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 https://opensource.org/licenses/CDDL-1.0.
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 memcpy 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/multilist.h>
300 #include <sys/fm/fs/zfs.h>
301 #include <sys/callb.h>
302 #include <sys/kstat.h>
303 #include <sys/zthr.h>
304 #include <zfs_fletcher.h>
305 #include <sys/arc_impl.h>
306 #include <sys/trace_zfs.h>
307 #include <sys/aggsum.h>
308 #include <sys/wmsum.h>
309 #include <cityhash.h>
310 #include <sys/vdev_trim.h>
311 #include <sys/zfs_racct.h>
312 #include <sys/zstd/zstd.h>
315 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
316 boolean_t arc_watch = B_FALSE;
320 * This thread's job is to keep enough free memory in the system, by
321 * calling arc_kmem_reap_soon() plus arc_reduce_target_size(), which improves
322 * arc_available_memory().
324 static zthr_t *arc_reap_zthr;
327 * This thread's job is to keep arc_size under arc_c, by calling
328 * arc_evict(), which improves arc_is_overflowing().
330 static zthr_t *arc_evict_zthr;
331 static arc_buf_hdr_t **arc_state_evict_markers;
332 static int arc_state_evict_marker_count;
334 static kmutex_t arc_evict_lock;
335 static boolean_t arc_evict_needed = B_FALSE;
338 * Count of bytes evicted since boot.
340 static uint64_t arc_evict_count;
343 * List of arc_evict_waiter_t's, representing threads waiting for the
344 * arc_evict_count to reach specific values.
346 static list_t arc_evict_waiters;
349 * When arc_is_overflowing(), arc_get_data_impl() waits for this percent of
350 * the requested amount of data to be evicted. For example, by default for
351 * every 2KB that's evicted, 1KB of it may be "reused" by a new allocation.
352 * Since this is above 100%, it ensures that progress is made towards getting
353 * arc_size under arc_c. Since this is finite, it ensures that allocations
354 * can still happen, even during the potentially long time that arc_size is
357 static uint_t zfs_arc_eviction_pct = 200;
360 * The number of headers to evict in arc_evict_state_impl() before
361 * dropping the sublist lock and evicting from another sublist. A lower
362 * value means we're more likely to evict the "correct" header (i.e. the
363 * oldest header in the arc state), but comes with higher overhead
364 * (i.e. more invocations of arc_evict_state_impl()).
366 static uint_t zfs_arc_evict_batch_limit = 10;
368 /* number of seconds before growing cache again */
369 uint_t arc_grow_retry = 5;
372 * Minimum time between calls to arc_kmem_reap_soon().
374 static const int arc_kmem_cache_reap_retry_ms = 1000;
376 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
377 static int zfs_arc_overflow_shift = 8;
379 /* shift of arc_c for calculating both min and max arc_p */
380 static uint_t arc_p_min_shift = 4;
382 /* log2(fraction of arc to reclaim) */
383 uint_t arc_shrink_shift = 7;
385 /* percent of pagecache to reclaim arc to */
387 uint_t zfs_arc_pc_percent = 0;
391 * log2(fraction of ARC which must be free to allow growing).
392 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
393 * when reading a new block into the ARC, we will evict an equal-sized block
396 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
397 * we will still not allow it to grow.
399 uint_t arc_no_grow_shift = 5;
403 * minimum lifespan of a prefetch block in clock ticks
404 * (initialized in arc_init())
406 static uint_t arc_min_prefetch_ms;
407 static uint_t arc_min_prescient_prefetch_ms;
410 * If this percent of memory is free, don't throttle.
412 uint_t arc_lotsfree_percent = 10;
415 * The arc has filled available memory and has now warmed up.
420 * These tunables are for performance analysis.
422 uint64_t zfs_arc_max = 0;
423 uint64_t zfs_arc_min = 0;
424 uint64_t zfs_arc_meta_limit = 0;
425 uint64_t zfs_arc_meta_min = 0;
426 static uint64_t zfs_arc_dnode_limit = 0;
427 static uint_t zfs_arc_dnode_reduce_percent = 10;
428 static uint_t zfs_arc_grow_retry = 0;
429 static uint_t zfs_arc_shrink_shift = 0;
430 static uint_t zfs_arc_p_min_shift = 0;
431 uint_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
434 * ARC dirty data constraints for arc_tempreserve_space() throttle:
435 * * total dirty data limit
436 * * anon block dirty limit
437 * * each pool's anon allowance
439 static const unsigned long zfs_arc_dirty_limit_percent = 50;
440 static const unsigned long zfs_arc_anon_limit_percent = 25;
441 static const unsigned long zfs_arc_pool_dirty_percent = 20;
444 * Enable or disable compressed arc buffers.
446 int zfs_compressed_arc_enabled = B_TRUE;
449 * ARC will evict meta buffers that exceed arc_meta_limit. This
450 * tunable make arc_meta_limit adjustable for different workloads.
452 static uint64_t zfs_arc_meta_limit_percent = 75;
455 * Percentage that can be consumed by dnodes of ARC meta buffers.
457 static uint_t zfs_arc_dnode_limit_percent = 10;
460 * These tunables are Linux-specific
462 static uint64_t zfs_arc_sys_free = 0;
463 static uint_t zfs_arc_min_prefetch_ms = 0;
464 static uint_t zfs_arc_min_prescient_prefetch_ms = 0;
465 static int zfs_arc_p_dampener_disable = 1;
466 static uint_t zfs_arc_meta_prune = 10000;
467 static uint_t zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED;
468 static uint_t zfs_arc_meta_adjust_restarts = 4096;
469 static uint_t zfs_arc_lotsfree_percent = 10;
472 * Number of arc_prune threads
474 static int zfs_arc_prune_task_threads = 1;
477 arc_state_t ARC_anon;
479 arc_state_t ARC_mru_ghost;
481 arc_state_t ARC_mfu_ghost;
482 arc_state_t ARC_l2c_only;
484 arc_stats_t arc_stats = {
485 { "hits", KSTAT_DATA_UINT64 },
486 { "misses", KSTAT_DATA_UINT64 },
487 { "demand_data_hits", KSTAT_DATA_UINT64 },
488 { "demand_data_misses", KSTAT_DATA_UINT64 },
489 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
490 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
491 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
492 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
493 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
494 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
495 { "mru_hits", KSTAT_DATA_UINT64 },
496 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
497 { "mfu_hits", KSTAT_DATA_UINT64 },
498 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
499 { "deleted", KSTAT_DATA_UINT64 },
500 { "mutex_miss", KSTAT_DATA_UINT64 },
501 { "access_skip", KSTAT_DATA_UINT64 },
502 { "evict_skip", KSTAT_DATA_UINT64 },
503 { "evict_not_enough", KSTAT_DATA_UINT64 },
504 { "evict_l2_cached", KSTAT_DATA_UINT64 },
505 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
506 { "evict_l2_eligible_mfu", KSTAT_DATA_UINT64 },
507 { "evict_l2_eligible_mru", KSTAT_DATA_UINT64 },
508 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
509 { "evict_l2_skip", KSTAT_DATA_UINT64 },
510 { "hash_elements", KSTAT_DATA_UINT64 },
511 { "hash_elements_max", KSTAT_DATA_UINT64 },
512 { "hash_collisions", KSTAT_DATA_UINT64 },
513 { "hash_chains", KSTAT_DATA_UINT64 },
514 { "hash_chain_max", KSTAT_DATA_UINT64 },
515 { "p", KSTAT_DATA_UINT64 },
516 { "c", KSTAT_DATA_UINT64 },
517 { "c_min", KSTAT_DATA_UINT64 },
518 { "c_max", KSTAT_DATA_UINT64 },
519 { "size", KSTAT_DATA_UINT64 },
520 { "compressed_size", KSTAT_DATA_UINT64 },
521 { "uncompressed_size", KSTAT_DATA_UINT64 },
522 { "overhead_size", KSTAT_DATA_UINT64 },
523 { "hdr_size", KSTAT_DATA_UINT64 },
524 { "data_size", KSTAT_DATA_UINT64 },
525 { "metadata_size", KSTAT_DATA_UINT64 },
526 { "dbuf_size", KSTAT_DATA_UINT64 },
527 { "dnode_size", KSTAT_DATA_UINT64 },
528 { "bonus_size", KSTAT_DATA_UINT64 },
529 #if defined(COMPAT_FREEBSD11)
530 { "other_size", KSTAT_DATA_UINT64 },
532 { "anon_size", KSTAT_DATA_UINT64 },
533 { "anon_evictable_data", KSTAT_DATA_UINT64 },
534 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
535 { "mru_size", KSTAT_DATA_UINT64 },
536 { "mru_evictable_data", KSTAT_DATA_UINT64 },
537 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
538 { "mru_ghost_size", KSTAT_DATA_UINT64 },
539 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
540 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
541 { "mfu_size", KSTAT_DATA_UINT64 },
542 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
543 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
544 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
545 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
546 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
547 { "l2_hits", KSTAT_DATA_UINT64 },
548 { "l2_misses", KSTAT_DATA_UINT64 },
549 { "l2_prefetch_asize", KSTAT_DATA_UINT64 },
550 { "l2_mru_asize", KSTAT_DATA_UINT64 },
551 { "l2_mfu_asize", KSTAT_DATA_UINT64 },
552 { "l2_bufc_data_asize", KSTAT_DATA_UINT64 },
553 { "l2_bufc_metadata_asize", KSTAT_DATA_UINT64 },
554 { "l2_feeds", KSTAT_DATA_UINT64 },
555 { "l2_rw_clash", KSTAT_DATA_UINT64 },
556 { "l2_read_bytes", KSTAT_DATA_UINT64 },
557 { "l2_write_bytes", KSTAT_DATA_UINT64 },
558 { "l2_writes_sent", KSTAT_DATA_UINT64 },
559 { "l2_writes_done", KSTAT_DATA_UINT64 },
560 { "l2_writes_error", KSTAT_DATA_UINT64 },
561 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
562 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
563 { "l2_evict_reading", KSTAT_DATA_UINT64 },
564 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
565 { "l2_free_on_write", KSTAT_DATA_UINT64 },
566 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
567 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
568 { "l2_io_error", KSTAT_DATA_UINT64 },
569 { "l2_size", KSTAT_DATA_UINT64 },
570 { "l2_asize", KSTAT_DATA_UINT64 },
571 { "l2_hdr_size", KSTAT_DATA_UINT64 },
572 { "l2_log_blk_writes", KSTAT_DATA_UINT64 },
573 { "l2_log_blk_avg_asize", KSTAT_DATA_UINT64 },
574 { "l2_log_blk_asize", KSTAT_DATA_UINT64 },
575 { "l2_log_blk_count", KSTAT_DATA_UINT64 },
576 { "l2_data_to_meta_ratio", KSTAT_DATA_UINT64 },
577 { "l2_rebuild_success", KSTAT_DATA_UINT64 },
578 { "l2_rebuild_unsupported", KSTAT_DATA_UINT64 },
579 { "l2_rebuild_io_errors", KSTAT_DATA_UINT64 },
580 { "l2_rebuild_dh_errors", KSTAT_DATA_UINT64 },
581 { "l2_rebuild_cksum_lb_errors", KSTAT_DATA_UINT64 },
582 { "l2_rebuild_lowmem", KSTAT_DATA_UINT64 },
583 { "l2_rebuild_size", KSTAT_DATA_UINT64 },
584 { "l2_rebuild_asize", KSTAT_DATA_UINT64 },
585 { "l2_rebuild_bufs", KSTAT_DATA_UINT64 },
586 { "l2_rebuild_bufs_precached", KSTAT_DATA_UINT64 },
587 { "l2_rebuild_log_blks", KSTAT_DATA_UINT64 },
588 { "memory_throttle_count", KSTAT_DATA_UINT64 },
589 { "memory_direct_count", KSTAT_DATA_UINT64 },
590 { "memory_indirect_count", KSTAT_DATA_UINT64 },
591 { "memory_all_bytes", KSTAT_DATA_UINT64 },
592 { "memory_free_bytes", KSTAT_DATA_UINT64 },
593 { "memory_available_bytes", KSTAT_DATA_INT64 },
594 { "arc_no_grow", KSTAT_DATA_UINT64 },
595 { "arc_tempreserve", KSTAT_DATA_UINT64 },
596 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
597 { "arc_prune", KSTAT_DATA_UINT64 },
598 { "arc_meta_used", KSTAT_DATA_UINT64 },
599 { "arc_meta_limit", KSTAT_DATA_UINT64 },
600 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
601 { "arc_meta_max", KSTAT_DATA_UINT64 },
602 { "arc_meta_min", KSTAT_DATA_UINT64 },
603 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
604 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
605 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
606 { "arc_need_free", KSTAT_DATA_UINT64 },
607 { "arc_sys_free", KSTAT_DATA_UINT64 },
608 { "arc_raw_size", KSTAT_DATA_UINT64 },
609 { "cached_only_in_progress", KSTAT_DATA_UINT64 },
610 { "abd_chunk_waste_size", KSTAT_DATA_UINT64 },
615 #define ARCSTAT_MAX(stat, val) { \
617 while ((val) > (m = arc_stats.stat.value.ui64) && \
618 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
623 * We define a macro to allow ARC hits/misses to be easily broken down by
624 * two separate conditions, giving a total of four different subtypes for
625 * each of hits and misses (so eight statistics total).
627 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
630 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
632 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
636 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
638 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
643 * This macro allows us to use kstats as floating averages. Each time we
644 * update this kstat, we first factor it and the update value by
645 * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall
646 * average. This macro assumes that integer loads and stores are atomic, but
647 * is not safe for multiple writers updating the kstat in parallel (only the
648 * last writer's update will remain).
650 #define ARCSTAT_F_AVG_FACTOR 3
651 #define ARCSTAT_F_AVG(stat, value) \
653 uint64_t x = ARCSTAT(stat); \
654 x = x - x / ARCSTAT_F_AVG_FACTOR + \
655 (value) / ARCSTAT_F_AVG_FACTOR; \
659 static kstat_t *arc_ksp;
662 * There are several ARC variables that are critical to export as kstats --
663 * but we don't want to have to grovel around in the kstat whenever we wish to
664 * manipulate them. For these variables, we therefore define them to be in
665 * terms of the statistic variable. This assures that we are not introducing
666 * the possibility of inconsistency by having shadow copies of the variables,
667 * while still allowing the code to be readable.
669 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
670 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
671 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
672 /* max size for dnodes */
673 #define arc_dnode_size_limit ARCSTAT(arcstat_dnode_limit)
674 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
675 #define arc_need_free ARCSTAT(arcstat_need_free) /* waiting to be evicted */
677 hrtime_t arc_growtime;
678 list_t arc_prune_list;
679 kmutex_t arc_prune_mtx;
680 taskq_t *arc_prune_taskq;
682 #define GHOST_STATE(state) \
683 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
684 (state) == arc_l2c_only)
686 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
687 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
688 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
689 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
690 #define HDR_PRESCIENT_PREFETCH(hdr) \
691 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
692 #define HDR_COMPRESSION_ENABLED(hdr) \
693 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
695 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
696 #define HDR_L2_READING(hdr) \
697 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
698 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
699 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
700 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
701 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
702 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
703 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
704 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
706 #define HDR_ISTYPE_METADATA(hdr) \
707 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
708 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
710 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
711 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
712 #define HDR_HAS_RABD(hdr) \
713 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
714 (hdr)->b_crypt_hdr.b_rabd != NULL)
715 #define HDR_ENCRYPTED(hdr) \
716 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
717 #define HDR_AUTHENTICATED(hdr) \
718 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
720 /* For storing compression mode in b_flags */
721 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
723 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
724 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
725 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
726 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
728 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
729 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
730 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
731 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
737 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
738 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
739 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
742 * Hash table routines
745 #define BUF_LOCKS 2048
746 typedef struct buf_hash_table {
748 arc_buf_hdr_t **ht_table;
749 kmutex_t ht_locks[BUF_LOCKS] ____cacheline_aligned;
752 static buf_hash_table_t buf_hash_table;
754 #define BUF_HASH_INDEX(spa, dva, birth) \
755 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
756 #define BUF_HASH_LOCK(idx) (&buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
757 #define HDR_LOCK(hdr) \
758 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
760 uint64_t zfs_crc64_table[256];
766 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
767 #define L2ARC_HEADROOM 2 /* num of writes */
770 * If we discover during ARC scan any buffers to be compressed, we boost
771 * our headroom for the next scanning cycle by this percentage multiple.
773 #define L2ARC_HEADROOM_BOOST 200
774 #define L2ARC_FEED_SECS 1 /* caching interval secs */
775 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
778 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
779 * and each of the state has two types: data and metadata.
781 #define L2ARC_FEED_TYPES 4
783 /* L2ARC Performance Tunables */
784 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
785 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
786 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
787 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
788 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
789 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
790 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
791 int l2arc_feed_again = B_TRUE; /* turbo warmup */
792 int l2arc_norw = B_FALSE; /* no reads during writes */
793 static uint_t l2arc_meta_percent = 33; /* limit on headers size */
798 static list_t L2ARC_dev_list; /* device list */
799 static list_t *l2arc_dev_list; /* device list pointer */
800 static kmutex_t l2arc_dev_mtx; /* device list mutex */
801 static l2arc_dev_t *l2arc_dev_last; /* last device used */
802 static list_t L2ARC_free_on_write; /* free after write buf list */
803 static list_t *l2arc_free_on_write; /* free after write list ptr */
804 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
805 static uint64_t l2arc_ndev; /* number of devices */
807 typedef struct l2arc_read_callback {
808 arc_buf_hdr_t *l2rcb_hdr; /* read header */
809 blkptr_t l2rcb_bp; /* original blkptr */
810 zbookmark_phys_t l2rcb_zb; /* original bookmark */
811 int l2rcb_flags; /* original flags */
812 abd_t *l2rcb_abd; /* temporary buffer */
813 } l2arc_read_callback_t;
815 typedef struct l2arc_data_free {
816 /* protected by l2arc_free_on_write_mtx */
819 arc_buf_contents_t l2df_type;
820 list_node_t l2df_list_node;
823 typedef enum arc_fill_flags {
824 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
825 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
826 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
827 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
828 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
831 typedef enum arc_ovf_level {
832 ARC_OVF_NONE, /* ARC within target size. */
833 ARC_OVF_SOME, /* ARC is slightly overflowed. */
834 ARC_OVF_SEVERE /* ARC is severely overflowed. */
837 static kmutex_t l2arc_feed_thr_lock;
838 static kcondvar_t l2arc_feed_thr_cv;
839 static uint8_t l2arc_thread_exit;
841 static kmutex_t l2arc_rebuild_thr_lock;
842 static kcondvar_t l2arc_rebuild_thr_cv;
844 enum arc_hdr_alloc_flags {
845 ARC_HDR_ALLOC_RDATA = 0x1,
846 ARC_HDR_DO_ADAPT = 0x2,
847 ARC_HDR_USE_RESERVE = 0x4,
851 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, const void *, int);
852 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, const void *);
853 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, const void *, int);
854 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, const void *);
855 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, const void *);
856 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size,
858 static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t);
859 static void arc_hdr_alloc_abd(arc_buf_hdr_t *, int);
860 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
861 static void arc_buf_watch(arc_buf_t *);
863 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
864 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
865 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
866 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
868 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
869 static void l2arc_read_done(zio_t *);
870 static void l2arc_do_free_on_write(void);
871 static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
872 boolean_t state_only);
874 #define l2arc_hdr_arcstats_increment(hdr) \
875 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_FALSE)
876 #define l2arc_hdr_arcstats_decrement(hdr) \
877 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_FALSE)
878 #define l2arc_hdr_arcstats_increment_state(hdr) \
879 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_TRUE)
880 #define l2arc_hdr_arcstats_decrement_state(hdr) \
881 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_TRUE)
884 * l2arc_exclude_special : A zfs module parameter that controls whether buffers
885 * present on special vdevs are eligibile for caching in L2ARC. If
886 * set to 1, exclude dbufs on special vdevs from being cached to
889 int l2arc_exclude_special = 0;
892 * l2arc_mfuonly : A ZFS module parameter that controls whether only MFU
893 * metadata and data are cached from ARC into L2ARC.
895 static int l2arc_mfuonly = 0;
899 * l2arc_trim_ahead : A ZFS module parameter that controls how much ahead of
900 * the current write size (l2arc_write_max) we should TRIM if we
901 * have filled the device. It is defined as a percentage of the
902 * write size. If set to 100 we trim twice the space required to
903 * accommodate upcoming writes. A minimum of 64MB will be trimmed.
904 * It also enables TRIM of the whole L2ARC device upon creation or
905 * addition to an existing pool or if the header of the device is
906 * invalid upon importing a pool or onlining a cache device. The
907 * default is 0, which disables TRIM on L2ARC altogether as it can
908 * put significant stress on the underlying storage devices. This
909 * will vary depending of how well the specific device handles
912 static uint64_t l2arc_trim_ahead = 0;
915 * Performance tuning of L2ARC persistence:
917 * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding
918 * an L2ARC device (either at pool import or later) will attempt
919 * to rebuild L2ARC buffer contents.
920 * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls
921 * whether log blocks are written to the L2ARC device. If the L2ARC
922 * device is less than 1GB, the amount of data l2arc_evict()
923 * evicts is significant compared to the amount of restored L2ARC
924 * data. In this case do not write log blocks in L2ARC in order
925 * not to waste space.
927 static int l2arc_rebuild_enabled = B_TRUE;
928 static uint64_t l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024;
930 /* L2ARC persistence rebuild control routines. */
931 void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen);
932 static __attribute__((noreturn)) void l2arc_dev_rebuild_thread(void *arg);
933 static int l2arc_rebuild(l2arc_dev_t *dev);
935 /* L2ARC persistence read I/O routines. */
936 static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
937 static int l2arc_log_blk_read(l2arc_dev_t *dev,
938 const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
939 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
940 zio_t *this_io, zio_t **next_io);
941 static zio_t *l2arc_log_blk_fetch(vdev_t *vd,
942 const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb);
943 static void l2arc_log_blk_fetch_abort(zio_t *zio);
945 /* L2ARC persistence block restoration routines. */
946 static void l2arc_log_blk_restore(l2arc_dev_t *dev,
947 const l2arc_log_blk_phys_t *lb, uint64_t lb_asize);
948 static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
951 /* L2ARC persistence write I/O routines. */
952 static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
953 l2arc_write_callback_t *cb);
955 /* L2ARC persistence auxiliary routines. */
956 boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
957 const l2arc_log_blkptr_t *lbp);
958 static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
959 const arc_buf_hdr_t *ab);
960 boolean_t l2arc_range_check_overlap(uint64_t bottom,
961 uint64_t top, uint64_t check);
962 static void l2arc_blk_fetch_done(zio_t *zio);
963 static inline uint64_t
964 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev);
967 * We use Cityhash for this. It's fast, and has good hash properties without
968 * requiring any large static buffers.
971 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
973 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
976 #define HDR_EMPTY(hdr) \
977 ((hdr)->b_dva.dva_word[0] == 0 && \
978 (hdr)->b_dva.dva_word[1] == 0)
980 #define HDR_EMPTY_OR_LOCKED(hdr) \
981 (HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr)))
983 #define HDR_EQUAL(spa, dva, birth, hdr) \
984 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
985 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
986 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
989 buf_discard_identity(arc_buf_hdr_t *hdr)
991 hdr->b_dva.dva_word[0] = 0;
992 hdr->b_dva.dva_word[1] = 0;
996 static arc_buf_hdr_t *
997 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
999 const dva_t *dva = BP_IDENTITY(bp);
1000 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1001 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1002 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1005 mutex_enter(hash_lock);
1006 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1007 hdr = hdr->b_hash_next) {
1008 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1013 mutex_exit(hash_lock);
1019 * Insert an entry into the hash table. If there is already an element
1020 * equal to elem in the hash table, then the already existing element
1021 * will be returned and the new element will not be inserted.
1022 * Otherwise returns NULL.
1023 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1025 static arc_buf_hdr_t *
1026 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1028 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1029 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1030 arc_buf_hdr_t *fhdr;
1033 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1034 ASSERT(hdr->b_birth != 0);
1035 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1037 if (lockp != NULL) {
1039 mutex_enter(hash_lock);
1041 ASSERT(MUTEX_HELD(hash_lock));
1044 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1045 fhdr = fhdr->b_hash_next, i++) {
1046 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1050 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1051 buf_hash_table.ht_table[idx] = hdr;
1052 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1054 /* collect some hash table performance data */
1056 ARCSTAT_BUMP(arcstat_hash_collisions);
1058 ARCSTAT_BUMP(arcstat_hash_chains);
1060 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1062 uint64_t he = atomic_inc_64_nv(
1063 &arc_stats.arcstat_hash_elements.value.ui64);
1064 ARCSTAT_MAX(arcstat_hash_elements_max, he);
1070 buf_hash_remove(arc_buf_hdr_t *hdr)
1072 arc_buf_hdr_t *fhdr, **hdrp;
1073 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1075 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1076 ASSERT(HDR_IN_HASH_TABLE(hdr));
1078 hdrp = &buf_hash_table.ht_table[idx];
1079 while ((fhdr = *hdrp) != hdr) {
1080 ASSERT3P(fhdr, !=, NULL);
1081 hdrp = &fhdr->b_hash_next;
1083 *hdrp = hdr->b_hash_next;
1084 hdr->b_hash_next = NULL;
1085 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1087 /* collect some hash table performance data */
1088 atomic_dec_64(&arc_stats.arcstat_hash_elements.value.ui64);
1090 if (buf_hash_table.ht_table[idx] &&
1091 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1092 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1096 * Global data structures and functions for the buf kmem cache.
1099 static kmem_cache_t *hdr_full_cache;
1100 static kmem_cache_t *hdr_full_crypt_cache;
1101 static kmem_cache_t *hdr_l2only_cache;
1102 static kmem_cache_t *buf_cache;
1107 #if defined(_KERNEL)
1109 * Large allocations which do not require contiguous pages
1110 * should be using vmem_free() in the linux kernel\
1112 vmem_free(buf_hash_table.ht_table,
1113 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1115 kmem_free(buf_hash_table.ht_table,
1116 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1118 for (int i = 0; i < BUF_LOCKS; i++)
1119 mutex_destroy(BUF_HASH_LOCK(i));
1120 kmem_cache_destroy(hdr_full_cache);
1121 kmem_cache_destroy(hdr_full_crypt_cache);
1122 kmem_cache_destroy(hdr_l2only_cache);
1123 kmem_cache_destroy(buf_cache);
1127 * Constructor callback - called when the cache is empty
1128 * and a new buf is requested.
1131 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1133 (void) unused, (void) kmflag;
1134 arc_buf_hdr_t *hdr = vbuf;
1136 memset(hdr, 0, HDR_FULL_SIZE);
1137 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1138 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1139 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1140 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1141 list_link_init(&hdr->b_l1hdr.b_arc_node);
1142 list_link_init(&hdr->b_l2hdr.b_l2node);
1143 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1144 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1150 hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
1153 arc_buf_hdr_t *hdr = vbuf;
1155 hdr_full_cons(vbuf, unused, kmflag);
1156 memset(&hdr->b_crypt_hdr, 0, sizeof (hdr->b_crypt_hdr));
1157 arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1163 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1165 (void) unused, (void) kmflag;
1166 arc_buf_hdr_t *hdr = vbuf;
1168 memset(hdr, 0, HDR_L2ONLY_SIZE);
1169 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1175 buf_cons(void *vbuf, void *unused, int kmflag)
1177 (void) unused, (void) kmflag;
1178 arc_buf_t *buf = vbuf;
1180 memset(buf, 0, sizeof (arc_buf_t));
1181 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1182 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1188 * Destructor callback - called when a cached buf is
1189 * no longer required.
1192 hdr_full_dest(void *vbuf, void *unused)
1195 arc_buf_hdr_t *hdr = vbuf;
1197 ASSERT(HDR_EMPTY(hdr));
1198 cv_destroy(&hdr->b_l1hdr.b_cv);
1199 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1200 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1201 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1202 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1206 hdr_full_crypt_dest(void *vbuf, void *unused)
1208 (void) vbuf, (void) unused;
1210 hdr_full_dest(vbuf, unused);
1211 arc_space_return(sizeof (((arc_buf_hdr_t *)NULL)->b_crypt_hdr),
1216 hdr_l2only_dest(void *vbuf, void *unused)
1219 arc_buf_hdr_t *hdr = vbuf;
1221 ASSERT(HDR_EMPTY(hdr));
1222 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1226 buf_dest(void *vbuf, void *unused)
1229 arc_buf_t *buf = vbuf;
1231 mutex_destroy(&buf->b_evict_lock);
1232 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1238 uint64_t *ct = NULL;
1239 uint64_t hsize = 1ULL << 12;
1243 * The hash table is big enough to fill all of physical memory
1244 * with an average block size of zfs_arc_average_blocksize (default 8K).
1245 * By default, the table will take up
1246 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1248 while (hsize * zfs_arc_average_blocksize < arc_all_memory())
1251 buf_hash_table.ht_mask = hsize - 1;
1252 #if defined(_KERNEL)
1254 * Large allocations which do not require contiguous pages
1255 * should be using vmem_alloc() in the linux kernel
1257 buf_hash_table.ht_table =
1258 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
1260 buf_hash_table.ht_table =
1261 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1263 if (buf_hash_table.ht_table == NULL) {
1264 ASSERT(hsize > (1ULL << 8));
1269 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1270 0, hdr_full_cons, hdr_full_dest, NULL, NULL, NULL, 0);
1271 hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
1272 HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
1273 NULL, NULL, NULL, 0);
1274 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1275 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, NULL,
1277 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1278 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1280 for (i = 0; i < 256; i++)
1281 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1282 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1284 for (i = 0; i < BUF_LOCKS; i++)
1285 mutex_init(BUF_HASH_LOCK(i), NULL, MUTEX_DEFAULT, NULL);
1288 #define ARC_MINTIME (hz>>4) /* 62 ms */
1291 * This is the size that the buf occupies in memory. If the buf is compressed,
1292 * it will correspond to the compressed size. You should use this method of
1293 * getting the buf size unless you explicitly need the logical size.
1296 arc_buf_size(arc_buf_t *buf)
1298 return (ARC_BUF_COMPRESSED(buf) ?
1299 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1303 arc_buf_lsize(arc_buf_t *buf)
1305 return (HDR_GET_LSIZE(buf->b_hdr));
1309 * This function will return B_TRUE if the buffer is encrypted in memory.
1310 * This buffer can be decrypted by calling arc_untransform().
1313 arc_is_encrypted(arc_buf_t *buf)
1315 return (ARC_BUF_ENCRYPTED(buf) != 0);
1319 * Returns B_TRUE if the buffer represents data that has not had its MAC
1323 arc_is_unauthenticated(arc_buf_t *buf)
1325 return (HDR_NOAUTH(buf->b_hdr) != 0);
1329 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1330 uint8_t *iv, uint8_t *mac)
1332 arc_buf_hdr_t *hdr = buf->b_hdr;
1334 ASSERT(HDR_PROTECTED(hdr));
1336 memcpy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
1337 memcpy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
1338 memcpy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
1339 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1340 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1344 * Indicates how this buffer is compressed in memory. If it is not compressed
1345 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1346 * arc_untransform() as long as it is also unencrypted.
1349 arc_get_compression(arc_buf_t *buf)
1351 return (ARC_BUF_COMPRESSED(buf) ?
1352 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1356 * Return the compression algorithm used to store this data in the ARC. If ARC
1357 * compression is enabled or this is an encrypted block, this will be the same
1358 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1360 static inline enum zio_compress
1361 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1363 return (HDR_COMPRESSION_ENABLED(hdr) ?
1364 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1368 arc_get_complevel(arc_buf_t *buf)
1370 return (buf->b_hdr->b_complevel);
1373 static inline boolean_t
1374 arc_buf_is_shared(arc_buf_t *buf)
1376 boolean_t shared = (buf->b_data != NULL &&
1377 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1378 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1379 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1380 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1381 IMPLY(shared, ARC_BUF_SHARED(buf));
1382 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1385 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1386 * already being shared" requirement prevents us from doing that.
1393 * Free the checksum associated with this header. If there is no checksum, this
1397 arc_cksum_free(arc_buf_hdr_t *hdr)
1399 ASSERT(HDR_HAS_L1HDR(hdr));
1401 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1402 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1403 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1404 hdr->b_l1hdr.b_freeze_cksum = NULL;
1406 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1410 * Return true iff at least one of the bufs on hdr is not compressed.
1411 * Encrypted buffers count as compressed.
1414 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1416 ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr));
1418 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1419 if (!ARC_BUF_COMPRESSED(b)) {
1428 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1429 * matches the checksum that is stored in the hdr. If there is no checksum,
1430 * or if the buf is compressed, this is a no-op.
1433 arc_cksum_verify(arc_buf_t *buf)
1435 arc_buf_hdr_t *hdr = buf->b_hdr;
1438 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1441 if (ARC_BUF_COMPRESSED(buf))
1444 ASSERT(HDR_HAS_L1HDR(hdr));
1446 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1448 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1449 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1453 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1454 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1455 panic("buffer modified while frozen!");
1456 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1460 * This function makes the assumption that data stored in the L2ARC
1461 * will be transformed exactly as it is in the main pool. Because of
1462 * this we can verify the checksum against the reading process's bp.
1465 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1467 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1468 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1471 * Block pointers always store the checksum for the logical data.
1472 * If the block pointer has the gang bit set, then the checksum
1473 * it represents is for the reconstituted data and not for an
1474 * individual gang member. The zio pipeline, however, must be able to
1475 * determine the checksum of each of the gang constituents so it
1476 * treats the checksum comparison differently than what we need
1477 * for l2arc blocks. This prevents us from using the
1478 * zio_checksum_error() interface directly. Instead we must call the
1479 * zio_checksum_error_impl() so that we can ensure the checksum is
1480 * generated using the correct checksum algorithm and accounts for the
1481 * logical I/O size and not just a gang fragment.
1483 return (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1484 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1485 zio->io_offset, NULL) == 0);
1489 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1490 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1491 * isn't modified later on. If buf is compressed or there is already a checksum
1492 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1495 arc_cksum_compute(arc_buf_t *buf)
1497 arc_buf_hdr_t *hdr = buf->b_hdr;
1499 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1502 ASSERT(HDR_HAS_L1HDR(hdr));
1504 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1505 if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1506 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1510 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1511 ASSERT(!ARC_BUF_COMPRESSED(buf));
1512 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1514 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1515 hdr->b_l1hdr.b_freeze_cksum);
1516 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1522 arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
1524 (void) sig, (void) unused;
1525 panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr);
1530 arc_buf_unwatch(arc_buf_t *buf)
1534 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1535 PROT_READ | PROT_WRITE));
1543 arc_buf_watch(arc_buf_t *buf)
1547 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1554 static arc_buf_contents_t
1555 arc_buf_type(arc_buf_hdr_t *hdr)
1557 arc_buf_contents_t type;
1558 if (HDR_ISTYPE_METADATA(hdr)) {
1559 type = ARC_BUFC_METADATA;
1561 type = ARC_BUFC_DATA;
1563 VERIFY3U(hdr->b_type, ==, type);
1568 arc_is_metadata(arc_buf_t *buf)
1570 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1574 arc_bufc_to_flags(arc_buf_contents_t type)
1578 /* metadata field is 0 if buffer contains normal data */
1580 case ARC_BUFC_METADATA:
1581 return (ARC_FLAG_BUFC_METADATA);
1585 panic("undefined ARC buffer type!");
1586 return ((uint32_t)-1);
1590 arc_buf_thaw(arc_buf_t *buf)
1592 arc_buf_hdr_t *hdr = buf->b_hdr;
1594 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1595 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1597 arc_cksum_verify(buf);
1600 * Compressed buffers do not manipulate the b_freeze_cksum.
1602 if (ARC_BUF_COMPRESSED(buf))
1605 ASSERT(HDR_HAS_L1HDR(hdr));
1606 arc_cksum_free(hdr);
1607 arc_buf_unwatch(buf);
1611 arc_buf_freeze(arc_buf_t *buf)
1613 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1616 if (ARC_BUF_COMPRESSED(buf))
1619 ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1620 arc_cksum_compute(buf);
1624 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1625 * the following functions should be used to ensure that the flags are
1626 * updated in a thread-safe way. When manipulating the flags either
1627 * the hash_lock must be held or the hdr must be undiscoverable. This
1628 * ensures that we're not racing with any other threads when updating
1632 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1634 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1635 hdr->b_flags |= flags;
1639 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1641 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1642 hdr->b_flags &= ~flags;
1646 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1647 * done in a special way since we have to clear and set bits
1648 * at the same time. Consumers that wish to set the compression bits
1649 * must use this function to ensure that the flags are updated in
1650 * thread-safe manner.
1653 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1655 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1658 * Holes and embedded blocks will always have a psize = 0 so
1659 * we ignore the compression of the blkptr and set the
1660 * want to uncompress them. Mark them as uncompressed.
1662 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1663 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1664 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1666 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1667 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1670 HDR_SET_COMPRESS(hdr, cmp);
1671 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1675 * Looks for another buf on the same hdr which has the data decompressed, copies
1676 * from it, and returns true. If no such buf exists, returns false.
1679 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1681 arc_buf_hdr_t *hdr = buf->b_hdr;
1682 boolean_t copied = B_FALSE;
1684 ASSERT(HDR_HAS_L1HDR(hdr));
1685 ASSERT3P(buf->b_data, !=, NULL);
1686 ASSERT(!ARC_BUF_COMPRESSED(buf));
1688 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1689 from = from->b_next) {
1690 /* can't use our own data buffer */
1695 if (!ARC_BUF_COMPRESSED(from)) {
1696 memcpy(buf->b_data, from->b_data, arc_buf_size(buf));
1703 * There were no decompressed bufs, so there should not be a
1704 * checksum on the hdr either.
1706 if (zfs_flags & ZFS_DEBUG_MODIFY)
1707 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1713 * Allocates an ARC buf header that's in an evicted & L2-cached state.
1714 * This is used during l2arc reconstruction to make empty ARC buffers
1715 * which circumvent the regular disk->arc->l2arc path and instead come
1716 * into being in the reverse order, i.e. l2arc->arc.
1718 static arc_buf_hdr_t *
1719 arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
1720 dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth,
1721 enum zio_compress compress, uint8_t complevel, boolean_t protected,
1722 boolean_t prefetch, arc_state_type_t arcs_state)
1727 hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP);
1728 hdr->b_birth = birth;
1731 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR);
1732 HDR_SET_LSIZE(hdr, size);
1733 HDR_SET_PSIZE(hdr, psize);
1734 arc_hdr_set_compress(hdr, compress);
1735 hdr->b_complevel = complevel;
1737 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
1739 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
1740 hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa);
1744 hdr->b_l2hdr.b_dev = dev;
1745 hdr->b_l2hdr.b_daddr = daddr;
1746 hdr->b_l2hdr.b_arcs_state = arcs_state;
1752 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1755 arc_hdr_size(arc_buf_hdr_t *hdr)
1759 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1760 HDR_GET_PSIZE(hdr) > 0) {
1761 size = HDR_GET_PSIZE(hdr);
1763 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1764 size = HDR_GET_LSIZE(hdr);
1770 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1774 uint64_t lsize = HDR_GET_LSIZE(hdr);
1775 uint64_t psize = HDR_GET_PSIZE(hdr);
1776 void *tmpbuf = NULL;
1777 abd_t *abd = hdr->b_l1hdr.b_pabd;
1779 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1780 ASSERT(HDR_AUTHENTICATED(hdr));
1781 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1784 * The MAC is calculated on the compressed data that is stored on disk.
1785 * However, if compressed arc is disabled we will only have the
1786 * decompressed data available to us now. Compress it into a temporary
1787 * abd so we can verify the MAC. The performance overhead of this will
1788 * be relatively low, since most objects in an encrypted objset will
1789 * be encrypted (instead of authenticated) anyway.
