4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2018, Joyent, Inc.
24 * Copyright (c) 2011, 2020, Delphix. All rights reserved.
25 * Copyright (c) 2014, Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2017, Nexenta Systems, Inc. All rights reserved.
27 * Copyright (c) 2019, loli10K <ezomori.nozomu@gmail.com>. All rights reserved.
28 * Copyright (c) 2020, George Amanakis. All rights reserved.
29 * Copyright (c) 2019, Klara Inc.
30 * Copyright (c) 2019, Allan Jude
31 * Copyright (c) 2020, The FreeBSD Foundation [1]
33 * [1] Portions of this software were developed by Allan Jude
34 * under sponsorship from the FreeBSD Foundation.
38 * DVA-based Adjustable Replacement Cache
40 * While much of the theory of operation used here is
41 * based on the self-tuning, low overhead replacement cache
42 * presented by Megiddo and Modha at FAST 2003, there are some
43 * significant differences:
45 * 1. The Megiddo and Modha model assumes any page is evictable.
46 * Pages in its cache cannot be "locked" into memory. This makes
47 * the eviction algorithm simple: evict the last page in the list.
48 * This also make the performance characteristics easy to reason
49 * about. Our cache is not so simple. At any given moment, some
50 * subset of the blocks in the cache are un-evictable because we
51 * have handed out a reference to them. Blocks are only evictable
52 * when there are no external references active. This makes
53 * eviction far more problematic: we choose to evict the evictable
54 * blocks that are the "lowest" in the list.
56 * There are times when it is not possible to evict the requested
57 * space. In these circumstances we are unable to adjust the cache
58 * size. To prevent the cache growing unbounded at these times we
59 * implement a "cache throttle" that slows the flow of new data
60 * into the cache until we can make space available.
62 * 2. The Megiddo and Modha model assumes a fixed cache size.
63 * Pages are evicted when the cache is full and there is a cache
64 * miss. Our model has a variable sized cache. It grows with
65 * high use, but also tries to react to memory pressure from the
66 * operating system: decreasing its size when system memory is
69 * 3. The Megiddo and Modha model assumes a fixed page size. All
70 * elements of the cache are therefore exactly the same size. So
71 * when adjusting the cache size following a cache miss, its simply
72 * a matter of choosing a single page to evict. In our model, we
73 * have variable sized cache blocks (ranging from 512 bytes to
74 * 128K bytes). We therefore choose a set of blocks to evict to make
75 * space for a cache miss that approximates as closely as possible
76 * the space used by the new block.
78 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
79 * by N. Megiddo & D. Modha, FAST 2003
85 * A new reference to a cache buffer can be obtained in two
86 * ways: 1) via a hash table lookup using the DVA as a key,
87 * or 2) via one of the ARC lists. The arc_read() interface
88 * uses method 1, while the internal ARC algorithms for
89 * adjusting the cache use method 2. We therefore provide two
90 * types of locks: 1) the hash table lock array, and 2) the
93 * Buffers do not have their own mutexes, rather they rely on the
94 * hash table mutexes for the bulk of their protection (i.e. most
95 * fields in the arc_buf_hdr_t are protected by these mutexes).
97 * buf_hash_find() returns the appropriate mutex (held) when it
98 * locates the requested buffer in the hash table. It returns
99 * NULL for the mutex if the buffer was not in the table.
101 * buf_hash_remove() expects the appropriate hash mutex to be
102 * already held before it is invoked.
104 * Each ARC state also has a mutex which is used to protect the
105 * buffer list associated with the state. When attempting to
106 * obtain a hash table lock while holding an ARC list lock you
107 * must use: mutex_tryenter() to avoid deadlock. Also note that
108 * the active state mutex must be held before the ghost state mutex.
110 * It as also possible to register a callback which is run when the
111 * arc_meta_limit is reached and no buffers can be safely evicted. In
112 * this case the arc user should drop a reference on some arc buffers so
113 * they can be reclaimed and the arc_meta_limit honored. For example,
114 * when using the ZPL each dentry holds a references on a znode. These
115 * dentries must be pruned before the arc buffer holding the znode can
118 * Note that the majority of the performance stats are manipulated
119 * with atomic operations.
121 * The L2ARC uses the l2ad_mtx on each vdev for the following:
123 * - L2ARC buflist creation
124 * - L2ARC buflist eviction
125 * - L2ARC write completion, which walks L2ARC buflists
126 * - ARC header destruction, as it removes from L2ARC buflists
127 * - ARC header release, as it removes from L2ARC buflists
133 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
134 * This structure can point either to a block that is still in the cache or to
135 * one that is only accessible in an L2 ARC device, or it can provide
136 * information about a block that was recently evicted. If a block is
137 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
138 * information to retrieve it from the L2ARC device. This information is
139 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
140 * that is in this state cannot access the data directly.
142 * Blocks that are actively being referenced or have not been evicted
143 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
144 * the arc_buf_hdr_t that will point to the data block in memory. A block can
145 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
146 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
147 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
149 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
150 * ability to store the physical data (b_pabd) associated with the DVA of the
151 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
152 * it will match its on-disk compression characteristics. This behavior can be
153 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
154 * compressed ARC functionality is disabled, the b_pabd will point to an
155 * uncompressed version of the on-disk data.
157 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
158 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
159 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
160 * consumer. The ARC will provide references to this data and will keep it
161 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
162 * data block and will evict any arc_buf_t that is no longer referenced. The
163 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
164 * "overhead_size" kstat.
166 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
167 * compressed form. The typical case is that consumers will want uncompressed
168 * data, and when that happens a new data buffer is allocated where the data is
169 * decompressed for them to use. Currently the only consumer who wants
170 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
171 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
172 * with the arc_buf_hdr_t.
174 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
175 * first one is owned by a compressed send consumer (and therefore references
176 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
177 * used by any other consumer (and has its own uncompressed copy of the data
192 * | b_buf +------------>+-----------+ arc_buf_t
193 * | b_pabd +-+ |b_next +---->+-----------+
194 * +-----------+ | |-----------| |b_next +-->NULL
195 * | |b_comp = T | +-----------+
196 * | |b_data +-+ |b_comp = F |
197 * | +-----------+ | |b_data +-+
198 * +->+------+ | +-----------+ |
200 * data | |<--------------+ | uncompressed
201 * +------+ compressed, | data
202 * shared +-->+------+
207 * When a consumer reads a block, the ARC must first look to see if the
208 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
209 * arc_buf_t and either copies uncompressed data into a new data buffer from an
210 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
211 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
212 * hdr is compressed and the desired compression characteristics of the
213 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
214 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
215 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
216 * be anywhere in the hdr's list.
218 * The diagram below shows an example of an uncompressed ARC hdr that is
219 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
220 * the last element in the buf list):
232 * | | arc_buf_t (shared)
233 * | b_buf +------------>+---------+ arc_buf_t
234 * | | |b_next +---->+---------+
235 * | b_pabd +-+ |---------| |b_next +-->NULL
236 * +-----------+ | | | +---------+
238 * | +---------+ | |b_data +-+
239 * +->+------+ | +---------+ |
241 * uncompressed | | | |
244 * | uncompressed | | |
247 * +---------------------------------+
249 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
250 * since the physical block is about to be rewritten. The new data contents
251 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
252 * it may compress the data before writing it to disk. The ARC will be called
253 * with the transformed data and will bcopy the transformed on-disk block into
254 * a newly allocated b_pabd. Writes are always done into buffers which have
255 * either been loaned (and hence are new and don't have other readers) or
256 * buffers which have been released (and hence have their own hdr, if there
257 * were originally other readers of the buf's original hdr). This ensures that
258 * the ARC only needs to update a single buf and its hdr after a write occurs.
260 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
261 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
262 * that when compressed ARC is enabled that the L2ARC blocks are identical
263 * to the on-disk block in the main data pool. This provides a significant
264 * advantage since the ARC can leverage the bp's checksum when reading from the
265 * L2ARC to determine if the contents are valid. However, if the compressed
266 * ARC is disabled, then the L2ARC's block must be transformed to look
267 * like the physical block in the main data pool before comparing the
268 * checksum and determining its validity.
270 * The L1ARC has a slightly different system for storing encrypted data.
271 * Raw (encrypted + possibly compressed) data has a few subtle differences from
272 * data that is just compressed. The biggest difference is that it is not
273 * possible to decrypt encrypted data (or vice-versa) if the keys aren't loaded.
274 * The other difference is that encryption cannot be treated as a suggestion.
275 * If a caller would prefer compressed data, but they actually wind up with
276 * uncompressed data the worst thing that could happen is there might be a
277 * performance hit. If the caller requests encrypted data, however, we must be
278 * sure they actually get it or else secret information could be leaked. Raw
279 * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore,
280 * may have both an encrypted version and a decrypted version of its data at
281 * once. When a caller needs a raw arc_buf_t, it is allocated and the data is
282 * copied out of this header. To avoid complications with b_pabd, raw buffers
288 #include <sys/spa_impl.h>
289 #include <sys/zio_compress.h>
290 #include <sys/zio_checksum.h>
291 #include <sys/zfs_context.h>
293 #include <sys/zfs_refcount.h>
294 #include <sys/vdev.h>
295 #include <sys/vdev_impl.h>
296 #include <sys/dsl_pool.h>
297 #include <sys/multilist.h>
300 #include <sys/fm/fs/zfs.h>
301 #include <sys/callb.h>
302 #include <sys/kstat.h>
303 #include <sys/zthr.h>
304 #include <zfs_fletcher.h>
305 #include <sys/arc_impl.h>
306 #include <sys/trace_zfs.h>
307 #include <sys/aggsum.h>
308 #include <sys/wmsum.h>
309 #include <cityhash.h>
310 #include <sys/vdev_trim.h>
311 #include <sys/zfs_racct.h>
312 #include <sys/zstd/zstd.h>
315 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
316 boolean_t arc_watch = B_FALSE;
320 * This thread's job is to keep enough free memory in the system, by
321 * calling arc_kmem_reap_soon() plus arc_reduce_target_size(), which improves
322 * arc_available_memory().
324 static zthr_t *arc_reap_zthr;
327 * This thread's job is to keep arc_size under arc_c, by calling
328 * arc_evict(), which improves arc_is_overflowing().
330 static zthr_t *arc_evict_zthr;
331 static arc_buf_hdr_t **arc_state_evict_markers;
332 static int arc_state_evict_marker_count;
334 static kmutex_t arc_evict_lock;
335 static boolean_t arc_evict_needed = B_FALSE;
338 * Count of bytes evicted since boot.
340 static uint64_t arc_evict_count;
343 * List of arc_evict_waiter_t's, representing threads waiting for the
344 * arc_evict_count to reach specific values.
346 static list_t arc_evict_waiters;
349 * When arc_is_overflowing(), arc_get_data_impl() waits for this percent of
350 * the requested amount of data to be evicted. For example, by default for
351 * every 2KB that's evicted, 1KB of it may be "reused" by a new allocation.
352 * Since this is above 100%, it ensures that progress is made towards getting
353 * arc_size under arc_c. Since this is finite, it ensures that allocations
354 * can still happen, even during the potentially long time that arc_size is
357 int 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 int zfs_arc_evict_batch_limit = 10;
368 /* number of seconds before growing cache again */
369 int arc_grow_retry = 5;
372 * Minimum time between calls to arc_kmem_reap_soon().
374 int arc_kmem_cache_reap_retry_ms = 1000;
376 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
377 int zfs_arc_overflow_shift = 8;
379 /* shift of arc_c for calculating both min and max arc_p */
380 int arc_p_min_shift = 4;
382 /* log2(fraction of arc to reclaim) */
383 int arc_shrink_shift = 7;
385 /* percent of pagecache to reclaim arc to */
387 uint_t zfs_arc_pc_percent = 0;
391 * log2(fraction of ARC which must be free to allow growing).
392 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
393 * when reading a new block into the ARC, we will evict an equal-sized block
396 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
397 * we will still not allow it to grow.
399 int arc_no_grow_shift = 5;
403 * minimum lifespan of a prefetch block in clock ticks
404 * (initialized in arc_init())
406 static int arc_min_prefetch_ms;
407 static int arc_min_prescient_prefetch_ms;
410 * If this percent of memory is free, don't throttle.
412 int arc_lotsfree_percent = 10;
415 * The arc has filled available memory and has now warmed up.
420 * These tunables are for performance analysis.
422 unsigned long zfs_arc_max = 0;
423 unsigned long zfs_arc_min = 0;
424 unsigned long zfs_arc_meta_limit = 0;
425 unsigned long zfs_arc_meta_min = 0;
426 unsigned long zfs_arc_dnode_limit = 0;
427 unsigned long zfs_arc_dnode_reduce_percent = 10;
428 int zfs_arc_grow_retry = 0;
429 int zfs_arc_shrink_shift = 0;
430 int zfs_arc_p_min_shift = 0;
431 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
434 * ARC dirty data constraints for arc_tempreserve_space() throttle.
436 unsigned long zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */
437 unsigned long zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */
438 unsigned long zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */
441 * Enable or disable compressed arc buffers.
443 int zfs_compressed_arc_enabled = B_TRUE;
446 * ARC will evict meta buffers that exceed arc_meta_limit. This
447 * tunable make arc_meta_limit adjustable for different workloads.
449 unsigned long zfs_arc_meta_limit_percent = 75;
452 * Percentage that can be consumed by dnodes of ARC meta buffers.
454 unsigned long zfs_arc_dnode_limit_percent = 10;
457 * These tunables are Linux specific
459 unsigned long zfs_arc_sys_free = 0;
460 int zfs_arc_min_prefetch_ms = 0;
461 int zfs_arc_min_prescient_prefetch_ms = 0;
462 int zfs_arc_p_dampener_disable = 1;
463 int zfs_arc_meta_prune = 10000;
464 int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED;
465 int zfs_arc_meta_adjust_restarts = 4096;
466 int zfs_arc_lotsfree_percent = 10;
469 arc_state_t ARC_anon;
471 arc_state_t ARC_mru_ghost;
473 arc_state_t ARC_mfu_ghost;
474 arc_state_t ARC_l2c_only;
476 arc_stats_t arc_stats = {
477 { "hits", KSTAT_DATA_UINT64 },
478 { "misses", KSTAT_DATA_UINT64 },
479 { "demand_data_hits", KSTAT_DATA_UINT64 },
480 { "demand_data_misses", KSTAT_DATA_UINT64 },
481 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
482 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
483 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
484 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
485 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
486 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
487 { "mru_hits", KSTAT_DATA_UINT64 },
488 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
489 { "mfu_hits", KSTAT_DATA_UINT64 },
490 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
491 { "deleted", KSTAT_DATA_UINT64 },
492 { "mutex_miss", KSTAT_DATA_UINT64 },
493 { "access_skip", KSTAT_DATA_UINT64 },
494 { "evict_skip", KSTAT_DATA_UINT64 },
495 { "evict_not_enough", KSTAT_DATA_UINT64 },
496 { "evict_l2_cached", KSTAT_DATA_UINT64 },
497 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
498 { "evict_l2_eligible_mfu", KSTAT_DATA_UINT64 },
499 { "evict_l2_eligible_mru", KSTAT_DATA_UINT64 },
500 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
501 { "evict_l2_skip", KSTAT_DATA_UINT64 },
502 { "hash_elements", KSTAT_DATA_UINT64 },
503 { "hash_elements_max", KSTAT_DATA_UINT64 },
504 { "hash_collisions", KSTAT_DATA_UINT64 },
505 { "hash_chains", KSTAT_DATA_UINT64 },
506 { "hash_chain_max", KSTAT_DATA_UINT64 },
507 { "p", KSTAT_DATA_UINT64 },
508 { "c", KSTAT_DATA_UINT64 },
509 { "c_min", KSTAT_DATA_UINT64 },
510 { "c_max", KSTAT_DATA_UINT64 },
511 { "size", KSTAT_DATA_UINT64 },
512 { "compressed_size", KSTAT_DATA_UINT64 },
513 { "uncompressed_size", KSTAT_DATA_UINT64 },
514 { "overhead_size", KSTAT_DATA_UINT64 },
515 { "hdr_size", KSTAT_DATA_UINT64 },
516 { "data_size", KSTAT_DATA_UINT64 },
517 { "metadata_size", KSTAT_DATA_UINT64 },
518 { "dbuf_size", KSTAT_DATA_UINT64 },
519 { "dnode_size", KSTAT_DATA_UINT64 },
520 { "bonus_size", KSTAT_DATA_UINT64 },
521 #if defined(COMPAT_FREEBSD11)
522 { "other_size", KSTAT_DATA_UINT64 },
524 { "anon_size", KSTAT_DATA_UINT64 },
525 { "anon_evictable_data", KSTAT_DATA_UINT64 },
526 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
527 { "mru_size", KSTAT_DATA_UINT64 },
528 { "mru_evictable_data", KSTAT_DATA_UINT64 },
529 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
530 { "mru_ghost_size", KSTAT_DATA_UINT64 },
531 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
532 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
533 { "mfu_size", KSTAT_DATA_UINT64 },
534 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
535 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
536 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
537 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
538 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
539 { "l2_hits", KSTAT_DATA_UINT64 },
540 { "l2_misses", KSTAT_DATA_UINT64 },
541 { "l2_prefetch_asize", KSTAT_DATA_UINT64 },
542 { "l2_mru_asize", KSTAT_DATA_UINT64 },
543 { "l2_mfu_asize", KSTAT_DATA_UINT64 },
544 { "l2_bufc_data_asize", KSTAT_DATA_UINT64 },
545 { "l2_bufc_metadata_asize", KSTAT_DATA_UINT64 },
546 { "l2_feeds", KSTAT_DATA_UINT64 },
547 { "l2_rw_clash", KSTAT_DATA_UINT64 },
548 { "l2_read_bytes", KSTAT_DATA_UINT64 },
549 { "l2_write_bytes", KSTAT_DATA_UINT64 },
550 { "l2_writes_sent", KSTAT_DATA_UINT64 },
551 { "l2_writes_done", KSTAT_DATA_UINT64 },
552 { "l2_writes_error", KSTAT_DATA_UINT64 },
553 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
554 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
555 { "l2_evict_reading", KSTAT_DATA_UINT64 },
556 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
557 { "l2_free_on_write", KSTAT_DATA_UINT64 },
558 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
559 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
560 { "l2_io_error", KSTAT_DATA_UINT64 },
561 { "l2_size", KSTAT_DATA_UINT64 },
562 { "l2_asize", KSTAT_DATA_UINT64 },
563 { "l2_hdr_size", KSTAT_DATA_UINT64 },
564 { "l2_log_blk_writes", KSTAT_DATA_UINT64 },
565 { "l2_log_blk_avg_asize", KSTAT_DATA_UINT64 },
566 { "l2_log_blk_asize", KSTAT_DATA_UINT64 },
567 { "l2_log_blk_count", KSTAT_DATA_UINT64 },
568 { "l2_data_to_meta_ratio", KSTAT_DATA_UINT64 },
569 { "l2_rebuild_success", KSTAT_DATA_UINT64 },
570 { "l2_rebuild_unsupported", KSTAT_DATA_UINT64 },
571 { "l2_rebuild_io_errors", KSTAT_DATA_UINT64 },
572 { "l2_rebuild_dh_errors", KSTAT_DATA_UINT64 },
573 { "l2_rebuild_cksum_lb_errors", KSTAT_DATA_UINT64 },
574 { "l2_rebuild_lowmem", KSTAT_DATA_UINT64 },
575 { "l2_rebuild_size", KSTAT_DATA_UINT64 },
576 { "l2_rebuild_asize", KSTAT_DATA_UINT64 },
577 { "l2_rebuild_bufs", KSTAT_DATA_UINT64 },
578 { "l2_rebuild_bufs_precached", KSTAT_DATA_UINT64 },
579 { "l2_rebuild_log_blks", KSTAT_DATA_UINT64 },
580 { "memory_throttle_count", KSTAT_DATA_UINT64 },
581 { "memory_direct_count", KSTAT_DATA_UINT64 },
582 { "memory_indirect_count", KSTAT_DATA_UINT64 },
583 { "memory_all_bytes", KSTAT_DATA_UINT64 },
584 { "memory_free_bytes", KSTAT_DATA_UINT64 },
585 { "memory_available_bytes", KSTAT_DATA_INT64 },
586 { "arc_no_grow", KSTAT_DATA_UINT64 },
587 { "arc_tempreserve", KSTAT_DATA_UINT64 },
588 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
589 { "arc_prune", KSTAT_DATA_UINT64 },
590 { "arc_meta_used", KSTAT_DATA_UINT64 },
591 { "arc_meta_limit", KSTAT_DATA_UINT64 },
592 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
593 { "arc_meta_max", KSTAT_DATA_UINT64 },
594 { "arc_meta_min", KSTAT_DATA_UINT64 },
595 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
596 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
597 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
598 { "arc_need_free", KSTAT_DATA_UINT64 },
599 { "arc_sys_free", KSTAT_DATA_UINT64 },
600 { "arc_raw_size", KSTAT_DATA_UINT64 },
601 { "cached_only_in_progress", KSTAT_DATA_UINT64 },
602 { "abd_chunk_waste_size", KSTAT_DATA_UINT64 },
607 #define ARCSTAT_MAX(stat, val) { \
609 while ((val) > (m = arc_stats.stat.value.ui64) && \
610 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
615 * We define a macro to allow ARC hits/misses to be easily broken down by
616 * two separate conditions, giving a total of four different subtypes for
617 * each of hits and misses (so eight statistics total).
619 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
622 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
624 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
628 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
630 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
635 * This macro allows us to use kstats as floating averages. Each time we
636 * update this kstat, we first factor it and the update value by
637 * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall
638 * average. This macro assumes that integer loads and stores are atomic, but
639 * is not safe for multiple writers updating the kstat in parallel (only the
640 * last writer's update will remain).
642 #define ARCSTAT_F_AVG_FACTOR 3
643 #define ARCSTAT_F_AVG(stat, value) \
645 uint64_t x = ARCSTAT(stat); \
646 x = x - x / ARCSTAT_F_AVG_FACTOR + \
647 (value) / ARCSTAT_F_AVG_FACTOR; \
655 * There are several ARC variables that are critical to export as kstats --
656 * but we don't want to have to grovel around in the kstat whenever we wish to
657 * manipulate them. For these variables, we therefore define them to be in
658 * terms of the statistic variable. This assures that we are not introducing
659 * the possibility of inconsistency by having shadow copies of the variables,
660 * while still allowing the code to be readable.
662 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
663 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
664 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
665 /* max size for dnodes */
666 #define arc_dnode_size_limit ARCSTAT(arcstat_dnode_limit)
667 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
668 #define arc_need_free ARCSTAT(arcstat_need_free) /* waiting to be evicted */
670 hrtime_t arc_growtime;
671 list_t arc_prune_list;
672 kmutex_t arc_prune_mtx;
673 taskq_t *arc_prune_taskq;
675 #define GHOST_STATE(state) \
676 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
677 (state) == arc_l2c_only)
679 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
680 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
681 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
682 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
683 #define HDR_PRESCIENT_PREFETCH(hdr) \
684 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
685 #define HDR_COMPRESSION_ENABLED(hdr) \
686 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
688 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
689 #define HDR_L2_READING(hdr) \
690 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
691 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
692 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
693 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
694 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
695 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
696 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
697 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
699 #define HDR_ISTYPE_METADATA(hdr) \
700 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
701 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
703 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
704 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
705 #define HDR_HAS_RABD(hdr) \
706 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
707 (hdr)->b_crypt_hdr.b_rabd != NULL)
708 #define HDR_ENCRYPTED(hdr) \
709 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
710 #define HDR_AUTHENTICATED(hdr) \
711 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
713 /* For storing compression mode in b_flags */
714 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
716 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
717 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
718 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
719 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
721 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
722 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
723 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
724 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
730 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
731 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
732 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
735 * Hash table routines
738 #define BUF_LOCKS 2048
739 typedef struct buf_hash_table {
741 arc_buf_hdr_t **ht_table;
742 kmutex_t ht_locks[BUF_LOCKS] ____cacheline_aligned;
745 static buf_hash_table_t buf_hash_table;
747 #define BUF_HASH_INDEX(spa, dva, birth) \
748 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
749 #define BUF_HASH_LOCK(idx) (&buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
750 #define HDR_LOCK(hdr) \
751 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
753 uint64_t zfs_crc64_table[256];
759 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
760 #define L2ARC_HEADROOM 2 /* num of writes */
763 * If we discover during ARC scan any buffers to be compressed, we boost
764 * our headroom for the next scanning cycle by this percentage multiple.
766 #define L2ARC_HEADROOM_BOOST 200
767 #define L2ARC_FEED_SECS 1 /* caching interval secs */
768 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
771 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
772 * and each of the state has two types: data and metadata.
774 #define L2ARC_FEED_TYPES 4
776 /* L2ARC Performance Tunables */
777 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
778 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
779 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
780 unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
781 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
782 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
783 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
784 int l2arc_feed_again = B_TRUE; /* turbo warmup */
785 int l2arc_norw = B_FALSE; /* no reads during writes */
786 int l2arc_meta_percent = 33; /* limit on headers size */
791 static list_t L2ARC_dev_list; /* device list */
792 static list_t *l2arc_dev_list; /* device list pointer */
793 static kmutex_t l2arc_dev_mtx; /* device list mutex */
794 static l2arc_dev_t *l2arc_dev_last; /* last device used */
795 static list_t L2ARC_free_on_write; /* free after write buf list */
796 static list_t *l2arc_free_on_write; /* free after write list ptr */
797 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
798 static uint64_t l2arc_ndev; /* number of devices */
800 typedef struct l2arc_read_callback {
801 arc_buf_hdr_t *l2rcb_hdr; /* read header */
802 blkptr_t l2rcb_bp; /* original blkptr */
803 zbookmark_phys_t l2rcb_zb; /* original bookmark */
804 int l2rcb_flags; /* original flags */
805 abd_t *l2rcb_abd; /* temporary buffer */
806 } l2arc_read_callback_t;
808 typedef struct l2arc_data_free {
809 /* protected by l2arc_free_on_write_mtx */
812 arc_buf_contents_t l2df_type;
813 list_node_t l2df_list_node;
816 typedef enum arc_fill_flags {
817 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
818 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
819 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
820 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
821 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
824 typedef enum arc_ovf_level {
825 ARC_OVF_NONE, /* ARC within target size. */
826 ARC_OVF_SOME, /* ARC is slightly overflowed. */
827 ARC_OVF_SEVERE /* ARC is severely overflowed. */
830 static kmutex_t l2arc_feed_thr_lock;
831 static kcondvar_t l2arc_feed_thr_cv;
832 static uint8_t l2arc_thread_exit;
834 static kmutex_t l2arc_rebuild_thr_lock;
835 static kcondvar_t l2arc_rebuild_thr_cv;
837 enum arc_hdr_alloc_flags {
838 ARC_HDR_ALLOC_RDATA = 0x1,
839 ARC_HDR_DO_ADAPT = 0x2,
840 ARC_HDR_USE_RESERVE = 0x4,
844 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *, int);
845 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
846 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *, int);
847 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
848 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
849 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
850 static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t);
851 static void arc_hdr_alloc_abd(arc_buf_hdr_t *, int);
852 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
853 static void arc_buf_watch(arc_buf_t *);
855 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
856 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
857 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
858 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
860 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
861 static void l2arc_read_done(zio_t *);
862 static void l2arc_do_free_on_write(void);
863 static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
864 boolean_t state_only);
866 #define l2arc_hdr_arcstats_increment(hdr) \
867 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_FALSE)
868 #define l2arc_hdr_arcstats_decrement(hdr) \
869 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_FALSE)
870 #define l2arc_hdr_arcstats_increment_state(hdr) \
871 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_TRUE)
872 #define l2arc_hdr_arcstats_decrement_state(hdr) \
873 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_TRUE)
876 * l2arc_mfuonly : A ZFS module parameter that controls whether only MFU
877 * metadata and data are cached from ARC into L2ARC.
879 int l2arc_mfuonly = 0;
883 * l2arc_trim_ahead : A ZFS module parameter that controls how much ahead of
884 * the current write size (l2arc_write_max) we should TRIM if we
885 * have filled the device. It is defined as a percentage of the
886 * write size. If set to 100 we trim twice the space required to
887 * accommodate upcoming writes. A minimum of 64MB will be trimmed.
888 * It also enables TRIM of the whole L2ARC device upon creation or
889 * addition to an existing pool or if the header of the device is
890 * invalid upon importing a pool or onlining a cache device. The
891 * default is 0, which disables TRIM on L2ARC altogether as it can
892 * put significant stress on the underlying storage devices. This
893 * will vary depending of how well the specific device handles
896 unsigned long l2arc_trim_ahead = 0;
899 * Performance tuning of L2ARC persistence:
901 * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding
902 * an L2ARC device (either at pool import or later) will attempt
903 * to rebuild L2ARC buffer contents.
904 * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls
905 * whether log blocks are written to the L2ARC device. If the L2ARC
906 * device is less than 1GB, the amount of data l2arc_evict()
907 * evicts is significant compared to the amount of restored L2ARC
908 * data. In this case do not write log blocks in L2ARC in order
909 * not to waste space.
911 int l2arc_rebuild_enabled = B_TRUE;
912 unsigned long l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024;
914 /* L2ARC persistence rebuild control routines. */
915 void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen);
916 static void l2arc_dev_rebuild_thread(void *arg);
917 static int l2arc_rebuild(l2arc_dev_t *dev);
919 /* L2ARC persistence read I/O routines. */
920 static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
921 static int l2arc_log_blk_read(l2arc_dev_t *dev,
922 const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
923 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
924 zio_t *this_io, zio_t **next_io);
925 static zio_t *l2arc_log_blk_fetch(vdev_t *vd,
926 const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb);
927 static void l2arc_log_blk_fetch_abort(zio_t *zio);
929 /* L2ARC persistence block restoration routines. */
930 static void l2arc_log_blk_restore(l2arc_dev_t *dev,
931 const l2arc_log_blk_phys_t *lb, uint64_t lb_asize);
932 static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
935 /* L2ARC persistence write I/O routines. */
936 static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
937 l2arc_write_callback_t *cb);
939 /* L2ARC persistence auxiliary routines. */
940 boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
941 const l2arc_log_blkptr_t *lbp);
942 static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
943 const arc_buf_hdr_t *ab);
944 boolean_t l2arc_range_check_overlap(uint64_t bottom,
945 uint64_t top, uint64_t check);
946 static void l2arc_blk_fetch_done(zio_t *zio);
947 static inline uint64_t
948 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev);
951 * We use Cityhash for this. It's fast, and has good hash properties without
952 * requiring any large static buffers.
955 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
957 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
960 #define HDR_EMPTY(hdr) \
961 ((hdr)->b_dva.dva_word[0] == 0 && \
962 (hdr)->b_dva.dva_word[1] == 0)
964 #define HDR_EMPTY_OR_LOCKED(hdr) \
965 (HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr)))
967 #define HDR_EQUAL(spa, dva, birth, hdr) \
968 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
969 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
970 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
973 buf_discard_identity(arc_buf_hdr_t *hdr)
975 hdr->b_dva.dva_word[0] = 0;
976 hdr->b_dva.dva_word[1] = 0;
980 static arc_buf_hdr_t *
981 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
983 const dva_t *dva = BP_IDENTITY(bp);
984 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
985 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
986 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
989 mutex_enter(hash_lock);
990 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
991 hdr = hdr->b_hash_next) {
992 if (HDR_EQUAL(spa, dva, birth, hdr)) {
997 mutex_exit(hash_lock);
1003 * Insert an entry into the hash table. If there is already an element
1004 * equal to elem in the hash table, then the already existing element
1005 * will be returned and the new element will not be inserted.
1006 * Otherwise returns NULL.
1007 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1009 static arc_buf_hdr_t *
1010 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1012 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1013 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1014 arc_buf_hdr_t *fhdr;
1017 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1018 ASSERT(hdr->b_birth != 0);
1019 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1021 if (lockp != NULL) {
1023 mutex_enter(hash_lock);
1025 ASSERT(MUTEX_HELD(hash_lock));
1028 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1029 fhdr = fhdr->b_hash_next, i++) {
1030 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1034 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1035 buf_hash_table.ht_table[idx] = hdr;
1036 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1038 /* collect some hash table performance data */
1040 ARCSTAT_BUMP(arcstat_hash_collisions);
1042 ARCSTAT_BUMP(arcstat_hash_chains);
1044 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1046 uint64_t he = atomic_inc_64_nv(
1047 &arc_stats.arcstat_hash_elements.value.ui64);
1048 ARCSTAT_MAX(arcstat_hash_elements_max, he);
1054 buf_hash_remove(arc_buf_hdr_t *hdr)
1056 arc_buf_hdr_t *fhdr, **hdrp;
1057 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1059 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1060 ASSERT(HDR_IN_HASH_TABLE(hdr));
1062 hdrp = &buf_hash_table.ht_table[idx];
1063 while ((fhdr = *hdrp) != hdr) {
1064 ASSERT3P(fhdr, !=, NULL);
1065 hdrp = &fhdr->b_hash_next;
1067 *hdrp = hdr->b_hash_next;
1068 hdr->b_hash_next = NULL;
1069 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1071 /* collect some hash table performance data */
1072 atomic_dec_64(&arc_stats.arcstat_hash_elements.value.ui64);
1074 if (buf_hash_table.ht_table[idx] &&
1075 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1076 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1080 * Global data structures and functions for the buf kmem cache.
1083 static kmem_cache_t *hdr_full_cache;
1084 static kmem_cache_t *hdr_full_crypt_cache;
1085 static kmem_cache_t *hdr_l2only_cache;
1086 static kmem_cache_t *buf_cache;
1093 #if defined(_KERNEL)
1095 * Large allocations which do not require contiguous pages
1096 * should be using vmem_free() in the linux kernel\
1098 vmem_free(buf_hash_table.ht_table,
1099 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1101 kmem_free(buf_hash_table.ht_table,
1102 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1104 for (i = 0; i < BUF_LOCKS; i++)
1105 mutex_destroy(BUF_HASH_LOCK(i));
1106 kmem_cache_destroy(hdr_full_cache);
1107 kmem_cache_destroy(hdr_full_crypt_cache);
1108 kmem_cache_destroy(hdr_l2only_cache);
1109 kmem_cache_destroy(buf_cache);
1113 * Constructor callback - called when the cache is empty
1114 * and a new buf is requested.
1118 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1120 arc_buf_hdr_t *hdr = vbuf;
1122 bzero(hdr, HDR_FULL_SIZE);
1123 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1124 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1125 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1126 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1127 list_link_init(&hdr->b_l1hdr.b_arc_node);
1128 list_link_init(&hdr->b_l2hdr.b_l2node);
1129 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1130 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1137 hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
1139 arc_buf_hdr_t *hdr = vbuf;
1141 hdr_full_cons(vbuf, unused, kmflag);
1142 bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr));
1143 arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1150 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1152 arc_buf_hdr_t *hdr = vbuf;
1154 bzero(hdr, HDR_L2ONLY_SIZE);
1155 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1162 buf_cons(void *vbuf, void *unused, int kmflag)
1164 arc_buf_t *buf = vbuf;
1166 bzero(buf, sizeof (arc_buf_t));
1167 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1168 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1174 * Destructor callback - called when a cached buf is
1175 * no longer required.
1179 hdr_full_dest(void *vbuf, void *unused)
1181 arc_buf_hdr_t *hdr = vbuf;
1183 ASSERT(HDR_EMPTY(hdr));
1184 cv_destroy(&hdr->b_l1hdr.b_cv);
1185 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1186 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1187 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1188 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1193 hdr_full_crypt_dest(void *vbuf, void *unused)
1195 arc_buf_hdr_t *hdr = vbuf;
1197 hdr_full_dest(vbuf, unused);
1198 arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1203 hdr_l2only_dest(void *vbuf, void *unused)
1205 arc_buf_hdr_t *hdr __maybe_unused = vbuf;
1207 ASSERT(HDR_EMPTY(hdr));
1208 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1213 buf_dest(void *vbuf, void *unused)
1215 arc_buf_t *buf = vbuf;
1217 mutex_destroy(&buf->b_evict_lock);
1218 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1224 uint64_t *ct = NULL;
1225 uint64_t hsize = 1ULL << 12;
1229 * The hash table is big enough to fill all of physical memory
1230 * with an average block size of zfs_arc_average_blocksize (default 8K).
1231 * By default, the table will take up
1232 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1234 while (hsize * zfs_arc_average_blocksize < arc_all_memory())
1237 buf_hash_table.ht_mask = hsize - 1;
1238 #if defined(_KERNEL)
1240 * Large allocations which do not require contiguous pages
1241 * should be using vmem_alloc() in the linux kernel
1243 buf_hash_table.ht_table =
1244 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
1246 buf_hash_table.ht_table =
1247 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1249 if (buf_hash_table.ht_table == NULL) {
1250 ASSERT(hsize > (1ULL << 8));
1255 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1256 0, hdr_full_cons, hdr_full_dest, NULL, NULL, NULL, 0);
1257 hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
1258 HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
1259 NULL, NULL, NULL, 0);
1260 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1261 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, NULL,
1263 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1264 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1266 for (i = 0; i < 256; i++)
1267 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1268 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1270 for (i = 0; i < BUF_LOCKS; i++)
1271 mutex_init(BUF_HASH_LOCK(i), NULL, MUTEX_DEFAULT, NULL);
1274 #define ARC_MINTIME (hz>>4) /* 62 ms */
1277 * This is the size that the buf occupies in memory. If the buf is compressed,
1278 * it will correspond to the compressed size. You should use this method of
1279 * getting the buf size unless you explicitly need the logical size.
1282 arc_buf_size(arc_buf_t *buf)
1284 return (ARC_BUF_COMPRESSED(buf) ?
1285 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1289 arc_buf_lsize(arc_buf_t *buf)
1291 return (HDR_GET_LSIZE(buf->b_hdr));
1295 * This function will return B_TRUE if the buffer is encrypted in memory.
1296 * This buffer can be decrypted by calling arc_untransform().
1299 arc_is_encrypted(arc_buf_t *buf)
1301 return (ARC_BUF_ENCRYPTED(buf) != 0);
1305 * Returns B_TRUE if the buffer represents data that has not had its MAC
1309 arc_is_unauthenticated(arc_buf_t *buf)
1311 return (HDR_NOAUTH(buf->b_hdr) != 0);
1315 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1316 uint8_t *iv, uint8_t *mac)
1318 arc_buf_hdr_t *hdr = buf->b_hdr;
1320 ASSERT(HDR_PROTECTED(hdr));
1322 bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
1323 bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
1324 bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
1325 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1326 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1330 * Indicates how this buffer is compressed in memory. If it is not compressed
1331 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1332 * arc_untransform() as long as it is also unencrypted.
1335 arc_get_compression(arc_buf_t *buf)
1337 return (ARC_BUF_COMPRESSED(buf) ?
1338 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1342 * Return the compression algorithm used to store this data in the ARC. If ARC
1343 * compression is enabled or this is an encrypted block, this will be the same
1344 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1346 static inline enum zio_compress
1347 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1349 return (HDR_COMPRESSION_ENABLED(hdr) ?
1350 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1354 arc_get_complevel(arc_buf_t *buf)
1356 return (buf->b_hdr->b_complevel);
1359 static inline boolean_t
1360 arc_buf_is_shared(arc_buf_t *buf)
1362 boolean_t shared = (buf->b_data != NULL &&
1363 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1364 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1365 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1366 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1367 IMPLY(shared, ARC_BUF_SHARED(buf));
1368 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1371 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1372 * already being shared" requirement prevents us from doing that.
1379 * Free the checksum associated with this header. If there is no checksum, this
1383 arc_cksum_free(arc_buf_hdr_t *hdr)
1385 ASSERT(HDR_HAS_L1HDR(hdr));
1387 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1388 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1389 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1390 hdr->b_l1hdr.b_freeze_cksum = NULL;
1392 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1396 * Return true iff at least one of the bufs on hdr is not compressed.
1397 * Encrypted buffers count as compressed.
1400 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1402 ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr));
1404 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1405 if (!ARC_BUF_COMPRESSED(b)) {
1414 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1415 * matches the checksum that is stored in the hdr. If there is no checksum,
1416 * or if the buf is compressed, this is a no-op.
1419 arc_cksum_verify(arc_buf_t *buf)
1421 arc_buf_hdr_t *hdr = buf->b_hdr;
1424 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1427 if (ARC_BUF_COMPRESSED(buf))
1430 ASSERT(HDR_HAS_L1HDR(hdr));
1432 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1434 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1435 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1439 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1440 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1441 panic("buffer modified while frozen!");
1442 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1446 * This function makes the assumption that data stored in the L2ARC
1447 * will be transformed exactly as it is in the main pool. Because of
1448 * this we can verify the checksum against the reading process's bp.
