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) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
30 * DVA-based Adjustable Replacement Cache
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal ARC algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
96 * Each ARC state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an ARC list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
102 * It as also possible to register a callback which is run when the
103 * arc_meta_limit is reached and no buffers can be safely evicted. In
104 * this case the arc user should drop a reference on some arc buffers so
105 * they can be reclaimed and the arc_meta_limit honored. For example,
106 * when using the ZPL each dentry holds a references on a znode. These
107 * dentries must be pruned before the arc buffer holding the znode can
110 * Note that the majority of the performance stats are manipulated
111 * with atomic operations.
113 * The L2ARC uses the l2ad_mtx on each vdev for the following:
115 * - L2ARC buflist creation
116 * - L2ARC buflist eviction
117 * - L2ARC write completion, which walks L2ARC buflists
118 * - ARC header destruction, as it removes from L2ARC buflists
119 * - ARC header release, as it removes from L2ARC buflists
125 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
126 * This structure can point either to a block that is still in the cache or to
127 * one that is only accessible in an L2 ARC device, or it can provide
128 * information about a block that was recently evicted. If a block is
129 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
130 * information to retrieve it from the L2ARC device. This information is
131 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
132 * that is in this state cannot access the data directly.
134 * Blocks that are actively being referenced or have not been evicted
135 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
136 * the arc_buf_hdr_t that will point to the data block in memory. A block can
137 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
138 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
139 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
141 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
142 * ability to store the physical data (b_pabd) associated with the DVA of the
143 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
144 * it will match its on-disk compression characteristics. This behavior can be
145 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
146 * compressed ARC functionality is disabled, the b_pabd will point to an
147 * uncompressed version of the on-disk data.
149 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
150 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
151 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
152 * consumer. The ARC will provide references to this data and will keep it
153 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
154 * data block and will evict any arc_buf_t that is no longer referenced. The
155 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
156 * "overhead_size" kstat.
158 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
159 * compressed form. The typical case is that consumers will want uncompressed
160 * data, and when that happens a new data buffer is allocated where the data is
161 * decompressed for them to use. Currently the only consumer who wants
162 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
163 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
164 * with the arc_buf_hdr_t.
166 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
167 * first one is owned by a compressed send consumer (and therefore references
168 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
169 * used by any other consumer (and has its own uncompressed copy of the data
184 * | b_buf +------------>+-----------+ arc_buf_t
185 * | b_pabd +-+ |b_next +---->+-----------+
186 * +-----------+ | |-----------| |b_next +-->NULL
187 * | |b_comp = T | +-----------+
188 * | |b_data +-+ |b_comp = F |
189 * | +-----------+ | |b_data +-+
190 * +->+------+ | +-----------+ |
192 * data | |<--------------+ | uncompressed
193 * +------+ compressed, | data
194 * shared +-->+------+
199 * When a consumer reads a block, the ARC must first look to see if the
200 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
201 * arc_buf_t and either copies uncompressed data into a new data buffer from an
202 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
203 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
204 * hdr is compressed and the desired compression characteristics of the
205 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
206 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
207 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
208 * be anywhere in the hdr's list.
210 * The diagram below shows an example of an uncompressed ARC hdr that is
211 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
212 * the last element in the buf list):
224 * | | arc_buf_t (shared)
225 * | b_buf +------------>+---------+ arc_buf_t
226 * | | |b_next +---->+---------+
227 * | b_pabd +-+ |---------| |b_next +-->NULL
228 * +-----------+ | | | +---------+
230 * | +---------+ | |b_data +-+
231 * +->+------+ | +---------+ |
233 * uncompressed | | | |
236 * | uncompressed | | |
239 * +---------------------------------+
241 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
242 * since the physical block is about to be rewritten. The new data contents
243 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
244 * it may compress the data before writing it to disk. The ARC will be called
245 * with the transformed data and will bcopy the transformed on-disk block into
246 * a newly allocated b_pabd. Writes are always done into buffers which have
247 * either been loaned (and hence are new and don't have other readers) or
248 * buffers which have been released (and hence have their own hdr, if there
249 * were originally other readers of the buf's original hdr). This ensures that
250 * the ARC only needs to update a single buf and its hdr after a write occurs.
252 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
253 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
254 * that when compressed ARC is enabled that the L2ARC blocks are identical
255 * to the on-disk block in the main data pool. This provides a significant
256 * advantage since the ARC can leverage the bp's checksum when reading from the
257 * L2ARC to determine if the contents are valid. However, if the compressed
258 * ARC is disabled, then the L2ARC's block must be transformed to look
259 * like the physical block in the main data pool before comparing the
260 * checksum and determining its validity.
265 #include <sys/spa_impl.h>
266 #include <sys/zio_compress.h>
267 #include <sys/zio_checksum.h>
268 #include <sys/zfs_context.h>
270 #include <sys/refcount.h>
271 #include <sys/vdev.h>
272 #include <sys/vdev_impl.h>
273 #include <sys/dsl_pool.h>
274 #include <sys/zio_checksum.h>
275 #include <sys/multilist.h>
278 #include <sys/vmsystm.h>
280 #include <sys/fs/swapnode.h>
282 #include <linux/mm_compat.h>
284 #include <sys/callb.h>
285 #include <sys/kstat.h>
286 #include <sys/dmu_tx.h>
287 #include <zfs_fletcher.h>
288 #include <sys/arc_impl.h>
289 #include <sys/trace_arc.h>
292 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
293 boolean_t arc_watch
= B_FALSE
;
296 static kmutex_t arc_reclaim_lock
;
297 static kcondvar_t arc_reclaim_thread_cv
;
298 static boolean_t arc_reclaim_thread_exit
;
299 static kcondvar_t arc_reclaim_waiters_cv
;
302 * The number of headers to evict in arc_evict_state_impl() before
303 * dropping the sublist lock and evicting from another sublist. A lower
304 * value means we're more likely to evict the "correct" header (i.e. the
305 * oldest header in the arc state), but comes with higher overhead
306 * (i.e. more invocations of arc_evict_state_impl()).
308 int zfs_arc_evict_batch_limit
= 10;
310 /* number of seconds before growing cache again */
311 static int arc_grow_retry
= 5;
313 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
314 int zfs_arc_overflow_shift
= 8;
316 /* shift of arc_c for calculating both min and max arc_p */
317 static int arc_p_min_shift
= 4;
319 /* log2(fraction of arc to reclaim) */
320 static int arc_shrink_shift
= 7;
322 /* percent of pagecache to reclaim arc to */
324 static uint_t zfs_arc_pc_percent
= 0;
328 * log2(fraction of ARC which must be free to allow growing).
329 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
330 * when reading a new block into the ARC, we will evict an equal-sized block
333 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
334 * we will still not allow it to grow.
336 int arc_no_grow_shift
= 5;
340 * minimum lifespan of a prefetch block in clock ticks
341 * (initialized in arc_init())
343 static int arc_min_prefetch_lifespan
;
346 * If this percent of memory is free, don't throttle.
348 int arc_lotsfree_percent
= 10;
353 * The arc has filled available memory and has now warmed up.
355 static boolean_t arc_warm
;
358 * log2 fraction of the zio arena to keep free.
360 int arc_zio_arena_free_shift
= 2;
363 * These tunables are for performance analysis.
365 unsigned long zfs_arc_max
= 0;
366 unsigned long zfs_arc_min
= 0;
367 unsigned long zfs_arc_meta_limit
= 0;
368 unsigned long zfs_arc_meta_min
= 0;
369 unsigned long zfs_arc_dnode_limit
= 0;
370 unsigned long zfs_arc_dnode_reduce_percent
= 10;
371 int zfs_arc_grow_retry
= 0;
372 int zfs_arc_shrink_shift
= 0;
373 int zfs_arc_p_min_shift
= 0;
374 int zfs_arc_average_blocksize
= 8 * 1024; /* 8KB */
376 int zfs_compressed_arc_enabled
= B_TRUE
;
379 * ARC will evict meta buffers that exceed arc_meta_limit. This
380 * tunable make arc_meta_limit adjustable for different workloads.
382 unsigned long zfs_arc_meta_limit_percent
= 75;
385 * Percentage that can be consumed by dnodes of ARC meta buffers.
387 unsigned long zfs_arc_dnode_limit_percent
= 10;
390 * These tunables are Linux specific
392 unsigned long zfs_arc_sys_free
= 0;
393 int zfs_arc_min_prefetch_lifespan
= 0;
394 int zfs_arc_p_aggressive_disable
= 1;
395 int zfs_arc_p_dampener_disable
= 1;
396 int zfs_arc_meta_prune
= 10000;
397 int zfs_arc_meta_strategy
= ARC_STRATEGY_META_BALANCED
;
398 int zfs_arc_meta_adjust_restarts
= 4096;
399 int zfs_arc_lotsfree_percent
= 10;
402 static arc_state_t ARC_anon
;
403 static arc_state_t ARC_mru
;
404 static arc_state_t ARC_mru_ghost
;
405 static arc_state_t ARC_mfu
;
406 static arc_state_t ARC_mfu_ghost
;
407 static arc_state_t ARC_l2c_only
;
409 typedef struct arc_stats
{
410 kstat_named_t arcstat_hits
;
411 kstat_named_t arcstat_misses
;
412 kstat_named_t arcstat_demand_data_hits
;
413 kstat_named_t arcstat_demand_data_misses
;
414 kstat_named_t arcstat_demand_metadata_hits
;
415 kstat_named_t arcstat_demand_metadata_misses
;
416 kstat_named_t arcstat_prefetch_data_hits
;
417 kstat_named_t arcstat_prefetch_data_misses
;
418 kstat_named_t arcstat_prefetch_metadata_hits
;
419 kstat_named_t arcstat_prefetch_metadata_misses
;
420 kstat_named_t arcstat_mru_hits
;
421 kstat_named_t arcstat_mru_ghost_hits
;
422 kstat_named_t arcstat_mfu_hits
;
423 kstat_named_t arcstat_mfu_ghost_hits
;
424 kstat_named_t arcstat_deleted
;
426 * Number of buffers that could not be evicted because the hash lock
427 * was held by another thread. The lock may not necessarily be held
428 * by something using the same buffer, since hash locks are shared
429 * by multiple buffers.
431 kstat_named_t arcstat_mutex_miss
;
433 * Number of buffers skipped because they have I/O in progress, are
434 * indrect prefetch buffers that have not lived long enough, or are
435 * not from the spa we're trying to evict from.
437 kstat_named_t arcstat_evict_skip
;
439 * Number of times arc_evict_state() was unable to evict enough
440 * buffers to reach its target amount.
442 kstat_named_t arcstat_evict_not_enough
;
443 kstat_named_t arcstat_evict_l2_cached
;
444 kstat_named_t arcstat_evict_l2_eligible
;
445 kstat_named_t arcstat_evict_l2_ineligible
;
446 kstat_named_t arcstat_evict_l2_skip
;
447 kstat_named_t arcstat_hash_elements
;
448 kstat_named_t arcstat_hash_elements_max
;
449 kstat_named_t arcstat_hash_collisions
;
450 kstat_named_t arcstat_hash_chains
;
451 kstat_named_t arcstat_hash_chain_max
;
452 kstat_named_t arcstat_p
;
453 kstat_named_t arcstat_c
;
454 kstat_named_t arcstat_c_min
;
455 kstat_named_t arcstat_c_max
;
456 kstat_named_t arcstat_size
;
458 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
459 * Note that the compressed bytes may match the uncompressed bytes
460 * if the block is either not compressed or compressed arc is disabled.
462 kstat_named_t arcstat_compressed_size
;
464 * Uncompressed size of the data stored in b_pabd. If compressed
465 * arc is disabled then this value will be identical to the stat
468 kstat_named_t arcstat_uncompressed_size
;
470 * Number of bytes stored in all the arc_buf_t's. This is classified
471 * as "overhead" since this data is typically short-lived and will
472 * be evicted from the arc when it becomes unreferenced unless the
473 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
474 * values have been set (see comment in dbuf.c for more information).
476 kstat_named_t arcstat_overhead_size
;
478 * Number of bytes consumed by internal ARC structures necessary
479 * for tracking purposes; these structures are not actually
480 * backed by ARC buffers. This includes arc_buf_hdr_t structures
481 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
482 * caches), and arc_buf_t structures (allocated via arc_buf_t
485 kstat_named_t arcstat_hdr_size
;
487 * Number of bytes consumed by ARC buffers of type equal to
488 * ARC_BUFC_DATA. This is generally consumed by buffers backing
489 * on disk user data (e.g. plain file contents).
491 kstat_named_t arcstat_data_size
;
493 * Number of bytes consumed by ARC buffers of type equal to
494 * ARC_BUFC_METADATA. This is generally consumed by buffers
495 * backing on disk data that is used for internal ZFS
496 * structures (e.g. ZAP, dnode, indirect blocks, etc).
498 kstat_named_t arcstat_metadata_size
;
500 * Number of bytes consumed by dmu_buf_impl_t objects.
502 kstat_named_t arcstat_dbuf_size
;
504 * Number of bytes consumed by dnode_t objects.
506 kstat_named_t arcstat_dnode_size
;
508 * Number of bytes consumed by bonus buffers.
510 kstat_named_t arcstat_bonus_size
;
512 * Total number of bytes consumed by ARC buffers residing in the
513 * arc_anon state. This includes *all* buffers in the arc_anon
514 * state; e.g. data, metadata, evictable, and unevictable buffers
515 * are all included in this value.
517 kstat_named_t arcstat_anon_size
;
519 * Number of bytes consumed by ARC buffers that meet the
520 * following criteria: backing buffers of type ARC_BUFC_DATA,
521 * residing in the arc_anon state, and are eligible for eviction
522 * (e.g. have no outstanding holds on the buffer).
524 kstat_named_t arcstat_anon_evictable_data
;
526 * Number of bytes consumed by ARC buffers that meet the
527 * following criteria: backing buffers of type ARC_BUFC_METADATA,
528 * residing in the arc_anon state, and are eligible for eviction
529 * (e.g. have no outstanding holds on the buffer).
531 kstat_named_t arcstat_anon_evictable_metadata
;
533 * Total number of bytes consumed by ARC buffers residing in the
534 * arc_mru state. This includes *all* buffers in the arc_mru
535 * state; e.g. data, metadata, evictable, and unevictable buffers
536 * are all included in this value.
538 kstat_named_t arcstat_mru_size
;
540 * Number of bytes consumed by ARC buffers that meet the
541 * following criteria: backing buffers of type ARC_BUFC_DATA,
542 * residing in the arc_mru state, and are eligible for eviction
543 * (e.g. have no outstanding holds on the buffer).
545 kstat_named_t arcstat_mru_evictable_data
;
547 * Number of bytes consumed by ARC buffers that meet the
548 * following criteria: backing buffers of type ARC_BUFC_METADATA,
549 * residing in the arc_mru state, and are eligible for eviction
550 * (e.g. have no outstanding holds on the buffer).
552 kstat_named_t arcstat_mru_evictable_metadata
;
554 * Total number of bytes that *would have been* consumed by ARC
555 * buffers in the arc_mru_ghost state. The key thing to note
556 * here, is the fact that this size doesn't actually indicate
557 * RAM consumption. The ghost lists only consist of headers and
558 * don't actually have ARC buffers linked off of these headers.
559 * Thus, *if* the headers had associated ARC buffers, these
560 * buffers *would have* consumed this number of bytes.
562 kstat_named_t arcstat_mru_ghost_size
;
564 * Number of bytes that *would have been* consumed by ARC
565 * buffers that are eligible for eviction, of type
566 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
568 kstat_named_t arcstat_mru_ghost_evictable_data
;
570 * Number of bytes that *would have been* consumed by ARC
571 * buffers that are eligible for eviction, of type
572 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
574 kstat_named_t arcstat_mru_ghost_evictable_metadata
;
576 * Total number of bytes consumed by ARC buffers residing in the
577 * arc_mfu state. This includes *all* buffers in the arc_mfu
578 * state; e.g. data, metadata, evictable, and unevictable buffers
579 * are all included in this value.
581 kstat_named_t arcstat_mfu_size
;
583 * Number of bytes consumed by ARC buffers that are eligible for
584 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
587 kstat_named_t arcstat_mfu_evictable_data
;
589 * Number of bytes consumed by ARC buffers that are eligible for
590 * eviction, of type ARC_BUFC_METADATA, and reside in the
593 kstat_named_t arcstat_mfu_evictable_metadata
;
595 * Total number of bytes that *would have been* consumed by ARC
596 * buffers in the arc_mfu_ghost state. See the comment above
597 * arcstat_mru_ghost_size for more details.
599 kstat_named_t arcstat_mfu_ghost_size
;
601 * Number of bytes that *would have been* consumed by ARC
602 * buffers that are eligible for eviction, of type
603 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
605 kstat_named_t arcstat_mfu_ghost_evictable_data
;
607 * Number of bytes that *would have been* consumed by ARC
608 * buffers that are eligible for eviction, of type
609 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
611 kstat_named_t arcstat_mfu_ghost_evictable_metadata
;
612 kstat_named_t arcstat_l2_hits
;
613 kstat_named_t arcstat_l2_misses
;
614 kstat_named_t arcstat_l2_feeds
;
615 kstat_named_t arcstat_l2_rw_clash
;
616 kstat_named_t arcstat_l2_read_bytes
;
617 kstat_named_t arcstat_l2_write_bytes
;
618 kstat_named_t arcstat_l2_writes_sent
;
619 kstat_named_t arcstat_l2_writes_done
;
620 kstat_named_t arcstat_l2_writes_error
;
621 kstat_named_t arcstat_l2_writes_lock_retry
;
622 kstat_named_t arcstat_l2_evict_lock_retry
;
623 kstat_named_t arcstat_l2_evict_reading
;
624 kstat_named_t arcstat_l2_evict_l1cached
;
625 kstat_named_t arcstat_l2_free_on_write
;
626 kstat_named_t arcstat_l2_abort_lowmem
;
627 kstat_named_t arcstat_l2_cksum_bad
;
628 kstat_named_t arcstat_l2_io_error
;
629 kstat_named_t arcstat_l2_lsize
;
630 kstat_named_t arcstat_l2_psize
;
631 kstat_named_t arcstat_l2_hdr_size
;
632 kstat_named_t arcstat_memory_throttle_count
;
633 kstat_named_t arcstat_memory_direct_count
;
634 kstat_named_t arcstat_memory_indirect_count
;
635 kstat_named_t arcstat_memory_all_bytes
;
636 kstat_named_t arcstat_memory_free_bytes
;
637 kstat_named_t arcstat_memory_available_bytes
;
638 kstat_named_t arcstat_no_grow
;
639 kstat_named_t arcstat_tempreserve
;
640 kstat_named_t arcstat_loaned_bytes
;
641 kstat_named_t arcstat_prune
;
642 kstat_named_t arcstat_meta_used
;
643 kstat_named_t arcstat_meta_limit
;
644 kstat_named_t arcstat_dnode_limit
;
645 kstat_named_t arcstat_meta_max
;
646 kstat_named_t arcstat_meta_min
;
647 kstat_named_t arcstat_sync_wait_for_async
;
648 kstat_named_t arcstat_demand_hit_predictive_prefetch
;
649 kstat_named_t arcstat_need_free
;
650 kstat_named_t arcstat_sys_free
;
653 static arc_stats_t arc_stats
= {
654 { "hits", KSTAT_DATA_UINT64
},
655 { "misses", KSTAT_DATA_UINT64
},
656 { "demand_data_hits", KSTAT_DATA_UINT64
},
657 { "demand_data_misses", KSTAT_DATA_UINT64
},
658 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
659 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
660 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
661 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
662 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
663 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
664 { "mru_hits", KSTAT_DATA_UINT64
},
665 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
666 { "mfu_hits", KSTAT_DATA_UINT64
},
667 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
668 { "deleted", KSTAT_DATA_UINT64
},
669 { "mutex_miss", KSTAT_DATA_UINT64
},
670 { "evict_skip", KSTAT_DATA_UINT64
},
671 { "evict_not_enough", KSTAT_DATA_UINT64
},
672 { "evict_l2_cached", KSTAT_DATA_UINT64
},
673 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
674 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
675 { "evict_l2_skip", KSTAT_DATA_UINT64
},
676 { "hash_elements", KSTAT_DATA_UINT64
},
677 { "hash_elements_max", KSTAT_DATA_UINT64
},
678 { "hash_collisions", KSTAT_DATA_UINT64
},
679 { "hash_chains", KSTAT_DATA_UINT64
},
680 { "hash_chain_max", KSTAT_DATA_UINT64
},
681 { "p", KSTAT_DATA_UINT64
},
682 { "c", KSTAT_DATA_UINT64
},
683 { "c_min", KSTAT_DATA_UINT64
},
684 { "c_max", KSTAT_DATA_UINT64
},
685 { "size", KSTAT_DATA_UINT64
},
686 { "compressed_size", KSTAT_DATA_UINT64
},
687 { "uncompressed_size", KSTAT_DATA_UINT64
},
688 { "overhead_size", KSTAT_DATA_UINT64
},
689 { "hdr_size", KSTAT_DATA_UINT64
},
690 { "data_size", KSTAT_DATA_UINT64
},
691 { "metadata_size", KSTAT_DATA_UINT64
},
692 { "dbuf_size", KSTAT_DATA_UINT64
},
693 { "dnode_size", KSTAT_DATA_UINT64
},
694 { "bonus_size", KSTAT_DATA_UINT64
},
695 { "anon_size", KSTAT_DATA_UINT64
},
696 { "anon_evictable_data", KSTAT_DATA_UINT64
},
697 { "anon_evictable_metadata", KSTAT_DATA_UINT64
},
698 { "mru_size", KSTAT_DATA_UINT64
},
699 { "mru_evictable_data", KSTAT_DATA_UINT64
},
700 { "mru_evictable_metadata", KSTAT_DATA_UINT64
},
701 { "mru_ghost_size", KSTAT_DATA_UINT64
},
702 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64
},
703 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
704 { "mfu_size", KSTAT_DATA_UINT64
},
705 { "mfu_evictable_data", KSTAT_DATA_UINT64
},
706 { "mfu_evictable_metadata", KSTAT_DATA_UINT64
},
707 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
708 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64
},
709 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
710 { "l2_hits", KSTAT_DATA_UINT64
},
711 { "l2_misses", KSTAT_DATA_UINT64
},
712 { "l2_feeds", KSTAT_DATA_UINT64
},
713 { "l2_rw_clash", KSTAT_DATA_UINT64
},
714 { "l2_read_bytes", KSTAT_DATA_UINT64
},
715 { "l2_write_bytes", KSTAT_DATA_UINT64
},
716 { "l2_writes_sent", KSTAT_DATA_UINT64
},
717 { "l2_writes_done", KSTAT_DATA_UINT64
},
718 { "l2_writes_error", KSTAT_DATA_UINT64
},
719 { "l2_writes_lock_retry", KSTAT_DATA_UINT64
},
720 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
721 { "l2_evict_reading", KSTAT_DATA_UINT64
},
722 { "l2_evict_l1cached", KSTAT_DATA_UINT64
},
723 { "l2_free_on_write", KSTAT_DATA_UINT64
},
724 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
725 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
726 { "l2_io_error", KSTAT_DATA_UINT64
},
727 { "l2_size", KSTAT_DATA_UINT64
},
728 { "l2_asize", KSTAT_DATA_UINT64
},
729 { "l2_hdr_size", KSTAT_DATA_UINT64
},
730 { "memory_throttle_count", KSTAT_DATA_UINT64
},
731 { "memory_direct_count", KSTAT_DATA_UINT64
},
732 { "memory_indirect_count", KSTAT_DATA_UINT64
},
733 { "memory_all_bytes", KSTAT_DATA_UINT64
},
734 { "memory_free_bytes", KSTAT_DATA_UINT64
},
735 { "memory_available_bytes", KSTAT_DATA_INT64
},
736 { "arc_no_grow", KSTAT_DATA_UINT64
},
737 { "arc_tempreserve", KSTAT_DATA_UINT64
},
738 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
739 { "arc_prune", KSTAT_DATA_UINT64
},
740 { "arc_meta_used", KSTAT_DATA_UINT64
},
741 { "arc_meta_limit", KSTAT_DATA_UINT64
},
742 { "arc_dnode_limit", KSTAT_DATA_UINT64
},
743 { "arc_meta_max", KSTAT_DATA_UINT64
},
744 { "arc_meta_min", KSTAT_DATA_UINT64
},
745 { "sync_wait_for_async", KSTAT_DATA_UINT64
},
746 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64
},
747 { "arc_need_free", KSTAT_DATA_UINT64
},
748 { "arc_sys_free", KSTAT_DATA_UINT64
}
751 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
753 #define ARCSTAT_INCR(stat, val) \
754 atomic_add_64(&arc_stats.stat.value.ui64, (val))
756 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
757 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
759 #define ARCSTAT_MAX(stat, val) { \
761 while ((val) > (m = arc_stats.stat.value.ui64) && \
762 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
766 #define ARCSTAT_MAXSTAT(stat) \
767 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
770 * We define a macro to allow ARC hits/misses to be easily broken down by
771 * two separate conditions, giving a total of four different subtypes for
772 * each of hits and misses (so eight statistics total).
774 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
777 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
779 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
783 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
785 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
790 static arc_state_t
*arc_anon
;
791 static arc_state_t
*arc_mru
;
792 static arc_state_t
*arc_mru_ghost
;
793 static arc_state_t
*arc_mfu
;
794 static arc_state_t
*arc_mfu_ghost
;
795 static arc_state_t
*arc_l2c_only
;
798 * There are several ARC variables that are critical to export as kstats --
799 * but we don't want to have to grovel around in the kstat whenever we wish to
800 * manipulate them. For these variables, we therefore define them to be in
801 * terms of the statistic variable. This assures that we are not introducing
802 * the possibility of inconsistency by having shadow copies of the variables,
803 * while still allowing the code to be readable.
805 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
806 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
807 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
808 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
809 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
810 #define arc_no_grow ARCSTAT(arcstat_no_grow) /* do not grow cache size */
811 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
812 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
813 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
814 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
815 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
816 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
817 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
818 #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
819 #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */
820 #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
821 #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */
822 #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */
824 /* compressed size of entire arc */
825 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
826 /* uncompressed size of entire arc */
827 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
828 /* number of bytes in the arc from arc_buf_t's */
829 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
831 static list_t arc_prune_list
;
832 static kmutex_t arc_prune_mtx
;
833 static taskq_t
*arc_prune_taskq
;
835 #define GHOST_STATE(state) \
836 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
837 (state) == arc_l2c_only)
839 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
840 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
841 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
842 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
843 #define HDR_COMPRESSION_ENABLED(hdr) \
844 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
846 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
847 #define HDR_L2_READING(hdr) \
848 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
849 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
850 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
851 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
852 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
853 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
855 #define HDR_ISTYPE_METADATA(hdr) \
856 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
857 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
859 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
860 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
862 /* For storing compression mode in b_flags */
863 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
865 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
866 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
867 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
868 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
870 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
871 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
872 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
878 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
879 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
882 * Hash table routines
885 #define HT_LOCK_ALIGN 64
886 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
891 unsigned char pad
[HT_LOCK_PAD
];
895 #define BUF_LOCKS 8192
896 typedef struct buf_hash_table
{
898 arc_buf_hdr_t
**ht_table
;
899 struct ht_lock ht_locks
[BUF_LOCKS
];
902 static buf_hash_table_t buf_hash_table
;
904 #define BUF_HASH_INDEX(spa, dva, birth) \
905 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
906 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
907 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
908 #define HDR_LOCK(hdr) \
909 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
911 uint64_t zfs_crc64_table
[256];
917 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
918 #define L2ARC_HEADROOM 2 /* num of writes */
921 * If we discover during ARC scan any buffers to be compressed, we boost
922 * our headroom for the next scanning cycle by this percentage multiple.
924 #define L2ARC_HEADROOM_BOOST 200
925 #define L2ARC_FEED_SECS 1 /* caching interval secs */
926 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
929 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
930 * and each of the state has two types: data and metadata.
932 #define L2ARC_FEED_TYPES 4
934 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
935 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
937 /* L2ARC Performance Tunables */
938 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
939 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
940 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
941 unsigned long l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
942 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
943 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
944 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
945 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
946 int l2arc_norw
= B_FALSE
; /* no reads during writes */
951 static list_t L2ARC_dev_list
; /* device list */
952 static list_t
*l2arc_dev_list
; /* device list pointer */
953 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
954 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
955 static list_t L2ARC_free_on_write
; /* free after write buf list */
956 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
957 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
958 static uint64_t l2arc_ndev
; /* number of devices */
960 typedef struct l2arc_read_callback
{
961 arc_buf_hdr_t
*l2rcb_hdr
; /* read header */
962 blkptr_t l2rcb_bp
; /* original blkptr */
963 zbookmark_phys_t l2rcb_zb
; /* original bookmark */
964 int l2rcb_flags
; /* original flags */
965 abd_t
*l2rcb_abd
; /* temporary buffer */
966 } l2arc_read_callback_t
;
968 typedef struct l2arc_data_free
{
969 /* protected by l2arc_free_on_write_mtx */
972 arc_buf_contents_t l2df_type
;
973 list_node_t l2df_list_node
;
976 static kmutex_t l2arc_feed_thr_lock
;
977 static kcondvar_t l2arc_feed_thr_cv
;
978 static uint8_t l2arc_thread_exit
;
980 static abd_t
*arc_get_data_abd(arc_buf_hdr_t
*, uint64_t, void *);
981 static void *arc_get_data_buf(arc_buf_hdr_t
*, uint64_t, void *);
982 static void arc_get_data_impl(arc_buf_hdr_t
*, uint64_t, void *);
983 static void arc_free_data_abd(arc_buf_hdr_t
*, abd_t
*, uint64_t, void *);
984 static void arc_free_data_buf(arc_buf_hdr_t
*, void *, uint64_t, void *);
985 static void arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
);
986 static void arc_hdr_free_pabd(arc_buf_hdr_t
*);
987 static void arc_hdr_alloc_pabd(arc_buf_hdr_t
*);
988 static void arc_access(arc_buf_hdr_t
*, kmutex_t
*);
989 static boolean_t
arc_is_overflowing(void);
990 static void arc_buf_watch(arc_buf_t
*);
991 static void arc_tuning_update(void);
992 static void arc_prune_async(int64_t);
993 static uint64_t arc_all_memory(void);
995 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t
*);
996 static uint32_t arc_bufc_to_flags(arc_buf_contents_t
);
997 static inline void arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
998 static inline void arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
1000 static boolean_t
l2arc_write_eligible(uint64_t, arc_buf_hdr_t
*);
1001 static void l2arc_read_done(zio_t
*);
1004 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
1006 uint8_t *vdva
= (uint8_t *)dva
;
1007 uint64_t crc
= -1ULL;
1010 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
1012 for (i
= 0; i
< sizeof (dva_t
); i
++)
1013 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
1015 crc
^= (spa
>>8) ^ birth
;
1020 #define HDR_EMPTY(hdr) \
1021 ((hdr)->b_dva.dva_word[0] == 0 && \
1022 (hdr)->b_dva.dva_word[1] == 0)
1024 #define HDR_EQUAL(spa, dva, birth, hdr) \
1025 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1026 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1027 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1030 buf_discard_identity(arc_buf_hdr_t
*hdr
)
1032 hdr
->b_dva
.dva_word
[0] = 0;
1033 hdr
->b_dva
.dva_word
[1] = 0;
1037 static arc_buf_hdr_t
*
1038 buf_hash_find(uint64_t spa
, const blkptr_t
*bp
, kmutex_t
**lockp
)
1040 const dva_t
*dva
= BP_IDENTITY(bp
);
1041 uint64_t birth
= BP_PHYSICAL_BIRTH(bp
);
1042 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
1043 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1046 mutex_enter(hash_lock
);
1047 for (hdr
= buf_hash_table
.ht_table
[idx
]; hdr
!= NULL
;
1048 hdr
= hdr
->b_hash_next
) {
1049 if (HDR_EQUAL(spa
, dva
, birth
, hdr
)) {
1054 mutex_exit(hash_lock
);
1060 * Insert an entry into the hash table. If there is already an element
1061 * equal to elem in the hash table, then the already existing element
1062 * will be returned and the new element will not be inserted.
