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>
283 #include <linux/page_compat.h>
285 #include <sys/callb.h>
286 #include <sys/kstat.h>
287 #include <sys/dmu_tx.h>
288 #include <zfs_fletcher.h>
289 #include <sys/arc_impl.h>
290 #include <sys/trace_arc.h>
293 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
294 boolean_t arc_watch
= B_FALSE
;
297 static kmutex_t arc_reclaim_lock
;
298 static kcondvar_t arc_reclaim_thread_cv
;
299 static boolean_t arc_reclaim_thread_exit
;
300 static kcondvar_t arc_reclaim_waiters_cv
;
303 * The number of headers to evict in arc_evict_state_impl() before
304 * dropping the sublist lock and evicting from another sublist. A lower
305 * value means we're more likely to evict the "correct" header (i.e. the
306 * oldest header in the arc state), but comes with higher overhead
307 * (i.e. more invocations of arc_evict_state_impl()).
309 int zfs_arc_evict_batch_limit
= 10;
311 /* number of seconds before growing cache again */
312 static int arc_grow_retry
= 5;
314 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
315 int zfs_arc_overflow_shift
= 8;
317 /* shift of arc_c for calculating both min and max arc_p */
318 static int arc_p_min_shift
= 4;
320 /* log2(fraction of arc to reclaim) */
321 static int arc_shrink_shift
= 7;
323 /* percent of pagecache to reclaim arc to */
325 static uint_t zfs_arc_pc_percent
= 0;
329 * log2(fraction of ARC which must be free to allow growing).
330 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
331 * when reading a new block into the ARC, we will evict an equal-sized block
334 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
335 * we will still not allow it to grow.
337 int arc_no_grow_shift
= 5;
341 * minimum lifespan of a prefetch block in clock ticks
342 * (initialized in arc_init())
344 static int arc_min_prefetch_lifespan
;
347 * If this percent of memory is free, don't throttle.
349 int arc_lotsfree_percent
= 10;
354 * The arc has filled available memory and has now warmed up.
356 static boolean_t arc_warm
;
359 * log2 fraction of the zio arena to keep free.
361 int arc_zio_arena_free_shift
= 2;
364 * These tunables are for performance analysis.
366 unsigned long zfs_arc_max
= 0;
367 unsigned long zfs_arc_min
= 0;
368 unsigned long zfs_arc_meta_limit
= 0;
369 unsigned long zfs_arc_meta_min
= 0;
370 unsigned long zfs_arc_dnode_limit
= 0;
371 unsigned long zfs_arc_dnode_reduce_percent
= 10;
372 int zfs_arc_grow_retry
= 0;
373 int zfs_arc_shrink_shift
= 0;
374 int zfs_arc_p_min_shift
= 0;
375 int zfs_arc_average_blocksize
= 8 * 1024; /* 8KB */
377 int zfs_compressed_arc_enabled
= B_TRUE
;
380 * ARC will evict meta buffers that exceed arc_meta_limit. This
381 * tunable make arc_meta_limit adjustable for different workloads.
383 unsigned long zfs_arc_meta_limit_percent
= 75;
386 * Percentage that can be consumed by dnodes of ARC meta buffers.
388 unsigned long zfs_arc_dnode_limit_percent
= 10;
391 * These tunables are Linux specific
393 unsigned long zfs_arc_sys_free
= 0;
394 int zfs_arc_min_prefetch_lifespan
= 0;
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 when updating the access state due to the
434 * header having already been released after acquiring the hash lock.
436 kstat_named_t arcstat_access_skip
;
438 * Number of buffers skipped because they have I/O in progress, are
439 * indirect prefetch buffers that have not lived long enough, or are
440 * not from the spa we're trying to evict from.
442 kstat_named_t arcstat_evict_skip
;
444 * Number of times arc_evict_state() was unable to evict enough
445 * buffers to reach its target amount.
447 kstat_named_t arcstat_evict_not_enough
;
448 kstat_named_t arcstat_evict_l2_cached
;
449 kstat_named_t arcstat_evict_l2_eligible
;
450 kstat_named_t arcstat_evict_l2_ineligible
;
451 kstat_named_t arcstat_evict_l2_skip
;
452 kstat_named_t arcstat_hash_elements
;
453 kstat_named_t arcstat_hash_elements_max
;
454 kstat_named_t arcstat_hash_collisions
;
455 kstat_named_t arcstat_hash_chains
;
456 kstat_named_t arcstat_hash_chain_max
;
457 kstat_named_t arcstat_p
;
458 kstat_named_t arcstat_c
;
459 kstat_named_t arcstat_c_min
;
460 kstat_named_t arcstat_c_max
;
461 kstat_named_t arcstat_size
;
463 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
464 * Note that the compressed bytes may match the uncompressed bytes
465 * if the block is either not compressed or compressed arc is disabled.
467 kstat_named_t arcstat_compressed_size
;
469 * Uncompressed size of the data stored in b_pabd. If compressed
470 * arc is disabled then this value will be identical to the stat
473 kstat_named_t arcstat_uncompressed_size
;
475 * Number of bytes stored in all the arc_buf_t's. This is classified
476 * as "overhead" since this data is typically short-lived and will
477 * be evicted from the arc when it becomes unreferenced unless the
478 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
479 * values have been set (see comment in dbuf.c for more information).
481 kstat_named_t arcstat_overhead_size
;
483 * Number of bytes consumed by internal ARC structures necessary
484 * for tracking purposes; these structures are not actually
485 * backed by ARC buffers. This includes arc_buf_hdr_t structures
486 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
487 * caches), and arc_buf_t structures (allocated via arc_buf_t
490 kstat_named_t arcstat_hdr_size
;
492 * Number of bytes consumed by ARC buffers of type equal to
493 * ARC_BUFC_DATA. This is generally consumed by buffers backing
494 * on disk user data (e.g. plain file contents).
496 kstat_named_t arcstat_data_size
;
498 * Number of bytes consumed by ARC buffers of type equal to
499 * ARC_BUFC_METADATA. This is generally consumed by buffers
500 * backing on disk data that is used for internal ZFS
501 * structures (e.g. ZAP, dnode, indirect blocks, etc).
503 kstat_named_t arcstat_metadata_size
;
505 * Number of bytes consumed by dmu_buf_impl_t objects.
507 kstat_named_t arcstat_dbuf_size
;
509 * Number of bytes consumed by dnode_t objects.
511 kstat_named_t arcstat_dnode_size
;
513 * Number of bytes consumed by bonus buffers.
515 kstat_named_t arcstat_bonus_size
;
517 * Total number of bytes consumed by ARC buffers residing in the
518 * arc_anon state. This includes *all* buffers in the arc_anon
519 * state; e.g. data, metadata, evictable, and unevictable buffers
520 * are all included in this value.
522 kstat_named_t arcstat_anon_size
;
524 * Number of bytes consumed by ARC buffers that meet the
525 * following criteria: backing buffers of type ARC_BUFC_DATA,
526 * residing in the arc_anon state, and are eligible for eviction
527 * (e.g. have no outstanding holds on the buffer).
529 kstat_named_t arcstat_anon_evictable_data
;
531 * Number of bytes consumed by ARC buffers that meet the
532 * following criteria: backing buffers of type ARC_BUFC_METADATA,
533 * residing in the arc_anon state, and are eligible for eviction
534 * (e.g. have no outstanding holds on the buffer).
536 kstat_named_t arcstat_anon_evictable_metadata
;
538 * Total number of bytes consumed by ARC buffers residing in the
539 * arc_mru state. This includes *all* buffers in the arc_mru
540 * state; e.g. data, metadata, evictable, and unevictable buffers
541 * are all included in this value.
543 kstat_named_t arcstat_mru_size
;
545 * Number of bytes consumed by ARC buffers that meet the
546 * following criteria: backing buffers of type ARC_BUFC_DATA,
547 * residing in the arc_mru state, and are eligible for eviction
548 * (e.g. have no outstanding holds on the buffer).
550 kstat_named_t arcstat_mru_evictable_data
;
552 * Number of bytes consumed by ARC buffers that meet the
553 * following criteria: backing buffers of type ARC_BUFC_METADATA,
554 * residing in the arc_mru state, and are eligible for eviction
555 * (e.g. have no outstanding holds on the buffer).
557 kstat_named_t arcstat_mru_evictable_metadata
;
559 * Total number of bytes that *would have been* consumed by ARC
560 * buffers in the arc_mru_ghost state. The key thing to note
561 * here, is the fact that this size doesn't actually indicate
562 * RAM consumption. The ghost lists only consist of headers and
563 * don't actually have ARC buffers linked off of these headers.
564 * Thus, *if* the headers had associated ARC buffers, these
565 * buffers *would have* consumed this number of bytes.
567 kstat_named_t arcstat_mru_ghost_size
;
569 * Number of bytes that *would have been* consumed by ARC
570 * buffers that are eligible for eviction, of type
571 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
573 kstat_named_t arcstat_mru_ghost_evictable_data
;
575 * Number of bytes that *would have been* consumed by ARC
576 * buffers that are eligible for eviction, of type
577 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
579 kstat_named_t arcstat_mru_ghost_evictable_metadata
;
581 * Total number of bytes consumed by ARC buffers residing in the
582 * arc_mfu state. This includes *all* buffers in the arc_mfu
583 * state; e.g. data, metadata, evictable, and unevictable buffers
584 * are all included in this value.
586 kstat_named_t arcstat_mfu_size
;
588 * Number of bytes consumed by ARC buffers that are eligible for
589 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
592 kstat_named_t arcstat_mfu_evictable_data
;
594 * Number of bytes consumed by ARC buffers that are eligible for
595 * eviction, of type ARC_BUFC_METADATA, and reside in the
598 kstat_named_t arcstat_mfu_evictable_metadata
;
600 * Total number of bytes that *would have been* consumed by ARC
601 * buffers in the arc_mfu_ghost state. See the comment above
602 * arcstat_mru_ghost_size for more details.
604 kstat_named_t arcstat_mfu_ghost_size
;
606 * Number of bytes that *would have been* consumed by ARC
607 * buffers that are eligible for eviction, of type
608 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
610 kstat_named_t arcstat_mfu_ghost_evictable_data
;
612 * Number of bytes that *would have been* consumed by ARC
613 * buffers that are eligible for eviction, of type
614 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
616 kstat_named_t arcstat_mfu_ghost_evictable_metadata
;
617 kstat_named_t arcstat_l2_hits
;
618 kstat_named_t arcstat_l2_misses
;
619 kstat_named_t arcstat_l2_feeds
;
620 kstat_named_t arcstat_l2_rw_clash
;
621 kstat_named_t arcstat_l2_read_bytes
;
622 kstat_named_t arcstat_l2_write_bytes
;
623 kstat_named_t arcstat_l2_writes_sent
;
624 kstat_named_t arcstat_l2_writes_done
;
625 kstat_named_t arcstat_l2_writes_error
;
626 kstat_named_t arcstat_l2_writes_lock_retry
;
627 kstat_named_t arcstat_l2_evict_lock_retry
;
628 kstat_named_t arcstat_l2_evict_reading
;
629 kstat_named_t arcstat_l2_evict_l1cached
;
630 kstat_named_t arcstat_l2_free_on_write
;
631 kstat_named_t arcstat_l2_abort_lowmem
;
632 kstat_named_t arcstat_l2_cksum_bad
;
633 kstat_named_t arcstat_l2_io_error
;
634 kstat_named_t arcstat_l2_lsize
;
635 kstat_named_t arcstat_l2_psize
;
636 kstat_named_t arcstat_l2_hdr_size
;
637 kstat_named_t arcstat_memory_throttle_count
;
638 kstat_named_t arcstat_memory_direct_count
;
639 kstat_named_t arcstat_memory_indirect_count
;
640 kstat_named_t arcstat_memory_all_bytes
;
641 kstat_named_t arcstat_memory_free_bytes
;
642 kstat_named_t arcstat_memory_available_bytes
;
643 kstat_named_t arcstat_no_grow
;
644 kstat_named_t arcstat_tempreserve
;
645 kstat_named_t arcstat_loaned_bytes
;
646 kstat_named_t arcstat_prune
;
647 kstat_named_t arcstat_meta_used
;
648 kstat_named_t arcstat_meta_limit
;
649 kstat_named_t arcstat_dnode_limit
;
650 kstat_named_t arcstat_meta_max
;
651 kstat_named_t arcstat_meta_min
;
652 kstat_named_t arcstat_sync_wait_for_async
;
653 kstat_named_t arcstat_demand_hit_predictive_prefetch
;
654 kstat_named_t arcstat_need_free
;
655 kstat_named_t arcstat_sys_free
;
658 static arc_stats_t arc_stats
= {
659 { "hits", KSTAT_DATA_UINT64
},
660 { "misses", KSTAT_DATA_UINT64
},
661 { "demand_data_hits", KSTAT_DATA_UINT64
},
662 { "demand_data_misses", KSTAT_DATA_UINT64
},
663 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
664 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
665 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
666 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
667 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
668 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
669 { "mru_hits", KSTAT_DATA_UINT64
},
670 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
671 { "mfu_hits", KSTAT_DATA_UINT64
},
672 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
673 { "deleted", KSTAT_DATA_UINT64
},
674 { "mutex_miss", KSTAT_DATA_UINT64
},
675 { "access_skip", KSTAT_DATA_UINT64
},
676 { "evict_skip", KSTAT_DATA_UINT64
},
677 { "evict_not_enough", KSTAT_DATA_UINT64
},
678 { "evict_l2_cached", KSTAT_DATA_UINT64
},
679 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
680 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
681 { "evict_l2_skip", KSTAT_DATA_UINT64
},
682 { "hash_elements", KSTAT_DATA_UINT64
},
683 { "hash_elements_max", KSTAT_DATA_UINT64
},
684 { "hash_collisions", KSTAT_DATA_UINT64
},
685 { "hash_chains", KSTAT_DATA_UINT64
},
686 { "hash_chain_max", KSTAT_DATA_UINT64
},
687 { "p", KSTAT_DATA_UINT64
},
688 { "c", KSTAT_DATA_UINT64
},
689 { "c_min", KSTAT_DATA_UINT64
},
690 { "c_max", KSTAT_DATA_UINT64
},
691 { "size", KSTAT_DATA_UINT64
},
692 { "compressed_size", KSTAT_DATA_UINT64
},
693 { "uncompressed_size", KSTAT_DATA_UINT64
},
694 { "overhead_size", KSTAT_DATA_UINT64
},
695 { "hdr_size", KSTAT_DATA_UINT64
},
696 { "data_size", KSTAT_DATA_UINT64
},
697 { "metadata_size", KSTAT_DATA_UINT64
},
698 { "dbuf_size", KSTAT_DATA_UINT64
},
699 { "dnode_size", KSTAT_DATA_UINT64
},
700 { "bonus_size", KSTAT_DATA_UINT64
},
701 { "anon_size", KSTAT_DATA_UINT64
},
702 { "anon_evictable_data", KSTAT_DATA_UINT64
},
703 { "anon_evictable_metadata", KSTAT_DATA_UINT64
},
704 { "mru_size", KSTAT_DATA_UINT64
},
705 { "mru_evictable_data", KSTAT_DATA_UINT64
},
706 { "mru_evictable_metadata", KSTAT_DATA_UINT64
},
707 { "mru_ghost_size", KSTAT_DATA_UINT64
},
708 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64
},
709 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
710 { "mfu_size", KSTAT_DATA_UINT64
},
711 { "mfu_evictable_data", KSTAT_DATA_UINT64
},
712 { "mfu_evictable_metadata", KSTAT_DATA_UINT64
},
713 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
714 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64
},
715 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
716 { "l2_hits", KSTAT_DATA_UINT64
},
717 { "l2_misses", KSTAT_DATA_UINT64
},
718 { "l2_feeds", KSTAT_DATA_UINT64
},
719 { "l2_rw_clash", KSTAT_DATA_UINT64
},
720 { "l2_read_bytes", KSTAT_DATA_UINT64
},
721 { "l2_write_bytes", KSTAT_DATA_UINT64
},
722 { "l2_writes_sent", KSTAT_DATA_UINT64
},
723 { "l2_writes_done", KSTAT_DATA_UINT64
},
724 { "l2_writes_error", KSTAT_DATA_UINT64
},
725 { "l2_writes_lock_retry", KSTAT_DATA_UINT64
},
726 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
727 { "l2_evict_reading", KSTAT_DATA_UINT64
},
728 { "l2_evict_l1cached", KSTAT_DATA_UINT64
},
729 { "l2_free_on_write", KSTAT_DATA_UINT64
},
730 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
731 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
732 { "l2_io_error", KSTAT_DATA_UINT64
},
733 { "l2_size", KSTAT_DATA_UINT64
},
734 { "l2_asize", KSTAT_DATA_UINT64
},
735 { "l2_hdr_size", KSTAT_DATA_UINT64
},
736 { "memory_throttle_count", KSTAT_DATA_UINT64
},
737 { "memory_direct_count", KSTAT_DATA_UINT64
},
738 { "memory_indirect_count", KSTAT_DATA_UINT64
},
739 { "memory_all_bytes", KSTAT_DATA_UINT64
},
740 { "memory_free_bytes", KSTAT_DATA_UINT64
},
741 { "memory_available_bytes", KSTAT_DATA_INT64
},
742 { "arc_no_grow", KSTAT_DATA_UINT64
},
743 { "arc_tempreserve", KSTAT_DATA_UINT64
},
744 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
745 { "arc_prune", KSTAT_DATA_UINT64
},
746 { "arc_meta_used", KSTAT_DATA_UINT64
},
747 { "arc_meta_limit", KSTAT_DATA_UINT64
},
748 { "arc_dnode_limit", KSTAT_DATA_UINT64
},
749 { "arc_meta_max", KSTAT_DATA_UINT64
},
750 { "arc_meta_min", KSTAT_DATA_UINT64
},
751 { "sync_wait_for_async", KSTAT_DATA_UINT64
},
752 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64
},
753 { "arc_need_free", KSTAT_DATA_UINT64
},
754 { "arc_sys_free", KSTAT_DATA_UINT64
}
757 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
759 #define ARCSTAT_INCR(stat, val) \
760 atomic_add_64(&arc_stats.stat.value.ui64, (val))
762 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
763 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
765 #define ARCSTAT_MAX(stat, val) { \
767 while ((val) > (m = arc_stats.stat.value.ui64) && \
768 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
772 #define ARCSTAT_MAXSTAT(stat) \
773 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
776 * We define a macro to allow ARC hits/misses to be easily broken down by
777 * two separate conditions, giving a total of four different subtypes for
778 * each of hits and misses (so eight statistics total).
780 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
783 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
785 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
789 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
791 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
796 static arc_state_t
*arc_anon
;
797 static arc_state_t
*arc_mru
;
798 static arc_state_t
*arc_mru_ghost
;
799 static arc_state_t
*arc_mfu
;
800 static arc_state_t
*arc_mfu_ghost
;
801 static arc_state_t
*arc_l2c_only
;
804 * There are several ARC variables that are critical to export as kstats --
805 * but we don't want to have to grovel around in the kstat whenever we wish to
806 * manipulate them. For these variables, we therefore define them to be in
807 * terms of the statistic variable. This assures that we are not introducing
808 * the possibility of inconsistency by having shadow copies of the variables,
809 * while still allowing the code to be readable.
811 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
812 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
813 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
814 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
815 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
816 #define arc_no_grow ARCSTAT(arcstat_no_grow) /* do not grow cache size */
817 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
818 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
819 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
820 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
821 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
822 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
823 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
824 #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */
825 #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */
826 #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */
827 #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */
828 #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */
830 /* compressed size of entire arc */
831 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
832 /* uncompressed size of entire arc */
833 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
834 /* number of bytes in the arc from arc_buf_t's */
835 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
837 static list_t arc_prune_list
;
838 static kmutex_t arc_prune_mtx
;
839 static taskq_t
*arc_prune_taskq
;
841 #define GHOST_STATE(state) \
842 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
843 (state) == arc_l2c_only)
845 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
846 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
847 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
848 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
849 #define HDR_COMPRESSION_ENABLED(hdr) \
850 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
852 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
853 #define HDR_L2_READING(hdr) \
854 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
855 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
856 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
857 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
858 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
859 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
861 #define HDR_ISTYPE_METADATA(hdr) \
862 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
863 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
865 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
866 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
868 /* For storing compression mode in b_flags */
869 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
871 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
872 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
873 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
874 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
876 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
877 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
878 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
884 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
885 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
888 * Hash table routines
891 #define HT_LOCK_ALIGN 64
892 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
897 unsigned char pad
[HT_LOCK_PAD
];
901 #define BUF_LOCKS 8192
902 typedef struct buf_hash_table
{
904 arc_buf_hdr_t
**ht_table
;
905 struct ht_lock ht_locks
[BUF_LOCKS
];
908 static buf_hash_table_t buf_hash_table
;
910 #define BUF_HASH_INDEX(spa, dva, birth) \
911 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
912 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
913 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
914 #define HDR_LOCK(hdr) \
915 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
917 uint64_t zfs_crc64_table
[256];
923 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
924 #define L2ARC_HEADROOM 2 /* num of writes */
927 * If we discover during ARC scan any buffers to be compressed, we boost
928 * our headroom for the next scanning cycle by this percentage multiple.
930 #define L2ARC_HEADROOM_BOOST 200
931 #define L2ARC_FEED_SECS 1 /* caching interval secs */
932 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
935 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
936 * and each of the state has two types: data and metadata.
938 #define L2ARC_FEED_TYPES 4
940 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
941 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
943 /* L2ARC Performance Tunables */
944 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
945 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
946 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
947 unsigned long l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
948 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
949 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
950 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
951 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
952 int l2arc_norw
= B_FALSE
; /* no reads during writes */
957 static list_t L2ARC_dev_list
; /* device list */
958 static list_t
*l2arc_dev_list
; /* device list pointer */
959 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
960 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
961 static list_t L2ARC_free_on_write
; /* free after write buf list */
962 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
963 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
964 static uint64_t l2arc_ndev
; /* number of devices */
966 typedef struct l2arc_read_callback
{
967 arc_buf_hdr_t
*l2rcb_hdr
; /* read header */
968 blkptr_t l2rcb_bp
; /* original blkptr */
969 zbookmark_phys_t l2rcb_zb
; /* original bookmark */
970 int l2rcb_flags
; /* original flags */
971 abd_t
*l2rcb_abd
; /* temporary buffer */
972 } l2arc_read_callback_t
;
974 typedef struct l2arc_data_free
{
975 /* protected by l2arc_free_on_write_mtx */
978 arc_buf_contents_t l2df_type
;
979 list_node_t l2df_list_node
;
982 static kmutex_t l2arc_feed_thr_lock
;
983 static kcondvar_t l2arc_feed_thr_cv
;
984 static uint8_t l2arc_thread_exit
;
986 static abd_t
*arc_get_data_abd(arc_buf_hdr_t
*, uint64_t, void *);
987 static void *arc_get_data_buf(arc_buf_hdr_t
*, uint64_t, void *);
988 static void arc_get_data_impl(arc_buf_hdr_t
*, uint64_t, void *);
989 static void arc_free_data_abd(arc_buf_hdr_t
*, abd_t
*, uint64_t, void *);
990 static void arc_free_data_buf(arc_buf_hdr_t
*, void *, uint64_t, void *);
991 static void arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
);
992 static void arc_hdr_free_pabd(arc_buf_hdr_t
*);
993 static void arc_hdr_alloc_pabd(arc_buf_hdr_t
*);
994 static void arc_access(arc_buf_hdr_t
*, kmutex_t
*);
995 static boolean_t
arc_is_overflowing(void);
996 static void arc_buf_watch(arc_buf_t
*);
997 static void arc_tuning_update(void);
998 static void arc_prune_async(int64_t);
999 static uint64_t arc_all_memory(void);
1001 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t
*);
1002 static uint32_t arc_bufc_to_flags(arc_buf_contents_t
);
1003 static inline void arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
1004 static inline void arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
1006 static boolean_t
l2arc_write_eligible(uint64_t, arc_buf_hdr_t
*);
1007 static void l2arc_read_done(zio_t
*);
1010 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
1012 uint8_t *vdva
= (uint8_t *)dva
;
1013 uint64_t crc
= -1ULL;
1016 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
1018 for (i
= 0; i
< sizeof (dva_t
); i
++)
1019 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
1021 crc
^= (spa
>>8) ^ birth
;
1026 #define HDR_EMPTY(hdr) \
1027 ((hdr)->b_dva.dva_word[0] == 0 && \
1028 (hdr)->b_dva.dva_word[1] == 0)
1030 #define HDR_EQUAL(spa, dva, birth, hdr) \
1031 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1032 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1033 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1036 buf_discard_identity(arc_buf_hdr_t
*hdr
)
1038 hdr
->b_dva
.dva_word
[0] = 0;
1039 hdr
->b_dva
.dva_word
[1] = 0;
1043 static arc_buf_hdr_t
*
1044 buf_hash_find(uint64_t spa
, const blkptr_t
*bp
, kmutex_t
**lockp
)
1046 const dva_t
*dva
= BP_IDENTITY(bp
);
1047 uint64_t birth
= BP_PHYSICAL_BIRTH(bp
);
1048 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
1049 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1052 mutex_enter(hash_lock
);
1053 for (hdr
= buf_hash_table
.ht_table
[idx
]; hdr
!= NULL
;
1054 hdr
= hdr
->b_hash_next
) {
1055 if (HDR_EQUAL(spa
, dva
, birth
, hdr
)) {
1060 mutex_exit(hash_lock
);
1066 * Insert an entry into the hash table. If there is already an element
1067 * equal to elem in the hash table, then the already existing element
1068 * will be returned and the new element will not be inserted.
1069 * Otherwise returns NULL.
1070 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1072 static arc_buf_hdr_t
*
1073 buf_hash_insert(arc_buf_hdr_t
*hdr
, kmutex_t
**lockp
)
1075 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1076 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1077 arc_buf_hdr_t
*fhdr
;
1080 ASSERT(!DVA_IS_EMPTY(&hdr
->b_dva
));
1081 ASSERT(hdr
->b_birth
!= 0);
1082 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
1084 if (lockp
!= NULL
) {
1086 mutex_enter(hash_lock
);
1088 ASSERT(MUTEX_HELD(hash_lock
));
1091 for (fhdr
= buf_hash_table
.ht_table
[idx
], i
= 0; fhdr
!= NULL
;
1092 fhdr
= fhdr
->b_hash_next
, i
++) {
1093 if (HDR_EQUAL(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
, fhdr
))
1097 hdr
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
1098 buf_hash_table
.ht_table
[idx
] = hdr
;
1099 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1101 /* collect some hash table performance data */
1103 ARCSTAT_BUMP(arcstat_hash_collisions
);
1105 ARCSTAT_BUMP(arcstat_hash_chains
);
1107 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
1110 ARCSTAT_BUMP(arcstat_hash_elements
);
1111 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
1117 buf_hash_remove(arc_buf_hdr_t
*hdr
)
1119 arc_buf_hdr_t
*fhdr
, **hdrp
;
1120 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1122 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
1123 ASSERT(HDR_IN_HASH_TABLE(hdr
));
1125 hdrp
= &buf_hash_table
.ht_table
[idx
];
1126 while ((fhdr
= *hdrp
) != hdr
) {
1127 ASSERT3P(fhdr
, !=, NULL
);
1128 hdrp
= &fhdr
->b_hash_next
;
1130 *hdrp
= hdr
->b_hash_next
;
1131 hdr
->b_hash_next
= NULL
;
1132 arc_hdr_clear_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1134 /* collect some hash table performance data */
1135 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
1137 if (buf_hash_table
.ht_table
[idx
] &&
1138 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
1139 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
1143 * Global data structures and functions for the buf kmem cache.
1145 static kmem_cache_t
*hdr_full_cache
;
1146 static kmem_cache_t
*hdr_l2only_cache
;
1147 static kmem_cache_t
*buf_cache
;
1154 #if defined(_KERNEL) && defined(HAVE_SPL)
1156 * Large allocations which do not require contiguous pages
1157 * should be using vmem_free() in the linux kernel\
1159 vmem_free(buf_hash_table
.ht_table
,
1160 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1162 kmem_free(buf_hash_table
.ht_table
,
1163 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1165 for (i
= 0; i
< BUF_LOCKS
; i
++)
1166 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
1167 kmem_cache_destroy(hdr_full_cache
);
1168 kmem_cache_destroy(hdr_l2only_cache
);
1169 kmem_cache_destroy(buf_cache
);
1173 * Constructor callback - called when the cache is empty
1174 * and a new buf is requested.
1178 hdr_full_cons(void *vbuf
, void *unused
, int kmflag
)
1180 arc_buf_hdr_t
*hdr
= vbuf
;
1182 bzero(hdr
, HDR_FULL_SIZE
);
1183 cv_init(&hdr
->b_l1hdr
.b_cv
, NULL
, CV_DEFAULT
, NULL
);
1184 zfs_refcount_create(&hdr
->b_l1hdr
.b_refcnt
);
1185 mutex_init(&hdr
->b_l1hdr
.b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1186 list_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1187 list_link_init(&hdr
->b_l2hdr
.b_l2node
);
1188 multilist_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1189 arc_space_consume(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1196 hdr_l2only_cons(void *vbuf
, void *unused
, int kmflag
)
1198 arc_buf_hdr_t
*hdr
= vbuf
;
1200 bzero(hdr
, HDR_L2ONLY_SIZE
);
1201 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1208 buf_cons(void *vbuf
, void *unused
, int kmflag
)
1210 arc_buf_t
*buf
= vbuf
;
1212 bzero(buf
, sizeof (arc_buf_t
));
1213 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1214 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1220 * Destructor callback - called when a cached buf is
1221 * no longer required.
