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;
332 static kmutex_t arc_evict_lock;
333 static boolean_t arc_evict_needed = B_FALSE;
336 * Count of bytes evicted since boot.
338 static uint64_t arc_evict_count;
341 * List of arc_evict_waiter_t's, representing threads waiting for the
342 * arc_evict_count to reach specific values.
344 static list_t arc_evict_waiters;
347 * When arc_is_overflowing(), arc_get_data_impl() waits for this percent of
348 * the requested amount of data to be evicted. For example, by default for
349 * every 2KB that's evicted, 1KB of it may be "reused" by a new allocation.
350 * Since this is above 100%, it ensures that progress is made towards getting
351 * arc_size under arc_c. Since this is finite, it ensures that allocations
352 * can still happen, even during the potentially long time that arc_size is
355 int zfs_arc_eviction_pct = 200;
358 * The number of headers to evict in arc_evict_state_impl() before
359 * dropping the sublist lock and evicting from another sublist. A lower
360 * value means we're more likely to evict the "correct" header (i.e. the
361 * oldest header in the arc state), but comes with higher overhead
362 * (i.e. more invocations of arc_evict_state_impl()).
364 int zfs_arc_evict_batch_limit = 10;
366 /* number of seconds before growing cache again */
367 int arc_grow_retry = 5;
370 * Minimum time between calls to arc_kmem_reap_soon().
372 int arc_kmem_cache_reap_retry_ms = 1000;
374 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
375 int zfs_arc_overflow_shift = 8;
377 /* shift of arc_c for calculating both min and max arc_p */
378 int arc_p_min_shift = 4;
380 /* log2(fraction of arc to reclaim) */
381 int arc_shrink_shift = 7;
383 /* percent of pagecache to reclaim arc to */
385 uint_t zfs_arc_pc_percent = 0;
389 * log2(fraction of ARC which must be free to allow growing).
390 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
391 * when reading a new block into the ARC, we will evict an equal-sized block
394 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
395 * we will still not allow it to grow.
397 int arc_no_grow_shift = 5;
401 * minimum lifespan of a prefetch block in clock ticks
402 * (initialized in arc_init())
404 static int arc_min_prefetch_ms;
405 static int arc_min_prescient_prefetch_ms;
408 * If this percent of memory is free, don't throttle.
410 int arc_lotsfree_percent = 10;
413 * The arc has filled available memory and has now warmed up.
418 * These tunables are for performance analysis.
420 unsigned long zfs_arc_max = 0;
421 unsigned long zfs_arc_min = 0;
422 unsigned long zfs_arc_meta_limit = 0;
423 unsigned long zfs_arc_meta_min = 0;
424 unsigned long zfs_arc_dnode_limit = 0;
425 unsigned long zfs_arc_dnode_reduce_percent = 10;
426 int zfs_arc_grow_retry = 0;
427 int zfs_arc_shrink_shift = 0;
428 int zfs_arc_p_min_shift = 0;
429 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
432 * ARC dirty data constraints for arc_tempreserve_space() throttle.
434 unsigned long zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */
435 unsigned long zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */
436 unsigned long zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */
439 * Enable or disable compressed arc buffers.
441 int zfs_compressed_arc_enabled = B_TRUE;
444 * ARC will evict meta buffers that exceed arc_meta_limit. This
445 * tunable make arc_meta_limit adjustable for different workloads.
447 unsigned long zfs_arc_meta_limit_percent = 75;
450 * Percentage that can be consumed by dnodes of ARC meta buffers.
452 unsigned long zfs_arc_dnode_limit_percent = 10;
455 * These tunables are Linux specific
457 unsigned long zfs_arc_sys_free = 0;
458 int zfs_arc_min_prefetch_ms = 0;
459 int zfs_arc_min_prescient_prefetch_ms = 0;
460 int zfs_arc_p_dampener_disable = 1;
461 int zfs_arc_meta_prune = 10000;
462 int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED;
463 int zfs_arc_meta_adjust_restarts = 4096;
464 int zfs_arc_lotsfree_percent = 10;
467 arc_state_t ARC_anon;
469 arc_state_t ARC_mru_ghost;
471 arc_state_t ARC_mfu_ghost;
472 arc_state_t ARC_l2c_only;
474 arc_stats_t arc_stats = {
475 { "hits", KSTAT_DATA_UINT64 },
476 { "misses", KSTAT_DATA_UINT64 },
477 { "demand_data_hits", KSTAT_DATA_UINT64 },
478 { "demand_data_misses", KSTAT_DATA_UINT64 },
479 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
480 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
481 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
482 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
483 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
484 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
485 { "mru_hits", KSTAT_DATA_UINT64 },
486 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
487 { "mfu_hits", KSTAT_DATA_UINT64 },
488 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
489 { "deleted", KSTAT_DATA_UINT64 },
490 { "mutex_miss", KSTAT_DATA_UINT64 },
491 { "access_skip", KSTAT_DATA_UINT64 },
492 { "evict_skip", KSTAT_DATA_UINT64 },
493 { "evict_not_enough", KSTAT_DATA_UINT64 },
494 { "evict_l2_cached", KSTAT_DATA_UINT64 },
495 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
496 { "evict_l2_eligible_mfu", KSTAT_DATA_UINT64 },
497 { "evict_l2_eligible_mru", KSTAT_DATA_UINT64 },
498 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
499 { "evict_l2_skip", KSTAT_DATA_UINT64 },
500 { "hash_elements", KSTAT_DATA_UINT64 },
501 { "hash_elements_max", KSTAT_DATA_UINT64 },
502 { "hash_collisions", KSTAT_DATA_UINT64 },
503 { "hash_chains", KSTAT_DATA_UINT64 },
504 { "hash_chain_max", KSTAT_DATA_UINT64 },
505 { "p", KSTAT_DATA_UINT64 },
506 { "c", KSTAT_DATA_UINT64 },
507 { "c_min", KSTAT_DATA_UINT64 },
508 { "c_max", KSTAT_DATA_UINT64 },
509 { "size", KSTAT_DATA_UINT64 },
510 { "compressed_size", KSTAT_DATA_UINT64 },
511 { "uncompressed_size", KSTAT_DATA_UINT64 },
512 { "overhead_size", KSTAT_DATA_UINT64 },
513 { "hdr_size", KSTAT_DATA_UINT64 },
514 { "data_size", KSTAT_DATA_UINT64 },
515 { "metadata_size", KSTAT_DATA_UINT64 },
516 { "dbuf_size", KSTAT_DATA_UINT64 },
517 { "dnode_size", KSTAT_DATA_UINT64 },
518 { "bonus_size", KSTAT_DATA_UINT64 },
519 #if defined(COMPAT_FREEBSD11)
520 { "other_size", KSTAT_DATA_UINT64 },
522 { "anon_size", KSTAT_DATA_UINT64 },
523 { "anon_evictable_data", KSTAT_DATA_UINT64 },
524 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
525 { "mru_size", KSTAT_DATA_UINT64 },
526 { "mru_evictable_data", KSTAT_DATA_UINT64 },
527 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
528 { "mru_ghost_size", KSTAT_DATA_UINT64 },
529 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
530 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
531 { "mfu_size", KSTAT_DATA_UINT64 },
532 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
533 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
534 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
535 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
536 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
537 { "l2_hits", KSTAT_DATA_UINT64 },
538 { "l2_misses", KSTAT_DATA_UINT64 },
539 { "l2_prefetch_asize", KSTAT_DATA_UINT64 },
540 { "l2_mru_asize", KSTAT_DATA_UINT64 },
541 { "l2_mfu_asize", KSTAT_DATA_UINT64 },
542 { "l2_bufc_data_asize", KSTAT_DATA_UINT64 },
543 { "l2_bufc_metadata_asize", KSTAT_DATA_UINT64 },
544 { "l2_feeds", KSTAT_DATA_UINT64 },
545 { "l2_rw_clash", KSTAT_DATA_UINT64 },
546 { "l2_read_bytes", KSTAT_DATA_UINT64 },
547 { "l2_write_bytes", KSTAT_DATA_UINT64 },
548 { "l2_writes_sent", KSTAT_DATA_UINT64 },
549 { "l2_writes_done", KSTAT_DATA_UINT64 },
550 { "l2_writes_error", KSTAT_DATA_UINT64 },
551 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
552 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
553 { "l2_evict_reading", KSTAT_DATA_UINT64 },
554 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
555 { "l2_free_on_write", KSTAT_DATA_UINT64 },
556 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
557 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
558 { "l2_io_error", KSTAT_DATA_UINT64 },
559 { "l2_size", KSTAT_DATA_UINT64 },
560 { "l2_asize", KSTAT_DATA_UINT64 },
561 { "l2_hdr_size", KSTAT_DATA_UINT64 },
562 { "l2_log_blk_writes", KSTAT_DATA_UINT64 },
563 { "l2_log_blk_avg_asize", KSTAT_DATA_UINT64 },
564 { "l2_log_blk_asize", KSTAT_DATA_UINT64 },
565 { "l2_log_blk_count", KSTAT_DATA_UINT64 },
566 { "l2_data_to_meta_ratio", KSTAT_DATA_UINT64 },
567 { "l2_rebuild_success", KSTAT_DATA_UINT64 },
568 { "l2_rebuild_unsupported", KSTAT_DATA_UINT64 },
569 { "l2_rebuild_io_errors", KSTAT_DATA_UINT64 },
570 { "l2_rebuild_dh_errors", KSTAT_DATA_UINT64 },
571 { "l2_rebuild_cksum_lb_errors", KSTAT_DATA_UINT64 },
572 { "l2_rebuild_lowmem", KSTAT_DATA_UINT64 },
573 { "l2_rebuild_size", KSTAT_DATA_UINT64 },
574 { "l2_rebuild_asize", KSTAT_DATA_UINT64 },
575 { "l2_rebuild_bufs", KSTAT_DATA_UINT64 },
576 { "l2_rebuild_bufs_precached", KSTAT_DATA_UINT64 },
577 { "l2_rebuild_log_blks", KSTAT_DATA_UINT64 },
578 { "memory_throttle_count", KSTAT_DATA_UINT64 },
579 { "memory_direct_count", KSTAT_DATA_UINT64 },
580 { "memory_indirect_count", KSTAT_DATA_UINT64 },
581 { "memory_all_bytes", KSTAT_DATA_UINT64 },
582 { "memory_free_bytes", KSTAT_DATA_UINT64 },
583 { "memory_available_bytes", KSTAT_DATA_INT64 },
584 { "arc_no_grow", KSTAT_DATA_UINT64 },
585 { "arc_tempreserve", KSTAT_DATA_UINT64 },
586 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
587 { "arc_prune", KSTAT_DATA_UINT64 },
588 { "arc_meta_used", KSTAT_DATA_UINT64 },
589 { "arc_meta_limit", KSTAT_DATA_UINT64 },
590 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
591 { "arc_meta_max", KSTAT_DATA_UINT64 },
592 { "arc_meta_min", KSTAT_DATA_UINT64 },
593 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
594 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
595 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
596 { "arc_need_free", KSTAT_DATA_UINT64 },
597 { "arc_sys_free", KSTAT_DATA_UINT64 },
598 { "arc_raw_size", KSTAT_DATA_UINT64 },
599 { "cached_only_in_progress", KSTAT_DATA_UINT64 },
600 { "abd_chunk_waste_size", KSTAT_DATA_UINT64 },
605 #define ARCSTAT_MAX(stat, val) { \
607 while ((val) > (m = arc_stats.stat.value.ui64) && \
608 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
613 * We define a macro to allow ARC hits/misses to be easily broken down by
614 * two separate conditions, giving a total of four different subtypes for
615 * each of hits and misses (so eight statistics total).
617 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
620 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
622 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
626 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
628 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
633 * This macro allows us to use kstats as floating averages. Each time we
634 * update this kstat, we first factor it and the update value by
635 * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall
636 * average. This macro assumes that integer loads and stores are atomic, but
637 * is not safe for multiple writers updating the kstat in parallel (only the
638 * last writer's update will remain).
640 #define ARCSTAT_F_AVG_FACTOR 3
641 #define ARCSTAT_F_AVG(stat, value) \
643 uint64_t x = ARCSTAT(stat); \
644 x = x - x / ARCSTAT_F_AVG_FACTOR + \
645 (value) / ARCSTAT_F_AVG_FACTOR; \
652 * There are several ARC variables that are critical to export as kstats --
653 * but we don't want to have to grovel around in the kstat whenever we wish to
654 * manipulate them. For these variables, we therefore define them to be in
655 * terms of the statistic variable. This assures that we are not introducing
656 * the possibility of inconsistency by having shadow copies of the variables,
657 * while still allowing the code to be readable.
659 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
660 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
661 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
662 /* max size for dnodes */
663 #define arc_dnode_size_limit ARCSTAT(arcstat_dnode_limit)
664 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
665 #define arc_need_free ARCSTAT(arcstat_need_free) /* waiting to be evicted */
667 hrtime_t arc_growtime;
668 list_t arc_prune_list;
669 kmutex_t arc_prune_mtx;
670 taskq_t *arc_prune_taskq;
672 #define GHOST_STATE(state) \
673 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
674 (state) == arc_l2c_only)
676 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
677 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
678 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
679 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
680 #define HDR_PRESCIENT_PREFETCH(hdr) \
681 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
682 #define HDR_COMPRESSION_ENABLED(hdr) \
683 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
685 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
686 #define HDR_L2_READING(hdr) \
687 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
688 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
689 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
690 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
691 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
692 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
693 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
694 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
696 #define HDR_ISTYPE_METADATA(hdr) \
697 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
698 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
700 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
701 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
702 #define HDR_HAS_RABD(hdr) \
703 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
704 (hdr)->b_crypt_hdr.b_rabd != NULL)
705 #define HDR_ENCRYPTED(hdr) \
706 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
707 #define HDR_AUTHENTICATED(hdr) \
708 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
710 /* For storing compression mode in b_flags */
711 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
713 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
714 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
715 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
716 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
718 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
719 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
720 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
721 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
727 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
728 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
729 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
732 * Hash table routines
735 #define BUF_LOCKS 2048
736 typedef struct buf_hash_table {
738 arc_buf_hdr_t **ht_table;
739 kmutex_t ht_locks[BUF_LOCKS] ____cacheline_aligned;
742 static buf_hash_table_t buf_hash_table;
744 #define BUF_HASH_INDEX(spa, dva, birth) \
745 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
746 #define BUF_HASH_LOCK(idx) (&buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
747 #define HDR_LOCK(hdr) \
748 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
750 uint64_t zfs_crc64_table[256];
756 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
757 #define L2ARC_HEADROOM 2 /* num of writes */
760 * If we discover during ARC scan any buffers to be compressed, we boost
761 * our headroom for the next scanning cycle by this percentage multiple.
763 #define L2ARC_HEADROOM_BOOST 200
764 #define L2ARC_FEED_SECS 1 /* caching interval secs */
765 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
768 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
769 * and each of the state has two types: data and metadata.
771 #define L2ARC_FEED_TYPES 4
773 /* L2ARC Performance Tunables */
774 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
775 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
776 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
777 unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
778 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
779 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
780 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
781 int l2arc_feed_again = B_TRUE; /* turbo warmup */
782 int l2arc_norw = B_FALSE; /* no reads during writes */
783 int l2arc_meta_percent = 33; /* limit on headers size */
788 static list_t L2ARC_dev_list; /* device list */
789 static list_t *l2arc_dev_list; /* device list pointer */
790 static kmutex_t l2arc_dev_mtx; /* device list mutex */
791 static l2arc_dev_t *l2arc_dev_last; /* last device used */
792 static list_t L2ARC_free_on_write; /* free after write buf list */
793 static list_t *l2arc_free_on_write; /* free after write list ptr */
794 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
795 static uint64_t l2arc_ndev; /* number of devices */
797 typedef struct l2arc_read_callback {
798 arc_buf_hdr_t *l2rcb_hdr; /* read header */
799 blkptr_t l2rcb_bp; /* original blkptr */
800 zbookmark_phys_t l2rcb_zb; /* original bookmark */
801 int l2rcb_flags; /* original flags */
802 abd_t *l2rcb_abd; /* temporary buffer */
803 } l2arc_read_callback_t;
805 typedef struct l2arc_data_free {
806 /* protected by l2arc_free_on_write_mtx */
809 arc_buf_contents_t l2df_type;
810 list_node_t l2df_list_node;
813 typedef enum arc_fill_flags {
814 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
815 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
816 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
817 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
818 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
821 typedef enum arc_ovf_level {
822 ARC_OVF_NONE, /* ARC within target size. */
823 ARC_OVF_SOME, /* ARC is slightly overflowed. */
824 ARC_OVF_SEVERE /* ARC is severely overflowed. */
827 static kmutex_t l2arc_feed_thr_lock;
828 static kcondvar_t l2arc_feed_thr_cv;
829 static uint8_t l2arc_thread_exit;
831 static kmutex_t l2arc_rebuild_thr_lock;
832 static kcondvar_t l2arc_rebuild_thr_cv;
834 enum arc_hdr_alloc_flags {
835 ARC_HDR_ALLOC_RDATA = 0x1,
836 ARC_HDR_DO_ADAPT = 0x2,
837 ARC_HDR_USE_RESERVE = 0x4,
841 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *, int);
842 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
843 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *, int);
844 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
845 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
846 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
847 static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t);
848 static void arc_hdr_alloc_abd(arc_buf_hdr_t *, int);
849 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
850 static void arc_buf_watch(arc_buf_t *);
852 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
853 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
854 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
855 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
857 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
858 static void l2arc_read_done(zio_t *);
859 static void l2arc_do_free_on_write(void);
860 static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
861 boolean_t state_only);
863 #define l2arc_hdr_arcstats_increment(hdr) \
864 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_FALSE)
865 #define l2arc_hdr_arcstats_decrement(hdr) \
866 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_FALSE)
867 #define l2arc_hdr_arcstats_increment_state(hdr) \
868 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_TRUE)
869 #define l2arc_hdr_arcstats_decrement_state(hdr) \
870 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_TRUE)
873 * l2arc_mfuonly : A ZFS module parameter that controls whether only MFU
874 * metadata and data are cached from ARC into L2ARC.
876 int l2arc_mfuonly = 0;
880 * l2arc_trim_ahead : A ZFS module parameter that controls how much ahead of
881 * the current write size (l2arc_write_max) we should TRIM if we
882 * have filled the device. It is defined as a percentage of the
883 * write size. If set to 100 we trim twice the space required to
884 * accommodate upcoming writes. A minimum of 64MB will be trimmed.
885 * It also enables TRIM of the whole L2ARC device upon creation or
886 * addition to an existing pool or if the header of the device is
887 * invalid upon importing a pool or onlining a cache device. The
888 * default is 0, which disables TRIM on L2ARC altogether as it can
889 * put significant stress on the underlying storage devices. This
890 * will vary depending of how well the specific device handles
893 unsigned long l2arc_trim_ahead = 0;
896 * Performance tuning of L2ARC persistence:
898 * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding
899 * an L2ARC device (either at pool import or later) will attempt
900 * to rebuild L2ARC buffer contents.
901 * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls
902 * whether log blocks are written to the L2ARC device. If the L2ARC
903 * device is less than 1GB, the amount of data l2arc_evict()
904 * evicts is significant compared to the amount of restored L2ARC
905 * data. In this case do not write log blocks in L2ARC in order
906 * not to waste space.
908 int l2arc_rebuild_enabled = B_TRUE;
909 unsigned long l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024;
911 /* L2ARC persistence rebuild control routines. */
912 void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen);
913 static void l2arc_dev_rebuild_thread(void *arg);
914 static int l2arc_rebuild(l2arc_dev_t *dev);
916 /* L2ARC persistence read I/O routines. */
917 static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
918 static int l2arc_log_blk_read(l2arc_dev_t *dev,
919 const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
920 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
921 zio_t *this_io, zio_t **next_io);
922 static zio_t *l2arc_log_blk_fetch(vdev_t *vd,
923 const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb);
924 static void l2arc_log_blk_fetch_abort(zio_t *zio);
926 /* L2ARC persistence block restoration routines. */
927 static void l2arc_log_blk_restore(l2arc_dev_t *dev,
928 const l2arc_log_blk_phys_t *lb, uint64_t lb_asize);
929 static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
932 /* L2ARC persistence write I/O routines. */
933 static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
934 l2arc_write_callback_t *cb);
936 /* L2ARC persistence auxiliary routines. */
937 boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
938 const l2arc_log_blkptr_t *lbp);
939 static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
940 const arc_buf_hdr_t *ab);
941 boolean_t l2arc_range_check_overlap(uint64_t bottom,
942 uint64_t top, uint64_t check);
943 static void l2arc_blk_fetch_done(zio_t *zio);
944 static inline uint64_t
945 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev);
948 * We use Cityhash for this. It's fast, and has good hash properties without
949 * requiring any large static buffers.
952 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
954 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
957 #define HDR_EMPTY(hdr) \
958 ((hdr)->b_dva.dva_word[0] == 0 && \
959 (hdr)->b_dva.dva_word[1] == 0)
961 #define HDR_EMPTY_OR_LOCKED(hdr) \
962 (HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr)))
964 #define HDR_EQUAL(spa, dva, birth, hdr) \
965 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
966 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
967 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
970 buf_discard_identity(arc_buf_hdr_t *hdr)
972 hdr->b_dva.dva_word[0] = 0;
973 hdr->b_dva.dva_word[1] = 0;
977 static arc_buf_hdr_t *
978 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
980 const dva_t *dva = BP_IDENTITY(bp);
981 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
982 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
983 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
986 mutex_enter(hash_lock);
987 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
988 hdr = hdr->b_hash_next) {
989 if (HDR_EQUAL(spa, dva, birth, hdr)) {
994 mutex_exit(hash_lock);
1000 * Insert an entry into the hash table. If there is already an element
1001 * equal to elem in the hash table, then the already existing element
1002 * will be returned and the new element will not be inserted.
1003 * Otherwise returns NULL.
1004 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1006 static arc_buf_hdr_t *
1007 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1009 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1010 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1011 arc_buf_hdr_t *fhdr;
1014 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1015 ASSERT(hdr->b_birth != 0);
1016 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1018 if (lockp != NULL) {
1020 mutex_enter(hash_lock);
1022 ASSERT(MUTEX_HELD(hash_lock));
1025 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1026 fhdr = fhdr->b_hash_next, i++) {
1027 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1031 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1032 buf_hash_table.ht_table[idx] = hdr;
1033 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1035 /* collect some hash table performance data */
1037 ARCSTAT_BUMP(arcstat_hash_collisions);
1039 ARCSTAT_BUMP(arcstat_hash_chains);
1041 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1043 uint64_t he = atomic_inc_64_nv(
1044 &arc_stats.arcstat_hash_elements.value.ui64);
1045 ARCSTAT_MAX(arcstat_hash_elements_max, he);
1051 buf_hash_remove(arc_buf_hdr_t *hdr)
1053 arc_buf_hdr_t *fhdr, **hdrp;
1054 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1056 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1057 ASSERT(HDR_IN_HASH_TABLE(hdr));
1059 hdrp = &buf_hash_table.ht_table[idx];
1060 while ((fhdr = *hdrp) != hdr) {
1061 ASSERT3P(fhdr, !=, NULL);
1062 hdrp = &fhdr->b_hash_next;
1064 *hdrp = hdr->b_hash_next;
1065 hdr->b_hash_next = NULL;
1066 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1068 /* collect some hash table performance data */
1069 atomic_dec_64(&arc_stats.arcstat_hash_elements.value.ui64);
1071 if (buf_hash_table.ht_table[idx] &&
1072 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1073 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1077 * Global data structures and functions for the buf kmem cache.
1080 static kmem_cache_t *hdr_full_cache;
1081 static kmem_cache_t *hdr_full_crypt_cache;
1082 static kmem_cache_t *hdr_l2only_cache;
1083 static kmem_cache_t *buf_cache;
1090 #if defined(_KERNEL)
1092 * Large allocations which do not require contiguous pages
1093 * should be using vmem_free() in the linux kernel\
1095 vmem_free(buf_hash_table.ht_table,
1096 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1098 kmem_free(buf_hash_table.ht_table,
1099 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1101 for (i = 0; i < BUF_LOCKS; i++)
1102 mutex_destroy(BUF_HASH_LOCK(i));
1103 kmem_cache_destroy(hdr_full_cache);
1104 kmem_cache_destroy(hdr_full_crypt_cache);
1105 kmem_cache_destroy(hdr_l2only_cache);
1106 kmem_cache_destroy(buf_cache);
1110 * Constructor callback - called when the cache is empty
1111 * and a new buf is requested.
1115 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1117 arc_buf_hdr_t *hdr = vbuf;
1119 bzero(hdr, HDR_FULL_SIZE);
1120 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1121 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1122 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1123 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1124 list_link_init(&hdr->b_l1hdr.b_arc_node);
1125 list_link_init(&hdr->b_l2hdr.b_l2node);
1126 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1127 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1134 hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
1136 arc_buf_hdr_t *hdr = vbuf;
1138 hdr_full_cons(vbuf, unused, kmflag);
1139 bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr));
1140 arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1147 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1149 arc_buf_hdr_t *hdr = vbuf;
1151 bzero(hdr, HDR_L2ONLY_SIZE);
1152 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1159 buf_cons(void *vbuf, void *unused, int kmflag)
1161 arc_buf_t *buf = vbuf;
1163 bzero(buf, sizeof (arc_buf_t));
1164 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1165 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1171 * Destructor callback - called when a cached buf is
1172 * no longer required.
1176 hdr_full_dest(void *vbuf, void *unused)
1178 arc_buf_hdr_t *hdr = vbuf;
1180 ASSERT(HDR_EMPTY(hdr));
1181 cv_destroy(&hdr->b_l1hdr.b_cv);
1182 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1183 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1184 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1185 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1190 hdr_full_crypt_dest(void *vbuf, void *unused)
1192 arc_buf_hdr_t *hdr = vbuf;
1194 hdr_full_dest(vbuf, unused);
1195 arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1200 hdr_l2only_dest(void *vbuf, void *unused)
1202 arc_buf_hdr_t *hdr __maybe_unused = vbuf;
1204 ASSERT(HDR_EMPTY(hdr));
1205 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1210 buf_dest(void *vbuf, void *unused)
1212 arc_buf_t *buf = vbuf;
1214 mutex_destroy(&buf->b_evict_lock);
1215 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1221 uint64_t *ct = NULL;
1222 uint64_t hsize = 1ULL << 12;
1226 * The hash table is big enough to fill all of physical memory
1227 * with an average block size of zfs_arc_average_blocksize (default 8K).
1228 * By default, the table will take up
1229 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1231 while (hsize * zfs_arc_average_blocksize < arc_all_memory())
1234 buf_hash_table.ht_mask = hsize - 1;
1235 #if defined(_KERNEL)
1237 * Large allocations which do not require contiguous pages
1238 * should be using vmem_alloc() in the linux kernel
1240 buf_hash_table.ht_table =
1241 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
1243 buf_hash_table.ht_table =
1244 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1246 if (buf_hash_table.ht_table == NULL) {
1247 ASSERT(hsize > (1ULL << 8));
1252 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1253 0, hdr_full_cons, hdr_full_dest, NULL, NULL, NULL, 0);
1254 hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
1255 HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
1256 NULL, NULL, NULL, 0);
1257 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1258 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, NULL,
1260 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1261 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1263 for (i = 0; i < 256; i++)
1264 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1265 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1267 for (i = 0; i < BUF_LOCKS; i++)
1268 mutex_init(BUF_HASH_LOCK(i), NULL, MUTEX_DEFAULT, NULL);
1271 #define ARC_MINTIME (hz>>4) /* 62 ms */
1274 * This is the size that the buf occupies in memory. If the buf is compressed,
1275 * it will correspond to the compressed size. You should use this method of
1276 * getting the buf size unless you explicitly need the logical size.
1279 arc_buf_size(arc_buf_t *buf)
1281 return (ARC_BUF_COMPRESSED(buf) ?
1282 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1286 arc_buf_lsize(arc_buf_t *buf)
1288 return (HDR_GET_LSIZE(buf->b_hdr));
1292 * This function will return B_TRUE if the buffer is encrypted in memory.
1293 * This buffer can be decrypted by calling arc_untransform().
1296 arc_is_encrypted(arc_buf_t *buf)
1298 return (ARC_BUF_ENCRYPTED(buf) != 0);
1302 * Returns B_TRUE if the buffer represents data that has not had its MAC
1306 arc_is_unauthenticated(arc_buf_t *buf)
1308 return (HDR_NOAUTH(buf->b_hdr) != 0);
1312 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1313 uint8_t *iv, uint8_t *mac)
1315 arc_buf_hdr_t *hdr = buf->b_hdr;
1317 ASSERT(HDR_PROTECTED(hdr));
1319 bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
1320 bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
1321 bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
1322 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1323 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1327 * Indicates how this buffer is compressed in memory. If it is not compressed
1328 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1329 * arc_untransform() as long as it is also unencrypted.
1332 arc_get_compression(arc_buf_t *buf)
1334 return (ARC_BUF_COMPRESSED(buf) ?
1335 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1339 * Return the compression algorithm used to store this data in the ARC. If ARC
1340 * compression is enabled or this is an encrypted block, this will be the same
1341 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1343 static inline enum zio_compress
1344 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1346 return (HDR_COMPRESSION_ENABLED(hdr) ?
1347 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1351 arc_get_complevel(arc_buf_t *buf)
1353 return (buf->b_hdr->b_complevel);
1356 static inline boolean_t
1357 arc_buf_is_shared(arc_buf_t *buf)
1359 boolean_t shared = (buf->b_data != NULL &&
1360 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1361 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1362 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1363 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1364 IMPLY(shared, ARC_BUF_SHARED(buf));
1365 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1368 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1369 * already being shared" requirement prevents us from doing that.
1376 * Free the checksum associated with this header. If there is no checksum, this
1380 arc_cksum_free(arc_buf_hdr_t *hdr)
1382 ASSERT(HDR_HAS_L1HDR(hdr));
1384 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1385 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1386 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1387 hdr->b_l1hdr.b_freeze_cksum = NULL;
1389 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1393 * Return true iff at least one of the bufs on hdr is not compressed.
1394 * Encrypted buffers count as compressed.
1397 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1399 ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr));
1401 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1402 if (!ARC_BUF_COMPRESSED(b)) {
1411 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1412 * matches the checksum that is stored in the hdr. If there is no checksum,
1413 * or if the buf is compressed, this is a no-op.
1416 arc_cksum_verify(arc_buf_t *buf)
1418 arc_buf_hdr_t *hdr = buf->b_hdr;
1421 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1424 if (ARC_BUF_COMPRESSED(buf))
1427 ASSERT(HDR_HAS_L1HDR(hdr));
1429 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1431 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1432 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1436 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1437 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1438 panic("buffer modified while frozen!");
1439 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1443 * This function makes the assumption that data stored in the L2ARC
1444 * will be transformed exactly as it is in the main pool. Because of
1445 * this we can verify the checksum against the reading process's bp.
1448 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1450 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1451 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1454 * Block pointers always store the checksum for the logical data.
1455 * If the block pointer has the gang bit set, then the checksum
1456 * it represents is for the reconstituted data and not for an
1457 * individual gang member. The zio pipeline, however, must be able to
1458 * determine the checksum of each of the gang constituents so it
1459 * treats the checksum comparison differently than what we need
1460 * for l2arc blocks. This prevents us from using the
1461 * zio_checksum_error() interface directly. Instead we must call the
1462 * zio_checksum_error_impl() so that we can ensure the checksum is
1463 * generated using the correct checksum algorithm and accounts for the
1464 * logical I/O size and not just a gang fragment.
1466 return (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1467 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1468 zio->io_offset, NULL) == 0);
1472 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1473 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1474 * isn't modified later on. If buf is compressed or there is already a checksum
1475 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1478 arc_cksum_compute(arc_buf_t *buf)
1480 arc_buf_hdr_t *hdr = buf->b_hdr;
1482 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1485 ASSERT(HDR_HAS_L1HDR(hdr));
1487 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1488 if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1489 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1493 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1494 ASSERT(!ARC_BUF_COMPRESSED(buf));
1495 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1497 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1498 hdr->b_l1hdr.b_freeze_cksum);
1499 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1505 arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
1507 panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr);
1513 arc_buf_unwatch(arc_buf_t *buf)
1517 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1518 PROT_READ | PROT_WRITE));
1525 arc_buf_watch(arc_buf_t *buf)
1529 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1534 static arc_buf_contents_t
1535 arc_buf_type(arc_buf_hdr_t *hdr)
1537 arc_buf_contents_t type;
1538 if (HDR_ISTYPE_METADATA(hdr)) {
1539 type = ARC_BUFC_METADATA;
1541 type = ARC_BUFC_DATA;
1543 VERIFY3U(hdr->b_type, ==, type);
1548 arc_is_metadata(arc_buf_t *buf)
1550 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1554 arc_bufc_to_flags(arc_buf_contents_t type)
1558 /* metadata field is 0 if buffer contains normal data */
1560 case ARC_BUFC_METADATA:
1561 return (ARC_FLAG_BUFC_METADATA);
1565 panic("undefined ARC buffer type!");
1566 return ((uint32_t)-1);
1570 arc_buf_thaw(arc_buf_t *buf)
1572 arc_buf_hdr_t *hdr = buf->b_hdr;
1574 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1575 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1577 arc_cksum_verify(buf);
1580 * Compressed buffers do not manipulate the b_freeze_cksum.
1582 if (ARC_BUF_COMPRESSED(buf))
1585 ASSERT(HDR_HAS_L1HDR(hdr));
1586 arc_cksum_free(hdr);
1587 arc_buf_unwatch(buf);
1591 arc_buf_freeze(arc_buf_t *buf)
1593 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1596 if (ARC_BUF_COMPRESSED(buf))
1599 ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1600 arc_cksum_compute(buf);
1604 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1605 * the following functions should be used to ensure that the flags are
1606 * updated in a thread-safe way. When manipulating the flags either
1607 * the hash_lock must be held or the hdr must be undiscoverable. This
1608 * ensures that we're not racing with any other threads when updating
1612 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1614 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1615 hdr->b_flags |= flags;
1619 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1621 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1622 hdr->b_flags &= ~flags;
1626 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1627 * done in a special way since we have to clear and set bits
1628 * at the same time. Consumers that wish to set the compression bits
1629 * must use this function to ensure that the flags are updated in
1630 * thread-safe manner.
1633 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1635 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1638 * Holes and embedded blocks will always have a psize = 0 so
1639 * we ignore the compression of the blkptr and set the
1640 * want to uncompress them. Mark them as uncompressed.
1642 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1643 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1644 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1646 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1647 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1650 HDR_SET_COMPRESS(hdr, cmp);
1651 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1655 * Looks for another buf on the same hdr which has the data decompressed, copies
1656 * from it, and returns true. If no such buf exists, returns false.
1659 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1661 arc_buf_hdr_t *hdr = buf->b_hdr;
1662 boolean_t copied = B_FALSE;
1664 ASSERT(HDR_HAS_L1HDR(hdr));
1665 ASSERT3P(buf->b_data, !=, NULL);
1666 ASSERT(!ARC_BUF_COMPRESSED(buf));
1668 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1669 from = from->b_next) {
1670 /* can't use our own data buffer */
1675 if (!ARC_BUF_COMPRESSED(from)) {
1676 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1683 * There were no decompressed bufs, so there should not be a
1684 * checksum on the hdr either.
1686 if (zfs_flags & ZFS_DEBUG_MODIFY)
1687 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1693 * Allocates an ARC buf header that's in an evicted & L2-cached state.
1694 * This is used during l2arc reconstruction to make empty ARC buffers
1695 * which circumvent the regular disk->arc->l2arc path and instead come
1696 * into being in the reverse order, i.e. l2arc->arc.
1698 static arc_buf_hdr_t *
1699 arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
1700 dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth,
1701 enum zio_compress compress, uint8_t complevel, boolean_t protected,
1702 boolean_t prefetch, arc_state_type_t arcs_state)
1707 hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP);
1708 hdr->b_birth = birth;
1711 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR);
1712 HDR_SET_LSIZE(hdr, size);
1713 HDR_SET_PSIZE(hdr, psize);
1714 arc_hdr_set_compress(hdr, compress);
1715 hdr->b_complevel = complevel;
1717 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
1719 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
1720 hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa);
1724 hdr->b_l2hdr.b_dev = dev;
1725 hdr->b_l2hdr.b_daddr = daddr;
1726 hdr->b_l2hdr.b_arcs_state = arcs_state;
1732 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1735 arc_hdr_size(arc_buf_hdr_t *hdr)
1739 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1740 HDR_GET_PSIZE(hdr) > 0) {
1741 size = HDR_GET_PSIZE(hdr);
1743 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1744 size = HDR_GET_LSIZE(hdr);
1750 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1754 uint64_t lsize = HDR_GET_LSIZE(hdr);
1755 uint64_t psize = HDR_GET_PSIZE(hdr);
1756 void *tmpbuf = NULL;
1757 abd_t *abd = hdr->b_l1hdr.b_pabd;
1759 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1760 ASSERT(HDR_AUTHENTICATED(hdr));
1761 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1764 * The MAC is calculated on the compressed data that is stored on disk.
1765 * However, if compressed arc is disabled we will only have the
1766 * decompressed data available to us now. Compress it into a temporary
1767 * abd so we can verify the MAC. The performance overhead of this will
1768 * be relatively low, since most objects in an encrypted objset will
1769 * be encrypted (instead of authenticated) anyway.
1771 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1772 !HDR_COMPRESSION_ENABLED(hdr)) {
1773 tmpbuf = zio_buf_alloc(lsize);
1774 abd = abd_get_from_buf(tmpbuf, lsize);
1775 abd_take_ownership_of_buf(abd, B_TRUE);
1776 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1777 hdr->b_l1hdr.b_pabd, tmpbuf, lsize, hdr->b_complevel);
1778 ASSERT3U(csize, <=, psize);
1779 abd_zero_off(abd, csize, psize - csize);
1783 * Authentication is best effort. We authenticate whenever the key is
1784 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1786 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1787 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1788 ASSERT3U(lsize, ==, psize);
1789 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1790 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1792 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1793 hdr->b_crypt_hdr.b_mac);
1797 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1798 else if (ret != ENOENT)
1814 * This function will take a header that only has raw encrypted data in
1815 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1816 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1817 * also decompress the data.
1820 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1825 boolean_t no_crypt = B_FALSE;
1826 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1828 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1829 ASSERT(HDR_ENCRYPTED(hdr));
1831 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT);
1833 ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1834 B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1835 hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1836 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1841 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1842 HDR_GET_PSIZE(hdr));
1846 * If this header has disabled arc compression but the b_pabd is
1847 * compressed after decrypting it, we need to decompress the newly
1850 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1851 !HDR_COMPRESSION_ENABLED(hdr)) {
1853 * We want to make sure that we are correctly honoring the
1854 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1855 * and then loan a buffer from it, rather than allocating a
1856 * linear buffer and wrapping it in an abd later.
1858 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
1860 tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
1862 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1863 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
1864 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
1866 abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
1870 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
1871 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1872 arc_hdr_size(hdr), hdr);
1873 hdr->b_l1hdr.b_pabd = cabd;
1879 arc_hdr_free_abd(hdr, B_FALSE);
1881 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
1887 * This function is called during arc_buf_fill() to prepare the header's
1888 * abd plaintext pointer for use. This involves authenticated protected
1889 * data and decrypting encrypted data into the plaintext abd.