1063 * Otherwise returns NULL.
1064 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1066 static arc_buf_hdr_t
*
1067 buf_hash_insert(arc_buf_hdr_t
*hdr
, kmutex_t
**lockp
)
1069 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1070 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1071 arc_buf_hdr_t
*fhdr
;
1074 ASSERT(!DVA_IS_EMPTY(&hdr
->b_dva
));
1075 ASSERT(hdr
->b_birth
!= 0);
1076 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
1078 if (lockp
!= NULL
) {
1080 mutex_enter(hash_lock
);
1082 ASSERT(MUTEX_HELD(hash_lock
));
1085 for (fhdr
= buf_hash_table
.ht_table
[idx
], i
= 0; fhdr
!= NULL
;
1086 fhdr
= fhdr
->b_hash_next
, i
++) {
1087 if (HDR_EQUAL(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
, fhdr
))
1091 hdr
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
1092 buf_hash_table
.ht_table
[idx
] = hdr
;
1093 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1095 /* collect some hash table performance data */
1097 ARCSTAT_BUMP(arcstat_hash_collisions
);
1099 ARCSTAT_BUMP(arcstat_hash_chains
);
1101 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
1104 ARCSTAT_BUMP(arcstat_hash_elements
);
1105 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
1111 buf_hash_remove(arc_buf_hdr_t
*hdr
)
1113 arc_buf_hdr_t
*fhdr
, **hdrp
;
1114 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1116 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
1117 ASSERT(HDR_IN_HASH_TABLE(hdr
));
1119 hdrp
= &buf_hash_table
.ht_table
[idx
];
1120 while ((fhdr
= *hdrp
) != hdr
) {
1121 ASSERT3P(fhdr
, !=, NULL
);
1122 hdrp
= &fhdr
->b_hash_next
;
1124 *hdrp
= hdr
->b_hash_next
;
1125 hdr
->b_hash_next
= NULL
;
1126 arc_hdr_clear_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1128 /* collect some hash table performance data */
1129 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
1131 if (buf_hash_table
.ht_table
[idx
] &&
1132 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
1133 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
1137 * Global data structures and functions for the buf kmem cache.
1139 static kmem_cache_t
*hdr_full_cache
;
1140 static kmem_cache_t
*hdr_l2only_cache
;
1141 static kmem_cache_t
*buf_cache
;
1148 #if defined(_KERNEL) && defined(HAVE_SPL)
1150 * Large allocations which do not require contiguous pages
1151 * should be using vmem_free() in the linux kernel\
1153 vmem_free(buf_hash_table
.ht_table
,
1154 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1156 kmem_free(buf_hash_table
.ht_table
,
1157 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1159 for (i
= 0; i
< BUF_LOCKS
; i
++)
1160 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
1161 kmem_cache_destroy(hdr_full_cache
);
1162 kmem_cache_destroy(hdr_l2only_cache
);
1163 kmem_cache_destroy(buf_cache
);
1167 * Constructor callback - called when the cache is empty
1168 * and a new buf is requested.
1172 hdr_full_cons(void *vbuf
, void *unused
, int kmflag
)
1174 arc_buf_hdr_t
*hdr
= vbuf
;
1176 bzero(hdr
, HDR_FULL_SIZE
);
1177 cv_init(&hdr
->b_l1hdr
.b_cv
, NULL
, CV_DEFAULT
, NULL
);
1178 refcount_create(&hdr
->b_l1hdr
.b_refcnt
);
1179 mutex_init(&hdr
->b_l1hdr
.b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1180 list_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1181 list_link_init(&hdr
->b_l2hdr
.b_l2node
);
1182 multilist_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1183 arc_space_consume(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1190 hdr_l2only_cons(void *vbuf
, void *unused
, int kmflag
)
1192 arc_buf_hdr_t
*hdr
= vbuf
;
1194 bzero(hdr
, HDR_L2ONLY_SIZE
);
1195 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1202 buf_cons(void *vbuf
, void *unused
, int kmflag
)
1204 arc_buf_t
*buf
= vbuf
;
1206 bzero(buf
, sizeof (arc_buf_t
));
1207 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1208 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1214 * Destructor callback - called when a cached buf is
1215 * no longer required.
1219 hdr_full_dest(void *vbuf
, void *unused
)
1221 arc_buf_hdr_t
*hdr
= vbuf
;
1223 ASSERT(HDR_EMPTY(hdr
));
1224 cv_destroy(&hdr
->b_l1hdr
.b_cv
);
1225 refcount_destroy(&hdr
->b_l1hdr
.b_refcnt
);
1226 mutex_destroy(&hdr
->b_l1hdr
.b_freeze_lock
);
1227 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1228 arc_space_return(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1233 hdr_l2only_dest(void *vbuf
, void *unused
)
1235 ASSERTV(arc_buf_hdr_t
*hdr
= vbuf
);
1237 ASSERT(HDR_EMPTY(hdr
));
1238 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1243 buf_dest(void *vbuf
, void *unused
)
1245 arc_buf_t
*buf
= vbuf
;
1247 mutex_destroy(&buf
->b_evict_lock
);
1248 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1252 * Reclaim callback -- invoked when memory is low.
1256 hdr_recl(void *unused
)
1258 dprintf("hdr_recl called\n");
1260 * umem calls the reclaim func when we destroy the buf cache,
1261 * which is after we do arc_fini().
1264 cv_signal(&arc_reclaim_thread_cv
);
1270 uint64_t *ct
= NULL
;
1271 uint64_t hsize
= 1ULL << 12;
1275 * The hash table is big enough to fill all of physical memory
1276 * with an average block size of zfs_arc_average_blocksize (default 8K).
1277 * By default, the table will take up
1278 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1280 while (hsize
* zfs_arc_average_blocksize
< arc_all_memory())
1283 buf_hash_table
.ht_mask
= hsize
- 1;
1284 #if defined(_KERNEL) && defined(HAVE_SPL)
1286 * Large allocations which do not require contiguous pages
1287 * should be using vmem_alloc() in the linux kernel
1289 buf_hash_table
.ht_table
=
1290 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
1292 buf_hash_table
.ht_table
=
1293 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
1295 if (buf_hash_table
.ht_table
== NULL
) {
1296 ASSERT(hsize
> (1ULL << 8));
1301 hdr_full_cache
= kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE
,
1302 0, hdr_full_cons
, hdr_full_dest
, hdr_recl
, NULL
, NULL
, 0);
1303 hdr_l2only_cache
= kmem_cache_create("arc_buf_hdr_t_l2only",
1304 HDR_L2ONLY_SIZE
, 0, hdr_l2only_cons
, hdr_l2only_dest
, hdr_recl
,
1306 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
1307 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
1309 for (i
= 0; i
< 256; i
++)
1310 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
1311 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
1313 for (i
= 0; i
< BUF_LOCKS
; i
++) {
1314 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
1315 NULL
, MUTEX_DEFAULT
, NULL
);
1319 #define ARC_MINTIME (hz>>4) /* 62 ms */
1322 * This is the size that the buf occupies in memory. If the buf is compressed,
1323 * it will correspond to the compressed size. You should use this method of
1324 * getting the buf size unless you explicitly need the logical size.
1327 arc_buf_size(arc_buf_t
*buf
)
1329 return (ARC_BUF_COMPRESSED(buf
) ?
1330 HDR_GET_PSIZE(buf
->b_hdr
) : HDR_GET_LSIZE(buf
->b_hdr
));
1334 arc_buf_lsize(arc_buf_t
*buf
)
1336 return (HDR_GET_LSIZE(buf
->b_hdr
));
1340 arc_get_compression(arc_buf_t
*buf
)
1342 return (ARC_BUF_COMPRESSED(buf
) ?
1343 HDR_GET_COMPRESS(buf
->b_hdr
) : ZIO_COMPRESS_OFF
);
1346 static inline boolean_t
1347 arc_buf_is_shared(arc_buf_t
*buf
)
1349 boolean_t shared
= (buf
->b_data
!= NULL
&&
1350 buf
->b_hdr
->b_l1hdr
.b_pabd
!= NULL
&&
1351 abd_is_linear(buf
->b_hdr
->b_l1hdr
.b_pabd
) &&
1352 buf
->b_data
== abd_to_buf(buf
->b_hdr
->b_l1hdr
.b_pabd
));
1353 IMPLY(shared
, HDR_SHARED_DATA(buf
->b_hdr
));
1354 IMPLY(shared
, ARC_BUF_SHARED(buf
));
1355 IMPLY(shared
, ARC_BUF_COMPRESSED(buf
) || ARC_BUF_LAST(buf
));
1358 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1359 * already being shared" requirement prevents us from doing that.
1366 * Free the checksum associated with this header. If there is no checksum, this
1370 arc_cksum_free(arc_buf_hdr_t
*hdr
)
1372 ASSERT(HDR_HAS_L1HDR(hdr
));
1373 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1374 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1375 kmem_free(hdr
->b_l1hdr
.b_freeze_cksum
, sizeof (zio_cksum_t
));
1376 hdr
->b_l1hdr
.b_freeze_cksum
= NULL
;
1378 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1382 * Return true iff at least one of the bufs on hdr is not compressed.
1385 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t
*hdr
)
1387 for (arc_buf_t
*b
= hdr
->b_l1hdr
.b_buf
; b
!= NULL
; b
= b
->b_next
) {
1388 if (!ARC_BUF_COMPRESSED(b
)) {
1397 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1398 * matches the checksum that is stored in the hdr. If there is no checksum,
1399 * or if the buf is compressed, this is a no-op.
1402 arc_cksum_verify(arc_buf_t
*buf
)
1404 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1407 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1410 if (ARC_BUF_COMPRESSED(buf
)) {
1411 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1412 arc_hdr_has_uncompressed_buf(hdr
));
1416 ASSERT(HDR_HAS_L1HDR(hdr
));
1418 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1419 if (hdr
->b_l1hdr
.b_freeze_cksum
== NULL
|| HDR_IO_ERROR(hdr
)) {
1420 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1424 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
, &zc
);
1425 if (!ZIO_CHECKSUM_EQUAL(*hdr
->b_l1hdr
.b_freeze_cksum
, zc
))
1426 panic("buffer modified while frozen!");
1427 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1431 arc_cksum_is_equal(arc_buf_hdr_t
*hdr
, zio_t
*zio
)
1433 enum zio_compress compress
= BP_GET_COMPRESS(zio
->io_bp
);
1434 boolean_t valid_cksum
;
1436 ASSERT(!BP_IS_EMBEDDED(zio
->io_bp
));
1437 VERIFY3U(BP_GET_PSIZE(zio
->io_bp
), ==, HDR_GET_PSIZE(hdr
));
1440 * We rely on the blkptr's checksum to determine if the block
1441 * is valid or not. When compressed arc is enabled, the l2arc
1442 * writes the block to the l2arc just as it appears in the pool.
1443 * This allows us to use the blkptr's checksum to validate the
1444 * data that we just read off of the l2arc without having to store
1445 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1446 * arc is disabled, then the data written to the l2arc is always
1447 * uncompressed and won't match the block as it exists in the main
1448 * pool. When this is the case, we must first compress it if it is
1449 * compressed on the main pool before we can validate the checksum.
1451 if (!HDR_COMPRESSION_ENABLED(hdr
) && compress
!= ZIO_COMPRESS_OFF
) {
1455 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1457 cbuf
= zio_buf_alloc(HDR_GET_PSIZE(hdr
));
1458 lsize
= HDR_GET_LSIZE(hdr
);
1459 csize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1461 ASSERT3U(csize
, <=, HDR_GET_PSIZE(hdr
));
1462 if (csize
< HDR_GET_PSIZE(hdr
)) {
1464 * Compressed blocks are always a multiple of the
1465 * smallest ashift in the pool. Ideally, we would
1466 * like to round up the csize to the next
1467 * spa_min_ashift but that value may have changed
1468 * since the block was last written. Instead,
1469 * we rely on the fact that the hdr's psize
1470 * was set to the psize of the block when it was
1471 * last written. We set the csize to that value
1472 * and zero out any part that should not contain
1475 bzero((char *)cbuf
+ csize
, HDR_GET_PSIZE(hdr
) - csize
);
1476 csize
= HDR_GET_PSIZE(hdr
);
1478 zio_push_transform(zio
, cbuf
, csize
, HDR_GET_PSIZE(hdr
), NULL
);
1482 * Block pointers always store the checksum for the logical data.
1483 * If the block pointer has the gang bit set, then the checksum
1484 * it represents is for the reconstituted data and not for an
1485 * individual gang member. The zio pipeline, however, must be able to
1486 * determine the checksum of each of the gang constituents so it
1487 * treats the checksum comparison differently than what we need
1488 * for l2arc blocks. This prevents us from using the
1489 * zio_checksum_error() interface directly. Instead we must call the
1490 * zio_checksum_error_impl() so that we can ensure the checksum is
1491 * generated using the correct checksum algorithm and accounts for the
1492 * logical I/O size and not just a gang fragment.
1494 valid_cksum
= (zio_checksum_error_impl(zio
->io_spa
, zio
->io_bp
,
1495 BP_GET_CHECKSUM(zio
->io_bp
), zio
->io_abd
, zio
->io_size
,
1496 zio
->io_offset
, NULL
) == 0);
1497 zio_pop_transforms(zio
);
1498 return (valid_cksum
);
1502 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1503 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1504 * isn't modified later on. If buf is compressed or there is already a checksum
1505 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1508 arc_cksum_compute(arc_buf_t
*buf
)
1510 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1512 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1515 ASSERT(HDR_HAS_L1HDR(hdr
));
1517 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1518 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1519 ASSERT(arc_hdr_has_uncompressed_buf(hdr
));
1520 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1522 } else if (ARC_BUF_COMPRESSED(buf
)) {
1523 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1527 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1528 hdr
->b_l1hdr
.b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
),
1530 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
,
1531 hdr
->b_l1hdr
.b_freeze_cksum
);
1532 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1538 arc_buf_sigsegv(int sig
, siginfo_t
*si
, void *unused
)
1540 panic("Got SIGSEGV at address: 0x%lx\n", (long)si
->si_addr
);
1546 arc_buf_unwatch(arc_buf_t
*buf
)
1550 ASSERT0(mprotect(buf
->b_data
, arc_buf_size(buf
),
1551 PROT_READ
| PROT_WRITE
));
1558 arc_buf_watch(arc_buf_t
*buf
)
1562 ASSERT0(mprotect(buf
->b_data
, arc_buf_size(buf
),
1567 static arc_buf_contents_t
1568 arc_buf_type(arc_buf_hdr_t
*hdr
)
1570 arc_buf_contents_t type
;
1571 if (HDR_ISTYPE_METADATA(hdr
)) {
1572 type
= ARC_BUFC_METADATA
;
1574 type
= ARC_BUFC_DATA
;
1576 VERIFY3U(hdr
->b_type
, ==, type
);
1581 arc_is_metadata(arc_buf_t
*buf
)
1583 return (HDR_ISTYPE_METADATA(buf
->b_hdr
) != 0);
1587 arc_bufc_to_flags(arc_buf_contents_t type
)
1591 /* metadata field is 0 if buffer contains normal data */
1593 case ARC_BUFC_METADATA
:
1594 return (ARC_FLAG_BUFC_METADATA
);
1598 panic("undefined ARC buffer type!");
1599 return ((uint32_t)-1);
1603 arc_buf_thaw(arc_buf_t
*buf
)
1605 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1607 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
1608 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
1610 arc_cksum_verify(buf
);
1613 * Compressed buffers do not manipulate the b_freeze_cksum or
1614 * allocate b_thawed.
1616 if (ARC_BUF_COMPRESSED(buf
)) {
1617 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1618 arc_hdr_has_uncompressed_buf(hdr
));
1622 ASSERT(HDR_HAS_L1HDR(hdr
));
1623 arc_cksum_free(hdr
);
1624 arc_buf_unwatch(buf
);
1628 arc_buf_freeze(arc_buf_t
*buf
)
1630 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1631 kmutex_t
*hash_lock
;
1633 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1636 if (ARC_BUF_COMPRESSED(buf
)) {
1637 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1638 arc_hdr_has_uncompressed_buf(hdr
));
1642 hash_lock
= HDR_LOCK(hdr
);
1643 mutex_enter(hash_lock
);
1645 ASSERT(HDR_HAS_L1HDR(hdr
));
1646 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
||
1647 hdr
->b_l1hdr
.b_state
== arc_anon
);
1648 arc_cksum_compute(buf
);
1649 mutex_exit(hash_lock
);
1653 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1654 * the following functions should be used to ensure that the flags are
1655 * updated in a thread-safe way. When manipulating the flags either
1656 * the hash_lock must be held or the hdr must be undiscoverable. This
1657 * ensures that we're not racing with any other threads when updating
1661 arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1663 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1664 hdr
->b_flags
|= flags
;
1668 arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1670 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1671 hdr
->b_flags
&= ~flags
;
1675 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1676 * done in a special way since we have to clear and set bits
1677 * at the same time. Consumers that wish to set the compression bits
1678 * must use this function to ensure that the flags are updated in
1679 * thread-safe manner.
1682 arc_hdr_set_compress(arc_buf_hdr_t
*hdr
, enum zio_compress cmp
)
1684 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1687 * Holes and embedded blocks will always have a psize = 0 so
1688 * we ignore the compression of the blkptr and set the
1689 * want to uncompress them. Mark them as uncompressed.
1691 if (!zfs_compressed_arc_enabled
|| HDR_GET_PSIZE(hdr
) == 0) {
1692 arc_hdr_clear_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1693 HDR_SET_COMPRESS(hdr
, ZIO_COMPRESS_OFF
);
1694 ASSERT(!HDR_COMPRESSION_ENABLED(hdr
));
1695 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1697 arc_hdr_set_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1698 HDR_SET_COMPRESS(hdr
, cmp
);
1699 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, cmp
);
1700 ASSERT(HDR_COMPRESSION_ENABLED(hdr
));
1705 * Looks for another buf on the same hdr which has the data decompressed, copies
1706 * from it, and returns true. If no such buf exists, returns false.
1709 arc_buf_try_copy_decompressed_data(arc_buf_t
*buf
)
1711 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1712 boolean_t copied
= B_FALSE
;
1714 ASSERT(HDR_HAS_L1HDR(hdr
));
1715 ASSERT3P(buf
->b_data
, !=, NULL
);
1716 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1718 for (arc_buf_t
*from
= hdr
->b_l1hdr
.b_buf
; from
!= NULL
;
1719 from
= from
->b_next
) {
1720 /* can't use our own data buffer */
1725 if (!ARC_BUF_COMPRESSED(from
)) {
1726 bcopy(from
->b_data
, buf
->b_data
, arc_buf_size(buf
));
1733 * There were no decompressed bufs, so there should not be a
1734 * checksum on the hdr either.
1736 EQUIV(!copied
, hdr
->b_l1hdr
.b_freeze_cksum
== NULL
);
1742 * Given a buf that has a data buffer attached to it, this function will
1743 * efficiently fill the buf with data of the specified compression setting from
1744 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1745 * are already sharing a data buf, no copy is performed.
1747 * If the buf is marked as compressed but uncompressed data was requested, this
1748 * will allocate a new data buffer for the buf, remove that flag, and fill the
1749 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1750 * uncompressed data, and (since we haven't added support for it yet) if you
1751 * want compressed data your buf must already be marked as compressed and have
1752 * the correct-sized data buffer.
1755 arc_buf_fill(arc_buf_t
*buf
, boolean_t compressed
)
1757 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1758 boolean_t hdr_compressed
= (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
1759 dmu_object_byteswap_t bswap
= hdr
->b_l1hdr
.b_byteswap
;
1761 ASSERT3P(buf
->b_data
, !=, NULL
);
1762 IMPLY(compressed
, hdr_compressed
);
1763 IMPLY(compressed
, ARC_BUF_COMPRESSED(buf
));
1765 if (hdr_compressed
== compressed
) {
1766 if (!arc_buf_is_shared(buf
)) {
1767 abd_copy_to_buf(buf
->b_data
, hdr
->b_l1hdr
.b_pabd
,
1771 ASSERT(hdr_compressed
);
1772 ASSERT(!compressed
);
1773 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, HDR_GET_PSIZE(hdr
));
1776 * If the buf is sharing its data with the hdr, unlink it and
1777 * allocate a new data buffer for the buf.
1779 if (arc_buf_is_shared(buf
)) {
1780 ASSERT(ARC_BUF_COMPRESSED(buf
));
1782 /* We need to give the buf it's own b_data */
1783 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
1785 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
1786 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
1788 /* Previously overhead was 0; just add new overhead */
1789 ARCSTAT_INCR(arcstat_overhead_size
, HDR_GET_LSIZE(hdr
));
1790 } else if (ARC_BUF_COMPRESSED(buf
)) {
1791 /* We need to reallocate the buf's b_data */
1792 arc_free_data_buf(hdr
, buf
->b_data
, HDR_GET_PSIZE(hdr
),
1795 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
1797 /* We increased the size of b_data; update overhead */
1798 ARCSTAT_INCR(arcstat_overhead_size
,
1799 HDR_GET_LSIZE(hdr
) - HDR_GET_PSIZE(hdr
));
1803 * Regardless of the buf's previous compression settings, it
1804 * should not be compressed at the end of this function.
1806 buf
->b_flags
&= ~ARC_BUF_FLAG_COMPRESSED
;
1809 * Try copying the data from another buf which already has a
1810 * decompressed version. If that's not possible, it's time to
1811 * bite the bullet and decompress the data from the hdr.
1813 if (arc_buf_try_copy_decompressed_data(buf
)) {
1814 /* Skip byteswapping and checksumming (already done) */
1815 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, !=, NULL
);
1818 int error
= zio_decompress_data(HDR_GET_COMPRESS(hdr
),
1819 hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
1820 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
1823 * Absent hardware errors or software bugs, this should
1824 * be impossible, but log it anyway so we can debug it.
1828 "hdr %p, compress %d, psize %d, lsize %d",
1829 hdr
, HDR_GET_COMPRESS(hdr
),
1830 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
1831 return (SET_ERROR(EIO
));
1836 /* Byteswap the buf's data if necessary */
1837 if (bswap
!= DMU_BSWAP_NUMFUNCS
) {
1838 ASSERT(!HDR_SHARED_DATA(hdr
));
1839 ASSERT3U(bswap
, <, DMU_BSWAP_NUMFUNCS
);
1840 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, HDR_GET_LSIZE(hdr
));
1843 /* Compute the hdr's checksum if necessary */
1844 arc_cksum_compute(buf
);
1850 arc_decompress(arc_buf_t
*buf
)
1852 return (arc_buf_fill(buf
, B_FALSE
));
1856 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1859 arc_hdr_size(arc_buf_hdr_t
*hdr
)
1863 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
1864 HDR_GET_PSIZE(hdr
) > 0) {
1865 size
= HDR_GET_PSIZE(hdr
);
1867 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, 0);
1868 size
= HDR_GET_LSIZE(hdr
);
1874 * Increment the amount of evictable space in the arc_state_t's refcount.
1875 * We account for the space used by the hdr and the arc buf individually
1876 * so that we can add and remove them from the refcount individually.
1879 arc_evictable_space_increment(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
1881 arc_buf_contents_t type
= arc_buf_type(hdr
);
1884 ASSERT(HDR_HAS_L1HDR(hdr
));
1886 if (GHOST_STATE(state
)) {
1887 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
1888 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1889 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
1890 (void) refcount_add_many(&state
->arcs_esize
[type
],
1891 HDR_GET_LSIZE(hdr
), hdr
);
1895 ASSERT(!GHOST_STATE(state
));
1896 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
1897 (void) refcount_add_many(&state
->arcs_esize
[type
],
1898 arc_hdr_size(hdr
), hdr
);
1900 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
; buf
= buf
->b_next
) {
1901 if (arc_buf_is_shared(buf
))
1903 (void) refcount_add_many(&state
->arcs_esize
[type
],
1904 arc_buf_size(buf
), buf
);
1909 * Decrement the amount of evictable space in the arc_state_t's refcount.
1910 * We account for the space used by the hdr and the arc buf individually
1911 * so that we can add and remove them from the refcount individually.
1914 arc_evictable_space_decrement(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
1916 arc_buf_contents_t type
= arc_buf_type(hdr
);
1919 ASSERT(HDR_HAS_L1HDR(hdr
));
1921 if (GHOST_STATE(state
)) {
1922 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
1923 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1924 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
1925 (void) refcount_remove_many(&state
->arcs_esize
[type
],
1926 HDR_GET_LSIZE(hdr
), hdr
);
1930 ASSERT(!GHOST_STATE(state
));
1931 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
1932 (void) refcount_remove_many(&state
->arcs_esize
[type
],
1933 arc_hdr_size(hdr
), hdr
);
1935 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
; buf
= buf
->b_next
) {
1936 if (arc_buf_is_shared(buf
))
1938 (void) refcount_remove_many(&state
->arcs_esize
[type
],
1939 arc_buf_size(buf
), buf
);
1944 * Add a reference to this hdr indicating that someone is actively
1945 * referencing that memory. When the refcount transitions from 0 to 1,
1946 * we remove it from the respective arc_state_t list to indicate that
1947 * it is not evictable.
1950 add_reference(arc_buf_hdr_t
*hdr
, void *tag
)
1954 ASSERT(HDR_HAS_L1HDR(hdr
));
1955 if (!MUTEX_HELD(HDR_LOCK(hdr
))) {
1956 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
1957 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
1958 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1961 state
= hdr
->b_l1hdr
.b_state
;
1963 if ((refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
) == 1) &&
1964 (state
!= arc_anon
)) {
1965 /* We don't use the L2-only state list. */
1966 if (state
!= arc_l2c_only
) {
1967 multilist_remove(state
->arcs_list
[arc_buf_type(hdr
)],
1969 arc_evictable_space_decrement(hdr
, state
);
1971 /* remove the prefetch flag if we get a reference */
1972 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
1977 * Remove a reference from this hdr. When the reference transitions from
1978 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
1979 * list making it eligible for eviction.
1982 remove_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1985 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
1987 ASSERT(HDR_HAS_L1HDR(hdr
));
1988 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
1989 ASSERT(!GHOST_STATE(state
));
1992 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1993 * check to prevent usage of the arc_l2c_only list.
1995 if (((cnt
= refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
)) == 0) &&
1996 (state
!= arc_anon
)) {
1997 multilist_insert(state
->arcs_list
[arc_buf_type(hdr
)], hdr
);
1998 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
1999 arc_evictable_space_increment(hdr
, state
);
2005 * Returns detailed information about a specific arc buffer. When the
2006 * state_index argument is set the function will calculate the arc header
2007 * list position for its arc state. Since this requires a linear traversal
2008 * callers are strongly encourage not to do this. However, it can be helpful
2009 * for targeted analysis so the functionality is provided.
2012 arc_buf_info(arc_buf_t
*ab
, arc_buf_info_t
*abi
, int state_index
)
2014 arc_buf_hdr_t
*hdr
= ab
->b_hdr
;
2015 l1arc_buf_hdr_t
*l1hdr
= NULL
;
2016 l2arc_buf_hdr_t
*l2hdr
= NULL
;
2017 arc_state_t
*state
= NULL
;
2019 memset(abi
, 0, sizeof (arc_buf_info_t
));
2024 abi
->abi_flags
= hdr
->b_flags
;
2026 if (HDR_HAS_L1HDR(hdr
)) {
2027 l1hdr
= &hdr
->b_l1hdr
;
2028 state
= l1hdr
->b_state
;
2030 if (HDR_HAS_L2HDR(hdr
))
2031 l2hdr
= &hdr
->b_l2hdr
;
2034 abi
->abi_bufcnt
= l1hdr
->b_bufcnt
;
2035 abi
->abi_access
= l1hdr
->b_arc_access
;
2036 abi
->abi_mru_hits
= l1hdr
->b_mru_hits
;
2037 abi
->abi_mru_ghost_hits
= l1hdr
->b_mru_ghost_hits
;
2038 abi
->abi_mfu_hits
= l1hdr
->b_mfu_hits
;
2039 abi
->abi_mfu_ghost_hits
= l1hdr
->b_mfu_ghost_hits
;
2040 abi
->abi_holds
= refcount_count(&l1hdr
->b_refcnt
);
2044 abi
->abi_l2arc_dattr
= l2hdr
->b_daddr
;
2045 abi
->abi_l2arc_hits
= l2hdr
->b_hits
;
2048 abi
->abi_state_type
= state
? state
->arcs_state
: ARC_STATE_ANON
;
2049 abi
->abi_state_contents
= arc_buf_type(hdr
);
2050 abi
->abi_size
= arc_hdr_size(hdr
);
2054 * Move the supplied buffer to the indicated state. The hash lock
2055 * for the buffer must be held by the caller.
