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, 2020, Delphix. All rights reserved.
25 * Copyright (c) 2014, Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2017, Nexenta Systems, Inc. All rights reserved.
27 * Copyright (c) 2019, loli10K <ezomori.nozomu@gmail.com>. All rights reserved.
28 * Copyright (c) 2020, George Amanakis. All rights reserved.
29 * Copyright (c) 2019, Klara Inc.
30 * Copyright (c) 2019, Allan Jude
31 * Copyright (c) 2020, The FreeBSD Foundation [1]
33 * [1] Portions of this software were developed by Allan Jude
34 * under sponsorship from the FreeBSD Foundation.
38 * DVA-based Adjustable Replacement Cache
40 * While much of the theory of operation used here is
41 * based on the self-tuning, low overhead replacement cache
42 * presented by Megiddo and Modha at FAST 2003, there are some
43 * significant differences:
45 * 1. The Megiddo and Modha model assumes any page is evictable.
46 * Pages in its cache cannot be "locked" into memory. This makes
47 * the eviction algorithm simple: evict the last page in the list.
48 * This also make the performance characteristics easy to reason
49 * about. Our cache is not so simple. At any given moment, some
50 * subset of the blocks in the cache are un-evictable because we
51 * have handed out a reference to them. Blocks are only evictable
52 * when there are no external references active. This makes
53 * eviction far more problematic: we choose to evict the evictable
54 * blocks that are the "lowest" in the list.
56 * There are times when it is not possible to evict the requested
57 * space. In these circumstances we are unable to adjust the cache
58 * size. To prevent the cache growing unbounded at these times we
59 * implement a "cache throttle" that slows the flow of new data
60 * into the cache until we can make space available.
62 * 2. The Megiddo and Modha model assumes a fixed cache size.
63 * Pages are evicted when the cache is full and there is a cache
64 * miss. Our model has a variable sized cache. It grows with
65 * high use, but also tries to react to memory pressure from the
66 * operating system: decreasing its size when system memory is
69 * 3. The Megiddo and Modha model assumes a fixed page size. All
70 * elements of the cache are therefore exactly the same size. So
71 * when adjusting the cache size following a cache miss, its simply
72 * a matter of choosing a single page to evict. In our model, we
73 * have variable sized cache blocks (ranging from 512 bytes to
74 * 128K bytes). We therefore choose a set of blocks to evict to make
75 * space for a cache miss that approximates as closely as possible
76 * the space used by the new block.
78 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
79 * by N. Megiddo & D. Modha, FAST 2003
85 * A new reference to a cache buffer can be obtained in two
86 * ways: 1) via a hash table lookup using the DVA as a key,
87 * or 2) via one of the ARC lists. The arc_read() interface
88 * uses method 1, while the internal ARC algorithms for
89 * adjusting the cache use method 2. We therefore provide two
90 * types of locks: 1) the hash table lock array, and 2) the
93 * Buffers do not have their own mutexes, rather they rely on the
94 * hash table mutexes for the bulk of their protection (i.e. most
95 * fields in the arc_buf_hdr_t are protected by these mutexes).
97 * buf_hash_find() returns the appropriate mutex (held) when it
98 * locates the requested buffer in the hash table. It returns
99 * NULL for the mutex if the buffer was not in the table.
101 * buf_hash_remove() expects the appropriate hash mutex to be
102 * already held before it is invoked.
104 * Each ARC state also has a mutex which is used to protect the
105 * buffer list associated with the state. When attempting to
106 * obtain a hash table lock while holding an ARC list lock you
107 * must use: mutex_tryenter() to avoid deadlock. Also note that
108 * the active state mutex must be held before the ghost state mutex.
110 * It as also possible to register a callback which is run when the
111 * arc_meta_limit is reached and no buffers can be safely evicted. In
112 * this case the arc user should drop a reference on some arc buffers so
113 * they can be reclaimed and the arc_meta_limit honored. For example,
114 * when using the ZPL each dentry holds a references on a znode. These
115 * dentries must be pruned before the arc buffer holding the znode can
118 * Note that the majority of the performance stats are manipulated
119 * with atomic operations.
121 * The L2ARC uses the l2ad_mtx on each vdev for the following:
123 * - L2ARC buflist creation
124 * - L2ARC buflist eviction
125 * - L2ARC write completion, which walks L2ARC buflists
126 * - ARC header destruction, as it removes from L2ARC buflists
127 * - ARC header release, as it removes from L2ARC buflists
133 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
134 * This structure can point either to a block that is still in the cache or to
135 * one that is only accessible in an L2 ARC device, or it can provide
136 * information about a block that was recently evicted. If a block is
137 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
138 * information to retrieve it from the L2ARC device. This information is
139 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
140 * that is in this state cannot access the data directly.
142 * Blocks that are actively being referenced or have not been evicted
143 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
144 * the arc_buf_hdr_t that will point to the data block in memory. A block can
145 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
146 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
147 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
149 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
150 * ability to store the physical data (b_pabd) associated with the DVA of the
151 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
152 * it will match its on-disk compression characteristics. This behavior can be
153 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
154 * compressed ARC functionality is disabled, the b_pabd will point to an
155 * uncompressed version of the on-disk data.
157 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
158 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
159 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
160 * consumer. The ARC will provide references to this data and will keep it
161 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
162 * data block and will evict any arc_buf_t that is no longer referenced. The
163 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
164 * "overhead_size" kstat.
166 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
167 * compressed form. The typical case is that consumers will want uncompressed
168 * data, and when that happens a new data buffer is allocated where the data is
169 * decompressed for them to use. Currently the only consumer who wants
170 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
171 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
172 * with the arc_buf_hdr_t.
174 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
175 * first one is owned by a compressed send consumer (and therefore references
176 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
177 * used by any other consumer (and has its own uncompressed copy of the data
192 * | b_buf +------------>+-----------+ arc_buf_t
193 * | b_pabd +-+ |b_next +---->+-----------+
194 * +-----------+ | |-----------| |b_next +-->NULL
195 * | |b_comp = T | +-----------+
196 * | |b_data +-+ |b_comp = F |
197 * | +-----------+ | |b_data +-+
198 * +->+------+ | +-----------+ |
200 * data | |<--------------+ | uncompressed
201 * +------+ compressed, | data
202 * shared +-->+------+
207 * When a consumer reads a block, the ARC must first look to see if the
208 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
209 * arc_buf_t and either copies uncompressed data into a new data buffer from an
210 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
211 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
212 * hdr is compressed and the desired compression characteristics of the
213 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
214 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
215 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
216 * be anywhere in the hdr's list.
218 * The diagram below shows an example of an uncompressed ARC hdr that is
219 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
220 * the last element in the buf list):
232 * | | arc_buf_t (shared)
233 * | b_buf +------------>+---------+ arc_buf_t
234 * | | |b_next +---->+---------+
235 * | b_pabd +-+ |---------| |b_next +-->NULL
236 * +-----------+ | | | +---------+
238 * | +---------+ | |b_data +-+
239 * +->+------+ | +---------+ |
241 * uncompressed | | | |
244 * | uncompressed | | |
247 * +---------------------------------+
249 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
250 * since the physical block is about to be rewritten. The new data contents
251 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
252 * it may compress the data before writing it to disk. The ARC will be called
253 * with the transformed data and will bcopy the transformed on-disk block into
254 * a newly allocated b_pabd. Writes are always done into buffers which have
255 * either been loaned (and hence are new and don't have other readers) or
256 * buffers which have been released (and hence have their own hdr, if there
257 * were originally other readers of the buf's original hdr). This ensures that
258 * the ARC only needs to update a single buf and its hdr after a write occurs.
260 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
261 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
262 * that when compressed ARC is enabled that the L2ARC blocks are identical
263 * to the on-disk block in the main data pool. This provides a significant
264 * advantage since the ARC can leverage the bp's checksum when reading from the
265 * L2ARC to determine if the contents are valid. However, if the compressed
266 * ARC is disabled, then the L2ARC's block must be transformed to look
267 * like the physical block in the main data pool before comparing the
268 * checksum and determining its validity.
270 * The L1ARC has a slightly different system for storing encrypted data.
271 * Raw (encrypted + possibly compressed) data has a few subtle differences from
272 * data that is just compressed. The biggest difference is that it is not
273 * possible to decrypt encrypted data (or vice-versa) if the keys aren't loaded.
274 * The other difference is that encryption cannot be treated as a suggestion.
275 * If a caller would prefer compressed data, but they actually wind up with
276 * uncompressed data the worst thing that could happen is there might be a
277 * performance hit. If the caller requests encrypted data, however, we must be
278 * sure they actually get it or else secret information could be leaked. Raw
279 * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore,
280 * may have both an encrypted version and a decrypted version of its data at
281 * once. When a caller needs a raw arc_buf_t, it is allocated and the data is
282 * copied out of this header. To avoid complications with b_pabd, raw buffers
288 #include <sys/spa_impl.h>
289 #include <sys/zio_compress.h>
290 #include <sys/zio_checksum.h>
291 #include <sys/zfs_context.h>
293 #include <sys/zfs_refcount.h>
294 #include <sys/vdev.h>
295 #include <sys/vdev_impl.h>
296 #include <sys/dsl_pool.h>
297 #include <sys/zio_checksum.h>
298 #include <sys/multilist.h>
301 #include <sys/fm/fs/zfs.h>
302 #include <sys/callb.h>
303 #include <sys/kstat.h>
304 #include <sys/zthr.h>
305 #include <zfs_fletcher.h>
306 #include <sys/arc_impl.h>
307 #include <sys/trace_zfs.h>
308 #include <sys/aggsum.h>
309 #include <cityhash.h>
310 #include <sys/vdev_trim.h>
311 #include <sys/zstd/zstd.h>
314 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
315 boolean_t arc_watch = B_FALSE;
319 * This thread's job is to keep enough free memory in the system, by
320 * calling arc_kmem_reap_soon() plus arc_reduce_target_size(), which improves
321 * arc_available_memory().
323 static zthr_t *arc_reap_zthr;
326 * This thread's job is to keep arc_size under arc_c, by calling
327 * arc_evict(), which improves arc_is_overflowing().
329 static zthr_t *arc_evict_zthr;
331 static kmutex_t arc_evict_lock;
332 static boolean_t arc_evict_needed = B_FALSE;
335 * Count of bytes evicted since boot.
337 static uint64_t arc_evict_count;
340 * List of arc_evict_waiter_t's, representing threads waiting for the
341 * arc_evict_count to reach specific values.
343 static list_t arc_evict_waiters;
346 * When arc_is_overflowing(), arc_get_data_impl() waits for this percent of
347 * the requested amount of data to be evicted. For example, by default for
348 * every 2KB that's evicted, 1KB of it may be "reused" by a new allocation.
349 * Since this is above 100%, it ensures that progress is made towards getting
350 * arc_size under arc_c. Since this is finite, it ensures that allocations
351 * can still happen, even during the potentially long time that arc_size is
354 int zfs_arc_eviction_pct = 200;
357 * The number of headers to evict in arc_evict_state_impl() before
358 * dropping the sublist lock and evicting from another sublist. A lower
359 * value means we're more likely to evict the "correct" header (i.e. the
360 * oldest header in the arc state), but comes with higher overhead
361 * (i.e. more invocations of arc_evict_state_impl()).
363 int zfs_arc_evict_batch_limit = 10;
365 /* number of seconds before growing cache again */
366 int arc_grow_retry = 5;
369 * Minimum time between calls to arc_kmem_reap_soon().
371 int arc_kmem_cache_reap_retry_ms = 1000;
373 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
374 int zfs_arc_overflow_shift = 8;
376 /* shift of arc_c for calculating both min and max arc_p */
377 int arc_p_min_shift = 4;
379 /* log2(fraction of arc to reclaim) */
380 int arc_shrink_shift = 7;
382 /* percent of pagecache to reclaim arc to */
384 uint_t zfs_arc_pc_percent = 0;
388 * log2(fraction of ARC which must be free to allow growing).
389 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
390 * when reading a new block into the ARC, we will evict an equal-sized block
393 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
394 * we will still not allow it to grow.
396 int arc_no_grow_shift = 5;
400 * minimum lifespan of a prefetch block in clock ticks
401 * (initialized in arc_init())
403 static int arc_min_prefetch_ms;
404 static int arc_min_prescient_prefetch_ms;
407 * If this percent of memory is free, don't throttle.
409 int arc_lotsfree_percent = 10;
412 * The arc has filled available memory and has now warmed up.
417 * These tunables are for performance analysis.
419 unsigned long zfs_arc_max = 0;
420 unsigned long zfs_arc_min = 0;
421 unsigned long zfs_arc_meta_limit = 0;
422 unsigned long zfs_arc_meta_min = 0;
423 unsigned long zfs_arc_dnode_limit = 0;
424 unsigned long zfs_arc_dnode_reduce_percent = 10;
425 int zfs_arc_grow_retry = 0;
426 int zfs_arc_shrink_shift = 0;
427 int zfs_arc_p_min_shift = 0;
428 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
431 * ARC dirty data constraints for arc_tempreserve_space() throttle.
433 unsigned long zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */
434 unsigned long zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */
435 unsigned long zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */
438 * Enable or disable compressed arc buffers.
440 int zfs_compressed_arc_enabled = B_TRUE;
443 * ARC will evict meta buffers that exceed arc_meta_limit. This
444 * tunable make arc_meta_limit adjustable for different workloads.
446 unsigned long zfs_arc_meta_limit_percent = 75;
449 * Percentage that can be consumed by dnodes of ARC meta buffers.
451 unsigned long zfs_arc_dnode_limit_percent = 10;
454 * These tunables are Linux specific
456 unsigned long zfs_arc_sys_free = 0;
457 int zfs_arc_min_prefetch_ms = 0;
458 int zfs_arc_min_prescient_prefetch_ms = 0;
459 int zfs_arc_p_dampener_disable = 1;
460 int zfs_arc_meta_prune = 10000;
461 int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED;
462 int zfs_arc_meta_adjust_restarts = 4096;
463 int zfs_arc_lotsfree_percent = 10;
466 arc_state_t ARC_anon;
468 arc_state_t ARC_mru_ghost;
470 arc_state_t ARC_mfu_ghost;
471 arc_state_t ARC_l2c_only;
473 arc_stats_t arc_stats = {
474 { "hits", KSTAT_DATA_UINT64 },
475 { "misses", KSTAT_DATA_UINT64 },
476 { "demand_data_hits", KSTAT_DATA_UINT64 },
477 { "demand_data_misses", KSTAT_DATA_UINT64 },
478 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
479 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
480 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
481 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
482 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
483 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
484 { "mru_hits", KSTAT_DATA_UINT64 },
485 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
486 { "mfu_hits", KSTAT_DATA_UINT64 },
487 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
488 { "deleted", KSTAT_DATA_UINT64 },
489 { "mutex_miss", KSTAT_DATA_UINT64 },
490 { "access_skip", KSTAT_DATA_UINT64 },
491 { "evict_skip", KSTAT_DATA_UINT64 },
492 { "evict_not_enough", KSTAT_DATA_UINT64 },
493 { "evict_l2_cached", KSTAT_DATA_UINT64 },
494 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
495 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
496 { "evict_l2_skip", KSTAT_DATA_UINT64 },
497 { "hash_elements", KSTAT_DATA_UINT64 },
498 { "hash_elements_max", KSTAT_DATA_UINT64 },
499 { "hash_collisions", KSTAT_DATA_UINT64 },
500 { "hash_chains", KSTAT_DATA_UINT64 },
501 { "hash_chain_max", KSTAT_DATA_UINT64 },
502 { "p", KSTAT_DATA_UINT64 },
503 { "c", KSTAT_DATA_UINT64 },
504 { "c_min", KSTAT_DATA_UINT64 },
505 { "c_max", KSTAT_DATA_UINT64 },
506 { "size", KSTAT_DATA_UINT64 },
507 { "compressed_size", KSTAT_DATA_UINT64 },
508 { "uncompressed_size", KSTAT_DATA_UINT64 },
509 { "overhead_size", KSTAT_DATA_UINT64 },
510 { "hdr_size", KSTAT_DATA_UINT64 },
511 { "data_size", KSTAT_DATA_UINT64 },
512 { "metadata_size", KSTAT_DATA_UINT64 },
513 { "dbuf_size", KSTAT_DATA_UINT64 },
514 { "dnode_size", KSTAT_DATA_UINT64 },
515 { "bonus_size", KSTAT_DATA_UINT64 },
516 #if defined(COMPAT_FREEBSD11)
517 { "other_size", KSTAT_DATA_UINT64 },
519 { "anon_size", KSTAT_DATA_UINT64 },
520 { "anon_evictable_data", KSTAT_DATA_UINT64 },
521 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
522 { "mru_size", KSTAT_DATA_UINT64 },
523 { "mru_evictable_data", KSTAT_DATA_UINT64 },
524 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
525 { "mru_ghost_size", KSTAT_DATA_UINT64 },
526 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
527 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
528 { "mfu_size", KSTAT_DATA_UINT64 },
529 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
530 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
531 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
532 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
533 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
534 { "l2_hits", KSTAT_DATA_UINT64 },
535 { "l2_misses", KSTAT_DATA_UINT64 },
536 { "l2_feeds", KSTAT_DATA_UINT64 },
537 { "l2_rw_clash", KSTAT_DATA_UINT64 },
538 { "l2_read_bytes", KSTAT_DATA_UINT64 },
539 { "l2_write_bytes", KSTAT_DATA_UINT64 },
540 { "l2_writes_sent", KSTAT_DATA_UINT64 },
541 { "l2_writes_done", KSTAT_DATA_UINT64 },
542 { "l2_writes_error", KSTAT_DATA_UINT64 },
543 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
544 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
545 { "l2_evict_reading", KSTAT_DATA_UINT64 },
546 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
547 { "l2_free_on_write", KSTAT_DATA_UINT64 },
548 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
549 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
550 { "l2_io_error", KSTAT_DATA_UINT64 },
551 { "l2_size", KSTAT_DATA_UINT64 },
552 { "l2_asize", KSTAT_DATA_UINT64 },
553 { "l2_hdr_size", KSTAT_DATA_UINT64 },
554 { "l2_log_blk_writes", KSTAT_DATA_UINT64 },
555 { "l2_log_blk_avg_asize", KSTAT_DATA_UINT64 },
556 { "l2_log_blk_asize", KSTAT_DATA_UINT64 },
557 { "l2_log_blk_count", KSTAT_DATA_UINT64 },
558 { "l2_data_to_meta_ratio", KSTAT_DATA_UINT64 },
559 { "l2_rebuild_success", KSTAT_DATA_UINT64 },
560 { "l2_rebuild_unsupported", KSTAT_DATA_UINT64 },
561 { "l2_rebuild_io_errors", KSTAT_DATA_UINT64 },
562 { "l2_rebuild_dh_errors", KSTAT_DATA_UINT64 },
563 { "l2_rebuild_cksum_lb_errors", KSTAT_DATA_UINT64 },
564 { "l2_rebuild_lowmem", KSTAT_DATA_UINT64 },
565 { "l2_rebuild_size", KSTAT_DATA_UINT64 },
566 { "l2_rebuild_asize", KSTAT_DATA_UINT64 },
567 { "l2_rebuild_bufs", KSTAT_DATA_UINT64 },
568 { "l2_rebuild_bufs_precached", KSTAT_DATA_UINT64 },
569 { "l2_rebuild_log_blks", KSTAT_DATA_UINT64 },
570 { "memory_throttle_count", KSTAT_DATA_UINT64 },
571 { "memory_direct_count", KSTAT_DATA_UINT64 },
572 { "memory_indirect_count", KSTAT_DATA_UINT64 },
573 { "memory_all_bytes", KSTAT_DATA_UINT64 },
574 { "memory_free_bytes", KSTAT_DATA_UINT64 },
575 { "memory_available_bytes", KSTAT_DATA_INT64 },
576 { "arc_no_grow", KSTAT_DATA_UINT64 },
577 { "arc_tempreserve", KSTAT_DATA_UINT64 },
578 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
579 { "arc_prune", KSTAT_DATA_UINT64 },
580 { "arc_meta_used", KSTAT_DATA_UINT64 },
581 { "arc_meta_limit", KSTAT_DATA_UINT64 },
582 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
583 { "arc_meta_max", KSTAT_DATA_UINT64 },
584 { "arc_meta_min", KSTAT_DATA_UINT64 },
585 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
586 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
587 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
588 { "arc_need_free", KSTAT_DATA_UINT64 },
589 { "arc_sys_free", KSTAT_DATA_UINT64 },
590 { "arc_raw_size", KSTAT_DATA_UINT64 },
591 { "cached_only_in_progress", KSTAT_DATA_UINT64 },
592 { "abd_chunk_waste_size", KSTAT_DATA_UINT64 },
595 #define ARCSTAT_MAX(stat, val) { \
597 while ((val) > (m = arc_stats.stat.value.ui64) && \
598 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
602 #define ARCSTAT_MAXSTAT(stat) \
603 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
606 * We define a macro to allow ARC hits/misses to be easily broken down by
607 * two separate conditions, giving a total of four different subtypes for
608 * each of hits and misses (so eight statistics total).
610 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
613 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
615 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
619 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
621 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
626 * This macro allows us to use kstats as floating averages. Each time we
627 * update this kstat, we first factor it and the update value by
628 * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall
629 * average. This macro assumes that integer loads and stores are atomic, but
630 * is not safe for multiple writers updating the kstat in parallel (only the
631 * last writer's update will remain).
633 #define ARCSTAT_F_AVG_FACTOR 3
634 #define ARCSTAT_F_AVG(stat, value) \
636 uint64_t x = ARCSTAT(stat); \
637 x = x - x / ARCSTAT_F_AVG_FACTOR + \
638 (value) / ARCSTAT_F_AVG_FACTOR; \
644 static arc_state_t *arc_anon;
645 static arc_state_t *arc_mru_ghost;
646 static arc_state_t *arc_mfu_ghost;
647 static arc_state_t *arc_l2c_only;
649 arc_state_t *arc_mru;
650 arc_state_t *arc_mfu;
653 * There are several ARC variables that are critical to export as kstats --
654 * but we don't want to have to grovel around in the kstat whenever we wish to
655 * manipulate them. For these variables, we therefore define them to be in
656 * terms of the statistic variable. This assures that we are not introducing
657 * the possibility of inconsistency by having shadow copies of the variables,
658 * while still allowing the code to be readable.
660 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
661 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
662 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
663 /* max size for dnodes */
664 #define arc_dnode_size_limit ARCSTAT(arcstat_dnode_limit)
665 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
666 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
667 #define arc_need_free ARCSTAT(arcstat_need_free) /* waiting to be evicted */
669 /* size of all b_rabd's in entire arc */
670 #define arc_raw_size ARCSTAT(arcstat_raw_size)
671 /* compressed size of entire arc */
672 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
673 /* uncompressed size of entire arc */
674 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
675 /* number of bytes in the arc from arc_buf_t's */
676 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
679 * There are also some ARC variables that we want to export, but that are
680 * updated so often that having the canonical representation be the statistic
681 * variable causes a performance bottleneck. We want to use aggsum_t's for these
682 * instead, but still be able to export the kstat in the same way as before.
683 * The solution is to always use the aggsum version, except in the kstat update
687 aggsum_t arc_meta_used;
688 aggsum_t astat_data_size;
689 aggsum_t astat_metadata_size;
690 aggsum_t astat_dbuf_size;
691 aggsum_t astat_dnode_size;
692 aggsum_t astat_bonus_size;
693 aggsum_t astat_hdr_size;
694 aggsum_t astat_l2_hdr_size;
695 aggsum_t astat_abd_chunk_waste_size;
697 hrtime_t arc_growtime;
698 list_t arc_prune_list;
699 kmutex_t arc_prune_mtx;
700 taskq_t *arc_prune_taskq;
702 #define GHOST_STATE(state) \
703 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
704 (state) == arc_l2c_only)
706 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
707 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
708 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
709 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
710 #define HDR_PRESCIENT_PREFETCH(hdr) \
711 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
712 #define HDR_COMPRESSION_ENABLED(hdr) \
713 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
715 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
716 #define HDR_L2_READING(hdr) \
717 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
718 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
719 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
720 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
721 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
722 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
723 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
724 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
726 #define HDR_ISTYPE_METADATA(hdr) \
727 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
728 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
730 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
731 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
732 #define HDR_HAS_RABD(hdr) \
733 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
734 (hdr)->b_crypt_hdr.b_rabd != NULL)
735 #define HDR_ENCRYPTED(hdr) \
736 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
737 #define HDR_AUTHENTICATED(hdr) \
738 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
740 /* For storing compression mode in b_flags */
741 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
743 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
744 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
745 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
746 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
748 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
749 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
750 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
751 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
757 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
758 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
759 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
762 * Hash table routines
765 #define HT_LOCK_ALIGN 64
766 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
771 unsigned char pad[HT_LOCK_PAD];
775 #define BUF_LOCKS 8192
776 typedef struct buf_hash_table {
778 arc_buf_hdr_t **ht_table;
779 struct ht_lock ht_locks[BUF_LOCKS];
782 static buf_hash_table_t buf_hash_table;
784 #define BUF_HASH_INDEX(spa, dva, birth) \
785 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
786 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
787 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
788 #define HDR_LOCK(hdr) \
789 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
791 uint64_t zfs_crc64_table[256];
797 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
798 #define L2ARC_HEADROOM 2 /* num of writes */
801 * If we discover during ARC scan any buffers to be compressed, we boost
802 * our headroom for the next scanning cycle by this percentage multiple.
804 #define L2ARC_HEADROOM_BOOST 200
805 #define L2ARC_FEED_SECS 1 /* caching interval secs */
806 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
809 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
810 * and each of the state has two types: data and metadata.
812 #define L2ARC_FEED_TYPES 4
814 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
815 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
817 /* L2ARC Performance Tunables */
818 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
819 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
820 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
821 unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
822 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
823 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
824 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
825 int l2arc_feed_again = B_TRUE; /* turbo warmup */
826 int l2arc_norw = B_FALSE; /* no reads during writes */
827 int l2arc_meta_percent = 33; /* limit on headers size */
832 static list_t L2ARC_dev_list; /* device list */
833 static list_t *l2arc_dev_list; /* device list pointer */
834 static kmutex_t l2arc_dev_mtx; /* device list mutex */
835 static l2arc_dev_t *l2arc_dev_last; /* last device used */
836 static list_t L2ARC_free_on_write; /* free after write buf list */
837 static list_t *l2arc_free_on_write; /* free after write list ptr */
838 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
839 static uint64_t l2arc_ndev; /* number of devices */
841 typedef struct l2arc_read_callback {
842 arc_buf_hdr_t *l2rcb_hdr; /* read header */
843 blkptr_t l2rcb_bp; /* original blkptr */
844 zbookmark_phys_t l2rcb_zb; /* original bookmark */
845 int l2rcb_flags; /* original flags */
846 abd_t *l2rcb_abd; /* temporary buffer */
847 } l2arc_read_callback_t;
849 typedef struct l2arc_data_free {
850 /* protected by l2arc_free_on_write_mtx */
853 arc_buf_contents_t l2df_type;
854 list_node_t l2df_list_node;
857 typedef enum arc_fill_flags {
858 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
859 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
860 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
861 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
862 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
865 static kmutex_t l2arc_feed_thr_lock;
866 static kcondvar_t l2arc_feed_thr_cv;
867 static uint8_t l2arc_thread_exit;
869 static kmutex_t l2arc_rebuild_thr_lock;
870 static kcondvar_t l2arc_rebuild_thr_cv;
872 enum arc_hdr_alloc_flags {
873 ARC_HDR_ALLOC_RDATA = 0x1,
874 ARC_HDR_DO_ADAPT = 0x2,
878 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *, boolean_t);
879 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
880 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *, boolean_t);
881 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
882 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
883 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
884 static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t);
885 static void arc_hdr_alloc_abd(arc_buf_hdr_t *, int);
886 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
887 static void arc_buf_watch(arc_buf_t *);
889 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
890 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
891 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
892 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
894 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
895 static void l2arc_read_done(zio_t *);
896 static void l2arc_do_free_on_write(void);
899 * l2arc_mfuonly : A ZFS module parameter that controls whether only MFU
900 * metadata and data are cached from ARC into L2ARC.
902 int l2arc_mfuonly = 0;
906 * l2arc_trim_ahead : A ZFS module parameter that controls how much ahead of
907 * the current write size (l2arc_write_max) we should TRIM if we
908 * have filled the device. It is defined as a percentage of the
909 * write size. If set to 100 we trim twice the space required to
910 * accommodate upcoming writes. A minimum of 64MB will be trimmed.
911 * It also enables TRIM of the whole L2ARC device upon creation or
912 * addition to an existing pool or if the header of the device is
913 * invalid upon importing a pool or onlining a cache device. The
914 * default is 0, which disables TRIM on L2ARC altogether as it can
915 * put significant stress on the underlying storage devices. This
916 * will vary depending of how well the specific device handles
919 unsigned long l2arc_trim_ahead = 0;
922 * Performance tuning of L2ARC persistence:
924 * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding
925 * an L2ARC device (either at pool import or later) will attempt
926 * to rebuild L2ARC buffer contents.
927 * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls
928 * whether log blocks are written to the L2ARC device. If the L2ARC
929 * device is less than 1GB, the amount of data l2arc_evict()
930 * evicts is significant compared to the amount of restored L2ARC
931 * data. In this case do not write log blocks in L2ARC in order
932 * not to waste space.
934 int l2arc_rebuild_enabled = B_TRUE;
935 unsigned long l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024;
937 /* L2ARC persistence rebuild control routines. */
938 void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen);
939 static void l2arc_dev_rebuild_thread(void *arg);
940 static int l2arc_rebuild(l2arc_dev_t *dev);
942 /* L2ARC persistence read I/O routines. */
943 static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
944 static int l2arc_log_blk_read(l2arc_dev_t *dev,
945 const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
946 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
947 zio_t *this_io, zio_t **next_io);
948 static zio_t *l2arc_log_blk_fetch(vdev_t *vd,
949 const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb);
950 static void l2arc_log_blk_fetch_abort(zio_t *zio);
952 /* L2ARC persistence block restoration routines. */
953 static void l2arc_log_blk_restore(l2arc_dev_t *dev,
954 const l2arc_log_blk_phys_t *lb, uint64_t lb_asize, uint64_t lb_daddr);
955 static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
958 /* L2ARC persistence write I/O routines. */
959 static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
960 l2arc_write_callback_t *cb);
962 /* L2ARC persistence auxiliary routines. */
963 boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
964 const l2arc_log_blkptr_t *lbp);
965 static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
966 const arc_buf_hdr_t *ab);
967 boolean_t l2arc_range_check_overlap(uint64_t bottom,
968 uint64_t top, uint64_t check);
969 static void l2arc_blk_fetch_done(zio_t *zio);
970 static inline uint64_t
971 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev);
974 * We use Cityhash for this. It's fast, and has good hash properties without
975 * requiring any large static buffers.
978 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
980 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
983 #define HDR_EMPTY(hdr) \
984 ((hdr)->b_dva.dva_word[0] == 0 && \
985 (hdr)->b_dva.dva_word[1] == 0)
987 #define HDR_EMPTY_OR_LOCKED(hdr) \
988 (HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr)))
990 #define HDR_EQUAL(spa, dva, birth, hdr) \
991 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
992 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
993 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
996 buf_discard_identity(arc_buf_hdr_t *hdr)
998 hdr->b_dva.dva_word[0] = 0;
999 hdr->b_dva.dva_word[1] = 0;
1003 static arc_buf_hdr_t *
1004 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1006 const dva_t *dva = BP_IDENTITY(bp);
1007 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1008 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1009 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1012 mutex_enter(hash_lock);
1013 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1014 hdr = hdr->b_hash_next) {
1015 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1020 mutex_exit(hash_lock);
1026 * Insert an entry into the hash table. If there is already an element
1027 * equal to elem in the hash table, then the already existing element
1028 * will be returned and the new element will not be inserted.
1029 * Otherwise returns NULL.
1030 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1032 static arc_buf_hdr_t *
1033 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1035 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1036 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1037 arc_buf_hdr_t *fhdr;
1040 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1041 ASSERT(hdr->b_birth != 0);
1042 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1044 if (lockp != NULL) {
1046 mutex_enter(hash_lock);
1048 ASSERT(MUTEX_HELD(hash_lock));
1051 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1052 fhdr = fhdr->b_hash_next, i++) {
1053 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1057 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1058 buf_hash_table.ht_table[idx] = hdr;
1059 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1061 /* collect some hash table performance data */
1063 ARCSTAT_BUMP(arcstat_hash_collisions);
1065 ARCSTAT_BUMP(arcstat_hash_chains);
1067 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1070 ARCSTAT_BUMP(arcstat_hash_elements);
1071 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1077 buf_hash_remove(arc_buf_hdr_t *hdr)
1079 arc_buf_hdr_t *fhdr, **hdrp;
1080 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1082 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1083 ASSERT(HDR_IN_HASH_TABLE(hdr));
1085 hdrp = &buf_hash_table.ht_table[idx];
1086 while ((fhdr = *hdrp) != hdr) {
1087 ASSERT3P(fhdr, !=, NULL);
1088 hdrp = &fhdr->b_hash_next;
1090 *hdrp = hdr->b_hash_next;
1091 hdr->b_hash_next = NULL;
1092 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1094 /* collect some hash table performance data */
1095 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1097 if (buf_hash_table.ht_table[idx] &&
1098 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1099 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1103 * Global data structures and functions for the buf kmem cache.
1106 static kmem_cache_t *hdr_full_cache;
1107 static kmem_cache_t *hdr_full_crypt_cache;
1108 static kmem_cache_t *hdr_l2only_cache;
1109 static kmem_cache_t *buf_cache;
1116 #if defined(_KERNEL)
1118 * Large allocations which do not require contiguous pages
1119 * should be using vmem_free() in the linux kernel\
1121 vmem_free(buf_hash_table.ht_table,
1122 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1124 kmem_free(buf_hash_table.ht_table,
1125 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1127 for (i = 0; i < BUF_LOCKS; i++)
1128 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1129 kmem_cache_destroy(hdr_full_cache);
1130 kmem_cache_destroy(hdr_full_crypt_cache);
1131 kmem_cache_destroy(hdr_l2only_cache);
1132 kmem_cache_destroy(buf_cache);
1136 * Constructor callback - called when the cache is empty
1137 * and a new buf is requested.
1141 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1143 arc_buf_hdr_t *hdr = vbuf;
1145 bzero(hdr, HDR_FULL_SIZE);
1146 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1147 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1148 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1149 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1150 list_link_init(&hdr->b_l1hdr.b_arc_node);
1151 list_link_init(&hdr->b_l2hdr.b_l2node);
1152 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1153 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1160 hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
1162 arc_buf_hdr_t *hdr = vbuf;
1164 hdr_full_cons(vbuf, unused, kmflag);
1165 bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr));
1166 arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1173 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1175 arc_buf_hdr_t *hdr = vbuf;
1177 bzero(hdr, HDR_L2ONLY_SIZE);
1178 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1185 buf_cons(void *vbuf, void *unused, int kmflag)
1187 arc_buf_t *buf = vbuf;
1189 bzero(buf, sizeof (arc_buf_t));
1190 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1191 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1197 * Destructor callback - called when a cached buf is
1198 * no longer required.
1202 hdr_full_dest(void *vbuf, void *unused)
1204 arc_buf_hdr_t *hdr = vbuf;
1206 ASSERT(HDR_EMPTY(hdr));
1207 cv_destroy(&hdr->b_l1hdr.b_cv);
1208 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1209 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1210 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1211 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1216 hdr_full_crypt_dest(void *vbuf, void *unused)
1218 arc_buf_hdr_t *hdr = vbuf;
1220 hdr_full_dest(vbuf, unused);
1221 arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1226 hdr_l2only_dest(void *vbuf, void *unused)
1228 arc_buf_hdr_t *hdr __maybe_unused = vbuf;
1230 ASSERT(HDR_EMPTY(hdr));
1231 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1236 buf_dest(void *vbuf, void *unused)
1238 arc_buf_t *buf = vbuf;
1240 mutex_destroy(&buf->b_evict_lock);
1241 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1247 uint64_t *ct = NULL;
1248 uint64_t hsize = 1ULL << 12;
1252 * The hash table is big enough to fill all of physical memory
1253 * with an average block size of zfs_arc_average_blocksize (default 8K).
