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 * metadata 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. For example, when using the ZPL each dentry
114 * holds a references on a znode. These dentries must be pruned before
115 * the arc buffer holding the znode can be safely evicted.
117 * Note that the majority of the performance stats are manipulated
118 * with atomic operations.
120 * The L2ARC uses the l2ad_mtx on each vdev for the following:
122 * - L2ARC buflist creation
123 * - L2ARC buflist eviction
124 * - L2ARC write completion, which walks L2ARC buflists
125 * - ARC header destruction, as it removes from L2ARC buflists
126 * - ARC header release, as it removes from L2ARC buflists
132 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
133 * This structure can point either to a block that is still in the cache or to
134 * one that is only accessible in an L2 ARC device, or it can provide
135 * information about a block that was recently evicted. If a block is
136 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
137 * information to retrieve it from the L2ARC device. This information is
138 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
139 * that is in this state cannot access the data directly.
141 * Blocks that are actively being referenced or have not been evicted
142 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
143 * the arc_buf_hdr_t that will point to the data block in memory. A block can
144 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
145 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
146 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
148 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
149 * ability to store the physical data (b_pabd) associated with the DVA of the
150 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
151 * it will match its on-disk compression characteristics. This behavior can be
152 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
153 * compressed ARC functionality is disabled, the b_pabd will point to an
154 * uncompressed version of the on-disk data.
156 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
157 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
158 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
159 * consumer. The ARC will provide references to this data and will keep it
160 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
161 * data block and will evict any arc_buf_t that is no longer referenced. The
162 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
163 * "overhead_size" kstat.
165 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
166 * compressed form. The typical case is that consumers will want uncompressed
167 * data, and when that happens a new data buffer is allocated where the data is
168 * decompressed for them to use. Currently the only consumer who wants
169 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
170 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
171 * with the arc_buf_hdr_t.
173 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
174 * first one is owned by a compressed send consumer (and therefore references
175 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
176 * used by any other consumer (and has its own uncompressed copy of the data
191 * | b_buf +------------>+-----------+ arc_buf_t
192 * | b_pabd +-+ |b_next +---->+-----------+
193 * +-----------+ | |-----------| |b_next +-->NULL
194 * | |b_comp = T | +-----------+
195 * | |b_data +-+ |b_comp = F |
196 * | +-----------+ | |b_data +-+
197 * +->+------+ | +-----------+ |
199 * data | |<--------------+ | uncompressed
200 * +------+ compressed, | data
201 * shared +-->+------+
206 * When a consumer reads a block, the ARC must first look to see if the
207 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
208 * arc_buf_t and either copies uncompressed data into a new data buffer from an
209 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
210 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
211 * hdr is compressed and the desired compression characteristics of the
212 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
213 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
214 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
215 * be anywhere in the hdr's list.
217 * The diagram below shows an example of an uncompressed ARC hdr that is
218 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
219 * the last element in the buf list):
231 * | | arc_buf_t (shared)
232 * | b_buf +------------>+---------+ arc_buf_t
233 * | | |b_next +---->+---------+
234 * | b_pabd +-+ |---------| |b_next +-->NULL
235 * +-----------+ | | | +---------+
237 * | +---------+ | |b_data +-+
238 * +->+------+ | +---------+ |
240 * uncompressed | | | |
243 * | uncompressed | | |
246 * +---------------------------------+
248 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
249 * since the physical block is about to be rewritten. The new data contents
250 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
251 * it may compress the data before writing it to disk. The ARC will be called
252 * with the transformed data and will memcpy the transformed on-disk block into
253 * a newly allocated b_pabd. Writes are always done into buffers which have
254 * either been loaned (and hence are new and don't have other readers) or
255 * buffers which have been released (and hence have their own hdr, if there
256 * were originally other readers of the buf's original hdr). This ensures that
257 * the ARC only needs to update a single buf and its hdr after a write occurs.
259 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
260 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
261 * that when compressed ARC is enabled that the L2ARC blocks are identical
262 * to the on-disk block in the main data pool. This provides a significant
263 * advantage since the ARC can leverage the bp's checksum when reading from the
264 * L2ARC to determine if the contents are valid. However, if the compressed
265 * ARC is disabled, then the L2ARC's block must be transformed to look
266 * like the physical block in the main data pool before comparing the
267 * checksum and determining its validity.
269 * The L1ARC has a slightly different system for storing encrypted data.
270 * Raw (encrypted + possibly compressed) data has a few subtle differences from
271 * data that is just compressed. The biggest difference is that it is not
272 * possible to decrypt encrypted data (or vice-versa) if the keys aren't loaded.
273 * The other difference is that encryption cannot be treated as a suggestion.
274 * If a caller would prefer compressed data, but they actually wind up with
275 * uncompressed data the worst thing that could happen is there might be a
276 * performance hit. If the caller requests encrypted data, however, we must be
277 * sure they actually get it or else secret information could be leaked. Raw
278 * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore,
279 * may have both an encrypted version and a decrypted version of its data at
280 * once. When a caller needs a raw arc_buf_t, it is allocated and the data is
281 * copied out of this header. To avoid complications with b_pabd, raw buffers
287 #include <sys/spa_impl.h>
288 #include <sys/zio_compress.h>
289 #include <sys/zio_checksum.h>
290 #include <sys/zfs_context.h>
292 #include <sys/zfs_refcount.h>
293 #include <sys/vdev.h>
294 #include <sys/vdev_impl.h>
295 #include <sys/dsl_pool.h>
296 #include <sys/multilist.h>
299 #include <sys/fm/fs/zfs.h>
300 #include <sys/callb.h>
301 #include <sys/kstat.h>
302 #include <sys/zthr.h>
303 #include <zfs_fletcher.h>
304 #include <sys/arc_impl.h>
305 #include <sys/trace_zfs.h>
306 #include <sys/aggsum.h>
307 #include <sys/wmsum.h>
308 #include <cityhash.h>
309 #include <sys/vdev_trim.h>
310 #include <sys/zfs_racct.h>
311 #include <sys/zstd/zstd.h>
314 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
315 boolean_t arc_watch = B_FALSE;
319 * This thread's job is to keep enough free memory in the system, by
320 * calling arc_kmem_reap_soon() plus arc_reduce_target_size(), which improves
321 * arc_available_memory().
323 static zthr_t *arc_reap_zthr;
326 * This thread's job is to keep arc_size under arc_c, by calling
327 * arc_evict(), which improves arc_is_overflowing().
329 static zthr_t *arc_evict_zthr;
330 static arc_buf_hdr_t **arc_state_evict_markers;
331 static int arc_state_evict_marker_count;
333 static kmutex_t arc_evict_lock;
334 static boolean_t arc_evict_needed = B_FALSE;
335 static clock_t arc_last_uncached_flush;
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 /* log2(fraction of arc to reclaim) */
380 uint_t arc_shrink_shift = 7;
382 /* percent of pagecache to reclaim arc to */
384 uint_t zfs_arc_pc_percent = 0;
388 * log2(fraction of ARC which must be free to allow growing).
389 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
390 * when reading a new block into the ARC, we will evict an equal-sized block
393 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
394 * we will still not allow it to grow.
396 uint_t arc_no_grow_shift = 5;
400 * minimum lifespan of a prefetch block in clock ticks
401 * (initialized in arc_init())
403 static uint_t arc_min_prefetch_ms;
404 static uint_t arc_min_prescient_prefetch_ms;
407 * If this percent of memory is free, don't throttle.
409 uint_t arc_lotsfree_percent = 10;
412 * The arc has filled available memory and has now warmed up.
417 * These tunables are for performance analysis.
419 uint64_t zfs_arc_max = 0;
420 uint64_t zfs_arc_min = 0;
421 static uint64_t zfs_arc_dnode_limit = 0;
422 static uint_t zfs_arc_dnode_reduce_percent = 10;
423 static uint_t zfs_arc_grow_retry = 0;
424 static uint_t zfs_arc_shrink_shift = 0;
425 uint_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
428 * ARC dirty data constraints for arc_tempreserve_space() throttle:
429 * * total dirty data limit
430 * * anon block dirty limit
431 * * each pool's anon allowance
433 static const unsigned long zfs_arc_dirty_limit_percent = 50;
434 static const unsigned long zfs_arc_anon_limit_percent = 25;
435 static const unsigned long zfs_arc_pool_dirty_percent = 20;
438 * Enable or disable compressed arc buffers.
440 int zfs_compressed_arc_enabled = B_TRUE;
443 * Balance between metadata and data on ghost hits. Values above 100
444 * increase metadata caching by proportionally reducing effect of ghost
445 * data hits on target data/metadata rate.
447 static uint_t zfs_arc_meta_balance = 500;
450 * Percentage that can be consumed by dnodes of ARC meta buffers.
452 static uint_t zfs_arc_dnode_limit_percent = 10;
455 * These tunables are Linux-specific
457 static uint64_t zfs_arc_sys_free = 0;
458 static uint_t zfs_arc_min_prefetch_ms = 0;
459 static uint_t zfs_arc_min_prescient_prefetch_ms = 0;
460 static uint_t zfs_arc_lotsfree_percent = 10;
463 * Number of arc_prune threads
465 static int zfs_arc_prune_task_threads = 1;
468 arc_state_t ARC_anon;
470 arc_state_t ARC_mru_ghost;
472 arc_state_t ARC_mfu_ghost;
473 arc_state_t ARC_l2c_only;
474 arc_state_t ARC_uncached;
476 arc_stats_t arc_stats = {
477 { "hits", KSTAT_DATA_UINT64 },
478 { "iohits", KSTAT_DATA_UINT64 },
479 { "misses", KSTAT_DATA_UINT64 },
480 { "demand_data_hits", KSTAT_DATA_UINT64 },
481 { "demand_data_iohits", KSTAT_DATA_UINT64 },
482 { "demand_data_misses", KSTAT_DATA_UINT64 },
483 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
484 { "demand_metadata_iohits", KSTAT_DATA_UINT64 },
485 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
486 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
487 { "prefetch_data_iohits", KSTAT_DATA_UINT64 },
488 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
489 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
490 { "prefetch_metadata_iohits", KSTAT_DATA_UINT64 },
491 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
492 { "mru_hits", KSTAT_DATA_UINT64 },
493 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
494 { "mfu_hits", KSTAT_DATA_UINT64 },
495 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
496 { "uncached_hits", KSTAT_DATA_UINT64 },
497 { "deleted", KSTAT_DATA_UINT64 },
498 { "mutex_miss", KSTAT_DATA_UINT64 },
499 { "access_skip", KSTAT_DATA_UINT64 },
500 { "evict_skip", KSTAT_DATA_UINT64 },
501 { "evict_not_enough", KSTAT_DATA_UINT64 },
502 { "evict_l2_cached", KSTAT_DATA_UINT64 },
503 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
504 { "evict_l2_eligible_mfu", KSTAT_DATA_UINT64 },
505 { "evict_l2_eligible_mru", KSTAT_DATA_UINT64 },
506 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
507 { "evict_l2_skip", KSTAT_DATA_UINT64 },
508 { "hash_elements", KSTAT_DATA_UINT64 },
509 { "hash_elements_max", KSTAT_DATA_UINT64 },
510 { "hash_collisions", KSTAT_DATA_UINT64 },
511 { "hash_chains", KSTAT_DATA_UINT64 },
512 { "hash_chain_max", KSTAT_DATA_UINT64 },
513 { "meta", KSTAT_DATA_UINT64 },
514 { "pd", KSTAT_DATA_UINT64 },
515 { "pm", 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_data", KSTAT_DATA_UINT64 },
534 { "anon_metadata", KSTAT_DATA_UINT64 },
535 { "anon_evictable_data", KSTAT_DATA_UINT64 },
536 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
537 { "mru_size", KSTAT_DATA_UINT64 },
538 { "mru_data", KSTAT_DATA_UINT64 },
539 { "mru_metadata", KSTAT_DATA_UINT64 },
540 { "mru_evictable_data", KSTAT_DATA_UINT64 },
541 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
542 { "mru_ghost_size", KSTAT_DATA_UINT64 },
543 { "mru_ghost_data", KSTAT_DATA_UINT64 },
544 { "mru_ghost_metadata", KSTAT_DATA_UINT64 },
545 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
546 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
547 { "mfu_size", KSTAT_DATA_UINT64 },
548 { "mfu_data", KSTAT_DATA_UINT64 },
549 { "mfu_metadata", KSTAT_DATA_UINT64 },
550 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
551 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
552 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
553 { "mfu_ghost_data", KSTAT_DATA_UINT64 },
554 { "mfu_ghost_metadata", KSTAT_DATA_UINT64 },
555 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
556 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
557 { "uncached_size", KSTAT_DATA_UINT64 },
558 { "uncached_data", KSTAT_DATA_UINT64 },
559 { "uncached_metadata", KSTAT_DATA_UINT64 },
560 { "uncached_evictable_data", KSTAT_DATA_UINT64 },
561 { "uncached_evictable_metadata", KSTAT_DATA_UINT64 },
562 { "l2_hits", KSTAT_DATA_UINT64 },
563 { "l2_misses", KSTAT_DATA_UINT64 },
564 { "l2_prefetch_asize", KSTAT_DATA_UINT64 },
565 { "l2_mru_asize", KSTAT_DATA_UINT64 },
566 { "l2_mfu_asize", KSTAT_DATA_UINT64 },
567 { "l2_bufc_data_asize", KSTAT_DATA_UINT64 },
568 { "l2_bufc_metadata_asize", KSTAT_DATA_UINT64 },
569 { "l2_feeds", KSTAT_DATA_UINT64 },
570 { "l2_rw_clash", KSTAT_DATA_UINT64 },
571 { "l2_read_bytes", KSTAT_DATA_UINT64 },
572 { "l2_write_bytes", KSTAT_DATA_UINT64 },
573 { "l2_writes_sent", KSTAT_DATA_UINT64 },
574 { "l2_writes_done", KSTAT_DATA_UINT64 },
575 { "l2_writes_error", KSTAT_DATA_UINT64 },
576 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
577 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
578 { "l2_evict_reading", KSTAT_DATA_UINT64 },
579 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
580 { "l2_free_on_write", KSTAT_DATA_UINT64 },
581 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
582 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
583 { "l2_io_error", KSTAT_DATA_UINT64 },
584 { "l2_size", KSTAT_DATA_UINT64 },
585 { "l2_asize", KSTAT_DATA_UINT64 },
586 { "l2_hdr_size", KSTAT_DATA_UINT64 },
587 { "l2_log_blk_writes", KSTAT_DATA_UINT64 },
588 { "l2_log_blk_avg_asize", KSTAT_DATA_UINT64 },
589 { "l2_log_blk_asize", KSTAT_DATA_UINT64 },
590 { "l2_log_blk_count", KSTAT_DATA_UINT64 },
591 { "l2_data_to_meta_ratio", KSTAT_DATA_UINT64 },
592 { "l2_rebuild_success", KSTAT_DATA_UINT64 },
593 { "l2_rebuild_unsupported", KSTAT_DATA_UINT64 },
594 { "l2_rebuild_io_errors", KSTAT_DATA_UINT64 },
595 { "l2_rebuild_dh_errors", KSTAT_DATA_UINT64 },
596 { "l2_rebuild_cksum_lb_errors", KSTAT_DATA_UINT64 },
597 { "l2_rebuild_lowmem", KSTAT_DATA_UINT64 },
598 { "l2_rebuild_size", KSTAT_DATA_UINT64 },
599 { "l2_rebuild_asize", KSTAT_DATA_UINT64 },
600 { "l2_rebuild_bufs", KSTAT_DATA_UINT64 },
601 { "l2_rebuild_bufs_precached", KSTAT_DATA_UINT64 },
602 { "l2_rebuild_log_blks", KSTAT_DATA_UINT64 },
603 { "memory_throttle_count", KSTAT_DATA_UINT64 },
604 { "memory_direct_count", KSTAT_DATA_UINT64 },
605 { "memory_indirect_count", KSTAT_DATA_UINT64 },
606 { "memory_all_bytes", KSTAT_DATA_UINT64 },
607 { "memory_free_bytes", KSTAT_DATA_UINT64 },
608 { "memory_available_bytes", KSTAT_DATA_INT64 },
609 { "arc_no_grow", KSTAT_DATA_UINT64 },
610 { "arc_tempreserve", KSTAT_DATA_UINT64 },
611 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
612 { "arc_prune", KSTAT_DATA_UINT64 },
613 { "arc_meta_used", KSTAT_DATA_UINT64 },
614 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
615 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
616 { "predictive_prefetch", KSTAT_DATA_UINT64 },
617 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
618 { "demand_iohit_predictive_prefetch", KSTAT_DATA_UINT64 },
619 { "prescient_prefetch", KSTAT_DATA_UINT64 },
620 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
621 { "demand_iohit_prescient_prefetch", KSTAT_DATA_UINT64 },
622 { "arc_need_free", KSTAT_DATA_UINT64 },
623 { "arc_sys_free", KSTAT_DATA_UINT64 },
624 { "arc_raw_size", KSTAT_DATA_UINT64 },
625 { "cached_only_in_progress", KSTAT_DATA_UINT64 },
626 { "abd_chunk_waste_size", KSTAT_DATA_UINT64 },
631 #define ARCSTAT_MAX(stat, val) { \
633 while ((val) > (m = arc_stats.stat.value.ui64) && \
634 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
639 * We define a macro to allow ARC hits/misses to be easily broken down by
640 * two separate conditions, giving a total of four different subtypes for
641 * each of hits and misses (so eight statistics total).
643 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
646 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
648 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
652 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
654 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
659 * This macro allows us to use kstats as floating averages. Each time we
660 * update this kstat, we first factor it and the update value by
661 * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall
662 * average. This macro assumes that integer loads and stores are atomic, but
663 * is not safe for multiple writers updating the kstat in parallel (only the
664 * last writer's update will remain).
666 #define ARCSTAT_F_AVG_FACTOR 3
667 #define ARCSTAT_F_AVG(stat, value) \
669 uint64_t x = ARCSTAT(stat); \
670 x = x - x / ARCSTAT_F_AVG_FACTOR + \
671 (value) / ARCSTAT_F_AVG_FACTOR; \
675 static kstat_t *arc_ksp;
678 * There are several ARC variables that are critical to export as kstats --
679 * but we don't want to have to grovel around in the kstat whenever we wish to
680 * manipulate them. For these variables, we therefore define them to be in
681 * terms of the statistic variable. This assures that we are not introducing
682 * the possibility of inconsistency by having shadow copies of the variables,
683 * while still allowing the code to be readable.
685 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
686 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
687 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
688 #define arc_need_free ARCSTAT(arcstat_need_free) /* waiting to be evicted */
690 hrtime_t arc_growtime;
691 list_t arc_prune_list;
692 kmutex_t arc_prune_mtx;
693 taskq_t *arc_prune_taskq;
695 #define GHOST_STATE(state) \
696 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
697 (state) == arc_l2c_only)
699 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
700 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
701 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
702 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
703 #define HDR_PRESCIENT_PREFETCH(hdr) \
704 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
705 #define HDR_COMPRESSION_ENABLED(hdr) \
706 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
708 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
709 #define HDR_UNCACHED(hdr) ((hdr)->b_flags & ARC_FLAG_UNCACHED)
710 #define HDR_L2_READING(hdr) \
711 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
712 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
713 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
714 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
715 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
716 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
717 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
718 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
720 #define HDR_ISTYPE_METADATA(hdr) \
721 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
722 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
724 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
725 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
726 #define HDR_HAS_RABD(hdr) \
727 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
728 (hdr)->b_crypt_hdr.b_rabd != NULL)
729 #define HDR_ENCRYPTED(hdr) \
730 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
731 #define HDR_AUTHENTICATED(hdr) \
732 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
734 /* For storing compression mode in b_flags */
735 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
737 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
738 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
739 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
740 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
742 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
743 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
744 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
745 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
751 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
752 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
753 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
756 * Hash table routines
759 #define BUF_LOCKS 2048
760 typedef struct buf_hash_table {
762 arc_buf_hdr_t **ht_table;
763 kmutex_t ht_locks[BUF_LOCKS] ____cacheline_aligned;
766 static buf_hash_table_t buf_hash_table;
768 #define BUF_HASH_INDEX(spa, dva, birth) \
769 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
770 #define BUF_HASH_LOCK(idx) (&buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
771 #define HDR_LOCK(hdr) \
772 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
774 uint64_t zfs_crc64_table[256];
780 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
781 #define L2ARC_HEADROOM 2 /* num of writes */
784 * If we discover during ARC scan any buffers to be compressed, we boost
785 * our headroom for the next scanning cycle by this percentage multiple.
787 #define L2ARC_HEADROOM_BOOST 200
788 #define L2ARC_FEED_SECS 1 /* caching interval secs */
789 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
792 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
793 * and each of the state has two types: data and metadata.
795 #define L2ARC_FEED_TYPES 4
797 /* L2ARC Performance Tunables */
798 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
799 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
800 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
801 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
802 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
803 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
804 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
805 int l2arc_feed_again = B_TRUE; /* turbo warmup */
806 int l2arc_norw = B_FALSE; /* no reads during writes */
807 static uint_t l2arc_meta_percent = 33; /* limit on headers size */
812 static list_t L2ARC_dev_list; /* device list */
813 static list_t *l2arc_dev_list; /* device list pointer */
814 static kmutex_t l2arc_dev_mtx; /* device list mutex */
815 static l2arc_dev_t *l2arc_dev_last; /* last device used */
816 static list_t L2ARC_free_on_write; /* free after write buf list */
817 static list_t *l2arc_free_on_write; /* free after write list ptr */
818 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
819 static uint64_t l2arc_ndev; /* number of devices */
821 typedef struct l2arc_read_callback {
822 arc_buf_hdr_t *l2rcb_hdr; /* read header */
823 blkptr_t l2rcb_bp; /* original blkptr */
824 zbookmark_phys_t l2rcb_zb; /* original bookmark */
825 int l2rcb_flags; /* original flags */
826 abd_t *l2rcb_abd; /* temporary buffer */
827 } l2arc_read_callback_t;
829 typedef struct l2arc_data_free {
830 /* protected by l2arc_free_on_write_mtx */
833 arc_buf_contents_t l2df_type;
834 list_node_t l2df_list_node;
837 typedef enum arc_fill_flags {
838 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
839 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
840 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
841 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
842 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
845 typedef enum arc_ovf_level {
846 ARC_OVF_NONE, /* ARC within target size. */
847 ARC_OVF_SOME, /* ARC is slightly overflowed. */
848 ARC_OVF_SEVERE /* ARC is severely overflowed. */
851 static kmutex_t l2arc_feed_thr_lock;
852 static kcondvar_t l2arc_feed_thr_cv;
853 static uint8_t l2arc_thread_exit;
855 static kmutex_t l2arc_rebuild_thr_lock;
856 static kcondvar_t l2arc_rebuild_thr_cv;
858 enum arc_hdr_alloc_flags {
859 ARC_HDR_ALLOC_RDATA = 0x1,
860 ARC_HDR_USE_RESERVE = 0x4,
861 ARC_HDR_ALLOC_LINEAR = 0x8,
865 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, const void *, int);
866 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, const void *);
867 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, const void *, int);
868 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, const void *);
869 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, const void *);
870 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size,
872 static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t);
873 static void arc_hdr_alloc_abd(arc_buf_hdr_t *, int);
874 static void arc_hdr_destroy(arc_buf_hdr_t *);
875 static void arc_access(arc_buf_hdr_t *, arc_flags_t, boolean_t);
876 static void arc_buf_watch(arc_buf_t *);
877 static void arc_change_state(arc_state_t *, arc_buf_hdr_t *);
879 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
880 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
881 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
882 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
884 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
885 static void l2arc_read_done(zio_t *);
886 static void l2arc_do_free_on_write(void);
887 static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
888 boolean_t state_only);
890 #define l2arc_hdr_arcstats_increment(hdr) \
891 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_FALSE)
892 #define l2arc_hdr_arcstats_decrement(hdr) \
893 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_FALSE)
894 #define l2arc_hdr_arcstats_increment_state(hdr) \
895 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_TRUE)
896 #define l2arc_hdr_arcstats_decrement_state(hdr) \
897 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_TRUE)
900 * l2arc_exclude_special : A zfs module parameter that controls whether buffers
901 * present on special vdevs are eligibile for caching in L2ARC. If
902 * set to 1, exclude dbufs on special vdevs from being cached to
905 int l2arc_exclude_special = 0;
908 * l2arc_mfuonly : A ZFS module parameter that controls whether only MFU
909 * metadata and data are cached from ARC into L2ARC.
911 static int l2arc_mfuonly = 0;
915 * l2arc_trim_ahead : A ZFS module parameter that controls how much ahead of
916 * the current write size (l2arc_write_max) we should TRIM if we
917 * have filled the device. It is defined as a percentage of the
918 * write size. If set to 100 we trim twice the space required to
919 * accommodate upcoming writes. A minimum of 64MB will be trimmed.
920 * It also enables TRIM of the whole L2ARC device upon creation or
921 * addition to an existing pool or if the header of the device is
922 * invalid upon importing a pool or onlining a cache device. The
923 * default is 0, which disables TRIM on L2ARC altogether as it can
924 * put significant stress on the underlying storage devices. This
925 * will vary depending of how well the specific device handles
928 static uint64_t l2arc_trim_ahead = 0;
931 * Performance tuning of L2ARC persistence:
933 * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding
934 * an L2ARC device (either at pool import or later) will attempt
935 * to rebuild L2ARC buffer contents.
936 * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls
937 * whether log blocks are written to the L2ARC device. If the L2ARC
938 * device is less than 1GB, the amount of data l2arc_evict()
939 * evicts is significant compared to the amount of restored L2ARC
940 * data. In this case do not write log blocks in L2ARC in order
941 * not to waste space.
943 static int l2arc_rebuild_enabled = B_TRUE;
944 static uint64_t l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024;
946 /* L2ARC persistence rebuild control routines. */
947 void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen);
948 static __attribute__((noreturn)) void l2arc_dev_rebuild_thread(void *arg);
949 static int l2arc_rebuild(l2arc_dev_t *dev);
951 /* L2ARC persistence read I/O routines. */
952 static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
953 static int l2arc_log_blk_read(l2arc_dev_t *dev,
954 const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
955 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
956 zio_t *this_io, zio_t **next_io);
957 static zio_t *l2arc_log_blk_fetch(vdev_t *vd,
958 const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb);
959 static void l2arc_log_blk_fetch_abort(zio_t *zio);
961 /* L2ARC persistence block restoration routines. */
962 static void l2arc_log_blk_restore(l2arc_dev_t *dev,
963 const l2arc_log_blk_phys_t *lb, uint64_t lb_asize);
964 static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
967 /* L2ARC persistence write I/O routines. */
968 static uint64_t l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
969 l2arc_write_callback_t *cb);
971 /* L2ARC persistence auxiliary routines. */
972 boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
973 const l2arc_log_blkptr_t *lbp);
974 static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
975 const arc_buf_hdr_t *ab);
976 boolean_t l2arc_range_check_overlap(uint64_t bottom,
977 uint64_t top, uint64_t check);
978 static void l2arc_blk_fetch_done(zio_t *zio);
979 static inline uint64_t
980 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev);
983 * We use Cityhash for this. It's fast, and has good hash properties without
984 * requiring any large static buffers.
987 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
989 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
992 #define HDR_EMPTY(hdr) \
993 ((hdr)->b_dva.dva_word[0] == 0 && \
994 (hdr)->b_dva.dva_word[1] == 0)
996 #define HDR_EMPTY_OR_LOCKED(hdr) \
997 (HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr)))
999 #define HDR_EQUAL(spa, dva, birth, hdr) \
1000 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1001 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1002 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1005 buf_discard_identity(arc_buf_hdr_t *hdr)
1007 hdr->b_dva.dva_word[0] = 0;
1008 hdr->b_dva.dva_word[1] = 0;
1012 static arc_buf_hdr_t *
1013 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1015 const dva_t *dva = BP_IDENTITY(bp);
1016 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1017 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1018 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1021 mutex_enter(hash_lock);
1022 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1023 hdr = hdr->b_hash_next) {
1024 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1029 mutex_exit(hash_lock);
1035 * Insert an entry into the hash table. If there is already an element
1036 * equal to elem in the hash table, then the already existing element
1037 * will be returned and the new element will not be inserted.
1038 * Otherwise returns NULL.
1039 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1041 static arc_buf_hdr_t *
1042 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1044 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1045 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1046 arc_buf_hdr_t *fhdr;
1049 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1050 ASSERT(hdr->b_birth != 0);
1051 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1053 if (lockp != NULL) {
1055 mutex_enter(hash_lock);
1057 ASSERT(MUTEX_HELD(hash_lock));
1060 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1061 fhdr = fhdr->b_hash_next, i++) {
1062 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1066 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1067 buf_hash_table.ht_table[idx] = hdr;
1068 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1070 /* collect some hash table performance data */
1072 ARCSTAT_BUMP(arcstat_hash_collisions);
1074 ARCSTAT_BUMP(arcstat_hash_chains);
1076 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1078 uint64_t he = atomic_inc_64_nv(
1079 &arc_stats.arcstat_hash_elements.value.ui64);
1080 ARCSTAT_MAX(arcstat_hash_elements_max, he);
1086 buf_hash_remove(arc_buf_hdr_t *hdr)
1088 arc_buf_hdr_t *fhdr, **hdrp;
1089 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1091 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1092 ASSERT(HDR_IN_HASH_TABLE(hdr));
1094 hdrp = &buf_hash_table.ht_table[idx];
1095 while ((fhdr = *hdrp) != hdr) {
1096 ASSERT3P(fhdr, !=, NULL);
1097 hdrp = &fhdr->b_hash_next;
1099 *hdrp = hdr->b_hash_next;
1100 hdr->b_hash_next = NULL;
1101 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1103 /* collect some hash table performance data */
1104 atomic_dec_64(&arc_stats.arcstat_hash_elements.value.ui64);
1106 if (buf_hash_table.ht_table[idx] &&
1107 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1108 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1112 * Global data structures and functions for the buf kmem cache.
1115 static kmem_cache_t *hdr_full_cache;
1116 static kmem_cache_t *hdr_full_crypt_cache;
1117 static kmem_cache_t *hdr_l2only_cache;
1118 static kmem_cache_t *buf_cache;
1123 #if defined(_KERNEL)
1125 * Large allocations which do not require contiguous pages
1126 * should be using vmem_free() in the linux kernel\
1128 vmem_free(buf_hash_table.ht_table,
1129 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1131 kmem_free(buf_hash_table.ht_table,
1132 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1134 for (int i = 0; i < BUF_LOCKS; i++)
1135 mutex_destroy(BUF_HASH_LOCK(i));
1136 kmem_cache_destroy(hdr_full_cache);
1137 kmem_cache_destroy(hdr_full_crypt_cache);
1138 kmem_cache_destroy(hdr_l2only_cache);
1139 kmem_cache_destroy(buf_cache);
1143 * Constructor callback - called when the cache is empty
1144 * and a new buf is requested.
1147 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1149 (void) unused, (void) kmflag;
1150 arc_buf_hdr_t *hdr = vbuf;
1152 memset(hdr, 0, HDR_FULL_SIZE);
1153 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1154 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1155 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1157 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1159 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1160 list_link_init(&hdr->b_l2hdr.b_l2node);
1161 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1167 hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
1170 arc_buf_hdr_t *hdr = vbuf;
1172 hdr_full_cons(vbuf, unused, kmflag);
1173 memset(&hdr->b_crypt_hdr, 0, sizeof (hdr->b_crypt_hdr));
1174 arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1180 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1182 (void) unused, (void) kmflag;
1183 arc_buf_hdr_t *hdr = vbuf;
1185 memset(hdr, 0, HDR_L2ONLY_SIZE);
1186 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1192 buf_cons(void *vbuf, void *unused, int kmflag)
1194 (void) unused, (void) kmflag;
1195 arc_buf_t *buf = vbuf;
1197 memset(buf, 0, sizeof (arc_buf_t));
1198 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1204 * Destructor callback - called when a cached buf is
1205 * no longer required.
1208 hdr_full_dest(void *vbuf, void *unused)
1211 arc_buf_hdr_t *hdr = vbuf;
1213 ASSERT(HDR_EMPTY(hdr));
1214 cv_destroy(&hdr->b_l1hdr.b_cv);
1215 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1217 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1219 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1220 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1224 hdr_full_crypt_dest(void *vbuf, void *unused)
1226 (void) vbuf, (void) unused;
1228 hdr_full_dest(vbuf, unused);
1229 arc_space_return(sizeof (((arc_buf_hdr_t *)NULL)->b_crypt_hdr),
1234 hdr_l2only_dest(void *vbuf, void *unused)
1237 arc_buf_hdr_t *hdr = vbuf;
1239 ASSERT(HDR_EMPTY(hdr));
1240 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1244 buf_dest(void *vbuf, void *unused)
1249 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1255 uint64_t *ct = NULL;
1256 uint64_t hsize = 1ULL << 12;
1260 * The hash table is big enough to fill all of physical memory
1261 * with an average block size of zfs_arc_average_blocksize (default 8K).
1262 * By default, the table will take up
1263 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1265 while (hsize * zfs_arc_average_blocksize < arc_all_memory())
1268 buf_hash_table.ht_mask = hsize - 1;
1269 #if defined(_KERNEL)
1271 * Large allocations which do not require contiguous pages
1272 * should be using vmem_alloc() in the linux kernel
1274 buf_hash_table.ht_table =
1275 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
1277 buf_hash_table.ht_table =
1278 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1280 if (buf_hash_table.ht_table == NULL) {
1281 ASSERT(hsize > (1ULL << 8));
1286 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1287 0, hdr_full_cons, hdr_full_dest, NULL, NULL, NULL, 0);
1288 hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
1289 HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
1290 NULL, NULL, NULL, 0);
1291 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1292 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, NULL,
1294 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1295 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1297 for (i = 0; i < 256; i++)
1298 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1299 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1301 for (i = 0; i < BUF_LOCKS; i++)
1302 mutex_init(BUF_HASH_LOCK(i), NULL, MUTEX_DEFAULT, NULL);
1305 #define ARC_MINTIME (hz>>4) /* 62 ms */
1308 * This is the size that the buf occupies in memory. If the buf is compressed,
1309 * it will correspond to the compressed size. You should use this method of
1310 * getting the buf size unless you explicitly need the logical size.
1313 arc_buf_size(arc_buf_t *buf)
1315 return (ARC_BUF_COMPRESSED(buf) ?
1316 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1320 arc_buf_lsize(arc_buf_t *buf)
1322 return (HDR_GET_LSIZE(buf->b_hdr));
1326 * This function will return B_TRUE if the buffer is encrypted in memory.
1327 * This buffer can be decrypted by calling arc_untransform().
1330 arc_is_encrypted(arc_buf_t *buf)
1332 return (ARC_BUF_ENCRYPTED(buf) != 0);
1336 * Returns B_TRUE if the buffer represents data that has not had its MAC
1340 arc_is_unauthenticated(arc_buf_t *buf)
1342 return (HDR_NOAUTH(buf->b_hdr) != 0);
1346 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1347 uint8_t *iv, uint8_t *mac)
1349 arc_buf_hdr_t *hdr = buf->b_hdr;
1351 ASSERT(HDR_PROTECTED(hdr));
1353 memcpy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
1354 memcpy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
1355 memcpy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
1356 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1357 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1361 * Indicates how this buffer is compressed in memory. If it is not compressed
1362 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1363 * arc_untransform() as long as it is also unencrypted.
1366 arc_get_compression(arc_buf_t *buf)
1368 return (ARC_BUF_COMPRESSED(buf) ?
1369 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1373 * Return the compression algorithm used to store this data in the ARC. If ARC
1374 * compression is enabled or this is an encrypted block, this will be the same
1375 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1377 static inline enum zio_compress
1378 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1380 return (HDR_COMPRESSION_ENABLED(hdr) ?
1381 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1385 arc_get_complevel(arc_buf_t *buf)
1387 return (buf->b_hdr->b_complevel);
1390 static inline boolean_t
1391 arc_buf_is_shared(arc_buf_t *buf)
1393 boolean_t shared = (buf->b_data != NULL &&
1394 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1395 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1396 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1397 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1398 IMPLY(shared, ARC_BUF_SHARED(buf));
1399 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1402 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1403 * already being shared" requirement prevents us from doing that.
1410 * Free the checksum associated with this header. If there is no checksum, this
1414 arc_cksum_free(arc_buf_hdr_t *hdr)
1417 ASSERT(HDR_HAS_L1HDR(hdr));
1419 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1420 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1421 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1422 hdr->b_l1hdr.b_freeze_cksum = NULL;
1424 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1429 * Return true iff at least one of the bufs on hdr is not compressed.
1430 * Encrypted buffers count as compressed.
1433 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1435 ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr));
1437 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1438 if (!ARC_BUF_COMPRESSED(b)) {
1447 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1448 * matches the checksum that is stored in the hdr. If there is no checksum,
1449 * or if the buf is compressed, this is a no-op.
1452 arc_cksum_verify(arc_buf_t *buf)
1455 arc_buf_hdr_t *hdr = buf->b_hdr;
1458 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1461 if (ARC_BUF_COMPRESSED(buf))
1464 ASSERT(HDR_HAS_L1HDR(hdr));
1466 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1468 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1469 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1473 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1474 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1475 panic("buffer modified while frozen!");
1476 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1481 * This function makes the assumption that data stored in the L2ARC
1482 * will be transformed exactly as it is in the main pool. Because of
1483 * this we can verify the checksum against the reading process's bp.
1486 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1488 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1489 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1492 * Block pointers always store the checksum for the logical data.
1493 * If the block pointer has the gang bit set, then the checksum
1494 * it represents is for the reconstituted data and not for an
1495 * individual gang member. The zio pipeline, however, must be able to
1496 * determine the checksum of each of the gang constituents so it
1497 * treats the checksum comparison differently than what we need
1498 * for l2arc blocks. This prevents us from using the
1499 * zio_checksum_error() interface directly. Instead we must call the
1500 * zio_checksum_error_impl() so that we can ensure the checksum is
1501 * generated using the correct checksum algorithm and accounts for the
1502 * logical I/O size and not just a gang fragment.
1504 return (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1505 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1506 zio->io_offset, NULL) == 0);
1510 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1511 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1512 * isn't modified later on. If buf is compressed or there is already a checksum
1513 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1516 arc_cksum_compute(arc_buf_t *buf)
1518 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1522 arc_buf_hdr_t *hdr = buf->b_hdr;
1523 ASSERT(HDR_HAS_L1HDR(hdr));
1524 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1525 if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1526 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1530 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1531 ASSERT(!ARC_BUF_COMPRESSED(buf));
1532 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1534 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1535 hdr->b_l1hdr.b_freeze_cksum);
1536 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1543 arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
1545 (void) sig, (void) unused;
1546 panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr);
1551 arc_buf_unwatch(arc_buf_t *buf)
1555 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1556 PROT_READ | PROT_WRITE));
1564 arc_buf_watch(arc_buf_t *buf)
1568 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1575 static arc_buf_contents_t
1576 arc_buf_type(arc_buf_hdr_t *hdr)
1578 arc_buf_contents_t type;
1579 if (HDR_ISTYPE_METADATA(hdr)) {
1580 type = ARC_BUFC_METADATA;
1582 type = ARC_BUFC_DATA;
1584 VERIFY3U(hdr->b_type, ==, type);
1589 arc_is_metadata(arc_buf_t *buf)
1591 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1595 arc_bufc_to_flags(arc_buf_contents_t type)
1599 /* metadata field is 0 if buffer contains normal data */
1601 case ARC_BUFC_METADATA:
1602 return (ARC_FLAG_BUFC_METADATA);
1606 panic("undefined ARC buffer type!");
1607 return ((uint32_t)-1);
1611 arc_buf_thaw(arc_buf_t *buf)
1613 arc_buf_hdr_t *hdr = buf->b_hdr;
1615 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1616 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1618 arc_cksum_verify(buf);
1621 * Compressed buffers do not manipulate the b_freeze_cksum.
1623 if (ARC_BUF_COMPRESSED(buf))
1626 ASSERT(HDR_HAS_L1HDR(hdr));
1627 arc_cksum_free(hdr);
1628 arc_buf_unwatch(buf);
1632 arc_buf_freeze(arc_buf_t *buf)
1634 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1637 if (ARC_BUF_COMPRESSED(buf))
1640 ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1641 arc_cksum_compute(buf);
1645 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1646 * the following functions should be used to ensure that the flags are
1647 * updated in a thread-safe way. When manipulating the flags either
1648 * the hash_lock must be held or the hdr must be undiscoverable. This
1649 * ensures that we're not racing with any other threads when updating
1653 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1655 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1656 hdr->b_flags |= flags;
1660 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1662 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1663 hdr->b_flags &= ~flags;
1667 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1668 * done in a special way since we have to clear and set bits
1669 * at the same time. Consumers that wish to set the compression bits
1670 * must use this function to ensure that the flags are updated in
1671 * thread-safe manner.