2058 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*hdr
,
2059 kmutex_t
*hash_lock
)
2061 arc_state_t
*old_state
;
2064 boolean_t update_old
, update_new
;
2065 arc_buf_contents_t buftype
= arc_buf_type(hdr
);
2068 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2069 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2070 * L1 hdr doesn't always exist when we change state to arc_anon before
2071 * destroying a header, in which case reallocating to add the L1 hdr is
2074 if (HDR_HAS_L1HDR(hdr
)) {
2075 old_state
= hdr
->b_l1hdr
.b_state
;
2076 refcnt
= refcount_count(&hdr
->b_l1hdr
.b_refcnt
);
2077 bufcnt
= hdr
->b_l1hdr
.b_bufcnt
;
2078 update_old
= (bufcnt
> 0 || hdr
->b_l1hdr
.b_pabd
!= NULL
);
2080 old_state
= arc_l2c_only
;
2083 update_old
= B_FALSE
;
2085 update_new
= update_old
;
2087 ASSERT(MUTEX_HELD(hash_lock
));
2088 ASSERT3P(new_state
, !=, old_state
);
2089 ASSERT(!GHOST_STATE(new_state
) || bufcnt
== 0);
2090 ASSERT(old_state
!= arc_anon
|| bufcnt
<= 1);
2093 * If this buffer is evictable, transfer it from the
2094 * old state list to the new state list.
2097 if (old_state
!= arc_anon
&& old_state
!= arc_l2c_only
) {
2098 ASSERT(HDR_HAS_L1HDR(hdr
));
2099 multilist_remove(old_state
->arcs_list
[buftype
], hdr
);
2101 if (GHOST_STATE(old_state
)) {
2103 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2104 update_old
= B_TRUE
;
2106 arc_evictable_space_decrement(hdr
, old_state
);
2108 if (new_state
!= arc_anon
&& new_state
!= arc_l2c_only
) {
2110 * An L1 header always exists here, since if we're
2111 * moving to some L1-cached state (i.e. not l2c_only or
2112 * anonymous), we realloc the header to add an L1hdr
2115 ASSERT(HDR_HAS_L1HDR(hdr
));
2116 multilist_insert(new_state
->arcs_list
[buftype
], hdr
);
2118 if (GHOST_STATE(new_state
)) {
2120 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2121 update_new
= B_TRUE
;
2123 arc_evictable_space_increment(hdr
, new_state
);
2127 ASSERT(!HDR_EMPTY(hdr
));
2128 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(hdr
))
2129 buf_hash_remove(hdr
);
2131 /* adjust state sizes (ignore arc_l2c_only) */
2133 if (update_new
&& new_state
!= arc_l2c_only
) {
2134 ASSERT(HDR_HAS_L1HDR(hdr
));
2135 if (GHOST_STATE(new_state
)) {
2139 * When moving a header to a ghost state, we first
2140 * remove all arc buffers. Thus, we'll have a
2141 * bufcnt of zero, and no arc buffer to use for
2142 * the reference. As a result, we use the arc
2143 * header pointer for the reference.
2145 (void) refcount_add_many(&new_state
->arcs_size
,
2146 HDR_GET_LSIZE(hdr
), hdr
);
2147 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2150 uint32_t buffers
= 0;
2153 * Each individual buffer holds a unique reference,
2154 * thus we must remove each of these references one
2157 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2158 buf
= buf
->b_next
) {
2159 ASSERT3U(bufcnt
, !=, 0);
2163 * When the arc_buf_t is sharing the data
2164 * block with the hdr, the owner of the
2165 * reference belongs to the hdr. Only
2166 * add to the refcount if the arc_buf_t is
2169 if (arc_buf_is_shared(buf
))
2172 (void) refcount_add_many(&new_state
->arcs_size
,
2173 arc_buf_size(buf
), buf
);
2175 ASSERT3U(bufcnt
, ==, buffers
);
2177 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2178 (void) refcount_add_many(&new_state
->arcs_size
,
2179 arc_hdr_size(hdr
), hdr
);
2181 ASSERT(GHOST_STATE(old_state
));
2186 if (update_old
&& old_state
!= arc_l2c_only
) {
2187 ASSERT(HDR_HAS_L1HDR(hdr
));
2188 if (GHOST_STATE(old_state
)) {
2190 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2193 * When moving a header off of a ghost state,
2194 * the header will not contain any arc buffers.
2195 * We use the arc header pointer for the reference
2196 * which is exactly what we did when we put the
2197 * header on the ghost state.
2200 (void) refcount_remove_many(&old_state
->arcs_size
,
2201 HDR_GET_LSIZE(hdr
), hdr
);
2204 uint32_t buffers
= 0;
2207 * Each individual buffer holds a unique reference,
2208 * thus we must remove each of these references one
2211 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2212 buf
= buf
->b_next
) {
2213 ASSERT3U(bufcnt
, !=, 0);
2217 * When the arc_buf_t is sharing the data
2218 * block with the hdr, the owner of the
2219 * reference belongs to the hdr. Only
2220 * add to the refcount if the arc_buf_t is
2223 if (arc_buf_is_shared(buf
))
2226 (void) refcount_remove_many(
2227 &old_state
->arcs_size
, arc_buf_size(buf
),
2230 ASSERT3U(bufcnt
, ==, buffers
);
2231 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2232 (void) refcount_remove_many(
2233 &old_state
->arcs_size
, arc_hdr_size(hdr
), hdr
);
2237 if (HDR_HAS_L1HDR(hdr
))
2238 hdr
->b_l1hdr
.b_state
= new_state
;
2241 * L2 headers should never be on the L2 state list since they don't
2242 * have L1 headers allocated.
2244 ASSERT(multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]) &&
2245 multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]));
2249 arc_space_consume(uint64_t space
, arc_space_type_t type
)
2251 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2256 case ARC_SPACE_DATA
:
2257 ARCSTAT_INCR(arcstat_data_size
, space
);
2259 case ARC_SPACE_META
:
2260 ARCSTAT_INCR(arcstat_metadata_size
, space
);
2262 case ARC_SPACE_BONUS
:
2263 ARCSTAT_INCR(arcstat_bonus_size
, space
);
2265 case ARC_SPACE_DNODE
:
2266 ARCSTAT_INCR(arcstat_dnode_size
, space
);
2268 case ARC_SPACE_DBUF
:
2269 ARCSTAT_INCR(arcstat_dbuf_size
, space
);
2271 case ARC_SPACE_HDRS
:
2272 ARCSTAT_INCR(arcstat_hdr_size
, space
);
2274 case ARC_SPACE_L2HDRS
:
2275 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
2279 if (type
!= ARC_SPACE_DATA
)
2280 ARCSTAT_INCR(arcstat_meta_used
, space
);
2282 atomic_add_64(&arc_size
, space
);
2286 arc_space_return(uint64_t space
, arc_space_type_t type
)
2288 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2293 case ARC_SPACE_DATA
:
2294 ARCSTAT_INCR(arcstat_data_size
, -space
);
2296 case ARC_SPACE_META
:
2297 ARCSTAT_INCR(arcstat_metadata_size
, -space
);
2299 case ARC_SPACE_BONUS
:
2300 ARCSTAT_INCR(arcstat_bonus_size
, -space
);
2302 case ARC_SPACE_DNODE
:
2303 ARCSTAT_INCR(arcstat_dnode_size
, -space
);
2305 case ARC_SPACE_DBUF
:
2306 ARCSTAT_INCR(arcstat_dbuf_size
, -space
);
2308 case ARC_SPACE_HDRS
:
2309 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
2311 case ARC_SPACE_L2HDRS
:
2312 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
2316 if (type
!= ARC_SPACE_DATA
) {
2317 ASSERT(arc_meta_used
>= space
);
2318 if (arc_meta_max
< arc_meta_used
)
2319 arc_meta_max
= arc_meta_used
;
2320 ARCSTAT_INCR(arcstat_meta_used
, -space
);
2323 ASSERT(arc_size
>= space
);
2324 atomic_add_64(&arc_size
, -space
);
2328 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2329 * with the hdr's b_pabd.
2332 arc_can_share(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2335 * The criteria for sharing a hdr's data are:
2336 * 1. the hdr's compression matches the buf's compression
2337 * 2. the hdr doesn't need to be byteswapped
2338 * 3. the hdr isn't already being shared
2339 * 4. the buf is either compressed or it is the last buf in the hdr list
2341 * Criterion #4 maintains the invariant that shared uncompressed
2342 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2343 * might ask, "if a compressed buf is allocated first, won't that be the
2344 * last thing in the list?", but in that case it's impossible to create
2345 * a shared uncompressed buf anyway (because the hdr must be compressed
2346 * to have the compressed buf). You might also think that #3 is
2347 * sufficient to make this guarantee, however it's possible
2348 * (specifically in the rare L2ARC write race mentioned in
2349 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2350 * is sharable, but wasn't at the time of its allocation. Rather than
2351 * allow a new shared uncompressed buf to be created and then shuffle
2352 * the list around to make it the last element, this simply disallows
2353 * sharing if the new buf isn't the first to be added.
2355 ASSERT3P(buf
->b_hdr
, ==, hdr
);
2356 boolean_t hdr_compressed
= HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
;
2357 boolean_t buf_compressed
= ARC_BUF_COMPRESSED(buf
) != 0;
2358 return (buf_compressed
== hdr_compressed
&&
2359 hdr
->b_l1hdr
.b_byteswap
== DMU_BSWAP_NUMFUNCS
&&
2360 !HDR_SHARED_DATA(hdr
) &&
2361 (ARC_BUF_LAST(buf
) || ARC_BUF_COMPRESSED(buf
)));
2365 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2366 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2367 * copy was made successfully, or an error code otherwise.
2370 arc_buf_alloc_impl(arc_buf_hdr_t
*hdr
, void *tag
, boolean_t compressed
,
2371 boolean_t fill
, arc_buf_t
**ret
)
2375 ASSERT(HDR_HAS_L1HDR(hdr
));
2376 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2377 VERIFY(hdr
->b_type
== ARC_BUFC_DATA
||
2378 hdr
->b_type
== ARC_BUFC_METADATA
);
2379 ASSERT3P(ret
, !=, NULL
);
2380 ASSERT3P(*ret
, ==, NULL
);
2382 hdr
->b_l1hdr
.b_mru_hits
= 0;
2383 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
2384 hdr
->b_l1hdr
.b_mfu_hits
= 0;
2385 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
2386 hdr
->b_l1hdr
.b_l2_hits
= 0;
2388 buf
= *ret
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
2391 buf
->b_next
= hdr
->b_l1hdr
.b_buf
;
2394 add_reference(hdr
, tag
);
2397 * We're about to change the hdr's b_flags. We must either
2398 * hold the hash_lock or be undiscoverable.
2400 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2403 * Only honor requests for compressed bufs if the hdr is actually
2406 if (compressed
&& HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
)
2407 buf
->b_flags
|= ARC_BUF_FLAG_COMPRESSED
;
2410 * If the hdr's data can be shared then we share the data buffer and
2411 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2412 * allocate a new buffer to store the buf's data.
2414 * There are two additional restrictions here because we're sharing
2415 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2416 * actively involved in an L2ARC write, because if this buf is used by
2417 * an arc_write() then the hdr's data buffer will be released when the
2418 * write completes, even though the L2ARC write might still be using it.
2419 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2420 * need to be ABD-aware.
2422 boolean_t can_share
= arc_can_share(hdr
, buf
) && !HDR_L2_WRITING(hdr
) &&
2423 abd_is_linear(hdr
->b_l1hdr
.b_pabd
);
2425 /* Set up b_data and sharing */
2427 buf
->b_data
= abd_to_buf(hdr
->b_l1hdr
.b_pabd
);
2428 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2429 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2432 arc_get_data_buf(hdr
, arc_buf_size(buf
), buf
);
2433 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2435 VERIFY3P(buf
->b_data
, !=, NULL
);
2437 hdr
->b_l1hdr
.b_buf
= buf
;
2438 hdr
->b_l1hdr
.b_bufcnt
+= 1;
2441 * If the user wants the data from the hdr, we need to either copy or
2442 * decompress the data.
2445 return (arc_buf_fill(buf
, ARC_BUF_COMPRESSED(buf
) != 0));
2451 static char *arc_onloan_tag
= "onloan";
2454 arc_loaned_bytes_update(int64_t delta
)
2456 atomic_add_64(&arc_loaned_bytes
, delta
);
2458 /* assert that it did not wrap around */
2459 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
2463 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2464 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2465 * buffers must be returned to the arc before they can be used by the DMU or
2469 arc_loan_buf(spa_t
*spa
, boolean_t is_metadata
, int size
)
2471 arc_buf_t
*buf
= arc_alloc_buf(spa
, arc_onloan_tag
,
2472 is_metadata
? ARC_BUFC_METADATA
: ARC_BUFC_DATA
, size
);
2474 arc_loaned_bytes_update(size
);
2480 arc_loan_compressed_buf(spa_t
*spa
, uint64_t psize
, uint64_t lsize
,
2481 enum zio_compress compression_type
)
2483 arc_buf_t
*buf
= arc_alloc_compressed_buf(spa
, arc_onloan_tag
,
2484 psize
, lsize
, compression_type
);
2486 arc_loaned_bytes_update(psize
);
2493 * Return a loaned arc buffer to the arc.
2496 arc_return_buf(arc_buf_t
*buf
, void *tag
)
2498 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2500 ASSERT3P(buf
->b_data
, !=, NULL
);
2501 ASSERT(HDR_HAS_L1HDR(hdr
));
2502 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2503 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2505 arc_loaned_bytes_update(-arc_buf_size(buf
));
2508 /* Detach an arc_buf from a dbuf (tag) */
2510 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
2512 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2514 ASSERT3P(buf
->b_data
, !=, NULL
);
2515 ASSERT(HDR_HAS_L1HDR(hdr
));
2516 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2517 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2519 arc_loaned_bytes_update(arc_buf_size(buf
));
2523 l2arc_free_abd_on_write(abd_t
*abd
, size_t size
, arc_buf_contents_t type
)
2525 l2arc_data_free_t
*df
= kmem_alloc(sizeof (*df
), KM_SLEEP
);
2528 df
->l2df_size
= size
;
2529 df
->l2df_type
= type
;
2530 mutex_enter(&l2arc_free_on_write_mtx
);
2531 list_insert_head(l2arc_free_on_write
, df
);
2532 mutex_exit(&l2arc_free_on_write_mtx
);
2536 arc_hdr_free_on_write(arc_buf_hdr_t
*hdr
)
2538 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2539 arc_buf_contents_t type
= arc_buf_type(hdr
);
2540 uint64_t size
= arc_hdr_size(hdr
);
2542 /* protected by hash lock, if in the hash table */
2543 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
2544 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2545 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
2547 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2550 (void) refcount_remove_many(&state
->arcs_size
, size
, hdr
);
2551 if (type
== ARC_BUFC_METADATA
) {
2552 arc_space_return(size
, ARC_SPACE_META
);
2554 ASSERT(type
== ARC_BUFC_DATA
);
2555 arc_space_return(size
, ARC_SPACE_DATA
);
2558 l2arc_free_abd_on_write(hdr
->b_l1hdr
.b_pabd
, size
, type
);
2562 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2563 * data buffer, we transfer the refcount ownership to the hdr and update
2564 * the appropriate kstats.
2567 arc_share_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2569 ASSERT(arc_can_share(hdr
, buf
));
2570 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2571 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2574 * Start sharing the data buffer. We transfer the
2575 * refcount ownership to the hdr since it always owns
2576 * the refcount whenever an arc_buf_t is shared.
2578 refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, buf
, hdr
);
2579 hdr
->b_l1hdr
.b_pabd
= abd_get_from_buf(buf
->b_data
, arc_buf_size(buf
));
2580 abd_take_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
,
2581 HDR_ISTYPE_METADATA(hdr
));
2582 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2583 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2586 * Since we've transferred ownership to the hdr we need
2587 * to increment its compressed and uncompressed kstats and
2588 * decrement the overhead size.
2590 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
2591 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
2592 ARCSTAT_INCR(arcstat_overhead_size
, -arc_buf_size(buf
));
2596 arc_unshare_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2598 ASSERT(arc_buf_is_shared(buf
));
2599 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2600 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2603 * We are no longer sharing this buffer so we need
2604 * to transfer its ownership to the rightful owner.
2606 refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, hdr
, buf
);
2607 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2608 abd_release_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
);
2609 abd_put(hdr
->b_l1hdr
.b_pabd
);
2610 hdr
->b_l1hdr
.b_pabd
= NULL
;
2611 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
2614 * Since the buffer is no longer shared between
2615 * the arc buf and the hdr, count it as overhead.
2617 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
2618 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
2619 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2623 * Remove an arc_buf_t from the hdr's buf list and return the last
2624 * arc_buf_t on the list. If no buffers remain on the list then return
2628 arc_buf_remove(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2630 ASSERT(HDR_HAS_L1HDR(hdr
));
2631 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2633 arc_buf_t
**bufp
= &hdr
->b_l1hdr
.b_buf
;
2634 arc_buf_t
*lastbuf
= NULL
;
2637 * Remove the buf from the hdr list and locate the last
2638 * remaining buffer on the list.
2640 while (*bufp
!= NULL
) {
2642 *bufp
= buf
->b_next
;
2645 * If we've removed a buffer in the middle of
2646 * the list then update the lastbuf and update
2649 if (*bufp
!= NULL
) {
2651 bufp
= &(*bufp
)->b_next
;
2655 ASSERT3P(lastbuf
, !=, buf
);
2656 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, lastbuf
!= NULL
);
2657 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, hdr
->b_l1hdr
.b_buf
!= NULL
);
2658 IMPLY(lastbuf
!= NULL
, ARC_BUF_LAST(lastbuf
));
2664 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
2668 arc_buf_destroy_impl(arc_buf_t
*buf
)
2670 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2673 * Free up the data associated with the buf but only if we're not
2674 * sharing this with the hdr. If we are sharing it with the hdr, the
2675 * hdr is responsible for doing the free.
2677 if (buf
->b_data
!= NULL
) {
2679 * We're about to change the hdr's b_flags. We must either
2680 * hold the hash_lock or be undiscoverable.
2682 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2684 arc_cksum_verify(buf
);
2685 arc_buf_unwatch(buf
);
2687 if (arc_buf_is_shared(buf
)) {
2688 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2690 uint64_t size
= arc_buf_size(buf
);
2691 arc_free_data_buf(hdr
, buf
->b_data
, size
, buf
);
2692 ARCSTAT_INCR(arcstat_overhead_size
, -size
);
2696 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
2697 hdr
->b_l1hdr
.b_bufcnt
-= 1;
2700 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
2702 if (ARC_BUF_SHARED(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
2704 * If the current arc_buf_t is sharing its data buffer with the
2705 * hdr, then reassign the hdr's b_pabd to share it with the new
2706 * buffer at the end of the list. The shared buffer is always
2707 * the last one on the hdr's buffer list.
2709 * There is an equivalent case for compressed bufs, but since
2710 * they aren't guaranteed to be the last buf in the list and
2711 * that is an exceedingly rare case, we just allow that space be
2712 * wasted temporarily.
2714 if (lastbuf
!= NULL
) {
2715 /* Only one buf can be shared at once */
2716 VERIFY(!arc_buf_is_shared(lastbuf
));
2717 /* hdr is uncompressed so can't have compressed buf */
2718 VERIFY(!ARC_BUF_COMPRESSED(lastbuf
));
2720 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2721 arc_hdr_free_pabd(hdr
);
2724 * We must setup a new shared block between the
2725 * last buffer and the hdr. The data would have
2726 * been allocated by the arc buf so we need to transfer
2727 * ownership to the hdr since it's now being shared.
2729 arc_share_buf(hdr
, lastbuf
);
2731 } else if (HDR_SHARED_DATA(hdr
)) {
2733 * Uncompressed shared buffers are always at the end
2734 * of the list. Compressed buffers don't have the
2735 * same requirements. This makes it hard to
2736 * simply assert that the lastbuf is shared so
2737 * we rely on the hdr's compression flags to determine
2738 * if we have a compressed, shared buffer.
2740 ASSERT3P(lastbuf
, !=, NULL
);
2741 ASSERT(arc_buf_is_shared(lastbuf
) ||
2742 HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
2746 * Free the checksum if we're removing the last uncompressed buf from
2749 if (!arc_hdr_has_uncompressed_buf(hdr
)) {
2750 arc_cksum_free(hdr
);
2753 /* clean up the buf */
2755 kmem_cache_free(buf_cache
, buf
);
2759 arc_hdr_alloc_pabd(arc_buf_hdr_t
*hdr
)
2761 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2762 ASSERT(HDR_HAS_L1HDR(hdr
));
2763 ASSERT(!HDR_SHARED_DATA(hdr
));
2765 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2766 hdr
->b_l1hdr
.b_pabd
= arc_get_data_abd(hdr
, arc_hdr_size(hdr
), hdr
);
2767 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
2768 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2770 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
2771 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
2775 arc_hdr_free_pabd(arc_buf_hdr_t
*hdr
)
2777 ASSERT(HDR_HAS_L1HDR(hdr
));
2778 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2781 * If the hdr is currently being written to the l2arc then
2782 * we defer freeing the data by adding it to the l2arc_free_on_write
2783 * list. The l2arc will free the data once it's finished
2784 * writing it to the l2arc device.
2786 if (HDR_L2_WRITING(hdr
)) {
2787 arc_hdr_free_on_write(hdr
);
2788 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
2790 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
,
2791 arc_hdr_size(hdr
), hdr
);
2793 hdr
->b_l1hdr
.b_pabd
= NULL
;
2794 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
2796 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
2797 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
2800 static arc_buf_hdr_t
*
2801 arc_hdr_alloc(uint64_t spa
, int32_t psize
, int32_t lsize
,
2802 enum zio_compress compression_type
, arc_buf_contents_t type
)
2806 VERIFY(type
== ARC_BUFC_DATA
|| type
== ARC_BUFC_METADATA
);
2808 hdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
2809 ASSERT(HDR_EMPTY(hdr
));
2810 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2811 HDR_SET_PSIZE(hdr
, psize
);
2812 HDR_SET_LSIZE(hdr
, lsize
);
2816 arc_hdr_set_flags(hdr
, arc_bufc_to_flags(type
) | ARC_FLAG_HAS_L1HDR
);
2817 arc_hdr_set_compress(hdr
, compression_type
);
2819 hdr
->b_l1hdr
.b_state
= arc_anon
;
2820 hdr
->b_l1hdr
.b_arc_access
= 0;
2821 hdr
->b_l1hdr
.b_bufcnt
= 0;
2822 hdr
->b_l1hdr
.b_buf
= NULL
;
2825 * Allocate the hdr's buffer. This will contain either
2826 * the compressed or uncompressed data depending on the block
2827 * it references and compressed arc enablement.
2829 arc_hdr_alloc_pabd(hdr
);
2830 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2836 * Transition between the two allocation states for the arc_buf_hdr struct.
2837 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
2838 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
2839 * version is used when a cache buffer is only in the L2ARC in order to reduce
2842 static arc_buf_hdr_t
*
2843 arc_hdr_realloc(arc_buf_hdr_t
*hdr
, kmem_cache_t
*old
, kmem_cache_t
*new)
2845 arc_buf_hdr_t
*nhdr
;
2846 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
2848 ASSERT(HDR_HAS_L2HDR(hdr
));
2849 ASSERT((old
== hdr_full_cache
&& new == hdr_l2only_cache
) ||
2850 (old
== hdr_l2only_cache
&& new == hdr_full_cache
));
2852 nhdr
= kmem_cache_alloc(new, KM_PUSHPAGE
);
2854 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
2855 buf_hash_remove(hdr
);
2857 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
2859 if (new == hdr_full_cache
) {
2860 arc_hdr_set_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
2862 * arc_access and arc_change_state need to be aware that a
2863 * header has just come out of L2ARC, so we set its state to
2864 * l2c_only even though it's about to change.
2866 nhdr
->b_l1hdr
.b_state
= arc_l2c_only
;
2868 /* Verify previous threads set to NULL before freeing */
2869 ASSERT3P(nhdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2871 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2872 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
2873 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2876 * If we've reached here, We must have been called from
2877 * arc_evict_hdr(), as such we should have already been
2878 * removed from any ghost list we were previously on
2879 * (which protects us from racing with arc_evict_state),
2880 * thus no locking is needed during this check.
2882 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
2885 * A buffer must not be moved into the arc_l2c_only
2886 * state if it's not finished being written out to the
2887 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
2888 * might try to be accessed, even though it was removed.
2890 VERIFY(!HDR_L2_WRITING(hdr
));
2891 VERIFY3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2893 arc_hdr_clear_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
2896 * The header has been reallocated so we need to re-insert it into any
2899 (void) buf_hash_insert(nhdr
, NULL
);
2901 ASSERT(list_link_active(&hdr
->b_l2hdr
.b_l2node
));
2903 mutex_enter(&dev
->l2ad_mtx
);
2906 * We must place the realloc'ed header back into the list at
2907 * the same spot. Otherwise, if it's placed earlier in the list,
2908 * l2arc_write_buffers() could find it during the function's
2909 * write phase, and try to write it out to the l2arc.
2911 list_insert_after(&dev
->l2ad_buflist
, hdr
, nhdr
);
2912 list_remove(&dev
->l2ad_buflist
, hdr
);
2914 mutex_exit(&dev
->l2ad_mtx
);
2917 * Since we're using the pointer address as the tag when
2918 * incrementing and decrementing the l2ad_alloc refcount, we
2919 * must remove the old pointer (that we're about to destroy) and
2920 * add the new pointer to the refcount. Otherwise we'd remove
2921 * the wrong pointer address when calling arc_hdr_destroy() later.
2924 (void) refcount_remove_many(&dev
->l2ad_alloc
, arc_hdr_size(hdr
), hdr
);
2925 (void) refcount_add_many(&dev
->l2ad_alloc
, arc_hdr_size(nhdr
), nhdr
);
2927 buf_discard_identity(hdr
);
2928 kmem_cache_free(old
, hdr
);
2934 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
2935 * The buf is returned thawed since we expect the consumer to modify it.
2938 arc_alloc_buf(spa_t
*spa
, void *tag
, arc_buf_contents_t type
, int32_t size
)
2940 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), size
, size
,
2941 ZIO_COMPRESS_OFF
, type
);
2942 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
2944 arc_buf_t
*buf
= NULL
;
2945 VERIFY0(arc_buf_alloc_impl(hdr
, tag
, B_FALSE
, B_FALSE
, &buf
));
2952 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
2953 * for bufs containing metadata.
2956 arc_alloc_compressed_buf(spa_t
*spa
, void *tag
, uint64_t psize
, uint64_t lsize
,
2957 enum zio_compress compression_type
)
2959 ASSERT3U(lsize
, >, 0);
2960 ASSERT3U(lsize
, >=, psize
);
2961 ASSERT(compression_type
> ZIO_COMPRESS_OFF
);
2962 ASSERT(compression_type
< ZIO_COMPRESS_FUNCTIONS
);
2964 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
2965 compression_type
, ARC_BUFC_DATA
);
2966 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
2968 arc_buf_t
*buf
= NULL
;
2969 VERIFY0(arc_buf_alloc_impl(hdr
, tag
, B_TRUE
, B_FALSE
, &buf
));
2971 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2973 if (!arc_buf_is_shared(buf
)) {
2975 * To ensure that the hdr has the correct data in it if we call
2976 * arc_decompress() on this buf before it's been written to
2977 * disk, it's easiest if we just set up sharing between the
2980 ASSERT(!abd_is_linear(hdr
->b_l1hdr
.b_pabd
));
2981 arc_hdr_free_pabd(hdr
);
2982 arc_share_buf(hdr
, buf
);
2989 arc_hdr_l2hdr_destroy(arc_buf_hdr_t
*hdr
)
2991 l2arc_buf_hdr_t
*l2hdr
= &hdr
->b_l2hdr
;
2992 l2arc_dev_t
*dev
= l2hdr
->b_dev
;
2993 uint64_t psize
= arc_hdr_size(hdr
);
2995 ASSERT(MUTEX_HELD(&dev
->l2ad_mtx
));
2996 ASSERT(HDR_HAS_L2HDR(hdr
));
2998 list_remove(&dev
->l2ad_buflist
, hdr
);
3000 ARCSTAT_INCR(arcstat_l2_psize
, -psize
);
3001 ARCSTAT_INCR(arcstat_l2_lsize
, -HDR_GET_LSIZE(hdr
));
3003 vdev_space_update(dev
->l2ad_vdev
, -psize
, 0, 0);
3005 (void) refcount_remove_many(&dev
->l2ad_alloc
, psize
, hdr
);
3006 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
3010 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
3012 if (HDR_HAS_L1HDR(hdr
)) {
3013 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
||
3014 hdr
->b_l1hdr
.b_bufcnt
> 0);
3015 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3016 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
3018 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3019 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
3021 if (!HDR_EMPTY(hdr
))
3022 buf_discard_identity(hdr
);
3024 if (HDR_HAS_L2HDR(hdr
)) {
3025 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
3026 boolean_t buflist_held
= MUTEX_HELD(&dev
->l2ad_mtx
);
3029 mutex_enter(&dev
->l2ad_mtx
);
3032 * Even though we checked this conditional above, we
3033 * need to check this again now that we have the
3034 * l2ad_mtx. This is because we could be racing with
3035 * another thread calling l2arc_evict() which might have
3036 * destroyed this header's L2 portion as we were waiting
3037 * to acquire the l2ad_mtx. If that happens, we don't
3038 * want to re-destroy the header's L2 portion.
