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, 2017 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/dnlc.h>
271 #include <sys/racct.h>
273 #include <sys/callb.h>
274 #include <sys/kstat.h>
275 #include <sys/trim_map.h>
276 #include <zfs_fletcher.h>
278 #include <sys/aggsum.h>
279 #include <sys/cityhash.h>
281 #include <machine/vmparam.h>
285 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
286 boolean_t arc_watch = B_FALSE;
291 static kmutex_t arc_reclaim_lock;
292 static kcondvar_t arc_reclaim_thread_cv;
293 static boolean_t arc_reclaim_thread_exit;
294 static kcondvar_t arc_reclaim_waiters_cv;
296 static kmutex_t arc_dnlc_evicts_lock;
297 static kcondvar_t arc_dnlc_evicts_cv;
298 static boolean_t arc_dnlc_evicts_thread_exit;
300 uint_t arc_reduce_dnlc_percent = 3;
303 * The number of headers to evict in arc_evict_state_impl() before
304 * dropping the sublist lock and evicting from another sublist. A lower
305 * value means we're more likely to evict the "correct" header (i.e. the
306 * oldest header in the arc state), but comes with higher overhead
307 * (i.e. more invocations of arc_evict_state_impl()).
309 int zfs_arc_evict_batch_limit = 10;
311 /* number of seconds before growing cache again */
312 static int arc_grow_retry = 60;
314 /* number of milliseconds before attempting a kmem-cache-reap */
315 static int arc_kmem_cache_reap_retry_ms = 1000;
317 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
318 int zfs_arc_overflow_shift = 8;
320 /* shift of arc_c for calculating both min and max arc_p */
321 static int arc_p_min_shift = 4;
323 /* log2(fraction of arc to reclaim) */
324 static int arc_shrink_shift = 7;
327 * log2(fraction of ARC which must be free to allow growing).
328 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
329 * when reading a new block into the ARC, we will evict an equal-sized block
332 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
333 * we will still not allow it to grow.
335 int arc_no_grow_shift = 5;
339 * minimum lifespan of a prefetch block in clock ticks
340 * (initialized in arc_init())
342 static int arc_min_prefetch_lifespan;
345 * If this percent of memory is free, don't throttle.
347 int arc_lotsfree_percent = 10;
350 extern boolean_t zfs_prefetch_disable;
353 * The arc has filled available memory and has now warmed up.
355 static boolean_t arc_warm;
358 * log2 fraction of the zio arena to keep free.
360 int arc_zio_arena_free_shift = 2;
363 * These tunables are for performance analysis.
365 uint64_t zfs_arc_max;
366 uint64_t zfs_arc_min;
367 uint64_t zfs_arc_meta_limit = 0;
368 uint64_t zfs_arc_meta_min = 0;
369 int zfs_arc_grow_retry = 0;
370 int zfs_arc_shrink_shift = 0;
371 int zfs_arc_no_grow_shift = 0;
372 int zfs_arc_p_min_shift = 0;
373 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
374 u_int zfs_arc_free_target = 0;
376 /* Absolute min for arc min / max is 16MB. */
377 static uint64_t arc_abs_min = 16 << 20;
379 boolean_t zfs_compressed_arc_enabled = B_TRUE;
381 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
382 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
383 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
384 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
385 static int sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS);
387 #if defined(__FreeBSD__) && defined(_KERNEL)
389 arc_free_target_init(void *unused __unused)
392 zfs_arc_free_target = vm_pageout_wakeup_thresh;
394 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
395 arc_free_target_init, NULL);
397 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
398 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
399 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
400 TUNABLE_INT("vfs.zfs.arc_grow_retry", &zfs_arc_grow_retry);
401 TUNABLE_INT("vfs.zfs.arc_no_grow_shift", &zfs_arc_no_grow_shift);
402 SYSCTL_DECL(_vfs_zfs);
403 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
404 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
405 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
406 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
407 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_no_grow_shift, CTLTYPE_U32 | CTLFLAG_RWTUN,
408 0, sizeof(uint32_t), sysctl_vfs_zfs_arc_no_grow_shift, "U",
409 "log2(fraction of ARC which must be free to allow growing)");
410 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
411 &zfs_arc_average_blocksize, 0,
412 "ARC average blocksize");
413 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
414 &arc_shrink_shift, 0,
415 "log2(fraction of arc to reclaim)");
416 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_grow_retry, CTLFLAG_RW,
418 "Wait in seconds before considering growing ARC");
419 SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN,
420 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC");
423 * We don't have a tunable for arc_free_target due to the dependency on
424 * pagedaemon initialisation.
426 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
427 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
428 sysctl_vfs_zfs_arc_free_target, "IU",
429 "Desired number of free pages below which ARC triggers reclaim");
432 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
437 val = zfs_arc_free_target;
438 err = sysctl_handle_int(oidp, &val, 0, req);
439 if (err != 0 || req->newptr == NULL)
444 if (val > vm_cnt.v_page_count)
447 zfs_arc_free_target = val;
453 * Must be declared here, before the definition of corresponding kstat
454 * macro which uses the same names will confuse the compiler.
456 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
457 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
458 sysctl_vfs_zfs_arc_meta_limit, "QU",
459 "ARC metadata limit");
463 * Note that buffers can be in one of 6 states:
464 * ARC_anon - anonymous (discussed below)
465 * ARC_mru - recently used, currently cached
466 * ARC_mru_ghost - recentely used, no longer in cache
467 * ARC_mfu - frequently used, currently cached
468 * ARC_mfu_ghost - frequently used, no longer in cache
469 * ARC_l2c_only - exists in L2ARC but not other states
470 * When there are no active references to the buffer, they are
471 * are linked onto a list in one of these arc states. These are
472 * the only buffers that can be evicted or deleted. Within each
473 * state there are multiple lists, one for meta-data and one for
474 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
475 * etc.) is tracked separately so that it can be managed more
476 * explicitly: favored over data, limited explicitly.
478 * Anonymous buffers are buffers that are not associated with
479 * a DVA. These are buffers that hold dirty block copies
480 * before they are written to stable storage. By definition,
481 * they are "ref'd" and are considered part of arc_mru
482 * that cannot be freed. Generally, they will aquire a DVA
483 * as they are written and migrate onto the arc_mru list.
485 * The ARC_l2c_only state is for buffers that are in the second
486 * level ARC but no longer in any of the ARC_m* lists. The second
487 * level ARC itself may also contain buffers that are in any of
488 * the ARC_m* states - meaning that a buffer can exist in two
489 * places. The reason for the ARC_l2c_only state is to keep the
490 * buffer header in the hash table, so that reads that hit the
491 * second level ARC benefit from these fast lookups.
494 typedef struct arc_state {
496 * list of evictable buffers
498 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
500 * total amount of evictable data in this state
502 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
504 * total amount of data in this state; this includes: evictable,
505 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
507 refcount_t arcs_size;
511 static arc_state_t ARC_anon;
512 static arc_state_t ARC_mru;
513 static arc_state_t ARC_mru_ghost;
514 static arc_state_t ARC_mfu;
515 static arc_state_t ARC_mfu_ghost;
516 static arc_state_t ARC_l2c_only;
518 typedef struct arc_stats {
519 kstat_named_t arcstat_hits;
520 kstat_named_t arcstat_misses;
521 kstat_named_t arcstat_demand_data_hits;
522 kstat_named_t arcstat_demand_data_misses;
523 kstat_named_t arcstat_demand_metadata_hits;
524 kstat_named_t arcstat_demand_metadata_misses;
525 kstat_named_t arcstat_prefetch_data_hits;
526 kstat_named_t arcstat_prefetch_data_misses;
527 kstat_named_t arcstat_prefetch_metadata_hits;
528 kstat_named_t arcstat_prefetch_metadata_misses;
529 kstat_named_t arcstat_mru_hits;
530 kstat_named_t arcstat_mru_ghost_hits;
531 kstat_named_t arcstat_mfu_hits;
532 kstat_named_t arcstat_mfu_ghost_hits;
533 kstat_named_t arcstat_allocated;
534 kstat_named_t arcstat_deleted;
536 * Number of buffers that could not be evicted because the hash lock
537 * was held by another thread. The lock may not necessarily be held
538 * by something using the same buffer, since hash locks are shared
539 * by multiple buffers.
541 kstat_named_t arcstat_mutex_miss;
543 * Number of buffers skipped because they have I/O in progress, are
544 * indrect prefetch buffers that have not lived long enough, or are
545 * not from the spa we're trying to evict from.
547 kstat_named_t arcstat_evict_skip;
549 * Number of times arc_evict_state() was unable to evict enough
550 * buffers to reach it's target amount.
552 kstat_named_t arcstat_evict_not_enough;
553 kstat_named_t arcstat_evict_l2_cached;
554 kstat_named_t arcstat_evict_l2_eligible;
555 kstat_named_t arcstat_evict_l2_ineligible;
556 kstat_named_t arcstat_evict_l2_skip;
557 kstat_named_t arcstat_hash_elements;
558 kstat_named_t arcstat_hash_elements_max;
559 kstat_named_t arcstat_hash_collisions;
560 kstat_named_t arcstat_hash_chains;
561 kstat_named_t arcstat_hash_chain_max;
562 kstat_named_t arcstat_p;
563 kstat_named_t arcstat_c;
564 kstat_named_t arcstat_c_min;
565 kstat_named_t arcstat_c_max;
566 /* Not updated directly; only synced in arc_kstat_update. */
567 kstat_named_t arcstat_size;
569 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
570 * Note that the compressed bytes may match the uncompressed bytes
571 * if the block is either not compressed or compressed arc is disabled.
573 kstat_named_t arcstat_compressed_size;
575 * Uncompressed size of the data stored in b_pabd. If compressed
576 * arc is disabled then this value will be identical to the stat
579 kstat_named_t arcstat_uncompressed_size;
581 * Number of bytes stored in all the arc_buf_t's. This is classified
582 * as "overhead" since this data is typically short-lived and will
583 * be evicted from the arc when it becomes unreferenced unless the
584 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
585 * values have been set (see comment in dbuf.c for more information).
587 kstat_named_t arcstat_overhead_size;
589 * Number of bytes consumed by internal ARC structures necessary
590 * for tracking purposes; these structures are not actually
591 * backed by ARC buffers. This includes arc_buf_hdr_t structures
592 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
593 * caches), and arc_buf_t structures (allocated via arc_buf_t
595 * Not updated directly; only synced in arc_kstat_update.
597 kstat_named_t arcstat_hdr_size;
599 * Number of bytes consumed by ARC buffers of type equal to
600 * ARC_BUFC_DATA. This is generally consumed by buffers backing
601 * on disk user data (e.g. plain file contents).
602 * Not updated directly; only synced in arc_kstat_update.
604 kstat_named_t arcstat_data_size;
606 * Number of bytes consumed by ARC buffers of type equal to
607 * ARC_BUFC_METADATA. This is generally consumed by buffers
608 * backing on disk data that is used for internal ZFS
609 * structures (e.g. ZAP, dnode, indirect blocks, etc).
610 * Not updated directly; only synced in arc_kstat_update.
612 kstat_named_t arcstat_metadata_size;
614 * Number of bytes consumed by various buffers and structures
615 * not actually backed with ARC buffers. This includes bonus
616 * buffers (allocated directly via zio_buf_* functions),
617 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
618 * cache), and dnode_t structures (allocated via dnode_t cache).
619 * Not updated directly; only synced in arc_kstat_update.
621 kstat_named_t arcstat_other_size;
623 * Total number of bytes consumed by ARC buffers residing in the
624 * arc_anon state. This includes *all* buffers in the arc_anon
625 * state; e.g. data, metadata, evictable, and unevictable buffers
626 * are all included in this value.
627 * Not updated directly; only synced in arc_kstat_update.
629 kstat_named_t arcstat_anon_size;
631 * Number of bytes consumed by ARC buffers that meet the
632 * following criteria: backing buffers of type ARC_BUFC_DATA,
633 * residing in the arc_anon state, and are eligible for eviction
634 * (e.g. have no outstanding holds on the buffer).
635 * Not updated directly; only synced in arc_kstat_update.
637 kstat_named_t arcstat_anon_evictable_data;
639 * Number of bytes consumed by ARC buffers that meet the
640 * following criteria: backing buffers of type ARC_BUFC_METADATA,
641 * residing in the arc_anon state, and are eligible for eviction
642 * (e.g. have no outstanding holds on the buffer).
