 kstat_named_t arcstat_raw_size;
733 static arc_stats_t arc_stats = {
734 { "hits", KSTAT_DATA_UINT64 },
735 { "misses", KSTAT_DATA_UINT64 },
736 { "demand_data_hits", KSTAT_DATA_UINT64 },
737 { "demand_data_misses", KSTAT_DATA_UINT64 },
738 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
739 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
740 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
741 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
742 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
743 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
744 { "mru_hits", KSTAT_DATA_UINT64 },
745 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
746 { "mfu_hits", KSTAT_DATA_UINT64 },
747 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
748 { "deleted", KSTAT_DATA_UINT64 },
749 { "mutex_miss", KSTAT_DATA_UINT64 },
750 { "access_skip", KSTAT_DATA_UINT64 },
751 { "evict_skip", KSTAT_DATA_UINT64 },
752 { "evict_not_enough", KSTAT_DATA_UINT64 },
753 { "evict_l2_cached", KSTAT_DATA_UINT64 },
754 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
755 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
756 { "evict_l2_skip", KSTAT_DATA_UINT64 },
757 { "hash_elements", KSTAT_DATA_UINT64 },
758 { "hash_elements_max", KSTAT_DATA_UINT64 },
759 { "hash_collisions", KSTAT_DATA_UINT64 },
760 { "hash_chains", KSTAT_DATA_UINT64 },
761 { "hash_chain_max", KSTAT_DATA_UINT64 },
762 { "p", KSTAT_DATA_UINT64 },
763 { "c", KSTAT_DATA_UINT64 },
764 { "c_min", KSTAT_DATA_UINT64 },
765 { "c_max", KSTAT_DATA_UINT64 },
766 { "size", KSTAT_DATA_UINT64 },
767 { "compressed_size", KSTAT_DATA_UINT64 },
768 { "uncompressed_size", KSTAT_DATA_UINT64 },
769 { "overhead_size", KSTAT_DATA_UINT64 },
770 { "hdr_size", KSTAT_DATA_UINT64 },
771 { "data_size", KSTAT_DATA_UINT64 },
772 { "metadata_size", KSTAT_DATA_UINT64 },
773 { "dbuf_size", KSTAT_DATA_UINT64 },
774 { "dnode_size", KSTAT_DATA_UINT64 },
775 { "bonus_size", KSTAT_DATA_UINT64 },
776 { "anon_size", KSTAT_DATA_UINT64 },
777 { "anon_evictable_data", KSTAT_DATA_UINT64 },
778 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
779 { "mru_size", KSTAT_DATA_UINT64 },
780 { "mru_evictable_data", KSTAT_DATA_UINT64 },
781 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
782 { "mru_ghost_size", KSTAT_DATA_UINT64 },
783 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
784 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
785 { "mfu_size", KSTAT_DATA_UINT64 },
786 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
787 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
788 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
789 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
790 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
791 { "l2_hits", KSTAT_DATA_UINT64 },
792 { "l2_misses", KSTAT_DATA_UINT64 },
793 { "l2_feeds", KSTAT_DATA_UINT64 },
794 { "l2_rw_clash", KSTAT_DATA_UINT64 },
795 { "l2_read_bytes", KSTAT_DATA_UINT64 },
796 { "l2_write_bytes", KSTAT_DATA_UINT64 },
797 { "l2_writes_sent", KSTAT_DATA_UINT64 },
798 { "l2_writes_done", KSTAT_DATA_UINT64 },
799 { "l2_writes_error", KSTAT_DATA_UINT64 },
800 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
801 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
802 { "l2_evict_reading", KSTAT_DATA_UINT64 },
803 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
804 { "l2_free_on_write", KSTAT_DATA_UINT64 },
805 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
806 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
807 { "l2_io_error", KSTAT_DATA_UINT64 },
808 { "l2_size", KSTAT_DATA_UINT64 },
809 { "l2_asize", KSTAT_DATA_UINT64 },
810 { "l2_hdr_size", KSTAT_DATA_UINT64 },
811 { "memory_throttle_count", KSTAT_DATA_UINT64 },
812 { "memory_direct_count", KSTAT_DATA_UINT64 },
813 { "memory_indirect_count", KSTAT_DATA_UINT64 },
814 { "memory_all_bytes", KSTAT_DATA_UINT64 },
815 { "memory_free_bytes", KSTAT_DATA_UINT64 },
816 { "memory_available_bytes", KSTAT_DATA_INT64 },
817 { "arc_no_grow", KSTAT_DATA_UINT64 },
818 { "arc_tempreserve", KSTAT_DATA_UINT64 },
819 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
820 { "arc_prune", KSTAT_DATA_UINT64 },
821 { "arc_meta_used", KSTAT_DATA_UINT64 },
822 { "arc_meta_limit", KSTAT_DATA_UINT64 },
823 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
824 { "arc_meta_max", KSTAT_DATA_UINT64 },
825 { "arc_meta_min", KSTAT_DATA_UINT64 },
826 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
827 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
828 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
829 { "arc_need_free", KSTAT_DATA_UINT64 },
830 { "arc_sys_free", KSTAT_DATA_UINT64 },
831 { "arc_raw_size", KSTAT_DATA_UINT64 }
834 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
836 #define ARCSTAT_INCR(stat, val) \
837 atomic_add_64(&arc_stats.stat.value.ui64, (val))
839 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
840 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
842 #define ARCSTAT_MAX(stat, val) { \
844 while ((val) > (m = arc_stats.stat.value.ui64) && \
845 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
849 #define ARCSTAT_MAXSTAT(stat) \
850 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
853 * We define a macro to allow ARC hits/misses to be easily broken down by
854 * two separate conditions, giving a total of four different subtypes for
855 * each of hits and misses (so eight statistics total).
857 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
860 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
862 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
866 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
868 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
873 static arc_state_t *arc_anon;
874 static arc_state_t *arc_mru;
875 static arc_state_t *arc_mru_ghost;
876 static arc_state_t *arc_mfu;
877 static arc_state_t *arc_mfu_ghost;
878 static arc_state_t *arc_l2c_only;
881 * There are several ARC variables that are critical to export as kstats --
882 * but we don't want to have to grovel around in the kstat whenever we wish to
883 * manipulate them. For these variables, we therefore define them to be in
884 * terms of the statistic variable. This assures that we are not introducing
885 * the possibility of inconsistency by having shadow copies of the variables,
886 * while still allowing the code to be readable.
888 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
889 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
890 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
891 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
892 #define arc_no_grow ARCSTAT(arcstat_no_grow) /* do not grow cache size */
893 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
894 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
895 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
896 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
897 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
898 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
899 #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */
900 #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */
902 /* size of all b_rabd's in entire arc */
903 #define arc_raw_size ARCSTAT(arcstat_raw_size)
904 /* compressed size of entire arc */
905 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
906 /* uncompressed size of entire arc */
907 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
908 /* number of bytes in the arc from arc_buf_t's */
909 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
912 * There are also some ARC variables that we want to export, but that are
913 * updated so often that having the canonical representation be the statistic
914 * variable causes a performance bottleneck. We want to use aggsum_t's for these
915 * instead, but still be able to export the kstat in the same way as before.
916 * The solution is to always use the aggsum version, except in the kstat update
920 aggsum_t arc_meta_used;
921 aggsum_t astat_data_size;
922 aggsum_t astat_metadata_size;
923 aggsum_t astat_dbuf_size;
924 aggsum_t astat_dnode_size;
925 aggsum_t astat_bonus_size;
926 aggsum_t astat_hdr_size;
927 aggsum_t astat_l2_hdr_size;
929 static hrtime_t arc_growtime;
930 static list_t arc_prune_list;
931 static kmutex_t arc_prune_mtx;
932 static taskq_t *arc_prune_taskq;
934 #define GHOST_STATE(state) \
935 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
936 (state) == arc_l2c_only)
938 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
939 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
940 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
941 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
942 #define HDR_PRESCIENT_PREFETCH(hdr) \
943 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
944 #define HDR_COMPRESSION_ENABLED(hdr) \
945 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
947 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
948 #define HDR_L2_READING(hdr) \
949 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
950 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
951 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
952 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
953 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
954 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
955 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
956 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
958 #define HDR_ISTYPE_METADATA(hdr) \
959 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
960 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
962 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
963 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
964 #define HDR_HAS_RABD(hdr) \
965 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
966 (hdr)->b_crypt_hdr.b_rabd != NULL)
967 #define HDR_ENCRYPTED(hdr) \
968 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
969 #define HDR_AUTHENTICATED(hdr) \
970 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
972 /* For storing compression mode in b_flags */
973 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
975 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
976 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
977 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
978 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
980 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
981 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
982 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
983 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
989 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
990 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
991 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
994 * Hash table routines
997 #define HT_LOCK_ALIGN 64
998 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
1003 unsigned char pad[HT_LOCK_PAD];
1007 #define BUF_LOCKS 8192
1008 typedef struct buf_hash_table {
1010 arc_buf_hdr_t **ht_table;
1011 struct ht_lock ht_locks[BUF_LOCKS];
1014 static buf_hash_table_t buf_hash_table;
1016 #define BUF_HASH_INDEX(spa, dva, birth) \
1017 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1018 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1019 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1020 #define HDR_LOCK(hdr) \
1021 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1023 uint64_t zfs_crc64_table[256];
1029 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1030 #define L2ARC_HEADROOM 2 /* num of writes */
1033 * If we discover during ARC scan any buffers to be compressed, we boost
1034 * our headroom for the next scanning cycle by this percentage multiple.
1036 #define L2ARC_HEADROOM_BOOST 200
1037 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1038 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1041 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
1042 * and each of the state has two types: data and metadata.
1044 #define L2ARC_FEED_TYPES 4
1046 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1047 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1049 /* L2ARC Performance Tunables */
1050 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
1051 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
1052 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
1053 unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1054 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1055 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
1056 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1057 int l2arc_feed_again = B_TRUE; /* turbo warmup */
1058 int l2arc_norw = B_FALSE; /* no reads during writes */
1063 static list_t L2ARC_dev_list; /* device list */
1064 static list_t *l2arc_dev_list; /* device list pointer */
1065 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1066 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1067 static list_t L2ARC_free_on_write; /* free after write buf list */
1068 static list_t *l2arc_free_on_write; /* free after write list ptr */
1069 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1070 static uint64_t l2arc_ndev; /* number of devices */
1072 typedef struct l2arc_read_callback {
1073 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1074 blkptr_t l2rcb_bp; /* original blkptr */
1075 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1076 int l2rcb_flags; /* original flags */
1077 abd_t *l2rcb_abd; /* temporary buffer */
1078 } l2arc_read_callback_t;
1080 typedef struct l2arc_data_free {
1081 /* protected by l2arc_free_on_write_mtx */
1084 arc_buf_contents_t l2df_type;
1085 list_node_t l2df_list_node;
1086 } l2arc_data_free_t;
1088 typedef enum arc_fill_flags {
1089 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
1090 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
1091 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
1092 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
1093 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
1096 static kmutex_t l2arc_feed_thr_lock;
1097 static kcondvar_t l2arc_feed_thr_cv;
1098 static uint8_t l2arc_thread_exit;
1100 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1101 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1102 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1103 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1104 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1105 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1106 static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t);
1107 static void arc_hdr_alloc_abd(arc_buf_hdr_t *, boolean_t);
1108 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1109 static boolean_t arc_is_overflowing(void);
1110 static void arc_buf_watch(arc_buf_t *);
1111 static void arc_tuning_update(void);
1112 static void arc_prune_async(int64_t);
1113 static uint64_t arc_all_memory(void);
1115 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1116 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1117 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1118 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1120 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1121 static void l2arc_read_done(zio_t *);
1125 * We use Cityhash for this. It's fast, and has good hash properties without
1126 * requiring any large static buffers.
1129 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1131 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1134 #define HDR_EMPTY(hdr) \
1135 ((hdr)->b_dva.dva_word[0] == 0 && \
1136 (hdr)->b_dva.dva_word[1] == 0)
1138 #define HDR_EQUAL(spa, dva, birth, hdr) \
1139 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1140 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1141 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1144 buf_discard_identity(arc_buf_hdr_t *hdr)
1146 hdr->b_dva.dva_word[0] = 0;
1147 hdr->b_dva.dva_word[1] = 0;
1151 static arc_buf_hdr_t *
1152 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1154 const dva_t *dva = BP_IDENTITY(bp);
1155 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1156 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1157 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1160 mutex_enter(hash_lock);
1161 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1162 hdr = hdr->b_hash_next) {
1163 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1168 mutex_exit(hash_lock);
1174 * Insert an entry into the hash table. If there is already an element
1175 * equal to elem in the hash table, then the already existing element
1176 * will be returned and the new element will not be inserted.
1177 * Otherwise returns NULL.
1178 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1180 static arc_buf_hdr_t *
1181 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1183 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1184 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1185 arc_buf_hdr_t *fhdr;
1188 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1189 ASSERT(hdr->b_birth != 0);
1190 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1192 if (lockp != NULL) {
1194 mutex_enter(hash_lock);
1196 ASSERT(MUTEX_HELD(hash_lock));
1199 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1200 fhdr = fhdr->b_hash_next, i++) {
1201 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1205 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1206 buf_hash_table.ht_table[idx] = hdr;
1207 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1209 /* collect some hash table performance data */
1211 ARCSTAT_BUMP(arcstat_hash_collisions);
1213 ARCSTAT_BUMP(arcstat_hash_chains);
1215 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1218 ARCSTAT_BUMP(arcstat_hash_elements);
1219 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1225 buf_hash_remove(arc_buf_hdr_t *hdr)
1227 arc_buf_hdr_t *fhdr, **hdrp;
1228 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1230 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1231 ASSERT(HDR_IN_HASH_TABLE(hdr));
1233 hdrp = &buf_hash_table.ht_table[idx];
1234 while ((fhdr = *hdrp) != hdr) {
1235 ASSERT3P(fhdr, !=, NULL);
1236 hdrp = &fhdr->b_hash_next;
1238 *hdrp = hdr->b_hash_next;
1239 hdr->b_hash_next = NULL;
1240 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1242 /* collect some hash table performance data */
1243 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1245 if (buf_hash_table.ht_table[idx] &&
1246 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1247 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1251 * Global data structures and functions for the buf kmem cache.
1254 static kmem_cache_t *hdr_full_cache;
1255 static kmem_cache_t *hdr_full_crypt_cache;
1256 static kmem_cache_t *hdr_l2only_cache;
1257 static kmem_cache_t *buf_cache;
1264 #if defined(_KERNEL)
1266 * Large allocations which do not require contiguous pages
1267 * should be using vmem_free() in the linux kernel\
1269 vmem_free(buf_hash_table.ht_table,
1270 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1272 kmem_free(buf_hash_table.ht_table,
1273 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1275 for (i = 0; i < BUF_LOCKS; i++)
1276 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1277 kmem_cache_destroy(hdr_full_cache);
1278 kmem_cache_destroy(hdr_full_crypt_cache);
1279 kmem_cache_destroy(hdr_l2only_cache);
1280 kmem_cache_destroy(buf_cache);
1284 * Constructor callback - called when the cache is empty
1285 * and a new buf is requested.
1289 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1291 arc_buf_hdr_t *hdr = vbuf;
1293 bzero(hdr, HDR_FULL_SIZE);
1294 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1295 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1296 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1297 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1298 list_link_init(&hdr->b_l1hdr.b_arc_node);
1299 list_link_init(&hdr->b_l2hdr.b_l2node);
1300 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1301 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1308 hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
1310 arc_buf_hdr_t *hdr = vbuf;
1312 hdr_full_cons(vbuf, unused, kmflag);
1313 bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr));
1314 arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1321 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1323 arc_buf_hdr_t *hdr = vbuf;
1325 bzero(hdr, HDR_L2ONLY_SIZE);
1326 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1333 buf_cons(void *vbuf, void *unused, int kmflag)
1335 arc_buf_t *buf = vbuf;
1337 bzero(buf, sizeof (arc_buf_t));
1338 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1339 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1345 * Destructor callback - called when a cached buf is
1346 * no longer required.
1350 hdr_full_dest(void *vbuf, void *unused)
1352 arc_buf_hdr_t *hdr = vbuf;
1354 ASSERT(HDR_EMPTY(hdr));
1355 cv_destroy(&hdr->b_l1hdr.b_cv);
1356 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1357 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1358 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1359 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1364 hdr_full_crypt_dest(void *vbuf, void *unused)
1366 arc_buf_hdr_t *hdr = vbuf;
1368 hdr_full_dest(vbuf, unused);
1369 arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1374 hdr_l2only_dest(void *vbuf, void *unused)
1376 ASSERTV(arc_buf_hdr_t *hdr = vbuf);
1378 ASSERT(HDR_EMPTY(hdr));
1379 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1384 buf_dest(void *vbuf, void *unused)
1386 arc_buf_t *buf = vbuf;
1388 mutex_destroy(&buf->b_evict_lock);
1389 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1393 * Reclaim callback -- invoked when memory is low.
1397 hdr_recl(void *unused)
1399 dprintf("hdr_recl called\n");
1401 * umem calls the reclaim func when we destroy the buf cache,
1402 * which is after we do arc_fini().
1404 if (arc_initialized)
1405 zthr_wakeup(arc_reap_zthr);
1411 uint64_t *ct = NULL;
1412 uint64_t hsize = 1ULL << 12;
1416 * The hash table is big enough to fill all of physical memory
1417 * with an average block size of zfs_arc_average_blocksize (default 8K).
1418 * By default, the table will take up
1419 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1421 while (hsize * zfs_arc_average_blocksize < arc_all_memory())
1424 buf_hash_table.ht_mask = hsize - 1;
1425 #if defined(_KERNEL)
1427 * Large allocations which do not require contiguous pages
1428 * should be using vmem_alloc() in the linux kernel
1430 buf_hash_table.ht_table =
1431 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
1433 buf_hash_table.ht_table =
1434 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1436 if (buf_hash_table.ht_table == NULL) {
1437 ASSERT(hsize > (1ULL << 8));
1442 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1443 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1444 hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
1445 HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
1446 hdr_recl, NULL, NULL, 0);
1447 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1448 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1450 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1451 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1453 for (i = 0; i < 256; i++)
1454 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1455 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1457 for (i = 0; i < BUF_LOCKS; i++) {
1458 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1459 NULL, MUTEX_DEFAULT, NULL);
1463 #define ARC_MINTIME (hz>>4) /* 62 ms */
1466 * This is the size that the buf occupies in memory. If the buf is compressed,
1467 * it will correspond to the compressed size. You should use this method of
1468 * getting the buf size unless you explicitly need the logical size.
1471 arc_buf_size(arc_buf_t *buf)
1473 return (ARC_BUF_COMPRESSED(buf) ?
1474 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1478 arc_buf_lsize(arc_buf_t *buf)
1480 return (HDR_GET_LSIZE(buf->b_hdr));
1484 * This function will return B_TRUE if the buffer is encrypted in memory.
1485 * This buffer can be decrypted by calling arc_untransform().
1488 arc_is_encrypted(arc_buf_t *buf)
1490 return (ARC_BUF_ENCRYPTED(buf) != 0);
1494 * Returns B_TRUE if the buffer represents data that has not had its MAC
1498 arc_is_unauthenticated(arc_buf_t *buf)
1500 return (HDR_NOAUTH(buf->b_hdr) != 0);
1504 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1505 uint8_t *iv, uint8_t *mac)
1507 arc_buf_hdr_t *hdr = buf->b_hdr;
1509 ASSERT(HDR_PROTECTED(hdr));
1511 bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
1512 bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
1513 bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
1514 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1515 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1519 * Indicates how this buffer is compressed in memory. If it is not compressed
1520 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1521 * arc_untransform() as long as it is also unencrypted.
1524 arc_get_compression(arc_buf_t *buf)
1526 return (ARC_BUF_COMPRESSED(buf) ?
1527 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1531 * Return the compression algorithm used to store this data in the ARC. If ARC
1532 * compression is enabled or this is an encrypted block, this will be the same
1533 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1535 static inline enum zio_compress
1536 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1538 return (HDR_COMPRESSION_ENABLED(hdr) ?
1539 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1542 static inline boolean_t
1543 arc_buf_is_shared(arc_buf_t *buf)
1545 boolean_t shared = (buf->b_data != NULL &&
1546 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1547 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1548 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1549 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1550 IMPLY(shared, ARC_BUF_SHARED(buf));
1551 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1554 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1555 * already being shared" requirement prevents us from doing that.
1562 * Free the checksum associated with this header. If there is no checksum, this
1566 arc_cksum_free(arc_buf_hdr_t *hdr)
1568 ASSERT(HDR_HAS_L1HDR(hdr));
1570 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1571 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1572 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1573 hdr->b_l1hdr.b_freeze_cksum = NULL;
1575 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1579 * Return true iff at least one of the bufs on hdr is not compressed.
1580 * Encrypted buffers count as compressed.
1583 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1585 ASSERT(hdr->b_l1hdr.b_state == arc_anon ||
1586 MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1588 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1589 if (!ARC_BUF_COMPRESSED(b)) {
1598 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1599 * matches the checksum that is stored in the hdr. If there is no checksum,
1600 * or if the buf is compressed, this is a no-op.
1603 arc_cksum_verify(arc_buf_t *buf)
1605 arc_buf_hdr_t *hdr = buf->b_hdr;
1608 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1611 if (ARC_BUF_COMPRESSED(buf))
1614 ASSERT(HDR_HAS_L1HDR(hdr));
1616 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1618 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1619 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1623 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1624 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1625 panic("buffer modified while frozen!");
1626 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1630 * This function makes the assumption that data stored in the L2ARC
1631 * will be transformed exactly as it is in the main pool. Because of
1632 * this we can verify the checksum against the reading process's bp.
1635 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1637 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1638 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1641 * Block pointers always store the checksum for the logical data.
1642 * If the block pointer has the gang bit set, then the checksum
1643 * it represents is for the reconstituted data and not for an
1644 * individual gang member. The zio pipeline, however, must be able to
1645 * determine the checksum of each of the gang constituents so it
1646 * treats the checksum comparison differently than what we need
1647 * for l2arc blocks. This prevents us from using the
1648 * zio_checksum_error() interface directly. Instead we must call the
1649 * zio_checksum_error_impl() so that we can ensure the checksum is
1650 * generated using the correct checksum algorithm and accounts for the
1651 * logical I/O size and not just a gang fragment.
1653 return (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1654 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1655 zio->io_offset, NULL) == 0);
1659 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1660 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1661 * isn't modified later on. If buf is compressed or there is already a checksum
1662 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1665 arc_cksum_compute(arc_buf_t *buf)
1667 arc_buf_hdr_t *hdr = buf->b_hdr;
1669 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1672 ASSERT(HDR_HAS_L1HDR(hdr));
1674 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1675 if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1676 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1680 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1681 ASSERT(!ARC_BUF_COMPRESSED(buf));
1682 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1684 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1685 hdr->b_l1hdr.b_freeze_cksum);
1686 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1692 arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
1694 panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr);
1700 arc_buf_unwatch(arc_buf_t *buf)
1704 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1705 PROT_READ | PROT_WRITE));
1712 arc_buf_watch(arc_buf_t *buf)
1716 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1721 static arc_buf_contents_t
1722 arc_buf_type(arc_buf_hdr_t *hdr)
1724 arc_buf_contents_t type;
1725 if (HDR_ISTYPE_METADATA(hdr)) {
1726 type = ARC_BUFC_METADATA;
1728 type = ARC_BUFC_DATA;
1730 VERIFY3U(hdr->b_type, ==, type);
1735 arc_is_metadata(arc_buf_t *buf)
1737 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1741 arc_bufc_to_flags(arc_buf_contents_t type)
1745 /* metadata field is 0 if buffer contains normal data */
1747 case ARC_BUFC_METADATA:
1748 return (ARC_FLAG_BUFC_METADATA);
1752 panic("undefined ARC buffer type!");
1753 return ((uint32_t)-1);
1757 arc_buf_thaw(arc_buf_t *buf)
1759 arc_buf_hdr_t *hdr = buf->b_hdr;
1761 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1762 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1764 arc_cksum_verify(buf);
1767 * Compressed buffers do not manipulate the b_freeze_cksum.
1769 if (ARC_BUF_COMPRESSED(buf))
1772 ASSERT(HDR_HAS_L1HDR(hdr));
1773 arc_cksum_free(hdr);
1774 arc_buf_unwatch(buf);
1778 arc_buf_freeze(arc_buf_t *buf)
1780 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1783 if (ARC_BUF_COMPRESSED(buf))
1786 ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1787 arc_cksum_compute(buf);
1791 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1792 * the following functions should be used to ensure that the flags are
1793 * updated in a thread-safe way. When manipulating the flags either
1794 * the hash_lock must be held or the hdr must be undiscoverable. This
1795 * ensures that we're not racing with any other threads when updating
1799 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1801 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1802 hdr->b_flags |= flags;
1806 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1808 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1809 hdr->b_flags &= ~flags;
1813 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1814 * done in a special way since we have to clear and set bits
1815 * at the same time. Consumers that wish to set the compression bits
1816 * must use this function to ensure that the flags are updated in
1817 * thread-safe manner.
1820 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1822 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1825 * Holes and embedded blocks will always have a psize = 0 so
1826 * we ignore the compression of the blkptr and set the
1827 * want to uncompress them. Mark them as uncompressed.
1829 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1830 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1831 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1833 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1834 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1837 HDR_SET_COMPRESS(hdr, cmp);
1838 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1842 * Looks for another buf on the same hdr which has the data decompressed, copies
1843 * from it, and returns true. If no such buf exists, returns false.
1846 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1848 arc_buf_hdr_t *hdr = buf->b_hdr;
1849 boolean_t copied = B_FALSE;
1851 ASSERT(HDR_HAS_L1HDR(hdr));
1852 ASSERT3P(buf->b_data, !=, NULL);
1853 ASSERT(!ARC_BUF_COMPRESSED(buf));
1855 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1856 from = from->b_next) {
1857 /* can't use our own data buffer */
1862 if (!ARC_BUF_COMPRESSED(from)) {
1863 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1870 * There were no decompressed bufs, so there should not be a
1871 * checksum on the hdr either.
1873 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1879 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1882 arc_hdr_size(arc_buf_hdr_t *hdr)
1886 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1887 HDR_GET_PSIZE(hdr) > 0) {
1888 size = HDR_GET_PSIZE(hdr);
1890 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1891 size = HDR_GET_LSIZE(hdr);
1897 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1901 uint64_t lsize = HDR_GET_LSIZE(hdr);
1902 uint64_t psize = HDR_GET_PSIZE(hdr);
1903 void *tmpbuf = NULL;
1904 abd_t *abd = hdr->b_l1hdr.b_pabd;
1906 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1907 ASSERT(HDR_AUTHENTICATED(hdr));
1908 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1911 * The MAC is calculated on the compressed data that is stored on disk.
1912 * However, if compressed arc is disabled we will only have the
1913 * decompressed data available to us now. Compress it into a temporary
1914 * abd so we can verify the MAC. The performance overhead of this will
1915 * be relatively low, since most objects in an encrypted objset will
1916 * be encrypted (instead of authenticated) anyway.
1918 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1919 !HDR_COMPRESSION_ENABLED(hdr)) {
1920 tmpbuf = zio_buf_alloc(lsize);
1921 abd = abd_get_from_buf(tmpbuf, lsize);
1922 abd_take_ownership_of_buf(abd, B_TRUE);
1924 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1925 hdr->b_l1hdr.b_pabd, tmpbuf, lsize);
1926 ASSERT3U(csize, <=, psize);
1927 abd_zero_off(abd, csize, psize - csize);
1931 * Authentication is best effort. We authenticate whenever the key is
1932 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1934 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1935 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1936 ASSERT3U(lsize, ==, psize);
1937 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1938 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1940 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1941 hdr->b_crypt_hdr.b_mac);
1945 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1946 else if (ret != ENOENT)
1962 * This function will take a header that only has raw encrypted data in
1963 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1964 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1965 * also decompress the data.
1968 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1973 boolean_t no_crypt = B_FALSE;
1974 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1976 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1977 ASSERT(HDR_ENCRYPTED(hdr));
1979 arc_hdr_alloc_abd(hdr, B_FALSE);
1981 ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1982 B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1983 hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1984 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1989 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1990 HDR_GET_PSIZE(hdr));
1994 * If this header has disabled arc compression but the b_pabd is
1995 * compressed after decrypting it, we need to decompress the newly
1998 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1999 !HDR_COMPRESSION_ENABLED(hdr)) {
2001 * We want to make sure that we are correctly honoring the
2002 * zfs_abd_scatter_enabled setting, so we allocate an abd here
2003 * and then loan a buffer from it, rather than allocating a
2004 * linear buffer and wrapping it in an abd later.