3040 if (HDR_HAS_L2HDR(hdr
))
3041 arc_hdr_l2hdr_destroy(hdr
);
3044 mutex_exit(&dev
->l2ad_mtx
);
3047 if (HDR_HAS_L1HDR(hdr
)) {
3048 arc_cksum_free(hdr
);
3050 while (hdr
->b_l1hdr
.b_buf
!= NULL
)
3051 arc_buf_destroy_impl(hdr
->b_l1hdr
.b_buf
);
3053 if (hdr
->b_l1hdr
.b_pabd
!= NULL
)
3054 arc_hdr_free_pabd(hdr
);
3057 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
3058 if (HDR_HAS_L1HDR(hdr
)) {
3059 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3060 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
3061 kmem_cache_free(hdr_full_cache
, hdr
);
3063 kmem_cache_free(hdr_l2only_cache
, hdr
);
3068 arc_buf_destroy(arc_buf_t
*buf
, void* tag
)
3070 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3071 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
3073 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
3074 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
3075 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3076 VERIFY0(remove_reference(hdr
, NULL
, tag
));
3077 arc_hdr_destroy(hdr
);
3081 mutex_enter(hash_lock
);
3082 ASSERT3P(hdr
, ==, buf
->b_hdr
);
3083 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
3084 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3085 ASSERT3P(hdr
->b_l1hdr
.b_state
, !=, arc_anon
);
3086 ASSERT3P(buf
->b_data
, !=, NULL
);
3088 (void) remove_reference(hdr
, hash_lock
, tag
);
3089 arc_buf_destroy_impl(buf
);
3090 mutex_exit(hash_lock
);
3094 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3095 * state of the header is dependent on its state prior to entering this
3096 * function. The following transitions are possible:
3098 * - arc_mru -> arc_mru_ghost
3099 * - arc_mfu -> arc_mfu_ghost
3100 * - arc_mru_ghost -> arc_l2c_only
3101 * - arc_mru_ghost -> deleted
3102 * - arc_mfu_ghost -> arc_l2c_only
3103 * - arc_mfu_ghost -> deleted
3106 arc_evict_hdr(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
3108 arc_state_t
*evicted_state
, *state
;
3109 int64_t bytes_evicted
= 0;
3111 ASSERT(MUTEX_HELD(hash_lock
));
3112 ASSERT(HDR_HAS_L1HDR(hdr
));
3114 state
= hdr
->b_l1hdr
.b_state
;
3115 if (GHOST_STATE(state
)) {
3116 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3117 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
3120 * l2arc_write_buffers() relies on a header's L1 portion
3121 * (i.e. its b_pabd field) during it's write phase.
3122 * Thus, we cannot push a header onto the arc_l2c_only
3123 * state (removing its L1 piece) until the header is
3124 * done being written to the l2arc.
3126 if (HDR_HAS_L2HDR(hdr
) && HDR_L2_WRITING(hdr
)) {
3127 ARCSTAT_BUMP(arcstat_evict_l2_skip
);
3128 return (bytes_evicted
);
3131 ARCSTAT_BUMP(arcstat_deleted
);
3132 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3134 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, hdr
);
3136 if (HDR_HAS_L2HDR(hdr
)) {
3137 ASSERT(hdr
->b_l1hdr
.b_pabd
== NULL
);
3139 * This buffer is cached on the 2nd Level ARC;
3140 * don't destroy the header.
3142 arc_change_state(arc_l2c_only
, hdr
, hash_lock
);
3144 * dropping from L1+L2 cached to L2-only,
3145 * realloc to remove the L1 header.
3147 hdr
= arc_hdr_realloc(hdr
, hdr_full_cache
,
3150 arc_change_state(arc_anon
, hdr
, hash_lock
);
3151 arc_hdr_destroy(hdr
);
3153 return (bytes_evicted
);
3156 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
3157 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
3159 /* prefetch buffers have a minimum lifespan */
3160 if (HDR_IO_IN_PROGRESS(hdr
) ||
3161 ((hdr
->b_flags
& (ARC_FLAG_PREFETCH
| ARC_FLAG_INDIRECT
)) &&
3162 ddi_get_lbolt() - hdr
->b_l1hdr
.b_arc_access
<
3163 arc_min_prefetch_lifespan
)) {
3164 ARCSTAT_BUMP(arcstat_evict_skip
);
3165 return (bytes_evicted
);
3168 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
3169 while (hdr
->b_l1hdr
.b_buf
) {
3170 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
3171 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
3172 ARCSTAT_BUMP(arcstat_mutex_miss
);
3175 if (buf
->b_data
!= NULL
)
3176 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3177 mutex_exit(&buf
->b_evict_lock
);
3178 arc_buf_destroy_impl(buf
);
3181 if (HDR_HAS_L2HDR(hdr
)) {
3182 ARCSTAT_INCR(arcstat_evict_l2_cached
, HDR_GET_LSIZE(hdr
));
3184 if (l2arc_write_eligible(hdr
->b_spa
, hdr
)) {
3185 ARCSTAT_INCR(arcstat_evict_l2_eligible
,
3186 HDR_GET_LSIZE(hdr
));
3188 ARCSTAT_INCR(arcstat_evict_l2_ineligible
,
3189 HDR_GET_LSIZE(hdr
));
3193 if (hdr
->b_l1hdr
.b_bufcnt
== 0) {
3194 arc_cksum_free(hdr
);
3196 bytes_evicted
+= arc_hdr_size(hdr
);
3199 * If this hdr is being evicted and has a compressed
3200 * buffer then we discard it here before we change states.
3201 * This ensures that the accounting is updated correctly
3202 * in arc_free_data_impl().
3204 arc_hdr_free_pabd(hdr
);
3206 arc_change_state(evicted_state
, hdr
, hash_lock
);
3207 ASSERT(HDR_IN_HASH_TABLE(hdr
));
3208 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
3209 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, hdr
);
3212 return (bytes_evicted
);
3216 arc_evict_state_impl(multilist_t
*ml
, int idx
, arc_buf_hdr_t
*marker
,
3217 uint64_t spa
, int64_t bytes
)
3219 multilist_sublist_t
*mls
;
3220 uint64_t bytes_evicted
= 0;
3222 kmutex_t
*hash_lock
;
3223 int evict_count
= 0;
3225 ASSERT3P(marker
, !=, NULL
);
3226 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3228 mls
= multilist_sublist_lock(ml
, idx
);
3230 for (hdr
= multilist_sublist_prev(mls
, marker
); hdr
!= NULL
;
3231 hdr
= multilist_sublist_prev(mls
, marker
)) {
3232 if ((bytes
!= ARC_EVICT_ALL
&& bytes_evicted
>= bytes
) ||
3233 (evict_count
>= zfs_arc_evict_batch_limit
))
3237 * To keep our iteration location, move the marker
3238 * forward. Since we're not holding hdr's hash lock, we
3239 * must be very careful and not remove 'hdr' from the
3240 * sublist. Otherwise, other consumers might mistake the
3241 * 'hdr' as not being on a sublist when they call the
3242 * multilist_link_active() function (they all rely on
3243 * the hash lock protecting concurrent insertions and
3244 * removals). multilist_sublist_move_forward() was
3245 * specifically implemented to ensure this is the case
3246 * (only 'marker' will be removed and re-inserted).
3248 multilist_sublist_move_forward(mls
, marker
);
3251 * The only case where the b_spa field should ever be
3252 * zero, is the marker headers inserted by
3253 * arc_evict_state(). It's possible for multiple threads
3254 * to be calling arc_evict_state() concurrently (e.g.
3255 * dsl_pool_close() and zio_inject_fault()), so we must
3256 * skip any markers we see from these other threads.
3258 if (hdr
->b_spa
== 0)
3261 /* we're only interested in evicting buffers of a certain spa */
3262 if (spa
!= 0 && hdr
->b_spa
!= spa
) {
3263 ARCSTAT_BUMP(arcstat_evict_skip
);
3267 hash_lock
= HDR_LOCK(hdr
);
3270 * We aren't calling this function from any code path
3271 * that would already be holding a hash lock, so we're
3272 * asserting on this assumption to be defensive in case
3273 * this ever changes. Without this check, it would be
3274 * possible to incorrectly increment arcstat_mutex_miss
3275 * below (e.g. if the code changed such that we called
3276 * this function with a hash lock held).
3278 ASSERT(!MUTEX_HELD(hash_lock
));
3280 if (mutex_tryenter(hash_lock
)) {
3281 uint64_t evicted
= arc_evict_hdr(hdr
, hash_lock
);
3282 mutex_exit(hash_lock
);
3284 bytes_evicted
+= evicted
;
3287 * If evicted is zero, arc_evict_hdr() must have
3288 * decided to skip this header, don't increment
3289 * evict_count in this case.
3295 * If arc_size isn't overflowing, signal any
3296 * threads that might happen to be waiting.
3298 * For each header evicted, we wake up a single
3299 * thread. If we used cv_broadcast, we could
3300 * wake up "too many" threads causing arc_size
3301 * to significantly overflow arc_c; since
3302 * arc_get_data_impl() doesn't check for overflow
3303 * when it's woken up (it doesn't because it's
3304 * possible for the ARC to be overflowing while
3305 * full of un-evictable buffers, and the
3306 * function should proceed in this case).
3308 * If threads are left sleeping, due to not
3309 * using cv_broadcast, they will be woken up
3310 * just before arc_reclaim_thread() sleeps.
3312 mutex_enter(&arc_reclaim_lock
);
3313 if (!arc_is_overflowing())
3314 cv_signal(&arc_reclaim_waiters_cv
);
3315 mutex_exit(&arc_reclaim_lock
);
3317 ARCSTAT_BUMP(arcstat_mutex_miss
);
3321 multilist_sublist_unlock(mls
);
3323 return (bytes_evicted
);
3327 * Evict buffers from the given arc state, until we've removed the
3328 * specified number of bytes. Move the removed buffers to the
3329 * appropriate evict state.
3331 * This function makes a "best effort". It skips over any buffers
3332 * it can't get a hash_lock on, and so, may not catch all candidates.
3333 * It may also return without evicting as much space as requested.
3335 * If bytes is specified using the special value ARC_EVICT_ALL, this
3336 * will evict all available (i.e. unlocked and evictable) buffers from
3337 * the given arc state; which is used by arc_flush().
3340 arc_evict_state(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
3341 arc_buf_contents_t type
)
3343 uint64_t total_evicted
= 0;
3344 multilist_t
*ml
= state
->arcs_list
[type
];
3346 arc_buf_hdr_t
**markers
;
3349 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3351 num_sublists
= multilist_get_num_sublists(ml
);
3354 * If we've tried to evict from each sublist, made some
3355 * progress, but still have not hit the target number of bytes
3356 * to evict, we want to keep trying. The markers allow us to
3357 * pick up where we left off for each individual sublist, rather
3358 * than starting from the tail each time.
3360 markers
= kmem_zalloc(sizeof (*markers
) * num_sublists
, KM_SLEEP
);
3361 for (i
= 0; i
< num_sublists
; i
++) {
3362 multilist_sublist_t
*mls
;
3364 markers
[i
] = kmem_cache_alloc(hdr_full_cache
, KM_SLEEP
);
3367 * A b_spa of 0 is used to indicate that this header is
3368 * a marker. This fact is used in arc_adjust_type() and
3369 * arc_evict_state_impl().
3371 markers
[i
]->b_spa
= 0;
3373 mls
= multilist_sublist_lock(ml
, i
);
3374 multilist_sublist_insert_tail(mls
, markers
[i
]);
3375 multilist_sublist_unlock(mls
);
3379 * While we haven't hit our target number of bytes to evict, or
3380 * we're evicting all available buffers.
3382 while (total_evicted
< bytes
|| bytes
== ARC_EVICT_ALL
) {
3383 int sublist_idx
= multilist_get_random_index(ml
);
3384 uint64_t scan_evicted
= 0;
3387 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
3388 * Request that 10% of the LRUs be scanned by the superblock
3391 if (type
== ARC_BUFC_DATA
&& arc_dnode_size
> arc_dnode_limit
)
3392 arc_prune_async((arc_dnode_size
- arc_dnode_limit
) /
3393 sizeof (dnode_t
) / zfs_arc_dnode_reduce_percent
);
3396 * Start eviction using a randomly selected sublist,
3397 * this is to try and evenly balance eviction across all
3398 * sublists. Always starting at the same sublist
3399 * (e.g. index 0) would cause evictions to favor certain
3400 * sublists over others.
3402 for (i
= 0; i
< num_sublists
; i
++) {
3403 uint64_t bytes_remaining
;
3404 uint64_t bytes_evicted
;
3406 if (bytes
== ARC_EVICT_ALL
)
3407 bytes_remaining
= ARC_EVICT_ALL
;
3408 else if (total_evicted
< bytes
)
3409 bytes_remaining
= bytes
- total_evicted
;
3413 bytes_evicted
= arc_evict_state_impl(ml
, sublist_idx
,
3414 markers
[sublist_idx
], spa
, bytes_remaining
);
3416 scan_evicted
+= bytes_evicted
;
3417 total_evicted
+= bytes_evicted
;
3419 /* we've reached the end, wrap to the beginning */
3420 if (++sublist_idx
>= num_sublists
)
3425 * If we didn't evict anything during this scan, we have
3426 * no reason to believe we'll evict more during another
3427 * scan, so break the loop.
3429 if (scan_evicted
== 0) {
3430 /* This isn't possible, let's make that obvious */
3431 ASSERT3S(bytes
, !=, 0);
3434 * When bytes is ARC_EVICT_ALL, the only way to
3435 * break the loop is when scan_evicted is zero.
3436 * In that case, we actually have evicted enough,
3437 * so we don't want to increment the kstat.
3439 if (bytes
!= ARC_EVICT_ALL
) {
3440 ASSERT3S(total_evicted
, <, bytes
);
3441 ARCSTAT_BUMP(arcstat_evict_not_enough
);
3448 for (i
= 0; i
< num_sublists
; i
++) {
3449 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
3450 multilist_sublist_remove(mls
, markers
[i
]);
3451 multilist_sublist_unlock(mls
);
3453 kmem_cache_free(hdr_full_cache
, markers
[i
]);
3455 kmem_free(markers
, sizeof (*markers
) * num_sublists
);
3457 return (total_evicted
);
3461 * Flush all "evictable" data of the given type from the arc state
3462 * specified. This will not evict any "active" buffers (i.e. referenced).
3464 * When 'retry' is set to B_FALSE, the function will make a single pass
3465 * over the state and evict any buffers that it can. Since it doesn't
3466 * continually retry the eviction, it might end up leaving some buffers
3467 * in the ARC due to lock misses.
3469 * When 'retry' is set to B_TRUE, the function will continually retry the
3470 * eviction until *all* evictable buffers have been removed from the
3471 * state. As a result, if concurrent insertions into the state are
3472 * allowed (e.g. if the ARC isn't shutting down), this function might
3473 * wind up in an infinite loop, continually trying to evict buffers.
3476 arc_flush_state(arc_state_t
*state
, uint64_t spa
, arc_buf_contents_t type
,
3479 uint64_t evicted
= 0;
3481 while (refcount_count(&state
->arcs_esize
[type
]) != 0) {
3482 evicted
+= arc_evict_state(state
, spa
, ARC_EVICT_ALL
, type
);
3492 * Helper function for arc_prune_async() it is responsible for safely
3493 * handling the execution of a registered arc_prune_func_t.
3496 arc_prune_task(void *ptr
)
3498 arc_prune_t
*ap
= (arc_prune_t
*)ptr
;
3499 arc_prune_func_t
*func
= ap
->p_pfunc
;
3502 func(ap
->p_adjust
, ap
->p_private
);
3504 refcount_remove(&ap
->p_refcnt
, func
);
3508 * Notify registered consumers they must drop holds on a portion of the ARC
3509 * buffered they reference. This provides a mechanism to ensure the ARC can
3510 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
3511 * is analogous to dnlc_reduce_cache() but more generic.
3513 * This operation is performed asynchronously so it may be safely called
3514 * in the context of the arc_reclaim_thread(). A reference is taken here
3515 * for each registered arc_prune_t and the arc_prune_task() is responsible
3516 * for releasing it once the registered arc_prune_func_t has completed.
3519 arc_prune_async(int64_t adjust
)
3523 mutex_enter(&arc_prune_mtx
);
3524 for (ap
= list_head(&arc_prune_list
); ap
!= NULL
;
3525 ap
= list_next(&arc_prune_list
, ap
)) {
3527 if (refcount_count(&ap
->p_refcnt
) >= 2)
3530 refcount_add(&ap
->p_refcnt
, ap
->p_pfunc
);
3531 ap
->p_adjust
= adjust
;
3532 if (taskq_dispatch(arc_prune_taskq
, arc_prune_task
,
3533 ap
, TQ_SLEEP
) == TASKQID_INVALID
) {
3534 refcount_remove(&ap
->p_refcnt
, ap
->p_pfunc
);
3537 ARCSTAT_BUMP(arcstat_prune
);
3539 mutex_exit(&arc_prune_mtx
);
3543 * Evict the specified number of bytes from the state specified,
3544 * restricting eviction to the spa and type given. This function
3545 * prevents us from trying to evict more from a state's list than
3546 * is "evictable", and to skip evicting altogether when passed a
3547 * negative value for "bytes". In contrast, arc_evict_state() will
3548 * evict everything it can, when passed a negative value for "bytes".
3551 arc_adjust_impl(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
3552 arc_buf_contents_t type
)
3556 if (bytes
> 0 && refcount_count(&state
->arcs_esize
[type
]) > 0) {
3557 delta
= MIN(refcount_count(&state
->arcs_esize
[type
]), bytes
);
3558 return (arc_evict_state(state
, spa
, delta
, type
));
3565 * The goal of this function is to evict enough meta data buffers from the
3566 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
3567 * more complicated than it appears because it is common for data buffers
3568 * to have holds on meta data buffers. In addition, dnode meta data buffers
3569 * will be held by the dnodes in the block preventing them from being freed.
3570 * This means we can't simply traverse the ARC and expect to always find
3571 * enough unheld meta data buffer to release.
3573 * Therefore, this function has been updated to make alternating passes
3574 * over the ARC releasing data buffers and then newly unheld meta data
3575 * buffers. This ensures forward progress is maintained and arc_meta_used
3576 * will decrease. Normally this is sufficient, but if required the ARC
3577 * will call the registered prune callbacks causing dentry and inodes to
3578 * be dropped from the VFS cache. This will make dnode meta data buffers
3579 * available for reclaim.
3582 arc_adjust_meta_balanced(void)
3584 int64_t delta
, prune
= 0, adjustmnt
;
3585 uint64_t total_evicted
= 0;
3586 arc_buf_contents_t type
= ARC_BUFC_DATA
;
3587 int restarts
= MAX(zfs_arc_meta_adjust_restarts
, 0);
3591 * This slightly differs than the way we evict from the mru in
3592 * arc_adjust because we don't have a "target" value (i.e. no
3593 * "meta" arc_p). As a result, I think we can completely
3594 * cannibalize the metadata in the MRU before we evict the
3595 * metadata from the MFU. I think we probably need to implement a
3596 * "metadata arc_p" value to do this properly.
3598 adjustmnt
= arc_meta_used
- arc_meta_limit
;
3600 if (adjustmnt
> 0 && refcount_count(&arc_mru
->arcs_esize
[type
]) > 0) {
3601 delta
= MIN(refcount_count(&arc_mru
->arcs_esize
[type
]),
3603 total_evicted
+= arc_adjust_impl(arc_mru
, 0, delta
, type
);
3608 * We can't afford to recalculate adjustmnt here. If we do,
3609 * new metadata buffers can sneak into the MRU or ANON lists,
3610 * thus penalize the MFU metadata. Although the fudge factor is
3611 * small, it has been empirically shown to be significant for
3612 * certain workloads (e.g. creating many empty directories). As
3613 * such, we use the original calculation for adjustmnt, and
3614 * simply decrement the amount of data evicted from the MRU.
3617 if (adjustmnt
> 0 && refcount_count(&arc_mfu
->arcs_esize
[type
]) > 0) {
3618 delta
= MIN(refcount_count(&arc_mfu
->arcs_esize
[type
]),
3620 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, delta
, type
);
3623 adjustmnt
= arc_meta_used
- arc_meta_limit
;
3625 if (adjustmnt
> 0 &&
3626 refcount_count(&arc_mru_ghost
->arcs_esize
[type
]) > 0) {
3627 delta
= MIN(adjustmnt
,
3628 refcount_count(&arc_mru_ghost
->arcs_esize
[type
]));
3629 total_evicted
+= arc_adjust_impl(arc_mru_ghost
, 0, delta
, type
);
3633 if (adjustmnt
> 0 &&
3634 refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]) > 0) {
3635 delta
= MIN(adjustmnt
,
3636 refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]));
3637 total_evicted
+= arc_adjust_impl(arc_mfu_ghost
, 0, delta
, type
);
3641 * If after attempting to make the requested adjustment to the ARC
3642 * the meta limit is still being exceeded then request that the
3643 * higher layers drop some cached objects which have holds on ARC
3644 * meta buffers. Requests to the upper layers will be made with
3645 * increasingly large scan sizes until the ARC is below the limit.
3647 if (arc_meta_used
> arc_meta_limit
) {
3648 if (type
== ARC_BUFC_DATA
) {
3649 type
= ARC_BUFC_METADATA
;
3651 type
= ARC_BUFC_DATA
;
3653 if (zfs_arc_meta_prune
) {
3654 prune
+= zfs_arc_meta_prune
;
3655 arc_prune_async(prune
);
3664 return (total_evicted
);
3668 * Evict metadata buffers from the cache, such that arc_meta_used is
3669 * capped by the arc_meta_limit tunable.
3672 arc_adjust_meta_only(void)
3674 uint64_t total_evicted
= 0;
3678 * If we're over the meta limit, we want to evict enough
3679 * metadata to get back under the meta limit. We don't want to
3680 * evict so much that we drop the MRU below arc_p, though. If
3681 * we're over the meta limit more than we're over arc_p, we
3682 * evict some from the MRU here, and some from the MFU below.
3684 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
3685 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
3686 refcount_count(&arc_mru
->arcs_size
) - arc_p
));
3688 total_evicted
+= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3691 * Similar to the above, we want to evict enough bytes to get us
3692 * below the meta limit, but not so much as to drop us below the
3693 * space allotted to the MFU (which is defined as arc_c - arc_p).
3695 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
3696 (int64_t)(refcount_count(&arc_mfu
->arcs_size
) - (arc_c
- arc_p
)));
3698 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3700 return (total_evicted
);
3704 arc_adjust_meta(void)
3706 if (zfs_arc_meta_strategy
== ARC_STRATEGY_META_ONLY
)
3707 return (arc_adjust_meta_only());
3709 return (arc_adjust_meta_balanced());
3713 * Return the type of the oldest buffer in the given arc state
3715 * This function will select a random sublist of type ARC_BUFC_DATA and
3716 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3717 * is compared, and the type which contains the "older" buffer will be
3720 static arc_buf_contents_t
3721 arc_adjust_type(arc_state_t
*state
)
3723 multilist_t
*data_ml
= state
->arcs_list
[ARC_BUFC_DATA
];
3724 multilist_t
*meta_ml
= state
->arcs_list
[ARC_BUFC_METADATA
];
3725 int data_idx
= multilist_get_random_index(data_ml
);
3726 int meta_idx
= multilist_get_random_index(meta_ml
);
3727 multilist_sublist_t
*data_mls
;
3728 multilist_sublist_t
*meta_mls
;
3729 arc_buf_contents_t type
;
3730 arc_buf_hdr_t
*data_hdr
;
3731 arc_buf_hdr_t
*meta_hdr
;
3734 * We keep the sublist lock until we're finished, to prevent
3735 * the headers from being destroyed via arc_evict_state().
3737 data_mls
= multilist_sublist_lock(data_ml
, data_idx
);
3738 meta_mls
= multilist_sublist_lock(meta_ml
, meta_idx
);
3741 * These two loops are to ensure we skip any markers that
3742 * might be at the tail of the lists due to arc_evict_state().
3745 for (data_hdr
= multilist_sublist_tail(data_mls
); data_hdr
!= NULL
;
3746 data_hdr
= multilist_sublist_prev(data_mls
, data_hdr
)) {
3747 if (data_hdr
->b_spa
!= 0)
3751 for (meta_hdr
= multilist_sublist_tail(meta_mls
); meta_hdr
!= NULL
;
3752 meta_hdr
= multilist_sublist_prev(meta_mls
, meta_hdr
)) {
3753 if (meta_hdr
->b_spa
!= 0)
3757 if (data_hdr
== NULL
&& meta_hdr
== NULL
) {
3758 type
= ARC_BUFC_DATA
;
3759 } else if (data_hdr
== NULL
) {
3760 ASSERT3P(meta_hdr
, !=, NULL
);
3761 type
= ARC_BUFC_METADATA
;
3762 } else if (meta_hdr
== NULL
) {
3763 ASSERT3P(data_hdr
, !=, NULL
);
3764 type
= ARC_BUFC_DATA
;
3766 ASSERT3P(data_hdr
, !=, NULL
);
3767 ASSERT3P(meta_hdr
, !=, NULL
);
3769 /* The headers can't be on the sublist without an L1 header */
3770 ASSERT(HDR_HAS_L1HDR(data_hdr
));
3771 ASSERT(HDR_HAS_L1HDR(meta_hdr
));
3773 if (data_hdr
->b_l1hdr
.b_arc_access
<
3774 meta_hdr
->b_l1hdr
.b_arc_access
) {
3775 type
= ARC_BUFC_DATA
;
3777 type
= ARC_BUFC_METADATA
;
3781 multilist_sublist_unlock(meta_mls
);
3782 multilist_sublist_unlock(data_mls
);
3788 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3793 uint64_t total_evicted
= 0;
3798 * If we're over arc_meta_limit, we want to correct that before
3799 * potentially evicting data buffers below.
3801 total_evicted
+= arc_adjust_meta();
3806 * If we're over the target cache size, we want to evict enough
3807 * from the list to get back to our target size. We don't want
3808 * to evict too much from the MRU, such that it drops below
3809 * arc_p. So, if we're over our target cache size more than
3810 * the MRU is over arc_p, we'll evict enough to get back to
3811 * arc_p here, and then evict more from the MFU below.
3813 target
= MIN((int64_t)(arc_size
- arc_c
),
3814 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
3815 refcount_count(&arc_mru
->arcs_size
) + arc_meta_used
- arc_p
));
3818 * If we're below arc_meta_min, always prefer to evict data.
3819 * Otherwise, try to satisfy the requested number of bytes to
3820 * evict from the type which contains older buffers; in an
3821 * effort to keep newer buffers in the cache regardless of their
3822 * type. If we cannot satisfy the number of bytes from this
3823 * type, spill over into the next type.
3825 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
3826 arc_meta_used
> arc_meta_min
) {
3827 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3828 total_evicted
+= bytes
;
3831 * If we couldn't evict our target number of bytes from
3832 * metadata, we try to get the rest from data.
3837 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3839 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3840 total_evicted
+= bytes
;
3843 * If we couldn't evict our target number of bytes from
3844 * data, we try to get the rest from metadata.
3849 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3855 * Now that we've tried to evict enough from the MRU to get its
3856 * size back to arc_p, if we're still above the target cache
3857 * size, we evict the rest from the MFU.
3859 target
= arc_size
- arc_c
;
3861 if (arc_adjust_type(arc_mfu
) == ARC_BUFC_METADATA
&&
3862 arc_meta_used
> arc_meta_min
) {
3863 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3864 total_evicted
+= bytes
;
3867 * If we couldn't evict our target number of bytes from
3868 * metadata, we try to get the rest from data.
3873 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3875 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3876 total_evicted
+= bytes
;
3879 * If we couldn't evict our target number of bytes from
3880 * data, we try to get the rest from data.
3885 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3889 * Adjust ghost lists
3891 * In addition to the above, the ARC also defines target values
3892 * for the ghost lists. The sum of the mru list and mru ghost
3893 * list should never exceed the target size of the cache, and
3894 * the sum of the mru list, mfu list, mru ghost list, and mfu
3895 * ghost list should never exceed twice the target size of the
3896 * cache. The following logic enforces these limits on the ghost
3897 * caches, and evicts from them as needed.
3899 target
= refcount_count(&arc_mru
->arcs_size
) +
3900 refcount_count(&arc_mru_ghost
->arcs_size
) - arc_c
;
3902 bytes
= arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_DATA
);
3903 total_evicted
+= bytes
;
3908 arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_METADATA
);
3911 * We assume the sum of the mru list and mfu list is less than
3912 * or equal to arc_c (we enforced this above), which means we
3913 * can use the simpler of the two equations below:
3915 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3916 * mru ghost + mfu ghost <= arc_c
3918 target
= refcount_count(&arc_mru_ghost
->arcs_size
) +
3919 refcount_count(&arc_mfu_ghost
->arcs_size
) - arc_c
;
3921 bytes
= arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_DATA
);
3922 total_evicted
+= bytes
;
3927 arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_METADATA
);
3929 return (total_evicted
);
3933 arc_flush(spa_t
*spa
, boolean_t retry
)
3938 * If retry is B_TRUE, a spa must not be specified since we have
3939 * no good way to determine if all of a spa's buffers have been
3940 * evicted from an arc state.
3942 ASSERT(!retry
|| spa
== 0);
3945 guid
= spa_load_guid(spa
);
3947 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_DATA
, retry
);
3948 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_METADATA
, retry
);
3950 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_DATA
, retry
);
3951 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_METADATA
, retry
);
3953 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3954 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3956 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3957 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3961 arc_shrink(int64_t to_free
)
3965 if (c
> to_free
&& c
- to_free
> arc_c_min
) {
3966 arc_c
= c
- to_free
;
3967 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
3968 if (arc_c
> arc_size
)
3969 arc_c
= MAX(arc_size
, arc_c_min
);
3971 arc_p
= (arc_c
>> 1);
3972 ASSERT(arc_c
>= arc_c_min
);
3973 ASSERT((int64_t)arc_p
>= 0);
3978 if (arc_size
> arc_c
)
3979 (void) arc_adjust();
3983 * Return maximum amount of memory that we could possibly use. Reduced
3984 * to half of all memory in user space which is primarily used for testing.
3987 arc_all_memory(void)
3990 #ifdef CONFIG_HIGHMEM
3991 return (ptob(totalram_pages
- totalhigh_pages
));
3993 return (ptob(totalram_pages
));
3994 #endif /* CONFIG_HIGHMEM */
3996 return (ptob(physmem
) / 2);
3997 #endif /* _KERNEL */
4001 * Return the amount of memory that is considered free. In user space
4002 * which is primarily used for testing we pretend that free memory ranges
4003 * from 0-20% of all memory.
4006 arc_free_memory(void)
4009 #ifdef CONFIG_HIGHMEM
4012 return (ptob(si
.freeram
- si
.freehigh
));
4014 #ifdef ZFS_GLOBAL_NODE_PAGE_STATE
4015 return (ptob(nr_free_pages() +
4016 global_node_page_state(NR_INACTIVE_FILE
) +
4017 global_node_page_state(NR_INACTIVE_ANON
) +
4018 global_node_page_state(NR_SLAB_RECLAIMABLE
)));
4020 return (ptob(nr_free_pages() +
4021 global_page_state(NR_INACTIVE_FILE
) +
4022 global_page_state(NR_INACTIVE_ANON
) +
4023 global_page_state(NR_SLAB_RECLAIMABLE
)));
4024 #endif /* ZFS_GLOBAL_NODE_PAGE_STATE */
4025 #endif /* CONFIG_HIGHMEM */
4027 return (spa_get_random(arc_all_memory() * 20 / 100));
4028 #endif /* _KERNEL */
4031 typedef enum free_memory_reason_t
{
4036 FMR_PAGES_PP_MAXIMUM
,
4039 } free_memory_reason_t
;
4041 int64_t last_free_memory
;
4042 free_memory_reason_t last_free_reason
;
4046 * Additional reserve of pages for pp_reserve.