1451 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1453 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1454 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1457 * Block pointers always store the checksum for the logical data.
1458 * If the block pointer has the gang bit set, then the checksum
1459 * it represents is for the reconstituted data and not for an
1460 * individual gang member. The zio pipeline, however, must be able to
1461 * determine the checksum of each of the gang constituents so it
1462 * treats the checksum comparison differently than what we need
1463 * for l2arc blocks. This prevents us from using the
1464 * zio_checksum_error() interface directly. Instead we must call the
1465 * zio_checksum_error_impl() so that we can ensure the checksum is
1466 * generated using the correct checksum algorithm and accounts for the
1467 * logical I/O size and not just a gang fragment.
1469 return (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1470 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1471 zio->io_offset, NULL) == 0);
1475 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1476 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1477 * isn't modified later on. If buf is compressed or there is already a checksum
1478 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1481 arc_cksum_compute(arc_buf_t *buf)
1483 arc_buf_hdr_t *hdr = buf->b_hdr;
1485 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1488 ASSERT(HDR_HAS_L1HDR(hdr));
1490 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1491 if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1492 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1496 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1497 ASSERT(!ARC_BUF_COMPRESSED(buf));
1498 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1500 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1501 hdr->b_l1hdr.b_freeze_cksum);
1502 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1508 arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
1510 panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr);
1516 arc_buf_unwatch(arc_buf_t *buf)
1520 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1521 PROT_READ | PROT_WRITE));
1528 arc_buf_watch(arc_buf_t *buf)
1532 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1537 static arc_buf_contents_t
1538 arc_buf_type(arc_buf_hdr_t *hdr)
1540 arc_buf_contents_t type;
1541 if (HDR_ISTYPE_METADATA(hdr)) {
1542 type = ARC_BUFC_METADATA;
1544 type = ARC_BUFC_DATA;
1546 VERIFY3U(hdr->b_type, ==, type);
1551 arc_is_metadata(arc_buf_t *buf)
1553 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1557 arc_bufc_to_flags(arc_buf_contents_t type)
1561 /* metadata field is 0 if buffer contains normal data */
1563 case ARC_BUFC_METADATA:
1564 return (ARC_FLAG_BUFC_METADATA);
1568 panic("undefined ARC buffer type!");
1569 return ((uint32_t)-1);
1573 arc_buf_thaw(arc_buf_t *buf)
1575 arc_buf_hdr_t *hdr = buf->b_hdr;
1577 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1578 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1580 arc_cksum_verify(buf);
1583 * Compressed buffers do not manipulate the b_freeze_cksum.
1585 if (ARC_BUF_COMPRESSED(buf))
1588 ASSERT(HDR_HAS_L1HDR(hdr));
1589 arc_cksum_free(hdr);
1590 arc_buf_unwatch(buf);
1594 arc_buf_freeze(arc_buf_t *buf)
1596 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1599 if (ARC_BUF_COMPRESSED(buf))
1602 ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1603 arc_cksum_compute(buf);
1607 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1608 * the following functions should be used to ensure that the flags are
1609 * updated in a thread-safe way. When manipulating the flags either
1610 * the hash_lock must be held or the hdr must be undiscoverable. This
1611 * ensures that we're not racing with any other threads when updating
1615 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1617 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1618 hdr->b_flags |= flags;
1622 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1624 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1625 hdr->b_flags &= ~flags;
1629 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1630 * done in a special way since we have to clear and set bits
1631 * at the same time. Consumers that wish to set the compression bits
1632 * must use this function to ensure that the flags are updated in
1633 * thread-safe manner.
1636 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1638 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1641 * Holes and embedded blocks will always have a psize = 0 so
1642 * we ignore the compression of the blkptr and set the
1643 * want to uncompress them. Mark them as uncompressed.
1645 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1646 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1647 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1649 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1650 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1653 HDR_SET_COMPRESS(hdr, cmp);
1654 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1658 * Looks for another buf on the same hdr which has the data decompressed, copies
1659 * from it, and returns true. If no such buf exists, returns false.
1662 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1664 arc_buf_hdr_t *hdr = buf->b_hdr;
1665 boolean_t copied = B_FALSE;
1667 ASSERT(HDR_HAS_L1HDR(hdr));
1668 ASSERT3P(buf->b_data, !=, NULL);
1669 ASSERT(!ARC_BUF_COMPRESSED(buf));
1671 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1672 from = from->b_next) {
1673 /* can't use our own data buffer */
1678 if (!ARC_BUF_COMPRESSED(from)) {
1679 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1686 * There were no decompressed bufs, so there should not be a
1687 * checksum on the hdr either.
1689 if (zfs_flags & ZFS_DEBUG_MODIFY)
1690 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1696 * Allocates an ARC buf header that's in an evicted & L2-cached state.
1697 * This is used during l2arc reconstruction to make empty ARC buffers
1698 * which circumvent the regular disk->arc->l2arc path and instead come
1699 * into being in the reverse order, i.e. l2arc->arc.
1701 static arc_buf_hdr_t *
1702 arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
1703 dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth,
1704 enum zio_compress compress, uint8_t complevel, boolean_t protected,
1705 boolean_t prefetch, arc_state_type_t arcs_state)
1710 hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP);
1711 hdr->b_birth = birth;
1714 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR);
1715 HDR_SET_LSIZE(hdr, size);
1716 HDR_SET_PSIZE(hdr, psize);
1717 arc_hdr_set_compress(hdr, compress);
1718 hdr->b_complevel = complevel;
1720 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
1722 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
1723 hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa);
1727 hdr->b_l2hdr.b_dev = dev;
1728 hdr->b_l2hdr.b_daddr = daddr;
1729 hdr->b_l2hdr.b_arcs_state = arcs_state;
1735 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1738 arc_hdr_size(arc_buf_hdr_t *hdr)
1742 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1743 HDR_GET_PSIZE(hdr) > 0) {
1744 size = HDR_GET_PSIZE(hdr);
1746 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1747 size = HDR_GET_LSIZE(hdr);
1753 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1757 uint64_t lsize = HDR_GET_LSIZE(hdr);
1758 uint64_t psize = HDR_GET_PSIZE(hdr);
1759 void *tmpbuf = NULL;
1760 abd_t *abd = hdr->b_l1hdr.b_pabd;
1762 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1763 ASSERT(HDR_AUTHENTICATED(hdr));
1764 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1767 * The MAC is calculated on the compressed data that is stored on disk.
1768 * However, if compressed arc is disabled we will only have the
1769 * decompressed data available to us now. Compress it into a temporary
1770 * abd so we can verify the MAC. The performance overhead of this will
1771 * be relatively low, since most objects in an encrypted objset will
1772 * be encrypted (instead of authenticated) anyway.
1774 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1775 !HDR_COMPRESSION_ENABLED(hdr)) {
1776 tmpbuf = zio_buf_alloc(lsize);
1777 abd = abd_get_from_buf(tmpbuf, lsize);
1778 abd_take_ownership_of_buf(abd, B_TRUE);
1779 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1780 hdr->b_l1hdr.b_pabd, tmpbuf, lsize, hdr->b_complevel);
1781 ASSERT3U(csize, <=, psize);
1782 abd_zero_off(abd, csize, psize - csize);
1786 * Authentication is best effort. We authenticate whenever the key is
1787 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1789 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1790 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1791 ASSERT3U(lsize, ==, psize);
1792 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1793 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1795 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1796 hdr->b_crypt_hdr.b_mac);
1800 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1801 else if (ret != ENOENT)
1817 * This function will take a header that only has raw encrypted data in
1818 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1819 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1820 * also decompress the data.
1823 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1828 boolean_t no_crypt = B_FALSE;
1829 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1831 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1832 ASSERT(HDR_ENCRYPTED(hdr));
1834 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT);
1836 ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1837 B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1838 hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1839 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1844 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1845 HDR_GET_PSIZE(hdr));
1849 * If this header has disabled arc compression but the b_pabd is
1850 * compressed after decrypting it, we need to decompress the newly
1853 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1854 !HDR_COMPRESSION_ENABLED(hdr)) {
1856 * We want to make sure that we are correctly honoring the
1857 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1858 * and then loan a buffer from it, rather than allocating a
1859 * linear buffer and wrapping it in an abd later.
1861 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
1863 tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
1865 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1866 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
1867 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
1869 abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
1873 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
1874 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1875 arc_hdr_size(hdr), hdr);
1876 hdr->b_l1hdr.b_pabd = cabd;
1882 arc_hdr_free_abd(hdr, B_FALSE);
1884 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
1890 * This function is called during arc_buf_fill() to prepare the header's
1891 * abd plaintext pointer for use. This involves authenticated protected
1892 * data and decrypting encrypted data into the plaintext abd.
1895 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
1896 const zbookmark_phys_t *zb, boolean_t noauth)
1900 ASSERT(HDR_PROTECTED(hdr));
1902 if (hash_lock != NULL)
1903 mutex_enter(hash_lock);
1905 if (HDR_NOAUTH(hdr) && !noauth) {
1907 * The caller requested authenticated data but our data has
1908 * not been authenticated yet. Verify the MAC now if we can.
1910 ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
1913 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
1915 * If we only have the encrypted version of the data, but the
1916 * unencrypted version was requested we take this opportunity
1917 * to store the decrypted version in the header for future use.
1919 ret = arc_hdr_decrypt(hdr, spa, zb);
1924 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1926 if (hash_lock != NULL)
1927 mutex_exit(hash_lock);
1932 if (hash_lock != NULL)
1933 mutex_exit(hash_lock);
1939 * This function is used by the dbuf code to decrypt bonus buffers in place.
1940 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
1941 * block, so we use the hash lock here to protect against concurrent calls to
1945 arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock)
1947 arc_buf_hdr_t *hdr = buf->b_hdr;
1949 ASSERT(HDR_ENCRYPTED(hdr));
1950 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1951 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1952 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1954 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
1956 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
1957 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1958 hdr->b_crypt_hdr.b_ebufcnt -= 1;
1962 * Given a buf that has a data buffer attached to it, this function will
1963 * efficiently fill the buf with data of the specified compression setting from
1964 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1965 * are already sharing a data buf, no copy is performed.
1967 * If the buf is marked as compressed but uncompressed data was requested, this
1968 * will allocate a new data buffer for the buf, remove that flag, and fill the
1969 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1970 * uncompressed data, and (since we haven't added support for it yet) if you
1971 * want compressed data your buf must already be marked as compressed and have
1972 * the correct-sized data buffer.
1975 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
1976 arc_fill_flags_t flags)
1979 arc_buf_hdr_t *hdr = buf->b_hdr;
1980 boolean_t hdr_compressed =
1981 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
1982 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
1983 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
1984 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
1985 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
1987 ASSERT3P(buf->b_data, !=, NULL);
1988 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
1989 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
1990 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
1991 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
1992 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
1993 IMPLY(encrypted, !ARC_BUF_SHARED(buf));
1996 * If the caller wanted encrypted data we just need to copy it from
1997 * b_rabd and potentially byteswap it. We won't be able to do any
1998 * further transforms on it.
2001 ASSERT(HDR_HAS_RABD(hdr));
2002 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
2003 HDR_GET_PSIZE(hdr));
2008 * Adjust encrypted and authenticated headers to accommodate
2009 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
2010 * allowed to fail decryption due to keys not being loaded
2011 * without being marked as an IO error.
2013 if (HDR_PROTECTED(hdr)) {
2014 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
2015 zb, !!(flags & ARC_FILL_NOAUTH));
2016 if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
2018 } else if (error != 0) {
2019 if (hash_lock != NULL)
2020 mutex_enter(hash_lock);
2021 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2022 if (hash_lock != NULL)
2023 mutex_exit(hash_lock);
2029 * There is a special case here for dnode blocks which are
2030 * decrypting their bonus buffers. These blocks may request to
2031 * be decrypted in-place. This is necessary because there may
2032 * be many dnodes pointing into this buffer and there is
2033 * currently no method to synchronize replacing the backing
2034 * b_data buffer and updating all of the pointers. Here we use
2035 * the hash lock to ensure there are no races. If the need
2036 * arises for other types to be decrypted in-place, they must
2037 * add handling here as well.
2039 if ((flags & ARC_FILL_IN_PLACE) != 0) {
2040 ASSERT(!hdr_compressed);
2041 ASSERT(!compressed);
2044 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
2045 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2047 if (hash_lock != NULL)
2048 mutex_enter(hash_lock);
2049 arc_buf_untransform_in_place(buf, hash_lock);
2050 if (hash_lock != NULL)
2051 mutex_exit(hash_lock);
2053 /* Compute the hdr's checksum if necessary */
2054 arc_cksum_compute(buf);
2060 if (hdr_compressed == compressed) {
2061 if (!arc_buf_is_shared(buf)) {
2062 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2066 ASSERT(hdr_compressed);
2067 ASSERT(!compressed);
2070 * If the buf is sharing its data with the hdr, unlink it and
2071 * allocate a new data buffer for the buf.
2073 if (arc_buf_is_shared(buf)) {
2074 ASSERT(ARC_BUF_COMPRESSED(buf));
2076 /* We need to give the buf its own b_data */
2077 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2079 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2080 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2082 /* Previously overhead was 0; just add new overhead */
2083 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2084 } else if (ARC_BUF_COMPRESSED(buf)) {
2085 /* We need to reallocate the buf's b_data */
2086 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2089 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2091 /* We increased the size of b_data; update overhead */
2092 ARCSTAT_INCR(arcstat_overhead_size,
2093 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2097 * Regardless of the buf's previous compression settings, it
2098 * should not be compressed at the end of this function.
2100 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2103 * Try copying the data from another buf which already has a
2104 * decompressed version. If that's not possible, it's time to
2105 * bite the bullet and decompress the data from the hdr.
2107 if (arc_buf_try_copy_decompressed_data(buf)) {
2108 /* Skip byteswapping and checksumming (already done) */
2111 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2112 hdr->b_l1hdr.b_pabd, buf->b_data,
2113 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr),
2117 * Absent hardware errors or software bugs, this should
2118 * be impossible, but log it anyway so we can debug it.
2122 "hdr %px, compress %d, psize %d, lsize %d",
2123 hdr, arc_hdr_get_compress(hdr),
2124 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2125 if (hash_lock != NULL)
2126 mutex_enter(hash_lock);
2127 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2128 if (hash_lock != NULL)
2129 mutex_exit(hash_lock);
2130 return (SET_ERROR(EIO));
2136 /* Byteswap the buf's data if necessary */
2137 if (bswap != DMU_BSWAP_NUMFUNCS) {
2138 ASSERT(!HDR_SHARED_DATA(hdr));
2139 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2140 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2143 /* Compute the hdr's checksum if necessary */
2144 arc_cksum_compute(buf);
2150 * If this function is being called to decrypt an encrypted buffer or verify an
2151 * authenticated one, the key must be loaded and a mapping must be made
2152 * available in the keystore via spa_keystore_create_mapping() or one of its
2156 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2160 arc_fill_flags_t flags = 0;
2163 flags |= ARC_FILL_IN_PLACE;
2165 ret = arc_buf_fill(buf, spa, zb, flags);
2166 if (ret == ECKSUM) {
2168 * Convert authentication and decryption errors to EIO
2169 * (and generate an ereport) before leaving the ARC.
2171 ret = SET_ERROR(EIO);
2172 spa_log_error(spa, zb);
2173 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2174 spa, NULL, zb, NULL, 0);
2181 * Increment the amount of evictable space in the arc_state_t's refcount.
2182 * We account for the space used by the hdr and the arc buf individually
2183 * so that we can add and remove them from the refcount individually.
2186 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2188 arc_buf_contents_t type = arc_buf_type(hdr);
2190 ASSERT(HDR_HAS_L1HDR(hdr));
2192 if (GHOST_STATE(state)) {
2193 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2194 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2195 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2196 ASSERT(!HDR_HAS_RABD(hdr));
2197 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2198 HDR_GET_LSIZE(hdr), hdr);
2202 if (hdr->b_l1hdr.b_pabd != NULL) {
2203 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2204 arc_hdr_size(hdr), hdr);
2206 if (HDR_HAS_RABD(hdr)) {
2207 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2208 HDR_GET_PSIZE(hdr), hdr);
2211 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2212 buf = buf->b_next) {
2213 if (arc_buf_is_shared(buf))
2215 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2216 arc_buf_size(buf), buf);
2221 * Decrement the amount of evictable space in the arc_state_t's refcount.
2222 * We account for the space used by the hdr and the arc buf individually
2223 * so that we can add and remove them from the refcount individually.
2226 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2228 arc_buf_contents_t type = arc_buf_type(hdr);
2230 ASSERT(HDR_HAS_L1HDR(hdr));
2232 if (GHOST_STATE(state)) {
2233 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2234 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2235 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2236 ASSERT(!HDR_HAS_RABD(hdr));
2237 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2238 HDR_GET_LSIZE(hdr), hdr);
2242 if (hdr->b_l1hdr.b_pabd != NULL) {
2243 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2244 arc_hdr_size(hdr), hdr);
2246 if (HDR_HAS_RABD(hdr)) {
2247 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2248 HDR_GET_PSIZE(hdr), hdr);
2251 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2252 buf = buf->b_next) {
2253 if (arc_buf_is_shared(buf))
2255 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2256 arc_buf_size(buf), buf);
2261 * Add a reference to this hdr indicating that someone is actively
2262 * referencing that memory. When the refcount transitions from 0 to 1,
2263 * we remove it from the respective arc_state_t list to indicate that
2264 * it is not evictable.
2267 add_reference(arc_buf_hdr_t *hdr, void *tag)
2271 ASSERT(HDR_HAS_L1HDR(hdr));
2272 if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) {
2273 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2274 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2275 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2278 state = hdr->b_l1hdr.b_state;
2280 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2281 (state != arc_anon)) {
2282 /* We don't use the L2-only state list. */
2283 if (state != arc_l2c_only) {
2284 multilist_remove(&state->arcs_list[arc_buf_type(hdr)],
2286 arc_evictable_space_decrement(hdr, state);
2288 /* remove the prefetch flag if we get a reference */
2289 if (HDR_HAS_L2HDR(hdr))
2290 l2arc_hdr_arcstats_decrement_state(hdr);
2291 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2292 if (HDR_HAS_L2HDR(hdr))
2293 l2arc_hdr_arcstats_increment_state(hdr);
2298 * Remove a reference from this hdr. When the reference transitions from
2299 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2300 * list making it eligible for eviction.
2303 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2306 arc_state_t *state = hdr->b_l1hdr.b_state;
2308 ASSERT(HDR_HAS_L1HDR(hdr));
2309 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2310 ASSERT(!GHOST_STATE(state));
2313 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2314 * check to prevent usage of the arc_l2c_only list.
2316 if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2317 (state != arc_anon)) {
2318 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
2319 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2320 arc_evictable_space_increment(hdr, state);
2326 * Returns detailed information about a specific arc buffer. When the
2327 * state_index argument is set the function will calculate the arc header
2328 * list position for its arc state. Since this requires a linear traversal
2329 * callers are strongly encourage not to do this. However, it can be helpful
2330 * for targeted analysis so the functionality is provided.
2333 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2335 arc_buf_hdr_t *hdr = ab->b_hdr;
2336 l1arc_buf_hdr_t *l1hdr = NULL;
2337 l2arc_buf_hdr_t *l2hdr = NULL;
2338 arc_state_t *state = NULL;
2340 memset(abi, 0, sizeof (arc_buf_info_t));
2345 abi->abi_flags = hdr->b_flags;
2347 if (HDR_HAS_L1HDR(hdr)) {
2348 l1hdr = &hdr->b_l1hdr;
2349 state = l1hdr->b_state;
2351 if (HDR_HAS_L2HDR(hdr))
2352 l2hdr = &hdr->b_l2hdr;
2355 abi->abi_bufcnt = l1hdr->b_bufcnt;
2356 abi->abi_access = l1hdr->b_arc_access;
2357 abi->abi_mru_hits = l1hdr->b_mru_hits;
2358 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2359 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2360 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2361 abi->abi_holds = zfs_refcount_count(&l1hdr->b_refcnt);
2365 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2366 abi->abi_l2arc_hits = l2hdr->b_hits;
2369 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2370 abi->abi_state_contents = arc_buf_type(hdr);
2371 abi->abi_size = arc_hdr_size(hdr);
2375 * Move the supplied buffer to the indicated state. The hash lock
2376 * for the buffer must be held by the caller.
2379 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2380 kmutex_t *hash_lock)
2382 arc_state_t *old_state;
2385 boolean_t update_old, update_new;
2386 arc_buf_contents_t buftype = arc_buf_type(hdr);
2389 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2390 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2391 * L1 hdr doesn't always exist when we change state to arc_anon before
2392 * destroying a header, in which case reallocating to add the L1 hdr is
2395 if (HDR_HAS_L1HDR(hdr)) {
2396 old_state = hdr->b_l1hdr.b_state;
2397 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2398 bufcnt = hdr->b_l1hdr.b_bufcnt;
2399 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
2402 old_state = arc_l2c_only;
2405 update_old = B_FALSE;
2407 update_new = update_old;
2409 ASSERT(MUTEX_HELD(hash_lock));
2410 ASSERT3P(new_state, !=, old_state);
2411 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2412 ASSERT(old_state != arc_anon || bufcnt <= 1);
2415 * If this buffer is evictable, transfer it from the
2416 * old state list to the new state list.
2419 if (old_state != arc_anon && old_state != arc_l2c_only) {
2420 ASSERT(HDR_HAS_L1HDR(hdr));
2421 multilist_remove(&old_state->arcs_list[buftype], hdr);
2423 if (GHOST_STATE(old_state)) {
2425 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2426 update_old = B_TRUE;
2428 arc_evictable_space_decrement(hdr, old_state);
2430 if (new_state != arc_anon && new_state != arc_l2c_only) {
2432 * An L1 header always exists here, since if we're
2433 * moving to some L1-cached state (i.e. not l2c_only or
2434 * anonymous), we realloc the header to add an L1hdr
2437 ASSERT(HDR_HAS_L1HDR(hdr));
2438 multilist_insert(&new_state->arcs_list[buftype], hdr);
2440 if (GHOST_STATE(new_state)) {
2442 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2443 update_new = B_TRUE;
2445 arc_evictable_space_increment(hdr, new_state);
2449 ASSERT(!HDR_EMPTY(hdr));
2450 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2451 buf_hash_remove(hdr);
2453 /* adjust state sizes (ignore arc_l2c_only) */
2455 if (update_new && new_state != arc_l2c_only) {
2456 ASSERT(HDR_HAS_L1HDR(hdr));
2457 if (GHOST_STATE(new_state)) {
2461 * When moving a header to a ghost state, we first
2462 * remove all arc buffers. Thus, we'll have a
2463 * bufcnt of zero, and no arc buffer to use for
2464 * the reference. As a result, we use the arc
2465 * header pointer for the reference.
2467 (void) zfs_refcount_add_many(&new_state->arcs_size,
2468 HDR_GET_LSIZE(hdr), hdr);
2469 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2470 ASSERT(!HDR_HAS_RABD(hdr));
2472 uint32_t buffers = 0;
2475 * Each individual buffer holds a unique reference,
2476 * thus we must remove each of these references one
2479 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2480 buf = buf->b_next) {
2481 ASSERT3U(bufcnt, !=, 0);
2485 * When the arc_buf_t is sharing the data
2486 * block with the hdr, the owner of the
2487 * reference belongs to the hdr. Only
2488 * add to the refcount if the arc_buf_t is
2491 if (arc_buf_is_shared(buf))
2494 (void) zfs_refcount_add_many(
2495 &new_state->arcs_size,
2496 arc_buf_size(buf), buf);
2498 ASSERT3U(bufcnt, ==, buffers);
2500 if (hdr->b_l1hdr.b_pabd != NULL) {
2501 (void) zfs_refcount_add_many(
2502 &new_state->arcs_size,
2503 arc_hdr_size(hdr), hdr);
2506 if (HDR_HAS_RABD(hdr)) {
2507 (void) zfs_refcount_add_many(
2508 &new_state->arcs_size,
2509 HDR_GET_PSIZE(hdr), hdr);
2514 if (update_old && old_state != arc_l2c_only) {
2515 ASSERT(HDR_HAS_L1HDR(hdr));
2516 if (GHOST_STATE(old_state)) {
2518 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2519 ASSERT(!HDR_HAS_RABD(hdr));
2522 * When moving a header off of a ghost state,
2523 * the header will not contain any arc buffers.
2524 * We use the arc header pointer for the reference
2525 * which is exactly what we did when we put the
2526 * header on the ghost state.
2529 (void) zfs_refcount_remove_many(&old_state->arcs_size,
2530 HDR_GET_LSIZE(hdr), hdr);
2532 uint32_t buffers = 0;
2535 * Each individual buffer holds a unique reference,
2536 * thus we must remove each of these references one
2539 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2540 buf = buf->b_next) {
2541 ASSERT3U(bufcnt, !=, 0);
2545 * When the arc_buf_t is sharing the data
2546 * block with the hdr, the owner of the
2547 * reference belongs to the hdr. Only
2548 * add to the refcount if the arc_buf_t is
2551 if (arc_buf_is_shared(buf))
2554 (void) zfs_refcount_remove_many(
2555 &old_state->arcs_size, arc_buf_size(buf),
2558 ASSERT3U(bufcnt, ==, buffers);
2559 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2562 if (hdr->b_l1hdr.b_pabd != NULL) {
2563 (void) zfs_refcount_remove_many(
2564 &old_state->arcs_size, arc_hdr_size(hdr),
2568 if (HDR_HAS_RABD(hdr)) {
2569 (void) zfs_refcount_remove_many(
2570 &old_state->arcs_size, HDR_GET_PSIZE(hdr),
2576 if (HDR_HAS_L1HDR(hdr)) {
2577 hdr->b_l1hdr.b_state = new_state;
2579 if (HDR_HAS_L2HDR(hdr) && new_state != arc_l2c_only) {
2580 l2arc_hdr_arcstats_decrement_state(hdr);
2581 hdr->b_l2hdr.b_arcs_state = new_state->arcs_state;
2582 l2arc_hdr_arcstats_increment_state(hdr);
2588 arc_space_consume(uint64_t space, arc_space_type_t type)
2590 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2595 case ARC_SPACE_DATA:
2596 ARCSTAT_INCR(arcstat_data_size, space);
2598 case ARC_SPACE_META:
2599 ARCSTAT_INCR(arcstat_metadata_size, space);
2601 case ARC_SPACE_BONUS:
2602 ARCSTAT_INCR(arcstat_bonus_size, space);
2604 case ARC_SPACE_DNODE:
2605 aggsum_add(&arc_sums.arcstat_dnode_size, space);
2607 case ARC_SPACE_DBUF:
2608 ARCSTAT_INCR(arcstat_dbuf_size, space);
2610 case ARC_SPACE_HDRS:
2611 ARCSTAT_INCR(arcstat_hdr_size, space);
2613 case ARC_SPACE_L2HDRS:
2614 aggsum_add(&arc_sums.arcstat_l2_hdr_size, space);
2616 case ARC_SPACE_ABD_CHUNK_WASTE:
2618 * Note: this includes space wasted by all scatter ABD's, not
2619 * just those allocated by the ARC. But the vast majority of
2620 * scatter ABD's come from the ARC, because other users are
2623 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, space);
2627 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE)
2628 aggsum_add(&arc_sums.arcstat_meta_used, space);
2630 aggsum_add(&arc_sums.arcstat_size, space);
2634 arc_space_return(uint64_t space, arc_space_type_t type)
2636 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2641 case ARC_SPACE_DATA:
2642 ARCSTAT_INCR(arcstat_data_size, -space);
2644 case ARC_SPACE_META:
2645 ARCSTAT_INCR(arcstat_metadata_size, -space);
2647 case ARC_SPACE_BONUS:
2648 ARCSTAT_INCR(arcstat_bonus_size, -space);
2650 case ARC_SPACE_DNODE:
2651 aggsum_add(&arc_sums.arcstat_dnode_size, -space);
2653 case ARC_SPACE_DBUF:
2654 ARCSTAT_INCR(arcstat_dbuf_size, -space);
2656 case ARC_SPACE_HDRS:
2657 ARCSTAT_INCR(arcstat_hdr_size, -space);
2659 case ARC_SPACE_L2HDRS:
2660 aggsum_add(&arc_sums.arcstat_l2_hdr_size, -space);
2662 case ARC_SPACE_ABD_CHUNK_WASTE:
2663 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, -space);
2667 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE) {
2668 ASSERT(aggsum_compare(&arc_sums.arcstat_meta_used,
2670 ARCSTAT_MAX(arcstat_meta_max,
2671 aggsum_upper_bound(&arc_sums.arcstat_meta_used));
2672 aggsum_add(&arc_sums.arcstat_meta_used, -space);
2675 ASSERT(aggsum_compare(&arc_sums.arcstat_size, space) >= 0);
2676 aggsum_add(&arc_sums.arcstat_size, -space);
2680 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2681 * with the hdr's b_pabd.
2684 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2687 * The criteria for sharing a hdr's data are:
2688 * 1. the buffer is not encrypted
2689 * 2. the hdr's compression matches the buf's compression
2690 * 3. the hdr doesn't need to be byteswapped
2691 * 4. the hdr isn't already being shared
2692 * 5. the buf is either compressed or it is the last buf in the hdr list
2694 * Criterion #5 maintains the invariant that shared uncompressed
2695 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2696 * might ask, "if a compressed buf is allocated first, won't that be the
2697 * last thing in the list?", but in that case it's impossible to create
2698 * a shared uncompressed buf anyway (because the hdr must be compressed
2699 * to have the compressed buf). You might also think that #3 is
2700 * sufficient to make this guarantee, however it's possible
2701 * (specifically in the rare L2ARC write race mentioned in
2702 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2703 * is shareable, but wasn't at the time of its allocation. Rather than
2704 * allow a new shared uncompressed buf to be created and then shuffle
2705 * the list around to make it the last element, this simply disallows
2706 * sharing if the new buf isn't the first to be added.
2708 ASSERT3P(buf->b_hdr, ==, hdr);
2709 boolean_t hdr_compressed =
2710 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF;
2711 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2712 return (!ARC_BUF_ENCRYPTED(buf) &&
2713 buf_compressed == hdr_compressed &&
2714 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2715 !HDR_SHARED_DATA(hdr) &&
2716 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2720 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2721 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2722 * copy was made successfully, or an error code otherwise.
2725 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2726 void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth,
2727 boolean_t fill, arc_buf_t **ret)
2730 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2732 ASSERT(HDR_HAS_L1HDR(hdr));
2733 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2734 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2735 hdr->b_type == ARC_BUFC_METADATA);
2736 ASSERT3P(ret, !=, NULL);
2737 ASSERT3P(*ret, ==, NULL);
2738 IMPLY(encrypted, compressed);
2740 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2743 buf->b_next = hdr->b_l1hdr.b_buf;
2746 add_reference(hdr, tag);
2749 * We're about to change the hdr's b_flags. We must either
2750 * hold the hash_lock or be undiscoverable.
2752 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2755 * Only honor requests for compressed bufs if the hdr is actually
2756 * compressed. This must be overridden if the buffer is encrypted since
2757 * encrypted buffers cannot be decompressed.
2760 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2761 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2762 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2763 } else if (compressed &&
2764 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2765 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2766 flags |= ARC_FILL_COMPRESSED;
2771 flags |= ARC_FILL_NOAUTH;
2775 * If the hdr's data can be shared then we share the data buffer and
2776 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2777 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2778 * buffer to store the buf's data.
2780 * There are two additional restrictions here because we're sharing
2781 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2782 * actively involved in an L2ARC write, because if this buf is used by
2783 * an arc_write() then the hdr's data buffer will be released when the
2784 * write completes, even though the L2ARC write might still be using it.
2785 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2786 * need to be ABD-aware. It must be allocated via
2787 * zio_[data_]buf_alloc(), not as a page, because we need to be able
2788 * to abd_release_ownership_of_buf(), which isn't allowed on "linear
2789 * page" buffers because the ABD code needs to handle freeing them
2792 boolean_t can_share = arc_can_share(hdr, buf) &&
2793 !HDR_L2_WRITING(hdr) &&
2794 hdr->b_l1hdr.b_pabd != NULL &&
2795 abd_is_linear(hdr->b_l1hdr.b_pabd) &&
2796 !abd_is_linear_page(hdr->b_l1hdr.b_pabd);
2798 /* Set up b_data and sharing */
2800 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2801 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2802 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2805 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2806 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2808 VERIFY3P(buf->b_data, !=, NULL);
2810 hdr->b_l1hdr.b_buf = buf;
2811 hdr->b_l1hdr.b_bufcnt += 1;
2813 hdr->b_crypt_hdr.b_ebufcnt += 1;
2816 * If the user wants the data from the hdr, we need to either copy or
2817 * decompress the data.
2820 ASSERT3P(zb, !=, NULL);
2821 return (arc_buf_fill(buf, spa, zb, flags));
2827 static char *arc_onloan_tag = "onloan";
2830 arc_loaned_bytes_update(int64_t delta)
2832 atomic_add_64(&arc_loaned_bytes, delta);
2834 /* assert that it did not wrap around */
2835 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2839 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2840 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2841 * buffers must be returned to the arc before they can be used by the DMU or
2845 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2847 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2848 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2850 arc_loaned_bytes_update(arc_buf_size(buf));
2856 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2857 enum zio_compress compression_type, uint8_t complevel)
2859 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2860 psize, lsize, compression_type, complevel);
2862 arc_loaned_bytes_update(arc_buf_size(buf));
2868 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2869 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2870 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2871 enum zio_compress compression_type, uint8_t complevel)
2873 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2874 byteorder, salt, iv, mac, ot, psize, lsize, compression_type,
2877 atomic_add_64(&arc_loaned_bytes, psize);
2883 * Return a loaned arc buffer to the arc.
2886 arc_return_buf(arc_buf_t *buf, 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, tag);
2893 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2895 arc_loaned_bytes_update(-arc_buf_size(buf));
2898 /* Detach an arc_buf from a dbuf (tag) */
2900 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2902 arc_buf_hdr_t *hdr = buf->b_hdr;
2904 ASSERT3P(buf->b_data, !=, NULL);
2905 ASSERT(HDR_HAS_L1HDR(hdr));
2906 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2907 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2909 arc_loaned_bytes_update(arc_buf_size(buf));
2913 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2915 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2918 df->l2df_size = size;
2919 df->l2df_type = type;
2920 mutex_enter(&l2arc_free_on_write_mtx);
2921 list_insert_head(l2arc_free_on_write, df);
2922 mutex_exit(&l2arc_free_on_write_mtx);
2926 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
2928 arc_state_t *state = hdr->b_l1hdr.b_state;
2929 arc_buf_contents_t type = arc_buf_type(hdr);
2930 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
2932 /* protected by hash lock, if in the hash table */
2933 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2934 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2935 ASSERT(state != arc_anon && state != arc_l2c_only);
2937 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2940 (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr);
2941 if (type == ARC_BUFC_METADATA) {
2942 arc_space_return(size, ARC_SPACE_META);
2944 ASSERT(type == ARC_BUFC_DATA);
2945 arc_space_return(size, ARC_SPACE_DATA);
2949 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
2951 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2956 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2957 * data buffer, we transfer the refcount ownership to the hdr and update
2958 * the appropriate kstats.
2961 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2963 ASSERT(arc_can_share(hdr, buf));
2964 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2965 ASSERT(!ARC_BUF_ENCRYPTED(buf));
2966 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2969 * Start sharing the data buffer. We transfer the
2970 * refcount ownership to the hdr since it always owns
2971 * the refcount whenever an arc_buf_t is shared.
2973 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
2974 arc_hdr_size(hdr), buf, hdr);
2975 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2976 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2977 HDR_ISTYPE_METADATA(hdr));
2978 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2979 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2982 * Since we've transferred ownership to the hdr we need
2983 * to increment its compressed and uncompressed kstats and
2984 * decrement the overhead size.
2986 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2987 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2988 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2992 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2994 ASSERT(arc_buf_is_shared(buf));
2995 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2996 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2999 * We are no longer sharing this buffer so we need
3000 * to transfer its ownership to the rightful owner.
3002 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
3003 arc_hdr_size(hdr), hdr, buf);
3004 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3005 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3006 abd_free(hdr->b_l1hdr.b_pabd);
3007 hdr->b_l1hdr.b_pabd = NULL;
3008 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3011 * Since the buffer is no longer shared between
3012 * the arc buf and the hdr, count it as overhead.
3014 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3015 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3016 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3020 * Remove an arc_buf_t from the hdr's buf list and return the last
3021 * arc_buf_t on the list. If no buffers remain on the list then return
3025 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3027 ASSERT(HDR_HAS_L1HDR(hdr));
3028 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3030 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3031 arc_buf_t *lastbuf = NULL;
3034 * Remove the buf from the hdr list and locate the last
3035 * remaining buffer on the list.
3037 while (*bufp != NULL) {
3039 *bufp = buf->b_next;
3042 * If we've removed a buffer in the middle of
3043 * the list then update the lastbuf and update
3046 if (*bufp != NULL) {
3048 bufp = &(*bufp)->b_next;
3052 ASSERT3P(lastbuf, !=, buf);
3053 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3054 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3055 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3061 * Free up buf->b_data and pull the arc_buf_t off of the arc_buf_hdr_t's
3065 arc_buf_destroy_impl(arc_buf_t *buf)
3067 arc_buf_hdr_t *hdr = buf->b_hdr;
3070 * Free up the data associated with the buf but only if we're not
3071 * sharing this with the hdr. If we are sharing it with the hdr, the
3072 * hdr is responsible for doing the free.
3074 if (buf->b_data != NULL) {
3076 * We're about to change the hdr's b_flags. We must either
3077 * hold the hash_lock or be undiscoverable.
3079 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3081 arc_cksum_verify(buf);
3082 arc_buf_unwatch(buf);
3084 if (arc_buf_is_shared(buf)) {
3085 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3087 uint64_t size = arc_buf_size(buf);
3088 arc_free_data_buf(hdr, buf->b_data, size, buf);
3089 ARCSTAT_INCR(arcstat_overhead_size, -size);
3093 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3094 hdr->b_l1hdr.b_bufcnt -= 1;
3096 if (ARC_BUF_ENCRYPTED(buf)) {
3097 hdr->b_crypt_hdr.b_ebufcnt -= 1;
3100 * If we have no more encrypted buffers and we've
3101 * already gotten a copy of the decrypted data we can
3102 * free b_rabd to save some space.
3104 if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
3105 HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
3106 !HDR_IO_IN_PROGRESS(hdr)) {
3107 arc_hdr_free_abd(hdr, B_TRUE);
3112 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3114 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3116 * If the current arc_buf_t is sharing its data buffer with the
3117 * hdr, then reassign the hdr's b_pabd to share it with the new
3118 * buffer at the end of the list. The shared buffer is always
3119 * the last one on the hdr's buffer list.
3121 * There is an equivalent case for compressed bufs, but since
3122 * they aren't guaranteed to be the last buf in the list and
3123 * that is an exceedingly rare case, we just allow that space be
3124 * wasted temporarily. We must also be careful not to share
3125 * encrypted buffers, since they cannot be shared.
3127 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
3128 /* Only one buf can be shared at once */
3129 VERIFY(!arc_buf_is_shared(lastbuf));
3130 /* hdr is uncompressed so can't have compressed buf */
3131 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3133 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3134 arc_hdr_free_abd(hdr, B_FALSE);
3137 * We must setup a new shared block between the
3138 * last buffer and the hdr. The data would have
3139 * been allocated by the arc buf so we need to transfer
3140 * ownership to the hdr since it's now being shared.
3142 arc_share_buf(hdr, lastbuf);
3144 } else if (HDR_SHARED_DATA(hdr)) {
3146 * Uncompressed shared buffers are always at the end
3147 * of the list. Compressed buffers don't have the
3148 * same requirements. This makes it hard to
3149 * simply assert that the lastbuf is shared so
3150 * we rely on the hdr's compression flags to determine
3151 * if we have a compressed, shared buffer.
3153 ASSERT3P(lastbuf, !=, NULL);
3154 ASSERT(arc_buf_is_shared(lastbuf) ||
3155 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3159 * Free the checksum if we're removing the last uncompressed buf from
3162 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3163 arc_cksum_free(hdr);
3166 /* clean up the buf */
3168 kmem_cache_free(buf_cache, buf);
3172 arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, int alloc_flags)
3175 boolean_t alloc_rdata = ((alloc_flags & ARC_HDR_ALLOC_RDATA) != 0);
3177 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3178 ASSERT(HDR_HAS_L1HDR(hdr));
3179 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3180 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3183 size = HDR_GET_PSIZE(hdr);
3184 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3185 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr,
3187 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3188 ARCSTAT_INCR(arcstat_raw_size, size);
3190 size = arc_hdr_size(hdr);
3191 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3192 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr,
3194 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3197 ARCSTAT_INCR(arcstat_compressed_size, size);
3198 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3202 arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3204 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3206 ASSERT(HDR_HAS_L1HDR(hdr));
3207 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3208 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3211 * If the hdr is currently being written to the l2arc then
3212 * we defer freeing the data by adding it to the l2arc_free_on_write
3213 * list. The l2arc will free the data once it's finished
3214 * writing it to the l2arc device.
