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_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
752 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
755 * Hash table routines
758 #define BUF_LOCKS 2048
759 typedef struct buf_hash_table {
761 arc_buf_hdr_t **ht_table;
762 kmutex_t ht_locks[BUF_LOCKS] ____cacheline_aligned;
765 static buf_hash_table_t buf_hash_table;
767 #define BUF_HASH_INDEX(spa, dva, birth) \
768 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
769 #define BUF_HASH_LOCK(idx) (&buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
770 #define HDR_LOCK(hdr) \
771 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
773 uint64_t zfs_crc64_table[256];
779 #define L2ARC_WRITE_SIZE (32 * 1024 * 1024) /* initial write max */
780 #define L2ARC_HEADROOM 8 /* num of writes */
783 * If we discover during ARC scan any buffers to be compressed, we boost
784 * our headroom for the next scanning cycle by this percentage multiple.
786 #define L2ARC_HEADROOM_BOOST 200
787 #define L2ARC_FEED_SECS 1 /* caching interval secs */
788 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
791 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
792 * and each of the state has two types: data and metadata.
794 #define L2ARC_FEED_TYPES 4
796 /* L2ARC Performance Tunables */
797 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
798 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
799 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
800 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
801 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
802 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
803 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
804 int l2arc_feed_again = B_TRUE; /* turbo warmup */
805 int l2arc_norw = B_FALSE; /* no reads during writes */
806 static uint_t l2arc_meta_percent = 33; /* limit on headers size */
811 static list_t L2ARC_dev_list; /* device list */
812 static list_t *l2arc_dev_list; /* device list pointer */
813 static kmutex_t l2arc_dev_mtx; /* device list mutex */
814 static l2arc_dev_t *l2arc_dev_last; /* last device used */
815 static list_t L2ARC_free_on_write; /* free after write buf list */
816 static list_t *l2arc_free_on_write; /* free after write list ptr */
817 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
818 static uint64_t l2arc_ndev; /* number of devices */
820 typedef struct l2arc_read_callback {
821 arc_buf_hdr_t *l2rcb_hdr; /* read header */
822 blkptr_t l2rcb_bp; /* original blkptr */
823 zbookmark_phys_t l2rcb_zb; /* original bookmark */
824 int l2rcb_flags; /* original flags */
825 abd_t *l2rcb_abd; /* temporary buffer */
826 } l2arc_read_callback_t;
828 typedef struct l2arc_data_free {
829 /* protected by l2arc_free_on_write_mtx */
832 arc_buf_contents_t l2df_type;
833 list_node_t l2df_list_node;
836 typedef enum arc_fill_flags {
837 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
838 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
839 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
840 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
841 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
844 typedef enum arc_ovf_level {
845 ARC_OVF_NONE, /* ARC within target size. */
846 ARC_OVF_SOME, /* ARC is slightly overflowed. */
847 ARC_OVF_SEVERE /* ARC is severely overflowed. */
850 static kmutex_t l2arc_feed_thr_lock;
851 static kcondvar_t l2arc_feed_thr_cv;
852 static uint8_t l2arc_thread_exit;
854 static kmutex_t l2arc_rebuild_thr_lock;
855 static kcondvar_t l2arc_rebuild_thr_cv;
857 enum arc_hdr_alloc_flags {
858 ARC_HDR_ALLOC_RDATA = 0x1,
859 ARC_HDR_USE_RESERVE = 0x4,
860 ARC_HDR_ALLOC_LINEAR = 0x8,
864 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, const void *, int);
865 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, const void *);
866 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, const void *, int);
867 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, const void *);
868 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, const void *);
869 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size,
871 static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t);
872 static void arc_hdr_alloc_abd(arc_buf_hdr_t *, int);
873 static void arc_hdr_destroy(arc_buf_hdr_t *);
874 static void arc_access(arc_buf_hdr_t *, arc_flags_t, boolean_t);
875 static void arc_buf_watch(arc_buf_t *);
876 static void arc_change_state(arc_state_t *, arc_buf_hdr_t *);
878 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
879 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
880 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
881 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
883 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
884 static void l2arc_read_done(zio_t *);
885 static void l2arc_do_free_on_write(void);
886 static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
887 boolean_t state_only);
889 static void arc_prune_async(uint64_t adjust);
891 #define l2arc_hdr_arcstats_increment(hdr) \
892 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_FALSE)
893 #define l2arc_hdr_arcstats_decrement(hdr) \
894 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_FALSE)
895 #define l2arc_hdr_arcstats_increment_state(hdr) \
896 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_TRUE)
897 #define l2arc_hdr_arcstats_decrement_state(hdr) \
898 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_TRUE)
901 * l2arc_exclude_special : A zfs module parameter that controls whether buffers
902 * present on special vdevs are eligibile for caching in L2ARC. If
903 * set to 1, exclude dbufs on special vdevs from being cached to
906 int l2arc_exclude_special = 0;
909 * l2arc_mfuonly : A ZFS module parameter that controls whether only MFU
910 * metadata and data are cached from ARC into L2ARC.
912 static int l2arc_mfuonly = 0;
916 * l2arc_trim_ahead : A ZFS module parameter that controls how much ahead of
917 * the current write size (l2arc_write_max) we should TRIM if we
918 * have filled the device. It is defined as a percentage of the
919 * write size. If set to 100 we trim twice the space required to
920 * accommodate upcoming writes. A minimum of 64MB will be trimmed.
921 * It also enables TRIM of the whole L2ARC device upon creation or
922 * addition to an existing pool or if the header of the device is
923 * invalid upon importing a pool or onlining a cache device. The
924 * default is 0, which disables TRIM on L2ARC altogether as it can
925 * put significant stress on the underlying storage devices. This
926 * will vary depending of how well the specific device handles
929 static uint64_t l2arc_trim_ahead = 0;
932 * Performance tuning of L2ARC persistence:
934 * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding
935 * an L2ARC device (either at pool import or later) will attempt
936 * to rebuild L2ARC buffer contents.
937 * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls
938 * whether log blocks are written to the L2ARC device. If the L2ARC
939 * device is less than 1GB, the amount of data l2arc_evict()
940 * evicts is significant compared to the amount of restored L2ARC
941 * data. In this case do not write log blocks in L2ARC in order
942 * not to waste space.
944 static int l2arc_rebuild_enabled = B_TRUE;
945 static uint64_t l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024;
947 /* L2ARC persistence rebuild control routines. */
948 void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen);
949 static __attribute__((noreturn)) void l2arc_dev_rebuild_thread(void *arg);
950 static int l2arc_rebuild(l2arc_dev_t *dev);
952 /* L2ARC persistence read I/O routines. */
953 static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
954 static int l2arc_log_blk_read(l2arc_dev_t *dev,
955 const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
956 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
957 zio_t *this_io, zio_t **next_io);
958 static zio_t *l2arc_log_blk_fetch(vdev_t *vd,
959 const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb);
960 static void l2arc_log_blk_fetch_abort(zio_t *zio);
962 /* L2ARC persistence block restoration routines. */
963 static void l2arc_log_blk_restore(l2arc_dev_t *dev,
964 const l2arc_log_blk_phys_t *lb, uint64_t lb_asize);
965 static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
968 /* L2ARC persistence write I/O routines. */
969 static uint64_t l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
970 l2arc_write_callback_t *cb);
972 /* L2ARC persistence auxiliary routines. */
973 boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
974 const l2arc_log_blkptr_t *lbp);
975 static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
976 const arc_buf_hdr_t *ab);
977 boolean_t l2arc_range_check_overlap(uint64_t bottom,
978 uint64_t top, uint64_t check);
979 static void l2arc_blk_fetch_done(zio_t *zio);
980 static inline uint64_t
981 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev);
984 * We use Cityhash for this. It's fast, and has good hash properties without
985 * requiring any large static buffers.
988 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
990 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
993 #define HDR_EMPTY(hdr) \
994 ((hdr)->b_dva.dva_word[0] == 0 && \
995 (hdr)->b_dva.dva_word[1] == 0)
997 #define HDR_EMPTY_OR_LOCKED(hdr) \
998 (HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr)))
1000 #define HDR_EQUAL(spa, dva, birth, hdr) \
1001 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1002 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1003 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1006 buf_discard_identity(arc_buf_hdr_t *hdr)
1008 hdr->b_dva.dva_word[0] = 0;
1009 hdr->b_dva.dva_word[1] = 0;
1013 static arc_buf_hdr_t *
1014 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1016 const dva_t *dva = BP_IDENTITY(bp);
1017 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1018 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1019 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1022 mutex_enter(hash_lock);
1023 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1024 hdr = hdr->b_hash_next) {
1025 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1030 mutex_exit(hash_lock);
1036 * Insert an entry into the hash table. If there is already an element
1037 * equal to elem in the hash table, then the already existing element
1038 * will be returned and the new element will not be inserted.
1039 * Otherwise returns NULL.
1040 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1042 static arc_buf_hdr_t *
1043 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1045 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1046 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1047 arc_buf_hdr_t *fhdr;
1050 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1051 ASSERT(hdr->b_birth != 0);
1052 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1054 if (lockp != NULL) {
1056 mutex_enter(hash_lock);
1058 ASSERT(MUTEX_HELD(hash_lock));
1061 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1062 fhdr = fhdr->b_hash_next, i++) {
1063 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1067 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1068 buf_hash_table.ht_table[idx] = hdr;
1069 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1071 /* collect some hash table performance data */
1073 ARCSTAT_BUMP(arcstat_hash_collisions);
1075 ARCSTAT_BUMP(arcstat_hash_chains);
1077 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1079 uint64_t he = atomic_inc_64_nv(
1080 &arc_stats.arcstat_hash_elements.value.ui64);
1081 ARCSTAT_MAX(arcstat_hash_elements_max, he);
1087 buf_hash_remove(arc_buf_hdr_t *hdr)
1089 arc_buf_hdr_t *fhdr, **hdrp;
1090 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1092 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1093 ASSERT(HDR_IN_HASH_TABLE(hdr));
1095 hdrp = &buf_hash_table.ht_table[idx];
1096 while ((fhdr = *hdrp) != hdr) {
1097 ASSERT3P(fhdr, !=, NULL);
1098 hdrp = &fhdr->b_hash_next;
1100 *hdrp = hdr->b_hash_next;
1101 hdr->b_hash_next = NULL;
1102 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1104 /* collect some hash table performance data */
1105 atomic_dec_64(&arc_stats.arcstat_hash_elements.value.ui64);
1107 if (buf_hash_table.ht_table[idx] &&
1108 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1109 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1113 * Global data structures and functions for the buf kmem cache.
1116 static kmem_cache_t *hdr_full_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_l2only_cache);
1138 kmem_cache_destroy(buf_cache);
1142 * Constructor callback - called when the cache is empty
1143 * and a new buf is requested.
1146 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1148 (void) unused, (void) kmflag;
1149 arc_buf_hdr_t *hdr = vbuf;
1151 memset(hdr, 0, HDR_FULL_SIZE);
1152 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1153 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1155 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1157 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1158 list_link_init(&hdr->b_l2hdr.b_l2node);
1159 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1165 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1167 (void) unused, (void) kmflag;
1168 arc_buf_hdr_t *hdr = vbuf;
1170 memset(hdr, 0, HDR_L2ONLY_SIZE);
1171 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1177 buf_cons(void *vbuf, void *unused, int kmflag)
1179 (void) unused, (void) kmflag;
1180 arc_buf_t *buf = vbuf;
1182 memset(buf, 0, sizeof (arc_buf_t));
1183 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1189 * Destructor callback - called when a cached buf is
1190 * no longer required.
1193 hdr_full_dest(void *vbuf, void *unused)
1196 arc_buf_hdr_t *hdr = vbuf;
1198 ASSERT(HDR_EMPTY(hdr));
1199 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1201 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1203 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1204 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1208 hdr_l2only_dest(void *vbuf, void *unused)
1211 arc_buf_hdr_t *hdr = vbuf;
1213 ASSERT(HDR_EMPTY(hdr));
1214 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1218 buf_dest(void *vbuf, void *unused)
1223 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1229 uint64_t *ct = NULL;
1230 uint64_t hsize = 1ULL << 12;
1234 * The hash table is big enough to fill all of physical memory
1235 * with an average block size of zfs_arc_average_blocksize (default 8K).
1236 * By default, the table will take up
1237 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1239 while (hsize * zfs_arc_average_blocksize < arc_all_memory())
1242 buf_hash_table.ht_mask = hsize - 1;
1243 #if defined(_KERNEL)
1245 * Large allocations which do not require contiguous pages
1246 * should be using vmem_alloc() in the linux kernel
1248 buf_hash_table.ht_table =
1249 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
1251 buf_hash_table.ht_table =
1252 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1254 if (buf_hash_table.ht_table == NULL) {
1255 ASSERT(hsize > (1ULL << 8));
1260 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1261 0, hdr_full_cons, hdr_full_dest, NULL, NULL, NULL, 0);
1262 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1263 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, NULL,
1265 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1266 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1268 for (i = 0; i < 256; i++)
1269 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1270 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1272 for (i = 0; i < BUF_LOCKS; i++)
1273 mutex_init(BUF_HASH_LOCK(i), NULL, MUTEX_DEFAULT, NULL);
1276 #define ARC_MINTIME (hz>>4) /* 62 ms */
1279 * This is the size that the buf occupies in memory. If the buf is compressed,
1280 * it will correspond to the compressed size. You should use this method of
1281 * getting the buf size unless you explicitly need the logical size.
1284 arc_buf_size(arc_buf_t *buf)
1286 return (ARC_BUF_COMPRESSED(buf) ?
1287 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1291 arc_buf_lsize(arc_buf_t *buf)
1293 return (HDR_GET_LSIZE(buf->b_hdr));
1297 * This function will return B_TRUE if the buffer is encrypted in memory.
1298 * This buffer can be decrypted by calling arc_untransform().
1301 arc_is_encrypted(arc_buf_t *buf)
1303 return (ARC_BUF_ENCRYPTED(buf) != 0);
1307 * Returns B_TRUE if the buffer represents data that has not had its MAC
1311 arc_is_unauthenticated(arc_buf_t *buf)
1313 return (HDR_NOAUTH(buf->b_hdr) != 0);
1317 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1318 uint8_t *iv, uint8_t *mac)
1320 arc_buf_hdr_t *hdr = buf->b_hdr;
1322 ASSERT(HDR_PROTECTED(hdr));
1324 memcpy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
1325 memcpy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
1326 memcpy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
1327 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1328 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1332 * Indicates how this buffer is compressed in memory. If it is not compressed
1333 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1334 * arc_untransform() as long as it is also unencrypted.
1337 arc_get_compression(arc_buf_t *buf)
1339 return (ARC_BUF_COMPRESSED(buf) ?
1340 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1344 * Return the compression algorithm used to store this data in the ARC. If ARC
1345 * compression is enabled or this is an encrypted block, this will be the same
1346 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1348 static inline enum zio_compress
1349 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1351 return (HDR_COMPRESSION_ENABLED(hdr) ?
1352 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1356 arc_get_complevel(arc_buf_t *buf)
1358 return (buf->b_hdr->b_complevel);
1361 static inline boolean_t
1362 arc_buf_is_shared(arc_buf_t *buf)
1364 boolean_t shared = (buf->b_data != NULL &&
1365 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1366 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1367 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1368 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1369 EQUIV(shared, ARC_BUF_SHARED(buf));
1370 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1373 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1374 * already being shared" requirement prevents us from doing that.
1381 * Free the checksum associated with this header. If there is no checksum, this
1385 arc_cksum_free(arc_buf_hdr_t *hdr)
1388 ASSERT(HDR_HAS_L1HDR(hdr));
1390 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1391 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1392 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1393 hdr->b_l1hdr.b_freeze_cksum = NULL;
1395 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1400 * Return true iff at least one of the bufs on hdr is not compressed.
1401 * Encrypted buffers count as compressed.
1404 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1406 ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr));
1408 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1409 if (!ARC_BUF_COMPRESSED(b)) {
1418 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1419 * matches the checksum that is stored in the hdr. If there is no checksum,
1420 * or if the buf is compressed, this is a no-op.
1423 arc_cksum_verify(arc_buf_t *buf)
1426 arc_buf_hdr_t *hdr = buf->b_hdr;
1429 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1432 if (ARC_BUF_COMPRESSED(buf))
1435 ASSERT(HDR_HAS_L1HDR(hdr));
1437 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1439 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1440 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1444 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1445 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1446 panic("buffer modified while frozen!");
1447 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1452 * This function makes the assumption that data stored in the L2ARC
1453 * will be transformed exactly as it is in the main pool. Because of
1454 * this we can verify the checksum against the reading process's bp.
1457 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1459 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1460 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1463 * Block pointers always store the checksum for the logical data.
1464 * If the block pointer has the gang bit set, then the checksum
1465 * it represents is for the reconstituted data and not for an
1466 * individual gang member. The zio pipeline, however, must be able to
1467 * determine the checksum of each of the gang constituents so it
1468 * treats the checksum comparison differently than what we need
1469 * for l2arc blocks. This prevents us from using the
1470 * zio_checksum_error() interface directly. Instead we must call the
1471 * zio_checksum_error_impl() so that we can ensure the checksum is
1472 * generated using the correct checksum algorithm and accounts for the
1473 * logical I/O size and not just a gang fragment.
1475 return (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1476 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1477 zio->io_offset, NULL) == 0);
1481 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1482 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1483 * isn't modified later on. If buf is compressed or there is already a checksum
1484 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1487 arc_cksum_compute(arc_buf_t *buf)
1489 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1493 arc_buf_hdr_t *hdr = buf->b_hdr;
1494 ASSERT(HDR_HAS_L1HDR(hdr));
1495 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1496 if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1497 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1501 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1502 ASSERT(!ARC_BUF_COMPRESSED(buf));
1503 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1505 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1506 hdr->b_l1hdr.b_freeze_cksum);
1507 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1514 arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
1516 (void) sig, (void) unused;
1517 panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr);
1522 arc_buf_unwatch(arc_buf_t *buf)
1526 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1527 PROT_READ | PROT_WRITE));
1535 arc_buf_watch(arc_buf_t *buf)
1539 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1546 static arc_buf_contents_t
1547 arc_buf_type(arc_buf_hdr_t *hdr)
1549 arc_buf_contents_t type;
1550 if (HDR_ISTYPE_METADATA(hdr)) {
1551 type = ARC_BUFC_METADATA;
1553 type = ARC_BUFC_DATA;
1555 VERIFY3U(hdr->b_type, ==, type);
1560 arc_is_metadata(arc_buf_t *buf)
1562 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1566 arc_bufc_to_flags(arc_buf_contents_t type)
1570 /* metadata field is 0 if buffer contains normal data */
1572 case ARC_BUFC_METADATA:
1573 return (ARC_FLAG_BUFC_METADATA);
1577 panic("undefined ARC buffer type!");
1578 return ((uint32_t)-1);
1582 arc_buf_thaw(arc_buf_t *buf)
1584 arc_buf_hdr_t *hdr = buf->b_hdr;
1586 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1587 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1589 arc_cksum_verify(buf);
1592 * Compressed buffers do not manipulate the b_freeze_cksum.
1594 if (ARC_BUF_COMPRESSED(buf))
1597 ASSERT(HDR_HAS_L1HDR(hdr));
1598 arc_cksum_free(hdr);
1599 arc_buf_unwatch(buf);
1603 arc_buf_freeze(arc_buf_t *buf)
1605 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1608 if (ARC_BUF_COMPRESSED(buf))
1611 ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1612 arc_cksum_compute(buf);
1616 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1617 * the following functions should be used to ensure that the flags are
1618 * updated in a thread-safe way. When manipulating the flags either
1619 * the hash_lock must be held or the hdr must be undiscoverable. This
1620 * ensures that we're not racing with any other threads when updating
1624 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1626 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1627 hdr->b_flags |= flags;
1631 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1633 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1634 hdr->b_flags &= ~flags;
1638 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1639 * done in a special way since we have to clear and set bits
1640 * at the same time. Consumers that wish to set the compression bits
1641 * must use this function to ensure that the flags are updated in
1642 * thread-safe manner.
1645 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1647 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1650 * Holes and embedded blocks will always have a psize = 0 so
1651 * we ignore the compression of the blkptr and set the
1652 * want to uncompress them. Mark them as uncompressed.
1654 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1655 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1656 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1658 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1659 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1662 HDR_SET_COMPRESS(hdr, cmp);
1663 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1667 * Looks for another buf on the same hdr which has the data decompressed, copies
1668 * from it, and returns true. If no such buf exists, returns false.
1671 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1673 arc_buf_hdr_t *hdr = buf->b_hdr;
1674 boolean_t copied = B_FALSE;
1676 ASSERT(HDR_HAS_L1HDR(hdr));
1677 ASSERT3P(buf->b_data, !=, NULL);
1678 ASSERT(!ARC_BUF_COMPRESSED(buf));
1680 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1681 from = from->b_next) {
1682 /* can't use our own data buffer */
1687 if (!ARC_BUF_COMPRESSED(from)) {
1688 memcpy(buf->b_data, from->b_data, arc_buf_size(buf));
1696 * There were no decompressed bufs, so there should not be a
1697 * checksum on the hdr either.
1699 if (zfs_flags & ZFS_DEBUG_MODIFY)
1700 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1707 * Allocates an ARC buf header that's in an evicted & L2-cached state.
1708 * This is used during l2arc reconstruction to make empty ARC buffers
1709 * which circumvent the regular disk->arc->l2arc path and instead come
1710 * into being in the reverse order, i.e. l2arc->arc.
1712 static arc_buf_hdr_t *
1713 arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
1714 dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth,
1715 enum zio_compress compress, uint8_t complevel, boolean_t protected,
1716 boolean_t prefetch, arc_state_type_t arcs_state)
1721 hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP);
1722 hdr->b_birth = birth;
1725 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR);
1726 HDR_SET_LSIZE(hdr, size);
1727 HDR_SET_PSIZE(hdr, psize);
1728 arc_hdr_set_compress(hdr, compress);
1729 hdr->b_complevel = complevel;
1731 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
1733 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
1734 hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa);
1738 hdr->b_l2hdr.b_dev = dev;
1739 hdr->b_l2hdr.b_daddr = daddr;
1740 hdr->b_l2hdr.b_arcs_state = arcs_state;
1746 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1749 arc_hdr_size(arc_buf_hdr_t *hdr)
1753 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1754 HDR_GET_PSIZE(hdr) > 0) {
1755 size = HDR_GET_PSIZE(hdr);
1757 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1758 size = HDR_GET_LSIZE(hdr);
1764 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1768 uint64_t lsize = HDR_GET_LSIZE(hdr);
1769 uint64_t psize = HDR_GET_PSIZE(hdr);
1770 void *tmpbuf = NULL;
1771 abd_t *abd = hdr->b_l1hdr.b_pabd;
1773 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1774 ASSERT(HDR_AUTHENTICATED(hdr));
1775 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1778 * The MAC is calculated on the compressed data that is stored on disk.
1779 * However, if compressed arc is disabled we will only have the
1780 * decompressed data available to us now. Compress it into a temporary
1781 * abd so we can verify the MAC. The performance overhead of this will
1782 * be relatively low, since most objects in an encrypted objset will
1783 * be encrypted (instead of authenticated) anyway.
1785 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1786 !HDR_COMPRESSION_ENABLED(hdr)) {
1788 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1789 hdr->b_l1hdr.b_pabd, &tmpbuf, lsize, hdr->b_complevel);
1790 ASSERT3P(tmpbuf, !=, NULL);
1791 ASSERT3U(csize, <=, psize);
1792 abd = abd_get_from_buf(tmpbuf, lsize);
1793 abd_take_ownership_of_buf(abd, B_TRUE);
1794 abd_zero_off(abd, csize, psize - csize);
1798 * Authentication is best effort. We authenticate whenever the key is
1799 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1801 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1802 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1803 ASSERT3U(lsize, ==, psize);
1804 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1805 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1807 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1808 hdr->b_crypt_hdr.b_mac);
1812 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1813 else if (ret != ENOENT)
1829 * This function will take a header that only has raw encrypted data in
1830 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1831 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1832 * also decompress the data.
1835 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1840 boolean_t no_crypt = B_FALSE;
1841 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1843 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1844 ASSERT(HDR_ENCRYPTED(hdr));
1846 arc_hdr_alloc_abd(hdr, 0);
1848 ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1849 B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1850 hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1851 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1856 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1857 HDR_GET_PSIZE(hdr));
1861 * If this header has disabled arc compression but the b_pabd is
1862 * compressed after decrypting it, we need to decompress the newly
1865 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1866 !HDR_COMPRESSION_ENABLED(hdr)) {
1868 * We want to make sure that we are correctly honoring the
1869 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1870 * and then loan a buffer from it, rather than allocating a
1871 * linear buffer and wrapping it in an abd later.
1873 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr, 0);
1874 tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
1876 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1877 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
1878 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
1880 abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
1884 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
1885 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1886 arc_hdr_size(hdr), hdr);
1887 hdr->b_l1hdr.b_pabd = cabd;
1893 arc_hdr_free_abd(hdr, B_FALSE);
1895 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
1901 * This function is called during arc_buf_fill() to prepare the header's
1902 * abd plaintext pointer for use. This involves authenticated protected
1903 * data and decrypting encrypted data into the plaintext abd.
1906 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
1907 const zbookmark_phys_t *zb, boolean_t noauth)
1911 ASSERT(HDR_PROTECTED(hdr));
1913 if (hash_lock != NULL)
1914 mutex_enter(hash_lock);
1916 if (HDR_NOAUTH(hdr) && !noauth) {
1918 * The caller requested authenticated data but our data has
1919 * not been authenticated yet. Verify the MAC now if we can.
1921 ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
1924 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
1926 * If we only have the encrypted version of the data, but the
1927 * unencrypted version was requested we take this opportunity
1928 * to store the decrypted version in the header for future use.
1930 ret = arc_hdr_decrypt(hdr, spa, zb);
1935 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1937 if (hash_lock != NULL)
1938 mutex_exit(hash_lock);
1943 if (hash_lock != NULL)
1944 mutex_exit(hash_lock);
1950 * This function is used by the dbuf code to decrypt bonus buffers in place.
1951 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
1952 * block, so we use the hash lock here to protect against concurrent calls to
1956 arc_buf_untransform_in_place(arc_buf_t *buf)
1958 arc_buf_hdr_t *hdr = buf->b_hdr;
1960 ASSERT(HDR_ENCRYPTED(hdr));
1961 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1962 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1963 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1965 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
1967 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
1968 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1972 * Given a buf that has a data buffer attached to it, this function will
1973 * efficiently fill the buf with data of the specified compression setting from
1974 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1975 * are already sharing a data buf, no copy is performed.
1977 * If the buf is marked as compressed but uncompressed data was requested, this
1978 * will allocate a new data buffer for the buf, remove that flag, and fill the
1979 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1980 * uncompressed data, and (since we haven't added support for it yet) if you
1981 * want compressed data your buf must already be marked as compressed and have
1982 * the correct-sized data buffer.
1985 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
1986 arc_fill_flags_t flags)
1989 arc_buf_hdr_t *hdr = buf->b_hdr;
1990 boolean_t hdr_compressed =
1991 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
1992 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
1993 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
1994 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
1995 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
1997 ASSERT3P(buf->b_data, !=, NULL);
1998 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
1999 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2000 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
2001 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
2002 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
2003 IMPLY(encrypted, !arc_buf_is_shared(buf));
2006 * If the caller wanted encrypted data we just need to copy it from
2007 * b_rabd and potentially byteswap it. We won't be able to do any
2008 * further transforms on it.
2011 ASSERT(HDR_HAS_RABD(hdr));
2012 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
2013 HDR_GET_PSIZE(hdr));
2018 * Adjust encrypted and authenticated headers to accommodate
2019 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
2020 * allowed to fail decryption due to keys not being loaded
2021 * without being marked as an IO error.
2023 if (HDR_PROTECTED(hdr)) {
2024 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
2025 zb, !!(flags & ARC_FILL_NOAUTH));
2026 if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
2028 } else if (error != 0) {
2029 if (hash_lock != NULL)
2030 mutex_enter(hash_lock);
2031 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2032 if (hash_lock != NULL)
2033 mutex_exit(hash_lock);
2039 * There is a special case here for dnode blocks which are
2040 * decrypting their bonus buffers. These blocks may request to
2041 * be decrypted in-place. This is necessary because there may
2042 * be many dnodes pointing into this buffer and there is
2043 * currently no method to synchronize replacing the backing
2044 * b_data buffer and updating all of the pointers. Here we use
2045 * the hash lock to ensure there are no races. If the need
2046 * arises for other types to be decrypted in-place, they must
2047 * add handling here as well.
2049 if ((flags & ARC_FILL_IN_PLACE) != 0) {
2050 ASSERT(!hdr_compressed);
2051 ASSERT(!compressed);
2054 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
2055 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2057 if (hash_lock != NULL)
2058 mutex_enter(hash_lock);
2059 arc_buf_untransform_in_place(buf);
2060 if (hash_lock != NULL)
2061 mutex_exit(hash_lock);
2063 /* Compute the hdr's checksum if necessary */
2064 arc_cksum_compute(buf);
2070 if (hdr_compressed == compressed) {
2071 if (ARC_BUF_SHARED(buf)) {
2072 ASSERT(arc_buf_is_shared(buf));
2074 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2078 ASSERT(hdr_compressed);
2079 ASSERT(!compressed);
2082 * If the buf is sharing its data with the hdr, unlink it and
2083 * allocate a new data buffer for the buf.
2085 if (ARC_BUF_SHARED(buf)) {
2086 ASSERT(ARC_BUF_COMPRESSED(buf));
2088 /* We need to give the buf its own b_data */
2089 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2091 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2092 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2094 /* Previously overhead was 0; just add new overhead */
2095 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2096 } else if (ARC_BUF_COMPRESSED(buf)) {
2097 ASSERT(!arc_buf_is_shared(buf));
2099 /* We need to reallocate the buf's b_data */
2100 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2103 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2105 /* We increased the size of b_data; update overhead */
2106 ARCSTAT_INCR(arcstat_overhead_size,
2107 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2111 * Regardless of the buf's previous compression settings, it
2112 * should not be compressed at the end of this function.
2114 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2117 * Try copying the data from another buf which already has a
2118 * decompressed version. If that's not possible, it's time to
2119 * bite the bullet and decompress the data from the hdr.
2121 if (arc_buf_try_copy_decompressed_data(buf)) {
2122 /* Skip byteswapping and checksumming (already done) */
2125 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2126 hdr->b_l1hdr.b_pabd, buf->b_data,
2127 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr),
2131 * Absent hardware errors or software bugs, this should
2132 * be impossible, but log it anyway so we can debug it.
2136 "hdr %px, compress %d, psize %d, lsize %d",
2137 hdr, arc_hdr_get_compress(hdr),
2138 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2139 if (hash_lock != NULL)
2140 mutex_enter(hash_lock);
2141 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2142 if (hash_lock != NULL)
2143 mutex_exit(hash_lock);
2144 return (SET_ERROR(EIO));
2150 /* Byteswap the buf's data if necessary */
2151 if (bswap != DMU_BSWAP_NUMFUNCS) {
2152 ASSERT(!HDR_SHARED_DATA(hdr));
2153 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2154 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2157 /* Compute the hdr's checksum if necessary */
2158 arc_cksum_compute(buf);
2164 * If this function is being called to decrypt an encrypted buffer or verify an
2165 * authenticated one, the key must be loaded and a mapping must be made
2166 * available in the keystore via spa_keystore_create_mapping() or one of its
2170 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2174 arc_fill_flags_t flags = 0;
2177 flags |= ARC_FILL_IN_PLACE;
2179 ret = arc_buf_fill(buf, spa, zb, flags);
2180 if (ret == ECKSUM) {
2182 * Convert authentication and decryption errors to EIO
2183 * (and generate an ereport) before leaving the ARC.
2185 ret = SET_ERROR(EIO);
2186 spa_log_error(spa, zb, &buf->b_hdr->b_birth);
2187 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2188 spa, NULL, zb, NULL, 0);
2195 * Increment the amount of evictable space in the arc_state_t's refcount.
2196 * We account for the space used by the hdr and the arc buf individually
2197 * so that we can add and remove them from the refcount individually.
2200 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2202 arc_buf_contents_t type = arc_buf_type(hdr);
2204 ASSERT(HDR_HAS_L1HDR(hdr));
2206 if (GHOST_STATE(state)) {
2207 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2208 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2209 ASSERT(!HDR_HAS_RABD(hdr));
2210 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2211 HDR_GET_LSIZE(hdr), hdr);
2215 if (hdr->b_l1hdr.b_pabd != NULL) {
2216 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2217 arc_hdr_size(hdr), hdr);
2219 if (HDR_HAS_RABD(hdr)) {
2220 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2221 HDR_GET_PSIZE(hdr), hdr);
2224 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2225 buf = buf->b_next) {
2226 if (ARC_BUF_SHARED(buf))
2228 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2229 arc_buf_size(buf), buf);
2234 * Decrement the amount of evictable space in the arc_state_t's refcount.
2235 * We account for the space used by the hdr and the arc buf individually
2236 * so that we can add and remove them from the refcount individually.
2239 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2241 arc_buf_contents_t type = arc_buf_type(hdr);
2243 ASSERT(HDR_HAS_L1HDR(hdr));
2245 if (GHOST_STATE(state)) {
2246 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2247 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2248 ASSERT(!HDR_HAS_RABD(hdr));
2249 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2250 HDR_GET_LSIZE(hdr), hdr);
2254 if (hdr->b_l1hdr.b_pabd != NULL) {
2255 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2256 arc_hdr_size(hdr), hdr);
2258 if (HDR_HAS_RABD(hdr)) {
2259 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2260 HDR_GET_PSIZE(hdr), hdr);
2263 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2264 buf = buf->b_next) {
2265 if (ARC_BUF_SHARED(buf))
2267 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2268 arc_buf_size(buf), buf);
2273 * Add a reference to this hdr indicating that someone is actively
2274 * referencing that memory. When the refcount transitions from 0 to 1,
2275 * we remove it from the respective arc_state_t list to indicate that
2276 * it is not evictable.
2279 add_reference(arc_buf_hdr_t *hdr, const void *tag)
2281 arc_state_t *state = hdr->b_l1hdr.b_state;
2283 ASSERT(HDR_HAS_L1HDR(hdr));
2284 if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) {
2285 ASSERT(state == arc_anon);
2286 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2287 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2290 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2291 state != arc_anon && state != arc_l2c_only) {
2292 /* We don't use the L2-only state list. */
2293 multilist_remove(&state->arcs_list[arc_buf_type(hdr)], hdr);
2294 arc_evictable_space_decrement(hdr, state);
2299 * Remove a reference from this hdr. When the reference transitions from
2300 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2301 * list making it eligible for eviction.
2304 remove_reference(arc_buf_hdr_t *hdr, const void *tag)
2307 arc_state_t *state = hdr->b_l1hdr.b_state;
2309 ASSERT(HDR_HAS_L1HDR(hdr));
2310 ASSERT(state == arc_anon || MUTEX_HELD(HDR_LOCK(hdr)));
2311 ASSERT(!GHOST_STATE(state)); /* arc_l2c_only counts as a ghost. */
2313 if ((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) != 0)
2316 if (state == arc_anon) {
2317 arc_hdr_destroy(hdr);
2320 if (state == arc_uncached && !HDR_PREFETCH(hdr)) {
2321 arc_change_state(arc_anon, hdr);
2322 arc_hdr_destroy(hdr);
2325 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
2326 arc_evictable_space_increment(hdr, state);
2331 * Returns detailed information about a specific arc buffer. When the
2332 * state_index argument is set the function will calculate the arc header
2333 * list position for its arc state. Since this requires a linear traversal
2334 * callers are strongly encourage not to do this. However, it can be helpful
2335 * for targeted analysis so the functionality is provided.
2338 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2341 arc_buf_hdr_t *hdr = ab->b_hdr;
2342 l1arc_buf_hdr_t *l1hdr = NULL;
2343 l2arc_buf_hdr_t *l2hdr = NULL;
2344 arc_state_t *state = NULL;
2346 memset(abi, 0, sizeof (arc_buf_info_t));
2351 abi->abi_flags = hdr->b_flags;
2353 if (HDR_HAS_L1HDR(hdr)) {
2354 l1hdr = &hdr->b_l1hdr;
2355 state = l1hdr->b_state;
2357 if (HDR_HAS_L2HDR(hdr))
2358 l2hdr = &hdr->b_l2hdr;
2361 abi->abi_bufcnt = 0;
2362 for (arc_buf_t *buf = l1hdr->b_buf; buf; buf = buf->b_next)
2364 abi->abi_access = l1hdr->b_arc_access;
2365 abi->abi_mru_hits = l1hdr->b_mru_hits;
2366 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2367 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2368 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2369 abi->abi_holds = zfs_refcount_count(&l1hdr->b_refcnt);
2373 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2374 abi->abi_l2arc_hits = l2hdr->b_hits;
2377 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2378 abi->abi_state_contents = arc_buf_type(hdr);
2379 abi->abi_size = arc_hdr_size(hdr);
2383 * Move the supplied buffer to the indicated state. The hash lock
2384 * for the buffer must be held by the caller.
2387 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr)
2389 arc_state_t *old_state;
2391 boolean_t update_old, update_new;
2392 arc_buf_contents_t type = arc_buf_type(hdr);
2395 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2396 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2397 * L1 hdr doesn't always exist when we change state to arc_anon before
2398 * destroying a header, in which case reallocating to add the L1 hdr is
2401 if (HDR_HAS_L1HDR(hdr)) {
2402 old_state = hdr->b_l1hdr.b_state;
2403 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2404 update_old = (hdr->b_l1hdr.b_buf != NULL ||
2405 hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
2407 IMPLY(GHOST_STATE(old_state), hdr->b_l1hdr.b_buf == NULL);
2408 IMPLY(GHOST_STATE(new_state), hdr->b_l1hdr.b_buf == NULL);
2409 IMPLY(old_state == arc_anon, hdr->b_l1hdr.b_buf == NULL ||
2410 ARC_BUF_LAST(hdr->b_l1hdr.b_buf));
2412 old_state = arc_l2c_only;
2414 update_old = B_FALSE;
2416 update_new = update_old;
2417 if (GHOST_STATE(old_state))
2418 update_old = B_TRUE;
2419 if (GHOST_STATE(new_state))
2420 update_new = B_TRUE;
2422 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
2423 ASSERT3P(new_state, !=, old_state);
2426 * If this buffer is evictable, transfer it from the
2427 * old state list to the new state list.