4048 int64_t arc_pages_pp_reserve
= 64;
4051 * Additional reserve of pages for swapfs.
4053 int64_t arc_swapfs_reserve
= 64;
4054 #endif /* _KERNEL */
4057 * Return the amount of memory that can be consumed before reclaim will be
4058 * needed. Positive if there is sufficient free memory, negative indicates
4059 * the amount of memory that needs to be freed up.
4062 arc_available_memory(void)
4064 int64_t lowest
= INT64_MAX
;
4065 free_memory_reason_t r
= FMR_UNKNOWN
;
4072 pgcnt_t needfree
= btop(arc_need_free
);
4073 pgcnt_t lotsfree
= btop(arc_sys_free
);
4074 pgcnt_t desfree
= 0;
4075 pgcnt_t freemem
= btop(arc_free_memory());
4079 n
= PAGESIZE
* (-needfree
);
4087 * check that we're out of range of the pageout scanner. It starts to
4088 * schedule paging if freemem is less than lotsfree and needfree.
4089 * lotsfree is the high-water mark for pageout, and needfree is the
4090 * number of needed free pages. We add extra pages here to make sure
4091 * the scanner doesn't start up while we're freeing memory.
4093 n
= PAGESIZE
* (freemem
- lotsfree
- needfree
- desfree
);
4101 * check to make sure that swapfs has enough space so that anon
4102 * reservations can still succeed. anon_resvmem() checks that the
4103 * availrmem is greater than swapfs_minfree, and the number of reserved
4104 * swap pages. We also add a bit of extra here just to prevent
4105 * circumstances from getting really dire.
4107 n
= PAGESIZE
* (availrmem
- swapfs_minfree
- swapfs_reserve
-
4108 desfree
- arc_swapfs_reserve
);
4111 r
= FMR_SWAPFS_MINFREE
;
4115 * Check that we have enough availrmem that memory locking (e.g., via
4116 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4117 * stores the number of pages that cannot be locked; when availrmem
4118 * drops below pages_pp_maximum, page locking mechanisms such as
4119 * page_pp_lock() will fail.)
4121 n
= PAGESIZE
* (availrmem
- pages_pp_maximum
-
4122 arc_pages_pp_reserve
);
4125 r
= FMR_PAGES_PP_MAXIMUM
;
4131 * If we're on a 32-bit platform, it's possible that we'll exhaust the
4132 * kernel heap space before we ever run out of available physical
4133 * memory. Most checks of the size of the heap_area compare against
4134 * tune.t_minarmem, which is the minimum available real memory that we
4135 * can have in the system. However, this is generally fixed at 25 pages
4136 * which is so low that it's useless. In this comparison, we seek to
4137 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4138 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4141 n
= vmem_size(heap_arena
, VMEM_FREE
) -
4142 (vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
) >> 2);
4150 * If zio data pages are being allocated out of a separate heap segment,
4151 * then enforce that the size of available vmem for this arena remains
4152 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4154 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4155 * memory (in the zio_arena) free, which can avoid memory
4156 * fragmentation issues.
4158 if (zio_arena
!= NULL
) {
4159 n
= (int64_t)vmem_size(zio_arena
, VMEM_FREE
) -
4160 (vmem_size(zio_arena
, VMEM_ALLOC
) >>
4161 arc_zio_arena_free_shift
);
4168 /* Every 100 calls, free a small amount */
4169 if (spa_get_random(100) == 0)
4171 #endif /* _KERNEL */
4173 last_free_memory
= lowest
;
4174 last_free_reason
= r
;
4180 * Determine if the system is under memory pressure and is asking
4181 * to reclaim memory. A return value of B_TRUE indicates that the system
4182 * is under memory pressure and that the arc should adjust accordingly.
4185 arc_reclaim_needed(void)
4187 return (arc_available_memory() < 0);
4191 arc_kmem_reap_now(void)
4194 kmem_cache_t
*prev_cache
= NULL
;
4195 kmem_cache_t
*prev_data_cache
= NULL
;
4196 extern kmem_cache_t
*zio_buf_cache
[];
4197 extern kmem_cache_t
*zio_data_buf_cache
[];
4198 extern kmem_cache_t
*range_seg_cache
;
4201 if ((arc_meta_used
>= arc_meta_limit
) && zfs_arc_meta_prune
) {
4203 * We are exceeding our meta-data cache limit.
4204 * Prune some entries to release holds on meta-data.
4206 arc_prune_async(zfs_arc_meta_prune
);
4210 * Reclaim unused memory from all kmem caches.
4216 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
4218 /* reach upper limit of cache size on 32-bit */
4219 if (zio_buf_cache
[i
] == NULL
)
4222 if (zio_buf_cache
[i
] != prev_cache
) {
4223 prev_cache
= zio_buf_cache
[i
];
4224 kmem_cache_reap_now(zio_buf_cache
[i
]);
4226 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
4227 prev_data_cache
= zio_data_buf_cache
[i
];
4228 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
4231 kmem_cache_reap_now(buf_cache
);
4232 kmem_cache_reap_now(hdr_full_cache
);
4233 kmem_cache_reap_now(hdr_l2only_cache
);
4234 kmem_cache_reap_now(range_seg_cache
);
4236 if (zio_arena
!= NULL
) {
4238 * Ask the vmem arena to reclaim unused memory from its
4241 vmem_qcache_reap(zio_arena
);
4246 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4247 * enough data and signal them to proceed. When this happens, the threads in
4248 * arc_get_data_impl() are sleeping while holding the hash lock for their
4249 * particular arc header. Thus, we must be careful to never sleep on a
4250 * hash lock in this thread. This is to prevent the following deadlock:
4252 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4253 * waiting for the reclaim thread to signal it.
4255 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4256 * fails, and goes to sleep forever.
4258 * This possible deadlock is avoided by always acquiring a hash lock
4259 * using mutex_tryenter() from arc_reclaim_thread().
4262 arc_reclaim_thread(void)
4264 fstrans_cookie_t cookie
= spl_fstrans_mark();
4265 hrtime_t growtime
= 0;
4268 CALLB_CPR_INIT(&cpr
, &arc_reclaim_lock
, callb_generic_cpr
, FTAG
);
4270 mutex_enter(&arc_reclaim_lock
);
4271 while (!arc_reclaim_thread_exit
) {
4273 uint64_t evicted
= 0;
4274 uint64_t need_free
= arc_need_free
;
4275 arc_tuning_update();
4278 * This is necessary in order for the mdb ::arc dcmd to
4279 * show up to date information. Since the ::arc command
4280 * does not call the kstat's update function, without
4281 * this call, the command may show stale stats for the
4282 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4283 * with this change, the data might be up to 1 second
4284 * out of date; but that should suffice. The arc_state_t
4285 * structures can be queried directly if more accurate
4286 * information is needed.
4289 if (arc_ksp
!= NULL
)
4290 arc_ksp
->ks_update(arc_ksp
, KSTAT_READ
);
4292 mutex_exit(&arc_reclaim_lock
);
4295 * We call arc_adjust() before (possibly) calling
4296 * arc_kmem_reap_now(), so that we can wake up
4297 * arc_get_data_buf() sooner.
4299 evicted
= arc_adjust();
4301 int64_t free_memory
= arc_available_memory();
4302 if (free_memory
< 0) {
4304 arc_no_grow
= B_TRUE
;
4308 * Wait at least zfs_grow_retry (default 5) seconds
4309 * before considering growing.
4311 growtime
= gethrtime() + SEC2NSEC(arc_grow_retry
);
4313 arc_kmem_reap_now();
4316 * If we are still low on memory, shrink the ARC
4317 * so that we have arc_shrink_min free space.
4319 free_memory
= arc_available_memory();
4321 to_free
= (arc_c
>> arc_shrink_shift
) - free_memory
;
4324 to_free
= MAX(to_free
, need_free
);
4326 arc_shrink(to_free
);
4328 } else if (free_memory
< arc_c
>> arc_no_grow_shift
) {
4329 arc_no_grow
= B_TRUE
;
4330 } else if (gethrtime() >= growtime
) {
4331 arc_no_grow
= B_FALSE
;
4334 mutex_enter(&arc_reclaim_lock
);
4337 * If evicted is zero, we couldn't evict anything via
4338 * arc_adjust(). This could be due to hash lock
4339 * collisions, but more likely due to the majority of
4340 * arc buffers being unevictable. Therefore, even if
4341 * arc_size is above arc_c, another pass is unlikely to
4342 * be helpful and could potentially cause us to enter an
4345 if (arc_size
<= arc_c
|| evicted
== 0) {
4347 * We're either no longer overflowing, or we
4348 * can't evict anything more, so we should wake
4349 * up any threads before we go to sleep and remove
4350 * the bytes we were working on from arc_need_free
4351 * since nothing more will be done here.
4353 cv_broadcast(&arc_reclaim_waiters_cv
);
4354 ARCSTAT_INCR(arcstat_need_free
, -need_free
);
4357 * Block until signaled, or after one second (we
4358 * might need to perform arc_kmem_reap_now()
4359 * even if we aren't being signalled)
4361 CALLB_CPR_SAFE_BEGIN(&cpr
);
4362 (void) cv_timedwait_sig_hires(&arc_reclaim_thread_cv
,
4363 &arc_reclaim_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
4364 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_lock
);
4368 arc_reclaim_thread_exit
= B_FALSE
;
4369 cv_broadcast(&arc_reclaim_thread_cv
);
4370 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_lock */
4371 spl_fstrans_unmark(cookie
);
4377 * Determine the amount of memory eligible for eviction contained in the
4378 * ARC. All clean data reported by the ghost lists can always be safely
4379 * evicted. Due to arc_c_min, the same does not hold for all clean data
4380 * contained by the regular mru and mfu lists.
4382 * In the case of the regular mru and mfu lists, we need to report as
4383 * much clean data as possible, such that evicting that same reported
4384 * data will not bring arc_size below arc_c_min. Thus, in certain
4385 * circumstances, the total amount of clean data in the mru and mfu
4386 * lists might not actually be evictable.
4388 * The following two distinct cases are accounted for:
4390 * 1. The sum of the amount of dirty data contained by both the mru and
4391 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4392 * is greater than or equal to arc_c_min.
4393 * (i.e. amount of dirty data >= arc_c_min)
4395 * This is the easy case; all clean data contained by the mru and mfu
4396 * lists is evictable. Evicting all clean data can only drop arc_size
4397 * to the amount of dirty data, which is greater than arc_c_min.
4399 * 2. The sum of the amount of dirty data contained by both the mru and
4400 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4401 * is less than arc_c_min.
4402 * (i.e. arc_c_min > amount of dirty data)
4404 * 2.1. arc_size is greater than or equal arc_c_min.
4405 * (i.e. arc_size >= arc_c_min > amount of dirty data)
4407 * In this case, not all clean data from the regular mru and mfu
4408 * lists is actually evictable; we must leave enough clean data
4409 * to keep arc_size above arc_c_min. Thus, the maximum amount of
4410 * evictable data from the two lists combined, is exactly the
4411 * difference between arc_size and arc_c_min.
4413 * 2.2. arc_size is less than arc_c_min
4414 * (i.e. arc_c_min > arc_size > amount of dirty data)
4416 * In this case, none of the data contained in the mru and mfu
4417 * lists is evictable, even if it's clean. Since arc_size is
4418 * already below arc_c_min, evicting any more would only
4419 * increase this negative difference.
4422 arc_evictable_memory(void)
4424 uint64_t arc_clean
=
4425 refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]) +
4426 refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]) +
4427 refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]) +
4428 refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
4429 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
4432 * Scale reported evictable memory in proportion to page cache, cap
4433 * at specified min/max.
4435 #ifdef ZFS_GLOBAL_NODE_PAGE_STATE
4436 uint64_t min
= (ptob(global_node_page_state(NR_FILE_PAGES
)) / 100) *
4439 uint64_t min
= (ptob(global_page_state(NR_FILE_PAGES
)) / 100) *
4442 min
= MAX(arc_c_min
, MIN(arc_c_max
, min
));
4444 if (arc_dirty
>= min
)
4447 return (MAX((int64_t)arc_size
- (int64_t)min
, 0));
4451 * If sc->nr_to_scan is zero, the caller is requesting a query of the
4452 * number of objects which can potentially be freed. If it is nonzero,
4453 * the request is to free that many objects.
4455 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
4456 * in struct shrinker and also require the shrinker to return the number
4459 * Older kernels require the shrinker to return the number of freeable
4460 * objects following the freeing of nr_to_free.
4462 static spl_shrinker_t
4463 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
4467 /* The arc is considered warm once reclaim has occurred */
4468 if (unlikely(arc_warm
== B_FALSE
))
4471 /* Return the potential number of reclaimable pages */
4472 pages
= btop((int64_t)arc_evictable_memory());
4473 if (sc
->nr_to_scan
== 0)
4476 /* Not allowed to perform filesystem reclaim */
4477 if (!(sc
->gfp_mask
& __GFP_FS
))
4478 return (SHRINK_STOP
);
4480 /* Reclaim in progress */
4481 if (mutex_tryenter(&arc_reclaim_lock
) == 0) {
4482 ARCSTAT_INCR(arcstat_need_free
, ptob(sc
->nr_to_scan
));
4486 mutex_exit(&arc_reclaim_lock
);
4489 * Evict the requested number of pages by shrinking arc_c the
4493 arc_shrink(ptob(sc
->nr_to_scan
));
4494 if (current_is_kswapd())
4495 arc_kmem_reap_now();
4496 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
4497 pages
= MAX((int64_t)pages
-
4498 (int64_t)btop(arc_evictable_memory()), 0);
4500 pages
= btop(arc_evictable_memory());
4503 * We've shrunk what we can, wake up threads.
4505 cv_broadcast(&arc_reclaim_waiters_cv
);
4507 pages
= SHRINK_STOP
;
4510 * When direct reclaim is observed it usually indicates a rapid
4511 * increase in memory pressure. This occurs because the kswapd
4512 * threads were unable to asynchronously keep enough free memory
4513 * available. In this case set arc_no_grow to briefly pause arc
4514 * growth to avoid compounding the memory pressure.
4516 if (current_is_kswapd()) {
4517 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
4519 arc_no_grow
= B_TRUE
;
4520 arc_kmem_reap_now();
4521 ARCSTAT_BUMP(arcstat_memory_direct_count
);
4526 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
4528 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
4529 #endif /* _KERNEL */
4532 * Adapt arc info given the number of bytes we are trying to add and
4533 * the state that we are coming from. This function is only called
4534 * when we are adding new content to the cache.
4537 arc_adapt(int bytes
, arc_state_t
*state
)
4540 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
4541 int64_t mrug_size
= refcount_count(&arc_mru_ghost
->arcs_size
);
4542 int64_t mfug_size
= refcount_count(&arc_mfu_ghost
->arcs_size
);
4544 if (state
== arc_l2c_only
)
4549 * Adapt the target size of the MRU list:
4550 * - if we just hit in the MRU ghost list, then increase
4551 * the target size of the MRU list.
4552 * - if we just hit in the MFU ghost list, then increase
4553 * the target size of the MFU list by decreasing the
4554 * target size of the MRU list.
4556 if (state
== arc_mru_ghost
) {
4557 mult
= (mrug_size
>= mfug_size
) ? 1 : (mfug_size
/ mrug_size
);
4558 if (!zfs_arc_p_dampener_disable
)
4559 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
4561 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
4562 } else if (state
== arc_mfu_ghost
) {
4565 mult
= (mfug_size
>= mrug_size
) ? 1 : (mrug_size
/ mfug_size
);
4566 if (!zfs_arc_p_dampener_disable
)
4567 mult
= MIN(mult
, 10);
4569 delta
= MIN(bytes
* mult
, arc_p
);
4570 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
4572 ASSERT((int64_t)arc_p
>= 0);
4574 if (arc_reclaim_needed()) {
4575 cv_signal(&arc_reclaim_thread_cv
);
4582 if (arc_c
>= arc_c_max
)
4586 * If we're within (2 * maxblocksize) bytes of the target
4587 * cache size, increment the target cache size
4589 ASSERT3U(arc_c
, >=, 2ULL << SPA_MAXBLOCKSHIFT
);
4590 if (arc_size
>= arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
4591 atomic_add_64(&arc_c
, (int64_t)bytes
);
4592 if (arc_c
> arc_c_max
)
4594 else if (state
== arc_anon
)
4595 atomic_add_64(&arc_p
, (int64_t)bytes
);
4599 ASSERT((int64_t)arc_p
>= 0);
4603 * Check if arc_size has grown past our upper threshold, determined by
4604 * zfs_arc_overflow_shift.
4607 arc_is_overflowing(void)
4609 /* Always allow at least one block of overflow */
4610 uint64_t overflow
= MAX(SPA_MAXBLOCKSIZE
,
4611 arc_c
>> zfs_arc_overflow_shift
);
4613 return (arc_size
>= arc_c
+ overflow
);
4617 arc_get_data_abd(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4619 arc_buf_contents_t type
= arc_buf_type(hdr
);
4621 arc_get_data_impl(hdr
, size
, tag
);
4622 if (type
== ARC_BUFC_METADATA
) {
4623 return (abd_alloc(size
, B_TRUE
));
4625 ASSERT(type
== ARC_BUFC_DATA
);
4626 return (abd_alloc(size
, B_FALSE
));
4631 arc_get_data_buf(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4633 arc_buf_contents_t type
= arc_buf_type(hdr
);
4635 arc_get_data_impl(hdr
, size
, tag
);
4636 if (type
== ARC_BUFC_METADATA
) {
4637 return (zio_buf_alloc(size
));
4639 ASSERT(type
== ARC_BUFC_DATA
);
4640 return (zio_data_buf_alloc(size
));
4645 * Allocate a block and return it to the caller. If we are hitting the
4646 * hard limit for the cache size, we must sleep, waiting for the eviction
4647 * thread to catch up. If we're past the target size but below the hard
4648 * limit, we'll only signal the reclaim thread and continue on.
4651 arc_get_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4653 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
4654 arc_buf_contents_t type
= arc_buf_type(hdr
);
4656 arc_adapt(size
, state
);
4659 * If arc_size is currently overflowing, and has grown past our
4660 * upper limit, we must be adding data faster than the evict
4661 * thread can evict. Thus, to ensure we don't compound the
4662 * problem by adding more data and forcing arc_size to grow even
4663 * further past it's target size, we halt and wait for the
4664 * eviction thread to catch up.
4666 * It's also possible that the reclaim thread is unable to evict
4667 * enough buffers to get arc_size below the overflow limit (e.g.
4668 * due to buffers being un-evictable, or hash lock collisions).
4669 * In this case, we want to proceed regardless if we're
4670 * overflowing; thus we don't use a while loop here.
4672 if (arc_is_overflowing()) {
4673 mutex_enter(&arc_reclaim_lock
);
4676 * Now that we've acquired the lock, we may no longer be
4677 * over the overflow limit, lets check.
4679 * We're ignoring the case of spurious wake ups. If that
4680 * were to happen, it'd let this thread consume an ARC
4681 * buffer before it should have (i.e. before we're under
4682 * the overflow limit and were signalled by the reclaim
4683 * thread). As long as that is a rare occurrence, it
4684 * shouldn't cause any harm.
4686 if (arc_is_overflowing()) {
4687 cv_signal(&arc_reclaim_thread_cv
);
4688 cv_wait(&arc_reclaim_waiters_cv
, &arc_reclaim_lock
);
4691 mutex_exit(&arc_reclaim_lock
);
4694 VERIFY3U(hdr
->b_type
, ==, type
);
4695 if (type
== ARC_BUFC_METADATA
) {
4696 arc_space_consume(size
, ARC_SPACE_META
);
4698 arc_space_consume(size
, ARC_SPACE_DATA
);
4702 * Update the state size. Note that ghost states have a
4703 * "ghost size" and so don't need to be updated.
4705 if (!GHOST_STATE(state
)) {
4707 (void) refcount_add_many(&state
->arcs_size
, size
, tag
);
4710 * If this is reached via arc_read, the link is
4711 * protected by the hash lock. If reached via
4712 * arc_buf_alloc, the header should not be accessed by
4713 * any other thread. And, if reached via arc_read_done,
4714 * the hash lock will protect it if it's found in the
4715 * hash table; otherwise no other thread should be
4716 * trying to [add|remove]_reference it.
4718 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
4719 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4720 (void) refcount_add_many(&state
->arcs_esize
[type
],
4725 * If we are growing the cache, and we are adding anonymous
4726 * data, and we have outgrown arc_p, update arc_p
4728 if (arc_size
< arc_c
&& hdr
->b_l1hdr
.b_state
== arc_anon
&&
4729 (refcount_count(&arc_anon
->arcs_size
) +
4730 refcount_count(&arc_mru
->arcs_size
) > arc_p
))
4731 arc_p
= MIN(arc_c
, arc_p
+ size
);
4736 arc_free_data_abd(arc_buf_hdr_t
*hdr
, abd_t
*abd
, uint64_t size
, void *tag
)
4738 arc_free_data_impl(hdr
, size
, tag
);
4743 arc_free_data_buf(arc_buf_hdr_t
*hdr
, void *buf
, uint64_t size
, void *tag
)
4745 arc_buf_contents_t type
= arc_buf_type(hdr
);
4747 arc_free_data_impl(hdr
, size
, tag
);
4748 if (type
== ARC_BUFC_METADATA
) {
4749 zio_buf_free(buf
, size
);
4751 ASSERT(type
== ARC_BUFC_DATA
);
4752 zio_data_buf_free(buf
, size
);
4757 * Free the arc data buffer.
4760 arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4762 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
4763 arc_buf_contents_t type
= arc_buf_type(hdr
);
4765 /* protected by hash lock, if in the hash table */
4766 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
4767 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4768 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
4770 (void) refcount_remove_many(&state
->arcs_esize
[type
],
4773 (void) refcount_remove_many(&state
->arcs_size
, size
, tag
);
4775 VERIFY3U(hdr
->b_type
, ==, type
);
4776 if (type
== ARC_BUFC_METADATA
) {
4777 arc_space_return(size
, ARC_SPACE_META
);
4779 ASSERT(type
== ARC_BUFC_DATA
);
4780 arc_space_return(size
, ARC_SPACE_DATA
);
4785 * This routine is called whenever a buffer is accessed.
4786 * NOTE: the hash lock is dropped in this function.
4789 arc_access(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
4793 ASSERT(MUTEX_HELD(hash_lock
));
4794 ASSERT(HDR_HAS_L1HDR(hdr
));
4796 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
4798 * This buffer is not in the cache, and does not
4799 * appear in our "ghost" list. Add the new buffer
4803 ASSERT0(hdr
->b_l1hdr
.b_arc_access
);
4804 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4805 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
4806 arc_change_state(arc_mru
, hdr
, hash_lock
);
4808 } else if (hdr
->b_l1hdr
.b_state
== arc_mru
) {
4809 now
= ddi_get_lbolt();
4812 * If this buffer is here because of a prefetch, then either:
4813 * - clear the flag if this is a "referencing" read
4814 * (any subsequent access will bump this into the MFU state).
4816 * - move the buffer to the head of the list if this is
4817 * another prefetch (to make it less likely to be evicted).
4819 if (HDR_PREFETCH(hdr
)) {
4820 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
4821 /* link protected by hash lock */
4822 ASSERT(multilist_link_active(
4823 &hdr
->b_l1hdr
.b_arc_node
));
4825 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
4826 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
4827 ARCSTAT_BUMP(arcstat_mru_hits
);
4829 hdr
->b_l1hdr
.b_arc_access
= now
;
4834 * This buffer has been "accessed" only once so far,
4835 * but it is still in the cache. Move it to the MFU
4838 if (ddi_time_after(now
, hdr
->b_l1hdr
.b_arc_access
+
4841 * More than 125ms have passed since we
4842 * instantiated this buffer. Move it to the
4843 * most frequently used state.
4845 hdr
->b_l1hdr
.b_arc_access
= now
;
4846 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4847 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4849 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
4850 ARCSTAT_BUMP(arcstat_mru_hits
);
4851 } else if (hdr
->b_l1hdr
.b_state
== arc_mru_ghost
) {
4852 arc_state_t
*new_state
;
4854 * This buffer has been "accessed" recently, but
4855 * was evicted from the cache. Move it to the
4859 if (HDR_PREFETCH(hdr
)) {
4860 new_state
= arc_mru
;
4861 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0)
4862 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
4863 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
4865 new_state
= arc_mfu
;
4866 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4869 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4870 arc_change_state(new_state
, hdr
, hash_lock
);
4872 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_ghost_hits
);
4873 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
4874 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu
) {
4876 * This buffer has been accessed more than once and is
4877 * still in the cache. Keep it in the MFU state.
4879 * NOTE: an add_reference() that occurred when we did
4880 * the arc_read() will have kicked this off the list.
4881 * If it was a prefetch, we will explicitly move it to
4882 * the head of the list now.
4884 if ((HDR_PREFETCH(hdr
)) != 0) {
4885 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4886 /* link protected by hash_lock */
4887 ASSERT(multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4889 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_hits
);
4890 ARCSTAT_BUMP(arcstat_mfu_hits
);
4891 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4892 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu_ghost
) {
4893 arc_state_t
*new_state
= arc_mfu
;
4895 * This buffer has been accessed more than once but has
4896 * been evicted from the cache. Move it back to the
4900 if (HDR_PREFETCH(hdr
)) {
4902 * This is a prefetch access...
4903 * move this block back to the MRU state.
4905 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
4906 new_state
= arc_mru
;
4909 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4910 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4911 arc_change_state(new_state
, hdr
, hash_lock
);
4913 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_ghost_hits
);
4914 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
4915 } else if (hdr
->b_l1hdr
.b_state
== arc_l2c_only
) {
4917 * This buffer is on the 2nd Level ARC.
4920 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4921 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4922 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4924 cmn_err(CE_PANIC
, "invalid arc state 0x%p",
4925 hdr
->b_l1hdr
.b_state
);
4929 /* a generic arc_done_func_t which you can use */
4932 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4934 if (zio
== NULL
|| zio
->io_error
== 0)
4935 bcopy(buf
->b_data
, arg
, arc_buf_size(buf
));
4936 arc_buf_destroy(buf
, arg
);
4939 /* a generic arc_done_func_t */
4941 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4943 arc_buf_t
**bufp
= arg
;
4944 if (zio
&& zio
->io_error
) {
4945 arc_buf_destroy(buf
, arg
);
4949 ASSERT(buf
->b_data
);
4954 arc_hdr_verify(arc_buf_hdr_t
*hdr
, blkptr_t
*bp
)
4956 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
)) {
4957 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, 0);
4958 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
4960 if (HDR_COMPRESSION_ENABLED(hdr
)) {
4961 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==,
4962 BP_GET_COMPRESS(bp
));
4964 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(bp
));
4965 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, BP_GET_PSIZE(bp
));
4970 arc_read_done(zio_t
*zio
)
4972 arc_buf_hdr_t
*hdr
= zio
->io_private
;
4973 kmutex_t
*hash_lock
= NULL
;
4974 arc_callback_t
*callback_list
;
4975 arc_callback_t
*acb
;
4976 boolean_t freeable
= B_FALSE
;
4977 boolean_t no_zio_error
= (zio
->io_error
== 0);
4980 * The hdr was inserted into hash-table and removed from lists
4981 * prior to starting I/O. We should find this header, since
4982 * it's in the hash table, and it should be legit since it's
4983 * not possible to evict it during the I/O. The only possible
4984 * reason for it not to be found is if we were freed during the
4987 if (HDR_IN_HASH_TABLE(hdr
)) {
4988 arc_buf_hdr_t
*found
;
4990 ASSERT3U(hdr
->b_birth
, ==, BP_PHYSICAL_BIRTH(zio
->io_bp
));
4991 ASSERT3U(hdr
->b_dva
.dva_word
[0], ==,
4992 BP_IDENTITY(zio
->io_bp
)->dva_word
[0]);
4993 ASSERT3U(hdr
->b_dva
.dva_word
[1], ==,
4994 BP_IDENTITY(zio
->io_bp
)->dva_word
[1]);
4996 found
= buf_hash_find(hdr
->b_spa
, zio
->io_bp
, &hash_lock
);
4998 ASSERT((found
== hdr
&&
4999 DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
5000 (found
== hdr
&& HDR_L2_READING(hdr
)));
5001 ASSERT3P(hash_lock
, !=, NULL
);
5005 /* byteswap if necessary */
5006 if (BP_SHOULD_BYTESWAP(zio
->io_bp
)) {
5007 if (BP_GET_LEVEL(zio
->io_bp
) > 0) {
5008 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_UINT64
;
5010 hdr
->b_l1hdr
.b_byteswap
=
5011 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
5014 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
5018 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_EVICTED
);
5019 if (l2arc_noprefetch
&& HDR_PREFETCH(hdr
))
5020 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2CACHE
);
5022 callback_list
= hdr
->b_l1hdr
.b_acb
;
5023 ASSERT3P(callback_list
, !=, NULL
);
5025 if (hash_lock
&& no_zio_error
&& hdr
->b_l1hdr
.b_state
== arc_anon
) {
5027 * Only call arc_access on anonymous buffers. This is because
5028 * if we've issued an I/O for an evicted buffer, we've already
5029 * called arc_access (to prevent any simultaneous readers from
5030 * getting confused).
5032 arc_access(hdr
, hash_lock
);
5036 * If a read request has a callback (i.e. acb_done is not NULL), then we
5037 * make a buf containing the data according to the parameters which were
5038 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5039 * aren't needlessly decompressing the data multiple times.
5041 int callback_cnt
= 0;
5042 for (acb
= callback_list
; acb
!= NULL
; acb
= acb
->acb_next
) {
5046 /* This is a demand read since prefetches don't use callbacks */
5049 int error
= arc_buf_alloc_impl(hdr
, acb
->acb_private
,
5050 acb
->acb_compressed
, no_zio_error
, &acb
->acb_buf
);
5052 zio
->io_error
= error
;
5055 hdr
->b_l1hdr
.b_acb
= NULL
;
5056 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5057 if (callback_cnt
== 0) {
5058 ASSERT(HDR_PREFETCH(hdr
));
5059 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
5060 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5063 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
) ||
5064 callback_list
!= NULL
);
5067 arc_hdr_verify(hdr
, zio
->io_bp
);
5069 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_ERROR
);
5070 if (hdr
->b_l1hdr
.b_state
!= arc_anon
)
5071 arc_change_state(arc_anon
, hdr
, hash_lock
);
5072 if (HDR_IN_HASH_TABLE(hdr
))
5073 buf_hash_remove(hdr
);
5074 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
5078 * Broadcast before we drop the hash_lock to avoid the possibility
5079 * that the hdr (and hence the cv) might be freed before we get to
5080 * the cv_broadcast().
