4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2018, Joyent, Inc.
24 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2017 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 * Note that the majority of the performance stats are manipulated
103 * with atomic operations.
105 * The L2ARC uses the l2ad_mtx on each vdev for the following:
107 * - L2ARC buflist creation
108 * - L2ARC buflist eviction
109 * - L2ARC write completion, which walks L2ARC buflists
110 * - ARC header destruction, as it removes from L2ARC buflists
111 * - ARC header release, as it removes from L2ARC buflists
117 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
118 * This structure can point either to a block that is still in the cache or to
119 * one that is only accessible in an L2 ARC device, or it can provide
120 * information about a block that was recently evicted. If a block is
121 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
122 * information to retrieve it from the L2ARC device. This information is
123 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
124 * that is in this state cannot access the data directly.
126 * Blocks that are actively being referenced or have not been evicted
127 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
128 * the arc_buf_hdr_t that will point to the data block in memory. A block can
129 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
130 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
131 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
133 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
134 * ability to store the physical data (b_pabd) associated with the DVA of the
135 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
136 * it will match its on-disk compression characteristics. This behavior can be
137 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
138 * compressed ARC functionality is disabled, the b_pabd will point to an
139 * uncompressed version of the on-disk data.
141 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
142 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
143 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
144 * consumer. The ARC will provide references to this data and will keep it
145 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
146 * data block and will evict any arc_buf_t that is no longer referenced. The
147 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
148 * "overhead_size" kstat.
150 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
151 * compressed form. The typical case is that consumers will want uncompressed
152 * data, and when that happens a new data buffer is allocated where the data is
153 * decompressed for them to use. Currently the only consumer who wants
154 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
155 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
156 * with the arc_buf_hdr_t.
158 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
159 * first one is owned by a compressed send consumer (and therefore references
160 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
161 * used by any other consumer (and has its own uncompressed copy of the data
176 * | b_buf +------------>+-----------+ arc_buf_t
177 * | b_pabd +-+ |b_next +---->+-----------+
178 * +-----------+ | |-----------| |b_next +-->NULL
179 * | |b_comp = T | +-----------+
180 * | |b_data +-+ |b_comp = F |
181 * | +-----------+ | |b_data +-+
182 * +->+------+ | +-----------+ |
184 * data | |<--------------+ | uncompressed
185 * +------+ compressed, | data
186 * shared +-->+------+
191 * When a consumer reads a block, the ARC must first look to see if the
192 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
193 * arc_buf_t and either copies uncompressed data into a new data buffer from an
194 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
195 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
196 * hdr is compressed and the desired compression characteristics of the
197 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
198 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
199 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
200 * be anywhere in the hdr's list.
202 * The diagram below shows an example of an uncompressed ARC hdr that is
203 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
204 * the last element in the buf list):
216 * | | arc_buf_t (shared)
217 * | b_buf +------------>+---------+ arc_buf_t
218 * | | |b_next +---->+---------+
219 * | b_pabd +-+ |---------| |b_next +-->NULL
220 * +-----------+ | | | +---------+
222 * | +---------+ | |b_data +-+
223 * +->+------+ | +---------+ |
225 * uncompressed | | | |
228 * | uncompressed | | |
231 * +---------------------------------+
233 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
234 * since the physical block is about to be rewritten. The new data contents
235 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
236 * it may compress the data before writing it to disk. The ARC will be called
237 * with the transformed data and will bcopy the transformed on-disk block into
238 * a newly allocated b_pabd. Writes are always done into buffers which have
239 * either been loaned (and hence are new and don't have other readers) or
240 * buffers which have been released (and hence have their own hdr, if there
241 * were originally other readers of the buf's original hdr). This ensures that
242 * the ARC only needs to update a single buf and its hdr after a write occurs.
244 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
245 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
246 * that when compressed ARC is enabled that the L2ARC blocks are identical
247 * to the on-disk block in the main data pool. This provides a significant
248 * advantage since the ARC can leverage the bp's checksum when reading from the
249 * L2ARC to determine if the contents are valid. However, if the compressed
250 * ARC is disabled, then the L2ARC's block must be transformed to look
251 * like the physical block in the main data pool before comparing the
252 * checksum and determining its validity.
257 #include <sys/spa_impl.h>
258 #include <sys/zio_compress.h>
259 #include <sys/zio_checksum.h>
260 #include <sys/zfs_context.h>
262 #include <sys/refcount.h>
263 #include <sys/vdev.h>
264 #include <sys/vdev_impl.h>
265 #include <sys/dsl_pool.h>
266 #include <sys/zio_checksum.h>
267 #include <sys/multilist.h>
270 #include <sys/vmsystm.h>
272 #include <sys/fs/swapnode.h>
273 #include <sys/dnlc.h>
275 #include <sys/callb.h>
276 #include <sys/kstat.h>
277 #include <sys/zthr.h>
278 #include <zfs_fletcher.h>
279 #include <sys/aggsum.h>
280 #include <sys/cityhash.h>
283 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
284 boolean_t arc_watch
= B_FALSE
;
289 * This thread's job is to keep enough free memory in the system, by
290 * calling arc_kmem_reap_now() plus arc_shrink(), which improves
291 * arc_available_memory().
293 static zthr_t
*arc_reap_zthr
;
296 * This thread's job is to keep arc_size under arc_c, by calling
297 * arc_adjust(), which improves arc_is_overflowing().
299 static zthr_t
*arc_adjust_zthr
;
301 static kmutex_t arc_adjust_lock
;
302 static kcondvar_t arc_adjust_waiters_cv
;
303 static boolean_t arc_adjust_needed
= B_FALSE
;
305 uint_t arc_reduce_dnlc_percent
= 3;
308 * The number of headers to evict in arc_evict_state_impl() before
309 * dropping the sublist lock and evicting from another sublist. A lower
310 * value means we're more likely to evict the "correct" header (i.e. the
311 * oldest header in the arc state), but comes with higher overhead
312 * (i.e. more invocations of arc_evict_state_impl()).
314 int zfs_arc_evict_batch_limit
= 10;
316 /* number of seconds before growing cache again */
317 int arc_grow_retry
= 60;
320 * Minimum time between calls to arc_kmem_reap_soon(). Note that this will
321 * be converted to ticks, so with the default hz=100, a setting of 15 ms
322 * will actually wait 2 ticks, or 20ms.
324 int arc_kmem_cache_reap_retry_ms
= 1000;
326 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
327 int zfs_arc_overflow_shift
= 8;
329 /* shift of arc_c for calculating both min and max arc_p */
330 int arc_p_min_shift
= 4;
332 /* log2(fraction of arc to reclaim) */
333 int arc_shrink_shift
= 7;
336 * log2(fraction of ARC which must be free to allow growing).
337 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
338 * when reading a new block into the ARC, we will evict an equal-sized block
341 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
342 * we will still not allow it to grow.
344 int arc_no_grow_shift
= 5;
348 * minimum lifespan of a prefetch block in clock ticks
349 * (initialized in arc_init())
351 static int arc_min_prefetch_lifespan
;
354 * If this percent of memory is free, don't throttle.
356 int arc_lotsfree_percent
= 10;
358 static boolean_t arc_initialized
;
361 * The arc has filled available memory and has now warmed up.
363 static boolean_t arc_warm
;
366 * log2 fraction of the zio arena to keep free.
368 int arc_zio_arena_free_shift
= 2;
371 * These tunables are for performance analysis.
373 uint64_t zfs_arc_max
;
374 uint64_t zfs_arc_min
;
375 uint64_t zfs_arc_meta_limit
= 0;
376 uint64_t zfs_arc_meta_min
= 0;
377 int zfs_arc_grow_retry
= 0;
378 int zfs_arc_shrink_shift
= 0;
379 int zfs_arc_p_min_shift
= 0;
380 int zfs_arc_average_blocksize
= 8 * 1024; /* 8KB */
383 * ARC dirty data constraints for arc_tempreserve_space() throttle
385 uint_t zfs_arc_dirty_limit_percent
= 50; /* total dirty data limit */
386 uint_t zfs_arc_anon_limit_percent
= 25; /* anon block dirty limit */
387 uint_t zfs_arc_pool_dirty_percent
= 20; /* each pool's anon allowance */
389 boolean_t zfs_compressed_arc_enabled
= B_TRUE
;
392 * Note that buffers can be in one of 6 states:
393 * ARC_anon - anonymous (discussed below)
394 * ARC_mru - recently used, currently cached
395 * ARC_mru_ghost - recentely used, no longer in cache
396 * ARC_mfu - frequently used, currently cached
397 * ARC_mfu_ghost - frequently used, no longer in cache
398 * ARC_l2c_only - exists in L2ARC but not other states
399 * When there are no active references to the buffer, they are
400 * are linked onto a list in one of these arc states. These are
401 * the only buffers that can be evicted or deleted. Within each
402 * state there are multiple lists, one for meta-data and one for
403 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
404 * etc.) is tracked separately so that it can be managed more
405 * explicitly: favored over data, limited explicitly.
407 * Anonymous buffers are buffers that are not associated with
408 * a DVA. These are buffers that hold dirty block copies
409 * before they are written to stable storage. By definition,
410 * they are "ref'd" and are considered part of arc_mru
411 * that cannot be freed. Generally, they will aquire a DVA
412 * as they are written and migrate onto the arc_mru list.
414 * The ARC_l2c_only state is for buffers that are in the second
415 * level ARC but no longer in any of the ARC_m* lists. The second
416 * level ARC itself may also contain buffers that are in any of
417 * the ARC_m* states - meaning that a buffer can exist in two
418 * places. The reason for the ARC_l2c_only state is to keep the
419 * buffer header in the hash table, so that reads that hit the
420 * second level ARC benefit from these fast lookups.
423 typedef struct arc_state
{
425 * list of evictable buffers
427 multilist_t
*arcs_list
[ARC_BUFC_NUMTYPES
];
429 * total amount of evictable data in this state
431 refcount_t arcs_esize
[ARC_BUFC_NUMTYPES
];
433 * total amount of data in this state; this includes: evictable,
434 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
436 refcount_t arcs_size
;
440 static arc_state_t ARC_anon
;
441 static arc_state_t ARC_mru
;
442 static arc_state_t ARC_mru_ghost
;
443 static arc_state_t ARC_mfu
;
444 static arc_state_t ARC_mfu_ghost
;
445 static arc_state_t ARC_l2c_only
;
447 typedef struct arc_stats
{
448 kstat_named_t arcstat_hits
;
449 kstat_named_t arcstat_misses
;
450 kstat_named_t arcstat_demand_data_hits
;
451 kstat_named_t arcstat_demand_data_misses
;
452 kstat_named_t arcstat_demand_metadata_hits
;
453 kstat_named_t arcstat_demand_metadata_misses
;
454 kstat_named_t arcstat_prefetch_data_hits
;
455 kstat_named_t arcstat_prefetch_data_misses
;
456 kstat_named_t arcstat_prefetch_metadata_hits
;
457 kstat_named_t arcstat_prefetch_metadata_misses
;
458 kstat_named_t arcstat_mru_hits
;
459 kstat_named_t arcstat_mru_ghost_hits
;
460 kstat_named_t arcstat_mfu_hits
;
461 kstat_named_t arcstat_mfu_ghost_hits
;
462 kstat_named_t arcstat_deleted
;
464 * Number of buffers that could not be evicted because the hash lock
465 * was held by another thread. The lock may not necessarily be held
466 * by something using the same buffer, since hash locks are shared
467 * by multiple buffers.
469 kstat_named_t arcstat_mutex_miss
;
471 * Number of buffers skipped because they have I/O in progress, are
472 * indrect prefetch buffers that have not lived long enough, or are
473 * not from the spa we're trying to evict from.
475 kstat_named_t arcstat_evict_skip
;
477 * Number of times arc_evict_state() was unable to evict enough
478 * buffers to reach it's target amount.
480 kstat_named_t arcstat_evict_not_enough
;
481 kstat_named_t arcstat_evict_l2_cached
;
482 kstat_named_t arcstat_evict_l2_eligible
;
483 kstat_named_t arcstat_evict_l2_ineligible
;
484 kstat_named_t arcstat_evict_l2_skip
;
485 kstat_named_t arcstat_hash_elements
;
486 kstat_named_t arcstat_hash_elements_max
;
487 kstat_named_t arcstat_hash_collisions
;
488 kstat_named_t arcstat_hash_chains
;
489 kstat_named_t arcstat_hash_chain_max
;
490 kstat_named_t arcstat_p
;
491 kstat_named_t arcstat_c
;
492 kstat_named_t arcstat_c_min
;
493 kstat_named_t arcstat_c_max
;
494 /* Not updated directly; only synced in arc_kstat_update. */
495 kstat_named_t arcstat_size
;
497 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
498 * Note that the compressed bytes may match the uncompressed bytes
499 * if the block is either not compressed or compressed arc is disabled.
501 kstat_named_t arcstat_compressed_size
;
503 * Uncompressed size of the data stored in b_pabd. If compressed
504 * arc is disabled then this value will be identical to the stat
507 kstat_named_t arcstat_uncompressed_size
;
509 * Number of bytes stored in all the arc_buf_t's. This is classified
510 * as "overhead" since this data is typically short-lived and will
511 * be evicted from the arc when it becomes unreferenced unless the
512 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
513 * values have been set (see comment in dbuf.c for more information).
515 kstat_named_t arcstat_overhead_size
;
517 * Number of bytes consumed by internal ARC structures necessary
518 * for tracking purposes; these structures are not actually
519 * backed by ARC buffers. This includes arc_buf_hdr_t structures
520 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
521 * caches), and arc_buf_t structures (allocated via arc_buf_t
523 * Not updated directly; only synced in arc_kstat_update.
525 kstat_named_t arcstat_hdr_size
;
527 * Number of bytes consumed by ARC buffers of type equal to
528 * ARC_BUFC_DATA. This is generally consumed by buffers backing
529 * on disk user data (e.g. plain file contents).
530 * Not updated directly; only synced in arc_kstat_update.
532 kstat_named_t arcstat_data_size
;
534 * Number of bytes consumed by ARC buffers of type equal to
535 * ARC_BUFC_METADATA. This is generally consumed by buffers
536 * backing on disk data that is used for internal ZFS
537 * structures (e.g. ZAP, dnode, indirect blocks, etc).
538 * Not updated directly; only synced in arc_kstat_update.
540 kstat_named_t arcstat_metadata_size
;
542 * Number of bytes consumed by various buffers and structures
543 * not actually backed with ARC buffers. This includes bonus
544 * buffers (allocated directly via zio_buf_* functions),
545 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
546 * cache), and dnode_t structures (allocated via dnode_t cache).
547 * Not updated directly; only synced in arc_kstat_update.
549 kstat_named_t arcstat_other_size
;
551 * Total number of bytes consumed by ARC buffers residing in the
552 * arc_anon state. This includes *all* buffers in the arc_anon
553 * state; e.g. data, metadata, evictable, and unevictable buffers
554 * are all included in this value.
555 * Not updated directly; only synced in arc_kstat_update.
557 kstat_named_t arcstat_anon_size
;
559 * Number of bytes consumed by ARC buffers that meet the
560 * following criteria: backing buffers of type ARC_BUFC_DATA,
561 * residing in the arc_anon state, and are eligible for eviction
562 * (e.g. have no outstanding holds on the buffer).
563 * Not updated directly; only synced in arc_kstat_update.
565 kstat_named_t arcstat_anon_evictable_data
;
567 * Number of bytes consumed by ARC buffers that meet the
568 * following criteria: backing buffers of type ARC_BUFC_METADATA,
569 * residing in the arc_anon state, and are eligible for eviction
570 * (e.g. have no outstanding holds on the buffer).
571 * Not updated directly; only synced in arc_kstat_update.
573 kstat_named_t arcstat_anon_evictable_metadata
;
575 * Total number of bytes consumed by ARC buffers residing in the
576 * arc_mru state. This includes *all* buffers in the arc_mru
577 * state; e.g. data, metadata, evictable, and unevictable buffers
578 * are all included in this value.
579 * Not updated directly; only synced in arc_kstat_update.
581 kstat_named_t arcstat_mru_size
;
583 * Number of bytes consumed by ARC buffers that meet the
584 * following criteria: backing buffers of type ARC_BUFC_DATA,
585 * residing in the arc_mru state, and are eligible for eviction
586 * (e.g. have no outstanding holds on the buffer).
587 * Not updated directly; only synced in arc_kstat_update.
589 kstat_named_t arcstat_mru_evictable_data
;
591 * Number of bytes consumed by ARC buffers that meet the
592 * following criteria: backing buffers of type ARC_BUFC_METADATA,
593 * residing in the arc_mru state, and are eligible for eviction
594 * (e.g. have no outstanding holds on the buffer).
595 * Not updated directly; only synced in arc_kstat_update.
597 kstat_named_t arcstat_mru_evictable_metadata
;
599 * Total number of bytes that *would have been* consumed by ARC
600 * buffers in the arc_mru_ghost state. The key thing to note
601 * here, is the fact that this size doesn't actually indicate
602 * RAM consumption. The ghost lists only consist of headers and
603 * don't actually have ARC buffers linked off of these headers.
604 * Thus, *if* the headers had associated ARC buffers, these
605 * buffers *would have* consumed this number of bytes.
606 * Not updated directly; only synced in arc_kstat_update.
608 kstat_named_t arcstat_mru_ghost_size
;
610 * Number of bytes that *would have been* consumed by ARC
611 * buffers that are eligible for eviction, of type
612 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
613 * Not updated directly; only synced in arc_kstat_update.
615 kstat_named_t arcstat_mru_ghost_evictable_data
;
617 * Number of bytes that *would have been* consumed by ARC
618 * buffers that are eligible for eviction, of type
619 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
620 * Not updated directly; only synced in arc_kstat_update.
622 kstat_named_t arcstat_mru_ghost_evictable_metadata
;
624 * Total number of bytes consumed by ARC buffers residing in the
625 * arc_mfu state. This includes *all* buffers in the arc_mfu
626 * state; e.g. data, metadata, evictable, and unevictable buffers
627 * are all included in this value.
628 * Not updated directly; only synced in arc_kstat_update.
630 kstat_named_t arcstat_mfu_size
;
632 * Number of bytes consumed by ARC buffers that are eligible for
633 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
635 * Not updated directly; only synced in arc_kstat_update.
637 kstat_named_t arcstat_mfu_evictable_data
;
639 * Number of bytes consumed by ARC buffers that are eligible for
640 * eviction, of type ARC_BUFC_METADATA, and reside in the
642 * Not updated directly; only synced in arc_kstat_update.
644 kstat_named_t arcstat_mfu_evictable_metadata
;
646 * Total number of bytes that *would have been* consumed by ARC
647 * buffers in the arc_mfu_ghost state. See the comment above
648 * arcstat_mru_ghost_size for more details.
649 * Not updated directly; only synced in arc_kstat_update.
651 kstat_named_t arcstat_mfu_ghost_size
;
653 * Number of bytes that *would have been* consumed by ARC
654 * buffers that are eligible for eviction, of type
655 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
656 * Not updated directly; only synced in arc_kstat_update.
658 kstat_named_t arcstat_mfu_ghost_evictable_data
;
660 * Number of bytes that *would have been* consumed by ARC
661 * buffers that are eligible for eviction, of type
662 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
663 * Not updated directly; only synced in arc_kstat_update.
665 kstat_named_t arcstat_mfu_ghost_evictable_metadata
;
666 kstat_named_t arcstat_l2_hits
;
667 kstat_named_t arcstat_l2_misses
;
668 kstat_named_t arcstat_l2_feeds
;
669 kstat_named_t arcstat_l2_rw_clash
;
670 kstat_named_t arcstat_l2_read_bytes
;
671 kstat_named_t arcstat_l2_write_bytes
;
672 kstat_named_t arcstat_l2_writes_sent
;
673 kstat_named_t arcstat_l2_writes_done
;
674 kstat_named_t arcstat_l2_writes_error
;
675 kstat_named_t arcstat_l2_writes_lock_retry
;
676 kstat_named_t arcstat_l2_evict_lock_retry
;
677 kstat_named_t arcstat_l2_evict_reading
;
678 kstat_named_t arcstat_l2_evict_l1cached
;
679 kstat_named_t arcstat_l2_free_on_write
;
680 kstat_named_t arcstat_l2_abort_lowmem
;
681 kstat_named_t arcstat_l2_cksum_bad
;
682 kstat_named_t arcstat_l2_io_error
;
683 kstat_named_t arcstat_l2_lsize
;
684 kstat_named_t arcstat_l2_psize
;
685 /* Not updated directly; only synced in arc_kstat_update. */
686 kstat_named_t arcstat_l2_hdr_size
;
687 kstat_named_t arcstat_memory_throttle_count
;
688 /* Not updated directly; only synced in arc_kstat_update. */
689 kstat_named_t arcstat_meta_used
;
690 kstat_named_t arcstat_meta_limit
;
691 kstat_named_t arcstat_meta_max
;
692 kstat_named_t arcstat_meta_min
;
693 kstat_named_t arcstat_sync_wait_for_async
;
694 kstat_named_t arcstat_demand_hit_predictive_prefetch
;
697 static arc_stats_t arc_stats
= {
698 { "hits", KSTAT_DATA_UINT64
},
699 { "misses", KSTAT_DATA_UINT64
},
700 { "demand_data_hits", KSTAT_DATA_UINT64
},
701 { "demand_data_misses", KSTAT_DATA_UINT64
},
702 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
703 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
704 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
705 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
706 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
707 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
708 { "mru_hits", KSTAT_DATA_UINT64
},
709 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
710 { "mfu_hits", KSTAT_DATA_UINT64
},
711 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
712 { "deleted", KSTAT_DATA_UINT64
},
713 { "mutex_miss", KSTAT_DATA_UINT64
},
714 { "evict_skip", KSTAT_DATA_UINT64
},
715 { "evict_not_enough", KSTAT_DATA_UINT64
},
716 { "evict_l2_cached", KSTAT_DATA_UINT64
},
717 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
718 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
719 { "evict_l2_skip", KSTAT_DATA_UINT64
},
720 { "hash_elements", KSTAT_DATA_UINT64
},
721 { "hash_elements_max", KSTAT_DATA_UINT64
},
722 { "hash_collisions", KSTAT_DATA_UINT64
},
723 { "hash_chains", KSTAT_DATA_UINT64
},
724 { "hash_chain_max", KSTAT_DATA_UINT64
},
725 { "p", KSTAT_DATA_UINT64
},
726 { "c", KSTAT_DATA_UINT64
},
727 { "c_min", KSTAT_DATA_UINT64
},
728 { "c_max", KSTAT_DATA_UINT64
},
729 { "size", KSTAT_DATA_UINT64
},
730 { "compressed_size", KSTAT_DATA_UINT64
},
731 { "uncompressed_size", KSTAT_DATA_UINT64
},
732 { "overhead_size", KSTAT_DATA_UINT64
},
733 { "hdr_size", KSTAT_DATA_UINT64
},
734 { "data_size", KSTAT_DATA_UINT64
},
735 { "metadata_size", KSTAT_DATA_UINT64
},
736 { "other_size", KSTAT_DATA_UINT64
},
737 { "anon_size", KSTAT_DATA_UINT64
},
738 { "anon_evictable_data", KSTAT_DATA_UINT64
},
739 { "anon_evictable_metadata", KSTAT_DATA_UINT64
},
740 { "mru_size", KSTAT_DATA_UINT64
},
741 { "mru_evictable_data", KSTAT_DATA_UINT64
},
742 { "mru_evictable_metadata", KSTAT_DATA_UINT64
},
743 { "mru_ghost_size", KSTAT_DATA_UINT64
},
744 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64
},
745 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
746 { "mfu_size", KSTAT_DATA_UINT64
},
747 { "mfu_evictable_data", KSTAT_DATA_UINT64
},
748 { "mfu_evictable_metadata", KSTAT_DATA_UINT64
},
749 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
750 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64
},
751 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
752 { "l2_hits", KSTAT_DATA_UINT64
},
753 { "l2_misses", KSTAT_DATA_UINT64
},
754 { "l2_feeds", KSTAT_DATA_UINT64
},
755 { "l2_rw_clash", KSTAT_DATA_UINT64
},
756 { "l2_read_bytes", KSTAT_DATA_UINT64
},
757 { "l2_write_bytes", KSTAT_DATA_UINT64
},
758 { "l2_writes_sent", KSTAT_DATA_UINT64
},
759 { "l2_writes_done", KSTAT_DATA_UINT64
},
760 { "l2_writes_error", KSTAT_DATA_UINT64
},
761 { "l2_writes_lock_retry", KSTAT_DATA_UINT64
},
762 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
763 { "l2_evict_reading", KSTAT_DATA_UINT64
},
764 { "l2_evict_l1cached", KSTAT_DATA_UINT64
},
765 { "l2_free_on_write", KSTAT_DATA_UINT64
},
766 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
767 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
768 { "l2_io_error", KSTAT_DATA_UINT64
},
769 { "l2_size", KSTAT_DATA_UINT64
},
770 { "l2_asize", KSTAT_DATA_UINT64
},
771 { "l2_hdr_size", KSTAT_DATA_UINT64
},
772 { "memory_throttle_count", KSTAT_DATA_UINT64
},
773 { "arc_meta_used", KSTAT_DATA_UINT64
},
774 { "arc_meta_limit", KSTAT_DATA_UINT64
},
775 { "arc_meta_max", KSTAT_DATA_UINT64
},
776 { "arc_meta_min", KSTAT_DATA_UINT64
},
777 { "sync_wait_for_async", KSTAT_DATA_UINT64
},
778 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64
},
781 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
783 #define ARCSTAT_INCR(stat, val) \
784 atomic_add_64(&arc_stats.stat.value.ui64, (val))
786 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
787 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
789 #define ARCSTAT_MAX(stat, val) { \
791 while ((val) > (m = arc_stats.stat.value.ui64) && \
792 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
796 #define ARCSTAT_MAXSTAT(stat) \
797 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
800 * We define a macro to allow ARC hits/misses to be easily broken down by
801 * two separate conditions, giving a total of four different subtypes for
802 * each of hits and misses (so eight statistics total).
804 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
807 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
809 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
813 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
815 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
820 static arc_state_t
*arc_anon
;
821 static arc_state_t
*arc_mru
;
822 static arc_state_t
*arc_mru_ghost
;
823 static arc_state_t
*arc_mfu
;
824 static arc_state_t
*arc_mfu_ghost
;
825 static arc_state_t
*arc_l2c_only
;
828 * There are several ARC variables that are critical to export as kstats --
829 * but we don't want to have to grovel around in the kstat whenever we wish to
830 * manipulate them. For these variables, we therefore define them to be in
831 * terms of the statistic variable. This assures that we are not introducing
832 * the possibility of inconsistency by having shadow copies of the variables,
833 * while still allowing the code to be readable.
835 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
836 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
837 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
838 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
839 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
840 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
841 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
843 /* compressed size of entire arc */
844 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
845 /* uncompressed size of entire arc */
846 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
847 /* number of bytes in the arc from arc_buf_t's */
848 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
851 * There are also some ARC variables that we want to export, but that are
852 * updated so often that having the canonical representation be the statistic
853 * variable causes a performance bottleneck. We want to use aggsum_t's for these
854 * instead, but still be able to export the kstat in the same way as before.
855 * The solution is to always use the aggsum version, except in the kstat update
859 aggsum_t arc_meta_used
;
860 aggsum_t astat_data_size
;
861 aggsum_t astat_metadata_size
;
862 aggsum_t astat_hdr_size
;
863 aggsum_t astat_other_size
;
864 aggsum_t astat_l2_hdr_size
;
866 static int arc_no_grow
; /* Don't try to grow cache size */
867 static hrtime_t arc_growtime
;
868 static uint64_t arc_tempreserve
;
869 static uint64_t arc_loaned_bytes
;
871 typedef struct arc_callback arc_callback_t
;
873 struct arc_callback
{
875 arc_done_func_t
*acb_done
;
877 boolean_t acb_compressed
;
878 zio_t
*acb_zio_dummy
;
879 arc_callback_t
*acb_next
;
882 typedef struct arc_write_callback arc_write_callback_t
;
884 struct arc_write_callback
{
886 arc_done_func_t
*awcb_ready
;
887 arc_done_func_t
*awcb_children_ready
;
888 arc_done_func_t
*awcb_physdone
;
889 arc_done_func_t
*awcb_done
;
894 * ARC buffers are separated into multiple structs as a memory saving measure:
895 * - Common fields struct, always defined, and embedded within it:
896 * - L2-only fields, always allocated but undefined when not in L2ARC
897 * - L1-only fields, only allocated when in L1ARC
899 * Buffer in L1 Buffer only in L2
900 * +------------------------+ +------------------------+
901 * | arc_buf_hdr_t | | arc_buf_hdr_t |
905 * +------------------------+ +------------------------+
906 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
907 * | (undefined if L1-only) | | |
908 * +------------------------+ +------------------------+
909 * | l1arc_buf_hdr_t |
914 * +------------------------+
916 * Because it's possible for the L2ARC to become extremely large, we can wind
917 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
918 * is minimized by only allocating the fields necessary for an L1-cached buffer
919 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
920 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
921 * words in pointers. arc_hdr_realloc() is used to switch a header between
922 * these two allocation states.