2430 if (old_state != arc_anon && old_state != arc_l2c_only) {
2431 ASSERT(HDR_HAS_L1HDR(hdr));
2432 /* remove_reference() saves on insert. */
2433 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2434 multilist_remove(&old_state->arcs_list[type],
2436 arc_evictable_space_decrement(hdr, old_state);
2439 if (new_state != arc_anon && new_state != arc_l2c_only) {
2441 * An L1 header always exists here, since if we're
2442 * moving to some L1-cached state (i.e. not l2c_only or
2443 * anonymous), we realloc the header to add an L1hdr
2446 ASSERT(HDR_HAS_L1HDR(hdr));
2447 multilist_insert(&new_state->arcs_list[type], hdr);
2448 arc_evictable_space_increment(hdr, new_state);
2452 ASSERT(!HDR_EMPTY(hdr));
2453 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2454 buf_hash_remove(hdr);
2456 /* adjust state sizes (ignore arc_l2c_only) */
2458 if (update_new && new_state != arc_l2c_only) {
2459 ASSERT(HDR_HAS_L1HDR(hdr));
2460 if (GHOST_STATE(new_state)) {
2463 * When moving a header to a ghost state, we first
2464 * remove all arc buffers. Thus, we'll have no arc
2465 * buffer to use for the reference. As a result, we
2466 * use the arc header pointer for the reference.
2468 (void) zfs_refcount_add_many(
2469 &new_state->arcs_size[type],
2470 HDR_GET_LSIZE(hdr), hdr);
2471 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2472 ASSERT(!HDR_HAS_RABD(hdr));
2476 * Each individual buffer holds a unique reference,
2477 * thus we must remove each of these references one
2480 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2481 buf = buf->b_next) {
2484 * When the arc_buf_t is sharing the data
2485 * block with the hdr, the owner of the
2486 * reference belongs to the hdr. Only
2487 * add to the refcount if the arc_buf_t is
2490 if (ARC_BUF_SHARED(buf))
2493 (void) zfs_refcount_add_many(
2494 &new_state->arcs_size[type],
2495 arc_buf_size(buf), buf);
2498 if (hdr->b_l1hdr.b_pabd != NULL) {
2499 (void) zfs_refcount_add_many(
2500 &new_state->arcs_size[type],
2501 arc_hdr_size(hdr), hdr);
2504 if (HDR_HAS_RABD(hdr)) {
2505 (void) zfs_refcount_add_many(
2506 &new_state->arcs_size[type],
2507 HDR_GET_PSIZE(hdr), hdr);
2512 if (update_old && old_state != arc_l2c_only) {
2513 ASSERT(HDR_HAS_L1HDR(hdr));
2514 if (GHOST_STATE(old_state)) {
2515 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2516 ASSERT(!HDR_HAS_RABD(hdr));
2519 * When moving a header off of a ghost state,
2520 * the header will not contain any arc buffers.
2521 * We use the arc header pointer for the reference
2522 * which is exactly what we did when we put the
2523 * header on the ghost state.
2526 (void) zfs_refcount_remove_many(
2527 &old_state->arcs_size[type],
2528 HDR_GET_LSIZE(hdr), hdr);
2532 * Each individual buffer holds a unique reference,
2533 * thus we must remove each of these references one
2536 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2537 buf = buf->b_next) {
2540 * When the arc_buf_t is sharing the data
2541 * block with the hdr, the owner of the
2542 * reference belongs to the hdr. Only
2543 * add to the refcount if the arc_buf_t is
2546 if (ARC_BUF_SHARED(buf))
2549 (void) zfs_refcount_remove_many(
2550 &old_state->arcs_size[type],
2551 arc_buf_size(buf), buf);
2553 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2556 if (hdr->b_l1hdr.b_pabd != NULL) {
2557 (void) zfs_refcount_remove_many(
2558 &old_state->arcs_size[type],
2559 arc_hdr_size(hdr), hdr);
2562 if (HDR_HAS_RABD(hdr)) {
2563 (void) zfs_refcount_remove_many(
2564 &old_state->arcs_size[type],
2565 HDR_GET_PSIZE(hdr), hdr);
2570 if (HDR_HAS_L1HDR(hdr)) {
2571 hdr->b_l1hdr.b_state = new_state;
2573 if (HDR_HAS_L2HDR(hdr) && new_state != arc_l2c_only) {
2574 l2arc_hdr_arcstats_decrement_state(hdr);
2575 hdr->b_l2hdr.b_arcs_state = new_state->arcs_state;
2576 l2arc_hdr_arcstats_increment_state(hdr);
2582 arc_space_consume(uint64_t space, arc_space_type_t type)
2584 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2589 case ARC_SPACE_DATA:
2590 ARCSTAT_INCR(arcstat_data_size, space);
2592 case ARC_SPACE_META:
2593 ARCSTAT_INCR(arcstat_metadata_size, space);
2595 case ARC_SPACE_BONUS:
2596 ARCSTAT_INCR(arcstat_bonus_size, space);
2598 case ARC_SPACE_DNODE:
2599 ARCSTAT_INCR(arcstat_dnode_size, space);
2601 case ARC_SPACE_DBUF:
2602 ARCSTAT_INCR(arcstat_dbuf_size, space);
2604 case ARC_SPACE_HDRS:
2605 ARCSTAT_INCR(arcstat_hdr_size, space);
2607 case ARC_SPACE_L2HDRS:
2608 aggsum_add(&arc_sums.arcstat_l2_hdr_size, space);
2610 case ARC_SPACE_ABD_CHUNK_WASTE:
2612 * Note: this includes space wasted by all scatter ABD's, not
2613 * just those allocated by the ARC. But the vast majority of
2614 * scatter ABD's come from the ARC, because other users are
2617 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, space);
2621 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE)
2622 ARCSTAT_INCR(arcstat_meta_used, space);
2624 aggsum_add(&arc_sums.arcstat_size, space);
2628 arc_space_return(uint64_t space, arc_space_type_t type)
2630 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2635 case ARC_SPACE_DATA:
2636 ARCSTAT_INCR(arcstat_data_size, -space);
2638 case ARC_SPACE_META:
2639 ARCSTAT_INCR(arcstat_metadata_size, -space);
2641 case ARC_SPACE_BONUS:
2642 ARCSTAT_INCR(arcstat_bonus_size, -space);
2644 case ARC_SPACE_DNODE:
2645 ARCSTAT_INCR(arcstat_dnode_size, -space);
2647 case ARC_SPACE_DBUF:
2648 ARCSTAT_INCR(arcstat_dbuf_size, -space);
2650 case ARC_SPACE_HDRS:
2651 ARCSTAT_INCR(arcstat_hdr_size, -space);
2653 case ARC_SPACE_L2HDRS:
2654 aggsum_add(&arc_sums.arcstat_l2_hdr_size, -space);
2656 case ARC_SPACE_ABD_CHUNK_WASTE:
2657 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, -space);
2661 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE)
2662 ARCSTAT_INCR(arcstat_meta_used, -space);
2664 ASSERT(aggsum_compare(&arc_sums.arcstat_size, space) >= 0);
2665 aggsum_add(&arc_sums.arcstat_size, -space);
2669 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2670 * with the hdr's b_pabd.
2673 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2676 * The criteria for sharing a hdr's data are:
2677 * 1. the buffer is not encrypted
2678 * 2. the hdr's compression matches the buf's compression
2679 * 3. the hdr doesn't need to be byteswapped
2680 * 4. the hdr isn't already being shared
2681 * 5. the buf is either compressed or it is the last buf in the hdr list
2683 * Criterion #5 maintains the invariant that shared uncompressed
2684 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2685 * might ask, "if a compressed buf is allocated first, won't that be the
2686 * last thing in the list?", but in that case it's impossible to create
2687 * a shared uncompressed buf anyway (because the hdr must be compressed
2688 * to have the compressed buf). You might also think that #3 is
2689 * sufficient to make this guarantee, however it's possible
2690 * (specifically in the rare L2ARC write race mentioned in
2691 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2692 * is shareable, but wasn't at the time of its allocation. Rather than
2693 * allow a new shared uncompressed buf to be created and then shuffle
2694 * the list around to make it the last element, this simply disallows
2695 * sharing if the new buf isn't the first to be added.
2697 ASSERT3P(buf->b_hdr, ==, hdr);
2698 boolean_t hdr_compressed =
2699 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF;
2700 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2701 return (!ARC_BUF_ENCRYPTED(buf) &&
2702 buf_compressed == hdr_compressed &&
2703 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2704 !HDR_SHARED_DATA(hdr) &&
2705 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2709 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2710 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2711 * copy was made successfully, or an error code otherwise.
2714 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2715 const void *tag, boolean_t encrypted, boolean_t compressed,
2716 boolean_t noauth, boolean_t fill, arc_buf_t **ret)
2719 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2721 ASSERT(HDR_HAS_L1HDR(hdr));
2722 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2723 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2724 hdr->b_type == ARC_BUFC_METADATA);
2725 ASSERT3P(ret, !=, NULL);
2726 ASSERT3P(*ret, ==, NULL);
2727 IMPLY(encrypted, compressed);
2729 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2732 buf->b_next = hdr->b_l1hdr.b_buf;
2735 add_reference(hdr, tag);
2738 * We're about to change the hdr's b_flags. We must either
2739 * hold the hash_lock or be undiscoverable.
2741 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2744 * Only honor requests for compressed bufs if the hdr is actually
2745 * compressed. This must be overridden if the buffer is encrypted since
2746 * encrypted buffers cannot be decompressed.
2749 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2750 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2751 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2752 } else if (compressed &&
2753 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2754 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2755 flags |= ARC_FILL_COMPRESSED;
2760 flags |= ARC_FILL_NOAUTH;
2764 * If the hdr's data can be shared then we share the data buffer and
2765 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2766 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2767 * buffer to store the buf's data.
2769 * There are two additional restrictions here because we're sharing
2770 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2771 * actively involved in an L2ARC write, because if this buf is used by
2772 * an arc_write() then the hdr's data buffer will be released when the
2773 * write completes, even though the L2ARC write might still be using it.
2774 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2775 * need to be ABD-aware. It must be allocated via
2776 * zio_[data_]buf_alloc(), not as a page, because we need to be able
2777 * to abd_release_ownership_of_buf(), which isn't allowed on "linear
2778 * page" buffers because the ABD code needs to handle freeing them
2781 boolean_t can_share = arc_can_share(hdr, buf) &&
2782 !HDR_L2_WRITING(hdr) &&
2783 hdr->b_l1hdr.b_pabd != NULL &&
2784 abd_is_linear(hdr->b_l1hdr.b_pabd) &&
2785 !abd_is_linear_page(hdr->b_l1hdr.b_pabd);
2787 /* Set up b_data and sharing */
2789 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2790 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2791 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2794 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2795 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2797 VERIFY3P(buf->b_data, !=, NULL);
2799 hdr->b_l1hdr.b_buf = buf;
2802 * If the user wants the data from the hdr, we need to either copy or
2803 * decompress the data.
2806 ASSERT3P(zb, !=, NULL);
2807 return (arc_buf_fill(buf, spa, zb, flags));
2813 static const char *arc_onloan_tag = "onloan";
2816 arc_loaned_bytes_update(int64_t delta)
2818 atomic_add_64(&arc_loaned_bytes, delta);
2820 /* assert that it did not wrap around */
2821 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2825 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2826 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2827 * buffers must be returned to the arc before they can be used by the DMU or
2831 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2833 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2834 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2836 arc_loaned_bytes_update(arc_buf_size(buf));
2842 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2843 enum zio_compress compression_type, uint8_t complevel)
2845 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2846 psize, lsize, compression_type, complevel);
2848 arc_loaned_bytes_update(arc_buf_size(buf));
2854 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2855 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2856 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2857 enum zio_compress compression_type, uint8_t complevel)
2859 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2860 byteorder, salt, iv, mac, ot, psize, lsize, compression_type,
2863 atomic_add_64(&arc_loaned_bytes, psize);
2869 * Return a loaned arc buffer to the arc.
2872 arc_return_buf(arc_buf_t *buf, const void *tag)
2874 arc_buf_hdr_t *hdr = buf->b_hdr;
2876 ASSERT3P(buf->b_data, !=, NULL);
2877 ASSERT(HDR_HAS_L1HDR(hdr));
2878 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2879 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2881 arc_loaned_bytes_update(-arc_buf_size(buf));
2884 /* Detach an arc_buf from a dbuf (tag) */
2886 arc_loan_inuse_buf(arc_buf_t *buf, const void *tag)
2888 arc_buf_hdr_t *hdr = buf->b_hdr;
2890 ASSERT3P(buf->b_data, !=, NULL);
2891 ASSERT(HDR_HAS_L1HDR(hdr));
2892 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2893 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2895 arc_loaned_bytes_update(arc_buf_size(buf));
2899 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2901 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2904 df->l2df_size = size;
2905 df->l2df_type = type;
2906 mutex_enter(&l2arc_free_on_write_mtx);
2907 list_insert_head(l2arc_free_on_write, df);
2908 mutex_exit(&l2arc_free_on_write_mtx);
2912 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
2914 arc_state_t *state = hdr->b_l1hdr.b_state;
2915 arc_buf_contents_t type = arc_buf_type(hdr);
2916 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
2918 /* protected by hash lock, if in the hash table */
2919 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2920 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2921 ASSERT(state != arc_anon && state != arc_l2c_only);
2923 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2926 (void) zfs_refcount_remove_many(&state->arcs_size[type], size, hdr);
2927 if (type == ARC_BUFC_METADATA) {
2928 arc_space_return(size, ARC_SPACE_META);
2930 ASSERT(type == ARC_BUFC_DATA);
2931 arc_space_return(size, ARC_SPACE_DATA);
2935 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
2937 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2942 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2943 * data buffer, we transfer the refcount ownership to the hdr and update
2944 * the appropriate kstats.
2947 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2949 ASSERT(arc_can_share(hdr, buf));
2950 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2951 ASSERT(!ARC_BUF_ENCRYPTED(buf));
2952 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2955 * Start sharing the data buffer. We transfer the
2956 * refcount ownership to the hdr since it always owns
2957 * the refcount whenever an arc_buf_t is shared.
2959 zfs_refcount_transfer_ownership_many(
2960 &hdr->b_l1hdr.b_state->arcs_size[arc_buf_type(hdr)],
2961 arc_hdr_size(hdr), buf, hdr);
2962 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2963 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2964 HDR_ISTYPE_METADATA(hdr));
2965 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2966 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2969 * Since we've transferred ownership to the hdr we need
2970 * to increment its compressed and uncompressed kstats and
2971 * decrement the overhead size.
2973 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2974 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2975 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2979 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2981 ASSERT(arc_buf_is_shared(buf));
2982 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2983 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2986 * We are no longer sharing this buffer so we need
2987 * to transfer its ownership to the rightful owner.
2989 zfs_refcount_transfer_ownership_many(
2990 &hdr->b_l1hdr.b_state->arcs_size[arc_buf_type(hdr)],
2991 arc_hdr_size(hdr), hdr, buf);
2992 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2993 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
2994 abd_free(hdr->b_l1hdr.b_pabd);
2995 hdr->b_l1hdr.b_pabd = NULL;
2996 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2999 * Since the buffer is no longer shared between
3000 * the arc buf and the hdr, count it as overhead.
3002 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3003 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3004 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3008 * Remove an arc_buf_t from the hdr's buf list and return the last
3009 * arc_buf_t on the list. If no buffers remain on the list then return
3013 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3015 ASSERT(HDR_HAS_L1HDR(hdr));
3016 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3018 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3019 arc_buf_t *lastbuf = NULL;
3022 * Remove the buf from the hdr list and locate the last
3023 * remaining buffer on the list.
3025 while (*bufp != NULL) {
3027 *bufp = buf->b_next;
3030 * If we've removed a buffer in the middle of
3031 * the list then update the lastbuf and update
3034 if (*bufp != NULL) {
3036 bufp = &(*bufp)->b_next;
3040 ASSERT3P(lastbuf, !=, buf);
3041 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3047 * Free up buf->b_data and pull the arc_buf_t off of the arc_buf_hdr_t's
3051 arc_buf_destroy_impl(arc_buf_t *buf)
3053 arc_buf_hdr_t *hdr = buf->b_hdr;
3056 * Free up the data associated with the buf but only if we're not
3057 * sharing this with the hdr. If we are sharing it with the hdr, the
3058 * hdr is responsible for doing the free.
3060 if (buf->b_data != NULL) {
3062 * We're about to change the hdr's b_flags. We must either
3063 * hold the hash_lock or be undiscoverable.
3065 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3067 arc_cksum_verify(buf);
3068 arc_buf_unwatch(buf);
3070 if (ARC_BUF_SHARED(buf)) {
3071 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3073 ASSERT(!arc_buf_is_shared(buf));
3074 uint64_t size = arc_buf_size(buf);
3075 arc_free_data_buf(hdr, buf->b_data, size, buf);
3076 ARCSTAT_INCR(arcstat_overhead_size, -size);
3081 * If we have no more encrypted buffers and we've already
3082 * gotten a copy of the decrypted data we can free b_rabd
3083 * to save some space.
3085 if (ARC_BUF_ENCRYPTED(buf) && HDR_HAS_RABD(hdr) &&
3086 hdr->b_l1hdr.b_pabd != NULL && !HDR_IO_IN_PROGRESS(hdr)) {
3088 for (b = hdr->b_l1hdr.b_buf; b; b = b->b_next) {
3089 if (b != buf && ARC_BUF_ENCRYPTED(b))
3093 arc_hdr_free_abd(hdr, B_TRUE);
3097 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3099 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3101 * If the current arc_buf_t is sharing its data buffer with the
3102 * hdr, then reassign the hdr's b_pabd to share it with the new
3103 * buffer at the end of the list. The shared buffer is always
3104 * the last one on the hdr's buffer list.
3106 * There is an equivalent case for compressed bufs, but since
3107 * they aren't guaranteed to be the last buf in the list and
3108 * that is an exceedingly rare case, we just allow that space be
3109 * wasted temporarily. We must also be careful not to share
3110 * encrypted buffers, since they cannot be shared.
3112 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
3113 /* Only one buf can be shared at once */
3114 ASSERT(!arc_buf_is_shared(lastbuf));
3115 /* hdr is uncompressed so can't have compressed buf */
3116 ASSERT(!ARC_BUF_COMPRESSED(lastbuf));
3118 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3119 arc_hdr_free_abd(hdr, B_FALSE);
3122 * We must setup a new shared block between the
3123 * last buffer and the hdr. The data would have
3124 * been allocated by the arc buf so we need to transfer
3125 * ownership to the hdr since it's now being shared.
3127 arc_share_buf(hdr, lastbuf);
3129 } else if (HDR_SHARED_DATA(hdr)) {
3131 * Uncompressed shared buffers are always at the end
3132 * of the list. Compressed buffers don't have the
3133 * same requirements. This makes it hard to
3134 * simply assert that the lastbuf is shared so
3135 * we rely on the hdr's compression flags to determine
3136 * if we have a compressed, shared buffer.
3138 ASSERT3P(lastbuf, !=, NULL);
3139 ASSERT(arc_buf_is_shared(lastbuf) ||
3140 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3144 * Free the checksum if we're removing the last uncompressed buf from
3147 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3148 arc_cksum_free(hdr);
3151 /* clean up the buf */
3153 kmem_cache_free(buf_cache, buf);
3157 arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, int alloc_flags)
3160 boolean_t alloc_rdata = ((alloc_flags & ARC_HDR_ALLOC_RDATA) != 0);
3162 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3163 ASSERT(HDR_HAS_L1HDR(hdr));
3164 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3165 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3168 size = HDR_GET_PSIZE(hdr);
3169 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3170 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr,
3172 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3173 ARCSTAT_INCR(arcstat_raw_size, size);
3175 size = arc_hdr_size(hdr);
3176 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3177 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr,
3179 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3182 ARCSTAT_INCR(arcstat_compressed_size, size);
3183 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3187 arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3189 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3191 ASSERT(HDR_HAS_L1HDR(hdr));
3192 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3193 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3196 * If the hdr is currently being written to the l2arc then
3197 * we defer freeing the data by adding it to the l2arc_free_on_write
3198 * list. The l2arc will free the data once it's finished
3199 * writing it to the l2arc device.
3201 if (HDR_L2_WRITING(hdr)) {
3202 arc_hdr_free_on_write(hdr, free_rdata);
3203 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3204 } else if (free_rdata) {
3205 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3207 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr);
3211 hdr->b_crypt_hdr.b_rabd = NULL;
3212 ARCSTAT_INCR(arcstat_raw_size, -size);
3214 hdr->b_l1hdr.b_pabd = NULL;
3217 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3218 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3220 ARCSTAT_INCR(arcstat_compressed_size, -size);
3221 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3225 * Allocate empty anonymous ARC header. The header will get its identity
3226 * assigned and buffers attached later as part of read or write operations.
3228 * In case of read arc_read() assigns header its identify (b_dva + b_birth),
3229 * inserts it into ARC hash to become globally visible and allocates physical
3230 * (b_pabd) or raw (b_rabd) ABD buffer to read into from disk. On disk read
3231 * completion arc_read_done() allocates ARC buffer(s) as needed, potentially
3232 * sharing one of them with the physical ABD buffer.
3234 * In case of write arc_alloc_buf() allocates ARC buffer to be filled with
3235 * data. Then after compression and/or encryption arc_write_ready() allocates
3236 * and fills (or potentially shares) physical (b_pabd) or raw (b_rabd) ABD
3237 * buffer. On disk write completion arc_write_done() assigns the header its
3238 * new identity (b_dva + b_birth) and inserts into ARC hash.
3240 * In case of partial overwrite the old data is read first as described. Then
3241 * arc_release() either allocates new anonymous ARC header and moves the ARC
3242 * buffer to it, or reuses the old ARC header by discarding its identity and
3243 * removing it from ARC hash. After buffer modification normal write process
3244 * follows as described.
3246 static arc_buf_hdr_t *
3247 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3248 boolean_t protected, enum zio_compress compression_type, uint8_t complevel,
3249 arc_buf_contents_t type)
3253 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3254 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3256 ASSERT(HDR_EMPTY(hdr));
3258 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3260 HDR_SET_PSIZE(hdr, psize);
3261 HDR_SET_LSIZE(hdr, lsize);
3265 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3266 arc_hdr_set_compress(hdr, compression_type);
3267 hdr->b_complevel = complevel;
3269 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3271 hdr->b_l1hdr.b_state = arc_anon;
3272 hdr->b_l1hdr.b_arc_access = 0;
3273 hdr->b_l1hdr.b_mru_hits = 0;
3274 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3275 hdr->b_l1hdr.b_mfu_hits = 0;
3276 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3277 hdr->b_l1hdr.b_buf = NULL;
3279 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3285 * Transition between the two allocation states for the arc_buf_hdr struct.
3286 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3287 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3288 * version is used when a cache buffer is only in the L2ARC in order to reduce
3291 static arc_buf_hdr_t *
3292 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3294 ASSERT(HDR_HAS_L2HDR(hdr));
3296 arc_buf_hdr_t *nhdr;
3297 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3299 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3300 (old == hdr_l2only_cache && new == hdr_full_cache));
3302 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3304 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3305 buf_hash_remove(hdr);
3307 memcpy(nhdr, hdr, HDR_L2ONLY_SIZE);
3309 if (new == hdr_full_cache) {
3310 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3312 * arc_access and arc_change_state need to be aware that a
3313 * header has just come out of L2ARC, so we set its state to
3314 * l2c_only even though it's about to change.
3316 nhdr->b_l1hdr.b_state = arc_l2c_only;
3318 /* Verify previous threads set to NULL before freeing */
3319 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3320 ASSERT(!HDR_HAS_RABD(hdr));
3322 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3324 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3328 * If we've reached here, We must have been called from
3329 * arc_evict_hdr(), as such we should have already been
3330 * removed from any ghost list we were previously on
3331 * (which protects us from racing with arc_evict_state),
3332 * thus no locking is needed during this check.
3334 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3337 * A buffer must not be moved into the arc_l2c_only
3338 * state if it's not finished being written out to the
3339 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3340 * might try to be accessed, even though it was removed.
3342 VERIFY(!HDR_L2_WRITING(hdr));
3343 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3344 ASSERT(!HDR_HAS_RABD(hdr));
3346 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3349 * The header has been reallocated so we need to re-insert it into any
3352 (void) buf_hash_insert(nhdr, NULL);
3354 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3356 mutex_enter(&dev->l2ad_mtx);
3359 * We must place the realloc'ed header back into the list at
3360 * the same spot. Otherwise, if it's placed earlier in the list,
3361 * l2arc_write_buffers() could find it during the function's
3362 * write phase, and try to write it out to the l2arc.
3364 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3365 list_remove(&dev->l2ad_buflist, hdr);
3367 mutex_exit(&dev->l2ad_mtx);
3370 * Since we're using the pointer address as the tag when
3371 * incrementing and decrementing the l2ad_alloc refcount, we
3372 * must remove the old pointer (that we're about to destroy) and
3373 * add the new pointer to the refcount. Otherwise we'd remove
3374 * the wrong pointer address when calling arc_hdr_destroy() later.
3377 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
3378 arc_hdr_size(hdr), hdr);
3379 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
3380 arc_hdr_size(nhdr), nhdr);
3382 buf_discard_identity(hdr);
3383 kmem_cache_free(old, hdr);
3389 * This function is used by the send / receive code to convert a newly
3390 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3391 * is also used to allow the root objset block to be updated without altering
3392 * its embedded MACs. Both block types will always be uncompressed so we do not
3393 * have to worry about compression type or psize.
3396 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3397 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3400 arc_buf_hdr_t *hdr = buf->b_hdr;
3402 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3403 ASSERT(HDR_HAS_L1HDR(hdr));
3404 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3406 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3407 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3408 hdr->b_crypt_hdr.b_dsobj = dsobj;
3409 hdr->b_crypt_hdr.b_ot = ot;
3410 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3411 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3412 if (!arc_hdr_has_uncompressed_buf(hdr))
3413 arc_cksum_free(hdr);
3416 memcpy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
3418 memcpy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
3420 memcpy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
3424 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3425 * The buf is returned thawed since we expect the consumer to modify it.
3428 arc_alloc_buf(spa_t *spa, const void *tag, arc_buf_contents_t type,
3431 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3432 B_FALSE, ZIO_COMPRESS_OFF, 0, type);
3434 arc_buf_t *buf = NULL;
3435 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3436 B_FALSE, B_FALSE, &buf));
3443 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3444 * for bufs containing metadata.
3447 arc_alloc_compressed_buf(spa_t *spa, const void *tag, uint64_t psize,
3448 uint64_t lsize, enum zio_compress compression_type, uint8_t complevel)
3450 ASSERT3U(lsize, >, 0);
3451 ASSERT3U(lsize, >=, psize);
3452 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3453 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3455 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3456 B_FALSE, compression_type, complevel, ARC_BUFC_DATA);
3458 arc_buf_t *buf = NULL;
3459 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3460 B_TRUE, B_FALSE, B_FALSE, &buf));
3464 * To ensure that the hdr has the correct data in it if we call
3465 * arc_untransform() on this buf before it's been written to disk,
3466 * it's easiest if we just set up sharing between the buf and the hdr.
3468 arc_share_buf(hdr, buf);
3474 arc_alloc_raw_buf(spa_t *spa, const void *tag, uint64_t dsobj,
3475 boolean_t byteorder, const uint8_t *salt, const uint8_t *iv,
3476 const uint8_t *mac, dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3477 enum zio_compress compression_type, uint8_t complevel)
3481 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3482 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3484 ASSERT3U(lsize, >, 0);
3485 ASSERT3U(lsize, >=, psize);
3486 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3487 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3489 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3490 compression_type, complevel, type);
3492 hdr->b_crypt_hdr.b_dsobj = dsobj;
3493 hdr->b_crypt_hdr.b_ot = ot;
3494 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3495 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3496 memcpy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
3497 memcpy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
3498 memcpy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
3501 * This buffer will be considered encrypted even if the ot is not an
3502 * encrypted type. It will become authenticated instead in
3503 * arc_write_ready().
3506 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3507 B_FALSE, B_FALSE, &buf));
3514 l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
3515 boolean_t state_only)
3517 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3518 l2arc_dev_t *dev = l2hdr->b_dev;
3519 uint64_t lsize = HDR_GET_LSIZE(hdr);
3520 uint64_t psize = HDR_GET_PSIZE(hdr);
3521 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3522 arc_buf_contents_t type = hdr->b_type;
3537 /* If the buffer is a prefetch, count it as such. */
3538 if (HDR_PREFETCH(hdr)) {
3539 ARCSTAT_INCR(arcstat_l2_prefetch_asize, asize_s);
3542 * We use the value stored in the L2 header upon initial
3543 * caching in L2ARC. This value will be updated in case
3544 * an MRU/MRU_ghost buffer transitions to MFU but the L2ARC
3545 * metadata (log entry) cannot currently be updated. Having
3546 * the ARC state in the L2 header solves the problem of a
3547 * possibly absent L1 header (apparent in buffers restored
3548 * from persistent L2ARC).
3550 switch (hdr->b_l2hdr.b_arcs_state) {
3551 case ARC_STATE_MRU_GHOST:
3553 ARCSTAT_INCR(arcstat_l2_mru_asize, asize_s);
3555 case ARC_STATE_MFU_GHOST:
3557 ARCSTAT_INCR(arcstat_l2_mfu_asize, asize_s);
3567 ARCSTAT_INCR(arcstat_l2_psize, psize_s);
3568 ARCSTAT_INCR(arcstat_l2_lsize, lsize_s);
3572 ARCSTAT_INCR(arcstat_l2_bufc_data_asize, asize_s);
3574 case ARC_BUFC_METADATA:
3575 ARCSTAT_INCR(arcstat_l2_bufc_metadata_asize, asize_s);
3584 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3586 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3587 l2arc_dev_t *dev = l2hdr->b_dev;
3588 uint64_t psize = HDR_GET_PSIZE(hdr);
3589 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3591 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3592 ASSERT(HDR_HAS_L2HDR(hdr));
3594 list_remove(&dev->l2ad_buflist, hdr);
3596 l2arc_hdr_arcstats_decrement(hdr);
3597 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3599 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3601 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3605 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3607 if (HDR_HAS_L1HDR(hdr)) {
3608 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3609 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3611 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3612 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3614 if (HDR_HAS_L2HDR(hdr)) {
3615 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3616 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3619 mutex_enter(&dev->l2ad_mtx);
3622 * Even though we checked this conditional above, we
3623 * need to check this again now that we have the
3624 * l2ad_mtx. This is because we could be racing with
3625 * another thread calling l2arc_evict() which might have
3626 * destroyed this header's L2 portion as we were waiting
3627 * to acquire the l2ad_mtx. If that happens, we don't
3628 * want to re-destroy the header's L2 portion.
3630 if (HDR_HAS_L2HDR(hdr)) {
3632 if (!HDR_EMPTY(hdr))
3633 buf_discard_identity(hdr);
3635 arc_hdr_l2hdr_destroy(hdr);
3639 mutex_exit(&dev->l2ad_mtx);
3643 * The header's identify can only be safely discarded once it is no
3644 * longer discoverable. This requires removing it from the hash table
3645 * and the l2arc header list. After this point the hash lock can not
3646 * be used to protect the header.
3648 if (!HDR_EMPTY(hdr))
3649 buf_discard_identity(hdr);
3651 if (HDR_HAS_L1HDR(hdr)) {
3652 arc_cksum_free(hdr);
3654 while (hdr->b_l1hdr.b_buf != NULL)
3655 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3657 if (hdr->b_l1hdr.b_pabd != NULL)
3658 arc_hdr_free_abd(hdr, B_FALSE);
3660 if (HDR_HAS_RABD(hdr))
3661 arc_hdr_free_abd(hdr, B_TRUE);
3664 ASSERT3P(hdr->b_hash_next, ==, NULL);
3665 if (HDR_HAS_L1HDR(hdr)) {
3666 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3667 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3669 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3671 kmem_cache_free(hdr_full_cache, hdr);
3673 kmem_cache_free(hdr_l2only_cache, hdr);
3678 arc_buf_destroy(arc_buf_t *buf, const void *tag)
3680 arc_buf_hdr_t *hdr = buf->b_hdr;
3682 if (hdr->b_l1hdr.b_state == arc_anon) {
3683 ASSERT3P(hdr->b_l1hdr.b_buf, ==, buf);
3684 ASSERT(ARC_BUF_LAST(buf));
3685 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3686 VERIFY0(remove_reference(hdr, tag));
3690 kmutex_t *hash_lock = HDR_LOCK(hdr);
3691 mutex_enter(hash_lock);
3693 ASSERT3P(hdr, ==, buf->b_hdr);
3694 ASSERT3P(hdr->b_l1hdr.b_buf, !=, NULL);
3695 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3696 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3697 ASSERT3P(buf->b_data, !=, NULL);
3699 arc_buf_destroy_impl(buf);
3700 (void) remove_reference(hdr, tag);
3701 mutex_exit(hash_lock);
3705 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3706 * state of the header is dependent on its state prior to entering this
3707 * function. The following transitions are possible:
3709 * - arc_mru -> arc_mru_ghost
3710 * - arc_mfu -> arc_mfu_ghost
3711 * - arc_mru_ghost -> arc_l2c_only
3712 * - arc_mru_ghost -> deleted
3713 * - arc_mfu_ghost -> arc_l2c_only
3714 * - arc_mfu_ghost -> deleted
3715 * - arc_uncached -> deleted
3717 * Return total size of evicted data buffers for eviction progress tracking.
3718 * When evicting from ghost states return logical buffer size to make eviction
3719 * progress at the same (or at least comparable) rate as from non-ghost states.
3721 * Return *real_evicted for actual ARC size reduction to wake up threads
3722 * waiting for it. For non-ghost states it includes size of evicted data
3723 * buffers (the headers are not freed there). For ghost states it includes
3724 * only the evicted headers size.
3727 arc_evict_hdr(arc_buf_hdr_t *hdr, uint64_t *real_evicted)
3729 arc_state_t *evicted_state, *state;
3730 int64_t bytes_evicted = 0;
3731 uint_t min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3732 arc_min_prescient_prefetch_ms : arc_min_prefetch_ms;
3734 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3735 ASSERT(HDR_HAS_L1HDR(hdr));
3736 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3737 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3738 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3741 state = hdr->b_l1hdr.b_state;
3742 if (GHOST_STATE(state)) {
3745 * l2arc_write_buffers() relies on a header's L1 portion
3746 * (i.e. its b_pabd field) during it's write phase.
3747 * Thus, we cannot push a header onto the arc_l2c_only
3748 * state (removing its L1 piece) until the header is
3749 * done being written to the l2arc.
3751 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3752 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3753 return (bytes_evicted);
3756 ARCSTAT_BUMP(arcstat_deleted);
3757 bytes_evicted += HDR_GET_LSIZE(hdr);
3759 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3761 if (HDR_HAS_L2HDR(hdr)) {
3762 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3763 ASSERT(!HDR_HAS_RABD(hdr));
3765 * This buffer is cached on the 2nd Level ARC;
3766 * don't destroy the header.
3768 arc_change_state(arc_l2c_only, hdr);
3770 * dropping from L1+L2 cached to L2-only,
3771 * realloc to remove the L1 header.
3773 (void) arc_hdr_realloc(hdr, hdr_full_cache,
3775 *real_evicted += HDR_FULL_SIZE - HDR_L2ONLY_SIZE;
3777 arc_change_state(arc_anon, hdr);
3778 arc_hdr_destroy(hdr);
3779 *real_evicted += HDR_FULL_SIZE;
3781 return (bytes_evicted);
3784 ASSERT(state == arc_mru || state == arc_mfu || state == arc_uncached);
3785 evicted_state = (state == arc_uncached) ? arc_anon :
3786 ((state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost);
3788 /* prefetch buffers have a minimum lifespan */
3789 if ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3790 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3791 MSEC_TO_TICK(min_lifetime)) {
3792 ARCSTAT_BUMP(arcstat_evict_skip);
3793 return (bytes_evicted);
3796 if (HDR_HAS_L2HDR(hdr)) {
3797 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3799 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3800 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3801 HDR_GET_LSIZE(hdr));
3803 switch (state->arcs_state) {
3806 arcstat_evict_l2_eligible_mru,
3807 HDR_GET_LSIZE(hdr));
3811 arcstat_evict_l2_eligible_mfu,
3812 HDR_GET_LSIZE(hdr));
3818 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3819 HDR_GET_LSIZE(hdr));
3823 bytes_evicted += arc_hdr_size(hdr);
3824 *real_evicted += arc_hdr_size(hdr);
3827 * If this hdr is being evicted and has a compressed buffer then we
3828 * discard it here before we change states. This ensures that the
3829 * accounting is updated correctly in arc_free_data_impl().