1254 * By default, the table will take up
1255 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1257 while (hsize * zfs_arc_average_blocksize < arc_all_memory())
1260 buf_hash_table.ht_mask = hsize - 1;
1261 #if defined(_KERNEL)
1263 * Large allocations which do not require contiguous pages
1264 * should be using vmem_alloc() in the linux kernel
1266 buf_hash_table.ht_table =
1267 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
1269 buf_hash_table.ht_table =
1270 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1272 if (buf_hash_table.ht_table == NULL) {
1273 ASSERT(hsize > (1ULL << 8));
1278 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1279 0, hdr_full_cons, hdr_full_dest, NULL, NULL, NULL, 0);
1280 hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
1281 HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
1282 NULL, NULL, NULL, 0);
1283 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1284 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, NULL,
1286 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1287 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1289 for (i = 0; i < 256; i++)
1290 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1291 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1293 for (i = 0; i < BUF_LOCKS; i++) {
1294 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1295 NULL, MUTEX_DEFAULT, NULL);
1299 #define ARC_MINTIME (hz>>4) /* 62 ms */
1302 * This is the size that the buf occupies in memory. If the buf is compressed,
1303 * it will correspond to the compressed size. You should use this method of
1304 * getting the buf size unless you explicitly need the logical size.
1307 arc_buf_size(arc_buf_t *buf)
1309 return (ARC_BUF_COMPRESSED(buf) ?
1310 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1314 arc_buf_lsize(arc_buf_t *buf)
1316 return (HDR_GET_LSIZE(buf->b_hdr));
1320 * This function will return B_TRUE if the buffer is encrypted in memory.
1321 * This buffer can be decrypted by calling arc_untransform().
1324 arc_is_encrypted(arc_buf_t *buf)
1326 return (ARC_BUF_ENCRYPTED(buf) != 0);
1330 * Returns B_TRUE if the buffer represents data that has not had its MAC
1334 arc_is_unauthenticated(arc_buf_t *buf)
1336 return (HDR_NOAUTH(buf->b_hdr) != 0);
1340 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1341 uint8_t *iv, uint8_t *mac)
1343 arc_buf_hdr_t *hdr = buf->b_hdr;
1345 ASSERT(HDR_PROTECTED(hdr));
1347 bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
1348 bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
1349 bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
1350 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1351 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1355 * Indicates how this buffer is compressed in memory. If it is not compressed
1356 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1357 * arc_untransform() as long as it is also unencrypted.
1360 arc_get_compression(arc_buf_t *buf)
1362 return (ARC_BUF_COMPRESSED(buf) ?
1363 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1367 * Return the compression algorithm used to store this data in the ARC. If ARC
1368 * compression is enabled or this is an encrypted block, this will be the same
1369 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1371 static inline enum zio_compress
1372 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1374 return (HDR_COMPRESSION_ENABLED(hdr) ?
1375 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1379 arc_get_complevel(arc_buf_t *buf)
1381 return (buf->b_hdr->b_complevel);
1384 static inline boolean_t
1385 arc_buf_is_shared(arc_buf_t *buf)
1387 boolean_t shared = (buf->b_data != NULL &&
1388 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1389 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1390 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1391 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1392 IMPLY(shared, ARC_BUF_SHARED(buf));
1393 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1396 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1397 * already being shared" requirement prevents us from doing that.
1404 * Free the checksum associated with this header. If there is no checksum, this
1408 arc_cksum_free(arc_buf_hdr_t *hdr)
1410 ASSERT(HDR_HAS_L1HDR(hdr));
1412 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1413 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1414 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1415 hdr->b_l1hdr.b_freeze_cksum = NULL;
1417 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1421 * Return true iff at least one of the bufs on hdr is not compressed.
1422 * Encrypted buffers count as compressed.
1425 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1427 ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr));
1429 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1430 if (!ARC_BUF_COMPRESSED(b)) {
1439 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1440 * matches the checksum that is stored in the hdr. If there is no checksum,
1441 * or if the buf is compressed, this is a no-op.
1444 arc_cksum_verify(arc_buf_t *buf)
1446 arc_buf_hdr_t *hdr = buf->b_hdr;
1449 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1452 if (ARC_BUF_COMPRESSED(buf))
1455 ASSERT(HDR_HAS_L1HDR(hdr));
1457 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1459 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1460 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1464 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1465 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1466 panic("buffer modified while frozen!");
1467 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1471 * This function makes the assumption that data stored in the L2ARC
1472 * will be transformed exactly as it is in the main pool. Because of
1473 * this we can verify the checksum against the reading process's bp.
1476 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1478 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1479 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1482 * Block pointers always store the checksum for the logical data.
1483 * If the block pointer has the gang bit set, then the checksum
1484 * it represents is for the reconstituted data and not for an
1485 * individual gang member. The zio pipeline, however, must be able to
1486 * determine the checksum of each of the gang constituents so it
1487 * treats the checksum comparison differently than what we need
1488 * for l2arc blocks. This prevents us from using the
1489 * zio_checksum_error() interface directly. Instead we must call the
1490 * zio_checksum_error_impl() so that we can ensure the checksum is
1491 * generated using the correct checksum algorithm and accounts for the
1492 * logical I/O size and not just a gang fragment.
1494 return (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1495 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1496 zio->io_offset, NULL) == 0);
1500 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1501 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1502 * isn't modified later on. If buf is compressed or there is already a checksum
1503 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1506 arc_cksum_compute(arc_buf_t *buf)
1508 arc_buf_hdr_t *hdr = buf->b_hdr;
1510 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1513 ASSERT(HDR_HAS_L1HDR(hdr));
1515 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1516 if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1517 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1521 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1522 ASSERT(!ARC_BUF_COMPRESSED(buf));
1523 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1525 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1526 hdr->b_l1hdr.b_freeze_cksum);
1527 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1533 arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
1535 panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr);
1541 arc_buf_unwatch(arc_buf_t *buf)
1545 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1546 PROT_READ | PROT_WRITE));
1553 arc_buf_watch(arc_buf_t *buf)
1557 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1562 static arc_buf_contents_t
1563 arc_buf_type(arc_buf_hdr_t *hdr)
1565 arc_buf_contents_t type;
1566 if (HDR_ISTYPE_METADATA(hdr)) {
1567 type = ARC_BUFC_METADATA;
1569 type = ARC_BUFC_DATA;
1571 VERIFY3U(hdr->b_type, ==, type);
1576 arc_is_metadata(arc_buf_t *buf)
1578 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1582 arc_bufc_to_flags(arc_buf_contents_t type)
1586 /* metadata field is 0 if buffer contains normal data */
1588 case ARC_BUFC_METADATA:
1589 return (ARC_FLAG_BUFC_METADATA);
1593 panic("undefined ARC buffer type!");
1594 return ((uint32_t)-1);
1598 arc_buf_thaw(arc_buf_t *buf)
1600 arc_buf_hdr_t *hdr = buf->b_hdr;
1602 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1603 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1605 arc_cksum_verify(buf);
1608 * Compressed buffers do not manipulate the b_freeze_cksum.
1610 if (ARC_BUF_COMPRESSED(buf))
1613 ASSERT(HDR_HAS_L1HDR(hdr));
1614 arc_cksum_free(hdr);
1615 arc_buf_unwatch(buf);
1619 arc_buf_freeze(arc_buf_t *buf)
1621 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1624 if (ARC_BUF_COMPRESSED(buf))
1627 ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1628 arc_cksum_compute(buf);
1632 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1633 * the following functions should be used to ensure that the flags are
1634 * updated in a thread-safe way. When manipulating the flags either
1635 * the hash_lock must be held or the hdr must be undiscoverable. This
1636 * ensures that we're not racing with any other threads when updating
1640 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1642 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1643 hdr->b_flags |= flags;
1647 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1649 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1650 hdr->b_flags &= ~flags;
1654 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1655 * done in a special way since we have to clear and set bits
1656 * at the same time. Consumers that wish to set the compression bits
1657 * must use this function to ensure that the flags are updated in
1658 * thread-safe manner.
1661 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1663 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1666 * Holes and embedded blocks will always have a psize = 0 so
1667 * we ignore the compression of the blkptr and set the
1668 * want to uncompress them. Mark them as uncompressed.
1670 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1671 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1672 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1674 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1675 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1678 HDR_SET_COMPRESS(hdr, cmp);
1679 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1683 * Looks for another buf on the same hdr which has the data decompressed, copies
1684 * from it, and returns true. If no such buf exists, returns false.
1687 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1689 arc_buf_hdr_t *hdr = buf->b_hdr;
1690 boolean_t copied = B_FALSE;
1692 ASSERT(HDR_HAS_L1HDR(hdr));
1693 ASSERT3P(buf->b_data, !=, NULL);
1694 ASSERT(!ARC_BUF_COMPRESSED(buf));
1696 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1697 from = from->b_next) {
1698 /* can't use our own data buffer */
1703 if (!ARC_BUF_COMPRESSED(from)) {
1704 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1711 * There were no decompressed bufs, so there should not be a
1712 * checksum on the hdr either.
1714 if (zfs_flags & ZFS_DEBUG_MODIFY)
1715 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1721 * Allocates an ARC buf header that's in an evicted & L2-cached state.
1722 * This is used during l2arc reconstruction to make empty ARC buffers
1723 * which circumvent the regular disk->arc->l2arc path and instead come
1724 * into being in the reverse order, i.e. l2arc->arc.
1726 static arc_buf_hdr_t *
1727 arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
1728 dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth,
1729 enum zio_compress compress, uint8_t complevel, boolean_t protected,
1735 hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP);
1736 hdr->b_birth = birth;
1739 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR);
1740 HDR_SET_LSIZE(hdr, size);
1741 HDR_SET_PSIZE(hdr, psize);
1742 arc_hdr_set_compress(hdr, compress);
1743 hdr->b_complevel = complevel;
1745 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
1747 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
1748 hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa);
1752 hdr->b_l2hdr.b_dev = dev;
1753 hdr->b_l2hdr.b_daddr = daddr;
1759 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1762 arc_hdr_size(arc_buf_hdr_t *hdr)
1766 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1767 HDR_GET_PSIZE(hdr) > 0) {
1768 size = HDR_GET_PSIZE(hdr);
1770 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1771 size = HDR_GET_LSIZE(hdr);
1777 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1781 uint64_t lsize = HDR_GET_LSIZE(hdr);
1782 uint64_t psize = HDR_GET_PSIZE(hdr);
1783 void *tmpbuf = NULL;
1784 abd_t *abd = hdr->b_l1hdr.b_pabd;
1786 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1787 ASSERT(HDR_AUTHENTICATED(hdr));
1788 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1791 * The MAC is calculated on the compressed data that is stored on disk.
1792 * However, if compressed arc is disabled we will only have the
1793 * decompressed data available to us now. Compress it into a temporary
1794 * abd so we can verify the MAC. The performance overhead of this will
1795 * be relatively low, since most objects in an encrypted objset will
1796 * be encrypted (instead of authenticated) anyway.
1798 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1799 !HDR_COMPRESSION_ENABLED(hdr)) {
1800 tmpbuf = zio_buf_alloc(lsize);
1801 abd = abd_get_from_buf(tmpbuf, lsize);
1802 abd_take_ownership_of_buf(abd, B_TRUE);
1803 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1804 hdr->b_l1hdr.b_pabd, tmpbuf, lsize, hdr->b_complevel);
1805 ASSERT3U(csize, <=, psize);
1806 abd_zero_off(abd, csize, psize - csize);
1810 * Authentication is best effort. We authenticate whenever the key is
1811 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1813 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1814 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1815 ASSERT3U(lsize, ==, psize);
1816 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1817 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1819 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1820 hdr->b_crypt_hdr.b_mac);
1824 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1825 else if (ret != ENOENT)
1841 * This function will take a header that only has raw encrypted data in
1842 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1843 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1844 * also decompress the data.
1847 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1852 boolean_t no_crypt = B_FALSE;
1853 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1855 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1856 ASSERT(HDR_ENCRYPTED(hdr));
1858 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT);
1860 ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1861 B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1862 hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1863 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1868 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1869 HDR_GET_PSIZE(hdr));
1873 * If this header has disabled arc compression but the b_pabd is
1874 * compressed after decrypting it, we need to decompress the newly
1877 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1878 !HDR_COMPRESSION_ENABLED(hdr)) {
1880 * We want to make sure that we are correctly honoring the
1881 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1882 * and then loan a buffer from it, rather than allocating a
1883 * linear buffer and wrapping it in an abd later.
1885 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr, B_TRUE);
1886 tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
1888 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1889 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
1890 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
1892 abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
1896 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
1897 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1898 arc_hdr_size(hdr), hdr);
1899 hdr->b_l1hdr.b_pabd = cabd;
1905 arc_hdr_free_abd(hdr, B_FALSE);
1907 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
1913 * This function is called during arc_buf_fill() to prepare the header's
1914 * abd plaintext pointer for use. This involves authenticated protected
1915 * data and decrypting encrypted data into the plaintext abd.
1918 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
1919 const zbookmark_phys_t *zb, boolean_t noauth)
1923 ASSERT(HDR_PROTECTED(hdr));
1925 if (hash_lock != NULL)
1926 mutex_enter(hash_lock);
1928 if (HDR_NOAUTH(hdr) && !noauth) {
1930 * The caller requested authenticated data but our data has
1931 * not been authenticated yet. Verify the MAC now if we can.
1933 ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
1936 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
1938 * If we only have the encrypted version of the data, but the
1939 * unencrypted version was requested we take this opportunity
1940 * to store the decrypted version in the header for future use.
1942 ret = arc_hdr_decrypt(hdr, spa, zb);
1947 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1949 if (hash_lock != NULL)
1950 mutex_exit(hash_lock);
1955 if (hash_lock != NULL)
1956 mutex_exit(hash_lock);
1962 * This function is used by the dbuf code to decrypt bonus buffers in place.
1963 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
1964 * block, so we use the hash lock here to protect against concurrent calls to
1968 arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock)
1970 arc_buf_hdr_t *hdr = buf->b_hdr;
1972 ASSERT(HDR_ENCRYPTED(hdr));
1973 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1974 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1975 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1977 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
1979 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
1980 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1981 hdr->b_crypt_hdr.b_ebufcnt -= 1;
1985 * Given a buf that has a data buffer attached to it, this function will
1986 * efficiently fill the buf with data of the specified compression setting from
1987 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1988 * are already sharing a data buf, no copy is performed.
1990 * If the buf is marked as compressed but uncompressed data was requested, this
1991 * will allocate a new data buffer for the buf, remove that flag, and fill the
1992 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1993 * uncompressed data, and (since we haven't added support for it yet) if you
1994 * want compressed data your buf must already be marked as compressed and have
1995 * the correct-sized data buffer.
1998 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
1999 arc_fill_flags_t flags)
2002 arc_buf_hdr_t *hdr = buf->b_hdr;
2003 boolean_t hdr_compressed =
2004 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
2005 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
2006 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
2007 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2008 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
2010 ASSERT3P(buf->b_data, !=, NULL);
2011 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
2012 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2013 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
2014 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
2015 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
2016 IMPLY(encrypted, !ARC_BUF_SHARED(buf));
2019 * If the caller wanted encrypted data we just need to copy it from
2020 * b_rabd and potentially byteswap it. We won't be able to do any
2021 * further transforms on it.
2024 ASSERT(HDR_HAS_RABD(hdr));
2025 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
2026 HDR_GET_PSIZE(hdr));
2031 * Adjust encrypted and authenticated headers to accommodate
2032 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
2033 * allowed to fail decryption due to keys not being loaded
2034 * without being marked as an IO error.
2036 if (HDR_PROTECTED(hdr)) {
2037 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
2038 zb, !!(flags & ARC_FILL_NOAUTH));
2039 if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
2041 } else if (error != 0) {
2042 if (hash_lock != NULL)
2043 mutex_enter(hash_lock);
2044 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2045 if (hash_lock != NULL)
2046 mutex_exit(hash_lock);
2052 * There is a special case here for dnode blocks which are
2053 * decrypting their bonus buffers. These blocks may request to
2054 * be decrypted in-place. This is necessary because there may
2055 * be many dnodes pointing into this buffer and there is
2056 * currently no method to synchronize replacing the backing
2057 * b_data buffer and updating all of the pointers. Here we use
2058 * the hash lock to ensure there are no races. If the need
2059 * arises for other types to be decrypted in-place, they must
2060 * add handling here as well.
2062 if ((flags & ARC_FILL_IN_PLACE) != 0) {
2063 ASSERT(!hdr_compressed);
2064 ASSERT(!compressed);
2067 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
2068 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2070 if (hash_lock != NULL)
2071 mutex_enter(hash_lock);
2072 arc_buf_untransform_in_place(buf, hash_lock);
2073 if (hash_lock != NULL)
2074 mutex_exit(hash_lock);
2076 /* Compute the hdr's checksum if necessary */
2077 arc_cksum_compute(buf);
2083 if (hdr_compressed == compressed) {
2084 if (!arc_buf_is_shared(buf)) {
2085 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2089 ASSERT(hdr_compressed);
2090 ASSERT(!compressed);
2091 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2094 * If the buf is sharing its data with the hdr, unlink it and
2095 * allocate a new data buffer for the buf.
2097 if (arc_buf_is_shared(buf)) {
2098 ASSERT(ARC_BUF_COMPRESSED(buf));
2100 /* We need to give the buf its own b_data */
2101 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2103 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2104 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2106 /* Previously overhead was 0; just add new overhead */
2107 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2108 } else if (ARC_BUF_COMPRESSED(buf)) {
2109 /* We need to reallocate the buf's b_data */
2110 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2113 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2115 /* We increased the size of b_data; update overhead */
2116 ARCSTAT_INCR(arcstat_overhead_size,
2117 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2121 * Regardless of the buf's previous compression settings, it
2122 * should not be compressed at the end of this function.
2124 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2127 * Try copying the data from another buf which already has a
2128 * decompressed version. If that's not possible, it's time to
2129 * bite the bullet and decompress the data from the hdr.
2131 if (arc_buf_try_copy_decompressed_data(buf)) {
2132 /* Skip byteswapping and checksumming (already done) */
2135 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2136 hdr->b_l1hdr.b_pabd, buf->b_data,
2137 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr),
2141 * Absent hardware errors or software bugs, this should
2142 * be impossible, but log it anyway so we can debug it.
2146 "hdr %px, compress %d, psize %d, lsize %d",
2147 hdr, arc_hdr_get_compress(hdr),
2148 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2149 if (hash_lock != NULL)
2150 mutex_enter(hash_lock);
2151 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2152 if (hash_lock != NULL)
2153 mutex_exit(hash_lock);
2154 return (SET_ERROR(EIO));
2160 /* Byteswap the buf's data if necessary */
2161 if (bswap != DMU_BSWAP_NUMFUNCS) {
2162 ASSERT(!HDR_SHARED_DATA(hdr));
2163 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2164 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2167 /* Compute the hdr's checksum if necessary */
2168 arc_cksum_compute(buf);
2174 * If this function is being called to decrypt an encrypted buffer or verify an
2175 * authenticated one, the key must be loaded and a mapping must be made
2176 * available in the keystore via spa_keystore_create_mapping() or one of its
2180 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2184 arc_fill_flags_t flags = 0;
2187 flags |= ARC_FILL_IN_PLACE;
2189 ret = arc_buf_fill(buf, spa, zb, flags);
2190 if (ret == ECKSUM) {
2192 * Convert authentication and decryption errors to EIO
2193 * (and generate an ereport) before leaving the ARC.
2195 ret = SET_ERROR(EIO);
2196 spa_log_error(spa, zb);
2197 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2198 spa, NULL, zb, NULL, 0);
2205 * Increment the amount of evictable space in the arc_state_t's refcount.
2206 * We account for the space used by the hdr and the arc buf individually
2207 * so that we can add and remove them from the refcount individually.
2210 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2212 arc_buf_contents_t type = arc_buf_type(hdr);
2214 ASSERT(HDR_HAS_L1HDR(hdr));
2216 if (GHOST_STATE(state)) {
2217 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2218 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2219 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2220 ASSERT(!HDR_HAS_RABD(hdr));
2221 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2222 HDR_GET_LSIZE(hdr), hdr);
2226 ASSERT(!GHOST_STATE(state));
2227 if (hdr->b_l1hdr.b_pabd != NULL) {
2228 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2229 arc_hdr_size(hdr), hdr);
2231 if (HDR_HAS_RABD(hdr)) {
2232 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2233 HDR_GET_PSIZE(hdr), hdr);
2236 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2237 buf = buf->b_next) {
2238 if (arc_buf_is_shared(buf))
2240 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2241 arc_buf_size(buf), buf);
2246 * Decrement the amount of evictable space in the arc_state_t's refcount.
2247 * We account for the space used by the hdr and the arc buf individually
2248 * so that we can add and remove them from the refcount individually.
2251 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2253 arc_buf_contents_t type = arc_buf_type(hdr);
2255 ASSERT(HDR_HAS_L1HDR(hdr));
2257 if (GHOST_STATE(state)) {
2258 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2259 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2260 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2261 ASSERT(!HDR_HAS_RABD(hdr));
2262 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2263 HDR_GET_LSIZE(hdr), hdr);
2267 ASSERT(!GHOST_STATE(state));
2268 if (hdr->b_l1hdr.b_pabd != NULL) {
2269 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2270 arc_hdr_size(hdr), hdr);
2272 if (HDR_HAS_RABD(hdr)) {
2273 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2274 HDR_GET_PSIZE(hdr), hdr);
2277 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2278 buf = buf->b_next) {
2279 if (arc_buf_is_shared(buf))
2281 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2282 arc_buf_size(buf), buf);
2287 * Add a reference to this hdr indicating that someone is actively
2288 * referencing that memory. When the refcount transitions from 0 to 1,
2289 * we remove it from the respective arc_state_t list to indicate that
2290 * it is not evictable.
2293 add_reference(arc_buf_hdr_t *hdr, void *tag)
2297 ASSERT(HDR_HAS_L1HDR(hdr));
2298 if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) {
2299 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2300 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2301 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2304 state = hdr->b_l1hdr.b_state;
2306 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2307 (state != arc_anon)) {
2308 /* We don't use the L2-only state list. */
2309 if (state != arc_l2c_only) {
2310 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2312 arc_evictable_space_decrement(hdr, state);
2314 /* remove the prefetch flag if we get a reference */
2315 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2320 * Remove a reference from this hdr. When the reference transitions from
2321 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2322 * list making it eligible for eviction.
2325 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2328 arc_state_t *state = hdr->b_l1hdr.b_state;
2330 ASSERT(HDR_HAS_L1HDR(hdr));
2331 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2332 ASSERT(!GHOST_STATE(state));
2335 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2336 * check to prevent usage of the arc_l2c_only list.
2338 if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2339 (state != arc_anon)) {
2340 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2341 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2342 arc_evictable_space_increment(hdr, state);
2348 * Returns detailed information about a specific arc buffer. When the
2349 * state_index argument is set the function will calculate the arc header
2350 * list position for its arc state. Since this requires a linear traversal
2351 * callers are strongly encourage not to do this. However, it can be helpful
2352 * for targeted analysis so the functionality is provided.
2355 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2357 arc_buf_hdr_t *hdr = ab->b_hdr;
2358 l1arc_buf_hdr_t *l1hdr = NULL;
2359 l2arc_buf_hdr_t *l2hdr = NULL;
2360 arc_state_t *state = NULL;
2362 memset(abi, 0, sizeof (arc_buf_info_t));
2367 abi->abi_flags = hdr->b_flags;
2369 if (HDR_HAS_L1HDR(hdr)) {
2370 l1hdr = &hdr->b_l1hdr;
2371 state = l1hdr->b_state;
2373 if (HDR_HAS_L2HDR(hdr))
2374 l2hdr = &hdr->b_l2hdr;
2377 abi->abi_bufcnt = l1hdr->b_bufcnt;
2378 abi->abi_access = l1hdr->b_arc_access;
2379 abi->abi_mru_hits = l1hdr->b_mru_hits;
2380 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2381 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2382 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2383 abi->abi_holds = zfs_refcount_count(&l1hdr->b_refcnt);
2387 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2388 abi->abi_l2arc_hits = l2hdr->b_hits;
2391 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2392 abi->abi_state_contents = arc_buf_type(hdr);
2393 abi->abi_size = arc_hdr_size(hdr);
2397 * Move the supplied buffer to the indicated state. The hash lock
2398 * for the buffer must be held by the caller.
2401 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2402 kmutex_t *hash_lock)
2404 arc_state_t *old_state;
2407 boolean_t update_old, update_new;
2408 arc_buf_contents_t buftype = arc_buf_type(hdr);
2411 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2412 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2413 * L1 hdr doesn't always exist when we change state to arc_anon before
2414 * destroying a header, in which case reallocating to add the L1 hdr is
2417 if (HDR_HAS_L1HDR(hdr)) {
2418 old_state = hdr->b_l1hdr.b_state;
2419 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2420 bufcnt = hdr->b_l1hdr.b_bufcnt;
2421 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
2424 old_state = arc_l2c_only;
2427 update_old = B_FALSE;
2429 update_new = update_old;
2431 ASSERT(MUTEX_HELD(hash_lock));
2432 ASSERT3P(new_state, !=, old_state);
2433 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2434 ASSERT(old_state != arc_anon || bufcnt <= 1);
2437 * If this buffer is evictable, transfer it from the
2438 * old state list to the new state list.
2441 if (old_state != arc_anon && old_state != arc_l2c_only) {
2442 ASSERT(HDR_HAS_L1HDR(hdr));
2443 multilist_remove(old_state->arcs_list[buftype], hdr);
2445 if (GHOST_STATE(old_state)) {
2447 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2448 update_old = B_TRUE;
2450 arc_evictable_space_decrement(hdr, old_state);
2452 if (new_state != arc_anon && new_state != arc_l2c_only) {
2454 * An L1 header always exists here, since if we're
2455 * moving to some L1-cached state (i.e. not l2c_only or
2456 * anonymous), we realloc the header to add an L1hdr
2459 ASSERT(HDR_HAS_L1HDR(hdr));
2460 multilist_insert(new_state->arcs_list[buftype], hdr);
2462 if (GHOST_STATE(new_state)) {
2464 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2465 update_new = B_TRUE;
2467 arc_evictable_space_increment(hdr, new_state);
2471 ASSERT(!HDR_EMPTY(hdr));
2472 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2473 buf_hash_remove(hdr);
2475 /* adjust state sizes (ignore arc_l2c_only) */
2477 if (update_new && new_state != arc_l2c_only) {
2478 ASSERT(HDR_HAS_L1HDR(hdr));
2479 if (GHOST_STATE(new_state)) {
2483 * When moving a header to a ghost state, we first
2484 * remove all arc buffers. Thus, we'll have a
2485 * bufcnt of zero, and no arc buffer to use for
2486 * the reference. As a result, we use the arc
2487 * header pointer for the reference.
2489 (void) zfs_refcount_add_many(&new_state->arcs_size,
2490 HDR_GET_LSIZE(hdr), hdr);
2491 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2492 ASSERT(!HDR_HAS_RABD(hdr));
2494 uint32_t buffers = 0;
2497 * Each individual buffer holds a unique reference,
2498 * thus we must remove each of these references one
2501 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2502 buf = buf->b_next) {
2503 ASSERT3U(bufcnt, !=, 0);
2507 * When the arc_buf_t is sharing the data
2508 * block with the hdr, the owner of the
2509 * reference belongs to the hdr. Only
2510 * add to the refcount if the arc_buf_t is
2513 if (arc_buf_is_shared(buf))
2516 (void) zfs_refcount_add_many(
2517 &new_state->arcs_size,
2518 arc_buf_size(buf), buf);
2520 ASSERT3U(bufcnt, ==, buffers);
2522 if (hdr->b_l1hdr.b_pabd != NULL) {
2523 (void) zfs_refcount_add_many(
2524 &new_state->arcs_size,
2525 arc_hdr_size(hdr), hdr);
2528 if (HDR_HAS_RABD(hdr)) {
2529 (void) zfs_refcount_add_many(
2530 &new_state->arcs_size,
2531 HDR_GET_PSIZE(hdr), hdr);
2536 if (update_old && old_state != arc_l2c_only) {
2537 ASSERT(HDR_HAS_L1HDR(hdr));
2538 if (GHOST_STATE(old_state)) {
2540 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2541 ASSERT(!HDR_HAS_RABD(hdr));
2544 * When moving a header off of a ghost state,
2545 * the header will not contain any arc buffers.
2546 * We use the arc header pointer for the reference
2547 * which is exactly what we did when we put the
2548 * header on the ghost state.
2551 (void) zfs_refcount_remove_many(&old_state->arcs_size,
2552 HDR_GET_LSIZE(hdr), hdr);
2554 uint32_t buffers = 0;
2557 * Each individual buffer holds a unique reference,
2558 * thus we must remove each of these references one
2561 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2562 buf = buf->b_next) {
2563 ASSERT3U(bufcnt, !=, 0);
2567 * When the arc_buf_t is sharing the data
2568 * block with the hdr, the owner of the
2569 * reference belongs to the hdr. Only
2570 * add to the refcount if the arc_buf_t is
2573 if (arc_buf_is_shared(buf))
2576 (void) zfs_refcount_remove_many(
2577 &old_state->arcs_size, arc_buf_size(buf),
2580 ASSERT3U(bufcnt, ==, buffers);
2581 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2584 if (hdr->b_l1hdr.b_pabd != NULL) {
2585 (void) zfs_refcount_remove_many(
2586 &old_state->arcs_size, arc_hdr_size(hdr),
2590 if (HDR_HAS_RABD(hdr)) {
2591 (void) zfs_refcount_remove_many(
2592 &old_state->arcs_size, HDR_GET_PSIZE(hdr),
2598 if (HDR_HAS_L1HDR(hdr))
2599 hdr->b_l1hdr.b_state = new_state;
2602 * L2 headers should never be on the L2 state list since they don't
2603 * have L1 headers allocated.
2605 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2606 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2610 arc_space_consume(uint64_t space, arc_space_type_t type)
2612 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2617 case ARC_SPACE_DATA:
2618 aggsum_add(&astat_data_size, space);
2620 case ARC_SPACE_META:
2621 aggsum_add(&astat_metadata_size, space);
2623 case ARC_SPACE_BONUS:
2624 aggsum_add(&astat_bonus_size, space);
2626 case ARC_SPACE_DNODE:
2627 aggsum_add(&astat_dnode_size, space);
2629 case ARC_SPACE_DBUF:
2630 aggsum_add(&astat_dbuf_size, space);
2632 case ARC_SPACE_HDRS:
2633 aggsum_add(&astat_hdr_size, space);
2635 case ARC_SPACE_L2HDRS:
2636 aggsum_add(&astat_l2_hdr_size, space);
2638 case ARC_SPACE_ABD_CHUNK_WASTE:
2640 * Note: this includes space wasted by all scatter ABD's, not
2641 * just those allocated by the ARC. But the vast majority of
2642 * scatter ABD's come from the ARC, because other users are
2645 aggsum_add(&astat_abd_chunk_waste_size, space);
2649 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE)
2650 aggsum_add(&arc_meta_used, space);
2652 aggsum_add(&arc_size, space);
2656 arc_space_return(uint64_t space, arc_space_type_t type)
2658 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2663 case ARC_SPACE_DATA:
2664 aggsum_add(&astat_data_size, -space);
2666 case ARC_SPACE_META:
2667 aggsum_add(&astat_metadata_size, -space);
2669 case ARC_SPACE_BONUS:
2670 aggsum_add(&astat_bonus_size, -space);
2672 case ARC_SPACE_DNODE:
2673 aggsum_add(&astat_dnode_size, -space);
2675 case ARC_SPACE_DBUF:
2676 aggsum_add(&astat_dbuf_size, -space);
2678 case ARC_SPACE_HDRS:
2679 aggsum_add(&astat_hdr_size, -space);
2681 case ARC_SPACE_L2HDRS:
2682 aggsum_add(&astat_l2_hdr_size, -space);
2684 case ARC_SPACE_ABD_CHUNK_WASTE:
2685 aggsum_add(&astat_abd_chunk_waste_size, -space);
2689 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE) {
2690 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2692 * We use the upper bound here rather than the precise value
2693 * because the arc_meta_max value doesn't need to be
2694 * precise. It's only consumed by humans via arcstats.
2696 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2697 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2698 aggsum_add(&arc_meta_used, -space);
2701 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2702 aggsum_add(&arc_size, -space);
2706 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2707 * with the hdr's b_pabd.
2710 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2713 * The criteria for sharing a hdr's data are:
2714 * 1. the buffer is not encrypted
2715 * 2. the hdr's compression matches the buf's compression
2716 * 3. the hdr doesn't need to be byteswapped
2717 * 4. the hdr isn't already being shared
2718 * 5. the buf is either compressed or it is the last buf in the hdr list
2720 * Criterion #5 maintains the invariant that shared uncompressed
2721 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2722 * might ask, "if a compressed buf is allocated first, won't that be the
2723 * last thing in the list?", but in that case it's impossible to create
2724 * a shared uncompressed buf anyway (because the hdr must be compressed
2725 * to have the compressed buf). You might also think that #3 is
2726 * sufficient to make this guarantee, however it's possible
2727 * (specifically in the rare L2ARC write race mentioned in
2728 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2729 * is shareable, but wasn't at the time of its allocation. Rather than
2730 * allow a new shared uncompressed buf to be created and then shuffle
2731 * the list around to make it the last element, this simply disallows
2732 * sharing if the new buf isn't the first to be added.
2734 ASSERT3P(buf->b_hdr, ==, hdr);
2735 boolean_t hdr_compressed =
2736 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF;
2737 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2738 return (!ARC_BUF_ENCRYPTED(buf) &&
2739 buf_compressed == hdr_compressed &&
2740 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2741 !HDR_SHARED_DATA(hdr) &&
2742 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2746 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2747 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2748 * copy was made successfully, or an error code otherwise.
2751 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2752 void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth,
2753 boolean_t fill, arc_buf_t **ret)
2756 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2758 ASSERT(HDR_HAS_L1HDR(hdr));
2759 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2760 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2761 hdr->b_type == ARC_BUFC_METADATA);
2762 ASSERT3P(ret, !=, NULL);
2763 ASSERT3P(*ret, ==, NULL);
2764 IMPLY(encrypted, compressed);
2766 hdr->b_l1hdr.b_mru_hits = 0;
2767 hdr->b_l1hdr.b_mru_ghost_hits = 0;
2768 hdr->b_l1hdr.b_mfu_hits = 0;
2769 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
2770 hdr->b_l1hdr.b_l2_hits = 0;
2772 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2775 buf->b_next = hdr->b_l1hdr.b_buf;
2778 add_reference(hdr, tag);
2781 * We're about to change the hdr's b_flags. We must either
2782 * hold the hash_lock or be undiscoverable.
2784 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2787 * Only honor requests for compressed bufs if the hdr is actually
2788 * compressed. This must be overridden if the buffer is encrypted since
2789 * encrypted buffers cannot be decompressed.
2792 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2793 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2794 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2795 } else if (compressed &&
2796 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2797 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2798 flags |= ARC_FILL_COMPRESSED;
2803 flags |= ARC_FILL_NOAUTH;
2807 * If the hdr's data can be shared then we share the data buffer and
2808 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2809 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2810 * buffer to store the buf's data.
2812 * There are two additional restrictions here because we're sharing
2813 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2814 * actively involved in an L2ARC write, because if this buf is used by
2815 * an arc_write() then the hdr's data buffer will be released when the
2816 * write completes, even though the L2ARC write might still be using it.
2817 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2818 * need to be ABD-aware. It must be allocated via
2819 * zio_[data_]buf_alloc(), not as a page, because we need to be able
2820 * to abd_release_ownership_of_buf(), which isn't allowed on "linear
2821 * page" buffers because the ABD code needs to handle freeing them
2824 boolean_t can_share = arc_can_share(hdr, buf) &&
2825 !HDR_L2_WRITING(hdr) &&
2826 hdr->b_l1hdr.b_pabd != NULL &&
2827 abd_is_linear(hdr->b_l1hdr.b_pabd) &&
2828 !abd_is_linear_page(hdr->b_l1hdr.b_pabd);
2830 /* Set up b_data and sharing */
2832 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2833 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2834 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2837 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2838 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2840 VERIFY3P(buf->b_data, !=, NULL);
2842 hdr->b_l1hdr.b_buf = buf;
2843 hdr->b_l1hdr.b_bufcnt += 1;
2845 hdr->b_crypt_hdr.b_ebufcnt += 1;
2848 * If the user wants the data from the hdr, we need to either copy or
2849 * decompress the data.