924 typedef struct l1arc_buf_hdr
{
925 kmutex_t b_freeze_lock
;
926 zio_cksum_t
*b_freeze_cksum
;
929 * Used for debugging with kmem_flags - by allocating and freeing
930 * b_thawed when the buffer is thawed, we get a record of the stack
931 * trace that thawed it.
938 /* for waiting on writes to complete */
942 /* protected by arc state mutex */
943 arc_state_t
*b_state
;
944 multilist_node_t b_arc_node
;
946 /* updated atomically */
947 clock_t b_arc_access
;
949 /* self protecting */
952 arc_callback_t
*b_acb
;
956 typedef struct l2arc_dev l2arc_dev_t
;
958 typedef struct l2arc_buf_hdr
{
959 /* protected by arc_buf_hdr mutex */
960 l2arc_dev_t
*b_dev
; /* L2ARC device */
961 uint64_t b_daddr
; /* disk address, offset byte */
963 list_node_t b_l2node
;
967 /* protected by hash lock */
971 arc_buf_contents_t b_type
;
972 arc_buf_hdr_t
*b_hash_next
;
976 * This field stores the size of the data buffer after
977 * compression, and is set in the arc's zio completion handlers.
978 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
980 * While the block pointers can store up to 32MB in their psize
981 * field, we can only store up to 32MB minus 512B. This is due
982 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
983 * a field of zeros represents 512B in the bp). We can't use a
984 * bias of 1 since we need to reserve a psize of zero, here, to
985 * represent holes and embedded blocks.
987 * This isn't a problem in practice, since the maximum size of a
988 * buffer is limited to 16MB, so we never need to store 32MB in
989 * this field. Even in the upstream illumos code base, the
990 * maximum size of a buffer is limited to 16MB.
995 * This field stores the size of the data buffer before
996 * compression, and cannot change once set. It is in units
997 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
999 uint16_t b_lsize
; /* immutable */
1000 uint64_t b_spa
; /* immutable */
1002 /* L2ARC fields. Undefined when not in L2ARC. */
1003 l2arc_buf_hdr_t b_l2hdr
;
1004 /* L1ARC fields. Undefined when in l2arc_only state */
1005 l1arc_buf_hdr_t b_l1hdr
;
1008 #define GHOST_STATE(state) \
1009 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1010 (state) == arc_l2c_only)
1012 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1013 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1014 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1015 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1016 #define HDR_COMPRESSION_ENABLED(hdr) \
1017 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1019 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1020 #define HDR_L2_READING(hdr) \
1021 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1022 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1023 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1024 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1025 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1026 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1028 #define HDR_ISTYPE_METADATA(hdr) \
1029 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1030 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1032 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1033 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1035 /* For storing compression mode in b_flags */
1036 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1038 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1039 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1040 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1041 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1043 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1044 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1045 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1051 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1052 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1055 * Hash table routines
1058 #define HT_LOCK_PAD 64
1063 unsigned char pad
[(HT_LOCK_PAD
- sizeof (kmutex_t
))];
1067 #define BUF_LOCKS 256
1068 typedef struct buf_hash_table
{
1070 arc_buf_hdr_t
**ht_table
;
1071 struct ht_lock ht_locks
[BUF_LOCKS
];
1074 static buf_hash_table_t buf_hash_table
;
1076 #define BUF_HASH_INDEX(spa, dva, birth) \
1077 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1078 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1079 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1080 #define HDR_LOCK(hdr) \
1081 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1083 uint64_t zfs_crc64_table
[256];
1089 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1090 #define L2ARC_HEADROOM 2 /* num of writes */
1092 * If we discover during ARC scan any buffers to be compressed, we boost
1093 * our headroom for the next scanning cycle by this percentage multiple.
1095 #define L2ARC_HEADROOM_BOOST 200
1096 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1097 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1099 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1100 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1102 /* L2ARC Performance Tunables */
1103 uint64_t l2arc_write_max
= L2ARC_WRITE_SIZE
; /* default max write size */
1104 uint64_t l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra write during warmup */
1105 uint64_t l2arc_headroom
= L2ARC_HEADROOM
; /* number of dev writes */
1106 uint64_t l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
1107 uint64_t l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
1108 uint64_t l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval milliseconds */
1109 boolean_t l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
1110 boolean_t l2arc_feed_again
= B_TRUE
; /* turbo warmup */
1111 boolean_t l2arc_norw
= B_TRUE
; /* no reads during writes */
1117 vdev_t
*l2ad_vdev
; /* vdev */
1118 spa_t
*l2ad_spa
; /* spa */
1119 uint64_t l2ad_hand
; /* next write location */
1120 uint64_t l2ad_start
; /* first addr on device */
1121 uint64_t l2ad_end
; /* last addr on device */
1122 boolean_t l2ad_first
; /* first sweep through */
1123 boolean_t l2ad_writing
; /* currently writing */
1124 kmutex_t l2ad_mtx
; /* lock for buffer list */
1125 list_t l2ad_buflist
; /* buffer list */
1126 list_node_t l2ad_node
; /* device list node */
1127 refcount_t l2ad_alloc
; /* allocated bytes */
1130 static list_t L2ARC_dev_list
; /* device list */
1131 static list_t
*l2arc_dev_list
; /* device list pointer */
1132 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
1133 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
1134 static list_t L2ARC_free_on_write
; /* free after write buf list */
1135 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
1136 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
1137 static uint64_t l2arc_ndev
; /* number of devices */
1139 typedef struct l2arc_read_callback
{
1140 arc_buf_hdr_t
*l2rcb_hdr
; /* read header */
1141 blkptr_t l2rcb_bp
; /* original blkptr */
1142 zbookmark_phys_t l2rcb_zb
; /* original bookmark */
1143 int l2rcb_flags
; /* original flags */
1144 abd_t
*l2rcb_abd
; /* temporary buffer */
1145 } l2arc_read_callback_t
;
1147 typedef struct l2arc_write_callback
{
1148 l2arc_dev_t
*l2wcb_dev
; /* device info */
1149 arc_buf_hdr_t
*l2wcb_head
; /* head of write buflist */
1150 } l2arc_write_callback_t
;
1152 typedef struct l2arc_data_free
{
1153 /* protected by l2arc_free_on_write_mtx */
1156 arc_buf_contents_t l2df_type
;
1157 list_node_t l2df_list_node
;
1158 } l2arc_data_free_t
;
1160 static kmutex_t l2arc_feed_thr_lock
;
1161 static kcondvar_t l2arc_feed_thr_cv
;
1162 static uint8_t l2arc_thread_exit
;
1164 static abd_t
*arc_get_data_abd(arc_buf_hdr_t
*, uint64_t, void *);
1165 static void *arc_get_data_buf(arc_buf_hdr_t
*, uint64_t, void *);
1166 static void arc_get_data_impl(arc_buf_hdr_t
*, uint64_t, void *);
1167 static void arc_free_data_abd(arc_buf_hdr_t
*, abd_t
*, uint64_t, void *);
1168 static void arc_free_data_buf(arc_buf_hdr_t
*, void *, uint64_t, void *);
1169 static void arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
);
1170 static void arc_hdr_free_pabd(arc_buf_hdr_t
*);
1171 static void arc_hdr_alloc_pabd(arc_buf_hdr_t
*);
1172 static void arc_access(arc_buf_hdr_t
*, kmutex_t
*);
1173 static boolean_t
arc_is_overflowing();
1174 static void arc_buf_watch(arc_buf_t
*);
1176 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t
*);
1177 static uint32_t arc_bufc_to_flags(arc_buf_contents_t
);
1178 static inline void arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
1179 static inline void arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
);
1181 static boolean_t
l2arc_write_eligible(uint64_t, arc_buf_hdr_t
*);
1182 static void l2arc_read_done(zio_t
*);
1186 * We use Cityhash for this. It's fast, and has good hash properties without
1187 * requiring any large static buffers.
1190 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
1192 return (cityhash4(spa
, dva
->dva_word
[0], dva
->dva_word
[1], birth
));
1195 #define HDR_EMPTY(hdr) \
1196 ((hdr)->b_dva.dva_word[0] == 0 && \
1197 (hdr)->b_dva.dva_word[1] == 0)
1199 #define HDR_EQUAL(spa, dva, birth, hdr) \
1200 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1201 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1202 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1205 buf_discard_identity(arc_buf_hdr_t
*hdr
)
1207 hdr
->b_dva
.dva_word
[0] = 0;
1208 hdr
->b_dva
.dva_word
[1] = 0;
1212 static arc_buf_hdr_t
*
1213 buf_hash_find(uint64_t spa
, const blkptr_t
*bp
, kmutex_t
**lockp
)
1215 const dva_t
*dva
= BP_IDENTITY(bp
);
1216 uint64_t birth
= BP_PHYSICAL_BIRTH(bp
);
1217 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
1218 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1221 mutex_enter(hash_lock
);
1222 for (hdr
= buf_hash_table
.ht_table
[idx
]; hdr
!= NULL
;
1223 hdr
= hdr
->b_hash_next
) {
1224 if (HDR_EQUAL(spa
, dva
, birth
, hdr
)) {
1229 mutex_exit(hash_lock
);
1235 * Insert an entry into the hash table. If there is already an element
1236 * equal to elem in the hash table, then the already existing element
1237 * will be returned and the new element will not be inserted.
1238 * Otherwise returns NULL.
1239 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1241 static arc_buf_hdr_t
*
1242 buf_hash_insert(arc_buf_hdr_t
*hdr
, kmutex_t
**lockp
)
1244 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1245 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
1246 arc_buf_hdr_t
*fhdr
;
1249 ASSERT(!DVA_IS_EMPTY(&hdr
->b_dva
));
1250 ASSERT(hdr
->b_birth
!= 0);
1251 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
1253 if (lockp
!= NULL
) {
1255 mutex_enter(hash_lock
);
1257 ASSERT(MUTEX_HELD(hash_lock
));
1260 for (fhdr
= buf_hash_table
.ht_table
[idx
], i
= 0; fhdr
!= NULL
;
1261 fhdr
= fhdr
->b_hash_next
, i
++) {
1262 if (HDR_EQUAL(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
, fhdr
))
1266 hdr
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
1267 buf_hash_table
.ht_table
[idx
] = hdr
;
1268 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1270 /* collect some hash table performance data */
1272 ARCSTAT_BUMP(arcstat_hash_collisions
);
1274 ARCSTAT_BUMP(arcstat_hash_chains
);
1276 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
1279 ARCSTAT_BUMP(arcstat_hash_elements
);
1280 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
1286 buf_hash_remove(arc_buf_hdr_t
*hdr
)
1288 arc_buf_hdr_t
*fhdr
, **hdrp
;
1289 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
1291 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
1292 ASSERT(HDR_IN_HASH_TABLE(hdr
));
1294 hdrp
= &buf_hash_table
.ht_table
[idx
];
1295 while ((fhdr
= *hdrp
) != hdr
) {
1296 ASSERT3P(fhdr
, !=, NULL
);
1297 hdrp
= &fhdr
->b_hash_next
;
1299 *hdrp
= hdr
->b_hash_next
;
1300 hdr
->b_hash_next
= NULL
;
1301 arc_hdr_clear_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
1303 /* collect some hash table performance data */
1304 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
1306 if (buf_hash_table
.ht_table
[idx
] &&
1307 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
1308 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
1312 * Global data structures and functions for the buf kmem cache.
1314 static kmem_cache_t
*hdr_full_cache
;
1315 static kmem_cache_t
*hdr_l2only_cache
;
1316 static kmem_cache_t
*buf_cache
;
1323 kmem_free(buf_hash_table
.ht_table
,
1324 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
1325 for (i
= 0; i
< BUF_LOCKS
; i
++)
1326 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
1327 kmem_cache_destroy(hdr_full_cache
);
1328 kmem_cache_destroy(hdr_l2only_cache
);
1329 kmem_cache_destroy(buf_cache
);
1333 * Constructor callback - called when the cache is empty
1334 * and a new buf is requested.
1338 hdr_full_cons(void *vbuf
, void *unused
, int kmflag
)
1340 arc_buf_hdr_t
*hdr
= vbuf
;
1342 bzero(hdr
, HDR_FULL_SIZE
);
1343 cv_init(&hdr
->b_l1hdr
.b_cv
, NULL
, CV_DEFAULT
, NULL
);
1344 refcount_create(&hdr
->b_l1hdr
.b_refcnt
);
1345 mutex_init(&hdr
->b_l1hdr
.b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1346 multilist_link_init(&hdr
->b_l1hdr
.b_arc_node
);
1347 arc_space_consume(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1354 hdr_l2only_cons(void *vbuf
, void *unused
, int kmflag
)
1356 arc_buf_hdr_t
*hdr
= vbuf
;
1358 bzero(hdr
, HDR_L2ONLY_SIZE
);
1359 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1366 buf_cons(void *vbuf
, void *unused
, int kmflag
)
1368 arc_buf_t
*buf
= vbuf
;
1370 bzero(buf
, sizeof (arc_buf_t
));
1371 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1372 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1378 * Destructor callback - called when a cached buf is
1379 * no longer required.
1383 hdr_full_dest(void *vbuf
, void *unused
)
1385 arc_buf_hdr_t
*hdr
= vbuf
;
1387 ASSERT(HDR_EMPTY(hdr
));
1388 cv_destroy(&hdr
->b_l1hdr
.b_cv
);
1389 refcount_destroy(&hdr
->b_l1hdr
.b_refcnt
);
1390 mutex_destroy(&hdr
->b_l1hdr
.b_freeze_lock
);
1391 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1392 arc_space_return(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1397 hdr_l2only_dest(void *vbuf
, void *unused
)
1399 arc_buf_hdr_t
*hdr
= vbuf
;
1401 ASSERT(HDR_EMPTY(hdr
));
1402 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1407 buf_dest(void *vbuf
, void *unused
)
1409 arc_buf_t
*buf
= vbuf
;
1411 mutex_destroy(&buf
->b_evict_lock
);
1412 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1416 * Reclaim callback -- invoked when memory is low.
1420 hdr_recl(void *unused
)
1422 dprintf("hdr_recl called\n");
1424 * umem calls the reclaim func when we destroy the buf cache,
1425 * which is after we do arc_fini().
1427 if (arc_initialized
)
1428 zthr_wakeup(arc_reap_zthr
);
1435 uint64_t hsize
= 1ULL << 12;
1439 * The hash table is big enough to fill all of physical memory
1440 * with an average block size of zfs_arc_average_blocksize (default 8K).
1441 * By default, the table will take up
1442 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1444 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
1447 buf_hash_table
.ht_mask
= hsize
- 1;
1448 buf_hash_table
.ht_table
=
1449 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
1450 if (buf_hash_table
.ht_table
== NULL
) {
1451 ASSERT(hsize
> (1ULL << 8));
1456 hdr_full_cache
= kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE
,
1457 0, hdr_full_cons
, hdr_full_dest
, hdr_recl
, NULL
, NULL
, 0);
1458 hdr_l2only_cache
= kmem_cache_create("arc_buf_hdr_t_l2only",
1459 HDR_L2ONLY_SIZE
, 0, hdr_l2only_cons
, hdr_l2only_dest
, hdr_recl
,
1461 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
1462 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
1464 for (i
= 0; i
< 256; i
++)
1465 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
1466 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
1468 for (i
= 0; i
< BUF_LOCKS
; i
++) {
1469 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
1470 NULL
, MUTEX_DEFAULT
, NULL
);
1475 * This is the size that the buf occupies in memory. If the buf is compressed,
1476 * it will correspond to the compressed size. You should use this method of
1477 * getting the buf size unless you explicitly need the logical size.
1480 arc_buf_size(arc_buf_t
*buf
)
1482 return (ARC_BUF_COMPRESSED(buf
) ?
1483 HDR_GET_PSIZE(buf
->b_hdr
) : HDR_GET_LSIZE(buf
->b_hdr
));
1487 arc_buf_lsize(arc_buf_t
*buf
)
1489 return (HDR_GET_LSIZE(buf
->b_hdr
));
1493 arc_get_compression(arc_buf_t
*buf
)
1495 return (ARC_BUF_COMPRESSED(buf
) ?
1496 HDR_GET_COMPRESS(buf
->b_hdr
) : ZIO_COMPRESS_OFF
);
1499 #define ARC_MINTIME (hz>>4) /* 62 ms */
1501 static inline boolean_t
1502 arc_buf_is_shared(arc_buf_t
*buf
)
1504 boolean_t shared
= (buf
->b_data
!= NULL
&&
1505 buf
->b_hdr
->b_l1hdr
.b_pabd
!= NULL
&&
1506 abd_is_linear(buf
->b_hdr
->b_l1hdr
.b_pabd
) &&
1507 buf
->b_data
== abd_to_buf(buf
->b_hdr
->b_l1hdr
.b_pabd
));
1508 IMPLY(shared
, HDR_SHARED_DATA(buf
->b_hdr
));
1509 IMPLY(shared
, ARC_BUF_SHARED(buf
));
1510 IMPLY(shared
, ARC_BUF_COMPRESSED(buf
) || ARC_BUF_LAST(buf
));
1513 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1514 * already being shared" requirement prevents us from doing that.
1521 * Free the checksum associated with this header. If there is no checksum, this
1525 arc_cksum_free(arc_buf_hdr_t
*hdr
)
1527 ASSERT(HDR_HAS_L1HDR(hdr
));
1528 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1529 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1530 kmem_free(hdr
->b_l1hdr
.b_freeze_cksum
, sizeof (zio_cksum_t
));
1531 hdr
->b_l1hdr
.b_freeze_cksum
= NULL
;
1533 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1537 * Return true iff at least one of the bufs on hdr is not compressed.
1540 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t
*hdr
)
1542 for (arc_buf_t
*b
= hdr
->b_l1hdr
.b_buf
; b
!= NULL
; b
= b
->b_next
) {
1543 if (!ARC_BUF_COMPRESSED(b
)) {
1551 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1552 * matches the checksum that is stored in the hdr. If there is no checksum,
1553 * or if the buf is compressed, this is a no-op.
1556 arc_cksum_verify(arc_buf_t
*buf
)
1558 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1561 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1564 if (ARC_BUF_COMPRESSED(buf
)) {
1565 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1566 arc_hdr_has_uncompressed_buf(hdr
));
1570 ASSERT(HDR_HAS_L1HDR(hdr
));
1572 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1573 if (hdr
->b_l1hdr
.b_freeze_cksum
== NULL
|| HDR_IO_ERROR(hdr
)) {
1574 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1578 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
, &zc
);
1579 if (!ZIO_CHECKSUM_EQUAL(*hdr
->b_l1hdr
.b_freeze_cksum
, zc
))
1580 panic("buffer modified while frozen!");
1581 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1585 arc_cksum_is_equal(arc_buf_hdr_t
*hdr
, zio_t
*zio
)
1587 enum zio_compress compress
= BP_GET_COMPRESS(zio
->io_bp
);
1588 boolean_t valid_cksum
;
1590 ASSERT(!BP_IS_EMBEDDED(zio
->io_bp
));
1591 VERIFY3U(BP_GET_PSIZE(zio
->io_bp
), ==, HDR_GET_PSIZE(hdr
));
1594 * We rely on the blkptr's checksum to determine if the block
1595 * is valid or not. When compressed arc is enabled, the l2arc
1596 * writes the block to the l2arc just as it appears in the pool.
1597 * This allows us to use the blkptr's checksum to validate the
1598 * data that we just read off of the l2arc without having to store
1599 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1600 * arc is disabled, then the data written to the l2arc is always
1601 * uncompressed and won't match the block as it exists in the main
1602 * pool. When this is the case, we must first compress it if it is
1603 * compressed on the main pool before we can validate the checksum.
1605 if (!HDR_COMPRESSION_ENABLED(hdr
) && compress
!= ZIO_COMPRESS_OFF
) {
1606 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1607 uint64_t lsize
= HDR_GET_LSIZE(hdr
);
1610 abd_t
*cdata
= abd_alloc_linear(HDR_GET_PSIZE(hdr
), B_TRUE
);
1611 csize
= zio_compress_data(compress
, zio
->io_abd
,
1612 abd_to_buf(cdata
), lsize
);
1614 ASSERT3U(csize
, <=, HDR_GET_PSIZE(hdr
));
1615 if (csize
< HDR_GET_PSIZE(hdr
)) {
1617 * Compressed blocks are always a multiple of the
1618 * smallest ashift in the pool. Ideally, we would
1619 * like to round up the csize to the next
1620 * spa_min_ashift but that value may have changed
1621 * since the block was last written. Instead,
1622 * we rely on the fact that the hdr's psize
1623 * was set to the psize of the block when it was
1624 * last written. We set the csize to that value
1625 * and zero out any part that should not contain
1628 abd_zero_off(cdata
, csize
, HDR_GET_PSIZE(hdr
) - csize
);
1629 csize
= HDR_GET_PSIZE(hdr
);
1631 zio_push_transform(zio
, cdata
, csize
, HDR_GET_PSIZE(hdr
), NULL
);
1635 * Block pointers always store the checksum for the logical data.
1636 * If the block pointer has the gang bit set, then the checksum
1637 * it represents is for the reconstituted data and not for an
1638 * individual gang member. The zio pipeline, however, must be able to
1639 * determine the checksum of each of the gang constituents so it
1640 * treats the checksum comparison differently than what we need
1641 * for l2arc blocks. This prevents us from using the
1642 * zio_checksum_error() interface directly. Instead we must call the
1643 * zio_checksum_error_impl() so that we can ensure the checksum is
1644 * generated using the correct checksum algorithm and accounts for the
1645 * logical I/O size and not just a gang fragment.
1647 valid_cksum
= (zio_checksum_error_impl(zio
->io_spa
, zio
->io_bp
,
1648 BP_GET_CHECKSUM(zio
->io_bp
), zio
->io_abd
, zio
->io_size
,
1649 zio
->io_offset
, NULL
) == 0);
1650 zio_pop_transforms(zio
);
1651 return (valid_cksum
);
1655 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1656 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1657 * isn't modified later on. If buf is compressed or there is already a checksum
1658 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1661 arc_cksum_compute(arc_buf_t
*buf
)
1663 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1665 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1668 ASSERT(HDR_HAS_L1HDR(hdr
));
1670 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1671 if (hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
) {
1672 ASSERT(arc_hdr_has_uncompressed_buf(hdr
));
1673 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1675 } else if (ARC_BUF_COMPRESSED(buf
)) {
1676 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1680 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1681 hdr
->b_l1hdr
.b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
),
1683 fletcher_2_native(buf
->b_data
, arc_buf_size(buf
), NULL
,
1684 hdr
->b_l1hdr
.b_freeze_cksum
);
1685 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1690 typedef struct procctl
{
1698 arc_buf_unwatch(arc_buf_t
*buf
)
1705 ctl
.prwatch
.pr_vaddr
= (uintptr_t)buf
->b_data
;
1706 ctl
.prwatch
.pr_size
= 0;
1707 ctl
.prwatch
.pr_wflags
= 0;
1708 result
= write(arc_procfd
, &ctl
, sizeof (ctl
));
1709 ASSERT3U(result
, ==, sizeof (ctl
));
1716 arc_buf_watch(arc_buf_t
*buf
)
1723 ctl
.prwatch
.pr_vaddr
= (uintptr_t)buf
->b_data
;
1724 ctl
.prwatch
.pr_size
= arc_buf_size(buf
);
1725 ctl
.prwatch
.pr_wflags
= WA_WRITE
;
1726 result
= write(arc_procfd
, &ctl
, sizeof (ctl
));
1727 ASSERT3U(result
, ==, sizeof (ctl
));
1732 static arc_buf_contents_t
1733 arc_buf_type(arc_buf_hdr_t
*hdr
)
1735 arc_buf_contents_t type
;
1736 if (HDR_ISTYPE_METADATA(hdr
)) {
1737 type
= ARC_BUFC_METADATA
;
1739 type
= ARC_BUFC_DATA
;
1741 VERIFY3U(hdr
->b_type
, ==, type
);
1746 arc_is_metadata(arc_buf_t
*buf
)
1748 return (HDR_ISTYPE_METADATA(buf
->b_hdr
) != 0);
1752 arc_bufc_to_flags(arc_buf_contents_t type
)
1756 /* metadata field is 0 if buffer contains normal data */
1758 case ARC_BUFC_METADATA
:
1759 return (ARC_FLAG_BUFC_METADATA
);
1763 panic("undefined ARC buffer type!");
1764 return ((uint32_t)-1);
1768 arc_buf_thaw(arc_buf_t
*buf
)
1770 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1772 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
1773 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
1775 arc_cksum_verify(buf
);
1778 * Compressed buffers do not manipulate the b_freeze_cksum or
1779 * allocate b_thawed.
1781 if (ARC_BUF_COMPRESSED(buf
)) {
1782 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1783 arc_hdr_has_uncompressed_buf(hdr
));
1787 ASSERT(HDR_HAS_L1HDR(hdr
));
1788 arc_cksum_free(hdr
);
1790 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
1792 if (zfs_flags
& ZFS_DEBUG_MODIFY
) {
1793 if (hdr
->b_l1hdr
.b_thawed
!= NULL
)
1794 kmem_free(hdr
->b_l1hdr
.b_thawed
, 1);
1795 hdr
->b_l1hdr
.b_thawed
= kmem_alloc(1, KM_SLEEP
);
1799 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
1801 arc_buf_unwatch(buf
);
1805 arc_buf_freeze(arc_buf_t
*buf
)
1807 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1808 kmutex_t
*hash_lock
;
1810 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1813 if (ARC_BUF_COMPRESSED(buf
)) {
1814 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
== NULL
||
1815 arc_hdr_has_uncompressed_buf(hdr
));
1819 hash_lock
= HDR_LOCK(hdr
);
1820 mutex_enter(hash_lock
);
1822 ASSERT(HDR_HAS_L1HDR(hdr
));
1823 ASSERT(hdr
->b_l1hdr
.b_freeze_cksum
!= NULL
||
1824 hdr
->b_l1hdr
.b_state
== arc_anon
);
1825 arc_cksum_compute(buf
);
1826 mutex_exit(hash_lock
);
1830 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1831 * the following functions should be used to ensure that the flags are
1832 * updated in a thread-safe way. When manipulating the flags either
1833 * the hash_lock must be held or the hdr must be undiscoverable. This
1834 * ensures that we're not racing with any other threads when updating
1838 arc_hdr_set_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1840 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1841 hdr
->b_flags
|= flags
;
1845 arc_hdr_clear_flags(arc_buf_hdr_t
*hdr
, arc_flags_t flags
)
1847 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1848 hdr
->b_flags
&= ~flags
;
1852 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1853 * done in a special way since we have to clear and set bits
1854 * at the same time. Consumers that wish to set the compression bits
1855 * must use this function to ensure that the flags are updated in
1856 * thread-safe manner.
1859 arc_hdr_set_compress(arc_buf_hdr_t
*hdr
, enum zio_compress cmp
)
1861 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
1864 * Holes and embedded blocks will always have a psize = 0 so
1865 * we ignore the compression of the blkptr and set the
1866 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
1867 * Holes and embedded blocks remain anonymous so we don't
1868 * want to uncompress them. Mark them as uncompressed.
1870 if (!zfs_compressed_arc_enabled
|| HDR_GET_PSIZE(hdr
) == 0) {
1871 arc_hdr_clear_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1872 HDR_SET_COMPRESS(hdr
, ZIO_COMPRESS_OFF
);
1873 ASSERT(!HDR_COMPRESSION_ENABLED(hdr
));
1874 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
1876 arc_hdr_set_flags(hdr
, ARC_FLAG_COMPRESSED_ARC
);
1877 HDR_SET_COMPRESS(hdr
, cmp
);
1878 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, cmp
);
1879 ASSERT(HDR_COMPRESSION_ENABLED(hdr
));
1884 * Looks for another buf on the same hdr which has the data decompressed, copies
1885 * from it, and returns true. If no such buf exists, returns false.
1888 arc_buf_try_copy_decompressed_data(arc_buf_t
*buf
)
1890 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1891 boolean_t copied
= B_FALSE
;
1893 ASSERT(HDR_HAS_L1HDR(hdr
));
1894 ASSERT3P(buf
->b_data
, !=, NULL
);
1895 ASSERT(!ARC_BUF_COMPRESSED(buf
));
1897 for (arc_buf_t
*from
= hdr
->b_l1hdr
.b_buf
; from
!= NULL
;
1898 from
= from
->b_next
) {
1899 /* can't use our own data buffer */
1904 if (!ARC_BUF_COMPRESSED(from
)) {
1905 bcopy(from
->b_data
, buf
->b_data
, arc_buf_size(buf
));
1912 * There were no decompressed bufs, so there should not be a
1913 * checksum on the hdr either.
1915 EQUIV(!copied
, hdr
->b_l1hdr
.b_freeze_cksum
== NULL
);
1921 * Given a buf that has a data buffer attached to it, this function will
1922 * efficiently fill the buf with data of the specified compression setting from
1923 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1924 * are already sharing a data buf, no copy is performed.
1926 * If the buf is marked as compressed but uncompressed data was requested, this
1927 * will allocate a new data buffer for the buf, remove that flag, and fill the
1928 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1929 * uncompressed data, and (since we haven't added support for it yet) if you
1930 * want compressed data your buf must already be marked as compressed and have
1931 * the correct-sized data buffer.