2006 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
2007 tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
2009 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2010 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
2011 HDR_GET_LSIZE(hdr));
2013 abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
2017 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
2018 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
2019 arc_hdr_size(hdr), hdr);
2020 hdr->b_l1hdr.b_pabd = cabd;
2026 arc_hdr_free_abd(hdr, B_FALSE);
2028 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
2034 * This function is called during arc_buf_fill() to prepare the header's
2035 * abd plaintext pointer for use. This involves authenticated protected
2036 * data and decrypting encrypted data into the plaintext abd.
2039 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
2040 const zbookmark_phys_t *zb, boolean_t noauth)
2044 ASSERT(HDR_PROTECTED(hdr));
2046 if (hash_lock != NULL)
2047 mutex_enter(hash_lock);
2049 if (HDR_NOAUTH(hdr) && !noauth) {
2051 * The caller requested authenticated data but our data has
2052 * not been authenticated yet. Verify the MAC now if we can.
2054 ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
2057 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
2059 * If we only have the encrypted version of the data, but the
2060 * unencrypted version was requested we take this opportunity
2061 * to store the decrypted version in the header for future use.
2063 ret = arc_hdr_decrypt(hdr, spa, zb);
2068 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2070 if (hash_lock != NULL)
2071 mutex_exit(hash_lock);
2076 if (hash_lock != NULL)
2077 mutex_exit(hash_lock);
2083 * This function is used by the dbuf code to decrypt bonus buffers in place.
2084 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
2085 * block, so we use the hash lock here to protect against concurrent calls to
2089 arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock)
2091 arc_buf_hdr_t *hdr = buf->b_hdr;
2093 ASSERT(HDR_ENCRYPTED(hdr));
2094 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2095 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2096 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2098 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
2100 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
2101 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2102 hdr->b_crypt_hdr.b_ebufcnt -= 1;
2106 * Given a buf that has a data buffer attached to it, this function will
2107 * efficiently fill the buf with data of the specified compression setting from
2108 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2109 * are already sharing a data buf, no copy is performed.
2111 * If the buf is marked as compressed but uncompressed data was requested, this
2112 * will allocate a new data buffer for the buf, remove that flag, and fill the
2113 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2114 * uncompressed data, and (since we haven't added support for it yet) if you
2115 * want compressed data your buf must already be marked as compressed and have
2116 * the correct-sized data buffer.
2119 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2120 arc_fill_flags_t flags)
2123 arc_buf_hdr_t *hdr = buf->b_hdr;
2124 boolean_t hdr_compressed =
2125 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
2126 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
2127 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
2128 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2129 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
2131 ASSERT3P(buf->b_data, !=, NULL);
2132 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
2133 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2134 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
2135 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
2136 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
2137 IMPLY(encrypted, !ARC_BUF_SHARED(buf));
2140 * If the caller wanted encrypted data we just need to copy it from
2141 * b_rabd and potentially byteswap it. We won't be able to do any
2142 * further transforms on it.
2145 ASSERT(HDR_HAS_RABD(hdr));
2146 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
2147 HDR_GET_PSIZE(hdr));
2152 * Adjust encrypted and authenticated headers to accomodate
2153 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
2154 * allowed to fail decryption due to keys not being loaded
2155 * without being marked as an IO error.
2157 if (HDR_PROTECTED(hdr)) {
2158 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
2159 zb, !!(flags & ARC_FILL_NOAUTH));
2160 if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
2162 } else if (error != 0) {
2163 if (hash_lock != NULL)
2164 mutex_enter(hash_lock);
2165 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2166 if (hash_lock != NULL)
2167 mutex_exit(hash_lock);
2173 * There is a special case here for dnode blocks which are
2174 * decrypting their bonus buffers. These blocks may request to
2175 * be decrypted in-place. This is necessary because there may
2176 * be many dnodes pointing into this buffer and there is
2177 * currently no method to synchronize replacing the backing
2178 * b_data buffer and updating all of the pointers. Here we use
2179 * the hash lock to ensure there are no races. If the need
2180 * arises for other types to be decrypted in-place, they must
2181 * add handling here as well.
2183 if ((flags & ARC_FILL_IN_PLACE) != 0) {
2184 ASSERT(!hdr_compressed);
2185 ASSERT(!compressed);
2188 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
2189 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2191 if (hash_lock != NULL)
2192 mutex_enter(hash_lock);
2193 arc_buf_untransform_in_place(buf, hash_lock);
2194 if (hash_lock != NULL)
2195 mutex_exit(hash_lock);
2197 /* Compute the hdr's checksum if necessary */
2198 arc_cksum_compute(buf);
2204 if (hdr_compressed == compressed) {
2205 if (!arc_buf_is_shared(buf)) {
2206 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2210 ASSERT(hdr_compressed);
2211 ASSERT(!compressed);
2212 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2215 * If the buf is sharing its data with the hdr, unlink it and
2216 * allocate a new data buffer for the buf.
2218 if (arc_buf_is_shared(buf)) {
2219 ASSERT(ARC_BUF_COMPRESSED(buf));
2221 /* We need to give the buf it's own b_data */
2222 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2224 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2225 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2227 /* Previously overhead was 0; just add new overhead */
2228 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2229 } else if (ARC_BUF_COMPRESSED(buf)) {
2230 /* We need to reallocate the buf's b_data */
2231 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2234 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2236 /* We increased the size of b_data; update overhead */
2237 ARCSTAT_INCR(arcstat_overhead_size,
2238 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2242 * Regardless of the buf's previous compression settings, it
2243 * should not be compressed at the end of this function.
2245 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2248 * Try copying the data from another buf which already has a
2249 * decompressed version. If that's not possible, it's time to
2250 * bite the bullet and decompress the data from the hdr.
2252 if (arc_buf_try_copy_decompressed_data(buf)) {
2253 /* Skip byteswapping and checksumming (already done) */
2254 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2257 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2258 hdr->b_l1hdr.b_pabd, buf->b_data,
2259 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2262 * Absent hardware errors or software bugs, this should
2263 * be impossible, but log it anyway so we can debug it.
2267 "hdr %p, compress %d, psize %d, lsize %d",
2268 hdr, arc_hdr_get_compress(hdr),
2269 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2270 if (hash_lock != NULL)
2271 mutex_enter(hash_lock);
2272 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2273 if (hash_lock != NULL)
2274 mutex_exit(hash_lock);
2275 return (SET_ERROR(EIO));
2281 /* Byteswap the buf's data if necessary */
2282 if (bswap != DMU_BSWAP_NUMFUNCS) {
2283 ASSERT(!HDR_SHARED_DATA(hdr));
2284 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2285 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2288 /* Compute the hdr's checksum if necessary */
2289 arc_cksum_compute(buf);
2295 * If this function is being called to decrypt an encrypted buffer or verify an
2296 * authenticated one, the key must be loaded and a mapping must be made
2297 * available in the keystore via spa_keystore_create_mapping() or one of its
2301 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2305 arc_fill_flags_t flags = 0;
2308 flags |= ARC_FILL_IN_PLACE;
2310 ret = arc_buf_fill(buf, spa, zb, flags);
2311 if (ret == ECKSUM) {
2313 * Convert authentication and decryption errors to EIO
2314 * (and generate an ereport) before leaving the ARC.
2316 ret = SET_ERROR(EIO);
2317 spa_log_error(spa, zb);
2318 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2319 spa, NULL, zb, NULL, 0, 0);
2326 * Increment the amount of evictable space in the arc_state_t's refcount.
2327 * We account for the space used by the hdr and the arc buf individually
2328 * so that we can add and remove them from the refcount individually.
2331 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2333 arc_buf_contents_t type = arc_buf_type(hdr);
2335 ASSERT(HDR_HAS_L1HDR(hdr));
2337 if (GHOST_STATE(state)) {
2338 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2339 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2340 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2341 ASSERT(!HDR_HAS_RABD(hdr));
2342 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2343 HDR_GET_LSIZE(hdr), hdr);
2347 ASSERT(!GHOST_STATE(state));
2348 if (hdr->b_l1hdr.b_pabd != NULL) {
2349 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2350 arc_hdr_size(hdr), hdr);
2352 if (HDR_HAS_RABD(hdr)) {
2353 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2354 HDR_GET_PSIZE(hdr), hdr);
2357 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2358 buf = buf->b_next) {
2359 if (arc_buf_is_shared(buf))
2361 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2362 arc_buf_size(buf), buf);
2367 * Decrement the amount of evictable space in the arc_state_t's refcount.
2368 * We account for the space used by the hdr and the arc buf individually
2369 * so that we can add and remove them from the refcount individually.
2372 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2374 arc_buf_contents_t type = arc_buf_type(hdr);
2376 ASSERT(HDR_HAS_L1HDR(hdr));
2378 if (GHOST_STATE(state)) {
2379 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2380 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2381 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2382 ASSERT(!HDR_HAS_RABD(hdr));
2383 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2384 HDR_GET_LSIZE(hdr), hdr);
2388 ASSERT(!GHOST_STATE(state));
2389 if (hdr->b_l1hdr.b_pabd != NULL) {
2390 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2391 arc_hdr_size(hdr), hdr);
2393 if (HDR_HAS_RABD(hdr)) {
2394 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2395 HDR_GET_PSIZE(hdr), hdr);
2398 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2399 buf = buf->b_next) {
2400 if (arc_buf_is_shared(buf))
2402 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2403 arc_buf_size(buf), buf);
2408 * Add a reference to this hdr indicating that someone is actively
2409 * referencing that memory. When the refcount transitions from 0 to 1,
2410 * we remove it from the respective arc_state_t list to indicate that
2411 * it is not evictable.
2414 add_reference(arc_buf_hdr_t *hdr, void *tag)
2418 ASSERT(HDR_HAS_L1HDR(hdr));
2419 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2420 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2421 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2422 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2425 state = hdr->b_l1hdr.b_state;
2427 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2428 (state != arc_anon)) {
2429 /* We don't use the L2-only state list. */
2430 if (state != arc_l2c_only) {
2431 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2433 arc_evictable_space_decrement(hdr, state);
2435 /* remove the prefetch flag if we get a reference */
2436 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2441 * Remove a reference from this hdr. When the reference transitions from
2442 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2443 * list making it eligible for eviction.
2446 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2449 arc_state_t *state = hdr->b_l1hdr.b_state;
2451 ASSERT(HDR_HAS_L1HDR(hdr));
2452 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2453 ASSERT(!GHOST_STATE(state));
2456 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2457 * check to prevent usage of the arc_l2c_only list.
2459 if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2460 (state != arc_anon)) {
2461 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2462 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2463 arc_evictable_space_increment(hdr, state);
2469 * Returns detailed information about a specific arc buffer. When the
2470 * state_index argument is set the function will calculate the arc header
2471 * list position for its arc state. Since this requires a linear traversal
2472 * callers are strongly encourage not to do this. However, it can be helpful
2473 * for targeted analysis so the functionality is provided.
2476 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2478 arc_buf_hdr_t *hdr = ab->b_hdr;
2479 l1arc_buf_hdr_t *l1hdr = NULL;
2480 l2arc_buf_hdr_t *l2hdr = NULL;
2481 arc_state_t *state = NULL;
2483 memset(abi, 0, sizeof (arc_buf_info_t));
2488 abi->abi_flags = hdr->b_flags;
2490 if (HDR_HAS_L1HDR(hdr)) {
2491 l1hdr = &hdr->b_l1hdr;
2492 state = l1hdr->b_state;
2494 if (HDR_HAS_L2HDR(hdr))
2495 l2hdr = &hdr->b_l2hdr;
2498 abi->abi_bufcnt = l1hdr->b_bufcnt;
2499 abi->abi_access = l1hdr->b_arc_access;
2500 abi->abi_mru_hits = l1hdr->b_mru_hits;
2501 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2502 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2503 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2504 abi->abi_holds = zfs_refcount_count(&l1hdr->b_refcnt);
2508 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2509 abi->abi_l2arc_hits = l2hdr->b_hits;
2512 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2513 abi->abi_state_contents = arc_buf_type(hdr);
2514 abi->abi_size = arc_hdr_size(hdr);
2518 * Move the supplied buffer to the indicated state. The hash lock
2519 * for the buffer must be held by the caller.
2522 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2523 kmutex_t *hash_lock)
2525 arc_state_t *old_state;
2528 boolean_t update_old, update_new;
2529 arc_buf_contents_t buftype = arc_buf_type(hdr);
2532 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2533 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2534 * L1 hdr doesn't always exist when we change state to arc_anon before
2535 * destroying a header, in which case reallocating to add the L1 hdr is
2538 if (HDR_HAS_L1HDR(hdr)) {
2539 old_state = hdr->b_l1hdr.b_state;
2540 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2541 bufcnt = hdr->b_l1hdr.b_bufcnt;
2542 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
2545 old_state = arc_l2c_only;
2548 update_old = B_FALSE;
2550 update_new = update_old;
2552 ASSERT(MUTEX_HELD(hash_lock));
2553 ASSERT3P(new_state, !=, old_state);
2554 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2555 ASSERT(old_state != arc_anon || bufcnt <= 1);
2558 * If this buffer is evictable, transfer it from the
2559 * old state list to the new state list.
2562 if (old_state != arc_anon && old_state != arc_l2c_only) {
2563 ASSERT(HDR_HAS_L1HDR(hdr));
2564 multilist_remove(old_state->arcs_list[buftype], hdr);
2566 if (GHOST_STATE(old_state)) {
2568 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2569 update_old = B_TRUE;
2571 arc_evictable_space_decrement(hdr, old_state);
2573 if (new_state != arc_anon && new_state != arc_l2c_only) {
2575 * An L1 header always exists here, since if we're
2576 * moving to some L1-cached state (i.e. not l2c_only or
2577 * anonymous), we realloc the header to add an L1hdr
2580 ASSERT(HDR_HAS_L1HDR(hdr));
2581 multilist_insert(new_state->arcs_list[buftype], hdr);
2583 if (GHOST_STATE(new_state)) {
2585 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2586 update_new = B_TRUE;
2588 arc_evictable_space_increment(hdr, new_state);
2592 ASSERT(!HDR_EMPTY(hdr));
2593 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2594 buf_hash_remove(hdr);
2596 /* adjust state sizes (ignore arc_l2c_only) */
2598 if (update_new && new_state != arc_l2c_only) {
2599 ASSERT(HDR_HAS_L1HDR(hdr));
2600 if (GHOST_STATE(new_state)) {
2604 * When moving a header to a ghost state, we first
2605 * remove all arc buffers. Thus, we'll have a
2606 * bufcnt of zero, and no arc buffer to use for
2607 * the reference. As a result, we use the arc
2608 * header pointer for the reference.
2610 (void) zfs_refcount_add_many(&new_state->arcs_size,
2611 HDR_GET_LSIZE(hdr), hdr);
2612 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2613 ASSERT(!HDR_HAS_RABD(hdr));
2615 uint32_t buffers = 0;
2618 * Each individual buffer holds a unique reference,
2619 * thus we must remove each of these references one
2622 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2623 buf = buf->b_next) {
2624 ASSERT3U(bufcnt, !=, 0);
2628 * When the arc_buf_t is sharing the data
2629 * block with the hdr, the owner of the
2630 * reference belongs to the hdr. Only
2631 * add to the refcount if the arc_buf_t is
2634 if (arc_buf_is_shared(buf))
2637 (void) zfs_refcount_add_many(
2638 &new_state->arcs_size,
2639 arc_buf_size(buf), buf);
2641 ASSERT3U(bufcnt, ==, buffers);
2643 if (hdr->b_l1hdr.b_pabd != NULL) {
2644 (void) zfs_refcount_add_many(
2645 &new_state->arcs_size,
2646 arc_hdr_size(hdr), hdr);
2649 if (HDR_HAS_RABD(hdr)) {
2650 (void) zfs_refcount_add_many(
2651 &new_state->arcs_size,
2652 HDR_GET_PSIZE(hdr), hdr);
2657 if (update_old && old_state != arc_l2c_only) {
2658 ASSERT(HDR_HAS_L1HDR(hdr));
2659 if (GHOST_STATE(old_state)) {
2661 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2662 ASSERT(!HDR_HAS_RABD(hdr));
2665 * When moving a header off of a ghost state,
2666 * the header will not contain any arc buffers.
2667 * We use the arc header pointer for the reference
2668 * which is exactly what we did when we put the
2669 * header on the ghost state.
2672 (void) zfs_refcount_remove_many(&old_state->arcs_size,
2673 HDR_GET_LSIZE(hdr), hdr);
2675 uint32_t buffers = 0;
2678 * Each individual buffer holds a unique reference,
2679 * thus we must remove each of these references one
2682 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2683 buf = buf->b_next) {
2684 ASSERT3U(bufcnt, !=, 0);
2688 * When the arc_buf_t is sharing the data
2689 * block with the hdr, the owner of the
2690 * reference belongs to the hdr. Only
2691 * add to the refcount if the arc_buf_t is
2694 if (arc_buf_is_shared(buf))
2697 (void) zfs_refcount_remove_many(
2698 &old_state->arcs_size, arc_buf_size(buf),
2701 ASSERT3U(bufcnt, ==, buffers);
2702 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2705 if (hdr->b_l1hdr.b_pabd != NULL) {
2706 (void) zfs_refcount_remove_many(
2707 &old_state->arcs_size, arc_hdr_size(hdr),
2711 if (HDR_HAS_RABD(hdr)) {
2712 (void) zfs_refcount_remove_many(
2713 &old_state->arcs_size, HDR_GET_PSIZE(hdr),
2719 if (HDR_HAS_L1HDR(hdr))
2720 hdr->b_l1hdr.b_state = new_state;
2723 * L2 headers should never be on the L2 state list since they don't
2724 * have L1 headers allocated.
2726 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2727 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2731 arc_space_consume(uint64_t space, arc_space_type_t type)
2733 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2738 case ARC_SPACE_DATA:
2739 aggsum_add(&astat_data_size, space);
2741 case ARC_SPACE_META:
2742 aggsum_add(&astat_metadata_size, space);
2744 case ARC_SPACE_BONUS:
2745 aggsum_add(&astat_bonus_size, space);
2747 case ARC_SPACE_DNODE:
2748 aggsum_add(&astat_dnode_size, space);
2750 case ARC_SPACE_DBUF:
2751 aggsum_add(&astat_dbuf_size, space);
2753 case ARC_SPACE_HDRS:
2754 aggsum_add(&astat_hdr_size, space);
2756 case ARC_SPACE_L2HDRS:
2757 aggsum_add(&astat_l2_hdr_size, space);
2761 if (type != ARC_SPACE_DATA)
2762 aggsum_add(&arc_meta_used, space);
2764 aggsum_add(&arc_size, space);
2768 arc_space_return(uint64_t space, arc_space_type_t type)
2770 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2775 case ARC_SPACE_DATA:
2776 aggsum_add(&astat_data_size, -space);
2778 case ARC_SPACE_META:
2779 aggsum_add(&astat_metadata_size, -space);
2781 case ARC_SPACE_BONUS:
2782 aggsum_add(&astat_bonus_size, -space);
2784 case ARC_SPACE_DNODE:
2785 aggsum_add(&astat_dnode_size, -space);
2787 case ARC_SPACE_DBUF:
2788 aggsum_add(&astat_dbuf_size, -space);
2790 case ARC_SPACE_HDRS:
2791 aggsum_add(&astat_hdr_size, -space);
2793 case ARC_SPACE_L2HDRS:
2794 aggsum_add(&astat_l2_hdr_size, -space);
2798 if (type != ARC_SPACE_DATA) {
2799 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2801 * We use the upper bound here rather than the precise value
2802 * because the arc_meta_max value doesn't need to be
2803 * precise. It's only consumed by humans via arcstats.
2805 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2806 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2807 aggsum_add(&arc_meta_used, -space);
2810 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2811 aggsum_add(&arc_size, -space);
2815 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2816 * with the hdr's b_pabd.
2819 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2822 * The criteria for sharing a hdr's data are:
2823 * 1. the buffer is not encrypted
2824 * 2. the hdr's compression matches the buf's compression
2825 * 3. the hdr doesn't need to be byteswapped
2826 * 4. the hdr isn't already being shared
2827 * 5. the buf is either compressed or it is the last buf in the hdr list
2829 * Criterion #5 maintains the invariant that shared uncompressed
2830 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2831 * might ask, "if a compressed buf is allocated first, won't that be the
2832 * last thing in the list?", but in that case it's impossible to create
2833 * a shared uncompressed buf anyway (because the hdr must be compressed
2834 * to have the compressed buf). You might also think that #3 is
2835 * sufficient to make this guarantee, however it's possible
2836 * (specifically in the rare L2ARC write race mentioned in
2837 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2838 * is sharable, but wasn't at the time of its allocation. Rather than
2839 * allow a new shared uncompressed buf to be created and then shuffle
2840 * the list around to make it the last element, this simply disallows
2841 * sharing if the new buf isn't the first to be added.
2843 ASSERT3P(buf->b_hdr, ==, hdr);
2844 boolean_t hdr_compressed =
2845 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF;
2846 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2847 return (!ARC_BUF_ENCRYPTED(buf) &&
2848 buf_compressed == hdr_compressed &&
2849 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2850 !HDR_SHARED_DATA(hdr) &&
2851 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2855 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2856 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2857 * copy was made successfully, or an error code otherwise.
2860 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2861 void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth,
2862 boolean_t fill, arc_buf_t **ret)
2865 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2867 ASSERT(HDR_HAS_L1HDR(hdr));
2868 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2869 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2870 hdr->b_type == ARC_BUFC_METADATA);
2871 ASSERT3P(ret, !=, NULL);
2872 ASSERT3P(*ret, ==, NULL);
2873 IMPLY(encrypted, compressed);
2875 hdr->b_l1hdr.b_mru_hits = 0;
2876 hdr->b_l1hdr.b_mru_ghost_hits = 0;
2877 hdr->b_l1hdr.b_mfu_hits = 0;
2878 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
2879 hdr->b_l1hdr.b_l2_hits = 0;
2881 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2884 buf->b_next = hdr->b_l1hdr.b_buf;
2887 add_reference(hdr, tag);
2890 * We're about to change the hdr's b_flags. We must either
2891 * hold the hash_lock or be undiscoverable.
2893 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2896 * Only honor requests for compressed bufs if the hdr is actually
2897 * compressed. This must be overriden if the buffer is encrypted since
2898 * encrypted buffers cannot be decompressed.
2901 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2902 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2903 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2904 } else if (compressed &&
2905 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2906 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2907 flags |= ARC_FILL_COMPRESSED;
2912 flags |= ARC_FILL_NOAUTH;
2916 * If the hdr's data can be shared then we share the data buffer and
2917 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2918 * allocate a new buffer to store the buf's data.
2920 * There are two additional restrictions here because we're sharing
2921 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2922 * actively involved in an L2ARC write, because if this buf is used by
2923 * an arc_write() then the hdr's data buffer will be released when the
2924 * write completes, even though the L2ARC write might still be using it.
2925 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2926 * need to be ABD-aware.
2928 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2929 hdr->b_l1hdr.b_pabd != NULL && abd_is_linear(hdr->b_l1hdr.b_pabd);
2931 /* Set up b_data and sharing */
2933 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2934 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2935 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2938 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2939 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2941 VERIFY3P(buf->b_data, !=, NULL);
2943 hdr->b_l1hdr.b_buf = buf;
2944 hdr->b_l1hdr.b_bufcnt += 1;
2946 hdr->b_crypt_hdr.b_ebufcnt += 1;
2949 * If the user wants the data from the hdr, we need to either copy or
2950 * decompress the data.
2953 ASSERT3P(zb, !=, NULL);
2954 return (arc_buf_fill(buf, spa, zb, flags));
2960 static char *arc_onloan_tag = "onloan";
2963 arc_loaned_bytes_update(int64_t delta)
2965 atomic_add_64(&arc_loaned_bytes, delta);
2967 /* assert that it did not wrap around */
2968 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2972 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2973 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2974 * buffers must be returned to the arc before they can be used by the DMU or
2978 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2980 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2981 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2983 arc_loaned_bytes_update(arc_buf_size(buf));
2989 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2990 enum zio_compress compression_type)
2992 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2993 psize, lsize, compression_type);
2995 arc_loaned_bytes_update(arc_buf_size(buf));
3001 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
3002 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
3003 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3004 enum zio_compress compression_type)
3006 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
3007 byteorder, salt, iv, mac, ot, psize, lsize, compression_type);
3009 atomic_add_64(&arc_loaned_bytes, psize);
3015 * Return a loaned arc buffer to the arc.
3018 arc_return_buf(arc_buf_t *buf, void *tag)
3020 arc_buf_hdr_t *hdr = buf->b_hdr;
3022 ASSERT3P(buf->b_data, !=, NULL);
3023 ASSERT(HDR_HAS_L1HDR(hdr));
3024 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
3025 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3027 arc_loaned_bytes_update(-arc_buf_size(buf));
3030 /* Detach an arc_buf from a dbuf (tag) */
3032 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
3034 arc_buf_hdr_t *hdr = buf->b_hdr;
3036 ASSERT3P(buf->b_data, !=, NULL);
3037 ASSERT(HDR_HAS_L1HDR(hdr));
3038 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
3039 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
3041 arc_loaned_bytes_update(arc_buf_size(buf));
3045 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
3047 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
3050 df->l2df_size = size;
3051 df->l2df_type = type;
3052 mutex_enter(&l2arc_free_on_write_mtx);
3053 list_insert_head(l2arc_free_on_write, df);
3054 mutex_exit(&l2arc_free_on_write_mtx);
3058 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3060 arc_state_t *state = hdr->b_l1hdr.b_state;
3061 arc_buf_contents_t type = arc_buf_type(hdr);
3062 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3064 /* protected by hash lock, if in the hash table */
3065 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3066 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3067 ASSERT(state != arc_anon && state != arc_l2c_only);
3069 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
3072 (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr);
3073 if (type == ARC_BUFC_METADATA) {
3074 arc_space_return(size, ARC_SPACE_META);
3076 ASSERT(type == ARC_BUFC_DATA);
3077 arc_space_return(size, ARC_SPACE_DATA);
3081 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
3083 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
3088 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
3089 * data buffer, we transfer the refcount ownership to the hdr and update
3090 * the appropriate kstats.
3093 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3095 ASSERT(arc_can_share(hdr, buf));
3096 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3097 ASSERT(!ARC_BUF_ENCRYPTED(buf));
3098 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3101 * Start sharing the data buffer. We transfer the
3102 * refcount ownership to the hdr since it always owns
3103 * the refcount whenever an arc_buf_t is shared.
3105 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
3106 arc_hdr_size(hdr), buf, hdr);
3107 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
3108 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
3109 HDR_ISTYPE_METADATA(hdr));
3110 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3111 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3114 * Since we've transferred ownership to the hdr we need
3115 * to increment its compressed and uncompressed kstats and
3116 * decrement the overhead size.
3118 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3119 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3120 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3124 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3126 ASSERT(arc_buf_is_shared(buf));
3127 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3128 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3131 * We are no longer sharing this buffer so we need
3132 * to transfer its ownership to the rightful owner.
3134 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
3135 arc_hdr_size(hdr), hdr, buf);
3136 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3137 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3138 abd_put(hdr->b_l1hdr.b_pabd);
3139 hdr->b_l1hdr.b_pabd = NULL;
3140 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3143 * Since the buffer is no longer shared between
3144 * the arc buf and the hdr, count it as overhead.
3146 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3147 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3148 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3152 * Remove an arc_buf_t from the hdr's buf list and return the last
3153 * arc_buf_t on the list. If no buffers remain on the list then return
3157 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3159 ASSERT(HDR_HAS_L1HDR(hdr));
3160 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3162 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3163 arc_buf_t *lastbuf = NULL;
3166 * Remove the buf from the hdr list and locate the last
3167 * remaining buffer on the list.
3169 while (*bufp != NULL) {
3171 *bufp = buf->b_next;
3174 * If we've removed a buffer in the middle of
3175 * the list then update the lastbuf and update
3178 if (*bufp != NULL) {
3180 bufp = &(*bufp)->b_next;
3184 ASSERT3P(lastbuf, !=, buf);
3185 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3186 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3187 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3193 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3197 arc_buf_destroy_impl(arc_buf_t *buf)
3199 arc_buf_hdr_t *hdr = buf->b_hdr;
3202 * Free up the data associated with the buf but only if we're not
3203 * sharing this with the hdr. If we are sharing it with the hdr, the
3204 * hdr is responsible for doing the free.
3206 if (buf->b_data != NULL) {
3208 * We're about to change the hdr's b_flags. We must either
3209 * hold the hash_lock or be undiscoverable.