3216 if (HDR_L2_WRITING(hdr)) {
3217 arc_hdr_free_on_write(hdr, free_rdata);
3218 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3219 } else if (free_rdata) {
3220 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3222 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr);
3226 hdr->b_crypt_hdr.b_rabd = NULL;
3227 ARCSTAT_INCR(arcstat_raw_size, -size);
3229 hdr->b_l1hdr.b_pabd = NULL;
3232 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3233 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3235 ARCSTAT_INCR(arcstat_compressed_size, -size);
3236 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3240 * Allocate empty anonymous ARC header. The header will get its identity
3241 * assigned and buffers attached later as part of read or write operations.
3243 * In case of read arc_read() assigns header its identify (b_dva + b_birth),
3244 * inserts it into ARC hash to become globally visible and allocates physical
3245 * (b_pabd) or raw (b_rabd) ABD buffer to read into from disk. On disk read
3246 * completion arc_read_done() allocates ARC buffer(s) as needed, potentially
3247 * sharing one of them with the physical ABD buffer.
3249 * In case of write arc_alloc_buf() allocates ARC buffer to be filled with
3250 * data. Then after compression and/or encryption arc_write_ready() allocates
3251 * and fills (or potentially shares) physical (b_pabd) or raw (b_rabd) ABD
3252 * buffer. On disk write completion arc_write_done() assigns the header its
3253 * new identity (b_dva + b_birth) and inserts into ARC hash.
3255 * In case of partial overwrite the old data is read first as described. Then
3256 * arc_release() either allocates new anonymous ARC header and moves the ARC
3257 * buffer to it, or reuses the old ARC header by discarding its identity and
3258 * removing it from ARC hash. After buffer modification normal write process
3259 * follows as described.
3261 static arc_buf_hdr_t *
3262 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3263 boolean_t protected, enum zio_compress compression_type, uint8_t complevel,
3264 arc_buf_contents_t type)
3268 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3270 hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
3272 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3275 ASSERT(HDR_EMPTY(hdr));
3276 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3277 HDR_SET_PSIZE(hdr, psize);
3278 HDR_SET_LSIZE(hdr, lsize);
3282 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3283 arc_hdr_set_compress(hdr, compression_type);
3284 hdr->b_complevel = complevel;
3286 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3288 hdr->b_l1hdr.b_state = arc_anon;
3289 hdr->b_l1hdr.b_arc_access = 0;
3290 hdr->b_l1hdr.b_mru_hits = 0;
3291 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3292 hdr->b_l1hdr.b_mfu_hits = 0;
3293 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3294 hdr->b_l1hdr.b_bufcnt = 0;
3295 hdr->b_l1hdr.b_buf = NULL;
3297 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3303 * Transition between the two allocation states for the arc_buf_hdr struct.
3304 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3305 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3306 * version is used when a cache buffer is only in the L2ARC in order to reduce
3309 static arc_buf_hdr_t *
3310 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3312 ASSERT(HDR_HAS_L2HDR(hdr));
3314 arc_buf_hdr_t *nhdr;
3315 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3317 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3318 (old == hdr_l2only_cache && new == hdr_full_cache));
3321 * if the caller wanted a new full header and the header is to be
3322 * encrypted we will actually allocate the header from the full crypt
3323 * cache instead. The same applies to freeing from the old cache.
3325 if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
3326 new = hdr_full_crypt_cache;
3327 if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
3328 old = hdr_full_crypt_cache;
3330 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3332 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3333 buf_hash_remove(hdr);
3335 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3337 if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
3338 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3340 * arc_access and arc_change_state need to be aware that a
3341 * header has just come out of L2ARC, so we set its state to
3342 * l2c_only even though it's about to change.
3344 nhdr->b_l1hdr.b_state = arc_l2c_only;
3346 /* Verify previous threads set to NULL before freeing */
3347 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3348 ASSERT(!HDR_HAS_RABD(hdr));
3350 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3351 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3352 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3355 * If we've reached here, We must have been called from
3356 * arc_evict_hdr(), as such we should have already been
3357 * removed from any ghost list we were previously on
3358 * (which protects us from racing with arc_evict_state),
3359 * thus no locking is needed during this check.
3361 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3364 * A buffer must not be moved into the arc_l2c_only
3365 * state if it's not finished being written out to the
3366 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3367 * might try to be accessed, even though it was removed.
3369 VERIFY(!HDR_L2_WRITING(hdr));
3370 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3371 ASSERT(!HDR_HAS_RABD(hdr));
3373 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3376 * The header has been reallocated so we need to re-insert it into any
3379 (void) buf_hash_insert(nhdr, NULL);
3381 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3383 mutex_enter(&dev->l2ad_mtx);
3386 * We must place the realloc'ed header back into the list at
3387 * the same spot. Otherwise, if it's placed earlier in the list,
3388 * l2arc_write_buffers() could find it during the function's
3389 * write phase, and try to write it out to the l2arc.
3391 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3392 list_remove(&dev->l2ad_buflist, hdr);
3394 mutex_exit(&dev->l2ad_mtx);
3397 * Since we're using the pointer address as the tag when
3398 * incrementing and decrementing the l2ad_alloc refcount, we
3399 * must remove the old pointer (that we're about to destroy) and
3400 * add the new pointer to the refcount. Otherwise we'd remove
3401 * the wrong pointer address when calling arc_hdr_destroy() later.
3404 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
3405 arc_hdr_size(hdr), hdr);
3406 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
3407 arc_hdr_size(nhdr), nhdr);
3409 buf_discard_identity(hdr);
3410 kmem_cache_free(old, hdr);
3416 * This function allows an L1 header to be reallocated as a crypt
3417 * header and vice versa. If we are going to a crypt header, the
3418 * new fields will be zeroed out.
3420 static arc_buf_hdr_t *
3421 arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
3423 arc_buf_hdr_t *nhdr;
3425 kmem_cache_t *ncache, *ocache;
3426 unsigned nsize, osize;
3429 * This function requires that hdr is in the arc_anon state.
3430 * Therefore it won't have any L2ARC data for us to worry
3433 ASSERT(HDR_HAS_L1HDR(hdr));
3434 ASSERT(!HDR_HAS_L2HDR(hdr));
3435 ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
3436 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3437 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3438 ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node));
3439 ASSERT3P(hdr->b_hash_next, ==, NULL);
3442 ncache = hdr_full_crypt_cache;
3443 nsize = sizeof (hdr->b_crypt_hdr);
3444 ocache = hdr_full_cache;
3445 osize = HDR_FULL_SIZE;
3447 ncache = hdr_full_cache;
3448 nsize = HDR_FULL_SIZE;
3449 ocache = hdr_full_crypt_cache;
3450 osize = sizeof (hdr->b_crypt_hdr);
3453 nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);
3456 * Copy all members that aren't locks or condvars to the new header.
3457 * No lists are pointing to us (as we asserted above), so we don't
3458 * need to worry about the list nodes.
3460 nhdr->b_dva = hdr->b_dva;
3461 nhdr->b_birth = hdr->b_birth;
3462 nhdr->b_type = hdr->b_type;
3463 nhdr->b_flags = hdr->b_flags;
3464 nhdr->b_psize = hdr->b_psize;
3465 nhdr->b_lsize = hdr->b_lsize;
3466 nhdr->b_spa = hdr->b_spa;
3467 nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
3468 nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
3469 nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
3470 nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
3471 nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
3472 nhdr->b_l1hdr.b_mru_hits = hdr->b_l1hdr.b_mru_hits;
3473 nhdr->b_l1hdr.b_mru_ghost_hits = hdr->b_l1hdr.b_mru_ghost_hits;
3474 nhdr->b_l1hdr.b_mfu_hits = hdr->b_l1hdr.b_mfu_hits;
3475 nhdr->b_l1hdr.b_mfu_ghost_hits = hdr->b_l1hdr.b_mfu_ghost_hits;
3476 nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
3477 nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
3480 * This zfs_refcount_add() exists only to ensure that the individual
3481 * arc buffers always point to a header that is referenced, avoiding
3482 * a small race condition that could trigger ASSERTs.
3484 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
3485 nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
3486 for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) {
3487 mutex_enter(&buf->b_evict_lock);
3489 mutex_exit(&buf->b_evict_lock);
3492 zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
3493 (void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
3494 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3497 arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
3499 arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
3502 /* unset all members of the original hdr */
3503 bzero(&hdr->b_dva, sizeof (dva_t));
3505 hdr->b_type = ARC_BUFC_INVALID;
3510 hdr->b_l1hdr.b_freeze_cksum = NULL;
3511 hdr->b_l1hdr.b_buf = NULL;
3512 hdr->b_l1hdr.b_bufcnt = 0;
3513 hdr->b_l1hdr.b_byteswap = 0;
3514 hdr->b_l1hdr.b_state = NULL;
3515 hdr->b_l1hdr.b_arc_access = 0;
3516 hdr->b_l1hdr.b_mru_hits = 0;
3517 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3518 hdr->b_l1hdr.b_mfu_hits = 0;
3519 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3520 hdr->b_l1hdr.b_acb = NULL;
3521 hdr->b_l1hdr.b_pabd = NULL;
3523 if (ocache == hdr_full_crypt_cache) {
3524 ASSERT(!HDR_HAS_RABD(hdr));
3525 hdr->b_crypt_hdr.b_ot = DMU_OT_NONE;
3526 hdr->b_crypt_hdr.b_ebufcnt = 0;
3527 hdr->b_crypt_hdr.b_dsobj = 0;
3528 bzero(hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3529 bzero(hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3530 bzero(hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3533 buf_discard_identity(hdr);
3534 kmem_cache_free(ocache, hdr);
3540 * This function is used by the send / receive code to convert a newly
3541 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3542 * is also used to allow the root objset block to be updated without altering
3543 * its embedded MACs. Both block types will always be uncompressed so we do not
3544 * have to worry about compression type or psize.
3547 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3548 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3551 arc_buf_hdr_t *hdr = buf->b_hdr;
3553 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3554 ASSERT(HDR_HAS_L1HDR(hdr));
3555 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3557 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3558 if (!HDR_PROTECTED(hdr))
3559 hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
3560 hdr->b_crypt_hdr.b_dsobj = dsobj;
3561 hdr->b_crypt_hdr.b_ot = ot;
3562 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3563 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3564 if (!arc_hdr_has_uncompressed_buf(hdr))
3565 arc_cksum_free(hdr);
3568 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3570 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3572 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3576 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3577 * The buf is returned thawed since we expect the consumer to modify it.
3580 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3582 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3583 B_FALSE, ZIO_COMPRESS_OFF, 0, type);
3585 arc_buf_t *buf = NULL;
3586 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3587 B_FALSE, B_FALSE, &buf));
3594 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3595 * for bufs containing metadata.
3598 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3599 enum zio_compress compression_type, uint8_t complevel)
3601 ASSERT3U(lsize, >, 0);
3602 ASSERT3U(lsize, >=, psize);
3603 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3604 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3606 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3607 B_FALSE, compression_type, complevel, ARC_BUFC_DATA);
3609 arc_buf_t *buf = NULL;
3610 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3611 B_TRUE, B_FALSE, B_FALSE, &buf));
3613 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3616 * To ensure that the hdr has the correct data in it if we call
3617 * arc_untransform() on this buf before it's been written to disk,
3618 * it's easiest if we just set up sharing between the buf and the hdr.
3620 arc_share_buf(hdr, buf);
3626 arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder,
3627 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
3628 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3629 enum zio_compress compression_type, uint8_t complevel)
3633 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3634 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3636 ASSERT3U(lsize, >, 0);
3637 ASSERT3U(lsize, >=, psize);
3638 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3639 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3641 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3642 compression_type, complevel, type);
3644 hdr->b_crypt_hdr.b_dsobj = dsobj;
3645 hdr->b_crypt_hdr.b_ot = ot;
3646 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3647 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3648 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3649 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3650 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3653 * This buffer will be considered encrypted even if the ot is not an
3654 * encrypted type. It will become authenticated instead in
3655 * arc_write_ready().
3658 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3659 B_FALSE, B_FALSE, &buf));
3661 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3667 l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
3668 boolean_t state_only)
3670 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3671 l2arc_dev_t *dev = l2hdr->b_dev;
3672 uint64_t lsize = HDR_GET_LSIZE(hdr);
3673 uint64_t psize = HDR_GET_PSIZE(hdr);
3674 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3675 arc_buf_contents_t type = hdr->b_type;
3690 /* If the buffer is a prefetch, count it as such. */
3691 if (HDR_PREFETCH(hdr)) {
3692 ARCSTAT_INCR(arcstat_l2_prefetch_asize, asize_s);
3695 * We use the value stored in the L2 header upon initial
3696 * caching in L2ARC. This value will be updated in case
3697 * an MRU/MRU_ghost buffer transitions to MFU but the L2ARC
3698 * metadata (log entry) cannot currently be updated. Having
3699 * the ARC state in the L2 header solves the problem of a
3700 * possibly absent L1 header (apparent in buffers restored
3701 * from persistent L2ARC).
3703 switch (hdr->b_l2hdr.b_arcs_state) {
3704 case ARC_STATE_MRU_GHOST:
3706 ARCSTAT_INCR(arcstat_l2_mru_asize, asize_s);
3708 case ARC_STATE_MFU_GHOST:
3710 ARCSTAT_INCR(arcstat_l2_mfu_asize, asize_s);
3720 ARCSTAT_INCR(arcstat_l2_psize, psize_s);
3721 ARCSTAT_INCR(arcstat_l2_lsize, lsize_s);
3725 ARCSTAT_INCR(arcstat_l2_bufc_data_asize, asize_s);
3727 case ARC_BUFC_METADATA:
3728 ARCSTAT_INCR(arcstat_l2_bufc_metadata_asize, asize_s);
3737 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3739 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3740 l2arc_dev_t *dev = l2hdr->b_dev;
3741 uint64_t psize = HDR_GET_PSIZE(hdr);
3742 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3744 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3745 ASSERT(HDR_HAS_L2HDR(hdr));
3747 list_remove(&dev->l2ad_buflist, hdr);
3749 l2arc_hdr_arcstats_decrement(hdr);
3750 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3752 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3754 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3758 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3760 if (HDR_HAS_L1HDR(hdr)) {
3761 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3762 hdr->b_l1hdr.b_bufcnt > 0);
3763 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3764 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3766 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3767 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3769 if (HDR_HAS_L2HDR(hdr)) {
3770 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3771 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3774 mutex_enter(&dev->l2ad_mtx);
3777 * Even though we checked this conditional above, we
3778 * need to check this again now that we have the
3779 * l2ad_mtx. This is because we could be racing with
3780 * another thread calling l2arc_evict() which might have
3781 * destroyed this header's L2 portion as we were waiting
3782 * to acquire the l2ad_mtx. If that happens, we don't
3783 * want to re-destroy the header's L2 portion.
3785 if (HDR_HAS_L2HDR(hdr))
3786 arc_hdr_l2hdr_destroy(hdr);
3789 mutex_exit(&dev->l2ad_mtx);
3793 * The header's identify can only be safely discarded once it is no
3794 * longer discoverable. This requires removing it from the hash table
3795 * and the l2arc header list. After this point the hash lock can not
3796 * be used to protect the header.
3798 if (!HDR_EMPTY(hdr))
3799 buf_discard_identity(hdr);
3801 if (HDR_HAS_L1HDR(hdr)) {
3802 arc_cksum_free(hdr);
3804 while (hdr->b_l1hdr.b_buf != NULL)
3805 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3807 if (hdr->b_l1hdr.b_pabd != NULL)
3808 arc_hdr_free_abd(hdr, B_FALSE);
3810 if (HDR_HAS_RABD(hdr))
3811 arc_hdr_free_abd(hdr, B_TRUE);
3814 ASSERT3P(hdr->b_hash_next, ==, NULL);
3815 if (HDR_HAS_L1HDR(hdr)) {
3816 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3817 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3819 if (!HDR_PROTECTED(hdr)) {
3820 kmem_cache_free(hdr_full_cache, hdr);
3822 kmem_cache_free(hdr_full_crypt_cache, hdr);
3825 kmem_cache_free(hdr_l2only_cache, hdr);
3830 arc_buf_destroy(arc_buf_t *buf, void* tag)
3832 arc_buf_hdr_t *hdr = buf->b_hdr;
3834 if (hdr->b_l1hdr.b_state == arc_anon) {
3835 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3836 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3837 VERIFY0(remove_reference(hdr, NULL, tag));
3838 arc_hdr_destroy(hdr);
3842 kmutex_t *hash_lock = HDR_LOCK(hdr);
3843 mutex_enter(hash_lock);
3845 ASSERT3P(hdr, ==, buf->b_hdr);
3846 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3847 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3848 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3849 ASSERT3P(buf->b_data, !=, NULL);
3851 (void) remove_reference(hdr, hash_lock, tag);
3852 arc_buf_destroy_impl(buf);
3853 mutex_exit(hash_lock);
3857 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3858 * state of the header is dependent on its state prior to entering this
3859 * function. The following transitions are possible:
3861 * - arc_mru -> arc_mru_ghost
3862 * - arc_mfu -> arc_mfu_ghost
3863 * - arc_mru_ghost -> arc_l2c_only
3864 * - arc_mru_ghost -> deleted
3865 * - arc_mfu_ghost -> arc_l2c_only
3866 * - arc_mfu_ghost -> deleted
3868 * Return total size of evicted data buffers for eviction progress tracking.
3869 * When evicting from ghost states return logical buffer size to make eviction
3870 * progress at the same (or at least comparable) rate as from non-ghost states.
3872 * Return *real_evicted for actual ARC size reduction to wake up threads
3873 * waiting for it. For non-ghost states it includes size of evicted data
3874 * buffers (the headers are not freed there). For ghost states it includes
3875 * only the evicted headers size.
3878 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, uint64_t *real_evicted)
3880 arc_state_t *evicted_state, *state;
3881 int64_t bytes_evicted = 0;
3882 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3883 arc_min_prescient_prefetch_ms : arc_min_prefetch_ms;
3885 ASSERT(MUTEX_HELD(hash_lock));
3886 ASSERT(HDR_HAS_L1HDR(hdr));
3889 state = hdr->b_l1hdr.b_state;
3890 if (GHOST_STATE(state)) {
3891 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3892 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3895 * l2arc_write_buffers() relies on a header's L1 portion
3896 * (i.e. its b_pabd field) during it's write phase.
3897 * Thus, we cannot push a header onto the arc_l2c_only
3898 * state (removing its L1 piece) until the header is
3899 * done being written to the l2arc.
3901 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3902 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3903 return (bytes_evicted);
3906 ARCSTAT_BUMP(arcstat_deleted);
3907 bytes_evicted += HDR_GET_LSIZE(hdr);
3909 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3911 if (HDR_HAS_L2HDR(hdr)) {
3912 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3913 ASSERT(!HDR_HAS_RABD(hdr));
3915 * This buffer is cached on the 2nd Level ARC;
3916 * don't destroy the header.
3918 arc_change_state(arc_l2c_only, hdr, hash_lock);
3920 * dropping from L1+L2 cached to L2-only,
3921 * realloc to remove the L1 header.
3923 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3925 *real_evicted += HDR_FULL_SIZE - HDR_L2ONLY_SIZE;
3927 arc_change_state(arc_anon, hdr, hash_lock);
3928 arc_hdr_destroy(hdr);
3929 *real_evicted += HDR_FULL_SIZE;
3931 return (bytes_evicted);
3934 ASSERT(state == arc_mru || state == arc_mfu);
3935 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3937 /* prefetch buffers have a minimum lifespan */
3938 if (HDR_IO_IN_PROGRESS(hdr) ||
3939 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3940 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3941 MSEC_TO_TICK(min_lifetime))) {
3942 ARCSTAT_BUMP(arcstat_evict_skip);
3943 return (bytes_evicted);
3946 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3947 while (hdr->b_l1hdr.b_buf) {
3948 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3949 if (!mutex_tryenter(&buf->b_evict_lock)) {
3950 ARCSTAT_BUMP(arcstat_mutex_miss);
3953 if (buf->b_data != NULL) {
3954 bytes_evicted += HDR_GET_LSIZE(hdr);
3955 *real_evicted += HDR_GET_LSIZE(hdr);
3957 mutex_exit(&buf->b_evict_lock);
3958 arc_buf_destroy_impl(buf);
3961 if (HDR_HAS_L2HDR(hdr)) {
3962 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3964 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3965 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3966 HDR_GET_LSIZE(hdr));
3968 switch (state->arcs_state) {
3971 arcstat_evict_l2_eligible_mru,
3972 HDR_GET_LSIZE(hdr));
3976 arcstat_evict_l2_eligible_mfu,
3977 HDR_GET_LSIZE(hdr));
3983 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3984 HDR_GET_LSIZE(hdr));
3988 if (hdr->b_l1hdr.b_bufcnt == 0) {
3989 arc_cksum_free(hdr);
3991 bytes_evicted += arc_hdr_size(hdr);
3992 *real_evicted += arc_hdr_size(hdr);
3995 * If this hdr is being evicted and has a compressed
3996 * buffer then we discard it here before we change states.
3997 * This ensures that the accounting is updated correctly
3998 * in arc_free_data_impl().
4000 if (hdr->b_l1hdr.b_pabd != NULL)
4001 arc_hdr_free_abd(hdr, B_FALSE);
4003 if (HDR_HAS_RABD(hdr))
4004 arc_hdr_free_abd(hdr, B_TRUE);
4006 arc_change_state(evicted_state, hdr, hash_lock);
4007 ASSERT(HDR_IN_HASH_TABLE(hdr));
4008 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
4009 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
4012 return (bytes_evicted);
4016 arc_set_need_free(void)
4018 ASSERT(MUTEX_HELD(&arc_evict_lock));
4019 int64_t remaining = arc_free_memory() - arc_sys_free / 2;
4020 arc_evict_waiter_t *aw = list_tail(&arc_evict_waiters);
4022 arc_need_free = MAX(-remaining, 0);
4025 MAX(-remaining, (int64_t)(aw->aew_count - arc_evict_count));
4030 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
4031 uint64_t spa, uint64_t bytes)
4033 multilist_sublist_t *mls;
4034 uint64_t bytes_evicted = 0, real_evicted = 0;
4036 kmutex_t *hash_lock;
4037 int evict_count = zfs_arc_evict_batch_limit;
4039 ASSERT3P(marker, !=, NULL);
4041 mls = multilist_sublist_lock(ml, idx);
4043 for (hdr = multilist_sublist_prev(mls, marker); likely(hdr != NULL);
4044 hdr = multilist_sublist_prev(mls, marker)) {
4045 if ((evict_count <= 0) || (bytes_evicted >= bytes))
4049 * To keep our iteration location, move the marker
4050 * forward. Since we're not holding hdr's hash lock, we
4051 * must be very careful and not remove 'hdr' from the
4052 * sublist. Otherwise, other consumers might mistake the
4053 * 'hdr' as not being on a sublist when they call the
4054 * multilist_link_active() function (they all rely on
4055 * the hash lock protecting concurrent insertions and
4056 * removals). multilist_sublist_move_forward() was
4057 * specifically implemented to ensure this is the case
4058 * (only 'marker' will be removed and re-inserted).
4060 multilist_sublist_move_forward(mls, marker);
4063 * The only case where the b_spa field should ever be
4064 * zero, is the marker headers inserted by
4065 * arc_evict_state(). It's possible for multiple threads
4066 * to be calling arc_evict_state() concurrently (e.g.
4067 * dsl_pool_close() and zio_inject_fault()), so we must
4068 * skip any markers we see from these other threads.
4070 if (hdr->b_spa == 0)
4073 /* we're only interested in evicting buffers of a certain spa */
4074 if (spa != 0 && hdr->b_spa != spa) {
4075 ARCSTAT_BUMP(arcstat_evict_skip);
4079 hash_lock = HDR_LOCK(hdr);
4082 * We aren't calling this function from any code path
4083 * that would already be holding a hash lock, so we're
4084 * asserting on this assumption to be defensive in case
4085 * this ever changes. Without this check, it would be
4086 * possible to incorrectly increment arcstat_mutex_miss
4087 * below (e.g. if the code changed such that we called
4088 * this function with a hash lock held).
4090 ASSERT(!MUTEX_HELD(hash_lock));
4092 if (mutex_tryenter(hash_lock)) {
4094 uint64_t evicted = arc_evict_hdr(hdr, hash_lock,
4096 mutex_exit(hash_lock);
4098 bytes_evicted += evicted;
4099 real_evicted += revicted;
4102 * If evicted is zero, arc_evict_hdr() must have
4103 * decided to skip this header, don't increment
4104 * evict_count in this case.
4110 ARCSTAT_BUMP(arcstat_mutex_miss);
4114 multilist_sublist_unlock(mls);
4117 * Increment the count of evicted bytes, and wake up any threads that
4118 * are waiting for the count to reach this value. Since the list is
4119 * ordered by ascending aew_count, we pop off the beginning of the
4120 * list until we reach the end, or a waiter that's past the current
4121 * "count". Doing this outside the loop reduces the number of times
4122 * we need to acquire the global arc_evict_lock.
4124 * Only wake when there's sufficient free memory in the system
4125 * (specifically, arc_sys_free/2, which by default is a bit more than
4126 * 1/64th of RAM). See the comments in arc_wait_for_eviction().
4128 mutex_enter(&arc_evict_lock);
4129 arc_evict_count += real_evicted;
4131 if (arc_free_memory() > arc_sys_free / 2) {
4132 arc_evict_waiter_t *aw;
4133 while ((aw = list_head(&arc_evict_waiters)) != NULL &&
4134 aw->aew_count <= arc_evict_count) {
4135 list_remove(&arc_evict_waiters, aw);
4136 cv_broadcast(&aw->aew_cv);
4139 arc_set_need_free();
4140 mutex_exit(&arc_evict_lock);
4143 * If the ARC size is reduced from arc_c_max to arc_c_min (especially
4144 * if the average cached block is small), eviction can be on-CPU for
4145 * many seconds. To ensure that other threads that may be bound to
4146 * this CPU are able to make progress, make a voluntary preemption
4151 return (bytes_evicted);
4155 * Allocate an array of buffer headers used as placeholders during arc state
4158 static arc_buf_hdr_t **
4159 arc_state_alloc_markers(int count)
4161 arc_buf_hdr_t **markers;
4163 markers = kmem_zalloc(sizeof (*markers) * count, KM_SLEEP);
4164 for (int i = 0; i < count; i++) {
4165 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
4168 * A b_spa of 0 is used to indicate that this header is
4169 * a marker. This fact is used in arc_evict_type() and
4170 * arc_evict_state_impl().
4172 markers[i]->b_spa = 0;
4179 arc_state_free_markers(arc_buf_hdr_t **markers, int count)
4181 for (int i = 0; i < count; i++)
4182 kmem_cache_free(hdr_full_cache, markers[i]);
4183 kmem_free(markers, sizeof (*markers) * count);
4187 * Evict buffers from the given arc state, until we've removed the
4188 * specified number of bytes. Move the removed buffers to the
4189 * appropriate evict state.
4191 * This function makes a "best effort". It skips over any buffers
4192 * it can't get a hash_lock on, and so, may not catch all candidates.
4193 * It may also return without evicting as much space as requested.
4195 * If bytes is specified using the special value ARC_EVICT_ALL, this
4196 * will evict all available (i.e. unlocked and evictable) buffers from
4197 * the given arc state; which is used by arc_flush().
4200 arc_evict_state(arc_state_t *state, uint64_t spa, uint64_t bytes,
4201 arc_buf_contents_t type)
4203 uint64_t total_evicted = 0;
4204 multilist_t *ml = &state->arcs_list[type];
4206 arc_buf_hdr_t **markers;
4208 num_sublists = multilist_get_num_sublists(ml);
4211 * If we've tried to evict from each sublist, made some
4212 * progress, but still have not hit the target number of bytes
4213 * to evict, we want to keep trying. The markers allow us to
4214 * pick up where we left off for each individual sublist, rather
4215 * than starting from the tail each time.
4217 if (zthr_iscurthread(arc_evict_zthr)) {
4218 markers = arc_state_evict_markers;
4219 ASSERT3S(num_sublists, <=, arc_state_evict_marker_count);
4221 markers = arc_state_alloc_markers(num_sublists);
4223 for (int i = 0; i < num_sublists; i++) {
4224 multilist_sublist_t *mls;
4226 mls = multilist_sublist_lock(ml, i);
4227 multilist_sublist_insert_tail(mls, markers[i]);
4228 multilist_sublist_unlock(mls);
4232 * While we haven't hit our target number of bytes to evict, or
4233 * we're evicting all available buffers.
4235 while (total_evicted < bytes) {
4236 int sublist_idx = multilist_get_random_index(ml);
4237 uint64_t scan_evicted = 0;
4240 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
4241 * Request that 10% of the LRUs be scanned by the superblock
4244 if (type == ARC_BUFC_DATA && aggsum_compare(
4245 &arc_sums.arcstat_dnode_size, arc_dnode_size_limit) > 0) {
4246 arc_prune_async((aggsum_upper_bound(
4247 &arc_sums.arcstat_dnode_size) -
4248 arc_dnode_size_limit) / sizeof (dnode_t) /
4249 zfs_arc_dnode_reduce_percent);
4253 * Start eviction using a randomly selected sublist,
4254 * this is to try and evenly balance eviction across all
4255 * sublists. Always starting at the same sublist
4256 * (e.g. index 0) would cause evictions to favor certain
4257 * sublists over others.
4259 for (int i = 0; i < num_sublists; i++) {
4260 uint64_t bytes_remaining;
4261 uint64_t bytes_evicted;
4263 if (total_evicted < bytes)
4264 bytes_remaining = bytes - total_evicted;
4268 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4269 markers[sublist_idx], spa, bytes_remaining);
4271 scan_evicted += bytes_evicted;
4272 total_evicted += bytes_evicted;
4274 /* we've reached the end, wrap to the beginning */
4275 if (++sublist_idx >= num_sublists)
4280 * If we didn't evict anything during this scan, we have
4281 * no reason to believe we'll evict more during another
4282 * scan, so break the loop.
4284 if (scan_evicted == 0) {
4285 /* This isn't possible, let's make that obvious */
4286 ASSERT3S(bytes, !=, 0);
4289 * When bytes is ARC_EVICT_ALL, the only way to
4290 * break the loop is when scan_evicted is zero.
4291 * In that case, we actually have evicted enough,
4292 * so we don't want to increment the kstat.
4294 if (bytes != ARC_EVICT_ALL) {
4295 ASSERT3S(total_evicted, <, bytes);
4296 ARCSTAT_BUMP(arcstat_evict_not_enough);
4303 for (int i = 0; i < num_sublists; i++) {
4304 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4305 multilist_sublist_remove(mls, markers[i]);
4306 multilist_sublist_unlock(mls);
4308 if (markers != arc_state_evict_markers)
4309 arc_state_free_markers(markers, num_sublists);
4311 return (total_evicted);
4315 * Flush all "evictable" data of the given type from the arc state
4316 * specified. This will not evict any "active" buffers (i.e. referenced).
4318 * When 'retry' is set to B_FALSE, the function will make a single pass
4319 * over the state and evict any buffers that it can. Since it doesn't
4320 * continually retry the eviction, it might end up leaving some buffers
4321 * in the ARC due to lock misses.
4323 * When 'retry' is set to B_TRUE, the function will continually retry the
4324 * eviction until *all* evictable buffers have been removed from the
4325 * state. As a result, if concurrent insertions into the state are
4326 * allowed (e.g. if the ARC isn't shutting down), this function might
4327 * wind up in an infinite loop, continually trying to evict buffers.
4330 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4333 uint64_t evicted = 0;
4335 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4336 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4346 * Evict the specified number of bytes from the state specified,
4347 * restricting eviction to the spa and type given. This function
4348 * prevents us from trying to evict more from a state's list than
4349 * is "evictable", and to skip evicting altogether when passed a
4350 * negative value for "bytes". In contrast, arc_evict_state() will
4351 * evict everything it can, when passed a negative value for "bytes".
4354 arc_evict_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4355 arc_buf_contents_t type)
4359 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4360 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4362 return (arc_evict_state(state, spa, delta, type));
4369 * The goal of this function is to evict enough meta data buffers from the
4370 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4371 * more complicated than it appears because it is common for data buffers
4372 * to have holds on meta data buffers. In addition, dnode meta data buffers
4373 * will be held by the dnodes in the block preventing them from being freed.
4374 * This means we can't simply traverse the ARC and expect to always find
4375 * enough unheld meta data buffer to release.
4377 * Therefore, this function has been updated to make alternating passes
4378 * over the ARC releasing data buffers and then newly unheld meta data
4379 * buffers. This ensures forward progress is maintained and meta_used
4380 * will decrease. Normally this is sufficient, but if required the ARC
4381 * will call the registered prune callbacks causing dentry and inodes to
4382 * be dropped from the VFS cache. This will make dnode meta data buffers
4383 * available for reclaim.
4386 arc_evict_meta_balanced(uint64_t meta_used)
4388 int64_t delta, prune = 0, adjustmnt;
4389 uint64_t total_evicted = 0;
4390 arc_buf_contents_t type = ARC_BUFC_DATA;
4391 int restarts = MAX(zfs_arc_meta_adjust_restarts, 0);
4395 * This slightly differs than the way we evict from the mru in
4396 * arc_evict because we don't have a "target" value (i.e. no
4397 * "meta" arc_p). As a result, I think we can completely
4398 * cannibalize the metadata in the MRU before we evict the
4399 * metadata from the MFU. I think we probably need to implement a
4400 * "metadata arc_p" value to do this properly.
4402 adjustmnt = meta_used - arc_meta_limit;
4404 if (adjustmnt > 0 &&
4405 zfs_refcount_count(&arc_mru->arcs_esize[type]) > 0) {
4406 delta = MIN(zfs_refcount_count(&arc_mru->arcs_esize[type]),
4408 total_evicted += arc_evict_impl(arc_mru, 0, delta, type);
4413 * We can't afford to recalculate adjustmnt here. If we do,
4414 * new metadata buffers can sneak into the MRU or ANON lists,
4415 * thus penalize the MFU metadata. Although the fudge factor is
4416 * small, it has been empirically shown to be significant for
4417 * certain workloads (e.g. creating many empty directories). As
4418 * such, we use the original calculation for adjustmnt, and
4419 * simply decrement the amount of data evicted from the MRU.
4422 if (adjustmnt > 0 &&
4423 zfs_refcount_count(&arc_mfu->arcs_esize[type]) > 0) {
4424 delta = MIN(zfs_refcount_count(&arc_mfu->arcs_esize[type]),
4426 total_evicted += arc_evict_impl(arc_mfu, 0, delta, type);
4429 adjustmnt = meta_used - arc_meta_limit;
4431 if (adjustmnt > 0 &&
4432 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) {
4433 delta = MIN(adjustmnt,
4434 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]));
4435 total_evicted += arc_evict_impl(arc_mru_ghost, 0, delta, type);
4439 if (adjustmnt > 0 &&
4440 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) {
4441 delta = MIN(adjustmnt,
4442 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]));
4443 total_evicted += arc_evict_impl(arc_mfu_ghost, 0, delta, type);
4447 * If after attempting to make the requested adjustment to the ARC
4448 * the meta limit is still being exceeded then request that the
4449 * higher layers drop some cached objects which have holds on ARC
4450 * meta buffers. Requests to the upper layers will be made with
4451 * increasingly large scan sizes until the ARC is below the limit.
4453 if (meta_used > arc_meta_limit) {
4454 if (type == ARC_BUFC_DATA) {
4455 type = ARC_BUFC_METADATA;
4457 type = ARC_BUFC_DATA;
4459 if (zfs_arc_meta_prune) {
4460 prune += zfs_arc_meta_prune;
4461 arc_prune_async(prune);
4470 return (total_evicted);
4474 * Evict metadata buffers from the cache, such that arcstat_meta_used is
4475 * capped by the arc_meta_limit tunable.
4478 arc_evict_meta_only(uint64_t meta_used)
4480 uint64_t total_evicted = 0;
4484 * If we're over the meta limit, we want to evict enough
4485 * metadata to get back under the meta limit. We don't want to
4486 * evict so much that we drop the MRU below arc_p, though. If
4487 * we're over the meta limit more than we're over arc_p, we
4488 * evict some from the MRU here, and some from the MFU below.
4490 target = MIN((int64_t)(meta_used - arc_meta_limit),
4491 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4492 zfs_refcount_count(&arc_mru->arcs_size) - arc_p));
4494 total_evicted += arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4497 * Similar to the above, we want to evict enough bytes to get us
4498 * below the meta limit, but not so much as to drop us below the
4499 * space allotted to the MFU (which is defined as arc_c - arc_p).
4501 target = MIN((int64_t)(meta_used - arc_meta_limit),
4502 (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) -
4505 total_evicted += arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4507 return (total_evicted);
4511 arc_evict_meta(uint64_t meta_used)
4513 if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY)
4514 return (arc_evict_meta_only(meta_used));
4516 return (arc_evict_meta_balanced(meta_used));
4520 * Return the type of the oldest buffer in the given arc state
4522 * This function will select a random sublist of type ARC_BUFC_DATA and
4523 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4524 * is compared, and the type which contains the "older" buffer will be
4527 static arc_buf_contents_t
4528 arc_evict_type(arc_state_t *state)
4530 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
4531 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
4532 int data_idx = multilist_get_random_index(data_ml);
4533 int meta_idx = multilist_get_random_index(meta_ml);
4534 multilist_sublist_t *data_mls;
4535 multilist_sublist_t *meta_mls;
4536 arc_buf_contents_t type;
4537 arc_buf_hdr_t *data_hdr;
4538 arc_buf_hdr_t *meta_hdr;
4541 * We keep the sublist lock until we're finished, to prevent
4542 * the headers from being destroyed via arc_evict_state().
4544 data_mls = multilist_sublist_lock(data_ml, data_idx);
4545 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4548 * These two loops are to ensure we skip any markers that
4549 * might be at the tail of the lists due to arc_evict_state().
4552 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4553 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4554 if (data_hdr->b_spa != 0)
4558 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4559 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4560 if (meta_hdr->b_spa != 0)
4564 if (data_hdr == NULL && meta_hdr == NULL) {
4565 type = ARC_BUFC_DATA;
4566 } else if (data_hdr == NULL) {
4567 ASSERT3P(meta_hdr, !=, NULL);
4568 type = ARC_BUFC_METADATA;
4569 } else if (meta_hdr == NULL) {
4570 ASSERT3P(data_hdr, !=, NULL);
4571 type = ARC_BUFC_DATA;
4573 ASSERT3P(data_hdr, !=, NULL);
4574 ASSERT3P(meta_hdr, !=, NULL);
4576 /* The headers can't be on the sublist without an L1 header */
4577 ASSERT(HDR_HAS_L1HDR(data_hdr));
4578 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4580 if (data_hdr->b_l1hdr.b_arc_access <
4581 meta_hdr->b_l1hdr.b_arc_access) {
4582 type = ARC_BUFC_DATA;
4584 type = ARC_BUFC_METADATA;
4588 multilist_sublist_unlock(meta_mls);
4589 multilist_sublist_unlock(data_mls);
4595 * Evict buffers from the cache, such that arcstat_size is capped by arc_c.
4600 uint64_t total_evicted = 0;
4603 uint64_t asize = aggsum_value(&arc_sums.arcstat_size);
4604 uint64_t ameta = aggsum_value(&arc_sums.arcstat_meta_used);
4607 * If we're over arc_meta_limit, we want to correct that before
4608 * potentially evicting data buffers below.
4610 total_evicted += arc_evict_meta(ameta);
4615 * If we're over the target cache size, we want to evict enough
4616 * from the list to get back to our target size. We don't want
4617 * to evict too much from the MRU, such that it drops below
4618 * arc_p. So, if we're over our target cache size more than
4619 * the MRU is over arc_p, we'll evict enough to get back to
4620 * arc_p here, and then evict more from the MFU below.
4622 target = MIN((int64_t)(asize - arc_c),
4623 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4624 zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4627 * If we're below arc_meta_min, always prefer to evict data.
4628 * Otherwise, try to satisfy the requested number of bytes to
4629 * evict from the type which contains older buffers; in an
4630 * effort to keep newer buffers in the cache regardless of their
4631 * type. If we cannot satisfy the number of bytes from this
4632 * type, spill over into the next type.