1934 arc_buf_fill(arc_buf_t
*buf
, boolean_t compressed
)
1936 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1937 boolean_t hdr_compressed
= (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
1938 dmu_object_byteswap_t bswap
= hdr
->b_l1hdr
.b_byteswap
;
1940 ASSERT3P(buf
->b_data
, !=, NULL
);
1941 IMPLY(compressed
, hdr_compressed
);
1942 IMPLY(compressed
, ARC_BUF_COMPRESSED(buf
));
1944 if (hdr_compressed
== compressed
) {
1945 if (!arc_buf_is_shared(buf
)) {
1946 abd_copy_to_buf(buf
->b_data
, hdr
->b_l1hdr
.b_pabd
,
1950 ASSERT(hdr_compressed
);
1951 ASSERT(!compressed
);
1952 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, HDR_GET_PSIZE(hdr
));
1955 * If the buf is sharing its data with the hdr, unlink it and
1956 * allocate a new data buffer for the buf.
1958 if (arc_buf_is_shared(buf
)) {
1959 ASSERT(ARC_BUF_COMPRESSED(buf
));
1961 /* We need to give the buf it's own b_data */
1962 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
1964 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
1965 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
1967 /* Previously overhead was 0; just add new overhead */
1968 ARCSTAT_INCR(arcstat_overhead_size
, HDR_GET_LSIZE(hdr
));
1969 } else if (ARC_BUF_COMPRESSED(buf
)) {
1970 /* We need to reallocate the buf's b_data */
1971 arc_free_data_buf(hdr
, buf
->b_data
, HDR_GET_PSIZE(hdr
),
1974 arc_get_data_buf(hdr
, HDR_GET_LSIZE(hdr
), buf
);
1976 /* We increased the size of b_data; update overhead */
1977 ARCSTAT_INCR(arcstat_overhead_size
,
1978 HDR_GET_LSIZE(hdr
) - HDR_GET_PSIZE(hdr
));
1982 * Regardless of the buf's previous compression settings, it
1983 * should not be compressed at the end of this function.
1985 buf
->b_flags
&= ~ARC_BUF_FLAG_COMPRESSED
;
1988 * Try copying the data from another buf which already has a
1989 * decompressed version. If that's not possible, it's time to
1990 * bite the bullet and decompress the data from the hdr.
1992 if (arc_buf_try_copy_decompressed_data(buf
)) {
1993 /* Skip byteswapping and checksumming (already done) */
1994 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, !=, NULL
);
1997 int error
= zio_decompress_data(HDR_GET_COMPRESS(hdr
),
1998 hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
1999 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
2002 * Absent hardware errors or software bugs, this should
2003 * be impossible, but log it anyway so we can debug it.
2007 "hdr %p, compress %d, psize %d, lsize %d",
2008 hdr
, HDR_GET_COMPRESS(hdr
),
2009 HDR_GET_PSIZE(hdr
), HDR_GET_LSIZE(hdr
));
2010 return (SET_ERROR(EIO
));
2015 /* Byteswap the buf's data if necessary */
2016 if (bswap
!= DMU_BSWAP_NUMFUNCS
) {
2017 ASSERT(!HDR_SHARED_DATA(hdr
));
2018 ASSERT3U(bswap
, <, DMU_BSWAP_NUMFUNCS
);
2019 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, HDR_GET_LSIZE(hdr
));
2022 /* Compute the hdr's checksum if necessary */
2023 arc_cksum_compute(buf
);
2029 arc_decompress(arc_buf_t
*buf
)
2031 return (arc_buf_fill(buf
, B_FALSE
));
2035 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2038 arc_hdr_size(arc_buf_hdr_t
*hdr
)
2042 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
&&
2043 HDR_GET_PSIZE(hdr
) > 0) {
2044 size
= HDR_GET_PSIZE(hdr
);
2046 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, 0);
2047 size
= HDR_GET_LSIZE(hdr
);
2053 * Increment the amount of evictable space in the arc_state_t's refcount.
2054 * We account for the space used by the hdr and the arc buf individually
2055 * so that we can add and remove them from the refcount individually.
2058 arc_evictable_space_increment(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
2060 arc_buf_contents_t type
= arc_buf_type(hdr
);
2062 ASSERT(HDR_HAS_L1HDR(hdr
));
2064 if (GHOST_STATE(state
)) {
2065 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
2066 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2067 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2068 (void) refcount_add_many(&state
->arcs_esize
[type
],
2069 HDR_GET_LSIZE(hdr
), hdr
);
2073 ASSERT(!GHOST_STATE(state
));
2074 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2075 (void) refcount_add_many(&state
->arcs_esize
[type
],
2076 arc_hdr_size(hdr
), hdr
);
2078 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2079 buf
= buf
->b_next
) {
2080 if (arc_buf_is_shared(buf
))
2082 (void) refcount_add_many(&state
->arcs_esize
[type
],
2083 arc_buf_size(buf
), buf
);
2088 * Decrement the amount of evictable space in the arc_state_t's refcount.
2089 * We account for the space used by the hdr and the arc buf individually
2090 * so that we can add and remove them from the refcount individually.
2093 arc_evictable_space_decrement(arc_buf_hdr_t
*hdr
, arc_state_t
*state
)
2095 arc_buf_contents_t type
= arc_buf_type(hdr
);
2097 ASSERT(HDR_HAS_L1HDR(hdr
));
2099 if (GHOST_STATE(state
)) {
2100 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
2101 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2102 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2103 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2104 HDR_GET_LSIZE(hdr
), hdr
);
2108 ASSERT(!GHOST_STATE(state
));
2109 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2110 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2111 arc_hdr_size(hdr
), hdr
);
2113 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2114 buf
= buf
->b_next
) {
2115 if (arc_buf_is_shared(buf
))
2117 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2118 arc_buf_size(buf
), buf
);
2123 * Add a reference to this hdr indicating that someone is actively
2124 * referencing that memory. When the refcount transitions from 0 to 1,
2125 * we remove it from the respective arc_state_t list to indicate that
2126 * it is not evictable.
2129 add_reference(arc_buf_hdr_t
*hdr
, void *tag
)
2131 ASSERT(HDR_HAS_L1HDR(hdr
));
2132 if (!MUTEX_HELD(HDR_LOCK(hdr
))) {
2133 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
2134 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2135 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2138 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2140 if ((refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
) == 1) &&
2141 (state
!= arc_anon
)) {
2142 /* We don't use the L2-only state list. */
2143 if (state
!= arc_l2c_only
) {
2144 multilist_remove(state
->arcs_list
[arc_buf_type(hdr
)],
2146 arc_evictable_space_decrement(hdr
, state
);
2148 /* remove the prefetch flag if we get a reference */
2149 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
2154 * Remove a reference from this hdr. When the reference transitions from
2155 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2156 * list making it eligible for eviction.
2159 remove_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
2162 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2164 ASSERT(HDR_HAS_L1HDR(hdr
));
2165 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
2166 ASSERT(!GHOST_STATE(state
));
2169 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2170 * check to prevent usage of the arc_l2c_only list.
2172 if (((cnt
= refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
)) == 0) &&
2173 (state
!= arc_anon
)) {
2174 multilist_insert(state
->arcs_list
[arc_buf_type(hdr
)], hdr
);
2175 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
2176 arc_evictable_space_increment(hdr
, state
);
2182 * Move the supplied buffer to the indicated state. The hash lock
2183 * for the buffer must be held by the caller.
2186 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*hdr
,
2187 kmutex_t
*hash_lock
)
2189 arc_state_t
*old_state
;
2192 boolean_t update_old
, update_new
;
2193 arc_buf_contents_t buftype
= arc_buf_type(hdr
);
2196 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2197 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2198 * L1 hdr doesn't always exist when we change state to arc_anon before
2199 * destroying a header, in which case reallocating to add the L1 hdr is
2202 if (HDR_HAS_L1HDR(hdr
)) {
2203 old_state
= hdr
->b_l1hdr
.b_state
;
2204 refcnt
= refcount_count(&hdr
->b_l1hdr
.b_refcnt
);
2205 bufcnt
= hdr
->b_l1hdr
.b_bufcnt
;
2206 update_old
= (bufcnt
> 0 || hdr
->b_l1hdr
.b_pabd
!= NULL
);
2208 old_state
= arc_l2c_only
;
2211 update_old
= B_FALSE
;
2213 update_new
= update_old
;
2215 ASSERT(MUTEX_HELD(hash_lock
));
2216 ASSERT3P(new_state
, !=, old_state
);
2217 ASSERT(!GHOST_STATE(new_state
) || bufcnt
== 0);
2218 ASSERT(old_state
!= arc_anon
|| bufcnt
<= 1);
2221 * If this buffer is evictable, transfer it from the
2222 * old state list to the new state list.
2225 if (old_state
!= arc_anon
&& old_state
!= arc_l2c_only
) {
2226 ASSERT(HDR_HAS_L1HDR(hdr
));
2227 multilist_remove(old_state
->arcs_list
[buftype
], hdr
);
2229 if (GHOST_STATE(old_state
)) {
2231 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2232 update_old
= B_TRUE
;
2234 arc_evictable_space_decrement(hdr
, old_state
);
2236 if (new_state
!= arc_anon
&& new_state
!= arc_l2c_only
) {
2239 * An L1 header always exists here, since if we're
2240 * moving to some L1-cached state (i.e. not l2c_only or
2241 * anonymous), we realloc the header to add an L1hdr
2244 ASSERT(HDR_HAS_L1HDR(hdr
));
2245 multilist_insert(new_state
->arcs_list
[buftype
], hdr
);
2247 if (GHOST_STATE(new_state
)) {
2249 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2250 update_new
= B_TRUE
;
2252 arc_evictable_space_increment(hdr
, new_state
);
2256 ASSERT(!HDR_EMPTY(hdr
));
2257 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(hdr
))
2258 buf_hash_remove(hdr
);
2260 /* adjust state sizes (ignore arc_l2c_only) */
2262 if (update_new
&& new_state
!= arc_l2c_only
) {
2263 ASSERT(HDR_HAS_L1HDR(hdr
));
2264 if (GHOST_STATE(new_state
)) {
2268 * When moving a header to a ghost state, we first
2269 * remove all arc buffers. Thus, we'll have a
2270 * bufcnt of zero, and no arc buffer to use for
2271 * the reference. As a result, we use the arc
2272 * header pointer for the reference.
2274 (void) refcount_add_many(&new_state
->arcs_size
,
2275 HDR_GET_LSIZE(hdr
), hdr
);
2276 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2278 uint32_t buffers
= 0;
2281 * Each individual buffer holds a unique reference,
2282 * thus we must remove each of these references one
2285 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2286 buf
= buf
->b_next
) {
2287 ASSERT3U(bufcnt
, !=, 0);
2291 * When the arc_buf_t is sharing the data
2292 * block with the hdr, the owner of the
2293 * reference belongs to the hdr. Only
2294 * add to the refcount if the arc_buf_t is
2297 if (arc_buf_is_shared(buf
))
2300 (void) refcount_add_many(&new_state
->arcs_size
,
2301 arc_buf_size(buf
), buf
);
2303 ASSERT3U(bufcnt
, ==, buffers
);
2305 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
2306 (void) refcount_add_many(&new_state
->arcs_size
,
2307 arc_hdr_size(hdr
), hdr
);
2309 ASSERT(GHOST_STATE(old_state
));
2314 if (update_old
&& old_state
!= arc_l2c_only
) {
2315 ASSERT(HDR_HAS_L1HDR(hdr
));
2316 if (GHOST_STATE(old_state
)) {
2318 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2321 * When moving a header off of a ghost state,
2322 * the header will not contain any arc buffers.
2323 * We use the arc header pointer for the reference
2324 * which is exactly what we did when we put the
2325 * header on the ghost state.
2328 (void) refcount_remove_many(&old_state
->arcs_size
,
2329 HDR_GET_LSIZE(hdr
), hdr
);
2331 uint32_t buffers
= 0;
2334 * Each individual buffer holds a unique reference,
2335 * thus we must remove each of these references one
2338 for (arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
2339 buf
= buf
->b_next
) {
2340 ASSERT3U(bufcnt
, !=, 0);
2344 * When the arc_buf_t is sharing the data
2345 * block with the hdr, the owner of the
2346 * reference belongs to the hdr. Only
2347 * add to the refcount if the arc_buf_t is
2350 if (arc_buf_is_shared(buf
))
2353 (void) refcount_remove_many(
2354 &old_state
->arcs_size
, arc_buf_size(buf
),
2357 ASSERT3U(bufcnt
, ==, buffers
);
2358 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2359 (void) refcount_remove_many(
2360 &old_state
->arcs_size
, arc_hdr_size(hdr
), hdr
);
2364 if (HDR_HAS_L1HDR(hdr
))
2365 hdr
->b_l1hdr
.b_state
= new_state
;
2368 * L2 headers should never be on the L2 state list since they don't
2369 * have L1 headers allocated.
2371 ASSERT(multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]) &&
2372 multilist_is_empty(arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]));
2376 arc_space_consume(uint64_t space
, arc_space_type_t type
)
2378 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2381 case ARC_SPACE_DATA
:
2382 aggsum_add(&astat_data_size
, space
);
2384 case ARC_SPACE_META
:
2385 aggsum_add(&astat_metadata_size
, space
);
2387 case ARC_SPACE_OTHER
:
2388 aggsum_add(&astat_other_size
, space
);
2390 case ARC_SPACE_HDRS
:
2391 aggsum_add(&astat_hdr_size
, space
);
2393 case ARC_SPACE_L2HDRS
:
2394 aggsum_add(&astat_l2_hdr_size
, space
);
2398 if (type
!= ARC_SPACE_DATA
)
2399 aggsum_add(&arc_meta_used
, space
);
2401 aggsum_add(&arc_size
, space
);
2405 arc_space_return(uint64_t space
, arc_space_type_t type
)
2407 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
2410 case ARC_SPACE_DATA
:
2411 aggsum_add(&astat_data_size
, -space
);
2413 case ARC_SPACE_META
:
2414 aggsum_add(&astat_metadata_size
, -space
);
2416 case ARC_SPACE_OTHER
:
2417 aggsum_add(&astat_other_size
, -space
);
2419 case ARC_SPACE_HDRS
:
2420 aggsum_add(&astat_hdr_size
, -space
);
2422 case ARC_SPACE_L2HDRS
:
2423 aggsum_add(&astat_l2_hdr_size
, -space
);
2427 if (type
!= ARC_SPACE_DATA
) {
2428 ASSERT(aggsum_compare(&arc_meta_used
, space
) >= 0);
2430 * We use the upper bound here rather than the precise value
2431 * because the arc_meta_max value doesn't need to be
2432 * precise. It's only consumed by humans via arcstats.
2434 if (arc_meta_max
< aggsum_upper_bound(&arc_meta_used
))
2435 arc_meta_max
= aggsum_upper_bound(&arc_meta_used
);
2436 aggsum_add(&arc_meta_used
, -space
);
2439 ASSERT(aggsum_compare(&arc_size
, space
) >= 0);
2440 aggsum_add(&arc_size
, -space
);
2444 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2445 * with the hdr's b_pabd.
2448 arc_can_share(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2451 * The criteria for sharing a hdr's data are:
2452 * 1. the hdr's compression matches the buf's compression
2453 * 2. the hdr doesn't need to be byteswapped
2454 * 3. the hdr isn't already being shared
2455 * 4. the buf is either compressed or it is the last buf in the hdr list
2457 * Criterion #4 maintains the invariant that shared uncompressed
2458 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2459 * might ask, "if a compressed buf is allocated first, won't that be the
2460 * last thing in the list?", but in that case it's impossible to create
2461 * a shared uncompressed buf anyway (because the hdr must be compressed
2462 * to have the compressed buf). You might also think that #3 is
2463 * sufficient to make this guarantee, however it's possible
2464 * (specifically in the rare L2ARC write race mentioned in
2465 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2466 * is sharable, but wasn't at the time of its allocation. Rather than
2467 * allow a new shared uncompressed buf to be created and then shuffle
2468 * the list around to make it the last element, this simply disallows
2469 * sharing if the new buf isn't the first to be added.
2471 ASSERT3P(buf
->b_hdr
, ==, hdr
);
2472 boolean_t hdr_compressed
= HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
;
2473 boolean_t buf_compressed
= ARC_BUF_COMPRESSED(buf
) != 0;
2474 return (buf_compressed
== hdr_compressed
&&
2475 hdr
->b_l1hdr
.b_byteswap
== DMU_BSWAP_NUMFUNCS
&&
2476 !HDR_SHARED_DATA(hdr
) &&
2477 (ARC_BUF_LAST(buf
) || ARC_BUF_COMPRESSED(buf
)));
2481 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2482 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2483 * copy was made successfully, or an error code otherwise.
2486 arc_buf_alloc_impl(arc_buf_hdr_t
*hdr
, void *tag
, boolean_t compressed
,
2487 boolean_t fill
, arc_buf_t
**ret
)
2491 ASSERT(HDR_HAS_L1HDR(hdr
));
2492 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2493 VERIFY(hdr
->b_type
== ARC_BUFC_DATA
||
2494 hdr
->b_type
== ARC_BUFC_METADATA
);
2495 ASSERT3P(ret
, !=, NULL
);
2496 ASSERT3P(*ret
, ==, NULL
);
2498 buf
= *ret
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
2501 buf
->b_next
= hdr
->b_l1hdr
.b_buf
;
2504 add_reference(hdr
, tag
);
2507 * We're about to change the hdr's b_flags. We must either
2508 * hold the hash_lock or be undiscoverable.
2510 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2513 * Only honor requests for compressed bufs if the hdr is actually
2516 if (compressed
&& HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
)
2517 buf
->b_flags
|= ARC_BUF_FLAG_COMPRESSED
;
2520 * If the hdr's data can be shared then we share the data buffer and
2521 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2522 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2523 * buffer to store the buf's data.
2525 * There are two additional restrictions here because we're sharing
2526 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2527 * actively involved in an L2ARC write, because if this buf is used by
2528 * an arc_write() then the hdr's data buffer will be released when the
2529 * write completes, even though the L2ARC write might still be using it.
2530 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2531 * need to be ABD-aware.
2533 boolean_t can_share
= arc_can_share(hdr
, buf
) && !HDR_L2_WRITING(hdr
) &&
2534 abd_is_linear(hdr
->b_l1hdr
.b_pabd
);
2536 /* Set up b_data and sharing */
2538 buf
->b_data
= abd_to_buf(hdr
->b_l1hdr
.b_pabd
);
2539 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2540 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2543 arc_get_data_buf(hdr
, arc_buf_size(buf
), buf
);
2544 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2546 VERIFY3P(buf
->b_data
, !=, NULL
);
2548 hdr
->b_l1hdr
.b_buf
= buf
;
2549 hdr
->b_l1hdr
.b_bufcnt
+= 1;
2552 * If the user wants the data from the hdr, we need to either copy or
2553 * decompress the data.
2556 return (arc_buf_fill(buf
, ARC_BUF_COMPRESSED(buf
) != 0));
2562 static char *arc_onloan_tag
= "onloan";
2565 arc_loaned_bytes_update(int64_t delta
)
2567 atomic_add_64(&arc_loaned_bytes
, delta
);
2569 /* assert that it did not wrap around */
2570 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
2574 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2575 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2576 * buffers must be returned to the arc before they can be used by the DMU or
2580 arc_loan_buf(spa_t
*spa
, boolean_t is_metadata
, int size
)
2582 arc_buf_t
*buf
= arc_alloc_buf(spa
, arc_onloan_tag
,
2583 is_metadata
? ARC_BUFC_METADATA
: ARC_BUFC_DATA
, size
);
2585 arc_loaned_bytes_update(arc_buf_size(buf
));
2591 arc_loan_compressed_buf(spa_t
*spa
, uint64_t psize
, uint64_t lsize
,
2592 enum zio_compress compression_type
)
2594 arc_buf_t
*buf
= arc_alloc_compressed_buf(spa
, arc_onloan_tag
,
2595 psize
, lsize
, compression_type
);
2597 arc_loaned_bytes_update(arc_buf_size(buf
));
2604 * Return a loaned arc buffer to the arc.
2607 arc_return_buf(arc_buf_t
*buf
, void *tag
)
2609 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2611 ASSERT3P(buf
->b_data
, !=, NULL
);
2612 ASSERT(HDR_HAS_L1HDR(hdr
));
2613 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2614 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2616 arc_loaned_bytes_update(-arc_buf_size(buf
));
2619 /* Detach an arc_buf from a dbuf (tag) */
2621 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
2623 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2625 ASSERT3P(buf
->b_data
, !=, NULL
);
2626 ASSERT(HDR_HAS_L1HDR(hdr
));
2627 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
2628 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
);
2630 arc_loaned_bytes_update(arc_buf_size(buf
));
2634 l2arc_free_abd_on_write(abd_t
*abd
, size_t size
, arc_buf_contents_t type
)
2636 l2arc_data_free_t
*df
= kmem_alloc(sizeof (*df
), KM_SLEEP
);
2639 df
->l2df_size
= size
;
2640 df
->l2df_type
= type
;
2641 mutex_enter(&l2arc_free_on_write_mtx
);
2642 list_insert_head(l2arc_free_on_write
, df
);
2643 mutex_exit(&l2arc_free_on_write_mtx
);
2647 arc_hdr_free_on_write(arc_buf_hdr_t
*hdr
)
2649 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2650 arc_buf_contents_t type
= arc_buf_type(hdr
);
2651 uint64_t size
= arc_hdr_size(hdr
);
2653 /* protected by hash lock, if in the hash table */
2654 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
2655 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2656 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
2658 (void) refcount_remove_many(&state
->arcs_esize
[type
],
2661 (void) refcount_remove_many(&state
->arcs_size
, size
, hdr
);
2662 if (type
== ARC_BUFC_METADATA
) {
2663 arc_space_return(size
, ARC_SPACE_META
);
2665 ASSERT(type
== ARC_BUFC_DATA
);
2666 arc_space_return(size
, ARC_SPACE_DATA
);
2669 l2arc_free_abd_on_write(hdr
->b_l1hdr
.b_pabd
, size
, type
);
2673 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2674 * data buffer, we transfer the refcount ownership to the hdr and update
2675 * the appropriate kstats.
2678 arc_share_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2680 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2682 ASSERT(arc_can_share(hdr
, buf
));
2683 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2684 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2687 * Start sharing the data buffer. We transfer the
2688 * refcount ownership to the hdr since it always owns
2689 * the refcount whenever an arc_buf_t is shared.
2691 refcount_transfer_ownership(&state
->arcs_size
, buf
, hdr
);
2692 hdr
->b_l1hdr
.b_pabd
= abd_get_from_buf(buf
->b_data
, arc_buf_size(buf
));
2693 abd_take_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
,
2694 HDR_ISTYPE_METADATA(hdr
));
2695 arc_hdr_set_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2696 buf
->b_flags
|= ARC_BUF_FLAG_SHARED
;
2699 * Since we've transferred ownership to the hdr we need
2700 * to increment its compressed and uncompressed kstats and
2701 * decrement the overhead size.
2703 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
2704 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
2705 ARCSTAT_INCR(arcstat_overhead_size
, -arc_buf_size(buf
));
2709 arc_unshare_buf(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2711 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
2713 ASSERT(arc_buf_is_shared(buf
));
2714 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2715 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2718 * We are no longer sharing this buffer so we need
2719 * to transfer its ownership to the rightful owner.
2721 refcount_transfer_ownership(&state
->arcs_size
, hdr
, buf
);
2722 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2723 abd_release_ownership_of_buf(hdr
->b_l1hdr
.b_pabd
);
2724 abd_put(hdr
->b_l1hdr
.b_pabd
);
2725 hdr
->b_l1hdr
.b_pabd
= NULL
;
2726 buf
->b_flags
&= ~ARC_BUF_FLAG_SHARED
;
2729 * Since the buffer is no longer shared between
2730 * the arc buf and the hdr, count it as overhead.
2732 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
2733 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
2734 ARCSTAT_INCR(arcstat_overhead_size
, arc_buf_size(buf
));
2738 * Remove an arc_buf_t from the hdr's buf list and return the last
2739 * arc_buf_t on the list. If no buffers remain on the list then return
2743 arc_buf_remove(arc_buf_hdr_t
*hdr
, arc_buf_t
*buf
)
2745 ASSERT(HDR_HAS_L1HDR(hdr
));
2746 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2748 arc_buf_t
**bufp
= &hdr
->b_l1hdr
.b_buf
;
2749 arc_buf_t
*lastbuf
= NULL
;
2752 * Remove the buf from the hdr list and locate the last
2753 * remaining buffer on the list.
2755 while (*bufp
!= NULL
) {
2757 *bufp
= buf
->b_next
;
2760 * If we've removed a buffer in the middle of
2761 * the list then update the lastbuf and update
2764 if (*bufp
!= NULL
) {
2766 bufp
= &(*bufp
)->b_next
;
2770 ASSERT3P(lastbuf
, !=, buf
);
2771 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, lastbuf
!= NULL
);
2772 IMPLY(hdr
->b_l1hdr
.b_bufcnt
> 0, hdr
->b_l1hdr
.b_buf
!= NULL
);
2773 IMPLY(lastbuf
!= NULL
, ARC_BUF_LAST(lastbuf
));
2779 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
2783 arc_buf_destroy_impl(arc_buf_t
*buf
)
2785 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2788 * Free up the data associated with the buf but only if we're not
2789 * sharing this with the hdr. If we are sharing it with the hdr, the
2790 * hdr is responsible for doing the free.
2792 if (buf
->b_data
!= NULL
) {
2794 * We're about to change the hdr's b_flags. We must either
2795 * hold the hash_lock or be undiscoverable.
2797 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)) || HDR_EMPTY(hdr
));
2799 arc_cksum_verify(buf
);
2800 arc_buf_unwatch(buf
);
2802 if (arc_buf_is_shared(buf
)) {
2803 arc_hdr_clear_flags(hdr
, ARC_FLAG_SHARED_DATA
);
2805 uint64_t size
= arc_buf_size(buf
);
2806 arc_free_data_buf(hdr
, buf
->b_data
, size
, buf
);
2807 ARCSTAT_INCR(arcstat_overhead_size
, -size
);
2811 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
2812 hdr
->b_l1hdr
.b_bufcnt
-= 1;
2815 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
2817 if (ARC_BUF_SHARED(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
2819 * If the current arc_buf_t is sharing its data buffer with the
2820 * hdr, then reassign the hdr's b_pabd to share it with the new
2821 * buffer at the end of the list. The shared buffer is always
2822 * the last one on the hdr's buffer list.
2824 * There is an equivalent case for compressed bufs, but since
2825 * they aren't guaranteed to be the last buf in the list and
2826 * that is an exceedingly rare case, we just allow that space be
2827 * wasted temporarily.
2829 if (lastbuf
!= NULL
) {
2830 /* Only one buf can be shared at once */
2831 VERIFY(!arc_buf_is_shared(lastbuf
));
2832 /* hdr is uncompressed so can't have compressed buf */
2833 VERIFY(!ARC_BUF_COMPRESSED(lastbuf
));
2835 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2836 arc_hdr_free_pabd(hdr
);
2839 * We must setup a new shared block between the
2840 * last buffer and the hdr. The data would have
2841 * been allocated by the arc buf so we need to transfer
2842 * ownership to the hdr since it's now being shared.
2844 arc_share_buf(hdr
, lastbuf
);
2846 } else if (HDR_SHARED_DATA(hdr
)) {
2848 * Uncompressed shared buffers are always at the end
2849 * of the list. Compressed buffers don't have the
2850 * same requirements. This makes it hard to
2851 * simply assert that the lastbuf is shared so
2852 * we rely on the hdr's compression flags to determine
2853 * if we have a compressed, shared buffer.
2855 ASSERT3P(lastbuf
, !=, NULL
);
2856 ASSERT(arc_buf_is_shared(lastbuf
) ||
2857 HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
2861 * Free the checksum if we're removing the last uncompressed buf from
2864 if (!arc_hdr_has_uncompressed_buf(hdr
)) {
2865 arc_cksum_free(hdr
);
2868 /* clean up the buf */
2870 kmem_cache_free(buf_cache
, buf
);
2874 arc_hdr_alloc_pabd(arc_buf_hdr_t
*hdr
)
2876 ASSERT3U(HDR_GET_LSIZE(hdr
), >, 0);
2877 ASSERT(HDR_HAS_L1HDR(hdr
));
2878 ASSERT(!HDR_SHARED_DATA(hdr
));
2880 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2881 hdr
->b_l1hdr
.b_pabd
= arc_get_data_abd(hdr
, arc_hdr_size(hdr
), hdr
);
2882 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
2883 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2885 ARCSTAT_INCR(arcstat_compressed_size
, arc_hdr_size(hdr
));
2886 ARCSTAT_INCR(arcstat_uncompressed_size
, HDR_GET_LSIZE(hdr
));
2890 arc_hdr_free_pabd(arc_buf_hdr_t
*hdr
)
2892 ASSERT(HDR_HAS_L1HDR(hdr
));
2893 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
2896 * If the hdr is currently being written to the l2arc then
2897 * we defer freeing the data by adding it to the l2arc_free_on_write
2898 * list. The l2arc will free the data once it's finished
2899 * writing it to the l2arc device.