3211 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3213 arc_cksum_verify(buf);
3214 arc_buf_unwatch(buf);
3216 if (arc_buf_is_shared(buf)) {
3217 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3219 uint64_t size = arc_buf_size(buf);
3220 arc_free_data_buf(hdr, buf->b_data, size, buf);
3221 ARCSTAT_INCR(arcstat_overhead_size, -size);
3225 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3226 hdr->b_l1hdr.b_bufcnt -= 1;
3228 if (ARC_BUF_ENCRYPTED(buf)) {
3229 hdr->b_crypt_hdr.b_ebufcnt -= 1;
3232 * If we have no more encrypted buffers and we've
3233 * already gotten a copy of the decrypted data we can
3234 * free b_rabd to save some space.
3236 if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
3237 HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
3238 !HDR_IO_IN_PROGRESS(hdr)) {
3239 arc_hdr_free_abd(hdr, B_TRUE);
3244 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3246 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3248 * If the current arc_buf_t is sharing its data buffer with the
3249 * hdr, then reassign the hdr's b_pabd to share it with the new
3250 * buffer at the end of the list. The shared buffer is always
3251 * the last one on the hdr's buffer list.
3253 * There is an equivalent case for compressed bufs, but since
3254 * they aren't guaranteed to be the last buf in the list and
3255 * that is an exceedingly rare case, we just allow that space be
3256 * wasted temporarily. We must also be careful not to share
3257 * encrypted buffers, since they cannot be shared.
3259 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
3260 /* Only one buf can be shared at once */
3261 VERIFY(!arc_buf_is_shared(lastbuf));
3262 /* hdr is uncompressed so can't have compressed buf */
3263 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3265 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3266 arc_hdr_free_abd(hdr, B_FALSE);
3269 * We must setup a new shared block between the
3270 * last buffer and the hdr. The data would have
3271 * been allocated by the arc buf so we need to transfer
3272 * ownership to the hdr since it's now being shared.
3274 arc_share_buf(hdr, lastbuf);
3276 } else if (HDR_SHARED_DATA(hdr)) {
3278 * Uncompressed shared buffers are always at the end
3279 * of the list. Compressed buffers don't have the
3280 * same requirements. This makes it hard to
3281 * simply assert that the lastbuf is shared so
3282 * we rely on the hdr's compression flags to determine
3283 * if we have a compressed, shared buffer.
3285 ASSERT3P(lastbuf, !=, NULL);
3286 ASSERT(arc_buf_is_shared(lastbuf) ||
3287 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3291 * Free the checksum if we're removing the last uncompressed buf from
3294 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3295 arc_cksum_free(hdr);
3298 /* clean up the buf */
3300 kmem_cache_free(buf_cache, buf);
3304 arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, boolean_t alloc_rdata)
3308 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3309 ASSERT(HDR_HAS_L1HDR(hdr));
3310 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3311 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3314 size = HDR_GET_PSIZE(hdr);
3315 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3316 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr);
3317 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3318 ARCSTAT_INCR(arcstat_raw_size, size);
3320 size = arc_hdr_size(hdr);
3321 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3322 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr);
3323 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3326 ARCSTAT_INCR(arcstat_compressed_size, size);
3327 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3331 arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3333 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3335 ASSERT(HDR_HAS_L1HDR(hdr));
3336 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3337 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3340 * If the hdr is currently being written to the l2arc then
3341 * we defer freeing the data by adding it to the l2arc_free_on_write
3342 * list. The l2arc will free the data once it's finished
3343 * writing it to the l2arc device.
3345 if (HDR_L2_WRITING(hdr)) {
3346 arc_hdr_free_on_write(hdr, free_rdata);
3347 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3348 } else if (free_rdata) {
3349 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3351 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr);
3355 hdr->b_crypt_hdr.b_rabd = NULL;
3356 ARCSTAT_INCR(arcstat_raw_size, -size);
3358 hdr->b_l1hdr.b_pabd = NULL;
3361 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3362 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3364 ARCSTAT_INCR(arcstat_compressed_size, -size);
3365 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3368 static arc_buf_hdr_t *
3369 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3370 boolean_t protected, enum zio_compress compression_type,
3371 arc_buf_contents_t type, boolean_t alloc_rdata)
3375 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3377 hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
3379 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3382 ASSERT(HDR_EMPTY(hdr));
3383 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3384 HDR_SET_PSIZE(hdr, psize);
3385 HDR_SET_LSIZE(hdr, lsize);
3389 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3390 arc_hdr_set_compress(hdr, compression_type);
3392 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3394 hdr->b_l1hdr.b_state = arc_anon;
3395 hdr->b_l1hdr.b_arc_access = 0;
3396 hdr->b_l1hdr.b_bufcnt = 0;
3397 hdr->b_l1hdr.b_buf = NULL;
3400 * Allocate the hdr's buffer. This will contain either
3401 * the compressed or uncompressed data depending on the block
3402 * it references and compressed arc enablement.
3404 arc_hdr_alloc_abd(hdr, alloc_rdata);
3405 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3411 * Transition between the two allocation states for the arc_buf_hdr struct.
3412 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3413 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3414 * version is used when a cache buffer is only in the L2ARC in order to reduce
3417 static arc_buf_hdr_t *
3418 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3420 ASSERT(HDR_HAS_L2HDR(hdr));
3422 arc_buf_hdr_t *nhdr;
3423 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3425 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3426 (old == hdr_l2only_cache && new == hdr_full_cache));
3429 * if the caller wanted a new full header and the header is to be
3430 * encrypted we will actually allocate the header from the full crypt
3431 * cache instead. The same applies to freeing from the old cache.
3433 if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
3434 new = hdr_full_crypt_cache;
3435 if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
3436 old = hdr_full_crypt_cache;
3438 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3440 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3441 buf_hash_remove(hdr);
3443 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3445 if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
3446 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3448 * arc_access and arc_change_state need to be aware that a
3449 * header has just come out of L2ARC, so we set its state to
3450 * l2c_only even though it's about to change.
3452 nhdr->b_l1hdr.b_state = arc_l2c_only;
3454 /* Verify previous threads set to NULL before freeing */
3455 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3456 ASSERT(!HDR_HAS_RABD(hdr));
3458 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3459 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3460 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3463 * If we've reached here, We must have been called from
3464 * arc_evict_hdr(), as such we should have already been
3465 * removed from any ghost list we were previously on
3466 * (which protects us from racing with arc_evict_state),
3467 * thus no locking is needed during this check.
3469 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3472 * A buffer must not be moved into the arc_l2c_only
3473 * state if it's not finished being written out to the
3474 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3475 * might try to be accessed, even though it was removed.
3477 VERIFY(!HDR_L2_WRITING(hdr));
3478 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3479 ASSERT(!HDR_HAS_RABD(hdr));
3481 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3484 * The header has been reallocated so we need to re-insert it into any
3487 (void) buf_hash_insert(nhdr, NULL);
3489 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3491 mutex_enter(&dev->l2ad_mtx);
3494 * We must place the realloc'ed header back into the list at
3495 * the same spot. Otherwise, if it's placed earlier in the list,
3496 * l2arc_write_buffers() could find it during the function's
3497 * write phase, and try to write it out to the l2arc.
3499 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3500 list_remove(&dev->l2ad_buflist, hdr);
3502 mutex_exit(&dev->l2ad_mtx);
3505 * Since we're using the pointer address as the tag when
3506 * incrementing and decrementing the l2ad_alloc refcount, we
3507 * must remove the old pointer (that we're about to destroy) and
3508 * add the new pointer to the refcount. Otherwise we'd remove
3509 * the wrong pointer address when calling arc_hdr_destroy() later.
3512 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
3513 arc_hdr_size(hdr), hdr);
3514 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
3515 arc_hdr_size(nhdr), nhdr);
3517 buf_discard_identity(hdr);
3518 kmem_cache_free(old, hdr);
3524 * This function allows an L1 header to be reallocated as a crypt
3525 * header and vice versa. If we are going to a crypt header, the
3526 * new fields will be zeroed out.
3528 static arc_buf_hdr_t *
3529 arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
3531 arc_buf_hdr_t *nhdr;
3533 kmem_cache_t *ncache, *ocache;
3534 unsigned nsize, osize;
3537 * This function requires that hdr is in the arc_anon state.
3538 * Therefore it won't have any L2ARC data for us to worry
3541 ASSERT(HDR_HAS_L1HDR(hdr));
3542 ASSERT(!HDR_HAS_L2HDR(hdr));
3543 ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
3544 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3545 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3546 ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node));
3547 ASSERT3P(hdr->b_hash_next, ==, NULL);
3550 ncache = hdr_full_crypt_cache;
3551 nsize = sizeof (hdr->b_crypt_hdr);
3552 ocache = hdr_full_cache;
3553 osize = HDR_FULL_SIZE;
3555 ncache = hdr_full_cache;
3556 nsize = HDR_FULL_SIZE;
3557 ocache = hdr_full_crypt_cache;
3558 osize = sizeof (hdr->b_crypt_hdr);
3561 nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);
3564 * Copy all members that aren't locks or condvars to the new header.
3565 * No lists are pointing to us (as we asserted above), so we don't
3566 * need to worry about the list nodes.
3568 nhdr->b_dva = hdr->b_dva;
3569 nhdr->b_birth = hdr->b_birth;
3570 nhdr->b_type = hdr->b_type;
3571 nhdr->b_flags = hdr->b_flags;
3572 nhdr->b_psize = hdr->b_psize;
3573 nhdr->b_lsize = hdr->b_lsize;
3574 nhdr->b_spa = hdr->b_spa;
3575 nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
3576 nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
3577 nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
3578 nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
3579 nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
3580 nhdr->b_l1hdr.b_mru_hits = hdr->b_l1hdr.b_mru_hits;
3581 nhdr->b_l1hdr.b_mru_ghost_hits = hdr->b_l1hdr.b_mru_ghost_hits;
3582 nhdr->b_l1hdr.b_mfu_hits = hdr->b_l1hdr.b_mfu_hits;
3583 nhdr->b_l1hdr.b_mfu_ghost_hits = hdr->b_l1hdr.b_mfu_ghost_hits;
3584 nhdr->b_l1hdr.b_l2_hits = hdr->b_l1hdr.b_l2_hits;
3585 nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
3586 nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
3589 * This zfs_refcount_add() exists only to ensure that the individual
3590 * arc buffers always point to a header that is referenced, avoiding
3591 * a small race condition that could trigger ASSERTs.
3593 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
3594 nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
3595 for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) {
3596 mutex_enter(&buf->b_evict_lock);
3598 mutex_exit(&buf->b_evict_lock);
3601 zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
3602 (void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
3603 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3606 arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
3608 arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
3611 /* unset all members of the original hdr */
3612 bzero(&hdr->b_dva, sizeof (dva_t));
3614 hdr->b_type = ARC_BUFC_INVALID;
3619 hdr->b_l1hdr.b_freeze_cksum = NULL;
3620 hdr->b_l1hdr.b_buf = NULL;
3621 hdr->b_l1hdr.b_bufcnt = 0;
3622 hdr->b_l1hdr.b_byteswap = 0;
3623 hdr->b_l1hdr.b_state = NULL;
3624 hdr->b_l1hdr.b_arc_access = 0;
3625 hdr->b_l1hdr.b_mru_hits = 0;
3626 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3627 hdr->b_l1hdr.b_mfu_hits = 0;
3628 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3629 hdr->b_l1hdr.b_l2_hits = 0;
3630 hdr->b_l1hdr.b_acb = NULL;
3631 hdr->b_l1hdr.b_pabd = NULL;
3633 if (ocache == hdr_full_crypt_cache) {
3634 ASSERT(!HDR_HAS_RABD(hdr));
3635 hdr->b_crypt_hdr.b_ot = DMU_OT_NONE;
3636 hdr->b_crypt_hdr.b_ebufcnt = 0;
3637 hdr->b_crypt_hdr.b_dsobj = 0;
3638 bzero(hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3639 bzero(hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3640 bzero(hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3643 buf_discard_identity(hdr);
3644 kmem_cache_free(ocache, hdr);
3650 * This function is used by the send / receive code to convert a newly
3651 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3652 * is also used to allow the root objset block to be uupdated without altering
3653 * its embedded MACs. Both block types will always be uncompressed so we do not
3654 * have to worry about compression type or psize.
3657 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3658 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3661 arc_buf_hdr_t *hdr = buf->b_hdr;
3663 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3664 ASSERT(HDR_HAS_L1HDR(hdr));
3665 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3667 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3668 if (!HDR_PROTECTED(hdr))
3669 hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
3670 hdr->b_crypt_hdr.b_dsobj = dsobj;
3671 hdr->b_crypt_hdr.b_ot = ot;
3672 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3673 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3674 if (!arc_hdr_has_uncompressed_buf(hdr))
3675 arc_cksum_free(hdr);
3678 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3680 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3682 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3686 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3687 * The buf is returned thawed since we expect the consumer to modify it.
3690 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3692 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3693 B_FALSE, ZIO_COMPRESS_OFF, type, B_FALSE);
3694 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3696 arc_buf_t *buf = NULL;
3697 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3698 B_FALSE, B_FALSE, &buf));
3705 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3706 * for bufs containing metadata.
3709 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3710 enum zio_compress compression_type)
3712 ASSERT3U(lsize, >, 0);
3713 ASSERT3U(lsize, >=, psize);
3714 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3715 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3717 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3718 B_FALSE, compression_type, ARC_BUFC_DATA, B_FALSE);
3719 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3721 arc_buf_t *buf = NULL;
3722 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3723 B_TRUE, B_FALSE, B_FALSE, &buf));
3725 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3727 if (!arc_buf_is_shared(buf)) {
3729 * To ensure that the hdr has the correct data in it if we call
3730 * arc_untransform() on this buf before it's been written to
3731 * disk, it's easiest if we just set up sharing between the
3734 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3735 arc_hdr_free_abd(hdr, B_FALSE);
3736 arc_share_buf(hdr, buf);
3743 arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder,
3744 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
3745 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3746 enum zio_compress compression_type)
3750 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3751 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3753 ASSERT3U(lsize, >, 0);
3754 ASSERT3U(lsize, >=, psize);
3755 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3756 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3758 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3759 compression_type, type, B_TRUE);
3760 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3762 hdr->b_crypt_hdr.b_dsobj = dsobj;
3763 hdr->b_crypt_hdr.b_ot = ot;
3764 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3765 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3766 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3767 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3768 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3771 * This buffer will be considered encrypted even if the ot is not an
3772 * encrypted type. It will become authenticated instead in
3773 * arc_write_ready().
3776 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3777 B_FALSE, B_FALSE, &buf));
3779 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3785 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3787 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3788 l2arc_dev_t *dev = l2hdr->b_dev;
3789 uint64_t psize = HDR_GET_PSIZE(hdr);
3790 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3792 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3793 ASSERT(HDR_HAS_L2HDR(hdr));
3795 list_remove(&dev->l2ad_buflist, hdr);
3797 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3798 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3800 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3802 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3804 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3808 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3810 if (HDR_HAS_L1HDR(hdr)) {
3811 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3812 hdr->b_l1hdr.b_bufcnt > 0);
3813 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3814 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3816 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3817 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3819 if (!HDR_EMPTY(hdr))
3820 buf_discard_identity(hdr);
3822 if (HDR_HAS_L2HDR(hdr)) {
3823 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3824 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3827 mutex_enter(&dev->l2ad_mtx);
3830 * Even though we checked this conditional above, we
3831 * need to check this again now that we have the
3832 * l2ad_mtx. This is because we could be racing with
3833 * another thread calling l2arc_evict() which might have
3834 * destroyed this header's L2 portion as we were waiting
3835 * to acquire the l2ad_mtx. If that happens, we don't
3836 * want to re-destroy the header's L2 portion.
3838 if (HDR_HAS_L2HDR(hdr))
3839 arc_hdr_l2hdr_destroy(hdr);
3842 mutex_exit(&dev->l2ad_mtx);
3845 if (HDR_HAS_L1HDR(hdr)) {
3846 arc_cksum_free(hdr);
3848 while (hdr->b_l1hdr.b_buf != NULL)
3849 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3851 if (hdr->b_l1hdr.b_pabd != NULL) {
3852 arc_hdr_free_abd(hdr, B_FALSE);
3855 if (HDR_HAS_RABD(hdr))
3856 arc_hdr_free_abd(hdr, B_TRUE);
3859 ASSERT3P(hdr->b_hash_next, ==, NULL);
3860 if (HDR_HAS_L1HDR(hdr)) {
3861 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3862 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3864 if (!HDR_PROTECTED(hdr)) {
3865 kmem_cache_free(hdr_full_cache, hdr);
3867 kmem_cache_free(hdr_full_crypt_cache, hdr);
3870 kmem_cache_free(hdr_l2only_cache, hdr);
3875 arc_buf_destroy(arc_buf_t *buf, void* tag)
3877 arc_buf_hdr_t *hdr = buf->b_hdr;
3878 kmutex_t *hash_lock = HDR_LOCK(hdr);
3880 if (hdr->b_l1hdr.b_state == arc_anon) {
3881 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3882 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3883 VERIFY0(remove_reference(hdr, NULL, tag));
3884 arc_hdr_destroy(hdr);
3888 mutex_enter(hash_lock);
3889 ASSERT3P(hdr, ==, buf->b_hdr);
3890 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3891 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3892 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3893 ASSERT3P(buf->b_data, !=, NULL);
3895 (void) remove_reference(hdr, hash_lock, tag);
3896 arc_buf_destroy_impl(buf);
3897 mutex_exit(hash_lock);
3901 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3902 * state of the header is dependent on its state prior to entering this
3903 * function. The following transitions are possible:
3905 * - arc_mru -> arc_mru_ghost
3906 * - arc_mfu -> arc_mfu_ghost
3907 * - arc_mru_ghost -> arc_l2c_only
3908 * - arc_mru_ghost -> deleted
3909 * - arc_mfu_ghost -> arc_l2c_only
3910 * - arc_mfu_ghost -> deleted
3913 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3915 arc_state_t *evicted_state, *state;
3916 int64_t bytes_evicted = 0;
3917 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3918 arc_min_prescient_prefetch_ms : arc_min_prefetch_ms;
3920 ASSERT(MUTEX_HELD(hash_lock));
3921 ASSERT(HDR_HAS_L1HDR(hdr));
3923 state = hdr->b_l1hdr.b_state;
3924 if (GHOST_STATE(state)) {
3925 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3926 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3929 * l2arc_write_buffers() relies on a header's L1 portion
3930 * (i.e. its b_pabd field) during it's write phase.
3931 * Thus, we cannot push a header onto the arc_l2c_only
3932 * state (removing its L1 piece) until the header is
3933 * done being written to the l2arc.
3935 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3936 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3937 return (bytes_evicted);
3940 ARCSTAT_BUMP(arcstat_deleted);
3941 bytes_evicted += HDR_GET_LSIZE(hdr);
3943 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3945 if (HDR_HAS_L2HDR(hdr)) {
3946 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3947 ASSERT(!HDR_HAS_RABD(hdr));
3949 * This buffer is cached on the 2nd Level ARC;
3950 * don't destroy the header.
3952 arc_change_state(arc_l2c_only, hdr, hash_lock);
3954 * dropping from L1+L2 cached to L2-only,
3955 * realloc to remove the L1 header.
3957 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3960 arc_change_state(arc_anon, hdr, hash_lock);
3961 arc_hdr_destroy(hdr);
3963 return (bytes_evicted);
3966 ASSERT(state == arc_mru || state == arc_mfu);
3967 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3969 /* prefetch buffers have a minimum lifespan */
3970 if (HDR_IO_IN_PROGRESS(hdr) ||
3971 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3972 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3973 MSEC_TO_TICK(min_lifetime))) {
3974 ARCSTAT_BUMP(arcstat_evict_skip);
3975 return (bytes_evicted);
3978 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3979 while (hdr->b_l1hdr.b_buf) {
3980 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3981 if (!mutex_tryenter(&buf->b_evict_lock)) {
3982 ARCSTAT_BUMP(arcstat_mutex_miss);
3985 if (buf->b_data != NULL)
3986 bytes_evicted += HDR_GET_LSIZE(hdr);
3987 mutex_exit(&buf->b_evict_lock);
3988 arc_buf_destroy_impl(buf);
3991 if (HDR_HAS_L2HDR(hdr)) {
3992 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3994 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3995 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3996 HDR_GET_LSIZE(hdr));
3998 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3999 HDR_GET_LSIZE(hdr));
4003 if (hdr->b_l1hdr.b_bufcnt == 0) {
4004 arc_cksum_free(hdr);
4006 bytes_evicted += arc_hdr_size(hdr);
4009 * If this hdr is being evicted and has a compressed
4010 * buffer then we discard it here before we change states.
4011 * This ensures that the accounting is updated correctly
4012 * in arc_free_data_impl().
4014 if (hdr->b_l1hdr.b_pabd != NULL)
4015 arc_hdr_free_abd(hdr, B_FALSE);
4017 if (HDR_HAS_RABD(hdr))
4018 arc_hdr_free_abd(hdr, B_TRUE);
4020 arc_change_state(evicted_state, hdr, hash_lock);
4021 ASSERT(HDR_IN_HASH_TABLE(hdr));
4022 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
4023 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
4026 return (bytes_evicted);
4030 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
4031 uint64_t spa, int64_t bytes)
4033 multilist_sublist_t *mls;
4034 uint64_t bytes_evicted = 0;
4036 kmutex_t *hash_lock;
4037 int evict_count = 0;
4039 ASSERT3P(marker, !=, NULL);
4040 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
4042 mls = multilist_sublist_lock(ml, idx);
4044 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
4045 hdr = multilist_sublist_prev(mls, marker)) {
4046 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
4047 (evict_count >= zfs_arc_evict_batch_limit))
4051 * To keep our iteration location, move the marker
4052 * forward. Since we're not holding hdr's hash lock, we
4053 * must be very careful and not remove 'hdr' from the
4054 * sublist. Otherwise, other consumers might mistake the
4055 * 'hdr' as not being on a sublist when they call the
4056 * multilist_link_active() function (they all rely on
4057 * the hash lock protecting concurrent insertions and
4058 * removals). multilist_sublist_move_forward() was
4059 * specifically implemented to ensure this is the case
4060 * (only 'marker' will be removed and re-inserted).
4062 multilist_sublist_move_forward(mls, marker);
4065 * The only case where the b_spa field should ever be
4066 * zero, is the marker headers inserted by
4067 * arc_evict_state(). It's possible for multiple threads
4068 * to be calling arc_evict_state() concurrently (e.g.
4069 * dsl_pool_close() and zio_inject_fault()), so we must
4070 * skip any markers we see from these other threads.
4072 if (hdr->b_spa == 0)
4075 /* we're only interested in evicting buffers of a certain spa */
4076 if (spa != 0 && hdr->b_spa != spa) {
4077 ARCSTAT_BUMP(arcstat_evict_skip);
4081 hash_lock = HDR_LOCK(hdr);
4084 * We aren't calling this function from any code path
4085 * that would already be holding a hash lock, so we're
4086 * asserting on this assumption to be defensive in case
4087 * this ever changes. Without this check, it would be
4088 * possible to incorrectly increment arcstat_mutex_miss
4089 * below (e.g. if the code changed such that we called
4090 * this function with a hash lock held).
4092 ASSERT(!MUTEX_HELD(hash_lock));
4094 if (mutex_tryenter(hash_lock)) {
4095 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
4096 mutex_exit(hash_lock);
4098 bytes_evicted += evicted;
4101 * If evicted is zero, arc_evict_hdr() must have
4102 * decided to skip this header, don't increment
4103 * evict_count in this case.
4109 * If arc_size isn't overflowing, signal any
4110 * threads that might happen to be waiting.
4112 * For each header evicted, we wake up a single
4113 * thread. If we used cv_broadcast, we could
4114 * wake up "too many" threads causing arc_size
4115 * to significantly overflow arc_c; since
4116 * arc_get_data_impl() doesn't check for overflow
4117 * when it's woken up (it doesn't because it's
4118 * possible for the ARC to be overflowing while
4119 * full of un-evictable buffers, and the
4120 * function should proceed in this case).
4122 * If threads are left sleeping, due to not
4123 * using cv_broadcast here, they will be woken
4124 * up via cv_broadcast in arc_adjust_cb() just
4125 * before arc_adjust_zthr sleeps.
4127 mutex_enter(&arc_adjust_lock);
4128 if (!arc_is_overflowing())
4129 cv_signal(&arc_adjust_waiters_cv);
4130 mutex_exit(&arc_adjust_lock);
4132 ARCSTAT_BUMP(arcstat_mutex_miss);
4136 multilist_sublist_unlock(mls);
4138 return (bytes_evicted);
4142 * Evict buffers from the given arc state, until we've removed the
4143 * specified number of bytes. Move the removed buffers to the
4144 * appropriate evict state.
4146 * This function makes a "best effort". It skips over any buffers
4147 * it can't get a hash_lock on, and so, may not catch all candidates.
4148 * It may also return without evicting as much space as requested.
4150 * If bytes is specified using the special value ARC_EVICT_ALL, this
4151 * will evict all available (i.e. unlocked and evictable) buffers from
4152 * the given arc state; which is used by arc_flush().
4155 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
4156 arc_buf_contents_t type)
4158 uint64_t total_evicted = 0;
4159 multilist_t *ml = state->arcs_list[type];
4161 arc_buf_hdr_t **markers;
4163 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
4165 num_sublists = multilist_get_num_sublists(ml);
4168 * If we've tried to evict from each sublist, made some
4169 * progress, but still have not hit the target number of bytes
4170 * to evict, we want to keep trying. The markers allow us to
4171 * pick up where we left off for each individual sublist, rather
4172 * than starting from the tail each time.
4174 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
4175 for (int i = 0; i < num_sublists; i++) {
4176 multilist_sublist_t *mls;
4178 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
4181 * A b_spa of 0 is used to indicate that this header is
4182 * a marker. This fact is used in arc_adjust_type() and
4183 * arc_evict_state_impl().
4185 markers[i]->b_spa = 0;
4187 mls = multilist_sublist_lock(ml, i);
4188 multilist_sublist_insert_tail(mls, markers[i]);
4189 multilist_sublist_unlock(mls);
4193 * While we haven't hit our target number of bytes to evict, or
4194 * we're evicting all available buffers.
4196 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
4197 int sublist_idx = multilist_get_random_index(ml);
4198 uint64_t scan_evicted = 0;
4201 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
4202 * Request that 10% of the LRUs be scanned by the superblock
4205 if (type == ARC_BUFC_DATA && aggsum_compare(&astat_dnode_size,
4206 arc_dnode_limit) > 0) {
4207 arc_prune_async((aggsum_upper_bound(&astat_dnode_size) -
4208 arc_dnode_limit) / sizeof (dnode_t) /
4209 zfs_arc_dnode_reduce_percent);
4213 * Start eviction using a randomly selected sublist,
4214 * this is to try and evenly balance eviction across all
4215 * sublists. Always starting at the same sublist
4216 * (e.g. index 0) would cause evictions to favor certain
4217 * sublists over others.
4219 for (int i = 0; i < num_sublists; i++) {
4220 uint64_t bytes_remaining;
4221 uint64_t bytes_evicted;
4223 if (bytes == ARC_EVICT_ALL)
4224 bytes_remaining = ARC_EVICT_ALL;
4225 else if (total_evicted < bytes)
4226 bytes_remaining = bytes - total_evicted;
4230 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4231 markers[sublist_idx], spa, bytes_remaining);
4233 scan_evicted += bytes_evicted;
4234 total_evicted += bytes_evicted;
4236 /* we've reached the end, wrap to the beginning */
4237 if (++sublist_idx >= num_sublists)
4242 * If we didn't evict anything during this scan, we have
4243 * no reason to believe we'll evict more during another
4244 * scan, so break the loop.
4246 if (scan_evicted == 0) {
4247 /* This isn't possible, let's make that obvious */
4248 ASSERT3S(bytes, !=, 0);
4251 * When bytes is ARC_EVICT_ALL, the only way to
4252 * break the loop is when scan_evicted is zero.
4253 * In that case, we actually have evicted enough,
4254 * so we don't want to increment the kstat.
4256 if (bytes != ARC_EVICT_ALL) {
4257 ASSERT3S(total_evicted, <, bytes);
4258 ARCSTAT_BUMP(arcstat_evict_not_enough);
4265 for (int i = 0; i < num_sublists; i++) {
4266 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4267 multilist_sublist_remove(mls, markers[i]);
4268 multilist_sublist_unlock(mls);
4270 kmem_cache_free(hdr_full_cache, markers[i]);
4272 kmem_free(markers, sizeof (*markers) * num_sublists);
4274 return (total_evicted);
4278 * Flush all "evictable" data of the given type from the arc state
4279 * specified. This will not evict any "active" buffers (i.e. referenced).