4634 if (arc_evict_type(arc_mru) == ARC_BUFC_METADATA &&
4635 ameta > arc_meta_min) {
4636 bytes = arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4637 total_evicted += bytes;
4640 * If we couldn't evict our target number of bytes from
4641 * metadata, we try to get the rest from data.
4646 arc_evict_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4648 bytes = arc_evict_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4649 total_evicted += bytes;
4652 * If we couldn't evict our target number of bytes from
4653 * data, we try to get the rest from metadata.
4658 arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4662 * Re-sum ARC stats after the first round of evictions.
4664 asize = aggsum_value(&arc_sums.arcstat_size);
4665 ameta = aggsum_value(&arc_sums.arcstat_meta_used);
4671 * Now that we've tried to evict enough from the MRU to get its
4672 * size back to arc_p, if we're still above the target cache
4673 * size, we evict the rest from the MFU.
4675 target = asize - arc_c;
4677 if (arc_evict_type(arc_mfu) == ARC_BUFC_METADATA &&
4678 ameta > arc_meta_min) {
4679 bytes = arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4680 total_evicted += bytes;
4683 * If we couldn't evict our target number of bytes from
4684 * metadata, we try to get the rest from data.
4689 arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4691 bytes = arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4692 total_evicted += bytes;
4695 * If we couldn't evict our target number of bytes from
4696 * data, we try to get the rest from data.
4701 arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4705 * Adjust ghost lists
4707 * In addition to the above, the ARC also defines target values
4708 * for the ghost lists. The sum of the mru list and mru ghost
4709 * list should never exceed the target size of the cache, and
4710 * the sum of the mru list, mfu list, mru ghost list, and mfu
4711 * ghost list should never exceed twice the target size of the
4712 * cache. The following logic enforces these limits on the ghost
4713 * caches, and evicts from them as needed.
4715 target = zfs_refcount_count(&arc_mru->arcs_size) +
4716 zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4718 bytes = arc_evict_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4719 total_evicted += bytes;
4724 arc_evict_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4727 * We assume the sum of the mru list and mfu list is less than
4728 * or equal to arc_c (we enforced this above), which means we
4729 * can use the simpler of the two equations below:
4731 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4732 * mru ghost + mfu ghost <= arc_c
4734 target = zfs_refcount_count(&arc_mru_ghost->arcs_size) +
4735 zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4737 bytes = arc_evict_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4738 total_evicted += bytes;
4743 arc_evict_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4745 return (total_evicted);
4749 arc_flush(spa_t *spa, boolean_t retry)
4754 * If retry is B_TRUE, a spa must not be specified since we have
4755 * no good way to determine if all of a spa's buffers have been
4756 * evicted from an arc state.
4758 ASSERT(!retry || spa == 0);
4761 guid = spa_load_guid(spa);
4763 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4764 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4766 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4767 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4769 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4770 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4772 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4773 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4777 arc_reduce_target_size(int64_t to_free)
4779 uint64_t asize = aggsum_value(&arc_sums.arcstat_size);
4782 * All callers want the ARC to actually evict (at least) this much
4783 * memory. Therefore we reduce from the lower of the current size and
4784 * the target size. This way, even if arc_c is much higher than
4785 * arc_size (as can be the case after many calls to arc_freed(), we will
4786 * immediately have arc_c < arc_size and therefore the arc_evict_zthr
4789 uint64_t c = MIN(arc_c, asize);
4791 if (c > to_free && c - to_free > arc_c_min) {
4792 arc_c = c - to_free;
4793 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4795 arc_p = (arc_c >> 1);
4796 ASSERT(arc_c >= arc_c_min);
4797 ASSERT((int64_t)arc_p >= 0);
4802 if (asize > arc_c) {
4803 /* See comment in arc_evict_cb_check() on why lock+flag */
4804 mutex_enter(&arc_evict_lock);
4805 arc_evict_needed = B_TRUE;
4806 mutex_exit(&arc_evict_lock);
4807 zthr_wakeup(arc_evict_zthr);
4812 * Determine if the system is under memory pressure and is asking
4813 * to reclaim memory. A return value of B_TRUE indicates that the system
4814 * is under memory pressure and that the arc should adjust accordingly.
4817 arc_reclaim_needed(void)
4819 return (arc_available_memory() < 0);
4823 arc_kmem_reap_soon(void)
4826 kmem_cache_t *prev_cache = NULL;
4827 kmem_cache_t *prev_data_cache = NULL;
4828 extern kmem_cache_t *zio_buf_cache[];
4829 extern kmem_cache_t *zio_data_buf_cache[];
4832 if ((aggsum_compare(&arc_sums.arcstat_meta_used,
4833 arc_meta_limit) >= 0) && zfs_arc_meta_prune) {
4835 * We are exceeding our meta-data cache limit.
4836 * Prune some entries to release holds on meta-data.
4838 arc_prune_async(zfs_arc_meta_prune);
4842 * Reclaim unused memory from all kmem caches.
4848 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4850 /* reach upper limit of cache size on 32-bit */
4851 if (zio_buf_cache[i] == NULL)
4854 if (zio_buf_cache[i] != prev_cache) {
4855 prev_cache = zio_buf_cache[i];
4856 kmem_cache_reap_now(zio_buf_cache[i]);
4858 if (zio_data_buf_cache[i] != prev_data_cache) {
4859 prev_data_cache = zio_data_buf_cache[i];
4860 kmem_cache_reap_now(zio_data_buf_cache[i]);
4863 kmem_cache_reap_now(buf_cache);
4864 kmem_cache_reap_now(hdr_full_cache);
4865 kmem_cache_reap_now(hdr_l2only_cache);
4866 kmem_cache_reap_now(zfs_btree_leaf_cache);
4867 abd_cache_reap_now();
4872 arc_evict_cb_check(void *arg, zthr_t *zthr)
4876 * This is necessary in order to keep the kstat information
4877 * up to date for tools that display kstat data such as the
4878 * mdb ::arc dcmd and the Linux crash utility. These tools
4879 * typically do not call kstat's update function, but simply
4880 * dump out stats from the most recent update. Without
4881 * this call, these commands may show stale stats for the
4882 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4883 * with this call, the data might be out of date if the
4884 * evict thread hasn't been woken recently; but that should
4885 * suffice. The arc_state_t structures can be queried
4886 * directly if more accurate information is needed.
4888 if (arc_ksp != NULL)
4889 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4893 * We have to rely on arc_wait_for_eviction() to tell us when to
4894 * evict, rather than checking if we are overflowing here, so that we
4895 * are sure to not leave arc_wait_for_eviction() waiting on aew_cv.
4896 * If we have become "not overflowing" since arc_wait_for_eviction()
4897 * checked, we need to wake it up. We could broadcast the CV here,
4898 * but arc_wait_for_eviction() may have not yet gone to sleep. We
4899 * would need to use a mutex to ensure that this function doesn't
4900 * broadcast until arc_wait_for_eviction() has gone to sleep (e.g.
4901 * the arc_evict_lock). However, the lock ordering of such a lock
4902 * would necessarily be incorrect with respect to the zthr_lock,
4903 * which is held before this function is called, and is held by
4904 * arc_wait_for_eviction() when it calls zthr_wakeup().
4906 return (arc_evict_needed);
4910 * Keep arc_size under arc_c by running arc_evict which evicts data
4915 arc_evict_cb(void *arg, zthr_t *zthr)
4917 uint64_t evicted = 0;
4918 fstrans_cookie_t cookie = spl_fstrans_mark();
4920 /* Evict from cache */
4921 evicted = arc_evict();
4924 * If evicted is zero, we couldn't evict anything
4925 * via arc_evict(). This could be due to hash lock
4926 * collisions, but more likely due to the majority of
4927 * arc buffers being unevictable. Therefore, even if
4928 * arc_size is above arc_c, another pass is unlikely to
4929 * be helpful and could potentially cause us to enter an
4930 * infinite loop. Additionally, zthr_iscancelled() is
4931 * checked here so that if the arc is shutting down, the
4932 * broadcast will wake any remaining arc evict waiters.
4934 mutex_enter(&arc_evict_lock);
4935 arc_evict_needed = !zthr_iscancelled(arc_evict_zthr) &&
4936 evicted > 0 && aggsum_compare(&arc_sums.arcstat_size, arc_c) > 0;
4937 if (!arc_evict_needed) {
4939 * We're either no longer overflowing, or we
4940 * can't evict anything more, so we should wake
4941 * arc_get_data_impl() sooner.
4943 arc_evict_waiter_t *aw;
4944 while ((aw = list_remove_head(&arc_evict_waiters)) != NULL) {
4945 cv_broadcast(&aw->aew_cv);
4947 arc_set_need_free();
4949 mutex_exit(&arc_evict_lock);
4950 spl_fstrans_unmark(cookie);
4955 arc_reap_cb_check(void *arg, zthr_t *zthr)
4957 int64_t free_memory = arc_available_memory();
4958 static int reap_cb_check_counter = 0;
4961 * If a kmem reap is already active, don't schedule more. We must
4962 * check for this because kmem_cache_reap_soon() won't actually
4963 * block on the cache being reaped (this is to prevent callers from
4964 * becoming implicitly blocked by a system-wide kmem reap -- which,
4965 * on a system with many, many full magazines, can take minutes).
4967 if (!kmem_cache_reap_active() && free_memory < 0) {
4969 arc_no_grow = B_TRUE;
4972 * Wait at least zfs_grow_retry (default 5) seconds
4973 * before considering growing.
4975 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4977 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4978 arc_no_grow = B_TRUE;
4979 } else if (gethrtime() >= arc_growtime) {
4980 arc_no_grow = B_FALSE;
4984 * Called unconditionally every 60 seconds to reclaim unused
4985 * zstd compression and decompression context. This is done
4986 * here to avoid the need for an independent thread.
4988 if (!((reap_cb_check_counter++) % 60))
4989 zfs_zstd_cache_reap_now();
4995 * Keep enough free memory in the system by reaping the ARC's kmem
4996 * caches. To cause more slabs to be reapable, we may reduce the
4997 * target size of the cache (arc_c), causing the arc_evict_cb()
4998 * to free more buffers.
5002 arc_reap_cb(void *arg, zthr_t *zthr)
5004 int64_t free_memory;
5005 fstrans_cookie_t cookie = spl_fstrans_mark();
5008 * Kick off asynchronous kmem_reap()'s of all our caches.
5010 arc_kmem_reap_soon();
5013 * Wait at least arc_kmem_cache_reap_retry_ms between
5014 * arc_kmem_reap_soon() calls. Without this check it is possible to
5015 * end up in a situation where we spend lots of time reaping
5016 * caches, while we're near arc_c_min. Waiting here also gives the
5017 * subsequent free memory check a chance of finding that the
5018 * asynchronous reap has already freed enough memory, and we don't
5019 * need to call arc_reduce_target_size().
5021 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
5024 * Reduce the target size as needed to maintain the amount of free
5025 * memory in the system at a fraction of the arc_size (1/128th by
5026 * default). If oversubscribed (free_memory < 0) then reduce the
5027 * target arc_size by the deficit amount plus the fractional
5028 * amount. If free memory is positive but less than the fractional
5029 * amount, reduce by what is needed to hit the fractional amount.
5031 free_memory = arc_available_memory();
5034 (arc_c >> arc_shrink_shift) - free_memory;
5036 arc_reduce_target_size(to_free);
5038 spl_fstrans_unmark(cookie);
5043 * Determine the amount of memory eligible for eviction contained in the
5044 * ARC. All clean data reported by the ghost lists can always be safely
5045 * evicted. Due to arc_c_min, the same does not hold for all clean data
5046 * contained by the regular mru and mfu lists.
5048 * In the case of the regular mru and mfu lists, we need to report as
5049 * much clean data as possible, such that evicting that same reported
5050 * data will not bring arc_size below arc_c_min. Thus, in certain
5051 * circumstances, the total amount of clean data in the mru and mfu
5052 * lists might not actually be evictable.
5054 * The following two distinct cases are accounted for:
5056 * 1. The sum of the amount of dirty data contained by both the mru and
5057 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5058 * is greater than or equal to arc_c_min.
5059 * (i.e. amount of dirty data >= arc_c_min)
5061 * This is the easy case; all clean data contained by the mru and mfu
5062 * lists is evictable. Evicting all clean data can only drop arc_size
5063 * to the amount of dirty data, which is greater than arc_c_min.
5065 * 2. The sum of the amount of dirty data contained by both the mru and
5066 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5067 * is less than arc_c_min.
5068 * (i.e. arc_c_min > amount of dirty data)
5070 * 2.1. arc_size is greater than or equal arc_c_min.
5071 * (i.e. arc_size >= arc_c_min > amount of dirty data)
5073 * In this case, not all clean data from the regular mru and mfu
5074 * lists is actually evictable; we must leave enough clean data
5075 * to keep arc_size above arc_c_min. Thus, the maximum amount of
5076 * evictable data from the two lists combined, is exactly the
5077 * difference between arc_size and arc_c_min.
5079 * 2.2. arc_size is less than arc_c_min
5080 * (i.e. arc_c_min > arc_size > amount of dirty data)
5082 * In this case, none of the data contained in the mru and mfu
5083 * lists is evictable, even if it's clean. Since arc_size is
5084 * already below arc_c_min, evicting any more would only
5085 * increase this negative difference.
5088 #endif /* _KERNEL */
5091 * Adapt arc info given the number of bytes we are trying to add and
5092 * the state that we are coming from. This function is only called
5093 * when we are adding new content to the cache.
5096 arc_adapt(int bytes, arc_state_t *state)
5099 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
5100 int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size);
5101 int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size);
5105 * Adapt the target size of the MRU list:
5106 * - if we just hit in the MRU ghost list, then increase
5107 * the target size of the MRU list.
5108 * - if we just hit in the MFU ghost list, then increase
5109 * the target size of the MFU list by decreasing the
5110 * target size of the MRU list.
5112 if (state == arc_mru_ghost) {
5113 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
5114 if (!zfs_arc_p_dampener_disable)
5115 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
5117 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
5118 } else if (state == arc_mfu_ghost) {
5121 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
5122 if (!zfs_arc_p_dampener_disable)
5123 mult = MIN(mult, 10);
5125 delta = MIN(bytes * mult, arc_p);
5126 arc_p = MAX(arc_p_min, arc_p - delta);
5128 ASSERT((int64_t)arc_p >= 0);
5131 * Wake reap thread if we do not have any available memory
5133 if (arc_reclaim_needed()) {
5134 zthr_wakeup(arc_reap_zthr);
5141 if (arc_c >= arc_c_max)
5145 * If we're within (2 * maxblocksize) bytes of the target
5146 * cache size, increment the target cache size
5148 ASSERT3U(arc_c, >=, 2ULL << SPA_MAXBLOCKSHIFT);
5149 if (aggsum_upper_bound(&arc_sums.arcstat_size) >=
5150 arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
5151 atomic_add_64(&arc_c, (int64_t)bytes);
5152 if (arc_c > arc_c_max)
5154 else if (state == arc_anon)
5155 atomic_add_64(&arc_p, (int64_t)bytes);
5159 ASSERT((int64_t)arc_p >= 0);
5163 * Check if arc_size has grown past our upper threshold, determined by
5164 * zfs_arc_overflow_shift.
5166 static arc_ovf_level_t
5167 arc_is_overflowing(boolean_t use_reserve)
5169 /* Always allow at least one block of overflow */
5170 int64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5171 arc_c >> zfs_arc_overflow_shift);
5174 * We just compare the lower bound here for performance reasons. Our
5175 * primary goals are to make sure that the arc never grows without
5176 * bound, and that it can reach its maximum size. This check
5177 * accomplishes both goals. The maximum amount we could run over by is
5178 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
5179 * in the ARC. In practice, that's in the tens of MB, which is low
5180 * enough to be safe.
5182 int64_t over = aggsum_lower_bound(&arc_sums.arcstat_size) -
5183 arc_c - overflow / 2;
5186 return (over < 0 ? ARC_OVF_NONE :
5187 over < overflow ? ARC_OVF_SOME : ARC_OVF_SEVERE);
5191 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag,
5194 arc_buf_contents_t type = arc_buf_type(hdr);
5196 arc_get_data_impl(hdr, size, tag, alloc_flags);
5197 if (type == ARC_BUFC_METADATA) {
5198 return (abd_alloc(size, B_TRUE));
5200 ASSERT(type == ARC_BUFC_DATA);
5201 return (abd_alloc(size, B_FALSE));
5206 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5208 arc_buf_contents_t type = arc_buf_type(hdr);
5210 arc_get_data_impl(hdr, size, tag, ARC_HDR_DO_ADAPT);
5211 if (type == ARC_BUFC_METADATA) {
5212 return (zio_buf_alloc(size));
5214 ASSERT(type == ARC_BUFC_DATA);
5215 return (zio_data_buf_alloc(size));
5220 * Wait for the specified amount of data (in bytes) to be evicted from the
5221 * ARC, and for there to be sufficient free memory in the system. Waiting for
5222 * eviction ensures that the memory used by the ARC decreases. Waiting for
5223 * free memory ensures that the system won't run out of free pages, regardless
5224 * of ARC behavior and settings. See arc_lowmem_init().
5227 arc_wait_for_eviction(uint64_t amount, boolean_t use_reserve)
5229 switch (arc_is_overflowing(use_reserve)) {
5234 * This is a bit racy without taking arc_evict_lock, but the
5235 * worst that can happen is we either call zthr_wakeup() extra
5236 * time due to race with other thread here, or the set flag
5237 * get cleared by arc_evict_cb(), which is unlikely due to
5238 * big hysteresis, but also not important since at this level
5239 * of overflow the eviction is purely advisory. Same time
5240 * taking the global lock here every time without waiting for
5241 * the actual eviction creates a significant lock contention.
5243 if (!arc_evict_needed) {
5244 arc_evict_needed = B_TRUE;
5245 zthr_wakeup(arc_evict_zthr);
5248 case ARC_OVF_SEVERE:
5251 arc_evict_waiter_t aw;
5252 list_link_init(&aw.aew_node);
5253 cv_init(&aw.aew_cv, NULL, CV_DEFAULT, NULL);
5255 uint64_t last_count = 0;
5256 mutex_enter(&arc_evict_lock);
5257 if (!list_is_empty(&arc_evict_waiters)) {
5258 arc_evict_waiter_t *last =
5259 list_tail(&arc_evict_waiters);
5260 last_count = last->aew_count;
5261 } else if (!arc_evict_needed) {
5262 arc_evict_needed = B_TRUE;
5263 zthr_wakeup(arc_evict_zthr);
5266 * Note, the last waiter's count may be less than
5267 * arc_evict_count if we are low on memory in which
5268 * case arc_evict_state_impl() may have deferred
5269 * wakeups (but still incremented arc_evict_count).
5271 aw.aew_count = MAX(last_count, arc_evict_count) + amount;
5273 list_insert_tail(&arc_evict_waiters, &aw);
5275 arc_set_need_free();
5277 DTRACE_PROBE3(arc__wait__for__eviction,
5279 uint64_t, arc_evict_count,
5280 uint64_t, aw.aew_count);
5283 * We will be woken up either when arc_evict_count reaches
5284 * aew_count, or when the ARC is no longer overflowing and
5285 * eviction completes.
5286 * In case of "false" wakeup, we will still be on the list.
5289 cv_wait(&aw.aew_cv, &arc_evict_lock);
5290 } while (list_link_active(&aw.aew_node));
5291 mutex_exit(&arc_evict_lock);
5293 cv_destroy(&aw.aew_cv);
5299 * Allocate a block and return it to the caller. If we are hitting the
5300 * hard limit for the cache size, we must sleep, waiting for the eviction
5301 * thread to catch up. If we're past the target size but below the hard
5302 * limit, we'll only signal the reclaim thread and continue on.
5305 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag,
5308 arc_state_t *state = hdr->b_l1hdr.b_state;
5309 arc_buf_contents_t type = arc_buf_type(hdr);
5311 if (alloc_flags & ARC_HDR_DO_ADAPT)
5312 arc_adapt(size, state);
5315 * If arc_size is currently overflowing, we must be adding data
5316 * faster than we are evicting. To ensure we don't compound the
5317 * problem by adding more data and forcing arc_size to grow even
5318 * further past it's target size, we wait for the eviction thread to
5319 * make some progress. We also wait for there to be sufficient free
5320 * memory in the system, as measured by arc_free_memory().
5322 * Specifically, we wait for zfs_arc_eviction_pct percent of the
5323 * requested size to be evicted. This should be more than 100%, to
5324 * ensure that that progress is also made towards getting arc_size
5325 * under arc_c. See the comment above zfs_arc_eviction_pct.
5327 arc_wait_for_eviction(size * zfs_arc_eviction_pct / 100,
5328 alloc_flags & ARC_HDR_USE_RESERVE);
5330 VERIFY3U(hdr->b_type, ==, type);
5331 if (type == ARC_BUFC_METADATA) {
5332 arc_space_consume(size, ARC_SPACE_META);
5334 arc_space_consume(size, ARC_SPACE_DATA);
5338 * Update the state size. Note that ghost states have a
5339 * "ghost size" and so don't need to be updated.
5341 if (!GHOST_STATE(state)) {
5343 (void) zfs_refcount_add_many(&state->arcs_size, size, tag);
5346 * If this is reached via arc_read, the link is
5347 * protected by the hash lock. If reached via
5348 * arc_buf_alloc, the header should not be accessed by
5349 * any other thread. And, if reached via arc_read_done,
5350 * the hash lock will protect it if it's found in the
5351 * hash table; otherwise no other thread should be
5352 * trying to [add|remove]_reference it.
5354 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5355 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5356 (void) zfs_refcount_add_many(&state->arcs_esize[type],
5361 * If we are growing the cache, and we are adding anonymous
5362 * data, and we have outgrown arc_p, update arc_p
5364 if (aggsum_upper_bound(&arc_sums.arcstat_size) < arc_c &&
5365 hdr->b_l1hdr.b_state == arc_anon &&
5366 (zfs_refcount_count(&arc_anon->arcs_size) +
5367 zfs_refcount_count(&arc_mru->arcs_size) > arc_p))
5368 arc_p = MIN(arc_c, arc_p + size);
5373 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5375 arc_free_data_impl(hdr, size, tag);
5380 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5382 arc_buf_contents_t type = arc_buf_type(hdr);
5384 arc_free_data_impl(hdr, size, tag);
5385 if (type == ARC_BUFC_METADATA) {
5386 zio_buf_free(buf, size);
5388 ASSERT(type == ARC_BUFC_DATA);
5389 zio_data_buf_free(buf, size);
5394 * Free the arc data buffer.
5397 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5399 arc_state_t *state = hdr->b_l1hdr.b_state;
5400 arc_buf_contents_t type = arc_buf_type(hdr);
5402 /* protected by hash lock, if in the hash table */
5403 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5404 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5405 ASSERT(state != arc_anon && state != arc_l2c_only);
5407 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
5410 (void) zfs_refcount_remove_many(&state->arcs_size, size, tag);
5412 VERIFY3U(hdr->b_type, ==, type);
5413 if (type == ARC_BUFC_METADATA) {
5414 arc_space_return(size, ARC_SPACE_META);
5416 ASSERT(type == ARC_BUFC_DATA);
5417 arc_space_return(size, ARC_SPACE_DATA);
5422 * This routine is called whenever a buffer is accessed.
5423 * NOTE: the hash lock is dropped in this function.
5426 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5430 ASSERT(MUTEX_HELD(hash_lock));
5431 ASSERT(HDR_HAS_L1HDR(hdr));
5433 if (hdr->b_l1hdr.b_state == arc_anon) {
5435 * This buffer is not in the cache, and does not
5436 * appear in our "ghost" list. Add the new buffer
5440 ASSERT0(hdr->b_l1hdr.b_arc_access);
5441 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5442 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5443 arc_change_state(arc_mru, hdr, hash_lock);
5445 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5446 now = ddi_get_lbolt();
5449 * If this buffer is here because of a prefetch, then either:
5450 * - clear the flag if this is a "referencing" read
5451 * (any subsequent access will bump this into the MFU state).
5453 * - move the buffer to the head of the list if this is
5454 * another prefetch (to make it less likely to be evicted).
5456 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5457 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5458 /* link protected by hash lock */
5459 ASSERT(multilist_link_active(
5460 &hdr->b_l1hdr.b_arc_node));
5462 if (HDR_HAS_L2HDR(hdr))
5463 l2arc_hdr_arcstats_decrement_state(hdr);
5464 arc_hdr_clear_flags(hdr,
5466 ARC_FLAG_PRESCIENT_PREFETCH);
5467 hdr->b_l1hdr.b_mru_hits++;
5468 ARCSTAT_BUMP(arcstat_mru_hits);
5469 if (HDR_HAS_L2HDR(hdr))
5470 l2arc_hdr_arcstats_increment_state(hdr);
5472 hdr->b_l1hdr.b_arc_access = now;
5477 * This buffer has been "accessed" only once so far,
5478 * but it is still in the cache. Move it to the MFU
5481 if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access +
5484 * More than 125ms have passed since we
5485 * instantiated this buffer. Move it to the
5486 * most frequently used state.
5488 hdr->b_l1hdr.b_arc_access = now;
5489 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5490 arc_change_state(arc_mfu, hdr, hash_lock);
5492 hdr->b_l1hdr.b_mru_hits++;
5493 ARCSTAT_BUMP(arcstat_mru_hits);
5494 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5495 arc_state_t *new_state;
5497 * This buffer has been "accessed" recently, but
5498 * was evicted from the cache. Move it to the
5501 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5502 new_state = arc_mru;
5503 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5504 if (HDR_HAS_L2HDR(hdr))
5505 l2arc_hdr_arcstats_decrement_state(hdr);
5506 arc_hdr_clear_flags(hdr,
5508 ARC_FLAG_PRESCIENT_PREFETCH);
5509 if (HDR_HAS_L2HDR(hdr))
5510 l2arc_hdr_arcstats_increment_state(hdr);
5512 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5514 new_state = arc_mfu;
5515 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5518 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5519 arc_change_state(new_state, hdr, hash_lock);
5521 hdr->b_l1hdr.b_mru_ghost_hits++;
5522 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5523 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5525 * This buffer has been accessed more than once and is
5526 * still in the cache. Keep it in the MFU state.
5528 * NOTE: an add_reference() that occurred when we did
5529 * the arc_read() will have kicked this off the list.
5530 * If it was a prefetch, we will explicitly move it to
5531 * the head of the list now.
5534 hdr->b_l1hdr.b_mfu_hits++;
5535 ARCSTAT_BUMP(arcstat_mfu_hits);
5536 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5537 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5538 arc_state_t *new_state = arc_mfu;
5540 * This buffer has been accessed more than once but has
5541 * been evicted from the cache. Move it back to the
5545 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5547 * This is a prefetch access...
5548 * move this block back to the MRU state.
5550 new_state = arc_mru;
5553 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5554 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5555 arc_change_state(new_state, hdr, hash_lock);
5557 hdr->b_l1hdr.b_mfu_ghost_hits++;
5558 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5559 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5561 * This buffer is on the 2nd Level ARC.
5564 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5565 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5566 arc_change_state(arc_mfu, hdr, hash_lock);
5568 cmn_err(CE_PANIC, "invalid arc state 0x%p",
5569 hdr->b_l1hdr.b_state);
5574 * This routine is called by dbuf_hold() to update the arc_access() state
5575 * which otherwise would be skipped for entries in the dbuf cache.
5578 arc_buf_access(arc_buf_t *buf)
5580 mutex_enter(&buf->b_evict_lock);
5581 arc_buf_hdr_t *hdr = buf->b_hdr;
5584 * Avoid taking the hash_lock when possible as an optimization.
5585 * The header must be checked again under the hash_lock in order
5586 * to handle the case where it is concurrently being released.
5588 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5589 mutex_exit(&buf->b_evict_lock);
5593 kmutex_t *hash_lock = HDR_LOCK(hdr);
5594 mutex_enter(hash_lock);
5596 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5597 mutex_exit(hash_lock);
5598 mutex_exit(&buf->b_evict_lock);
5599 ARCSTAT_BUMP(arcstat_access_skip);
5603 mutex_exit(&buf->b_evict_lock);
5605 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5606 hdr->b_l1hdr.b_state == arc_mfu);
5608 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5609 arc_access(hdr, hash_lock);
5610 mutex_exit(hash_lock);
5612 ARCSTAT_BUMP(arcstat_hits);
5613 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr) && !HDR_PRESCIENT_PREFETCH(hdr),
5614 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5617 /* a generic arc_read_done_func_t which you can use */
5620 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5621 arc_buf_t *buf, void *arg)
5626 bcopy(buf->b_data, arg, arc_buf_size(buf));
5627 arc_buf_destroy(buf, arg);
5630 /* a generic arc_read_done_func_t */
5633 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5634 arc_buf_t *buf, void *arg)
5636 arc_buf_t **bufp = arg;
5639 ASSERT(zio == NULL || zio->io_error != 0);
5642 ASSERT(zio == NULL || zio->io_error == 0);
5644 ASSERT(buf->b_data != NULL);
5649 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5651 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5652 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5653 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5655 if (HDR_COMPRESSION_ENABLED(hdr)) {
5656 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5657 BP_GET_COMPRESS(bp));
5659 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5660 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5661 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5666 arc_read_done(zio_t *zio)
5668 blkptr_t *bp = zio->io_bp;
5669 arc_buf_hdr_t *hdr = zio->io_private;
5670 kmutex_t *hash_lock = NULL;
5671 arc_callback_t *callback_list;
5672 arc_callback_t *acb;
5673 boolean_t freeable = B_FALSE;
5676 * The hdr was inserted into hash-table and removed from lists
5677 * prior to starting I/O. We should find this header, since
5678 * it's in the hash table, and it should be legit since it's
5679 * not possible to evict it during the I/O. The only possible
5680 * reason for it not to be found is if we were freed during the
5683 if (HDR_IN_HASH_TABLE(hdr)) {
5684 arc_buf_hdr_t *found;
5686 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5687 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5688 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5689 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5690 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5692 found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock);
5694 ASSERT((found == hdr &&
5695 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5696 (found == hdr && HDR_L2_READING(hdr)));
5697 ASSERT3P(hash_lock, !=, NULL);
5700 if (BP_IS_PROTECTED(bp)) {
5701 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5702 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5703 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5704 hdr->b_crypt_hdr.b_iv);
5706 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5709 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5710 sizeof (zil_chain_t));
5711 zio_crypt_decode_mac_zil(tmpbuf,
5712 hdr->b_crypt_hdr.b_mac);
5713 abd_return_buf(zio->io_abd, tmpbuf,
5714 sizeof (zil_chain_t));
5716 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
5720 if (zio->io_error == 0) {
5721 /* byteswap if necessary */
5722 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5723 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5724 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5726 hdr->b_l1hdr.b_byteswap =
5727 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5730 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5732 if (!HDR_L2_READING(hdr)) {
5733 hdr->b_complevel = zio->io_prop.zp_complevel;
5737 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5738 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5739 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5741 callback_list = hdr->b_l1hdr.b_acb;
5742 ASSERT3P(callback_list, !=, NULL);
5744 if (hash_lock && zio->io_error == 0 &&
5745 hdr->b_l1hdr.b_state == arc_anon) {
5747 * Only call arc_access on anonymous buffers. This is because
5748 * if we've issued an I/O for an evicted buffer, we've already
5749 * called arc_access (to prevent any simultaneous readers from
5750 * getting confused).
5752 arc_access(hdr, hash_lock);
5756 * If a read request has a callback (i.e. acb_done is not NULL), then we
5757 * make a buf containing the data according to the parameters which were
5758 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5759 * aren't needlessly decompressing the data multiple times.
5761 int callback_cnt = 0;
5762 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5763 if (!acb->acb_done || acb->acb_nobuf)
5768 if (zio->io_error != 0)
5771 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5772 &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
5773 acb->acb_compressed, acb->acb_noauth, B_TRUE,
5777 * Assert non-speculative zios didn't fail because an
5778 * encryption key wasn't loaded
5780 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
5784 * If we failed to decrypt, report an error now (as the zio
5785 * layer would have done if it had done the transforms).
5787 if (error == ECKSUM) {
5788 ASSERT(BP_IS_PROTECTED(bp));
5789 error = SET_ERROR(EIO);
5790 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5791 spa_log_error(zio->io_spa, &acb->acb_zb);
5792 (void) zfs_ereport_post(
5793 FM_EREPORT_ZFS_AUTHENTICATION,
5794 zio->io_spa, NULL, &acb->acb_zb, zio, 0);
5800 * Decompression or decryption failed. Set
5801 * io_error so that when we call acb_done
5802 * (below), we will indicate that the read
5803 * failed. Note that in the unusual case
5804 * where one callback is compressed and another
5805 * uncompressed, we will mark all of them
5806 * as failed, even though the uncompressed
5807 * one can't actually fail. In this case,
5808 * the hdr will not be anonymous, because
5809 * if there are multiple callbacks, it's
5810 * because multiple threads found the same
5811 * arc buf in the hash table.
5813 zio->io_error = error;
5818 * If there are multiple callbacks, we must have the hash lock,
5819 * because the only way for multiple threads to find this hdr is
5820 * in the hash table. This ensures that if there are multiple
5821 * callbacks, the hdr is not anonymous. If it were anonymous,
5822 * we couldn't use arc_buf_destroy() in the error case below.
5824 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5826 hdr->b_l1hdr.b_acb = NULL;
5827 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5828 if (callback_cnt == 0)
5829 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
5831 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5832 callback_list != NULL);
5834 if (zio->io_error == 0) {
5835 arc_hdr_verify(hdr, zio->io_bp);
5837 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5838 if (hdr->b_l1hdr.b_state != arc_anon)
5839 arc_change_state(arc_anon, hdr, hash_lock);
5840 if (HDR_IN_HASH_TABLE(hdr))
5841 buf_hash_remove(hdr);
5842 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5846 * Broadcast before we drop the hash_lock to avoid the possibility
5847 * that the hdr (and hence the cv) might be freed before we get to
5848 * the cv_broadcast().
5850 cv_broadcast(&hdr->b_l1hdr.b_cv);
5852 if (hash_lock != NULL) {
5853 mutex_exit(hash_lock);
5856 * This block was freed while we waited for the read to
5857 * complete. It has been removed from the hash table and
5858 * moved to the anonymous state (so that it won't show up
5861 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5862 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5865 /* execute each callback and free its structure */
5866 while ((acb = callback_list) != NULL) {
5867 if (acb->acb_done != NULL) {
5868 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5870 * If arc_buf_alloc_impl() fails during
5871 * decompression, the buf will still be
5872 * allocated, and needs to be freed here.
5874 arc_buf_destroy(acb->acb_buf,
5876 acb->acb_buf = NULL;
5878 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5879 acb->acb_buf, acb->acb_private);
5882 if (acb->acb_zio_dummy != NULL) {
5883 acb->acb_zio_dummy->io_error = zio->io_error;
5884 zio_nowait(acb->acb_zio_dummy);
5887 callback_list = acb->acb_next;
5888 kmem_free(acb, sizeof (arc_callback_t));
5892 arc_hdr_destroy(hdr);
5896 * "Read" the block at the specified DVA (in bp) via the
5897 * cache. If the block is found in the cache, invoke the provided
5898 * callback immediately and return. Note that the `zio' parameter
5899 * in the callback will be NULL in this case, since no IO was
5900 * required. If the block is not in the cache pass the read request
5901 * on to the spa with a substitute callback function, so that the
5902 * requested block will be added to the cache.
5904 * If a read request arrives for a block that has a read in-progress,
5905 * either wait for the in-progress read to complete (and return the
5906 * results); or, if this is a read with a "done" func, add a record
5907 * to the read to invoke the "done" func when the read completes,
5908 * and return; or just return.
5910 * arc_read_done() will invoke all the requested "done" functions
5911 * for readers of this block.
5914 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
5915 arc_read_done_func_t *done, void *private, zio_priority_t priority,
5916 int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5918 arc_buf_hdr_t *hdr = NULL;
5919 kmutex_t *hash_lock = NULL;
5921 uint64_t guid = spa_load_guid(spa);
5922 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
5923 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
5924 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5925 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
5926 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5927 boolean_t embedded_bp = !!BP_IS_EMBEDDED(bp);
5928 boolean_t no_buf = *arc_flags & ARC_FLAG_NO_BUF;
5931 ASSERT(!embedded_bp ||
5932 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5933 ASSERT(!BP_IS_HOLE(bp));
5934 ASSERT(!BP_IS_REDACTED(bp));
5937 * Normally SPL_FSTRANS will already be set since kernel threads which
5938 * expect to call the DMU interfaces will set it when created. System
5939 * calls are similarly handled by setting/cleaning the bit in the
5940 * registered callback (module/os/.../zfs/zpl_*).
5942 * External consumers such as Lustre which call the exported DMU
5943 * interfaces may not have set SPL_FSTRANS. To avoid a deadlock
5944 * on the hash_lock always set and clear the bit.
5946 fstrans_cookie_t cookie = spl_fstrans_mark();
5949 * Verify the block pointer contents are reasonable. This should
5950 * always be the case since the blkptr is protected by a checksum.
5951 * However, if there is damage it's desirable to detect this early
5952 * and treat it as a checksum error. This allows an alternate blkptr
5953 * to be tried when one is available (e.g. ditto blocks).
5955 if (!zfs_blkptr_verify(spa, bp, zio_flags & ZIO_FLAG_CONFIG_WRITER,
5957 rc = SET_ERROR(ECKSUM);
5963 * Embedded BP's have no DVA and require no I/O to "read".
5964 * Create an anonymous arc buf to back it.
5966 hdr = buf_hash_find(guid, bp, &hash_lock);
5970 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5971 * we maintain encrypted data separately from compressed / uncompressed
5972 * data. If the user is requesting raw encrypted data and we don't have
5973 * that in the header we will read from disk to guarantee that we can
5974 * get it even if the encryption keys aren't loaded.
5976 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
5977 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
5978 arc_buf_t *buf = NULL;
5979 *arc_flags |= ARC_FLAG_CACHED;
5981 if (HDR_IO_IN_PROGRESS(hdr)) {
5982 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5984 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
5985 mutex_exit(hash_lock);
5986 ARCSTAT_BUMP(arcstat_cached_only_in_progress);
5987 rc = SET_ERROR(ENOENT);
5991 ASSERT3P(head_zio, !=, NULL);
5992 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5993 priority == ZIO_PRIORITY_SYNC_READ) {
5995 * This is a sync read that needs to wait for
5996 * an in-flight async read. Request that the
5997 * zio have its priority upgraded.
5999 zio_change_priority(head_zio, priority);
6000 DTRACE_PROBE1(arc__async__upgrade__sync,
6001 arc_buf_hdr_t *, hdr);
6002 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
6004 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
6005 arc_hdr_clear_flags(hdr,
6006 ARC_FLAG_PREDICTIVE_PREFETCH);
6009 if (*arc_flags & ARC_FLAG_WAIT) {
6010 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6011 mutex_exit(hash_lock);
6014 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6017 arc_callback_t *acb = NULL;
6019 acb = kmem_zalloc(sizeof (arc_callback_t),
6021 acb->acb_done = done;
6022 acb->acb_private = private;
6023 acb->acb_compressed = compressed_read;
6024 acb->acb_encrypted = encrypted_read;
6025 acb->acb_noauth = noauth_read;
6026 acb->acb_nobuf = no_buf;
6029 acb->acb_zio_dummy = zio_null(pio,
6030 spa, NULL, NULL, NULL, zio_flags);
6032 ASSERT3P(acb->acb_done, !=, NULL);
6033 acb->acb_zio_head = head_zio;
6034 acb->acb_next = hdr->b_l1hdr.b_acb;
6035 hdr->b_l1hdr.b_acb = acb;
6037 mutex_exit(hash_lock);
6041 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
6042 hdr->b_l1hdr.b_state == arc_mfu);
6044 if (done && !no_buf) {
6045 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
6047 * This is a demand read which does not have to
6048 * wait for i/o because we did a predictive
6049 * prefetch i/o for it, which has completed.
6052 arc__demand__hit__predictive__prefetch,
6053 arc_buf_hdr_t *, hdr);
6055 arcstat_demand_hit_predictive_prefetch);
6056 arc_hdr_clear_flags(hdr,
6057 ARC_FLAG_PREDICTIVE_PREFETCH);
6060 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
6062 arcstat_demand_hit_prescient_prefetch);
6063 arc_hdr_clear_flags(hdr,
6064 ARC_FLAG_PRESCIENT_PREFETCH);
6067 ASSERT(!embedded_bp || !BP_IS_HOLE(bp));
6069 /* Get a buf with the desired data in it. */
6070 rc = arc_buf_alloc_impl(hdr, spa, zb, private,
6071 encrypted_read, compressed_read, noauth_read,
6075 * Convert authentication and decryption errors
6076 * to EIO (and generate an ereport if needed)
6077 * before leaving the ARC.