1791 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1792 !HDR_COMPRESSION_ENABLED(hdr)) {
1793 tmpbuf = zio_buf_alloc(lsize);
1794 abd = abd_get_from_buf(tmpbuf, lsize);
1795 abd_take_ownership_of_buf(abd, B_TRUE);
1796 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1797 hdr->b_l1hdr.b_pabd, tmpbuf, lsize, hdr->b_complevel);
1798 ASSERT3U(csize, <=, psize);
1799 abd_zero_off(abd, csize, psize - csize);
1803 * Authentication is best effort. We authenticate whenever the key is
1804 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1806 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1807 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1808 ASSERT3U(lsize, ==, psize);
1809 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1810 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1812 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1813 hdr->b_crypt_hdr.b_mac);
1817 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1818 else if (ret != ENOENT)
1834 * This function will take a header that only has raw encrypted data in
1835 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1836 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1837 * also decompress the data.
1840 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1845 boolean_t no_crypt = B_FALSE;
1846 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1848 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1849 ASSERT(HDR_ENCRYPTED(hdr));
1851 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT);
1853 ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1854 B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1855 hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1856 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1861 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1862 HDR_GET_PSIZE(hdr));
1866 * If this header has disabled arc compression but the b_pabd is
1867 * compressed after decrypting it, we need to decompress the newly
1870 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1871 !HDR_COMPRESSION_ENABLED(hdr)) {
1873 * We want to make sure that we are correctly honoring the
1874 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1875 * and then loan a buffer from it, rather than allocating a
1876 * linear buffer and wrapping it in an abd later.
1878 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
1880 tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
1882 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1883 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
1884 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
1886 abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
1890 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
1891 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1892 arc_hdr_size(hdr), hdr);
1893 hdr->b_l1hdr.b_pabd = cabd;
1899 arc_hdr_free_abd(hdr, B_FALSE);
1901 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
1907 * This function is called during arc_buf_fill() to prepare the header's
1908 * abd plaintext pointer for use. This involves authenticated protected
1909 * data and decrypting encrypted data into the plaintext abd.
1912 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
1913 const zbookmark_phys_t *zb, boolean_t noauth)
1917 ASSERT(HDR_PROTECTED(hdr));
1919 if (hash_lock != NULL)
1920 mutex_enter(hash_lock);
1922 if (HDR_NOAUTH(hdr) && !noauth) {
1924 * The caller requested authenticated data but our data has
1925 * not been authenticated yet. Verify the MAC now if we can.
1927 ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
1930 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
1932 * If we only have the encrypted version of the data, but the
1933 * unencrypted version was requested we take this opportunity
1934 * to store the decrypted version in the header for future use.
1936 ret = arc_hdr_decrypt(hdr, spa, zb);
1941 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1943 if (hash_lock != NULL)
1944 mutex_exit(hash_lock);
1949 if (hash_lock != NULL)
1950 mutex_exit(hash_lock);
1956 * This function is used by the dbuf code to decrypt bonus buffers in place.
1957 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
1958 * block, so we use the hash lock here to protect against concurrent calls to
1962 arc_buf_untransform_in_place(arc_buf_t *buf)
1964 arc_buf_hdr_t *hdr = buf->b_hdr;
1966 ASSERT(HDR_ENCRYPTED(hdr));
1967 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1968 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1969 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1971 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
1973 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
1974 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1975 hdr->b_crypt_hdr.b_ebufcnt -= 1;
1979 * Given a buf that has a data buffer attached to it, this function will
1980 * efficiently fill the buf with data of the specified compression setting from
1981 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1982 * are already sharing a data buf, no copy is performed.
1984 * If the buf is marked as compressed but uncompressed data was requested, this
1985 * will allocate a new data buffer for the buf, remove that flag, and fill the
1986 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1987 * uncompressed data, and (since we haven't added support for it yet) if you
1988 * want compressed data your buf must already be marked as compressed and have
1989 * the correct-sized data buffer.
1992 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
1993 arc_fill_flags_t flags)
1996 arc_buf_hdr_t *hdr = buf->b_hdr;
1997 boolean_t hdr_compressed =
1998 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
1999 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
2000 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
2001 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2002 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
2004 ASSERT3P(buf->b_data, !=, NULL);
2005 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
2006 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2007 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
2008 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
2009 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
2010 IMPLY(encrypted, !ARC_BUF_SHARED(buf));
2013 * If the caller wanted encrypted data we just need to copy it from
2014 * b_rabd and potentially byteswap it. We won't be able to do any
2015 * further transforms on it.
2018 ASSERT(HDR_HAS_RABD(hdr));
2019 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
2020 HDR_GET_PSIZE(hdr));
2025 * Adjust encrypted and authenticated headers to accommodate
2026 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
2027 * allowed to fail decryption due to keys not being loaded
2028 * without being marked as an IO error.
2030 if (HDR_PROTECTED(hdr)) {
2031 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
2032 zb, !!(flags & ARC_FILL_NOAUTH));
2033 if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
2035 } else if (error != 0) {
2036 if (hash_lock != NULL)
2037 mutex_enter(hash_lock);
2038 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2039 if (hash_lock != NULL)
2040 mutex_exit(hash_lock);
2046 * There is a special case here for dnode blocks which are
2047 * decrypting their bonus buffers. These blocks may request to
2048 * be decrypted in-place. This is necessary because there may
2049 * be many dnodes pointing into this buffer and there is
2050 * currently no method to synchronize replacing the backing
2051 * b_data buffer and updating all of the pointers. Here we use
2052 * the hash lock to ensure there are no races. If the need
2053 * arises for other types to be decrypted in-place, they must
2054 * add handling here as well.
2056 if ((flags & ARC_FILL_IN_PLACE) != 0) {
2057 ASSERT(!hdr_compressed);
2058 ASSERT(!compressed);
2061 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
2062 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2064 if (hash_lock != NULL)
2065 mutex_enter(hash_lock);
2066 arc_buf_untransform_in_place(buf);
2067 if (hash_lock != NULL)
2068 mutex_exit(hash_lock);
2070 /* Compute the hdr's checksum if necessary */
2071 arc_cksum_compute(buf);
2077 if (hdr_compressed == compressed) {
2078 if (!arc_buf_is_shared(buf)) {
2079 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2083 ASSERT(hdr_compressed);
2084 ASSERT(!compressed);
2087 * If the buf is sharing its data with the hdr, unlink it and
2088 * allocate a new data buffer for the buf.
2090 if (arc_buf_is_shared(buf)) {
2091 ASSERT(ARC_BUF_COMPRESSED(buf));
2093 /* We need to give the buf its own b_data */
2094 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2096 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2097 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2099 /* Previously overhead was 0; just add new overhead */
2100 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2101 } else if (ARC_BUF_COMPRESSED(buf)) {
2102 /* We need to reallocate the buf's b_data */
2103 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2106 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2108 /* We increased the size of b_data; update overhead */
2109 ARCSTAT_INCR(arcstat_overhead_size,
2110 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2114 * Regardless of the buf's previous compression settings, it
2115 * should not be compressed at the end of this function.
2117 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2120 * Try copying the data from another buf which already has a
2121 * decompressed version. If that's not possible, it's time to
2122 * bite the bullet and decompress the data from the hdr.
2124 if (arc_buf_try_copy_decompressed_data(buf)) {
2125 /* Skip byteswapping and checksumming (already done) */
2128 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2129 hdr->b_l1hdr.b_pabd, buf->b_data,
2130 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr),
2134 * Absent hardware errors or software bugs, this should
2135 * be impossible, but log it anyway so we can debug it.
2139 "hdr %px, compress %d, psize %d, lsize %d",
2140 hdr, arc_hdr_get_compress(hdr),
2141 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2142 if (hash_lock != NULL)
2143 mutex_enter(hash_lock);
2144 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2145 if (hash_lock != NULL)
2146 mutex_exit(hash_lock);
2147 return (SET_ERROR(EIO));
2153 /* Byteswap the buf's data if necessary */
2154 if (bswap != DMU_BSWAP_NUMFUNCS) {
2155 ASSERT(!HDR_SHARED_DATA(hdr));
2156 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2157 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2160 /* Compute the hdr's checksum if necessary */
2161 arc_cksum_compute(buf);
2167 * If this function is being called to decrypt an encrypted buffer or verify an
2168 * authenticated one, the key must be loaded and a mapping must be made
2169 * available in the keystore via spa_keystore_create_mapping() or one of its
2173 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2177 arc_fill_flags_t flags = 0;
2180 flags |= ARC_FILL_IN_PLACE;
2182 ret = arc_buf_fill(buf, spa, zb, flags);
2183 if (ret == ECKSUM) {
2185 * Convert authentication and decryption errors to EIO
2186 * (and generate an ereport) before leaving the ARC.
2188 ret = SET_ERROR(EIO);
2189 spa_log_error(spa, zb);
2190 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2191 spa, NULL, zb, NULL, 0);
2198 * Increment the amount of evictable space in the arc_state_t's refcount.
2199 * We account for the space used by the hdr and the arc buf individually
2200 * so that we can add and remove them from the refcount individually.
2203 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2205 arc_buf_contents_t type = arc_buf_type(hdr);
2207 ASSERT(HDR_HAS_L1HDR(hdr));
2209 if (GHOST_STATE(state)) {
2210 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2211 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2212 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2213 ASSERT(!HDR_HAS_RABD(hdr));
2214 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2215 HDR_GET_LSIZE(hdr), hdr);
2219 if (hdr->b_l1hdr.b_pabd != NULL) {
2220 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2221 arc_hdr_size(hdr), hdr);
2223 if (HDR_HAS_RABD(hdr)) {
2224 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2225 HDR_GET_PSIZE(hdr), hdr);
2228 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2229 buf = buf->b_next) {
2230 if (arc_buf_is_shared(buf))
2232 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2233 arc_buf_size(buf), buf);
2238 * Decrement the amount of evictable space in the arc_state_t's refcount.
2239 * We account for the space used by the hdr and the arc buf individually
2240 * so that we can add and remove them from the refcount individually.
2243 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2245 arc_buf_contents_t type = arc_buf_type(hdr);
2247 ASSERT(HDR_HAS_L1HDR(hdr));
2249 if (GHOST_STATE(state)) {
2250 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2251 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2252 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2253 ASSERT(!HDR_HAS_RABD(hdr));
2254 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2255 HDR_GET_LSIZE(hdr), hdr);
2259 if (hdr->b_l1hdr.b_pabd != NULL) {
2260 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2261 arc_hdr_size(hdr), hdr);
2263 if (HDR_HAS_RABD(hdr)) {
2264 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2265 HDR_GET_PSIZE(hdr), hdr);
2268 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2269 buf = buf->b_next) {
2270 if (arc_buf_is_shared(buf))
2272 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2273 arc_buf_size(buf), buf);
2278 * Add a reference to this hdr indicating that someone is actively
2279 * referencing that memory. When the refcount transitions from 0 to 1,
2280 * we remove it from the respective arc_state_t list to indicate that
2281 * it is not evictable.
2284 add_reference(arc_buf_hdr_t *hdr, const void *tag)
2288 ASSERT(HDR_HAS_L1HDR(hdr));
2289 if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) {
2290 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2291 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2292 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2295 state = hdr->b_l1hdr.b_state;
2297 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2298 (state != arc_anon)) {
2299 /* We don't use the L2-only state list. */
2300 if (state != arc_l2c_only) {
2301 multilist_remove(&state->arcs_list[arc_buf_type(hdr)],
2303 arc_evictable_space_decrement(hdr, state);
2305 /* remove the prefetch flag if we get a reference */
2306 if (HDR_HAS_L2HDR(hdr))
2307 l2arc_hdr_arcstats_decrement_state(hdr);
2308 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2309 if (HDR_HAS_L2HDR(hdr))
2310 l2arc_hdr_arcstats_increment_state(hdr);
2315 * Remove a reference from this hdr. When the reference transitions from
2316 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2317 * list making it eligible for eviction.
2320 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, const void *tag)
2323 arc_state_t *state = hdr->b_l1hdr.b_state;
2325 ASSERT(HDR_HAS_L1HDR(hdr));
2326 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2327 ASSERT(!GHOST_STATE(state));
2330 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2331 * check to prevent usage of the arc_l2c_only list.
2333 if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2334 (state != arc_anon)) {
2335 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
2336 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2337 arc_evictable_space_increment(hdr, state);
2343 * Returns detailed information about a specific arc buffer. When the
2344 * state_index argument is set the function will calculate the arc header
2345 * list position for its arc state. Since this requires a linear traversal
2346 * callers are strongly encourage not to do this. However, it can be helpful
2347 * for targeted analysis so the functionality is provided.
2350 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2353 arc_buf_hdr_t *hdr = ab->b_hdr;
2354 l1arc_buf_hdr_t *l1hdr = NULL;
2355 l2arc_buf_hdr_t *l2hdr = NULL;
2356 arc_state_t *state = NULL;
2358 memset(abi, 0, sizeof (arc_buf_info_t));
2363 abi->abi_flags = hdr->b_flags;
2365 if (HDR_HAS_L1HDR(hdr)) {
2366 l1hdr = &hdr->b_l1hdr;
2367 state = l1hdr->b_state;
2369 if (HDR_HAS_L2HDR(hdr))
2370 l2hdr = &hdr->b_l2hdr;
2373 abi->abi_bufcnt = l1hdr->b_bufcnt;
2374 abi->abi_access = l1hdr->b_arc_access;
2375 abi->abi_mru_hits = l1hdr->b_mru_hits;
2376 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2377 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2378 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2379 abi->abi_holds = zfs_refcount_count(&l1hdr->b_refcnt);
2383 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2384 abi->abi_l2arc_hits = l2hdr->b_hits;
2387 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2388 abi->abi_state_contents = arc_buf_type(hdr);
2389 abi->abi_size = arc_hdr_size(hdr);
2393 * Move the supplied buffer to the indicated state. The hash lock
2394 * for the buffer must be held by the caller.
2397 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2398 kmutex_t *hash_lock)
2400 arc_state_t *old_state;
2403 boolean_t update_old, update_new;
2404 arc_buf_contents_t buftype = arc_buf_type(hdr);
2407 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2408 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2409 * L1 hdr doesn't always exist when we change state to arc_anon before
2410 * destroying a header, in which case reallocating to add the L1 hdr is
2413 if (HDR_HAS_L1HDR(hdr)) {
2414 old_state = hdr->b_l1hdr.b_state;
2415 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2416 bufcnt = hdr->b_l1hdr.b_bufcnt;
2417 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
2420 old_state = arc_l2c_only;
2423 update_old = B_FALSE;
2425 update_new = update_old;
2427 ASSERT(MUTEX_HELD(hash_lock));
2428 ASSERT3P(new_state, !=, old_state);
2429 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2430 ASSERT(old_state != arc_anon || bufcnt <= 1);
2433 * If this buffer is evictable, transfer it from the
2434 * old state list to the new state list.
2437 if (old_state != arc_anon && old_state != arc_l2c_only) {
2438 ASSERT(HDR_HAS_L1HDR(hdr));
2439 multilist_remove(&old_state->arcs_list[buftype], hdr);
2441 if (GHOST_STATE(old_state)) {
2443 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2444 update_old = B_TRUE;
2446 arc_evictable_space_decrement(hdr, old_state);
2448 if (new_state != arc_anon && new_state != arc_l2c_only) {
2450 * An L1 header always exists here, since if we're
2451 * moving to some L1-cached state (i.e. not l2c_only or
2452 * anonymous), we realloc the header to add an L1hdr
2455 ASSERT(HDR_HAS_L1HDR(hdr));
2456 multilist_insert(&new_state->arcs_list[buftype], hdr);
2458 if (GHOST_STATE(new_state)) {
2460 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2461 update_new = B_TRUE;
2463 arc_evictable_space_increment(hdr, new_state);
2467 ASSERT(!HDR_EMPTY(hdr));
2468 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2469 buf_hash_remove(hdr);
2471 /* adjust state sizes (ignore arc_l2c_only) */
2473 if (update_new && new_state != arc_l2c_only) {
2474 ASSERT(HDR_HAS_L1HDR(hdr));
2475 if (GHOST_STATE(new_state)) {
2479 * When moving a header to a ghost state, we first
2480 * remove all arc buffers. Thus, we'll have a
2481 * bufcnt of zero, and no arc buffer to use for
2482 * the reference. As a result, we use the arc
2483 * header pointer for the reference.
2485 (void) zfs_refcount_add_many(&new_state->arcs_size,
2486 HDR_GET_LSIZE(hdr), hdr);
2487 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2488 ASSERT(!HDR_HAS_RABD(hdr));
2490 uint32_t buffers = 0;
2493 * Each individual buffer holds a unique reference,
2494 * thus we must remove each of these references one
2497 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2498 buf = buf->b_next) {
2499 ASSERT3U(bufcnt, !=, 0);
2503 * When the arc_buf_t is sharing the data
2504 * block with the hdr, the owner of the
2505 * reference belongs to the hdr. Only
2506 * add to the refcount if the arc_buf_t is
2509 if (arc_buf_is_shared(buf))
2512 (void) zfs_refcount_add_many(
2513 &new_state->arcs_size,
2514 arc_buf_size(buf), buf);
2516 ASSERT3U(bufcnt, ==, buffers);
2518 if (hdr->b_l1hdr.b_pabd != NULL) {
2519 (void) zfs_refcount_add_many(
2520 &new_state->arcs_size,
2521 arc_hdr_size(hdr), hdr);
2524 if (HDR_HAS_RABD(hdr)) {
2525 (void) zfs_refcount_add_many(
2526 &new_state->arcs_size,
2527 HDR_GET_PSIZE(hdr), hdr);
2532 if (update_old && old_state != arc_l2c_only) {
2533 ASSERT(HDR_HAS_L1HDR(hdr));
2534 if (GHOST_STATE(old_state)) {
2536 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2537 ASSERT(!HDR_HAS_RABD(hdr));
2540 * When moving a header off of a ghost state,
2541 * the header will not contain any arc buffers.
2542 * We use the arc header pointer for the reference
2543 * which is exactly what we did when we put the
2544 * header on the ghost state.
2547 (void) zfs_refcount_remove_many(&old_state->arcs_size,
2548 HDR_GET_LSIZE(hdr), hdr);
2550 uint32_t buffers = 0;
2553 * Each individual buffer holds a unique reference,
2554 * thus we must remove each of these references one
2557 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2558 buf = buf->b_next) {
2559 ASSERT3U(bufcnt, !=, 0);
2563 * When the arc_buf_t is sharing the data
2564 * block with the hdr, the owner of the
2565 * reference belongs to the hdr. Only
2566 * add to the refcount if the arc_buf_t is
2569 if (arc_buf_is_shared(buf))
2572 (void) zfs_refcount_remove_many(
2573 &old_state->arcs_size, arc_buf_size(buf),
2576 ASSERT3U(bufcnt, ==, buffers);
2577 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2580 if (hdr->b_l1hdr.b_pabd != NULL) {
2581 (void) zfs_refcount_remove_many(
2582 &old_state->arcs_size, arc_hdr_size(hdr),
2586 if (HDR_HAS_RABD(hdr)) {
2587 (void) zfs_refcount_remove_many(
2588 &old_state->arcs_size, HDR_GET_PSIZE(hdr),
2594 if (HDR_HAS_L1HDR(hdr)) {
2595 hdr->b_l1hdr.b_state = new_state;
2597 if (HDR_HAS_L2HDR(hdr) && new_state != arc_l2c_only) {
2598 l2arc_hdr_arcstats_decrement_state(hdr);
2599 hdr->b_l2hdr.b_arcs_state = new_state->arcs_state;
2600 l2arc_hdr_arcstats_increment_state(hdr);
2606 arc_space_consume(uint64_t space, arc_space_type_t type)
2608 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2613 case ARC_SPACE_DATA:
2614 ARCSTAT_INCR(arcstat_data_size, space);
2616 case ARC_SPACE_META:
2617 ARCSTAT_INCR(arcstat_metadata_size, space);
2619 case ARC_SPACE_BONUS:
2620 ARCSTAT_INCR(arcstat_bonus_size, space);
2622 case ARC_SPACE_DNODE:
2623 aggsum_add(&arc_sums.arcstat_dnode_size, space);
2625 case ARC_SPACE_DBUF:
2626 ARCSTAT_INCR(arcstat_dbuf_size, space);
2628 case ARC_SPACE_HDRS:
2629 ARCSTAT_INCR(arcstat_hdr_size, space);
2631 case ARC_SPACE_L2HDRS:
2632 aggsum_add(&arc_sums.arcstat_l2_hdr_size, space);
2634 case ARC_SPACE_ABD_CHUNK_WASTE:
2636 * Note: this includes space wasted by all scatter ABD's, not
2637 * just those allocated by the ARC. But the vast majority of
2638 * scatter ABD's come from the ARC, because other users are
2641 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, space);
2645 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE)
2646 aggsum_add(&arc_sums.arcstat_meta_used, space);
2648 aggsum_add(&arc_sums.arcstat_size, space);
2652 arc_space_return(uint64_t space, arc_space_type_t type)
2654 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2659 case ARC_SPACE_DATA:
2660 ARCSTAT_INCR(arcstat_data_size, -space);
2662 case ARC_SPACE_META:
2663 ARCSTAT_INCR(arcstat_metadata_size, -space);
2665 case ARC_SPACE_BONUS:
2666 ARCSTAT_INCR(arcstat_bonus_size, -space);
2668 case ARC_SPACE_DNODE:
2669 aggsum_add(&arc_sums.arcstat_dnode_size, -space);
2671 case ARC_SPACE_DBUF:
2672 ARCSTAT_INCR(arcstat_dbuf_size, -space);
2674 case ARC_SPACE_HDRS:
2675 ARCSTAT_INCR(arcstat_hdr_size, -space);
2677 case ARC_SPACE_L2HDRS:
2678 aggsum_add(&arc_sums.arcstat_l2_hdr_size, -space);
2680 case ARC_SPACE_ABD_CHUNK_WASTE:
2681 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, -space);
2685 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE) {
2686 ASSERT(aggsum_compare(&arc_sums.arcstat_meta_used,
2688 ARCSTAT_MAX(arcstat_meta_max,
2689 aggsum_upper_bound(&arc_sums.arcstat_meta_used));
2690 aggsum_add(&arc_sums.arcstat_meta_used, -space);
2693 ASSERT(aggsum_compare(&arc_sums.arcstat_size, space) >= 0);
2694 aggsum_add(&arc_sums.arcstat_size, -space);
2698 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2699 * with the hdr's b_pabd.
2702 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2705 * The criteria for sharing a hdr's data are:
2706 * 1. the buffer is not encrypted
2707 * 2. the hdr's compression matches the buf's compression
2708 * 3. the hdr doesn't need to be byteswapped
2709 * 4. the hdr isn't already being shared
2710 * 5. the buf is either compressed or it is the last buf in the hdr list
2712 * Criterion #5 maintains the invariant that shared uncompressed
2713 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2714 * might ask, "if a compressed buf is allocated first, won't that be the
2715 * last thing in the list?", but in that case it's impossible to create
2716 * a shared uncompressed buf anyway (because the hdr must be compressed
2717 * to have the compressed buf). You might also think that #3 is
2718 * sufficient to make this guarantee, however it's possible
2719 * (specifically in the rare L2ARC write race mentioned in
2720 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2721 * is shareable, but wasn't at the time of its allocation. Rather than
2722 * allow a new shared uncompressed buf to be created and then shuffle
2723 * the list around to make it the last element, this simply disallows
2724 * sharing if the new buf isn't the first to be added.
2726 ASSERT3P(buf->b_hdr, ==, hdr);
2727 boolean_t hdr_compressed =
2728 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF;
2729 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2730 return (!ARC_BUF_ENCRYPTED(buf) &&
2731 buf_compressed == hdr_compressed &&
2732 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2733 !HDR_SHARED_DATA(hdr) &&
2734 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2738 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2739 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2740 * copy was made successfully, or an error code otherwise.
2743 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2744 const void *tag, boolean_t encrypted, boolean_t compressed,
2745 boolean_t noauth, boolean_t fill, arc_buf_t **ret)
2748 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2750 ASSERT(HDR_HAS_L1HDR(hdr));
2751 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2752 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2753 hdr->b_type == ARC_BUFC_METADATA);
2754 ASSERT3P(ret, !=, NULL);
2755 ASSERT3P(*ret, ==, NULL);
2756 IMPLY(encrypted, compressed);
2758 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2761 buf->b_next = hdr->b_l1hdr.b_buf;
2764 add_reference(hdr, tag);
2767 * We're about to change the hdr's b_flags. We must either
2768 * hold the hash_lock or be undiscoverable.
2770 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2773 * Only honor requests for compressed bufs if the hdr is actually
2774 * compressed. This must be overridden if the buffer is encrypted since
2775 * encrypted buffers cannot be decompressed.
2778 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2779 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2780 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2781 } else if (compressed &&
2782 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2783 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2784 flags |= ARC_FILL_COMPRESSED;
2789 flags |= ARC_FILL_NOAUTH;
2793 * If the hdr's data can be shared then we share the data buffer and
2794 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2795 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2796 * buffer to store the buf's data.
2798 * There are two additional restrictions here because we're sharing
2799 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2800 * actively involved in an L2ARC write, because if this buf is used by
2801 * an arc_write() then the hdr's data buffer will be released when the
2802 * write completes, even though the L2ARC write might still be using it.
2803 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2804 * need to be ABD-aware. It must be allocated via
2805 * zio_[data_]buf_alloc(), not as a page, because we need to be able
2806 * to abd_release_ownership_of_buf(), which isn't allowed on "linear
2807 * page" buffers because the ABD code needs to handle freeing them
2810 boolean_t can_share = arc_can_share(hdr, buf) &&
2811 !HDR_L2_WRITING(hdr) &&
2812 hdr->b_l1hdr.b_pabd != NULL &&
2813 abd_is_linear(hdr->b_l1hdr.b_pabd) &&
2814 !abd_is_linear_page(hdr->b_l1hdr.b_pabd);
2816 /* Set up b_data and sharing */
2818 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2819 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2820 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2823 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2824 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2826 VERIFY3P(buf->b_data, !=, NULL);
2828 hdr->b_l1hdr.b_buf = buf;
2829 hdr->b_l1hdr.b_bufcnt += 1;
2831 hdr->b_crypt_hdr.b_ebufcnt += 1;
2834 * If the user wants the data from the hdr, we need to either copy or
2835 * decompress the data.
2838 ASSERT3P(zb, !=, NULL);
2839 return (arc_buf_fill(buf, spa, zb, flags));
2845 static const char *arc_onloan_tag = "onloan";
2848 arc_loaned_bytes_update(int64_t delta)
2850 atomic_add_64(&arc_loaned_bytes, delta);
2852 /* assert that it did not wrap around */
2853 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2857 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2858 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2859 * buffers must be returned to the arc before they can be used by the DMU or
2863 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2865 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2866 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2868 arc_loaned_bytes_update(arc_buf_size(buf));
2874 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2875 enum zio_compress compression_type, uint8_t complevel)
2877 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2878 psize, lsize, compression_type, complevel);
2880 arc_loaned_bytes_update(arc_buf_size(buf));
2886 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2887 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2888 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2889 enum zio_compress compression_type, uint8_t complevel)
2891 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2892 byteorder, salt, iv, mac, ot, psize, lsize, compression_type,
2895 atomic_add_64(&arc_loaned_bytes, psize);
2901 * Return a loaned arc buffer to the arc.
2904 arc_return_buf(arc_buf_t *buf, const void *tag)
2906 arc_buf_hdr_t *hdr = buf->b_hdr;
2908 ASSERT3P(buf->b_data, !=, NULL);
2909 ASSERT(HDR_HAS_L1HDR(hdr));
2910 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2911 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2913 arc_loaned_bytes_update(-arc_buf_size(buf));
2916 /* Detach an arc_buf from a dbuf (tag) */
2918 arc_loan_inuse_buf(arc_buf_t *buf, const 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, arc_onloan_tag);
2925 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2927 arc_loaned_bytes_update(arc_buf_size(buf));
2931 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2933 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2936 df->l2df_size = size;
2937 df->l2df_type = type;
2938 mutex_enter(&l2arc_free_on_write_mtx);
2939 list_insert_head(l2arc_free_on_write, df);
2940 mutex_exit(&l2arc_free_on_write_mtx);
2944 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
2946 arc_state_t *state = hdr->b_l1hdr.b_state;
2947 arc_buf_contents_t type = arc_buf_type(hdr);
2948 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
2950 /* protected by hash lock, if in the hash table */
2951 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2952 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2953 ASSERT(state != arc_anon && state != arc_l2c_only);
2955 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2958 (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr);
2959 if (type == ARC_BUFC_METADATA) {
2960 arc_space_return(size, ARC_SPACE_META);
2962 ASSERT(type == ARC_BUFC_DATA);
2963 arc_space_return(size, ARC_SPACE_DATA);
2967 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
2969 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2974 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2975 * data buffer, we transfer the refcount ownership to the hdr and update
2976 * the appropriate kstats.
2979 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2981 ASSERT(arc_can_share(hdr, buf));
2982 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2983 ASSERT(!ARC_BUF_ENCRYPTED(buf));
2984 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2987 * Start sharing the data buffer. We transfer the
2988 * refcount ownership to the hdr since it always owns
2989 * the refcount whenever an arc_buf_t is shared.
2991 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
2992 arc_hdr_size(hdr), buf, hdr);
2993 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2994 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2995 HDR_ISTYPE_METADATA(hdr));
2996 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2997 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3000 * Since we've transferred ownership to the hdr we need
3001 * to increment its compressed and uncompressed kstats and
3002 * decrement the overhead size.
3004 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3005 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3006 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3010 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3012 ASSERT(arc_buf_is_shared(buf));
3013 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3014 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3017 * We are no longer sharing this buffer so we need
3018 * to transfer its ownership to the rightful owner.
3020 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
3021 arc_hdr_size(hdr), hdr, buf);
3022 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3023 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3024 abd_free(hdr->b_l1hdr.b_pabd);
3025 hdr->b_l1hdr.b_pabd = NULL;
3026 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3029 * Since the buffer is no longer shared between
3030 * the arc buf and the hdr, count it as overhead.
3032 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3033 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3034 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3038 * Remove an arc_buf_t from the hdr's buf list and return the last
3039 * arc_buf_t on the list. If no buffers remain on the list then return
3043 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3045 ASSERT(HDR_HAS_L1HDR(hdr));
3046 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3048 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3049 arc_buf_t *lastbuf = NULL;
3052 * Remove the buf from the hdr list and locate the last
3053 * remaining buffer on the list.
3055 while (*bufp != NULL) {
3057 *bufp = buf->b_next;
3060 * If we've removed a buffer in the middle of
3061 * the list then update the lastbuf and update
3064 if (*bufp != NULL) {
3066 bufp = &(*bufp)->b_next;
3070 ASSERT3P(lastbuf, !=, buf);
3071 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3072 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3073 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3079 * Free up buf->b_data and pull the arc_buf_t off of the arc_buf_hdr_t's
3083 arc_buf_destroy_impl(arc_buf_t *buf)
3085 arc_buf_hdr_t *hdr = buf->b_hdr;
3088 * Free up the data associated with the buf but only if we're not
3089 * sharing this with the hdr. If we are sharing it with the hdr, the
3090 * hdr is responsible for doing the free.
3092 if (buf->b_data != NULL) {
3094 * We're about to change the hdr's b_flags. We must either
3095 * hold the hash_lock or be undiscoverable.
3097 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3099 arc_cksum_verify(buf);
3100 arc_buf_unwatch(buf);
3102 if (arc_buf_is_shared(buf)) {
3103 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3105 uint64_t size = arc_buf_size(buf);
3106 arc_free_data_buf(hdr, buf->b_data, size, buf);
3107 ARCSTAT_INCR(arcstat_overhead_size, -size);
3111 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3112 hdr->b_l1hdr.b_bufcnt -= 1;
3114 if (ARC_BUF_ENCRYPTED(buf)) {
3115 hdr->b_crypt_hdr.b_ebufcnt -= 1;
3118 * If we have no more encrypted buffers and we've
3119 * already gotten a copy of the decrypted data we can
3120 * free b_rabd to save some space.
3122 if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
3123 HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
3124 !HDR_IO_IN_PROGRESS(hdr)) {
3125 arc_hdr_free_abd(hdr, B_TRUE);
3130 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3132 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3134 * If the current arc_buf_t is sharing its data buffer with the
3135 * hdr, then reassign the hdr's b_pabd to share it with the new
3136 * buffer at the end of the list. The shared buffer is always
3137 * the last one on the hdr's buffer list.
3139 * There is an equivalent case for compressed bufs, but since
3140 * they aren't guaranteed to be the last buf in the list and
3141 * that is an exceedingly rare case, we just allow that space be
3142 * wasted temporarily. We must also be careful not to share
3143 * encrypted buffers, since they cannot be shared.
3145 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
3146 /* Only one buf can be shared at once */
3147 VERIFY(!arc_buf_is_shared(lastbuf));
3148 /* hdr is uncompressed so can't have compressed buf */
3149 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3151 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3152 arc_hdr_free_abd(hdr, B_FALSE);
3155 * We must setup a new shared block between the
3156 * last buffer and the hdr. The data would have
3157 * been allocated by the arc buf so we need to transfer
3158 * ownership to the hdr since it's now being shared.
3160 arc_share_buf(hdr, lastbuf);
3162 } else if (HDR_SHARED_DATA(hdr)) {
3164 * Uncompressed shared buffers are always at the end
3165 * of the list. Compressed buffers don't have the
3166 * same requirements. This makes it hard to
3167 * simply assert that the lastbuf is shared so
3168 * we rely on the hdr's compression flags to determine
3169 * if we have a compressed, shared buffer.
3171 ASSERT3P(lastbuf, !=, NULL);
3172 ASSERT(arc_buf_is_shared(lastbuf) ||
3173 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3177 * Free the checksum if we're removing the last uncompressed buf from
3180 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3181 arc_cksum_free(hdr);
3184 /* clean up the buf */
3186 kmem_cache_free(buf_cache, buf);
3190 arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, int alloc_flags)
3193 boolean_t alloc_rdata = ((alloc_flags & ARC_HDR_ALLOC_RDATA) != 0);
3195 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3196 ASSERT(HDR_HAS_L1HDR(hdr));
3197 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3198 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3201 size = HDR_GET_PSIZE(hdr);
3202 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3203 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr,
3205 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3206 ARCSTAT_INCR(arcstat_raw_size, size);
3208 size = arc_hdr_size(hdr);
3209 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3210 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr,
3212 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3215 ARCSTAT_INCR(arcstat_compressed_size, size);
3216 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3220 arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3222 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3224 ASSERT(HDR_HAS_L1HDR(hdr));
3225 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3226 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3229 * If the hdr is currently being written to the l2arc then
3230 * we defer freeing the data by adding it to the l2arc_free_on_write
3231 * list. The l2arc will free the data once it's finished
3232 * writing it to the l2arc device.
3234 if (HDR_L2_WRITING(hdr)) {
3235 arc_hdr_free_on_write(hdr, free_rdata);
3236 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3237 } else if (free_rdata) {
3238 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3240 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr);
3244 hdr->b_crypt_hdr.b_rabd = NULL;
3245 ARCSTAT_INCR(arcstat_raw_size, -size);
3247 hdr->b_l1hdr.b_pabd = NULL;
3250 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3251 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3253 ARCSTAT_INCR(arcstat_compressed_size, -size);
3254 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3258 * Allocate empty anonymous ARC header. The header will get its identity
3259 * assigned and buffers attached later as part of read or write operations.
3261 * In case of read arc_read() assigns header its identify (b_dva + b_birth),
3262 * inserts it into ARC hash to become globally visible and allocates physical
3263 * (b_pabd) or raw (b_rabd) ABD buffer to read into from disk. On disk read
3264 * completion arc_read_done() allocates ARC buffer(s) as needed, potentially
3265 * sharing one of them with the physical ABD buffer.
3267 * In case of write arc_alloc_buf() allocates ARC buffer to be filled with
3268 * data. Then after compression and/or encryption arc_write_ready() allocates
3269 * and fills (or potentially shares) physical (b_pabd) or raw (b_rabd) ABD
3270 * buffer. On disk write completion arc_write_done() assigns the header its
3271 * new identity (b_dva + b_birth) and inserts into ARC hash.
3273 * In case of partial overwrite the old data is read first as described. Then
3274 * arc_release() either allocates new anonymous ARC header and moves the ARC
3275 * buffer to it, or reuses the old ARC header by discarding its identity and
3276 * removing it from ARC hash. After buffer modification normal write process
3277 * follows as described.
3279 static arc_buf_hdr_t *
3280 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3281 boolean_t protected, enum zio_compress compression_type, uint8_t complevel,
3282 arc_buf_contents_t type)
3286 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3288 hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
3290 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3293 ASSERT(HDR_EMPTY(hdr));
3294 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3295 HDR_SET_PSIZE(hdr, psize);
3296 HDR_SET_LSIZE(hdr, lsize);
3300 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3301 arc_hdr_set_compress(hdr, compression_type);
3302 hdr->b_complevel = complevel;
3304 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3306 hdr->b_l1hdr.b_state = arc_anon;
3307 hdr->b_l1hdr.b_arc_access = 0;
3308 hdr->b_l1hdr.b_mru_hits = 0;
3309 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3310 hdr->b_l1hdr.b_mfu_hits = 0;
3311 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3312 hdr->b_l1hdr.b_bufcnt = 0;
3313 hdr->b_l1hdr.b_buf = NULL;
3315 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3321 * Transition between the two allocation states for the arc_buf_hdr struct.
3322 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3323 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3324 * version is used when a cache buffer is only in the L2ARC in order to reduce
3327 static arc_buf_hdr_t *
3328 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3330 ASSERT(HDR_HAS_L2HDR(hdr));
3332 arc_buf_hdr_t *nhdr;
3333 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3335 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3336 (old == hdr_l2only_cache && new == hdr_full_cache));
3339 * if the caller wanted a new full header and the header is to be
3340 * encrypted we will actually allocate the header from the full crypt
3341 * cache instead. The same applies to freeing from the old cache.
3343 if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
3344 new = hdr_full_crypt_cache;
3345 if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
3346 old = hdr_full_crypt_cache;
3348 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3350 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3351 buf_hash_remove(hdr);
3353 memcpy(nhdr, hdr, HDR_L2ONLY_SIZE);
3355 if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
3356 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3358 * arc_access and arc_change_state need to be aware that a
3359 * header has just come out of L2ARC, so we set its state to
3360 * l2c_only even though it's about to change.
3362 nhdr->b_l1hdr.b_state = arc_l2c_only;
3364 /* Verify previous threads set to NULL before freeing */
3365 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3366 ASSERT(!HDR_HAS_RABD(hdr));
3368 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3369 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3370 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3373 * If we've reached here, We must have been called from
3374 * arc_evict_hdr(), as such we should have already been
3375 * removed from any ghost list we were previously on
3376 * (which protects us from racing with arc_evict_state),
3377 * thus no locking is needed during this check.
3379 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3382 * A buffer must not be moved into the arc_l2c_only
3383 * state if it's not finished being written out to the
3384 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3385 * might try to be accessed, even though it was removed.
3387 VERIFY(!HDR_L2_WRITING(hdr));
3388 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3389 ASSERT(!HDR_HAS_RABD(hdr));
3391 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3394 * The header has been reallocated so we need to re-insert it into any
3397 (void) buf_hash_insert(nhdr, NULL);
3399 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3401 mutex_enter(&dev->l2ad_mtx);
3404 * We must place the realloc'ed header back into the list at
3405 * the same spot. Otherwise, if it's placed earlier in the list,
3406 * l2arc_write_buffers() could find it during the function's
3407 * write phase, and try to write it out to the l2arc.
3409 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3410 list_remove(&dev->l2ad_buflist, hdr);
3412 mutex_exit(&dev->l2ad_mtx);
3415 * Since we're using the pointer address as the tag when
3416 * incrementing and decrementing the l2ad_alloc refcount, we
3417 * must remove the old pointer (that we're about to destroy) and
3418 * add the new pointer to the refcount. Otherwise we'd remove
3419 * the wrong pointer address when calling arc_hdr_destroy() later.
3422 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
3423 arc_hdr_size(hdr), hdr);
3424 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
3425 arc_hdr_size(nhdr), nhdr);
3427 buf_discard_identity(hdr);
3428 kmem_cache_free(old, hdr);
3434 * This function allows an L1 header to be reallocated as a crypt
3435 * header and vice versa. If we are going to a crypt header, the
3436 * new fields will be zeroed out.
3438 static arc_buf_hdr_t *
3439 arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
3441 arc_buf_hdr_t *nhdr;
3443 kmem_cache_t *ncache, *ocache;
3446 * This function requires that hdr is in the arc_anon state.