4281 * When 'retry' is set to B_FALSE, the function will make a single pass
4282 * over the state and evict any buffers that it can. Since it doesn't
4283 * continually retry the eviction, it might end up leaving some buffers
4284 * in the ARC due to lock misses.
4286 * When 'retry' is set to B_TRUE, the function will continually retry the
4287 * eviction until *all* evictable buffers have been removed from the
4288 * state. As a result, if concurrent insertions into the state are
4289 * allowed (e.g. if the ARC isn't shutting down), this function might
4290 * wind up in an infinite loop, continually trying to evict buffers.
4293 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4296 uint64_t evicted = 0;
4298 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4299 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4309 * Helper function for arc_prune_async() it is responsible for safely
4310 * handling the execution of a registered arc_prune_func_t.
4313 arc_prune_task(void *ptr)
4315 arc_prune_t *ap = (arc_prune_t *)ptr;
4316 arc_prune_func_t *func = ap->p_pfunc;
4319 func(ap->p_adjust, ap->p_private);
4321 zfs_refcount_remove(&ap->p_refcnt, func);
4325 * Notify registered consumers they must drop holds on a portion of the ARC
4326 * buffered they reference. This provides a mechanism to ensure the ARC can
4327 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
4328 * is analogous to dnlc_reduce_cache() but more generic.
4330 * This operation is performed asynchronously so it may be safely called
4331 * in the context of the arc_reclaim_thread(). A reference is taken here
4332 * for each registered arc_prune_t and the arc_prune_task() is responsible
4333 * for releasing it once the registered arc_prune_func_t has completed.
4336 arc_prune_async(int64_t adjust)
4340 mutex_enter(&arc_prune_mtx);
4341 for (ap = list_head(&arc_prune_list); ap != NULL;
4342 ap = list_next(&arc_prune_list, ap)) {
4344 if (zfs_refcount_count(&ap->p_refcnt) >= 2)
4347 zfs_refcount_add(&ap->p_refcnt, ap->p_pfunc);
4348 ap->p_adjust = adjust;
4349 if (taskq_dispatch(arc_prune_taskq, arc_prune_task,
4350 ap, TQ_SLEEP) == TASKQID_INVALID) {
4351 zfs_refcount_remove(&ap->p_refcnt, ap->p_pfunc);
4354 ARCSTAT_BUMP(arcstat_prune);
4356 mutex_exit(&arc_prune_mtx);
4360 * Evict the specified number of bytes from the state specified,
4361 * restricting eviction to the spa and type given. This function
4362 * prevents us from trying to evict more from a state's list than
4363 * is "evictable", and to skip evicting altogether when passed a
4364 * negative value for "bytes". In contrast, arc_evict_state() will
4365 * evict everything it can, when passed a negative value for "bytes".
4368 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4369 arc_buf_contents_t type)
4373 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4374 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4376 return (arc_evict_state(state, spa, delta, type));
4383 * The goal of this function is to evict enough meta data buffers from the
4384 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4385 * more complicated than it appears because it is common for data buffers
4386 * to have holds on meta data buffers. In addition, dnode meta data buffers
4387 * will be held by the dnodes in the block preventing them from being freed.
4388 * This means we can't simply traverse the ARC and expect to always find
4389 * enough unheld meta data buffer to release.
4391 * Therefore, this function has been updated to make alternating passes
4392 * over the ARC releasing data buffers and then newly unheld meta data
4393 * buffers. This ensures forward progress is maintained and meta_used
4394 * will decrease. Normally this is sufficient, but if required the ARC
4395 * will call the registered prune callbacks causing dentry and inodes to
4396 * be dropped from the VFS cache. This will make dnode meta data buffers
4397 * available for reclaim.
4400 arc_adjust_meta_balanced(uint64_t meta_used)
4402 int64_t delta, prune = 0, adjustmnt;
4403 uint64_t total_evicted = 0;
4404 arc_buf_contents_t type = ARC_BUFC_DATA;
4405 int restarts = MAX(zfs_arc_meta_adjust_restarts, 0);
4409 * This slightly differs than the way we evict from the mru in
4410 * arc_adjust because we don't have a "target" value (i.e. no
4411 * "meta" arc_p). As a result, I think we can completely
4412 * cannibalize the metadata in the MRU before we evict the
4413 * metadata from the MFU. I think we probably need to implement a
4414 * "metadata arc_p" value to do this properly.
4416 adjustmnt = meta_used - arc_meta_limit;
4418 if (adjustmnt > 0 &&
4419 zfs_refcount_count(&arc_mru->arcs_esize[type]) > 0) {
4420 delta = MIN(zfs_refcount_count(&arc_mru->arcs_esize[type]),
4422 total_evicted += arc_adjust_impl(arc_mru, 0, delta, type);
4427 * We can't afford to recalculate adjustmnt here. If we do,
4428 * new metadata buffers can sneak into the MRU or ANON lists,
4429 * thus penalize the MFU metadata. Although the fudge factor is
4430 * small, it has been empirically shown to be significant for
4431 * certain workloads (e.g. creating many empty directories). As
4432 * such, we use the original calculation for adjustmnt, and
4433 * simply decrement the amount of data evicted from the MRU.
4436 if (adjustmnt > 0 &&
4437 zfs_refcount_count(&arc_mfu->arcs_esize[type]) > 0) {
4438 delta = MIN(zfs_refcount_count(&arc_mfu->arcs_esize[type]),
4440 total_evicted += arc_adjust_impl(arc_mfu, 0, delta, type);
4443 adjustmnt = meta_used - arc_meta_limit;
4445 if (adjustmnt > 0 &&
4446 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) {
4447 delta = MIN(adjustmnt,
4448 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]));
4449 total_evicted += arc_adjust_impl(arc_mru_ghost, 0, delta, type);
4453 if (adjustmnt > 0 &&
4454 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) {
4455 delta = MIN(adjustmnt,
4456 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]));
4457 total_evicted += arc_adjust_impl(arc_mfu_ghost, 0, delta, type);
4461 * If after attempting to make the requested adjustment to the ARC
4462 * the meta limit is still being exceeded then request that the
4463 * higher layers drop some cached objects which have holds on ARC
4464 * meta buffers. Requests to the upper layers will be made with
4465 * increasingly large scan sizes until the ARC is below the limit.
4467 if (meta_used > arc_meta_limit) {
4468 if (type == ARC_BUFC_DATA) {
4469 type = ARC_BUFC_METADATA;
4471 type = ARC_BUFC_DATA;
4473 if (zfs_arc_meta_prune) {
4474 prune += zfs_arc_meta_prune;
4475 arc_prune_async(prune);
4484 return (total_evicted);
4488 * Evict metadata buffers from the cache, such that arc_meta_used is
4489 * capped by the arc_meta_limit tunable.
4492 arc_adjust_meta_only(uint64_t meta_used)
4494 uint64_t total_evicted = 0;
4498 * If we're over the meta limit, we want to evict enough
4499 * metadata to get back under the meta limit. We don't want to
4500 * evict so much that we drop the MRU below arc_p, though. If
4501 * we're over the meta limit more than we're over arc_p, we
4502 * evict some from the MRU here, and some from the MFU below.
4504 target = MIN((int64_t)(meta_used - arc_meta_limit),
4505 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4506 zfs_refcount_count(&arc_mru->arcs_size) - arc_p));
4508 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4511 * Similar to the above, we want to evict enough bytes to get us
4512 * below the meta limit, but not so much as to drop us below the
4513 * space allotted to the MFU (which is defined as arc_c - arc_p).
4515 target = MIN((int64_t)(meta_used - arc_meta_limit),
4516 (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) -
4519 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4521 return (total_evicted);
4525 arc_adjust_meta(uint64_t meta_used)
4527 if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY)
4528 return (arc_adjust_meta_only(meta_used));
4530 return (arc_adjust_meta_balanced(meta_used));
4534 * Return the type of the oldest buffer in the given arc state
4536 * This function will select a random sublist of type ARC_BUFC_DATA and
4537 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4538 * is compared, and the type which contains the "older" buffer will be
4541 static arc_buf_contents_t
4542 arc_adjust_type(arc_state_t *state)
4544 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4545 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4546 int data_idx = multilist_get_random_index(data_ml);
4547 int meta_idx = multilist_get_random_index(meta_ml);
4548 multilist_sublist_t *data_mls;
4549 multilist_sublist_t *meta_mls;
4550 arc_buf_contents_t type;
4551 arc_buf_hdr_t *data_hdr;
4552 arc_buf_hdr_t *meta_hdr;
4555 * We keep the sublist lock until we're finished, to prevent
4556 * the headers from being destroyed via arc_evict_state().
4558 data_mls = multilist_sublist_lock(data_ml, data_idx);
4559 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4562 * These two loops are to ensure we skip any markers that
4563 * might be at the tail of the lists due to arc_evict_state().
4566 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4567 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4568 if (data_hdr->b_spa != 0)
4572 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4573 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4574 if (meta_hdr->b_spa != 0)
4578 if (data_hdr == NULL && meta_hdr == NULL) {
4579 type = ARC_BUFC_DATA;
4580 } else if (data_hdr == NULL) {
4581 ASSERT3P(meta_hdr, !=, NULL);
4582 type = ARC_BUFC_METADATA;
4583 } else if (meta_hdr == NULL) {
4584 ASSERT3P(data_hdr, !=, NULL);
4585 type = ARC_BUFC_DATA;
4587 ASSERT3P(data_hdr, !=, NULL);
4588 ASSERT3P(meta_hdr, !=, NULL);
4590 /* The headers can't be on the sublist without an L1 header */
4591 ASSERT(HDR_HAS_L1HDR(data_hdr));
4592 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4594 if (data_hdr->b_l1hdr.b_arc_access <
4595 meta_hdr->b_l1hdr.b_arc_access) {
4596 type = ARC_BUFC_DATA;
4598 type = ARC_BUFC_METADATA;
4602 multilist_sublist_unlock(meta_mls);
4603 multilist_sublist_unlock(data_mls);
4609 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4614 uint64_t total_evicted = 0;
4617 uint64_t asize = aggsum_value(&arc_size);
4618 uint64_t ameta = aggsum_value(&arc_meta_used);
4621 * If we're over arc_meta_limit, we want to correct that before
4622 * potentially evicting data buffers below.
4624 total_evicted += arc_adjust_meta(ameta);
4629 * If we're over the target cache size, we want to evict enough
4630 * from the list to get back to our target size. We don't want
4631 * to evict too much from the MRU, such that it drops below
4632 * arc_p. So, if we're over our target cache size more than
4633 * the MRU is over arc_p, we'll evict enough to get back to
4634 * arc_p here, and then evict more from the MFU below.
4636 target = MIN((int64_t)(asize - arc_c),
4637 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4638 zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4641 * If we're below arc_meta_min, always prefer to evict data.
4642 * Otherwise, try to satisfy the requested number of bytes to
4643 * evict from the type which contains older buffers; in an
4644 * effort to keep newer buffers in the cache regardless of their
4645 * type. If we cannot satisfy the number of bytes from this
4646 * type, spill over into the next type.
4648 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4649 ameta > arc_meta_min) {
4650 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4651 total_evicted += bytes;
4654 * If we couldn't evict our target number of bytes from
4655 * metadata, we try to get the rest from data.
4660 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4662 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4663 total_evicted += bytes;
4666 * If we couldn't evict our target number of bytes from
4667 * data, we try to get the rest from metadata.
4672 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4676 * Re-sum ARC stats after the first round of evictions.
4678 asize = aggsum_value(&arc_size);
4679 ameta = aggsum_value(&arc_meta_used);
4685 * Now that we've tried to evict enough from the MRU to get its
4686 * size back to arc_p, if we're still above the target cache
4687 * size, we evict the rest from the MFU.
4689 target = asize - arc_c;
4691 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4692 ameta > arc_meta_min) {
4693 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4694 total_evicted += bytes;
4697 * If we couldn't evict our target number of bytes from
4698 * metadata, we try to get the rest from data.
4703 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4705 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4706 total_evicted += bytes;
4709 * If we couldn't evict our target number of bytes from
4710 * data, we try to get the rest from data.
4715 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4719 * Adjust ghost lists
4721 * In addition to the above, the ARC also defines target values
4722 * for the ghost lists. The sum of the mru list and mru ghost
4723 * list should never exceed the target size of the cache, and
4724 * the sum of the mru list, mfu list, mru ghost list, and mfu
4725 * ghost list should never exceed twice the target size of the
4726 * cache. The following logic enforces these limits on the ghost
4727 * caches, and evicts from them as needed.
4729 target = zfs_refcount_count(&arc_mru->arcs_size) +
4730 zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4732 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4733 total_evicted += bytes;
4738 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4741 * We assume the sum of the mru list and mfu list is less than
4742 * or equal to arc_c (we enforced this above), which means we
4743 * can use the simpler of the two equations below:
4745 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4746 * mru ghost + mfu ghost <= arc_c
4748 target = zfs_refcount_count(&arc_mru_ghost->arcs_size) +
4749 zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4751 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4752 total_evicted += bytes;
4757 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4759 return (total_evicted);
4763 arc_flush(spa_t *spa, boolean_t retry)
4768 * If retry is B_TRUE, a spa must not be specified since we have
4769 * no good way to determine if all of a spa's buffers have been
4770 * evicted from an arc state.
4772 ASSERT(!retry || spa == 0);
4775 guid = spa_load_guid(spa);
4777 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4778 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4780 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4781 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4783 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4784 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4786 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4787 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4791 arc_reduce_target_size(int64_t to_free)
4793 uint64_t asize = aggsum_value(&arc_size);
4796 if (c > to_free && c - to_free > arc_c_min) {
4797 arc_c = c - to_free;
4798 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4800 arc_c = MAX(asize, arc_c_min);
4802 arc_p = (arc_c >> 1);
4803 ASSERT(arc_c >= arc_c_min);
4804 ASSERT((int64_t)arc_p >= 0);
4809 if (asize > arc_c) {
4810 /* See comment in arc_adjust_cb_check() on why lock+flag */
4811 mutex_enter(&arc_adjust_lock);
4812 arc_adjust_needed = B_TRUE;
4813 mutex_exit(&arc_adjust_lock);
4814 zthr_wakeup(arc_adjust_zthr);
4818 * Return maximum amount of memory that we could possibly use. Reduced
4819 * to half of all memory in user space which is primarily used for testing.
4822 arc_all_memory(void)
4825 #ifdef CONFIG_HIGHMEM
4826 return (ptob(zfs_totalram_pages - totalhigh_pages));
4828 return (ptob(zfs_totalram_pages));
4829 #endif /* CONFIG_HIGHMEM */
4831 return (ptob(physmem) / 2);
4832 #endif /* _KERNEL */
4836 * Return the amount of memory that is considered free. In user space
4837 * which is primarily used for testing we pretend that free memory ranges
4838 * from 0-20% of all memory.
4841 arc_free_memory(void)
4844 #ifdef CONFIG_HIGHMEM
4847 return (ptob(si.freeram - si.freehigh));
4849 return (ptob(nr_free_pages() +
4850 nr_inactive_file_pages() +
4851 nr_inactive_anon_pages() +
4852 nr_slab_reclaimable_pages()));
4854 #endif /* CONFIG_HIGHMEM */
4856 return (spa_get_random(arc_all_memory() * 20 / 100));
4857 #endif /* _KERNEL */
4860 typedef enum free_memory_reason_t {
4865 FMR_PAGES_PP_MAXIMUM,
4868 } free_memory_reason_t;
4870 int64_t last_free_memory;
4871 free_memory_reason_t last_free_reason;
4875 * Additional reserve of pages for pp_reserve.
4877 int64_t arc_pages_pp_reserve = 64;
4880 * Additional reserve of pages for swapfs.
4882 int64_t arc_swapfs_reserve = 64;
4883 #endif /* _KERNEL */
4886 * Return the amount of memory that can be consumed before reclaim will be
4887 * needed. Positive if there is sufficient free memory, negative indicates
4888 * the amount of memory that needs to be freed up.
4891 arc_available_memory(void)
4893 int64_t lowest = INT64_MAX;
4894 free_memory_reason_t r = FMR_UNKNOWN;
4901 pgcnt_t needfree = btop(arc_need_free);
4902 pgcnt_t lotsfree = btop(arc_sys_free);
4903 pgcnt_t desfree = 0;
4904 pgcnt_t freemem = btop(arc_free_memory());
4908 n = PAGESIZE * (-needfree);
4916 * check that we're out of range of the pageout scanner. It starts to
4917 * schedule paging if freemem is less than lotsfree and needfree.
4918 * lotsfree is the high-water mark for pageout, and needfree is the
4919 * number of needed free pages. We add extra pages here to make sure
4920 * the scanner doesn't start up while we're freeing memory.
4922 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4930 * check to make sure that swapfs has enough space so that anon
4931 * reservations can still succeed. anon_resvmem() checks that the
4932 * availrmem is greater than swapfs_minfree, and the number of reserved
4933 * swap pages. We also add a bit of extra here just to prevent
4934 * circumstances from getting really dire.
4936 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4937 desfree - arc_swapfs_reserve);
4940 r = FMR_SWAPFS_MINFREE;
4944 * Check that we have enough availrmem that memory locking (e.g., via
4945 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4946 * stores the number of pages that cannot be locked; when availrmem
4947 * drops below pages_pp_maximum, page locking mechanisms such as
4948 * page_pp_lock() will fail.)
4950 n = PAGESIZE * (availrmem - pages_pp_maximum -
4951 arc_pages_pp_reserve);
4954 r = FMR_PAGES_PP_MAXIMUM;
4960 * If we're on a 32-bit platform, it's possible that we'll exhaust the
4961 * kernel heap space before we ever run out of available physical
4962 * memory. Most checks of the size of the heap_area compare against
4963 * tune.t_minarmem, which is the minimum available real memory that we
4964 * can have in the system. However, this is generally fixed at 25 pages
4965 * which is so low that it's useless. In this comparison, we seek to
4966 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4967 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4970 n = vmem_size(heap_arena, VMEM_FREE) -
4971 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
4979 * If zio data pages are being allocated out of a separate heap segment,
4980 * then enforce that the size of available vmem for this arena remains
4981 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4983 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4984 * memory (in the zio_arena) free, which can avoid memory
4985 * fragmentation issues.
4987 if (zio_arena != NULL) {
4988 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4989 (vmem_size(zio_arena, VMEM_ALLOC) >>
4990 arc_zio_arena_free_shift);
4997 /* Every 100 calls, free a small amount */
4998 if (spa_get_random(100) == 0)
5000 #endif /* _KERNEL */
5002 last_free_memory = lowest;
5003 last_free_reason = r;
5009 * Determine if the system is under memory pressure and is asking
5010 * to reclaim memory. A return value of B_TRUE indicates that the system
5011 * is under memory pressure and that the arc should adjust accordingly.
5014 arc_reclaim_needed(void)
5016 return (arc_available_memory() < 0);
5020 arc_kmem_reap_soon(void)
5023 kmem_cache_t *prev_cache = NULL;
5024 kmem_cache_t *prev_data_cache = NULL;
5025 extern kmem_cache_t *zio_buf_cache[];
5026 extern kmem_cache_t *zio_data_buf_cache[];
5027 extern kmem_cache_t *range_seg_cache;
5030 if ((aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) &&
5031 zfs_arc_meta_prune) {
5033 * We are exceeding our meta-data cache limit.
5034 * Prune some entries to release holds on meta-data.
5036 arc_prune_async(zfs_arc_meta_prune);
5040 * Reclaim unused memory from all kmem caches.
5046 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
5048 /* reach upper limit of cache size on 32-bit */
5049 if (zio_buf_cache[i] == NULL)
5052 if (zio_buf_cache[i] != prev_cache) {
5053 prev_cache = zio_buf_cache[i];
5054 kmem_cache_reap_now(zio_buf_cache[i]);
5056 if (zio_data_buf_cache[i] != prev_data_cache) {
5057 prev_data_cache = zio_data_buf_cache[i];
5058 kmem_cache_reap_now(zio_data_buf_cache[i]);
5061 kmem_cache_reap_now(buf_cache);
5062 kmem_cache_reap_now(hdr_full_cache);
5063 kmem_cache_reap_now(hdr_l2only_cache);
5064 kmem_cache_reap_now(range_seg_cache);
5066 if (zio_arena != NULL) {
5068 * Ask the vmem arena to reclaim unused memory from its
5071 vmem_qcache_reap(zio_arena);
5077 arc_adjust_cb_check(void *arg, zthr_t *zthr)
5080 * This is necessary so that any changes which may have been made to
5081 * many of the zfs_arc_* module parameters will be propagated to
5082 * their actual internal variable counterparts. Without this,
5083 * changing those module params at runtime would have no effect.
5085 arc_tuning_update();
5088 * This is necessary in order to keep the kstat information
5089 * up to date for tools that display kstat data such as the
5090 * mdb ::arc dcmd and the Linux crash utility. These tools
5091 * typically do not call kstat's update function, but simply
5092 * dump out stats from the most recent update. Without
5093 * this call, these commands may show stale stats for the
5094 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
5095 * with this change, the data might be up to 1 second
5096 * out of date(the arc_adjust_zthr has a maximum sleep
5097 * time of 1 second); but that should suffice. The
5098 * arc_state_t structures can be queried directly if more
5099 * accurate information is needed.
5101 if (arc_ksp != NULL)
5102 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
5105 * We have to rely on arc_get_data_impl() to tell us when to adjust,
5106 * rather than checking if we are overflowing here, so that we are
5107 * sure to not leave arc_get_data_impl() waiting on
5108 * arc_adjust_waiters_cv. If we have become "not overflowing" since
5109 * arc_get_data_impl() checked, we need to wake it up. We could
5110 * broadcast the CV here, but arc_get_data_impl() may have not yet
5111 * gone to sleep. We would need to use a mutex to ensure that this
5112 * function doesn't broadcast until arc_get_data_impl() has gone to
5113 * sleep (e.g. the arc_adjust_lock). However, the lock ordering of
5114 * such a lock would necessarily be incorrect with respect to the
5115 * zthr_lock, which is held before this function is called, and is
5116 * held by arc_get_data_impl() when it calls zthr_wakeup().
5118 return (arc_adjust_needed);
5122 * Keep arc_size under arc_c by running arc_adjust which evicts data
5127 arc_adjust_cb(void *arg, zthr_t *zthr)
5129 uint64_t evicted = 0;
5130 fstrans_cookie_t cookie = spl_fstrans_mark();
5132 /* Evict from cache */
5133 evicted = arc_adjust();
5136 * If evicted is zero, we couldn't evict anything
5137 * via arc_adjust(). This could be due to hash lock
5138 * collisions, but more likely due to the majority of
5139 * arc buffers being unevictable. Therefore, even if
5140 * arc_size is above arc_c, another pass is unlikely to
5141 * be helpful and could potentially cause us to enter an
5142 * infinite loop. Additionally, zthr_iscancelled() is
5143 * checked here so that if the arc is shutting down, the
5144 * broadcast will wake any remaining arc adjust waiters.
5146 mutex_enter(&arc_adjust_lock);
5147 arc_adjust_needed = !zthr_iscancelled(arc_adjust_zthr) &&
5148 evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0;
5149 if (!arc_adjust_needed) {
5151 * We're either no longer overflowing, or we
5152 * can't evict anything more, so we should wake
5153 * arc_get_data_impl() sooner.
5155 cv_broadcast(&arc_adjust_waiters_cv);
5158 mutex_exit(&arc_adjust_lock);
5159 spl_fstrans_unmark(cookie);
5164 arc_reap_cb_check(void *arg, zthr_t *zthr)
5166 int64_t free_memory = arc_available_memory();
5169 * If a kmem reap is already active, don't schedule more. We must
5170 * check for this because kmem_cache_reap_soon() won't actually
5171 * block on the cache being reaped (this is to prevent callers from
5172 * becoming implicitly blocked by a system-wide kmem reap -- which,
5173 * on a system with many, many full magazines, can take minutes).
5175 if (!kmem_cache_reap_active() && free_memory < 0) {
5177 arc_no_grow = B_TRUE;
5180 * Wait at least zfs_grow_retry (default 5) seconds
5181 * before considering growing.
5183 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
5185 } else if (free_memory < arc_c >> arc_no_grow_shift) {
5186 arc_no_grow = B_TRUE;
5187 } else if (gethrtime() >= arc_growtime) {
5188 arc_no_grow = B_FALSE;
5195 * Keep enough free memory in the system by reaping the ARC's kmem
5196 * caches. To cause more slabs to be reapable, we may reduce the
5197 * target size of the cache (arc_c), causing the arc_adjust_cb()
5198 * to free more buffers.
5202 arc_reap_cb(void *arg, zthr_t *zthr)
5204 int64_t free_memory;
5205 fstrans_cookie_t cookie = spl_fstrans_mark();
5208 * Kick off asynchronous kmem_reap()'s of all our caches.
5210 arc_kmem_reap_soon();
5213 * Wait at least arc_kmem_cache_reap_retry_ms between
5214 * arc_kmem_reap_soon() calls. Without this check it is possible to
5215 * end up in a situation where we spend lots of time reaping
5216 * caches, while we're near arc_c_min. Waiting here also gives the
5217 * subsequent free memory check a chance of finding that the
5218 * asynchronous reap has already freed enough memory, and we don't
5219 * need to call arc_reduce_target_size().
5221 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
5224 * Reduce the target size as needed to maintain the amount of free
5225 * memory in the system at a fraction of the arc_size (1/128th by
5226 * default). If oversubscribed (free_memory < 0) then reduce the
5227 * target arc_size by the deficit amount plus the fractional
5228 * amount. If free memory is positive but less then the fractional
5229 * amount, reduce by what is needed to hit the fractional amount.
5231 free_memory = arc_available_memory();
5234 (arc_c >> arc_shrink_shift) - free_memory;
5237 to_free = MAX(to_free, arc_need_free);
5239 arc_reduce_target_size(to_free);
5241 spl_fstrans_unmark(cookie);
5246 * Determine the amount of memory eligible for eviction contained in the
5247 * ARC. All clean data reported by the ghost lists can always be safely
5248 * evicted. Due to arc_c_min, the same does not hold for all clean data
5249 * contained by the regular mru and mfu lists.
5251 * In the case of the regular mru and mfu lists, we need to report as
5252 * much clean data as possible, such that evicting that same reported
5253 * data will not bring arc_size below arc_c_min. Thus, in certain
5254 * circumstances, the total amount of clean data in the mru and mfu
5255 * lists might not actually be evictable.
5257 * The following two distinct cases are accounted for:
5259 * 1. The sum of the amount of dirty data contained by both the mru and
5260 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5261 * is greater than or equal to arc_c_min.
5262 * (i.e. amount of dirty data >= arc_c_min)
5264 * This is the easy case; all clean data contained by the mru and mfu
5265 * lists is evictable. Evicting all clean data can only drop arc_size
5266 * to the amount of dirty data, which is greater than arc_c_min.
5268 * 2. The sum of the amount of dirty data contained by both the mru and
5269 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5270 * is less than arc_c_min.
5271 * (i.e. arc_c_min > amount of dirty data)
5273 * 2.1. arc_size is greater than or equal arc_c_min.
5274 * (i.e. arc_size >= arc_c_min > amount of dirty data)
5276 * In this case, not all clean data from the regular mru and mfu
5277 * lists is actually evictable; we must leave enough clean data
5278 * to keep arc_size above arc_c_min. Thus, the maximum amount of
5279 * evictable data from the two lists combined, is exactly the
5280 * difference between arc_size and arc_c_min.
5282 * 2.2. arc_size is less than arc_c_min
5283 * (i.e. arc_c_min > arc_size > amount of dirty data)
5285 * In this case, none of the data contained in the mru and mfu
5286 * lists is evictable, even if it's clean. Since arc_size is
5287 * already below arc_c_min, evicting any more would only
5288 * increase this negative difference.
5291 arc_evictable_memory(void)
5293 int64_t asize = aggsum_value(&arc_size);
5294 uint64_t arc_clean =
5295 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_DATA]) +
5296 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA]) +
5297 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_DATA]) +
5298 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
5299 uint64_t arc_dirty = MAX((int64_t)asize - (int64_t)arc_clean, 0);
5302 * Scale reported evictable memory in proportion to page cache, cap
5303 * at specified min/max.
5305 uint64_t min = (ptob(nr_file_pages()) / 100) * zfs_arc_pc_percent;
5306 min = MAX(arc_c_min, MIN(arc_c_max, min));
5308 if (arc_dirty >= min)
5311 return (MAX((int64_t)asize - (int64_t)min, 0));
5315 * If sc->nr_to_scan is zero, the caller is requesting a query of the
5316 * number of objects which can potentially be freed. If it is nonzero,
5317 * the request is to free that many objects.
5319 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
5320 * in struct shrinker and also require the shrinker to return the number
5323 * Older kernels require the shrinker to return the number of freeable
5324 * objects following the freeing of nr_to_free.