2901 if (HDR_L2_WRITING(hdr
)) {
2902 arc_hdr_free_on_write(hdr
);
2903 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
2905 arc_free_data_abd(hdr
, hdr
->b_l1hdr
.b_pabd
,
2906 arc_hdr_size(hdr
), hdr
);
2908 hdr
->b_l1hdr
.b_pabd
= NULL
;
2909 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
2911 ARCSTAT_INCR(arcstat_compressed_size
, -arc_hdr_size(hdr
));
2912 ARCSTAT_INCR(arcstat_uncompressed_size
, -HDR_GET_LSIZE(hdr
));
2915 static arc_buf_hdr_t
*
2916 arc_hdr_alloc(uint64_t spa
, int32_t psize
, int32_t lsize
,
2917 enum zio_compress compression_type
, arc_buf_contents_t type
)
2921 VERIFY(type
== ARC_BUFC_DATA
|| type
== ARC_BUFC_METADATA
);
2923 hdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
2924 ASSERT(HDR_EMPTY(hdr
));
2925 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2926 ASSERT3P(hdr
->b_l1hdr
.b_thawed
, ==, NULL
);
2927 HDR_SET_PSIZE(hdr
, psize
);
2928 HDR_SET_LSIZE(hdr
, lsize
);
2932 arc_hdr_set_flags(hdr
, arc_bufc_to_flags(type
) | ARC_FLAG_HAS_L1HDR
);
2933 arc_hdr_set_compress(hdr
, compression_type
);
2935 hdr
->b_l1hdr
.b_state
= arc_anon
;
2936 hdr
->b_l1hdr
.b_arc_access
= 0;
2937 hdr
->b_l1hdr
.b_bufcnt
= 0;
2938 hdr
->b_l1hdr
.b_buf
= NULL
;
2941 * Allocate the hdr's buffer. This will contain either
2942 * the compressed or uncompressed data depending on the block
2943 * it references and compressed arc enablement.
2945 arc_hdr_alloc_pabd(hdr
);
2946 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2952 * Transition between the two allocation states for the arc_buf_hdr struct.
2953 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
2954 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
2955 * version is used when a cache buffer is only in the L2ARC in order to reduce
2958 static arc_buf_hdr_t
*
2959 arc_hdr_realloc(arc_buf_hdr_t
*hdr
, kmem_cache_t
*old
, kmem_cache_t
*new)
2961 ASSERT(HDR_HAS_L2HDR(hdr
));
2963 arc_buf_hdr_t
*nhdr
;
2964 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
2966 ASSERT((old
== hdr_full_cache
&& new == hdr_l2only_cache
) ||
2967 (old
== hdr_l2only_cache
&& new == hdr_full_cache
));
2969 nhdr
= kmem_cache_alloc(new, KM_PUSHPAGE
);
2971 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
2972 buf_hash_remove(hdr
);
2974 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
2976 if (new == hdr_full_cache
) {
2977 arc_hdr_set_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
2979 * arc_access and arc_change_state need to be aware that a
2980 * header has just come out of L2ARC, so we set its state to
2981 * l2c_only even though it's about to change.
2983 nhdr
->b_l1hdr
.b_state
= arc_l2c_only
;
2985 /* Verify previous threads set to NULL before freeing */
2986 ASSERT3P(nhdr
->b_l1hdr
.b_pabd
, ==, NULL
);
2988 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
2989 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
2990 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
2993 * If we've reached here, We must have been called from
2994 * arc_evict_hdr(), as such we should have already been
2995 * removed from any ghost list we were previously on
2996 * (which protects us from racing with arc_evict_state),
2997 * thus no locking is needed during this check.
2999 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3002 * A buffer must not be moved into the arc_l2c_only
3003 * state if it's not finished being written out to the
3004 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3005 * might try to be accessed, even though it was removed.
3007 VERIFY(!HDR_L2_WRITING(hdr
));
3008 VERIFY3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
3011 if (hdr
->b_l1hdr
.b_thawed
!= NULL
) {
3012 kmem_free(hdr
->b_l1hdr
.b_thawed
, 1);
3013 hdr
->b_l1hdr
.b_thawed
= NULL
;
3017 arc_hdr_clear_flags(nhdr
, ARC_FLAG_HAS_L1HDR
);
3020 * The header has been reallocated so we need to re-insert it into any
3023 (void) buf_hash_insert(nhdr
, NULL
);
3025 ASSERT(list_link_active(&hdr
->b_l2hdr
.b_l2node
));
3027 mutex_enter(&dev
->l2ad_mtx
);
3030 * We must place the realloc'ed header back into the list at
3031 * the same spot. Otherwise, if it's placed earlier in the list,
3032 * l2arc_write_buffers() could find it during the function's
3033 * write phase, and try to write it out to the l2arc.
3035 list_insert_after(&dev
->l2ad_buflist
, hdr
, nhdr
);
3036 list_remove(&dev
->l2ad_buflist
, hdr
);
3038 mutex_exit(&dev
->l2ad_mtx
);
3041 * Since we're using the pointer address as the tag when
3042 * incrementing and decrementing the l2ad_alloc refcount, we
3043 * must remove the old pointer (that we're about to destroy) and
3044 * add the new pointer to the refcount. Otherwise we'd remove
3045 * the wrong pointer address when calling arc_hdr_destroy() later.
3048 (void) refcount_remove_many(&dev
->l2ad_alloc
, arc_hdr_size(hdr
), hdr
);
3049 (void) refcount_add_many(&dev
->l2ad_alloc
, arc_hdr_size(nhdr
), nhdr
);
3051 buf_discard_identity(hdr
);
3052 kmem_cache_free(old
, hdr
);
3058 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3059 * The buf is returned thawed since we expect the consumer to modify it.
3062 arc_alloc_buf(spa_t
*spa
, void *tag
, arc_buf_contents_t type
, int32_t size
)
3064 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), size
, size
,
3065 ZIO_COMPRESS_OFF
, type
);
3066 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
3068 arc_buf_t
*buf
= NULL
;
3069 VERIFY0(arc_buf_alloc_impl(hdr
, tag
, B_FALSE
, B_FALSE
, &buf
));
3076 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3077 * for bufs containing metadata.
3080 arc_alloc_compressed_buf(spa_t
*spa
, void *tag
, uint64_t psize
, uint64_t lsize
,
3081 enum zio_compress compression_type
)
3083 ASSERT3U(lsize
, >, 0);
3084 ASSERT3U(lsize
, >=, psize
);
3085 ASSERT(compression_type
> ZIO_COMPRESS_OFF
);
3086 ASSERT(compression_type
< ZIO_COMPRESS_FUNCTIONS
);
3088 arc_buf_hdr_t
*hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
3089 compression_type
, ARC_BUFC_DATA
);
3090 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr
)));
3092 arc_buf_t
*buf
= NULL
;
3093 VERIFY0(arc_buf_alloc_impl(hdr
, tag
, B_TRUE
, B_FALSE
, &buf
));
3095 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
3097 if (!arc_buf_is_shared(buf
)) {
3099 * To ensure that the hdr has the correct data in it if we call
3100 * arc_decompress() on this buf before it's been written to
3101 * disk, it's easiest if we just set up sharing between the
3104 ASSERT(!abd_is_linear(hdr
->b_l1hdr
.b_pabd
));
3105 arc_hdr_free_pabd(hdr
);
3106 arc_share_buf(hdr
, buf
);
3113 arc_hdr_l2hdr_destroy(arc_buf_hdr_t
*hdr
)
3115 l2arc_buf_hdr_t
*l2hdr
= &hdr
->b_l2hdr
;
3116 l2arc_dev_t
*dev
= l2hdr
->b_dev
;
3117 uint64_t psize
= arc_hdr_size(hdr
);
3119 ASSERT(MUTEX_HELD(&dev
->l2ad_mtx
));
3120 ASSERT(HDR_HAS_L2HDR(hdr
));
3122 list_remove(&dev
->l2ad_buflist
, hdr
);
3124 ARCSTAT_INCR(arcstat_l2_psize
, -psize
);
3125 ARCSTAT_INCR(arcstat_l2_lsize
, -HDR_GET_LSIZE(hdr
));
3127 vdev_space_update(dev
->l2ad_vdev
, -psize
, 0, 0);
3129 (void) refcount_remove_many(&dev
->l2ad_alloc
, psize
, hdr
);
3130 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
3134 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
3136 if (HDR_HAS_L1HDR(hdr
)) {
3137 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
||
3138 hdr
->b_l1hdr
.b_bufcnt
> 0);
3139 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3140 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
3142 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3143 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
3145 if (!HDR_EMPTY(hdr
))
3146 buf_discard_identity(hdr
);
3148 if (HDR_HAS_L2HDR(hdr
)) {
3149 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
3150 boolean_t buflist_held
= MUTEX_HELD(&dev
->l2ad_mtx
);
3153 mutex_enter(&dev
->l2ad_mtx
);
3156 * Even though we checked this conditional above, we
3157 * need to check this again now that we have the
3158 * l2ad_mtx. This is because we could be racing with
3159 * another thread calling l2arc_evict() which might have
3160 * destroyed this header's L2 portion as we were waiting
3161 * to acquire the l2ad_mtx. If that happens, we don't
3162 * want to re-destroy the header's L2 portion.
3164 if (HDR_HAS_L2HDR(hdr
))
3165 arc_hdr_l2hdr_destroy(hdr
);
3168 mutex_exit(&dev
->l2ad_mtx
);
3171 if (HDR_HAS_L1HDR(hdr
)) {
3172 arc_cksum_free(hdr
);
3174 while (hdr
->b_l1hdr
.b_buf
!= NULL
)
3175 arc_buf_destroy_impl(hdr
->b_l1hdr
.b_buf
);
3178 if (hdr
->b_l1hdr
.b_thawed
!= NULL
) {
3179 kmem_free(hdr
->b_l1hdr
.b_thawed
, 1);
3180 hdr
->b_l1hdr
.b_thawed
= NULL
;
3184 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
3185 arc_hdr_free_pabd(hdr
);
3189 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
3190 if (HDR_HAS_L1HDR(hdr
)) {
3191 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3192 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
3193 kmem_cache_free(hdr_full_cache
, hdr
);
3195 kmem_cache_free(hdr_l2only_cache
, hdr
);
3200 arc_buf_destroy(arc_buf_t
*buf
, void* tag
)
3202 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3203 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
3205 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
3206 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
3207 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3208 VERIFY0(remove_reference(hdr
, NULL
, tag
));
3209 arc_hdr_destroy(hdr
);
3213 mutex_enter(hash_lock
);
3214 ASSERT3P(hdr
, ==, buf
->b_hdr
);
3215 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
3216 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3217 ASSERT3P(hdr
->b_l1hdr
.b_state
, !=, arc_anon
);
3218 ASSERT3P(buf
->b_data
, !=, NULL
);
3220 (void) remove_reference(hdr
, hash_lock
, tag
);
3221 arc_buf_destroy_impl(buf
);
3222 mutex_exit(hash_lock
);
3226 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3227 * state of the header is dependent on it's state prior to entering this
3228 * function. The following transitions are possible:
3230 * - arc_mru -> arc_mru_ghost
3231 * - arc_mfu -> arc_mfu_ghost
3232 * - arc_mru_ghost -> arc_l2c_only
3233 * - arc_mru_ghost -> deleted
3234 * - arc_mfu_ghost -> arc_l2c_only
3235 * - arc_mfu_ghost -> deleted
3238 arc_evict_hdr(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
3240 arc_state_t
*evicted_state
, *state
;
3241 int64_t bytes_evicted
= 0;
3243 ASSERT(MUTEX_HELD(hash_lock
));
3244 ASSERT(HDR_HAS_L1HDR(hdr
));
3246 state
= hdr
->b_l1hdr
.b_state
;
3247 if (GHOST_STATE(state
)) {
3248 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3249 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
3252 * l2arc_write_buffers() relies on a header's L1 portion
3253 * (i.e. its b_pabd field) during it's write phase.
3254 * Thus, we cannot push a header onto the arc_l2c_only
3255 * state (removing it's L1 piece) until the header is
3256 * done being written to the l2arc.
3258 if (HDR_HAS_L2HDR(hdr
) && HDR_L2_WRITING(hdr
)) {
3259 ARCSTAT_BUMP(arcstat_evict_l2_skip
);
3260 return (bytes_evicted
);
3263 ARCSTAT_BUMP(arcstat_deleted
);
3264 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3266 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, hdr
);
3268 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
3269 if (HDR_HAS_L2HDR(hdr
)) {
3271 * This buffer is cached on the 2nd Level ARC;
3272 * don't destroy the header.
3274 arc_change_state(arc_l2c_only
, hdr
, hash_lock
);
3276 * dropping from L1+L2 cached to L2-only,
3277 * realloc to remove the L1 header.
3279 hdr
= arc_hdr_realloc(hdr
, hdr_full_cache
,
3282 arc_change_state(arc_anon
, hdr
, hash_lock
);
3283 arc_hdr_destroy(hdr
);
3285 return (bytes_evicted
);
3288 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
3289 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
3291 /* prefetch buffers have a minimum lifespan */
3292 if (HDR_IO_IN_PROGRESS(hdr
) ||
3293 ((hdr
->b_flags
& (ARC_FLAG_PREFETCH
| ARC_FLAG_INDIRECT
)) &&
3294 ddi_get_lbolt() - hdr
->b_l1hdr
.b_arc_access
<
3295 arc_min_prefetch_lifespan
)) {
3296 ARCSTAT_BUMP(arcstat_evict_skip
);
3297 return (bytes_evicted
);
3300 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
3301 while (hdr
->b_l1hdr
.b_buf
) {
3302 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
3303 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
3304 ARCSTAT_BUMP(arcstat_mutex_miss
);
3307 if (buf
->b_data
!= NULL
)
3308 bytes_evicted
+= HDR_GET_LSIZE(hdr
);
3309 mutex_exit(&buf
->b_evict_lock
);
3310 arc_buf_destroy_impl(buf
);
3313 if (HDR_HAS_L2HDR(hdr
)) {
3314 ARCSTAT_INCR(arcstat_evict_l2_cached
, HDR_GET_LSIZE(hdr
));
3316 if (l2arc_write_eligible(hdr
->b_spa
, hdr
)) {
3317 ARCSTAT_INCR(arcstat_evict_l2_eligible
,
3318 HDR_GET_LSIZE(hdr
));
3320 ARCSTAT_INCR(arcstat_evict_l2_ineligible
,
3321 HDR_GET_LSIZE(hdr
));
3325 if (hdr
->b_l1hdr
.b_bufcnt
== 0) {
3326 arc_cksum_free(hdr
);
3328 bytes_evicted
+= arc_hdr_size(hdr
);
3331 * If this hdr is being evicted and has a compressed
3332 * buffer then we discard it here before we change states.
3333 * This ensures that the accounting is updated correctly
3334 * in arc_free_data_impl().
3336 arc_hdr_free_pabd(hdr
);
3338 arc_change_state(evicted_state
, hdr
, hash_lock
);
3339 ASSERT(HDR_IN_HASH_TABLE(hdr
));
3340 arc_hdr_set_flags(hdr
, ARC_FLAG_IN_HASH_TABLE
);
3341 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, hdr
);
3344 return (bytes_evicted
);
3348 arc_evict_state_impl(multilist_t
*ml
, int idx
, arc_buf_hdr_t
*marker
,
3349 uint64_t spa
, int64_t bytes
)
3351 multilist_sublist_t
*mls
;
3352 uint64_t bytes_evicted
= 0;
3354 kmutex_t
*hash_lock
;
3355 int evict_count
= 0;
3357 ASSERT3P(marker
, !=, NULL
);
3358 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3360 mls
= multilist_sublist_lock(ml
, idx
);
3362 for (hdr
= multilist_sublist_prev(mls
, marker
); hdr
!= NULL
;
3363 hdr
= multilist_sublist_prev(mls
, marker
)) {
3364 if ((bytes
!= ARC_EVICT_ALL
&& bytes_evicted
>= bytes
) ||
3365 (evict_count
>= zfs_arc_evict_batch_limit
))
3369 * To keep our iteration location, move the marker
3370 * forward. Since we're not holding hdr's hash lock, we
3371 * must be very careful and not remove 'hdr' from the
3372 * sublist. Otherwise, other consumers might mistake the
3373 * 'hdr' as not being on a sublist when they call the
3374 * multilist_link_active() function (they all rely on
3375 * the hash lock protecting concurrent insertions and
3376 * removals). multilist_sublist_move_forward() was
3377 * specifically implemented to ensure this is the case
3378 * (only 'marker' will be removed and re-inserted).
3380 multilist_sublist_move_forward(mls
, marker
);
3383 * The only case where the b_spa field should ever be
3384 * zero, is the marker headers inserted by
3385 * arc_evict_state(). It's possible for multiple threads
3386 * to be calling arc_evict_state() concurrently (e.g.
3387 * dsl_pool_close() and zio_inject_fault()), so we must
3388 * skip any markers we see from these other threads.
3390 if (hdr
->b_spa
== 0)
3393 /* we're only interested in evicting buffers of a certain spa */
3394 if (spa
!= 0 && hdr
->b_spa
!= spa
) {
3395 ARCSTAT_BUMP(arcstat_evict_skip
);
3399 hash_lock
= HDR_LOCK(hdr
);
3402 * We aren't calling this function from any code path
3403 * that would already be holding a hash lock, so we're
3404 * asserting on this assumption to be defensive in case
3405 * this ever changes. Without this check, it would be
3406 * possible to incorrectly increment arcstat_mutex_miss
3407 * below (e.g. if the code changed such that we called
3408 * this function with a hash lock held).
3410 ASSERT(!MUTEX_HELD(hash_lock
));
3412 if (mutex_tryenter(hash_lock
)) {
3413 uint64_t evicted
= arc_evict_hdr(hdr
, hash_lock
);
3414 mutex_exit(hash_lock
);
3416 bytes_evicted
+= evicted
;
3419 * If evicted is zero, arc_evict_hdr() must have
3420 * decided to skip this header, don't increment
3421 * evict_count in this case.
3427 * If arc_size isn't overflowing, signal any
3428 * threads that might happen to be waiting.
3430 * For each header evicted, we wake up a single
3431 * thread. If we used cv_broadcast, we could
3432 * wake up "too many" threads causing arc_size
3433 * to significantly overflow arc_c; since
3434 * arc_get_data_impl() doesn't check for overflow
3435 * when it's woken up (it doesn't because it's
3436 * possible for the ARC to be overflowing while
3437 * full of un-evictable buffers, and the
3438 * function should proceed in this case).
3440 * If threads are left sleeping, due to not
3441 * using cv_broadcast here, they will be woken
3442 * up via cv_broadcast in arc_adjust_cb() just
3443 * before arc_adjust_zthr sleeps.
3445 mutex_enter(&arc_adjust_lock
);
3446 if (!arc_is_overflowing())
3447 cv_signal(&arc_adjust_waiters_cv
);
3448 mutex_exit(&arc_adjust_lock
);
3450 ARCSTAT_BUMP(arcstat_mutex_miss
);
3454 multilist_sublist_unlock(mls
);
3456 return (bytes_evicted
);
3460 * Evict buffers from the given arc state, until we've removed the
3461 * specified number of bytes. Move the removed buffers to the
3462 * appropriate evict state.
3464 * This function makes a "best effort". It skips over any buffers
3465 * it can't get a hash_lock on, and so, may not catch all candidates.
3466 * It may also return without evicting as much space as requested.
3468 * If bytes is specified using the special value ARC_EVICT_ALL, this
3469 * will evict all available (i.e. unlocked and evictable) buffers from
3470 * the given arc state; which is used by arc_flush().
3473 arc_evict_state(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
3474 arc_buf_contents_t type
)
3476 uint64_t total_evicted
= 0;
3477 multilist_t
*ml
= state
->arcs_list
[type
];
3479 arc_buf_hdr_t
**markers
;
3481 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
3483 num_sublists
= multilist_get_num_sublists(ml
);
3486 * If we've tried to evict from each sublist, made some
3487 * progress, but still have not hit the target number of bytes
3488 * to evict, we want to keep trying. The markers allow us to
3489 * pick up where we left off for each individual sublist, rather
3490 * than starting from the tail each time.
3492 markers
= kmem_zalloc(sizeof (*markers
) * num_sublists
, KM_SLEEP
);
3493 for (int i
= 0; i
< num_sublists
; i
++) {
3494 markers
[i
] = kmem_cache_alloc(hdr_full_cache
, KM_SLEEP
);
3497 * A b_spa of 0 is used to indicate that this header is
3498 * a marker. This fact is used in arc_adjust_type() and
3499 * arc_evict_state_impl().
3501 markers
[i
]->b_spa
= 0;
3503 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
3504 multilist_sublist_insert_tail(mls
, markers
[i
]);
3505 multilist_sublist_unlock(mls
);
3509 * While we haven't hit our target number of bytes to evict, or
3510 * we're evicting all available buffers.
3512 while (total_evicted
< bytes
|| bytes
== ARC_EVICT_ALL
) {
3514 * Start eviction using a randomly selected sublist,
3515 * this is to try and evenly balance eviction across all
3516 * sublists. Always starting at the same sublist
3517 * (e.g. index 0) would cause evictions to favor certain
3518 * sublists over others.
3520 int sublist_idx
= multilist_get_random_index(ml
);
3521 uint64_t scan_evicted
= 0;
3523 for (int i
= 0; i
< num_sublists
; i
++) {
3524 uint64_t bytes_remaining
;
3525 uint64_t bytes_evicted
;
3527 if (bytes
== ARC_EVICT_ALL
)
3528 bytes_remaining
= ARC_EVICT_ALL
;
3529 else if (total_evicted
< bytes
)
3530 bytes_remaining
= bytes
- total_evicted
;
3534 bytes_evicted
= arc_evict_state_impl(ml
, sublist_idx
,
3535 markers
[sublist_idx
], spa
, bytes_remaining
);
3537 scan_evicted
+= bytes_evicted
;
3538 total_evicted
+= bytes_evicted
;
3540 /* we've reached the end, wrap to the beginning */
3541 if (++sublist_idx
>= num_sublists
)
3546 * If we didn't evict anything during this scan, we have
3547 * no reason to believe we'll evict more during another
3548 * scan, so break the loop.
3550 if (scan_evicted
== 0) {
3551 /* This isn't possible, let's make that obvious */
3552 ASSERT3S(bytes
, !=, 0);
3555 * When bytes is ARC_EVICT_ALL, the only way to
3556 * break the loop is when scan_evicted is zero.
3557 * In that case, we actually have evicted enough,
3558 * so we don't want to increment the kstat.
3560 if (bytes
!= ARC_EVICT_ALL
) {
3561 ASSERT3S(total_evicted
, <, bytes
);
3562 ARCSTAT_BUMP(arcstat_evict_not_enough
);
3569 for (int i
= 0; i
< num_sublists
; i
++) {
3570 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
3571 multilist_sublist_remove(mls
, markers
[i
]);
3572 multilist_sublist_unlock(mls
);
3574 kmem_cache_free(hdr_full_cache
, markers
[i
]);
3576 kmem_free(markers
, sizeof (*markers
) * num_sublists
);
3578 return (total_evicted
);
3582 * Flush all "evictable" data of the given type from the arc state
3583 * specified. This will not evict any "active" buffers (i.e. referenced).
3585 * When 'retry' is set to B_FALSE, the function will make a single pass
3586 * over the state and evict any buffers that it can. Since it doesn't
3587 * continually retry the eviction, it might end up leaving some buffers
3588 * in the ARC due to lock misses.
3590 * When 'retry' is set to B_TRUE, the function will continually retry the
3591 * eviction until *all* evictable buffers have been removed from the
3592 * state. As a result, if concurrent insertions into the state are
3593 * allowed (e.g. if the ARC isn't shutting down), this function might
3594 * wind up in an infinite loop, continually trying to evict buffers.
3597 arc_flush_state(arc_state_t
*state
, uint64_t spa
, arc_buf_contents_t type
,
3600 uint64_t evicted
= 0;
3602 while (refcount_count(&state
->arcs_esize
[type
]) != 0) {
3603 evicted
+= arc_evict_state(state
, spa
, ARC_EVICT_ALL
, type
);
3613 * Evict the specified number of bytes from the state specified,
3614 * restricting eviction to the spa and type given. This function
3615 * prevents us from trying to evict more from a state's list than
3616 * is "evictable", and to skip evicting altogether when passed a
3617 * negative value for "bytes". In contrast, arc_evict_state() will
3618 * evict everything it can, when passed a negative value for "bytes".
3621 arc_adjust_impl(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
3622 arc_buf_contents_t type
)
3626 if (bytes
> 0 && refcount_count(&state
->arcs_esize
[type
]) > 0) {
3627 delta
= MIN(refcount_count(&state
->arcs_esize
[type
]), bytes
);
3628 return (arc_evict_state(state
, spa
, delta
, type
));
3635 * Evict metadata buffers from the cache, such that arc_meta_used is
3636 * capped by the arc_meta_limit tunable.
3639 arc_adjust_meta(uint64_t meta_used
)
3641 uint64_t total_evicted
= 0;
3645 * If we're over the meta limit, we want to evict enough
3646 * metadata to get back under the meta limit. We don't want to
3647 * evict so much that we drop the MRU below arc_p, though. If
3648 * we're over the meta limit more than we're over arc_p, we
3649 * evict some from the MRU here, and some from the MFU below.
3651 target
= MIN((int64_t)(meta_used
- arc_meta_limit
),
3652 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
3653 refcount_count(&arc_mru
->arcs_size
) - arc_p
));
3655 total_evicted
+= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3658 * Similar to the above, we want to evict enough bytes to get us
3659 * below the meta limit, but not so much as to drop us below the
3660 * space allotted to the MFU (which is defined as arc_c - arc_p).
3662 target
= MIN((int64_t)(meta_used
- arc_meta_limit
),
3663 (int64_t)(refcount_count(&arc_mfu
->arcs_size
) -
3666 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3668 return (total_evicted
);
3672 * Return the type of the oldest buffer in the given arc state
3674 * This function will select a random sublist of type ARC_BUFC_DATA and
3675 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3676 * is compared, and the type which contains the "older" buffer will be
3679 static arc_buf_contents_t
3680 arc_adjust_type(arc_state_t
*state
)
3682 multilist_t
*data_ml
= state
->arcs_list
[ARC_BUFC_DATA
];
3683 multilist_t
*meta_ml
= state
->arcs_list
[ARC_BUFC_METADATA
];
3684 int data_idx
= multilist_get_random_index(data_ml
);
3685 int meta_idx
= multilist_get_random_index(meta_ml
);
3686 multilist_sublist_t
*data_mls
;
3687 multilist_sublist_t
*meta_mls
;
3688 arc_buf_contents_t type
;
3689 arc_buf_hdr_t
*data_hdr
;
3690 arc_buf_hdr_t
*meta_hdr
;
3693 * We keep the sublist lock until we're finished, to prevent
3694 * the headers from being destroyed via arc_evict_state().
3696 data_mls
= multilist_sublist_lock(data_ml
, data_idx
);
3697 meta_mls
= multilist_sublist_lock(meta_ml
, meta_idx
);
3700 * These two loops are to ensure we skip any markers that
3701 * might be at the tail of the lists due to arc_evict_state().
3704 for (data_hdr
= multilist_sublist_tail(data_mls
); data_hdr
!= NULL
;
3705 data_hdr
= multilist_sublist_prev(data_mls
, data_hdr
)) {
3706 if (data_hdr
->b_spa
!= 0)
3710 for (meta_hdr
= multilist_sublist_tail(meta_mls
); meta_hdr
!= NULL
;
3711 meta_hdr
= multilist_sublist_prev(meta_mls
, meta_hdr
)) {
3712 if (meta_hdr
->b_spa
!= 0)
3716 if (data_hdr
== NULL
&& meta_hdr
== NULL
) {
3717 type
= ARC_BUFC_DATA
;
3718 } else if (data_hdr
== NULL
) {
3719 ASSERT3P(meta_hdr
, !=, NULL
);
3720 type
= ARC_BUFC_METADATA
;
3721 } else if (meta_hdr
== NULL
) {
3722 ASSERT3P(data_hdr
, !=, NULL
);
3723 type
= ARC_BUFC_DATA
;
3725 ASSERT3P(data_hdr
, !=, NULL
);
3726 ASSERT3P(meta_hdr
, !=, NULL
);
3728 /* The headers can't be on the sublist without an L1 header */
3729 ASSERT(HDR_HAS_L1HDR(data_hdr
));
3730 ASSERT(HDR_HAS_L1HDR(meta_hdr
));
3732 if (data_hdr
->b_l1hdr
.b_arc_access
<
3733 meta_hdr
->b_l1hdr
.b_arc_access
) {
3734 type
= ARC_BUFC_DATA
;
3736 type
= ARC_BUFC_METADATA
;
3740 multilist_sublist_unlock(meta_mls
);
3741 multilist_sublist_unlock(data_mls
);
3747 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3752 uint64_t total_evicted
= 0;
3755 uint64_t asize
= aggsum_value(&arc_size
);
3756 uint64_t ameta
= aggsum_value(&arc_meta_used
);
3759 * If we're over arc_meta_limit, we want to correct that before
3760 * potentially evicting data buffers below.