1674 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1676 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1679 * Holes and embedded blocks will always have a psize = 0 so
1680 * we ignore the compression of the blkptr and set the
1681 * want to uncompress them. Mark them as uncompressed.
1683 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1684 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1685 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1687 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1688 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1691 HDR_SET_COMPRESS(hdr, cmp);
1692 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1696 * Looks for another buf on the same hdr which has the data decompressed, copies
1697 * from it, and returns true. If no such buf exists, returns false.
1700 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1702 arc_buf_hdr_t *hdr = buf->b_hdr;
1703 boolean_t copied = B_FALSE;
1705 ASSERT(HDR_HAS_L1HDR(hdr));
1706 ASSERT3P(buf->b_data, !=, NULL);
1707 ASSERT(!ARC_BUF_COMPRESSED(buf));
1709 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1710 from = from->b_next) {
1711 /* can't use our own data buffer */
1716 if (!ARC_BUF_COMPRESSED(from)) {
1717 memcpy(buf->b_data, from->b_data, arc_buf_size(buf));
1725 * There were no decompressed bufs, so there should not be a
1726 * checksum on the hdr either.
1728 if (zfs_flags & ZFS_DEBUG_MODIFY)
1729 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1736 * Allocates an ARC buf header that's in an evicted & L2-cached state.
1737 * This is used during l2arc reconstruction to make empty ARC buffers
1738 * which circumvent the regular disk->arc->l2arc path and instead come
1739 * into being in the reverse order, i.e. l2arc->arc.
1741 static arc_buf_hdr_t *
1742 arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
1743 dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth,
1744 enum zio_compress compress, uint8_t complevel, boolean_t protected,
1745 boolean_t prefetch, arc_state_type_t arcs_state)
1750 hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP);
1751 hdr->b_birth = birth;
1754 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR);
1755 HDR_SET_LSIZE(hdr, size);
1756 HDR_SET_PSIZE(hdr, psize);
1757 arc_hdr_set_compress(hdr, compress);
1758 hdr->b_complevel = complevel;
1760 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
1762 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
1763 hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa);
1767 hdr->b_l2hdr.b_dev = dev;
1768 hdr->b_l2hdr.b_daddr = daddr;
1769 hdr->b_l2hdr.b_arcs_state = arcs_state;
1775 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1778 arc_hdr_size(arc_buf_hdr_t *hdr)
1782 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1783 HDR_GET_PSIZE(hdr) > 0) {
1784 size = HDR_GET_PSIZE(hdr);
1786 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1787 size = HDR_GET_LSIZE(hdr);
1793 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1797 uint64_t lsize = HDR_GET_LSIZE(hdr);
1798 uint64_t psize = HDR_GET_PSIZE(hdr);
1799 void *tmpbuf = NULL;
1800 abd_t *abd = hdr->b_l1hdr.b_pabd;
1802 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1803 ASSERT(HDR_AUTHENTICATED(hdr));
1804 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1807 * The MAC is calculated on the compressed data that is stored on disk.
1808 * However, if compressed arc is disabled we will only have the
1809 * decompressed data available to us now. Compress it into a temporary
1810 * abd so we can verify the MAC. The performance overhead of this will
1811 * be relatively low, since most objects in an encrypted objset will
1812 * be encrypted (instead of authenticated) anyway.
1814 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1815 !HDR_COMPRESSION_ENABLED(hdr)) {
1817 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1818 hdr->b_l1hdr.b_pabd, &tmpbuf, lsize, hdr->b_complevel);
1819 ASSERT3P(tmpbuf, !=, NULL);
1820 ASSERT3U(csize, <=, psize);
1821 abd = abd_get_from_buf(tmpbuf, lsize);
1822 abd_take_ownership_of_buf(abd, B_TRUE);
1823 abd_zero_off(abd, csize, psize - csize);
1827 * Authentication is best effort. We authenticate whenever the key is
1828 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1830 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1831 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1832 ASSERT3U(lsize, ==, psize);
1833 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1834 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1836 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1837 hdr->b_crypt_hdr.b_mac);
1841 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1842 else if (ret != ENOENT)
1858 * This function will take a header that only has raw encrypted data in
1859 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1860 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1861 * also decompress the data.
1864 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1869 boolean_t no_crypt = B_FALSE;
1870 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1872 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1873 ASSERT(HDR_ENCRYPTED(hdr));
1875 arc_hdr_alloc_abd(hdr, 0);
1877 ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1878 B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1879 hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1880 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1885 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1886 HDR_GET_PSIZE(hdr));
1890 * If this header has disabled arc compression but the b_pabd is
1891 * compressed after decrypting it, we need to decompress the newly
1894 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1895 !HDR_COMPRESSION_ENABLED(hdr)) {
1897 * We want to make sure that we are correctly honoring the
1898 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1899 * and then loan a buffer from it, rather than allocating a
1900 * linear buffer and wrapping it in an abd later.
1902 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr, 0);
1903 tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
1905 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1906 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
1907 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
1909 abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
1913 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
1914 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1915 arc_hdr_size(hdr), hdr);
1916 hdr->b_l1hdr.b_pabd = cabd;
1922 arc_hdr_free_abd(hdr, B_FALSE);
1924 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
1930 * This function is called during arc_buf_fill() to prepare the header's
1931 * abd plaintext pointer for use. This involves authenticated protected
1932 * data and decrypting encrypted data into the plaintext abd.
1935 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
1936 const zbookmark_phys_t *zb, boolean_t noauth)
1940 ASSERT(HDR_PROTECTED(hdr));
1942 if (hash_lock != NULL)
1943 mutex_enter(hash_lock);
1945 if (HDR_NOAUTH(hdr) && !noauth) {
1947 * The caller requested authenticated data but our data has
1948 * not been authenticated yet. Verify the MAC now if we can.
1950 ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
1953 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
1955 * If we only have the encrypted version of the data, but the
1956 * unencrypted version was requested we take this opportunity
1957 * to store the decrypted version in the header for future use.
1959 ret = arc_hdr_decrypt(hdr, spa, zb);
1964 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1966 if (hash_lock != NULL)
1967 mutex_exit(hash_lock);
1972 if (hash_lock != NULL)
1973 mutex_exit(hash_lock);
1979 * This function is used by the dbuf code to decrypt bonus buffers in place.
1980 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
1981 * block, so we use the hash lock here to protect against concurrent calls to
1985 arc_buf_untransform_in_place(arc_buf_t *buf)
1987 arc_buf_hdr_t *hdr = buf->b_hdr;
1989 ASSERT(HDR_ENCRYPTED(hdr));
1990 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1991 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1992 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1994 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
1996 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
1997 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1998 hdr->b_crypt_hdr.b_ebufcnt -= 1;
2002 * Given a buf that has a data buffer attached to it, this function will
2003 * efficiently fill the buf with data of the specified compression setting from
2004 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2005 * are already sharing a data buf, no copy is performed.
2007 * If the buf is marked as compressed but uncompressed data was requested, this
2008 * will allocate a new data buffer for the buf, remove that flag, and fill the
2009 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2010 * uncompressed data, and (since we haven't added support for it yet) if you
2011 * want compressed data your buf must already be marked as compressed and have
2012 * the correct-sized data buffer.
2015 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2016 arc_fill_flags_t flags)
2019 arc_buf_hdr_t *hdr = buf->b_hdr;
2020 boolean_t hdr_compressed =
2021 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
2022 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
2023 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
2024 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2025 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
2027 ASSERT3P(buf->b_data, !=, NULL);
2028 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
2029 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2030 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
2031 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
2032 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
2033 IMPLY(encrypted, !ARC_BUF_SHARED(buf));
2036 * If the caller wanted encrypted data we just need to copy it from
2037 * b_rabd and potentially byteswap it. We won't be able to do any
2038 * further transforms on it.
2041 ASSERT(HDR_HAS_RABD(hdr));
2042 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
2043 HDR_GET_PSIZE(hdr));
2048 * Adjust encrypted and authenticated headers to accommodate
2049 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
2050 * allowed to fail decryption due to keys not being loaded
2051 * without being marked as an IO error.
2053 if (HDR_PROTECTED(hdr)) {
2054 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
2055 zb, !!(flags & ARC_FILL_NOAUTH));
2056 if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
2058 } else if (error != 0) {
2059 if (hash_lock != NULL)
2060 mutex_enter(hash_lock);
2061 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2062 if (hash_lock != NULL)
2063 mutex_exit(hash_lock);
2069 * There is a special case here for dnode blocks which are
2070 * decrypting their bonus buffers. These blocks may request to
2071 * be decrypted in-place. This is necessary because there may
2072 * be many dnodes pointing into this buffer and there is
2073 * currently no method to synchronize replacing the backing
2074 * b_data buffer and updating all of the pointers. Here we use
2075 * the hash lock to ensure there are no races. If the need
2076 * arises for other types to be decrypted in-place, they must
2077 * add handling here as well.
2079 if ((flags & ARC_FILL_IN_PLACE) != 0) {
2080 ASSERT(!hdr_compressed);
2081 ASSERT(!compressed);
2084 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
2085 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2087 if (hash_lock != NULL)
2088 mutex_enter(hash_lock);
2089 arc_buf_untransform_in_place(buf);
2090 if (hash_lock != NULL)
2091 mutex_exit(hash_lock);
2093 /* Compute the hdr's checksum if necessary */
2094 arc_cksum_compute(buf);
2100 if (hdr_compressed == compressed) {
2101 if (!arc_buf_is_shared(buf)) {
2102 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2106 ASSERT(hdr_compressed);
2107 ASSERT(!compressed);
2110 * If the buf is sharing its data with the hdr, unlink it and
2111 * allocate a new data buffer for the buf.
2113 if (arc_buf_is_shared(buf)) {
2114 ASSERT(ARC_BUF_COMPRESSED(buf));
2116 /* We need to give the buf its own b_data */
2117 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2119 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2120 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2122 /* Previously overhead was 0; just add new overhead */
2123 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2124 } else if (ARC_BUF_COMPRESSED(buf)) {
2125 /* We need to reallocate the buf's b_data */
2126 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2129 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2131 /* We increased the size of b_data; update overhead */
2132 ARCSTAT_INCR(arcstat_overhead_size,
2133 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2137 * Regardless of the buf's previous compression settings, it
2138 * should not be compressed at the end of this function.
2140 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2143 * Try copying the data from another buf which already has a
2144 * decompressed version. If that's not possible, it's time to
2145 * bite the bullet and decompress the data from the hdr.
2147 if (arc_buf_try_copy_decompressed_data(buf)) {
2148 /* Skip byteswapping and checksumming (already done) */
2151 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2152 hdr->b_l1hdr.b_pabd, buf->b_data,
2153 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr),
2157 * Absent hardware errors or software bugs, this should
2158 * be impossible, but log it anyway so we can debug it.
2162 "hdr %px, compress %d, psize %d, lsize %d",
2163 hdr, arc_hdr_get_compress(hdr),
2164 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2165 if (hash_lock != NULL)
2166 mutex_enter(hash_lock);
2167 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2168 if (hash_lock != NULL)
2169 mutex_exit(hash_lock);
2170 return (SET_ERROR(EIO));
2176 /* Byteswap the buf's data if necessary */
2177 if (bswap != DMU_BSWAP_NUMFUNCS) {
2178 ASSERT(!HDR_SHARED_DATA(hdr));
2179 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2180 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2183 /* Compute the hdr's checksum if necessary */
2184 arc_cksum_compute(buf);
2190 * If this function is being called to decrypt an encrypted buffer or verify an
2191 * authenticated one, the key must be loaded and a mapping must be made
2192 * available in the keystore via spa_keystore_create_mapping() or one of its
2196 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2200 arc_fill_flags_t flags = 0;
2203 flags |= ARC_FILL_IN_PLACE;
2205 ret = arc_buf_fill(buf, spa, zb, flags);
2206 if (ret == ECKSUM) {
2208 * Convert authentication and decryption errors to EIO
2209 * (and generate an ereport) before leaving the ARC.
2211 ret = SET_ERROR(EIO);
2212 spa_log_error(spa, zb, &buf->b_hdr->b_birth);
2213 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2214 spa, NULL, zb, NULL, 0);
2221 * Increment the amount of evictable space in the arc_state_t's refcount.
2222 * We account for the space used by the hdr and the arc buf individually
2223 * so that we can add and remove them from the refcount individually.
2226 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2228 arc_buf_contents_t type = arc_buf_type(hdr);
2230 ASSERT(HDR_HAS_L1HDR(hdr));
2232 if (GHOST_STATE(state)) {
2233 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2234 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2235 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2236 ASSERT(!HDR_HAS_RABD(hdr));
2237 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2238 HDR_GET_LSIZE(hdr), hdr);
2242 if (hdr->b_l1hdr.b_pabd != NULL) {
2243 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2244 arc_hdr_size(hdr), hdr);
2246 if (HDR_HAS_RABD(hdr)) {
2247 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2248 HDR_GET_PSIZE(hdr), hdr);
2251 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2252 buf = buf->b_next) {
2253 if (arc_buf_is_shared(buf))
2255 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2256 arc_buf_size(buf), buf);
2261 * Decrement the amount of evictable space in the arc_state_t's refcount.
2262 * We account for the space used by the hdr and the arc buf individually
2263 * so that we can add and remove them from the refcount individually.
2266 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2268 arc_buf_contents_t type = arc_buf_type(hdr);
2270 ASSERT(HDR_HAS_L1HDR(hdr));
2272 if (GHOST_STATE(state)) {
2273 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2274 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2275 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2276 ASSERT(!HDR_HAS_RABD(hdr));
2277 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2278 HDR_GET_LSIZE(hdr), hdr);
2282 if (hdr->b_l1hdr.b_pabd != NULL) {
2283 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2284 arc_hdr_size(hdr), hdr);
2286 if (HDR_HAS_RABD(hdr)) {
2287 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2288 HDR_GET_PSIZE(hdr), hdr);
2291 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2292 buf = buf->b_next) {
2293 if (arc_buf_is_shared(buf))
2295 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2296 arc_buf_size(buf), buf);
2301 * Add a reference to this hdr indicating that someone is actively
2302 * referencing that memory. When the refcount transitions from 0 to 1,
2303 * we remove it from the respective arc_state_t list to indicate that
2304 * it is not evictable.
2307 add_reference(arc_buf_hdr_t *hdr, const void *tag)
2309 arc_state_t *state = hdr->b_l1hdr.b_state;
2311 ASSERT(HDR_HAS_L1HDR(hdr));
2312 if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) {
2313 ASSERT(state == arc_anon);
2314 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2315 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2318 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2319 state != arc_anon && state != arc_l2c_only) {
2320 /* We don't use the L2-only state list. */
2321 multilist_remove(&state->arcs_list[arc_buf_type(hdr)], hdr);
2322 arc_evictable_space_decrement(hdr, state);
2327 * Remove a reference from this hdr. When the reference transitions from
2328 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2329 * list making it eligible for eviction.
2332 remove_reference(arc_buf_hdr_t *hdr, const void *tag)
2335 arc_state_t *state = hdr->b_l1hdr.b_state;
2337 ASSERT(HDR_HAS_L1HDR(hdr));
2338 ASSERT(state == arc_anon || MUTEX_HELD(HDR_LOCK(hdr)));
2339 ASSERT(!GHOST_STATE(state)); /* arc_l2c_only counts as a ghost. */
2341 if ((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) != 0)
2344 if (state == arc_anon) {
2345 arc_hdr_destroy(hdr);
2348 if (state == arc_uncached && !HDR_PREFETCH(hdr)) {
2349 arc_change_state(arc_anon, hdr);
2350 arc_hdr_destroy(hdr);
2353 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
2354 arc_evictable_space_increment(hdr, state);
2359 * Returns detailed information about a specific arc buffer. When the
2360 * state_index argument is set the function will calculate the arc header
2361 * list position for its arc state. Since this requires a linear traversal
2362 * callers are strongly encourage not to do this. However, it can be helpful
2363 * for targeted analysis so the functionality is provided.
2366 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2369 arc_buf_hdr_t *hdr = ab->b_hdr;
2370 l1arc_buf_hdr_t *l1hdr = NULL;
2371 l2arc_buf_hdr_t *l2hdr = NULL;
2372 arc_state_t *state = NULL;
2374 memset(abi, 0, sizeof (arc_buf_info_t));
2379 abi->abi_flags = hdr->b_flags;
2381 if (HDR_HAS_L1HDR(hdr)) {
2382 l1hdr = &hdr->b_l1hdr;
2383 state = l1hdr->b_state;
2385 if (HDR_HAS_L2HDR(hdr))
2386 l2hdr = &hdr->b_l2hdr;
2389 abi->abi_bufcnt = l1hdr->b_bufcnt;
2390 abi->abi_access = l1hdr->b_arc_access;
2391 abi->abi_mru_hits = l1hdr->b_mru_hits;
2392 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2393 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2394 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2395 abi->abi_holds = zfs_refcount_count(&l1hdr->b_refcnt);
2399 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2400 abi->abi_l2arc_hits = l2hdr->b_hits;
2403 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2404 abi->abi_state_contents = arc_buf_type(hdr);
2405 abi->abi_size = arc_hdr_size(hdr);
2409 * Move the supplied buffer to the indicated state. The hash lock
2410 * for the buffer must be held by the caller.
2413 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr)
2415 arc_state_t *old_state;
2418 boolean_t update_old, update_new;
2419 arc_buf_contents_t type = arc_buf_type(hdr);
2422 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2423 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2424 * L1 hdr doesn't always exist when we change state to arc_anon before
2425 * destroying a header, in which case reallocating to add the L1 hdr is
2428 if (HDR_HAS_L1HDR(hdr)) {
2429 old_state = hdr->b_l1hdr.b_state;
2430 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2431 bufcnt = hdr->b_l1hdr.b_bufcnt;
2432 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
2435 IMPLY(GHOST_STATE(old_state), bufcnt == 0);
2436 IMPLY(GHOST_STATE(new_state), bufcnt == 0);
2437 IMPLY(GHOST_STATE(old_state), hdr->b_l1hdr.b_buf == NULL);
2438 IMPLY(GHOST_STATE(new_state), hdr->b_l1hdr.b_buf == NULL);
2439 IMPLY(old_state == arc_anon, bufcnt <= 1);
2441 old_state = arc_l2c_only;
2444 update_old = B_FALSE;
2446 update_new = update_old;
2447 if (GHOST_STATE(old_state))
2448 update_old = B_TRUE;
2449 if (GHOST_STATE(new_state))
2450 update_new = B_TRUE;
2452 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
2453 ASSERT3P(new_state, !=, old_state);
2456 * If this buffer is evictable, transfer it from the
2457 * old state list to the new state list.
2460 if (old_state != arc_anon && old_state != arc_l2c_only) {
2461 ASSERT(HDR_HAS_L1HDR(hdr));
2462 /* remove_reference() saves on insert. */
2463 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2464 multilist_remove(&old_state->arcs_list[type],
2466 arc_evictable_space_decrement(hdr, old_state);
2469 if (new_state != arc_anon && new_state != arc_l2c_only) {
2471 * An L1 header always exists here, since if we're
2472 * moving to some L1-cached state (i.e. not l2c_only or
2473 * anonymous), we realloc the header to add an L1hdr
2476 ASSERT(HDR_HAS_L1HDR(hdr));
2477 multilist_insert(&new_state->arcs_list[type], hdr);
2478 arc_evictable_space_increment(hdr, new_state);
2482 ASSERT(!HDR_EMPTY(hdr));
2483 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2484 buf_hash_remove(hdr);
2486 /* adjust state sizes (ignore arc_l2c_only) */
2488 if (update_new && new_state != arc_l2c_only) {
2489 ASSERT(HDR_HAS_L1HDR(hdr));
2490 if (GHOST_STATE(new_state)) {
2494 * When moving a header to a ghost state, we first
2495 * remove all arc buffers. Thus, we'll have a
2496 * bufcnt of zero, and no arc buffer to use for
2497 * the reference. As a result, we use the arc
2498 * header pointer for the reference.
2500 (void) zfs_refcount_add_many(
2501 &new_state->arcs_size[type],
2502 HDR_GET_LSIZE(hdr), hdr);
2503 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2504 ASSERT(!HDR_HAS_RABD(hdr));
2506 uint32_t buffers = 0;
2509 * Each individual buffer holds a unique reference,
2510 * thus we must remove each of these references one
2513 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2514 buf = buf->b_next) {
2515 ASSERT3U(bufcnt, !=, 0);
2519 * When the arc_buf_t is sharing the data
2520 * block with the hdr, the owner of the
2521 * reference belongs to the hdr. Only
2522 * add to the refcount if the arc_buf_t is
2525 if (arc_buf_is_shared(buf))
2528 (void) zfs_refcount_add_many(
2529 &new_state->arcs_size[type],
2530 arc_buf_size(buf), buf);
2532 ASSERT3U(bufcnt, ==, buffers);
2534 if (hdr->b_l1hdr.b_pabd != NULL) {
2535 (void) zfs_refcount_add_many(
2536 &new_state->arcs_size[type],
2537 arc_hdr_size(hdr), hdr);
2540 if (HDR_HAS_RABD(hdr)) {
2541 (void) zfs_refcount_add_many(
2542 &new_state->arcs_size[type],
2543 HDR_GET_PSIZE(hdr), hdr);
2548 if (update_old && old_state != arc_l2c_only) {
2549 ASSERT(HDR_HAS_L1HDR(hdr));
2550 if (GHOST_STATE(old_state)) {
2552 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2553 ASSERT(!HDR_HAS_RABD(hdr));
2556 * When moving a header off of a ghost state,
2557 * the header will not contain any arc buffers.
2558 * We use the arc header pointer for the reference
2559 * which is exactly what we did when we put the
2560 * header on the ghost state.
2563 (void) zfs_refcount_remove_many(
2564 &old_state->arcs_size[type],
2565 HDR_GET_LSIZE(hdr), hdr);
2567 uint32_t buffers = 0;
2570 * Each individual buffer holds a unique reference,
2571 * thus we must remove each of these references one
2574 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2575 buf = buf->b_next) {
2576 ASSERT3U(bufcnt, !=, 0);
2580 * When the arc_buf_t is sharing the data
2581 * block with the hdr, the owner of the
2582 * reference belongs to the hdr. Only
2583 * add to the refcount if the arc_buf_t is
2586 if (arc_buf_is_shared(buf))
2589 (void) zfs_refcount_remove_many(
2590 &old_state->arcs_size[type],
2591 arc_buf_size(buf), buf);
2593 ASSERT3U(bufcnt, ==, buffers);
2594 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2597 if (hdr->b_l1hdr.b_pabd != NULL) {
2598 (void) zfs_refcount_remove_many(
2599 &old_state->arcs_size[type],
2600 arc_hdr_size(hdr), hdr);
2603 if (HDR_HAS_RABD(hdr)) {
2604 (void) zfs_refcount_remove_many(
2605 &old_state->arcs_size[type],
2606 HDR_GET_PSIZE(hdr), hdr);
2611 if (HDR_HAS_L1HDR(hdr)) {
2612 hdr->b_l1hdr.b_state = new_state;
2614 if (HDR_HAS_L2HDR(hdr) && new_state != arc_l2c_only) {
2615 l2arc_hdr_arcstats_decrement_state(hdr);
2616 hdr->b_l2hdr.b_arcs_state = new_state->arcs_state;
2617 l2arc_hdr_arcstats_increment_state(hdr);
2623 arc_space_consume(uint64_t space, arc_space_type_t type)
2625 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2630 case ARC_SPACE_DATA:
2631 ARCSTAT_INCR(arcstat_data_size, space);
2633 case ARC_SPACE_META:
2634 ARCSTAT_INCR(arcstat_metadata_size, space);
2636 case ARC_SPACE_BONUS:
2637 ARCSTAT_INCR(arcstat_bonus_size, space);
2639 case ARC_SPACE_DNODE:
2640 ARCSTAT_INCR(arcstat_dnode_size, space);
2642 case ARC_SPACE_DBUF:
2643 ARCSTAT_INCR(arcstat_dbuf_size, space);
2645 case ARC_SPACE_HDRS:
2646 ARCSTAT_INCR(arcstat_hdr_size, space);
2648 case ARC_SPACE_L2HDRS:
2649 aggsum_add(&arc_sums.arcstat_l2_hdr_size, space);
2651 case ARC_SPACE_ABD_CHUNK_WASTE:
2653 * Note: this includes space wasted by all scatter ABD's, not
2654 * just those allocated by the ARC. But the vast majority of
2655 * scatter ABD's come from the ARC, because other users are
2658 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, space);
2662 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE)
2663 ARCSTAT_INCR(arcstat_meta_used, space);
2665 aggsum_add(&arc_sums.arcstat_size, space);
2669 arc_space_return(uint64_t space, arc_space_type_t type)
2671 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2676 case ARC_SPACE_DATA:
2677 ARCSTAT_INCR(arcstat_data_size, -space);
2679 case ARC_SPACE_META:
2680 ARCSTAT_INCR(arcstat_metadata_size, -space);
2682 case ARC_SPACE_BONUS:
2683 ARCSTAT_INCR(arcstat_bonus_size, -space);
2685 case ARC_SPACE_DNODE:
2686 ARCSTAT_INCR(arcstat_dnode_size, -space);
2688 case ARC_SPACE_DBUF:
2689 ARCSTAT_INCR(arcstat_dbuf_size, -space);
2691 case ARC_SPACE_HDRS:
2692 ARCSTAT_INCR(arcstat_hdr_size, -space);
2694 case ARC_SPACE_L2HDRS:
2695 aggsum_add(&arc_sums.arcstat_l2_hdr_size, -space);
2697 case ARC_SPACE_ABD_CHUNK_WASTE:
2698 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, -space);
2702 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE)
2703 ARCSTAT_INCR(arcstat_meta_used, -space);
2705 ASSERT(aggsum_compare(&arc_sums.arcstat_size, space) >= 0);
2706 aggsum_add(&arc_sums.arcstat_size, -space);
2710 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2711 * with the hdr's b_pabd.
2714 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2717 * The criteria for sharing a hdr's data are:
2718 * 1. the buffer is not encrypted
2719 * 2. the hdr's compression matches the buf's compression
2720 * 3. the hdr doesn't need to be byteswapped
2721 * 4. the hdr isn't already being shared
2722 * 5. the buf is either compressed or it is the last buf in the hdr list
2724 * Criterion #5 maintains the invariant that shared uncompressed
2725 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2726 * might ask, "if a compressed buf is allocated first, won't that be the
2727 * last thing in the list?", but in that case it's impossible to create
2728 * a shared uncompressed buf anyway (because the hdr must be compressed
2729 * to have the compressed buf). You might also think that #3 is
2730 * sufficient to make this guarantee, however it's possible
2731 * (specifically in the rare L2ARC write race mentioned in
2732 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2733 * is shareable, but wasn't at the time of its allocation. Rather than
2734 * allow a new shared uncompressed buf to be created and then shuffle
2735 * the list around to make it the last element, this simply disallows
2736 * sharing if the new buf isn't the first to be added.
2738 ASSERT3P(buf->b_hdr, ==, hdr);
2739 boolean_t hdr_compressed =
2740 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF;
2741 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2742 return (!ARC_BUF_ENCRYPTED(buf) &&
2743 buf_compressed == hdr_compressed &&
2744 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2745 !HDR_SHARED_DATA(hdr) &&
2746 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2750 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2751 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2752 * copy was made successfully, or an error code otherwise.
2755 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2756 const void *tag, boolean_t encrypted, boolean_t compressed,
2757 boolean_t noauth, boolean_t fill, arc_buf_t **ret)
2760 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2762 ASSERT(HDR_HAS_L1HDR(hdr));
2763 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2764 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2765 hdr->b_type == ARC_BUFC_METADATA);
2766 ASSERT3P(ret, !=, NULL);
2767 ASSERT3P(*ret, ==, NULL);
2768 IMPLY(encrypted, compressed);
2770 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2773 buf->b_next = hdr->b_l1hdr.b_buf;
2776 add_reference(hdr, tag);
2779 * We're about to change the hdr's b_flags. We must either
2780 * hold the hash_lock or be undiscoverable.
2782 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2785 * Only honor requests for compressed bufs if the hdr is actually
2786 * compressed. This must be overridden if the buffer is encrypted since
2787 * encrypted buffers cannot be decompressed.
2790 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2791 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2792 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2793 } else if (compressed &&
2794 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2795 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2796 flags |= ARC_FILL_COMPRESSED;
2801 flags |= ARC_FILL_NOAUTH;
2805 * If the hdr's data can be shared then we share the data buffer and
2806 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2807 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2808 * buffer to store the buf's data.
2810 * There are two additional restrictions here because we're sharing
2811 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2812 * actively involved in an L2ARC write, because if this buf is used by
2813 * an arc_write() then the hdr's data buffer will be released when the
2814 * write completes, even though the L2ARC write might still be using it.
2815 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2816 * need to be ABD-aware. It must be allocated via
2817 * zio_[data_]buf_alloc(), not as a page, because we need to be able
2818 * to abd_release_ownership_of_buf(), which isn't allowed on "linear
2819 * page" buffers because the ABD code needs to handle freeing them
2822 boolean_t can_share = arc_can_share(hdr, buf) &&
2823 !HDR_L2_WRITING(hdr) &&
2824 hdr->b_l1hdr.b_pabd != NULL &&
2825 abd_is_linear(hdr->b_l1hdr.b_pabd) &&
2826 !abd_is_linear_page(hdr->b_l1hdr.b_pabd);
2828 /* Set up b_data and sharing */
2830 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2831 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2832 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2835 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2836 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2838 VERIFY3P(buf->b_data, !=, NULL);
2840 hdr->b_l1hdr.b_buf = buf;
2841 hdr->b_l1hdr.b_bufcnt += 1;
2843 hdr->b_crypt_hdr.b_ebufcnt += 1;
2846 * If the user wants the data from the hdr, we need to either copy or
2847 * decompress the data.
2850 ASSERT3P(zb, !=, NULL);
2851 return (arc_buf_fill(buf, spa, zb, flags));
2857 static const char *arc_onloan_tag = "onloan";
2860 arc_loaned_bytes_update(int64_t delta)
2862 atomic_add_64(&arc_loaned_bytes, delta);
2864 /* assert that it did not wrap around */
2865 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2869 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2870 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2871 * buffers must be returned to the arc before they can be used by the DMU or
2875 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2877 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2878 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2880 arc_loaned_bytes_update(arc_buf_size(buf));
2886 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2887 enum zio_compress compression_type, uint8_t complevel)
2889 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2890 psize, lsize, compression_type, complevel);
2892 arc_loaned_bytes_update(arc_buf_size(buf));
2898 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2899 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2900 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2901 enum zio_compress compression_type, uint8_t complevel)
2903 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2904 byteorder, salt, iv, mac, ot, psize, lsize, compression_type,
2907 atomic_add_64(&arc_loaned_bytes, psize);
2913 * Return a loaned arc buffer to the arc.
2916 arc_return_buf(arc_buf_t *buf, const void *tag)
2918 arc_buf_hdr_t *hdr = buf->b_hdr;
2920 ASSERT3P(buf->b_data, !=, NULL);
2921 ASSERT(HDR_HAS_L1HDR(hdr));
2922 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2923 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2925 arc_loaned_bytes_update(-arc_buf_size(buf));
2928 /* Detach an arc_buf from a dbuf (tag) */
2930 arc_loan_inuse_buf(arc_buf_t *buf, const void *tag)
2932 arc_buf_hdr_t *hdr = buf->b_hdr;
2934 ASSERT3P(buf->b_data, !=, NULL);
2935 ASSERT(HDR_HAS_L1HDR(hdr));
2936 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2937 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2939 arc_loaned_bytes_update(arc_buf_size(buf));
2943 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2945 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2948 df->l2df_size = size;
2949 df->l2df_type = type;
2950 mutex_enter(&l2arc_free_on_write_mtx);
2951 list_insert_head(l2arc_free_on_write, df);
2952 mutex_exit(&l2arc_free_on_write_mtx);
2956 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
2958 arc_state_t *state = hdr->b_l1hdr.b_state;
2959 arc_buf_contents_t type = arc_buf_type(hdr);
2960 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
2962 /* protected by hash lock, if in the hash table */
2963 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2964 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2965 ASSERT(state != arc_anon && state != arc_l2c_only);
2967 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2970 (void) zfs_refcount_remove_many(&state->arcs_size[type], size, hdr);
2971 if (type == ARC_BUFC_METADATA) {
2972 arc_space_return(size, ARC_SPACE_META);
2974 ASSERT(type == ARC_BUFC_DATA);
2975 arc_space_return(size, ARC_SPACE_DATA);
2979 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
2981 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2986 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2987 * data buffer, we transfer the refcount ownership to the hdr and update
2988 * the appropriate kstats.
2991 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2993 ASSERT(arc_can_share(hdr, buf));
2994 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2995 ASSERT(!ARC_BUF_ENCRYPTED(buf));
2996 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2999 * Start sharing the data buffer. We transfer the
3000 * refcount ownership to the hdr since it always owns
3001 * the refcount whenever an arc_buf_t is shared.
3003 zfs_refcount_transfer_ownership_many(
3004 &hdr->b_l1hdr.b_state->arcs_size[arc_buf_type(hdr)],
3005 arc_hdr_size(hdr), buf, hdr);
3006 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
3007 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
3008 HDR_ISTYPE_METADATA(hdr));
3009 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3010 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3013 * Since we've transferred ownership to the hdr we need
3014 * to increment its compressed and uncompressed kstats and
3015 * decrement the overhead size.
3017 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3018 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3019 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3023 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3025 ASSERT(arc_buf_is_shared(buf));
3026 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3027 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3030 * We are no longer sharing this buffer so we need
3031 * to transfer its ownership to the rightful owner.
3033 zfs_refcount_transfer_ownership_many(
3034 &hdr->b_l1hdr.b_state->arcs_size[arc_buf_type(hdr)],
3035 arc_hdr_size(hdr), hdr, buf);
3036 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3037 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3038 abd_free(hdr->b_l1hdr.b_pabd);
3039 hdr->b_l1hdr.b_pabd = NULL;
3040 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3043 * Since the buffer is no longer shared between
3044 * the arc buf and the hdr, count it as overhead.
3046 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3047 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3048 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3052 * Remove an arc_buf_t from the hdr's buf list and return the last
3053 * arc_buf_t on the list. If no buffers remain on the list then return
3057 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3059 ASSERT(HDR_HAS_L1HDR(hdr));
3060 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3062 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3063 arc_buf_t *lastbuf = NULL;
3066 * Remove the buf from the hdr list and locate the last
3067 * remaining buffer on the list.
3069 while (*bufp != NULL) {
3071 *bufp = buf->b_next;
3074 * If we've removed a buffer in the middle of
3075 * the list then update the lastbuf and update
3078 if (*bufp != NULL) {
3080 bufp = &(*bufp)->b_next;
3084 ASSERT3P(lastbuf, !=, buf);
3085 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3086 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3087 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3093 * Free up buf->b_data and pull the arc_buf_t off of the arc_buf_hdr_t's
3097 arc_buf_destroy_impl(arc_buf_t *buf)
3099 arc_buf_hdr_t *hdr = buf->b_hdr;
3102 * Free up the data associated with the buf but only if we're not
3103 * sharing this with the hdr. If we are sharing it with the hdr, the
3104 * hdr is responsible for doing the free.
3106 if (buf->b_data != NULL) {
3108 * We're about to change the hdr's b_flags. We must either
3109 * hold the hash_lock or be undiscoverable.
3111 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3113 arc_cksum_verify(buf);
3114 arc_buf_unwatch(buf);
3116 if (arc_buf_is_shared(buf)) {
3117 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3119 uint64_t size = arc_buf_size(buf);
3120 arc_free_data_buf(hdr, buf->b_data, size, buf);
3121 ARCSTAT_INCR(arcstat_overhead_size, -size);
3125 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3126 hdr->b_l1hdr.b_bufcnt -= 1;
3128 if (ARC_BUF_ENCRYPTED(buf)) {
3129 hdr->b_crypt_hdr.b_ebufcnt -= 1;
3132 * If we have no more encrypted buffers and we've
3133 * already gotten a copy of the decrypted data we can
3134 * free b_rabd to save some space.
3136 if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
3137 HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
3138 !HDR_IO_IN_PROGRESS(hdr)) {
3139 arc_hdr_free_abd(hdr, B_TRUE);
3144 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3146 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3148 * If the current arc_buf_t is sharing its data buffer with the
3149 * hdr, then reassign the hdr's b_pabd to share it with the new
3150 * buffer at the end of the list. The shared buffer is always
3151 * the last one on the hdr's buffer list.
3153 * There is an equivalent case for compressed bufs, but since
3154 * they aren't guaranteed to be the last buf in the list and
3155 * that is an exceedingly rare case, we just allow that space be
3156 * wasted temporarily. We must also be careful not to share
3157 * encrypted buffers, since they cannot be shared.
3159 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
3160 /* Only one buf can be shared at once */
3161 VERIFY(!arc_buf_is_shared(lastbuf));
3162 /* hdr is uncompressed so can't have compressed buf */
3163 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3165 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3166 arc_hdr_free_abd(hdr, B_FALSE);
3169 * We must setup a new shared block between the
3170 * last buffer and the hdr. The data would have
3171 * been allocated by the arc buf so we need to transfer
3172 * ownership to the hdr since it's now being shared.
3174 arc_share_buf(hdr, lastbuf);
3176 } else if (HDR_SHARED_DATA(hdr)) {
3178 * Uncompressed shared buffers are always at the end
3179 * of the list. Compressed buffers don't have the
3180 * same requirements. This makes it hard to
3181 * simply assert that the lastbuf is shared so
3182 * we rely on the hdr's compression flags to determine
3183 * if we have a compressed, shared buffer.
3185 ASSERT3P(lastbuf, !=, NULL);
3186 ASSERT(arc_buf_is_shared(lastbuf) ||
3187 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3191 * Free the checksum if we're removing the last uncompressed buf from
3194 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3195 arc_cksum_free(hdr);
3198 /* clean up the buf */
3200 kmem_cache_free(buf_cache, buf);
3204 arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, int alloc_flags)
3207 boolean_t alloc_rdata = ((alloc_flags & ARC_HDR_ALLOC_RDATA) != 0);
3209 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3210 ASSERT(HDR_HAS_L1HDR(hdr));
3211 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3212 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3215 size = HDR_GET_PSIZE(hdr);
3216 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3217 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr,
3219 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3220 ARCSTAT_INCR(arcstat_raw_size, size);
3222 size = arc_hdr_size(hdr);
3223 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3224 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr,
3226 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3229 ARCSTAT_INCR(arcstat_compressed_size, size);
3230 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3234 arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3236 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3238 ASSERT(HDR_HAS_L1HDR(hdr));
3239 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3240 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3243 * If the hdr is currently being written to the l2arc then
3244 * we defer freeing the data by adding it to the l2arc_free_on_write
3245 * list. The l2arc will free the data once it's finished
3246 * writing it to the l2arc device.
3248 if (HDR_L2_WRITING(hdr)) {
3249 arc_hdr_free_on_write(hdr, free_rdata);
3250 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3251 } else if (free_rdata) {
3252 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3254 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr);
3258 hdr->b_crypt_hdr.b_rabd = NULL;
3259 ARCSTAT_INCR(arcstat_raw_size, -size);
3261 hdr->b_l1hdr.b_pabd = NULL;
3264 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3265 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3267 ARCSTAT_INCR(arcstat_compressed_size, -size);
3268 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3272 * Allocate empty anonymous ARC header. The header will get its identity
3273 * assigned and buffers attached later as part of read or write operations.
3275 * In case of read arc_read() assigns header its identify (b_dva + b_birth),
3276 * inserts it into ARC hash to become globally visible and allocates physical
3277 * (b_pabd) or raw (b_rabd) ABD buffer to read into from disk. On disk read
3278 * completion arc_read_done() allocates ARC buffer(s) as needed, potentially
3279 * sharing one of them with the physical ABD buffer.
3281 * In case of write arc_alloc_buf() allocates ARC buffer to be filled with
3282 * data. Then after compression and/or encryption arc_write_ready() allocates
3283 * and fills (or potentially shares) physical (b_pabd) or raw (b_rabd) ABD
3284 * buffer. On disk write completion arc_write_done() assigns the header its
3285 * new identity (b_dva + b_birth) and inserts into ARC hash.
3287 * In case of partial overwrite the old data is read first as described. Then
3288 * arc_release() either allocates new anonymous ARC header and moves the ARC
3289 * buffer to it, or reuses the old ARC header by discarding its identity and
3290 * removing it from ARC hash. After buffer modification normal write process
3291 * follows as described.