5326 static spl_shrinker_t
5327 __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc)
5331 /* The arc is considered warm once reclaim has occurred */
5332 if (unlikely(arc_warm == B_FALSE))
5335 /* Return the potential number of reclaimable pages */
5336 pages = btop((int64_t)arc_evictable_memory());
5337 if (sc->nr_to_scan == 0)
5340 /* Not allowed to perform filesystem reclaim */
5341 if (!(sc->gfp_mask & __GFP_FS))
5342 return (SHRINK_STOP);
5344 /* Reclaim in progress */
5345 if (mutex_tryenter(&arc_adjust_lock) == 0) {
5346 ARCSTAT_INCR(arcstat_need_free, ptob(sc->nr_to_scan));
5350 mutex_exit(&arc_adjust_lock);
5353 * Evict the requested number of pages by shrinking arc_c the
5357 arc_reduce_target_size(ptob(sc->nr_to_scan));
5358 if (current_is_kswapd())
5359 arc_kmem_reap_soon();
5360 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
5361 pages = MAX((int64_t)pages -
5362 (int64_t)btop(arc_evictable_memory()), 0);
5364 pages = btop(arc_evictable_memory());
5367 * We've shrunk what we can, wake up threads.
5369 cv_broadcast(&arc_adjust_waiters_cv);
5371 pages = SHRINK_STOP;
5374 * When direct reclaim is observed it usually indicates a rapid
5375 * increase in memory pressure. This occurs because the kswapd
5376 * threads were unable to asynchronously keep enough free memory
5377 * available. In this case set arc_no_grow to briefly pause arc
5378 * growth to avoid compounding the memory pressure.
5380 if (current_is_kswapd()) {
5381 ARCSTAT_BUMP(arcstat_memory_indirect_count);
5383 arc_no_grow = B_TRUE;
5384 arc_kmem_reap_soon();
5385 ARCSTAT_BUMP(arcstat_memory_direct_count);
5390 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func);
5392 SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS);
5393 #endif /* _KERNEL */
5396 * Adapt arc info given the number of bytes we are trying to add and
5397 * the state that we are coming from. This function is only called
5398 * when we are adding new content to the cache.
5401 arc_adapt(int bytes, arc_state_t *state)
5404 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
5405 int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size);
5406 int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size);
5408 if (state == arc_l2c_only)
5413 * Adapt the target size of the MRU list:
5414 * - if we just hit in the MRU ghost list, then increase
5415 * the target size of the MRU list.
5416 * - if we just hit in the MFU ghost list, then increase
5417 * the target size of the MFU list by decreasing the
5418 * target size of the MRU list.
5420 if (state == arc_mru_ghost) {
5421 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
5422 if (!zfs_arc_p_dampener_disable)
5423 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
5425 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
5426 } else if (state == arc_mfu_ghost) {
5429 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
5430 if (!zfs_arc_p_dampener_disable)
5431 mult = MIN(mult, 10);
5433 delta = MIN(bytes * mult, arc_p);
5434 arc_p = MAX(arc_p_min, arc_p - delta);
5436 ASSERT((int64_t)arc_p >= 0);
5439 * Wake reap thread if we do not have any available memory
5441 if (arc_reclaim_needed()) {
5442 zthr_wakeup(arc_reap_zthr);
5449 if (arc_c >= arc_c_max)
5453 * If we're within (2 * maxblocksize) bytes of the target
5454 * cache size, increment the target cache size
5456 ASSERT3U(arc_c, >=, 2ULL << SPA_MAXBLOCKSHIFT);
5457 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >=
5459 atomic_add_64(&arc_c, (int64_t)bytes);
5460 if (arc_c > arc_c_max)
5462 else if (state == arc_anon)
5463 atomic_add_64(&arc_p, (int64_t)bytes);
5467 ASSERT((int64_t)arc_p >= 0);
5471 * Check if arc_size has grown past our upper threshold, determined by
5472 * zfs_arc_overflow_shift.
5475 arc_is_overflowing(void)
5477 /* Always allow at least one block of overflow */
5478 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5479 arc_c >> zfs_arc_overflow_shift);
5482 * We just compare the lower bound here for performance reasons. Our
5483 * primary goals are to make sure that the arc never grows without
5484 * bound, and that it can reach its maximum size. This check
5485 * accomplishes both goals. The maximum amount we could run over by is
5486 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
5487 * in the ARC. In practice, that's in the tens of MB, which is low
5488 * enough to be safe.
5490 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
5494 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5496 arc_buf_contents_t type = arc_buf_type(hdr);
5498 arc_get_data_impl(hdr, size, tag);
5499 if (type == ARC_BUFC_METADATA) {
5500 return (abd_alloc(size, B_TRUE));
5502 ASSERT(type == ARC_BUFC_DATA);
5503 return (abd_alloc(size, B_FALSE));
5508 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5510 arc_buf_contents_t type = arc_buf_type(hdr);
5512 arc_get_data_impl(hdr, size, tag);
5513 if (type == ARC_BUFC_METADATA) {
5514 return (zio_buf_alloc(size));
5516 ASSERT(type == ARC_BUFC_DATA);
5517 return (zio_data_buf_alloc(size));
5522 * Allocate a block and return it to the caller. If we are hitting the
5523 * hard limit for the cache size, we must sleep, waiting for the eviction
5524 * thread to catch up. If we're past the target size but below the hard
5525 * limit, we'll only signal the reclaim thread and continue on.
5528 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5530 arc_state_t *state = hdr->b_l1hdr.b_state;
5531 arc_buf_contents_t type = arc_buf_type(hdr);
5533 arc_adapt(size, state);
5536 * If arc_size is currently overflowing, and has grown past our
5537 * upper limit, we must be adding data faster than the evict
5538 * thread can evict. Thus, to ensure we don't compound the
5539 * problem by adding more data and forcing arc_size to grow even
5540 * further past it's target size, we halt and wait for the
5541 * eviction thread to catch up.
5543 * It's also possible that the reclaim thread is unable to evict
5544 * enough buffers to get arc_size below the overflow limit (e.g.
5545 * due to buffers being un-evictable, or hash lock collisions).
5546 * In this case, we want to proceed regardless if we're
5547 * overflowing; thus we don't use a while loop here.
5549 if (arc_is_overflowing()) {
5550 mutex_enter(&arc_adjust_lock);
5553 * Now that we've acquired the lock, we may no longer be
5554 * over the overflow limit, lets check.
5556 * We're ignoring the case of spurious wake ups. If that
5557 * were to happen, it'd let this thread consume an ARC
5558 * buffer before it should have (i.e. before we're under
5559 * the overflow limit and were signalled by the reclaim
5560 * thread). As long as that is a rare occurrence, it
5561 * shouldn't cause any harm.
5563 if (arc_is_overflowing()) {
5564 arc_adjust_needed = B_TRUE;
5565 zthr_wakeup(arc_adjust_zthr);
5566 (void) cv_wait(&arc_adjust_waiters_cv,
5569 mutex_exit(&arc_adjust_lock);
5572 VERIFY3U(hdr->b_type, ==, type);
5573 if (type == ARC_BUFC_METADATA) {
5574 arc_space_consume(size, ARC_SPACE_META);
5576 arc_space_consume(size, ARC_SPACE_DATA);
5580 * Update the state size. Note that ghost states have a
5581 * "ghost size" and so don't need to be updated.
5583 if (!GHOST_STATE(state)) {
5585 (void) zfs_refcount_add_many(&state->arcs_size, size, tag);
5588 * If this is reached via arc_read, the link is
5589 * protected by the hash lock. If reached via
5590 * arc_buf_alloc, the header should not be accessed by
5591 * any other thread. And, if reached via arc_read_done,
5592 * the hash lock will protect it if it's found in the
5593 * hash table; otherwise no other thread should be
5594 * trying to [add|remove]_reference it.
5596 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5597 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5598 (void) zfs_refcount_add_many(&state->arcs_esize[type],
5603 * If we are growing the cache, and we are adding anonymous
5604 * data, and we have outgrown arc_p, update arc_p
5606 if (aggsum_compare(&arc_size, arc_c) < 0 &&
5607 hdr->b_l1hdr.b_state == arc_anon &&
5608 (zfs_refcount_count(&arc_anon->arcs_size) +
5609 zfs_refcount_count(&arc_mru->arcs_size) > arc_p))
5610 arc_p = MIN(arc_c, arc_p + size);
5615 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5617 arc_free_data_impl(hdr, size, tag);
5622 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5624 arc_buf_contents_t type = arc_buf_type(hdr);
5626 arc_free_data_impl(hdr, size, tag);
5627 if (type == ARC_BUFC_METADATA) {
5628 zio_buf_free(buf, size);
5630 ASSERT(type == ARC_BUFC_DATA);
5631 zio_data_buf_free(buf, size);
5636 * Free the arc data buffer.
5639 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5641 arc_state_t *state = hdr->b_l1hdr.b_state;
5642 arc_buf_contents_t type = arc_buf_type(hdr);
5644 /* protected by hash lock, if in the hash table */
5645 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5646 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5647 ASSERT(state != arc_anon && state != arc_l2c_only);
5649 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
5652 (void) zfs_refcount_remove_many(&state->arcs_size, size, tag);
5654 VERIFY3U(hdr->b_type, ==, type);
5655 if (type == ARC_BUFC_METADATA) {
5656 arc_space_return(size, ARC_SPACE_META);
5658 ASSERT(type == ARC_BUFC_DATA);
5659 arc_space_return(size, ARC_SPACE_DATA);
5664 * This routine is called whenever a buffer is accessed.
5665 * NOTE: the hash lock is dropped in this function.
5668 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5672 ASSERT(MUTEX_HELD(hash_lock));
5673 ASSERT(HDR_HAS_L1HDR(hdr));
5675 if (hdr->b_l1hdr.b_state == arc_anon) {
5677 * This buffer is not in the cache, and does not
5678 * appear in our "ghost" list. Add the new buffer
5682 ASSERT0(hdr->b_l1hdr.b_arc_access);
5683 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5684 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5685 arc_change_state(arc_mru, hdr, hash_lock);
5687 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5688 now = ddi_get_lbolt();
5691 * If this buffer is here because of a prefetch, then either:
5692 * - clear the flag if this is a "referencing" read
5693 * (any subsequent access will bump this into the MFU state).
5695 * - move the buffer to the head of the list if this is
5696 * another prefetch (to make it less likely to be evicted).
5698 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5699 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5700 /* link protected by hash lock */
5701 ASSERT(multilist_link_active(
5702 &hdr->b_l1hdr.b_arc_node));
5704 arc_hdr_clear_flags(hdr,
5706 ARC_FLAG_PRESCIENT_PREFETCH);
5707 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5708 ARCSTAT_BUMP(arcstat_mru_hits);
5710 hdr->b_l1hdr.b_arc_access = now;
5715 * This buffer has been "accessed" only once so far,
5716 * but it is still in the cache. Move it to the MFU
5719 if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access +
5722 * More than 125ms have passed since we
5723 * instantiated this buffer. Move it to the
5724 * most frequently used state.
5726 hdr->b_l1hdr.b_arc_access = now;
5727 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5728 arc_change_state(arc_mfu, hdr, hash_lock);
5730 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5731 ARCSTAT_BUMP(arcstat_mru_hits);
5732 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5733 arc_state_t *new_state;
5735 * This buffer has been "accessed" recently, but
5736 * was evicted from the cache. Move it to the
5740 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5741 new_state = arc_mru;
5742 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5743 arc_hdr_clear_flags(hdr,
5745 ARC_FLAG_PRESCIENT_PREFETCH);
5747 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5749 new_state = arc_mfu;
5750 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5753 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5754 arc_change_state(new_state, hdr, hash_lock);
5756 atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits);
5757 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5758 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5760 * This buffer has been accessed more than once and is
5761 * still in the cache. Keep it in the MFU state.
5763 * NOTE: an add_reference() that occurred when we did
5764 * the arc_read() will have kicked this off the list.
5765 * If it was a prefetch, we will explicitly move it to
5766 * the head of the list now.
5769 atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits);
5770 ARCSTAT_BUMP(arcstat_mfu_hits);
5771 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5772 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5773 arc_state_t *new_state = arc_mfu;
5775 * This buffer has been accessed more than once but has
5776 * been evicted from the cache. Move it back to the
5780 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5782 * This is a prefetch access...
5783 * move this block back to the MRU state.
5785 new_state = arc_mru;
5788 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5789 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5790 arc_change_state(new_state, hdr, hash_lock);
5792 atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits);
5793 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5794 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5796 * This buffer is on the 2nd Level ARC.
5799 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5800 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5801 arc_change_state(arc_mfu, hdr, hash_lock);
5803 cmn_err(CE_PANIC, "invalid arc state 0x%p",
5804 hdr->b_l1hdr.b_state);
5809 * This routine is called by dbuf_hold() to update the arc_access() state
5810 * which otherwise would be skipped for entries in the dbuf cache.
5813 arc_buf_access(arc_buf_t *buf)
5815 mutex_enter(&buf->b_evict_lock);
5816 arc_buf_hdr_t *hdr = buf->b_hdr;
5819 * Avoid taking the hash_lock when possible as an optimization.
5820 * The header must be checked again under the hash_lock in order
5821 * to handle the case where it is concurrently being released.
5823 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5824 mutex_exit(&buf->b_evict_lock);
5828 kmutex_t *hash_lock = HDR_LOCK(hdr);
5829 mutex_enter(hash_lock);
5831 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5832 mutex_exit(hash_lock);
5833 mutex_exit(&buf->b_evict_lock);
5834 ARCSTAT_BUMP(arcstat_access_skip);
5838 mutex_exit(&buf->b_evict_lock);
5840 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5841 hdr->b_l1hdr.b_state == arc_mfu);
5843 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5844 arc_access(hdr, hash_lock);
5845 mutex_exit(hash_lock);
5847 ARCSTAT_BUMP(arcstat_hits);
5848 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr) && !HDR_PRESCIENT_PREFETCH(hdr),
5849 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5852 /* a generic arc_read_done_func_t which you can use */
5855 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5856 arc_buf_t *buf, void *arg)
5861 bcopy(buf->b_data, arg, arc_buf_size(buf));
5862 arc_buf_destroy(buf, arg);
5865 /* a generic arc_read_done_func_t */
5868 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5869 arc_buf_t *buf, void *arg)
5871 arc_buf_t **bufp = arg;
5874 ASSERT(zio == NULL || zio->io_error != 0);
5877 ASSERT(zio == NULL || zio->io_error == 0);
5879 ASSERT(buf->b_data != NULL);
5884 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5886 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5887 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5888 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5890 if (HDR_COMPRESSION_ENABLED(hdr)) {
5891 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5892 BP_GET_COMPRESS(bp));
5894 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5895 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5896 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5901 arc_read_done(zio_t *zio)
5903 blkptr_t *bp = zio->io_bp;
5904 arc_buf_hdr_t *hdr = zio->io_private;
5905 kmutex_t *hash_lock = NULL;
5906 arc_callback_t *callback_list;
5907 arc_callback_t *acb;
5908 boolean_t freeable = B_FALSE;
5911 * The hdr was inserted into hash-table and removed from lists
5912 * prior to starting I/O. We should find this header, since
5913 * it's in the hash table, and it should be legit since it's
5914 * not possible to evict it during the I/O. The only possible
5915 * reason for it not to be found is if we were freed during the
5918 if (HDR_IN_HASH_TABLE(hdr)) {
5919 arc_buf_hdr_t *found;
5921 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5922 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5923 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5924 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5925 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5927 found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock);
5929 ASSERT((found == hdr &&
5930 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5931 (found == hdr && HDR_L2_READING(hdr)));
5932 ASSERT3P(hash_lock, !=, NULL);
5935 if (BP_IS_PROTECTED(bp)) {
5936 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5937 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5938 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5939 hdr->b_crypt_hdr.b_iv);
5941 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5944 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5945 sizeof (zil_chain_t));
5946 zio_crypt_decode_mac_zil(tmpbuf,
5947 hdr->b_crypt_hdr.b_mac);
5948 abd_return_buf(zio->io_abd, tmpbuf,
5949 sizeof (zil_chain_t));
5951 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
5955 if (zio->io_error == 0) {
5956 /* byteswap if necessary */
5957 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5958 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5959 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5961 hdr->b_l1hdr.b_byteswap =
5962 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5965 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5969 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5970 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5971 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5973 callback_list = hdr->b_l1hdr.b_acb;
5974 ASSERT3P(callback_list, !=, NULL);
5976 if (hash_lock && zio->io_error == 0 &&
5977 hdr->b_l1hdr.b_state == arc_anon) {
5979 * Only call arc_access on anonymous buffers. This is because
5980 * if we've issued an I/O for an evicted buffer, we've already
5981 * called arc_access (to prevent any simultaneous readers from
5982 * getting confused).
5984 arc_access(hdr, hash_lock);
5988 * If a read request has a callback (i.e. acb_done is not NULL), then we
5989 * make a buf containing the data according to the parameters which were
5990 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5991 * aren't needlessly decompressing the data multiple times.
5993 int callback_cnt = 0;
5994 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
6000 if (zio->io_error != 0)
6003 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
6004 &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
6005 acb->acb_compressed, acb->acb_noauth, B_TRUE,
6009 * Assert non-speculative zios didn't fail because an
6010 * encryption key wasn't loaded
6012 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
6016 * If we failed to decrypt, report an error now (as the zio
6017 * layer would have done if it had done the transforms).
6019 if (error == ECKSUM) {
6020 ASSERT(BP_IS_PROTECTED(bp));
6021 error = SET_ERROR(EIO);
6022 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
6023 spa_log_error(zio->io_spa, &acb->acb_zb);
6024 zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
6025 zio->io_spa, NULL, &acb->acb_zb, zio, 0, 0);
6031 * Decompression or decryption failed. Set
6032 * io_error so that when we call acb_done
6033 * (below), we will indicate that the read
6034 * failed. Note that in the unusual case
6035 * where one callback is compressed and another
6036 * uncompressed, we will mark all of them
6037 * as failed, even though the uncompressed
6038 * one can't actually fail. In this case,
6039 * the hdr will not be anonymous, because
6040 * if there are multiple callbacks, it's
6041 * because multiple threads found the same
6042 * arc buf in the hash table.
6044 zio->io_error = error;
6049 * If there are multiple callbacks, we must have the hash lock,
6050 * because the only way for multiple threads to find this hdr is
6051 * in the hash table. This ensures that if there are multiple
6052 * callbacks, the hdr is not anonymous. If it were anonymous,
6053 * we couldn't use arc_buf_destroy() in the error case below.
6055 ASSERT(callback_cnt < 2 || hash_lock != NULL);
6057 hdr->b_l1hdr.b_acb = NULL;
6058 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6059 if (callback_cnt == 0)
6060 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6062 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
6063 callback_list != NULL);
6065 if (zio->io_error == 0) {
6066 arc_hdr_verify(hdr, zio->io_bp);
6068 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
6069 if (hdr->b_l1hdr.b_state != arc_anon)
6070 arc_change_state(arc_anon, hdr, hash_lock);
6071 if (HDR_IN_HASH_TABLE(hdr))
6072 buf_hash_remove(hdr);
6073 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
6077 * Broadcast before we drop the hash_lock to avoid the possibility
6078 * that the hdr (and hence the cv) might be freed before we get to
6079 * the cv_broadcast().
6081 cv_broadcast(&hdr->b_l1hdr.b_cv);
6083 if (hash_lock != NULL) {
6084 mutex_exit(hash_lock);
6087 * This block was freed while we waited for the read to
6088 * complete. It has been removed from the hash table and
6089 * moved to the anonymous state (so that it won't show up
6092 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
6093 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
6096 /* execute each callback and free its structure */
6097 while ((acb = callback_list) != NULL) {
6098 if (acb->acb_done != NULL) {
6099 if (zio->io_error != 0 && acb->acb_buf != NULL) {
6101 * If arc_buf_alloc_impl() fails during
6102 * decompression, the buf will still be
6103 * allocated, and needs to be freed here.
6105 arc_buf_destroy(acb->acb_buf,
6107 acb->acb_buf = NULL;
6109 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
6110 acb->acb_buf, acb->acb_private);
6113 if (acb->acb_zio_dummy != NULL) {
6114 acb->acb_zio_dummy->io_error = zio->io_error;
6115 zio_nowait(acb->acb_zio_dummy);
6118 callback_list = acb->acb_next;
6119 kmem_free(acb, sizeof (arc_callback_t));
6123 arc_hdr_destroy(hdr);
6127 * "Read" the block at the specified DVA (in bp) via the
6128 * cache. If the block is found in the cache, invoke the provided
6129 * callback immediately and return. Note that the `zio' parameter
6130 * in the callback will be NULL in this case, since no IO was
6131 * required. If the block is not in the cache pass the read request
6132 * on to the spa with a substitute callback function, so that the
6133 * requested block will be added to the cache.
6135 * If a read request arrives for a block that has a read in-progress,
6136 * either wait for the in-progress read to complete (and return the
6137 * results); or, if this is a read with a "done" func, add a record
6138 * to the read to invoke the "done" func when the read completes,
6139 * and return; or just return.
6141 * arc_read_done() will invoke all the requested "done" functions
6142 * for readers of this block.
6145 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
6146 arc_read_done_func_t *done, void *private, zio_priority_t priority,
6147 int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
6149 arc_buf_hdr_t *hdr = NULL;
6150 kmutex_t *hash_lock = NULL;
6152 uint64_t guid = spa_load_guid(spa);
6153 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
6154 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
6155 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
6156 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
6157 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
6158 boolean_t embedded_bp = !!BP_IS_EMBEDDED(bp);
6161 ASSERT(!embedded_bp ||
6162 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
6167 * Embedded BP's have no DVA and require no I/O to "read".
6168 * Create an anonymous arc buf to back it.
6170 hdr = buf_hash_find(guid, bp, &hash_lock);
6174 * Determine if we have an L1 cache hit or a cache miss. For simplicity
6175 * we maintain encrypted data seperately from compressed / uncompressed
6176 * data. If the user is requesting raw encrypted data and we don't have
6177 * that in the header we will read from disk to guarantee that we can
6178 * get it even if the encryption keys aren't loaded.
6180 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
6181 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
6182 arc_buf_t *buf = NULL;
6183 *arc_flags |= ARC_FLAG_CACHED;
6185 if (HDR_IO_IN_PROGRESS(hdr)) {
6186 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
6188 ASSERT3P(head_zio, !=, NULL);
6189 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
6190 priority == ZIO_PRIORITY_SYNC_READ) {
6192 * This is a sync read that needs to wait for
6193 * an in-flight async read. Request that the
6194 * zio have its priority upgraded.
6196 zio_change_priority(head_zio, priority);
6197 DTRACE_PROBE1(arc__async__upgrade__sync,
6198 arc_buf_hdr_t *, hdr);
6199 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
6201 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
6202 arc_hdr_clear_flags(hdr,
6203 ARC_FLAG_PREDICTIVE_PREFETCH);
6206 if (*arc_flags & ARC_FLAG_WAIT) {
6207 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6208 mutex_exit(hash_lock);
6211 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6214 arc_callback_t *acb = NULL;
6216 acb = kmem_zalloc(sizeof (arc_callback_t),
6218 acb->acb_done = done;
6219 acb->acb_private = private;
6220 acb->acb_compressed = compressed_read;
6221 acb->acb_encrypted = encrypted_read;
6222 acb->acb_noauth = noauth_read;
6225 acb->acb_zio_dummy = zio_null(pio,
6226 spa, NULL, NULL, NULL, zio_flags);
6228 ASSERT3P(acb->acb_done, !=, NULL);
6229 acb->acb_zio_head = head_zio;
6230 acb->acb_next = hdr->b_l1hdr.b_acb;
6231 hdr->b_l1hdr.b_acb = acb;
6232 mutex_exit(hash_lock);
6235 mutex_exit(hash_lock);
6239 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
6240 hdr->b_l1hdr.b_state == arc_mfu);
6243 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
6245 * This is a demand read which does not have to
6246 * wait for i/o because we did a predictive
6247 * prefetch i/o for it, which has completed.
6250 arc__demand__hit__predictive__prefetch,
6251 arc_buf_hdr_t *, hdr);
6253 arcstat_demand_hit_predictive_prefetch);
6254 arc_hdr_clear_flags(hdr,
6255 ARC_FLAG_PREDICTIVE_PREFETCH);
6258 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
6260 arcstat_demand_hit_prescient_prefetch);
6261 arc_hdr_clear_flags(hdr,
6262 ARC_FLAG_PRESCIENT_PREFETCH);
6265 ASSERT(!embedded_bp || !BP_IS_HOLE(bp));
6267 /* Get a buf with the desired data in it. */
6268 rc = arc_buf_alloc_impl(hdr, spa, zb, private,
6269 encrypted_read, compressed_read, noauth_read,
6273 * Convert authentication and decryption errors
6274 * to EIO (and generate an ereport if needed)
6275 * before leaving the ARC.
6277 rc = SET_ERROR(EIO);
6278 if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
6279 spa_log_error(spa, zb);
6281 FM_EREPORT_ZFS_AUTHENTICATION,
6282 spa, NULL, zb, NULL, 0, 0);
6286 (void) remove_reference(hdr, hash_lock,
6288 arc_buf_destroy_impl(buf);
6292 /* assert any errors weren't due to unloaded keys */
6293 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
6295 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
6296 zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
6297 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6299 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
6300 arc_access(hdr, hash_lock);
6301 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6302 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6303 if (*arc_flags & ARC_FLAG_L2CACHE)
6304 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6305 mutex_exit(hash_lock);
6306 ARCSTAT_BUMP(arcstat_hits);
6307 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6308 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
6309 data, metadata, hits);
6312 done(NULL, zb, bp, buf, private);
6314 uint64_t lsize = BP_GET_LSIZE(bp);
6315 uint64_t psize = BP_GET_PSIZE(bp);
6316 arc_callback_t *acb;
6319 boolean_t devw = B_FALSE;
6324 * Gracefully handle a damaged logical block size as a
6327 if (lsize > spa_maxblocksize(spa)) {
6328 rc = SET_ERROR(ECKSUM);
6334 * This block is not in the cache or it has
6337 arc_buf_hdr_t *exists = NULL;
6338 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
6339 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
6340 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), type,
6344 hdr->b_dva = *BP_IDENTITY(bp);
6345 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
6346 exists = buf_hash_insert(hdr, &hash_lock);
6348 if (exists != NULL) {
6349 /* somebody beat us to the hash insert */
6350 mutex_exit(hash_lock);
6351 buf_discard_identity(hdr);
6352 arc_hdr_destroy(hdr);
6353 goto top; /* restart the IO request */
6357 * This block is in the ghost cache or encrypted data
6358 * was requested and we didn't have it. If it was
6359 * L2-only (and thus didn't have an L1 hdr),
6360 * we realloc the header to add an L1 hdr.
6362 if (!HDR_HAS_L1HDR(hdr)) {
6363 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
6367 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
6368 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6369 ASSERT(!HDR_HAS_RABD(hdr));
6370 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6371 ASSERT0(zfs_refcount_count(
6372 &hdr->b_l1hdr.b_refcnt));
6373 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
6374 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
6375 } else if (HDR_IO_IN_PROGRESS(hdr)) {
6377 * If this header already had an IO in progress
6378 * and we are performing another IO to fetch
6379 * encrypted data we must wait until the first
6380 * IO completes so as not to confuse
6381 * arc_read_done(). This should be very rare
6382 * and so the performance impact shouldn't
6385 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6386 mutex_exit(hash_lock);
6391 * This is a delicate dance that we play here.
6392 * This hdr might be in the ghost list so we access
6393 * it to move it out of the ghost list before we
6394 * initiate the read. If it's a prefetch then
6395 * it won't have a callback so we'll remove the
6396 * reference that arc_buf_alloc_impl() created. We
6397 * do this after we've called arc_access() to
6398 * avoid hitting an assert in remove_reference().
6400 arc_access(hdr, hash_lock);
6401 arc_hdr_alloc_abd(hdr, encrypted_read);
6404 if (encrypted_read) {
6405 ASSERT(HDR_HAS_RABD(hdr));
6406 size = HDR_GET_PSIZE(hdr);
6407 hdr_abd = hdr->b_crypt_hdr.b_rabd;
6408 zio_flags |= ZIO_FLAG_RAW;
6410 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6411 size = arc_hdr_size(hdr);
6412 hdr_abd = hdr->b_l1hdr.b_pabd;
6414 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
6415 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6419 * For authenticated bp's, we do not ask the ZIO layer
6420 * to authenticate them since this will cause the entire
6421 * IO to fail if the key isn't loaded. Instead, we
6422 * defer authentication until arc_buf_fill(), which will
6423 * verify the data when the key is available.