1225 hdr_full_dest(void *vbuf
, void *unused
)
1227 arc_buf_hdr_t
*hdr
= vbuf
;
1229 ASSERT(HDR_EMPTY(hdr
));
1230 cv_destroy(&hdr
->b_l1hdr
.b_cv
);
1231 zfs_refcount_destroy(&hdr
->b_l1hdr
.b_refcnt
);
1232 mutex_destroy(&hdr
->b_l1hdr
.b_freeze_lock
);
1233 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1234 arc_space_return(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1239 hdr_l2only_dest(void *vbuf
, void *unused
)
1241 ASSERTV(arc_buf_hdr_t
*hdr
= vbuf
);
1243 ASSERT(HDR_EMPTY(hdr
));
1244 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1249 buf_dest(void *vbuf
, void *unused
)
1251 arc_buf_t
*buf
= vbuf
;
1253 mutex_destroy(&buf
->b_evict_lock
);
1254 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1258 * Reclaim callback -- invoked when memory is low.
1262 hdr_recl(void *unused
)
1264 dprintf("hdr_recl called\n");
1266 * umem calls the reclaim func when we destroy the buf cache,
1267 * which is after we do arc_fini().
1270 cv_signal(&arc_reclaim_thread_cv
);
1276 uint64_t *ct
= NULL
;
1277 uint64_t hsize
= 1ULL << 12;
1281 * The hash table is big enough to fill all of physical memory
1282 * with an average block size of zfs_arc_average_blocksize (default 8K).
1283 * By default, the table will take up
1284 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1286 while (hsize
* zfs_arc_average_blocksize
< arc_all_memory())
1289 buf_hash_table
.ht_mask
= hsize
- 1;
1290 #if defined(_KERNEL) && defined(HAVE_SPL)
1292 * Large allocations which do not require contiguous pages
1293 * should be using vmem_alloc() in the linux kernel
1295 buf_hash_table
.ht_table
=
1296 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
1298 buf_hash_table
.ht_table
=
1299 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
1301 if (buf_hash_table
.ht_table
== NULL
) {
1302 ASSERT(hsize
> (1ULL << 8));
1307 hdr_full_cache
= kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE
,
1308 0, hdr_full_cons
, hdr_full_dest
, hdr_recl
, NULL
, NULL
, 0);
1309 hdr_l2only_cache
= kmem_cache_create("arc_buf_hdr_t_l2only",
1310 HDR_L2ONLY_SIZE
, 0, hdr_l2only_cons
, hdr_l2only_dest
, hdr_recl
,
1312 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
1313 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
1315 for (i
= 0; i
< 256; i
++)
1316 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
1317 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
1319 for (i
= 0; i
< BUF_LOCKS
; i
++) {
1320 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
1321 NULL
, MUTEX_DEFAULT
, NULL
);
1325 #define ARC_MINTIME (hz>>4) /* 62 ms */
1328 * This is the size that the buf occupies in memory. If the buf is compressed,
1329 * it will correspond to the compressed size. You should use this method of
1330 * getting the buf size unless you explicitly need the logical size.
1333 arc_buf_size(arc_buf_t
*buf
)
1335 return (ARC_BUF_COMPRESSED(buf
) ?
1336 HDR_GET_PSIZE(buf
->b_hdr
) : HDR_GET_LSIZE(buf
->b_hdr
));
1340 arc_buf_lsize(arc_buf_t
*buf
)
1342 return (HDR_GET_LSIZE(buf
->b_hdr
));
1346 arc_get_compression(arc_buf_t
*buf
)
1348 return (ARC_BUF_COMPRESSED(buf
) ?
1349 HDR_GET_COMPRESS(buf
->b_hdr
) : ZIO_COMPRESS_OFF
);
1352 static inline boolean_t
1353 arc_buf_is_shared(arc_buf_t
*buf
)
1355 boolean_t shared
= (buf
->b_data
!= NULL
&&
1356 buf
->b_hdr
->b_l1hdr
.b_pabd
!= NULL
&&
1357 abd_is_linear(buf
->b_hdr
->b_l1hdr
.b_pabd
) &&
1358 buf
->b_data
== abd_to_buf(buf
->b_hdr
->b_l1hdr
.b_pabd
));
1359 IMPLY(shared
, HDR_SHARED_DATA(buf
->b_hdr
));
1360 IMPLY(shared
, ARC_BUF_SHARED(buf
));
1361 IMPLY(shared
, ARC_BUF_COMPRESSED(buf
) || ARC_BUF_LAST(buf
));
1364 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1365 * already being shared" requirement prevents us from doing that.
1372 * Free the checksum associated with this header. If there is no checksum, this
1376 arc_cksum_free(arc_buf_hdr_t
*hdr
)
1378 ASSERT(HDR_HAS_L1HDR(hdr
));
1379 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1380 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1381 kmem_free(hdr
->b_l1hdr
.b_freeze_cksum
, sizeof (zio_cksum_t
));
1382 hdr
->b_l1hdr
.b_freeze_cksum
= NULL
;
1384 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1388 * Return true iff at least one of the bufs on hdr is not compressed.
1391 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t
*hdr
)
1393 for (arc_buf_t
*b
= hdr
->b_l1hdr
.b_buf
; b
!= NULL
; b
= b
->b_next
) {
1394 if (!ARC_BUF_COMPRESSED(b
)) {
1403 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1404 * matches the checksum that is stored in the hdr. If there is no checksum,
1405 * or if the buf is compressed, this is a no-op.
1408 arc_cksum_verify(arc_buf_t
*buf
)
1410 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1413 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1416 if (ARC_BUF_COMPRESSED(buf
)) {
1417 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1418 arc_hdr_has_uncompressed_buf(hdr
));
1422 ASSERT(HDR_HAS_L1HDR(hdr
));
1424 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1425 if (hdr
->b_l1hdr
.b_freeze_cksum
== NULL
|| HDR_IO_ERROR(hdr
)) {
1426 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1430 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
, &zc
);
1431 if (!ZIO_CHECKSUM_EQUAL(*hdr
->b_l1hdr
.b_freeze_cksum
, zc
))
1432 panic("buffer modified while frozen!");
1433 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1437 arc_cksum_is_equal(arc_buf_hdr_t
*hdr
, zio_t
*zio
)
1439 enum zio_compress compress
= BP_GET_COMPRESS(zio
->io_bp
);
1440 boolean_t valid_cksum
;
1442 ASSERT(!BP_IS_EMBEDDED(zio
->io_bp
));
1443 VERIFY3U(BP_GET_PSIZE(zio
->io_bp
), ==, HDR_GET_PSIZE(hdr
));
1446 * We rely on the blkptr's checksum to determine if the block
1447 * is valid or not. When compressed arc is enabled, the l2arc
1448 * writes the block to the l2arc just as it appears in the pool.
1449 * This allows us to use the blkptr's checksum to validate the
1450 * data that we just read off of the l2arc without having to store
1451 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1452 * arc is disabled, then the data written to the l2arc is always
1453 * uncompressed and won't match the block as it exists in the main
1454 * pool. When this is the case, we must first compress it if it is
1455 * compressed on the main pool before we can validate the checksum.
1457 if (!HDR_COMPRESSION_ENABLED(hdr
) && compress
!= ZIO_COMPRESS_OFF
) {
1461 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1463 cbuf
= zio_buf_alloc(HDR_GET_PSIZE(hdr
));
1464 lsize
= HDR_GET_LSIZE(hdr
);
1465 csize
= zio_compress_data(compress
, zio
->io_abd
, cbuf
, lsize
);
1467 ASSERT3U(csize
, <=, HDR_GET_PSIZE(hdr
));
1468 if (csize
< HDR_GET_PSIZE(hdr
)) {
1470 * Compressed blocks are always a multiple of the
1471 * smallest ashift in the pool. Ideally, we would
1472 * like to round up the csize to the next
1473 * spa_min_ashift but that value may have changed
1474 * since the block was last written. Instead,
1475 * we rely on the fact that the hdr's psize
1476 * was set to the psize of the block when it was
1477 * last written. We set the csize to that value
1478 * and zero out any part that should not contain
1481 bzero((char *)cbuf
+ csize
, HDR_GET_PSIZE(hdr
) - csize
);
1482 csize
= HDR_GET_PSIZE(hdr
);
1484 zio_push_transform(zio
, cbuf
, csize
, HDR_GET_PSIZE(hdr
), NULL
);
1488 * Block pointers always store the checksum for the logical data.
1489 * If the block pointer has the gang bit set, then the checksum
1490 * it represents is for the reconstituted data and not for an
1491 * individual gang member. The zio pipeline, however, must be able to
1492 * determine the checksum of each of the gang constituents so it
1493 * treats the checksum comparison differently than what we need
1494 * for l2arc blocks. This prevents us from using the
1495 * zio_checksum_error() interface directly. Instead we must call the
1496 * zio_checksum_error_impl() so that we can ensure the checksum is
1497 * generated using the correct checksum algorithm and accounts for the
1498 * logical I/O size and not just a gang fragment.
1500 valid_cksum
= (zio_checksum_error_impl(zio
->io_spa
, zio
->io_bp
,
1501 BP_GET_CHECKSUM(zio
->io_bp
), zio
->io_abd
, zio
->io_size
,
1502 zio
->io_offset
, NULL
) == 0);
1503 zio_pop_transforms(zio
);
1504 return (valid_cksum
);
1508 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1509 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1510 * isn't modified later on. If buf is compressed or there is already a checksum
1511 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1514 arc_cksum_compute(arc_buf_t
*buf
)
1516 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1518 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1521 ASSERT(HDR_HAS_L1HDR(hdr
));
1523 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1524 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1525 ASSERT(arc_hdr_has_uncompressed_buf(hdr
));
1526 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1528 } else if (ARC_BUF_COMPRESSED(buf
)) {
1529 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1533 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1534 hdr
->b_l1hdr
.b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
),
1536 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
,
1537 hdr
->b_l1hdr
.b_freeze_cksum
);
1538 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1544 arc_buf_sigsegv(int sig
, siginfo_t
*si
, void *unused
)
1546 panic("Got SIGSEGV at address: 0x%lx\n", (long)si
->si_addr
);
1552 arc_buf_unwatch(arc_buf_t
*buf
)
1556 ASSERT0(mprotect(buf
->b_data
, arc_buf_size(buf
),
1557 PROT_READ
| PROT_WRITE
));
1564 arc_buf_watch(arc_buf_t
*buf
)
1568 ASSERT0(mprotect(buf
->b_data
, arc_buf_size(buf
),
1573 static arc_buf_contents_t
1574 arc_buf_type(arc_buf_hdr_t
*hdr
)
1576 arc_buf_contents_t type
;
1577 if (HDR_ISTYPE_METADATA(hdr
)) {
1578 type
= ARC_BUFC_METADATA
;
1580 type
= ARC_BUFC_DATA
;
1582 VERIFY3U(hdr
->b_type
, ==, type
);
1587 arc_is_metadata(arc_buf_t
*buf
)
1589 return (HDR_ISTYPE_METADATA(buf
->b_hdr
) != 0);
1593 arc_bufc_to_flags(arc_buf_contents_t type
)
1597 /* metadata field is 0 if buffer contains normal data */
1599 case ARC_BUFC_METADATA
:
1600 return (ARC_FLAG_BUFC_METADATA
);
1604 panic("undefined ARC buffer type!");
1605 return ((uint32_t)-1);
1609 arc_buf_thaw(arc_buf_t
*buf
)
1611 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1613 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
1614 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
1616 arc_cksum_verify(buf
);
1619 * Compressed buffers do not manipulate the b_freeze_cksum or
1620 * allocate b_thawed.
1622 if (ARC_BUF_COMPRESSED(buf
)) {
1623 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1624 arc_hdr_has_uncompressed_buf(hdr
));
1628 ASSERT(HDR_HAS_L1HDR(hdr
));
1629 arc_cksum_free(hdr
);
1630 arc_buf_unwatch(buf
);
1634 arc_buf_freeze(arc_buf_t
*buf
)
1636 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1637 kmutex_t
*hash_lock
;
1639 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1642 if (ARC_BUF_COMPRESSED(buf
)) {
1643 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1644 arc_hdr_has_uncompressed_buf(hdr
));
1648 hash_lock
= HDR_LOCK(hdr
);
1649 mutex_enter(hash_lock
);
1651 ASSERT(HDR_HAS_L1HDR(hdr
));
1652 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
||
1653 hdr
->b_l1hdr
.b_state
== arc_anon
);
1654 arc_cksum_compute(buf
);
1655 mutex_exit(hash_lock
);
1659 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1660 * the following functions should be used to ensure that the flags are
1661 * updated in a thread-safe way. When manipulating the flags either
1662 * the hash_lock must be held or the hdr must be undiscoverable. This
1663 * ensures that we're not racing with any other threads when updating
1667 arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1669 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1670 hdr
->b_flags
|= flags
;
1674 arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1676 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1677 hdr
->b_flags
&= ~flags
;
1681 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1682 * done in a special way since we have to clear and set bits
1683 * at the same time. Consumers that wish to set the compression bits
1684 * must use this function to ensure that the flags are updated in
1685 * thread-safe manner.
1688 arc_hdr_set_compress(arc_buf_hdr_t
*hdr
, enum zio_compress cmp
)
1690 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1693 * Holes and embedded blocks will always have a psize = 0 so
1694 * we ignore the compression of the blkptr and set the
1695 * want to uncompress them. Mark them as uncompressed.
1697 if (!zfs_compressed_arc_enabled
|| HDR_GET_PSIZE(hdr
) == 0) {
1698 arc_hdr_clear_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1699 HDR_SET_COMPRESS(hdr
, ZIO_COMPRESS_OFF
);
1700 ASSERT(!HDR_COMPRESSION_ENABLED(hdr
));
1701 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1703 arc_hdr_set_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1704 HDR_SET_COMPRESS(hdr
, cmp
);
1705 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, cmp
);
1706 ASSERT(HDR_COMPRESSION_ENABLED(hdr
));
1711 * Looks for another buf on the same hdr which has the data decompressed, copies
1712 * from it, and returns true. If no such buf exists, returns false.
1715 arc_buf_try_copy_decompressed_data(arc_buf_t
*buf
)
1717 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1718 boolean_t copied
= B_FALSE
;
1720 ASSERT(HDR_HAS_L1HDR(hdr
));
1721 ASSERT3P(buf
->b_data
, !=, NULL
);
1722 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1724 for (arc_buf_t
*from
= hdr
->b_l1hdr
.b_buf
; from
!= NULL
;
1725 from
= from
->b_next
) {
1726 /* can't use our own data buffer */
1731 if (!ARC_BUF_COMPRESSED(from
)) {
1732 bcopy(from
->b_data
, buf
->b_data
, arc_buf_size(buf
));
1739 * There were no decompressed bufs, so there should not be a
1740 * checksum on the hdr either.
1742 EQUIV(!copied
, hdr
->b_l1hdr
.b_freeze_cksum
== NULL
);
1748 * Given a buf that has a data buffer attached to it, this function will
1749 * efficiently fill the buf with data of the specified compression setting from
1750 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1751 * are already sharing a data buf, no copy is performed.
1753 * If the buf is marked as compressed but uncompressed data was requested, this
1754 * will allocate a new data buffer for the buf, remove that flag, and fill the
1755 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1756 * uncompressed data, and (since we haven't added support for it yet) if you
1757 * want compressed data your buf must already be marked as compressed and have
1758 * the correct-sized data buffer.
1761 arc_buf_fill(arc_buf_t
*buf
, boolean_t compressed
)
1763 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1764 boolean_t hdr_compressed
= (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
1765 dmu_object_byteswap_t bswap
= hdr
->b_l1hdr
.b_byteswap
;
1767 ASSERT3P(buf
->b_data
, !=, NULL
);
1768 IMPLY(compressed
, hdr_compressed
);
1769 IMPLY(compressed
, ARC_BUF_COMPRESSED(buf
));
1771 if (hdr_compressed
== compressed
) {
1772 if (!arc_buf_is_shared(buf
)) {
1773 abd_copy_to_buf(buf
->b_data
, hdr
->b_l1hdr
.b_pabd
,
1777 ASSERT(hdr_compressed
);
1778 ASSERT(!compressed
);
1779 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, HDR_GET_PSIZE(hdr
));
1782 * If the buf is sharing its data with the hdr, unlink it and
1783 * allocate a new data buffer for the buf.
1785 if (arc_buf_is_shared(buf
)) {
1786 ASSERT(ARC_BUF_COMPRESSED(buf
));
1788 /* We need to give the buf it's own b_data */
1789 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
1791 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
1792 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
1794 /* Previously overhead was 0; just add new overhead */
1795 ARCSTAT_INCR(arcstat_overhead_size
, HDR_GET_LSIZE(hdr
));
1796 } else if (ARC_BUF_COMPRESSED(buf
)) {
1797 /* We need to reallocate the buf's b_data */
1798 arc_free_data_buf(hdr
, buf
->b_data
, HDR_GET_PSIZE(hdr
),
1801 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
1803 /* We increased the size of b_data; update overhead */
1804 ARCSTAT_INCR(arcstat_overhead_size
,
1805 HDR_GET_LSIZE(hdr
) - HDR_GET_PSIZE(hdr
));
1809 * Regardless of the buf's previous compression settings, it
1810 * should not be compressed at the end of this function.
1812 buf
->b_flags
&= ~ARC_BUF_FLAG_COMPRESSED
;
1815 * Try copying the data from another buf which already has a
1816 * decompressed version. If that's not possible, it's time to
1817 * bite the bullet and decompress the data from the hdr.
1819 if (arc_buf_try_copy_decompressed_data(buf
)) {
1820 /* Skip byteswapping and checksumming (already done) */
1821 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, !=, NULL
);
1824 int error
= zio_decompress_data(HDR_GET_COMPRESS(hdr
),
1825 hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
1826 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
1829 * Absent hardware errors or software bugs, this should
1830 * be impossible, but log it anyway so we can debug it.
1834 "hdr %p, compress %d, psize %d, lsize %d",
1835 hdr
, HDR_GET_COMPRESS(hdr
),
1836 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
1837 return (SET_ERROR(EIO
));
1842 /* Byteswap the buf's data if necessary */
1843 if (bswap
!= DMU_BSWAP_NUMFUNCS
) {
1844 ASSERT(!HDR_SHARED_DATA(hdr
));
1845 ASSERT3U(bswap
, <, DMU_BSWAP_NUMFUNCS
);
1846 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, HDR_GET_LSIZE(hdr
));
1849 /* Compute the hdr's checksum if necessary */
1850 arc_cksum_compute(buf
);
1856 arc_decompress(arc_buf_t
*buf
)
1858 return (arc_buf_fill(buf
, B_FALSE
));
1862 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1865 arc_hdr_size(arc_buf_hdr_t
*hdr
)
1869 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
1870 HDR_GET_PSIZE(hdr
) > 0) {
1871 size
= HDR_GET_PSIZE(hdr
);
1873 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, 0);
1874 size
= HDR_GET_LSIZE(hdr
);
1880 * Increment the amount of evictable space in the arc_state_t's refcount.
1881 * We account for the space used by the hdr and the arc buf individually
1882 * so that we can add and remove them from the refcount individually.
1885 arc_evictable_space_increment(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
1887 arc_buf_contents_t type
= arc_buf_type(hdr
);
1890 ASSERT(HDR_HAS_L1HDR(hdr
));
1892 if (GHOST_STATE(state
)) {
1893 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
1894 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1895 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
1896 (void) zfs_refcount_add_many(&state
->arcs_esize
[type
],
1897 HDR_GET_LSIZE(hdr
), hdr
);
1901 ASSERT(!GHOST_STATE(state
));
1902 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
1903 (void) zfs_refcount_add_many(&state
->arcs_esize
[type
],
1904 arc_hdr_size(hdr
), hdr
);
1906 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
; buf
= buf
->b_next
) {
1907 if (arc_buf_is_shared(buf
))
1909 (void) zfs_refcount_add_many(&state
->arcs_esize
[type
],
1910 arc_buf_size(buf
), buf
);
1915 * Decrement the amount of evictable space in the arc_state_t's refcount.
1916 * We account for the space used by the hdr and the arc buf individually
1917 * so that we can add and remove them from the refcount individually.
1920 arc_evictable_space_decrement(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
1922 arc_buf_contents_t type
= arc_buf_type(hdr
);
1925 ASSERT(HDR_HAS_L1HDR(hdr
));
1927 if (GHOST_STATE(state
)) {
1928 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
1929 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1930 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
1931 (void) zfs_refcount_remove_many(&state
->arcs_esize
[type
],
1932 HDR_GET_LSIZE(hdr
), hdr
);
1936 ASSERT(!GHOST_STATE(state
));
1937 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
1938 (void) zfs_refcount_remove_many(&state
->arcs_esize
[type
],
1939 arc_hdr_size(hdr
), hdr
);
1941 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
; buf
= buf
->b_next
) {
1942 if (arc_buf_is_shared(buf
))
1944 (void) zfs_refcount_remove_many(&state
->arcs_esize
[type
],
1945 arc_buf_size(buf
), buf
);
1950 * Add a reference to this hdr indicating that someone is actively
1951 * referencing that memory. When the refcount transitions from 0 to 1,
1952 * we remove it from the respective arc_state_t list to indicate that
1953 * it is not evictable.
1956 add_reference(arc_buf_hdr_t
*hdr
, void *tag
)
1960 ASSERT(HDR_HAS_L1HDR(hdr
));
1961 if (!MUTEX_HELD(HDR_LOCK(hdr
))) {
1962 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
1963 ASSERT(zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
1964 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1967 state
= hdr
->b_l1hdr
.b_state
;
1969 if ((zfs_refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
) == 1) &&
1970 (state
!= arc_anon
)) {
1971 /* We don't use the L2-only state list. */
1972 if (state
!= arc_l2c_only
) {
1973 multilist_remove(state
->arcs_list
[arc_buf_type(hdr
)],
1975 arc_evictable_space_decrement(hdr
, state
);
1977 /* remove the prefetch flag if we get a reference */
1978 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
1983 * Remove a reference from this hdr. When the reference transitions from
1984 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
1985 * list making it eligible for eviction.
1988 remove_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1991 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
1993 ASSERT(HDR_HAS_L1HDR(hdr
));
1994 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
1995 ASSERT(!GHOST_STATE(state
));
1998 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1999 * check to prevent usage of the arc_l2c_only list.
2001 if (((cnt
= zfs_refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
)) == 0) &&
2002 (state
!= arc_anon
)) {
2003 multilist_insert(state
->arcs_list
[arc_buf_type(hdr
)], hdr
);
2004 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
2005 arc_evictable_space_increment(hdr
, state
);
2011 * Returns detailed information about a specific arc buffer. When the
2012 * state_index argument is set the function will calculate the arc header
2013 * list position for its arc state. Since this requires a linear traversal
2014 * callers are strongly encourage not to do this. However, it can be helpful
2015 * for targeted analysis so the functionality is provided.
2018 arc_buf_info(arc_buf_t
*ab
, arc_buf_info_t
*abi
, int state_index
)
2020 arc_buf_hdr_t
*hdr
= ab
->b_hdr
;
2021 l1arc_buf_hdr_t
*l1hdr
= NULL
;
2022 l2arc_buf_hdr_t
*l2hdr
= NULL
;
2023 arc_state_t
*state
= NULL
;
2025 memset(abi
, 0, sizeof (arc_buf_info_t
));
2030 abi
->abi_flags
= hdr
->b_flags
;
2032 if (HDR_HAS_L1HDR(hdr
)) {
2033 l1hdr
= &hdr
->b_l1hdr
;
2034 state
= l1hdr
->b_state
;
2036 if (HDR_HAS_L2HDR(hdr
))
2037 l2hdr
= &hdr
->b_l2hdr
;
2040 abi
->abi_bufcnt
= l1hdr
->b_bufcnt
;
2041 abi
->abi_access
= l1hdr
->b_arc_access
;
2042 abi
->abi_mru_hits
= l1hdr
->b_mru_hits
;
2043 abi
->abi_mru_ghost_hits
= l1hdr
->b_mru_ghost_hits
;
2044 abi
->abi_mfu_hits
= l1hdr
->b_mfu_hits
;
2045 abi
->abi_mfu_ghost_hits
= l1hdr
->b_mfu_ghost_hits
;
2046 abi
->abi_holds
= zfs_refcount_count(&l1hdr
->b_refcnt
);
2050 abi
->abi_l2arc_dattr
= l2hdr
->b_daddr
;
2051 abi
->abi_l2arc_hits
= l2hdr
->b_hits
;
2054 abi
->abi_state_type
= state
? state
->arcs_state
: ARC_STATE_ANON
;
2055 abi
->abi_state_contents
= arc_buf_type(hdr
);
2056 abi
->abi_size
= arc_hdr_size(hdr
);
2060 * Move the supplied buffer to the indicated state. The hash lock
2061 * for the buffer must be held by the caller.
2064 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*hdr
,
2065 kmutex_t
*hash_lock
)
2067 arc_state_t
*old_state
;
2070 boolean_t update_old
, update_new
;
2071 arc_buf_contents_t buftype
= arc_buf_type(hdr
);
2074 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2075 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2076 * L1 hdr doesn't always exist when we change state to arc_anon before
2077 * destroying a header, in which case reallocating to add the L1 hdr is
2080 if (HDR_HAS_L1HDR(hdr
)) {
2081 old_state
= hdr
->b_l1hdr
.b_state
;
2082 refcnt
= zfs_refcount_count(&hdr
->b_l1hdr
.b_refcnt
);
2083 bufcnt
= hdr
->b_l1hdr
.b_bufcnt
;
2084 update_old
= (bufcnt
> 0 || hdr
->b_l1hdr
.b_pabd
!= NULL
);
2086 old_state
= arc_l2c_only
;
2089 update_old
= B_FALSE
;
2091 update_new
= update_old
;
2093 ASSERT(MUTEX_HELD(hash_lock
));
2094 ASSERT3P(new_state
, !=, old_state
);
2095 ASSERT(!GHOST_STATE(new_state
) || bufcnt
== 0);
2096 ASSERT(old_state
!= arc_anon
|| bufcnt
<= 1);
2099 * If this buffer is evictable, transfer it from the
2100 * old state list to the new state list.
2103 if (old_state
!= arc_anon
&& old_state
!= arc_l2c_only
) {
2104 ASSERT(HDR_HAS_L1HDR(hdr
));
2105 multilist_remove(old_state
->arcs_list
[buftype
], hdr
);
2107 if (GHOST_STATE(old_state
)) {
2109 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2110 update_old
= B_TRUE
;
2112 arc_evictable_space_decrement(hdr
, old_state
);
2114 if (new_state
!= arc_anon
&& new_state
!= arc_l2c_only
) {
2116 * An L1 header always exists here, since if we're
2117 * moving to some L1-cached state (i.e. not l2c_only or
2118 * anonymous), we realloc the header to add an L1hdr
2121 ASSERT(HDR_HAS_L1HDR(hdr
));
2122 multilist_insert(new_state
->arcs_list
[buftype
], hdr
);
2124 if (GHOST_STATE(new_state
)) {
2126 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2127 update_new
= B_TRUE
;
2129 arc_evictable_space_increment(hdr
, new_state
);
2133 ASSERT(!HDR_EMPTY(hdr
));
2134 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(hdr
))
2135 buf_hash_remove(hdr
);
2137 /* adjust state sizes (ignore arc_l2c_only) */
2139 if (update_new
&& new_state
!= arc_l2c_only
) {
2140 ASSERT(HDR_HAS_L1HDR(hdr
));
2141 if (GHOST_STATE(new_state
)) {
2145 * When moving a header to a ghost state, we first
2146 * remove all arc buffers. Thus, we'll have a
2147 * bufcnt of zero, and no arc buffer to use for
2148 * the reference. As a result, we use the arc
2149 * header pointer for the reference.
2151 (void) zfs_refcount_add_many(&new_state
->arcs_size
,
2152 HDR_GET_LSIZE(hdr
), hdr
);
2153 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2156 uint32_t buffers
= 0;
2159 * Each individual buffer holds a unique reference,
2160 * thus we must remove each of these references one
2163 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2164 buf
= buf
->b_next
) {
2165 ASSERT3U(bufcnt
, !=, 0);
2169 * When the arc_buf_t is sharing the data
2170 * block with the hdr, the owner of the
2171 * reference belongs to the hdr. Only
2172 * add to the refcount if the arc_buf_t is
2175 if (arc_buf_is_shared(buf
))
2178 (void) zfs_refcount_add_many(
2179 &new_state
->arcs_size
,
2180 arc_buf_size(buf
), buf
);
2182 ASSERT3U(bufcnt
, ==, buffers
);
2184 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2185 (void) zfs_refcount_add_many(
2186 &new_state
->arcs_size
,
2187 arc_hdr_size(hdr
), hdr
);
2189 ASSERT(GHOST_STATE(old_state
));
2194 if (update_old
&& old_state
!= arc_l2c_only
) {
2195 ASSERT(HDR_HAS_L1HDR(hdr
));
2196 if (GHOST_STATE(old_state
)) {
2198 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2201 * When moving a header off of a ghost state,
2202 * the header will not contain any arc buffers.
2203 * We use the arc header pointer for the reference
2204 * which is exactly what we did when we put the
2205 * header on the ghost state.
2208 (void) zfs_refcount_remove_many(&old_state
->arcs_size
,
2209 HDR_GET_LSIZE(hdr
), hdr
);
2212 uint32_t buffers
= 0;
2215 * Each individual buffer holds a unique reference,
2216 * thus we must remove each of these references one
2219 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2220 buf
= buf
->b_next
) {
2221 ASSERT3U(bufcnt
, !=, 0);
2225 * When the arc_buf_t is sharing the data
2226 * block with the hdr, the owner of the
2227 * reference belongs to the hdr. Only
2228 * add to the refcount if the arc_buf_t is
2231 if (arc_buf_is_shared(buf
))
2234 (void) zfs_refcount_remove_many(
2235 &old_state
->arcs_size
, arc_buf_size(buf
),
2238 ASSERT3U(bufcnt
, ==, buffers
);
2239 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2240 (void) zfs_refcount_remove_many(
2241 &old_state
->arcs_size
, arc_hdr_size(hdr
), hdr
);
2245 if (HDR_HAS_L1HDR(hdr
))
2246 hdr
->b_l1hdr
.b_state
= new_state
;
2249 * L2 headers should never be on the L2 state list since they don't
2250 * have L1 headers allocated.