2852 ASSERT3P(zb, !=, NULL);
2853 return (arc_buf_fill(buf, spa, zb, flags));
2859 static char *arc_onloan_tag = "onloan";
2862 arc_loaned_bytes_update(int64_t delta)
2864 atomic_add_64(&arc_loaned_bytes, delta);
2866 /* assert that it did not wrap around */
2867 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2871 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2872 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2873 * buffers must be returned to the arc before they can be used by the DMU or
2877 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2879 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2880 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2882 arc_loaned_bytes_update(arc_buf_size(buf));
2888 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2889 enum zio_compress compression_type, uint8_t complevel)
2891 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2892 psize, lsize, compression_type, complevel);
2894 arc_loaned_bytes_update(arc_buf_size(buf));
2900 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2901 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2902 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2903 enum zio_compress compression_type, uint8_t complevel)
2905 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2906 byteorder, salt, iv, mac, ot, psize, lsize, compression_type,
2909 atomic_add_64(&arc_loaned_bytes, psize);
2915 * Return a loaned arc buffer to the arc.
2918 arc_return_buf(arc_buf_t *buf, void *tag)
2920 arc_buf_hdr_t *hdr = buf->b_hdr;
2922 ASSERT3P(buf->b_data, !=, NULL);
2923 ASSERT(HDR_HAS_L1HDR(hdr));
2924 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2925 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2927 arc_loaned_bytes_update(-arc_buf_size(buf));
2930 /* Detach an arc_buf from a dbuf (tag) */
2932 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2934 arc_buf_hdr_t *hdr = buf->b_hdr;
2936 ASSERT3P(buf->b_data, !=, NULL);
2937 ASSERT(HDR_HAS_L1HDR(hdr));
2938 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2939 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2941 arc_loaned_bytes_update(arc_buf_size(buf));
2945 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2947 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2950 df->l2df_size = size;
2951 df->l2df_type = type;
2952 mutex_enter(&l2arc_free_on_write_mtx);
2953 list_insert_head(l2arc_free_on_write, df);
2954 mutex_exit(&l2arc_free_on_write_mtx);
2958 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
2960 arc_state_t *state = hdr->b_l1hdr.b_state;
2961 arc_buf_contents_t type = arc_buf_type(hdr);
2962 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
2964 /* protected by hash lock, if in the hash table */
2965 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2966 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2967 ASSERT(state != arc_anon && state != arc_l2c_only);
2969 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2972 (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr);
2973 if (type == ARC_BUFC_METADATA) {
2974 arc_space_return(size, ARC_SPACE_META);
2976 ASSERT(type == ARC_BUFC_DATA);
2977 arc_space_return(size, ARC_SPACE_DATA);
2981 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
2983 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2988 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2989 * data buffer, we transfer the refcount ownership to the hdr and update
2990 * the appropriate kstats.
2993 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2995 ASSERT(arc_can_share(hdr, buf));
2996 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2997 ASSERT(!ARC_BUF_ENCRYPTED(buf));
2998 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3001 * Start sharing the data buffer. We transfer the
3002 * refcount ownership to the hdr since it always owns
3003 * the refcount whenever an arc_buf_t is shared.
3005 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
3006 arc_hdr_size(hdr), buf, hdr);
3007 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
3008 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
3009 HDR_ISTYPE_METADATA(hdr));
3010 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3011 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3014 * Since we've transferred ownership to the hdr we need
3015 * to increment its compressed and uncompressed kstats and
3016 * decrement the overhead size.
3018 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3019 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3020 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3024 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3026 ASSERT(arc_buf_is_shared(buf));
3027 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3028 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3031 * We are no longer sharing this buffer so we need
3032 * to transfer its ownership to the rightful owner.
3034 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
3035 arc_hdr_size(hdr), hdr, buf);
3036 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3037 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3038 abd_put(hdr->b_l1hdr.b_pabd);
3039 hdr->b_l1hdr.b_pabd = NULL;
3040 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3043 * Since the buffer is no longer shared between
3044 * the arc buf and the hdr, count it as overhead.
3046 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3047 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3048 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3052 * Remove an arc_buf_t from the hdr's buf list and return the last
3053 * arc_buf_t on the list. If no buffers remain on the list then return
3057 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3059 ASSERT(HDR_HAS_L1HDR(hdr));
3060 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3062 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3063 arc_buf_t *lastbuf = NULL;
3066 * Remove the buf from the hdr list and locate the last
3067 * remaining buffer on the list.
3069 while (*bufp != NULL) {
3071 *bufp = buf->b_next;
3074 * If we've removed a buffer in the middle of
3075 * the list then update the lastbuf and update
3078 if (*bufp != NULL) {
3080 bufp = &(*bufp)->b_next;
3084 ASSERT3P(lastbuf, !=, buf);
3085 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3086 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3087 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3093 * Free up buf->b_data and pull the arc_buf_t off of the arc_buf_hdr_t's
3097 arc_buf_destroy_impl(arc_buf_t *buf)
3099 arc_buf_hdr_t *hdr = buf->b_hdr;
3102 * Free up the data associated with the buf but only if we're not
3103 * sharing this with the hdr. If we are sharing it with the hdr, the
3104 * hdr is responsible for doing the free.
3106 if (buf->b_data != NULL) {
3108 * We're about to change the hdr's b_flags. We must either
3109 * hold the hash_lock or be undiscoverable.
3111 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3113 arc_cksum_verify(buf);
3114 arc_buf_unwatch(buf);
3116 if (arc_buf_is_shared(buf)) {
3117 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3119 uint64_t size = arc_buf_size(buf);
3120 arc_free_data_buf(hdr, buf->b_data, size, buf);
3121 ARCSTAT_INCR(arcstat_overhead_size, -size);
3125 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3126 hdr->b_l1hdr.b_bufcnt -= 1;
3128 if (ARC_BUF_ENCRYPTED(buf)) {
3129 hdr->b_crypt_hdr.b_ebufcnt -= 1;
3132 * If we have no more encrypted buffers and we've
3133 * already gotten a copy of the decrypted data we can
3134 * free b_rabd to save some space.
3136 if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
3137 HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
3138 !HDR_IO_IN_PROGRESS(hdr)) {
3139 arc_hdr_free_abd(hdr, B_TRUE);
3144 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3146 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3148 * If the current arc_buf_t is sharing its data buffer with the
3149 * hdr, then reassign the hdr's b_pabd to share it with the new
3150 * buffer at the end of the list. The shared buffer is always
3151 * the last one on the hdr's buffer list.
3153 * There is an equivalent case for compressed bufs, but since
3154 * they aren't guaranteed to be the last buf in the list and
3155 * that is an exceedingly rare case, we just allow that space be
3156 * wasted temporarily. We must also be careful not to share
3157 * encrypted buffers, since they cannot be shared.
3159 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
3160 /* Only one buf can be shared at once */
3161 VERIFY(!arc_buf_is_shared(lastbuf));
3162 /* hdr is uncompressed so can't have compressed buf */
3163 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3165 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3166 arc_hdr_free_abd(hdr, B_FALSE);
3169 * We must setup a new shared block between the
3170 * last buffer and the hdr. The data would have
3171 * been allocated by the arc buf so we need to transfer
3172 * ownership to the hdr since it's now being shared.
3174 arc_share_buf(hdr, lastbuf);
3176 } else if (HDR_SHARED_DATA(hdr)) {
3178 * Uncompressed shared buffers are always at the end
3179 * of the list. Compressed buffers don't have the
3180 * same requirements. This makes it hard to
3181 * simply assert that the lastbuf is shared so
3182 * we rely on the hdr's compression flags to determine
3183 * if we have a compressed, shared buffer.
3185 ASSERT3P(lastbuf, !=, NULL);
3186 ASSERT(arc_buf_is_shared(lastbuf) ||
3187 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3191 * Free the checksum if we're removing the last uncompressed buf from
3194 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3195 arc_cksum_free(hdr);
3198 /* clean up the buf */
3200 kmem_cache_free(buf_cache, buf);
3204 arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, int alloc_flags)
3207 boolean_t alloc_rdata = ((alloc_flags & ARC_HDR_ALLOC_RDATA) != 0);
3208 boolean_t do_adapt = ((alloc_flags & ARC_HDR_DO_ADAPT) != 0);
3210 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3211 ASSERT(HDR_HAS_L1HDR(hdr));
3212 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3213 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3216 size = HDR_GET_PSIZE(hdr);
3217 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3218 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr,
3220 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3221 ARCSTAT_INCR(arcstat_raw_size, size);
3223 size = arc_hdr_size(hdr);
3224 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3225 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr,
3227 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3230 ARCSTAT_INCR(arcstat_compressed_size, size);
3231 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3235 arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3237 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3239 ASSERT(HDR_HAS_L1HDR(hdr));
3240 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3241 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3244 * If the hdr is currently being written to the l2arc then
3245 * we defer freeing the data by adding it to the l2arc_free_on_write
3246 * list. The l2arc will free the data once it's finished
3247 * writing it to the l2arc device.
3249 if (HDR_L2_WRITING(hdr)) {
3250 arc_hdr_free_on_write(hdr, free_rdata);
3251 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3252 } else if (free_rdata) {
3253 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3255 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr);
3259 hdr->b_crypt_hdr.b_rabd = NULL;
3260 ARCSTAT_INCR(arcstat_raw_size, -size);
3262 hdr->b_l1hdr.b_pabd = NULL;
3265 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3266 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3268 ARCSTAT_INCR(arcstat_compressed_size, -size);
3269 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3272 static arc_buf_hdr_t *
3273 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3274 boolean_t protected, enum zio_compress compression_type, uint8_t complevel,
3275 arc_buf_contents_t type, boolean_t alloc_rdata)
3278 int flags = ARC_HDR_DO_ADAPT;
3280 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3282 hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
3284 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3286 flags |= alloc_rdata ? ARC_HDR_ALLOC_RDATA : 0;
3288 ASSERT(HDR_EMPTY(hdr));
3289 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3290 HDR_SET_PSIZE(hdr, psize);
3291 HDR_SET_LSIZE(hdr, lsize);
3295 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3296 arc_hdr_set_compress(hdr, compression_type);
3297 hdr->b_complevel = complevel;
3299 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3301 hdr->b_l1hdr.b_state = arc_anon;
3302 hdr->b_l1hdr.b_arc_access = 0;
3303 hdr->b_l1hdr.b_bufcnt = 0;
3304 hdr->b_l1hdr.b_buf = NULL;
3307 * Allocate the hdr's buffer. This will contain either
3308 * the compressed or uncompressed data depending on the block
3309 * it references and compressed arc enablement.
3311 arc_hdr_alloc_abd(hdr, flags);
3312 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3318 * Transition between the two allocation states for the arc_buf_hdr struct.
3319 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3320 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3321 * version is used when a cache buffer is only in the L2ARC in order to reduce
3324 static arc_buf_hdr_t *
3325 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3327 ASSERT(HDR_HAS_L2HDR(hdr));
3329 arc_buf_hdr_t *nhdr;
3330 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3332 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3333 (old == hdr_l2only_cache && new == hdr_full_cache));
3336 * if the caller wanted a new full header and the header is to be
3337 * encrypted we will actually allocate the header from the full crypt
3338 * cache instead. The same applies to freeing from the old cache.
3340 if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
3341 new = hdr_full_crypt_cache;
3342 if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
3343 old = hdr_full_crypt_cache;
3345 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3347 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3348 buf_hash_remove(hdr);
3350 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3352 if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
3353 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3355 * arc_access and arc_change_state need to be aware that a
3356 * header has just come out of L2ARC, so we set its state to
3357 * l2c_only even though it's about to change.
3359 nhdr->b_l1hdr.b_state = arc_l2c_only;
3361 /* Verify previous threads set to NULL before freeing */
3362 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3363 ASSERT(!HDR_HAS_RABD(hdr));
3365 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3366 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3367 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3370 * If we've reached here, We must have been called from
3371 * arc_evict_hdr(), as such we should have already been
3372 * removed from any ghost list we were previously on
3373 * (which protects us from racing with arc_evict_state),
3374 * thus no locking is needed during this check.
3376 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3379 * A buffer must not be moved into the arc_l2c_only
3380 * state if it's not finished being written out to the
3381 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3382 * might try to be accessed, even though it was removed.
3384 VERIFY(!HDR_L2_WRITING(hdr));
3385 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3386 ASSERT(!HDR_HAS_RABD(hdr));
3388 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3391 * The header has been reallocated so we need to re-insert it into any
3394 (void) buf_hash_insert(nhdr, NULL);
3396 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3398 mutex_enter(&dev->l2ad_mtx);
3401 * We must place the realloc'ed header back into the list at
3402 * the same spot. Otherwise, if it's placed earlier in the list,
3403 * l2arc_write_buffers() could find it during the function's
3404 * write phase, and try to write it out to the l2arc.
3406 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3407 list_remove(&dev->l2ad_buflist, hdr);
3409 mutex_exit(&dev->l2ad_mtx);
3412 * Since we're using the pointer address as the tag when
3413 * incrementing and decrementing the l2ad_alloc refcount, we
3414 * must remove the old pointer (that we're about to destroy) and
3415 * add the new pointer to the refcount. Otherwise we'd remove
3416 * the wrong pointer address when calling arc_hdr_destroy() later.
3419 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
3420 arc_hdr_size(hdr), hdr);
3421 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
3422 arc_hdr_size(nhdr), nhdr);
3424 buf_discard_identity(hdr);
3425 kmem_cache_free(old, hdr);
3431 * This function allows an L1 header to be reallocated as a crypt
3432 * header and vice versa. If we are going to a crypt header, the
3433 * new fields will be zeroed out.
3435 static arc_buf_hdr_t *
3436 arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
3438 arc_buf_hdr_t *nhdr;
3440 kmem_cache_t *ncache, *ocache;
3441 unsigned nsize, osize;
3444 * This function requires that hdr is in the arc_anon state.
3445 * Therefore it won't have any L2ARC data for us to worry
3448 ASSERT(HDR_HAS_L1HDR(hdr));
3449 ASSERT(!HDR_HAS_L2HDR(hdr));
3450 ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
3451 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3452 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3453 ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node));
3454 ASSERT3P(hdr->b_hash_next, ==, NULL);
3457 ncache = hdr_full_crypt_cache;
3458 nsize = sizeof (hdr->b_crypt_hdr);
3459 ocache = hdr_full_cache;
3460 osize = HDR_FULL_SIZE;
3462 ncache = hdr_full_cache;
3463 nsize = HDR_FULL_SIZE;
3464 ocache = hdr_full_crypt_cache;
3465 osize = sizeof (hdr->b_crypt_hdr);
3468 nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);
3471 * Copy all members that aren't locks or condvars to the new header.
3472 * No lists are pointing to us (as we asserted above), so we don't
3473 * need to worry about the list nodes.
3475 nhdr->b_dva = hdr->b_dva;
3476 nhdr->b_birth = hdr->b_birth;
3477 nhdr->b_type = hdr->b_type;
3478 nhdr->b_flags = hdr->b_flags;
3479 nhdr->b_psize = hdr->b_psize;
3480 nhdr->b_lsize = hdr->b_lsize;
3481 nhdr->b_spa = hdr->b_spa;
3482 nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
3483 nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
3484 nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
3485 nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
3486 nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
3487 nhdr->b_l1hdr.b_mru_hits = hdr->b_l1hdr.b_mru_hits;
3488 nhdr->b_l1hdr.b_mru_ghost_hits = hdr->b_l1hdr.b_mru_ghost_hits;
3489 nhdr->b_l1hdr.b_mfu_hits = hdr->b_l1hdr.b_mfu_hits;
3490 nhdr->b_l1hdr.b_mfu_ghost_hits = hdr->b_l1hdr.b_mfu_ghost_hits;
3491 nhdr->b_l1hdr.b_l2_hits = hdr->b_l1hdr.b_l2_hits;
3492 nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
3493 nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
3496 * This zfs_refcount_add() exists only to ensure that the individual
3497 * arc buffers always point to a header that is referenced, avoiding
3498 * a small race condition that could trigger ASSERTs.
3500 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
3501 nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
3502 for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) {
3503 mutex_enter(&buf->b_evict_lock);
3505 mutex_exit(&buf->b_evict_lock);
3508 zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
3509 (void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
3510 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3513 arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
3515 arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
3518 /* unset all members of the original hdr */
3519 bzero(&hdr->b_dva, sizeof (dva_t));
3521 hdr->b_type = ARC_BUFC_INVALID;
3526 hdr->b_l1hdr.b_freeze_cksum = NULL;
3527 hdr->b_l1hdr.b_buf = NULL;
3528 hdr->b_l1hdr.b_bufcnt = 0;
3529 hdr->b_l1hdr.b_byteswap = 0;
3530 hdr->b_l1hdr.b_state = NULL;
3531 hdr->b_l1hdr.b_arc_access = 0;
3532 hdr->b_l1hdr.b_mru_hits = 0;
3533 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3534 hdr->b_l1hdr.b_mfu_hits = 0;
3535 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3536 hdr->b_l1hdr.b_l2_hits = 0;
3537 hdr->b_l1hdr.b_acb = NULL;
3538 hdr->b_l1hdr.b_pabd = NULL;
3540 if (ocache == hdr_full_crypt_cache) {
3541 ASSERT(!HDR_HAS_RABD(hdr));
3542 hdr->b_crypt_hdr.b_ot = DMU_OT_NONE;
3543 hdr->b_crypt_hdr.b_ebufcnt = 0;
3544 hdr->b_crypt_hdr.b_dsobj = 0;
3545 bzero(hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3546 bzero(hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3547 bzero(hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3550 buf_discard_identity(hdr);
3551 kmem_cache_free(ocache, hdr);
3557 * This function is used by the send / receive code to convert a newly
3558 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3559 * is also used to allow the root objset block to be updated without altering
3560 * its embedded MACs. Both block types will always be uncompressed so we do not
3561 * have to worry about compression type or psize.
3564 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3565 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3568 arc_buf_hdr_t *hdr = buf->b_hdr;
3570 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3571 ASSERT(HDR_HAS_L1HDR(hdr));
3572 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3574 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3575 if (!HDR_PROTECTED(hdr))
3576 hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
3577 hdr->b_crypt_hdr.b_dsobj = dsobj;
3578 hdr->b_crypt_hdr.b_ot = ot;
3579 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3580 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3581 if (!arc_hdr_has_uncompressed_buf(hdr))
3582 arc_cksum_free(hdr);
3585 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3587 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3589 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3593 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3594 * The buf is returned thawed since we expect the consumer to modify it.
3597 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3599 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3600 B_FALSE, ZIO_COMPRESS_OFF, 0, type, B_FALSE);
3602 arc_buf_t *buf = NULL;
3603 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3604 B_FALSE, B_FALSE, &buf));
3611 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3612 * for bufs containing metadata.
3615 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3616 enum zio_compress compression_type, uint8_t complevel)
3618 ASSERT3U(lsize, >, 0);
3619 ASSERT3U(lsize, >=, psize);
3620 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3621 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3623 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3624 B_FALSE, compression_type, complevel, ARC_BUFC_DATA, B_FALSE);
3626 arc_buf_t *buf = NULL;
3627 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3628 B_TRUE, B_FALSE, B_FALSE, &buf));
3630 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3632 if (!arc_buf_is_shared(buf)) {
3634 * To ensure that the hdr has the correct data in it if we call
3635 * arc_untransform() on this buf before it's been written to
3636 * disk, it's easiest if we just set up sharing between the
3639 arc_hdr_free_abd(hdr, B_FALSE);
3640 arc_share_buf(hdr, buf);
3647 arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder,
3648 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
3649 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3650 enum zio_compress compression_type, uint8_t complevel)
3654 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3655 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3657 ASSERT3U(lsize, >, 0);
3658 ASSERT3U(lsize, >=, psize);
3659 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3660 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3662 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3663 compression_type, complevel, type, B_TRUE);
3665 hdr->b_crypt_hdr.b_dsobj = dsobj;
3666 hdr->b_crypt_hdr.b_ot = ot;
3667 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3668 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3669 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3670 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3671 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3674 * This buffer will be considered encrypted even if the ot is not an
3675 * encrypted type. It will become authenticated instead in
3676 * arc_write_ready().
3679 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3680 B_FALSE, B_FALSE, &buf));
3682 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3688 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3690 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3691 l2arc_dev_t *dev = l2hdr->b_dev;
3692 uint64_t psize = HDR_GET_PSIZE(hdr);
3693 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3695 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3696 ASSERT(HDR_HAS_L2HDR(hdr));
3698 list_remove(&dev->l2ad_buflist, hdr);
3700 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3701 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3703 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3705 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3707 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3711 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3713 if (HDR_HAS_L1HDR(hdr)) {
3714 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3715 hdr->b_l1hdr.b_bufcnt > 0);
3716 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3717 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3719 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3720 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3722 if (HDR_HAS_L2HDR(hdr)) {
3723 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3724 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3727 mutex_enter(&dev->l2ad_mtx);
3730 * Even though we checked this conditional above, we
3731 * need to check this again now that we have the
3732 * l2ad_mtx. This is because we could be racing with
3733 * another thread calling l2arc_evict() which might have
3734 * destroyed this header's L2 portion as we were waiting
3735 * to acquire the l2ad_mtx. If that happens, we don't
3736 * want to re-destroy the header's L2 portion.
3738 if (HDR_HAS_L2HDR(hdr))
3739 arc_hdr_l2hdr_destroy(hdr);
3742 mutex_exit(&dev->l2ad_mtx);
3746 * The header's identify can only be safely discarded once it is no
3747 * longer discoverable. This requires removing it from the hash table
3748 * and the l2arc header list. After this point the hash lock can not
3749 * be used to protect the header.
3751 if (!HDR_EMPTY(hdr))
3752 buf_discard_identity(hdr);
3754 if (HDR_HAS_L1HDR(hdr)) {
3755 arc_cksum_free(hdr);
3757 while (hdr->b_l1hdr.b_buf != NULL)
3758 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3760 if (hdr->b_l1hdr.b_pabd != NULL)
3761 arc_hdr_free_abd(hdr, B_FALSE);
3763 if (HDR_HAS_RABD(hdr))
3764 arc_hdr_free_abd(hdr, B_TRUE);
3767 ASSERT3P(hdr->b_hash_next, ==, NULL);
3768 if (HDR_HAS_L1HDR(hdr)) {
3769 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3770 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3772 if (!HDR_PROTECTED(hdr)) {
3773 kmem_cache_free(hdr_full_cache, hdr);
3775 kmem_cache_free(hdr_full_crypt_cache, hdr);
3778 kmem_cache_free(hdr_l2only_cache, hdr);
3783 arc_buf_destroy(arc_buf_t *buf, void* tag)
3785 arc_buf_hdr_t *hdr = buf->b_hdr;
3787 if (hdr->b_l1hdr.b_state == arc_anon) {
3788 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3789 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3790 VERIFY0(remove_reference(hdr, NULL, tag));
3791 arc_hdr_destroy(hdr);
3795 kmutex_t *hash_lock = HDR_LOCK(hdr);
3796 mutex_enter(hash_lock);
3798 ASSERT3P(hdr, ==, buf->b_hdr);
3799 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3800 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3801 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3802 ASSERT3P(buf->b_data, !=, NULL);
3804 (void) remove_reference(hdr, hash_lock, tag);
3805 arc_buf_destroy_impl(buf);
3806 mutex_exit(hash_lock);
3810 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3811 * state of the header is dependent on its state prior to entering this
3812 * function. The following transitions are possible:
3814 * - arc_mru -> arc_mru_ghost
3815 * - arc_mfu -> arc_mfu_ghost
3816 * - arc_mru_ghost -> arc_l2c_only
3817 * - arc_mru_ghost -> deleted
3818 * - arc_mfu_ghost -> arc_l2c_only
3819 * - arc_mfu_ghost -> deleted
3822 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3824 arc_state_t *evicted_state, *state;
3825 int64_t bytes_evicted = 0;
3826 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3827 arc_min_prescient_prefetch_ms : arc_min_prefetch_ms;
3829 ASSERT(MUTEX_HELD(hash_lock));
3830 ASSERT(HDR_HAS_L1HDR(hdr));
3832 state = hdr->b_l1hdr.b_state;
3833 if (GHOST_STATE(state)) {
3834 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3835 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3838 * l2arc_write_buffers() relies on a header's L1 portion
3839 * (i.e. its b_pabd field) during it's write phase.
3840 * Thus, we cannot push a header onto the arc_l2c_only
3841 * state (removing its L1 piece) until the header is
3842 * done being written to the l2arc.
3844 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3845 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3846 return (bytes_evicted);
3849 ARCSTAT_BUMP(arcstat_deleted);
3850 bytes_evicted += HDR_GET_LSIZE(hdr);
3852 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3854 if (HDR_HAS_L2HDR(hdr)) {
3855 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3856 ASSERT(!HDR_HAS_RABD(hdr));
3858 * This buffer is cached on the 2nd Level ARC;
3859 * don't destroy the header.
3861 arc_change_state(arc_l2c_only, hdr, hash_lock);
3863 * dropping from L1+L2 cached to L2-only,
3864 * realloc to remove the L1 header.
3866 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3869 arc_change_state(arc_anon, hdr, hash_lock);
3870 arc_hdr_destroy(hdr);
3872 return (bytes_evicted);
3875 ASSERT(state == arc_mru || state == arc_mfu);
3876 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3878 /* prefetch buffers have a minimum lifespan */
3879 if (HDR_IO_IN_PROGRESS(hdr) ||
3880 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3881 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3882 MSEC_TO_TICK(min_lifetime))) {
3883 ARCSTAT_BUMP(arcstat_evict_skip);
3884 return (bytes_evicted);
3887 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3888 while (hdr->b_l1hdr.b_buf) {
3889 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3890 if (!mutex_tryenter(&buf->b_evict_lock)) {
3891 ARCSTAT_BUMP(arcstat_mutex_miss);
3894 if (buf->b_data != NULL)
3895 bytes_evicted += HDR_GET_LSIZE(hdr);
3896 mutex_exit(&buf->b_evict_lock);
3897 arc_buf_destroy_impl(buf);
3900 if (HDR_HAS_L2HDR(hdr)) {
3901 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3903 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3904 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3905 HDR_GET_LSIZE(hdr));
3907 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3908 HDR_GET_LSIZE(hdr));
3912 if (hdr->b_l1hdr.b_bufcnt == 0) {
3913 arc_cksum_free(hdr);
3915 bytes_evicted += arc_hdr_size(hdr);
3918 * If this hdr is being evicted and has a compressed
3919 * buffer then we discard it here before we change states.
3920 * This ensures that the accounting is updated correctly
3921 * in arc_free_data_impl().
3923 if (hdr->b_l1hdr.b_pabd != NULL)
3924 arc_hdr_free_abd(hdr, B_FALSE);
3926 if (HDR_HAS_RABD(hdr))
3927 arc_hdr_free_abd(hdr, B_TRUE);
3929 arc_change_state(evicted_state, hdr, hash_lock);
3930 ASSERT(HDR_IN_HASH_TABLE(hdr));
3931 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3932 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3935 return (bytes_evicted);
3939 arc_set_need_free(void)
3941 ASSERT(MUTEX_HELD(&arc_evict_lock));
3942 int64_t remaining = arc_free_memory() - arc_sys_free / 2;
3943 arc_evict_waiter_t *aw = list_tail(&arc_evict_waiters);
3945 arc_need_free = MAX(-remaining, 0);
3948 MAX(-remaining, (int64_t)(aw->aew_count - arc_evict_count));
3953 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3954 uint64_t spa, int64_t bytes)
3956 multilist_sublist_t *mls;
3957 uint64_t bytes_evicted = 0;
3959 kmutex_t *hash_lock;
3960 int evict_count = 0;
3962 ASSERT3P(marker, !=, NULL);
3963 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3965 mls = multilist_sublist_lock(ml, idx);
3967 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3968 hdr = multilist_sublist_prev(mls, marker)) {
3969 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3970 (evict_count >= zfs_arc_evict_batch_limit))
3974 * To keep our iteration location, move the marker
3975 * forward. Since we're not holding hdr's hash lock, we
3976 * must be very careful and not remove 'hdr' from the
3977 * sublist. Otherwise, other consumers might mistake the
3978 * 'hdr' as not being on a sublist when they call the
3979 * multilist_link_active() function (they all rely on
3980 * the hash lock protecting concurrent insertions and
3981 * removals). multilist_sublist_move_forward() was
3982 * specifically implemented to ensure this is the case
3983 * (only 'marker' will be removed and re-inserted).
3985 multilist_sublist_move_forward(mls, marker);
3988 * The only case where the b_spa field should ever be
3989 * zero, is the marker headers inserted by
3990 * arc_evict_state(). It's possible for multiple threads
3991 * to be calling arc_evict_state() concurrently (e.g.
3992 * dsl_pool_close() and zio_inject_fault()), so we must
3993 * skip any markers we see from these other threads.
3995 if (hdr->b_spa == 0)
3998 /* we're only interested in evicting buffers of a certain spa */
3999 if (spa != 0 && hdr->b_spa != spa) {
4000 ARCSTAT_BUMP(arcstat_evict_skip);
4004 hash_lock = HDR_LOCK(hdr);
4007 * We aren't calling this function from any code path
4008 * that would already be holding a hash lock, so we're
4009 * asserting on this assumption to be defensive in case
4010 * this ever changes. Without this check, it would be
4011 * possible to incorrectly increment arcstat_mutex_miss
4012 * below (e.g. if the code changed such that we called
4013 * this function with a hash lock held).
4015 ASSERT(!MUTEX_HELD(hash_lock));
4017 if (mutex_tryenter(hash_lock)) {
4018 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
4019 mutex_exit(hash_lock);
4021 bytes_evicted += evicted;
4024 * If evicted is zero, arc_evict_hdr() must have
4025 * decided to skip this header, don't increment
4026 * evict_count in this case.
4032 ARCSTAT_BUMP(arcstat_mutex_miss);
4036 multilist_sublist_unlock(mls);
4039 * Increment the count of evicted bytes, and wake up any threads that
4040 * are waiting for the count to reach this value. Since the list is
4041 * ordered by ascending aew_count, we pop off the beginning of the
4042 * list until we reach the end, or a waiter that's past the current
4043 * "count". Doing this outside the loop reduces the number of times
4044 * we need to acquire the global arc_evict_lock.
4046 * Only wake when there's sufficient free memory in the system
4047 * (specifically, arc_sys_free/2, which by default is a bit more than
4048 * 1/64th of RAM). See the comments in arc_wait_for_eviction().
4050 mutex_enter(&arc_evict_lock);
4051 arc_evict_count += bytes_evicted;
4053 if ((int64_t)(arc_free_memory() - arc_sys_free / 2) > 0) {
4054 arc_evict_waiter_t *aw;
4055 while ((aw = list_head(&arc_evict_waiters)) != NULL &&
4056 aw->aew_count <= arc_evict_count) {
4057 list_remove(&arc_evict_waiters, aw);
4058 cv_broadcast(&aw->aew_cv);
4061 arc_set_need_free();
4062 mutex_exit(&arc_evict_lock);
4065 * If the ARC size is reduced from arc_c_max to arc_c_min (especially
4066 * if the average cached block is small), eviction can be on-CPU for
4067 * many seconds. To ensure that other threads that may be bound to
4068 * this CPU are able to make progress, make a voluntary preemption
4073 return (bytes_evicted);
4077 * Evict buffers from the given arc state, until we've removed the
4078 * specified number of bytes. Move the removed buffers to the
4079 * appropriate evict state.
4081 * This function makes a "best effort". It skips over any buffers
4082 * it can't get a hash_lock on, and so, may not catch all candidates.
4083 * It may also return without evicting as much space as requested.
4085 * If bytes is specified using the special value ARC_EVICT_ALL, this
4086 * will evict all available (i.e. unlocked and evictable) buffers from
4087 * the given arc state; which is used by arc_flush().
4090 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
4091 arc_buf_contents_t type)
4093 uint64_t total_evicted = 0;
4094 multilist_t *ml = state->arcs_list[type];
4096 arc_buf_hdr_t **markers;
4098 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
4100 num_sublists = multilist_get_num_sublists(ml);
4103 * If we've tried to evict from each sublist, made some
4104 * progress, but still have not hit the target number of bytes
4105 * to evict, we want to keep trying. The markers allow us to
4106 * pick up where we left off for each individual sublist, rather
4107 * than starting from the tail each time.
4109 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
4110 for (int i = 0; i < num_sublists; i++) {
4111 multilist_sublist_t *mls;
4113 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
4116 * A b_spa of 0 is used to indicate that this header is
4117 * a marker. This fact is used in arc_evict_type() and
4118 * arc_evict_state_impl().
4120 markers[i]->b_spa = 0;
4122 mls = multilist_sublist_lock(ml, i);
4123 multilist_sublist_insert_tail(mls, markers[i]);
4124 multilist_sublist_unlock(mls);
4128 * While we haven't hit our target number of bytes to evict, or
4129 * we're evicting all available buffers.
4131 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
4132 int sublist_idx = multilist_get_random_index(ml);
4133 uint64_t scan_evicted = 0;
4136 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
4137 * Request that 10% of the LRUs be scanned by the superblock
4140 if (type == ARC_BUFC_DATA && aggsum_compare(&astat_dnode_size,
4141 arc_dnode_size_limit) > 0) {
4142 arc_prune_async((aggsum_upper_bound(&astat_dnode_size) -
4143 arc_dnode_size_limit) / sizeof (dnode_t) /
4144 zfs_arc_dnode_reduce_percent);
4148 * Start eviction using a randomly selected sublist,
4149 * this is to try and evenly balance eviction across all
4150 * sublists. Always starting at the same sublist
4151 * (e.g. index 0) would cause evictions to favor certain
4152 * sublists over others.
4154 for (int i = 0; i < num_sublists; i++) {
4155 uint64_t bytes_remaining;
4156 uint64_t bytes_evicted;
4158 if (bytes == ARC_EVICT_ALL)
4159 bytes_remaining = ARC_EVICT_ALL;
4160 else if (total_evicted < bytes)
4161 bytes_remaining = bytes - total_evicted;
4165 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4166 markers[sublist_idx], spa, bytes_remaining);
4168 scan_evicted += bytes_evicted;
4169 total_evicted += bytes_evicted;
4171 /* we've reached the end, wrap to the beginning */
4172 if (++sublist_idx >= num_sublists)
4177 * If we didn't evict anything during this scan, we have
4178 * no reason to believe we'll evict more during another
4179 * scan, so break the loop.
4181 if (scan_evicted == 0) {
4182 /* This isn't possible, let's make that obvious */
4183 ASSERT3S(bytes, !=, 0);
4186 * When bytes is ARC_EVICT_ALL, the only way to
4187 * break the loop is when scan_evicted is zero.
4188 * In that case, we actually have evicted enough,
4189 * so we don't want to increment the kstat.
4191 if (bytes != ARC_EVICT_ALL) {
4192 ASSERT3S(total_evicted, <, bytes);
4193 ARCSTAT_BUMP(arcstat_evict_not_enough);
4200 for (int i = 0; i < num_sublists; i++) {
4201 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4202 multilist_sublist_remove(mls, markers[i]);
4203 multilist_sublist_unlock(mls);
4205 kmem_cache_free(hdr_full_cache, markers[i]);
4207 kmem_free(markers, sizeof (*markers) * num_sublists);
4209 return (total_evicted);
4213 * Flush all "evictable" data of the given type from the arc state
4214 * specified. This will not evict any "active" buffers (i.e. referenced).
4216 * When 'retry' is set to B_FALSE, the function will make a single pass
4217 * over the state and evict any buffers that it can. Since it doesn't
4218 * continually retry the eviction, it might end up leaving some buffers
4219 * in the ARC due to lock misses.
4221 * When 'retry' is set to B_TRUE, the function will continually retry the
4222 * eviction until *all* evictable buffers have been removed from the
4223 * state. As a result, if concurrent insertions into the state are
4224 * allowed (e.g. if the ARC isn't shutting down), this function might
4225 * wind up in an infinite loop, continually trying to evict buffers.
4228 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4231 uint64_t evicted = 0;
4233 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4234 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4244 * Evict the specified number of bytes from the state specified,
4245 * restricting eviction to the spa and type given. This function
4246 * prevents us from trying to evict more from a state's list than
4247 * is "evictable", and to skip evicting altogether when passed a
4248 * negative value for "bytes". In contrast, arc_evict_state() will
4249 * evict everything it can, when passed a negative value for "bytes".
4252 arc_evict_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4253 arc_buf_contents_t type)
4257 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4258 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4260 return (arc_evict_state(state, spa, delta, type));
4267 * The goal of this function is to evict enough meta data buffers from the
4268 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4269 * more complicated than it appears because it is common for data buffers
4270 * to have holds on meta data buffers. In addition, dnode meta data buffers
4271 * will be held by the dnodes in the block preventing them from being freed.
4272 * This means we can't simply traverse the ARC and expect to always find
4273 * enough unheld meta data buffer to release.