6425 if (BP_IS_AUTHENTICATED(bp))
6426 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
6429 if (*arc_flags & ARC_FLAG_PREFETCH &&
6430 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))
6431 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6432 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6433 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6434 if (*arc_flags & ARC_FLAG_L2CACHE)
6435 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6436 if (BP_IS_AUTHENTICATED(bp))
6437 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6438 if (BP_GET_LEVEL(bp) > 0)
6439 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
6440 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
6441 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
6442 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
6444 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
6445 acb->acb_done = done;
6446 acb->acb_private = private;
6447 acb->acb_compressed = compressed_read;
6448 acb->acb_encrypted = encrypted_read;
6449 acb->acb_noauth = noauth_read;
6452 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6453 hdr->b_l1hdr.b_acb = acb;
6454 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6456 if (HDR_HAS_L2HDR(hdr) &&
6457 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
6458 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
6459 addr = hdr->b_l2hdr.b_daddr;
6461 * Lock out L2ARC device removal.
6463 if (vdev_is_dead(vd) ||
6464 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
6469 * We count both async reads and scrub IOs as asynchronous so
6470 * that both can be upgraded in the event of a cache hit while
6471 * the read IO is still in-flight.
6473 if (priority == ZIO_PRIORITY_ASYNC_READ ||
6474 priority == ZIO_PRIORITY_SCRUB)
6475 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6477 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6480 * At this point, we have a level 1 cache miss or a blkptr
6481 * with embedded data. Try again in L2ARC if possible.
6483 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
6486 * Skip ARC stat bump for block pointers with embedded
6487 * data. The data are read from the blkptr itself via
6488 * decode_embedded_bp_compressed().
6491 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr,
6492 blkptr_t *, bp, uint64_t, lsize,
6493 zbookmark_phys_t *, zb);
6494 ARCSTAT_BUMP(arcstat_misses);
6495 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6496 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data,
6500 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
6502 * Read from the L2ARC if the following are true:
6503 * 1. The L2ARC vdev was previously cached.
6504 * 2. This buffer still has L2ARC metadata.
6505 * 3. This buffer isn't currently writing to the L2ARC.
6506 * 4. The L2ARC entry wasn't evicted, which may
6507 * also have invalidated the vdev.
6508 * 5. This isn't prefetch and l2arc_noprefetch is set.
6510 if (HDR_HAS_L2HDR(hdr) &&
6511 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
6512 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
6513 l2arc_read_callback_t *cb;
6517 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
6518 ARCSTAT_BUMP(arcstat_l2_hits);
6519 atomic_inc_32(&hdr->b_l2hdr.b_hits);
6521 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
6523 cb->l2rcb_hdr = hdr;
6526 cb->l2rcb_flags = zio_flags;
6528 asize = vdev_psize_to_asize(vd, size);
6529 if (asize != size) {
6530 abd = abd_alloc_for_io(asize,
6531 HDR_ISTYPE_METADATA(hdr));
6532 cb->l2rcb_abd = abd;
6537 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6538 addr + asize <= vd->vdev_psize -
6539 VDEV_LABEL_END_SIZE);
6542 * l2arc read. The SCL_L2ARC lock will be
6543 * released by l2arc_read_done().
6544 * Issue a null zio if the underlying buffer
6545 * was squashed to zero size by compression.
6547 ASSERT3U(arc_hdr_get_compress(hdr), !=,
6548 ZIO_COMPRESS_EMPTY);
6549 rzio = zio_read_phys(pio, vd, addr,
6552 l2arc_read_done, cb, priority,
6553 zio_flags | ZIO_FLAG_DONT_CACHE |
6555 ZIO_FLAG_DONT_PROPAGATE |
6556 ZIO_FLAG_DONT_RETRY, B_FALSE);
6557 acb->acb_zio_head = rzio;
6559 if (hash_lock != NULL)
6560 mutex_exit(hash_lock);
6562 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6564 ARCSTAT_INCR(arcstat_l2_read_bytes,
6565 HDR_GET_PSIZE(hdr));
6567 if (*arc_flags & ARC_FLAG_NOWAIT) {
6572 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6573 if (zio_wait(rzio) == 0)
6576 /* l2arc read error; goto zio_read() */
6577 if (hash_lock != NULL)
6578 mutex_enter(hash_lock);
6580 DTRACE_PROBE1(l2arc__miss,
6581 arc_buf_hdr_t *, hdr);
6582 ARCSTAT_BUMP(arcstat_l2_misses);
6583 if (HDR_L2_WRITING(hdr))
6584 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6585 spa_config_exit(spa, SCL_L2ARC, vd);
6589 spa_config_exit(spa, SCL_L2ARC, vd);
6591 * Skip ARC stat bump for block pointers with
6592 * embedded data. The data are read from the blkptr
6593 * itself via decode_embedded_bp_compressed().
6595 if (l2arc_ndev != 0 && !embedded_bp) {
6596 DTRACE_PROBE1(l2arc__miss,
6597 arc_buf_hdr_t *, hdr);
6598 ARCSTAT_BUMP(arcstat_l2_misses);
6602 rzio = zio_read(pio, spa, bp, hdr_abd, size,
6603 arc_read_done, hdr, priority, zio_flags, zb);
6604 acb->acb_zio_head = rzio;
6606 if (hash_lock != NULL)
6607 mutex_exit(hash_lock);
6609 if (*arc_flags & ARC_FLAG_WAIT) {
6610 rc = zio_wait(rzio);
6614 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6619 /* embedded bps don't actually go to disk */
6621 spa_read_history_add(spa, zb, *arc_flags);
6626 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6630 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6632 p->p_private = private;
6633 list_link_init(&p->p_node);
6634 zfs_refcount_create(&p->p_refcnt);
6636 mutex_enter(&arc_prune_mtx);
6637 zfs_refcount_add(&p->p_refcnt, &arc_prune_list);
6638 list_insert_head(&arc_prune_list, p);
6639 mutex_exit(&arc_prune_mtx);
6645 arc_remove_prune_callback(arc_prune_t *p)
6647 boolean_t wait = B_FALSE;
6648 mutex_enter(&arc_prune_mtx);
6649 list_remove(&arc_prune_list, p);
6650 if (zfs_refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6652 mutex_exit(&arc_prune_mtx);
6654 /* wait for arc_prune_task to finish */
6656 taskq_wait_outstanding(arc_prune_taskq, 0);
6657 ASSERT0(zfs_refcount_count(&p->p_refcnt));
6658 zfs_refcount_destroy(&p->p_refcnt);
6659 kmem_free(p, sizeof (*p));
6663 * Notify the arc that a block was freed, and thus will never be used again.
6666 arc_freed(spa_t *spa, const blkptr_t *bp)
6669 kmutex_t *hash_lock;
6670 uint64_t guid = spa_load_guid(spa);
6672 ASSERT(!BP_IS_EMBEDDED(bp));
6674 hdr = buf_hash_find(guid, bp, &hash_lock);
6679 * We might be trying to free a block that is still doing I/O
6680 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6681 * dmu_sync-ed block). If this block is being prefetched, then it
6682 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6683 * until the I/O completes. A block may also have a reference if it is
6684 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6685 * have written the new block to its final resting place on disk but
6686 * without the dedup flag set. This would have left the hdr in the MRU
6687 * state and discoverable. When the txg finally syncs it detects that
6688 * the block was overridden in open context and issues an override I/O.
6689 * Since this is a dedup block, the override I/O will determine if the
6690 * block is already in the DDT. If so, then it will replace the io_bp
6691 * with the bp from the DDT and allow the I/O to finish. When the I/O
6692 * reaches the done callback, dbuf_write_override_done, it will
6693 * check to see if the io_bp and io_bp_override are identical.
6694 * If they are not, then it indicates that the bp was replaced with
6695 * the bp in the DDT and the override bp is freed. This allows
6696 * us to arrive here with a reference on a block that is being
6697 * freed. So if we have an I/O in progress, or a reference to
6698 * this hdr, then we don't destroy the hdr.
6700 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6701 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6702 arc_change_state(arc_anon, hdr, hash_lock);
6703 arc_hdr_destroy(hdr);
6704 mutex_exit(hash_lock);
6706 mutex_exit(hash_lock);
6712 * Release this buffer from the cache, making it an anonymous buffer. This
6713 * must be done after a read and prior to modifying the buffer contents.
6714 * If the buffer has more than one reference, we must make
6715 * a new hdr for the buffer.
6718 arc_release(arc_buf_t *buf, void *tag)
6720 arc_buf_hdr_t *hdr = buf->b_hdr;
6723 * It would be nice to assert that if its DMU metadata (level >
6724 * 0 || it's the dnode file), then it must be syncing context.
6725 * But we don't know that information at this level.
6728 mutex_enter(&buf->b_evict_lock);
6730 ASSERT(HDR_HAS_L1HDR(hdr));
6733 * We don't grab the hash lock prior to this check, because if
6734 * the buffer's header is in the arc_anon state, it won't be
6735 * linked into the hash table.
6737 if (hdr->b_l1hdr.b_state == arc_anon) {
6738 mutex_exit(&buf->b_evict_lock);
6739 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6740 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6741 ASSERT(!HDR_HAS_L2HDR(hdr));
6742 ASSERT(HDR_EMPTY(hdr));
6744 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6745 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6746 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6748 hdr->b_l1hdr.b_arc_access = 0;
6751 * If the buf is being overridden then it may already
6752 * have a hdr that is not empty.
6754 buf_discard_identity(hdr);
6760 kmutex_t *hash_lock = HDR_LOCK(hdr);
6761 mutex_enter(hash_lock);
6764 * This assignment is only valid as long as the hash_lock is
6765 * held, we must be careful not to reference state or the
6766 * b_state field after dropping the lock.
6768 arc_state_t *state = hdr->b_l1hdr.b_state;
6769 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6770 ASSERT3P(state, !=, arc_anon);
6772 /* this buffer is not on any list */
6773 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6775 if (HDR_HAS_L2HDR(hdr)) {
6776 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6779 * We have to recheck this conditional again now that
6780 * we're holding the l2ad_mtx to prevent a race with
6781 * another thread which might be concurrently calling
6782 * l2arc_evict(). In that case, l2arc_evict() might have
6783 * destroyed the header's L2 portion as we were waiting
6784 * to acquire the l2ad_mtx.
6786 if (HDR_HAS_L2HDR(hdr))
6787 arc_hdr_l2hdr_destroy(hdr);
6789 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6793 * Do we have more than one buf?
6795 if (hdr->b_l1hdr.b_bufcnt > 1) {
6796 arc_buf_hdr_t *nhdr;
6797 uint64_t spa = hdr->b_spa;
6798 uint64_t psize = HDR_GET_PSIZE(hdr);
6799 uint64_t lsize = HDR_GET_LSIZE(hdr);
6800 boolean_t protected = HDR_PROTECTED(hdr);
6801 enum zio_compress compress = arc_hdr_get_compress(hdr);
6802 arc_buf_contents_t type = arc_buf_type(hdr);
6803 VERIFY3U(hdr->b_type, ==, type);
6805 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6806 (void) remove_reference(hdr, hash_lock, tag);
6808 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6809 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6810 ASSERT(ARC_BUF_LAST(buf));
6814 * Pull the data off of this hdr and attach it to
6815 * a new anonymous hdr. Also find the last buffer
6816 * in the hdr's buffer list.
6818 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6819 ASSERT3P(lastbuf, !=, NULL);
6822 * If the current arc_buf_t and the hdr are sharing their data
6823 * buffer, then we must stop sharing that block.
6825 if (arc_buf_is_shared(buf)) {
6826 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6827 VERIFY(!arc_buf_is_shared(lastbuf));
6830 * First, sever the block sharing relationship between
6831 * buf and the arc_buf_hdr_t.
6833 arc_unshare_buf(hdr, buf);
6836 * Now we need to recreate the hdr's b_pabd. Since we
6837 * have lastbuf handy, we try to share with it, but if
6838 * we can't then we allocate a new b_pabd and copy the
6839 * data from buf into it.
6841 if (arc_can_share(hdr, lastbuf)) {
6842 arc_share_buf(hdr, lastbuf);
6844 arc_hdr_alloc_abd(hdr, B_FALSE);
6845 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6846 buf->b_data, psize);
6848 VERIFY3P(lastbuf->b_data, !=, NULL);
6849 } else if (HDR_SHARED_DATA(hdr)) {
6851 * Uncompressed shared buffers are always at the end
6852 * of the list. Compressed buffers don't have the
6853 * same requirements. This makes it hard to
6854 * simply assert that the lastbuf is shared so
6855 * we rely on the hdr's compression flags to determine
6856 * if we have a compressed, shared buffer.
6858 ASSERT(arc_buf_is_shared(lastbuf) ||
6859 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6860 ASSERT(!ARC_BUF_SHARED(buf));
6863 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6864 ASSERT3P(state, !=, arc_l2c_only);
6866 (void) zfs_refcount_remove_many(&state->arcs_size,
6867 arc_buf_size(buf), buf);
6869 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6870 ASSERT3P(state, !=, arc_l2c_only);
6871 (void) zfs_refcount_remove_many(
6872 &state->arcs_esize[type],
6873 arc_buf_size(buf), buf);
6876 hdr->b_l1hdr.b_bufcnt -= 1;
6877 if (ARC_BUF_ENCRYPTED(buf))
6878 hdr->b_crypt_hdr.b_ebufcnt -= 1;
6880 arc_cksum_verify(buf);
6881 arc_buf_unwatch(buf);
6883 /* if this is the last uncompressed buf free the checksum */
6884 if (!arc_hdr_has_uncompressed_buf(hdr))
6885 arc_cksum_free(hdr);
6887 mutex_exit(hash_lock);
6890 * Allocate a new hdr. The new hdr will contain a b_pabd
6891 * buffer which will be freed in arc_write().
6893 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6894 compress, type, HDR_HAS_RABD(hdr));
6895 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6896 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6897 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6898 VERIFY3U(nhdr->b_type, ==, type);
6899 ASSERT(!HDR_SHARED_DATA(nhdr));
6901 nhdr->b_l1hdr.b_buf = buf;
6902 nhdr->b_l1hdr.b_bufcnt = 1;
6903 if (ARC_BUF_ENCRYPTED(buf))
6904 nhdr->b_crypt_hdr.b_ebufcnt = 1;
6905 nhdr->b_l1hdr.b_mru_hits = 0;
6906 nhdr->b_l1hdr.b_mru_ghost_hits = 0;
6907 nhdr->b_l1hdr.b_mfu_hits = 0;
6908 nhdr->b_l1hdr.b_mfu_ghost_hits = 0;
6909 nhdr->b_l1hdr.b_l2_hits = 0;
6910 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6913 mutex_exit(&buf->b_evict_lock);
6914 (void) zfs_refcount_add_many(&arc_anon->arcs_size,
6915 arc_buf_size(buf), buf);
6917 mutex_exit(&buf->b_evict_lock);
6918 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6919 /* protected by hash lock, or hdr is on arc_anon */
6920 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6921 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6922 hdr->b_l1hdr.b_mru_hits = 0;
6923 hdr->b_l1hdr.b_mru_ghost_hits = 0;
6924 hdr->b_l1hdr.b_mfu_hits = 0;
6925 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
6926 hdr->b_l1hdr.b_l2_hits = 0;
6927 arc_change_state(arc_anon, hdr, hash_lock);
6928 hdr->b_l1hdr.b_arc_access = 0;
6930 mutex_exit(hash_lock);
6931 buf_discard_identity(hdr);
6937 arc_released(arc_buf_t *buf)
6941 mutex_enter(&buf->b_evict_lock);
6942 released = (buf->b_data != NULL &&
6943 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6944 mutex_exit(&buf->b_evict_lock);
6950 arc_referenced(arc_buf_t *buf)
6954 mutex_enter(&buf->b_evict_lock);
6955 referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6956 mutex_exit(&buf->b_evict_lock);
6957 return (referenced);
6962 arc_write_ready(zio_t *zio)
6964 arc_write_callback_t *callback = zio->io_private;
6965 arc_buf_t *buf = callback->awcb_buf;
6966 arc_buf_hdr_t *hdr = buf->b_hdr;
6967 blkptr_t *bp = zio->io_bp;
6968 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6969 fstrans_cookie_t cookie = spl_fstrans_mark();
6971 ASSERT(HDR_HAS_L1HDR(hdr));
6972 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6973 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6976 * If we're reexecuting this zio because the pool suspended, then
6977 * cleanup any state that was previously set the first time the
6978 * callback was invoked.
6980 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6981 arc_cksum_free(hdr);
6982 arc_buf_unwatch(buf);
6983 if (hdr->b_l1hdr.b_pabd != NULL) {
6984 if (arc_buf_is_shared(buf)) {
6985 arc_unshare_buf(hdr, buf);
6987 arc_hdr_free_abd(hdr, B_FALSE);
6991 if (HDR_HAS_RABD(hdr))
6992 arc_hdr_free_abd(hdr, B_TRUE);
6994 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6995 ASSERT(!HDR_HAS_RABD(hdr));
6996 ASSERT(!HDR_SHARED_DATA(hdr));
6997 ASSERT(!arc_buf_is_shared(buf));
6999 callback->awcb_ready(zio, buf, callback->awcb_private);
7001 if (HDR_IO_IN_PROGRESS(hdr))
7002 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
7004 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
7006 if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr))
7007 hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp));
7009 if (BP_IS_PROTECTED(bp)) {
7010 /* ZIL blocks are written through zio_rewrite */
7011 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
7012 ASSERT(HDR_PROTECTED(hdr));
7014 if (BP_SHOULD_BYTESWAP(bp)) {
7015 if (BP_GET_LEVEL(bp) > 0) {
7016 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
7018 hdr->b_l1hdr.b_byteswap =
7019 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
7022 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
7025 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
7026 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
7027 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
7028 hdr->b_crypt_hdr.b_iv);
7029 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
7033 * If this block was written for raw encryption but the zio layer
7034 * ended up only authenticating it, adjust the buffer flags now.
7036 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
7037 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
7038 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
7039 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
7040 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
7041 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
7042 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
7043 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
7046 /* this must be done after the buffer flags are adjusted */
7047 arc_cksum_compute(buf);
7049 enum zio_compress compress;
7050 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
7051 compress = ZIO_COMPRESS_OFF;
7053 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
7054 compress = BP_GET_COMPRESS(bp);
7056 HDR_SET_PSIZE(hdr, psize);
7057 arc_hdr_set_compress(hdr, compress);
7059 if (zio->io_error != 0 || psize == 0)
7063 * Fill the hdr with data. If the buffer is encrypted we have no choice
7064 * but to copy the data into b_radb. If the hdr is compressed, the data
7065 * we want is available from the zio, otherwise we can take it from
7068 * We might be able to share the buf's data with the hdr here. However,
7069 * doing so would cause the ARC to be full of linear ABDs if we write a
7070 * lot of shareable data. As a compromise, we check whether scattered
7071 * ABDs are allowed, and assume that if they are then the user wants
7072 * the ARC to be primarily filled with them regardless of the data being
7073 * written. Therefore, if they're allowed then we allocate one and copy
7074 * the data into it; otherwise, we share the data directly if we can.
7076 if (ARC_BUF_ENCRYPTED(buf)) {
7077 ASSERT3U(psize, >, 0);
7078 ASSERT(ARC_BUF_COMPRESSED(buf));
7079 arc_hdr_alloc_abd(hdr, B_TRUE);
7080 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
7081 } else if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
7083 * Ideally, we would always copy the io_abd into b_pabd, but the
7084 * user may have disabled compressed ARC, thus we must check the
7085 * hdr's compression setting rather than the io_bp's.
7087 if (BP_IS_ENCRYPTED(bp)) {
7088 ASSERT3U(psize, >, 0);
7089 arc_hdr_alloc_abd(hdr, B_TRUE);
7090 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
7091 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
7092 !ARC_BUF_COMPRESSED(buf)) {
7093 ASSERT3U(psize, >, 0);
7094 arc_hdr_alloc_abd(hdr, B_FALSE);
7095 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
7097 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
7098 arc_hdr_alloc_abd(hdr, B_FALSE);
7099 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
7103 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
7104 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
7105 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
7107 arc_share_buf(hdr, buf);
7111 arc_hdr_verify(hdr, bp);
7112 spl_fstrans_unmark(cookie);
7116 arc_write_children_ready(zio_t *zio)
7118 arc_write_callback_t *callback = zio->io_private;
7119 arc_buf_t *buf = callback->awcb_buf;
7121 callback->awcb_children_ready(zio, buf, callback->awcb_private);
7125 * The SPA calls this callback for each physical write that happens on behalf
7126 * of a logical write. See the comment in dbuf_write_physdone() for details.
7129 arc_write_physdone(zio_t *zio)
7131 arc_write_callback_t *cb = zio->io_private;
7132 if (cb->awcb_physdone != NULL)
7133 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
7137 arc_write_done(zio_t *zio)
7139 arc_write_callback_t *callback = zio->io_private;
7140 arc_buf_t *buf = callback->awcb_buf;
7141 arc_buf_hdr_t *hdr = buf->b_hdr;
7143 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
7145 if (zio->io_error == 0) {
7146 arc_hdr_verify(hdr, zio->io_bp);
7148 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
7149 buf_discard_identity(hdr);
7151 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
7152 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
7155 ASSERT(HDR_EMPTY(hdr));
7159 * If the block to be written was all-zero or compressed enough to be
7160 * embedded in the BP, no write was performed so there will be no
7161 * dva/birth/checksum. The buffer must therefore remain anonymous
7164 if (!HDR_EMPTY(hdr)) {
7165 arc_buf_hdr_t *exists;
7166 kmutex_t *hash_lock;
7168 ASSERT3U(zio->io_error, ==, 0);
7170 arc_cksum_verify(buf);
7172 exists = buf_hash_insert(hdr, &hash_lock);
7173 if (exists != NULL) {
7175 * This can only happen if we overwrite for
7176 * sync-to-convergence, because we remove
7177 * buffers from the hash table when we arc_free().
7179 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
7180 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
7181 panic("bad overwrite, hdr=%p exists=%p",
7182 (void *)hdr, (void *)exists);
7183 ASSERT(zfs_refcount_is_zero(
7184 &exists->b_l1hdr.b_refcnt));
7185 arc_change_state(arc_anon, exists, hash_lock);
7186 mutex_exit(hash_lock);
7187 arc_hdr_destroy(exists);
7188 exists = buf_hash_insert(hdr, &hash_lock);
7189 ASSERT3P(exists, ==, NULL);
7190 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
7192 ASSERT(zio->io_prop.zp_nopwrite);
7193 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
7194 panic("bad nopwrite, hdr=%p exists=%p",
7195 (void *)hdr, (void *)exists);
7198 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
7199 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
7200 ASSERT(BP_GET_DEDUP(zio->io_bp));
7201 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
7204 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
7205 /* if it's not anon, we are doing a scrub */
7206 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
7207 arc_access(hdr, hash_lock);
7208 mutex_exit(hash_lock);
7210 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
7213 ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
7214 callback->awcb_done(zio, buf, callback->awcb_private);
7216 abd_put(zio->io_abd);
7217 kmem_free(callback, sizeof (arc_write_callback_t));
7221 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
7222 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc,
7223 const zio_prop_t *zp, arc_write_done_func_t *ready,
7224 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
7225 arc_write_done_func_t *done, void *private, zio_priority_t priority,
7226 int zio_flags, const zbookmark_phys_t *zb)
7228 arc_buf_hdr_t *hdr = buf->b_hdr;
7229 arc_write_callback_t *callback;
7231 zio_prop_t localprop = *zp;
7233 ASSERT3P(ready, !=, NULL);
7234 ASSERT3P(done, !=, NULL);
7235 ASSERT(!HDR_IO_ERROR(hdr));
7236 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
7237 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
7238 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
7240 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
7242 if (ARC_BUF_ENCRYPTED(buf)) {
7243 ASSERT(ARC_BUF_COMPRESSED(buf));
7244 localprop.zp_encrypt = B_TRUE;
7245 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
7246 localprop.zp_byteorder =
7247 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
7248 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
7249 bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt,
7251 bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv,
7253 bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac,
7255 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
7256 localprop.zp_nopwrite = B_FALSE;
7257 localprop.zp_copies =
7258 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
7260 zio_flags |= ZIO_FLAG_RAW;
7261 } else if (ARC_BUF_COMPRESSED(buf)) {
7262 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
7263 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
7264 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
7266 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
7267 callback->awcb_ready = ready;
7268 callback->awcb_children_ready = children_ready;
7269 callback->awcb_physdone = physdone;
7270 callback->awcb_done = done;
7271 callback->awcb_private = private;
7272 callback->awcb_buf = buf;
7275 * The hdr's b_pabd is now stale, free it now. A new data block
7276 * will be allocated when the zio pipeline calls arc_write_ready().
7278 if (hdr->b_l1hdr.b_pabd != NULL) {
7280 * If the buf is currently sharing the data block with
7281 * the hdr then we need to break that relationship here.
7282 * The hdr will remain with a NULL data pointer and the
7283 * buf will take sole ownership of the block.
7285 if (arc_buf_is_shared(buf)) {
7286 arc_unshare_buf(hdr, buf);
7288 arc_hdr_free_abd(hdr, B_FALSE);
7290 VERIFY3P(buf->b_data, !=, NULL);
7293 if (HDR_HAS_RABD(hdr))
7294 arc_hdr_free_abd(hdr, B_TRUE);
7296 if (!(zio_flags & ZIO_FLAG_RAW))
7297 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
7299 ASSERT(!arc_buf_is_shared(buf));
7300 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
7302 zio = zio_write(pio, spa, txg, bp,
7303 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
7304 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
7305 (children_ready != NULL) ? arc_write_children_ready : NULL,
7306 arc_write_physdone, arc_write_done, callback,
7307 priority, zio_flags, zb);
7313 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
7316 uint64_t available_memory = arc_free_memory();
7320 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
7323 if (available_memory > arc_all_memory() * arc_lotsfree_percent / 100)
7326 if (txg > spa->spa_lowmem_last_txg) {
7327 spa->spa_lowmem_last_txg = txg;
7328 spa->spa_lowmem_page_load = 0;
7331 * If we are in pageout, we know that memory is already tight,
7332 * the arc is already going to be evicting, so we just want to
7333 * continue to let page writes occur as quickly as possible.
7335 if (current_is_kswapd()) {
7336 if (spa->spa_lowmem_page_load >
7337 MAX(arc_sys_free / 4, available_memory) / 4) {
7338 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
7339 return (SET_ERROR(ERESTART));
7341 /* Note: reserve is inflated, so we deflate */
7342 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
7344 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
7345 /* memory is low, delay before restarting */
7346 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
7347 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
7348 return (SET_ERROR(EAGAIN));
7350 spa->spa_lowmem_page_load = 0;
7351 #endif /* _KERNEL */
7356 arc_tempreserve_clear(uint64_t reserve)
7358 atomic_add_64(&arc_tempreserve, -reserve);
7359 ASSERT((int64_t)arc_tempreserve >= 0);
7363 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
7369 reserve > arc_c/4 &&
7370 reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT))
7371 arc_c = MIN(arc_c_max, reserve * 4);
7374 * Throttle when the calculated memory footprint for the TXG
7375 * exceeds the target ARC size.
7377 if (reserve > arc_c) {
7378 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
7379 return (SET_ERROR(ERESTART));
7383 * Don't count loaned bufs as in flight dirty data to prevent long
7384 * network delays from blocking transactions that are ready to be
7385 * assigned to a txg.
7388 /* assert that it has not wrapped around */
7389 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
7391 anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) -
7392 arc_loaned_bytes), 0);
7395 * Writes will, almost always, require additional memory allocations
7396 * in order to compress/encrypt/etc the data. We therefore need to
7397 * make sure that there is sufficient available memory for this.
7399 error = arc_memory_throttle(spa, reserve, txg);
7404 * Throttle writes when the amount of dirty data in the cache
7405 * gets too large. We try to keep the cache less than half full
7406 * of dirty blocks so that our sync times don't grow too large.
7408 * In the case of one pool being built on another pool, we want
7409 * to make sure we don't end up throttling the lower (backing)
7410 * pool when the upper pool is the majority contributor to dirty
7411 * data. To insure we make forward progress during throttling, we
7412 * also check the current pool's net dirty data and only throttle
7413 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
7414 * data in the cache.
7416 * Note: if two requests come in concurrently, we might let them
7417 * both succeed, when one of them should fail. Not a huge deal.