2252 ASSERT(multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]) &&
2253 multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]));
2257 arc_space_consume(uint64_t space
, arc_space_type_t type
)
2259 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2264 case ARC_SPACE_DATA
:
2265 ARCSTAT_INCR(arcstat_data_size
, space
);
2267 case ARC_SPACE_META
:
2268 ARCSTAT_INCR(arcstat_metadata_size
, space
);
2270 case ARC_SPACE_BONUS
:
2271 ARCSTAT_INCR(arcstat_bonus_size
, space
);
2273 case ARC_SPACE_DNODE
:
2274 ARCSTAT_INCR(arcstat_dnode_size
, space
);
2276 case ARC_SPACE_DBUF
:
2277 ARCSTAT_INCR(arcstat_dbuf_size
, space
);
2279 case ARC_SPACE_HDRS
:
2280 ARCSTAT_INCR(arcstat_hdr_size
, space
);
2282 case ARC_SPACE_L2HDRS
:
2283 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
2287 if (type
!= ARC_SPACE_DATA
)
2288 ARCSTAT_INCR(arcstat_meta_used
, space
);
2290 atomic_add_64(&arc_size
, space
);
2294 arc_space_return(uint64_t space
, arc_space_type_t type
)
2296 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2301 case ARC_SPACE_DATA
:
2302 ARCSTAT_INCR(arcstat_data_size
, -space
);
2304 case ARC_SPACE_META
:
2305 ARCSTAT_INCR(arcstat_metadata_size
, -space
);
2307 case ARC_SPACE_BONUS
:
2308 ARCSTAT_INCR(arcstat_bonus_size
, -space
);
2310 case ARC_SPACE_DNODE
:
2311 ARCSTAT_INCR(arcstat_dnode_size
, -space
);
2313 case ARC_SPACE_DBUF
:
2314 ARCSTAT_INCR(arcstat_dbuf_size
, -space
);
2316 case ARC_SPACE_HDRS
:
2317 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
2319 case ARC_SPACE_L2HDRS
:
2320 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
2324 if (type
!= ARC_SPACE_DATA
) {
2325 ASSERT(arc_meta_used
>= space
);
2326 if (arc_meta_max
< arc_meta_used
)
2327 arc_meta_max
= arc_meta_used
;
2328 ARCSTAT_INCR(arcstat_meta_used
, -space
);
2331 ASSERT(arc_size
>= space
);
2332 atomic_add_64(&arc_size
, -space
);
2336 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2337 * with the hdr's b_pabd.
2340 arc_can_share(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2343 * The criteria for sharing a hdr's data are:
2344 * 1. the hdr's compression matches the buf's compression
2345 * 2. the hdr doesn't need to be byteswapped
2346 * 3. the hdr isn't already being shared
2347 * 4. the buf is either compressed or it is the last buf in the hdr list
2349 * Criterion #4 maintains the invariant that shared uncompressed
2350 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2351 * might ask, "if a compressed buf is allocated first, won't that be the
2352 * last thing in the list?", but in that case it's impossible to create
2353 * a shared uncompressed buf anyway (because the hdr must be compressed
2354 * to have the compressed buf). You might also think that #3 is
2355 * sufficient to make this guarantee, however it's possible
2356 * (specifically in the rare L2ARC write race mentioned in
2357 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2358 * is sharable, but wasn't at the time of its allocation. Rather than
2359 * allow a new shared uncompressed buf to be created and then shuffle
2360 * the list around to make it the last element, this simply disallows
2361 * sharing if the new buf isn't the first to be added.
2363 ASSERT3P(buf
->b_hdr
, ==, hdr
);
2364 boolean_t hdr_compressed
= HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
;
2365 boolean_t buf_compressed
= ARC_BUF_COMPRESSED(buf
) != 0;
2366 return (buf_compressed
== hdr_compressed
&&
2367 hdr
->b_l1hdr
.b_byteswap
== DMU_BSWAP_NUMFUNCS
&&
2368 !HDR_SHARED_DATA(hdr
) &&
2369 (ARC_BUF_LAST(buf
) || ARC_BUF_COMPRESSED(buf
)));
2373 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2374 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2375 * copy was made successfully, or an error code otherwise.
2378 arc_buf_alloc_impl(arc_buf_hdr_t
*hdr
, void *tag
, boolean_t compressed
,
2379 boolean_t fill
, arc_buf_t
**ret
)
2383 ASSERT(HDR_HAS_L1HDR(hdr
));
2384 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2385 VERIFY(hdr
->b_type
== ARC_BUFC_DATA
||
2386 hdr
->b_type
== ARC_BUFC_METADATA
);
2387 ASSERT3P(ret
, !=, NULL
);
2388 ASSERT3P(*ret
, ==, NULL
);
2390 hdr
->b_l1hdr
.b_mru_hits
= 0;
2391 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
2392 hdr
->b_l1hdr
.b_mfu_hits
= 0;
2393 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
2394 hdr
->b_l1hdr
.b_l2_hits
= 0;
2396 buf
= *ret
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
2399 buf
->b_next
= hdr
->b_l1hdr
.b_buf
;
2402 add_reference(hdr
, tag
);
2405 * We're about to change the hdr's b_flags. We must either
2406 * hold the hash_lock or be undiscoverable.
2408 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2411 * Only honor requests for compressed bufs if the hdr is actually
2414 if (compressed
&& HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
)
2415 buf
->b_flags
|= ARC_BUF_FLAG_COMPRESSED
;
2418 * If the hdr's data can be shared then we share the data buffer and
2419 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2420 * allocate a new buffer to store the buf's data.
2422 * There are two additional restrictions here because we're sharing
2423 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2424 * actively involved in an L2ARC write, because if this buf is used by
2425 * an arc_write() then the hdr's data buffer will be released when the
2426 * write completes, even though the L2ARC write might still be using it.
2427 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2428 * need to be ABD-aware.
2430 boolean_t can_share
= arc_can_share(hdr
, buf
) && !HDR_L2_WRITING(hdr
) &&
2431 abd_is_linear(hdr
->b_l1hdr
.b_pabd
);
2433 /* Set up b_data and sharing */
2435 buf
->b_data
= abd_to_buf(hdr
->b_l1hdr
.b_pabd
);
2436 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2437 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2440 arc_get_data_buf(hdr
, arc_buf_size(buf
), buf
);
2441 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2443 VERIFY3P(buf
->b_data
, !=, NULL
);
2445 hdr
->b_l1hdr
.b_buf
= buf
;
2446 hdr
->b_l1hdr
.b_bufcnt
+= 1;
2449 * If the user wants the data from the hdr, we need to either copy or
2450 * decompress the data.
2453 return (arc_buf_fill(buf
, ARC_BUF_COMPRESSED(buf
) != 0));
2459 static char *arc_onloan_tag
= "onloan";
2462 arc_loaned_bytes_update(int64_t delta
)
2464 atomic_add_64(&arc_loaned_bytes
, delta
);
2466 /* assert that it did not wrap around */
2467 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
2471 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2472 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2473 * buffers must be returned to the arc before they can be used by the DMU or
2477 arc_loan_buf(spa_t
*spa
, boolean_t is_metadata
, int size
)
2479 arc_buf_t
*buf
= arc_alloc_buf(spa
, arc_onloan_tag
,
2480 is_metadata
? ARC_BUFC_METADATA
: ARC_BUFC_DATA
, size
);
2482 arc_loaned_bytes_update(size
);
2488 arc_loan_compressed_buf(spa_t
*spa
, uint64_t psize
, uint64_t lsize
,
2489 enum zio_compress compression_type
)
2491 arc_buf_t
*buf
= arc_alloc_compressed_buf(spa
, arc_onloan_tag
,
2492 psize
, lsize
, compression_type
);
2494 arc_loaned_bytes_update(psize
);
2501 * Return a loaned arc buffer to the arc.
2504 arc_return_buf(arc_buf_t
*buf
, void *tag
)
2506 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2508 ASSERT3P(buf
->b_data
, !=, NULL
);
2509 ASSERT(HDR_HAS_L1HDR(hdr
));
2510 (void) zfs_refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2511 (void) zfs_refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2513 arc_loaned_bytes_update(-arc_buf_size(buf
));
2516 /* Detach an arc_buf from a dbuf (tag) */
2518 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
2520 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2522 ASSERT3P(buf
->b_data
, !=, NULL
);
2523 ASSERT(HDR_HAS_L1HDR(hdr
));
2524 (void) zfs_refcount_add(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2525 (void) zfs_refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2527 arc_loaned_bytes_update(arc_buf_size(buf
));
2531 l2arc_free_abd_on_write(abd_t
*abd
, size_t size
, arc_buf_contents_t type
)
2533 l2arc_data_free_t
*df
= kmem_alloc(sizeof (*df
), KM_SLEEP
);
2536 df
->l2df_size
= size
;
2537 df
->l2df_type
= type
;
2538 mutex_enter(&l2arc_free_on_write_mtx
);
2539 list_insert_head(l2arc_free_on_write
, df
);
2540 mutex_exit(&l2arc_free_on_write_mtx
);
2544 arc_hdr_free_on_write(arc_buf_hdr_t
*hdr
)
2546 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2547 arc_buf_contents_t type
= arc_buf_type(hdr
);
2548 uint64_t size
= arc_hdr_size(hdr
);
2550 /* protected by hash lock, if in the hash table */
2551 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
2552 ASSERT(zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2553 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
2555 (void) zfs_refcount_remove_many(&state
->arcs_esize
[type
],
2558 (void) zfs_refcount_remove_many(&state
->arcs_size
, size
, hdr
);
2559 if (type
== ARC_BUFC_METADATA
) {
2560 arc_space_return(size
, ARC_SPACE_META
);
2562 ASSERT(type
== ARC_BUFC_DATA
);
2563 arc_space_return(size
, ARC_SPACE_DATA
);
2566 l2arc_free_abd_on_write(hdr
->b_l1hdr
.b_pabd
, size
, type
);
2570 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2571 * data buffer, we transfer the refcount ownership to the hdr and update
2572 * the appropriate kstats.
2575 arc_share_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2577 ASSERT(arc_can_share(hdr
, buf
));
2578 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2579 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2582 * Start sharing the data buffer. We transfer the
2583 * refcount ownership to the hdr since it always owns
2584 * the refcount whenever an arc_buf_t is shared.
2586 zfs_refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, buf
,
2588 hdr
->b_l1hdr
.b_pabd
= abd_get_from_buf(buf
->b_data
, arc_buf_size(buf
));
2589 abd_take_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
,
2590 HDR_ISTYPE_METADATA(hdr
));
2591 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2592 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2595 * Since we've transferred ownership to the hdr we need
2596 * to increment its compressed and uncompressed kstats and
2597 * decrement the overhead size.
2599 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
2600 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
2601 ARCSTAT_INCR(arcstat_overhead_size
, -arc_buf_size(buf
));
2605 arc_unshare_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2607 ASSERT(arc_buf_is_shared(buf
));
2608 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2609 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2612 * We are no longer sharing this buffer so we need
2613 * to transfer its ownership to the rightful owner.
2615 zfs_refcount_transfer_ownership(&hdr
->b_l1hdr
.b_state
->arcs_size
, hdr
,
2617 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2618 abd_release_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
);
2619 abd_put(hdr
->b_l1hdr
.b_pabd
);
2620 hdr
->b_l1hdr
.b_pabd
= NULL
;
2621 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
2624 * Since the buffer is no longer shared between
2625 * the arc buf and the hdr, count it as overhead.
2627 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
2628 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
2629 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2633 * Remove an arc_buf_t from the hdr's buf list and return the last
2634 * arc_buf_t on the list. If no buffers remain on the list then return
2638 arc_buf_remove(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2640 ASSERT(HDR_HAS_L1HDR(hdr
));
2641 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2643 arc_buf_t
**bufp
= &hdr
->b_l1hdr
.b_buf
;
2644 arc_buf_t
*lastbuf
= NULL
;
2647 * Remove the buf from the hdr list and locate the last
2648 * remaining buffer on the list.
2650 while (*bufp
!= NULL
) {
2652 *bufp
= buf
->b_next
;
2655 * If we've removed a buffer in the middle of
2656 * the list then update the lastbuf and update
2659 if (*bufp
!= NULL
) {
2661 bufp
= &(*bufp
)->b_next
;
2665 ASSERT3P(lastbuf
, !=, buf
);
2666 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, lastbuf
!= NULL
);
2667 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, hdr
->b_l1hdr
.b_buf
!= NULL
);
2668 IMPLY(lastbuf
!= NULL
, ARC_BUF_LAST(lastbuf
));
2674 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
2678 arc_buf_destroy_impl(arc_buf_t
*buf
)
2680 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2683 * Free up the data associated with the buf but only if we're not
2684 * sharing this with the hdr. If we are sharing it with the hdr, the
2685 * hdr is responsible for doing the free.
2687 if (buf
->b_data
!= NULL
) {
2689 * We're about to change the hdr's b_flags. We must either
2690 * hold the hash_lock or be undiscoverable.
2692 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2694 arc_cksum_verify(buf
);
2695 arc_buf_unwatch(buf
);
2697 if (arc_buf_is_shared(buf
)) {
2698 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2700 uint64_t size
= arc_buf_size(buf
);
2701 arc_free_data_buf(hdr
, buf
->b_data
, size
, buf
);
2702 ARCSTAT_INCR(arcstat_overhead_size
, -size
);
2706 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
2707 hdr
->b_l1hdr
.b_bufcnt
-= 1;
2710 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
2712 if (ARC_BUF_SHARED(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
2714 * If the current arc_buf_t is sharing its data buffer with the
2715 * hdr, then reassign the hdr's b_pabd to share it with the new
2716 * buffer at the end of the list. The shared buffer is always
2717 * the last one on the hdr's buffer list.
2719 * There is an equivalent case for compressed bufs, but since
2720 * they aren't guaranteed to be the last buf in the list and
2721 * that is an exceedingly rare case, we just allow that space be
2722 * wasted temporarily.
2724 if (lastbuf
!= NULL
) {
2725 /* Only one buf can be shared at once */
2726 VERIFY(!arc_buf_is_shared(lastbuf
));
2727 /* hdr is uncompressed so can't have compressed buf */
2728 VERIFY(!ARC_BUF_COMPRESSED(lastbuf
));
2730 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2731 arc_hdr_free_pabd(hdr
);
2734 * We must setup a new shared block between the
2735 * last buffer and the hdr. The data would have
2736 * been allocated by the arc buf so we need to transfer
2737 * ownership to the hdr since it's now being shared.
2739 arc_share_buf(hdr
, lastbuf
);
2741 } else if (HDR_SHARED_DATA(hdr
)) {
2743 * Uncompressed shared buffers are always at the end
2744 * of the list. Compressed buffers don't have the
2745 * same requirements. This makes it hard to
2746 * simply assert that the lastbuf is shared so
2747 * we rely on the hdr's compression flags to determine
2748 * if we have a compressed, shared buffer.
2750 ASSERT3P(lastbuf
, !=, NULL
);
2751 ASSERT(arc_buf_is_shared(lastbuf
) ||
2752 HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
2756 * Free the checksum if we're removing the last uncompressed buf from
2759 if (!arc_hdr_has_uncompressed_buf(hdr
)) {
2760 arc_cksum_free(hdr
);
2763 /* clean up the buf */
2765 kmem_cache_free(buf_cache
, buf
);
2769 arc_hdr_alloc_pabd(arc_buf_hdr_t
*hdr
)
2771 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2772 ASSERT(HDR_HAS_L1HDR(hdr
));
2773 ASSERT(!HDR_SHARED_DATA(hdr
));
2775 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2776 hdr
->b_l1hdr
.b_pabd
= arc_get_data_abd(hdr
, arc_hdr_size(hdr
), hdr
);
2777 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
2778 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2780 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
2781 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
2785 arc_hdr_free_pabd(arc_buf_hdr_t
*hdr
)
2787 ASSERT(HDR_HAS_L1HDR(hdr
));
2788 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2791 * If the hdr is currently being written to the l2arc then
2792 * we defer freeing the data by adding it to the l2arc_free_on_write
2793 * list. The l2arc will free the data once it's finished
2794 * writing it to the l2arc device.
2796 if (HDR_L2_WRITING(hdr
)) {
2797 arc_hdr_free_on_write(hdr
);
2798 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
2800 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
,
2801 arc_hdr_size(hdr
), hdr
);
2803 hdr
->b_l1hdr
.b_pabd
= NULL
;
2804 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
2806 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
2807 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
2810 static arc_buf_hdr_t
*
2811 arc_hdr_alloc(uint64_t spa
, int32_t psize
, int32_t lsize
,
2812 enum zio_compress compression_type
, arc_buf_contents_t type
)
2816 VERIFY(type
== ARC_BUFC_DATA
|| type
== ARC_BUFC_METADATA
);
2818 hdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
2819 ASSERT(HDR_EMPTY(hdr
));
2820 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2821 HDR_SET_PSIZE(hdr
, psize
);
2822 HDR_SET_LSIZE(hdr
, lsize
);
2826 arc_hdr_set_flags(hdr
, arc_bufc_to_flags(type
) | ARC_FLAG_HAS_L1HDR
);
2827 arc_hdr_set_compress(hdr
, compression_type
);
2829 hdr
->b_l1hdr
.b_state
= arc_anon
;
2830 hdr
->b_l1hdr
.b_arc_access
= 0;
2831 hdr
->b_l1hdr
.b_bufcnt
= 0;
2832 hdr
->b_l1hdr
.b_buf
= NULL
;
2835 * Allocate the hdr's buffer. This will contain either
2836 * the compressed or uncompressed data depending on the block
2837 * it references and compressed arc enablement.
2839 arc_hdr_alloc_pabd(hdr
);
2840 ASSERT(zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2846 * Transition between the two allocation states for the arc_buf_hdr struct.
2847 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
2848 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
2849 * version is used when a cache buffer is only in the L2ARC in order to reduce
2852 static arc_buf_hdr_t
*
2853 arc_hdr_realloc(arc_buf_hdr_t
*hdr
, kmem_cache_t
*old
, kmem_cache_t
*new)
2855 arc_buf_hdr_t
*nhdr
;
2856 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
2858 ASSERT(HDR_HAS_L2HDR(hdr
));
2859 ASSERT((old
== hdr_full_cache
&& new == hdr_l2only_cache
) ||
2860 (old
== hdr_l2only_cache
&& new == hdr_full_cache
));
2862 nhdr
= kmem_cache_alloc(new, KM_PUSHPAGE
);
2864 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
2865 buf_hash_remove(hdr
);
2867 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
2869 if (new == hdr_full_cache
) {
2870 arc_hdr_set_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
2872 * arc_access and arc_change_state need to be aware that a
2873 * header has just come out of L2ARC, so we set its state to
2874 * l2c_only even though it's about to change.
2876 nhdr
->b_l1hdr
.b_state
= arc_l2c_only
;
2878 /* Verify previous threads set to NULL before freeing */
2879 ASSERT3P(nhdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2881 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2882 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
2883 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2886 * If we've reached here, We must have been called from
2887 * arc_evict_hdr(), as such we should have already been
2888 * removed from any ghost list we were previously on
2889 * (which protects us from racing with arc_evict_state),
2890 * thus no locking is needed during this check.
2892 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
2895 * A buffer must not be moved into the arc_l2c_only
2896 * state if it's not finished being written out to the
2897 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
2898 * might try to be accessed, even though it was removed.
2900 VERIFY(!HDR_L2_WRITING(hdr
));
2901 VERIFY3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2903 arc_hdr_clear_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
2906 * The header has been reallocated so we need to re-insert it into any
2909 (void) buf_hash_insert(nhdr
, NULL
);
2911 ASSERT(list_link_active(&hdr
->b_l2hdr
.b_l2node
));
2913 mutex_enter(&dev
->l2ad_mtx
);
2916 * We must place the realloc'ed header back into the list at
2917 * the same spot. Otherwise, if it's placed earlier in the list,
2918 * l2arc_write_buffers() could find it during the function's
2919 * write phase, and try to write it out to the l2arc.
2921 list_insert_after(&dev
->l2ad_buflist
, hdr
, nhdr
);
2922 list_remove(&dev
->l2ad_buflist
, hdr
);
2924 mutex_exit(&dev
->l2ad_mtx
);
2927 * Since we're using the pointer address as the tag when
2928 * incrementing and decrementing the l2ad_alloc refcount, we
2929 * must remove the old pointer (that we're about to destroy) and
2930 * add the new pointer to the refcount. Otherwise we'd remove
2931 * the wrong pointer address when calling arc_hdr_destroy() later.
2934 (void) zfs_refcount_remove_many(&dev
->l2ad_alloc
, arc_hdr_size(hdr
),
2936 (void) zfs_refcount_add_many(&dev
->l2ad_alloc
, arc_hdr_size(nhdr
),
2939 buf_discard_identity(hdr
);
2940 kmem_cache_free(old
, hdr
);
2946 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
2947 * The buf is returned thawed since we expect the consumer to modify it.
2950 arc_alloc_buf(spa_t
*spa
, void *tag
, arc_buf_contents_t type
, int32_t size
)
2952 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), size
, size
,
2953 ZIO_COMPRESS_OFF
, type
);
2954 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
2956 arc_buf_t
*buf
= NULL
;
2957 VERIFY0(arc_buf_alloc_impl(hdr
, tag
, B_FALSE
, B_FALSE
, &buf
));
2964 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
2965 * for bufs containing metadata.
2968 arc_alloc_compressed_buf(spa_t
*spa
, void *tag
, uint64_t psize
, uint64_t lsize
,
2969 enum zio_compress compression_type
)
2971 ASSERT3U(lsize
, >, 0);
2972 ASSERT3U(lsize
, >=, psize
);
2973 ASSERT(compression_type
> ZIO_COMPRESS_OFF
);
2974 ASSERT(compression_type
< ZIO_COMPRESS_FUNCTIONS
);
2976 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
2977 compression_type
, ARC_BUFC_DATA
);
2978 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
2980 arc_buf_t
*buf
= NULL
;
2981 VERIFY0(arc_buf_alloc_impl(hdr
, tag
, B_TRUE
, B_FALSE
, &buf
));
2983 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2985 if (!arc_buf_is_shared(buf
)) {
2987 * To ensure that the hdr has the correct data in it if we call
2988 * arc_decompress() on this buf before it's been written to
2989 * disk, it's easiest if we just set up sharing between the
2992 ASSERT(!abd_is_linear(hdr
->b_l1hdr
.b_pabd
));
2993 arc_hdr_free_pabd(hdr
);
2994 arc_share_buf(hdr
, buf
);
3001 arc_hdr_l2hdr_destroy(arc_buf_hdr_t
*hdr
)
3003 l2arc_buf_hdr_t
*l2hdr
= &hdr
->b_l2hdr
;
3004 l2arc_dev_t
*dev
= l2hdr
->b_dev
;
3005 uint64_t psize
= arc_hdr_size(hdr
);
3007 ASSERT(MUTEX_HELD(&dev
->l2ad_mtx
));
3008 ASSERT(HDR_HAS_L2HDR(hdr
));
3010 list_remove(&dev
->l2ad_buflist
, hdr
);
3012 ARCSTAT_INCR(arcstat_l2_psize
, -psize
);
3013 ARCSTAT_INCR(arcstat_l2_lsize
, -HDR_GET_LSIZE(hdr
));
3015 vdev_space_update(dev
->l2ad_vdev
, -psize
, 0, 0);
3017 (void) zfs_refcount_remove_many(&dev
->l2ad_alloc
, psize
, hdr
);
3018 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
3022 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
3024 if (HDR_HAS_L1HDR(hdr
)) {
3025 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
||
3026 hdr
->b_l1hdr
.b_bufcnt
> 0);
3027 ASSERT(zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3028 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
3030 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3031 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
3033 if (!HDR_EMPTY(hdr
))
3034 buf_discard_identity(hdr
);
3036 if (HDR_HAS_L2HDR(hdr
)) {
3037 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
3038 boolean_t buflist_held
= MUTEX_HELD(&dev
->l2ad_mtx
);
3041 mutex_enter(&dev
->l2ad_mtx
);
3044 * Even though we checked this conditional above, we
3045 * need to check this again now that we have the
3046 * l2ad_mtx. This is because we could be racing with
3047 * another thread calling l2arc_evict() which might have
3048 * destroyed this header's L2 portion as we were waiting
3049 * to acquire the l2ad_mtx. If that happens, we don't
3050 * want to re-destroy the header's L2 portion.
3052 if (HDR_HAS_L2HDR(hdr
))
3053 arc_hdr_l2hdr_destroy(hdr
);
3056 mutex_exit(&dev
->l2ad_mtx
);
3059 if (HDR_HAS_L1HDR(hdr
)) {
3060 arc_cksum_free(hdr
);
3062 while (hdr
->b_l1hdr
.b_buf
!= NULL
)
3063 arc_buf_destroy_impl(hdr
->b_l1hdr
.b_buf
);
3065 if (hdr
->b_l1hdr
.b_pabd
!= NULL
)
3066 arc_hdr_free_pabd(hdr
);
3069 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
3070 if (HDR_HAS_L1HDR(hdr
)) {
3071 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3072 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
3073 kmem_cache_free(hdr_full_cache
, hdr
);
3075 kmem_cache_free(hdr_l2only_cache
, hdr
);
3080 arc_buf_destroy(arc_buf_t
*buf
, void* tag
)
3082 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3083 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
3085 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
3086 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
3087 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3088 VERIFY0(remove_reference(hdr
, NULL
, tag
));
3089 arc_hdr_destroy(hdr
);
3093 mutex_enter(hash_lock
);
3094 ASSERT3P(hdr
, ==, buf
->b_hdr
);
3095 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
3096 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3097 ASSERT3P(hdr
->b_l1hdr
.b_state
, !=, arc_anon
);
3098 ASSERT3P(buf
->b_data
, !=, NULL
);
3100 (void) remove_reference(hdr
, hash_lock
, tag
);
3101 arc_buf_destroy_impl(buf
);
3102 mutex_exit(hash_lock
);
3106 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3107 * state of the header is dependent on its state prior to entering this
3108 * function. The following transitions are possible:
3110 * - arc_mru -> arc_mru_ghost
3111 * - arc_mfu -> arc_mfu_ghost
3112 * - arc_mru_ghost -> arc_l2c_only
3113 * - arc_mru_ghost -> deleted
3114 * - arc_mfu_ghost -> arc_l2c_only
3115 * - arc_mfu_ghost -> deleted
3118 arc_evict_hdr(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
3120 arc_state_t
*evicted_state
, *state
;
3121 int64_t bytes_evicted
= 0;
3123 ASSERT(MUTEX_HELD(hash_lock
));
3124 ASSERT(HDR_HAS_L1HDR(hdr
));
3126 state
= hdr
->b_l1hdr
.b_state
;
3127 if (GHOST_STATE(state
)) {
3128 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3129 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
3132 * l2arc_write_buffers() relies on a header's L1 portion
3133 * (i.e. its b_pabd field) during it's write phase.
3134 * Thus, we cannot push a header onto the arc_l2c_only
3135 * state (removing its L1 piece) until the header is
3136 * done being written to the l2arc.
3138 if (HDR_HAS_L2HDR(hdr
) && HDR_L2_WRITING(hdr
)) {
3139 ARCSTAT_BUMP(arcstat_evict_l2_skip
);
3140 return (bytes_evicted
);
3143 ARCSTAT_BUMP(arcstat_deleted
);
3144 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3146 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, hdr
);
3148 if (HDR_HAS_L2HDR(hdr
)) {
3149 ASSERT(hdr
->b_l1hdr
.b_pabd
== NULL
);
3151 * This buffer is cached on the 2nd Level ARC;
3152 * don't destroy the header.
3154 arc_change_state(arc_l2c_only
, hdr
, hash_lock
);
3156 * dropping from L1+L2 cached to L2-only,
3157 * realloc to remove the L1 header.
3159 hdr
= arc_hdr_realloc(hdr
, hdr_full_cache
,
3162 arc_change_state(arc_anon
, hdr
, hash_lock
);
3163 arc_hdr_destroy(hdr
);
3165 return (bytes_evicted
);
3168 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
3169 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
3171 /* prefetch buffers have a minimum lifespan */
3172 if (HDR_IO_IN_PROGRESS(hdr
) ||
3173 ((hdr
->b_flags
& (ARC_FLAG_PREFETCH
| ARC_FLAG_INDIRECT
)) &&
3174 ddi_get_lbolt() - hdr
->b_l1hdr
.b_arc_access
<
3175 arc_min_prefetch_lifespan
)) {
3176 ARCSTAT_BUMP(arcstat_evict_skip
);
3177 return (bytes_evicted
);
3180 ASSERT0(zfs_refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
3181 while (hdr
->b_l1hdr
.b_buf
) {
3182 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
3183 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
3184 ARCSTAT_BUMP(arcstat_mutex_miss
);
3187 if (buf
->b_data
!= NULL
)
3188 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3189 mutex_exit(&buf
->b_evict_lock
);
3190 arc_buf_destroy_impl(buf
);
3193 if (HDR_HAS_L2HDR(hdr
)) {
3194 ARCSTAT_INCR(arcstat_evict_l2_cached
, HDR_GET_LSIZE(hdr
));
3196 if (l2arc_write_eligible(hdr
->b_spa
, hdr
)) {
3197 ARCSTAT_INCR(arcstat_evict_l2_eligible
,
3198 HDR_GET_LSIZE(hdr
));
3200 ARCSTAT_INCR(arcstat_evict_l2_ineligible
,
3201 HDR_GET_LSIZE(hdr
));
3205 if (hdr
->b_l1hdr
.b_bufcnt
== 0) {
3206 arc_cksum_free(hdr
);
3208 bytes_evicted
+= arc_hdr_size(hdr
);
3211 * If this hdr is being evicted and has a compressed
3212 * buffer then we discard it here before we change states.
3213 * This ensures that the accounting is updated correctly
3214 * in arc_free_data_impl().