4275 * Therefore, this function has been updated to make alternating passes
4276 * over the ARC releasing data buffers and then newly unheld meta data
4277 * buffers. This ensures forward progress is maintained and meta_used
4278 * will decrease. Normally this is sufficient, but if required the ARC
4279 * will call the registered prune callbacks causing dentry and inodes to
4280 * be dropped from the VFS cache. This will make dnode meta data buffers
4281 * available for reclaim.
4284 arc_evict_meta_balanced(uint64_t meta_used)
4286 int64_t delta, prune = 0, adjustmnt;
4287 uint64_t total_evicted = 0;
4288 arc_buf_contents_t type = ARC_BUFC_DATA;
4289 int restarts = MAX(zfs_arc_meta_adjust_restarts, 0);
4293 * This slightly differs than the way we evict from the mru in
4294 * arc_evict because we don't have a "target" value (i.e. no
4295 * "meta" arc_p). As a result, I think we can completely
4296 * cannibalize the metadata in the MRU before we evict the
4297 * metadata from the MFU. I think we probably need to implement a
4298 * "metadata arc_p" value to do this properly.
4300 adjustmnt = meta_used - arc_meta_limit;
4302 if (adjustmnt > 0 &&
4303 zfs_refcount_count(&arc_mru->arcs_esize[type]) > 0) {
4304 delta = MIN(zfs_refcount_count(&arc_mru->arcs_esize[type]),
4306 total_evicted += arc_evict_impl(arc_mru, 0, delta, type);
4311 * We can't afford to recalculate adjustmnt here. If we do,
4312 * new metadata buffers can sneak into the MRU or ANON lists,
4313 * thus penalize the MFU metadata. Although the fudge factor is
4314 * small, it has been empirically shown to be significant for
4315 * certain workloads (e.g. creating many empty directories). As
4316 * such, we use the original calculation for adjustmnt, and
4317 * simply decrement the amount of data evicted from the MRU.
4320 if (adjustmnt > 0 &&
4321 zfs_refcount_count(&arc_mfu->arcs_esize[type]) > 0) {
4322 delta = MIN(zfs_refcount_count(&arc_mfu->arcs_esize[type]),
4324 total_evicted += arc_evict_impl(arc_mfu, 0, delta, type);
4327 adjustmnt = meta_used - arc_meta_limit;
4329 if (adjustmnt > 0 &&
4330 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) {
4331 delta = MIN(adjustmnt,
4332 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]));
4333 total_evicted += arc_evict_impl(arc_mru_ghost, 0, delta, type);
4337 if (adjustmnt > 0 &&
4338 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) {
4339 delta = MIN(adjustmnt,
4340 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]));
4341 total_evicted += arc_evict_impl(arc_mfu_ghost, 0, delta, type);
4345 * If after attempting to make the requested adjustment to the ARC
4346 * the meta limit is still being exceeded then request that the
4347 * higher layers drop some cached objects which have holds on ARC
4348 * meta buffers. Requests to the upper layers will be made with
4349 * increasingly large scan sizes until the ARC is below the limit.
4351 if (meta_used > arc_meta_limit) {
4352 if (type == ARC_BUFC_DATA) {
4353 type = ARC_BUFC_METADATA;
4355 type = ARC_BUFC_DATA;
4357 if (zfs_arc_meta_prune) {
4358 prune += zfs_arc_meta_prune;
4359 arc_prune_async(prune);
4368 return (total_evicted);
4372 * Evict metadata buffers from the cache, such that arc_meta_used is
4373 * capped by the arc_meta_limit tunable.
4376 arc_evict_meta_only(uint64_t meta_used)
4378 uint64_t total_evicted = 0;
4382 * If we're over the meta limit, we want to evict enough
4383 * metadata to get back under the meta limit. We don't want to
4384 * evict so much that we drop the MRU below arc_p, though. If
4385 * we're over the meta limit more than we're over arc_p, we
4386 * evict some from the MRU here, and some from the MFU below.
4388 target = MIN((int64_t)(meta_used - arc_meta_limit),
4389 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4390 zfs_refcount_count(&arc_mru->arcs_size) - arc_p));
4392 total_evicted += arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4395 * Similar to the above, we want to evict enough bytes to get us
4396 * below the meta limit, but not so much as to drop us below the
4397 * space allotted to the MFU (which is defined as arc_c - arc_p).
4399 target = MIN((int64_t)(meta_used - arc_meta_limit),
4400 (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) -
4403 total_evicted += arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4405 return (total_evicted);
4409 arc_evict_meta(uint64_t meta_used)
4411 if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY)
4412 return (arc_evict_meta_only(meta_used));
4414 return (arc_evict_meta_balanced(meta_used));
4418 * Return the type of the oldest buffer in the given arc state
4420 * This function will select a random sublist of type ARC_BUFC_DATA and
4421 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4422 * is compared, and the type which contains the "older" buffer will be
4425 static arc_buf_contents_t
4426 arc_evict_type(arc_state_t *state)
4428 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4429 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4430 int data_idx = multilist_get_random_index(data_ml);
4431 int meta_idx = multilist_get_random_index(meta_ml);
4432 multilist_sublist_t *data_mls;
4433 multilist_sublist_t *meta_mls;
4434 arc_buf_contents_t type;
4435 arc_buf_hdr_t *data_hdr;
4436 arc_buf_hdr_t *meta_hdr;
4439 * We keep the sublist lock until we're finished, to prevent
4440 * the headers from being destroyed via arc_evict_state().
4442 data_mls = multilist_sublist_lock(data_ml, data_idx);
4443 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4446 * These two loops are to ensure we skip any markers that
4447 * might be at the tail of the lists due to arc_evict_state().
4450 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4451 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4452 if (data_hdr->b_spa != 0)
4456 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4457 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4458 if (meta_hdr->b_spa != 0)
4462 if (data_hdr == NULL && meta_hdr == NULL) {
4463 type = ARC_BUFC_DATA;
4464 } else if (data_hdr == NULL) {
4465 ASSERT3P(meta_hdr, !=, NULL);
4466 type = ARC_BUFC_METADATA;
4467 } else if (meta_hdr == NULL) {
4468 ASSERT3P(data_hdr, !=, NULL);
4469 type = ARC_BUFC_DATA;
4471 ASSERT3P(data_hdr, !=, NULL);
4472 ASSERT3P(meta_hdr, !=, NULL);
4474 /* The headers can't be on the sublist without an L1 header */
4475 ASSERT(HDR_HAS_L1HDR(data_hdr));
4476 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4478 if (data_hdr->b_l1hdr.b_arc_access <
4479 meta_hdr->b_l1hdr.b_arc_access) {
4480 type = ARC_BUFC_DATA;
4482 type = ARC_BUFC_METADATA;
4486 multilist_sublist_unlock(meta_mls);
4487 multilist_sublist_unlock(data_mls);
4493 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4498 uint64_t total_evicted = 0;
4501 uint64_t asize = aggsum_value(&arc_size);
4502 uint64_t ameta = aggsum_value(&arc_meta_used);
4505 * If we're over arc_meta_limit, we want to correct that before
4506 * potentially evicting data buffers below.
4508 total_evicted += arc_evict_meta(ameta);
4513 * If we're over the target cache size, we want to evict enough
4514 * from the list to get back to our target size. We don't want
4515 * to evict too much from the MRU, such that it drops below
4516 * arc_p. So, if we're over our target cache size more than
4517 * the MRU is over arc_p, we'll evict enough to get back to
4518 * arc_p here, and then evict more from the MFU below.
4520 target = MIN((int64_t)(asize - arc_c),
4521 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4522 zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4525 * If we're below arc_meta_min, always prefer to evict data.
4526 * Otherwise, try to satisfy the requested number of bytes to
4527 * evict from the type which contains older buffers; in an
4528 * effort to keep newer buffers in the cache regardless of their
4529 * type. If we cannot satisfy the number of bytes from this
4530 * type, spill over into the next type.
4532 if (arc_evict_type(arc_mru) == ARC_BUFC_METADATA &&
4533 ameta > arc_meta_min) {
4534 bytes = arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4535 total_evicted += bytes;
4538 * If we couldn't evict our target number of bytes from
4539 * metadata, we try to get the rest from data.
4544 arc_evict_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4546 bytes = arc_evict_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4547 total_evicted += bytes;
4550 * If we couldn't evict our target number of bytes from
4551 * data, we try to get the rest from metadata.
4556 arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4560 * Re-sum ARC stats after the first round of evictions.
4562 asize = aggsum_value(&arc_size);
4563 ameta = aggsum_value(&arc_meta_used);
4569 * Now that we've tried to evict enough from the MRU to get its
4570 * size back to arc_p, if we're still above the target cache
4571 * size, we evict the rest from the MFU.
4573 target = asize - arc_c;
4575 if (arc_evict_type(arc_mfu) == ARC_BUFC_METADATA &&
4576 ameta > arc_meta_min) {
4577 bytes = arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4578 total_evicted += bytes;
4581 * If we couldn't evict our target number of bytes from
4582 * metadata, we try to get the rest from data.
4587 arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4589 bytes = arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4590 total_evicted += bytes;
4593 * If we couldn't evict our target number of bytes from
4594 * data, we try to get the rest from data.
4599 arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4603 * Adjust ghost lists
4605 * In addition to the above, the ARC also defines target values
4606 * for the ghost lists. The sum of the mru list and mru ghost
4607 * list should never exceed the target size of the cache, and
4608 * the sum of the mru list, mfu list, mru ghost list, and mfu
4609 * ghost list should never exceed twice the target size of the
4610 * cache. The following logic enforces these limits on the ghost
4611 * caches, and evicts from them as needed.
4613 target = zfs_refcount_count(&arc_mru->arcs_size) +
4614 zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4616 bytes = arc_evict_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4617 total_evicted += bytes;
4622 arc_evict_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4625 * We assume the sum of the mru list and mfu list is less than
4626 * or equal to arc_c (we enforced this above), which means we
4627 * can use the simpler of the two equations below:
4629 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4630 * mru ghost + mfu ghost <= arc_c
4632 target = zfs_refcount_count(&arc_mru_ghost->arcs_size) +
4633 zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4635 bytes = arc_evict_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4636 total_evicted += bytes;
4641 arc_evict_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4643 return (total_evicted);
4647 arc_flush(spa_t *spa, boolean_t retry)
4652 * If retry is B_TRUE, a spa must not be specified since we have
4653 * no good way to determine if all of a spa's buffers have been
4654 * evicted from an arc state.
4656 ASSERT(!retry || spa == 0);
4659 guid = spa_load_guid(spa);
4661 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4662 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4664 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4665 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4667 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4668 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4670 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4671 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4675 arc_reduce_target_size(int64_t to_free)
4677 uint64_t asize = aggsum_value(&arc_size);
4680 * All callers want the ARC to actually evict (at least) this much
4681 * memory. Therefore we reduce from the lower of the current size and
4682 * the target size. This way, even if arc_c is much higher than
4683 * arc_size (as can be the case after many calls to arc_freed(), we will
4684 * immediately have arc_c < arc_size and therefore the arc_evict_zthr
4687 uint64_t c = MIN(arc_c, asize);
4689 if (c > to_free && c - to_free > arc_c_min) {
4690 arc_c = c - to_free;
4691 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4693 arc_p = (arc_c >> 1);
4694 ASSERT(arc_c >= arc_c_min);
4695 ASSERT((int64_t)arc_p >= 0);
4700 if (asize > arc_c) {
4701 /* See comment in arc_evict_cb_check() on why lock+flag */
4702 mutex_enter(&arc_evict_lock);
4703 arc_evict_needed = B_TRUE;
4704 mutex_exit(&arc_evict_lock);
4705 zthr_wakeup(arc_evict_zthr);
4710 * Determine if the system is under memory pressure and is asking
4711 * to reclaim memory. A return value of B_TRUE indicates that the system
4712 * is under memory pressure and that the arc should adjust accordingly.
4715 arc_reclaim_needed(void)
4717 return (arc_available_memory() < 0);
4721 arc_kmem_reap_soon(void)
4724 kmem_cache_t *prev_cache = NULL;
4725 kmem_cache_t *prev_data_cache = NULL;
4726 extern kmem_cache_t *zio_buf_cache[];
4727 extern kmem_cache_t *zio_data_buf_cache[];
4730 if ((aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) &&
4731 zfs_arc_meta_prune) {
4733 * We are exceeding our meta-data cache limit.
4734 * Prune some entries to release holds on meta-data.
4736 arc_prune_async(zfs_arc_meta_prune);
4740 * Reclaim unused memory from all kmem caches.
4746 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4748 /* reach upper limit of cache size on 32-bit */
4749 if (zio_buf_cache[i] == NULL)
4752 if (zio_buf_cache[i] != prev_cache) {
4753 prev_cache = zio_buf_cache[i];
4754 kmem_cache_reap_now(zio_buf_cache[i]);
4756 if (zio_data_buf_cache[i] != prev_data_cache) {
4757 prev_data_cache = zio_data_buf_cache[i];
4758 kmem_cache_reap_now(zio_data_buf_cache[i]);
4761 kmem_cache_reap_now(buf_cache);
4762 kmem_cache_reap_now(hdr_full_cache);
4763 kmem_cache_reap_now(hdr_l2only_cache);
4764 kmem_cache_reap_now(zfs_btree_leaf_cache);
4765 abd_cache_reap_now();
4770 arc_evict_cb_check(void *arg, zthr_t *zthr)
4773 * This is necessary so that any changes which may have been made to
4774 * many of the zfs_arc_* module parameters will be propagated to
4775 * their actual internal variable counterparts. Without this,
4776 * changing those module params at runtime would have no effect.
4778 arc_tuning_update(B_FALSE);
4781 * This is necessary in order to keep the kstat information
4782 * up to date for tools that display kstat data such as the
4783 * mdb ::arc dcmd and the Linux crash utility. These tools
4784 * typically do not call kstat's update function, but simply
4785 * dump out stats from the most recent update. Without
4786 * this call, these commands may show stale stats for the
4787 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4788 * with this change, the data might be up to 1 second
4789 * out of date(the arc_evict_zthr has a maximum sleep
4790 * time of 1 second); but that should suffice. The
4791 * arc_state_t structures can be queried directly if more
4792 * accurate information is needed.
4794 if (arc_ksp != NULL)
4795 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4798 * We have to rely on arc_wait_for_eviction() to tell us when to
4799 * evict, rather than checking if we are overflowing here, so that we
4800 * are sure to not leave arc_wait_for_eviction() waiting on aew_cv.
4801 * If we have become "not overflowing" since arc_wait_for_eviction()
4802 * checked, we need to wake it up. We could broadcast the CV here,
4803 * but arc_wait_for_eviction() may have not yet gone to sleep. We
4804 * would need to use a mutex to ensure that this function doesn't
4805 * broadcast until arc_wait_for_eviction() has gone to sleep (e.g.
4806 * the arc_evict_lock). However, the lock ordering of such a lock
4807 * would necessarily be incorrect with respect to the zthr_lock,
4808 * which is held before this function is called, and is held by
4809 * arc_wait_for_eviction() when it calls zthr_wakeup().
4811 return (arc_evict_needed);
4815 * Keep arc_size under arc_c by running arc_evict which evicts data
4820 arc_evict_cb(void *arg, zthr_t *zthr)
4822 uint64_t evicted = 0;
4823 fstrans_cookie_t cookie = spl_fstrans_mark();
4825 /* Evict from cache */
4826 evicted = arc_evict();
4829 * If evicted is zero, we couldn't evict anything
4830 * via arc_evict(). This could be due to hash lock
4831 * collisions, but more likely due to the majority of
4832 * arc buffers being unevictable. Therefore, even if
4833 * arc_size is above arc_c, another pass is unlikely to
4834 * be helpful and could potentially cause us to enter an
4835 * infinite loop. Additionally, zthr_iscancelled() is
4836 * checked here so that if the arc is shutting down, the
4837 * broadcast will wake any remaining arc evict waiters.
4839 mutex_enter(&arc_evict_lock);
4840 arc_evict_needed = !zthr_iscancelled(arc_evict_zthr) &&
4841 evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0;
4842 if (!arc_evict_needed) {
4844 * We're either no longer overflowing, or we
4845 * can't evict anything more, so we should wake
4846 * arc_get_data_impl() sooner.
4848 arc_evict_waiter_t *aw;
4849 while ((aw = list_remove_head(&arc_evict_waiters)) != NULL) {
4850 cv_broadcast(&aw->aew_cv);
4852 arc_set_need_free();
4854 mutex_exit(&arc_evict_lock);
4855 spl_fstrans_unmark(cookie);
4860 arc_reap_cb_check(void *arg, zthr_t *zthr)
4862 int64_t free_memory = arc_available_memory();
4863 static int reap_cb_check_counter = 0;
4866 * If a kmem reap is already active, don't schedule more. We must
4867 * check for this because kmem_cache_reap_soon() won't actually
4868 * block on the cache being reaped (this is to prevent callers from
4869 * becoming implicitly blocked by a system-wide kmem reap -- which,
4870 * on a system with many, many full magazines, can take minutes).
4872 if (!kmem_cache_reap_active() && free_memory < 0) {
4874 arc_no_grow = B_TRUE;
4877 * Wait at least zfs_grow_retry (default 5) seconds
4878 * before considering growing.
4880 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4882 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4883 arc_no_grow = B_TRUE;
4884 } else if (gethrtime() >= arc_growtime) {
4885 arc_no_grow = B_FALSE;
4889 * Called unconditionally every 60 seconds to reclaim unused
4890 * zstd compression and decompression context. This is done
4891 * here to avoid the need for an independent thread.
4893 if (!((reap_cb_check_counter++) % 60))
4894 zfs_zstd_cache_reap_now();
4900 * Keep enough free memory in the system by reaping the ARC's kmem
4901 * caches. To cause more slabs to be reapable, we may reduce the
4902 * target size of the cache (arc_c), causing the arc_evict_cb()
4903 * to free more buffers.
4907 arc_reap_cb(void *arg, zthr_t *zthr)
4909 int64_t free_memory;
4910 fstrans_cookie_t cookie = spl_fstrans_mark();
4913 * Kick off asynchronous kmem_reap()'s of all our caches.
4915 arc_kmem_reap_soon();
4918 * Wait at least arc_kmem_cache_reap_retry_ms between
4919 * arc_kmem_reap_soon() calls. Without this check it is possible to
4920 * end up in a situation where we spend lots of time reaping
4921 * caches, while we're near arc_c_min. Waiting here also gives the
4922 * subsequent free memory check a chance of finding that the
4923 * asynchronous reap has already freed enough memory, and we don't
4924 * need to call arc_reduce_target_size().
4926 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
4929 * Reduce the target size as needed to maintain the amount of free
4930 * memory in the system at a fraction of the arc_size (1/128th by
4931 * default). If oversubscribed (free_memory < 0) then reduce the
4932 * target arc_size by the deficit amount plus the fractional
4933 * amount. If free memory is positive but less then the fractional
4934 * amount, reduce by what is needed to hit the fractional amount.
4936 free_memory = arc_available_memory();
4939 (arc_c >> arc_shrink_shift) - free_memory;
4941 arc_reduce_target_size(to_free);
4943 spl_fstrans_unmark(cookie);
4948 * Determine the amount of memory eligible for eviction contained in the
4949 * ARC. All clean data reported by the ghost lists can always be safely
4950 * evicted. Due to arc_c_min, the same does not hold for all clean data
4951 * contained by the regular mru and mfu lists.
4953 * In the case of the regular mru and mfu lists, we need to report as
4954 * much clean data as possible, such that evicting that same reported
4955 * data will not bring arc_size below arc_c_min. Thus, in certain
4956 * circumstances, the total amount of clean data in the mru and mfu
4957 * lists might not actually be evictable.
4959 * The following two distinct cases are accounted for:
4961 * 1. The sum of the amount of dirty data contained by both the mru and
4962 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4963 * is greater than or equal to arc_c_min.
4964 * (i.e. amount of dirty data >= arc_c_min)
4966 * This is the easy case; all clean data contained by the mru and mfu
4967 * lists is evictable. Evicting all clean data can only drop arc_size
4968 * to the amount of dirty data, which is greater than arc_c_min.
4970 * 2. The sum of the amount of dirty data contained by both the mru and
4971 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4972 * is less than arc_c_min.
4973 * (i.e. arc_c_min > amount of dirty data)
4975 * 2.1. arc_size is greater than or equal arc_c_min.
4976 * (i.e. arc_size >= arc_c_min > amount of dirty data)
4978 * In this case, not all clean data from the regular mru and mfu
4979 * lists is actually evictable; we must leave enough clean data
4980 * to keep arc_size above arc_c_min. Thus, the maximum amount of
4981 * evictable data from the two lists combined, is exactly the
4982 * difference between arc_size and arc_c_min.
4984 * 2.2. arc_size is less than arc_c_min
4985 * (i.e. arc_c_min > arc_size > amount of dirty data)
4987 * In this case, none of the data contained in the mru and mfu
4988 * lists is evictable, even if it's clean. Since arc_size is
4989 * already below arc_c_min, evicting any more would only
4990 * increase this negative difference.
4993 #endif /* _KERNEL */
4996 * Adapt arc info given the number of bytes we are trying to add and
4997 * the state that we are coming from. This function is only called
4998 * when we are adding new content to the cache.
5001 arc_adapt(int bytes, arc_state_t *state)
5004 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
5005 int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size);
5006 int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size);
5010 * Adapt the target size of the MRU list:
5011 * - if we just hit in the MRU ghost list, then increase
5012 * the target size of the MRU list.
5013 * - if we just hit in the MFU ghost list, then increase
5014 * the target size of the MFU list by decreasing the
5015 * target size of the MRU list.
5017 if (state == arc_mru_ghost) {
5018 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
5019 if (!zfs_arc_p_dampener_disable)
5020 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
5022 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
5023 } else if (state == arc_mfu_ghost) {
5026 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
5027 if (!zfs_arc_p_dampener_disable)
5028 mult = MIN(mult, 10);
5030 delta = MIN(bytes * mult, arc_p);
5031 arc_p = MAX(arc_p_min, arc_p - delta);
5033 ASSERT((int64_t)arc_p >= 0);
5036 * Wake reap thread if we do not have any available memory
5038 if (arc_reclaim_needed()) {
5039 zthr_wakeup(arc_reap_zthr);
5046 if (arc_c >= arc_c_max)
5050 * If we're within (2 * maxblocksize) bytes of the target
5051 * cache size, increment the target cache size
5053 ASSERT3U(arc_c, >=, 2ULL << SPA_MAXBLOCKSHIFT);
5054 if (aggsum_upper_bound(&arc_size) >=
5055 arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
5056 atomic_add_64(&arc_c, (int64_t)bytes);
5057 if (arc_c > arc_c_max)
5059 else if (state == arc_anon)
5060 atomic_add_64(&arc_p, (int64_t)bytes);
5064 ASSERT((int64_t)arc_p >= 0);
5068 * Check if arc_size has grown past our upper threshold, determined by
5069 * zfs_arc_overflow_shift.
5072 arc_is_overflowing(void)
5074 /* Always allow at least one block of overflow */
5075 int64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5076 arc_c >> zfs_arc_overflow_shift);
5079 * We just compare the lower bound here for performance reasons. Our
5080 * primary goals are to make sure that the arc never grows without
5081 * bound, and that it can reach its maximum size. This check
5082 * accomplishes both goals. The maximum amount we could run over by is
5083 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
5084 * in the ARC. In practice, that's in the tens of MB, which is low
5085 * enough to be safe.
5087 return (aggsum_lower_bound(&arc_size) >= (int64_t)arc_c + overflow);
5091 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag,
5094 arc_buf_contents_t type = arc_buf_type(hdr);
5096 arc_get_data_impl(hdr, size, tag, do_adapt);
5097 if (type == ARC_BUFC_METADATA) {
5098 return (abd_alloc(size, B_TRUE));
5100 ASSERT(type == ARC_BUFC_DATA);
5101 return (abd_alloc(size, B_FALSE));
5106 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5108 arc_buf_contents_t type = arc_buf_type(hdr);
5110 arc_get_data_impl(hdr, size, tag, B_TRUE);
5111 if (type == ARC_BUFC_METADATA) {
5112 return (zio_buf_alloc(size));
5114 ASSERT(type == ARC_BUFC_DATA);
5115 return (zio_data_buf_alloc(size));
5120 * Wait for the specified amount of data (in bytes) to be evicted from the
5121 * ARC, and for there to be sufficient free memory in the system. Waiting for
5122 * eviction ensures that the memory used by the ARC decreases. Waiting for
5123 * free memory ensures that the system won't run out of free pages, regardless
5124 * of ARC behavior and settings. See arc_lowmem_init().
5127 arc_wait_for_eviction(uint64_t amount)
5129 mutex_enter(&arc_evict_lock);
5130 if (arc_is_overflowing()) {
5131 arc_evict_needed = B_TRUE;
5132 zthr_wakeup(arc_evict_zthr);
5135 arc_evict_waiter_t aw;
5136 list_link_init(&aw.aew_node);
5137 cv_init(&aw.aew_cv, NULL, CV_DEFAULT, NULL);
5139 arc_evict_waiter_t *last =
5140 list_tail(&arc_evict_waiters);
5142 ASSERT3U(last->aew_count, >, arc_evict_count);
5143 aw.aew_count = last->aew_count + amount;
5145 aw.aew_count = arc_evict_count + amount;
5148 list_insert_tail(&arc_evict_waiters, &aw);
5150 arc_set_need_free();
5152 DTRACE_PROBE3(arc__wait__for__eviction,
5154 uint64_t, arc_evict_count,
5155 uint64_t, aw.aew_count);
5158 * We will be woken up either when arc_evict_count
5159 * reaches aew_count, or when the ARC is no longer
5160 * overflowing and eviction completes.
5162 cv_wait(&aw.aew_cv, &arc_evict_lock);
5165 * In case of "false" wakeup, we will still be on the
5168 if (list_link_active(&aw.aew_node))
5169 list_remove(&arc_evict_waiters, &aw);
5171 cv_destroy(&aw.aew_cv);
5174 mutex_exit(&arc_evict_lock);
5178 * Allocate a block and return it to the caller. If we are hitting the
5179 * hard limit for the cache size, we must sleep, waiting for the eviction
5180 * thread to catch up. If we're past the target size but below the hard
5181 * limit, we'll only signal the reclaim thread and continue on.
5184 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag,
5187 arc_state_t *state = hdr->b_l1hdr.b_state;
5188 arc_buf_contents_t type = arc_buf_type(hdr);
5191 arc_adapt(size, state);
5194 * If arc_size is currently overflowing, we must be adding data
5195 * faster than we are evicting. To ensure we don't compound the
5196 * problem by adding more data and forcing arc_size to grow even
5197 * further past it's target size, we wait for the eviction thread to
5198 * make some progress. We also wait for there to be sufficient free
5199 * memory in the system, as measured by arc_free_memory().
5201 * Specifically, we wait for zfs_arc_eviction_pct percent of the
5202 * requested size to be evicted. This should be more than 100%, to
5203 * ensure that that progress is also made towards getting arc_size
5204 * under arc_c. See the comment above zfs_arc_eviction_pct.
5206 * We do the overflowing check without holding the arc_evict_lock to
5207 * reduce lock contention in this hot path. Note that
5208 * arc_wait_for_eviction() will acquire the lock and check again to
5209 * ensure we are truly overflowing before blocking.
5211 if (arc_is_overflowing()) {
5212 arc_wait_for_eviction(size *
5213 zfs_arc_eviction_pct / 100);
5216 VERIFY3U(hdr->b_type, ==, type);
5217 if (type == ARC_BUFC_METADATA) {
5218 arc_space_consume(size, ARC_SPACE_META);
5220 arc_space_consume(size, ARC_SPACE_DATA);
5224 * Update the state size. Note that ghost states have a
5225 * "ghost size" and so don't need to be updated.
5227 if (!GHOST_STATE(state)) {
5229 (void) zfs_refcount_add_many(&state->arcs_size, size, tag);
5232 * If this is reached via arc_read, the link is
5233 * protected by the hash lock. If reached via
5234 * arc_buf_alloc, the header should not be accessed by
5235 * any other thread. And, if reached via arc_read_done,
5236 * the hash lock will protect it if it's found in the
5237 * hash table; otherwise no other thread should be
5238 * trying to [add|remove]_reference it.
5240 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5241 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5242 (void) zfs_refcount_add_many(&state->arcs_esize[type],
5247 * If we are growing the cache, and we are adding anonymous
5248 * data, and we have outgrown arc_p, update arc_p
5250 if (aggsum_upper_bound(&arc_size) < arc_c &&
5251 hdr->b_l1hdr.b_state == arc_anon &&
5252 (zfs_refcount_count(&arc_anon->arcs_size) +
5253 zfs_refcount_count(&arc_mru->arcs_size) > arc_p))
5254 arc_p = MIN(arc_c, arc_p + size);
5259 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5261 arc_free_data_impl(hdr, size, tag);
5266 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5268 arc_buf_contents_t type = arc_buf_type(hdr);
5270 arc_free_data_impl(hdr, size, tag);
5271 if (type == ARC_BUFC_METADATA) {
5272 zio_buf_free(buf, size);
5274 ASSERT(type == ARC_BUFC_DATA);
5275 zio_data_buf_free(buf, size);
5280 * Free the arc data buffer.
5283 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5285 arc_state_t *state = hdr->b_l1hdr.b_state;
5286 arc_buf_contents_t type = arc_buf_type(hdr);
5288 /* protected by hash lock, if in the hash table */
5289 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5290 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5291 ASSERT(state != arc_anon && state != arc_l2c_only);
5293 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
5296 (void) zfs_refcount_remove_many(&state->arcs_size, size, tag);
5298 VERIFY3U(hdr->b_type, ==, type);
5299 if (type == ARC_BUFC_METADATA) {
5300 arc_space_return(size, ARC_SPACE_META);
5302 ASSERT(type == ARC_BUFC_DATA);
5303 arc_space_return(size, ARC_SPACE_DATA);
5308 * This routine is called whenever a buffer is accessed.
5309 * NOTE: the hash lock is dropped in this function.
5312 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5316 ASSERT(MUTEX_HELD(hash_lock));
5317 ASSERT(HDR_HAS_L1HDR(hdr));
5319 if (hdr->b_l1hdr.b_state == arc_anon) {
5321 * This buffer is not in the cache, and does not
5322 * appear in our "ghost" list. Add the new buffer
5326 ASSERT0(hdr->b_l1hdr.b_arc_access);
5327 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5328 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5329 arc_change_state(arc_mru, hdr, hash_lock);
5331 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5332 now = ddi_get_lbolt();
5335 * If this buffer is here because of a prefetch, then either:
5336 * - clear the flag if this is a "referencing" read
5337 * (any subsequent access will bump this into the MFU state).
5339 * - move the buffer to the head of the list if this is
5340 * another prefetch (to make it less likely to be evicted).
5342 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5343 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5344 /* link protected by hash lock */
5345 ASSERT(multilist_link_active(
5346 &hdr->b_l1hdr.b_arc_node));
5348 arc_hdr_clear_flags(hdr,
5350 ARC_FLAG_PRESCIENT_PREFETCH);
5351 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5352 ARCSTAT_BUMP(arcstat_mru_hits);
5354 hdr->b_l1hdr.b_arc_access = now;
5359 * This buffer has been "accessed" only once so far,
5360 * but it is still in the cache. Move it to the MFU
5363 if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access +
5366 * More than 125ms have passed since we
5367 * instantiated this buffer. Move it to the
5368 * most frequently used state.
5370 hdr->b_l1hdr.b_arc_access = now;
5371 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5372 arc_change_state(arc_mfu, hdr, hash_lock);
5374 atomic_inc_32(&hdr->b_l1hdr.b_mru_hits);
5375 ARCSTAT_BUMP(arcstat_mru_hits);
5376 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5377 arc_state_t *new_state;
5379 * This buffer has been "accessed" recently, but
5380 * was evicted from the cache. Move it to the
5384 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5385 new_state = arc_mru;
5386 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5387 arc_hdr_clear_flags(hdr,
5389 ARC_FLAG_PRESCIENT_PREFETCH);
5391 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5393 new_state = arc_mfu;
5394 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5397 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5398 arc_change_state(new_state, hdr, hash_lock);
5400 atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits);
5401 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5402 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5404 * This buffer has been accessed more than once and is
5405 * still in the cache. Keep it in the MFU state.
5407 * NOTE: an add_reference() that occurred when we did
5408 * the arc_read() will have kicked this off the list.
5409 * If it was a prefetch, we will explicitly move it to
5410 * the head of the list now.
5413 atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits);
5414 ARCSTAT_BUMP(arcstat_mfu_hits);
5415 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5416 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5417 arc_state_t *new_state = arc_mfu;
5419 * This buffer has been accessed more than once but has
5420 * been evicted from the cache. Move it back to the
5424 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5426 * This is a prefetch access...
5427 * move this block back to the MRU state.
5429 new_state = arc_mru;
5432 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5433 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5434 arc_change_state(new_state, hdr, hash_lock);
5436 atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits);
5437 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5438 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5440 * This buffer is on the 2nd Level ARC.
5443 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5444 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5445 arc_change_state(arc_mfu, hdr, hash_lock);
5447 cmn_err(CE_PANIC, "invalid arc state 0x%p",
5448 hdr->b_l1hdr.b_state);
5453 * This routine is called by dbuf_hold() to update the arc_access() state
5454 * which otherwise would be skipped for entries in the dbuf cache.
5457 arc_buf_access(arc_buf_t *buf)
5459 mutex_enter(&buf->b_evict_lock);
5460 arc_buf_hdr_t *hdr = buf->b_hdr;
5463 * Avoid taking the hash_lock when possible as an optimization.
5464 * The header must be checked again under the hash_lock in order
5465 * to handle the case where it is concurrently being released.
5467 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5468 mutex_exit(&buf->b_evict_lock);
5472 kmutex_t *hash_lock = HDR_LOCK(hdr);
5473 mutex_enter(hash_lock);
5475 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5476 mutex_exit(hash_lock);
5477 mutex_exit(&buf->b_evict_lock);
5478 ARCSTAT_BUMP(arcstat_access_skip);
5482 mutex_exit(&buf->b_evict_lock);
5484 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5485 hdr->b_l1hdr.b_state == arc_mfu);
5487 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5488 arc_access(hdr, hash_lock);
5489 mutex_exit(hash_lock);
5491 ARCSTAT_BUMP(arcstat_hits);
5492 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr) && !HDR_PRESCIENT_PREFETCH(hdr),
5493 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5496 /* a generic arc_read_done_func_t which you can use */
5499 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5500 arc_buf_t *buf, void *arg)
5505 bcopy(buf->b_data, arg, arc_buf_size(buf));
5506 arc_buf_destroy(buf, arg);
5509 /* a generic arc_read_done_func_t */
5512 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5513 arc_buf_t *buf, void *arg)
5515 arc_buf_t **bufp = arg;
5518 ASSERT(zio == NULL || zio->io_error != 0);
5521 ASSERT(zio == NULL || zio->io_error == 0);
5523 ASSERT(buf->b_data != NULL);
5528 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5530 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5531 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5532 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5534 if (HDR_COMPRESSION_ENABLED(hdr)) {
5535 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5536 BP_GET_COMPRESS(bp));
5538 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5539 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5540 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5545 arc_read_done(zio_t *zio)
5547 blkptr_t *bp = zio->io_bp;
5548 arc_buf_hdr_t *hdr = zio->io_private;
5549 kmutex_t *hash_lock = NULL;
5550 arc_callback_t *callback_list;
5551 arc_callback_t *acb;
5552 boolean_t freeable = B_FALSE;
5555 * The hdr was inserted into hash-table and removed from lists
5556 * prior to starting I/O. We should find this header, since
5557 * it's in the hash table, and it should be legit since it's
5558 * not possible to evict it during the I/O. The only possible
5559 * reason for it not to be found is if we were freed during the
5562 if (HDR_IN_HASH_TABLE(hdr)) {
5563 arc_buf_hdr_t *found;
5565 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5566 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5567 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5568 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5569 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5571 found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock);
5573 ASSERT((found == hdr &&
5574 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5575 (found == hdr && HDR_L2_READING(hdr)));
5576 ASSERT3P(hash_lock, !=, NULL);
5579 if (BP_IS_PROTECTED(bp)) {
5580 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5581 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5582 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5583 hdr->b_crypt_hdr.b_iv);
5585 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5588 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5589 sizeof (zil_chain_t));
5590 zio_crypt_decode_mac_zil(tmpbuf,
5591 hdr->b_crypt_hdr.b_mac);
5592 abd_return_buf(zio->io_abd, tmpbuf,
5593 sizeof (zil_chain_t));
5595 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
5599 if (zio->io_error == 0) {
5600 /* byteswap if necessary */
5601 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5602 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5603 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5605 hdr->b_l1hdr.b_byteswap =
5606 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5609 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5611 if (!HDR_L2_READING(hdr)) {
5612 hdr->b_complevel = zio->io_prop.zp_complevel;
5616 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5617 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5618 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5620 callback_list = hdr->b_l1hdr.b_acb;
5621 ASSERT3P(callback_list, !=, NULL);
5623 if (hash_lock && zio->io_error == 0 &&
5624 hdr->b_l1hdr.b_state == arc_anon) {
5626 * Only call arc_access on anonymous buffers. This is because
5627 * if we've issued an I/O for an evicted buffer, we've already
5628 * called arc_access (to prevent any simultaneous readers from
5629 * getting confused).