1892 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
1893 const zbookmark_phys_t *zb, boolean_t noauth)
1897 ASSERT(HDR_PROTECTED(hdr));
1899 if (hash_lock != NULL)
1900 mutex_enter(hash_lock);
1902 if (HDR_NOAUTH(hdr) && !noauth) {
1904 * The caller requested authenticated data but our data has
1905 * not been authenticated yet. Verify the MAC now if we can.
1907 ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
1910 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
1912 * If we only have the encrypted version of the data, but the
1913 * unencrypted version was requested we take this opportunity
1914 * to store the decrypted version in the header for future use.
1916 ret = arc_hdr_decrypt(hdr, spa, zb);
1921 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1923 if (hash_lock != NULL)
1924 mutex_exit(hash_lock);
1929 if (hash_lock != NULL)
1930 mutex_exit(hash_lock);
1936 * This function is used by the dbuf code to decrypt bonus buffers in place.
1937 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
1938 * block, so we use the hash lock here to protect against concurrent calls to
1942 arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock)
1944 arc_buf_hdr_t *hdr = buf->b_hdr;
1946 ASSERT(HDR_ENCRYPTED(hdr));
1947 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1948 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1949 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1951 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
1953 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
1954 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1955 hdr->b_crypt_hdr.b_ebufcnt -= 1;
1959 * Given a buf that has a data buffer attached to it, this function will
1960 * efficiently fill the buf with data of the specified compression setting from
1961 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1962 * are already sharing a data buf, no copy is performed.
1964 * If the buf is marked as compressed but uncompressed data was requested, this
1965 * will allocate a new data buffer for the buf, remove that flag, and fill the
1966 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1967 * uncompressed data, and (since we haven't added support for it yet) if you
1968 * want compressed data your buf must already be marked as compressed and have
1969 * the correct-sized data buffer.
1972 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
1973 arc_fill_flags_t flags)
1976 arc_buf_hdr_t *hdr = buf->b_hdr;
1977 boolean_t hdr_compressed =
1978 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
1979 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
1980 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
1981 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
1982 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
1984 ASSERT3P(buf->b_data, !=, NULL);
1985 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
1986 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
1987 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
1988 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
1989 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
1990 IMPLY(encrypted, !ARC_BUF_SHARED(buf));
1993 * If the caller wanted encrypted data we just need to copy it from
1994 * b_rabd and potentially byteswap it. We won't be able to do any
1995 * further transforms on it.
1998 ASSERT(HDR_HAS_RABD(hdr));
1999 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
2000 HDR_GET_PSIZE(hdr));
2005 * Adjust encrypted and authenticated headers to accommodate
2006 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
2007 * allowed to fail decryption due to keys not being loaded
2008 * without being marked as an IO error.
2010 if (HDR_PROTECTED(hdr)) {
2011 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
2012 zb, !!(flags & ARC_FILL_NOAUTH));
2013 if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
2015 } else if (error != 0) {
2016 if (hash_lock != NULL)
2017 mutex_enter(hash_lock);
2018 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2019 if (hash_lock != NULL)
2020 mutex_exit(hash_lock);
2026 * There is a special case here for dnode blocks which are
2027 * decrypting their bonus buffers. These blocks may request to
2028 * be decrypted in-place. This is necessary because there may
2029 * be many dnodes pointing into this buffer and there is
2030 * currently no method to synchronize replacing the backing
2031 * b_data buffer and updating all of the pointers. Here we use
2032 * the hash lock to ensure there are no races. If the need
2033 * arises for other types to be decrypted in-place, they must
2034 * add handling here as well.
2036 if ((flags & ARC_FILL_IN_PLACE) != 0) {
2037 ASSERT(!hdr_compressed);
2038 ASSERT(!compressed);
2041 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
2042 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2044 if (hash_lock != NULL)
2045 mutex_enter(hash_lock);
2046 arc_buf_untransform_in_place(buf, hash_lock);
2047 if (hash_lock != NULL)
2048 mutex_exit(hash_lock);
2050 /* Compute the hdr's checksum if necessary */
2051 arc_cksum_compute(buf);
2057 if (hdr_compressed == compressed) {
2058 if (!arc_buf_is_shared(buf)) {
2059 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2063 ASSERT(hdr_compressed);
2064 ASSERT(!compressed);
2065 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2068 * If the buf is sharing its data with the hdr, unlink it and
2069 * allocate a new data buffer for the buf.
2071 if (arc_buf_is_shared(buf)) {
2072 ASSERT(ARC_BUF_COMPRESSED(buf));
2074 /* We need to give the buf its own b_data */
2075 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2077 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2078 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2080 /* Previously overhead was 0; just add new overhead */
2081 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2082 } else if (ARC_BUF_COMPRESSED(buf)) {
2083 /* We need to reallocate the buf's b_data */
2084 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2087 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2089 /* We increased the size of b_data; update overhead */
2090 ARCSTAT_INCR(arcstat_overhead_size,
2091 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2095 * Regardless of the buf's previous compression settings, it
2096 * should not be compressed at the end of this function.
2098 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2101 * Try copying the data from another buf which already has a
2102 * decompressed version. If that's not possible, it's time to
2103 * bite the bullet and decompress the data from the hdr.
2105 if (arc_buf_try_copy_decompressed_data(buf)) {
2106 /* Skip byteswapping and checksumming (already done) */
2109 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2110 hdr->b_l1hdr.b_pabd, buf->b_data,
2111 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr),
2115 * Absent hardware errors or software bugs, this should
2116 * be impossible, but log it anyway so we can debug it.
2120 "hdr %px, compress %d, psize %d, lsize %d",
2121 hdr, arc_hdr_get_compress(hdr),
2122 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2123 if (hash_lock != NULL)
2124 mutex_enter(hash_lock);
2125 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2126 if (hash_lock != NULL)
2127 mutex_exit(hash_lock);
2128 return (SET_ERROR(EIO));
2134 /* Byteswap the buf's data if necessary */
2135 if (bswap != DMU_BSWAP_NUMFUNCS) {
2136 ASSERT(!HDR_SHARED_DATA(hdr));
2137 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2138 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2141 /* Compute the hdr's checksum if necessary */
2142 arc_cksum_compute(buf);
2148 * If this function is being called to decrypt an encrypted buffer or verify an
2149 * authenticated one, the key must be loaded and a mapping must be made
2150 * available in the keystore via spa_keystore_create_mapping() or one of its
2154 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2158 arc_fill_flags_t flags = 0;
2161 flags |= ARC_FILL_IN_PLACE;
2163 ret = arc_buf_fill(buf, spa, zb, flags);
2164 if (ret == ECKSUM) {
2166 * Convert authentication and decryption errors to EIO
2167 * (and generate an ereport) before leaving the ARC.
2169 ret = SET_ERROR(EIO);
2170 spa_log_error(spa, zb);
2171 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2172 spa, NULL, zb, NULL, 0);
2179 * Increment the amount of evictable space in the arc_state_t's refcount.
2180 * We account for the space used by the hdr and the arc buf individually
2181 * so that we can add and remove them from the refcount individually.
2184 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2186 arc_buf_contents_t type = arc_buf_type(hdr);
2188 ASSERT(HDR_HAS_L1HDR(hdr));
2190 if (GHOST_STATE(state)) {
2191 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2192 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2193 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2194 ASSERT(!HDR_HAS_RABD(hdr));
2195 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2196 HDR_GET_LSIZE(hdr), hdr);
2200 if (hdr->b_l1hdr.b_pabd != NULL) {
2201 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2202 arc_hdr_size(hdr), hdr);
2204 if (HDR_HAS_RABD(hdr)) {
2205 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2206 HDR_GET_PSIZE(hdr), hdr);
2209 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2210 buf = buf->b_next) {
2211 if (arc_buf_is_shared(buf))
2213 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2214 arc_buf_size(buf), buf);
2219 * Decrement the amount of evictable space in the arc_state_t's refcount.
2220 * We account for the space used by the hdr and the arc buf individually
2221 * so that we can add and remove them from the refcount individually.
2224 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2226 arc_buf_contents_t type = arc_buf_type(hdr);
2228 ASSERT(HDR_HAS_L1HDR(hdr));
2230 if (GHOST_STATE(state)) {
2231 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2232 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2233 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2234 ASSERT(!HDR_HAS_RABD(hdr));
2235 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2236 HDR_GET_LSIZE(hdr), hdr);
2240 if (hdr->b_l1hdr.b_pabd != NULL) {
2241 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2242 arc_hdr_size(hdr), hdr);
2244 if (HDR_HAS_RABD(hdr)) {
2245 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2246 HDR_GET_PSIZE(hdr), hdr);
2249 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2250 buf = buf->b_next) {
2251 if (arc_buf_is_shared(buf))
2253 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2254 arc_buf_size(buf), buf);
2259 * Add a reference to this hdr indicating that someone is actively
2260 * referencing that memory. When the refcount transitions from 0 to 1,
2261 * we remove it from the respective arc_state_t list to indicate that
2262 * it is not evictable.
2265 add_reference(arc_buf_hdr_t *hdr, void *tag)
2269 ASSERT(HDR_HAS_L1HDR(hdr));
2270 if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) {
2271 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2272 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2273 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2276 state = hdr->b_l1hdr.b_state;
2278 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2279 (state != arc_anon)) {
2280 /* We don't use the L2-only state list. */
2281 if (state != arc_l2c_only) {
2282 multilist_remove(&state->arcs_list[arc_buf_type(hdr)],
2284 arc_evictable_space_decrement(hdr, state);
2286 /* remove the prefetch flag if we get a reference */
2287 if (HDR_HAS_L2HDR(hdr))
2288 l2arc_hdr_arcstats_decrement_state(hdr);
2289 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2290 if (HDR_HAS_L2HDR(hdr))
2291 l2arc_hdr_arcstats_increment_state(hdr);
2296 * Remove a reference from this hdr. When the reference transitions from
2297 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2298 * list making it eligible for eviction.
2301 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2304 arc_state_t *state = hdr->b_l1hdr.b_state;
2306 ASSERT(HDR_HAS_L1HDR(hdr));
2307 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2308 ASSERT(!GHOST_STATE(state));
2311 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2312 * check to prevent usage of the arc_l2c_only list.
2314 if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2315 (state != arc_anon)) {
2316 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
2317 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2318 arc_evictable_space_increment(hdr, state);
2324 * Returns detailed information about a specific arc buffer. When the
2325 * state_index argument is set the function will calculate the arc header
2326 * list position for its arc state. Since this requires a linear traversal
2327 * callers are strongly encourage not to do this. However, it can be helpful
2328 * for targeted analysis so the functionality is provided.
2331 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2333 arc_buf_hdr_t *hdr = ab->b_hdr;
2334 l1arc_buf_hdr_t *l1hdr = NULL;
2335 l2arc_buf_hdr_t *l2hdr = NULL;
2336 arc_state_t *state = NULL;
2338 memset(abi, 0, sizeof (arc_buf_info_t));
2343 abi->abi_flags = hdr->b_flags;
2345 if (HDR_HAS_L1HDR(hdr)) {
2346 l1hdr = &hdr->b_l1hdr;
2347 state = l1hdr->b_state;
2349 if (HDR_HAS_L2HDR(hdr))
2350 l2hdr = &hdr->b_l2hdr;
2353 abi->abi_bufcnt = l1hdr->b_bufcnt;
2354 abi->abi_access = l1hdr->b_arc_access;
2355 abi->abi_mru_hits = l1hdr->b_mru_hits;
2356 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2357 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2358 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2359 abi->abi_holds = zfs_refcount_count(&l1hdr->b_refcnt);
2363 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2364 abi->abi_l2arc_hits = l2hdr->b_hits;
2367 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2368 abi->abi_state_contents = arc_buf_type(hdr);
2369 abi->abi_size = arc_hdr_size(hdr);
2373 * Move the supplied buffer to the indicated state. The hash lock
2374 * for the buffer must be held by the caller.
2377 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2378 kmutex_t *hash_lock)
2380 arc_state_t *old_state;
2383 boolean_t update_old, update_new;
2384 arc_buf_contents_t buftype = arc_buf_type(hdr);
2387 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2388 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2389 * L1 hdr doesn't always exist when we change state to arc_anon before
2390 * destroying a header, in which case reallocating to add the L1 hdr is
2393 if (HDR_HAS_L1HDR(hdr)) {
2394 old_state = hdr->b_l1hdr.b_state;
2395 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2396 bufcnt = hdr->b_l1hdr.b_bufcnt;
2397 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
2400 old_state = arc_l2c_only;
2403 update_old = B_FALSE;
2405 update_new = update_old;
2407 ASSERT(MUTEX_HELD(hash_lock));
2408 ASSERT3P(new_state, !=, old_state);
2409 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2410 ASSERT(old_state != arc_anon || bufcnt <= 1);
2413 * If this buffer is evictable, transfer it from the
2414 * old state list to the new state list.
2417 if (old_state != arc_anon && old_state != arc_l2c_only) {
2418 ASSERT(HDR_HAS_L1HDR(hdr));
2419 multilist_remove(&old_state->arcs_list[buftype], hdr);
2421 if (GHOST_STATE(old_state)) {
2423 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2424 update_old = B_TRUE;
2426 arc_evictable_space_decrement(hdr, old_state);
2428 if (new_state != arc_anon && new_state != arc_l2c_only) {
2430 * An L1 header always exists here, since if we're
2431 * moving to some L1-cached state (i.e. not l2c_only or
2432 * anonymous), we realloc the header to add an L1hdr
2435 ASSERT(HDR_HAS_L1HDR(hdr));
2436 multilist_insert(&new_state->arcs_list[buftype], hdr);
2438 if (GHOST_STATE(new_state)) {
2440 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2441 update_new = B_TRUE;
2443 arc_evictable_space_increment(hdr, new_state);
2447 ASSERT(!HDR_EMPTY(hdr));
2448 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2449 buf_hash_remove(hdr);
2451 /* adjust state sizes (ignore arc_l2c_only) */
2453 if (update_new && new_state != arc_l2c_only) {
2454 ASSERT(HDR_HAS_L1HDR(hdr));
2455 if (GHOST_STATE(new_state)) {
2459 * When moving a header to a ghost state, we first
2460 * remove all arc buffers. Thus, we'll have a
2461 * bufcnt of zero, and no arc buffer to use for
2462 * the reference. As a result, we use the arc
2463 * header pointer for the reference.
2465 (void) zfs_refcount_add_many(&new_state->arcs_size,
2466 HDR_GET_LSIZE(hdr), hdr);
2467 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2468 ASSERT(!HDR_HAS_RABD(hdr));
2470 uint32_t buffers = 0;
2473 * Each individual buffer holds a unique reference,
2474 * thus we must remove each of these references one
2477 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2478 buf = buf->b_next) {
2479 ASSERT3U(bufcnt, !=, 0);
2483 * When the arc_buf_t is sharing the data
2484 * block with the hdr, the owner of the
2485 * reference belongs to the hdr. Only
2486 * add to the refcount if the arc_buf_t is
2489 if (arc_buf_is_shared(buf))
2492 (void) zfs_refcount_add_many(
2493 &new_state->arcs_size,
2494 arc_buf_size(buf), buf);
2496 ASSERT3U(bufcnt, ==, buffers);
2498 if (hdr->b_l1hdr.b_pabd != NULL) {
2499 (void) zfs_refcount_add_many(
2500 &new_state->arcs_size,
2501 arc_hdr_size(hdr), hdr);
2504 if (HDR_HAS_RABD(hdr)) {
2505 (void) zfs_refcount_add_many(
2506 &new_state->arcs_size,
2507 HDR_GET_PSIZE(hdr), hdr);
2512 if (update_old && old_state != arc_l2c_only) {
2513 ASSERT(HDR_HAS_L1HDR(hdr));
2514 if (GHOST_STATE(old_state)) {
2516 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2517 ASSERT(!HDR_HAS_RABD(hdr));
2520 * When moving a header off of a ghost state,
2521 * the header will not contain any arc buffers.
2522 * We use the arc header pointer for the reference
2523 * which is exactly what we did when we put the
2524 * header on the ghost state.
2527 (void) zfs_refcount_remove_many(&old_state->arcs_size,
2528 HDR_GET_LSIZE(hdr), hdr);
2530 uint32_t buffers = 0;
2533 * Each individual buffer holds a unique reference,
2534 * thus we must remove each of these references one
2537 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2538 buf = buf->b_next) {
2539 ASSERT3U(bufcnt, !=, 0);
2543 * When the arc_buf_t is sharing the data
2544 * block with the hdr, the owner of the
2545 * reference belongs to the hdr. Only
2546 * add to the refcount if the arc_buf_t is
2549 if (arc_buf_is_shared(buf))
2552 (void) zfs_refcount_remove_many(
2553 &old_state->arcs_size, arc_buf_size(buf),
2556 ASSERT3U(bufcnt, ==, buffers);
2557 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2560 if (hdr->b_l1hdr.b_pabd != NULL) {
2561 (void) zfs_refcount_remove_many(
2562 &old_state->arcs_size, arc_hdr_size(hdr),
2566 if (HDR_HAS_RABD(hdr)) {
2567 (void) zfs_refcount_remove_many(
2568 &old_state->arcs_size, HDR_GET_PSIZE(hdr),
2574 if (HDR_HAS_L1HDR(hdr)) {
2575 hdr->b_l1hdr.b_state = new_state;
2577 if (HDR_HAS_L2HDR(hdr) && new_state != arc_l2c_only) {
2578 l2arc_hdr_arcstats_decrement_state(hdr);
2579 hdr->b_l2hdr.b_arcs_state = new_state->arcs_state;
2580 l2arc_hdr_arcstats_increment_state(hdr);
2586 arc_space_consume(uint64_t space, arc_space_type_t type)
2588 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2593 case ARC_SPACE_DATA:
2594 ARCSTAT_INCR(arcstat_data_size, space);
2596 case ARC_SPACE_META:
2597 ARCSTAT_INCR(arcstat_metadata_size, space);
2599 case ARC_SPACE_BONUS:
2600 ARCSTAT_INCR(arcstat_bonus_size, space);
2602 case ARC_SPACE_DNODE:
2603 aggsum_add(&arc_sums.arcstat_dnode_size, space);
2605 case ARC_SPACE_DBUF:
2606 ARCSTAT_INCR(arcstat_dbuf_size, space);
2608 case ARC_SPACE_HDRS:
2609 ARCSTAT_INCR(arcstat_hdr_size, space);
2611 case ARC_SPACE_L2HDRS:
2612 aggsum_add(&arc_sums.arcstat_l2_hdr_size, space);
2614 case ARC_SPACE_ABD_CHUNK_WASTE:
2616 * Note: this includes space wasted by all scatter ABD's, not
2617 * just those allocated by the ARC. But the vast majority of
2618 * scatter ABD's come from the ARC, because other users are
2621 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, space);
2625 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE)
2626 aggsum_add(&arc_sums.arcstat_meta_used, space);
2628 aggsum_add(&arc_sums.arcstat_size, space);
2632 arc_space_return(uint64_t space, arc_space_type_t type)
2634 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2639 case ARC_SPACE_DATA:
2640 ARCSTAT_INCR(arcstat_data_size, -space);
2642 case ARC_SPACE_META:
2643 ARCSTAT_INCR(arcstat_metadata_size, -space);
2645 case ARC_SPACE_BONUS:
2646 ARCSTAT_INCR(arcstat_bonus_size, -space);
2648 case ARC_SPACE_DNODE:
2649 aggsum_add(&arc_sums.arcstat_dnode_size, -space);
2651 case ARC_SPACE_DBUF:
2652 ARCSTAT_INCR(arcstat_dbuf_size, -space);
2654 case ARC_SPACE_HDRS:
2655 ARCSTAT_INCR(arcstat_hdr_size, -space);
2657 case ARC_SPACE_L2HDRS:
2658 aggsum_add(&arc_sums.arcstat_l2_hdr_size, -space);
2660 case ARC_SPACE_ABD_CHUNK_WASTE:
2661 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, -space);
2665 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE) {
2666 ASSERT(aggsum_compare(&arc_sums.arcstat_meta_used,
2668 ARCSTAT_MAX(arcstat_meta_max,
2669 aggsum_upper_bound(&arc_sums.arcstat_meta_used));
2670 aggsum_add(&arc_sums.arcstat_meta_used, -space);
2673 ASSERT(aggsum_compare(&arc_sums.arcstat_size, space) >= 0);
2674 aggsum_add(&arc_sums.arcstat_size, -space);
2678 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2679 * with the hdr's b_pabd.
2682 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2685 * The criteria for sharing a hdr's data are:
2686 * 1. the buffer is not encrypted
2687 * 2. the hdr's compression matches the buf's compression
2688 * 3. the hdr doesn't need to be byteswapped
2689 * 4. the hdr isn't already being shared
2690 * 5. the buf is either compressed or it is the last buf in the hdr list
2692 * Criterion #5 maintains the invariant that shared uncompressed
2693 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2694 * might ask, "if a compressed buf is allocated first, won't that be the
2695 * last thing in the list?", but in that case it's impossible to create
2696 * a shared uncompressed buf anyway (because the hdr must be compressed
2697 * to have the compressed buf). You might also think that #3 is
2698 * sufficient to make this guarantee, however it's possible
2699 * (specifically in the rare L2ARC write race mentioned in
2700 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2701 * is shareable, but wasn't at the time of its allocation. Rather than
2702 * allow a new shared uncompressed buf to be created and then shuffle
2703 * the list around to make it the last element, this simply disallows
2704 * sharing if the new buf isn't the first to be added.
2706 ASSERT3P(buf->b_hdr, ==, hdr);
2707 boolean_t hdr_compressed =
2708 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF;
2709 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2710 return (!ARC_BUF_ENCRYPTED(buf) &&
2711 buf_compressed == hdr_compressed &&
2712 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2713 !HDR_SHARED_DATA(hdr) &&
2714 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2718 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2719 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2720 * copy was made successfully, or an error code otherwise.
2723 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2724 void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth,
2725 boolean_t fill, arc_buf_t **ret)
2728 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2730 ASSERT(HDR_HAS_L1HDR(hdr));
2731 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2732 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2733 hdr->b_type == ARC_BUFC_METADATA);
2734 ASSERT3P(ret, !=, NULL);
2735 ASSERT3P(*ret, ==, NULL);
2736 IMPLY(encrypted, compressed);
2738 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2741 buf->b_next = hdr->b_l1hdr.b_buf;
2744 add_reference(hdr, tag);
2747 * We're about to change the hdr's b_flags. We must either
2748 * hold the hash_lock or be undiscoverable.
2750 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2753 * Only honor requests for compressed bufs if the hdr is actually
2754 * compressed. This must be overridden if the buffer is encrypted since
2755 * encrypted buffers cannot be decompressed.
2758 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2759 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2760 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2761 } else if (compressed &&
2762 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2763 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2764 flags |= ARC_FILL_COMPRESSED;
2769 flags |= ARC_FILL_NOAUTH;
2773 * If the hdr's data can be shared then we share the data buffer and
2774 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2775 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2776 * buffer to store the buf's data.
2778 * There are two additional restrictions here because we're sharing
2779 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2780 * actively involved in an L2ARC write, because if this buf is used by
2781 * an arc_write() then the hdr's data buffer will be released when the
2782 * write completes, even though the L2ARC write might still be using it.
2783 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2784 * need to be ABD-aware. It must be allocated via
2785 * zio_[data_]buf_alloc(), not as a page, because we need to be able
2786 * to abd_release_ownership_of_buf(), which isn't allowed on "linear
2787 * page" buffers because the ABD code needs to handle freeing them
2790 boolean_t can_share = arc_can_share(hdr, buf) &&
2791 !HDR_L2_WRITING(hdr) &&
2792 hdr->b_l1hdr.b_pabd != NULL &&
2793 abd_is_linear(hdr->b_l1hdr.b_pabd) &&
2794 !abd_is_linear_page(hdr->b_l1hdr.b_pabd);
2796 /* Set up b_data and sharing */
2798 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2799 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2800 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2803 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2804 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2806 VERIFY3P(buf->b_data, !=, NULL);
2808 hdr->b_l1hdr.b_buf = buf;
2809 hdr->b_l1hdr.b_bufcnt += 1;
2811 hdr->b_crypt_hdr.b_ebufcnt += 1;
2814 * If the user wants the data from the hdr, we need to either copy or
2815 * decompress the data.
2818 ASSERT3P(zb, !=, NULL);
2819 return (arc_buf_fill(buf, spa, zb, flags));
2825 static char *arc_onloan_tag = "onloan";
2828 arc_loaned_bytes_update(int64_t delta)
2830 atomic_add_64(&arc_loaned_bytes, delta);
2832 /* assert that it did not wrap around */
2833 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2837 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2838 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2839 * buffers must be returned to the arc before they can be used by the DMU or
2843 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2845 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2846 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2848 arc_loaned_bytes_update(arc_buf_size(buf));
2854 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2855 enum zio_compress compression_type, uint8_t complevel)
2857 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2858 psize, lsize, compression_type, complevel);
2860 arc_loaned_bytes_update(arc_buf_size(buf));
2866 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2867 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2868 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2869 enum zio_compress compression_type, uint8_t complevel)
2871 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2872 byteorder, salt, iv, mac, ot, psize, lsize, compression_type,
2875 atomic_add_64(&arc_loaned_bytes, psize);
2881 * Return a loaned arc buffer to the arc.
2884 arc_return_buf(arc_buf_t *buf, void *tag)
2886 arc_buf_hdr_t *hdr = buf->b_hdr;
2888 ASSERT3P(buf->b_data, !=, NULL);
2889 ASSERT(HDR_HAS_L1HDR(hdr));
2890 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2891 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2893 arc_loaned_bytes_update(-arc_buf_size(buf));
2896 /* Detach an arc_buf from a dbuf (tag) */
2898 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2900 arc_buf_hdr_t *hdr = buf->b_hdr;
2902 ASSERT3P(buf->b_data, !=, NULL);
2903 ASSERT(HDR_HAS_L1HDR(hdr));
2904 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2905 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2907 arc_loaned_bytes_update(arc_buf_size(buf));
2911 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2913 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2916 df->l2df_size = size;
2917 df->l2df_type = type;
2918 mutex_enter(&l2arc_free_on_write_mtx);
2919 list_insert_head(l2arc_free_on_write, df);
2920 mutex_exit(&l2arc_free_on_write_mtx);
2924 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
2926 arc_state_t *state = hdr->b_l1hdr.b_state;
2927 arc_buf_contents_t type = arc_buf_type(hdr);
2928 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
2930 /* protected by hash lock, if in the hash table */
2931 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2932 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2933 ASSERT(state != arc_anon && state != arc_l2c_only);
2935 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2938 (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr);
2939 if (type == ARC_BUFC_METADATA) {
2940 arc_space_return(size, ARC_SPACE_META);
2942 ASSERT(type == ARC_BUFC_DATA);
2943 arc_space_return(size, ARC_SPACE_DATA);
2947 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
2949 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2954 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2955 * data buffer, we transfer the refcount ownership to the hdr and update
2956 * the appropriate kstats.
2959 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2961 ASSERT(arc_can_share(hdr, buf));
2962 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2963 ASSERT(!ARC_BUF_ENCRYPTED(buf));
2964 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2967 * Start sharing the data buffer. We transfer the
2968 * refcount ownership to the hdr since it always owns
2969 * the refcount whenever an arc_buf_t is shared.
2971 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
2972 arc_hdr_size(hdr), buf, hdr);
2973 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2974 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2975 HDR_ISTYPE_METADATA(hdr));
2976 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2977 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2980 * Since we've transferred ownership to the hdr we need
2981 * to increment its compressed and uncompressed kstats and
2982 * decrement the overhead size.
2984 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2985 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2986 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2990 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2992 ASSERT(arc_buf_is_shared(buf));
2993 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2994 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2997 * We are no longer sharing this buffer so we need
2998 * to transfer its ownership to the rightful owner.
3000 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
3001 arc_hdr_size(hdr), hdr, buf);
3002 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3003 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3004 abd_free(hdr->b_l1hdr.b_pabd);
3005 hdr->b_l1hdr.b_pabd = NULL;
3006 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3009 * Since the buffer is no longer shared between
3010 * the arc buf and the hdr, count it as overhead.
3012 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3013 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3014 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3018 * Remove an arc_buf_t from the hdr's buf list and return the last
3019 * arc_buf_t on the list. If no buffers remain on the list then return
3023 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3025 ASSERT(HDR_HAS_L1HDR(hdr));
3026 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3028 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3029 arc_buf_t *lastbuf = NULL;
3032 * Remove the buf from the hdr list and locate the last
3033 * remaining buffer on the list.
3035 while (*bufp != NULL) {
3037 *bufp = buf->b_next;
3040 * If we've removed a buffer in the middle of
3041 * the list then update the lastbuf and update
3044 if (*bufp != NULL) {
3046 bufp = &(*bufp)->b_next;
3050 ASSERT3P(lastbuf, !=, buf);
3051 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3052 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3053 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3059 * Free up buf->b_data and pull the arc_buf_t off of the arc_buf_hdr_t's
3063 arc_buf_destroy_impl(arc_buf_t *buf)
3065 arc_buf_hdr_t *hdr = buf->b_hdr;
3068 * Free up the data associated with the buf but only if we're not
3069 * sharing this with the hdr. If we are sharing it with the hdr, the
3070 * hdr is responsible for doing the free.
3072 if (buf->b_data != NULL) {
3074 * We're about to change the hdr's b_flags. We must either
3075 * hold the hash_lock or be undiscoverable.
3077 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3079 arc_cksum_verify(buf);
3080 arc_buf_unwatch(buf);
3082 if (arc_buf_is_shared(buf)) {
3083 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3085 uint64_t size = arc_buf_size(buf);
3086 arc_free_data_buf(hdr, buf->b_data, size, buf);
3087 ARCSTAT_INCR(arcstat_overhead_size, -size);
3091 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3092 hdr->b_l1hdr.b_bufcnt -= 1;
3094 if (ARC_BUF_ENCRYPTED(buf)) {
3095 hdr->b_crypt_hdr.b_ebufcnt -= 1;
3098 * If we have no more encrypted buffers and we've
3099 * already gotten a copy of the decrypted data we can
3100 * free b_rabd to save some space.
3102 if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
3103 HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
3104 !HDR_IO_IN_PROGRESS(hdr)) {
3105 arc_hdr_free_abd(hdr, B_TRUE);
3110 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3112 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3114 * If the current arc_buf_t is sharing its data buffer with the
3115 * hdr, then reassign the hdr's b_pabd to share it with the new
3116 * buffer at the end of the list. The shared buffer is always
3117 * the last one on the hdr's buffer list.
3119 * There is an equivalent case for compressed bufs, but since
3120 * they aren't guaranteed to be the last buf in the list and
3121 * that is an exceedingly rare case, we just allow that space be
3122 * wasted temporarily. We must also be careful not to share
3123 * encrypted buffers, since they cannot be shared.
3125 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
3126 /* Only one buf can be shared at once */
3127 VERIFY(!arc_buf_is_shared(lastbuf));
3128 /* hdr is uncompressed so can't have compressed buf */
3129 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3131 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3132 arc_hdr_free_abd(hdr, B_FALSE);
3135 * We must setup a new shared block between the
3136 * last buffer and the hdr. The data would have
3137 * been allocated by the arc buf so we need to transfer
3138 * ownership to the hdr since it's now being shared.
3140 arc_share_buf(hdr, lastbuf);
3142 } else if (HDR_SHARED_DATA(hdr)) {
3144 * Uncompressed shared buffers are always at the end
3145 * of the list. Compressed buffers don't have the
3146 * same requirements. This makes it hard to
3147 * simply assert that the lastbuf is shared so
3148 * we rely on the hdr's compression flags to determine
3149 * if we have a compressed, shared buffer.
3151 ASSERT3P(lastbuf, !=, NULL);
3152 ASSERT(arc_buf_is_shared(lastbuf) ||
3153 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3157 * Free the checksum if we're removing the last uncompressed buf from
3160 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3161 arc_cksum_free(hdr);
3164 /* clean up the buf */
3166 kmem_cache_free(buf_cache, buf);
3170 arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, int alloc_flags)
3173 boolean_t alloc_rdata = ((alloc_flags & ARC_HDR_ALLOC_RDATA) != 0);
3175 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3176 ASSERT(HDR_HAS_L1HDR(hdr));
3177 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3178 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3181 size = HDR_GET_PSIZE(hdr);
3182 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3183 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr,
3185 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3186 ARCSTAT_INCR(arcstat_raw_size, size);
3188 size = arc_hdr_size(hdr);
3189 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3190 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr,
3192 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3195 ARCSTAT_INCR(arcstat_compressed_size, size);
3196 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3200 arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3202 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3204 ASSERT(HDR_HAS_L1HDR(hdr));
3205 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3206 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3209 * If the hdr is currently being written to the l2arc then
3210 * we defer freeing the data by adding it to the l2arc_free_on_write
3211 * list. The l2arc will free the data once it's finished
3212 * writing it to the l2arc device.
3214 if (HDR_L2_WRITING(hdr)) {
3215 arc_hdr_free_on_write(hdr, free_rdata);
3216 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3217 } else if (free_rdata) {
3218 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3220 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr);
3224 hdr->b_crypt_hdr.b_rabd = NULL;
3225 ARCSTAT_INCR(arcstat_raw_size, -size);
3227 hdr->b_l1hdr.b_pabd = NULL;
3230 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3231 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3233 ARCSTAT_INCR(arcstat_compressed_size, -size);
3234 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3238 * Allocate empty anonymous ARC header. The header will get its identity
3239 * assigned and buffers attached later as part of read or write operations.
3241 * In case of read arc_read() assigns header its identify (b_dva + b_birth),
3242 * inserts it into ARC hash to become globally visible and allocates physical
3243 * (b_pabd) or raw (b_rabd) ABD buffer to read into from disk. On disk read
3244 * completion arc_read_done() allocates ARC buffer(s) as needed, potentially
3245 * sharing one of them with the physical ABD buffer.
3247 * In case of write arc_alloc_buf() allocates ARC buffer to be filled with
3248 * data. Then after compression and/or encryption arc_write_ready() allocates
3249 * and fills (or potentially shares) physical (b_pabd) or raw (b_rabd) ABD
3250 * buffer. On disk write completion arc_write_done() assigns the header its
3251 * new identity (b_dva + b_birth) and inserts into ARC hash.
3253 * In case of partial overwrite the old data is read first as described. Then
3254 * arc_release() either allocates new anonymous ARC header and moves the ARC
3255 * buffer to it, or reuses the old ARC header by discarding its identity and
3256 * removing it from ARC hash. After buffer modification normal write process
3257 * follows as described.
3259 static arc_buf_hdr_t *
3260 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3261 boolean_t protected, enum zio_compress compression_type, uint8_t complevel,
3262 arc_buf_contents_t type)
3266 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3268 hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
3270 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3273 ASSERT(HDR_EMPTY(hdr));
3274 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3275 HDR_SET_PSIZE(hdr, psize);
3276 HDR_SET_LSIZE(hdr, lsize);
3280 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3281 arc_hdr_set_compress(hdr, compression_type);
3282 hdr->b_complevel = complevel;
3284 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3286 hdr->b_l1hdr.b_state = arc_anon;
3287 hdr->b_l1hdr.b_arc_access = 0;
3288 hdr->b_l1hdr.b_mru_hits = 0;
3289 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3290 hdr->b_l1hdr.b_mfu_hits = 0;
3291 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3292 hdr->b_l1hdr.b_bufcnt = 0;
3293 hdr->b_l1hdr.b_buf = NULL;
3295 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3301 * Transition between the two allocation states for the arc_buf_hdr struct.
3302 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3303 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3304 * version is used when a cache buffer is only in the L2ARC in order to reduce
3307 static arc_buf_hdr_t *
3308 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3310 ASSERT(HDR_HAS_L2HDR(hdr));
3312 arc_buf_hdr_t *nhdr;
3313 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3315 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3316 (old == hdr_l2only_cache && new == hdr_full_cache));
3319 * if the caller wanted a new full header and the header is to be
3320 * encrypted we will actually allocate the header from the full crypt
3321 * cache instead. The same applies to freeing from the old cache.
3323 if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
3324 new = hdr_full_crypt_cache;
3325 if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
3326 old = hdr_full_crypt_cache;
3328 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3330 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3331 buf_hash_remove(hdr);
3333 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3335 if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
3336 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3338 * arc_access and arc_change_state need to be aware that a
3339 * header has just come out of L2ARC, so we set its state to
3340 * l2c_only even though it's about to change.
3342 nhdr->b_l1hdr.b_state = arc_l2c_only;
3344 /* Verify previous threads set to NULL before freeing */
3345 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3346 ASSERT(!HDR_HAS_RABD(hdr));
3348 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3349 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3350 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3353 * If we've reached here, We must have been called from
3354 * arc_evict_hdr(), as such we should have already been
3355 * removed from any ghost list we were previously on
3356 * (which protects us from racing with arc_evict_state),
3357 * thus no locking is needed during this check.
3359 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3362 * A buffer must not be moved into the arc_l2c_only
3363 * state if it's not finished being written out to the
3364 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3365 * might try to be accessed, even though it was removed.
3367 VERIFY(!HDR_L2_WRITING(hdr));
3368 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3369 ASSERT(!HDR_HAS_RABD(hdr));
3371 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3374 * The header has been reallocated so we need to re-insert it into any
3377 (void) buf_hash_insert(nhdr, NULL);
3379 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3381 mutex_enter(&dev->l2ad_mtx);
3384 * We must place the realloc'ed header back into the list at
3385 * the same spot. Otherwise, if it's placed earlier in the list,
3386 * l2arc_write_buffers() could find it during the function's
3387 * write phase, and try to write it out to the l2arc.
3389 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3390 list_remove(&dev->l2ad_buflist, hdr);
3392 mutex_exit(&dev->l2ad_mtx);
3395 * Since we're using the pointer address as the tag when
3396 * incrementing and decrementing the l2ad_alloc refcount, we
3397 * must remove the old pointer (that we're about to destroy) and
3398 * add the new pointer to the refcount. Otherwise we'd remove
3399 * the wrong pointer address when calling arc_hdr_destroy() later.
3402 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
3403 arc_hdr_size(hdr), hdr);
3404 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
3405 arc_hdr_size(nhdr), nhdr);
3407 buf_discard_identity(hdr);
3408 kmem_cache_free(old, hdr);
3414 * This function allows an L1 header to be reallocated as a crypt
3415 * header and vice versa. If we are going to a crypt header, the
3416 * new fields will be zeroed out.
3418 static arc_buf_hdr_t *
3419 arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
3421 arc_buf_hdr_t *nhdr;
3423 kmem_cache_t *ncache, *ocache;
3426 * This function requires that hdr is in the arc_anon state.
3427 * Therefore it won't have any L2ARC data for us to worry
3430 ASSERT(HDR_HAS_L1HDR(hdr));
3431 ASSERT(!HDR_HAS_L2HDR(hdr));
3432 ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
3433 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3434 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3435 ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node));
3436 ASSERT3P(hdr->b_hash_next, ==, NULL);
3439 ncache = hdr_full_crypt_cache;
3440 ocache = hdr_full_cache;
3442 ncache = hdr_full_cache;
3443 ocache = hdr_full_crypt_cache;
3446 nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);
3449 * Copy all members that aren't locks or condvars to the new header.
3450 * No lists are pointing to us (as we asserted above), so we don't
3451 * need to worry about the list nodes.