5082 cv_broadcast(&hdr
->b_l1hdr
.b_cv
);
5084 if (hash_lock
!= NULL
) {
5085 mutex_exit(hash_lock
);
5088 * This block was freed while we waited for the read to
5089 * complete. It has been removed from the hash table and
5090 * moved to the anonymous state (so that it won't show up
5093 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
5094 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
5097 /* execute each callback and free its structure */
5098 while ((acb
= callback_list
) != NULL
) {
5100 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
5102 if (acb
->acb_zio_dummy
!= NULL
) {
5103 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
5104 zio_nowait(acb
->acb_zio_dummy
);
5107 callback_list
= acb
->acb_next
;
5108 kmem_free(acb
, sizeof (arc_callback_t
));
5112 arc_hdr_destroy(hdr
);
5116 * "Read" the block at the specified DVA (in bp) via the
5117 * cache. If the block is found in the cache, invoke the provided
5118 * callback immediately and return. Note that the `zio' parameter
5119 * in the callback will be NULL in this case, since no IO was
5120 * required. If the block is not in the cache pass the read request
5121 * on to the spa with a substitute callback function, so that the
5122 * requested block will be added to the cache.
5124 * If a read request arrives for a block that has a read in-progress,
5125 * either wait for the in-progress read to complete (and return the
5126 * results); or, if this is a read with a "done" func, add a record
5127 * to the read to invoke the "done" func when the read completes,
5128 * and return; or just return.
5130 * arc_read_done() will invoke all the requested "done" functions
5131 * for readers of this block.
5134 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_done_func_t
*done
,
5135 void *private, zio_priority_t priority
, int zio_flags
,
5136 arc_flags_t
*arc_flags
, const zbookmark_phys_t
*zb
)
5138 arc_buf_hdr_t
*hdr
= NULL
;
5139 kmutex_t
*hash_lock
= NULL
;
5141 uint64_t guid
= spa_load_guid(spa
);
5142 boolean_t compressed_read
= (zio_flags
& ZIO_FLAG_RAW
) != 0;
5145 ASSERT(!BP_IS_EMBEDDED(bp
) ||
5146 BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
);
5149 if (!BP_IS_EMBEDDED(bp
)) {
5151 * Embedded BP's have no DVA and require no I/O to "read".
5152 * Create an anonymous arc buf to back it.
5154 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
5157 if (hdr
!= NULL
&& HDR_HAS_L1HDR(hdr
) && hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5158 arc_buf_t
*buf
= NULL
;
5159 *arc_flags
|= ARC_FLAG_CACHED
;
5161 if (HDR_IO_IN_PROGRESS(hdr
)) {
5163 if ((hdr
->b_flags
& ARC_FLAG_PRIO_ASYNC_READ
) &&
5164 priority
== ZIO_PRIORITY_SYNC_READ
) {
5166 * This sync read must wait for an
5167 * in-progress async read (e.g. a predictive
5168 * prefetch). Async reads are queued
5169 * separately at the vdev_queue layer, so
5170 * this is a form of priority inversion.
5171 * Ideally, we would "inherit" the demand
5172 * i/o's priority by moving the i/o from
5173 * the async queue to the synchronous queue,
5174 * but there is currently no mechanism to do
5175 * so. Track this so that we can evaluate
5176 * the magnitude of this potential performance
5179 * Note that if the prefetch i/o is already
5180 * active (has been issued to the device),
5181 * the prefetch improved performance, because
5182 * we issued it sooner than we would have
5183 * without the prefetch.
5185 DTRACE_PROBE1(arc__sync__wait__for__async
,
5186 arc_buf_hdr_t
*, hdr
);
5187 ARCSTAT_BUMP(arcstat_sync_wait_for_async
);
5189 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5190 arc_hdr_clear_flags(hdr
,
5191 ARC_FLAG_PREDICTIVE_PREFETCH
);
5194 if (*arc_flags
& ARC_FLAG_WAIT
) {
5195 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
5196 mutex_exit(hash_lock
);
5199 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5202 arc_callback_t
*acb
= NULL
;
5204 acb
= kmem_zalloc(sizeof (arc_callback_t
),
5206 acb
->acb_done
= done
;
5207 acb
->acb_private
= private;
5208 acb
->acb_compressed
= compressed_read
;
5210 acb
->acb_zio_dummy
= zio_null(pio
,
5211 spa
, NULL
, NULL
, NULL
, zio_flags
);
5213 ASSERT3P(acb
->acb_done
, !=, NULL
);
5214 acb
->acb_next
= hdr
->b_l1hdr
.b_acb
;
5215 hdr
->b_l1hdr
.b_acb
= acb
;
5216 mutex_exit(hash_lock
);
5219 mutex_exit(hash_lock
);
5223 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
5224 hdr
->b_l1hdr
.b_state
== arc_mfu
);
5227 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5229 * This is a demand read which does not have to
5230 * wait for i/o because we did a predictive
5231 * prefetch i/o for it, which has completed.
5234 arc__demand__hit__predictive__prefetch
,
5235 arc_buf_hdr_t
*, hdr
);
5237 arcstat_demand_hit_predictive_prefetch
);
5238 arc_hdr_clear_flags(hdr
,
5239 ARC_FLAG_PREDICTIVE_PREFETCH
);
5241 ASSERT(!BP_IS_EMBEDDED(bp
) || !BP_IS_HOLE(bp
));
5243 /* Get a buf with the desired data in it. */
5244 VERIFY0(arc_buf_alloc_impl(hdr
, private,
5245 compressed_read
, B_TRUE
, &buf
));
5246 } else if (*arc_flags
& ARC_FLAG_PREFETCH
&&
5247 refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
5248 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
5250 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
5251 arc_access(hdr
, hash_lock
);
5252 if (*arc_flags
& ARC_FLAG_L2CACHE
)
5253 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5254 mutex_exit(hash_lock
);
5255 ARCSTAT_BUMP(arcstat_hits
);
5256 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
5257 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
5258 data
, metadata
, hits
);
5261 done(NULL
, buf
, private);
5263 uint64_t lsize
= BP_GET_LSIZE(bp
);
5264 uint64_t psize
= BP_GET_PSIZE(bp
);
5265 arc_callback_t
*acb
;
5268 boolean_t devw
= B_FALSE
;
5272 * Gracefully handle a damaged logical block size as a
5275 if (lsize
> spa_maxblocksize(spa
)) {
5276 rc
= SET_ERROR(ECKSUM
);
5281 /* this block is not in the cache */
5282 arc_buf_hdr_t
*exists
= NULL
;
5283 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
5284 hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
5285 BP_GET_COMPRESS(bp
), type
);
5287 if (!BP_IS_EMBEDDED(bp
)) {
5288 hdr
->b_dva
= *BP_IDENTITY(bp
);
5289 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
5290 exists
= buf_hash_insert(hdr
, &hash_lock
);
5292 if (exists
!= NULL
) {
5293 /* somebody beat us to the hash insert */
5294 mutex_exit(hash_lock
);
5295 buf_discard_identity(hdr
);
5296 arc_hdr_destroy(hdr
);
5297 goto top
; /* restart the IO request */
5301 * This block is in the ghost cache. If it was L2-only
5302 * (and thus didn't have an L1 hdr), we realloc the
5303 * header to add an L1 hdr.
5305 if (!HDR_HAS_L1HDR(hdr
)) {
5306 hdr
= arc_hdr_realloc(hdr
, hdr_l2only_cache
,
5310 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5311 ASSERT(GHOST_STATE(hdr
->b_l1hdr
.b_state
));
5312 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5313 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5314 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
5315 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
5318 * This is a delicate dance that we play here.
5319 * This hdr is in the ghost list so we access it
5320 * to move it out of the ghost list before we
5321 * initiate the read. If it's a prefetch then
5322 * it won't have a callback so we'll remove the
5323 * reference that arc_buf_alloc_impl() created. We
5324 * do this after we've called arc_access() to
5325 * avoid hitting an assert in remove_reference().
5327 arc_access(hdr
, hash_lock
);
5328 arc_hdr_alloc_pabd(hdr
);
5330 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5331 size
= arc_hdr_size(hdr
);
5334 * If compression is enabled on the hdr, then will do
5335 * RAW I/O and will store the compressed data in the hdr's
5336 * data block. Otherwise, the hdr's data block will contain
5337 * the uncompressed data.
5339 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
) {
5340 zio_flags
|= ZIO_FLAG_RAW
;
5343 if (*arc_flags
& ARC_FLAG_PREFETCH
)
5344 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
5345 if (*arc_flags
& ARC_FLAG_L2CACHE
)
5346 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5347 if (BP_GET_LEVEL(bp
) > 0)
5348 arc_hdr_set_flags(hdr
, ARC_FLAG_INDIRECT
);
5349 if (*arc_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
)
5350 arc_hdr_set_flags(hdr
, ARC_FLAG_PREDICTIVE_PREFETCH
);
5351 ASSERT(!GHOST_STATE(hdr
->b_l1hdr
.b_state
));
5353 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
5354 acb
->acb_done
= done
;
5355 acb
->acb_private
= private;
5356 acb
->acb_compressed
= compressed_read
;
5358 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5359 hdr
->b_l1hdr
.b_acb
= acb
;
5360 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5362 if (HDR_HAS_L2HDR(hdr
) &&
5363 (vd
= hdr
->b_l2hdr
.b_dev
->l2ad_vdev
) != NULL
) {
5364 devw
= hdr
->b_l2hdr
.b_dev
->l2ad_writing
;
5365 addr
= hdr
->b_l2hdr
.b_daddr
;
5367 * Lock out device removal.
5369 if (vdev_is_dead(vd
) ||
5370 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
5374 if (priority
== ZIO_PRIORITY_ASYNC_READ
)
5375 arc_hdr_set_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
5377 arc_hdr_clear_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
5379 if (hash_lock
!= NULL
)
5380 mutex_exit(hash_lock
);
5383 * At this point, we have a level 1 cache miss. Try again in
5384 * L2ARC if possible.
5386 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, lsize
);
5388 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
5389 uint64_t, lsize
, zbookmark_phys_t
*, zb
);
5390 ARCSTAT_BUMP(arcstat_misses
);
5391 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
5392 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
5393 data
, metadata
, misses
);
5395 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
5397 * Read from the L2ARC if the following are true:
5398 * 1. The L2ARC vdev was previously cached.
5399 * 2. This buffer still has L2ARC metadata.
5400 * 3. This buffer isn't currently writing to the L2ARC.
5401 * 4. The L2ARC entry wasn't evicted, which may
5402 * also have invalidated the vdev.
5403 * 5. This isn't prefetch and l2arc_noprefetch is set.
5405 if (HDR_HAS_L2HDR(hdr
) &&
5406 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
5407 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
5408 l2arc_read_callback_t
*cb
;
5412 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
5413 ARCSTAT_BUMP(arcstat_l2_hits
);
5414 atomic_inc_32(&hdr
->b_l2hdr
.b_hits
);
5416 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
5418 cb
->l2rcb_hdr
= hdr
;
5421 cb
->l2rcb_flags
= zio_flags
;
5423 asize
= vdev_psize_to_asize(vd
, size
);
5424 if (asize
!= size
) {
5425 abd
= abd_alloc_for_io(asize
,
5426 HDR_ISTYPE_METADATA(hdr
));
5427 cb
->l2rcb_abd
= abd
;
5429 abd
= hdr
->b_l1hdr
.b_pabd
;
5432 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
5433 addr
+ asize
<= vd
->vdev_psize
-
5434 VDEV_LABEL_END_SIZE
);
5437 * l2arc read. The SCL_L2ARC lock will be
5438 * released by l2arc_read_done().
5439 * Issue a null zio if the underlying buffer
5440 * was squashed to zero size by compression.
5442 ASSERT3U(HDR_GET_COMPRESS(hdr
), !=,
5443 ZIO_COMPRESS_EMPTY
);
5444 rzio
= zio_read_phys(pio
, vd
, addr
,
5447 l2arc_read_done
, cb
, priority
,
5448 zio_flags
| ZIO_FLAG_DONT_CACHE
|
5450 ZIO_FLAG_DONT_PROPAGATE
|
5451 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
5453 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
5455 ARCSTAT_INCR(arcstat_l2_read_bytes
, size
);
5457 if (*arc_flags
& ARC_FLAG_NOWAIT
) {
5462 ASSERT(*arc_flags
& ARC_FLAG_WAIT
);
5463 if (zio_wait(rzio
) == 0)
5466 /* l2arc read error; goto zio_read() */
5468 DTRACE_PROBE1(l2arc__miss
,
5469 arc_buf_hdr_t
*, hdr
);
5470 ARCSTAT_BUMP(arcstat_l2_misses
);
5471 if (HDR_L2_WRITING(hdr
))
5472 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
5473 spa_config_exit(spa
, SCL_L2ARC
, vd
);
5477 spa_config_exit(spa
, SCL_L2ARC
, vd
);
5478 if (l2arc_ndev
!= 0) {
5479 DTRACE_PROBE1(l2arc__miss
,
5480 arc_buf_hdr_t
*, hdr
);
5481 ARCSTAT_BUMP(arcstat_l2_misses
);
5485 rzio
= zio_read(pio
, spa
, bp
, hdr
->b_l1hdr
.b_pabd
, size
,
5486 arc_read_done
, hdr
, priority
, zio_flags
, zb
);
5488 if (*arc_flags
& ARC_FLAG_WAIT
) {
5489 rc
= zio_wait(rzio
);
5493 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5498 spa_read_history_add(spa
, zb
, *arc_flags
);
5503 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
5507 p
= kmem_alloc(sizeof (*p
), KM_SLEEP
);
5509 p
->p_private
= private;
5510 list_link_init(&p
->p_node
);
5511 refcount_create(&p
->p_refcnt
);
5513 mutex_enter(&arc_prune_mtx
);
5514 refcount_add(&p
->p_refcnt
, &arc_prune_list
);
5515 list_insert_head(&arc_prune_list
, p
);
5516 mutex_exit(&arc_prune_mtx
);
5522 arc_remove_prune_callback(arc_prune_t
*p
)
5524 boolean_t wait
= B_FALSE
;
5525 mutex_enter(&arc_prune_mtx
);
5526 list_remove(&arc_prune_list
, p
);
5527 if (refcount_remove(&p
->p_refcnt
, &arc_prune_list
) > 0)
5529 mutex_exit(&arc_prune_mtx
);
5531 /* wait for arc_prune_task to finish */
5533 taskq_wait_outstanding(arc_prune_taskq
, 0);
5534 ASSERT0(refcount_count(&p
->p_refcnt
));
5535 refcount_destroy(&p
->p_refcnt
);
5536 kmem_free(p
, sizeof (*p
));
5540 * Notify the arc that a block was freed, and thus will never be used again.
5543 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
5546 kmutex_t
*hash_lock
;
5547 uint64_t guid
= spa_load_guid(spa
);
5549 ASSERT(!BP_IS_EMBEDDED(bp
));
5551 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
5556 * We might be trying to free a block that is still doing I/O
5557 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5558 * dmu_sync-ed block). If this block is being prefetched, then it
5559 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5560 * until the I/O completes. A block may also have a reference if it is
5561 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5562 * have written the new block to its final resting place on disk but
5563 * without the dedup flag set. This would have left the hdr in the MRU
5564 * state and discoverable. When the txg finally syncs it detects that
5565 * the block was overridden in open context and issues an override I/O.
5566 * Since this is a dedup block, the override I/O will determine if the
5567 * block is already in the DDT. If so, then it will replace the io_bp
5568 * with the bp from the DDT and allow the I/O to finish. When the I/O
5569 * reaches the done callback, dbuf_write_override_done, it will
5570 * check to see if the io_bp and io_bp_override are identical.
5571 * If they are not, then it indicates that the bp was replaced with
5572 * the bp in the DDT and the override bp is freed. This allows
5573 * us to arrive here with a reference on a block that is being
5574 * freed. So if we have an I/O in progress, or a reference to
5575 * this hdr, then we don't destroy the hdr.
5577 if (!HDR_HAS_L1HDR(hdr
) || (!HDR_IO_IN_PROGRESS(hdr
) &&
5578 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
))) {
5579 arc_change_state(arc_anon
, hdr
, hash_lock
);
5580 arc_hdr_destroy(hdr
);
5581 mutex_exit(hash_lock
);
5583 mutex_exit(hash_lock
);
5589 * Release this buffer from the cache, making it an anonymous buffer. This
5590 * must be done after a read and prior to modifying the buffer contents.
5591 * If the buffer has more than one reference, we must make
5592 * a new hdr for the buffer.
5595 arc_release(arc_buf_t
*buf
, void *tag
)
5597 kmutex_t
*hash_lock
;
5599 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5602 * It would be nice to assert that if its DMU metadata (level >
5603 * 0 || it's the dnode file), then it must be syncing context.
5604 * But we don't know that information at this level.
5607 mutex_enter(&buf
->b_evict_lock
);
5609 ASSERT(HDR_HAS_L1HDR(hdr
));
5612 * We don't grab the hash lock prior to this check, because if
5613 * the buffer's header is in the arc_anon state, it won't be
5614 * linked into the hash table.
5616 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
5617 mutex_exit(&buf
->b_evict_lock
);
5618 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5619 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
5620 ASSERT(!HDR_HAS_L2HDR(hdr
));
5621 ASSERT(HDR_EMPTY(hdr
));
5623 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
5624 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), ==, 1);
5625 ASSERT(!list_link_active(&hdr
->b_l1hdr
.b_arc_node
));
5627 hdr
->b_l1hdr
.b_arc_access
= 0;
5630 * If the buf is being overridden then it may already
5631 * have a hdr that is not empty.
5633 buf_discard_identity(hdr
);
5639 hash_lock
= HDR_LOCK(hdr
);
5640 mutex_enter(hash_lock
);
5643 * This assignment is only valid as long as the hash_lock is
5644 * held, we must be careful not to reference state or the
5645 * b_state field after dropping the lock.
5647 state
= hdr
->b_l1hdr
.b_state
;
5648 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
5649 ASSERT3P(state
, !=, arc_anon
);
5651 /* this buffer is not on any list */
5652 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), >, 0);
5654 if (HDR_HAS_L2HDR(hdr
)) {
5655 mutex_enter(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
5658 * We have to recheck this conditional again now that
5659 * we're holding the l2ad_mtx to prevent a race with
5660 * another thread which might be concurrently calling
5661 * l2arc_evict(). In that case, l2arc_evict() might have
5662 * destroyed the header's L2 portion as we were waiting
5663 * to acquire the l2ad_mtx.
5665 if (HDR_HAS_L2HDR(hdr
))
5666 arc_hdr_l2hdr_destroy(hdr
);
5668 mutex_exit(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
5672 * Do we have more than one buf?
5674 if (hdr
->b_l1hdr
.b_bufcnt
> 1) {
5675 arc_buf_hdr_t
*nhdr
;
5676 uint64_t spa
= hdr
->b_spa
;
5677 uint64_t psize
= HDR_GET_PSIZE(hdr
);
5678 uint64_t lsize
= HDR_GET_LSIZE(hdr
);
5679 enum zio_compress compress
= HDR_GET_COMPRESS(hdr
);
5680 arc_buf_contents_t type
= arc_buf_type(hdr
);
5681 VERIFY3U(hdr
->b_type
, ==, type
);
5683 ASSERT(hdr
->b_l1hdr
.b_buf
!= buf
|| buf
->b_next
!= NULL
);
5684 (void) remove_reference(hdr
, hash_lock
, tag
);
5686 if (arc_buf_is_shared(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
5687 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
5688 ASSERT(ARC_BUF_LAST(buf
));
5692 * Pull the data off of this hdr and attach it to
5693 * a new anonymous hdr. Also find the last buffer
5694 * in the hdr's buffer list.
5696 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
5697 ASSERT3P(lastbuf
, !=, NULL
);
5700 * If the current arc_buf_t and the hdr are sharing their data
5701 * buffer, then we must stop sharing that block.
5703 if (arc_buf_is_shared(buf
)) {
5704 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
5705 VERIFY(!arc_buf_is_shared(lastbuf
));
5708 * First, sever the block sharing relationship between
5709 * buf and the arc_buf_hdr_t.
5711 arc_unshare_buf(hdr
, buf
);
5714 * Now we need to recreate the hdr's b_pabd. Since we
5715 * have lastbuf handy, we try to share with it, but if
5716 * we can't then we allocate a new b_pabd and copy the
5717 * data from buf into it.
5719 if (arc_can_share(hdr
, lastbuf
)) {
5720 arc_share_buf(hdr
, lastbuf
);
5722 arc_hdr_alloc_pabd(hdr
);
5723 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
,
5724 buf
->b_data
, psize
);
5726 VERIFY3P(lastbuf
->b_data
, !=, NULL
);
5727 } else if (HDR_SHARED_DATA(hdr
)) {
5729 * Uncompressed shared buffers are always at the end
5730 * of the list. Compressed buffers don't have the
5731 * same requirements. This makes it hard to
5732 * simply assert that the lastbuf is shared so
5733 * we rely on the hdr's compression flags to determine
5734 * if we have a compressed, shared buffer.
5736 ASSERT(arc_buf_is_shared(lastbuf
) ||
5737 HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
5738 ASSERT(!ARC_BUF_SHARED(buf
));
5740 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5741 ASSERT3P(state
, !=, arc_l2c_only
);
5743 (void) refcount_remove_many(&state
->arcs_size
,
5744 arc_buf_size(buf
), buf
);
5746 if (refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
)) {
5747 ASSERT3P(state
, !=, arc_l2c_only
);
5748 (void) refcount_remove_many(&state
->arcs_esize
[type
],
5749 arc_buf_size(buf
), buf
);
5752 hdr
->b_l1hdr
.b_bufcnt
-= 1;
5753 arc_cksum_verify(buf
);
5754 arc_buf_unwatch(buf
);
5756 /* if this is the last uncompressed buf free the checksum */
5757 if (!arc_hdr_has_uncompressed_buf(hdr
))
5758 arc_cksum_free(hdr
);
5760 mutex_exit(hash_lock
);
5763 * Allocate a new hdr. The new hdr will contain a b_pabd
5764 * buffer which will be freed in arc_write().
5766 nhdr
= arc_hdr_alloc(spa
, psize
, lsize
, compress
, type
);
5767 ASSERT3P(nhdr
->b_l1hdr
.b_buf
, ==, NULL
);
5768 ASSERT0(nhdr
->b_l1hdr
.b_bufcnt
);
5769 ASSERT0(refcount_count(&nhdr
->b_l1hdr
.b_refcnt
));
5770 VERIFY3U(nhdr
->b_type
, ==, type
);
5771 ASSERT(!HDR_SHARED_DATA(nhdr
));
5773 nhdr
->b_l1hdr
.b_buf
= buf
;
5774 nhdr
->b_l1hdr
.b_bufcnt
= 1;
5775 nhdr
->b_l1hdr
.b_mru_hits
= 0;
5776 nhdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
5777 nhdr
->b_l1hdr
.b_mfu_hits
= 0;
5778 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
5779 nhdr
->b_l1hdr
.b_l2_hits
= 0;
5780 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, tag
);
5783 mutex_exit(&buf
->b_evict_lock
);
5784 (void) refcount_add_many(&arc_anon
->arcs_size
,
5785 HDR_GET_LSIZE(nhdr
), buf
);
5787 mutex_exit(&buf
->b_evict_lock
);
5788 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 1);
5789 /* protected by hash lock, or hdr is on arc_anon */
5790 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
5791 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5792 hdr
->b_l1hdr
.b_mru_hits
= 0;
5793 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
5794 hdr
->b_l1hdr
.b_mfu_hits
= 0;
5795 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
5796 hdr
->b_l1hdr
.b_l2_hits
= 0;
5797 arc_change_state(arc_anon
, hdr
, hash_lock
);
5798 hdr
->b_l1hdr
.b_arc_access
= 0;
5799 mutex_exit(hash_lock
);
5801 buf_discard_identity(hdr
);
5807 arc_released(arc_buf_t
*buf
)
5811 mutex_enter(&buf
->b_evict_lock
);
5812 released
= (buf
->b_data
!= NULL
&&
5813 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
5814 mutex_exit(&buf
->b_evict_lock
);
5820 arc_referenced(arc_buf_t
*buf
)
5824 mutex_enter(&buf
->b_evict_lock
);
5825 referenced
= (refcount_count(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
5826 mutex_exit(&buf
->b_evict_lock
);
5827 return (referenced
);
5832 arc_write_ready(zio_t
*zio
)
5834 arc_write_callback_t
*callback
= zio
->io_private
;
5835 arc_buf_t
*buf
= callback
->awcb_buf
;
5836 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5837 uint64_t psize
= BP_IS_HOLE(zio
->io_bp
) ? 0 : BP_GET_PSIZE(zio
->io_bp
);
5838 enum zio_compress compress
;
5839 fstrans_cookie_t cookie
= spl_fstrans_mark();
5841 ASSERT(HDR_HAS_L1HDR(hdr
));
5842 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
5843 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
5846 * If we're reexecuting this zio because the pool suspended, then
5847 * cleanup any state that was previously set the first time the
5848 * callback was invoked.
5850 if (zio
->io_flags
& ZIO_FLAG_REEXECUTED
) {
5851 arc_cksum_free(hdr
);
5852 arc_buf_unwatch(buf
);
5853 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5854 if (arc_buf_is_shared(buf
)) {
5855 arc_unshare_buf(hdr
, buf
);
5857 arc_hdr_free_pabd(hdr
);
5861 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5862 ASSERT(!HDR_SHARED_DATA(hdr
));
5863 ASSERT(!arc_buf_is_shared(buf
));
5865 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
5867 if (HDR_IO_IN_PROGRESS(hdr
))
5868 ASSERT(zio
->io_flags
& ZIO_FLAG_REEXECUTED
);
5870 arc_cksum_compute(buf
);
5871 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5873 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
5874 compress
= ZIO_COMPRESS_OFF
;
5876 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(zio
->io_bp
));
5877 compress
= BP_GET_COMPRESS(zio
->io_bp
);
5879 HDR_SET_PSIZE(hdr
, psize
);
5880 arc_hdr_set_compress(hdr
, compress
);
5883 * Fill the hdr with data. If the hdr is compressed, the data we want
5884 * is available from the zio, otherwise we can take it from the buf.
5886 * We might be able to share the buf's data with the hdr here. However,
5887 * doing so would cause the ARC to be full of linear ABDs if we write a
5888 * lot of shareable data. As a compromise, we check whether scattered
5889 * ABDs are allowed, and assume that if they are then the user wants
5890 * the ARC to be primarily filled with them regardless of the data being
5891 * written. Therefore, if they're allowed then we allocate one and copy
5892 * the data into it; otherwise, we share the data directly if we can.
5894 if (zfs_abd_scatter_enabled
|| !arc_can_share(hdr
, buf
)) {
5895 arc_hdr_alloc_pabd(hdr
);
5898 * Ideally, we would always copy the io_abd into b_pabd, but the
5899 * user may have disabled compressed ARC, thus we must check the
5900 * hdr's compression setting rather than the io_bp's.
5902 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
) {
5903 ASSERT3U(BP_GET_COMPRESS(zio
->io_bp
), !=,
5905 ASSERT3U(psize
, >, 0);
5907 abd_copy(hdr
->b_l1hdr
.b_pabd
, zio
->io_abd
, psize
);
5909 ASSERT3U(zio
->io_orig_size
, ==, arc_hdr_size(hdr
));
5911 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
5915 ASSERT3P(buf
->b_data
, ==, abd_to_buf(zio
->io_orig_abd
));
5916 ASSERT3U(zio
->io_orig_size
, ==, arc_buf_size(buf
));
5917 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
5919 arc_share_buf(hdr
, buf
);
5922 arc_hdr_verify(hdr
, zio
->io_bp
);
5923 spl_fstrans_unmark(cookie
);
5927 arc_write_children_ready(zio_t
*zio
)
5929 arc_write_callback_t
*callback
= zio
->io_private
;
5930 arc_buf_t
*buf
= callback
->awcb_buf
;
5932 callback
->awcb_children_ready(zio
, buf
, callback
->awcb_private
);
5936 * The SPA calls this callback for each physical write that happens on behalf
5937 * of a logical write. See the comment in dbuf_write_physdone() for details.
5940 arc_write_physdone(zio_t
*zio
)
5942 arc_write_callback_t
*cb
= zio
->io_private
;
5943 if (cb
->awcb_physdone
!= NULL
)
5944 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
5948 arc_write_done(zio_t
*zio
)
5950 arc_write_callback_t
*callback
= zio
->io_private
;
5951 arc_buf_t
*buf
= callback
->awcb_buf
;
5952 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5954 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5956 if (zio
->io_error
== 0) {
5957 arc_hdr_verify(hdr
, zio
->io_bp
);
5959 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
5960 buf_discard_identity(hdr
);
5962 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
5963 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
5966 ASSERT(HDR_EMPTY(hdr
));
5970 * If the block to be written was all-zero or compressed enough to be
5971 * embedded in the BP, no write was performed so there will be no
5972 * dva/birth/checksum. The buffer must therefore remain anonymous
5975 if (!HDR_EMPTY(hdr
)) {
5976 arc_buf_hdr_t
*exists
;
5977 kmutex_t
*hash_lock
;
5979 ASSERT3U(zio
->io_error
, ==, 0);
5981 arc_cksum_verify(buf
);
5983 exists
= buf_hash_insert(hdr
, &hash_lock
);
5984 if (exists
!= NULL
) {
5986 * This can only happen if we overwrite for
5987 * sync-to-convergence, because we remove
5988 * buffers from the hash table when we arc_free().