643 * Not updated directly; only synced in arc_kstat_update.
645 kstat_named_t arcstat_anon_evictable_metadata;
647 * Total number of bytes consumed by ARC buffers residing in the
648 * arc_mru state. This includes *all* buffers in the arc_mru
649 * state; e.g. data, metadata, evictable, and unevictable buffers
650 * are all included in this value.
651 * Not updated directly; only synced in arc_kstat_update.
653 kstat_named_t arcstat_mru_size;
655 * Number of bytes consumed by ARC buffers that meet the
656 * following criteria: backing buffers of type ARC_BUFC_DATA,
657 * residing in the arc_mru state, and are eligible for eviction
658 * (e.g. have no outstanding holds on the buffer).
659 * Not updated directly; only synced in arc_kstat_update.
661 kstat_named_t arcstat_mru_evictable_data;
663 * Number of bytes consumed by ARC buffers that meet the
664 * following criteria: backing buffers of type ARC_BUFC_METADATA,
665 * residing in the arc_mru state, and are eligible for eviction
666 * (e.g. have no outstanding holds on the buffer).
667 * Not updated directly; only synced in arc_kstat_update.
669 kstat_named_t arcstat_mru_evictable_metadata;
671 * Total number of bytes that *would have been* consumed by ARC
672 * buffers in the arc_mru_ghost state. The key thing to note
673 * here, is the fact that this size doesn't actually indicate
674 * RAM consumption. The ghost lists only consist of headers and
675 * don't actually have ARC buffers linked off of these headers.
676 * Thus, *if* the headers had associated ARC buffers, these
677 * buffers *would have* consumed this number of bytes.
678 * Not updated directly; only synced in arc_kstat_update.
680 kstat_named_t arcstat_mru_ghost_size;
682 * Number of bytes that *would have been* consumed by ARC
683 * buffers that are eligible for eviction, of type
684 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
685 * Not updated directly; only synced in arc_kstat_update.
687 kstat_named_t arcstat_mru_ghost_evictable_data;
689 * Number of bytes that *would have been* consumed by ARC
690 * buffers that are eligible for eviction, of type
691 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
692 * Not updated directly; only synced in arc_kstat_update.
694 kstat_named_t arcstat_mru_ghost_evictable_metadata;
696 * Total number of bytes consumed by ARC buffers residing in the
697 * arc_mfu state. This includes *all* buffers in the arc_mfu
698 * state; e.g. data, metadata, evictable, and unevictable buffers
699 * are all included in this value.
700 * Not updated directly; only synced in arc_kstat_update.
702 kstat_named_t arcstat_mfu_size;
704 * Number of bytes consumed by ARC buffers that are eligible for
705 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
707 * Not updated directly; only synced in arc_kstat_update.
709 kstat_named_t arcstat_mfu_evictable_data;
711 * Number of bytes consumed by ARC buffers that are eligible for
712 * eviction, of type ARC_BUFC_METADATA, and reside in the
714 * Not updated directly; only synced in arc_kstat_update.
716 kstat_named_t arcstat_mfu_evictable_metadata;
718 * Total number of bytes that *would have been* consumed by ARC
719 * buffers in the arc_mfu_ghost state. See the comment above
720 * arcstat_mru_ghost_size for more details.
721 * Not updated directly; only synced in arc_kstat_update.
723 kstat_named_t arcstat_mfu_ghost_size;
725 * Number of bytes that *would have been* consumed by ARC
726 * buffers that are eligible for eviction, of type
727 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
728 * Not updated directly; only synced in arc_kstat_update.
730 kstat_named_t arcstat_mfu_ghost_evictable_data;
732 * Number of bytes that *would have been* consumed by ARC
733 * buffers that are eligible for eviction, of type
734 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
735 * Not updated directly; only synced in arc_kstat_update.
737 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
738 kstat_named_t arcstat_l2_hits;
739 kstat_named_t arcstat_l2_misses;
740 kstat_named_t arcstat_l2_feeds;
741 kstat_named_t arcstat_l2_rw_clash;
742 kstat_named_t arcstat_l2_read_bytes;
743 kstat_named_t arcstat_l2_write_bytes;
744 kstat_named_t arcstat_l2_writes_sent;
745 kstat_named_t arcstat_l2_writes_done;
746 kstat_named_t arcstat_l2_writes_error;
747 kstat_named_t arcstat_l2_writes_lock_retry;
748 kstat_named_t arcstat_l2_evict_lock_retry;
749 kstat_named_t arcstat_l2_evict_reading;
750 kstat_named_t arcstat_l2_evict_l1cached;
751 kstat_named_t arcstat_l2_free_on_write;
752 kstat_named_t arcstat_l2_abort_lowmem;
753 kstat_named_t arcstat_l2_cksum_bad;
754 kstat_named_t arcstat_l2_io_error;
755 kstat_named_t arcstat_l2_lsize;
756 kstat_named_t arcstat_l2_psize;
757 /* Not updated directly; only synced in arc_kstat_update. */
758 kstat_named_t arcstat_l2_hdr_size;
759 kstat_named_t arcstat_l2_write_trylock_fail;
760 kstat_named_t arcstat_l2_write_passed_headroom;
761 kstat_named_t arcstat_l2_write_spa_mismatch;
762 kstat_named_t arcstat_l2_write_in_l2;
763 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
764 kstat_named_t arcstat_l2_write_not_cacheable;
765 kstat_named_t arcstat_l2_write_full;
766 kstat_named_t arcstat_l2_write_buffer_iter;
767 kstat_named_t arcstat_l2_write_pios;
768 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
769 kstat_named_t arcstat_l2_write_buffer_list_iter;
770 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
771 kstat_named_t arcstat_memory_throttle_count;
772 /* Not updated directly; only synced in arc_kstat_update. */
773 kstat_named_t arcstat_meta_used;
774 kstat_named_t arcstat_meta_limit;
775 kstat_named_t arcstat_meta_max;
776 kstat_named_t arcstat_meta_min;
777 kstat_named_t arcstat_sync_wait_for_async;
778 kstat_named_t arcstat_demand_hit_predictive_prefetch;
781 static arc_stats_t arc_stats = {
782 { "hits", KSTAT_DATA_UINT64 },
783 { "misses", KSTAT_DATA_UINT64 },
784 { "demand_data_hits", KSTAT_DATA_UINT64 },
785 { "demand_data_misses", KSTAT_DATA_UINT64 },
786 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
787 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
788 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
789 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
790 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
791 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
792 { "mru_hits", KSTAT_DATA_UINT64 },
793 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
794 { "mfu_hits", KSTAT_DATA_UINT64 },
795 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
796 { "allocated", KSTAT_DATA_UINT64 },
797 { "deleted", KSTAT_DATA_UINT64 },
798 { "mutex_miss", KSTAT_DATA_UINT64 },
799 { "evict_skip", KSTAT_DATA_UINT64 },
800 { "evict_not_enough", KSTAT_DATA_UINT64 },
801 { "evict_l2_cached", KSTAT_DATA_UINT64 },
802 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
803 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
804 { "evict_l2_skip", KSTAT_DATA_UINT64 },
805 { "hash_elements", KSTAT_DATA_UINT64 },
806 { "hash_elements_max", KSTAT_DATA_UINT64 },
807 { "hash_collisions", KSTAT_DATA_UINT64 },
808 { "hash_chains", KSTAT_DATA_UINT64 },
809 { "hash_chain_max", KSTAT_DATA_UINT64 },
810 { "p", KSTAT_DATA_UINT64 },
811 { "c", KSTAT_DATA_UINT64 },
812 { "c_min", KSTAT_DATA_UINT64 },
813 { "c_max", KSTAT_DATA_UINT64 },
814 { "size", KSTAT_DATA_UINT64 },
815 { "compressed_size", KSTAT_DATA_UINT64 },
816 { "uncompressed_size", KSTAT_DATA_UINT64 },
817 { "overhead_size", KSTAT_DATA_UINT64 },
818 { "hdr_size", KSTAT_DATA_UINT64 },
819 { "data_size", KSTAT_DATA_UINT64 },
820 { "metadata_size", KSTAT_DATA_UINT64 },
821 { "other_size", KSTAT_DATA_UINT64 },
822 { "anon_size", KSTAT_DATA_UINT64 },
823 { "anon_evictable_data", KSTAT_DATA_UINT64 },
824 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
825 { "mru_size", KSTAT_DATA_UINT64 },
826 { "mru_evictable_data", KSTAT_DATA_UINT64 },
827 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
828 { "mru_ghost_size", KSTAT_DATA_UINT64 },
829 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
830 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
831 { "mfu_size", KSTAT_DATA_UINT64 },
832 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
833 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
834 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
835 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
836 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
837 { "l2_hits", KSTAT_DATA_UINT64 },
838 { "l2_misses", KSTAT_DATA_UINT64 },
839 { "l2_feeds", KSTAT_DATA_UINT64 },
840 { "l2_rw_clash", KSTAT_DATA_UINT64 },
841 { "l2_read_bytes", KSTAT_DATA_UINT64 },
842 { "l2_write_bytes", KSTAT_DATA_UINT64 },
843 { "l2_writes_sent", KSTAT_DATA_UINT64 },
844 { "l2_writes_done", KSTAT_DATA_UINT64 },
845 { "l2_writes_error", KSTAT_DATA_UINT64 },
846 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
847 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
848 { "l2_evict_reading", KSTAT_DATA_UINT64 },
849 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
850 { "l2_free_on_write", KSTAT_DATA_UINT64 },
851 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
852 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
853 { "l2_io_error", KSTAT_DATA_UINT64 },
854 { "l2_size", KSTAT_DATA_UINT64 },
855 { "l2_asize", KSTAT_DATA_UINT64 },
856 { "l2_hdr_size", KSTAT_DATA_UINT64 },
857 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
858 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
859 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
860 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
861 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
862 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
863 { "l2_write_full", KSTAT_DATA_UINT64 },
864 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
865 { "l2_write_pios", KSTAT_DATA_UINT64 },
866 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
867 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
868 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
869 { "memory_throttle_count", KSTAT_DATA_UINT64 },
870 { "arc_meta_used", KSTAT_DATA_UINT64 },
871 { "arc_meta_limit", KSTAT_DATA_UINT64 },
872 { "arc_meta_max", KSTAT_DATA_UINT64 },
873 { "arc_meta_min", KSTAT_DATA_UINT64 },
874 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
875 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
878 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
880 #define ARCSTAT_INCR(stat, val) \
881 atomic_add_64(&arc_stats.stat.value.ui64, (val))
883 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
884 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
886 #define ARCSTAT_MAX(stat, val) { \
888 while ((val) > (m = arc_stats.stat.value.ui64) && \
889 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
893 #define ARCSTAT_MAXSTAT(stat) \
894 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
897 * We define a macro to allow ARC hits/misses to be easily broken down by
898 * two separate conditions, giving a total of four different subtypes for
899 * each of hits and misses (so eight statistics total).
901 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
904 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
906 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
910 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
912 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
917 static arc_state_t *arc_anon;
918 static arc_state_t *arc_mru;
919 static arc_state_t *arc_mru_ghost;
920 static arc_state_t *arc_mfu;
921 static arc_state_t *arc_mfu_ghost;
922 static arc_state_t *arc_l2c_only;
925 * There are several ARC variables that are critical to export as kstats --
926 * but we don't want to have to grovel around in the kstat whenever we wish to
927 * manipulate them. For these variables, we therefore define them to be in
928 * terms of the statistic variable. This assures that we are not introducing
929 * the possibility of inconsistency by having shadow copies of the variables,
930 * while still allowing the code to be readable.
932 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
933 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
934 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
935 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
936 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
937 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
938 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
940 /* compressed size of entire arc */
941 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
942 /* uncompressed size of entire arc */
943 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
944 /* number of bytes in the arc from arc_buf_t's */
945 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
948 * There are also some ARC variables that we want to export, but that are
949 * updated so often that having the canonical representation be the statistic
950 * variable causes a performance bottleneck. We want to use aggsum_t's for these
951 * instead, but still be able to export the kstat in the same way as before.