5631 arc_access(hdr, hash_lock);
5635 * If a read request has a callback (i.e. acb_done is not NULL), then we
5636 * make a buf containing the data according to the parameters which were
5637 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5638 * aren't needlessly decompressing the data multiple times.
5640 int callback_cnt = 0;
5641 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5647 if (zio->io_error != 0)
5650 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5651 &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
5652 acb->acb_compressed, acb->acb_noauth, B_TRUE,
5656 * Assert non-speculative zios didn't fail because an
5657 * encryption key wasn't loaded
5659 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
5663 * If we failed to decrypt, report an error now (as the zio
5664 * layer would have done if it had done the transforms).
5666 if (error == ECKSUM) {
5667 ASSERT(BP_IS_PROTECTED(bp));
5668 error = SET_ERROR(EIO);
5669 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5670 spa_log_error(zio->io_spa, &acb->acb_zb);
5671 (void) zfs_ereport_post(
5672 FM_EREPORT_ZFS_AUTHENTICATION,
5673 zio->io_spa, NULL, &acb->acb_zb, zio, 0);
5679 * Decompression or decryption failed. Set
5680 * io_error so that when we call acb_done
5681 * (below), we will indicate that the read
5682 * failed. Note that in the unusual case
5683 * where one callback is compressed and another
5684 * uncompressed, we will mark all of them
5685 * as failed, even though the uncompressed
5686 * one can't actually fail. In this case,
5687 * the hdr will not be anonymous, because
5688 * if there are multiple callbacks, it's
5689 * because multiple threads found the same
5690 * arc buf in the hash table.
5692 zio->io_error = error;
5697 * If there are multiple callbacks, we must have the hash lock,
5698 * because the only way for multiple threads to find this hdr is
5699 * in the hash table. This ensures that if there are multiple
5700 * callbacks, the hdr is not anonymous. If it were anonymous,
5701 * we couldn't use arc_buf_destroy() in the error case below.
5703 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5705 hdr->b_l1hdr.b_acb = NULL;
5706 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5707 if (callback_cnt == 0)
5708 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
5710 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5711 callback_list != NULL);
5713 if (zio->io_error == 0) {
5714 arc_hdr_verify(hdr, zio->io_bp);
5716 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5717 if (hdr->b_l1hdr.b_state != arc_anon)
5718 arc_change_state(arc_anon, hdr, hash_lock);
5719 if (HDR_IN_HASH_TABLE(hdr))
5720 buf_hash_remove(hdr);
5721 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5725 * Broadcast before we drop the hash_lock to avoid the possibility
5726 * that the hdr (and hence the cv) might be freed before we get to
5727 * the cv_broadcast().
5729 cv_broadcast(&hdr->b_l1hdr.b_cv);
5731 if (hash_lock != NULL) {
5732 mutex_exit(hash_lock);
5735 * This block was freed while we waited for the read to
5736 * complete. It has been removed from the hash table and
5737 * moved to the anonymous state (so that it won't show up
5740 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5741 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5744 /* execute each callback and free its structure */
5745 while ((acb = callback_list) != NULL) {
5746 if (acb->acb_done != NULL) {
5747 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5749 * If arc_buf_alloc_impl() fails during
5750 * decompression, the buf will still be
5751 * allocated, and needs to be freed here.
5753 arc_buf_destroy(acb->acb_buf,
5755 acb->acb_buf = NULL;
5757 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5758 acb->acb_buf, acb->acb_private);
5761 if (acb->acb_zio_dummy != NULL) {
5762 acb->acb_zio_dummy->io_error = zio->io_error;
5763 zio_nowait(acb->acb_zio_dummy);
5766 callback_list = acb->acb_next;
5767 kmem_free(acb, sizeof (arc_callback_t));
5771 arc_hdr_destroy(hdr);
5775 * "Read" the block at the specified DVA (in bp) via the
5776 * cache. If the block is found in the cache, invoke the provided
5777 * callback immediately and return. Note that the `zio' parameter
5778 * in the callback will be NULL in this case, since no IO was
5779 * required. If the block is not in the cache pass the read request
5780 * on to the spa with a substitute callback function, so that the
5781 * requested block will be added to the cache.
5783 * If a read request arrives for a block that has a read in-progress,
5784 * either wait for the in-progress read to complete (and return the
5785 * results); or, if this is a read with a "done" func, add a record
5786 * to the read to invoke the "done" func when the read completes,
5787 * and return; or just return.
5789 * arc_read_done() will invoke all the requested "done" functions
5790 * for readers of this block.
5793 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
5794 arc_read_done_func_t *done, void *private, zio_priority_t priority,
5795 int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5797 arc_buf_hdr_t *hdr = NULL;
5798 kmutex_t *hash_lock = NULL;
5800 uint64_t guid = spa_load_guid(spa);
5801 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
5802 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
5803 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5804 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
5805 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5806 boolean_t embedded_bp = !!BP_IS_EMBEDDED(bp);
5809 ASSERT(!embedded_bp ||
5810 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5811 ASSERT(!BP_IS_HOLE(bp));
5812 ASSERT(!BP_IS_REDACTED(bp));
5815 * Normally SPL_FSTRANS will already be set since kernel threads which
5816 * expect to call the DMU interfaces will set it when created. System
5817 * calls are similarly handled by setting/cleaning the bit in the
5818 * registered callback (module/os/.../zfs/zpl_*).
5820 * External consumers such as Lustre which call the exported DMU
5821 * interfaces may not have set SPL_FSTRANS. To avoid a deadlock
5822 * on the hash_lock always set and clear the bit.
5824 fstrans_cookie_t cookie = spl_fstrans_mark();
5828 * Embedded BP's have no DVA and require no I/O to "read".
5829 * Create an anonymous arc buf to back it.
5831 hdr = buf_hash_find(guid, bp, &hash_lock);
5835 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5836 * we maintain encrypted data separately from compressed / uncompressed
5837 * data. If the user is requesting raw encrypted data and we don't have
5838 * that in the header we will read from disk to guarantee that we can
5839 * get it even if the encryption keys aren't loaded.
5841 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
5842 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
5843 arc_buf_t *buf = NULL;
5844 *arc_flags |= ARC_FLAG_CACHED;
5846 if (HDR_IO_IN_PROGRESS(hdr)) {
5847 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5849 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
5850 mutex_exit(hash_lock);
5851 ARCSTAT_BUMP(arcstat_cached_only_in_progress);
5852 rc = SET_ERROR(ENOENT);
5856 ASSERT3P(head_zio, !=, NULL);
5857 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5858 priority == ZIO_PRIORITY_SYNC_READ) {
5860 * This is a sync read that needs to wait for
5861 * an in-flight async read. Request that the
5862 * zio have its priority upgraded.
5864 zio_change_priority(head_zio, priority);
5865 DTRACE_PROBE1(arc__async__upgrade__sync,
5866 arc_buf_hdr_t *, hdr);
5867 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5869 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5870 arc_hdr_clear_flags(hdr,
5871 ARC_FLAG_PREDICTIVE_PREFETCH);
5874 if (*arc_flags & ARC_FLAG_WAIT) {
5875 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5876 mutex_exit(hash_lock);
5879 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5882 arc_callback_t *acb = NULL;
5884 acb = kmem_zalloc(sizeof (arc_callback_t),
5886 acb->acb_done = done;
5887 acb->acb_private = private;
5888 acb->acb_compressed = compressed_read;
5889 acb->acb_encrypted = encrypted_read;
5890 acb->acb_noauth = noauth_read;
5893 acb->acb_zio_dummy = zio_null(pio,
5894 spa, NULL, NULL, NULL, zio_flags);
5896 ASSERT3P(acb->acb_done, !=, NULL);
5897 acb->acb_zio_head = head_zio;
5898 acb->acb_next = hdr->b_l1hdr.b_acb;
5899 hdr->b_l1hdr.b_acb = acb;
5900 mutex_exit(hash_lock);
5903 mutex_exit(hash_lock);
5907 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5908 hdr->b_l1hdr.b_state == arc_mfu);
5911 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5913 * This is a demand read which does not have to
5914 * wait for i/o because we did a predictive
5915 * prefetch i/o for it, which has completed.
5918 arc__demand__hit__predictive__prefetch,
5919 arc_buf_hdr_t *, hdr);
5921 arcstat_demand_hit_predictive_prefetch);
5922 arc_hdr_clear_flags(hdr,
5923 ARC_FLAG_PREDICTIVE_PREFETCH);
5926 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5928 arcstat_demand_hit_prescient_prefetch);
5929 arc_hdr_clear_flags(hdr,
5930 ARC_FLAG_PRESCIENT_PREFETCH);
5933 ASSERT(!embedded_bp || !BP_IS_HOLE(bp));
5935 /* Get a buf with the desired data in it. */
5936 rc = arc_buf_alloc_impl(hdr, spa, zb, private,
5937 encrypted_read, compressed_read, noauth_read,
5941 * Convert authentication and decryption errors
5942 * to EIO (and generate an ereport if needed)
5943 * before leaving the ARC.
5945 rc = SET_ERROR(EIO);
5946 if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5947 spa_log_error(spa, zb);
5948 (void) zfs_ereport_post(
5949 FM_EREPORT_ZFS_AUTHENTICATION,
5950 spa, NULL, zb, NULL, 0);
5954 (void) remove_reference(hdr, hash_lock,
5956 arc_buf_destroy_impl(buf);
5960 /* assert any errors weren't due to unloaded keys */
5961 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5963 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5964 zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5965 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5967 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5968 arc_access(hdr, hash_lock);
5969 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5970 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5971 if (*arc_flags & ARC_FLAG_L2CACHE)
5972 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5973 mutex_exit(hash_lock);
5974 ARCSTAT_BUMP(arcstat_hits);
5975 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5976 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5977 data, metadata, hits);
5980 done(NULL, zb, bp, buf, private);
5982 uint64_t lsize = BP_GET_LSIZE(bp);
5983 uint64_t psize = BP_GET_PSIZE(bp);
5984 arc_callback_t *acb;
5987 boolean_t devw = B_FALSE;
5990 int alloc_flags = encrypted_read ? ARC_HDR_ALLOC_RDATA : 0;
5992 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
5993 rc = SET_ERROR(ENOENT);
5994 if (hash_lock != NULL)
5995 mutex_exit(hash_lock);
6000 * Gracefully handle a damaged logical block size as a
6003 if (lsize > spa_maxblocksize(spa)) {
6004 rc = SET_ERROR(ECKSUM);
6005 if (hash_lock != NULL)
6006 mutex_exit(hash_lock);
6012 * This block is not in the cache or it has
6015 arc_buf_hdr_t *exists = NULL;
6016 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
6017 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
6018 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), 0, type,
6022 hdr->b_dva = *BP_IDENTITY(bp);
6023 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
6024 exists = buf_hash_insert(hdr, &hash_lock);
6026 if (exists != NULL) {
6027 /* somebody beat us to the hash insert */
6028 mutex_exit(hash_lock);
6029 buf_discard_identity(hdr);
6030 arc_hdr_destroy(hdr);
6031 goto top; /* restart the IO request */
6035 * This block is in the ghost cache or encrypted data
6036 * was requested and we didn't have it. If it was
6037 * L2-only (and thus didn't have an L1 hdr),
6038 * we realloc the header to add an L1 hdr.
6040 if (!HDR_HAS_L1HDR(hdr)) {
6041 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
6045 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
6046 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6047 ASSERT(!HDR_HAS_RABD(hdr));
6048 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6049 ASSERT0(zfs_refcount_count(
6050 &hdr->b_l1hdr.b_refcnt));
6051 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
6052 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
6053 } else if (HDR_IO_IN_PROGRESS(hdr)) {
6055 * If this header already had an IO in progress
6056 * and we are performing another IO to fetch
6057 * encrypted data we must wait until the first
6058 * IO completes so as not to confuse
6059 * arc_read_done(). This should be very rare
6060 * and so the performance impact shouldn't
6063 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6064 mutex_exit(hash_lock);
6069 * This is a delicate dance that we play here.
6070 * This hdr might be in the ghost list so we access
6071 * it to move it out of the ghost list before we
6072 * initiate the read. If it's a prefetch then
6073 * it won't have a callback so we'll remove the
6074 * reference that arc_buf_alloc_impl() created. We
6075 * do this after we've called arc_access() to
6076 * avoid hitting an assert in remove_reference().
6078 arc_adapt(arc_hdr_size(hdr), hdr->b_l1hdr.b_state);
6079 arc_access(hdr, hash_lock);
6080 arc_hdr_alloc_abd(hdr, alloc_flags);
6083 if (encrypted_read) {
6084 ASSERT(HDR_HAS_RABD(hdr));
6085 size = HDR_GET_PSIZE(hdr);
6086 hdr_abd = hdr->b_crypt_hdr.b_rabd;
6087 zio_flags |= ZIO_FLAG_RAW;
6089 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6090 size = arc_hdr_size(hdr);
6091 hdr_abd = hdr->b_l1hdr.b_pabd;
6093 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
6094 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6098 * For authenticated bp's, we do not ask the ZIO layer
6099 * to authenticate them since this will cause the entire
6100 * IO to fail if the key isn't loaded. Instead, we
6101 * defer authentication until arc_buf_fill(), which will
6102 * verify the data when the key is available.
6104 if (BP_IS_AUTHENTICATED(bp))
6105 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
6108 if (*arc_flags & ARC_FLAG_PREFETCH &&
6109 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))
6110 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6111 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6112 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6113 if (*arc_flags & ARC_FLAG_L2CACHE)
6114 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6115 if (BP_IS_AUTHENTICATED(bp))
6116 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6117 if (BP_GET_LEVEL(bp) > 0)
6118 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
6119 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
6120 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
6121 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
6123 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
6124 acb->acb_done = done;
6125 acb->acb_private = private;
6126 acb->acb_compressed = compressed_read;
6127 acb->acb_encrypted = encrypted_read;
6128 acb->acb_noauth = noauth_read;
6131 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6132 hdr->b_l1hdr.b_acb = acb;
6133 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6135 if (HDR_HAS_L2HDR(hdr) &&
6136 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
6137 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
6138 addr = hdr->b_l2hdr.b_daddr;
6140 * Lock out L2ARC device removal.
6142 if (vdev_is_dead(vd) ||
6143 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
6148 * We count both async reads and scrub IOs as asynchronous so
6149 * that both can be upgraded in the event of a cache hit while
6150 * the read IO is still in-flight.
6152 if (priority == ZIO_PRIORITY_ASYNC_READ ||
6153 priority == ZIO_PRIORITY_SCRUB)
6154 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6156 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6159 * At this point, we have a level 1 cache miss or a blkptr
6160 * with embedded data. Try again in L2ARC if possible.
6162 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
6165 * Skip ARC stat bump for block pointers with embedded
6166 * data. The data are read from the blkptr itself via
6167 * decode_embedded_bp_compressed().
6170 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr,
6171 blkptr_t *, bp, uint64_t, lsize,
6172 zbookmark_phys_t *, zb);
6173 ARCSTAT_BUMP(arcstat_misses);
6174 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6175 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data,
6179 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
6181 * Read from the L2ARC if the following are true:
6182 * 1. The L2ARC vdev was previously cached.
6183 * 2. This buffer still has L2ARC metadata.
6184 * 3. This buffer isn't currently writing to the L2ARC.
6185 * 4. The L2ARC entry wasn't evicted, which may
6186 * also have invalidated the vdev.
6187 * 5. This isn't prefetch and l2arc_noprefetch is set.
6189 if (HDR_HAS_L2HDR(hdr) &&
6190 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
6191 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
6192 l2arc_read_callback_t *cb;
6196 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
6197 ARCSTAT_BUMP(arcstat_l2_hits);
6198 atomic_inc_32(&hdr->b_l2hdr.b_hits);
6200 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
6202 cb->l2rcb_hdr = hdr;
6205 cb->l2rcb_flags = zio_flags;
6208 * When Compressed ARC is disabled, but the
6209 * L2ARC block is compressed, arc_hdr_size()
6210 * will have returned LSIZE rather than PSIZE.
6212 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
6213 !HDR_COMPRESSION_ENABLED(hdr) &&
6214 HDR_GET_PSIZE(hdr) != 0) {
6215 size = HDR_GET_PSIZE(hdr);
6218 asize = vdev_psize_to_asize(vd, size);
6219 if (asize != size) {
6220 abd = abd_alloc_for_io(asize,
6221 HDR_ISTYPE_METADATA(hdr));
6222 cb->l2rcb_abd = abd;
6227 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6228 addr + asize <= vd->vdev_psize -
6229 VDEV_LABEL_END_SIZE);
6232 * l2arc read. The SCL_L2ARC lock will be
6233 * released by l2arc_read_done().
6234 * Issue a null zio if the underlying buffer
6235 * was squashed to zero size by compression.
6237 ASSERT3U(arc_hdr_get_compress(hdr), !=,
6238 ZIO_COMPRESS_EMPTY);
6239 rzio = zio_read_phys(pio, vd, addr,
6242 l2arc_read_done, cb, priority,
6243 zio_flags | ZIO_FLAG_DONT_CACHE |
6245 ZIO_FLAG_DONT_PROPAGATE |
6246 ZIO_FLAG_DONT_RETRY, B_FALSE);
6247 acb->acb_zio_head = rzio;
6249 if (hash_lock != NULL)
6250 mutex_exit(hash_lock);
6252 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6254 ARCSTAT_INCR(arcstat_l2_read_bytes,
6255 HDR_GET_PSIZE(hdr));
6257 if (*arc_flags & ARC_FLAG_NOWAIT) {
6262 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6263 if (zio_wait(rzio) == 0)
6266 /* l2arc read error; goto zio_read() */
6267 if (hash_lock != NULL)
6268 mutex_enter(hash_lock);
6270 DTRACE_PROBE1(l2arc__miss,
6271 arc_buf_hdr_t *, hdr);
6272 ARCSTAT_BUMP(arcstat_l2_misses);
6273 if (HDR_L2_WRITING(hdr))
6274 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6275 spa_config_exit(spa, SCL_L2ARC, vd);
6279 spa_config_exit(spa, SCL_L2ARC, vd);
6281 * Skip ARC stat bump for block pointers with
6282 * embedded data. The data are read from the blkptr
6283 * itself via decode_embedded_bp_compressed().
6285 if (l2arc_ndev != 0 && !embedded_bp) {
6286 DTRACE_PROBE1(l2arc__miss,
6287 arc_buf_hdr_t *, hdr);
6288 ARCSTAT_BUMP(arcstat_l2_misses);
6292 rzio = zio_read(pio, spa, bp, hdr_abd, size,
6293 arc_read_done, hdr, priority, zio_flags, zb);
6294 acb->acb_zio_head = rzio;
6296 if (hash_lock != NULL)
6297 mutex_exit(hash_lock);
6299 if (*arc_flags & ARC_FLAG_WAIT) {
6300 rc = zio_wait(rzio);
6304 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6309 /* embedded bps don't actually go to disk */
6311 spa_read_history_add(spa, zb, *arc_flags);
6312 spl_fstrans_unmark(cookie);
6317 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6321 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6323 p->p_private = private;
6324 list_link_init(&p->p_node);
6325 zfs_refcount_create(&p->p_refcnt);
6327 mutex_enter(&arc_prune_mtx);
6328 zfs_refcount_add(&p->p_refcnt, &arc_prune_list);
6329 list_insert_head(&arc_prune_list, p);
6330 mutex_exit(&arc_prune_mtx);
6336 arc_remove_prune_callback(arc_prune_t *p)
6338 boolean_t wait = B_FALSE;
6339 mutex_enter(&arc_prune_mtx);
6340 list_remove(&arc_prune_list, p);
6341 if (zfs_refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6343 mutex_exit(&arc_prune_mtx);
6345 /* wait for arc_prune_task to finish */
6347 taskq_wait_outstanding(arc_prune_taskq, 0);
6348 ASSERT0(zfs_refcount_count(&p->p_refcnt));
6349 zfs_refcount_destroy(&p->p_refcnt);
6350 kmem_free(p, sizeof (*p));
6354 * Notify the arc that a block was freed, and thus will never be used again.
6357 arc_freed(spa_t *spa, const blkptr_t *bp)
6360 kmutex_t *hash_lock;
6361 uint64_t guid = spa_load_guid(spa);
6363 ASSERT(!BP_IS_EMBEDDED(bp));
6365 hdr = buf_hash_find(guid, bp, &hash_lock);
6370 * We might be trying to free a block that is still doing I/O
6371 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6372 * dmu_sync-ed block). If this block is being prefetched, then it
6373 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6374 * until the I/O completes. A block may also have a reference if it is
6375 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6376 * have written the new block to its final resting place on disk but
6377 * without the dedup flag set. This would have left the hdr in the MRU
6378 * state and discoverable. When the txg finally syncs it detects that
6379 * the block was overridden in open context and issues an override I/O.
6380 * Since this is a dedup block, the override I/O will determine if the
6381 * block is already in the DDT. If so, then it will replace the io_bp
6382 * with the bp from the DDT and allow the I/O to finish. When the I/O
6383 * reaches the done callback, dbuf_write_override_done, it will
6384 * check to see if the io_bp and io_bp_override are identical.
6385 * If they are not, then it indicates that the bp was replaced with
6386 * the bp in the DDT and the override bp is freed. This allows
6387 * us to arrive here with a reference on a block that is being
6388 * freed. So if we have an I/O in progress, or a reference to
6389 * this hdr, then we don't destroy the hdr.
6391 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6392 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6393 arc_change_state(arc_anon, hdr, hash_lock);
6394 arc_hdr_destroy(hdr);
6395 mutex_exit(hash_lock);
6397 mutex_exit(hash_lock);
6403 * Release this buffer from the cache, making it an anonymous buffer. This
6404 * must be done after a read and prior to modifying the buffer contents.
6405 * If the buffer has more than one reference, we must make
6406 * a new hdr for the buffer.
6409 arc_release(arc_buf_t *buf, void *tag)
6411 arc_buf_hdr_t *hdr = buf->b_hdr;
6414 * It would be nice to assert that if its DMU metadata (level >
6415 * 0 || it's the dnode file), then it must be syncing context.
6416 * But we don't know that information at this level.
6419 mutex_enter(&buf->b_evict_lock);
6421 ASSERT(HDR_HAS_L1HDR(hdr));
6424 * We don't grab the hash lock prior to this check, because if
6425 * the buffer's header is in the arc_anon state, it won't be
6426 * linked into the hash table.
6428 if (hdr->b_l1hdr.b_state == arc_anon) {
6429 mutex_exit(&buf->b_evict_lock);
6430 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6431 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6432 ASSERT(!HDR_HAS_L2HDR(hdr));
6433 ASSERT(HDR_EMPTY(hdr));
6435 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6436 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6437 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6439 hdr->b_l1hdr.b_arc_access = 0;
6442 * If the buf is being overridden then it may already
6443 * have a hdr that is not empty.
6445 buf_discard_identity(hdr);
6451 kmutex_t *hash_lock = HDR_LOCK(hdr);
6452 mutex_enter(hash_lock);
6455 * This assignment is only valid as long as the hash_lock is
6456 * held, we must be careful not to reference state or the
6457 * b_state field after dropping the lock.
6459 arc_state_t *state = hdr->b_l1hdr.b_state;
6460 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6461 ASSERT3P(state, !=, arc_anon);
6463 /* this buffer is not on any list */
6464 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6466 if (HDR_HAS_L2HDR(hdr)) {
6467 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6470 * We have to recheck this conditional again now that
6471 * we're holding the l2ad_mtx to prevent a race with
6472 * another thread which might be concurrently calling
6473 * l2arc_evict(). In that case, l2arc_evict() might have
6474 * destroyed the header's L2 portion as we were waiting
6475 * to acquire the l2ad_mtx.
6477 if (HDR_HAS_L2HDR(hdr))
6478 arc_hdr_l2hdr_destroy(hdr);
6480 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6484 * Do we have more than one buf?
6486 if (hdr->b_l1hdr.b_bufcnt > 1) {
6487 arc_buf_hdr_t *nhdr;
6488 uint64_t spa = hdr->b_spa;
6489 uint64_t psize = HDR_GET_PSIZE(hdr);
6490 uint64_t lsize = HDR_GET_LSIZE(hdr);
6491 boolean_t protected = HDR_PROTECTED(hdr);
6492 enum zio_compress compress = arc_hdr_get_compress(hdr);
6493 arc_buf_contents_t type = arc_buf_type(hdr);
6494 VERIFY3U(hdr->b_type, ==, type);
6496 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6497 (void) remove_reference(hdr, hash_lock, tag);
6499 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6500 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6501 ASSERT(ARC_BUF_LAST(buf));
6505 * Pull the data off of this hdr and attach it to
6506 * a new anonymous hdr. Also find the last buffer
6507 * in the hdr's buffer list.
6509 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6510 ASSERT3P(lastbuf, !=, NULL);
6513 * If the current arc_buf_t and the hdr are sharing their data
6514 * buffer, then we must stop sharing that block.
6516 if (arc_buf_is_shared(buf)) {
6517 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6518 VERIFY(!arc_buf_is_shared(lastbuf));
6521 * First, sever the block sharing relationship between
6522 * buf and the arc_buf_hdr_t.
6524 arc_unshare_buf(hdr, buf);
6527 * Now we need to recreate the hdr's b_pabd. Since we
6528 * have lastbuf handy, we try to share with it, but if
6529 * we can't then we allocate a new b_pabd and copy the
6530 * data from buf into it.
6532 if (arc_can_share(hdr, lastbuf)) {
6533 arc_share_buf(hdr, lastbuf);
6535 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT);
6536 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6537 buf->b_data, psize);
6539 VERIFY3P(lastbuf->b_data, !=, NULL);
6540 } else if (HDR_SHARED_DATA(hdr)) {
6542 * Uncompressed shared buffers are always at the end
6543 * of the list. Compressed buffers don't have the
6544 * same requirements. This makes it hard to
6545 * simply assert that the lastbuf is shared so
6546 * we rely on the hdr's compression flags to determine
6547 * if we have a compressed, shared buffer.
6549 ASSERT(arc_buf_is_shared(lastbuf) ||
6550 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6551 ASSERT(!ARC_BUF_SHARED(buf));
6554 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6555 ASSERT3P(state, !=, arc_l2c_only);
6557 (void) zfs_refcount_remove_many(&state->arcs_size,
6558 arc_buf_size(buf), buf);
6560 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6561 ASSERT3P(state, !=, arc_l2c_only);
6562 (void) zfs_refcount_remove_many(
6563 &state->arcs_esize[type],
6564 arc_buf_size(buf), buf);
6567 hdr->b_l1hdr.b_bufcnt -= 1;
6568 if (ARC_BUF_ENCRYPTED(buf))
6569 hdr->b_crypt_hdr.b_ebufcnt -= 1;
6571 arc_cksum_verify(buf);
6572 arc_buf_unwatch(buf);
6574 /* if this is the last uncompressed buf free the checksum */
6575 if (!arc_hdr_has_uncompressed_buf(hdr))
6576 arc_cksum_free(hdr);
6578 mutex_exit(hash_lock);
6581 * Allocate a new hdr. The new hdr will contain a b_pabd
6582 * buffer which will be freed in arc_write().
6584 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6585 compress, hdr->b_complevel, type, HDR_HAS_RABD(hdr));
6586 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6587 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6588 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6589 VERIFY3U(nhdr->b_type, ==, type);
6590 ASSERT(!HDR_SHARED_DATA(nhdr));
6592 nhdr->b_l1hdr.b_buf = buf;
6593 nhdr->b_l1hdr.b_bufcnt = 1;
6594 if (ARC_BUF_ENCRYPTED(buf))
6595 nhdr->b_crypt_hdr.b_ebufcnt = 1;
6596 nhdr->b_l1hdr.b_mru_hits = 0;
6597 nhdr->b_l1hdr.b_mru_ghost_hits = 0;
6598 nhdr->b_l1hdr.b_mfu_hits = 0;
6599 nhdr->b_l1hdr.b_mfu_ghost_hits = 0;
6600 nhdr->b_l1hdr.b_l2_hits = 0;
6601 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6604 mutex_exit(&buf->b_evict_lock);
6605 (void) zfs_refcount_add_many(&arc_anon->arcs_size,
6606 arc_buf_size(buf), buf);
6608 mutex_exit(&buf->b_evict_lock);
6609 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6610 /* protected by hash lock, or hdr is on arc_anon */
6611 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6612 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6613 hdr->b_l1hdr.b_mru_hits = 0;
6614 hdr->b_l1hdr.b_mru_ghost_hits = 0;
6615 hdr->b_l1hdr.b_mfu_hits = 0;
6616 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
6617 hdr->b_l1hdr.b_l2_hits = 0;
6618 arc_change_state(arc_anon, hdr, hash_lock);
6619 hdr->b_l1hdr.b_arc_access = 0;
6621 mutex_exit(hash_lock);
6622 buf_discard_identity(hdr);
6628 arc_released(arc_buf_t *buf)
6632 mutex_enter(&buf->b_evict_lock);
6633 released = (buf->b_data != NULL &&
6634 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6635 mutex_exit(&buf->b_evict_lock);
6641 arc_referenced(arc_buf_t *buf)
6645 mutex_enter(&buf->b_evict_lock);
6646 referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6647 mutex_exit(&buf->b_evict_lock);
6648 return (referenced);
6653 arc_write_ready(zio_t *zio)
6655 arc_write_callback_t *callback = zio->io_private;
6656 arc_buf_t *buf = callback->awcb_buf;
6657 arc_buf_hdr_t *hdr = buf->b_hdr;
6658 blkptr_t *bp = zio->io_bp;
6659 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6660 fstrans_cookie_t cookie = spl_fstrans_mark();
6662 ASSERT(HDR_HAS_L1HDR(hdr));
6663 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6664 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6667 * If we're reexecuting this zio because the pool suspended, then
6668 * cleanup any state that was previously set the first time the
6669 * callback was invoked.
6671 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6672 arc_cksum_free(hdr);
6673 arc_buf_unwatch(buf);
6674 if (hdr->b_l1hdr.b_pabd != NULL) {
6675 if (arc_buf_is_shared(buf)) {
6676 arc_unshare_buf(hdr, buf);
6678 arc_hdr_free_abd(hdr, B_FALSE);
6682 if (HDR_HAS_RABD(hdr))
6683 arc_hdr_free_abd(hdr, B_TRUE);
6685 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6686 ASSERT(!HDR_HAS_RABD(hdr));
6687 ASSERT(!HDR_SHARED_DATA(hdr));
6688 ASSERT(!arc_buf_is_shared(buf));
6690 callback->awcb_ready(zio, buf, callback->awcb_private);
6692 if (HDR_IO_IN_PROGRESS(hdr))
6693 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6695 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6697 if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr))
6698 hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp));
6700 if (BP_IS_PROTECTED(bp)) {
6701 /* ZIL blocks are written through zio_rewrite */
6702 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6703 ASSERT(HDR_PROTECTED(hdr));
6705 if (BP_SHOULD_BYTESWAP(bp)) {
6706 if (BP_GET_LEVEL(bp) > 0) {
6707 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6709 hdr->b_l1hdr.b_byteswap =
6710 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6713 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6716 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6717 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6718 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6719 hdr->b_crypt_hdr.b_iv);
6720 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6724 * If this block was written for raw encryption but the zio layer
6725 * ended up only authenticating it, adjust the buffer flags now.
6727 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6728 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6729 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6730 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6731 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6732 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6733 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6734 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6737 /* this must be done after the buffer flags are adjusted */
6738 arc_cksum_compute(buf);
6740 enum zio_compress compress;
6741 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6742 compress = ZIO_COMPRESS_OFF;
6744 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6745 compress = BP_GET_COMPRESS(bp);
6747 HDR_SET_PSIZE(hdr, psize);
6748 arc_hdr_set_compress(hdr, compress);
6749 hdr->b_complevel = zio->io_prop.zp_complevel;
6751 if (zio->io_error != 0 || psize == 0)
6755 * Fill the hdr with data. If the buffer is encrypted we have no choice
6756 * but to copy the data into b_radb. If the hdr is compressed, the data
6757 * we want is available from the zio, otherwise we can take it from
6760 * We might be able to share the buf's data with the hdr here. However,
6761 * doing so would cause the ARC to be full of linear ABDs if we write a
6762 * lot of shareable data. As a compromise, we check whether scattered
6763 * ABDs are allowed, and assume that if they are then the user wants
6764 * the ARC to be primarily filled with them regardless of the data being
6765 * written. Therefore, if they're allowed then we allocate one and copy
6766 * the data into it; otherwise, we share the data directly if we can.
6768 if (ARC_BUF_ENCRYPTED(buf)) {
6769 ASSERT3U(psize, >, 0);
6770 ASSERT(ARC_BUF_COMPRESSED(buf));
6771 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT|ARC_HDR_ALLOC_RDATA);
6772 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6773 } else if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6775 * Ideally, we would always copy the io_abd into b_pabd, but the
6776 * user may have disabled compressed ARC, thus we must check the
6777 * hdr's compression setting rather than the io_bp's.
6779 if (BP_IS_ENCRYPTED(bp)) {
6780 ASSERT3U(psize, >, 0);
6781 arc_hdr_alloc_abd(hdr,
6782 ARC_HDR_DO_ADAPT|ARC_HDR_ALLOC_RDATA);
6783 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6784 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
6785 !ARC_BUF_COMPRESSED(buf)) {
6786 ASSERT3U(psize, >, 0);
6787 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT);
6788 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6790 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6791 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT);
6792 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6796 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6797 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6798 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6800 arc_share_buf(hdr, buf);
6804 arc_hdr_verify(hdr, bp);
6805 spl_fstrans_unmark(cookie);
6809 arc_write_children_ready(zio_t *zio)
6811 arc_write_callback_t *callback = zio->io_private;
6812 arc_buf_t *buf = callback->awcb_buf;
6814 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6818 * The SPA calls this callback for each physical write that happens on behalf
6819 * of a logical write. See the comment in dbuf_write_physdone() for details.
6822 arc_write_physdone(zio_t *zio)
6824 arc_write_callback_t *cb = zio->io_private;
6825 if (cb->awcb_physdone != NULL)
6826 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6830 arc_write_done(zio_t *zio)
6832 arc_write_callback_t *callback = zio->io_private;
6833 arc_buf_t *buf = callback->awcb_buf;
6834 arc_buf_hdr_t *hdr = buf->b_hdr;
6836 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6838 if (zio->io_error == 0) {
6839 arc_hdr_verify(hdr, zio->io_bp);
6841 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6842 buf_discard_identity(hdr);
6844 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6845 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6848 ASSERT(HDR_EMPTY(hdr));
6852 * If the block to be written was all-zero or compressed enough to be
6853 * embedded in the BP, no write was performed so there will be no
6854 * dva/birth/checksum. The buffer must therefore remain anonymous
6857 if (!HDR_EMPTY(hdr)) {
6858 arc_buf_hdr_t *exists;
6859 kmutex_t *hash_lock;
6861 ASSERT3U(zio->io_error, ==, 0);
6863 arc_cksum_verify(buf);
6865 exists = buf_hash_insert(hdr, &hash_lock);
6866 if (exists != NULL) {
6868 * This can only happen if we overwrite for
6869 * sync-to-convergence, because we remove
6870 * buffers from the hash table when we arc_free().