5990 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
5991 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
5992 panic("bad overwrite, hdr=%p exists=%p",
5993 (void *)hdr
, (void *)exists
);
5994 ASSERT(refcount_is_zero(
5995 &exists
->b_l1hdr
.b_refcnt
));
5996 arc_change_state(arc_anon
, exists
, hash_lock
);
5997 mutex_exit(hash_lock
);
5998 arc_hdr_destroy(exists
);
5999 exists
= buf_hash_insert(hdr
, &hash_lock
);
6000 ASSERT3P(exists
, ==, NULL
);
6001 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
6003 ASSERT(zio
->io_prop
.zp_nopwrite
);
6004 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
6005 panic("bad nopwrite, hdr=%p exists=%p",
6006 (void *)hdr
, (void *)exists
);
6009 ASSERT(hdr
->b_l1hdr
.b_bufcnt
== 1);
6010 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
6011 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
6012 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
6015 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6016 /* if it's not anon, we are doing a scrub */
6017 if (exists
== NULL
&& hdr
->b_l1hdr
.b_state
== arc_anon
)
6018 arc_access(hdr
, hash_lock
);
6019 mutex_exit(hash_lock
);
6021 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6024 ASSERT(!refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
6025 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
6027 abd_put(zio
->io_abd
);
6028 kmem_free(callback
, sizeof (arc_write_callback_t
));
6032 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
6033 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
,
6034 const zio_prop_t
*zp
, arc_done_func_t
*ready
,
6035 arc_done_func_t
*children_ready
, arc_done_func_t
*physdone
,
6036 arc_done_func_t
*done
, void *private, zio_priority_t priority
,
6037 int zio_flags
, const zbookmark_phys_t
*zb
)
6039 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
6040 arc_write_callback_t
*callback
;
6042 zio_prop_t localprop
= *zp
;
6044 ASSERT3P(ready
, !=, NULL
);
6045 ASSERT3P(done
, !=, NULL
);
6046 ASSERT(!HDR_IO_ERROR(hdr
));
6047 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
6048 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
6049 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
6051 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
6052 if (ARC_BUF_COMPRESSED(buf
)) {
6054 * We're writing a pre-compressed buffer. Make the
6055 * compression algorithm requested by the zio_prop_t match
6056 * the pre-compressed buffer's compression algorithm.
6058 localprop
.zp_compress
= HDR_GET_COMPRESS(hdr
);
6060 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, arc_buf_size(buf
));
6061 zio_flags
|= ZIO_FLAG_RAW
;
6063 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
6064 callback
->awcb_ready
= ready
;
6065 callback
->awcb_children_ready
= children_ready
;
6066 callback
->awcb_physdone
= physdone
;
6067 callback
->awcb_done
= done
;
6068 callback
->awcb_private
= private;
6069 callback
->awcb_buf
= buf
;
6072 * The hdr's b_pabd is now stale, free it now. A new data block
6073 * will be allocated when the zio pipeline calls arc_write_ready().
6075 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
6077 * If the buf is currently sharing the data block with
6078 * the hdr then we need to break that relationship here.
6079 * The hdr will remain with a NULL data pointer and the
6080 * buf will take sole ownership of the block.
6082 if (arc_buf_is_shared(buf
)) {
6083 arc_unshare_buf(hdr
, buf
);
6085 arc_hdr_free_pabd(hdr
);
6087 VERIFY3P(buf
->b_data
, !=, NULL
);
6088 arc_hdr_set_compress(hdr
, ZIO_COMPRESS_OFF
);
6090 ASSERT(!arc_buf_is_shared(buf
));
6091 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
6093 zio
= zio_write(pio
, spa
, txg
, bp
,
6094 abd_get_from_buf(buf
->b_data
, HDR_GET_LSIZE(hdr
)),
6095 HDR_GET_LSIZE(hdr
), arc_buf_size(buf
), &localprop
, arc_write_ready
,
6096 (children_ready
!= NULL
) ? arc_write_children_ready
: NULL
,
6097 arc_write_physdone
, arc_write_done
, callback
,
6098 priority
, zio_flags
, zb
);
6104 arc_memory_throttle(uint64_t reserve
, uint64_t txg
)
6107 uint64_t available_memory
= arc_free_memory();
6108 static uint64_t page_load
= 0;
6109 static uint64_t last_txg
= 0;
6113 MIN(available_memory
, vmem_size(heap_arena
, VMEM_FREE
));
6116 if (available_memory
> arc_all_memory() * arc_lotsfree_percent
/ 100)
6119 if (txg
> last_txg
) {
6124 * If we are in pageout, we know that memory is already tight,
6125 * the arc is already going to be evicting, so we just want to
6126 * continue to let page writes occur as quickly as possible.
6128 if (current_is_kswapd()) {
6129 if (page_load
> MAX(arc_sys_free
/ 4, available_memory
) / 4) {
6130 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
6131 return (SET_ERROR(ERESTART
));
6133 /* Note: reserve is inflated, so we deflate */
6134 page_load
+= reserve
/ 8;
6136 } else if (page_load
> 0 && arc_reclaim_needed()) {
6137 /* memory is low, delay before restarting */
6138 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
6139 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
6140 return (SET_ERROR(EAGAIN
));
6148 arc_tempreserve_clear(uint64_t reserve
)
6150 atomic_add_64(&arc_tempreserve
, -reserve
);
6151 ASSERT((int64_t)arc_tempreserve
>= 0);
6155 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
6161 reserve
> arc_c
/4 &&
6162 reserve
* 4 > (2ULL << SPA_MAXBLOCKSHIFT
))
6163 arc_c
= MIN(arc_c_max
, reserve
* 4);
6166 * Throttle when the calculated memory footprint for the TXG
6167 * exceeds the target ARC size.
6169 if (reserve
> arc_c
) {
6170 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
6171 return (SET_ERROR(ERESTART
));
6175 * Don't count loaned bufs as in flight dirty data to prevent long
6176 * network delays from blocking transactions that are ready to be
6177 * assigned to a txg.
6180 /* assert that it has not wrapped around */
6181 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
6183 anon_size
= MAX((int64_t)(refcount_count(&arc_anon
->arcs_size
) -
6184 arc_loaned_bytes
), 0);
6187 * Writes will, almost always, require additional memory allocations
6188 * in order to compress/encrypt/etc the data. We therefore need to
6189 * make sure that there is sufficient available memory for this.
6191 error
= arc_memory_throttle(reserve
, txg
);
6196 * Throttle writes when the amount of dirty data in the cache
6197 * gets too large. We try to keep the cache less than half full
6198 * of dirty blocks so that our sync times don't grow too large.
6199 * Note: if two requests come in concurrently, we might let them
6200 * both succeed, when one of them should fail. Not a huge deal.
6203 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
6204 anon_size
> arc_c
/ 4) {
6205 uint64_t meta_esize
=
6206 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6207 uint64_t data_esize
=
6208 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6209 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6210 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6211 arc_tempreserve
>> 10, meta_esize
>> 10,
6212 data_esize
>> 10, reserve
>> 10, arc_c
>> 10);
6213 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
6214 return (SET_ERROR(ERESTART
));
6216 atomic_add_64(&arc_tempreserve
, reserve
);
6221 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
6222 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
6224 size
->value
.ui64
= refcount_count(&state
->arcs_size
);
6225 evict_data
->value
.ui64
=
6226 refcount_count(&state
->arcs_esize
[ARC_BUFC_DATA
]);
6227 evict_metadata
->value
.ui64
=
6228 refcount_count(&state
->arcs_esize
[ARC_BUFC_METADATA
]);
6232 arc_kstat_update(kstat_t
*ksp
, int rw
)
6234 arc_stats_t
*as
= ksp
->ks_data
;
6236 if (rw
== KSTAT_WRITE
) {
6239 arc_kstat_update_state(arc_anon
,
6240 &as
->arcstat_anon_size
,
6241 &as
->arcstat_anon_evictable_data
,
6242 &as
->arcstat_anon_evictable_metadata
);
6243 arc_kstat_update_state(arc_mru
,
6244 &as
->arcstat_mru_size
,
6245 &as
->arcstat_mru_evictable_data
,
6246 &as
->arcstat_mru_evictable_metadata
);
6247 arc_kstat_update_state(arc_mru_ghost
,
6248 &as
->arcstat_mru_ghost_size
,
6249 &as
->arcstat_mru_ghost_evictable_data
,
6250 &as
->arcstat_mru_ghost_evictable_metadata
);
6251 arc_kstat_update_state(arc_mfu
,
6252 &as
->arcstat_mfu_size
,
6253 &as
->arcstat_mfu_evictable_data
,
6254 &as
->arcstat_mfu_evictable_metadata
);
6255 arc_kstat_update_state(arc_mfu_ghost
,
6256 &as
->arcstat_mfu_ghost_size
,
6257 &as
->arcstat_mfu_ghost_evictable_data
,
6258 &as
->arcstat_mfu_ghost_evictable_metadata
);
6260 as
->arcstat_memory_all_bytes
.value
.ui64
=
6262 as
->arcstat_memory_free_bytes
.value
.ui64
=
6264 as
->arcstat_memory_available_bytes
.value
.i64
=
6265 arc_available_memory();
6272 * This function *must* return indices evenly distributed between all
6273 * sublists of the multilist. This is needed due to how the ARC eviction
6274 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6275 * distributed between all sublists and uses this assumption when
6276 * deciding which sublist to evict from and how much to evict from it.
6279 arc_state_multilist_index_func(multilist_t
*ml
, void *obj
)
6281 arc_buf_hdr_t
*hdr
= obj
;
6284 * We rely on b_dva to generate evenly distributed index
6285 * numbers using buf_hash below. So, as an added precaution,
6286 * let's make sure we never add empty buffers to the arc lists.
6288 ASSERT(!HDR_EMPTY(hdr
));
6291 * The assumption here, is the hash value for a given
6292 * arc_buf_hdr_t will remain constant throughout its lifetime
6293 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
6294 * Thus, we don't need to store the header's sublist index
6295 * on insertion, as this index can be recalculated on removal.
6297 * Also, the low order bits of the hash value are thought to be
6298 * distributed evenly. Otherwise, in the case that the multilist
6299 * has a power of two number of sublists, each sublists' usage
6300 * would not be evenly distributed.
6302 return (buf_hash(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
) %
6303 multilist_get_num_sublists(ml
));
6307 * Called during module initialization and periodically thereafter to
6308 * apply reasonable changes to the exposed performance tunings. Non-zero
6309 * zfs_* values which differ from the currently set values will be applied.
6312 arc_tuning_update(void)
6314 uint64_t allmem
= arc_all_memory();
6315 unsigned long limit
;
6317 /* Valid range: 64M - <all physical memory> */
6318 if ((zfs_arc_max
) && (zfs_arc_max
!= arc_c_max
) &&
6319 (zfs_arc_max
> 64 << 20) && (zfs_arc_max
< allmem
) &&
6320 (zfs_arc_max
> arc_c_min
)) {
6321 arc_c_max
= zfs_arc_max
;
6323 arc_p
= (arc_c
>> 1);
6324 if (arc_meta_limit
> arc_c_max
)
6325 arc_meta_limit
= arc_c_max
;
6326 if (arc_dnode_limit
> arc_meta_limit
)
6327 arc_dnode_limit
= arc_meta_limit
;
6330 /* Valid range: 32M - <arc_c_max> */
6331 if ((zfs_arc_min
) && (zfs_arc_min
!= arc_c_min
) &&
6332 (zfs_arc_min
>= 2ULL << SPA_MAXBLOCKSHIFT
) &&
6333 (zfs_arc_min
<= arc_c_max
)) {
6334 arc_c_min
= zfs_arc_min
;
6335 arc_c
= MAX(arc_c
, arc_c_min
);
6338 /* Valid range: 16M - <arc_c_max> */
6339 if ((zfs_arc_meta_min
) && (zfs_arc_meta_min
!= arc_meta_min
) &&
6340 (zfs_arc_meta_min
>= 1ULL << SPA_MAXBLOCKSHIFT
) &&
6341 (zfs_arc_meta_min
<= arc_c_max
)) {
6342 arc_meta_min
= zfs_arc_meta_min
;
6343 if (arc_meta_limit
< arc_meta_min
)
6344 arc_meta_limit
= arc_meta_min
;
6345 if (arc_dnode_limit
< arc_meta_min
)
6346 arc_dnode_limit
= arc_meta_min
;
6349 /* Valid range: <arc_meta_min> - <arc_c_max> */
6350 limit
= zfs_arc_meta_limit
? zfs_arc_meta_limit
:
6351 MIN(zfs_arc_meta_limit_percent
, 100) * arc_c_max
/ 100;
6352 if ((limit
!= arc_meta_limit
) &&
6353 (limit
>= arc_meta_min
) &&
6354 (limit
<= arc_c_max
))
6355 arc_meta_limit
= limit
;
6357 /* Valid range: <arc_meta_min> - <arc_meta_limit> */
6358 limit
= zfs_arc_dnode_limit
? zfs_arc_dnode_limit
:
6359 MIN(zfs_arc_dnode_limit_percent
, 100) * arc_meta_limit
/ 100;
6360 if ((limit
!= arc_dnode_limit
) &&
6361 (limit
>= arc_meta_min
) &&
6362 (limit
<= arc_meta_limit
))
6363 arc_dnode_limit
= limit
;
6365 /* Valid range: 1 - N */
6366 if (zfs_arc_grow_retry
)
6367 arc_grow_retry
= zfs_arc_grow_retry
;
6369 /* Valid range: 1 - N */
6370 if (zfs_arc_shrink_shift
) {
6371 arc_shrink_shift
= zfs_arc_shrink_shift
;
6372 arc_no_grow_shift
= MIN(arc_no_grow_shift
, arc_shrink_shift
-1);
6375 /* Valid range: 1 - N */
6376 if (zfs_arc_p_min_shift
)
6377 arc_p_min_shift
= zfs_arc_p_min_shift
;
6379 /* Valid range: 1 - N ticks */
6380 if (zfs_arc_min_prefetch_lifespan
)
6381 arc_min_prefetch_lifespan
= zfs_arc_min_prefetch_lifespan
;
6383 /* Valid range: 0 - 100 */
6384 if ((zfs_arc_lotsfree_percent
>= 0) &&
6385 (zfs_arc_lotsfree_percent
<= 100))
6386 arc_lotsfree_percent
= zfs_arc_lotsfree_percent
;
6388 /* Valid range: 0 - <all physical memory> */
6389 if ((zfs_arc_sys_free
) && (zfs_arc_sys_free
!= arc_sys_free
))
6390 arc_sys_free
= MIN(MAX(zfs_arc_sys_free
, 0), allmem
);
6395 arc_state_init(void)
6397 arc_anon
= &ARC_anon
;
6399 arc_mru_ghost
= &ARC_mru_ghost
;
6401 arc_mfu_ghost
= &ARC_mfu_ghost
;
6402 arc_l2c_only
= &ARC_l2c_only
;
6404 arc_mru
->arcs_list
[ARC_BUFC_METADATA
] =
6405 multilist_create(sizeof (arc_buf_hdr_t
),
6406 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6407 arc_state_multilist_index_func
);
6408 arc_mru
->arcs_list
[ARC_BUFC_DATA
] =
6409 multilist_create(sizeof (arc_buf_hdr_t
),
6410 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6411 arc_state_multilist_index_func
);
6412 arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
6413 multilist_create(sizeof (arc_buf_hdr_t
),
6414 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6415 arc_state_multilist_index_func
);
6416 arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
] =
6417 multilist_create(sizeof (arc_buf_hdr_t
),
6418 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6419 arc_state_multilist_index_func
);
6420 arc_mfu
->arcs_list
[ARC_BUFC_METADATA
] =
6421 multilist_create(sizeof (arc_buf_hdr_t
),
6422 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6423 arc_state_multilist_index_func
);
6424 arc_mfu
->arcs_list
[ARC_BUFC_DATA
] =
6425 multilist_create(sizeof (arc_buf_hdr_t
),
6426 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6427 arc_state_multilist_index_func
);
6428 arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
6429 multilist_create(sizeof (arc_buf_hdr_t
),
6430 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6431 arc_state_multilist_index_func
);
6432 arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
] =
6433 multilist_create(sizeof (arc_buf_hdr_t
),
6434 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6435 arc_state_multilist_index_func
);
6436 arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
] =
6437 multilist_create(sizeof (arc_buf_hdr_t
),
6438 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6439 arc_state_multilist_index_func
);
6440 arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
] =
6441 multilist_create(sizeof (arc_buf_hdr_t
),
6442 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6443 arc_state_multilist_index_func
);
6445 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6446 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6447 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
6448 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
6449 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6450 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6451 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
6452 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
6453 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6454 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6455 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
6456 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
6458 refcount_create(&arc_anon
->arcs_size
);
6459 refcount_create(&arc_mru
->arcs_size
);
6460 refcount_create(&arc_mru_ghost
->arcs_size
);
6461 refcount_create(&arc_mfu
->arcs_size
);
6462 refcount_create(&arc_mfu_ghost
->arcs_size
);
6463 refcount_create(&arc_l2c_only
->arcs_size
);
6465 arc_anon
->arcs_state
= ARC_STATE_ANON
;
6466 arc_mru
->arcs_state
= ARC_STATE_MRU
;
6467 arc_mru_ghost
->arcs_state
= ARC_STATE_MRU_GHOST
;
6468 arc_mfu
->arcs_state
= ARC_STATE_MFU
;
6469 arc_mfu_ghost
->arcs_state
= ARC_STATE_MFU_GHOST
;
6470 arc_l2c_only
->arcs_state
= ARC_STATE_L2C_ONLY
;
6474 arc_state_fini(void)
6476 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6477 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6478 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
6479 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
6480 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6481 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6482 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
6483 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
6484 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6485 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6486 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
6487 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
6489 refcount_destroy(&arc_anon
->arcs_size
);
6490 refcount_destroy(&arc_mru
->arcs_size
);
6491 refcount_destroy(&arc_mru_ghost
->arcs_size
);
6492 refcount_destroy(&arc_mfu
->arcs_size
);
6493 refcount_destroy(&arc_mfu_ghost
->arcs_size
);
6494 refcount_destroy(&arc_l2c_only
->arcs_size
);
6496 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
6497 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
6498 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
6499 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
6500 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
6501 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
6502 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
6503 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
6504 multilist_destroy(arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]);
6505 multilist_destroy(arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]);
6509 arc_target_bytes(void)
6517 uint64_t percent
, allmem
= arc_all_memory();
6519 mutex_init(&arc_reclaim_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
6520 cv_init(&arc_reclaim_thread_cv
, NULL
, CV_DEFAULT
, NULL
);
6521 cv_init(&arc_reclaim_waiters_cv
, NULL
, CV_DEFAULT
, NULL
);
6523 /* Convert seconds to clock ticks */
6524 arc_min_prefetch_lifespan
= 1 * hz
;
6528 * Register a shrinker to support synchronous (direct) memory
6529 * reclaim from the arc. This is done to prevent kswapd from
6530 * swapping out pages when it is preferable to shrink the arc.
6532 spl_register_shrinker(&arc_shrinker
);
6534 /* Set to 1/64 of all memory or a minimum of 512K */
6535 arc_sys_free
= MAX(allmem
/ 64, (512 * 1024));
6539 /* Set max to 1/2 of all memory */
6540 arc_c_max
= allmem
/ 2;
6543 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more */
6544 arc_c_min
= MAX(allmem
/ 32, 2ULL << SPA_MAXBLOCKSHIFT
);
6547 * In userland, there's only the memory pressure that we artificially
6548 * create (see arc_available_memory()). Don't let arc_c get too
6549 * small, because it can cause transactions to be larger than
6550 * arc_c, causing arc_tempreserve_space() to fail.
6552 arc_c_min
= MAX(arc_c_max
/ 2, 2ULL << SPA_MAXBLOCKSHIFT
);
6556 arc_p
= (arc_c
>> 1);
6559 /* Set min to 1/2 of arc_c_min */
6560 arc_meta_min
= 1ULL << SPA_MAXBLOCKSHIFT
;
6561 /* Initialize maximum observed usage to zero */
6564 * Set arc_meta_limit to a percent of arc_c_max with a floor of
6565 * arc_meta_min, and a ceiling of arc_c_max.
6567 percent
= MIN(zfs_arc_meta_limit_percent
, 100);
6568 arc_meta_limit
= MAX(arc_meta_min
, (percent
* arc_c_max
) / 100);
6569 percent
= MIN(zfs_arc_dnode_limit_percent
, 100);
6570 arc_dnode_limit
= (percent
* arc_meta_limit
) / 100;
6572 /* Apply user specified tunings */
6573 arc_tuning_update();
6575 /* if kmem_flags are set, lets try to use less memory */
6576 if (kmem_debugging())
6578 if (arc_c
< arc_c_min
)
6584 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
6585 offsetof(arc_prune_t
, p_node
));
6586 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6588 arc_prune_taskq
= taskq_create("arc_prune", max_ncpus
, defclsyspri
,
6589 max_ncpus
, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
6591 arc_reclaim_thread_exit
= B_FALSE
;
6593 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
6594 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
6596 if (arc_ksp
!= NULL
) {
6597 arc_ksp
->ks_data
= &arc_stats
;
6598 arc_ksp
->ks_update
= arc_kstat_update
;
6599 kstat_install(arc_ksp
);
6602 (void) thread_create(NULL
, 0, arc_reclaim_thread
, NULL
, 0, &p0
,
6603 TS_RUN
, defclsyspri
);
6609 * Calculate maximum amount of dirty data per pool.
6611 * If it has been set by a module parameter, take that.
6612 * Otherwise, use a percentage of physical memory defined by
6613 * zfs_dirty_data_max_percent (default 10%) with a cap at
6614 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
6616 if (zfs_dirty_data_max_max
== 0)
6617 zfs_dirty_data_max_max
= MIN(4ULL * 1024 * 1024 * 1024,
6618 allmem
* zfs_dirty_data_max_max_percent
/ 100);
6620 if (zfs_dirty_data_max
== 0) {
6621 zfs_dirty_data_max
= allmem
*
6622 zfs_dirty_data_max_percent
/ 100;
6623 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
6624 zfs_dirty_data_max_max
);
6634 spl_unregister_shrinker(&arc_shrinker
);
6635 #endif /* _KERNEL */
6637 mutex_enter(&arc_reclaim_lock
);
6638 arc_reclaim_thread_exit
= B_TRUE
;
6640 * The reclaim thread will set arc_reclaim_thread_exit back to
6641 * B_FALSE when it is finished exiting; we're waiting for that.
6643 while (arc_reclaim_thread_exit
) {
6644 cv_signal(&arc_reclaim_thread_cv
);
6645 cv_wait(&arc_reclaim_thread_cv
, &arc_reclaim_lock
);
6647 mutex_exit(&arc_reclaim_lock
);
6649 /* Use B_TRUE to ensure *all* buffers are evicted */
6650 arc_flush(NULL
, B_TRUE
);
6654 if (arc_ksp
!= NULL
) {
6655 kstat_delete(arc_ksp
);
6659 taskq_wait(arc_prune_taskq
);
6660 taskq_destroy(arc_prune_taskq
);
6662 mutex_enter(&arc_prune_mtx
);
6663 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
6664 list_remove(&arc_prune_list
, p
);
6665 refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
6666 refcount_destroy(&p
->p_refcnt
);
6667 kmem_free(p
, sizeof (*p
));
6669 mutex_exit(&arc_prune_mtx
);
6671 list_destroy(&arc_prune_list
);
6672 mutex_destroy(&arc_prune_mtx
);
6673 mutex_destroy(&arc_reclaim_lock
);
6674 cv_destroy(&arc_reclaim_thread_cv
);
6675 cv_destroy(&arc_reclaim_waiters_cv
);
6680 ASSERT0(arc_loaned_bytes
);
6686 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6687 * It uses dedicated storage devices to hold cached data, which are populated
6688 * using large infrequent writes. The main role of this cache is to boost
6689 * the performance of random read workloads. The intended L2ARC devices
6690 * include short-stroked disks, solid state disks, and other media with
6691 * substantially faster read latency than disk.
6693 * +-----------------------+
6695 * +-----------------------+
6698 * l2arc_feed_thread() arc_read()
6702 * +---------------+ |
6704 * +---------------+ |
6709 * +-------+ +-------+
6711 * | cache | | cache |
6712 * +-------+ +-------+
6713 * +=========+ .-----.
6714 * : L2ARC : |-_____-|
6715 * : devices : | Disks |
6716 * +=========+ `-_____-'
6718 * Read requests are satisfied from the following sources, in order:
6721 * 2) vdev cache of L2ARC devices
6723 * 4) vdev cache of disks
6726 * Some L2ARC device types exhibit extremely slow write performance.
6727 * To accommodate for this there are some significant differences between
6728 * the L2ARC and traditional cache design:
6730 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6731 * the ARC behave as usual, freeing buffers and placing headers on ghost
6732 * lists. The ARC does not send buffers to the L2ARC during eviction as
6733 * this would add inflated write latencies for all ARC memory pressure.
6735 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6736 * It does this by periodically scanning buffers from the eviction-end of
6737 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6738 * not already there. It scans until a headroom of buffers is satisfied,
6739 * which itself is a buffer for ARC eviction. If a compressible buffer is
6740 * found during scanning and selected for writing to an L2ARC device, we
6741 * temporarily boost scanning headroom during the next scan cycle to make
6742 * sure we adapt to compression effects (which might significantly reduce
6743 * the data volume we write to L2ARC). The thread that does this is
6744 * l2arc_feed_thread(), illustrated below; example sizes are included to
6745 * provide a better sense of ratio than this diagram:
6748 * +---------------------+----------+
6749 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6750 * +---------------------+----------+ | o L2ARC eligible
6751 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6752 * +---------------------+----------+ |
6753 * 15.9 Gbytes ^ 32 Mbytes |
6755 * l2arc_feed_thread()
6757 * l2arc write hand <--[oooo]--'
6761 * +==============================+
6762 * L2ARC dev |####|#|###|###| |####| ... |
6763 * +==============================+
6766 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6767 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6768 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6769 * safe to say that this is an uncommon case, since buffers at the end of
6770 * the ARC lists have moved there due to inactivity.
6772 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6773 * then the L2ARC simply misses copying some buffers. This serves as a
6774 * pressure valve to prevent heavy read workloads from both stalling the ARC
6775 * with waits and clogging the L2ARC with writes. This also helps prevent
6776 * the potential for the L2ARC to churn if it attempts to cache content too
6777 * quickly, such as during backups of the entire pool.
6779 * 5. After system boot and before the ARC has filled main memory, there are
6780 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6781 * lists can remain mostly static. Instead of searching from tail of these
6782 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6783 * for eligible buffers, greatly increasing its chance of finding them.
6785 * The L2ARC device write speed is also boosted during this time so that
6786 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6787 * there are no L2ARC reads, and no fear of degrading read performance
6788 * through increased writes.
6790 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6791 * the vdev queue can aggregate them into larger and fewer writes. Each
6792 * device is written to in a rotor fashion, sweeping writes through
6793 * available space then repeating.
6795 * 7. The L2ARC does not store dirty content. It never needs to flush
6796 * write buffers back to disk based storage.
6798 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6799 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6801 * The performance of the L2ARC can be tweaked by a number of tunables, which
6802 * may be necessary for different workloads:
6804 * l2arc_write_max max write bytes per interval
6805 * l2arc_write_boost extra write bytes during device warmup
6806 * l2arc_noprefetch skip caching prefetched buffers
6807 * l2arc_headroom number of max device writes to precache
6808 * l2arc_headroom_boost when we find compressed buffers during ARC
6809 * scanning, we multiply headroom by this
6810 * percentage factor for the next scan cycle,
6811 * since more compressed buffers are likely to
6813 * l2arc_feed_secs seconds between L2ARC writing
6815 * Tunables may be removed or added as future performance improvements are
6816 * integrated, and also may become zpool properties.
6818 * There are three key functions that control how the L2ARC warms up:
6820 * l2arc_write_eligible() check if a buffer is eligible to cache
6821 * l2arc_write_size() calculate how much to write
6822 * l2arc_write_interval() calculate sleep delay between writes
6824 * These three functions determine what to write, how much, and how quickly
6829 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*hdr
)
6832 * A buffer is *not* eligible for the L2ARC if it:
6833 * 1. belongs to a different spa.
6834 * 2. is already cached on the L2ARC.
6835 * 3. has an I/O in progress (it may be an incomplete read).
6836 * 4. is flagged not eligible (zfs property).
6838 if (hdr
->b_spa
!= spa_guid
|| HDR_HAS_L2HDR(hdr
) ||
6839 HDR_IO_IN_PROGRESS(hdr
) || !HDR_L2CACHE(hdr
))
6846 l2arc_write_size(void)
6851 * Make sure our globals have meaningful values in case the user
6854 size
= l2arc_write_max
;
6856 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
6857 "be greater than zero, resetting it to the default (%d)",
6859 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
6862 if (arc_warm
== B_FALSE
)
6863 size
+= l2arc_write_boost
;
6870 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
6872 clock_t interval
, next
, now
;
6875 * If the ARC lists are busy, increase our write rate; if the
6876 * lists are stale, idle back. This is achieved by checking
6877 * how much we previously wrote - if it was more than half of
6878 * what we wanted, schedule the next write much sooner.
6880 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
6881 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
6883 interval
= hz
* l2arc_feed_secs
;
6885 now
= ddi_get_lbolt();
6886 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
6892 * Cycle through L2ARC devices. This is how L2ARC load balances.
6893 * If a device is returned, this also returns holding the spa config lock.
6895 static l2arc_dev_t
*
6896 l2arc_dev_get_next(void)
6898 l2arc_dev_t
*first
, *next
= NULL
;
6901 * Lock out the removal of spas (spa_namespace_lock), then removal
6902 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6903 * both locks will be dropped and a spa config lock held instead.
6905 mutex_enter(&spa_namespace_lock
);
6906 mutex_enter(&l2arc_dev_mtx
);
6908 /* if there are no vdevs, there is nothing to do */
6909 if (l2arc_ndev
== 0)
6913 next
= l2arc_dev_last
;
6915 /* loop around the list looking for a non-faulted vdev */
6917 next
= list_head(l2arc_dev_list
);
6919 next
= list_next(l2arc_dev_list
, next
);
6921 next
= list_head(l2arc_dev_list
);
6924 /* if we have come back to the start, bail out */
6927 else if (next
== first
)
6930 } while (vdev_is_dead(next
->l2ad_vdev
));
6932 /* if we were unable to find any usable vdevs, return NULL */
6933 if (vdev_is_dead(next
->l2ad_vdev
))
6936 l2arc_dev_last
= next
;
6939 mutex_exit(&l2arc_dev_mtx
);
6942 * Grab the config lock to prevent the 'next' device from being
6943 * removed while we are writing to it.
6946 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
6947 mutex_exit(&spa_namespace_lock
);
6953 * Free buffers that were tagged for destruction.