952 * The solution is to always use the aggsum version, except in the kstat update
956 aggsum_t arc_meta_used;
957 aggsum_t astat_data_size;
958 aggsum_t astat_metadata_size;
959 aggsum_t astat_hdr_size;
960 aggsum_t astat_other_size;
961 aggsum_t astat_l2_hdr_size;
963 static int arc_no_grow; /* Don't try to grow cache size */
964 static uint64_t arc_tempreserve;
965 static uint64_t arc_loaned_bytes;
967 typedef struct arc_callback arc_callback_t;
969 struct arc_callback {
971 arc_done_func_t *acb_done;
973 boolean_t acb_compressed;
974 zio_t *acb_zio_dummy;
975 arc_callback_t *acb_next;
978 typedef struct arc_write_callback arc_write_callback_t;
980 struct arc_write_callback {
982 arc_done_func_t *awcb_ready;
983 arc_done_func_t *awcb_children_ready;
984 arc_done_func_t *awcb_physdone;
985 arc_done_func_t *awcb_done;
990 * ARC buffers are separated into multiple structs as a memory saving measure:
991 * - Common fields struct, always defined, and embedded within it:
992 * - L2-only fields, always allocated but undefined when not in L2ARC
993 * - L1-only fields, only allocated when in L1ARC
995 * Buffer in L1 Buffer only in L2
996 * +------------------------+ +------------------------+
997 * | arc_buf_hdr_t | | arc_buf_hdr_t |
1001 * +------------------------+ +------------------------+
1002 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
1003 * | (undefined if L1-only) | | |
1004 * +------------------------+ +------------------------+
1005 * | l1arc_buf_hdr_t |
1010 * +------------------------+
1012 * Because it's possible for the L2ARC to become extremely large, we can wind
1013 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
1014 * is minimized by only allocating the fields necessary for an L1-cached buffer
1015 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
1016 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
1017 * words in pointers. arc_hdr_realloc() is used to switch a header between
1018 * these two allocation states.
1020 typedef struct l1arc_buf_hdr {
1021 kmutex_t b_freeze_lock;
1022 zio_cksum_t *b_freeze_cksum;
1025 * Used for debugging with kmem_flags - by allocating and freeing
1026 * b_thawed when the buffer is thawed, we get a record of the stack
1027 * trace that thawed it.
1034 /* for waiting on writes to complete */
1038 /* protected by arc state mutex */
1039 arc_state_t *b_state;
1040 multilist_node_t b_arc_node;
1042 /* updated atomically */
1043 clock_t b_arc_access;
1045 /* self protecting */
1046 refcount_t b_refcnt;
1048 arc_callback_t *b_acb;
1052 typedef struct l2arc_dev l2arc_dev_t;
1054 typedef struct l2arc_buf_hdr {
1055 /* protected by arc_buf_hdr mutex */
1056 l2arc_dev_t *b_dev; /* L2ARC device */
1057 uint64_t b_daddr; /* disk address, offset byte */
1059 list_node_t b_l2node;
1062 struct arc_buf_hdr {
1063 /* protected by hash lock */
1067 arc_buf_contents_t b_type;
1068 arc_buf_hdr_t *b_hash_next;
1069 arc_flags_t b_flags;
1072 * This field stores the size of the data buffer after
1073 * compression, and is set in the arc's zio completion handlers.
1074 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1076 * While the block pointers can store up to 32MB in their psize
1077 * field, we can only store up to 32MB minus 512B. This is due
1078 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1079 * a field of zeros represents 512B in the bp). We can't use a
1080 * bias of 1 since we need to reserve a psize of zero, here, to
1081 * represent holes and embedded blocks.
1083 * This isn't a problem in practice, since the maximum size of a
1084 * buffer is limited to 16MB, so we never need to store 32MB in
1085 * this field. Even in the upstream illumos code base, the
1086 * maximum size of a buffer is limited to 16MB.
1091 * This field stores the size of the data buffer before
1092 * compression, and cannot change once set. It is in units
1093 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1095 uint16_t b_lsize; /* immutable */
1096 uint64_t b_spa; /* immutable */
1098 /* L2ARC fields. Undefined when not in L2ARC. */
1099 l2arc_buf_hdr_t b_l2hdr;
1100 /* L1ARC fields. Undefined when in l2arc_only state */
1101 l1arc_buf_hdr_t b_l1hdr;
1104 #if defined(__FreeBSD__) && defined(_KERNEL)
1106 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1111 val = arc_meta_limit;
1112 err = sysctl_handle_64(oidp, &val, 0, req);
1113 if (err != 0 || req->newptr == NULL)
1116 if (val <= 0 || val > arc_c_max)
1119 arc_meta_limit = val;
1124 sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS)
1129 val = arc_no_grow_shift;
1130 err = sysctl_handle_32(oidp, &val, 0, req);
1131 if (err != 0 || req->newptr == NULL)
1134 if (val >= arc_shrink_shift)
1137 arc_no_grow_shift = val;
1142 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1148 err = sysctl_handle_64(oidp, &val, 0, req);
1149 if (err != 0 || req->newptr == NULL)
1152 if (zfs_arc_max == 0) {
1153 /* Loader tunable so blindly set */
1158 if (val < arc_abs_min || val > kmem_size())
1160 if (val < arc_c_min)
1162 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1168 arc_p = (arc_c >> 1);
1170 if (zfs_arc_meta_limit == 0) {
1171 /* limit meta-data to 1/4 of the arc capacity */
1172 arc_meta_limit = arc_c_max / 4;
1175 /* if kmem_flags are set, lets try to use less memory */
1176 if (kmem_debugging())
1179 zfs_arc_max = arc_c;
1185 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1191 err = sysctl_handle_64(oidp, &val, 0, req);
1192 if (err != 0 || req->newptr == NULL)
1195 if (zfs_arc_min == 0) {
1196 /* Loader tunable so blindly set */
1201 if (val < arc_abs_min || val > arc_c_max)
1206 if (zfs_arc_meta_min == 0)
1207 arc_meta_min = arc_c_min / 2;
1209 if (arc_c < arc_c_min)
1212 zfs_arc_min = arc_c_min;
1218 #define GHOST_STATE(state) \
1219 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1220 (state) == arc_l2c_only)
1222 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1223 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1224 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1225 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1226 #define HDR_COMPRESSION_ENABLED(hdr) \
1227 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1229 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1230 #define HDR_L2_READING(hdr) \
1231 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1232 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1233 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1234 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1235 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1236 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1238 #define HDR_ISTYPE_METADATA(hdr) \
1239 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1240 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1242 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1243 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1245 /* For storing compression mode in b_flags */
1246 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1248 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1249 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1250 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1251 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1253 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1254 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1255 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1261 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1262 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1265 * Hash table routines
1268 #define HT_LOCK_PAD CACHE_LINE_SIZE
1273 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1277 #define BUF_LOCKS 256
1278 typedef struct buf_hash_table {
1280 arc_buf_hdr_t **ht_table;
1281 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1284 static buf_hash_table_t buf_hash_table;
1286 #define BUF_HASH_INDEX(spa, dva, birth) \
1287 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1288 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1289 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1290 #define HDR_LOCK(hdr) \
1291 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1293 uint64_t zfs_crc64_table[256];
1299 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1300 #define L2ARC_HEADROOM 2 /* num of writes */
1302 * If we discover during ARC scan any buffers to be compressed, we boost
1303 * our headroom for the next scanning cycle by this percentage multiple.
1305 #define L2ARC_HEADROOM_BOOST 200
1306 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1307 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1309 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1310 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1312 /* L2ARC Performance Tunables */
1313 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1314 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1315 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1316 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1317 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1318 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1319 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1320 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1321 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1323 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1324 &l2arc_write_max, 0, "max write size");
1325 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1326 &l2arc_write_boost, 0, "extra write during warmup");
1327 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1328 &l2arc_headroom, 0, "number of dev writes");
1329 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1330 &l2arc_feed_secs, 0, "interval seconds");
1331 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1332 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1334 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1335 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1336 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1337 &l2arc_feed_again, 0, "turbo warmup");
1338 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1339 &l2arc_norw, 0, "no reads during writes");
1341 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1342 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1343 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1344 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1345 "size of anonymous state");
1346 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1347 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1348 "size of anonymous state");
1350 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1351 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1352 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1353 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1354 "size of metadata in mru state");
1355 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1356 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1357 "size of data in mru state");
1359 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1360 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1361 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1362 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1363 "size of metadata in mru ghost state");
1364 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1365 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1366 "size of data in mru ghost state");
1368 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1369 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1370 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1371 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1372 "size of metadata in mfu state");
1373 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1374 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1375 "size of data in mfu state");
1377 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1378 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1379 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1380 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1381 "size of metadata in mfu ghost state");
1382 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1383 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1384 "size of data in mfu ghost state");
1386 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1387 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1393 vdev_t *l2ad_vdev; /* vdev */
1394 spa_t *l2ad_spa; /* spa */
1395 uint64_t l2ad_hand; /* next write location */
1396 uint64_t l2ad_start; /* first addr on device */
1397 uint64_t l2ad_end; /* last addr on device */
1398 boolean_t l2ad_first; /* first sweep through */
1399 boolean_t l2ad_writing; /* currently writing */
1400 kmutex_t l2ad_mtx; /* lock for buffer list */
1401 list_t l2ad_buflist; /* buffer list */
1402 list_node_t l2ad_node; /* device list node */
1403 refcount_t l2ad_alloc; /* allocated bytes */
1406 static list_t L2ARC_dev_list; /* device list */
1407 static list_t *l2arc_dev_list; /* device list pointer */
1408 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1409 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1410 static list_t L2ARC_free_on_write; /* free after write buf list */
1411 static list_t *l2arc_free_on_write; /* free after write list ptr */
1412 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1413 static uint64_t l2arc_ndev; /* number of devices */
1415 typedef struct l2arc_read_callback {
1416 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1417 blkptr_t l2rcb_bp; /* original blkptr */
1418 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1419 int l2rcb_flags; /* original flags */
1420 abd_t *l2rcb_abd; /* temporary buffer */
1421 } l2arc_read_callback_t;
1423 typedef struct l2arc_write_callback {
1424 l2arc_dev_t *l2wcb_dev; /* device info */
1425 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1426 } l2arc_write_callback_t;
1428 typedef struct l2arc_data_free {
1429 /* protected by l2arc_free_on_write_mtx */
1432 arc_buf_contents_t l2df_type;
1433 list_node_t l2df_list_node;
1434 } l2arc_data_free_t;
1436 static kmutex_t l2arc_feed_thr_lock;
1437 static kcondvar_t l2arc_feed_thr_cv;
1438 static uint8_t l2arc_thread_exit;
1440 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1441 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1442 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1443 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1444 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1445 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1446 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1447 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1448 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1449 static boolean_t arc_is_overflowing();
1450 static void arc_buf_watch(arc_buf_t *);
1452 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1453 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1454 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1455 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1457 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1458 static void l2arc_read_done(zio_t *);
1461 l2arc_trim(const arc_buf_hdr_t *hdr)
1463 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1465 ASSERT(HDR_HAS_L2HDR(hdr));
1466 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1468 if (HDR_GET_PSIZE(hdr) != 0) {
1469 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1470 HDR_GET_PSIZE(hdr), 0);
1475 * We use Cityhash for this. It's fast, and has good hash properties without
1476 * requiring any large static buffers.
1479 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1481 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1484 #define HDR_EMPTY(hdr) \
1485 ((hdr)->b_dva.dva_word[0] == 0 && \
1486 (hdr)->b_dva.dva_word[1] == 0)
1488 #define HDR_EQUAL(spa, dva, birth, hdr) \
1489 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1490 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1491 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1494 buf_discard_identity(arc_buf_hdr_t *hdr)
1496 hdr->b_dva.dva_word[0] = 0;
1497 hdr->b_dva.dva_word[1] = 0;
1501 static arc_buf_hdr_t *
1502 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1504 const dva_t *dva = BP_IDENTITY(bp);
1505 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1506 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1507 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1510 mutex_enter(hash_lock);
1511 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1512 hdr = hdr->b_hash_next) {
1513 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1518 mutex_exit(hash_lock);
1524 * Insert an entry into the hash table. If there is already an element
1525 * equal to elem in the hash table, then the already existing element
1526 * will be returned and the new element will not be inserted.
1527 * Otherwise returns NULL.