6872 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6873 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6874 panic("bad overwrite, hdr=%p exists=%p",
6875 (void *)hdr, (void *)exists);
6876 ASSERT(zfs_refcount_is_zero(
6877 &exists->b_l1hdr.b_refcnt));
6878 arc_change_state(arc_anon, exists, hash_lock);
6879 arc_hdr_destroy(exists);
6880 mutex_exit(hash_lock);
6881 exists = buf_hash_insert(hdr, &hash_lock);
6882 ASSERT3P(exists, ==, NULL);
6883 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6885 ASSERT(zio->io_prop.zp_nopwrite);
6886 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6887 panic("bad nopwrite, hdr=%p exists=%p",
6888 (void *)hdr, (void *)exists);
6891 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6892 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6893 ASSERT(BP_GET_DEDUP(zio->io_bp));
6894 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6897 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6898 /* if it's not anon, we are doing a scrub */
6899 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6900 arc_access(hdr, hash_lock);
6901 mutex_exit(hash_lock);
6903 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6906 ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6907 callback->awcb_done(zio, buf, callback->awcb_private);
6909 abd_put(zio->io_abd);
6910 kmem_free(callback, sizeof (arc_write_callback_t));
6914 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
6915 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc,
6916 const zio_prop_t *zp, arc_write_done_func_t *ready,
6917 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
6918 arc_write_done_func_t *done, void *private, zio_priority_t priority,
6919 int zio_flags, const zbookmark_phys_t *zb)
6921 arc_buf_hdr_t *hdr = buf->b_hdr;
6922 arc_write_callback_t *callback;
6924 zio_prop_t localprop = *zp;
6926 ASSERT3P(ready, !=, NULL);
6927 ASSERT3P(done, !=, NULL);
6928 ASSERT(!HDR_IO_ERROR(hdr));
6929 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6930 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6931 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6933 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6935 if (ARC_BUF_ENCRYPTED(buf)) {
6936 ASSERT(ARC_BUF_COMPRESSED(buf));
6937 localprop.zp_encrypt = B_TRUE;
6938 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6939 localprop.zp_complevel = hdr->b_complevel;
6940 localprop.zp_byteorder =
6941 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
6942 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
6943 bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt,
6945 bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv,
6947 bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac,
6949 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
6950 localprop.zp_nopwrite = B_FALSE;
6951 localprop.zp_copies =
6952 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
6954 zio_flags |= ZIO_FLAG_RAW;
6955 } else if (ARC_BUF_COMPRESSED(buf)) {
6956 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6957 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6958 localprop.zp_complevel = hdr->b_complevel;
6959 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6961 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6962 callback->awcb_ready = ready;
6963 callback->awcb_children_ready = children_ready;
6964 callback->awcb_physdone = physdone;
6965 callback->awcb_done = done;
6966 callback->awcb_private = private;
6967 callback->awcb_buf = buf;
6970 * The hdr's b_pabd is now stale, free it now. A new data block
6971 * will be allocated when the zio pipeline calls arc_write_ready().
6973 if (hdr->b_l1hdr.b_pabd != NULL) {
6975 * If the buf is currently sharing the data block with
6976 * the hdr then we need to break that relationship here.
6977 * The hdr will remain with a NULL data pointer and the
6978 * buf will take sole ownership of the block.
6980 if (arc_buf_is_shared(buf)) {
6981 arc_unshare_buf(hdr, buf);
6983 arc_hdr_free_abd(hdr, B_FALSE);
6985 VERIFY3P(buf->b_data, !=, NULL);
6988 if (HDR_HAS_RABD(hdr))
6989 arc_hdr_free_abd(hdr, B_TRUE);
6991 if (!(zio_flags & ZIO_FLAG_RAW))
6992 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6994 ASSERT(!arc_buf_is_shared(buf));
6995 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6997 zio = zio_write(pio, spa, txg, bp,
6998 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6999 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
7000 (children_ready != NULL) ? arc_write_children_ready : NULL,
7001 arc_write_physdone, arc_write_done, callback,
7002 priority, zio_flags, zb);
7008 arc_tempreserve_clear(uint64_t reserve)
7010 atomic_add_64(&arc_tempreserve, -reserve);
7011 ASSERT((int64_t)arc_tempreserve >= 0);
7015 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
7021 reserve > arc_c/4 &&
7022 reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT))
7023 arc_c = MIN(arc_c_max, reserve * 4);
7026 * Throttle when the calculated memory footprint for the TXG
7027 * exceeds the target ARC size.
7029 if (reserve > arc_c) {
7030 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
7031 return (SET_ERROR(ERESTART));
7035 * Don't count loaned bufs as in flight dirty data to prevent long
7036 * network delays from blocking transactions that are ready to be
7037 * assigned to a txg.
7040 /* assert that it has not wrapped around */
7041 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
7043 anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) -
7044 arc_loaned_bytes), 0);
7047 * Writes will, almost always, require additional memory allocations
7048 * in order to compress/encrypt/etc the data. We therefore need to
7049 * make sure that there is sufficient available memory for this.
7051 error = arc_memory_throttle(spa, reserve, txg);
7056 * Throttle writes when the amount of dirty data in the cache
7057 * gets too large. We try to keep the cache less than half full
7058 * of dirty blocks so that our sync times don't grow too large.
7060 * In the case of one pool being built on another pool, we want
7061 * to make sure we don't end up throttling the lower (backing)
7062 * pool when the upper pool is the majority contributor to dirty
7063 * data. To insure we make forward progress during throttling, we
7064 * also check the current pool's net dirty data and only throttle
7065 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
7066 * data in the cache.
7068 * Note: if two requests come in concurrently, we might let them
7069 * both succeed, when one of them should fail. Not a huge deal.
7071 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
7072 uint64_t spa_dirty_anon = spa_dirty_data(spa);
7074 if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
7075 anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
7076 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
7078 uint64_t meta_esize = zfs_refcount_count(
7079 &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7080 uint64_t data_esize =
7081 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7082 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
7083 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
7084 arc_tempreserve >> 10, meta_esize >> 10,
7085 data_esize >> 10, reserve >> 10, arc_c >> 10);
7087 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
7088 return (SET_ERROR(ERESTART));
7090 atomic_add_64(&arc_tempreserve, reserve);
7095 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
7096 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
7098 size->value.ui64 = zfs_refcount_count(&state->arcs_size);
7099 evict_data->value.ui64 =
7100 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
7101 evict_metadata->value.ui64 =
7102 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
7106 arc_kstat_update(kstat_t *ksp, int rw)
7108 arc_stats_t *as = ksp->ks_data;
7110 if (rw == KSTAT_WRITE) {
7111 return (SET_ERROR(EACCES));
7113 arc_kstat_update_state(arc_anon,
7114 &as->arcstat_anon_size,
7115 &as->arcstat_anon_evictable_data,
7116 &as->arcstat_anon_evictable_metadata);
7117 arc_kstat_update_state(arc_mru,
7118 &as->arcstat_mru_size,
7119 &as->arcstat_mru_evictable_data,
7120 &as->arcstat_mru_evictable_metadata);
7121 arc_kstat_update_state(arc_mru_ghost,
7122 &as->arcstat_mru_ghost_size,
7123 &as->arcstat_mru_ghost_evictable_data,
7124 &as->arcstat_mru_ghost_evictable_metadata);
7125 arc_kstat_update_state(arc_mfu,
7126 &as->arcstat_mfu_size,
7127 &as->arcstat_mfu_evictable_data,
7128 &as->arcstat_mfu_evictable_metadata);
7129 arc_kstat_update_state(arc_mfu_ghost,
7130 &as->arcstat_mfu_ghost_size,
7131 &as->arcstat_mfu_ghost_evictable_data,
7132 &as->arcstat_mfu_ghost_evictable_metadata);
7134 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
7135 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
7136 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
7137 ARCSTAT(arcstat_metadata_size) =
7138 aggsum_value(&astat_metadata_size);
7139 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
7140 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
7141 ARCSTAT(arcstat_dbuf_size) = aggsum_value(&astat_dbuf_size);
7142 #if defined(COMPAT_FREEBSD11)
7143 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_bonus_size) +
7144 aggsum_value(&astat_dnode_size) +
7145 aggsum_value(&astat_dbuf_size);
7147 ARCSTAT(arcstat_dnode_size) = aggsum_value(&astat_dnode_size);
7148 ARCSTAT(arcstat_bonus_size) = aggsum_value(&astat_bonus_size);
7149 ARCSTAT(arcstat_abd_chunk_waste_size) =
7150 aggsum_value(&astat_abd_chunk_waste_size);
7152 as->arcstat_memory_all_bytes.value.ui64 =
7154 as->arcstat_memory_free_bytes.value.ui64 =
7156 as->arcstat_memory_available_bytes.value.i64 =
7157 arc_available_memory();
7164 * This function *must* return indices evenly distributed between all
7165 * sublists of the multilist. This is needed due to how the ARC eviction
7166 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7167 * distributed between all sublists and uses this assumption when
7168 * deciding which sublist to evict from and how much to evict from it.
7171 arc_state_multilist_index_func(multilist_t *ml, void *obj)
7173 arc_buf_hdr_t *hdr = obj;
7176 * We rely on b_dva to generate evenly distributed index
7177 * numbers using buf_hash below. So, as an added precaution,
7178 * let's make sure we never add empty buffers to the arc lists.
7180 ASSERT(!HDR_EMPTY(hdr));
7183 * The assumption here, is the hash value for a given
7184 * arc_buf_hdr_t will remain constant throughout its lifetime
7185 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7186 * Thus, we don't need to store the header's sublist index
7187 * on insertion, as this index can be recalculated on removal.
7189 * Also, the low order bits of the hash value are thought to be
7190 * distributed evenly. Otherwise, in the case that the multilist
7191 * has a power of two number of sublists, each sublists' usage
7192 * would not be evenly distributed.
7194 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
7195 multilist_get_num_sublists(ml));
7198 #define WARN_IF_TUNING_IGNORED(tuning, value, do_warn) do { \
7199 if ((do_warn) && (tuning) && ((tuning) != (value))) { \
7201 "ignoring tunable %s (using %llu instead)", \
7202 (#tuning), (value)); \
7207 * Called during module initialization and periodically thereafter to
7208 * apply reasonable changes to the exposed performance tunings. Can also be
7209 * called explicitly by param_set_arc_*() functions when ARC tunables are
7210 * updated manually. Non-zero zfs_* values which differ from the currently set
7211 * values will be applied.
7214 arc_tuning_update(boolean_t verbose)
7216 uint64_t allmem = arc_all_memory();
7217 unsigned long limit;
7219 /* Valid range: 32M - <arc_c_max> */
7220 if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) &&
7221 (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) &&
7222 (zfs_arc_min <= arc_c_max)) {
7223 arc_c_min = zfs_arc_min;
7224 arc_c = MAX(arc_c, arc_c_min);
7226 WARN_IF_TUNING_IGNORED(zfs_arc_min, arc_c_min, verbose);
7228 /* Valid range: 64M - <all physical memory> */
7229 if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) &&
7230 (zfs_arc_max >= 64 << 20) && (zfs_arc_max < allmem) &&
7231 (zfs_arc_max > arc_c_min)) {
7232 arc_c_max = zfs_arc_max;
7233 arc_c = MIN(arc_c, arc_c_max);
7234 arc_p = (arc_c >> 1);
7235 if (arc_meta_limit > arc_c_max)
7236 arc_meta_limit = arc_c_max;
7237 if (arc_dnode_size_limit > arc_meta_limit)
7238 arc_dnode_size_limit = arc_meta_limit;
7240 WARN_IF_TUNING_IGNORED(zfs_arc_max, arc_c_max, verbose);
7242 /* Valid range: 16M - <arc_c_max> */
7243 if ((zfs_arc_meta_min) && (zfs_arc_meta_min != arc_meta_min) &&
7244 (zfs_arc_meta_min >= 1ULL << SPA_MAXBLOCKSHIFT) &&
7245 (zfs_arc_meta_min <= arc_c_max)) {
7246 arc_meta_min = zfs_arc_meta_min;
7247 if (arc_meta_limit < arc_meta_min)
7248 arc_meta_limit = arc_meta_min;
7249 if (arc_dnode_size_limit < arc_meta_min)
7250 arc_dnode_size_limit = arc_meta_min;
7252 WARN_IF_TUNING_IGNORED(zfs_arc_meta_min, arc_meta_min, verbose);
7254 /* Valid range: <arc_meta_min> - <arc_c_max> */
7255 limit = zfs_arc_meta_limit ? zfs_arc_meta_limit :
7256 MIN(zfs_arc_meta_limit_percent, 100) * arc_c_max / 100;
7257 if ((limit != arc_meta_limit) &&
7258 (limit >= arc_meta_min) &&
7259 (limit <= arc_c_max))
7260 arc_meta_limit = limit;
7261 WARN_IF_TUNING_IGNORED(zfs_arc_meta_limit, arc_meta_limit, verbose);
7263 /* Valid range: <arc_meta_min> - <arc_meta_limit> */
7264 limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit :
7265 MIN(zfs_arc_dnode_limit_percent, 100) * arc_meta_limit / 100;
7266 if ((limit != arc_dnode_size_limit) &&
7267 (limit >= arc_meta_min) &&
7268 (limit <= arc_meta_limit))
7269 arc_dnode_size_limit = limit;
7270 WARN_IF_TUNING_IGNORED(zfs_arc_dnode_limit, arc_dnode_size_limit,
7273 /* Valid range: 1 - N */
7274 if (zfs_arc_grow_retry)
7275 arc_grow_retry = zfs_arc_grow_retry;
7277 /* Valid range: 1 - N */
7278 if (zfs_arc_shrink_shift) {
7279 arc_shrink_shift = zfs_arc_shrink_shift;
7280 arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1);
7283 /* Valid range: 1 - N */
7284 if (zfs_arc_p_min_shift)
7285 arc_p_min_shift = zfs_arc_p_min_shift;
7287 /* Valid range: 1 - N ms */
7288 if (zfs_arc_min_prefetch_ms)
7289 arc_min_prefetch_ms = zfs_arc_min_prefetch_ms;
7291 /* Valid range: 1 - N ms */
7292 if (zfs_arc_min_prescient_prefetch_ms) {
7293 arc_min_prescient_prefetch_ms =
7294 zfs_arc_min_prescient_prefetch_ms;
7297 /* Valid range: 0 - 100 */
7298 if ((zfs_arc_lotsfree_percent >= 0) &&
7299 (zfs_arc_lotsfree_percent <= 100))
7300 arc_lotsfree_percent = zfs_arc_lotsfree_percent;
7301 WARN_IF_TUNING_IGNORED(zfs_arc_lotsfree_percent, arc_lotsfree_percent,
7304 /* Valid range: 0 - <all physical memory> */
7305 if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free))
7306 arc_sys_free = MIN(MAX(zfs_arc_sys_free, 0), allmem);
7307 WARN_IF_TUNING_IGNORED(zfs_arc_sys_free, arc_sys_free, verbose);
7311 arc_state_init(void)
7313 arc_anon = &ARC_anon;
7315 arc_mru_ghost = &ARC_mru_ghost;
7317 arc_mfu_ghost = &ARC_mfu_ghost;
7318 arc_l2c_only = &ARC_l2c_only;
7320 arc_mru->arcs_list[ARC_BUFC_METADATA] =
7321 multilist_create(sizeof (arc_buf_hdr_t),
7322 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7323 arc_state_multilist_index_func);
7324 arc_mru->arcs_list[ARC_BUFC_DATA] =
7325 multilist_create(sizeof (arc_buf_hdr_t),
7326 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7327 arc_state_multilist_index_func);
7328 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
7329 multilist_create(sizeof (arc_buf_hdr_t),
7330 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7331 arc_state_multilist_index_func);
7332 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
7333 multilist_create(sizeof (arc_buf_hdr_t),
7334 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7335 arc_state_multilist_index_func);
7336 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
7337 multilist_create(sizeof (arc_buf_hdr_t),
7338 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7339 arc_state_multilist_index_func);
7340 arc_mfu->arcs_list[ARC_BUFC_DATA] =
7341 multilist_create(sizeof (arc_buf_hdr_t),
7342 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7343 arc_state_multilist_index_func);
7344 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
7345 multilist_create(sizeof (arc_buf_hdr_t),
7346 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7347 arc_state_multilist_index_func);
7348 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
7349 multilist_create(sizeof (arc_buf_hdr_t),
7350 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7351 arc_state_multilist_index_func);
7352 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
7353 multilist_create(sizeof (arc_buf_hdr_t),
7354 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7355 arc_state_multilist_index_func);
7356 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
7357 multilist_create(sizeof (arc_buf_hdr_t),
7358 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7359 arc_state_multilist_index_func);
7361 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7362 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7363 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7364 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7365 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7366 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7367 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7368 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7369 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7370 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7371 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7372 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7374 zfs_refcount_create(&arc_anon->arcs_size);
7375 zfs_refcount_create(&arc_mru->arcs_size);
7376 zfs_refcount_create(&arc_mru_ghost->arcs_size);
7377 zfs_refcount_create(&arc_mfu->arcs_size);
7378 zfs_refcount_create(&arc_mfu_ghost->arcs_size);
7379 zfs_refcount_create(&arc_l2c_only->arcs_size);
7381 aggsum_init(&arc_meta_used, 0);
7382 aggsum_init(&arc_size, 0);
7383 aggsum_init(&astat_data_size, 0);
7384 aggsum_init(&astat_metadata_size, 0);
7385 aggsum_init(&astat_hdr_size, 0);
7386 aggsum_init(&astat_l2_hdr_size, 0);
7387 aggsum_init(&astat_bonus_size, 0);
7388 aggsum_init(&astat_dnode_size, 0);
7389 aggsum_init(&astat_dbuf_size, 0);
7390 aggsum_init(&astat_abd_chunk_waste_size, 0);
7392 arc_anon->arcs_state = ARC_STATE_ANON;
7393 arc_mru->arcs_state = ARC_STATE_MRU;
7394 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7395 arc_mfu->arcs_state = ARC_STATE_MFU;
7396 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7397 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7401 arc_state_fini(void)
7403 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7404 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7405 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7406 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7407 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7408 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7409 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7410 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7411 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7412 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7413 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7414 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7416 zfs_refcount_destroy(&arc_anon->arcs_size);
7417 zfs_refcount_destroy(&arc_mru->arcs_size);
7418 zfs_refcount_destroy(&arc_mru_ghost->arcs_size);
7419 zfs_refcount_destroy(&arc_mfu->arcs_size);
7420 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
7421 zfs_refcount_destroy(&arc_l2c_only->arcs_size);
7423 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
7424 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7425 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7426 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7427 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
7428 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7429 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
7430 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7431 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7432 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7434 aggsum_fini(&arc_meta_used);
7435 aggsum_fini(&arc_size);
7436 aggsum_fini(&astat_data_size);
7437 aggsum_fini(&astat_metadata_size);
7438 aggsum_fini(&astat_hdr_size);
7439 aggsum_fini(&astat_l2_hdr_size);
7440 aggsum_fini(&astat_bonus_size);
7441 aggsum_fini(&astat_dnode_size);
7442 aggsum_fini(&astat_dbuf_size);
7443 aggsum_fini(&astat_abd_chunk_waste_size);
7447 arc_target_bytes(void)
7455 uint64_t percent, allmem = arc_all_memory();
7456 mutex_init(&arc_evict_lock, NULL, MUTEX_DEFAULT, NULL);
7457 list_create(&arc_evict_waiters, sizeof (arc_evict_waiter_t),
7458 offsetof(arc_evict_waiter_t, aew_node));
7460 arc_min_prefetch_ms = 1000;
7461 arc_min_prescient_prefetch_ms = 6000;
7463 #if defined(_KERNEL)
7467 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more. */
7468 arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT);
7470 /* How to set default max varies by platform. */
7471 arc_c_max = arc_default_max(arc_c_min, allmem);
7475 * In userland, there's only the memory pressure that we artificially
7476 * create (see arc_available_memory()). Don't let arc_c get too
7477 * small, because it can cause transactions to be larger than
7478 * arc_c, causing arc_tempreserve_space() to fail.
7480 arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT);
7484 arc_p = (arc_c >> 1);
7486 /* Set min to 1/2 of arc_c_min */
7487 arc_meta_min = 1ULL << SPA_MAXBLOCKSHIFT;
7488 /* Initialize maximum observed usage to zero */
7491 * Set arc_meta_limit to a percent of arc_c_max with a floor of
7492 * arc_meta_min, and a ceiling of arc_c_max.
7494 percent = MIN(zfs_arc_meta_limit_percent, 100);
7495 arc_meta_limit = MAX(arc_meta_min, (percent * arc_c_max) / 100);
7496 percent = MIN(zfs_arc_dnode_limit_percent, 100);
7497 arc_dnode_size_limit = (percent * arc_meta_limit) / 100;
7499 /* Apply user specified tunings */
7500 arc_tuning_update(B_TRUE);
7502 /* if kmem_flags are set, lets try to use less memory */
7503 if (kmem_debugging())
7505 if (arc_c < arc_c_min)
7512 list_create(&arc_prune_list, sizeof (arc_prune_t),
7513 offsetof(arc_prune_t, p_node));
7514 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7516 arc_prune_taskq = taskq_create("arc_prune", boot_ncpus, defclsyspri,
7517 boot_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
7519 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7520 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7522 if (arc_ksp != NULL) {
7523 arc_ksp->ks_data = &arc_stats;
7524 arc_ksp->ks_update = arc_kstat_update;
7525 kstat_install(arc_ksp);
7528 arc_evict_zthr = zthr_create_timer("arc_evict",
7529 arc_evict_cb_check, arc_evict_cb, NULL, SEC2NSEC(1));
7530 arc_reap_zthr = zthr_create_timer("arc_reap",
7531 arc_reap_cb_check, arc_reap_cb, NULL, SEC2NSEC(1));
7536 * Calculate maximum amount of dirty data per pool.
7538 * If it has been set by a module parameter, take that.
7539 * Otherwise, use a percentage of physical memory defined by
7540 * zfs_dirty_data_max_percent (default 10%) with a cap at
7541 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
7544 if (zfs_dirty_data_max_max == 0)
7545 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
7546 allmem * zfs_dirty_data_max_max_percent / 100);
7548 if (zfs_dirty_data_max_max == 0)
7549 zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024,
7550 allmem * zfs_dirty_data_max_max_percent / 100);
7553 if (zfs_dirty_data_max == 0) {
7554 zfs_dirty_data_max = allmem *
7555 zfs_dirty_data_max_percent / 100;
7556 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7557 zfs_dirty_data_max_max);
7568 #endif /* _KERNEL */
7570 /* Use B_TRUE to ensure *all* buffers are evicted */
7571 arc_flush(NULL, B_TRUE);
7573 if (arc_ksp != NULL) {
7574 kstat_delete(arc_ksp);
7578 taskq_wait(arc_prune_taskq);
7579 taskq_destroy(arc_prune_taskq);
7581 mutex_enter(&arc_prune_mtx);
7582 while ((p = list_head(&arc_prune_list)) != NULL) {
7583 list_remove(&arc_prune_list, p);
7584 zfs_refcount_remove(&p->p_refcnt, &arc_prune_list);
7585 zfs_refcount_destroy(&p->p_refcnt);
7586 kmem_free(p, sizeof (*p));
7588 mutex_exit(&arc_prune_mtx);
7590 list_destroy(&arc_prune_list);
7591 mutex_destroy(&arc_prune_mtx);
7593 (void) zthr_cancel(arc_evict_zthr);
7594 (void) zthr_cancel(arc_reap_zthr);
7596 mutex_destroy(&arc_evict_lock);
7597 list_destroy(&arc_evict_waiters);
7600 * Free any buffers that were tagged for destruction. This needs
7601 * to occur before arc_state_fini() runs and destroys the aggsum
7602 * values which are updated when freeing scatter ABDs.
7604 l2arc_do_free_on_write();
7607 * buf_fini() must proceed arc_state_fini() because buf_fin() may
7608 * trigger the release of kmem magazines, which can callback to
7609 * arc_space_return() which accesses aggsums freed in act_state_fini().
7615 * We destroy the zthrs after all the ARC state has been
7616 * torn down to avoid the case of them receiving any
7617 * wakeup() signals after they are destroyed.
7619 zthr_destroy(arc_evict_zthr);
7620 zthr_destroy(arc_reap_zthr);
7622 ASSERT0(arc_loaned_bytes);
7628 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7629 * It uses dedicated storage devices to hold cached data, which are populated
7630 * using large infrequent writes. The main role of this cache is to boost
7631 * the performance of random read workloads. The intended L2ARC devices
7632 * include short-stroked disks, solid state disks, and other media with
7633 * substantially faster read latency than disk.
7635 * +-----------------------+
7637 * +-----------------------+
7640 * l2arc_feed_thread() arc_read()
7644 * +---------------+ |
7646 * +---------------+ |
7651 * +-------+ +-------+
7653 * | cache | | cache |
7654 * +-------+ +-------+
7655 * +=========+ .-----.
7656 * : L2ARC : |-_____-|
7657 * : devices : | Disks |
7658 * +=========+ `-_____-'
7660 * Read requests are satisfied from the following sources, in order:
7663 * 2) vdev cache of L2ARC devices
7665 * 4) vdev cache of disks
7668 * Some L2ARC device types exhibit extremely slow write performance.
7669 * To accommodate for this there are some significant differences between
7670 * the L2ARC and traditional cache design:
7672 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7673 * the ARC behave as usual, freeing buffers and placing headers on ghost
7674 * lists. The ARC does not send buffers to the L2ARC during eviction as
7675 * this would add inflated write latencies for all ARC memory pressure.
7677 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7678 * It does this by periodically scanning buffers from the eviction-end of
7679 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7680 * not already there. It scans until a headroom of buffers is satisfied,
7681 * which itself is a buffer for ARC eviction. If a compressible buffer is
7682 * found during scanning and selected for writing to an L2ARC device, we
7683 * temporarily boost scanning headroom during the next scan cycle to make
7684 * sure we adapt to compression effects (which might significantly reduce
7685 * the data volume we write to L2ARC). The thread that does this is
7686 * l2arc_feed_thread(), illustrated below; example sizes are included to
7687 * provide a better sense of ratio than this diagram:
7690 * +---------------------+----------+
7691 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7692 * +---------------------+----------+ | o L2ARC eligible
7693 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7694 * +---------------------+----------+ |
7695 * 15.9 Gbytes ^ 32 Mbytes |
7697 * l2arc_feed_thread()
7699 * l2arc write hand <--[oooo]--'
7703 * +==============================+
7704 * L2ARC dev |####|#|###|###| |####| ... |
7705 * +==============================+
7708 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7709 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7710 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7711 * safe to say that this is an uncommon case, since buffers at the end of
7712 * the ARC lists have moved there due to inactivity.
7714 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7715 * then the L2ARC simply misses copying some buffers. This serves as a
7716 * pressure valve to prevent heavy read workloads from both stalling the ARC
7717 * with waits and clogging the L2ARC with writes. This also helps prevent
7718 * the potential for the L2ARC to churn if it attempts to cache content too
7719 * quickly, such as during backups of the entire pool.
7721 * 5. After system boot and before the ARC has filled main memory, there are
7722 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7723 * lists can remain mostly static. Instead of searching from tail of these
7724 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7725 * for eligible buffers, greatly increasing its chance of finding them.
7727 * The L2ARC device write speed is also boosted during this time so that
7728 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7729 * there are no L2ARC reads, and no fear of degrading read performance
7730 * through increased writes.
7732 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7733 * the vdev queue can aggregate them into larger and fewer writes. Each
7734 * device is written to in a rotor fashion, sweeping writes through
7735 * available space then repeating.
7737 * 7. The L2ARC does not store dirty content. It never needs to flush
7738 * write buffers back to disk based storage.
7740 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7741 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7743 * The performance of the L2ARC can be tweaked by a number of tunables, which
7744 * may be necessary for different workloads:
7746 * l2arc_write_max max write bytes per interval
7747 * l2arc_write_boost extra write bytes during device warmup
7748 * l2arc_noprefetch skip caching prefetched buffers
7749 * l2arc_headroom number of max device writes to precache
7750 * l2arc_headroom_boost when we find compressed buffers during ARC
7751 * scanning, we multiply headroom by this
7752 * percentage factor for the next scan cycle,
7753 * since more compressed buffers are likely to
7755 * l2arc_feed_secs seconds between L2ARC writing
7757 * Tunables may be removed or added as future performance improvements are
7758 * integrated, and also may become zpool properties.
7760 * There are three key functions that control how the L2ARC warms up:
7762 * l2arc_write_eligible() check if a buffer is eligible to cache
7763 * l2arc_write_size() calculate how much to write
7764 * l2arc_write_interval() calculate sleep delay between writes
7766 * These three functions determine what to write, how much, and how quickly
7769 * L2ARC persistence:
7771 * When writing buffers to L2ARC, we periodically add some metadata to
7772 * make sure we can pick them up after reboot, thus dramatically reducing
7773 * the impact that any downtime has on the performance of storage systems
7774 * with large caches.
7776 * The implementation works fairly simply by integrating the following two
7779 * *) When writing to the L2ARC, we occasionally write a "l2arc log block",
7780 * which is an additional piece of metadata which describes what's been
7781 * written. This allows us to rebuild the arc_buf_hdr_t structures of the
7782 * main ARC buffers. There are 2 linked-lists of log blocks headed by
7783 * dh_start_lbps[2]. We alternate which chain we append to, so they are
7784 * time-wise and offset-wise interleaved, but that is an optimization rather
7785 * than for correctness. The log block also includes a pointer to the
7786 * previous block in its chain.
7788 * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device
7789 * for our header bookkeeping purposes. This contains a device header,
7790 * which contains our top-level reference structures. We update it each
7791 * time we write a new log block, so that we're able to locate it in the
7792 * L2ARC device. If this write results in an inconsistent device header
7793 * (e.g. due to power failure), we detect this by verifying the header's
7794 * checksum and simply fail to reconstruct the L2ARC after reboot.
7796 * Implementation diagram:
7798 * +=== L2ARC device (not to scale) ======================================+
7799 * | ___two newest log block pointers__.__________ |
7800 * | / \dh_start_lbps[1] |
7801 * | / \ \dh_start_lbps[0]|
7803 * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---|
7804 * || hdr| ^ /^ /^ / / |
7805 * |+------+ ...--\-------/ \-----/--\------/ / |
7806 * | \--------------/ \--------------/ |
7807 * +======================================================================+
7809 * As can be seen on the diagram, rather than using a simple linked list,
7810 * we use a pair of linked lists with alternating elements. This is a
7811 * performance enhancement due to the fact that we only find out the
7812 * address of the next log block access once the current block has been
7813 * completely read in. Obviously, this hurts performance, because we'd be
7814 * keeping the device's I/O queue at only a 1 operation deep, thus
7815 * incurring a large amount of I/O round-trip latency. Having two lists
7816 * allows us to fetch two log blocks ahead of where we are currently
7817 * rebuilding L2ARC buffers.
7819 * On-device data structures:
7821 * L2ARC device header: l2arc_dev_hdr_phys_t
7822 * L2ARC log block: l2arc_log_blk_phys_t
7824 * L2ARC reconstruction:
7826 * When writing data, we simply write in the standard rotary fashion,
7827 * evicting buffers as we go and simply writing new data over them (writing
7828 * a new log block every now and then). This obviously means that once we
7829 * loop around the end of the device, we will start cutting into an already
7830 * committed log block (and its referenced data buffers), like so:
7832 * current write head__ __old tail
7835 * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |-->
7836 * ^ ^^^^^^^^^___________________________________
7838 * <<nextwrite>> may overwrite this blk and/or its bufs --'
7840 * When importing the pool, we detect this situation and use it to stop
7841 * our scanning process (see l2arc_rebuild).
7843 * There is one significant caveat to consider when rebuilding ARC contents
7844 * from an L2ARC device: what about invalidated buffers? Given the above
7845 * construction, we cannot update blocks which we've already written to amend
7846 * them to remove buffers which were invalidated. Thus, during reconstruction,
7847 * we might be populating the cache with buffers for data that's not on the
7848 * main pool anymore, or may have been overwritten!
7850 * As it turns out, this isn't a problem. Every arc_read request includes
7851 * both the DVA and, crucially, the birth TXG of the BP the caller is
7852 * looking for. So even if the cache were populated by completely rotten
7853 * blocks for data that had been long deleted and/or overwritten, we'll
7854 * never actually return bad data from the cache, since the DVA with the
7855 * birth TXG uniquely identify a block in space and time - once created,
7856 * a block is immutable on disk. The worst thing we have done is wasted
7857 * some time and memory at l2arc rebuild to reconstruct outdated ARC
7858 * entries that will get dropped from the l2arc as it is being updated
7861 * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write
7862 * hand are not restored. This is done by saving the offset (in bytes)
7863 * l2arc_evict() has evicted to in the L2ARC device header and taking it
7864 * into account when restoring buffers.
7868 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
7871 * A buffer is *not* eligible for the L2ARC if it:
7872 * 1. belongs to a different spa.
7873 * 2. is already cached on the L2ARC.
7874 * 3. has an I/O in progress (it may be an incomplete read).
7875 * 4. is flagged not eligible (zfs property).
7877 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
7878 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
7885 l2arc_write_size(l2arc_dev_t *dev)
7887 uint64_t size, dev_size, tsize;
7890 * Make sure our globals have meaningful values in case the user
7893 size = l2arc_write_max;
7895 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
7896 "be greater than zero, resetting it to the default (%d)",
7898 size = l2arc_write_max = L2ARC_WRITE_SIZE;
7901 if (arc_warm == B_FALSE)
7902 size += l2arc_write_boost;
7905 * Make sure the write size does not exceed the size of the cache
7906 * device. This is important in l2arc_evict(), otherwise infinite
7907 * iteration can occur.
7909 dev_size = dev->l2ad_end - dev->l2ad_start;
7910 tsize = size + l2arc_log_blk_overhead(size, dev);
7911 if (dev->l2ad_vdev->vdev_has_trim && l2arc_trim_ahead > 0)
7912 tsize += MAX(64 * 1024 * 1024,
7913 (tsize * l2arc_trim_ahead) / 100);
7915 if (tsize >= dev_size) {
7916 cmn_err(CE_NOTE, "l2arc_write_max or l2arc_write_boost "
7917 "plus the overhead of log blocks (persistent L2ARC, "
7918 "%llu bytes) exceeds the size of the cache device "
7919 "(guid %llu), resetting them to the default (%d)",
7920 l2arc_log_blk_overhead(size, dev),
7921 dev->l2ad_vdev->vdev_guid, L2ARC_WRITE_SIZE);
7922 size = l2arc_write_max = l2arc_write_boost = L2ARC_WRITE_SIZE;
7924 if (arc_warm == B_FALSE)
7925 size += l2arc_write_boost;
7933 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
7935 clock_t interval, next, now;
7938 * If the ARC lists are busy, increase our write rate; if the
7939 * lists are stale, idle back. This is achieved by checking
7940 * how much we previously wrote - if it was more than half of
7941 * what we wanted, schedule the next write much sooner.
7943 if (l2arc_feed_again && wrote > (wanted / 2))
7944 interval = (hz * l2arc_feed_min_ms) / 1000;
7946 interval = hz * l2arc_feed_secs;
7948 now = ddi_get_lbolt();
7949 next = MAX(now, MIN(now + interval, began + interval));
7955 * Cycle through L2ARC devices. This is how L2ARC load balances.
7956 * If a device is returned, this also returns holding the spa config lock.
7958 static l2arc_dev_t *
7959 l2arc_dev_get_next(void)
7961 l2arc_dev_t *first, *next = NULL;
7964 * Lock out the removal of spas (spa_namespace_lock), then removal
7965 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7966 * both locks will be dropped and a spa config lock held instead.
7968 mutex_enter(&spa_namespace_lock);
7969 mutex_enter(&l2arc_dev_mtx);
7971 /* if there are no vdevs, there is nothing to do */
7972 if (l2arc_ndev == 0)
7976 next = l2arc_dev_last;
7978 /* loop around the list looking for a non-faulted vdev */
7980 next = list_head(l2arc_dev_list);
7982 next = list_next(l2arc_dev_list, next);
7984 next = list_head(l2arc_dev_list);
7987 /* if we have come back to the start, bail out */
7990 else if (next == first)
7993 } while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
7994 next->l2ad_trim_all);
7996 /* if we were unable to find any usable vdevs, return NULL */
7997 if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
7998 next->l2ad_trim_all)
8001 l2arc_dev_last = next;
8004 mutex_exit(&l2arc_dev_mtx);
8007 * Grab the config lock to prevent the 'next' device from being
8008 * removed while we are writing to it.
8011 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
8012 mutex_exit(&spa_namespace_lock);
8018 * Free buffers that were tagged for destruction.
8021 l2arc_do_free_on_write(void)
8024 l2arc_data_free_t *df, *df_prev;
8026 mutex_enter(&l2arc_free_on_write_mtx);
8027 buflist = l2arc_free_on_write;
8029 for (df = list_tail(buflist); df; df = df_prev) {
8030 df_prev = list_prev(buflist, df);
8031 ASSERT3P(df->l2df_abd, !=, NULL);
8032 abd_free(df->l2df_abd);
8033 list_remove(buflist, df);
8034 kmem_free(df, sizeof (l2arc_data_free_t));
8037 mutex_exit(&l2arc_free_on_write_mtx);
8041 * A write to a cache device has completed. Update all headers to allow
8042 * reads from these buffers to begin.