6079 rc = SET_ERROR(EIO);
6080 if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
6081 spa_log_error(spa, zb);
6082 (void) zfs_ereport_post(
6083 FM_EREPORT_ZFS_AUTHENTICATION,
6084 spa, NULL, zb, NULL, 0);
6088 (void) remove_reference(hdr, hash_lock,
6090 arc_buf_destroy_impl(buf);
6094 /* assert any errors weren't due to unloaded keys */
6095 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
6097 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
6098 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6099 if (HDR_HAS_L2HDR(hdr))
6100 l2arc_hdr_arcstats_decrement_state(hdr);
6101 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6102 if (HDR_HAS_L2HDR(hdr))
6103 l2arc_hdr_arcstats_increment_state(hdr);
6105 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
6106 arc_access(hdr, hash_lock);
6107 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6108 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6109 if (*arc_flags & ARC_FLAG_L2CACHE)
6110 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6111 mutex_exit(hash_lock);
6112 ARCSTAT_BUMP(arcstat_hits);
6113 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6114 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
6115 data, metadata, hits);
6118 done(NULL, zb, bp, buf, private);
6120 uint64_t lsize = BP_GET_LSIZE(bp);
6121 uint64_t psize = BP_GET_PSIZE(bp);
6122 arc_callback_t *acb;
6125 boolean_t devw = B_FALSE;
6128 int alloc_flags = encrypted_read ? ARC_HDR_ALLOC_RDATA : 0;
6130 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
6131 rc = SET_ERROR(ENOENT);
6132 if (hash_lock != NULL)
6133 mutex_exit(hash_lock);
6139 * This block is not in the cache or it has
6142 arc_buf_hdr_t *exists = NULL;
6143 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
6144 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
6145 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), 0, type);
6148 hdr->b_dva = *BP_IDENTITY(bp);
6149 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
6150 exists = buf_hash_insert(hdr, &hash_lock);
6152 if (exists != NULL) {
6153 /* somebody beat us to the hash insert */
6154 mutex_exit(hash_lock);
6155 buf_discard_identity(hdr);
6156 arc_hdr_destroy(hdr);
6157 goto top; /* restart the IO request */
6159 alloc_flags |= ARC_HDR_DO_ADAPT;
6162 * This block is in the ghost cache or encrypted data
6163 * was requested and we didn't have it. If it was
6164 * L2-only (and thus didn't have an L1 hdr),
6165 * we realloc the header to add an L1 hdr.
6167 if (!HDR_HAS_L1HDR(hdr)) {
6168 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
6172 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
6173 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6174 ASSERT(!HDR_HAS_RABD(hdr));
6175 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6176 ASSERT0(zfs_refcount_count(
6177 &hdr->b_l1hdr.b_refcnt));
6178 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
6179 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
6180 } else if (HDR_IO_IN_PROGRESS(hdr)) {
6182 * If this header already had an IO in progress
6183 * and we are performing another IO to fetch
6184 * encrypted data we must wait until the first
6185 * IO completes so as not to confuse
6186 * arc_read_done(). This should be very rare
6187 * and so the performance impact shouldn't
6190 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6191 mutex_exit(hash_lock);
6196 * This is a delicate dance that we play here.
6197 * This hdr might be in the ghost list so we access
6198 * it to move it out of the ghost list before we
6199 * initiate the read. If it's a prefetch then
6200 * it won't have a callback so we'll remove the
6201 * reference that arc_buf_alloc_impl() created. We
6202 * do this after we've called arc_access() to
6203 * avoid hitting an assert in remove_reference().
6205 arc_adapt(arc_hdr_size(hdr), hdr->b_l1hdr.b_state);
6206 arc_access(hdr, hash_lock);
6209 arc_hdr_alloc_abd(hdr, alloc_flags);
6210 if (encrypted_read) {
6211 ASSERT(HDR_HAS_RABD(hdr));
6212 size = HDR_GET_PSIZE(hdr);
6213 hdr_abd = hdr->b_crypt_hdr.b_rabd;
6214 zio_flags |= ZIO_FLAG_RAW;
6216 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6217 size = arc_hdr_size(hdr);
6218 hdr_abd = hdr->b_l1hdr.b_pabd;
6220 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
6221 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6225 * For authenticated bp's, we do not ask the ZIO layer
6226 * to authenticate them since this will cause the entire
6227 * IO to fail if the key isn't loaded. Instead, we
6228 * defer authentication until arc_buf_fill(), which will
6229 * verify the data when the key is available.
6231 if (BP_IS_AUTHENTICATED(bp))
6232 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
6235 if (*arc_flags & ARC_FLAG_PREFETCH &&
6236 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6237 if (HDR_HAS_L2HDR(hdr))
6238 l2arc_hdr_arcstats_decrement_state(hdr);
6239 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6240 if (HDR_HAS_L2HDR(hdr))
6241 l2arc_hdr_arcstats_increment_state(hdr);
6243 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6244 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6245 if (*arc_flags & ARC_FLAG_L2CACHE)
6246 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6247 if (BP_IS_AUTHENTICATED(bp))
6248 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6249 if (BP_GET_LEVEL(bp) > 0)
6250 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
6251 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
6252 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
6253 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
6255 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
6256 acb->acb_done = done;
6257 acb->acb_private = private;
6258 acb->acb_compressed = compressed_read;
6259 acb->acb_encrypted = encrypted_read;
6260 acb->acb_noauth = noauth_read;
6263 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6264 hdr->b_l1hdr.b_acb = acb;
6265 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6267 if (HDR_HAS_L2HDR(hdr) &&
6268 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
6269 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
6270 addr = hdr->b_l2hdr.b_daddr;
6272 * Lock out L2ARC device removal.
6274 if (vdev_is_dead(vd) ||
6275 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
6280 * We count both async reads and scrub IOs as asynchronous so
6281 * that both can be upgraded in the event of a cache hit while
6282 * the read IO is still in-flight.
6284 if (priority == ZIO_PRIORITY_ASYNC_READ ||
6285 priority == ZIO_PRIORITY_SCRUB)
6286 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6288 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6291 * At this point, we have a level 1 cache miss or a blkptr
6292 * with embedded data. Try again in L2ARC if possible.
6294 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
6297 * Skip ARC stat bump for block pointers with embedded
6298 * data. The data are read from the blkptr itself via
6299 * decode_embedded_bp_compressed().
6302 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr,
6303 blkptr_t *, bp, uint64_t, lsize,
6304 zbookmark_phys_t *, zb);
6305 ARCSTAT_BUMP(arcstat_misses);
6306 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6307 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data,
6309 zfs_racct_read(size, 1);
6312 /* Check if the spa even has l2 configured */
6313 const boolean_t spa_has_l2 = l2arc_ndev != 0 &&
6314 spa->spa_l2cache.sav_count > 0;
6316 if (vd != NULL && spa_has_l2 && !(l2arc_norw && devw)) {
6318 * Read from the L2ARC if the following are true:
6319 * 1. The L2ARC vdev was previously cached.
6320 * 2. This buffer still has L2ARC metadata.
6321 * 3. This buffer isn't currently writing to the L2ARC.
6322 * 4. The L2ARC entry wasn't evicted, which may
6323 * also have invalidated the vdev.
6324 * 5. This isn't prefetch or l2arc_noprefetch is 0.
6326 if (HDR_HAS_L2HDR(hdr) &&
6327 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
6328 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
6329 l2arc_read_callback_t *cb;
6333 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
6334 ARCSTAT_BUMP(arcstat_l2_hits);
6335 hdr->b_l2hdr.b_hits++;
6337 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
6339 cb->l2rcb_hdr = hdr;
6342 cb->l2rcb_flags = zio_flags;
6345 * When Compressed ARC is disabled, but the
6346 * L2ARC block is compressed, arc_hdr_size()
6347 * will have returned LSIZE rather than PSIZE.
6349 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
6350 !HDR_COMPRESSION_ENABLED(hdr) &&
6351 HDR_GET_PSIZE(hdr) != 0) {
6352 size = HDR_GET_PSIZE(hdr);
6355 asize = vdev_psize_to_asize(vd, size);
6356 if (asize != size) {
6357 abd = abd_alloc_for_io(asize,
6358 HDR_ISTYPE_METADATA(hdr));
6359 cb->l2rcb_abd = abd;
6364 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6365 addr + asize <= vd->vdev_psize -
6366 VDEV_LABEL_END_SIZE);
6369 * l2arc read. The SCL_L2ARC lock will be
6370 * released by l2arc_read_done().
6371 * Issue a null zio if the underlying buffer
6372 * was squashed to zero size by compression.
6374 ASSERT3U(arc_hdr_get_compress(hdr), !=,
6375 ZIO_COMPRESS_EMPTY);
6376 rzio = zio_read_phys(pio, vd, addr,
6379 l2arc_read_done, cb, priority,
6380 zio_flags | ZIO_FLAG_DONT_CACHE |
6382 ZIO_FLAG_DONT_PROPAGATE |
6383 ZIO_FLAG_DONT_RETRY, B_FALSE);
6384 acb->acb_zio_head = rzio;
6386 if (hash_lock != NULL)
6387 mutex_exit(hash_lock);
6389 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6391 ARCSTAT_INCR(arcstat_l2_read_bytes,
6392 HDR_GET_PSIZE(hdr));
6394 if (*arc_flags & ARC_FLAG_NOWAIT) {
6399 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6400 if (zio_wait(rzio) == 0)
6403 /* l2arc read error; goto zio_read() */
6404 if (hash_lock != NULL)
6405 mutex_enter(hash_lock);
6407 DTRACE_PROBE1(l2arc__miss,
6408 arc_buf_hdr_t *, hdr);
6409 ARCSTAT_BUMP(arcstat_l2_misses);
6410 if (HDR_L2_WRITING(hdr))
6411 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6412 spa_config_exit(spa, SCL_L2ARC, vd);
6416 spa_config_exit(spa, SCL_L2ARC, vd);
6419 * Only a spa with l2 should contribute to l2
6420 * miss stats. (Including the case of having a
6421 * faulted cache device - that's also a miss.)
6425 * Skip ARC stat bump for block pointers with
6426 * embedded data. The data are read from the
6428 * decode_embedded_bp_compressed().
6431 DTRACE_PROBE1(l2arc__miss,
6432 arc_buf_hdr_t *, hdr);
6433 ARCSTAT_BUMP(arcstat_l2_misses);
6438 rzio = zio_read(pio, spa, bp, hdr_abd, size,
6439 arc_read_done, hdr, priority, zio_flags, zb);
6440 acb->acb_zio_head = rzio;
6442 if (hash_lock != NULL)
6443 mutex_exit(hash_lock);
6445 if (*arc_flags & ARC_FLAG_WAIT) {
6446 rc = zio_wait(rzio);
6450 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6455 /* embedded bps don't actually go to disk */
6457 spa_read_history_add(spa, zb, *arc_flags);
6458 spl_fstrans_unmark(cookie);
6463 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6467 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6469 p->p_private = private;
6470 list_link_init(&p->p_node);
6471 zfs_refcount_create(&p->p_refcnt);
6473 mutex_enter(&arc_prune_mtx);
6474 zfs_refcount_add(&p->p_refcnt, &arc_prune_list);
6475 list_insert_head(&arc_prune_list, p);
6476 mutex_exit(&arc_prune_mtx);
6482 arc_remove_prune_callback(arc_prune_t *p)
6484 boolean_t wait = B_FALSE;
6485 mutex_enter(&arc_prune_mtx);
6486 list_remove(&arc_prune_list, p);
6487 if (zfs_refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6489 mutex_exit(&arc_prune_mtx);
6491 /* wait for arc_prune_task to finish */
6493 taskq_wait_outstanding(arc_prune_taskq, 0);
6494 ASSERT0(zfs_refcount_count(&p->p_refcnt));
6495 zfs_refcount_destroy(&p->p_refcnt);
6496 kmem_free(p, sizeof (*p));
6500 * Notify the arc that a block was freed, and thus will never be used again.
6503 arc_freed(spa_t *spa, const blkptr_t *bp)
6506 kmutex_t *hash_lock;
6507 uint64_t guid = spa_load_guid(spa);
6509 ASSERT(!BP_IS_EMBEDDED(bp));
6511 hdr = buf_hash_find(guid, bp, &hash_lock);
6516 * We might be trying to free a block that is still doing I/O
6517 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6518 * dmu_sync-ed block). If this block is being prefetched, then it
6519 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6520 * until the I/O completes. A block may also have a reference if it is
6521 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6522 * have written the new block to its final resting place on disk but
6523 * without the dedup flag set. This would have left the hdr in the MRU
6524 * state and discoverable. When the txg finally syncs it detects that
6525 * the block was overridden in open context and issues an override I/O.
6526 * Since this is a dedup block, the override I/O will determine if the
6527 * block is already in the DDT. If so, then it will replace the io_bp
6528 * with the bp from the DDT and allow the I/O to finish. When the I/O
6529 * reaches the done callback, dbuf_write_override_done, it will
6530 * check to see if the io_bp and io_bp_override are identical.
6531 * If they are not, then it indicates that the bp was replaced with
6532 * the bp in the DDT and the override bp is freed. This allows
6533 * us to arrive here with a reference on a block that is being
6534 * freed. So if we have an I/O in progress, or a reference to
6535 * this hdr, then we don't destroy the hdr.
6537 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6538 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6539 arc_change_state(arc_anon, hdr, hash_lock);
6540 arc_hdr_destroy(hdr);
6541 mutex_exit(hash_lock);
6543 mutex_exit(hash_lock);
6549 * Release this buffer from the cache, making it an anonymous buffer. This
6550 * must be done after a read and prior to modifying the buffer contents.
6551 * If the buffer has more than one reference, we must make
6552 * a new hdr for the buffer.
6555 arc_release(arc_buf_t *buf, void *tag)
6557 arc_buf_hdr_t *hdr = buf->b_hdr;
6560 * It would be nice to assert that if its DMU metadata (level >
6561 * 0 || it's the dnode file), then it must be syncing context.
6562 * But we don't know that information at this level.
6565 mutex_enter(&buf->b_evict_lock);
6567 ASSERT(HDR_HAS_L1HDR(hdr));
6570 * We don't grab the hash lock prior to this check, because if
6571 * the buffer's header is in the arc_anon state, it won't be
6572 * linked into the hash table.
6574 if (hdr->b_l1hdr.b_state == arc_anon) {
6575 mutex_exit(&buf->b_evict_lock);
6576 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6577 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6578 ASSERT(!HDR_HAS_L2HDR(hdr));
6580 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6581 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6582 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6584 hdr->b_l1hdr.b_arc_access = 0;
6587 * If the buf is being overridden then it may already
6588 * have a hdr that is not empty.
6590 buf_discard_identity(hdr);
6596 kmutex_t *hash_lock = HDR_LOCK(hdr);
6597 mutex_enter(hash_lock);
6600 * This assignment is only valid as long as the hash_lock is
6601 * held, we must be careful not to reference state or the
6602 * b_state field after dropping the lock.
6604 arc_state_t *state = hdr->b_l1hdr.b_state;
6605 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6606 ASSERT3P(state, !=, arc_anon);
6608 /* this buffer is not on any list */
6609 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6611 if (HDR_HAS_L2HDR(hdr)) {
6612 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6615 * We have to recheck this conditional again now that
6616 * we're holding the l2ad_mtx to prevent a race with
6617 * another thread which might be concurrently calling
6618 * l2arc_evict(). In that case, l2arc_evict() might have
6619 * destroyed the header's L2 portion as we were waiting
6620 * to acquire the l2ad_mtx.
6622 if (HDR_HAS_L2HDR(hdr))
6623 arc_hdr_l2hdr_destroy(hdr);
6625 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6629 * Do we have more than one buf?
6631 if (hdr->b_l1hdr.b_bufcnt > 1) {
6632 arc_buf_hdr_t *nhdr;
6633 uint64_t spa = hdr->b_spa;
6634 uint64_t psize = HDR_GET_PSIZE(hdr);
6635 uint64_t lsize = HDR_GET_LSIZE(hdr);
6636 boolean_t protected = HDR_PROTECTED(hdr);
6637 enum zio_compress compress = arc_hdr_get_compress(hdr);
6638 arc_buf_contents_t type = arc_buf_type(hdr);
6639 VERIFY3U(hdr->b_type, ==, type);
6641 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6642 (void) remove_reference(hdr, hash_lock, tag);
6644 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6645 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6646 ASSERT(ARC_BUF_LAST(buf));
6650 * Pull the data off of this hdr and attach it to
6651 * a new anonymous hdr. Also find the last buffer
6652 * in the hdr's buffer list.
6654 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6655 ASSERT3P(lastbuf, !=, NULL);
6658 * If the current arc_buf_t and the hdr are sharing their data
6659 * buffer, then we must stop sharing that block.
6661 if (arc_buf_is_shared(buf)) {
6662 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6663 VERIFY(!arc_buf_is_shared(lastbuf));
6666 * First, sever the block sharing relationship between
6667 * buf and the arc_buf_hdr_t.
6669 arc_unshare_buf(hdr, buf);
6672 * Now we need to recreate the hdr's b_pabd. Since we
6673 * have lastbuf handy, we try to share with it, but if
6674 * we can't then we allocate a new b_pabd and copy the
6675 * data from buf into it.
6677 if (arc_can_share(hdr, lastbuf)) {
6678 arc_share_buf(hdr, lastbuf);
6680 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT);
6681 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6682 buf->b_data, psize);
6684 VERIFY3P(lastbuf->b_data, !=, NULL);
6685 } else if (HDR_SHARED_DATA(hdr)) {
6687 * Uncompressed shared buffers are always at the end
6688 * of the list. Compressed buffers don't have the
6689 * same requirements. This makes it hard to
6690 * simply assert that the lastbuf is shared so
6691 * we rely on the hdr's compression flags to determine
6692 * if we have a compressed, shared buffer.
6694 ASSERT(arc_buf_is_shared(lastbuf) ||
6695 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6696 ASSERT(!ARC_BUF_SHARED(buf));
6699 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6700 ASSERT3P(state, !=, arc_l2c_only);
6702 (void) zfs_refcount_remove_many(&state->arcs_size,
6703 arc_buf_size(buf), buf);
6705 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6706 ASSERT3P(state, !=, arc_l2c_only);
6707 (void) zfs_refcount_remove_many(
6708 &state->arcs_esize[type],
6709 arc_buf_size(buf), buf);
6712 hdr->b_l1hdr.b_bufcnt -= 1;
6713 if (ARC_BUF_ENCRYPTED(buf))
6714 hdr->b_crypt_hdr.b_ebufcnt -= 1;
6716 arc_cksum_verify(buf);
6717 arc_buf_unwatch(buf);
6719 /* if this is the last uncompressed buf free the checksum */
6720 if (!arc_hdr_has_uncompressed_buf(hdr))
6721 arc_cksum_free(hdr);
6723 mutex_exit(hash_lock);
6726 * Allocate a new hdr. The new hdr will contain a b_pabd
6727 * buffer which will be freed in arc_write().
6729 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6730 compress, hdr->b_complevel, type);
6731 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6732 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6733 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6734 VERIFY3U(nhdr->b_type, ==, type);
6735 ASSERT(!HDR_SHARED_DATA(nhdr));
6737 nhdr->b_l1hdr.b_buf = buf;
6738 nhdr->b_l1hdr.b_bufcnt = 1;
6739 if (ARC_BUF_ENCRYPTED(buf))
6740 nhdr->b_crypt_hdr.b_ebufcnt = 1;
6741 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6744 mutex_exit(&buf->b_evict_lock);
6745 (void) zfs_refcount_add_many(&arc_anon->arcs_size,
6746 arc_buf_size(buf), buf);
6748 mutex_exit(&buf->b_evict_lock);
6749 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6750 /* protected by hash lock, or hdr is on arc_anon */
6751 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6752 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6753 hdr->b_l1hdr.b_mru_hits = 0;
6754 hdr->b_l1hdr.b_mru_ghost_hits = 0;
6755 hdr->b_l1hdr.b_mfu_hits = 0;
6756 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
6757 arc_change_state(arc_anon, hdr, hash_lock);
6758 hdr->b_l1hdr.b_arc_access = 0;
6760 mutex_exit(hash_lock);
6761 buf_discard_identity(hdr);
6767 arc_released(arc_buf_t *buf)
6771 mutex_enter(&buf->b_evict_lock);
6772 released = (buf->b_data != NULL &&
6773 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6774 mutex_exit(&buf->b_evict_lock);
6780 arc_referenced(arc_buf_t *buf)
6784 mutex_enter(&buf->b_evict_lock);
6785 referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6786 mutex_exit(&buf->b_evict_lock);
6787 return (referenced);
6792 arc_write_ready(zio_t *zio)
6794 arc_write_callback_t *callback = zio->io_private;
6795 arc_buf_t *buf = callback->awcb_buf;
6796 arc_buf_hdr_t *hdr = buf->b_hdr;
6797 blkptr_t *bp = zio->io_bp;
6798 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6799 fstrans_cookie_t cookie = spl_fstrans_mark();
6801 ASSERT(HDR_HAS_L1HDR(hdr));
6802 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6803 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6806 * If we're reexecuting this zio because the pool suspended, then
6807 * cleanup any state that was previously set the first time the
6808 * callback was invoked.
6810 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6811 arc_cksum_free(hdr);
6812 arc_buf_unwatch(buf);
6813 if (hdr->b_l1hdr.b_pabd != NULL) {
6814 if (arc_buf_is_shared(buf)) {
6815 arc_unshare_buf(hdr, buf);
6817 arc_hdr_free_abd(hdr, B_FALSE);
6821 if (HDR_HAS_RABD(hdr))
6822 arc_hdr_free_abd(hdr, B_TRUE);
6824 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6825 ASSERT(!HDR_HAS_RABD(hdr));
6826 ASSERT(!HDR_SHARED_DATA(hdr));
6827 ASSERT(!arc_buf_is_shared(buf));
6829 callback->awcb_ready(zio, buf, callback->awcb_private);
6831 if (HDR_IO_IN_PROGRESS(hdr))
6832 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6834 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6836 if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr))
6837 hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp));
6839 if (BP_IS_PROTECTED(bp)) {
6840 /* ZIL blocks are written through zio_rewrite */
6841 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6842 ASSERT(HDR_PROTECTED(hdr));
6844 if (BP_SHOULD_BYTESWAP(bp)) {
6845 if (BP_GET_LEVEL(bp) > 0) {
6846 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6848 hdr->b_l1hdr.b_byteswap =
6849 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6852 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6855 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6856 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6857 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6858 hdr->b_crypt_hdr.b_iv);
6859 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6863 * If this block was written for raw encryption but the zio layer
6864 * ended up only authenticating it, adjust the buffer flags now.
6866 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6867 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6868 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6869 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6870 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6871 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6872 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6873 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6876 /* this must be done after the buffer flags are adjusted */
6877 arc_cksum_compute(buf);
6879 enum zio_compress compress;
6880 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6881 compress = ZIO_COMPRESS_OFF;
6883 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6884 compress = BP_GET_COMPRESS(bp);
6886 HDR_SET_PSIZE(hdr, psize);
6887 arc_hdr_set_compress(hdr, compress);
6888 hdr->b_complevel = zio->io_prop.zp_complevel;
6890 if (zio->io_error != 0 || psize == 0)
6894 * Fill the hdr with data. If the buffer is encrypted we have no choice
6895 * but to copy the data into b_radb. If the hdr is compressed, the data
6896 * we want is available from the zio, otherwise we can take it from
6899 * We might be able to share the buf's data with the hdr here. However,
6900 * doing so would cause the ARC to be full of linear ABDs if we write a
6901 * lot of shareable data. As a compromise, we check whether scattered
6902 * ABDs are allowed, and assume that if they are then the user wants
6903 * the ARC to be primarily filled with them regardless of the data being
6904 * written. Therefore, if they're allowed then we allocate one and copy
6905 * the data into it; otherwise, we share the data directly if we can.
6907 if (ARC_BUF_ENCRYPTED(buf)) {
6908 ASSERT3U(psize, >, 0);
6909 ASSERT(ARC_BUF_COMPRESSED(buf));
6910 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT | ARC_HDR_ALLOC_RDATA |
6911 ARC_HDR_USE_RESERVE);
6912 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6913 } else if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6915 * Ideally, we would always copy the io_abd into b_pabd, but the
6916 * user may have disabled compressed ARC, thus we must check the
6917 * hdr's compression setting rather than the io_bp's.
6919 if (BP_IS_ENCRYPTED(bp)) {
6920 ASSERT3U(psize, >, 0);
6921 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT |
6922 ARC_HDR_ALLOC_RDATA | ARC_HDR_USE_RESERVE);
6923 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6924 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
6925 !ARC_BUF_COMPRESSED(buf)) {
6926 ASSERT3U(psize, >, 0);
6927 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT |
6928 ARC_HDR_USE_RESERVE);
6929 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6931 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6932 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT |
6933 ARC_HDR_USE_RESERVE);
6934 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6938 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6939 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6940 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6942 arc_share_buf(hdr, buf);
6946 arc_hdr_verify(hdr, bp);
6947 spl_fstrans_unmark(cookie);
6951 arc_write_children_ready(zio_t *zio)
6953 arc_write_callback_t *callback = zio->io_private;
6954 arc_buf_t *buf = callback->awcb_buf;
6956 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6960 * The SPA calls this callback for each physical write that happens on behalf
6961 * of a logical write. See the comment in dbuf_write_physdone() for details.
6964 arc_write_physdone(zio_t *zio)
6966 arc_write_callback_t *cb = zio->io_private;
6967 if (cb->awcb_physdone != NULL)
6968 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6972 arc_write_done(zio_t *zio)
6974 arc_write_callback_t *callback = zio->io_private;
6975 arc_buf_t *buf = callback->awcb_buf;
6976 arc_buf_hdr_t *hdr = buf->b_hdr;
6978 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6980 if (zio->io_error == 0) {
6981 arc_hdr_verify(hdr, zio->io_bp);
6983 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6984 buf_discard_identity(hdr);
6986 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6987 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6990 ASSERT(HDR_EMPTY(hdr));
6994 * If the block to be written was all-zero or compressed enough to be
6995 * embedded in the BP, no write was performed so there will be no
6996 * dva/birth/checksum. The buffer must therefore remain anonymous
6999 if (!HDR_EMPTY(hdr)) {
7000 arc_buf_hdr_t *exists;
7001 kmutex_t *hash_lock;
7003 ASSERT3U(zio->io_error, ==, 0);
7005 arc_cksum_verify(buf);
7007 exists = buf_hash_insert(hdr, &hash_lock);
7008 if (exists != NULL) {
7010 * This can only happen if we overwrite for
7011 * sync-to-convergence, because we remove
7012 * buffers from the hash table when we arc_free().
7014 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
7015 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
7016 panic("bad overwrite, hdr=%p exists=%p",
7017 (void *)hdr, (void *)exists);
7018 ASSERT(zfs_refcount_is_zero(
7019 &exists->b_l1hdr.b_refcnt));
7020 arc_change_state(arc_anon, exists, hash_lock);
7021 arc_hdr_destroy(exists);
7022 mutex_exit(hash_lock);
7023 exists = buf_hash_insert(hdr, &hash_lock);
7024 ASSERT3P(exists, ==, NULL);
7025 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
7027 ASSERT(zio->io_prop.zp_nopwrite);
7028 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
7029 panic("bad nopwrite, hdr=%p exists=%p",
7030 (void *)hdr, (void *)exists);
7033 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
7034 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
7035 ASSERT(BP_GET_DEDUP(zio->io_bp));
7036 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
7039 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
7040 /* if it's not anon, we are doing a scrub */
7041 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
7042 arc_access(hdr, hash_lock);
7043 mutex_exit(hash_lock);
7045 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
7048 ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
7049 callback->awcb_done(zio, buf, callback->awcb_private);
7051 abd_free(zio->io_abd);
7052 kmem_free(callback, sizeof (arc_write_callback_t));
7056 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
7057 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc,
7058 const zio_prop_t *zp, arc_write_done_func_t *ready,
7059 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
7060 arc_write_done_func_t *done, void *private, zio_priority_t priority,
7061 int zio_flags, const zbookmark_phys_t *zb)
7063 arc_buf_hdr_t *hdr = buf->b_hdr;
7064 arc_write_callback_t *callback;
7066 zio_prop_t localprop = *zp;
7068 ASSERT3P(ready, !=, NULL);
7069 ASSERT3P(done, !=, NULL);
7070 ASSERT(!HDR_IO_ERROR(hdr));
7071 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
7072 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
7073 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
7075 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
7077 if (ARC_BUF_ENCRYPTED(buf)) {
7078 ASSERT(ARC_BUF_COMPRESSED(buf));
7079 localprop.zp_encrypt = B_TRUE;
7080 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
7081 localprop.zp_complevel = hdr->b_complevel;
7082 localprop.zp_byteorder =
7083 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
7084 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
7085 bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt,
7087 bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv,
7089 bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac,
7091 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
7092 localprop.zp_nopwrite = B_FALSE;
7093 localprop.zp_copies =
7094 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
7096 zio_flags |= ZIO_FLAG_RAW;
7097 } else if (ARC_BUF_COMPRESSED(buf)) {
7098 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
7099 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
7100 localprop.zp_complevel = hdr->b_complevel;
7101 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
7103 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
7104 callback->awcb_ready = ready;
7105 callback->awcb_children_ready = children_ready;
7106 callback->awcb_physdone = physdone;
7107 callback->awcb_done = done;
7108 callback->awcb_private = private;
7109 callback->awcb_buf = buf;
7112 * The hdr's b_pabd is now stale, free it now. A new data block
7113 * will be allocated when the zio pipeline calls arc_write_ready().
7115 if (hdr->b_l1hdr.b_pabd != NULL) {
7117 * If the buf is currently sharing the data block with
7118 * the hdr then we need to break that relationship here.
7119 * The hdr will remain with a NULL data pointer and the
7120 * buf will take sole ownership of the block.
7122 if (arc_buf_is_shared(buf)) {
7123 arc_unshare_buf(hdr, buf);
7125 arc_hdr_free_abd(hdr, B_FALSE);
7127 VERIFY3P(buf->b_data, !=, NULL);
7130 if (HDR_HAS_RABD(hdr))
7131 arc_hdr_free_abd(hdr, B_TRUE);
7133 if (!(zio_flags & ZIO_FLAG_RAW))
7134 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
7136 ASSERT(!arc_buf_is_shared(buf));
7137 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
7139 zio = zio_write(pio, spa, txg, bp,
7140 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
7141 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
7142 (children_ready != NULL) ? arc_write_children_ready : NULL,
7143 arc_write_physdone, arc_write_done, callback,
7144 priority, zio_flags, zb);
7150 arc_tempreserve_clear(uint64_t reserve)
7152 atomic_add_64(&arc_tempreserve, -reserve);
7153 ASSERT((int64_t)arc_tempreserve >= 0);
7157 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
7163 reserve > arc_c/4 &&
7164 reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT))
7165 arc_c = MIN(arc_c_max, reserve * 4);
7168 * Throttle when the calculated memory footprint for the TXG
7169 * exceeds the target ARC size.
7171 if (reserve > arc_c) {
7172 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
7173 return (SET_ERROR(ERESTART));
7177 * Don't count loaned bufs as in flight dirty data to prevent long
7178 * network delays from blocking transactions that are ready to be
7179 * assigned to a txg.
7182 /* assert that it has not wrapped around */
7183 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
7185 anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) -
7186 arc_loaned_bytes), 0);
7189 * Writes will, almost always, require additional memory allocations
7190 * in order to compress/encrypt/etc the data. We therefore need to
7191 * make sure that there is sufficient available memory for this.
7193 error = arc_memory_throttle(spa, reserve, txg);
7198 * Throttle writes when the amount of dirty data in the cache
7199 * gets too large. We try to keep the cache less than half full
7200 * of dirty blocks so that our sync times don't grow too large.
7202 * In the case of one pool being built on another pool, we want
7203 * to make sure we don't end up throttling the lower (backing)
7204 * pool when the upper pool is the majority contributor to dirty
7205 * data. To insure we make forward progress during throttling, we
7206 * also check the current pool's net dirty data and only throttle
7207 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
7208 * data in the cache.
7210 * Note: if two requests come in concurrently, we might let them
7211 * both succeed, when one of them should fail. Not a huge deal.
7213 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
7214 uint64_t spa_dirty_anon = spa_dirty_data(spa);
7215 uint64_t rarc_c = arc_warm ? arc_c : arc_c_max;
7216 if (total_dirty > rarc_c * zfs_arc_dirty_limit_percent / 100 &&
7217 anon_size > rarc_c * zfs_arc_anon_limit_percent / 100 &&
7218 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
7220 uint64_t meta_esize = zfs_refcount_count(
7221 &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7222 uint64_t data_esize =
7223 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7224 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
7225 "anon_data=%lluK tempreserve=%lluK rarc_c=%lluK\n",
7226 (u_longlong_t)arc_tempreserve >> 10,
7227 (u_longlong_t)meta_esize >> 10,
7228 (u_longlong_t)data_esize >> 10,
7229 (u_longlong_t)reserve >> 10,
7230 (u_longlong_t)rarc_c >> 10);
7232 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
7233 return (SET_ERROR(ERESTART));
7235 atomic_add_64(&arc_tempreserve, reserve);
7240 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
7241 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
7243 size->value.ui64 = zfs_refcount_count(&state->arcs_size);
7244 evict_data->value.ui64 =
7245 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
7246 evict_metadata->value.ui64 =
7247 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
7251 arc_kstat_update(kstat_t *ksp, int rw)
7253 arc_stats_t *as = ksp->ks_data;
7255 if (rw == KSTAT_WRITE)
7256 return (SET_ERROR(EACCES));
7258 as->arcstat_hits.value.ui64 =
7259 wmsum_value(&arc_sums.arcstat_hits);
7260 as->arcstat_misses.value.ui64 =
7261 wmsum_value(&arc_sums.arcstat_misses);
7262 as->arcstat_demand_data_hits.value.ui64 =
7263 wmsum_value(&arc_sums.arcstat_demand_data_hits);
7264 as->arcstat_demand_data_misses.value.ui64 =
7265 wmsum_value(&arc_sums.arcstat_demand_data_misses);
7266 as->arcstat_demand_metadata_hits.value.ui64 =
7267 wmsum_value(&arc_sums.arcstat_demand_metadata_hits);
7268 as->arcstat_demand_metadata_misses.value.ui64 =
7269 wmsum_value(&arc_sums.arcstat_demand_metadata_misses);
7270 as->arcstat_prefetch_data_hits.value.ui64 =
7271 wmsum_value(&arc_sums.arcstat_prefetch_data_hits);
7272 as->arcstat_prefetch_data_misses.value.ui64 =
7273 wmsum_value(&arc_sums.arcstat_prefetch_data_misses);
7274 as->arcstat_prefetch_metadata_hits.value.ui64 =
7275 wmsum_value(&arc_sums.arcstat_prefetch_metadata_hits);
7276 as->arcstat_prefetch_metadata_misses.value.ui64 =
7277 wmsum_value(&arc_sums.arcstat_prefetch_metadata_misses);
7278 as->arcstat_mru_hits.value.ui64 =
7279 wmsum_value(&arc_sums.arcstat_mru_hits);
7280 as->arcstat_mru_ghost_hits.value.ui64 =
7281 wmsum_value(&arc_sums.arcstat_mru_ghost_hits);
7282 as->arcstat_mfu_hits.value.ui64 =
7283 wmsum_value(&arc_sums.arcstat_mfu_hits);
7284 as->arcstat_mfu_ghost_hits.value.ui64 =
7285 wmsum_value(&arc_sums.arcstat_mfu_ghost_hits);
7286 as->arcstat_deleted.value.ui64 =
7287 wmsum_value(&arc_sums.arcstat_deleted);
7288 as->arcstat_mutex_miss.value.ui64 =
7289 wmsum_value(&arc_sums.arcstat_mutex_miss);
7290 as->arcstat_access_skip.value.ui64 =
7291 wmsum_value(&arc_sums.arcstat_access_skip);
7292 as->arcstat_evict_skip.value.ui64 =
7293 wmsum_value(&arc_sums.arcstat_evict_skip);
7294 as->arcstat_evict_not_enough.value.ui64 =
7295 wmsum_value(&arc_sums.arcstat_evict_not_enough);
7296 as->arcstat_evict_l2_cached.value.ui64 =
7297 wmsum_value(&arc_sums.arcstat_evict_l2_cached);
7298 as->arcstat_evict_l2_eligible.value.ui64 =
7299 wmsum_value(&arc_sums.arcstat_evict_l2_eligible);
7300 as->arcstat_evict_l2_eligible_mfu.value.ui64 =
7301 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mfu);
7302 as->arcstat_evict_l2_eligible_mru.value.ui64 =
7303 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mru);
7304 as->arcstat_evict_l2_ineligible.value.ui64 =
7305 wmsum_value(&arc_sums.arcstat_evict_l2_ineligible);
7306 as->arcstat_evict_l2_skip.value.ui64 =
7307 wmsum_value(&arc_sums.arcstat_evict_l2_skip);
7308 as->arcstat_hash_collisions.value.ui64 =
7309 wmsum_value(&arc_sums.arcstat_hash_collisions);
7310 as->arcstat_hash_chains.value.ui64 =
7311 wmsum_value(&arc_sums.arcstat_hash_chains);
7312 as->arcstat_size.value.ui64 =
7313 aggsum_value(&arc_sums.arcstat_size);
7314 as->arcstat_compressed_size.value.ui64 =
7315 wmsum_value(&arc_sums.arcstat_compressed_size);
7316 as->arcstat_uncompressed_size.value.ui64 =
7317 wmsum_value(&arc_sums.arcstat_uncompressed_size);
7318 as->arcstat_overhead_size.value.ui64 =
7319 wmsum_value(&arc_sums.arcstat_overhead_size);
7320 as->arcstat_hdr_size.value.ui64 =
7321 wmsum_value(&arc_sums.arcstat_hdr_size);
7322 as->arcstat_data_size.value.ui64 =
7323 wmsum_value(&arc_sums.arcstat_data_size);
7324 as->arcstat_metadata_size.value.ui64 =
7325 wmsum_value(&arc_sums.arcstat_metadata_size);
7326 as->arcstat_dbuf_size.value.ui64 =
7327 wmsum_value(&arc_sums.arcstat_dbuf_size);
7328 #if defined(COMPAT_FREEBSD11)
7329 as->arcstat_other_size.value.ui64 =
7330 wmsum_value(&arc_sums.arcstat_bonus_size) +
7331 aggsum_value(&arc_sums.arcstat_dnode_size) +
7332 wmsum_value(&arc_sums.arcstat_dbuf_size);
7335 arc_kstat_update_state(arc_anon,
7336 &as->arcstat_anon_size,
7337 &as->arcstat_anon_evictable_data,
7338 &as->arcstat_anon_evictable_metadata);
7339 arc_kstat_update_state(arc_mru,
7340 &as->arcstat_mru_size,
7341 &as->arcstat_mru_evictable_data,
7342 &as->arcstat_mru_evictable_metadata);
7343 arc_kstat_update_state(arc_mru_ghost,
7344 &as->arcstat_mru_ghost_size,
7345 &as->arcstat_mru_ghost_evictable_data,
7346 &as->arcstat_mru_ghost_evictable_metadata);
7347 arc_kstat_update_state(arc_mfu,
7348 &as->arcstat_mfu_size,
7349 &as->arcstat_mfu_evictable_data,
7350 &as->arcstat_mfu_evictable_metadata);
7351 arc_kstat_update_state(arc_mfu_ghost,
7352 &as->arcstat_mfu_ghost_size,
7353 &as->arcstat_mfu_ghost_evictable_data,
7354 &as->arcstat_mfu_ghost_evictable_metadata);
7356 as->arcstat_dnode_size.value.ui64 =
7357 aggsum_value(&arc_sums.arcstat_dnode_size);
7358 as->arcstat_bonus_size.value.ui64 =
7359 wmsum_value(&arc_sums.arcstat_bonus_size);
7360 as->arcstat_l2_hits.value.ui64 =
7361 wmsum_value(&arc_sums.arcstat_l2_hits);
7362 as->arcstat_l2_misses.value.ui64 =
7363 wmsum_value(&arc_sums.arcstat_l2_misses);
7364 as->arcstat_l2_prefetch_asize.value.ui64 =
7365 wmsum_value(&arc_sums.arcstat_l2_prefetch_asize);
7366 as->arcstat_l2_mru_asize.value.ui64 =
7367 wmsum_value(&arc_sums.arcstat_l2_mru_asize);
7368 as->arcstat_l2_mfu_asize.value.ui64 =
7369 wmsum_value(&arc_sums.arcstat_l2_mfu_asize);
7370 as->arcstat_l2_bufc_data_asize.value.ui64 =
7371 wmsum_value(&arc_sums.arcstat_l2_bufc_data_asize);
7372 as->arcstat_l2_bufc_metadata_asize.value.ui64 =
7373 wmsum_value(&arc_sums.arcstat_l2_bufc_metadata_asize);
7374 as->arcstat_l2_feeds.value.ui64 =
7375 wmsum_value(&arc_sums.arcstat_l2_feeds);
7376 as->arcstat_l2_rw_clash.value.ui64 =
7377 wmsum_value(&arc_sums.arcstat_l2_rw_clash);
7378 as->arcstat_l2_read_bytes.value.ui64 =
7379 wmsum_value(&arc_sums.arcstat_l2_read_bytes);
7380 as->arcstat_l2_write_bytes.value.ui64 =
7381 wmsum_value(&arc_sums.arcstat_l2_write_bytes);
7382 as->arcstat_l2_writes_sent.value.ui64 =
7383 wmsum_value(&arc_sums.arcstat_l2_writes_sent);
7384 as->arcstat_l2_writes_done.value.ui64 =
7385 wmsum_value(&arc_sums.arcstat_l2_writes_done);
7386 as->arcstat_l2_writes_error.value.ui64 =
7387 wmsum_value(&arc_sums.arcstat_l2_writes_error);
7388 as->arcstat_l2_writes_lock_retry.value.ui64 =
7389 wmsum_value(&arc_sums.arcstat_l2_writes_lock_retry);
7390 as->arcstat_l2_evict_lock_retry.value.ui64 =
7391 wmsum_value(&arc_sums.arcstat_l2_evict_lock_retry);
7392 as->arcstat_l2_evict_reading.value.ui64 =
7393 wmsum_value(&arc_sums.arcstat_l2_evict_reading);
7394 as->arcstat_l2_evict_l1cached.value.ui64 =
7395 wmsum_value(&arc_sums.arcstat_l2_evict_l1cached);
7396 as->arcstat_l2_free_on_write.value.ui64 =
7397 wmsum_value(&arc_sums.arcstat_l2_free_on_write);
7398 as->arcstat_l2_abort_lowmem.value.ui64 =
7399 wmsum_value(&arc_sums.arcstat_l2_abort_lowmem);
7400 as->arcstat_l2_cksum_bad.value.ui64 =
7401 wmsum_value(&arc_sums.arcstat_l2_cksum_bad);
7402 as->arcstat_l2_io_error.value.ui64 =
7403 wmsum_value(&arc_sums.arcstat_l2_io_error);
7404 as->arcstat_l2_lsize.value.ui64 =
7405 wmsum_value(&arc_sums.arcstat_l2_lsize);
7406 as->arcstat_l2_psize.value.ui64 =
7407 wmsum_value(&arc_sums.arcstat_l2_psize);
7408 as->arcstat_l2_hdr_size.value.ui64 =
7409 aggsum_value(&arc_sums.arcstat_l2_hdr_size);
7410 as->arcstat_l2_log_blk_writes.value.ui64 =
7411 wmsum_value(&arc_sums.arcstat_l2_log_blk_writes);
7412 as->arcstat_l2_log_blk_asize.value.ui64 =
7413 wmsum_value(&arc_sums.arcstat_l2_log_blk_asize);
7414 as->arcstat_l2_log_blk_count.value.ui64 =
7415 wmsum_value(&arc_sums.arcstat_l2_log_blk_count);
7416 as->arcstat_l2_rebuild_success.value.ui64 =
7417 wmsum_value(&arc_sums.arcstat_l2_rebuild_success);
7418 as->arcstat_l2_rebuild_abort_unsupported.value.ui64 =
7419 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7420 as->arcstat_l2_rebuild_abort_io_errors.value.ui64 =
7421 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7422 as->arcstat_l2_rebuild_abort_dh_errors.value.ui64 =
7423 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7424 as->arcstat_l2_rebuild_abort_cksum_lb_errors.value.ui64 =
7425 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7426 as->arcstat_l2_rebuild_abort_lowmem.value.ui64 =
7427 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7428 as->arcstat_l2_rebuild_size.value.ui64 =
7429 wmsum_value(&arc_sums.arcstat_l2_rebuild_size);
7430 as->arcstat_l2_rebuild_asize.value.ui64 =
7431 wmsum_value(&arc_sums.arcstat_l2_rebuild_asize);
7432 as->arcstat_l2_rebuild_bufs.value.ui64 =
7433 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs);
7434 as->arcstat_l2_rebuild_bufs_precached.value.ui64 =
7435 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7436 as->arcstat_l2_rebuild_log_blks.value.ui64 =
7437 wmsum_value(&arc_sums.arcstat_l2_rebuild_log_blks);
7438 as->arcstat_memory_throttle_count.value.ui64 =
7439 wmsum_value(&arc_sums.arcstat_memory_throttle_count);
7440 as->arcstat_memory_direct_count.value.ui64 =
7441 wmsum_value(&arc_sums.arcstat_memory_direct_count);
7442 as->arcstat_memory_indirect_count.value.ui64 =
7443 wmsum_value(&arc_sums.arcstat_memory_indirect_count);
7445 as->arcstat_memory_all_bytes.value.ui64 =
7447 as->arcstat_memory_free_bytes.value.ui64 =
7449 as->arcstat_memory_available_bytes.value.i64 =
7450 arc_available_memory();
7452 as->arcstat_prune.value.ui64 =
7453 wmsum_value(&arc_sums.arcstat_prune);
7454 as->arcstat_meta_used.value.ui64 =
7455 aggsum_value(&arc_sums.arcstat_meta_used);
7456 as->arcstat_async_upgrade_sync.value.ui64 =
7457 wmsum_value(&arc_sums.arcstat_async_upgrade_sync);
7458 as->arcstat_demand_hit_predictive_prefetch.value.ui64 =
7459 wmsum_value(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7460 as->arcstat_demand_hit_prescient_prefetch.value.ui64 =
7461 wmsum_value(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7462 as->arcstat_raw_size.value.ui64 =
7463 wmsum_value(&arc_sums.arcstat_raw_size);
7464 as->arcstat_cached_only_in_progress.value.ui64 =
7465 wmsum_value(&arc_sums.arcstat_cached_only_in_progress);
7466 as->arcstat_abd_chunk_waste_size.value.ui64 =
7467 wmsum_value(&arc_sums.arcstat_abd_chunk_waste_size);
7473 * This function *must* return indices evenly distributed between all
7474 * sublists of the multilist. This is needed due to how the ARC eviction
7475 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7476 * distributed between all sublists and uses this assumption when
7477 * deciding which sublist to evict from and how much to evict from it.