1528 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1530 static arc_buf_hdr_t *
1531 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1533 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1534 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1535 arc_buf_hdr_t *fhdr;
1538 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1539 ASSERT(hdr->b_birth != 0);
1540 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1542 if (lockp != NULL) {
1544 mutex_enter(hash_lock);
1546 ASSERT(MUTEX_HELD(hash_lock));
1549 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1550 fhdr = fhdr->b_hash_next, i++) {
1551 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1555 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1556 buf_hash_table.ht_table[idx] = hdr;
1557 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1559 /* collect some hash table performance data */
1561 ARCSTAT_BUMP(arcstat_hash_collisions);
1563 ARCSTAT_BUMP(arcstat_hash_chains);
1565 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1568 ARCSTAT_BUMP(arcstat_hash_elements);
1569 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1575 buf_hash_remove(arc_buf_hdr_t *hdr)
1577 arc_buf_hdr_t *fhdr, **hdrp;
1578 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1580 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1581 ASSERT(HDR_IN_HASH_TABLE(hdr));
1583 hdrp = &buf_hash_table.ht_table[idx];
1584 while ((fhdr = *hdrp) != hdr) {
1585 ASSERT3P(fhdr, !=, NULL);
1586 hdrp = &fhdr->b_hash_next;
1588 *hdrp = hdr->b_hash_next;
1589 hdr->b_hash_next = NULL;
1590 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1592 /* collect some hash table performance data */
1593 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1595 if (buf_hash_table.ht_table[idx] &&
1596 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1597 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1601 * Global data structures and functions for the buf kmem cache.
1603 static kmem_cache_t *hdr_full_cache;
1604 static kmem_cache_t *hdr_l2only_cache;
1605 static kmem_cache_t *buf_cache;
1612 kmem_free(buf_hash_table.ht_table,
1613 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1614 for (i = 0; i < BUF_LOCKS; i++)
1615 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1616 kmem_cache_destroy(hdr_full_cache);
1617 kmem_cache_destroy(hdr_l2only_cache);
1618 kmem_cache_destroy(buf_cache);
1622 * Constructor callback - called when the cache is empty
1623 * and a new buf is requested.
1627 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1629 arc_buf_hdr_t *hdr = vbuf;
1631 bzero(hdr, HDR_FULL_SIZE);
1632 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1633 refcount_create(&hdr->b_l1hdr.b_refcnt);
1634 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1635 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1636 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1643 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1645 arc_buf_hdr_t *hdr = vbuf;
1647 bzero(hdr, HDR_L2ONLY_SIZE);
1648 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1655 buf_cons(void *vbuf, void *unused, int kmflag)
1657 arc_buf_t *buf = vbuf;
1659 bzero(buf, sizeof (arc_buf_t));
1660 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1661 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1667 * Destructor callback - called when a cached buf is
1668 * no longer required.
1672 hdr_full_dest(void *vbuf, void *unused)
1674 arc_buf_hdr_t *hdr = vbuf;
1676 ASSERT(HDR_EMPTY(hdr));
1677 cv_destroy(&hdr->b_l1hdr.b_cv);
1678 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1679 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1680 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1681 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1686 hdr_l2only_dest(void *vbuf, void *unused)
1688 arc_buf_hdr_t *hdr = vbuf;
1690 ASSERT(HDR_EMPTY(hdr));
1691 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1696 buf_dest(void *vbuf, void *unused)
1698 arc_buf_t *buf = vbuf;
1700 mutex_destroy(&buf->b_evict_lock);
1701 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1705 * Reclaim callback -- invoked when memory is low.
1709 hdr_recl(void *unused)
1711 dprintf("hdr_recl called\n");
1713 * umem calls the reclaim func when we destroy the buf cache,
1714 * which is after we do arc_fini().
1717 cv_signal(&arc_reclaim_thread_cv);
1724 uint64_t hsize = 1ULL << 12;
1728 * The hash table is big enough to fill all of physical memory
1729 * with an average block size of zfs_arc_average_blocksize (default 8K).
1730 * By default, the table will take up
1731 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1733 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1736 buf_hash_table.ht_mask = hsize - 1;
1737 buf_hash_table.ht_table =
1738 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1739 if (buf_hash_table.ht_table == NULL) {
1740 ASSERT(hsize > (1ULL << 8));
1745 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1746 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1747 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1748 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1750 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1751 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1753 for (i = 0; i < 256; i++)
1754 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1755 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1757 for (i = 0; i < BUF_LOCKS; i++) {
1758 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1759 NULL, MUTEX_DEFAULT, NULL);
1764 * This is the size that the buf occupies in memory. If the buf is compressed,
1765 * it will correspond to the compressed size. You should use this method of
1766 * getting the buf size unless you explicitly need the logical size.
1769 arc_buf_size(arc_buf_t *buf)
1771 return (ARC_BUF_COMPRESSED(buf) ?
1772 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1776 arc_buf_lsize(arc_buf_t *buf)
1778 return (HDR_GET_LSIZE(buf->b_hdr));
1782 arc_get_compression(arc_buf_t *buf)
1784 return (ARC_BUF_COMPRESSED(buf) ?
1785 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1788 #define ARC_MINTIME (hz>>4) /* 62 ms */
1790 static inline boolean_t
1791 arc_buf_is_shared(arc_buf_t *buf)
1793 boolean_t shared = (buf->b_data != NULL &&
1794 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1795 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1796 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1797 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1798 IMPLY(shared, ARC_BUF_SHARED(buf));
1799 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1802 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1803 * already being shared" requirement prevents us from doing that.
1810 * Free the checksum associated with this header. If there is no checksum, this
1814 arc_cksum_free(arc_buf_hdr_t *hdr)
1816 ASSERT(HDR_HAS_L1HDR(hdr));
1817 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1818 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1819 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1820 hdr->b_l1hdr.b_freeze_cksum = NULL;
1822 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1826 * Return true iff at least one of the bufs on hdr is not compressed.
1829 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1831 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1832 if (!ARC_BUF_COMPRESSED(b)) {
1840 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1841 * matches the checksum that is stored in the hdr. If there is no checksum,
1842 * or if the buf is compressed, this is a no-op.
1845 arc_cksum_verify(arc_buf_t *buf)
1847 arc_buf_hdr_t *hdr = buf->b_hdr;
1850 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1853 if (ARC_BUF_COMPRESSED(buf)) {
1854 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1855 arc_hdr_has_uncompressed_buf(hdr));
1859 ASSERT(HDR_HAS_L1HDR(hdr));
1861 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1862 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1863 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1867 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1868 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1869 panic("buffer modified while frozen!");
1870 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1874 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1876 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1877 boolean_t valid_cksum;
1879 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1880 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1883 * We rely on the blkptr's checksum to determine if the block
1884 * is valid or not. When compressed arc is enabled, the l2arc
1885 * writes the block to the l2arc just as it appears in the pool.
1886 * This allows us to use the blkptr's checksum to validate the
1887 * data that we just read off of the l2arc without having to store
1888 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1889 * arc is disabled, then the data written to the l2arc is always
1890 * uncompressed and won't match the block as it exists in the main
1891 * pool. When this is the case, we must first compress it if it is
1892 * compressed on the main pool before we can validate the checksum.
1894 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1895 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1896 uint64_t lsize = HDR_GET_LSIZE(hdr);
1899 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1900 csize = zio_compress_data(compress, zio->io_abd,
1901 abd_to_buf(cdata), lsize);
1903 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1904 if (csize < HDR_GET_PSIZE(hdr)) {
1906 * Compressed blocks are always a multiple of the
1907 * smallest ashift in the pool. Ideally, we would
1908 * like to round up the csize to the next
1909 * spa_min_ashift but that value may have changed
1910 * since the block was last written. Instead,
1911 * we rely on the fact that the hdr's psize
1912 * was set to the psize of the block when it was
1913 * last written. We set the csize to that value
1914 * and zero out any part that should not contain
1917 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1918 csize = HDR_GET_PSIZE(hdr);
1920 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1924 * Block pointers always store the checksum for the logical data.
1925 * If the block pointer has the gang bit set, then the checksum
1926 * it represents is for the reconstituted data and not for an
1927 * individual gang member. The zio pipeline, however, must be able to
1928 * determine the checksum of each of the gang constituents so it
1929 * treats the checksum comparison differently than what we need
1930 * for l2arc blocks. This prevents us from using the
1931 * zio_checksum_error() interface directly. Instead we must call the
1932 * zio_checksum_error_impl() so that we can ensure the checksum is
1933 * generated using the correct checksum algorithm and accounts for the
1934 * logical I/O size and not just a gang fragment.
1936 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1937 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1938 zio->io_offset, NULL) == 0);
1939 zio_pop_transforms(zio);
1940 return (valid_cksum);
1944 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1945 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1946 * isn't modified later on. If buf is compressed or there is already a checksum
1947 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1950 arc_cksum_compute(arc_buf_t *buf)
1952 arc_buf_hdr_t *hdr = buf->b_hdr;
1954 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1957 ASSERT(HDR_HAS_L1HDR(hdr));
1959 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1960 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1961 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1962 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1964 } else if (ARC_BUF_COMPRESSED(buf)) {
1965 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1969 ASSERT(!ARC_BUF_COMPRESSED(buf));
1970 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1972 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1973 hdr->b_l1hdr.b_freeze_cksum);
1974 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1982 typedef struct procctl {
1990 arc_buf_unwatch(arc_buf_t *buf)
1997 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1998 ctl.prwatch.pr_size = 0;
1999 ctl.prwatch.pr_wflags = 0;
2000 result = write(arc_procfd, &ctl, sizeof (ctl));
2001 ASSERT3U(result, ==, sizeof (ctl));
2008 arc_buf_watch(arc_buf_t *buf)
2015 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2016 ctl.prwatch.pr_size = arc_buf_size(buf);
2017 ctl.prwatch.pr_wflags = WA_WRITE;
2018 result = write(arc_procfd, &ctl, sizeof (ctl));
2019 ASSERT3U(result, ==, sizeof (ctl));
2023 #endif /* illumos */
2025 static arc_buf_contents_t
2026 arc_buf_type(arc_buf_hdr_t *hdr)
2028 arc_buf_contents_t type;
2029 if (HDR_ISTYPE_METADATA(hdr)) {
2030 type = ARC_BUFC_METADATA;
2032 type = ARC_BUFC_DATA;
2034 VERIFY3U(hdr->b_type, ==, type);
2039 arc_is_metadata(arc_buf_t *buf)
2041 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2045 arc_bufc_to_flags(arc_buf_contents_t type)
2049 /* metadata field is 0 if buffer contains normal data */
2051 case ARC_BUFC_METADATA:
2052 return (ARC_FLAG_BUFC_METADATA);
2056 panic("undefined ARC buffer type!");
2057 return ((uint32_t)-1);
2061 arc_buf_thaw(arc_buf_t *buf)
2063 arc_buf_hdr_t *hdr = buf->b_hdr;
2065 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2066 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2068 arc_cksum_verify(buf);
2071 * Compressed buffers do not manipulate the b_freeze_cksum or
2072 * allocate b_thawed.
2074 if (ARC_BUF_COMPRESSED(buf)) {
2075 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2076 arc_hdr_has_uncompressed_buf(hdr));
2080 ASSERT(HDR_HAS_L1HDR(hdr));
2081 arc_cksum_free(hdr);
2083 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2085 if (zfs_flags & ZFS_DEBUG_MODIFY) {
2086 if (hdr->b_l1hdr.b_thawed != NULL)
2087 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2088 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2092 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2095 arc_buf_unwatch(buf);
2100 arc_buf_freeze(arc_buf_t *buf)
2102 arc_buf_hdr_t *hdr = buf->b_hdr;
2103 kmutex_t *hash_lock;
2105 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2108 if (ARC_BUF_COMPRESSED(buf)) {
2109 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2110 arc_hdr_has_uncompressed_buf(hdr));
2114 hash_lock = HDR_LOCK(hdr);
2115 mutex_enter(hash_lock);
2117 ASSERT(HDR_HAS_L1HDR(hdr));
2118 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
2119 hdr->b_l1hdr.b_state == arc_anon);
2120 arc_cksum_compute(buf);
2121 mutex_exit(hash_lock);
2125 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2126 * the following functions should be used to ensure that the flags are
2127 * updated in a thread-safe way. When manipulating the flags either
2128 * the hash_lock must be held or the hdr must be undiscoverable. This
2129 * ensures that we're not racing with any other threads when updating
2133 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2135 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2136 hdr->b_flags |= flags;
2140 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2142 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2143 hdr->b_flags &= ~flags;
2147 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2148 * done in a special way since we have to clear and set bits
2149 * at the same time. Consumers that wish to set the compression bits
2150 * must use this function to ensure that the flags are updated in
2151 * thread-safe manner.