3447 * Therefore it won't have any L2ARC data for us to worry
3450 ASSERT(HDR_HAS_L1HDR(hdr));
3451 ASSERT(!HDR_HAS_L2HDR(hdr));
3452 ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
3453 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3454 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3455 ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node));
3456 ASSERT3P(hdr->b_hash_next, ==, NULL);
3459 ncache = hdr_full_crypt_cache;
3460 ocache = hdr_full_cache;
3462 ncache = hdr_full_cache;
3463 ocache = hdr_full_crypt_cache;
3466 nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);
3469 * Copy all members that aren't locks or condvars to the new header.
3470 * No lists are pointing to us (as we asserted above), so we don't
3471 * need to worry about the list nodes.
3473 nhdr->b_dva = hdr->b_dva;
3474 nhdr->b_birth = hdr->b_birth;
3475 nhdr->b_type = hdr->b_type;
3476 nhdr->b_flags = hdr->b_flags;
3477 nhdr->b_psize = hdr->b_psize;
3478 nhdr->b_lsize = hdr->b_lsize;
3479 nhdr->b_spa = hdr->b_spa;
3480 nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
3481 nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
3482 nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
3483 nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
3484 nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
3485 nhdr->b_l1hdr.b_mru_hits = hdr->b_l1hdr.b_mru_hits;
3486 nhdr->b_l1hdr.b_mru_ghost_hits = hdr->b_l1hdr.b_mru_ghost_hits;
3487 nhdr->b_l1hdr.b_mfu_hits = hdr->b_l1hdr.b_mfu_hits;
3488 nhdr->b_l1hdr.b_mfu_ghost_hits = hdr->b_l1hdr.b_mfu_ghost_hits;
3489 nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
3490 nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
3493 * This zfs_refcount_add() exists only to ensure that the individual
3494 * arc buffers always point to a header that is referenced, avoiding
3495 * a small race condition that could trigger ASSERTs.
3497 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
3498 nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
3499 for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) {
3500 mutex_enter(&buf->b_evict_lock);
3502 mutex_exit(&buf->b_evict_lock);
3505 zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
3506 (void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
3507 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3510 arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
3512 arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
3515 /* unset all members of the original hdr */
3516 memset(&hdr->b_dva, 0, sizeof (dva_t));
3518 hdr->b_type = ARC_BUFC_INVALID;
3523 hdr->b_l1hdr.b_freeze_cksum = NULL;
3524 hdr->b_l1hdr.b_buf = NULL;
3525 hdr->b_l1hdr.b_bufcnt = 0;
3526 hdr->b_l1hdr.b_byteswap = 0;
3527 hdr->b_l1hdr.b_state = NULL;
3528 hdr->b_l1hdr.b_arc_access = 0;
3529 hdr->b_l1hdr.b_mru_hits = 0;
3530 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3531 hdr->b_l1hdr.b_mfu_hits = 0;
3532 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3533 hdr->b_l1hdr.b_acb = NULL;
3534 hdr->b_l1hdr.b_pabd = NULL;
3536 if (ocache == hdr_full_crypt_cache) {
3537 ASSERT(!HDR_HAS_RABD(hdr));
3538 hdr->b_crypt_hdr.b_ot = DMU_OT_NONE;
3539 hdr->b_crypt_hdr.b_ebufcnt = 0;
3540 hdr->b_crypt_hdr.b_dsobj = 0;
3541 memset(hdr->b_crypt_hdr.b_salt, 0, ZIO_DATA_SALT_LEN);
3542 memset(hdr->b_crypt_hdr.b_iv, 0, ZIO_DATA_IV_LEN);
3543 memset(hdr->b_crypt_hdr.b_mac, 0, ZIO_DATA_MAC_LEN);
3546 buf_discard_identity(hdr);
3547 kmem_cache_free(ocache, hdr);
3553 * This function is used by the send / receive code to convert a newly
3554 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3555 * is also used to allow the root objset block to be updated without altering
3556 * its embedded MACs. Both block types will always be uncompressed so we do not
3557 * have to worry about compression type or psize.
3560 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3561 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3564 arc_buf_hdr_t *hdr = buf->b_hdr;
3566 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3567 ASSERT(HDR_HAS_L1HDR(hdr));
3568 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3570 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3571 if (!HDR_PROTECTED(hdr))
3572 hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
3573 hdr->b_crypt_hdr.b_dsobj = dsobj;
3574 hdr->b_crypt_hdr.b_ot = ot;
3575 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3576 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3577 if (!arc_hdr_has_uncompressed_buf(hdr))
3578 arc_cksum_free(hdr);
3581 memcpy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
3583 memcpy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
3585 memcpy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
3589 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3590 * The buf is returned thawed since we expect the consumer to modify it.
3593 arc_alloc_buf(spa_t *spa, const void *tag, arc_buf_contents_t type,
3596 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3597 B_FALSE, ZIO_COMPRESS_OFF, 0, type);
3599 arc_buf_t *buf = NULL;
3600 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3601 B_FALSE, B_FALSE, &buf));
3608 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3609 * for bufs containing metadata.
3612 arc_alloc_compressed_buf(spa_t *spa, const void *tag, uint64_t psize,
3613 uint64_t lsize, enum zio_compress compression_type, uint8_t complevel)
3615 ASSERT3U(lsize, >, 0);
3616 ASSERT3U(lsize, >=, psize);
3617 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3618 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3620 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3621 B_FALSE, compression_type, complevel, ARC_BUFC_DATA);
3623 arc_buf_t *buf = NULL;
3624 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3625 B_TRUE, B_FALSE, B_FALSE, &buf));
3627 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3630 * To ensure that the hdr has the correct data in it if we call
3631 * arc_untransform() on this buf before it's been written to disk,
3632 * it's easiest if we just set up sharing between the buf and the hdr.
3634 arc_share_buf(hdr, buf);
3640 arc_alloc_raw_buf(spa_t *spa, const void *tag, uint64_t dsobj,
3641 boolean_t byteorder, const uint8_t *salt, const uint8_t *iv,
3642 const uint8_t *mac, dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3643 enum zio_compress compression_type, uint8_t complevel)
3647 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3648 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3650 ASSERT3U(lsize, >, 0);
3651 ASSERT3U(lsize, >=, psize);
3652 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3653 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3655 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3656 compression_type, complevel, type);
3658 hdr->b_crypt_hdr.b_dsobj = dsobj;
3659 hdr->b_crypt_hdr.b_ot = ot;
3660 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3661 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3662 memcpy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
3663 memcpy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
3664 memcpy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
3667 * This buffer will be considered encrypted even if the ot is not an
3668 * encrypted type. It will become authenticated instead in
3669 * arc_write_ready().
3672 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3673 B_FALSE, B_FALSE, &buf));
3675 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3681 l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
3682 boolean_t state_only)
3684 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3685 l2arc_dev_t *dev = l2hdr->b_dev;
3686 uint64_t lsize = HDR_GET_LSIZE(hdr);
3687 uint64_t psize = HDR_GET_PSIZE(hdr);
3688 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3689 arc_buf_contents_t type = hdr->b_type;
3704 /* If the buffer is a prefetch, count it as such. */
3705 if (HDR_PREFETCH(hdr)) {
3706 ARCSTAT_INCR(arcstat_l2_prefetch_asize, asize_s);
3709 * We use the value stored in the L2 header upon initial
3710 * caching in L2ARC. This value will be updated in case
3711 * an MRU/MRU_ghost buffer transitions to MFU but the L2ARC
3712 * metadata (log entry) cannot currently be updated. Having
3713 * the ARC state in the L2 header solves the problem of a
3714 * possibly absent L1 header (apparent in buffers restored
3715 * from persistent L2ARC).
3717 switch (hdr->b_l2hdr.b_arcs_state) {
3718 case ARC_STATE_MRU_GHOST:
3720 ARCSTAT_INCR(arcstat_l2_mru_asize, asize_s);
3722 case ARC_STATE_MFU_GHOST:
3724 ARCSTAT_INCR(arcstat_l2_mfu_asize, asize_s);
3734 ARCSTAT_INCR(arcstat_l2_psize, psize_s);
3735 ARCSTAT_INCR(arcstat_l2_lsize, lsize_s);
3739 ARCSTAT_INCR(arcstat_l2_bufc_data_asize, asize_s);
3741 case ARC_BUFC_METADATA:
3742 ARCSTAT_INCR(arcstat_l2_bufc_metadata_asize, asize_s);
3751 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3753 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3754 l2arc_dev_t *dev = l2hdr->b_dev;
3755 uint64_t psize = HDR_GET_PSIZE(hdr);
3756 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3758 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3759 ASSERT(HDR_HAS_L2HDR(hdr));
3761 list_remove(&dev->l2ad_buflist, hdr);
3763 l2arc_hdr_arcstats_decrement(hdr);
3764 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3766 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3768 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3772 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3774 if (HDR_HAS_L1HDR(hdr)) {
3775 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3776 hdr->b_l1hdr.b_bufcnt > 0);
3777 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3778 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3780 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3781 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3783 if (HDR_HAS_L2HDR(hdr)) {
3784 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3785 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3788 mutex_enter(&dev->l2ad_mtx);
3791 * Even though we checked this conditional above, we
3792 * need to check this again now that we have the
3793 * l2ad_mtx. This is because we could be racing with
3794 * another thread calling l2arc_evict() which might have
3795 * destroyed this header's L2 portion as we were waiting
3796 * to acquire the l2ad_mtx. If that happens, we don't
3797 * want to re-destroy the header's L2 portion.
3799 if (HDR_HAS_L2HDR(hdr)) {
3801 if (!HDR_EMPTY(hdr))
3802 buf_discard_identity(hdr);
3804 arc_hdr_l2hdr_destroy(hdr);
3808 mutex_exit(&dev->l2ad_mtx);
3812 * The header's identify can only be safely discarded once it is no
3813 * longer discoverable. This requires removing it from the hash table
3814 * and the l2arc header list. After this point the hash lock can not
3815 * be used to protect the header.
3817 if (!HDR_EMPTY(hdr))
3818 buf_discard_identity(hdr);
3820 if (HDR_HAS_L1HDR(hdr)) {
3821 arc_cksum_free(hdr);
3823 while (hdr->b_l1hdr.b_buf != NULL)
3824 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3826 if (hdr->b_l1hdr.b_pabd != NULL)
3827 arc_hdr_free_abd(hdr, B_FALSE);
3829 if (HDR_HAS_RABD(hdr))
3830 arc_hdr_free_abd(hdr, B_TRUE);
3833 ASSERT3P(hdr->b_hash_next, ==, NULL);
3834 if (HDR_HAS_L1HDR(hdr)) {
3835 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3836 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3838 if (!HDR_PROTECTED(hdr)) {
3839 kmem_cache_free(hdr_full_cache, hdr);
3841 kmem_cache_free(hdr_full_crypt_cache, hdr);
3844 kmem_cache_free(hdr_l2only_cache, hdr);
3849 arc_buf_destroy(arc_buf_t *buf, const void *tag)
3851 arc_buf_hdr_t *hdr = buf->b_hdr;
3853 if (hdr->b_l1hdr.b_state == arc_anon) {
3854 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3855 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3856 VERIFY0(remove_reference(hdr, NULL, tag));
3857 arc_hdr_destroy(hdr);
3861 kmutex_t *hash_lock = HDR_LOCK(hdr);
3862 mutex_enter(hash_lock);
3864 ASSERT3P(hdr, ==, buf->b_hdr);
3865 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3866 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3867 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3868 ASSERT3P(buf->b_data, !=, NULL);
3870 (void) remove_reference(hdr, hash_lock, tag);
3871 arc_buf_destroy_impl(buf);
3872 mutex_exit(hash_lock);
3876 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3877 * state of the header is dependent on its state prior to entering this
3878 * function. The following transitions are possible:
3880 * - arc_mru -> arc_mru_ghost
3881 * - arc_mfu -> arc_mfu_ghost
3882 * - arc_mru_ghost -> arc_l2c_only
3883 * - arc_mru_ghost -> deleted
3884 * - arc_mfu_ghost -> arc_l2c_only
3885 * - arc_mfu_ghost -> deleted
3887 * Return total size of evicted data buffers for eviction progress tracking.
3888 * When evicting from ghost states return logical buffer size to make eviction
3889 * progress at the same (or at least comparable) rate as from non-ghost states.
3891 * Return *real_evicted for actual ARC size reduction to wake up threads
3892 * waiting for it. For non-ghost states it includes size of evicted data
3893 * buffers (the headers are not freed there). For ghost states it includes
3894 * only the evicted headers size.
3897 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, uint64_t *real_evicted)
3899 arc_state_t *evicted_state, *state;
3900 int64_t bytes_evicted = 0;
3901 uint_t min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3902 arc_min_prescient_prefetch_ms : arc_min_prefetch_ms;
3904 ASSERT(MUTEX_HELD(hash_lock));
3905 ASSERT(HDR_HAS_L1HDR(hdr));
3908 state = hdr->b_l1hdr.b_state;
3909 if (GHOST_STATE(state)) {
3910 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3911 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3914 * l2arc_write_buffers() relies on a header's L1 portion
3915 * (i.e. its b_pabd field) during it's write phase.
3916 * Thus, we cannot push a header onto the arc_l2c_only
3917 * state (removing its L1 piece) until the header is
3918 * done being written to the l2arc.
3920 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3921 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3922 return (bytes_evicted);
3925 ARCSTAT_BUMP(arcstat_deleted);
3926 bytes_evicted += HDR_GET_LSIZE(hdr);
3928 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3930 if (HDR_HAS_L2HDR(hdr)) {
3931 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3932 ASSERT(!HDR_HAS_RABD(hdr));
3934 * This buffer is cached on the 2nd Level ARC;
3935 * don't destroy the header.
3937 arc_change_state(arc_l2c_only, hdr, hash_lock);
3939 * dropping from L1+L2 cached to L2-only,
3940 * realloc to remove the L1 header.
3942 (void) arc_hdr_realloc(hdr, hdr_full_cache,
3944 *real_evicted += HDR_FULL_SIZE - HDR_L2ONLY_SIZE;
3946 arc_change_state(arc_anon, hdr, hash_lock);
3947 arc_hdr_destroy(hdr);
3948 *real_evicted += HDR_FULL_SIZE;
3950 return (bytes_evicted);
3953 ASSERT(state == arc_mru || state == arc_mfu);
3954 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3956 /* prefetch buffers have a minimum lifespan */
3957 if (HDR_IO_IN_PROGRESS(hdr) ||
3958 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3959 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3960 MSEC_TO_TICK(min_lifetime))) {
3961 ARCSTAT_BUMP(arcstat_evict_skip);
3962 return (bytes_evicted);
3965 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3966 while (hdr->b_l1hdr.b_buf) {
3967 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3968 if (!mutex_tryenter(&buf->b_evict_lock)) {
3969 ARCSTAT_BUMP(arcstat_mutex_miss);
3972 if (buf->b_data != NULL) {
3973 bytes_evicted += HDR_GET_LSIZE(hdr);
3974 *real_evicted += HDR_GET_LSIZE(hdr);
3976 mutex_exit(&buf->b_evict_lock);
3977 arc_buf_destroy_impl(buf);
3980 if (HDR_HAS_L2HDR(hdr)) {
3981 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3983 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3984 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3985 HDR_GET_LSIZE(hdr));
3987 switch (state->arcs_state) {
3990 arcstat_evict_l2_eligible_mru,
3991 HDR_GET_LSIZE(hdr));
3995 arcstat_evict_l2_eligible_mfu,
3996 HDR_GET_LSIZE(hdr));
4002 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
4003 HDR_GET_LSIZE(hdr));
4007 if (hdr->b_l1hdr.b_bufcnt == 0) {
4008 arc_cksum_free(hdr);
4010 bytes_evicted += arc_hdr_size(hdr);
4011 *real_evicted += arc_hdr_size(hdr);
4014 * If this hdr is being evicted and has a compressed
4015 * buffer then we discard it here before we change states.
4016 * This ensures that the accounting is updated correctly
4017 * in arc_free_data_impl().
4019 if (hdr->b_l1hdr.b_pabd != NULL)
4020 arc_hdr_free_abd(hdr, B_FALSE);
4022 if (HDR_HAS_RABD(hdr))
4023 arc_hdr_free_abd(hdr, B_TRUE);
4025 arc_change_state(evicted_state, hdr, hash_lock);
4026 ASSERT(HDR_IN_HASH_TABLE(hdr));
4027 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
4028 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
4031 return (bytes_evicted);
4035 arc_set_need_free(void)
4037 ASSERT(MUTEX_HELD(&arc_evict_lock));
4038 int64_t remaining = arc_free_memory() - arc_sys_free / 2;
4039 arc_evict_waiter_t *aw = list_tail(&arc_evict_waiters);
4041 arc_need_free = MAX(-remaining, 0);
4044 MAX(-remaining, (int64_t)(aw->aew_count - arc_evict_count));
4049 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
4050 uint64_t spa, uint64_t bytes)
4052 multilist_sublist_t *mls;
4053 uint64_t bytes_evicted = 0, real_evicted = 0;
4055 kmutex_t *hash_lock;
4056 uint_t evict_count = zfs_arc_evict_batch_limit;
4058 ASSERT3P(marker, !=, NULL);
4060 mls = multilist_sublist_lock(ml, idx);
4062 for (hdr = multilist_sublist_prev(mls, marker); likely(hdr != NULL);
4063 hdr = multilist_sublist_prev(mls, marker)) {
4064 if ((evict_count == 0) || (bytes_evicted >= bytes))
4068 * To keep our iteration location, move the marker
4069 * forward. Since we're not holding hdr's hash lock, we
4070 * must be very careful and not remove 'hdr' from the
4071 * sublist. Otherwise, other consumers might mistake the
4072 * 'hdr' as not being on a sublist when they call the
4073 * multilist_link_active() function (they all rely on
4074 * the hash lock protecting concurrent insertions and
4075 * removals). multilist_sublist_move_forward() was
4076 * specifically implemented to ensure this is the case
4077 * (only 'marker' will be removed and re-inserted).
4079 multilist_sublist_move_forward(mls, marker);
4082 * The only case where the b_spa field should ever be
4083 * zero, is the marker headers inserted by
4084 * arc_evict_state(). It's possible for multiple threads
4085 * to be calling arc_evict_state() concurrently (e.g.
4086 * dsl_pool_close() and zio_inject_fault()), so we must
4087 * skip any markers we see from these other threads.
4089 if (hdr->b_spa == 0)
4092 /* we're only interested in evicting buffers of a certain spa */
4093 if (spa != 0 && hdr->b_spa != spa) {
4094 ARCSTAT_BUMP(arcstat_evict_skip);
4098 hash_lock = HDR_LOCK(hdr);
4101 * We aren't calling this function from any code path
4102 * that would already be holding a hash lock, so we're
4103 * asserting on this assumption to be defensive in case
4104 * this ever changes. Without this check, it would be
4105 * possible to incorrectly increment arcstat_mutex_miss
4106 * below (e.g. if the code changed such that we called
4107 * this function with a hash lock held).
4109 ASSERT(!MUTEX_HELD(hash_lock));
4111 if (mutex_tryenter(hash_lock)) {
4113 uint64_t evicted = arc_evict_hdr(hdr, hash_lock,
4115 mutex_exit(hash_lock);
4117 bytes_evicted += evicted;
4118 real_evicted += revicted;
4121 * If evicted is zero, arc_evict_hdr() must have
4122 * decided to skip this header, don't increment
4123 * evict_count in this case.
4129 ARCSTAT_BUMP(arcstat_mutex_miss);
4133 multilist_sublist_unlock(mls);
4136 * Increment the count of evicted bytes, and wake up any threads that
4137 * are waiting for the count to reach this value. Since the list is
4138 * ordered by ascending aew_count, we pop off the beginning of the
4139 * list until we reach the end, or a waiter that's past the current
4140 * "count". Doing this outside the loop reduces the number of times
4141 * we need to acquire the global arc_evict_lock.
4143 * Only wake when there's sufficient free memory in the system
4144 * (specifically, arc_sys_free/2, which by default is a bit more than
4145 * 1/64th of RAM). See the comments in arc_wait_for_eviction().
4147 mutex_enter(&arc_evict_lock);
4148 arc_evict_count += real_evicted;
4150 if (arc_free_memory() > arc_sys_free / 2) {
4151 arc_evict_waiter_t *aw;
4152 while ((aw = list_head(&arc_evict_waiters)) != NULL &&
4153 aw->aew_count <= arc_evict_count) {
4154 list_remove(&arc_evict_waiters, aw);
4155 cv_broadcast(&aw->aew_cv);
4158 arc_set_need_free();
4159 mutex_exit(&arc_evict_lock);
4162 * If the ARC size is reduced from arc_c_max to arc_c_min (especially
4163 * if the average cached block is small), eviction can be on-CPU for
4164 * many seconds. To ensure that other threads that may be bound to
4165 * this CPU are able to make progress, make a voluntary preemption
4168 kpreempt(KPREEMPT_SYNC);
4170 return (bytes_evicted);
4174 * Allocate an array of buffer headers used as placeholders during arc state
4177 static arc_buf_hdr_t **
4178 arc_state_alloc_markers(int count)
4180 arc_buf_hdr_t **markers;
4182 markers = kmem_zalloc(sizeof (*markers) * count, KM_SLEEP);
4183 for (int i = 0; i < count; i++) {
4184 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
4187 * A b_spa of 0 is used to indicate that this header is
4188 * a marker. This fact is used in arc_evict_type() and
4189 * arc_evict_state_impl().
4191 markers[i]->b_spa = 0;
4198 arc_state_free_markers(arc_buf_hdr_t **markers, int count)
4200 for (int i = 0; i < count; i++)
4201 kmem_cache_free(hdr_full_cache, markers[i]);
4202 kmem_free(markers, sizeof (*markers) * count);
4206 * Evict buffers from the given arc state, until we've removed the
4207 * specified number of bytes. Move the removed buffers to the
4208 * appropriate evict state.
4210 * This function makes a "best effort". It skips over any buffers
4211 * it can't get a hash_lock on, and so, may not catch all candidates.
4212 * It may also return without evicting as much space as requested.
4214 * If bytes is specified using the special value ARC_EVICT_ALL, this
4215 * will evict all available (i.e. unlocked and evictable) buffers from
4216 * the given arc state; which is used by arc_flush().
4219 arc_evict_state(arc_state_t *state, uint64_t spa, uint64_t bytes,
4220 arc_buf_contents_t type)
4222 uint64_t total_evicted = 0;
4223 multilist_t *ml = &state->arcs_list[type];
4225 arc_buf_hdr_t **markers;
4227 num_sublists = multilist_get_num_sublists(ml);
4230 * If we've tried to evict from each sublist, made some
4231 * progress, but still have not hit the target number of bytes
4232 * to evict, we want to keep trying. The markers allow us to
4233 * pick up where we left off for each individual sublist, rather
4234 * than starting from the tail each time.
4236 if (zthr_iscurthread(arc_evict_zthr)) {
4237 markers = arc_state_evict_markers;
4238 ASSERT3S(num_sublists, <=, arc_state_evict_marker_count);
4240 markers = arc_state_alloc_markers(num_sublists);
4242 for (int i = 0; i < num_sublists; i++) {
4243 multilist_sublist_t *mls;
4245 mls = multilist_sublist_lock(ml, i);
4246 multilist_sublist_insert_tail(mls, markers[i]);
4247 multilist_sublist_unlock(mls);
4251 * While we haven't hit our target number of bytes to evict, or
4252 * we're evicting all available buffers.
4254 while (total_evicted < bytes) {
4255 int sublist_idx = multilist_get_random_index(ml);
4256 uint64_t scan_evicted = 0;
4259 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
4260 * Request that 10% of the LRUs be scanned by the superblock
4263 if (type == ARC_BUFC_DATA && aggsum_compare(
4264 &arc_sums.arcstat_dnode_size, arc_dnode_size_limit) > 0) {
4265 arc_prune_async((aggsum_upper_bound(
4266 &arc_sums.arcstat_dnode_size) -
4267 arc_dnode_size_limit) / sizeof (dnode_t) /
4268 zfs_arc_dnode_reduce_percent);
4272 * Start eviction using a randomly selected sublist,
4273 * this is to try and evenly balance eviction across all
4274 * sublists. Always starting at the same sublist
4275 * (e.g. index 0) would cause evictions to favor certain
4276 * sublists over others.
4278 for (int i = 0; i < num_sublists; i++) {
4279 uint64_t bytes_remaining;
4280 uint64_t bytes_evicted;
4282 if (total_evicted < bytes)
4283 bytes_remaining = bytes - total_evicted;
4287 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4288 markers[sublist_idx], spa, bytes_remaining);
4290 scan_evicted += bytes_evicted;
4291 total_evicted += bytes_evicted;
4293 /* we've reached the end, wrap to the beginning */
4294 if (++sublist_idx >= num_sublists)
4299 * If we didn't evict anything during this scan, we have
4300 * no reason to believe we'll evict more during another
4301 * scan, so break the loop.
4303 if (scan_evicted == 0) {
4304 /* This isn't possible, let's make that obvious */
4305 ASSERT3S(bytes, !=, 0);
4308 * When bytes is ARC_EVICT_ALL, the only way to
4309 * break the loop is when scan_evicted is zero.
4310 * In that case, we actually have evicted enough,
4311 * so we don't want to increment the kstat.
4313 if (bytes != ARC_EVICT_ALL) {
4314 ASSERT3S(total_evicted, <, bytes);
4315 ARCSTAT_BUMP(arcstat_evict_not_enough);
4322 for (int i = 0; i < num_sublists; i++) {
4323 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4324 multilist_sublist_remove(mls, markers[i]);
4325 multilist_sublist_unlock(mls);
4327 if (markers != arc_state_evict_markers)
4328 arc_state_free_markers(markers, num_sublists);
4330 return (total_evicted);
4334 * Flush all "evictable" data of the given type from the arc state
4335 * specified. This will not evict any "active" buffers (i.e. referenced).
4337 * When 'retry' is set to B_FALSE, the function will make a single pass
4338 * over the state and evict any buffers that it can. Since it doesn't
4339 * continually retry the eviction, it might end up leaving some buffers
4340 * in the ARC due to lock misses.
4342 * When 'retry' is set to B_TRUE, the function will continually retry the
4343 * eviction until *all* evictable buffers have been removed from the
4344 * state. As a result, if concurrent insertions into the state are
4345 * allowed (e.g. if the ARC isn't shutting down), this function might
4346 * wind up in an infinite loop, continually trying to evict buffers.
4349 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4352 uint64_t evicted = 0;
4354 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4355 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4365 * Evict the specified number of bytes from the state specified,
4366 * restricting eviction to the spa and type given. This function
4367 * prevents us from trying to evict more from a state's list than
4368 * is "evictable", and to skip evicting altogether when passed a
4369 * negative value for "bytes". In contrast, arc_evict_state() will
4370 * evict everything it can, when passed a negative value for "bytes".
4373 arc_evict_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4374 arc_buf_contents_t type)
4378 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4379 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4381 return (arc_evict_state(state, spa, delta, type));
4388 * The goal of this function is to evict enough meta data buffers from the
4389 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4390 * more complicated than it appears because it is common for data buffers
4391 * to have holds on meta data buffers. In addition, dnode meta data buffers
4392 * will be held by the dnodes in the block preventing them from being freed.
4393 * This means we can't simply traverse the ARC and expect to always find
4394 * enough unheld meta data buffer to release.
4396 * Therefore, this function has been updated to make alternating passes
4397 * over the ARC releasing data buffers and then newly unheld meta data
4398 * buffers. This ensures forward progress is maintained and meta_used
4399 * will decrease. Normally this is sufficient, but if required the ARC
4400 * will call the registered prune callbacks causing dentry and inodes to
4401 * be dropped from the VFS cache. This will make dnode meta data buffers
4402 * available for reclaim.
4405 arc_evict_meta_balanced(uint64_t meta_used)
4407 int64_t delta, adjustmnt;
4408 uint64_t total_evicted = 0, prune = 0;
4409 arc_buf_contents_t type = ARC_BUFC_DATA;
4410 uint_t restarts = zfs_arc_meta_adjust_restarts;
4414 * This slightly differs than the way we evict from the mru in
4415 * arc_evict because we don't have a "target" value (i.e. no
4416 * "meta" arc_p). As a result, I think we can completely
4417 * cannibalize the metadata in the MRU before we evict the
4418 * metadata from the MFU. I think we probably need to implement a
4419 * "metadata arc_p" value to do this properly.
4421 adjustmnt = meta_used - arc_meta_limit;
4423 if (adjustmnt > 0 &&
4424 zfs_refcount_count(&arc_mru->arcs_esize[type]) > 0) {
4425 delta = MIN(zfs_refcount_count(&arc_mru->arcs_esize[type]),
4427 total_evicted += arc_evict_impl(arc_mru, 0, delta, type);
4432 * We can't afford to recalculate adjustmnt here. If we do,
4433 * new metadata buffers can sneak into the MRU or ANON lists,
4434 * thus penalize the MFU metadata. Although the fudge factor is
4435 * small, it has been empirically shown to be significant for
4436 * certain workloads (e.g. creating many empty directories). As
4437 * such, we use the original calculation for adjustmnt, and
4438 * simply decrement the amount of data evicted from the MRU.
4441 if (adjustmnt > 0 &&
4442 zfs_refcount_count(&arc_mfu->arcs_esize[type]) > 0) {
4443 delta = MIN(zfs_refcount_count(&arc_mfu->arcs_esize[type]),
4445 total_evicted += arc_evict_impl(arc_mfu, 0, delta, type);
4448 adjustmnt = meta_used - arc_meta_limit;
4450 if (adjustmnt > 0 &&
4451 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) {
4452 delta = MIN(adjustmnt,
4453 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]));
4454 total_evicted += arc_evict_impl(arc_mru_ghost, 0, delta, type);
4458 if (adjustmnt > 0 &&
4459 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) {
4460 delta = MIN(adjustmnt,
4461 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]));
4462 total_evicted += arc_evict_impl(arc_mfu_ghost, 0, delta, type);
4466 * If after attempting to make the requested adjustment to the ARC
4467 * the meta limit is still being exceeded then request that the
4468 * higher layers drop some cached objects which have holds on ARC
4469 * meta buffers. Requests to the upper layers will be made with
4470 * increasingly large scan sizes until the ARC is below the limit.
4472 if (meta_used > arc_meta_limit || arc_available_memory() < 0) {
4473 if (type == ARC_BUFC_DATA) {
4474 type = ARC_BUFC_METADATA;
4476 type = ARC_BUFC_DATA;
4478 if (zfs_arc_meta_prune) {
4479 prune += zfs_arc_meta_prune;
4480 arc_prune_async(prune);
4489 return (total_evicted);
4493 * Evict metadata buffers from the cache, such that arcstat_meta_used is
4494 * capped by the arc_meta_limit tunable.
4497 arc_evict_meta_only(uint64_t meta_used)
4499 uint64_t total_evicted = 0;
4503 * If we're over the meta limit, we want to evict enough
4504 * metadata to get back under the meta limit. We don't want to
4505 * evict so much that we drop the MRU below arc_p, though. If
4506 * we're over the meta limit more than we're over arc_p, we
4507 * evict some from the MRU here, and some from the MFU below.
4509 target = MIN((int64_t)(meta_used - arc_meta_limit),
4510 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4511 zfs_refcount_count(&arc_mru->arcs_size) - arc_p));
4513 total_evicted += arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4516 * Similar to the above, we want to evict enough bytes to get us
4517 * below the meta limit, but not so much as to drop us below the
4518 * space allotted to the MFU (which is defined as arc_c - arc_p).
4520 target = MIN((int64_t)(meta_used - arc_meta_limit),
4521 (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) -
4524 total_evicted += arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4526 return (total_evicted);
4530 arc_evict_meta(uint64_t meta_used)
4532 if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY)
4533 return (arc_evict_meta_only(meta_used));
4535 return (arc_evict_meta_balanced(meta_used));
4539 * Return the type of the oldest buffer in the given arc state
4541 * This function will select a random sublist of type ARC_BUFC_DATA and
4542 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4543 * is compared, and the type which contains the "older" buffer will be
4546 static arc_buf_contents_t
4547 arc_evict_type(arc_state_t *state)
4549 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
4550 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
4551 int data_idx = multilist_get_random_index(data_ml);
4552 int meta_idx = multilist_get_random_index(meta_ml);
4553 multilist_sublist_t *data_mls;
4554 multilist_sublist_t *meta_mls;
4555 arc_buf_contents_t type;
4556 arc_buf_hdr_t *data_hdr;
4557 arc_buf_hdr_t *meta_hdr;
4560 * We keep the sublist lock until we're finished, to prevent
4561 * the headers from being destroyed via arc_evict_state().
4563 data_mls = multilist_sublist_lock(data_ml, data_idx);
4564 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4567 * These two loops are to ensure we skip any markers that
4568 * might be at the tail of the lists due to arc_evict_state().
4571 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4572 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4573 if (data_hdr->b_spa != 0)
4577 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4578 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4579 if (meta_hdr->b_spa != 0)
4583 if (data_hdr == NULL && meta_hdr == NULL) {
4584 type = ARC_BUFC_DATA;
4585 } else if (data_hdr == NULL) {
4586 ASSERT3P(meta_hdr, !=, NULL);
4587 type = ARC_BUFC_METADATA;
4588 } else if (meta_hdr == NULL) {
4589 ASSERT3P(data_hdr, !=, NULL);
4590 type = ARC_BUFC_DATA;
4592 ASSERT3P(data_hdr, !=, NULL);
4593 ASSERT3P(meta_hdr, !=, NULL);
4595 /* The headers can't be on the sublist without an L1 header */
4596 ASSERT(HDR_HAS_L1HDR(data_hdr));
4597 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4599 if (data_hdr->b_l1hdr.b_arc_access <
4600 meta_hdr->b_l1hdr.b_arc_access) {
4601 type = ARC_BUFC_DATA;
4603 type = ARC_BUFC_METADATA;
4607 multilist_sublist_unlock(meta_mls);
4608 multilist_sublist_unlock(data_mls);
4614 * Evict buffers from the cache, such that arcstat_size is capped by arc_c.
4619 uint64_t total_evicted = 0;
4622 uint64_t asize = aggsum_value(&arc_sums.arcstat_size);
4623 uint64_t ameta = aggsum_value(&arc_sums.arcstat_meta_used);
4626 * If we're over arc_meta_limit, we want to correct that before
4627 * potentially evicting data buffers below.
4629 total_evicted += arc_evict_meta(ameta);
4634 * If we're over the target cache size, we want to evict enough
4635 * from the list to get back to our target size. We don't want
4636 * to evict too much from the MRU, such that it drops below
4637 * arc_p. So, if we're over our target cache size more than
4638 * the MRU is over arc_p, we'll evict enough to get back to
4639 * arc_p here, and then evict more from the MFU below.
4641 target = MIN((int64_t)(asize - arc_c),
4642 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4643 zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4646 * If we're below arc_meta_min, always prefer to evict data.
4647 * Otherwise, try to satisfy the requested number of bytes to
4648 * evict from the type which contains older buffers; in an
4649 * effort to keep newer buffers in the cache regardless of their
4650 * type. If we cannot satisfy the number of bytes from this
4651 * type, spill over into the next type.
4653 if (arc_evict_type(arc_mru) == ARC_BUFC_METADATA &&
4654 ameta > arc_meta_min) {
4655 bytes = arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4656 total_evicted += bytes;
4659 * If we couldn't evict our target number of bytes from
4660 * metadata, we try to get the rest from data.
4665 arc_evict_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4667 bytes = arc_evict_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4668 total_evicted += bytes;
4671 * If we couldn't evict our target number of bytes from
4672 * data, we try to get the rest from metadata.
4677 arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4681 * Re-sum ARC stats after the first round of evictions.
4683 asize = aggsum_value(&arc_sums.arcstat_size);
4684 ameta = aggsum_value(&arc_sums.arcstat_meta_used);
4690 * Now that we've tried to evict enough from the MRU to get its
4691 * size back to arc_p, if we're still above the target cache
4692 * size, we evict the rest from the MFU.
4694 target = asize - arc_c;
4696 if (arc_evict_type(arc_mfu) == ARC_BUFC_METADATA &&
4697 ameta > arc_meta_min) {
4698 bytes = arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4699 total_evicted += bytes;
4702 * If we couldn't evict our target number of bytes from
4703 * metadata, we try to get the rest from data.
4708 arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4710 bytes = arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4711 total_evicted += bytes;
4714 * If we couldn't evict our target number of bytes from
4715 * data, we try to get the rest from data.
4720 arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4724 * Adjust ghost lists
4726 * In addition to the above, the ARC also defines target values
4727 * for the ghost lists. The sum of the mru list and mru ghost
4728 * list should never exceed the target size of the cache, and
4729 * the sum of the mru list, mfu list, mru ghost list, and mfu
4730 * ghost list should never exceed twice the target size of the
4731 * cache. The following logic enforces these limits on the ghost
4732 * caches, and evicts from them as needed.
4734 target = zfs_refcount_count(&arc_mru->arcs_size) +
4735 zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4737 bytes = arc_evict_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4738 total_evicted += bytes;
4743 arc_evict_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4746 * We assume the sum of the mru list and mfu list is less than
4747 * or equal to arc_c (we enforced this above), which means we
4748 * can use the simpler of the two equations below:
4750 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4751 * mru ghost + mfu ghost <= arc_c
4753 target = zfs_refcount_count(&arc_mru_ghost->arcs_size) +
4754 zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4756 bytes = arc_evict_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4757 total_evicted += bytes;
4762 arc_evict_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4764 return (total_evicted);
4768 arc_flush(spa_t *spa, boolean_t retry)
4773 * If retry is B_TRUE, a spa must not be specified since we have
4774 * no good way to determine if all of a spa's buffers have been
4775 * evicted from an arc state.
4777 ASSERT(!retry || spa == 0);
4780 guid = spa_load_guid(spa);
4782 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4783 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4785 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4786 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4788 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4789 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4791 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4792 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4796 arc_reduce_target_size(int64_t to_free)
4798 uint64_t asize = aggsum_value(&arc_sums.arcstat_size);
4801 * All callers want the ARC to actually evict (at least) this much
4802 * memory. Therefore we reduce from the lower of the current size and
4803 * the target size. This way, even if arc_c is much higher than
4804 * arc_size (as can be the case after many calls to arc_freed(), we will
4805 * immediately have arc_c < arc_size and therefore the arc_evict_zthr
4808 uint64_t c = MIN(arc_c, asize);
4810 if (c > to_free && c - to_free > arc_c_min) {
4811 arc_c = c - to_free;
4812 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4814 arc_p = (arc_c >> 1);
4815 ASSERT(arc_c >= arc_c_min);
4816 ASSERT((int64_t)arc_p >= 0);
4821 if (asize > arc_c) {
4822 /* See comment in arc_evict_cb_check() on why lock+flag */
4823 mutex_enter(&arc_evict_lock);
4824 arc_evict_needed = B_TRUE;
4825 mutex_exit(&arc_evict_lock);
4826 zthr_wakeup(arc_evict_zthr);
4831 * Determine if the system is under memory pressure and is asking
4832 * to reclaim memory. A return value of B_TRUE indicates that the system
4833 * is under memory pressure and that the arc should adjust accordingly.
4836 arc_reclaim_needed(void)
4838 return (arc_available_memory() < 0);
4842 arc_kmem_reap_soon(void)
4845 kmem_cache_t *prev_cache = NULL;
4846 kmem_cache_t *prev_data_cache = NULL;
4849 if ((aggsum_compare(&arc_sums.arcstat_meta_used,
4850 arc_meta_limit) >= 0) && zfs_arc_meta_prune) {
4852 * We are exceeding our meta-data cache limit.
4853 * Prune some entries to release holds on meta-data.
4855 arc_prune_async(zfs_arc_meta_prune);
4859 * Reclaim unused memory from all kmem caches.