3453 nhdr->b_dva = hdr->b_dva;
3454 nhdr->b_birth = hdr->b_birth;
3455 nhdr->b_type = hdr->b_type;
3456 nhdr->b_flags = hdr->b_flags;
3457 nhdr->b_psize = hdr->b_psize;
3458 nhdr->b_lsize = hdr->b_lsize;
3459 nhdr->b_spa = hdr->b_spa;
3460 nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
3461 nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
3462 nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
3463 nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
3464 nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
3465 nhdr->b_l1hdr.b_mru_hits = hdr->b_l1hdr.b_mru_hits;
3466 nhdr->b_l1hdr.b_mru_ghost_hits = hdr->b_l1hdr.b_mru_ghost_hits;
3467 nhdr->b_l1hdr.b_mfu_hits = hdr->b_l1hdr.b_mfu_hits;
3468 nhdr->b_l1hdr.b_mfu_ghost_hits = hdr->b_l1hdr.b_mfu_ghost_hits;
3469 nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
3470 nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
3473 * This zfs_refcount_add() exists only to ensure that the individual
3474 * arc buffers always point to a header that is referenced, avoiding
3475 * a small race condition that could trigger ASSERTs.
3477 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
3478 nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
3479 for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) {
3480 mutex_enter(&buf->b_evict_lock);
3482 mutex_exit(&buf->b_evict_lock);
3485 zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
3486 (void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
3487 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3490 arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
3492 arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
3495 /* unset all members of the original hdr */
3496 bzero(&hdr->b_dva, sizeof (dva_t));
3498 hdr->b_type = ARC_BUFC_INVALID;
3503 hdr->b_l1hdr.b_freeze_cksum = NULL;
3504 hdr->b_l1hdr.b_buf = NULL;
3505 hdr->b_l1hdr.b_bufcnt = 0;
3506 hdr->b_l1hdr.b_byteswap = 0;
3507 hdr->b_l1hdr.b_state = NULL;
3508 hdr->b_l1hdr.b_arc_access = 0;
3509 hdr->b_l1hdr.b_mru_hits = 0;
3510 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3511 hdr->b_l1hdr.b_mfu_hits = 0;
3512 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3513 hdr->b_l1hdr.b_acb = NULL;
3514 hdr->b_l1hdr.b_pabd = NULL;
3516 if (ocache == hdr_full_crypt_cache) {
3517 ASSERT(!HDR_HAS_RABD(hdr));
3518 hdr->b_crypt_hdr.b_ot = DMU_OT_NONE;
3519 hdr->b_crypt_hdr.b_ebufcnt = 0;
3520 hdr->b_crypt_hdr.b_dsobj = 0;
3521 bzero(hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3522 bzero(hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3523 bzero(hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3526 buf_discard_identity(hdr);
3527 kmem_cache_free(ocache, hdr);
3533 * This function is used by the send / receive code to convert a newly
3534 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3535 * is also used to allow the root objset block to be updated without altering
3536 * its embedded MACs. Both block types will always be uncompressed so we do not
3537 * have to worry about compression type or psize.
3540 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3541 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3544 arc_buf_hdr_t *hdr = buf->b_hdr;
3546 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3547 ASSERT(HDR_HAS_L1HDR(hdr));
3548 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3550 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3551 if (!HDR_PROTECTED(hdr))
3552 hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
3553 hdr->b_crypt_hdr.b_dsobj = dsobj;
3554 hdr->b_crypt_hdr.b_ot = ot;
3555 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3556 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3557 if (!arc_hdr_has_uncompressed_buf(hdr))
3558 arc_cksum_free(hdr);
3561 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3563 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3565 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3569 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3570 * The buf is returned thawed since we expect the consumer to modify it.
3573 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3575 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3576 B_FALSE, ZIO_COMPRESS_OFF, 0, type);
3578 arc_buf_t *buf = NULL;
3579 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3580 B_FALSE, B_FALSE, &buf));
3587 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3588 * for bufs containing metadata.
3591 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3592 enum zio_compress compression_type, uint8_t complevel)
3594 ASSERT3U(lsize, >, 0);
3595 ASSERT3U(lsize, >=, psize);
3596 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3597 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3599 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3600 B_FALSE, compression_type, complevel, ARC_BUFC_DATA);
3602 arc_buf_t *buf = NULL;
3603 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3604 B_TRUE, B_FALSE, B_FALSE, &buf));
3606 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3609 * To ensure that the hdr has the correct data in it if we call
3610 * arc_untransform() on this buf before it's been written to disk,
3611 * it's easiest if we just set up sharing between the buf and the hdr.
3613 arc_share_buf(hdr, buf);
3619 arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder,
3620 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
3621 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3622 enum zio_compress compression_type, uint8_t complevel)
3626 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3627 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3629 ASSERT3U(lsize, >, 0);
3630 ASSERT3U(lsize, >=, psize);
3631 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3632 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3634 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3635 compression_type, complevel, type);
3637 hdr->b_crypt_hdr.b_dsobj = dsobj;
3638 hdr->b_crypt_hdr.b_ot = ot;
3639 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3640 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3641 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3642 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3643 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3646 * This buffer will be considered encrypted even if the ot is not an
3647 * encrypted type. It will become authenticated instead in
3648 * arc_write_ready().
3651 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3652 B_FALSE, B_FALSE, &buf));
3654 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3660 l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
3661 boolean_t state_only)
3663 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3664 l2arc_dev_t *dev = l2hdr->b_dev;
3665 uint64_t lsize = HDR_GET_LSIZE(hdr);
3666 uint64_t psize = HDR_GET_PSIZE(hdr);
3667 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3668 arc_buf_contents_t type = hdr->b_type;
3683 /* If the buffer is a prefetch, count it as such. */
3684 if (HDR_PREFETCH(hdr)) {
3685 ARCSTAT_INCR(arcstat_l2_prefetch_asize, asize_s);
3688 * We use the value stored in the L2 header upon initial
3689 * caching in L2ARC. This value will be updated in case
3690 * an MRU/MRU_ghost buffer transitions to MFU but the L2ARC
3691 * metadata (log entry) cannot currently be updated. Having
3692 * the ARC state in the L2 header solves the problem of a
3693 * possibly absent L1 header (apparent in buffers restored
3694 * from persistent L2ARC).
3696 switch (hdr->b_l2hdr.b_arcs_state) {
3697 case ARC_STATE_MRU_GHOST:
3699 ARCSTAT_INCR(arcstat_l2_mru_asize, asize_s);
3701 case ARC_STATE_MFU_GHOST:
3703 ARCSTAT_INCR(arcstat_l2_mfu_asize, asize_s);
3713 ARCSTAT_INCR(arcstat_l2_psize, psize_s);
3714 ARCSTAT_INCR(arcstat_l2_lsize, lsize_s);
3718 ARCSTAT_INCR(arcstat_l2_bufc_data_asize, asize_s);
3720 case ARC_BUFC_METADATA:
3721 ARCSTAT_INCR(arcstat_l2_bufc_metadata_asize, asize_s);
3730 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3732 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3733 l2arc_dev_t *dev = l2hdr->b_dev;
3734 uint64_t psize = HDR_GET_PSIZE(hdr);
3735 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3737 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3738 ASSERT(HDR_HAS_L2HDR(hdr));
3740 list_remove(&dev->l2ad_buflist, hdr);
3742 l2arc_hdr_arcstats_decrement(hdr);
3743 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3745 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3747 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3751 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3753 if (HDR_HAS_L1HDR(hdr)) {
3754 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3755 hdr->b_l1hdr.b_bufcnt > 0);
3756 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3757 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3759 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3760 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3762 if (HDR_HAS_L2HDR(hdr)) {
3763 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3764 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3767 mutex_enter(&dev->l2ad_mtx);
3770 * Even though we checked this conditional above, we
3771 * need to check this again now that we have the
3772 * l2ad_mtx. This is because we could be racing with
3773 * another thread calling l2arc_evict() which might have
3774 * destroyed this header's L2 portion as we were waiting
3775 * to acquire the l2ad_mtx. If that happens, we don't
3776 * want to re-destroy the header's L2 portion.
3778 if (HDR_HAS_L2HDR(hdr))
3779 arc_hdr_l2hdr_destroy(hdr);
3782 mutex_exit(&dev->l2ad_mtx);
3786 * The header's identify can only be safely discarded once it is no
3787 * longer discoverable. This requires removing it from the hash table
3788 * and the l2arc header list. After this point the hash lock can not
3789 * be used to protect the header.
3791 if (!HDR_EMPTY(hdr))
3792 buf_discard_identity(hdr);
3794 if (HDR_HAS_L1HDR(hdr)) {
3795 arc_cksum_free(hdr);
3797 while (hdr->b_l1hdr.b_buf != NULL)
3798 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3800 if (hdr->b_l1hdr.b_pabd != NULL)
3801 arc_hdr_free_abd(hdr, B_FALSE);
3803 if (HDR_HAS_RABD(hdr))
3804 arc_hdr_free_abd(hdr, B_TRUE);
3807 ASSERT3P(hdr->b_hash_next, ==, NULL);
3808 if (HDR_HAS_L1HDR(hdr)) {
3809 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3810 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3812 if (!HDR_PROTECTED(hdr)) {
3813 kmem_cache_free(hdr_full_cache, hdr);
3815 kmem_cache_free(hdr_full_crypt_cache, hdr);
3818 kmem_cache_free(hdr_l2only_cache, hdr);
3823 arc_buf_destroy(arc_buf_t *buf, void* tag)
3825 arc_buf_hdr_t *hdr = buf->b_hdr;
3827 if (hdr->b_l1hdr.b_state == arc_anon) {
3828 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3829 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3830 VERIFY0(remove_reference(hdr, NULL, tag));
3831 arc_hdr_destroy(hdr);
3835 kmutex_t *hash_lock = HDR_LOCK(hdr);
3836 mutex_enter(hash_lock);
3838 ASSERT3P(hdr, ==, buf->b_hdr);
3839 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3840 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3841 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3842 ASSERT3P(buf->b_data, !=, NULL);
3844 (void) remove_reference(hdr, hash_lock, tag);
3845 arc_buf_destroy_impl(buf);
3846 mutex_exit(hash_lock);
3850 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3851 * state of the header is dependent on its state prior to entering this
3852 * function. The following transitions are possible:
3854 * - arc_mru -> arc_mru_ghost
3855 * - arc_mfu -> arc_mfu_ghost
3856 * - arc_mru_ghost -> arc_l2c_only
3857 * - arc_mru_ghost -> deleted
3858 * - arc_mfu_ghost -> arc_l2c_only
3859 * - arc_mfu_ghost -> deleted
3861 * Return total size of evicted data buffers for eviction progress tracking.
3862 * When evicting from ghost states return logical buffer size to make eviction
3863 * progress at the same (or at least comparable) rate as from non-ghost states.
3865 * Return *real_evicted for actual ARC size reduction to wake up threads
3866 * waiting for it. For non-ghost states it includes size of evicted data
3867 * buffers (the headers are not freed there). For ghost states it includes
3868 * only the evicted headers size.
3871 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, uint64_t *real_evicted)
3873 arc_state_t *evicted_state, *state;
3874 int64_t bytes_evicted = 0;
3875 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3876 arc_min_prescient_prefetch_ms : arc_min_prefetch_ms;
3878 ASSERT(MUTEX_HELD(hash_lock));
3879 ASSERT(HDR_HAS_L1HDR(hdr));
3882 state = hdr->b_l1hdr.b_state;
3883 if (GHOST_STATE(state)) {
3884 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3885 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3888 * l2arc_write_buffers() relies on a header's L1 portion
3889 * (i.e. its b_pabd field) during it's write phase.
3890 * Thus, we cannot push a header onto the arc_l2c_only
3891 * state (removing its L1 piece) until the header is
3892 * done being written to the l2arc.
3894 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3895 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3896 return (bytes_evicted);
3899 ARCSTAT_BUMP(arcstat_deleted);
3900 bytes_evicted += HDR_GET_LSIZE(hdr);
3902 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3904 if (HDR_HAS_L2HDR(hdr)) {
3905 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3906 ASSERT(!HDR_HAS_RABD(hdr));
3908 * This buffer is cached on the 2nd Level ARC;
3909 * don't destroy the header.
3911 arc_change_state(arc_l2c_only, hdr, hash_lock);
3913 * dropping from L1+L2 cached to L2-only,
3914 * realloc to remove the L1 header.
3916 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3918 *real_evicted += HDR_FULL_SIZE - HDR_L2ONLY_SIZE;
3920 arc_change_state(arc_anon, hdr, hash_lock);
3921 arc_hdr_destroy(hdr);
3922 *real_evicted += HDR_FULL_SIZE;
3924 return (bytes_evicted);
3927 ASSERT(state == arc_mru || state == arc_mfu);
3928 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3930 /* prefetch buffers have a minimum lifespan */
3931 if (HDR_IO_IN_PROGRESS(hdr) ||
3932 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3933 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3934 MSEC_TO_TICK(min_lifetime))) {
3935 ARCSTAT_BUMP(arcstat_evict_skip);
3936 return (bytes_evicted);
3939 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3940 while (hdr->b_l1hdr.b_buf) {
3941 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3942 if (!mutex_tryenter(&buf->b_evict_lock)) {
3943 ARCSTAT_BUMP(arcstat_mutex_miss);
3946 if (buf->b_data != NULL) {
3947 bytes_evicted += HDR_GET_LSIZE(hdr);
3948 *real_evicted += HDR_GET_LSIZE(hdr);
3950 mutex_exit(&buf->b_evict_lock);
3951 arc_buf_destroy_impl(buf);
3954 if (HDR_HAS_L2HDR(hdr)) {
3955 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3957 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3958 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3959 HDR_GET_LSIZE(hdr));
3961 switch (state->arcs_state) {
3964 arcstat_evict_l2_eligible_mru,
3965 HDR_GET_LSIZE(hdr));
3969 arcstat_evict_l2_eligible_mfu,
3970 HDR_GET_LSIZE(hdr));
3976 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3977 HDR_GET_LSIZE(hdr));
3981 if (hdr->b_l1hdr.b_bufcnt == 0) {
3982 arc_cksum_free(hdr);
3984 bytes_evicted += arc_hdr_size(hdr);
3985 *real_evicted += arc_hdr_size(hdr);
3988 * If this hdr is being evicted and has a compressed
3989 * buffer then we discard it here before we change states.
3990 * This ensures that the accounting is updated correctly
3991 * in arc_free_data_impl().
3993 if (hdr->b_l1hdr.b_pabd != NULL)
3994 arc_hdr_free_abd(hdr, B_FALSE);
3996 if (HDR_HAS_RABD(hdr))
3997 arc_hdr_free_abd(hdr, B_TRUE);
3999 arc_change_state(evicted_state, hdr, hash_lock);
4000 ASSERT(HDR_IN_HASH_TABLE(hdr));
4001 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
4002 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
4005 return (bytes_evicted);
4009 arc_set_need_free(void)
4011 ASSERT(MUTEX_HELD(&arc_evict_lock));
4012 int64_t remaining = arc_free_memory() - arc_sys_free / 2;
4013 arc_evict_waiter_t *aw = list_tail(&arc_evict_waiters);
4015 arc_need_free = MAX(-remaining, 0);
4018 MAX(-remaining, (int64_t)(aw->aew_count - arc_evict_count));
4023 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
4024 uint64_t spa, uint64_t bytes)
4026 multilist_sublist_t *mls;
4027 uint64_t bytes_evicted = 0, real_evicted = 0;
4029 kmutex_t *hash_lock;
4030 int evict_count = zfs_arc_evict_batch_limit;
4032 ASSERT3P(marker, !=, NULL);
4034 mls = multilist_sublist_lock(ml, idx);
4036 for (hdr = multilist_sublist_prev(mls, marker); likely(hdr != NULL);
4037 hdr = multilist_sublist_prev(mls, marker)) {
4038 if ((evict_count <= 0) || (bytes_evicted >= bytes))
4042 * To keep our iteration location, move the marker
4043 * forward. Since we're not holding hdr's hash lock, we
4044 * must be very careful and not remove 'hdr' from the
4045 * sublist. Otherwise, other consumers might mistake the
4046 * 'hdr' as not being on a sublist when they call the
4047 * multilist_link_active() function (they all rely on
4048 * the hash lock protecting concurrent insertions and
4049 * removals). multilist_sublist_move_forward() was
4050 * specifically implemented to ensure this is the case
4051 * (only 'marker' will be removed and re-inserted).
4053 multilist_sublist_move_forward(mls, marker);
4056 * The only case where the b_spa field should ever be
4057 * zero, is the marker headers inserted by
4058 * arc_evict_state(). It's possible for multiple threads
4059 * to be calling arc_evict_state() concurrently (e.g.
4060 * dsl_pool_close() and zio_inject_fault()), so we must
4061 * skip any markers we see from these other threads.
4063 if (hdr->b_spa == 0)
4066 /* we're only interested in evicting buffers of a certain spa */
4067 if (spa != 0 && hdr->b_spa != spa) {
4068 ARCSTAT_BUMP(arcstat_evict_skip);
4072 hash_lock = HDR_LOCK(hdr);
4075 * We aren't calling this function from any code path
4076 * that would already be holding a hash lock, so we're
4077 * asserting on this assumption to be defensive in case
4078 * this ever changes. Without this check, it would be
4079 * possible to incorrectly increment arcstat_mutex_miss
4080 * below (e.g. if the code changed such that we called
4081 * this function with a hash lock held).
4083 ASSERT(!MUTEX_HELD(hash_lock));
4085 if (mutex_tryenter(hash_lock)) {
4087 uint64_t evicted = arc_evict_hdr(hdr, hash_lock,
4089 mutex_exit(hash_lock);
4091 bytes_evicted += evicted;
4092 real_evicted += revicted;
4095 * If evicted is zero, arc_evict_hdr() must have
4096 * decided to skip this header, don't increment
4097 * evict_count in this case.
4103 ARCSTAT_BUMP(arcstat_mutex_miss);
4107 multilist_sublist_unlock(mls);
4110 * Increment the count of evicted bytes, and wake up any threads that
4111 * are waiting for the count to reach this value. Since the list is
4112 * ordered by ascending aew_count, we pop off the beginning of the
4113 * list until we reach the end, or a waiter that's past the current
4114 * "count". Doing this outside the loop reduces the number of times
4115 * we need to acquire the global arc_evict_lock.
4117 * Only wake when there's sufficient free memory in the system
4118 * (specifically, arc_sys_free/2, which by default is a bit more than
4119 * 1/64th of RAM). See the comments in arc_wait_for_eviction().
4121 mutex_enter(&arc_evict_lock);
4122 arc_evict_count += real_evicted;
4124 if (arc_free_memory() > arc_sys_free / 2) {
4125 arc_evict_waiter_t *aw;
4126 while ((aw = list_head(&arc_evict_waiters)) != NULL &&
4127 aw->aew_count <= arc_evict_count) {
4128 list_remove(&arc_evict_waiters, aw);
4129 cv_broadcast(&aw->aew_cv);
4132 arc_set_need_free();
4133 mutex_exit(&arc_evict_lock);
4136 * If the ARC size is reduced from arc_c_max to arc_c_min (especially
4137 * if the average cached block is small), eviction can be on-CPU for
4138 * many seconds. To ensure that other threads that may be bound to
4139 * this CPU are able to make progress, make a voluntary preemption
4144 return (bytes_evicted);
4148 * Evict buffers from the given arc state, until we've removed the
4149 * specified number of bytes. Move the removed buffers to the
4150 * appropriate evict state.
4152 * This function makes a "best effort". It skips over any buffers
4153 * it can't get a hash_lock on, and so, may not catch all candidates.
4154 * It may also return without evicting as much space as requested.
4156 * If bytes is specified using the special value ARC_EVICT_ALL, this
4157 * will evict all available (i.e. unlocked and evictable) buffers from
4158 * the given arc state; which is used by arc_flush().
4161 arc_evict_state(arc_state_t *state, uint64_t spa, uint64_t bytes,
4162 arc_buf_contents_t type)
4164 uint64_t total_evicted = 0;
4165 multilist_t *ml = &state->arcs_list[type];
4167 arc_buf_hdr_t **markers;
4169 num_sublists = multilist_get_num_sublists(ml);
4172 * If we've tried to evict from each sublist, made some
4173 * progress, but still have not hit the target number of bytes
4174 * to evict, we want to keep trying. The markers allow us to
4175 * pick up where we left off for each individual sublist, rather
4176 * than starting from the tail each time.
4178 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
4179 for (int i = 0; i < num_sublists; i++) {
4180 multilist_sublist_t *mls;
4182 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
4185 * A b_spa of 0 is used to indicate that this header is
4186 * a marker. This fact is used in arc_evict_type() and
4187 * arc_evict_state_impl().
4189 markers[i]->b_spa = 0;
4191 mls = multilist_sublist_lock(ml, i);
4192 multilist_sublist_insert_tail(mls, markers[i]);
4193 multilist_sublist_unlock(mls);
4197 * While we haven't hit our target number of bytes to evict, or
4198 * we're evicting all available buffers.
4200 while (total_evicted < bytes) {
4201 int sublist_idx = multilist_get_random_index(ml);
4202 uint64_t scan_evicted = 0;
4205 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
4206 * Request that 10% of the LRUs be scanned by the superblock
4209 if (type == ARC_BUFC_DATA && aggsum_compare(
4210 &arc_sums.arcstat_dnode_size, arc_dnode_size_limit) > 0) {
4211 arc_prune_async((aggsum_upper_bound(
4212 &arc_sums.arcstat_dnode_size) -
4213 arc_dnode_size_limit) / sizeof (dnode_t) /
4214 zfs_arc_dnode_reduce_percent);
4218 * Start eviction using a randomly selected sublist,
4219 * this is to try and evenly balance eviction across all
4220 * sublists. Always starting at the same sublist
4221 * (e.g. index 0) would cause evictions to favor certain
4222 * sublists over others.
4224 for (int i = 0; i < num_sublists; i++) {
4225 uint64_t bytes_remaining;
4226 uint64_t bytes_evicted;
4228 if (total_evicted < bytes)
4229 bytes_remaining = bytes - total_evicted;
4233 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4234 markers[sublist_idx], spa, bytes_remaining);
4236 scan_evicted += bytes_evicted;
4237 total_evicted += bytes_evicted;
4239 /* we've reached the end, wrap to the beginning */
4240 if (++sublist_idx >= num_sublists)
4245 * If we didn't evict anything during this scan, we have
4246 * no reason to believe we'll evict more during another
4247 * scan, so break the loop.
4249 if (scan_evicted == 0) {
4250 /* This isn't possible, let's make that obvious */
4251 ASSERT3S(bytes, !=, 0);
4254 * When bytes is ARC_EVICT_ALL, the only way to
4255 * break the loop is when scan_evicted is zero.
4256 * In that case, we actually have evicted enough,
4257 * so we don't want to increment the kstat.
4259 if (bytes != ARC_EVICT_ALL) {
4260 ASSERT3S(total_evicted, <, bytes);
4261 ARCSTAT_BUMP(arcstat_evict_not_enough);
4268 for (int i = 0; i < num_sublists; i++) {
4269 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4270 multilist_sublist_remove(mls, markers[i]);
4271 multilist_sublist_unlock(mls);
4273 kmem_cache_free(hdr_full_cache, markers[i]);
4275 kmem_free(markers, sizeof (*markers) * num_sublists);
4277 return (total_evicted);
4281 * Flush all "evictable" data of the given type from the arc state
4282 * specified. This will not evict any "active" buffers (i.e. referenced).
4284 * When 'retry' is set to B_FALSE, the function will make a single pass
4285 * over the state and evict any buffers that it can. Since it doesn't
4286 * continually retry the eviction, it might end up leaving some buffers
4287 * in the ARC due to lock misses.
4289 * When 'retry' is set to B_TRUE, the function will continually retry the
4290 * eviction until *all* evictable buffers have been removed from the
4291 * state. As a result, if concurrent insertions into the state are
4292 * allowed (e.g. if the ARC isn't shutting down), this function might
4293 * wind up in an infinite loop, continually trying to evict buffers.
4296 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4299 uint64_t evicted = 0;
4301 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4302 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4312 * Evict the specified number of bytes from the state specified,
4313 * restricting eviction to the spa and type given. This function
4314 * prevents us from trying to evict more from a state's list than
4315 * is "evictable", and to skip evicting altogether when passed a
4316 * negative value for "bytes". In contrast, arc_evict_state() will
4317 * evict everything it can, when passed a negative value for "bytes".
4320 arc_evict_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4321 arc_buf_contents_t type)
4325 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4326 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4328 return (arc_evict_state(state, spa, delta, type));
4335 * The goal of this function is to evict enough meta data buffers from the
4336 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
4337 * more complicated than it appears because it is common for data buffers
4338 * to have holds on meta data buffers. In addition, dnode meta data buffers
4339 * will be held by the dnodes in the block preventing them from being freed.
4340 * This means we can't simply traverse the ARC and expect to always find
4341 * enough unheld meta data buffer to release.
4343 * Therefore, this function has been updated to make alternating passes
4344 * over the ARC releasing data buffers and then newly unheld meta data
4345 * buffers. This ensures forward progress is maintained and meta_used
4346 * will decrease. Normally this is sufficient, but if required the ARC
4347 * will call the registered prune callbacks causing dentry and inodes to
4348 * be dropped from the VFS cache. This will make dnode meta data buffers
4349 * available for reclaim.
4352 arc_evict_meta_balanced(uint64_t meta_used)
4354 int64_t delta, prune = 0, adjustmnt;
4355 uint64_t total_evicted = 0;
4356 arc_buf_contents_t type = ARC_BUFC_DATA;
4357 int restarts = MAX(zfs_arc_meta_adjust_restarts, 0);
4361 * This slightly differs than the way we evict from the mru in
4362 * arc_evict because we don't have a "target" value (i.e. no
4363 * "meta" arc_p). As a result, I think we can completely
4364 * cannibalize the metadata in the MRU before we evict the
4365 * metadata from the MFU. I think we probably need to implement a
4366 * "metadata arc_p" value to do this properly.
4368 adjustmnt = meta_used - arc_meta_limit;
4370 if (adjustmnt > 0 &&
4371 zfs_refcount_count(&arc_mru->arcs_esize[type]) > 0) {
4372 delta = MIN(zfs_refcount_count(&arc_mru->arcs_esize[type]),
4374 total_evicted += arc_evict_impl(arc_mru, 0, delta, type);
4379 * We can't afford to recalculate adjustmnt here. If we do,
4380 * new metadata buffers can sneak into the MRU or ANON lists,
4381 * thus penalize the MFU metadata. Although the fudge factor is
4382 * small, it has been empirically shown to be significant for
4383 * certain workloads (e.g. creating many empty directories). As
4384 * such, we use the original calculation for adjustmnt, and
4385 * simply decrement the amount of data evicted from the MRU.
4388 if (adjustmnt > 0 &&
4389 zfs_refcount_count(&arc_mfu->arcs_esize[type]) > 0) {
4390 delta = MIN(zfs_refcount_count(&arc_mfu->arcs_esize[type]),
4392 total_evicted += arc_evict_impl(arc_mfu, 0, delta, type);
4395 adjustmnt = meta_used - arc_meta_limit;
4397 if (adjustmnt > 0 &&
4398 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) {
4399 delta = MIN(adjustmnt,
4400 zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]));
4401 total_evicted += arc_evict_impl(arc_mru_ghost, 0, delta, type);
4405 if (adjustmnt > 0 &&
4406 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) {
4407 delta = MIN(adjustmnt,
4408 zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]));
4409 total_evicted += arc_evict_impl(arc_mfu_ghost, 0, delta, type);
4413 * If after attempting to make the requested adjustment to the ARC
4414 * the meta limit is still being exceeded then request that the
4415 * higher layers drop some cached objects which have holds on ARC
4416 * meta buffers. Requests to the upper layers will be made with
4417 * increasingly large scan sizes until the ARC is below the limit.
4419 if (meta_used > arc_meta_limit) {
4420 if (type == ARC_BUFC_DATA) {
4421 type = ARC_BUFC_METADATA;
4423 type = ARC_BUFC_DATA;
4425 if (zfs_arc_meta_prune) {
4426 prune += zfs_arc_meta_prune;
4427 arc_prune_async(prune);
4436 return (total_evicted);
4440 * Evict metadata buffers from the cache, such that arcstat_meta_used is
4441 * capped by the arc_meta_limit tunable.
4444 arc_evict_meta_only(uint64_t meta_used)
4446 uint64_t total_evicted = 0;
4450 * If we're over the meta limit, we want to evict enough
4451 * metadata to get back under the meta limit. We don't want to
4452 * evict so much that we drop the MRU below arc_p, though. If
4453 * we're over the meta limit more than we're over arc_p, we
4454 * evict some from the MRU here, and some from the MFU below.
4456 target = MIN((int64_t)(meta_used - arc_meta_limit),
4457 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4458 zfs_refcount_count(&arc_mru->arcs_size) - arc_p));
4460 total_evicted += arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4463 * Similar to the above, we want to evict enough bytes to get us
4464 * below the meta limit, but not so much as to drop us below the
4465 * space allotted to the MFU (which is defined as arc_c - arc_p).
4467 target = MIN((int64_t)(meta_used - arc_meta_limit),
4468 (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) -
4471 total_evicted += arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4473 return (total_evicted);
4477 arc_evict_meta(uint64_t meta_used)
4479 if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY)
4480 return (arc_evict_meta_only(meta_used));
4482 return (arc_evict_meta_balanced(meta_used));
4486 * Return the type of the oldest buffer in the given arc state
4488 * This function will select a random sublist of type ARC_BUFC_DATA and
4489 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4490 * is compared, and the type which contains the "older" buffer will be
4493 static arc_buf_contents_t
4494 arc_evict_type(arc_state_t *state)
4496 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
4497 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
4498 int data_idx = multilist_get_random_index(data_ml);
4499 int meta_idx = multilist_get_random_index(meta_ml);
4500 multilist_sublist_t *data_mls;
4501 multilist_sublist_t *meta_mls;
4502 arc_buf_contents_t type;
4503 arc_buf_hdr_t *data_hdr;
4504 arc_buf_hdr_t *meta_hdr;
4507 * We keep the sublist lock until we're finished, to prevent
4508 * the headers from being destroyed via arc_evict_state().
4510 data_mls = multilist_sublist_lock(data_ml, data_idx);
4511 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4514 * These two loops are to ensure we skip any markers that
4515 * might be at the tail of the lists due to arc_evict_state().
4518 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4519 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4520 if (data_hdr->b_spa != 0)
4524 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4525 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4526 if (meta_hdr->b_spa != 0)
4530 if (data_hdr == NULL && meta_hdr == NULL) {
4531 type = ARC_BUFC_DATA;
4532 } else if (data_hdr == NULL) {
4533 ASSERT3P(meta_hdr, !=, NULL);
4534 type = ARC_BUFC_METADATA;
4535 } else if (meta_hdr == NULL) {
4536 ASSERT3P(data_hdr, !=, NULL);
4537 type = ARC_BUFC_DATA;
4539 ASSERT3P(data_hdr, !=, NULL);
4540 ASSERT3P(meta_hdr, !=, NULL);
4542 /* The headers can't be on the sublist without an L1 header */
4543 ASSERT(HDR_HAS_L1HDR(data_hdr));
4544 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4546 if (data_hdr->b_l1hdr.b_arc_access <
4547 meta_hdr->b_l1hdr.b_arc_access) {
4548 type = ARC_BUFC_DATA;
4550 type = ARC_BUFC_METADATA;
4554 multilist_sublist_unlock(meta_mls);
4555 multilist_sublist_unlock(data_mls);
4561 * Evict buffers from the cache, such that arcstat_size is capped by arc_c.
4566 uint64_t total_evicted = 0;
4569 uint64_t asize = aggsum_value(&arc_sums.arcstat_size);
4570 uint64_t ameta = aggsum_value(&arc_sums.arcstat_meta_used);
4573 * If we're over arc_meta_limit, we want to correct that before
4574 * potentially evicting data buffers below.
4576 total_evicted += arc_evict_meta(ameta);
4581 * If we're over the target cache size, we want to evict enough
4582 * from the list to get back to our target size. We don't want
4583 * to evict too much from the MRU, such that it drops below
4584 * arc_p. So, if we're over our target cache size more than
4585 * the MRU is over arc_p, we'll evict enough to get back to
4586 * arc_p here, and then evict more from the MFU below.
4588 target = MIN((int64_t)(asize - arc_c),
4589 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4590 zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4593 * If we're below arc_meta_min, always prefer to evict data.
4594 * Otherwise, try to satisfy the requested number of bytes to
4595 * evict from the type which contains older buffers; in an
4596 * effort to keep newer buffers in the cache regardless of their
4597 * type. If we cannot satisfy the number of bytes from this
4598 * type, spill over into the next type.
4600 if (arc_evict_type(arc_mru) == ARC_BUFC_METADATA &&
4601 ameta > arc_meta_min) {
4602 bytes = arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4603 total_evicted += bytes;
4606 * If we couldn't evict our target number of bytes from
4607 * metadata, we try to get the rest from data.
4612 arc_evict_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4614 bytes = arc_evict_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4615 total_evicted += bytes;
4618 * If we couldn't evict our target number of bytes from
4619 * data, we try to get the rest from metadata.
4624 arc_evict_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4628 * Re-sum ARC stats after the first round of evictions.
4630 asize = aggsum_value(&arc_sums.arcstat_size);
4631 ameta = aggsum_value(&arc_sums.arcstat_meta_used);
4637 * Now that we've tried to evict enough from the MRU to get its
4638 * size back to arc_p, if we're still above the target cache
4639 * size, we evict the rest from the MFU.
4641 target = asize - arc_c;
4643 if (arc_evict_type(arc_mfu) == ARC_BUFC_METADATA &&
4644 ameta > arc_meta_min) {
4645 bytes = arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4646 total_evicted += bytes;
4649 * If we couldn't evict our target number of bytes from
4650 * metadata, we try to get the rest from data.
4655 arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4657 bytes = arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4658 total_evicted += bytes;
4661 * If we couldn't evict our target number of bytes from
4662 * data, we try to get the rest from data.
4667 arc_evict_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4671 * Adjust ghost lists
4673 * In addition to the above, the ARC also defines target values
4674 * for the ghost lists. The sum of the mru list and mru ghost
4675 * list should never exceed the target size of the cache, and
4676 * the sum of the mru list, mfu list, mru ghost list, and mfu
4677 * ghost list should never exceed twice the target size of the
4678 * cache. The following logic enforces these limits on the ghost
4679 * caches, and evicts from them as needed.
4681 target = zfs_refcount_count(&arc_mru->arcs_size) +
4682 zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4684 bytes = arc_evict_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4685 total_evicted += bytes;
4690 arc_evict_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4693 * We assume the sum of the mru list and mfu list is less than
4694 * or equal to arc_c (we enforced this above), which means we
4695 * can use the simpler of the two equations below:
4697 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4698 * mru ghost + mfu ghost <= arc_c
4700 target = zfs_refcount_count(&arc_mru_ghost->arcs_size) +
4701 zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4703 bytes = arc_evict_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4704 total_evicted += bytes;
4709 arc_evict_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4711 return (total_evicted);
4715 arc_flush(spa_t *spa, boolean_t retry)
4720 * If retry is B_TRUE, a spa must not be specified since we have
4721 * no good way to determine if all of a spa's buffers have been
4722 * evicted from an arc state.
4724 ASSERT(!retry || spa == 0);
4727 guid = spa_load_guid(spa);
4729 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4730 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4732 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4733 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4735 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4736 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4738 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4739 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4743 arc_reduce_target_size(int64_t to_free)
4745 uint64_t asize = aggsum_value(&arc_sums.arcstat_size);
4748 * All callers want the ARC to actually evict (at least) this much
4749 * memory. Therefore we reduce from the lower of the current size and
4750 * the target size. This way, even if arc_c is much higher than
4751 * arc_size (as can be the case after many calls to arc_freed(), we will
4752 * immediately have arc_c < arc_size and therefore the arc_evict_zthr
4755 uint64_t c = MIN(arc_c, asize);
4757 if (c > to_free && c - to_free > arc_c_min) {
4758 arc_c = c - to_free;
4759 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4761 arc_p = (arc_c >> 1);
4762 ASSERT(arc_c >= arc_c_min);
4763 ASSERT((int64_t)arc_p >= 0);
4768 if (asize > arc_c) {
4769 /* See comment in arc_evict_cb_check() on why lock+flag */
4770 mutex_enter(&arc_evict_lock);
4771 arc_evict_needed = B_TRUE;
4772 mutex_exit(&arc_evict_lock);
4773 zthr_wakeup(arc_evict_zthr);
4778 * Determine if the system is under memory pressure and is asking
4779 * to reclaim memory. A return value of B_TRUE indicates that the system
4780 * is under memory pressure and that the arc should adjust accordingly.
4783 arc_reclaim_needed(void)
4785 return (arc_available_memory() < 0);
4789 arc_kmem_reap_soon(void)
4792 kmem_cache_t *prev_cache = NULL;
4793 kmem_cache_t *prev_data_cache = NULL;
4794 extern kmem_cache_t *zio_buf_cache[];
4795 extern kmem_cache_t *zio_data_buf_cache[];
4798 if ((aggsum_compare(&arc_sums.arcstat_meta_used,
4799 arc_meta_limit) >= 0) && zfs_arc_meta_prune) {
4801 * We are exceeding our meta-data cache limit.
4802 * Prune some entries to release holds on meta-data.
4804 arc_prune_async(zfs_arc_meta_prune);
4808 * Reclaim unused memory from all kmem caches.
4814 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4816 /* reach upper limit of cache size on 32-bit */
4817 if (zio_buf_cache[i] == NULL)
4820 if (zio_buf_cache[i] != prev_cache) {
4821 prev_cache = zio_buf_cache[i];
4822 kmem_cache_reap_now(zio_buf_cache[i]);
4824 if (zio_data_buf_cache[i] != prev_data_cache) {
4825 prev_data_cache = zio_data_buf_cache[i];
4826 kmem_cache_reap_now(zio_data_buf_cache[i]);
4829 kmem_cache_reap_now(buf_cache);
4830 kmem_cache_reap_now(hdr_full_cache);
4831 kmem_cache_reap_now(hdr_l2only_cache);
4832 kmem_cache_reap_now(zfs_btree_leaf_cache);
4833 abd_cache_reap_now();
4838 arc_evict_cb_check(void *arg, zthr_t *zthr)
4842 * This is necessary in order to keep the kstat information
4843 * up to date for tools that display kstat data such as the
4844 * mdb ::arc dcmd and the Linux crash utility. These tools
4845 * typically do not call kstat's update function, but simply
4846 * dump out stats from the most recent update. Without
4847 * this call, these commands may show stale stats for the
4848 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4849 * with this call, the data might be out of date if the
4850 * evict thread hasn't been woken recently; but that should
4851 * suffice. The arc_state_t structures can be queried
4852 * directly if more accurate information is needed.
4854 if (arc_ksp != NULL)
4855 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4859 * We have to rely on arc_wait_for_eviction() to tell us when to
4860 * evict, rather than checking if we are overflowing here, so that we
4861 * are sure to not leave arc_wait_for_eviction() waiting on aew_cv.
4862 * If we have become "not overflowing" since arc_wait_for_eviction()
4863 * checked, we need to wake it up. We could broadcast the CV here,
4864 * but arc_wait_for_eviction() may have not yet gone to sleep. We
4865 * would need to use a mutex to ensure that this function doesn't
4866 * broadcast until arc_wait_for_eviction() has gone to sleep (e.g.