3762 total_evicted
+= arc_adjust_meta(ameta
);
3767 * If we're over the target cache size, we want to evict enough
3768 * from the list to get back to our target size. We don't want
3769 * to evict too much from the MRU, such that it drops below
3770 * arc_p. So, if we're over our target cache size more than
3771 * the MRU is over arc_p, we'll evict enough to get back to
3772 * arc_p here, and then evict more from the MFU below.
3774 target
= MIN((int64_t)(asize
- arc_c
),
3775 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
3776 refcount_count(&arc_mru
->arcs_size
) + ameta
- arc_p
));
3779 * If we're below arc_meta_min, always prefer to evict data.
3780 * Otherwise, try to satisfy the requested number of bytes to
3781 * evict from the type which contains older buffers; in an
3782 * effort to keep newer buffers in the cache regardless of their
3783 * type. If we cannot satisfy the number of bytes from this
3784 * type, spill over into the next type.
3786 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
3787 ameta
> arc_meta_min
) {
3788 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3789 total_evicted
+= bytes
;
3792 * If we couldn't evict our target number of bytes from
3793 * metadata, we try to get the rest from data.
3798 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3800 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3801 total_evicted
+= bytes
;
3804 * If we couldn't evict our target number of bytes from
3805 * data, we try to get the rest from metadata.
3810 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3816 * Now that we've tried to evict enough from the MRU to get its
3817 * size back to arc_p, if we're still above the target cache
3818 * size, we evict the rest from the MFU.
3820 target
= asize
- arc_c
;
3822 if (arc_adjust_type(arc_mfu
) == ARC_BUFC_METADATA
&&
3823 ameta
> arc_meta_min
) {
3824 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3825 total_evicted
+= bytes
;
3828 * If we couldn't evict our target number of bytes from
3829 * metadata, we try to get the rest from data.
3834 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3836 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3837 total_evicted
+= bytes
;
3840 * If we couldn't evict our target number of bytes from
3841 * data, we try to get the rest from data.
3846 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3850 * Adjust ghost lists
3852 * In addition to the above, the ARC also defines target values
3853 * for the ghost lists. The sum of the mru list and mru ghost
3854 * list should never exceed the target size of the cache, and
3855 * the sum of the mru list, mfu list, mru ghost list, and mfu
3856 * ghost list should never exceed twice the target size of the
3857 * cache. The following logic enforces these limits on the ghost
3858 * caches, and evicts from them as needed.
3860 target
= refcount_count(&arc_mru
->arcs_size
) +
3861 refcount_count(&arc_mru_ghost
->arcs_size
) - arc_c
;
3863 bytes
= arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_DATA
);
3864 total_evicted
+= bytes
;
3869 arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_METADATA
);
3872 * We assume the sum of the mru list and mfu list is less than
3873 * or equal to arc_c (we enforced this above), which means we
3874 * can use the simpler of the two equations below:
3876 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3877 * mru ghost + mfu ghost <= arc_c
3879 target
= refcount_count(&arc_mru_ghost
->arcs_size
) +
3880 refcount_count(&arc_mfu_ghost
->arcs_size
) - arc_c
;
3882 bytes
= arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_DATA
);
3883 total_evicted
+= bytes
;
3888 arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_METADATA
);
3890 return (total_evicted
);
3894 arc_flush(spa_t
*spa
, boolean_t retry
)
3899 * If retry is B_TRUE, a spa must not be specified since we have
3900 * no good way to determine if all of a spa's buffers have been
3901 * evicted from an arc state.
3903 ASSERT(!retry
|| spa
== 0);
3906 guid
= spa_load_guid(spa
);
3908 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_DATA
, retry
);
3909 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_METADATA
, retry
);
3911 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_DATA
, retry
);
3912 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_METADATA
, retry
);
3914 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3915 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3917 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3918 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3922 arc_reduce_target_size(int64_t to_free
)
3924 uint64_t asize
= aggsum_value(&arc_size
);
3925 if (arc_c
> arc_c_min
) {
3927 if (arc_c
> arc_c_min
+ to_free
)
3928 atomic_add_64(&arc_c
, -to_free
);
3932 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
3934 arc_c
= MAX(asize
, arc_c_min
);
3936 arc_p
= (arc_c
>> 1);
3937 ASSERT(arc_c
>= arc_c_min
);
3938 ASSERT((int64_t)arc_p
>= 0);
3941 if (asize
> arc_c
) {
3942 /* See comment in arc_adjust_cb_check() on why lock+flag */
3943 mutex_enter(&arc_adjust_lock
);
3944 arc_adjust_needed
= B_TRUE
;
3945 mutex_exit(&arc_adjust_lock
);
3946 zthr_wakeup(arc_adjust_zthr
);
3950 typedef enum free_memory_reason_t
{
3955 FMR_PAGES_PP_MAXIMUM
,
3958 } free_memory_reason_t
;
3960 int64_t last_free_memory
;
3961 free_memory_reason_t last_free_reason
;
3964 * Additional reserve of pages for pp_reserve.
3966 int64_t arc_pages_pp_reserve
= 64;
3969 * Additional reserve of pages for swapfs.
3971 int64_t arc_swapfs_reserve
= 64;
3974 * Return the amount of memory that can be consumed before reclaim will be
3975 * needed. Positive if there is sufficient free memory, negative indicates
3976 * the amount of memory that needs to be freed up.
3979 arc_available_memory(void)
3981 int64_t lowest
= INT64_MAX
;
3983 free_memory_reason_t r
= FMR_UNKNOWN
;
3987 n
= PAGESIZE
* (-needfree
);
3995 * check that we're out of range of the pageout scanner. It starts to
3996 * schedule paging if freemem is less than lotsfree and needfree.
3997 * lotsfree is the high-water mark for pageout, and needfree is the
3998 * number of needed free pages. We add extra pages here to make sure
3999 * the scanner doesn't start up while we're freeing memory.
4001 n
= PAGESIZE
* (freemem
- lotsfree
- needfree
- desfree
);
4008 * check to make sure that swapfs has enough space so that anon
4009 * reservations can still succeed. anon_resvmem() checks that the
4010 * availrmem is greater than swapfs_minfree, and the number of reserved
4011 * swap pages. We also add a bit of extra here just to prevent
4012 * circumstances from getting really dire.
4014 n
= PAGESIZE
* (availrmem
- swapfs_minfree
- swapfs_reserve
-
4015 desfree
- arc_swapfs_reserve
);
4018 r
= FMR_SWAPFS_MINFREE
;
4023 * Check that we have enough availrmem that memory locking (e.g., via
4024 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4025 * stores the number of pages that cannot be locked; when availrmem
4026 * drops below pages_pp_maximum, page locking mechanisms such as
4027 * page_pp_lock() will fail.)
4029 n
= PAGESIZE
* (availrmem
- pages_pp_maximum
-
4030 arc_pages_pp_reserve
);
4033 r
= FMR_PAGES_PP_MAXIMUM
;
4038 * If we're on an i386 platform, it's possible that we'll exhaust the
4039 * kernel heap space before we ever run out of available physical
4040 * memory. Most checks of the size of the heap_area compare against
4041 * tune.t_minarmem, which is the minimum available real memory that we
4042 * can have in the system. However, this is generally fixed at 25 pages
4043 * which is so low that it's useless. In this comparison, we seek to
4044 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4045 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4048 n
= (int64_t)vmem_size(heap_arena
, VMEM_FREE
) -
4049 (vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
) >> 2);
4057 * If zio data pages are being allocated out of a separate heap segment,
4058 * then enforce that the size of available vmem for this arena remains
4059 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4061 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4062 * memory (in the zio_arena) free, which can avoid memory
4063 * fragmentation issues.
4065 if (zio_arena
!= NULL
) {
4066 n
= (int64_t)vmem_size(zio_arena
, VMEM_FREE
) -
4067 (vmem_size(zio_arena
, VMEM_ALLOC
) >>
4068 arc_zio_arena_free_shift
);
4075 /* Every 100 calls, free a small amount */
4076 if (spa_get_random(100) == 0)
4080 last_free_memory
= lowest
;
4081 last_free_reason
= r
;
4088 * Determine if the system is under memory pressure and is asking
4089 * to reclaim memory. A return value of B_TRUE indicates that the system
4090 * is under memory pressure and that the arc should adjust accordingly.
4093 arc_reclaim_needed(void)
4095 return (arc_available_memory() < 0);
4099 arc_kmem_reap_soon(void)
4102 kmem_cache_t
*prev_cache
= NULL
;
4103 kmem_cache_t
*prev_data_cache
= NULL
;
4104 extern kmem_cache_t
*zio_buf_cache
[];
4105 extern kmem_cache_t
*zio_data_buf_cache
[];
4106 extern kmem_cache_t
*range_seg_cache
;
4107 extern kmem_cache_t
*abd_chunk_cache
;
4110 if (aggsum_compare(&arc_meta_used
, arc_meta_limit
) >= 0) {
4112 * We are exceeding our meta-data cache limit.
4113 * Purge some DNLC entries to release holds on meta-data.
4115 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent
);
4119 * Reclaim unused memory from all kmem caches.
4125 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
4126 if (zio_buf_cache
[i
] != prev_cache
) {
4127 prev_cache
= zio_buf_cache
[i
];
4128 kmem_cache_reap_soon(zio_buf_cache
[i
]);
4130 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
4131 prev_data_cache
= zio_data_buf_cache
[i
];
4132 kmem_cache_reap_soon(zio_data_buf_cache
[i
]);
4135 kmem_cache_reap_soon(abd_chunk_cache
);
4136 kmem_cache_reap_soon(buf_cache
);
4137 kmem_cache_reap_soon(hdr_full_cache
);
4138 kmem_cache_reap_soon(hdr_l2only_cache
);
4139 kmem_cache_reap_soon(range_seg_cache
);
4141 if (zio_arena
!= NULL
) {
4143 * Ask the vmem arena to reclaim unused memory from its
4146 vmem_qcache_reap(zio_arena
);
4152 arc_adjust_cb_check(void *arg
, zthr_t
*zthr
)
4155 * This is necessary in order for the mdb ::arc dcmd to
4156 * show up to date information. Since the ::arc command
4157 * does not call the kstat's update function, without
4158 * this call, the command may show stale stats for the
4159 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4160 * with this change, the data might be up to 1 second
4161 * out of date(the arc_adjust_zthr has a maximum sleep
4162 * time of 1 second); but that should suffice. The
4163 * arc_state_t structures can be queried directly if more
4164 * accurate information is needed.
4166 if (arc_ksp
!= NULL
)
4167 arc_ksp
->ks_update(arc_ksp
, KSTAT_READ
);
4170 * We have to rely on arc_get_data_impl() to tell us when to adjust,
4171 * rather than checking if we are overflowing here, so that we are
4172 * sure to not leave arc_get_data_impl() waiting on
4173 * arc_adjust_waiters_cv. If we have become "not overflowing" since
4174 * arc_get_data_impl() checked, we need to wake it up. We could
4175 * broadcast the CV here, but arc_get_data_impl() may have not yet
4176 * gone to sleep. We would need to use a mutex to ensure that this
4177 * function doesn't broadcast until arc_get_data_impl() has gone to
4178 * sleep (e.g. the arc_adjust_lock). However, the lock ordering of
4179 * such a lock would necessarily be incorrect with respect to the
4180 * zthr_lock, which is held before this function is called, and is
4181 * held by arc_get_data_impl() when it calls zthr_wakeup().
4183 return (arc_adjust_needed
);
4187 * Keep arc_size under arc_c by running arc_adjust which evicts data
4192 arc_adjust_cb(void *arg
, zthr_t
*zthr
)
4194 uint64_t evicted
= 0;
4196 /* Evict from cache */
4197 evicted
= arc_adjust();
4200 * If evicted is zero, we couldn't evict anything
4201 * via arc_adjust(). This could be due to hash lock
4202 * collisions, but more likely due to the majority of
4203 * arc buffers being unevictable. Therefore, even if
4204 * arc_size is above arc_c, another pass is unlikely to
4205 * be helpful and could potentially cause us to enter an
4206 * infinite loop. Additionally, zthr_iscancelled() is
4207 * checked here so that if the arc is shutting down, the
4208 * broadcast will wake any remaining arc adjust waiters.
4210 mutex_enter(&arc_adjust_lock
);
4211 arc_adjust_needed
= !zthr_iscancelled(arc_adjust_zthr
) &&
4212 evicted
> 0 && aggsum_compare(&arc_size
, arc_c
) > 0;
4213 if (!arc_adjust_needed
) {
4215 * We're either no longer overflowing, or we
4216 * can't evict anything more, so we should wake
4219 cv_broadcast(&arc_adjust_waiters_cv
);
4221 mutex_exit(&arc_adjust_lock
);
4228 arc_reap_cb_check(void *arg
, zthr_t
*zthr
)
4230 int64_t free_memory
= arc_available_memory();
4233 * If a kmem reap is already active, don't schedule more. We must
4234 * check for this because kmem_cache_reap_soon() won't actually
4235 * block on the cache being reaped (this is to prevent callers from
4236 * becoming implicitly blocked by a system-wide kmem reap -- which,
4237 * on a system with many, many full magazines, can take minutes).
4239 if (!kmem_cache_reap_active() &&
4241 arc_no_grow
= B_TRUE
;
4244 * Wait at least zfs_grow_retry (default 60) seconds
4245 * before considering growing.
4247 arc_growtime
= gethrtime() + SEC2NSEC(arc_grow_retry
);
4249 } else if (free_memory
< arc_c
>> arc_no_grow_shift
) {
4250 arc_no_grow
= B_TRUE
;
4251 } else if (gethrtime() >= arc_growtime
) {
4252 arc_no_grow
= B_FALSE
;
4259 * Keep enough free memory in the system by reaping the ARC's kmem
4260 * caches. To cause more slabs to be reapable, we may reduce the
4261 * target size of the cache (arc_c), causing the arc_adjust_cb()
4262 * to free more buffers.
4266 arc_reap_cb(void *arg
, zthr_t
*zthr
)
4268 int64_t free_memory
;
4271 * Kick off asynchronous kmem_reap()'s of all our caches.
4273 arc_kmem_reap_soon();
4276 * Wait at least arc_kmem_cache_reap_retry_ms between
4277 * arc_kmem_reap_soon() calls. Without this check it is possible to
4278 * end up in a situation where we spend lots of time reaping
4279 * caches, while we're near arc_c_min. Waiting here also gives the
4280 * subsequent free memory check a chance of finding that the
4281 * asynchronous reap has already freed enough memory, and we don't
4282 * need to call arc_reduce_target_size().
4284 delay((hz
* arc_kmem_cache_reap_retry_ms
+ 999) / 1000);
4287 * Reduce the target size as needed to maintain the amount of free
4288 * memory in the system at a fraction of the arc_size (1/128th by
4289 * default). If oversubscribed (free_memory < 0) then reduce the
4290 * target arc_size by the deficit amount plus the fractional
4291 * amount. If free memory is positive but less then the fractional
4292 * amount, reduce by what is needed to hit the fractional amount.
4294 free_memory
= arc_available_memory();
4297 (arc_c
>> arc_shrink_shift
) - free_memory
;
4300 to_free
= MAX(to_free
, ptob(needfree
));
4302 arc_reduce_target_size(to_free
);
4309 * Adapt arc info given the number of bytes we are trying to add and
4310 * the state that we are comming from. This function is only called
4311 * when we are adding new content to the cache.
4314 arc_adapt(int bytes
, arc_state_t
*state
)
4317 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
4318 int64_t mrug_size
= refcount_count(&arc_mru_ghost
->arcs_size
);
4319 int64_t mfug_size
= refcount_count(&arc_mfu_ghost
->arcs_size
);
4321 if (state
== arc_l2c_only
)
4326 * Adapt the target size of the MRU list:
4327 * - if we just hit in the MRU ghost list, then increase
4328 * the target size of the MRU list.
4329 * - if we just hit in the MFU ghost list, then increase
4330 * the target size of the MFU list by decreasing the
4331 * target size of the MRU list.
4333 if (state
== arc_mru_ghost
) {
4334 mult
= (mrug_size
>= mfug_size
) ? 1 : (mfug_size
/ mrug_size
);
4335 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
4337 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
4338 } else if (state
== arc_mfu_ghost
) {
4341 mult
= (mfug_size
>= mrug_size
) ? 1 : (mrug_size
/ mfug_size
);
4342 mult
= MIN(mult
, 10);
4344 delta
= MIN(bytes
* mult
, arc_p
);
4345 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
4347 ASSERT((int64_t)arc_p
>= 0);
4350 * Wake reap thread if we do not have any available memory
4352 if (arc_reclaim_needed()) {
4353 zthr_wakeup(arc_reap_zthr
);
4361 if (arc_c
>= arc_c_max
)
4365 * If we're within (2 * maxblocksize) bytes of the target
4366 * cache size, increment the target cache size
4368 if (aggsum_compare(&arc_size
, arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) >
4370 atomic_add_64(&arc_c
, (int64_t)bytes
);
4371 if (arc_c
> arc_c_max
)
4373 else if (state
== arc_anon
)
4374 atomic_add_64(&arc_p
, (int64_t)bytes
);
4378 ASSERT((int64_t)arc_p
>= 0);
4382 * Check if arc_size has grown past our upper threshold, determined by
4383 * zfs_arc_overflow_shift.
4386 arc_is_overflowing(void)
4388 /* Always allow at least one block of overflow */
4389 uint64_t overflow
= MAX(SPA_MAXBLOCKSIZE
,
4390 arc_c
>> zfs_arc_overflow_shift
);
4393 * We just compare the lower bound here for performance reasons. Our
4394 * primary goals are to make sure that the arc never grows without
4395 * bound, and that it can reach its maximum size. This check
4396 * accomplishes both goals. The maximum amount we could run over by is
4397 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4398 * in the ARC. In practice, that's in the tens of MB, which is low
4399 * enough to be safe.
4401 return (aggsum_lower_bound(&arc_size
) >= arc_c
+ overflow
);
4405 arc_get_data_abd(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4407 arc_buf_contents_t type
= arc_buf_type(hdr
);
4409 arc_get_data_impl(hdr
, size
, tag
);
4410 if (type
== ARC_BUFC_METADATA
) {
4411 return (abd_alloc(size
, B_TRUE
));
4413 ASSERT(type
== ARC_BUFC_DATA
);
4414 return (abd_alloc(size
, B_FALSE
));
4419 arc_get_data_buf(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4421 arc_buf_contents_t type
= arc_buf_type(hdr
);
4423 arc_get_data_impl(hdr
, size
, tag
);
4424 if (type
== ARC_BUFC_METADATA
) {
4425 return (zio_buf_alloc(size
));
4427 ASSERT(type
== ARC_BUFC_DATA
);
4428 return (zio_data_buf_alloc(size
));
4433 * Allocate a block and return it to the caller. If we are hitting the
4434 * hard limit for the cache size, we must sleep, waiting for the eviction
4435 * thread to catch up. If we're past the target size but below the hard
4436 * limit, we'll only signal the reclaim thread and continue on.
4439 arc_get_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4441 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
4442 arc_buf_contents_t type
= arc_buf_type(hdr
);
4444 arc_adapt(size
, state
);
4447 * If arc_size is currently overflowing, and has grown past our
4448 * upper limit, we must be adding data faster than the evict
4449 * thread can evict. Thus, to ensure we don't compound the
4450 * problem by adding more data and forcing arc_size to grow even
4451 * further past it's target size, we halt and wait for the
4452 * eviction thread to catch up.
4454 * It's also possible that the reclaim thread is unable to evict
4455 * enough buffers to get arc_size below the overflow limit (e.g.
4456 * due to buffers being un-evictable, or hash lock collisions).
4457 * In this case, we want to proceed regardless if we're
4458 * overflowing; thus we don't use a while loop here.
4460 if (arc_is_overflowing()) {
4461 mutex_enter(&arc_adjust_lock
);
4464 * Now that we've acquired the lock, we may no longer be
4465 * over the overflow limit, lets check.
4467 * We're ignoring the case of spurious wake ups. If that
4468 * were to happen, it'd let this thread consume an ARC
4469 * buffer before it should have (i.e. before we're under
4470 * the overflow limit and were signalled by the reclaim
4471 * thread). As long as that is a rare occurrence, it
4472 * shouldn't cause any harm.
4474 if (arc_is_overflowing()) {
4475 arc_adjust_needed
= B_TRUE
;
4476 zthr_wakeup(arc_adjust_zthr
);
4477 (void) cv_wait(&arc_adjust_waiters_cv
,
4480 mutex_exit(&arc_adjust_lock
);
4483 VERIFY3U(hdr
->b_type
, ==, type
);
4484 if (type
== ARC_BUFC_METADATA
) {
4485 arc_space_consume(size
, ARC_SPACE_META
);
4487 arc_space_consume(size
, ARC_SPACE_DATA
);
4491 * Update the state size. Note that ghost states have a
4492 * "ghost size" and so don't need to be updated.
4494 if (!GHOST_STATE(state
)) {
4496 (void) refcount_add_many(&state
->arcs_size
, size
, tag
);
4499 * If this is reached via arc_read, the link is
4500 * protected by the hash lock. If reached via
4501 * arc_buf_alloc, the header should not be accessed by
4502 * any other thread. And, if reached via arc_read_done,
4503 * the hash lock will protect it if it's found in the
4504 * hash table; otherwise no other thread should be
4505 * trying to [add|remove]_reference it.
4507 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
4508 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4509 (void) refcount_add_many(&state
->arcs_esize
[type
],
4514 * If we are growing the cache, and we are adding anonymous
4515 * data, and we have outgrown arc_p, update arc_p
4517 if (aggsum_compare(&arc_size
, arc_c
) < 0 &&
4518 hdr
->b_l1hdr
.b_state
== arc_anon
&&
4519 (refcount_count(&arc_anon
->arcs_size
) +
4520 refcount_count(&arc_mru
->arcs_size
) > arc_p
))
4521 arc_p
= MIN(arc_c
, arc_p
+ size
);
4526 arc_free_data_abd(arc_buf_hdr_t
*hdr
, abd_t
*abd
, uint64_t size
, void *tag
)
4528 arc_free_data_impl(hdr
, size
, tag
);
4533 arc_free_data_buf(arc_buf_hdr_t
*hdr
, void *buf
, uint64_t size
, void *tag
)
4535 arc_buf_contents_t type
= arc_buf_type(hdr
);
4537 arc_free_data_impl(hdr
, size
, tag
);
4538 if (type
== ARC_BUFC_METADATA
) {
4539 zio_buf_free(buf
, size
);
4541 ASSERT(type
== ARC_BUFC_DATA
);
4542 zio_data_buf_free(buf
, size
);
4547 * Free the arc data buffer.
4550 arc_free_data_impl(arc_buf_hdr_t
*hdr
, uint64_t size
, void *tag
)
4552 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
4553 arc_buf_contents_t type
= arc_buf_type(hdr
);
4555 /* protected by hash lock, if in the hash table */
4556 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
4557 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4558 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
4560 (void) refcount_remove_many(&state
->arcs_esize
[type
],
4563 (void) refcount_remove_many(&state
->arcs_size
, size
, tag
);
4565 VERIFY3U(hdr
->b_type
, ==, type
);
4566 if (type
== ARC_BUFC_METADATA
) {
4567 arc_space_return(size
, ARC_SPACE_META
);
4569 ASSERT(type
== ARC_BUFC_DATA
);
4570 arc_space_return(size
, ARC_SPACE_DATA
);
4575 * This routine is called whenever a buffer is accessed.
4576 * NOTE: the hash lock is dropped in this function.
4579 arc_access(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
4583 ASSERT(MUTEX_HELD(hash_lock
));
4584 ASSERT(HDR_HAS_L1HDR(hdr
));
4586 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
4588 * This buffer is not in the cache, and does not
4589 * appear in our "ghost" list. Add the new buffer
4593 ASSERT0(hdr
->b_l1hdr
.b_arc_access
);
4594 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4595 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
4596 arc_change_state(arc_mru
, hdr
, hash_lock
);
4598 } else if (hdr
->b_l1hdr
.b_state
== arc_mru
) {
4599 now
= ddi_get_lbolt();
4602 * If this buffer is here because of a prefetch, then either:
4603 * - clear the flag if this is a "referencing" read
4604 * (any subsequent access will bump this into the MFU state).
4606 * - move the buffer to the head of the list if this is
4607 * another prefetch (to make it less likely to be evicted).
4609 if (HDR_PREFETCH(hdr
)) {
4610 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
4611 /* link protected by hash lock */
4612 ASSERT(multilist_link_active(
4613 &hdr
->b_l1hdr
.b_arc_node
));
4615 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
4616 ARCSTAT_BUMP(arcstat_mru_hits
);
4618 hdr
->b_l1hdr
.b_arc_access
= now
;
4623 * This buffer has been "accessed" only once so far,
4624 * but it is still in the cache. Move it to the MFU
4627 if (now
> hdr
->b_l1hdr
.b_arc_access
+ ARC_MINTIME
) {
4629 * More than 125ms have passed since we
4630 * instantiated this buffer. Move it to the
4631 * most frequently used state.
4633 hdr
->b_l1hdr
.b_arc_access
= now
;
4634 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4635 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4637 ARCSTAT_BUMP(arcstat_mru_hits
);
4638 } else if (hdr
->b_l1hdr
.b_state
== arc_mru_ghost
) {
4639 arc_state_t
*new_state
;
4641 * This buffer has been "accessed" recently, but
4642 * was evicted from the cache. Move it to the
4646 if (HDR_PREFETCH(hdr
)) {
4647 new_state
= arc_mru
;
4648 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0)
4649 arc_hdr_clear_flags(hdr
, ARC_FLAG_PREFETCH
);
4650 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
4652 new_state
= arc_mfu
;
4653 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4656 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4657 arc_change_state(new_state
, hdr
, hash_lock
);
4659 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
4660 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu
) {
4662 * This buffer has been accessed more than once and is
4663 * still in the cache. Keep it in the MFU state.
4665 * NOTE: an add_reference() that occurred when we did
4666 * the arc_read() will have kicked this off the list.
4667 * If it was a prefetch, we will explicitly move it to
4668 * the head of the list now.
4670 if ((HDR_PREFETCH(hdr
)) != 0) {
4671 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4672 /* link protected by hash_lock */
4673 ASSERT(multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4675 ARCSTAT_BUMP(arcstat_mfu_hits
);
4676 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4677 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu_ghost
) {
4678 arc_state_t
*new_state
= arc_mfu
;
4680 * This buffer has been accessed more than once but has
4681 * been evicted from the cache. Move it back to the
4685 if (HDR_PREFETCH(hdr
)) {
4687 * This is a prefetch access...
4688 * move this block back to the MRU state.
4690 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
4691 new_state
= arc_mru
;
4694 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4695 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4696 arc_change_state(new_state
, hdr
, hash_lock
);
4698 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
4699 } else if (hdr
->b_l1hdr
.b_state
== arc_l2c_only
) {
4701 * This buffer is on the 2nd Level ARC.
4704 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4705 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4706 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4708 ASSERT(!"invalid arc state");
4712 /* a generic arc_done_func_t which you can use */
4715 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4717 if (zio
== NULL
|| zio
->io_error
== 0)
4718 bcopy(buf
->b_data
, arg
, arc_buf_size(buf
));
4719 arc_buf_destroy(buf
, arg
);
4722 /* a generic arc_done_func_t */
4724 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4726 arc_buf_t
**bufp
= arg
;
4728 ASSERT(zio
== NULL
|| zio
->io_error
!= 0);
4731 ASSERT(zio
== NULL
|| zio
->io_error
== 0);
4733 ASSERT(buf
->b_data
!= NULL
);
4738 arc_hdr_verify(arc_buf_hdr_t
*hdr
, blkptr_t
*bp
)
4740 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
)) {
4741 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, 0);
4742 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==, ZIO_COMPRESS_OFF
);
4744 if (HDR_COMPRESSION_ENABLED(hdr
)) {
4745 ASSERT3U(HDR_GET_COMPRESS(hdr
), ==,
4746 BP_GET_COMPRESS(bp
));
4748 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(bp
));
4749 ASSERT3U(HDR_GET_PSIZE(hdr
), ==, BP_GET_PSIZE(bp
));
4754 arc_read_done(zio_t
*zio
)
4756 arc_buf_hdr_t
*hdr
= zio
->io_private
;
4757 kmutex_t
*hash_lock
= NULL
;
4758 arc_callback_t
*callback_list
;
4759 arc_callback_t
*acb
;
4760 boolean_t freeable
= B_FALSE
;
4761 boolean_t no_zio_error
= (zio
->io_error
== 0);
4764 * The hdr was inserted into hash-table and removed from lists
4765 * prior to starting I/O. We should find this header, since
4766 * it's in the hash table, and it should be legit since it's
4767 * not possible to evict it during the I/O. The only possible
4768 * reason for it not to be found is if we were freed during the
4771 if (HDR_IN_HASH_TABLE(hdr
)) {
4772 ASSERT3U(hdr
->b_birth
, ==, BP_PHYSICAL_BIRTH(zio
->io_bp
));
4773 ASSERT3U(hdr
->b_dva
.dva_word
[0], ==,
4774 BP_IDENTITY(zio
->io_bp
)->dva_word
[0]);
4775 ASSERT3U(hdr
->b_dva
.dva_word
[1], ==,
4776 BP_IDENTITY(zio
->io_bp
)->dva_word
[1]);
4778 arc_buf_hdr_t
*found
= buf_hash_find(hdr
->b_spa
, zio
->io_bp
,
4781 ASSERT((found
== hdr
&&
4782 DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
4783 (found
== hdr
&& HDR_L2_READING(hdr
)));
4784 ASSERT3P(hash_lock
, !=, NULL
);
4788 /* byteswap if necessary */
4789 if (BP_SHOULD_BYTESWAP(zio
->io_bp
)) {
4790 if (BP_GET_LEVEL(zio
->io_bp
) > 0) {
4791 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_UINT64
;
4793 hdr
->b_l1hdr
.b_byteswap
=
4794 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
4797 hdr
->b_l1hdr
.b_byteswap
= DMU_BSWAP_NUMFUNCS
;
4801 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_EVICTED
);
4802 if (l2arc_noprefetch
&& HDR_PREFETCH(hdr
))
4803 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2CACHE
);
4805 callback_list
= hdr
->b_l1hdr
.b_acb
;
4806 ASSERT3P(callback_list
, !=, NULL
);
4808 if (hash_lock
&& no_zio_error
&& hdr
->b_l1hdr
.b_state
== arc_anon
) {
4810 * Only call arc_access on anonymous buffers. This is because
4811 * if we've issued an I/O for an evicted buffer, we've already
4812 * called arc_access (to prevent any simultaneous readers from
4813 * getting confused).