2154 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2156 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2159 * Holes and embedded blocks will always have a psize = 0 so
2160 * we ignore the compression of the blkptr and set the
2161 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2162 * Holes and embedded blocks remain anonymous so we don't
2163 * want to uncompress them. Mark them as uncompressed.
2165 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2166 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2167 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2168 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2169 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2171 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2172 HDR_SET_COMPRESS(hdr, cmp);
2173 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2174 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2179 * Looks for another buf on the same hdr which has the data decompressed, copies
2180 * from it, and returns true. If no such buf exists, returns false.
2183 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
2185 arc_buf_hdr_t *hdr = buf->b_hdr;
2186 boolean_t copied = B_FALSE;
2188 ASSERT(HDR_HAS_L1HDR(hdr));
2189 ASSERT3P(buf->b_data, !=, NULL);
2190 ASSERT(!ARC_BUF_COMPRESSED(buf));
2192 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2193 from = from->b_next) {
2194 /* can't use our own data buffer */
2199 if (!ARC_BUF_COMPRESSED(from)) {
2200 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2207 * There were no decompressed bufs, so there should not be a
2208 * checksum on the hdr either.
2210 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
2216 * Given a buf that has a data buffer attached to it, this function will
2217 * efficiently fill the buf with data of the specified compression setting from
2218 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2219 * are already sharing a data buf, no copy is performed.
2221 * If the buf is marked as compressed but uncompressed data was requested, this
2222 * will allocate a new data buffer for the buf, remove that flag, and fill the
2223 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2224 * uncompressed data, and (since we haven't added support for it yet) if you
2225 * want compressed data your buf must already be marked as compressed and have
2226 * the correct-sized data buffer.
2229 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2231 arc_buf_hdr_t *hdr = buf->b_hdr;
2232 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2233 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2235 ASSERT3P(buf->b_data, !=, NULL);
2236 IMPLY(compressed, hdr_compressed);
2237 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2239 if (hdr_compressed == compressed) {
2240 if (!arc_buf_is_shared(buf)) {
2241 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2245 ASSERT(hdr_compressed);
2246 ASSERT(!compressed);
2247 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2250 * If the buf is sharing its data with the hdr, unlink it and
2251 * allocate a new data buffer for the buf.
2253 if (arc_buf_is_shared(buf)) {
2254 ASSERT(ARC_BUF_COMPRESSED(buf));
2256 /* We need to give the buf it's own b_data */
2257 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2259 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2260 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2262 /* Previously overhead was 0; just add new overhead */
2263 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2264 } else if (ARC_BUF_COMPRESSED(buf)) {
2265 /* We need to reallocate the buf's b_data */
2266 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2269 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2271 /* We increased the size of b_data; update overhead */
2272 ARCSTAT_INCR(arcstat_overhead_size,
2273 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2277 * Regardless of the buf's previous compression settings, it
2278 * should not be compressed at the end of this function.
2280 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2283 * Try copying the data from another buf which already has a
2284 * decompressed version. If that's not possible, it's time to
2285 * bite the bullet and decompress the data from the hdr.
2287 if (arc_buf_try_copy_decompressed_data(buf)) {
2288 /* Skip byteswapping and checksumming (already done) */
2289 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2292 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2293 hdr->b_l1hdr.b_pabd, buf->b_data,
2294 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2297 * Absent hardware errors or software bugs, this should
2298 * be impossible, but log it anyway so we can debug it.
2302 "hdr %p, compress %d, psize %d, lsize %d",
2303 hdr, HDR_GET_COMPRESS(hdr),
2304 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2305 return (SET_ERROR(EIO));
2310 /* Byteswap the buf's data if necessary */
2311 if (bswap != DMU_BSWAP_NUMFUNCS) {
2312 ASSERT(!HDR_SHARED_DATA(hdr));
2313 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2314 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2317 /* Compute the hdr's checksum if necessary */
2318 arc_cksum_compute(buf);
2324 arc_decompress(arc_buf_t *buf)
2326 return (arc_buf_fill(buf, B_FALSE));
2330 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2333 arc_hdr_size(arc_buf_hdr_t *hdr)
2337 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2338 HDR_GET_PSIZE(hdr) > 0) {
2339 size = HDR_GET_PSIZE(hdr);
2341 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2342 size = HDR_GET_LSIZE(hdr);
2348 * Increment the amount of evictable space in the arc_state_t's refcount.
2349 * We account for the space used by the hdr and the arc buf individually
2350 * so that we can add and remove them from the refcount individually.
2353 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2355 arc_buf_contents_t type = arc_buf_type(hdr);
2357 ASSERT(HDR_HAS_L1HDR(hdr));
2359 if (GHOST_STATE(state)) {
2360 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2361 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2362 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2363 (void) refcount_add_many(&state->arcs_esize[type],
2364 HDR_GET_LSIZE(hdr), hdr);
2368 ASSERT(!GHOST_STATE(state));
2369 if (hdr->b_l1hdr.b_pabd != NULL) {
2370 (void) refcount_add_many(&state->arcs_esize[type],
2371 arc_hdr_size(hdr), hdr);
2373 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2374 buf = buf->b_next) {
2375 if (arc_buf_is_shared(buf))
2377 (void) refcount_add_many(&state->arcs_esize[type],
2378 arc_buf_size(buf), buf);
2383 * Decrement the amount of evictable space in the arc_state_t's refcount.
2384 * We account for the space used by the hdr and the arc buf individually
2385 * so that we can add and remove them from the refcount individually.
2388 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2390 arc_buf_contents_t type = arc_buf_type(hdr);
2392 ASSERT(HDR_HAS_L1HDR(hdr));
2394 if (GHOST_STATE(state)) {
2395 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2396 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2397 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2398 (void) refcount_remove_many(&state->arcs_esize[type],
2399 HDR_GET_LSIZE(hdr), hdr);
2403 ASSERT(!GHOST_STATE(state));
2404 if (hdr->b_l1hdr.b_pabd != NULL) {
2405 (void) refcount_remove_many(&state->arcs_esize[type],
2406 arc_hdr_size(hdr), hdr);
2408 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2409 buf = buf->b_next) {
2410 if (arc_buf_is_shared(buf))
2412 (void) refcount_remove_many(&state->arcs_esize[type],
2413 arc_buf_size(buf), buf);
2418 * Add a reference to this hdr indicating that someone is actively
2419 * referencing that memory. When the refcount transitions from 0 to 1,
2420 * we remove it from the respective arc_state_t list to indicate that
2421 * it is not evictable.
2424 add_reference(arc_buf_hdr_t *hdr, void *tag)
2426 ASSERT(HDR_HAS_L1HDR(hdr));
2427 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2428 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2429 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2430 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2433 arc_state_t *state = hdr->b_l1hdr.b_state;
2435 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2436 (state != arc_anon)) {
2437 /* We don't use the L2-only state list. */
2438 if (state != arc_l2c_only) {
2439 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2441 arc_evictable_space_decrement(hdr, state);
2443 /* remove the prefetch flag if we get a reference */
2444 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2449 * Remove a reference from this hdr. When the reference transitions from
2450 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2451 * list making it eligible for eviction.
2454 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2457 arc_state_t *state = hdr->b_l1hdr.b_state;
2459 ASSERT(HDR_HAS_L1HDR(hdr));
2460 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2461 ASSERT(!GHOST_STATE(state));
2464 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2465 * check to prevent usage of the arc_l2c_only list.
2467 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2468 (state != arc_anon)) {
2469 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2470 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2471 arc_evictable_space_increment(hdr, state);
2477 * Move the supplied buffer to the indicated state. The hash lock
2478 * for the buffer must be held by the caller.
2481 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2482 kmutex_t *hash_lock)
2484 arc_state_t *old_state;
2487 boolean_t update_old, update_new;
2488 arc_buf_contents_t buftype = arc_buf_type(hdr);
2491 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2492 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2493 * L1 hdr doesn't always exist when we change state to arc_anon before
2494 * destroying a header, in which case reallocating to add the L1 hdr is
2497 if (HDR_HAS_L1HDR(hdr)) {
2498 old_state = hdr->b_l1hdr.b_state;
2499 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2500 bufcnt = hdr->b_l1hdr.b_bufcnt;
2501 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2503 old_state = arc_l2c_only;
2506 update_old = B_FALSE;
2508 update_new = update_old;
2510 ASSERT(MUTEX_HELD(hash_lock));
2511 ASSERT3P(new_state, !=, old_state);
2512 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2513 ASSERT(old_state != arc_anon || bufcnt <= 1);
2516 * If this buffer is evictable, transfer it from the
2517 * old state list to the new state list.
2520 if (old_state != arc_anon && old_state != arc_l2c_only) {
2521 ASSERT(HDR_HAS_L1HDR(hdr));
2522 multilist_remove(old_state->arcs_list[buftype], hdr);
2524 if (GHOST_STATE(old_state)) {
2526 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2527 update_old = B_TRUE;
2529 arc_evictable_space_decrement(hdr, old_state);
2531 if (new_state != arc_anon && new_state != arc_l2c_only) {
2534 * An L1 header always exists here, since if we're
2535 * moving to some L1-cached state (i.e. not l2c_only or
2536 * anonymous), we realloc the header to add an L1hdr
2539 ASSERT(HDR_HAS_L1HDR(hdr));
2540 multilist_insert(new_state->arcs_list[buftype], hdr);
2542 if (GHOST_STATE(new_state)) {
2544 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2545 update_new = B_TRUE;
2547 arc_evictable_space_increment(hdr, new_state);
2551 ASSERT(!HDR_EMPTY(hdr));
2552 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2553 buf_hash_remove(hdr);
2555 /* adjust state sizes (ignore arc_l2c_only) */
2557 if (update_new && new_state != arc_l2c_only) {
2558 ASSERT(HDR_HAS_L1HDR(hdr));
2559 if (GHOST_STATE(new_state)) {
2563 * When moving a header to a ghost state, we first
2564 * remove all arc buffers. Thus, we'll have a
2565 * bufcnt of zero, and no arc buffer to use for
2566 * the reference. As a result, we use the arc
2567 * header pointer for the reference.
2569 (void) refcount_add_many(&new_state->arcs_size,
2570 HDR_GET_LSIZE(hdr), hdr);
2571 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2573 uint32_t buffers = 0;
2576 * Each individual buffer holds a unique reference,
2577 * thus we must remove each of these references one
2580 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2581 buf = buf->b_next) {
2582 ASSERT3U(bufcnt, !=, 0);
2586 * When the arc_buf_t is sharing the data
2587 * block with the hdr, the owner of the
2588 * reference belongs to the hdr. Only
2589 * add to the refcount if the arc_buf_t is
2592 if (arc_buf_is_shared(buf))
2595 (void) refcount_add_many(&new_state->arcs_size,
2596 arc_buf_size(buf), buf);
2598 ASSERT3U(bufcnt, ==, buffers);
2600 if (hdr->b_l1hdr.b_pabd != NULL) {
2601 (void) refcount_add_many(&new_state->arcs_size,
2602 arc_hdr_size(hdr), hdr);
2604 ASSERT(GHOST_STATE(old_state));
2609 if (update_old && old_state != arc_l2c_only) {
2610 ASSERT(HDR_HAS_L1HDR(hdr));
2611 if (GHOST_STATE(old_state)) {
2613 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2616 * When moving a header off of a ghost state,
2617 * the header will not contain any arc buffers.
2618 * We use the arc header pointer for the reference
2619 * which is exactly what we did when we put the
2620 * header on the ghost state.