3293 static arc_buf_hdr_t *
3294 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3295 boolean_t protected, enum zio_compress compression_type, uint8_t complevel,
3296 arc_buf_contents_t type)
3300 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3302 hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
3304 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3307 ASSERT(HDR_EMPTY(hdr));
3309 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3311 HDR_SET_PSIZE(hdr, psize);
3312 HDR_SET_LSIZE(hdr, lsize);
3316 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3317 arc_hdr_set_compress(hdr, compression_type);
3318 hdr->b_complevel = complevel;
3320 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3322 hdr->b_l1hdr.b_state = arc_anon;
3323 hdr->b_l1hdr.b_arc_access = 0;
3324 hdr->b_l1hdr.b_mru_hits = 0;
3325 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3326 hdr->b_l1hdr.b_mfu_hits = 0;
3327 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3328 hdr->b_l1hdr.b_bufcnt = 0;
3329 hdr->b_l1hdr.b_buf = NULL;
3331 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3337 * Transition between the two allocation states for the arc_buf_hdr struct.
3338 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3339 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3340 * version is used when a cache buffer is only in the L2ARC in order to reduce
3343 static arc_buf_hdr_t *
3344 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3346 ASSERT(HDR_HAS_L2HDR(hdr));
3348 arc_buf_hdr_t *nhdr;
3349 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3351 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3352 (old == hdr_l2only_cache && new == hdr_full_cache));
3355 * if the caller wanted a new full header and the header is to be
3356 * encrypted we will actually allocate the header from the full crypt
3357 * cache instead. The same applies to freeing from the old cache.
3359 if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
3360 new = hdr_full_crypt_cache;
3361 if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
3362 old = hdr_full_crypt_cache;
3364 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3366 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3367 buf_hash_remove(hdr);
3369 memcpy(nhdr, hdr, HDR_L2ONLY_SIZE);
3371 if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
3372 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3374 * arc_access and arc_change_state need to be aware that a
3375 * header has just come out of L2ARC, so we set its state to
3376 * l2c_only even though it's about to change.
3378 nhdr->b_l1hdr.b_state = arc_l2c_only;
3380 /* Verify previous threads set to NULL before freeing */
3381 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3382 ASSERT(!HDR_HAS_RABD(hdr));
3384 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3385 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3387 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3391 * If we've reached here, We must have been called from
3392 * arc_evict_hdr(), as such we should have already been
3393 * removed from any ghost list we were previously on
3394 * (which protects us from racing with arc_evict_state),
3395 * thus no locking is needed during this check.
3397 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3400 * A buffer must not be moved into the arc_l2c_only
3401 * state if it's not finished being written out to the
3402 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3403 * might try to be accessed, even though it was removed.
3405 VERIFY(!HDR_L2_WRITING(hdr));
3406 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3407 ASSERT(!HDR_HAS_RABD(hdr));
3409 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3412 * The header has been reallocated so we need to re-insert it into any
3415 (void) buf_hash_insert(nhdr, NULL);
3417 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3419 mutex_enter(&dev->l2ad_mtx);
3422 * We must place the realloc'ed header back into the list at
3423 * the same spot. Otherwise, if it's placed earlier in the list,
3424 * l2arc_write_buffers() could find it during the function's
3425 * write phase, and try to write it out to the l2arc.
3427 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3428 list_remove(&dev->l2ad_buflist, hdr);
3430 mutex_exit(&dev->l2ad_mtx);
3433 * Since we're using the pointer address as the tag when
3434 * incrementing and decrementing the l2ad_alloc refcount, we
3435 * must remove the old pointer (that we're about to destroy) and
3436 * add the new pointer to the refcount. Otherwise we'd remove
3437 * the wrong pointer address when calling arc_hdr_destroy() later.
3440 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
3441 arc_hdr_size(hdr), hdr);
3442 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
3443 arc_hdr_size(nhdr), nhdr);
3445 buf_discard_identity(hdr);
3446 kmem_cache_free(old, hdr);
3452 * This function allows an L1 header to be reallocated as a crypt
3453 * header and vice versa. If we are going to a crypt header, the
3454 * new fields will be zeroed out.
3456 static arc_buf_hdr_t *
3457 arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
3459 arc_buf_hdr_t *nhdr;
3461 kmem_cache_t *ncache, *ocache;
3464 * This function requires that hdr is in the arc_anon state.
3465 * Therefore it won't have any L2ARC data for us to worry
3468 ASSERT(HDR_HAS_L1HDR(hdr));
3469 ASSERT(!HDR_HAS_L2HDR(hdr));
3470 ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
3471 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3472 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3473 ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node));
3474 ASSERT3P(hdr->b_hash_next, ==, NULL);
3477 ncache = hdr_full_crypt_cache;
3478 ocache = hdr_full_cache;
3480 ncache = hdr_full_cache;
3481 ocache = hdr_full_crypt_cache;
3484 nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);
3487 * Copy all members that aren't locks or condvars to the new header.
3488 * No lists are pointing to us (as we asserted above), so we don't
3489 * need to worry about the list nodes.
3491 nhdr->b_dva = hdr->b_dva;
3492 nhdr->b_birth = hdr->b_birth;
3493 nhdr->b_type = hdr->b_type;
3494 nhdr->b_flags = hdr->b_flags;
3495 nhdr->b_psize = hdr->b_psize;
3496 nhdr->b_lsize = hdr->b_lsize;
3497 nhdr->b_spa = hdr->b_spa;
3499 nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
3501 nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
3502 nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
3503 nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
3504 nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
3505 nhdr->b_l1hdr.b_mru_hits = hdr->b_l1hdr.b_mru_hits;
3506 nhdr->b_l1hdr.b_mru_ghost_hits = hdr->b_l1hdr.b_mru_ghost_hits;
3507 nhdr->b_l1hdr.b_mfu_hits = hdr->b_l1hdr.b_mfu_hits;
3508 nhdr->b_l1hdr.b_mfu_ghost_hits = hdr->b_l1hdr.b_mfu_ghost_hits;
3509 nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
3510 nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
3513 * This zfs_refcount_add() exists only to ensure that the individual
3514 * arc buffers always point to a header that is referenced, avoiding
3515 * a small race condition that could trigger ASSERTs.
3517 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
3518 nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
3519 for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next)
3522 zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
3523 (void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
3524 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3527 arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
3529 arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
3532 /* unset all members of the original hdr */
3533 memset(&hdr->b_dva, 0, sizeof (dva_t));
3541 hdr->b_l1hdr.b_freeze_cksum = NULL;
3543 hdr->b_l1hdr.b_buf = NULL;
3544 hdr->b_l1hdr.b_bufcnt = 0;
3545 hdr->b_l1hdr.b_byteswap = 0;
3546 hdr->b_l1hdr.b_state = NULL;
3547 hdr->b_l1hdr.b_arc_access = 0;
3548 hdr->b_l1hdr.b_mru_hits = 0;
3549 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3550 hdr->b_l1hdr.b_mfu_hits = 0;
3551 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3552 hdr->b_l1hdr.b_acb = NULL;
3553 hdr->b_l1hdr.b_pabd = NULL;
3555 if (ocache == hdr_full_crypt_cache) {
3556 ASSERT(!HDR_HAS_RABD(hdr));
3557 hdr->b_crypt_hdr.b_ot = DMU_OT_NONE;
3558 hdr->b_crypt_hdr.b_ebufcnt = 0;
3559 hdr->b_crypt_hdr.b_dsobj = 0;
3560 memset(hdr->b_crypt_hdr.b_salt, 0, ZIO_DATA_SALT_LEN);
3561 memset(hdr->b_crypt_hdr.b_iv, 0, ZIO_DATA_IV_LEN);
3562 memset(hdr->b_crypt_hdr.b_mac, 0, ZIO_DATA_MAC_LEN);
3565 buf_discard_identity(hdr);
3566 kmem_cache_free(ocache, hdr);
3572 * This function is used by the send / receive code to convert a newly
3573 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3574 * is also used to allow the root objset block to be updated without altering
3575 * its embedded MACs. Both block types will always be uncompressed so we do not
3576 * have to worry about compression type or psize.
3579 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3580 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3583 arc_buf_hdr_t *hdr = buf->b_hdr;
3585 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3586 ASSERT(HDR_HAS_L1HDR(hdr));
3587 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3589 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3590 if (!HDR_PROTECTED(hdr))
3591 hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
3592 hdr->b_crypt_hdr.b_dsobj = dsobj;
3593 hdr->b_crypt_hdr.b_ot = ot;
3594 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3595 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3596 if (!arc_hdr_has_uncompressed_buf(hdr))
3597 arc_cksum_free(hdr);
3600 memcpy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
3602 memcpy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
3604 memcpy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
3608 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3609 * The buf is returned thawed since we expect the consumer to modify it.
3612 arc_alloc_buf(spa_t *spa, const void *tag, arc_buf_contents_t type,
3615 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3616 B_FALSE, ZIO_COMPRESS_OFF, 0, type);
3618 arc_buf_t *buf = NULL;
3619 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3620 B_FALSE, B_FALSE, &buf));
3627 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3628 * for bufs containing metadata.
3631 arc_alloc_compressed_buf(spa_t *spa, const void *tag, uint64_t psize,
3632 uint64_t lsize, enum zio_compress compression_type, uint8_t complevel)
3634 ASSERT3U(lsize, >, 0);
3635 ASSERT3U(lsize, >=, psize);
3636 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3637 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3639 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3640 B_FALSE, compression_type, complevel, ARC_BUFC_DATA);
3642 arc_buf_t *buf = NULL;
3643 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3644 B_TRUE, B_FALSE, B_FALSE, &buf));
3648 * To ensure that the hdr has the correct data in it if we call
3649 * arc_untransform() on this buf before it's been written to disk,
3650 * it's easiest if we just set up sharing between the buf and the hdr.
3652 arc_share_buf(hdr, buf);
3658 arc_alloc_raw_buf(spa_t *spa, const void *tag, uint64_t dsobj,
3659 boolean_t byteorder, const uint8_t *salt, const uint8_t *iv,
3660 const uint8_t *mac, dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3661 enum zio_compress compression_type, uint8_t complevel)
3665 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3666 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3668 ASSERT3U(lsize, >, 0);
3669 ASSERT3U(lsize, >=, psize);
3670 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3671 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3673 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3674 compression_type, complevel, type);
3676 hdr->b_crypt_hdr.b_dsobj = dsobj;
3677 hdr->b_crypt_hdr.b_ot = ot;
3678 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3679 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3680 memcpy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
3681 memcpy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
3682 memcpy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
3685 * This buffer will be considered encrypted even if the ot is not an
3686 * encrypted type. It will become authenticated instead in
3687 * arc_write_ready().
3690 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3691 B_FALSE, B_FALSE, &buf));
3698 l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
3699 boolean_t state_only)
3701 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3702 l2arc_dev_t *dev = l2hdr->b_dev;
3703 uint64_t lsize = HDR_GET_LSIZE(hdr);
3704 uint64_t psize = HDR_GET_PSIZE(hdr);
3705 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3706 arc_buf_contents_t type = hdr->b_type;
3721 /* If the buffer is a prefetch, count it as such. */
3722 if (HDR_PREFETCH(hdr)) {
3723 ARCSTAT_INCR(arcstat_l2_prefetch_asize, asize_s);
3726 * We use the value stored in the L2 header upon initial
3727 * caching in L2ARC. This value will be updated in case
3728 * an MRU/MRU_ghost buffer transitions to MFU but the L2ARC
3729 * metadata (log entry) cannot currently be updated. Having
3730 * the ARC state in the L2 header solves the problem of a
3731 * possibly absent L1 header (apparent in buffers restored
3732 * from persistent L2ARC).
3734 switch (hdr->b_l2hdr.b_arcs_state) {
3735 case ARC_STATE_MRU_GHOST:
3737 ARCSTAT_INCR(arcstat_l2_mru_asize, asize_s);
3739 case ARC_STATE_MFU_GHOST:
3741 ARCSTAT_INCR(arcstat_l2_mfu_asize, asize_s);
3751 ARCSTAT_INCR(arcstat_l2_psize, psize_s);
3752 ARCSTAT_INCR(arcstat_l2_lsize, lsize_s);
3756 ARCSTAT_INCR(arcstat_l2_bufc_data_asize, asize_s);
3758 case ARC_BUFC_METADATA:
3759 ARCSTAT_INCR(arcstat_l2_bufc_metadata_asize, asize_s);
3768 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3770 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3771 l2arc_dev_t *dev = l2hdr->b_dev;
3772 uint64_t psize = HDR_GET_PSIZE(hdr);
3773 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3775 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3776 ASSERT(HDR_HAS_L2HDR(hdr));
3778 list_remove(&dev->l2ad_buflist, hdr);
3780 l2arc_hdr_arcstats_decrement(hdr);
3781 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3783 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3785 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3789 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3791 if (HDR_HAS_L1HDR(hdr)) {
3792 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3793 hdr->b_l1hdr.b_bufcnt > 0);
3794 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3795 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3797 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3798 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3800 if (HDR_HAS_L2HDR(hdr)) {
3801 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3802 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3805 mutex_enter(&dev->l2ad_mtx);
3808 * Even though we checked this conditional above, we
3809 * need to check this again now that we have the
3810 * l2ad_mtx. This is because we could be racing with
3811 * another thread calling l2arc_evict() which might have
3812 * destroyed this header's L2 portion as we were waiting
3813 * to acquire the l2ad_mtx. If that happens, we don't
3814 * want to re-destroy the header's L2 portion.
3816 if (HDR_HAS_L2HDR(hdr)) {
3818 if (!HDR_EMPTY(hdr))
3819 buf_discard_identity(hdr);
3821 arc_hdr_l2hdr_destroy(hdr);
3825 mutex_exit(&dev->l2ad_mtx);
3829 * The header's identify can only be safely discarded once it is no
3830 * longer discoverable. This requires removing it from the hash table
3831 * and the l2arc header list. After this point the hash lock can not
3832 * be used to protect the header.
3834 if (!HDR_EMPTY(hdr))
3835 buf_discard_identity(hdr);
3837 if (HDR_HAS_L1HDR(hdr)) {
3838 arc_cksum_free(hdr);
3840 while (hdr->b_l1hdr.b_buf != NULL)
3841 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3843 if (hdr->b_l1hdr.b_pabd != NULL)
3844 arc_hdr_free_abd(hdr, B_FALSE);
3846 if (HDR_HAS_RABD(hdr))
3847 arc_hdr_free_abd(hdr, B_TRUE);
3850 ASSERT3P(hdr->b_hash_next, ==, NULL);
3851 if (HDR_HAS_L1HDR(hdr)) {
3852 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3853 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3855 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3858 if (!HDR_PROTECTED(hdr)) {
3859 kmem_cache_free(hdr_full_cache, hdr);
3861 kmem_cache_free(hdr_full_crypt_cache, hdr);
3864 kmem_cache_free(hdr_l2only_cache, hdr);
3869 arc_buf_destroy(arc_buf_t *buf, const void *tag)
3871 arc_buf_hdr_t *hdr = buf->b_hdr;
3873 if (hdr->b_l1hdr.b_state == arc_anon) {
3874 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3875 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3876 VERIFY0(remove_reference(hdr, tag));
3880 kmutex_t *hash_lock = HDR_LOCK(hdr);
3881 mutex_enter(hash_lock);
3883 ASSERT3P(hdr, ==, buf->b_hdr);
3884 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3885 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3886 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3887 ASSERT3P(buf->b_data, !=, NULL);
3889 arc_buf_destroy_impl(buf);
3890 (void) remove_reference(hdr, tag);
3891 mutex_exit(hash_lock);
3895 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3896 * state of the header is dependent on its state prior to entering this
3897 * function. The following transitions are possible:
3899 * - arc_mru -> arc_mru_ghost
3900 * - arc_mfu -> arc_mfu_ghost
3901 * - arc_mru_ghost -> arc_l2c_only
3902 * - arc_mru_ghost -> deleted
3903 * - arc_mfu_ghost -> arc_l2c_only
3904 * - arc_mfu_ghost -> deleted
3905 * - arc_uncached -> deleted
3907 * Return total size of evicted data buffers for eviction progress tracking.
3908 * When evicting from ghost states return logical buffer size to make eviction
3909 * progress at the same (or at least comparable) rate as from non-ghost states.
3911 * Return *real_evicted for actual ARC size reduction to wake up threads
3912 * waiting for it. For non-ghost states it includes size of evicted data
3913 * buffers (the headers are not freed there). For ghost states it includes
3914 * only the evicted headers size.
3917 arc_evict_hdr(arc_buf_hdr_t *hdr, uint64_t *real_evicted)
3919 arc_state_t *evicted_state, *state;
3920 int64_t bytes_evicted = 0;
3921 uint_t min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3922 arc_min_prescient_prefetch_ms : arc_min_prefetch_ms;
3924 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3925 ASSERT(HDR_HAS_L1HDR(hdr));
3926 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3927 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3928 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3929 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3932 state = hdr->b_l1hdr.b_state;
3933 if (GHOST_STATE(state)) {
3936 * l2arc_write_buffers() relies on a header's L1 portion
3937 * (i.e. its b_pabd field) during it's write phase.
3938 * Thus, we cannot push a header onto the arc_l2c_only
3939 * state (removing its L1 piece) until the header is
3940 * done being written to the l2arc.
3942 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3943 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3944 return (bytes_evicted);
3947 ARCSTAT_BUMP(arcstat_deleted);
3948 bytes_evicted += HDR_GET_LSIZE(hdr);
3950 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3952 if (HDR_HAS_L2HDR(hdr)) {
3953 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3954 ASSERT(!HDR_HAS_RABD(hdr));
3956 * This buffer is cached on the 2nd Level ARC;
3957 * don't destroy the header.
3959 arc_change_state(arc_l2c_only, hdr);
3961 * dropping from L1+L2 cached to L2-only,
3962 * realloc to remove the L1 header.
3964 (void) arc_hdr_realloc(hdr, hdr_full_cache,
3966 *real_evicted += HDR_FULL_SIZE - HDR_L2ONLY_SIZE;
3968 arc_change_state(arc_anon, hdr);
3969 arc_hdr_destroy(hdr);
3970 *real_evicted += HDR_FULL_SIZE;
3972 return (bytes_evicted);
3975 ASSERT(state == arc_mru || state == arc_mfu || state == arc_uncached);
3976 evicted_state = (state == arc_uncached) ? arc_anon :
3977 ((state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost);
3979 /* prefetch buffers have a minimum lifespan */
3980 if ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3981 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3982 MSEC_TO_TICK(min_lifetime)) {
3983 ARCSTAT_BUMP(arcstat_evict_skip);
3984 return (bytes_evicted);
3987 if (HDR_HAS_L2HDR(hdr)) {
3988 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3990 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3991 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3992 HDR_GET_LSIZE(hdr));
3994 switch (state->arcs_state) {
3997 arcstat_evict_l2_eligible_mru,
3998 HDR_GET_LSIZE(hdr));
4002 arcstat_evict_l2_eligible_mfu,
4003 HDR_GET_LSIZE(hdr));
4009 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
4010 HDR_GET_LSIZE(hdr));
4014 bytes_evicted += arc_hdr_size(hdr);
4015 *real_evicted += arc_hdr_size(hdr);
4018 * If this hdr is being evicted and has a compressed buffer then we
4019 * discard it here before we change states. This ensures that the
4020 * accounting is updated correctly in arc_free_data_impl().
4022 if (hdr->b_l1hdr.b_pabd != NULL)
4023 arc_hdr_free_abd(hdr, B_FALSE);
4025 if (HDR_HAS_RABD(hdr))
4026 arc_hdr_free_abd(hdr, B_TRUE);
4028 arc_change_state(evicted_state, hdr);
4029 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
4030 if (evicted_state == arc_anon) {
4031 arc_hdr_destroy(hdr);
4032 *real_evicted += HDR_FULL_SIZE;
4034 ASSERT(HDR_IN_HASH_TABLE(hdr));
4037 return (bytes_evicted);
4041 arc_set_need_free(void)
4043 ASSERT(MUTEX_HELD(&arc_evict_lock));
4044 int64_t remaining = arc_free_memory() - arc_sys_free / 2;
4045 arc_evict_waiter_t *aw = list_tail(&arc_evict_waiters);
4047 arc_need_free = MAX(-remaining, 0);
4050 MAX(-remaining, (int64_t)(aw->aew_count - arc_evict_count));
4055 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
4056 uint64_t spa, uint64_t bytes)
4058 multilist_sublist_t *mls;
4059 uint64_t bytes_evicted = 0, real_evicted = 0;
4061 kmutex_t *hash_lock;
4062 uint_t evict_count = zfs_arc_evict_batch_limit;
4064 ASSERT3P(marker, !=, NULL);
4066 mls = multilist_sublist_lock(ml, idx);
4068 for (hdr = multilist_sublist_prev(mls, marker); likely(hdr != NULL);
4069 hdr = multilist_sublist_prev(mls, marker)) {
4070 if ((evict_count == 0) || (bytes_evicted >= bytes))
4074 * To keep our iteration location, move the marker
4075 * forward. Since we're not holding hdr's hash lock, we
4076 * must be very careful and not remove 'hdr' from the
4077 * sublist. Otherwise, other consumers might mistake the
4078 * 'hdr' as not being on a sublist when they call the
4079 * multilist_link_active() function (they all rely on
4080 * the hash lock protecting concurrent insertions and
4081 * removals). multilist_sublist_move_forward() was
4082 * specifically implemented to ensure this is the case
4083 * (only 'marker' will be removed and re-inserted).
4085 multilist_sublist_move_forward(mls, marker);
4088 * The only case where the b_spa field should ever be
4089 * zero, is the marker headers inserted by
4090 * arc_evict_state(). It's possible for multiple threads
4091 * to be calling arc_evict_state() concurrently (e.g.
4092 * dsl_pool_close() and zio_inject_fault()), so we must
4093 * skip any markers we see from these other threads.
4095 if (hdr->b_spa == 0)
4098 /* we're only interested in evicting buffers of a certain spa */
4099 if (spa != 0 && hdr->b_spa != spa) {
4100 ARCSTAT_BUMP(arcstat_evict_skip);
4104 hash_lock = HDR_LOCK(hdr);
4107 * We aren't calling this function from any code path
4108 * that would already be holding a hash lock, so we're
4109 * asserting on this assumption to be defensive in case
4110 * this ever changes. Without this check, it would be
4111 * possible to incorrectly increment arcstat_mutex_miss
4112 * below (e.g. if the code changed such that we called
4113 * this function with a hash lock held).
4115 ASSERT(!MUTEX_HELD(hash_lock));
4117 if (mutex_tryenter(hash_lock)) {
4119 uint64_t evicted = arc_evict_hdr(hdr, &revicted);
4120 mutex_exit(hash_lock);
4122 bytes_evicted += evicted;
4123 real_evicted += revicted;
4126 * If evicted is zero, arc_evict_hdr() must have
4127 * decided to skip this header, don't increment
4128 * evict_count in this case.
4134 ARCSTAT_BUMP(arcstat_mutex_miss);
4138 multilist_sublist_unlock(mls);
4141 * Increment the count of evicted bytes, and wake up any threads that
4142 * are waiting for the count to reach this value. Since the list is
4143 * ordered by ascending aew_count, we pop off the beginning of the
4144 * list until we reach the end, or a waiter that's past the current
4145 * "count". Doing this outside the loop reduces the number of times
4146 * we need to acquire the global arc_evict_lock.
4148 * Only wake when there's sufficient free memory in the system
4149 * (specifically, arc_sys_free/2, which by default is a bit more than
4150 * 1/64th of RAM). See the comments in arc_wait_for_eviction().
4152 mutex_enter(&arc_evict_lock);
4153 arc_evict_count += real_evicted;
4155 if (arc_free_memory() > arc_sys_free / 2) {
4156 arc_evict_waiter_t *aw;
4157 while ((aw = list_head(&arc_evict_waiters)) != NULL &&
4158 aw->aew_count <= arc_evict_count) {
4159 list_remove(&arc_evict_waiters, aw);
4160 cv_broadcast(&aw->aew_cv);
4163 arc_set_need_free();
4164 mutex_exit(&arc_evict_lock);
4167 * If the ARC size is reduced from arc_c_max to arc_c_min (especially
4168 * if the average cached block is small), eviction can be on-CPU for
4169 * many seconds. To ensure that other threads that may be bound to
4170 * this CPU are able to make progress, make a voluntary preemption
4173 kpreempt(KPREEMPT_SYNC);
4175 return (bytes_evicted);
4179 * Allocate an array of buffer headers used as placeholders during arc state
4182 static arc_buf_hdr_t **
4183 arc_state_alloc_markers(int count)
4185 arc_buf_hdr_t **markers;
4187 markers = kmem_zalloc(sizeof (*markers) * count, KM_SLEEP);
4188 for (int i = 0; i < count; i++) {
4189 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
4192 * A b_spa of 0 is used to indicate that this header is
4193 * a marker. This fact is used in arc_evict_state_impl().
4195 markers[i]->b_spa = 0;
4202 arc_state_free_markers(arc_buf_hdr_t **markers, int count)
4204 for (int i = 0; i < count; i++)
4205 kmem_cache_free(hdr_full_cache, markers[i]);
4206 kmem_free(markers, sizeof (*markers) * count);
4210 * Evict buffers from the given arc state, until we've removed the
4211 * specified number of bytes. Move the removed buffers to the
4212 * appropriate evict state.
4214 * This function makes a "best effort". It skips over any buffers
4215 * it can't get a hash_lock on, and so, may not catch all candidates.
4216 * It may also return without evicting as much space as requested.
4218 * If bytes is specified using the special value ARC_EVICT_ALL, this
4219 * will evict all available (i.e. unlocked and evictable) buffers from
4220 * the given arc state; which is used by arc_flush().
4223 arc_evict_state(arc_state_t *state, arc_buf_contents_t type, uint64_t spa,
4226 uint64_t total_evicted = 0;
4227 multilist_t *ml = &state->arcs_list[type];
4229 arc_buf_hdr_t **markers;
4231 num_sublists = multilist_get_num_sublists(ml);
4234 * If we've tried to evict from each sublist, made some
4235 * progress, but still have not hit the target number of bytes
4236 * to evict, we want to keep trying. The markers allow us to
4237 * pick up where we left off for each individual sublist, rather
4238 * than starting from the tail each time.
4240 if (zthr_iscurthread(arc_evict_zthr)) {
4241 markers = arc_state_evict_markers;
4242 ASSERT3S(num_sublists, <=, arc_state_evict_marker_count);
4244 markers = arc_state_alloc_markers(num_sublists);
4246 for (int i = 0; i < num_sublists; i++) {
4247 multilist_sublist_t *mls;
4249 mls = multilist_sublist_lock(ml, i);
4250 multilist_sublist_insert_tail(mls, markers[i]);
4251 multilist_sublist_unlock(mls);
4255 * While we haven't hit our target number of bytes to evict, or
4256 * we're evicting all available buffers.
4258 while (total_evicted < bytes) {
4259 int sublist_idx = multilist_get_random_index(ml);
4260 uint64_t scan_evicted = 0;
4263 * Start eviction using a randomly selected sublist,
4264 * this is to try and evenly balance eviction across all
4265 * sublists. Always starting at the same sublist
4266 * (e.g. index 0) would cause evictions to favor certain
4267 * sublists over others.
4269 for (int i = 0; i < num_sublists; i++) {
4270 uint64_t bytes_remaining;
4271 uint64_t bytes_evicted;
4273 if (total_evicted < bytes)
4274 bytes_remaining = bytes - total_evicted;
4278 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4279 markers[sublist_idx], spa, bytes_remaining);
4281 scan_evicted += bytes_evicted;
4282 total_evicted += bytes_evicted;
4284 /* we've reached the end, wrap to the beginning */
4285 if (++sublist_idx >= num_sublists)
4290 * If we didn't evict anything during this scan, we have
4291 * no reason to believe we'll evict more during another
4292 * scan, so break the loop.
4294 if (scan_evicted == 0) {
4295 /* This isn't possible, let's make that obvious */
4296 ASSERT3S(bytes, !=, 0);
4299 * When bytes is ARC_EVICT_ALL, the only way to
4300 * break the loop is when scan_evicted is zero.
4301 * In that case, we actually have evicted enough,
4302 * so we don't want to increment the kstat.
4304 if (bytes != ARC_EVICT_ALL) {
4305 ASSERT3S(total_evicted, <, bytes);
4306 ARCSTAT_BUMP(arcstat_evict_not_enough);
4313 for (int i = 0; i < num_sublists; i++) {
4314 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4315 multilist_sublist_remove(mls, markers[i]);
4316 multilist_sublist_unlock(mls);
4318 if (markers != arc_state_evict_markers)
4319 arc_state_free_markers(markers, num_sublists);
4321 return (total_evicted);
4325 * Flush all "evictable" data of the given type from the arc state
4326 * specified. This will not evict any "active" buffers (i.e. referenced).
4328 * When 'retry' is set to B_FALSE, the function will make a single pass
4329 * over the state and evict any buffers that it can. Since it doesn't
4330 * continually retry the eviction, it might end up leaving some buffers
4331 * in the ARC due to lock misses.
4333 * When 'retry' is set to B_TRUE, the function will continually retry the
4334 * eviction until *all* evictable buffers have been removed from the
4335 * state. As a result, if concurrent insertions into the state are
4336 * allowed (e.g. if the ARC isn't shutting down), this function might
4337 * wind up in an infinite loop, continually trying to evict buffers.
4340 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4343 uint64_t evicted = 0;
4345 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4346 evicted += arc_evict_state(state, type, spa, ARC_EVICT_ALL);
4356 * Evict the specified number of bytes from the state specified. This
4357 * function prevents us from trying to evict more from a state's list
4358 * than is "evictable", and to skip evicting altogether when passed a
4359 * negative value for "bytes". In contrast, arc_evict_state() will
4360 * evict everything it can, when passed a negative value for "bytes".
4363 arc_evict_impl(arc_state_t *state, arc_buf_contents_t type, int64_t bytes)
4367 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4368 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4370 return (arc_evict_state(state, type, 0, delta));
4377 * Adjust specified fraction, taking into account initial ghost state(s) size,
4378 * ghost hit bytes towards increasing the fraction, ghost hit bytes towards
4379 * decreasing it, plus a balance factor, controlling the decrease rate, used
4380 * to balance metadata vs data.
4383 arc_evict_adj(uint64_t frac, uint64_t total, uint64_t up, uint64_t down,
4386 if (total < 8 || up + down == 0)
4390 * We should not have more ghost hits than ghost size, but they
4391 * may get close. Restrict maximum adjustment in that case.
4393 if (up + down >= total / 4) {
4394 uint64_t scale = (up + down) / (total / 8);
4399 /* Get maximal dynamic range by choosing optimal shifts. */
4400 int s = highbit64(total);
4401 s = MIN(64 - s, 32);
4403 uint64_t ofrac = (1ULL << 32) - frac;
4405 if (frac >= 4 * ofrac)
4406 up /= frac / (2 * ofrac + 1);
4407 up = (up << s) / (total >> (32 - s));
4408 if (ofrac >= 4 * frac)
4409 down /= ofrac / (2 * frac + 1);
4410 down = (down << s) / (total >> (32 - s));
4411 down = down * 100 / balance;
4413 return (frac + up - down);
4417 * Evict buffers from the cache, such that arcstat_size is capped by arc_c.
4422 uint64_t asize, bytes, total_evicted = 0;
4423 int64_t e, mrud, mrum, mfud, mfum, w;
4424 static uint64_t ogrd, ogrm, ogfd, ogfm;
4425 static uint64_t gsrd, gsrm, gsfd, gsfm;
4426 uint64_t ngrd, ngrm, ngfd, ngfm;
4428 /* Get current size of ARC states we can evict from. */
4429 mrud = zfs_refcount_count(&arc_mru->arcs_size[ARC_BUFC_DATA]) +
4430 zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_DATA]);
4431 mrum = zfs_refcount_count(&arc_mru->arcs_size[ARC_BUFC_METADATA]) +
4432 zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_METADATA]);
4433 mfud = zfs_refcount_count(&arc_mfu->arcs_size[ARC_BUFC_DATA]);
4434 mfum = zfs_refcount_count(&arc_mfu->arcs_size[ARC_BUFC_METADATA]);
4435 uint64_t d = mrud + mfud;
4436 uint64_t m = mrum + mfum;
4439 /* Get ARC ghost hits since last eviction. */
4440 ngrd = wmsum_value(&arc_mru_ghost->arcs_hits[ARC_BUFC_DATA]);
4441 uint64_t grd = ngrd - ogrd;
4443 ngrm = wmsum_value(&arc_mru_ghost->arcs_hits[ARC_BUFC_METADATA]);
4444 uint64_t grm = ngrm - ogrm;
4446 ngfd = wmsum_value(&arc_mfu_ghost->arcs_hits[ARC_BUFC_DATA]);
4447 uint64_t gfd = ngfd - ogfd;
4449 ngfm = wmsum_value(&arc_mfu_ghost->arcs_hits[ARC_BUFC_METADATA]);
4450 uint64_t gfm = ngfm - ogfm;
4453 /* Adjust ARC states balance based on ghost hits. */
4454 arc_meta = arc_evict_adj(arc_meta, gsrd + gsrm + gsfd + gsfm,
4455 grm + gfm, grd + gfd, zfs_arc_meta_balance);
4456 arc_pd = arc_evict_adj(arc_pd, gsrd + gsfd, grd, gfd, 100);
4457 arc_pm = arc_evict_adj(arc_pm, gsrm + gsfm, grm, gfm, 100);
4459 asize = aggsum_value(&arc_sums.arcstat_size);
4460 int64_t wt = t - (asize - arc_c);
4463 * Try to reduce pinned dnodes if more than 3/4 of wanted metadata
4464 * target is not evictable or if they go over arc_dnode_limit.
4467 int64_t dn = wmsum_value(&arc_sums.arcstat_dnode_size);
4468 w = wt * (int64_t)(arc_meta >> 16) >> 16;
4469 if (zfs_refcount_count(&arc_mru->arcs_size[ARC_BUFC_METADATA]) +
4470 zfs_refcount_count(&arc_mfu->arcs_size[ARC_BUFC_METADATA]) -
4471 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA]) -
4472 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]) >
4474 prune = dn / sizeof (dnode_t) *
4475 zfs_arc_dnode_reduce_percent / 100;
4476 } else if (dn > arc_dnode_limit) {
4477 prune = (dn - arc_dnode_limit) / sizeof (dnode_t) *
4478 zfs_arc_dnode_reduce_percent / 100;
4481 arc_prune_async(prune);
4483 /* Evict MRU metadata. */
4484 w = wt * (int64_t)(arc_meta * arc_pm >> 48) >> 16;
4485 e = MIN((int64_t)(asize - arc_c), (int64_t)(mrum - w));
4486 bytes = arc_evict_impl(arc_mru, ARC_BUFC_METADATA, e);
4487 total_evicted += bytes;
4491 /* Evict MFU metadata. */
4492 w = wt * (int64_t)(arc_meta >> 16) >> 16;
4493 e = MIN((int64_t)(asize - arc_c), (int64_t)(m - w));
4494 bytes = arc_evict_impl(arc_mfu, ARC_BUFC_METADATA, e);
4495 total_evicted += bytes;
4499 /* Evict MRU data. */
4500 wt -= m - total_evicted;
4501 w = wt * (int64_t)(arc_pd >> 16) >> 16;
4502 e = MIN((int64_t)(asize - arc_c), (int64_t)(mrud - w));
4503 bytes = arc_evict_impl(arc_mru, ARC_BUFC_DATA, e);
4504 total_evicted += bytes;
4508 /* Evict MFU data. */
4510 bytes = arc_evict_impl(arc_mfu, ARC_BUFC_DATA, e);
4512 total_evicted += bytes;
4517 * Size of each state's ghost list represents how much that state
4518 * may grow by shrinking the other states. Would it need to shrink
4519 * other states to zero (that is unlikely), its ghost size would be
4520 * equal to sum of other three state sizes. But excessive ghost
4521 * size may result in false ghost hits (too far back), that may
4522 * never result in real cache hits if several states are competing.
4523 * So choose some arbitraty point of 1/2 of other state sizes.
4525 gsrd = (mrum + mfud + mfum) / 2;
4526 e = zfs_refcount_count(&arc_mru_ghost->arcs_size[ARC_BUFC_DATA]) -
4528 (void) arc_evict_impl(arc_mru_ghost, ARC_BUFC_DATA, e);
4530 gsrm = (mrud + mfud + mfum) / 2;
4531 e = zfs_refcount_count(&arc_mru_ghost->arcs_size[ARC_BUFC_METADATA]) -
4533 (void) arc_evict_impl(arc_mru_ghost, ARC_BUFC_METADATA, e);
4535 gsfd = (mrud + mrum + mfum) / 2;
4536 e = zfs_refcount_count(&arc_mfu_ghost->arcs_size[ARC_BUFC_DATA]) -
4538 (void) arc_evict_impl(arc_mfu_ghost, ARC_BUFC_DATA, e);
4540 gsfm = (mrud + mrum + mfud) / 2;
4541 e = zfs_refcount_count(&arc_mfu_ghost->arcs_size[ARC_BUFC_METADATA]) -
4543 (void) arc_evict_impl(arc_mfu_ghost, ARC_BUFC_METADATA, e);
4545 return (total_evicted);
4549 arc_flush(spa_t *spa, boolean_t retry)
4554 * If retry is B_TRUE, a spa must not be specified since we have
4555 * no good way to determine if all of a spa's buffers have been
4556 * evicted from an arc state.
4558 ASSERT(!retry || spa == NULL);
4561 guid = spa_load_guid(spa);
4563 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4564 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4566 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4567 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4569 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4570 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4572 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4573 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4575 (void) arc_flush_state(arc_uncached, guid, ARC_BUFC_DATA, retry);
4576 (void) arc_flush_state(arc_uncached, guid, ARC_BUFC_METADATA, retry);
4580 arc_reduce_target_size(int64_t to_free)
4588 * All callers want the ARC to actually evict (at least) this much
4589 * memory. Therefore we reduce from the lower of the current size and
4590 * the target size. This way, even if arc_c is much higher than
4591 * arc_size (as can be the case after many calls to arc_freed(), we will
4592 * immediately have arc_c < arc_size and therefore the arc_evict_zthr
4595 uint64_t asize = aggsum_value(&arc_sums.arcstat_size);
4597 to_free += c - asize;
4598 arc_c = MAX((int64_t)c - to_free, (int64_t)arc_c_min);
4600 /* See comment in arc_evict_cb_check() on why lock+flag */
4601 mutex_enter(&arc_evict_lock);
4602 arc_evict_needed = B_TRUE;
4603 mutex_exit(&arc_evict_lock);
4604 zthr_wakeup(arc_evict_zthr);
4608 * Determine if the system is under memory pressure and is asking
4609 * to reclaim memory. A return value of B_TRUE indicates that the system
4610 * is under memory pressure and that the arc should adjust accordingly.
4613 arc_reclaim_needed(void)
4615 return (arc_available_memory() < 0);
4619 arc_kmem_reap_soon(void)
4622 kmem_cache_t *prev_cache = NULL;
4623 kmem_cache_t *prev_data_cache = NULL;
4628 * Reclaim unused memory from all kmem caches.
4634 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4636 /* reach upper limit of cache size on 32-bit */
4637 if (zio_buf_cache[i] == NULL)
4640 if (zio_buf_cache[i] != prev_cache) {
4641 prev_cache = zio_buf_cache[i];
4642 kmem_cache_reap_now(zio_buf_cache[i]);
4644 if (zio_data_buf_cache[i] != prev_data_cache) {
4645 prev_data_cache = zio_data_buf_cache[i];
4646 kmem_cache_reap_now(zio_data_buf_cache[i]);
4649 kmem_cache_reap_now(buf_cache);
4650 kmem_cache_reap_now(hdr_full_cache);
4651 kmem_cache_reap_now(hdr_l2only_cache);
4652 kmem_cache_reap_now(zfs_btree_leaf_cache);
4653 abd_cache_reap_now();
4657 arc_evict_cb_check(void *arg, zthr_t *zthr)
4659 (void) arg, (void) zthr;
4663 * This is necessary in order to keep the kstat information
4664 * up to date for tools that display kstat data such as the
4665 * mdb ::arc dcmd and the Linux crash utility. These tools
4666 * typically do not call kstat's update function, but simply
4667 * dump out stats from the most recent update. Without
4668 * this call, these commands may show stale stats for the
4669 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4670 * with this call, the data might be out of date if the
4671 * evict thread hasn't been woken recently; but that should
4672 * suffice. The arc_state_t structures can be queried
4673 * directly if more accurate information is needed.
4675 if (arc_ksp != NULL)
4676 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4680 * We have to rely on arc_wait_for_eviction() to tell us when to
4681 * evict, rather than checking if we are overflowing here, so that we
4682 * are sure to not leave arc_wait_for_eviction() waiting on aew_cv.
4683 * If we have become "not overflowing" since arc_wait_for_eviction()
4684 * checked, we need to wake it up. We could broadcast the CV here,
4685 * but arc_wait_for_eviction() may have not yet gone to sleep. We
4686 * would need to use a mutex to ensure that this function doesn't
4687 * broadcast until arc_wait_for_eviction() has gone to sleep (e.g.