7419 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
7420 uint64_t spa_dirty_anon = spa_dirty_data(spa);
7422 if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
7423 anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
7424 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
7426 uint64_t meta_esize = zfs_refcount_count(
7427 &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7428 uint64_t data_esize =
7429 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7430 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
7431 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
7432 arc_tempreserve >> 10, meta_esize >> 10,
7433 data_esize >> 10, reserve >> 10, arc_c >> 10);
7435 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
7436 return (SET_ERROR(ERESTART));
7438 atomic_add_64(&arc_tempreserve, reserve);
7443 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
7444 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
7446 size->value.ui64 = zfs_refcount_count(&state->arcs_size);
7447 evict_data->value.ui64 =
7448 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
7449 evict_metadata->value.ui64 =
7450 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
7454 arc_kstat_update(kstat_t *ksp, int rw)
7456 arc_stats_t *as = ksp->ks_data;
7458 if (rw == KSTAT_WRITE) {
7459 return (SET_ERROR(EACCES));
7461 arc_kstat_update_state(arc_anon,
7462 &as->arcstat_anon_size,
7463 &as->arcstat_anon_evictable_data,
7464 &as->arcstat_anon_evictable_metadata);
7465 arc_kstat_update_state(arc_mru,
7466 &as->arcstat_mru_size,
7467 &as->arcstat_mru_evictable_data,
7468 &as->arcstat_mru_evictable_metadata);
7469 arc_kstat_update_state(arc_mru_ghost,
7470 &as->arcstat_mru_ghost_size,
7471 &as->arcstat_mru_ghost_evictable_data,
7472 &as->arcstat_mru_ghost_evictable_metadata);
7473 arc_kstat_update_state(arc_mfu,
7474 &as->arcstat_mfu_size,
7475 &as->arcstat_mfu_evictable_data,
7476 &as->arcstat_mfu_evictable_metadata);
7477 arc_kstat_update_state(arc_mfu_ghost,
7478 &as->arcstat_mfu_ghost_size,
7479 &as->arcstat_mfu_ghost_evictable_data,
7480 &as->arcstat_mfu_ghost_evictable_metadata);
7482 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
7483 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
7484 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
7485 ARCSTAT(arcstat_metadata_size) =
7486 aggsum_value(&astat_metadata_size);
7487 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
7488 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
7489 ARCSTAT(arcstat_dbuf_size) = aggsum_value(&astat_dbuf_size);
7490 ARCSTAT(arcstat_dnode_size) = aggsum_value(&astat_dnode_size);
7491 ARCSTAT(arcstat_bonus_size) = aggsum_value(&astat_bonus_size);
7493 as->arcstat_memory_all_bytes.value.ui64 =
7495 as->arcstat_memory_free_bytes.value.ui64 =
7497 as->arcstat_memory_available_bytes.value.i64 =
7498 arc_available_memory();
7505 * This function *must* return indices evenly distributed between all
7506 * sublists of the multilist. This is needed due to how the ARC eviction
7507 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7508 * distributed between all sublists and uses this assumption when
7509 * deciding which sublist to evict from and how much to evict from it.
7512 arc_state_multilist_index_func(multilist_t *ml, void *obj)
7514 arc_buf_hdr_t *hdr = obj;
7517 * We rely on b_dva to generate evenly distributed index
7518 * numbers using buf_hash below. So, as an added precaution,
7519 * let's make sure we never add empty buffers to the arc lists.
7521 ASSERT(!HDR_EMPTY(hdr));
7524 * The assumption here, is the hash value for a given
7525 * arc_buf_hdr_t will remain constant throughout its lifetime
7526 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7527 * Thus, we don't need to store the header's sublist index
7528 * on insertion, as this index can be recalculated on removal.
7530 * Also, the low order bits of the hash value are thought to be
7531 * distributed evenly. Otherwise, in the case that the multilist
7532 * has a power of two number of sublists, each sublists' usage
7533 * would not be evenly distributed.
7535 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
7536 multilist_get_num_sublists(ml));
7540 * Called during module initialization and periodically thereafter to
7541 * apply reasonable changes to the exposed performance tunings. Non-zero
7542 * zfs_* values which differ from the currently set values will be applied.
7545 arc_tuning_update(void)
7547 uint64_t allmem = arc_all_memory();
7548 unsigned long limit;
7550 /* Valid range: 64M - <all physical memory> */
7551 if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) &&
7552 (zfs_arc_max >= 64 << 20) && (zfs_arc_max < allmem) &&
7553 (zfs_arc_max > arc_c_min)) {
7554 arc_c_max = zfs_arc_max;
7556 arc_p = (arc_c >> 1);
7557 if (arc_meta_limit > arc_c_max)
7558 arc_meta_limit = arc_c_max;
7559 if (arc_dnode_limit > arc_meta_limit)
7560 arc_dnode_limit = arc_meta_limit;
7563 /* Valid range: 32M - <arc_c_max> */
7564 if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) &&
7565 (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) &&
7566 (zfs_arc_min <= arc_c_max)) {
7567 arc_c_min = zfs_arc_min;
7568 arc_c = MAX(arc_c, arc_c_min);
7571 /* Valid range: 16M - <arc_c_max> */
7572 if ((zfs_arc_meta_min) && (zfs_arc_meta_min != arc_meta_min) &&
7573 (zfs_arc_meta_min >= 1ULL << SPA_MAXBLOCKSHIFT) &&
7574 (zfs_arc_meta_min <= arc_c_max)) {
7575 arc_meta_min = zfs_arc_meta_min;
7576 if (arc_meta_limit < arc_meta_min)
7577 arc_meta_limit = arc_meta_min;
7578 if (arc_dnode_limit < arc_meta_min)
7579 arc_dnode_limit = arc_meta_min;
7582 /* Valid range: <arc_meta_min> - <arc_c_max> */
7583 limit = zfs_arc_meta_limit ? zfs_arc_meta_limit :
7584 MIN(zfs_arc_meta_limit_percent, 100) * arc_c_max / 100;
7585 if ((limit != arc_meta_limit) &&
7586 (limit >= arc_meta_min) &&
7587 (limit <= arc_c_max))
7588 arc_meta_limit = limit;
7590 /* Valid range: <arc_meta_min> - <arc_meta_limit> */
7591 limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit :
7592 MIN(zfs_arc_dnode_limit_percent, 100) * arc_meta_limit / 100;
7593 if ((limit != arc_dnode_limit) &&
7594 (limit >= arc_meta_min) &&
7595 (limit <= arc_meta_limit))
7596 arc_dnode_limit = limit;
7598 /* Valid range: 1 - N */
7599 if (zfs_arc_grow_retry)
7600 arc_grow_retry = zfs_arc_grow_retry;
7602 /* Valid range: 1 - N */
7603 if (zfs_arc_shrink_shift) {
7604 arc_shrink_shift = zfs_arc_shrink_shift;
7605 arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1);
7608 /* Valid range: 1 - N */
7609 if (zfs_arc_p_min_shift)
7610 arc_p_min_shift = zfs_arc_p_min_shift;
7612 /* Valid range: 1 - N ms */
7613 if (zfs_arc_min_prefetch_ms)
7614 arc_min_prefetch_ms = zfs_arc_min_prefetch_ms;
7616 /* Valid range: 1 - N ms */
7617 if (zfs_arc_min_prescient_prefetch_ms) {
7618 arc_min_prescient_prefetch_ms =
7619 zfs_arc_min_prescient_prefetch_ms;
7622 /* Valid range: 0 - 100 */
7623 if ((zfs_arc_lotsfree_percent >= 0) &&
7624 (zfs_arc_lotsfree_percent <= 100))
7625 arc_lotsfree_percent = zfs_arc_lotsfree_percent;
7627 /* Valid range: 0 - <all physical memory> */
7628 if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free))
7629 arc_sys_free = MIN(MAX(zfs_arc_sys_free, 0), allmem);
7634 arc_state_init(void)
7636 arc_anon = &ARC_anon;
7638 arc_mru_ghost = &ARC_mru_ghost;
7640 arc_mfu_ghost = &ARC_mfu_ghost;
7641 arc_l2c_only = &ARC_l2c_only;
7643 arc_mru->arcs_list[ARC_BUFC_METADATA] =
7644 multilist_create(sizeof (arc_buf_hdr_t),
7645 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7646 arc_state_multilist_index_func);
7647 arc_mru->arcs_list[ARC_BUFC_DATA] =
7648 multilist_create(sizeof (arc_buf_hdr_t),
7649 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7650 arc_state_multilist_index_func);
7651 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
7652 multilist_create(sizeof (arc_buf_hdr_t),
7653 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7654 arc_state_multilist_index_func);
7655 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
7656 multilist_create(sizeof (arc_buf_hdr_t),
7657 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7658 arc_state_multilist_index_func);
7659 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
7660 multilist_create(sizeof (arc_buf_hdr_t),
7661 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7662 arc_state_multilist_index_func);
7663 arc_mfu->arcs_list[ARC_BUFC_DATA] =
7664 multilist_create(sizeof (arc_buf_hdr_t),
7665 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7666 arc_state_multilist_index_func);
7667 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
7668 multilist_create(sizeof (arc_buf_hdr_t),
7669 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7670 arc_state_multilist_index_func);
7671 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
7672 multilist_create(sizeof (arc_buf_hdr_t),
7673 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7674 arc_state_multilist_index_func);
7675 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
7676 multilist_create(sizeof (arc_buf_hdr_t),
7677 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7678 arc_state_multilist_index_func);
7679 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
7680 multilist_create(sizeof (arc_buf_hdr_t),
7681 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7682 arc_state_multilist_index_func);
7684 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7685 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7686 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7687 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7688 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7689 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7690 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7691 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7692 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7693 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7694 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7695 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7697 zfs_refcount_create(&arc_anon->arcs_size);
7698 zfs_refcount_create(&arc_mru->arcs_size);
7699 zfs_refcount_create(&arc_mru_ghost->arcs_size);
7700 zfs_refcount_create(&arc_mfu->arcs_size);
7701 zfs_refcount_create(&arc_mfu_ghost->arcs_size);
7702 zfs_refcount_create(&arc_l2c_only->arcs_size);
7704 aggsum_init(&arc_meta_used, 0);
7705 aggsum_init(&arc_size, 0);
7706 aggsum_init(&astat_data_size, 0);
7707 aggsum_init(&astat_metadata_size, 0);
7708 aggsum_init(&astat_hdr_size, 0);
7709 aggsum_init(&astat_l2_hdr_size, 0);
7710 aggsum_init(&astat_bonus_size, 0);
7711 aggsum_init(&astat_dnode_size, 0);
7712 aggsum_init(&astat_dbuf_size, 0);
7714 arc_anon->arcs_state = ARC_STATE_ANON;
7715 arc_mru->arcs_state = ARC_STATE_MRU;
7716 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7717 arc_mfu->arcs_state = ARC_STATE_MFU;
7718 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7719 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7723 arc_state_fini(void)
7725 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7726 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7727 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7728 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7729 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7730 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7731 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7732 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7733 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7734 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7735 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7736 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7738 zfs_refcount_destroy(&arc_anon->arcs_size);
7739 zfs_refcount_destroy(&arc_mru->arcs_size);
7740 zfs_refcount_destroy(&arc_mru_ghost->arcs_size);
7741 zfs_refcount_destroy(&arc_mfu->arcs_size);
7742 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
7743 zfs_refcount_destroy(&arc_l2c_only->arcs_size);
7745 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
7746 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7747 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7748 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7749 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
7750 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7751 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
7752 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7753 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7754 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7756 aggsum_fini(&arc_meta_used);
7757 aggsum_fini(&arc_size);
7758 aggsum_fini(&astat_data_size);
7759 aggsum_fini(&astat_metadata_size);
7760 aggsum_fini(&astat_hdr_size);
7761 aggsum_fini(&astat_l2_hdr_size);
7762 aggsum_fini(&astat_bonus_size);
7763 aggsum_fini(&astat_dnode_size);
7764 aggsum_fini(&astat_dbuf_size);
7768 arc_target_bytes(void)
7776 uint64_t percent, allmem = arc_all_memory();
7777 mutex_init(&arc_adjust_lock, NULL, MUTEX_DEFAULT, NULL);
7778 cv_init(&arc_adjust_waiters_cv, NULL, CV_DEFAULT, NULL);
7780 arc_min_prefetch_ms = 1000;
7781 arc_min_prescient_prefetch_ms = 6000;
7785 * Register a shrinker to support synchronous (direct) memory
7786 * reclaim from the arc. This is done to prevent kswapd from
7787 * swapping out pages when it is preferable to shrink the arc.
7789 spl_register_shrinker(&arc_shrinker);
7791 /* Set to 1/64 of all memory or a minimum of 512K */
7792 arc_sys_free = MAX(allmem / 64, (512 * 1024));
7796 /* Set max to 1/2 of all memory */
7797 arc_c_max = allmem / 2;
7800 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more */
7801 arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT);
7804 * In userland, there's only the memory pressure that we artificially
7805 * create (see arc_available_memory()). Don't let arc_c get too
7806 * small, because it can cause transactions to be larger than
7807 * arc_c, causing arc_tempreserve_space() to fail.
7809 arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT);
7813 arc_p = (arc_c >> 1);
7815 /* Set min to 1/2 of arc_c_min */
7816 arc_meta_min = 1ULL << SPA_MAXBLOCKSHIFT;
7817 /* Initialize maximum observed usage to zero */
7820 * Set arc_meta_limit to a percent of arc_c_max with a floor of
7821 * arc_meta_min, and a ceiling of arc_c_max.
7823 percent = MIN(zfs_arc_meta_limit_percent, 100);
7824 arc_meta_limit = MAX(arc_meta_min, (percent * arc_c_max) / 100);
7825 percent = MIN(zfs_arc_dnode_limit_percent, 100);
7826 arc_dnode_limit = (percent * arc_meta_limit) / 100;
7828 /* Apply user specified tunings */
7829 arc_tuning_update();
7831 /* if kmem_flags are set, lets try to use less memory */
7832 if (kmem_debugging())
7834 if (arc_c < arc_c_min)
7840 * The arc must be "uninitialized", so that hdr_recl() (which is
7841 * registered by buf_init()) will not access arc_reap_zthr before
7844 ASSERT(!arc_initialized);
7847 list_create(&arc_prune_list, sizeof (arc_prune_t),
7848 offsetof(arc_prune_t, p_node));
7849 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7851 arc_prune_taskq = taskq_create("arc_prune", max_ncpus, defclsyspri,
7852 max_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
7854 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7855 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7857 if (arc_ksp != NULL) {
7858 arc_ksp->ks_data = &arc_stats;
7859 arc_ksp->ks_update = arc_kstat_update;
7860 kstat_install(arc_ksp);
7863 arc_adjust_zthr = zthr_create(arc_adjust_cb_check,
7864 arc_adjust_cb, NULL);
7865 arc_reap_zthr = zthr_create_timer(arc_reap_cb_check,
7866 arc_reap_cb, NULL, SEC2NSEC(1));
7868 arc_initialized = B_TRUE;
7872 * Calculate maximum amount of dirty data per pool.
7874 * If it has been set by a module parameter, take that.
7875 * Otherwise, use a percentage of physical memory defined by
7876 * zfs_dirty_data_max_percent (default 10%) with a cap at
7877 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
7879 if (zfs_dirty_data_max_max == 0)
7880 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
7881 allmem * zfs_dirty_data_max_max_percent / 100);
7883 if (zfs_dirty_data_max == 0) {
7884 zfs_dirty_data_max = allmem *
7885 zfs_dirty_data_max_percent / 100;
7886 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7887 zfs_dirty_data_max_max);
7897 spl_unregister_shrinker(&arc_shrinker);
7898 #endif /* _KERNEL */
7900 /* Use B_TRUE to ensure *all* buffers are evicted */
7901 arc_flush(NULL, B_TRUE);
7903 arc_initialized = B_FALSE;
7905 if (arc_ksp != NULL) {
7906 kstat_delete(arc_ksp);
7910 taskq_wait(arc_prune_taskq);
7911 taskq_destroy(arc_prune_taskq);
7913 mutex_enter(&arc_prune_mtx);
7914 while ((p = list_head(&arc_prune_list)) != NULL) {
7915 list_remove(&arc_prune_list, p);
7916 zfs_refcount_remove(&p->p_refcnt, &arc_prune_list);
7917 zfs_refcount_destroy(&p->p_refcnt);
7918 kmem_free(p, sizeof (*p));
7920 mutex_exit(&arc_prune_mtx);
7922 list_destroy(&arc_prune_list);
7923 mutex_destroy(&arc_prune_mtx);
7924 (void) zthr_cancel(arc_adjust_zthr);
7925 zthr_destroy(arc_adjust_zthr);
7927 (void) zthr_cancel(arc_reap_zthr);
7928 zthr_destroy(arc_reap_zthr);
7930 mutex_destroy(&arc_adjust_lock);
7931 cv_destroy(&arc_adjust_waiters_cv);
7934 * buf_fini() must proceed arc_state_fini() because buf_fin() may
7935 * trigger the release of kmem magazines, which can callback to
7936 * arc_space_return() which accesses aggsums freed in act_state_fini().
7941 ASSERT0(arc_loaned_bytes);
7947 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7948 * It uses dedicated storage devices to hold cached data, which are populated
7949 * using large infrequent writes. The main role of this cache is to boost
7950 * the performance of random read workloads. The intended L2ARC devices
7951 * include short-stroked disks, solid state disks, and other media with
7952 * substantially faster read latency than disk.
7954 * +-----------------------+
7956 * +-----------------------+
7959 * l2arc_feed_thread() arc_read()
7963 * +---------------+ |
7965 * +---------------+ |
7970 * +-------+ +-------+
7972 * | cache | | cache |
7973 * +-------+ +-------+
7974 * +=========+ .-----.
7975 * : L2ARC : |-_____-|
7976 * : devices : | Disks |
7977 * +=========+ `-_____-'
7979 * Read requests are satisfied from the following sources, in order:
7982 * 2) vdev cache of L2ARC devices
7984 * 4) vdev cache of disks
7987 * Some L2ARC device types exhibit extremely slow write performance.
7988 * To accommodate for this there are some significant differences between
7989 * the L2ARC and traditional cache design:
7991 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7992 * the ARC behave as usual, freeing buffers and placing headers on ghost
7993 * lists. The ARC does not send buffers to the L2ARC during eviction as
7994 * this would add inflated write latencies for all ARC memory pressure.
7996 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7997 * It does this by periodically scanning buffers from the eviction-end of
7998 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7999 * not already there. It scans until a headroom of buffers is satisfied,
8000 * which itself is a buffer for ARC eviction. If a compressible buffer is
8001 * found during scanning and selected for writing to an L2ARC device, we
8002 * temporarily boost scanning headroom during the next scan cycle to make
8003 * sure we adapt to compression effects (which might significantly reduce
8004 * the data volume we write to L2ARC). The thread that does this is
8005 * l2arc_feed_thread(), illustrated below; example sizes are included to
8006 * provide a better sense of ratio than this diagram:
8009 * +---------------------+----------+
8010 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
8011 * +---------------------+----------+ | o L2ARC eligible
8012 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
8013 * +---------------------+----------+ |
8014 * 15.9 Gbytes ^ 32 Mbytes |
8016 * l2arc_feed_thread()
8018 * l2arc write hand <--[oooo]--'
8022 * +==============================+
8023 * L2ARC dev |####|#|###|###| |####| ... |
8024 * +==============================+
8027 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
8028 * evicted, then the L2ARC has cached a buffer much sooner than it probably
8029 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
8030 * safe to say that this is an uncommon case, since buffers at the end of
8031 * the ARC lists have moved there due to inactivity.
8033 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
8034 * then the L2ARC simply misses copying some buffers. This serves as a
8035 * pressure valve to prevent heavy read workloads from both stalling the ARC
8036 * with waits and clogging the L2ARC with writes. This also helps prevent
8037 * the potential for the L2ARC to churn if it attempts to cache content too
8038 * quickly, such as during backups of the entire pool.
8040 * 5. After system boot and before the ARC has filled main memory, there are
8041 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
8042 * lists can remain mostly static. Instead of searching from tail of these
8043 * lists as pictured, the l2arc_feed_thread() will search from the list heads
8044 * for eligible buffers, greatly increasing its chance of finding them.
8046 * The L2ARC device write speed is also boosted during this time so that
8047 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
8048 * there are no L2ARC reads, and no fear of degrading read performance
8049 * through increased writes.
8051 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
8052 * the vdev queue can aggregate them into larger and fewer writes. Each
8053 * device is written to in a rotor fashion, sweeping writes through
8054 * available space then repeating.
8056 * 7. The L2ARC does not store dirty content. It never needs to flush
8057 * write buffers back to disk based storage.
8059 * 8. If an ARC buffer is written (and dirtied) which also exists in the
8060 * L2ARC, the now stale L2ARC buffer is immediately dropped.
8062 * The performance of the L2ARC can be tweaked by a number of tunables, which
8063 * may be necessary for different workloads:
8065 * l2arc_write_max max write bytes per interval
8066 * l2arc_write_boost extra write bytes during device warmup
8067 * l2arc_noprefetch skip caching prefetched buffers
8068 * l2arc_headroom number of max device writes to precache
8069 * l2arc_headroom_boost when we find compressed buffers during ARC
8070 * scanning, we multiply headroom by this
8071 * percentage factor for the next scan cycle,
8072 * since more compressed buffers are likely to
8074 * l2arc_feed_secs seconds between L2ARC writing
8076 * Tunables may be removed or added as future performance improvements are
8077 * integrated, and also may become zpool properties.
8079 * There are three key functions that control how the L2ARC warms up:
8081 * l2arc_write_eligible() check if a buffer is eligible to cache
8082 * l2arc_write_size() calculate how much to write
8083 * l2arc_write_interval() calculate sleep delay between writes
8085 * These three functions determine what to write, how much, and how quickly
8090 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
8093 * A buffer is *not* eligible for the L2ARC if it:
8094 * 1. belongs to a different spa.
8095 * 2. is already cached on the L2ARC.
8096 * 3. has an I/O in progress (it may be an incomplete read).
8097 * 4. is flagged not eligible (zfs property).
8099 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
8100 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
8107 l2arc_write_size(void)
8112 * Make sure our globals have meaningful values in case the user
8115 size = l2arc_write_max;
8117 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
8118 "be greater than zero, resetting it to the default (%d)",
8120 size = l2arc_write_max = L2ARC_WRITE_SIZE;
8123 if (arc_warm == B_FALSE)
8124 size += l2arc_write_boost;
8131 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
8133 clock_t interval, next, now;
8136 * If the ARC lists are busy, increase our write rate; if the
8137 * lists are stale, idle back. This is achieved by checking
8138 * how much we previously wrote - if it was more than half of
8139 * what we wanted, schedule the next write much sooner.
8141 if (l2arc_feed_again && wrote > (wanted / 2))
8142 interval = (hz * l2arc_feed_min_ms) / 1000;
8144 interval = hz * l2arc_feed_secs;
8146 now = ddi_get_lbolt();
8147 next = MAX(now, MIN(now + interval, began + interval));
8153 * Cycle through L2ARC devices. This is how L2ARC load balances.
8154 * If a device is returned, this also returns holding the spa config lock.
8156 static l2arc_dev_t *
8157 l2arc_dev_get_next(void)
8159 l2arc_dev_t *first, *next = NULL;
8162 * Lock out the removal of spas (spa_namespace_lock), then removal
8163 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
8164 * both locks will be dropped and a spa config lock held instead.
8166 mutex_enter(&spa_namespace_lock);
8167 mutex_enter(&l2arc_dev_mtx);
8169 /* if there are no vdevs, there is nothing to do */
8170 if (l2arc_ndev == 0)
8174 next = l2arc_dev_last;
8176 /* loop around the list looking for a non-faulted vdev */
8178 next = list_head(l2arc_dev_list);
8180 next = list_next(l2arc_dev_list, next);
8182 next = list_head(l2arc_dev_list);
8185 /* if we have come back to the start, bail out */
8188 else if (next == first)
8191 } while (vdev_is_dead(next->l2ad_vdev));
8193 /* if we were unable to find any usable vdevs, return NULL */
8194 if (vdev_is_dead(next->l2ad_vdev))
8197 l2arc_dev_last = next;
8200 mutex_exit(&l2arc_dev_mtx);
8203 * Grab the config lock to prevent the 'next' device from being
8204 * removed while we are writing to it.
8207 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
8208 mutex_exit(&spa_namespace_lock);
8214 * Free buffers that were tagged for destruction.
8217 l2arc_do_free_on_write(void)
8220 l2arc_data_free_t *df, *df_prev;
8222 mutex_enter(&l2arc_free_on_write_mtx);
8223 buflist = l2arc_free_on_write;
8225 for (df = list_tail(buflist); df; df = df_prev) {
8226 df_prev = list_prev(buflist, df);
8227 ASSERT3P(df->l2df_abd, !=, NULL);
8228 abd_free(df->l2df_abd);
8229 list_remove(buflist, df);
8230 kmem_free(df, sizeof (l2arc_data_free_t));
8233 mutex_exit(&l2arc_free_on_write_mtx);
8237 * A write to a cache device has completed. Update all headers to allow
8238 * reads from these buffers to begin.
8241 l2arc_write_done(zio_t *zio)
8243 l2arc_write_callback_t *cb;
8246 arc_buf_hdr_t *head, *hdr, *hdr_prev;
8247 kmutex_t *hash_lock;
8248 int64_t bytes_dropped = 0;
8250 cb = zio->io_private;
8251 ASSERT3P(cb, !=, NULL);
8252 dev = cb->l2wcb_dev;
8253 ASSERT3P(dev, !=, NULL);
8254 head = cb->l2wcb_head;
8255 ASSERT3P(head, !=, NULL);
8256 buflist = &dev->l2ad_buflist;
8257 ASSERT3P(buflist, !=, NULL);
8258 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
8259 l2arc_write_callback_t *, cb);
8261 if (zio->io_error != 0)
8262 ARCSTAT_BUMP(arcstat_l2_writes_error);
8265 * All writes completed, or an error was hit.
8268 mutex_enter(&dev->l2ad_mtx);
8269 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
8270 hdr_prev = list_prev(buflist, hdr);
8272 hash_lock = HDR_LOCK(hdr);
8275 * We cannot use mutex_enter or else we can deadlock
8276 * with l2arc_write_buffers (due to swapping the order
8277 * the hash lock and l2ad_mtx are taken).
8279 if (!mutex_tryenter(hash_lock)) {
8281 * Missed the hash lock. We must retry so we
8282 * don't leave the ARC_FLAG_L2_WRITING bit set.
8284 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
8287 * We don't want to rescan the headers we've
8288 * already marked as having been written out, so
8289 * we reinsert the head node so we can pick up
8290 * where we left off.
8292 list_remove(buflist, head);
8293 list_insert_after(buflist, hdr, head);
8295 mutex_exit(&dev->l2ad_mtx);
8298 * We wait for the hash lock to become available
8299 * to try and prevent busy waiting, and increase
8300 * the chance we'll be able to acquire the lock
8301 * the next time around.
8303 mutex_enter(hash_lock);
8304 mutex_exit(hash_lock);
8309 * We could not have been moved into the arc_l2c_only
8310 * state while in-flight due to our ARC_FLAG_L2_WRITING
8311 * bit being set. Let's just ensure that's being enforced.
8313 ASSERT(HDR_HAS_L1HDR(hdr));
8316 * Skipped - drop L2ARC entry and mark the header as no
8317 * longer L2 eligibile.
8319 if (zio->io_error != 0) {
8321 * Error - drop L2ARC entry.
8323 list_remove(buflist, hdr);
8324 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
8326 uint64_t psize = HDR_GET_PSIZE(hdr);
8327 ARCSTAT_INCR(arcstat_l2_psize, -psize);
8328 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
8331 vdev_psize_to_asize(dev->l2ad_vdev, psize);
8332 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
8333 arc_hdr_size(hdr), hdr);
8337 * Allow ARC to begin reads and ghost list evictions to
8340 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
8342 mutex_exit(hash_lock);
8345 atomic_inc_64(&l2arc_writes_done);
8346 list_remove(buflist, head);
8347 ASSERT(!HDR_HAS_L1HDR(head));
8348 kmem_cache_free(hdr_l2only_cache, head);
8349 mutex_exit(&dev->l2ad_mtx);
8351 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
8353 l2arc_do_free_on_write();
8355 kmem_free(cb, sizeof (l2arc_write_callback_t));
8359 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
8362 spa_t *spa = zio->io_spa;
8363 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
8364 blkptr_t *bp = zio->io_bp;
8365 uint8_t salt[ZIO_DATA_SALT_LEN];
8366 uint8_t iv[ZIO_DATA_IV_LEN];
8367 uint8_t mac[ZIO_DATA_MAC_LEN];
8368 boolean_t no_crypt = B_FALSE;
8371 * ZIL data is never be written to the L2ARC, so we don't need
8372 * special handling for its unique MAC storage.