2623 (void) refcount_remove_many(&old_state->arcs_size,
2624 HDR_GET_LSIZE(hdr), hdr);
2626 uint32_t buffers = 0;
2629 * Each individual buffer holds a unique reference,
2630 * thus we must remove each of these references one
2633 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2634 buf = buf->b_next) {
2635 ASSERT3U(bufcnt, !=, 0);
2639 * When the arc_buf_t is sharing the data
2640 * block with the hdr, the owner of the
2641 * reference belongs to the hdr. Only
2642 * add to the refcount if the arc_buf_t is
2645 if (arc_buf_is_shared(buf))
2648 (void) refcount_remove_many(
2649 &old_state->arcs_size, arc_buf_size(buf),
2652 ASSERT3U(bufcnt, ==, buffers);
2653 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2654 (void) refcount_remove_many(
2655 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2659 if (HDR_HAS_L1HDR(hdr))
2660 hdr->b_l1hdr.b_state = new_state;
2663 * L2 headers should never be on the L2 state list since they don't
2664 * have L1 headers allocated.
2666 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2667 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2671 arc_space_consume(uint64_t space, arc_space_type_t type)
2673 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2676 case ARC_SPACE_DATA:
2677 aggsum_add(&astat_data_size, space);
2679 case ARC_SPACE_META:
2680 aggsum_add(&astat_metadata_size, space);
2682 case ARC_SPACE_OTHER:
2683 aggsum_add(&astat_other_size, space);
2685 case ARC_SPACE_HDRS:
2686 aggsum_add(&astat_hdr_size, space);
2688 case ARC_SPACE_L2HDRS:
2689 aggsum_add(&astat_l2_hdr_size, space);
2693 if (type != ARC_SPACE_DATA)
2694 aggsum_add(&arc_meta_used, space);
2696 aggsum_add(&arc_size, space);
2700 arc_space_return(uint64_t space, arc_space_type_t type)
2702 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2705 case ARC_SPACE_DATA:
2706 aggsum_add(&astat_data_size, -space);
2708 case ARC_SPACE_META:
2709 aggsum_add(&astat_metadata_size, -space);
2711 case ARC_SPACE_OTHER:
2712 aggsum_add(&astat_other_size, -space);
2714 case ARC_SPACE_HDRS:
2715 aggsum_add(&astat_hdr_size, -space);
2717 case ARC_SPACE_L2HDRS:
2718 aggsum_add(&astat_l2_hdr_size, -space);
2722 if (type != ARC_SPACE_DATA) {
2723 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2725 * We use the upper bound here rather than the precise value
2726 * because the arc_meta_max value doesn't need to be
2727 * precise. It's only consumed by humans via arcstats.
2729 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2730 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2731 aggsum_add(&arc_meta_used, -space);
2734 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2735 aggsum_add(&arc_size, -space);
2739 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2740 * with the hdr's b_pabd.
2743 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2746 * The criteria for sharing a hdr's data are:
2747 * 1. the hdr's compression matches the buf's compression
2748 * 2. the hdr doesn't need to be byteswapped
2749 * 3. the hdr isn't already being shared
2750 * 4. the buf is either compressed or it is the last buf in the hdr list
2752 * Criterion #4 maintains the invariant that shared uncompressed
2753 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2754 * might ask, "if a compressed buf is allocated first, won't that be the
2755 * last thing in the list?", but in that case it's impossible to create
2756 * a shared uncompressed buf anyway (because the hdr must be compressed
2757 * to have the compressed buf). You might also think that #3 is
2758 * sufficient to make this guarantee, however it's possible
2759 * (specifically in the rare L2ARC write race mentioned in
2760 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2761 * is sharable, but wasn't at the time of its allocation. Rather than
2762 * allow a new shared uncompressed buf to be created and then shuffle
2763 * the list around to make it the last element, this simply disallows
2764 * sharing if the new buf isn't the first to be added.
2766 ASSERT3P(buf->b_hdr, ==, hdr);
2767 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2768 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2769 return (buf_compressed == hdr_compressed &&
2770 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2771 !HDR_SHARED_DATA(hdr) &&
2772 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2776 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2777 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2778 * copy was made successfully, or an error code otherwise.
2781 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2782 boolean_t fill, arc_buf_t **ret)
2786 ASSERT(HDR_HAS_L1HDR(hdr));
2787 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2788 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2789 hdr->b_type == ARC_BUFC_METADATA);
2790 ASSERT3P(ret, !=, NULL);
2791 ASSERT3P(*ret, ==, NULL);
2793 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2796 buf->b_next = hdr->b_l1hdr.b_buf;
2799 add_reference(hdr, tag);
2802 * We're about to change the hdr's b_flags. We must either
2803 * hold the hash_lock or be undiscoverable.
2805 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2808 * Only honor requests for compressed bufs if the hdr is actually
2811 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2812 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2815 * If the hdr's data can be shared then we share the data buffer and
2816 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2817 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2818 * buffer to store the buf's data.
2820 * There are two additional restrictions here because we're sharing
2821 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2822 * actively involved in an L2ARC write, because if this buf is used by
2823 * an arc_write() then the hdr's data buffer will be released when the
2824 * write completes, even though the L2ARC write might still be using it.
2825 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2826 * need to be ABD-aware.
2828 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2829 abd_is_linear(hdr->b_l1hdr.b_pabd);
2831 /* Set up b_data and sharing */
2833 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2834 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2835 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2838 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2839 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2841 VERIFY3P(buf->b_data, !=, NULL);
2843 hdr->b_l1hdr.b_buf = buf;
2844 hdr->b_l1hdr.b_bufcnt += 1;
2847 * If the user wants the data from the hdr, we need to either copy or
2848 * decompress the data.
2851 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2857 static char *arc_onloan_tag = "onloan";
2860 arc_loaned_bytes_update(int64_t delta)
2862 atomic_add_64(&arc_loaned_bytes, delta);
2864 /* assert that it did not wrap around */
2865 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2869 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2870 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2871 * buffers must be returned to the arc before they can be used by the DMU or
2875 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2877 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2878 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2880 arc_loaned_bytes_update(size);
2886 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2887 enum zio_compress compression_type)
2889 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2890 psize, lsize, compression_type);
2892 arc_loaned_bytes_update(psize);
2899 * Return a loaned arc buffer to the arc.
2902 arc_return_buf(arc_buf_t *buf, void *tag)
2904 arc_buf_hdr_t *hdr = buf->b_hdr;
2906 ASSERT3P(buf->b_data, !=, NULL);
2907 ASSERT(HDR_HAS_L1HDR(hdr));
2908 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2909 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2911 arc_loaned_bytes_update(-arc_buf_size(buf));
2914 /* Detach an arc_buf from a dbuf (tag) */
2916 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2918 arc_buf_hdr_t *hdr = buf->b_hdr;
2920 ASSERT3P(buf->b_data, !=, NULL);
2921 ASSERT(HDR_HAS_L1HDR(hdr));
2922 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2923 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2925 arc_loaned_bytes_update(arc_buf_size(buf));
2929 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2931 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2934 df->l2df_size = size;
2935 df->l2df_type = type;
2936 mutex_enter(&l2arc_free_on_write_mtx);
2937 list_insert_head(l2arc_free_on_write, df);
2938 mutex_exit(&l2arc_free_on_write_mtx);
2942 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2944 arc_state_t *state = hdr->b_l1hdr.b_state;
2945 arc_buf_contents_t type = arc_buf_type(hdr);
2946 uint64_t size = arc_hdr_size(hdr);
2948 /* protected by hash lock, if in the hash table */
2949 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2950 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2951 ASSERT(state != arc_anon && state != arc_l2c_only);
2953 (void) refcount_remove_many(&state->arcs_esize[type],
2956 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2957 if (type == ARC_BUFC_METADATA) {
2958 arc_space_return(size, ARC_SPACE_META);
2960 ASSERT(type == ARC_BUFC_DATA);
2961 arc_space_return(size, ARC_SPACE_DATA);
2964 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2968 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2969 * data buffer, we transfer the refcount ownership to the hdr and update
2970 * the appropriate kstats.
2973 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2975 arc_state_t *state = hdr->b_l1hdr.b_state;
2977 ASSERT(arc_can_share(hdr, buf));
2978 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2979 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2982 * Start sharing the data buffer. We transfer the
2983 * refcount ownership to the hdr since it always owns
2984 * the refcount whenever an arc_buf_t is shared.
2986 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2987 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2988 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2989 HDR_ISTYPE_METADATA(hdr));
2990 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2991 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2994 * Since we've transferred ownership to the hdr we need
2995 * to increment its compressed and uncompressed kstats and
2996 * decrement the overhead size.
2998 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2999 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3000 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3004 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3006 arc_state_t *state = hdr->b_l1hdr.b_state;
3008 ASSERT(arc_buf_is_shared(buf));
3009 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3010 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3013 * We are no longer sharing this buffer so we need
3014 * to transfer its ownership to the rightful owner.
3016 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
3017 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3018 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3019 abd_put(hdr->b_l1hdr.b_pabd);
3020 hdr->b_l1hdr.b_pabd = NULL;
3021 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3024 * Since the buffer is no longer shared between
3025 * the arc buf and the hdr, count it as overhead.
3027 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3028 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3029 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3033 * Remove an arc_buf_t from the hdr's buf list and return the last
3034 * arc_buf_t on the list. If no buffers remain on the list then return
3038 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3040 ASSERT(HDR_HAS_L1HDR(hdr));
3041 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3043 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3044 arc_buf_t *lastbuf = NULL;
3047 * Remove the buf from the hdr list and locate the last
3048 * remaining buffer on the list.
3050 while (*bufp != NULL) {
3052 *bufp = buf->b_next;
3055 * If we've removed a buffer in the middle of
3056 * the list then update the lastbuf and update
3059 if (*bufp != NULL) {
3061 bufp = &(*bufp)->b_next;
3065 ASSERT3P(lastbuf, !=, buf);
3066 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3067 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3068 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3074 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3078 arc_buf_destroy_impl(arc_buf_t *buf)
3080 arc_buf_hdr_t *hdr = buf->b_hdr;
3083 * Free up the data associated with the buf but only if we're not
3084 * sharing this with the hdr. If we are sharing it with the hdr, the
3085 * hdr is responsible for doing the free.
3087 if (buf->b_data != NULL) {
3089 * We're about to change the hdr's b_flags. We must either
3090 * hold the hash_lock or be undiscoverable.
3092 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3094 arc_cksum_verify(buf);
3096 arc_buf_unwatch(buf);
3099 if (arc_buf_is_shared(buf)) {
3100 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3102 uint64_t size = arc_buf_size(buf);
3103 arc_free_data_buf(hdr, buf->b_data, size, buf);
3104 ARCSTAT_INCR(arcstat_overhead_size, -size);
3108 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3109 hdr->b_l1hdr.b_bufcnt -= 1;
3112 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3114 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3116 * If the current arc_buf_t is sharing its data buffer with the
3117 * hdr, then reassign the hdr's b_pabd to share it with the new
3118 * buffer at the end of the list. The shared buffer is always
3119 * the last one on the hdr's buffer list.
3121 * There is an equivalent case for compressed bufs, but since
3122 * they aren't guaranteed to be the last buf in the list and
3123 * that is an exceedingly rare case, we just allow that space be
3124 * wasted temporarily.
3126 if (lastbuf != NULL) {
3127 /* Only one buf can be shared at once */
3128 VERIFY(!arc_buf_is_shared(lastbuf));
3129 /* hdr is uncompressed so can't have compressed buf */
3130 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3132 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3133 arc_hdr_free_pabd(hdr);
3136 * We must setup a new shared block between the
3137 * last buffer and the hdr. The data would have
3138 * been allocated by the arc buf so we need to transfer
3139 * ownership to the hdr since it's now being shared.
3141 arc_share_buf(hdr, lastbuf);
3143 } else if (HDR_SHARED_DATA(hdr)) {
3145 * Uncompressed shared buffers are always at the end
3146 * of the list. Compressed buffers don't have the
3147 * same requirements. This makes it hard to
3148 * simply assert that the lastbuf is shared so
3149 * we rely on the hdr's compression flags to determine
3150 * if we have a compressed, shared buffer.