4688 * the arc_evict_lock). However, the lock ordering of such a lock
4689 * would necessarily be incorrect with respect to the zthr_lock,
4690 * which is held before this function is called, and is held by
4691 * arc_wait_for_eviction() when it calls zthr_wakeup().
4693 if (arc_evict_needed)
4697 * If we have buffers in uncached state, evict them periodically.
4699 return ((zfs_refcount_count(&arc_uncached->arcs_esize[ARC_BUFC_DATA]) +
4700 zfs_refcount_count(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]) &&
4701 ddi_get_lbolt() - arc_last_uncached_flush >
4702 MSEC_TO_TICK(arc_min_prefetch_ms / 2)));
4706 * Keep arc_size under arc_c by running arc_evict which evicts data
4710 arc_evict_cb(void *arg, zthr_t *zthr)
4712 (void) arg, (void) zthr;
4714 uint64_t evicted = 0;
4715 fstrans_cookie_t cookie = spl_fstrans_mark();
4717 /* Always try to evict from uncached state. */
4718 arc_last_uncached_flush = ddi_get_lbolt();
4719 evicted += arc_flush_state(arc_uncached, 0, ARC_BUFC_DATA, B_FALSE);
4720 evicted += arc_flush_state(arc_uncached, 0, ARC_BUFC_METADATA, B_FALSE);
4722 /* Evict from other states only if told to. */
4723 if (arc_evict_needed)
4724 evicted += arc_evict();
4727 * If evicted is zero, we couldn't evict anything
4728 * via arc_evict(). This could be due to hash lock
4729 * collisions, but more likely due to the majority of
4730 * arc buffers being unevictable. Therefore, even if
4731 * arc_size is above arc_c, another pass is unlikely to
4732 * be helpful and could potentially cause us to enter an
4733 * infinite loop. Additionally, zthr_iscancelled() is
4734 * checked here so that if the arc is shutting down, the
4735 * broadcast will wake any remaining arc evict waiters.
4737 mutex_enter(&arc_evict_lock);
4738 arc_evict_needed = !zthr_iscancelled(arc_evict_zthr) &&
4739 evicted > 0 && aggsum_compare(&arc_sums.arcstat_size, arc_c) > 0;
4740 if (!arc_evict_needed) {
4742 * We're either no longer overflowing, or we
4743 * can't evict anything more, so we should wake
4744 * arc_get_data_impl() sooner.
4746 arc_evict_waiter_t *aw;
4747 while ((aw = list_remove_head(&arc_evict_waiters)) != NULL) {
4748 cv_broadcast(&aw->aew_cv);
4750 arc_set_need_free();
4752 mutex_exit(&arc_evict_lock);
4753 spl_fstrans_unmark(cookie);
4757 arc_reap_cb_check(void *arg, zthr_t *zthr)
4759 (void) arg, (void) zthr;
4761 int64_t free_memory = arc_available_memory();
4762 static int reap_cb_check_counter = 0;
4765 * If a kmem reap is already active, don't schedule more. We must
4766 * check for this because kmem_cache_reap_soon() won't actually
4767 * block on the cache being reaped (this is to prevent callers from
4768 * becoming implicitly blocked by a system-wide kmem reap -- which,
4769 * on a system with many, many full magazines, can take minutes).
4771 if (!kmem_cache_reap_active() && free_memory < 0) {
4773 arc_no_grow = B_TRUE;
4776 * Wait at least zfs_grow_retry (default 5) seconds
4777 * before considering growing.
4779 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4781 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4782 arc_no_grow = B_TRUE;
4783 } else if (gethrtime() >= arc_growtime) {
4784 arc_no_grow = B_FALSE;
4788 * Called unconditionally every 60 seconds to reclaim unused
4789 * zstd compression and decompression context. This is done
4790 * here to avoid the need for an independent thread.
4792 if (!((reap_cb_check_counter++) % 60))
4793 zfs_zstd_cache_reap_now();
4799 * Keep enough free memory in the system by reaping the ARC's kmem
4800 * caches. To cause more slabs to be reapable, we may reduce the
4801 * target size of the cache (arc_c), causing the arc_evict_cb()
4802 * to free more buffers.
4805 arc_reap_cb(void *arg, zthr_t *zthr)
4807 (void) arg, (void) zthr;
4809 int64_t free_memory;
4810 fstrans_cookie_t cookie = spl_fstrans_mark();
4813 * Kick off asynchronous kmem_reap()'s of all our caches.
4815 arc_kmem_reap_soon();
4818 * Wait at least arc_kmem_cache_reap_retry_ms between
4819 * arc_kmem_reap_soon() calls. Without this check it is possible to
4820 * end up in a situation where we spend lots of time reaping
4821 * caches, while we're near arc_c_min. Waiting here also gives the
4822 * subsequent free memory check a chance of finding that the
4823 * asynchronous reap has already freed enough memory, and we don't
4824 * need to call arc_reduce_target_size().
4826 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
4829 * Reduce the target size as needed to maintain the amount of free
4830 * memory in the system at a fraction of the arc_size (1/128th by
4831 * default). If oversubscribed (free_memory < 0) then reduce the
4832 * target arc_size by the deficit amount plus the fractional
4833 * amount. If free memory is positive but less than the fractional
4834 * amount, reduce by what is needed to hit the fractional amount.
4836 free_memory = arc_available_memory();
4838 int64_t can_free = arc_c - arc_c_min;
4840 int64_t to_free = (can_free >> arc_shrink_shift) - free_memory;
4842 arc_reduce_target_size(to_free);
4844 spl_fstrans_unmark(cookie);
4849 * Determine the amount of memory eligible for eviction contained in the
4850 * ARC. All clean data reported by the ghost lists can always be safely
4851 * evicted. Due to arc_c_min, the same does not hold for all clean data
4852 * contained by the regular mru and mfu lists.
4854 * In the case of the regular mru and mfu lists, we need to report as
4855 * much clean data as possible, such that evicting that same reported
4856 * data will not bring arc_size below arc_c_min. Thus, in certain
4857 * circumstances, the total amount of clean data in the mru and mfu
4858 * lists might not actually be evictable.
4860 * The following two distinct cases are accounted for:
4862 * 1. The sum of the amount of dirty data contained by both the mru and
4863 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4864 * is greater than or equal to arc_c_min.
4865 * (i.e. amount of dirty data >= arc_c_min)
4867 * This is the easy case; all clean data contained by the mru and mfu
4868 * lists is evictable. Evicting all clean data can only drop arc_size
4869 * to the amount of dirty data, which is greater than arc_c_min.
4871 * 2. The sum of the amount of dirty data contained by both the mru and
4872 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4873 * is less than arc_c_min.
4874 * (i.e. arc_c_min > amount of dirty data)
4876 * 2.1. arc_size is greater than or equal arc_c_min.
4877 * (i.e. arc_size >= arc_c_min > amount of dirty data)
4879 * In this case, not all clean data from the regular mru and mfu
4880 * lists is actually evictable; we must leave enough clean data
4881 * to keep arc_size above arc_c_min. Thus, the maximum amount of
4882 * evictable data from the two lists combined, is exactly the
4883 * difference between arc_size and arc_c_min.
4885 * 2.2. arc_size is less than arc_c_min
4886 * (i.e. arc_c_min > arc_size > amount of dirty data)
4888 * In this case, none of the data contained in the mru and mfu
4889 * lists is evictable, even if it's clean. Since arc_size is
4890 * already below arc_c_min, evicting any more would only
4891 * increase this negative difference.
4894 #endif /* _KERNEL */
4897 * Adapt arc info given the number of bytes we are trying to add and
4898 * the state that we are coming from. This function is only called
4899 * when we are adding new content to the cache.
4902 arc_adapt(uint64_t bytes)
4905 * Wake reap thread if we do not have any available memory
4907 if (arc_reclaim_needed()) {
4908 zthr_wakeup(arc_reap_zthr);
4915 if (arc_c >= arc_c_max)
4919 * If we're within (2 * maxblocksize) bytes of the target
4920 * cache size, increment the target cache size
4922 if (aggsum_upper_bound(&arc_sums.arcstat_size) +
4923 2 * SPA_MAXBLOCKSIZE >= arc_c) {
4924 uint64_t dc = MAX(bytes, SPA_OLD_MAXBLOCKSIZE);
4925 if (atomic_add_64_nv(&arc_c, dc) > arc_c_max)
4931 * Check if arc_size has grown past our upper threshold, determined by
4932 * zfs_arc_overflow_shift.
4934 static arc_ovf_level_t
4935 arc_is_overflowing(boolean_t use_reserve)
4937 /* Always allow at least one block of overflow */
4938 int64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4939 arc_c >> zfs_arc_overflow_shift);
4942 * We just compare the lower bound here for performance reasons. Our
4943 * primary goals are to make sure that the arc never grows without
4944 * bound, and that it can reach its maximum size. This check
4945 * accomplishes both goals. The maximum amount we could run over by is
4946 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4947 * in the ARC. In practice, that's in the tens of MB, which is low
4948 * enough to be safe.
4950 int64_t over = aggsum_lower_bound(&arc_sums.arcstat_size) -
4951 arc_c - overflow / 2;
4954 return (over < 0 ? ARC_OVF_NONE :
4955 over < overflow ? ARC_OVF_SOME : ARC_OVF_SEVERE);
4959 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, const void *tag,
4962 arc_buf_contents_t type = arc_buf_type(hdr);
4964 arc_get_data_impl(hdr, size, tag, alloc_flags);
4965 if (alloc_flags & ARC_HDR_ALLOC_LINEAR)
4966 return (abd_alloc_linear(size, type == ARC_BUFC_METADATA));
4968 return (abd_alloc(size, type == ARC_BUFC_METADATA));
4972 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, const void *tag)
4974 arc_buf_contents_t type = arc_buf_type(hdr);
4976 arc_get_data_impl(hdr, size, tag, 0);
4977 if (type == ARC_BUFC_METADATA) {
4978 return (zio_buf_alloc(size));
4980 ASSERT(type == ARC_BUFC_DATA);
4981 return (zio_data_buf_alloc(size));
4986 * Wait for the specified amount of data (in bytes) to be evicted from the
4987 * ARC, and for there to be sufficient free memory in the system. Waiting for
4988 * eviction ensures that the memory used by the ARC decreases. Waiting for
4989 * free memory ensures that the system won't run out of free pages, regardless
4990 * of ARC behavior and settings. See arc_lowmem_init().
4993 arc_wait_for_eviction(uint64_t amount, boolean_t use_reserve)
4995 switch (arc_is_overflowing(use_reserve)) {
5000 * This is a bit racy without taking arc_evict_lock, but the
5001 * worst that can happen is we either call zthr_wakeup() extra
5002 * time due to race with other thread here, or the set flag
5003 * get cleared by arc_evict_cb(), which is unlikely due to
5004 * big hysteresis, but also not important since at this level
5005 * of overflow the eviction is purely advisory. Same time
5006 * taking the global lock here every time without waiting for
5007 * the actual eviction creates a significant lock contention.
5009 if (!arc_evict_needed) {
5010 arc_evict_needed = B_TRUE;
5011 zthr_wakeup(arc_evict_zthr);
5014 case ARC_OVF_SEVERE:
5017 arc_evict_waiter_t aw;
5018 list_link_init(&aw.aew_node);
5019 cv_init(&aw.aew_cv, NULL, CV_DEFAULT, NULL);
5021 uint64_t last_count = 0;
5022 mutex_enter(&arc_evict_lock);
5023 if (!list_is_empty(&arc_evict_waiters)) {
5024 arc_evict_waiter_t *last =
5025 list_tail(&arc_evict_waiters);
5026 last_count = last->aew_count;
5027 } else if (!arc_evict_needed) {
5028 arc_evict_needed = B_TRUE;
5029 zthr_wakeup(arc_evict_zthr);
5032 * Note, the last waiter's count may be less than
5033 * arc_evict_count if we are low on memory in which
5034 * case arc_evict_state_impl() may have deferred
5035 * wakeups (but still incremented arc_evict_count).
5037 aw.aew_count = MAX(last_count, arc_evict_count) + amount;
5039 list_insert_tail(&arc_evict_waiters, &aw);
5041 arc_set_need_free();
5043 DTRACE_PROBE3(arc__wait__for__eviction,
5045 uint64_t, arc_evict_count,
5046 uint64_t, aw.aew_count);
5049 * We will be woken up either when arc_evict_count reaches
5050 * aew_count, or when the ARC is no longer overflowing and
5051 * eviction completes.
5052 * In case of "false" wakeup, we will still be on the list.
5055 cv_wait(&aw.aew_cv, &arc_evict_lock);
5056 } while (list_link_active(&aw.aew_node));
5057 mutex_exit(&arc_evict_lock);
5059 cv_destroy(&aw.aew_cv);
5065 * Allocate a block and return it to the caller. If we are hitting the
5066 * hard limit for the cache size, we must sleep, waiting for the eviction
5067 * thread to catch up. If we're past the target size but below the hard
5068 * limit, we'll only signal the reclaim thread and continue on.
5071 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, const void *tag,
5077 * If arc_size is currently overflowing, we must be adding data
5078 * faster than we are evicting. To ensure we don't compound the
5079 * problem by adding more data and forcing arc_size to grow even
5080 * further past it's target size, we wait for the eviction thread to
5081 * make some progress. We also wait for there to be sufficient free
5082 * memory in the system, as measured by arc_free_memory().
5084 * Specifically, we wait for zfs_arc_eviction_pct percent of the
5085 * requested size to be evicted. This should be more than 100%, to
5086 * ensure that that progress is also made towards getting arc_size
5087 * under arc_c. See the comment above zfs_arc_eviction_pct.
5089 arc_wait_for_eviction(size * zfs_arc_eviction_pct / 100,
5090 alloc_flags & ARC_HDR_USE_RESERVE);
5092 arc_buf_contents_t type = arc_buf_type(hdr);
5093 if (type == ARC_BUFC_METADATA) {
5094 arc_space_consume(size, ARC_SPACE_META);
5096 arc_space_consume(size, ARC_SPACE_DATA);
5100 * Update the state size. Note that ghost states have a
5101 * "ghost size" and so don't need to be updated.
5103 arc_state_t *state = hdr->b_l1hdr.b_state;
5104 if (!GHOST_STATE(state)) {
5106 (void) zfs_refcount_add_many(&state->arcs_size[type], size,
5110 * If this is reached via arc_read, the link is
5111 * protected by the hash lock. If reached via
5112 * arc_buf_alloc, the header should not be accessed by
5113 * any other thread. And, if reached via arc_read_done,
5114 * the hash lock will protect it if it's found in the
5115 * hash table; otherwise no other thread should be
5116 * trying to [add|remove]_reference it.
5118 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5119 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5120 (void) zfs_refcount_add_many(&state->arcs_esize[type],
5127 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size,
5130 arc_free_data_impl(hdr, size, tag);
5135 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, const void *tag)
5137 arc_buf_contents_t type = arc_buf_type(hdr);
5139 arc_free_data_impl(hdr, size, tag);
5140 if (type == ARC_BUFC_METADATA) {
5141 zio_buf_free(buf, size);
5143 ASSERT(type == ARC_BUFC_DATA);
5144 zio_data_buf_free(buf, size);
5149 * Free the arc data buffer.
5152 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, const void *tag)
5154 arc_state_t *state = hdr->b_l1hdr.b_state;
5155 arc_buf_contents_t type = arc_buf_type(hdr);
5157 /* protected by hash lock, if in the hash table */
5158 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5159 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5160 ASSERT(state != arc_anon && state != arc_l2c_only);
5162 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
5165 (void) zfs_refcount_remove_many(&state->arcs_size[type], size, tag);
5167 VERIFY3U(hdr->b_type, ==, type);
5168 if (type == ARC_BUFC_METADATA) {
5169 arc_space_return(size, ARC_SPACE_META);
5171 ASSERT(type == ARC_BUFC_DATA);
5172 arc_space_return(size, ARC_SPACE_DATA);
5177 * This routine is called whenever a buffer is accessed.
5180 arc_access(arc_buf_hdr_t *hdr, arc_flags_t arc_flags, boolean_t hit)
5182 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
5183 ASSERT(HDR_HAS_L1HDR(hdr));
5186 * Update buffer prefetch status.
5188 boolean_t was_prefetch = HDR_PREFETCH(hdr);
5189 boolean_t now_prefetch = arc_flags & ARC_FLAG_PREFETCH;
5190 if (was_prefetch != now_prefetch) {
5192 ARCSTAT_CONDSTAT(hit, demand_hit, demand_iohit,
5193 HDR_PRESCIENT_PREFETCH(hdr), prescient, predictive,
5196 if (HDR_HAS_L2HDR(hdr))
5197 l2arc_hdr_arcstats_decrement_state(hdr);
5199 arc_hdr_clear_flags(hdr,
5200 ARC_FLAG_PREFETCH | ARC_FLAG_PRESCIENT_PREFETCH);
5202 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5204 if (HDR_HAS_L2HDR(hdr))
5205 l2arc_hdr_arcstats_increment_state(hdr);
5208 if (arc_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5209 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5210 ARCSTAT_BUMP(arcstat_prescient_prefetch);
5212 ARCSTAT_BUMP(arcstat_predictive_prefetch);
5215 if (arc_flags & ARC_FLAG_L2CACHE)
5216 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5218 clock_t now = ddi_get_lbolt();
5219 if (hdr->b_l1hdr.b_state == arc_anon) {
5220 arc_state_t *new_state;
5222 * This buffer is not in the cache, and does not appear in
5223 * our "ghost" lists. Add it to the MRU or uncached state.
5225 ASSERT0(hdr->b_l1hdr.b_arc_access);
5226 hdr->b_l1hdr.b_arc_access = now;
5227 if (HDR_UNCACHED(hdr)) {
5228 new_state = arc_uncached;
5229 DTRACE_PROBE1(new_state__uncached, arc_buf_hdr_t *,
5232 new_state = arc_mru;
5233 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5235 arc_change_state(new_state, hdr);
5236 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5238 * This buffer has been accessed once recently and either
5239 * its read is still in progress or it is in the cache.
5241 if (HDR_IO_IN_PROGRESS(hdr)) {
5242 hdr->b_l1hdr.b_arc_access = now;
5245 hdr->b_l1hdr.b_mru_hits++;
5246 ARCSTAT_BUMP(arcstat_mru_hits);
5249 * If the previous access was a prefetch, then it already
5250 * handled possible promotion, so nothing more to do for now.
5253 hdr->b_l1hdr.b_arc_access = now;
5258 * If more than ARC_MINTIME have passed from the previous
5259 * hit, promote the buffer to the MFU state.
5261 if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access +
5263 hdr->b_l1hdr.b_arc_access = now;
5264 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5265 arc_change_state(arc_mfu, hdr);
5267 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5268 arc_state_t *new_state;
5270 * This buffer has been accessed once recently, but was
5271 * evicted from the cache. Would we have bigger MRU, it
5272 * would be an MRU hit, so handle it the same way, except
5273 * we don't need to check the previous access time.
5275 hdr->b_l1hdr.b_mru_ghost_hits++;
5276 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5277 hdr->b_l1hdr.b_arc_access = now;
5278 wmsum_add(&arc_mru_ghost->arcs_hits[arc_buf_type(hdr)],
5281 new_state = arc_mru;
5282 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5284 new_state = arc_mfu;
5285 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5287 arc_change_state(new_state, hdr);
5288 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5290 * This buffer has been accessed more than once and either
5291 * still in the cache or being restored from one of ghosts.
5293 if (!HDR_IO_IN_PROGRESS(hdr)) {
5294 hdr->b_l1hdr.b_mfu_hits++;
5295 ARCSTAT_BUMP(arcstat_mfu_hits);
5297 hdr->b_l1hdr.b_arc_access = now;
5298 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5300 * This buffer has been accessed more than once recently, but
5301 * has been evicted from the cache. Would we have bigger MFU
5302 * it would stay in cache, so move it back to MFU state.
5304 hdr->b_l1hdr.b_mfu_ghost_hits++;
5305 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5306 hdr->b_l1hdr.b_arc_access = now;
5307 wmsum_add(&arc_mfu_ghost->arcs_hits[arc_buf_type(hdr)],
5309 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5310 arc_change_state(arc_mfu, hdr);
5311 } else if (hdr->b_l1hdr.b_state == arc_uncached) {
5313 * This buffer is uncacheable, but we got a hit. Probably
5314 * a demand read after prefetch. Nothing more to do here.
5316 if (!HDR_IO_IN_PROGRESS(hdr))
5317 ARCSTAT_BUMP(arcstat_uncached_hits);
5318 hdr->b_l1hdr.b_arc_access = now;
5319 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5321 * This buffer is on the 2nd Level ARC and was not accessed
5322 * for a long time, so treat it as new and put into MRU.
5324 hdr->b_l1hdr.b_arc_access = now;
5325 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5326 arc_change_state(arc_mru, hdr);
5328 cmn_err(CE_PANIC, "invalid arc state 0x%p",
5329 hdr->b_l1hdr.b_state);
5334 * This routine is called by dbuf_hold() to update the arc_access() state
5335 * which otherwise would be skipped for entries in the dbuf cache.
5338 arc_buf_access(arc_buf_t *buf)
5340 arc_buf_hdr_t *hdr = buf->b_hdr;
5343 * Avoid taking the hash_lock when possible as an optimization.
5344 * The header must be checked again under the hash_lock in order
5345 * to handle the case where it is concurrently being released.
5347 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr))
5350 kmutex_t *hash_lock = HDR_LOCK(hdr);
5351 mutex_enter(hash_lock);
5353 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5354 mutex_exit(hash_lock);
5355 ARCSTAT_BUMP(arcstat_access_skip);
5359 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5360 hdr->b_l1hdr.b_state == arc_mfu ||
5361 hdr->b_l1hdr.b_state == arc_uncached);
5363 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5364 arc_access(hdr, 0, B_TRUE);
5365 mutex_exit(hash_lock);
5367 ARCSTAT_BUMP(arcstat_hits);
5368 ARCSTAT_CONDSTAT(B_TRUE /* demand */, demand, prefetch,
5369 !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5372 /* a generic arc_read_done_func_t which you can use */
5374 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5375 arc_buf_t *buf, void *arg)
5377 (void) zio, (void) zb, (void) bp;
5382 memcpy(arg, buf->b_data, arc_buf_size(buf));
5383 arc_buf_destroy(buf, arg);
5386 /* a generic arc_read_done_func_t */
5388 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5389 arc_buf_t *buf, void *arg)
5391 (void) zb, (void) bp;
5392 arc_buf_t **bufp = arg;
5395 ASSERT(zio == NULL || zio->io_error != 0);
5398 ASSERT(zio == NULL || zio->io_error == 0);
5400 ASSERT(buf->b_data != NULL);
5405 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5407 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5408 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5409 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5411 if (HDR_COMPRESSION_ENABLED(hdr)) {
5412 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5413 BP_GET_COMPRESS(bp));
5415 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5416 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5417 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5422 arc_read_done(zio_t *zio)
5424 blkptr_t *bp = zio->io_bp;
5425 arc_buf_hdr_t *hdr = zio->io_private;
5426 kmutex_t *hash_lock = NULL;
5427 arc_callback_t *callback_list;
5428 arc_callback_t *acb;
5431 * The hdr was inserted into hash-table and removed from lists
5432 * prior to starting I/O. We should find this header, since
5433 * it's in the hash table, and it should be legit since it's
5434 * not possible to evict it during the I/O. The only possible
5435 * reason for it not to be found is if we were freed during the
5438 if (HDR_IN_HASH_TABLE(hdr)) {
5439 arc_buf_hdr_t *found;
5441 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5442 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5443 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5444 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5445 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5447 found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock);
5449 ASSERT((found == hdr &&
5450 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5451 (found == hdr && HDR_L2_READING(hdr)));
5452 ASSERT3P(hash_lock, !=, NULL);
5455 if (BP_IS_PROTECTED(bp)) {
5456 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5457 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5458 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5459 hdr->b_crypt_hdr.b_iv);
5461 if (zio->io_error == 0) {
5462 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5465 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5466 sizeof (zil_chain_t));
5467 zio_crypt_decode_mac_zil(tmpbuf,
5468 hdr->b_crypt_hdr.b_mac);
5469 abd_return_buf(zio->io_abd, tmpbuf,
5470 sizeof (zil_chain_t));
5472 zio_crypt_decode_mac_bp(bp,
5473 hdr->b_crypt_hdr.b_mac);
5478 if (zio->io_error == 0) {
5479 /* byteswap if necessary */
5480 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5481 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5482 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5484 hdr->b_l1hdr.b_byteswap =
5485 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5488 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5490 if (!HDR_L2_READING(hdr)) {
5491 hdr->b_complevel = zio->io_prop.zp_complevel;
5495 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5496 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5497 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5499 callback_list = hdr->b_l1hdr.b_acb;
5500 ASSERT3P(callback_list, !=, NULL);
5501 hdr->b_l1hdr.b_acb = NULL;
5504 * If a read request has a callback (i.e. acb_done is not NULL), then we
5505 * make a buf containing the data according to the parameters which were
5506 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5507 * aren't needlessly decompressing the data multiple times.
5509 int callback_cnt = 0;
5510 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5512 /* We need the last one to call below in original order. */
5513 callback_list = acb;
5515 if (!acb->acb_done || acb->acb_nobuf)
5520 if (zio->io_error != 0)
5523 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5524 &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
5525 acb->acb_compressed, acb->acb_noauth, B_TRUE,
5529 * Assert non-speculative zios didn't fail because an
5530 * encryption key wasn't loaded
5532 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
5536 * If we failed to decrypt, report an error now (as the zio
5537 * layer would have done if it had done the transforms).
5539 if (error == ECKSUM) {
5540 ASSERT(BP_IS_PROTECTED(bp));
5541 error = SET_ERROR(EIO);
5542 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5543 spa_log_error(zio->io_spa, &acb->acb_zb,
5544 &zio->io_bp->blk_birth);
5545 (void) zfs_ereport_post(
5546 FM_EREPORT_ZFS_AUTHENTICATION,
5547 zio->io_spa, NULL, &acb->acb_zb, zio, 0);
5553 * Decompression or decryption failed. Set
5554 * io_error so that when we call acb_done
5555 * (below), we will indicate that the read
5556 * failed. Note that in the unusual case
5557 * where one callback is compressed and another
5558 * uncompressed, we will mark all of them
5559 * as failed, even though the uncompressed
5560 * one can't actually fail. In this case,
5561 * the hdr will not be anonymous, because
5562 * if there are multiple callbacks, it's
5563 * because multiple threads found the same
5564 * arc buf in the hash table.
5566 zio->io_error = error;
5571 * If there are multiple callbacks, we must have the hash lock,
5572 * because the only way for multiple threads to find this hdr is
5573 * in the hash table. This ensures that if there are multiple
5574 * callbacks, the hdr is not anonymous. If it were anonymous,
5575 * we couldn't use arc_buf_destroy() in the error case below.
5577 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5579 if (zio->io_error == 0) {
5580 arc_hdr_verify(hdr, zio->io_bp);
5582 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5583 if (hdr->b_l1hdr.b_state != arc_anon)
5584 arc_change_state(arc_anon, hdr);
5585 if (HDR_IN_HASH_TABLE(hdr))
5586 buf_hash_remove(hdr);
5590 * Broadcast before we drop the hash_lock to avoid the possibility
5591 * that the hdr (and hence the cv) might be freed before we get to
5592 * the cv_broadcast().
5594 cv_broadcast(&hdr->b_l1hdr.b_cv);
5596 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5597 (void) remove_reference(hdr, hdr);
5599 if (hash_lock != NULL)
5600 mutex_exit(hash_lock);
5602 /* execute each callback and free its structure */
5603 while ((acb = callback_list) != NULL) {
5604 if (acb->acb_done != NULL) {
5605 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5607 * If arc_buf_alloc_impl() fails during
5608 * decompression, the buf will still be
5609 * allocated, and needs to be freed here.
5611 arc_buf_destroy(acb->acb_buf,
5613 acb->acb_buf = NULL;
5615 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5616 acb->acb_buf, acb->acb_private);
5619 if (acb->acb_zio_dummy != NULL) {
5620 acb->acb_zio_dummy->io_error = zio->io_error;
5621 zio_nowait(acb->acb_zio_dummy);
5624 callback_list = acb->acb_prev;
5625 if (acb->acb_wait) {
5626 mutex_enter(&acb->acb_wait_lock);
5627 acb->acb_wait_error = zio->io_error;
5628 acb->acb_wait = B_FALSE;
5629 cv_signal(&acb->acb_wait_cv);
5630 mutex_exit(&acb->acb_wait_lock);
5631 /* acb will be freed by the waiting thread. */
5633 kmem_free(acb, sizeof (arc_callback_t));
5639 * "Read" the block at the specified DVA (in bp) via the
5640 * cache. If the block is found in the cache, invoke the provided
5641 * callback immediately and return. Note that the `zio' parameter
5642 * in the callback will be NULL in this case, since no IO was
5643 * required. If the block is not in the cache pass the read request
5644 * on to the spa with a substitute callback function, so that the
5645 * requested block will be added to the cache.
5647 * If a read request arrives for a block that has a read in-progress,
5648 * either wait for the in-progress read to complete (and return the
5649 * results); or, if this is a read with a "done" func, add a record
5650 * to the read to invoke the "done" func when the read completes,
5651 * and return; or just return.
5653 * arc_read_done() will invoke all the requested "done" functions
5654 * for readers of this block.
5657 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
5658 arc_read_done_func_t *done, void *private, zio_priority_t priority,
5659 int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5661 arc_buf_hdr_t *hdr = NULL;
5662 kmutex_t *hash_lock = NULL;
5664 uint64_t guid = spa_load_guid(spa);
5665 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
5666 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
5667 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5668 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
5669 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5670 boolean_t embedded_bp = !!BP_IS_EMBEDDED(bp);
5671 boolean_t no_buf = *arc_flags & ARC_FLAG_NO_BUF;
5672 arc_buf_t *buf = NULL;
5675 ASSERT(!embedded_bp ||
5676 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5677 ASSERT(!BP_IS_HOLE(bp));
5678 ASSERT(!BP_IS_REDACTED(bp));
5681 * Normally SPL_FSTRANS will already be set since kernel threads which
5682 * expect to call the DMU interfaces will set it when created. System
5683 * calls are similarly handled by setting/cleaning the bit in the
5684 * registered callback (module/os/.../zfs/zpl_*).
5686 * External consumers such as Lustre which call the exported DMU
5687 * interfaces may not have set SPL_FSTRANS. To avoid a deadlock
5688 * on the hash_lock always set and clear the bit.
5690 fstrans_cookie_t cookie = spl_fstrans_mark();
5693 * Verify the block pointer contents are reasonable. This should
5694 * always be the case since the blkptr is protected by a checksum.
5695 * However, if there is damage it's desirable to detect this early
5696 * and treat it as a checksum error. This allows an alternate blkptr
5697 * to be tried when one is available (e.g. ditto blocks).
5699 if (!zfs_blkptr_verify(spa, bp, (zio_flags & ZIO_FLAG_CONFIG_WRITER) ?
5700 BLK_CONFIG_HELD : BLK_CONFIG_NEEDED, BLK_VERIFY_LOG)) {
5701 rc = SET_ERROR(ECKSUM);
5707 * Embedded BP's have no DVA and require no I/O to "read".
5708 * Create an anonymous arc buf to back it.
5710 hdr = buf_hash_find(guid, bp, &hash_lock);
5714 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5715 * we maintain encrypted data separately from compressed / uncompressed
5716 * data. If the user is requesting raw encrypted data and we don't have
5717 * that in the header we will read from disk to guarantee that we can
5718 * get it even if the encryption keys aren't loaded.
5720 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
5721 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
5722 boolean_t is_data = !HDR_ISTYPE_METADATA(hdr);
5724 if (HDR_IO_IN_PROGRESS(hdr)) {
5725 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
5726 mutex_exit(hash_lock);
5727 ARCSTAT_BUMP(arcstat_cached_only_in_progress);
5728 rc = SET_ERROR(ENOENT);
5732 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5733 ASSERT3P(head_zio, !=, NULL);
5734 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5735 priority == ZIO_PRIORITY_SYNC_READ) {
5737 * This is a sync read that needs to wait for
5738 * an in-flight async read. Request that the
5739 * zio have its priority upgraded.
5741 zio_change_priority(head_zio, priority);
5742 DTRACE_PROBE1(arc__async__upgrade__sync,
5743 arc_buf_hdr_t *, hdr);
5744 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5747 DTRACE_PROBE1(arc__iohit, arc_buf_hdr_t *, hdr);
5748 arc_access(hdr, *arc_flags, B_FALSE);
5751 * If there are multiple threads reading the same block
5752 * and that block is not yet in the ARC, then only one
5753 * thread will do the physical I/O and all other
5754 * threads will wait until that I/O completes.
5755 * Synchronous reads use the acb_wait_cv whereas nowait
5756 * reads register a callback. Both are signalled/called
5759 * Errors of the physical I/O may need to be propagated.
5760 * Synchronous read errors are returned here from
5761 * arc_read_done via acb_wait_error. Nowait reads
5762 * attach the acb_zio_dummy zio to pio and
5763 * arc_read_done propagates the physical I/O's io_error
5764 * to acb_zio_dummy, and thereby to pio.
5766 arc_callback_t *acb = NULL;
5767 if (done || pio || *arc_flags & ARC_FLAG_WAIT) {
5768 acb = kmem_zalloc(sizeof (arc_callback_t),
5770 acb->acb_done = done;
5771 acb->acb_private = private;
5772 acb->acb_compressed = compressed_read;
5773 acb->acb_encrypted = encrypted_read;
5774 acb->acb_noauth = noauth_read;
5775 acb->acb_nobuf = no_buf;
5776 if (*arc_flags & ARC_FLAG_WAIT) {
5777 acb->acb_wait = B_TRUE;
5778 mutex_init(&acb->acb_wait_lock, NULL,
5779 MUTEX_DEFAULT, NULL);
5780 cv_init(&acb->acb_wait_cv, NULL,
5785 acb->acb_zio_dummy = zio_null(pio,
5786 spa, NULL, NULL, NULL, zio_flags);
5788 acb->acb_zio_head = head_zio;
5789 acb->acb_next = hdr->b_l1hdr.b_acb;
5790 if (hdr->b_l1hdr.b_acb)
5791 hdr->b_l1hdr.b_acb->acb_prev = acb;
5792 hdr->b_l1hdr.b_acb = acb;
5794 mutex_exit(hash_lock);
5796 ARCSTAT_BUMP(arcstat_iohits);
5797 ARCSTAT_CONDSTAT(!(*arc_flags & ARC_FLAG_PREFETCH),
5798 demand, prefetch, is_data, data, metadata, iohits);
5800 if (*arc_flags & ARC_FLAG_WAIT) {
5801 mutex_enter(&acb->acb_wait_lock);
5802 while (acb->acb_wait) {
5803 cv_wait(&acb->acb_wait_cv,
5804 &acb->acb_wait_lock);
5806 rc = acb->acb_wait_error;
5807 mutex_exit(&acb->acb_wait_lock);
5808 mutex_destroy(&acb->acb_wait_lock);
5809 cv_destroy(&acb->acb_wait_cv);
5810 kmem_free(acb, sizeof (arc_callback_t));
5815 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5816 hdr->b_l1hdr.b_state == arc_mfu ||
5817 hdr->b_l1hdr.b_state == arc_uncached);
5819 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5820 arc_access(hdr, *arc_flags, B_TRUE);
5822 if (done && !no_buf) {
5823 ASSERT(!embedded_bp || !BP_IS_HOLE(bp));
5825 /* Get a buf with the desired data in it. */
5826 rc = arc_buf_alloc_impl(hdr, spa, zb, private,
5827 encrypted_read, compressed_read, noauth_read,
5831 * Convert authentication and decryption errors
5832 * to EIO (and generate an ereport if needed)
5833 * before leaving the ARC.
5835 rc = SET_ERROR(EIO);
5836 if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5837 spa_log_error(spa, zb, &hdr->b_birth);
5838 (void) zfs_ereport_post(
5839 FM_EREPORT_ZFS_AUTHENTICATION,
5840 spa, NULL, zb, NULL, 0);
5844 arc_buf_destroy_impl(buf);
5846 (void) remove_reference(hdr, private);
5849 /* assert any errors weren't due to unloaded keys */
5850 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5853 mutex_exit(hash_lock);
5854 ARCSTAT_BUMP(arcstat_hits);
5855 ARCSTAT_CONDSTAT(!(*arc_flags & ARC_FLAG_PREFETCH),
5856 demand, prefetch, is_data, data, metadata, hits);
5857 *arc_flags |= ARC_FLAG_CACHED;
5860 uint64_t lsize = BP_GET_LSIZE(bp);
5861 uint64_t psize = BP_GET_PSIZE(bp);
5862 arc_callback_t *acb;
5865 boolean_t devw = B_FALSE;
5868 int alloc_flags = encrypted_read ? ARC_HDR_ALLOC_RDATA : 0;
5869 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5871 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
5872 if (hash_lock != NULL)
5873 mutex_exit(hash_lock);
5874 rc = SET_ERROR(ENOENT);
5880 * This block is not in the cache or it has
5883 arc_buf_hdr_t *exists = NULL;
5884 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5885 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), 0, type);
5888 hdr->b_dva = *BP_IDENTITY(bp);
5889 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5890 exists = buf_hash_insert(hdr, &hash_lock);
5892 if (exists != NULL) {
5893 /* somebody beat us to the hash insert */
5894 mutex_exit(hash_lock);
5895 buf_discard_identity(hdr);
5896 arc_hdr_destroy(hdr);
5897 goto top; /* restart the IO request */
5901 * This block is in the ghost cache or encrypted data
5902 * was requested and we didn't have it. If it was
5903 * L2-only (and thus didn't have an L1 hdr),
5904 * we realloc the header to add an L1 hdr.
5906 if (!HDR_HAS_L1HDR(hdr)) {
5907 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5911 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
5912 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5913 ASSERT(!HDR_HAS_RABD(hdr));
5914 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5915 ASSERT0(zfs_refcount_count(
5916 &hdr->b_l1hdr.b_refcnt));
5917 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5919 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5921 } else if (HDR_IO_IN_PROGRESS(hdr)) {
5923 * If this header already had an IO in progress
5924 * and we are performing another IO to fetch
5925 * encrypted data we must wait until the first
5926 * IO completes so as not to confuse
5927 * arc_read_done(). This should be very rare
5928 * and so the performance impact shouldn't
5931 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5932 mutex_exit(hash_lock);
5936 if (*arc_flags & ARC_FLAG_UNCACHED) {
5937 arc_hdr_set_flags(hdr, ARC_FLAG_UNCACHED);
5938 if (!encrypted_read)
5939 alloc_flags |= ARC_HDR_ALLOC_LINEAR;
5943 * Take additional reference for IO_IN_PROGRESS. It stops
5944 * arc_access() from putting this header without any buffers
5945 * and so other references but obviously nonevictable onto
5946 * the evictable list of MRU or MFU state.
5948 add_reference(hdr, hdr);
5950 arc_access(hdr, *arc_flags, B_FALSE);
5951 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5952 arc_hdr_alloc_abd(hdr, alloc_flags);
5953 if (encrypted_read) {
5954 ASSERT(HDR_HAS_RABD(hdr));
5955 size = HDR_GET_PSIZE(hdr);
5956 hdr_abd = hdr->b_crypt_hdr.b_rabd;
5957 zio_flags |= ZIO_FLAG_RAW;
5959 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5960 size = arc_hdr_size(hdr);
5961 hdr_abd = hdr->b_l1hdr.b_pabd;
5963 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
5964 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
5968 * For authenticated bp's, we do not ask the ZIO layer
5969 * to authenticate them since this will cause the entire
5970 * IO to fail if the key isn't loaded. Instead, we
5971 * defer authentication until arc_buf_fill(), which will
5972 * verify the data when the key is available.
5974 if (BP_IS_AUTHENTICATED(bp))
5975 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
5978 if (BP_IS_AUTHENTICATED(bp))
5979 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
5980 if (BP_GET_LEVEL(bp) > 0)
5981 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5982 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5984 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5985 acb->acb_done = done;
5986 acb->acb_private = private;
5987 acb->acb_compressed = compressed_read;
5988 acb->acb_encrypted = encrypted_read;
5989 acb->acb_noauth = noauth_read;
5992 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5993 hdr->b_l1hdr.b_acb = acb;
5995 if (HDR_HAS_L2HDR(hdr) &&
5996 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5997 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5998 addr = hdr->b_l2hdr.b_daddr;
6000 * Lock out L2ARC device removal.
6002 if (vdev_is_dead(vd) ||
6003 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
6008 * We count both async reads and scrub IOs as asynchronous so
6009 * that both can be upgraded in the event of a cache hit while
6010 * the read IO is still in-flight.
6012 if (priority == ZIO_PRIORITY_ASYNC_READ ||
6013 priority == ZIO_PRIORITY_SCRUB)
6014 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6016 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6019 * At this point, we have a level 1 cache miss or a blkptr
6020 * with embedded data. Try again in L2ARC if possible.