7480 arc_state_multilist_index_func(multilist_t *ml, void *obj)
7482 arc_buf_hdr_t *hdr = obj;
7485 * We rely on b_dva to generate evenly distributed index
7486 * numbers using buf_hash below. So, as an added precaution,
7487 * let's make sure we never add empty buffers to the arc lists.
7489 ASSERT(!HDR_EMPTY(hdr));
7492 * The assumption here, is the hash value for a given
7493 * arc_buf_hdr_t will remain constant throughout its lifetime
7494 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7495 * Thus, we don't need to store the header's sublist index
7496 * on insertion, as this index can be recalculated on removal.
7498 * Also, the low order bits of the hash value are thought to be
7499 * distributed evenly. Otherwise, in the case that the multilist
7500 * has a power of two number of sublists, each sublists' usage
7501 * would not be evenly distributed. In this context full 64bit
7502 * division would be a waste of time, so limit it to 32 bits.
7504 return ((unsigned int)buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
7505 multilist_get_num_sublists(ml));
7509 arc_state_l2c_multilist_index_func(multilist_t *ml, void *obj)
7511 panic("Header %p insert into arc_l2c_only %p", obj, ml);
7514 #define WARN_IF_TUNING_IGNORED(tuning, value, do_warn) do { \
7515 if ((do_warn) && (tuning) && ((tuning) != (value))) { \
7517 "ignoring tunable %s (using %llu instead)", \
7518 (#tuning), (value)); \
7523 * Called during module initialization and periodically thereafter to
7524 * apply reasonable changes to the exposed performance tunings. Can also be
7525 * called explicitly by param_set_arc_*() functions when ARC tunables are
7526 * updated manually. Non-zero zfs_* values which differ from the currently set
7527 * values will be applied.
7530 arc_tuning_update(boolean_t verbose)
7532 uint64_t allmem = arc_all_memory();
7533 unsigned long limit;
7535 /* Valid range: 32M - <arc_c_max> */
7536 if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) &&
7537 (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) &&
7538 (zfs_arc_min <= arc_c_max)) {
7539 arc_c_min = zfs_arc_min;
7540 arc_c = MAX(arc_c, arc_c_min);
7542 WARN_IF_TUNING_IGNORED(zfs_arc_min, arc_c_min, verbose);
7544 /* Valid range: 64M - <all physical memory> */
7545 if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) &&
7546 (zfs_arc_max >= MIN_ARC_MAX) && (zfs_arc_max < allmem) &&
7547 (zfs_arc_max > arc_c_min)) {
7548 arc_c_max = zfs_arc_max;
7549 arc_c = MIN(arc_c, arc_c_max);
7550 arc_p = (arc_c >> 1);
7551 if (arc_meta_limit > arc_c_max)
7552 arc_meta_limit = arc_c_max;
7553 if (arc_dnode_size_limit > arc_meta_limit)
7554 arc_dnode_size_limit = arc_meta_limit;
7556 WARN_IF_TUNING_IGNORED(zfs_arc_max, arc_c_max, verbose);
7558 /* Valid range: 16M - <arc_c_max> */
7559 if ((zfs_arc_meta_min) && (zfs_arc_meta_min != arc_meta_min) &&
7560 (zfs_arc_meta_min >= 1ULL << SPA_MAXBLOCKSHIFT) &&
7561 (zfs_arc_meta_min <= arc_c_max)) {
7562 arc_meta_min = zfs_arc_meta_min;
7563 if (arc_meta_limit < arc_meta_min)
7564 arc_meta_limit = arc_meta_min;
7565 if (arc_dnode_size_limit < arc_meta_min)
7566 arc_dnode_size_limit = arc_meta_min;
7568 WARN_IF_TUNING_IGNORED(zfs_arc_meta_min, arc_meta_min, verbose);
7570 /* Valid range: <arc_meta_min> - <arc_c_max> */
7571 limit = zfs_arc_meta_limit ? zfs_arc_meta_limit :
7572 MIN(zfs_arc_meta_limit_percent, 100) * arc_c_max / 100;
7573 if ((limit != arc_meta_limit) &&
7574 (limit >= arc_meta_min) &&
7575 (limit <= arc_c_max))
7576 arc_meta_limit = limit;
7577 WARN_IF_TUNING_IGNORED(zfs_arc_meta_limit, arc_meta_limit, verbose);
7579 /* Valid range: <arc_meta_min> - <arc_meta_limit> */
7580 limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit :
7581 MIN(zfs_arc_dnode_limit_percent, 100) * arc_meta_limit / 100;
7582 if ((limit != arc_dnode_size_limit) &&
7583 (limit >= arc_meta_min) &&
7584 (limit <= arc_meta_limit))
7585 arc_dnode_size_limit = limit;
7586 WARN_IF_TUNING_IGNORED(zfs_arc_dnode_limit, arc_dnode_size_limit,
7589 /* Valid range: 1 - N */
7590 if (zfs_arc_grow_retry)
7591 arc_grow_retry = zfs_arc_grow_retry;
7593 /* Valid range: 1 - N */
7594 if (zfs_arc_shrink_shift) {
7595 arc_shrink_shift = zfs_arc_shrink_shift;
7596 arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1);
7599 /* Valid range: 1 - N */
7600 if (zfs_arc_p_min_shift)
7601 arc_p_min_shift = zfs_arc_p_min_shift;
7603 /* Valid range: 1 - N ms */
7604 if (zfs_arc_min_prefetch_ms)
7605 arc_min_prefetch_ms = zfs_arc_min_prefetch_ms;
7607 /* Valid range: 1 - N ms */
7608 if (zfs_arc_min_prescient_prefetch_ms) {
7609 arc_min_prescient_prefetch_ms =
7610 zfs_arc_min_prescient_prefetch_ms;
7613 /* Valid range: 0 - 100 */
7614 if ((zfs_arc_lotsfree_percent >= 0) &&
7615 (zfs_arc_lotsfree_percent <= 100))
7616 arc_lotsfree_percent = zfs_arc_lotsfree_percent;
7617 WARN_IF_TUNING_IGNORED(zfs_arc_lotsfree_percent, arc_lotsfree_percent,
7620 /* Valid range: 0 - <all physical memory> */
7621 if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free))
7622 arc_sys_free = MIN(MAX(zfs_arc_sys_free, 0), allmem);
7623 WARN_IF_TUNING_IGNORED(zfs_arc_sys_free, arc_sys_free, verbose);
7627 arc_state_multilist_init(multilist_t *ml,
7628 multilist_sublist_index_func_t *index_func, int *maxcountp)
7630 multilist_create(ml, sizeof (arc_buf_hdr_t),
7631 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), index_func);
7632 *maxcountp = MAX(*maxcountp, multilist_get_num_sublists(ml));
7636 arc_state_init(void)
7638 int num_sublists = 0;
7640 arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_METADATA],
7641 arc_state_multilist_index_func, &num_sublists);
7642 arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_DATA],
7643 arc_state_multilist_index_func, &num_sublists);
7644 arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
7645 arc_state_multilist_index_func, &num_sublists);
7646 arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
7647 arc_state_multilist_index_func, &num_sublists);
7648 arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
7649 arc_state_multilist_index_func, &num_sublists);
7650 arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_DATA],
7651 arc_state_multilist_index_func, &num_sublists);
7652 arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
7653 arc_state_multilist_index_func, &num_sublists);
7654 arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
7655 arc_state_multilist_index_func, &num_sublists);
7658 * L2 headers should never be on the L2 state list since they don't
7659 * have L1 headers allocated. Special index function asserts that.
7661 arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
7662 arc_state_l2c_multilist_index_func, &num_sublists);
7663 arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
7664 arc_state_l2c_multilist_index_func, &num_sublists);
7667 * Keep track of the number of markers needed to reclaim buffers from
7668 * any ARC state. The markers will be pre-allocated so as to minimize
7669 * the number of memory allocations performed by the eviction thread.
7671 arc_state_evict_marker_count = num_sublists;
7673 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7674 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7675 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7676 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7677 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7678 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7679 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7680 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7681 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7682 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7683 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7684 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7686 zfs_refcount_create(&arc_anon->arcs_size);
7687 zfs_refcount_create(&arc_mru->arcs_size);
7688 zfs_refcount_create(&arc_mru_ghost->arcs_size);
7689 zfs_refcount_create(&arc_mfu->arcs_size);
7690 zfs_refcount_create(&arc_mfu_ghost->arcs_size);
7691 zfs_refcount_create(&arc_l2c_only->arcs_size);
7693 wmsum_init(&arc_sums.arcstat_hits, 0);
7694 wmsum_init(&arc_sums.arcstat_misses, 0);
7695 wmsum_init(&arc_sums.arcstat_demand_data_hits, 0);
7696 wmsum_init(&arc_sums.arcstat_demand_data_misses, 0);
7697 wmsum_init(&arc_sums.arcstat_demand_metadata_hits, 0);
7698 wmsum_init(&arc_sums.arcstat_demand_metadata_misses, 0);
7699 wmsum_init(&arc_sums.arcstat_prefetch_data_hits, 0);
7700 wmsum_init(&arc_sums.arcstat_prefetch_data_misses, 0);
7701 wmsum_init(&arc_sums.arcstat_prefetch_metadata_hits, 0);
7702 wmsum_init(&arc_sums.arcstat_prefetch_metadata_misses, 0);
7703 wmsum_init(&arc_sums.arcstat_mru_hits, 0);
7704 wmsum_init(&arc_sums.arcstat_mru_ghost_hits, 0);
7705 wmsum_init(&arc_sums.arcstat_mfu_hits, 0);
7706 wmsum_init(&arc_sums.arcstat_mfu_ghost_hits, 0);
7707 wmsum_init(&arc_sums.arcstat_deleted, 0);
7708 wmsum_init(&arc_sums.arcstat_mutex_miss, 0);
7709 wmsum_init(&arc_sums.arcstat_access_skip, 0);
7710 wmsum_init(&arc_sums.arcstat_evict_skip, 0);
7711 wmsum_init(&arc_sums.arcstat_evict_not_enough, 0);
7712 wmsum_init(&arc_sums.arcstat_evict_l2_cached, 0);
7713 wmsum_init(&arc_sums.arcstat_evict_l2_eligible, 0);
7714 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mfu, 0);
7715 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mru, 0);
7716 wmsum_init(&arc_sums.arcstat_evict_l2_ineligible, 0);
7717 wmsum_init(&arc_sums.arcstat_evict_l2_skip, 0);
7718 wmsum_init(&arc_sums.arcstat_hash_collisions, 0);
7719 wmsum_init(&arc_sums.arcstat_hash_chains, 0);
7720 aggsum_init(&arc_sums.arcstat_size, 0);
7721 wmsum_init(&arc_sums.arcstat_compressed_size, 0);
7722 wmsum_init(&arc_sums.arcstat_uncompressed_size, 0);
7723 wmsum_init(&arc_sums.arcstat_overhead_size, 0);
7724 wmsum_init(&arc_sums.arcstat_hdr_size, 0);
7725 wmsum_init(&arc_sums.arcstat_data_size, 0);
7726 wmsum_init(&arc_sums.arcstat_metadata_size, 0);
7727 wmsum_init(&arc_sums.arcstat_dbuf_size, 0);
7728 aggsum_init(&arc_sums.arcstat_dnode_size, 0);
7729 wmsum_init(&arc_sums.arcstat_bonus_size, 0);
7730 wmsum_init(&arc_sums.arcstat_l2_hits, 0);
7731 wmsum_init(&arc_sums.arcstat_l2_misses, 0);
7732 wmsum_init(&arc_sums.arcstat_l2_prefetch_asize, 0);
7733 wmsum_init(&arc_sums.arcstat_l2_mru_asize, 0);
7734 wmsum_init(&arc_sums.arcstat_l2_mfu_asize, 0);
7735 wmsum_init(&arc_sums.arcstat_l2_bufc_data_asize, 0);
7736 wmsum_init(&arc_sums.arcstat_l2_bufc_metadata_asize, 0);
7737 wmsum_init(&arc_sums.arcstat_l2_feeds, 0);
7738 wmsum_init(&arc_sums.arcstat_l2_rw_clash, 0);
7739 wmsum_init(&arc_sums.arcstat_l2_read_bytes, 0);
7740 wmsum_init(&arc_sums.arcstat_l2_write_bytes, 0);
7741 wmsum_init(&arc_sums.arcstat_l2_writes_sent, 0);
7742 wmsum_init(&arc_sums.arcstat_l2_writes_done, 0);
7743 wmsum_init(&arc_sums.arcstat_l2_writes_error, 0);
7744 wmsum_init(&arc_sums.arcstat_l2_writes_lock_retry, 0);
7745 wmsum_init(&arc_sums.arcstat_l2_evict_lock_retry, 0);
7746 wmsum_init(&arc_sums.arcstat_l2_evict_reading, 0);
7747 wmsum_init(&arc_sums.arcstat_l2_evict_l1cached, 0);
7748 wmsum_init(&arc_sums.arcstat_l2_free_on_write, 0);
7749 wmsum_init(&arc_sums.arcstat_l2_abort_lowmem, 0);
7750 wmsum_init(&arc_sums.arcstat_l2_cksum_bad, 0);
7751 wmsum_init(&arc_sums.arcstat_l2_io_error, 0);
7752 wmsum_init(&arc_sums.arcstat_l2_lsize, 0);
7753 wmsum_init(&arc_sums.arcstat_l2_psize, 0);
7754 aggsum_init(&arc_sums.arcstat_l2_hdr_size, 0);
7755 wmsum_init(&arc_sums.arcstat_l2_log_blk_writes, 0);
7756 wmsum_init(&arc_sums.arcstat_l2_log_blk_asize, 0);
7757 wmsum_init(&arc_sums.arcstat_l2_log_blk_count, 0);
7758 wmsum_init(&arc_sums.arcstat_l2_rebuild_success, 0);
7759 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_unsupported, 0);
7760 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_io_errors, 0);
7761 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_dh_errors, 0);
7762 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors, 0);
7763 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_lowmem, 0);
7764 wmsum_init(&arc_sums.arcstat_l2_rebuild_size, 0);
7765 wmsum_init(&arc_sums.arcstat_l2_rebuild_asize, 0);
7766 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs, 0);
7767 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs_precached, 0);
7768 wmsum_init(&arc_sums.arcstat_l2_rebuild_log_blks, 0);
7769 wmsum_init(&arc_sums.arcstat_memory_throttle_count, 0);
7770 wmsum_init(&arc_sums.arcstat_memory_direct_count, 0);
7771 wmsum_init(&arc_sums.arcstat_memory_indirect_count, 0);
7772 wmsum_init(&arc_sums.arcstat_prune, 0);
7773 aggsum_init(&arc_sums.arcstat_meta_used, 0);
7774 wmsum_init(&arc_sums.arcstat_async_upgrade_sync, 0);
7775 wmsum_init(&arc_sums.arcstat_demand_hit_predictive_prefetch, 0);
7776 wmsum_init(&arc_sums.arcstat_demand_hit_prescient_prefetch, 0);
7777 wmsum_init(&arc_sums.arcstat_raw_size, 0);
7778 wmsum_init(&arc_sums.arcstat_cached_only_in_progress, 0);
7779 wmsum_init(&arc_sums.arcstat_abd_chunk_waste_size, 0);
7781 arc_anon->arcs_state = ARC_STATE_ANON;
7782 arc_mru->arcs_state = ARC_STATE_MRU;
7783 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7784 arc_mfu->arcs_state = ARC_STATE_MFU;
7785 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7786 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7790 arc_state_fini(void)
7792 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7793 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7794 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7795 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7796 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7797 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7798 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7799 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7800 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7801 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7802 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7803 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7805 zfs_refcount_destroy(&arc_anon->arcs_size);
7806 zfs_refcount_destroy(&arc_mru->arcs_size);
7807 zfs_refcount_destroy(&arc_mru_ghost->arcs_size);
7808 zfs_refcount_destroy(&arc_mfu->arcs_size);
7809 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
7810 zfs_refcount_destroy(&arc_l2c_only->arcs_size);
7812 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
7813 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7814 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7815 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7816 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
7817 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7818 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
7819 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7820 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7821 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7823 wmsum_fini(&arc_sums.arcstat_hits);
7824 wmsum_fini(&arc_sums.arcstat_misses);
7825 wmsum_fini(&arc_sums.arcstat_demand_data_hits);
7826 wmsum_fini(&arc_sums.arcstat_demand_data_misses);
7827 wmsum_fini(&arc_sums.arcstat_demand_metadata_hits);
7828 wmsum_fini(&arc_sums.arcstat_demand_metadata_misses);
7829 wmsum_fini(&arc_sums.arcstat_prefetch_data_hits);
7830 wmsum_fini(&arc_sums.arcstat_prefetch_data_misses);
7831 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_hits);
7832 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_misses);
7833 wmsum_fini(&arc_sums.arcstat_mru_hits);
7834 wmsum_fini(&arc_sums.arcstat_mru_ghost_hits);
7835 wmsum_fini(&arc_sums.arcstat_mfu_hits);
7836 wmsum_fini(&arc_sums.arcstat_mfu_ghost_hits);
7837 wmsum_fini(&arc_sums.arcstat_deleted);
7838 wmsum_fini(&arc_sums.arcstat_mutex_miss);
7839 wmsum_fini(&arc_sums.arcstat_access_skip);
7840 wmsum_fini(&arc_sums.arcstat_evict_skip);
7841 wmsum_fini(&arc_sums.arcstat_evict_not_enough);
7842 wmsum_fini(&arc_sums.arcstat_evict_l2_cached);
7843 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible);
7844 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mfu);
7845 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mru);
7846 wmsum_fini(&arc_sums.arcstat_evict_l2_ineligible);
7847 wmsum_fini(&arc_sums.arcstat_evict_l2_skip);
7848 wmsum_fini(&arc_sums.arcstat_hash_collisions);
7849 wmsum_fini(&arc_sums.arcstat_hash_chains);
7850 aggsum_fini(&arc_sums.arcstat_size);
7851 wmsum_fini(&arc_sums.arcstat_compressed_size);
7852 wmsum_fini(&arc_sums.arcstat_uncompressed_size);
7853 wmsum_fini(&arc_sums.arcstat_overhead_size);
7854 wmsum_fini(&arc_sums.arcstat_hdr_size);
7855 wmsum_fini(&arc_sums.arcstat_data_size);
7856 wmsum_fini(&arc_sums.arcstat_metadata_size);
7857 wmsum_fini(&arc_sums.arcstat_dbuf_size);
7858 aggsum_fini(&arc_sums.arcstat_dnode_size);
7859 wmsum_fini(&arc_sums.arcstat_bonus_size);
7860 wmsum_fini(&arc_sums.arcstat_l2_hits);
7861 wmsum_fini(&arc_sums.arcstat_l2_misses);
7862 wmsum_fini(&arc_sums.arcstat_l2_prefetch_asize);
7863 wmsum_fini(&arc_sums.arcstat_l2_mru_asize);
7864 wmsum_fini(&arc_sums.arcstat_l2_mfu_asize);
7865 wmsum_fini(&arc_sums.arcstat_l2_bufc_data_asize);
7866 wmsum_fini(&arc_sums.arcstat_l2_bufc_metadata_asize);
7867 wmsum_fini(&arc_sums.arcstat_l2_feeds);
7868 wmsum_fini(&arc_sums.arcstat_l2_rw_clash);
7869 wmsum_fini(&arc_sums.arcstat_l2_read_bytes);
7870 wmsum_fini(&arc_sums.arcstat_l2_write_bytes);
7871 wmsum_fini(&arc_sums.arcstat_l2_writes_sent);
7872 wmsum_fini(&arc_sums.arcstat_l2_writes_done);
7873 wmsum_fini(&arc_sums.arcstat_l2_writes_error);
7874 wmsum_fini(&arc_sums.arcstat_l2_writes_lock_retry);
7875 wmsum_fini(&arc_sums.arcstat_l2_evict_lock_retry);
7876 wmsum_fini(&arc_sums.arcstat_l2_evict_reading);
7877 wmsum_fini(&arc_sums.arcstat_l2_evict_l1cached);
7878 wmsum_fini(&arc_sums.arcstat_l2_free_on_write);
7879 wmsum_fini(&arc_sums.arcstat_l2_abort_lowmem);
7880 wmsum_fini(&arc_sums.arcstat_l2_cksum_bad);
7881 wmsum_fini(&arc_sums.arcstat_l2_io_error);
7882 wmsum_fini(&arc_sums.arcstat_l2_lsize);
7883 wmsum_fini(&arc_sums.arcstat_l2_psize);
7884 aggsum_fini(&arc_sums.arcstat_l2_hdr_size);
7885 wmsum_fini(&arc_sums.arcstat_l2_log_blk_writes);
7886 wmsum_fini(&arc_sums.arcstat_l2_log_blk_asize);
7887 wmsum_fini(&arc_sums.arcstat_l2_log_blk_count);
7888 wmsum_fini(&arc_sums.arcstat_l2_rebuild_success);
7889 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7890 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7891 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7892 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7893 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7894 wmsum_fini(&arc_sums.arcstat_l2_rebuild_size);
7895 wmsum_fini(&arc_sums.arcstat_l2_rebuild_asize);
7896 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs);
7897 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7898 wmsum_fini(&arc_sums.arcstat_l2_rebuild_log_blks);
7899 wmsum_fini(&arc_sums.arcstat_memory_throttle_count);
7900 wmsum_fini(&arc_sums.arcstat_memory_direct_count);
7901 wmsum_fini(&arc_sums.arcstat_memory_indirect_count);
7902 wmsum_fini(&arc_sums.arcstat_prune);
7903 aggsum_fini(&arc_sums.arcstat_meta_used);
7904 wmsum_fini(&arc_sums.arcstat_async_upgrade_sync);
7905 wmsum_fini(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7906 wmsum_fini(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7907 wmsum_fini(&arc_sums.arcstat_raw_size);
7908 wmsum_fini(&arc_sums.arcstat_cached_only_in_progress);
7909 wmsum_fini(&arc_sums.arcstat_abd_chunk_waste_size);
7913 arc_target_bytes(void)
7919 arc_set_limits(uint64_t allmem)
7921 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more. */
7922 arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT);
7924 /* How to set default max varies by platform. */
7925 arc_c_max = arc_default_max(arc_c_min, allmem);
7930 uint64_t percent, allmem = arc_all_memory();
7931 mutex_init(&arc_evict_lock, NULL, MUTEX_DEFAULT, NULL);
7932 list_create(&arc_evict_waiters, sizeof (arc_evict_waiter_t),
7933 offsetof(arc_evict_waiter_t, aew_node));
7935 arc_min_prefetch_ms = 1000;
7936 arc_min_prescient_prefetch_ms = 6000;
7938 #if defined(_KERNEL)
7942 arc_set_limits(allmem);
7946 * If zfs_arc_max is non-zero at init, meaning it was set in the kernel
7947 * environment before the module was loaded, don't block setting the
7948 * maximum because it is less than arc_c_min, instead, reset arc_c_min
7950 * zfs_arc_min will be handled by arc_tuning_update().
7952 if (zfs_arc_max != 0 && zfs_arc_max >= MIN_ARC_MAX &&
7953 zfs_arc_max < allmem) {
7954 arc_c_max = zfs_arc_max;
7955 if (arc_c_min >= arc_c_max) {
7956 arc_c_min = MAX(zfs_arc_max / 2,
7957 2ULL << SPA_MAXBLOCKSHIFT);
7962 * In userland, there's only the memory pressure that we artificially
7963 * create (see arc_available_memory()). Don't let arc_c get too
7964 * small, because it can cause transactions to be larger than
7965 * arc_c, causing arc_tempreserve_space() to fail.
7967 arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT);
7971 arc_p = (arc_c >> 1);
7973 /* Set min to 1/2 of arc_c_min */
7974 arc_meta_min = 1ULL << SPA_MAXBLOCKSHIFT;
7976 * Set arc_meta_limit to a percent of arc_c_max with a floor of
7977 * arc_meta_min, and a ceiling of arc_c_max.
7979 percent = MIN(zfs_arc_meta_limit_percent, 100);
7980 arc_meta_limit = MAX(arc_meta_min, (percent * arc_c_max) / 100);
7981 percent = MIN(zfs_arc_dnode_limit_percent, 100);
7982 arc_dnode_size_limit = (percent * arc_meta_limit) / 100;
7984 /* Apply user specified tunings */
7985 arc_tuning_update(B_TRUE);
7987 /* if kmem_flags are set, lets try to use less memory */
7988 if (kmem_debugging())
7990 if (arc_c < arc_c_min)
7993 arc_register_hotplug();
7999 list_create(&arc_prune_list, sizeof (arc_prune_t),
8000 offsetof(arc_prune_t, p_node));
8001 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
8003 arc_prune_taskq = taskq_create("arc_prune", 100, defclsyspri,
8004 boot_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC |
8005 TASKQ_THREADS_CPU_PCT);
8007 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
8008 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
8010 if (arc_ksp != NULL) {
8011 arc_ksp->ks_data = &arc_stats;
8012 arc_ksp->ks_update = arc_kstat_update;
8013 kstat_install(arc_ksp);
8016 arc_state_evict_markers =
8017 arc_state_alloc_markers(arc_state_evict_marker_count);
8018 arc_evict_zthr = zthr_create("arc_evict",
8019 arc_evict_cb_check, arc_evict_cb, NULL, defclsyspri);
8020 arc_reap_zthr = zthr_create_timer("arc_reap",
8021 arc_reap_cb_check, arc_reap_cb, NULL, SEC2NSEC(1), minclsyspri);
8026 * Calculate maximum amount of dirty data per pool.
8028 * If it has been set by a module parameter, take that.
8029 * Otherwise, use a percentage of physical memory defined by
8030 * zfs_dirty_data_max_percent (default 10%) with a cap at
8031 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
8034 if (zfs_dirty_data_max_max == 0)
8035 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
8036 allmem * zfs_dirty_data_max_max_percent / 100);
8038 if (zfs_dirty_data_max_max == 0)
8039 zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024,
8040 allmem * zfs_dirty_data_max_max_percent / 100);
8043 if (zfs_dirty_data_max == 0) {
8044 zfs_dirty_data_max = allmem *
8045 zfs_dirty_data_max_percent / 100;
8046 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
8047 zfs_dirty_data_max_max);
8058 #endif /* _KERNEL */
8060 /* Use B_TRUE to ensure *all* buffers are evicted */
8061 arc_flush(NULL, B_TRUE);
8063 if (arc_ksp != NULL) {
8064 kstat_delete(arc_ksp);
8068 taskq_wait(arc_prune_taskq);
8069 taskq_destroy(arc_prune_taskq);
8071 mutex_enter(&arc_prune_mtx);
8072 while ((p = list_head(&arc_prune_list)) != NULL) {
8073 list_remove(&arc_prune_list, p);
8074 zfs_refcount_remove(&p->p_refcnt, &arc_prune_list);
8075 zfs_refcount_destroy(&p->p_refcnt);
8076 kmem_free(p, sizeof (*p));
8078 mutex_exit(&arc_prune_mtx);
8080 list_destroy(&arc_prune_list);
8081 mutex_destroy(&arc_prune_mtx);
8083 (void) zthr_cancel(arc_evict_zthr);
8084 (void) zthr_cancel(arc_reap_zthr);
8085 arc_state_free_markers(arc_state_evict_markers,
8086 arc_state_evict_marker_count);
8088 mutex_destroy(&arc_evict_lock);
8089 list_destroy(&arc_evict_waiters);
8092 * Free any buffers that were tagged for destruction. This needs
8093 * to occur before arc_state_fini() runs and destroys the aggsum
8094 * values which are updated when freeing scatter ABDs.
8096 l2arc_do_free_on_write();
8099 * buf_fini() must proceed arc_state_fini() because buf_fin() may
8100 * trigger the release of kmem magazines, which can callback to
8101 * arc_space_return() which accesses aggsums freed in act_state_fini().
8106 arc_unregister_hotplug();
8109 * We destroy the zthrs after all the ARC state has been
8110 * torn down to avoid the case of them receiving any
8111 * wakeup() signals after they are destroyed.
8113 zthr_destroy(arc_evict_zthr);
8114 zthr_destroy(arc_reap_zthr);
8116 ASSERT0(arc_loaned_bytes);
8122 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
8123 * It uses dedicated storage devices to hold cached data, which are populated
8124 * using large infrequent writes. The main role of this cache is to boost
8125 * the performance of random read workloads. The intended L2ARC devices
8126 * include short-stroked disks, solid state disks, and other media with
8127 * substantially faster read latency than disk.
8129 * +-----------------------+
8131 * +-----------------------+
8134 * l2arc_feed_thread() arc_read()
8138 * +---------------+ |
8140 * +---------------+ |
8145 * +-------+ +-------+
8147 * | cache | | cache |
8148 * +-------+ +-------+
8149 * +=========+ .-----.
8150 * : L2ARC : |-_____-|
8151 * : devices : | Disks |
8152 * +=========+ `-_____-'
8154 * Read requests are satisfied from the following sources, in order:
8157 * 2) vdev cache of L2ARC devices
8159 * 4) vdev cache of disks
8162 * Some L2ARC device types exhibit extremely slow write performance.
8163 * To accommodate for this there are some significant differences between
8164 * the L2ARC and traditional cache design:
8166 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
8167 * the ARC behave as usual, freeing buffers and placing headers on ghost
8168 * lists. The ARC does not send buffers to the L2ARC during eviction as
8169 * this would add inflated write latencies for all ARC memory pressure.
8171 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
8172 * It does this by periodically scanning buffers from the eviction-end of
8173 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
8174 * not already there. It scans until a headroom of buffers is satisfied,
8175 * which itself is a buffer for ARC eviction. If a compressible buffer is
8176 * found during scanning and selected for writing to an L2ARC device, we
8177 * temporarily boost scanning headroom during the next scan cycle to make
8178 * sure we adapt to compression effects (which might significantly reduce
8179 * the data volume we write to L2ARC). The thread that does this is
8180 * l2arc_feed_thread(), illustrated below; example sizes are included to
8181 * provide a better sense of ratio than this diagram:
8184 * +---------------------+----------+
8185 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
8186 * +---------------------+----------+ | o L2ARC eligible
8187 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
8188 * +---------------------+----------+ |
8189 * 15.9 Gbytes ^ 32 Mbytes |
8191 * l2arc_feed_thread()
8193 * l2arc write hand <--[oooo]--'
8197 * +==============================+
8198 * L2ARC dev |####|#|###|###| |####| ... |
8199 * +==============================+
8202 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
8203 * evicted, then the L2ARC has cached a buffer much sooner than it probably
8204 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
8205 * safe to say that this is an uncommon case, since buffers at the end of
8206 * the ARC lists have moved there due to inactivity.
8208 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
8209 * then the L2ARC simply misses copying some buffers. This serves as a
8210 * pressure valve to prevent heavy read workloads from both stalling the ARC
8211 * with waits and clogging the L2ARC with writes. This also helps prevent
8212 * the potential for the L2ARC to churn if it attempts to cache content too
8213 * quickly, such as during backups of the entire pool.
8215 * 5. After system boot and before the ARC has filled main memory, there are
8216 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
8217 * lists can remain mostly static. Instead of searching from tail of these
8218 * lists as pictured, the l2arc_feed_thread() will search from the list heads
8219 * for eligible buffers, greatly increasing its chance of finding them.
8221 * The L2ARC device write speed is also boosted during this time so that
8222 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
8223 * there are no L2ARC reads, and no fear of degrading read performance
8224 * through increased writes.
8226 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
8227 * the vdev queue can aggregate them into larger and fewer writes. Each
8228 * device is written to in a rotor fashion, sweeping writes through
8229 * available space then repeating.
8231 * 7. The L2ARC does not store dirty content. It never needs to flush
8232 * write buffers back to disk based storage.
8234 * 8. If an ARC buffer is written (and dirtied) which also exists in the
8235 * L2ARC, the now stale L2ARC buffer is immediately dropped.
8237 * The performance of the L2ARC can be tweaked by a number of tunables, which
8238 * may be necessary for different workloads:
8240 * l2arc_write_max max write bytes per interval
8241 * l2arc_write_boost extra write bytes during device warmup
8242 * l2arc_noprefetch skip caching prefetched buffers
8243 * l2arc_headroom number of max device writes to precache
8244 * l2arc_headroom_boost when we find compressed buffers during ARC
8245 * scanning, we multiply headroom by this
8246 * percentage factor for the next scan cycle,
8247 * since more compressed buffers are likely to
8249 * l2arc_feed_secs seconds between L2ARC writing
8251 * Tunables may be removed or added as future performance improvements are
8252 * integrated, and also may become zpool properties.
8254 * There are three key functions that control how the L2ARC warms up:
8256 * l2arc_write_eligible() check if a buffer is eligible to cache
8257 * l2arc_write_size() calculate how much to write
8258 * l2arc_write_interval() calculate sleep delay between writes
8260 * These three functions determine what to write, how much, and how quickly
8263 * L2ARC persistence:
8265 * When writing buffers to L2ARC, we periodically add some metadata to
8266 * make sure we can pick them up after reboot, thus dramatically reducing
8267 * the impact that any downtime has on the performance of storage systems
8268 * with large caches.
8270 * The implementation works fairly simply by integrating the following two
8273 * *) When writing to the L2ARC, we occasionally write a "l2arc log block",
8274 * which is an additional piece of metadata which describes what's been
8275 * written. This allows us to rebuild the arc_buf_hdr_t structures of the
8276 * main ARC buffers. There are 2 linked-lists of log blocks headed by
8277 * dh_start_lbps[2]. We alternate which chain we append to, so they are
8278 * time-wise and offset-wise interleaved, but that is an optimization rather
8279 * than for correctness. The log block also includes a pointer to the
8280 * previous block in its chain.
8282 * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device
8283 * for our header bookkeeping purposes. This contains a device header,
8284 * which contains our top-level reference structures. We update it each
8285 * time we write a new log block, so that we're able to locate it in the
8286 * L2ARC device. If this write results in an inconsistent device header
8287 * (e.g. due to power failure), we detect this by verifying the header's
8288 * checksum and simply fail to reconstruct the L2ARC after reboot.