4867 * the arc_evict_lock). However, the lock ordering of such a lock
4868 * would necessarily be incorrect with respect to the zthr_lock,
4869 * which is held before this function is called, and is held by
4870 * arc_wait_for_eviction() when it calls zthr_wakeup().
4872 return (arc_evict_needed);
4876 * Keep arc_size under arc_c by running arc_evict which evicts data
4881 arc_evict_cb(void *arg, zthr_t *zthr)
4883 uint64_t evicted = 0;
4884 fstrans_cookie_t cookie = spl_fstrans_mark();
4886 /* Evict from cache */
4887 evicted = arc_evict();
4890 * If evicted is zero, we couldn't evict anything
4891 * via arc_evict(). This could be due to hash lock
4892 * collisions, but more likely due to the majority of
4893 * arc buffers being unevictable. Therefore, even if
4894 * arc_size is above arc_c, another pass is unlikely to
4895 * be helpful and could potentially cause us to enter an
4896 * infinite loop. Additionally, zthr_iscancelled() is
4897 * checked here so that if the arc is shutting down, the
4898 * broadcast will wake any remaining arc evict waiters.
4900 mutex_enter(&arc_evict_lock);
4901 arc_evict_needed = !zthr_iscancelled(arc_evict_zthr) &&
4902 evicted > 0 && aggsum_compare(&arc_sums.arcstat_size, arc_c) > 0;
4903 if (!arc_evict_needed) {
4905 * We're either no longer overflowing, or we
4906 * can't evict anything more, so we should wake
4907 * arc_get_data_impl() sooner.
4909 arc_evict_waiter_t *aw;
4910 while ((aw = list_remove_head(&arc_evict_waiters)) != NULL) {
4911 cv_broadcast(&aw->aew_cv);
4913 arc_set_need_free();
4915 mutex_exit(&arc_evict_lock);
4916 spl_fstrans_unmark(cookie);
4921 arc_reap_cb_check(void *arg, zthr_t *zthr)
4923 int64_t free_memory = arc_available_memory();
4924 static int reap_cb_check_counter = 0;
4927 * If a kmem reap is already active, don't schedule more. We must
4928 * check for this because kmem_cache_reap_soon() won't actually
4929 * block on the cache being reaped (this is to prevent callers from
4930 * becoming implicitly blocked by a system-wide kmem reap -- which,
4931 * on a system with many, many full magazines, can take minutes).
4933 if (!kmem_cache_reap_active() && free_memory < 0) {
4935 arc_no_grow = B_TRUE;
4938 * Wait at least zfs_grow_retry (default 5) seconds
4939 * before considering growing.
4941 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4943 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4944 arc_no_grow = B_TRUE;
4945 } else if (gethrtime() >= arc_growtime) {
4946 arc_no_grow = B_FALSE;
4950 * Called unconditionally every 60 seconds to reclaim unused
4951 * zstd compression and decompression context. This is done
4952 * here to avoid the need for an independent thread.
4954 if (!((reap_cb_check_counter++) % 60))
4955 zfs_zstd_cache_reap_now();
4961 * Keep enough free memory in the system by reaping the ARC's kmem
4962 * caches. To cause more slabs to be reapable, we may reduce the
4963 * target size of the cache (arc_c), causing the arc_evict_cb()
4964 * to free more buffers.
4968 arc_reap_cb(void *arg, zthr_t *zthr)
4970 int64_t free_memory;
4971 fstrans_cookie_t cookie = spl_fstrans_mark();
4974 * Kick off asynchronous kmem_reap()'s of all our caches.
4976 arc_kmem_reap_soon();
4979 * Wait at least arc_kmem_cache_reap_retry_ms between
4980 * arc_kmem_reap_soon() calls. Without this check it is possible to
4981 * end up in a situation where we spend lots of time reaping
4982 * caches, while we're near arc_c_min. Waiting here also gives the
4983 * subsequent free memory check a chance of finding that the
4984 * asynchronous reap has already freed enough memory, and we don't
4985 * need to call arc_reduce_target_size().
4987 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
4990 * Reduce the target size as needed to maintain the amount of free
4991 * memory in the system at a fraction of the arc_size (1/128th by
4992 * default). If oversubscribed (free_memory < 0) then reduce the
4993 * target arc_size by the deficit amount plus the fractional
4994 * amount. If free memory is positive but less than the fractional
4995 * amount, reduce by what is needed to hit the fractional amount.
4997 free_memory = arc_available_memory();
5000 (arc_c >> arc_shrink_shift) - free_memory;
5002 arc_reduce_target_size(to_free);
5004 spl_fstrans_unmark(cookie);
5009 * Determine the amount of memory eligible for eviction contained in the
5010 * ARC. All clean data reported by the ghost lists can always be safely
5011 * evicted. Due to arc_c_min, the same does not hold for all clean data
5012 * contained by the regular mru and mfu lists.
5014 * In the case of the regular mru and mfu lists, we need to report as
5015 * much clean data as possible, such that evicting that same reported
5016 * data will not bring arc_size below arc_c_min. Thus, in certain
5017 * circumstances, the total amount of clean data in the mru and mfu
5018 * lists might not actually be evictable.
5020 * The following two distinct cases are accounted for:
5022 * 1. The sum of the amount of dirty data contained by both the mru and
5023 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5024 * is greater than or equal to arc_c_min.
5025 * (i.e. amount of dirty data >= arc_c_min)
5027 * This is the easy case; all clean data contained by the mru and mfu
5028 * lists is evictable. Evicting all clean data can only drop arc_size
5029 * to the amount of dirty data, which is greater than arc_c_min.
5031 * 2. The sum of the amount of dirty data contained by both the mru and
5032 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
5033 * is less than arc_c_min.
5034 * (i.e. arc_c_min > amount of dirty data)
5036 * 2.1. arc_size is greater than or equal arc_c_min.
5037 * (i.e. arc_size >= arc_c_min > amount of dirty data)
5039 * In this case, not all clean data from the regular mru and mfu
5040 * lists is actually evictable; we must leave enough clean data
5041 * to keep arc_size above arc_c_min. Thus, the maximum amount of
5042 * evictable data from the two lists combined, is exactly the
5043 * difference between arc_size and arc_c_min.
5045 * 2.2. arc_size is less than arc_c_min
5046 * (i.e. arc_c_min > arc_size > amount of dirty data)
5048 * In this case, none of the data contained in the mru and mfu
5049 * lists is evictable, even if it's clean. Since arc_size is
5050 * already below arc_c_min, evicting any more would only
5051 * increase this negative difference.
5054 #endif /* _KERNEL */
5057 * Adapt arc info given the number of bytes we are trying to add and
5058 * the state that we are coming from. This function is only called
5059 * when we are adding new content to the cache.
5062 arc_adapt(int bytes, arc_state_t *state)
5065 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
5066 int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size);
5067 int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size);
5071 * Adapt the target size of the MRU list:
5072 * - if we just hit in the MRU ghost list, then increase
5073 * the target size of the MRU list.
5074 * - if we just hit in the MFU ghost list, then increase
5075 * the target size of the MFU list by decreasing the
5076 * target size of the MRU list.
5078 if (state == arc_mru_ghost) {
5079 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
5080 if (!zfs_arc_p_dampener_disable)
5081 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
5083 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
5084 } else if (state == arc_mfu_ghost) {
5087 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
5088 if (!zfs_arc_p_dampener_disable)
5089 mult = MIN(mult, 10);
5091 delta = MIN(bytes * mult, arc_p);
5092 arc_p = MAX(arc_p_min, arc_p - delta);
5094 ASSERT((int64_t)arc_p >= 0);
5097 * Wake reap thread if we do not have any available memory
5099 if (arc_reclaim_needed()) {
5100 zthr_wakeup(arc_reap_zthr);
5107 if (arc_c >= arc_c_max)
5111 * If we're within (2 * maxblocksize) bytes of the target
5112 * cache size, increment the target cache size
5114 ASSERT3U(arc_c, >=, 2ULL << SPA_MAXBLOCKSHIFT);
5115 if (aggsum_upper_bound(&arc_sums.arcstat_size) >=
5116 arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
5117 atomic_add_64(&arc_c, (int64_t)bytes);
5118 if (arc_c > arc_c_max)
5120 else if (state == arc_anon)
5121 atomic_add_64(&arc_p, (int64_t)bytes);
5125 ASSERT((int64_t)arc_p >= 0);
5129 * Check if arc_size has grown past our upper threshold, determined by
5130 * zfs_arc_overflow_shift.
5132 static arc_ovf_level_t
5133 arc_is_overflowing(boolean_t use_reserve)
5135 /* Always allow at least one block of overflow */
5136 int64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5137 arc_c >> zfs_arc_overflow_shift);
5140 * We just compare the lower bound here for performance reasons. Our
5141 * primary goals are to make sure that the arc never grows without
5142 * bound, and that it can reach its maximum size. This check
5143 * accomplishes both goals. The maximum amount we could run over by is
5144 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
5145 * in the ARC. In practice, that's in the tens of MB, which is low
5146 * enough to be safe.
5148 int64_t over = aggsum_lower_bound(&arc_sums.arcstat_size) -
5149 arc_c - overflow / 2;
5152 return (over < 0 ? ARC_OVF_NONE :
5153 over < overflow ? ARC_OVF_SOME : ARC_OVF_SEVERE);
5157 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag,
5160 arc_buf_contents_t type = arc_buf_type(hdr);
5162 arc_get_data_impl(hdr, size, tag, alloc_flags);
5163 if (type == ARC_BUFC_METADATA) {
5164 return (abd_alloc(size, B_TRUE));
5166 ASSERT(type == ARC_BUFC_DATA);
5167 return (abd_alloc(size, B_FALSE));
5172 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5174 arc_buf_contents_t type = arc_buf_type(hdr);
5176 arc_get_data_impl(hdr, size, tag, ARC_HDR_DO_ADAPT);
5177 if (type == ARC_BUFC_METADATA) {
5178 return (zio_buf_alloc(size));
5180 ASSERT(type == ARC_BUFC_DATA);
5181 return (zio_data_buf_alloc(size));
5186 * Wait for the specified amount of data (in bytes) to be evicted from the
5187 * ARC, and for there to be sufficient free memory in the system. Waiting for
5188 * eviction ensures that the memory used by the ARC decreases. Waiting for
5189 * free memory ensures that the system won't run out of free pages, regardless
5190 * of ARC behavior and settings. See arc_lowmem_init().
5193 arc_wait_for_eviction(uint64_t amount, boolean_t use_reserve)
5195 switch (arc_is_overflowing(use_reserve)) {
5200 * This is a bit racy without taking arc_evict_lock, but the
5201 * worst that can happen is we either call zthr_wakeup() extra
5202 * time due to race with other thread here, or the set flag
5203 * get cleared by arc_evict_cb(), which is unlikely due to
5204 * big hysteresis, but also not important since at this level
5205 * of overflow the eviction is purely advisory. Same time
5206 * taking the global lock here every time without waiting for
5207 * the actual eviction creates a significant lock contention.
5209 if (!arc_evict_needed) {
5210 arc_evict_needed = B_TRUE;
5211 zthr_wakeup(arc_evict_zthr);
5214 case ARC_OVF_SEVERE:
5217 arc_evict_waiter_t aw;
5218 list_link_init(&aw.aew_node);
5219 cv_init(&aw.aew_cv, NULL, CV_DEFAULT, NULL);
5221 uint64_t last_count = 0;
5222 mutex_enter(&arc_evict_lock);
5223 if (!list_is_empty(&arc_evict_waiters)) {
5224 arc_evict_waiter_t *last =
5225 list_tail(&arc_evict_waiters);
5226 last_count = last->aew_count;
5227 } else if (!arc_evict_needed) {
5228 arc_evict_needed = B_TRUE;
5229 zthr_wakeup(arc_evict_zthr);
5232 * Note, the last waiter's count may be less than
5233 * arc_evict_count if we are low on memory in which
5234 * case arc_evict_state_impl() may have deferred
5235 * wakeups (but still incremented arc_evict_count).
5237 aw.aew_count = MAX(last_count, arc_evict_count) + amount;
5239 list_insert_tail(&arc_evict_waiters, &aw);
5241 arc_set_need_free();
5243 DTRACE_PROBE3(arc__wait__for__eviction,
5245 uint64_t, arc_evict_count,
5246 uint64_t, aw.aew_count);
5249 * We will be woken up either when arc_evict_count reaches
5250 * aew_count, or when the ARC is no longer overflowing and
5251 * eviction completes.
5252 * In case of "false" wakeup, we will still be on the list.
5255 cv_wait(&aw.aew_cv, &arc_evict_lock);
5256 } while (list_link_active(&aw.aew_node));
5257 mutex_exit(&arc_evict_lock);
5259 cv_destroy(&aw.aew_cv);
5265 * Allocate a block and return it to the caller. If we are hitting the
5266 * hard limit for the cache size, we must sleep, waiting for the eviction
5267 * thread to catch up. If we're past the target size but below the hard
5268 * limit, we'll only signal the reclaim thread and continue on.
5271 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag,
5274 arc_state_t *state = hdr->b_l1hdr.b_state;
5275 arc_buf_contents_t type = arc_buf_type(hdr);
5277 if (alloc_flags & ARC_HDR_DO_ADAPT)
5278 arc_adapt(size, state);
5281 * If arc_size is currently overflowing, we must be adding data
5282 * faster than we are evicting. To ensure we don't compound the
5283 * problem by adding more data and forcing arc_size to grow even
5284 * further past it's target size, we wait for the eviction thread to
5285 * make some progress. We also wait for there to be sufficient free
5286 * memory in the system, as measured by arc_free_memory().
5288 * Specifically, we wait for zfs_arc_eviction_pct percent of the
5289 * requested size to be evicted. This should be more than 100%, to
5290 * ensure that that progress is also made towards getting arc_size
5291 * under arc_c. See the comment above zfs_arc_eviction_pct.
5293 arc_wait_for_eviction(size * zfs_arc_eviction_pct / 100,
5294 alloc_flags & ARC_HDR_USE_RESERVE);
5296 VERIFY3U(hdr->b_type, ==, type);
5297 if (type == ARC_BUFC_METADATA) {
5298 arc_space_consume(size, ARC_SPACE_META);
5300 arc_space_consume(size, ARC_SPACE_DATA);
5304 * Update the state size. Note that ghost states have a
5305 * "ghost size" and so don't need to be updated.
5307 if (!GHOST_STATE(state)) {
5309 (void) zfs_refcount_add_many(&state->arcs_size, size, tag);
5312 * If this is reached via arc_read, the link is
5313 * protected by the hash lock. If reached via
5314 * arc_buf_alloc, the header should not be accessed by
5315 * any other thread. And, if reached via arc_read_done,
5316 * the hash lock will protect it if it's found in the
5317 * hash table; otherwise no other thread should be
5318 * trying to [add|remove]_reference it.
5320 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5321 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5322 (void) zfs_refcount_add_many(&state->arcs_esize[type],
5327 * If we are growing the cache, and we are adding anonymous
5328 * data, and we have outgrown arc_p, update arc_p
5330 if (aggsum_upper_bound(&arc_sums.arcstat_size) < arc_c &&
5331 hdr->b_l1hdr.b_state == arc_anon &&
5332 (zfs_refcount_count(&arc_anon->arcs_size) +
5333 zfs_refcount_count(&arc_mru->arcs_size) > arc_p))
5334 arc_p = MIN(arc_c, arc_p + size);
5339 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5341 arc_free_data_impl(hdr, size, tag);
5346 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5348 arc_buf_contents_t type = arc_buf_type(hdr);
5350 arc_free_data_impl(hdr, size, tag);
5351 if (type == ARC_BUFC_METADATA) {
5352 zio_buf_free(buf, size);
5354 ASSERT(type == ARC_BUFC_DATA);
5355 zio_data_buf_free(buf, size);
5360 * Free the arc data buffer.
5363 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5365 arc_state_t *state = hdr->b_l1hdr.b_state;
5366 arc_buf_contents_t type = arc_buf_type(hdr);
5368 /* protected by hash lock, if in the hash table */
5369 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5370 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5371 ASSERT(state != arc_anon && state != arc_l2c_only);
5373 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
5376 (void) zfs_refcount_remove_many(&state->arcs_size, size, tag);
5378 VERIFY3U(hdr->b_type, ==, type);
5379 if (type == ARC_BUFC_METADATA) {
5380 arc_space_return(size, ARC_SPACE_META);
5382 ASSERT(type == ARC_BUFC_DATA);
5383 arc_space_return(size, ARC_SPACE_DATA);
5388 * This routine is called whenever a buffer is accessed.
5389 * NOTE: the hash lock is dropped in this function.
5392 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5396 ASSERT(MUTEX_HELD(hash_lock));
5397 ASSERT(HDR_HAS_L1HDR(hdr));
5399 if (hdr->b_l1hdr.b_state == arc_anon) {
5401 * This buffer is not in the cache, and does not
5402 * appear in our "ghost" list. Add the new buffer
5406 ASSERT0(hdr->b_l1hdr.b_arc_access);
5407 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5408 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5409 arc_change_state(arc_mru, hdr, hash_lock);
5411 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5412 now = ddi_get_lbolt();
5415 * If this buffer is here because of a prefetch, then either:
5416 * - clear the flag if this is a "referencing" read
5417 * (any subsequent access will bump this into the MFU state).
5419 * - move the buffer to the head of the list if this is
5420 * another prefetch (to make it less likely to be evicted).
5422 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5423 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5424 /* link protected by hash lock */
5425 ASSERT(multilist_link_active(
5426 &hdr->b_l1hdr.b_arc_node));
5428 if (HDR_HAS_L2HDR(hdr))
5429 l2arc_hdr_arcstats_decrement_state(hdr);
5430 arc_hdr_clear_flags(hdr,
5432 ARC_FLAG_PRESCIENT_PREFETCH);
5433 hdr->b_l1hdr.b_mru_hits++;
5434 ARCSTAT_BUMP(arcstat_mru_hits);
5435 if (HDR_HAS_L2HDR(hdr))
5436 l2arc_hdr_arcstats_increment_state(hdr);
5438 hdr->b_l1hdr.b_arc_access = now;
5443 * This buffer has been "accessed" only once so far,
5444 * but it is still in the cache. Move it to the MFU
5447 if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access +
5450 * More than 125ms have passed since we
5451 * instantiated this buffer. Move it to the
5452 * most frequently used state.
5454 hdr->b_l1hdr.b_arc_access = now;
5455 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5456 arc_change_state(arc_mfu, hdr, hash_lock);
5458 hdr->b_l1hdr.b_mru_hits++;
5459 ARCSTAT_BUMP(arcstat_mru_hits);
5460 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5461 arc_state_t *new_state;
5463 * This buffer has been "accessed" recently, but
5464 * was evicted from the cache. Move it to the
5467 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5468 new_state = arc_mru;
5469 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5470 if (HDR_HAS_L2HDR(hdr))
5471 l2arc_hdr_arcstats_decrement_state(hdr);
5472 arc_hdr_clear_flags(hdr,
5474 ARC_FLAG_PRESCIENT_PREFETCH);
5475 if (HDR_HAS_L2HDR(hdr))
5476 l2arc_hdr_arcstats_increment_state(hdr);
5478 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5480 new_state = arc_mfu;
5481 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5484 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5485 arc_change_state(new_state, hdr, hash_lock);
5487 hdr->b_l1hdr.b_mru_ghost_hits++;
5488 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5489 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5491 * This buffer has been accessed more than once and is
5492 * still in the cache. Keep it in the MFU state.
5494 * NOTE: an add_reference() that occurred when we did
5495 * the arc_read() will have kicked this off the list.
5496 * If it was a prefetch, we will explicitly move it to
5497 * the head of the list now.
5500 hdr->b_l1hdr.b_mfu_hits++;
5501 ARCSTAT_BUMP(arcstat_mfu_hits);
5502 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5503 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5504 arc_state_t *new_state = arc_mfu;
5506 * This buffer has been accessed more than once but has
5507 * been evicted from the cache. Move it back to the
5511 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5513 * This is a prefetch access...
5514 * move this block back to the MRU state.
5516 new_state = arc_mru;
5519 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5520 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5521 arc_change_state(new_state, hdr, hash_lock);
5523 hdr->b_l1hdr.b_mfu_ghost_hits++;
5524 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5525 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5527 * This buffer is on the 2nd Level ARC.
5530 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5531 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5532 arc_change_state(arc_mfu, hdr, hash_lock);
5534 cmn_err(CE_PANIC, "invalid arc state 0x%p",
5535 hdr->b_l1hdr.b_state);
5540 * This routine is called by dbuf_hold() to update the arc_access() state
5541 * which otherwise would be skipped for entries in the dbuf cache.
5544 arc_buf_access(arc_buf_t *buf)
5546 mutex_enter(&buf->b_evict_lock);
5547 arc_buf_hdr_t *hdr = buf->b_hdr;
5550 * Avoid taking the hash_lock when possible as an optimization.
5551 * The header must be checked again under the hash_lock in order
5552 * to handle the case where it is concurrently being released.
5554 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5555 mutex_exit(&buf->b_evict_lock);
5559 kmutex_t *hash_lock = HDR_LOCK(hdr);
5560 mutex_enter(hash_lock);
5562 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5563 mutex_exit(hash_lock);
5564 mutex_exit(&buf->b_evict_lock);
5565 ARCSTAT_BUMP(arcstat_access_skip);
5569 mutex_exit(&buf->b_evict_lock);
5571 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5572 hdr->b_l1hdr.b_state == arc_mfu);
5574 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5575 arc_access(hdr, hash_lock);
5576 mutex_exit(hash_lock);
5578 ARCSTAT_BUMP(arcstat_hits);
5579 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr) && !HDR_PRESCIENT_PREFETCH(hdr),
5580 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5583 /* a generic arc_read_done_func_t which you can use */
5586 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5587 arc_buf_t *buf, void *arg)
5592 bcopy(buf->b_data, arg, arc_buf_size(buf));
5593 arc_buf_destroy(buf, arg);
5596 /* a generic arc_read_done_func_t */
5599 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5600 arc_buf_t *buf, void *arg)
5602 arc_buf_t **bufp = arg;
5605 ASSERT(zio == NULL || zio->io_error != 0);
5608 ASSERT(zio == NULL || zio->io_error == 0);
5610 ASSERT(buf->b_data != NULL);
5615 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5617 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5618 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5619 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5621 if (HDR_COMPRESSION_ENABLED(hdr)) {
5622 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5623 BP_GET_COMPRESS(bp));
5625 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5626 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5627 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5632 arc_read_done(zio_t *zio)
5634 blkptr_t *bp = zio->io_bp;
5635 arc_buf_hdr_t *hdr = zio->io_private;
5636 kmutex_t *hash_lock = NULL;
5637 arc_callback_t *callback_list;
5638 arc_callback_t *acb;
5639 boolean_t freeable = B_FALSE;
5642 * The hdr was inserted into hash-table and removed from lists
5643 * prior to starting I/O. We should find this header, since
5644 * it's in the hash table, and it should be legit since it's
5645 * not possible to evict it during the I/O. The only possible
5646 * reason for it not to be found is if we were freed during the
5649 if (HDR_IN_HASH_TABLE(hdr)) {
5650 arc_buf_hdr_t *found;
5652 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5653 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5654 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5655 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5656 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5658 found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock);
5660 ASSERT((found == hdr &&
5661 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5662 (found == hdr && HDR_L2_READING(hdr)));
5663 ASSERT3P(hash_lock, !=, NULL);
5666 if (BP_IS_PROTECTED(bp)) {
5667 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5668 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5669 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5670 hdr->b_crypt_hdr.b_iv);
5672 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5675 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5676 sizeof (zil_chain_t));
5677 zio_crypt_decode_mac_zil(tmpbuf,
5678 hdr->b_crypt_hdr.b_mac);
5679 abd_return_buf(zio->io_abd, tmpbuf,
5680 sizeof (zil_chain_t));
5682 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
5686 if (zio->io_error == 0) {
5687 /* byteswap if necessary */
5688 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5689 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5690 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5692 hdr->b_l1hdr.b_byteswap =
5693 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5696 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5698 if (!HDR_L2_READING(hdr)) {
5699 hdr->b_complevel = zio->io_prop.zp_complevel;
5703 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5704 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5705 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5707 callback_list = hdr->b_l1hdr.b_acb;
5708 ASSERT3P(callback_list, !=, NULL);
5710 if (hash_lock && zio->io_error == 0 &&
5711 hdr->b_l1hdr.b_state == arc_anon) {
5713 * Only call arc_access on anonymous buffers. This is because
5714 * if we've issued an I/O for an evicted buffer, we've already
5715 * called arc_access (to prevent any simultaneous readers from
5716 * getting confused).
5718 arc_access(hdr, hash_lock);
5722 * If a read request has a callback (i.e. acb_done is not NULL), then we
5723 * make a buf containing the data according to the parameters which were
5724 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5725 * aren't needlessly decompressing the data multiple times.
5727 int callback_cnt = 0;
5728 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5729 if (!acb->acb_done || acb->acb_nobuf)
5734 if (zio->io_error != 0)
5737 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5738 &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
5739 acb->acb_compressed, acb->acb_noauth, B_TRUE,
5743 * Assert non-speculative zios didn't fail because an
5744 * encryption key wasn't loaded
5746 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
5750 * If we failed to decrypt, report an error now (as the zio
5751 * layer would have done if it had done the transforms).
5753 if (error == ECKSUM) {
5754 ASSERT(BP_IS_PROTECTED(bp));
5755 error = SET_ERROR(EIO);
5756 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5757 spa_log_error(zio->io_spa, &acb->acb_zb);
5758 (void) zfs_ereport_post(
5759 FM_EREPORT_ZFS_AUTHENTICATION,
5760 zio->io_spa, NULL, &acb->acb_zb, zio, 0);
5766 * Decompression or decryption failed. Set
5767 * io_error so that when we call acb_done
5768 * (below), we will indicate that the read
5769 * failed. Note that in the unusual case
5770 * where one callback is compressed and another
5771 * uncompressed, we will mark all of them
5772 * as failed, even though the uncompressed
5773 * one can't actually fail. In this case,
5774 * the hdr will not be anonymous, because
5775 * if there are multiple callbacks, it's
5776 * because multiple threads found the same
5777 * arc buf in the hash table.
5779 zio->io_error = error;
5784 * If there are multiple callbacks, we must have the hash lock,
5785 * because the only way for multiple threads to find this hdr is
5786 * in the hash table. This ensures that if there are multiple
5787 * callbacks, the hdr is not anonymous. If it were anonymous,
5788 * we couldn't use arc_buf_destroy() in the error case below.
5790 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5792 hdr->b_l1hdr.b_acb = NULL;
5793 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5794 if (callback_cnt == 0)
5795 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
5797 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5798 callback_list != NULL);
5800 if (zio->io_error == 0) {
5801 arc_hdr_verify(hdr, zio->io_bp);
5803 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5804 if (hdr->b_l1hdr.b_state != arc_anon)
5805 arc_change_state(arc_anon, hdr, hash_lock);
5806 if (HDR_IN_HASH_TABLE(hdr))
5807 buf_hash_remove(hdr);
5808 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5812 * Broadcast before we drop the hash_lock to avoid the possibility
5813 * that the hdr (and hence the cv) might be freed before we get to
5814 * the cv_broadcast().
5816 cv_broadcast(&hdr->b_l1hdr.b_cv);
5818 if (hash_lock != NULL) {
5819 mutex_exit(hash_lock);
5822 * This block was freed while we waited for the read to
5823 * complete. It has been removed from the hash table and
5824 * moved to the anonymous state (so that it won't show up
5827 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5828 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5831 /* execute each callback and free its structure */
5832 while ((acb = callback_list) != NULL) {
5833 if (acb->acb_done != NULL) {
5834 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5836 * If arc_buf_alloc_impl() fails during
5837 * decompression, the buf will still be
5838 * allocated, and needs to be freed here.
5840 arc_buf_destroy(acb->acb_buf,
5842 acb->acb_buf = NULL;
5844 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5845 acb->acb_buf, acb->acb_private);
5848 if (acb->acb_zio_dummy != NULL) {
5849 acb->acb_zio_dummy->io_error = zio->io_error;
5850 zio_nowait(acb->acb_zio_dummy);
5853 callback_list = acb->acb_next;
5854 kmem_free(acb, sizeof (arc_callback_t));
5858 arc_hdr_destroy(hdr);
5862 * "Read" the block at the specified DVA (in bp) via the
5863 * cache. If the block is found in the cache, invoke the provided
5864 * callback immediately and return. Note that the `zio' parameter
5865 * in the callback will be NULL in this case, since no IO was
5866 * required. If the block is not in the cache pass the read request
5867 * on to the spa with a substitute callback function, so that the
5868 * requested block will be added to the cache.
5870 * If a read request arrives for a block that has a read in-progress,
5871 * either wait for the in-progress read to complete (and return the
5872 * results); or, if this is a read with a "done" func, add a record
5873 * to the read to invoke the "done" func when the read completes,
5874 * and return; or just return.
5876 * arc_read_done() will invoke all the requested "done" functions
5877 * for readers of this block.
5880 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
5881 arc_read_done_func_t *done, void *private, zio_priority_t priority,
5882 int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5884 arc_buf_hdr_t *hdr = NULL;
5885 kmutex_t *hash_lock = NULL;
5887 uint64_t guid = spa_load_guid(spa);
5888 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
5889 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
5890 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5891 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
5892 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5893 boolean_t embedded_bp = !!BP_IS_EMBEDDED(bp);
5894 boolean_t no_buf = *arc_flags & ARC_FLAG_NO_BUF;
5897 ASSERT(!embedded_bp ||
5898 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5899 ASSERT(!BP_IS_HOLE(bp));
5900 ASSERT(!BP_IS_REDACTED(bp));
5903 * Normally SPL_FSTRANS will already be set since kernel threads which
5904 * expect to call the DMU interfaces will set it when created. System
5905 * calls are similarly handled by setting/cleaning the bit in the
5906 * registered callback (module/os/.../zfs/zpl_*).
5908 * External consumers such as Lustre which call the exported DMU
5909 * interfaces may not have set SPL_FSTRANS. To avoid a deadlock
5910 * on the hash_lock always set and clear the bit.
5912 fstrans_cookie_t cookie = spl_fstrans_mark();
5915 * Verify the block pointer contents are reasonable. This should
5916 * always be the case since the blkptr is protected by a checksum.
5917 * However, if there is damage it's desirable to detect this early
5918 * and treat it as a checksum error. This allows an alternate blkptr
5919 * to be tried when one is available (e.g. ditto blocks).
5921 if (!zfs_blkptr_verify(spa, bp, zio_flags & ZIO_FLAG_CONFIG_WRITER,
5923 rc = SET_ERROR(ECKSUM);
5929 * Embedded BP's have no DVA and require no I/O to "read".
5930 * Create an anonymous arc buf to back it.
5932 hdr = buf_hash_find(guid, bp, &hash_lock);
5936 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5937 * we maintain encrypted data separately from compressed / uncompressed
5938 * data. If the user is requesting raw encrypted data and we don't have
5939 * that in the header we will read from disk to guarantee that we can
5940 * get it even if the encryption keys aren't loaded.
5942 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
5943 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
5944 arc_buf_t *buf = NULL;
5945 *arc_flags |= ARC_FLAG_CACHED;
5947 if (HDR_IO_IN_PROGRESS(hdr)) {
5948 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5950 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
5951 mutex_exit(hash_lock);
5952 ARCSTAT_BUMP(arcstat_cached_only_in_progress);
5953 rc = SET_ERROR(ENOENT);
5957 ASSERT3P(head_zio, !=, NULL);
5958 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5959 priority == ZIO_PRIORITY_SYNC_READ) {
5961 * This is a sync read that needs to wait for
5962 * an in-flight async read. Request that the
5963 * zio have its priority upgraded.
5965 zio_change_priority(head_zio, priority);
5966 DTRACE_PROBE1(arc__async__upgrade__sync,
5967 arc_buf_hdr_t *, hdr);
5968 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5970 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5971 arc_hdr_clear_flags(hdr,
5972 ARC_FLAG_PREDICTIVE_PREFETCH);
5975 if (*arc_flags & ARC_FLAG_WAIT) {
5976 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5977 mutex_exit(hash_lock);
5980 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5983 arc_callback_t *acb = NULL;
5985 acb = kmem_zalloc(sizeof (arc_callback_t),
5987 acb->acb_done = done;
5988 acb->acb_private = private;
5989 acb->acb_compressed = compressed_read;
5990 acb->acb_encrypted = encrypted_read;
5991 acb->acb_noauth = noauth_read;
5992 acb->acb_nobuf = no_buf;
5995 acb->acb_zio_dummy = zio_null(pio,
5996 spa, NULL, NULL, NULL, zio_flags);
5998 ASSERT3P(acb->acb_done, !=, NULL);
5999 acb->acb_zio_head = head_zio;
6000 acb->acb_next = hdr->b_l1hdr.b_acb;
6001 hdr->b_l1hdr.b_acb = acb;
6003 mutex_exit(hash_lock);
6007 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
6008 hdr->b_l1hdr.b_state == arc_mfu);
6010 if (done && !no_buf) {
6011 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
6013 * This is a demand read which does not have to
6014 * wait for i/o because we did a predictive
6015 * prefetch i/o for it, which has completed.
6018 arc__demand__hit__predictive__prefetch,
6019 arc_buf_hdr_t *, hdr);
6021 arcstat_demand_hit_predictive_prefetch);
6022 arc_hdr_clear_flags(hdr,
6023 ARC_FLAG_PREDICTIVE_PREFETCH);
6026 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
6028 arcstat_demand_hit_prescient_prefetch);
6029 arc_hdr_clear_flags(hdr,
6030 ARC_FLAG_PRESCIENT_PREFETCH);
6033 ASSERT(!embedded_bp || !BP_IS_HOLE(bp));
6035 /* Get a buf with the desired data in it. */
6036 rc = arc_buf_alloc_impl(hdr, spa, zb, private,
6037 encrypted_read, compressed_read, noauth_read,
6041 * Convert authentication and decryption errors
6042 * to EIO (and generate an ereport if needed)
6043 * before leaving the ARC.
6045 rc = SET_ERROR(EIO);
6046 if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
6047 spa_log_error(spa, zb);
6048 (void) zfs_ereport_post(
6049 FM_EREPORT_ZFS_AUTHENTICATION,
6050 spa, NULL, zb, NULL, 0);
6054 (void) remove_reference(hdr, hash_lock,
6056 arc_buf_destroy_impl(buf);
6060 /* assert any errors weren't due to unloaded keys */
6061 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
6063 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
6064 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6065 if (HDR_HAS_L2HDR(hdr))
6066 l2arc_hdr_arcstats_decrement_state(hdr);
6067 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6068 if (HDR_HAS_L2HDR(hdr))
6069 l2arc_hdr_arcstats_increment_state(hdr);
6071 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
6072 arc_access(hdr, hash_lock);
6073 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6074 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6075 if (*arc_flags & ARC_FLAG_L2CACHE)
6076 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6077 mutex_exit(hash_lock);
6078 ARCSTAT_BUMP(arcstat_hits);
6079 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6080 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
6081 data, metadata, hits);
6084 done(NULL, zb, bp, buf, private);
6086 uint64_t lsize = BP_GET_LSIZE(bp);
6087 uint64_t psize = BP_GET_PSIZE(bp);
6088 arc_callback_t *acb;
6091 boolean_t devw = B_FALSE;
6094 int alloc_flags = encrypted_read ? ARC_HDR_ALLOC_RDATA : 0;
6096 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
6097 rc = SET_ERROR(ENOENT);
6098 if (hash_lock != NULL)
6099 mutex_exit(hash_lock);
6105 * This block is not in the cache or it has
6108 arc_buf_hdr_t *exists = NULL;
6109 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
6110 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
6111 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), 0, type);
6114 hdr->b_dva = *BP_IDENTITY(bp);
6115 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
6116 exists = buf_hash_insert(hdr, &hash_lock);
6118 if (exists != NULL) {
6119 /* somebody beat us to the hash insert */
6120 mutex_exit(hash_lock);
6121 buf_discard_identity(hdr);
6122 arc_hdr_destroy(hdr);
6123 goto top; /* restart the IO request */
6125 alloc_flags |= ARC_HDR_DO_ADAPT;
6128 * This block is in the ghost cache or encrypted data
6129 * was requested and we didn't have it. If it was
6130 * L2-only (and thus didn't have an L1 hdr),
6131 * we realloc the header to add an L1 hdr.
6133 if (!HDR_HAS_L1HDR(hdr)) {
6134 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
6138 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
6139 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6140 ASSERT(!HDR_HAS_RABD(hdr));
6141 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6142 ASSERT0(zfs_refcount_count(
6143 &hdr->b_l1hdr.b_refcnt));
6144 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
6145 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
6146 } else if (HDR_IO_IN_PROGRESS(hdr)) {
6148 * If this header already had an IO in progress
6149 * and we are performing another IO to fetch
6150 * encrypted data we must wait until the first
6151 * IO completes so as not to confuse
6152 * arc_read_done(). This should be very rare
6153 * and so the performance impact shouldn't
6156 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
6157 mutex_exit(hash_lock);
6162 * This is a delicate dance that we play here.
6163 * This hdr might be in the ghost list so we access
6164 * it to move it out of the ghost list before we
6165 * initiate the read. If it's a prefetch then
6166 * it won't have a callback so we'll remove the
6167 * reference that arc_buf_alloc_impl() created. We
6168 * do this after we've called arc_access() to
6169 * avoid hitting an assert in remove_reference().
6171 arc_adapt(arc_hdr_size(hdr), hdr->b_l1hdr.b_state);
6172 arc_access(hdr, hash_lock);
6175 arc_hdr_alloc_abd(hdr, alloc_flags);
6176 if (encrypted_read) {
6177 ASSERT(HDR_HAS_RABD(hdr));
6178 size = HDR_GET_PSIZE(hdr);
6179 hdr_abd = hdr->b_crypt_hdr.b_rabd;
6180 zio_flags |= ZIO_FLAG_RAW;
6182 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6183 size = arc_hdr_size(hdr);
6184 hdr_abd = hdr->b_l1hdr.b_pabd;
6186 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
6187 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6191 * For authenticated bp's, we do not ask the ZIO layer
6192 * to authenticate them since this will cause the entire
6193 * IO to fail if the key isn't loaded. Instead, we
6194 * defer authentication until arc_buf_fill(), which will
6195 * verify the data when the key is available.
6197 if (BP_IS_AUTHENTICATED(bp))
6198 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
6201 if (*arc_flags & ARC_FLAG_PREFETCH &&
6202 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6203 if (HDR_HAS_L2HDR(hdr))
6204 l2arc_hdr_arcstats_decrement_state(hdr);
6205 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6206 if (HDR_HAS_L2HDR(hdr))
6207 l2arc_hdr_arcstats_increment_state(hdr);
6209 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6210 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6211 if (*arc_flags & ARC_FLAG_L2CACHE)
6212 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6213 if (BP_IS_AUTHENTICATED(bp))
6214 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6215 if (BP_GET_LEVEL(bp) > 0)
6216 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
6217 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
6218 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
6219 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
6221 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
6222 acb->acb_done = done;
6223 acb->acb_private = private;
6224 acb->acb_compressed = compressed_read;
6225 acb->acb_encrypted = encrypted_read;
6226 acb->acb_noauth = noauth_read;
6229 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6230 hdr->b_l1hdr.b_acb = acb;
6231 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6233 if (HDR_HAS_L2HDR(hdr) &&
6234 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
6235 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
6236 addr = hdr->b_l2hdr.b_daddr;
6238 * Lock out L2ARC device removal.