6022 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
6025 * Skip ARC stat bump for block pointers with embedded
6026 * data. The data are read from the blkptr itself via
6027 * decode_embedded_bp_compressed().
6030 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr,
6031 blkptr_t *, bp, uint64_t, lsize,
6032 zbookmark_phys_t *, zb);
6033 ARCSTAT_BUMP(arcstat_misses);
6034 ARCSTAT_CONDSTAT(!(*arc_flags & ARC_FLAG_PREFETCH),
6035 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data,
6037 zfs_racct_read(size, 1);
6040 /* Check if the spa even has l2 configured */
6041 const boolean_t spa_has_l2 = l2arc_ndev != 0 &&
6042 spa->spa_l2cache.sav_count > 0;
6044 if (vd != NULL && spa_has_l2 && !(l2arc_norw && devw)) {
6046 * Read from the L2ARC if the following are true:
6047 * 1. The L2ARC vdev was previously cached.
6048 * 2. This buffer still has L2ARC metadata.
6049 * 3. This buffer isn't currently writing to the L2ARC.
6050 * 4. The L2ARC entry wasn't evicted, which may
6051 * also have invalidated the vdev.
6052 * 5. This isn't prefetch or l2arc_noprefetch is 0.
6054 if (HDR_HAS_L2HDR(hdr) &&
6055 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
6056 !(l2arc_noprefetch &&
6057 (*arc_flags & ARC_FLAG_PREFETCH))) {
6058 l2arc_read_callback_t *cb;
6062 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
6063 ARCSTAT_BUMP(arcstat_l2_hits);
6064 hdr->b_l2hdr.b_hits++;
6066 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
6068 cb->l2rcb_hdr = hdr;
6071 cb->l2rcb_flags = zio_flags;
6074 * When Compressed ARC is disabled, but the
6075 * L2ARC block is compressed, arc_hdr_size()
6076 * will have returned LSIZE rather than PSIZE.
6078 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
6079 !HDR_COMPRESSION_ENABLED(hdr) &&
6080 HDR_GET_PSIZE(hdr) != 0) {
6081 size = HDR_GET_PSIZE(hdr);
6084 asize = vdev_psize_to_asize(vd, size);
6085 if (asize != size) {
6086 abd = abd_alloc_for_io(asize,
6087 HDR_ISTYPE_METADATA(hdr));
6088 cb->l2rcb_abd = abd;
6093 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6094 addr + asize <= vd->vdev_psize -
6095 VDEV_LABEL_END_SIZE);
6098 * l2arc read. The SCL_L2ARC lock will be
6099 * released by l2arc_read_done().
6100 * Issue a null zio if the underlying buffer
6101 * was squashed to zero size by compression.
6103 ASSERT3U(arc_hdr_get_compress(hdr), !=,
6104 ZIO_COMPRESS_EMPTY);
6105 rzio = zio_read_phys(pio, vd, addr,
6108 l2arc_read_done, cb, priority,
6109 zio_flags | ZIO_FLAG_CANFAIL |
6110 ZIO_FLAG_DONT_PROPAGATE |
6111 ZIO_FLAG_DONT_RETRY, B_FALSE);
6112 acb->acb_zio_head = rzio;
6114 if (hash_lock != NULL)
6115 mutex_exit(hash_lock);
6117 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6119 ARCSTAT_INCR(arcstat_l2_read_bytes,
6120 HDR_GET_PSIZE(hdr));
6122 if (*arc_flags & ARC_FLAG_NOWAIT) {
6127 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6128 if (zio_wait(rzio) == 0)
6131 /* l2arc read error; goto zio_read() */
6132 if (hash_lock != NULL)
6133 mutex_enter(hash_lock);
6135 DTRACE_PROBE1(l2arc__miss,
6136 arc_buf_hdr_t *, hdr);
6137 ARCSTAT_BUMP(arcstat_l2_misses);
6138 if (HDR_L2_WRITING(hdr))
6139 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6140 spa_config_exit(spa, SCL_L2ARC, vd);
6144 spa_config_exit(spa, SCL_L2ARC, vd);
6147 * Only a spa with l2 should contribute to l2
6148 * miss stats. (Including the case of having a
6149 * faulted cache device - that's also a miss.)
6153 * Skip ARC stat bump for block pointers with
6154 * embedded data. The data are read from the
6156 * decode_embedded_bp_compressed().
6159 DTRACE_PROBE1(l2arc__miss,
6160 arc_buf_hdr_t *, hdr);
6161 ARCSTAT_BUMP(arcstat_l2_misses);
6166 rzio = zio_read(pio, spa, bp, hdr_abd, size,
6167 arc_read_done, hdr, priority, zio_flags, zb);
6168 acb->acb_zio_head = rzio;
6170 if (hash_lock != NULL)
6171 mutex_exit(hash_lock);
6173 if (*arc_flags & ARC_FLAG_WAIT) {
6174 rc = zio_wait(rzio);
6178 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6183 /* embedded bps don't actually go to disk */
6185 spa_read_history_add(spa, zb, *arc_flags);
6186 spl_fstrans_unmark(cookie);
6191 done(NULL, zb, bp, buf, private);
6192 if (pio && rc != 0) {
6193 zio_t *zio = zio_null(pio, spa, NULL, NULL, NULL, zio_flags);
6201 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6205 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6207 p->p_private = private;
6208 list_link_init(&p->p_node);
6209 zfs_refcount_create(&p->p_refcnt);
6211 mutex_enter(&arc_prune_mtx);
6212 zfs_refcount_add(&p->p_refcnt, &arc_prune_list);
6213 list_insert_head(&arc_prune_list, p);
6214 mutex_exit(&arc_prune_mtx);
6220 arc_remove_prune_callback(arc_prune_t *p)
6222 boolean_t wait = B_FALSE;
6223 mutex_enter(&arc_prune_mtx);
6224 list_remove(&arc_prune_list, p);
6225 if (zfs_refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6227 mutex_exit(&arc_prune_mtx);
6229 /* wait for arc_prune_task to finish */
6231 taskq_wait_outstanding(arc_prune_taskq, 0);
6232 ASSERT0(zfs_refcount_count(&p->p_refcnt));
6233 zfs_refcount_destroy(&p->p_refcnt);
6234 kmem_free(p, sizeof (*p));
6238 * Notify the arc that a block was freed, and thus will never be used again.
6241 arc_freed(spa_t *spa, const blkptr_t *bp)
6244 kmutex_t *hash_lock;
6245 uint64_t guid = spa_load_guid(spa);
6247 ASSERT(!BP_IS_EMBEDDED(bp));
6249 hdr = buf_hash_find(guid, bp, &hash_lock);
6254 * We might be trying to free a block that is still doing I/O
6255 * (i.e. prefetch) or has some other reference (i.e. a dedup-ed,
6256 * dmu_sync-ed block). A block may also have a reference if it is
6257 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6258 * have written the new block to its final resting place on disk but
6259 * without the dedup flag set. This would have left the hdr in the MRU
6260 * state and discoverable. When the txg finally syncs it detects that
6261 * the block was overridden in open context and issues an override I/O.
6262 * Since this is a dedup block, the override I/O will determine if the
6263 * block is already in the DDT. If so, then it will replace the io_bp
6264 * with the bp from the DDT and allow the I/O to finish. When the I/O
6265 * reaches the done callback, dbuf_write_override_done, it will
6266 * check to see if the io_bp and io_bp_override are identical.
6267 * If they are not, then it indicates that the bp was replaced with
6268 * the bp in the DDT and the override bp is freed. This allows
6269 * us to arrive here with a reference on a block that is being
6270 * freed. So if we have an I/O in progress, or a reference to
6271 * this hdr, then we don't destroy the hdr.
6273 if (!HDR_HAS_L1HDR(hdr) ||
6274 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6275 arc_change_state(arc_anon, hdr);
6276 arc_hdr_destroy(hdr);
6277 mutex_exit(hash_lock);
6279 mutex_exit(hash_lock);
6285 * Release this buffer from the cache, making it an anonymous buffer. This
6286 * must be done after a read and prior to modifying the buffer contents.
6287 * If the buffer has more than one reference, we must make
6288 * a new hdr for the buffer.
6291 arc_release(arc_buf_t *buf, const void *tag)
6293 arc_buf_hdr_t *hdr = buf->b_hdr;
6296 * It would be nice to assert that if its DMU metadata (level >
6297 * 0 || it's the dnode file), then it must be syncing context.
6298 * But we don't know that information at this level.
6301 ASSERT(HDR_HAS_L1HDR(hdr));
6304 * We don't grab the hash lock prior to this check, because if
6305 * the buffer's header is in the arc_anon state, it won't be
6306 * linked into the hash table.
6308 if (hdr->b_l1hdr.b_state == arc_anon) {
6309 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6310 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6311 ASSERT(!HDR_HAS_L2HDR(hdr));
6313 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6314 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6315 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6317 hdr->b_l1hdr.b_arc_access = 0;
6320 * If the buf is being overridden then it may already
6321 * have a hdr that is not empty.
6323 buf_discard_identity(hdr);
6329 kmutex_t *hash_lock = HDR_LOCK(hdr);
6330 mutex_enter(hash_lock);
6333 * This assignment is only valid as long as the hash_lock is
6334 * held, we must be careful not to reference state or the
6335 * b_state field after dropping the lock.
6337 arc_state_t *state = hdr->b_l1hdr.b_state;
6338 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6339 ASSERT3P(state, !=, arc_anon);
6341 /* this buffer is not on any list */
6342 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6344 if (HDR_HAS_L2HDR(hdr)) {
6345 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6348 * We have to recheck this conditional again now that
6349 * we're holding the l2ad_mtx to prevent a race with
6350 * another thread which might be concurrently calling
6351 * l2arc_evict(). In that case, l2arc_evict() might have
6352 * destroyed the header's L2 portion as we were waiting
6353 * to acquire the l2ad_mtx.
6355 if (HDR_HAS_L2HDR(hdr))
6356 arc_hdr_l2hdr_destroy(hdr);
6358 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6362 * Do we have more than one buf?
6364 if (hdr->b_l1hdr.b_bufcnt > 1) {
6365 arc_buf_hdr_t *nhdr;
6366 uint64_t spa = hdr->b_spa;
6367 uint64_t psize = HDR_GET_PSIZE(hdr);
6368 uint64_t lsize = HDR_GET_LSIZE(hdr);
6369 boolean_t protected = HDR_PROTECTED(hdr);
6370 enum zio_compress compress = arc_hdr_get_compress(hdr);
6371 arc_buf_contents_t type = arc_buf_type(hdr);
6372 VERIFY3U(hdr->b_type, ==, type);
6374 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6375 VERIFY3S(remove_reference(hdr, tag), >, 0);
6377 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6378 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6379 ASSERT(ARC_BUF_LAST(buf));
6383 * Pull the data off of this hdr and attach it to
6384 * a new anonymous hdr. Also find the last buffer
6385 * in the hdr's buffer list.
6387 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6388 ASSERT3P(lastbuf, !=, NULL);
6391 * If the current arc_buf_t and the hdr are sharing their data
6392 * buffer, then we must stop sharing that block.
6394 if (arc_buf_is_shared(buf)) {
6395 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6396 VERIFY(!arc_buf_is_shared(lastbuf));
6399 * First, sever the block sharing relationship between
6400 * buf and the arc_buf_hdr_t.
6402 arc_unshare_buf(hdr, buf);
6405 * Now we need to recreate the hdr's b_pabd. Since we
6406 * have lastbuf handy, we try to share with it, but if
6407 * we can't then we allocate a new b_pabd and copy the
6408 * data from buf into it.
6410 if (arc_can_share(hdr, lastbuf)) {
6411 arc_share_buf(hdr, lastbuf);
6413 arc_hdr_alloc_abd(hdr, 0);
6414 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6415 buf->b_data, psize);
6417 VERIFY3P(lastbuf->b_data, !=, NULL);
6418 } else if (HDR_SHARED_DATA(hdr)) {
6420 * Uncompressed shared buffers are always at the end
6421 * of the list. Compressed buffers don't have the
6422 * same requirements. This makes it hard to
6423 * simply assert that the lastbuf is shared so
6424 * we rely on the hdr's compression flags to determine
6425 * if we have a compressed, shared buffer.
6427 ASSERT(arc_buf_is_shared(lastbuf) ||
6428 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6429 ASSERT(!ARC_BUF_SHARED(buf));
6432 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6433 ASSERT3P(state, !=, arc_l2c_only);
6435 (void) zfs_refcount_remove_many(&state->arcs_size[type],
6436 arc_buf_size(buf), buf);
6438 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6439 ASSERT3P(state, !=, arc_l2c_only);
6440 (void) zfs_refcount_remove_many(
6441 &state->arcs_esize[type],
6442 arc_buf_size(buf), buf);
6445 hdr->b_l1hdr.b_bufcnt -= 1;
6446 if (ARC_BUF_ENCRYPTED(buf))
6447 hdr->b_crypt_hdr.b_ebufcnt -= 1;
6449 arc_cksum_verify(buf);
6450 arc_buf_unwatch(buf);
6452 /* if this is the last uncompressed buf free the checksum */
6453 if (!arc_hdr_has_uncompressed_buf(hdr))
6454 arc_cksum_free(hdr);
6456 mutex_exit(hash_lock);
6458 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6459 compress, hdr->b_complevel, type);
6460 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6461 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6462 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6463 VERIFY3U(nhdr->b_type, ==, type);
6464 ASSERT(!HDR_SHARED_DATA(nhdr));
6466 nhdr->b_l1hdr.b_buf = buf;
6467 nhdr->b_l1hdr.b_bufcnt = 1;
6468 if (ARC_BUF_ENCRYPTED(buf))
6469 nhdr->b_crypt_hdr.b_ebufcnt = 1;
6470 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6473 (void) zfs_refcount_add_many(&arc_anon->arcs_size[type],
6474 arc_buf_size(buf), buf);
6476 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6477 /* protected by hash lock, or hdr is on arc_anon */
6478 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6479 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6480 hdr->b_l1hdr.b_mru_hits = 0;
6481 hdr->b_l1hdr.b_mru_ghost_hits = 0;
6482 hdr->b_l1hdr.b_mfu_hits = 0;
6483 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
6484 arc_change_state(arc_anon, hdr);
6485 hdr->b_l1hdr.b_arc_access = 0;
6487 mutex_exit(hash_lock);
6488 buf_discard_identity(hdr);
6494 arc_released(arc_buf_t *buf)
6496 return (buf->b_data != NULL &&
6497 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6502 arc_referenced(arc_buf_t *buf)
6504 return (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6509 arc_write_ready(zio_t *zio)
6511 arc_write_callback_t *callback = zio->io_private;
6512 arc_buf_t *buf = callback->awcb_buf;
6513 arc_buf_hdr_t *hdr = buf->b_hdr;
6514 blkptr_t *bp = zio->io_bp;
6515 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6516 fstrans_cookie_t cookie = spl_fstrans_mark();
6518 ASSERT(HDR_HAS_L1HDR(hdr));
6519 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6520 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6523 * If we're reexecuting this zio because the pool suspended, then
6524 * cleanup any state that was previously set the first time the
6525 * callback was invoked.
6527 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6528 arc_cksum_free(hdr);
6529 arc_buf_unwatch(buf);
6530 if (hdr->b_l1hdr.b_pabd != NULL) {
6531 if (arc_buf_is_shared(buf)) {
6532 arc_unshare_buf(hdr, buf);
6534 arc_hdr_free_abd(hdr, B_FALSE);
6538 if (HDR_HAS_RABD(hdr))
6539 arc_hdr_free_abd(hdr, B_TRUE);
6541 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6542 ASSERT(!HDR_HAS_RABD(hdr));
6543 ASSERT(!HDR_SHARED_DATA(hdr));
6544 ASSERT(!arc_buf_is_shared(buf));
6546 callback->awcb_ready(zio, buf, callback->awcb_private);
6548 if (HDR_IO_IN_PROGRESS(hdr)) {
6549 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6551 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6552 add_reference(hdr, hdr); /* For IO_IN_PROGRESS. */
6555 if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr))
6556 hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp));
6558 if (BP_IS_PROTECTED(bp)) {
6559 /* ZIL blocks are written through zio_rewrite */
6560 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6561 ASSERT(HDR_PROTECTED(hdr));
6563 if (BP_SHOULD_BYTESWAP(bp)) {
6564 if (BP_GET_LEVEL(bp) > 0) {
6565 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6567 hdr->b_l1hdr.b_byteswap =
6568 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6571 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6574 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6575 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6576 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6577 hdr->b_crypt_hdr.b_iv);
6578 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6582 * If this block was written for raw encryption but the zio layer
6583 * ended up only authenticating it, adjust the buffer flags now.
6585 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6586 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6587 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6588 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6589 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6590 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6591 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6592 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6595 /* this must be done after the buffer flags are adjusted */
6596 arc_cksum_compute(buf);
6598 enum zio_compress compress;
6599 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6600 compress = ZIO_COMPRESS_OFF;
6602 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6603 compress = BP_GET_COMPRESS(bp);
6605 HDR_SET_PSIZE(hdr, psize);
6606 arc_hdr_set_compress(hdr, compress);
6607 hdr->b_complevel = zio->io_prop.zp_complevel;
6609 if (zio->io_error != 0 || psize == 0)
6613 * Fill the hdr with data. If the buffer is encrypted we have no choice
6614 * but to copy the data into b_radb. If the hdr is compressed, the data
6615 * we want is available from the zio, otherwise we can take it from
6618 * We might be able to share the buf's data with the hdr here. However,
6619 * doing so would cause the ARC to be full of linear ABDs if we write a
6620 * lot of shareable data. As a compromise, we check whether scattered
6621 * ABDs are allowed, and assume that if they are then the user wants
6622 * the ARC to be primarily filled with them regardless of the data being
6623 * written. Therefore, if they're allowed then we allocate one and copy
6624 * the data into it; otherwise, we share the data directly if we can.
6626 if (ARC_BUF_ENCRYPTED(buf)) {
6627 ASSERT3U(psize, >, 0);
6628 ASSERT(ARC_BUF_COMPRESSED(buf));
6629 arc_hdr_alloc_abd(hdr, ARC_HDR_ALLOC_RDATA |
6630 ARC_HDR_USE_RESERVE);
6631 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6632 } else if (!(HDR_UNCACHED(hdr) ||
6633 abd_size_alloc_linear(arc_buf_size(buf))) ||
6634 !arc_can_share(hdr, buf)) {
6636 * Ideally, we would always copy the io_abd into b_pabd, but the
6637 * user may have disabled compressed ARC, thus we must check the
6638 * hdr's compression setting rather than the io_bp's.
6640 if (BP_IS_ENCRYPTED(bp)) {
6641 ASSERT3U(psize, >, 0);
6642 arc_hdr_alloc_abd(hdr, ARC_HDR_ALLOC_RDATA |
6643 ARC_HDR_USE_RESERVE);
6644 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6645 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
6646 !ARC_BUF_COMPRESSED(buf)) {
6647 ASSERT3U(psize, >, 0);
6648 arc_hdr_alloc_abd(hdr, ARC_HDR_USE_RESERVE);
6649 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6651 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6652 arc_hdr_alloc_abd(hdr, ARC_HDR_USE_RESERVE);
6653 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6657 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6658 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6659 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6661 arc_share_buf(hdr, buf);
6665 arc_hdr_verify(hdr, bp);
6666 spl_fstrans_unmark(cookie);
6670 arc_write_children_ready(zio_t *zio)
6672 arc_write_callback_t *callback = zio->io_private;
6673 arc_buf_t *buf = callback->awcb_buf;
6675 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6679 arc_write_done(zio_t *zio)
6681 arc_write_callback_t *callback = zio->io_private;
6682 arc_buf_t *buf = callback->awcb_buf;
6683 arc_buf_hdr_t *hdr = buf->b_hdr;
6685 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6687 if (zio->io_error == 0) {
6688 arc_hdr_verify(hdr, zio->io_bp);
6690 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6691 buf_discard_identity(hdr);
6693 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6694 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6697 ASSERT(HDR_EMPTY(hdr));
6701 * If the block to be written was all-zero or compressed enough to be
6702 * embedded in the BP, no write was performed so there will be no
6703 * dva/birth/checksum. The buffer must therefore remain anonymous
6706 if (!HDR_EMPTY(hdr)) {
6707 arc_buf_hdr_t *exists;
6708 kmutex_t *hash_lock;
6710 ASSERT3U(zio->io_error, ==, 0);
6712 arc_cksum_verify(buf);
6714 exists = buf_hash_insert(hdr, &hash_lock);
6715 if (exists != NULL) {
6717 * This can only happen if we overwrite for
6718 * sync-to-convergence, because we remove
6719 * buffers from the hash table when we arc_free().
6721 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6722 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6723 panic("bad overwrite, hdr=%p exists=%p",
6724 (void *)hdr, (void *)exists);
6725 ASSERT(zfs_refcount_is_zero(
6726 &exists->b_l1hdr.b_refcnt));
6727 arc_change_state(arc_anon, exists);
6728 arc_hdr_destroy(exists);
6729 mutex_exit(hash_lock);
6730 exists = buf_hash_insert(hdr, &hash_lock);
6731 ASSERT3P(exists, ==, NULL);
6732 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6734 ASSERT(zio->io_prop.zp_nopwrite);
6735 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6736 panic("bad nopwrite, hdr=%p exists=%p",
6737 (void *)hdr, (void *)exists);
6740 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6741 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6742 ASSERT(BP_GET_DEDUP(zio->io_bp));
6743 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6746 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6747 VERIFY3S(remove_reference(hdr, hdr), >, 0);
6748 /* if it's not anon, we are doing a scrub */
6749 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6750 arc_access(hdr, 0, B_FALSE);
6751 mutex_exit(hash_lock);
6753 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6754 VERIFY3S(remove_reference(hdr, hdr), >, 0);
6757 callback->awcb_done(zio, buf, callback->awcb_private);
6759 abd_free(zio->io_abd);
6760 kmem_free(callback, sizeof (arc_write_callback_t));
6764 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
6765 blkptr_t *bp, arc_buf_t *buf, boolean_t uncached, boolean_t l2arc,
6766 const zio_prop_t *zp, arc_write_done_func_t *ready,
6767 arc_write_done_func_t *children_ready, arc_write_done_func_t *done,
6768 void *private, zio_priority_t priority, int zio_flags,
6769 const zbookmark_phys_t *zb)
6771 arc_buf_hdr_t *hdr = buf->b_hdr;
6772 arc_write_callback_t *callback;
6774 zio_prop_t localprop = *zp;
6776 ASSERT3P(ready, !=, NULL);
6777 ASSERT3P(done, !=, NULL);
6778 ASSERT(!HDR_IO_ERROR(hdr));
6779 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6780 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6781 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6783 arc_hdr_set_flags(hdr, ARC_FLAG_UNCACHED);
6785 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6787 if (ARC_BUF_ENCRYPTED(buf)) {
6788 ASSERT(ARC_BUF_COMPRESSED(buf));
6789 localprop.zp_encrypt = B_TRUE;
6790 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6791 localprop.zp_complevel = hdr->b_complevel;
6792 localprop.zp_byteorder =
6793 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
6794 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
6795 memcpy(localprop.zp_salt, hdr->b_crypt_hdr.b_salt,
6797 memcpy(localprop.zp_iv, hdr->b_crypt_hdr.b_iv,
6799 memcpy(localprop.zp_mac, hdr->b_crypt_hdr.b_mac,
6801 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
6802 localprop.zp_nopwrite = B_FALSE;
6803 localprop.zp_copies =
6804 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
6806 zio_flags |= ZIO_FLAG_RAW;
6807 } else if (ARC_BUF_COMPRESSED(buf)) {
6808 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6809 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6810 localprop.zp_complevel = hdr->b_complevel;
6811 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6813 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6814 callback->awcb_ready = ready;
6815 callback->awcb_children_ready = children_ready;
6816 callback->awcb_done = done;
6817 callback->awcb_private = private;
6818 callback->awcb_buf = buf;
6821 * The hdr's b_pabd is now stale, free it now. A new data block
6822 * will be allocated when the zio pipeline calls arc_write_ready().
6824 if (hdr->b_l1hdr.b_pabd != NULL) {
6826 * If the buf is currently sharing the data block with
6827 * the hdr then we need to break that relationship here.
6828 * The hdr will remain with a NULL data pointer and the
6829 * buf will take sole ownership of the block.
6831 if (arc_buf_is_shared(buf)) {
6832 arc_unshare_buf(hdr, buf);
6834 arc_hdr_free_abd(hdr, B_FALSE);
6836 VERIFY3P(buf->b_data, !=, NULL);
6839 if (HDR_HAS_RABD(hdr))
6840 arc_hdr_free_abd(hdr, B_TRUE);
6842 if (!(zio_flags & ZIO_FLAG_RAW))
6843 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6845 ASSERT(!arc_buf_is_shared(buf));
6846 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6848 zio = zio_write(pio, spa, txg, bp,
6849 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6850 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6851 (children_ready != NULL) ? arc_write_children_ready : NULL,
6852 arc_write_done, callback, priority, zio_flags, zb);
6858 arc_tempreserve_clear(uint64_t reserve)
6860 atomic_add_64(&arc_tempreserve, -reserve);
6861 ASSERT((int64_t)arc_tempreserve >= 0);
6865 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
6871 reserve > arc_c/4 &&
6872 reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT))
6873 arc_c = MIN(arc_c_max, reserve * 4);
6876 * Throttle when the calculated memory footprint for the TXG
6877 * exceeds the target ARC size.
6879 if (reserve > arc_c) {
6880 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
6881 return (SET_ERROR(ERESTART));
6885 * Don't count loaned bufs as in flight dirty data to prevent long
6886 * network delays from blocking transactions that are ready to be
6887 * assigned to a txg.
6890 /* assert that it has not wrapped around */
6891 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6893 anon_size = MAX((int64_t)
6894 (zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_DATA]) +
6895 zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_METADATA]) -
6896 arc_loaned_bytes), 0);
6899 * Writes will, almost always, require additional memory allocations
6900 * in order to compress/encrypt/etc the data. We therefore need to
6901 * make sure that there is sufficient available memory for this.
6903 error = arc_memory_throttle(spa, reserve, txg);
6908 * Throttle writes when the amount of dirty data in the cache
6909 * gets too large. We try to keep the cache less than half full
6910 * of dirty blocks so that our sync times don't grow too large.
6912 * In the case of one pool being built on another pool, we want
6913 * to make sure we don't end up throttling the lower (backing)
6914 * pool when the upper pool is the majority contributor to dirty
6915 * data. To insure we make forward progress during throttling, we
6916 * also check the current pool's net dirty data and only throttle
6917 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
6918 * data in the cache.
6920 * Note: if two requests come in concurrently, we might let them
6921 * both succeed, when one of them should fail. Not a huge deal.
6923 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
6924 uint64_t spa_dirty_anon = spa_dirty_data(spa);
6925 uint64_t rarc_c = arc_warm ? arc_c : arc_c_max;
6926 if (total_dirty > rarc_c * zfs_arc_dirty_limit_percent / 100 &&
6927 anon_size > rarc_c * zfs_arc_anon_limit_percent / 100 &&
6928 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
6930 uint64_t meta_esize = zfs_refcount_count(
6931 &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6932 uint64_t data_esize =
6933 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6934 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6935 "anon_data=%lluK tempreserve=%lluK rarc_c=%lluK\n",
6936 (u_longlong_t)arc_tempreserve >> 10,
6937 (u_longlong_t)meta_esize >> 10,
6938 (u_longlong_t)data_esize >> 10,
6939 (u_longlong_t)reserve >> 10,
6940 (u_longlong_t)rarc_c >> 10);
6942 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
6943 return (SET_ERROR(ERESTART));
6945 atomic_add_64(&arc_tempreserve, reserve);
6950 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6951 kstat_named_t *data, kstat_named_t *metadata,
6952 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6955 zfs_refcount_count(&state->arcs_size[ARC_BUFC_DATA]);
6956 metadata->value.ui64 =
6957 zfs_refcount_count(&state->arcs_size[ARC_BUFC_METADATA]);
6958 size->value.ui64 = data->value.ui64 + metadata->value.ui64;
6959 evict_data->value.ui64 =
6960 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6961 evict_metadata->value.ui64 =
6962 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6966 arc_kstat_update(kstat_t *ksp, int rw)
6968 arc_stats_t *as = ksp->ks_data;
6970 if (rw == KSTAT_WRITE)
6971 return (SET_ERROR(EACCES));
6973 as->arcstat_hits.value.ui64 =
6974 wmsum_value(&arc_sums.arcstat_hits);
6975 as->arcstat_iohits.value.ui64 =
6976 wmsum_value(&arc_sums.arcstat_iohits);
6977 as->arcstat_misses.value.ui64 =
6978 wmsum_value(&arc_sums.arcstat_misses);
6979 as->arcstat_demand_data_hits.value.ui64 =
6980 wmsum_value(&arc_sums.arcstat_demand_data_hits);
6981 as->arcstat_demand_data_iohits.value.ui64 =
6982 wmsum_value(&arc_sums.arcstat_demand_data_iohits);
6983 as->arcstat_demand_data_misses.value.ui64 =
6984 wmsum_value(&arc_sums.arcstat_demand_data_misses);
6985 as->arcstat_demand_metadata_hits.value.ui64 =
6986 wmsum_value(&arc_sums.arcstat_demand_metadata_hits);
6987 as->arcstat_demand_metadata_iohits.value.ui64 =
6988 wmsum_value(&arc_sums.arcstat_demand_metadata_iohits);
6989 as->arcstat_demand_metadata_misses.value.ui64 =
6990 wmsum_value(&arc_sums.arcstat_demand_metadata_misses);
6991 as->arcstat_prefetch_data_hits.value.ui64 =
6992 wmsum_value(&arc_sums.arcstat_prefetch_data_hits);
6993 as->arcstat_prefetch_data_iohits.value.ui64 =
6994 wmsum_value(&arc_sums.arcstat_prefetch_data_iohits);
6995 as->arcstat_prefetch_data_misses.value.ui64 =
6996 wmsum_value(&arc_sums.arcstat_prefetch_data_misses);
6997 as->arcstat_prefetch_metadata_hits.value.ui64 =
6998 wmsum_value(&arc_sums.arcstat_prefetch_metadata_hits);
6999 as->arcstat_prefetch_metadata_iohits.value.ui64 =
7000 wmsum_value(&arc_sums.arcstat_prefetch_metadata_iohits);
7001 as->arcstat_prefetch_metadata_misses.value.ui64 =
7002 wmsum_value(&arc_sums.arcstat_prefetch_metadata_misses);
7003 as->arcstat_mru_hits.value.ui64 =
7004 wmsum_value(&arc_sums.arcstat_mru_hits);
7005 as->arcstat_mru_ghost_hits.value.ui64 =
7006 wmsum_value(&arc_sums.arcstat_mru_ghost_hits);
7007 as->arcstat_mfu_hits.value.ui64 =
7008 wmsum_value(&arc_sums.arcstat_mfu_hits);
7009 as->arcstat_mfu_ghost_hits.value.ui64 =
7010 wmsum_value(&arc_sums.arcstat_mfu_ghost_hits);
7011 as->arcstat_uncached_hits.value.ui64 =
7012 wmsum_value(&arc_sums.arcstat_uncached_hits);
7013 as->arcstat_deleted.value.ui64 =
7014 wmsum_value(&arc_sums.arcstat_deleted);
7015 as->arcstat_mutex_miss.value.ui64 =
7016 wmsum_value(&arc_sums.arcstat_mutex_miss);
7017 as->arcstat_access_skip.value.ui64 =
7018 wmsum_value(&arc_sums.arcstat_access_skip);
7019 as->arcstat_evict_skip.value.ui64 =
7020 wmsum_value(&arc_sums.arcstat_evict_skip);
7021 as->arcstat_evict_not_enough.value.ui64 =
7022 wmsum_value(&arc_sums.arcstat_evict_not_enough);
7023 as->arcstat_evict_l2_cached.value.ui64 =
7024 wmsum_value(&arc_sums.arcstat_evict_l2_cached);
7025 as->arcstat_evict_l2_eligible.value.ui64 =
7026 wmsum_value(&arc_sums.arcstat_evict_l2_eligible);
7027 as->arcstat_evict_l2_eligible_mfu.value.ui64 =
7028 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mfu);
7029 as->arcstat_evict_l2_eligible_mru.value.ui64 =
7030 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mru);
7031 as->arcstat_evict_l2_ineligible.value.ui64 =
7032 wmsum_value(&arc_sums.arcstat_evict_l2_ineligible);
7033 as->arcstat_evict_l2_skip.value.ui64 =
7034 wmsum_value(&arc_sums.arcstat_evict_l2_skip);
7035 as->arcstat_hash_collisions.value.ui64 =
7036 wmsum_value(&arc_sums.arcstat_hash_collisions);
7037 as->arcstat_hash_chains.value.ui64 =
7038 wmsum_value(&arc_sums.arcstat_hash_chains);
7039 as->arcstat_size.value.ui64 =
7040 aggsum_value(&arc_sums.arcstat_size);
7041 as->arcstat_compressed_size.value.ui64 =
7042 wmsum_value(&arc_sums.arcstat_compressed_size);
7043 as->arcstat_uncompressed_size.value.ui64 =
7044 wmsum_value(&arc_sums.arcstat_uncompressed_size);
7045 as->arcstat_overhead_size.value.ui64 =
7046 wmsum_value(&arc_sums.arcstat_overhead_size);
7047 as->arcstat_hdr_size.value.ui64 =
7048 wmsum_value(&arc_sums.arcstat_hdr_size);
7049 as->arcstat_data_size.value.ui64 =
7050 wmsum_value(&arc_sums.arcstat_data_size);
7051 as->arcstat_metadata_size.value.ui64 =
7052 wmsum_value(&arc_sums.arcstat_metadata_size);
7053 as->arcstat_dbuf_size.value.ui64 =
7054 wmsum_value(&arc_sums.arcstat_dbuf_size);
7055 #if defined(COMPAT_FREEBSD11)
7056 as->arcstat_other_size.value.ui64 =
7057 wmsum_value(&arc_sums.arcstat_bonus_size) +
7058 wmsum_value(&arc_sums.arcstat_dnode_size) +
7059 wmsum_value(&arc_sums.arcstat_dbuf_size);
7062 arc_kstat_update_state(arc_anon,
7063 &as->arcstat_anon_size,
7064 &as->arcstat_anon_data,
7065 &as->arcstat_anon_metadata,
7066 &as->arcstat_anon_evictable_data,
7067 &as->arcstat_anon_evictable_metadata);
7068 arc_kstat_update_state(arc_mru,
7069 &as->arcstat_mru_size,
7070 &as->arcstat_mru_data,
7071 &as->arcstat_mru_metadata,
7072 &as->arcstat_mru_evictable_data,
7073 &as->arcstat_mru_evictable_metadata);
7074 arc_kstat_update_state(arc_mru_ghost,
7075 &as->arcstat_mru_ghost_size,
7076 &as->arcstat_mru_ghost_data,
7077 &as->arcstat_mru_ghost_metadata,
7078 &as->arcstat_mru_ghost_evictable_data,
7079 &as->arcstat_mru_ghost_evictable_metadata);
7080 arc_kstat_update_state(arc_mfu,
7081 &as->arcstat_mfu_size,
7082 &as->arcstat_mfu_data,
7083 &as->arcstat_mfu_metadata,
7084 &as->arcstat_mfu_evictable_data,
7085 &as->arcstat_mfu_evictable_metadata);
7086 arc_kstat_update_state(arc_mfu_ghost,
7087 &as->arcstat_mfu_ghost_size,
7088 &as->arcstat_mfu_ghost_data,
7089 &as->arcstat_mfu_ghost_metadata,
7090 &as->arcstat_mfu_ghost_evictable_data,
7091 &as->arcstat_mfu_ghost_evictable_metadata);
7092 arc_kstat_update_state(arc_uncached,
7093 &as->arcstat_uncached_size,
7094 &as->arcstat_uncached_data,
7095 &as->arcstat_uncached_metadata,
7096 &as->arcstat_uncached_evictable_data,
7097 &as->arcstat_uncached_evictable_metadata);
7099 as->arcstat_dnode_size.value.ui64 =
7100 wmsum_value(&arc_sums.arcstat_dnode_size);
7101 as->arcstat_bonus_size.value.ui64 =
7102 wmsum_value(&arc_sums.arcstat_bonus_size);
7103 as->arcstat_l2_hits.value.ui64 =
7104 wmsum_value(&arc_sums.arcstat_l2_hits);
7105 as->arcstat_l2_misses.value.ui64 =
7106 wmsum_value(&arc_sums.arcstat_l2_misses);
7107 as->arcstat_l2_prefetch_asize.value.ui64 =
7108 wmsum_value(&arc_sums.arcstat_l2_prefetch_asize);
7109 as->arcstat_l2_mru_asize.value.ui64 =
7110 wmsum_value(&arc_sums.arcstat_l2_mru_asize);
7111 as->arcstat_l2_mfu_asize.value.ui64 =
7112 wmsum_value(&arc_sums.arcstat_l2_mfu_asize);
7113 as->arcstat_l2_bufc_data_asize.value.ui64 =
7114 wmsum_value(&arc_sums.arcstat_l2_bufc_data_asize);
7115 as->arcstat_l2_bufc_metadata_asize.value.ui64 =
7116 wmsum_value(&arc_sums.arcstat_l2_bufc_metadata_asize);
7117 as->arcstat_l2_feeds.value.ui64 =
7118 wmsum_value(&arc_sums.arcstat_l2_feeds);
7119 as->arcstat_l2_rw_clash.value.ui64 =
7120 wmsum_value(&arc_sums.arcstat_l2_rw_clash);
7121 as->arcstat_l2_read_bytes.value.ui64 =
7122 wmsum_value(&arc_sums.arcstat_l2_read_bytes);
7123 as->arcstat_l2_write_bytes.value.ui64 =
7124 wmsum_value(&arc_sums.arcstat_l2_write_bytes);
7125 as->arcstat_l2_writes_sent.value.ui64 =
7126 wmsum_value(&arc_sums.arcstat_l2_writes_sent);
7127 as->arcstat_l2_writes_done.value.ui64 =
7128 wmsum_value(&arc_sums.arcstat_l2_writes_done);
7129 as->arcstat_l2_writes_error.value.ui64 =
7130 wmsum_value(&arc_sums.arcstat_l2_writes_error);
7131 as->arcstat_l2_writes_lock_retry.value.ui64 =
7132 wmsum_value(&arc_sums.arcstat_l2_writes_lock_retry);
7133 as->arcstat_l2_evict_lock_retry.value.ui64 =
7134 wmsum_value(&arc_sums.arcstat_l2_evict_lock_retry);
7135 as->arcstat_l2_evict_reading.value.ui64 =
7136 wmsum_value(&arc_sums.arcstat_l2_evict_reading);
7137 as->arcstat_l2_evict_l1cached.value.ui64 =
7138 wmsum_value(&arc_sums.arcstat_l2_evict_l1cached);
7139 as->arcstat_l2_free_on_write.value.ui64 =
7140 wmsum_value(&arc_sums.arcstat_l2_free_on_write);
7141 as->arcstat_l2_abort_lowmem.value.ui64 =
7142 wmsum_value(&arc_sums.arcstat_l2_abort_lowmem);
7143 as->arcstat_l2_cksum_bad.value.ui64 =
7144 wmsum_value(&arc_sums.arcstat_l2_cksum_bad);
7145 as->arcstat_l2_io_error.value.ui64 =
7146 wmsum_value(&arc_sums.arcstat_l2_io_error);
7147 as->arcstat_l2_lsize.value.ui64 =
7148 wmsum_value(&arc_sums.arcstat_l2_lsize);
7149 as->arcstat_l2_psize.value.ui64 =
7150 wmsum_value(&arc_sums.arcstat_l2_psize);
7151 as->arcstat_l2_hdr_size.value.ui64 =
7152 aggsum_value(&arc_sums.arcstat_l2_hdr_size);
7153 as->arcstat_l2_log_blk_writes.value.ui64 =
7154 wmsum_value(&arc_sums.arcstat_l2_log_blk_writes);
7155 as->arcstat_l2_log_blk_asize.value.ui64 =
7156 wmsum_value(&arc_sums.arcstat_l2_log_blk_asize);
7157 as->arcstat_l2_log_blk_count.value.ui64 =
7158 wmsum_value(&arc_sums.arcstat_l2_log_blk_count);
7159 as->arcstat_l2_rebuild_success.value.ui64 =
7160 wmsum_value(&arc_sums.arcstat_l2_rebuild_success);
7161 as->arcstat_l2_rebuild_abort_unsupported.value.ui64 =
7162 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7163 as->arcstat_l2_rebuild_abort_io_errors.value.ui64 =
7164 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7165 as->arcstat_l2_rebuild_abort_dh_errors.value.ui64 =
7166 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7167 as->arcstat_l2_rebuild_abort_cksum_lb_errors.value.ui64 =
7168 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7169 as->arcstat_l2_rebuild_abort_lowmem.value.ui64 =
7170 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7171 as->arcstat_l2_rebuild_size.value.ui64 =
7172 wmsum_value(&arc_sums.arcstat_l2_rebuild_size);
7173 as->arcstat_l2_rebuild_asize.value.ui64 =
7174 wmsum_value(&arc_sums.arcstat_l2_rebuild_asize);
7175 as->arcstat_l2_rebuild_bufs.value.ui64 =
7176 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs);
7177 as->arcstat_l2_rebuild_bufs_precached.value.ui64 =
7178 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7179 as->arcstat_l2_rebuild_log_blks.value.ui64 =
7180 wmsum_value(&arc_sums.arcstat_l2_rebuild_log_blks);
7181 as->arcstat_memory_throttle_count.value.ui64 =
7182 wmsum_value(&arc_sums.arcstat_memory_throttle_count);
7183 as->arcstat_memory_direct_count.value.ui64 =
7184 wmsum_value(&arc_sums.arcstat_memory_direct_count);
7185 as->arcstat_memory_indirect_count.value.ui64 =
7186 wmsum_value(&arc_sums.arcstat_memory_indirect_count);
7188 as->arcstat_memory_all_bytes.value.ui64 =
7190 as->arcstat_memory_free_bytes.value.ui64 =
7192 as->arcstat_memory_available_bytes.value.i64 =
7193 arc_available_memory();
7195 as->arcstat_prune.value.ui64 =
7196 wmsum_value(&arc_sums.arcstat_prune);
7197 as->arcstat_meta_used.value.ui64 =
7198 wmsum_value(&arc_sums.arcstat_meta_used);
7199 as->arcstat_async_upgrade_sync.value.ui64 =
7200 wmsum_value(&arc_sums.arcstat_async_upgrade_sync);
7201 as->arcstat_predictive_prefetch.value.ui64 =
7202 wmsum_value(&arc_sums.arcstat_predictive_prefetch);
7203 as->arcstat_demand_hit_predictive_prefetch.value.ui64 =
7204 wmsum_value(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7205 as->arcstat_demand_iohit_predictive_prefetch.value.ui64 =
7206 wmsum_value(&arc_sums.arcstat_demand_iohit_predictive_prefetch);
7207 as->arcstat_prescient_prefetch.value.ui64 =
7208 wmsum_value(&arc_sums.arcstat_prescient_prefetch);
7209 as->arcstat_demand_hit_prescient_prefetch.value.ui64 =
7210 wmsum_value(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7211 as->arcstat_demand_iohit_prescient_prefetch.value.ui64 =
7212 wmsum_value(&arc_sums.arcstat_demand_iohit_prescient_prefetch);
7213 as->arcstat_raw_size.value.ui64 =
7214 wmsum_value(&arc_sums.arcstat_raw_size);
7215 as->arcstat_cached_only_in_progress.value.ui64 =
7216 wmsum_value(&arc_sums.arcstat_cached_only_in_progress);
7217 as->arcstat_abd_chunk_waste_size.value.ui64 =
7218 wmsum_value(&arc_sums.arcstat_abd_chunk_waste_size);
7224 * This function *must* return indices evenly distributed between all
7225 * sublists of the multilist. This is needed due to how the ARC eviction
7226 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7227 * distributed between all sublists and uses this assumption when
7228 * deciding which sublist to evict from and how much to evict from it.