3152 ASSERT3P(lastbuf, !=, NULL);
3153 ASSERT(arc_buf_is_shared(lastbuf) ||
3154 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
3158 * Free the checksum if we're removing the last uncompressed buf from
3161 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3162 arc_cksum_free(hdr);
3165 /* clean up the buf */
3167 kmem_cache_free(buf_cache, buf);
3171 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
3173 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3174 ASSERT(HDR_HAS_L1HDR(hdr));
3175 ASSERT(!HDR_SHARED_DATA(hdr));
3177 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3178 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
3179 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3180 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3182 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3183 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3187 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3189 ASSERT(HDR_HAS_L1HDR(hdr));
3190 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3193 * If the hdr is currently being written to the l2arc then
3194 * we defer freeing the data by adding it to the l2arc_free_on_write
3195 * list. The l2arc will free the data once it's finished
3196 * writing it to the l2arc device.
3198 if (HDR_L2_WRITING(hdr)) {
3199 arc_hdr_free_on_write(hdr);
3200 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3202 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3203 arc_hdr_size(hdr), hdr);
3205 hdr->b_l1hdr.b_pabd = NULL;
3206 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3208 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3209 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3212 static arc_buf_hdr_t *
3213 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3214 enum zio_compress compression_type, arc_buf_contents_t type)
3218 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3220 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3221 ASSERT(HDR_EMPTY(hdr));
3222 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3223 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3224 HDR_SET_PSIZE(hdr, psize);
3225 HDR_SET_LSIZE(hdr, lsize);
3229 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3230 arc_hdr_set_compress(hdr, compression_type);
3232 hdr->b_l1hdr.b_state = arc_anon;
3233 hdr->b_l1hdr.b_arc_access = 0;
3234 hdr->b_l1hdr.b_bufcnt = 0;
3235 hdr->b_l1hdr.b_buf = NULL;
3238 * Allocate the hdr's buffer. This will contain either
3239 * the compressed or uncompressed data depending on the block
3240 * it references and compressed arc enablement.
3242 arc_hdr_alloc_pabd(hdr);
3243 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3249 * Transition between the two allocation states for the arc_buf_hdr struct.
3250 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3251 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3252 * version is used when a cache buffer is only in the L2ARC in order to reduce
3255 static arc_buf_hdr_t *
3256 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3258 ASSERT(HDR_HAS_L2HDR(hdr));
3260 arc_buf_hdr_t *nhdr;
3261 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3263 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3264 (old == hdr_l2only_cache && new == hdr_full_cache));
3266 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3268 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3269 buf_hash_remove(hdr);
3271 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3273 if (new == hdr_full_cache) {
3274 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3276 * arc_access and arc_change_state need to be aware that a
3277 * header has just come out of L2ARC, so we set its state to
3278 * l2c_only even though it's about to change.
3280 nhdr->b_l1hdr.b_state = arc_l2c_only;
3282 /* Verify previous threads set to NULL before freeing */
3283 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3285 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3286 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3287 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3290 * If we've reached here, We must have been called from
3291 * arc_evict_hdr(), as such we should have already been
3292 * removed from any ghost list we were previously on
3293 * (which protects us from racing with arc_evict_state),
3294 * thus no locking is needed during this check.
3296 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3299 * A buffer must not be moved into the arc_l2c_only
3300 * state if it's not finished being written out to the
3301 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3302 * might try to be accessed, even though it was removed.
3304 VERIFY(!HDR_L2_WRITING(hdr));
3305 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3308 if (hdr->b_l1hdr.b_thawed != NULL) {
3309 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3310 hdr->b_l1hdr.b_thawed = NULL;
3314 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3317 * The header has been reallocated so we need to re-insert it into any
3320 (void) buf_hash_insert(nhdr, NULL);
3322 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3324 mutex_enter(&dev->l2ad_mtx);
3327 * We must place the realloc'ed header back into the list at
3328 * the same spot. Otherwise, if it's placed earlier in the list,
3329 * l2arc_write_buffers() could find it during the function's
3330 * write phase, and try to write it out to the l2arc.
3332 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3333 list_remove(&dev->l2ad_buflist, hdr);
3335 mutex_exit(&dev->l2ad_mtx);
3338 * Since we're using the pointer address as the tag when
3339 * incrementing and decrementing the l2ad_alloc refcount, we
3340 * must remove the old pointer (that we're about to destroy) and
3341 * add the new pointer to the refcount. Otherwise we'd remove
3342 * the wrong pointer address when calling arc_hdr_destroy() later.
3345 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3346 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3348 buf_discard_identity(hdr);
3349 kmem_cache_free(old, hdr);
3355 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3356 * The buf is returned thawed since we expect the consumer to modify it.
3359 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3361 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3362 ZIO_COMPRESS_OFF, type);
3363 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3365 arc_buf_t *buf = NULL;
3366 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3373 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3374 * for bufs containing metadata.
3377 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3378 enum zio_compress compression_type)
3380 ASSERT3U(lsize, >, 0);
3381 ASSERT3U(lsize, >=, psize);
3382 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3383 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3385 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3386 compression_type, ARC_BUFC_DATA);
3387 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3389 arc_buf_t *buf = NULL;
3390 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3392 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3394 if (!arc_buf_is_shared(buf)) {
3396 * To ensure that the hdr has the correct data in it if we call
3397 * arc_decompress() on this buf before it's been written to
3398 * disk, it's easiest if we just set up sharing between the
3401 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3402 arc_hdr_free_pabd(hdr);
3403 arc_share_buf(hdr, buf);
3410 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3412 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3413 l2arc_dev_t *dev = l2hdr->b_dev;
3414 uint64_t psize = arc_hdr_size(hdr);
3416 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3417 ASSERT(HDR_HAS_L2HDR(hdr));
3419 list_remove(&dev->l2ad_buflist, hdr);
3421 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3422 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3424 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3426 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3427 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3431 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3433 if (HDR_HAS_L1HDR(hdr)) {
3434 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3435 hdr->b_l1hdr.b_bufcnt > 0);
3436 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3437 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3439 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3440 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3442 if (!HDR_EMPTY(hdr))
3443 buf_discard_identity(hdr);
3445 if (HDR_HAS_L2HDR(hdr)) {
3446 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3447 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3450 mutex_enter(&dev->l2ad_mtx);
3453 * Even though we checked this conditional above, we
3454 * need to check this again now that we have the
3455 * l2ad_mtx. This is because we could be racing with
3456 * another thread calling l2arc_evict() which might have
3457 * destroyed this header's L2 portion as we were waiting
3458 * to acquire the l2ad_mtx. If that happens, we don't
3459 * want to re-destroy the header's L2 portion.
3461 if (HDR_HAS_L2HDR(hdr)) {
3463 arc_hdr_l2hdr_destroy(hdr);
3467 mutex_exit(&dev->l2ad_mtx);
3470 if (HDR_HAS_L1HDR(hdr)) {
3471 arc_cksum_free(hdr);
3473 while (hdr->b_l1hdr.b_buf != NULL)
3474 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3477 if (hdr->b_l1hdr.b_thawed != NULL) {
3478 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3479 hdr->b_l1hdr.b_thawed = NULL;
3483 if (hdr->b_l1hdr.b_pabd != NULL) {
3484 arc_hdr_free_pabd(hdr);
3488 ASSERT3P(hdr->b_hash_next, ==, NULL);
3489 if (HDR_HAS_L1HDR(hdr)) {
3490 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3491 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3492 kmem_cache_free(hdr_full_cache, hdr);
3494 kmem_cache_free(hdr_l2only_cache, hdr);
3499 arc_buf_destroy(arc_buf_t *buf, void* tag)
3501 arc_buf_hdr_t *hdr = buf->b_hdr;
3502 kmutex_t *hash_lock = HDR_LOCK(hdr);
3504 if (hdr->b_l1hdr.b_state == arc_anon) {
3505 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3506 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3507 VERIFY0(remove_reference(hdr, NULL, tag));
3508 arc_hdr_destroy(hdr);
3512 mutex_enter(hash_lock);
3513 ASSERT3P(hdr, ==, buf->b_hdr);
3514 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3515 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3516 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3517 ASSERT3P(buf->b_data, !=, NULL);
3519 (void) remove_reference(hdr, hash_lock, tag);
3520 arc_buf_destroy_impl(buf);
3521 mutex_exit(hash_lock);
3525 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3526 * state of the header is dependent on it's state prior to entering this
3527 * function. The following transitions are possible:
3529 * - arc_mru -> arc_mru_ghost
3530 * - arc_mfu -> arc_mfu_ghost
3531 * - arc_mru_ghost -> arc_l2c_only
3532 * - arc_mru_ghost -> deleted
3533 * - arc_mfu_ghost -> arc_l2c_only
3534 * - arc_mfu_ghost -> deleted
3537 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3539 arc_state_t *evicted_state, *state;
3540 int64_t bytes_evicted = 0;
3542 ASSERT(MUTEX_HELD(hash_lock));
3543 ASSERT(HDR_HAS_L1HDR(hdr));
3545 state = hdr->b_l1hdr.b_state;
3546 if (GHOST_STATE(state)) {
3547 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3548 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3551 * l2arc_write_buffers() relies on a header's L1 portion
3552 * (i.e. its b_pabd field) during it's write phase.
3553 * Thus, we cannot push a header onto the arc_l2c_only
3554 * state (removing it's L1 piece) until the header is
3555 * done being written to the l2arc.
3557 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3558 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3559 return (bytes_evicted);
3562 ARCSTAT_BUMP(arcstat_deleted);
3563 bytes_evicted += HDR_GET_LSIZE(hdr);
3565 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3567 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3568 if (HDR_HAS_L2HDR(hdr)) {
3570 * This buffer is cached on the 2nd Level ARC;
3571 * don't destroy the header.
3573 arc_change_state(arc_l2c_only, hdr, hash_lock);
3575 * dropping from L1+L2 cached to L2-only,
3576 * realloc to remove the L1 header.
3578 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3581 arc_change_state(arc_anon, hdr, hash_lock);
3582 arc_hdr_destroy(hdr);
3584 return (bytes_evicted);
3587 ASSERT(state == arc_mru || state == arc_mfu);
3588 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3590 /* prefetch buffers have a minimum lifespan */
3591 if (HDR_IO_IN_PROGRESS(hdr) ||
3592 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3593 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3594 arc_min_prefetch_lifespan)) {
3595 ARCSTAT_BUMP(arcstat_evict_skip);
3596 return (bytes_evicted);
3599 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3600 while (hdr->b_l1hdr.b_buf) {
3601 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3602 if (!mutex_tryenter(&buf->b_evict_lock)) {
3603 ARCSTAT_BUMP(arcstat_mutex_miss);
3606 if (buf->b_data != NULL)
3607 bytes_evicted += HDR_GET_LSIZE(hdr);
3608 mutex_exit(&buf->b_evict_lock);
3609 arc_buf_destroy_impl(buf);
3612 if (HDR_HAS_L2HDR(hdr)) {
3613 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3615 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3616 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3617 HDR_GET_LSIZE(hdr));
3619 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3620 HDR_GET_LSIZE(hdr));
3624 if (hdr->b_l1hdr.b_bufcnt == 0) {
3625 arc_cksum_free(hdr);
3627 bytes_evicted += arc_hdr_size(hdr);
3630 * If this hdr is being evicted and has a compressed
3631 * buffer then we discard it here before we change states.
3632 * This ensures that the accounting is updated correctly
3633 * in arc_free_data_impl().
3635 arc_hdr_free_pabd(hdr);
3637 arc_change_state(evicted_state, hdr, hash_lock);
3638 ASSERT(HDR_IN_HASH_TABLE(hdr));
3639 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3640 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3643 return (bytes_evicted);
3647 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3648 uint64_t spa, int64_t bytes)
3650 multilist_sublist_t *mls;
3651 uint64_t bytes_evicted = 0;
3653 kmutex_t *hash_lock;
3654 int evict_count = 0;
3656 ASSERT3P(marker, !=, NULL);
3657 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3659 mls = multilist_sublist_lock(ml, idx);
3661 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3662 hdr = multilist_sublist_prev(mls, marker)) {
3663 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3664 (evict_count >= zfs_arc_evict_batch_limit))
3668 * To keep our iteration location, move the marker
3669 * forward. Since we're not holding hdr's hash lock, we
3670 * must be very careful and not remove 'hdr' from the
3671 * sublist. Otherwise, other consumers might mistake the
3672 * 'hdr' as not being on a sublist when they call the
3673 * multilist_link_active() function (they all rely on
3674 * the hash lock protecting concurrent insertions and
3675 * removals). multilist_sublist_move_forward() was
3676 * specifically implemented to ensure this is the case
3677 * (only 'marker' will be removed and re-inserted).