3216 arc_hdr_free_pabd(hdr
);
3218 arc_change_state(evicted_state
, hdr
, hash_lock
);
3219 ASSERT(HDR_IN_HASH_TABLE(hdr
));
3220 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
3221 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, hdr
);
3224 return (bytes_evicted
);
3228 arc_evict_state_impl(multilist_t
*ml
, int idx
, arc_buf_hdr_t
*marker
,
3229 uint64_t spa
, int64_t bytes
)
3231 multilist_sublist_t
*mls
;
3232 uint64_t bytes_evicted
= 0;
3234 kmutex_t
*hash_lock
;
3235 int evict_count
= 0;
3237 ASSERT3P(marker
, !=, NULL
);
3238 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3240 mls
= multilist_sublist_lock(ml
, idx
);
3242 for (hdr
= multilist_sublist_prev(mls
, marker
); hdr
!= NULL
;
3243 hdr
= multilist_sublist_prev(mls
, marker
)) {
3244 if ((bytes
!= ARC_EVICT_ALL
&& bytes_evicted
>= bytes
) ||
3245 (evict_count
>= zfs_arc_evict_batch_limit
))
3249 * To keep our iteration location, move the marker
3250 * forward. Since we're not holding hdr's hash lock, we
3251 * must be very careful and not remove 'hdr' from the
3252 * sublist. Otherwise, other consumers might mistake the
3253 * 'hdr' as not being on a sublist when they call the
3254 * multilist_link_active() function (they all rely on
3255 * the hash lock protecting concurrent insertions and
3256 * removals). multilist_sublist_move_forward() was
3257 * specifically implemented to ensure this is the case
3258 * (only 'marker' will be removed and re-inserted).
3260 multilist_sublist_move_forward(mls
, marker
);
3263 * The only case where the b_spa field should ever be
3264 * zero, is the marker headers inserted by
3265 * arc_evict_state(). It's possible for multiple threads
3266 * to be calling arc_evict_state() concurrently (e.g.
3267 * dsl_pool_close() and zio_inject_fault()), so we must
3268 * skip any markers we see from these other threads.
3270 if (hdr
->b_spa
== 0)
3273 /* we're only interested in evicting buffers of a certain spa */
3274 if (spa
!= 0 && hdr
->b_spa
!= spa
) {
3275 ARCSTAT_BUMP(arcstat_evict_skip
);
3279 hash_lock
= HDR_LOCK(hdr
);
3282 * We aren't calling this function from any code path
3283 * that would already be holding a hash lock, so we're
3284 * asserting on this assumption to be defensive in case
3285 * this ever changes. Without this check, it would be
3286 * possible to incorrectly increment arcstat_mutex_miss
3287 * below (e.g. if the code changed such that we called
3288 * this function with a hash lock held).
3290 ASSERT(!MUTEX_HELD(hash_lock
));
3292 if (mutex_tryenter(hash_lock
)) {
3293 uint64_t evicted
= arc_evict_hdr(hdr
, hash_lock
);
3294 mutex_exit(hash_lock
);
3296 bytes_evicted
+= evicted
;
3299 * If evicted is zero, arc_evict_hdr() must have
3300 * decided to skip this header, don't increment
3301 * evict_count in this case.
3307 * If arc_size isn't overflowing, signal any
3308 * threads that might happen to be waiting.
3310 * For each header evicted, we wake up a single
3311 * thread. If we used cv_broadcast, we could
3312 * wake up "too many" threads causing arc_size
3313 * to significantly overflow arc_c; since
3314 * arc_get_data_impl() doesn't check for overflow
3315 * when it's woken up (it doesn't because it's
3316 * possible for the ARC to be overflowing while
3317 * full of un-evictable buffers, and the
3318 * function should proceed in this case).
3320 * If threads are left sleeping, due to not
3321 * using cv_broadcast, they will be woken up
3322 * just before arc_reclaim_thread() sleeps.
3324 mutex_enter(&arc_reclaim_lock
);
3325 if (!arc_is_overflowing())
3326 cv_signal(&arc_reclaim_waiters_cv
);
3327 mutex_exit(&arc_reclaim_lock
);
3329 ARCSTAT_BUMP(arcstat_mutex_miss
);
3333 multilist_sublist_unlock(mls
);
3335 return (bytes_evicted
);
3339 * Evict buffers from the given arc state, until we've removed the
3340 * specified number of bytes. Move the removed buffers to the
3341 * appropriate evict state.
3343 * This function makes a "best effort". It skips over any buffers
3344 * it can't get a hash_lock on, and so, may not catch all candidates.
3345 * It may also return without evicting as much space as requested.
3347 * If bytes is specified using the special value ARC_EVICT_ALL, this
3348 * will evict all available (i.e. unlocked and evictable) buffers from
3349 * the given arc state; which is used by arc_flush().
3352 arc_evict_state(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
3353 arc_buf_contents_t type
)
3355 uint64_t total_evicted
= 0;
3356 multilist_t
*ml
= state
->arcs_list
[type
];
3358 arc_buf_hdr_t
**markers
;
3361 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3363 num_sublists
= multilist_get_num_sublists(ml
);
3366 * If we've tried to evict from each sublist, made some
3367 * progress, but still have not hit the target number of bytes
3368 * to evict, we want to keep trying. The markers allow us to
3369 * pick up where we left off for each individual sublist, rather
3370 * than starting from the tail each time.
3372 markers
= kmem_zalloc(sizeof (*markers
) * num_sublists
, KM_SLEEP
);
3373 for (i
= 0; i
< num_sublists
; i
++) {
3374 multilist_sublist_t
*mls
;
3376 markers
[i
] = kmem_cache_alloc(hdr_full_cache
, KM_SLEEP
);
3379 * A b_spa of 0 is used to indicate that this header is
3380 * a marker. This fact is used in arc_adjust_type() and
3381 * arc_evict_state_impl().
3383 markers
[i
]->b_spa
= 0;
3385 mls
= multilist_sublist_lock(ml
, i
);
3386 multilist_sublist_insert_tail(mls
, markers
[i
]);
3387 multilist_sublist_unlock(mls
);
3391 * While we haven't hit our target number of bytes to evict, or
3392 * we're evicting all available buffers.
3394 while (total_evicted
< bytes
|| bytes
== ARC_EVICT_ALL
) {
3395 int sublist_idx
= multilist_get_random_index(ml
);
3396 uint64_t scan_evicted
= 0;
3399 * Try to reduce pinned dnodes with a floor of arc_dnode_limit.
3400 * Request that 10% of the LRUs be scanned by the superblock
3403 if (type
== ARC_BUFC_DATA
&& arc_dnode_size
> arc_dnode_limit
)
3404 arc_prune_async((arc_dnode_size
- arc_dnode_limit
) /
3405 sizeof (dnode_t
) / zfs_arc_dnode_reduce_percent
);
3408 * Start eviction using a randomly selected sublist,
3409 * this is to try and evenly balance eviction across all
3410 * sublists. Always starting at the same sublist
3411 * (e.g. index 0) would cause evictions to favor certain
3412 * sublists over others.
3414 for (i
= 0; i
< num_sublists
; i
++) {
3415 uint64_t bytes_remaining
;
3416 uint64_t bytes_evicted
;
3418 if (bytes
== ARC_EVICT_ALL
)
3419 bytes_remaining
= ARC_EVICT_ALL
;
3420 else if (total_evicted
< bytes
)
3421 bytes_remaining
= bytes
- total_evicted
;
3425 bytes_evicted
= arc_evict_state_impl(ml
, sublist_idx
,
3426 markers
[sublist_idx
], spa
, bytes_remaining
);
3428 scan_evicted
+= bytes_evicted
;
3429 total_evicted
+= bytes_evicted
;
3431 /* we've reached the end, wrap to the beginning */
3432 if (++sublist_idx
>= num_sublists
)
3437 * If we didn't evict anything during this scan, we have
3438 * no reason to believe we'll evict more during another
3439 * scan, so break the loop.
3441 if (scan_evicted
== 0) {
3442 /* This isn't possible, let's make that obvious */
3443 ASSERT3S(bytes
, !=, 0);
3446 * When bytes is ARC_EVICT_ALL, the only way to
3447 * break the loop is when scan_evicted is zero.
3448 * In that case, we actually have evicted enough,
3449 * so we don't want to increment the kstat.
3451 if (bytes
!= ARC_EVICT_ALL
) {
3452 ASSERT3S(total_evicted
, <, bytes
);
3453 ARCSTAT_BUMP(arcstat_evict_not_enough
);
3460 for (i
= 0; i
< num_sublists
; i
++) {
3461 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
3462 multilist_sublist_remove(mls
, markers
[i
]);
3463 multilist_sublist_unlock(mls
);
3465 kmem_cache_free(hdr_full_cache
, markers
[i
]);
3467 kmem_free(markers
, sizeof (*markers
) * num_sublists
);
3469 return (total_evicted
);
3473 * Flush all "evictable" data of the given type from the arc state
3474 * specified. This will not evict any "active" buffers (i.e. referenced).
3476 * When 'retry' is set to B_FALSE, the function will make a single pass
3477 * over the state and evict any buffers that it can. Since it doesn't
3478 * continually retry the eviction, it might end up leaving some buffers
3479 * in the ARC due to lock misses.
3481 * When 'retry' is set to B_TRUE, the function will continually retry the
3482 * eviction until *all* evictable buffers have been removed from the
3483 * state. As a result, if concurrent insertions into the state are
3484 * allowed (e.g. if the ARC isn't shutting down), this function might
3485 * wind up in an infinite loop, continually trying to evict buffers.
3488 arc_flush_state(arc_state_t
*state
, uint64_t spa
, arc_buf_contents_t type
,
3491 uint64_t evicted
= 0;
3493 while (zfs_refcount_count(&state
->arcs_esize
[type
]) != 0) {
3494 evicted
+= arc_evict_state(state
, spa
, ARC_EVICT_ALL
, type
);
3504 * Helper function for arc_prune_async() it is responsible for safely
3505 * handling the execution of a registered arc_prune_func_t.
3508 arc_prune_task(void *ptr
)
3510 arc_prune_t
*ap
= (arc_prune_t
*)ptr
;
3511 arc_prune_func_t
*func
= ap
->p_pfunc
;
3514 func(ap
->p_adjust
, ap
->p_private
);
3516 zfs_refcount_remove(&ap
->p_refcnt
, func
);
3520 * Notify registered consumers they must drop holds on a portion of the ARC
3521 * buffered they reference. This provides a mechanism to ensure the ARC can
3522 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
3523 * is analogous to dnlc_reduce_cache() but more generic.
3525 * This operation is performed asynchronously so it may be safely called
3526 * in the context of the arc_reclaim_thread(). A reference is taken here
3527 * for each registered arc_prune_t and the arc_prune_task() is responsible
3528 * for releasing it once the registered arc_prune_func_t has completed.
3531 arc_prune_async(int64_t adjust
)
3535 mutex_enter(&arc_prune_mtx
);
3536 for (ap
= list_head(&arc_prune_list
); ap
!= NULL
;
3537 ap
= list_next(&arc_prune_list
, ap
)) {
3539 if (zfs_refcount_count(&ap
->p_refcnt
) >= 2)
3542 zfs_refcount_add(&ap
->p_refcnt
, ap
->p_pfunc
);
3543 ap
->p_adjust
= adjust
;
3544 if (taskq_dispatch(arc_prune_taskq
, arc_prune_task
,
3545 ap
, TQ_SLEEP
) == TASKQID_INVALID
) {
3546 zfs_refcount_remove(&ap
->p_refcnt
, ap
->p_pfunc
);
3549 ARCSTAT_BUMP(arcstat_prune
);
3551 mutex_exit(&arc_prune_mtx
);
3555 * Evict the specified number of bytes from the state specified,
3556 * restricting eviction to the spa and type given. This function
3557 * prevents us from trying to evict more from a state's list than
3558 * is "evictable", and to skip evicting altogether when passed a
3559 * negative value for "bytes". In contrast, arc_evict_state() will
3560 * evict everything it can, when passed a negative value for "bytes".
3563 arc_adjust_impl(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
3564 arc_buf_contents_t type
)
3568 if (bytes
> 0 && zfs_refcount_count(&state
->arcs_esize
[type
]) > 0) {
3569 delta
= MIN(zfs_refcount_count(&state
->arcs_esize
[type
]),
3571 return (arc_evict_state(state
, spa
, delta
, type
));
3578 * The goal of this function is to evict enough meta data buffers from the
3579 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
3580 * more complicated than it appears because it is common for data buffers
3581 * to have holds on meta data buffers. In addition, dnode meta data buffers
3582 * will be held by the dnodes in the block preventing them from being freed.
3583 * This means we can't simply traverse the ARC and expect to always find
3584 * enough unheld meta data buffer to release.
3586 * Therefore, this function has been updated to make alternating passes
3587 * over the ARC releasing data buffers and then newly unheld meta data
3588 * buffers. This ensures forward progress is maintained and arc_meta_used
3589 * will decrease. Normally this is sufficient, but if required the ARC
3590 * will call the registered prune callbacks causing dentry and inodes to
3591 * be dropped from the VFS cache. This will make dnode meta data buffers
3592 * available for reclaim.
3595 arc_adjust_meta_balanced(void)
3597 int64_t delta
, prune
= 0, adjustmnt
;
3598 uint64_t total_evicted
= 0;
3599 arc_buf_contents_t type
= ARC_BUFC_DATA
;
3600 int restarts
= MAX(zfs_arc_meta_adjust_restarts
, 0);
3604 * This slightly differs than the way we evict from the mru in
3605 * arc_adjust because we don't have a "target" value (i.e. no
3606 * "meta" arc_p). As a result, I think we can completely
3607 * cannibalize the metadata in the MRU before we evict the
3608 * metadata from the MFU. I think we probably need to implement a
3609 * "metadata arc_p" value to do this properly.
3611 adjustmnt
= arc_meta_used
- arc_meta_limit
;
3613 if (adjustmnt
> 0 &&
3614 zfs_refcount_count(&arc_mru
->arcs_esize
[type
]) > 0) {
3615 delta
= MIN(zfs_refcount_count(&arc_mru
->arcs_esize
[type
]),
3617 total_evicted
+= arc_adjust_impl(arc_mru
, 0, delta
, type
);
3622 * We can't afford to recalculate adjustmnt here. If we do,
3623 * new metadata buffers can sneak into the MRU or ANON lists,
3624 * thus penalize the MFU metadata. Although the fudge factor is
3625 * small, it has been empirically shown to be significant for
3626 * certain workloads (e.g. creating many empty directories). As
3627 * such, we use the original calculation for adjustmnt, and
3628 * simply decrement the amount of data evicted from the MRU.
3631 if (adjustmnt
> 0 &&
3632 zfs_refcount_count(&arc_mfu
->arcs_esize
[type
]) > 0) {
3633 delta
= MIN(zfs_refcount_count(&arc_mfu
->arcs_esize
[type
]),
3635 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, delta
, type
);
3638 adjustmnt
= arc_meta_used
- arc_meta_limit
;
3640 if (adjustmnt
> 0 &&
3641 zfs_refcount_count(&arc_mru_ghost
->arcs_esize
[type
]) > 0) {
3642 delta
= MIN(adjustmnt
,
3643 zfs_refcount_count(&arc_mru_ghost
->arcs_esize
[type
]));
3644 total_evicted
+= arc_adjust_impl(arc_mru_ghost
, 0, delta
, type
);
3648 if (adjustmnt
> 0 &&
3649 zfs_refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]) > 0) {
3650 delta
= MIN(adjustmnt
,
3651 zfs_refcount_count(&arc_mfu_ghost
->arcs_esize
[type
]));
3652 total_evicted
+= arc_adjust_impl(arc_mfu_ghost
, 0, delta
, type
);
3656 * If after attempting to make the requested adjustment to the ARC
3657 * the meta limit is still being exceeded then request that the
3658 * higher layers drop some cached objects which have holds on ARC
3659 * meta buffers. Requests to the upper layers will be made with
3660 * increasingly large scan sizes until the ARC is below the limit.
3662 if (arc_meta_used
> arc_meta_limit
) {
3663 if (type
== ARC_BUFC_DATA
) {
3664 type
= ARC_BUFC_METADATA
;
3666 type
= ARC_BUFC_DATA
;
3668 if (zfs_arc_meta_prune
) {
3669 prune
+= zfs_arc_meta_prune
;
3670 arc_prune_async(prune
);
3679 return (total_evicted
);
3683 * Evict metadata buffers from the cache, such that arc_meta_used is
3684 * capped by the arc_meta_limit tunable.
3687 arc_adjust_meta_only(void)
3689 uint64_t total_evicted
= 0;
3693 * If we're over the meta limit, we want to evict enough
3694 * metadata to get back under the meta limit. We don't want to
3695 * evict so much that we drop the MRU below arc_p, though. If
3696 * we're over the meta limit more than we're over arc_p, we
3697 * evict some from the MRU here, and some from the MFU below.
3699 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
3700 (int64_t)(zfs_refcount_count(&arc_anon
->arcs_size
) +
3701 zfs_refcount_count(&arc_mru
->arcs_size
) - arc_p
));
3703 total_evicted
+= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3706 * Similar to the above, we want to evict enough bytes to get us
3707 * below the meta limit, but not so much as to drop us below the
3708 * space allotted to the MFU (which is defined as arc_c - arc_p).
3710 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
3711 (int64_t)(zfs_refcount_count(&arc_mfu
->arcs_size
) - (arc_c
-
3714 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3716 return (total_evicted
);
3720 arc_adjust_meta(void)
3722 if (zfs_arc_meta_strategy
== ARC_STRATEGY_META_ONLY
)
3723 return (arc_adjust_meta_only());
3725 return (arc_adjust_meta_balanced());
3729 * Return the type of the oldest buffer in the given arc state
3731 * This function will select a random sublist of type ARC_BUFC_DATA and
3732 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3733 * is compared, and the type which contains the "older" buffer will be
3736 static arc_buf_contents_t
3737 arc_adjust_type(arc_state_t
*state
)
3739 multilist_t
*data_ml
= state
->arcs_list
[ARC_BUFC_DATA
];
3740 multilist_t
*meta_ml
= state
->arcs_list
[ARC_BUFC_METADATA
];
3741 int data_idx
= multilist_get_random_index(data_ml
);
3742 int meta_idx
= multilist_get_random_index(meta_ml
);
3743 multilist_sublist_t
*data_mls
;
3744 multilist_sublist_t
*meta_mls
;
3745 arc_buf_contents_t type
;
3746 arc_buf_hdr_t
*data_hdr
;
3747 arc_buf_hdr_t
*meta_hdr
;
3750 * We keep the sublist lock until we're finished, to prevent
3751 * the headers from being destroyed via arc_evict_state().
3753 data_mls
= multilist_sublist_lock(data_ml
, data_idx
);
3754 meta_mls
= multilist_sublist_lock(meta_ml
, meta_idx
);
3757 * These two loops are to ensure we skip any markers that
3758 * might be at the tail of the lists due to arc_evict_state().
3761 for (data_hdr
= multilist_sublist_tail(data_mls
); data_hdr
!= NULL
;
3762 data_hdr
= multilist_sublist_prev(data_mls
, data_hdr
)) {
3763 if (data_hdr
->b_spa
!= 0)
3767 for (meta_hdr
= multilist_sublist_tail(meta_mls
); meta_hdr
!= NULL
;
3768 meta_hdr
= multilist_sublist_prev(meta_mls
, meta_hdr
)) {
3769 if (meta_hdr
->b_spa
!= 0)
3773 if (data_hdr
== NULL
&& meta_hdr
== NULL
) {
3774 type
= ARC_BUFC_DATA
;
3775 } else if (data_hdr
== NULL
) {
3776 ASSERT3P(meta_hdr
, !=, NULL
);
3777 type
= ARC_BUFC_METADATA
;
3778 } else if (meta_hdr
== NULL
) {
3779 ASSERT3P(data_hdr
, !=, NULL
);
3780 type
= ARC_BUFC_DATA
;
3782 ASSERT3P(data_hdr
, !=, NULL
);
3783 ASSERT3P(meta_hdr
, !=, NULL
);
3785 /* The headers can't be on the sublist without an L1 header */
3786 ASSERT(HDR_HAS_L1HDR(data_hdr
));
3787 ASSERT(HDR_HAS_L1HDR(meta_hdr
));
3789 if (data_hdr
->b_l1hdr
.b_arc_access
<
3790 meta_hdr
->b_l1hdr
.b_arc_access
) {
3791 type
= ARC_BUFC_DATA
;
3793 type
= ARC_BUFC_METADATA
;
3797 multilist_sublist_unlock(meta_mls
);
3798 multilist_sublist_unlock(data_mls
);
3804 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3809 uint64_t total_evicted
= 0;
3814 * If we're over arc_meta_limit, we want to correct that before
3815 * potentially evicting data buffers below.
3817 total_evicted
+= arc_adjust_meta();
3822 * If we're over the target cache size, we want to evict enough
3823 * from the list to get back to our target size. We don't want
3824 * to evict too much from the MRU, such that it drops below
3825 * arc_p. So, if we're over our target cache size more than
3826 * the MRU is over arc_p, we'll evict enough to get back to
3827 * arc_p here, and then evict more from the MFU below.
3829 target
= MIN((int64_t)(arc_size
- arc_c
),
3830 (int64_t)(zfs_refcount_count(&arc_anon
->arcs_size
) +
3831 zfs_refcount_count(&arc_mru
->arcs_size
) + arc_meta_used
- arc_p
));
3834 * If we're below arc_meta_min, always prefer to evict data.
3835 * Otherwise, try to satisfy the requested number of bytes to
3836 * evict from the type which contains older buffers; in an
3837 * effort to keep newer buffers in the cache regardless of their
3838 * type. If we cannot satisfy the number of bytes from this
3839 * type, spill over into the next type.
3841 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
3842 arc_meta_used
> arc_meta_min
) {
3843 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3844 total_evicted
+= bytes
;
3847 * If we couldn't evict our target number of bytes from
3848 * metadata, we try to get the rest from data.
3853 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3855 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3856 total_evicted
+= bytes
;
3859 * If we couldn't evict our target number of bytes from
3860 * data, we try to get the rest from metadata.
3865 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3871 * Now that we've tried to evict enough from the MRU to get its
3872 * size back to arc_p, if we're still above the target cache
3873 * size, we evict the rest from the MFU.
3875 target
= arc_size
- arc_c
;
3877 if (arc_adjust_type(arc_mfu
) == ARC_BUFC_METADATA
&&
3878 arc_meta_used
> arc_meta_min
) {
3879 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3880 total_evicted
+= bytes
;
3883 * If we couldn't evict our target number of bytes from
3884 * metadata, we try to get the rest from data.
3889 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3891 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3892 total_evicted
+= bytes
;
3895 * If we couldn't evict our target number of bytes from
3896 * data, we try to get the rest from data.
3901 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3905 * Adjust ghost lists
3907 * In addition to the above, the ARC also defines target values
3908 * for the ghost lists. The sum of the mru list and mru ghost
3909 * list should never exceed the target size of the cache, and
3910 * the sum of the mru list, mfu list, mru ghost list, and mfu
3911 * ghost list should never exceed twice the target size of the
3912 * cache. The following logic enforces these limits on the ghost
3913 * caches, and evicts from them as needed.
3915 target
= zfs_refcount_count(&arc_mru
->arcs_size
) +
3916 zfs_refcount_count(&arc_mru_ghost
->arcs_size
) - arc_c
;
3918 bytes
= arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_DATA
);
3919 total_evicted
+= bytes
;
3924 arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_METADATA
);
3927 * We assume the sum of the mru list and mfu list is less than
3928 * or equal to arc_c (we enforced this above), which means we
3929 * can use the simpler of the two equations below:
3931 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3932 * mru ghost + mfu ghost <= arc_c
3934 target
= zfs_refcount_count(&arc_mru_ghost
->arcs_size
) +
3935 zfs_refcount_count(&arc_mfu_ghost
->arcs_size
) - arc_c
;
3937 bytes
= arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_DATA
);
3938 total_evicted
+= bytes
;
3943 arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_METADATA
);
3945 return (total_evicted
);
3949 arc_flush(spa_t
*spa
, boolean_t retry
)
3954 * If retry is B_TRUE, a spa must not be specified since we have
3955 * no good way to determine if all of a spa's buffers have been
3956 * evicted from an arc state.
3958 ASSERT(!retry
|| spa
== 0);
3961 guid
= spa_load_guid(spa
);
3963 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_DATA
, retry
);
3964 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_METADATA
, retry
);
3966 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_DATA
, retry
);
3967 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_METADATA
, retry
);
3969 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3970 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3972 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3973 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3977 arc_shrink(int64_t to_free
)
3981 if (c
> to_free
&& c
- to_free
> arc_c_min
) {
3982 arc_c
= c
- to_free
;
3983 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
3984 if (arc_c
> arc_size
)
3985 arc_c
= MAX(arc_size
, arc_c_min
);
3987 arc_p
= (arc_c
>> 1);
3988 ASSERT(arc_c
>= arc_c_min
);
3989 ASSERT((int64_t)arc_p
>= 0);
3994 if (arc_size
> arc_c
)
3995 (void) arc_adjust();
3999 * Return maximum amount of memory that we could possibly use. Reduced
4000 * to half of all memory in user space which is primarily used for testing.
4003 arc_all_memory(void)
4006 #ifdef CONFIG_HIGHMEM
4007 return (ptob(zfs_totalram_pages
- totalhigh_pages
));
4009 return (ptob(zfs_totalram_pages
));
4010 #endif /* CONFIG_HIGHMEM */
4012 return (ptob(physmem
) / 2);
4013 #endif /* _KERNEL */
4017 * Return the amount of memory that is considered free. In user space
4018 * which is primarily used for testing we pretend that free memory ranges
4019 * from 0-20% of all memory.
4022 arc_free_memory(void)
4025 #ifdef CONFIG_HIGHMEM
4028 return (ptob(si
.freeram
- si
.freehigh
));
4030 return (ptob(nr_free_pages() +
4031 nr_inactive_file_pages() +
4032 nr_inactive_anon_pages() +
4033 nr_slab_reclaimable_pages()));
4035 #endif /* CONFIG_HIGHMEM */
4037 return (spa_get_random(arc_all_memory() * 20 / 100));
4038 #endif /* _KERNEL */
4041 typedef enum free_memory_reason_t
{
4046 FMR_PAGES_PP_MAXIMUM
,
4049 } free_memory_reason_t
;
4051 int64_t last_free_memory
;
4052 free_memory_reason_t last_free_reason
;
4056 * Additional reserve of pages for pp_reserve.
4058 int64_t arc_pages_pp_reserve
= 64;
4061 * Additional reserve of pages for swapfs.
4063 int64_t arc_swapfs_reserve
= 64;
4064 #endif /* _KERNEL */
4067 * Return the amount of memory that can be consumed before reclaim will be
4068 * needed. Positive if there is sufficient free memory, negative indicates
4069 * the amount of memory that needs to be freed up.
4072 arc_available_memory(void)
4074 int64_t lowest
= INT64_MAX
;
4075 free_memory_reason_t r
= FMR_UNKNOWN
;
4082 pgcnt_t needfree
= btop(arc_need_free
);
4083 pgcnt_t lotsfree
= btop(arc_sys_free
);
4084 pgcnt_t desfree
= 0;
4085 pgcnt_t freemem
= btop(arc_free_memory());
4089 n
= PAGESIZE
* (-needfree
);
4097 * check that we're out of range of the pageout scanner. It starts to
4098 * schedule paging if freemem is less than lotsfree and needfree.
4099 * lotsfree is the high-water mark for pageout, and needfree is the
4100 * number of needed free pages. We add extra pages here to make sure
4101 * the scanner doesn't start up while we're freeing memory.
4103 n
= PAGESIZE
* (freemem
- lotsfree
- needfree
- desfree
);
4111 * check to make sure that swapfs has enough space so that anon
4112 * reservations can still succeed. anon_resvmem() checks that the
4113 * availrmem is greater than swapfs_minfree, and the number of reserved
4114 * swap pages. We also add a bit of extra here just to prevent
4115 * circumstances from getting really dire.
4117 n
= PAGESIZE
* (availrmem
- swapfs_minfree
- swapfs_reserve
-
4118 desfree
- arc_swapfs_reserve
);
4121 r
= FMR_SWAPFS_MINFREE
;
4125 * Check that we have enough availrmem that memory locking (e.g., via
4126 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4127 * stores the number of pages that cannot be locked; when availrmem
4128 * drops below pages_pp_maximum, page locking mechanisms such as
4129 * page_pp_lock() will fail.)
4131 n
= PAGESIZE
* (availrmem
- pages_pp_maximum
-
4132 arc_pages_pp_reserve
);
4135 r
= FMR_PAGES_PP_MAXIMUM
;
4141 * If we're on a 32-bit platform, it's possible that we'll exhaust the
4142 * kernel heap space before we ever run out of available physical
4143 * memory. Most checks of the size of the heap_area compare against
4144 * tune.t_minarmem, which is the minimum available real memory that we
4145 * can have in the system. However, this is generally fixed at 25 pages
4146 * which is so low that it's useless. In this comparison, we seek to
4147 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4148 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4151 n
= vmem_size(heap_arena
, VMEM_FREE
) -
4152 (vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
) >> 2);
4160 * If zio data pages are being allocated out of a separate heap segment,
4161 * then enforce that the size of available vmem for this arena remains
4162 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4164 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4165 * memory (in the zio_arena) free, which can avoid memory
4166 * fragmentation issues.
4168 if (zio_arena
!= NULL
) {
4169 n
= (int64_t)vmem_size(zio_arena
, VMEM_FREE
) -
4170 (vmem_size(zio_arena
, VMEM_ALLOC
) >>
4171 arc_zio_arena_free_shift
);
4178 /* Every 100 calls, free a small amount */
4179 if (spa_get_random(100) == 0)
4181 #endif /* _KERNEL */
4183 last_free_memory
= lowest
;
4184 last_free_reason
= r
;
4190 * Determine if the system is under memory pressure and is asking
4191 * to reclaim memory. A return value of B_TRUE indicates that the system
4192 * is under memory pressure and that the arc should adjust accordingly.
4195 arc_reclaim_needed(void)
4197 return (arc_available_memory() < 0);
4201 arc_kmem_reap_now(void)
4204 kmem_cache_t
*prev_cache
= NULL
;
4205 kmem_cache_t
*prev_data_cache
= NULL
;
4206 extern kmem_cache_t
*zio_buf_cache
[];
4207 extern kmem_cache_t
*zio_data_buf_cache
[];
4208 extern kmem_cache_t
*range_seg_cache
;
4211 if ((arc_meta_used
>= arc_meta_limit
) && zfs_arc_meta_prune
) {
4213 * We are exceeding our meta-data cache limit.