7231 arc_state_multilist_index_func(multilist_t *ml, void *obj)
7233 arc_buf_hdr_t *hdr = obj;
7236 * We rely on b_dva to generate evenly distributed index
7237 * numbers using buf_hash below. So, as an added precaution,
7238 * let's make sure we never add empty buffers to the arc lists.
7240 ASSERT(!HDR_EMPTY(hdr));
7243 * The assumption here, is the hash value for a given
7244 * arc_buf_hdr_t will remain constant throughout its lifetime
7245 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7246 * Thus, we don't need to store the header's sublist index
7247 * on insertion, as this index can be recalculated on removal.
7249 * Also, the low order bits of the hash value are thought to be
7250 * distributed evenly. Otherwise, in the case that the multilist
7251 * has a power of two number of sublists, each sublists' usage
7252 * would not be evenly distributed. In this context full 64bit
7253 * division would be a waste of time, so limit it to 32 bits.
7255 return ((unsigned int)buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
7256 multilist_get_num_sublists(ml));
7260 arc_state_l2c_multilist_index_func(multilist_t *ml, void *obj)
7262 panic("Header %p insert into arc_l2c_only %p", obj, ml);
7265 #define WARN_IF_TUNING_IGNORED(tuning, value, do_warn) do { \
7266 if ((do_warn) && (tuning) && ((tuning) != (value))) { \
7268 "ignoring tunable %s (using %llu instead)", \
7269 (#tuning), (u_longlong_t)(value)); \
7274 * Called during module initialization and periodically thereafter to
7275 * apply reasonable changes to the exposed performance tunings. Can also be
7276 * called explicitly by param_set_arc_*() functions when ARC tunables are
7277 * updated manually. Non-zero zfs_* values which differ from the currently set
7278 * values will be applied.
7281 arc_tuning_update(boolean_t verbose)
7283 uint64_t allmem = arc_all_memory();
7285 /* Valid range: 32M - <arc_c_max> */
7286 if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) &&
7287 (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) &&
7288 (zfs_arc_min <= arc_c_max)) {
7289 arc_c_min = zfs_arc_min;
7290 arc_c = MAX(arc_c, arc_c_min);
7292 WARN_IF_TUNING_IGNORED(zfs_arc_min, arc_c_min, verbose);
7294 /* Valid range: 64M - <all physical memory> */
7295 if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) &&
7296 (zfs_arc_max >= MIN_ARC_MAX) && (zfs_arc_max < allmem) &&
7297 (zfs_arc_max > arc_c_min)) {
7298 arc_c_max = zfs_arc_max;
7299 arc_c = MIN(arc_c, arc_c_max);
7300 if (arc_dnode_limit > arc_c_max)
7301 arc_dnode_limit = arc_c_max;
7303 WARN_IF_TUNING_IGNORED(zfs_arc_max, arc_c_max, verbose);
7305 /* Valid range: 0 - <all physical memory> */
7306 arc_dnode_limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit :
7307 MIN(zfs_arc_dnode_limit_percent, 100) * arc_c_max / 100;
7308 WARN_IF_TUNING_IGNORED(zfs_arc_dnode_limit, arc_dnode_limit, verbose);
7310 /* Valid range: 1 - N */
7311 if (zfs_arc_grow_retry)
7312 arc_grow_retry = zfs_arc_grow_retry;
7314 /* Valid range: 1 - N */
7315 if (zfs_arc_shrink_shift) {
7316 arc_shrink_shift = zfs_arc_shrink_shift;
7317 arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1);
7320 /* Valid range: 1 - N ms */
7321 if (zfs_arc_min_prefetch_ms)
7322 arc_min_prefetch_ms = zfs_arc_min_prefetch_ms;
7324 /* Valid range: 1 - N ms */
7325 if (zfs_arc_min_prescient_prefetch_ms) {
7326 arc_min_prescient_prefetch_ms =
7327 zfs_arc_min_prescient_prefetch_ms;
7330 /* Valid range: 0 - 100 */
7331 if (zfs_arc_lotsfree_percent <= 100)
7332 arc_lotsfree_percent = zfs_arc_lotsfree_percent;
7333 WARN_IF_TUNING_IGNORED(zfs_arc_lotsfree_percent, arc_lotsfree_percent,
7336 /* Valid range: 0 - <all physical memory> */
7337 if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free))
7338 arc_sys_free = MIN(zfs_arc_sys_free, allmem);
7339 WARN_IF_TUNING_IGNORED(zfs_arc_sys_free, arc_sys_free, verbose);
7343 arc_state_multilist_init(multilist_t *ml,
7344 multilist_sublist_index_func_t *index_func, int *maxcountp)
7346 multilist_create(ml, sizeof (arc_buf_hdr_t),
7347 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), index_func);
7348 *maxcountp = MAX(*maxcountp, multilist_get_num_sublists(ml));
7352 arc_state_init(void)
7354 int num_sublists = 0;
7356 arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_METADATA],
7357 arc_state_multilist_index_func, &num_sublists);
7358 arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_DATA],
7359 arc_state_multilist_index_func, &num_sublists);
7360 arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
7361 arc_state_multilist_index_func, &num_sublists);
7362 arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
7363 arc_state_multilist_index_func, &num_sublists);
7364 arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
7365 arc_state_multilist_index_func, &num_sublists);
7366 arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_DATA],
7367 arc_state_multilist_index_func, &num_sublists);
7368 arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
7369 arc_state_multilist_index_func, &num_sublists);
7370 arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
7371 arc_state_multilist_index_func, &num_sublists);
7372 arc_state_multilist_init(&arc_uncached->arcs_list[ARC_BUFC_METADATA],
7373 arc_state_multilist_index_func, &num_sublists);
7374 arc_state_multilist_init(&arc_uncached->arcs_list[ARC_BUFC_DATA],
7375 arc_state_multilist_index_func, &num_sublists);
7378 * L2 headers should never be on the L2 state list since they don't
7379 * have L1 headers allocated. Special index function asserts that.
7381 arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
7382 arc_state_l2c_multilist_index_func, &num_sublists);
7383 arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
7384 arc_state_l2c_multilist_index_func, &num_sublists);
7387 * Keep track of the number of markers needed to reclaim buffers from
7388 * any ARC state. The markers will be pre-allocated so as to minimize
7389 * the number of memory allocations performed by the eviction thread.
7391 arc_state_evict_marker_count = num_sublists;
7393 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7394 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7395 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7396 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7397 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7398 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7399 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7400 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7401 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7402 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7403 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7404 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7405 zfs_refcount_create(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]);
7406 zfs_refcount_create(&arc_uncached->arcs_esize[ARC_BUFC_DATA]);
7408 zfs_refcount_create(&arc_anon->arcs_size[ARC_BUFC_DATA]);
7409 zfs_refcount_create(&arc_anon->arcs_size[ARC_BUFC_METADATA]);
7410 zfs_refcount_create(&arc_mru->arcs_size[ARC_BUFC_DATA]);
7411 zfs_refcount_create(&arc_mru->arcs_size[ARC_BUFC_METADATA]);
7412 zfs_refcount_create(&arc_mru_ghost->arcs_size[ARC_BUFC_DATA]);
7413 zfs_refcount_create(&arc_mru_ghost->arcs_size[ARC_BUFC_METADATA]);
7414 zfs_refcount_create(&arc_mfu->arcs_size[ARC_BUFC_DATA]);
7415 zfs_refcount_create(&arc_mfu->arcs_size[ARC_BUFC_METADATA]);
7416 zfs_refcount_create(&arc_mfu_ghost->arcs_size[ARC_BUFC_DATA]);
7417 zfs_refcount_create(&arc_mfu_ghost->arcs_size[ARC_BUFC_METADATA]);
7418 zfs_refcount_create(&arc_l2c_only->arcs_size[ARC_BUFC_DATA]);
7419 zfs_refcount_create(&arc_l2c_only->arcs_size[ARC_BUFC_METADATA]);
7420 zfs_refcount_create(&arc_uncached->arcs_size[ARC_BUFC_DATA]);
7421 zfs_refcount_create(&arc_uncached->arcs_size[ARC_BUFC_METADATA]);
7423 wmsum_init(&arc_mru_ghost->arcs_hits[ARC_BUFC_DATA], 0);
7424 wmsum_init(&arc_mru_ghost->arcs_hits[ARC_BUFC_METADATA], 0);
7425 wmsum_init(&arc_mfu_ghost->arcs_hits[ARC_BUFC_DATA], 0);
7426 wmsum_init(&arc_mfu_ghost->arcs_hits[ARC_BUFC_METADATA], 0);
7428 wmsum_init(&arc_sums.arcstat_hits, 0);
7429 wmsum_init(&arc_sums.arcstat_iohits, 0);
7430 wmsum_init(&arc_sums.arcstat_misses, 0);
7431 wmsum_init(&arc_sums.arcstat_demand_data_hits, 0);
7432 wmsum_init(&arc_sums.arcstat_demand_data_iohits, 0);
7433 wmsum_init(&arc_sums.arcstat_demand_data_misses, 0);
7434 wmsum_init(&arc_sums.arcstat_demand_metadata_hits, 0);
7435 wmsum_init(&arc_sums.arcstat_demand_metadata_iohits, 0);
7436 wmsum_init(&arc_sums.arcstat_demand_metadata_misses, 0);
7437 wmsum_init(&arc_sums.arcstat_prefetch_data_hits, 0);
7438 wmsum_init(&arc_sums.arcstat_prefetch_data_iohits, 0);
7439 wmsum_init(&arc_sums.arcstat_prefetch_data_misses, 0);
7440 wmsum_init(&arc_sums.arcstat_prefetch_metadata_hits, 0);
7441 wmsum_init(&arc_sums.arcstat_prefetch_metadata_iohits, 0);
7442 wmsum_init(&arc_sums.arcstat_prefetch_metadata_misses, 0);
7443 wmsum_init(&arc_sums.arcstat_mru_hits, 0);
7444 wmsum_init(&arc_sums.arcstat_mru_ghost_hits, 0);
7445 wmsum_init(&arc_sums.arcstat_mfu_hits, 0);
7446 wmsum_init(&arc_sums.arcstat_mfu_ghost_hits, 0);
7447 wmsum_init(&arc_sums.arcstat_uncached_hits, 0);
7448 wmsum_init(&arc_sums.arcstat_deleted, 0);
7449 wmsum_init(&arc_sums.arcstat_mutex_miss, 0);
7450 wmsum_init(&arc_sums.arcstat_access_skip, 0);
7451 wmsum_init(&arc_sums.arcstat_evict_skip, 0);
7452 wmsum_init(&arc_sums.arcstat_evict_not_enough, 0);
7453 wmsum_init(&arc_sums.arcstat_evict_l2_cached, 0);
7454 wmsum_init(&arc_sums.arcstat_evict_l2_eligible, 0);
7455 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mfu, 0);
7456 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mru, 0);
7457 wmsum_init(&arc_sums.arcstat_evict_l2_ineligible, 0);
7458 wmsum_init(&arc_sums.arcstat_evict_l2_skip, 0);
7459 wmsum_init(&arc_sums.arcstat_hash_collisions, 0);
7460 wmsum_init(&arc_sums.arcstat_hash_chains, 0);
7461 aggsum_init(&arc_sums.arcstat_size, 0);
7462 wmsum_init(&arc_sums.arcstat_compressed_size, 0);
7463 wmsum_init(&arc_sums.arcstat_uncompressed_size, 0);
7464 wmsum_init(&arc_sums.arcstat_overhead_size, 0);
7465 wmsum_init(&arc_sums.arcstat_hdr_size, 0);
7466 wmsum_init(&arc_sums.arcstat_data_size, 0);
7467 wmsum_init(&arc_sums.arcstat_metadata_size, 0);
7468 wmsum_init(&arc_sums.arcstat_dbuf_size, 0);
7469 wmsum_init(&arc_sums.arcstat_dnode_size, 0);
7470 wmsum_init(&arc_sums.arcstat_bonus_size, 0);
7471 wmsum_init(&arc_sums.arcstat_l2_hits, 0);
7472 wmsum_init(&arc_sums.arcstat_l2_misses, 0);
7473 wmsum_init(&arc_sums.arcstat_l2_prefetch_asize, 0);
7474 wmsum_init(&arc_sums.arcstat_l2_mru_asize, 0);
7475 wmsum_init(&arc_sums.arcstat_l2_mfu_asize, 0);
7476 wmsum_init(&arc_sums.arcstat_l2_bufc_data_asize, 0);
7477 wmsum_init(&arc_sums.arcstat_l2_bufc_metadata_asize, 0);
7478 wmsum_init(&arc_sums.arcstat_l2_feeds, 0);
7479 wmsum_init(&arc_sums.arcstat_l2_rw_clash, 0);
7480 wmsum_init(&arc_sums.arcstat_l2_read_bytes, 0);
7481 wmsum_init(&arc_sums.arcstat_l2_write_bytes, 0);
7482 wmsum_init(&arc_sums.arcstat_l2_writes_sent, 0);
7483 wmsum_init(&arc_sums.arcstat_l2_writes_done, 0);
7484 wmsum_init(&arc_sums.arcstat_l2_writes_error, 0);
7485 wmsum_init(&arc_sums.arcstat_l2_writes_lock_retry, 0);
7486 wmsum_init(&arc_sums.arcstat_l2_evict_lock_retry, 0);
7487 wmsum_init(&arc_sums.arcstat_l2_evict_reading, 0);
7488 wmsum_init(&arc_sums.arcstat_l2_evict_l1cached, 0);
7489 wmsum_init(&arc_sums.arcstat_l2_free_on_write, 0);
7490 wmsum_init(&arc_sums.arcstat_l2_abort_lowmem, 0);
7491 wmsum_init(&arc_sums.arcstat_l2_cksum_bad, 0);
7492 wmsum_init(&arc_sums.arcstat_l2_io_error, 0);
7493 wmsum_init(&arc_sums.arcstat_l2_lsize, 0);
7494 wmsum_init(&arc_sums.arcstat_l2_psize, 0);
7495 aggsum_init(&arc_sums.arcstat_l2_hdr_size, 0);
7496 wmsum_init(&arc_sums.arcstat_l2_log_blk_writes, 0);
7497 wmsum_init(&arc_sums.arcstat_l2_log_blk_asize, 0);
7498 wmsum_init(&arc_sums.arcstat_l2_log_blk_count, 0);
7499 wmsum_init(&arc_sums.arcstat_l2_rebuild_success, 0);
7500 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_unsupported, 0);
7501 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_io_errors, 0);
7502 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_dh_errors, 0);
7503 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors, 0);
7504 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_lowmem, 0);
7505 wmsum_init(&arc_sums.arcstat_l2_rebuild_size, 0);
7506 wmsum_init(&arc_sums.arcstat_l2_rebuild_asize, 0);
7507 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs, 0);
7508 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs_precached, 0);
7509 wmsum_init(&arc_sums.arcstat_l2_rebuild_log_blks, 0);
7510 wmsum_init(&arc_sums.arcstat_memory_throttle_count, 0);
7511 wmsum_init(&arc_sums.arcstat_memory_direct_count, 0);
7512 wmsum_init(&arc_sums.arcstat_memory_indirect_count, 0);
7513 wmsum_init(&arc_sums.arcstat_prune, 0);
7514 wmsum_init(&arc_sums.arcstat_meta_used, 0);
7515 wmsum_init(&arc_sums.arcstat_async_upgrade_sync, 0);
7516 wmsum_init(&arc_sums.arcstat_predictive_prefetch, 0);
7517 wmsum_init(&arc_sums.arcstat_demand_hit_predictive_prefetch, 0);
7518 wmsum_init(&arc_sums.arcstat_demand_iohit_predictive_prefetch, 0);
7519 wmsum_init(&arc_sums.arcstat_prescient_prefetch, 0);
7520 wmsum_init(&arc_sums.arcstat_demand_hit_prescient_prefetch, 0);
7521 wmsum_init(&arc_sums.arcstat_demand_iohit_prescient_prefetch, 0);
7522 wmsum_init(&arc_sums.arcstat_raw_size, 0);
7523 wmsum_init(&arc_sums.arcstat_cached_only_in_progress, 0);
7524 wmsum_init(&arc_sums.arcstat_abd_chunk_waste_size, 0);
7526 arc_anon->arcs_state = ARC_STATE_ANON;
7527 arc_mru->arcs_state = ARC_STATE_MRU;
7528 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7529 arc_mfu->arcs_state = ARC_STATE_MFU;
7530 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7531 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7532 arc_uncached->arcs_state = ARC_STATE_UNCACHED;
7536 arc_state_fini(void)
7538 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7539 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7540 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7541 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7542 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7543 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7544 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7545 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7546 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7547 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7548 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7549 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7550 zfs_refcount_destroy(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]);
7551 zfs_refcount_destroy(&arc_uncached->arcs_esize[ARC_BUFC_DATA]);
7553 zfs_refcount_destroy(&arc_anon->arcs_size[ARC_BUFC_DATA]);
7554 zfs_refcount_destroy(&arc_anon->arcs_size[ARC_BUFC_METADATA]);
7555 zfs_refcount_destroy(&arc_mru->arcs_size[ARC_BUFC_DATA]);
7556 zfs_refcount_destroy(&arc_mru->arcs_size[ARC_BUFC_METADATA]);
7557 zfs_refcount_destroy(&arc_mru_ghost->arcs_size[ARC_BUFC_DATA]);
7558 zfs_refcount_destroy(&arc_mru_ghost->arcs_size[ARC_BUFC_METADATA]);
7559 zfs_refcount_destroy(&arc_mfu->arcs_size[ARC_BUFC_DATA]);
7560 zfs_refcount_destroy(&arc_mfu->arcs_size[ARC_BUFC_METADATA]);
7561 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size[ARC_BUFC_DATA]);
7562 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size[ARC_BUFC_METADATA]);
7563 zfs_refcount_destroy(&arc_l2c_only->arcs_size[ARC_BUFC_DATA]);
7564 zfs_refcount_destroy(&arc_l2c_only->arcs_size[ARC_BUFC_METADATA]);
7565 zfs_refcount_destroy(&arc_uncached->arcs_size[ARC_BUFC_DATA]);
7566 zfs_refcount_destroy(&arc_uncached->arcs_size[ARC_BUFC_METADATA]);
7568 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
7569 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7570 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7571 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7572 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
7573 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7574 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
7575 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7576 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7577 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7578 multilist_destroy(&arc_uncached->arcs_list[ARC_BUFC_METADATA]);
7579 multilist_destroy(&arc_uncached->arcs_list[ARC_BUFC_DATA]);
7581 wmsum_fini(&arc_mru_ghost->arcs_hits[ARC_BUFC_DATA]);
7582 wmsum_fini(&arc_mru_ghost->arcs_hits[ARC_BUFC_METADATA]);
7583 wmsum_fini(&arc_mfu_ghost->arcs_hits[ARC_BUFC_DATA]);
7584 wmsum_fini(&arc_mfu_ghost->arcs_hits[ARC_BUFC_METADATA]);
7586 wmsum_fini(&arc_sums.arcstat_hits);
7587 wmsum_fini(&arc_sums.arcstat_iohits);
7588 wmsum_fini(&arc_sums.arcstat_misses);
7589 wmsum_fini(&arc_sums.arcstat_demand_data_hits);
7590 wmsum_fini(&arc_sums.arcstat_demand_data_iohits);
7591 wmsum_fini(&arc_sums.arcstat_demand_data_misses);
7592 wmsum_fini(&arc_sums.arcstat_demand_metadata_hits);
7593 wmsum_fini(&arc_sums.arcstat_demand_metadata_iohits);
7594 wmsum_fini(&arc_sums.arcstat_demand_metadata_misses);
7595 wmsum_fini(&arc_sums.arcstat_prefetch_data_hits);
7596 wmsum_fini(&arc_sums.arcstat_prefetch_data_iohits);
7597 wmsum_fini(&arc_sums.arcstat_prefetch_data_misses);
7598 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_hits);
7599 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_iohits);
7600 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_misses);
7601 wmsum_fini(&arc_sums.arcstat_mru_hits);
7602 wmsum_fini(&arc_sums.arcstat_mru_ghost_hits);
7603 wmsum_fini(&arc_sums.arcstat_mfu_hits);
7604 wmsum_fini(&arc_sums.arcstat_mfu_ghost_hits);
7605 wmsum_fini(&arc_sums.arcstat_uncached_hits);
7606 wmsum_fini(&arc_sums.arcstat_deleted);
7607 wmsum_fini(&arc_sums.arcstat_mutex_miss);
7608 wmsum_fini(&arc_sums.arcstat_access_skip);
7609 wmsum_fini(&arc_sums.arcstat_evict_skip);
7610 wmsum_fini(&arc_sums.arcstat_evict_not_enough);
7611 wmsum_fini(&arc_sums.arcstat_evict_l2_cached);
7612 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible);
7613 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mfu);
7614 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mru);
7615 wmsum_fini(&arc_sums.arcstat_evict_l2_ineligible);
7616 wmsum_fini(&arc_sums.arcstat_evict_l2_skip);
7617 wmsum_fini(&arc_sums.arcstat_hash_collisions);
7618 wmsum_fini(&arc_sums.arcstat_hash_chains);
7619 aggsum_fini(&arc_sums.arcstat_size);
7620 wmsum_fini(&arc_sums.arcstat_compressed_size);
7621 wmsum_fini(&arc_sums.arcstat_uncompressed_size);
7622 wmsum_fini(&arc_sums.arcstat_overhead_size);
7623 wmsum_fini(&arc_sums.arcstat_hdr_size);
7624 wmsum_fini(&arc_sums.arcstat_data_size);
7625 wmsum_fini(&arc_sums.arcstat_metadata_size);
7626 wmsum_fini(&arc_sums.arcstat_dbuf_size);
7627 wmsum_fini(&arc_sums.arcstat_dnode_size);
7628 wmsum_fini(&arc_sums.arcstat_bonus_size);
7629 wmsum_fini(&arc_sums.arcstat_l2_hits);
7630 wmsum_fini(&arc_sums.arcstat_l2_misses);
7631 wmsum_fini(&arc_sums.arcstat_l2_prefetch_asize);
7632 wmsum_fini(&arc_sums.arcstat_l2_mru_asize);
7633 wmsum_fini(&arc_sums.arcstat_l2_mfu_asize);
7634 wmsum_fini(&arc_sums.arcstat_l2_bufc_data_asize);
7635 wmsum_fini(&arc_sums.arcstat_l2_bufc_metadata_asize);
7636 wmsum_fini(&arc_sums.arcstat_l2_feeds);
7637 wmsum_fini(&arc_sums.arcstat_l2_rw_clash);
7638 wmsum_fini(&arc_sums.arcstat_l2_read_bytes);
7639 wmsum_fini(&arc_sums.arcstat_l2_write_bytes);
7640 wmsum_fini(&arc_sums.arcstat_l2_writes_sent);
7641 wmsum_fini(&arc_sums.arcstat_l2_writes_done);
7642 wmsum_fini(&arc_sums.arcstat_l2_writes_error);
7643 wmsum_fini(&arc_sums.arcstat_l2_writes_lock_retry);
7644 wmsum_fini(&arc_sums.arcstat_l2_evict_lock_retry);
7645 wmsum_fini(&arc_sums.arcstat_l2_evict_reading);
7646 wmsum_fini(&arc_sums.arcstat_l2_evict_l1cached);
7647 wmsum_fini(&arc_sums.arcstat_l2_free_on_write);
7648 wmsum_fini(&arc_sums.arcstat_l2_abort_lowmem);
7649 wmsum_fini(&arc_sums.arcstat_l2_cksum_bad);
7650 wmsum_fini(&arc_sums.arcstat_l2_io_error);
7651 wmsum_fini(&arc_sums.arcstat_l2_lsize);
7652 wmsum_fini(&arc_sums.arcstat_l2_psize);
7653 aggsum_fini(&arc_sums.arcstat_l2_hdr_size);
7654 wmsum_fini(&arc_sums.arcstat_l2_log_blk_writes);
7655 wmsum_fini(&arc_sums.arcstat_l2_log_blk_asize);
7656 wmsum_fini(&arc_sums.arcstat_l2_log_blk_count);
7657 wmsum_fini(&arc_sums.arcstat_l2_rebuild_success);
7658 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7659 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7660 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7661 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7662 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7663 wmsum_fini(&arc_sums.arcstat_l2_rebuild_size);
7664 wmsum_fini(&arc_sums.arcstat_l2_rebuild_asize);
7665 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs);
7666 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7667 wmsum_fini(&arc_sums.arcstat_l2_rebuild_log_blks);
7668 wmsum_fini(&arc_sums.arcstat_memory_throttle_count);
7669 wmsum_fini(&arc_sums.arcstat_memory_direct_count);
7670 wmsum_fini(&arc_sums.arcstat_memory_indirect_count);
7671 wmsum_fini(&arc_sums.arcstat_prune);
7672 wmsum_fini(&arc_sums.arcstat_meta_used);
7673 wmsum_fini(&arc_sums.arcstat_async_upgrade_sync);
7674 wmsum_fini(&arc_sums.arcstat_predictive_prefetch);
7675 wmsum_fini(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7676 wmsum_fini(&arc_sums.arcstat_demand_iohit_predictive_prefetch);
7677 wmsum_fini(&arc_sums.arcstat_prescient_prefetch);
7678 wmsum_fini(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7679 wmsum_fini(&arc_sums.arcstat_demand_iohit_prescient_prefetch);
7680 wmsum_fini(&arc_sums.arcstat_raw_size);
7681 wmsum_fini(&arc_sums.arcstat_cached_only_in_progress);
7682 wmsum_fini(&arc_sums.arcstat_abd_chunk_waste_size);
7686 arc_target_bytes(void)
7692 arc_set_limits(uint64_t allmem)
7694 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more. */
7695 arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT);
7697 /* How to set default max varies by platform. */
7698 arc_c_max = arc_default_max(arc_c_min, allmem);
7703 uint64_t percent, allmem = arc_all_memory();
7704 mutex_init(&arc_evict_lock, NULL, MUTEX_DEFAULT, NULL);
7705 list_create(&arc_evict_waiters, sizeof (arc_evict_waiter_t),
7706 offsetof(arc_evict_waiter_t, aew_node));
7708 arc_min_prefetch_ms = 1000;
7709 arc_min_prescient_prefetch_ms = 6000;
7711 #if defined(_KERNEL)
7715 arc_set_limits(allmem);
7719 * If zfs_arc_max is non-zero at init, meaning it was set in the kernel
7720 * environment before the module was loaded, don't block setting the
7721 * maximum because it is less than arc_c_min, instead, reset arc_c_min
7723 * zfs_arc_min will be handled by arc_tuning_update().
7725 if (zfs_arc_max != 0 && zfs_arc_max >= MIN_ARC_MAX &&
7726 zfs_arc_max < allmem) {
7727 arc_c_max = zfs_arc_max;
7728 if (arc_c_min >= arc_c_max) {
7729 arc_c_min = MAX(zfs_arc_max / 2,
7730 2ULL << SPA_MAXBLOCKSHIFT);
7735 * In userland, there's only the memory pressure that we artificially
7736 * create (see arc_available_memory()). Don't let arc_c get too
7737 * small, because it can cause transactions to be larger than
7738 * arc_c, causing arc_tempreserve_space() to fail.
7740 arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT);
7745 * 32-bit fixed point fractions of metadata from total ARC size,
7746 * MRU data from all data and MRU metadata from all metadata.
7748 arc_meta = (1ULL << 32) / 4; /* Metadata is 25% of arc_c. */
7749 arc_pd = (1ULL << 32) / 2; /* Data MRU is 50% of data. */
7750 arc_pm = (1ULL << 32) / 2; /* Metadata MRU is 50% of metadata. */
7752 percent = MIN(zfs_arc_dnode_limit_percent, 100);
7753 arc_dnode_limit = arc_c_max * percent / 100;
7755 /* Apply user specified tunings */
7756 arc_tuning_update(B_TRUE);
7758 /* if kmem_flags are set, lets try to use less memory */
7759 if (kmem_debugging())
7761 if (arc_c < arc_c_min)
7764 arc_register_hotplug();
7770 list_create(&arc_prune_list, sizeof (arc_prune_t),
7771 offsetof(arc_prune_t, p_node));
7772 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7774 arc_prune_taskq = taskq_create("arc_prune", zfs_arc_prune_task_threads,
7775 defclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
7777 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7778 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7780 if (arc_ksp != NULL) {
7781 arc_ksp->ks_data = &arc_stats;
7782 arc_ksp->ks_update = arc_kstat_update;
7783 kstat_install(arc_ksp);
7786 arc_state_evict_markers =
7787 arc_state_alloc_markers(arc_state_evict_marker_count);
7788 arc_evict_zthr = zthr_create_timer("arc_evict",
7789 arc_evict_cb_check, arc_evict_cb, NULL, SEC2NSEC(1), defclsyspri);
7790 arc_reap_zthr = zthr_create_timer("arc_reap",
7791 arc_reap_cb_check, arc_reap_cb, NULL, SEC2NSEC(1), minclsyspri);
7796 * Calculate maximum amount of dirty data per pool.
7798 * If it has been set by a module parameter, take that.
7799 * Otherwise, use a percentage of physical memory defined by
7800 * zfs_dirty_data_max_percent (default 10%) with a cap at
7801 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
7804 if (zfs_dirty_data_max_max == 0)
7805 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
7806 allmem * zfs_dirty_data_max_max_percent / 100);
7808 if (zfs_dirty_data_max_max == 0)
7809 zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024,
7810 allmem * zfs_dirty_data_max_max_percent / 100);
7813 if (zfs_dirty_data_max == 0) {
7814 zfs_dirty_data_max = allmem *
7815 zfs_dirty_data_max_percent / 100;
7816 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7817 zfs_dirty_data_max_max);
7820 if (zfs_wrlog_data_max == 0) {
7823 * dp_wrlog_total is reduced for each txg at the end of
7824 * spa_sync(). However, dp_dirty_total is reduced every time
7825 * a block is written out. Thus under normal operation,
7826 * dp_wrlog_total could grow 2 times as big as
7827 * zfs_dirty_data_max.
7829 zfs_wrlog_data_max = zfs_dirty_data_max * 2;
7840 #endif /* _KERNEL */
7842 /* Use B_TRUE to ensure *all* buffers are evicted */
7843 arc_flush(NULL, B_TRUE);
7845 if (arc_ksp != NULL) {
7846 kstat_delete(arc_ksp);
7850 taskq_wait(arc_prune_taskq);
7851 taskq_destroy(arc_prune_taskq);
7853 mutex_enter(&arc_prune_mtx);
7854 while ((p = list_remove_head(&arc_prune_list)) != NULL) {
7855 zfs_refcount_remove(&p->p_refcnt, &arc_prune_list);
7856 zfs_refcount_destroy(&p->p_refcnt);
7857 kmem_free(p, sizeof (*p));
7859 mutex_exit(&arc_prune_mtx);
7861 list_destroy(&arc_prune_list);
7862 mutex_destroy(&arc_prune_mtx);
7864 (void) zthr_cancel(arc_evict_zthr);
7865 (void) zthr_cancel(arc_reap_zthr);
7866 arc_state_free_markers(arc_state_evict_markers,
7867 arc_state_evict_marker_count);
7869 mutex_destroy(&arc_evict_lock);
7870 list_destroy(&arc_evict_waiters);
7873 * Free any buffers that were tagged for destruction. This needs
7874 * to occur before arc_state_fini() runs and destroys the aggsum
7875 * values which are updated when freeing scatter ABDs.
7877 l2arc_do_free_on_write();
7880 * buf_fini() must proceed arc_state_fini() because buf_fin() may
7881 * trigger the release of kmem magazines, which can callback to
7882 * arc_space_return() which accesses aggsums freed in act_state_fini().
7887 arc_unregister_hotplug();
7890 * We destroy the zthrs after all the ARC state has been
7891 * torn down to avoid the case of them receiving any
7892 * wakeup() signals after they are destroyed.
7894 zthr_destroy(arc_evict_zthr);
7895 zthr_destroy(arc_reap_zthr);
7897 ASSERT0(arc_loaned_bytes);
7903 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7904 * It uses dedicated storage devices to hold cached data, which are populated
7905 * using large infrequent writes. The main role of this cache is to boost
7906 * the performance of random read workloads. The intended L2ARC devices
7907 * include short-stroked disks, solid state disks, and other media with
7908 * substantially faster read latency than disk.
7910 * +-----------------------+
7912 * +-----------------------+
7915 * l2arc_feed_thread() arc_read()
7919 * +---------------+ |
7921 * +---------------+ |
7926 * +-------+ +-------+
7928 * | cache | | cache |
7929 * +-------+ +-------+
7930 * +=========+ .-----.
7931 * : L2ARC : |-_____-|
7932 * : devices : | Disks |
7933 * +=========+ `-_____-'
7935 * Read requests are satisfied from the following sources, in order:
7938 * 2) vdev cache of L2ARC devices
7940 * 4) vdev cache of disks
7943 * Some L2ARC device types exhibit extremely slow write performance.
7944 * To accommodate for this there are some significant differences between
7945 * the L2ARC and traditional cache design:
7947 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7948 * the ARC behave as usual, freeing buffers and placing headers on ghost
7949 * lists. The ARC does not send buffers to the L2ARC during eviction as
7950 * this would add inflated write latencies for all ARC memory pressure.
7952 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7953 * It does this by periodically scanning buffers from the eviction-end of
7954 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7955 * not already there. It scans until a headroom of buffers is satisfied,
7956 * which itself is a buffer for ARC eviction. If a compressible buffer is
7957 * found during scanning and selected for writing to an L2ARC device, we
7958 * temporarily boost scanning headroom during the next scan cycle to make
7959 * sure we adapt to compression effects (which might significantly reduce
7960 * the data volume we write to L2ARC). The thread that does this is
7961 * l2arc_feed_thread(), illustrated below; example sizes are included to
7962 * provide a better sense of ratio than this diagram:
7965 * +---------------------+----------+
7966 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7967 * +---------------------+----------+ | o L2ARC eligible
7968 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7969 * +---------------------+----------+ |
7970 * 15.9 Gbytes ^ 32 Mbytes |
7972 * l2arc_feed_thread()
7974 * l2arc write hand <--[oooo]--'
7978 * +==============================+
7979 * L2ARC dev |####|#|###|###| |####| ... |
7980 * +==============================+
7983 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7984 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7985 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7986 * safe to say that this is an uncommon case, since buffers at the end of
7987 * the ARC lists have moved there due to inactivity.
7989 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7990 * then the L2ARC simply misses copying some buffers. This serves as a
7991 * pressure valve to prevent heavy read workloads from both stalling the ARC
7992 * with waits and clogging the L2ARC with writes. This also helps prevent
7993 * the potential for the L2ARC to churn if it attempts to cache content too
7994 * quickly, such as during backups of the entire pool.
7996 * 5. After system boot and before the ARC has filled main memory, there are
7997 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7998 * lists can remain mostly static. Instead of searching from tail of these
7999 * lists as pictured, the l2arc_feed_thread() will search from the list heads
8000 * for eligible buffers, greatly increasing its chance of finding them.
8002 * The L2ARC device write speed is also boosted during this time so that
8003 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
8004 * there are no L2ARC reads, and no fear of degrading read performance
8005 * through increased writes.
8007 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
8008 * the vdev queue can aggregate them into larger and fewer writes. Each
8009 * device is written to in a rotor fashion, sweeping writes through
8010 * available space then repeating.
8012 * 7. The L2ARC does not store dirty content. It never needs to flush
8013 * write buffers back to disk based storage.
8015 * 8. If an ARC buffer is written (and dirtied) which also exists in the
8016 * L2ARC, the now stale L2ARC buffer is immediately dropped.
8018 * The performance of the L2ARC can be tweaked by a number of tunables, which
8019 * may be necessary for different workloads:
8021 * l2arc_write_max max write bytes per interval
8022 * l2arc_write_boost extra write bytes during device warmup
8023 * l2arc_noprefetch skip caching prefetched buffers
8024 * l2arc_headroom number of max device writes to precache
8025 * l2arc_headroom_boost when we find compressed buffers during ARC
8026 * scanning, we multiply headroom by this
8027 * percentage factor for the next scan cycle,
8028 * since more compressed buffers are likely to
8030 * l2arc_feed_secs seconds between L2ARC writing
8032 * Tunables may be removed or added as future performance improvements are
8033 * integrated, and also may become zpool properties.
8035 * There are three key functions that control how the L2ARC warms up:
8037 * l2arc_write_eligible() check if a buffer is eligible to cache
8038 * l2arc_write_size() calculate how much to write
8039 * l2arc_write_interval() calculate sleep delay between writes
8041 * These three functions determine what to write, how much, and how quickly
8044 * L2ARC persistence:
8046 * When writing buffers to L2ARC, we periodically add some metadata to
8047 * make sure we can pick them up after reboot, thus dramatically reducing
8048 * the impact that any downtime has on the performance of storage systems
8049 * with large caches.
8051 * The implementation works fairly simply by integrating the following two
8054 * *) When writing to the L2ARC, we occasionally write a "l2arc log block",
8055 * which is an additional piece of metadata which describes what's been
8056 * written. This allows us to rebuild the arc_buf_hdr_t structures of the
8057 * main ARC buffers. There are 2 linked-lists of log blocks headed by
8058 * dh_start_lbps[2]. We alternate which chain we append to, so they are
8059 * time-wise and offset-wise interleaved, but that is an optimization rather
8060 * than for correctness. The log block also includes a pointer to the
8061 * previous block in its chain.