4865 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4867 /* reach upper limit of cache size on 32-bit */
4868 if (zio_buf_cache[i] == NULL)
4871 if (zio_buf_cache[i] != prev_cache) {
4872 prev_cache = zio_buf_cache[i];
4873 kmem_cache_reap_now(zio_buf_cache[i]);
4875 if (zio_data_buf_cache[i] != prev_data_cache) {
4876 prev_data_cache = zio_data_buf_cache[i];
4877 kmem_cache_reap_now(zio_data_buf_cache[i]);
4880 kmem_cache_reap_now(buf_cache);
4881 kmem_cache_reap_now(hdr_full_cache);
4882 kmem_cache_reap_now(hdr_l2only_cache);
4883 kmem_cache_reap_now(zfs_btree_leaf_cache);
4884 abd_cache_reap_now();
4888 arc_evict_cb_check(void *arg, zthr_t *zthr)
4890 (void) arg, (void) zthr;
4894 * This is necessary in order to keep the kstat information
4895 * up to date for tools that display kstat data such as the
4896 * mdb ::arc dcmd and the Linux crash utility. These tools
4897 * typically do not call kstat's update function, but simply
4898 * dump out stats from the most recent update. Without
4899 * this call, these commands may show stale stats for the
4900 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4901 * with this call, the data might be out of date if the
4902 * evict thread hasn't been woken recently; but that should
4903 * suffice. The arc_state_t structures can be queried
4904 * directly if more accurate information is needed.
4906 if (arc_ksp != NULL)
4907 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4911 * We have to rely on arc_wait_for_eviction() to tell us when to
4912 * evict, rather than checking if we are overflowing here, so that we
4913 * are sure to not leave arc_wait_for_eviction() waiting on aew_cv.
4914 * If we have become "not overflowing" since arc_wait_for_eviction()
4915 * checked, we need to wake it up. We could broadcast the CV here,
4916 * but arc_wait_for_eviction() may have not yet gone to sleep. We
4917 * would need to use a mutex to ensure that this function doesn't
4918 * broadcast until arc_wait_for_eviction() has gone to sleep (e.g.
4919 * the arc_evict_lock). However, the lock ordering of such a lock
4920 * would necessarily be incorrect with respect to the zthr_lock,
4921 * which is held before this function is called, and is held by
4922 * arc_wait_for_eviction() when it calls zthr_wakeup().
4924 return (arc_evict_needed);
4928 * Keep arc_size under arc_c by running arc_evict which evicts data
4932 arc_evict_cb(void *arg, zthr_t *zthr)
4934 (void) arg, (void) zthr;
4936 uint64_t evicted = 0;
4937 fstrans_cookie_t cookie = spl_fstrans_mark();
4939 /* Evict from cache */
4940 evicted = arc_evict();
4943 * If evicted is zero, we couldn't evict anything
4944 * via arc_evict(). This could be due to hash lock
4945 * collisions, but more likely due to the majority of
4946 * arc buffers being unevictable. Therefore, even if
4947 * arc_size is above arc_c, another pass is unlikely to
4948 * be helpful and could potentially cause us to enter an
4949 * infinite loop. Additionally, zthr_iscancelled() is
4950 * checked here so that if the arc is shutting down, the
4951 * broadcast will wake any remaining arc evict waiters.
4953 mutex_enter(&arc_evict_lock);
4954 arc_evict_needed = !zthr_iscancelled(arc_evict_zthr) &&
4955 evicted > 0 && aggsum_compare(&arc_sums.arcstat_size, arc_c) > 0;
4956 if (!arc_evict_needed) {
4958 * We're either no longer overflowing, or we
4959 * can't evict anything more, so we should wake
4960 * arc_get_data_impl() sooner.
4962 arc_evict_waiter_t *aw;
4963 while ((aw = list_remove_head(&arc_evict_waiters)) != NULL) {
4964 cv_broadcast(&aw->aew_cv);
4966 arc_set_need_free();
4968 mutex_exit(&arc_evict_lock);
4969 spl_fstrans_unmark(cookie);
4973 arc_reap_cb_check(void *arg, zthr_t *zthr)
4975 (void) arg, (void) zthr;
4977 int64_t free_memory = arc_available_memory();
4978 static int reap_cb_check_counter = 0;
4981 * If a kmem reap is already active, don't schedule more. We must
4982 * check for this because kmem_cache_reap_soon() won't actually
4983 * block on the cache being reaped (this is to prevent callers from
4984 * becoming implicitly blocked by a system-wide kmem reap -- which,
4985 * on a system with many, many full magazines, can take minutes).
4987 if (!kmem_cache_reap_active() && free_memory < 0) {
4989 arc_no_grow = B_TRUE;
4992 * Wait at least zfs_grow_retry (default 5) seconds
4993 * before considering growing.
4995 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4997 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4998 arc_no_grow = B_TRUE;
4999 } else if (gethrtime() >= arc_growtime) {
5000 arc_no_grow = B_FALSE;
5004 * Called unconditionally every 60 seconds to reclaim unused
5005 * zstd compression and decompression context. This is done
5006 * here to avoid the need for an independent thread.
5008 if (!((reap_cb_check_counter++) % 60))
5009 zfs_zstd_cache_reap_now();
5015 * Keep enough free memory in the system by reaping the ARC's kmem
5016 * caches. To cause more slabs to be reapable, we may reduce the
5017 * target size of the cache (arc_c), causing the arc_evict_cb()
5018 * to free more buffers.
5021 arc_reap_cb(void *arg, zthr_t *zthr)
5023 (void) arg, (void) zthr;
5025 int64_t free_memory;
5026 fstrans_cookie_t cookie = spl_fstrans_mark();
5029 * Kick off asynchronous kmem_reap()'s of all our caches.
5031 arc_kmem_reap_soon();
5034 * Wait at least arc_kmem_cache_reap_retry_ms between
5035 * arc_kmem_reap_soon() calls. Without this check it is possible to
5036 * end up in a situation where we spend lots of time reaping
5037 * caches, while we're near arc_c_min. Waiting here also gives the
5038 * subsequent free memory check a chance of finding that the
5039 * asynchronous reap has already freed enough memory, and we don't
5040 * need to call arc_reduce_target_size().
5042 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
5045 * Reduce the target size as needed to maintain the amount of free
5046 * memory in the system at a fraction of the arc_size (1/128th by
5047 * default). If oversubscribed (free_memory < 0) then reduce the
5048 * target arc_size by the deficit amount plus the fractional
5049 * amount. If free memory is positive but less than the fractional
5050 * amount, reduce by what is needed to hit the fractional amount.
5052 free_memory = arc_available_memory();
5054 int64_t can_free = arc_c - arc_c_min;
5056 int64_t to_free = (can_free >> arc_shrink_shift) - free_memory;
5058 arc_reduce_target_size(to_free);
5060 spl_fstrans_unmark(cookie);
5065 * Determine the amount of memory eligible for eviction contained in the
5066 * ARC. All clean data reported by the ghost lists can always be safely
5067 * evicted. Due to arc_c_min, the same does not hold for all clean data
5068 * contained by the regular mru and mfu lists.
5070 * In the case of the regular mru and mfu lists, we need to report as
5071 * much clean data as possible, such that evicting that same reported
5072 * data will not bring arc_size below arc_c_min. Thus, in certain
5073 * circumstances, the total amount of clean data in the mru and mfu
5074 * lists might not actually be evictable.
5076 * The following two distinct cases are accounted for:
5078 * 1. The sum of the amount of dirty data contained by both the mru and
5079 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5080 * is greater than or equal to arc_c_min.
5081 * (i.e. amount of dirty data >= arc_c_min)
5083 * This is the easy case; all clean data contained by the mru and mfu
5084 * lists is evictable. Evicting all clean data can only drop arc_size
5085 * to the amount of dirty data, which is greater than arc_c_min.
5087 * 2. The sum of the amount of dirty data contained by both the mru and
5088 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5089 * is less than arc_c_min.
5090 * (i.e. arc_c_min > amount of dirty data)
5092 * 2.1. arc_size is greater than or equal arc_c_min.
5093 * (i.e. arc_size >= arc_c_min > amount of dirty data)
5095 * In this case, not all clean data from the regular mru and mfu
5096 * lists is actually evictable; we must leave enough clean data
5097 * to keep arc_size above arc_c_min. Thus, the maximum amount of
5098 * evictable data from the two lists combined, is exactly the
5099 * difference between arc_size and arc_c_min.
5101 * 2.2. arc_size is less than arc_c_min
5102 * (i.e. arc_c_min > arc_size > amount of dirty data)
5104 * In this case, none of the data contained in the mru and mfu
5105 * lists is evictable, even if it's clean. Since arc_size is
5106 * already below arc_c_min, evicting any more would only
5107 * increase this negative difference.
5110 #endif /* _KERNEL */
5113 * Adapt arc info given the number of bytes we are trying to add and
5114 * the state that we are coming from. This function is only called
5115 * when we are adding new content to the cache.
5118 arc_adapt(int bytes, arc_state_t *state)
5121 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
5122 int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size);
5123 int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size);
5127 * Adapt the target size of the MRU list:
5128 * - if we just hit in the MRU ghost list, then increase
5129 * the target size of the MRU list.
5130 * - if we just hit in the MFU ghost list, then increase
5131 * the target size of the MFU list by decreasing the
5132 * target size of the MRU list.
5134 if (state == arc_mru_ghost) {
5135 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
5136 if (!zfs_arc_p_dampener_disable)
5137 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
5139 arc_p = MIN(arc_c - arc_p_min, arc_p + (uint64_t)bytes * mult);
5140 } else if (state == arc_mfu_ghost) {
5143 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
5144 if (!zfs_arc_p_dampener_disable)
5145 mult = MIN(mult, 10);
5147 delta = MIN(bytes * mult, arc_p);
5148 arc_p = MAX(arc_p_min, arc_p - delta);
5150 ASSERT((int64_t)arc_p >= 0);
5153 * Wake reap thread if we do not have any available memory
5155 if (arc_reclaim_needed()) {
5156 zthr_wakeup(arc_reap_zthr);
5163 if (arc_c >= arc_c_max)
5167 * If we're within (2 * maxblocksize) bytes of the target
5168 * cache size, increment the target cache size
5170 ASSERT3U(arc_c, >=, 2ULL << SPA_MAXBLOCKSHIFT);
5171 if (aggsum_upper_bound(&arc_sums.arcstat_size) >=
5172 arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
5173 atomic_add_64(&arc_c, (int64_t)bytes);
5174 if (arc_c > arc_c_max)
5176 else if (state == arc_anon && arc_p < arc_c >> 1)
5177 atomic_add_64(&arc_p, (int64_t)bytes);
5181 ASSERT((int64_t)arc_p >= 0);
5185 * Check if arc_size has grown past our upper threshold, determined by
5186 * zfs_arc_overflow_shift.
5188 static arc_ovf_level_t
5189 arc_is_overflowing(boolean_t use_reserve)
5191 /* Always allow at least one block of overflow */
5192 int64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5193 arc_c >> zfs_arc_overflow_shift);
5196 * We just compare the lower bound here for performance reasons. Our
5197 * primary goals are to make sure that the arc never grows without
5198 * bound, and that it can reach its maximum size. This check
5199 * accomplishes both goals. The maximum amount we could run over by is
5200 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
5201 * in the ARC. In practice, that's in the tens of MB, which is low
5202 * enough to be safe.
5204 int64_t over = aggsum_lower_bound(&arc_sums.arcstat_size) -
5205 arc_c - overflow / 2;
5208 return (over < 0 ? ARC_OVF_NONE :
5209 over < overflow ? ARC_OVF_SOME : ARC_OVF_SEVERE);
5213 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, const void *tag,
5216 arc_buf_contents_t type = arc_buf_type(hdr);
5218 arc_get_data_impl(hdr, size, tag, alloc_flags);
5219 if (type == ARC_BUFC_METADATA) {
5220 return (abd_alloc(size, B_TRUE));
5222 ASSERT(type == ARC_BUFC_DATA);
5223 return (abd_alloc(size, B_FALSE));
5228 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, const void *tag)
5230 arc_buf_contents_t type = arc_buf_type(hdr);
5232 arc_get_data_impl(hdr, size, tag, ARC_HDR_DO_ADAPT);
5233 if (type == ARC_BUFC_METADATA) {
5234 return (zio_buf_alloc(size));
5236 ASSERT(type == ARC_BUFC_DATA);
5237 return (zio_data_buf_alloc(size));
5242 * Wait for the specified amount of data (in bytes) to be evicted from the
5243 * ARC, and for there to be sufficient free memory in the system. Waiting for
5244 * eviction ensures that the memory used by the ARC decreases. Waiting for
5245 * free memory ensures that the system won't run out of free pages, regardless
5246 * of ARC behavior and settings. See arc_lowmem_init().
5249 arc_wait_for_eviction(uint64_t amount, boolean_t use_reserve)
5251 switch (arc_is_overflowing(use_reserve)) {
5256 * This is a bit racy without taking arc_evict_lock, but the
5257 * worst that can happen is we either call zthr_wakeup() extra
5258 * time due to race with other thread here, or the set flag
5259 * get cleared by arc_evict_cb(), which is unlikely due to
5260 * big hysteresis, but also not important since at this level
5261 * of overflow the eviction is purely advisory. Same time
5262 * taking the global lock here every time without waiting for
5263 * the actual eviction creates a significant lock contention.
5265 if (!arc_evict_needed) {
5266 arc_evict_needed = B_TRUE;
5267 zthr_wakeup(arc_evict_zthr);
5270 case ARC_OVF_SEVERE:
5273 arc_evict_waiter_t aw;
5274 list_link_init(&aw.aew_node);
5275 cv_init(&aw.aew_cv, NULL, CV_DEFAULT, NULL);
5277 uint64_t last_count = 0;
5278 mutex_enter(&arc_evict_lock);
5279 if (!list_is_empty(&arc_evict_waiters)) {
5280 arc_evict_waiter_t *last =
5281 list_tail(&arc_evict_waiters);
5282 last_count = last->aew_count;
5283 } else if (!arc_evict_needed) {
5284 arc_evict_needed = B_TRUE;
5285 zthr_wakeup(arc_evict_zthr);
5288 * Note, the last waiter's count may be less than
5289 * arc_evict_count if we are low on memory in which
5290 * case arc_evict_state_impl() may have deferred
5291 * wakeups (but still incremented arc_evict_count).
5293 aw.aew_count = MAX(last_count, arc_evict_count) + amount;
5295 list_insert_tail(&arc_evict_waiters, &aw);
5297 arc_set_need_free();
5299 DTRACE_PROBE3(arc__wait__for__eviction,
5301 uint64_t, arc_evict_count,
5302 uint64_t, aw.aew_count);
5305 * We will be woken up either when arc_evict_count reaches
5306 * aew_count, or when the ARC is no longer overflowing and
5307 * eviction completes.
5308 * In case of "false" wakeup, we will still be on the list.
5311 cv_wait(&aw.aew_cv, &arc_evict_lock);
5312 } while (list_link_active(&aw.aew_node));
5313 mutex_exit(&arc_evict_lock);
5315 cv_destroy(&aw.aew_cv);
5321 * Allocate a block and return it to the caller. If we are hitting the
5322 * hard limit for the cache size, we must sleep, waiting for the eviction
5323 * thread to catch up. If we're past the target size but below the hard
5324 * limit, we'll only signal the reclaim thread and continue on.
5327 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, const void *tag,
5330 arc_state_t *state = hdr->b_l1hdr.b_state;
5331 arc_buf_contents_t type = arc_buf_type(hdr);
5333 if (alloc_flags & ARC_HDR_DO_ADAPT)
5334 arc_adapt(size, state);
5337 * If arc_size is currently overflowing, we must be adding data
5338 * faster than we are evicting. To ensure we don't compound the
5339 * problem by adding more data and forcing arc_size to grow even
5340 * further past it's target size, we wait for the eviction thread to
5341 * make some progress. We also wait for there to be sufficient free
5342 * memory in the system, as measured by arc_free_memory().
5344 * Specifically, we wait for zfs_arc_eviction_pct percent of the
5345 * requested size to be evicted. This should be more than 100%, to
5346 * ensure that that progress is also made towards getting arc_size
5347 * under arc_c. See the comment above zfs_arc_eviction_pct.
5349 arc_wait_for_eviction(size * zfs_arc_eviction_pct / 100,
5350 alloc_flags & ARC_HDR_USE_RESERVE);
5352 VERIFY3U(hdr->b_type, ==, type);
5353 if (type == ARC_BUFC_METADATA) {
5354 arc_space_consume(size, ARC_SPACE_META);
5356 arc_space_consume(size, ARC_SPACE_DATA);
5360 * Update the state size. Note that ghost states have a
5361 * "ghost size" and so don't need to be updated.
5363 if (!GHOST_STATE(state)) {
5365 (void) zfs_refcount_add_many(&state->arcs_size, size, tag);
5368 * If this is reached via arc_read, the link is
5369 * protected by the hash lock. If reached via
5370 * arc_buf_alloc, the header should not be accessed by
5371 * any other thread. And, if reached via arc_read_done,
5372 * the hash lock will protect it if it's found in the
5373 * hash table; otherwise no other thread should be
5374 * trying to [add|remove]_reference it.
5376 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5377 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5378 (void) zfs_refcount_add_many(&state->arcs_esize[type],
5383 * If we are growing the cache, and we are adding anonymous
5384 * data, and we have outgrown arc_p, update arc_p
5386 if (aggsum_upper_bound(&arc_sums.arcstat_size) < arc_c &&
5387 hdr->b_l1hdr.b_state == arc_anon &&
5388 (zfs_refcount_count(&arc_anon->arcs_size) +
5389 zfs_refcount_count(&arc_mru->arcs_size) > arc_p &&
5390 arc_p < arc_c >> 1))
5391 arc_p = MIN(arc_c, arc_p + size);
5396 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size,
5399 arc_free_data_impl(hdr, size, tag);
5404 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, const void *tag)
5406 arc_buf_contents_t type = arc_buf_type(hdr);
5408 arc_free_data_impl(hdr, size, tag);
5409 if (type == ARC_BUFC_METADATA) {
5410 zio_buf_free(buf, size);
5412 ASSERT(type == ARC_BUFC_DATA);
5413 zio_data_buf_free(buf, size);
5418 * Free the arc data buffer.
5421 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, const void *tag)
5423 arc_state_t *state = hdr->b_l1hdr.b_state;
5424 arc_buf_contents_t type = arc_buf_type(hdr);
5426 /* protected by hash lock, if in the hash table */
5427 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5428 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5429 ASSERT(state != arc_anon && state != arc_l2c_only);
5431 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
5434 (void) zfs_refcount_remove_many(&state->arcs_size, size, tag);
5436 VERIFY3U(hdr->b_type, ==, type);
5437 if (type == ARC_BUFC_METADATA) {
5438 arc_space_return(size, ARC_SPACE_META);
5440 ASSERT(type == ARC_BUFC_DATA);
5441 arc_space_return(size, ARC_SPACE_DATA);
5446 * This routine is called whenever a buffer is accessed.
5447 * NOTE: the hash lock is dropped in this function.
5450 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5454 ASSERT(MUTEX_HELD(hash_lock));
5455 ASSERT(HDR_HAS_L1HDR(hdr));
5457 if (hdr->b_l1hdr.b_state == arc_anon) {
5459 * This buffer is not in the cache, and does not
5460 * appear in our "ghost" list. Add the new buffer
5464 ASSERT0(hdr->b_l1hdr.b_arc_access);
5465 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5466 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5467 arc_change_state(arc_mru, hdr, hash_lock);
5469 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5470 now = ddi_get_lbolt();
5473 * If this buffer is here because of a prefetch, then either:
5474 * - clear the flag if this is a "referencing" read
5475 * (any subsequent access will bump this into the MFU state).
5477 * - move the buffer to the head of the list if this is
5478 * another prefetch (to make it less likely to be evicted).
5480 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5481 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5482 /* link protected by hash lock */
5483 ASSERT(multilist_link_active(
5484 &hdr->b_l1hdr.b_arc_node));
5486 if (HDR_HAS_L2HDR(hdr))
5487 l2arc_hdr_arcstats_decrement_state(hdr);
5488 arc_hdr_clear_flags(hdr,
5490 ARC_FLAG_PRESCIENT_PREFETCH);
5491 hdr->b_l1hdr.b_mru_hits++;
5492 ARCSTAT_BUMP(arcstat_mru_hits);
5493 if (HDR_HAS_L2HDR(hdr))
5494 l2arc_hdr_arcstats_increment_state(hdr);
5496 hdr->b_l1hdr.b_arc_access = now;
5501 * This buffer has been "accessed" only once so far,
5502 * but it is still in the cache. Move it to the MFU
5505 if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access +
5508 * More than 125ms have passed since we
5509 * instantiated this buffer. Move it to the
5510 * most frequently used state.
5512 hdr->b_l1hdr.b_arc_access = now;
5513 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5514 arc_change_state(arc_mfu, hdr, hash_lock);
5516 hdr->b_l1hdr.b_mru_hits++;
5517 ARCSTAT_BUMP(arcstat_mru_hits);
5518 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5519 arc_state_t *new_state;
5521 * This buffer has been "accessed" recently, but
5522 * was evicted from the cache. Move it to the
5525 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5526 new_state = arc_mru;
5527 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5528 if (HDR_HAS_L2HDR(hdr))
5529 l2arc_hdr_arcstats_decrement_state(hdr);
5530 arc_hdr_clear_flags(hdr,
5532 ARC_FLAG_PRESCIENT_PREFETCH);
5533 if (HDR_HAS_L2HDR(hdr))
5534 l2arc_hdr_arcstats_increment_state(hdr);
5536 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5538 new_state = arc_mfu;
5539 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5542 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5543 arc_change_state(new_state, hdr, hash_lock);
5545 hdr->b_l1hdr.b_mru_ghost_hits++;
5546 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5547 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5549 * This buffer has been accessed more than once and is
5550 * still in the cache. Keep it in the MFU state.
5552 * NOTE: an add_reference() that occurred when we did
5553 * the arc_read() will have kicked this off the list.
5554 * If it was a prefetch, we will explicitly move it to
5555 * the head of the list now.
5558 hdr->b_l1hdr.b_mfu_hits++;
5559 ARCSTAT_BUMP(arcstat_mfu_hits);
5560 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5561 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5562 arc_state_t *new_state = arc_mfu;
5564 * This buffer has been accessed more than once but has
5565 * been evicted from the cache. Move it back to the
5569 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5571 * This is a prefetch access...
5572 * move this block back to the MRU state.
5574 new_state = arc_mru;
5577 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5578 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5579 arc_change_state(new_state, hdr, hash_lock);
5581 hdr->b_l1hdr.b_mfu_ghost_hits++;
5582 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5583 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5585 * This buffer is on the 2nd Level ARC.
5588 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5589 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5590 arc_change_state(arc_mfu, hdr, hash_lock);
5592 cmn_err(CE_PANIC, "invalid arc state 0x%p",
5593 hdr->b_l1hdr.b_state);
5598 * This routine is called by dbuf_hold() to update the arc_access() state
5599 * which otherwise would be skipped for entries in the dbuf cache.
5602 arc_buf_access(arc_buf_t *buf)
5604 mutex_enter(&buf->b_evict_lock);
5605 arc_buf_hdr_t *hdr = buf->b_hdr;
5608 * Avoid taking the hash_lock when possible as an optimization.
5609 * The header must be checked again under the hash_lock in order
5610 * to handle the case where it is concurrently being released.
5612 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5613 mutex_exit(&buf->b_evict_lock);
5617 kmutex_t *hash_lock = HDR_LOCK(hdr);
5618 mutex_enter(hash_lock);
5620 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5621 mutex_exit(hash_lock);
5622 mutex_exit(&buf->b_evict_lock);
5623 ARCSTAT_BUMP(arcstat_access_skip);
5627 mutex_exit(&buf->b_evict_lock);
5629 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5630 hdr->b_l1hdr.b_state == arc_mfu);
5632 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5633 arc_access(hdr, hash_lock);
5634 mutex_exit(hash_lock);
5636 ARCSTAT_BUMP(arcstat_hits);
5637 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr) && !HDR_PRESCIENT_PREFETCH(hdr),
5638 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5641 /* a generic arc_read_done_func_t which you can use */
5643 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5644 arc_buf_t *buf, void *arg)
5646 (void) zio, (void) zb, (void) bp;
5651 memcpy(arg, buf->b_data, arc_buf_size(buf));
5652 arc_buf_destroy(buf, arg);
5655 /* a generic arc_read_done_func_t */
5657 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5658 arc_buf_t *buf, void *arg)
5660 (void) zb, (void) bp;
5661 arc_buf_t **bufp = arg;
5664 ASSERT(zio == NULL || zio->io_error != 0);
5667 ASSERT(zio == NULL || zio->io_error == 0);
5669 ASSERT(buf->b_data != NULL);
5674 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5676 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5677 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5678 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5680 if (HDR_COMPRESSION_ENABLED(hdr)) {
5681 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5682 BP_GET_COMPRESS(bp));
5684 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5685 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5686 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5691 arc_read_done(zio_t *zio)
5693 blkptr_t *bp = zio->io_bp;
5694 arc_buf_hdr_t *hdr = zio->io_private;
5695 kmutex_t *hash_lock = NULL;
5696 arc_callback_t *callback_list;
5697 arc_callback_t *acb;
5698 boolean_t freeable = B_FALSE;
5701 * The hdr was inserted into hash-table and removed from lists
5702 * prior to starting I/O. We should find this header, since
5703 * it's in the hash table, and it should be legit since it's
5704 * not possible to evict it during the I/O. The only possible
5705 * reason for it not to be found is if we were freed during the
5708 if (HDR_IN_HASH_TABLE(hdr)) {
5709 arc_buf_hdr_t *found;
5711 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5712 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5713 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5714 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5715 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5717 found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock);
5719 ASSERT((found == hdr &&
5720 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5721 (found == hdr && HDR_L2_READING(hdr)));
5722 ASSERT3P(hash_lock, !=, NULL);
5725 if (BP_IS_PROTECTED(bp)) {
5726 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5727 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5728 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5729 hdr->b_crypt_hdr.b_iv);
5731 if (zio->io_error == 0) {
5732 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5735 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5736 sizeof (zil_chain_t));
5737 zio_crypt_decode_mac_zil(tmpbuf,
5738 hdr->b_crypt_hdr.b_mac);
5739 abd_return_buf(zio->io_abd, tmpbuf,
5740 sizeof (zil_chain_t));
5742 zio_crypt_decode_mac_bp(bp,
5743 hdr->b_crypt_hdr.b_mac);
5748 if (zio->io_error == 0) {
5749 /* byteswap if necessary */
5750 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5751 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5752 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5754 hdr->b_l1hdr.b_byteswap =
5755 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5758 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5760 if (!HDR_L2_READING(hdr)) {
5761 hdr->b_complevel = zio->io_prop.zp_complevel;
5765 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5766 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5767 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5769 callback_list = hdr->b_l1hdr.b_acb;
5770 ASSERT3P(callback_list, !=, NULL);
5772 if (hash_lock && zio->io_error == 0 &&
5773 hdr->b_l1hdr.b_state == arc_anon) {
5775 * Only call arc_access on anonymous buffers. This is because
5776 * if we've issued an I/O for an evicted buffer, we've already
5777 * called arc_access (to prevent any simultaneous readers from
5778 * getting confused).
5780 arc_access(hdr, hash_lock);
5784 * If a read request has a callback (i.e. acb_done is not NULL), then we
5785 * make a buf containing the data according to the parameters which were
5786 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5787 * aren't needlessly decompressing the data multiple times.
5789 int callback_cnt = 0;
5790 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5791 if (!acb->acb_done || acb->acb_nobuf)
5796 if (zio->io_error != 0)
5799 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5800 &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
5801 acb->acb_compressed, acb->acb_noauth, B_TRUE,
5805 * Assert non-speculative zios didn't fail because an
5806 * encryption key wasn't loaded
5808 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
5812 * If we failed to decrypt, report an error now (as the zio
5813 * layer would have done if it had done the transforms).
5815 if (error == ECKSUM) {
5816 ASSERT(BP_IS_PROTECTED(bp));
5817 error = SET_ERROR(EIO);
5818 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5819 spa_log_error(zio->io_spa, &acb->acb_zb);
5820 (void) zfs_ereport_post(
5821 FM_EREPORT_ZFS_AUTHENTICATION,
5822 zio->io_spa, NULL, &acb->acb_zb, zio, 0);
5828 * Decompression or decryption failed. Set
5829 * io_error so that when we call acb_done
5830 * (below), we will indicate that the read
5831 * failed. Note that in the unusual case
5832 * where one callback is compressed and another
5833 * uncompressed, we will mark all of them
5834 * as failed, even though the uncompressed
5835 * one can't actually fail. In this case,
5836 * the hdr will not be anonymous, because
5837 * if there are multiple callbacks, it's
5838 * because multiple threads found the same
5839 * arc buf in the hash table.
5841 zio->io_error = error;
5846 * If there are multiple callbacks, we must have the hash lock,
5847 * because the only way for multiple threads to find this hdr is
5848 * in the hash table. This ensures that if there are multiple
5849 * callbacks, the hdr is not anonymous. If it were anonymous,
5850 * we couldn't use arc_buf_destroy() in the error case below.
5852 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5854 hdr->b_l1hdr.b_acb = NULL;
5855 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5856 if (callback_cnt == 0)
5857 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
5859 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5860 callback_list != NULL);
5862 if (zio->io_error == 0) {
5863 arc_hdr_verify(hdr, zio->io_bp);
5865 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5866 if (hdr->b_l1hdr.b_state != arc_anon)
5867 arc_change_state(arc_anon, hdr, hash_lock);
5868 if (HDR_IN_HASH_TABLE(hdr))
5869 buf_hash_remove(hdr);
5870 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5874 * Broadcast before we drop the hash_lock to avoid the possibility
5875 * that the hdr (and hence the cv) might be freed before we get to
5876 * the cv_broadcast().
5878 cv_broadcast(&hdr->b_l1hdr.b_cv);
5880 if (hash_lock != NULL) {
5881 mutex_exit(hash_lock);
5884 * This block was freed while we waited for the read to
5885 * complete. It has been removed from the hash table and
5886 * moved to the anonymous state (so that it won't show up
5889 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5890 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5893 /* execute each callback and free its structure */
5894 while ((acb = callback_list) != NULL) {
5895 if (acb->acb_done != NULL) {
5896 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5898 * If arc_buf_alloc_impl() fails during
5899 * decompression, the buf will still be
5900 * allocated, and needs to be freed here.
5902 arc_buf_destroy(acb->acb_buf,
5904 acb->acb_buf = NULL;
5906 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5907 acb->acb_buf, acb->acb_private);
5910 if (acb->acb_zio_dummy != NULL) {
5911 acb->acb_zio_dummy->io_error = zio->io_error;
5912 zio_nowait(acb->acb_zio_dummy);
5915 callback_list = acb->acb_next;
5916 kmem_free(acb, sizeof (arc_callback_t));
5920 arc_hdr_destroy(hdr);
5924 * "Read" the block at the specified DVA (in bp) via the
5925 * cache. If the block is found in the cache, invoke the provided
5926 * callback immediately and return. Note that the `zio' parameter
5927 * in the callback will be NULL in this case, since no IO was
5928 * required. If the block is not in the cache pass the read request
5929 * on to the spa with a substitute callback function, so that the
5930 * requested block will be added to the cache.
5932 * If a read request arrives for a block that has a read in-progress,
5933 * either wait for the in-progress read to complete (and return the
5934 * results); or, if this is a read with a "done" func, add a record
5935 * to the read to invoke the "done" func when the read completes,
5936 * and return; or just return.
5938 * arc_read_done() will invoke all the requested "done" functions
5939 * for readers of this block.
5942 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
5943 arc_read_done_func_t *done, void *private, zio_priority_t priority,
5944 int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5946 arc_buf_hdr_t *hdr = NULL;
5947 kmutex_t *hash_lock = NULL;
5949 uint64_t guid = spa_load_guid(spa);
5950 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
5951 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
5952 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5953 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
5954 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5955 boolean_t embedded_bp = !!BP_IS_EMBEDDED(bp);
5956 boolean_t no_buf = *arc_flags & ARC_FLAG_NO_BUF;
5959 ASSERT(!embedded_bp ||
5960 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5961 ASSERT(!BP_IS_HOLE(bp));
5962 ASSERT(!BP_IS_REDACTED(bp));
5965 * Normally SPL_FSTRANS will already be set since kernel threads which
5966 * expect to call the DMU interfaces will set it when created. System
5967 * calls are similarly handled by setting/cleaning the bit in the
5968 * registered callback (module/os/.../zfs/zpl_*).
5970 * External consumers such as Lustre which call the exported DMU
5971 * interfaces may not have set SPL_FSTRANS. To avoid a deadlock
5972 * on the hash_lock always set and clear the bit.
5974 fstrans_cookie_t cookie = spl_fstrans_mark();
5977 * Verify the block pointer contents are reasonable. This should
5978 * always be the case since the blkptr is protected by a checksum.
5979 * However, if there is damage it's desirable to detect this early
5980 * and treat it as a checksum error. This allows an alternate blkptr
5981 * to be tried when one is available (e.g. ditto blocks).
5983 if (!zfs_blkptr_verify(spa, bp, zio_flags & ZIO_FLAG_CONFIG_WRITER,
5985 rc = SET_ERROR(ECKSUM);
5991 * Embedded BP's have no DVA and require no I/O to "read".
5992 * Create an anonymous arc buf to back it.
5994 hdr = buf_hash_find(guid, bp, &hash_lock);
5998 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5999 * we maintain encrypted data separately from compressed / uncompressed
6000 * data. If the user is requesting raw encrypted data and we don't have
6001 * that in the header we will read from disk to guarantee that we can
6002 * get it even if the encryption keys aren't loaded.
6004 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
6005 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
6006 arc_buf_t *buf = NULL;
6007 *arc_flags |= ARC_FLAG_CACHED;
6009 if (HDR_IO_IN_PROGRESS(hdr)) {
6010 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
6012 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
6013 mutex_exit(hash_lock);
6014 ARCSTAT_BUMP(arcstat_cached_only_in_progress);
6015 rc = SET_ERROR(ENOENT);
6019 ASSERT3P(head_zio, !=, NULL);
6020 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
6021 priority == ZIO_PRIORITY_SYNC_READ) {
6023 * This is a sync read that needs to wait for
6024 * an in-flight async read. Request that the
6025 * zio have its priority upgraded.
6027 zio_change_priority(head_zio, priority);
6028 DTRACE_PROBE1(arc__async__upgrade__sync,
6029 arc_buf_hdr_t *, hdr);
6030 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
6032 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
6033 arc_hdr_clear_flags(hdr,
6034 ARC_FLAG_PREDICTIVE_PREFETCH);
6038 * If there are multiple threads reading the same block
6039 * and that block is not yet in the ARC, then only one
6040 * thread will do the physical I/O and all other
6041 * threads will wait until that I/O completes.
6042 * Synchronous reads use the b_cv whereas nowait reads
6043 * register a callback. Both are signalled/called in
6046 * Errors of the physical I/O may need to be propagated
6047 * to the pio. For synchronous reads, we simply restart
6048 * this function and it will reassess. Nowait reads
6049 * attach the acb_zio_dummy zio to pio and
6050 * arc_read_done propagates the physical I/O's io_error
6051 * to acb_zio_dummy, and thereby to pio.
6054 if (*arc_flags & ARC_FLAG_WAIT) {
6055 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6056 mutex_exit(hash_lock);
6059 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6062 arc_callback_t *acb = NULL;
6064 acb = kmem_zalloc(sizeof (arc_callback_t),
6066 acb->acb_done = done;
6067 acb->acb_private = private;
6068 acb->acb_compressed = compressed_read;
6069 acb->acb_encrypted = encrypted_read;
6070 acb->acb_noauth = noauth_read;
6071 acb->acb_nobuf = no_buf;
6074 acb->acb_zio_dummy = zio_null(pio,
6075 spa, NULL, NULL, NULL, zio_flags);
6077 ASSERT3P(acb->acb_done, !=, NULL);
6078 acb->acb_zio_head = head_zio;
6079 acb->acb_next = hdr->b_l1hdr.b_acb;
6080 hdr->b_l1hdr.b_acb = acb;
6082 mutex_exit(hash_lock);
6086 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
6087 hdr->b_l1hdr.b_state == arc_mfu);
6089 if (done && !no_buf) {
6090 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
6092 * This is a demand read which does not have to
6093 * wait for i/o because we did a predictive
6094 * prefetch i/o for it, which has completed.
6097 arc__demand__hit__predictive__prefetch,
6098 arc_buf_hdr_t *, hdr);
6100 arcstat_demand_hit_predictive_prefetch);
6101 arc_hdr_clear_flags(hdr,
6102 ARC_FLAG_PREDICTIVE_PREFETCH);
6105 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
6107 arcstat_demand_hit_prescient_prefetch);
6108 arc_hdr_clear_flags(hdr,
6109 ARC_FLAG_PRESCIENT_PREFETCH);
6112 ASSERT(!embedded_bp || !BP_IS_HOLE(bp));
6114 /* Get a buf with the desired data in it. */
6115 rc = arc_buf_alloc_impl(hdr, spa, zb, private,
6116 encrypted_read, compressed_read, noauth_read,
6120 * Convert authentication and decryption errors
6121 * to EIO (and generate an ereport if needed)
6122 * before leaving the ARC.
6124 rc = SET_ERROR(EIO);
6125 if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
6126 spa_log_error(spa, zb);
6127 (void) zfs_ereport_post(
6128 FM_EREPORT_ZFS_AUTHENTICATION,
6129 spa, NULL, zb, NULL, 0);
6133 (void) remove_reference(hdr, hash_lock,
6135 arc_buf_destroy_impl(buf);
6139 /* assert any errors weren't due to unloaded keys */
6140 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
6142 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
6143 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6144 if (HDR_HAS_L2HDR(hdr))
6145 l2arc_hdr_arcstats_decrement_state(hdr);
6146 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6147 if (HDR_HAS_L2HDR(hdr))
6148 l2arc_hdr_arcstats_increment_state(hdr);
6150 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
6151 arc_access(hdr, hash_lock);
6152 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6153 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6154 if (*arc_flags & ARC_FLAG_L2CACHE)
6155 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6156 mutex_exit(hash_lock);
6157 ARCSTAT_BUMP(arcstat_hits);
6158 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6159 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
6160 data, metadata, hits);
6163 done(NULL, zb, bp, buf, private);
6165 uint64_t lsize = BP_GET_LSIZE(bp);
6166 uint64_t psize = BP_GET_PSIZE(bp);
6167 arc_callback_t *acb;
6170 boolean_t devw = B_FALSE;
6173 int alloc_flags = encrypted_read ? ARC_HDR_ALLOC_RDATA : 0;
6175 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
6176 rc = SET_ERROR(ENOENT);
6177 if (hash_lock != NULL)
6178 mutex_exit(hash_lock);
6184 * This block is not in the cache or it has
6187 arc_buf_hdr_t *exists = NULL;
6188 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
6189 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
6190 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), 0, type);
6193 hdr->b_dva = *BP_IDENTITY(bp);
6194 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
6195 exists = buf_hash_insert(hdr, &hash_lock);
6197 if (exists != NULL) {
6198 /* somebody beat us to the hash insert */
6199 mutex_exit(hash_lock);
6200 buf_discard_identity(hdr);
6201 arc_hdr_destroy(hdr);
6202 goto top; /* restart the IO request */
6204 alloc_flags |= ARC_HDR_DO_ADAPT;
6207 * This block is in the ghost cache or encrypted data
6208 * was requested and we didn't have it. If it was
6209 * L2-only (and thus didn't have an L1 hdr),
6210 * we realloc the header to add an L1 hdr.
6212 if (!HDR_HAS_L1HDR(hdr)) {
6213 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
6217 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
6218 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6219 ASSERT(!HDR_HAS_RABD(hdr));
6220 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6221 ASSERT0(zfs_refcount_count(
6222 &hdr->b_l1hdr.b_refcnt));
6223 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
6224 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
6225 } else if (HDR_IO_IN_PROGRESS(hdr)) {
6227 * If this header already had an IO in progress
6228 * and we are performing another IO to fetch
6229 * encrypted data we must wait until the first
6230 * IO completes so as not to confuse
6231 * arc_read_done(). This should be very rare
6232 * and so the performance impact shouldn't
6235 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6236 mutex_exit(hash_lock);
6241 * This is a delicate dance that we play here.
6242 * This hdr might be in the ghost list so we access
6243 * it to move it out of the ghost list before we
6244 * initiate the read. If it's a prefetch then
6245 * it won't have a callback so we'll remove the
6246 * reference that arc_buf_alloc_impl() created. We
6247 * do this after we've called arc_access() to
6248 * avoid hitting an assert in remove_reference().