4214 * Prune some entries to release holds on meta-data.
4216 arc_prune_async(zfs_arc_meta_prune
);
4220 * Reclaim unused memory from all kmem caches.
4226 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
4228 /* reach upper limit of cache size on 32-bit */
4229 if (zio_buf_cache
[i
] == NULL
)
4232 if (zio_buf_cache
[i
] != prev_cache
) {
4233 prev_cache
= zio_buf_cache
[i
];
4234 kmem_cache_reap_now(zio_buf_cache
[i
]);
4236 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
4237 prev_data_cache
= zio_data_buf_cache
[i
];
4238 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
4241 kmem_cache_reap_now(buf_cache
);
4242 kmem_cache_reap_now(hdr_full_cache
);
4243 kmem_cache_reap_now(hdr_l2only_cache
);
4244 kmem_cache_reap_now(range_seg_cache
);
4246 if (zio_arena
!= NULL
) {
4248 * Ask the vmem arena to reclaim unused memory from its
4251 vmem_qcache_reap(zio_arena
);
4256 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4257 * enough data and signal them to proceed. When this happens, the threads in
4258 * arc_get_data_impl() are sleeping while holding the hash lock for their
4259 * particular arc header. Thus, we must be careful to never sleep on a
4260 * hash lock in this thread. This is to prevent the following deadlock:
4262 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4263 * waiting for the reclaim thread to signal it.
4265 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4266 * fails, and goes to sleep forever.
4268 * This possible deadlock is avoided by always acquiring a hash lock
4269 * using mutex_tryenter() from arc_reclaim_thread().
4272 arc_reclaim_thread(void)
4274 fstrans_cookie_t cookie
= spl_fstrans_mark();
4275 hrtime_t growtime
= 0;
4278 CALLB_CPR_INIT(&cpr
, &arc_reclaim_lock
, callb_generic_cpr
, FTAG
);
4280 mutex_enter(&arc_reclaim_lock
);
4281 while (!arc_reclaim_thread_exit
) {
4283 uint64_t evicted
= 0;
4284 uint64_t need_free
= arc_need_free
;
4285 arc_tuning_update();
4288 * This is necessary in order for the mdb ::arc dcmd to
4289 * show up to date information. Since the ::arc command
4290 * does not call the kstat's update function, without
4291 * this call, the command may show stale stats for the
4292 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4293 * with this change, the data might be up to 1 second
4294 * out of date; but that should suffice. The arc_state_t
4295 * structures can be queried directly if more accurate
4296 * information is needed.
4299 if (arc_ksp
!= NULL
)
4300 arc_ksp
->ks_update(arc_ksp
, KSTAT_READ
);
4302 mutex_exit(&arc_reclaim_lock
);
4305 * We call arc_adjust() before (possibly) calling
4306 * arc_kmem_reap_now(), so that we can wake up
4307 * arc_get_data_buf() sooner.
4309 evicted
= arc_adjust();
4311 int64_t free_memory
= arc_available_memory();
4312 if (free_memory
< 0) {
4314 arc_no_grow
= B_TRUE
;
4318 * Wait at least zfs_grow_retry (default 5) seconds
4319 * before considering growing.
4321 growtime
= gethrtime() + SEC2NSEC(arc_grow_retry
);
4323 arc_kmem_reap_now();
4326 * If we are still low on memory, shrink the ARC
4327 * so that we have arc_shrink_min free space.
4329 free_memory
= arc_available_memory();
4331 to_free
= (arc_c
>> arc_shrink_shift
) - free_memory
;
4334 to_free
= MAX(to_free
, need_free
);
4336 arc_shrink(to_free
);
4338 } else if (free_memory
< arc_c
>> arc_no_grow_shift
) {
4339 arc_no_grow
= B_TRUE
;
4340 } else if (gethrtime() >= growtime
) {
4341 arc_no_grow
= B_FALSE
;
4344 mutex_enter(&arc_reclaim_lock
);
4347 * If evicted is zero, we couldn't evict anything via
4348 * arc_adjust(). This could be due to hash lock
4349 * collisions, but more likely due to the majority of
4350 * arc buffers being unevictable. Therefore, even if
4351 * arc_size is above arc_c, another pass is unlikely to
4352 * be helpful and could potentially cause us to enter an
4355 if (arc_size
<= arc_c
|| evicted
== 0) {
4357 * We're either no longer overflowing, or we
4358 * can't evict anything more, so we should wake
4359 * up any threads before we go to sleep and remove
4360 * the bytes we were working on from arc_need_free
4361 * since nothing more will be done here.
4363 cv_broadcast(&arc_reclaim_waiters_cv
);
4364 ARCSTAT_INCR(arcstat_need_free
, -need_free
);
4367 * Block until signaled, or after one second (we
4368 * might need to perform arc_kmem_reap_now()
4369 * even if we aren't being signalled)
4371 CALLB_CPR_SAFE_BEGIN(&cpr
);
4372 (void) cv_timedwait_sig_hires(&arc_reclaim_thread_cv
,
4373 &arc_reclaim_lock
, SEC2NSEC(1), MSEC2NSEC(1), 0);
4374 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_lock
);
4378 arc_reclaim_thread_exit
= B_FALSE
;
4379 cv_broadcast(&arc_reclaim_thread_cv
);
4380 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_lock */
4381 spl_fstrans_unmark(cookie
);
4387 * Determine the amount of memory eligible for eviction contained in the
4388 * ARC. All clean data reported by the ghost lists can always be safely
4389 * evicted. Due to arc_c_min, the same does not hold for all clean data
4390 * contained by the regular mru and mfu lists.
4392 * In the case of the regular mru and mfu lists, we need to report as
4393 * much clean data as possible, such that evicting that same reported
4394 * data will not bring arc_size below arc_c_min. Thus, in certain
4395 * circumstances, the total amount of clean data in the mru and mfu
4396 * lists might not actually be evictable.
4398 * The following two distinct cases are accounted for:
4400 * 1. The sum of the amount of dirty data contained by both the mru and
4401 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4402 * is greater than or equal to arc_c_min.
4403 * (i.e. amount of dirty data >= arc_c_min)
4405 * This is the easy case; all clean data contained by the mru and mfu
4406 * lists is evictable. Evicting all clean data can only drop arc_size
4407 * to the amount of dirty data, which is greater than arc_c_min.
4409 * 2. The sum of the amount of dirty data contained by both the mru and
4410 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4411 * is less than arc_c_min.
4412 * (i.e. arc_c_min > amount of dirty data)
4414 * 2.1. arc_size is greater than or equal arc_c_min.
4415 * (i.e. arc_size >= arc_c_min > amount of dirty data)
4417 * In this case, not all clean data from the regular mru and mfu
4418 * lists is actually evictable; we must leave enough clean data
4419 * to keep arc_size above arc_c_min. Thus, the maximum amount of
4420 * evictable data from the two lists combined, is exactly the
4421 * difference between arc_size and arc_c_min.
4423 * 2.2. arc_size is less than arc_c_min
4424 * (i.e. arc_c_min > arc_size > amount of dirty data)
4426 * In this case, none of the data contained in the mru and mfu
4427 * lists is evictable, even if it's clean. Since arc_size is
4428 * already below arc_c_min, evicting any more would only
4429 * increase this negative difference.
4432 arc_evictable_memory(void)
4434 uint64_t arc_clean
=
4435 zfs_refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]) +
4436 zfs_refcount_count(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]) +
4437 zfs_refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]) +
4438 zfs_refcount_count(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
4439 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
4442 * Scale reported evictable memory in proportion to page cache, cap
4443 * at specified min/max.
4445 uint64_t min
= (ptob(nr_file_pages()) / 100) * zfs_arc_pc_percent
;
4446 min
= MAX(arc_c_min
, MIN(arc_c_max
, min
));
4448 if (arc_dirty
>= min
)
4451 return (MAX((int64_t)arc_size
- (int64_t)min
, 0));
4455 * If sc->nr_to_scan is zero, the caller is requesting a query of the
4456 * number of objects which can potentially be freed. If it is nonzero,
4457 * the request is to free that many objects.
4459 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
4460 * in struct shrinker and also require the shrinker to return the number
4463 * Older kernels require the shrinker to return the number of freeable
4464 * objects following the freeing of nr_to_free.
4466 static spl_shrinker_t
4467 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
4471 /* The arc is considered warm once reclaim has occurred */
4472 if (unlikely(arc_warm
== B_FALSE
))
4475 /* Return the potential number of reclaimable pages */
4476 pages
= btop((int64_t)arc_evictable_memory());
4477 if (sc
->nr_to_scan
== 0)
4480 /* Not allowed to perform filesystem reclaim */
4481 if (!(sc
->gfp_mask
& __GFP_FS
))
4482 return (SHRINK_STOP
);
4484 /* Reclaim in progress */
4485 if (mutex_tryenter(&arc_reclaim_lock
) == 0) {
4486 ARCSTAT_INCR(arcstat_need_free
, ptob(sc
->nr_to_scan
));
4490 mutex_exit(&arc_reclaim_lock
);
4493 * Evict the requested number of pages by shrinking arc_c the
4497 arc_shrink(ptob(sc
->nr_to_scan
));
4498 if (current_is_kswapd())
4499 arc_kmem_reap_now();
4500 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
4501 pages
= MAX((int64_t)pages
-
4502 (int64_t)btop(arc_evictable_memory()), 0);
4504 pages
= btop(arc_evictable_memory());
4507 * We've shrunk what we can, wake up threads.
4509 cv_broadcast(&arc_reclaim_waiters_cv
);
4511 pages
= SHRINK_STOP
;
4514 * When direct reclaim is observed it usually indicates a rapid
4515 * increase in memory pressure. This occurs because the kswapd
4516 * threads were unable to asynchronously keep enough free memory
4517 * available. In this case set arc_no_grow to briefly pause arc
4518 * growth to avoid compounding the memory pressure.
4520 if (current_is_kswapd()) {
4521 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
4523 arc_no_grow
= B_TRUE
;
4524 arc_kmem_reap_now();
4525 ARCSTAT_BUMP(arcstat_memory_direct_count
);
4530 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
4532 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
4533 #endif /* _KERNEL */
4536 * Adapt arc info given the number of bytes we are trying to add and
4537 * the state that we are coming from. This function is only called
4538 * when we are adding new content to the cache.
4541 arc_adapt(int bytes
, arc_state_t
*state
)
4544 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
4545 int64_t mrug_size
= zfs_refcount_count(&arc_mru_ghost
->arcs_size
);
4546 int64_t mfug_size
= zfs_refcount_count(&arc_mfu_ghost
->arcs_size
);
4548 if (state
== arc_l2c_only
)
4553 * Adapt the target size of the MRU list:
4554 * - if we just hit in the MRU ghost list, then increase
4555 * the target size of the MRU list.
4556 * - if we just hit in the MFU ghost list, then increase
4557 * the target size of the MFU list by decreasing the
4558 * target size of the MRU list.
4560 if (state
== arc_mru_ghost
) {
4561 mult
= (mrug_size
>= mfug_size
) ? 1 : (mfug_size
/ mrug_size
);
4562 if (!zfs_arc_p_dampener_disable
)
4563 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
4565 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
4566 } else if (state
== arc_mfu_ghost
) {
4569 mult
= (mfug_size
>= mrug_size
) ? 1 : (mrug_size
/ mfug_size
);
4570 if (!zfs_arc_p_dampener_disable
)
4571 mult
= MIN(mult
, 10);
4573 delta
= MIN(bytes
* mult
, arc_p
);
4574 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
4576 ASSERT((int64_t)arc_p
>= 0);
4578 if (arc_reclaim_needed()) {
4579 cv_signal(&arc_reclaim_thread_cv
);
4586 if (arc_c
>= arc_c_max
)
4590 * If we're within (2 * maxblocksize) bytes of the target
4591 * cache size, increment the target cache size
4593 ASSERT3U(arc_c
, >=, 2ULL << SPA_MAXBLOCKSHIFT
);
4594 if (arc_size
>= arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
4595 atomic_add_64(&arc_c
, (int64_t)bytes
);
4596 if (arc_c
> arc_c_max
)
4598 else if (state
== arc_anon
)
4599 atomic_add_64(&arc_p
, (int64_t)bytes
);
4603 ASSERT((int64_t)arc_p
>= 0);
4607 * Check if arc_size has grown past our upper threshold, determined by
4608 * zfs_arc_overflow_shift.
4611 arc_is_overflowing(void)
4613 /* Always allow at least one block of overflow */
4614 uint64_t overflow
= MAX(SPA_MAXBLOCKSIZE
,
4615 arc_c
>> zfs_arc_overflow_shift
);
4617 return (arc_size
>= arc_c
+ overflow
);
4621 arc_get_data_abd(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4623 arc_buf_contents_t type
= arc_buf_type(hdr
);
4625 arc_get_data_impl(hdr
, size
, tag
);
4626 if (type
== ARC_BUFC_METADATA
) {
4627 return (abd_alloc(size
, B_TRUE
));
4629 ASSERT(type
== ARC_BUFC_DATA
);
4630 return (abd_alloc(size
, B_FALSE
));
4635 arc_get_data_buf(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4637 arc_buf_contents_t type
= arc_buf_type(hdr
);
4639 arc_get_data_impl(hdr
, size
, tag
);
4640 if (type
== ARC_BUFC_METADATA
) {
4641 return (zio_buf_alloc(size
));
4643 ASSERT(type
== ARC_BUFC_DATA
);
4644 return (zio_data_buf_alloc(size
));
4649 * Allocate a block and return it to the caller. If we are hitting the
4650 * hard limit for the cache size, we must sleep, waiting for the eviction
4651 * thread to catch up. If we're past the target size but below the hard
4652 * limit, we'll only signal the reclaim thread and continue on.
4655 arc_get_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4657 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
4658 arc_buf_contents_t type
= arc_buf_type(hdr
);
4660 arc_adapt(size
, state
);
4663 * If arc_size is currently overflowing, and has grown past our
4664 * upper limit, we must be adding data faster than the evict
4665 * thread can evict. Thus, to ensure we don't compound the
4666 * problem by adding more data and forcing arc_size to grow even
4667 * further past it's target size, we halt and wait for the
4668 * eviction thread to catch up.
4670 * It's also possible that the reclaim thread is unable to evict
4671 * enough buffers to get arc_size below the overflow limit (e.g.
4672 * due to buffers being un-evictable, or hash lock collisions).
4673 * In this case, we want to proceed regardless if we're
4674 * overflowing; thus we don't use a while loop here.
4676 if (arc_is_overflowing()) {
4677 mutex_enter(&arc_reclaim_lock
);
4680 * Now that we've acquired the lock, we may no longer be
4681 * over the overflow limit, lets check.
4683 * We're ignoring the case of spurious wake ups. If that
4684 * were to happen, it'd let this thread consume an ARC
4685 * buffer before it should have (i.e. before we're under
4686 * the overflow limit and were signalled by the reclaim
4687 * thread). As long as that is a rare occurrence, it
4688 * shouldn't cause any harm.
4690 if (arc_is_overflowing()) {
4691 cv_signal(&arc_reclaim_thread_cv
);
4692 cv_wait(&arc_reclaim_waiters_cv
, &arc_reclaim_lock
);
4695 mutex_exit(&arc_reclaim_lock
);
4698 VERIFY3U(hdr
->b_type
, ==, type
);
4699 if (type
== ARC_BUFC_METADATA
) {
4700 arc_space_consume(size
, ARC_SPACE_META
);
4702 arc_space_consume(size
, ARC_SPACE_DATA
);
4706 * Update the state size. Note that ghost states have a
4707 * "ghost size" and so don't need to be updated.
4709 if (!GHOST_STATE(state
)) {
4711 (void) zfs_refcount_add_many(&state
->arcs_size
, size
, tag
);
4714 * If this is reached via arc_read, the link is
4715 * protected by the hash lock. If reached via
4716 * arc_buf_alloc, the header should not be accessed by
4717 * any other thread. And, if reached via arc_read_done,
4718 * the hash lock will protect it if it's found in the
4719 * hash table; otherwise no other thread should be
4720 * trying to [add|remove]_reference it.
4722 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
4723 ASSERT(zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4724 (void) zfs_refcount_add_many(&state
->arcs_esize
[type
],
4729 * If we are growing the cache, and we are adding anonymous
4730 * data, and we have outgrown arc_p, update arc_p
4732 if (arc_size
< arc_c
&& hdr
->b_l1hdr
.b_state
== arc_anon
&&
4733 (zfs_refcount_count(&arc_anon
->arcs_size
) +
4734 zfs_refcount_count(&arc_mru
->arcs_size
) > arc_p
))
4735 arc_p
= MIN(arc_c
, arc_p
+ size
);
4740 arc_free_data_abd(arc_buf_hdr_t
*hdr
, abd_t
*abd
, uint64_t size
, void *tag
)
4742 arc_free_data_impl(hdr
, size
, tag
);
4747 arc_free_data_buf(arc_buf_hdr_t
*hdr
, void *buf
, uint64_t size
, void *tag
)
4749 arc_buf_contents_t type
= arc_buf_type(hdr
);
4751 arc_free_data_impl(hdr
, size
, tag
);
4752 if (type
== ARC_BUFC_METADATA
) {
4753 zio_buf_free(buf
, size
);
4755 ASSERT(type
== ARC_BUFC_DATA
);
4756 zio_data_buf_free(buf
, size
);
4761 * Free the arc data buffer.
4764 arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4766 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
4767 arc_buf_contents_t type
= arc_buf_type(hdr
);
4769 /* protected by hash lock, if in the hash table */
4770 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
4771 ASSERT(zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4772 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
4774 (void) zfs_refcount_remove_many(&state
->arcs_esize
[type
],
4777 (void) zfs_refcount_remove_many(&state
->arcs_size
, size
, tag
);
4779 VERIFY3U(hdr
->b_type
, ==, type
);
4780 if (type
== ARC_BUFC_METADATA
) {
4781 arc_space_return(size
, ARC_SPACE_META
);
4783 ASSERT(type
== ARC_BUFC_DATA
);
4784 arc_space_return(size
, ARC_SPACE_DATA
);
4789 * This routine is called whenever a buffer is accessed.
4790 * NOTE: the hash lock is dropped in this function.
4793 arc_access(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
4797 ASSERT(MUTEX_HELD(hash_lock
));
4798 ASSERT(HDR_HAS_L1HDR(hdr
));
4800 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
4802 * This buffer is not in the cache, and does not
4803 * appear in our "ghost" list. Add the new buffer
4807 ASSERT0(hdr
->b_l1hdr
.b_arc_access
);
4808 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4809 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
4810 arc_change_state(arc_mru
, hdr
, hash_lock
);
4812 } else if (hdr
->b_l1hdr
.b_state
== arc_mru
) {
4813 now
= ddi_get_lbolt();
4816 * If this buffer is here because of a prefetch, then either:
4817 * - clear the flag if this is a "referencing" read
4818 * (any subsequent access will bump this into the MFU state).
4820 * - move the buffer to the head of the list if this is
4821 * another prefetch (to make it less likely to be evicted).
4823 if (HDR_PREFETCH(hdr
)) {
4824 if (zfs_refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
4825 /* link protected by hash lock */
4826 ASSERT(multilist_link_active(
4827 &hdr
->b_l1hdr
.b_arc_node
));
4829 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
4830 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
4831 ARCSTAT_BUMP(arcstat_mru_hits
);
4833 hdr
->b_l1hdr
.b_arc_access
= now
;
4838 * This buffer has been "accessed" only once so far,
4839 * but it is still in the cache. Move it to the MFU
4842 if (ddi_time_after(now
, hdr
->b_l1hdr
.b_arc_access
+
4845 * More than 125ms have passed since we
4846 * instantiated this buffer. Move it to the
4847 * most frequently used state.
4849 hdr
->b_l1hdr
.b_arc_access
= now
;
4850 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4851 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4853 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
4854 ARCSTAT_BUMP(arcstat_mru_hits
);
4855 } else if (hdr
->b_l1hdr
.b_state
== arc_mru_ghost
) {
4856 arc_state_t
*new_state
;
4858 * This buffer has been "accessed" recently, but
4859 * was evicted from the cache. Move it to the
4863 if (HDR_PREFETCH(hdr
)) {
4864 new_state
= arc_mru
;
4865 if (zfs_refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0)
4866 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
4867 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
4869 new_state
= arc_mfu
;
4870 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4873 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4874 arc_change_state(new_state
, hdr
, hash_lock
);
4876 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_ghost_hits
);
4877 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
4878 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu
) {
4880 * This buffer has been accessed more than once and is
4881 * still in the cache. Keep it in the MFU state.
4883 * NOTE: an add_reference() that occurred when we did
4884 * the arc_read() will have kicked this off the list.
4885 * If it was a prefetch, we will explicitly move it to
4886 * the head of the list now.
4888 if ((HDR_PREFETCH(hdr
)) != 0) {
4889 ASSERT(zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4890 /* link protected by hash_lock */
4891 ASSERT(multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4893 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_hits
);
4894 ARCSTAT_BUMP(arcstat_mfu_hits
);
4895 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4896 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu_ghost
) {
4897 arc_state_t
*new_state
= arc_mfu
;
4899 * This buffer has been accessed more than once but has
4900 * been evicted from the cache. Move it back to the
4904 if (HDR_PREFETCH(hdr
)) {
4906 * This is a prefetch access...
4907 * move this block back to the MRU state.
4909 ASSERT0(zfs_refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
4910 new_state
= arc_mru
;
4913 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4914 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4915 arc_change_state(new_state
, hdr
, hash_lock
);
4917 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_ghost_hits
);
4918 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
4919 } else if (hdr
->b_l1hdr
.b_state
== arc_l2c_only
) {
4921 * This buffer is on the 2nd Level ARC.
4924 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4925 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4926 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4928 cmn_err(CE_PANIC
, "invalid arc state 0x%p",
4929 hdr
->b_l1hdr
.b_state
);
4934 * This routine is called by dbuf_hold() to update the arc_access() state
4935 * which otherwise would be skipped for entries in the dbuf cache.
4938 arc_buf_access(arc_buf_t
*buf
)
4940 mutex_enter(&buf
->b_evict_lock
);
4941 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4944 * Avoid taking the hash_lock when possible as an optimization.
4945 * The header must be checked again under the hash_lock in order
4946 * to handle the case where it is concurrently being released.
4948 if (hdr
->b_l1hdr
.b_state
== arc_anon
|| HDR_EMPTY(hdr
)) {
4949 mutex_exit(&buf
->b_evict_lock
);
4953 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
4954 mutex_enter(hash_lock
);
4956 if (hdr
->b_l1hdr
.b_state
== arc_anon
|| HDR_EMPTY(hdr
)) {
4957 mutex_exit(hash_lock
);
4958 mutex_exit(&buf
->b_evict_lock
);
4959 ARCSTAT_BUMP(arcstat_access_skip
);
4963 mutex_exit(&buf
->b_evict_lock
);
4965 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
4966 hdr
->b_l1hdr
.b_state
== arc_mfu
);
4968 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
4969 arc_access(hdr
, hash_lock
);
4970 mutex_exit(hash_lock
);
4972 ARCSTAT_BUMP(arcstat_hits
);
4973 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
), demand
, prefetch
,
4974 !HDR_ISTYPE_METADATA(hdr
), data
, metadata
, hits
);
4977 /* a generic arc_read_done_func_t which you can use */
4980 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4982 if (zio
== NULL
|| zio
->io_error
== 0)
4983 bcopy(buf
->b_data
, arg
, arc_buf_size(buf
));
4984 arc_buf_destroy(buf
, arg
);
4987 /* a generic arc_done_func_t */
4989 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4991 arc_buf_t
**bufp
= arg
;
4992 if (zio
&& zio
->io_error
) {
4993 arc_buf_destroy(buf
, arg
);
4997 ASSERT(buf
->b_data
);
5002 arc_hdr_verify(arc_buf_hdr_t
*hdr
, blkptr_t
*bp
)
5004 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
)) {
5005 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, 0);
5006 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
5008 if (HDR_COMPRESSION_ENABLED(hdr
)) {
5009 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==,
5010 BP_GET_COMPRESS(bp
));
5012 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(bp
));
5013 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, BP_GET_PSIZE(bp
));
5018 arc_read_done(zio_t
*zio
)
5020 arc_buf_hdr_t
*hdr
= zio
->io_private
;
5021 kmutex_t
*hash_lock
= NULL
;
5022 arc_callback_t
*callback_list
;
5023 arc_callback_t
*acb
;
5024 boolean_t freeable
= B_FALSE
;
5025 boolean_t no_zio_error
= (zio
->io_error
== 0);
5028 * The hdr was inserted into hash-table and removed from lists
5029 * prior to starting I/O. We should find this header, since
5030 * it's in the hash table, and it should be legit since it's
5031 * not possible to evict it during the I/O. The only possible
5032 * reason for it not to be found is if we were freed during the
5035 if (HDR_IN_HASH_TABLE(hdr
)) {
5036 arc_buf_hdr_t
*found
;
5038 ASSERT3U(hdr
->b_birth
, ==, BP_PHYSICAL_BIRTH(zio
->io_bp
));
5039 ASSERT3U(hdr
->b_dva
.dva_word
[0], ==,
5040 BP_IDENTITY(zio
->io_bp
)->dva_word
[0]);
5041 ASSERT3U(hdr
->b_dva
.dva_word
[1], ==,
5042 BP_IDENTITY(zio
->io_bp
)->dva_word
[1]);
5044 found
= buf_hash_find(hdr
->b_spa
, zio
->io_bp
, &hash_lock
);
5046 ASSERT((found
== hdr
&&
5047 DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
5048 (found
== hdr
&& HDR_L2_READING(hdr
)));
5049 ASSERT3P(hash_lock
, !=, NULL
);
5053 /* byteswap if necessary */
5054 if (BP_SHOULD_BYTESWAP(zio
->io_bp
)) {
5055 if (BP_GET_LEVEL(zio
->io_bp
) > 0) {
5056 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_UINT64
;
5058 hdr
->b_l1hdr
.b_byteswap
=
5059 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
5062 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
5066 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_EVICTED
);
5067 if (l2arc_noprefetch
&& HDR_PREFETCH(hdr
))
5068 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2CACHE
);
5070 callback_list
= hdr
->b_l1hdr
.b_acb
;
5071 ASSERT3P(callback_list
, !=, NULL
);
5073 if (hash_lock
&& no_zio_error
&& hdr
->b_l1hdr
.b_state
== arc_anon
) {
5075 * Only call arc_access on anonymous buffers. This is because
5076 * if we've issued an I/O for an evicted buffer, we've already
5077 * called arc_access (to prevent any simultaneous readers from
5078 * getting confused).
5080 arc_access(hdr
, hash_lock
);
5084 * If a read request has a callback (i.e. acb_done is not NULL), then we
5085 * make a buf containing the data according to the parameters which were
5086 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5087 * aren't needlessly decompressing the data multiple times.
5089 int callback_cnt
= 0;
5090 for (acb
= callback_list
; acb
!= NULL
; acb
= acb
->acb_next
) {
5094 /* This is a demand read since prefetches don't use callbacks */
5097 int error
= arc_buf_alloc_impl(hdr
, acb
->acb_private
,
5098 acb
->acb_compressed
, no_zio_error
, &acb
->acb_buf
);
5100 zio
->io_error
= error
;
5103 hdr
->b_l1hdr
.b_acb
= NULL
;
5104 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5105 if (callback_cnt
== 0) {
5106 ASSERT(HDR_PREFETCH(hdr
));
5107 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
5108 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5111 ASSERT(zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
) ||
5112 callback_list
!= NULL
);
5115 arc_hdr_verify(hdr
, zio
->io_bp
);
5117 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_ERROR
);
5118 if (hdr
->b_l1hdr
.b_state
!= arc_anon
)
5119 arc_change_state(arc_anon
, hdr
, hash_lock
);
5120 if (HDR_IN_HASH_TABLE(hdr
))
5121 buf_hash_remove(hdr
);
5122 freeable
= zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
5126 * Broadcast before we drop the hash_lock to avoid the possibility
5127 * that the hdr (and hence the cv) might be freed before we get to
5128 * the cv_broadcast().
5130 cv_broadcast(&hdr
->b_l1hdr
.b_cv
);
5132 if (hash_lock
!= NULL
) {
5133 mutex_exit(hash_lock
);
5136 * This block was freed while we waited for the read to
5137 * complete. It has been removed from the hash table and
5138 * moved to the anonymous state (so that it won't show up
5141 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
5142 freeable
= zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
5145 /* execute each callback and free its structure */
5146 while ((acb
= callback_list
) != NULL
) {
5148 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
5150 if (acb
->acb_zio_dummy
!= NULL
) {
5151 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
5152 zio_nowait(acb
->acb_zio_dummy
);
5155 callback_list
= acb
->acb_next
;
5156 kmem_free(acb
, sizeof (arc_callback_t
));
5160 arc_hdr_destroy(hdr
);
5164 * "Read" the block at the specified DVA (in bp) via the
5165 * cache. If the block is found in the cache, invoke the provided
5166 * callback immediately and return. Note that the `zio' parameter
5167 * in the callback will be NULL in this case, since no IO was
5168 * required. If the block is not in the cache pass the read request
5169 * on to the spa with a substitute callback function, so that the
5170 * requested block will be added to the cache.
5172 * If a read request arrives for a block that has a read in-progress,
5173 * either wait for the in-progress read to complete (and return the
5174 * results); or, if this is a read with a "done" func, add a record
5175 * to the read to invoke the "done" func when the read completes,
5176 * and return; or just return.
5178 * arc_read_done() will invoke all the requested "done" functions
5179 * for readers of this block.