8063 * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device
8064 * for our header bookkeeping purposes. This contains a device header,
8065 * which contains our top-level reference structures. We update it each
8066 * time we write a new log block, so that we're able to locate it in the
8067 * L2ARC device. If this write results in an inconsistent device header
8068 * (e.g. due to power failure), we detect this by verifying the header's
8069 * checksum and simply fail to reconstruct the L2ARC after reboot.
8071 * Implementation diagram:
8073 * +=== L2ARC device (not to scale) ======================================+
8074 * | ___two newest log block pointers__.__________ |
8075 * | / \dh_start_lbps[1] |
8076 * | / \ \dh_start_lbps[0]|
8078 * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---|
8079 * || hdr| ^ /^ /^ / / |
8080 * |+------+ ...--\-------/ \-----/--\------/ / |
8081 * | \--------------/ \--------------/ |
8082 * +======================================================================+
8084 * As can be seen on the diagram, rather than using a simple linked list,
8085 * we use a pair of linked lists with alternating elements. This is a
8086 * performance enhancement due to the fact that we only find out the
8087 * address of the next log block access once the current block has been
8088 * completely read in. Obviously, this hurts performance, because we'd be
8089 * keeping the device's I/O queue at only a 1 operation deep, thus
8090 * incurring a large amount of I/O round-trip latency. Having two lists
8091 * allows us to fetch two log blocks ahead of where we are currently
8092 * rebuilding L2ARC buffers.
8094 * On-device data structures:
8096 * L2ARC device header: l2arc_dev_hdr_phys_t
8097 * L2ARC log block: l2arc_log_blk_phys_t
8099 * L2ARC reconstruction:
8101 * When writing data, we simply write in the standard rotary fashion,
8102 * evicting buffers as we go and simply writing new data over them (writing
8103 * a new log block every now and then). This obviously means that once we
8104 * loop around the end of the device, we will start cutting into an already
8105 * committed log block (and its referenced data buffers), like so:
8107 * current write head__ __old tail
8110 * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |-->
8111 * ^ ^^^^^^^^^___________________________________
8113 * <<nextwrite>> may overwrite this blk and/or its bufs --'
8115 * When importing the pool, we detect this situation and use it to stop
8116 * our scanning process (see l2arc_rebuild).
8118 * There is one significant caveat to consider when rebuilding ARC contents
8119 * from an L2ARC device: what about invalidated buffers? Given the above
8120 * construction, we cannot update blocks which we've already written to amend
8121 * them to remove buffers which were invalidated. Thus, during reconstruction,
8122 * we might be populating the cache with buffers for data that's not on the
8123 * main pool anymore, or may have been overwritten!
8125 * As it turns out, this isn't a problem. Every arc_read request includes
8126 * both the DVA and, crucially, the birth TXG of the BP the caller is
8127 * looking for. So even if the cache were populated by completely rotten
8128 * blocks for data that had been long deleted and/or overwritten, we'll
8129 * never actually return bad data from the cache, since the DVA with the
8130 * birth TXG uniquely identify a block in space and time - once created,
8131 * a block is immutable on disk. The worst thing we have done is wasted
8132 * some time and memory at l2arc rebuild to reconstruct outdated ARC
8133 * entries that will get dropped from the l2arc as it is being updated
8136 * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write
8137 * hand are not restored. This is done by saving the offset (in bytes)
8138 * l2arc_evict() has evicted to in the L2ARC device header and taking it
8139 * into account when restoring buffers.
8143 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
8146 * A buffer is *not* eligible for the L2ARC if it:
8147 * 1. belongs to a different spa.
8148 * 2. is already cached on the L2ARC.
8149 * 3. has an I/O in progress (it may be an incomplete read).
8150 * 4. is flagged not eligible (zfs property).
8152 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
8153 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
8160 l2arc_write_size(l2arc_dev_t *dev)
8165 * Make sure our globals have meaningful values in case the user
8168 size = l2arc_write_max;
8170 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
8171 "be greater than zero, resetting it to the default (%d)",
8173 size = l2arc_write_max = L2ARC_WRITE_SIZE;
8176 if (arc_warm == B_FALSE)
8177 size += l2arc_write_boost;
8179 /* We need to add in the worst case scenario of log block overhead. */
8180 size += l2arc_log_blk_overhead(size, dev);
8181 if (dev->l2ad_vdev->vdev_has_trim && l2arc_trim_ahead > 0) {
8183 * Trim ahead of the write size 64MB or (l2arc_trim_ahead/100)
8184 * times the writesize, whichever is greater.
8186 size += MAX(64 * 1024 * 1024,
8187 (size * l2arc_trim_ahead) / 100);
8191 * Make sure the write size does not exceed the size of the cache
8192 * device. This is important in l2arc_evict(), otherwise infinite
8193 * iteration can occur.
8195 if (size > dev->l2ad_end - dev->l2ad_start) {
8196 cmn_err(CE_NOTE, "l2arc_write_max or l2arc_write_boost "
8197 "plus the overhead of log blocks (persistent L2ARC, "
8198 "%llu bytes) exceeds the size of the cache device "
8199 "(guid %llu), resetting them to the default (%d)",
8200 (u_longlong_t)l2arc_log_blk_overhead(size, dev),
8201 (u_longlong_t)dev->l2ad_vdev->vdev_guid, L2ARC_WRITE_SIZE);
8203 size = l2arc_write_max = l2arc_write_boost = L2ARC_WRITE_SIZE;
8205 if (l2arc_trim_ahead > 1) {
8206 cmn_err(CE_NOTE, "l2arc_trim_ahead set to 1");
8207 l2arc_trim_ahead = 1;
8210 if (arc_warm == B_FALSE)
8211 size += l2arc_write_boost;
8213 size += l2arc_log_blk_overhead(size, dev);
8214 if (dev->l2ad_vdev->vdev_has_trim && l2arc_trim_ahead > 0) {
8215 size += MAX(64 * 1024 * 1024,
8216 (size * l2arc_trim_ahead) / 100);
8225 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
8227 clock_t interval, next, now;
8230 * If the ARC lists are busy, increase our write rate; if the
8231 * lists are stale, idle back. This is achieved by checking
8232 * how much we previously wrote - if it was more than half of
8233 * what we wanted, schedule the next write much sooner.
8235 if (l2arc_feed_again && wrote > (wanted / 2))
8236 interval = (hz * l2arc_feed_min_ms) / 1000;
8238 interval = hz * l2arc_feed_secs;
8240 now = ddi_get_lbolt();
8241 next = MAX(now, MIN(now + interval, began + interval));
8247 * Cycle through L2ARC devices. This is how L2ARC load balances.
8248 * If a device is returned, this also returns holding the spa config lock.
8250 static l2arc_dev_t *
8251 l2arc_dev_get_next(void)
8253 l2arc_dev_t *first, *next = NULL;
8256 * Lock out the removal of spas (spa_namespace_lock), then removal
8257 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
8258 * both locks will be dropped and a spa config lock held instead.
8260 mutex_enter(&spa_namespace_lock);
8261 mutex_enter(&l2arc_dev_mtx);
8263 /* if there are no vdevs, there is nothing to do */
8264 if (l2arc_ndev == 0)
8268 next = l2arc_dev_last;
8270 /* loop around the list looking for a non-faulted vdev */
8272 next = list_head(l2arc_dev_list);
8274 next = list_next(l2arc_dev_list, next);
8276 next = list_head(l2arc_dev_list);
8279 /* if we have come back to the start, bail out */
8282 else if (next == first)
8285 ASSERT3P(next, !=, NULL);
8286 } while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8287 next->l2ad_trim_all);
8289 /* if we were unable to find any usable vdevs, return NULL */
8290 if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8291 next->l2ad_trim_all)
8294 l2arc_dev_last = next;
8297 mutex_exit(&l2arc_dev_mtx);
8300 * Grab the config lock to prevent the 'next' device from being
8301 * removed while we are writing to it.
8304 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
8305 mutex_exit(&spa_namespace_lock);
8311 * Free buffers that were tagged for destruction.
8314 l2arc_do_free_on_write(void)
8316 l2arc_data_free_t *df;
8318 mutex_enter(&l2arc_free_on_write_mtx);
8319 while ((df = list_remove_head(l2arc_free_on_write)) != NULL) {
8320 ASSERT3P(df->l2df_abd, !=, NULL);
8321 abd_free(df->l2df_abd);
8322 kmem_free(df, sizeof (l2arc_data_free_t));
8324 mutex_exit(&l2arc_free_on_write_mtx);
8328 * A write to a cache device has completed. Update all headers to allow
8329 * reads from these buffers to begin.
8332 l2arc_write_done(zio_t *zio)
8334 l2arc_write_callback_t *cb;
8335 l2arc_lb_abd_buf_t *abd_buf;
8336 l2arc_lb_ptr_buf_t *lb_ptr_buf;
8338 l2arc_dev_hdr_phys_t *l2dhdr;
8340 arc_buf_hdr_t *head, *hdr, *hdr_prev;
8341 kmutex_t *hash_lock;
8342 int64_t bytes_dropped = 0;
8344 cb = zio->io_private;
8345 ASSERT3P(cb, !=, NULL);
8346 dev = cb->l2wcb_dev;
8347 l2dhdr = dev->l2ad_dev_hdr;
8348 ASSERT3P(dev, !=, NULL);
8349 head = cb->l2wcb_head;
8350 ASSERT3P(head, !=, NULL);
8351 buflist = &dev->l2ad_buflist;
8352 ASSERT3P(buflist, !=, NULL);
8353 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
8354 l2arc_write_callback_t *, cb);
8357 * All writes completed, or an error was hit.
8360 mutex_enter(&dev->l2ad_mtx);
8361 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
8362 hdr_prev = list_prev(buflist, hdr);
8364 hash_lock = HDR_LOCK(hdr);
8367 * We cannot use mutex_enter or else we can deadlock
8368 * with l2arc_write_buffers (due to swapping the order
8369 * the hash lock and l2ad_mtx are taken).
8371 if (!mutex_tryenter(hash_lock)) {
8373 * Missed the hash lock. We must retry so we
8374 * don't leave the ARC_FLAG_L2_WRITING bit set.
8376 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
8379 * We don't want to rescan the headers we've
8380 * already marked as having been written out, so
8381 * we reinsert the head node so we can pick up
8382 * where we left off.
8384 list_remove(buflist, head);
8385 list_insert_after(buflist, hdr, head);
8387 mutex_exit(&dev->l2ad_mtx);
8390 * We wait for the hash lock to become available
8391 * to try and prevent busy waiting, and increase
8392 * the chance we'll be able to acquire the lock
8393 * the next time around.
8395 mutex_enter(hash_lock);
8396 mutex_exit(hash_lock);
8401 * We could not have been moved into the arc_l2c_only
8402 * state while in-flight due to our ARC_FLAG_L2_WRITING
8403 * bit being set. Let's just ensure that's being enforced.
8405 ASSERT(HDR_HAS_L1HDR(hdr));
8408 * Skipped - drop L2ARC entry and mark the header as no
8409 * longer L2 eligibile.
8411 if (zio->io_error != 0) {
8413 * Error - drop L2ARC entry.
8415 list_remove(buflist, hdr);
8416 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
8418 uint64_t psize = HDR_GET_PSIZE(hdr);
8419 l2arc_hdr_arcstats_decrement(hdr);
8422 vdev_psize_to_asize(dev->l2ad_vdev, psize);
8423 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
8424 arc_hdr_size(hdr), hdr);
8428 * Allow ARC to begin reads and ghost list evictions to
8431 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
8433 mutex_exit(hash_lock);
8437 * Free the allocated abd buffers for writing the log blocks.
8438 * If the zio failed reclaim the allocated space and remove the
8439 * pointers to these log blocks from the log block pointer list
8440 * of the L2ARC device.
8442 while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) {
8443 abd_free(abd_buf->abd);
8444 zio_buf_free(abd_buf, sizeof (*abd_buf));
8445 if (zio->io_error != 0) {
8446 lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list);
8448 * L2BLK_GET_PSIZE returns aligned size for log
8452 L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop);
8453 bytes_dropped += asize;
8454 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8455 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8456 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8458 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8459 kmem_free(lb_ptr_buf->lb_ptr,
8460 sizeof (l2arc_log_blkptr_t));
8461 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8464 list_destroy(&cb->l2wcb_abd_list);
8466 if (zio->io_error != 0) {
8467 ARCSTAT_BUMP(arcstat_l2_writes_error);
8470 * Restore the lbps array in the header to its previous state.
8471 * If the list of log block pointers is empty, zero out the
8472 * log block pointers in the device header.
8474 lb_ptr_buf = list_head(&dev->l2ad_lbptr_list);
8475 for (int i = 0; i < 2; i++) {
8476 if (lb_ptr_buf == NULL) {
8478 * If the list is empty zero out the device
8479 * header. Otherwise zero out the second log
8480 * block pointer in the header.
8484 dev->l2ad_dev_hdr_asize);
8486 memset(&l2dhdr->dh_start_lbps[i], 0,
8487 sizeof (l2arc_log_blkptr_t));
8491 memcpy(&l2dhdr->dh_start_lbps[i], lb_ptr_buf->lb_ptr,
8492 sizeof (l2arc_log_blkptr_t));
8493 lb_ptr_buf = list_next(&dev->l2ad_lbptr_list,
8498 ARCSTAT_BUMP(arcstat_l2_writes_done);
8499 list_remove(buflist, head);
8500 ASSERT(!HDR_HAS_L1HDR(head));
8501 kmem_cache_free(hdr_l2only_cache, head);
8502 mutex_exit(&dev->l2ad_mtx);
8504 ASSERT(dev->l2ad_vdev != NULL);
8505 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
8507 l2arc_do_free_on_write();
8509 kmem_free(cb, sizeof (l2arc_write_callback_t));
8513 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
8516 spa_t *spa = zio->io_spa;
8517 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
8518 blkptr_t *bp = zio->io_bp;
8519 uint8_t salt[ZIO_DATA_SALT_LEN];
8520 uint8_t iv[ZIO_DATA_IV_LEN];
8521 uint8_t mac[ZIO_DATA_MAC_LEN];
8522 boolean_t no_crypt = B_FALSE;
8525 * ZIL data is never be written to the L2ARC, so we don't need
8526 * special handling for its unique MAC storage.
8528 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
8529 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
8530 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8533 * If the data was encrypted, decrypt it now. Note that
8534 * we must check the bp here and not the hdr, since the
8535 * hdr does not have its encryption parameters updated
8536 * until arc_read_done().
8538 if (BP_IS_ENCRYPTED(bp)) {
8539 abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8540 ARC_HDR_USE_RESERVE);
8542 zio_crypt_decode_params_bp(bp, salt, iv);
8543 zio_crypt_decode_mac_bp(bp, mac);
8545 ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
8546 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
8547 salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
8548 hdr->b_l1hdr.b_pabd, &no_crypt);
8550 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8555 * If we actually performed decryption, replace b_pabd
8556 * with the decrypted data. Otherwise we can just throw
8557 * our decryption buffer away.
8560 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8561 arc_hdr_size(hdr), hdr);
8562 hdr->b_l1hdr.b_pabd = eabd;
8565 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8570 * If the L2ARC block was compressed, but ARC compression
8571 * is disabled we decompress the data into a new buffer and
8572 * replace the existing data.
8574 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8575 !HDR_COMPRESSION_ENABLED(hdr)) {
8576 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8577 ARC_HDR_USE_RESERVE);
8578 void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
8580 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
8581 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
8582 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
8584 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8585 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
8589 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8590 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8591 arc_hdr_size(hdr), hdr);
8592 hdr->b_l1hdr.b_pabd = cabd;
8594 zio->io_size = HDR_GET_LSIZE(hdr);
8605 * A read to a cache device completed. Validate buffer contents before
8606 * handing over to the regular ARC routines.
8609 l2arc_read_done(zio_t *zio)
8612 l2arc_read_callback_t *cb = zio->io_private;
8614 kmutex_t *hash_lock;
8615 boolean_t valid_cksum;
8616 boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
8617 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
8619 ASSERT3P(zio->io_vd, !=, NULL);
8620 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
8622 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
8624 ASSERT3P(cb, !=, NULL);
8625 hdr = cb->l2rcb_hdr;
8626 ASSERT3P(hdr, !=, NULL);
8628 hash_lock = HDR_LOCK(hdr);
8629 mutex_enter(hash_lock);
8630 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
8633 * If the data was read into a temporary buffer,
8634 * move it and free the buffer.
8636 if (cb->l2rcb_abd != NULL) {
8637 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
8638 if (zio->io_error == 0) {
8640 abd_copy(hdr->b_crypt_hdr.b_rabd,
8641 cb->l2rcb_abd, arc_hdr_size(hdr));
8643 abd_copy(hdr->b_l1hdr.b_pabd,
8644 cb->l2rcb_abd, arc_hdr_size(hdr));
8649 * The following must be done regardless of whether
8650 * there was an error:
8651 * - free the temporary buffer
8652 * - point zio to the real ARC buffer
8653 * - set zio size accordingly
8654 * These are required because zio is either re-used for
8655 * an I/O of the block in the case of the error
8656 * or the zio is passed to arc_read_done() and it
8659 abd_free(cb->l2rcb_abd);
8660 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
8663 ASSERT(HDR_HAS_RABD(hdr));
8664 zio->io_abd = zio->io_orig_abd =
8665 hdr->b_crypt_hdr.b_rabd;
8667 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8668 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8672 ASSERT3P(zio->io_abd, !=, NULL);
8675 * Check this survived the L2ARC journey.
8677 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8678 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8679 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
8680 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
8681 zio->io_prop.zp_complevel = hdr->b_complevel;
8683 valid_cksum = arc_cksum_is_equal(hdr, zio);
8686 * b_rabd will always match the data as it exists on disk if it is
8687 * being used. Therefore if we are reading into b_rabd we do not
8688 * attempt to untransform the data.
8690 if (valid_cksum && !using_rdata)
8691 tfm_error = l2arc_untransform(zio, cb);
8693 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8694 !HDR_L2_EVICTED(hdr)) {
8695 mutex_exit(hash_lock);
8696 zio->io_private = hdr;
8700 * Buffer didn't survive caching. Increment stats and
8701 * reissue to the original storage device.
8703 if (zio->io_error != 0) {
8704 ARCSTAT_BUMP(arcstat_l2_io_error);
8706 zio->io_error = SET_ERROR(EIO);
8708 if (!valid_cksum || tfm_error != 0)
8709 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8712 * If there's no waiter, issue an async i/o to the primary
8713 * storage now. If there *is* a waiter, the caller must
8714 * issue the i/o in a context where it's OK to block.
8716 if (zio->io_waiter == NULL) {
8717 zio_t *pio = zio_unique_parent(zio);
8718 void *abd = (using_rdata) ?
8719 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8721 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8723 zio = zio_read(pio, zio->io_spa, zio->io_bp,
8724 abd, zio->io_size, arc_read_done,
8725 hdr, zio->io_priority, cb->l2rcb_flags,
8729 * Original ZIO will be freed, so we need to update
8730 * ARC header with the new ZIO pointer to be used
8731 * by zio_change_priority() in arc_read().
8733 for (struct arc_callback *acb = hdr->b_l1hdr.b_acb;
8734 acb != NULL; acb = acb->acb_next)
8735 acb->acb_zio_head = zio;
8737 mutex_exit(hash_lock);
8740 mutex_exit(hash_lock);
8744 kmem_free(cb, sizeof (l2arc_read_callback_t));
8748 * This is the list priority from which the L2ARC will search for pages to
8749 * cache. This is used within loops (0..3) to cycle through lists in the
8750 * desired order. This order can have a significant effect on cache
8753 * Currently the metadata lists are hit first, MFU then MRU, followed by
8754 * the data lists. This function returns a locked list, and also returns
8757 static multilist_sublist_t *
8758 l2arc_sublist_lock(int list_num)
8760 multilist_t *ml = NULL;
8763 ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
8767 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
8770 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
8773 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
8776 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
8783 * Return a randomly-selected sublist. This is acceptable
8784 * because the caller feeds only a little bit of data for each
8785 * call (8MB). Subsequent calls will result in different
8786 * sublists being selected.
8788 idx = multilist_get_random_index(ml);
8789 return (multilist_sublist_lock(ml, idx));
8793 * Calculates the maximum overhead of L2ARC metadata log blocks for a given
8794 * L2ARC write size. l2arc_evict and l2arc_write_size need to include this
8795 * overhead in processing to make sure there is enough headroom available
8796 * when writing buffers.
8798 static inline uint64_t
8799 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev)
8801 if (dev->l2ad_log_entries == 0) {
8804 uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT;
8806 uint64_t log_blocks = (log_entries +
8807 dev->l2ad_log_entries - 1) /
8808 dev->l2ad_log_entries;
8810 return (vdev_psize_to_asize(dev->l2ad_vdev,
8811 sizeof (l2arc_log_blk_phys_t)) * log_blocks);
8816 * Evict buffers from the device write hand to the distance specified in
8817 * bytes. This distance may span populated buffers, it may span nothing.
8818 * This is clearing a region on the L2ARC device ready for writing.
8819 * If the 'all' boolean is set, every buffer is evicted.
8822 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
8825 arc_buf_hdr_t *hdr, *hdr_prev;
8826 kmutex_t *hash_lock;
8828 l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev;
8829 vdev_t *vd = dev->l2ad_vdev;
8832 buflist = &dev->l2ad_buflist;
8836 if (dev->l2ad_hand + distance > dev->l2ad_end) {
8838 * When there is no space to accommodate upcoming writes,
8839 * evict to the end. Then bump the write and evict hands
8840 * to the start and iterate. This iteration does not
8841 * happen indefinitely as we make sure in
8842 * l2arc_write_size() that when the write hand is reset,
8843 * the write size does not exceed the end of the device.
8846 taddr = dev->l2ad_end;
8848 taddr = dev->l2ad_hand + distance;
8850 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
8851 uint64_t, taddr, boolean_t, all);
8855 * This check has to be placed after deciding whether to
8858 if (dev->l2ad_first) {
8860 * This is the first sweep through the device. There is
8861 * nothing to evict. We have already trimmmed the
8867 * Trim the space to be evicted.
8869 if (vd->vdev_has_trim && dev->l2ad_evict < taddr &&
8870 l2arc_trim_ahead > 0) {
8872 * We have to drop the spa_config lock because
8873 * vdev_trim_range() will acquire it.
8874 * l2ad_evict already accounts for the label
8875 * size. To prevent vdev_trim_ranges() from
8876 * adding it again, we subtract it from
8879 spa_config_exit(dev->l2ad_spa, SCL_L2ARC, dev);
8880 vdev_trim_simple(vd,
8881 dev->l2ad_evict - VDEV_LABEL_START_SIZE,
8882 taddr - dev->l2ad_evict);
8883 spa_config_enter(dev->l2ad_spa, SCL_L2ARC, dev,
8888 * When rebuilding L2ARC we retrieve the evict hand
8889 * from the header of the device. Of note, l2arc_evict()
8890 * does not actually delete buffers from the cache
8891 * device, but trimming may do so depending on the
8892 * hardware implementation. Thus keeping track of the
8893 * evict hand is useful.
8895 dev->l2ad_evict = MAX(dev->l2ad_evict, taddr);
8900 mutex_enter(&dev->l2ad_mtx);
8902 * We have to account for evicted log blocks. Run vdev_space_update()
8903 * on log blocks whose offset (in bytes) is before the evicted offset
8904 * (in bytes) by searching in the list of pointers to log blocks
8905 * present in the L2ARC device.
8907 for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf;
8908 lb_ptr_buf = lb_ptr_buf_prev) {
8910 lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf);
8912 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
8913 uint64_t asize = L2BLK_GET_PSIZE(
8914 (lb_ptr_buf->lb_ptr)->lbp_prop);
8917 * We don't worry about log blocks left behind (ie
8918 * lbp_payload_start < l2ad_hand) because l2arc_write_buffers()
8919 * will never write more than l2arc_evict() evicts.
8921 if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) {
8924 vdev_space_update(vd, -asize, 0, 0);
8925 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8926 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8927 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8929 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8930 list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf);
8931 kmem_free(lb_ptr_buf->lb_ptr,
8932 sizeof (l2arc_log_blkptr_t));
8933 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8937 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
8938 hdr_prev = list_prev(buflist, hdr);
8940 ASSERT(!HDR_EMPTY(hdr));
8941 hash_lock = HDR_LOCK(hdr);
8944 * We cannot use mutex_enter or else we can deadlock
8945 * with l2arc_write_buffers (due to swapping the order
8946 * the hash lock and l2ad_mtx are taken).
8948 if (!mutex_tryenter(hash_lock)) {
8950 * Missed the hash lock. Retry.
8952 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
8953 mutex_exit(&dev->l2ad_mtx);
8954 mutex_enter(hash_lock);
8955 mutex_exit(hash_lock);
8960 * A header can't be on this list if it doesn't have L2 header.
8962 ASSERT(HDR_HAS_L2HDR(hdr));
8964 /* Ensure this header has finished being written. */
8965 ASSERT(!HDR_L2_WRITING(hdr));
8966 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
8968 if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict ||
8969 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
8971 * We've evicted to the target address,
8972 * or the end of the device.
8974 mutex_exit(hash_lock);
8978 if (!HDR_HAS_L1HDR(hdr)) {
8979 ASSERT(!HDR_L2_READING(hdr));
8981 * This doesn't exist in the ARC. Destroy.
8982 * arc_hdr_destroy() will call list_remove()
8983 * and decrement arcstat_l2_lsize.
8985 arc_change_state(arc_anon, hdr);
8986 arc_hdr_destroy(hdr);
8988 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
8989 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
8991 * Invalidate issued or about to be issued
8992 * reads, since we may be about to write
8993 * over this location.
8995 if (HDR_L2_READING(hdr)) {
8996 ARCSTAT_BUMP(arcstat_l2_evict_reading);
8997 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
9000 arc_hdr_l2hdr_destroy(hdr);
9002 mutex_exit(hash_lock);
9004 mutex_exit(&dev->l2ad_mtx);
9008 * We need to check if we evict all buffers, otherwise we may iterate
9011 if (!all && rerun) {
9013 * Bump device hand to the device start if it is approaching the
9014 * end. l2arc_evict() has already evicted ahead for this case.
9016 dev->l2ad_hand = dev->l2ad_start;
9017 dev->l2ad_evict = dev->l2ad_start;
9018 dev->l2ad_first = B_FALSE;
9024 * In case of cache device removal (all) the following
9025 * assertions may be violated without functional consequences
9026 * as the device is about to be removed.
9028 ASSERT3U(dev->l2ad_hand + distance, <, dev->l2ad_end);
9029 if (!dev->l2ad_first)
9030 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
9035 * Handle any abd transforms that might be required for writing to the L2ARC.
9036 * If successful, this function will always return an abd with the data
9037 * transformed as it is on disk in a new abd of asize bytes.
9040 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
9045 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
9046 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
9047 uint64_t psize = HDR_GET_PSIZE(hdr);
9048 uint64_t size = arc_hdr_size(hdr);
9049 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
9050 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
9051 dsl_crypto_key_t *dck = NULL;
9052 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
9053 boolean_t no_crypt = B_FALSE;
9055 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
9056 !HDR_COMPRESSION_ENABLED(hdr)) ||
9057 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
9058 ASSERT3U(psize, <=, asize);
9061 * If this data simply needs its own buffer, we simply allocate it
9062 * and copy the data. This may be done to eliminate a dependency on a
9063 * shared buffer or to reallocate the buffer to match asize.
9065 if (HDR_HAS_RABD(hdr) && asize != psize) {
9066 ASSERT3U(asize, >=, psize);
9067 to_write = abd_alloc_for_io(asize, ismd);
9068 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
9070 abd_zero_off(to_write, psize, asize - psize);
9074 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
9075 !HDR_ENCRYPTED(hdr)) {
9076 ASSERT3U(size, ==, psize);
9077 to_write = abd_alloc_for_io(asize, ismd);
9078 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
9080 abd_zero_off(to_write, size, asize - size);
9084 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
9086 * In some cases, we can wind up with size > asize, so
9087 * we need to opt for the larger allocation option here.
9089 * (We also need abd_return_buf_copy in all cases because
9090 * it's an ASSERT() to modify the buffer before returning it
9091 * with arc_return_buf(), and all the compressors
9092 * write things before deciding to fail compression in nearly
9095 cabd = abd_alloc_for_io(size, ismd);
9096 tmp = abd_borrow_buf(cabd, size);
9098 psize = zio_compress_data(compress, to_write, &tmp, size,
9101 if (psize >= asize) {
9102 psize = HDR_GET_PSIZE(hdr);
9103 abd_return_buf_copy(cabd, tmp, size);
9104 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
9106 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
9108 abd_zero_off(to_write, psize, asize - psize);
9111 ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
9113 memset((char *)tmp + psize, 0, asize - psize);
9114 psize = HDR_GET_PSIZE(hdr);
9115 abd_return_buf_copy(cabd, tmp, size);
9120 if (HDR_ENCRYPTED(hdr)) {
9121 eabd = abd_alloc_for_io(asize, ismd);
9124 * If the dataset was disowned before the buffer
9125 * made it to this point, the key to re-encrypt
9126 * it won't be available. In this case we simply
9127 * won't write the buffer to the L2ARC.
9129 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
9134 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
9135 hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
9136 hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
9142 abd_copy(eabd, to_write, psize);
9145 abd_zero_off(eabd, psize, asize - psize);
9147 /* assert that the MAC we got here matches the one we saved */
9148 ASSERT0(memcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
9149 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9151 if (to_write == cabd)
9158 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
9159 *abd_out = to_write;
9164 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9175 l2arc_blk_fetch_done(zio_t *zio)
9177 l2arc_read_callback_t *cb;
9179 cb = zio->io_private;
9180 if (cb->l2rcb_abd != NULL)
9181 abd_free(cb->l2rcb_abd);
9182 kmem_free(cb, sizeof (l2arc_read_callback_t));
9186 * Find and write ARC buffers to the L2ARC device.
9188 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
9189 * for reading until they have completed writing.
9190 * The headroom_boost is an in-out parameter used to maintain headroom boost
9191 * state between calls to this function.
9193 * Returns the number of bytes actually written (which may be smaller than
9194 * the delta by which the device hand has changed due to alignment and the
9195 * writing of log blocks).
9198 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
9200 arc_buf_hdr_t *hdr, *hdr_prev, *head;
9201 uint64_t write_asize, write_psize, write_lsize, headroom;
9203 l2arc_write_callback_t *cb = NULL;
9205 uint64_t guid = spa_load_guid(spa);
9206 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9208 ASSERT3P(dev->l2ad_vdev, !=, NULL);
9211 write_lsize = write_asize = write_psize = 0;
9213 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
9214 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
9217 * Copy buffers for L2ARC writing.
9219 for (int pass = 0; pass < L2ARC_FEED_TYPES; pass++) {
9221 * If pass == 1 or 3, we cache MRU metadata and data
9224 if (l2arc_mfuonly) {
9225 if (pass == 1 || pass == 3)
9229 multilist_sublist_t *mls = l2arc_sublist_lock(pass);
9230 uint64_t passed_sz = 0;
9232 VERIFY3P(mls, !=, NULL);
9235 * L2ARC fast warmup.
9237 * Until the ARC is warm and starts to evict, read from the
9238 * head of the ARC lists rather than the tail.
9240 if (arc_warm == B_FALSE)
9241 hdr = multilist_sublist_head(mls);
9243 hdr = multilist_sublist_tail(mls);
9245 headroom = target_sz * l2arc_headroom;
9246 if (zfs_compressed_arc_enabled)
9247 headroom = (headroom * l2arc_headroom_boost) / 100;
9249 for (; hdr; hdr = hdr_prev) {
9250 kmutex_t *hash_lock;
9251 abd_t *to_write = NULL;
9253 if (arc_warm == B_FALSE)
9254 hdr_prev = multilist_sublist_next(mls, hdr);
9256 hdr_prev = multilist_sublist_prev(mls, hdr);
9258 hash_lock = HDR_LOCK(hdr);
9259 if (!mutex_tryenter(hash_lock)) {
9261 * Skip this buffer rather than waiting.
9266 passed_sz += HDR_GET_LSIZE(hdr);
9267 if (l2arc_headroom != 0 && passed_sz > headroom) {
9271 mutex_exit(hash_lock);
9275 if (!l2arc_write_eligible(guid, hdr)) {
9276 mutex_exit(hash_lock);
9280 ASSERT(HDR_HAS_L1HDR(hdr));
9282 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
9283 ASSERT3U(arc_hdr_size(hdr), >, 0);
9284 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
9286 uint64_t psize = HDR_GET_PSIZE(hdr);
9287 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
9291 * If the allocated size of this buffer plus the max
9292 * size for the pending log block exceeds the evicted
9293 * target size, terminate writing buffers for this run.
9295 if (write_asize + asize +
9296 sizeof (l2arc_log_blk_phys_t) > target_sz) {
9298 mutex_exit(hash_lock);
9303 * We rely on the L1 portion of the header below, so
9304 * it's invalid for this header to have been evicted out
9305 * of the ghost cache, prior to being written out. The
9306 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
9308 arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
9311 * If this header has b_rabd, we can use this since it
9312 * must always match the data exactly as it exists on
9313 * disk. Otherwise, the L2ARC can normally use the
9314 * hdr's data, but if we're sharing data between the
9315 * hdr and one of its bufs, L2ARC needs its own copy of
9316 * the data so that the ZIO below can't race with the
9317 * buf consumer. To ensure that this copy will be
9318 * available for the lifetime of the ZIO and be cleaned
9319 * up afterwards, we add it to the l2arc_free_on_write
9320 * queue. If we need to apply any transforms to the
9321 * data (compression, encryption) we will also need the
9324 if (HDR_HAS_RABD(hdr) && psize == asize) {
9325 to_write = hdr->b_crypt_hdr.b_rabd;
9326 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
9327 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
9328 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
9330 to_write = hdr->b_l1hdr.b_pabd;
9333 arc_buf_contents_t type = arc_buf_type(hdr);
9335 ret = l2arc_apply_transforms(spa, hdr, asize,
9338 arc_hdr_clear_flags(hdr,
9339 ARC_FLAG_L2_WRITING);
9340 mutex_exit(hash_lock);
9344 l2arc_free_abd_on_write(to_write, asize, type);
9349 * Insert a dummy header on the buflist so
9350 * l2arc_write_done() can find where the
9351 * write buffers begin without searching.
9353 mutex_enter(&dev->l2ad_mtx);
9354 list_insert_head(&dev->l2ad_buflist, head);
9355 mutex_exit(&dev->l2ad_mtx);
9358 sizeof (l2arc_write_callback_t), KM_SLEEP);
9359 cb->l2wcb_dev = dev;
9360 cb->l2wcb_head = head;
9362 * Create a list to save allocated abd buffers
9363 * for l2arc_log_blk_commit().
9365 list_create(&cb->l2wcb_abd_list,
9366 sizeof (l2arc_lb_abd_buf_t),
9367 offsetof(l2arc_lb_abd_buf_t, node));
9368 pio = zio_root(spa, l2arc_write_done, cb,
9372 hdr->b_l2hdr.b_dev = dev;
9373 hdr->b_l2hdr.b_hits = 0;
9375 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
9376 hdr->b_l2hdr.b_arcs_state =
9377 hdr->b_l1hdr.b_state->arcs_state;
9378 arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR);
9380 mutex_enter(&dev->l2ad_mtx);
9381 list_insert_head(&dev->l2ad_buflist, hdr);
9382 mutex_exit(&dev->l2ad_mtx);
9384 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
9385 arc_hdr_size(hdr), hdr);
9387 wzio = zio_write_phys(pio, dev->l2ad_vdev,
9388 hdr->b_l2hdr.b_daddr, asize, to_write,
9389 ZIO_CHECKSUM_OFF, NULL, hdr,
9390 ZIO_PRIORITY_ASYNC_WRITE,
9391 ZIO_FLAG_CANFAIL, B_FALSE);
9393 write_lsize += HDR_GET_LSIZE(hdr);
9394 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
9397 write_psize += psize;
9398 write_asize += asize;
9399 dev->l2ad_hand += asize;
9400 l2arc_hdr_arcstats_increment(hdr);
9401 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9403 mutex_exit(hash_lock);
9406 * Append buf info to current log and commit if full.
9407 * arcstat_l2_{size,asize} kstats are updated
9410 if (l2arc_log_blk_insert(dev, hdr)) {
9412 * l2ad_hand will be adjusted in
9413 * l2arc_log_blk_commit().
9416 l2arc_log_blk_commit(dev, pio, cb);
9422 multilist_sublist_unlock(mls);
9428 /* No buffers selected for writing? */
9430 ASSERT0(write_lsize);
9431 ASSERT(!HDR_HAS_L1HDR(head));
9432 kmem_cache_free(hdr_l2only_cache, head);
9435 * Although we did not write any buffers l2ad_evict may
9438 if (dev->l2ad_evict != l2dhdr->dh_evict)
9439 l2arc_dev_hdr_update(dev);
9444 if (!dev->l2ad_first)
9445 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
9447 ASSERT3U(write_asize, <=, target_sz);
9448 ARCSTAT_BUMP(arcstat_l2_writes_sent);
9449 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
9451 dev->l2ad_writing = B_TRUE;
9452 (void) zio_wait(pio);
9453 dev->l2ad_writing = B_FALSE;
9456 * Update the device header after the zio completes as
9457 * l2arc_write_done() may have updated the memory holding the log block
9458 * pointers in the device header.
9460 l2arc_dev_hdr_update(dev);
9462 return (write_asize);
9466 l2arc_hdr_limit_reached(void)
9468 int64_t s = aggsum_upper_bound(&arc_sums.arcstat_l2_hdr_size);
9470 return (arc_reclaim_needed() ||
9471 (s > (arc_warm ? arc_c : arc_c_max) * l2arc_meta_percent / 100));
9475 * This thread feeds the L2ARC at regular intervals. This is the beating
9476 * heart of the L2ARC.
9478 static __attribute__((noreturn)) void
9479 l2arc_feed_thread(void *unused)
9485 uint64_t size, wrote;
9486 clock_t begin, next = ddi_get_lbolt();
9487 fstrans_cookie_t cookie;
9489 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
9491 mutex_enter(&l2arc_feed_thr_lock);
9493 cookie = spl_fstrans_mark();
9494 while (l2arc_thread_exit == 0) {
9495 CALLB_CPR_SAFE_BEGIN(&cpr);
9496 (void) cv_timedwait_idle(&l2arc_feed_thr_cv,
9497 &l2arc_feed_thr_lock, next);
9498 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
9499 next = ddi_get_lbolt() + hz;
9502 * Quick check for L2ARC devices.
9504 mutex_enter(&l2arc_dev_mtx);
9505 if (l2arc_ndev == 0) {
9506 mutex_exit(&l2arc_dev_mtx);
9509 mutex_exit(&l2arc_dev_mtx);
9510 begin = ddi_get_lbolt();
9513 * This selects the next l2arc device to write to, and in
9514 * doing so the next spa to feed from: dev->l2ad_spa. This
9515 * will return NULL if there are now no l2arc devices or if
9516 * they are all faulted.
9518 * If a device is returned, its spa's config lock is also
9519 * held to prevent device removal. l2arc_dev_get_next()
9520 * will grab and release l2arc_dev_mtx.
9522 if ((dev = l2arc_dev_get_next()) == NULL)
9525 spa = dev->l2ad_spa;
9526 ASSERT3P(spa, !=, NULL);
9529 * If the pool is read-only then force the feed thread to
9530 * sleep a little longer.
9532 if (!spa_writeable(spa)) {
9533 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
9534 spa_config_exit(spa, SCL_L2ARC, dev);
9539 * Avoid contributing to memory pressure.
9541 if (l2arc_hdr_limit_reached()) {
9542 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
9543 spa_config_exit(spa, SCL_L2ARC, dev);
9547 ARCSTAT_BUMP(arcstat_l2_feeds);
9549 size = l2arc_write_size(dev);
9552 * Evict L2ARC buffers that will be overwritten.
9554 l2arc_evict(dev, size, B_FALSE);
9557 * Write ARC buffers.
9559 wrote = l2arc_write_buffers(spa, dev, size);
9562 * Calculate interval between writes.
9564 next = l2arc_write_interval(begin, size, wrote);
9565 spa_config_exit(spa, SCL_L2ARC, dev);
9567 spl_fstrans_unmark(cookie);
9569 l2arc_thread_exit = 0;
9570 cv_broadcast(&l2arc_feed_thr_cv);
9571 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
9576 l2arc_vdev_present(vdev_t *vd)
9578 return (l2arc_vdev_get(vd) != NULL);
9582 * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if
9583 * the vdev_t isn't an L2ARC device.