4815 arc_access(hdr
, hash_lock
);
4819 * If a read request has a callback (i.e. acb_done is not NULL), then we
4820 * make a buf containing the data according to the parameters which were
4821 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
4822 * aren't needlessly decompressing the data multiple times.
4824 int callback_cnt
= 0;
4825 for (acb
= callback_list
; acb
!= NULL
; acb
= acb
->acb_next
) {
4829 /* This is a demand read since prefetches don't use callbacks */
4833 int error
= arc_buf_alloc_impl(hdr
, acb
->acb_private
,
4834 acb
->acb_compressed
, zio
->io_error
== 0,
4838 * Decompression failed. Set io_error
4839 * so that when we call acb_done (below),
4840 * we will indicate that the read failed.
4841 * Note that in the unusual case where one
4842 * callback is compressed and another
4843 * uncompressed, we will mark all of them
4844 * as failed, even though the uncompressed
4845 * one can't actually fail. In this case,
4846 * the hdr will not be anonymous, because
4847 * if there are multiple callbacks, it's
4848 * because multiple threads found the same
4849 * arc buf in the hash table.
4851 zio
->io_error
= error
;
4856 * If there are multiple callbacks, we must have the hash lock,
4857 * because the only way for multiple threads to find this hdr is
4858 * in the hash table. This ensures that if there are multiple
4859 * callbacks, the hdr is not anonymous. If it were anonymous,
4860 * we couldn't use arc_buf_destroy() in the error case below.
4862 ASSERT(callback_cnt
< 2 || hash_lock
!= NULL
);
4864 hdr
->b_l1hdr
.b_acb
= NULL
;
4865 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
4866 if (callback_cnt
== 0) {
4867 ASSERT(HDR_PREFETCH(hdr
));
4868 ASSERT0(hdr
->b_l1hdr
.b_bufcnt
);
4869 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
4872 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
) ||
4873 callback_list
!= NULL
);
4876 arc_hdr_verify(hdr
, zio
->io_bp
);
4878 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_ERROR
);
4879 if (hdr
->b_l1hdr
.b_state
!= arc_anon
)
4880 arc_change_state(arc_anon
, hdr
, hash_lock
);
4881 if (HDR_IN_HASH_TABLE(hdr
))
4882 buf_hash_remove(hdr
);
4883 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
4887 * Broadcast before we drop the hash_lock to avoid the possibility
4888 * that the hdr (and hence the cv) might be freed before we get to
4889 * the cv_broadcast().
4891 cv_broadcast(&hdr
->b_l1hdr
.b_cv
);
4893 if (hash_lock
!= NULL
) {
4894 mutex_exit(hash_lock
);
4897 * This block was freed while we waited for the read to
4898 * complete. It has been removed from the hash table and
4899 * moved to the anonymous state (so that it won't show up
4902 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
4903 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
4906 /* execute each callback and free its structure */
4907 while ((acb
= callback_list
) != NULL
) {
4908 if (acb
->acb_done
!= NULL
) {
4909 if (zio
->io_error
!= 0 && acb
->acb_buf
!= NULL
) {
4911 * If arc_buf_alloc_impl() fails during
4912 * decompression, the buf will still be
4913 * allocated, and needs to be freed here.
4915 arc_buf_destroy(acb
->acb_buf
, acb
->acb_private
);
4916 acb
->acb_buf
= NULL
;
4918 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
4921 if (acb
->acb_zio_dummy
!= NULL
) {
4922 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
4923 zio_nowait(acb
->acb_zio_dummy
);
4926 callback_list
= acb
->acb_next
;
4927 kmem_free(acb
, sizeof (arc_callback_t
));
4931 arc_hdr_destroy(hdr
);
4935 * "Read" the block at the specified DVA (in bp) via the
4936 * cache. If the block is found in the cache, invoke the provided
4937 * callback immediately and return. Note that the `zio' parameter
4938 * in the callback will be NULL in this case, since no IO was
4939 * required. If the block is not in the cache pass the read request
4940 * on to the spa with a substitute callback function, so that the
4941 * requested block will be added to the cache.
4943 * If a read request arrives for a block that has a read in-progress,
4944 * either wait for the in-progress read to complete (and return the
4945 * results); or, if this is a read with a "done" func, add a record
4946 * to the read to invoke the "done" func when the read completes,
4947 * and return; or just return.
4949 * arc_read_done() will invoke all the requested "done" functions
4950 * for readers of this block.
4953 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_done_func_t
*done
,
4954 void *private, zio_priority_t priority
, int zio_flags
,
4955 arc_flags_t
*arc_flags
, const zbookmark_phys_t
*zb
)
4957 arc_buf_hdr_t
*hdr
= NULL
;
4958 kmutex_t
*hash_lock
= NULL
;
4960 uint64_t guid
= spa_load_guid(spa
);
4961 boolean_t compressed_read
= (zio_flags
& ZIO_FLAG_RAW
) != 0;
4963 ASSERT(!BP_IS_EMBEDDED(bp
) ||
4964 BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
);
4967 if (!BP_IS_EMBEDDED(bp
)) {
4969 * Embedded BP's have no DVA and require no I/O to "read".
4970 * Create an anonymous arc buf to back it.
4972 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
4975 if (hdr
!= NULL
&& HDR_HAS_L1HDR(hdr
) && hdr
->b_l1hdr
.b_pabd
!= NULL
) {
4976 arc_buf_t
*buf
= NULL
;
4977 *arc_flags
|= ARC_FLAG_CACHED
;
4979 if (HDR_IO_IN_PROGRESS(hdr
)) {
4981 if ((hdr
->b_flags
& ARC_FLAG_PRIO_ASYNC_READ
) &&
4982 priority
== ZIO_PRIORITY_SYNC_READ
) {
4984 * This sync read must wait for an
4985 * in-progress async read (e.g. a predictive
4986 * prefetch). Async reads are queued
4987 * separately at the vdev_queue layer, so
4988 * this is a form of priority inversion.
4989 * Ideally, we would "inherit" the demand
4990 * i/o's priority by moving the i/o from
4991 * the async queue to the synchronous queue,
4992 * but there is currently no mechanism to do
4993 * so. Track this so that we can evaluate
4994 * the magnitude of this potential performance
4997 * Note that if the prefetch i/o is already
4998 * active (has been issued to the device),
4999 * the prefetch improved performance, because
5000 * we issued it sooner than we would have
5001 * without the prefetch.
5003 DTRACE_PROBE1(arc__sync__wait__for__async
,
5004 arc_buf_hdr_t
*, hdr
);
5005 ARCSTAT_BUMP(arcstat_sync_wait_for_async
);
5007 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5008 arc_hdr_clear_flags(hdr
,
5009 ARC_FLAG_PREDICTIVE_PREFETCH
);
5012 if (*arc_flags
& ARC_FLAG_WAIT
) {
5013 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
5014 mutex_exit(hash_lock
);
5017 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5020 arc_callback_t
*acb
= NULL
;
5022 acb
= kmem_zalloc(sizeof (arc_callback_t
),
5024 acb
->acb_done
= done
;
5025 acb
->acb_private
= private;
5026 acb
->acb_compressed
= compressed_read
;
5028 acb
->acb_zio_dummy
= zio_null(pio
,
5029 spa
, NULL
, NULL
, NULL
, zio_flags
);
5031 ASSERT3P(acb
->acb_done
, !=, NULL
);
5032 acb
->acb_next
= hdr
->b_l1hdr
.b_acb
;
5033 hdr
->b_l1hdr
.b_acb
= acb
;
5034 mutex_exit(hash_lock
);
5037 mutex_exit(hash_lock
);
5041 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
5042 hdr
->b_l1hdr
.b_state
== arc_mfu
);
5045 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
5047 * This is a demand read which does not have to
5048 * wait for i/o because we did a predictive
5049 * prefetch i/o for it, which has completed.
5052 arc__demand__hit__predictive__prefetch
,
5053 arc_buf_hdr_t
*, hdr
);
5055 arcstat_demand_hit_predictive_prefetch
);
5056 arc_hdr_clear_flags(hdr
,
5057 ARC_FLAG_PREDICTIVE_PREFETCH
);
5059 ASSERT(!BP_IS_EMBEDDED(bp
) || !BP_IS_HOLE(bp
));
5061 /* Get a buf with the desired data in it. */
5062 VERIFY0(arc_buf_alloc_impl(hdr
, private,
5063 compressed_read
, B_TRUE
, &buf
));
5064 } else if (*arc_flags
& ARC_FLAG_PREFETCH
&&
5065 refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
5066 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
5068 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
5069 arc_access(hdr
, hash_lock
);
5070 if (*arc_flags
& ARC_FLAG_L2CACHE
)
5071 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5072 mutex_exit(hash_lock
);
5073 ARCSTAT_BUMP(arcstat_hits
);
5074 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
5075 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
5076 data
, metadata
, hits
);
5079 done(NULL
, buf
, private);
5081 uint64_t lsize
= BP_GET_LSIZE(bp
);
5082 uint64_t psize
= BP_GET_PSIZE(bp
);
5083 arc_callback_t
*acb
;
5086 boolean_t devw
= B_FALSE
;
5090 /* this block is not in the cache */
5091 arc_buf_hdr_t
*exists
= NULL
;
5092 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
5093 hdr
= arc_hdr_alloc(spa_load_guid(spa
), psize
, lsize
,
5094 BP_GET_COMPRESS(bp
), type
);
5096 if (!BP_IS_EMBEDDED(bp
)) {
5097 hdr
->b_dva
= *BP_IDENTITY(bp
);
5098 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
5099 exists
= buf_hash_insert(hdr
, &hash_lock
);
5101 if (exists
!= NULL
) {
5102 /* somebody beat us to the hash insert */
5103 mutex_exit(hash_lock
);
5104 buf_discard_identity(hdr
);
5105 arc_hdr_destroy(hdr
);
5106 goto top
; /* restart the IO request */
5110 * This block is in the ghost cache. If it was L2-only
5111 * (and thus didn't have an L1 hdr), we realloc the
5112 * header to add an L1 hdr.
5114 if (!HDR_HAS_L1HDR(hdr
)) {
5115 hdr
= arc_hdr_realloc(hdr
, hdr_l2only_cache
,
5118 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5119 ASSERT(GHOST_STATE(hdr
->b_l1hdr
.b_state
));
5120 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5121 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5122 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
5123 ASSERT3P(hdr
->b_l1hdr
.b_freeze_cksum
, ==, NULL
);
5126 * This is a delicate dance that we play here.
5127 * This hdr is in the ghost list so we access it
5128 * to move it out of the ghost list before we
5129 * initiate the read. If it's a prefetch then
5130 * it won't have a callback so we'll remove the
5131 * reference that arc_buf_alloc_impl() created. We
5132 * do this after we've called arc_access() to
5133 * avoid hitting an assert in remove_reference().
5135 arc_access(hdr
, hash_lock
);
5136 arc_hdr_alloc_pabd(hdr
);
5138 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5139 size
= arc_hdr_size(hdr
);
5142 * If compression is enabled on the hdr, then will do
5143 * RAW I/O and will store the compressed data in the hdr's
5144 * data block. Otherwise, the hdr's data block will contain
5145 * the uncompressed data.
5147 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
) {
5148 zio_flags
|= ZIO_FLAG_RAW
;
5151 if (*arc_flags
& ARC_FLAG_PREFETCH
)
5152 arc_hdr_set_flags(hdr
, ARC_FLAG_PREFETCH
);
5153 if (*arc_flags
& ARC_FLAG_L2CACHE
)
5154 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5155 if (BP_GET_LEVEL(bp
) > 0)
5156 arc_hdr_set_flags(hdr
, ARC_FLAG_INDIRECT
);
5157 if (*arc_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
)
5158 arc_hdr_set_flags(hdr
, ARC_FLAG_PREDICTIVE_PREFETCH
);
5159 ASSERT(!GHOST_STATE(hdr
->b_l1hdr
.b_state
));
5161 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
5162 acb
->acb_done
= done
;
5163 acb
->acb_private
= private;
5164 acb
->acb_compressed
= compressed_read
;
5166 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5167 hdr
->b_l1hdr
.b_acb
= acb
;
5168 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5170 if (HDR_HAS_L2HDR(hdr
) &&
5171 (vd
= hdr
->b_l2hdr
.b_dev
->l2ad_vdev
) != NULL
) {
5172 devw
= hdr
->b_l2hdr
.b_dev
->l2ad_writing
;
5173 addr
= hdr
->b_l2hdr
.b_daddr
;
5175 * Lock out L2ARC device removal.
5177 if (vdev_is_dead(vd
) ||
5178 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
5182 if (priority
== ZIO_PRIORITY_ASYNC_READ
)
5183 arc_hdr_set_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
5185 arc_hdr_clear_flags(hdr
, ARC_FLAG_PRIO_ASYNC_READ
);
5187 if (hash_lock
!= NULL
)
5188 mutex_exit(hash_lock
);
5191 * At this point, we have a level 1 cache miss. Try again in
5192 * L2ARC if possible.
5194 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, lsize
);
5196 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
5197 uint64_t, lsize
, zbookmark_phys_t
*, zb
);
5198 ARCSTAT_BUMP(arcstat_misses
);
5199 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
5200 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
5201 data
, metadata
, misses
);
5203 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
5205 * Read from the L2ARC if the following are true:
5206 * 1. The L2ARC vdev was previously cached.
5207 * 2. This buffer still has L2ARC metadata.
5208 * 3. This buffer isn't currently writing to the L2ARC.
5209 * 4. The L2ARC entry wasn't evicted, which may
5210 * also have invalidated the vdev.
5211 * 5. This isn't prefetch and l2arc_noprefetch is set.
5213 if (HDR_HAS_L2HDR(hdr
) &&
5214 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
5215 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
5216 l2arc_read_callback_t
*cb
;
5220 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
5221 ARCSTAT_BUMP(arcstat_l2_hits
);
5223 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
5225 cb
->l2rcb_hdr
= hdr
;
5228 cb
->l2rcb_flags
= zio_flags
;
5230 asize
= vdev_psize_to_asize(vd
, size
);
5231 if (asize
!= size
) {
5232 abd
= abd_alloc_for_io(asize
,
5233 HDR_ISTYPE_METADATA(hdr
));
5234 cb
->l2rcb_abd
= abd
;
5236 abd
= hdr
->b_l1hdr
.b_pabd
;
5239 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
5240 addr
+ asize
<= vd
->vdev_psize
-
5241 VDEV_LABEL_END_SIZE
);
5244 * l2arc read. The SCL_L2ARC lock will be
5245 * released by l2arc_read_done().
5246 * Issue a null zio if the underlying buffer
5247 * was squashed to zero size by compression.
5249 ASSERT3U(HDR_GET_COMPRESS(hdr
), !=,
5250 ZIO_COMPRESS_EMPTY
);
5251 rzio
= zio_read_phys(pio
, vd
, addr
,
5254 l2arc_read_done
, cb
, priority
,
5255 zio_flags
| ZIO_FLAG_DONT_CACHE
|
5257 ZIO_FLAG_DONT_PROPAGATE
|
5258 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
5259 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
5261 ARCSTAT_INCR(arcstat_l2_read_bytes
, size
);
5263 if (*arc_flags
& ARC_FLAG_NOWAIT
) {
5268 ASSERT(*arc_flags
& ARC_FLAG_WAIT
);
5269 if (zio_wait(rzio
) == 0)
5272 /* l2arc read error; goto zio_read() */
5274 DTRACE_PROBE1(l2arc__miss
,
5275 arc_buf_hdr_t
*, hdr
);
5276 ARCSTAT_BUMP(arcstat_l2_misses
);
5277 if (HDR_L2_WRITING(hdr
))
5278 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
5279 spa_config_exit(spa
, SCL_L2ARC
, vd
);
5283 spa_config_exit(spa
, SCL_L2ARC
, vd
);
5284 if (l2arc_ndev
!= 0) {
5285 DTRACE_PROBE1(l2arc__miss
,
5286 arc_buf_hdr_t
*, hdr
);
5287 ARCSTAT_BUMP(arcstat_l2_misses
);
5291 rzio
= zio_read(pio
, spa
, bp
, hdr
->b_l1hdr
.b_pabd
, size
,
5292 arc_read_done
, hdr
, priority
, zio_flags
, zb
);
5294 if (*arc_flags
& ARC_FLAG_WAIT
)
5295 return (zio_wait(rzio
));
5297 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
5304 * Notify the arc that a block was freed, and thus will never be used again.
5307 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
5310 kmutex_t
*hash_lock
;
5311 uint64_t guid
= spa_load_guid(spa
);
5313 ASSERT(!BP_IS_EMBEDDED(bp
));
5315 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
5320 * We might be trying to free a block that is still doing I/O
5321 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5322 * dmu_sync-ed block). If this block is being prefetched, then it
5323 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5324 * until the I/O completes. A block may also have a reference if it is
5325 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5326 * have written the new block to its final resting place on disk but
5327 * without the dedup flag set. This would have left the hdr in the MRU
5328 * state and discoverable. When the txg finally syncs it detects that
5329 * the block was overridden in open context and issues an override I/O.
5330 * Since this is a dedup block, the override I/O will determine if the
5331 * block is already in the DDT. If so, then it will replace the io_bp
5332 * with the bp from the DDT and allow the I/O to finish. When the I/O
5333 * reaches the done callback, dbuf_write_override_done, it will
5334 * check to see if the io_bp and io_bp_override are identical.
5335 * If they are not, then it indicates that the bp was replaced with
5336 * the bp in the DDT and the override bp is freed. This allows
5337 * us to arrive here with a reference on a block that is being
5338 * freed. So if we have an I/O in progress, or a reference to
5339 * this hdr, then we don't destroy the hdr.
5341 if (!HDR_HAS_L1HDR(hdr
) || (!HDR_IO_IN_PROGRESS(hdr
) &&
5342 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
))) {
5343 arc_change_state(arc_anon
, hdr
, hash_lock
);
5344 arc_hdr_destroy(hdr
);
5345 mutex_exit(hash_lock
);
5347 mutex_exit(hash_lock
);
5353 * Release this buffer from the cache, making it an anonymous buffer. This
5354 * must be done after a read and prior to modifying the buffer contents.
5355 * If the buffer has more than one reference, we must make
5356 * a new hdr for the buffer.
5359 arc_release(arc_buf_t
*buf
, void *tag
)
5361 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5364 * It would be nice to assert that if it's DMU metadata (level >
5365 * 0 || it's the dnode file), then it must be syncing context.
5366 * But we don't know that information at this level.
5369 mutex_enter(&buf
->b_evict_lock
);
5371 ASSERT(HDR_HAS_L1HDR(hdr
));
5374 * We don't grab the hash lock prior to this check, because if
5375 * the buffer's header is in the arc_anon state, it won't be
5376 * linked into the hash table.
5378 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
5379 mutex_exit(&buf
->b_evict_lock
);
5380 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5381 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
5382 ASSERT(!HDR_HAS_L2HDR(hdr
));
5383 ASSERT(HDR_EMPTY(hdr
));
5385 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
5386 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), ==, 1);
5387 ASSERT(!list_link_active(&hdr
->b_l1hdr
.b_arc_node
));
5389 hdr
->b_l1hdr
.b_arc_access
= 0;
5392 * If the buf is being overridden then it may already
5393 * have a hdr that is not empty.
5395 buf_discard_identity(hdr
);
5401 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
5402 mutex_enter(hash_lock
);
5405 * This assignment is only valid as long as the hash_lock is
5406 * held, we must be careful not to reference state or the
5407 * b_state field after dropping the lock.
5409 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
5410 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
5411 ASSERT3P(state
, !=, arc_anon
);
5413 /* this buffer is not on any list */
5414 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), >, 0);
5416 if (HDR_HAS_L2HDR(hdr
)) {
5417 mutex_enter(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
5420 * We have to recheck this conditional again now that
5421 * we're holding the l2ad_mtx to prevent a race with
5422 * another thread which might be concurrently calling
5423 * l2arc_evict(). In that case, l2arc_evict() might have
5424 * destroyed the header's L2 portion as we were waiting
5425 * to acquire the l2ad_mtx.
5427 if (HDR_HAS_L2HDR(hdr
))
5428 arc_hdr_l2hdr_destroy(hdr
);
5430 mutex_exit(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
5434 * Do we have more than one buf?
5436 if (hdr
->b_l1hdr
.b_bufcnt
> 1) {
5437 arc_buf_hdr_t
*nhdr
;
5438 uint64_t spa
= hdr
->b_spa
;
5439 uint64_t psize
= HDR_GET_PSIZE(hdr
);
5440 uint64_t lsize
= HDR_GET_LSIZE(hdr
);
5441 enum zio_compress compress
= HDR_GET_COMPRESS(hdr
);
5442 arc_buf_contents_t type
= arc_buf_type(hdr
);
5443 VERIFY3U(hdr
->b_type
, ==, type
);
5445 ASSERT(hdr
->b_l1hdr
.b_buf
!= buf
|| buf
->b_next
!= NULL
);
5446 (void) remove_reference(hdr
, hash_lock
, tag
);
5448 if (arc_buf_is_shared(buf
) && !ARC_BUF_COMPRESSED(buf
)) {
5449 ASSERT3P(hdr
->b_l1hdr
.b_buf
, !=, buf
);
5450 ASSERT(ARC_BUF_LAST(buf
));
5454 * Pull the data off of this hdr and attach it to
5455 * a new anonymous hdr. Also find the last buffer
5456 * in the hdr's buffer list.
5458 arc_buf_t
*lastbuf
= arc_buf_remove(hdr
, buf
);
5459 ASSERT3P(lastbuf
, !=, NULL
);
5462 * If the current arc_buf_t and the hdr are sharing their data
5463 * buffer, then we must stop sharing that block.
5465 if (arc_buf_is_shared(buf
)) {
5466 VERIFY(!arc_buf_is_shared(lastbuf
));
5469 * First, sever the block sharing relationship between
5470 * buf and the arc_buf_hdr_t.
5472 arc_unshare_buf(hdr
, buf
);
5475 * Now we need to recreate the hdr's b_pabd. Since we
5476 * have lastbuf handy, we try to share with it, but if
5477 * we can't then we allocate a new b_pabd and copy the
5478 * data from buf into it.
5480 if (arc_can_share(hdr
, lastbuf
)) {
5481 arc_share_buf(hdr
, lastbuf
);
5483 arc_hdr_alloc_pabd(hdr
);
5484 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
,
5485 buf
->b_data
, psize
);
5487 VERIFY3P(lastbuf
->b_data
, !=, NULL
);
5488 } else if (HDR_SHARED_DATA(hdr
)) {
5490 * Uncompressed shared buffers are always at the end
5491 * of the list. Compressed buffers don't have the
5492 * same requirements. This makes it hard to
5493 * simply assert that the lastbuf is shared so
5494 * we rely on the hdr's compression flags to determine
5495 * if we have a compressed, shared buffer.
5497 ASSERT(arc_buf_is_shared(lastbuf
) ||
5498 HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
);
5499 ASSERT(!ARC_BUF_SHARED(buf
));
5501 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
5502 ASSERT3P(state
, !=, arc_l2c_only
);
5504 (void) refcount_remove_many(&state
->arcs_size
,
5505 arc_buf_size(buf
), buf
);
5507 if (refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
)) {
5508 ASSERT3P(state
, !=, arc_l2c_only
);
5509 (void) refcount_remove_many(&state
->arcs_esize
[type
],
5510 arc_buf_size(buf
), buf
);
5513 hdr
->b_l1hdr
.b_bufcnt
-= 1;
5514 arc_cksum_verify(buf
);
5515 arc_buf_unwatch(buf
);
5517 mutex_exit(hash_lock
);
5520 * Allocate a new hdr. The new hdr will contain a b_pabd
5521 * buffer which will be freed in arc_write().
5523 nhdr
= arc_hdr_alloc(spa
, psize
, lsize
, compress
, type
);
5524 ASSERT3P(nhdr
->b_l1hdr
.b_buf
, ==, NULL
);
5525 ASSERT0(nhdr
->b_l1hdr
.b_bufcnt
);
5526 ASSERT0(refcount_count(&nhdr
->b_l1hdr
.b_refcnt
));
5527 VERIFY3U(nhdr
->b_type
, ==, type
);
5528 ASSERT(!HDR_SHARED_DATA(nhdr
));
5530 nhdr
->b_l1hdr
.b_buf
= buf
;
5531 nhdr
->b_l1hdr
.b_bufcnt
= 1;
5532 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, tag
);
5535 mutex_exit(&buf
->b_evict_lock
);
5536 (void) refcount_add_many(&arc_anon
->arcs_size
,
5537 arc_buf_size(buf
), buf
);
5539 mutex_exit(&buf
->b_evict_lock
);
5540 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 1);
5541 /* protected by hash lock, or hdr is on arc_anon */
5542 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
5543 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5544 arc_change_state(arc_anon
, hdr
, hash_lock
);
5545 hdr
->b_l1hdr
.b_arc_access
= 0;
5546 mutex_exit(hash_lock
);
5548 buf_discard_identity(hdr
);
5554 arc_released(arc_buf_t
*buf
)
5558 mutex_enter(&buf
->b_evict_lock
);
5559 released
= (buf
->b_data
!= NULL
&&
5560 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
5561 mutex_exit(&buf
->b_evict_lock
);
5567 arc_referenced(arc_buf_t
*buf
)
5571 mutex_enter(&buf
->b_evict_lock
);
5572 referenced
= (refcount_count(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
5573 mutex_exit(&buf
->b_evict_lock
);
5574 return (referenced
);
5579 arc_write_ready(zio_t
*zio
)
5581 arc_write_callback_t
*callback
= zio
->io_private
;
5582 arc_buf_t
*buf
= callback
->awcb_buf
;
5583 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5584 uint64_t psize
= BP_IS_HOLE(zio
->io_bp
) ? 0 : BP_GET_PSIZE(zio
->io_bp
);
5586 ASSERT(HDR_HAS_L1HDR(hdr
));
5587 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
5588 ASSERT(hdr
->b_l1hdr
.b_bufcnt
> 0);
5591 * If we're reexecuting this zio because the pool suspended, then
5592 * cleanup any state that was previously set the first time the
5593 * callback was invoked.
5595 if (zio
->io_flags
& ZIO_FLAG_REEXECUTED
) {
5596 arc_cksum_free(hdr
);
5597 arc_buf_unwatch(buf
);
5598 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5599 if (arc_buf_is_shared(buf
)) {
5600 arc_unshare_buf(hdr
, buf
);
5602 arc_hdr_free_pabd(hdr
);
5606 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5607 ASSERT(!HDR_SHARED_DATA(hdr
));
5608 ASSERT(!arc_buf_is_shared(buf
));
5610 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
5612 if (HDR_IO_IN_PROGRESS(hdr
))
5613 ASSERT(zio
->io_flags
& ZIO_FLAG_REEXECUTED
);
5615 arc_cksum_compute(buf
);
5616 arc_hdr_set_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5618 enum zio_compress compress
;
5619 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
5620 compress
= ZIO_COMPRESS_OFF
;
5622 ASSERT3U(HDR_GET_LSIZE(hdr
), ==, BP_GET_LSIZE(zio
->io_bp
));
5623 compress
= BP_GET_COMPRESS(zio
->io_bp
);
5625 HDR_SET_PSIZE(hdr
, psize
);
5626 arc_hdr_set_compress(hdr
, compress
);
5630 * Fill the hdr with data. If the hdr is compressed, the data we want
5631 * is available from the zio, otherwise we can take it from the buf.