6240 if (vdev_is_dead(vd) ||
6241 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
6246 * We count both async reads and scrub IOs as asynchronous so
6247 * that both can be upgraded in the event of a cache hit while
6248 * the read IO is still in-flight.
6250 if (priority == ZIO_PRIORITY_ASYNC_READ ||
6251 priority == ZIO_PRIORITY_SCRUB)
6252 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6254 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6257 * At this point, we have a level 1 cache miss or a blkptr
6258 * with embedded data. Try again in L2ARC if possible.
6260 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
6263 * Skip ARC stat bump for block pointers with embedded
6264 * data. The data are read from the blkptr itself via
6265 * decode_embedded_bp_compressed().
6268 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr,
6269 blkptr_t *, bp, uint64_t, lsize,
6270 zbookmark_phys_t *, zb);
6271 ARCSTAT_BUMP(arcstat_misses);
6272 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6273 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data,
6275 zfs_racct_read(size, 1);
6278 /* Check if the spa even has l2 configured */
6279 const boolean_t spa_has_l2 = l2arc_ndev != 0 &&
6280 spa->spa_l2cache.sav_count > 0;
6282 if (vd != NULL && spa_has_l2 && !(l2arc_norw && devw)) {
6284 * Read from the L2ARC if the following are true:
6285 * 1. The L2ARC vdev was previously cached.
6286 * 2. This buffer still has L2ARC metadata.
6287 * 3. This buffer isn't currently writing to the L2ARC.
6288 * 4. The L2ARC entry wasn't evicted, which may
6289 * also have invalidated the vdev.
6290 * 5. This isn't prefetch or l2arc_noprefetch is 0.
6292 if (HDR_HAS_L2HDR(hdr) &&
6293 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
6294 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
6295 l2arc_read_callback_t *cb;
6299 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
6300 ARCSTAT_BUMP(arcstat_l2_hits);
6301 hdr->b_l2hdr.b_hits++;
6303 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
6305 cb->l2rcb_hdr = hdr;
6308 cb->l2rcb_flags = zio_flags;
6311 * When Compressed ARC is disabled, but the
6312 * L2ARC block is compressed, arc_hdr_size()
6313 * will have returned LSIZE rather than PSIZE.
6315 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
6316 !HDR_COMPRESSION_ENABLED(hdr) &&
6317 HDR_GET_PSIZE(hdr) != 0) {
6318 size = HDR_GET_PSIZE(hdr);
6321 asize = vdev_psize_to_asize(vd, size);
6322 if (asize != size) {
6323 abd = abd_alloc_for_io(asize,
6324 HDR_ISTYPE_METADATA(hdr));
6325 cb->l2rcb_abd = abd;
6330 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6331 addr + asize <= vd->vdev_psize -
6332 VDEV_LABEL_END_SIZE);
6335 * l2arc read. The SCL_L2ARC lock will be
6336 * released by l2arc_read_done().
6337 * Issue a null zio if the underlying buffer
6338 * was squashed to zero size by compression.
6340 ASSERT3U(arc_hdr_get_compress(hdr), !=,
6341 ZIO_COMPRESS_EMPTY);
6342 rzio = zio_read_phys(pio, vd, addr,
6345 l2arc_read_done, cb, priority,
6346 zio_flags | ZIO_FLAG_DONT_CACHE |
6348 ZIO_FLAG_DONT_PROPAGATE |
6349 ZIO_FLAG_DONT_RETRY, B_FALSE);
6350 acb->acb_zio_head = rzio;
6352 if (hash_lock != NULL)
6353 mutex_exit(hash_lock);
6355 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6357 ARCSTAT_INCR(arcstat_l2_read_bytes,
6358 HDR_GET_PSIZE(hdr));
6360 if (*arc_flags & ARC_FLAG_NOWAIT) {
6365 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6366 if (zio_wait(rzio) == 0)
6369 /* l2arc read error; goto zio_read() */
6370 if (hash_lock != NULL)
6371 mutex_enter(hash_lock);
6373 DTRACE_PROBE1(l2arc__miss,
6374 arc_buf_hdr_t *, hdr);
6375 ARCSTAT_BUMP(arcstat_l2_misses);
6376 if (HDR_L2_WRITING(hdr))
6377 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6378 spa_config_exit(spa, SCL_L2ARC, vd);
6382 spa_config_exit(spa, SCL_L2ARC, vd);
6385 * Only a spa with l2 should contribute to l2
6386 * miss stats. (Including the case of having a
6387 * faulted cache device - that's also a miss.)
6391 * Skip ARC stat bump for block pointers with
6392 * embedded data. The data are read from the
6394 * decode_embedded_bp_compressed().
6397 DTRACE_PROBE1(l2arc__miss,
6398 arc_buf_hdr_t *, hdr);
6399 ARCSTAT_BUMP(arcstat_l2_misses);
6404 rzio = zio_read(pio, spa, bp, hdr_abd, size,
6405 arc_read_done, hdr, priority, zio_flags, zb);
6406 acb->acb_zio_head = rzio;
6408 if (hash_lock != NULL)
6409 mutex_exit(hash_lock);
6411 if (*arc_flags & ARC_FLAG_WAIT) {
6412 rc = zio_wait(rzio);
6416 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6421 /* embedded bps don't actually go to disk */
6423 spa_read_history_add(spa, zb, *arc_flags);
6424 spl_fstrans_unmark(cookie);
6429 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6433 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6435 p->p_private = private;
6436 list_link_init(&p->p_node);
6437 zfs_refcount_create(&p->p_refcnt);
6439 mutex_enter(&arc_prune_mtx);
6440 zfs_refcount_add(&p->p_refcnt, &arc_prune_list);
6441 list_insert_head(&arc_prune_list, p);
6442 mutex_exit(&arc_prune_mtx);
6448 arc_remove_prune_callback(arc_prune_t *p)
6450 boolean_t wait = B_FALSE;
6451 mutex_enter(&arc_prune_mtx);
6452 list_remove(&arc_prune_list, p);
6453 if (zfs_refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6455 mutex_exit(&arc_prune_mtx);
6457 /* wait for arc_prune_task to finish */
6459 taskq_wait_outstanding(arc_prune_taskq, 0);
6460 ASSERT0(zfs_refcount_count(&p->p_refcnt));
6461 zfs_refcount_destroy(&p->p_refcnt);
6462 kmem_free(p, sizeof (*p));
6466 * Notify the arc that a block was freed, and thus will never be used again.
6469 arc_freed(spa_t *spa, const blkptr_t *bp)
6472 kmutex_t *hash_lock;
6473 uint64_t guid = spa_load_guid(spa);
6475 ASSERT(!BP_IS_EMBEDDED(bp));
6477 hdr = buf_hash_find(guid, bp, &hash_lock);
6482 * We might be trying to free a block that is still doing I/O
6483 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6484 * dmu_sync-ed block). If this block is being prefetched, then it
6485 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6486 * until the I/O completes. A block may also have a reference if it is
6487 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6488 * have written the new block to its final resting place on disk but
6489 * without the dedup flag set. This would have left the hdr in the MRU
6490 * state and discoverable. When the txg finally syncs it detects that
6491 * the block was overridden in open context and issues an override I/O.
6492 * Since this is a dedup block, the override I/O will determine if the
6493 * block is already in the DDT. If so, then it will replace the io_bp
6494 * with the bp from the DDT and allow the I/O to finish. When the I/O
6495 * reaches the done callback, dbuf_write_override_done, it will
6496 * check to see if the io_bp and io_bp_override are identical.
6497 * If they are not, then it indicates that the bp was replaced with
6498 * the bp in the DDT and the override bp is freed. This allows
6499 * us to arrive here with a reference on a block that is being
6500 * freed. So if we have an I/O in progress, or a reference to
6501 * this hdr, then we don't destroy the hdr.
6503 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6504 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6505 arc_change_state(arc_anon, hdr, hash_lock);
6506 arc_hdr_destroy(hdr);
6507 mutex_exit(hash_lock);
6509 mutex_exit(hash_lock);
6515 * Release this buffer from the cache, making it an anonymous buffer. This
6516 * must be done after a read and prior to modifying the buffer contents.
6517 * If the buffer has more than one reference, we must make
6518 * a new hdr for the buffer.
6521 arc_release(arc_buf_t *buf, void *tag)
6523 arc_buf_hdr_t *hdr = buf->b_hdr;
6526 * It would be nice to assert that if its DMU metadata (level >
6527 * 0 || it's the dnode file), then it must be syncing context.
6528 * But we don't know that information at this level.
6531 mutex_enter(&buf->b_evict_lock);
6533 ASSERT(HDR_HAS_L1HDR(hdr));
6536 * We don't grab the hash lock prior to this check, because if
6537 * the buffer's header is in the arc_anon state, it won't be
6538 * linked into the hash table.
6540 if (hdr->b_l1hdr.b_state == arc_anon) {
6541 mutex_exit(&buf->b_evict_lock);
6542 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6543 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6544 ASSERT(!HDR_HAS_L2HDR(hdr));
6546 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6547 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6548 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6550 hdr->b_l1hdr.b_arc_access = 0;
6553 * If the buf is being overridden then it may already
6554 * have a hdr that is not empty.
6556 buf_discard_identity(hdr);
6562 kmutex_t *hash_lock = HDR_LOCK(hdr);
6563 mutex_enter(hash_lock);
6566 * This assignment is only valid as long as the hash_lock is
6567 * held, we must be careful not to reference state or the
6568 * b_state field after dropping the lock.
6570 arc_state_t *state = hdr->b_l1hdr.b_state;
6571 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6572 ASSERT3P(state, !=, arc_anon);
6574 /* this buffer is not on any list */
6575 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6577 if (HDR_HAS_L2HDR(hdr)) {
6578 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6581 * We have to recheck this conditional again now that
6582 * we're holding the l2ad_mtx to prevent a race with
6583 * another thread which might be concurrently calling
6584 * l2arc_evict(). In that case, l2arc_evict() might have
6585 * destroyed the header's L2 portion as we were waiting
6586 * to acquire the l2ad_mtx.
6588 if (HDR_HAS_L2HDR(hdr))
6589 arc_hdr_l2hdr_destroy(hdr);
6591 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6595 * Do we have more than one buf?
6597 if (hdr->b_l1hdr.b_bufcnt > 1) {
6598 arc_buf_hdr_t *nhdr;
6599 uint64_t spa = hdr->b_spa;
6600 uint64_t psize = HDR_GET_PSIZE(hdr);
6601 uint64_t lsize = HDR_GET_LSIZE(hdr);
6602 boolean_t protected = HDR_PROTECTED(hdr);
6603 enum zio_compress compress = arc_hdr_get_compress(hdr);
6604 arc_buf_contents_t type = arc_buf_type(hdr);
6605 VERIFY3U(hdr->b_type, ==, type);
6607 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6608 (void) remove_reference(hdr, hash_lock, tag);
6610 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6611 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6612 ASSERT(ARC_BUF_LAST(buf));
6616 * Pull the data off of this hdr and attach it to
6617 * a new anonymous hdr. Also find the last buffer
6618 * in the hdr's buffer list.
6620 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6621 ASSERT3P(lastbuf, !=, NULL);
6624 * If the current arc_buf_t and the hdr are sharing their data
6625 * buffer, then we must stop sharing that block.
6627 if (arc_buf_is_shared(buf)) {
6628 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6629 VERIFY(!arc_buf_is_shared(lastbuf));
6632 * First, sever the block sharing relationship between
6633 * buf and the arc_buf_hdr_t.
6635 arc_unshare_buf(hdr, buf);
6638 * Now we need to recreate the hdr's b_pabd. Since we
6639 * have lastbuf handy, we try to share with it, but if
6640 * we can't then we allocate a new b_pabd and copy the
6641 * data from buf into it.
6643 if (arc_can_share(hdr, lastbuf)) {
6644 arc_share_buf(hdr, lastbuf);
6646 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT);
6647 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6648 buf->b_data, psize);
6650 VERIFY3P(lastbuf->b_data, !=, NULL);
6651 } else if (HDR_SHARED_DATA(hdr)) {
6653 * Uncompressed shared buffers are always at the end
6654 * of the list. Compressed buffers don't have the
6655 * same requirements. This makes it hard to
6656 * simply assert that the lastbuf is shared so
6657 * we rely on the hdr's compression flags to determine
6658 * if we have a compressed, shared buffer.
6660 ASSERT(arc_buf_is_shared(lastbuf) ||
6661 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6662 ASSERT(!ARC_BUF_SHARED(buf));
6665 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6666 ASSERT3P(state, !=, arc_l2c_only);
6668 (void) zfs_refcount_remove_many(&state->arcs_size,
6669 arc_buf_size(buf), buf);
6671 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6672 ASSERT3P(state, !=, arc_l2c_only);
6673 (void) zfs_refcount_remove_many(
6674 &state->arcs_esize[type],
6675 arc_buf_size(buf), buf);
6678 hdr->b_l1hdr.b_bufcnt -= 1;
6679 if (ARC_BUF_ENCRYPTED(buf))
6680 hdr->b_crypt_hdr.b_ebufcnt -= 1;
6682 arc_cksum_verify(buf);
6683 arc_buf_unwatch(buf);
6685 /* if this is the last uncompressed buf free the checksum */
6686 if (!arc_hdr_has_uncompressed_buf(hdr))
6687 arc_cksum_free(hdr);
6689 mutex_exit(hash_lock);
6692 * Allocate a new hdr. The new hdr will contain a b_pabd
6693 * buffer which will be freed in arc_write().
6695 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6696 compress, hdr->b_complevel, type);
6697 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6698 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6699 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6700 VERIFY3U(nhdr->b_type, ==, type);
6701 ASSERT(!HDR_SHARED_DATA(nhdr));
6703 nhdr->b_l1hdr.b_buf = buf;
6704 nhdr->b_l1hdr.b_bufcnt = 1;
6705 if (ARC_BUF_ENCRYPTED(buf))
6706 nhdr->b_crypt_hdr.b_ebufcnt = 1;
6707 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6710 mutex_exit(&buf->b_evict_lock);
6711 (void) zfs_refcount_add_many(&arc_anon->arcs_size,
6712 arc_buf_size(buf), buf);
6714 mutex_exit(&buf->b_evict_lock);
6715 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6716 /* protected by hash lock, or hdr is on arc_anon */
6717 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6718 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6719 hdr->b_l1hdr.b_mru_hits = 0;
6720 hdr->b_l1hdr.b_mru_ghost_hits = 0;
6721 hdr->b_l1hdr.b_mfu_hits = 0;
6722 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
6723 arc_change_state(arc_anon, hdr, hash_lock);
6724 hdr->b_l1hdr.b_arc_access = 0;
6726 mutex_exit(hash_lock);
6727 buf_discard_identity(hdr);
6733 arc_released(arc_buf_t *buf)
6737 mutex_enter(&buf->b_evict_lock);
6738 released = (buf->b_data != NULL &&
6739 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6740 mutex_exit(&buf->b_evict_lock);
6746 arc_referenced(arc_buf_t *buf)
6750 mutex_enter(&buf->b_evict_lock);
6751 referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6752 mutex_exit(&buf->b_evict_lock);
6753 return (referenced);
6758 arc_write_ready(zio_t *zio)
6760 arc_write_callback_t *callback = zio->io_private;
6761 arc_buf_t *buf = callback->awcb_buf;
6762 arc_buf_hdr_t *hdr = buf->b_hdr;
6763 blkptr_t *bp = zio->io_bp;
6764 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6765 fstrans_cookie_t cookie = spl_fstrans_mark();
6767 ASSERT(HDR_HAS_L1HDR(hdr));
6768 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6769 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6772 * If we're reexecuting this zio because the pool suspended, then
6773 * cleanup any state that was previously set the first time the
6774 * callback was invoked.
6776 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6777 arc_cksum_free(hdr);
6778 arc_buf_unwatch(buf);
6779 if (hdr->b_l1hdr.b_pabd != NULL) {
6780 if (arc_buf_is_shared(buf)) {
6781 arc_unshare_buf(hdr, buf);
6783 arc_hdr_free_abd(hdr, B_FALSE);
6787 if (HDR_HAS_RABD(hdr))
6788 arc_hdr_free_abd(hdr, B_TRUE);
6790 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6791 ASSERT(!HDR_HAS_RABD(hdr));
6792 ASSERT(!HDR_SHARED_DATA(hdr));
6793 ASSERT(!arc_buf_is_shared(buf));
6795 callback->awcb_ready(zio, buf, callback->awcb_private);
6797 if (HDR_IO_IN_PROGRESS(hdr))
6798 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6800 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6802 if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr))
6803 hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp));
6805 if (BP_IS_PROTECTED(bp)) {
6806 /* ZIL blocks are written through zio_rewrite */
6807 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6808 ASSERT(HDR_PROTECTED(hdr));
6810 if (BP_SHOULD_BYTESWAP(bp)) {
6811 if (BP_GET_LEVEL(bp) > 0) {
6812 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6814 hdr->b_l1hdr.b_byteswap =
6815 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6818 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6821 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6822 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6823 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6824 hdr->b_crypt_hdr.b_iv);
6825 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6829 * If this block was written for raw encryption but the zio layer
6830 * ended up only authenticating it, adjust the buffer flags now.
6832 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6833 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6834 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6835 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6836 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6837 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6838 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6839 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6842 /* this must be done after the buffer flags are adjusted */
6843 arc_cksum_compute(buf);
6845 enum zio_compress compress;
6846 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6847 compress = ZIO_COMPRESS_OFF;
6849 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6850 compress = BP_GET_COMPRESS(bp);
6852 HDR_SET_PSIZE(hdr, psize);
6853 arc_hdr_set_compress(hdr, compress);
6854 hdr->b_complevel = zio->io_prop.zp_complevel;
6856 if (zio->io_error != 0 || psize == 0)
6860 * Fill the hdr with data. If the buffer is encrypted we have no choice
6861 * but to copy the data into b_radb. If the hdr is compressed, the data
6862 * we want is available from the zio, otherwise we can take it from
6865 * We might be able to share the buf's data with the hdr here. However,
6866 * doing so would cause the ARC to be full of linear ABDs if we write a
6867 * lot of shareable data. As a compromise, we check whether scattered
6868 * ABDs are allowed, and assume that if they are then the user wants
6869 * the ARC to be primarily filled with them regardless of the data being
6870 * written. Therefore, if they're allowed then we allocate one and copy
6871 * the data into it; otherwise, we share the data directly if we can.
6873 if (ARC_BUF_ENCRYPTED(buf)) {
6874 ASSERT3U(psize, >, 0);
6875 ASSERT(ARC_BUF_COMPRESSED(buf));
6876 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT | ARC_HDR_ALLOC_RDATA |
6877 ARC_HDR_USE_RESERVE);
6878 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6879 } else if (!abd_size_alloc_linear(arc_buf_size(buf)) ||
6880 !arc_can_share(hdr, buf)) {
6882 * Ideally, we would always copy the io_abd into b_pabd, but the
6883 * user may have disabled compressed ARC, thus we must check the
6884 * hdr's compression setting rather than the io_bp's.
6886 if (BP_IS_ENCRYPTED(bp)) {
6887 ASSERT3U(psize, >, 0);
6888 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT |
6889 ARC_HDR_ALLOC_RDATA | ARC_HDR_USE_RESERVE);
6890 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6891 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
6892 !ARC_BUF_COMPRESSED(buf)) {
6893 ASSERT3U(psize, >, 0);
6894 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT |
6895 ARC_HDR_USE_RESERVE);
6896 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6898 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6899 arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT |
6900 ARC_HDR_USE_RESERVE);
6901 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6905 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6906 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6907 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6909 arc_share_buf(hdr, buf);
6913 arc_hdr_verify(hdr, bp);
6914 spl_fstrans_unmark(cookie);
6918 arc_write_children_ready(zio_t *zio)
6920 arc_write_callback_t *callback = zio->io_private;
6921 arc_buf_t *buf = callback->awcb_buf;
6923 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6927 * The SPA calls this callback for each physical write that happens on behalf
6928 * of a logical write. See the comment in dbuf_write_physdone() for details.
6931 arc_write_physdone(zio_t *zio)
6933 arc_write_callback_t *cb = zio->io_private;
6934 if (cb->awcb_physdone != NULL)
6935 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6939 arc_write_done(zio_t *zio)
6941 arc_write_callback_t *callback = zio->io_private;
6942 arc_buf_t *buf = callback->awcb_buf;
6943 arc_buf_hdr_t *hdr = buf->b_hdr;
6945 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6947 if (zio->io_error == 0) {
6948 arc_hdr_verify(hdr, zio->io_bp);
6950 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6951 buf_discard_identity(hdr);
6953 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6954 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6957 ASSERT(HDR_EMPTY(hdr));
6961 * If the block to be written was all-zero or compressed enough to be
6962 * embedded in the BP, no write was performed so there will be no
6963 * dva/birth/checksum. The buffer must therefore remain anonymous
6966 if (!HDR_EMPTY(hdr)) {
6967 arc_buf_hdr_t *exists;
6968 kmutex_t *hash_lock;
6970 ASSERT3U(zio->io_error, ==, 0);
6972 arc_cksum_verify(buf);
6974 exists = buf_hash_insert(hdr, &hash_lock);
6975 if (exists != NULL) {
6977 * This can only happen if we overwrite for
6978 * sync-to-convergence, because we remove
6979 * buffers from the hash table when we arc_free().
6981 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6982 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6983 panic("bad overwrite, hdr=%p exists=%p",
6984 (void *)hdr, (void *)exists);
6985 ASSERT(zfs_refcount_is_zero(
6986 &exists->b_l1hdr.b_refcnt));
6987 arc_change_state(arc_anon, exists, hash_lock);
6988 arc_hdr_destroy(exists);
6989 mutex_exit(hash_lock);
6990 exists = buf_hash_insert(hdr, &hash_lock);
6991 ASSERT3P(exists, ==, NULL);
6992 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6994 ASSERT(zio->io_prop.zp_nopwrite);
6995 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6996 panic("bad nopwrite, hdr=%p exists=%p",
6997 (void *)hdr, (void *)exists);
7000 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
7001 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
7002 ASSERT(BP_GET_DEDUP(zio->io_bp));
7003 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
7006 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
7007 /* if it's not anon, we are doing a scrub */
7008 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
7009 arc_access(hdr, hash_lock);
7010 mutex_exit(hash_lock);
7012 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
7015 ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
7016 callback->awcb_done(zio, buf, callback->awcb_private);
7018 abd_free(zio->io_abd);
7019 kmem_free(callback, sizeof (arc_write_callback_t));
7023 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
7024 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc,
7025 const zio_prop_t *zp, arc_write_done_func_t *ready,
7026 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
7027 arc_write_done_func_t *done, void *private, zio_priority_t priority,
7028 int zio_flags, const zbookmark_phys_t *zb)
7030 arc_buf_hdr_t *hdr = buf->b_hdr;
7031 arc_write_callback_t *callback;
7033 zio_prop_t localprop = *zp;
7035 ASSERT3P(ready, !=, NULL);
7036 ASSERT3P(done, !=, NULL);
7037 ASSERT(!HDR_IO_ERROR(hdr));
7038 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
7039 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
7040 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
7042 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
7044 if (ARC_BUF_ENCRYPTED(buf)) {
7045 ASSERT(ARC_BUF_COMPRESSED(buf));
7046 localprop.zp_encrypt = B_TRUE;
7047 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
7048 localprop.zp_complevel = hdr->b_complevel;
7049 localprop.zp_byteorder =
7050 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
7051 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
7052 bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt,
7054 bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv,
7056 bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac,
7058 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
7059 localprop.zp_nopwrite = B_FALSE;
7060 localprop.zp_copies =
7061 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
7063 zio_flags |= ZIO_FLAG_RAW;
7064 } else if (ARC_BUF_COMPRESSED(buf)) {
7065 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
7066 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
7067 localprop.zp_complevel = hdr->b_complevel;
7068 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
7070 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
7071 callback->awcb_ready = ready;
7072 callback->awcb_children_ready = children_ready;
7073 callback->awcb_physdone = physdone;
7074 callback->awcb_done = done;
7075 callback->awcb_private = private;
7076 callback->awcb_buf = buf;
7079 * The hdr's b_pabd is now stale, free it now. A new data block
7080 * will be allocated when the zio pipeline calls arc_write_ready().
7082 if (hdr->b_l1hdr.b_pabd != NULL) {
7084 * If the buf is currently sharing the data block with
7085 * the hdr then we need to break that relationship here.
7086 * The hdr will remain with a NULL data pointer and the
7087 * buf will take sole ownership of the block.
7089 if (arc_buf_is_shared(buf)) {
7090 arc_unshare_buf(hdr, buf);
7092 arc_hdr_free_abd(hdr, B_FALSE);
7094 VERIFY3P(buf->b_data, !=, NULL);
7097 if (HDR_HAS_RABD(hdr))
7098 arc_hdr_free_abd(hdr, B_TRUE);
7100 if (!(zio_flags & ZIO_FLAG_RAW))
7101 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
7103 ASSERT(!arc_buf_is_shared(buf));
7104 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
7106 zio = zio_write(pio, spa, txg, bp,
7107 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
7108 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
7109 (children_ready != NULL) ? arc_write_children_ready : NULL,
7110 arc_write_physdone, arc_write_done, callback,
7111 priority, zio_flags, zb);
7117 arc_tempreserve_clear(uint64_t reserve)
7119 atomic_add_64(&arc_tempreserve, -reserve);
7120 ASSERT((int64_t)arc_tempreserve >= 0);
7124 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
7130 reserve > arc_c/4 &&
7131 reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT))
7132 arc_c = MIN(arc_c_max, reserve * 4);
7135 * Throttle when the calculated memory footprint for the TXG
7136 * exceeds the target ARC size.
7138 if (reserve > arc_c) {
7139 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
7140 return (SET_ERROR(ERESTART));
7144 * Don't count loaned bufs as in flight dirty data to prevent long
7145 * network delays from blocking transactions that are ready to be
7146 * assigned to a txg.
7149 /* assert that it has not wrapped around */
7150 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
7152 anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) -
7153 arc_loaned_bytes), 0);
7156 * Writes will, almost always, require additional memory allocations
7157 * in order to compress/encrypt/etc the data. We therefore need to
7158 * make sure that there is sufficient available memory for this.
7160 error = arc_memory_throttle(spa, reserve, txg);
7165 * Throttle writes when the amount of dirty data in the cache
7166 * gets too large. We try to keep the cache less than half full
7167 * of dirty blocks so that our sync times don't grow too large.
7169 * In the case of one pool being built on another pool, we want
7170 * to make sure we don't end up throttling the lower (backing)
7171 * pool when the upper pool is the majority contributor to dirty
7172 * data. To insure we make forward progress during throttling, we
7173 * also check the current pool's net dirty data and only throttle
7174 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
7175 * data in the cache.
7177 * Note: if two requests come in concurrently, we might let them
7178 * both succeed, when one of them should fail. Not a huge deal.
7180 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
7181 uint64_t spa_dirty_anon = spa_dirty_data(spa);
7182 uint64_t rarc_c = arc_warm ? arc_c : arc_c_max;
7183 if (total_dirty > rarc_c * zfs_arc_dirty_limit_percent / 100 &&
7184 anon_size > rarc_c * zfs_arc_anon_limit_percent / 100 &&
7185 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
7187 uint64_t meta_esize = zfs_refcount_count(
7188 &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7189 uint64_t data_esize =
7190 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7191 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
7192 "anon_data=%lluK tempreserve=%lluK rarc_c=%lluK\n",
7193 (u_longlong_t)arc_tempreserve >> 10,
7194 (u_longlong_t)meta_esize >> 10,
7195 (u_longlong_t)data_esize >> 10,
7196 (u_longlong_t)reserve >> 10,
7197 (u_longlong_t)rarc_c >> 10);
7199 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
7200 return (SET_ERROR(ERESTART));
7202 atomic_add_64(&arc_tempreserve, reserve);
7207 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
7208 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
7210 size->value.ui64 = zfs_refcount_count(&state->arcs_size);
7211 evict_data->value.ui64 =
7212 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
7213 evict_metadata->value.ui64 =
7214 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
7218 arc_kstat_update(kstat_t *ksp, int rw)
7220 arc_stats_t *as = ksp->ks_data;
7222 if (rw == KSTAT_WRITE)
7223 return (SET_ERROR(EACCES));
7225 as->arcstat_hits.value.ui64 =
7226 wmsum_value(&arc_sums.arcstat_hits);
7227 as->arcstat_misses.value.ui64 =
7228 wmsum_value(&arc_sums.arcstat_misses);
7229 as->arcstat_demand_data_hits.value.ui64 =
7230 wmsum_value(&arc_sums.arcstat_demand_data_hits);
7231 as->arcstat_demand_data_misses.value.ui64 =
7232 wmsum_value(&arc_sums.arcstat_demand_data_misses);
7233 as->arcstat_demand_metadata_hits.value.ui64 =
7234 wmsum_value(&arc_sums.arcstat_demand_metadata_hits);
7235 as->arcstat_demand_metadata_misses.value.ui64 =
7236 wmsum_value(&arc_sums.arcstat_demand_metadata_misses);
7237 as->arcstat_prefetch_data_hits.value.ui64 =
7238 wmsum_value(&arc_sums.arcstat_prefetch_data_hits);
7239 as->arcstat_prefetch_data_misses.value.ui64 =
7240 wmsum_value(&arc_sums.arcstat_prefetch_data_misses);
7241 as->arcstat_prefetch_metadata_hits.value.ui64 =
7242 wmsum_value(&arc_sums.arcstat_prefetch_metadata_hits);
7243 as->arcstat_prefetch_metadata_misses.value.ui64 =
7244 wmsum_value(&arc_sums.arcstat_prefetch_metadata_misses);
7245 as->arcstat_mru_hits.value.ui64 =
7246 wmsum_value(&arc_sums.arcstat_mru_hits);
7247 as->arcstat_mru_ghost_hits.value.ui64 =
7248 wmsum_value(&arc_sums.arcstat_mru_ghost_hits);
7249 as->arcstat_mfu_hits.value.ui64 =
7250 wmsum_value(&arc_sums.arcstat_mfu_hits);
7251 as->arcstat_mfu_ghost_hits.value.ui64 =
7252 wmsum_value(&arc_sums.arcstat_mfu_ghost_hits);
7253 as->arcstat_deleted.value.ui64 =
7254 wmsum_value(&arc_sums.arcstat_deleted);
7255 as->arcstat_mutex_miss.value.ui64 =
7256 wmsum_value(&arc_sums.arcstat_mutex_miss);
7257 as->arcstat_access_skip.value.ui64 =
7258 wmsum_value(&arc_sums.arcstat_access_skip);
7259 as->arcstat_evict_skip.value.ui64 =
7260 wmsum_value(&arc_sums.arcstat_evict_skip);
7261 as->arcstat_evict_not_enough.value.ui64 =
7262 wmsum_value(&arc_sums.arcstat_evict_not_enough);
7263 as->arcstat_evict_l2_cached.value.ui64 =
7264 wmsum_value(&arc_sums.arcstat_evict_l2_cached);
7265 as->arcstat_evict_l2_eligible.value.ui64 =
7266 wmsum_value(&arc_sums.arcstat_evict_l2_eligible);
7267 as->arcstat_evict_l2_eligible_mfu.value.ui64 =
7268 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mfu);
7269 as->arcstat_evict_l2_eligible_mru.value.ui64 =
7270 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mru);
7271 as->arcstat_evict_l2_ineligible.value.ui64 =
7272 wmsum_value(&arc_sums.arcstat_evict_l2_ineligible);
7273 as->arcstat_evict_l2_skip.value.ui64 =
7274 wmsum_value(&arc_sums.arcstat_evict_l2_skip);
7275 as->arcstat_hash_collisions.value.ui64 =
7276 wmsum_value(&arc_sums.arcstat_hash_collisions);
7277 as->arcstat_hash_chains.value.ui64 =
7278 wmsum_value(&arc_sums.arcstat_hash_chains);
7279 as->arcstat_size.value.ui64 =
7280 aggsum_value(&arc_sums.arcstat_size);
7281 as->arcstat_compressed_size.value.ui64 =
7282 wmsum_value(&arc_sums.arcstat_compressed_size);
7283 as->arcstat_uncompressed_size.value.ui64 =
7284 wmsum_value(&arc_sums.arcstat_uncompressed_size);
7285 as->arcstat_overhead_size.value.ui64 =
7286 wmsum_value(&arc_sums.arcstat_overhead_size);
7287 as->arcstat_hdr_size.value.ui64 =
7288 wmsum_value(&arc_sums.arcstat_hdr_size);
7289 as->arcstat_data_size.value.ui64 =
7290 wmsum_value(&arc_sums.arcstat_data_size);
7291 as->arcstat_metadata_size.value.ui64 =
7292 wmsum_value(&arc_sums.arcstat_metadata_size);
7293 as->arcstat_dbuf_size.value.ui64 =
7294 wmsum_value(&arc_sums.arcstat_dbuf_size);
7295 #if defined(COMPAT_FREEBSD11)
7296 as->arcstat_other_size.value.ui64 =
7297 wmsum_value(&arc_sums.arcstat_bonus_size) +
7298 aggsum_value(&arc_sums.arcstat_dnode_size) +
7299 wmsum_value(&arc_sums.arcstat_dbuf_size);
7302 arc_kstat_update_state(arc_anon,
7303 &as->arcstat_anon_size,
7304 &as->arcstat_anon_evictable_data,
7305 &as->arcstat_anon_evictable_metadata);
7306 arc_kstat_update_state(arc_mru,
7307 &as->arcstat_mru_size,
7308 &as->arcstat_mru_evictable_data,
7309 &as->arcstat_mru_evictable_metadata);
7310 arc_kstat_update_state(arc_mru_ghost,
7311 &as->arcstat_mru_ghost_size,
7312 &as->arcstat_mru_ghost_evictable_data,
7313 &as->arcstat_mru_ghost_evictable_metadata);
7314 arc_kstat_update_state(arc_mfu,
7315 &as->arcstat_mfu_size,
7316 &as->arcstat_mfu_evictable_data,
7317 &as->arcstat_mfu_evictable_metadata);
7318 arc_kstat_update_state(arc_mfu_ghost,
7319 &as->arcstat_mfu_ghost_size,
7320 &as->arcstat_mfu_ghost_evictable_data,
7321 &as->arcstat_mfu_ghost_evictable_metadata);
7323 as->arcstat_dnode_size.value.ui64 =
7324 aggsum_value(&arc_sums.arcstat_dnode_size);
7325 as->arcstat_bonus_size.value.ui64 =
7326 wmsum_value(&arc_sums.arcstat_bonus_size);
7327 as->arcstat_l2_hits.value.ui64 =
7328 wmsum_value(&arc_sums.arcstat_l2_hits);
7329 as->arcstat_l2_misses.value.ui64 =
7330 wmsum_value(&arc_sums.arcstat_l2_misses);
7331 as->arcstat_l2_prefetch_asize.value.ui64 =
7332 wmsum_value(&arc_sums.arcstat_l2_prefetch_asize);
7333 as->arcstat_l2_mru_asize.value.ui64 =
7334 wmsum_value(&arc_sums.arcstat_l2_mru_asize);
7335 as->arcstat_l2_mfu_asize.value.ui64 =
7336 wmsum_value(&arc_sums.arcstat_l2_mfu_asize);
7337 as->arcstat_l2_bufc_data_asize.value.ui64 =
7338 wmsum_value(&arc_sums.arcstat_l2_bufc_data_asize);
7339 as->arcstat_l2_bufc_metadata_asize.value.ui64 =
7340 wmsum_value(&arc_sums.arcstat_l2_bufc_metadata_asize);
7341 as->arcstat_l2_feeds.value.ui64 =
7342 wmsum_value(&arc_sums.arcstat_l2_feeds);
7343 as->arcstat_l2_rw_clash.value.ui64 =
7344 wmsum_value(&arc_sums.arcstat_l2_rw_clash);
7345 as->arcstat_l2_read_bytes.value.ui64 =
7346 wmsum_value(&arc_sums.arcstat_l2_read_bytes);
7347 as->arcstat_l2_write_bytes.value.ui64 =
7348 wmsum_value(&arc_sums.arcstat_l2_write_bytes);
7349 as->arcstat_l2_writes_sent.value.ui64 =
7350 wmsum_value(&arc_sums.arcstat_l2_writes_sent);
7351 as->arcstat_l2_writes_done.value.ui64 =
7352 wmsum_value(&arc_sums.arcstat_l2_writes_done);
7353 as->arcstat_l2_writes_error.value.ui64 =
7354 wmsum_value(&arc_sums.arcstat_l2_writes_error);
7355 as->arcstat_l2_writes_lock_retry.value.ui64 =
7356 wmsum_value(&arc_sums.arcstat_l2_writes_lock_retry);
7357 as->arcstat_l2_evict_lock_retry.value.ui64 =
7358 wmsum_value(&arc_sums.arcstat_l2_evict_lock_retry);
7359 as->arcstat_l2_evict_reading.value.ui64 =
7360 wmsum_value(&arc_sums.arcstat_l2_evict_reading);
7361 as->arcstat_l2_evict_l1cached.value.ui64 =
7362 wmsum_value(&arc_sums.arcstat_l2_evict_l1cached);
7363 as->arcstat_l2_free_on_write.value.ui64 =
7364 wmsum_value(&arc_sums.arcstat_l2_free_on_write);
7365 as->arcstat_l2_abort_lowmem.value.ui64 =
7366 wmsum_value(&arc_sums.arcstat_l2_abort_lowmem);
7367 as->arcstat_l2_cksum_bad.value.ui64 =
7368 wmsum_value(&arc_sums.arcstat_l2_cksum_bad);
7369 as->arcstat_l2_io_error.value.ui64 =
7370 wmsum_value(&arc_sums.arcstat_l2_io_error);
7371 as->arcstat_l2_lsize.value.ui64 =
7372 wmsum_value(&arc_sums.arcstat_l2_lsize);
7373 as->arcstat_l2_psize.value.ui64 =
7374 wmsum_value(&arc_sums.arcstat_l2_psize);
7375 as->arcstat_l2_hdr_size.value.ui64 =
7376 aggsum_value(&arc_sums.arcstat_l2_hdr_size);
7377 as->arcstat_l2_log_blk_writes.value.ui64 =
7378 wmsum_value(&arc_sums.arcstat_l2_log_blk_writes);
7379 as->arcstat_l2_log_blk_asize.value.ui64 =
7380 wmsum_value(&arc_sums.arcstat_l2_log_blk_asize);
7381 as->arcstat_l2_log_blk_count.value.ui64 =
7382 wmsum_value(&arc_sums.arcstat_l2_log_blk_count);
7383 as->arcstat_l2_rebuild_success.value.ui64 =
7384 wmsum_value(&arc_sums.arcstat_l2_rebuild_success);
7385 as->arcstat_l2_rebuild_abort_unsupported.value.ui64 =
7386 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7387 as->arcstat_l2_rebuild_abort_io_errors.value.ui64 =
7388 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7389 as->arcstat_l2_rebuild_abort_dh_errors.value.ui64 =
7390 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7391 as->arcstat_l2_rebuild_abort_cksum_lb_errors.value.ui64 =
7392 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7393 as->arcstat_l2_rebuild_abort_lowmem.value.ui64 =
7394 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7395 as->arcstat_l2_rebuild_size.value.ui64 =
7396 wmsum_value(&arc_sums.arcstat_l2_rebuild_size);
7397 as->arcstat_l2_rebuild_asize.value.ui64 =
7398 wmsum_value(&arc_sums.arcstat_l2_rebuild_asize);
7399 as->arcstat_l2_rebuild_bufs.value.ui64 =
7400 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs);
7401 as->arcstat_l2_rebuild_bufs_precached.value.ui64 =
7402 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7403 as->arcstat_l2_rebuild_log_blks.value.ui64 =
7404 wmsum_value(&arc_sums.arcstat_l2_rebuild_log_blks);
7405 as->arcstat_memory_throttle_count.value.ui64 =
7406 wmsum_value(&arc_sums.arcstat_memory_throttle_count);
7407 as->arcstat_memory_direct_count.value.ui64 =
7408 wmsum_value(&arc_sums.arcstat_memory_direct_count);
7409 as->arcstat_memory_indirect_count.value.ui64 =
7410 wmsum_value(&arc_sums.arcstat_memory_indirect_count);
7412 as->arcstat_memory_all_bytes.value.ui64 =
7414 as->arcstat_memory_free_bytes.value.ui64 =
7416 as->arcstat_memory_available_bytes.value.i64 =
7417 arc_available_memory();
7419 as->arcstat_prune.value.ui64 =
7420 wmsum_value(&arc_sums.arcstat_prune);
7421 as->arcstat_meta_used.value.ui64 =
7422 aggsum_value(&arc_sums.arcstat_meta_used);
7423 as->arcstat_async_upgrade_sync.value.ui64 =
7424 wmsum_value(&arc_sums.arcstat_async_upgrade_sync);
7425 as->arcstat_demand_hit_predictive_prefetch.value.ui64 =
7426 wmsum_value(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7427 as->arcstat_demand_hit_prescient_prefetch.value.ui64 =
7428 wmsum_value(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7429 as->arcstat_raw_size.value.ui64 =
7430 wmsum_value(&arc_sums.arcstat_raw_size);
7431 as->arcstat_cached_only_in_progress.value.ui64 =
7432 wmsum_value(&arc_sums.arcstat_cached_only_in_progress);
7433 as->arcstat_abd_chunk_waste_size.value.ui64 =
7434 wmsum_value(&arc_sums.arcstat_abd_chunk_waste_size);
7440 * This function *must* return indices evenly distributed between all
7441 * sublists of the multilist. This is needed due to how the ARC eviction
7442 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7443 * distributed between all sublists and uses this assumption when
7444 * deciding which sublist to evict from and how much to evict from it.