3831 if (hdr->b_l1hdr.b_pabd != NULL)
3832 arc_hdr_free_abd(hdr, B_FALSE);
3834 if (HDR_HAS_RABD(hdr))
3835 arc_hdr_free_abd(hdr, B_TRUE);
3837 arc_change_state(evicted_state, hdr);
3838 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3839 if (evicted_state == arc_anon) {
3840 arc_hdr_destroy(hdr);
3841 *real_evicted += HDR_FULL_SIZE;
3843 ASSERT(HDR_IN_HASH_TABLE(hdr));
3846 return (bytes_evicted);
3850 arc_set_need_free(void)
3852 ASSERT(MUTEX_HELD(&arc_evict_lock));
3853 int64_t remaining = arc_free_memory() - arc_sys_free / 2;
3854 arc_evict_waiter_t *aw = list_tail(&arc_evict_waiters);
3856 arc_need_free = MAX(-remaining, 0);
3859 MAX(-remaining, (int64_t)(aw->aew_count - arc_evict_count));
3864 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3865 uint64_t spa, uint64_t bytes)
3867 multilist_sublist_t *mls;
3868 uint64_t bytes_evicted = 0, real_evicted = 0;
3870 kmutex_t *hash_lock;
3871 uint_t evict_count = zfs_arc_evict_batch_limit;
3873 ASSERT3P(marker, !=, NULL);
3875 mls = multilist_sublist_lock(ml, idx);
3877 for (hdr = multilist_sublist_prev(mls, marker); likely(hdr != NULL);
3878 hdr = multilist_sublist_prev(mls, marker)) {
3879 if ((evict_count == 0) || (bytes_evicted >= bytes))
3883 * To keep our iteration location, move the marker
3884 * forward. Since we're not holding hdr's hash lock, we
3885 * must be very careful and not remove 'hdr' from the
3886 * sublist. Otherwise, other consumers might mistake the
3887 * 'hdr' as not being on a sublist when they call the
3888 * multilist_link_active() function (they all rely on
3889 * the hash lock protecting concurrent insertions and
3890 * removals). multilist_sublist_move_forward() was
3891 * specifically implemented to ensure this is the case
3892 * (only 'marker' will be removed and re-inserted).
3894 multilist_sublist_move_forward(mls, marker);
3897 * The only case where the b_spa field should ever be
3898 * zero, is the marker headers inserted by
3899 * arc_evict_state(). It's possible for multiple threads
3900 * to be calling arc_evict_state() concurrently (e.g.
3901 * dsl_pool_close() and zio_inject_fault()), so we must
3902 * skip any markers we see from these other threads.
3904 if (hdr->b_spa == 0)
3907 /* we're only interested in evicting buffers of a certain spa */
3908 if (spa != 0 && hdr->b_spa != spa) {
3909 ARCSTAT_BUMP(arcstat_evict_skip);
3913 hash_lock = HDR_LOCK(hdr);
3916 * We aren't calling this function from any code path
3917 * that would already be holding a hash lock, so we're
3918 * asserting on this assumption to be defensive in case
3919 * this ever changes. Without this check, it would be
3920 * possible to incorrectly increment arcstat_mutex_miss
3921 * below (e.g. if the code changed such that we called
3922 * this function with a hash lock held).
3924 ASSERT(!MUTEX_HELD(hash_lock));
3926 if (mutex_tryenter(hash_lock)) {
3928 uint64_t evicted = arc_evict_hdr(hdr, &revicted);
3929 mutex_exit(hash_lock);
3931 bytes_evicted += evicted;
3932 real_evicted += revicted;
3935 * If evicted is zero, arc_evict_hdr() must have
3936 * decided to skip this header, don't increment
3937 * evict_count in this case.
3943 ARCSTAT_BUMP(arcstat_mutex_miss);
3947 multilist_sublist_unlock(mls);
3950 * Increment the count of evicted bytes, and wake up any threads that
3951 * are waiting for the count to reach this value. Since the list is
3952 * ordered by ascending aew_count, we pop off the beginning of the
3953 * list until we reach the end, or a waiter that's past the current
3954 * "count". Doing this outside the loop reduces the number of times
3955 * we need to acquire the global arc_evict_lock.
3957 * Only wake when there's sufficient free memory in the system
3958 * (specifically, arc_sys_free/2, which by default is a bit more than
3959 * 1/64th of RAM). See the comments in arc_wait_for_eviction().
3961 mutex_enter(&arc_evict_lock);
3962 arc_evict_count += real_evicted;
3964 if (arc_free_memory() > arc_sys_free / 2) {
3965 arc_evict_waiter_t *aw;
3966 while ((aw = list_head(&arc_evict_waiters)) != NULL &&
3967 aw->aew_count <= arc_evict_count) {
3968 list_remove(&arc_evict_waiters, aw);
3969 cv_broadcast(&aw->aew_cv);
3972 arc_set_need_free();
3973 mutex_exit(&arc_evict_lock);
3976 * If the ARC size is reduced from arc_c_max to arc_c_min (especially
3977 * if the average cached block is small), eviction can be on-CPU for
3978 * many seconds. To ensure that other threads that may be bound to
3979 * this CPU are able to make progress, make a voluntary preemption
3982 kpreempt(KPREEMPT_SYNC);
3984 return (bytes_evicted);
3988 * Allocate an array of buffer headers used as placeholders during arc state
3991 static arc_buf_hdr_t **
3992 arc_state_alloc_markers(int count)
3994 arc_buf_hdr_t **markers;
3996 markers = kmem_zalloc(sizeof (*markers) * count, KM_SLEEP);
3997 for (int i = 0; i < count; i++) {
3998 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
4001 * A b_spa of 0 is used to indicate that this header is
4002 * a marker. This fact is used in arc_evict_state_impl().
4004 markers[i]->b_spa = 0;
4011 arc_state_free_markers(arc_buf_hdr_t **markers, int count)
4013 for (int i = 0; i < count; i++)
4014 kmem_cache_free(hdr_full_cache, markers[i]);
4015 kmem_free(markers, sizeof (*markers) * count);
4019 * Evict buffers from the given arc state, until we've removed the
4020 * specified number of bytes. Move the removed buffers to the
4021 * appropriate evict state.
4023 * This function makes a "best effort". It skips over any buffers
4024 * it can't get a hash_lock on, and so, may not catch all candidates.
4025 * It may also return without evicting as much space as requested.
4027 * If bytes is specified using the special value ARC_EVICT_ALL, this
4028 * will evict all available (i.e. unlocked and evictable) buffers from
4029 * the given arc state; which is used by arc_flush().
4032 arc_evict_state(arc_state_t *state, arc_buf_contents_t type, uint64_t spa,
4035 uint64_t total_evicted = 0;
4036 multilist_t *ml = &state->arcs_list[type];
4038 arc_buf_hdr_t **markers;
4040 num_sublists = multilist_get_num_sublists(ml);
4043 * If we've tried to evict from each sublist, made some
4044 * progress, but still have not hit the target number of bytes
4045 * to evict, we want to keep trying. The markers allow us to
4046 * pick up where we left off for each individual sublist, rather
4047 * than starting from the tail each time.
4049 if (zthr_iscurthread(arc_evict_zthr)) {
4050 markers = arc_state_evict_markers;
4051 ASSERT3S(num_sublists, <=, arc_state_evict_marker_count);
4053 markers = arc_state_alloc_markers(num_sublists);
4055 for (int i = 0; i < num_sublists; i++) {
4056 multilist_sublist_t *mls;
4058 mls = multilist_sublist_lock(ml, i);
4059 multilist_sublist_insert_tail(mls, markers[i]);
4060 multilist_sublist_unlock(mls);
4064 * While we haven't hit our target number of bytes to evict, or
4065 * we're evicting all available buffers.
4067 while (total_evicted < bytes) {
4068 int sublist_idx = multilist_get_random_index(ml);
4069 uint64_t scan_evicted = 0;
4072 * Start eviction using a randomly selected sublist,
4073 * this is to try and evenly balance eviction across all
4074 * sublists. Always starting at the same sublist
4075 * (e.g. index 0) would cause evictions to favor certain
4076 * sublists over others.
4078 for (int i = 0; i < num_sublists; i++) {
4079 uint64_t bytes_remaining;
4080 uint64_t bytes_evicted;
4082 if (total_evicted < bytes)
4083 bytes_remaining = bytes - total_evicted;
4087 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4088 markers[sublist_idx], spa, bytes_remaining);
4090 scan_evicted += bytes_evicted;
4091 total_evicted += bytes_evicted;
4093 /* we've reached the end, wrap to the beginning */
4094 if (++sublist_idx >= num_sublists)
4099 * If we didn't evict anything during this scan, we have
4100 * no reason to believe we'll evict more during another
4101 * scan, so break the loop.
4103 if (scan_evicted == 0) {
4104 /* This isn't possible, let's make that obvious */
4105 ASSERT3S(bytes, !=, 0);
4108 * When bytes is ARC_EVICT_ALL, the only way to
4109 * break the loop is when scan_evicted is zero.
4110 * In that case, we actually have evicted enough,
4111 * so we don't want to increment the kstat.
4113 if (bytes != ARC_EVICT_ALL) {
4114 ASSERT3S(total_evicted, <, bytes);
4115 ARCSTAT_BUMP(arcstat_evict_not_enough);
4122 for (int i = 0; i < num_sublists; i++) {
4123 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4124 multilist_sublist_remove(mls, markers[i]);
4125 multilist_sublist_unlock(mls);
4127 if (markers != arc_state_evict_markers)
4128 arc_state_free_markers(markers, num_sublists);
4130 return (total_evicted);
4134 * Flush all "evictable" data of the given type from the arc state
4135 * specified. This will not evict any "active" buffers (i.e. referenced).
4137 * When 'retry' is set to B_FALSE, the function will make a single pass
4138 * over the state and evict any buffers that it can. Since it doesn't
4139 * continually retry the eviction, it might end up leaving some buffers
4140 * in the ARC due to lock misses.
4142 * When 'retry' is set to B_TRUE, the function will continually retry the
4143 * eviction until *all* evictable buffers have been removed from the
4144 * state. As a result, if concurrent insertions into the state are
4145 * allowed (e.g. if the ARC isn't shutting down), this function might
4146 * wind up in an infinite loop, continually trying to evict buffers.
4149 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4152 uint64_t evicted = 0;
4154 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4155 evicted += arc_evict_state(state, type, spa, ARC_EVICT_ALL);
4165 * Evict the specified number of bytes from the state specified. This
4166 * function prevents us from trying to evict more from a state's list
4167 * than is "evictable", and to skip evicting altogether when passed a
4168 * negative value for "bytes". In contrast, arc_evict_state() will
4169 * evict everything it can, when passed a negative value for "bytes".
4172 arc_evict_impl(arc_state_t *state, arc_buf_contents_t type, int64_t bytes)
4176 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4177 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4179 return (arc_evict_state(state, type, 0, delta));
4186 * Adjust specified fraction, taking into account initial ghost state(s) size,
4187 * ghost hit bytes towards increasing the fraction, ghost hit bytes towards
4188 * decreasing it, plus a balance factor, controlling the decrease rate, used
4189 * to balance metadata vs data.
4192 arc_evict_adj(uint64_t frac, uint64_t total, uint64_t up, uint64_t down,
4195 if (total < 8 || up + down == 0)
4199 * We should not have more ghost hits than ghost size, but they
4200 * may get close. Restrict maximum adjustment in that case.
4202 if (up + down >= total / 4) {
4203 uint64_t scale = (up + down) / (total / 8);
4208 /* Get maximal dynamic range by choosing optimal shifts. */
4209 int s = highbit64(total);
4210 s = MIN(64 - s, 32);
4212 uint64_t ofrac = (1ULL << 32) - frac;
4214 if (frac >= 4 * ofrac)
4215 up /= frac / (2 * ofrac + 1);
4216 up = (up << s) / (total >> (32 - s));
4217 if (ofrac >= 4 * frac)
4218 down /= ofrac / (2 * frac + 1);
4219 down = (down << s) / (total >> (32 - s));
4220 down = down * 100 / balance;
4222 return (frac + up - down);
4226 * Evict buffers from the cache, such that arcstat_size is capped by arc_c.
4231 uint64_t asize, bytes, total_evicted = 0;
4232 int64_t e, mrud, mrum, mfud, mfum, w;
4233 static uint64_t ogrd, ogrm, ogfd, ogfm;
4234 static uint64_t gsrd, gsrm, gsfd, gsfm;
4235 uint64_t ngrd, ngrm, ngfd, ngfm;
4237 /* Get current size of ARC states we can evict from. */
4238 mrud = zfs_refcount_count(&arc_mru->arcs_size[ARC_BUFC_DATA]) +
4239 zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_DATA]);
4240 mrum = zfs_refcount_count(&arc_mru->arcs_size[ARC_BUFC_METADATA]) +
4241 zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_METADATA]);
4242 mfud = zfs_refcount_count(&arc_mfu->arcs_size[ARC_BUFC_DATA]);
4243 mfum = zfs_refcount_count(&arc_mfu->arcs_size[ARC_BUFC_METADATA]);
4244 uint64_t d = mrud + mfud;
4245 uint64_t m = mrum + mfum;
4248 /* Get ARC ghost hits since last eviction. */
4249 ngrd = wmsum_value(&arc_mru_ghost->arcs_hits[ARC_BUFC_DATA]);
4250 uint64_t grd = ngrd - ogrd;
4252 ngrm = wmsum_value(&arc_mru_ghost->arcs_hits[ARC_BUFC_METADATA]);
4253 uint64_t grm = ngrm - ogrm;
4255 ngfd = wmsum_value(&arc_mfu_ghost->arcs_hits[ARC_BUFC_DATA]);
4256 uint64_t gfd = ngfd - ogfd;
4258 ngfm = wmsum_value(&arc_mfu_ghost->arcs_hits[ARC_BUFC_METADATA]);
4259 uint64_t gfm = ngfm - ogfm;
4262 /* Adjust ARC states balance based on ghost hits. */
4263 arc_meta = arc_evict_adj(arc_meta, gsrd + gsrm + gsfd + gsfm,
4264 grm + gfm, grd + gfd, zfs_arc_meta_balance);
4265 arc_pd = arc_evict_adj(arc_pd, gsrd + gsfd, grd, gfd, 100);
4266 arc_pm = arc_evict_adj(arc_pm, gsrm + gsfm, grm, gfm, 100);
4268 asize = aggsum_value(&arc_sums.arcstat_size);
4269 int64_t wt = t - (asize - arc_c);
4272 * Try to reduce pinned dnodes if more than 3/4 of wanted metadata
4273 * target is not evictable or if they go over arc_dnode_limit.
4276 int64_t dn = wmsum_value(&arc_sums.arcstat_dnode_size);
4277 w = wt * (int64_t)(arc_meta >> 16) >> 16;
4278 if (zfs_refcount_count(&arc_mru->arcs_size[ARC_BUFC_METADATA]) +
4279 zfs_refcount_count(&arc_mfu->arcs_size[ARC_BUFC_METADATA]) -
4280 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA]) -
4281 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]) >
4283 prune = dn / sizeof (dnode_t) *
4284 zfs_arc_dnode_reduce_percent / 100;
4285 } else if (dn > arc_dnode_limit) {
4286 prune = (dn - arc_dnode_limit) / sizeof (dnode_t) *
4287 zfs_arc_dnode_reduce_percent / 100;
4290 arc_prune_async(prune);
4292 /* Evict MRU metadata. */
4293 w = wt * (int64_t)(arc_meta * arc_pm >> 48) >> 16;
4294 e = MIN((int64_t)(asize - arc_c), (int64_t)(mrum - w));
4295 bytes = arc_evict_impl(arc_mru, ARC_BUFC_METADATA, e);
4296 total_evicted += bytes;
4300 /* Evict MFU metadata. */
4301 w = wt * (int64_t)(arc_meta >> 16) >> 16;
4302 e = MIN((int64_t)(asize - arc_c), (int64_t)(m - w));
4303 bytes = arc_evict_impl(arc_mfu, ARC_BUFC_METADATA, e);
4304 total_evicted += bytes;
4308 /* Evict MRU data. */
4309 wt -= m - total_evicted;
4310 w = wt * (int64_t)(arc_pd >> 16) >> 16;
4311 e = MIN((int64_t)(asize - arc_c), (int64_t)(mrud - w));
4312 bytes = arc_evict_impl(arc_mru, ARC_BUFC_DATA, e);
4313 total_evicted += bytes;
4317 /* Evict MFU data. */
4319 bytes = arc_evict_impl(arc_mfu, ARC_BUFC_DATA, e);
4321 total_evicted += bytes;
4326 * Size of each state's ghost list represents how much that state
4327 * may grow by shrinking the other states. Would it need to shrink
4328 * other states to zero (that is unlikely), its ghost size would be
4329 * equal to sum of other three state sizes. But excessive ghost
4330 * size may result in false ghost hits (too far back), that may
4331 * never result in real cache hits if several states are competing.
4332 * So choose some arbitraty point of 1/2 of other state sizes.
4334 gsrd = (mrum + mfud + mfum) / 2;
4335 e = zfs_refcount_count(&arc_mru_ghost->arcs_size[ARC_BUFC_DATA]) -
4337 (void) arc_evict_impl(arc_mru_ghost, ARC_BUFC_DATA, e);
4339 gsrm = (mrud + mfud + mfum) / 2;
4340 e = zfs_refcount_count(&arc_mru_ghost->arcs_size[ARC_BUFC_METADATA]) -
4342 (void) arc_evict_impl(arc_mru_ghost, ARC_BUFC_METADATA, e);
4344 gsfd = (mrud + mrum + mfum) / 2;
4345 e = zfs_refcount_count(&arc_mfu_ghost->arcs_size[ARC_BUFC_DATA]) -
4347 (void) arc_evict_impl(arc_mfu_ghost, ARC_BUFC_DATA, e);
4349 gsfm = (mrud + mrum + mfud) / 2;
4350 e = zfs_refcount_count(&arc_mfu_ghost->arcs_size[ARC_BUFC_METADATA]) -
4352 (void) arc_evict_impl(arc_mfu_ghost, ARC_BUFC_METADATA, e);
4354 return (total_evicted);
4358 arc_flush(spa_t *spa, boolean_t retry)
4363 * If retry is B_TRUE, a spa must not be specified since we have
4364 * no good way to determine if all of a spa's buffers have been
4365 * evicted from an arc state.
4367 ASSERT(!retry || spa == NULL);
4370 guid = spa_load_guid(spa);
4372 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4373 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4375 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4376 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4378 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4379 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4381 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4382 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4384 (void) arc_flush_state(arc_uncached, guid, ARC_BUFC_DATA, retry);
4385 (void) arc_flush_state(arc_uncached, guid, ARC_BUFC_METADATA, retry);
4389 arc_reduce_target_size(int64_t to_free)
4397 * All callers want the ARC to actually evict (at least) this much
4398 * memory. Therefore we reduce from the lower of the current size and
4399 * the target size. This way, even if arc_c is much higher than
4400 * arc_size (as can be the case after many calls to arc_freed(), we will
4401 * immediately have arc_c < arc_size and therefore the arc_evict_zthr
4404 uint64_t asize = aggsum_value(&arc_sums.arcstat_size);
4406 to_free += c - asize;
4407 arc_c = MAX((int64_t)c - to_free, (int64_t)arc_c_min);
4409 /* See comment in arc_evict_cb_check() on why lock+flag */
4410 mutex_enter(&arc_evict_lock);
4411 arc_evict_needed = B_TRUE;
4412 mutex_exit(&arc_evict_lock);
4413 zthr_wakeup(arc_evict_zthr);
4417 * Determine if the system is under memory pressure and is asking
4418 * to reclaim memory. A return value of B_TRUE indicates that the system
4419 * is under memory pressure and that the arc should adjust accordingly.
4422 arc_reclaim_needed(void)
4424 return (arc_available_memory() < 0);
4428 arc_kmem_reap_soon(void)
4431 kmem_cache_t *prev_cache = NULL;
4432 kmem_cache_t *prev_data_cache = NULL;
4437 * Reclaim unused memory from all kmem caches.
4443 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4445 /* reach upper limit of cache size on 32-bit */
4446 if (zio_buf_cache[i] == NULL)
4449 if (zio_buf_cache[i] != prev_cache) {
4450 prev_cache = zio_buf_cache[i];
4451 kmem_cache_reap_now(zio_buf_cache[i]);
4453 if (zio_data_buf_cache[i] != prev_data_cache) {
4454 prev_data_cache = zio_data_buf_cache[i];
4455 kmem_cache_reap_now(zio_data_buf_cache[i]);
4458 kmem_cache_reap_now(buf_cache);
4459 kmem_cache_reap_now(hdr_full_cache);
4460 kmem_cache_reap_now(hdr_l2only_cache);
4461 kmem_cache_reap_now(zfs_btree_leaf_cache);
4462 abd_cache_reap_now();
4466 arc_evict_cb_check(void *arg, zthr_t *zthr)
4468 (void) arg, (void) zthr;
4472 * This is necessary in order to keep the kstat information
4473 * up to date for tools that display kstat data such as the
4474 * mdb ::arc dcmd and the Linux crash utility. These tools
4475 * typically do not call kstat's update function, but simply
4476 * dump out stats from the most recent update. Without
4477 * this call, these commands may show stale stats for the
4478 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4479 * with this call, the data might be out of date if the
4480 * evict thread hasn't been woken recently; but that should
4481 * suffice. The arc_state_t structures can be queried
4482 * directly if more accurate information is needed.
4484 if (arc_ksp != NULL)
4485 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4489 * We have to rely on arc_wait_for_eviction() to tell us when to
4490 * evict, rather than checking if we are overflowing here, so that we
4491 * are sure to not leave arc_wait_for_eviction() waiting on aew_cv.
4492 * If we have become "not overflowing" since arc_wait_for_eviction()
4493 * checked, we need to wake it up. We could broadcast the CV here,
4494 * but arc_wait_for_eviction() may have not yet gone to sleep. We
4495 * would need to use a mutex to ensure that this function doesn't
4496 * broadcast until arc_wait_for_eviction() has gone to sleep (e.g.
4497 * the arc_evict_lock). However, the lock ordering of such a lock
4498 * would necessarily be incorrect with respect to the zthr_lock,
4499 * which is held before this function is called, and is held by
4500 * arc_wait_for_eviction() when it calls zthr_wakeup().
4502 if (arc_evict_needed)
4506 * If we have buffers in uncached state, evict them periodically.
4508 return ((zfs_refcount_count(&arc_uncached->arcs_esize[ARC_BUFC_DATA]) +
4509 zfs_refcount_count(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]) &&
4510 ddi_get_lbolt() - arc_last_uncached_flush >
4511 MSEC_TO_TICK(arc_min_prefetch_ms / 2)));
4515 * Keep arc_size under arc_c by running arc_evict which evicts data
4519 arc_evict_cb(void *arg, zthr_t *zthr)
4523 uint64_t evicted = 0;
4524 fstrans_cookie_t cookie = spl_fstrans_mark();
4526 /* Always try to evict from uncached state. */
4527 arc_last_uncached_flush = ddi_get_lbolt();
4528 evicted += arc_flush_state(arc_uncached, 0, ARC_BUFC_DATA, B_FALSE);
4529 evicted += arc_flush_state(arc_uncached, 0, ARC_BUFC_METADATA, B_FALSE);
4531 /* Evict from other states only if told to. */
4532 if (arc_evict_needed)
4533 evicted += arc_evict();
4536 * If evicted is zero, we couldn't evict anything
4537 * via arc_evict(). This could be due to hash lock
4538 * collisions, but more likely due to the majority of
4539 * arc buffers being unevictable. Therefore, even if
4540 * arc_size is above arc_c, another pass is unlikely to
4541 * be helpful and could potentially cause us to enter an
4542 * infinite loop. Additionally, zthr_iscancelled() is
4543 * checked here so that if the arc is shutting down, the
4544 * broadcast will wake any remaining arc evict waiters.
4546 * Note we cancel using zthr instead of arc_evict_zthr
4547 * because the latter may not yet be initializd when the
4548 * callback is first invoked.
4550 mutex_enter(&arc_evict_lock);
4551 arc_evict_needed = !zthr_iscancelled(zthr) &&
4552 evicted > 0 && aggsum_compare(&arc_sums.arcstat_size, arc_c) > 0;
4553 if (!arc_evict_needed) {
4555 * We're either no longer overflowing, or we
4556 * can't evict anything more, so we should wake
4557 * arc_get_data_impl() sooner.
4559 arc_evict_waiter_t *aw;
4560 while ((aw = list_remove_head(&arc_evict_waiters)) != NULL) {
4561 cv_broadcast(&aw->aew_cv);
4563 arc_set_need_free();
4565 mutex_exit(&arc_evict_lock);
4566 spl_fstrans_unmark(cookie);
4570 arc_reap_cb_check(void *arg, zthr_t *zthr)
4572 (void) arg, (void) zthr;
4574 int64_t free_memory = arc_available_memory();
4575 static int reap_cb_check_counter = 0;
4578 * If a kmem reap is already active, don't schedule more. We must
4579 * check for this because kmem_cache_reap_soon() won't actually
4580 * block on the cache being reaped (this is to prevent callers from
4581 * becoming implicitly blocked by a system-wide kmem reap -- which,
4582 * on a system with many, many full magazines, can take minutes).
4584 if (!kmem_cache_reap_active() && free_memory < 0) {
4586 arc_no_grow = B_TRUE;
4589 * Wait at least zfs_grow_retry (default 5) seconds
4590 * before considering growing.
4592 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4594 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4595 arc_no_grow = B_TRUE;
4596 } else if (gethrtime() >= arc_growtime) {
4597 arc_no_grow = B_FALSE;
4601 * Called unconditionally every 60 seconds to reclaim unused
4602 * zstd compression and decompression context. This is done
4603 * here to avoid the need for an independent thread.
4605 if (!((reap_cb_check_counter++) % 60))
4606 zfs_zstd_cache_reap_now();
4612 * Keep enough free memory in the system by reaping the ARC's kmem
4613 * caches. To cause more slabs to be reapable, we may reduce the
4614 * target size of the cache (arc_c), causing the arc_evict_cb()
4615 * to free more buffers.
4618 arc_reap_cb(void *arg, zthr_t *zthr)
4620 (void) arg, (void) zthr;
4622 int64_t free_memory;
4623 fstrans_cookie_t cookie = spl_fstrans_mark();
4626 * Kick off asynchronous kmem_reap()'s of all our caches.
4628 arc_kmem_reap_soon();
4631 * Wait at least arc_kmem_cache_reap_retry_ms between
4632 * arc_kmem_reap_soon() calls. Without this check it is possible to
4633 * end up in a situation where we spend lots of time reaping
4634 * caches, while we're near arc_c_min. Waiting here also gives the
4635 * subsequent free memory check a chance of finding that the
4636 * asynchronous reap has already freed enough memory, and we don't
4637 * need to call arc_reduce_target_size().
4639 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
4642 * Reduce the target size as needed to maintain the amount of free
4643 * memory in the system at a fraction of the arc_size (1/128th by
4644 * default). If oversubscribed (free_memory < 0) then reduce the
4645 * target arc_size by the deficit amount plus the fractional
4646 * amount. If free memory is positive but less than the fractional
4647 * amount, reduce by what is needed to hit the fractional amount.
4649 free_memory = arc_available_memory();
4651 int64_t can_free = arc_c - arc_c_min;
4653 int64_t to_free = (can_free >> arc_shrink_shift) - free_memory;
4655 arc_reduce_target_size(to_free);
4657 spl_fstrans_unmark(cookie);
4662 * Determine the amount of memory eligible for eviction contained in the
4663 * ARC. All clean data reported by the ghost lists can always be safely
4664 * evicted. Due to arc_c_min, the same does not hold for all clean data
4665 * contained by the regular mru and mfu lists.
4667 * In the case of the regular mru and mfu lists, we need to report as
4668 * much clean data as possible, such that evicting that same reported
4669 * data will not bring arc_size below arc_c_min. Thus, in certain
4670 * circumstances, the total amount of clean data in the mru and mfu
4671 * lists might not actually be evictable.
4673 * The following two distinct cases are accounted for:
4675 * 1. The sum of the amount of dirty data contained by both the mru and
4676 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4677 * is greater than or equal to arc_c_min.
4678 * (i.e. amount of dirty data >= arc_c_min)
4680 * This is the easy case; all clean data contained by the mru and mfu
4681 * lists is evictable. Evicting all clean data can only drop arc_size
4682 * to the amount of dirty data, which is greater than arc_c_min.
4684 * 2. The sum of the amount of dirty data contained by both the mru and
4685 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4686 * is less than arc_c_min.
4687 * (i.e. arc_c_min > amount of dirty data)
4689 * 2.1. arc_size is greater than or equal arc_c_min.
4690 * (i.e. arc_size >= arc_c_min > amount of dirty data)
4692 * In this case, not all clean data from the regular mru and mfu
4693 * lists is actually evictable; we must leave enough clean data
4694 * to keep arc_size above arc_c_min. Thus, the maximum amount of
4695 * evictable data from the two lists combined, is exactly the
4696 * difference between arc_size and arc_c_min.
4698 * 2.2. arc_size is less than arc_c_min
4699 * (i.e. arc_c_min > arc_size > amount of dirty data)
4701 * In this case, none of the data contained in the mru and mfu
4702 * lists is evictable, even if it's clean. Since arc_size is
4703 * already below arc_c_min, evicting any more would only
4704 * increase this negative difference.
4707 #endif /* _KERNEL */
4710 * Adapt arc info given the number of bytes we are trying to add and
4711 * the state that we are coming from. This function is only called
4712 * when we are adding new content to the cache.
4715 arc_adapt(uint64_t bytes)
4718 * Wake reap thread if we do not have any available memory
4720 if (arc_reclaim_needed()) {
4721 zthr_wakeup(arc_reap_zthr);
4728 if (arc_c >= arc_c_max)
4732 * If we're within (2 * maxblocksize) bytes of the target
4733 * cache size, increment the target cache size
4735 if (aggsum_upper_bound(&arc_sums.arcstat_size) +
4736 2 * SPA_MAXBLOCKSIZE >= arc_c) {
4737 uint64_t dc = MAX(bytes, SPA_OLD_MAXBLOCKSIZE);
4738 if (atomic_add_64_nv(&arc_c, dc) > arc_c_max)
4744 * Check if arc_size has grown past our upper threshold, determined by
4745 * zfs_arc_overflow_shift.
4747 static arc_ovf_level_t
4748 arc_is_overflowing(boolean_t use_reserve)
4750 /* Always allow at least one block of overflow */
4751 int64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4752 arc_c >> zfs_arc_overflow_shift);
4755 * We just compare the lower bound here for performance reasons. Our
4756 * primary goals are to make sure that the arc never grows without
4757 * bound, and that it can reach its maximum size. This check
4758 * accomplishes both goals. The maximum amount we could run over by is
4759 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4760 * in the ARC. In practice, that's in the tens of MB, which is low
4761 * enough to be safe.
4763 int64_t over = aggsum_lower_bound(&arc_sums.arcstat_size) -
4764 arc_c - overflow / 2;
4767 return (over < 0 ? ARC_OVF_NONE :
4768 over < overflow ? ARC_OVF_SOME : ARC_OVF_SEVERE);
4772 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, const void *tag,
4775 arc_buf_contents_t type = arc_buf_type(hdr);
4777 arc_get_data_impl(hdr, size, tag, alloc_flags);
4778 if (alloc_flags & ARC_HDR_ALLOC_LINEAR)
4779 return (abd_alloc_linear(size, type == ARC_BUFC_METADATA));
4781 return (abd_alloc(size, type == ARC_BUFC_METADATA));
4785 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, const void *tag)
4787 arc_buf_contents_t type = arc_buf_type(hdr);
4789 arc_get_data_impl(hdr, size, tag, 0);
4790 if (type == ARC_BUFC_METADATA) {
4791 return (zio_buf_alloc(size));
4793 ASSERT(type == ARC_BUFC_DATA);
4794 return (zio_data_buf_alloc(size));
4799 * Wait for the specified amount of data (in bytes) to be evicted from the
4800 * ARC, and for there to be sufficient free memory in the system. Waiting for
4801 * eviction ensures that the memory used by the ARC decreases. Waiting for
4802 * free memory ensures that the system won't run out of free pages, regardless
4803 * of ARC behavior and settings. See arc_lowmem_init().
4806 arc_wait_for_eviction(uint64_t amount, boolean_t use_reserve)
4808 switch (arc_is_overflowing(use_reserve)) {
4813 * This is a bit racy without taking arc_evict_lock, but the
4814 * worst that can happen is we either call zthr_wakeup() extra
4815 * time due to race with other thread here, or the set flag
4816 * get cleared by arc_evict_cb(), which is unlikely due to
4817 * big hysteresis, but also not important since at this level
4818 * of overflow the eviction is purely advisory. Same time
4819 * taking the global lock here every time without waiting for
4820 * the actual eviction creates a significant lock contention.
4822 if (!arc_evict_needed) {
4823 arc_evict_needed = B_TRUE;
4824 zthr_wakeup(arc_evict_zthr);
4827 case ARC_OVF_SEVERE:
4830 arc_evict_waiter_t aw;
4831 list_link_init(&aw.aew_node);
4832 cv_init(&aw.aew_cv, NULL, CV_DEFAULT, NULL);
4834 uint64_t last_count = 0;
4835 mutex_enter(&arc_evict_lock);
4836 if (!list_is_empty(&arc_evict_waiters)) {
4837 arc_evict_waiter_t *last =
4838 list_tail(&arc_evict_waiters);
4839 last_count = last->aew_count;
4840 } else if (!arc_evict_needed) {
4841 arc_evict_needed = B_TRUE;
4842 zthr_wakeup(arc_evict_zthr);
4845 * Note, the last waiter's count may be less than
4846 * arc_evict_count if we are low on memory in which
4847 * case arc_evict_state_impl() may have deferred
4848 * wakeups (but still incremented arc_evict_count).
4850 aw.aew_count = MAX(last_count, arc_evict_count) + amount;
4852 list_insert_tail(&arc_evict_waiters, &aw);
4854 arc_set_need_free();
4856 DTRACE_PROBE3(arc__wait__for__eviction,
4858 uint64_t, arc_evict_count,
4859 uint64_t, aw.aew_count);
4862 * We will be woken up either when arc_evict_count reaches
4863 * aew_count, or when the ARC is no longer overflowing and
4864 * eviction completes.
4865 * In case of "false" wakeup, we will still be on the list.
4868 cv_wait(&aw.aew_cv, &arc_evict_lock);
4869 } while (list_link_active(&aw.aew_node));
4870 mutex_exit(&arc_evict_lock);
4872 cv_destroy(&aw.aew_cv);
4878 * Allocate a block and return it to the caller. If we are hitting the
4879 * hard limit for the cache size, we must sleep, waiting for the eviction
4880 * thread to catch up. If we're past the target size but below the hard
4881 * limit, we'll only signal the reclaim thread and continue on.
4884 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, const void *tag,
4890 * If arc_size is currently overflowing, we must be adding data
4891 * faster than we are evicting. To ensure we don't compound the
4892 * problem by adding more data and forcing arc_size to grow even
4893 * further past it's target size, we wait for the eviction thread to
4894 * make some progress. We also wait for there to be sufficient free
4895 * memory in the system, as measured by arc_free_memory().
4897 * Specifically, we wait for zfs_arc_eviction_pct percent of the
4898 * requested size to be evicted. This should be more than 100%, to
4899 * ensure that that progress is also made towards getting arc_size
4900 * under arc_c. See the comment above zfs_arc_eviction_pct.
4902 arc_wait_for_eviction(size * zfs_arc_eviction_pct / 100,
4903 alloc_flags & ARC_HDR_USE_RESERVE);
4905 arc_buf_contents_t type = arc_buf_type(hdr);
4906 if (type == ARC_BUFC_METADATA) {
4907 arc_space_consume(size, ARC_SPACE_META);
4909 arc_space_consume(size, ARC_SPACE_DATA);
4913 * Update the state size. Note that ghost states have a
4914 * "ghost size" and so don't need to be updated.
4916 arc_state_t *state = hdr->b_l1hdr.b_state;
4917 if (!GHOST_STATE(state)) {
4919 (void) zfs_refcount_add_many(&state->arcs_size[type], size,
4923 * If this is reached via arc_read, the link is
4924 * protected by the hash lock. If reached via
4925 * arc_buf_alloc, the header should not be accessed by
4926 * any other thread. And, if reached via arc_read_done,
4927 * the hash lock will protect it if it's found in the
4928 * hash table; otherwise no other thread should be
4929 * trying to [add|remove]_reference it.
4931 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4932 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4933 (void) zfs_refcount_add_many(&state->arcs_esize[type],
4940 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size,
4943 arc_free_data_impl(hdr, size, tag);
4948 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, const void *tag)
4950 arc_buf_contents_t type = arc_buf_type(hdr);
4952 arc_free_data_impl(hdr, size, tag);
4953 if (type == ARC_BUFC_METADATA) {
4954 zio_buf_free(buf, size);
4956 ASSERT(type == ARC_BUFC_DATA);
4957 zio_data_buf_free(buf, size);
4962 * Free the arc data buffer.
4965 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, const void *tag)
4967 arc_state_t *state = hdr->b_l1hdr.b_state;
4968 arc_buf_contents_t type = arc_buf_type(hdr);
4970 /* protected by hash lock, if in the hash table */
4971 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4972 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4973 ASSERT(state != arc_anon && state != arc_l2c_only);
4975 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
4978 (void) zfs_refcount_remove_many(&state->arcs_size[type], size, tag);
4980 VERIFY3U(hdr->b_type, ==, type);
4981 if (type == ARC_BUFC_METADATA) {
4982 arc_space_return(size, ARC_SPACE_META);
4984 ASSERT(type == ARC_BUFC_DATA);
4985 arc_space_return(size, ARC_SPACE_DATA);
4990 * This routine is called whenever a buffer is accessed.
4993 arc_access(arc_buf_hdr_t *hdr, arc_flags_t arc_flags, boolean_t hit)
4995 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
4996 ASSERT(HDR_HAS_L1HDR(hdr));
4999 * Update buffer prefetch status.
5001 boolean_t was_prefetch = HDR_PREFETCH(hdr);
5002 boolean_t now_prefetch = arc_flags & ARC_FLAG_PREFETCH;
5003 if (was_prefetch != now_prefetch) {
5005 ARCSTAT_CONDSTAT(hit, demand_hit, demand_iohit,
5006 HDR_PRESCIENT_PREFETCH(hdr), prescient, predictive,
5009 if (HDR_HAS_L2HDR(hdr))
5010 l2arc_hdr_arcstats_decrement_state(hdr);
5012 arc_hdr_clear_flags(hdr,
5013 ARC_FLAG_PREFETCH | ARC_FLAG_PRESCIENT_PREFETCH);
5015 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5017 if (HDR_HAS_L2HDR(hdr))
5018 l2arc_hdr_arcstats_increment_state(hdr);
5021 if (arc_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5022 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5023 ARCSTAT_BUMP(arcstat_prescient_prefetch);
5025 ARCSTAT_BUMP(arcstat_predictive_prefetch);
5028 if (arc_flags & ARC_FLAG_L2CACHE)
5029 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5031 clock_t now = ddi_get_lbolt();
5032 if (hdr->b_l1hdr.b_state == arc_anon) {
5033 arc_state_t *new_state;
5035 * This buffer is not in the cache, and does not appear in
5036 * our "ghost" lists. Add it to the MRU or uncached state.