9586 l2arc_vdev_get(vdev_t *vd)
9590 mutex_enter(&l2arc_dev_mtx);
9591 for (dev = list_head(l2arc_dev_list); dev != NULL;
9592 dev = list_next(l2arc_dev_list, dev)) {
9593 if (dev->l2ad_vdev == vd)
9596 mutex_exit(&l2arc_dev_mtx);
9602 l2arc_rebuild_dev(l2arc_dev_t *dev, boolean_t reopen)
9604 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9605 uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9606 spa_t *spa = dev->l2ad_spa;
9609 * The L2ARC has to hold at least the payload of one log block for
9610 * them to be restored (persistent L2ARC). The payload of a log block
9611 * depends on the amount of its log entries. We always write log blocks
9612 * with 1022 entries. How many of them are committed or restored depends
9613 * on the size of the L2ARC device. Thus the maximum payload of
9614 * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device
9615 * is less than that, we reduce the amount of committed and restored
9616 * log entries per block so as to enable persistence.
9618 if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) {
9619 dev->l2ad_log_entries = 0;
9621 dev->l2ad_log_entries = MIN((dev->l2ad_end -
9622 dev->l2ad_start) >> SPA_MAXBLOCKSHIFT,
9623 L2ARC_LOG_BLK_MAX_ENTRIES);
9627 * Read the device header, if an error is returned do not rebuild L2ARC.
9629 if (l2arc_dev_hdr_read(dev) == 0 && dev->l2ad_log_entries > 0) {
9631 * If we are onlining a cache device (vdev_reopen) that was
9632 * still present (l2arc_vdev_present()) and rebuild is enabled,
9633 * we should evict all ARC buffers and pointers to log blocks
9634 * and reclaim their space before restoring its contents to
9638 if (!l2arc_rebuild_enabled) {
9641 l2arc_evict(dev, 0, B_TRUE);
9642 /* start a new log block */
9643 dev->l2ad_log_ent_idx = 0;
9644 dev->l2ad_log_blk_payload_asize = 0;
9645 dev->l2ad_log_blk_payload_start = 0;
9649 * Just mark the device as pending for a rebuild. We won't
9650 * be starting a rebuild in line here as it would block pool
9651 * import. Instead spa_load_impl will hand that off to an
9652 * async task which will call l2arc_spa_rebuild_start.
9654 dev->l2ad_rebuild = B_TRUE;
9655 } else if (spa_writeable(spa)) {
9657 * In this case TRIM the whole device if l2arc_trim_ahead > 0,
9658 * otherwise create a new header. We zero out the memory holding
9659 * the header to reset dh_start_lbps. If we TRIM the whole
9660 * device the new header will be written by
9661 * vdev_trim_l2arc_thread() at the end of the TRIM to update the
9662 * trim_state in the header too. When reading the header, if
9663 * trim_state is not VDEV_TRIM_COMPLETE and l2arc_trim_ahead > 0
9664 * we opt to TRIM the whole device again.
9666 if (l2arc_trim_ahead > 0) {
9667 dev->l2ad_trim_all = B_TRUE;
9669 memset(l2dhdr, 0, l2dhdr_asize);
9670 l2arc_dev_hdr_update(dev);
9676 * Add a vdev for use by the L2ARC. By this point the spa has already
9677 * validated the vdev and opened it.
9680 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
9682 l2arc_dev_t *adddev;
9683 uint64_t l2dhdr_asize;
9685 ASSERT(!l2arc_vdev_present(vd));
9688 * Create a new l2arc device entry.
9690 adddev = vmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
9691 adddev->l2ad_spa = spa;
9692 adddev->l2ad_vdev = vd;
9693 /* leave extra size for an l2arc device header */
9694 l2dhdr_asize = adddev->l2ad_dev_hdr_asize =
9695 MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift);
9696 adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize;
9697 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
9698 ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end);
9699 adddev->l2ad_hand = adddev->l2ad_start;
9700 adddev->l2ad_evict = adddev->l2ad_start;
9701 adddev->l2ad_first = B_TRUE;
9702 adddev->l2ad_writing = B_FALSE;
9703 adddev->l2ad_trim_all = B_FALSE;
9704 list_link_init(&adddev->l2ad_node);
9705 adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP);
9707 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
9709 * This is a list of all ARC buffers that are still valid on the
9712 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
9713 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
9716 * This is a list of pointers to log blocks that are still present
9719 list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t),
9720 offsetof(l2arc_lb_ptr_buf_t, node));
9722 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
9723 zfs_refcount_create(&adddev->l2ad_alloc);
9724 zfs_refcount_create(&adddev->l2ad_lb_asize);
9725 zfs_refcount_create(&adddev->l2ad_lb_count);
9728 * Decide if dev is eligible for L2ARC rebuild or whole device
9729 * trimming. This has to happen before the device is added in the
9730 * cache device list and l2arc_dev_mtx is released. Otherwise
9731 * l2arc_feed_thread() might already start writing on the
9734 l2arc_rebuild_dev(adddev, B_FALSE);
9737 * Add device to global list
9739 mutex_enter(&l2arc_dev_mtx);
9740 list_insert_head(l2arc_dev_list, adddev);
9741 atomic_inc_64(&l2arc_ndev);
9742 mutex_exit(&l2arc_dev_mtx);
9746 * Decide if a vdev is eligible for L2ARC rebuild, called from vdev_reopen()
9747 * in case of onlining a cache device.
9750 l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen)
9752 l2arc_dev_t *dev = NULL;
9754 dev = l2arc_vdev_get(vd);
9755 ASSERT3P(dev, !=, NULL);
9758 * In contrast to l2arc_add_vdev() we do not have to worry about
9759 * l2arc_feed_thread() invalidating previous content when onlining a
9760 * cache device. The device parameters (l2ad*) are not cleared when
9761 * offlining the device and writing new buffers will not invalidate
9762 * all previous content. In worst case only buffers that have not had
9763 * their log block written to the device will be lost.
9764 * When onlining the cache device (ie offline->online without exporting
9765 * the pool in between) this happens:
9766 * vdev_reopen() -> vdev_open() -> l2arc_rebuild_vdev()
9768 * vdev_is_dead() = B_FALSE l2ad_rebuild = B_TRUE
9769 * During the time where vdev_is_dead = B_FALSE and until l2ad_rebuild
9770 * is set to B_TRUE we might write additional buffers to the device.
9772 l2arc_rebuild_dev(dev, reopen);
9776 * Remove a vdev from the L2ARC.
9779 l2arc_remove_vdev(vdev_t *vd)
9781 l2arc_dev_t *remdev = NULL;
9784 * Find the device by vdev
9786 remdev = l2arc_vdev_get(vd);
9787 ASSERT3P(remdev, !=, NULL);
9790 * Cancel any ongoing or scheduled rebuild.
9792 mutex_enter(&l2arc_rebuild_thr_lock);
9793 if (remdev->l2ad_rebuild_began == B_TRUE) {
9794 remdev->l2ad_rebuild_cancel = B_TRUE;
9795 while (remdev->l2ad_rebuild == B_TRUE)
9796 cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock);
9798 mutex_exit(&l2arc_rebuild_thr_lock);
9801 * Remove device from global list
9803 mutex_enter(&l2arc_dev_mtx);
9804 list_remove(l2arc_dev_list, remdev);
9805 l2arc_dev_last = NULL; /* may have been invalidated */
9806 atomic_dec_64(&l2arc_ndev);
9807 mutex_exit(&l2arc_dev_mtx);
9810 * Clear all buflists and ARC references. L2ARC device flush.
9812 l2arc_evict(remdev, 0, B_TRUE);
9813 list_destroy(&remdev->l2ad_buflist);
9814 ASSERT(list_is_empty(&remdev->l2ad_lbptr_list));
9815 list_destroy(&remdev->l2ad_lbptr_list);
9816 mutex_destroy(&remdev->l2ad_mtx);
9817 zfs_refcount_destroy(&remdev->l2ad_alloc);
9818 zfs_refcount_destroy(&remdev->l2ad_lb_asize);
9819 zfs_refcount_destroy(&remdev->l2ad_lb_count);
9820 kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize);
9821 vmem_free(remdev, sizeof (l2arc_dev_t));
9827 l2arc_thread_exit = 0;
9830 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9831 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
9832 mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9833 cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL);
9834 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
9835 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
9837 l2arc_dev_list = &L2ARC_dev_list;
9838 l2arc_free_on_write = &L2ARC_free_on_write;
9839 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
9840 offsetof(l2arc_dev_t, l2ad_node));
9841 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
9842 offsetof(l2arc_data_free_t, l2df_list_node));
9848 mutex_destroy(&l2arc_feed_thr_lock);
9849 cv_destroy(&l2arc_feed_thr_cv);
9850 mutex_destroy(&l2arc_rebuild_thr_lock);
9851 cv_destroy(&l2arc_rebuild_thr_cv);
9852 mutex_destroy(&l2arc_dev_mtx);
9853 mutex_destroy(&l2arc_free_on_write_mtx);
9855 list_destroy(l2arc_dev_list);
9856 list_destroy(l2arc_free_on_write);
9862 if (!(spa_mode_global & SPA_MODE_WRITE))
9865 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
9866 TS_RUN, defclsyspri);
9872 if (!(spa_mode_global & SPA_MODE_WRITE))
9875 mutex_enter(&l2arc_feed_thr_lock);
9876 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
9877 l2arc_thread_exit = 1;
9878 while (l2arc_thread_exit != 0)
9879 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
9880 mutex_exit(&l2arc_feed_thr_lock);
9884 * Punches out rebuild threads for the L2ARC devices in a spa. This should
9885 * be called after pool import from the spa async thread, since starting
9886 * these threads directly from spa_import() will make them part of the
9887 * "zpool import" context and delay process exit (and thus pool import).
9890 l2arc_spa_rebuild_start(spa_t *spa)
9892 ASSERT(MUTEX_HELD(&spa_namespace_lock));
9895 * Locate the spa's l2arc devices and kick off rebuild threads.
9897 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
9899 l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]);
9901 /* Don't attempt a rebuild if the vdev is UNAVAIL */
9904 mutex_enter(&l2arc_rebuild_thr_lock);
9905 if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) {
9906 dev->l2ad_rebuild_began = B_TRUE;
9907 (void) thread_create(NULL, 0, l2arc_dev_rebuild_thread,
9908 dev, 0, &p0, TS_RUN, minclsyspri);
9910 mutex_exit(&l2arc_rebuild_thr_lock);
9915 * Main entry point for L2ARC rebuilding.
9917 static __attribute__((noreturn)) void
9918 l2arc_dev_rebuild_thread(void *arg)
9920 l2arc_dev_t *dev = arg;
9922 VERIFY(!dev->l2ad_rebuild_cancel);
9923 VERIFY(dev->l2ad_rebuild);
9924 (void) l2arc_rebuild(dev);
9925 mutex_enter(&l2arc_rebuild_thr_lock);
9926 dev->l2ad_rebuild_began = B_FALSE;
9927 dev->l2ad_rebuild = B_FALSE;
9928 mutex_exit(&l2arc_rebuild_thr_lock);
9934 * This function implements the actual L2ARC metadata rebuild. It:
9935 * starts reading the log block chain and restores each block's contents
9936 * to memory (reconstructing arc_buf_hdr_t's).
9938 * Operation stops under any of the following conditions:
9940 * 1) We reach the end of the log block chain.
9941 * 2) We encounter *any* error condition (cksum errors, io errors)
9944 l2arc_rebuild(l2arc_dev_t *dev)
9946 vdev_t *vd = dev->l2ad_vdev;
9947 spa_t *spa = vd->vdev_spa;
9949 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9950 l2arc_log_blk_phys_t *this_lb, *next_lb;
9951 zio_t *this_io = NULL, *next_io = NULL;
9952 l2arc_log_blkptr_t lbps[2];
9953 l2arc_lb_ptr_buf_t *lb_ptr_buf;
9954 boolean_t lock_held;
9956 this_lb = vmem_zalloc(sizeof (*this_lb), KM_SLEEP);
9957 next_lb = vmem_zalloc(sizeof (*next_lb), KM_SLEEP);
9960 * We prevent device removal while issuing reads to the device,
9961 * then during the rebuilding phases we drop this lock again so
9962 * that a spa_unload or device remove can be initiated - this is
9963 * safe, because the spa will signal us to stop before removing
9964 * our device and wait for us to stop.
9966 spa_config_enter(spa, SCL_L2ARC, vd, RW_READER);
9970 * Retrieve the persistent L2ARC device state.
9971 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9973 dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start);
9974 dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr +
9975 L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop),
9977 dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST);
9979 vd->vdev_trim_action_time = l2dhdr->dh_trim_action_time;
9980 vd->vdev_trim_state = l2dhdr->dh_trim_state;
9983 * In case the zfs module parameter l2arc_rebuild_enabled is false
9984 * we do not start the rebuild process.
9986 if (!l2arc_rebuild_enabled)
9989 /* Prepare the rebuild process */
9990 memcpy(lbps, l2dhdr->dh_start_lbps, sizeof (lbps));
9992 /* Start the rebuild process */
9994 if (!l2arc_log_blkptr_valid(dev, &lbps[0]))
9997 if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1],
9998 this_lb, next_lb, this_io, &next_io)) != 0)
10002 * Our memory pressure valve. If the system is running low
10003 * on memory, rather than swamping memory with new ARC buf
10004 * hdrs, we opt not to rebuild the L2ARC. At this point,
10005 * however, we have already set up our L2ARC dev to chain in
10006 * new metadata log blocks, so the user may choose to offline/
10007 * online the L2ARC dev at a later time (or re-import the pool)
10008 * to reconstruct it (when there's less memory pressure).
10010 if (l2arc_hdr_limit_reached()) {
10011 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem);
10012 cmn_err(CE_NOTE, "System running low on memory, "
10013 "aborting L2ARC rebuild.");
10014 err = SET_ERROR(ENOMEM);
10018 spa_config_exit(spa, SCL_L2ARC, vd);
10019 lock_held = B_FALSE;
10022 * Now that we know that the next_lb checks out alright, we
10023 * can start reconstruction from this log block.
10024 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10026 uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop);
10027 l2arc_log_blk_restore(dev, this_lb, asize);
10030 * log block restored, include its pointer in the list of
10031 * pointers to log blocks present in the L2ARC device.
10033 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
10034 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t),
10036 memcpy(lb_ptr_buf->lb_ptr, &lbps[0],
10037 sizeof (l2arc_log_blkptr_t));
10038 mutex_enter(&dev->l2ad_mtx);
10039 list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf);
10040 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10041 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10042 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10043 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10044 mutex_exit(&dev->l2ad_mtx);
10045 vdev_space_update(vd, asize, 0, 0);
10048 * Protection against loops of log blocks:
10050 * l2ad_hand l2ad_evict
10052 * l2ad_start |=======================================| l2ad_end
10053 * -----|||----|||---|||----|||
10055 * ---|||---|||----|||---|||
10058 * In this situation the pointer of log block (4) passes
10059 * l2arc_log_blkptr_valid() but the log block should not be
10060 * restored as it is overwritten by the payload of log block
10061 * (0). Only log blocks (0)-(3) should be restored. We check
10062 * whether l2ad_evict lies in between the payload starting
10063 * offset of the next log block (lbps[1].lbp_payload_start)
10064 * and the payload starting offset of the present log block
10065 * (lbps[0].lbp_payload_start). If true and this isn't the
10066 * first pass, we are looping from the beginning and we should
10069 if (l2arc_range_check_overlap(lbps[1].lbp_payload_start,
10070 lbps[0].lbp_payload_start, dev->l2ad_evict) &&
10074 kpreempt(KPREEMPT_SYNC);
10076 mutex_enter(&l2arc_rebuild_thr_lock);
10077 if (dev->l2ad_rebuild_cancel) {
10078 dev->l2ad_rebuild = B_FALSE;
10079 cv_signal(&l2arc_rebuild_thr_cv);
10080 mutex_exit(&l2arc_rebuild_thr_lock);
10081 err = SET_ERROR(ECANCELED);
10084 mutex_exit(&l2arc_rebuild_thr_lock);
10085 if (spa_config_tryenter(spa, SCL_L2ARC, vd,
10087 lock_held = B_TRUE;
10091 * L2ARC config lock held by somebody in writer,
10092 * possibly due to them trying to remove us. They'll
10093 * likely to want us to shut down, so after a little
10094 * delay, we check l2ad_rebuild_cancel and retry
10101 * Continue with the next log block.
10104 lbps[1] = this_lb->lb_prev_lbp;
10105 PTR_SWAP(this_lb, next_lb);
10110 if (this_io != NULL)
10111 l2arc_log_blk_fetch_abort(this_io);
10113 if (next_io != NULL)
10114 l2arc_log_blk_fetch_abort(next_io);
10115 vmem_free(this_lb, sizeof (*this_lb));
10116 vmem_free(next_lb, sizeof (*next_lb));
10118 if (!l2arc_rebuild_enabled) {
10119 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10121 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) {
10122 ARCSTAT_BUMP(arcstat_l2_rebuild_success);
10123 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10124 "successful, restored %llu blocks",
10125 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10126 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) {
10128 * No error but also nothing restored, meaning the lbps array
10129 * in the device header points to invalid/non-present log
10130 * blocks. Reset the header.
10132 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10133 "no valid log blocks");
10134 memset(l2dhdr, 0, dev->l2ad_dev_hdr_asize);
10135 l2arc_dev_hdr_update(dev);
10136 } else if (err == ECANCELED) {
10138 * In case the rebuild was canceled do not log to spa history
10139 * log as the pool may be in the process of being removed.
10141 zfs_dbgmsg("L2ARC rebuild aborted, restored %llu blocks",
10142 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10143 } else if (err != 0) {
10144 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10145 "aborted, restored %llu blocks",
10146 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10150 spa_config_exit(spa, SCL_L2ARC, vd);
10156 * Attempts to read the device header on the provided L2ARC device and writes
10157 * it to `hdr'. On success, this function returns 0, otherwise the appropriate
10158 * error code is returned.
10161 l2arc_dev_hdr_read(l2arc_dev_t *dev)
10165 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10166 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10169 guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10171 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10173 err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev,
10174 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd,
10175 ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
10176 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY |
10177 ZIO_FLAG_SPECULATIVE, B_FALSE));
10182 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors);
10183 zfs_dbgmsg("L2ARC IO error (%d) while reading device header, "
10184 "vdev guid: %llu", err,
10185 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10189 if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC))
10190 byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr));
10192 if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC ||
10193 l2dhdr->dh_spa_guid != guid ||
10194 l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid ||
10195 l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION ||
10196 l2dhdr->dh_log_entries != dev->l2ad_log_entries ||
10197 l2dhdr->dh_end != dev->l2ad_end ||
10198 !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end,
10199 l2dhdr->dh_evict) ||
10200 (l2dhdr->dh_trim_state != VDEV_TRIM_COMPLETE &&
10201 l2arc_trim_ahead > 0)) {
10203 * Attempt to rebuild a device containing no actual dev hdr
10204 * or containing a header from some other pool or from another
10205 * version of persistent L2ARC.
10207 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported);
10208 return (SET_ERROR(ENOTSUP));
10215 * Reads L2ARC log blocks from storage and validates their contents.
10217 * This function implements a simple fetcher to make sure that while
10218 * we're processing one buffer the L2ARC is already fetching the next
10219 * one in the chain.
10221 * The arguments this_lp and next_lp point to the current and next log block
10222 * address in the block chain. Similarly, this_lb and next_lb hold the
10223 * l2arc_log_blk_phys_t's of the current and next L2ARC blk.
10225 * The `this_io' and `next_io' arguments are used for block fetching.
10226 * When issuing the first blk IO during rebuild, you should pass NULL for
10227 * `this_io'. This function will then issue a sync IO to read the block and
10228 * also issue an async IO to fetch the next block in the block chain. The
10229 * fetched IO is returned in `next_io'. On subsequent calls to this
10230 * function, pass the value returned in `next_io' from the previous call
10231 * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO.
10232 * Prior to the call, you should initialize your `next_io' pointer to be
10233 * NULL. If no fetch IO was issued, the pointer is left set at NULL.
10235 * On success, this function returns 0, otherwise it returns an appropriate
10236 * error code. On error the fetching IO is aborted and cleared before
10237 * returning from this function. Therefore, if we return `success', the
10238 * caller can assume that we have taken care of cleanup of fetch IOs.
10241 l2arc_log_blk_read(l2arc_dev_t *dev,
10242 const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp,
10243 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
10244 zio_t *this_io, zio_t **next_io)
10251 ASSERT(this_lbp != NULL && next_lbp != NULL);
10252 ASSERT(this_lb != NULL && next_lb != NULL);
10253 ASSERT(next_io != NULL && *next_io == NULL);
10254 ASSERT(l2arc_log_blkptr_valid(dev, this_lbp));
10257 * Check to see if we have issued the IO for this log block in a
10258 * previous run. If not, this is the first call, so issue it now.
10260 if (this_io == NULL) {
10261 this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp,
10266 * Peek to see if we can start issuing the next IO immediately.
10268 if (l2arc_log_blkptr_valid(dev, next_lbp)) {
10270 * Start issuing IO for the next log block early - this
10271 * should help keep the L2ARC device busy while we
10272 * decompress and restore this log block.
10274 *next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp,
10278 /* Wait for the IO to read this log block to complete */
10279 if ((err = zio_wait(this_io)) != 0) {
10280 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
10281 zfs_dbgmsg("L2ARC IO error (%d) while reading log block, "
10282 "offset: %llu, vdev guid: %llu", err,
10283 (u_longlong_t)this_lbp->lbp_daddr,
10284 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10289 * Make sure the buffer checks out.
10290 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10292 asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop);
10293 fletcher_4_native(this_lb, asize, NULL, &cksum);
10294 if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) {
10295 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors);
10296 zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, "
10297 "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu",
10298 (u_longlong_t)this_lbp->lbp_daddr,
10299 (u_longlong_t)dev->l2ad_vdev->vdev_guid,
10300 (u_longlong_t)dev->l2ad_hand,
10301 (u_longlong_t)dev->l2ad_evict);
10302 err = SET_ERROR(ECKSUM);
10306 /* Now we can take our time decoding this buffer */
10307 switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) {
10308 case ZIO_COMPRESS_OFF:
10310 case ZIO_COMPRESS_LZ4:
10311 abd = abd_alloc_for_io(asize, B_TRUE);
10312 abd_copy_from_buf_off(abd, this_lb, 0, asize);
10313 if ((err = zio_decompress_data(
10314 L2BLK_GET_COMPRESS((this_lbp)->lbp_prop),
10315 abd, this_lb, asize, sizeof (*this_lb), NULL)) != 0) {
10316 err = SET_ERROR(EINVAL);
10321 err = SET_ERROR(EINVAL);
10324 if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC))
10325 byteswap_uint64_array(this_lb, sizeof (*this_lb));
10326 if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) {
10327 err = SET_ERROR(EINVAL);
10331 /* Abort an in-flight fetch I/O in case of error */
10332 if (err != 0 && *next_io != NULL) {
10333 l2arc_log_blk_fetch_abort(*next_io);
10342 * Restores the payload of a log block to ARC. This creates empty ARC hdr
10343 * entries which only contain an l2arc hdr, essentially restoring the
10344 * buffers to their L2ARC evicted state. This function also updates space
10345 * usage on the L2ARC vdev to make sure it tracks restored buffers.
10348 l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb,
10351 uint64_t size = 0, asize = 0;
10352 uint64_t log_entries = dev->l2ad_log_entries;
10355 * Usually arc_adapt() is called only for data, not headers, but
10356 * since we may allocate significant amount of memory here, let ARC
10359 arc_adapt(log_entries * HDR_L2ONLY_SIZE);
10361 for (int i = log_entries - 1; i >= 0; i--) {
10363 * Restore goes in the reverse temporal direction to preserve
10364 * correct temporal ordering of buffers in the l2ad_buflist.
10365 * l2arc_hdr_restore also does a list_insert_tail instead of
10366 * list_insert_head on the l2ad_buflist:
10368 * LIST l2ad_buflist LIST
10369 * HEAD <------ (time) ------ TAIL
10370 * direction +-----+-----+-----+-----+-----+ direction
10371 * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild
10372 * fill +-----+-----+-----+-----+-----+
10376 * l2arc_feed_thread l2arc_rebuild
10377 * will place new bufs here restores bufs here
10379 * During l2arc_rebuild() the device is not used by
10380 * l2arc_feed_thread() as dev->l2ad_rebuild is set to true.
10382 size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop);
10383 asize += vdev_psize_to_asize(dev->l2ad_vdev,
10384 L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop));
10385 l2arc_hdr_restore(&lb->lb_entries[i], dev);
10389 * Record rebuild stats:
10390 * size Logical size of restored buffers in the L2ARC
10391 * asize Aligned size of restored buffers in the L2ARC
10393 ARCSTAT_INCR(arcstat_l2_rebuild_size, size);
10394 ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize);
10395 ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries);
10396 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize);
10397 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize);
10398 ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks);
10402 * Restores a single ARC buf hdr from a log entry. The ARC buffer is put
10403 * into a state indicating that it has been evicted to L2ARC.
10406 l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
10408 arc_buf_hdr_t *hdr, *exists;
10409 kmutex_t *hash_lock;
10410 arc_buf_contents_t type = L2BLK_GET_TYPE((le)->le_prop);
10414 * Do all the allocation before grabbing any locks, this lets us
10415 * sleep if memory is full and we don't have to deal with failed
10418 hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type,
10419 dev, le->le_dva, le->le_daddr,
10420 L2BLK_GET_PSIZE((le)->le_prop), le->le_birth,
10421 L2BLK_GET_COMPRESS((le)->le_prop), le->le_complevel,
10422 L2BLK_GET_PROTECTED((le)->le_prop),
10423 L2BLK_GET_PREFETCH((le)->le_prop),
10424 L2BLK_GET_STATE((le)->le_prop));
10425 asize = vdev_psize_to_asize(dev->l2ad_vdev,
10426 L2BLK_GET_PSIZE((le)->le_prop));
10429 * vdev_space_update() has to be called before arc_hdr_destroy() to
10430 * avoid underflow since the latter also calls vdev_space_update().
10432 l2arc_hdr_arcstats_increment(hdr);
10433 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10435 mutex_enter(&dev->l2ad_mtx);
10436 list_insert_tail(&dev->l2ad_buflist, hdr);
10437 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
10438 mutex_exit(&dev->l2ad_mtx);
10440 exists = buf_hash_insert(hdr, &hash_lock);
10442 /* Buffer was already cached, no need to restore it. */
10443 arc_hdr_destroy(hdr);
10445 * If the buffer is already cached, check whether it has
10446 * L2ARC metadata. If not, enter them and update the flag.
10447 * This is important is case of onlining a cache device, since
10448 * we previously evicted all L2ARC metadata from ARC.
10450 if (!HDR_HAS_L2HDR(exists)) {
10451 arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR);
10452 exists->b_l2hdr.b_dev = dev;
10453 exists->b_l2hdr.b_daddr = le->le_daddr;
10454 exists->b_l2hdr.b_arcs_state =
10455 L2BLK_GET_STATE((le)->le_prop);
10456 mutex_enter(&dev->l2ad_mtx);
10457 list_insert_tail(&dev->l2ad_buflist, exists);
10458 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
10459 arc_hdr_size(exists), exists);
10460 mutex_exit(&dev->l2ad_mtx);
10461 l2arc_hdr_arcstats_increment(exists);
10462 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10464 ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
10467 mutex_exit(hash_lock);
10471 * Starts an asynchronous read IO to read a log block. This is used in log
10472 * block reconstruction to start reading the next block before we are done
10473 * decoding and reconstructing the current block, to keep the l2arc device
10474 * nice and hot with read IO to process.
10475 * The returned zio will contain a newly allocated memory buffers for the IO
10476 * data which should then be freed by the caller once the zio is no longer
10477 * needed (i.e. due to it having completed). If you wish to abort this
10478 * zio, you should do so using l2arc_log_blk_fetch_abort, which takes
10479 * care of disposing of the allocated buffers correctly.
10482 l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp,
10483 l2arc_log_blk_phys_t *lb)
10487 l2arc_read_callback_t *cb;
10489 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10490 asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10491 ASSERT(asize <= sizeof (l2arc_log_blk_phys_t));
10493 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP);
10494 cb->l2rcb_abd = abd_get_from_buf(lb, asize);
10495 pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb,
10496 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
10497 (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize,
10498 cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10499 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL |
10500 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
10506 * Aborts a zio returned from l2arc_log_blk_fetch and frees the data
10507 * buffers allocated for it.
10510 l2arc_log_blk_fetch_abort(zio_t *zio)
10512 (void) zio_wait(zio);
10516 * Creates a zio to update the device header on an l2arc device.
10519 l2arc_dev_hdr_update(l2arc_dev_t *dev)
10521 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10522 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10526 VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER));
10528 l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC;
10529 l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION;
10530 l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10531 l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid;
10532 l2dhdr->dh_log_entries = dev->l2ad_log_entries;
10533 l2dhdr->dh_evict = dev->l2ad_evict;
10534 l2dhdr->dh_start = dev->l2ad_start;
10535 l2dhdr->dh_end = dev->l2ad_end;
10536 l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize);
10537 l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count);
10538 l2dhdr->dh_flags = 0;
10539 l2dhdr->dh_trim_action_time = dev->l2ad_vdev->vdev_trim_action_time;
10540 l2dhdr->dh_trim_state = dev->l2ad_vdev->vdev_trim_state;
10541 if (dev->l2ad_first)
10542 l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST;
10544 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10546 err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev,
10547 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL,
10548 NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE));
10553 zfs_dbgmsg("L2ARC IO error (%d) while writing device header, "
10554 "vdev guid: %llu", err,
10555 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10560 * Commits a log block to the L2ARC device. This routine is invoked from
10561 * l2arc_write_buffers when the log block fills up.
10562 * This function allocates some memory to temporarily hold the serialized
10563 * buffer to be written. This is then released in l2arc_write_done.
10566 l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb)
10568 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10569 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10570 uint64_t psize, asize;
10572 l2arc_lb_abd_buf_t *abd_buf;
10573 uint8_t *tmpbuf = NULL;
10574 l2arc_lb_ptr_buf_t *lb_ptr_buf;
10576 VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries);
10578 abd_buf = zio_buf_alloc(sizeof (*abd_buf));
10579 abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb));
10580 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
10581 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP);
10583 /* link the buffer into the block chain */
10584 lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1];
10585 lb->lb_magic = L2ARC_LOG_BLK_MAGIC;
10588 * l2arc_log_blk_commit() may be called multiple times during a single
10589 * l2arc_write_buffers() call. Save the allocated abd buffers in a list
10590 * so we can free them in l2arc_write_done() later on.
10592 list_insert_tail(&cb->l2wcb_abd_list, abd_buf);
10594 /* try to compress the buffer */
10595 psize = zio_compress_data(ZIO_COMPRESS_LZ4,
10596 abd_buf->abd, (void **) &tmpbuf, sizeof (*lb), 0);
10598 /* a log block is never entirely zero */
10599 ASSERT(psize != 0);
10600 asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
10601 ASSERT(asize <= sizeof (*lb));
10604 * Update the start log block pointer in the device header to point
10605 * to the log block we're about to write.
10607 l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0];
10608 l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand;
10609 l2dhdr->dh_start_lbps[0].lbp_payload_asize =
10610 dev->l2ad_log_blk_payload_asize;
10611 l2dhdr->dh_start_lbps[0].lbp_payload_start =
10612 dev->l2ad_log_blk_payload_start;
10614 (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb));
10616 (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize);
10617 L2BLK_SET_CHECKSUM(
10618 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10619 ZIO_CHECKSUM_FLETCHER_4);
10620 if (asize < sizeof (*lb)) {
10621 /* compression succeeded */
10622 memset(tmpbuf + psize, 0, asize - psize);
10623 L2BLK_SET_COMPRESS(
10624 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10627 /* compression failed */
10628 memcpy(tmpbuf, lb, sizeof (*lb));
10629 L2BLK_SET_COMPRESS(
10630 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10634 /* checksum what we're about to write */
10635 fletcher_4_native(tmpbuf, asize, NULL,
10636 &l2dhdr->dh_start_lbps[0].lbp_cksum);
10638 abd_free(abd_buf->abd);
10640 /* perform the write itself */
10641 abd_buf->abd = abd_get_from_buf(tmpbuf, sizeof (*lb));
10642 abd_take_ownership_of_buf(abd_buf->abd, B_TRUE);
10643 wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand,
10644 asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10645 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
10646 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
10647 (void) zio_nowait(wzio);
10649 dev->l2ad_hand += asize;
10651 * Include the committed log block's pointer in the list of pointers
10652 * to log blocks present in the L2ARC device.
10654 memcpy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[0],
10655 sizeof (l2arc_log_blkptr_t));
10656 mutex_enter(&dev->l2ad_mtx);
10657 list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf);
10658 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10659 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10660 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10661 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10662 mutex_exit(&dev->l2ad_mtx);
10663 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10665 /* bump the kstats */
10666 ARCSTAT_INCR(arcstat_l2_write_bytes, asize);
10667 ARCSTAT_BUMP(arcstat_l2_log_blk_writes);
10668 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize);
10669 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio,
10670 dev->l2ad_log_blk_payload_asize / asize);
10672 /* start a new log block */
10673 dev->l2ad_log_ent_idx = 0;
10674 dev->l2ad_log_blk_payload_asize = 0;
10675 dev->l2ad_log_blk_payload_start = 0;
10681 * Validates an L2ARC log block address to make sure that it can be read
10682 * from the provided L2ARC device.
10685 l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp)
10687 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10688 uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10689 uint64_t end = lbp->lbp_daddr + asize - 1;
10690 uint64_t start = lbp->lbp_payload_start;
10691 boolean_t evicted = B_FALSE;
10694 * A log block is valid if all of the following conditions are true:
10695 * - it fits entirely (including its payload) between l2ad_start and
10697 * - it has a valid size
10698 * - neither the log block itself nor part of its payload was evicted
10699 * by l2arc_evict():
10701 * l2ad_hand l2ad_evict
10706 * l2ad_start ============================================ l2ad_end
10707 * --------------------------||||
10714 l2arc_range_check_overlap(start, end, dev->l2ad_hand) ||
10715 l2arc_range_check_overlap(start, end, dev->l2ad_evict) ||
10716 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) ||
10717 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end);
10719 return (start >= dev->l2ad_start && end <= dev->l2ad_end &&
10720 asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) &&
10721 (!evicted || dev->l2ad_first));
10725 * Inserts ARC buffer header `hdr' into the current L2ARC log block on
10726 * the device. The buffer being inserted must be present in L2ARC.
10727 * Returns B_TRUE if the L2ARC log block is full and needs to be committed
10728 * to L2ARC, or B_FALSE if it still has room for more ARC buffers.
10731 l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr)
10733 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10734 l2arc_log_ent_phys_t *le;
10736 if (dev->l2ad_log_entries == 0)
10739 int index = dev->l2ad_log_ent_idx++;
10741 ASSERT3S(index, <, dev->l2ad_log_entries);
10742 ASSERT(HDR_HAS_L2HDR(hdr));
10744 le = &lb->lb_entries[index];
10745 memset(le, 0, sizeof (*le));
10746 le->le_dva = hdr->b_dva;
10747 le->le_birth = hdr->b_birth;
10748 le->le_daddr = hdr->b_l2hdr.b_daddr;
10750 dev->l2ad_log_blk_payload_start = le->le_daddr;
10751 L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr));
10752 L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr));
10753 L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr));
10754 le->le_complevel = hdr->b_complevel;
10755 L2BLK_SET_TYPE((le)->le_prop, hdr->b_type);
10756 L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr)));
10757 L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr)));
10758 L2BLK_SET_STATE((le)->le_prop, hdr->b_l1hdr.b_state->arcs_state);
10760 dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev,
10761 HDR_GET_PSIZE(hdr));
10763 return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries);
10767 * Checks whether a given L2ARC device address sits in a time-sequential
10768 * range. The trick here is that the L2ARC is a rotary buffer, so we can't
10769 * just do a range comparison, we need to handle the situation in which the
10770 * range wraps around the end of the L2ARC device. Arguments:
10771 * bottom -- Lower end of the range to check (written to earlier).
10772 * top -- Upper end of the range to check (written to later).
10773 * check -- The address for which we want to determine if it sits in
10774 * between the top and bottom.
10776 * The 3-way conditional below represents the following cases:
10778 * bottom < top : Sequentially ordered case:
10779 * <check>--------+-------------------+
10780 * | (overlap here?) |
10782 * |---------------<bottom>============<top>--------------|
10784 * bottom > top: Looped-around case:
10785 * <check>--------+------------------+
10786 * | (overlap here?) |
10788 * |===============<top>---------------<bottom>===========|
10791 * +---------------+---------<check>
10793 * top == bottom : Just a single address comparison.
10796 l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check)
10799 return (bottom <= check && check <= top);
10800 else if (bottom > top)
10801 return (check <= top || bottom <= check);
10803 return (check == top);
10806 EXPORT_SYMBOL(arc_buf_size);
10807 EXPORT_SYMBOL(arc_write);
10808 EXPORT_SYMBOL(arc_read);
10809 EXPORT_SYMBOL(arc_buf_info);
10810 EXPORT_SYMBOL(arc_getbuf_func);
10811 EXPORT_SYMBOL(arc_add_prune_callback);
10812 EXPORT_SYMBOL(arc_remove_prune_callback);
10814 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min, param_set_arc_min,
10815 spl_param_get_u64, ZMOD_RW, "Minimum ARC size in bytes");
10817 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, max, param_set_arc_max,
10818 spl_param_get_u64, ZMOD_RW, "Maximum ARC size in bytes");
10820 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_balance, UINT, ZMOD_RW,
10821 "Balance between metadata and data on ghost hits.");
10823 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, grow_retry, param_set_arc_int,
10824 param_get_uint, ZMOD_RW, "Seconds before growing ARC size");
10826 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, shrink_shift, param_set_arc_int,
10827 param_get_uint, ZMOD_RW, "log2(fraction of ARC to reclaim)");
10829 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, pc_percent, UINT, ZMOD_RW,
10830 "Percent of pagecache to reclaim ARC to");
10832 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, average_blocksize, UINT, ZMOD_RD,
10833 "Target average block size");
10835 ZFS_MODULE_PARAM(zfs, zfs_, compressed_arc_enabled, INT, ZMOD_RW,
10836 "Disable compressed ARC buffers");
10838 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prefetch_ms, param_set_arc_int,
10839 param_get_uint, ZMOD_RW, "Min life of prefetch block in ms");
10841 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prescient_prefetch_ms,
10842 param_set_arc_int, param_get_uint, ZMOD_RW,
10843 "Min life of prescient prefetched block in ms");
10845 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_max, U64, ZMOD_RW,
10846 "Max write bytes per interval");
10848 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_boost, U64, ZMOD_RW,
10849 "Extra write bytes during device warmup");
10851 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom, U64, ZMOD_RW,
10852 "Number of max device writes to precache");
10854 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom_boost, U64, ZMOD_RW,
10855 "Compressed l2arc_headroom multiplier");
10857 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, trim_ahead, U64, ZMOD_RW,
10858 "TRIM ahead L2ARC write size multiplier");
10860 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_secs, U64, ZMOD_RW,
10861 "Seconds between L2ARC writing");
10863 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_min_ms, U64, ZMOD_RW,
10864 "Min feed interval in milliseconds");
10866 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, noprefetch, INT, ZMOD_RW,
10867 "Skip caching prefetched buffers");
10869 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_again, INT, ZMOD_RW,
10870 "Turbo L2ARC warmup");
10872 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, norw, INT, ZMOD_RW,
10873 "No reads during writes");
10875 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, meta_percent, UINT, ZMOD_RW,
10876 "Percent of ARC size allowed for L2ARC-only headers");
10878 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_enabled, INT, ZMOD_RW,
10879 "Rebuild the L2ARC when importing a pool");
10881 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_blocks_min_l2size, U64, ZMOD_RW,
10882 "Min size in bytes to write rebuild log blocks in L2ARC");
10884 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, mfuonly, INT, ZMOD_RW,
10885 "Cache only MFU data from ARC into L2ARC");
10887 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, exclude_special, INT, ZMOD_RW,
10888 "Exclude dbufs on special vdevs from being cached to L2ARC if set.");
10890 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, lotsfree_percent, param_set_arc_int,
10891 param_get_uint, ZMOD_RW, "System free memory I/O throttle in bytes");
10893 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, sys_free, param_set_arc_u64,
10894 spl_param_get_u64, ZMOD_RW, "System free memory target size in bytes");
10896 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit, param_set_arc_u64,
10897 spl_param_get_u64, ZMOD_RW, "Minimum bytes of dnodes in ARC");
10899 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit_percent,
10900 param_set_arc_int, param_get_uint, ZMOD_RW,
10901 "Percent of ARC meta buffers for dnodes");
10903 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, dnode_reduce_percent, UINT, ZMOD_RW,
10904 "Percentage of excess dnodes to try to unpin");
10906 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, eviction_pct, UINT, ZMOD_RW,
10907 "When full, ARC allocation waits for eviction of this % of alloc size");
10909 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, evict_batch_limit, UINT, ZMOD_RW,
10910 "The number of headers to evict per sublist before moving to the next");
10912 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, prune_task_threads, INT, ZMOD_RW,
10913 "Number of arc_prune threads");