6956 l2arc_do_free_on_write(void)
6959 l2arc_data_free_t
*df
, *df_prev
;
6961 mutex_enter(&l2arc_free_on_write_mtx
);
6962 buflist
= l2arc_free_on_write
;
6964 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
6965 df_prev
= list_prev(buflist
, df
);
6966 ASSERT3P(df
->l2df_abd
, !=, NULL
);
6967 abd_free(df
->l2df_abd
);
6968 list_remove(buflist
, df
);
6969 kmem_free(df
, sizeof (l2arc_data_free_t
));
6972 mutex_exit(&l2arc_free_on_write_mtx
);
6976 * A write to a cache device has completed. Update all headers to allow
6977 * reads from these buffers to begin.
6980 l2arc_write_done(zio_t
*zio
)
6982 l2arc_write_callback_t
*cb
;
6985 arc_buf_hdr_t
*head
, *hdr
, *hdr_prev
;
6986 kmutex_t
*hash_lock
;
6987 int64_t bytes_dropped
= 0;
6989 cb
= zio
->io_private
;
6990 ASSERT3P(cb
, !=, NULL
);
6991 dev
= cb
->l2wcb_dev
;
6992 ASSERT3P(dev
, !=, NULL
);
6993 head
= cb
->l2wcb_head
;
6994 ASSERT3P(head
, !=, NULL
);
6995 buflist
= &dev
->l2ad_buflist
;
6996 ASSERT3P(buflist
, !=, NULL
);
6997 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
6998 l2arc_write_callback_t
*, cb
);
7000 if (zio
->io_error
!= 0)
7001 ARCSTAT_BUMP(arcstat_l2_writes_error
);
7004 * All writes completed, or an error was hit.
7007 mutex_enter(&dev
->l2ad_mtx
);
7008 for (hdr
= list_prev(buflist
, head
); hdr
; hdr
= hdr_prev
) {
7009 hdr_prev
= list_prev(buflist
, hdr
);
7011 hash_lock
= HDR_LOCK(hdr
);
7014 * We cannot use mutex_enter or else we can deadlock
7015 * with l2arc_write_buffers (due to swapping the order
7016 * the hash lock and l2ad_mtx are taken).
7018 if (!mutex_tryenter(hash_lock
)) {
7020 * Missed the hash lock. We must retry so we
7021 * don't leave the ARC_FLAG_L2_WRITING bit set.
7023 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry
);
7026 * We don't want to rescan the headers we've
7027 * already marked as having been written out, so
7028 * we reinsert the head node so we can pick up
7029 * where we left off.
7031 list_remove(buflist
, head
);
7032 list_insert_after(buflist
, hdr
, head
);
7034 mutex_exit(&dev
->l2ad_mtx
);
7037 * We wait for the hash lock to become available
7038 * to try and prevent busy waiting, and increase
7039 * the chance we'll be able to acquire the lock
7040 * the next time around.
7042 mutex_enter(hash_lock
);
7043 mutex_exit(hash_lock
);
7048 * We could not have been moved into the arc_l2c_only
7049 * state while in-flight due to our ARC_FLAG_L2_WRITING
7050 * bit being set. Let's just ensure that's being enforced.
7052 ASSERT(HDR_HAS_L1HDR(hdr
));
7055 * Skipped - drop L2ARC entry and mark the header as no
7056 * longer L2 eligibile.
7058 if (zio
->io_error
!= 0) {
7060 * Error - drop L2ARC entry.
7062 list_remove(buflist
, hdr
);
7063 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
7065 ARCSTAT_INCR(arcstat_l2_psize
, -arc_hdr_size(hdr
));
7066 ARCSTAT_INCR(arcstat_l2_lsize
, -HDR_GET_LSIZE(hdr
));
7068 bytes_dropped
+= arc_hdr_size(hdr
);
7069 (void) refcount_remove_many(&dev
->l2ad_alloc
,
7070 arc_hdr_size(hdr
), hdr
);
7074 * Allow ARC to begin reads and ghost list evictions to
7077 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_WRITING
);
7079 mutex_exit(hash_lock
);
7082 atomic_inc_64(&l2arc_writes_done
);
7083 list_remove(buflist
, head
);
7084 ASSERT(!HDR_HAS_L1HDR(head
));
7085 kmem_cache_free(hdr_l2only_cache
, head
);
7086 mutex_exit(&dev
->l2ad_mtx
);
7088 vdev_space_update(dev
->l2ad_vdev
, -bytes_dropped
, 0, 0);
7090 l2arc_do_free_on_write();
7092 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
7096 * A read to a cache device completed. Validate buffer contents before
7097 * handing over to the regular ARC routines.
7100 l2arc_read_done(zio_t
*zio
)
7102 l2arc_read_callback_t
*cb
;
7104 kmutex_t
*hash_lock
;
7105 boolean_t valid_cksum
;
7107 ASSERT3P(zio
->io_vd
, !=, NULL
);
7108 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
7110 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
7112 cb
= zio
->io_private
;
7113 ASSERT3P(cb
, !=, NULL
);
7114 hdr
= cb
->l2rcb_hdr
;
7115 ASSERT3P(hdr
, !=, NULL
);
7117 hash_lock
= HDR_LOCK(hdr
);
7118 mutex_enter(hash_lock
);
7119 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
7122 * If the data was read into a temporary buffer,
7123 * move it and free the buffer.
7125 if (cb
->l2rcb_abd
!= NULL
) {
7126 ASSERT3U(arc_hdr_size(hdr
), <, zio
->io_size
);
7127 if (zio
->io_error
== 0) {
7128 abd_copy(hdr
->b_l1hdr
.b_pabd
, cb
->l2rcb_abd
,
7133 * The following must be done regardless of whether
7134 * there was an error:
7135 * - free the temporary buffer
7136 * - point zio to the real ARC buffer
7137 * - set zio size accordingly
7138 * These are required because zio is either re-used for
7139 * an I/O of the block in the case of the error
7140 * or the zio is passed to arc_read_done() and it
7143 abd_free(cb
->l2rcb_abd
);
7144 zio
->io_size
= zio
->io_orig_size
= arc_hdr_size(hdr
);
7145 zio
->io_abd
= zio
->io_orig_abd
= hdr
->b_l1hdr
.b_pabd
;
7148 ASSERT3P(zio
->io_abd
, !=, NULL
);
7151 * Check this survived the L2ARC journey.
7153 ASSERT3P(zio
->io_abd
, ==, hdr
->b_l1hdr
.b_pabd
);
7154 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
7155 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
7157 valid_cksum
= arc_cksum_is_equal(hdr
, zio
);
7158 if (valid_cksum
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
7159 mutex_exit(hash_lock
);
7160 zio
->io_private
= hdr
;
7163 mutex_exit(hash_lock
);
7165 * Buffer didn't survive caching. Increment stats and
7166 * reissue to the original storage device.
7168 if (zio
->io_error
!= 0) {
7169 ARCSTAT_BUMP(arcstat_l2_io_error
);
7171 zio
->io_error
= SET_ERROR(EIO
);
7174 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
7177 * If there's no waiter, issue an async i/o to the primary
7178 * storage now. If there *is* a waiter, the caller must
7179 * issue the i/o in a context where it's OK to block.
7181 if (zio
->io_waiter
== NULL
) {
7182 zio_t
*pio
= zio_unique_parent(zio
);
7184 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
7186 zio_nowait(zio_read(pio
, zio
->io_spa
, zio
->io_bp
,
7187 hdr
->b_l1hdr
.b_pabd
, zio
->io_size
, arc_read_done
,
7188 hdr
, zio
->io_priority
, cb
->l2rcb_flags
,
7193 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
7197 * This is the list priority from which the L2ARC will search for pages to
7198 * cache. This is used within loops (0..3) to cycle through lists in the
7199 * desired order. This order can have a significant effect on cache
7202 * Currently the metadata lists are hit first, MFU then MRU, followed by
7203 * the data lists. This function returns a locked list, and also returns
7206 static multilist_sublist_t
*
7207 l2arc_sublist_lock(int list_num
)
7209 multilist_t
*ml
= NULL
;
7212 ASSERT(list_num
>= 0 && list_num
< L2ARC_FEED_TYPES
);
7216 ml
= arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
7219 ml
= arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
7222 ml
= arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
7225 ml
= arc_mru
->arcs_list
[ARC_BUFC_DATA
];
7232 * Return a randomly-selected sublist. This is acceptable
7233 * because the caller feeds only a little bit of data for each
7234 * call (8MB). Subsequent calls will result in different
7235 * sublists being selected.
7237 idx
= multilist_get_random_index(ml
);
7238 return (multilist_sublist_lock(ml
, idx
));
7242 * Evict buffers from the device write hand to the distance specified in
7243 * bytes. This distance may span populated buffers, it may span nothing.
7244 * This is clearing a region on the L2ARC device ready for writing.
7245 * If the 'all' boolean is set, every buffer is evicted.
7248 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
7251 arc_buf_hdr_t
*hdr
, *hdr_prev
;
7252 kmutex_t
*hash_lock
;
7255 buflist
= &dev
->l2ad_buflist
;
7257 if (!all
&& dev
->l2ad_first
) {
7259 * This is the first sweep through the device. There is
7265 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
7267 * When nearing the end of the device, evict to the end
7268 * before the device write hand jumps to the start.
7270 taddr
= dev
->l2ad_end
;
7272 taddr
= dev
->l2ad_hand
+ distance
;
7274 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
7275 uint64_t, taddr
, boolean_t
, all
);
7278 mutex_enter(&dev
->l2ad_mtx
);
7279 for (hdr
= list_tail(buflist
); hdr
; hdr
= hdr_prev
) {
7280 hdr_prev
= list_prev(buflist
, hdr
);
7282 hash_lock
= HDR_LOCK(hdr
);
7285 * We cannot use mutex_enter or else we can deadlock
7286 * with l2arc_write_buffers (due to swapping the order
7287 * the hash lock and l2ad_mtx are taken).
7289 if (!mutex_tryenter(hash_lock
)) {
7291 * Missed the hash lock. Retry.
7293 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
7294 mutex_exit(&dev
->l2ad_mtx
);
7295 mutex_enter(hash_lock
);
7296 mutex_exit(hash_lock
);
7300 if (HDR_L2_WRITE_HEAD(hdr
)) {
7302 * We hit a write head node. Leave it for
7303 * l2arc_write_done().
7305 list_remove(buflist
, hdr
);
7306 mutex_exit(hash_lock
);
7310 if (!all
&& HDR_HAS_L2HDR(hdr
) &&
7311 (hdr
->b_l2hdr
.b_daddr
> taddr
||
7312 hdr
->b_l2hdr
.b_daddr
< dev
->l2ad_hand
)) {
7314 * We've evicted to the target address,
7315 * or the end of the device.
7317 mutex_exit(hash_lock
);
7321 ASSERT(HDR_HAS_L2HDR(hdr
));
7322 if (!HDR_HAS_L1HDR(hdr
)) {
7323 ASSERT(!HDR_L2_READING(hdr
));
7325 * This doesn't exist in the ARC. Destroy.
7326 * arc_hdr_destroy() will call list_remove()
7327 * and decrement arcstat_l2_lsize.
7329 arc_change_state(arc_anon
, hdr
, hash_lock
);
7330 arc_hdr_destroy(hdr
);
7332 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_l2c_only
);
7333 ARCSTAT_BUMP(arcstat_l2_evict_l1cached
);
7335 * Invalidate issued or about to be issued
7336 * reads, since we may be about to write
7337 * over this location.
7339 if (HDR_L2_READING(hdr
)) {
7340 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
7341 arc_hdr_set_flags(hdr
, ARC_FLAG_L2_EVICTED
);
7344 /* Ensure this header has finished being written */
7345 ASSERT(!HDR_L2_WRITING(hdr
));
7347 arc_hdr_l2hdr_destroy(hdr
);
7349 mutex_exit(hash_lock
);
7351 mutex_exit(&dev
->l2ad_mtx
);
7355 * Find and write ARC buffers to the L2ARC device.
7357 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7358 * for reading until they have completed writing.
7359 * The headroom_boost is an in-out parameter used to maintain headroom boost
7360 * state between calls to this function.
7362 * Returns the number of bytes actually written (which may be smaller than
7363 * the delta by which the device hand has changed due to alignment).
7366 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
)
7368 arc_buf_hdr_t
*hdr
, *hdr_prev
, *head
;
7369 uint64_t write_asize
, write_psize
, write_lsize
, headroom
;
7371 l2arc_write_callback_t
*cb
;
7373 uint64_t guid
= spa_load_guid(spa
);
7376 ASSERT3P(dev
->l2ad_vdev
, !=, NULL
);
7379 write_lsize
= write_asize
= write_psize
= 0;
7381 head
= kmem_cache_alloc(hdr_l2only_cache
, KM_PUSHPAGE
);
7382 arc_hdr_set_flags(head
, ARC_FLAG_L2_WRITE_HEAD
| ARC_FLAG_HAS_L2HDR
);
7385 * Copy buffers for L2ARC writing.
7387 for (try = 0; try < L2ARC_FEED_TYPES
; try++) {
7388 multilist_sublist_t
*mls
= l2arc_sublist_lock(try);
7389 uint64_t passed_sz
= 0;
7391 VERIFY3P(mls
, !=, NULL
);
7394 * L2ARC fast warmup.
7396 * Until the ARC is warm and starts to evict, read from the
7397 * head of the ARC lists rather than the tail.
7399 if (arc_warm
== B_FALSE
)
7400 hdr
= multilist_sublist_head(mls
);
7402 hdr
= multilist_sublist_tail(mls
);
7404 headroom
= target_sz
* l2arc_headroom
;
7405 if (zfs_compressed_arc_enabled
)
7406 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
7408 for (; hdr
; hdr
= hdr_prev
) {
7409 kmutex_t
*hash_lock
;
7411 if (arc_warm
== B_FALSE
)
7412 hdr_prev
= multilist_sublist_next(mls
, hdr
);
7414 hdr_prev
= multilist_sublist_prev(mls
, hdr
);
7416 hash_lock
= HDR_LOCK(hdr
);
7417 if (!mutex_tryenter(hash_lock
)) {
7419 * Skip this buffer rather than waiting.
7424 passed_sz
+= HDR_GET_LSIZE(hdr
);
7425 if (passed_sz
> headroom
) {
7429 mutex_exit(hash_lock
);
7433 if (!l2arc_write_eligible(guid
, hdr
)) {
7434 mutex_exit(hash_lock
);
7439 * We rely on the L1 portion of the header below, so
7440 * it's invalid for this header to have been evicted out
7441 * of the ghost cache, prior to being written out. The
7442 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7444 ASSERT(HDR_HAS_L1HDR(hdr
));
7446 ASSERT3U(HDR_GET_PSIZE(hdr
), >, 0);
7447 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
7448 ASSERT3U(arc_hdr_size(hdr
), >, 0);
7449 uint64_t psize
= arc_hdr_size(hdr
);
7450 uint64_t asize
= vdev_psize_to_asize(dev
->l2ad_vdev
,
7453 if ((write_asize
+ asize
) > target_sz
) {
7455 mutex_exit(hash_lock
);
7461 * Insert a dummy header on the buflist so
7462 * l2arc_write_done() can find where the
7463 * write buffers begin without searching.
7465 mutex_enter(&dev
->l2ad_mtx
);
7466 list_insert_head(&dev
->l2ad_buflist
, head
);
7467 mutex_exit(&dev
->l2ad_mtx
);
7470 sizeof (l2arc_write_callback_t
), KM_SLEEP
);
7471 cb
->l2wcb_dev
= dev
;
7472 cb
->l2wcb_head
= head
;
7473 pio
= zio_root(spa
, l2arc_write_done
, cb
,
7477 hdr
->b_l2hdr
.b_dev
= dev
;
7478 hdr
->b_l2hdr
.b_hits
= 0;
7480 hdr
->b_l2hdr
.b_daddr
= dev
->l2ad_hand
;
7481 arc_hdr_set_flags(hdr
,
7482 ARC_FLAG_L2_WRITING
| ARC_FLAG_HAS_L2HDR
);
7484 mutex_enter(&dev
->l2ad_mtx
);
7485 list_insert_head(&dev
->l2ad_buflist
, hdr
);
7486 mutex_exit(&dev
->l2ad_mtx
);
7488 (void) refcount_add_many(&dev
->l2ad_alloc
, psize
, hdr
);
7491 * Normally the L2ARC can use the hdr's data, but if
7492 * we're sharing data between the hdr and one of its
7493 * bufs, L2ARC needs its own copy of the data so that
7494 * the ZIO below can't race with the buf consumer.
7495 * Another case where we need to create a copy of the
7496 * data is when the buffer size is not device-aligned
7497 * and we need to pad the block to make it such.
7498 * That also keeps the clock hand suitably aligned.
7500 * To ensure that the copy will be available for the
7501 * lifetime of the ZIO and be cleaned up afterwards, we
7502 * add it to the l2arc_free_on_write queue.
7505 if (!HDR_SHARED_DATA(hdr
) && psize
== asize
) {
7506 to_write
= hdr
->b_l1hdr
.b_pabd
;
7508 to_write
= abd_alloc_for_io(asize
,
7509 HDR_ISTYPE_METADATA(hdr
));
7510 abd_copy(to_write
, hdr
->b_l1hdr
.b_pabd
, psize
);
7511 if (asize
!= psize
) {
7512 abd_zero_off(to_write
, psize
,
7515 l2arc_free_abd_on_write(to_write
, asize
,
7518 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
7519 hdr
->b_l2hdr
.b_daddr
, asize
, to_write
,
7520 ZIO_CHECKSUM_OFF
, NULL
, hdr
,
7521 ZIO_PRIORITY_ASYNC_WRITE
,
7522 ZIO_FLAG_CANFAIL
, B_FALSE
);
7524 write_lsize
+= HDR_GET_LSIZE(hdr
);
7525 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
7528 write_psize
+= psize
;
7529 write_asize
+= asize
;
7530 dev
->l2ad_hand
+= asize
;
7532 mutex_exit(hash_lock
);
7534 (void) zio_nowait(wzio
);
7537 multilist_sublist_unlock(mls
);
7543 /* No buffers selected for writing? */
7545 ASSERT0(write_lsize
);
7546 ASSERT(!HDR_HAS_L1HDR(head
));
7547 kmem_cache_free(hdr_l2only_cache
, head
);
7551 ASSERT3U(write_asize
, <=, target_sz
);
7552 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
7553 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_psize
);
7554 ARCSTAT_INCR(arcstat_l2_lsize
, write_lsize
);
7555 ARCSTAT_INCR(arcstat_l2_psize
, write_psize
);
7556 vdev_space_update(dev
->l2ad_vdev
, write_psize
, 0, 0);
7559 * Bump device hand to the device start if it is approaching the end.
7560 * l2arc_evict() will already have evicted ahead for this case.
7562 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
7563 dev
->l2ad_hand
= dev
->l2ad_start
;
7564 dev
->l2ad_first
= B_FALSE
;
7567 dev
->l2ad_writing
= B_TRUE
;
7568 (void) zio_wait(pio
);
7569 dev
->l2ad_writing
= B_FALSE
;
7571 return (write_asize
);
7575 * This thread feeds the L2ARC at regular intervals. This is the beating
7576 * heart of the L2ARC.
7579 l2arc_feed_thread(void)
7584 uint64_t size
, wrote
;
7585 clock_t begin
, next
= ddi_get_lbolt();
7586 fstrans_cookie_t cookie
;
7588 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
7590 mutex_enter(&l2arc_feed_thr_lock
);
7592 cookie
= spl_fstrans_mark();
7593 while (l2arc_thread_exit
== 0) {
7594 CALLB_CPR_SAFE_BEGIN(&cpr
);
7595 (void) cv_timedwait_sig(&l2arc_feed_thr_cv
,
7596 &l2arc_feed_thr_lock
, next
);
7597 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
7598 next
= ddi_get_lbolt() + hz
;
7601 * Quick check for L2ARC devices.
7603 mutex_enter(&l2arc_dev_mtx
);
7604 if (l2arc_ndev
== 0) {
7605 mutex_exit(&l2arc_dev_mtx
);
7608 mutex_exit(&l2arc_dev_mtx
);
7609 begin
= ddi_get_lbolt();
7612 * This selects the next l2arc device to write to, and in
7613 * doing so the next spa to feed from: dev->l2ad_spa. This
7614 * will return NULL if there are now no l2arc devices or if
7615 * they are all faulted.
7617 * If a device is returned, its spa's config lock is also
7618 * held to prevent device removal. l2arc_dev_get_next()
7619 * will grab and release l2arc_dev_mtx.
7621 if ((dev
= l2arc_dev_get_next()) == NULL
)
7624 spa
= dev
->l2ad_spa
;
7625 ASSERT3P(spa
, !=, NULL
);
7628 * If the pool is read-only then force the feed thread to
7629 * sleep a little longer.
7631 if (!spa_writeable(spa
)) {
7632 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
7633 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7638 * Avoid contributing to memory pressure.
7640 if (arc_reclaim_needed()) {
7641 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
7642 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7646 ARCSTAT_BUMP(arcstat_l2_feeds
);
7648 size
= l2arc_write_size();
7651 * Evict L2ARC buffers that will be overwritten.
7653 l2arc_evict(dev
, size
, B_FALSE
);
7656 * Write ARC buffers.
7658 wrote
= l2arc_write_buffers(spa
, dev
, size
);
7661 * Calculate interval between writes.
7663 next
= l2arc_write_interval(begin
, size
, wrote
);
7664 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7666 spl_fstrans_unmark(cookie
);
7668 l2arc_thread_exit
= 0;
7669 cv_broadcast(&l2arc_feed_thr_cv
);
7670 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
7675 l2arc_vdev_present(vdev_t
*vd
)
7679 mutex_enter(&l2arc_dev_mtx
);
7680 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
7681 dev
= list_next(l2arc_dev_list
, dev
)) {
7682 if (dev
->l2ad_vdev
== vd
)
7685 mutex_exit(&l2arc_dev_mtx
);
7687 return (dev
!= NULL
);
7691 * Add a vdev for use by the L2ARC. By this point the spa has already
7692 * validated the vdev and opened it.
7695 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
7697 l2arc_dev_t
*adddev
;
7699 ASSERT(!l2arc_vdev_present(vd
));
7702 * Create a new l2arc device entry.
7704 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
7705 adddev
->l2ad_spa
= spa
;
7706 adddev
->l2ad_vdev
= vd
;
7707 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
7708 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
7709 adddev
->l2ad_hand
= adddev
->l2ad_start
;
7710 adddev
->l2ad_first
= B_TRUE
;
7711 adddev
->l2ad_writing
= B_FALSE
;
7712 list_link_init(&adddev
->l2ad_node
);
7714 mutex_init(&adddev
->l2ad_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7716 * This is a list of all ARC buffers that are still valid on the
7719 list_create(&adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
7720 offsetof(arc_buf_hdr_t
, b_l2hdr
.b_l2node
));
7722 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
7723 refcount_create(&adddev
->l2ad_alloc
);
7726 * Add device to global list
7728 mutex_enter(&l2arc_dev_mtx
);
7729 list_insert_head(l2arc_dev_list
, adddev
);
7730 atomic_inc_64(&l2arc_ndev
);
7731 mutex_exit(&l2arc_dev_mtx
);
7735 * Remove a vdev from the L2ARC.
7738 l2arc_remove_vdev(vdev_t
*vd
)
7740 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
7743 * Find the device by vdev
7745 mutex_enter(&l2arc_dev_mtx
);
7746 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
7747 nextdev
= list_next(l2arc_dev_list
, dev
);
7748 if (vd
== dev
->l2ad_vdev
) {
7753 ASSERT3P(remdev
, !=, NULL
);
7756 * Remove device from global list
7758 list_remove(l2arc_dev_list
, remdev
);
7759 l2arc_dev_last
= NULL
; /* may have been invalidated */
7760 atomic_dec_64(&l2arc_ndev
);
7761 mutex_exit(&l2arc_dev_mtx
);
7764 * Clear all buflists and ARC references. L2ARC device flush.
7766 l2arc_evict(remdev
, 0, B_TRUE
);
7767 list_destroy(&remdev
->l2ad_buflist
);
7768 mutex_destroy(&remdev
->l2ad_mtx
);
7769 refcount_destroy(&remdev
->l2ad_alloc
);
7770 kmem_free(remdev
, sizeof (l2arc_dev_t
));
7776 l2arc_thread_exit
= 0;
7778 l2arc_writes_sent
= 0;
7779 l2arc_writes_done
= 0;
7781 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
7782 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
7783 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7784 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7786 l2arc_dev_list
= &L2ARC_dev_list
;
7787 l2arc_free_on_write
= &L2ARC_free_on_write
;
7788 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
7789 offsetof(l2arc_dev_t
, l2ad_node
));
7790 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
7791 offsetof(l2arc_data_free_t
, l2df_list_node
));
7798 * This is called from dmu_fini(), which is called from spa_fini();
7799 * Because of this, we can assume that all l2arc devices have
7800 * already been removed when the pools themselves were removed.
7803 l2arc_do_free_on_write();
7805 mutex_destroy(&l2arc_feed_thr_lock
);
7806 cv_destroy(&l2arc_feed_thr_cv
);
7807 mutex_destroy(&l2arc_dev_mtx
);
7808 mutex_destroy(&l2arc_free_on_write_mtx
);
7810 list_destroy(l2arc_dev_list
);
7811 list_destroy(l2arc_free_on_write
);
7817 if (!(spa_mode_global
& FWRITE
))
7820 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
7821 TS_RUN
, defclsyspri
);
7827 if (!(spa_mode_global
& FWRITE
))
7830 mutex_enter(&l2arc_feed_thr_lock
);
7831 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
7832 l2arc_thread_exit
= 1;
7833 while (l2arc_thread_exit
!= 0)
7834 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
7835 mutex_exit(&l2arc_feed_thr_lock
);
7838 #if defined(_KERNEL) && defined(HAVE_SPL)
7839 EXPORT_SYMBOL(arc_buf_size
);
7840 EXPORT_SYMBOL(arc_write
);
7841 EXPORT_SYMBOL(arc_read
);
7842 EXPORT_SYMBOL(arc_buf_info
);
7843 EXPORT_SYMBOL(arc_getbuf_func
);
7844 EXPORT_SYMBOL(arc_add_prune_callback
);
7845 EXPORT_SYMBOL(arc_remove_prune_callback
);
7848 module_param(zfs_arc_min
, ulong
, 0644);
7849 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
7851 module_param(zfs_arc_max
, ulong
, 0644);
7852 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
7854 module_param(zfs_arc_meta_limit
, ulong
, 0644);
7855 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
7857 module_param(zfs_arc_meta_limit_percent
, ulong
, 0644);
7858 MODULE_PARM_DESC(zfs_arc_meta_limit_percent
,
7859 "Percent of arc size for arc meta limit");
7861 module_param(zfs_arc_meta_min
, ulong
, 0644);
7862 MODULE_PARM_DESC(zfs_arc_meta_min
, "Min arc metadata");
7864 module_param(zfs_arc_meta_prune
, int, 0644);
7865 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Meta objects to scan for prune");
7867 module_param(zfs_arc_meta_adjust_restarts
, int, 0644);
7868 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts
,
7869 "Limit number of restarts in arc_adjust_meta");
7871 module_param(zfs_arc_meta_strategy
, int, 0644);
7872 MODULE_PARM_DESC(zfs_arc_meta_strategy
, "Meta reclaim strategy");
7874 module_param(zfs_arc_grow_retry
, int, 0644);
7875 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
7877 module_param(zfs_arc_p_aggressive_disable
, int, 0644);
7878 MODULE_PARM_DESC(zfs_arc_p_aggressive_disable
, "disable aggressive arc_p grow");
7880 module_param(zfs_arc_p_dampener_disable
, int, 0644);
7881 MODULE_PARM_DESC(zfs_arc_p_dampener_disable
, "disable arc_p adapt dampener");
7883 module_param(zfs_arc_shrink_shift
, int, 0644);
7884 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
7886 module_param(zfs_arc_pc_percent
, uint
, 0644);
7887 MODULE_PARM_DESC(zfs_arc_pc_percent
,
7888 "Percent of pagecache to reclaim arc to");
7890 module_param(zfs_arc_p_min_shift
, int, 0644);
7891 MODULE_PARM_DESC(zfs_arc_p_min_shift
, "arc_c shift to calc min/max arc_p");
7893 module_param(zfs_arc_average_blocksize
, int, 0444);
7894 MODULE_PARM_DESC(zfs_arc_average_blocksize
, "Target average block size");
7896 module_param(zfs_compressed_arc_enabled
, int, 0644);
7897 MODULE_PARM_DESC(zfs_compressed_arc_enabled
, "Disable compressed arc buffers");
7899 module_param(zfs_arc_min_prefetch_lifespan
, int, 0644);
7900 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan
, "Min life of prefetch block");
7902 module_param(l2arc_write_max
, ulong
, 0644);
7903 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
7905 module_param(l2arc_write_boost
, ulong
, 0644);
7906 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
7908 module_param(l2arc_headroom
, ulong
, 0644);
7909 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
7911 module_param(l2arc_headroom_boost
, ulong
, 0644);
7912 MODULE_PARM_DESC(l2arc_headroom_boost
, "Compressed l2arc_headroom multiplier");
7914 module_param(l2arc_feed_secs
, ulong
, 0644);
7915 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
7917 module_param(l2arc_feed_min_ms
, ulong
, 0644);
7918 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
7920 module_param(l2arc_noprefetch
, int, 0644);
7921 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
7923 module_param(l2arc_feed_again
, int, 0644);
7924 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
7926 module_param(l2arc_norw
, int, 0644);
7927 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");
7929 module_param(zfs_arc_lotsfree_percent
, int, 0644);
7930 MODULE_PARM_DESC(zfs_arc_lotsfree_percent
,
7931 "System free memory I/O throttle in bytes");
7933 module_param(zfs_arc_sys_free
, ulong
, 0644);
7934 MODULE_PARM_DESC(zfs_arc_sys_free
, "System free memory target size in bytes");
7936 module_param(zfs_arc_dnode_limit
, ulong
, 0644);
7937 MODULE_PARM_DESC(zfs_arc_dnode_limit
, "Minimum bytes of dnodes in arc");
7939 module_param(zfs_arc_dnode_limit_percent
, ulong
, 0644);
7940 MODULE_PARM_DESC(zfs_arc_dnode_limit_percent
,
7941 "Percent of ARC meta buffers for dnodes");
7943 module_param(zfs_arc_dnode_reduce_percent
, ulong
, 0644);
7944 MODULE_PARM_DESC(zfs_arc_dnode_reduce_percent
,
7945 "Percentage of excess dnodes to try to unpin");