5038 ASSERT0(hdr->b_l1hdr.b_arc_access);
5039 hdr->b_l1hdr.b_arc_access = now;
5040 if (HDR_UNCACHED(hdr)) {
5041 new_state = arc_uncached;
5042 DTRACE_PROBE1(new_state__uncached, arc_buf_hdr_t *,
5045 new_state = arc_mru;
5046 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5048 arc_change_state(new_state, hdr);
5049 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5051 * This buffer has been accessed once recently and either
5052 * its read is still in progress or it is in the cache.
5054 if (HDR_IO_IN_PROGRESS(hdr)) {
5055 hdr->b_l1hdr.b_arc_access = now;
5058 hdr->b_l1hdr.b_mru_hits++;
5059 ARCSTAT_BUMP(arcstat_mru_hits);
5062 * If the previous access was a prefetch, then it already
5063 * handled possible promotion, so nothing more to do for now.
5066 hdr->b_l1hdr.b_arc_access = now;
5071 * If more than ARC_MINTIME have passed from the previous
5072 * hit, promote the buffer to the MFU state.
5074 if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access +
5076 hdr->b_l1hdr.b_arc_access = now;
5077 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5078 arc_change_state(arc_mfu, hdr);
5080 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5081 arc_state_t *new_state;
5083 * This buffer has been accessed once recently, but was
5084 * evicted from the cache. Would we have bigger MRU, it
5085 * would be an MRU hit, so handle it the same way, except
5086 * we don't need to check the previous access time.
5088 hdr->b_l1hdr.b_mru_ghost_hits++;
5089 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5090 hdr->b_l1hdr.b_arc_access = now;
5091 wmsum_add(&arc_mru_ghost->arcs_hits[arc_buf_type(hdr)],
5094 new_state = arc_mru;
5095 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5097 new_state = arc_mfu;
5098 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5100 arc_change_state(new_state, hdr);
5101 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5103 * This buffer has been accessed more than once and either
5104 * still in the cache or being restored from one of ghosts.
5106 if (!HDR_IO_IN_PROGRESS(hdr)) {
5107 hdr->b_l1hdr.b_mfu_hits++;
5108 ARCSTAT_BUMP(arcstat_mfu_hits);
5110 hdr->b_l1hdr.b_arc_access = now;
5111 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5113 * This buffer has been accessed more than once recently, but
5114 * has been evicted from the cache. Would we have bigger MFU
5115 * it would stay in cache, so move it back to MFU state.
5117 hdr->b_l1hdr.b_mfu_ghost_hits++;
5118 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5119 hdr->b_l1hdr.b_arc_access = now;
5120 wmsum_add(&arc_mfu_ghost->arcs_hits[arc_buf_type(hdr)],
5122 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5123 arc_change_state(arc_mfu, hdr);
5124 } else if (hdr->b_l1hdr.b_state == arc_uncached) {
5126 * This buffer is uncacheable, but we got a hit. Probably
5127 * a demand read after prefetch. Nothing more to do here.
5129 if (!HDR_IO_IN_PROGRESS(hdr))
5130 ARCSTAT_BUMP(arcstat_uncached_hits);
5131 hdr->b_l1hdr.b_arc_access = now;
5132 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5134 * This buffer is on the 2nd Level ARC and was not accessed
5135 * for a long time, so treat it as new and put into MRU.
5137 hdr->b_l1hdr.b_arc_access = now;
5138 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5139 arc_change_state(arc_mru, hdr);
5141 cmn_err(CE_PANIC, "invalid arc state 0x%p",
5142 hdr->b_l1hdr.b_state);
5147 * This routine is called by dbuf_hold() to update the arc_access() state
5148 * which otherwise would be skipped for entries in the dbuf cache.
5151 arc_buf_access(arc_buf_t *buf)
5153 arc_buf_hdr_t *hdr = buf->b_hdr;
5156 * Avoid taking the hash_lock when possible as an optimization.
5157 * The header must be checked again under the hash_lock in order
5158 * to handle the case where it is concurrently being released.
5160 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr))
5163 kmutex_t *hash_lock = HDR_LOCK(hdr);
5164 mutex_enter(hash_lock);
5166 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5167 mutex_exit(hash_lock);
5168 ARCSTAT_BUMP(arcstat_access_skip);
5172 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5173 hdr->b_l1hdr.b_state == arc_mfu ||
5174 hdr->b_l1hdr.b_state == arc_uncached);
5176 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5177 arc_access(hdr, 0, B_TRUE);
5178 mutex_exit(hash_lock);
5180 ARCSTAT_BUMP(arcstat_hits);
5181 ARCSTAT_CONDSTAT(B_TRUE /* demand */, demand, prefetch,
5182 !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5185 /* a generic arc_read_done_func_t which you can use */
5187 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5188 arc_buf_t *buf, void *arg)
5190 (void) zio, (void) zb, (void) bp;
5195 memcpy(arg, buf->b_data, arc_buf_size(buf));
5196 arc_buf_destroy(buf, arg);
5199 /* a generic arc_read_done_func_t */
5201 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5202 arc_buf_t *buf, void *arg)
5204 (void) zb, (void) bp;
5205 arc_buf_t **bufp = arg;
5208 ASSERT(zio == NULL || zio->io_error != 0);
5211 ASSERT(zio == NULL || zio->io_error == 0);
5213 ASSERT(buf->b_data != NULL);
5218 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5220 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5221 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5222 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5224 if (HDR_COMPRESSION_ENABLED(hdr)) {
5225 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5226 BP_GET_COMPRESS(bp));
5228 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5229 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5230 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5235 arc_read_done(zio_t *zio)
5237 blkptr_t *bp = zio->io_bp;
5238 arc_buf_hdr_t *hdr = zio->io_private;
5239 kmutex_t *hash_lock = NULL;
5240 arc_callback_t *callback_list;
5241 arc_callback_t *acb;
5244 * The hdr was inserted into hash-table and removed from lists
5245 * prior to starting I/O. We should find this header, since
5246 * it's in the hash table, and it should be legit since it's
5247 * not possible to evict it during the I/O. The only possible
5248 * reason for it not to be found is if we were freed during the
5251 if (HDR_IN_HASH_TABLE(hdr)) {
5252 arc_buf_hdr_t *found;
5254 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5255 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5256 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5257 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5258 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5260 found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock);
5262 ASSERT((found == hdr &&
5263 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5264 (found == hdr && HDR_L2_READING(hdr)));
5265 ASSERT3P(hash_lock, !=, NULL);
5268 if (BP_IS_PROTECTED(bp)) {
5269 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5270 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5271 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5272 hdr->b_crypt_hdr.b_iv);
5274 if (zio->io_error == 0) {
5275 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5278 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5279 sizeof (zil_chain_t));
5280 zio_crypt_decode_mac_zil(tmpbuf,
5281 hdr->b_crypt_hdr.b_mac);
5282 abd_return_buf(zio->io_abd, tmpbuf,
5283 sizeof (zil_chain_t));
5285 zio_crypt_decode_mac_bp(bp,
5286 hdr->b_crypt_hdr.b_mac);
5291 if (zio->io_error == 0) {
5292 /* byteswap if necessary */
5293 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5294 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5295 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5297 hdr->b_l1hdr.b_byteswap =
5298 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5301 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5303 if (!HDR_L2_READING(hdr)) {
5304 hdr->b_complevel = zio->io_prop.zp_complevel;
5308 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5309 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5310 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5312 callback_list = hdr->b_l1hdr.b_acb;
5313 ASSERT3P(callback_list, !=, NULL);
5314 hdr->b_l1hdr.b_acb = NULL;
5317 * If a read request has a callback (i.e. acb_done is not NULL), then we
5318 * make a buf containing the data according to the parameters which were
5319 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5320 * aren't needlessly decompressing the data multiple times.
5322 int callback_cnt = 0;
5323 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5325 /* We need the last one to call below in original order. */
5326 callback_list = acb;
5328 if (!acb->acb_done || acb->acb_nobuf)
5333 if (zio->io_error != 0)
5336 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5337 &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
5338 acb->acb_compressed, acb->acb_noauth, B_TRUE,
5342 * Assert non-speculative zios didn't fail because an
5343 * encryption key wasn't loaded
5345 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
5349 * If we failed to decrypt, report an error now (as the zio
5350 * layer would have done if it had done the transforms).
5352 if (error == ECKSUM) {
5353 ASSERT(BP_IS_PROTECTED(bp));
5354 error = SET_ERROR(EIO);
5355 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5356 spa_log_error(zio->io_spa, &acb->acb_zb,
5357 &zio->io_bp->blk_birth);
5358 (void) zfs_ereport_post(
5359 FM_EREPORT_ZFS_AUTHENTICATION,
5360 zio->io_spa, NULL, &acb->acb_zb, zio, 0);
5366 * Decompression or decryption failed. Set
5367 * io_error so that when we call acb_done
5368 * (below), we will indicate that the read
5369 * failed. Note that in the unusual case
5370 * where one callback is compressed and another
5371 * uncompressed, we will mark all of them
5372 * as failed, even though the uncompressed
5373 * one can't actually fail. In this case,
5374 * the hdr will not be anonymous, because
5375 * if there are multiple callbacks, it's
5376 * because multiple threads found the same
5377 * arc buf in the hash table.
5379 zio->io_error = error;
5384 * If there are multiple callbacks, we must have the hash lock,
5385 * because the only way for multiple threads to find this hdr is
5386 * in the hash table. This ensures that if there are multiple
5387 * callbacks, the hdr is not anonymous. If it were anonymous,
5388 * we couldn't use arc_buf_destroy() in the error case below.
5390 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5392 if (zio->io_error == 0) {
5393 arc_hdr_verify(hdr, zio->io_bp);
5395 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5396 if (hdr->b_l1hdr.b_state != arc_anon)
5397 arc_change_state(arc_anon, hdr);
5398 if (HDR_IN_HASH_TABLE(hdr))
5399 buf_hash_remove(hdr);
5402 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5403 (void) remove_reference(hdr, hdr);
5405 if (hash_lock != NULL)
5406 mutex_exit(hash_lock);
5408 /* execute each callback and free its structure */
5409 while ((acb = callback_list) != NULL) {
5410 if (acb->acb_done != NULL) {
5411 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5413 * If arc_buf_alloc_impl() fails during
5414 * decompression, the buf will still be
5415 * allocated, and needs to be freed here.
5417 arc_buf_destroy(acb->acb_buf,
5419 acb->acb_buf = NULL;
5421 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5422 acb->acb_buf, acb->acb_private);
5425 if (acb->acb_zio_dummy != NULL) {
5426 acb->acb_zio_dummy->io_error = zio->io_error;
5427 zio_nowait(acb->acb_zio_dummy);
5430 callback_list = acb->acb_prev;
5431 if (acb->acb_wait) {
5432 mutex_enter(&acb->acb_wait_lock);
5433 acb->acb_wait_error = zio->io_error;
5434 acb->acb_wait = B_FALSE;
5435 cv_signal(&acb->acb_wait_cv);
5436 mutex_exit(&acb->acb_wait_lock);
5437 /* acb will be freed by the waiting thread. */
5439 kmem_free(acb, sizeof (arc_callback_t));
5445 * "Read" the block at the specified DVA (in bp) via the
5446 * cache. If the block is found in the cache, invoke the provided
5447 * callback immediately and return. Note that the `zio' parameter
5448 * in the callback will be NULL in this case, since no IO was
5449 * required. If the block is not in the cache pass the read request
5450 * on to the spa with a substitute callback function, so that the
5451 * requested block will be added to the cache.
5453 * If a read request arrives for a block that has a read in-progress,
5454 * either wait for the in-progress read to complete (and return the
5455 * results); or, if this is a read with a "done" func, add a record
5456 * to the read to invoke the "done" func when the read completes,
5457 * and return; or just return.
5459 * arc_read_done() will invoke all the requested "done" functions
5460 * for readers of this block.
5463 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
5464 arc_read_done_func_t *done, void *private, zio_priority_t priority,
5465 int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5467 arc_buf_hdr_t *hdr = NULL;
5468 kmutex_t *hash_lock = NULL;
5470 uint64_t guid = spa_load_guid(spa);
5471 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
5472 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
5473 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5474 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
5475 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5476 boolean_t embedded_bp = !!BP_IS_EMBEDDED(bp);
5477 boolean_t no_buf = *arc_flags & ARC_FLAG_NO_BUF;
5478 arc_buf_t *buf = NULL;
5481 ASSERT(!embedded_bp ||
5482 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5483 ASSERT(!BP_IS_HOLE(bp));
5484 ASSERT(!BP_IS_REDACTED(bp));
5487 * Normally SPL_FSTRANS will already be set since kernel threads which
5488 * expect to call the DMU interfaces will set it when created. System
5489 * calls are similarly handled by setting/cleaning the bit in the
5490 * registered callback (module/os/.../zfs/zpl_*).
5492 * External consumers such as Lustre which call the exported DMU
5493 * interfaces may not have set SPL_FSTRANS. To avoid a deadlock
5494 * on the hash_lock always set and clear the bit.
5496 fstrans_cookie_t cookie = spl_fstrans_mark();
5499 * Verify the block pointer contents are reasonable. This should
5500 * always be the case since the blkptr is protected by a checksum.
5501 * However, if there is damage it's desirable to detect this early
5502 * and treat it as a checksum error. This allows an alternate blkptr
5503 * to be tried when one is available (e.g. ditto blocks).
5505 if (!zfs_blkptr_verify(spa, bp, (zio_flags & ZIO_FLAG_CONFIG_WRITER) ?
5506 BLK_CONFIG_HELD : BLK_CONFIG_NEEDED, BLK_VERIFY_LOG)) {
5507 rc = SET_ERROR(ECKSUM);
5513 * Embedded BP's have no DVA and require no I/O to "read".
5514 * Create an anonymous arc buf to back it.
5516 hdr = buf_hash_find(guid, bp, &hash_lock);
5520 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5521 * we maintain encrypted data separately from compressed / uncompressed
5522 * data. If the user is requesting raw encrypted data and we don't have
5523 * that in the header we will read from disk to guarantee that we can
5524 * get it even if the encryption keys aren't loaded.
5526 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
5527 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
5528 boolean_t is_data = !HDR_ISTYPE_METADATA(hdr);
5530 if (HDR_IO_IN_PROGRESS(hdr)) {
5531 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
5532 mutex_exit(hash_lock);
5533 ARCSTAT_BUMP(arcstat_cached_only_in_progress);
5534 rc = SET_ERROR(ENOENT);
5538 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5539 ASSERT3P(head_zio, !=, NULL);
5540 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5541 priority == ZIO_PRIORITY_SYNC_READ) {
5543 * This is a sync read that needs to wait for
5544 * an in-flight async read. Request that the
5545 * zio have its priority upgraded.
5547 zio_change_priority(head_zio, priority);
5548 DTRACE_PROBE1(arc__async__upgrade__sync,
5549 arc_buf_hdr_t *, hdr);
5550 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5553 DTRACE_PROBE1(arc__iohit, arc_buf_hdr_t *, hdr);
5554 arc_access(hdr, *arc_flags, B_FALSE);
5557 * If there are multiple threads reading the same block
5558 * and that block is not yet in the ARC, then only one
5559 * thread will do the physical I/O and all other
5560 * threads will wait until that I/O completes.
5561 * Synchronous reads use the acb_wait_cv whereas nowait
5562 * reads register a callback. Both are signalled/called
5565 * Errors of the physical I/O may need to be propagated.
5566 * Synchronous read errors are returned here from
5567 * arc_read_done via acb_wait_error. Nowait reads
5568 * attach the acb_zio_dummy zio to pio and
5569 * arc_read_done propagates the physical I/O's io_error
5570 * to acb_zio_dummy, and thereby to pio.
5572 arc_callback_t *acb = NULL;
5573 if (done || pio || *arc_flags & ARC_FLAG_WAIT) {
5574 acb = kmem_zalloc(sizeof (arc_callback_t),
5576 acb->acb_done = done;
5577 acb->acb_private = private;
5578 acb->acb_compressed = compressed_read;
5579 acb->acb_encrypted = encrypted_read;
5580 acb->acb_noauth = noauth_read;
5581 acb->acb_nobuf = no_buf;
5582 if (*arc_flags & ARC_FLAG_WAIT) {
5583 acb->acb_wait = B_TRUE;
5584 mutex_init(&acb->acb_wait_lock, NULL,
5585 MUTEX_DEFAULT, NULL);
5586 cv_init(&acb->acb_wait_cv, NULL,
5591 acb->acb_zio_dummy = zio_null(pio,
5592 spa, NULL, NULL, NULL, zio_flags);
5594 acb->acb_zio_head = head_zio;
5595 acb->acb_next = hdr->b_l1hdr.b_acb;
5596 hdr->b_l1hdr.b_acb->acb_prev = acb;
5597 hdr->b_l1hdr.b_acb = acb;
5599 mutex_exit(hash_lock);
5601 ARCSTAT_BUMP(arcstat_iohits);
5602 ARCSTAT_CONDSTAT(!(*arc_flags & ARC_FLAG_PREFETCH),
5603 demand, prefetch, is_data, data, metadata, iohits);
5605 if (*arc_flags & ARC_FLAG_WAIT) {
5606 mutex_enter(&acb->acb_wait_lock);
5607 while (acb->acb_wait) {
5608 cv_wait(&acb->acb_wait_cv,
5609 &acb->acb_wait_lock);
5611 rc = acb->acb_wait_error;
5612 mutex_exit(&acb->acb_wait_lock);
5613 mutex_destroy(&acb->acb_wait_lock);
5614 cv_destroy(&acb->acb_wait_cv);
5615 kmem_free(acb, sizeof (arc_callback_t));
5620 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5621 hdr->b_l1hdr.b_state == arc_mfu ||
5622 hdr->b_l1hdr.b_state == arc_uncached);
5624 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5625 arc_access(hdr, *arc_flags, B_TRUE);
5627 if (done && !no_buf) {
5628 ASSERT(!embedded_bp || !BP_IS_HOLE(bp));
5630 /* Get a buf with the desired data in it. */
5631 rc = arc_buf_alloc_impl(hdr, spa, zb, private,
5632 encrypted_read, compressed_read, noauth_read,
5636 * Convert authentication and decryption errors
5637 * to EIO (and generate an ereport if needed)
5638 * before leaving the ARC.
5640 rc = SET_ERROR(EIO);
5641 if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5642 spa_log_error(spa, zb, &hdr->b_birth);
5643 (void) zfs_ereport_post(
5644 FM_EREPORT_ZFS_AUTHENTICATION,
5645 spa, NULL, zb, NULL, 0);
5649 arc_buf_destroy_impl(buf);
5651 (void) remove_reference(hdr, private);
5654 /* assert any errors weren't due to unloaded keys */
5655 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5658 mutex_exit(hash_lock);
5659 ARCSTAT_BUMP(arcstat_hits);
5660 ARCSTAT_CONDSTAT(!(*arc_flags & ARC_FLAG_PREFETCH),
5661 demand, prefetch, is_data, data, metadata, hits);
5662 *arc_flags |= ARC_FLAG_CACHED;
5665 uint64_t lsize = BP_GET_LSIZE(bp);
5666 uint64_t psize = BP_GET_PSIZE(bp);
5667 arc_callback_t *acb;
5670 boolean_t devw = B_FALSE;
5673 int alloc_flags = encrypted_read ? ARC_HDR_ALLOC_RDATA : 0;
5674 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5676 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
5677 if (hash_lock != NULL)
5678 mutex_exit(hash_lock);
5679 rc = SET_ERROR(ENOENT);
5685 * This block is not in the cache or it has
5688 arc_buf_hdr_t *exists = NULL;
5689 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5690 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), 0, type);
5693 hdr->b_dva = *BP_IDENTITY(bp);
5694 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5695 exists = buf_hash_insert(hdr, &hash_lock);
5697 if (exists != NULL) {
5698 /* somebody beat us to the hash insert */
5699 mutex_exit(hash_lock);
5700 buf_discard_identity(hdr);
5701 arc_hdr_destroy(hdr);
5702 goto top; /* restart the IO request */
5706 * This block is in the ghost cache or encrypted data
5707 * was requested and we didn't have it. If it was
5708 * L2-only (and thus didn't have an L1 hdr),
5709 * we realloc the header to add an L1 hdr.
5711 if (!HDR_HAS_L1HDR(hdr)) {
5712 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5716 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
5717 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5718 ASSERT(!HDR_HAS_RABD(hdr));
5719 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5720 ASSERT0(zfs_refcount_count(
5721 &hdr->b_l1hdr.b_refcnt));
5722 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5724 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5726 } else if (HDR_IO_IN_PROGRESS(hdr)) {
5728 * If this header already had an IO in progress
5729 * and we are performing another IO to fetch
5730 * encrypted data we must wait until the first
5731 * IO completes so as not to confuse
5732 * arc_read_done(). This should be very rare
5733 * and so the performance impact shouldn't
5736 arc_callback_t *acb = kmem_zalloc(
5737 sizeof (arc_callback_t), KM_SLEEP);
5738 acb->acb_wait = B_TRUE;
5739 mutex_init(&acb->acb_wait_lock, NULL,
5740 MUTEX_DEFAULT, NULL);
5741 cv_init(&acb->acb_wait_cv, NULL, CV_DEFAULT,
5744 hdr->b_l1hdr.b_acb->acb_zio_head;
5745 acb->acb_next = hdr->b_l1hdr.b_acb;
5746 hdr->b_l1hdr.b_acb->acb_prev = acb;
5747 hdr->b_l1hdr.b_acb = acb;
5748 mutex_exit(hash_lock);
5749 mutex_enter(&acb->acb_wait_lock);
5750 while (acb->acb_wait) {
5751 cv_wait(&acb->acb_wait_cv,
5752 &acb->acb_wait_lock);
5754 mutex_exit(&acb->acb_wait_lock);
5755 mutex_destroy(&acb->acb_wait_lock);
5756 cv_destroy(&acb->acb_wait_cv);
5757 kmem_free(acb, sizeof (arc_callback_t));
5761 if (*arc_flags & ARC_FLAG_UNCACHED) {
5762 arc_hdr_set_flags(hdr, ARC_FLAG_UNCACHED);
5763 if (!encrypted_read)
5764 alloc_flags |= ARC_HDR_ALLOC_LINEAR;
5768 * Take additional reference for IO_IN_PROGRESS. It stops
5769 * arc_access() from putting this header without any buffers
5770 * and so other references but obviously nonevictable onto
5771 * the evictable list of MRU or MFU state.
5773 add_reference(hdr, hdr);
5775 arc_access(hdr, *arc_flags, B_FALSE);
5776 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5777 arc_hdr_alloc_abd(hdr, alloc_flags);
5778 if (encrypted_read) {
5779 ASSERT(HDR_HAS_RABD(hdr));
5780 size = HDR_GET_PSIZE(hdr);
5781 hdr_abd = hdr->b_crypt_hdr.b_rabd;
5782 zio_flags |= ZIO_FLAG_RAW;
5784 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5785 size = arc_hdr_size(hdr);
5786 hdr_abd = hdr->b_l1hdr.b_pabd;
5788 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
5789 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
5793 * For authenticated bp's, we do not ask the ZIO layer
5794 * to authenticate them since this will cause the entire
5795 * IO to fail if the key isn't loaded. Instead, we
5796 * defer authentication until arc_buf_fill(), which will
5797 * verify the data when the key is available.
5799 if (BP_IS_AUTHENTICATED(bp))
5800 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
5803 if (BP_IS_AUTHENTICATED(bp))
5804 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
5805 if (BP_GET_LEVEL(bp) > 0)
5806 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5807 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5809 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5810 acb->acb_done = done;
5811 acb->acb_private = private;
5812 acb->acb_compressed = compressed_read;
5813 acb->acb_encrypted = encrypted_read;
5814 acb->acb_noauth = noauth_read;
5817 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5818 hdr->b_l1hdr.b_acb = acb;
5820 if (HDR_HAS_L2HDR(hdr) &&
5821 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5822 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5823 addr = hdr->b_l2hdr.b_daddr;
5825 * Lock out L2ARC device removal.
5827 if (vdev_is_dead(vd) ||
5828 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5833 * We count both async reads and scrub IOs as asynchronous so
5834 * that both can be upgraded in the event of a cache hit while
5835 * the read IO is still in-flight.
5837 if (priority == ZIO_PRIORITY_ASYNC_READ ||
5838 priority == ZIO_PRIORITY_SCRUB)
5839 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5841 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5844 * At this point, we have a level 1 cache miss or a blkptr
5845 * with embedded data. Try again in L2ARC if possible.
5847 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5850 * Skip ARC stat bump for block pointers with embedded
5851 * data. The data are read from the blkptr itself via
5852 * decode_embedded_bp_compressed().
5855 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr,
5856 blkptr_t *, bp, uint64_t, lsize,
5857 zbookmark_phys_t *, zb);
5858 ARCSTAT_BUMP(arcstat_misses);
5859 ARCSTAT_CONDSTAT(!(*arc_flags & ARC_FLAG_PREFETCH),
5860 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data,
5862 zfs_racct_read(size, 1);
5865 /* Check if the spa even has l2 configured */
5866 const boolean_t spa_has_l2 = l2arc_ndev != 0 &&
5867 spa->spa_l2cache.sav_count > 0;
5869 if (vd != NULL && spa_has_l2 && !(l2arc_norw && devw)) {
5871 * Read from the L2ARC if the following are true:
5872 * 1. The L2ARC vdev was previously cached.
5873 * 2. This buffer still has L2ARC metadata.
5874 * 3. This buffer isn't currently writing to the L2ARC.
5875 * 4. The L2ARC entry wasn't evicted, which may
5876 * also have invalidated the vdev.
5878 if (HDR_HAS_L2HDR(hdr) &&
5879 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr)) {
5880 l2arc_read_callback_t *cb;
5884 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5885 ARCSTAT_BUMP(arcstat_l2_hits);
5886 hdr->b_l2hdr.b_hits++;
5888 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5890 cb->l2rcb_hdr = hdr;
5893 cb->l2rcb_flags = zio_flags;
5896 * When Compressed ARC is disabled, but the
5897 * L2ARC block is compressed, arc_hdr_size()
5898 * will have returned LSIZE rather than PSIZE.
5900 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
5901 !HDR_COMPRESSION_ENABLED(hdr) &&
5902 HDR_GET_PSIZE(hdr) != 0) {
5903 size = HDR_GET_PSIZE(hdr);
5906 asize = vdev_psize_to_asize(vd, size);
5907 if (asize != size) {
5908 abd = abd_alloc_for_io(asize,
5909 HDR_ISTYPE_METADATA(hdr));
5910 cb->l2rcb_abd = abd;
5915 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5916 addr + asize <= vd->vdev_psize -
5917 VDEV_LABEL_END_SIZE);
5920 * l2arc read. The SCL_L2ARC lock will be
5921 * released by l2arc_read_done().
5922 * Issue a null zio if the underlying buffer
5923 * was squashed to zero size by compression.
5925 ASSERT3U(arc_hdr_get_compress(hdr), !=,
5926 ZIO_COMPRESS_EMPTY);
5927 rzio = zio_read_phys(pio, vd, addr,
5930 l2arc_read_done, cb, priority,
5931 zio_flags | ZIO_FLAG_CANFAIL |
5932 ZIO_FLAG_DONT_PROPAGATE |
5933 ZIO_FLAG_DONT_RETRY, B_FALSE);
5934 acb->acb_zio_head = rzio;
5936 if (hash_lock != NULL)
5937 mutex_exit(hash_lock);
5939 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5941 ARCSTAT_INCR(arcstat_l2_read_bytes,
5942 HDR_GET_PSIZE(hdr));
5944 if (*arc_flags & ARC_FLAG_NOWAIT) {
5949 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5950 if (zio_wait(rzio) == 0)
5953 /* l2arc read error; goto zio_read() */
5954 if (hash_lock != NULL)
5955 mutex_enter(hash_lock);
5957 DTRACE_PROBE1(l2arc__miss,
5958 arc_buf_hdr_t *, hdr);
5959 ARCSTAT_BUMP(arcstat_l2_misses);
5960 if (HDR_L2_WRITING(hdr))
5961 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5962 spa_config_exit(spa, SCL_L2ARC, vd);
5966 spa_config_exit(spa, SCL_L2ARC, vd);
5969 * Only a spa with l2 should contribute to l2
5970 * miss stats. (Including the case of having a
5971 * faulted cache device - that's also a miss.)
5975 * Skip ARC stat bump for block pointers with
5976 * embedded data. The data are read from the
5978 * decode_embedded_bp_compressed().
5981 DTRACE_PROBE1(l2arc__miss,
5982 arc_buf_hdr_t *, hdr);
5983 ARCSTAT_BUMP(arcstat_l2_misses);
5988 rzio = zio_read(pio, spa, bp, hdr_abd, size,
5989 arc_read_done, hdr, priority, zio_flags, zb);
5990 acb->acb_zio_head = rzio;
5992 if (hash_lock != NULL)
5993 mutex_exit(hash_lock);
5995 if (*arc_flags & ARC_FLAG_WAIT) {
5996 rc = zio_wait(rzio);
6000 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6005 /* embedded bps don't actually go to disk */
6007 spa_read_history_add(spa, zb, *arc_flags);
6008 spl_fstrans_unmark(cookie);
6013 done(NULL, zb, bp, buf, private);
6014 if (pio && rc != 0) {
6015 zio_t *zio = zio_null(pio, spa, NULL, NULL, NULL, zio_flags);
6023 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6027 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6029 p->p_private = private;
6030 list_link_init(&p->p_node);
6031 zfs_refcount_create(&p->p_refcnt);
6033 mutex_enter(&arc_prune_mtx);
6034 zfs_refcount_add(&p->p_refcnt, &arc_prune_list);
6035 list_insert_head(&arc_prune_list, p);
6036 mutex_exit(&arc_prune_mtx);
6042 arc_remove_prune_callback(arc_prune_t *p)
6044 boolean_t wait = B_FALSE;
6045 mutex_enter(&arc_prune_mtx);
6046 list_remove(&arc_prune_list, p);
6047 if (zfs_refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6049 mutex_exit(&arc_prune_mtx);
6051 /* wait for arc_prune_task to finish */
6053 taskq_wait_outstanding(arc_prune_taskq, 0);
6054 ASSERT0(zfs_refcount_count(&p->p_refcnt));
6055 zfs_refcount_destroy(&p->p_refcnt);
6056 kmem_free(p, sizeof (*p));
6060 * Helper function for arc_prune_async() it is responsible for safely
6061 * handling the execution of a registered arc_prune_func_t.
6064 arc_prune_task(void *ptr)
6066 arc_prune_t *ap = (arc_prune_t *)ptr;
6067 arc_prune_func_t *func = ap->p_pfunc;
6070 func(ap->p_adjust, ap->p_private);
6072 zfs_refcount_remove(&ap->p_refcnt, func);
6076 * Notify registered consumers they must drop holds on a portion of the ARC
6077 * buffers they reference. This provides a mechanism to ensure the ARC can
6078 * honor the metadata limit and reclaim otherwise pinned ARC buffers.
6080 * This operation is performed asynchronously so it may be safely called
6081 * in the context of the arc_reclaim_thread(). A reference is taken here
6082 * for each registered arc_prune_t and the arc_prune_task() is responsible
6083 * for releasing it once the registered arc_prune_func_t has completed.
6086 arc_prune_async(uint64_t adjust)
6090 mutex_enter(&arc_prune_mtx);
6091 for (ap = list_head(&arc_prune_list); ap != NULL;
6092 ap = list_next(&arc_prune_list, ap)) {
6094 if (zfs_refcount_count(&ap->p_refcnt) >= 2)
6097 zfs_refcount_add(&ap->p_refcnt, ap->p_pfunc);
6098 ap->p_adjust = adjust;
6099 if (taskq_dispatch(arc_prune_taskq, arc_prune_task,
6100 ap, TQ_SLEEP) == TASKQID_INVALID) {
6101 zfs_refcount_remove(&ap->p_refcnt, ap->p_pfunc);
6104 ARCSTAT_BUMP(arcstat_prune);
6106 mutex_exit(&arc_prune_mtx);
6110 * Notify the arc that a block was freed, and thus will never be used again.
6113 arc_freed(spa_t *spa, const blkptr_t *bp)
6116 kmutex_t *hash_lock;
6117 uint64_t guid = spa_load_guid(spa);
6119 ASSERT(!BP_IS_EMBEDDED(bp));
6121 hdr = buf_hash_find(guid, bp, &hash_lock);
6126 * We might be trying to free a block that is still doing I/O
6127 * (i.e. prefetch) or has some other reference (i.e. a dedup-ed,
6128 * dmu_sync-ed block). A block may also have a reference if it is
6129 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6130 * have written the new block to its final resting place on disk but
6131 * without the dedup flag set. This would have left the hdr in the MRU
6132 * state and discoverable. When the txg finally syncs it detects that
6133 * the block was overridden in open context and issues an override I/O.
6134 * Since this is a dedup block, the override I/O will determine if the
6135 * block is already in the DDT. If so, then it will replace the io_bp
6136 * with the bp from the DDT and allow the I/O to finish. When the I/O
6137 * reaches the done callback, dbuf_write_override_done, it will
6138 * check to see if the io_bp and io_bp_override are identical.
6139 * If they are not, then it indicates that the bp was replaced with
6140 * the bp in the DDT and the override bp is freed. This allows
6141 * us to arrive here with a reference on a block that is being
6142 * freed. So if we have an I/O in progress, or a reference to
6143 * this hdr, then we don't destroy the hdr.
6145 if (!HDR_HAS_L1HDR(hdr) ||
6146 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6147 arc_change_state(arc_anon, hdr);
6148 arc_hdr_destroy(hdr);
6149 mutex_exit(hash_lock);
6151 mutex_exit(hash_lock);
6157 * Release this buffer from the cache, making it an anonymous buffer. This
6158 * must be done after a read and prior to modifying the buffer contents.
6159 * If the buffer has more than one reference, we must make
6160 * a new hdr for the buffer.
6163 arc_release(arc_buf_t *buf, const void *tag)
6165 arc_buf_hdr_t *hdr = buf->b_hdr;
6168 * It would be nice to assert that if its DMU metadata (level >
6169 * 0 || it's the dnode file), then it must be syncing context.
6170 * But we don't know that information at this level.
6173 ASSERT(HDR_HAS_L1HDR(hdr));
6176 * We don't grab the hash lock prior to this check, because if
6177 * the buffer's header is in the arc_anon state, it won't be
6178 * linked into the hash table.
6180 if (hdr->b_l1hdr.b_state == arc_anon) {
6181 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6182 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6183 ASSERT(!HDR_HAS_L2HDR(hdr));
6185 ASSERT3P(hdr->b_l1hdr.b_buf, ==, buf);
6186 ASSERT(ARC_BUF_LAST(buf));
6187 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6188 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6190 hdr->b_l1hdr.b_arc_access = 0;
6193 * If the buf is being overridden then it may already
6194 * have a hdr that is not empty.
6196 buf_discard_identity(hdr);
6202 kmutex_t *hash_lock = HDR_LOCK(hdr);
6203 mutex_enter(hash_lock);
6206 * This assignment is only valid as long as the hash_lock is
6207 * held, we must be careful not to reference state or the
6208 * b_state field after dropping the lock.
6210 arc_state_t *state = hdr->b_l1hdr.b_state;
6211 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6212 ASSERT3P(state, !=, arc_anon);
6214 /* this buffer is not on any list */
6215 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6217 if (HDR_HAS_L2HDR(hdr)) {
6218 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6221 * We have to recheck this conditional again now that
6222 * we're holding the l2ad_mtx to prevent a race with
6223 * another thread which might be concurrently calling
6224 * l2arc_evict(). In that case, l2arc_evict() might have
6225 * destroyed the header's L2 portion as we were waiting
6226 * to acquire the l2ad_mtx.
6228 if (HDR_HAS_L2HDR(hdr))
6229 arc_hdr_l2hdr_destroy(hdr);
6231 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6235 * Do we have more than one buf?
6237 if (hdr->b_l1hdr.b_buf != buf || !ARC_BUF_LAST(buf)) {
6238 arc_buf_hdr_t *nhdr;
6239 uint64_t spa = hdr->b_spa;
6240 uint64_t psize = HDR_GET_PSIZE(hdr);
6241 uint64_t lsize = HDR_GET_LSIZE(hdr);
6242 boolean_t protected = HDR_PROTECTED(hdr);
6243 enum zio_compress compress = arc_hdr_get_compress(hdr);
6244 arc_buf_contents_t type = arc_buf_type(hdr);
6245 VERIFY3U(hdr->b_type, ==, type);
6247 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6248 VERIFY3S(remove_reference(hdr, tag), >, 0);
6250 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
6251 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6252 ASSERT(ARC_BUF_LAST(buf));
6256 * Pull the data off of this hdr and attach it to
6257 * a new anonymous hdr. Also find the last buffer
6258 * in the hdr's buffer list.
6260 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6261 ASSERT3P(lastbuf, !=, NULL);
6264 * If the current arc_buf_t and the hdr are sharing their data
6265 * buffer, then we must stop sharing that block.
6267 if (ARC_BUF_SHARED(buf)) {
6268 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6269 ASSERT(!arc_buf_is_shared(lastbuf));
6272 * First, sever the block sharing relationship between
6273 * buf and the arc_buf_hdr_t.
6275 arc_unshare_buf(hdr, buf);
6278 * Now we need to recreate the hdr's b_pabd. Since we
6279 * have lastbuf handy, we try to share with it, but if
6280 * we can't then we allocate a new b_pabd and copy the
6281 * data from buf into it.