7447 arc_state_multilist_index_func(multilist_t *ml, void *obj)
7449 arc_buf_hdr_t *hdr = obj;
7452 * We rely on b_dva to generate evenly distributed index
7453 * numbers using buf_hash below. So, as an added precaution,
7454 * let's make sure we never add empty buffers to the arc lists.
7456 ASSERT(!HDR_EMPTY(hdr));
7459 * The assumption here, is the hash value for a given
7460 * arc_buf_hdr_t will remain constant throughout its lifetime
7461 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7462 * Thus, we don't need to store the header's sublist index
7463 * on insertion, as this index can be recalculated on removal.
7465 * Also, the low order bits of the hash value are thought to be
7466 * distributed evenly. Otherwise, in the case that the multilist
7467 * has a power of two number of sublists, each sublists' usage
7468 * would not be evenly distributed. In this context full 64bit
7469 * division would be a waste of time, so limit it to 32 bits.
7471 return ((unsigned int)buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
7472 multilist_get_num_sublists(ml));
7476 arc_state_l2c_multilist_index_func(multilist_t *ml, void *obj)
7478 panic("Header %p insert into arc_l2c_only %p", obj, ml);
7481 #define WARN_IF_TUNING_IGNORED(tuning, value, do_warn) do { \
7482 if ((do_warn) && (tuning) && ((tuning) != (value))) { \
7484 "ignoring tunable %s (using %llu instead)", \
7485 (#tuning), (u_longlong_t)(value)); \
7490 * Called during module initialization and periodically thereafter to
7491 * apply reasonable changes to the exposed performance tunings. Can also be
7492 * called explicitly by param_set_arc_*() functions when ARC tunables are
7493 * updated manually. Non-zero zfs_* values which differ from the currently set
7494 * values will be applied.
7497 arc_tuning_update(boolean_t verbose)
7499 uint64_t allmem = arc_all_memory();
7500 unsigned long limit;
7502 /* Valid range: 32M - <arc_c_max> */
7503 if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) &&
7504 (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) &&
7505 (zfs_arc_min <= arc_c_max)) {
7506 arc_c_min = zfs_arc_min;
7507 arc_c = MAX(arc_c, arc_c_min);
7509 WARN_IF_TUNING_IGNORED(zfs_arc_min, arc_c_min, verbose);
7511 /* Valid range: 64M - <all physical memory> */
7512 if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) &&
7513 (zfs_arc_max >= MIN_ARC_MAX) && (zfs_arc_max < allmem) &&
7514 (zfs_arc_max > arc_c_min)) {
7515 arc_c_max = zfs_arc_max;
7516 arc_c = MIN(arc_c, arc_c_max);
7517 arc_p = (arc_c >> 1);
7518 if (arc_meta_limit > arc_c_max)
7519 arc_meta_limit = arc_c_max;
7520 if (arc_dnode_size_limit > arc_meta_limit)
7521 arc_dnode_size_limit = arc_meta_limit;
7523 WARN_IF_TUNING_IGNORED(zfs_arc_max, arc_c_max, verbose);
7525 /* Valid range: 16M - <arc_c_max> */
7526 if ((zfs_arc_meta_min) && (zfs_arc_meta_min != arc_meta_min) &&
7527 (zfs_arc_meta_min >= 1ULL << SPA_MAXBLOCKSHIFT) &&
7528 (zfs_arc_meta_min <= arc_c_max)) {
7529 arc_meta_min = zfs_arc_meta_min;
7530 if (arc_meta_limit < arc_meta_min)
7531 arc_meta_limit = arc_meta_min;
7532 if (arc_dnode_size_limit < arc_meta_min)
7533 arc_dnode_size_limit = arc_meta_min;
7535 WARN_IF_TUNING_IGNORED(zfs_arc_meta_min, arc_meta_min, verbose);
7537 /* Valid range: <arc_meta_min> - <arc_c_max> */
7538 limit = zfs_arc_meta_limit ? zfs_arc_meta_limit :
7539 MIN(zfs_arc_meta_limit_percent, 100) * arc_c_max / 100;
7540 if ((limit != arc_meta_limit) &&
7541 (limit >= arc_meta_min) &&
7542 (limit <= arc_c_max))
7543 arc_meta_limit = limit;
7544 WARN_IF_TUNING_IGNORED(zfs_arc_meta_limit, arc_meta_limit, verbose);
7546 /* Valid range: <arc_meta_min> - <arc_meta_limit> */
7547 limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit :
7548 MIN(zfs_arc_dnode_limit_percent, 100) * arc_meta_limit / 100;
7549 if ((limit != arc_dnode_size_limit) &&
7550 (limit >= arc_meta_min) &&
7551 (limit <= arc_meta_limit))
7552 arc_dnode_size_limit = limit;
7553 WARN_IF_TUNING_IGNORED(zfs_arc_dnode_limit, arc_dnode_size_limit,
7556 /* Valid range: 1 - N */
7557 if (zfs_arc_grow_retry)
7558 arc_grow_retry = zfs_arc_grow_retry;
7560 /* Valid range: 1 - N */
7561 if (zfs_arc_shrink_shift) {
7562 arc_shrink_shift = zfs_arc_shrink_shift;
7563 arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1);
7566 /* Valid range: 1 - N */
7567 if (zfs_arc_p_min_shift)
7568 arc_p_min_shift = zfs_arc_p_min_shift;
7570 /* Valid range: 1 - N ms */
7571 if (zfs_arc_min_prefetch_ms)
7572 arc_min_prefetch_ms = zfs_arc_min_prefetch_ms;
7574 /* Valid range: 1 - N ms */
7575 if (zfs_arc_min_prescient_prefetch_ms) {
7576 arc_min_prescient_prefetch_ms =
7577 zfs_arc_min_prescient_prefetch_ms;
7580 /* Valid range: 0 - 100 */
7581 if ((zfs_arc_lotsfree_percent >= 0) &&
7582 (zfs_arc_lotsfree_percent <= 100))
7583 arc_lotsfree_percent = zfs_arc_lotsfree_percent;
7584 WARN_IF_TUNING_IGNORED(zfs_arc_lotsfree_percent, arc_lotsfree_percent,
7587 /* Valid range: 0 - <all physical memory> */
7588 if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free))
7589 arc_sys_free = MIN(MAX(zfs_arc_sys_free, 0), allmem);
7590 WARN_IF_TUNING_IGNORED(zfs_arc_sys_free, arc_sys_free, verbose);
7594 arc_state_init(void)
7596 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
7597 sizeof (arc_buf_hdr_t),
7598 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7599 arc_state_multilist_index_func);
7600 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
7601 sizeof (arc_buf_hdr_t),
7602 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7603 arc_state_multilist_index_func);
7604 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
7605 sizeof (arc_buf_hdr_t),
7606 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7607 arc_state_multilist_index_func);
7608 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
7609 sizeof (arc_buf_hdr_t),
7610 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7611 arc_state_multilist_index_func);
7612 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
7613 sizeof (arc_buf_hdr_t),
7614 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7615 arc_state_multilist_index_func);
7616 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
7617 sizeof (arc_buf_hdr_t),
7618 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7619 arc_state_multilist_index_func);
7620 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
7621 sizeof (arc_buf_hdr_t),
7622 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7623 arc_state_multilist_index_func);
7624 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
7625 sizeof (arc_buf_hdr_t),
7626 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7627 arc_state_multilist_index_func);
7629 * L2 headers should never be on the L2 state list since they don't
7630 * have L1 headers allocated. Special index function asserts that.
7632 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
7633 sizeof (arc_buf_hdr_t),
7634 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7635 arc_state_l2c_multilist_index_func);
7636 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
7637 sizeof (arc_buf_hdr_t),
7638 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7639 arc_state_l2c_multilist_index_func);
7641 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7642 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7643 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7644 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7645 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7646 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7647 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7648 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7649 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7650 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7651 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7652 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7654 zfs_refcount_create(&arc_anon->arcs_size);
7655 zfs_refcount_create(&arc_mru->arcs_size);
7656 zfs_refcount_create(&arc_mru_ghost->arcs_size);
7657 zfs_refcount_create(&arc_mfu->arcs_size);
7658 zfs_refcount_create(&arc_mfu_ghost->arcs_size);
7659 zfs_refcount_create(&arc_l2c_only->arcs_size);
7661 wmsum_init(&arc_sums.arcstat_hits, 0);
7662 wmsum_init(&arc_sums.arcstat_misses, 0);
7663 wmsum_init(&arc_sums.arcstat_demand_data_hits, 0);
7664 wmsum_init(&arc_sums.arcstat_demand_data_misses, 0);
7665 wmsum_init(&arc_sums.arcstat_demand_metadata_hits, 0);
7666 wmsum_init(&arc_sums.arcstat_demand_metadata_misses, 0);
7667 wmsum_init(&arc_sums.arcstat_prefetch_data_hits, 0);
7668 wmsum_init(&arc_sums.arcstat_prefetch_data_misses, 0);
7669 wmsum_init(&arc_sums.arcstat_prefetch_metadata_hits, 0);
7670 wmsum_init(&arc_sums.arcstat_prefetch_metadata_misses, 0);
7671 wmsum_init(&arc_sums.arcstat_mru_hits, 0);
7672 wmsum_init(&arc_sums.arcstat_mru_ghost_hits, 0);
7673 wmsum_init(&arc_sums.arcstat_mfu_hits, 0);
7674 wmsum_init(&arc_sums.arcstat_mfu_ghost_hits, 0);
7675 wmsum_init(&arc_sums.arcstat_deleted, 0);
7676 wmsum_init(&arc_sums.arcstat_mutex_miss, 0);
7677 wmsum_init(&arc_sums.arcstat_access_skip, 0);
7678 wmsum_init(&arc_sums.arcstat_evict_skip, 0);
7679 wmsum_init(&arc_sums.arcstat_evict_not_enough, 0);
7680 wmsum_init(&arc_sums.arcstat_evict_l2_cached, 0);
7681 wmsum_init(&arc_sums.arcstat_evict_l2_eligible, 0);
7682 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mfu, 0);
7683 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mru, 0);
7684 wmsum_init(&arc_sums.arcstat_evict_l2_ineligible, 0);
7685 wmsum_init(&arc_sums.arcstat_evict_l2_skip, 0);
7686 wmsum_init(&arc_sums.arcstat_hash_collisions, 0);
7687 wmsum_init(&arc_sums.arcstat_hash_chains, 0);
7688 aggsum_init(&arc_sums.arcstat_size, 0);
7689 wmsum_init(&arc_sums.arcstat_compressed_size, 0);
7690 wmsum_init(&arc_sums.arcstat_uncompressed_size, 0);
7691 wmsum_init(&arc_sums.arcstat_overhead_size, 0);
7692 wmsum_init(&arc_sums.arcstat_hdr_size, 0);
7693 wmsum_init(&arc_sums.arcstat_data_size, 0);
7694 wmsum_init(&arc_sums.arcstat_metadata_size, 0);
7695 wmsum_init(&arc_sums.arcstat_dbuf_size, 0);
7696 aggsum_init(&arc_sums.arcstat_dnode_size, 0);
7697 wmsum_init(&arc_sums.arcstat_bonus_size, 0);
7698 wmsum_init(&arc_sums.arcstat_l2_hits, 0);
7699 wmsum_init(&arc_sums.arcstat_l2_misses, 0);
7700 wmsum_init(&arc_sums.arcstat_l2_prefetch_asize, 0);
7701 wmsum_init(&arc_sums.arcstat_l2_mru_asize, 0);
7702 wmsum_init(&arc_sums.arcstat_l2_mfu_asize, 0);
7703 wmsum_init(&arc_sums.arcstat_l2_bufc_data_asize, 0);
7704 wmsum_init(&arc_sums.arcstat_l2_bufc_metadata_asize, 0);
7705 wmsum_init(&arc_sums.arcstat_l2_feeds, 0);
7706 wmsum_init(&arc_sums.arcstat_l2_rw_clash, 0);
7707 wmsum_init(&arc_sums.arcstat_l2_read_bytes, 0);
7708 wmsum_init(&arc_sums.arcstat_l2_write_bytes, 0);
7709 wmsum_init(&arc_sums.arcstat_l2_writes_sent, 0);
7710 wmsum_init(&arc_sums.arcstat_l2_writes_done, 0);
7711 wmsum_init(&arc_sums.arcstat_l2_writes_error, 0);
7712 wmsum_init(&arc_sums.arcstat_l2_writes_lock_retry, 0);
7713 wmsum_init(&arc_sums.arcstat_l2_evict_lock_retry, 0);
7714 wmsum_init(&arc_sums.arcstat_l2_evict_reading, 0);
7715 wmsum_init(&arc_sums.arcstat_l2_evict_l1cached, 0);
7716 wmsum_init(&arc_sums.arcstat_l2_free_on_write, 0);
7717 wmsum_init(&arc_sums.arcstat_l2_abort_lowmem, 0);
7718 wmsum_init(&arc_sums.arcstat_l2_cksum_bad, 0);
7719 wmsum_init(&arc_sums.arcstat_l2_io_error, 0);
7720 wmsum_init(&arc_sums.arcstat_l2_lsize, 0);
7721 wmsum_init(&arc_sums.arcstat_l2_psize, 0);
7722 aggsum_init(&arc_sums.arcstat_l2_hdr_size, 0);
7723 wmsum_init(&arc_sums.arcstat_l2_log_blk_writes, 0);
7724 wmsum_init(&arc_sums.arcstat_l2_log_blk_asize, 0);
7725 wmsum_init(&arc_sums.arcstat_l2_log_blk_count, 0);
7726 wmsum_init(&arc_sums.arcstat_l2_rebuild_success, 0);
7727 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_unsupported, 0);
7728 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_io_errors, 0);
7729 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_dh_errors, 0);
7730 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors, 0);
7731 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_lowmem, 0);
7732 wmsum_init(&arc_sums.arcstat_l2_rebuild_size, 0);
7733 wmsum_init(&arc_sums.arcstat_l2_rebuild_asize, 0);
7734 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs, 0);
7735 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs_precached, 0);
7736 wmsum_init(&arc_sums.arcstat_l2_rebuild_log_blks, 0);
7737 wmsum_init(&arc_sums.arcstat_memory_throttle_count, 0);
7738 wmsum_init(&arc_sums.arcstat_memory_direct_count, 0);
7739 wmsum_init(&arc_sums.arcstat_memory_indirect_count, 0);
7740 wmsum_init(&arc_sums.arcstat_prune, 0);
7741 aggsum_init(&arc_sums.arcstat_meta_used, 0);
7742 wmsum_init(&arc_sums.arcstat_async_upgrade_sync, 0);
7743 wmsum_init(&arc_sums.arcstat_demand_hit_predictive_prefetch, 0);
7744 wmsum_init(&arc_sums.arcstat_demand_hit_prescient_prefetch, 0);
7745 wmsum_init(&arc_sums.arcstat_raw_size, 0);
7746 wmsum_init(&arc_sums.arcstat_cached_only_in_progress, 0);
7747 wmsum_init(&arc_sums.arcstat_abd_chunk_waste_size, 0);
7749 arc_anon->arcs_state = ARC_STATE_ANON;
7750 arc_mru->arcs_state = ARC_STATE_MRU;
7751 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7752 arc_mfu->arcs_state = ARC_STATE_MFU;
7753 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7754 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7758 arc_state_fini(void)
7760 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7761 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7762 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7763 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7764 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7765 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7766 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7767 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7768 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7769 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7770 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7771 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7773 zfs_refcount_destroy(&arc_anon->arcs_size);
7774 zfs_refcount_destroy(&arc_mru->arcs_size);
7775 zfs_refcount_destroy(&arc_mru_ghost->arcs_size);
7776 zfs_refcount_destroy(&arc_mfu->arcs_size);
7777 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
7778 zfs_refcount_destroy(&arc_l2c_only->arcs_size);
7780 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
7781 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7782 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7783 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7784 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
7785 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7786 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
7787 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7788 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7789 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7791 wmsum_fini(&arc_sums.arcstat_hits);
7792 wmsum_fini(&arc_sums.arcstat_misses);
7793 wmsum_fini(&arc_sums.arcstat_demand_data_hits);
7794 wmsum_fini(&arc_sums.arcstat_demand_data_misses);
7795 wmsum_fini(&arc_sums.arcstat_demand_metadata_hits);
7796 wmsum_fini(&arc_sums.arcstat_demand_metadata_misses);
7797 wmsum_fini(&arc_sums.arcstat_prefetch_data_hits);
7798 wmsum_fini(&arc_sums.arcstat_prefetch_data_misses);
7799 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_hits);
7800 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_misses);
7801 wmsum_fini(&arc_sums.arcstat_mru_hits);
7802 wmsum_fini(&arc_sums.arcstat_mru_ghost_hits);
7803 wmsum_fini(&arc_sums.arcstat_mfu_hits);
7804 wmsum_fini(&arc_sums.arcstat_mfu_ghost_hits);
7805 wmsum_fini(&arc_sums.arcstat_deleted);
7806 wmsum_fini(&arc_sums.arcstat_mutex_miss);
7807 wmsum_fini(&arc_sums.arcstat_access_skip);
7808 wmsum_fini(&arc_sums.arcstat_evict_skip);
7809 wmsum_fini(&arc_sums.arcstat_evict_not_enough);
7810 wmsum_fini(&arc_sums.arcstat_evict_l2_cached);
7811 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible);
7812 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mfu);
7813 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mru);
7814 wmsum_fini(&arc_sums.arcstat_evict_l2_ineligible);
7815 wmsum_fini(&arc_sums.arcstat_evict_l2_skip);
7816 wmsum_fini(&arc_sums.arcstat_hash_collisions);
7817 wmsum_fini(&arc_sums.arcstat_hash_chains);
7818 aggsum_fini(&arc_sums.arcstat_size);
7819 wmsum_fini(&arc_sums.arcstat_compressed_size);
7820 wmsum_fini(&arc_sums.arcstat_uncompressed_size);
7821 wmsum_fini(&arc_sums.arcstat_overhead_size);
7822 wmsum_fini(&arc_sums.arcstat_hdr_size);
7823 wmsum_fini(&arc_sums.arcstat_data_size);
7824 wmsum_fini(&arc_sums.arcstat_metadata_size);
7825 wmsum_fini(&arc_sums.arcstat_dbuf_size);
7826 aggsum_fini(&arc_sums.arcstat_dnode_size);
7827 wmsum_fini(&arc_sums.arcstat_bonus_size);
7828 wmsum_fini(&arc_sums.arcstat_l2_hits);
7829 wmsum_fini(&arc_sums.arcstat_l2_misses);
7830 wmsum_fini(&arc_sums.arcstat_l2_prefetch_asize);
7831 wmsum_fini(&arc_sums.arcstat_l2_mru_asize);
7832 wmsum_fini(&arc_sums.arcstat_l2_mfu_asize);
7833 wmsum_fini(&arc_sums.arcstat_l2_bufc_data_asize);
7834 wmsum_fini(&arc_sums.arcstat_l2_bufc_metadata_asize);
7835 wmsum_fini(&arc_sums.arcstat_l2_feeds);
7836 wmsum_fini(&arc_sums.arcstat_l2_rw_clash);
7837 wmsum_fini(&arc_sums.arcstat_l2_read_bytes);
7838 wmsum_fini(&arc_sums.arcstat_l2_write_bytes);
7839 wmsum_fini(&arc_sums.arcstat_l2_writes_sent);
7840 wmsum_fini(&arc_sums.arcstat_l2_writes_done);
7841 wmsum_fini(&arc_sums.arcstat_l2_writes_error);
7842 wmsum_fini(&arc_sums.arcstat_l2_writes_lock_retry);
7843 wmsum_fini(&arc_sums.arcstat_l2_evict_lock_retry);
7844 wmsum_fini(&arc_sums.arcstat_l2_evict_reading);
7845 wmsum_fini(&arc_sums.arcstat_l2_evict_l1cached);
7846 wmsum_fini(&arc_sums.arcstat_l2_free_on_write);
7847 wmsum_fini(&arc_sums.arcstat_l2_abort_lowmem);
7848 wmsum_fini(&arc_sums.arcstat_l2_cksum_bad);
7849 wmsum_fini(&arc_sums.arcstat_l2_io_error);
7850 wmsum_fini(&arc_sums.arcstat_l2_lsize);
7851 wmsum_fini(&arc_sums.arcstat_l2_psize);
7852 aggsum_fini(&arc_sums.arcstat_l2_hdr_size);
7853 wmsum_fini(&arc_sums.arcstat_l2_log_blk_writes);
7854 wmsum_fini(&arc_sums.arcstat_l2_log_blk_asize);
7855 wmsum_fini(&arc_sums.arcstat_l2_log_blk_count);
7856 wmsum_fini(&arc_sums.arcstat_l2_rebuild_success);
7857 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7858 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7859 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7860 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7861 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7862 wmsum_fini(&arc_sums.arcstat_l2_rebuild_size);
7863 wmsum_fini(&arc_sums.arcstat_l2_rebuild_asize);
7864 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs);
7865 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7866 wmsum_fini(&arc_sums.arcstat_l2_rebuild_log_blks);
7867 wmsum_fini(&arc_sums.arcstat_memory_throttle_count);
7868 wmsum_fini(&arc_sums.arcstat_memory_direct_count);
7869 wmsum_fini(&arc_sums.arcstat_memory_indirect_count);
7870 wmsum_fini(&arc_sums.arcstat_prune);
7871 aggsum_fini(&arc_sums.arcstat_meta_used);
7872 wmsum_fini(&arc_sums.arcstat_async_upgrade_sync);
7873 wmsum_fini(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7874 wmsum_fini(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7875 wmsum_fini(&arc_sums.arcstat_raw_size);
7876 wmsum_fini(&arc_sums.arcstat_cached_only_in_progress);
7877 wmsum_fini(&arc_sums.arcstat_abd_chunk_waste_size);
7881 arc_target_bytes(void)
7887 arc_set_limits(uint64_t allmem)
7889 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more. */
7890 arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT);
7892 /* How to set default max varies by platform. */
7893 arc_c_max = arc_default_max(arc_c_min, allmem);
7898 uint64_t percent, allmem = arc_all_memory();
7899 mutex_init(&arc_evict_lock, NULL, MUTEX_DEFAULT, NULL);
7900 list_create(&arc_evict_waiters, sizeof (arc_evict_waiter_t),
7901 offsetof(arc_evict_waiter_t, aew_node));
7903 arc_min_prefetch_ms = 1000;
7904 arc_min_prescient_prefetch_ms = 6000;
7906 #if defined(_KERNEL)
7910 arc_set_limits(allmem);
7914 * If zfs_arc_max is non-zero at init, meaning it was set in the kernel
7915 * environment before the module was loaded, don't block setting the
7916 * maximum because it is less than arc_c_min, instead, reset arc_c_min
7918 * zfs_arc_min will be handled by arc_tuning_update().
7920 if (zfs_arc_max != 0 && zfs_arc_max >= MIN_ARC_MAX &&
7921 zfs_arc_max < allmem) {
7922 arc_c_max = zfs_arc_max;
7923 if (arc_c_min >= arc_c_max) {
7924 arc_c_min = MAX(zfs_arc_max / 2,
7925 2ULL << SPA_MAXBLOCKSHIFT);
7930 * In userland, there's only the memory pressure that we artificially
7931 * create (see arc_available_memory()). Don't let arc_c get too
7932 * small, because it can cause transactions to be larger than
7933 * arc_c, causing arc_tempreserve_space() to fail.
7935 arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT);
7939 arc_p = (arc_c >> 1);
7941 /* Set min to 1/2 of arc_c_min */
7942 arc_meta_min = 1ULL << SPA_MAXBLOCKSHIFT;
7944 * Set arc_meta_limit to a percent of arc_c_max with a floor of
7945 * arc_meta_min, and a ceiling of arc_c_max.
7947 percent = MIN(zfs_arc_meta_limit_percent, 100);
7948 arc_meta_limit = MAX(arc_meta_min, (percent * arc_c_max) / 100);
7949 percent = MIN(zfs_arc_dnode_limit_percent, 100);
7950 arc_dnode_size_limit = (percent * arc_meta_limit) / 100;
7952 /* Apply user specified tunings */
7953 arc_tuning_update(B_TRUE);
7955 /* if kmem_flags are set, lets try to use less memory */
7956 if (kmem_debugging())
7958 if (arc_c < arc_c_min)
7961 arc_register_hotplug();
7967 list_create(&arc_prune_list, sizeof (arc_prune_t),
7968 offsetof(arc_prune_t, p_node));
7969 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7971 arc_prune_taskq = taskq_create("arc_prune", 100, defclsyspri,
7972 boot_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC |
7973 TASKQ_THREADS_CPU_PCT);
7975 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7976 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7978 if (arc_ksp != NULL) {
7979 arc_ksp->ks_data = &arc_stats;
7980 arc_ksp->ks_update = arc_kstat_update;
7981 kstat_install(arc_ksp);
7984 arc_evict_zthr = zthr_create("arc_evict",
7985 arc_evict_cb_check, arc_evict_cb, NULL, defclsyspri);
7986 arc_reap_zthr = zthr_create_timer("arc_reap",
7987 arc_reap_cb_check, arc_reap_cb, NULL, SEC2NSEC(1), minclsyspri);
7992 * Calculate maximum amount of dirty data per pool.
7994 * If it has been set by a module parameter, take that.
7995 * Otherwise, use a percentage of physical memory defined by
7996 * zfs_dirty_data_max_percent (default 10%) with a cap at
7997 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
8000 if (zfs_dirty_data_max_max == 0)
8001 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
8002 allmem * zfs_dirty_data_max_max_percent / 100);
8004 if (zfs_dirty_data_max_max == 0)
8005 zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024,
8006 allmem * zfs_dirty_data_max_max_percent / 100);
8009 if (zfs_dirty_data_max == 0) {
8010 zfs_dirty_data_max = allmem *
8011 zfs_dirty_data_max_percent / 100;
8012 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
8013 zfs_dirty_data_max_max);
8016 if (zfs_wrlog_data_max == 0) {
8019 * dp_wrlog_total is reduced for each txg at the end of
8020 * spa_sync(). However, dp_dirty_total is reduced every time
8021 * a block is written out. Thus under normal operation,
8022 * dp_wrlog_total could grow 2 times as big as
8023 * zfs_dirty_data_max.
8025 zfs_wrlog_data_max = zfs_dirty_data_max * 2;
8036 #endif /* _KERNEL */
8038 /* Use B_TRUE to ensure *all* buffers are evicted */
8039 arc_flush(NULL, B_TRUE);
8041 if (arc_ksp != NULL) {
8042 kstat_delete(arc_ksp);
8046 taskq_wait(arc_prune_taskq);
8047 taskq_destroy(arc_prune_taskq);
8049 mutex_enter(&arc_prune_mtx);
8050 while ((p = list_head(&arc_prune_list)) != NULL) {
8051 list_remove(&arc_prune_list, p);
8052 zfs_refcount_remove(&p->p_refcnt, &arc_prune_list);
8053 zfs_refcount_destroy(&p->p_refcnt);
8054 kmem_free(p, sizeof (*p));
8056 mutex_exit(&arc_prune_mtx);
8058 list_destroy(&arc_prune_list);
8059 mutex_destroy(&arc_prune_mtx);
8061 (void) zthr_cancel(arc_evict_zthr);
8062 (void) zthr_cancel(arc_reap_zthr);
8064 mutex_destroy(&arc_evict_lock);
8065 list_destroy(&arc_evict_waiters);
8068 * Free any buffers that were tagged for destruction. This needs
8069 * to occur before arc_state_fini() runs and destroys the aggsum
8070 * values which are updated when freeing scatter ABDs.
8072 l2arc_do_free_on_write();
8075 * buf_fini() must proceed arc_state_fini() because buf_fin() may
8076 * trigger the release of kmem magazines, which can callback to
8077 * arc_space_return() which accesses aggsums freed in act_state_fini().
8082 arc_unregister_hotplug();
8085 * We destroy the zthrs after all the ARC state has been
8086 * torn down to avoid the case of them receiving any
8087 * wakeup() signals after they are destroyed.
8089 zthr_destroy(arc_evict_zthr);
8090 zthr_destroy(arc_reap_zthr);
8092 ASSERT0(arc_loaned_bytes);
8098 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
8099 * It uses dedicated storage devices to hold cached data, which are populated
8100 * using large infrequent writes. The main role of this cache is to boost
8101 * the performance of random read workloads. The intended L2ARC devices
8102 * include short-stroked disks, solid state disks, and other media with
8103 * substantially faster read latency than disk.
8105 * +-----------------------+
8107 * +-----------------------+
8110 * l2arc_feed_thread() arc_read()
8114 * +---------------+ |
8116 * +---------------+ |
8121 * +-------+ +-------+
8123 * | cache | | cache |
8124 * +-------+ +-------+
8125 * +=========+ .-----.
8126 * : L2ARC : |-_____-|
8127 * : devices : | Disks |
8128 * +=========+ `-_____-'
8130 * Read requests are satisfied from the following sources, in order:
8133 * 2) vdev cache of L2ARC devices
8135 * 4) vdev cache of disks
8138 * Some L2ARC device types exhibit extremely slow write performance.
8139 * To accommodate for this there are some significant differences between
8140 * the L2ARC and traditional cache design:
8142 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
8143 * the ARC behave as usual, freeing buffers and placing headers on ghost
8144 * lists. The ARC does not send buffers to the L2ARC during eviction as
8145 * this would add inflated write latencies for all ARC memory pressure.
8147 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
8148 * It does this by periodically scanning buffers from the eviction-end of
8149 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
8150 * not already there. It scans until a headroom of buffers is satisfied,
8151 * which itself is a buffer for ARC eviction. If a compressible buffer is
8152 * found during scanning and selected for writing to an L2ARC device, we
8153 * temporarily boost scanning headroom during the next scan cycle to make
8154 * sure we adapt to compression effects (which might significantly reduce
8155 * the data volume we write to L2ARC). The thread that does this is
8156 * l2arc_feed_thread(), illustrated below; example sizes are included to
8157 * provide a better sense of ratio than this diagram:
8160 * +---------------------+----------+
8161 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
8162 * +---------------------+----------+ | o L2ARC eligible
8163 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
8164 * +---------------------+----------+ |
8165 * 15.9 Gbytes ^ 32 Mbytes |
8167 * l2arc_feed_thread()
8169 * l2arc write hand <--[oooo]--'
8173 * +==============================+
8174 * L2ARC dev |####|#|###|###| |####| ... |
8175 * +==============================+
8178 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
8179 * evicted, then the L2ARC has cached a buffer much sooner than it probably
8180 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
8181 * safe to say that this is an uncommon case, since buffers at the end of
8182 * the ARC lists have moved there due to inactivity.
8184 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
8185 * then the L2ARC simply misses copying some buffers. This serves as a
8186 * pressure valve to prevent heavy read workloads from both stalling the ARC
8187 * with waits and clogging the L2ARC with writes. This also helps prevent
8188 * the potential for the L2ARC to churn if it attempts to cache content too
8189 * quickly, such as during backups of the entire pool.
8191 * 5. After system boot and before the ARC has filled main memory, there are
8192 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
8193 * lists can remain mostly static. Instead of searching from tail of these
8194 * lists as pictured, the l2arc_feed_thread() will search from the list heads
8195 * for eligible buffers, greatly increasing its chance of finding them.
8197 * The L2ARC device write speed is also boosted during this time so that
8198 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
8199 * there are no L2ARC reads, and no fear of degrading read performance
8200 * through increased writes.
8202 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
8203 * the vdev queue can aggregate them into larger and fewer writes. Each
8204 * device is written to in a rotor fashion, sweeping writes through
8205 * available space then repeating.
8207 * 7. The L2ARC does not store dirty content. It never needs to flush
8208 * write buffers back to disk based storage.
8210 * 8. If an ARC buffer is written (and dirtied) which also exists in the
8211 * L2ARC, the now stale L2ARC buffer is immediately dropped.
8213 * The performance of the L2ARC can be tweaked by a number of tunables, which
8214 * may be necessary for different workloads:
8216 * l2arc_write_max max write bytes per interval
8217 * l2arc_write_boost extra write bytes during device warmup
8218 * l2arc_noprefetch skip caching prefetched buffers
8219 * l2arc_headroom number of max device writes to precache
8220 * l2arc_headroom_boost when we find compressed buffers during ARC
8221 * scanning, we multiply headroom by this
8222 * percentage factor for the next scan cycle,
8223 * since more compressed buffers are likely to
8225 * l2arc_feed_secs seconds between L2ARC writing
8227 * Tunables may be removed or added as future performance improvements are
8228 * integrated, and also may become zpool properties.
8230 * There are three key functions that control how the L2ARC warms up:
8232 * l2arc_write_eligible() check if a buffer is eligible to cache
8233 * l2arc_write_size() calculate how much to write
8234 * l2arc_write_interval() calculate sleep delay between writes
8236 * These three functions determine what to write, how much, and how quickly
8239 * L2ARC persistence:
8241 * When writing buffers to L2ARC, we periodically add some metadata to
8242 * make sure we can pick them up after reboot, thus dramatically reducing
8243 * the impact that any downtime has on the performance of storage systems
8244 * with large caches.
8246 * The implementation works fairly simply by integrating the following two
8249 * *) When writing to the L2ARC, we occasionally write a "l2arc log block",
8250 * which is an additional piece of metadata which describes what's been
8251 * written. This allows us to rebuild the arc_buf_hdr_t structures of the
8252 * main ARC buffers. There are 2 linked-lists of log blocks headed by
8253 * dh_start_lbps[2]. We alternate which chain we append to, so they are
8254 * time-wise and offset-wise interleaved, but that is an optimization rather
8255 * than for correctness. The log block also includes a pointer to the
8256 * previous block in its chain.
8258 * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device
8259 * for our header bookkeeping purposes. This contains a device header,
8260 * which contains our top-level reference structures. We update it each
8261 * time we write a new log block, so that we're able to locate it in the
8262 * L2ARC device. If this write results in an inconsistent device header
8263 * (e.g. due to power failure), we detect this by verifying the header's
8264 * checksum and simply fail to reconstruct the L2ARC after reboot.
8266 * Implementation diagram:
8268 * +=== L2ARC device (not to scale) ======================================+
8269 * | ___two newest log block pointers__.__________ |
8270 * | / \dh_start_lbps[1] |
8271 * | / \ \dh_start_lbps[0]|
8273 * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---|
8274 * || hdr| ^ /^ /^ / / |
8275 * |+------+ ...--\-------/ \-----/--\------/ / |
8276 * | \--------------/ \--------------/ |
8277 * +======================================================================+
8279 * As can be seen on the diagram, rather than using a simple linked list,
8280 * we use a pair of linked lists with alternating elements. This is a
8281 * performance enhancement due to the fact that we only find out the
8282 * address of the next log block access once the current block has been
8283 * completely read in. Obviously, this hurts performance, because we'd be
8284 * keeping the device's I/O queue at only a 1 operation deep, thus
8285 * incurring a large amount of I/O round-trip latency. Having two lists
8286 * allows us to fetch two log blocks ahead of where we are currently
8287 * rebuilding L2ARC buffers.
8289 * On-device data structures:
8291 * L2ARC device header: l2arc_dev_hdr_phys_t
8292 * L2ARC log block: l2arc_log_blk_phys_t
8294 * L2ARC reconstruction:
8296 * When writing data, we simply write in the standard rotary fashion,
8297 * evicting buffers as we go and simply writing new data over them (writing
8298 * a new log block every now and then). This obviously means that once we
8299 * loop around the end of the device, we will start cutting into an already
8300 * committed log block (and its referenced data buffers), like so:
8302 * current write head__ __old tail
8305 * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |-->
8306 * ^ ^^^^^^^^^___________________________________
8308 * <<nextwrite>> may overwrite this blk and/or its bufs --'
8310 * When importing the pool, we detect this situation and use it to stop
8311 * our scanning process (see l2arc_rebuild).
8313 * There is one significant caveat to consider when rebuilding ARC contents
8314 * from an L2ARC device: what about invalidated buffers? Given the above
8315 * construction, we cannot update blocks which we've already written to amend
8316 * them to remove buffers which were invalidated. Thus, during reconstruction,
8317 * we might be populating the cache with buffers for data that's not on the
8318 * main pool anymore, or may have been overwritten!
8320 * As it turns out, this isn't a problem. Every arc_read request includes
8321 * both the DVA and, crucially, the birth TXG of the BP the caller is
8322 * looking for. So even if the cache were populated by completely rotten
8323 * blocks for data that had been long deleted and/or overwritten, we'll
8324 * never actually return bad data from the cache, since the DVA with the
8325 * birth TXG uniquely identify a block in space and time - once created,
8326 * a block is immutable on disk. The worst thing we have done is wasted
8327 * some time and memory at l2arc rebuild to reconstruct outdated ARC
8328 * entries that will get dropped from the l2arc as it is being updated
8331 * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write
8332 * hand are not restored. This is done by saving the offset (in bytes)
8333 * l2arc_evict() has evicted to in the L2ARC device header and taking it
8334 * into account when restoring buffers.
8338 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
8341 * A buffer is *not* eligible for the L2ARC if it:
8342 * 1. belongs to a different spa.