8374 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
8375 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
8376 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8379 * If the data was encrypted, decrypt it now. Note that
8380 * we must check the bp here and not the hdr, since the
8381 * hdr does not have its encryption parameters updated
8382 * until arc_read_done().
8384 if (BP_IS_ENCRYPTED(bp)) {
8385 abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
8387 zio_crypt_decode_params_bp(bp, salt, iv);
8388 zio_crypt_decode_mac_bp(bp, mac);
8390 ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
8391 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
8392 salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
8393 hdr->b_l1hdr.b_pabd, &no_crypt);
8395 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8400 * If we actually performed decryption, replace b_pabd
8401 * with the decrypted data. Otherwise we can just throw
8402 * our decryption buffer away.
8405 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8406 arc_hdr_size(hdr), hdr);
8407 hdr->b_l1hdr.b_pabd = eabd;
8410 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8415 * If the L2ARC block was compressed, but ARC compression
8416 * is disabled we decompress the data into a new buffer and
8417 * replace the existing data.
8419 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8420 !HDR_COMPRESSION_ENABLED(hdr)) {
8421 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
8422 void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
8424 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
8425 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
8426 HDR_GET_LSIZE(hdr));
8428 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8429 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
8433 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8434 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8435 arc_hdr_size(hdr), hdr);
8436 hdr->b_l1hdr.b_pabd = cabd;
8438 zio->io_size = HDR_GET_LSIZE(hdr);
8449 * A read to a cache device completed. Validate buffer contents before
8450 * handing over to the regular ARC routines.
8453 l2arc_read_done(zio_t *zio)
8456 l2arc_read_callback_t *cb = zio->io_private;
8458 kmutex_t *hash_lock;
8459 boolean_t valid_cksum;
8460 boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
8461 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
8463 ASSERT3P(zio->io_vd, !=, NULL);
8464 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
8466 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
8468 ASSERT3P(cb, !=, NULL);
8469 hdr = cb->l2rcb_hdr;
8470 ASSERT3P(hdr, !=, NULL);
8472 hash_lock = HDR_LOCK(hdr);
8473 mutex_enter(hash_lock);
8474 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
8477 * If the data was read into a temporary buffer,
8478 * move it and free the buffer.
8480 if (cb->l2rcb_abd != NULL) {
8481 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
8482 if (zio->io_error == 0) {
8484 abd_copy(hdr->b_crypt_hdr.b_rabd,
8485 cb->l2rcb_abd, arc_hdr_size(hdr));
8487 abd_copy(hdr->b_l1hdr.b_pabd,
8488 cb->l2rcb_abd, arc_hdr_size(hdr));
8493 * The following must be done regardless of whether
8494 * there was an error:
8495 * - free the temporary buffer
8496 * - point zio to the real ARC buffer
8497 * - set zio size accordingly
8498 * These are required because zio is either re-used for
8499 * an I/O of the block in the case of the error
8500 * or the zio is passed to arc_read_done() and it
8503 abd_free(cb->l2rcb_abd);
8504 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
8507 ASSERT(HDR_HAS_RABD(hdr));
8508 zio->io_abd = zio->io_orig_abd =
8509 hdr->b_crypt_hdr.b_rabd;
8511 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8512 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8516 ASSERT3P(zio->io_abd, !=, NULL);
8519 * Check this survived the L2ARC journey.
8521 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8522 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8523 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
8524 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
8526 valid_cksum = arc_cksum_is_equal(hdr, zio);
8529 * b_rabd will always match the data as it exists on disk if it is
8530 * being used. Therefore if we are reading into b_rabd we do not
8531 * attempt to untransform the data.
8533 if (valid_cksum && !using_rdata)
8534 tfm_error = l2arc_untransform(zio, cb);
8536 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8537 !HDR_L2_EVICTED(hdr)) {
8538 mutex_exit(hash_lock);
8539 zio->io_private = hdr;
8542 mutex_exit(hash_lock);
8544 * Buffer didn't survive caching. Increment stats and
8545 * reissue to the original storage device.
8547 if (zio->io_error != 0) {
8548 ARCSTAT_BUMP(arcstat_l2_io_error);
8550 zio->io_error = SET_ERROR(EIO);
8552 if (!valid_cksum || tfm_error != 0)
8553 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8556 * If there's no waiter, issue an async i/o to the primary
8557 * storage now. If there *is* a waiter, the caller must
8558 * issue the i/o in a context where it's OK to block.
8560 if (zio->io_waiter == NULL) {
8561 zio_t *pio = zio_unique_parent(zio);
8562 void *abd = (using_rdata) ?
8563 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8565 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8567 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
8568 abd, zio->io_size, arc_read_done,
8569 hdr, zio->io_priority, cb->l2rcb_flags,
8574 kmem_free(cb, sizeof (l2arc_read_callback_t));
8578 * This is the list priority from which the L2ARC will search for pages to
8579 * cache. This is used within loops (0..3) to cycle through lists in the
8580 * desired order. This order can have a significant effect on cache
8583 * Currently the metadata lists are hit first, MFU then MRU, followed by
8584 * the data lists. This function returns a locked list, and also returns
8587 static multilist_sublist_t *
8588 l2arc_sublist_lock(int list_num)
8590 multilist_t *ml = NULL;
8593 ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
8597 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
8600 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
8603 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
8606 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
8613 * Return a randomly-selected sublist. This is acceptable
8614 * because the caller feeds only a little bit of data for each
8615 * call (8MB). Subsequent calls will result in different
8616 * sublists being selected.
8618 idx = multilist_get_random_index(ml);
8619 return (multilist_sublist_lock(ml, idx));
8623 * Evict buffers from the device write hand to the distance specified in
8624 * bytes. This distance may span populated buffers, it may span nothing.
8625 * This is clearing a region on the L2ARC device ready for writing.
8626 * If the 'all' boolean is set, every buffer is evicted.
8629 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
8632 arc_buf_hdr_t *hdr, *hdr_prev;
8633 kmutex_t *hash_lock;
8636 buflist = &dev->l2ad_buflist;
8638 if (!all && dev->l2ad_first) {
8640 * This is the first sweep through the device. There is
8646 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
8648 * When nearing the end of the device, evict to the end
8649 * before the device write hand jumps to the start.
8651 taddr = dev->l2ad_end;
8653 taddr = dev->l2ad_hand + distance;
8655 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
8656 uint64_t, taddr, boolean_t, all);
8659 mutex_enter(&dev->l2ad_mtx);
8660 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
8661 hdr_prev = list_prev(buflist, hdr);
8663 hash_lock = HDR_LOCK(hdr);
8666 * We cannot use mutex_enter or else we can deadlock
8667 * with l2arc_write_buffers (due to swapping the order
8668 * the hash lock and l2ad_mtx are taken).
8670 if (!mutex_tryenter(hash_lock)) {
8672 * Missed the hash lock. Retry.
8674 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
8675 mutex_exit(&dev->l2ad_mtx);
8676 mutex_enter(hash_lock);
8677 mutex_exit(hash_lock);
8682 * A header can't be on this list if it doesn't have L2 header.
8684 ASSERT(HDR_HAS_L2HDR(hdr));
8686 /* Ensure this header has finished being written. */
8687 ASSERT(!HDR_L2_WRITING(hdr));
8688 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
8690 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
8691 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
8693 * We've evicted to the target address,
8694 * or the end of the device.
8696 mutex_exit(hash_lock);
8700 if (!HDR_HAS_L1HDR(hdr)) {
8701 ASSERT(!HDR_L2_READING(hdr));
8703 * This doesn't exist in the ARC. Destroy.
8704 * arc_hdr_destroy() will call list_remove()
8705 * and decrement arcstat_l2_lsize.
8707 arc_change_state(arc_anon, hdr, hash_lock);
8708 arc_hdr_destroy(hdr);
8710 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
8711 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
8713 * Invalidate issued or about to be issued
8714 * reads, since we may be about to write
8715 * over this location.
8717 if (HDR_L2_READING(hdr)) {
8718 ARCSTAT_BUMP(arcstat_l2_evict_reading);
8719 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
8722 arc_hdr_l2hdr_destroy(hdr);
8724 mutex_exit(hash_lock);
8726 mutex_exit(&dev->l2ad_mtx);
8730 * Handle any abd transforms that might be required for writing to the L2ARC.
8731 * If successful, this function will always return an abd with the data
8732 * transformed as it is on disk in a new abd of asize bytes.
8735 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
8740 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
8741 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
8742 uint64_t psize = HDR_GET_PSIZE(hdr);
8743 uint64_t size = arc_hdr_size(hdr);
8744 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
8745 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
8746 dsl_crypto_key_t *dck = NULL;
8747 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
8748 boolean_t no_crypt = B_FALSE;
8750 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8751 !HDR_COMPRESSION_ENABLED(hdr)) ||
8752 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
8753 ASSERT3U(psize, <=, asize);
8756 * If this data simply needs its own buffer, we simply allocate it
8757 * and copy the data. This may be done to elimiate a depedency on a
8758 * shared buffer or to reallocate the buffer to match asize.
8760 if (HDR_HAS_RABD(hdr) && asize != psize) {
8761 ASSERT3U(asize, >=, psize);
8762 to_write = abd_alloc_for_io(asize, ismd);
8763 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
8765 abd_zero_off(to_write, psize, asize - psize);
8769 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
8770 !HDR_ENCRYPTED(hdr)) {
8771 ASSERT3U(size, ==, psize);
8772 to_write = abd_alloc_for_io(asize, ismd);
8773 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
8775 abd_zero_off(to_write, size, asize - size);
8779 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
8780 cabd = abd_alloc_for_io(asize, ismd);
8781 tmp = abd_borrow_buf(cabd, asize);
8783 psize = zio_compress_data(compress, to_write, tmp, size);
8784 ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
8786 bzero((char *)tmp + psize, asize - psize);
8787 psize = HDR_GET_PSIZE(hdr);
8788 abd_return_buf_copy(cabd, tmp, asize);
8792 if (HDR_ENCRYPTED(hdr)) {
8793 eabd = abd_alloc_for_io(asize, ismd);
8796 * If the dataset was disowned before the buffer
8797 * made it to this point, the key to re-encrypt
8798 * it won't be available. In this case we simply
8799 * won't write the buffer to the L2ARC.
8801 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
8806 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
8807 hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
8808 hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
8814 abd_copy(eabd, to_write, psize);
8817 abd_zero_off(eabd, psize, asize - psize);
8819 /* assert that the MAC we got here matches the one we saved */
8820 ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
8821 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8823 if (to_write == cabd)
8830 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
8831 *abd_out = to_write;
8836 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8847 * Find and write ARC buffers to the L2ARC device.
8849 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
8850 * for reading until they have completed writing.
8851 * The headroom_boost is an in-out parameter used to maintain headroom boost
8852 * state between calls to this function.
8854 * Returns the number of bytes actually written (which may be smaller than
8855 * the delta by which the device hand has changed due to alignment).
8858 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
8860 arc_buf_hdr_t *hdr, *hdr_prev, *head;
8861 uint64_t write_asize, write_psize, write_lsize, headroom;
8863 l2arc_write_callback_t *cb;
8865 uint64_t guid = spa_load_guid(spa);
8867 ASSERT3P(dev->l2ad_vdev, !=, NULL);
8870 write_lsize = write_asize = write_psize = 0;
8872 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
8873 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
8876 * Copy buffers for L2ARC writing.
8878 for (int try = 0; try < L2ARC_FEED_TYPES; try++) {
8879 multilist_sublist_t *mls = l2arc_sublist_lock(try);
8880 uint64_t passed_sz = 0;
8882 VERIFY3P(mls, !=, NULL);
8885 * L2ARC fast warmup.
8887 * Until the ARC is warm and starts to evict, read from the
8888 * head of the ARC lists rather than the tail.
8890 if (arc_warm == B_FALSE)
8891 hdr = multilist_sublist_head(mls);
8893 hdr = multilist_sublist_tail(mls);
8895 headroom = target_sz * l2arc_headroom;
8896 if (zfs_compressed_arc_enabled)
8897 headroom = (headroom * l2arc_headroom_boost) / 100;
8899 for (; hdr; hdr = hdr_prev) {
8900 kmutex_t *hash_lock;
8901 abd_t *to_write = NULL;
8903 if (arc_warm == B_FALSE)
8904 hdr_prev = multilist_sublist_next(mls, hdr);
8906 hdr_prev = multilist_sublist_prev(mls, hdr);
8908 hash_lock = HDR_LOCK(hdr);
8909 if (!mutex_tryenter(hash_lock)) {
8911 * Skip this buffer rather than waiting.
8916 passed_sz += HDR_GET_LSIZE(hdr);
8917 if (passed_sz > headroom) {
8921 mutex_exit(hash_lock);
8925 if (!l2arc_write_eligible(guid, hdr)) {
8926 mutex_exit(hash_lock);
8931 * We rely on the L1 portion of the header below, so
8932 * it's invalid for this header to have been evicted out
8933 * of the ghost cache, prior to being written out. The
8934 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8936 ASSERT(HDR_HAS_L1HDR(hdr));
8938 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8939 ASSERT3U(arc_hdr_size(hdr), >, 0);
8940 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
8942 uint64_t psize = HDR_GET_PSIZE(hdr);
8943 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
8946 if ((write_asize + asize) > target_sz) {
8948 mutex_exit(hash_lock);
8953 * We rely on the L1 portion of the header below, so
8954 * it's invalid for this header to have been evicted out
8955 * of the ghost cache, prior to being written out. The
8956 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8958 arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
8959 ASSERT(HDR_HAS_L1HDR(hdr));
8961 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8962 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
8964 ASSERT3U(arc_hdr_size(hdr), >, 0);
8967 * If this header has b_rabd, we can use this since it
8968 * must always match the data exactly as it exists on
8969 * disk. Otherwise, the L2ARC can normally use the
8970 * hdr's data, but if we're sharing data between the
8971 * hdr and one of its bufs, L2ARC needs its own copy of
8972 * the data so that the ZIO below can't race with the
8973 * buf consumer. To ensure that this copy will be
8974 * available for the lifetime of the ZIO and be cleaned
8975 * up afterwards, we add it to the l2arc_free_on_write
8976 * queue. If we need to apply any transforms to the
8977 * data (compression, encryption) we will also need the
8980 if (HDR_HAS_RABD(hdr) && psize == asize) {
8981 to_write = hdr->b_crypt_hdr.b_rabd;
8982 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
8983 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
8984 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
8986 to_write = hdr->b_l1hdr.b_pabd;
8989 arc_buf_contents_t type = arc_buf_type(hdr);
8991 ret = l2arc_apply_transforms(spa, hdr, asize,
8994 arc_hdr_clear_flags(hdr,
8995 ARC_FLAG_L2_WRITING);
8996 mutex_exit(hash_lock);
9000 l2arc_free_abd_on_write(to_write, asize, type);
9005 * Insert a dummy header on the buflist so
9006 * l2arc_write_done() can find where the
9007 * write buffers begin without searching.
9009 mutex_enter(&dev->l2ad_mtx);
9010 list_insert_head(&dev->l2ad_buflist, head);
9011 mutex_exit(&dev->l2ad_mtx);
9014 sizeof (l2arc_write_callback_t), KM_SLEEP);
9015 cb->l2wcb_dev = dev;
9016 cb->l2wcb_head = head;
9017 pio = zio_root(spa, l2arc_write_done, cb,
9021 hdr->b_l2hdr.b_dev = dev;
9022 hdr->b_l2hdr.b_hits = 0;
9024 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
9025 arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR);
9027 mutex_enter(&dev->l2ad_mtx);
9028 list_insert_head(&dev->l2ad_buflist, hdr);
9029 mutex_exit(&dev->l2ad_mtx);
9031 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
9032 arc_hdr_size(hdr), hdr);
9034 wzio = zio_write_phys(pio, dev->l2ad_vdev,
9035 hdr->b_l2hdr.b_daddr, asize, to_write,
9036 ZIO_CHECKSUM_OFF, NULL, hdr,
9037 ZIO_PRIORITY_ASYNC_WRITE,
9038 ZIO_FLAG_CANFAIL, B_FALSE);
9040 write_lsize += HDR_GET_LSIZE(hdr);
9041 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
9044 write_psize += psize;
9045 write_asize += asize;
9046 dev->l2ad_hand += asize;
9047 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9049 mutex_exit(hash_lock);
9051 (void) zio_nowait(wzio);
9054 multilist_sublist_unlock(mls);
9060 /* No buffers selected for writing? */
9062 ASSERT0(write_lsize);
9063 ASSERT(!HDR_HAS_L1HDR(head));
9064 kmem_cache_free(hdr_l2only_cache, head);
9068 ASSERT3U(write_asize, <=, target_sz);
9069 ARCSTAT_BUMP(arcstat_l2_writes_sent);
9070 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
9071 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
9072 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
9075 * Bump device hand to the device start if it is approaching the end.
9076 * l2arc_evict() will already have evicted ahead for this case.
9078 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
9079 dev->l2ad_hand = dev->l2ad_start;
9080 dev->l2ad_first = B_FALSE;
9083 dev->l2ad_writing = B_TRUE;
9084 (void) zio_wait(pio);
9085 dev->l2ad_writing = B_FALSE;
9087 return (write_asize);
9091 * This thread feeds the L2ARC at regular intervals. This is the beating
9092 * heart of the L2ARC.
9096 l2arc_feed_thread(void *unused)
9101 uint64_t size, wrote;
9102 clock_t begin, next = ddi_get_lbolt();
9103 fstrans_cookie_t cookie;
9105 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
9107 mutex_enter(&l2arc_feed_thr_lock);
9109 cookie = spl_fstrans_mark();
9110 while (l2arc_thread_exit == 0) {
9111 CALLB_CPR_SAFE_BEGIN(&cpr);
9112 (void) cv_timedwait_sig(&l2arc_feed_thr_cv,
9113 &l2arc_feed_thr_lock, next);
9114 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
9115 next = ddi_get_lbolt() + hz;
9118 * Quick check for L2ARC devices.
9120 mutex_enter(&l2arc_dev_mtx);
9121 if (l2arc_ndev == 0) {
9122 mutex_exit(&l2arc_dev_mtx);
9125 mutex_exit(&l2arc_dev_mtx);
9126 begin = ddi_get_lbolt();
9129 * This selects the next l2arc device to write to, and in
9130 * doing so the next spa to feed from: dev->l2ad_spa. This
9131 * will return NULL if there are now no l2arc devices or if
9132 * they are all faulted.
9134 * If a device is returned, its spa's config lock is also
9135 * held to prevent device removal. l2arc_dev_get_next()
9136 * will grab and release l2arc_dev_mtx.
9138 if ((dev = l2arc_dev_get_next()) == NULL)
9141 spa = dev->l2ad_spa;
9142 ASSERT3P(spa, !=, NULL);
9145 * If the pool is read-only then force the feed thread to
9146 * sleep a little longer.
9148 if (!spa_writeable(spa)) {
9149 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
9150 spa_config_exit(spa, SCL_L2ARC, dev);
9155 * Avoid contributing to memory pressure.
9157 if (arc_reclaim_needed()) {
9158 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
9159 spa_config_exit(spa, SCL_L2ARC, dev);
9163 ARCSTAT_BUMP(arcstat_l2_feeds);
9165 size = l2arc_write_size();
9168 * Evict L2ARC buffers that will be overwritten.
9170 l2arc_evict(dev, size, B_FALSE);
9173 * Write ARC buffers.
9175 wrote = l2arc_write_buffers(spa, dev, size);
9178 * Calculate interval between writes.
9180 next = l2arc_write_interval(begin, size, wrote);
9181 spa_config_exit(spa, SCL_L2ARC, dev);
9183 spl_fstrans_unmark(cookie);
9185 l2arc_thread_exit = 0;
9186 cv_broadcast(&l2arc_feed_thr_cv);
9187 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
9192 l2arc_vdev_present(vdev_t *vd)
9196 mutex_enter(&l2arc_dev_mtx);
9197 for (dev = list_head(l2arc_dev_list); dev != NULL;
9198 dev = list_next(l2arc_dev_list, dev)) {
9199 if (dev->l2ad_vdev == vd)
9202 mutex_exit(&l2arc_dev_mtx);
9204 return (dev != NULL);
9208 * Add a vdev for use by the L2ARC. By this point the spa has already
9209 * validated the vdev and opened it.
9212 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
9214 l2arc_dev_t *adddev;
9216 ASSERT(!l2arc_vdev_present(vd));
9219 * Create a new l2arc device entry.
9221 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
9222 adddev->l2ad_spa = spa;
9223 adddev->l2ad_vdev = vd;
9224 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
9225 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
9226 adddev->l2ad_hand = adddev->l2ad_start;
9227 adddev->l2ad_first = B_TRUE;
9228 adddev->l2ad_writing = B_FALSE;
9229 list_link_init(&adddev->l2ad_node);
9231 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
9233 * This is a list of all ARC buffers that are still valid on the
9236 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
9237 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
9239 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
9240 zfs_refcount_create(&adddev->l2ad_alloc);
9243 * Add device to global list
9245 mutex_enter(&l2arc_dev_mtx);
9246 list_insert_head(l2arc_dev_list, adddev);
9247 atomic_inc_64(&l2arc_ndev);
9248 mutex_exit(&l2arc_dev_mtx);
9252 * Remove a vdev from the L2ARC.
9255 l2arc_remove_vdev(vdev_t *vd)
9257 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
9260 * Find the device by vdev
9262 mutex_enter(&l2arc_dev_mtx);
9263 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
9264 nextdev = list_next(l2arc_dev_list, dev);
9265 if (vd == dev->l2ad_vdev) {
9270 ASSERT3P(remdev, !=, NULL);
9273 * Remove device from global list
9275 list_remove(l2arc_dev_list, remdev);
9276 l2arc_dev_last = NULL; /* may have been invalidated */
9277 atomic_dec_64(&l2arc_ndev);
9278 mutex_exit(&l2arc_dev_mtx);
9281 * Clear all buflists and ARC references. L2ARC device flush.
9283 l2arc_evict(remdev, 0, B_TRUE);
9284 list_destroy(&remdev->l2ad_buflist);
9285 mutex_destroy(&remdev->l2ad_mtx);
9286 zfs_refcount_destroy(&remdev->l2ad_alloc);
9287 kmem_free(remdev, sizeof (l2arc_dev_t));
9293 l2arc_thread_exit = 0;
9295 l2arc_writes_sent = 0;
9296 l2arc_writes_done = 0;
9298 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9299 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
9300 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
9301 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
9303 l2arc_dev_list = &L2ARC_dev_list;
9304 l2arc_free_on_write = &L2ARC_free_on_write;
9305 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
9306 offsetof(l2arc_dev_t, l2ad_node));
9307 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
9308 offsetof(l2arc_data_free_t, l2df_list_node));
9315 * This is called from dmu_fini(), which is called from spa_fini();
9316 * Because of this, we can assume that all l2arc devices have
9317 * already been removed when the pools themselves were removed.
9320 l2arc_do_free_on_write();
9322 mutex_destroy(&l2arc_feed_thr_lock);
9323 cv_destroy(&l2arc_feed_thr_cv);
9324 mutex_destroy(&l2arc_dev_mtx);
9325 mutex_destroy(&l2arc_free_on_write_mtx);
9327 list_destroy(l2arc_dev_list);
9328 list_destroy(l2arc_free_on_write);
9334 if (!(spa_mode_global & FWRITE))
9337 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
9338 TS_RUN, defclsyspri);
9344 if (!(spa_mode_global & FWRITE))
9347 mutex_enter(&l2arc_feed_thr_lock);
9348 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
9349 l2arc_thread_exit = 1;
9350 while (l2arc_thread_exit != 0)
9351 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
9352 mutex_exit(&l2arc_feed_thr_lock);
9355 #if defined(_KERNEL)
9356 EXPORT_SYMBOL(arc_buf_size);
9357 EXPORT_SYMBOL(arc_write);
9358 EXPORT_SYMBOL(arc_read);
9359 EXPORT_SYMBOL(arc_buf_info);
9360 EXPORT_SYMBOL(arc_getbuf_func);
9361 EXPORT_SYMBOL(arc_add_prune_callback);
9362 EXPORT_SYMBOL(arc_remove_prune_callback);
9365 module_param(zfs_arc_min, ulong, 0644);
9366 MODULE_PARM_DESC(zfs_arc_min, "Min arc size");
9368 module_param(zfs_arc_max, ulong, 0644);
9369 MODULE_PARM_DESC(zfs_arc_max, "Max arc size");
9371 module_param(zfs_arc_meta_limit, ulong, 0644);
9372 MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size");
9374 module_param(zfs_arc_meta_limit_percent, ulong, 0644);
9375 MODULE_PARM_DESC(zfs_arc_meta_limit_percent,
9376 "Percent of arc size for arc meta limit");
9378 module_param(zfs_arc_meta_min, ulong, 0644);
9379 MODULE_PARM_DESC(zfs_arc_meta_min, "Min arc metadata");
9381 module_param(zfs_arc_meta_prune, int, 0644);
9382 MODULE_PARM_DESC(zfs_arc_meta_prune, "Meta objects to scan for prune");
9384 module_param(zfs_arc_meta_adjust_restarts, int, 0644);
9385 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts,
9386 "Limit number of restarts in arc_adjust_meta");
9388 module_param(zfs_arc_meta_strategy, int, 0644);
9389 MODULE_PARM_DESC(zfs_arc_meta_strategy, "Meta reclaim strategy");
9391 module_param(zfs_arc_grow_retry, int, 0644);
9392 MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size");
9394 module_param(zfs_arc_p_dampener_disable, int, 0644);
9395 MODULE_PARM_DESC(zfs_arc_p_dampener_disable, "disable arc_p adapt dampener");
9397 module_param(zfs_arc_shrink_shift, int, 0644);
9398 MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)");
9400 module_param(zfs_arc_pc_percent, uint, 0644);
9401 MODULE_PARM_DESC(zfs_arc_pc_percent,
9402 "Percent of pagecache to reclaim arc to");
9404 module_param(zfs_arc_p_min_shift, int, 0644);
9405 MODULE_PARM_DESC(zfs_arc_p_min_shift, "arc_c shift to calc min/max arc_p");
9407 module_param(zfs_arc_average_blocksize, int, 0444);
9408 MODULE_PARM_DESC(zfs_arc_average_blocksize, "Target average block size");
9410 module_param(zfs_compressed_arc_enabled, int, 0644);
9411 MODULE_PARM_DESC(zfs_compressed_arc_enabled, "Disable compressed arc buffers");
9413 module_param(zfs_arc_min_prefetch_ms, int, 0644);
9414 MODULE_PARM_DESC(zfs_arc_min_prefetch_ms, "Min life of prefetch block in ms");
9416 module_param(zfs_arc_min_prescient_prefetch_ms, int, 0644);
9417 MODULE_PARM_DESC(zfs_arc_min_prescient_prefetch_ms,
9418 "Min life of prescient prefetched block in ms");
9420 module_param(l2arc_write_max, ulong, 0644);
9421 MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval");
9423 module_param(l2arc_write_boost, ulong, 0644);
9424 MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup");
9426 module_param(l2arc_headroom, ulong, 0644);
9427 MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache");
9429 module_param(l2arc_headroom_boost, ulong, 0644);
9430 MODULE_PARM_DESC(l2arc_headroom_boost, "Compressed l2arc_headroom multiplier");
9432 module_param(l2arc_feed_secs, ulong, 0644);
9433 MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing");
9435 module_param(l2arc_feed_min_ms, ulong, 0644);
9436 MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds");
9438 module_param(l2arc_noprefetch, int, 0644);
9439 MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers");
9441 module_param(l2arc_feed_again, int, 0644);
9442 MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup");
9444 module_param(l2arc_norw, int, 0644);
9445 MODULE_PARM_DESC(l2arc_norw, "No reads during writes");
9447 module_param(zfs_arc_lotsfree_percent, int, 0644);
9448 MODULE_PARM_DESC(zfs_arc_lotsfree_percent,
9449 "System free memory I/O throttle in bytes");
9451 module_param(zfs_arc_sys_free, ulong, 0644);
9452 MODULE_PARM_DESC(zfs_arc_sys_free, "System free memory target size in bytes");
9454 module_param(zfs_arc_dnode_limit, ulong, 0644);
9455 MODULE_PARM_DESC(zfs_arc_dnode_limit, "Minimum bytes of dnodes in arc");
9457 module_param(zfs_arc_dnode_limit_percent, ulong, 0644);
9458 MODULE_PARM_DESC(zfs_arc_dnode_limit_percent,
9459 "Percent of ARC meta buffers for dnodes");
9461 module_param(zfs_arc_dnode_reduce_percent, ulong, 0644);
9462 MODULE_PARM_DESC(zfs_arc_dnode_reduce_percent,
9463 "Percentage of excess dnodes to try to unpin");
projects / FreeBSD / FreeBSD.git / blob