6283 if (arc_can_share(hdr, lastbuf)) {
6284 arc_share_buf(hdr, lastbuf);
6286 arc_hdr_alloc_abd(hdr, 0);
6287 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6288 buf->b_data, psize);
6290 VERIFY3P(lastbuf->b_data, !=, NULL);
6291 } else if (HDR_SHARED_DATA(hdr)) {
6293 * Uncompressed shared buffers are always at the end
6294 * of the list. Compressed buffers don't have the
6295 * same requirements. This makes it hard to
6296 * simply assert that the lastbuf is shared so
6297 * we rely on the hdr's compression flags to determine
6298 * if we have a compressed, shared buffer.
6300 ASSERT(arc_buf_is_shared(lastbuf) ||
6301 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6302 ASSERT(!arc_buf_is_shared(buf));
6305 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6306 ASSERT3P(state, !=, arc_l2c_only);
6308 (void) zfs_refcount_remove_many(&state->arcs_size[type],
6309 arc_buf_size(buf), buf);
6311 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6312 ASSERT3P(state, !=, arc_l2c_only);
6313 (void) zfs_refcount_remove_many(
6314 &state->arcs_esize[type],
6315 arc_buf_size(buf), buf);
6318 arc_cksum_verify(buf);
6319 arc_buf_unwatch(buf);
6321 /* if this is the last uncompressed buf free the checksum */
6322 if (!arc_hdr_has_uncompressed_buf(hdr))
6323 arc_cksum_free(hdr);
6325 mutex_exit(hash_lock);
6327 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6328 compress, hdr->b_complevel, type);
6329 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6330 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6331 VERIFY3U(nhdr->b_type, ==, type);
6332 ASSERT(!HDR_SHARED_DATA(nhdr));
6334 nhdr->b_l1hdr.b_buf = buf;
6335 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6338 (void) zfs_refcount_add_many(&arc_anon->arcs_size[type],
6339 arc_buf_size(buf), buf);
6341 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6342 /* protected by hash lock, or hdr is on arc_anon */
6343 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6344 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6345 hdr->b_l1hdr.b_mru_hits = 0;
6346 hdr->b_l1hdr.b_mru_ghost_hits = 0;
6347 hdr->b_l1hdr.b_mfu_hits = 0;
6348 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
6349 arc_change_state(arc_anon, hdr);
6350 hdr->b_l1hdr.b_arc_access = 0;
6352 mutex_exit(hash_lock);
6353 buf_discard_identity(hdr);
6359 arc_released(arc_buf_t *buf)
6361 return (buf->b_data != NULL &&
6362 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6367 arc_referenced(arc_buf_t *buf)
6369 return (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6374 arc_write_ready(zio_t *zio)
6376 arc_write_callback_t *callback = zio->io_private;
6377 arc_buf_t *buf = callback->awcb_buf;
6378 arc_buf_hdr_t *hdr = buf->b_hdr;
6379 blkptr_t *bp = zio->io_bp;
6380 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6381 fstrans_cookie_t cookie = spl_fstrans_mark();
6383 ASSERT(HDR_HAS_L1HDR(hdr));
6384 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6385 ASSERT3P(hdr->b_l1hdr.b_buf, !=, NULL);
6388 * If we're reexecuting this zio because the pool suspended, then
6389 * cleanup any state that was previously set the first time the
6390 * callback was invoked.
6392 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6393 arc_cksum_free(hdr);
6394 arc_buf_unwatch(buf);
6395 if (hdr->b_l1hdr.b_pabd != NULL) {
6396 if (ARC_BUF_SHARED(buf)) {
6397 arc_unshare_buf(hdr, buf);
6399 ASSERT(!arc_buf_is_shared(buf));
6400 arc_hdr_free_abd(hdr, B_FALSE);
6404 if (HDR_HAS_RABD(hdr))
6405 arc_hdr_free_abd(hdr, B_TRUE);
6407 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6408 ASSERT(!HDR_HAS_RABD(hdr));
6409 ASSERT(!HDR_SHARED_DATA(hdr));
6410 ASSERT(!arc_buf_is_shared(buf));
6412 callback->awcb_ready(zio, buf, callback->awcb_private);
6414 if (HDR_IO_IN_PROGRESS(hdr)) {
6415 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6417 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6418 add_reference(hdr, hdr); /* For IO_IN_PROGRESS. */
6421 if (BP_IS_PROTECTED(bp)) {
6422 /* ZIL blocks are written through zio_rewrite */
6423 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6425 if (BP_SHOULD_BYTESWAP(bp)) {
6426 if (BP_GET_LEVEL(bp) > 0) {
6427 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6429 hdr->b_l1hdr.b_byteswap =
6430 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6433 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6436 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
6437 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6438 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6439 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6440 hdr->b_crypt_hdr.b_iv);
6441 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6443 arc_hdr_clear_flags(hdr, ARC_FLAG_PROTECTED);
6447 * If this block was written for raw encryption but the zio layer
6448 * ended up only authenticating it, adjust the buffer flags now.
6450 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6451 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6452 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6453 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6454 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6455 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6456 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6457 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6460 /* this must be done after the buffer flags are adjusted */
6461 arc_cksum_compute(buf);
6463 enum zio_compress compress;
6464 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6465 compress = ZIO_COMPRESS_OFF;
6467 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6468 compress = BP_GET_COMPRESS(bp);
6470 HDR_SET_PSIZE(hdr, psize);
6471 arc_hdr_set_compress(hdr, compress);
6472 hdr->b_complevel = zio->io_prop.zp_complevel;
6474 if (zio->io_error != 0 || psize == 0)
6478 * Fill the hdr with data. If the buffer is encrypted we have no choice
6479 * but to copy the data into b_radb. If the hdr is compressed, the data
6480 * we want is available from the zio, otherwise we can take it from
6483 * We might be able to share the buf's data with the hdr here. However,
6484 * doing so would cause the ARC to be full of linear ABDs if we write a
6485 * lot of shareable data. As a compromise, we check whether scattered
6486 * ABDs are allowed, and assume that if they are then the user wants
6487 * the ARC to be primarily filled with them regardless of the data being
6488 * written. Therefore, if they're allowed then we allocate one and copy
6489 * the data into it; otherwise, we share the data directly if we can.
6491 if (ARC_BUF_ENCRYPTED(buf)) {
6492 ASSERT3U(psize, >, 0);
6493 ASSERT(ARC_BUF_COMPRESSED(buf));
6494 arc_hdr_alloc_abd(hdr, ARC_HDR_ALLOC_RDATA |
6495 ARC_HDR_USE_RESERVE);
6496 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6497 } else if (!(HDR_UNCACHED(hdr) ||
6498 abd_size_alloc_linear(arc_buf_size(buf))) ||
6499 !arc_can_share(hdr, buf)) {
6501 * Ideally, we would always copy the io_abd into b_pabd, but the
6502 * user may have disabled compressed ARC, thus we must check the
6503 * hdr's compression setting rather than the io_bp's.
6505 if (BP_IS_ENCRYPTED(bp)) {
6506 ASSERT3U(psize, >, 0);
6507 arc_hdr_alloc_abd(hdr, ARC_HDR_ALLOC_RDATA |
6508 ARC_HDR_USE_RESERVE);
6509 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6510 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
6511 !ARC_BUF_COMPRESSED(buf)) {
6512 ASSERT3U(psize, >, 0);
6513 arc_hdr_alloc_abd(hdr, ARC_HDR_USE_RESERVE);
6514 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6516 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6517 arc_hdr_alloc_abd(hdr, ARC_HDR_USE_RESERVE);
6518 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6522 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6523 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6524 ASSERT3P(hdr->b_l1hdr.b_buf, ==, buf);
6525 ASSERT(ARC_BUF_LAST(buf));
6527 arc_share_buf(hdr, buf);
6531 arc_hdr_verify(hdr, bp);
6532 spl_fstrans_unmark(cookie);
6536 arc_write_children_ready(zio_t *zio)
6538 arc_write_callback_t *callback = zio->io_private;
6539 arc_buf_t *buf = callback->awcb_buf;
6541 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6545 arc_write_done(zio_t *zio)
6547 arc_write_callback_t *callback = zio->io_private;
6548 arc_buf_t *buf = callback->awcb_buf;
6549 arc_buf_hdr_t *hdr = buf->b_hdr;
6551 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6553 if (zio->io_error == 0) {
6554 arc_hdr_verify(hdr, zio->io_bp);
6556 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6557 buf_discard_identity(hdr);
6559 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6560 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6563 ASSERT(HDR_EMPTY(hdr));
6567 * If the block to be written was all-zero or compressed enough to be
6568 * embedded in the BP, no write was performed so there will be no
6569 * dva/birth/checksum. The buffer must therefore remain anonymous
6572 if (!HDR_EMPTY(hdr)) {
6573 arc_buf_hdr_t *exists;
6574 kmutex_t *hash_lock;
6576 ASSERT3U(zio->io_error, ==, 0);
6578 arc_cksum_verify(buf);
6580 exists = buf_hash_insert(hdr, &hash_lock);
6581 if (exists != NULL) {
6583 * This can only happen if we overwrite for
6584 * sync-to-convergence, because we remove
6585 * buffers from the hash table when we arc_free().
6587 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6588 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6589 panic("bad overwrite, hdr=%p exists=%p",
6590 (void *)hdr, (void *)exists);
6591 ASSERT(zfs_refcount_is_zero(
6592 &exists->b_l1hdr.b_refcnt));
6593 arc_change_state(arc_anon, exists);
6594 arc_hdr_destroy(exists);
6595 mutex_exit(hash_lock);
6596 exists = buf_hash_insert(hdr, &hash_lock);
6597 ASSERT3P(exists, ==, NULL);
6598 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6600 ASSERT(zio->io_prop.zp_nopwrite);
6601 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6602 panic("bad nopwrite, hdr=%p exists=%p",
6603 (void *)hdr, (void *)exists);
6606 ASSERT3P(hdr->b_l1hdr.b_buf, !=, NULL);
6607 ASSERT(ARC_BUF_LAST(hdr->b_l1hdr.b_buf));
6608 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6609 ASSERT(BP_GET_DEDUP(zio->io_bp));
6610 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6613 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6614 VERIFY3S(remove_reference(hdr, hdr), >, 0);
6615 /* if it's not anon, we are doing a scrub */
6616 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6617 arc_access(hdr, 0, B_FALSE);
6618 mutex_exit(hash_lock);
6620 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6621 VERIFY3S(remove_reference(hdr, hdr), >, 0);
6624 callback->awcb_done(zio, buf, callback->awcb_private);
6626 abd_free(zio->io_abd);
6627 kmem_free(callback, sizeof (arc_write_callback_t));
6631 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
6632 blkptr_t *bp, arc_buf_t *buf, boolean_t uncached, boolean_t l2arc,
6633 const zio_prop_t *zp, arc_write_done_func_t *ready,
6634 arc_write_done_func_t *children_ready, arc_write_done_func_t *done,
6635 void *private, zio_priority_t priority, int zio_flags,
6636 const zbookmark_phys_t *zb)
6638 arc_buf_hdr_t *hdr = buf->b_hdr;
6639 arc_write_callback_t *callback;
6641 zio_prop_t localprop = *zp;
6643 ASSERT3P(ready, !=, NULL);
6644 ASSERT3P(done, !=, NULL);
6645 ASSERT(!HDR_IO_ERROR(hdr));
6646 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6647 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6648 ASSERT3P(hdr->b_l1hdr.b_buf, !=, NULL);
6650 arc_hdr_set_flags(hdr, ARC_FLAG_UNCACHED);
6652 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6654 if (ARC_BUF_ENCRYPTED(buf)) {
6655 ASSERT(ARC_BUF_COMPRESSED(buf));
6656 localprop.zp_encrypt = B_TRUE;
6657 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6658 localprop.zp_complevel = hdr->b_complevel;
6659 localprop.zp_byteorder =
6660 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
6661 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
6662 memcpy(localprop.zp_salt, hdr->b_crypt_hdr.b_salt,
6664 memcpy(localprop.zp_iv, hdr->b_crypt_hdr.b_iv,
6666 memcpy(localprop.zp_mac, hdr->b_crypt_hdr.b_mac,
6668 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
6669 localprop.zp_nopwrite = B_FALSE;
6670 localprop.zp_copies =
6671 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
6673 zio_flags |= ZIO_FLAG_RAW;
6674 } else if (ARC_BUF_COMPRESSED(buf)) {
6675 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6676 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6677 localprop.zp_complevel = hdr->b_complevel;
6678 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6680 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6681 callback->awcb_ready = ready;
6682 callback->awcb_children_ready = children_ready;
6683 callback->awcb_done = done;
6684 callback->awcb_private = private;
6685 callback->awcb_buf = buf;
6688 * The hdr's b_pabd is now stale, free it now. A new data block
6689 * will be allocated when the zio pipeline calls arc_write_ready().
6691 if (hdr->b_l1hdr.b_pabd != NULL) {
6693 * If the buf is currently sharing the data block with
6694 * the hdr then we need to break that relationship here.
6695 * The hdr will remain with a NULL data pointer and the
6696 * buf will take sole ownership of the block.
6698 if (ARC_BUF_SHARED(buf)) {
6699 arc_unshare_buf(hdr, buf);
6701 ASSERT(!arc_buf_is_shared(buf));
6702 arc_hdr_free_abd(hdr, B_FALSE);
6704 VERIFY3P(buf->b_data, !=, NULL);
6707 if (HDR_HAS_RABD(hdr))
6708 arc_hdr_free_abd(hdr, B_TRUE);
6710 if (!(zio_flags & ZIO_FLAG_RAW))
6711 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6713 ASSERT(!arc_buf_is_shared(buf));
6714 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6716 zio = zio_write(pio, spa, txg, bp,
6717 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6718 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6719 (children_ready != NULL) ? arc_write_children_ready : NULL,
6720 arc_write_done, callback, priority, zio_flags, zb);
6726 arc_tempreserve_clear(uint64_t reserve)
6728 atomic_add_64(&arc_tempreserve, -reserve);
6729 ASSERT((int64_t)arc_tempreserve >= 0);
6733 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
6739 reserve > arc_c/4 &&
6740 reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT))
6741 arc_c = MIN(arc_c_max, reserve * 4);
6744 * Throttle when the calculated memory footprint for the TXG
6745 * exceeds the target ARC size.
6747 if (reserve > arc_c) {
6748 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
6749 return (SET_ERROR(ERESTART));
6753 * Don't count loaned bufs as in flight dirty data to prevent long
6754 * network delays from blocking transactions that are ready to be
6755 * assigned to a txg.
6758 /* assert that it has not wrapped around */
6759 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6761 anon_size = MAX((int64_t)
6762 (zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_DATA]) +
6763 zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_METADATA]) -
6764 arc_loaned_bytes), 0);
6767 * Writes will, almost always, require additional memory allocations
6768 * in order to compress/encrypt/etc the data. We therefore need to
6769 * make sure that there is sufficient available memory for this.
6771 error = arc_memory_throttle(spa, reserve, txg);
6776 * Throttle writes when the amount of dirty data in the cache
6777 * gets too large. We try to keep the cache less than half full
6778 * of dirty blocks so that our sync times don't grow too large.
6780 * In the case of one pool being built on another pool, we want
6781 * to make sure we don't end up throttling the lower (backing)
6782 * pool when the upper pool is the majority contributor to dirty
6783 * data. To insure we make forward progress during throttling, we
6784 * also check the current pool's net dirty data and only throttle
6785 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
6786 * data in the cache.
6788 * Note: if two requests come in concurrently, we might let them
6789 * both succeed, when one of them should fail. Not a huge deal.
6791 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
6792 uint64_t spa_dirty_anon = spa_dirty_data(spa);
6793 uint64_t rarc_c = arc_warm ? arc_c : arc_c_max;
6794 if (total_dirty > rarc_c * zfs_arc_dirty_limit_percent / 100 &&
6795 anon_size > rarc_c * zfs_arc_anon_limit_percent / 100 &&
6796 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
6798 uint64_t meta_esize = zfs_refcount_count(
6799 &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6800 uint64_t data_esize =
6801 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6802 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6803 "anon_data=%lluK tempreserve=%lluK rarc_c=%lluK\n",
6804 (u_longlong_t)arc_tempreserve >> 10,
6805 (u_longlong_t)meta_esize >> 10,
6806 (u_longlong_t)data_esize >> 10,
6807 (u_longlong_t)reserve >> 10,
6808 (u_longlong_t)rarc_c >> 10);
6810 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
6811 return (SET_ERROR(ERESTART));
6813 atomic_add_64(&arc_tempreserve, reserve);
6818 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6819 kstat_named_t *data, kstat_named_t *metadata,
6820 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6823 zfs_refcount_count(&state->arcs_size[ARC_BUFC_DATA]);
6824 metadata->value.ui64 =
6825 zfs_refcount_count(&state->arcs_size[ARC_BUFC_METADATA]);
6826 size->value.ui64 = data->value.ui64 + metadata->value.ui64;
6827 evict_data->value.ui64 =
6828 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6829 evict_metadata->value.ui64 =
6830 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6834 arc_kstat_update(kstat_t *ksp, int rw)
6836 arc_stats_t *as = ksp->ks_data;
6838 if (rw == KSTAT_WRITE)
6839 return (SET_ERROR(EACCES));
6841 as->arcstat_hits.value.ui64 =
6842 wmsum_value(&arc_sums.arcstat_hits);
6843 as->arcstat_iohits.value.ui64 =
6844 wmsum_value(&arc_sums.arcstat_iohits);
6845 as->arcstat_misses.value.ui64 =
6846 wmsum_value(&arc_sums.arcstat_misses);
6847 as->arcstat_demand_data_hits.value.ui64 =
6848 wmsum_value(&arc_sums.arcstat_demand_data_hits);
6849 as->arcstat_demand_data_iohits.value.ui64 =
6850 wmsum_value(&arc_sums.arcstat_demand_data_iohits);
6851 as->arcstat_demand_data_misses.value.ui64 =
6852 wmsum_value(&arc_sums.arcstat_demand_data_misses);
6853 as->arcstat_demand_metadata_hits.value.ui64 =
6854 wmsum_value(&arc_sums.arcstat_demand_metadata_hits);
6855 as->arcstat_demand_metadata_iohits.value.ui64 =
6856 wmsum_value(&arc_sums.arcstat_demand_metadata_iohits);
6857 as->arcstat_demand_metadata_misses.value.ui64 =
6858 wmsum_value(&arc_sums.arcstat_demand_metadata_misses);
6859 as->arcstat_prefetch_data_hits.value.ui64 =
6860 wmsum_value(&arc_sums.arcstat_prefetch_data_hits);
6861 as->arcstat_prefetch_data_iohits.value.ui64 =
6862 wmsum_value(&arc_sums.arcstat_prefetch_data_iohits);
6863 as->arcstat_prefetch_data_misses.value.ui64 =
6864 wmsum_value(&arc_sums.arcstat_prefetch_data_misses);
6865 as->arcstat_prefetch_metadata_hits.value.ui64 =
6866 wmsum_value(&arc_sums.arcstat_prefetch_metadata_hits);
6867 as->arcstat_prefetch_metadata_iohits.value.ui64 =
6868 wmsum_value(&arc_sums.arcstat_prefetch_metadata_iohits);
6869 as->arcstat_prefetch_metadata_misses.value.ui64 =
6870 wmsum_value(&arc_sums.arcstat_prefetch_metadata_misses);
6871 as->arcstat_mru_hits.value.ui64 =
6872 wmsum_value(&arc_sums.arcstat_mru_hits);
6873 as->arcstat_mru_ghost_hits.value.ui64 =
6874 wmsum_value(&arc_sums.arcstat_mru_ghost_hits);
6875 as->arcstat_mfu_hits.value.ui64 =
6876 wmsum_value(&arc_sums.arcstat_mfu_hits);
6877 as->arcstat_mfu_ghost_hits.value.ui64 =
6878 wmsum_value(&arc_sums.arcstat_mfu_ghost_hits);
6879 as->arcstat_uncached_hits.value.ui64 =
6880 wmsum_value(&arc_sums.arcstat_uncached_hits);
6881 as->arcstat_deleted.value.ui64 =
6882 wmsum_value(&arc_sums.arcstat_deleted);
6883 as->arcstat_mutex_miss.value.ui64 =
6884 wmsum_value(&arc_sums.arcstat_mutex_miss);
6885 as->arcstat_access_skip.value.ui64 =
6886 wmsum_value(&arc_sums.arcstat_access_skip);
6887 as->arcstat_evict_skip.value.ui64 =
6888 wmsum_value(&arc_sums.arcstat_evict_skip);
6889 as->arcstat_evict_not_enough.value.ui64 =
6890 wmsum_value(&arc_sums.arcstat_evict_not_enough);
6891 as->arcstat_evict_l2_cached.value.ui64 =
6892 wmsum_value(&arc_sums.arcstat_evict_l2_cached);
6893 as->arcstat_evict_l2_eligible.value.ui64 =
6894 wmsum_value(&arc_sums.arcstat_evict_l2_eligible);
6895 as->arcstat_evict_l2_eligible_mfu.value.ui64 =
6896 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mfu);
6897 as->arcstat_evict_l2_eligible_mru.value.ui64 =
6898 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mru);
6899 as->arcstat_evict_l2_ineligible.value.ui64 =
6900 wmsum_value(&arc_sums.arcstat_evict_l2_ineligible);
6901 as->arcstat_evict_l2_skip.value.ui64 =
6902 wmsum_value(&arc_sums.arcstat_evict_l2_skip);
6903 as->arcstat_hash_collisions.value.ui64 =
6904 wmsum_value(&arc_sums.arcstat_hash_collisions);
6905 as->arcstat_hash_chains.value.ui64 =
6906 wmsum_value(&arc_sums.arcstat_hash_chains);
6907 as->arcstat_size.value.ui64 =
6908 aggsum_value(&arc_sums.arcstat_size);
6909 as->arcstat_compressed_size.value.ui64 =
6910 wmsum_value(&arc_sums.arcstat_compressed_size);
6911 as->arcstat_uncompressed_size.value.ui64 =
6912 wmsum_value(&arc_sums.arcstat_uncompressed_size);
6913 as->arcstat_overhead_size.value.ui64 =
6914 wmsum_value(&arc_sums.arcstat_overhead_size);
6915 as->arcstat_hdr_size.value.ui64 =
6916 wmsum_value(&arc_sums.arcstat_hdr_size);
6917 as->arcstat_data_size.value.ui64 =
6918 wmsum_value(&arc_sums.arcstat_data_size);
6919 as->arcstat_metadata_size.value.ui64 =
6920 wmsum_value(&arc_sums.arcstat_metadata_size);
6921 as->arcstat_dbuf_size.value.ui64 =
6922 wmsum_value(&arc_sums.arcstat_dbuf_size);
6923 #if defined(COMPAT_FREEBSD11)
6924 as->arcstat_other_size.value.ui64 =
6925 wmsum_value(&arc_sums.arcstat_bonus_size) +
6926 wmsum_value(&arc_sums.arcstat_dnode_size) +
6927 wmsum_value(&arc_sums.arcstat_dbuf_size);
6930 arc_kstat_update_state(arc_anon,
6931 &as->arcstat_anon_size,
6932 &as->arcstat_anon_data,
6933 &as->arcstat_anon_metadata,
6934 &as->arcstat_anon_evictable_data,
6935 &as->arcstat_anon_evictable_metadata);
6936 arc_kstat_update_state(arc_mru,
6937 &as->arcstat_mru_size,
6938 &as->arcstat_mru_data,
6939 &as->arcstat_mru_metadata,
6940 &as->arcstat_mru_evictable_data,
6941 &as->arcstat_mru_evictable_metadata);
6942 arc_kstat_update_state(arc_mru_ghost,
6943 &as->arcstat_mru_ghost_size,
6944 &as->arcstat_mru_ghost_data,
6945 &as->arcstat_mru_ghost_metadata,
6946 &as->arcstat_mru_ghost_evictable_data,
6947 &as->arcstat_mru_ghost_evictable_metadata);
6948 arc_kstat_update_state(arc_mfu,
6949 &as->arcstat_mfu_size,
6950 &as->arcstat_mfu_data,
6951 &as->arcstat_mfu_metadata,
6952 &as->arcstat_mfu_evictable_data,
6953 &as->arcstat_mfu_evictable_metadata);
6954 arc_kstat_update_state(arc_mfu_ghost,
6955 &as->arcstat_mfu_ghost_size,
6956 &as->arcstat_mfu_ghost_data,
6957 &as->arcstat_mfu_ghost_metadata,
6958 &as->arcstat_mfu_ghost_evictable_data,
6959 &as->arcstat_mfu_ghost_evictable_metadata);
6960 arc_kstat_update_state(arc_uncached,
6961 &as->arcstat_uncached_size,
6962 &as->arcstat_uncached_data,
6963 &as->arcstat_uncached_metadata,
6964 &as->arcstat_uncached_evictable_data,
6965 &as->arcstat_uncached_evictable_metadata);
6967 as->arcstat_dnode_size.value.ui64 =
6968 wmsum_value(&arc_sums.arcstat_dnode_size);
6969 as->arcstat_bonus_size.value.ui64 =
6970 wmsum_value(&arc_sums.arcstat_bonus_size);
6971 as->arcstat_l2_hits.value.ui64 =
6972 wmsum_value(&arc_sums.arcstat_l2_hits);
6973 as->arcstat_l2_misses.value.ui64 =
6974 wmsum_value(&arc_sums.arcstat_l2_misses);
6975 as->arcstat_l2_prefetch_asize.value.ui64 =
6976 wmsum_value(&arc_sums.arcstat_l2_prefetch_asize);
6977 as->arcstat_l2_mru_asize.value.ui64 =
6978 wmsum_value(&arc_sums.arcstat_l2_mru_asize);
6979 as->arcstat_l2_mfu_asize.value.ui64 =
6980 wmsum_value(&arc_sums.arcstat_l2_mfu_asize);
6981 as->arcstat_l2_bufc_data_asize.value.ui64 =
6982 wmsum_value(&arc_sums.arcstat_l2_bufc_data_asize);
6983 as->arcstat_l2_bufc_metadata_asize.value.ui64 =
6984 wmsum_value(&arc_sums.arcstat_l2_bufc_metadata_asize);
6985 as->arcstat_l2_feeds.value.ui64 =
6986 wmsum_value(&arc_sums.arcstat_l2_feeds);
6987 as->arcstat_l2_rw_clash.value.ui64 =
6988 wmsum_value(&arc_sums.arcstat_l2_rw_clash);
6989 as->arcstat_l2_read_bytes.value.ui64 =
6990 wmsum_value(&arc_sums.arcstat_l2_read_bytes);
6991 as->arcstat_l2_write_bytes.value.ui64 =
6992 wmsum_value(&arc_sums.arcstat_l2_write_bytes);
6993 as->arcstat_l2_writes_sent.value.ui64 =
6994 wmsum_value(&arc_sums.arcstat_l2_writes_sent);
6995 as->arcstat_l2_writes_done.value.ui64 =
6996 wmsum_value(&arc_sums.arcstat_l2_writes_done);
6997 as->arcstat_l2_writes_error.value.ui64 =
6998 wmsum_value(&arc_sums.arcstat_l2_writes_error);
6999 as->arcstat_l2_writes_lock_retry.value.ui64 =
7000 wmsum_value(&arc_sums.arcstat_l2_writes_lock_retry);
7001 as->arcstat_l2_evict_lock_retry.value.ui64 =
7002 wmsum_value(&arc_sums.arcstat_l2_evict_lock_retry);
7003 as->arcstat_l2_evict_reading.value.ui64 =
7004 wmsum_value(&arc_sums.arcstat_l2_evict_reading);
7005 as->arcstat_l2_evict_l1cached.value.ui64 =
7006 wmsum_value(&arc_sums.arcstat_l2_evict_l1cached);
7007 as->arcstat_l2_free_on_write.value.ui64 =
7008 wmsum_value(&arc_sums.arcstat_l2_free_on_write);
7009 as->arcstat_l2_abort_lowmem.value.ui64 =
7010 wmsum_value(&arc_sums.arcstat_l2_abort_lowmem);
7011 as->arcstat_l2_cksum_bad.value.ui64 =
7012 wmsum_value(&arc_sums.arcstat_l2_cksum_bad);
7013 as->arcstat_l2_io_error.value.ui64 =
7014 wmsum_value(&arc_sums.arcstat_l2_io_error);
7015 as->arcstat_l2_lsize.value.ui64 =
7016 wmsum_value(&arc_sums.arcstat_l2_lsize);
7017 as->arcstat_l2_psize.value.ui64 =
7018 wmsum_value(&arc_sums.arcstat_l2_psize);
7019 as->arcstat_l2_hdr_size.value.ui64 =
7020 aggsum_value(&arc_sums.arcstat_l2_hdr_size);
7021 as->arcstat_l2_log_blk_writes.value.ui64 =
7022 wmsum_value(&arc_sums.arcstat_l2_log_blk_writes);
7023 as->arcstat_l2_log_blk_asize.value.ui64 =
7024 wmsum_value(&arc_sums.arcstat_l2_log_blk_asize);
7025 as->arcstat_l2_log_blk_count.value.ui64 =
7026 wmsum_value(&arc_sums.arcstat_l2_log_blk_count);
7027 as->arcstat_l2_rebuild_success.value.ui64 =
7028 wmsum_value(&arc_sums.arcstat_l2_rebuild_success);
7029 as->arcstat_l2_rebuild_abort_unsupported.value.ui64 =
7030 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7031 as->arcstat_l2_rebuild_abort_io_errors.value.ui64 =
7032 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7033 as->arcstat_l2_rebuild_abort_dh_errors.value.ui64 =
7034 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7035 as->arcstat_l2_rebuild_abort_cksum_lb_errors.value.ui64 =
7036 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7037 as->arcstat_l2_rebuild_abort_lowmem.value.ui64 =
7038 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7039 as->arcstat_l2_rebuild_size.value.ui64 =
7040 wmsum_value(&arc_sums.arcstat_l2_rebuild_size);
7041 as->arcstat_l2_rebuild_asize.value.ui64 =
7042 wmsum_value(&arc_sums.arcstat_l2_rebuild_asize);
7043 as->arcstat_l2_rebuild_bufs.value.ui64 =
7044 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs);
7045 as->arcstat_l2_rebuild_bufs_precached.value.ui64 =
7046 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7047 as->arcstat_l2_rebuild_log_blks.value.ui64 =
7048 wmsum_value(&arc_sums.arcstat_l2_rebuild_log_blks);
7049 as->arcstat_memory_throttle_count.value.ui64 =
7050 wmsum_value(&arc_sums.arcstat_memory_throttle_count);
7051 as->arcstat_memory_direct_count.value.ui64 =
7052 wmsum_value(&arc_sums.arcstat_memory_direct_count);
7053 as->arcstat_memory_indirect_count.value.ui64 =
7054 wmsum_value(&arc_sums.arcstat_memory_indirect_count);
7056 as->arcstat_memory_all_bytes.value.ui64 =
7058 as->arcstat_memory_free_bytes.value.ui64 =
7060 as->arcstat_memory_available_bytes.value.i64 =
7061 arc_available_memory();
7063 as->arcstat_prune.value.ui64 =
7064 wmsum_value(&arc_sums.arcstat_prune);
7065 as->arcstat_meta_used.value.ui64 =
7066 wmsum_value(&arc_sums.arcstat_meta_used);
7067 as->arcstat_async_upgrade_sync.value.ui64 =
7068 wmsum_value(&arc_sums.arcstat_async_upgrade_sync);
7069 as->arcstat_predictive_prefetch.value.ui64 =
7070 wmsum_value(&arc_sums.arcstat_predictive_prefetch);
7071 as->arcstat_demand_hit_predictive_prefetch.value.ui64 =
7072 wmsum_value(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7073 as->arcstat_demand_iohit_predictive_prefetch.value.ui64 =
7074 wmsum_value(&arc_sums.arcstat_demand_iohit_predictive_prefetch);
7075 as->arcstat_prescient_prefetch.value.ui64 =
7076 wmsum_value(&arc_sums.arcstat_prescient_prefetch);
7077 as->arcstat_demand_hit_prescient_prefetch.value.ui64 =
7078 wmsum_value(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7079 as->arcstat_demand_iohit_prescient_prefetch.value.ui64 =
7080 wmsum_value(&arc_sums.arcstat_demand_iohit_prescient_prefetch);
7081 as->arcstat_raw_size.value.ui64 =
7082 wmsum_value(&arc_sums.arcstat_raw_size);
7083 as->arcstat_cached_only_in_progress.value.ui64 =
7084 wmsum_value(&arc_sums.arcstat_cached_only_in_progress);
7085 as->arcstat_abd_chunk_waste_size.value.ui64 =
7086 wmsum_value(&arc_sums.arcstat_abd_chunk_waste_size);
7092 * This function *must* return indices evenly distributed between all
7093 * sublists of the multilist. This is needed due to how the ARC eviction
7094 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7095 * distributed between all sublists and uses this assumption when
7096 * deciding which sublist to evict from and how much to evict from it.
7099 arc_state_multilist_index_func(multilist_t *ml, void *obj)
7101 arc_buf_hdr_t *hdr = obj;
7104 * We rely on b_dva to generate evenly distributed index
7105 * numbers using buf_hash below. So, as an added precaution,
7106 * let's make sure we never add empty buffers to the arc lists.
7108 ASSERT(!HDR_EMPTY(hdr));
7111 * The assumption here, is the hash value for a given
7112 * arc_buf_hdr_t will remain constant throughout its lifetime
7113 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7114 * Thus, we don't need to store the header's sublist index
7115 * on insertion, as this index can be recalculated on removal.
7117 * Also, the low order bits of the hash value are thought to be
7118 * distributed evenly. Otherwise, in the case that the multilist
7119 * has a power of two number of sublists, each sublists' usage
7120 * would not be evenly distributed. In this context full 64bit
7121 * division would be a waste of time, so limit it to 32 bits.
7123 return ((unsigned int)buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
7124 multilist_get_num_sublists(ml));
7128 arc_state_l2c_multilist_index_func(multilist_t *ml, void *obj)
7130 panic("Header %p insert into arc_l2c_only %p", obj, ml);
7133 #define WARN_IF_TUNING_IGNORED(tuning, value, do_warn) do { \
7134 if ((do_warn) && (tuning) && ((tuning) != (value))) { \
7136 "ignoring tunable %s (using %llu instead)", \
7137 (#tuning), (u_longlong_t)(value)); \
7142 * Called during module initialization and periodically thereafter to
7143 * apply reasonable changes to the exposed performance tunings. Can also be
7144 * called explicitly by param_set_arc_*() functions when ARC tunables are
7145 * updated manually. Non-zero zfs_* values which differ from the currently set
7146 * values will be applied.
7149 arc_tuning_update(boolean_t verbose)
7151 uint64_t allmem = arc_all_memory();
7153 /* Valid range: 32M - <arc_c_max> */
7154 if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) &&
7155 (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) &&
7156 (zfs_arc_min <= arc_c_max)) {
7157 arc_c_min = zfs_arc_min;
7158 arc_c = MAX(arc_c, arc_c_min);
7160 WARN_IF_TUNING_IGNORED(zfs_arc_min, arc_c_min, verbose);
7162 /* Valid range: 64M - <all physical memory> */
7163 if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) &&
7164 (zfs_arc_max >= MIN_ARC_MAX) && (zfs_arc_max < allmem) &&
7165 (zfs_arc_max > arc_c_min)) {
7166 arc_c_max = zfs_arc_max;
7167 arc_c = MIN(arc_c, arc_c_max);
7168 if (arc_dnode_limit > arc_c_max)
7169 arc_dnode_limit = arc_c_max;
7171 WARN_IF_TUNING_IGNORED(zfs_arc_max, arc_c_max, verbose);
7173 /* Valid range: 0 - <all physical memory> */
7174 arc_dnode_limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit :
7175 MIN(zfs_arc_dnode_limit_percent, 100) * arc_c_max / 100;
7176 WARN_IF_TUNING_IGNORED(zfs_arc_dnode_limit, arc_dnode_limit, verbose);
7178 /* Valid range: 1 - N */
7179 if (zfs_arc_grow_retry)
7180 arc_grow_retry = zfs_arc_grow_retry;
7182 /* Valid range: 1 - N */
7183 if (zfs_arc_shrink_shift) {
7184 arc_shrink_shift = zfs_arc_shrink_shift;
7185 arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1);
7188 /* Valid range: 1 - N ms */
7189 if (zfs_arc_min_prefetch_ms)
7190 arc_min_prefetch_ms = zfs_arc_min_prefetch_ms;
7192 /* Valid range: 1 - N ms */
7193 if (zfs_arc_min_prescient_prefetch_ms) {
7194 arc_min_prescient_prefetch_ms =
7195 zfs_arc_min_prescient_prefetch_ms;
7198 /* Valid range: 0 - 100 */
7199 if (zfs_arc_lotsfree_percent <= 100)
7200 arc_lotsfree_percent = zfs_arc_lotsfree_percent;
7201 WARN_IF_TUNING_IGNORED(zfs_arc_lotsfree_percent, arc_lotsfree_percent,
7204 /* Valid range: 0 - <all physical memory> */
7205 if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free))
7206 arc_sys_free = MIN(zfs_arc_sys_free, allmem);
7207 WARN_IF_TUNING_IGNORED(zfs_arc_sys_free, arc_sys_free, verbose);
7211 arc_state_multilist_init(multilist_t *ml,
7212 multilist_sublist_index_func_t *index_func, int *maxcountp)
7214 multilist_create(ml, sizeof (arc_buf_hdr_t),
7215 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), index_func);
7216 *maxcountp = MAX(*maxcountp, multilist_get_num_sublists(ml));
7220 arc_state_init(void)
7222 int num_sublists = 0;
7224 arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_METADATA],
7225 arc_state_multilist_index_func, &num_sublists);
7226 arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_DATA],
7227 arc_state_multilist_index_func, &num_sublists);
7228 arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
7229 arc_state_multilist_index_func, &num_sublists);
7230 arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
7231 arc_state_multilist_index_func, &num_sublists);
7232 arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
7233 arc_state_multilist_index_func, &num_sublists);
7234 arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_DATA],
7235 arc_state_multilist_index_func, &num_sublists);
7236 arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
7237 arc_state_multilist_index_func, &num_sublists);
7238 arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
7239 arc_state_multilist_index_func, &num_sublists);
7240 arc_state_multilist_init(&arc_uncached->arcs_list[ARC_BUFC_METADATA],
7241 arc_state_multilist_index_func, &num_sublists);
7242 arc_state_multilist_init(&arc_uncached->arcs_list[ARC_BUFC_DATA],
7243 arc_state_multilist_index_func, &num_sublists);
7246 * L2 headers should never be on the L2 state list since they don't
7247 * have L1 headers allocated. Special index function asserts that.