6250 arc_adapt(arc_hdr_size(hdr), hdr->b_l1hdr.b_state);
6251 arc_access(hdr, hash_lock);
6254 arc_hdr_alloc_abd(hdr, alloc_flags);
6255 if (encrypted_read) {
6256 ASSERT(HDR_HAS_RABD(hdr));
6257 size = HDR_GET_PSIZE(hdr);
6258 hdr_abd = hdr->b_crypt_hdr.b_rabd;
6259 zio_flags |= ZIO_FLAG_RAW;
6261 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6262 size = arc_hdr_size(hdr);
6263 hdr_abd = hdr->b_l1hdr.b_pabd;
6265 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
6266 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6270 * For authenticated bp's, we do not ask the ZIO layer
6271 * to authenticate them since this will cause the entire
6272 * IO to fail if the key isn't loaded. Instead, we
6273 * defer authentication until arc_buf_fill(), which will
6274 * verify the data when the key is available.
6276 if (BP_IS_AUTHENTICATED(bp))
6277 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
6280 if (*arc_flags & ARC_FLAG_PREFETCH &&
6281 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6282 if (HDR_HAS_L2HDR(hdr))
6283 l2arc_hdr_arcstats_decrement_state(hdr);
6284 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6285 if (HDR_HAS_L2HDR(hdr))
6286 l2arc_hdr_arcstats_increment_state(hdr);
6288 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6289 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6290 if (*arc_flags & ARC_FLAG_L2CACHE)
6291 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6292 if (BP_IS_AUTHENTICATED(bp))
6293 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6294 if (BP_GET_LEVEL(bp) > 0)
6295 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
6296 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
6297 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
6298 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
6300 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
6301 acb->acb_done = done;
6302 acb->acb_private = private;
6303 acb->acb_compressed = compressed_read;
6304 acb->acb_encrypted = encrypted_read;
6305 acb->acb_noauth = noauth_read;
6308 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6309 hdr->b_l1hdr.b_acb = acb;
6310 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6312 if (HDR_HAS_L2HDR(hdr) &&
6313 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
6314 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
6315 addr = hdr->b_l2hdr.b_daddr;
6317 * Lock out L2ARC device removal.
6319 if (vdev_is_dead(vd) ||
6320 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
6325 * We count both async reads and scrub IOs as asynchronous so
6326 * that both can be upgraded in the event of a cache hit while
6327 * the read IO is still in-flight.
6329 if (priority == ZIO_PRIORITY_ASYNC_READ ||
6330 priority == ZIO_PRIORITY_SCRUB)
6331 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6333 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6336 * At this point, we have a level 1 cache miss or a blkptr
6337 * with embedded data. Try again in L2ARC if possible.
6339 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
6342 * Skip ARC stat bump for block pointers with embedded
6343 * data. The data are read from the blkptr itself via
6344 * decode_embedded_bp_compressed().
6347 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr,
6348 blkptr_t *, bp, uint64_t, lsize,
6349 zbookmark_phys_t *, zb);
6350 ARCSTAT_BUMP(arcstat_misses);
6351 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6352 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data,
6354 zfs_racct_read(size, 1);
6357 /* Check if the spa even has l2 configured */
6358 const boolean_t spa_has_l2 = l2arc_ndev != 0 &&
6359 spa->spa_l2cache.sav_count > 0;
6361 if (vd != NULL && spa_has_l2 && !(l2arc_norw && devw)) {
6363 * Read from the L2ARC if the following are true:
6364 * 1. The L2ARC vdev was previously cached.
6365 * 2. This buffer still has L2ARC metadata.
6366 * 3. This buffer isn't currently writing to the L2ARC.
6367 * 4. The L2ARC entry wasn't evicted, which may
6368 * also have invalidated the vdev.
6369 * 5. This isn't prefetch or l2arc_noprefetch is 0.
6371 if (HDR_HAS_L2HDR(hdr) &&
6372 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
6373 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
6374 l2arc_read_callback_t *cb;
6378 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
6379 ARCSTAT_BUMP(arcstat_l2_hits);
6380 hdr->b_l2hdr.b_hits++;
6382 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
6384 cb->l2rcb_hdr = hdr;
6387 cb->l2rcb_flags = zio_flags;
6390 * When Compressed ARC is disabled, but the
6391 * L2ARC block is compressed, arc_hdr_size()
6392 * will have returned LSIZE rather than PSIZE.
6394 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
6395 !HDR_COMPRESSION_ENABLED(hdr) &&
6396 HDR_GET_PSIZE(hdr) != 0) {
6397 size = HDR_GET_PSIZE(hdr);
6400 asize = vdev_psize_to_asize(vd, size);
6401 if (asize != size) {
6402 abd = abd_alloc_for_io(asize,
6403 HDR_ISTYPE_METADATA(hdr));
6404 cb->l2rcb_abd = abd;
6409 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6410 addr + asize <= vd->vdev_psize -
6411 VDEV_LABEL_END_SIZE);
6414 * l2arc read. The SCL_L2ARC lock will be
6415 * released by l2arc_read_done().
6416 * Issue a null zio if the underlying buffer
6417 * was squashed to zero size by compression.
6419 ASSERT3U(arc_hdr_get_compress(hdr), !=,
6420 ZIO_COMPRESS_EMPTY);
6421 rzio = zio_read_phys(pio, vd, addr,
6424 l2arc_read_done, cb, priority,
6425 zio_flags | ZIO_FLAG_DONT_CACHE |
6427 ZIO_FLAG_DONT_PROPAGATE |
6428 ZIO_FLAG_DONT_RETRY, B_FALSE);
6429 acb->acb_zio_head = rzio;
6431 if (hash_lock != NULL)
6432 mutex_exit(hash_lock);
6434 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6436 ARCSTAT_INCR(arcstat_l2_read_bytes,
6437 HDR_GET_PSIZE(hdr));
6439 if (*arc_flags & ARC_FLAG_NOWAIT) {
6444 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6445 if (zio_wait(rzio) == 0)
6448 /* l2arc read error; goto zio_read() */
6449 if (hash_lock != NULL)
6450 mutex_enter(hash_lock);
6452 DTRACE_PROBE1(l2arc__miss,
6453 arc_buf_hdr_t *, hdr);
6454 ARCSTAT_BUMP(arcstat_l2_misses);
6455 if (HDR_L2_WRITING(hdr))
6456 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6457 spa_config_exit(spa, SCL_L2ARC, vd);
6461 spa_config_exit(spa, SCL_L2ARC, vd);
6464 * Only a spa with l2 should contribute to l2
6465 * miss stats. (Including the case of having a
6466 * faulted cache device - that's also a miss.)
6470 * Skip ARC stat bump for block pointers with
6471 * embedded data. The data are read from the
6473 * decode_embedded_bp_compressed().
6476 DTRACE_PROBE1(l2arc__miss,
6477 arc_buf_hdr_t *, hdr);
6478 ARCSTAT_BUMP(arcstat_l2_misses);
6483 rzio = zio_read(pio, spa, bp, hdr_abd, size,
6484 arc_read_done, hdr, priority, zio_flags, zb);
6485 acb->acb_zio_head = rzio;
6487 if (hash_lock != NULL)
6488 mutex_exit(hash_lock);
6490 if (*arc_flags & ARC_FLAG_WAIT) {
6491 rc = zio_wait(rzio);
6495 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6500 /* embedded bps don't actually go to disk */
6502 spa_read_history_add(spa, zb, *arc_flags);
6503 spl_fstrans_unmark(cookie);
6508 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6512 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6514 p->p_private = private;
6515 list_link_init(&p->p_node);
6516 zfs_refcount_create(&p->p_refcnt);
6518 mutex_enter(&arc_prune_mtx);
6519 zfs_refcount_add(&p->p_refcnt, &arc_prune_list);
6520 list_insert_head(&arc_prune_list, p);
6521 mutex_exit(&arc_prune_mtx);
6527 arc_remove_prune_callback(arc_prune_t *p)
6529 boolean_t wait = B_FALSE;
6530 mutex_enter(&arc_prune_mtx);
6531 list_remove(&arc_prune_list, p);
6532 if (zfs_refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6534 mutex_exit(&arc_prune_mtx);
6536 /* wait for arc_prune_task to finish */
6538 taskq_wait_outstanding(arc_prune_taskq, 0);
6539 ASSERT0(zfs_refcount_count(&p->p_refcnt));
6540 zfs_refcount_destroy(&p->p_refcnt);
6541 kmem_free(p, sizeof (*p));
6545 * Notify the arc that a block was freed, and thus will never be used again.
6548 arc_freed(spa_t *spa, const blkptr_t *bp)
6551 kmutex_t *hash_lock;
6552 uint64_t guid = spa_load_guid(spa);
6554 ASSERT(!BP_IS_EMBEDDED(bp));
6556 hdr = buf_hash_find(guid, bp, &hash_lock);
6561 * We might be trying to free a block that is still doing I/O
6562 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6563 * dmu_sync-ed block). If this block is being prefetched, then it
6564 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6565 * until the I/O completes. A block may also have a reference if it is
6566 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6567 * have written the new block to its final resting place on disk but
6568 * without the dedup flag set. This would have left the hdr in the MRU
6569 * state and discoverable. When the txg finally syncs it detects that
6570 * the block was overridden in open context and issues an override I/O.
6571 * Since this is a dedup block, the override I/O will determine if the
6572 * block is already in the DDT. If so, then it will replace the io_bp
6573 * with the bp from the DDT and allow the I/O to finish. When the I/O
6574 * reaches the done callback, dbuf_write_override_done, it will
6575 * check to see if the io_bp and io_bp_override are identical.
6576 * If they are not, then it indicates that the bp was replaced with
6577 * the bp in the DDT and the override bp is freed. This allows
6578 * us to arrive here with a reference on a block that is being
6579 * freed. So if we have an I/O in progress, or a reference to
6580 * this hdr, then we don't destroy the hdr.
6582 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6583 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6584 arc_change_state(arc_anon, hdr, hash_lock);
6585 arc_hdr_destroy(hdr);
6586 mutex_exit(hash_lock);
6588 mutex_exit(hash_lock);
6594 * Release this buffer from the cache, making it an anonymous buffer. This
6595 * must be done after a read and prior to modifying the buffer contents.
6596 * If the buffer has more than one reference, we must make
6597 * a new hdr for the buffer.
6600 arc_release(arc_buf_t *buf, const void *tag)
6602 arc_buf_hdr_t *hdr = buf->b_hdr;
6605 * It would be nice to assert that if its DMU metadata (level >
6606 * 0 || it's the dnode file), then it must be syncing context.
6607 * But we don't know that information at this level.
6610 mutex_enter(&buf->b_evict_lock);
6612 ASSERT(HDR_HAS_L1HDR(hdr));
6615 * We don't grab the hash lock prior to this check, because if
6616 * the buffer's header is in the arc_anon state, it won't be
6617 * linked into the hash table.
6619 if (hdr->b_l1hdr.b_state == arc_anon) {
6620 mutex_exit(&buf->b_evict_lock);
6621 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6622 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6623 ASSERT(!HDR_HAS_L2HDR(hdr));
6625 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6626 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6627 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6629 hdr->b_l1hdr.b_arc_access = 0;
6632 * If the buf is being overridden then it may already
6633 * have a hdr that is not empty.
6635 buf_discard_identity(hdr);
6641 kmutex_t *hash_lock = HDR_LOCK(hdr);
6642 mutex_enter(hash_lock);
6645 * This assignment is only valid as long as the hash_lock is
6646 * held, we must be careful not to reference state or the
6647 * b_state field after dropping the lock.
6649 arc_state_t *state = hdr->b_l1hdr.b_state;
6650 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6651 ASSERT3P(state, !=, arc_anon);
6653 /* this buffer is not on any list */
6654 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6656 if (HDR_HAS_L2HDR(hdr)) {
6657 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6660 * We have to recheck this conditional again now that
6661 * we're holding the l2ad_mtx to prevent a race with
6662 * another thread which might be concurrently calling
6663 * l2arc_evict(). In that case, l2arc_evict() might have
6664 * destroyed the header's L2 portion as we were waiting
6665 * to acquire the l2ad_mtx.
6667 if (HDR_HAS_L2HDR(hdr))
6668 arc_hdr_l2hdr_destroy(hdr);
6670 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6674 * Do we have more than one buf?
6676 if (hdr->b_l1hdr.b_bufcnt > 1) {
6677 arc_buf_hdr_t *nhdr;
6678 uint64_t spa = hdr->b_spa;
6679 uint64_t psize = HDR_GET_PSIZE(hdr);
6680 uint64_t lsize = HDR_GET_LSIZE(hdr);
6681 boolean_t protected = HDR_PROTECTED(hdr);
6682 enum zio_compress compress = arc_hdr_get_compress(hdr);
6683 arc_buf_contents_t type = arc_buf_type(hdr);
6684 VERIFY3U(hdr->b_type, ==, type);
6686 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6687 (void) remove_reference(hdr, hash_lock, tag);
6689 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6690 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6691 ASSERT(ARC_BUF_LAST(buf));
6695 * Pull the data off of this hdr and attach it to
6696 * a new anonymous hdr. Also find the last buffer
6697 * in the hdr's buffer list.
6699 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6700 ASSERT3P(lastbuf, !=, NULL);
6703 * If the current arc_buf_t and the hdr are sharing their data
6704 * buffer, then we must stop sharing that block.
6706 if (arc_buf_is_shared(buf)) {
6707 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6708 VERIFY(!arc_buf_is_shared(lastbuf));
6711 * First, sever the block sharing relationship between
6712 * buf and the arc_buf_hdr_t.
6714 arc_unshare_buf(hdr, buf);
6717 * Now we need to recreate the hdr's b_pabd. Since we
6718 * have lastbuf handy, we try to share with it, but if
6719 * we can't then we allocate a new b_pabd and copy the
6720 * data from buf into it.
6722 if (arc_can_share(hdr, lastbuf)) {
6723 arc_share_buf(hdr, lastbuf);
6725 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT);
6726 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6727 buf->b_data, psize);
6729 VERIFY3P(lastbuf->b_data, !=, NULL);
6730 } else if (HDR_SHARED_DATA(hdr)) {
6732 * Uncompressed shared buffers are always at the end
6733 * of the list. Compressed buffers don't have the
6734 * same requirements. This makes it hard to
6735 * simply assert that the lastbuf is shared so
6736 * we rely on the hdr's compression flags to determine
6737 * if we have a compressed, shared buffer.
6739 ASSERT(arc_buf_is_shared(lastbuf) ||
6740 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6741 ASSERT(!ARC_BUF_SHARED(buf));
6744 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6745 ASSERT3P(state, !=, arc_l2c_only);
6747 (void) zfs_refcount_remove_many(&state->arcs_size,
6748 arc_buf_size(buf), buf);
6750 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6751 ASSERT3P(state, !=, arc_l2c_only);
6752 (void) zfs_refcount_remove_many(
6753 &state->arcs_esize[type],
6754 arc_buf_size(buf), buf);
6757 hdr->b_l1hdr.b_bufcnt -= 1;
6758 if (ARC_BUF_ENCRYPTED(buf))
6759 hdr->b_crypt_hdr.b_ebufcnt -= 1;
6761 arc_cksum_verify(buf);
6762 arc_buf_unwatch(buf);
6764 /* if this is the last uncompressed buf free the checksum */
6765 if (!arc_hdr_has_uncompressed_buf(hdr))
6766 arc_cksum_free(hdr);
6768 mutex_exit(hash_lock);
6771 * Allocate a new hdr. The new hdr will contain a b_pabd
6772 * buffer which will be freed in arc_write().
6774 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6775 compress, hdr->b_complevel, type);
6776 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6777 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6778 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6779 VERIFY3U(nhdr->b_type, ==, type);
6780 ASSERT(!HDR_SHARED_DATA(nhdr));
6782 nhdr->b_l1hdr.b_buf = buf;
6783 nhdr->b_l1hdr.b_bufcnt = 1;
6784 if (ARC_BUF_ENCRYPTED(buf))
6785 nhdr->b_crypt_hdr.b_ebufcnt = 1;
6786 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6789 mutex_exit(&buf->b_evict_lock);
6790 (void) zfs_refcount_add_many(&arc_anon->arcs_size,
6791 arc_buf_size(buf), buf);
6793 mutex_exit(&buf->b_evict_lock);
6794 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6795 /* protected by hash lock, or hdr is on arc_anon */
6796 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6797 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6798 hdr->b_l1hdr.b_mru_hits = 0;
6799 hdr->b_l1hdr.b_mru_ghost_hits = 0;
6800 hdr->b_l1hdr.b_mfu_hits = 0;
6801 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
6802 arc_change_state(arc_anon, hdr, hash_lock);
6803 hdr->b_l1hdr.b_arc_access = 0;
6805 mutex_exit(hash_lock);
6806 buf_discard_identity(hdr);
6812 arc_released(arc_buf_t *buf)
6816 mutex_enter(&buf->b_evict_lock);
6817 released = (buf->b_data != NULL &&
6818 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6819 mutex_exit(&buf->b_evict_lock);
6825 arc_referenced(arc_buf_t *buf)
6829 mutex_enter(&buf->b_evict_lock);
6830 referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6831 mutex_exit(&buf->b_evict_lock);
6832 return (referenced);
6837 arc_write_ready(zio_t *zio)
6839 arc_write_callback_t *callback = zio->io_private;
6840 arc_buf_t *buf = callback->awcb_buf;
6841 arc_buf_hdr_t *hdr = buf->b_hdr;
6842 blkptr_t *bp = zio->io_bp;
6843 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6844 fstrans_cookie_t cookie = spl_fstrans_mark();
6846 ASSERT(HDR_HAS_L1HDR(hdr));
6847 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6848 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6851 * If we're reexecuting this zio because the pool suspended, then
6852 * cleanup any state that was previously set the first time the
6853 * callback was invoked.
6855 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6856 arc_cksum_free(hdr);
6857 arc_buf_unwatch(buf);
6858 if (hdr->b_l1hdr.b_pabd != NULL) {
6859 if (arc_buf_is_shared(buf)) {
6860 arc_unshare_buf(hdr, buf);
6862 arc_hdr_free_abd(hdr, B_FALSE);
6866 if (HDR_HAS_RABD(hdr))
6867 arc_hdr_free_abd(hdr, B_TRUE);
6869 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6870 ASSERT(!HDR_HAS_RABD(hdr));
6871 ASSERT(!HDR_SHARED_DATA(hdr));
6872 ASSERT(!arc_buf_is_shared(buf));
6874 callback->awcb_ready(zio, buf, callback->awcb_private);
6876 if (HDR_IO_IN_PROGRESS(hdr))
6877 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6879 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6881 if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr))
6882 hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp));
6884 if (BP_IS_PROTECTED(bp)) {
6885 /* ZIL blocks are written through zio_rewrite */
6886 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6887 ASSERT(HDR_PROTECTED(hdr));
6889 if (BP_SHOULD_BYTESWAP(bp)) {
6890 if (BP_GET_LEVEL(bp) > 0) {
6891 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6893 hdr->b_l1hdr.b_byteswap =
6894 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6897 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6900 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6901 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6902 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6903 hdr->b_crypt_hdr.b_iv);
6904 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6908 * If this block was written for raw encryption but the zio layer
6909 * ended up only authenticating it, adjust the buffer flags now.
6911 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6912 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6913 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6914 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6915 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6916 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6917 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6918 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6921 /* this must be done after the buffer flags are adjusted */
6922 arc_cksum_compute(buf);
6924 enum zio_compress compress;
6925 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6926 compress = ZIO_COMPRESS_OFF;
6928 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6929 compress = BP_GET_COMPRESS(bp);
6931 HDR_SET_PSIZE(hdr, psize);
6932 arc_hdr_set_compress(hdr, compress);
6933 hdr->b_complevel = zio->io_prop.zp_complevel;
6935 if (zio->io_error != 0 || psize == 0)
6939 * Fill the hdr with data. If the buffer is encrypted we have no choice
6940 * but to copy the data into b_radb. If the hdr is compressed, the data
6941 * we want is available from the zio, otherwise we can take it from
6944 * We might be able to share the buf's data with the hdr here. However,
6945 * doing so would cause the ARC to be full of linear ABDs if we write a
6946 * lot of shareable data. As a compromise, we check whether scattered
6947 * ABDs are allowed, and assume that if they are then the user wants
6948 * the ARC to be primarily filled with them regardless of the data being
6949 * written. Therefore, if they're allowed then we allocate one and copy
6950 * the data into it; otherwise, we share the data directly if we can.
6952 if (ARC_BUF_ENCRYPTED(buf)) {
6953 ASSERT3U(psize, >, 0);
6954 ASSERT(ARC_BUF_COMPRESSED(buf));
6955 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT | ARC_HDR_ALLOC_RDATA |
6956 ARC_HDR_USE_RESERVE);
6957 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6958 } else if (!abd_size_alloc_linear(arc_buf_size(buf)) ||
6959 !arc_can_share(hdr, buf)) {
6961 * Ideally, we would always copy the io_abd into b_pabd, but the
6962 * user may have disabled compressed ARC, thus we must check the
6963 * hdr's compression setting rather than the io_bp's.
6965 if (BP_IS_ENCRYPTED(bp)) {
6966 ASSERT3U(psize, >, 0);
6967 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT |
6968 ARC_HDR_ALLOC_RDATA | ARC_HDR_USE_RESERVE);
6969 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6970 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
6971 !ARC_BUF_COMPRESSED(buf)) {
6972 ASSERT3U(psize, >, 0);
6973 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT |
6974 ARC_HDR_USE_RESERVE);
6975 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6977 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6978 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT |
6979 ARC_HDR_USE_RESERVE);
6980 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6984 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6985 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6986 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6988 arc_share_buf(hdr, buf);
6992 arc_hdr_verify(hdr, bp);
6993 spl_fstrans_unmark(cookie);
6997 arc_write_children_ready(zio_t *zio)
6999 arc_write_callback_t *callback = zio->io_private;
7000 arc_buf_t *buf = callback->awcb_buf;
7002 callback->awcb_children_ready(zio, buf, callback->awcb_private);
7006 * The SPA calls this callback for each physical write that happens on behalf
7007 * of a logical write. See the comment in dbuf_write_physdone() for details.
7010 arc_write_physdone(zio_t *zio)
7012 arc_write_callback_t *cb = zio->io_private;
7013 if (cb->awcb_physdone != NULL)
7014 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
7018 arc_write_done(zio_t *zio)
7020 arc_write_callback_t *callback = zio->io_private;
7021 arc_buf_t *buf = callback->awcb_buf;
7022 arc_buf_hdr_t *hdr = buf->b_hdr;
7024 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
7026 if (zio->io_error == 0) {
7027 arc_hdr_verify(hdr, zio->io_bp);
7029 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
7030 buf_discard_identity(hdr);
7032 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
7033 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
7036 ASSERT(HDR_EMPTY(hdr));
7040 * If the block to be written was all-zero or compressed enough to be
7041 * embedded in the BP, no write was performed so there will be no
7042 * dva/birth/checksum. The buffer must therefore remain anonymous
7045 if (!HDR_EMPTY(hdr)) {
7046 arc_buf_hdr_t *exists;
7047 kmutex_t *hash_lock;
7049 ASSERT3U(zio->io_error, ==, 0);
7051 arc_cksum_verify(buf);
7053 exists = buf_hash_insert(hdr, &hash_lock);
7054 if (exists != NULL) {
7056 * This can only happen if we overwrite for
7057 * sync-to-convergence, because we remove
7058 * buffers from the hash table when we arc_free().
7060 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
7061 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
7062 panic("bad overwrite, hdr=%p exists=%p",
7063 (void *)hdr, (void *)exists);
7064 ASSERT(zfs_refcount_is_zero(
7065 &exists->b_l1hdr.b_refcnt));
7066 arc_change_state(arc_anon, exists, hash_lock);
7067 arc_hdr_destroy(exists);
7068 mutex_exit(hash_lock);
7069 exists = buf_hash_insert(hdr, &hash_lock);
7070 ASSERT3P(exists, ==, NULL);
7071 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
7073 ASSERT(zio->io_prop.zp_nopwrite);
7074 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
7075 panic("bad nopwrite, hdr=%p exists=%p",
7076 (void *)hdr, (void *)exists);
7079 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
7080 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
7081 ASSERT(BP_GET_DEDUP(zio->io_bp));
7082 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
7085 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
7086 /* if it's not anon, we are doing a scrub */
7087 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
7088 arc_access(hdr, hash_lock);
7089 mutex_exit(hash_lock);
7091 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
7094 ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
7095 callback->awcb_done(zio, buf, callback->awcb_private);
7097 abd_free(zio->io_abd);
7098 kmem_free(callback, sizeof (arc_write_callback_t));
7102 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
7103 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc,
7104 const zio_prop_t *zp, arc_write_done_func_t *ready,
7105 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
7106 arc_write_done_func_t *done, void *private, zio_priority_t priority,
7107 int zio_flags, const zbookmark_phys_t *zb)
7109 arc_buf_hdr_t *hdr = buf->b_hdr;
7110 arc_write_callback_t *callback;
7112 zio_prop_t localprop = *zp;
7114 ASSERT3P(ready, !=, NULL);
7115 ASSERT3P(done, !=, NULL);
7116 ASSERT(!HDR_IO_ERROR(hdr));
7117 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
7118 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
7119 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
7121 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
7123 if (ARC_BUF_ENCRYPTED(buf)) {
7124 ASSERT(ARC_BUF_COMPRESSED(buf));
7125 localprop.zp_encrypt = B_TRUE;
7126 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
7127 localprop.zp_complevel = hdr->b_complevel;
7128 localprop.zp_byteorder =
7129 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
7130 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
7131 memcpy(localprop.zp_salt, hdr->b_crypt_hdr.b_salt,
7133 memcpy(localprop.zp_iv, hdr->b_crypt_hdr.b_iv,
7135 memcpy(localprop.zp_mac, hdr->b_crypt_hdr.b_mac,
7137 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
7138 localprop.zp_nopwrite = B_FALSE;
7139 localprop.zp_copies =
7140 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
7142 zio_flags |= ZIO_FLAG_RAW;
7143 } else if (ARC_BUF_COMPRESSED(buf)) {
7144 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
7145 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
7146 localprop.zp_complevel = hdr->b_complevel;
7147 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
7149 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
7150 callback->awcb_ready = ready;
7151 callback->awcb_children_ready = children_ready;
7152 callback->awcb_physdone = physdone;
7153 callback->awcb_done = done;
7154 callback->awcb_private = private;
7155 callback->awcb_buf = buf;
7158 * The hdr's b_pabd is now stale, free it now. A new data block
7159 * will be allocated when the zio pipeline calls arc_write_ready().
7161 if (hdr->b_l1hdr.b_pabd != NULL) {
7163 * If the buf is currently sharing the data block with
7164 * the hdr then we need to break that relationship here.
7165 * The hdr will remain with a NULL data pointer and the
7166 * buf will take sole ownership of the block.
7168 if (arc_buf_is_shared(buf)) {
7169 arc_unshare_buf(hdr, buf);
7171 arc_hdr_free_abd(hdr, B_FALSE);
7173 VERIFY3P(buf->b_data, !=, NULL);
7176 if (HDR_HAS_RABD(hdr))
7177 arc_hdr_free_abd(hdr, B_TRUE);
7179 if (!(zio_flags & ZIO_FLAG_RAW))
7180 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
7182 ASSERT(!arc_buf_is_shared(buf));
7183 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
7185 zio = zio_write(pio, spa, txg, bp,
7186 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
7187 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
7188 (children_ready != NULL) ? arc_write_children_ready : NULL,
7189 arc_write_physdone, arc_write_done, callback,
7190 priority, zio_flags, zb);
7196 arc_tempreserve_clear(uint64_t reserve)
7198 atomic_add_64(&arc_tempreserve, -reserve);
7199 ASSERT((int64_t)arc_tempreserve >= 0);
7203 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
7209 reserve > arc_c/4 &&
7210 reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT))
7211 arc_c = MIN(arc_c_max, reserve * 4);
7214 * Throttle when the calculated memory footprint for the TXG
7215 * exceeds the target ARC size.
7217 if (reserve > arc_c) {
7218 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
7219 return (SET_ERROR(ERESTART));
7223 * Don't count loaned bufs as in flight dirty data to prevent long
7224 * network delays from blocking transactions that are ready to be
7225 * assigned to a txg.
7228 /* assert that it has not wrapped around */
7229 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
7231 anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) -
7232 arc_loaned_bytes), 0);
7235 * Writes will, almost always, require additional memory allocations
7236 * in order to compress/encrypt/etc the data. We therefore need to
7237 * make sure that there is sufficient available memory for this.
7239 error = arc_memory_throttle(spa, reserve, txg);
7244 * Throttle writes when the amount of dirty data in the cache
7245 * gets too large. We try to keep the cache less than half full
7246 * of dirty blocks so that our sync times don't grow too large.
7248 * In the case of one pool being built on another pool, we want
7249 * to make sure we don't end up throttling the lower (backing)
7250 * pool when the upper pool is the majority contributor to dirty
7251 * data. To insure we make forward progress during throttling, we
7252 * also check the current pool's net dirty data and only throttle
7253 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
7254 * data in the cache.
7256 * Note: if two requests come in concurrently, we might let them
7257 * both succeed, when one of them should fail. Not a huge deal.
7259 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
7260 uint64_t spa_dirty_anon = spa_dirty_data(spa);
7261 uint64_t rarc_c = arc_warm ? arc_c : arc_c_max;
7262 if (total_dirty > rarc_c * zfs_arc_dirty_limit_percent / 100 &&
7263 anon_size > rarc_c * zfs_arc_anon_limit_percent / 100 &&
7264 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
7266 uint64_t meta_esize = zfs_refcount_count(
7267 &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7268 uint64_t data_esize =
7269 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7270 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
7271 "anon_data=%lluK tempreserve=%lluK rarc_c=%lluK\n",
7272 (u_longlong_t)arc_tempreserve >> 10,
7273 (u_longlong_t)meta_esize >> 10,
7274 (u_longlong_t)data_esize >> 10,
7275 (u_longlong_t)reserve >> 10,
7276 (u_longlong_t)rarc_c >> 10);
7278 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
7279 return (SET_ERROR(ERESTART));
7281 atomic_add_64(&arc_tempreserve, reserve);
7286 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
7287 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
7289 size->value.ui64 = zfs_refcount_count(&state->arcs_size);
7290 evict_data->value.ui64 =
7291 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
7292 evict_metadata->value.ui64 =
7293 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
7297 arc_kstat_update(kstat_t *ksp, int rw)
7299 arc_stats_t *as = ksp->ks_data;
7301 if (rw == KSTAT_WRITE)
7302 return (SET_ERROR(EACCES));
7304 as->arcstat_hits.value.ui64 =
7305 wmsum_value(&arc_sums.arcstat_hits);
7306 as->arcstat_misses.value.ui64 =
7307 wmsum_value(&arc_sums.arcstat_misses);
7308 as->arcstat_demand_data_hits.value.ui64 =
7309 wmsum_value(&arc_sums.arcstat_demand_data_hits);
7310 as->arcstat_demand_data_misses.value.ui64 =
7311 wmsum_value(&arc_sums.arcstat_demand_data_misses);
7312 as->arcstat_demand_metadata_hits.value.ui64 =
7313 wmsum_value(&arc_sums.arcstat_demand_metadata_hits);
7314 as->arcstat_demand_metadata_misses.value.ui64 =
7315 wmsum_value(&arc_sums.arcstat_demand_metadata_misses);
7316 as->arcstat_prefetch_data_hits.value.ui64 =
7317 wmsum_value(&arc_sums.arcstat_prefetch_data_hits);
7318 as->arcstat_prefetch_data_misses.value.ui64 =
7319 wmsum_value(&arc_sums.arcstat_prefetch_data_misses);
7320 as->arcstat_prefetch_metadata_hits.value.ui64 =
7321 wmsum_value(&arc_sums.arcstat_prefetch_metadata_hits);
7322 as->arcstat_prefetch_metadata_misses.value.ui64 =
7323 wmsum_value(&arc_sums.arcstat_prefetch_metadata_misses);
7324 as->arcstat_mru_hits.value.ui64 =
7325 wmsum_value(&arc_sums.arcstat_mru_hits);
7326 as->arcstat_mru_ghost_hits.value.ui64 =
7327 wmsum_value(&arc_sums.arcstat_mru_ghost_hits);
7328 as->arcstat_mfu_hits.value.ui64 =
7329 wmsum_value(&arc_sums.arcstat_mfu_hits);
7330 as->arcstat_mfu_ghost_hits.value.ui64 =
7331 wmsum_value(&arc_sums.arcstat_mfu_ghost_hits);
7332 as->arcstat_deleted.value.ui64 =
7333 wmsum_value(&arc_sums.arcstat_deleted);
7334 as->arcstat_mutex_miss.value.ui64 =
7335 wmsum_value(&arc_sums.arcstat_mutex_miss);
7336 as->arcstat_access_skip.value.ui64 =
7337 wmsum_value(&arc_sums.arcstat_access_skip);
7338 as->arcstat_evict_skip.value.ui64 =
7339 wmsum_value(&arc_sums.arcstat_evict_skip);
7340 as->arcstat_evict_not_enough.value.ui64 =
7341 wmsum_value(&arc_sums.arcstat_evict_not_enough);
7342 as->arcstat_evict_l2_cached.value.ui64 =
7343 wmsum_value(&arc_sums.arcstat_evict_l2_cached);
7344 as->arcstat_evict_l2_eligible.value.ui64 =
7345 wmsum_value(&arc_sums.arcstat_evict_l2_eligible);
7346 as->arcstat_evict_l2_eligible_mfu.value.ui64 =
7347 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mfu);
7348 as->arcstat_evict_l2_eligible_mru.value.ui64 =
7349 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mru);
7350 as->arcstat_evict_l2_ineligible.value.ui64 =
7351 wmsum_value(&arc_sums.arcstat_evict_l2_ineligible);
7352 as->arcstat_evict_l2_skip.value.ui64 =
7353 wmsum_value(&arc_sums.arcstat_evict_l2_skip);
7354 as->arcstat_hash_collisions.value.ui64 =
7355 wmsum_value(&arc_sums.arcstat_hash_collisions);
7356 as->arcstat_hash_chains.value.ui64 =
7357 wmsum_value(&arc_sums.arcstat_hash_chains);
7358 as->arcstat_size.value.ui64 =
7359 aggsum_value(&arc_sums.arcstat_size);
7360 as->arcstat_compressed_size.value.ui64 =
7361 wmsum_value(&arc_sums.arcstat_compressed_size);
7362 as->arcstat_uncompressed_size.value.ui64 =
7363 wmsum_value(&arc_sums.arcstat_uncompressed_size);
7364 as->arcstat_overhead_size.value.ui64 =
7365 wmsum_value(&arc_sums.arcstat_overhead_size);
7366 as->arcstat_hdr_size.value.ui64 =
7367 wmsum_value(&arc_sums.arcstat_hdr_size);
7368 as->arcstat_data_size.value.ui64 =
7369 wmsum_value(&arc_sums.arcstat_data_size);
7370 as->arcstat_metadata_size.value.ui64 =
7371 wmsum_value(&arc_sums.arcstat_metadata_size);
7372 as->arcstat_dbuf_size.value.ui64 =
7373 wmsum_value(&arc_sums.arcstat_dbuf_size);
7374 #if defined(COMPAT_FREEBSD11)
7375 as->arcstat_other_size.value.ui64 =
7376 wmsum_value(&arc_sums.arcstat_bonus_size) +
7377 aggsum_value(&arc_sums.arcstat_dnode_size) +
7378 wmsum_value(&arc_sums.arcstat_dbuf_size);
7381 arc_kstat_update_state(arc_anon,
7382 &as->arcstat_anon_size,
7383 &as->arcstat_anon_evictable_data,
7384 &as->arcstat_anon_evictable_metadata);
7385 arc_kstat_update_state(arc_mru,
7386 &as->arcstat_mru_size,
7387 &as->arcstat_mru_evictable_data,
7388 &as->arcstat_mru_evictable_metadata);
7389 arc_kstat_update_state(arc_mru_ghost,
7390 &as->arcstat_mru_ghost_size,
7391 &as->arcstat_mru_ghost_evictable_data,
7392 &as->arcstat_mru_ghost_evictable_metadata);
7393 arc_kstat_update_state(arc_mfu,
7394 &as->arcstat_mfu_size,
7395 &as->arcstat_mfu_evictable_data,
7396 &as->arcstat_mfu_evictable_metadata);
7397 arc_kstat_update_state(arc_mfu_ghost,
7398 &as->arcstat_mfu_ghost_size,
7399 &as->arcstat_mfu_ghost_evictable_data,
7400 &as->arcstat_mfu_ghost_evictable_metadata);
7402 as->arcstat_dnode_size.value.ui64 =
7403 aggsum_value(&arc_sums.arcstat_dnode_size);
7404 as->arcstat_bonus_size.value.ui64 =
7405 wmsum_value(&arc_sums.arcstat_bonus_size);
7406 as->arcstat_l2_hits.value.ui64 =
7407 wmsum_value(&arc_sums.arcstat_l2_hits);
7408 as->arcstat_l2_misses.value.ui64 =
7409 wmsum_value(&arc_sums.arcstat_l2_misses);
7410 as->arcstat_l2_prefetch_asize.value.ui64 =
7411 wmsum_value(&arc_sums.arcstat_l2_prefetch_asize);
7412 as->arcstat_l2_mru_asize.value.ui64 =
7413 wmsum_value(&arc_sums.arcstat_l2_mru_asize);
7414 as->arcstat_l2_mfu_asize.value.ui64 =
7415 wmsum_value(&arc_sums.arcstat_l2_mfu_asize);
7416 as->arcstat_l2_bufc_data_asize.value.ui64 =
7417 wmsum_value(&arc_sums.arcstat_l2_bufc_data_asize);
7418 as->arcstat_l2_bufc_metadata_asize.value.ui64 =
7419 wmsum_value(&arc_sums.arcstat_l2_bufc_metadata_asize);
7420 as->arcstat_l2_feeds.value.ui64 =
7421 wmsum_value(&arc_sums.arcstat_l2_feeds);
7422 as->arcstat_l2_rw_clash.value.ui64 =
7423 wmsum_value(&arc_sums.arcstat_l2_rw_clash);
7424 as->arcstat_l2_read_bytes.value.ui64 =
7425 wmsum_value(&arc_sums.arcstat_l2_read_bytes);
7426 as->arcstat_l2_write_bytes.value.ui64 =
7427 wmsum_value(&arc_sums.arcstat_l2_write_bytes);
7428 as->arcstat_l2_writes_sent.value.ui64 =
7429 wmsum_value(&arc_sums.arcstat_l2_writes_sent);
7430 as->arcstat_l2_writes_done.value.ui64 =
7431 wmsum_value(&arc_sums.arcstat_l2_writes_done);
7432 as->arcstat_l2_writes_error.value.ui64 =
7433 wmsum_value(&arc_sums.arcstat_l2_writes_error);
7434 as->arcstat_l2_writes_lock_retry.value.ui64 =
7435 wmsum_value(&arc_sums.arcstat_l2_writes_lock_retry);
7436 as->arcstat_l2_evict_lock_retry.value.ui64 =
7437 wmsum_value(&arc_sums.arcstat_l2_evict_lock_retry);
7438 as->arcstat_l2_evict_reading.value.ui64 =
7439 wmsum_value(&arc_sums.arcstat_l2_evict_reading);
7440 as->arcstat_l2_evict_l1cached.value.ui64 =
7441 wmsum_value(&arc_sums.arcstat_l2_evict_l1cached);
7442 as->arcstat_l2_free_on_write.value.ui64 =
7443 wmsum_value(&arc_sums.arcstat_l2_free_on_write);
7444 as->arcstat_l2_abort_lowmem.value.ui64 =
7445 wmsum_value(&arc_sums.arcstat_l2_abort_lowmem);
7446 as->arcstat_l2_cksum_bad.value.ui64 =
7447 wmsum_value(&arc_sums.arcstat_l2_cksum_bad);
7448 as->arcstat_l2_io_error.value.ui64 =
7449 wmsum_value(&arc_sums.arcstat_l2_io_error);
7450 as->arcstat_l2_lsize.value.ui64 =
7451 wmsum_value(&arc_sums.arcstat_l2_lsize);
7452 as->arcstat_l2_psize.value.ui64 =
7453 wmsum_value(&arc_sums.arcstat_l2_psize);
7454 as->arcstat_l2_hdr_size.value.ui64 =
7455 aggsum_value(&arc_sums.arcstat_l2_hdr_size);
7456 as->arcstat_l2_log_blk_writes.value.ui64 =
7457 wmsum_value(&arc_sums.arcstat_l2_log_blk_writes);
7458 as->arcstat_l2_log_blk_asize.value.ui64 =
7459 wmsum_value(&arc_sums.arcstat_l2_log_blk_asize);
7460 as->arcstat_l2_log_blk_count.value.ui64 =
7461 wmsum_value(&arc_sums.arcstat_l2_log_blk_count);
7462 as->arcstat_l2_rebuild_success.value.ui64 =
7463 wmsum_value(&arc_sums.arcstat_l2_rebuild_success);
7464 as->arcstat_l2_rebuild_abort_unsupported.value.ui64 =
7465 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7466 as->arcstat_l2_rebuild_abort_io_errors.value.ui64 =
7467 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7468 as->arcstat_l2_rebuild_abort_dh_errors.value.ui64 =
7469 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7470 as->arcstat_l2_rebuild_abort_cksum_lb_errors.value.ui64 =
7471 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7472 as->arcstat_l2_rebuild_abort_lowmem.value.ui64 =
7473 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7474 as->arcstat_l2_rebuild_size.value.ui64 =
7475 wmsum_value(&arc_sums.arcstat_l2_rebuild_size);
7476 as->arcstat_l2_rebuild_asize.value.ui64 =
7477 wmsum_value(&arc_sums.arcstat_l2_rebuild_asize);
7478 as->arcstat_l2_rebuild_bufs.value.ui64 =
7479 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs);
7480 as->arcstat_l2_rebuild_bufs_precached.value.ui64 =
7481 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7482 as->arcstat_l2_rebuild_log_blks.value.ui64 =
7483 wmsum_value(&arc_sums.arcstat_l2_rebuild_log_blks);
7484 as->arcstat_memory_throttle_count.value.ui64 =
7485 wmsum_value(&arc_sums.arcstat_memory_throttle_count);
7486 as->arcstat_memory_direct_count.value.ui64 =
7487 wmsum_value(&arc_sums.arcstat_memory_direct_count);
7488 as->arcstat_memory_indirect_count.value.ui64 =
7489 wmsum_value(&arc_sums.arcstat_memory_indirect_count);
7491 as->arcstat_memory_all_bytes.value.ui64 =
7493 as->arcstat_memory_free_bytes.value.ui64 =
7495 as->arcstat_memory_available_bytes.value.i64 =
7496 arc_available_memory();
7498 as->arcstat_prune.value.ui64 =
7499 wmsum_value(&arc_sums.arcstat_prune);
7500 as->arcstat_meta_used.value.ui64 =
7501 aggsum_value(&arc_sums.arcstat_meta_used);
7502 as->arcstat_async_upgrade_sync.value.ui64 =
7503 wmsum_value(&arc_sums.arcstat_async_upgrade_sync);
7504 as->arcstat_demand_hit_predictive_prefetch.value.ui64 =
7505 wmsum_value(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7506 as->arcstat_demand_hit_prescient_prefetch.value.ui64 =
7507 wmsum_value(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7508 as->arcstat_raw_size.value.ui64 =
7509 wmsum_value(&arc_sums.arcstat_raw_size);
7510 as->arcstat_cached_only_in_progress.value.ui64 =
7511 wmsum_value(&arc_sums.arcstat_cached_only_in_progress);
7512 as->arcstat_abd_chunk_waste_size.value.ui64 =
7513 wmsum_value(&arc_sums.arcstat_abd_chunk_waste_size);
7519 * This function *must* return indices evenly distributed between all
7520 * sublists of the multilist. This is needed due to how the ARC eviction
7521 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7522 * distributed between all sublists and uses this assumption when
7523 * deciding which sublist to evict from and how much to evict from it.