8343 * 2. is already cached on the L2ARC.
8344 * 3. has an I/O in progress (it may be an incomplete read).
8345 * 4. is flagged not eligible (zfs property).
8347 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
8348 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
8355 l2arc_write_size(l2arc_dev_t *dev)
8357 uint64_t size, dev_size, tsize;
8360 * Make sure our globals have meaningful values in case the user
8363 size = l2arc_write_max;
8365 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
8366 "be greater than zero, resetting it to the default (%d)",
8368 size = l2arc_write_max = L2ARC_WRITE_SIZE;
8371 if (arc_warm == B_FALSE)
8372 size += l2arc_write_boost;
8375 * Make sure the write size does not exceed the size of the cache
8376 * device. This is important in l2arc_evict(), otherwise infinite
8377 * iteration can occur.
8379 dev_size = dev->l2ad_end - dev->l2ad_start;
8380 tsize = size + l2arc_log_blk_overhead(size, dev);
8381 if (dev->l2ad_vdev->vdev_has_trim && l2arc_trim_ahead > 0)
8382 tsize += MAX(64 * 1024 * 1024,
8383 (tsize * l2arc_trim_ahead) / 100);
8385 if (tsize >= dev_size) {
8386 cmn_err(CE_NOTE, "l2arc_write_max or l2arc_write_boost "
8387 "plus the overhead of log blocks (persistent L2ARC, "
8388 "%llu bytes) exceeds the size of the cache device "
8389 "(guid %llu), resetting them to the default (%d)",
8390 (u_longlong_t)l2arc_log_blk_overhead(size, dev),
8391 (u_longlong_t)dev->l2ad_vdev->vdev_guid, L2ARC_WRITE_SIZE);
8392 size = l2arc_write_max = l2arc_write_boost = L2ARC_WRITE_SIZE;
8394 if (arc_warm == B_FALSE)
8395 size += l2arc_write_boost;
8403 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
8405 clock_t interval, next, now;
8408 * If the ARC lists are busy, increase our write rate; if the
8409 * lists are stale, idle back. This is achieved by checking
8410 * how much we previously wrote - if it was more than half of
8411 * what we wanted, schedule the next write much sooner.
8413 if (l2arc_feed_again && wrote > (wanted / 2))
8414 interval = (hz * l2arc_feed_min_ms) / 1000;
8416 interval = hz * l2arc_feed_secs;
8418 now = ddi_get_lbolt();
8419 next = MAX(now, MIN(now + interval, began + interval));
8425 * Cycle through L2ARC devices. This is how L2ARC load balances.
8426 * If a device is returned, this also returns holding the spa config lock.
8428 static l2arc_dev_t *
8429 l2arc_dev_get_next(void)
8431 l2arc_dev_t *first, *next = NULL;
8434 * Lock out the removal of spas (spa_namespace_lock), then removal
8435 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
8436 * both locks will be dropped and a spa config lock held instead.
8438 mutex_enter(&spa_namespace_lock);
8439 mutex_enter(&l2arc_dev_mtx);
8441 /* if there are no vdevs, there is nothing to do */
8442 if (l2arc_ndev == 0)
8446 next = l2arc_dev_last;
8448 /* loop around the list looking for a non-faulted vdev */
8450 next = list_head(l2arc_dev_list);
8452 next = list_next(l2arc_dev_list, next);
8454 next = list_head(l2arc_dev_list);
8457 /* if we have come back to the start, bail out */
8460 else if (next == first)
8463 } while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8464 next->l2ad_trim_all);
8466 /* if we were unable to find any usable vdevs, return NULL */
8467 if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8468 next->l2ad_trim_all)
8471 l2arc_dev_last = next;
8474 mutex_exit(&l2arc_dev_mtx);
8477 * Grab the config lock to prevent the 'next' device from being
8478 * removed while we are writing to it.
8481 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
8482 mutex_exit(&spa_namespace_lock);
8488 * Free buffers that were tagged for destruction.
8491 l2arc_do_free_on_write(void)
8494 l2arc_data_free_t *df, *df_prev;
8496 mutex_enter(&l2arc_free_on_write_mtx);
8497 buflist = l2arc_free_on_write;
8499 for (df = list_tail(buflist); df; df = df_prev) {
8500 df_prev = list_prev(buflist, df);
8501 ASSERT3P(df->l2df_abd, !=, NULL);
8502 abd_free(df->l2df_abd);
8503 list_remove(buflist, df);
8504 kmem_free(df, sizeof (l2arc_data_free_t));
8507 mutex_exit(&l2arc_free_on_write_mtx);
8511 * A write to a cache device has completed. Update all headers to allow
8512 * reads from these buffers to begin.
8515 l2arc_write_done(zio_t *zio)
8517 l2arc_write_callback_t *cb;
8518 l2arc_lb_abd_buf_t *abd_buf;
8519 l2arc_lb_ptr_buf_t *lb_ptr_buf;
8521 l2arc_dev_hdr_phys_t *l2dhdr;
8523 arc_buf_hdr_t *head, *hdr, *hdr_prev;
8524 kmutex_t *hash_lock;
8525 int64_t bytes_dropped = 0;
8527 cb = zio->io_private;
8528 ASSERT3P(cb, !=, NULL);
8529 dev = cb->l2wcb_dev;
8530 l2dhdr = dev->l2ad_dev_hdr;
8531 ASSERT3P(dev, !=, NULL);
8532 head = cb->l2wcb_head;
8533 ASSERT3P(head, !=, NULL);
8534 buflist = &dev->l2ad_buflist;
8535 ASSERT3P(buflist, !=, NULL);
8536 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
8537 l2arc_write_callback_t *, cb);
8540 * All writes completed, or an error was hit.
8543 mutex_enter(&dev->l2ad_mtx);
8544 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
8545 hdr_prev = list_prev(buflist, hdr);
8547 hash_lock = HDR_LOCK(hdr);
8550 * We cannot use mutex_enter or else we can deadlock
8551 * with l2arc_write_buffers (due to swapping the order
8552 * the hash lock and l2ad_mtx are taken).
8554 if (!mutex_tryenter(hash_lock)) {
8556 * Missed the hash lock. We must retry so we
8557 * don't leave the ARC_FLAG_L2_WRITING bit set.
8559 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
8562 * We don't want to rescan the headers we've
8563 * already marked as having been written out, so
8564 * we reinsert the head node so we can pick up
8565 * where we left off.
8567 list_remove(buflist, head);
8568 list_insert_after(buflist, hdr, head);
8570 mutex_exit(&dev->l2ad_mtx);
8573 * We wait for the hash lock to become available
8574 * to try and prevent busy waiting, and increase
8575 * the chance we'll be able to acquire the lock
8576 * the next time around.
8578 mutex_enter(hash_lock);
8579 mutex_exit(hash_lock);
8584 * We could not have been moved into the arc_l2c_only
8585 * state while in-flight due to our ARC_FLAG_L2_WRITING
8586 * bit being set. Let's just ensure that's being enforced.
8588 ASSERT(HDR_HAS_L1HDR(hdr));
8591 * Skipped - drop L2ARC entry and mark the header as no
8592 * longer L2 eligibile.
8594 if (zio->io_error != 0) {
8596 * Error - drop L2ARC entry.
8598 list_remove(buflist, hdr);
8599 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
8601 uint64_t psize = HDR_GET_PSIZE(hdr);
8602 l2arc_hdr_arcstats_decrement(hdr);
8605 vdev_psize_to_asize(dev->l2ad_vdev, psize);
8606 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
8607 arc_hdr_size(hdr), hdr);
8611 * Allow ARC to begin reads and ghost list evictions to
8614 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
8616 mutex_exit(hash_lock);
8620 * Free the allocated abd buffers for writing the log blocks.
8621 * If the zio failed reclaim the allocated space and remove the
8622 * pointers to these log blocks from the log block pointer list
8623 * of the L2ARC device.
8625 while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) {
8626 abd_free(abd_buf->abd);
8627 zio_buf_free(abd_buf, sizeof (*abd_buf));
8628 if (zio->io_error != 0) {
8629 lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list);
8631 * L2BLK_GET_PSIZE returns aligned size for log
8635 L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop);
8636 bytes_dropped += asize;
8637 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8638 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8639 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8641 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8642 kmem_free(lb_ptr_buf->lb_ptr,
8643 sizeof (l2arc_log_blkptr_t));
8644 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8647 list_destroy(&cb->l2wcb_abd_list);
8649 if (zio->io_error != 0) {
8650 ARCSTAT_BUMP(arcstat_l2_writes_error);
8653 * Restore the lbps array in the header to its previous state.
8654 * If the list of log block pointers is empty, zero out the
8655 * log block pointers in the device header.
8657 lb_ptr_buf = list_head(&dev->l2ad_lbptr_list);
8658 for (int i = 0; i < 2; i++) {
8659 if (lb_ptr_buf == NULL) {
8661 * If the list is empty zero out the device
8662 * header. Otherwise zero out the second log
8663 * block pointer in the header.
8666 bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
8668 bzero(&l2dhdr->dh_start_lbps[i],
8669 sizeof (l2arc_log_blkptr_t));
8673 bcopy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[i],
8674 sizeof (l2arc_log_blkptr_t));
8675 lb_ptr_buf = list_next(&dev->l2ad_lbptr_list,
8680 ARCSTAT_BUMP(arcstat_l2_writes_done);
8681 list_remove(buflist, head);
8682 ASSERT(!HDR_HAS_L1HDR(head));
8683 kmem_cache_free(hdr_l2only_cache, head);
8684 mutex_exit(&dev->l2ad_mtx);
8686 ASSERT(dev->l2ad_vdev != NULL);
8687 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
8689 l2arc_do_free_on_write();
8691 kmem_free(cb, sizeof (l2arc_write_callback_t));
8695 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
8698 spa_t *spa = zio->io_spa;
8699 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
8700 blkptr_t *bp = zio->io_bp;
8701 uint8_t salt[ZIO_DATA_SALT_LEN];
8702 uint8_t iv[ZIO_DATA_IV_LEN];
8703 uint8_t mac[ZIO_DATA_MAC_LEN];
8704 boolean_t no_crypt = B_FALSE;
8707 * ZIL data is never be written to the L2ARC, so we don't need
8708 * special handling for its unique MAC storage.
8710 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
8711 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
8712 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8715 * If the data was encrypted, decrypt it now. Note that
8716 * we must check the bp here and not the hdr, since the
8717 * hdr does not have its encryption parameters updated
8718 * until arc_read_done().
8720 if (BP_IS_ENCRYPTED(bp)) {
8721 abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8722 ARC_HDR_DO_ADAPT | ARC_HDR_USE_RESERVE);
8724 zio_crypt_decode_params_bp(bp, salt, iv);
8725 zio_crypt_decode_mac_bp(bp, mac);
8727 ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
8728 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
8729 salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
8730 hdr->b_l1hdr.b_pabd, &no_crypt);
8732 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8737 * If we actually performed decryption, replace b_pabd
8738 * with the decrypted data. Otherwise we can just throw
8739 * our decryption buffer away.
8742 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8743 arc_hdr_size(hdr), hdr);
8744 hdr->b_l1hdr.b_pabd = eabd;
8747 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8752 * If the L2ARC block was compressed, but ARC compression
8753 * is disabled we decompress the data into a new buffer and
8754 * replace the existing data.
8756 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8757 !HDR_COMPRESSION_ENABLED(hdr)) {
8758 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8759 ARC_HDR_DO_ADAPT | ARC_HDR_USE_RESERVE);
8760 void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
8762 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
8763 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
8764 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
8766 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8767 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
8771 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8772 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8773 arc_hdr_size(hdr), hdr);
8774 hdr->b_l1hdr.b_pabd = cabd;
8776 zio->io_size = HDR_GET_LSIZE(hdr);
8787 * A read to a cache device completed. Validate buffer contents before
8788 * handing over to the regular ARC routines.
8791 l2arc_read_done(zio_t *zio)
8794 l2arc_read_callback_t *cb = zio->io_private;
8796 kmutex_t *hash_lock;
8797 boolean_t valid_cksum;
8798 boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
8799 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
8801 ASSERT3P(zio->io_vd, !=, NULL);
8802 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
8804 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
8806 ASSERT3P(cb, !=, NULL);
8807 hdr = cb->l2rcb_hdr;
8808 ASSERT3P(hdr, !=, NULL);
8810 hash_lock = HDR_LOCK(hdr);
8811 mutex_enter(hash_lock);
8812 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
8815 * If the data was read into a temporary buffer,
8816 * move it and free the buffer.
8818 if (cb->l2rcb_abd != NULL) {
8819 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
8820 if (zio->io_error == 0) {
8822 abd_copy(hdr->b_crypt_hdr.b_rabd,
8823 cb->l2rcb_abd, arc_hdr_size(hdr));
8825 abd_copy(hdr->b_l1hdr.b_pabd,
8826 cb->l2rcb_abd, arc_hdr_size(hdr));
8831 * The following must be done regardless of whether
8832 * there was an error:
8833 * - free the temporary buffer
8834 * - point zio to the real ARC buffer
8835 * - set zio size accordingly
8836 * These are required because zio is either re-used for
8837 * an I/O of the block in the case of the error
8838 * or the zio is passed to arc_read_done() and it
8841 abd_free(cb->l2rcb_abd);
8842 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
8845 ASSERT(HDR_HAS_RABD(hdr));
8846 zio->io_abd = zio->io_orig_abd =
8847 hdr->b_crypt_hdr.b_rabd;
8849 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8850 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8854 ASSERT3P(zio->io_abd, !=, NULL);
8857 * Check this survived the L2ARC journey.
8859 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8860 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8861 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
8862 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
8863 zio->io_prop.zp_complevel = hdr->b_complevel;
8865 valid_cksum = arc_cksum_is_equal(hdr, zio);
8868 * b_rabd will always match the data as it exists on disk if it is
8869 * being used. Therefore if we are reading into b_rabd we do not
8870 * attempt to untransform the data.
8872 if (valid_cksum && !using_rdata)
8873 tfm_error = l2arc_untransform(zio, cb);
8875 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8876 !HDR_L2_EVICTED(hdr)) {
8877 mutex_exit(hash_lock);
8878 zio->io_private = hdr;
8882 * Buffer didn't survive caching. Increment stats and
8883 * reissue to the original storage device.
8885 if (zio->io_error != 0) {
8886 ARCSTAT_BUMP(arcstat_l2_io_error);
8888 zio->io_error = SET_ERROR(EIO);
8890 if (!valid_cksum || tfm_error != 0)
8891 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8894 * If there's no waiter, issue an async i/o to the primary
8895 * storage now. If there *is* a waiter, the caller must
8896 * issue the i/o in a context where it's OK to block.
8898 if (zio->io_waiter == NULL) {
8899 zio_t *pio = zio_unique_parent(zio);
8900 void *abd = (using_rdata) ?
8901 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8903 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8905 zio = zio_read(pio, zio->io_spa, zio->io_bp,
8906 abd, zio->io_size, arc_read_done,
8907 hdr, zio->io_priority, cb->l2rcb_flags,
8911 * Original ZIO will be freed, so we need to update
8912 * ARC header with the new ZIO pointer to be used
8913 * by zio_change_priority() in arc_read().
8915 for (struct arc_callback *acb = hdr->b_l1hdr.b_acb;
8916 acb != NULL; acb = acb->acb_next)
8917 acb->acb_zio_head = zio;
8919 mutex_exit(hash_lock);
8922 mutex_exit(hash_lock);
8926 kmem_free(cb, sizeof (l2arc_read_callback_t));
8930 * This is the list priority from which the L2ARC will search for pages to
8931 * cache. This is used within loops (0..3) to cycle through lists in the
8932 * desired order. This order can have a significant effect on cache
8935 * Currently the metadata lists are hit first, MFU then MRU, followed by
8936 * the data lists. This function returns a locked list, and also returns
8939 static multilist_sublist_t *
8940 l2arc_sublist_lock(int list_num)
8942 multilist_t *ml = NULL;
8945 ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
8949 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
8952 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
8955 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
8958 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
8965 * Return a randomly-selected sublist. This is acceptable
8966 * because the caller feeds only a little bit of data for each
8967 * call (8MB). Subsequent calls will result in different
8968 * sublists being selected.
8970 idx = multilist_get_random_index(ml);
8971 return (multilist_sublist_lock(ml, idx));
8975 * Calculates the maximum overhead of L2ARC metadata log blocks for a given
8976 * L2ARC write size. l2arc_evict and l2arc_write_size need to include this
8977 * overhead in processing to make sure there is enough headroom available
8978 * when writing buffers.
8980 static inline uint64_t
8981 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev)
8983 if (dev->l2ad_log_entries == 0) {
8986 uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT;
8988 uint64_t log_blocks = (log_entries +
8989 dev->l2ad_log_entries - 1) /
8990 dev->l2ad_log_entries;
8992 return (vdev_psize_to_asize(dev->l2ad_vdev,
8993 sizeof (l2arc_log_blk_phys_t)) * log_blocks);
8998 * Evict buffers from the device write hand to the distance specified in
8999 * bytes. This distance may span populated buffers, it may span nothing.
9000 * This is clearing a region on the L2ARC device ready for writing.
9001 * If the 'all' boolean is set, every buffer is evicted.
9004 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
9007 arc_buf_hdr_t *hdr, *hdr_prev;
9008 kmutex_t *hash_lock;
9010 l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev;
9011 vdev_t *vd = dev->l2ad_vdev;
9014 buflist = &dev->l2ad_buflist;
9017 * We need to add in the worst case scenario of log block overhead.
9019 distance += l2arc_log_blk_overhead(distance, dev);
9020 if (vd->vdev_has_trim && l2arc_trim_ahead > 0) {
9022 * Trim ahead of the write size 64MB or (l2arc_trim_ahead/100)
9023 * times the write size, whichever is greater.
9025 distance += MAX(64 * 1024 * 1024,
9026 (distance * l2arc_trim_ahead) / 100);
9031 if (dev->l2ad_hand >= (dev->l2ad_end - distance)) {
9033 * When there is no space to accommodate upcoming writes,
9034 * evict to the end. Then bump the write and evict hands
9035 * to the start and iterate. This iteration does not
9036 * happen indefinitely as we make sure in
9037 * l2arc_write_size() that when the write hand is reset,
9038 * the write size does not exceed the end of the device.
9041 taddr = dev->l2ad_end;
9043 taddr = dev->l2ad_hand + distance;
9045 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
9046 uint64_t, taddr, boolean_t, all);
9050 * This check has to be placed after deciding whether to
9053 if (dev->l2ad_first) {
9055 * This is the first sweep through the device. There is
9056 * nothing to evict. We have already trimmmed the
9062 * Trim the space to be evicted.
9064 if (vd->vdev_has_trim && dev->l2ad_evict < taddr &&
9065 l2arc_trim_ahead > 0) {
9067 * We have to drop the spa_config lock because
9068 * vdev_trim_range() will acquire it.
9069 * l2ad_evict already accounts for the label
9070 * size. To prevent vdev_trim_ranges() from
9071 * adding it again, we subtract it from
9074 spa_config_exit(dev->l2ad_spa, SCL_L2ARC, dev);
9075 vdev_trim_simple(vd,
9076 dev->l2ad_evict - VDEV_LABEL_START_SIZE,
9077 taddr - dev->l2ad_evict);
9078 spa_config_enter(dev->l2ad_spa, SCL_L2ARC, dev,
9083 * When rebuilding L2ARC we retrieve the evict hand
9084 * from the header of the device. Of note, l2arc_evict()
9085 * does not actually delete buffers from the cache
9086 * device, but trimming may do so depending on the
9087 * hardware implementation. Thus keeping track of the
9088 * evict hand is useful.
9090 dev->l2ad_evict = MAX(dev->l2ad_evict, taddr);
9095 mutex_enter(&dev->l2ad_mtx);
9097 * We have to account for evicted log blocks. Run vdev_space_update()
9098 * on log blocks whose offset (in bytes) is before the evicted offset
9099 * (in bytes) by searching in the list of pointers to log blocks
9100 * present in the L2ARC device.
9102 for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf;
9103 lb_ptr_buf = lb_ptr_buf_prev) {
9105 lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf);
9107 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
9108 uint64_t asize = L2BLK_GET_PSIZE(
9109 (lb_ptr_buf->lb_ptr)->lbp_prop);
9112 * We don't worry about log blocks left behind (ie
9113 * lbp_payload_start < l2ad_hand) because l2arc_write_buffers()
9114 * will never write more than l2arc_evict() evicts.
9116 if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) {
9119 vdev_space_update(vd, -asize, 0, 0);
9120 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
9121 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
9122 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
9124 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
9125 list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf);
9126 kmem_free(lb_ptr_buf->lb_ptr,
9127 sizeof (l2arc_log_blkptr_t));
9128 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
9132 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
9133 hdr_prev = list_prev(buflist, hdr);
9135 ASSERT(!HDR_EMPTY(hdr));
9136 hash_lock = HDR_LOCK(hdr);
9139 * We cannot use mutex_enter or else we can deadlock
9140 * with l2arc_write_buffers (due to swapping the order
9141 * the hash lock and l2ad_mtx are taken).
9143 if (!mutex_tryenter(hash_lock)) {
9145 * Missed the hash lock. Retry.
9147 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
9148 mutex_exit(&dev->l2ad_mtx);
9149 mutex_enter(hash_lock);
9150 mutex_exit(hash_lock);
9155 * A header can't be on this list if it doesn't have L2 header.
9157 ASSERT(HDR_HAS_L2HDR(hdr));
9159 /* Ensure this header has finished being written. */
9160 ASSERT(!HDR_L2_WRITING(hdr));
9161 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
9163 if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict ||
9164 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
9166 * We've evicted to the target address,
9167 * or the end of the device.
9169 mutex_exit(hash_lock);
9173 if (!HDR_HAS_L1HDR(hdr)) {
9174 ASSERT(!HDR_L2_READING(hdr));
9176 * This doesn't exist in the ARC. Destroy.
9177 * arc_hdr_destroy() will call list_remove()
9178 * and decrement arcstat_l2_lsize.
9180 arc_change_state(arc_anon, hdr, hash_lock);
9181 arc_hdr_destroy(hdr);
9183 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
9184 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
9186 * Invalidate issued or about to be issued
9187 * reads, since we may be about to write
9188 * over this location.
9190 if (HDR_L2_READING(hdr)) {
9191 ARCSTAT_BUMP(arcstat_l2_evict_reading);
9192 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
9195 arc_hdr_l2hdr_destroy(hdr);
9197 mutex_exit(hash_lock);
9199 mutex_exit(&dev->l2ad_mtx);
9203 * We need to check if we evict all buffers, otherwise we may iterate
9206 if (!all && rerun) {
9208 * Bump device hand to the device start if it is approaching the
9209 * end. l2arc_evict() has already evicted ahead for this case.
9211 dev->l2ad_hand = dev->l2ad_start;
9212 dev->l2ad_evict = dev->l2ad_start;
9213 dev->l2ad_first = B_FALSE;
9219 * In case of cache device removal (all) the following
9220 * assertions may be violated without functional consequences
9221 * as the device is about to be removed.
9223 ASSERT3U(dev->l2ad_hand + distance, <, dev->l2ad_end);
9224 if (!dev->l2ad_first)
9225 ASSERT3U(dev->l2ad_hand, <, dev->l2ad_evict);
9230 * Handle any abd transforms that might be required for writing to the L2ARC.
9231 * If successful, this function will always return an abd with the data
9232 * transformed as it is on disk in a new abd of asize bytes.
9235 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
9240 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
9241 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
9242 uint64_t psize = HDR_GET_PSIZE(hdr);
9243 uint64_t size = arc_hdr_size(hdr);
9244 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
9245 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
9246 dsl_crypto_key_t *dck = NULL;
9247 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
9248 boolean_t no_crypt = B_FALSE;
9250 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
9251 !HDR_COMPRESSION_ENABLED(hdr)) ||
9252 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
9253 ASSERT3U(psize, <=, asize);
9256 * If this data simply needs its own buffer, we simply allocate it
9257 * and copy the data. This may be done to eliminate a dependency on a
9258 * shared buffer or to reallocate the buffer to match asize.
9260 if (HDR_HAS_RABD(hdr) && asize != psize) {
9261 ASSERT3U(asize, >=, psize);
9262 to_write = abd_alloc_for_io(asize, ismd);
9263 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
9265 abd_zero_off(to_write, psize, asize - psize);
9269 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
9270 !HDR_ENCRYPTED(hdr)) {
9271 ASSERT3U(size, ==, psize);
9272 to_write = abd_alloc_for_io(asize, ismd);
9273 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
9275 abd_zero_off(to_write, size, asize - size);
9279 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
9280 cabd = abd_alloc_for_io(asize, ismd);
9281 tmp = abd_borrow_buf(cabd, asize);
9283 psize = zio_compress_data(compress, to_write, tmp, size,
9286 if (psize >= size) {
9287 abd_return_buf(cabd, tmp, asize);
9288 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
9290 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
9292 abd_zero_off(to_write, size, asize - size);
9295 ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
9297 bzero((char *)tmp + psize, asize - psize);
9298 psize = HDR_GET_PSIZE(hdr);
9299 abd_return_buf_copy(cabd, tmp, asize);
9304 if (HDR_ENCRYPTED(hdr)) {
9305 eabd = abd_alloc_for_io(asize, ismd);
9308 * If the dataset was disowned before the buffer
9309 * made it to this point, the key to re-encrypt
9310 * it won't be available. In this case we simply
9311 * won't write the buffer to the L2ARC.
9313 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
9318 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
9319 hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
9320 hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
9326 abd_copy(eabd, to_write, psize);
9329 abd_zero_off(eabd, psize, asize - psize);
9331 /* assert that the MAC we got here matches the one we saved */
9332 ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
9333 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9335 if (to_write == cabd)
9342 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
9343 *abd_out = to_write;
9348 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9359 l2arc_blk_fetch_done(zio_t *zio)
9361 l2arc_read_callback_t *cb;
9363 cb = zio->io_private;
9364 if (cb->l2rcb_abd != NULL)
9365 abd_free(cb->l2rcb_abd);
9366 kmem_free(cb, sizeof (l2arc_read_callback_t));
9370 * Find and write ARC buffers to the L2ARC device.
9372 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
9373 * for reading until they have completed writing.
9374 * The headroom_boost is an in-out parameter used to maintain headroom boost
9375 * state between calls to this function.
9377 * Returns the number of bytes actually written (which may be smaller than
9378 * the delta by which the device hand has changed due to alignment and the
9379 * writing of log blocks).
9382 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
9384 arc_buf_hdr_t *hdr, *hdr_prev, *head;
9385 uint64_t write_asize, write_psize, write_lsize, headroom;
9387 l2arc_write_callback_t *cb = NULL;
9389 uint64_t guid = spa_load_guid(spa);
9390 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9392 ASSERT3P(dev->l2ad_vdev, !=, NULL);
9395 write_lsize = write_asize = write_psize = 0;
9397 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
9398 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
9401 * Copy buffers for L2ARC writing.
9403 for (int pass = 0; pass < L2ARC_FEED_TYPES; pass++) {
9405 * If pass == 1 or 3, we cache MRU metadata and data
9408 if (l2arc_mfuonly) {
9409 if (pass == 1 || pass == 3)
9413 multilist_sublist_t *mls = l2arc_sublist_lock(pass);
9414 uint64_t passed_sz = 0;
9416 VERIFY3P(mls, !=, NULL);
9419 * L2ARC fast warmup.
9421 * Until the ARC is warm and starts to evict, read from the
9422 * head of the ARC lists rather than the tail.
9424 if (arc_warm == B_FALSE)
9425 hdr = multilist_sublist_head(mls);
9427 hdr = multilist_sublist_tail(mls);
9429 headroom = target_sz * l2arc_headroom;
9430 if (zfs_compressed_arc_enabled)
9431 headroom = (headroom * l2arc_headroom_boost) / 100;
9433 for (; hdr; hdr = hdr_prev) {
9434 kmutex_t *hash_lock;
9435 abd_t *to_write = NULL;
9437 if (arc_warm == B_FALSE)
9438 hdr_prev = multilist_sublist_next(mls, hdr);
9440 hdr_prev = multilist_sublist_prev(mls, hdr);
9442 hash_lock = HDR_LOCK(hdr);
9443 if (!mutex_tryenter(hash_lock)) {
9445 * Skip this buffer rather than waiting.
9450 passed_sz += HDR_GET_LSIZE(hdr);
9451 if (l2arc_headroom != 0 && passed_sz > headroom) {
9455 mutex_exit(hash_lock);
9459 if (!l2arc_write_eligible(guid, hdr)) {
9460 mutex_exit(hash_lock);
9465 * We rely on the L1 portion of the header below, so
9466 * it's invalid for this header to have been evicted out
9467 * of the ghost cache, prior to being written out. The
9468 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
9470 ASSERT(HDR_HAS_L1HDR(hdr));
9472 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
9473 ASSERT3U(arc_hdr_size(hdr), >, 0);
9474 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
9476 uint64_t psize = HDR_GET_PSIZE(hdr);
9477 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
9480 if ((write_asize + asize) > target_sz) {
9482 mutex_exit(hash_lock);
9487 * We rely on the L1 portion of the header below, so
9488 * it's invalid for this header to have been evicted out
9489 * of the ghost cache, prior to being written out. The
9490 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
9492 arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
9493 ASSERT(HDR_HAS_L1HDR(hdr));
9495 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
9496 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
9498 ASSERT3U(arc_hdr_size(hdr), >, 0);
9501 * If this header has b_rabd, we can use this since it
9502 * must always match the data exactly as it exists on
9503 * disk. Otherwise, the L2ARC can normally use the
9504 * hdr's data, but if we're sharing data between the
9505 * hdr and one of its bufs, L2ARC needs its own copy of
9506 * the data so that the ZIO below can't race with the
9507 * buf consumer. To ensure that this copy will be
9508 * available for the lifetime of the ZIO and be cleaned
9509 * up afterwards, we add it to the l2arc_free_on_write
9510 * queue. If we need to apply any transforms to the
9511 * data (compression, encryption) we will also need the
9514 if (HDR_HAS_RABD(hdr) && psize == asize) {
9515 to_write = hdr->b_crypt_hdr.b_rabd;
9516 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
9517 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
9518 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
9520 to_write = hdr->b_l1hdr.b_pabd;
9523 arc_buf_contents_t type = arc_buf_type(hdr);
9525 ret = l2arc_apply_transforms(spa, hdr, asize,
9528 arc_hdr_clear_flags(hdr,
9529 ARC_FLAG_L2_WRITING);
9530 mutex_exit(hash_lock);
9534 l2arc_free_abd_on_write(to_write, asize, type);
9539 * Insert a dummy header on the buflist so
9540 * l2arc_write_done() can find where the
9541 * write buffers begin without searching.
9543 mutex_enter(&dev->l2ad_mtx);
9544 list_insert_head(&dev->l2ad_buflist, head);
9545 mutex_exit(&dev->l2ad_mtx);
9548 sizeof (l2arc_write_callback_t), KM_SLEEP);
9549 cb->l2wcb_dev = dev;
9550 cb->l2wcb_head = head;
9552 * Create a list to save allocated abd buffers
9553 * for l2arc_log_blk_commit().
9555 list_create(&cb->l2wcb_abd_list,
9556 sizeof (l2arc_lb_abd_buf_t),
9557 offsetof(l2arc_lb_abd_buf_t, node));
9558 pio = zio_root(spa, l2arc_write_done, cb,
9562 hdr->b_l2hdr.b_dev = dev;
9563 hdr->b_l2hdr.b_hits = 0;
9565 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
9566 hdr->b_l2hdr.b_arcs_state =
9567 hdr->b_l1hdr.b_state->arcs_state;
9568 arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR);
9570 mutex_enter(&dev->l2ad_mtx);
9571 list_insert_head(&dev->l2ad_buflist, hdr);
9572 mutex_exit(&dev->l2ad_mtx);
9574 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
9575 arc_hdr_size(hdr), hdr);
9577 wzio = zio_write_phys(pio, dev->l2ad_vdev,
9578 hdr->b_l2hdr.b_daddr, asize, to_write,
9579 ZIO_CHECKSUM_OFF, NULL, hdr,
9580 ZIO_PRIORITY_ASYNC_WRITE,
9581 ZIO_FLAG_CANFAIL, B_FALSE);
9583 write_lsize += HDR_GET_LSIZE(hdr);
9584 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
9587 write_psize += psize;
9588 write_asize += asize;
9589 dev->l2ad_hand += asize;
9590 l2arc_hdr_arcstats_increment(hdr);
9591 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9593 mutex_exit(hash_lock);
9596 * Append buf info to current log and commit if full.
9597 * arcstat_l2_{size,asize} kstats are updated
9600 if (l2arc_log_blk_insert(dev, hdr))
9601 l2arc_log_blk_commit(dev, pio, cb);
9606 multilist_sublist_unlock(mls);
9612 /* No buffers selected for writing? */
9614 ASSERT0(write_lsize);
9615 ASSERT(!HDR_HAS_L1HDR(head));
9616 kmem_cache_free(hdr_l2only_cache, head);
9619 * Although we did not write any buffers l2ad_evict may
9622 if (dev->l2ad_evict != l2dhdr->dh_evict)
9623 l2arc_dev_hdr_update(dev);
9628 if (!dev->l2ad_first)
9629 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
9631 ASSERT3U(write_asize, <=, target_sz);
9632 ARCSTAT_BUMP(arcstat_l2_writes_sent);
9633 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
9635 dev->l2ad_writing = B_TRUE;
9636 (void) zio_wait(pio);
9637 dev->l2ad_writing = B_FALSE;
9640 * Update the device header after the zio completes as
9641 * l2arc_write_done() may have updated the memory holding the log block
9642 * pointers in the device header.
9644 l2arc_dev_hdr_update(dev);
9646 return (write_asize);
9650 l2arc_hdr_limit_reached(void)
9652 int64_t s = aggsum_upper_bound(&arc_sums.arcstat_l2_hdr_size);
9654 return (arc_reclaim_needed() || (s > arc_meta_limit * 3 / 4) ||
9655 (s > (arc_warm ? arc_c : arc_c_max) * l2arc_meta_percent / 100));
9659 * This thread feeds the L2ARC at regular intervals. This is the beating
9660 * heart of the L2ARC.
9664 l2arc_feed_thread(void *unused)
9669 uint64_t size, wrote;
9670 clock_t begin, next = ddi_get_lbolt();
9671 fstrans_cookie_t cookie;
9673 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
9675 mutex_enter(&l2arc_feed_thr_lock);
9677 cookie = spl_fstrans_mark();
9678 while (l2arc_thread_exit == 0) {
9679 CALLB_CPR_SAFE_BEGIN(&cpr);
9680 (void) cv_timedwait_idle(&l2arc_feed_thr_cv,
9681 &l2arc_feed_thr_lock, next);
9682 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
9683 next = ddi_get_lbolt() + hz;
9686 * Quick check for L2ARC devices.
9688 mutex_enter(&l2arc_dev_mtx);
9689 if (l2arc_ndev == 0) {
9690 mutex_exit(&l2arc_dev_mtx);
9693 mutex_exit(&l2arc_dev_mtx);
9694 begin = ddi_get_lbolt();
9697 * This selects the next l2arc device to write to, and in
9698 * doing so the next spa to feed from: dev->l2ad_spa. This
9699 * will return NULL if there are now no l2arc devices or if
9700 * they are all faulted.
9702 * If a device is returned, its spa's config lock is also
9703 * held to prevent device removal. l2arc_dev_get_next()
9704 * will grab and release l2arc_dev_mtx.
9706 if ((dev = l2arc_dev_get_next()) == NULL)
9709 spa = dev->l2ad_spa;
9710 ASSERT3P(spa, !=, NULL);
9713 * If the pool is read-only then force the feed thread to
9714 * sleep a little longer.
9716 if (!spa_writeable(spa)) {
9717 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
9718 spa_config_exit(spa, SCL_L2ARC, dev);
9723 * Avoid contributing to memory pressure.
9725 if (l2arc_hdr_limit_reached()) {
9726 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
9727 spa_config_exit(spa, SCL_L2ARC, dev);
9731 ARCSTAT_BUMP(arcstat_l2_feeds);
9733 size = l2arc_write_size(dev);
9736 * Evict L2ARC buffers that will be overwritten.
9738 l2arc_evict(dev, size, B_FALSE);
9741 * Write ARC buffers.
9743 wrote = l2arc_write_buffers(spa, dev, size);
9746 * Calculate interval between writes.
9748 next = l2arc_write_interval(begin, size, wrote);
9749 spa_config_exit(spa, SCL_L2ARC, dev);
9751 spl_fstrans_unmark(cookie);
9753 l2arc_thread_exit = 0;
9754 cv_broadcast(&l2arc_feed_thr_cv);
9755 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
9760 l2arc_vdev_present(vdev_t *vd)
9762 return (l2arc_vdev_get(vd) != NULL);
9766 * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if
9767 * the vdev_t isn't an L2ARC device.
9770 l2arc_vdev_get(vdev_t *vd)
9774 mutex_enter(&l2arc_dev_mtx);
9775 for (dev = list_head(l2arc_dev_list); dev != NULL;
9776 dev = list_next(l2arc_dev_list, dev)) {
9777 if (dev->l2ad_vdev == vd)
9780 mutex_exit(&l2arc_dev_mtx);
9786 l2arc_rebuild_dev(l2arc_dev_t *dev, boolean_t reopen)
9788 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9789 uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9790 spa_t *spa = dev->l2ad_spa;
9793 * The L2ARC has to hold at least the payload of one log block for
9794 * them to be restored (persistent L2ARC). The payload of a log block
9795 * depends on the amount of its log entries. We always write log blocks
9796 * with 1022 entries. How many of them are committed or restored depends
9797 * on the size of the L2ARC device. Thus the maximum payload of
9798 * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device
9799 * is less than that, we reduce the amount of committed and restored
9800 * log entries per block so as to enable persistence.
9802 if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) {
9803 dev->l2ad_log_entries = 0;
9805 dev->l2ad_log_entries = MIN((dev->l2ad_end -
9806 dev->l2ad_start) >> SPA_MAXBLOCKSHIFT,
9807 L2ARC_LOG_BLK_MAX_ENTRIES);
9811 * Read the device header, if an error is returned do not rebuild L2ARC.
9813 if (l2arc_dev_hdr_read(dev) == 0 && dev->l2ad_log_entries > 0) {
9815 * If we are onlining a cache device (vdev_reopen) that was
9816 * still present (l2arc_vdev_present()) and rebuild is enabled,
9817 * we should evict all ARC buffers and pointers to log blocks
9818 * and reclaim their space before restoring its contents to
9822 if (!l2arc_rebuild_enabled) {
9825 l2arc_evict(dev, 0, B_TRUE);
9826 /* start a new log block */
9827 dev->l2ad_log_ent_idx = 0;
9828 dev->l2ad_log_blk_payload_asize = 0;
9829 dev->l2ad_log_blk_payload_start = 0;
9833 * Just mark the device as pending for a rebuild. We won't
9834 * be starting a rebuild in line here as it would block pool
9835 * import. Instead spa_load_impl will hand that off to an
9836 * async task which will call l2arc_spa_rebuild_start.