8290 * Implementation diagram:
8292 * +=== L2ARC device (not to scale) ======================================+
8293 * | ___two newest log block pointers__.__________ |
8294 * | / \dh_start_lbps[1] |
8295 * | / \ \dh_start_lbps[0]|
8297 * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---|
8298 * || hdr| ^ /^ /^ / / |
8299 * |+------+ ...--\-------/ \-----/--\------/ / |
8300 * | \--------------/ \--------------/ |
8301 * +======================================================================+
8303 * As can be seen on the diagram, rather than using a simple linked list,
8304 * we use a pair of linked lists with alternating elements. This is a
8305 * performance enhancement due to the fact that we only find out the
8306 * address of the next log block access once the current block has been
8307 * completely read in. Obviously, this hurts performance, because we'd be
8308 * keeping the device's I/O queue at only a 1 operation deep, thus
8309 * incurring a large amount of I/O round-trip latency. Having two lists
8310 * allows us to fetch two log blocks ahead of where we are currently
8311 * rebuilding L2ARC buffers.
8313 * On-device data structures:
8315 * L2ARC device header: l2arc_dev_hdr_phys_t
8316 * L2ARC log block: l2arc_log_blk_phys_t
8318 * L2ARC reconstruction:
8320 * When writing data, we simply write in the standard rotary fashion,
8321 * evicting buffers as we go and simply writing new data over them (writing
8322 * a new log block every now and then). This obviously means that once we
8323 * loop around the end of the device, we will start cutting into an already
8324 * committed log block (and its referenced data buffers), like so:
8326 * current write head__ __old tail
8329 * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |-->
8330 * ^ ^^^^^^^^^___________________________________
8332 * <<nextwrite>> may overwrite this blk and/or its bufs --'
8334 * When importing the pool, we detect this situation and use it to stop
8335 * our scanning process (see l2arc_rebuild).
8337 * There is one significant caveat to consider when rebuilding ARC contents
8338 * from an L2ARC device: what about invalidated buffers? Given the above
8339 * construction, we cannot update blocks which we've already written to amend
8340 * them to remove buffers which were invalidated. Thus, during reconstruction,
8341 * we might be populating the cache with buffers for data that's not on the
8342 * main pool anymore, or may have been overwritten!
8344 * As it turns out, this isn't a problem. Every arc_read request includes
8345 * both the DVA and, crucially, the birth TXG of the BP the caller is
8346 * looking for. So even if the cache were populated by completely rotten
8347 * blocks for data that had been long deleted and/or overwritten, we'll
8348 * never actually return bad data from the cache, since the DVA with the
8349 * birth TXG uniquely identify a block in space and time - once created,
8350 * a block is immutable on disk. The worst thing we have done is wasted
8351 * some time and memory at l2arc rebuild to reconstruct outdated ARC
8352 * entries that will get dropped from the l2arc as it is being updated
8355 * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write
8356 * hand are not restored. This is done by saving the offset (in bytes)
8357 * l2arc_evict() has evicted to in the L2ARC device header and taking it
8358 * into account when restoring buffers.
8362 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
8365 * A buffer is *not* eligible for the L2ARC if it:
8366 * 1. belongs to a different spa.
8367 * 2. is already cached on the L2ARC.
8368 * 3. has an I/O in progress (it may be an incomplete read).
8369 * 4. is flagged not eligible (zfs property).
8371 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
8372 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
8379 l2arc_write_size(l2arc_dev_t *dev)
8381 uint64_t size, dev_size, tsize;
8384 * Make sure our globals have meaningful values in case the user
8387 size = l2arc_write_max;
8389 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
8390 "be greater than zero, resetting it to the default (%d)",
8392 size = l2arc_write_max = L2ARC_WRITE_SIZE;
8395 if (arc_warm == B_FALSE)
8396 size += l2arc_write_boost;
8399 * Make sure the write size does not exceed the size of the cache
8400 * device. This is important in l2arc_evict(), otherwise infinite
8401 * iteration can occur.
8403 dev_size = dev->l2ad_end - dev->l2ad_start;
8404 tsize = size + l2arc_log_blk_overhead(size, dev);
8405 if (dev->l2ad_vdev->vdev_has_trim && l2arc_trim_ahead > 0)
8406 tsize += MAX(64 * 1024 * 1024,
8407 (tsize * l2arc_trim_ahead) / 100);
8409 if (tsize >= dev_size) {
8410 cmn_err(CE_NOTE, "l2arc_write_max or l2arc_write_boost "
8411 "plus the overhead of log blocks (persistent L2ARC, "
8412 "%llu bytes) exceeds the size of the cache device "
8413 "(guid %llu), resetting them to the default (%d)",
8414 l2arc_log_blk_overhead(size, dev),
8415 dev->l2ad_vdev->vdev_guid, L2ARC_WRITE_SIZE);
8416 size = l2arc_write_max = l2arc_write_boost = L2ARC_WRITE_SIZE;
8418 if (arc_warm == B_FALSE)
8419 size += l2arc_write_boost;
8427 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
8429 clock_t interval, next, now;
8432 * If the ARC lists are busy, increase our write rate; if the
8433 * lists are stale, idle back. This is achieved by checking
8434 * how much we previously wrote - if it was more than half of
8435 * what we wanted, schedule the next write much sooner.
8437 if (l2arc_feed_again && wrote > (wanted / 2))
8438 interval = (hz * l2arc_feed_min_ms) / 1000;
8440 interval = hz * l2arc_feed_secs;
8442 now = ddi_get_lbolt();
8443 next = MAX(now, MIN(now + interval, began + interval));
8449 * Cycle through L2ARC devices. This is how L2ARC load balances.
8450 * If a device is returned, this also returns holding the spa config lock.
8452 static l2arc_dev_t *
8453 l2arc_dev_get_next(void)
8455 l2arc_dev_t *first, *next = NULL;
8458 * Lock out the removal of spas (spa_namespace_lock), then removal
8459 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
8460 * both locks will be dropped and a spa config lock held instead.
8462 mutex_enter(&spa_namespace_lock);
8463 mutex_enter(&l2arc_dev_mtx);
8465 /* if there are no vdevs, there is nothing to do */
8466 if (l2arc_ndev == 0)
8470 next = l2arc_dev_last;
8472 /* loop around the list looking for a non-faulted vdev */
8474 next = list_head(l2arc_dev_list);
8476 next = list_next(l2arc_dev_list, next);
8478 next = list_head(l2arc_dev_list);
8481 /* if we have come back to the start, bail out */
8484 else if (next == first)
8487 } while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8488 next->l2ad_trim_all);
8490 /* if we were unable to find any usable vdevs, return NULL */
8491 if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8492 next->l2ad_trim_all)
8495 l2arc_dev_last = next;
8498 mutex_exit(&l2arc_dev_mtx);
8501 * Grab the config lock to prevent the 'next' device from being
8502 * removed while we are writing to it.
8505 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
8506 mutex_exit(&spa_namespace_lock);
8512 * Free buffers that were tagged for destruction.
8515 l2arc_do_free_on_write(void)
8518 l2arc_data_free_t *df, *df_prev;
8520 mutex_enter(&l2arc_free_on_write_mtx);
8521 buflist = l2arc_free_on_write;
8523 for (df = list_tail(buflist); df; df = df_prev) {
8524 df_prev = list_prev(buflist, df);
8525 ASSERT3P(df->l2df_abd, !=, NULL);
8526 abd_free(df->l2df_abd);
8527 list_remove(buflist, df);
8528 kmem_free(df, sizeof (l2arc_data_free_t));
8531 mutex_exit(&l2arc_free_on_write_mtx);
8535 * A write to a cache device has completed. Update all headers to allow
8536 * reads from these buffers to begin.
8539 l2arc_write_done(zio_t *zio)
8541 l2arc_write_callback_t *cb;
8542 l2arc_lb_abd_buf_t *abd_buf;
8543 l2arc_lb_ptr_buf_t *lb_ptr_buf;
8545 l2arc_dev_hdr_phys_t *l2dhdr;
8547 arc_buf_hdr_t *head, *hdr, *hdr_prev;
8548 kmutex_t *hash_lock;
8549 int64_t bytes_dropped = 0;
8551 cb = zio->io_private;
8552 ASSERT3P(cb, !=, NULL);
8553 dev = cb->l2wcb_dev;
8554 l2dhdr = dev->l2ad_dev_hdr;
8555 ASSERT3P(dev, !=, NULL);
8556 head = cb->l2wcb_head;
8557 ASSERT3P(head, !=, NULL);
8558 buflist = &dev->l2ad_buflist;
8559 ASSERT3P(buflist, !=, NULL);
8560 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
8561 l2arc_write_callback_t *, cb);
8564 * All writes completed, or an error was hit.
8567 mutex_enter(&dev->l2ad_mtx);
8568 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
8569 hdr_prev = list_prev(buflist, hdr);
8571 hash_lock = HDR_LOCK(hdr);
8574 * We cannot use mutex_enter or else we can deadlock
8575 * with l2arc_write_buffers (due to swapping the order
8576 * the hash lock and l2ad_mtx are taken).
8578 if (!mutex_tryenter(hash_lock)) {
8580 * Missed the hash lock. We must retry so we
8581 * don't leave the ARC_FLAG_L2_WRITING bit set.
8583 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
8586 * We don't want to rescan the headers we've
8587 * already marked as having been written out, so
8588 * we reinsert the head node so we can pick up
8589 * where we left off.
8591 list_remove(buflist, head);
8592 list_insert_after(buflist, hdr, head);
8594 mutex_exit(&dev->l2ad_mtx);
8597 * We wait for the hash lock to become available
8598 * to try and prevent busy waiting, and increase
8599 * the chance we'll be able to acquire the lock
8600 * the next time around.
8602 mutex_enter(hash_lock);
8603 mutex_exit(hash_lock);
8608 * We could not have been moved into the arc_l2c_only
8609 * state while in-flight due to our ARC_FLAG_L2_WRITING
8610 * bit being set. Let's just ensure that's being enforced.
8612 ASSERT(HDR_HAS_L1HDR(hdr));
8615 * Skipped - drop L2ARC entry and mark the header as no
8616 * longer L2 eligibile.
8618 if (zio->io_error != 0) {
8620 * Error - drop L2ARC entry.
8622 list_remove(buflist, hdr);
8623 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
8625 uint64_t psize = HDR_GET_PSIZE(hdr);
8626 l2arc_hdr_arcstats_decrement(hdr);
8629 vdev_psize_to_asize(dev->l2ad_vdev, psize);
8630 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
8631 arc_hdr_size(hdr), hdr);
8635 * Allow ARC to begin reads and ghost list evictions to
8638 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
8640 mutex_exit(hash_lock);
8644 * Free the allocated abd buffers for writing the log blocks.
8645 * If the zio failed reclaim the allocated space and remove the
8646 * pointers to these log blocks from the log block pointer list
8647 * of the L2ARC device.
8649 while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) {
8650 abd_free(abd_buf->abd);
8651 zio_buf_free(abd_buf, sizeof (*abd_buf));
8652 if (zio->io_error != 0) {
8653 lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list);
8655 * L2BLK_GET_PSIZE returns aligned size for log
8659 L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop);
8660 bytes_dropped += asize;
8661 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8662 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8663 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8665 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8666 kmem_free(lb_ptr_buf->lb_ptr,
8667 sizeof (l2arc_log_blkptr_t));
8668 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8671 list_destroy(&cb->l2wcb_abd_list);
8673 if (zio->io_error != 0) {
8674 ARCSTAT_BUMP(arcstat_l2_writes_error);
8677 * Restore the lbps array in the header to its previous state.
8678 * If the list of log block pointers is empty, zero out the
8679 * log block pointers in the device header.
8681 lb_ptr_buf = list_head(&dev->l2ad_lbptr_list);
8682 for (int i = 0; i < 2; i++) {
8683 if (lb_ptr_buf == NULL) {
8685 * If the list is empty zero out the device
8686 * header. Otherwise zero out the second log
8687 * block pointer in the header.
8690 bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
8692 bzero(&l2dhdr->dh_start_lbps[i],
8693 sizeof (l2arc_log_blkptr_t));
8697 bcopy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[i],
8698 sizeof (l2arc_log_blkptr_t));
8699 lb_ptr_buf = list_next(&dev->l2ad_lbptr_list,
8704 ARCSTAT_BUMP(arcstat_l2_writes_done);
8705 list_remove(buflist, head);
8706 ASSERT(!HDR_HAS_L1HDR(head));
8707 kmem_cache_free(hdr_l2only_cache, head);
8708 mutex_exit(&dev->l2ad_mtx);
8710 ASSERT(dev->l2ad_vdev != NULL);
8711 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
8713 l2arc_do_free_on_write();
8715 kmem_free(cb, sizeof (l2arc_write_callback_t));
8719 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
8722 spa_t *spa = zio->io_spa;
8723 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
8724 blkptr_t *bp = zio->io_bp;
8725 uint8_t salt[ZIO_DATA_SALT_LEN];
8726 uint8_t iv[ZIO_DATA_IV_LEN];
8727 uint8_t mac[ZIO_DATA_MAC_LEN];
8728 boolean_t no_crypt = B_FALSE;
8731 * ZIL data is never be written to the L2ARC, so we don't need
8732 * special handling for its unique MAC storage.
8734 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
8735 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
8736 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8739 * If the data was encrypted, decrypt it now. Note that
8740 * we must check the bp here and not the hdr, since the
8741 * hdr does not have its encryption parameters updated
8742 * until arc_read_done().
8744 if (BP_IS_ENCRYPTED(bp)) {
8745 abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8746 ARC_HDR_DO_ADAPT | ARC_HDR_USE_RESERVE);
8748 zio_crypt_decode_params_bp(bp, salt, iv);
8749 zio_crypt_decode_mac_bp(bp, mac);
8751 ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
8752 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
8753 salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
8754 hdr->b_l1hdr.b_pabd, &no_crypt);
8756 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8761 * If we actually performed decryption, replace b_pabd
8762 * with the decrypted data. Otherwise we can just throw
8763 * our decryption buffer away.
8766 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8767 arc_hdr_size(hdr), hdr);
8768 hdr->b_l1hdr.b_pabd = eabd;
8771 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8776 * If the L2ARC block was compressed, but ARC compression
8777 * is disabled we decompress the data into a new buffer and
8778 * replace the existing data.
8780 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8781 !HDR_COMPRESSION_ENABLED(hdr)) {
8782 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8783 ARC_HDR_DO_ADAPT | ARC_HDR_USE_RESERVE);
8784 void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
8786 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
8787 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
8788 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
8790 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8791 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
8795 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8796 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8797 arc_hdr_size(hdr), hdr);
8798 hdr->b_l1hdr.b_pabd = cabd;
8800 zio->io_size = HDR_GET_LSIZE(hdr);
8811 * A read to a cache device completed. Validate buffer contents before
8812 * handing over to the regular ARC routines.
8815 l2arc_read_done(zio_t *zio)
8818 l2arc_read_callback_t *cb = zio->io_private;
8820 kmutex_t *hash_lock;
8821 boolean_t valid_cksum;
8822 boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
8823 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
8825 ASSERT3P(zio->io_vd, !=, NULL);
8826 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
8828 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
8830 ASSERT3P(cb, !=, NULL);
8831 hdr = cb->l2rcb_hdr;
8832 ASSERT3P(hdr, !=, NULL);
8834 hash_lock = HDR_LOCK(hdr);
8835 mutex_enter(hash_lock);
8836 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
8839 * If the data was read into a temporary buffer,
8840 * move it and free the buffer.
8842 if (cb->l2rcb_abd != NULL) {
8843 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
8844 if (zio->io_error == 0) {
8846 abd_copy(hdr->b_crypt_hdr.b_rabd,
8847 cb->l2rcb_abd, arc_hdr_size(hdr));
8849 abd_copy(hdr->b_l1hdr.b_pabd,
8850 cb->l2rcb_abd, arc_hdr_size(hdr));
8855 * The following must be done regardless of whether
8856 * there was an error:
8857 * - free the temporary buffer
8858 * - point zio to the real ARC buffer
8859 * - set zio size accordingly
8860 * These are required because zio is either re-used for
8861 * an I/O of the block in the case of the error
8862 * or the zio is passed to arc_read_done() and it
8865 abd_free(cb->l2rcb_abd);
8866 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
8869 ASSERT(HDR_HAS_RABD(hdr));
8870 zio->io_abd = zio->io_orig_abd =
8871 hdr->b_crypt_hdr.b_rabd;
8873 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8874 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8878 ASSERT3P(zio->io_abd, !=, NULL);
8881 * Check this survived the L2ARC journey.
8883 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8884 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8885 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
8886 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
8887 zio->io_prop.zp_complevel = hdr->b_complevel;
8889 valid_cksum = arc_cksum_is_equal(hdr, zio);
8892 * b_rabd will always match the data as it exists on disk if it is
8893 * being used. Therefore if we are reading into b_rabd we do not
8894 * attempt to untransform the data.
8896 if (valid_cksum && !using_rdata)
8897 tfm_error = l2arc_untransform(zio, cb);
8899 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8900 !HDR_L2_EVICTED(hdr)) {
8901 mutex_exit(hash_lock);
8902 zio->io_private = hdr;
8906 * Buffer didn't survive caching. Increment stats and
8907 * reissue to the original storage device.
8909 if (zio->io_error != 0) {
8910 ARCSTAT_BUMP(arcstat_l2_io_error);
8912 zio->io_error = SET_ERROR(EIO);
8914 if (!valid_cksum || tfm_error != 0)
8915 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8918 * If there's no waiter, issue an async i/o to the primary
8919 * storage now. If there *is* a waiter, the caller must
8920 * issue the i/o in a context where it's OK to block.
8922 if (zio->io_waiter == NULL) {
8923 zio_t *pio = zio_unique_parent(zio);
8924 void *abd = (using_rdata) ?
8925 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8927 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8929 zio = zio_read(pio, zio->io_spa, zio->io_bp,
8930 abd, zio->io_size, arc_read_done,
8931 hdr, zio->io_priority, cb->l2rcb_flags,
8935 * Original ZIO will be freed, so we need to update
8936 * ARC header with the new ZIO pointer to be used
8937 * by zio_change_priority() in arc_read().
8939 for (struct arc_callback *acb = hdr->b_l1hdr.b_acb;
8940 acb != NULL; acb = acb->acb_next)
8941 acb->acb_zio_head = zio;
8943 mutex_exit(hash_lock);
8946 mutex_exit(hash_lock);
8950 kmem_free(cb, sizeof (l2arc_read_callback_t));
8954 * This is the list priority from which the L2ARC will search for pages to
8955 * cache. This is used within loops (0..3) to cycle through lists in the
8956 * desired order. This order can have a significant effect on cache
8959 * Currently the metadata lists are hit first, MFU then MRU, followed by
8960 * the data lists. This function returns a locked list, and also returns
8963 static multilist_sublist_t *
8964 l2arc_sublist_lock(int list_num)
8966 multilist_t *ml = NULL;
8969 ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
8973 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
8976 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
8979 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
8982 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
8989 * Return a randomly-selected sublist. This is acceptable
8990 * because the caller feeds only a little bit of data for each
8991 * call (8MB). Subsequent calls will result in different
8992 * sublists being selected.
8994 idx = multilist_get_random_index(ml);
8995 return (multilist_sublist_lock(ml, idx));
8999 * Calculates the maximum overhead of L2ARC metadata log blocks for a given
9000 * L2ARC write size. l2arc_evict and l2arc_write_size need to include this
9001 * overhead in processing to make sure there is enough headroom available
9002 * when writing buffers.
9004 static inline uint64_t
9005 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev)
9007 if (dev->l2ad_log_entries == 0) {
9010 uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT;
9012 uint64_t log_blocks = (log_entries +
9013 dev->l2ad_log_entries - 1) /
9014 dev->l2ad_log_entries;
9016 return (vdev_psize_to_asize(dev->l2ad_vdev,
9017 sizeof (l2arc_log_blk_phys_t)) * log_blocks);
9022 * Evict buffers from the device write hand to the distance specified in
9023 * bytes. This distance may span populated buffers, it may span nothing.
9024 * This is clearing a region on the L2ARC device ready for writing.
9025 * If the 'all' boolean is set, every buffer is evicted.
9028 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
9031 arc_buf_hdr_t *hdr, *hdr_prev;
9032 kmutex_t *hash_lock;
9034 l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev;
9035 vdev_t *vd = dev->l2ad_vdev;
9038 buflist = &dev->l2ad_buflist;
9041 * We need to add in the worst case scenario of log block overhead.
9043 distance += l2arc_log_blk_overhead(distance, dev);
9044 if (vd->vdev_has_trim && l2arc_trim_ahead > 0) {
9046 * Trim ahead of the write size 64MB or (l2arc_trim_ahead/100)
9047 * times the write size, whichever is greater.
9049 distance += MAX(64 * 1024 * 1024,
9050 (distance * l2arc_trim_ahead) / 100);
9055 if (dev->l2ad_hand >= (dev->l2ad_end - distance)) {
9057 * When there is no space to accommodate upcoming writes,
9058 * evict to the end. Then bump the write and evict hands
9059 * to the start and iterate. This iteration does not
9060 * happen indefinitely as we make sure in
9061 * l2arc_write_size() that when the write hand is reset,
9062 * the write size does not exceed the end of the device.
9065 taddr = dev->l2ad_end;
9067 taddr = dev->l2ad_hand + distance;
9069 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
9070 uint64_t, taddr, boolean_t, all);
9074 * This check has to be placed after deciding whether to
9077 if (dev->l2ad_first) {
9079 * This is the first sweep through the device. There is
9080 * nothing to evict. We have already trimmmed the
9086 * Trim the space to be evicted.
9088 if (vd->vdev_has_trim && dev->l2ad_evict < taddr &&
9089 l2arc_trim_ahead > 0) {
9091 * We have to drop the spa_config lock because
9092 * vdev_trim_range() will acquire it.
9093 * l2ad_evict already accounts for the label
9094 * size. To prevent vdev_trim_ranges() from
9095 * adding it again, we subtract it from
9098 spa_config_exit(dev->l2ad_spa, SCL_L2ARC, dev);
9099 vdev_trim_simple(vd,
9100 dev->l2ad_evict - VDEV_LABEL_START_SIZE,
9101 taddr - dev->l2ad_evict);
9102 spa_config_enter(dev->l2ad_spa, SCL_L2ARC, dev,
9107 * When rebuilding L2ARC we retrieve the evict hand
9108 * from the header of the device. Of note, l2arc_evict()
9109 * does not actually delete buffers from the cache
9110 * device, but trimming may do so depending on the
9111 * hardware implementation. Thus keeping track of the
9112 * evict hand is useful.
9114 dev->l2ad_evict = MAX(dev->l2ad_evict, taddr);
9119 mutex_enter(&dev->l2ad_mtx);
9121 * We have to account for evicted log blocks. Run vdev_space_update()
9122 * on log blocks whose offset (in bytes) is before the evicted offset
9123 * (in bytes) by searching in the list of pointers to log blocks
9124 * present in the L2ARC device.
9126 for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf;
9127 lb_ptr_buf = lb_ptr_buf_prev) {
9129 lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf);
9131 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
9132 uint64_t asize = L2BLK_GET_PSIZE(
9133 (lb_ptr_buf->lb_ptr)->lbp_prop);
9136 * We don't worry about log blocks left behind (ie
9137 * lbp_payload_start < l2ad_hand) because l2arc_write_buffers()
9138 * will never write more than l2arc_evict() evicts.
9140 if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) {
9143 vdev_space_update(vd, -asize, 0, 0);
9144 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
9145 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
9146 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
9148 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
9149 list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf);
9150 kmem_free(lb_ptr_buf->lb_ptr,
9151 sizeof (l2arc_log_blkptr_t));
9152 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
9156 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
9157 hdr_prev = list_prev(buflist, hdr);
9159 ASSERT(!HDR_EMPTY(hdr));
9160 hash_lock = HDR_LOCK(hdr);
9163 * We cannot use mutex_enter or else we can deadlock
9164 * with l2arc_write_buffers (due to swapping the order
9165 * the hash lock and l2ad_mtx are taken).
9167 if (!mutex_tryenter(hash_lock)) {
9169 * Missed the hash lock. Retry.
9171 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
9172 mutex_exit(&dev->l2ad_mtx);
9173 mutex_enter(hash_lock);
9174 mutex_exit(hash_lock);
9179 * A header can't be on this list if it doesn't have L2 header.
9181 ASSERT(HDR_HAS_L2HDR(hdr));
9183 /* Ensure this header has finished being written. */
9184 ASSERT(!HDR_L2_WRITING(hdr));
9185 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
9187 if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict ||
9188 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
9190 * We've evicted to the target address,
9191 * or the end of the device.
9193 mutex_exit(hash_lock);
9197 if (!HDR_HAS_L1HDR(hdr)) {
9198 ASSERT(!HDR_L2_READING(hdr));
9200 * This doesn't exist in the ARC. Destroy.
9201 * arc_hdr_destroy() will call list_remove()
9202 * and decrement arcstat_l2_lsize.
9204 arc_change_state(arc_anon, hdr, hash_lock);
9205 arc_hdr_destroy(hdr);
9207 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
9208 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
9210 * Invalidate issued or about to be issued
9211 * reads, since we may be about to write
9212 * over this location.
9214 if (HDR_L2_READING(hdr)) {
9215 ARCSTAT_BUMP(arcstat_l2_evict_reading);
9216 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
9219 arc_hdr_l2hdr_destroy(hdr);
9221 mutex_exit(hash_lock);
9223 mutex_exit(&dev->l2ad_mtx);
9227 * We need to check if we evict all buffers, otherwise we may iterate
9230 if (!all && rerun) {
9232 * Bump device hand to the device start if it is approaching the
9233 * end. l2arc_evict() has already evicted ahead for this case.
9235 dev->l2ad_hand = dev->l2ad_start;
9236 dev->l2ad_evict = dev->l2ad_start;
9237 dev->l2ad_first = B_FALSE;
9243 * In case of cache device removal (all) the following
9244 * assertions may be violated without functional consequences
9245 * as the device is about to be removed.
9247 ASSERT3U(dev->l2ad_hand + distance, <, dev->l2ad_end);
9248 if (!dev->l2ad_first)
9249 ASSERT3U(dev->l2ad_hand, <, dev->l2ad_evict);
9254 * Handle any abd transforms that might be required for writing to the L2ARC.
9255 * If successful, this function will always return an abd with the data
9256 * transformed as it is on disk in a new abd of asize bytes.
9259 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
9264 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
9265 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
9266 uint64_t psize = HDR_GET_PSIZE(hdr);
9267 uint64_t size = arc_hdr_size(hdr);
9268 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
9269 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
9270 dsl_crypto_key_t *dck = NULL;
9271 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
9272 boolean_t no_crypt = B_FALSE;
9274 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
9275 !HDR_COMPRESSION_ENABLED(hdr)) ||
9276 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
9277 ASSERT3U(psize, <=, asize);
9280 * If this data simply needs its own buffer, we simply allocate it
9281 * and copy the data. This may be done to eliminate a dependency on a
9282 * shared buffer or to reallocate the buffer to match asize.
9284 if (HDR_HAS_RABD(hdr) && asize != psize) {
9285 ASSERT3U(asize, >=, psize);
9286 to_write = abd_alloc_for_io(asize, ismd);
9287 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
9289 abd_zero_off(to_write, psize, asize - psize);
9293 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
9294 !HDR_ENCRYPTED(hdr)) {
9295 ASSERT3U(size, ==, psize);
9296 to_write = abd_alloc_for_io(asize, ismd);
9297 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
9299 abd_zero_off(to_write, size, asize - size);
9303 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
9304 cabd = abd_alloc_for_io(asize, ismd);
9305 tmp = abd_borrow_buf(cabd, asize);
9307 psize = zio_compress_data(compress, to_write, tmp, size,
9310 if (psize >= size) {
9311 abd_return_buf(cabd, tmp, asize);
9312 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
9314 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
9316 abd_zero_off(to_write, size, asize - size);
9319 ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
9321 bzero((char *)tmp + psize, asize - psize);
9322 psize = HDR_GET_PSIZE(hdr);
9323 abd_return_buf_copy(cabd, tmp, asize);
9328 if (HDR_ENCRYPTED(hdr)) {
9329 eabd = abd_alloc_for_io(asize, ismd);
9332 * If the dataset was disowned before the buffer
9333 * made it to this point, the key to re-encrypt
9334 * it won't be available. In this case we simply
9335 * won't write the buffer to the L2ARC.
9337 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
9342 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
9343 hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
9344 hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
9350 abd_copy(eabd, to_write, psize);
9353 abd_zero_off(eabd, psize, asize - psize);
9355 /* assert that the MAC we got here matches the one we saved */
9356 ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
9357 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9359 if (to_write == cabd)
9366 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
9367 *abd_out = to_write;
9372 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9383 l2arc_blk_fetch_done(zio_t *zio)
9385 l2arc_read_callback_t *cb;
9387 cb = zio->io_private;
9388 if (cb->l2rcb_abd != NULL)
9389 abd_free(cb->l2rcb_abd);
9390 kmem_free(cb, sizeof (l2arc_read_callback_t));
9394 * Find and write ARC buffers to the L2ARC device.
9396 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
9397 * for reading until they have completed writing.
9398 * The headroom_boost is an in-out parameter used to maintain headroom boost
9399 * state between calls to this function.
9401 * Returns the number of bytes actually written (which may be smaller than
9402 * the delta by which the device hand has changed due to alignment and the
9403 * writing of log blocks).
9406 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
9408 arc_buf_hdr_t *hdr, *hdr_prev, *head;
9409 uint64_t write_asize, write_psize, write_lsize, headroom;
9411 l2arc_write_callback_t *cb = NULL;
9413 uint64_t guid = spa_load_guid(spa);
9414 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9416 ASSERT3P(dev->l2ad_vdev, !=, NULL);
9419 write_lsize = write_asize = write_psize = 0;
9421 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
9422 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
9425 * Copy buffers for L2ARC writing.
9427 for (int pass = 0; pass < L2ARC_FEED_TYPES; pass++) {
9429 * If pass == 1 or 3, we cache MRU metadata and data
9432 if (l2arc_mfuonly) {
9433 if (pass == 1 || pass == 3)
9437 multilist_sublist_t *mls = l2arc_sublist_lock(pass);
9438 uint64_t passed_sz = 0;
9440 VERIFY3P(mls, !=, NULL);
9443 * L2ARC fast warmup.
9445 * Until the ARC is warm and starts to evict, read from the
9446 * head of the ARC lists rather than the tail.
9448 if (arc_warm == B_FALSE)
9449 hdr = multilist_sublist_head(mls);
9451 hdr = multilist_sublist_tail(mls);
9453 headroom = target_sz * l2arc_headroom;
9454 if (zfs_compressed_arc_enabled)
9455 headroom = (headroom * l2arc_headroom_boost) / 100;
9457 for (; hdr; hdr = hdr_prev) {
9458 kmutex_t *hash_lock;
9459 abd_t *to_write = NULL;
9461 if (arc_warm == B_FALSE)
9462 hdr_prev = multilist_sublist_next(mls, hdr);
9464 hdr_prev = multilist_sublist_prev(mls, hdr);
9466 hash_lock = HDR_LOCK(hdr);
9467 if (!mutex_tryenter(hash_lock)) {
9469 * Skip this buffer rather than waiting.
9474 passed_sz += HDR_GET_LSIZE(hdr);
9475 if (l2arc_headroom != 0 && passed_sz > headroom) {
9479 mutex_exit(hash_lock);
9483 if (!l2arc_write_eligible(guid, hdr)) {
9484 mutex_exit(hash_lock);
9489 * We rely on the L1 portion of the header below, so
9490 * it's invalid for this header to have been evicted out
9491 * of the ghost cache, prior to being written out. The
9492 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
9494 ASSERT(HDR_HAS_L1HDR(hdr));
9496 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
9497 ASSERT3U(arc_hdr_size(hdr), >, 0);
9498 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
9500 uint64_t psize = HDR_GET_PSIZE(hdr);
9501 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
9504 if ((write_asize + asize) > target_sz) {
9506 mutex_exit(hash_lock);
9511 * We rely on the L1 portion of the header below, so
9512 * it's invalid for this header to have been evicted out
9513 * of the ghost cache, prior to being written out. The
9514 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
9516 arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
9517 ASSERT(HDR_HAS_L1HDR(hdr));
9519 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
9520 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
9522 ASSERT3U(arc_hdr_size(hdr), >, 0);
9525 * If this header has b_rabd, we can use this since it
9526 * must always match the data exactly as it exists on
9527 * disk. Otherwise, the L2ARC can normally use the
9528 * hdr's data, but if we're sharing data between the
9529 * hdr and one of its bufs, L2ARC needs its own copy of
9530 * the data so that the ZIO below can't race with the
9531 * buf consumer. To ensure that this copy will be
9532 * available for the lifetime of the ZIO and be cleaned
9533 * up afterwards, we add it to the l2arc_free_on_write
9534 * queue. If we need to apply any transforms to the
9535 * data (compression, encryption) we will also need the
9538 if (HDR_HAS_RABD(hdr) && psize == asize) {
9539 to_write = hdr->b_crypt_hdr.b_rabd;
9540 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
9541 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
9542 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
9544 to_write = hdr->b_l1hdr.b_pabd;
9547 arc_buf_contents_t type = arc_buf_type(hdr);
9549 ret = l2arc_apply_transforms(spa, hdr, asize,
9552 arc_hdr_clear_flags(hdr,
9553 ARC_FLAG_L2_WRITING);
9554 mutex_exit(hash_lock);
9558 l2arc_free_abd_on_write(to_write, asize, type);
9563 * Insert a dummy header on the buflist so
9564 * l2arc_write_done() can find where the
9565 * write buffers begin without searching.
9567 mutex_enter(&dev->l2ad_mtx);
9568 list_insert_head(&dev->l2ad_buflist, head);
9569 mutex_exit(&dev->l2ad_mtx);
9572 sizeof (l2arc_write_callback_t), KM_SLEEP);
9573 cb->l2wcb_dev = dev;
9574 cb->l2wcb_head = head;
9576 * Create a list to save allocated abd buffers
9577 * for l2arc_log_blk_commit().
9579 list_create(&cb->l2wcb_abd_list,
9580 sizeof (l2arc_lb_abd_buf_t),
9581 offsetof(l2arc_lb_abd_buf_t, node));
9582 pio = zio_root(spa, l2arc_write_done, cb,
9586 hdr->b_l2hdr.b_dev = dev;
9587 hdr->b_l2hdr.b_hits = 0;
9589 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
9590 hdr->b_l2hdr.b_arcs_state =
9591 hdr->b_l1hdr.b_state->arcs_state;
9592 arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR);
9594 mutex_enter(&dev->l2ad_mtx);
9595 list_insert_head(&dev->l2ad_buflist, hdr);
9596 mutex_exit(&dev->l2ad_mtx);
9598 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
9599 arc_hdr_size(hdr), hdr);
9601 wzio = zio_write_phys(pio, dev->l2ad_vdev,
9602 hdr->b_l2hdr.b_daddr, asize, to_write,
9603 ZIO_CHECKSUM_OFF, NULL, hdr,
9604 ZIO_PRIORITY_ASYNC_WRITE,
9605 ZIO_FLAG_CANFAIL, B_FALSE);
9607 write_lsize += HDR_GET_LSIZE(hdr);
9608 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
9611 write_psize += psize;
9612 write_asize += asize;
9613 dev->l2ad_hand += asize;
9614 l2arc_hdr_arcstats_increment(hdr);
9615 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9617 mutex_exit(hash_lock);
9620 * Append buf info to current log and commit if full.
9621 * arcstat_l2_{size,asize} kstats are updated
9624 if (l2arc_log_blk_insert(dev, hdr))
9625 l2arc_log_blk_commit(dev, pio, cb);
9630 multilist_sublist_unlock(mls);
9636 /* No buffers selected for writing? */
9638 ASSERT0(write_lsize);
9639 ASSERT(!HDR_HAS_L1HDR(head));
9640 kmem_cache_free(hdr_l2only_cache, head);
9643 * Although we did not write any buffers l2ad_evict may
9646 if (dev->l2ad_evict != l2dhdr->dh_evict)
9647 l2arc_dev_hdr_update(dev);
9652 if (!dev->l2ad_first)
9653 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
9655 ASSERT3U(write_asize, <=, target_sz);
9656 ARCSTAT_BUMP(arcstat_l2_writes_sent);
9657 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
9659 dev->l2ad_writing = B_TRUE;
9660 (void) zio_wait(pio);
9661 dev->l2ad_writing = B_FALSE;
9664 * Update the device header after the zio completes as
9665 * l2arc_write_done() may have updated the memory holding the log block
9666 * pointers in the device header.
9668 l2arc_dev_hdr_update(dev);
9670 return (write_asize);
9674 l2arc_hdr_limit_reached(void)
9676 int64_t s = aggsum_upper_bound(&arc_sums.arcstat_l2_hdr_size);
9678 return (arc_reclaim_needed() || (s > arc_meta_limit * 3 / 4) ||
9679 (s > (arc_warm ? arc_c : arc_c_max) * l2arc_meta_percent / 100));
9683 * This thread feeds the L2ARC at regular intervals. This is the beating
9684 * heart of the L2ARC.
9688 l2arc_feed_thread(void *unused)
9693 uint64_t size, wrote;
9694 clock_t begin, next = ddi_get_lbolt();
9695 fstrans_cookie_t cookie;
9697 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
9699 mutex_enter(&l2arc_feed_thr_lock);
9701 cookie = spl_fstrans_mark();
9702 while (l2arc_thread_exit == 0) {
9703 CALLB_CPR_SAFE_BEGIN(&cpr);
9704 (void) cv_timedwait_idle(&l2arc_feed_thr_cv,
9705 &l2arc_feed_thr_lock, next);
9706 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
9707 next = ddi_get_lbolt() + hz;
9710 * Quick check for L2ARC devices.
9712 mutex_enter(&l2arc_dev_mtx);
9713 if (l2arc_ndev == 0) {
9714 mutex_exit(&l2arc_dev_mtx);
9717 mutex_exit(&l2arc_dev_mtx);
9718 begin = ddi_get_lbolt();
9721 * This selects the next l2arc device to write to, and in
9722 * doing so the next spa to feed from: dev->l2ad_spa. This
9723 * will return NULL if there are now no l2arc devices or if
9724 * they are all faulted.
9726 * If a device is returned, its spa's config lock is also
9727 * held to prevent device removal. l2arc_dev_get_next()
9728 * will grab and release l2arc_dev_mtx.
9730 if ((dev = l2arc_dev_get_next()) == NULL)
9733 spa = dev->l2ad_spa;
9734 ASSERT3P(spa, !=, NULL);
9737 * If the pool is read-only then force the feed thread to
9738 * sleep a little longer.
9740 if (!spa_writeable(spa)) {
9741 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
9742 spa_config_exit(spa, SCL_L2ARC, dev);
9747 * Avoid contributing to memory pressure.
9749 if (l2arc_hdr_limit_reached()) {
9750 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
9751 spa_config_exit(spa, SCL_L2ARC, dev);
9755 ARCSTAT_BUMP(arcstat_l2_feeds);
9757 size = l2arc_write_size(dev);
9760 * Evict L2ARC buffers that will be overwritten.
9762 l2arc_evict(dev, size, B_FALSE);
9765 * Write ARC buffers.
9767 wrote = l2arc_write_buffers(spa, dev, size);
9770 * Calculate interval between writes.
9772 next = l2arc_write_interval(begin, size, wrote);
9773 spa_config_exit(spa, SCL_L2ARC, dev);
9775 spl_fstrans_unmark(cookie);
9777 l2arc_thread_exit = 0;
9778 cv_broadcast(&l2arc_feed_thr_cv);
9779 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
9784 l2arc_vdev_present(vdev_t *vd)
9786 return (l2arc_vdev_get(vd) != NULL);
9790 * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if
9791 * the vdev_t isn't an L2ARC device.
9794 l2arc_vdev_get(vdev_t *vd)
9798 mutex_enter(&l2arc_dev_mtx);
9799 for (dev = list_head(l2arc_dev_list); dev != NULL;
9800 dev = list_next(l2arc_dev_list, dev)) {
9801 if (dev->l2ad_vdev == vd)
9804 mutex_exit(&l2arc_dev_mtx);
9810 l2arc_rebuild_dev(l2arc_dev_t *dev, boolean_t reopen)
9812 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9813 uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9814 spa_t *spa = dev->l2ad_spa;
9817 * The L2ARC has to hold at least the payload of one log block for
9818 * them to be restored (persistent L2ARC). The payload of a log block
9819 * depends on the amount of its log entries. We always write log blocks
9820 * with 1022 entries. How many of them are committed or restored depends
9821 * on the size of the L2ARC device. Thus the maximum payload of
9822 * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device
9823 * is less than that, we reduce the amount of committed and restored
9824 * log entries per block so as to enable persistence.
9826 if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) {
9827 dev->l2ad_log_entries = 0;
9829 dev->l2ad_log_entries = MIN((dev->l2ad_end -
9830 dev->l2ad_start) >> SPA_MAXBLOCKSHIFT,
9831 L2ARC_LOG_BLK_MAX_ENTRIES);
9835 * Read the device header, if an error is returned do not rebuild L2ARC.