7249 arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
7250 arc_state_l2c_multilist_index_func, &num_sublists);
7251 arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
7252 arc_state_l2c_multilist_index_func, &num_sublists);
7255 * Keep track of the number of markers needed to reclaim buffers from
7256 * any ARC state. The markers will be pre-allocated so as to minimize
7257 * the number of memory allocations performed by the eviction thread.
7259 arc_state_evict_marker_count = num_sublists;
7261 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7262 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7263 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7264 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7265 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7266 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7267 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7268 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7269 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7270 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7271 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7272 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7273 zfs_refcount_create(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]);
7274 zfs_refcount_create(&arc_uncached->arcs_esize[ARC_BUFC_DATA]);
7276 zfs_refcount_create(&arc_anon->arcs_size[ARC_BUFC_DATA]);
7277 zfs_refcount_create(&arc_anon->arcs_size[ARC_BUFC_METADATA]);
7278 zfs_refcount_create(&arc_mru->arcs_size[ARC_BUFC_DATA]);
7279 zfs_refcount_create(&arc_mru->arcs_size[ARC_BUFC_METADATA]);
7280 zfs_refcount_create(&arc_mru_ghost->arcs_size[ARC_BUFC_DATA]);
7281 zfs_refcount_create(&arc_mru_ghost->arcs_size[ARC_BUFC_METADATA]);
7282 zfs_refcount_create(&arc_mfu->arcs_size[ARC_BUFC_DATA]);
7283 zfs_refcount_create(&arc_mfu->arcs_size[ARC_BUFC_METADATA]);
7284 zfs_refcount_create(&arc_mfu_ghost->arcs_size[ARC_BUFC_DATA]);
7285 zfs_refcount_create(&arc_mfu_ghost->arcs_size[ARC_BUFC_METADATA]);
7286 zfs_refcount_create(&arc_l2c_only->arcs_size[ARC_BUFC_DATA]);
7287 zfs_refcount_create(&arc_l2c_only->arcs_size[ARC_BUFC_METADATA]);
7288 zfs_refcount_create(&arc_uncached->arcs_size[ARC_BUFC_DATA]);
7289 zfs_refcount_create(&arc_uncached->arcs_size[ARC_BUFC_METADATA]);
7291 wmsum_init(&arc_mru_ghost->arcs_hits[ARC_BUFC_DATA], 0);
7292 wmsum_init(&arc_mru_ghost->arcs_hits[ARC_BUFC_METADATA], 0);
7293 wmsum_init(&arc_mfu_ghost->arcs_hits[ARC_BUFC_DATA], 0);
7294 wmsum_init(&arc_mfu_ghost->arcs_hits[ARC_BUFC_METADATA], 0);
7296 wmsum_init(&arc_sums.arcstat_hits, 0);
7297 wmsum_init(&arc_sums.arcstat_iohits, 0);
7298 wmsum_init(&arc_sums.arcstat_misses, 0);
7299 wmsum_init(&arc_sums.arcstat_demand_data_hits, 0);
7300 wmsum_init(&arc_sums.arcstat_demand_data_iohits, 0);
7301 wmsum_init(&arc_sums.arcstat_demand_data_misses, 0);
7302 wmsum_init(&arc_sums.arcstat_demand_metadata_hits, 0);
7303 wmsum_init(&arc_sums.arcstat_demand_metadata_iohits, 0);
7304 wmsum_init(&arc_sums.arcstat_demand_metadata_misses, 0);
7305 wmsum_init(&arc_sums.arcstat_prefetch_data_hits, 0);
7306 wmsum_init(&arc_sums.arcstat_prefetch_data_iohits, 0);
7307 wmsum_init(&arc_sums.arcstat_prefetch_data_misses, 0);
7308 wmsum_init(&arc_sums.arcstat_prefetch_metadata_hits, 0);
7309 wmsum_init(&arc_sums.arcstat_prefetch_metadata_iohits, 0);
7310 wmsum_init(&arc_sums.arcstat_prefetch_metadata_misses, 0);
7311 wmsum_init(&arc_sums.arcstat_mru_hits, 0);
7312 wmsum_init(&arc_sums.arcstat_mru_ghost_hits, 0);
7313 wmsum_init(&arc_sums.arcstat_mfu_hits, 0);
7314 wmsum_init(&arc_sums.arcstat_mfu_ghost_hits, 0);
7315 wmsum_init(&arc_sums.arcstat_uncached_hits, 0);
7316 wmsum_init(&arc_sums.arcstat_deleted, 0);
7317 wmsum_init(&arc_sums.arcstat_mutex_miss, 0);
7318 wmsum_init(&arc_sums.arcstat_access_skip, 0);
7319 wmsum_init(&arc_sums.arcstat_evict_skip, 0);
7320 wmsum_init(&arc_sums.arcstat_evict_not_enough, 0);
7321 wmsum_init(&arc_sums.arcstat_evict_l2_cached, 0);
7322 wmsum_init(&arc_sums.arcstat_evict_l2_eligible, 0);
7323 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mfu, 0);
7324 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mru, 0);
7325 wmsum_init(&arc_sums.arcstat_evict_l2_ineligible, 0);
7326 wmsum_init(&arc_sums.arcstat_evict_l2_skip, 0);
7327 wmsum_init(&arc_sums.arcstat_hash_collisions, 0);
7328 wmsum_init(&arc_sums.arcstat_hash_chains, 0);
7329 aggsum_init(&arc_sums.arcstat_size, 0);
7330 wmsum_init(&arc_sums.arcstat_compressed_size, 0);
7331 wmsum_init(&arc_sums.arcstat_uncompressed_size, 0);
7332 wmsum_init(&arc_sums.arcstat_overhead_size, 0);
7333 wmsum_init(&arc_sums.arcstat_hdr_size, 0);
7334 wmsum_init(&arc_sums.arcstat_data_size, 0);
7335 wmsum_init(&arc_sums.arcstat_metadata_size, 0);
7336 wmsum_init(&arc_sums.arcstat_dbuf_size, 0);
7337 wmsum_init(&arc_sums.arcstat_dnode_size, 0);
7338 wmsum_init(&arc_sums.arcstat_bonus_size, 0);
7339 wmsum_init(&arc_sums.arcstat_l2_hits, 0);
7340 wmsum_init(&arc_sums.arcstat_l2_misses, 0);
7341 wmsum_init(&arc_sums.arcstat_l2_prefetch_asize, 0);
7342 wmsum_init(&arc_sums.arcstat_l2_mru_asize, 0);
7343 wmsum_init(&arc_sums.arcstat_l2_mfu_asize, 0);
7344 wmsum_init(&arc_sums.arcstat_l2_bufc_data_asize, 0);
7345 wmsum_init(&arc_sums.arcstat_l2_bufc_metadata_asize, 0);
7346 wmsum_init(&arc_sums.arcstat_l2_feeds, 0);
7347 wmsum_init(&arc_sums.arcstat_l2_rw_clash, 0);
7348 wmsum_init(&arc_sums.arcstat_l2_read_bytes, 0);
7349 wmsum_init(&arc_sums.arcstat_l2_write_bytes, 0);
7350 wmsum_init(&arc_sums.arcstat_l2_writes_sent, 0);
7351 wmsum_init(&arc_sums.arcstat_l2_writes_done, 0);
7352 wmsum_init(&arc_sums.arcstat_l2_writes_error, 0);
7353 wmsum_init(&arc_sums.arcstat_l2_writes_lock_retry, 0);
7354 wmsum_init(&arc_sums.arcstat_l2_evict_lock_retry, 0);
7355 wmsum_init(&arc_sums.arcstat_l2_evict_reading, 0);
7356 wmsum_init(&arc_sums.arcstat_l2_evict_l1cached, 0);
7357 wmsum_init(&arc_sums.arcstat_l2_free_on_write, 0);
7358 wmsum_init(&arc_sums.arcstat_l2_abort_lowmem, 0);
7359 wmsum_init(&arc_sums.arcstat_l2_cksum_bad, 0);
7360 wmsum_init(&arc_sums.arcstat_l2_io_error, 0);
7361 wmsum_init(&arc_sums.arcstat_l2_lsize, 0);
7362 wmsum_init(&arc_sums.arcstat_l2_psize, 0);
7363 aggsum_init(&arc_sums.arcstat_l2_hdr_size, 0);
7364 wmsum_init(&arc_sums.arcstat_l2_log_blk_writes, 0);
7365 wmsum_init(&arc_sums.arcstat_l2_log_blk_asize, 0);
7366 wmsum_init(&arc_sums.arcstat_l2_log_blk_count, 0);
7367 wmsum_init(&arc_sums.arcstat_l2_rebuild_success, 0);
7368 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_unsupported, 0);
7369 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_io_errors, 0);
7370 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_dh_errors, 0);
7371 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors, 0);
7372 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_lowmem, 0);
7373 wmsum_init(&arc_sums.arcstat_l2_rebuild_size, 0);
7374 wmsum_init(&arc_sums.arcstat_l2_rebuild_asize, 0);
7375 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs, 0);
7376 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs_precached, 0);
7377 wmsum_init(&arc_sums.arcstat_l2_rebuild_log_blks, 0);
7378 wmsum_init(&arc_sums.arcstat_memory_throttle_count, 0);
7379 wmsum_init(&arc_sums.arcstat_memory_direct_count, 0);
7380 wmsum_init(&arc_sums.arcstat_memory_indirect_count, 0);
7381 wmsum_init(&arc_sums.arcstat_prune, 0);
7382 wmsum_init(&arc_sums.arcstat_meta_used, 0);
7383 wmsum_init(&arc_sums.arcstat_async_upgrade_sync, 0);
7384 wmsum_init(&arc_sums.arcstat_predictive_prefetch, 0);
7385 wmsum_init(&arc_sums.arcstat_demand_hit_predictive_prefetch, 0);
7386 wmsum_init(&arc_sums.arcstat_demand_iohit_predictive_prefetch, 0);
7387 wmsum_init(&arc_sums.arcstat_prescient_prefetch, 0);
7388 wmsum_init(&arc_sums.arcstat_demand_hit_prescient_prefetch, 0);
7389 wmsum_init(&arc_sums.arcstat_demand_iohit_prescient_prefetch, 0);
7390 wmsum_init(&arc_sums.arcstat_raw_size, 0);
7391 wmsum_init(&arc_sums.arcstat_cached_only_in_progress, 0);
7392 wmsum_init(&arc_sums.arcstat_abd_chunk_waste_size, 0);
7394 arc_anon->arcs_state = ARC_STATE_ANON;
7395 arc_mru->arcs_state = ARC_STATE_MRU;
7396 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7397 arc_mfu->arcs_state = ARC_STATE_MFU;
7398 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7399 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7400 arc_uncached->arcs_state = ARC_STATE_UNCACHED;
7404 arc_state_fini(void)
7406 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7407 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7408 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7409 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7410 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7411 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7412 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7413 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7414 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7415 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7416 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7417 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7418 zfs_refcount_destroy(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]);
7419 zfs_refcount_destroy(&arc_uncached->arcs_esize[ARC_BUFC_DATA]);
7421 zfs_refcount_destroy(&arc_anon->arcs_size[ARC_BUFC_DATA]);
7422 zfs_refcount_destroy(&arc_anon->arcs_size[ARC_BUFC_METADATA]);
7423 zfs_refcount_destroy(&arc_mru->arcs_size[ARC_BUFC_DATA]);
7424 zfs_refcount_destroy(&arc_mru->arcs_size[ARC_BUFC_METADATA]);
7425 zfs_refcount_destroy(&arc_mru_ghost->arcs_size[ARC_BUFC_DATA]);
7426 zfs_refcount_destroy(&arc_mru_ghost->arcs_size[ARC_BUFC_METADATA]);
7427 zfs_refcount_destroy(&arc_mfu->arcs_size[ARC_BUFC_DATA]);
7428 zfs_refcount_destroy(&arc_mfu->arcs_size[ARC_BUFC_METADATA]);
7429 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size[ARC_BUFC_DATA]);
7430 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size[ARC_BUFC_METADATA]);
7431 zfs_refcount_destroy(&arc_l2c_only->arcs_size[ARC_BUFC_DATA]);
7432 zfs_refcount_destroy(&arc_l2c_only->arcs_size[ARC_BUFC_METADATA]);
7433 zfs_refcount_destroy(&arc_uncached->arcs_size[ARC_BUFC_DATA]);
7434 zfs_refcount_destroy(&arc_uncached->arcs_size[ARC_BUFC_METADATA]);
7436 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
7437 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7438 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7439 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7440 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
7441 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7442 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
7443 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7444 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7445 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7446 multilist_destroy(&arc_uncached->arcs_list[ARC_BUFC_METADATA]);
7447 multilist_destroy(&arc_uncached->arcs_list[ARC_BUFC_DATA]);
7449 wmsum_fini(&arc_mru_ghost->arcs_hits[ARC_BUFC_DATA]);
7450 wmsum_fini(&arc_mru_ghost->arcs_hits[ARC_BUFC_METADATA]);
7451 wmsum_fini(&arc_mfu_ghost->arcs_hits[ARC_BUFC_DATA]);
7452 wmsum_fini(&arc_mfu_ghost->arcs_hits[ARC_BUFC_METADATA]);
7454 wmsum_fini(&arc_sums.arcstat_hits);
7455 wmsum_fini(&arc_sums.arcstat_iohits);
7456 wmsum_fini(&arc_sums.arcstat_misses);
7457 wmsum_fini(&arc_sums.arcstat_demand_data_hits);
7458 wmsum_fini(&arc_sums.arcstat_demand_data_iohits);
7459 wmsum_fini(&arc_sums.arcstat_demand_data_misses);
7460 wmsum_fini(&arc_sums.arcstat_demand_metadata_hits);
7461 wmsum_fini(&arc_sums.arcstat_demand_metadata_iohits);
7462 wmsum_fini(&arc_sums.arcstat_demand_metadata_misses);
7463 wmsum_fini(&arc_sums.arcstat_prefetch_data_hits);
7464 wmsum_fini(&arc_sums.arcstat_prefetch_data_iohits);
7465 wmsum_fini(&arc_sums.arcstat_prefetch_data_misses);
7466 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_hits);
7467 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_iohits);
7468 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_misses);
7469 wmsum_fini(&arc_sums.arcstat_mru_hits);
7470 wmsum_fini(&arc_sums.arcstat_mru_ghost_hits);
7471 wmsum_fini(&arc_sums.arcstat_mfu_hits);
7472 wmsum_fini(&arc_sums.arcstat_mfu_ghost_hits);
7473 wmsum_fini(&arc_sums.arcstat_uncached_hits);
7474 wmsum_fini(&arc_sums.arcstat_deleted);
7475 wmsum_fini(&arc_sums.arcstat_mutex_miss);
7476 wmsum_fini(&arc_sums.arcstat_access_skip);
7477 wmsum_fini(&arc_sums.arcstat_evict_skip);
7478 wmsum_fini(&arc_sums.arcstat_evict_not_enough);
7479 wmsum_fini(&arc_sums.arcstat_evict_l2_cached);
7480 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible);
7481 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mfu);
7482 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mru);
7483 wmsum_fini(&arc_sums.arcstat_evict_l2_ineligible);
7484 wmsum_fini(&arc_sums.arcstat_evict_l2_skip);
7485 wmsum_fini(&arc_sums.arcstat_hash_collisions);
7486 wmsum_fini(&arc_sums.arcstat_hash_chains);
7487 aggsum_fini(&arc_sums.arcstat_size);
7488 wmsum_fini(&arc_sums.arcstat_compressed_size);
7489 wmsum_fini(&arc_sums.arcstat_uncompressed_size);
7490 wmsum_fini(&arc_sums.arcstat_overhead_size);
7491 wmsum_fini(&arc_sums.arcstat_hdr_size);
7492 wmsum_fini(&arc_sums.arcstat_data_size);
7493 wmsum_fini(&arc_sums.arcstat_metadata_size);
7494 wmsum_fini(&arc_sums.arcstat_dbuf_size);
7495 wmsum_fini(&arc_sums.arcstat_dnode_size);
7496 wmsum_fini(&arc_sums.arcstat_bonus_size);
7497 wmsum_fini(&arc_sums.arcstat_l2_hits);
7498 wmsum_fini(&arc_sums.arcstat_l2_misses);
7499 wmsum_fini(&arc_sums.arcstat_l2_prefetch_asize);
7500 wmsum_fini(&arc_sums.arcstat_l2_mru_asize);
7501 wmsum_fini(&arc_sums.arcstat_l2_mfu_asize);
7502 wmsum_fini(&arc_sums.arcstat_l2_bufc_data_asize);
7503 wmsum_fini(&arc_sums.arcstat_l2_bufc_metadata_asize);
7504 wmsum_fini(&arc_sums.arcstat_l2_feeds);
7505 wmsum_fini(&arc_sums.arcstat_l2_rw_clash);
7506 wmsum_fini(&arc_sums.arcstat_l2_read_bytes);
7507 wmsum_fini(&arc_sums.arcstat_l2_write_bytes);
7508 wmsum_fini(&arc_sums.arcstat_l2_writes_sent);
7509 wmsum_fini(&arc_sums.arcstat_l2_writes_done);
7510 wmsum_fini(&arc_sums.arcstat_l2_writes_error);
7511 wmsum_fini(&arc_sums.arcstat_l2_writes_lock_retry);
7512 wmsum_fini(&arc_sums.arcstat_l2_evict_lock_retry);
7513 wmsum_fini(&arc_sums.arcstat_l2_evict_reading);
7514 wmsum_fini(&arc_sums.arcstat_l2_evict_l1cached);
7515 wmsum_fini(&arc_sums.arcstat_l2_free_on_write);
7516 wmsum_fini(&arc_sums.arcstat_l2_abort_lowmem);
7517 wmsum_fini(&arc_sums.arcstat_l2_cksum_bad);
7518 wmsum_fini(&arc_sums.arcstat_l2_io_error);
7519 wmsum_fini(&arc_sums.arcstat_l2_lsize);
7520 wmsum_fini(&arc_sums.arcstat_l2_psize);
7521 aggsum_fini(&arc_sums.arcstat_l2_hdr_size);
7522 wmsum_fini(&arc_sums.arcstat_l2_log_blk_writes);
7523 wmsum_fini(&arc_sums.arcstat_l2_log_blk_asize);
7524 wmsum_fini(&arc_sums.arcstat_l2_log_blk_count);
7525 wmsum_fini(&arc_sums.arcstat_l2_rebuild_success);
7526 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7527 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7528 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7529 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7530 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7531 wmsum_fini(&arc_sums.arcstat_l2_rebuild_size);
7532 wmsum_fini(&arc_sums.arcstat_l2_rebuild_asize);
7533 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs);
7534 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7535 wmsum_fini(&arc_sums.arcstat_l2_rebuild_log_blks);
7536 wmsum_fini(&arc_sums.arcstat_memory_throttle_count);
7537 wmsum_fini(&arc_sums.arcstat_memory_direct_count);
7538 wmsum_fini(&arc_sums.arcstat_memory_indirect_count);
7539 wmsum_fini(&arc_sums.arcstat_prune);
7540 wmsum_fini(&arc_sums.arcstat_meta_used);
7541 wmsum_fini(&arc_sums.arcstat_async_upgrade_sync);
7542 wmsum_fini(&arc_sums.arcstat_predictive_prefetch);
7543 wmsum_fini(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7544 wmsum_fini(&arc_sums.arcstat_demand_iohit_predictive_prefetch);
7545 wmsum_fini(&arc_sums.arcstat_prescient_prefetch);
7546 wmsum_fini(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7547 wmsum_fini(&arc_sums.arcstat_demand_iohit_prescient_prefetch);
7548 wmsum_fini(&arc_sums.arcstat_raw_size);
7549 wmsum_fini(&arc_sums.arcstat_cached_only_in_progress);
7550 wmsum_fini(&arc_sums.arcstat_abd_chunk_waste_size);
7554 arc_target_bytes(void)
7560 arc_set_limits(uint64_t allmem)
7562 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more. */
7563 arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT);
7565 /* How to set default max varies by platform. */
7566 arc_c_max = arc_default_max(arc_c_min, allmem);
7571 uint64_t percent, allmem = arc_all_memory();
7572 mutex_init(&arc_evict_lock, NULL, MUTEX_DEFAULT, NULL);
7573 list_create(&arc_evict_waiters, sizeof (arc_evict_waiter_t),
7574 offsetof(arc_evict_waiter_t, aew_node));
7576 arc_min_prefetch_ms = 1000;
7577 arc_min_prescient_prefetch_ms = 6000;
7579 #if defined(_KERNEL)
7583 arc_set_limits(allmem);
7587 * If zfs_arc_max is non-zero at init, meaning it was set in the kernel
7588 * environment before the module was loaded, don't block setting the
7589 * maximum because it is less than arc_c_min, instead, reset arc_c_min
7591 * zfs_arc_min will be handled by arc_tuning_update().
7593 if (zfs_arc_max != 0 && zfs_arc_max >= MIN_ARC_MAX &&
7594 zfs_arc_max < allmem) {
7595 arc_c_max = zfs_arc_max;
7596 if (arc_c_min >= arc_c_max) {
7597 arc_c_min = MAX(zfs_arc_max / 2,
7598 2ULL << SPA_MAXBLOCKSHIFT);
7603 * In userland, there's only the memory pressure that we artificially
7604 * create (see arc_available_memory()). Don't let arc_c get too
7605 * small, because it can cause transactions to be larger than
7606 * arc_c, causing arc_tempreserve_space() to fail.
7608 arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT);
7613 * 32-bit fixed point fractions of metadata from total ARC size,
7614 * MRU data from all data and MRU metadata from all metadata.
7616 arc_meta = (1ULL << 32) / 4; /* Metadata is 25% of arc_c. */
7617 arc_pd = (1ULL << 32) / 2; /* Data MRU is 50% of data. */
7618 arc_pm = (1ULL << 32) / 2; /* Metadata MRU is 50% of metadata. */
7620 percent = MIN(zfs_arc_dnode_limit_percent, 100);
7621 arc_dnode_limit = arc_c_max * percent / 100;
7623 /* Apply user specified tunings */
7624 arc_tuning_update(B_TRUE);
7626 /* if kmem_flags are set, lets try to use less memory */
7627 if (kmem_debugging())
7629 if (arc_c < arc_c_min)
7632 arc_register_hotplug();
7638 list_create(&arc_prune_list, sizeof (arc_prune_t),
7639 offsetof(arc_prune_t, p_node));
7640 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7642 arc_prune_taskq = taskq_create("arc_prune", zfs_arc_prune_task_threads,
7643 defclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
7645 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7646 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7648 if (arc_ksp != NULL) {
7649 arc_ksp->ks_data = &arc_stats;
7650 arc_ksp->ks_update = arc_kstat_update;
7651 kstat_install(arc_ksp);
7654 arc_state_evict_markers =
7655 arc_state_alloc_markers(arc_state_evict_marker_count);
7656 arc_evict_zthr = zthr_create_timer("arc_evict",
7657 arc_evict_cb_check, arc_evict_cb, NULL, SEC2NSEC(1), defclsyspri);
7658 arc_reap_zthr = zthr_create_timer("arc_reap",
7659 arc_reap_cb_check, arc_reap_cb, NULL, SEC2NSEC(1), minclsyspri);
7664 * Calculate maximum amount of dirty data per pool.
7666 * If it has been set by a module parameter, take that.
7667 * Otherwise, use a percentage of physical memory defined by
7668 * zfs_dirty_data_max_percent (default 10%) with a cap at
7669 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
7672 if (zfs_dirty_data_max_max == 0)
7673 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
7674 allmem * zfs_dirty_data_max_max_percent / 100);
7676 if (zfs_dirty_data_max_max == 0)
7677 zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024,
7678 allmem * zfs_dirty_data_max_max_percent / 100);
7681 if (zfs_dirty_data_max == 0) {
7682 zfs_dirty_data_max = allmem *
7683 zfs_dirty_data_max_percent / 100;
7684 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7685 zfs_dirty_data_max_max);
7688 if (zfs_wrlog_data_max == 0) {
7691 * dp_wrlog_total is reduced for each txg at the end of
7692 * spa_sync(). However, dp_dirty_total is reduced every time
7693 * a block is written out. Thus under normal operation,
7694 * dp_wrlog_total could grow 2 times as big as
7695 * zfs_dirty_data_max.
7697 zfs_wrlog_data_max = zfs_dirty_data_max * 2;
7708 #endif /* _KERNEL */
7710 /* Use B_TRUE to ensure *all* buffers are evicted */
7711 arc_flush(NULL, B_TRUE);
7713 if (arc_ksp != NULL) {
7714 kstat_delete(arc_ksp);
7718 taskq_wait(arc_prune_taskq);
7719 taskq_destroy(arc_prune_taskq);
7721 mutex_enter(&arc_prune_mtx);
7722 while ((p = list_remove_head(&arc_prune_list)) != NULL) {
7723 zfs_refcount_remove(&p->p_refcnt, &arc_prune_list);
7724 zfs_refcount_destroy(&p->p_refcnt);
7725 kmem_free(p, sizeof (*p));
7727 mutex_exit(&arc_prune_mtx);
7729 list_destroy(&arc_prune_list);
7730 mutex_destroy(&arc_prune_mtx);
7732 (void) zthr_cancel(arc_evict_zthr);
7733 (void) zthr_cancel(arc_reap_zthr);
7734 arc_state_free_markers(arc_state_evict_markers,
7735 arc_state_evict_marker_count);
7737 mutex_destroy(&arc_evict_lock);
7738 list_destroy(&arc_evict_waiters);
7741 * Free any buffers that were tagged for destruction. This needs
7742 * to occur before arc_state_fini() runs and destroys the aggsum
7743 * values which are updated when freeing scatter ABDs.
7745 l2arc_do_free_on_write();
7748 * buf_fini() must proceed arc_state_fini() because buf_fin() may
7749 * trigger the release of kmem magazines, which can callback to
7750 * arc_space_return() which accesses aggsums freed in act_state_fini().
7755 arc_unregister_hotplug();
7758 * We destroy the zthrs after all the ARC state has been
7759 * torn down to avoid the case of them receiving any
7760 * wakeup() signals after they are destroyed.
7762 zthr_destroy(arc_evict_zthr);
7763 zthr_destroy(arc_reap_zthr);
7765 ASSERT0(arc_loaned_bytes);
7771 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7772 * It uses dedicated storage devices to hold cached data, which are populated
7773 * using large infrequent writes. The main role of this cache is to boost
7774 * the performance of random read workloads. The intended L2ARC devices
7775 * include short-stroked disks, solid state disks, and other media with
7776 * substantially faster read latency than disk.
7778 * +-----------------------+
7780 * +-----------------------+
7783 * l2arc_feed_thread() arc_read()
7787 * +---------------+ |
7789 * +---------------+ |
7794 * +-------+ +-------+
7796 * | cache | | cache |
7797 * +-------+ +-------+
7798 * +=========+ .-----.
7799 * : L2ARC : |-_____-|
7800 * : devices : | Disks |
7801 * +=========+ `-_____-'
7803 * Read requests are satisfied from the following sources, in order:
7806 * 2) vdev cache of L2ARC devices
7808 * 4) vdev cache of disks
7811 * Some L2ARC device types exhibit extremely slow write performance.
7812 * To accommodate for this there are some significant differences between
7813 * the L2ARC and traditional cache design:
7815 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7816 * the ARC behave as usual, freeing buffers and placing headers on ghost
7817 * lists. The ARC does not send buffers to the L2ARC during eviction as
7818 * this would add inflated write latencies for all ARC memory pressure.
7820 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7821 * It does this by periodically scanning buffers from the eviction-end of
7822 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7823 * not already there. It scans until a headroom of buffers is satisfied,
7824 * which itself is a buffer for ARC eviction. If a compressible buffer is
7825 * found during scanning and selected for writing to an L2ARC device, we
7826 * temporarily boost scanning headroom during the next scan cycle to make
7827 * sure we adapt to compression effects (which might significantly reduce
7828 * the data volume we write to L2ARC). The thread that does this is
7829 * l2arc_feed_thread(), illustrated below; example sizes are included to
7830 * provide a better sense of ratio than this diagram:
7833 * +---------------------+----------+
7834 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7835 * +---------------------+----------+ | o L2ARC eligible
7836 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7837 * +---------------------+----------+ |
7838 * 15.9 Gbytes ^ 32 Mbytes |
7840 * l2arc_feed_thread()
7842 * l2arc write hand <--[oooo]--'
7846 * +==============================+
7847 * L2ARC dev |####|#|###|###| |####| ... |
7848 * +==============================+
7851 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7852 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7853 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7854 * safe to say that this is an uncommon case, since buffers at the end of
7855 * the ARC lists have moved there due to inactivity.
7857 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7858 * then the L2ARC simply misses copying some buffers. This serves as a
7859 * pressure valve to prevent heavy read workloads from both stalling the ARC
7860 * with waits and clogging the L2ARC with writes. This also helps prevent
7861 * the potential for the L2ARC to churn if it attempts to cache content too
7862 * quickly, such as during backups of the entire pool.
7864 * 5. After system boot and before the ARC has filled main memory, there are
7865 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7866 * lists can remain mostly static. Instead of searching from tail of these
7867 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7868 * for eligible buffers, greatly increasing its chance of finding them.
7870 * The L2ARC device write speed is also boosted during this time so that
7871 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7872 * there are no L2ARC reads, and no fear of degrading read performance
7873 * through increased writes.
7875 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7876 * the vdev queue can aggregate them into larger and fewer writes. Each
7877 * device is written to in a rotor fashion, sweeping writes through
7878 * available space then repeating.
7880 * 7. The L2ARC does not store dirty content. It never needs to flush
7881 * write buffers back to disk based storage.
7883 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7884 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7886 * The performance of the L2ARC can be tweaked by a number of tunables, which
7887 * may be necessary for different workloads:
7889 * l2arc_write_max max write bytes per interval
7890 * l2arc_write_boost extra write bytes during device warmup
7891 * l2arc_noprefetch skip caching prefetched buffers
7892 * l2arc_headroom number of max device writes to precache
7893 * l2arc_headroom_boost when we find compressed buffers during ARC
7894 * scanning, we multiply headroom by this
7895 * percentage factor for the next scan cycle,
7896 * since more compressed buffers are likely to
7898 * l2arc_feed_secs seconds between L2ARC writing
7900 * Tunables may be removed or added as future performance improvements are
7901 * integrated, and also may become zpool properties.
7903 * There are three key functions that control how the L2ARC warms up:
7905 * l2arc_write_eligible() check if a buffer is eligible to cache
7906 * l2arc_write_size() calculate how much to write
7907 * l2arc_write_interval() calculate sleep delay between writes
7909 * These three functions determine what to write, how much, and how quickly
7912 * L2ARC persistence:
7914 * When writing buffers to L2ARC, we periodically add some metadata to
7915 * make sure we can pick them up after reboot, thus dramatically reducing
7916 * the impact that any downtime has on the performance of storage systems
7917 * with large caches.
7919 * The implementation works fairly simply by integrating the following two
7922 * *) When writing to the L2ARC, we occasionally write a "l2arc log block",
7923 * which is an additional piece of metadata which describes what's been
7924 * written. This allows us to rebuild the arc_buf_hdr_t structures of the
7925 * main ARC buffers. There are 2 linked-lists of log blocks headed by
7926 * dh_start_lbps[2]. We alternate which chain we append to, so they are
7927 * time-wise and offset-wise interleaved, but that is an optimization rather
7928 * than for correctness. The log block also includes a pointer to the
7929 * previous block in its chain.
7931 * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device
7932 * for our header bookkeeping purposes. This contains a device header,
7933 * which contains our top-level reference structures. We update it each
7934 * time we write a new log block, so that we're able to locate it in the
7935 * L2ARC device. If this write results in an inconsistent device header
7936 * (e.g. due to power failure), we detect this by verifying the header's
7937 * checksum and simply fail to reconstruct the L2ARC after reboot.
7939 * Implementation diagram:
7941 * +=== L2ARC device (not to scale) ======================================+
7942 * | ___two newest log block pointers__.__________ |
7943 * | / \dh_start_lbps[1] |
7944 * | / \ \dh_start_lbps[0]|
7946 * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---|
7947 * || hdr| ^ /^ /^ / / |
7948 * |+------+ ...--\-------/ \-----/--\------/ / |
7949 * | \--------------/ \--------------/ |
7950 * +======================================================================+
7952 * As can be seen on the diagram, rather than using a simple linked list,
7953 * we use a pair of linked lists with alternating elements. This is a
7954 * performance enhancement due to the fact that we only find out the
7955 * address of the next log block access once the current block has been
7956 * completely read in. Obviously, this hurts performance, because we'd be
7957 * keeping the device's I/O queue at only a 1 operation deep, thus
7958 * incurring a large amount of I/O round-trip latency. Having two lists
7959 * allows us to fetch two log blocks ahead of where we are currently
7960 * rebuilding L2ARC buffers.
7962 * On-device data structures:
7964 * L2ARC device header: l2arc_dev_hdr_phys_t
7965 * L2ARC log block: l2arc_log_blk_phys_t
7967 * L2ARC reconstruction:
7969 * When writing data, we simply write in the standard rotary fashion,
7970 * evicting buffers as we go and simply writing new data over them (writing
7971 * a new log block every now and then). This obviously means that once we
7972 * loop around the end of the device, we will start cutting into an already
7973 * committed log block (and its referenced data buffers), like so:
7975 * current write head__ __old tail
7978 * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |-->
7979 * ^ ^^^^^^^^^___________________________________
7981 * <<nextwrite>> may overwrite this blk and/or its bufs --'
7983 * When importing the pool, we detect this situation and use it to stop
7984 * our scanning process (see l2arc_rebuild).
7986 * There is one significant caveat to consider when rebuilding ARC contents
7987 * from an L2ARC device: what about invalidated buffers? Given the above
7988 * construction, we cannot update blocks which we've already written to amend
7989 * them to remove buffers which were invalidated. Thus, during reconstruction,
7990 * we might be populating the cache with buffers for data that's not on the
7991 * main pool anymore, or may have been overwritten!
7993 * As it turns out, this isn't a problem. Every arc_read request includes
7994 * both the DVA and, crucially, the birth TXG of the BP the caller is
7995 * looking for. So even if the cache were populated by completely rotten
7996 * blocks for data that had been long deleted and/or overwritten, we'll
7997 * never actually return bad data from the cache, since the DVA with the
7998 * birth TXG uniquely identify a block in space and time - once created,
7999 * a block is immutable on disk. The worst thing we have done is wasted
8000 * some time and memory at l2arc rebuild to reconstruct outdated ARC
8001 * entries that will get dropped from the l2arc as it is being updated
8004 * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write
8005 * hand are not restored. This is done by saving the offset (in bytes)
8006 * l2arc_evict() has evicted to in the L2ARC device header and taking it
8007 * into account when restoring buffers.
8011 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
8014 * A buffer is *not* eligible for the L2ARC if it:
8015 * 1. belongs to a different spa.
8016 * 2. is already cached on the L2ARC.
8017 * 3. has an I/O in progress (it may be an incomplete read).
8018 * 4. is flagged not eligible (zfs property).
8020 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
8021 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
8028 l2arc_write_size(l2arc_dev_t *dev)
8033 * Make sure our globals have meaningful values in case the user
8036 size = l2arc_write_max;
8038 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
8039 "be greater than zero, resetting it to the default (%d)",
8041 size = l2arc_write_max = L2ARC_WRITE_SIZE;
8044 if (arc_warm == B_FALSE)
8045 size += l2arc_write_boost;
8047 /* We need to add in the worst case scenario of log block overhead. */
8048 size += l2arc_log_blk_overhead(size, dev);
8049 if (dev->l2ad_vdev->vdev_has_trim && l2arc_trim_ahead > 0) {
8051 * Trim ahead of the write size 64MB or (l2arc_trim_ahead/100)
8052 * times the writesize, whichever is greater.
8054 size += MAX(64 * 1024 * 1024,
8055 (size * l2arc_trim_ahead) / 100);
8059 * Make sure the write size does not exceed the size of the cache
8060 * device. This is important in l2arc_evict(), otherwise infinite
8061 * iteration can occur.
8063 if (size > dev->l2ad_end - dev->l2ad_start) {
8064 cmn_err(CE_NOTE, "l2arc_write_max or l2arc_write_boost "
8065 "plus the overhead of log blocks (persistent L2ARC, "
8066 "%llu bytes) exceeds the size of the cache device "
8067 "(guid %llu), resetting them to the default (%d)",
8068 (u_longlong_t)l2arc_log_blk_overhead(size, dev),
8069 (u_longlong_t)dev->l2ad_vdev->vdev_guid, L2ARC_WRITE_SIZE);
8071 size = l2arc_write_max = l2arc_write_boost = L2ARC_WRITE_SIZE;
8073 if (l2arc_trim_ahead > 1) {
8074 cmn_err(CE_NOTE, "l2arc_trim_ahead set to 1");
8075 l2arc_trim_ahead = 1;
8078 if (arc_warm == B_FALSE)
8079 size += l2arc_write_boost;
8081 size += l2arc_log_blk_overhead(size, dev);
8082 if (dev->l2ad_vdev->vdev_has_trim && l2arc_trim_ahead > 0) {
8083 size += MAX(64 * 1024 * 1024,
8084 (size * l2arc_trim_ahead) / 100);
8093 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
8095 clock_t interval, next, now;
8098 * If the ARC lists are busy, increase our write rate; if the
8099 * lists are stale, idle back. This is achieved by checking
8100 * how much we previously wrote - if it was more than half of
8101 * what we wanted, schedule the next write much sooner.