8045 l2arc_write_done(zio_t *zio)
8047 l2arc_write_callback_t *cb;
8048 l2arc_lb_abd_buf_t *abd_buf;
8049 l2arc_lb_ptr_buf_t *lb_ptr_buf;
8051 l2arc_dev_hdr_phys_t *l2dhdr;
8053 arc_buf_hdr_t *head, *hdr, *hdr_prev;
8054 kmutex_t *hash_lock;
8055 int64_t bytes_dropped = 0;
8057 cb = zio->io_private;
8058 ASSERT3P(cb, !=, NULL);
8059 dev = cb->l2wcb_dev;
8060 l2dhdr = dev->l2ad_dev_hdr;
8061 ASSERT3P(dev, !=, NULL);
8062 head = cb->l2wcb_head;
8063 ASSERT3P(head, !=, NULL);
8064 buflist = &dev->l2ad_buflist;
8065 ASSERT3P(buflist, !=, NULL);
8066 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
8067 l2arc_write_callback_t *, cb);
8069 if (zio->io_error != 0)
8070 ARCSTAT_BUMP(arcstat_l2_writes_error);
8073 * All writes completed, or an error was hit.
8076 mutex_enter(&dev->l2ad_mtx);
8077 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
8078 hdr_prev = list_prev(buflist, hdr);
8080 hash_lock = HDR_LOCK(hdr);
8083 * We cannot use mutex_enter or else we can deadlock
8084 * with l2arc_write_buffers (due to swapping the order
8085 * the hash lock and l2ad_mtx are taken).
8087 if (!mutex_tryenter(hash_lock)) {
8089 * Missed the hash lock. We must retry so we
8090 * don't leave the ARC_FLAG_L2_WRITING bit set.
8092 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
8095 * We don't want to rescan the headers we've
8096 * already marked as having been written out, so
8097 * we reinsert the head node so we can pick up
8098 * where we left off.
8100 list_remove(buflist, head);
8101 list_insert_after(buflist, hdr, head);
8103 mutex_exit(&dev->l2ad_mtx);
8106 * We wait for the hash lock to become available
8107 * to try and prevent busy waiting, and increase
8108 * the chance we'll be able to acquire the lock
8109 * the next time around.
8111 mutex_enter(hash_lock);
8112 mutex_exit(hash_lock);
8117 * We could not have been moved into the arc_l2c_only
8118 * state while in-flight due to our ARC_FLAG_L2_WRITING
8119 * bit being set. Let's just ensure that's being enforced.
8121 ASSERT(HDR_HAS_L1HDR(hdr));
8124 * Skipped - drop L2ARC entry and mark the header as no
8125 * longer L2 eligibile.
8127 if (zio->io_error != 0) {
8129 * Error - drop L2ARC entry.
8131 list_remove(buflist, hdr);
8132 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
8134 uint64_t psize = HDR_GET_PSIZE(hdr);
8135 ARCSTAT_INCR(arcstat_l2_psize, -psize);
8136 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
8139 vdev_psize_to_asize(dev->l2ad_vdev, psize);
8140 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
8141 arc_hdr_size(hdr), hdr);
8145 * Allow ARC to begin reads and ghost list evictions to
8148 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
8150 mutex_exit(hash_lock);
8154 * Free the allocated abd buffers for writing the log blocks.
8155 * If the zio failed reclaim the allocated space and remove the
8156 * pointers to these log blocks from the log block pointer list
8157 * of the L2ARC device.
8159 while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) {
8160 abd_free(abd_buf->abd);
8161 zio_buf_free(abd_buf, sizeof (*abd_buf));
8162 if (zio->io_error != 0) {
8163 lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list);
8165 * L2BLK_GET_PSIZE returns aligned size for log
8169 L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop);
8170 bytes_dropped += asize;
8171 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8172 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8173 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8175 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8176 kmem_free(lb_ptr_buf->lb_ptr,
8177 sizeof (l2arc_log_blkptr_t));
8178 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8181 list_destroy(&cb->l2wcb_abd_list);
8183 if (zio->io_error != 0) {
8185 * Restore the lbps array in the header to its previous state.
8186 * If the list of log block pointers is empty, zero out the
8187 * log block pointers in the device header.
8189 lb_ptr_buf = list_head(&dev->l2ad_lbptr_list);
8190 for (int i = 0; i < 2; i++) {
8191 if (lb_ptr_buf == NULL) {
8193 * If the list is empty zero out the device
8194 * header. Otherwise zero out the second log
8195 * block pointer in the header.
8198 bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
8200 bzero(&l2dhdr->dh_start_lbps[i],
8201 sizeof (l2arc_log_blkptr_t));
8205 bcopy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[i],
8206 sizeof (l2arc_log_blkptr_t));
8207 lb_ptr_buf = list_next(&dev->l2ad_lbptr_list,
8212 atomic_inc_64(&l2arc_writes_done);
8213 list_remove(buflist, head);
8214 ASSERT(!HDR_HAS_L1HDR(head));
8215 kmem_cache_free(hdr_l2only_cache, head);
8216 mutex_exit(&dev->l2ad_mtx);
8218 ASSERT(dev->l2ad_vdev != NULL);
8219 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
8221 l2arc_do_free_on_write();
8223 kmem_free(cb, sizeof (l2arc_write_callback_t));
8227 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
8230 spa_t *spa = zio->io_spa;
8231 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
8232 blkptr_t *bp = zio->io_bp;
8233 uint8_t salt[ZIO_DATA_SALT_LEN];
8234 uint8_t iv[ZIO_DATA_IV_LEN];
8235 uint8_t mac[ZIO_DATA_MAC_LEN];
8236 boolean_t no_crypt = B_FALSE;
8239 * ZIL data is never be written to the L2ARC, so we don't need
8240 * special handling for its unique MAC storage.
8242 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
8243 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
8244 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8247 * If the data was encrypted, decrypt it now. Note that
8248 * we must check the bp here and not the hdr, since the
8249 * hdr does not have its encryption parameters updated
8250 * until arc_read_done().
8252 if (BP_IS_ENCRYPTED(bp)) {
8253 abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8256 zio_crypt_decode_params_bp(bp, salt, iv);
8257 zio_crypt_decode_mac_bp(bp, mac);
8259 ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
8260 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
8261 salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
8262 hdr->b_l1hdr.b_pabd, &no_crypt);
8264 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8269 * If we actually performed decryption, replace b_pabd
8270 * with the decrypted data. Otherwise we can just throw
8271 * our decryption buffer away.
8274 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8275 arc_hdr_size(hdr), hdr);
8276 hdr->b_l1hdr.b_pabd = eabd;
8279 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8284 * If the L2ARC block was compressed, but ARC compression
8285 * is disabled we decompress the data into a new buffer and
8286 * replace the existing data.
8288 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8289 !HDR_COMPRESSION_ENABLED(hdr)) {
8290 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8292 void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
8294 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
8295 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
8296 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
8298 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8299 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
8303 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8304 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8305 arc_hdr_size(hdr), hdr);
8306 hdr->b_l1hdr.b_pabd = cabd;
8308 zio->io_size = HDR_GET_LSIZE(hdr);
8319 * A read to a cache device completed. Validate buffer contents before
8320 * handing over to the regular ARC routines.
8323 l2arc_read_done(zio_t *zio)
8326 l2arc_read_callback_t *cb = zio->io_private;
8328 kmutex_t *hash_lock;
8329 boolean_t valid_cksum;
8330 boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
8331 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
8333 ASSERT3P(zio->io_vd, !=, NULL);
8334 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
8336 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
8338 ASSERT3P(cb, !=, NULL);
8339 hdr = cb->l2rcb_hdr;
8340 ASSERT3P(hdr, !=, NULL);
8342 hash_lock = HDR_LOCK(hdr);
8343 mutex_enter(hash_lock);
8344 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
8347 * If the data was read into a temporary buffer,
8348 * move it and free the buffer.
8350 if (cb->l2rcb_abd != NULL) {
8351 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
8352 if (zio->io_error == 0) {
8354 abd_copy(hdr->b_crypt_hdr.b_rabd,
8355 cb->l2rcb_abd, arc_hdr_size(hdr));
8357 abd_copy(hdr->b_l1hdr.b_pabd,
8358 cb->l2rcb_abd, arc_hdr_size(hdr));
8363 * The following must be done regardless of whether
8364 * there was an error:
8365 * - free the temporary buffer
8366 * - point zio to the real ARC buffer
8367 * - set zio size accordingly
8368 * These are required because zio is either re-used for
8369 * an I/O of the block in the case of the error
8370 * or the zio is passed to arc_read_done() and it
8373 abd_free(cb->l2rcb_abd);
8374 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
8377 ASSERT(HDR_HAS_RABD(hdr));
8378 zio->io_abd = zio->io_orig_abd =
8379 hdr->b_crypt_hdr.b_rabd;
8381 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8382 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8386 ASSERT3P(zio->io_abd, !=, NULL);
8389 * Check this survived the L2ARC journey.
8391 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8392 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8393 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
8394 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
8395 zio->io_prop.zp_complevel = hdr->b_complevel;
8397 valid_cksum = arc_cksum_is_equal(hdr, zio);
8400 * b_rabd will always match the data as it exists on disk if it is
8401 * being used. Therefore if we are reading into b_rabd we do not
8402 * attempt to untransform the data.
8404 if (valid_cksum && !using_rdata)
8405 tfm_error = l2arc_untransform(zio, cb);
8407 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8408 !HDR_L2_EVICTED(hdr)) {
8409 mutex_exit(hash_lock);
8410 zio->io_private = hdr;
8414 * Buffer didn't survive caching. Increment stats and
8415 * reissue to the original storage device.
8417 if (zio->io_error != 0) {
8418 ARCSTAT_BUMP(arcstat_l2_io_error);
8420 zio->io_error = SET_ERROR(EIO);
8422 if (!valid_cksum || tfm_error != 0)
8423 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8426 * If there's no waiter, issue an async i/o to the primary
8427 * storage now. If there *is* a waiter, the caller must
8428 * issue the i/o in a context where it's OK to block.
8430 if (zio->io_waiter == NULL) {
8431 zio_t *pio = zio_unique_parent(zio);
8432 void *abd = (using_rdata) ?
8433 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8435 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8437 zio = zio_read(pio, zio->io_spa, zio->io_bp,
8438 abd, zio->io_size, arc_read_done,
8439 hdr, zio->io_priority, cb->l2rcb_flags,
8443 * Original ZIO will be freed, so we need to update
8444 * ARC header with the new ZIO pointer to be used
8445 * by zio_change_priority() in arc_read().
8447 for (struct arc_callback *acb = hdr->b_l1hdr.b_acb;
8448 acb != NULL; acb = acb->acb_next)
8449 acb->acb_zio_head = zio;
8451 mutex_exit(hash_lock);
8454 mutex_exit(hash_lock);
8458 kmem_free(cb, sizeof (l2arc_read_callback_t));
8462 * This is the list priority from which the L2ARC will search for pages to
8463 * cache. This is used within loops (0..3) to cycle through lists in the
8464 * desired order. This order can have a significant effect on cache
8467 * Currently the metadata lists are hit first, MFU then MRU, followed by
8468 * the data lists. This function returns a locked list, and also returns
8471 static multilist_sublist_t *
8472 l2arc_sublist_lock(int list_num)
8474 multilist_t *ml = NULL;
8477 ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
8481 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
8484 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
8487 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
8490 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
8497 * Return a randomly-selected sublist. This is acceptable
8498 * because the caller feeds only a little bit of data for each
8499 * call (8MB). Subsequent calls will result in different
8500 * sublists being selected.
8502 idx = multilist_get_random_index(ml);
8503 return (multilist_sublist_lock(ml, idx));
8507 * Calculates the maximum overhead of L2ARC metadata log blocks for a given
8508 * L2ARC write size. l2arc_evict and l2arc_write_size need to include this
8509 * overhead in processing to make sure there is enough headroom available
8510 * when writing buffers.
8512 static inline uint64_t
8513 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev)
8515 if (dev->l2ad_log_entries == 0) {
8518 uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT;
8520 uint64_t log_blocks = (log_entries +
8521 dev->l2ad_log_entries - 1) /
8522 dev->l2ad_log_entries;
8524 return (vdev_psize_to_asize(dev->l2ad_vdev,
8525 sizeof (l2arc_log_blk_phys_t)) * log_blocks);
8530 * Evict buffers from the device write hand to the distance specified in
8531 * bytes. This distance may span populated buffers, it may span nothing.
8532 * This is clearing a region on the L2ARC device ready for writing.
8533 * If the 'all' boolean is set, every buffer is evicted.
8536 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
8539 arc_buf_hdr_t *hdr, *hdr_prev;
8540 kmutex_t *hash_lock;
8542 l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev;
8543 vdev_t *vd = dev->l2ad_vdev;
8546 buflist = &dev->l2ad_buflist;
8549 * We need to add in the worst case scenario of log block overhead.
8551 distance += l2arc_log_blk_overhead(distance, dev);
8552 if (vd->vdev_has_trim && l2arc_trim_ahead > 0) {
8554 * Trim ahead of the write size 64MB or (l2arc_trim_ahead/100)
8555 * times the write size, whichever is greater.
8557 distance += MAX(64 * 1024 * 1024,
8558 (distance * l2arc_trim_ahead) / 100);
8563 if (dev->l2ad_hand >= (dev->l2ad_end - distance)) {
8565 * When there is no space to accommodate upcoming writes,
8566 * evict to the end. Then bump the write and evict hands
8567 * to the start and iterate. This iteration does not
8568 * happen indefinitely as we make sure in
8569 * l2arc_write_size() that when the write hand is reset,
8570 * the write size does not exceed the end of the device.
8573 taddr = dev->l2ad_end;
8575 taddr = dev->l2ad_hand + distance;
8577 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
8578 uint64_t, taddr, boolean_t, all);
8582 * This check has to be placed after deciding whether to
8585 if (dev->l2ad_first) {
8587 * This is the first sweep through the device. There is
8588 * nothing to evict. We have already trimmmed the
8594 * Trim the space to be evicted.
8596 if (vd->vdev_has_trim && dev->l2ad_evict < taddr &&
8597 l2arc_trim_ahead > 0) {
8599 * We have to drop the spa_config lock because
8600 * vdev_trim_range() will acquire it.
8601 * l2ad_evict already accounts for the label
8602 * size. To prevent vdev_trim_ranges() from
8603 * adding it again, we subtract it from
8606 spa_config_exit(dev->l2ad_spa, SCL_L2ARC, dev);
8607 vdev_trim_simple(vd,
8608 dev->l2ad_evict - VDEV_LABEL_START_SIZE,
8609 taddr - dev->l2ad_evict);
8610 spa_config_enter(dev->l2ad_spa, SCL_L2ARC, dev,
8615 * When rebuilding L2ARC we retrieve the evict hand
8616 * from the header of the device. Of note, l2arc_evict()
8617 * does not actually delete buffers from the cache
8618 * device, but trimming may do so depending on the
8619 * hardware implementation. Thus keeping track of the
8620 * evict hand is useful.
8622 dev->l2ad_evict = MAX(dev->l2ad_evict, taddr);
8627 mutex_enter(&dev->l2ad_mtx);
8629 * We have to account for evicted log blocks. Run vdev_space_update()
8630 * on log blocks whose offset (in bytes) is before the evicted offset
8631 * (in bytes) by searching in the list of pointers to log blocks
8632 * present in the L2ARC device.
8634 for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf;
8635 lb_ptr_buf = lb_ptr_buf_prev) {
8637 lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf);
8639 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
8640 uint64_t asize = L2BLK_GET_PSIZE(
8641 (lb_ptr_buf->lb_ptr)->lbp_prop);
8644 * We don't worry about log blocks left behind (ie
8645 * lbp_payload_start < l2ad_hand) because l2arc_write_buffers()
8646 * will never write more than l2arc_evict() evicts.
8648 if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) {
8651 vdev_space_update(vd, -asize, 0, 0);
8652 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8653 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8654 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8656 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8657 list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf);
8658 kmem_free(lb_ptr_buf->lb_ptr,
8659 sizeof (l2arc_log_blkptr_t));
8660 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8664 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
8665 hdr_prev = list_prev(buflist, hdr);
8667 ASSERT(!HDR_EMPTY(hdr));
8668 hash_lock = HDR_LOCK(hdr);
8671 * We cannot use mutex_enter or else we can deadlock
8672 * with l2arc_write_buffers (due to swapping the order
8673 * the hash lock and l2ad_mtx are taken).
8675 if (!mutex_tryenter(hash_lock)) {
8677 * Missed the hash lock. Retry.
8679 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
8680 mutex_exit(&dev->l2ad_mtx);
8681 mutex_enter(hash_lock);
8682 mutex_exit(hash_lock);
8687 * A header can't be on this list if it doesn't have L2 header.
8689 ASSERT(HDR_HAS_L2HDR(hdr));
8691 /* Ensure this header has finished being written. */
8692 ASSERT(!HDR_L2_WRITING(hdr));
8693 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
8695 if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict ||
8696 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
8698 * We've evicted to the target address,
8699 * or the end of the device.
8701 mutex_exit(hash_lock);
8705 if (!HDR_HAS_L1HDR(hdr)) {
8706 ASSERT(!HDR_L2_READING(hdr));
8708 * This doesn't exist in the ARC. Destroy.
8709 * arc_hdr_destroy() will call list_remove()
8710 * and decrement arcstat_l2_lsize.
8712 arc_change_state(arc_anon, hdr, hash_lock);
8713 arc_hdr_destroy(hdr);
8715 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
8716 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
8718 * Invalidate issued or about to be issued
8719 * reads, since we may be about to write
8720 * over this location.
8722 if (HDR_L2_READING(hdr)) {
8723 ARCSTAT_BUMP(arcstat_l2_evict_reading);
8724 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
8727 arc_hdr_l2hdr_destroy(hdr);
8729 mutex_exit(hash_lock);
8731 mutex_exit(&dev->l2ad_mtx);
8735 * We need to check if we evict all buffers, otherwise we may iterate
8738 if (!all && rerun) {
8740 * Bump device hand to the device start if it is approaching the
8741 * end. l2arc_evict() has already evicted ahead for this case.
8743 dev->l2ad_hand = dev->l2ad_start;
8744 dev->l2ad_evict = dev->l2ad_start;
8745 dev->l2ad_first = B_FALSE;
8749 ASSERT3U(dev->l2ad_hand + distance, <, dev->l2ad_end);
8750 if (!dev->l2ad_first)
8751 ASSERT3U(dev->l2ad_hand, <, dev->l2ad_evict);
8755 * Handle any abd transforms that might be required for writing to the L2ARC.
8756 * If successful, this function will always return an abd with the data
8757 * transformed as it is on disk in a new abd of asize bytes.
8760 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
8765 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
8766 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
8767 uint64_t psize = HDR_GET_PSIZE(hdr);
8768 uint64_t size = arc_hdr_size(hdr);
8769 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
8770 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
8771 dsl_crypto_key_t *dck = NULL;
8772 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
8773 boolean_t no_crypt = B_FALSE;
8775 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8776 !HDR_COMPRESSION_ENABLED(hdr)) ||
8777 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
8778 ASSERT3U(psize, <=, asize);
8781 * If this data simply needs its own buffer, we simply allocate it
8782 * and copy the data. This may be done to eliminate a dependency on a
8783 * shared buffer or to reallocate the buffer to match asize.
8785 if (HDR_HAS_RABD(hdr) && asize != psize) {
8786 ASSERT3U(asize, >=, psize);
8787 to_write = abd_alloc_for_io(asize, ismd);
8788 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
8790 abd_zero_off(to_write, psize, asize - psize);
8794 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
8795 !HDR_ENCRYPTED(hdr)) {
8796 ASSERT3U(size, ==, psize);
8797 to_write = abd_alloc_for_io(asize, ismd);
8798 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
8800 abd_zero_off(to_write, size, asize - size);
8804 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
8805 cabd = abd_alloc_for_io(asize, ismd);
8806 tmp = abd_borrow_buf(cabd, asize);
8808 psize = zio_compress_data(compress, to_write, tmp, size,
8811 if (psize >= size) {
8812 abd_return_buf(cabd, tmp, asize);
8813 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
8815 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
8817 abd_zero_off(to_write, size, asize - size);
8820 ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
8822 bzero((char *)tmp + psize, asize - psize);
8823 psize = HDR_GET_PSIZE(hdr);
8824 abd_return_buf_copy(cabd, tmp, asize);
8829 if (HDR_ENCRYPTED(hdr)) {
8830 eabd = abd_alloc_for_io(asize, ismd);
8833 * If the dataset was disowned before the buffer
8834 * made it to this point, the key to re-encrypt
8835 * it won't be available. In this case we simply
8836 * won't write the buffer to the L2ARC.
8838 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
8843 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
8844 hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
8845 hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
8851 abd_copy(eabd, to_write, psize);
8854 abd_zero_off(eabd, psize, asize - psize);
8856 /* assert that the MAC we got here matches the one we saved */
8857 ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
8858 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8860 if (to_write == cabd)
8867 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
8868 *abd_out = to_write;
8873 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8884 l2arc_blk_fetch_done(zio_t *zio)
8886 l2arc_read_callback_t *cb;
8888 cb = zio->io_private;
8889 if (cb->l2rcb_abd != NULL)
8890 abd_put(cb->l2rcb_abd);
8891 kmem_free(cb, sizeof (l2arc_read_callback_t));
8895 * Find and write ARC buffers to the L2ARC device.
8897 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
8898 * for reading until they have completed writing.
8899 * The headroom_boost is an in-out parameter used to maintain headroom boost
8900 * state between calls to this function.
8902 * Returns the number of bytes actually written (which may be smaller than
8903 * the delta by which the device hand has changed due to alignment and the
8904 * writing of log blocks).
8907 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
8909 arc_buf_hdr_t *hdr, *hdr_prev, *head;
8910 uint64_t write_asize, write_psize, write_lsize, headroom;
8912 l2arc_write_callback_t *cb = NULL;
8914 uint64_t guid = spa_load_guid(spa);
8916 ASSERT3P(dev->l2ad_vdev, !=, NULL);
8919 write_lsize = write_asize = write_psize = 0;
8921 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
8922 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
8925 * Copy buffers for L2ARC writing.
8927 for (int try = 0; try < L2ARC_FEED_TYPES; try++) {
8929 * If try == 1 or 3, we cache MRU metadata and data
8932 if (l2arc_mfuonly) {
8933 if (try == 1 || try == 3)
8937 multilist_sublist_t *mls = l2arc_sublist_lock(try);
8938 uint64_t passed_sz = 0;
8940 VERIFY3P(mls, !=, NULL);
8943 * L2ARC fast warmup.
8945 * Until the ARC is warm and starts to evict, read from the
8946 * head of the ARC lists rather than the tail.
8948 if (arc_warm == B_FALSE)
8949 hdr = multilist_sublist_head(mls);
8951 hdr = multilist_sublist_tail(mls);
8953 headroom = target_sz * l2arc_headroom;
8954 if (zfs_compressed_arc_enabled)
8955 headroom = (headroom * l2arc_headroom_boost) / 100;
8957 for (; hdr; hdr = hdr_prev) {
8958 kmutex_t *hash_lock;
8959 abd_t *to_write = NULL;
8961 if (arc_warm == B_FALSE)
8962 hdr_prev = multilist_sublist_next(mls, hdr);
8964 hdr_prev = multilist_sublist_prev(mls, hdr);
8966 hash_lock = HDR_LOCK(hdr);
8967 if (!mutex_tryenter(hash_lock)) {
8969 * Skip this buffer rather than waiting.
8974 passed_sz += HDR_GET_LSIZE(hdr);
8975 if (l2arc_headroom != 0 && passed_sz > headroom) {
8979 mutex_exit(hash_lock);
8983 if (!l2arc_write_eligible(guid, hdr)) {
8984 mutex_exit(hash_lock);
8989 * We rely on the L1 portion of the header below, so
8990 * it's invalid for this header to have been evicted out
8991 * of the ghost cache, prior to being written out. The
8992 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8994 ASSERT(HDR_HAS_L1HDR(hdr));
8996 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8997 ASSERT3U(arc_hdr_size(hdr), >, 0);
8998 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
9000 uint64_t psize = HDR_GET_PSIZE(hdr);
9001 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
9004 if ((write_asize + asize) > target_sz) {
9006 mutex_exit(hash_lock);
9011 * We rely on the L1 portion of the header below, so
9012 * it's invalid for this header to have been evicted out
9013 * of the ghost cache, prior to being written out. The
9014 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
9016 arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
9017 ASSERT(HDR_HAS_L1HDR(hdr));
9019 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
9020 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
9022 ASSERT3U(arc_hdr_size(hdr), >, 0);
9025 * If this header has b_rabd, we can use this since it
9026 * must always match the data exactly as it exists on
9027 * disk. Otherwise, the L2ARC can normally use the
9028 * hdr's data, but if we're sharing data between the
9029 * hdr and one of its bufs, L2ARC needs its own copy of
9030 * the data so that the ZIO below can't race with the
9031 * buf consumer. To ensure that this copy will be
9032 * available for the lifetime of the ZIO and be cleaned
9033 * up afterwards, we add it to the l2arc_free_on_write
9034 * queue. If we need to apply any transforms to the
9035 * data (compression, encryption) we will also need the
9038 if (HDR_HAS_RABD(hdr) && psize == asize) {
9039 to_write = hdr->b_crypt_hdr.b_rabd;
9040 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
9041 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
9042 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
9044 to_write = hdr->b_l1hdr.b_pabd;
9047 arc_buf_contents_t type = arc_buf_type(hdr);
9049 ret = l2arc_apply_transforms(spa, hdr, asize,
9052 arc_hdr_clear_flags(hdr,
9053 ARC_FLAG_L2_WRITING);
9054 mutex_exit(hash_lock);
9058 l2arc_free_abd_on_write(to_write, asize, type);
9063 * Insert a dummy header on the buflist so
9064 * l2arc_write_done() can find where the
9065 * write buffers begin without searching.
9067 mutex_enter(&dev->l2ad_mtx);
9068 list_insert_head(&dev->l2ad_buflist, head);
9069 mutex_exit(&dev->l2ad_mtx);
9072 sizeof (l2arc_write_callback_t), KM_SLEEP);
9073 cb->l2wcb_dev = dev;
9074 cb->l2wcb_head = head;
9076 * Create a list to save allocated abd buffers
9077 * for l2arc_log_blk_commit().
9079 list_create(&cb->l2wcb_abd_list,
9080 sizeof (l2arc_lb_abd_buf_t),
9081 offsetof(l2arc_lb_abd_buf_t, node));
9082 pio = zio_root(spa, l2arc_write_done, cb,
9086 hdr->b_l2hdr.b_dev = dev;
9087 hdr->b_l2hdr.b_hits = 0;
9089 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
9090 arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR);
9092 mutex_enter(&dev->l2ad_mtx);
9093 list_insert_head(&dev->l2ad_buflist, hdr);
9094 mutex_exit(&dev->l2ad_mtx);
9096 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
9097 arc_hdr_size(hdr), hdr);
9099 wzio = zio_write_phys(pio, dev->l2ad_vdev,
9100 hdr->b_l2hdr.b_daddr, asize, to_write,
9101 ZIO_CHECKSUM_OFF, NULL, hdr,
9102 ZIO_PRIORITY_ASYNC_WRITE,
9103 ZIO_FLAG_CANFAIL, B_FALSE);
9105 write_lsize += HDR_GET_LSIZE(hdr);
9106 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
9109 write_psize += psize;
9110 write_asize += asize;
9111 dev->l2ad_hand += asize;
9112 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9114 mutex_exit(hash_lock);
9117 * Append buf info to current log and commit if full.
9118 * arcstat_l2_{size,asize} kstats are updated
9121 if (l2arc_log_blk_insert(dev, hdr))
9122 l2arc_log_blk_commit(dev, pio, cb);
9127 multilist_sublist_unlock(mls);
9133 /* No buffers selected for writing? */
9135 ASSERT0(write_lsize);
9136 ASSERT(!HDR_HAS_L1HDR(head));
9137 kmem_cache_free(hdr_l2only_cache, head);
9140 * Although we did not write any buffers l2ad_evict may
9143 l2arc_dev_hdr_update(dev);
9148 if (!dev->l2ad_first)
9149 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
9151 ASSERT3U(write_asize, <=, target_sz);
9152 ARCSTAT_BUMP(arcstat_l2_writes_sent);
9153 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
9154 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
9155 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
9157 dev->l2ad_writing = B_TRUE;
9158 (void) zio_wait(pio);
9159 dev->l2ad_writing = B_FALSE;
9162 * Update the device header after the zio completes as
9163 * l2arc_write_done() may have updated the memory holding the log block
9164 * pointers in the device header.
9166 l2arc_dev_hdr_update(dev);
9168 return (write_asize);
9172 l2arc_hdr_limit_reached(void)
9174 int64_t s = aggsum_upper_bound(&astat_l2_hdr_size);
9176 return (arc_reclaim_needed() || (s > arc_meta_limit * 3 / 4) ||
9177 (s > (arc_warm ? arc_c : arc_c_max) * l2arc_meta_percent / 100));
9181 * This thread feeds the L2ARC at regular intervals. This is the beating
9182 * heart of the L2ARC.
9186 l2arc_feed_thread(void *unused)
9191 uint64_t size, wrote;
9192 clock_t begin, next = ddi_get_lbolt();
9193 fstrans_cookie_t cookie;
9195 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
9197 mutex_enter(&l2arc_feed_thr_lock);
9199 cookie = spl_fstrans_mark();
9200 while (l2arc_thread_exit == 0) {
9201 CALLB_CPR_SAFE_BEGIN(&cpr);
9202 (void) cv_timedwait_idle(&l2arc_feed_thr_cv,
9203 &l2arc_feed_thr_lock, next);
9204 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
9205 next = ddi_get_lbolt() + hz;
9208 * Quick check for L2ARC devices.
9210 mutex_enter(&l2arc_dev_mtx);
9211 if (l2arc_ndev == 0) {
9212 mutex_exit(&l2arc_dev_mtx);
9215 mutex_exit(&l2arc_dev_mtx);
9216 begin = ddi_get_lbolt();
9219 * This selects the next l2arc device to write to, and in
9220 * doing so the next spa to feed from: dev->l2ad_spa. This
9221 * will return NULL if there are now no l2arc devices or if
9222 * they are all faulted.
9224 * If a device is returned, its spa's config lock is also
9225 * held to prevent device removal. l2arc_dev_get_next()
9226 * will grab and release l2arc_dev_mtx.
9228 if ((dev = l2arc_dev_get_next()) == NULL)
9231 spa = dev->l2ad_spa;
9232 ASSERT3P(spa, !=, NULL);
9235 * If the pool is read-only then force the feed thread to
9236 * sleep a little longer.
9238 if (!spa_writeable(spa)) {
9239 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
9240 spa_config_exit(spa, SCL_L2ARC, dev);
9245 * Avoid contributing to memory pressure.
9247 if (l2arc_hdr_limit_reached()) {
9248 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
9249 spa_config_exit(spa, SCL_L2ARC, dev);
9253 ARCSTAT_BUMP(arcstat_l2_feeds);
9255 size = l2arc_write_size(dev);
9258 * Evict L2ARC buffers that will be overwritten.
9260 l2arc_evict(dev, size, B_FALSE);
9263 * Write ARC buffers.
9265 wrote = l2arc_write_buffers(spa, dev, size);
9268 * Calculate interval between writes.
9270 next = l2arc_write_interval(begin, size, wrote);
9271 spa_config_exit(spa, SCL_L2ARC, dev);
9273 spl_fstrans_unmark(cookie);
9275 l2arc_thread_exit = 0;
9276 cv_broadcast(&l2arc_feed_thr_cv);
9277 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
9282 l2arc_vdev_present(vdev_t *vd)
9284 return (l2arc_vdev_get(vd) != NULL);
9288 * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if
9289 * the vdev_t isn't an L2ARC device.
9292 l2arc_vdev_get(vdev_t *vd)
9296 mutex_enter(&l2arc_dev_mtx);
9297 for (dev = list_head(l2arc_dev_list); dev != NULL;
9298 dev = list_next(l2arc_dev_list, dev)) {
9299 if (dev->l2ad_vdev == vd)
9302 mutex_exit(&l2arc_dev_mtx);
9308 * Add a vdev for use by the L2ARC. By this point the spa has already
9309 * validated the vdev and opened it.
9312 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
9314 l2arc_dev_t *adddev;
9315 uint64_t l2dhdr_asize;
9317 ASSERT(!l2arc_vdev_present(vd));
9320 * Create a new l2arc device entry.
9322 adddev = vmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
9323 adddev->l2ad_spa = spa;
9324 adddev->l2ad_vdev = vd;
9325 /* leave extra size for an l2arc device header */
9326 l2dhdr_asize = adddev->l2ad_dev_hdr_asize =
9327 MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift);
9328 adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize;
9329 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
9330 ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end);
9331 adddev->l2ad_hand = adddev->l2ad_start;
9332 adddev->l2ad_evict = adddev->l2ad_start;
9333 adddev->l2ad_first = B_TRUE;
9334 adddev->l2ad_writing = B_FALSE;
9335 adddev->l2ad_trim_all = B_FALSE;
9336 list_link_init(&adddev->l2ad_node);
9337 adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP);
9339 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
9341 * This is a list of all ARC buffers that are still valid on the
9344 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
9345 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
9348 * This is a list of pointers to log blocks that are still present
9351 list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t),
9352 offsetof(l2arc_lb_ptr_buf_t, node));
9354 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
9355 zfs_refcount_create(&adddev->l2ad_alloc);
9356 zfs_refcount_create(&adddev->l2ad_lb_asize);
9357 zfs_refcount_create(&adddev->l2ad_lb_count);
9360 * Add device to global list
9362 mutex_enter(&l2arc_dev_mtx);
9363 list_insert_head(l2arc_dev_list, adddev);
9364 atomic_inc_64(&l2arc_ndev);
9365 mutex_exit(&l2arc_dev_mtx);
9368 * Decide if vdev is eligible for L2ARC rebuild
9370 l2arc_rebuild_vdev(adddev->l2ad_vdev, B_FALSE);
9374 l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen)
9376 l2arc_dev_t *dev = NULL;
9377 l2arc_dev_hdr_phys_t *l2dhdr;
9378 uint64_t l2dhdr_asize;
9381 boolean_t l2dhdr_valid = B_TRUE;
9383 dev = l2arc_vdev_get(vd);
9384 ASSERT3P(dev, !=, NULL);
9385 spa = dev->l2ad_spa;
9386 l2dhdr = dev->l2ad_dev_hdr;
9387 l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9390 * The L2ARC has to hold at least the payload of one log block for
9391 * them to be restored (persistent L2ARC). The payload of a log block
9392 * depends on the amount of its log entries. We always write log blocks
9393 * with 1022 entries. How many of them are committed or restored depends
9394 * on the size of the L2ARC device. Thus the maximum payload of
9395 * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device
9396 * is less than that, we reduce the amount of committed and restored
9397 * log entries per block so as to enable persistence.
9399 if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) {
9400 dev->l2ad_log_entries = 0;
9402 dev->l2ad_log_entries = MIN((dev->l2ad_end -
9403 dev->l2ad_start) >> SPA_MAXBLOCKSHIFT,
9404 L2ARC_LOG_BLK_MAX_ENTRIES);
9408 * Read the device header, if an error is returned do not rebuild L2ARC.
9410 if ((err = l2arc_dev_hdr_read(dev)) != 0)
9411 l2dhdr_valid = B_FALSE;
9413 if (l2dhdr_valid && dev->l2ad_log_entries > 0) {
9415 * If we are onlining a cache device (vdev_reopen) that was
9416 * still present (l2arc_vdev_present()) and rebuild is enabled,
9417 * we should evict all ARC buffers and pointers to log blocks
9418 * and reclaim their space before restoring its contents to
9422 if (!l2arc_rebuild_enabled) {
9425 l2arc_evict(dev, 0, B_TRUE);
9426 /* start a new log block */
9427 dev->l2ad_log_ent_idx = 0;
9428 dev->l2ad_log_blk_payload_asize = 0;
9429 dev->l2ad_log_blk_payload_start = 0;
9433 * Just mark the device as pending for a rebuild. We won't
9434 * be starting a rebuild in line here as it would block pool
9435 * import. Instead spa_load_impl will hand that off to an
9436 * async task which will call l2arc_spa_rebuild_start.