5633 * We might be able to share the buf's data with the hdr here. However,
5634 * doing so would cause the ARC to be full of linear ABDs if we write a
5635 * lot of shareable data. As a compromise, we check whether scattered
5636 * ABDs are allowed, and assume that if they are then the user wants
5637 * the ARC to be primarily filled with them regardless of the data being
5638 * written. Therefore, if they're allowed then we allocate one and copy
5639 * the data into it; otherwise, we share the data directly if we can.
5641 if (zfs_abd_scatter_enabled
|| !arc_can_share(hdr
, buf
)) {
5642 arc_hdr_alloc_pabd(hdr
);
5645 * Ideally, we would always copy the io_abd into b_pabd, but the
5646 * user may have disabled compressed ARC, thus we must check the
5647 * hdr's compression setting rather than the io_bp's.
5649 if (HDR_GET_COMPRESS(hdr
) != ZIO_COMPRESS_OFF
) {
5650 ASSERT3U(BP_GET_COMPRESS(zio
->io_bp
), !=,
5652 ASSERT3U(psize
, >, 0);
5654 abd_copy(hdr
->b_l1hdr
.b_pabd
, zio
->io_abd
, psize
);
5656 ASSERT3U(zio
->io_orig_size
, ==, arc_hdr_size(hdr
));
5658 abd_copy_from_buf(hdr
->b_l1hdr
.b_pabd
, buf
->b_data
,
5662 ASSERT3P(buf
->b_data
, ==, abd_to_buf(zio
->io_orig_abd
));
5663 ASSERT3U(zio
->io_orig_size
, ==, arc_buf_size(buf
));
5664 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, ==, 1);
5666 arc_share_buf(hdr
, buf
);
5669 arc_hdr_verify(hdr
, zio
->io_bp
);
5673 arc_write_children_ready(zio_t
*zio
)
5675 arc_write_callback_t
*callback
= zio
->io_private
;
5676 arc_buf_t
*buf
= callback
->awcb_buf
;
5678 callback
->awcb_children_ready(zio
, buf
, callback
->awcb_private
);
5682 * The SPA calls this callback for each physical write that happens on behalf
5683 * of a logical write. See the comment in dbuf_write_physdone() for details.
5686 arc_write_physdone(zio_t
*zio
)
5688 arc_write_callback_t
*cb
= zio
->io_private
;
5689 if (cb
->awcb_physdone
!= NULL
)
5690 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
5694 arc_write_done(zio_t
*zio
)
5696 arc_write_callback_t
*callback
= zio
->io_private
;
5697 arc_buf_t
*buf
= callback
->awcb_buf
;
5698 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5700 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5702 if (zio
->io_error
== 0) {
5703 arc_hdr_verify(hdr
, zio
->io_bp
);
5705 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
5706 buf_discard_identity(hdr
);
5708 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
5709 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
5712 ASSERT(HDR_EMPTY(hdr
));
5716 * If the block to be written was all-zero or compressed enough to be
5717 * embedded in the BP, no write was performed so there will be no
5718 * dva/birth/checksum. The buffer must therefore remain anonymous
5721 if (!HDR_EMPTY(hdr
)) {
5722 arc_buf_hdr_t
*exists
;
5723 kmutex_t
*hash_lock
;
5725 ASSERT3U(zio
->io_error
, ==, 0);
5727 arc_cksum_verify(buf
);
5729 exists
= buf_hash_insert(hdr
, &hash_lock
);
5730 if (exists
!= NULL
) {
5732 * This can only happen if we overwrite for
5733 * sync-to-convergence, because we remove
5734 * buffers from the hash table when we arc_free().
5736 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
5737 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
5738 panic("bad overwrite, hdr=%p exists=%p",
5739 (void *)hdr
, (void *)exists
);
5740 ASSERT(refcount_is_zero(
5741 &exists
->b_l1hdr
.b_refcnt
));
5742 arc_change_state(arc_anon
, exists
, hash_lock
);
5743 mutex_exit(hash_lock
);
5744 arc_hdr_destroy(exists
);
5745 exists
= buf_hash_insert(hdr
, &hash_lock
);
5746 ASSERT3P(exists
, ==, NULL
);
5747 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
5749 ASSERT(zio
->io_prop
.zp_nopwrite
);
5750 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
5751 panic("bad nopwrite, hdr=%p exists=%p",
5752 (void *)hdr
, (void *)exists
);
5755 ASSERT(hdr
->b_l1hdr
.b_bufcnt
== 1);
5756 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
5757 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
5758 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
5761 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5762 /* if it's not anon, we are doing a scrub */
5763 if (exists
== NULL
&& hdr
->b_l1hdr
.b_state
== arc_anon
)
5764 arc_access(hdr
, hash_lock
);
5765 mutex_exit(hash_lock
);
5767 arc_hdr_clear_flags(hdr
, ARC_FLAG_IO_IN_PROGRESS
);
5770 ASSERT(!refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5771 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
5773 abd_put(zio
->io_abd
);
5774 kmem_free(callback
, sizeof (arc_write_callback_t
));
5778 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
, blkptr_t
*bp
, arc_buf_t
*buf
,
5779 boolean_t l2arc
, const zio_prop_t
*zp
, arc_done_func_t
*ready
,
5780 arc_done_func_t
*children_ready
, arc_done_func_t
*physdone
,
5781 arc_done_func_t
*done
, void *private, zio_priority_t priority
,
5782 int zio_flags
, const zbookmark_phys_t
*zb
)
5784 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5785 arc_write_callback_t
*callback
;
5787 zio_prop_t localprop
= *zp
;
5789 ASSERT3P(ready
, !=, NULL
);
5790 ASSERT3P(done
, !=, NULL
);
5791 ASSERT(!HDR_IO_ERROR(hdr
));
5792 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5793 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
5794 ASSERT3U(hdr
->b_l1hdr
.b_bufcnt
, >, 0);
5796 arc_hdr_set_flags(hdr
, ARC_FLAG_L2CACHE
);
5797 if (ARC_BUF_COMPRESSED(buf
)) {
5799 * We're writing a pre-compressed buffer. Make the
5800 * compression algorithm requested by the zio_prop_t match
5801 * the pre-compressed buffer's compression algorithm.
5803 localprop
.zp_compress
= HDR_GET_COMPRESS(hdr
);
5805 ASSERT3U(HDR_GET_LSIZE(hdr
), !=, arc_buf_size(buf
));
5806 zio_flags
|= ZIO_FLAG_RAW
;
5808 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
5809 callback
->awcb_ready
= ready
;
5810 callback
->awcb_children_ready
= children_ready
;
5811 callback
->awcb_physdone
= physdone
;
5812 callback
->awcb_done
= done
;
5813 callback
->awcb_private
= private;
5814 callback
->awcb_buf
= buf
;
5817 * The hdr's b_pabd is now stale, free it now. A new data block
5818 * will be allocated when the zio pipeline calls arc_write_ready().
5820 if (hdr
->b_l1hdr
.b_pabd
!= NULL
) {
5822 * If the buf is currently sharing the data block with
5823 * the hdr then we need to break that relationship here.
5824 * The hdr will remain with a NULL data pointer and the
5825 * buf will take sole ownership of the block.
5827 if (arc_buf_is_shared(buf
)) {
5828 arc_unshare_buf(hdr
, buf
);
5830 arc_hdr_free_pabd(hdr
);
5832 VERIFY3P(buf
->b_data
, !=, NULL
);
5833 arc_hdr_set_compress(hdr
, ZIO_COMPRESS_OFF
);
5835 ASSERT(!arc_buf_is_shared(buf
));
5836 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, ==, NULL
);
5838 zio
= zio_write(pio
, spa
, txg
, bp
,
5839 abd_get_from_buf(buf
->b_data
, HDR_GET_LSIZE(hdr
)),
5840 HDR_GET_LSIZE(hdr
), arc_buf_size(buf
), &localprop
, arc_write_ready
,
5841 (children_ready
!= NULL
) ? arc_write_children_ready
: NULL
,
5842 arc_write_physdone
, arc_write_done
, callback
,
5843 priority
, zio_flags
, zb
);
5849 arc_memory_throttle(spa_t
*spa
, uint64_t reserve
, uint64_t txg
)
5852 uint64_t available_memory
= ptob(freemem
);
5856 MIN(available_memory
, vmem_size(heap_arena
, VMEM_FREE
));
5859 if (freemem
> physmem
* arc_lotsfree_percent
/ 100)
5862 if (txg
> spa
->spa_lowmem_last_txg
) {
5863 spa
->spa_lowmem_last_txg
= txg
;
5864 spa
->spa_lowmem_page_load
= 0;
5867 * If we are in pageout, we know that memory is already tight,
5868 * the arc is already going to be evicting, so we just want to
5869 * continue to let page writes occur as quickly as possible.
5871 if (curproc
== proc_pageout
) {
5872 if (spa
->spa_lowmem_page_load
>
5873 MAX(ptob(minfree
), available_memory
) / 4)
5874 return (SET_ERROR(ERESTART
));
5875 /* Note: reserve is inflated, so we deflate */
5876 atomic_add_64(&spa
->spa_lowmem_page_load
, reserve
/ 8);
5878 } else if (spa
->spa_lowmem_page_load
> 0 && arc_reclaim_needed()) {
5879 /* memory is low, delay before restarting */
5880 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
5881 return (SET_ERROR(EAGAIN
));
5883 spa
->spa_lowmem_page_load
= 0;
5884 #endif /* _KERNEL */
5889 arc_tempreserve_clear(uint64_t reserve
)
5891 atomic_add_64(&arc_tempreserve
, -reserve
);
5892 ASSERT((int64_t)arc_tempreserve
>= 0);
5896 arc_tempreserve_space(spa_t
*spa
, uint64_t reserve
, uint64_t txg
)
5901 if (reserve
> arc_c
/4 && !arc_no_grow
)
5902 arc_c
= MIN(arc_c_max
, reserve
* 4);
5903 if (reserve
> arc_c
)
5904 return (SET_ERROR(ENOMEM
));
5907 * Don't count loaned bufs as in flight dirty data to prevent long
5908 * network delays from blocking transactions that are ready to be
5909 * assigned to a txg.
5912 /* assert that it has not wrapped around */
5913 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes
, 0), >=, 0);
5915 anon_size
= MAX((int64_t)(refcount_count(&arc_anon
->arcs_size
) -
5916 arc_loaned_bytes
), 0);
5919 * Writes will, almost always, require additional memory allocations
5920 * in order to compress/encrypt/etc the data. We therefore need to
5921 * make sure that there is sufficient available memory for this.
5923 error
= arc_memory_throttle(spa
, reserve
, txg
);
5928 * Throttle writes when the amount of dirty data in the cache
5929 * gets too large. We try to keep the cache less than half full
5930 * of dirty blocks so that our sync times don't grow too large.
5932 * In the case of one pool being built on another pool, we want
5933 * to make sure we don't end up throttling the lower (backing)
5934 * pool when the upper pool is the majority contributor to dirty
5935 * data. To insure we make forward progress during throttling, we
5936 * also check the current pool's net dirty data and only throttle
5937 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
5938 * data in the cache.
5940 * Note: if two requests come in concurrently, we might let them
5941 * both succeed, when one of them should fail. Not a huge deal.
5943 uint64_t total_dirty
= reserve
+ arc_tempreserve
+ anon_size
;
5944 uint64_t spa_dirty_anon
= spa_dirty_data(spa
);
5946 if (total_dirty
> arc_c
* zfs_arc_dirty_limit_percent
/ 100 &&
5947 anon_size
> arc_c
* zfs_arc_anon_limit_percent
/ 100 &&
5948 spa_dirty_anon
> anon_size
* zfs_arc_pool_dirty_percent
/ 100) {
5949 uint64_t meta_esize
=
5950 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
5951 uint64_t data_esize
=
5952 refcount_count(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
5953 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5954 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5955 arc_tempreserve
>> 10, meta_esize
>> 10,
5956 data_esize
>> 10, reserve
>> 10, arc_c
>> 10);
5957 return (SET_ERROR(ERESTART
));
5959 atomic_add_64(&arc_tempreserve
, reserve
);
5964 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
5965 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
5967 size
->value
.ui64
= refcount_count(&state
->arcs_size
);
5968 evict_data
->value
.ui64
=
5969 refcount_count(&state
->arcs_esize
[ARC_BUFC_DATA
]);
5970 evict_metadata
->value
.ui64
=
5971 refcount_count(&state
->arcs_esize
[ARC_BUFC_METADATA
]);
5975 arc_kstat_update(kstat_t
*ksp
, int rw
)
5977 arc_stats_t
*as
= ksp
->ks_data
;
5979 if (rw
== KSTAT_WRITE
) {
5982 arc_kstat_update_state(arc_anon
,
5983 &as
->arcstat_anon_size
,
5984 &as
->arcstat_anon_evictable_data
,
5985 &as
->arcstat_anon_evictable_metadata
);
5986 arc_kstat_update_state(arc_mru
,
5987 &as
->arcstat_mru_size
,
5988 &as
->arcstat_mru_evictable_data
,
5989 &as
->arcstat_mru_evictable_metadata
);
5990 arc_kstat_update_state(arc_mru_ghost
,
5991 &as
->arcstat_mru_ghost_size
,
5992 &as
->arcstat_mru_ghost_evictable_data
,
5993 &as
->arcstat_mru_ghost_evictable_metadata
);
5994 arc_kstat_update_state(arc_mfu
,
5995 &as
->arcstat_mfu_size
,
5996 &as
->arcstat_mfu_evictable_data
,
5997 &as
->arcstat_mfu_evictable_metadata
);
5998 arc_kstat_update_state(arc_mfu_ghost
,
5999 &as
->arcstat_mfu_ghost_size
,
6000 &as
->arcstat_mfu_ghost_evictable_data
,
6001 &as
->arcstat_mfu_ghost_evictable_metadata
);
6003 ARCSTAT(arcstat_size
) = aggsum_value(&arc_size
);
6004 ARCSTAT(arcstat_meta_used
) = aggsum_value(&arc_meta_used
);
6005 ARCSTAT(arcstat_data_size
) = aggsum_value(&astat_data_size
);
6006 ARCSTAT(arcstat_metadata_size
) =
6007 aggsum_value(&astat_metadata_size
);
6008 ARCSTAT(arcstat_hdr_size
) = aggsum_value(&astat_hdr_size
);
6009 ARCSTAT(arcstat_other_size
) = aggsum_value(&astat_other_size
);
6010 ARCSTAT(arcstat_l2_hdr_size
) = aggsum_value(&astat_l2_hdr_size
);
6017 * This function *must* return indices evenly distributed between all
6018 * sublists of the multilist. This is needed due to how the ARC eviction
6019 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6020 * distributed between all sublists and uses this assumption when
6021 * deciding which sublist to evict from and how much to evict from it.
6024 arc_state_multilist_index_func(multilist_t
*ml
, void *obj
)
6026 arc_buf_hdr_t
*hdr
= obj
;
6029 * We rely on b_dva to generate evenly distributed index
6030 * numbers using buf_hash below. So, as an added precaution,
6031 * let's make sure we never add empty buffers to the arc lists.
6033 ASSERT(!HDR_EMPTY(hdr
));
6036 * The assumption here, is the hash value for a given
6037 * arc_buf_hdr_t will remain constant throughout it's lifetime
6038 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6039 * Thus, we don't need to store the header's sublist index
6040 * on insertion, as this index can be recalculated on removal.
6042 * Also, the low order bits of the hash value are thought to be
6043 * distributed evenly. Otherwise, in the case that the multilist
6044 * has a power of two number of sublists, each sublists' usage
6045 * would not be evenly distributed.
6047 return (buf_hash(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
) %
6048 multilist_get_num_sublists(ml
));
6052 arc_state_init(void)
6054 arc_anon
= &ARC_anon
;
6056 arc_mru_ghost
= &ARC_mru_ghost
;
6058 arc_mfu_ghost
= &ARC_mfu_ghost
;
6059 arc_l2c_only
= &ARC_l2c_only
;
6061 arc_mru
->arcs_list
[ARC_BUFC_METADATA
] =
6062 multilist_create(sizeof (arc_buf_hdr_t
),
6063 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6064 arc_state_multilist_index_func
);
6065 arc_mru
->arcs_list
[ARC_BUFC_DATA
] =
6066 multilist_create(sizeof (arc_buf_hdr_t
),
6067 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6068 arc_state_multilist_index_func
);
6069 arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
6070 multilist_create(sizeof (arc_buf_hdr_t
),
6071 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6072 arc_state_multilist_index_func
);
6073 arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
] =
6074 multilist_create(sizeof (arc_buf_hdr_t
),
6075 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6076 arc_state_multilist_index_func
);
6077 arc_mfu
->arcs_list
[ARC_BUFC_METADATA
] =
6078 multilist_create(sizeof (arc_buf_hdr_t
),
6079 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6080 arc_state_multilist_index_func
);
6081 arc_mfu
->arcs_list
[ARC_BUFC_DATA
] =
6082 multilist_create(sizeof (arc_buf_hdr_t
),
6083 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6084 arc_state_multilist_index_func
);
6085 arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
] =
6086 multilist_create(sizeof (arc_buf_hdr_t
),
6087 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6088 arc_state_multilist_index_func
);
6089 arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
] =
6090 multilist_create(sizeof (arc_buf_hdr_t
),
6091 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6092 arc_state_multilist_index_func
);
6093 arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
] =
6094 multilist_create(sizeof (arc_buf_hdr_t
),
6095 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6096 arc_state_multilist_index_func
);
6097 arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
] =
6098 multilist_create(sizeof (arc_buf_hdr_t
),
6099 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
6100 arc_state_multilist_index_func
);
6102 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6103 refcount_create(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6104 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
6105 refcount_create(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
6106 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6107 refcount_create(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6108 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
6109 refcount_create(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
6110 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6111 refcount_create(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6112 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
6113 refcount_create(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
6115 refcount_create(&arc_anon
->arcs_size
);
6116 refcount_create(&arc_mru
->arcs_size
);
6117 refcount_create(&arc_mru_ghost
->arcs_size
);
6118 refcount_create(&arc_mfu
->arcs_size
);
6119 refcount_create(&arc_mfu_ghost
->arcs_size
);
6120 refcount_create(&arc_l2c_only
->arcs_size
);
6122 aggsum_init(&arc_meta_used
, 0);
6123 aggsum_init(&arc_size
, 0);
6124 aggsum_init(&astat_data_size
, 0);
6125 aggsum_init(&astat_metadata_size
, 0);
6126 aggsum_init(&astat_hdr_size
, 0);
6127 aggsum_init(&astat_other_size
, 0);
6128 aggsum_init(&astat_l2_hdr_size
, 0);
6132 arc_state_fini(void)
6134 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_METADATA
]);
6135 refcount_destroy(&arc_anon
->arcs_esize
[ARC_BUFC_DATA
]);
6136 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_METADATA
]);
6137 refcount_destroy(&arc_mru
->arcs_esize
[ARC_BUFC_DATA
]);
6138 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6139 refcount_destroy(&arc_mru_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6140 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_METADATA
]);
6141 refcount_destroy(&arc_mfu
->arcs_esize
[ARC_BUFC_DATA
]);
6142 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_METADATA
]);
6143 refcount_destroy(&arc_mfu_ghost
->arcs_esize
[ARC_BUFC_DATA
]);
6144 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_METADATA
]);
6145 refcount_destroy(&arc_l2c_only
->arcs_esize
[ARC_BUFC_DATA
]);
6147 refcount_destroy(&arc_anon
->arcs_size
);
6148 refcount_destroy(&arc_mru
->arcs_size
);
6149 refcount_destroy(&arc_mru_ghost
->arcs_size
);
6150 refcount_destroy(&arc_mfu
->arcs_size
);
6151 refcount_destroy(&arc_mfu_ghost
->arcs_size
);
6152 refcount_destroy(&arc_l2c_only
->arcs_size
);
6154 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
6155 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
6156 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
6157 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
6158 multilist_destroy(arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
6159 multilist_destroy(arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
6160 multilist_destroy(arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
6161 multilist_destroy(arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
6163 aggsum_fini(&arc_meta_used
);
6164 aggsum_fini(&arc_size
);
6165 aggsum_fini(&astat_data_size
);
6166 aggsum_fini(&astat_metadata_size
);
6167 aggsum_fini(&astat_hdr_size
);
6168 aggsum_fini(&astat_other_size
);
6169 aggsum_fini(&astat_l2_hdr_size
);
6182 * allmem is "all memory that we could possibly use".
6185 uint64_t allmem
= ptob(physmem
- swapfs_minfree
);
6187 uint64_t allmem
= (physmem
* PAGESIZE
) / 2;
6189 mutex_init(&arc_adjust_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
6190 cv_init(&arc_adjust_waiters_cv
, NULL
, CV_DEFAULT
, NULL
);
6192 /* Convert seconds to clock ticks */
6193 arc_min_prefetch_lifespan
= 1 * hz
;
6195 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
6196 arc_c_min
= MAX(allmem
/ 32, 64 << 20);
6197 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
6198 if (allmem
>= 1 << 30)
6199 arc_c_max
= allmem
- (1 << 30);
6201 arc_c_max
= arc_c_min
;
6202 arc_c_max
= MAX(allmem
* 3 / 4, arc_c_max
);
6205 * In userland, there's only the memory pressure that we artificially
6206 * create (see arc_available_memory()). Don't let arc_c get too
6207 * small, because it can cause transactions to be larger than
6208 * arc_c, causing arc_tempreserve_space() to fail.
6211 arc_c_min
= arc_c_max
/ 2;
6215 * Allow the tunables to override our calculations if they are
6216 * reasonable (ie. over 64MB)
6218 if (zfs_arc_max
> 64 << 20 && zfs_arc_max
< allmem
) {
6219 arc_c_max
= zfs_arc_max
;
6220 arc_c_min
= MIN(arc_c_min
, arc_c_max
);
6222 if (zfs_arc_min
> 64 << 20 && zfs_arc_min
<= arc_c_max
)
6223 arc_c_min
= zfs_arc_min
;
6226 arc_p
= (arc_c
>> 1);
6228 /* limit meta-data to 1/4 of the arc capacity */
6229 arc_meta_limit
= arc_c_max
/ 4;
6233 * Metadata is stored in the kernel's heap. Don't let us
6234 * use more than half the heap for the ARC.
6236 arc_meta_limit
= MIN(arc_meta_limit
,
6237 vmem_size(heap_arena
, VMEM_ALLOC
| VMEM_FREE
) / 2);
6240 /* Allow the tunable to override if it is reasonable */
6241 if (zfs_arc_meta_limit
> 0 && zfs_arc_meta_limit
<= arc_c_max
)
6242 arc_meta_limit
= zfs_arc_meta_limit
;
6244 if (arc_c_min
< arc_meta_limit
/ 2 && zfs_arc_min
== 0)
6245 arc_c_min
= arc_meta_limit
/ 2;
6247 if (zfs_arc_meta_min
> 0) {
6248 arc_meta_min
= zfs_arc_meta_min
;
6250 arc_meta_min
= arc_c_min
/ 2;
6253 if (zfs_arc_grow_retry
> 0)
6254 arc_grow_retry
= zfs_arc_grow_retry
;
6256 if (zfs_arc_shrink_shift
> 0)
6257 arc_shrink_shift
= zfs_arc_shrink_shift
;
6260 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6262 if (arc_no_grow_shift
>= arc_shrink_shift
)
6263 arc_no_grow_shift
= arc_shrink_shift
- 1;
6265 if (zfs_arc_p_min_shift
> 0)
6266 arc_p_min_shift
= zfs_arc_p_min_shift
;
6268 /* if kmem_flags are set, lets try to use less memory */
6269 if (kmem_debugging())
6271 if (arc_c
< arc_c_min
)
6277 * The arc must be "uninitialized", so that hdr_recl() (which is
6278 * registered by buf_init()) will not access arc_reap_zthr before
6281 ASSERT(!arc_initialized
);
6284 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
6285 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
6287 if (arc_ksp
!= NULL
) {
6288 arc_ksp
->ks_data
= &arc_stats
;
6289 arc_ksp
->ks_update
= arc_kstat_update
;
6290 kstat_install(arc_ksp
);
6293 arc_adjust_zthr
= zthr_create(arc_adjust_cb_check
,
6294 arc_adjust_cb
, NULL
);
6295 arc_reap_zthr
= zthr_create_timer(arc_reap_cb_check
,
6296 arc_reap_cb
, NULL
, SEC2NSEC(1));
6298 arc_initialized
= B_TRUE
;
6302 * Calculate maximum amount of dirty data per pool.
6304 * If it has been set by /etc/system, take that.
6305 * Otherwise, use a percentage of physical memory defined by
6306 * zfs_dirty_data_max_percent (default 10%) with a cap at
6307 * zfs_dirty_data_max_max (default 4GB).
6309 if (zfs_dirty_data_max
== 0) {
6310 zfs_dirty_data_max
= physmem
* PAGESIZE
*
6311 zfs_dirty_data_max_percent
/ 100;
6312 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
6313 zfs_dirty_data_max_max
);
6320 /* Use B_TRUE to ensure *all* buffers are evicted */
6321 arc_flush(NULL
, B_TRUE
);
6323 arc_initialized
= B_FALSE
;
6325 if (arc_ksp
!= NULL
) {
6326 kstat_delete(arc_ksp
);
6330 (void) zthr_cancel(arc_adjust_zthr
);
6331 zthr_destroy(arc_adjust_zthr
);
6333 (void) zthr_cancel(arc_reap_zthr
);
6334 zthr_destroy(arc_reap_zthr
);
6336 mutex_destroy(&arc_adjust_lock
);
6337 cv_destroy(&arc_adjust_waiters_cv
);
6340 * buf_fini() must proceed arc_state_fini() because buf_fin() may
6341 * trigger the release of kmem magazines, which can callback to
6342 * arc_space_return() which accesses aggsums freed in act_state_fini().
6347 ASSERT0(arc_loaned_bytes
);
6353 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6354 * It uses dedicated storage devices to hold cached data, which are populated
6355 * using large infrequent writes. The main role of this cache is to boost
6356 * the performance of random read workloads. The intended L2ARC devices
6357 * include short-stroked disks, solid state disks, and other media with
6358 * substantially faster read latency than disk.
6360 * +-----------------------+
6362 * +-----------------------+
6365 * l2arc_feed_thread() arc_read()
6369 * +---------------+ |
6371 * +---------------+ |
6376 * +-------+ +-------+
6378 * | cache | | cache |
6379 * +-------+ +-------+
6380 * +=========+ .-----.
6381 * : L2ARC : |-_____-|
6382 * : devices : | Disks |
6383 * +=========+ `-_____-'
6385 * Read requests are satisfied from the following sources, in order:
6388 * 2) vdev cache of L2ARC devices
6390 * 4) vdev cache of disks
6393 * Some L2ARC device types exhibit extremely slow write performance.
6394 * To accommodate for this there are some significant differences between
6395 * the L2ARC and traditional cache design:
6397 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6398 * the ARC behave as usual, freeing buffers and placing headers on ghost
6399 * lists. The ARC does not send buffers to the L2ARC during eviction as
6400 * this would add inflated write latencies for all ARC memory pressure.
6402 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6403 * It does this by periodically scanning buffers from the eviction-end of
6404 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6405 * not already there. It scans until a headroom of buffers is satisfied,
6406 * which itself is a buffer for ARC eviction. If a compressible buffer is
6407 * found during scanning and selected for writing to an L2ARC device, we
6408 * temporarily boost scanning headroom during the next scan cycle to make
6409 * sure we adapt to compression effects (which might significantly reduce
6410 * the data volume we write to L2ARC). The thread that does this is
6411 * l2arc_feed_thread(), illustrated below; example sizes are included to
6412 * provide a better sense of ratio than this diagram:
6415 * +---------------------+----------+
6416 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6417 * +---------------------+----------+ | o L2ARC eligible
6418 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6419 * +---------------------+----------+ |
6420 * 15.9 Gbytes ^ 32 Mbytes |
6422 * l2arc_feed_thread()
6424 * l2arc write hand <--[oooo]--'
6428 * +==============================+
6429 * L2ARC dev |####|#|###|###| |####| ... |
6430 * +==============================+
6433 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6434 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6435 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6436 * safe to say that this is an uncommon case, since buffers at the end of
6437 * the ARC lists have moved there due to inactivity.
6439 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6440 * then the L2ARC simply misses copying some buffers. This serves as a
6441 * pressure valve to prevent heavy read workloads from both stalling the ARC
6442 * with waits and clogging the L2ARC with writes. This also helps prevent
6443 * the potential for the L2ARC to churn if it attempts to cache content too
6444 * quickly, such as during backups of the entire pool.
6446 * 5. After system boot and before the ARC has filled main memory, there are
6447 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6448 * lists can remain mostly static. Instead of searching from tail of these
6449 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6450 * for eligible buffers, greatly increasing its chance of finding them.
6452 * The L2ARC device write speed is also boosted during this time so that
6453 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6454 * there are no L2ARC reads, and no fear of degrading read performance
6455 * through increased writes.
6457 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6458 * the vdev queue can aggregate them into larger and fewer writes. Each
6459 * device is written to in a rotor fashion, sweeping writes through
6460 * available space then repeating.
6462 * 7. The L2ARC does not store dirty content. It never needs to flush
6463 * write buffers back to disk based storage.