8103 if (l2arc_feed_again && wrote > (wanted / 2))
8104 interval = (hz * l2arc_feed_min_ms) / 1000;
8106 interval = hz * l2arc_feed_secs;
8108 now = ddi_get_lbolt();
8109 next = MAX(now, MIN(now + interval, began + interval));
8115 * Cycle through L2ARC devices. This is how L2ARC load balances.
8116 * If a device is returned, this also returns holding the spa config lock.
8118 static l2arc_dev_t *
8119 l2arc_dev_get_next(void)
8121 l2arc_dev_t *first, *next = NULL;
8124 * Lock out the removal of spas (spa_namespace_lock), then removal
8125 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
8126 * both locks will be dropped and a spa config lock held instead.
8128 mutex_enter(&spa_namespace_lock);
8129 mutex_enter(&l2arc_dev_mtx);
8131 /* if there are no vdevs, there is nothing to do */
8132 if (l2arc_ndev == 0)
8136 next = l2arc_dev_last;
8138 /* loop around the list looking for a non-faulted vdev */
8140 next = list_head(l2arc_dev_list);
8142 next = list_next(l2arc_dev_list, next);
8144 next = list_head(l2arc_dev_list);
8147 /* if we have come back to the start, bail out */
8150 else if (next == first)
8153 ASSERT3P(next, !=, NULL);
8154 } while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8155 next->l2ad_trim_all);
8157 /* if we were unable to find any usable vdevs, return NULL */
8158 if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8159 next->l2ad_trim_all)
8162 l2arc_dev_last = next;
8165 mutex_exit(&l2arc_dev_mtx);
8168 * Grab the config lock to prevent the 'next' device from being
8169 * removed while we are writing to it.
8172 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
8173 mutex_exit(&spa_namespace_lock);
8179 * Free buffers that were tagged for destruction.
8182 l2arc_do_free_on_write(void)
8184 l2arc_data_free_t *df;
8186 mutex_enter(&l2arc_free_on_write_mtx);
8187 while ((df = list_remove_head(l2arc_free_on_write)) != NULL) {
8188 ASSERT3P(df->l2df_abd, !=, NULL);
8189 abd_free(df->l2df_abd);
8190 kmem_free(df, sizeof (l2arc_data_free_t));
8192 mutex_exit(&l2arc_free_on_write_mtx);
8196 * A write to a cache device has completed. Update all headers to allow
8197 * reads from these buffers to begin.
8200 l2arc_write_done(zio_t *zio)
8202 l2arc_write_callback_t *cb;
8203 l2arc_lb_abd_buf_t *abd_buf;
8204 l2arc_lb_ptr_buf_t *lb_ptr_buf;
8206 l2arc_dev_hdr_phys_t *l2dhdr;
8208 arc_buf_hdr_t *head, *hdr, *hdr_prev;
8209 kmutex_t *hash_lock;
8210 int64_t bytes_dropped = 0;
8212 cb = zio->io_private;
8213 ASSERT3P(cb, !=, NULL);
8214 dev = cb->l2wcb_dev;
8215 l2dhdr = dev->l2ad_dev_hdr;
8216 ASSERT3P(dev, !=, NULL);
8217 head = cb->l2wcb_head;
8218 ASSERT3P(head, !=, NULL);
8219 buflist = &dev->l2ad_buflist;
8220 ASSERT3P(buflist, !=, NULL);
8221 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
8222 l2arc_write_callback_t *, cb);
8225 * All writes completed, or an error was hit.
8228 mutex_enter(&dev->l2ad_mtx);
8229 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
8230 hdr_prev = list_prev(buflist, hdr);
8232 hash_lock = HDR_LOCK(hdr);
8235 * We cannot use mutex_enter or else we can deadlock
8236 * with l2arc_write_buffers (due to swapping the order
8237 * the hash lock and l2ad_mtx are taken).
8239 if (!mutex_tryenter(hash_lock)) {
8241 * Missed the hash lock. We must retry so we
8242 * don't leave the ARC_FLAG_L2_WRITING bit set.
8244 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
8247 * We don't want to rescan the headers we've
8248 * already marked as having been written out, so
8249 * we reinsert the head node so we can pick up
8250 * where we left off.
8252 list_remove(buflist, head);
8253 list_insert_after(buflist, hdr, head);
8255 mutex_exit(&dev->l2ad_mtx);
8258 * We wait for the hash lock to become available
8259 * to try and prevent busy waiting, and increase
8260 * the chance we'll be able to acquire the lock
8261 * the next time around.
8263 mutex_enter(hash_lock);
8264 mutex_exit(hash_lock);
8269 * We could not have been moved into the arc_l2c_only
8270 * state while in-flight due to our ARC_FLAG_L2_WRITING
8271 * bit being set. Let's just ensure that's being enforced.
8273 ASSERT(HDR_HAS_L1HDR(hdr));
8276 * Skipped - drop L2ARC entry and mark the header as no
8277 * longer L2 eligibile.
8279 if (zio->io_error != 0) {
8281 * Error - drop L2ARC entry.
8283 list_remove(buflist, hdr);
8284 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
8286 uint64_t psize = HDR_GET_PSIZE(hdr);
8287 l2arc_hdr_arcstats_decrement(hdr);
8290 vdev_psize_to_asize(dev->l2ad_vdev, psize);
8291 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
8292 arc_hdr_size(hdr), hdr);
8296 * Allow ARC to begin reads and ghost list evictions to
8299 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
8301 mutex_exit(hash_lock);
8305 * Free the allocated abd buffers for writing the log blocks.
8306 * If the zio failed reclaim the allocated space and remove the
8307 * pointers to these log blocks from the log block pointer list
8308 * of the L2ARC device.
8310 while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) {
8311 abd_free(abd_buf->abd);
8312 zio_buf_free(abd_buf, sizeof (*abd_buf));
8313 if (zio->io_error != 0) {
8314 lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list);
8316 * L2BLK_GET_PSIZE returns aligned size for log
8320 L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop);
8321 bytes_dropped += asize;
8322 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8323 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8324 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8326 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8327 kmem_free(lb_ptr_buf->lb_ptr,
8328 sizeof (l2arc_log_blkptr_t));
8329 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8332 list_destroy(&cb->l2wcb_abd_list);
8334 if (zio->io_error != 0) {
8335 ARCSTAT_BUMP(arcstat_l2_writes_error);
8338 * Restore the lbps array in the header to its previous state.
8339 * If the list of log block pointers is empty, zero out the
8340 * log block pointers in the device header.
8342 lb_ptr_buf = list_head(&dev->l2ad_lbptr_list);
8343 for (int i = 0; i < 2; i++) {
8344 if (lb_ptr_buf == NULL) {
8346 * If the list is empty zero out the device
8347 * header. Otherwise zero out the second log
8348 * block pointer in the header.
8352 dev->l2ad_dev_hdr_asize);
8354 memset(&l2dhdr->dh_start_lbps[i], 0,
8355 sizeof (l2arc_log_blkptr_t));
8359 memcpy(&l2dhdr->dh_start_lbps[i], lb_ptr_buf->lb_ptr,
8360 sizeof (l2arc_log_blkptr_t));
8361 lb_ptr_buf = list_next(&dev->l2ad_lbptr_list,
8366 ARCSTAT_BUMP(arcstat_l2_writes_done);
8367 list_remove(buflist, head);
8368 ASSERT(!HDR_HAS_L1HDR(head));
8369 kmem_cache_free(hdr_l2only_cache, head);
8370 mutex_exit(&dev->l2ad_mtx);
8372 ASSERT(dev->l2ad_vdev != NULL);
8373 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
8375 l2arc_do_free_on_write();
8377 kmem_free(cb, sizeof (l2arc_write_callback_t));
8381 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
8384 spa_t *spa = zio->io_spa;
8385 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
8386 blkptr_t *bp = zio->io_bp;
8387 uint8_t salt[ZIO_DATA_SALT_LEN];
8388 uint8_t iv[ZIO_DATA_IV_LEN];
8389 uint8_t mac[ZIO_DATA_MAC_LEN];
8390 boolean_t no_crypt = B_FALSE;
8393 * ZIL data is never be written to the L2ARC, so we don't need
8394 * special handling for its unique MAC storage.
8396 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
8397 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
8398 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8401 * If the data was encrypted, decrypt it now. Note that
8402 * we must check the bp here and not the hdr, since the
8403 * hdr does not have its encryption parameters updated
8404 * until arc_read_done().
8406 if (BP_IS_ENCRYPTED(bp)) {
8407 abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8408 ARC_HDR_USE_RESERVE);
8410 zio_crypt_decode_params_bp(bp, salt, iv);
8411 zio_crypt_decode_mac_bp(bp, mac);
8413 ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
8414 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
8415 salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
8416 hdr->b_l1hdr.b_pabd, &no_crypt);
8418 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8423 * If we actually performed decryption, replace b_pabd
8424 * with the decrypted data. Otherwise we can just throw
8425 * our decryption buffer away.
8428 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8429 arc_hdr_size(hdr), hdr);
8430 hdr->b_l1hdr.b_pabd = eabd;
8433 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8438 * If the L2ARC block was compressed, but ARC compression
8439 * is disabled we decompress the data into a new buffer and
8440 * replace the existing data.
8442 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8443 !HDR_COMPRESSION_ENABLED(hdr)) {
8444 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8445 ARC_HDR_USE_RESERVE);
8446 void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
8448 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
8449 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
8450 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
8452 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8453 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
8457 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8458 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8459 arc_hdr_size(hdr), hdr);
8460 hdr->b_l1hdr.b_pabd = cabd;
8462 zio->io_size = HDR_GET_LSIZE(hdr);
8473 * A read to a cache device completed. Validate buffer contents before
8474 * handing over to the regular ARC routines.
8477 l2arc_read_done(zio_t *zio)
8480 l2arc_read_callback_t *cb = zio->io_private;
8482 kmutex_t *hash_lock;
8483 boolean_t valid_cksum;
8484 boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
8485 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
8487 ASSERT3P(zio->io_vd, !=, NULL);
8488 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
8490 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
8492 ASSERT3P(cb, !=, NULL);
8493 hdr = cb->l2rcb_hdr;
8494 ASSERT3P(hdr, !=, NULL);
8496 hash_lock = HDR_LOCK(hdr);
8497 mutex_enter(hash_lock);
8498 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
8501 * If the data was read into a temporary buffer,
8502 * move it and free the buffer.
8504 if (cb->l2rcb_abd != NULL) {
8505 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
8506 if (zio->io_error == 0) {
8508 abd_copy(hdr->b_crypt_hdr.b_rabd,
8509 cb->l2rcb_abd, arc_hdr_size(hdr));
8511 abd_copy(hdr->b_l1hdr.b_pabd,
8512 cb->l2rcb_abd, arc_hdr_size(hdr));
8517 * The following must be done regardless of whether
8518 * there was an error:
8519 * - free the temporary buffer
8520 * - point zio to the real ARC buffer
8521 * - set zio size accordingly
8522 * These are required because zio is either re-used for
8523 * an I/O of the block in the case of the error
8524 * or the zio is passed to arc_read_done() and it
8527 abd_free(cb->l2rcb_abd);
8528 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
8531 ASSERT(HDR_HAS_RABD(hdr));
8532 zio->io_abd = zio->io_orig_abd =
8533 hdr->b_crypt_hdr.b_rabd;
8535 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8536 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8540 ASSERT3P(zio->io_abd, !=, NULL);
8543 * Check this survived the L2ARC journey.
8545 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8546 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8547 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
8548 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
8549 zio->io_prop.zp_complevel = hdr->b_complevel;
8551 valid_cksum = arc_cksum_is_equal(hdr, zio);
8554 * b_rabd will always match the data as it exists on disk if it is
8555 * being used. Therefore if we are reading into b_rabd we do not
8556 * attempt to untransform the data.
8558 if (valid_cksum && !using_rdata)
8559 tfm_error = l2arc_untransform(zio, cb);
8561 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8562 !HDR_L2_EVICTED(hdr)) {
8563 mutex_exit(hash_lock);
8564 zio->io_private = hdr;
8568 * Buffer didn't survive caching. Increment stats and
8569 * reissue to the original storage device.
8571 if (zio->io_error != 0) {
8572 ARCSTAT_BUMP(arcstat_l2_io_error);
8574 zio->io_error = SET_ERROR(EIO);
8576 if (!valid_cksum || tfm_error != 0)
8577 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8580 * If there's no waiter, issue an async i/o to the primary
8581 * storage now. If there *is* a waiter, the caller must
8582 * issue the i/o in a context where it's OK to block.
8584 if (zio->io_waiter == NULL) {
8585 zio_t *pio = zio_unique_parent(zio);
8586 void *abd = (using_rdata) ?
8587 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8589 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8591 zio = zio_read(pio, zio->io_spa, zio->io_bp,
8592 abd, zio->io_size, arc_read_done,
8593 hdr, zio->io_priority, cb->l2rcb_flags,
8597 * Original ZIO will be freed, so we need to update
8598 * ARC header with the new ZIO pointer to be used
8599 * by zio_change_priority() in arc_read().
8601 for (struct arc_callback *acb = hdr->b_l1hdr.b_acb;
8602 acb != NULL; acb = acb->acb_next)
8603 acb->acb_zio_head = zio;
8605 mutex_exit(hash_lock);
8608 mutex_exit(hash_lock);
8612 kmem_free(cb, sizeof (l2arc_read_callback_t));
8616 * This is the list priority from which the L2ARC will search for pages to
8617 * cache. This is used within loops (0..3) to cycle through lists in the
8618 * desired order. This order can have a significant effect on cache
8621 * Currently the metadata lists are hit first, MFU then MRU, followed by
8622 * the data lists. This function returns a locked list, and also returns
8625 static multilist_sublist_t *
8626 l2arc_sublist_lock(int list_num)
8628 multilist_t *ml = NULL;
8631 ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
8635 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
8638 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
8641 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
8644 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
8651 * Return a randomly-selected sublist. This is acceptable
8652 * because the caller feeds only a little bit of data for each
8653 * call (8MB). Subsequent calls will result in different
8654 * sublists being selected.
8656 idx = multilist_get_random_index(ml);
8657 return (multilist_sublist_lock(ml, idx));
8661 * Calculates the maximum overhead of L2ARC metadata log blocks for a given
8662 * L2ARC write size. l2arc_evict and l2arc_write_size need to include this
8663 * overhead in processing to make sure there is enough headroom available
8664 * when writing buffers.
8666 static inline uint64_t
8667 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev)
8669 if (dev->l2ad_log_entries == 0) {
8672 uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT;
8674 uint64_t log_blocks = (log_entries +
8675 dev->l2ad_log_entries - 1) /
8676 dev->l2ad_log_entries;
8678 return (vdev_psize_to_asize(dev->l2ad_vdev,
8679 sizeof (l2arc_log_blk_phys_t)) * log_blocks);
8684 * Evict buffers from the device write hand to the distance specified in
8685 * bytes. This distance may span populated buffers, it may span nothing.
8686 * This is clearing a region on the L2ARC device ready for writing.
8687 * If the 'all' boolean is set, every buffer is evicted.
8690 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
8693 arc_buf_hdr_t *hdr, *hdr_prev;
8694 kmutex_t *hash_lock;
8696 l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev;
8697 vdev_t *vd = dev->l2ad_vdev;
8700 buflist = &dev->l2ad_buflist;
8704 if (dev->l2ad_hand + distance > dev->l2ad_end) {
8706 * When there is no space to accommodate upcoming writes,
8707 * evict to the end. Then bump the write and evict hands
8708 * to the start and iterate. This iteration does not
8709 * happen indefinitely as we make sure in
8710 * l2arc_write_size() that when the write hand is reset,
8711 * the write size does not exceed the end of the device.
8714 taddr = dev->l2ad_end;
8716 taddr = dev->l2ad_hand + distance;
8718 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
8719 uint64_t, taddr, boolean_t, all);
8723 * This check has to be placed after deciding whether to
8726 if (dev->l2ad_first) {
8728 * This is the first sweep through the device. There is
8729 * nothing to evict. We have already trimmmed the
8735 * Trim the space to be evicted.
8737 if (vd->vdev_has_trim && dev->l2ad_evict < taddr &&
8738 l2arc_trim_ahead > 0) {
8740 * We have to drop the spa_config lock because
8741 * vdev_trim_range() will acquire it.
8742 * l2ad_evict already accounts for the label
8743 * size. To prevent vdev_trim_ranges() from
8744 * adding it again, we subtract it from
8747 spa_config_exit(dev->l2ad_spa, SCL_L2ARC, dev);
8748 vdev_trim_simple(vd,
8749 dev->l2ad_evict - VDEV_LABEL_START_SIZE,
8750 taddr - dev->l2ad_evict);
8751 spa_config_enter(dev->l2ad_spa, SCL_L2ARC, dev,
8756 * When rebuilding L2ARC we retrieve the evict hand
8757 * from the header of the device. Of note, l2arc_evict()
8758 * does not actually delete buffers from the cache
8759 * device, but trimming may do so depending on the
8760 * hardware implementation. Thus keeping track of the
8761 * evict hand is useful.
8763 dev->l2ad_evict = MAX(dev->l2ad_evict, taddr);
8768 mutex_enter(&dev->l2ad_mtx);
8770 * We have to account for evicted log blocks. Run vdev_space_update()
8771 * on log blocks whose offset (in bytes) is before the evicted offset
8772 * (in bytes) by searching in the list of pointers to log blocks
8773 * present in the L2ARC device.
8775 for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf;
8776 lb_ptr_buf = lb_ptr_buf_prev) {
8778 lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf);
8780 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
8781 uint64_t asize = L2BLK_GET_PSIZE(
8782 (lb_ptr_buf->lb_ptr)->lbp_prop);
8785 * We don't worry about log blocks left behind (ie
8786 * lbp_payload_start < l2ad_hand) because l2arc_write_buffers()
8787 * will never write more than l2arc_evict() evicts.
8789 if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) {
8792 vdev_space_update(vd, -asize, 0, 0);
8793 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8794 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8795 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8797 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8798 list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf);
8799 kmem_free(lb_ptr_buf->lb_ptr,
8800 sizeof (l2arc_log_blkptr_t));
8801 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8805 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
8806 hdr_prev = list_prev(buflist, hdr);
8808 ASSERT(!HDR_EMPTY(hdr));
8809 hash_lock = HDR_LOCK(hdr);
8812 * We cannot use mutex_enter or else we can deadlock
8813 * with l2arc_write_buffers (due to swapping the order
8814 * the hash lock and l2ad_mtx are taken).
8816 if (!mutex_tryenter(hash_lock)) {
8818 * Missed the hash lock. Retry.
8820 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
8821 mutex_exit(&dev->l2ad_mtx);
8822 mutex_enter(hash_lock);
8823 mutex_exit(hash_lock);
8828 * A header can't be on this list if it doesn't have L2 header.
8830 ASSERT(HDR_HAS_L2HDR(hdr));
8832 /* Ensure this header has finished being written. */
8833 ASSERT(!HDR_L2_WRITING(hdr));
8834 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
8836 if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict ||
8837 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
8839 * We've evicted to the target address,
8840 * or the end of the device.
8842 mutex_exit(hash_lock);
8846 if (!HDR_HAS_L1HDR(hdr)) {
8847 ASSERT(!HDR_L2_READING(hdr));
8849 * This doesn't exist in the ARC. Destroy.
8850 * arc_hdr_destroy() will call list_remove()
8851 * and decrement arcstat_l2_lsize.
8853 arc_change_state(arc_anon, hdr);
8854 arc_hdr_destroy(hdr);
8856 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
8857 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
8859 * Invalidate issued or about to be issued
8860 * reads, since we may be about to write
8861 * over this location.
8863 if (HDR_L2_READING(hdr)) {
8864 ARCSTAT_BUMP(arcstat_l2_evict_reading);
8865 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
8868 arc_hdr_l2hdr_destroy(hdr);
8870 mutex_exit(hash_lock);
8872 mutex_exit(&dev->l2ad_mtx);
8876 * We need to check if we evict all buffers, otherwise we may iterate
8879 if (!all && rerun) {
8881 * Bump device hand to the device start if it is approaching the
8882 * end. l2arc_evict() has already evicted ahead for this case.
8884 dev->l2ad_hand = dev->l2ad_start;
8885 dev->l2ad_evict = dev->l2ad_start;
8886 dev->l2ad_first = B_FALSE;
8892 * In case of cache device removal (all) the following
8893 * assertions may be violated without functional consequences
8894 * as the device is about to be removed.
8896 ASSERT3U(dev->l2ad_hand + distance, <, dev->l2ad_end);
8897 if (!dev->l2ad_first)
8898 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
8903 * Handle any abd transforms that might be required for writing to the L2ARC.
8904 * If successful, this function will always return an abd with the data
8905 * transformed as it is on disk in a new abd of asize bytes.
8908 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
8913 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
8914 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
8915 uint64_t psize = HDR_GET_PSIZE(hdr);
8916 uint64_t size = arc_hdr_size(hdr);
8917 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
8918 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
8919 dsl_crypto_key_t *dck = NULL;
8920 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
8921 boolean_t no_crypt = B_FALSE;
8923 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8924 !HDR_COMPRESSION_ENABLED(hdr)) ||
8925 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
8926 ASSERT3U(psize, <=, asize);
8929 * If this data simply needs its own buffer, we simply allocate it
8930 * and copy the data. This may be done to eliminate a dependency on a
8931 * shared buffer or to reallocate the buffer to match asize.
8933 if (HDR_HAS_RABD(hdr) && asize != psize) {
8934 ASSERT3U(asize, >=, psize);
8935 to_write = abd_alloc_for_io(asize, ismd);
8936 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
8938 abd_zero_off(to_write, psize, asize - psize);
8942 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
8943 !HDR_ENCRYPTED(hdr)) {
8944 ASSERT3U(size, ==, psize);
8945 to_write = abd_alloc_for_io(asize, ismd);
8946 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
8948 abd_zero_off(to_write, size, asize - size);
8952 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
8954 * In some cases, we can wind up with size > asize, so
8955 * we need to opt for the larger allocation option here.
8957 * (We also need abd_return_buf_copy in all cases because
8958 * it's an ASSERT() to modify the buffer before returning it
8959 * with arc_return_buf(), and all the compressors
8960 * write things before deciding to fail compression in nearly
8963 uint64_t bufsize = MAX(size, asize);
8964 cabd = abd_alloc_for_io(bufsize, ismd);
8965 tmp = abd_borrow_buf(cabd, bufsize);
8967 psize = zio_compress_data(compress, to_write, &tmp, size,
8970 if (psize >= asize) {
8971 psize = HDR_GET_PSIZE(hdr);
8972 abd_return_buf_copy(cabd, tmp, bufsize);
8973 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
8975 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
8977 abd_zero_off(to_write, psize, asize - psize);
8980 ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
8982 memset((char *)tmp + psize, 0, bufsize - psize);
8983 psize = HDR_GET_PSIZE(hdr);
8984 abd_return_buf_copy(cabd, tmp, bufsize);
8989 if (HDR_ENCRYPTED(hdr)) {
8990 eabd = abd_alloc_for_io(asize, ismd);
8993 * If the dataset was disowned before the buffer
8994 * made it to this point, the key to re-encrypt
8995 * it won't be available. In this case we simply
8996 * won't write the buffer to the L2ARC.
8998 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
9003 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
9004 hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
9005 hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
9011 abd_copy(eabd, to_write, psize);
9014 abd_zero_off(eabd, psize, asize - psize);
9016 /* assert that the MAC we got here matches the one we saved */
9017 ASSERT0(memcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
9018 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9020 if (to_write == cabd)
9027 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
9028 *abd_out = to_write;
9033 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9044 l2arc_blk_fetch_done(zio_t *zio)
9046 l2arc_read_callback_t *cb;
9048 cb = zio->io_private;
9049 if (cb->l2rcb_abd != NULL)
9050 abd_free(cb->l2rcb_abd);
9051 kmem_free(cb, sizeof (l2arc_read_callback_t));
9055 * Find and write ARC buffers to the L2ARC device.
9057 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
9058 * for reading until they have completed writing.
9059 * The headroom_boost is an in-out parameter used to maintain headroom boost
9060 * state between calls to this function.
9062 * Returns the number of bytes actually written (which may be smaller than
9063 * the delta by which the device hand has changed due to alignment and the
9064 * writing of log blocks).
9067 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
9069 arc_buf_hdr_t *hdr, *hdr_prev, *head;
9070 uint64_t write_asize, write_psize, write_lsize, headroom;
9072 l2arc_write_callback_t *cb = NULL;
9074 uint64_t guid = spa_load_guid(spa);
9075 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9077 ASSERT3P(dev->l2ad_vdev, !=, NULL);
9080 write_lsize = write_asize = write_psize = 0;
9082 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
9083 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
9086 * Copy buffers for L2ARC writing.
9088 for (int pass = 0; pass < L2ARC_FEED_TYPES; pass++) {
9090 * If pass == 1 or 3, we cache MRU metadata and data
9093 if (l2arc_mfuonly) {
9094 if (pass == 1 || pass == 3)
9098 multilist_sublist_t *mls = l2arc_sublist_lock(pass);
9099 uint64_t passed_sz = 0;
9101 VERIFY3P(mls, !=, NULL);
9104 * L2ARC fast warmup.
9106 * Until the ARC is warm and starts to evict, read from the
9107 * head of the ARC lists rather than the tail.
9109 if (arc_warm == B_FALSE)
9110 hdr = multilist_sublist_head(mls);
9112 hdr = multilist_sublist_tail(mls);
9114 headroom = target_sz * l2arc_headroom;
9115 if (zfs_compressed_arc_enabled)
9116 headroom = (headroom * l2arc_headroom_boost) / 100;
9118 for (; hdr; hdr = hdr_prev) {
9119 kmutex_t *hash_lock;
9120 abd_t *to_write = NULL;
9122 if (arc_warm == B_FALSE)
9123 hdr_prev = multilist_sublist_next(mls, hdr);
9125 hdr_prev = multilist_sublist_prev(mls, hdr);
9127 hash_lock = HDR_LOCK(hdr);
9128 if (!mutex_tryenter(hash_lock)) {
9130 * Skip this buffer rather than waiting.
9135 passed_sz += HDR_GET_LSIZE(hdr);
9136 if (l2arc_headroom != 0 && passed_sz > headroom) {
9140 mutex_exit(hash_lock);
9144 if (!l2arc_write_eligible(guid, hdr)) {
9145 mutex_exit(hash_lock);
9149 ASSERT(HDR_HAS_L1HDR(hdr));
9151 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
9152 ASSERT3U(arc_hdr_size(hdr), >, 0);
9153 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
9155 uint64_t psize = HDR_GET_PSIZE(hdr);
9156 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
9160 * If the allocated size of this buffer plus the max
9161 * size for the pending log block exceeds the evicted
9162 * target size, terminate writing buffers for this run.
9164 if (write_asize + asize +
9165 sizeof (l2arc_log_blk_phys_t) > target_sz) {
9167 mutex_exit(hash_lock);
9172 * We rely on the L1 portion of the header below, so
9173 * it's invalid for this header to have been evicted out
9174 * of the ghost cache, prior to being written out. The
9175 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
9177 arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
9180 * If this header has b_rabd, we can use this since it
9181 * must always match the data exactly as it exists on
9182 * disk. Otherwise, the L2ARC can normally use the
9183 * hdr's data, but if we're sharing data between the
9184 * hdr and one of its bufs, L2ARC needs its own copy of
9185 * the data so that the ZIO below can't race with the
9186 * buf consumer. To ensure that this copy will be
9187 * available for the lifetime of the ZIO and be cleaned
9188 * up afterwards, we add it to the l2arc_free_on_write
9189 * queue. If we need to apply any transforms to the
9190 * data (compression, encryption) we will also need the
9193 if (HDR_HAS_RABD(hdr) && psize == asize) {
9194 to_write = hdr->b_crypt_hdr.b_rabd;
9195 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
9196 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
9197 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
9199 to_write = hdr->b_l1hdr.b_pabd;
9202 arc_buf_contents_t type = arc_buf_type(hdr);
9204 ret = l2arc_apply_transforms(spa, hdr, asize,
9207 arc_hdr_clear_flags(hdr,
9208 ARC_FLAG_L2_WRITING);
9209 mutex_exit(hash_lock);
9213 l2arc_free_abd_on_write(to_write, asize, type);
9218 * Insert a dummy header on the buflist so
9219 * l2arc_write_done() can find where the
9220 * write buffers begin without searching.
9222 mutex_enter(&dev->l2ad_mtx);
9223 list_insert_head(&dev->l2ad_buflist, head);
9224 mutex_exit(&dev->l2ad_mtx);
9227 sizeof (l2arc_write_callback_t), KM_SLEEP);
9228 cb->l2wcb_dev = dev;
9229 cb->l2wcb_head = head;
9231 * Create a list to save allocated abd buffers
9232 * for l2arc_log_blk_commit().
9234 list_create(&cb->l2wcb_abd_list,
9235 sizeof (l2arc_lb_abd_buf_t),
9236 offsetof(l2arc_lb_abd_buf_t, node));
9237 pio = zio_root(spa, l2arc_write_done, cb,
9241 hdr->b_l2hdr.b_dev = dev;
9242 hdr->b_l2hdr.b_hits = 0;
9244 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
9245 hdr->b_l2hdr.b_arcs_state =
9246 hdr->b_l1hdr.b_state->arcs_state;
9247 arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR);
9249 mutex_enter(&dev->l2ad_mtx);
9250 list_insert_head(&dev->l2ad_buflist, hdr);
9251 mutex_exit(&dev->l2ad_mtx);
9253 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
9254 arc_hdr_size(hdr), hdr);
9256 wzio = zio_write_phys(pio, dev->l2ad_vdev,
9257 hdr->b_l2hdr.b_daddr, asize, to_write,
9258 ZIO_CHECKSUM_OFF, NULL, hdr,
9259 ZIO_PRIORITY_ASYNC_WRITE,
9260 ZIO_FLAG_CANFAIL, B_FALSE);
9262 write_lsize += HDR_GET_LSIZE(hdr);
9263 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
9266 write_psize += psize;
9267 write_asize += asize;
9268 dev->l2ad_hand += asize;
9269 l2arc_hdr_arcstats_increment(hdr);
9270 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9272 mutex_exit(hash_lock);
9275 * Append buf info to current log and commit if full.
9276 * arcstat_l2_{size,asize} kstats are updated
9279 if (l2arc_log_blk_insert(dev, hdr)) {
9281 * l2ad_hand will be adjusted in
9282 * l2arc_log_blk_commit().
9285 l2arc_log_blk_commit(dev, pio, cb);
9291 multilist_sublist_unlock(mls);
9297 /* No buffers selected for writing? */
9299 ASSERT0(write_lsize);
9300 ASSERT(!HDR_HAS_L1HDR(head));
9301 kmem_cache_free(hdr_l2only_cache, head);
9304 * Although we did not write any buffers l2ad_evict may
9307 if (dev->l2ad_evict != l2dhdr->dh_evict)
9308 l2arc_dev_hdr_update(dev);
9313 if (!dev->l2ad_first)
9314 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
9316 ASSERT3U(write_asize, <=, target_sz);
9317 ARCSTAT_BUMP(arcstat_l2_writes_sent);
9318 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
9320 dev->l2ad_writing = B_TRUE;
9321 (void) zio_wait(pio);
9322 dev->l2ad_writing = B_FALSE;
9325 * Update the device header after the zio completes as
9326 * l2arc_write_done() may have updated the memory holding the log block
9327 * pointers in the device header.
9329 l2arc_dev_hdr_update(dev);
9331 return (write_asize);
9335 l2arc_hdr_limit_reached(void)
9337 int64_t s = aggsum_upper_bound(&arc_sums.arcstat_l2_hdr_size);
9339 return (arc_reclaim_needed() ||
9340 (s > (arc_warm ? arc_c : arc_c_max) * l2arc_meta_percent / 100));
9344 * This thread feeds the L2ARC at regular intervals. This is the beating
9345 * heart of the L2ARC.
9347 static __attribute__((noreturn)) void
9348 l2arc_feed_thread(void *unused)
9354 uint64_t size, wrote;
9355 clock_t begin, next = ddi_get_lbolt();
9356 fstrans_cookie_t cookie;
9358 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
9360 mutex_enter(&l2arc_feed_thr_lock);
9362 cookie = spl_fstrans_mark();
9363 while (l2arc_thread_exit == 0) {
9364 CALLB_CPR_SAFE_BEGIN(&cpr);
9365 (void) cv_timedwait_idle(&l2arc_feed_thr_cv,
9366 &l2arc_feed_thr_lock, next);
9367 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
9368 next = ddi_get_lbolt() + hz;
9371 * Quick check for L2ARC devices.
9373 mutex_enter(&l2arc_dev_mtx);
9374 if (l2arc_ndev == 0) {
9375 mutex_exit(&l2arc_dev_mtx);
9378 mutex_exit(&l2arc_dev_mtx);
9379 begin = ddi_get_lbolt();
9382 * This selects the next l2arc device to write to, and in
9383 * doing so the next spa to feed from: dev->l2ad_spa. This
9384 * will return NULL if there are now no l2arc devices or if
9385 * they are all faulted.
9387 * If a device is returned, its spa's config lock is also
9388 * held to prevent device removal. l2arc_dev_get_next()
9389 * will grab and release l2arc_dev_mtx.
9391 if ((dev = l2arc_dev_get_next()) == NULL)
9394 spa = dev->l2ad_spa;
9395 ASSERT3P(spa, !=, NULL);
9398 * If the pool is read-only then force the feed thread to
9399 * sleep a little longer.
9401 if (!spa_writeable(spa)) {
9402 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
9403 spa_config_exit(spa, SCL_L2ARC, dev);
9408 * Avoid contributing to memory pressure.
9410 if (l2arc_hdr_limit_reached()) {
9411 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
9412 spa_config_exit(spa, SCL_L2ARC, dev);
9416 ARCSTAT_BUMP(arcstat_l2_feeds);
9418 size = l2arc_write_size(dev);
9421 * Evict L2ARC buffers that will be overwritten.
9423 l2arc_evict(dev, size, B_FALSE);
9426 * Write ARC buffers.
9428 wrote = l2arc_write_buffers(spa, dev, size);
9431 * Calculate interval between writes.
9433 next = l2arc_write_interval(begin, size, wrote);
9434 spa_config_exit(spa, SCL_L2ARC, dev);
9436 spl_fstrans_unmark(cookie);
9438 l2arc_thread_exit = 0;
9439 cv_broadcast(&l2arc_feed_thr_cv);
9440 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
9445 l2arc_vdev_present(vdev_t *vd)
9447 return (l2arc_vdev_get(vd) != NULL);
9451 * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if
9452 * the vdev_t isn't an L2ARC device.
9455 l2arc_vdev_get(vdev_t *vd)
9459 mutex_enter(&l2arc_dev_mtx);
9460 for (dev = list_head(l2arc_dev_list); dev != NULL;
9461 dev = list_next(l2arc_dev_list, dev)) {
9462 if (dev->l2ad_vdev == vd)
9465 mutex_exit(&l2arc_dev_mtx);
9471 l2arc_rebuild_dev(l2arc_dev_t *dev, boolean_t reopen)
9473 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9474 uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9475 spa_t *spa = dev->l2ad_spa;
9478 * The L2ARC has to hold at least the payload of one log block for
9479 * them to be restored (persistent L2ARC). The payload of a log block
9480 * depends on the amount of its log entries. We always write log blocks
9481 * with 1022 entries. How many of them are committed or restored depends
9482 * on the size of the L2ARC device. Thus the maximum payload of
9483 * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device
9484 * is less than that, we reduce the amount of committed and restored
9485 * log entries per block so as to enable persistence.
9487 if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) {
9488 dev->l2ad_log_entries = 0;
9490 dev->l2ad_log_entries = MIN((dev->l2ad_end -
9491 dev->l2ad_start) >> SPA_MAXBLOCKSHIFT,
9492 L2ARC_LOG_BLK_MAX_ENTRIES);
9496 * Read the device header, if an error is returned do not rebuild L2ARC.
9498 if (l2arc_dev_hdr_read(dev) == 0 && dev->l2ad_log_entries > 0) {
9500 * If we are onlining a cache device (vdev_reopen) that was
9501 * still present (l2arc_vdev_present()) and rebuild is enabled,
9502 * we should evict all ARC buffers and pointers to log blocks
9503 * and reclaim their space before restoring its contents to
9507 if (!l2arc_rebuild_enabled) {
9510 l2arc_evict(dev, 0, B_TRUE);
9511 /* start a new log block */
9512 dev->l2ad_log_ent_idx = 0;
9513 dev->l2ad_log_blk_payload_asize = 0;
9514 dev->l2ad_log_blk_payload_start = 0;
9518 * Just mark the device as pending for a rebuild. We won't
9519 * be starting a rebuild in line here as it would block pool
9520 * import. Instead spa_load_impl will hand that off to an
9521 * async task which will call l2arc_spa_rebuild_start.
9523 dev->l2ad_rebuild = B_TRUE;
9524 } else if (spa_writeable(spa)) {
9526 * In this case TRIM the whole device if l2arc_trim_ahead > 0,
9527 * otherwise create a new header. We zero out the memory holding
9528 * the header to reset dh_start_lbps. If we TRIM the whole
9529 * device the new header will be written by
9530 * vdev_trim_l2arc_thread() at the end of the TRIM to update the
9531 * trim_state in the header too. When reading the header, if
9532 * trim_state is not VDEV_TRIM_COMPLETE and l2arc_trim_ahead > 0
9533 * we opt to TRIM the whole device again.
9535 if (l2arc_trim_ahead > 0) {
9536 dev->l2ad_trim_all = B_TRUE;
9538 memset(l2dhdr, 0, l2dhdr_asize);
9539 l2arc_dev_hdr_update(dev);
9545 * Add a vdev for use by the L2ARC. By this point the spa has already
9546 * validated the vdev and opened it.
9549 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
9551 l2arc_dev_t *adddev;
9552 uint64_t l2dhdr_asize;
9554 ASSERT(!l2arc_vdev_present(vd));
9557 * Create a new l2arc device entry.