7526 arc_state_multilist_index_func(multilist_t *ml, void *obj)
7528 arc_buf_hdr_t *hdr = obj;
7531 * We rely on b_dva to generate evenly distributed index
7532 * numbers using buf_hash below. So, as an added precaution,
7533 * let's make sure we never add empty buffers to the arc lists.
7535 ASSERT(!HDR_EMPTY(hdr));
7538 * The assumption here, is the hash value for a given
7539 * arc_buf_hdr_t will remain constant throughout its lifetime
7540 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7541 * Thus, we don't need to store the header's sublist index
7542 * on insertion, as this index can be recalculated on removal.
7544 * Also, the low order bits of the hash value are thought to be
7545 * distributed evenly. Otherwise, in the case that the multilist
7546 * has a power of two number of sublists, each sublists' usage
7547 * would not be evenly distributed. In this context full 64bit
7548 * division would be a waste of time, so limit it to 32 bits.
7550 return ((unsigned int)buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
7551 multilist_get_num_sublists(ml));
7555 arc_state_l2c_multilist_index_func(multilist_t *ml, void *obj)
7557 panic("Header %p insert into arc_l2c_only %p", obj, ml);
7560 #define WARN_IF_TUNING_IGNORED(tuning, value, do_warn) do { \
7561 if ((do_warn) && (tuning) && ((tuning) != (value))) { \
7563 "ignoring tunable %s (using %llu instead)", \
7564 (#tuning), (u_longlong_t)(value)); \
7569 * Called during module initialization and periodically thereafter to
7570 * apply reasonable changes to the exposed performance tunings. Can also be
7571 * called explicitly by param_set_arc_*() functions when ARC tunables are
7572 * updated manually. Non-zero zfs_* values which differ from the currently set
7573 * values will be applied.
7576 arc_tuning_update(boolean_t verbose)
7578 uint64_t allmem = arc_all_memory();
7579 unsigned long limit;
7581 /* Valid range: 32M - <arc_c_max> */
7582 if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) &&
7583 (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) &&
7584 (zfs_arc_min <= arc_c_max)) {
7585 arc_c_min = zfs_arc_min;
7586 arc_c = MAX(arc_c, arc_c_min);
7588 WARN_IF_TUNING_IGNORED(zfs_arc_min, arc_c_min, verbose);
7590 /* Valid range: 64M - <all physical memory> */
7591 if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) &&
7592 (zfs_arc_max >= MIN_ARC_MAX) && (zfs_arc_max < allmem) &&
7593 (zfs_arc_max > arc_c_min)) {
7594 arc_c_max = zfs_arc_max;
7595 arc_c = MIN(arc_c, arc_c_max);
7596 arc_p = (arc_c >> 1);
7597 if (arc_meta_limit > arc_c_max)
7598 arc_meta_limit = arc_c_max;
7599 if (arc_dnode_size_limit > arc_meta_limit)
7600 arc_dnode_size_limit = arc_meta_limit;
7602 WARN_IF_TUNING_IGNORED(zfs_arc_max, arc_c_max, verbose);
7604 /* Valid range: 16M - <arc_c_max> */
7605 if ((zfs_arc_meta_min) && (zfs_arc_meta_min != arc_meta_min) &&
7606 (zfs_arc_meta_min >= 1ULL << SPA_MAXBLOCKSHIFT) &&
7607 (zfs_arc_meta_min <= arc_c_max)) {
7608 arc_meta_min = zfs_arc_meta_min;
7609 if (arc_meta_limit < arc_meta_min)
7610 arc_meta_limit = arc_meta_min;
7611 if (arc_dnode_size_limit < arc_meta_min)
7612 arc_dnode_size_limit = arc_meta_min;
7614 WARN_IF_TUNING_IGNORED(zfs_arc_meta_min, arc_meta_min, verbose);
7616 /* Valid range: <arc_meta_min> - <arc_c_max> */
7617 limit = zfs_arc_meta_limit ? zfs_arc_meta_limit :
7618 MIN(zfs_arc_meta_limit_percent, 100) * arc_c_max / 100;
7619 if ((limit != arc_meta_limit) &&
7620 (limit >= arc_meta_min) &&
7621 (limit <= arc_c_max))
7622 arc_meta_limit = limit;
7623 WARN_IF_TUNING_IGNORED(zfs_arc_meta_limit, arc_meta_limit, verbose);
7625 /* Valid range: <arc_meta_min> - <arc_meta_limit> */
7626 limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit :
7627 MIN(zfs_arc_dnode_limit_percent, 100) * arc_meta_limit / 100;
7628 if ((limit != arc_dnode_size_limit) &&
7629 (limit >= arc_meta_min) &&
7630 (limit <= arc_meta_limit))
7631 arc_dnode_size_limit = limit;
7632 WARN_IF_TUNING_IGNORED(zfs_arc_dnode_limit, arc_dnode_size_limit,
7635 /* Valid range: 1 - N */
7636 if (zfs_arc_grow_retry)
7637 arc_grow_retry = zfs_arc_grow_retry;
7639 /* Valid range: 1 - N */
7640 if (zfs_arc_shrink_shift) {
7641 arc_shrink_shift = zfs_arc_shrink_shift;
7642 arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1);
7645 /* Valid range: 1 - N */
7646 if (zfs_arc_p_min_shift)
7647 arc_p_min_shift = zfs_arc_p_min_shift;
7649 /* Valid range: 1 - N ms */
7650 if (zfs_arc_min_prefetch_ms)
7651 arc_min_prefetch_ms = zfs_arc_min_prefetch_ms;
7653 /* Valid range: 1 - N ms */
7654 if (zfs_arc_min_prescient_prefetch_ms) {
7655 arc_min_prescient_prefetch_ms =
7656 zfs_arc_min_prescient_prefetch_ms;
7659 /* Valid range: 0 - 100 */
7660 if (zfs_arc_lotsfree_percent <= 100)
7661 arc_lotsfree_percent = zfs_arc_lotsfree_percent;
7662 WARN_IF_TUNING_IGNORED(zfs_arc_lotsfree_percent, arc_lotsfree_percent,
7665 /* Valid range: 0 - <all physical memory> */
7666 if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free))
7667 arc_sys_free = MIN(zfs_arc_sys_free, allmem);
7668 WARN_IF_TUNING_IGNORED(zfs_arc_sys_free, arc_sys_free, verbose);
7672 arc_state_multilist_init(multilist_t *ml,
7673 multilist_sublist_index_func_t *index_func, int *maxcountp)
7675 multilist_create(ml, sizeof (arc_buf_hdr_t),
7676 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), index_func);
7677 *maxcountp = MAX(*maxcountp, multilist_get_num_sublists(ml));
7681 arc_state_init(void)
7683 int num_sublists = 0;
7685 arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_METADATA],
7686 arc_state_multilist_index_func, &num_sublists);
7687 arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_DATA],
7688 arc_state_multilist_index_func, &num_sublists);
7689 arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
7690 arc_state_multilist_index_func, &num_sublists);
7691 arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
7692 arc_state_multilist_index_func, &num_sublists);
7693 arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
7694 arc_state_multilist_index_func, &num_sublists);
7695 arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_DATA],
7696 arc_state_multilist_index_func, &num_sublists);
7697 arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
7698 arc_state_multilist_index_func, &num_sublists);
7699 arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
7700 arc_state_multilist_index_func, &num_sublists);
7703 * L2 headers should never be on the L2 state list since they don't
7704 * have L1 headers allocated. Special index function asserts that.
7706 arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
7707 arc_state_l2c_multilist_index_func, &num_sublists);
7708 arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
7709 arc_state_l2c_multilist_index_func, &num_sublists);
7712 * Keep track of the number of markers needed to reclaim buffers from
7713 * any ARC state. The markers will be pre-allocated so as to minimize
7714 * the number of memory allocations performed by the eviction thread.
7716 arc_state_evict_marker_count = num_sublists;
7718 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7719 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7720 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7721 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7722 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7723 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7724 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7725 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7726 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7727 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7728 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7729 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7731 zfs_refcount_create(&arc_anon->arcs_size);
7732 zfs_refcount_create(&arc_mru->arcs_size);
7733 zfs_refcount_create(&arc_mru_ghost->arcs_size);
7734 zfs_refcount_create(&arc_mfu->arcs_size);
7735 zfs_refcount_create(&arc_mfu_ghost->arcs_size);
7736 zfs_refcount_create(&arc_l2c_only->arcs_size);
7738 wmsum_init(&arc_sums.arcstat_hits, 0);
7739 wmsum_init(&arc_sums.arcstat_misses, 0);
7740 wmsum_init(&arc_sums.arcstat_demand_data_hits, 0);
7741 wmsum_init(&arc_sums.arcstat_demand_data_misses, 0);
7742 wmsum_init(&arc_sums.arcstat_demand_metadata_hits, 0);
7743 wmsum_init(&arc_sums.arcstat_demand_metadata_misses, 0);
7744 wmsum_init(&arc_sums.arcstat_prefetch_data_hits, 0);
7745 wmsum_init(&arc_sums.arcstat_prefetch_data_misses, 0);
7746 wmsum_init(&arc_sums.arcstat_prefetch_metadata_hits, 0);
7747 wmsum_init(&arc_sums.arcstat_prefetch_metadata_misses, 0);
7748 wmsum_init(&arc_sums.arcstat_mru_hits, 0);
7749 wmsum_init(&arc_sums.arcstat_mru_ghost_hits, 0);
7750 wmsum_init(&arc_sums.arcstat_mfu_hits, 0);
7751 wmsum_init(&arc_sums.arcstat_mfu_ghost_hits, 0);
7752 wmsum_init(&arc_sums.arcstat_deleted, 0);
7753 wmsum_init(&arc_sums.arcstat_mutex_miss, 0);
7754 wmsum_init(&arc_sums.arcstat_access_skip, 0);
7755 wmsum_init(&arc_sums.arcstat_evict_skip, 0);
7756 wmsum_init(&arc_sums.arcstat_evict_not_enough, 0);
7757 wmsum_init(&arc_sums.arcstat_evict_l2_cached, 0);
7758 wmsum_init(&arc_sums.arcstat_evict_l2_eligible, 0);
7759 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mfu, 0);
7760 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mru, 0);
7761 wmsum_init(&arc_sums.arcstat_evict_l2_ineligible, 0);
7762 wmsum_init(&arc_sums.arcstat_evict_l2_skip, 0);
7763 wmsum_init(&arc_sums.arcstat_hash_collisions, 0);
7764 wmsum_init(&arc_sums.arcstat_hash_chains, 0);
7765 aggsum_init(&arc_sums.arcstat_size, 0);
7766 wmsum_init(&arc_sums.arcstat_compressed_size, 0);
7767 wmsum_init(&arc_sums.arcstat_uncompressed_size, 0);
7768 wmsum_init(&arc_sums.arcstat_overhead_size, 0);
7769 wmsum_init(&arc_sums.arcstat_hdr_size, 0);
7770 wmsum_init(&arc_sums.arcstat_data_size, 0);
7771 wmsum_init(&arc_sums.arcstat_metadata_size, 0);
7772 wmsum_init(&arc_sums.arcstat_dbuf_size, 0);
7773 aggsum_init(&arc_sums.arcstat_dnode_size, 0);
7774 wmsum_init(&arc_sums.arcstat_bonus_size, 0);
7775 wmsum_init(&arc_sums.arcstat_l2_hits, 0);
7776 wmsum_init(&arc_sums.arcstat_l2_misses, 0);
7777 wmsum_init(&arc_sums.arcstat_l2_prefetch_asize, 0);
7778 wmsum_init(&arc_sums.arcstat_l2_mru_asize, 0);
7779 wmsum_init(&arc_sums.arcstat_l2_mfu_asize, 0);
7780 wmsum_init(&arc_sums.arcstat_l2_bufc_data_asize, 0);
7781 wmsum_init(&arc_sums.arcstat_l2_bufc_metadata_asize, 0);
7782 wmsum_init(&arc_sums.arcstat_l2_feeds, 0);
7783 wmsum_init(&arc_sums.arcstat_l2_rw_clash, 0);
7784 wmsum_init(&arc_sums.arcstat_l2_read_bytes, 0);
7785 wmsum_init(&arc_sums.arcstat_l2_write_bytes, 0);
7786 wmsum_init(&arc_sums.arcstat_l2_writes_sent, 0);
7787 wmsum_init(&arc_sums.arcstat_l2_writes_done, 0);
7788 wmsum_init(&arc_sums.arcstat_l2_writes_error, 0);
7789 wmsum_init(&arc_sums.arcstat_l2_writes_lock_retry, 0);
7790 wmsum_init(&arc_sums.arcstat_l2_evict_lock_retry, 0);
7791 wmsum_init(&arc_sums.arcstat_l2_evict_reading, 0);
7792 wmsum_init(&arc_sums.arcstat_l2_evict_l1cached, 0);
7793 wmsum_init(&arc_sums.arcstat_l2_free_on_write, 0);
7794 wmsum_init(&arc_sums.arcstat_l2_abort_lowmem, 0);
7795 wmsum_init(&arc_sums.arcstat_l2_cksum_bad, 0);
7796 wmsum_init(&arc_sums.arcstat_l2_io_error, 0);
7797 wmsum_init(&arc_sums.arcstat_l2_lsize, 0);
7798 wmsum_init(&arc_sums.arcstat_l2_psize, 0);
7799 aggsum_init(&arc_sums.arcstat_l2_hdr_size, 0);
7800 wmsum_init(&arc_sums.arcstat_l2_log_blk_writes, 0);
7801 wmsum_init(&arc_sums.arcstat_l2_log_blk_asize, 0);
7802 wmsum_init(&arc_sums.arcstat_l2_log_blk_count, 0);
7803 wmsum_init(&arc_sums.arcstat_l2_rebuild_success, 0);
7804 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_unsupported, 0);
7805 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_io_errors, 0);
7806 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_dh_errors, 0);
7807 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors, 0);
7808 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_lowmem, 0);
7809 wmsum_init(&arc_sums.arcstat_l2_rebuild_size, 0);
7810 wmsum_init(&arc_sums.arcstat_l2_rebuild_asize, 0);
7811 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs, 0);
7812 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs_precached, 0);
7813 wmsum_init(&arc_sums.arcstat_l2_rebuild_log_blks, 0);
7814 wmsum_init(&arc_sums.arcstat_memory_throttle_count, 0);
7815 wmsum_init(&arc_sums.arcstat_memory_direct_count, 0);
7816 wmsum_init(&arc_sums.arcstat_memory_indirect_count, 0);
7817 wmsum_init(&arc_sums.arcstat_prune, 0);
7818 aggsum_init(&arc_sums.arcstat_meta_used, 0);
7819 wmsum_init(&arc_sums.arcstat_async_upgrade_sync, 0);
7820 wmsum_init(&arc_sums.arcstat_demand_hit_predictive_prefetch, 0);
7821 wmsum_init(&arc_sums.arcstat_demand_hit_prescient_prefetch, 0);
7822 wmsum_init(&arc_sums.arcstat_raw_size, 0);
7823 wmsum_init(&arc_sums.arcstat_cached_only_in_progress, 0);
7824 wmsum_init(&arc_sums.arcstat_abd_chunk_waste_size, 0);
7826 arc_anon->arcs_state = ARC_STATE_ANON;
7827 arc_mru->arcs_state = ARC_STATE_MRU;
7828 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7829 arc_mfu->arcs_state = ARC_STATE_MFU;
7830 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7831 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7835 arc_state_fini(void)
7837 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7838 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7839 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7840 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7841 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7842 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7843 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7844 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7845 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7846 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7847 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7848 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7850 zfs_refcount_destroy(&arc_anon->arcs_size);
7851 zfs_refcount_destroy(&arc_mru->arcs_size);
7852 zfs_refcount_destroy(&arc_mru_ghost->arcs_size);
7853 zfs_refcount_destroy(&arc_mfu->arcs_size);
7854 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
7855 zfs_refcount_destroy(&arc_l2c_only->arcs_size);
7857 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
7858 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7859 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7860 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7861 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
7862 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7863 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
7864 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7865 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7866 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7868 wmsum_fini(&arc_sums.arcstat_hits);
7869 wmsum_fini(&arc_sums.arcstat_misses);
7870 wmsum_fini(&arc_sums.arcstat_demand_data_hits);
7871 wmsum_fini(&arc_sums.arcstat_demand_data_misses);
7872 wmsum_fini(&arc_sums.arcstat_demand_metadata_hits);
7873 wmsum_fini(&arc_sums.arcstat_demand_metadata_misses);
7874 wmsum_fini(&arc_sums.arcstat_prefetch_data_hits);
7875 wmsum_fini(&arc_sums.arcstat_prefetch_data_misses);
7876 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_hits);
7877 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_misses);
7878 wmsum_fini(&arc_sums.arcstat_mru_hits);
7879 wmsum_fini(&arc_sums.arcstat_mru_ghost_hits);
7880 wmsum_fini(&arc_sums.arcstat_mfu_hits);
7881 wmsum_fini(&arc_sums.arcstat_mfu_ghost_hits);
7882 wmsum_fini(&arc_sums.arcstat_deleted);
7883 wmsum_fini(&arc_sums.arcstat_mutex_miss);
7884 wmsum_fini(&arc_sums.arcstat_access_skip);
7885 wmsum_fini(&arc_sums.arcstat_evict_skip);
7886 wmsum_fini(&arc_sums.arcstat_evict_not_enough);
7887 wmsum_fini(&arc_sums.arcstat_evict_l2_cached);
7888 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible);
7889 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mfu);
7890 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mru);
7891 wmsum_fini(&arc_sums.arcstat_evict_l2_ineligible);
7892 wmsum_fini(&arc_sums.arcstat_evict_l2_skip);
7893 wmsum_fini(&arc_sums.arcstat_hash_collisions);
7894 wmsum_fini(&arc_sums.arcstat_hash_chains);
7895 aggsum_fini(&arc_sums.arcstat_size);
7896 wmsum_fini(&arc_sums.arcstat_compressed_size);
7897 wmsum_fini(&arc_sums.arcstat_uncompressed_size);
7898 wmsum_fini(&arc_sums.arcstat_overhead_size);
7899 wmsum_fini(&arc_sums.arcstat_hdr_size);
7900 wmsum_fini(&arc_sums.arcstat_data_size);
7901 wmsum_fini(&arc_sums.arcstat_metadata_size);
7902 wmsum_fini(&arc_sums.arcstat_dbuf_size);
7903 aggsum_fini(&arc_sums.arcstat_dnode_size);
7904 wmsum_fini(&arc_sums.arcstat_bonus_size);
7905 wmsum_fini(&arc_sums.arcstat_l2_hits);
7906 wmsum_fini(&arc_sums.arcstat_l2_misses);
7907 wmsum_fini(&arc_sums.arcstat_l2_prefetch_asize);
7908 wmsum_fini(&arc_sums.arcstat_l2_mru_asize);
7909 wmsum_fini(&arc_sums.arcstat_l2_mfu_asize);
7910 wmsum_fini(&arc_sums.arcstat_l2_bufc_data_asize);
7911 wmsum_fini(&arc_sums.arcstat_l2_bufc_metadata_asize);
7912 wmsum_fini(&arc_sums.arcstat_l2_feeds);
7913 wmsum_fini(&arc_sums.arcstat_l2_rw_clash);
7914 wmsum_fini(&arc_sums.arcstat_l2_read_bytes);
7915 wmsum_fini(&arc_sums.arcstat_l2_write_bytes);
7916 wmsum_fini(&arc_sums.arcstat_l2_writes_sent);
7917 wmsum_fini(&arc_sums.arcstat_l2_writes_done);
7918 wmsum_fini(&arc_sums.arcstat_l2_writes_error);
7919 wmsum_fini(&arc_sums.arcstat_l2_writes_lock_retry);
7920 wmsum_fini(&arc_sums.arcstat_l2_evict_lock_retry);
7921 wmsum_fini(&arc_sums.arcstat_l2_evict_reading);
7922 wmsum_fini(&arc_sums.arcstat_l2_evict_l1cached);
7923 wmsum_fini(&arc_sums.arcstat_l2_free_on_write);
7924 wmsum_fini(&arc_sums.arcstat_l2_abort_lowmem);
7925 wmsum_fini(&arc_sums.arcstat_l2_cksum_bad);
7926 wmsum_fini(&arc_sums.arcstat_l2_io_error);
7927 wmsum_fini(&arc_sums.arcstat_l2_lsize);
7928 wmsum_fini(&arc_sums.arcstat_l2_psize);
7929 aggsum_fini(&arc_sums.arcstat_l2_hdr_size);
7930 wmsum_fini(&arc_sums.arcstat_l2_log_blk_writes);
7931 wmsum_fini(&arc_sums.arcstat_l2_log_blk_asize);
7932 wmsum_fini(&arc_sums.arcstat_l2_log_blk_count);
7933 wmsum_fini(&arc_sums.arcstat_l2_rebuild_success);
7934 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7935 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7936 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7937 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7938 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7939 wmsum_fini(&arc_sums.arcstat_l2_rebuild_size);
7940 wmsum_fini(&arc_sums.arcstat_l2_rebuild_asize);
7941 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs);
7942 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7943 wmsum_fini(&arc_sums.arcstat_l2_rebuild_log_blks);
7944 wmsum_fini(&arc_sums.arcstat_memory_throttle_count);
7945 wmsum_fini(&arc_sums.arcstat_memory_direct_count);
7946 wmsum_fini(&arc_sums.arcstat_memory_indirect_count);
7947 wmsum_fini(&arc_sums.arcstat_prune);
7948 aggsum_fini(&arc_sums.arcstat_meta_used);
7949 wmsum_fini(&arc_sums.arcstat_async_upgrade_sync);
7950 wmsum_fini(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7951 wmsum_fini(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7952 wmsum_fini(&arc_sums.arcstat_raw_size);
7953 wmsum_fini(&arc_sums.arcstat_cached_only_in_progress);
7954 wmsum_fini(&arc_sums.arcstat_abd_chunk_waste_size);
7958 arc_target_bytes(void)
7964 arc_set_limits(uint64_t allmem)
7966 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more. */
7967 arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT);
7969 /* How to set default max varies by platform. */
7970 arc_c_max = arc_default_max(arc_c_min, allmem);
7975 uint64_t percent, allmem = arc_all_memory();
7976 mutex_init(&arc_evict_lock, NULL, MUTEX_DEFAULT, NULL);
7977 list_create(&arc_evict_waiters, sizeof (arc_evict_waiter_t),
7978 offsetof(arc_evict_waiter_t, aew_node));
7980 arc_min_prefetch_ms = 1000;
7981 arc_min_prescient_prefetch_ms = 6000;
7983 #if defined(_KERNEL)
7987 arc_set_limits(allmem);
7991 * If zfs_arc_max is non-zero at init, meaning it was set in the kernel
7992 * environment before the module was loaded, don't block setting the
7993 * maximum because it is less than arc_c_min, instead, reset arc_c_min
7995 * zfs_arc_min will be handled by arc_tuning_update().
7997 if (zfs_arc_max != 0 && zfs_arc_max >= MIN_ARC_MAX &&
7998 zfs_arc_max < allmem) {
7999 arc_c_max = zfs_arc_max;
8000 if (arc_c_min >= arc_c_max) {
8001 arc_c_min = MAX(zfs_arc_max / 2,
8002 2ULL << SPA_MAXBLOCKSHIFT);
8007 * In userland, there's only the memory pressure that we artificially
8008 * create (see arc_available_memory()). Don't let arc_c get too
8009 * small, because it can cause transactions to be larger than
8010 * arc_c, causing arc_tempreserve_space() to fail.
8012 arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT);
8016 arc_p = (arc_c >> 1);
8018 /* Set min to 1/2 of arc_c_min */
8019 arc_meta_min = 1ULL << SPA_MAXBLOCKSHIFT;
8021 * Set arc_meta_limit to a percent of arc_c_max with a floor of
8022 * arc_meta_min, and a ceiling of arc_c_max.
8024 percent = MIN(zfs_arc_meta_limit_percent, 100);
8025 arc_meta_limit = MAX(arc_meta_min, (percent * arc_c_max) / 100);
8026 percent = MIN(zfs_arc_dnode_limit_percent, 100);
8027 arc_dnode_size_limit = (percent * arc_meta_limit) / 100;
8029 /* Apply user specified tunings */
8030 arc_tuning_update(B_TRUE);
8032 /* if kmem_flags are set, lets try to use less memory */
8033 if (kmem_debugging())
8035 if (arc_c < arc_c_min)
8038 arc_register_hotplug();
8044 list_create(&arc_prune_list, sizeof (arc_prune_t),
8045 offsetof(arc_prune_t, p_node));
8046 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
8048 arc_prune_taskq = taskq_create("arc_prune", zfs_arc_prune_task_threads,
8049 defclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
8051 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
8052 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
8054 if (arc_ksp != NULL) {
8055 arc_ksp->ks_data = &arc_stats;
8056 arc_ksp->ks_update = arc_kstat_update;
8057 kstat_install(arc_ksp);
8060 arc_state_evict_markers =
8061 arc_state_alloc_markers(arc_state_evict_marker_count);
8062 arc_evict_zthr = zthr_create("arc_evict",
8063 arc_evict_cb_check, arc_evict_cb, NULL, defclsyspri);
8064 arc_reap_zthr = zthr_create_timer("arc_reap",
8065 arc_reap_cb_check, arc_reap_cb, NULL, SEC2NSEC(1), minclsyspri);
8070 * Calculate maximum amount of dirty data per pool.
8072 * If it has been set by a module parameter, take that.
8073 * Otherwise, use a percentage of physical memory defined by
8074 * zfs_dirty_data_max_percent (default 10%) with a cap at
8075 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
8078 if (zfs_dirty_data_max_max == 0)
8079 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
8080 allmem * zfs_dirty_data_max_max_percent / 100);
8082 if (zfs_dirty_data_max_max == 0)
8083 zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024,
8084 allmem * zfs_dirty_data_max_max_percent / 100);
8087 if (zfs_dirty_data_max == 0) {
8088 zfs_dirty_data_max = allmem *
8089 zfs_dirty_data_max_percent / 100;
8090 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
8091 zfs_dirty_data_max_max);
8094 if (zfs_wrlog_data_max == 0) {
8097 * dp_wrlog_total is reduced for each txg at the end of
8098 * spa_sync(). However, dp_dirty_total is reduced every time
8099 * a block is written out. Thus under normal operation,
8100 * dp_wrlog_total could grow 2 times as big as
8101 * zfs_dirty_data_max.
8103 zfs_wrlog_data_max = zfs_dirty_data_max * 2;
8114 #endif /* _KERNEL */
8116 /* Use B_TRUE to ensure *all* buffers are evicted */
8117 arc_flush(NULL, B_TRUE);
8119 if (arc_ksp != NULL) {
8120 kstat_delete(arc_ksp);
8124 taskq_wait(arc_prune_taskq);
8125 taskq_destroy(arc_prune_taskq);
8127 mutex_enter(&arc_prune_mtx);
8128 while ((p = list_head(&arc_prune_list)) != NULL) {
8129 list_remove(&arc_prune_list, p);
8130 zfs_refcount_remove(&p->p_refcnt, &arc_prune_list);
8131 zfs_refcount_destroy(&p->p_refcnt);
8132 kmem_free(p, sizeof (*p));
8134 mutex_exit(&arc_prune_mtx);
8136 list_destroy(&arc_prune_list);
8137 mutex_destroy(&arc_prune_mtx);
8139 (void) zthr_cancel(arc_evict_zthr);
8140 (void) zthr_cancel(arc_reap_zthr);
8141 arc_state_free_markers(arc_state_evict_markers,
8142 arc_state_evict_marker_count);
8144 mutex_destroy(&arc_evict_lock);
8145 list_destroy(&arc_evict_waiters);
8148 * Free any buffers that were tagged for destruction. This needs
8149 * to occur before arc_state_fini() runs and destroys the aggsum
8150 * values which are updated when freeing scatter ABDs.
8152 l2arc_do_free_on_write();
8155 * buf_fini() must proceed arc_state_fini() because buf_fin() may
8156 * trigger the release of kmem magazines, which can callback to
8157 * arc_space_return() which accesses aggsums freed in act_state_fini().
8162 arc_unregister_hotplug();
8165 * We destroy the zthrs after all the ARC state has been
8166 * torn down to avoid the case of them receiving any
8167 * wakeup() signals after they are destroyed.
8169 zthr_destroy(arc_evict_zthr);
8170 zthr_destroy(arc_reap_zthr);
8172 ASSERT0(arc_loaned_bytes);
8178 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
8179 * It uses dedicated storage devices to hold cached data, which are populated
8180 * using large infrequent writes. The main role of this cache is to boost
8181 * the performance of random read workloads. The intended L2ARC devices
8182 * include short-stroked disks, solid state disks, and other media with
8183 * substantially faster read latency than disk.
8185 * +-----------------------+
8187 * +-----------------------+
8190 * l2arc_feed_thread() arc_read()
8194 * +---------------+ |
8196 * +---------------+ |
8201 * +-------+ +-------+
8203 * | cache | | cache |
8204 * +-------+ +-------+
8205 * +=========+ .-----.
8206 * : L2ARC : |-_____-|
8207 * : devices : | Disks |
8208 * +=========+ `-_____-'
8210 * Read requests are satisfied from the following sources, in order:
8213 * 2) vdev cache of L2ARC devices
8215 * 4) vdev cache of disks
8218 * Some L2ARC device types exhibit extremely slow write performance.
8219 * To accommodate for this there are some significant differences between
8220 * the L2ARC and traditional cache design:
8222 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
8223 * the ARC behave as usual, freeing buffers and placing headers on ghost
8224 * lists. The ARC does not send buffers to the L2ARC during eviction as
8225 * this would add inflated write latencies for all ARC memory pressure.
8227 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
8228 * It does this by periodically scanning buffers from the eviction-end of
8229 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
8230 * not already there. It scans until a headroom of buffers is satisfied,
8231 * which itself is a buffer for ARC eviction. If a compressible buffer is
8232 * found during scanning and selected for writing to an L2ARC device, we
8233 * temporarily boost scanning headroom during the next scan cycle to make
8234 * sure we adapt to compression effects (which might significantly reduce
8235 * the data volume we write to L2ARC). The thread that does this is
8236 * l2arc_feed_thread(), illustrated below; example sizes are included to
8237 * provide a better sense of ratio than this diagram:
8240 * +---------------------+----------+
8241 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
8242 * +---------------------+----------+ | o L2ARC eligible
8243 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
8244 * +---------------------+----------+ |
8245 * 15.9 Gbytes ^ 32 Mbytes |
8247 * l2arc_feed_thread()
8249 * l2arc write hand <--[oooo]--'
8253 * +==============================+
8254 * L2ARC dev |####|#|###|###| |####| ... |
8255 * +==============================+
8258 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
8259 * evicted, then the L2ARC has cached a buffer much sooner than it probably
8260 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
8261 * safe to say that this is an uncommon case, since buffers at the end of
8262 * the ARC lists have moved there due to inactivity.
8264 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
8265 * then the L2ARC simply misses copying some buffers. This serves as a
8266 * pressure valve to prevent heavy read workloads from both stalling the ARC
8267 * with waits and clogging the L2ARC with writes. This also helps prevent
8268 * the potential for the L2ARC to churn if it attempts to cache content too
8269 * quickly, such as during backups of the entire pool.
8271 * 5. After system boot and before the ARC has filled main memory, there are
8272 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
8273 * lists can remain mostly static. Instead of searching from tail of these
8274 * lists as pictured, the l2arc_feed_thread() will search from the list heads
8275 * for eligible buffers, greatly increasing its chance of finding them.
8277 * The L2ARC device write speed is also boosted during this time so that
8278 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
8279 * there are no L2ARC reads, and no fear of degrading read performance
8280 * through increased writes.
8282 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
8283 * the vdev queue can aggregate them into larger and fewer writes. Each
8284 * device is written to in a rotor fashion, sweeping writes through
8285 * available space then repeating.
8287 * 7. The L2ARC does not store dirty content. It never needs to flush
8288 * write buffers back to disk based storage.
8290 * 8. If an ARC buffer is written (and dirtied) which also exists in the
8291 * L2ARC, the now stale L2ARC buffer is immediately dropped.
8293 * The performance of the L2ARC can be tweaked by a number of tunables, which
8294 * may be necessary for different workloads:
8296 * l2arc_write_max max write bytes per interval
8297 * l2arc_write_boost extra write bytes during device warmup
8298 * l2arc_noprefetch skip caching prefetched buffers
8299 * l2arc_headroom number of max device writes to precache
8300 * l2arc_headroom_boost when we find compressed buffers during ARC
8301 * scanning, we multiply headroom by this
8302 * percentage factor for the next scan cycle,
8303 * since more compressed buffers are likely to
8305 * l2arc_feed_secs seconds between L2ARC writing
8307 * Tunables may be removed or added as future performance improvements are
8308 * integrated, and also may become zpool properties.
8310 * There are three key functions that control how the L2ARC warms up:
8312 * l2arc_write_eligible() check if a buffer is eligible to cache
8313 * l2arc_write_size() calculate how much to write
8314 * l2arc_write_interval() calculate sleep delay between writes
8316 * These three functions determine what to write, how much, and how quickly
8319 * L2ARC persistence:
8321 * When writing buffers to L2ARC, we periodically add some metadata to
8322 * make sure we can pick them up after reboot, thus dramatically reducing
8323 * the impact that any downtime has on the performance of storage systems
8324 * with large caches.
8326 * The implementation works fairly simply by integrating the following two
8329 * *) When writing to the L2ARC, we occasionally write a "l2arc log block",
8330 * which is an additional piece of metadata which describes what's been
8331 * written. This allows us to rebuild the arc_buf_hdr_t structures of the
8332 * main ARC buffers. There are 2 linked-lists of log blocks headed by
8333 * dh_start_lbps[2]. We alternate which chain we append to, so they are
8334 * time-wise and offset-wise interleaved, but that is an optimization rather
8335 * than for correctness. The log block also includes a pointer to the
8336 * previous block in its chain.
8338 * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device
8339 * for our header bookkeeping purposes. This contains a device header,
8340 * which contains our top-level reference structures. We update it each
8341 * time we write a new log block, so that we're able to locate it in the
8342 * L2ARC device. If this write results in an inconsistent device header
8343 * (e.g. due to power failure), we detect this by verifying the header's
8344 * checksum and simply fail to reconstruct the L2ARC after reboot.
8346 * Implementation diagram:
8348 * +=== L2ARC device (not to scale) ======================================+
8349 * | ___two newest log block pointers__.__________ |
8350 * | / \dh_start_lbps[1] |
8351 * | / \ \dh_start_lbps[0]|
8353 * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---|
8354 * || hdr| ^ /^ /^ / / |
8355 * |+------+ ...--\-------/ \-----/--\------/ / |
8356 * | \--------------/ \--------------/ |
8357 * +======================================================================+
8359 * As can be seen on the diagram, rather than using a simple linked list,
8360 * we use a pair of linked lists with alternating elements. This is a
8361 * performance enhancement due to the fact that we only find out the
8362 * address of the next log block access once the current block has been
8363 * completely read in. Obviously, this hurts performance, because we'd be
8364 * keeping the device's I/O queue at only a 1 operation deep, thus
8365 * incurring a large amount of I/O round-trip latency. Having two lists
8366 * allows us to fetch two log blocks ahead of where we are currently
8367 * rebuilding L2ARC buffers.
8369 * On-device data structures:
8371 * L2ARC device header: l2arc_dev_hdr_phys_t
8372 * L2ARC log block: l2arc_log_blk_phys_t
8374 * L2ARC reconstruction:
8376 * When writing data, we simply write in the standard rotary fashion,
8377 * evicting buffers as we go and simply writing new data over them (writing
8378 * a new log block every now and then). This obviously means that once we
8379 * loop around the end of the device, we will start cutting into an already
8380 * committed log block (and its referenced data buffers), like so:
8382 * current write head__ __old tail
8385 * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |-->
8386 * ^ ^^^^^^^^^___________________________________
8388 * <<nextwrite>> may overwrite this blk and/or its bufs --'
8390 * When importing the pool, we detect this situation and use it to stop
8391 * our scanning process (see l2arc_rebuild).
8393 * There is one significant caveat to consider when rebuilding ARC contents
8394 * from an L2ARC device: what about invalidated buffers? Given the above
8395 * construction, we cannot update blocks which we've already written to amend
8396 * them to remove buffers which were invalidated. Thus, during reconstruction,
8397 * we might be populating the cache with buffers for data that's not on the
8398 * main pool anymore, or may have been overwritten!