6465 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6466 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6468 * The performance of the L2ARC can be tweaked by a number of tunables, which
6469 * may be necessary for different workloads:
6471 * l2arc_write_max max write bytes per interval
6472 * l2arc_write_boost extra write bytes during device warmup
6473 * l2arc_noprefetch skip caching prefetched buffers
6474 * l2arc_headroom number of max device writes to precache
6475 * l2arc_headroom_boost when we find compressed buffers during ARC
6476 * scanning, we multiply headroom by this
6477 * percentage factor for the next scan cycle,
6478 * since more compressed buffers are likely to
6480 * l2arc_feed_secs seconds between L2ARC writing
6482 * Tunables may be removed or added as future performance improvements are
6483 * integrated, and also may become zpool properties.
6485 * There are three key functions that control how the L2ARC warms up:
6487 * l2arc_write_eligible() check if a buffer is eligible to cache
6488 * l2arc_write_size() calculate how much to write
6489 * l2arc_write_interval() calculate sleep delay between writes
6491 * These three functions determine what to write, how much, and how quickly
6496 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*hdr
)
6499 * A buffer is *not* eligible for the L2ARC if it:
6500 * 1. belongs to a different spa.
6501 * 2. is already cached on the L2ARC.
6502 * 3. has an I/O in progress (it may be an incomplete read).
6503 * 4. is flagged not eligible (zfs property).
6505 if (hdr
->b_spa
!= spa_guid
|| HDR_HAS_L2HDR(hdr
) ||
6506 HDR_IO_IN_PROGRESS(hdr
) || !HDR_L2CACHE(hdr
))
6513 l2arc_write_size(void)
6518 * Make sure our globals have meaningful values in case the user
6521 size
= l2arc_write_max
;
6523 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
6524 "be greater than zero, resetting it to the default (%d)",
6526 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
6529 if (arc_warm
== B_FALSE
)
6530 size
+= l2arc_write_boost
;
6537 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
6539 clock_t interval
, next
, now
;
6542 * If the ARC lists are busy, increase our write rate; if the
6543 * lists are stale, idle back. This is achieved by checking
6544 * how much we previously wrote - if it was more than half of
6545 * what we wanted, schedule the next write much sooner.
6547 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
6548 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
6550 interval
= hz
* l2arc_feed_secs
;
6552 now
= ddi_get_lbolt();
6553 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
6559 * Cycle through L2ARC devices. This is how L2ARC load balances.
6560 * If a device is returned, this also returns holding the spa config lock.
6562 static l2arc_dev_t
*
6563 l2arc_dev_get_next(void)
6565 l2arc_dev_t
*first
, *next
= NULL
;
6568 * Lock out the removal of spas (spa_namespace_lock), then removal
6569 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6570 * both locks will be dropped and a spa config lock held instead.
6572 mutex_enter(&spa_namespace_lock
);
6573 mutex_enter(&l2arc_dev_mtx
);
6575 /* if there are no vdevs, there is nothing to do */
6576 if (l2arc_ndev
== 0)
6580 next
= l2arc_dev_last
;
6582 /* loop around the list looking for a non-faulted vdev */
6584 next
= list_head(l2arc_dev_list
);
6586 next
= list_next(l2arc_dev_list
, next
);
6588 next
= list_head(l2arc_dev_list
);
6591 /* if we have come back to the start, bail out */
6594 else if (next
== first
)
6597 } while (vdev_is_dead(next
->l2ad_vdev
));
6599 /* if we were unable to find any usable vdevs, return NULL */
6600 if (vdev_is_dead(next
->l2ad_vdev
))
6603 l2arc_dev_last
= next
;
6606 mutex_exit(&l2arc_dev_mtx
);
6609 * Grab the config lock to prevent the 'next' device from being
6610 * removed while we are writing to it.
6613 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
6614 mutex_exit(&spa_namespace_lock
);
6620 * Free buffers that were tagged for destruction.
6623 l2arc_do_free_on_write()
6626 l2arc_data_free_t
*df
, *df_prev
;
6628 mutex_enter(&l2arc_free_on_write_mtx
);
6629 buflist
= l2arc_free_on_write
;
6631 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
6632 df_prev
= list_prev(buflist
, df
);
6633 ASSERT3P(df
->l2df_abd
, !=, NULL
);
6634 abd_free(df
->l2df_abd
);
6635 list_remove(buflist
, df
);
6636 kmem_free(df
, sizeof (l2arc_data_free_t
));
6639 mutex_exit(&l2arc_free_on_write_mtx
);
6643 * A write to a cache device has completed. Update all headers to allow
6644 * reads from these buffers to begin.
6647 l2arc_write_done(zio_t
*zio
)
6649 l2arc_write_callback_t
*cb
;
6652 arc_buf_hdr_t
*head
, *hdr
, *hdr_prev
;
6653 kmutex_t
*hash_lock
;
6654 int64_t bytes_dropped
= 0;
6656 cb
= zio
->io_private
;
6657 ASSERT3P(cb
, !=, NULL
);
6658 dev
= cb
->l2wcb_dev
;
6659 ASSERT3P(dev
, !=, NULL
);
6660 head
= cb
->l2wcb_head
;
6661 ASSERT3P(head
, !=, NULL
);
6662 buflist
= &dev
->l2ad_buflist
;
6663 ASSERT3P(buflist
, !=, NULL
);
6664 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
6665 l2arc_write_callback_t
*, cb
);
6667 if (zio
->io_error
!= 0)
6668 ARCSTAT_BUMP(arcstat_l2_writes_error
);
6671 * All writes completed, or an error was hit.
6674 mutex_enter(&dev
->l2ad_mtx
);
6675 for (hdr
= list_prev(buflist
, head
); hdr
; hdr
= hdr_prev
) {
6676 hdr_prev
= list_prev(buflist
, hdr
);
6678 hash_lock
= HDR_LOCK(hdr
);
6681 * We cannot use mutex_enter or else we can deadlock
6682 * with l2arc_write_buffers (due to swapping the order
6683 * the hash lock and l2ad_mtx are taken).
6685 if (!mutex_tryenter(hash_lock
)) {
6687 * Missed the hash lock. We must retry so we
6688 * don't leave the ARC_FLAG_L2_WRITING bit set.
6690 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry
);
6693 * We don't want to rescan the headers we've
6694 * already marked as having been written out, so
6695 * we reinsert the head node so we can pick up
6696 * where we left off.
6698 list_remove(buflist
, head
);
6699 list_insert_after(buflist
, hdr
, head
);
6701 mutex_exit(&dev
->l2ad_mtx
);
6704 * We wait for the hash lock to become available
6705 * to try and prevent busy waiting, and increase
6706 * the chance we'll be able to acquire the lock
6707 * the next time around.
6709 mutex_enter(hash_lock
);
6710 mutex_exit(hash_lock
);
6715 * We could not have been moved into the arc_l2c_only
6716 * state while in-flight due to our ARC_FLAG_L2_WRITING
6717 * bit being set. Let's just ensure that's being enforced.
6719 ASSERT(HDR_HAS_L1HDR(hdr
));
6721 if (zio
->io_error
!= 0) {
6723 * Error - drop L2ARC entry.
6725 list_remove(buflist
, hdr
);
6726 arc_hdr_clear_flags(hdr
, ARC_FLAG_HAS_L2HDR
);
6728 ARCSTAT_INCR(arcstat_l2_psize
, -arc_hdr_size(hdr
));
6729 ARCSTAT_INCR(arcstat_l2_lsize
, -HDR_GET_LSIZE(hdr
));
6731 bytes_dropped
+= arc_hdr_size(hdr
);
6732 (void) refcount_remove_many(&dev
->l2ad_alloc
,
6733 arc_hdr_size(hdr
), hdr
);
6737 * Allow ARC to begin reads and ghost list evictions to
6740 arc_hdr_clear_flags(hdr
, ARC_FLAG_L2_WRITING
);
6742 mutex_exit(hash_lock
);
6745 atomic_inc_64(&l2arc_writes_done
);
6746 list_remove(buflist
, head
);
6747 ASSERT(!HDR_HAS_L1HDR(head
));
6748 kmem_cache_free(hdr_l2only_cache
, head
);
6749 mutex_exit(&dev
->l2ad_mtx
);
6751 vdev_space_update(dev
->l2ad_vdev
, -bytes_dropped
, 0, 0);
6753 l2arc_do_free_on_write();
6755 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
6759 * A read to a cache device completed. Validate buffer contents before
6760 * handing over to the regular ARC routines.
6763 l2arc_read_done(zio_t
*zio
)
6765 l2arc_read_callback_t
*cb
;
6767 kmutex_t
*hash_lock
;
6768 boolean_t valid_cksum
;
6770 ASSERT3P(zio
->io_vd
, !=, NULL
);
6771 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
6773 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
6775 cb
= zio
->io_private
;
6776 ASSERT3P(cb
, !=, NULL
);
6777 hdr
= cb
->l2rcb_hdr
;
6778 ASSERT3P(hdr
, !=, NULL
);
6780 hash_lock
= HDR_LOCK(hdr
);
6781 mutex_enter(hash_lock
);
6782 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
6785 * If the data was read into a temporary buffer,
6786 * move it and free the buffer.
6788 if (cb
->l2rcb_abd
!= NULL
) {
6789 ASSERT3U(arc_hdr_size(hdr
), <, zio
->io_size
);
6790 if (zio
->io_error
== 0) {
6791 abd_copy(hdr
->b_l1hdr
.b_pabd
, cb
->l2rcb_abd
,
6796 * The following must be done regardless of whether
6797 * there was an error:
6798 * - free the temporary buffer
6799 * - point zio to the real ARC buffer
6800 * - set zio size accordingly
6801 * These are required because zio is either re-used for
6802 * an I/O of the block in the case of the error
6803 * or the zio is passed to arc_read_done() and it
6806 abd_free(cb
->l2rcb_abd
);
6807 zio
->io_size
= zio
->io_orig_size
= arc_hdr_size(hdr
);
6808 zio
->io_abd
= zio
->io_orig_abd
= hdr
->b_l1hdr
.b_pabd
;
6811 ASSERT3P(zio
->io_abd
, !=, NULL
);
6814 * Check this survived the L2ARC journey.
6816 ASSERT3P(zio
->io_abd
, ==, hdr
->b_l1hdr
.b_pabd
);
6817 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
6818 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
6820 valid_cksum
= arc_cksum_is_equal(hdr
, zio
);
6821 if (valid_cksum
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
6822 mutex_exit(hash_lock
);
6823 zio
->io_private
= hdr
;
6826 mutex_exit(hash_lock
);
6828 * Buffer didn't survive caching. Increment stats and
6829 * reissue to the original storage device.
6831 if (zio
->io_error
!= 0) {
6832 ARCSTAT_BUMP(arcstat_l2_io_error
);
6834 zio
->io_error
= SET_ERROR(EIO
);
6837 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
6840 * If there's no waiter, issue an async i/o to the primary
6841 * storage now. If there *is* a waiter, the caller must
6842 * issue the i/o in a context where it's OK to block.
6844 if (zio
->io_waiter
== NULL
) {
6845 zio_t
*pio
= zio_unique_parent(zio
);
6847 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
6849 zio_nowait(zio_read(pio
, zio
->io_spa
, zio
->io_bp
,
6850 hdr
->b_l1hdr
.b_pabd
, zio
->io_size
, arc_read_done
,
6851 hdr
, zio
->io_priority
, cb
->l2rcb_flags
,
6856 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
6860 * This is the list priority from which the L2ARC will search for pages to
6861 * cache. This is used within loops (0..3) to cycle through lists in the
6862 * desired order. This order can have a significant effect on cache
6865 * Currently the metadata lists are hit first, MFU then MRU, followed by
6866 * the data lists. This function returns a locked list, and also returns
6869 static multilist_sublist_t
*
6870 l2arc_sublist_lock(int list_num
)
6872 multilist_t
*ml
= NULL
;
6875 ASSERT(list_num
>= 0 && list_num
<= 3);
6879 ml
= arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
6882 ml
= arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
6885 ml
= arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
6888 ml
= arc_mru
->arcs_list
[ARC_BUFC_DATA
];
6893 * Return a randomly-selected sublist. This is acceptable
6894 * because the caller feeds only a little bit of data for each
6895 * call (8MB). Subsequent calls will result in different
6896 * sublists being selected.
6898 idx
= multilist_get_random_index(ml
);
6899 return (multilist_sublist_lock(ml
, idx
));
6903 * Evict buffers from the device write hand to the distance specified in
6904 * bytes. This distance may span populated buffers, it may span nothing.
6905 * This is clearing a region on the L2ARC device ready for writing.
6906 * If the 'all' boolean is set, every buffer is evicted.
6909 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
6912 arc_buf_hdr_t
*hdr
, *hdr_prev
;
6913 kmutex_t
*hash_lock
;
6916 buflist
= &dev
->l2ad_buflist
;
6918 if (!all
&& dev
->l2ad_first
) {
6920 * This is the first sweep through the device. There is
6926 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
6928 * When nearing the end of the device, evict to the end
6929 * before the device write hand jumps to the start.
6931 taddr
= dev
->l2ad_end
;
6933 taddr
= dev
->l2ad_hand
+ distance
;
6935 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
6936 uint64_t, taddr
, boolean_t
, all
);
6939 mutex_enter(&dev
->l2ad_mtx
);
6940 for (hdr
= list_tail(buflist
); hdr
; hdr
= hdr_prev
) {
6941 hdr_prev
= list_prev(buflist
, hdr
);
6943 hash_lock
= HDR_LOCK(hdr
);
6946 * We cannot use mutex_enter or else we can deadlock
6947 * with l2arc_write_buffers (due to swapping the order
6948 * the hash lock and l2ad_mtx are taken).
6950 if (!mutex_tryenter(hash_lock
)) {
6952 * Missed the hash lock. Retry.
6954 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
6955 mutex_exit(&dev
->l2ad_mtx
);
6956 mutex_enter(hash_lock
);
6957 mutex_exit(hash_lock
);
6962 * A header can't be on this list if it doesn't have L2 header.
6964 ASSERT(HDR_HAS_L2HDR(hdr
));
6966 /* Ensure this header has finished being written. */
6967 ASSERT(!HDR_L2_WRITING(hdr
));
6968 ASSERT(!HDR_L2_WRITE_HEAD(hdr
));
6970 if (!all
&& (hdr
->b_l2hdr
.b_daddr
>= taddr
||
6971 hdr
->b_l2hdr
.b_daddr
< dev
->l2ad_hand
)) {
6973 * We've evicted to the target address,
6974 * or the end of the device.
6976 mutex_exit(hash_lock
);
6980 if (!HDR_HAS_L1HDR(hdr
)) {
6981 ASSERT(!HDR_L2_READING(hdr
));
6983 * This doesn't exist in the ARC. Destroy.
6984 * arc_hdr_destroy() will call list_remove()
6985 * and decrement arcstat_l2_lsize.
6987 arc_change_state(arc_anon
, hdr
, hash_lock
);
6988 arc_hdr_destroy(hdr
);
6990 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_l2c_only
);
6991 ARCSTAT_BUMP(arcstat_l2_evict_l1cached
);
6993 * Invalidate issued or about to be issued
6994 * reads, since we may be about to write
6995 * over this location.
6997 if (HDR_L2_READING(hdr
)) {
6998 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
6999 arc_hdr_set_flags(hdr
, ARC_FLAG_L2_EVICTED
);
7002 arc_hdr_l2hdr_destroy(hdr
);
7004 mutex_exit(hash_lock
);
7006 mutex_exit(&dev
->l2ad_mtx
);
7010 * Find and write ARC buffers to the L2ARC device.
7012 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7013 * for reading until they have completed writing.
7014 * The headroom_boost is an in-out parameter used to maintain headroom boost
7015 * state between calls to this function.
7017 * Returns the number of bytes actually written (which may be smaller than
7018 * the delta by which the device hand has changed due to alignment).
7021 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
)
7023 arc_buf_hdr_t
*hdr
, *hdr_prev
, *head
;
7024 uint64_t write_asize
, write_psize
, write_lsize
, headroom
;
7026 l2arc_write_callback_t
*cb
;
7028 uint64_t guid
= spa_load_guid(spa
);
7030 ASSERT3P(dev
->l2ad_vdev
, !=, NULL
);
7033 write_lsize
= write_asize
= write_psize
= 0;
7035 head
= kmem_cache_alloc(hdr_l2only_cache
, KM_PUSHPAGE
);
7036 arc_hdr_set_flags(head
, ARC_FLAG_L2_WRITE_HEAD
| ARC_FLAG_HAS_L2HDR
);
7039 * Copy buffers for L2ARC writing.
7041 for (int try = 0; try <= 3; try++) {
7042 multilist_sublist_t
*mls
= l2arc_sublist_lock(try);
7043 uint64_t passed_sz
= 0;
7046 * L2ARC fast warmup.
7048 * Until the ARC is warm and starts to evict, read from the
7049 * head of the ARC lists rather than the tail.
7051 if (arc_warm
== B_FALSE
)
7052 hdr
= multilist_sublist_head(mls
);
7054 hdr
= multilist_sublist_tail(mls
);
7056 headroom
= target_sz
* l2arc_headroom
;
7057 if (zfs_compressed_arc_enabled
)
7058 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
7060 for (; hdr
; hdr
= hdr_prev
) {
7061 kmutex_t
*hash_lock
;
7063 if (arc_warm
== B_FALSE
)
7064 hdr_prev
= multilist_sublist_next(mls
, hdr
);
7066 hdr_prev
= multilist_sublist_prev(mls
, hdr
);
7068 hash_lock
= HDR_LOCK(hdr
);
7069 if (!mutex_tryenter(hash_lock
)) {
7071 * Skip this buffer rather than waiting.
7076 passed_sz
+= HDR_GET_LSIZE(hdr
);
7077 if (passed_sz
> headroom
) {
7081 mutex_exit(hash_lock
);
7085 if (!l2arc_write_eligible(guid
, hdr
)) {
7086 mutex_exit(hash_lock
);
7091 * We rely on the L1 portion of the header below, so
7092 * it's invalid for this header to have been evicted out
7093 * of the ghost cache, prior to being written out. The
7094 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7096 ASSERT(HDR_HAS_L1HDR(hdr
));
7098 ASSERT3U(HDR_GET_PSIZE(hdr
), >, 0);
7099 ASSERT3P(hdr
->b_l1hdr
.b_pabd
, !=, NULL
);
7100 ASSERT3U(arc_hdr_size(hdr
), >, 0);
7101 uint64_t psize
= arc_hdr_size(hdr
);
7102 uint64_t asize
= vdev_psize_to_asize(dev
->l2ad_vdev
,
7105 if ((write_asize
+ asize
) > target_sz
) {
7107 mutex_exit(hash_lock
);
7113 * Insert a dummy header on the buflist so
7114 * l2arc_write_done() can find where the
7115 * write buffers begin without searching.
7117 mutex_enter(&dev
->l2ad_mtx
);
7118 list_insert_head(&dev
->l2ad_buflist
, head
);
7119 mutex_exit(&dev
->l2ad_mtx
);
7122 sizeof (l2arc_write_callback_t
), KM_SLEEP
);
7123 cb
->l2wcb_dev
= dev
;
7124 cb
->l2wcb_head
= head
;
7125 pio
= zio_root(spa
, l2arc_write_done
, cb
,
7129 hdr
->b_l2hdr
.b_dev
= dev
;
7130 hdr
->b_l2hdr
.b_daddr
= dev
->l2ad_hand
;
7131 arc_hdr_set_flags(hdr
,
7132 ARC_FLAG_L2_WRITING
| ARC_FLAG_HAS_L2HDR
);
7134 mutex_enter(&dev
->l2ad_mtx
);
7135 list_insert_head(&dev
->l2ad_buflist
, hdr
);
7136 mutex_exit(&dev
->l2ad_mtx
);
7138 (void) refcount_add_many(&dev
->l2ad_alloc
, psize
, hdr
);
7141 * Normally the L2ARC can use the hdr's data, but if
7142 * we're sharing data between the hdr and one of its
7143 * bufs, L2ARC needs its own copy of the data so that
7144 * the ZIO below can't race with the buf consumer.
7145 * Another case where we need to create a copy of the
7146 * data is when the buffer size is not device-aligned
7147 * and we need to pad the block to make it such.
7148 * That also keeps the clock hand suitably aligned.
7150 * To ensure that the copy will be available for the
7151 * lifetime of the ZIO and be cleaned up afterwards, we
7152 * add it to the l2arc_free_on_write queue.
7155 if (!HDR_SHARED_DATA(hdr
) && psize
== asize
) {
7156 to_write
= hdr
->b_l1hdr
.b_pabd
;
7158 to_write
= abd_alloc_for_io(asize
,
7159 HDR_ISTYPE_METADATA(hdr
));
7160 abd_copy(to_write
, hdr
->b_l1hdr
.b_pabd
, psize
);
7161 if (asize
!= psize
) {
7162 abd_zero_off(to_write
, psize
,
7165 l2arc_free_abd_on_write(to_write
, asize
,
7168 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
7169 hdr
->b_l2hdr
.b_daddr
, asize
, to_write
,
7170 ZIO_CHECKSUM_OFF
, NULL
, hdr
,
7171 ZIO_PRIORITY_ASYNC_WRITE
,
7172 ZIO_FLAG_CANFAIL
, B_FALSE
);
7174 write_lsize
+= HDR_GET_LSIZE(hdr
);
7175 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
7178 write_psize
+= psize
;
7179 write_asize
+= asize
;
7180 dev
->l2ad_hand
+= asize
;
7182 mutex_exit(hash_lock
);
7184 (void) zio_nowait(wzio
);
7187 multilist_sublist_unlock(mls
);
7193 /* No buffers selected for writing? */
7195 ASSERT0(write_lsize
);
7196 ASSERT(!HDR_HAS_L1HDR(head
));
7197 kmem_cache_free(hdr_l2only_cache
, head
);
7201 ASSERT3U(write_asize
, <=, target_sz
);
7202 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
7203 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_psize
);
7204 ARCSTAT_INCR(arcstat_l2_lsize
, write_lsize
);
7205 ARCSTAT_INCR(arcstat_l2_psize
, write_psize
);
7206 vdev_space_update(dev
->l2ad_vdev
, write_psize
, 0, 0);
7209 * Bump device hand to the device start if it is approaching the end.
7210 * l2arc_evict() will already have evicted ahead for this case.
7212 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
7213 dev
->l2ad_hand
= dev
->l2ad_start
;
7214 dev
->l2ad_first
= B_FALSE
;
7217 dev
->l2ad_writing
= B_TRUE
;
7218 (void) zio_wait(pio
);
7219 dev
->l2ad_writing
= B_FALSE
;
7221 return (write_asize
);
7225 * This thread feeds the L2ARC at regular intervals. This is the beating
7226 * heart of the L2ARC.
7230 l2arc_feed_thread(void *unused
)
7235 uint64_t size
, wrote
;
7236 clock_t begin
, next
= ddi_get_lbolt();
7238 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
7240 mutex_enter(&l2arc_feed_thr_lock
);
7242 while (l2arc_thread_exit
== 0) {
7243 CALLB_CPR_SAFE_BEGIN(&cpr
);
7244 (void) cv_timedwait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
,
7246 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
7247 next
= ddi_get_lbolt() + hz
;
7250 * Quick check for L2ARC devices.
7252 mutex_enter(&l2arc_dev_mtx
);
7253 if (l2arc_ndev
== 0) {
7254 mutex_exit(&l2arc_dev_mtx
);
7257 mutex_exit(&l2arc_dev_mtx
);
7258 begin
= ddi_get_lbolt();
7261 * This selects the next l2arc device to write to, and in
7262 * doing so the next spa to feed from: dev->l2ad_spa. This
7263 * will return NULL if there are now no l2arc devices or if
7264 * they are all faulted.
7266 * If a device is returned, its spa's config lock is also
7267 * held to prevent device removal. l2arc_dev_get_next()
7268 * will grab and release l2arc_dev_mtx.
7270 if ((dev
= l2arc_dev_get_next()) == NULL
)
7273 spa
= dev
->l2ad_spa
;
7274 ASSERT3P(spa
, !=, NULL
);
7277 * If the pool is read-only then force the feed thread to
7278 * sleep a little longer.
7280 if (!spa_writeable(spa
)) {
7281 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
7282 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7287 * Avoid contributing to memory pressure.
7289 if (arc_reclaim_needed()) {
7290 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
7291 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7295 ARCSTAT_BUMP(arcstat_l2_feeds
);
7297 size
= l2arc_write_size();
7300 * Evict L2ARC buffers that will be overwritten.
7302 l2arc_evict(dev
, size
, B_FALSE
);
7305 * Write ARC buffers.
7307 wrote
= l2arc_write_buffers(spa
, dev
, size
);
7310 * Calculate interval between writes.
7312 next
= l2arc_write_interval(begin
, size
, wrote
);
7313 spa_config_exit(spa
, SCL_L2ARC
, dev
);
7316 l2arc_thread_exit
= 0;
7317 cv_broadcast(&l2arc_feed_thr_cv
);
7318 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
7323 l2arc_vdev_present(vdev_t
*vd
)
7327 mutex_enter(&l2arc_dev_mtx
);
7328 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
7329 dev
= list_next(l2arc_dev_list
, dev
)) {
7330 if (dev
->l2ad_vdev
== vd
)
7333 mutex_exit(&l2arc_dev_mtx
);
7335 return (dev
!= NULL
);
7339 * Add a vdev for use by the L2ARC. By this point the spa has already
7340 * validated the vdev and opened it.
7343 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
7345 l2arc_dev_t
*adddev
;
7347 ASSERT(!l2arc_vdev_present(vd
));
7350 * Create a new l2arc device entry.
7352 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
7353 adddev
->l2ad_spa
= spa
;
7354 adddev
->l2ad_vdev
= vd
;
7355 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
7356 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
7357 adddev
->l2ad_hand
= adddev
->l2ad_start
;
7358 adddev
->l2ad_first
= B_TRUE
;
7359 adddev
->l2ad_writing
= B_FALSE
;
7361 mutex_init(&adddev
->l2ad_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7363 * This is a list of all ARC buffers that are still valid on the
7366 list_create(&adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
7367 offsetof(arc_buf_hdr_t
, b_l2hdr
.b_l2node
));
7369 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
7370 refcount_create(&adddev
->l2ad_alloc
);
7373 * Add device to global list
7375 mutex_enter(&l2arc_dev_mtx
);
7376 list_insert_head(l2arc_dev_list
, adddev
);
7377 atomic_inc_64(&l2arc_ndev
);
7378 mutex_exit(&l2arc_dev_mtx
);
7382 * Remove a vdev from the L2ARC.
7385 l2arc_remove_vdev(vdev_t
*vd
)
7387 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
7390 * Find the device by vdev
7392 mutex_enter(&l2arc_dev_mtx
);
7393 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
7394 nextdev
= list_next(l2arc_dev_list
, dev
);
7395 if (vd
== dev
->l2ad_vdev
) {
7400 ASSERT3P(remdev
, !=, NULL
);
7403 * Remove device from global list
7405 list_remove(l2arc_dev_list
, remdev
);
7406 l2arc_dev_last
= NULL
; /* may have been invalidated */
7407 atomic_dec_64(&l2arc_ndev
);
7408 mutex_exit(&l2arc_dev_mtx
);
7411 * Clear all buflists and ARC references. L2ARC device flush.
7413 l2arc_evict(remdev
, 0, B_TRUE
);
7414 list_destroy(&remdev
->l2ad_buflist
);
7415 mutex_destroy(&remdev
->l2ad_mtx
);
7416 refcount_destroy(&remdev
->l2ad_alloc
);
7417 kmem_free(remdev
, sizeof (l2arc_dev_t
));
7423 l2arc_thread_exit
= 0;
7425 l2arc_writes_sent
= 0;
7426 l2arc_writes_done
= 0;
7428 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
7429 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
7430 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7431 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7433 l2arc_dev_list
= &L2ARC_dev_list
;
7434 l2arc_free_on_write
= &L2ARC_free_on_write
;
7435 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
7436 offsetof(l2arc_dev_t
, l2ad_node
));
7437 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
7438 offsetof(l2arc_data_free_t
, l2df_list_node
));
7445 * This is called from dmu_fini(), which is called from spa_fini();
7446 * Because of this, we can assume that all l2arc devices have
7447 * already been removed when the pools themselves were removed.
7450 l2arc_do_free_on_write();
7452 mutex_destroy(&l2arc_feed_thr_lock
);
7453 cv_destroy(&l2arc_feed_thr_cv
);
7454 mutex_destroy(&l2arc_dev_mtx
);
7455 mutex_destroy(&l2arc_free_on_write_mtx
);
7457 list_destroy(l2arc_dev_list
);
7458 list_destroy(l2arc_free_on_write
);
7464 if (!(spa_mode_global
& FWRITE
))
7467 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
7468 TS_RUN
, minclsyspri
);
7474 if (!(spa_mode_global
& FWRITE
))
7477 mutex_enter(&l2arc_feed_thr_lock
);
7478 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
7479 l2arc_thread_exit
= 1;
7480 while (l2arc_thread_exit
!= 0)
7481 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
7482 mutex_exit(&l2arc_feed_thr_lock
);