9559 adddev = vmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
9560 adddev->l2ad_spa = spa;
9561 adddev->l2ad_vdev = vd;
9562 /* leave extra size for an l2arc device header */
9563 l2dhdr_asize = adddev->l2ad_dev_hdr_asize =
9564 MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift);
9565 adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize;
9566 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
9567 ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end);
9568 adddev->l2ad_hand = adddev->l2ad_start;
9569 adddev->l2ad_evict = adddev->l2ad_start;
9570 adddev->l2ad_first = B_TRUE;
9571 adddev->l2ad_writing = B_FALSE;
9572 adddev->l2ad_trim_all = B_FALSE;
9573 list_link_init(&adddev->l2ad_node);
9574 adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP);
9576 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
9578 * This is a list of all ARC buffers that are still valid on the
9581 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
9582 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
9585 * This is a list of pointers to log blocks that are still present
9588 list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t),
9589 offsetof(l2arc_lb_ptr_buf_t, node));
9591 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
9592 zfs_refcount_create(&adddev->l2ad_alloc);
9593 zfs_refcount_create(&adddev->l2ad_lb_asize);
9594 zfs_refcount_create(&adddev->l2ad_lb_count);
9597 * Decide if dev is eligible for L2ARC rebuild or whole device
9598 * trimming. This has to happen before the device is added in the
9599 * cache device list and l2arc_dev_mtx is released. Otherwise
9600 * l2arc_feed_thread() might already start writing on the
9603 l2arc_rebuild_dev(adddev, B_FALSE);
9606 * Add device to global list
9608 mutex_enter(&l2arc_dev_mtx);
9609 list_insert_head(l2arc_dev_list, adddev);
9610 atomic_inc_64(&l2arc_ndev);
9611 mutex_exit(&l2arc_dev_mtx);
9615 * Decide if a vdev is eligible for L2ARC rebuild, called from vdev_reopen()
9616 * in case of onlining a cache device.
9619 l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen)
9621 l2arc_dev_t *dev = NULL;
9623 dev = l2arc_vdev_get(vd);
9624 ASSERT3P(dev, !=, NULL);
9627 * In contrast to l2arc_add_vdev() we do not have to worry about
9628 * l2arc_feed_thread() invalidating previous content when onlining a
9629 * cache device. The device parameters (l2ad*) are not cleared when
9630 * offlining the device and writing new buffers will not invalidate
9631 * all previous content. In worst case only buffers that have not had
9632 * their log block written to the device will be lost.
9633 * When onlining the cache device (ie offline->online without exporting
9634 * the pool in between) this happens:
9635 * vdev_reopen() -> vdev_open() -> l2arc_rebuild_vdev()
9637 * vdev_is_dead() = B_FALSE l2ad_rebuild = B_TRUE
9638 * During the time where vdev_is_dead = B_FALSE and until l2ad_rebuild
9639 * is set to B_TRUE we might write additional buffers to the device.
9641 l2arc_rebuild_dev(dev, reopen);
9645 * Remove a vdev from the L2ARC.
9648 l2arc_remove_vdev(vdev_t *vd)
9650 l2arc_dev_t *remdev = NULL;
9653 * Find the device by vdev
9655 remdev = l2arc_vdev_get(vd);
9656 ASSERT3P(remdev, !=, NULL);
9659 * Cancel any ongoing or scheduled rebuild.
9661 mutex_enter(&l2arc_rebuild_thr_lock);
9662 if (remdev->l2ad_rebuild_began == B_TRUE) {
9663 remdev->l2ad_rebuild_cancel = B_TRUE;
9664 while (remdev->l2ad_rebuild == B_TRUE)
9665 cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock);
9667 mutex_exit(&l2arc_rebuild_thr_lock);
9670 * Remove device from global list
9672 mutex_enter(&l2arc_dev_mtx);
9673 list_remove(l2arc_dev_list, remdev);
9674 l2arc_dev_last = NULL; /* may have been invalidated */
9675 atomic_dec_64(&l2arc_ndev);
9676 mutex_exit(&l2arc_dev_mtx);
9679 * Clear all buflists and ARC references. L2ARC device flush.
9681 l2arc_evict(remdev, 0, B_TRUE);
9682 list_destroy(&remdev->l2ad_buflist);
9683 ASSERT(list_is_empty(&remdev->l2ad_lbptr_list));
9684 list_destroy(&remdev->l2ad_lbptr_list);
9685 mutex_destroy(&remdev->l2ad_mtx);
9686 zfs_refcount_destroy(&remdev->l2ad_alloc);
9687 zfs_refcount_destroy(&remdev->l2ad_lb_asize);
9688 zfs_refcount_destroy(&remdev->l2ad_lb_count);
9689 kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize);
9690 vmem_free(remdev, sizeof (l2arc_dev_t));
9696 l2arc_thread_exit = 0;
9699 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9700 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
9701 mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9702 cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL);
9703 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
9704 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
9706 l2arc_dev_list = &L2ARC_dev_list;
9707 l2arc_free_on_write = &L2ARC_free_on_write;
9708 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
9709 offsetof(l2arc_dev_t, l2ad_node));
9710 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
9711 offsetof(l2arc_data_free_t, l2df_list_node));
9717 mutex_destroy(&l2arc_feed_thr_lock);
9718 cv_destroy(&l2arc_feed_thr_cv);
9719 mutex_destroy(&l2arc_rebuild_thr_lock);
9720 cv_destroy(&l2arc_rebuild_thr_cv);
9721 mutex_destroy(&l2arc_dev_mtx);
9722 mutex_destroy(&l2arc_free_on_write_mtx);
9724 list_destroy(l2arc_dev_list);
9725 list_destroy(l2arc_free_on_write);
9731 if (!(spa_mode_global & SPA_MODE_WRITE))
9734 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
9735 TS_RUN, defclsyspri);
9741 if (!(spa_mode_global & SPA_MODE_WRITE))
9744 mutex_enter(&l2arc_feed_thr_lock);
9745 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
9746 l2arc_thread_exit = 1;
9747 while (l2arc_thread_exit != 0)
9748 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
9749 mutex_exit(&l2arc_feed_thr_lock);
9753 * Punches out rebuild threads for the L2ARC devices in a spa. This should
9754 * be called after pool import from the spa async thread, since starting
9755 * these threads directly from spa_import() will make them part of the
9756 * "zpool import" context and delay process exit (and thus pool import).
9759 l2arc_spa_rebuild_start(spa_t *spa)
9761 ASSERT(MUTEX_HELD(&spa_namespace_lock));
9764 * Locate the spa's l2arc devices and kick off rebuild threads.
9766 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
9768 l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]);
9770 /* Don't attempt a rebuild if the vdev is UNAVAIL */
9773 mutex_enter(&l2arc_rebuild_thr_lock);
9774 if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) {
9775 dev->l2ad_rebuild_began = B_TRUE;
9776 (void) thread_create(NULL, 0, l2arc_dev_rebuild_thread,
9777 dev, 0, &p0, TS_RUN, minclsyspri);
9779 mutex_exit(&l2arc_rebuild_thr_lock);
9784 * Main entry point for L2ARC rebuilding.
9786 static __attribute__((noreturn)) void
9787 l2arc_dev_rebuild_thread(void *arg)
9789 l2arc_dev_t *dev = arg;
9791 VERIFY(!dev->l2ad_rebuild_cancel);
9792 VERIFY(dev->l2ad_rebuild);
9793 (void) l2arc_rebuild(dev);
9794 mutex_enter(&l2arc_rebuild_thr_lock);
9795 dev->l2ad_rebuild_began = B_FALSE;
9796 dev->l2ad_rebuild = B_FALSE;
9797 mutex_exit(&l2arc_rebuild_thr_lock);
9803 * This function implements the actual L2ARC metadata rebuild. It:
9804 * starts reading the log block chain and restores each block's contents
9805 * to memory (reconstructing arc_buf_hdr_t's).
9807 * Operation stops under any of the following conditions:
9809 * 1) We reach the end of the log block chain.
9810 * 2) We encounter *any* error condition (cksum errors, io errors)
9813 l2arc_rebuild(l2arc_dev_t *dev)
9815 vdev_t *vd = dev->l2ad_vdev;
9816 spa_t *spa = vd->vdev_spa;
9818 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9819 l2arc_log_blk_phys_t *this_lb, *next_lb;
9820 zio_t *this_io = NULL, *next_io = NULL;
9821 l2arc_log_blkptr_t lbps[2];
9822 l2arc_lb_ptr_buf_t *lb_ptr_buf;
9823 boolean_t lock_held;
9825 this_lb = vmem_zalloc(sizeof (*this_lb), KM_SLEEP);
9826 next_lb = vmem_zalloc(sizeof (*next_lb), KM_SLEEP);
9829 * We prevent device removal while issuing reads to the device,
9830 * then during the rebuilding phases we drop this lock again so
9831 * that a spa_unload or device remove can be initiated - this is
9832 * safe, because the spa will signal us to stop before removing
9833 * our device and wait for us to stop.
9835 spa_config_enter(spa, SCL_L2ARC, vd, RW_READER);
9839 * Retrieve the persistent L2ARC device state.
9840 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9842 dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start);
9843 dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr +
9844 L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop),
9846 dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST);
9848 vd->vdev_trim_action_time = l2dhdr->dh_trim_action_time;
9849 vd->vdev_trim_state = l2dhdr->dh_trim_state;
9852 * In case the zfs module parameter l2arc_rebuild_enabled is false
9853 * we do not start the rebuild process.
9855 if (!l2arc_rebuild_enabled)
9858 /* Prepare the rebuild process */
9859 memcpy(lbps, l2dhdr->dh_start_lbps, sizeof (lbps));
9861 /* Start the rebuild process */
9863 if (!l2arc_log_blkptr_valid(dev, &lbps[0]))
9866 if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1],
9867 this_lb, next_lb, this_io, &next_io)) != 0)
9871 * Our memory pressure valve. If the system is running low
9872 * on memory, rather than swamping memory with new ARC buf
9873 * hdrs, we opt not to rebuild the L2ARC. At this point,
9874 * however, we have already set up our L2ARC dev to chain in
9875 * new metadata log blocks, so the user may choose to offline/
9876 * online the L2ARC dev at a later time (or re-import the pool)
9877 * to reconstruct it (when there's less memory pressure).
9879 if (l2arc_hdr_limit_reached()) {
9880 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem);
9881 cmn_err(CE_NOTE, "System running low on memory, "
9882 "aborting L2ARC rebuild.");
9883 err = SET_ERROR(ENOMEM);
9887 spa_config_exit(spa, SCL_L2ARC, vd);
9888 lock_held = B_FALSE;
9891 * Now that we know that the next_lb checks out alright, we
9892 * can start reconstruction from this log block.
9893 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9895 uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop);
9896 l2arc_log_blk_restore(dev, this_lb, asize);
9899 * log block restored, include its pointer in the list of
9900 * pointers to log blocks present in the L2ARC device.
9902 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
9903 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t),
9905 memcpy(lb_ptr_buf->lb_ptr, &lbps[0],
9906 sizeof (l2arc_log_blkptr_t));
9907 mutex_enter(&dev->l2ad_mtx);
9908 list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf);
9909 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
9910 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
9911 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
9912 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
9913 mutex_exit(&dev->l2ad_mtx);
9914 vdev_space_update(vd, asize, 0, 0);
9917 * Protection against loops of log blocks:
9919 * l2ad_hand l2ad_evict
9921 * l2ad_start |=======================================| l2ad_end
9922 * -----|||----|||---|||----|||
9924 * ---|||---|||----|||---|||
9927 * In this situation the pointer of log block (4) passes
9928 * l2arc_log_blkptr_valid() but the log block should not be
9929 * restored as it is overwritten by the payload of log block
9930 * (0). Only log blocks (0)-(3) should be restored. We check
9931 * whether l2ad_evict lies in between the payload starting
9932 * offset of the next log block (lbps[1].lbp_payload_start)
9933 * and the payload starting offset of the present log block
9934 * (lbps[0].lbp_payload_start). If true and this isn't the
9935 * first pass, we are looping from the beginning and we should
9938 if (l2arc_range_check_overlap(lbps[1].lbp_payload_start,
9939 lbps[0].lbp_payload_start, dev->l2ad_evict) &&
9943 kpreempt(KPREEMPT_SYNC);
9945 mutex_enter(&l2arc_rebuild_thr_lock);
9946 if (dev->l2ad_rebuild_cancel) {
9947 dev->l2ad_rebuild = B_FALSE;
9948 cv_signal(&l2arc_rebuild_thr_cv);
9949 mutex_exit(&l2arc_rebuild_thr_lock);
9950 err = SET_ERROR(ECANCELED);
9953 mutex_exit(&l2arc_rebuild_thr_lock);
9954 if (spa_config_tryenter(spa, SCL_L2ARC, vd,
9960 * L2ARC config lock held by somebody in writer,
9961 * possibly due to them trying to remove us. They'll
9962 * likely to want us to shut down, so after a little
9963 * delay, we check l2ad_rebuild_cancel and retry
9970 * Continue with the next log block.
9973 lbps[1] = this_lb->lb_prev_lbp;
9974 PTR_SWAP(this_lb, next_lb);
9979 if (this_io != NULL)
9980 l2arc_log_blk_fetch_abort(this_io);
9982 if (next_io != NULL)
9983 l2arc_log_blk_fetch_abort(next_io);
9984 vmem_free(this_lb, sizeof (*this_lb));
9985 vmem_free(next_lb, sizeof (*next_lb));
9987 if (!l2arc_rebuild_enabled) {
9988 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9990 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) {
9991 ARCSTAT_BUMP(arcstat_l2_rebuild_success);
9992 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9993 "successful, restored %llu blocks",
9994 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
9995 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) {
9997 * No error but also nothing restored, meaning the lbps array
9998 * in the device header points to invalid/non-present log
9999 * blocks. Reset the header.
10001 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10002 "no valid log blocks");
10003 memset(l2dhdr, 0, dev->l2ad_dev_hdr_asize);
10004 l2arc_dev_hdr_update(dev);
10005 } else if (err == ECANCELED) {
10007 * In case the rebuild was canceled do not log to spa history
10008 * log as the pool may be in the process of being removed.
10010 zfs_dbgmsg("L2ARC rebuild aborted, restored %llu blocks",
10011 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10012 } else if (err != 0) {
10013 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10014 "aborted, restored %llu blocks",
10015 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10019 spa_config_exit(spa, SCL_L2ARC, vd);
10025 * Attempts to read the device header on the provided L2ARC device and writes
10026 * it to `hdr'. On success, this function returns 0, otherwise the appropriate
10027 * error code is returned.
10030 l2arc_dev_hdr_read(l2arc_dev_t *dev)
10034 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10035 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10038 guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10040 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10042 err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev,
10043 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd,
10044 ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
10045 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY |
10046 ZIO_FLAG_SPECULATIVE, B_FALSE));
10051 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors);
10052 zfs_dbgmsg("L2ARC IO error (%d) while reading device header, "
10053 "vdev guid: %llu", err,
10054 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10058 if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC))
10059 byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr));
10061 if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC ||
10062 l2dhdr->dh_spa_guid != guid ||
10063 l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid ||
10064 l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION ||
10065 l2dhdr->dh_log_entries != dev->l2ad_log_entries ||
10066 l2dhdr->dh_end != dev->l2ad_end ||
10067 !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end,
10068 l2dhdr->dh_evict) ||
10069 (l2dhdr->dh_trim_state != VDEV_TRIM_COMPLETE &&
10070 l2arc_trim_ahead > 0)) {
10072 * Attempt to rebuild a device containing no actual dev hdr
10073 * or containing a header from some other pool or from another
10074 * version of persistent L2ARC.
10076 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported);
10077 return (SET_ERROR(ENOTSUP));
10084 * Reads L2ARC log blocks from storage and validates their contents.
10086 * This function implements a simple fetcher to make sure that while
10087 * we're processing one buffer the L2ARC is already fetching the next
10088 * one in the chain.
10090 * The arguments this_lp and next_lp point to the current and next log block
10091 * address in the block chain. Similarly, this_lb and next_lb hold the
10092 * l2arc_log_blk_phys_t's of the current and next L2ARC blk.
10094 * The `this_io' and `next_io' arguments are used for block fetching.
10095 * When issuing the first blk IO during rebuild, you should pass NULL for
10096 * `this_io'. This function will then issue a sync IO to read the block and
10097 * also issue an async IO to fetch the next block in the block chain. The
10098 * fetched IO is returned in `next_io'. On subsequent calls to this
10099 * function, pass the value returned in `next_io' from the previous call
10100 * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO.
10101 * Prior to the call, you should initialize your `next_io' pointer to be
10102 * NULL. If no fetch IO was issued, the pointer is left set at NULL.
10104 * On success, this function returns 0, otherwise it returns an appropriate
10105 * error code. On error the fetching IO is aborted and cleared before
10106 * returning from this function. Therefore, if we return `success', the
10107 * caller can assume that we have taken care of cleanup of fetch IOs.
10110 l2arc_log_blk_read(l2arc_dev_t *dev,
10111 const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp,
10112 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
10113 zio_t *this_io, zio_t **next_io)
10120 ASSERT(this_lbp != NULL && next_lbp != NULL);
10121 ASSERT(this_lb != NULL && next_lb != NULL);
10122 ASSERT(next_io != NULL && *next_io == NULL);
10123 ASSERT(l2arc_log_blkptr_valid(dev, this_lbp));
10126 * Check to see if we have issued the IO for this log block in a
10127 * previous run. If not, this is the first call, so issue it now.
10129 if (this_io == NULL) {
10130 this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp,
10135 * Peek to see if we can start issuing the next IO immediately.
10137 if (l2arc_log_blkptr_valid(dev, next_lbp)) {
10139 * Start issuing IO for the next log block early - this
10140 * should help keep the L2ARC device busy while we
10141 * decompress and restore this log block.
10143 *next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp,
10147 /* Wait for the IO to read this log block to complete */
10148 if ((err = zio_wait(this_io)) != 0) {
10149 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
10150 zfs_dbgmsg("L2ARC IO error (%d) while reading log block, "
10151 "offset: %llu, vdev guid: %llu", err,
10152 (u_longlong_t)this_lbp->lbp_daddr,
10153 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10158 * Make sure the buffer checks out.
10159 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10161 asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop);
10162 fletcher_4_native(this_lb, asize, NULL, &cksum);
10163 if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) {
10164 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors);
10165 zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, "
10166 "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu",
10167 (u_longlong_t)this_lbp->lbp_daddr,
10168 (u_longlong_t)dev->l2ad_vdev->vdev_guid,
10169 (u_longlong_t)dev->l2ad_hand,
10170 (u_longlong_t)dev->l2ad_evict);
10171 err = SET_ERROR(ECKSUM);
10175 /* Now we can take our time decoding this buffer */
10176 switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) {
10177 case ZIO_COMPRESS_OFF:
10179 case ZIO_COMPRESS_LZ4:
10180 abd = abd_alloc_for_io(asize, B_TRUE);
10181 abd_copy_from_buf_off(abd, this_lb, 0, asize);
10182 if ((err = zio_decompress_data(
10183 L2BLK_GET_COMPRESS((this_lbp)->lbp_prop),
10184 abd, this_lb, asize, sizeof (*this_lb), NULL)) != 0) {
10185 err = SET_ERROR(EINVAL);
10190 err = SET_ERROR(EINVAL);
10193 if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC))
10194 byteswap_uint64_array(this_lb, sizeof (*this_lb));
10195 if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) {
10196 err = SET_ERROR(EINVAL);
10200 /* Abort an in-flight fetch I/O in case of error */
10201 if (err != 0 && *next_io != NULL) {
10202 l2arc_log_blk_fetch_abort(*next_io);
10211 * Restores the payload of a log block to ARC. This creates empty ARC hdr
10212 * entries which only contain an l2arc hdr, essentially restoring the
10213 * buffers to their L2ARC evicted state. This function also updates space
10214 * usage on the L2ARC vdev to make sure it tracks restored buffers.
10217 l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb,
10220 uint64_t size = 0, asize = 0;
10221 uint64_t log_entries = dev->l2ad_log_entries;
10224 * Usually arc_adapt() is called only for data, not headers, but
10225 * since we may allocate significant amount of memory here, let ARC
10228 arc_adapt(log_entries * HDR_L2ONLY_SIZE);
10230 for (int i = log_entries - 1; i >= 0; i--) {
10232 * Restore goes in the reverse temporal direction to preserve
10233 * correct temporal ordering of buffers in the l2ad_buflist.
10234 * l2arc_hdr_restore also does a list_insert_tail instead of
10235 * list_insert_head on the l2ad_buflist:
10237 * LIST l2ad_buflist LIST
10238 * HEAD <------ (time) ------ TAIL
10239 * direction +-----+-----+-----+-----+-----+ direction
10240 * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild
10241 * fill +-----+-----+-----+-----+-----+
10245 * l2arc_feed_thread l2arc_rebuild
10246 * will place new bufs here restores bufs here
10248 * During l2arc_rebuild() the device is not used by
10249 * l2arc_feed_thread() as dev->l2ad_rebuild is set to true.
10251 size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop);
10252 asize += vdev_psize_to_asize(dev->l2ad_vdev,
10253 L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop));
10254 l2arc_hdr_restore(&lb->lb_entries[i], dev);
10258 * Record rebuild stats:
10259 * size Logical size of restored buffers in the L2ARC
10260 * asize Aligned size of restored buffers in the L2ARC
10262 ARCSTAT_INCR(arcstat_l2_rebuild_size, size);
10263 ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize);
10264 ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries);
10265 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize);
10266 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize);
10267 ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks);
10271 * Restores a single ARC buf hdr from a log entry. The ARC buffer is put
10272 * into a state indicating that it has been evicted to L2ARC.
10275 l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
10277 arc_buf_hdr_t *hdr, *exists;
10278 kmutex_t *hash_lock;
10279 arc_buf_contents_t type = L2BLK_GET_TYPE((le)->le_prop);
10283 * Do all the allocation before grabbing any locks, this lets us
10284 * sleep if memory is full and we don't have to deal with failed
10287 hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type,
10288 dev, le->le_dva, le->le_daddr,
10289 L2BLK_GET_PSIZE((le)->le_prop), le->le_birth,
10290 L2BLK_GET_COMPRESS((le)->le_prop), le->le_complevel,
10291 L2BLK_GET_PROTECTED((le)->le_prop),
10292 L2BLK_GET_PREFETCH((le)->le_prop),
10293 L2BLK_GET_STATE((le)->le_prop));
10294 asize = vdev_psize_to_asize(dev->l2ad_vdev,
10295 L2BLK_GET_PSIZE((le)->le_prop));
10298 * vdev_space_update() has to be called before arc_hdr_destroy() to
10299 * avoid underflow since the latter also calls vdev_space_update().
10301 l2arc_hdr_arcstats_increment(hdr);
10302 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10304 mutex_enter(&dev->l2ad_mtx);
10305 list_insert_tail(&dev->l2ad_buflist, hdr);
10306 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
10307 mutex_exit(&dev->l2ad_mtx);
10309 exists = buf_hash_insert(hdr, &hash_lock);
10311 /* Buffer was already cached, no need to restore it. */
10312 arc_hdr_destroy(hdr);
10314 * If the buffer is already cached, check whether it has
10315 * L2ARC metadata. If not, enter them and update the flag.
10316 * This is important is case of onlining a cache device, since
10317 * we previously evicted all L2ARC metadata from ARC.
10319 if (!HDR_HAS_L2HDR(exists)) {
10320 arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR);
10321 exists->b_l2hdr.b_dev = dev;
10322 exists->b_l2hdr.b_daddr = le->le_daddr;
10323 exists->b_l2hdr.b_arcs_state =
10324 L2BLK_GET_STATE((le)->le_prop);
10325 mutex_enter(&dev->l2ad_mtx);
10326 list_insert_tail(&dev->l2ad_buflist, exists);
10327 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
10328 arc_hdr_size(exists), exists);
10329 mutex_exit(&dev->l2ad_mtx);
10330 l2arc_hdr_arcstats_increment(exists);
10331 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10333 ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
10336 mutex_exit(hash_lock);
10340 * Starts an asynchronous read IO to read a log block. This is used in log
10341 * block reconstruction to start reading the next block before we are done
10342 * decoding and reconstructing the current block, to keep the l2arc device
10343 * nice and hot with read IO to process.
10344 * The returned zio will contain a newly allocated memory buffers for the IO
10345 * data which should then be freed by the caller once the zio is no longer
10346 * needed (i.e. due to it having completed). If you wish to abort this
10347 * zio, you should do so using l2arc_log_blk_fetch_abort, which takes
10348 * care of disposing of the allocated buffers correctly.
10351 l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp,
10352 l2arc_log_blk_phys_t *lb)
10356 l2arc_read_callback_t *cb;
10358 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10359 asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10360 ASSERT(asize <= sizeof (l2arc_log_blk_phys_t));
10362 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP);
10363 cb->l2rcb_abd = abd_get_from_buf(lb, asize);
10364 pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb,
10365 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
10366 (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize,
10367 cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10368 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL |
10369 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
10375 * Aborts a zio returned from l2arc_log_blk_fetch and frees the data
10376 * buffers allocated for it.
10379 l2arc_log_blk_fetch_abort(zio_t *zio)
10381 (void) zio_wait(zio);
10385 * Creates a zio to update the device header on an l2arc device.
10388 l2arc_dev_hdr_update(l2arc_dev_t *dev)
10390 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10391 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10395 VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER));
10397 l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC;
10398 l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION;
10399 l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10400 l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid;
10401 l2dhdr->dh_log_entries = dev->l2ad_log_entries;
10402 l2dhdr->dh_evict = dev->l2ad_evict;
10403 l2dhdr->dh_start = dev->l2ad_start;
10404 l2dhdr->dh_end = dev->l2ad_end;
10405 l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize);
10406 l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count);
10407 l2dhdr->dh_flags = 0;
10408 l2dhdr->dh_trim_action_time = dev->l2ad_vdev->vdev_trim_action_time;
10409 l2dhdr->dh_trim_state = dev->l2ad_vdev->vdev_trim_state;
10410 if (dev->l2ad_first)
10411 l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST;
10413 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10415 err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev,
10416 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL,
10417 NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE));
10422 zfs_dbgmsg("L2ARC IO error (%d) while writing device header, "
10423 "vdev guid: %llu", err,
10424 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10429 * Commits a log block to the L2ARC device. This routine is invoked from
10430 * l2arc_write_buffers when the log block fills up.
10431 * This function allocates some memory to temporarily hold the serialized
10432 * buffer to be written. This is then released in l2arc_write_done.
10435 l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb)
10437 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10438 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10439 uint64_t psize, asize;
10441 l2arc_lb_abd_buf_t *abd_buf;
10442 uint8_t *tmpbuf = NULL;
10443 l2arc_lb_ptr_buf_t *lb_ptr_buf;
10445 VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries);
10447 abd_buf = zio_buf_alloc(sizeof (*abd_buf));
10448 abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb));
10449 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
10450 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP);
10452 /* link the buffer into the block chain */
10453 lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1];
10454 lb->lb_magic = L2ARC_LOG_BLK_MAGIC;
10457 * l2arc_log_blk_commit() may be called multiple times during a single
10458 * l2arc_write_buffers() call. Save the allocated abd buffers in a list
10459 * so we can free them in l2arc_write_done() later on.
10461 list_insert_tail(&cb->l2wcb_abd_list, abd_buf);
10463 /* try to compress the buffer */
10464 psize = zio_compress_data(ZIO_COMPRESS_LZ4,
10465 abd_buf->abd, (void **) &tmpbuf, sizeof (*lb), 0);
10467 /* a log block is never entirely zero */
10468 ASSERT(psize != 0);
10469 asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
10470 ASSERT(asize <= sizeof (*lb));
10473 * Update the start log block pointer in the device header to point
10474 * to the log block we're about to write.
10476 l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0];
10477 l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand;
10478 l2dhdr->dh_start_lbps[0].lbp_payload_asize =
10479 dev->l2ad_log_blk_payload_asize;
10480 l2dhdr->dh_start_lbps[0].lbp_payload_start =
10481 dev->l2ad_log_blk_payload_start;
10483 (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb));
10485 (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize);
10486 L2BLK_SET_CHECKSUM(
10487 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10488 ZIO_CHECKSUM_FLETCHER_4);
10489 if (asize < sizeof (*lb)) {
10490 /* compression succeeded */
10491 memset(tmpbuf + psize, 0, asize - psize);
10492 L2BLK_SET_COMPRESS(
10493 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10496 /* compression failed */
10497 memcpy(tmpbuf, lb, sizeof (*lb));
10498 L2BLK_SET_COMPRESS(
10499 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10503 /* checksum what we're about to write */
10504 fletcher_4_native(tmpbuf, asize, NULL,
10505 &l2dhdr->dh_start_lbps[0].lbp_cksum);
10507 abd_free(abd_buf->abd);
10509 /* perform the write itself */
10510 abd_buf->abd = abd_get_from_buf(tmpbuf, sizeof (*lb));
10511 abd_take_ownership_of_buf(abd_buf->abd, B_TRUE);
10512 wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand,
10513 asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10514 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
10515 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
10516 (void) zio_nowait(wzio);
10518 dev->l2ad_hand += asize;
10520 * Include the committed log block's pointer in the list of pointers
10521 * to log blocks present in the L2ARC device.
10523 memcpy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[0],
10524 sizeof (l2arc_log_blkptr_t));
10525 mutex_enter(&dev->l2ad_mtx);
10526 list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf);
10527 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10528 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10529 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10530 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10531 mutex_exit(&dev->l2ad_mtx);
10532 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10534 /* bump the kstats */
10535 ARCSTAT_INCR(arcstat_l2_write_bytes, asize);
10536 ARCSTAT_BUMP(arcstat_l2_log_blk_writes);
10537 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize);
10538 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio,
10539 dev->l2ad_log_blk_payload_asize / asize);
10541 /* start a new log block */
10542 dev->l2ad_log_ent_idx = 0;
10543 dev->l2ad_log_blk_payload_asize = 0;
10544 dev->l2ad_log_blk_payload_start = 0;
10550 * Validates an L2ARC log block address to make sure that it can be read
10551 * from the provided L2ARC device.
10554 l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp)
10556 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10557 uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10558 uint64_t end = lbp->lbp_daddr + asize - 1;
10559 uint64_t start = lbp->lbp_payload_start;
10560 boolean_t evicted = B_FALSE;
10563 * A log block is valid if all of the following conditions are true:
10564 * - it fits entirely (including its payload) between l2ad_start and
10566 * - it has a valid size
10567 * - neither the log block itself nor part of its payload was evicted
10568 * by l2arc_evict():
10570 * l2ad_hand l2ad_evict
10575 * l2ad_start ============================================ l2ad_end
10576 * --------------------------||||
10583 l2arc_range_check_overlap(start, end, dev->l2ad_hand) ||
10584 l2arc_range_check_overlap(start, end, dev->l2ad_evict) ||
10585 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) ||
10586 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end);
10588 return (start >= dev->l2ad_start && end <= dev->l2ad_end &&
10589 asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) &&
10590 (!evicted || dev->l2ad_first));
10594 * Inserts ARC buffer header `hdr' into the current L2ARC log block on
10595 * the device. The buffer being inserted must be present in L2ARC.
10596 * Returns B_TRUE if the L2ARC log block is full and needs to be committed
10597 * to L2ARC, or B_FALSE if it still has room for more ARC buffers.
10600 l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr)
10602 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10603 l2arc_log_ent_phys_t *le;
10605 if (dev->l2ad_log_entries == 0)
10608 int index = dev->l2ad_log_ent_idx++;
10610 ASSERT3S(index, <, dev->l2ad_log_entries);
10611 ASSERT(HDR_HAS_L2HDR(hdr));
10613 le = &lb->lb_entries[index];
10614 memset(le, 0, sizeof (*le));
10615 le->le_dva = hdr->b_dva;
10616 le->le_birth = hdr->b_birth;
10617 le->le_daddr = hdr->b_l2hdr.b_daddr;
10619 dev->l2ad_log_blk_payload_start = le->le_daddr;
10620 L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr));
10621 L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr));
10622 L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr));
10623 le->le_complevel = hdr->b_complevel;
10624 L2BLK_SET_TYPE((le)->le_prop, hdr->b_type);
10625 L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr)));
10626 L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr)));
10627 L2BLK_SET_STATE((le)->le_prop, hdr->b_l1hdr.b_state->arcs_state);
10629 dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev,
10630 HDR_GET_PSIZE(hdr));
10632 return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries);
10636 * Checks whether a given L2ARC device address sits in a time-sequential
10637 * range. The trick here is that the L2ARC is a rotary buffer, so we can't
10638 * just do a range comparison, we need to handle the situation in which the
10639 * range wraps around the end of the L2ARC device. Arguments:
10640 * bottom -- Lower end of the range to check (written to earlier).
10641 * top -- Upper end of the range to check (written to later).
10642 * check -- The address for which we want to determine if it sits in
10643 * between the top and bottom.
10645 * The 3-way conditional below represents the following cases:
10647 * bottom < top : Sequentially ordered case:
10648 * <check>--------+-------------------+
10649 * | (overlap here?) |
10651 * |---------------<bottom>============<top>--------------|
10653 * bottom > top: Looped-around case:
10654 * <check>--------+------------------+
10655 * | (overlap here?) |
10657 * |===============<top>---------------<bottom>===========|
10660 * +---------------+---------<check>
10662 * top == bottom : Just a single address comparison.
10665 l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check)
10668 return (bottom <= check && check <= top);
10669 else if (bottom > top)
10670 return (check <= top || bottom <= check);
10672 return (check == top);
10675 EXPORT_SYMBOL(arc_buf_size);
10676 EXPORT_SYMBOL(arc_write);
10677 EXPORT_SYMBOL(arc_read);
10678 EXPORT_SYMBOL(arc_buf_info);
10679 EXPORT_SYMBOL(arc_getbuf_func);
10680 EXPORT_SYMBOL(arc_add_prune_callback);
10681 EXPORT_SYMBOL(arc_remove_prune_callback);
10683 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min, param_set_arc_min,
10684 spl_param_get_u64, ZMOD_RW, "Minimum ARC size in bytes");
10686 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, max, param_set_arc_max,
10687 spl_param_get_u64, ZMOD_RW, "Maximum ARC size in bytes");
10689 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_balance, UINT, ZMOD_RW,
10690 "Balance between metadata and data on ghost hits.");
10692 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, grow_retry, param_set_arc_int,
10693 param_get_uint, ZMOD_RW, "Seconds before growing ARC size");
10695 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, shrink_shift, param_set_arc_int,
10696 param_get_uint, ZMOD_RW, "log2(fraction of ARC to reclaim)");
10698 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, pc_percent, UINT, ZMOD_RW,
10699 "Percent of pagecache to reclaim ARC to");
10701 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, average_blocksize, UINT, ZMOD_RD,
10702 "Target average block size");
10704 ZFS_MODULE_PARAM(zfs, zfs_, compressed_arc_enabled, INT, ZMOD_RW,
10705 "Disable compressed ARC buffers");
10707 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prefetch_ms, param_set_arc_int,
10708 param_get_uint, ZMOD_RW, "Min life of prefetch block in ms");
10710 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prescient_prefetch_ms,
10711 param_set_arc_int, param_get_uint, ZMOD_RW,
10712 "Min life of prescient prefetched block in ms");
10714 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_max, U64, ZMOD_RW,
10715 "Max write bytes per interval");
10717 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_boost, U64, ZMOD_RW,
10718 "Extra write bytes during device warmup");
10720 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom, U64, ZMOD_RW,
10721 "Number of max device writes to precache");
10723 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom_boost, U64, ZMOD_RW,
10724 "Compressed l2arc_headroom multiplier");
10726 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, trim_ahead, U64, ZMOD_RW,
10727 "TRIM ahead L2ARC write size multiplier");
10729 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_secs, U64, ZMOD_RW,
10730 "Seconds between L2ARC writing");
10732 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_min_ms, U64, ZMOD_RW,
10733 "Min feed interval in milliseconds");
10735 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, noprefetch, INT, ZMOD_RW,
10736 "Skip caching prefetched buffers");
10738 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_again, INT, ZMOD_RW,
10739 "Turbo L2ARC warmup");
10741 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, norw, INT, ZMOD_RW,
10742 "No reads during writes");
10744 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, meta_percent, UINT, ZMOD_RW,
10745 "Percent of ARC size allowed for L2ARC-only headers");
10747 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_enabled, INT, ZMOD_RW,
10748 "Rebuild the L2ARC when importing a pool");
10750 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_blocks_min_l2size, U64, ZMOD_RW,
10751 "Min size in bytes to write rebuild log blocks in L2ARC");
10753 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, mfuonly, INT, ZMOD_RW,
10754 "Cache only MFU data from ARC into L2ARC");
10756 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, exclude_special, INT, ZMOD_RW,
10757 "Exclude dbufs on special vdevs from being cached to L2ARC if set.");
10759 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, lotsfree_percent, param_set_arc_int,
10760 param_get_uint, ZMOD_RW, "System free memory I/O throttle in bytes");
10762 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, sys_free, param_set_arc_u64,
10763 spl_param_get_u64, ZMOD_RW, "System free memory target size in bytes");
10765 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit, param_set_arc_u64,
10766 spl_param_get_u64, ZMOD_RW, "Minimum bytes of dnodes in ARC");
10768 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit_percent,
10769 param_set_arc_int, param_get_uint, ZMOD_RW,
10770 "Percent of ARC meta buffers for dnodes");
10772 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, dnode_reduce_percent, UINT, ZMOD_RW,
10773 "Percentage of excess dnodes to try to unpin");
10775 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, eviction_pct, UINT, ZMOD_RW,
10776 "When full, ARC allocation waits for eviction of this % of alloc size");
10778 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, evict_batch_limit, UINT, ZMOD_RW,
10779 "The number of headers to evict per sublist before moving to the next");
10781 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, prune_task_threads, INT, ZMOD_RW,
10782 "Number of arc_prune threads");