8400 * As it turns out, this isn't a problem. Every arc_read request includes
8401 * both the DVA and, crucially, the birth TXG of the BP the caller is
8402 * looking for. So even if the cache were populated by completely rotten
8403 * blocks for data that had been long deleted and/or overwritten, we'll
8404 * never actually return bad data from the cache, since the DVA with the
8405 * birth TXG uniquely identify a block in space and time - once created,
8406 * a block is immutable on disk. The worst thing we have done is wasted
8407 * some time and memory at l2arc rebuild to reconstruct outdated ARC
8408 * entries that will get dropped from the l2arc as it is being updated
8411 * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write
8412 * hand are not restored. This is done by saving the offset (in bytes)
8413 * l2arc_evict() has evicted to in the L2ARC device header and taking it
8414 * into account when restoring buffers.
8418 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
8421 * A buffer is *not* eligible for the L2ARC if it:
8422 * 1. belongs to a different spa.
8423 * 2. is already cached on the L2ARC.
8424 * 3. has an I/O in progress (it may be an incomplete read).
8425 * 4. is flagged not eligible (zfs property).
8427 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
8428 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
8435 l2arc_write_size(l2arc_dev_t *dev)
8437 uint64_t size, dev_size, tsize;
8440 * Make sure our globals have meaningful values in case the user
8443 size = l2arc_write_max;
8445 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
8446 "be greater than zero, resetting it to the default (%d)",
8448 size = l2arc_write_max = L2ARC_WRITE_SIZE;
8451 if (arc_warm == B_FALSE)
8452 size += l2arc_write_boost;
8455 * Make sure the write size does not exceed the size of the cache
8456 * device. This is important in l2arc_evict(), otherwise infinite
8457 * iteration can occur.
8459 dev_size = dev->l2ad_end - dev->l2ad_start;
8460 tsize = size + l2arc_log_blk_overhead(size, dev);
8461 if (dev->l2ad_vdev->vdev_has_trim && l2arc_trim_ahead > 0)
8462 tsize += MAX(64 * 1024 * 1024,
8463 (tsize * l2arc_trim_ahead) / 100);
8465 if (tsize >= dev_size) {
8466 cmn_err(CE_NOTE, "l2arc_write_max or l2arc_write_boost "
8467 "plus the overhead of log blocks (persistent L2ARC, "
8468 "%llu bytes) exceeds the size of the cache device "
8469 "(guid %llu), resetting them to the default (%d)",
8470 (u_longlong_t)l2arc_log_blk_overhead(size, dev),
8471 (u_longlong_t)dev->l2ad_vdev->vdev_guid, L2ARC_WRITE_SIZE);
8472 size = l2arc_write_max = l2arc_write_boost = L2ARC_WRITE_SIZE;
8474 if (arc_warm == B_FALSE)
8475 size += l2arc_write_boost;
8483 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
8485 clock_t interval, next, now;
8488 * If the ARC lists are busy, increase our write rate; if the
8489 * lists are stale, idle back. This is achieved by checking
8490 * how much we previously wrote - if it was more than half of
8491 * what we wanted, schedule the next write much sooner.
8493 if (l2arc_feed_again && wrote > (wanted / 2))
8494 interval = (hz * l2arc_feed_min_ms) / 1000;
8496 interval = hz * l2arc_feed_secs;
8498 now = ddi_get_lbolt();
8499 next = MAX(now, MIN(now + interval, began + interval));
8505 * Cycle through L2ARC devices. This is how L2ARC load balances.
8506 * If a device is returned, this also returns holding the spa config lock.
8508 static l2arc_dev_t *
8509 l2arc_dev_get_next(void)
8511 l2arc_dev_t *first, *next = NULL;
8514 * Lock out the removal of spas (spa_namespace_lock), then removal
8515 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
8516 * both locks will be dropped and a spa config lock held instead.
8518 mutex_enter(&spa_namespace_lock);
8519 mutex_enter(&l2arc_dev_mtx);
8521 /* if there are no vdevs, there is nothing to do */
8522 if (l2arc_ndev == 0)
8526 next = l2arc_dev_last;
8528 /* loop around the list looking for a non-faulted vdev */
8530 next = list_head(l2arc_dev_list);
8532 next = list_next(l2arc_dev_list, next);
8534 next = list_head(l2arc_dev_list);
8537 /* if we have come back to the start, bail out */
8540 else if (next == first)
8543 ASSERT3P(next, !=, NULL);
8544 } while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8545 next->l2ad_trim_all);
8547 /* if we were unable to find any usable vdevs, return NULL */
8548 if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8549 next->l2ad_trim_all)
8552 l2arc_dev_last = next;
8555 mutex_exit(&l2arc_dev_mtx);
8558 * Grab the config lock to prevent the 'next' device from being
8559 * removed while we are writing to it.
8562 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
8563 mutex_exit(&spa_namespace_lock);
8569 * Free buffers that were tagged for destruction.
8572 l2arc_do_free_on_write(void)
8575 l2arc_data_free_t *df, *df_prev;
8577 mutex_enter(&l2arc_free_on_write_mtx);
8578 buflist = l2arc_free_on_write;
8580 for (df = list_tail(buflist); df; df = df_prev) {
8581 df_prev = list_prev(buflist, df);
8582 ASSERT3P(df->l2df_abd, !=, NULL);
8583 abd_free(df->l2df_abd);
8584 list_remove(buflist, df);
8585 kmem_free(df, sizeof (l2arc_data_free_t));
8588 mutex_exit(&l2arc_free_on_write_mtx);
8592 * A write to a cache device has completed. Update all headers to allow
8593 * reads from these buffers to begin.
8596 l2arc_write_done(zio_t *zio)
8598 l2arc_write_callback_t *cb;
8599 l2arc_lb_abd_buf_t *abd_buf;
8600 l2arc_lb_ptr_buf_t *lb_ptr_buf;
8602 l2arc_dev_hdr_phys_t *l2dhdr;
8604 arc_buf_hdr_t *head, *hdr, *hdr_prev;
8605 kmutex_t *hash_lock;
8606 int64_t bytes_dropped = 0;
8608 cb = zio->io_private;
8609 ASSERT3P(cb, !=, NULL);
8610 dev = cb->l2wcb_dev;
8611 l2dhdr = dev->l2ad_dev_hdr;
8612 ASSERT3P(dev, !=, NULL);
8613 head = cb->l2wcb_head;
8614 ASSERT3P(head, !=, NULL);
8615 buflist = &dev->l2ad_buflist;
8616 ASSERT3P(buflist, !=, NULL);
8617 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
8618 l2arc_write_callback_t *, cb);
8621 * All writes completed, or an error was hit.
8624 mutex_enter(&dev->l2ad_mtx);
8625 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
8626 hdr_prev = list_prev(buflist, hdr);
8628 hash_lock = HDR_LOCK(hdr);
8631 * We cannot use mutex_enter or else we can deadlock
8632 * with l2arc_write_buffers (due to swapping the order
8633 * the hash lock and l2ad_mtx are taken).
8635 if (!mutex_tryenter(hash_lock)) {
8637 * Missed the hash lock. We must retry so we
8638 * don't leave the ARC_FLAG_L2_WRITING bit set.
8640 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
8643 * We don't want to rescan the headers we've
8644 * already marked as having been written out, so
8645 * we reinsert the head node so we can pick up
8646 * where we left off.
8648 list_remove(buflist, head);
8649 list_insert_after(buflist, hdr, head);
8651 mutex_exit(&dev->l2ad_mtx);
8654 * We wait for the hash lock to become available
8655 * to try and prevent busy waiting, and increase
8656 * the chance we'll be able to acquire the lock
8657 * the next time around.
8659 mutex_enter(hash_lock);
8660 mutex_exit(hash_lock);
8665 * We could not have been moved into the arc_l2c_only
8666 * state while in-flight due to our ARC_FLAG_L2_WRITING
8667 * bit being set. Let's just ensure that's being enforced.
8669 ASSERT(HDR_HAS_L1HDR(hdr));
8672 * Skipped - drop L2ARC entry and mark the header as no
8673 * longer L2 eligibile.
8675 if (zio->io_error != 0) {
8677 * Error - drop L2ARC entry.
8679 list_remove(buflist, hdr);
8680 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
8682 uint64_t psize = HDR_GET_PSIZE(hdr);
8683 l2arc_hdr_arcstats_decrement(hdr);
8686 vdev_psize_to_asize(dev->l2ad_vdev, psize);
8687 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
8688 arc_hdr_size(hdr), hdr);
8692 * Allow ARC to begin reads and ghost list evictions to
8695 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
8697 mutex_exit(hash_lock);
8701 * Free the allocated abd buffers for writing the log blocks.
8702 * If the zio failed reclaim the allocated space and remove the
8703 * pointers to these log blocks from the log block pointer list
8704 * of the L2ARC device.
8706 while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) {
8707 abd_free(abd_buf->abd);
8708 zio_buf_free(abd_buf, sizeof (*abd_buf));
8709 if (zio->io_error != 0) {
8710 lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list);
8712 * L2BLK_GET_PSIZE returns aligned size for log
8716 L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop);
8717 bytes_dropped += asize;
8718 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8719 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8720 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8722 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8723 kmem_free(lb_ptr_buf->lb_ptr,
8724 sizeof (l2arc_log_blkptr_t));
8725 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8728 list_destroy(&cb->l2wcb_abd_list);
8730 if (zio->io_error != 0) {
8731 ARCSTAT_BUMP(arcstat_l2_writes_error);
8734 * Restore the lbps array in the header to its previous state.
8735 * If the list of log block pointers is empty, zero out the
8736 * log block pointers in the device header.
8738 lb_ptr_buf = list_head(&dev->l2ad_lbptr_list);
8739 for (int i = 0; i < 2; i++) {
8740 if (lb_ptr_buf == NULL) {
8742 * If the list is empty zero out the device
8743 * header. Otherwise zero out the second log
8744 * block pointer in the header.
8748 dev->l2ad_dev_hdr_asize);
8750 memset(&l2dhdr->dh_start_lbps[i], 0,
8751 sizeof (l2arc_log_blkptr_t));
8755 memcpy(&l2dhdr->dh_start_lbps[i], lb_ptr_buf->lb_ptr,
8756 sizeof (l2arc_log_blkptr_t));
8757 lb_ptr_buf = list_next(&dev->l2ad_lbptr_list,
8762 ARCSTAT_BUMP(arcstat_l2_writes_done);
8763 list_remove(buflist, head);
8764 ASSERT(!HDR_HAS_L1HDR(head));
8765 kmem_cache_free(hdr_l2only_cache, head);
8766 mutex_exit(&dev->l2ad_mtx);
8768 ASSERT(dev->l2ad_vdev != NULL);
8769 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
8771 l2arc_do_free_on_write();
8773 kmem_free(cb, sizeof (l2arc_write_callback_t));
8777 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
8780 spa_t *spa = zio->io_spa;
8781 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
8782 blkptr_t *bp = zio->io_bp;
8783 uint8_t salt[ZIO_DATA_SALT_LEN];
8784 uint8_t iv[ZIO_DATA_IV_LEN];
8785 uint8_t mac[ZIO_DATA_MAC_LEN];
8786 boolean_t no_crypt = B_FALSE;
8789 * ZIL data is never be written to the L2ARC, so we don't need
8790 * special handling for its unique MAC storage.
8792 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
8793 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
8794 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8797 * If the data was encrypted, decrypt it now. Note that
8798 * we must check the bp here and not the hdr, since the
8799 * hdr does not have its encryption parameters updated
8800 * until arc_read_done().
8802 if (BP_IS_ENCRYPTED(bp)) {
8803 abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8804 ARC_HDR_DO_ADAPT | ARC_HDR_USE_RESERVE);
8806 zio_crypt_decode_params_bp(bp, salt, iv);
8807 zio_crypt_decode_mac_bp(bp, mac);
8809 ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
8810 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
8811 salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
8812 hdr->b_l1hdr.b_pabd, &no_crypt);
8814 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8819 * If we actually performed decryption, replace b_pabd
8820 * with the decrypted data. Otherwise we can just throw
8821 * our decryption buffer away.
8824 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8825 arc_hdr_size(hdr), hdr);
8826 hdr->b_l1hdr.b_pabd = eabd;
8829 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8834 * If the L2ARC block was compressed, but ARC compression
8835 * is disabled we decompress the data into a new buffer and
8836 * replace the existing data.
8838 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8839 !HDR_COMPRESSION_ENABLED(hdr)) {
8840 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8841 ARC_HDR_DO_ADAPT | ARC_HDR_USE_RESERVE);
8842 void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
8844 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
8845 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
8846 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
8848 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8849 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
8853 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8854 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8855 arc_hdr_size(hdr), hdr);
8856 hdr->b_l1hdr.b_pabd = cabd;
8858 zio->io_size = HDR_GET_LSIZE(hdr);
8869 * A read to a cache device completed. Validate buffer contents before
8870 * handing over to the regular ARC routines.
8873 l2arc_read_done(zio_t *zio)
8876 l2arc_read_callback_t *cb = zio->io_private;
8878 kmutex_t *hash_lock;
8879 boolean_t valid_cksum;
8880 boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
8881 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
8883 ASSERT3P(zio->io_vd, !=, NULL);
8884 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
8886 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
8888 ASSERT3P(cb, !=, NULL);
8889 hdr = cb->l2rcb_hdr;
8890 ASSERT3P(hdr, !=, NULL);
8892 hash_lock = HDR_LOCK(hdr);
8893 mutex_enter(hash_lock);
8894 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
8897 * If the data was read into a temporary buffer,
8898 * move it and free the buffer.
8900 if (cb->l2rcb_abd != NULL) {
8901 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
8902 if (zio->io_error == 0) {
8904 abd_copy(hdr->b_crypt_hdr.b_rabd,
8905 cb->l2rcb_abd, arc_hdr_size(hdr));
8907 abd_copy(hdr->b_l1hdr.b_pabd,
8908 cb->l2rcb_abd, arc_hdr_size(hdr));
8913 * The following must be done regardless of whether
8914 * there was an error:
8915 * - free the temporary buffer
8916 * - point zio to the real ARC buffer
8917 * - set zio size accordingly
8918 * These are required because zio is either re-used for
8919 * an I/O of the block in the case of the error
8920 * or the zio is passed to arc_read_done() and it
8923 abd_free(cb->l2rcb_abd);
8924 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
8927 ASSERT(HDR_HAS_RABD(hdr));
8928 zio->io_abd = zio->io_orig_abd =
8929 hdr->b_crypt_hdr.b_rabd;
8931 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8932 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8936 ASSERT3P(zio->io_abd, !=, NULL);
8939 * Check this survived the L2ARC journey.
8941 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8942 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8943 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
8944 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
8945 zio->io_prop.zp_complevel = hdr->b_complevel;
8947 valid_cksum = arc_cksum_is_equal(hdr, zio);
8950 * b_rabd will always match the data as it exists on disk if it is
8951 * being used. Therefore if we are reading into b_rabd we do not
8952 * attempt to untransform the data.
8954 if (valid_cksum && !using_rdata)
8955 tfm_error = l2arc_untransform(zio, cb);
8957 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8958 !HDR_L2_EVICTED(hdr)) {
8959 mutex_exit(hash_lock);
8960 zio->io_private = hdr;
8964 * Buffer didn't survive caching. Increment stats and
8965 * reissue to the original storage device.
8967 if (zio->io_error != 0) {
8968 ARCSTAT_BUMP(arcstat_l2_io_error);
8970 zio->io_error = SET_ERROR(EIO);
8972 if (!valid_cksum || tfm_error != 0)
8973 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8976 * If there's no waiter, issue an async i/o to the primary
8977 * storage now. If there *is* a waiter, the caller must
8978 * issue the i/o in a context where it's OK to block.
8980 if (zio->io_waiter == NULL) {
8981 zio_t *pio = zio_unique_parent(zio);
8982 void *abd = (using_rdata) ?
8983 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8985 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8987 zio = zio_read(pio, zio->io_spa, zio->io_bp,
8988 abd, zio->io_size, arc_read_done,
8989 hdr, zio->io_priority, cb->l2rcb_flags,
8993 * Original ZIO will be freed, so we need to update
8994 * ARC header with the new ZIO pointer to be used
8995 * by zio_change_priority() in arc_read().
8997 for (struct arc_callback *acb = hdr->b_l1hdr.b_acb;
8998 acb != NULL; acb = acb->acb_next)
8999 acb->acb_zio_head = zio;
9001 mutex_exit(hash_lock);
9004 mutex_exit(hash_lock);
9008 kmem_free(cb, sizeof (l2arc_read_callback_t));
9012 * This is the list priority from which the L2ARC will search for pages to
9013 * cache. This is used within loops (0..3) to cycle through lists in the
9014 * desired order. This order can have a significant effect on cache
9017 * Currently the metadata lists are hit first, MFU then MRU, followed by
9018 * the data lists. This function returns a locked list, and also returns
9021 static multilist_sublist_t *
9022 l2arc_sublist_lock(int list_num)
9024 multilist_t *ml = NULL;
9027 ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
9031 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
9034 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
9037 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
9040 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
9047 * Return a randomly-selected sublist. This is acceptable
9048 * because the caller feeds only a little bit of data for each
9049 * call (8MB). Subsequent calls will result in different
9050 * sublists being selected.
9052 idx = multilist_get_random_index(ml);
9053 return (multilist_sublist_lock(ml, idx));
9057 * Calculates the maximum overhead of L2ARC metadata log blocks for a given
9058 * L2ARC write size. l2arc_evict and l2arc_write_size need to include this
9059 * overhead in processing to make sure there is enough headroom available
9060 * when writing buffers.
9062 static inline uint64_t
9063 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev)
9065 if (dev->l2ad_log_entries == 0) {
9068 uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT;
9070 uint64_t log_blocks = (log_entries +
9071 dev->l2ad_log_entries - 1) /
9072 dev->l2ad_log_entries;
9074 return (vdev_psize_to_asize(dev->l2ad_vdev,
9075 sizeof (l2arc_log_blk_phys_t)) * log_blocks);
9080 * Evict buffers from the device write hand to the distance specified in
9081 * bytes. This distance may span populated buffers, it may span nothing.
9082 * This is clearing a region on the L2ARC device ready for writing.
9083 * If the 'all' boolean is set, every buffer is evicted.
9086 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
9089 arc_buf_hdr_t *hdr, *hdr_prev;
9090 kmutex_t *hash_lock;
9092 l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev;
9093 vdev_t *vd = dev->l2ad_vdev;
9096 buflist = &dev->l2ad_buflist;
9099 * We need to add in the worst case scenario of log block overhead.
9101 distance += l2arc_log_blk_overhead(distance, dev);
9102 if (vd->vdev_has_trim && l2arc_trim_ahead > 0) {
9104 * Trim ahead of the write size 64MB or (l2arc_trim_ahead/100)
9105 * times the write size, whichever is greater.
9107 distance += MAX(64 * 1024 * 1024,
9108 (distance * l2arc_trim_ahead) / 100);
9113 if (dev->l2ad_hand >= (dev->l2ad_end - distance)) {
9115 * When there is no space to accommodate upcoming writes,
9116 * evict to the end. Then bump the write and evict hands
9117 * to the start and iterate. This iteration does not
9118 * happen indefinitely as we make sure in
9119 * l2arc_write_size() that when the write hand is reset,
9120 * the write size does not exceed the end of the device.
9123 taddr = dev->l2ad_end;
9125 taddr = dev->l2ad_hand + distance;
9127 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
9128 uint64_t, taddr, boolean_t, all);
9132 * This check has to be placed after deciding whether to
9135 if (dev->l2ad_first) {
9137 * This is the first sweep through the device. There is
9138 * nothing to evict. We have already trimmmed the
9144 * Trim the space to be evicted.
9146 if (vd->vdev_has_trim && dev->l2ad_evict < taddr &&
9147 l2arc_trim_ahead > 0) {
9149 * We have to drop the spa_config lock because
9150 * vdev_trim_range() will acquire it.
9151 * l2ad_evict already accounts for the label
9152 * size. To prevent vdev_trim_ranges() from
9153 * adding it again, we subtract it from
9156 spa_config_exit(dev->l2ad_spa, SCL_L2ARC, dev);
9157 vdev_trim_simple(vd,
9158 dev->l2ad_evict - VDEV_LABEL_START_SIZE,
9159 taddr - dev->l2ad_evict);
9160 spa_config_enter(dev->l2ad_spa, SCL_L2ARC, dev,
9165 * When rebuilding L2ARC we retrieve the evict hand
9166 * from the header of the device. Of note, l2arc_evict()
9167 * does not actually delete buffers from the cache
9168 * device, but trimming may do so depending on the
9169 * hardware implementation. Thus keeping track of the
9170 * evict hand is useful.
9172 dev->l2ad_evict = MAX(dev->l2ad_evict, taddr);
9177 mutex_enter(&dev->l2ad_mtx);
9179 * We have to account for evicted log blocks. Run vdev_space_update()
9180 * on log blocks whose offset (in bytes) is before the evicted offset
9181 * (in bytes) by searching in the list of pointers to log blocks
9182 * present in the L2ARC device.
9184 for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf;
9185 lb_ptr_buf = lb_ptr_buf_prev) {
9187 lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf);
9189 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
9190 uint64_t asize = L2BLK_GET_PSIZE(
9191 (lb_ptr_buf->lb_ptr)->lbp_prop);
9194 * We don't worry about log blocks left behind (ie
9195 * lbp_payload_start < l2ad_hand) because l2arc_write_buffers()
9196 * will never write more than l2arc_evict() evicts.
9198 if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) {
9201 vdev_space_update(vd, -asize, 0, 0);
9202 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
9203 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
9204 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
9206 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
9207 list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf);
9208 kmem_free(lb_ptr_buf->lb_ptr,
9209 sizeof (l2arc_log_blkptr_t));
9210 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
9214 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
9215 hdr_prev = list_prev(buflist, hdr);
9217 ASSERT(!HDR_EMPTY(hdr));
9218 hash_lock = HDR_LOCK(hdr);
9221 * We cannot use mutex_enter or else we can deadlock
9222 * with l2arc_write_buffers (due to swapping the order
9223 * the hash lock and l2ad_mtx are taken).
9225 if (!mutex_tryenter(hash_lock)) {
9227 * Missed the hash lock. Retry.
9229 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
9230 mutex_exit(&dev->l2ad_mtx);
9231 mutex_enter(hash_lock);
9232 mutex_exit(hash_lock);
9237 * A header can't be on this list if it doesn't have L2 header.
9239 ASSERT(HDR_HAS_L2HDR(hdr));
9241 /* Ensure this header has finished being written. */
9242 ASSERT(!HDR_L2_WRITING(hdr));
9243 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
9245 if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict ||
9246 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
9248 * We've evicted to the target address,
9249 * or the end of the device.
9251 mutex_exit(hash_lock);
9255 if (!HDR_HAS_L1HDR(hdr)) {
9256 ASSERT(!HDR_L2_READING(hdr));
9258 * This doesn't exist in the ARC. Destroy.
9259 * arc_hdr_destroy() will call list_remove()
9260 * and decrement arcstat_l2_lsize.
9262 arc_change_state(arc_anon, hdr, hash_lock);
9263 arc_hdr_destroy(hdr);
9265 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
9266 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
9268 * Invalidate issued or about to be issued
9269 * reads, since we may be about to write
9270 * over this location.
9272 if (HDR_L2_READING(hdr)) {
9273 ARCSTAT_BUMP(arcstat_l2_evict_reading);
9274 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
9277 arc_hdr_l2hdr_destroy(hdr);
9279 mutex_exit(hash_lock);
9281 mutex_exit(&dev->l2ad_mtx);
9285 * We need to check if we evict all buffers, otherwise we may iterate
9288 if (!all && rerun) {
9290 * Bump device hand to the device start if it is approaching the
9291 * end. l2arc_evict() has already evicted ahead for this case.
9293 dev->l2ad_hand = dev->l2ad_start;
9294 dev->l2ad_evict = dev->l2ad_start;
9295 dev->l2ad_first = B_FALSE;
9301 * In case of cache device removal (all) the following
9302 * assertions may be violated without functional consequences
9303 * as the device is about to be removed.
9305 ASSERT3U(dev->l2ad_hand + distance, <, dev->l2ad_end);
9306 if (!dev->l2ad_first)
9307 ASSERT3U(dev->l2ad_hand, <, dev->l2ad_evict);
9312 * Handle any abd transforms that might be required for writing to the L2ARC.
9313 * If successful, this function will always return an abd with the data
9314 * transformed as it is on disk in a new abd of asize bytes.
9317 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
9322 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
9323 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
9324 uint64_t psize = HDR_GET_PSIZE(hdr);
9325 uint64_t size = arc_hdr_size(hdr);
9326 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
9327 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
9328 dsl_crypto_key_t *dck = NULL;
9329 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
9330 boolean_t no_crypt = B_FALSE;
9332 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
9333 !HDR_COMPRESSION_ENABLED(hdr)) ||
9334 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
9335 ASSERT3U(psize, <=, asize);
9338 * If this data simply needs its own buffer, we simply allocate it
9339 * and copy the data. This may be done to eliminate a dependency on a
9340 * shared buffer or to reallocate the buffer to match asize.
9342 if (HDR_HAS_RABD(hdr) && asize != psize) {
9343 ASSERT3U(asize, >=, psize);
9344 to_write = abd_alloc_for_io(asize, ismd);
9345 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
9347 abd_zero_off(to_write, psize, asize - psize);
9351 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
9352 !HDR_ENCRYPTED(hdr)) {
9353 ASSERT3U(size, ==, psize);
9354 to_write = abd_alloc_for_io(asize, ismd);
9355 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
9357 abd_zero_off(to_write, size, asize - size);
9361 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
9363 * In some cases, we can wind up with size > asize, so
9364 * we need to opt for the larger allocation option here.
9366 * (We also need abd_return_buf_copy in all cases because
9367 * it's an ASSERT() to modify the buffer before returning it
9368 * with arc_return_buf(), and all the compressors
9369 * write things before deciding to fail compression in nearly
9372 cabd = abd_alloc_for_io(size, ismd);
9373 tmp = abd_borrow_buf(cabd, size);
9375 psize = zio_compress_data(compress, to_write, tmp, size,
9378 if (psize >= asize) {
9379 psize = HDR_GET_PSIZE(hdr);
9380 abd_return_buf_copy(cabd, tmp, size);
9381 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
9383 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
9385 abd_zero_off(to_write, psize, asize - psize);
9388 ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
9390 memset((char *)tmp + psize, 0, asize - psize);
9391 psize = HDR_GET_PSIZE(hdr);
9392 abd_return_buf_copy(cabd, tmp, size);
9397 if (HDR_ENCRYPTED(hdr)) {
9398 eabd = abd_alloc_for_io(asize, ismd);
9401 * If the dataset was disowned before the buffer
9402 * made it to this point, the key to re-encrypt
9403 * it won't be available. In this case we simply
9404 * won't write the buffer to the L2ARC.
9406 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
9411 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
9412 hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
9413 hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
9419 abd_copy(eabd, to_write, psize);
9422 abd_zero_off(eabd, psize, asize - psize);
9424 /* assert that the MAC we got here matches the one we saved */
9425 ASSERT0(memcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
9426 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9428 if (to_write == cabd)
9435 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
9436 *abd_out = to_write;
9441 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9452 l2arc_blk_fetch_done(zio_t *zio)
9454 l2arc_read_callback_t *cb;
9456 cb = zio->io_private;
9457 if (cb->l2rcb_abd != NULL)
9458 abd_free(cb->l2rcb_abd);
9459 kmem_free(cb, sizeof (l2arc_read_callback_t));
9463 * Find and write ARC buffers to the L2ARC device.
9465 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
9466 * for reading until they have completed writing.
9467 * The headroom_boost is an in-out parameter used to maintain headroom boost
9468 * state between calls to this function.
9470 * Returns the number of bytes actually written (which may be smaller than
9471 * the delta by which the device hand has changed due to alignment and the
9472 * writing of log blocks).
9475 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
9477 arc_buf_hdr_t *hdr, *hdr_prev, *head;
9478 uint64_t write_asize, write_psize, write_lsize, headroom;
9480 l2arc_write_callback_t *cb = NULL;
9482 uint64_t guid = spa_load_guid(spa);
9483 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9485 ASSERT3P(dev->l2ad_vdev, !=, NULL);
9488 write_lsize = write_asize = write_psize = 0;
9490 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
9491 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
9494 * Copy buffers for L2ARC writing.
9496 for (int pass = 0; pass < L2ARC_FEED_TYPES; pass++) {
9498 * If pass == 1 or 3, we cache MRU metadata and data
9501 if (l2arc_mfuonly) {
9502 if (pass == 1 || pass == 3)
9506 multilist_sublist_t *mls = l2arc_sublist_lock(pass);
9507 uint64_t passed_sz = 0;
9509 VERIFY3P(mls, !=, NULL);
9512 * L2ARC fast warmup.
9514 * Until the ARC is warm and starts to evict, read from the
9515 * head of the ARC lists rather than the tail.
9517 if (arc_warm == B_FALSE)
9518 hdr = multilist_sublist_head(mls);
9520 hdr = multilist_sublist_tail(mls);
9522 headroom = target_sz * l2arc_headroom;
9523 if (zfs_compressed_arc_enabled)
9524 headroom = (headroom * l2arc_headroom_boost) / 100;
9526 for (; hdr; hdr = hdr_prev) {
9527 kmutex_t *hash_lock;
9528 abd_t *to_write = NULL;
9530 if (arc_warm == B_FALSE)
9531 hdr_prev = multilist_sublist_next(mls, hdr);
9533 hdr_prev = multilist_sublist_prev(mls, hdr);
9535 hash_lock = HDR_LOCK(hdr);
9536 if (!mutex_tryenter(hash_lock)) {
9538 * Skip this buffer rather than waiting.
9543 passed_sz += HDR_GET_LSIZE(hdr);
9544 if (l2arc_headroom != 0 && passed_sz > headroom) {
9548 mutex_exit(hash_lock);
9552 if (!l2arc_write_eligible(guid, hdr)) {
9553 mutex_exit(hash_lock);
9557 ASSERT(HDR_HAS_L1HDR(hdr));
9559 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
9560 ASSERT3U(arc_hdr_size(hdr), >, 0);
9561 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
9563 uint64_t psize = HDR_GET_PSIZE(hdr);
9564 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
9567 if ((write_asize + asize) > target_sz) {
9569 mutex_exit(hash_lock);
9574 * We rely on the L1 portion of the header below, so
9575 * it's invalid for this header to have been evicted out
9576 * of the ghost cache, prior to being written out. The
9577 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
9579 arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
9582 * If this header has b_rabd, we can use this since it
9583 * must always match the data exactly as it exists on
9584 * disk. Otherwise, the L2ARC can normally use the
9585 * hdr's data, but if we're sharing data between the
9586 * hdr and one of its bufs, L2ARC needs its own copy of
9587 * the data so that the ZIO below can't race with the
9588 * buf consumer. To ensure that this copy will be
9589 * available for the lifetime of the ZIO and be cleaned
9590 * up afterwards, we add it to the l2arc_free_on_write
9591 * queue. If we need to apply any transforms to the
9592 * data (compression, encryption) we will also need the
9595 if (HDR_HAS_RABD(hdr) && psize == asize) {
9596 to_write = hdr->b_crypt_hdr.b_rabd;
9597 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
9598 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
9599 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
9601 to_write = hdr->b_l1hdr.b_pabd;
9604 arc_buf_contents_t type = arc_buf_type(hdr);
9606 ret = l2arc_apply_transforms(spa, hdr, asize,
9609 arc_hdr_clear_flags(hdr,
9610 ARC_FLAG_L2_WRITING);
9611 mutex_exit(hash_lock);
9615 l2arc_free_abd_on_write(to_write, asize, type);
9620 * Insert a dummy header on the buflist so
9621 * l2arc_write_done() can find where the
9622 * write buffers begin without searching.
9624 mutex_enter(&dev->l2ad_mtx);
9625 list_insert_head(&dev->l2ad_buflist, head);
9626 mutex_exit(&dev->l2ad_mtx);
9629 sizeof (l2arc_write_callback_t), KM_SLEEP);
9630 cb->l2wcb_dev = dev;
9631 cb->l2wcb_head = head;
9633 * Create a list to save allocated abd buffers
9634 * for l2arc_log_blk_commit().
9636 list_create(&cb->l2wcb_abd_list,
9637 sizeof (l2arc_lb_abd_buf_t),
9638 offsetof(l2arc_lb_abd_buf_t, node));
9639 pio = zio_root(spa, l2arc_write_done, cb,
9643 hdr->b_l2hdr.b_dev = dev;
9644 hdr->b_l2hdr.b_hits = 0;
9646 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
9647 hdr->b_l2hdr.b_arcs_state =
9648 hdr->b_l1hdr.b_state->arcs_state;
9649 arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR);
9651 mutex_enter(&dev->l2ad_mtx);
9652 list_insert_head(&dev->l2ad_buflist, hdr);
9653 mutex_exit(&dev->l2ad_mtx);
9655 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
9656 arc_hdr_size(hdr), hdr);
9658 wzio = zio_write_phys(pio, dev->l2ad_vdev,
9659 hdr->b_l2hdr.b_daddr, asize, to_write,
9660 ZIO_CHECKSUM_OFF, NULL, hdr,
9661 ZIO_PRIORITY_ASYNC_WRITE,
9662 ZIO_FLAG_CANFAIL, B_FALSE);
9664 write_lsize += HDR_GET_LSIZE(hdr);
9665 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
9668 write_psize += psize;
9669 write_asize += asize;
9670 dev->l2ad_hand += asize;
9671 l2arc_hdr_arcstats_increment(hdr);
9672 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9674 mutex_exit(hash_lock);
9677 * Append buf info to current log and commit if full.
9678 * arcstat_l2_{size,asize} kstats are updated
9681 if (l2arc_log_blk_insert(dev, hdr))
9682 l2arc_log_blk_commit(dev, pio, cb);
9687 multilist_sublist_unlock(mls);
9693 /* No buffers selected for writing? */
9695 ASSERT0(write_lsize);
9696 ASSERT(!HDR_HAS_L1HDR(head));
9697 kmem_cache_free(hdr_l2only_cache, head);
9700 * Although we did not write any buffers l2ad_evict may
9703 if (dev->l2ad_evict != l2dhdr->dh_evict)
9704 l2arc_dev_hdr_update(dev);
9709 if (!dev->l2ad_first)
9710 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
9712 ASSERT3U(write_asize, <=, target_sz);
9713 ARCSTAT_BUMP(arcstat_l2_writes_sent);
9714 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
9716 dev->l2ad_writing = B_TRUE;
9717 (void) zio_wait(pio);
9718 dev->l2ad_writing = B_FALSE;
9721 * Update the device header after the zio completes as
9722 * l2arc_write_done() may have updated the memory holding the log block
9723 * pointers in the device header.
9725 l2arc_dev_hdr_update(dev);
9727 return (write_asize);
9731 l2arc_hdr_limit_reached(void)
9733 int64_t s = aggsum_upper_bound(&arc_sums.arcstat_l2_hdr_size);
9735 return (arc_reclaim_needed() || (s > arc_meta_limit * 3 / 4) ||
9736 (s > (arc_warm ? arc_c : arc_c_max) * l2arc_meta_percent / 100));
9740 * This thread feeds the L2ARC at regular intervals. This is the beating
9741 * heart of the L2ARC.
9743 static __attribute__((noreturn)) void
9744 l2arc_feed_thread(void *unused)
9750 uint64_t size, wrote;
9751 clock_t begin, next = ddi_get_lbolt();
9752 fstrans_cookie_t cookie;
9754 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
9756 mutex_enter(&l2arc_feed_thr_lock);
9758 cookie = spl_fstrans_mark();
9759 while (l2arc_thread_exit == 0) {
9760 CALLB_CPR_SAFE_BEGIN(&cpr);
9761 (void) cv_timedwait_idle(&l2arc_feed_thr_cv,
9762 &l2arc_feed_thr_lock, next);
9763 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
9764 next = ddi_get_lbolt() + hz;
9767 * Quick check for L2ARC devices.
9769 mutex_enter(&l2arc_dev_mtx);
9770 if (l2arc_ndev == 0) {
9771 mutex_exit(&l2arc_dev_mtx);
9774 mutex_exit(&l2arc_dev_mtx);
9775 begin = ddi_get_lbolt();
9778 * This selects the next l2arc device to write to, and in
9779 * doing so the next spa to feed from: dev->l2ad_spa. This
9780 * will return NULL if there are now no l2arc devices or if
9781 * they are all faulted.
9783 * If a device is returned, its spa's config lock is also
9784 * held to prevent device removal. l2arc_dev_get_next()
9785 * will grab and release l2arc_dev_mtx.
9787 if ((dev = l2arc_dev_get_next()) == NULL)
9790 spa = dev->l2ad_spa;
9791 ASSERT3P(spa, !=, NULL);
9794 * If the pool is read-only then force the feed thread to
9795 * sleep a little longer.
9797 if (!spa_writeable(spa)) {
9798 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
9799 spa_config_exit(spa, SCL_L2ARC, dev);
9804 * Avoid contributing to memory pressure.
9806 if (l2arc_hdr_limit_reached()) {
9807 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
9808 spa_config_exit(spa, SCL_L2ARC, dev);
9812 ARCSTAT_BUMP(arcstat_l2_feeds);
9814 size = l2arc_write_size(dev);
9817 * Evict L2ARC buffers that will be overwritten.
9819 l2arc_evict(dev, size, B_FALSE);
9822 * Write ARC buffers.
9824 wrote = l2arc_write_buffers(spa, dev, size);
9827 * Calculate interval between writes.
9829 next = l2arc_write_interval(begin, size, wrote);
9830 spa_config_exit(spa, SCL_L2ARC, dev);
9832 spl_fstrans_unmark(cookie);
9834 l2arc_thread_exit = 0;
9835 cv_broadcast(&l2arc_feed_thr_cv);
9836 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
9841 l2arc_vdev_present(vdev_t *vd)
9843 return (l2arc_vdev_get(vd) != NULL);
9847 * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if
9848 * the vdev_t isn't an L2ARC device.
9851 l2arc_vdev_get(vdev_t *vd)
9855 mutex_enter(&l2arc_dev_mtx);
9856 for (dev = list_head(l2arc_dev_list); dev != NULL;
9857 dev = list_next(l2arc_dev_list, dev)) {
9858 if (dev->l2ad_vdev == vd)
9861 mutex_exit(&l2arc_dev_mtx);
9867 l2arc_rebuild_dev(l2arc_dev_t *dev, boolean_t reopen)
9869 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9870 uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9871 spa_t *spa = dev->l2ad_spa;
9874 * The L2ARC has to hold at least the payload of one log block for
9875 * them to be restored (persistent L2ARC). The payload of a log block
9876 * depends on the amount of its log entries. We always write log blocks
9877 * with 1022 entries. How many of them are committed or restored depends
9878 * on the size of the L2ARC device. Thus the maximum payload of
9879 * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device
9880 * is less than that, we reduce the amount of committed and restored
9881 * log entries per block so as to enable persistence.
9883 if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) {
9884 dev->l2ad_log_entries = 0;
9886 dev->l2ad_log_entries = MIN((dev->l2ad_end -
9887 dev->l2ad_start) >> SPA_MAXBLOCKSHIFT,
9888 L2ARC_LOG_BLK_MAX_ENTRIES);
9892 * Read the device header, if an error is returned do not rebuild L2ARC.
9894 if (l2arc_dev_hdr_read(dev) == 0 && dev->l2ad_log_entries > 0) {
9896 * If we are onlining a cache device (vdev_reopen) that was
9897 * still present (l2arc_vdev_present()) and rebuild is enabled,
9898 * we should evict all ARC buffers and pointers to log blocks
9899 * and reclaim their space before restoring its contents to
9903 if (!l2arc_rebuild_enabled) {
9906 l2arc_evict(dev, 0, B_TRUE);
9907 /* start a new log block */
9908 dev->l2ad_log_ent_idx = 0;
9909 dev->l2ad_log_blk_payload_asize = 0;
9910 dev->l2ad_log_blk_payload_start = 0;
9914 * Just mark the device as pending for a rebuild. We won't
9915 * be starting a rebuild in line here as it would block pool
9916 * import. Instead spa_load_impl will hand that off to an
9917 * async task which will call l2arc_spa_rebuild_start.