5182 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_done_func_t
*done
,
5183 void *private, zio_priority_t priority
, int zio_flags
,
5184 arc_flags_t
*arc_flags
, const zbookmark_phys_t
*zb
)
5186 arc_buf_hdr_t
*hdr
= NULL
;
5187 kmutex_t
*hash_lock
= NULL
;
5189 uint64_t guid
= spa_load_guid(spa
);
5190 boolean_t compressed_read
= (zio_flags
& ZIO_FLAG_RAW
) != 0;
5193 ASSERT(!BP_IS_EMBEDDED(bp
) ||
5194 BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
);
5197 if (!BP_IS_EMBEDDED(bp
)) {
5199 * Embedded BP's have no DVA and require no I/O to "read".
5200 * Create an anonymous arc buf to back it.
5202 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
5205 if (hdr
!= NULL
&& HDR_HAS_L1HDR(hdr
) && hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5206 arc_buf_t
*buf
= NULL
;
5207 *arc_flags
|= ARC_FLAG_CACHED
;
5209 if (HDR_IO_IN_PROGRESS(hdr
)) {
5211 if ((hdr
->b_flags
& ARC_FLAG_PRIO_ASYNC_READ
) &&
5212 priority
== ZIO_PRIORITY_SYNC_READ
) {
5214 * This sync read must wait for an
5215 * in-progress async read (e.g. a predictive
5216 * prefetch). Async reads are queued
5217 * separately at the vdev_queue layer, so
5218 * this is a form of priority inversion.
5219 * Ideally, we would "inherit" the demand
5220 * i/o's priority by moving the i/o from
5221 * the async queue to the synchronous queue,
5222 * but there is currently no mechanism to do
5223 * so. Track this so that we can evaluate
5224 * the magnitude of this potential performance
5227 * Note that if the prefetch i/o is already
5228 * active (has been issued to the device),
5229 * the prefetch improved performance, because
5230 * we issued it sooner than we would have
5231 * without the prefetch.
5233 DTRACE_PROBE1(arc__sync__wait__for__async
,
5234 arc_buf_hdr_t
*, hdr
);
5235 ARCSTAT_BUMP(arcstat_sync_wait_for_async
);
5237 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5238 arc_hdr_clear_flags(hdr
,
5239 ARC_FLAG_PREDICTIVE_PREFETCH
);
5242 if (*arc_flags
& ARC_FLAG_WAIT
) {
5243 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
5244 mutex_exit(hash_lock
);
5247 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5250 arc_callback_t
*acb
= NULL
;
5252 acb
= kmem_zalloc(sizeof (arc_callback_t
),
5254 acb
->acb_done
= done
;
5255 acb
->acb_private
= private;
5256 acb
->acb_compressed
= compressed_read
;
5258 acb
->acb_zio_dummy
= zio_null(pio
,
5259 spa
, NULL
, NULL
, NULL
, zio_flags
);
5261 ASSERT3P(acb
->acb_done
, !=, NULL
);
5262 acb
->acb_next
= hdr
->b_l1hdr
.b_acb
;
5263 hdr
->b_l1hdr
.b_acb
= acb
;
5264 mutex_exit(hash_lock
);
5267 mutex_exit(hash_lock
);
5271 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
5272 hdr
->b_l1hdr
.b_state
== arc_mfu
);
5275 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5277 * This is a demand read which does not have to
5278 * wait for i/o because we did a predictive
5279 * prefetch i/o for it, which has completed.
5282 arc__demand__hit__predictive__prefetch
,
5283 arc_buf_hdr_t
*, hdr
);
5285 arcstat_demand_hit_predictive_prefetch
);
5286 arc_hdr_clear_flags(hdr
,
5287 ARC_FLAG_PREDICTIVE_PREFETCH
);
5289 ASSERT(!BP_IS_EMBEDDED(bp
) || !BP_IS_HOLE(bp
));
5291 /* Get a buf with the desired data in it. */
5292 VERIFY0(arc_buf_alloc_impl(hdr
, private,
5293 compressed_read
, B_TRUE
, &buf
));
5294 } else if (*arc_flags
& ARC_FLAG_PREFETCH
&&
5295 zfs_refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
5296 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
5298 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
5299 arc_access(hdr
, hash_lock
);
5300 if (*arc_flags
& ARC_FLAG_L2CACHE
)
5301 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5302 mutex_exit(hash_lock
);
5303 ARCSTAT_BUMP(arcstat_hits
);
5304 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
5305 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
5306 data
, metadata
, hits
);
5309 done(NULL
, buf
, private);
5311 uint64_t lsize
= BP_GET_LSIZE(bp
);
5312 uint64_t psize
= BP_GET_PSIZE(bp
);
5313 arc_callback_t
*acb
;
5316 boolean_t devw
= B_FALSE
;
5320 * Gracefully handle a damaged logical block size as a
5323 if (lsize
> spa_maxblocksize(spa
)) {
5324 rc
= SET_ERROR(ECKSUM
);
5329 /* this block is not in the cache */
5330 arc_buf_hdr_t
*exists
= NULL
;
5331 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
5332 hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
5333 BP_GET_COMPRESS(bp
), type
);
5335 if (!BP_IS_EMBEDDED(bp
)) {
5336 hdr
->b_dva
= *BP_IDENTITY(bp
);
5337 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
5338 exists
= buf_hash_insert(hdr
, &hash_lock
);
5340 if (exists
!= NULL
) {
5341 /* somebody beat us to the hash insert */
5342 mutex_exit(hash_lock
);
5343 buf_discard_identity(hdr
);
5344 arc_hdr_destroy(hdr
);
5345 goto top
; /* restart the IO request */
5349 * This block is in the ghost cache. If it was L2-only
5350 * (and thus didn't have an L1 hdr), we realloc the
5351 * header to add an L1 hdr.
5353 if (!HDR_HAS_L1HDR(hdr
)) {
5354 hdr
= arc_hdr_realloc(hdr
, hdr_l2only_cache
,
5358 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5359 ASSERT(GHOST_STATE(hdr
->b_l1hdr
.b_state
));
5360 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5361 ASSERT(zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5362 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
5363 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
5366 * This is a delicate dance that we play here.
5367 * This hdr is in the ghost list so we access it
5368 * to move it out of the ghost list before we
5369 * initiate the read. If it's a prefetch then
5370 * it won't have a callback so we'll remove the
5371 * reference that arc_buf_alloc_impl() created. We
5372 * do this after we've called arc_access() to
5373 * avoid hitting an assert in remove_reference().
5375 arc_access(hdr
, hash_lock
);
5376 arc_hdr_alloc_pabd(hdr
);
5378 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5379 size
= arc_hdr_size(hdr
);
5382 * If compression is enabled on the hdr, then will do
5383 * RAW I/O and will store the compressed data in the hdr's
5384 * data block. Otherwise, the hdr's data block will contain
5385 * the uncompressed data.
5387 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
) {
5388 zio_flags
|= ZIO_FLAG_RAW
;
5391 if (*arc_flags
& ARC_FLAG_PREFETCH
)
5392 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
5393 if (*arc_flags
& ARC_FLAG_L2CACHE
)
5394 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5395 if (BP_GET_LEVEL(bp
) > 0)
5396 arc_hdr_set_flags(hdr
, ARC_FLAG_INDIRECT
);
5397 if (*arc_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
)
5398 arc_hdr_set_flags(hdr
, ARC_FLAG_PREDICTIVE_PREFETCH
);
5399 ASSERT(!GHOST_STATE(hdr
->b_l1hdr
.b_state
));
5401 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
5402 acb
->acb_done
= done
;
5403 acb
->acb_private
= private;
5404 acb
->acb_compressed
= compressed_read
;
5406 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5407 hdr
->b_l1hdr
.b_acb
= acb
;
5408 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5410 if (HDR_HAS_L2HDR(hdr
) &&
5411 (vd
= hdr
->b_l2hdr
.b_dev
->l2ad_vdev
) != NULL
) {
5412 devw
= hdr
->b_l2hdr
.b_dev
->l2ad_writing
;
5413 addr
= hdr
->b_l2hdr
.b_daddr
;
5415 * Lock out device removal.
5417 if (vdev_is_dead(vd
) ||
5418 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
5422 if (priority
== ZIO_PRIORITY_ASYNC_READ
)
5423 arc_hdr_set_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
5425 arc_hdr_clear_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
5427 if (hash_lock
!= NULL
)
5428 mutex_exit(hash_lock
);
5431 * At this point, we have a level 1 cache miss. Try again in
5432 * L2ARC if possible.
5434 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, lsize
);
5436 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
5437 uint64_t, lsize
, zbookmark_phys_t
*, zb
);
5438 ARCSTAT_BUMP(arcstat_misses
);
5439 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
5440 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
5441 data
, metadata
, misses
);
5443 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
5445 * Read from the L2ARC if the following are true:
5446 * 1. The L2ARC vdev was previously cached.
5447 * 2. This buffer still has L2ARC metadata.
5448 * 3. This buffer isn't currently writing to the L2ARC.
5449 * 4. The L2ARC entry wasn't evicted, which may
5450 * also have invalidated the vdev.
5451 * 5. This isn't prefetch and l2arc_noprefetch is set.
5453 if (HDR_HAS_L2HDR(hdr
) &&
5454 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
5455 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
5456 l2arc_read_callback_t
*cb
;
5460 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
5461 ARCSTAT_BUMP(arcstat_l2_hits
);
5462 atomic_inc_32(&hdr
->b_l2hdr
.b_hits
);
5464 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
5466 cb
->l2rcb_hdr
= hdr
;
5469 cb
->l2rcb_flags
= zio_flags
;
5471 asize
= vdev_psize_to_asize(vd
, size
);
5472 if (asize
!= size
) {
5473 abd
= abd_alloc_for_io(asize
,
5474 HDR_ISTYPE_METADATA(hdr
));
5475 cb
->l2rcb_abd
= abd
;
5477 abd
= hdr
->b_l1hdr
.b_pabd
;
5480 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
5481 addr
+ asize
<= vd
->vdev_psize
-
5482 VDEV_LABEL_END_SIZE
);
5485 * l2arc read. The SCL_L2ARC lock will be
5486 * released by l2arc_read_done().
5487 * Issue a null zio if the underlying buffer
5488 * was squashed to zero size by compression.
5490 ASSERT3U(HDR_GET_COMPRESS(hdr
), !=,
5491 ZIO_COMPRESS_EMPTY
);
5492 rzio
= zio_read_phys(pio
, vd
, addr
,
5495 l2arc_read_done
, cb
, priority
,
5496 zio_flags
| ZIO_FLAG_DONT_CACHE
|
5498 ZIO_FLAG_DONT_PROPAGATE
|
5499 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
5501 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
5503 ARCSTAT_INCR(arcstat_l2_read_bytes
, size
);
5505 if (*arc_flags
& ARC_FLAG_NOWAIT
) {
5510 ASSERT(*arc_flags
& ARC_FLAG_WAIT
);
5511 if (zio_wait(rzio
) == 0)
5514 /* l2arc read error; goto zio_read() */
5516 DTRACE_PROBE1(l2arc__miss
,
5517 arc_buf_hdr_t
*, hdr
);
5518 ARCSTAT_BUMP(arcstat_l2_misses
);
5519 if (HDR_L2_WRITING(hdr
))
5520 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
5521 spa_config_exit(spa
, SCL_L2ARC
, vd
);
5525 spa_config_exit(spa
, SCL_L2ARC
, vd
);
5526 if (l2arc_ndev
!= 0) {
5527 DTRACE_PROBE1(l2arc__miss
,
5528 arc_buf_hdr_t
*, hdr
);
5529 ARCSTAT_BUMP(arcstat_l2_misses
);
5533 rzio
= zio_read(pio
, spa
, bp
, hdr
->b_l1hdr
.b_pabd
, size
,
5534 arc_read_done
, hdr
, priority
, zio_flags
, zb
);
5536 if (*arc_flags
& ARC_FLAG_WAIT
) {
5537 rc
= zio_wait(rzio
);
5541 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5546 spa_read_history_add(spa
, zb
, *arc_flags
);
5551 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
5555 p
= kmem_alloc(sizeof (*p
), KM_SLEEP
);
5557 p
->p_private
= private;
5558 list_link_init(&p
->p_node
);
5559 zfs_refcount_create(&p
->p_refcnt
);
5561 mutex_enter(&arc_prune_mtx
);
5562 zfs_refcount_add(&p
->p_refcnt
, &arc_prune_list
);
5563 list_insert_head(&arc_prune_list
, p
);
5564 mutex_exit(&arc_prune_mtx
);
5570 arc_remove_prune_callback(arc_prune_t
*p
)
5572 boolean_t wait
= B_FALSE
;
5573 mutex_enter(&arc_prune_mtx
);
5574 list_remove(&arc_prune_list
, p
);
5575 if (zfs_refcount_remove(&p
->p_refcnt
, &arc_prune_list
) > 0)
5577 mutex_exit(&arc_prune_mtx
);
5579 /* wait for arc_prune_task to finish */
5581 taskq_wait_outstanding(arc_prune_taskq
, 0);
5582 ASSERT0(zfs_refcount_count(&p
->p_refcnt
));
5583 zfs_refcount_destroy(&p
->p_refcnt
);
5584 kmem_free(p
, sizeof (*p
));
5588 * Notify the arc that a block was freed, and thus will never be used again.
5591 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
5594 kmutex_t
*hash_lock
;
5595 uint64_t guid
= spa_load_guid(spa
);
5597 ASSERT(!BP_IS_EMBEDDED(bp
));
5599 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
5604 * We might be trying to free a block that is still doing I/O
5605 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5606 * dmu_sync-ed block). If this block is being prefetched, then it
5607 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5608 * until the I/O completes. A block may also have a reference if it is
5609 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5610 * have written the new block to its final resting place on disk but
5611 * without the dedup flag set. This would have left the hdr in the MRU
5612 * state and discoverable. When the txg finally syncs it detects that
5613 * the block was overridden in open context and issues an override I/O.
5614 * Since this is a dedup block, the override I/O will determine if the
5615 * block is already in the DDT. If so, then it will replace the io_bp
5616 * with the bp from the DDT and allow the I/O to finish. When the I/O
5617 * reaches the done callback, dbuf_write_override_done, it will
5618 * check to see if the io_bp and io_bp_override are identical.
5619 * If they are not, then it indicates that the bp was replaced with
5620 * the bp in the DDT and the override bp is freed. This allows
5621 * us to arrive here with a reference on a block that is being
5622 * freed. So if we have an I/O in progress, or a reference to
5623 * this hdr, then we don't destroy the hdr.
5625 if (!HDR_HAS_L1HDR(hdr
) || (!HDR_IO_IN_PROGRESS(hdr
) &&
5626 zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
))) {
5627 arc_change_state(arc_anon
, hdr
, hash_lock
);
5628 arc_hdr_destroy(hdr
);
5629 mutex_exit(hash_lock
);
5631 mutex_exit(hash_lock
);
5637 * Release this buffer from the cache, making it an anonymous buffer. This
5638 * must be done after a read and prior to modifying the buffer contents.
5639 * If the buffer has more than one reference, we must make
5640 * a new hdr for the buffer.
5643 arc_release(arc_buf_t
*buf
, void *tag
)
5645 kmutex_t
*hash_lock
;
5647 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5650 * It would be nice to assert that if its DMU metadata (level >
5651 * 0 || it's the dnode file), then it must be syncing context.
5652 * But we don't know that information at this level.
5655 mutex_enter(&buf
->b_evict_lock
);
5657 ASSERT(HDR_HAS_L1HDR(hdr
));
5660 * We don't grab the hash lock prior to this check, because if
5661 * the buffer's header is in the arc_anon state, it won't be
5662 * linked into the hash table.
5664 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
5665 mutex_exit(&buf
->b_evict_lock
);
5666 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5667 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
5668 ASSERT(!HDR_HAS_L2HDR(hdr
));
5669 ASSERT(HDR_EMPTY(hdr
));
5671 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
5672 ASSERT3S(zfs_refcount_count(&hdr
->b_l1hdr
.b_refcnt
), ==, 1);
5673 ASSERT(!list_link_active(&hdr
->b_l1hdr
.b_arc_node
));
5675 hdr
->b_l1hdr
.b_arc_access
= 0;
5678 * If the buf is being overridden then it may already
5679 * have a hdr that is not empty.
5681 buf_discard_identity(hdr
);
5687 hash_lock
= HDR_LOCK(hdr
);
5688 mutex_enter(hash_lock
);
5691 * This assignment is only valid as long as the hash_lock is
5692 * held, we must be careful not to reference state or the
5693 * b_state field after dropping the lock.
5695 state
= hdr
->b_l1hdr
.b_state
;
5696 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
5697 ASSERT3P(state
, !=, arc_anon
);
5699 /* this buffer is not on any list */
5700 ASSERT3S(zfs_refcount_count(&hdr
->b_l1hdr
.b_refcnt
), >, 0);
5702 if (HDR_HAS_L2HDR(hdr
)) {
5703 mutex_enter(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
5706 * We have to recheck this conditional again now that
5707 * we're holding the l2ad_mtx to prevent a race with
5708 * another thread which might be concurrently calling
5709 * l2arc_evict(). In that case, l2arc_evict() might have
5710 * destroyed the header's L2 portion as we were waiting
5711 * to acquire the l2ad_mtx.
5713 if (HDR_HAS_L2HDR(hdr
))
5714 arc_hdr_l2hdr_destroy(hdr
);
5716 mutex_exit(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
5720 * Do we have more than one buf?
5722 if (hdr
->b_l1hdr
.b_bufcnt
> 1) {
5723 arc_buf_hdr_t
*nhdr
;
5724 uint64_t spa
= hdr
->b_spa
;
5725 uint64_t psize
= HDR_GET_PSIZE(hdr
);
5726 uint64_t lsize
= HDR_GET_LSIZE(hdr
);
5727 enum zio_compress compress
= HDR_GET_COMPRESS(hdr
);
5728 arc_buf_contents_t type
= arc_buf_type(hdr
);
5729 VERIFY3U(hdr
->b_type
, ==, type
);
5731 ASSERT(hdr
->b_l1hdr
.b_buf
!= buf
|| buf
->b_next
!= NULL
);
5732 (void) remove_reference(hdr
, hash_lock
, tag
);
5734 if (arc_buf_is_shared(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
5735 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
5736 ASSERT(ARC_BUF_LAST(buf
));
5740 * Pull the data off of this hdr and attach it to
5741 * a new anonymous hdr. Also find the last buffer
5742 * in the hdr's buffer list.
5744 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
5745 ASSERT3P(lastbuf
, !=, NULL
);
5748 * If the current arc_buf_t and the hdr are sharing their data
5749 * buffer, then we must stop sharing that block.
5751 if (arc_buf_is_shared(buf
)) {
5752 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
5753 VERIFY(!arc_buf_is_shared(lastbuf
));
5756 * First, sever the block sharing relationship between
5757 * buf and the arc_buf_hdr_t.
5759 arc_unshare_buf(hdr
, buf
);
5762 * Now we need to recreate the hdr's b_pabd. Since we
5763 * have lastbuf handy, we try to share with it, but if
5764 * we can't then we allocate a new b_pabd and copy the
5765 * data from buf into it.
5767 if (arc_can_share(hdr
, lastbuf
)) {
5768 arc_share_buf(hdr
, lastbuf
);
5770 arc_hdr_alloc_pabd(hdr
);
5771 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
,
5772 buf
->b_data
, psize
);
5774 VERIFY3P(lastbuf
->b_data
, !=, NULL
);
5775 } else if (HDR_SHARED_DATA(hdr
)) {
5777 * Uncompressed shared buffers are always at the end
5778 * of the list. Compressed buffers don't have the
5779 * same requirements. This makes it hard to
5780 * simply assert that the lastbuf is shared so
5781 * we rely on the hdr's compression flags to determine
5782 * if we have a compressed, shared buffer.
5784 ASSERT(arc_buf_is_shared(lastbuf
) ||
5785 HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
5786 ASSERT(!ARC_BUF_SHARED(buf
));
5788 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5789 ASSERT3P(state
, !=, arc_l2c_only
);
5791 (void) zfs_refcount_remove_many(&state
->arcs_size
,
5792 arc_buf_size(buf
), buf
);
5794 if (zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
)) {
5795 ASSERT3P(state
, !=, arc_l2c_only
);
5796 (void) zfs_refcount_remove_many(
5797 &state
->arcs_esize
[type
],
5798 arc_buf_size(buf
), buf
);
5801 hdr
->b_l1hdr
.b_bufcnt
-= 1;
5802 arc_cksum_verify(buf
);
5803 arc_buf_unwatch(buf
);
5805 /* if this is the last uncompressed buf free the checksum */
5806 if (!arc_hdr_has_uncompressed_buf(hdr
))
5807 arc_cksum_free(hdr
);
5809 mutex_exit(hash_lock
);
5812 * Allocate a new hdr. The new hdr will contain a b_pabd
5813 * buffer which will be freed in arc_write().
5815 nhdr
= arc_hdr_alloc(spa
, psize
, lsize
, compress
, type
);
5816 ASSERT3P(nhdr
->b_l1hdr
.b_buf
, ==, NULL
);
5817 ASSERT0(nhdr
->b_l1hdr
.b_bufcnt
);
5818 ASSERT0(zfs_refcount_count(&nhdr
->b_l1hdr
.b_refcnt
));
5819 VERIFY3U(nhdr
->b_type
, ==, type
);
5820 ASSERT(!HDR_SHARED_DATA(nhdr
));
5822 nhdr
->b_l1hdr
.b_buf
= buf
;
5823 nhdr
->b_l1hdr
.b_bufcnt
= 1;
5824 nhdr
->b_l1hdr
.b_mru_hits
= 0;
5825 nhdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
5826 nhdr
->b_l1hdr
.b_mfu_hits
= 0;
5827 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
5828 nhdr
->b_l1hdr
.b_l2_hits
= 0;
5829 (void) zfs_refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, tag
);
5832 mutex_exit(&buf
->b_evict_lock
);
5833 (void) zfs_refcount_add_many(&arc_anon
->arcs_size
,
5834 arc_buf_size(buf
), buf
);
5836 mutex_exit(&buf
->b_evict_lock
);
5837 ASSERT(zfs_refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 1);
5838 /* protected by hash lock, or hdr is on arc_anon */
5839 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
5840 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5841 hdr
->b_l1hdr
.b_mru_hits
= 0;
5842 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
5843 hdr
->b_l1hdr
.b_mfu_hits
= 0;
5844 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
5845 hdr
->b_l1hdr
.b_l2_hits
= 0;
5846 arc_change_state(arc_anon
, hdr
, hash_lock
);
5847 hdr
->b_l1hdr
.b_arc_access
= 0;
5848 mutex_exit(hash_lock
);
5850 buf_discard_identity(hdr
);
5856 arc_released(arc_buf_t
*buf
)
5860 mutex_enter(&buf
->b_evict_lock
);
5861 released
= (buf
->b_data
!= NULL
&&
5862 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
5863 mutex_exit(&buf
->b_evict_lock
);
5869 arc_referenced(arc_buf_t
*buf
)
5873 mutex_enter(&buf
->b_evict_lock
);
5874 referenced
= (zfs_refcount_count(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
5875 mutex_exit(&buf
->b_evict_lock
);
5876 return (referenced
);
5881 arc_write_ready(zio_t
*zio
)
5883 arc_write_callback_t
*callback
= zio
->io_private
;
5884 arc_buf_t
*buf
= callback
->awcb_buf
;
5885 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5886 uint64_t psize
= BP_IS_HOLE(zio
->io_bp
) ? 0 : BP_GET_PSIZE(zio
->io_bp
);
5887 enum zio_compress compress
;
5888 fstrans_cookie_t cookie
= spl_fstrans_mark();
5890 ASSERT(HDR_HAS_L1HDR(hdr
));
5891 ASSERT(!zfs_refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
5892 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
5895 * If we're reexecuting this zio because the pool suspended, then
5896 * cleanup any state that was previously set the first time the
5897 * callback was invoked.
5899 if (zio
->io_flags
& ZIO_FLAG_REEXECUTED
) {
5900 arc_cksum_free(hdr
);
5901 arc_buf_unwatch(buf
);
5902 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5903 if (arc_buf_is_shared(buf
)) {
5904 arc_unshare_buf(hdr
, buf
);
5906 arc_hdr_free_pabd(hdr
);
5910 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5911 ASSERT(!HDR_SHARED_DATA(hdr
));
5912 ASSERT(!arc_buf_is_shared(buf
));
5914 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
5916 if (HDR_IO_IN_PROGRESS(hdr
))
5917 ASSERT(zio
->io_flags
& ZIO_FLAG_REEXECUTED
);
5919 arc_cksum_compute(buf
);
5920 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5922 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
5923 compress
= ZIO_COMPRESS_OFF
;
5925 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(zio
->io_bp
));
5926 compress
= BP_GET_COMPRESS(zio
->io_bp
);
5928 HDR_SET_PSIZE(hdr
, psize
);
5929 arc_hdr_set_compress(hdr
, compress
);
5932 * Fill the hdr with data. If the hdr is compressed, the data we want
5933 * is available from the zio, otherwise we can take it from the buf.
5935 * We might be able to share the buf's data with the hdr here. However,
5936 * doing so would cause the ARC to be full of linear ABDs if we write a
5937 * lot of shareable data. As a compromise, we check whether scattered
5938 * ABDs are allowed, and assume that if they are then the user wants
5939 * the ARC to be primarily filled with them regardless of the data being
5940 * written. Therefore, if they're allowed then we allocate one and copy
5941 * the data into it; otherwise, we share the data directly if we can.
5943 if (zfs_abd_scatter_enabled
|| !arc_can_share(hdr
, buf
)) {
5944 arc_hdr_alloc_pabd(hdr
);
5947 * Ideally, we would always copy the io_abd into b_pabd, but the
5948 * user may have disabled compressed ARC, thus we must check the
5949 * hdr's compression setting rather than the io_bp's.
5951 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
) {
5952 ASSERT3U(BP_GET_COMPRESS(zio
->io_bp
), !=,
5954 ASSERT3U(psize
, >, 0);
5956 abd_copy(hdr
->b_l1hdr
.b_pabd
, zio
->io_abd
, psize
);
5958 ASSERT3U(zio
->io_orig_size
, ==, arc_hdr_size(hdr
));
5960 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
5964 ASSERT3P(buf
->b_data
, ==, abd_to_buf(zio
->io_orig_abd
));
5965 ASSERT3U(zio
->io_orig_size
, ==, arc_buf_size(buf
));
5966 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
5968 arc_share_buf(hdr
, buf
);
5971 arc_hdr_verify(hdr
, zio
->io_bp
);
5972 spl_fstrans_unmark(cookie
);
5976 arc_write_children_ready(zio_t
*zio
)
5978 arc_write_callback_t
*callback
= zio
->io_private
;
5979 arc_buf_t
*buf
= callback
->awcb_buf
;
5981 callback
->awcb_children_ready(zio
, buf
, callback
->awcb_private
);
5985 * The SPA calls this callback for each physical write that happens on behalf
5986 * of a logical write. See the comment in dbuf_write_physdone() for details.
5989 arc_write_physdone(zio_t
*zio
)
5991 arc_write_callback_t
*cb
= zio
->io_private
;
5992 if (cb
->awcb_physdone
!= NULL
)
5993 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
5997 arc_write_done(zio_t
*zio
)
5999 arc_write_callback_t
*callback
= zio
->io_private
;
6000 arc_buf_t
*buf
= callback
->awcb_buf
;
6001 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
6003 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
6005 if (zio
->io_error
== 0) {
6006 arc_hdr_verify(hdr
, zio
->io_bp
);
6008 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
6009 buf_discard_identity(hdr
);
6011 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
6012 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
6015 ASSERT(HDR_EMPTY(hdr
));
6019 * If the block to be written was all-zero or compressed enough to be
6020 * embedded in the BP, no write was performed so there will be no
6021 * dva/birth/checksum. The buffer must therefore remain anonymous
6024 if (!HDR_EMPTY(hdr
)) {
6025 arc_buf_hdr_t
*exists
;
6026 kmutex_t
*hash_lock
;
6028 ASSERT3U(zio
->io_error
, ==, 0);
6030 arc_cksum_verify(buf
);
6032 exists
= buf_hash_insert(hdr
, &hash_lock
);
6033 if (exists
!= NULL
) {
6035 * This can only happen if we overwrite for
6036 * sync-to-convergence, because we remove
6037 * buffers from the hash table when we arc_free().
6039 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
6040 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
6041 panic("bad overwrite, hdr=%p exists=%p",
6042 (void *)hdr
, (void *)exists
);
6043 ASSERT(zfs_refcount_is_zero(
6044 &exists
->b_l1hdr
.b_refcnt
));
6045 arc_change_state(arc_anon
, exists
, hash_lock
);
6046 mutex_exit(hash_lock
);
6047 arc_hdr_destroy(exists
);
6048 exists
= buf_hash_insert(hdr
, &hash_lock
);
6049 ASSERT3P(exists
, ==, NULL
);
6050 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
6052 ASSERT(zio
->io_prop
.zp_nopwrite
);
6053 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
6054 panic("bad nopwrite, hdr=%p exists=%p",
6055 (void *)hdr
, (void *)exists
);
6058 ASSERT(hdr
->b_l1hdr
.b_bufcnt
== 1);
6059 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
6060 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
6061 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
6064 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6065 /* if it's not anon, we are doing a scrub */
6066 if (exists
== NULL
&& hdr
->b_l1hdr
.b_state
== arc_anon
)
6067 arc_access(hdr
, hash_lock
);
6068 mutex_exit(hash_lock
);
6070 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
6073 ASSERT(!zfs_refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
6074 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
6076 abd_put(zio
->io_abd
);
6077 kmem_free(callback
, sizeof (arc_write_callback_t
));
6081 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
6082 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
,
6083 const zio_prop_t
*zp
, arc_done_func_t
*ready
,
6084 arc_done_func_t
*children_ready
, arc_done_func_t
*physdone
,
6085 arc_done_func_t
*done
, void *private, zio_priority_t priority
,
6086 int zio_flags
, const zbookmark_phys_t
*zb
)
6088 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
6089 arc_write_callback_t
*callback
;
6091 zio_prop_t localprop
= *zp
;
6093 ASSERT3P(ready
, !=, NULL
);
6094 ASSERT3P(done
, !=, NULL
);
6095 ASSERT(!HDR_IO_ERROR(hdr
));
6096 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
6097 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
6098 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
6100 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
6101 if (ARC_BUF_COMPRESSED(buf
)) {
6103 * We're writing a pre-compressed buffer. Make the
6104 * compression algorithm requested by the zio_prop_t match
6105 * the pre-compressed buffer's compression algorithm.