9438 dev->l2ad_rebuild = B_TRUE;
9439 } else if (spa_writeable(spa)) {
9441 * In this case TRIM the whole device if l2arc_trim_ahead > 0,
9442 * otherwise create a new header. We zero out the memory holding
9443 * the header to reset dh_start_lbps. If we TRIM the whole
9444 * device the new header will be written by
9445 * vdev_trim_l2arc_thread() at the end of the TRIM to update the
9446 * trim_state in the header too. When reading the header, if
9447 * trim_state is not VDEV_TRIM_COMPLETE and l2arc_trim_ahead > 0
9448 * we opt to TRIM the whole device again.
9450 if (l2arc_trim_ahead > 0) {
9451 dev->l2ad_trim_all = B_TRUE;
9453 bzero(l2dhdr, l2dhdr_asize);
9454 l2arc_dev_hdr_update(dev);
9460 * Remove a vdev from the L2ARC.
9463 l2arc_remove_vdev(vdev_t *vd)
9465 l2arc_dev_t *remdev = NULL;
9468 * Find the device by vdev
9470 remdev = l2arc_vdev_get(vd);
9471 ASSERT3P(remdev, !=, NULL);
9474 * Cancel any ongoing or scheduled rebuild.
9476 mutex_enter(&l2arc_rebuild_thr_lock);
9477 if (remdev->l2ad_rebuild_began == B_TRUE) {
9478 remdev->l2ad_rebuild_cancel = B_TRUE;
9479 while (remdev->l2ad_rebuild == B_TRUE)
9480 cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock);
9482 mutex_exit(&l2arc_rebuild_thr_lock);
9485 * Remove device from global list
9487 mutex_enter(&l2arc_dev_mtx);
9488 list_remove(l2arc_dev_list, remdev);
9489 l2arc_dev_last = NULL; /* may have been invalidated */
9490 atomic_dec_64(&l2arc_ndev);
9491 mutex_exit(&l2arc_dev_mtx);
9494 * Clear all buflists and ARC references. L2ARC device flush.
9496 l2arc_evict(remdev, 0, B_TRUE);
9497 list_destroy(&remdev->l2ad_buflist);
9498 ASSERT(list_is_empty(&remdev->l2ad_lbptr_list));
9499 list_destroy(&remdev->l2ad_lbptr_list);
9500 mutex_destroy(&remdev->l2ad_mtx);
9501 zfs_refcount_destroy(&remdev->l2ad_alloc);
9502 zfs_refcount_destroy(&remdev->l2ad_lb_asize);
9503 zfs_refcount_destroy(&remdev->l2ad_lb_count);
9504 kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize);
9505 vmem_free(remdev, sizeof (l2arc_dev_t));
9511 l2arc_thread_exit = 0;
9513 l2arc_writes_sent = 0;
9514 l2arc_writes_done = 0;
9516 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9517 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
9518 mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9519 cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL);
9520 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
9521 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
9523 l2arc_dev_list = &L2ARC_dev_list;
9524 l2arc_free_on_write = &L2ARC_free_on_write;
9525 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
9526 offsetof(l2arc_dev_t, l2ad_node));
9527 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
9528 offsetof(l2arc_data_free_t, l2df_list_node));
9534 mutex_destroy(&l2arc_feed_thr_lock);
9535 cv_destroy(&l2arc_feed_thr_cv);
9536 mutex_destroy(&l2arc_rebuild_thr_lock);
9537 cv_destroy(&l2arc_rebuild_thr_cv);
9538 mutex_destroy(&l2arc_dev_mtx);
9539 mutex_destroy(&l2arc_free_on_write_mtx);
9541 list_destroy(l2arc_dev_list);
9542 list_destroy(l2arc_free_on_write);
9548 if (!(spa_mode_global & SPA_MODE_WRITE))
9551 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
9552 TS_RUN, defclsyspri);
9558 if (!(spa_mode_global & SPA_MODE_WRITE))
9561 mutex_enter(&l2arc_feed_thr_lock);
9562 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
9563 l2arc_thread_exit = 1;
9564 while (l2arc_thread_exit != 0)
9565 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
9566 mutex_exit(&l2arc_feed_thr_lock);
9570 * Punches out rebuild threads for the L2ARC devices in a spa. This should
9571 * be called after pool import from the spa async thread, since starting
9572 * these threads directly from spa_import() will make them part of the
9573 * "zpool import" context and delay process exit (and thus pool import).
9576 l2arc_spa_rebuild_start(spa_t *spa)
9578 ASSERT(MUTEX_HELD(&spa_namespace_lock));
9581 * Locate the spa's l2arc devices and kick off rebuild threads.
9583 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
9585 l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]);
9587 /* Don't attempt a rebuild if the vdev is UNAVAIL */
9590 mutex_enter(&l2arc_rebuild_thr_lock);
9591 if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) {
9592 dev->l2ad_rebuild_began = B_TRUE;
9593 (void) thread_create(NULL, 0, l2arc_dev_rebuild_thread,
9594 dev, 0, &p0, TS_RUN, minclsyspri);
9596 mutex_exit(&l2arc_rebuild_thr_lock);
9601 * Main entry point for L2ARC rebuilding.
9604 l2arc_dev_rebuild_thread(void *arg)
9606 l2arc_dev_t *dev = arg;
9608 VERIFY(!dev->l2ad_rebuild_cancel);
9609 VERIFY(dev->l2ad_rebuild);
9610 (void) l2arc_rebuild(dev);
9611 mutex_enter(&l2arc_rebuild_thr_lock);
9612 dev->l2ad_rebuild_began = B_FALSE;
9613 dev->l2ad_rebuild = B_FALSE;
9614 mutex_exit(&l2arc_rebuild_thr_lock);
9620 * This function implements the actual L2ARC metadata rebuild. It:
9621 * starts reading the log block chain and restores each block's contents
9622 * to memory (reconstructing arc_buf_hdr_t's).
9624 * Operation stops under any of the following conditions:
9626 * 1) We reach the end of the log block chain.
9627 * 2) We encounter *any* error condition (cksum errors, io errors)
9630 l2arc_rebuild(l2arc_dev_t *dev)
9632 vdev_t *vd = dev->l2ad_vdev;
9633 spa_t *spa = vd->vdev_spa;
9635 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9636 l2arc_log_blk_phys_t *this_lb, *next_lb;
9637 zio_t *this_io = NULL, *next_io = NULL;
9638 l2arc_log_blkptr_t lbps[2];
9639 l2arc_lb_ptr_buf_t *lb_ptr_buf;
9640 boolean_t lock_held;
9642 this_lb = vmem_zalloc(sizeof (*this_lb), KM_SLEEP);
9643 next_lb = vmem_zalloc(sizeof (*next_lb), KM_SLEEP);
9646 * We prevent device removal while issuing reads to the device,
9647 * then during the rebuilding phases we drop this lock again so
9648 * that a spa_unload or device remove can be initiated - this is
9649 * safe, because the spa will signal us to stop before removing
9650 * our device and wait for us to stop.
9652 spa_config_enter(spa, SCL_L2ARC, vd, RW_READER);
9656 * Retrieve the persistent L2ARC device state.
9657 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9659 dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start);
9660 dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr +
9661 L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop),
9663 dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST);
9665 vd->vdev_trim_action_time = l2dhdr->dh_trim_action_time;
9666 vd->vdev_trim_state = l2dhdr->dh_trim_state;
9669 * In case the zfs module parameter l2arc_rebuild_enabled is false
9670 * we do not start the rebuild process.
9672 if (!l2arc_rebuild_enabled)
9675 /* Prepare the rebuild process */
9676 bcopy(l2dhdr->dh_start_lbps, lbps, sizeof (lbps));
9678 /* Start the rebuild process */
9680 if (!l2arc_log_blkptr_valid(dev, &lbps[0]))
9683 if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1],
9684 this_lb, next_lb, this_io, &next_io)) != 0)
9688 * Our memory pressure valve. If the system is running low
9689 * on memory, rather than swamping memory with new ARC buf
9690 * hdrs, we opt not to rebuild the L2ARC. At this point,
9691 * however, we have already set up our L2ARC dev to chain in
9692 * new metadata log blocks, so the user may choose to offline/
9693 * online the L2ARC dev at a later time (or re-import the pool)
9694 * to reconstruct it (when there's less memory pressure).
9696 if (l2arc_hdr_limit_reached()) {
9697 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem);
9698 cmn_err(CE_NOTE, "System running low on memory, "
9699 "aborting L2ARC rebuild.");
9700 err = SET_ERROR(ENOMEM);
9704 spa_config_exit(spa, SCL_L2ARC, vd);
9705 lock_held = B_FALSE;
9708 * Now that we know that the next_lb checks out alright, we
9709 * can start reconstruction from this log block.
9710 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9712 uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop);
9713 l2arc_log_blk_restore(dev, this_lb, asize, lbps[0].lbp_daddr);
9716 * log block restored, include its pointer in the list of
9717 * pointers to log blocks present in the L2ARC device.
9719 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
9720 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t),
9722 bcopy(&lbps[0], lb_ptr_buf->lb_ptr,
9723 sizeof (l2arc_log_blkptr_t));
9724 mutex_enter(&dev->l2ad_mtx);
9725 list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf);
9726 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
9727 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
9728 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
9729 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
9730 mutex_exit(&dev->l2ad_mtx);
9731 vdev_space_update(vd, asize, 0, 0);
9734 * Protection against loops of log blocks:
9736 * l2ad_hand l2ad_evict
9738 * l2ad_start |=======================================| l2ad_end
9739 * -----|||----|||---|||----|||
9741 * ---|||---|||----|||---|||
9744 * In this situation the pointer of log block (4) passes
9745 * l2arc_log_blkptr_valid() but the log block should not be
9746 * restored as it is overwritten by the payload of log block
9747 * (0). Only log blocks (0)-(3) should be restored. We check
9748 * whether l2ad_evict lies in between the payload starting
9749 * offset of the next log block (lbps[1].lbp_payload_start)
9750 * and the payload starting offset of the present log block
9751 * (lbps[0].lbp_payload_start). If true and this isn't the
9752 * first pass, we are looping from the beginning and we should
9755 if (l2arc_range_check_overlap(lbps[1].lbp_payload_start,
9756 lbps[0].lbp_payload_start, dev->l2ad_evict) &&
9761 mutex_enter(&l2arc_rebuild_thr_lock);
9762 if (dev->l2ad_rebuild_cancel) {
9763 dev->l2ad_rebuild = B_FALSE;
9764 cv_signal(&l2arc_rebuild_thr_cv);
9765 mutex_exit(&l2arc_rebuild_thr_lock);
9766 err = SET_ERROR(ECANCELED);
9769 mutex_exit(&l2arc_rebuild_thr_lock);
9770 if (spa_config_tryenter(spa, SCL_L2ARC, vd,
9776 * L2ARC config lock held by somebody in writer,
9777 * possibly due to them trying to remove us. They'll
9778 * likely to want us to shut down, so after a little
9779 * delay, we check l2ad_rebuild_cancel and retry
9786 * Continue with the next log block.
9789 lbps[1] = this_lb->lb_prev_lbp;
9790 PTR_SWAP(this_lb, next_lb);
9795 if (this_io != NULL)
9796 l2arc_log_blk_fetch_abort(this_io);
9798 if (next_io != NULL)
9799 l2arc_log_blk_fetch_abort(next_io);
9800 vmem_free(this_lb, sizeof (*this_lb));
9801 vmem_free(next_lb, sizeof (*next_lb));
9803 if (!l2arc_rebuild_enabled) {
9804 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9806 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) {
9807 ARCSTAT_BUMP(arcstat_l2_rebuild_success);
9808 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9809 "successful, restored %llu blocks",
9810 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
9811 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) {
9813 * No error but also nothing restored, meaning the lbps array
9814 * in the device header points to invalid/non-present log
9815 * blocks. Reset the header.
9817 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9818 "no valid log blocks");
9819 bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
9820 l2arc_dev_hdr_update(dev);
9821 } else if (err == ECANCELED) {
9823 * In case the rebuild was canceled do not log to spa history
9824 * log as the pool may be in the process of being removed.
9826 zfs_dbgmsg("L2ARC rebuild aborted, restored %llu blocks",
9827 zfs_refcount_count(&dev->l2ad_lb_count));
9828 } else if (err != 0) {
9829 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9830 "aborted, restored %llu blocks",
9831 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
9835 spa_config_exit(spa, SCL_L2ARC, vd);
9841 * Attempts to read the device header on the provided L2ARC device and writes
9842 * it to `hdr'. On success, this function returns 0, otherwise the appropriate
9843 * error code is returned.
9846 l2arc_dev_hdr_read(l2arc_dev_t *dev)
9850 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9851 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9854 guid = spa_guid(dev->l2ad_vdev->vdev_spa);
9856 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
9858 err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev,
9859 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd,
9860 ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
9861 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
9862 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY |
9863 ZIO_FLAG_SPECULATIVE, B_FALSE));
9868 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors);
9869 zfs_dbgmsg("L2ARC IO error (%d) while reading device header, "
9870 "vdev guid: %llu", err, dev->l2ad_vdev->vdev_guid);
9874 if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC))
9875 byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr));
9877 if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC ||
9878 l2dhdr->dh_spa_guid != guid ||
9879 l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid ||
9880 l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION ||
9881 l2dhdr->dh_log_entries != dev->l2ad_log_entries ||
9882 l2dhdr->dh_end != dev->l2ad_end ||
9883 !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end,
9884 l2dhdr->dh_evict) ||
9885 (l2dhdr->dh_trim_state != VDEV_TRIM_COMPLETE &&
9886 l2arc_trim_ahead > 0)) {
9888 * Attempt to rebuild a device containing no actual dev hdr
9889 * or containing a header from some other pool or from another
9890 * version of persistent L2ARC.
9892 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported);
9893 return (SET_ERROR(ENOTSUP));
9900 * Reads L2ARC log blocks from storage and validates their contents.
9902 * This function implements a simple fetcher to make sure that while
9903 * we're processing one buffer the L2ARC is already fetching the next
9906 * The arguments this_lp and next_lp point to the current and next log block
9907 * address in the block chain. Similarly, this_lb and next_lb hold the
9908 * l2arc_log_blk_phys_t's of the current and next L2ARC blk.
9910 * The `this_io' and `next_io' arguments are used for block fetching.
9911 * When issuing the first blk IO during rebuild, you should pass NULL for
9912 * `this_io'. This function will then issue a sync IO to read the block and
9913 * also issue an async IO to fetch the next block in the block chain. The
9914 * fetched IO is returned in `next_io'. On subsequent calls to this
9915 * function, pass the value returned in `next_io' from the previous call
9916 * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO.
9917 * Prior to the call, you should initialize your `next_io' pointer to be
9918 * NULL. If no fetch IO was issued, the pointer is left set at NULL.
9920 * On success, this function returns 0, otherwise it returns an appropriate
9921 * error code. On error the fetching IO is aborted and cleared before
9922 * returning from this function. Therefore, if we return `success', the
9923 * caller can assume that we have taken care of cleanup of fetch IOs.
9926 l2arc_log_blk_read(l2arc_dev_t *dev,
9927 const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp,
9928 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
9929 zio_t *this_io, zio_t **next_io)
9936 ASSERT(this_lbp != NULL && next_lbp != NULL);
9937 ASSERT(this_lb != NULL && next_lb != NULL);
9938 ASSERT(next_io != NULL && *next_io == NULL);
9939 ASSERT(l2arc_log_blkptr_valid(dev, this_lbp));
9942 * Check to see if we have issued the IO for this log block in a
9943 * previous run. If not, this is the first call, so issue it now.
9945 if (this_io == NULL) {
9946 this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp,
9951 * Peek to see if we can start issuing the next IO immediately.
9953 if (l2arc_log_blkptr_valid(dev, next_lbp)) {
9955 * Start issuing IO for the next log block early - this
9956 * should help keep the L2ARC device busy while we
9957 * decompress and restore this log block.
9959 *next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp,
9963 /* Wait for the IO to read this log block to complete */
9964 if ((err = zio_wait(this_io)) != 0) {
9965 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
9966 zfs_dbgmsg("L2ARC IO error (%d) while reading log block, "
9967 "offset: %llu, vdev guid: %llu", err, this_lbp->lbp_daddr,
9968 dev->l2ad_vdev->vdev_guid);
9973 * Make sure the buffer checks out.
9974 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9976 asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop);
9977 fletcher_4_native(this_lb, asize, NULL, &cksum);
9978 if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) {
9979 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors);
9980 zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, "
9981 "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu",
9982 this_lbp->lbp_daddr, dev->l2ad_vdev->vdev_guid,
9983 dev->l2ad_hand, dev->l2ad_evict);
9984 err = SET_ERROR(ECKSUM);
9988 /* Now we can take our time decoding this buffer */
9989 switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) {
9990 case ZIO_COMPRESS_OFF:
9992 case ZIO_COMPRESS_LZ4:
9993 abd = abd_alloc_for_io(asize, B_TRUE);
9994 abd_copy_from_buf_off(abd, this_lb, 0, asize);
9995 if ((err = zio_decompress_data(
9996 L2BLK_GET_COMPRESS((this_lbp)->lbp_prop),
9997 abd, this_lb, asize, sizeof (*this_lb), NULL)) != 0) {
9998 err = SET_ERROR(EINVAL);
10003 err = SET_ERROR(EINVAL);
10006 if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC))
10007 byteswap_uint64_array(this_lb, sizeof (*this_lb));
10008 if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) {
10009 err = SET_ERROR(EINVAL);
10013 /* Abort an in-flight fetch I/O in case of error */
10014 if (err != 0 && *next_io != NULL) {
10015 l2arc_log_blk_fetch_abort(*next_io);
10024 * Restores the payload of a log block to ARC. This creates empty ARC hdr
10025 * entries which only contain an l2arc hdr, essentially restoring the
10026 * buffers to their L2ARC evicted state. This function also updates space
10027 * usage on the L2ARC vdev to make sure it tracks restored buffers.
10030 l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb,
10031 uint64_t lb_asize, uint64_t lb_daddr)
10033 uint64_t size = 0, asize = 0;
10034 uint64_t log_entries = dev->l2ad_log_entries;
10037 * Usually arc_adapt() is called only for data, not headers, but
10038 * since we may allocate significant amount of memory here, let ARC
10041 arc_adapt(log_entries * HDR_L2ONLY_SIZE, arc_l2c_only);
10043 for (int i = log_entries - 1; i >= 0; i--) {
10045 * Restore goes in the reverse temporal direction to preserve
10046 * correct temporal ordering of buffers in the l2ad_buflist.
10047 * l2arc_hdr_restore also does a list_insert_tail instead of
10048 * list_insert_head on the l2ad_buflist:
10050 * LIST l2ad_buflist LIST
10051 * HEAD <------ (time) ------ TAIL
10052 * direction +-----+-----+-----+-----+-----+ direction
10053 * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild
10054 * fill +-----+-----+-----+-----+-----+
10058 * l2arc_feed_thread l2arc_rebuild
10059 * will place new bufs here restores bufs here
10061 * During l2arc_rebuild() the device is not used by
10062 * l2arc_feed_thread() as dev->l2ad_rebuild is set to true.
10064 size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop);
10065 asize += vdev_psize_to_asize(dev->l2ad_vdev,
10066 L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop));
10067 l2arc_hdr_restore(&lb->lb_entries[i], dev);
10071 * Record rebuild stats:
10072 * size Logical size of restored buffers in the L2ARC
10073 * asize Aligned size of restored buffers in the L2ARC
10075 ARCSTAT_INCR(arcstat_l2_rebuild_size, size);
10076 ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize);
10077 ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries);
10078 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize);
10079 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize);
10080 ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks);
10084 * Restores a single ARC buf hdr from a log entry. The ARC buffer is put
10085 * into a state indicating that it has been evicted to L2ARC.
10088 l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
10090 arc_buf_hdr_t *hdr, *exists;
10091 kmutex_t *hash_lock;
10092 arc_buf_contents_t type = L2BLK_GET_TYPE((le)->le_prop);
10096 * Do all the allocation before grabbing any locks, this lets us
10097 * sleep if memory is full and we don't have to deal with failed
10100 hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type,
10101 dev, le->le_dva, le->le_daddr,
10102 L2BLK_GET_PSIZE((le)->le_prop), le->le_birth,
10103 L2BLK_GET_COMPRESS((le)->le_prop), le->le_complevel,
10104 L2BLK_GET_PROTECTED((le)->le_prop),
10105 L2BLK_GET_PREFETCH((le)->le_prop));
10106 asize = vdev_psize_to_asize(dev->l2ad_vdev,
10107 L2BLK_GET_PSIZE((le)->le_prop));
10110 * vdev_space_update() has to be called before arc_hdr_destroy() to
10111 * avoid underflow since the latter also calls the former.
10113 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10115 ARCSTAT_INCR(arcstat_l2_lsize, HDR_GET_LSIZE(hdr));
10116 ARCSTAT_INCR(arcstat_l2_psize, HDR_GET_PSIZE(hdr));
10118 mutex_enter(&dev->l2ad_mtx);
10119 list_insert_tail(&dev->l2ad_buflist, hdr);
10120 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
10121 mutex_exit(&dev->l2ad_mtx);
10123 exists = buf_hash_insert(hdr, &hash_lock);
10125 /* Buffer was already cached, no need to restore it. */
10126 arc_hdr_destroy(hdr);
10128 * If the buffer is already cached, check whether it has
10129 * L2ARC metadata. If not, enter them and update the flag.
10130 * This is important is case of onlining a cache device, since
10131 * we previously evicted all L2ARC metadata from ARC.
10133 if (!HDR_HAS_L2HDR(exists)) {
10134 arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR);
10135 exists->b_l2hdr.b_dev = dev;
10136 exists->b_l2hdr.b_daddr = le->le_daddr;
10137 mutex_enter(&dev->l2ad_mtx);
10138 list_insert_tail(&dev->l2ad_buflist, exists);
10139 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
10140 arc_hdr_size(exists), exists);
10141 mutex_exit(&dev->l2ad_mtx);
10142 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10143 ARCSTAT_INCR(arcstat_l2_lsize, HDR_GET_LSIZE(exists));
10144 ARCSTAT_INCR(arcstat_l2_psize, HDR_GET_PSIZE(exists));
10146 ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
10149 mutex_exit(hash_lock);
10153 * Starts an asynchronous read IO to read a log block. This is used in log
10154 * block reconstruction to start reading the next block before we are done
10155 * decoding and reconstructing the current block, to keep the l2arc device
10156 * nice and hot with read IO to process.
10157 * The returned zio will contain a newly allocated memory buffers for the IO
10158 * data which should then be freed by the caller once the zio is no longer
10159 * needed (i.e. due to it having completed). If you wish to abort this
10160 * zio, you should do so using l2arc_log_blk_fetch_abort, which takes
10161 * care of disposing of the allocated buffers correctly.
10164 l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp,
10165 l2arc_log_blk_phys_t *lb)
10169 l2arc_read_callback_t *cb;
10171 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10172 asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10173 ASSERT(asize <= sizeof (l2arc_log_blk_phys_t));
10175 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP);
10176 cb->l2rcb_abd = abd_get_from_buf(lb, asize);
10177 pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb,
10178 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
10179 ZIO_FLAG_DONT_RETRY);
10180 (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize,
10181 cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10182 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
10183 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
10189 * Aborts a zio returned from l2arc_log_blk_fetch and frees the data
10190 * buffers allocated for it.
10193 l2arc_log_blk_fetch_abort(zio_t *zio)
10195 (void) zio_wait(zio);
10199 * Creates a zio to update the device header on an l2arc device.
10202 l2arc_dev_hdr_update(l2arc_dev_t *dev)
10204 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10205 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10209 VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER));
10211 l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC;
10212 l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION;
10213 l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10214 l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid;
10215 l2dhdr->dh_log_entries = dev->l2ad_log_entries;
10216 l2dhdr->dh_evict = dev->l2ad_evict;
10217 l2dhdr->dh_start = dev->l2ad_start;
10218 l2dhdr->dh_end = dev->l2ad_end;
10219 l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize);
10220 l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count);
10221 l2dhdr->dh_flags = 0;
10222 l2dhdr->dh_trim_action_time = dev->l2ad_vdev->vdev_trim_action_time;
10223 l2dhdr->dh_trim_state = dev->l2ad_vdev->vdev_trim_state;
10224 if (dev->l2ad_first)
10225 l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST;
10227 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10229 err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev,
10230 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL,
10231 NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE));
10236 zfs_dbgmsg("L2ARC IO error (%d) while writing device header, "
10237 "vdev guid: %llu", err, dev->l2ad_vdev->vdev_guid);
10242 * Commits a log block to the L2ARC device. This routine is invoked from
10243 * l2arc_write_buffers when the log block fills up.
10244 * This function allocates some memory to temporarily hold the serialized
10245 * buffer to be written. This is then released in l2arc_write_done.
10248 l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb)
10250 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10251 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10252 uint64_t psize, asize;
10254 l2arc_lb_abd_buf_t *abd_buf;
10256 l2arc_lb_ptr_buf_t *lb_ptr_buf;
10258 VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries);
10260 tmpbuf = zio_buf_alloc(sizeof (*lb));
10261 abd_buf = zio_buf_alloc(sizeof (*abd_buf));
10262 abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb));
10263 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
10264 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP);
10266 /* link the buffer into the block chain */
10267 lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1];
10268 lb->lb_magic = L2ARC_LOG_BLK_MAGIC;
10271 * l2arc_log_blk_commit() may be called multiple times during a single
10272 * l2arc_write_buffers() call. Save the allocated abd buffers in a list
10273 * so we can free them in l2arc_write_done() later on.
10275 list_insert_tail(&cb->l2wcb_abd_list, abd_buf);
10277 /* try to compress the buffer */
10278 psize = zio_compress_data(ZIO_COMPRESS_LZ4,
10279 abd_buf->abd, tmpbuf, sizeof (*lb), 0);
10281 /* a log block is never entirely zero */
10282 ASSERT(psize != 0);
10283 asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
10284 ASSERT(asize <= sizeof (*lb));
10287 * Update the start log block pointer in the device header to point
10288 * to the log block we're about to write.
10290 l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0];
10291 l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand;
10292 l2dhdr->dh_start_lbps[0].lbp_payload_asize =
10293 dev->l2ad_log_blk_payload_asize;
10294 l2dhdr->dh_start_lbps[0].lbp_payload_start =
10295 dev->l2ad_log_blk_payload_start;
10298 (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb));
10300 (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize);
10301 L2BLK_SET_CHECKSUM(
10302 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10303 ZIO_CHECKSUM_FLETCHER_4);
10304 if (asize < sizeof (*lb)) {
10305 /* compression succeeded */
10306 bzero(tmpbuf + psize, asize - psize);
10307 L2BLK_SET_COMPRESS(
10308 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10311 /* compression failed */
10312 bcopy(lb, tmpbuf, sizeof (*lb));
10313 L2BLK_SET_COMPRESS(
10314 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10318 /* checksum what we're about to write */
10319 fletcher_4_native(tmpbuf, asize, NULL,
10320 &l2dhdr->dh_start_lbps[0].lbp_cksum);
10322 abd_put(abd_buf->abd);
10324 /* perform the write itself */
10325 abd_buf->abd = abd_get_from_buf(tmpbuf, sizeof (*lb));
10326 abd_take_ownership_of_buf(abd_buf->abd, B_TRUE);
10327 wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand,
10328 asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10329 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
10330 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
10331 (void) zio_nowait(wzio);
10333 dev->l2ad_hand += asize;
10335 * Include the committed log block's pointer in the list of pointers
10336 * to log blocks present in the L2ARC device.
10338 bcopy(&l2dhdr->dh_start_lbps[0], lb_ptr_buf->lb_ptr,
10339 sizeof (l2arc_log_blkptr_t));
10340 mutex_enter(&dev->l2ad_mtx);
10341 list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf);
10342 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10343 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10344 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10345 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10346 mutex_exit(&dev->l2ad_mtx);
10347 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10349 /* bump the kstats */
10350 ARCSTAT_INCR(arcstat_l2_write_bytes, asize);
10351 ARCSTAT_BUMP(arcstat_l2_log_blk_writes);
10352 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize);
10353 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio,
10354 dev->l2ad_log_blk_payload_asize / asize);
10356 /* start a new log block */
10357 dev->l2ad_log_ent_idx = 0;
10358 dev->l2ad_log_blk_payload_asize = 0;
10359 dev->l2ad_log_blk_payload_start = 0;
10363 * Validates an L2ARC log block address to make sure that it can be read
10364 * from the provided L2ARC device.
10367 l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp)
10369 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10370 uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10371 uint64_t end = lbp->lbp_daddr + asize - 1;
10372 uint64_t start = lbp->lbp_payload_start;
10373 boolean_t evicted = B_FALSE;
10376 * A log block is valid if all of the following conditions are true:
10377 * - it fits entirely (including its payload) between l2ad_start and
10379 * - it has a valid size
10380 * - neither the log block itself nor part of its payload was evicted
10381 * by l2arc_evict():
10383 * l2ad_hand l2ad_evict
10388 * l2ad_start ============================================ l2ad_end
10389 * --------------------------||||
10396 l2arc_range_check_overlap(start, end, dev->l2ad_hand) ||
10397 l2arc_range_check_overlap(start, end, dev->l2ad_evict) ||
10398 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) ||
10399 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end);
10401 return (start >= dev->l2ad_start && end <= dev->l2ad_end &&
10402 asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) &&
10403 (!evicted || dev->l2ad_first));
10407 * Inserts ARC buffer header `hdr' into the current L2ARC log block on
10408 * the device. The buffer being inserted must be present in L2ARC.
10409 * Returns B_TRUE if the L2ARC log block is full and needs to be committed
10410 * to L2ARC, or B_FALSE if it still has room for more ARC buffers.
10413 l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr)
10415 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10416 l2arc_log_ent_phys_t *le;
10418 if (dev->l2ad_log_entries == 0)
10421 int index = dev->l2ad_log_ent_idx++;
10423 ASSERT3S(index, <, dev->l2ad_log_entries);
10424 ASSERT(HDR_HAS_L2HDR(hdr));
10426 le = &lb->lb_entries[index];
10427 bzero(le, sizeof (*le));
10428 le->le_dva = hdr->b_dva;
10429 le->le_birth = hdr->b_birth;
10430 le->le_daddr = hdr->b_l2hdr.b_daddr;
10432 dev->l2ad_log_blk_payload_start = le->le_daddr;
10433 L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr));
10434 L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr));
10435 L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr));
10436 le->le_complevel = hdr->b_complevel;
10437 L2BLK_SET_TYPE((le)->le_prop, hdr->b_type);
10438 L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr)));
10439 L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr)));
10441 dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev,
10442 HDR_GET_PSIZE(hdr));
10444 return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries);
10448 * Checks whether a given L2ARC device address sits in a time-sequential
10449 * range. The trick here is that the L2ARC is a rotary buffer, so we can't
10450 * just do a range comparison, we need to handle the situation in which the
10451 * range wraps around the end of the L2ARC device. Arguments:
10452 * bottom -- Lower end of the range to check (written to earlier).
10453 * top -- Upper end of the range to check (written to later).
10454 * check -- The address for which we want to determine if it sits in
10455 * between the top and bottom.
10457 * The 3-way conditional below represents the following cases:
10459 * bottom < top : Sequentially ordered case:
10460 * <check>--------+-------------------+
10461 * | (overlap here?) |
10463 * |---------------<bottom>============<top>--------------|
10465 * bottom > top: Looped-around case:
10466 * <check>--------+------------------+
10467 * | (overlap here?) |
10469 * |===============<top>---------------<bottom>===========|
10472 * +---------------+---------<check>
10474 * top == bottom : Just a single address comparison.
10477 l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check)
10480 return (bottom <= check && check <= top);
10481 else if (bottom > top)
10482 return (check <= top || bottom <= check);
10484 return (check == top);
10487 EXPORT_SYMBOL(arc_buf_size);
10488 EXPORT_SYMBOL(arc_write);
10489 EXPORT_SYMBOL(arc_read);
10490 EXPORT_SYMBOL(arc_buf_info);
10491 EXPORT_SYMBOL(arc_getbuf_func);
10492 EXPORT_SYMBOL(arc_add_prune_callback);
10493 EXPORT_SYMBOL(arc_remove_prune_callback);
10495 /* BEGIN CSTYLED */
10496 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min, param_set_arc_long,
10497 param_get_long, ZMOD_RW, "Min arc size");
10499 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, max, param_set_arc_long,
10500 param_get_long, ZMOD_RW, "Max arc size");
10502 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit, param_set_arc_long,
10503 param_get_long, ZMOD_RW, "Metadata limit for arc size");
10505 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit_percent,
10506 param_set_arc_long, param_get_long, ZMOD_RW,
10507 "Percent of arc size for arc meta limit");
10509 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_min, param_set_arc_long,
10510 param_get_long, ZMOD_RW, "Min arc metadata");
10512 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_prune, INT, ZMOD_RW,
10513 "Meta objects to scan for prune");
10515 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_adjust_restarts, INT, ZMOD_RW,
10516 "Limit number of restarts in arc_evict_meta");
10518 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_strategy, INT, ZMOD_RW,
10519 "Meta reclaim strategy");
10521 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, grow_retry, param_set_arc_int,
10522 param_get_int, ZMOD_RW, "Seconds before growing arc size");
10524 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, p_dampener_disable, INT, ZMOD_RW,
10525 "Disable arc_p adapt dampener");
10527 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, shrink_shift, param_set_arc_int,
10528 param_get_int, ZMOD_RW, "log2(fraction of arc to reclaim)");
10530 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, pc_percent, UINT, ZMOD_RW,
10531 "Percent of pagecache to reclaim arc to");
10533 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, p_min_shift, param_set_arc_int,
10534 param_get_int, ZMOD_RW, "arc_c shift to calc min/max arc_p");
10536 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, average_blocksize, INT, ZMOD_RD,
10537 "Target average block size");
10539 ZFS_MODULE_PARAM(zfs, zfs_, compressed_arc_enabled, INT, ZMOD_RW,
10540 "Disable compressed arc buffers");
10542 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prefetch_ms, param_set_arc_int,
10543 param_get_int, ZMOD_RW, "Min life of prefetch block in ms");
10545 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prescient_prefetch_ms,
10546 param_set_arc_int, param_get_int, ZMOD_RW,
10547 "Min life of prescient prefetched block in ms");
10549 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_max, ULONG, ZMOD_RW,
10550 "Max write bytes per interval");
10552 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_boost, ULONG, ZMOD_RW,
10553 "Extra write bytes during device warmup");
10555 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom, ULONG, ZMOD_RW,
10556 "Number of max device writes to precache");
10558 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom_boost, ULONG, ZMOD_RW,
10559 "Compressed l2arc_headroom multiplier");
10561 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, trim_ahead, ULONG, ZMOD_RW,
10562 "TRIM ahead L2ARC write size multiplier");
10564 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_secs, ULONG, ZMOD_RW,
10565 "Seconds between L2ARC writing");
10567 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_min_ms, ULONG, ZMOD_RW,
10568 "Min feed interval in milliseconds");
10570 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, noprefetch, INT, ZMOD_RW,
10571 "Skip caching prefetched buffers");
10573 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_again, INT, ZMOD_RW,
10574 "Turbo L2ARC warmup");
10576 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, norw, INT, ZMOD_RW,
10577 "No reads during writes");
10579 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, meta_percent, INT, ZMOD_RW,
10580 "Percent of ARC size allowed for L2ARC-only headers");
10582 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_enabled, INT, ZMOD_RW,
10583 "Rebuild the L2ARC when importing a pool");
10585 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_blocks_min_l2size, ULONG, ZMOD_RW,
10586 "Min size in bytes to write rebuild log blocks in L2ARC");
10588 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, mfuonly, INT, ZMOD_RW,
10589 "Cache only MFU data from ARC into L2ARC");
10591 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, lotsfree_percent, param_set_arc_int,
10592 param_get_int, ZMOD_RW, "System free memory I/O throttle in bytes");
10594 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, sys_free, param_set_arc_long,
10595 param_get_long, ZMOD_RW, "System free memory target size in bytes");
10597 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit, param_set_arc_long,
10598 param_get_long, ZMOD_RW, "Minimum bytes of dnodes in arc");
10600 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit_percent,
10601 param_set_arc_long, param_get_long, ZMOD_RW,
10602 "Percent of ARC meta buffers for dnodes");
10604 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, dnode_reduce_percent, ULONG, ZMOD_RW,
10605 "Percentage of excess dnodes to try to unpin");
10607 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, eviction_pct, INT, ZMOD_RW,
10608 "When full, ARC allocation waits for eviction of this % of alloc size");