9919 dev->l2ad_rebuild = B_TRUE;
9920 } else if (spa_writeable(spa)) {
9922 * In this case TRIM the whole device if l2arc_trim_ahead > 0,
9923 * otherwise create a new header. We zero out the memory holding
9924 * the header to reset dh_start_lbps. If we TRIM the whole
9925 * device the new header will be written by
9926 * vdev_trim_l2arc_thread() at the end of the TRIM to update the
9927 * trim_state in the header too. When reading the header, if
9928 * trim_state is not VDEV_TRIM_COMPLETE and l2arc_trim_ahead > 0
9929 * we opt to TRIM the whole device again.
9931 if (l2arc_trim_ahead > 0) {
9932 dev->l2ad_trim_all = B_TRUE;
9934 memset(l2dhdr, 0, l2dhdr_asize);
9935 l2arc_dev_hdr_update(dev);
9941 * Add a vdev for use by the L2ARC. By this point the spa has already
9942 * validated the vdev and opened it.
9945 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
9947 l2arc_dev_t *adddev;
9948 uint64_t l2dhdr_asize;
9950 ASSERT(!l2arc_vdev_present(vd));
9953 * Create a new l2arc device entry.
9955 adddev = vmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
9956 adddev->l2ad_spa = spa;
9957 adddev->l2ad_vdev = vd;
9958 /* leave extra size for an l2arc device header */
9959 l2dhdr_asize = adddev->l2ad_dev_hdr_asize =
9960 MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift);
9961 adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize;
9962 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
9963 ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end);
9964 adddev->l2ad_hand = adddev->l2ad_start;
9965 adddev->l2ad_evict = adddev->l2ad_start;
9966 adddev->l2ad_first = B_TRUE;
9967 adddev->l2ad_writing = B_FALSE;
9968 adddev->l2ad_trim_all = B_FALSE;
9969 list_link_init(&adddev->l2ad_node);
9970 adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP);
9972 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
9974 * This is a list of all ARC buffers that are still valid on the
9977 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
9978 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
9981 * This is a list of pointers to log blocks that are still present
9984 list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t),
9985 offsetof(l2arc_lb_ptr_buf_t, node));
9987 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
9988 zfs_refcount_create(&adddev->l2ad_alloc);
9989 zfs_refcount_create(&adddev->l2ad_lb_asize);
9990 zfs_refcount_create(&adddev->l2ad_lb_count);
9993 * Decide if dev is eligible for L2ARC rebuild or whole device
9994 * trimming. This has to happen before the device is added in the
9995 * cache device list and l2arc_dev_mtx is released. Otherwise
9996 * l2arc_feed_thread() might already start writing on the
9999 l2arc_rebuild_dev(adddev, B_FALSE);
10002 * Add device to global list
10004 mutex_enter(&l2arc_dev_mtx);
10005 list_insert_head(l2arc_dev_list, adddev);
10006 atomic_inc_64(&l2arc_ndev);
10007 mutex_exit(&l2arc_dev_mtx);
10011 * Decide if a vdev is eligible for L2ARC rebuild, called from vdev_reopen()
10012 * in case of onlining a cache device.
10015 l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen)
10017 l2arc_dev_t *dev = NULL;
10019 dev = l2arc_vdev_get(vd);
10020 ASSERT3P(dev, !=, NULL);
10023 * In contrast to l2arc_add_vdev() we do not have to worry about
10024 * l2arc_feed_thread() invalidating previous content when onlining a
10025 * cache device. The device parameters (l2ad*) are not cleared when
10026 * offlining the device and writing new buffers will not invalidate
10027 * all previous content. In worst case only buffers that have not had
10028 * their log block written to the device will be lost.
10029 * When onlining the cache device (ie offline->online without exporting
10030 * the pool in between) this happens:
10031 * vdev_reopen() -> vdev_open() -> l2arc_rebuild_vdev()
10033 * vdev_is_dead() = B_FALSE l2ad_rebuild = B_TRUE
10034 * During the time where vdev_is_dead = B_FALSE and until l2ad_rebuild
10035 * is set to B_TRUE we might write additional buffers to the device.
10037 l2arc_rebuild_dev(dev, reopen);
10041 * Remove a vdev from the L2ARC.
10044 l2arc_remove_vdev(vdev_t *vd)
10046 l2arc_dev_t *remdev = NULL;
10049 * Find the device by vdev
10051 remdev = l2arc_vdev_get(vd);
10052 ASSERT3P(remdev, !=, NULL);
10055 * Cancel any ongoing or scheduled rebuild.
10057 mutex_enter(&l2arc_rebuild_thr_lock);
10058 if (remdev->l2ad_rebuild_began == B_TRUE) {
10059 remdev->l2ad_rebuild_cancel = B_TRUE;
10060 while (remdev->l2ad_rebuild == B_TRUE)
10061 cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock);
10063 mutex_exit(&l2arc_rebuild_thr_lock);
10066 * Remove device from global list
10068 mutex_enter(&l2arc_dev_mtx);
10069 list_remove(l2arc_dev_list, remdev);
10070 l2arc_dev_last = NULL; /* may have been invalidated */
10071 atomic_dec_64(&l2arc_ndev);
10072 mutex_exit(&l2arc_dev_mtx);
10075 * Clear all buflists and ARC references. L2ARC device flush.
10077 l2arc_evict(remdev, 0, B_TRUE);
10078 list_destroy(&remdev->l2ad_buflist);
10079 ASSERT(list_is_empty(&remdev->l2ad_lbptr_list));
10080 list_destroy(&remdev->l2ad_lbptr_list);
10081 mutex_destroy(&remdev->l2ad_mtx);
10082 zfs_refcount_destroy(&remdev->l2ad_alloc);
10083 zfs_refcount_destroy(&remdev->l2ad_lb_asize);
10084 zfs_refcount_destroy(&remdev->l2ad_lb_count);
10085 kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize);
10086 vmem_free(remdev, sizeof (l2arc_dev_t));
10092 l2arc_thread_exit = 0;
10095 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
10096 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
10097 mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL);
10098 cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL);
10099 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
10100 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
10102 l2arc_dev_list = &L2ARC_dev_list;
10103 l2arc_free_on_write = &L2ARC_free_on_write;
10104 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
10105 offsetof(l2arc_dev_t, l2ad_node));
10106 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
10107 offsetof(l2arc_data_free_t, l2df_list_node));
10113 mutex_destroy(&l2arc_feed_thr_lock);
10114 cv_destroy(&l2arc_feed_thr_cv);
10115 mutex_destroy(&l2arc_rebuild_thr_lock);
10116 cv_destroy(&l2arc_rebuild_thr_cv);
10117 mutex_destroy(&l2arc_dev_mtx);
10118 mutex_destroy(&l2arc_free_on_write_mtx);
10120 list_destroy(l2arc_dev_list);
10121 list_destroy(l2arc_free_on_write);
10127 if (!(spa_mode_global & SPA_MODE_WRITE))
10130 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
10131 TS_RUN, defclsyspri);
10137 if (!(spa_mode_global & SPA_MODE_WRITE))
10140 mutex_enter(&l2arc_feed_thr_lock);
10141 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
10142 l2arc_thread_exit = 1;
10143 while (l2arc_thread_exit != 0)
10144 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
10145 mutex_exit(&l2arc_feed_thr_lock);
10149 * Punches out rebuild threads for the L2ARC devices in a spa. This should
10150 * be called after pool import from the spa async thread, since starting
10151 * these threads directly from spa_import() will make them part of the
10152 * "zpool import" context and delay process exit (and thus pool import).
10155 l2arc_spa_rebuild_start(spa_t *spa)
10157 ASSERT(MUTEX_HELD(&spa_namespace_lock));
10160 * Locate the spa's l2arc devices and kick off rebuild threads.
10162 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
10164 l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]);
10166 /* Don't attempt a rebuild if the vdev is UNAVAIL */
10169 mutex_enter(&l2arc_rebuild_thr_lock);
10170 if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) {
10171 dev->l2ad_rebuild_began = B_TRUE;
10172 (void) thread_create(NULL, 0, l2arc_dev_rebuild_thread,
10173 dev, 0, &p0, TS_RUN, minclsyspri);
10175 mutex_exit(&l2arc_rebuild_thr_lock);
10180 * Main entry point for L2ARC rebuilding.
10182 static __attribute__((noreturn)) void
10183 l2arc_dev_rebuild_thread(void *arg)
10185 l2arc_dev_t *dev = arg;
10187 VERIFY(!dev->l2ad_rebuild_cancel);
10188 VERIFY(dev->l2ad_rebuild);
10189 (void) l2arc_rebuild(dev);
10190 mutex_enter(&l2arc_rebuild_thr_lock);
10191 dev->l2ad_rebuild_began = B_FALSE;
10192 dev->l2ad_rebuild = B_FALSE;
10193 mutex_exit(&l2arc_rebuild_thr_lock);
10199 * This function implements the actual L2ARC metadata rebuild. It:
10200 * starts reading the log block chain and restores each block's contents
10201 * to memory (reconstructing arc_buf_hdr_t's).
10203 * Operation stops under any of the following conditions:
10205 * 1) We reach the end of the log block chain.
10206 * 2) We encounter *any* error condition (cksum errors, io errors)
10209 l2arc_rebuild(l2arc_dev_t *dev)
10211 vdev_t *vd = dev->l2ad_vdev;
10212 spa_t *spa = vd->vdev_spa;
10214 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10215 l2arc_log_blk_phys_t *this_lb, *next_lb;
10216 zio_t *this_io = NULL, *next_io = NULL;
10217 l2arc_log_blkptr_t lbps[2];
10218 l2arc_lb_ptr_buf_t *lb_ptr_buf;
10219 boolean_t lock_held;
10221 this_lb = vmem_zalloc(sizeof (*this_lb), KM_SLEEP);
10222 next_lb = vmem_zalloc(sizeof (*next_lb), KM_SLEEP);
10225 * We prevent device removal while issuing reads to the device,
10226 * then during the rebuilding phases we drop this lock again so
10227 * that a spa_unload or device remove can be initiated - this is
10228 * safe, because the spa will signal us to stop before removing
10229 * our device and wait for us to stop.
10231 spa_config_enter(spa, SCL_L2ARC, vd, RW_READER);
10232 lock_held = B_TRUE;
10235 * Retrieve the persistent L2ARC device state.
10236 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10238 dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start);
10239 dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr +
10240 L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop),
10242 dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST);
10244 vd->vdev_trim_action_time = l2dhdr->dh_trim_action_time;
10245 vd->vdev_trim_state = l2dhdr->dh_trim_state;
10248 * In case the zfs module parameter l2arc_rebuild_enabled is false
10249 * we do not start the rebuild process.
10251 if (!l2arc_rebuild_enabled)
10254 /* Prepare the rebuild process */
10255 memcpy(lbps, l2dhdr->dh_start_lbps, sizeof (lbps));
10257 /* Start the rebuild process */
10259 if (!l2arc_log_blkptr_valid(dev, &lbps[0]))
10262 if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1],
10263 this_lb, next_lb, this_io, &next_io)) != 0)
10267 * Our memory pressure valve. If the system is running low
10268 * on memory, rather than swamping memory with new ARC buf
10269 * hdrs, we opt not to rebuild the L2ARC. At this point,
10270 * however, we have already set up our L2ARC dev to chain in
10271 * new metadata log blocks, so the user may choose to offline/
10272 * online the L2ARC dev at a later time (or re-import the pool)
10273 * to reconstruct it (when there's less memory pressure).
10275 if (l2arc_hdr_limit_reached()) {
10276 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem);
10277 cmn_err(CE_NOTE, "System running low on memory, "
10278 "aborting L2ARC rebuild.");
10279 err = SET_ERROR(ENOMEM);
10283 spa_config_exit(spa, SCL_L2ARC, vd);
10284 lock_held = B_FALSE;
10287 * Now that we know that the next_lb checks out alright, we
10288 * can start reconstruction from this log block.
10289 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10291 uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop);
10292 l2arc_log_blk_restore(dev, this_lb, asize);
10295 * log block restored, include its pointer in the list of
10296 * pointers to log blocks present in the L2ARC device.
10298 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
10299 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t),
10301 memcpy(lb_ptr_buf->lb_ptr, &lbps[0],
10302 sizeof (l2arc_log_blkptr_t));
10303 mutex_enter(&dev->l2ad_mtx);
10304 list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf);
10305 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10306 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10307 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10308 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10309 mutex_exit(&dev->l2ad_mtx);
10310 vdev_space_update(vd, asize, 0, 0);
10313 * Protection against loops of log blocks:
10315 * l2ad_hand l2ad_evict
10317 * l2ad_start |=======================================| l2ad_end
10318 * -----|||----|||---|||----|||
10320 * ---|||---|||----|||---|||
10323 * In this situation the pointer of log block (4) passes
10324 * l2arc_log_blkptr_valid() but the log block should not be
10325 * restored as it is overwritten by the payload of log block
10326 * (0). Only log blocks (0)-(3) should be restored. We check
10327 * whether l2ad_evict lies in between the payload starting
10328 * offset of the next log block (lbps[1].lbp_payload_start)
10329 * and the payload starting offset of the present log block
10330 * (lbps[0].lbp_payload_start). If true and this isn't the
10331 * first pass, we are looping from the beginning and we should
10334 if (l2arc_range_check_overlap(lbps[1].lbp_payload_start,
10335 lbps[0].lbp_payload_start, dev->l2ad_evict) &&
10339 kpreempt(KPREEMPT_SYNC);
10341 mutex_enter(&l2arc_rebuild_thr_lock);
10342 if (dev->l2ad_rebuild_cancel) {
10343 dev->l2ad_rebuild = B_FALSE;
10344 cv_signal(&l2arc_rebuild_thr_cv);
10345 mutex_exit(&l2arc_rebuild_thr_lock);
10346 err = SET_ERROR(ECANCELED);
10349 mutex_exit(&l2arc_rebuild_thr_lock);
10350 if (spa_config_tryenter(spa, SCL_L2ARC, vd,
10352 lock_held = B_TRUE;
10356 * L2ARC config lock held by somebody in writer,
10357 * possibly due to them trying to remove us. They'll
10358 * likely to want us to shut down, so after a little
10359 * delay, we check l2ad_rebuild_cancel and retry
10366 * Continue with the next log block.
10369 lbps[1] = this_lb->lb_prev_lbp;
10370 PTR_SWAP(this_lb, next_lb);
10375 if (this_io != NULL)
10376 l2arc_log_blk_fetch_abort(this_io);
10378 if (next_io != NULL)
10379 l2arc_log_blk_fetch_abort(next_io);
10380 vmem_free(this_lb, sizeof (*this_lb));
10381 vmem_free(next_lb, sizeof (*next_lb));
10383 if (!l2arc_rebuild_enabled) {
10384 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10386 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) {
10387 ARCSTAT_BUMP(arcstat_l2_rebuild_success);
10388 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10389 "successful, restored %llu blocks",
10390 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10391 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) {
10393 * No error but also nothing restored, meaning the lbps array
10394 * in the device header points to invalid/non-present log
10395 * blocks. Reset the header.
10397 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10398 "no valid log blocks");
10399 memset(l2dhdr, 0, dev->l2ad_dev_hdr_asize);
10400 l2arc_dev_hdr_update(dev);
10401 } else if (err == ECANCELED) {
10403 * In case the rebuild was canceled do not log to spa history
10404 * log as the pool may be in the process of being removed.
10406 zfs_dbgmsg("L2ARC rebuild aborted, restored %llu blocks",
10407 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10408 } else if (err != 0) {
10409 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10410 "aborted, restored %llu blocks",
10411 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10415 spa_config_exit(spa, SCL_L2ARC, vd);
10421 * Attempts to read the device header on the provided L2ARC device and writes
10422 * it to `hdr'. On success, this function returns 0, otherwise the appropriate
10423 * error code is returned.
10426 l2arc_dev_hdr_read(l2arc_dev_t *dev)
10430 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10431 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10434 guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10436 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10438 err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev,
10439 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd,
10440 ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
10441 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
10442 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY |
10443 ZIO_FLAG_SPECULATIVE, B_FALSE));
10448 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors);
10449 zfs_dbgmsg("L2ARC IO error (%d) while reading device header, "
10450 "vdev guid: %llu", err,
10451 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10455 if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC))
10456 byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr));
10458 if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC ||
10459 l2dhdr->dh_spa_guid != guid ||
10460 l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid ||
10461 l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION ||
10462 l2dhdr->dh_log_entries != dev->l2ad_log_entries ||
10463 l2dhdr->dh_end != dev->l2ad_end ||
10464 !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end,
10465 l2dhdr->dh_evict) ||
10466 (l2dhdr->dh_trim_state != VDEV_TRIM_COMPLETE &&
10467 l2arc_trim_ahead > 0)) {
10469 * Attempt to rebuild a device containing no actual dev hdr
10470 * or containing a header from some other pool or from another
10471 * version of persistent L2ARC.
10473 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported);
10474 return (SET_ERROR(ENOTSUP));
10481 * Reads L2ARC log blocks from storage and validates their contents.
10483 * This function implements a simple fetcher to make sure that while
10484 * we're processing one buffer the L2ARC is already fetching the next
10485 * one in the chain.
10487 * The arguments this_lp and next_lp point to the current and next log block
10488 * address in the block chain. Similarly, this_lb and next_lb hold the
10489 * l2arc_log_blk_phys_t's of the current and next L2ARC blk.
10491 * The `this_io' and `next_io' arguments are used for block fetching.
10492 * When issuing the first blk IO during rebuild, you should pass NULL for
10493 * `this_io'. This function will then issue a sync IO to read the block and
10494 * also issue an async IO to fetch the next block in the block chain. The
10495 * fetched IO is returned in `next_io'. On subsequent calls to this
10496 * function, pass the value returned in `next_io' from the previous call
10497 * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO.
10498 * Prior to the call, you should initialize your `next_io' pointer to be
10499 * NULL. If no fetch IO was issued, the pointer is left set at NULL.
10501 * On success, this function returns 0, otherwise it returns an appropriate
10502 * error code. On error the fetching IO is aborted and cleared before
10503 * returning from this function. Therefore, if we return `success', the
10504 * caller can assume that we have taken care of cleanup of fetch IOs.
10507 l2arc_log_blk_read(l2arc_dev_t *dev,
10508 const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp,
10509 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
10510 zio_t *this_io, zio_t **next_io)
10517 ASSERT(this_lbp != NULL && next_lbp != NULL);
10518 ASSERT(this_lb != NULL && next_lb != NULL);
10519 ASSERT(next_io != NULL && *next_io == NULL);
10520 ASSERT(l2arc_log_blkptr_valid(dev, this_lbp));
10523 * Check to see if we have issued the IO for this log block in a
10524 * previous run. If not, this is the first call, so issue it now.
10526 if (this_io == NULL) {
10527 this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp,
10532 * Peek to see if we can start issuing the next IO immediately.
10534 if (l2arc_log_blkptr_valid(dev, next_lbp)) {
10536 * Start issuing IO for the next log block early - this
10537 * should help keep the L2ARC device busy while we
10538 * decompress and restore this log block.
10540 *next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp,
10544 /* Wait for the IO to read this log block to complete */
10545 if ((err = zio_wait(this_io)) != 0) {
10546 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
10547 zfs_dbgmsg("L2ARC IO error (%d) while reading log block, "
10548 "offset: %llu, vdev guid: %llu", err,
10549 (u_longlong_t)this_lbp->lbp_daddr,
10550 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10555 * Make sure the buffer checks out.
10556 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10558 asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop);
10559 fletcher_4_native(this_lb, asize, NULL, &cksum);
10560 if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) {
10561 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors);
10562 zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, "
10563 "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu",
10564 (u_longlong_t)this_lbp->lbp_daddr,
10565 (u_longlong_t)dev->l2ad_vdev->vdev_guid,
10566 (u_longlong_t)dev->l2ad_hand,
10567 (u_longlong_t)dev->l2ad_evict);
10568 err = SET_ERROR(ECKSUM);
10572 /* Now we can take our time decoding this buffer */
10573 switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) {
10574 case ZIO_COMPRESS_OFF:
10576 case ZIO_COMPRESS_LZ4:
10577 abd = abd_alloc_for_io(asize, B_TRUE);
10578 abd_copy_from_buf_off(abd, this_lb, 0, asize);
10579 if ((err = zio_decompress_data(
10580 L2BLK_GET_COMPRESS((this_lbp)->lbp_prop),
10581 abd, this_lb, asize, sizeof (*this_lb), NULL)) != 0) {
10582 err = SET_ERROR(EINVAL);
10587 err = SET_ERROR(EINVAL);
10590 if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC))
10591 byteswap_uint64_array(this_lb, sizeof (*this_lb));
10592 if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) {
10593 err = SET_ERROR(EINVAL);
10597 /* Abort an in-flight fetch I/O in case of error */
10598 if (err != 0 && *next_io != NULL) {
10599 l2arc_log_blk_fetch_abort(*next_io);
10608 * Restores the payload of a log block to ARC. This creates empty ARC hdr
10609 * entries which only contain an l2arc hdr, essentially restoring the
10610 * buffers to their L2ARC evicted state. This function also updates space
10611 * usage on the L2ARC vdev to make sure it tracks restored buffers.
10614 l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb,
10617 uint64_t size = 0, asize = 0;
10618 uint64_t log_entries = dev->l2ad_log_entries;
10621 * Usually arc_adapt() is called only for data, not headers, but
10622 * since we may allocate significant amount of memory here, let ARC
10625 arc_adapt(log_entries * HDR_L2ONLY_SIZE, arc_l2c_only);
10627 for (int i = log_entries - 1; i >= 0; i--) {
10629 * Restore goes in the reverse temporal direction to preserve
10630 * correct temporal ordering of buffers in the l2ad_buflist.
10631 * l2arc_hdr_restore also does a list_insert_tail instead of
10632 * list_insert_head on the l2ad_buflist:
10634 * LIST l2ad_buflist LIST
10635 * HEAD <------ (time) ------ TAIL
10636 * direction +-----+-----+-----+-----+-----+ direction
10637 * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild
10638 * fill +-----+-----+-----+-----+-----+
10642 * l2arc_feed_thread l2arc_rebuild
10643 * will place new bufs here restores bufs here
10645 * During l2arc_rebuild() the device is not used by
10646 * l2arc_feed_thread() as dev->l2ad_rebuild is set to true.
10648 size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop);
10649 asize += vdev_psize_to_asize(dev->l2ad_vdev,
10650 L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop));
10651 l2arc_hdr_restore(&lb->lb_entries[i], dev);
10655 * Record rebuild stats:
10656 * size Logical size of restored buffers in the L2ARC
10657 * asize Aligned size of restored buffers in the L2ARC
10659 ARCSTAT_INCR(arcstat_l2_rebuild_size, size);
10660 ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize);
10661 ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries);
10662 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize);
10663 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize);
10664 ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks);
10668 * Restores a single ARC buf hdr from a log entry. The ARC buffer is put
10669 * into a state indicating that it has been evicted to L2ARC.
10672 l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
10674 arc_buf_hdr_t *hdr, *exists;
10675 kmutex_t *hash_lock;
10676 arc_buf_contents_t type = L2BLK_GET_TYPE((le)->le_prop);
10680 * Do all the allocation before grabbing any locks, this lets us
10681 * sleep if memory is full and we don't have to deal with failed
10684 hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type,
10685 dev, le->le_dva, le->le_daddr,
10686 L2BLK_GET_PSIZE((le)->le_prop), le->le_birth,
10687 L2BLK_GET_COMPRESS((le)->le_prop), le->le_complevel,
10688 L2BLK_GET_PROTECTED((le)->le_prop),
10689 L2BLK_GET_PREFETCH((le)->le_prop),
10690 L2BLK_GET_STATE((le)->le_prop));
10691 asize = vdev_psize_to_asize(dev->l2ad_vdev,
10692 L2BLK_GET_PSIZE((le)->le_prop));
10695 * vdev_space_update() has to be called before arc_hdr_destroy() to
10696 * avoid underflow since the latter also calls vdev_space_update().
10698 l2arc_hdr_arcstats_increment(hdr);
10699 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10701 mutex_enter(&dev->l2ad_mtx);
10702 list_insert_tail(&dev->l2ad_buflist, hdr);
10703 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
10704 mutex_exit(&dev->l2ad_mtx);
10706 exists = buf_hash_insert(hdr, &hash_lock);
10708 /* Buffer was already cached, no need to restore it. */
10709 arc_hdr_destroy(hdr);
10711 * If the buffer is already cached, check whether it has
10712 * L2ARC metadata. If not, enter them and update the flag.
10713 * This is important is case of onlining a cache device, since
10714 * we previously evicted all L2ARC metadata from ARC.
10716 if (!HDR_HAS_L2HDR(exists)) {
10717 arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR);
10718 exists->b_l2hdr.b_dev = dev;
10719 exists->b_l2hdr.b_daddr = le->le_daddr;
10720 exists->b_l2hdr.b_arcs_state =
10721 L2BLK_GET_STATE((le)->le_prop);
10722 mutex_enter(&dev->l2ad_mtx);
10723 list_insert_tail(&dev->l2ad_buflist, exists);
10724 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
10725 arc_hdr_size(exists), exists);
10726 mutex_exit(&dev->l2ad_mtx);
10727 l2arc_hdr_arcstats_increment(exists);
10728 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10730 ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
10733 mutex_exit(hash_lock);
10737 * Starts an asynchronous read IO to read a log block. This is used in log
10738 * block reconstruction to start reading the next block before we are done
10739 * decoding and reconstructing the current block, to keep the l2arc device
10740 * nice and hot with read IO to process.
10741 * The returned zio will contain a newly allocated memory buffers for the IO
10742 * data which should then be freed by the caller once the zio is no longer
10743 * needed (i.e. due to it having completed). If you wish to abort this
10744 * zio, you should do so using l2arc_log_blk_fetch_abort, which takes
10745 * care of disposing of the allocated buffers correctly.
10748 l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp,
10749 l2arc_log_blk_phys_t *lb)
10753 l2arc_read_callback_t *cb;
10755 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10756 asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10757 ASSERT(asize <= sizeof (l2arc_log_blk_phys_t));
10759 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP);
10760 cb->l2rcb_abd = abd_get_from_buf(lb, asize);
10761 pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb,
10762 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
10763 ZIO_FLAG_DONT_RETRY);
10764 (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize,
10765 cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10766 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
10767 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
10773 * Aborts a zio returned from l2arc_log_blk_fetch and frees the data
10774 * buffers allocated for it.
10777 l2arc_log_blk_fetch_abort(zio_t *zio)
10779 (void) zio_wait(zio);
10783 * Creates a zio to update the device header on an l2arc device.
10786 l2arc_dev_hdr_update(l2arc_dev_t *dev)
10788 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10789 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10793 VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER));
10795 l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC;
10796 l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION;
10797 l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10798 l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid;
10799 l2dhdr->dh_log_entries = dev->l2ad_log_entries;
10800 l2dhdr->dh_evict = dev->l2ad_evict;
10801 l2dhdr->dh_start = dev->l2ad_start;
10802 l2dhdr->dh_end = dev->l2ad_end;
10803 l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize);
10804 l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count);
10805 l2dhdr->dh_flags = 0;
10806 l2dhdr->dh_trim_action_time = dev->l2ad_vdev->vdev_trim_action_time;
10807 l2dhdr->dh_trim_state = dev->l2ad_vdev->vdev_trim_state;
10808 if (dev->l2ad_first)
10809 l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST;
10811 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10813 err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev,
10814 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL,
10815 NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE));
10820 zfs_dbgmsg("L2ARC IO error (%d) while writing device header, "
10821 "vdev guid: %llu", err,
10822 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10827 * Commits a log block to the L2ARC device. This routine is invoked from
10828 * l2arc_write_buffers when the log block fills up.
10829 * This function allocates some memory to temporarily hold the serialized
10830 * buffer to be written. This is then released in l2arc_write_done.
10833 l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb)
10835 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10836 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10837 uint64_t psize, asize;
10839 l2arc_lb_abd_buf_t *abd_buf;
10841 l2arc_lb_ptr_buf_t *lb_ptr_buf;
10843 VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries);
10845 tmpbuf = zio_buf_alloc(sizeof (*lb));
10846 abd_buf = zio_buf_alloc(sizeof (*abd_buf));
10847 abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb));
10848 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
10849 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP);
10851 /* link the buffer into the block chain */
10852 lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1];
10853 lb->lb_magic = L2ARC_LOG_BLK_MAGIC;
10856 * l2arc_log_blk_commit() may be called multiple times during a single
10857 * l2arc_write_buffers() call. Save the allocated abd buffers in a list
10858 * so we can free them in l2arc_write_done() later on.
10860 list_insert_tail(&cb->l2wcb_abd_list, abd_buf);
10862 /* try to compress the buffer */
10863 psize = zio_compress_data(ZIO_COMPRESS_LZ4,
10864 abd_buf->abd, tmpbuf, sizeof (*lb), 0);
10866 /* a log block is never entirely zero */
10867 ASSERT(psize != 0);
10868 asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
10869 ASSERT(asize <= sizeof (*lb));
10872 * Update the start log block pointer in the device header to point
10873 * to the log block we're about to write.
10875 l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0];
10876 l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand;
10877 l2dhdr->dh_start_lbps[0].lbp_payload_asize =
10878 dev->l2ad_log_blk_payload_asize;
10879 l2dhdr->dh_start_lbps[0].lbp_payload_start =
10880 dev->l2ad_log_blk_payload_start;
10882 (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb));
10884 (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize);
10885 L2BLK_SET_CHECKSUM(
10886 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10887 ZIO_CHECKSUM_FLETCHER_4);
10888 if (asize < sizeof (*lb)) {
10889 /* compression succeeded */
10890 memset(tmpbuf + psize, 0, asize - psize);
10891 L2BLK_SET_COMPRESS(
10892 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10895 /* compression failed */
10896 memcpy(tmpbuf, lb, sizeof (*lb));
10897 L2BLK_SET_COMPRESS(
10898 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10902 /* checksum what we're about to write */
10903 fletcher_4_native(tmpbuf, asize, NULL,
10904 &l2dhdr->dh_start_lbps[0].lbp_cksum);
10906 abd_free(abd_buf->abd);
10908 /* perform the write itself */
10909 abd_buf->abd = abd_get_from_buf(tmpbuf, sizeof (*lb));
10910 abd_take_ownership_of_buf(abd_buf->abd, B_TRUE);
10911 wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand,
10912 asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10913 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
10914 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
10915 (void) zio_nowait(wzio);
10917 dev->l2ad_hand += asize;
10919 * Include the committed log block's pointer in the list of pointers
10920 * to log blocks present in the L2ARC device.
10922 memcpy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[0],
10923 sizeof (l2arc_log_blkptr_t));
10924 mutex_enter(&dev->l2ad_mtx);
10925 list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf);
10926 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10927 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10928 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10929 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10930 mutex_exit(&dev->l2ad_mtx);
10931 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10933 /* bump the kstats */
10934 ARCSTAT_INCR(arcstat_l2_write_bytes, asize);
10935 ARCSTAT_BUMP(arcstat_l2_log_blk_writes);
10936 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize);
10937 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio,
10938 dev->l2ad_log_blk_payload_asize / asize);
10940 /* start a new log block */
10941 dev->l2ad_log_ent_idx = 0;
10942 dev->l2ad_log_blk_payload_asize = 0;
10943 dev->l2ad_log_blk_payload_start = 0;
10947 * Validates an L2ARC log block address to make sure that it can be read
10948 * from the provided L2ARC device.
10951 l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp)
10953 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10954 uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10955 uint64_t end = lbp->lbp_daddr + asize - 1;
10956 uint64_t start = lbp->lbp_payload_start;
10957 boolean_t evicted = B_FALSE;
10960 * A log block is valid if all of the following conditions are true:
10961 * - it fits entirely (including its payload) between l2ad_start and
10963 * - it has a valid size
10964 * - neither the log block itself nor part of its payload was evicted
10965 * by l2arc_evict():
10967 * l2ad_hand l2ad_evict
10972 * l2ad_start ============================================ l2ad_end
10973 * --------------------------||||
10980 l2arc_range_check_overlap(start, end, dev->l2ad_hand) ||
10981 l2arc_range_check_overlap(start, end, dev->l2ad_evict) ||
10982 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) ||
10983 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end);
10985 return (start >= dev->l2ad_start && end <= dev->l2ad_end &&
10986 asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) &&
10987 (!evicted || dev->l2ad_first));
10991 * Inserts ARC buffer header `hdr' into the current L2ARC log block on
10992 * the device. The buffer being inserted must be present in L2ARC.
10993 * Returns B_TRUE if the L2ARC log block is full and needs to be committed
10994 * to L2ARC, or B_FALSE if it still has room for more ARC buffers.
10997 l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr)
10999 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
11000 l2arc_log_ent_phys_t *le;
11002 if (dev->l2ad_log_entries == 0)
11005 int index = dev->l2ad_log_ent_idx++;
11007 ASSERT3S(index, <, dev->l2ad_log_entries);
11008 ASSERT(HDR_HAS_L2HDR(hdr));
11010 le = &lb->lb_entries[index];
11011 memset(le, 0, sizeof (*le));
11012 le->le_dva = hdr->b_dva;
11013 le->le_birth = hdr->b_birth;
11014 le->le_daddr = hdr->b_l2hdr.b_daddr;
11016 dev->l2ad_log_blk_payload_start = le->le_daddr;
11017 L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr));
11018 L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr));
11019 L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr));
11020 le->le_complevel = hdr->b_complevel;
11021 L2BLK_SET_TYPE((le)->le_prop, hdr->b_type);
11022 L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr)));
11023 L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr)));
11024 L2BLK_SET_STATE((le)->le_prop, hdr->b_l1hdr.b_state->arcs_state);
11026 dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev,
11027 HDR_GET_PSIZE(hdr));
11029 return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries);
11033 * Checks whether a given L2ARC device address sits in a time-sequential
11034 * range. The trick here is that the L2ARC is a rotary buffer, so we can't
11035 * just do a range comparison, we need to handle the situation in which the
11036 * range wraps around the end of the L2ARC device. Arguments:
11037 * bottom -- Lower end of the range to check (written to earlier).
11038 * top -- Upper end of the range to check (written to later).
11039 * check -- The address for which we want to determine if it sits in
11040 * between the top and bottom.
11042 * The 3-way conditional below represents the following cases:
11044 * bottom < top : Sequentially ordered case:
11045 * <check>--------+-------------------+
11046 * | (overlap here?) |
11048 * |---------------<bottom>============<top>--------------|
11050 * bottom > top: Looped-around case:
11051 * <check>--------+------------------+
11052 * | (overlap here?) |
11054 * |===============<top>---------------<bottom>===========|
11057 * +---------------+---------<check>
11059 * top == bottom : Just a single address comparison.
11062 l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check)
11065 return (bottom <= check && check <= top);
11066 else if (bottom > top)
11067 return (check <= top || bottom <= check);
11069 return (check == top);
11072 EXPORT_SYMBOL(arc_buf_size);
11073 EXPORT_SYMBOL(arc_write);
11074 EXPORT_SYMBOL(arc_read);
11075 EXPORT_SYMBOL(arc_buf_info);
11076 EXPORT_SYMBOL(arc_getbuf_func);
11077 EXPORT_SYMBOL(arc_add_prune_callback);
11078 EXPORT_SYMBOL(arc_remove_prune_callback);
11080 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min, param_set_arc_min,
11081 spl_param_get_u64, ZMOD_RW, "Minimum ARC size in bytes");
11083 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, max, param_set_arc_max,
11084 spl_param_get_u64, ZMOD_RW, "Maximum ARC size in bytes");
11086 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit, param_set_arc_u64,
11087 spl_param_get_u64, ZMOD_RW, "Metadata limit for ARC size in bytes");
11089 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit_percent,
11090 param_set_arc_int, param_get_uint, ZMOD_RW,
11091 "Percent of ARC size for ARC meta limit");
11093 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_min, param_set_arc_u64,
11094 spl_param_get_u64, ZMOD_RW, "Minimum ARC metadata size in bytes");
11096 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_prune, INT, ZMOD_RW,
11097 "Meta objects to scan for prune");
11099 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_adjust_restarts, UINT, ZMOD_RW,
11100 "Limit number of restarts in arc_evict_meta");
11102 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_strategy, UINT, ZMOD_RW,
11103 "Meta reclaim strategy");
11105 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, grow_retry, param_set_arc_int,
11106 param_get_uint, ZMOD_RW, "Seconds before growing ARC size");
11108 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, p_dampener_disable, INT, ZMOD_RW,
11109 "Disable arc_p adapt dampener");
11111 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, shrink_shift, param_set_arc_int,
11112 param_get_uint, ZMOD_RW, "log2(fraction of ARC to reclaim)");
11114 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, pc_percent, UINT, ZMOD_RW,
11115 "Percent of pagecache to reclaim ARC to");
11117 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, p_min_shift, param_set_arc_int,
11118 param_get_uint, ZMOD_RW, "arc_c shift to calc min/max arc_p");
11120 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, average_blocksize, UINT, ZMOD_RD,
11121 "Target average block size");
11123 ZFS_MODULE_PARAM(zfs, zfs_, compressed_arc_enabled, INT, ZMOD_RW,
11124 "Disable compressed ARC buffers");
11126 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prefetch_ms, param_set_arc_int,
11127 param_get_uint, ZMOD_RW, "Min life of prefetch block in ms");
11129 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prescient_prefetch_ms,
11130 param_set_arc_int, param_get_uint, ZMOD_RW,
11131 "Min life of prescient prefetched block in ms");
11133 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_max, U64, ZMOD_RW,
11134 "Max write bytes per interval");
11136 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_boost, U64, ZMOD_RW,
11137 "Extra write bytes during device warmup");
11139 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom, U64, ZMOD_RW,
11140 "Number of max device writes to precache");
11142 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom_boost, U64, ZMOD_RW,
11143 "Compressed l2arc_headroom multiplier");
11145 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, trim_ahead, U64, ZMOD_RW,
11146 "TRIM ahead L2ARC write size multiplier");
11148 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_secs, U64, ZMOD_RW,
11149 "Seconds between L2ARC writing");
11151 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_min_ms, U64, ZMOD_RW,
11152 "Min feed interval in milliseconds");
11154 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, noprefetch, INT, ZMOD_RW,
11155 "Skip caching prefetched buffers");
11157 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_again, INT, ZMOD_RW,
11158 "Turbo L2ARC warmup");
11160 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, norw, INT, ZMOD_RW,
11161 "No reads during writes");
11163 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, meta_percent, UINT, ZMOD_RW,
11164 "Percent of ARC size allowed for L2ARC-only headers");
11166 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_enabled, INT, ZMOD_RW,
11167 "Rebuild the L2ARC when importing a pool");
11169 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_blocks_min_l2size, U64, ZMOD_RW,
11170 "Min size in bytes to write rebuild log blocks in L2ARC");
11172 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, mfuonly, INT, ZMOD_RW,
11173 "Cache only MFU data from ARC into L2ARC");
11175 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, exclude_special, INT, ZMOD_RW,
11176 "Exclude dbufs on special vdevs from being cached to L2ARC if set.");
11178 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, lotsfree_percent, param_set_arc_int,
11179 param_get_uint, ZMOD_RW, "System free memory I/O throttle in bytes");
11181 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, sys_free, param_set_arc_u64,
11182 spl_param_get_u64, ZMOD_RW, "System free memory target size in bytes");
11184 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit, param_set_arc_u64,
11185 spl_param_get_u64, ZMOD_RW, "Minimum bytes of dnodes in ARC");
11187 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit_percent,
11188 param_set_arc_int, param_get_uint, ZMOD_RW,
11189 "Percent of ARC meta buffers for dnodes");
11191 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, dnode_reduce_percent, UINT, ZMOD_RW,
11192 "Percentage of excess dnodes to try to unpin");
11194 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, eviction_pct, UINT, ZMOD_RW,
11195 "When full, ARC allocation waits for eviction of this % of alloc size");
11197 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, evict_batch_limit, UINT, ZMOD_RW,
11198 "The number of headers to evict per sublist before moving to the next");
11200 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, prune_task_threads, INT, ZMOD_RW,
11201 "Number of arc_prune threads");