6107 localprop
.zp_compress
= HDR_GET_COMPRESS(hdr
);
6109 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, arc_buf_size(buf
));
6110 zio_flags
|= ZIO_FLAG_RAW
;
6112 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
6113 callback
->awcb_ready
= ready
;
6114 callback
->awcb_children_ready
= children_ready
;
6115 callback
->awcb_physdone
= physdone
;
6116 callback
->awcb_done
= done
;
6117 callback
->awcb_private
= private;
6118 callback
->awcb_buf
= buf
;
6121 * The hdr's b_pabd is now stale, free it now. A new data block
6122 * will be allocated when the zio pipeline calls arc_write_ready().
6124 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
6126 * If the buf is currently sharing the data block with
6127 * the hdr then we need to break that relationship here.
6128 * The hdr will remain with a NULL data pointer and the
6129 * buf will take sole ownership of the block.
6131 if (arc_buf_is_shared(buf
)) {
6132 arc_unshare_buf(hdr
, buf
);
6134 arc_hdr_free_pabd(hdr
);
6136 VERIFY3P(buf
->b_data
, !=, NULL
);
6137 arc_hdr_set_compress(hdr
, ZIO_COMPRESS_OFF
);
6140 if (!(zio_flags
& ZIO_FLAG_RAW
))
6141 arc_hdr_set_compress(hdr
, ZIO_COMPRESS_OFF
);
6143 ASSERT(!arc_buf_is_shared(buf
));
6144 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
6146 zio
= zio_write(pio
, spa
, txg
, bp
,
6147 abd_get_from_buf(buf
->b_data
, HDR_GET_LSIZE(hdr
)),
6148 HDR_GET_LSIZE(hdr
), arc_buf_size(buf
), &localprop
, arc_write_ready
,
6149 (children_ready
!= NULL
) ? arc_write_children_ready
: NULL
,
6150 arc_write_physdone
, arc_write_done
, callback
,
6151 priority
, zio_flags
, zb
);
6157 arc_memory_throttle(uint64_t reserve
, uint64_t txg
)
6160 uint64_t available_memory
= arc_free_memory();
6161 static uint64_t page_load
= 0;
6162 static uint64_t last_txg
= 0;
6166 MIN(available_memory
, vmem_size(heap_arena
, VMEM_FREE
));
6169 if (available_memory
> arc_all_memory() * arc_lotsfree_percent
/ 100)
6172 if (txg
> last_txg
) {
6177 * If we are in pageout, we know that memory is already tight,
6178 * the arc is already going to be evicting, so we just want to
6179 * continue to let page writes occur as quickly as possible.
6181 if (current_is_kswapd()) {
6182 if (page_load
> MAX(arc_sys_free
/ 4, available_memory
) / 4) {
6183 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
6184 return (SET_ERROR(ERESTART
));
6186 /* Note: reserve is inflated, so we deflate */
6187 page_load
+= reserve
/ 8;
6189 } else if (page_load
> 0 && arc_reclaim_needed()) {
6190 /* memory is low, delay before restarting */
6191 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
6192 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
6193 return (SET_ERROR(EAGAIN
));
6201 arc_tempreserve_clear(uint64_t reserve
)
6203 atomic_add_64(&arc_tempreserve
, -reserve
);
6204 ASSERT((int64_t)arc_tempreserve
>= 0);
6208 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
6214 reserve
> arc_c
/4 &&
6215 reserve
* 4 > (2ULL << SPA_MAXBLOCKSHIFT
))
6216 arc_c
= MIN(arc_c_max
, reserve
* 4);
6219 * Throttle when the calculated memory footprint for the TXG
6220 * exceeds the target ARC size.
6222 if (reserve
> arc_c
) {
6223 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
6224 return (SET_ERROR(ERESTART
));
6228 * Don't count loaned bufs as in flight dirty data to prevent long
6229 * network delays from blocking transactions that are ready to be
6230 * assigned to a txg.
6233 /* assert that it has not wrapped around */
6234 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
6236 anon_size
= MAX((int64_t)(zfs_refcount_count(&arc_anon
->arcs_size
) -
6237 arc_loaned_bytes
), 0);
6240 * Writes will, almost always, require additional memory allocations
6241 * in order to compress/encrypt/etc the data. We therefore need to
6242 * make sure that there is sufficient available memory for this.
6244 error
= arc_memory_throttle(reserve
, txg
);
6249 * Throttle writes when the amount of dirty data in the cache
6250 * gets too large. We try to keep the cache less than half full
6251 * of dirty blocks so that our sync times don't grow too large.
6252 * Note: if two requests come in concurrently, we might let them
6253 * both succeed, when one of them should fail. Not a huge deal.
6256 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
6257 anon_size
> arc_c
/ 4) {
6258 uint64_t meta_esize
=
6260 &arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6261 uint64_t data_esize
=
6262 zfs_refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6263 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6264 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6265 arc_tempreserve
>> 10, meta_esize
>> 10,
6266 data_esize
>> 10, reserve
>> 10, arc_c
>> 10);
6267 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
6268 return (SET_ERROR(ERESTART
));
6270 atomic_add_64(&arc_tempreserve
, reserve
);
6275 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
6276 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
6278 size
->value
.ui64
= zfs_refcount_count(&state
->arcs_size
);
6279 evict_data
->value
.ui64
=
6280 zfs_refcount_count(&state
->arcs_esize
[ARC_BUFC_DATA
]);
6281 evict_metadata
->value
.ui64
=
6282 zfs_refcount_count(&state
->arcs_esize
[ARC_BUFC_METADATA
]);
6286 arc_kstat_update(kstat_t
*ksp
, int rw
)
6288 arc_stats_t
*as
= ksp
->ks_data
;
6290 if (rw
== KSTAT_WRITE
) {
6293 arc_kstat_update_state(arc_anon
,
6294 &as
->arcstat_anon_size
,
6295 &as
->arcstat_anon_evictable_data
,
6296 &as
->arcstat_anon_evictable_metadata
);
6297 arc_kstat_update_state(arc_mru
,
6298 &as
->arcstat_mru_size
,
6299 &as
->arcstat_mru_evictable_data
,
6300 &as
->arcstat_mru_evictable_metadata
);
6301 arc_kstat_update_state(arc_mru_ghost
,
6302 &as
->arcstat_mru_ghost_size
,
6303 &as
->arcstat_mru_ghost_evictable_data
,
6304 &as
->arcstat_mru_ghost_evictable_metadata
);
6305 arc_kstat_update_state(arc_mfu
,
6306 &as
->arcstat_mfu_size
,
6307 &as
->arcstat_mfu_evictable_data
,
6308 &as
->arcstat_mfu_evictable_metadata
);
6309 arc_kstat_update_state(arc_mfu_ghost
,
6310 &as
->arcstat_mfu_ghost_size
,
6311 &as
->arcstat_mfu_ghost_evictable_data
,
6312 &as
->arcstat_mfu_ghost_evictable_metadata
);
6314 as
->arcstat_memory_all_bytes
.value
.ui64
=
6316 as
->arcstat_memory_free_bytes
.value
.ui64
=
6318 as
->arcstat_memory_available_bytes
.value
.i64
=
6319 arc_available_memory();
6326 * This function *must* return indices evenly distributed between all
6327 * sublists of the multilist. This is needed due to how the ARC eviction
6328 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6329 * distributed between all sublists and uses this assumption when
6330 * deciding which sublist to evict from and how much to evict from it.
6333 arc_state_multilist_index_func(multilist_t
*ml
, void *obj
)
6335 arc_buf_hdr_t
*hdr
= obj
;
6338 * We rely on b_dva to generate evenly distributed index
6339 * numbers using buf_hash below. So, as an added precaution,
6340 * let's make sure we never add empty buffers to the arc lists.
6342 ASSERT(!HDR_EMPTY(hdr
));
6345 * The assumption here, is the hash value for a given
6346 * arc_buf_hdr_t will remain constant throughout its lifetime
6347 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
6348 * Thus, we don't need to store the header's sublist index
6349 * on insertion, as this index can be recalculated on removal.
6351 * Also, the low order bits of the hash value are thought to be
6352 * distributed evenly. Otherwise, in the case that the multilist
6353 * has a power of two number of sublists, each sublists' usage
6354 * would not be evenly distributed.
6356 return (buf_hash(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
) %
6357 multilist_get_num_sublists(ml
));
6361 * Called during module initialization and periodically thereafter to
6362 * apply reasonable changes to the exposed performance tunings. Non-zero
6363 * zfs_* values which differ from the currently set values will be applied.
6366 arc_tuning_update(void)
6368 uint64_t allmem
= arc_all_memory();
6369 unsigned long limit
;
6371 /* Valid range: 64M - <all physical memory> */
6372 if ((zfs_arc_max
) && (zfs_arc_max
!= arc_c_max
) &&
6373 (zfs_arc_max
>= 64 << 20) && (zfs_arc_max
< allmem
) &&
6374 (zfs_arc_max
> arc_c_min
)) {
6375 arc_c_max
= zfs_arc_max
;
6377 arc_p
= (arc_c
>> 1);
6378 if (arc_meta_limit
> arc_c_max
)
6379 arc_meta_limit
= arc_c_max
;
6380 if (arc_dnode_limit
> arc_meta_limit
)
6381 arc_dnode_limit
= arc_meta_limit
;
6384 /* Valid range: 32M - <arc_c_max> */
6385 if ((zfs_arc_min
) && (zfs_arc_min
!= arc_c_min
) &&
6386 (zfs_arc_min
>= 2ULL << SPA_MAXBLOCKSHIFT
) &&
6387 (zfs_arc_min
<= arc_c_max
)) {
6388 arc_c_min
= zfs_arc_min
;
6389 arc_c
= MAX(arc_c
, arc_c_min
);
6392 /* Valid range: 16M - <arc_c_max> */
6393 if ((zfs_arc_meta_min
) && (zfs_arc_meta_min
!= arc_meta_min
) &&
6394 (zfs_arc_meta_min
>= 1ULL << SPA_MAXBLOCKSHIFT
) &&
6395 (zfs_arc_meta_min
<= arc_c_max
)) {
6396 arc_meta_min
= zfs_arc_meta_min
;
6397 if (arc_meta_limit
< arc_meta_min
)
6398 arc_meta_limit
= arc_meta_min
;
6399 if (arc_dnode_limit
< arc_meta_min
)
6400 arc_dnode_limit
= arc_meta_min
;
6403 /* Valid range: <arc_meta_min> - <arc_c_max> */
6404 limit
= zfs_arc_meta_limit
? zfs_arc_meta_limit
:
6405 MIN(zfs_arc_meta_limit_percent
, 100) * arc_c_max
/ 100;
6406 if ((limit
!= arc_meta_limit
) &&
6407 (limit
>= arc_meta_min
) &&
6408 (limit
<= arc_c_max
))
6409 arc_meta_limit
= limit
;
6411 /* Valid range: <arc_meta_min> - <arc_meta_limit> */
6412 limit
= zfs_arc_dnode_limit
? zfs_arc_dnode_limit
:
6413 MIN(zfs_arc_dnode_limit_percent
, 100) * arc_meta_limit
/ 100;
6414 if ((limit
!= arc_dnode_limit
) &&
6415 (limit
>= arc_meta_min
) &&
6416 (limit
<= arc_meta_limit
))
6417 arc_dnode_limit
= limit
;
6419 /* Valid range: 1 - N */
6420 if (zfs_arc_grow_retry
)
6421 arc_grow_retry
= zfs_arc_grow_retry
;
6423 /* Valid range: 1 - N */
6424 if (zfs_arc_shrink_shift
) {
6425 arc_shrink_shift
= zfs_arc_shrink_shift
;
6426 arc_no_grow_shift
= MIN(arc_no_grow_shift
, arc_shrink_shift
-1);
6429 /* Valid range: 1 - N */
6430 if (zfs_arc_p_min_shift
)
6431 arc_p_min_shift
= zfs_arc_p_min_shift
;
6433 /* Valid range: 1 - N ticks */
6434 if (zfs_arc_min_prefetch_lifespan
)
6435 arc_min_prefetch_lifespan
= zfs_arc_min_prefetch_lifespan
;
6437 /* Valid range: 0 - 100 */
6438 if ((zfs_arc_lotsfree_percent
>= 0) &&
6439 (zfs_arc_lotsfree_percent
<= 100))
6440 arc_lotsfree_percent
= zfs_arc_lotsfree_percent
;
6442 /* Valid range: 0 - <all physical memory> */
6443 if ((zfs_arc_sys_free
) && (zfs_arc_sys_free
!= arc_sys_free
))
6444 arc_sys_free
= MIN(MAX(zfs_arc_sys_free
, 0), allmem
);
6449 arc_state_init(void)
6451 arc_anon
= &ARC_anon
;
6453 arc_mru_ghost
= &ARC_mru_ghost
;
6455 arc_mfu_ghost
= &ARC_mfu_ghost
;
6456 arc_l2c_only
= &ARC_l2c_only
;
6458 arc_mru
->arcs_list
[ARC_BUFC_METADATA
] =
6459 multilist_create(sizeof (arc_buf_hdr_t
),
6460 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6461 arc_state_multilist_index_func
);
6462 arc_mru
->arcs_list
[ARC_BUFC_DATA
] =
6463 multilist_create(sizeof (arc_buf_hdr_t
),
6464 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6465 arc_state_multilist_index_func
);
6466 arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
6467 multilist_create(sizeof (arc_buf_hdr_t
),
6468 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6469 arc_state_multilist_index_func
);
6470 arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
] =
6471 multilist_create(sizeof (arc_buf_hdr_t
),
6472 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6473 arc_state_multilist_index_func
);
6474 arc_mfu
->arcs_list
[ARC_BUFC_METADATA
] =
6475 multilist_create(sizeof (arc_buf_hdr_t
),
6476 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6477 arc_state_multilist_index_func
);
6478 arc_mfu
->arcs_list
[ARC_BUFC_DATA
] =
6479 multilist_create(sizeof (arc_buf_hdr_t
),
6480 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6481 arc_state_multilist_index_func
);
6482 arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
6483 multilist_create(sizeof (arc_buf_hdr_t
),
6484 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6485 arc_state_multilist_index_func
);
6486 arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
] =
6487 multilist_create(sizeof (arc_buf_hdr_t
),
6488 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6489 arc_state_multilist_index_func
);
6490 arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
] =
6491 multilist_create(sizeof (arc_buf_hdr_t
),
6492 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6493 arc_state_multilist_index_func
);
6494 arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
] =
6495 multilist_create(sizeof (arc_buf_hdr_t
),
6496 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6497 arc_state_multilist_index_func
);
6499 zfs_refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6500 zfs_refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6501 zfs_refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
6502 zfs_refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
6503 zfs_refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6504 zfs_refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6505 zfs_refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
6506 zfs_refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
6507 zfs_refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6508 zfs_refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6509 zfs_refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
6510 zfs_refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
6512 zfs_refcount_create(&arc_anon
->arcs_size
);
6513 zfs_refcount_create(&arc_mru
->arcs_size
);
6514 zfs_refcount_create(&arc_mru_ghost
->arcs_size
);
6515 zfs_refcount_create(&arc_mfu
->arcs_size
);
6516 zfs_refcount_create(&arc_mfu_ghost
->arcs_size
);
6517 zfs_refcount_create(&arc_l2c_only
->arcs_size
);
6519 arc_anon
->arcs_state
= ARC_STATE_ANON
;
6520 arc_mru
->arcs_state
= ARC_STATE_MRU
;
6521 arc_mru_ghost
->arcs_state
= ARC_STATE_MRU_GHOST
;
6522 arc_mfu
->arcs_state
= ARC_STATE_MFU
;
6523 arc_mfu_ghost
->arcs_state
= ARC_STATE_MFU_GHOST
;
6524 arc_l2c_only
->arcs_state
= ARC_STATE_L2C_ONLY
;
6528 arc_state_fini(void)
6530 zfs_refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6531 zfs_refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6532 zfs_refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
6533 zfs_refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
6534 zfs_refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6535 zfs_refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6536 zfs_refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
6537 zfs_refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
6538 zfs_refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6539 zfs_refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6540 zfs_refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
6541 zfs_refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
6543 zfs_refcount_destroy(&arc_anon
->arcs_size
);
6544 zfs_refcount_destroy(&arc_mru
->arcs_size
);
6545 zfs_refcount_destroy(&arc_mru_ghost
->arcs_size
);
6546 zfs_refcount_destroy(&arc_mfu
->arcs_size
);
6547 zfs_refcount_destroy(&arc_mfu_ghost
->arcs_size
);
6548 zfs_refcount_destroy(&arc_l2c_only
->arcs_size
);
6550 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
6551 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
6552 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
6553 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
6554 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
6555 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
6556 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
6557 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
6558 multilist_destroy(arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]);
6559 multilist_destroy(arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]);
6563 arc_target_bytes(void)
6571 uint64_t percent
, allmem
= arc_all_memory();
6573 mutex_init(&arc_reclaim_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
6574 cv_init(&arc_reclaim_thread_cv
, NULL
, CV_DEFAULT
, NULL
);
6575 cv_init(&arc_reclaim_waiters_cv
, NULL
, CV_DEFAULT
, NULL
);
6577 /* Convert seconds to clock ticks */
6578 arc_min_prefetch_lifespan
= 1 * hz
;
6582 * Register a shrinker to support synchronous (direct) memory
6583 * reclaim from the arc. This is done to prevent kswapd from
6584 * swapping out pages when it is preferable to shrink the arc.
6586 spl_register_shrinker(&arc_shrinker
);
6588 /* Set to 1/64 of all memory or a minimum of 512K */
6589 arc_sys_free
= MAX(allmem
/ 64, (512 * 1024));
6593 /* Set max to 1/2 of all memory */
6594 arc_c_max
= allmem
/ 2;
6597 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more */
6598 arc_c_min
= MAX(allmem
/ 32, 2ULL << SPA_MAXBLOCKSHIFT
);
6601 * In userland, there's only the memory pressure that we artificially
6602 * create (see arc_available_memory()). Don't let arc_c get too
6603 * small, because it can cause transactions to be larger than
6604 * arc_c, causing arc_tempreserve_space() to fail.
6606 arc_c_min
= MAX(arc_c_max
/ 2, 2ULL << SPA_MAXBLOCKSHIFT
);
6610 arc_p
= (arc_c
>> 1);
6613 /* Set min to 1/2 of arc_c_min */
6614 arc_meta_min
= 1ULL << SPA_MAXBLOCKSHIFT
;
6615 /* Initialize maximum observed usage to zero */
6618 * Set arc_meta_limit to a percent of arc_c_max with a floor of
6619 * arc_meta_min, and a ceiling of arc_c_max.
6621 percent
= MIN(zfs_arc_meta_limit_percent
, 100);
6622 arc_meta_limit
= MAX(arc_meta_min
, (percent
* arc_c_max
) / 100);
6623 percent
= MIN(zfs_arc_dnode_limit_percent
, 100);
6624 arc_dnode_limit
= (percent
* arc_meta_limit
) / 100;
6626 /* Apply user specified tunings */
6627 arc_tuning_update();
6629 /* if kmem_flags are set, lets try to use less memory */
6630 if (kmem_debugging())
6632 if (arc_c
< arc_c_min
)
6638 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
6639 offsetof(arc_prune_t
, p_node
));
6640 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6642 arc_prune_taskq
= taskq_create("arc_prune", max_ncpus
, defclsyspri
,
6643 max_ncpus
, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
6645 arc_reclaim_thread_exit
= B_FALSE
;
6647 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
6648 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
6650 if (arc_ksp
!= NULL
) {
6651 arc_ksp
->ks_data
= &arc_stats
;
6652 arc_ksp
->ks_update
= arc_kstat_update
;
6653 kstat_install(arc_ksp
);
6656 (void) thread_create(NULL
, 0, arc_reclaim_thread
, NULL
, 0, &p0
,
6657 TS_RUN
, defclsyspri
);
6663 * Calculate maximum amount of dirty data per pool.
6665 * If it has been set by a module parameter, take that.
6666 * Otherwise, use a percentage of physical memory defined by
6667 * zfs_dirty_data_max_percent (default 10%) with a cap at
6668 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
6670 if (zfs_dirty_data_max_max
== 0)
6671 zfs_dirty_data_max_max
= MIN(4ULL * 1024 * 1024 * 1024,
6672 allmem
* zfs_dirty_data_max_max_percent
/ 100);
6674 if (zfs_dirty_data_max
== 0) {
6675 zfs_dirty_data_max
= allmem
*
6676 zfs_dirty_data_max_percent
/ 100;
6677 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
6678 zfs_dirty_data_max_max
);
6688 spl_unregister_shrinker(&arc_shrinker
);
6689 #endif /* _KERNEL */
6691 mutex_enter(&arc_reclaim_lock
);
6692 arc_reclaim_thread_exit
= B_TRUE
;
6694 * The reclaim thread will set arc_reclaim_thread_exit back to
6695 * B_FALSE when it is finished exiting; we're waiting for that.
6697 while (arc_reclaim_thread_exit
) {
6698 cv_signal(&arc_reclaim_thread_cv
);
6699 cv_wait(&arc_reclaim_thread_cv
, &arc_reclaim_lock
);
6701 mutex_exit(&arc_reclaim_lock
);
6703 /* Use B_TRUE to ensure *all* buffers are evicted */
6704 arc_flush(NULL
, B_TRUE
);
6708 if (arc_ksp
!= NULL
) {
6709 kstat_delete(arc_ksp
);
6713 taskq_wait(arc_prune_taskq
);
6714 taskq_destroy(arc_prune_taskq
);
6716 mutex_enter(&arc_prune_mtx
);
6717 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
6718 list_remove(&arc_prune_list
, p
);
6719 zfs_refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
6720 zfs_refcount_destroy(&p
->p_refcnt
);
6721 kmem_free(p
, sizeof (*p
));
6723 mutex_exit(&arc_prune_mtx
);
6725 list_destroy(&arc_prune_list
);
6726 mutex_destroy(&arc_prune_mtx
);
6727 mutex_destroy(&arc_reclaim_lock
);
6728 cv_destroy(&arc_reclaim_thread_cv
);
6729 cv_destroy(&arc_reclaim_waiters_cv
);
6734 ASSERT0(arc_loaned_bytes
);
6740 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6741 * It uses dedicated storage devices to hold cached data, which are populated
6742 * using large infrequent writes. The main role of this cache is to boost
6743 * the performance of random read workloads. The intended L2ARC devices
6744 * include short-stroked disks, solid state disks, and other media with
6745 * substantially faster read latency than disk.
6747 * +-----------------------+
6749 * +-----------------------+
6752 * l2arc_feed_thread() arc_read()
6756 * +---------------+ |
6758 * +---------------+ |
6763 * +-------+ +-------+
6765 * | cache | | cache |
6766 * +-------+ +-------+
6767 * +=========+ .-----.
6768 * : L2ARC : |-_____-|
6769 * : devices : | Disks |
6770 * +=========+ `-_____-'
6772 * Read requests are satisfied from the following sources, in order:
6775 * 2) vdev cache of L2ARC devices
6777 * 4) vdev cache of disks
6780 * Some L2ARC device types exhibit extremely slow write performance.
6781 * To accommodate for this there are some significant differences between
6782 * the L2ARC and traditional cache design:
6784 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6785 * the ARC behave as usual, freeing buffers and placing headers on ghost
6786 * lists. The ARC does not send buffers to the L2ARC during eviction as
6787 * this would add inflated write latencies for all ARC memory pressure.
6789 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6790 * It does this by periodically scanning buffers from the eviction-end of
6791 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6792 * not already there. It scans until a headroom of buffers is satisfied,
6793 * which itself is a buffer for ARC eviction. If a compressible buffer is
6794 * found during scanning and selected for writing to an L2ARC device, we
6795 * temporarily boost scanning headroom during the next scan cycle to make
6796 * sure we adapt to compression effects (which might significantly reduce
6797 * the data volume we write to L2ARC). The thread that does this is
6798 * l2arc_feed_thread(), illustrated below; example sizes are included to
6799 * provide a better sense of ratio than this diagram:
6802 * +---------------------+----------+
6803 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6804 * +---------------------+----------+ | o L2ARC eligible
6805 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6806 * +---------------------+----------+ |
6807 * 15.9 Gbytes ^ 32 Mbytes |
6809 * l2arc_feed_thread()
6811 * l2arc write hand <--[oooo]--'
6815 * +==============================+
6816 * L2ARC dev |####|#|###|###| |####| ... |
6817 * +==============================+
6820 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6821 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6822 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6823 * safe to say that this is an uncommon case, since buffers at the end of
6824 * the ARC lists have moved there due to inactivity.
6826 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6827 * then the L2ARC simply misses copying some buffers. This serves as a
6828 * pressure valve to prevent heavy read workloads from both stalling the ARC
6829 * with waits and clogging the L2ARC with writes. This also helps prevent
6830 * the potential for the L2ARC to churn if it attempts to cache content too
6831 * quickly, such as during backups of the entire pool.
6833 * 5. After system boot and before the ARC has filled main memory, there are
6834 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6835 * lists can remain mostly static. Instead of searching from tail of these
6836 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6837 * for eligible buffers, greatly increasing its chance of finding them.
6839 * The L2ARC device write speed is also boosted during this time so that
6840 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6841 * there are no L2ARC reads, and no fear of degrading read performance
6842 * through increased writes.
6844 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6845 * the vdev queue can aggregate them into larger and fewer writes. Each
6846 * device is written to in a rotor fashion, sweeping writes through
6847 * available space then repeating.
6849 * 7. The L2ARC does not store dirty content. It never needs to flush
6850 * write buffers back to disk based storage.
6852 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6853 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6855 * The performance of the L2ARC can be tweaked by a number of tunables, which
6856 * may be necessary for different workloads:
6858 * l2arc_write_max max write bytes per interval
6859 * l2arc_write_boost extra write bytes during device warmup
6860 * l2arc_noprefetch skip caching prefetched buffers
6861 * l2arc_headroom number of max device writes to precache
6862 * l2arc_headroom_boost when we find compressed buffers during ARC
6863 * scanning, we multiply headroom by this
6864 * percentage factor for the next scan cycle,
6865 * since more compressed buffers are likely to
6867 * l2arc_feed_secs seconds between L2ARC writing
6869 * Tunables may be removed or added as future performance improvements are
6870 * integrated, and also may become zpool properties.
6872 * There are three key functions that control how the L2ARC warms up:
6874 * l2arc_write_eligible() check if a buffer is eligible to cache
6875 * l2arc_write_size() calculate how much to write
6876 * l2arc_write_interval() calculate sleep delay between writes
6878 * These three functions determine what to write, how much, and how quickly
6883 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*hdr
)
6886 * A buffer is *not* eligible for the L2ARC if it:
6887 * 1. belongs to a different spa.
6888 * 2. is already cached on the L2ARC.
6889 * 3. has an I/O in progress (it may be an incomplete read).
6890 * 4. is flagged not eligible (zfs property).
6892 if (hdr
->b_spa
!= spa_guid
|| HDR_HAS_L2HDR(hdr
) ||
6893 HDR_IO_IN_PROGRESS(hdr
) || !HDR_L2CACHE(hdr
))
6900 l2arc_write_size(void)
6905 * Make sure our globals have meaningful values in case the user
6908 size
= l2arc_write_max
;
6910 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
6911 "be greater than zero, resetting it to the default (%d)",
6913 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
6916 if (arc_warm
== B_FALSE
)
6917 size
+= l2arc_write_boost
;
6924 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
6926 clock_t interval
, next
, now
;
6929 * If the ARC lists are busy, increase our write rate; if the
6930 * lists are stale, idle back. This is achieved by checking
6931 * how much we previously wrote - if it was more than half of
6932 * what we wanted, schedule the next write much sooner.
6934 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
6935 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
6937 interval
= hz
* l2arc_feed_secs
;
6939 now
= ddi_get_lbolt();
6940 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
6946 * Cycle through L2ARC devices. This is how L2ARC load balances.
6947 * If a device is returned, this also returns holding the spa config lock.
6949 static l2arc_dev_t
*
6950 l2arc_dev_get_next(void)
6952 l2arc_dev_t
*first
, *next
= NULL
;
6955 * Lock out the removal of spas (spa_namespace_lock), then removal
6956 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6957 * both locks will be dropped and a spa config lock held instead.
6959 mutex_enter(&spa_namespace_lock
);
6960 mutex_enter(&l2arc_dev_mtx
);
6962 /* if there are no vdevs, there is nothing to do */
6963 if (l2arc_ndev
== 0)
6967 next
= l2arc_dev_last
;
6969 /* loop around the list looking for a non-faulted vdev */
6971 next
= list_head(l2arc_dev_list
);
6973 next
= list_next(l2arc_dev_list
, next
);
6975 next
= list_head(l2arc_dev_list
);
6978 /* if we have come back to the start, bail out */
6981 else if (next
== first
)
6984 } while (vdev_is_dead(next
->l2ad_vdev
));
6986 /* if we were unable to find any usable vdevs, return NULL */
6987 if (vdev_is_dead(next
->l2ad_vdev
))
6990 l2arc_dev_last
= next
;
6993 mutex_exit(&l2arc_dev_mtx
);
6996 * Grab the config lock to prevent the 'next' device from being
6997 * removed while we are writing to it.
7000 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
7001 mutex_exit(&spa_namespace_lock
);
7007 * Free buffers that were tagged for destruction.
7010 l2arc_do_free_on_write(void)
7013 l2arc_data_free_t
*df
, *df_prev
;
7015 mutex_enter(&l2arc_free_on_write_mtx
);
7016 buflist
= l2arc_free_on_write
;
7018 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
7019 df_prev
= list_prev(buflist
, df
);
7020 ASSERT3P(df
->l2df_abd
, !=, NULL
);
7021 abd_free(df
->l2df_abd
);
7022 list_remove(buflist
, df
);
7023 kmem_free(df
, sizeof (l2arc_data_free_t
));
7026 mutex_exit(&l2arc_free_on_write_mtx
);
7030 * A write to a cache device has completed. Update all headers to allow
7031 * reads from these buffers to begin.