9838 dev->l2ad_rebuild = B_TRUE;
9839 } else if (spa_writeable(spa)) {
9841 * In this case TRIM the whole device if l2arc_trim_ahead > 0,
9842 * otherwise create a new header. We zero out the memory holding
9843 * the header to reset dh_start_lbps. If we TRIM the whole
9844 * device the new header will be written by
9845 * vdev_trim_l2arc_thread() at the end of the TRIM to update the
9846 * trim_state in the header too. When reading the header, if
9847 * trim_state is not VDEV_TRIM_COMPLETE and l2arc_trim_ahead > 0
9848 * we opt to TRIM the whole device again.
9850 if (l2arc_trim_ahead > 0) {
9851 dev->l2ad_trim_all = B_TRUE;
9853 bzero(l2dhdr, l2dhdr_asize);
9854 l2arc_dev_hdr_update(dev);
9860 * Add a vdev for use by the L2ARC. By this point the spa has already
9861 * validated the vdev and opened it.
9864 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
9866 l2arc_dev_t *adddev;
9867 uint64_t l2dhdr_asize;
9869 ASSERT(!l2arc_vdev_present(vd));
9872 * Create a new l2arc device entry.
9874 adddev = vmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
9875 adddev->l2ad_spa = spa;
9876 adddev->l2ad_vdev = vd;
9877 /* leave extra size for an l2arc device header */
9878 l2dhdr_asize = adddev->l2ad_dev_hdr_asize =
9879 MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift);
9880 adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize;
9881 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
9882 ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end);
9883 adddev->l2ad_hand = adddev->l2ad_start;
9884 adddev->l2ad_evict = adddev->l2ad_start;
9885 adddev->l2ad_first = B_TRUE;
9886 adddev->l2ad_writing = B_FALSE;
9887 adddev->l2ad_trim_all = B_FALSE;
9888 list_link_init(&adddev->l2ad_node);
9889 adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP);
9891 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
9893 * This is a list of all ARC buffers that are still valid on the
9896 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
9897 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
9900 * This is a list of pointers to log blocks that are still present
9903 list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t),
9904 offsetof(l2arc_lb_ptr_buf_t, node));
9906 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
9907 zfs_refcount_create(&adddev->l2ad_alloc);
9908 zfs_refcount_create(&adddev->l2ad_lb_asize);
9909 zfs_refcount_create(&adddev->l2ad_lb_count);
9912 * Decide if dev is eligible for L2ARC rebuild or whole device
9913 * trimming. This has to happen before the device is added in the
9914 * cache device list and l2arc_dev_mtx is released. Otherwise
9915 * l2arc_feed_thread() might already start writing on the
9918 l2arc_rebuild_dev(adddev, B_FALSE);
9921 * Add device to global list
9923 mutex_enter(&l2arc_dev_mtx);
9924 list_insert_head(l2arc_dev_list, adddev);
9925 atomic_inc_64(&l2arc_ndev);
9926 mutex_exit(&l2arc_dev_mtx);
9930 * Decide if a vdev is eligible for L2ARC rebuild, called from vdev_reopen()
9931 * in case of onlining a cache device.
9934 l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen)
9936 l2arc_dev_t *dev = NULL;
9938 dev = l2arc_vdev_get(vd);
9939 ASSERT3P(dev, !=, NULL);
9942 * In contrast to l2arc_add_vdev() we do not have to worry about
9943 * l2arc_feed_thread() invalidating previous content when onlining a
9944 * cache device. The device parameters (l2ad*) are not cleared when
9945 * offlining the device and writing new buffers will not invalidate
9946 * all previous content. In worst case only buffers that have not had
9947 * their log block written to the device will be lost.
9948 * When onlining the cache device (ie offline->online without exporting
9949 * the pool in between) this happens:
9950 * vdev_reopen() -> vdev_open() -> l2arc_rebuild_vdev()
9952 * vdev_is_dead() = B_FALSE l2ad_rebuild = B_TRUE
9953 * During the time where vdev_is_dead = B_FALSE and until l2ad_rebuild
9954 * is set to B_TRUE we might write additional buffers to the device.
9956 l2arc_rebuild_dev(dev, reopen);
9960 * Remove a vdev from the L2ARC.
9963 l2arc_remove_vdev(vdev_t *vd)
9965 l2arc_dev_t *remdev = NULL;
9968 * Find the device by vdev
9970 remdev = l2arc_vdev_get(vd);
9971 ASSERT3P(remdev, !=, NULL);
9974 * Cancel any ongoing or scheduled rebuild.
9976 mutex_enter(&l2arc_rebuild_thr_lock);
9977 if (remdev->l2ad_rebuild_began == B_TRUE) {
9978 remdev->l2ad_rebuild_cancel = B_TRUE;
9979 while (remdev->l2ad_rebuild == B_TRUE)
9980 cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock);
9982 mutex_exit(&l2arc_rebuild_thr_lock);
9985 * Remove device from global list
9987 mutex_enter(&l2arc_dev_mtx);
9988 list_remove(l2arc_dev_list, remdev);
9989 l2arc_dev_last = NULL; /* may have been invalidated */
9990 atomic_dec_64(&l2arc_ndev);
9991 mutex_exit(&l2arc_dev_mtx);
9994 * Clear all buflists and ARC references. L2ARC device flush.
9996 l2arc_evict(remdev, 0, B_TRUE);
9997 list_destroy(&remdev->l2ad_buflist);
9998 ASSERT(list_is_empty(&remdev->l2ad_lbptr_list));
9999 list_destroy(&remdev->l2ad_lbptr_list);
10000 mutex_destroy(&remdev->l2ad_mtx);
10001 zfs_refcount_destroy(&remdev->l2ad_alloc);
10002 zfs_refcount_destroy(&remdev->l2ad_lb_asize);
10003 zfs_refcount_destroy(&remdev->l2ad_lb_count);
10004 kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize);
10005 vmem_free(remdev, sizeof (l2arc_dev_t));
10011 l2arc_thread_exit = 0;
10014 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
10015 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
10016 mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL);
10017 cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL);
10018 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
10019 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
10021 l2arc_dev_list = &L2ARC_dev_list;
10022 l2arc_free_on_write = &L2ARC_free_on_write;
10023 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
10024 offsetof(l2arc_dev_t, l2ad_node));
10025 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
10026 offsetof(l2arc_data_free_t, l2df_list_node));
10032 mutex_destroy(&l2arc_feed_thr_lock);
10033 cv_destroy(&l2arc_feed_thr_cv);
10034 mutex_destroy(&l2arc_rebuild_thr_lock);
10035 cv_destroy(&l2arc_rebuild_thr_cv);
10036 mutex_destroy(&l2arc_dev_mtx);
10037 mutex_destroy(&l2arc_free_on_write_mtx);
10039 list_destroy(l2arc_dev_list);
10040 list_destroy(l2arc_free_on_write);
10046 if (!(spa_mode_global & SPA_MODE_WRITE))
10049 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
10050 TS_RUN, defclsyspri);
10056 if (!(spa_mode_global & SPA_MODE_WRITE))
10059 mutex_enter(&l2arc_feed_thr_lock);
10060 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
10061 l2arc_thread_exit = 1;
10062 while (l2arc_thread_exit != 0)
10063 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
10064 mutex_exit(&l2arc_feed_thr_lock);
10068 * Punches out rebuild threads for the L2ARC devices in a spa. This should
10069 * be called after pool import from the spa async thread, since starting
10070 * these threads directly from spa_import() will make them part of the
10071 * "zpool import" context and delay process exit (and thus pool import).
10074 l2arc_spa_rebuild_start(spa_t *spa)
10076 ASSERT(MUTEX_HELD(&spa_namespace_lock));
10079 * Locate the spa's l2arc devices and kick off rebuild threads.
10081 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
10083 l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]);
10085 /* Don't attempt a rebuild if the vdev is UNAVAIL */
10088 mutex_enter(&l2arc_rebuild_thr_lock);
10089 if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) {
10090 dev->l2ad_rebuild_began = B_TRUE;
10091 (void) thread_create(NULL, 0, l2arc_dev_rebuild_thread,
10092 dev, 0, &p0, TS_RUN, minclsyspri);
10094 mutex_exit(&l2arc_rebuild_thr_lock);
10099 * Main entry point for L2ARC rebuilding.
10102 l2arc_dev_rebuild_thread(void *arg)
10104 l2arc_dev_t *dev = arg;
10106 VERIFY(!dev->l2ad_rebuild_cancel);
10107 VERIFY(dev->l2ad_rebuild);
10108 (void) l2arc_rebuild(dev);
10109 mutex_enter(&l2arc_rebuild_thr_lock);
10110 dev->l2ad_rebuild_began = B_FALSE;
10111 dev->l2ad_rebuild = B_FALSE;
10112 mutex_exit(&l2arc_rebuild_thr_lock);
10118 * This function implements the actual L2ARC metadata rebuild. It:
10119 * starts reading the log block chain and restores each block's contents
10120 * to memory (reconstructing arc_buf_hdr_t's).
10122 * Operation stops under any of the following conditions:
10124 * 1) We reach the end of the log block chain.
10125 * 2) We encounter *any* error condition (cksum errors, io errors)
10128 l2arc_rebuild(l2arc_dev_t *dev)
10130 vdev_t *vd = dev->l2ad_vdev;
10131 spa_t *spa = vd->vdev_spa;
10133 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10134 l2arc_log_blk_phys_t *this_lb, *next_lb;
10135 zio_t *this_io = NULL, *next_io = NULL;
10136 l2arc_log_blkptr_t lbps[2];
10137 l2arc_lb_ptr_buf_t *lb_ptr_buf;
10138 boolean_t lock_held;
10140 this_lb = vmem_zalloc(sizeof (*this_lb), KM_SLEEP);
10141 next_lb = vmem_zalloc(sizeof (*next_lb), KM_SLEEP);
10144 * We prevent device removal while issuing reads to the device,
10145 * then during the rebuilding phases we drop this lock again so
10146 * that a spa_unload or device remove can be initiated - this is
10147 * safe, because the spa will signal us to stop before removing
10148 * our device and wait for us to stop.
10150 spa_config_enter(spa, SCL_L2ARC, vd, RW_READER);
10151 lock_held = B_TRUE;
10154 * Retrieve the persistent L2ARC device state.
10155 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10157 dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start);
10158 dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr +
10159 L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop),
10161 dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST);
10163 vd->vdev_trim_action_time = l2dhdr->dh_trim_action_time;
10164 vd->vdev_trim_state = l2dhdr->dh_trim_state;
10167 * In case the zfs module parameter l2arc_rebuild_enabled is false
10168 * we do not start the rebuild process.
10170 if (!l2arc_rebuild_enabled)
10173 /* Prepare the rebuild process */
10174 bcopy(l2dhdr->dh_start_lbps, lbps, sizeof (lbps));
10176 /* Start the rebuild process */
10178 if (!l2arc_log_blkptr_valid(dev, &lbps[0]))
10181 if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1],
10182 this_lb, next_lb, this_io, &next_io)) != 0)
10186 * Our memory pressure valve. If the system is running low
10187 * on memory, rather than swamping memory with new ARC buf
10188 * hdrs, we opt not to rebuild the L2ARC. At this point,
10189 * however, we have already set up our L2ARC dev to chain in
10190 * new metadata log blocks, so the user may choose to offline/
10191 * online the L2ARC dev at a later time (or re-import the pool)
10192 * to reconstruct it (when there's less memory pressure).
10194 if (l2arc_hdr_limit_reached()) {
10195 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem);
10196 cmn_err(CE_NOTE, "System running low on memory, "
10197 "aborting L2ARC rebuild.");
10198 err = SET_ERROR(ENOMEM);
10202 spa_config_exit(spa, SCL_L2ARC, vd);
10203 lock_held = B_FALSE;
10206 * Now that we know that the next_lb checks out alright, we
10207 * can start reconstruction from this log block.
10208 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10210 uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop);
10211 l2arc_log_blk_restore(dev, this_lb, asize);
10214 * log block restored, include its pointer in the list of
10215 * pointers to log blocks present in the L2ARC device.
10217 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
10218 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t),
10220 bcopy(&lbps[0], lb_ptr_buf->lb_ptr,
10221 sizeof (l2arc_log_blkptr_t));
10222 mutex_enter(&dev->l2ad_mtx);
10223 list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf);
10224 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10225 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10226 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10227 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10228 mutex_exit(&dev->l2ad_mtx);
10229 vdev_space_update(vd, asize, 0, 0);
10232 * Protection against loops of log blocks:
10234 * l2ad_hand l2ad_evict
10236 * l2ad_start |=======================================| l2ad_end
10237 * -----|||----|||---|||----|||
10239 * ---|||---|||----|||---|||
10242 * In this situation the pointer of log block (4) passes
10243 * l2arc_log_blkptr_valid() but the log block should not be
10244 * restored as it is overwritten by the payload of log block
10245 * (0). Only log blocks (0)-(3) should be restored. We check
10246 * whether l2ad_evict lies in between the payload starting
10247 * offset of the next log block (lbps[1].lbp_payload_start)
10248 * and the payload starting offset of the present log block
10249 * (lbps[0].lbp_payload_start). If true and this isn't the
10250 * first pass, we are looping from the beginning and we should
10253 if (l2arc_range_check_overlap(lbps[1].lbp_payload_start,
10254 lbps[0].lbp_payload_start, dev->l2ad_evict) &&
10260 mutex_enter(&l2arc_rebuild_thr_lock);
10261 if (dev->l2ad_rebuild_cancel) {
10262 dev->l2ad_rebuild = B_FALSE;
10263 cv_signal(&l2arc_rebuild_thr_cv);
10264 mutex_exit(&l2arc_rebuild_thr_lock);
10265 err = SET_ERROR(ECANCELED);
10268 mutex_exit(&l2arc_rebuild_thr_lock);
10269 if (spa_config_tryenter(spa, SCL_L2ARC, vd,
10271 lock_held = B_TRUE;
10275 * L2ARC config lock held by somebody in writer,
10276 * possibly due to them trying to remove us. They'll
10277 * likely to want us to shut down, so after a little
10278 * delay, we check l2ad_rebuild_cancel and retry
10285 * Continue with the next log block.
10288 lbps[1] = this_lb->lb_prev_lbp;
10289 PTR_SWAP(this_lb, next_lb);
10294 if (this_io != NULL)
10295 l2arc_log_blk_fetch_abort(this_io);
10297 if (next_io != NULL)
10298 l2arc_log_blk_fetch_abort(next_io);
10299 vmem_free(this_lb, sizeof (*this_lb));
10300 vmem_free(next_lb, sizeof (*next_lb));
10302 if (!l2arc_rebuild_enabled) {
10303 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10305 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) {
10306 ARCSTAT_BUMP(arcstat_l2_rebuild_success);
10307 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10308 "successful, restored %llu blocks",
10309 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10310 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) {
10312 * No error but also nothing restored, meaning the lbps array
10313 * in the device header points to invalid/non-present log
10314 * blocks. Reset the header.
10316 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10317 "no valid log blocks");
10318 bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
10319 l2arc_dev_hdr_update(dev);
10320 } else if (err == ECANCELED) {
10322 * In case the rebuild was canceled do not log to spa history
10323 * log as the pool may be in the process of being removed.
10325 zfs_dbgmsg("L2ARC rebuild aborted, restored %llu blocks",
10326 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10327 } else if (err != 0) {
10328 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10329 "aborted, restored %llu blocks",
10330 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10334 spa_config_exit(spa, SCL_L2ARC, vd);
10340 * Attempts to read the device header on the provided L2ARC device and writes
10341 * it to `hdr'. On success, this function returns 0, otherwise the appropriate
10342 * error code is returned.
10345 l2arc_dev_hdr_read(l2arc_dev_t *dev)
10349 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10350 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10353 guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10355 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10357 err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev,
10358 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd,
10359 ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
10360 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
10361 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY |
10362 ZIO_FLAG_SPECULATIVE, B_FALSE));
10367 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors);
10368 zfs_dbgmsg("L2ARC IO error (%d) while reading device header, "
10369 "vdev guid: %llu", err,
10370 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10374 if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC))
10375 byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr));
10377 if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC ||
10378 l2dhdr->dh_spa_guid != guid ||
10379 l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid ||
10380 l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION ||
10381 l2dhdr->dh_log_entries != dev->l2ad_log_entries ||
10382 l2dhdr->dh_end != dev->l2ad_end ||
10383 !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end,
10384 l2dhdr->dh_evict) ||
10385 (l2dhdr->dh_trim_state != VDEV_TRIM_COMPLETE &&
10386 l2arc_trim_ahead > 0)) {
10388 * Attempt to rebuild a device containing no actual dev hdr
10389 * or containing a header from some other pool or from another
10390 * version of persistent L2ARC.
10392 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported);
10393 return (SET_ERROR(ENOTSUP));
10400 * Reads L2ARC log blocks from storage and validates their contents.
10402 * This function implements a simple fetcher to make sure that while
10403 * we're processing one buffer the L2ARC is already fetching the next
10404 * one in the chain.
10406 * The arguments this_lp and next_lp point to the current and next log block
10407 * address in the block chain. Similarly, this_lb and next_lb hold the
10408 * l2arc_log_blk_phys_t's of the current and next L2ARC blk.
10410 * The `this_io' and `next_io' arguments are used for block fetching.
10411 * When issuing the first blk IO during rebuild, you should pass NULL for
10412 * `this_io'. This function will then issue a sync IO to read the block and
10413 * also issue an async IO to fetch the next block in the block chain. The
10414 * fetched IO is returned in `next_io'. On subsequent calls to this
10415 * function, pass the value returned in `next_io' from the previous call
10416 * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO.
10417 * Prior to the call, you should initialize your `next_io' pointer to be
10418 * NULL. If no fetch IO was issued, the pointer is left set at NULL.
10420 * On success, this function returns 0, otherwise it returns an appropriate
10421 * error code. On error the fetching IO is aborted and cleared before
10422 * returning from this function. Therefore, if we return `success', the
10423 * caller can assume that we have taken care of cleanup of fetch IOs.
10426 l2arc_log_blk_read(l2arc_dev_t *dev,
10427 const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp,
10428 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
10429 zio_t *this_io, zio_t **next_io)
10436 ASSERT(this_lbp != NULL && next_lbp != NULL);
10437 ASSERT(this_lb != NULL && next_lb != NULL);
10438 ASSERT(next_io != NULL && *next_io == NULL);
10439 ASSERT(l2arc_log_blkptr_valid(dev, this_lbp));
10442 * Check to see if we have issued the IO for this log block in a
10443 * previous run. If not, this is the first call, so issue it now.
10445 if (this_io == NULL) {
10446 this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp,
10451 * Peek to see if we can start issuing the next IO immediately.
10453 if (l2arc_log_blkptr_valid(dev, next_lbp)) {
10455 * Start issuing IO for the next log block early - this
10456 * should help keep the L2ARC device busy while we
10457 * decompress and restore this log block.
10459 *next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp,
10463 /* Wait for the IO to read this log block to complete */
10464 if ((err = zio_wait(this_io)) != 0) {
10465 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
10466 zfs_dbgmsg("L2ARC IO error (%d) while reading log block, "
10467 "offset: %llu, vdev guid: %llu", err,
10468 (u_longlong_t)this_lbp->lbp_daddr,
10469 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10474 * Make sure the buffer checks out.
10475 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10477 asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop);
10478 fletcher_4_native(this_lb, asize, NULL, &cksum);
10479 if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) {
10480 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors);
10481 zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, "
10482 "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu",
10483 (u_longlong_t)this_lbp->lbp_daddr,
10484 (u_longlong_t)dev->l2ad_vdev->vdev_guid,
10485 (u_longlong_t)dev->l2ad_hand,
10486 (u_longlong_t)dev->l2ad_evict);
10487 err = SET_ERROR(ECKSUM);
10491 /* Now we can take our time decoding this buffer */
10492 switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) {
10493 case ZIO_COMPRESS_OFF:
10495 case ZIO_COMPRESS_LZ4:
10496 abd = abd_alloc_for_io(asize, B_TRUE);
10497 abd_copy_from_buf_off(abd, this_lb, 0, asize);
10498 if ((err = zio_decompress_data(
10499 L2BLK_GET_COMPRESS((this_lbp)->lbp_prop),
10500 abd, this_lb, asize, sizeof (*this_lb), NULL)) != 0) {
10501 err = SET_ERROR(EINVAL);
10506 err = SET_ERROR(EINVAL);
10509 if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC))
10510 byteswap_uint64_array(this_lb, sizeof (*this_lb));
10511 if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) {
10512 err = SET_ERROR(EINVAL);
10516 /* Abort an in-flight fetch I/O in case of error */
10517 if (err != 0 && *next_io != NULL) {
10518 l2arc_log_blk_fetch_abort(*next_io);
10527 * Restores the payload of a log block to ARC. This creates empty ARC hdr
10528 * entries which only contain an l2arc hdr, essentially restoring the
10529 * buffers to their L2ARC evicted state. This function also updates space
10530 * usage on the L2ARC vdev to make sure it tracks restored buffers.
10533 l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb,
10536 uint64_t size = 0, asize = 0;
10537 uint64_t log_entries = dev->l2ad_log_entries;
10540 * Usually arc_adapt() is called only for data, not headers, but
10541 * since we may allocate significant amount of memory here, let ARC
10544 arc_adapt(log_entries * HDR_L2ONLY_SIZE, arc_l2c_only);
10546 for (int i = log_entries - 1; i >= 0; i--) {
10548 * Restore goes in the reverse temporal direction to preserve
10549 * correct temporal ordering of buffers in the l2ad_buflist.
10550 * l2arc_hdr_restore also does a list_insert_tail instead of
10551 * list_insert_head on the l2ad_buflist:
10553 * LIST l2ad_buflist LIST
10554 * HEAD <------ (time) ------ TAIL
10555 * direction +-----+-----+-----+-----+-----+ direction
10556 * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild
10557 * fill +-----+-----+-----+-----+-----+
10561 * l2arc_feed_thread l2arc_rebuild
10562 * will place new bufs here restores bufs here
10564 * During l2arc_rebuild() the device is not used by
10565 * l2arc_feed_thread() as dev->l2ad_rebuild is set to true.
10567 size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop);
10568 asize += vdev_psize_to_asize(dev->l2ad_vdev,
10569 L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop));
10570 l2arc_hdr_restore(&lb->lb_entries[i], dev);
10574 * Record rebuild stats:
10575 * size Logical size of restored buffers in the L2ARC
10576 * asize Aligned size of restored buffers in the L2ARC
10578 ARCSTAT_INCR(arcstat_l2_rebuild_size, size);
10579 ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize);
10580 ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries);
10581 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize);
10582 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize);
10583 ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks);
10587 * Restores a single ARC buf hdr from a log entry. The ARC buffer is put
10588 * into a state indicating that it has been evicted to L2ARC.
10591 l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
10593 arc_buf_hdr_t *hdr, *exists;
10594 kmutex_t *hash_lock;
10595 arc_buf_contents_t type = L2BLK_GET_TYPE((le)->le_prop);
10599 * Do all the allocation before grabbing any locks, this lets us
10600 * sleep if memory is full and we don't have to deal with failed
10603 hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type,
10604 dev, le->le_dva, le->le_daddr,
10605 L2BLK_GET_PSIZE((le)->le_prop), le->le_birth,
10606 L2BLK_GET_COMPRESS((le)->le_prop), le->le_complevel,
10607 L2BLK_GET_PROTECTED((le)->le_prop),
10608 L2BLK_GET_PREFETCH((le)->le_prop),
10609 L2BLK_GET_STATE((le)->le_prop));
10610 asize = vdev_psize_to_asize(dev->l2ad_vdev,
10611 L2BLK_GET_PSIZE((le)->le_prop));
10614 * vdev_space_update() has to be called before arc_hdr_destroy() to
10615 * avoid underflow since the latter also calls vdev_space_update().
10617 l2arc_hdr_arcstats_increment(hdr);
10618 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10620 mutex_enter(&dev->l2ad_mtx);
10621 list_insert_tail(&dev->l2ad_buflist, hdr);
10622 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
10623 mutex_exit(&dev->l2ad_mtx);
10625 exists = buf_hash_insert(hdr, &hash_lock);
10627 /* Buffer was already cached, no need to restore it. */
10628 arc_hdr_destroy(hdr);
10630 * If the buffer is already cached, check whether it has
10631 * L2ARC metadata. If not, enter them and update the flag.
10632 * This is important is case of onlining a cache device, since
10633 * we previously evicted all L2ARC metadata from ARC.
10635 if (!HDR_HAS_L2HDR(exists)) {
10636 arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR);
10637 exists->b_l2hdr.b_dev = dev;
10638 exists->b_l2hdr.b_daddr = le->le_daddr;
10639 exists->b_l2hdr.b_arcs_state =
10640 L2BLK_GET_STATE((le)->le_prop);
10641 mutex_enter(&dev->l2ad_mtx);
10642 list_insert_tail(&dev->l2ad_buflist, exists);
10643 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
10644 arc_hdr_size(exists), exists);
10645 mutex_exit(&dev->l2ad_mtx);
10646 l2arc_hdr_arcstats_increment(exists);
10647 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10649 ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
10652 mutex_exit(hash_lock);
10656 * Starts an asynchronous read IO to read a log block. This is used in log
10657 * block reconstruction to start reading the next block before we are done
10658 * decoding and reconstructing the current block, to keep the l2arc device
10659 * nice and hot with read IO to process.
10660 * The returned zio will contain a newly allocated memory buffers for the IO
10661 * data which should then be freed by the caller once the zio is no longer
10662 * needed (i.e. due to it having completed). If you wish to abort this
10663 * zio, you should do so using l2arc_log_blk_fetch_abort, which takes
10664 * care of disposing of the allocated buffers correctly.
10667 l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp,
10668 l2arc_log_blk_phys_t *lb)
10672 l2arc_read_callback_t *cb;
10674 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10675 asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10676 ASSERT(asize <= sizeof (l2arc_log_blk_phys_t));
10678 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP);
10679 cb->l2rcb_abd = abd_get_from_buf(lb, asize);
10680 pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb,
10681 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
10682 ZIO_FLAG_DONT_RETRY);
10683 (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize,
10684 cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10685 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
10686 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
10692 * Aborts a zio returned from l2arc_log_blk_fetch and frees the data
10693 * buffers allocated for it.
10696 l2arc_log_blk_fetch_abort(zio_t *zio)
10698 (void) zio_wait(zio);
10702 * Creates a zio to update the device header on an l2arc device.
10705 l2arc_dev_hdr_update(l2arc_dev_t *dev)
10707 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10708 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10712 VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER));
10714 l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC;
10715 l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION;
10716 l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10717 l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid;
10718 l2dhdr->dh_log_entries = dev->l2ad_log_entries;
10719 l2dhdr->dh_evict = dev->l2ad_evict;
10720 l2dhdr->dh_start = dev->l2ad_start;
10721 l2dhdr->dh_end = dev->l2ad_end;
10722 l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize);
10723 l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count);
10724 l2dhdr->dh_flags = 0;
10725 l2dhdr->dh_trim_action_time = dev->l2ad_vdev->vdev_trim_action_time;
10726 l2dhdr->dh_trim_state = dev->l2ad_vdev->vdev_trim_state;
10727 if (dev->l2ad_first)
10728 l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST;
10730 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10732 err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev,
10733 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL,
10734 NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE));
10739 zfs_dbgmsg("L2ARC IO error (%d) while writing device header, "
10740 "vdev guid: %llu", err,
10741 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10746 * Commits a log block to the L2ARC device. This routine is invoked from
10747 * l2arc_write_buffers when the log block fills up.
10748 * This function allocates some memory to temporarily hold the serialized
10749 * buffer to be written. This is then released in l2arc_write_done.
10752 l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb)
10754 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10755 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10756 uint64_t psize, asize;
10758 l2arc_lb_abd_buf_t *abd_buf;
10760 l2arc_lb_ptr_buf_t *lb_ptr_buf;
10762 VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries);
10764 tmpbuf = zio_buf_alloc(sizeof (*lb));
10765 abd_buf = zio_buf_alloc(sizeof (*abd_buf));
10766 abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb));
10767 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
10768 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP);
10770 /* link the buffer into the block chain */
10771 lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1];
10772 lb->lb_magic = L2ARC_LOG_BLK_MAGIC;
10775 * l2arc_log_blk_commit() may be called multiple times during a single
10776 * l2arc_write_buffers() call. Save the allocated abd buffers in a list
10777 * so we can free them in l2arc_write_done() later on.
10779 list_insert_tail(&cb->l2wcb_abd_list, abd_buf);
10781 /* try to compress the buffer */
10782 psize = zio_compress_data(ZIO_COMPRESS_LZ4,
10783 abd_buf->abd, tmpbuf, sizeof (*lb), 0);
10785 /* a log block is never entirely zero */
10786 ASSERT(psize != 0);
10787 asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
10788 ASSERT(asize <= sizeof (*lb));
10791 * Update the start log block pointer in the device header to point
10792 * to the log block we're about to write.
10794 l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0];
10795 l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand;
10796 l2dhdr->dh_start_lbps[0].lbp_payload_asize =
10797 dev->l2ad_log_blk_payload_asize;
10798 l2dhdr->dh_start_lbps[0].lbp_payload_start =
10799 dev->l2ad_log_blk_payload_start;
10801 (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb));
10803 (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize);
10804 L2BLK_SET_CHECKSUM(
10805 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10806 ZIO_CHECKSUM_FLETCHER_4);
10807 if (asize < sizeof (*lb)) {
10808 /* compression succeeded */
10809 bzero(tmpbuf + psize, asize - psize);
10810 L2BLK_SET_COMPRESS(
10811 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10814 /* compression failed */
10815 bcopy(lb, tmpbuf, sizeof (*lb));
10816 L2BLK_SET_COMPRESS(
10817 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10821 /* checksum what we're about to write */
10822 fletcher_4_native(tmpbuf, asize, NULL,
10823 &l2dhdr->dh_start_lbps[0].lbp_cksum);
10825 abd_free(abd_buf->abd);
10827 /* perform the write itself */
10828 abd_buf->abd = abd_get_from_buf(tmpbuf, sizeof (*lb));
10829 abd_take_ownership_of_buf(abd_buf->abd, B_TRUE);
10830 wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand,
10831 asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10832 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
10833 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
10834 (void) zio_nowait(wzio);
10836 dev->l2ad_hand += asize;
10838 * Include the committed log block's pointer in the list of pointers
10839 * to log blocks present in the L2ARC device.
10841 bcopy(&l2dhdr->dh_start_lbps[0], lb_ptr_buf->lb_ptr,
10842 sizeof (l2arc_log_blkptr_t));
10843 mutex_enter(&dev->l2ad_mtx);
10844 list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf);
10845 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10846 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10847 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10848 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10849 mutex_exit(&dev->l2ad_mtx);
10850 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10852 /* bump the kstats */
10853 ARCSTAT_INCR(arcstat_l2_write_bytes, asize);
10854 ARCSTAT_BUMP(arcstat_l2_log_blk_writes);
10855 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize);
10856 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio,
10857 dev->l2ad_log_blk_payload_asize / asize);
10859 /* start a new log block */
10860 dev->l2ad_log_ent_idx = 0;
10861 dev->l2ad_log_blk_payload_asize = 0;
10862 dev->l2ad_log_blk_payload_start = 0;
10866 * Validates an L2ARC log block address to make sure that it can be read
10867 * from the provided L2ARC device.
10870 l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp)
10872 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10873 uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10874 uint64_t end = lbp->lbp_daddr + asize - 1;
10875 uint64_t start = lbp->lbp_payload_start;
10876 boolean_t evicted = B_FALSE;
10879 * A log block is valid if all of the following conditions are true:
10880 * - it fits entirely (including its payload) between l2ad_start and
10882 * - it has a valid size
10883 * - neither the log block itself nor part of its payload was evicted
10884 * by l2arc_evict():
10886 * l2ad_hand l2ad_evict
10891 * l2ad_start ============================================ l2ad_end
10892 * --------------------------||||
10899 l2arc_range_check_overlap(start, end, dev->l2ad_hand) ||
10900 l2arc_range_check_overlap(start, end, dev->l2ad_evict) ||
10901 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) ||
10902 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end);
10904 return (start >= dev->l2ad_start && end <= dev->l2ad_end &&
10905 asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) &&
10906 (!evicted || dev->l2ad_first));
10910 * Inserts ARC buffer header `hdr' into the current L2ARC log block on
10911 * the device. The buffer being inserted must be present in L2ARC.
10912 * Returns B_TRUE if the L2ARC log block is full and needs to be committed
10913 * to L2ARC, or B_FALSE if it still has room for more ARC buffers.
10916 l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr)
10918 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10919 l2arc_log_ent_phys_t *le;
10921 if (dev->l2ad_log_entries == 0)
10924 int index = dev->l2ad_log_ent_idx++;
10926 ASSERT3S(index, <, dev->l2ad_log_entries);
10927 ASSERT(HDR_HAS_L2HDR(hdr));
10929 le = &lb->lb_entries[index];
10930 bzero(le, sizeof (*le));
10931 le->le_dva = hdr->b_dva;
10932 le->le_birth = hdr->b_birth;
10933 le->le_daddr = hdr->b_l2hdr.b_daddr;
10935 dev->l2ad_log_blk_payload_start = le->le_daddr;
10936 L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr));
10937 L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr));
10938 L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr));
10939 le->le_complevel = hdr->b_complevel;
10940 L2BLK_SET_TYPE((le)->le_prop, hdr->b_type);
10941 L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr)));
10942 L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr)));
10943 L2BLK_SET_STATE((le)->le_prop, hdr->b_l1hdr.b_state->arcs_state);
10945 dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev,
10946 HDR_GET_PSIZE(hdr));
10948 return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries);
10952 * Checks whether a given L2ARC device address sits in a time-sequential
10953 * range. The trick here is that the L2ARC is a rotary buffer, so we can't
10954 * just do a range comparison, we need to handle the situation in which the
10955 * range wraps around the end of the L2ARC device. Arguments:
10956 * bottom -- Lower end of the range to check (written to earlier).
10957 * top -- Upper end of the range to check (written to later).
10958 * check -- The address for which we want to determine if it sits in
10959 * between the top and bottom.
10961 * The 3-way conditional below represents the following cases:
10963 * bottom < top : Sequentially ordered case:
10964 * <check>--------+-------------------+
10965 * | (overlap here?) |
10967 * |---------------<bottom>============<top>--------------|
10969 * bottom > top: Looped-around case:
10970 * <check>--------+------------------+
10971 * | (overlap here?) |
10973 * |===============<top>---------------<bottom>===========|
10976 * +---------------+---------<check>
10978 * top == bottom : Just a single address comparison.
10981 l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check)
10984 return (bottom <= check && check <= top);
10985 else if (bottom > top)
10986 return (check <= top || bottom <= check);
10988 return (check == top);
10991 EXPORT_SYMBOL(arc_buf_size);
10992 EXPORT_SYMBOL(arc_write);
10993 EXPORT_SYMBOL(arc_read);
10994 EXPORT_SYMBOL(arc_buf_info);
10995 EXPORT_SYMBOL(arc_getbuf_func);
10996 EXPORT_SYMBOL(arc_add_prune_callback);
10997 EXPORT_SYMBOL(arc_remove_prune_callback);
10999 /* BEGIN CSTYLED */
11000 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min, param_set_arc_min,
11001 param_get_long, ZMOD_RW, "Min arc size");
11003 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, max, param_set_arc_max,
11004 param_get_long, ZMOD_RW, "Max arc size");
11006 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit, param_set_arc_long,
11007 param_get_long, ZMOD_RW, "Metadata limit for arc size");
11009 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit_percent,
11010 param_set_arc_long, param_get_long, ZMOD_RW,
11011 "Percent of arc size for arc meta limit");
11013 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_min, param_set_arc_long,
11014 param_get_long, ZMOD_RW, "Min arc metadata");
11016 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_prune, INT, ZMOD_RW,
11017 "Meta objects to scan for prune");
11019 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_adjust_restarts, INT, ZMOD_RW,
11020 "Limit number of restarts in arc_evict_meta");
11022 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_strategy, INT, ZMOD_RW,
11023 "Meta reclaim strategy");
11025 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, grow_retry, param_set_arc_int,
11026 param_get_int, ZMOD_RW, "Seconds before growing arc size");
11028 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, p_dampener_disable, INT, ZMOD_RW,
11029 "Disable arc_p adapt dampener");
11031 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, shrink_shift, param_set_arc_int,
11032 param_get_int, ZMOD_RW, "log2(fraction of arc to reclaim)");
11034 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, pc_percent, UINT, ZMOD_RW,
11035 "Percent of pagecache to reclaim arc to");
11037 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, p_min_shift, param_set_arc_int,
11038 param_get_int, ZMOD_RW, "arc_c shift to calc min/max arc_p");
11040 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, average_blocksize, INT, ZMOD_RD,
11041 "Target average block size");
11043 ZFS_MODULE_PARAM(zfs, zfs_, compressed_arc_enabled, INT, ZMOD_RW,
11044 "Disable compressed arc buffers");
11046 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prefetch_ms, param_set_arc_int,
11047 param_get_int, ZMOD_RW, "Min life of prefetch block in ms");
11049 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prescient_prefetch_ms,
11050 param_set_arc_int, param_get_int, ZMOD_RW,
11051 "Min life of prescient prefetched block in ms");
11053 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_max, ULONG, ZMOD_RW,
11054 "Max write bytes per interval");
11056 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_boost, ULONG, ZMOD_RW,
11057 "Extra write bytes during device warmup");
11059 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom, ULONG, ZMOD_RW,
11060 "Number of max device writes to precache");
11062 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom_boost, ULONG, ZMOD_RW,
11063 "Compressed l2arc_headroom multiplier");
11065 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, trim_ahead, ULONG, ZMOD_RW,
11066 "TRIM ahead L2ARC write size multiplier");
11068 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_secs, ULONG, ZMOD_RW,
11069 "Seconds between L2ARC writing");
11071 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_min_ms, ULONG, ZMOD_RW,
11072 "Min feed interval in milliseconds");
11074 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, noprefetch, INT, ZMOD_RW,
11075 "Skip caching prefetched buffers");
11077 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_again, INT, ZMOD_RW,
11078 "Turbo L2ARC warmup");
11080 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, norw, INT, ZMOD_RW,
11081 "No reads during writes");
11083 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, meta_percent, INT, ZMOD_RW,
11084 "Percent of ARC size allowed for L2ARC-only headers");
11086 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_enabled, INT, ZMOD_RW,
11087 "Rebuild the L2ARC when importing a pool");
11089 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_blocks_min_l2size, ULONG, ZMOD_RW,
11090 "Min size in bytes to write rebuild log blocks in L2ARC");
11092 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, mfuonly, INT, ZMOD_RW,
11093 "Cache only MFU data from ARC into L2ARC");
11095 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, lotsfree_percent, param_set_arc_int,
11096 param_get_int, ZMOD_RW, "System free memory I/O throttle in bytes");
11098 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, sys_free, param_set_arc_long,
11099 param_get_long, ZMOD_RW, "System free memory target size in bytes");
11101 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit, param_set_arc_long,
11102 param_get_long, ZMOD_RW, "Minimum bytes of dnodes in arc");
11104 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit_percent,
11105 param_set_arc_long, param_get_long, ZMOD_RW,
11106 "Percent of ARC meta buffers for dnodes");
11108 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, dnode_reduce_percent, ULONG, ZMOD_RW,
11109 "Percentage of excess dnodes to try to unpin");
11111 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, eviction_pct, INT, ZMOD_RW,
11112 "When full, ARC allocation waits for eviction of this % of alloc size");
11114 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, evict_batch_limit, INT, ZMOD_RW,
11115 "The number of headers to evict per sublist before moving to the next");