9837 if (l2arc_dev_hdr_read(dev) == 0 && dev->l2ad_log_entries > 0) {
9839 * If we are onlining a cache device (vdev_reopen) that was
9840 * still present (l2arc_vdev_present()) and rebuild is enabled,
9841 * we should evict all ARC buffers and pointers to log blocks
9842 * and reclaim their space before restoring its contents to
9846 if (!l2arc_rebuild_enabled) {
9849 l2arc_evict(dev, 0, B_TRUE);
9850 /* start a new log block */
9851 dev->l2ad_log_ent_idx = 0;
9852 dev->l2ad_log_blk_payload_asize = 0;
9853 dev->l2ad_log_blk_payload_start = 0;
9857 * Just mark the device as pending for a rebuild. We won't
9858 * be starting a rebuild in line here as it would block pool
9859 * import. Instead spa_load_impl will hand that off to an
9860 * async task which will call l2arc_spa_rebuild_start.
9862 dev->l2ad_rebuild = B_TRUE;
9863 } else if (spa_writeable(spa)) {
9865 * In this case TRIM the whole device if l2arc_trim_ahead > 0,
9866 * otherwise create a new header. We zero out the memory holding
9867 * the header to reset dh_start_lbps. If we TRIM the whole
9868 * device the new header will be written by
9869 * vdev_trim_l2arc_thread() at the end of the TRIM to update the
9870 * trim_state in the header too. When reading the header, if
9871 * trim_state is not VDEV_TRIM_COMPLETE and l2arc_trim_ahead > 0
9872 * we opt to TRIM the whole device again.
9874 if (l2arc_trim_ahead > 0) {
9875 dev->l2ad_trim_all = B_TRUE;
9877 bzero(l2dhdr, l2dhdr_asize);
9878 l2arc_dev_hdr_update(dev);
9884 * Add a vdev for use by the L2ARC. By this point the spa has already
9885 * validated the vdev and opened it.
9888 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
9890 l2arc_dev_t *adddev;
9891 uint64_t l2dhdr_asize;
9893 ASSERT(!l2arc_vdev_present(vd));
9896 * Create a new l2arc device entry.
9898 adddev = vmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
9899 adddev->l2ad_spa = spa;
9900 adddev->l2ad_vdev = vd;
9901 /* leave extra size for an l2arc device header */
9902 l2dhdr_asize = adddev->l2ad_dev_hdr_asize =
9903 MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift);
9904 adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize;
9905 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
9906 ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end);
9907 adddev->l2ad_hand = adddev->l2ad_start;
9908 adddev->l2ad_evict = adddev->l2ad_start;
9909 adddev->l2ad_first = B_TRUE;
9910 adddev->l2ad_writing = B_FALSE;
9911 adddev->l2ad_trim_all = B_FALSE;
9912 list_link_init(&adddev->l2ad_node);
9913 adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP);
9915 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
9917 * This is a list of all ARC buffers that are still valid on the
9920 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
9921 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
9924 * This is a list of pointers to log blocks that are still present
9927 list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t),
9928 offsetof(l2arc_lb_ptr_buf_t, node));
9930 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
9931 zfs_refcount_create(&adddev->l2ad_alloc);
9932 zfs_refcount_create(&adddev->l2ad_lb_asize);
9933 zfs_refcount_create(&adddev->l2ad_lb_count);
9936 * Decide if dev is eligible for L2ARC rebuild or whole device
9937 * trimming. This has to happen before the device is added in the
9938 * cache device list and l2arc_dev_mtx is released. Otherwise
9939 * l2arc_feed_thread() might already start writing on the
9942 l2arc_rebuild_dev(adddev, B_FALSE);
9945 * Add device to global list
9947 mutex_enter(&l2arc_dev_mtx);
9948 list_insert_head(l2arc_dev_list, adddev);
9949 atomic_inc_64(&l2arc_ndev);
9950 mutex_exit(&l2arc_dev_mtx);
9954 * Decide if a vdev is eligible for L2ARC rebuild, called from vdev_reopen()
9955 * in case of onlining a cache device.
9958 l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen)
9960 l2arc_dev_t *dev = NULL;
9962 dev = l2arc_vdev_get(vd);
9963 ASSERT3P(dev, !=, NULL);
9966 * In contrast to l2arc_add_vdev() we do not have to worry about
9967 * l2arc_feed_thread() invalidating previous content when onlining a
9968 * cache device. The device parameters (l2ad*) are not cleared when
9969 * offlining the device and writing new buffers will not invalidate
9970 * all previous content. In worst case only buffers that have not had
9971 * their log block written to the device will be lost.
9972 * When onlining the cache device (ie offline->online without exporting
9973 * the pool in between) this happens:
9974 * vdev_reopen() -> vdev_open() -> l2arc_rebuild_vdev()
9976 * vdev_is_dead() = B_FALSE l2ad_rebuild = B_TRUE
9977 * During the time where vdev_is_dead = B_FALSE and until l2ad_rebuild
9978 * is set to B_TRUE we might write additional buffers to the device.
9980 l2arc_rebuild_dev(dev, reopen);
9984 * Remove a vdev from the L2ARC.
9987 l2arc_remove_vdev(vdev_t *vd)
9989 l2arc_dev_t *remdev = NULL;
9992 * Find the device by vdev
9994 remdev = l2arc_vdev_get(vd);
9995 ASSERT3P(remdev, !=, NULL);
9998 * Cancel any ongoing or scheduled rebuild.
10000 mutex_enter(&l2arc_rebuild_thr_lock);
10001 if (remdev->l2ad_rebuild_began == B_TRUE) {
10002 remdev->l2ad_rebuild_cancel = B_TRUE;
10003 while (remdev->l2ad_rebuild == B_TRUE)
10004 cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock);
10006 mutex_exit(&l2arc_rebuild_thr_lock);
10009 * Remove device from global list
10011 mutex_enter(&l2arc_dev_mtx);
10012 list_remove(l2arc_dev_list, remdev);
10013 l2arc_dev_last = NULL; /* may have been invalidated */
10014 atomic_dec_64(&l2arc_ndev);
10015 mutex_exit(&l2arc_dev_mtx);
10018 * Clear all buflists and ARC references. L2ARC device flush.
10020 l2arc_evict(remdev, 0, B_TRUE);
10021 list_destroy(&remdev->l2ad_buflist);
10022 ASSERT(list_is_empty(&remdev->l2ad_lbptr_list));
10023 list_destroy(&remdev->l2ad_lbptr_list);
10024 mutex_destroy(&remdev->l2ad_mtx);
10025 zfs_refcount_destroy(&remdev->l2ad_alloc);
10026 zfs_refcount_destroy(&remdev->l2ad_lb_asize);
10027 zfs_refcount_destroy(&remdev->l2ad_lb_count);
10028 kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize);
10029 vmem_free(remdev, sizeof (l2arc_dev_t));
10035 l2arc_thread_exit = 0;
10038 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
10039 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
10040 mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL);
10041 cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL);
10042 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
10043 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
10045 l2arc_dev_list = &L2ARC_dev_list;
10046 l2arc_free_on_write = &L2ARC_free_on_write;
10047 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
10048 offsetof(l2arc_dev_t, l2ad_node));
10049 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
10050 offsetof(l2arc_data_free_t, l2df_list_node));
10056 mutex_destroy(&l2arc_feed_thr_lock);
10057 cv_destroy(&l2arc_feed_thr_cv);
10058 mutex_destroy(&l2arc_rebuild_thr_lock);
10059 cv_destroy(&l2arc_rebuild_thr_cv);
10060 mutex_destroy(&l2arc_dev_mtx);
10061 mutex_destroy(&l2arc_free_on_write_mtx);
10063 list_destroy(l2arc_dev_list);
10064 list_destroy(l2arc_free_on_write);
10070 if (!(spa_mode_global & SPA_MODE_WRITE))
10073 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
10074 TS_RUN, defclsyspri);
10080 if (!(spa_mode_global & SPA_MODE_WRITE))
10083 mutex_enter(&l2arc_feed_thr_lock);
10084 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
10085 l2arc_thread_exit = 1;
10086 while (l2arc_thread_exit != 0)
10087 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
10088 mutex_exit(&l2arc_feed_thr_lock);
10092 * Punches out rebuild threads for the L2ARC devices in a spa. This should
10093 * be called after pool import from the spa async thread, since starting
10094 * these threads directly from spa_import() will make them part of the
10095 * "zpool import" context and delay process exit (and thus pool import).
10098 l2arc_spa_rebuild_start(spa_t *spa)
10100 ASSERT(MUTEX_HELD(&spa_namespace_lock));
10103 * Locate the spa's l2arc devices and kick off rebuild threads.
10105 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
10107 l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]);
10109 /* Don't attempt a rebuild if the vdev is UNAVAIL */
10112 mutex_enter(&l2arc_rebuild_thr_lock);
10113 if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) {
10114 dev->l2ad_rebuild_began = B_TRUE;
10115 (void) thread_create(NULL, 0, l2arc_dev_rebuild_thread,
10116 dev, 0, &p0, TS_RUN, minclsyspri);
10118 mutex_exit(&l2arc_rebuild_thr_lock);
10123 * Main entry point for L2ARC rebuilding.
10126 l2arc_dev_rebuild_thread(void *arg)
10128 l2arc_dev_t *dev = arg;
10130 VERIFY(!dev->l2ad_rebuild_cancel);
10131 VERIFY(dev->l2ad_rebuild);
10132 (void) l2arc_rebuild(dev);
10133 mutex_enter(&l2arc_rebuild_thr_lock);
10134 dev->l2ad_rebuild_began = B_FALSE;
10135 dev->l2ad_rebuild = B_FALSE;
10136 mutex_exit(&l2arc_rebuild_thr_lock);
10142 * This function implements the actual L2ARC metadata rebuild. It:
10143 * starts reading the log block chain and restores each block's contents
10144 * to memory (reconstructing arc_buf_hdr_t's).
10146 * Operation stops under any of the following conditions:
10148 * 1) We reach the end of the log block chain.
10149 * 2) We encounter *any* error condition (cksum errors, io errors)
10152 l2arc_rebuild(l2arc_dev_t *dev)
10154 vdev_t *vd = dev->l2ad_vdev;
10155 spa_t *spa = vd->vdev_spa;
10157 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10158 l2arc_log_blk_phys_t *this_lb, *next_lb;
10159 zio_t *this_io = NULL, *next_io = NULL;
10160 l2arc_log_blkptr_t lbps[2];
10161 l2arc_lb_ptr_buf_t *lb_ptr_buf;
10162 boolean_t lock_held;
10164 this_lb = vmem_zalloc(sizeof (*this_lb), KM_SLEEP);
10165 next_lb = vmem_zalloc(sizeof (*next_lb), KM_SLEEP);
10168 * We prevent device removal while issuing reads to the device,
10169 * then during the rebuilding phases we drop this lock again so
10170 * that a spa_unload or device remove can be initiated - this is
10171 * safe, because the spa will signal us to stop before removing
10172 * our device and wait for us to stop.
10174 spa_config_enter(spa, SCL_L2ARC, vd, RW_READER);
10175 lock_held = B_TRUE;
10178 * Retrieve the persistent L2ARC device state.
10179 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10181 dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start);
10182 dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr +
10183 L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop),
10185 dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST);
10187 vd->vdev_trim_action_time = l2dhdr->dh_trim_action_time;
10188 vd->vdev_trim_state = l2dhdr->dh_trim_state;
10191 * In case the zfs module parameter l2arc_rebuild_enabled is false
10192 * we do not start the rebuild process.
10194 if (!l2arc_rebuild_enabled)
10197 /* Prepare the rebuild process */
10198 bcopy(l2dhdr->dh_start_lbps, lbps, sizeof (lbps));
10200 /* Start the rebuild process */
10202 if (!l2arc_log_blkptr_valid(dev, &lbps[0]))
10205 if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1],
10206 this_lb, next_lb, this_io, &next_io)) != 0)
10210 * Our memory pressure valve. If the system is running low
10211 * on memory, rather than swamping memory with new ARC buf
10212 * hdrs, we opt not to rebuild the L2ARC. At this point,
10213 * however, we have already set up our L2ARC dev to chain in
10214 * new metadata log blocks, so the user may choose to offline/
10215 * online the L2ARC dev at a later time (or re-import the pool)
10216 * to reconstruct it (when there's less memory pressure).
10218 if (l2arc_hdr_limit_reached()) {
10219 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem);
10220 cmn_err(CE_NOTE, "System running low on memory, "
10221 "aborting L2ARC rebuild.");
10222 err = SET_ERROR(ENOMEM);
10226 spa_config_exit(spa, SCL_L2ARC, vd);
10227 lock_held = B_FALSE;
10230 * Now that we know that the next_lb checks out alright, we
10231 * can start reconstruction from this log block.
10232 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10234 uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop);
10235 l2arc_log_blk_restore(dev, this_lb, asize);
10238 * log block restored, include its pointer in the list of
10239 * pointers to log blocks present in the L2ARC device.
10241 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
10242 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t),
10244 bcopy(&lbps[0], lb_ptr_buf->lb_ptr,
10245 sizeof (l2arc_log_blkptr_t));
10246 mutex_enter(&dev->l2ad_mtx);
10247 list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf);
10248 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10249 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10250 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10251 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10252 mutex_exit(&dev->l2ad_mtx);
10253 vdev_space_update(vd, asize, 0, 0);
10256 * Protection against loops of log blocks:
10258 * l2ad_hand l2ad_evict
10260 * l2ad_start |=======================================| l2ad_end
10261 * -----|||----|||---|||----|||
10263 * ---|||---|||----|||---|||
10266 * In this situation the pointer of log block (4) passes
10267 * l2arc_log_blkptr_valid() but the log block should not be
10268 * restored as it is overwritten by the payload of log block
10269 * (0). Only log blocks (0)-(3) should be restored. We check
10270 * whether l2ad_evict lies in between the payload starting
10271 * offset of the next log block (lbps[1].lbp_payload_start)
10272 * and the payload starting offset of the present log block
10273 * (lbps[0].lbp_payload_start). If true and this isn't the
10274 * first pass, we are looping from the beginning and we should
10277 if (l2arc_range_check_overlap(lbps[1].lbp_payload_start,
10278 lbps[0].lbp_payload_start, dev->l2ad_evict) &&
10284 mutex_enter(&l2arc_rebuild_thr_lock);
10285 if (dev->l2ad_rebuild_cancel) {
10286 dev->l2ad_rebuild = B_FALSE;
10287 cv_signal(&l2arc_rebuild_thr_cv);
10288 mutex_exit(&l2arc_rebuild_thr_lock);
10289 err = SET_ERROR(ECANCELED);
10292 mutex_exit(&l2arc_rebuild_thr_lock);
10293 if (spa_config_tryenter(spa, SCL_L2ARC, vd,
10295 lock_held = B_TRUE;
10299 * L2ARC config lock held by somebody in writer,
10300 * possibly due to them trying to remove us. They'll
10301 * likely to want us to shut down, so after a little
10302 * delay, we check l2ad_rebuild_cancel and retry
10309 * Continue with the next log block.
10312 lbps[1] = this_lb->lb_prev_lbp;
10313 PTR_SWAP(this_lb, next_lb);
10318 if (this_io != NULL)
10319 l2arc_log_blk_fetch_abort(this_io);
10321 if (next_io != NULL)
10322 l2arc_log_blk_fetch_abort(next_io);
10323 vmem_free(this_lb, sizeof (*this_lb));
10324 vmem_free(next_lb, sizeof (*next_lb));
10326 if (!l2arc_rebuild_enabled) {
10327 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10329 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) {
10330 ARCSTAT_BUMP(arcstat_l2_rebuild_success);
10331 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10332 "successful, restored %llu blocks",
10333 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10334 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) {
10336 * No error but also nothing restored, meaning the lbps array
10337 * in the device header points to invalid/non-present log
10338 * blocks. Reset the header.
10340 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10341 "no valid log blocks");
10342 bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
10343 l2arc_dev_hdr_update(dev);
10344 } else if (err == ECANCELED) {
10346 * In case the rebuild was canceled do not log to spa history
10347 * log as the pool may be in the process of being removed.
10349 zfs_dbgmsg("L2ARC rebuild aborted, restored %llu blocks",
10350 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10351 } else if (err != 0) {
10352 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10353 "aborted, restored %llu blocks",
10354 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10358 spa_config_exit(spa, SCL_L2ARC, vd);
10364 * Attempts to read the device header on the provided L2ARC device and writes
10365 * it to `hdr'. On success, this function returns 0, otherwise the appropriate
10366 * error code is returned.
10369 l2arc_dev_hdr_read(l2arc_dev_t *dev)
10373 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10374 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10377 guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10379 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10381 err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev,
10382 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd,
10383 ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
10384 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
10385 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY |
10386 ZIO_FLAG_SPECULATIVE, B_FALSE));
10391 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors);
10392 zfs_dbgmsg("L2ARC IO error (%d) while reading device header, "
10393 "vdev guid: %llu", err,
10394 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10398 if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC))
10399 byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr));
10401 if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC ||
10402 l2dhdr->dh_spa_guid != guid ||
10403 l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid ||
10404 l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION ||
10405 l2dhdr->dh_log_entries != dev->l2ad_log_entries ||
10406 l2dhdr->dh_end != dev->l2ad_end ||
10407 !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end,
10408 l2dhdr->dh_evict) ||
10409 (l2dhdr->dh_trim_state != VDEV_TRIM_COMPLETE &&
10410 l2arc_trim_ahead > 0)) {
10412 * Attempt to rebuild a device containing no actual dev hdr
10413 * or containing a header from some other pool or from another
10414 * version of persistent L2ARC.
10416 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported);
10417 return (SET_ERROR(ENOTSUP));
10424 * Reads L2ARC log blocks from storage and validates their contents.
10426 * This function implements a simple fetcher to make sure that while
10427 * we're processing one buffer the L2ARC is already fetching the next
10428 * one in the chain.
10430 * The arguments this_lp and next_lp point to the current and next log block
10431 * address in the block chain. Similarly, this_lb and next_lb hold the
10432 * l2arc_log_blk_phys_t's of the current and next L2ARC blk.
10434 * The `this_io' and `next_io' arguments are used for block fetching.
10435 * When issuing the first blk IO during rebuild, you should pass NULL for
10436 * `this_io'. This function will then issue a sync IO to read the block and
10437 * also issue an async IO to fetch the next block in the block chain. The
10438 * fetched IO is returned in `next_io'. On subsequent calls to this
10439 * function, pass the value returned in `next_io' from the previous call
10440 * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO.
10441 * Prior to the call, you should initialize your `next_io' pointer to be
10442 * NULL. If no fetch IO was issued, the pointer is left set at NULL.
10444 * On success, this function returns 0, otherwise it returns an appropriate
10445 * error code. On error the fetching IO is aborted and cleared before
10446 * returning from this function. Therefore, if we return `success', the
10447 * caller can assume that we have taken care of cleanup of fetch IOs.
10450 l2arc_log_blk_read(l2arc_dev_t *dev,
10451 const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp,
10452 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
10453 zio_t *this_io, zio_t **next_io)
10460 ASSERT(this_lbp != NULL && next_lbp != NULL);
10461 ASSERT(this_lb != NULL && next_lb != NULL);
10462 ASSERT(next_io != NULL && *next_io == NULL);
10463 ASSERT(l2arc_log_blkptr_valid(dev, this_lbp));
10466 * Check to see if we have issued the IO for this log block in a
10467 * previous run. If not, this is the first call, so issue it now.
10469 if (this_io == NULL) {
10470 this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp,
10475 * Peek to see if we can start issuing the next IO immediately.
10477 if (l2arc_log_blkptr_valid(dev, next_lbp)) {
10479 * Start issuing IO for the next log block early - this
10480 * should help keep the L2ARC device busy while we
10481 * decompress and restore this log block.
10483 *next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp,
10487 /* Wait for the IO to read this log block to complete */
10488 if ((err = zio_wait(this_io)) != 0) {
10489 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
10490 zfs_dbgmsg("L2ARC IO error (%d) while reading log block, "
10491 "offset: %llu, vdev guid: %llu", err,
10492 (u_longlong_t)this_lbp->lbp_daddr,
10493 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10498 * Make sure the buffer checks out.
10499 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10501 asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop);
10502 fletcher_4_native(this_lb, asize, NULL, &cksum);
10503 if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) {
10504 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors);
10505 zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, "
10506 "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu",
10507 (u_longlong_t)this_lbp->lbp_daddr,
10508 (u_longlong_t)dev->l2ad_vdev->vdev_guid,
10509 (u_longlong_t)dev->l2ad_hand,
10510 (u_longlong_t)dev->l2ad_evict);
10511 err = SET_ERROR(ECKSUM);
10515 /* Now we can take our time decoding this buffer */
10516 switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) {
10517 case ZIO_COMPRESS_OFF:
10519 case ZIO_COMPRESS_LZ4:
10520 abd = abd_alloc_for_io(asize, B_TRUE);
10521 abd_copy_from_buf_off(abd, this_lb, 0, asize);
10522 if ((err = zio_decompress_data(
10523 L2BLK_GET_COMPRESS((this_lbp)->lbp_prop),
10524 abd, this_lb, asize, sizeof (*this_lb), NULL)) != 0) {
10525 err = SET_ERROR(EINVAL);
10530 err = SET_ERROR(EINVAL);
10533 if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC))
10534 byteswap_uint64_array(this_lb, sizeof (*this_lb));
10535 if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) {
10536 err = SET_ERROR(EINVAL);
10540 /* Abort an in-flight fetch I/O in case of error */
10541 if (err != 0 && *next_io != NULL) {
10542 l2arc_log_blk_fetch_abort(*next_io);
10551 * Restores the payload of a log block to ARC. This creates empty ARC hdr
10552 * entries which only contain an l2arc hdr, essentially restoring the
10553 * buffers to their L2ARC evicted state. This function also updates space
10554 * usage on the L2ARC vdev to make sure it tracks restored buffers.
10557 l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb,
10560 uint64_t size = 0, asize = 0;
10561 uint64_t log_entries = dev->l2ad_log_entries;
10564 * Usually arc_adapt() is called only for data, not headers, but
10565 * since we may allocate significant amount of memory here, let ARC
10568 arc_adapt(log_entries * HDR_L2ONLY_SIZE, arc_l2c_only);
10570 for (int i = log_entries - 1; i >= 0; i--) {
10572 * Restore goes in the reverse temporal direction to preserve
10573 * correct temporal ordering of buffers in the l2ad_buflist.
10574 * l2arc_hdr_restore also does a list_insert_tail instead of
10575 * list_insert_head on the l2ad_buflist:
10577 * LIST l2ad_buflist LIST
10578 * HEAD <------ (time) ------ TAIL
10579 * direction +-----+-----+-----+-----+-----+ direction
10580 * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild
10581 * fill +-----+-----+-----+-----+-----+
10585 * l2arc_feed_thread l2arc_rebuild
10586 * will place new bufs here restores bufs here
10588 * During l2arc_rebuild() the device is not used by
10589 * l2arc_feed_thread() as dev->l2ad_rebuild is set to true.
10591 size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop);
10592 asize += vdev_psize_to_asize(dev->l2ad_vdev,
10593 L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop));
10594 l2arc_hdr_restore(&lb->lb_entries[i], dev);
10598 * Record rebuild stats:
10599 * size Logical size of restored buffers in the L2ARC
10600 * asize Aligned size of restored buffers in the L2ARC
10602 ARCSTAT_INCR(arcstat_l2_rebuild_size, size);
10603 ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize);
10604 ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries);
10605 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize);
10606 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize);
10607 ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks);
10611 * Restores a single ARC buf hdr from a log entry. The ARC buffer is put
10612 * into a state indicating that it has been evicted to L2ARC.
10615 l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
10617 arc_buf_hdr_t *hdr, *exists;
10618 kmutex_t *hash_lock;
10619 arc_buf_contents_t type = L2BLK_GET_TYPE((le)->le_prop);
10623 * Do all the allocation before grabbing any locks, this lets us
10624 * sleep if memory is full and we don't have to deal with failed
10627 hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type,
10628 dev, le->le_dva, le->le_daddr,
10629 L2BLK_GET_PSIZE((le)->le_prop), le->le_birth,
10630 L2BLK_GET_COMPRESS((le)->le_prop), le->le_complevel,
10631 L2BLK_GET_PROTECTED((le)->le_prop),
10632 L2BLK_GET_PREFETCH((le)->le_prop),
10633 L2BLK_GET_STATE((le)->le_prop));
10634 asize = vdev_psize_to_asize(dev->l2ad_vdev,
10635 L2BLK_GET_PSIZE((le)->le_prop));
10638 * vdev_space_update() has to be called before arc_hdr_destroy() to
10639 * avoid underflow since the latter also calls vdev_space_update().
10641 l2arc_hdr_arcstats_increment(hdr);
10642 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10644 mutex_enter(&dev->l2ad_mtx);
10645 list_insert_tail(&dev->l2ad_buflist, hdr);
10646 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
10647 mutex_exit(&dev->l2ad_mtx);
10649 exists = buf_hash_insert(hdr, &hash_lock);
10651 /* Buffer was already cached, no need to restore it. */
10652 arc_hdr_destroy(hdr);
10654 * If the buffer is already cached, check whether it has
10655 * L2ARC metadata. If not, enter them and update the flag.
10656 * This is important is case of onlining a cache device, since
10657 * we previously evicted all L2ARC metadata from ARC.
10659 if (!HDR_HAS_L2HDR(exists)) {
10660 arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR);
10661 exists->b_l2hdr.b_dev = dev;
10662 exists->b_l2hdr.b_daddr = le->le_daddr;
10663 exists->b_l2hdr.b_arcs_state =
10664 L2BLK_GET_STATE((le)->le_prop);
10665 mutex_enter(&dev->l2ad_mtx);
10666 list_insert_tail(&dev->l2ad_buflist, exists);
10667 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
10668 arc_hdr_size(exists), exists);
10669 mutex_exit(&dev->l2ad_mtx);
10670 l2arc_hdr_arcstats_increment(exists);
10671 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10673 ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
10676 mutex_exit(hash_lock);
10680 * Starts an asynchronous read IO to read a log block. This is used in log
10681 * block reconstruction to start reading the next block before we are done
10682 * decoding and reconstructing the current block, to keep the l2arc device
10683 * nice and hot with read IO to process.
10684 * The returned zio will contain a newly allocated memory buffers for the IO
10685 * data which should then be freed by the caller once the zio is no longer
10686 * needed (i.e. due to it having completed). If you wish to abort this
10687 * zio, you should do so using l2arc_log_blk_fetch_abort, which takes
10688 * care of disposing of the allocated buffers correctly.
10691 l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp,
10692 l2arc_log_blk_phys_t *lb)
10696 l2arc_read_callback_t *cb;
10698 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10699 asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10700 ASSERT(asize <= sizeof (l2arc_log_blk_phys_t));
10702 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP);
10703 cb->l2rcb_abd = abd_get_from_buf(lb, asize);
10704 pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb,
10705 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
10706 ZIO_FLAG_DONT_RETRY);
10707 (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize,
10708 cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10709 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
10710 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
10716 * Aborts a zio returned from l2arc_log_blk_fetch and frees the data
10717 * buffers allocated for it.
10720 l2arc_log_blk_fetch_abort(zio_t *zio)
10722 (void) zio_wait(zio);
10726 * Creates a zio to update the device header on an l2arc device.
10729 l2arc_dev_hdr_update(l2arc_dev_t *dev)
10731 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10732 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10736 VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER));
10738 l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC;
10739 l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION;
10740 l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10741 l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid;
10742 l2dhdr->dh_log_entries = dev->l2ad_log_entries;
10743 l2dhdr->dh_evict = dev->l2ad_evict;
10744 l2dhdr->dh_start = dev->l2ad_start;
10745 l2dhdr->dh_end = dev->l2ad_end;
10746 l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize);
10747 l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count);
10748 l2dhdr->dh_flags = 0;
10749 l2dhdr->dh_trim_action_time = dev->l2ad_vdev->vdev_trim_action_time;
10750 l2dhdr->dh_trim_state = dev->l2ad_vdev->vdev_trim_state;
10751 if (dev->l2ad_first)
10752 l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST;
10754 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10756 err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev,
10757 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL,
10758 NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE));
10763 zfs_dbgmsg("L2ARC IO error (%d) while writing device header, "
10764 "vdev guid: %llu", err,
10765 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10770 * Commits a log block to the L2ARC device. This routine is invoked from
10771 * l2arc_write_buffers when the log block fills up.
10772 * This function allocates some memory to temporarily hold the serialized
10773 * buffer to be written. This is then released in l2arc_write_done.
10776 l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb)
10778 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10779 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10780 uint64_t psize, asize;
10782 l2arc_lb_abd_buf_t *abd_buf;
10784 l2arc_lb_ptr_buf_t *lb_ptr_buf;
10786 VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries);
10788 tmpbuf = zio_buf_alloc(sizeof (*lb));
10789 abd_buf = zio_buf_alloc(sizeof (*abd_buf));
10790 abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb));
10791 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
10792 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP);
10794 /* link the buffer into the block chain */
10795 lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1];
10796 lb->lb_magic = L2ARC_LOG_BLK_MAGIC;
10799 * l2arc_log_blk_commit() may be called multiple times during a single
10800 * l2arc_write_buffers() call. Save the allocated abd buffers in a list
10801 * so we can free them in l2arc_write_done() later on.
10803 list_insert_tail(&cb->l2wcb_abd_list, abd_buf);
10805 /* try to compress the buffer */
10806 psize = zio_compress_data(ZIO_COMPRESS_LZ4,
10807 abd_buf->abd, tmpbuf, sizeof (*lb), 0);
10809 /* a log block is never entirely zero */
10810 ASSERT(psize != 0);
10811 asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
10812 ASSERT(asize <= sizeof (*lb));
10815 * Update the start log block pointer in the device header to point
10816 * to the log block we're about to write.
10818 l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0];
10819 l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand;
10820 l2dhdr->dh_start_lbps[0].lbp_payload_asize =
10821 dev->l2ad_log_blk_payload_asize;
10822 l2dhdr->dh_start_lbps[0].lbp_payload_start =
10823 dev->l2ad_log_blk_payload_start;
10826 (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb));
10828 (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize);
10829 L2BLK_SET_CHECKSUM(
10830 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10831 ZIO_CHECKSUM_FLETCHER_4);
10832 if (asize < sizeof (*lb)) {
10833 /* compression succeeded */
10834 bzero(tmpbuf + psize, asize - psize);
10835 L2BLK_SET_COMPRESS(
10836 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10839 /* compression failed */
10840 bcopy(lb, tmpbuf, sizeof (*lb));
10841 L2BLK_SET_COMPRESS(
10842 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10846 /* checksum what we're about to write */
10847 fletcher_4_native(tmpbuf, asize, NULL,
10848 &l2dhdr->dh_start_lbps[0].lbp_cksum);
10850 abd_free(abd_buf->abd);
10852 /* perform the write itself */
10853 abd_buf->abd = abd_get_from_buf(tmpbuf, sizeof (*lb));
10854 abd_take_ownership_of_buf(abd_buf->abd, B_TRUE);
10855 wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand,
10856 asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10857 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
10858 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
10859 (void) zio_nowait(wzio);
10861 dev->l2ad_hand += asize;
10863 * Include the committed log block's pointer in the list of pointers
10864 * to log blocks present in the L2ARC device.
10866 bcopy(&l2dhdr->dh_start_lbps[0], lb_ptr_buf->lb_ptr,
10867 sizeof (l2arc_log_blkptr_t));
10868 mutex_enter(&dev->l2ad_mtx);
10869 list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf);
10870 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10871 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10872 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10873 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10874 mutex_exit(&dev->l2ad_mtx);
10875 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10877 /* bump the kstats */
10878 ARCSTAT_INCR(arcstat_l2_write_bytes, asize);
10879 ARCSTAT_BUMP(arcstat_l2_log_blk_writes);
10880 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize);
10881 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio,
10882 dev->l2ad_log_blk_payload_asize / asize);
10884 /* start a new log block */
10885 dev->l2ad_log_ent_idx = 0;
10886 dev->l2ad_log_blk_payload_asize = 0;
10887 dev->l2ad_log_blk_payload_start = 0;
10891 * Validates an L2ARC log block address to make sure that it can be read
10892 * from the provided L2ARC device.
10895 l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp)
10897 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10898 uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10899 uint64_t end = lbp->lbp_daddr + asize - 1;
10900 uint64_t start = lbp->lbp_payload_start;
10901 boolean_t evicted = B_FALSE;
10904 * A log block is valid if all of the following conditions are true:
10905 * - it fits entirely (including its payload) between l2ad_start and
10907 * - it has a valid size
10908 * - neither the log block itself nor part of its payload was evicted
10909 * by l2arc_evict():
10911 * l2ad_hand l2ad_evict
10916 * l2ad_start ============================================ l2ad_end
10917 * --------------------------||||
10924 l2arc_range_check_overlap(start, end, dev->l2ad_hand) ||
10925 l2arc_range_check_overlap(start, end, dev->l2ad_evict) ||
10926 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) ||
10927 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end);
10929 return (start >= dev->l2ad_start && end <= dev->l2ad_end &&
10930 asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) &&
10931 (!evicted || dev->l2ad_first));
10935 * Inserts ARC buffer header `hdr' into the current L2ARC log block on
10936 * the device. The buffer being inserted must be present in L2ARC.
10937 * Returns B_TRUE if the L2ARC log block is full and needs to be committed
10938 * to L2ARC, or B_FALSE if it still has room for more ARC buffers.
10941 l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr)
10943 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10944 l2arc_log_ent_phys_t *le;
10946 if (dev->l2ad_log_entries == 0)
10949 int index = dev->l2ad_log_ent_idx++;
10951 ASSERT3S(index, <, dev->l2ad_log_entries);
10952 ASSERT(HDR_HAS_L2HDR(hdr));
10954 le = &lb->lb_entries[index];
10955 bzero(le, sizeof (*le));
10956 le->le_dva = hdr->b_dva;
10957 le->le_birth = hdr->b_birth;
10958 le->le_daddr = hdr->b_l2hdr.b_daddr;
10960 dev->l2ad_log_blk_payload_start = le->le_daddr;
10961 L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr));
10962 L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr));
10963 L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr));
10964 le->le_complevel = hdr->b_complevel;
10965 L2BLK_SET_TYPE((le)->le_prop, hdr->b_type);
10966 L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr)));
10967 L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr)));
10968 L2BLK_SET_STATE((le)->le_prop, hdr->b_l1hdr.b_state->arcs_state);
10970 dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev,
10971 HDR_GET_PSIZE(hdr));
10973 return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries);
10977 * Checks whether a given L2ARC device address sits in a time-sequential
10978 * range. The trick here is that the L2ARC is a rotary buffer, so we can't
10979 * just do a range comparison, we need to handle the situation in which the
10980 * range wraps around the end of the L2ARC device. Arguments:
10981 * bottom -- Lower end of the range to check (written to earlier).
10982 * top -- Upper end of the range to check (written to later).
10983 * check -- The address for which we want to determine if it sits in
10984 * between the top and bottom.
10986 * The 3-way conditional below represents the following cases:
10988 * bottom < top : Sequentially ordered case:
10989 * <check>--------+-------------------+
10990 * | (overlap here?) |
10992 * |---------------<bottom>============<top>--------------|
10994 * bottom > top: Looped-around case:
10995 * <check>--------+------------------+
10996 * | (overlap here?) |
10998 * |===============<top>---------------<bottom>===========|
11001 * +---------------+---------<check>
11003 * top == bottom : Just a single address comparison.
11006 l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check)
11009 return (bottom <= check && check <= top);
11010 else if (bottom > top)
11011 return (check <= top || bottom <= check);
11013 return (check == top);
11016 EXPORT_SYMBOL(arc_buf_size);
11017 EXPORT_SYMBOL(arc_write);
11018 EXPORT_SYMBOL(arc_read);
11019 EXPORT_SYMBOL(arc_buf_info);
11020 EXPORT_SYMBOL(arc_getbuf_func);
11021 EXPORT_SYMBOL(arc_add_prune_callback);
11022 EXPORT_SYMBOL(arc_remove_prune_callback);
11024 /* BEGIN CSTYLED */
11025 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min, param_set_arc_min,
11026 param_get_long, ZMOD_RW, "Min arc size");
11028 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, max, param_set_arc_max,
11029 param_get_long, ZMOD_RW, "Max arc size");
11031 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit, param_set_arc_long,
11032 param_get_long, ZMOD_RW, "Metadata limit for arc size");
11034 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit_percent,
11035 param_set_arc_long, param_get_long, ZMOD_RW,
11036 "Percent of arc size for arc meta limit");
11038 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_min, param_set_arc_long,
11039 param_get_long, ZMOD_RW, "Min arc metadata");
11041 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_prune, INT, ZMOD_RW,
11042 "Meta objects to scan for prune");
11044 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_adjust_restarts, INT, ZMOD_RW,
11045 "Limit number of restarts in arc_evict_meta");
11047 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_strategy, INT, ZMOD_RW,
11048 "Meta reclaim strategy");
11050 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, grow_retry, param_set_arc_int,
11051 param_get_int, ZMOD_RW, "Seconds before growing arc size");
11053 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, p_dampener_disable, INT, ZMOD_RW,
11054 "Disable arc_p adapt dampener");
11056 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, shrink_shift, param_set_arc_int,
11057 param_get_int, ZMOD_RW, "log2(fraction of arc to reclaim)");
11059 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, pc_percent, UINT, ZMOD_RW,
11060 "Percent of pagecache to reclaim arc to");
11062 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, p_min_shift, param_set_arc_int,
11063 param_get_int, ZMOD_RW, "arc_c shift to calc min/max arc_p");
11065 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, average_blocksize, INT, ZMOD_RD,
11066 "Target average block size");
11068 ZFS_MODULE_PARAM(zfs, zfs_, compressed_arc_enabled, INT, ZMOD_RW,
11069 "Disable compressed arc buffers");
11071 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prefetch_ms, param_set_arc_int,
11072 param_get_int, ZMOD_RW, "Min life of prefetch block in ms");
11074 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prescient_prefetch_ms,
11075 param_set_arc_int, param_get_int, ZMOD_RW,
11076 "Min life of prescient prefetched block in ms");
11078 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_max, ULONG, ZMOD_RW,
11079 "Max write bytes per interval");
11081 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_boost, ULONG, ZMOD_RW,
11082 "Extra write bytes during device warmup");
11084 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom, ULONG, ZMOD_RW,
11085 "Number of max device writes to precache");
11087 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom_boost, ULONG, ZMOD_RW,
11088 "Compressed l2arc_headroom multiplier");
11090 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, trim_ahead, ULONG, ZMOD_RW,
11091 "TRIM ahead L2ARC write size multiplier");
11093 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_secs, ULONG, ZMOD_RW,
11094 "Seconds between L2ARC writing");
11096 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_min_ms, ULONG, ZMOD_RW,
11097 "Min feed interval in milliseconds");
11099 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, noprefetch, INT, ZMOD_RW,
11100 "Skip caching prefetched buffers");
11102 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_again, INT, ZMOD_RW,
11103 "Turbo L2ARC warmup");
11105 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, norw, INT, ZMOD_RW,
11106 "No reads during writes");
11108 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, meta_percent, INT, ZMOD_RW,
11109 "Percent of ARC size allowed for L2ARC-only headers");
11111 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_enabled, INT, ZMOD_RW,
11112 "Rebuild the L2ARC when importing a pool");
11114 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_blocks_min_l2size, ULONG, ZMOD_RW,
11115 "Min size in bytes to write rebuild log blocks in L2ARC");
11117 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, mfuonly, INT, ZMOD_RW,
11118 "Cache only MFU data from ARC into L2ARC");
11120 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, lotsfree_percent, param_set_arc_int,
11121 param_get_int, ZMOD_RW, "System free memory I/O throttle in bytes");
11123 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, sys_free, param_set_arc_long,
11124 param_get_long, ZMOD_RW, "System free memory target size in bytes");
11126 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit, param_set_arc_long,
11127 param_get_long, ZMOD_RW, "Minimum bytes of dnodes in arc");
11129 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit_percent,
11130 param_set_arc_long, param_get_long, ZMOD_RW,
11131 "Percent of ARC meta buffers for dnodes");
11133 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, dnode_reduce_percent, ULONG, ZMOD_RW,
11134 "Percentage of excess dnodes to try to unpin");
11136 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, eviction_pct, INT, ZMOD_RW,
11137 "When full, ARC allocation waits for eviction of this % of alloc size");
11139 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, evict_batch_limit, INT, ZMOD_RW,
11140 "The number of headers to evict per sublist before moving to the next");