7034 l2arc_write_done(zio_t
*zio
)
7036 l2arc_write_callback_t
*cb
;
7039 arc_buf_hdr_t
*head
, *hdr
, *hdr_prev
;
7040 kmutex_t
*hash_lock
;
7041 int64_t bytes_dropped
= 0;
7043 cb
= zio
->io_private
;
7044 ASSERT3P(cb
, !=, NULL
);
7045 dev
= cb
->l2wcb_dev
;
7046 ASSERT3P(dev
, !=, NULL
);
7047 head
= cb
->l2wcb_head
;
7048 ASSERT3P(head
, !=, NULL
);
7049 buflist
= &dev
->l2ad_buflist
;
7050 ASSERT3P(buflist
, !=, NULL
);
7051 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
7052 l2arc_write_callback_t
*, cb
);
7054 if (zio
->io_error
!= 0)
7055 ARCSTAT_BUMP(arcstat_l2_writes_error
);
7058 * All writes completed, or an error was hit.
7061 mutex_enter(&dev
->l2ad_mtx
);
7062 for (hdr
= list_prev(buflist
, head
); hdr
; hdr
= hdr_prev
) {
7063 hdr_prev
= list_prev(buflist
, hdr
);
7065 hash_lock
= HDR_LOCK(hdr
);
7068 * We cannot use mutex_enter or else we can deadlock
7069 * with l2arc_write_buffers (due to swapping the order
7070 * the hash lock and l2ad_mtx are taken).
7072 if (!mutex_tryenter(hash_lock
)) {
7074 * Missed the hash lock. We must retry so we
7075 * don't leave the ARC_FLAG_L2_WRITING bit set.
7077 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry
);
7080 * We don't want to rescan the headers we've
7081 * already marked as having been written out, so
7082 * we reinsert the head node so we can pick up
7083 * where we left off.
7085 list_remove(buflist
, head
);
7086 list_insert_after(buflist
, hdr
, head
);
7088 mutex_exit(&dev
->l2ad_mtx
);
7091 * We wait for the hash lock to become available
7092 * to try and prevent busy waiting, and increase
7093 * the chance we'll be able to acquire the lock
7094 * the next time around.
7096 mutex_enter(hash_lock
);
7097 mutex_exit(hash_lock
);
7102 * We could not have been moved into the arc_l2c_only
7103 * state while in-flight due to our ARC_FLAG_L2_WRITING
7104 * bit being set. Let's just ensure that's being enforced.
7106 ASSERT(HDR_HAS_L1HDR(hdr
));
7109 * Skipped - drop L2ARC entry and mark the header as no
7110 * longer L2 eligibile.
7112 if (zio
->io_error
!= 0) {
7114 * Error - drop L2ARC entry.
7116 list_remove(buflist
, hdr
);
7117 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
7119 ARCSTAT_INCR(arcstat_l2_psize
, -arc_hdr_size(hdr
));
7120 ARCSTAT_INCR(arcstat_l2_lsize
, -HDR_GET_LSIZE(hdr
));
7122 bytes_dropped
+= arc_hdr_size(hdr
);
7123 (void) zfs_refcount_remove_many(&dev
->l2ad_alloc
,
7124 arc_hdr_size(hdr
), hdr
);
7128 * Allow ARC to begin reads and ghost list evictions to
7131 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_WRITING
);
7133 mutex_exit(hash_lock
);
7136 atomic_inc_64(&l2arc_writes_done
);
7137 list_remove(buflist
, head
);
7138 ASSERT(!HDR_HAS_L1HDR(head
));
7139 kmem_cache_free(hdr_l2only_cache
, head
);
7140 mutex_exit(&dev
->l2ad_mtx
);
7142 vdev_space_update(dev
->l2ad_vdev
, -bytes_dropped
, 0, 0);
7144 l2arc_do_free_on_write();
7146 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
7150 * A read to a cache device completed. Validate buffer contents before
7151 * handing over to the regular ARC routines.
7154 l2arc_read_done(zio_t
*zio
)
7156 l2arc_read_callback_t
*cb
;
7158 kmutex_t
*hash_lock
;
7159 boolean_t valid_cksum
;
7161 ASSERT3P(zio
->io_vd
, !=, NULL
);
7162 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
7164 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
7166 cb
= zio
->io_private
;
7167 ASSERT3P(cb
, !=, NULL
);
7168 hdr
= cb
->l2rcb_hdr
;
7169 ASSERT3P(hdr
, !=, NULL
);
7171 hash_lock
= HDR_LOCK(hdr
);
7172 mutex_enter(hash_lock
);
7173 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
7176 * If the data was read into a temporary buffer,
7177 * move it and free the buffer.
7179 if (cb
->l2rcb_abd
!= NULL
) {
7180 ASSERT3U(arc_hdr_size(hdr
), <, zio
->io_size
);
7181 if (zio
->io_error
== 0) {
7182 abd_copy(hdr
->b_l1hdr
.b_pabd
, cb
->l2rcb_abd
,
7187 * The following must be done regardless of whether
7188 * there was an error:
7189 * - free the temporary buffer
7190 * - point zio to the real ARC buffer
7191 * - set zio size accordingly
7192 * These are required because zio is either re-used for
7193 * an I/O of the block in the case of the error
7194 * or the zio is passed to arc_read_done() and it
7197 abd_free(cb
->l2rcb_abd
);
7198 zio
->io_size
= zio
->io_orig_size
= arc_hdr_size(hdr
);
7199 zio
->io_abd
= zio
->io_orig_abd
= hdr
->b_l1hdr
.b_pabd
;
7202 ASSERT3P(zio
->io_abd
, !=, NULL
);
7205 * Check this survived the L2ARC journey.
7207 ASSERT3P(zio
->io_abd
, ==, hdr
->b_l1hdr
.b_pabd
);
7208 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
7209 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
7211 valid_cksum
= arc_cksum_is_equal(hdr
, zio
);
7212 if (valid_cksum
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
7213 mutex_exit(hash_lock
);
7214 zio
->io_private
= hdr
;
7217 mutex_exit(hash_lock
);
7219 * Buffer didn't survive caching. Increment stats and
7220 * reissue to the original storage device.
7222 if (zio
->io_error
!= 0) {
7223 ARCSTAT_BUMP(arcstat_l2_io_error
);
7225 zio
->io_error
= SET_ERROR(EIO
);
7228 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
7231 * If there's no waiter, issue an async i/o to the primary
7232 * storage now. If there *is* a waiter, the caller must
7233 * issue the i/o in a context where it's OK to block.
7235 if (zio
->io_waiter
== NULL
) {
7236 zio_t
*pio
= zio_unique_parent(zio
);
7238 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
7240 zio_nowait(zio_read(pio
, zio
->io_spa
, zio
->io_bp
,
7241 hdr
->b_l1hdr
.b_pabd
, zio
->io_size
, arc_read_done
,
7242 hdr
, zio
->io_priority
, cb
->l2rcb_flags
,
7247 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
7251 * This is the list priority from which the L2ARC will search for pages to
7252 * cache. This is used within loops (0..3) to cycle through lists in the
7253 * desired order. This order can have a significant effect on cache
7256 * Currently the metadata lists are hit first, MFU then MRU, followed by
7257 * the data lists. This function returns a locked list, and also returns
7260 static multilist_sublist_t
*
7261 l2arc_sublist_lock(int list_num
)
7263 multilist_t
*ml
= NULL
;
7266 ASSERT(list_num
>= 0 && list_num
< L2ARC_FEED_TYPES
);
7270 ml
= arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
7273 ml
= arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
7276 ml
= arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
7279 ml
= arc_mru
->arcs_list
[ARC_BUFC_DATA
];
7286 * Return a randomly-selected sublist. This is acceptable
7287 * because the caller feeds only a little bit of data for each
7288 * call (8MB). Subsequent calls will result in different
7289 * sublists being selected.
7291 idx
= multilist_get_random_index(ml
);
7292 return (multilist_sublist_lock(ml
, idx
));
7296 * Evict buffers from the device write hand to the distance specified in
7297 * bytes. This distance may span populated buffers, it may span nothing.
7298 * This is clearing a region on the L2ARC device ready for writing.
7299 * If the 'all' boolean is set, every buffer is evicted.
7302 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
7305 arc_buf_hdr_t
*hdr
, *hdr_prev
;
7306 kmutex_t
*hash_lock
;
7309 buflist
= &dev
->l2ad_buflist
;
7311 if (!all
&& dev
->l2ad_first
) {
7313 * This is the first sweep through the device. There is
7319 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
7321 * When nearing the end of the device, evict to the end
7322 * before the device write hand jumps to the start.
7324 taddr
= dev
->l2ad_end
;
7326 taddr
= dev
->l2ad_hand
+ distance
;
7328 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
7329 uint64_t, taddr
, boolean_t
, all
);
7332 mutex_enter(&dev
->l2ad_mtx
);
7333 for (hdr
= list_tail(buflist
); hdr
; hdr
= hdr_prev
) {
7334 hdr_prev
= list_prev(buflist
, hdr
);
7336 hash_lock
= HDR_LOCK(hdr
);
7339 * We cannot use mutex_enter or else we can deadlock
7340 * with l2arc_write_buffers (due to swapping the order
7341 * the hash lock and l2ad_mtx are taken).
7343 if (!mutex_tryenter(hash_lock
)) {
7345 * Missed the hash lock. Retry.
7347 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
7348 mutex_exit(&dev
->l2ad_mtx
);
7349 mutex_enter(hash_lock
);
7350 mutex_exit(hash_lock
);
7354 if (HDR_L2_WRITE_HEAD(hdr
)) {
7356 * We hit a write head node. Leave it for
7357 * l2arc_write_done().
7359 list_remove(buflist
, hdr
);
7360 mutex_exit(hash_lock
);
7364 if (!all
&& HDR_HAS_L2HDR(hdr
) &&
7365 (hdr
->b_l2hdr
.b_daddr
> taddr
||
7366 hdr
->b_l2hdr
.b_daddr
< dev
->l2ad_hand
)) {
7368 * We've evicted to the target address,
7369 * or the end of the device.
7371 mutex_exit(hash_lock
);
7375 ASSERT(HDR_HAS_L2HDR(hdr
));
7376 if (!HDR_HAS_L1HDR(hdr
)) {
7377 ASSERT(!HDR_L2_READING(hdr
));
7379 * This doesn't exist in the ARC. Destroy.
7380 * arc_hdr_destroy() will call list_remove()
7381 * and decrement arcstat_l2_lsize.
7383 arc_change_state(arc_anon
, hdr
, hash_lock
);
7384 arc_hdr_destroy(hdr
);
7386 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_l2c_only
);
7387 ARCSTAT_BUMP(arcstat_l2_evict_l1cached
);
7389 * Invalidate issued or about to be issued
7390 * reads, since we may be about to write
7391 * over this location.
7393 if (HDR_L2_READING(hdr
)) {
7394 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
7395 arc_hdr_set_flags(hdr
, ARC_FLAG_L2_EVICTED
);
7398 /* Ensure this header has finished being written */
7399 ASSERT(!HDR_L2_WRITING(hdr
));
7401 arc_hdr_l2hdr_destroy(hdr
);
7403 mutex_exit(hash_lock
);
7405 mutex_exit(&dev
->l2ad_mtx
);
7409 * Find and write ARC buffers to the L2ARC device.
7411 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7412 * for reading until they have completed writing.
7413 * The headroom_boost is an in-out parameter used to maintain headroom boost
7414 * state between calls to this function.
7416 * Returns the number of bytes actually written (which may be smaller than
7417 * the delta by which the device hand has changed due to alignment).
7420 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
)
7422 arc_buf_hdr_t
*hdr
, *hdr_prev
, *head
;
7423 uint64_t write_asize
, write_psize
, write_lsize
, headroom
;
7425 l2arc_write_callback_t
*cb
;
7427 uint64_t guid
= spa_load_guid(spa
);
7430 ASSERT3P(dev
->l2ad_vdev
, !=, NULL
);
7433 write_lsize
= write_asize
= write_psize
= 0;
7435 head
= kmem_cache_alloc(hdr_l2only_cache
, KM_PUSHPAGE
);
7436 arc_hdr_set_flags(head
, ARC_FLAG_L2_WRITE_HEAD
| ARC_FLAG_HAS_L2HDR
);
7439 * Copy buffers for L2ARC writing.
7441 for (try = 0; try < L2ARC_FEED_TYPES
; try++) {
7442 multilist_sublist_t
*mls
= l2arc_sublist_lock(try);
7443 uint64_t passed_sz
= 0;
7445 VERIFY3P(mls
, !=, NULL
);
7448 * L2ARC fast warmup.
7450 * Until the ARC is warm and starts to evict, read from the
7451 * head of the ARC lists rather than the tail.
7453 if (arc_warm
== B_FALSE
)
7454 hdr
= multilist_sublist_head(mls
);
7456 hdr
= multilist_sublist_tail(mls
);
7458 headroom
= target_sz
* l2arc_headroom
;
7459 if (zfs_compressed_arc_enabled
)
7460 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
7462 for (; hdr
; hdr
= hdr_prev
) {
7463 kmutex_t
*hash_lock
;
7465 if (arc_warm
== B_FALSE
)
7466 hdr_prev
= multilist_sublist_next(mls
, hdr
);
7468 hdr_prev
= multilist_sublist_prev(mls
, hdr
);
7470 hash_lock
= HDR_LOCK(hdr
);
7471 if (!mutex_tryenter(hash_lock
)) {
7473 * Skip this buffer rather than waiting.
7478 passed_sz
+= HDR_GET_LSIZE(hdr
);
7479 if (passed_sz
> headroom
) {
7483 mutex_exit(hash_lock
);
7487 if (!l2arc_write_eligible(guid
, hdr
)) {
7488 mutex_exit(hash_lock
);
7493 * We rely on the L1 portion of the header below, so
7494 * it's invalid for this header to have been evicted out
7495 * of the ghost cache, prior to being written out. The
7496 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7498 ASSERT(HDR_HAS_L1HDR(hdr
));
7500 ASSERT3U(HDR_GET_PSIZE(hdr
), >, 0);
7501 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
7502 ASSERT3U(arc_hdr_size(hdr
), >, 0);
7503 uint64_t psize
= arc_hdr_size(hdr
);
7504 uint64_t asize
= vdev_psize_to_asize(dev
->l2ad_vdev
,
7507 if ((write_asize
+ asize
) > target_sz
) {
7509 mutex_exit(hash_lock
);
7515 * Insert a dummy header on the buflist so
7516 * l2arc_write_done() can find where the
7517 * write buffers begin without searching.
7519 mutex_enter(&dev
->l2ad_mtx
);
7520 list_insert_head(&dev
->l2ad_buflist
, head
);
7521 mutex_exit(&dev
->l2ad_mtx
);
7524 sizeof (l2arc_write_callback_t
), KM_SLEEP
);
7525 cb
->l2wcb_dev
= dev
;
7526 cb
->l2wcb_head
= head
;
7527 pio
= zio_root(spa
, l2arc_write_done
, cb
,
7531 hdr
->b_l2hdr
.b_dev
= dev
;
7532 hdr
->b_l2hdr
.b_hits
= 0;
7534 hdr
->b_l2hdr
.b_daddr
= dev
->l2ad_hand
;
7535 arc_hdr_set_flags(hdr
,
7536 ARC_FLAG_L2_WRITING
| ARC_FLAG_HAS_L2HDR
);
7538 mutex_enter(&dev
->l2ad_mtx
);
7539 list_insert_head(&dev
->l2ad_buflist
, hdr
);
7540 mutex_exit(&dev
->l2ad_mtx
);
7542 (void) zfs_refcount_add_many(&dev
->l2ad_alloc
, psize
,
7546 * Normally the L2ARC can use the hdr's data, but if
7547 * we're sharing data between the hdr and one of its
7548 * bufs, L2ARC needs its own copy of the data so that
7549 * the ZIO below can't race with the buf consumer.
7550 * Another case where we need to create a copy of the
7551 * data is when the buffer size is not device-aligned
7552 * and we need to pad the block to make it such.
7553 * That also keeps the clock hand suitably aligned.
7555 * To ensure that the copy will be available for the
7556 * lifetime of the ZIO and be cleaned up afterwards, we
7557 * add it to the l2arc_free_on_write queue.
7560 if (!HDR_SHARED_DATA(hdr
) && psize
== asize
) {
7561 to_write
= hdr
->b_l1hdr
.b_pabd
;
7563 to_write
= abd_alloc_for_io(asize
,
7564 HDR_ISTYPE_METADATA(hdr
));
7565 abd_copy(to_write
, hdr
->b_l1hdr
.b_pabd
, psize
);
7566 if (asize
!= psize
) {
7567 abd_zero_off(to_write
, psize
,
7570 l2arc_free_abd_on_write(to_write
, asize
,
7573 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
7574 hdr
->b_l2hdr
.b_daddr
, asize
, to_write
,
7575 ZIO_CHECKSUM_OFF
, NULL
, hdr
,
7576 ZIO_PRIORITY_ASYNC_WRITE
,
7577 ZIO_FLAG_CANFAIL
, B_FALSE
);
7579 write_lsize
+= HDR_GET_LSIZE(hdr
);
7580 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
7583 write_psize
+= psize
;
7584 write_asize
+= asize
;
7585 dev
->l2ad_hand
+= asize
;
7587 mutex_exit(hash_lock
);
7589 (void) zio_nowait(wzio
);
7592 multilist_sublist_unlock(mls
);
7598 /* No buffers selected for writing? */
7600 ASSERT0(write_lsize
);
7601 ASSERT(!HDR_HAS_L1HDR(head
));
7602 kmem_cache_free(hdr_l2only_cache
, head
);
7606 ASSERT3U(write_asize
, <=, target_sz
);
7607 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
7608 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_psize
);
7609 ARCSTAT_INCR(arcstat_l2_lsize
, write_lsize
);
7610 ARCSTAT_INCR(arcstat_l2_psize
, write_psize
);
7611 vdev_space_update(dev
->l2ad_vdev
, write_psize
, 0, 0);
7614 * Bump device hand to the device start if it is approaching the end.
7615 * l2arc_evict() will already have evicted ahead for this case.
7617 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
7618 dev
->l2ad_hand
= dev
->l2ad_start
;
7619 dev
->l2ad_first
= B_FALSE
;
7622 dev
->l2ad_writing
= B_TRUE
;
7623 (void) zio_wait(pio
);
7624 dev
->l2ad_writing
= B_FALSE
;
7626 return (write_asize
);
7630 * This thread feeds the L2ARC at regular intervals. This is the beating
7631 * heart of the L2ARC.
7634 l2arc_feed_thread(void)
7639 uint64_t size
, wrote
;
7640 clock_t begin
, next
= ddi_get_lbolt();
7641 fstrans_cookie_t cookie
;
7643 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
7645 mutex_enter(&l2arc_feed_thr_lock
);
7647 cookie
= spl_fstrans_mark();
7648 while (l2arc_thread_exit
== 0) {
7649 CALLB_CPR_SAFE_BEGIN(&cpr
);
7650 (void) cv_timedwait_sig(&l2arc_feed_thr_cv
,
7651 &l2arc_feed_thr_lock
, next
);
7652 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
7653 next
= ddi_get_lbolt() + hz
;
7656 * Quick check for L2ARC devices.
7658 mutex_enter(&l2arc_dev_mtx
);
7659 if (l2arc_ndev
== 0) {
7660 mutex_exit(&l2arc_dev_mtx
);
7663 mutex_exit(&l2arc_dev_mtx
);
7664 begin
= ddi_get_lbolt();
7667 * This selects the next l2arc device to write to, and in
7668 * doing so the next spa to feed from: dev->l2ad_spa. This
7669 * will return NULL if there are now no l2arc devices or if
7670 * they are all faulted.
7672 * If a device is returned, its spa's config lock is also
7673 * held to prevent device removal. l2arc_dev_get_next()
7674 * will grab and release l2arc_dev_mtx.
7676 if ((dev
= l2arc_dev_get_next()) == NULL
)
7679 spa
= dev
->l2ad_spa
;
7680 ASSERT3P(spa
, !=, NULL
);
7683 * If the pool is read-only then force the feed thread to
7684 * sleep a little longer.
7686 if (!spa_writeable(spa
)) {
7687 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
7688 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7693 * Avoid contributing to memory pressure.
7695 if (arc_reclaim_needed()) {
7696 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
7697 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7701 ARCSTAT_BUMP(arcstat_l2_feeds
);
7703 size
= l2arc_write_size();
7706 * Evict L2ARC buffers that will be overwritten.
7708 l2arc_evict(dev
, size
, B_FALSE
);
7711 * Write ARC buffers.
7713 wrote
= l2arc_write_buffers(spa
, dev
, size
);
7716 * Calculate interval between writes.
7718 next
= l2arc_write_interval(begin
, size
, wrote
);
7719 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7721 spl_fstrans_unmark(cookie
);
7723 l2arc_thread_exit
= 0;
7724 cv_broadcast(&l2arc_feed_thr_cv
);
7725 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
7730 l2arc_vdev_present(vdev_t
*vd
)
7734 mutex_enter(&l2arc_dev_mtx
);
7735 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
7736 dev
= list_next(l2arc_dev_list
, dev
)) {
7737 if (dev
->l2ad_vdev
== vd
)
7740 mutex_exit(&l2arc_dev_mtx
);
7742 return (dev
!= NULL
);
7746 * Add a vdev for use by the L2ARC. By this point the spa has already
7747 * validated the vdev and opened it.
7750 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
7752 l2arc_dev_t
*adddev
;
7754 ASSERT(!l2arc_vdev_present(vd
));
7757 * Create a new l2arc device entry.
7759 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
7760 adddev
->l2ad_spa
= spa
;
7761 adddev
->l2ad_vdev
= vd
;
7762 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
7763 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
7764 adddev
->l2ad_hand
= adddev
->l2ad_start
;
7765 adddev
->l2ad_first
= B_TRUE
;
7766 adddev
->l2ad_writing
= B_FALSE
;
7767 list_link_init(&adddev
->l2ad_node
);
7769 mutex_init(&adddev
->l2ad_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7771 * This is a list of all ARC buffers that are still valid on the
7774 list_create(&adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
7775 offsetof(arc_buf_hdr_t
, b_l2hdr
.b_l2node
));
7777 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
7778 zfs_refcount_create(&adddev
->l2ad_alloc
);
7781 * Add device to global list
7783 mutex_enter(&l2arc_dev_mtx
);
7784 list_insert_head(l2arc_dev_list
, adddev
);
7785 atomic_inc_64(&l2arc_ndev
);
7786 mutex_exit(&l2arc_dev_mtx
);
7790 * Remove a vdev from the L2ARC.
7793 l2arc_remove_vdev(vdev_t
*vd
)
7795 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
7798 * Find the device by vdev
7800 mutex_enter(&l2arc_dev_mtx
);
7801 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
7802 nextdev
= list_next(l2arc_dev_list
, dev
);
7803 if (vd
== dev
->l2ad_vdev
) {
7808 ASSERT3P(remdev
, !=, NULL
);
7811 * Remove device from global list
7813 list_remove(l2arc_dev_list
, remdev
);
7814 l2arc_dev_last
= NULL
; /* may have been invalidated */
7815 atomic_dec_64(&l2arc_ndev
);
7816 mutex_exit(&l2arc_dev_mtx
);
7819 * Clear all buflists and ARC references. L2ARC device flush.
7821 l2arc_evict(remdev
, 0, B_TRUE
);
7822 list_destroy(&remdev
->l2ad_buflist
);
7823 mutex_destroy(&remdev
->l2ad_mtx
);
7824 zfs_refcount_destroy(&remdev
->l2ad_alloc
);
7825 kmem_free(remdev
, sizeof (l2arc_dev_t
));
7831 l2arc_thread_exit
= 0;
7833 l2arc_writes_sent
= 0;
7834 l2arc_writes_done
= 0;
7836 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
7837 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
7838 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7839 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7841 l2arc_dev_list
= &L2ARC_dev_list
;
7842 l2arc_free_on_write
= &L2ARC_free_on_write
;
7843 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
7844 offsetof(l2arc_dev_t
, l2ad_node
));
7845 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
7846 offsetof(l2arc_data_free_t
, l2df_list_node
));
7853 * This is called from dmu_fini(), which is called from spa_fini();
7854 * Because of this, we can assume that all l2arc devices have
7855 * already been removed when the pools themselves were removed.
7858 l2arc_do_free_on_write();
7860 mutex_destroy(&l2arc_feed_thr_lock
);
7861 cv_destroy(&l2arc_feed_thr_cv
);
7862 mutex_destroy(&l2arc_dev_mtx
);
7863 mutex_destroy(&l2arc_free_on_write_mtx
);
7865 list_destroy(l2arc_dev_list
);
7866 list_destroy(l2arc_free_on_write
);
7872 if (!(spa_mode_global
& FWRITE
))
7875 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
7876 TS_RUN
, defclsyspri
);
7882 if (!(spa_mode_global
& FWRITE
))
7885 mutex_enter(&l2arc_feed_thr_lock
);
7886 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
7887 l2arc_thread_exit
= 1;
7888 while (l2arc_thread_exit
!= 0)
7889 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
7890 mutex_exit(&l2arc_feed_thr_lock
);
7893 #if defined(_KERNEL) && defined(HAVE_SPL)
7894 EXPORT_SYMBOL(arc_buf_size
);
7895 EXPORT_SYMBOL(arc_write
);
7896 EXPORT_SYMBOL(arc_read
);
7897 EXPORT_SYMBOL(arc_buf_info
);
7898 EXPORT_SYMBOL(arc_getbuf_func
);
7899 EXPORT_SYMBOL(arc_add_prune_callback
);
7900 EXPORT_SYMBOL(arc_remove_prune_callback
);
7903 module_param(zfs_arc_min
, ulong
, 0644);
7904 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
7906 module_param(zfs_arc_max
, ulong
, 0644);
7907 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
7909 module_param(zfs_arc_meta_limit
, ulong
, 0644);
7910 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
7912 module_param(zfs_arc_meta_limit_percent
, ulong
, 0644);
7913 MODULE_PARM_DESC(zfs_arc_meta_limit_percent
,
7914 "Percent of arc size for arc meta limit");
7916 module_param(zfs_arc_meta_min
, ulong
, 0644);
7917 MODULE_PARM_DESC(zfs_arc_meta_min
, "Min arc metadata");
7919 module_param(zfs_arc_meta_prune
, int, 0644);
7920 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Meta objects to scan for prune");
7922 module_param(zfs_arc_meta_adjust_restarts
, int, 0644);
7923 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts
,
7924 "Limit number of restarts in arc_adjust_meta");
7926 module_param(zfs_arc_meta_strategy
, int, 0644);
7927 MODULE_PARM_DESC(zfs_arc_meta_strategy
, "Meta reclaim strategy");
7929 module_param(zfs_arc_grow_retry
, int, 0644);
7930 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
7932 module_param(zfs_arc_p_dampener_disable
, int, 0644);
7933 MODULE_PARM_DESC(zfs_arc_p_dampener_disable
, "disable arc_p adapt dampener");
7935 module_param(zfs_arc_shrink_shift
, int, 0644);
7936 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
7938 module_param(zfs_arc_pc_percent
, uint
, 0644);
7939 MODULE_PARM_DESC(zfs_arc_pc_percent
,
7940 "Percent of pagecache to reclaim arc to");
7942 module_param(zfs_arc_p_min_shift
, int, 0644);
7943 MODULE_PARM_DESC(zfs_arc_p_min_shift
, "arc_c shift to calc min/max arc_p");
7945 module_param(zfs_arc_average_blocksize
, int, 0444);
7946 MODULE_PARM_DESC(zfs_arc_average_blocksize
, "Target average block size");
7948 module_param(zfs_compressed_arc_enabled
, int, 0644);
7949 MODULE_PARM_DESC(zfs_compressed_arc_enabled
, "Disable compressed arc buffers");
7951 module_param(zfs_arc_min_prefetch_lifespan
, int, 0644);
7952 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan
, "Min life of prefetch block");
7954 module_param(l2arc_write_max
, ulong
, 0644);
7955 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
7957 module_param(l2arc_write_boost
, ulong
, 0644);
7958 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
7960 module_param(l2arc_headroom
, ulong
, 0644);
7961 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
7963 module_param(l2arc_headroom_boost
, ulong
, 0644);
7964 MODULE_PARM_DESC(l2arc_headroom_boost
, "Compressed l2arc_headroom multiplier");
7966 module_param(l2arc_feed_secs
, ulong
, 0644);
7967 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
7969 module_param(l2arc_feed_min_ms
, ulong
, 0644);
7970 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
7972 module_param(l2arc_noprefetch
, int, 0644);
7973 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
7975 module_param(l2arc_feed_again
, int, 0644);
7976 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
7978 module_param(l2arc_norw
, int, 0644);
7979 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");
7981 module_param(zfs_arc_lotsfree_percent
, int, 0644);
7982 MODULE_PARM_DESC(zfs_arc_lotsfree_percent
,
7983 "System free memory I/O throttle in bytes");
7985 module_param(zfs_arc_sys_free
, ulong
, 0644);
7986 MODULE_PARM_DESC(zfs_arc_sys_free
, "System free memory target size in bytes");
7988 module_param(zfs_arc_dnode_limit
, ulong
, 0644);
7989 MODULE_PARM_DESC(zfs_arc_dnode_limit
, "Minimum bytes of dnodes in arc");
7991 module_param(zfs_arc_dnode_limit_percent
, ulong
, 0644);
7992 MODULE_PARM_DESC(zfs_arc_dnode_limit_percent
,
7993 "Percent of ARC meta buffers for dnodes");
7995 module_param(zfs_arc_dnode_reduce_percent
, ulong
, 0644);
7996 MODULE_PARM_DESC(zfs_arc_dnode_reduce_percent
,
7997 "Percentage of excess dnodes to try to unpin");