3679 multilist_sublist_move_forward(mls, marker);
3682 * The only case where the b_spa field should ever be
3683 * zero, is the marker headers inserted by
3684 * arc_evict_state(). It's possible for multiple threads
3685 * to be calling arc_evict_state() concurrently (e.g.
3686 * dsl_pool_close() and zio_inject_fault()), so we must
3687 * skip any markers we see from these other threads.
3689 if (hdr->b_spa == 0)
3692 /* we're only interested in evicting buffers of a certain spa */
3693 if (spa != 0 && hdr->b_spa != spa) {
3694 ARCSTAT_BUMP(arcstat_evict_skip);
3698 hash_lock = HDR_LOCK(hdr);
3701 * We aren't calling this function from any code path
3702 * that would already be holding a hash lock, so we're
3703 * asserting on this assumption to be defensive in case
3704 * this ever changes. Without this check, it would be
3705 * possible to incorrectly increment arcstat_mutex_miss
3706 * below (e.g. if the code changed such that we called
3707 * this function with a hash lock held).
3709 ASSERT(!MUTEX_HELD(hash_lock));
3711 if (mutex_tryenter(hash_lock)) {
3712 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3713 mutex_exit(hash_lock);
3715 bytes_evicted += evicted;
3718 * If evicted is zero, arc_evict_hdr() must have
3719 * decided to skip this header, don't increment
3720 * evict_count in this case.
3726 * If arc_size isn't overflowing, signal any
3727 * threads that might happen to be waiting.
3729 * For each header evicted, we wake up a single
3730 * thread. If we used cv_broadcast, we could
3731 * wake up "too many" threads causing arc_size
3732 * to significantly overflow arc_c; since
3733 * arc_get_data_impl() doesn't check for overflow
3734 * when it's woken up (it doesn't because it's
3735 * possible for the ARC to be overflowing while
3736 * full of un-evictable buffers, and the
3737 * function should proceed in this case).
3739 * If threads are left sleeping, due to not
3740 * using cv_broadcast, they will be woken up
3741 * just before arc_reclaim_thread() sleeps.
3743 mutex_enter(&arc_reclaim_lock);
3744 if (!arc_is_overflowing())
3745 cv_signal(&arc_reclaim_waiters_cv);
3746 mutex_exit(&arc_reclaim_lock);
3748 ARCSTAT_BUMP(arcstat_mutex_miss);
3752 multilist_sublist_unlock(mls);
3754 return (bytes_evicted);
3758 * Evict buffers from the given arc state, until we've removed the
3759 * specified number of bytes. Move the removed buffers to the
3760 * appropriate evict state.
3762 * This function makes a "best effort". It skips over any buffers
3763 * it can't get a hash_lock on, and so, may not catch all candidates.
3764 * It may also return without evicting as much space as requested.
3766 * If bytes is specified using the special value ARC_EVICT_ALL, this
3767 * will evict all available (i.e. unlocked and evictable) buffers from
3768 * the given arc state; which is used by arc_flush().
3771 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3772 arc_buf_contents_t type)
3774 uint64_t total_evicted = 0;
3775 multilist_t *ml = state->arcs_list[type];
3777 arc_buf_hdr_t **markers;
3779 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3781 num_sublists = multilist_get_num_sublists(ml);
3784 * If we've tried to evict from each sublist, made some
3785 * progress, but still have not hit the target number of bytes
3786 * to evict, we want to keep trying. The markers allow us to
3787 * pick up where we left off for each individual sublist, rather
3788 * than starting from the tail each time.
3790 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3791 for (int i = 0; i < num_sublists; i++) {
3792 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3795 * A b_spa of 0 is used to indicate that this header is
3796 * a marker. This fact is used in arc_adjust_type() and
3797 * arc_evict_state_impl().
3799 markers[i]->b_spa = 0;
3801 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3802 multilist_sublist_insert_tail(mls, markers[i]);
3803 multilist_sublist_unlock(mls);
3807 * While we haven't hit our target number of bytes to evict, or
3808 * we're evicting all available buffers.
3810 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3812 * Start eviction using a randomly selected sublist,
3813 * this is to try and evenly balance eviction across all
3814 * sublists. Always starting at the same sublist
3815 * (e.g. index 0) would cause evictions to favor certain
3816 * sublists over others.
3818 int sublist_idx = multilist_get_random_index(ml);
3819 uint64_t scan_evicted = 0;
3821 for (int i = 0; i < num_sublists; i++) {
3822 uint64_t bytes_remaining;
3823 uint64_t bytes_evicted;
3825 if (bytes == ARC_EVICT_ALL)
3826 bytes_remaining = ARC_EVICT_ALL;
3827 else if (total_evicted < bytes)
3828 bytes_remaining = bytes - total_evicted;
3832 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3833 markers[sublist_idx], spa, bytes_remaining);
3835 scan_evicted += bytes_evicted;
3836 total_evicted += bytes_evicted;
3838 /* we've reached the end, wrap to the beginning */
3839 if (++sublist_idx >= num_sublists)
3844 * If we didn't evict anything during this scan, we have
3845 * no reason to believe we'll evict more during another
3846 * scan, so break the loop.
3848 if (scan_evicted == 0) {
3849 /* This isn't possible, let's make that obvious */
3850 ASSERT3S(bytes, !=, 0);
3853 * When bytes is ARC_EVICT_ALL, the only way to
3854 * break the loop is when scan_evicted is zero.
3855 * In that case, we actually have evicted enough,
3856 * so we don't want to increment the kstat.
3858 if (bytes != ARC_EVICT_ALL) {
3859 ASSERT3S(total_evicted, <, bytes);
3860 ARCSTAT_BUMP(arcstat_evict_not_enough);
3867 for (int i = 0; i < num_sublists; i++) {
3868 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3869 multilist_sublist_remove(mls, markers[i]);
3870 multilist_sublist_unlock(mls);
3872 kmem_cache_free(hdr_full_cache, markers[i]);
3874 kmem_free(markers, sizeof (*markers) * num_sublists);
3876 return (total_evicted);
3880 * Flush all "evictable" data of the given type from the arc state
3881 * specified. This will not evict any "active" buffers (i.e. referenced).
3883 * When 'retry' is set to B_FALSE, the function will make a single pass
3884 * over the state and evict any buffers that it can. Since it doesn't
3885 * continually retry the eviction, it might end up leaving some buffers
3886 * in the ARC due to lock misses.
3888 * When 'retry' is set to B_TRUE, the function will continually retry the
3889 * eviction until *all* evictable buffers have been removed from the
3890 * state. As a result, if concurrent insertions into the state are
3891 * allowed (e.g. if the ARC isn't shutting down), this function might
3892 * wind up in an infinite loop, continually trying to evict buffers.
3895 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3898 uint64_t evicted = 0;
3900 while (refcount_count(&state->arcs_esize[type]) != 0) {
3901 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3911 * Evict the specified number of bytes from the state specified,
3912 * restricting eviction to the spa and type given. This function
3913 * prevents us from trying to evict more from a state's list than
3914 * is "evictable", and to skip evicting altogether when passed a
3915 * negative value for "bytes". In contrast, arc_evict_state() will
3916 * evict everything it can, when passed a negative value for "bytes".
3919 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3920 arc_buf_contents_t type)
3924 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3925 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3926 return (arc_evict_state(state, spa, delta, type));
3933 * Evict metadata buffers from the cache, such that arc_meta_used is
3934 * capped by the arc_meta_limit tunable.
3937 arc_adjust_meta(uint64_t meta_used)
3939 uint64_t total_evicted = 0;
3943 * If we're over the meta limit, we want to evict enough
3944 * metadata to get back under the meta limit. We don't want to
3945 * evict so much that we drop the MRU below arc_p, though. If
3946 * we're over the meta limit more than we're over arc_p, we
3947 * evict some from the MRU here, and some from the MFU below.
3949 target = MIN((int64_t)(meta_used - arc_meta_limit),
3950 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3951 refcount_count(&arc_mru->arcs_size) - arc_p));
3953 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3956 * Similar to the above, we want to evict enough bytes to get us
3957 * below the meta limit, but not so much as to drop us below the
3958 * space allotted to the MFU (which is defined as arc_c - arc_p).
3960 target = MIN((int64_t)(meta_used - arc_meta_limit),
3961 (int64_t)(refcount_count(&arc_mfu->arcs_size) -
3964 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3966 return (total_evicted);
3970 * Return the type of the oldest buffer in the given arc state
3972 * This function will select a random sublist of type ARC_BUFC_DATA and
3973 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3974 * is compared, and the type which contains the "older" buffer will be
3977 static arc_buf_contents_t
3978 arc_adjust_type(arc_state_t *state)
3980 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
3981 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
3982 int data_idx = multilist_get_random_index(data_ml);
3983 int meta_idx = multilist_get_random_index(meta_ml);
3984 multilist_sublist_t *data_mls;
3985 multilist_sublist_t *meta_mls;
3986 arc_buf_contents_t type;
3987 arc_buf_hdr_t *data_hdr;
3988 arc_buf_hdr_t *meta_hdr;
3991 * We keep the sublist lock until we're finished, to prevent
3992 * the headers from being destroyed via arc_evict_state().
3994 data_mls = multilist_sublist_lock(data_ml, data_idx);
3995 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3998 * These two loops are to ensure we skip any markers that
3999 * might be at the tail of the lists due to arc_evict_state().
4002 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4003 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4004 if (data_hdr->b_spa != 0)
4008 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4009 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4010 if (meta_hdr->b_spa != 0)
4014 if (data_hdr == NULL && meta_hdr == NULL) {
4015 type = ARC_BUFC_DATA;
4016 } else if (data_hdr == NULL) {
4017 ASSERT3P(meta_hdr, !=, NULL);
4018 type = ARC_BUFC_METADATA;
4019 } else if (meta_hdr == NULL) {
4020 ASSERT3P(data_hdr, !=, NULL);
4021 type = ARC_BUFC_DATA;
4023 ASSERT3P(data_hdr, !=, NULL);
4024 ASSERT3P(meta_hdr, !=, NULL);
4026 /* The headers can't be on the sublist without an L1 header */
4027 ASSERT(HDR_HAS_L1HDR(data_hdr));
4028 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4030 if (data_hdr->b_l1hdr.b_arc_access <
4031 meta_hdr->b_l1hdr.b_arc_access) {
4032 type = ARC_BUFC_DATA;
4034 type = ARC_BUFC_METADATA;
4038 multilist_sublist_unlock(meta_mls);
4039 multilist_sublist_unlock(data_mls);
4045 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4050 uint64_t total_evicted = 0;
4053 uint64_t asize = aggsum_value(&arc_size);
4054 uint64_t ameta = aggsum_value(&arc_meta_used);
4057 * If we're over arc_meta_limit, we want to correct that before
4058 * potentially evicting data buffers below.
4060 total_evicted += arc_adjust_meta(ameta);
4065 * If we're over the target cache size, we want to evict enough
4066 * from the list to get back to our target size. We don't want
4067 * to evict too much from the MRU, such that it drops below
4068 * arc_p. So, if we're over our target cache size more than
4069 * the MRU is over arc_p, we'll evict enough to get back to
4070 * arc_p here, and then evict more from the MFU below.
4072 target = MIN((int64_t)(asize - arc_c),
4073 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4074 refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4077 * If we're below arc_meta_min, always prefer to evict data.
4078 * Otherwise, try to satisfy the requested number of bytes to
4079 * evict from the type which contains older buffers; in an
4080 * effort to keep newer buffers in the cache regardless of their
4081 * type. If we cannot satisfy the number of bytes from this
4082 * type, spill over into the next type.
4084 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4085 ameta > arc_meta_min) {
4086 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4087 total_evicted += bytes;
4090 * If we couldn't evict our target number of bytes from
4091 * metadata, we try to get the rest from data.
4096 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4098 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4099 total_evicted += bytes;
4102 * If we couldn't evict our target number of bytes from
4103 * data, we try to get the rest from metadata.
4108 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4114 * Now that we've tried to evict enough from the MRU to get its
4115 * size back to arc_p, if we're still above the target cache
4116 * size, we evict the rest from the MFU.
4118 target = asize - arc_c;
4120 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4121 ameta > arc_meta_min) {
4122 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4123 total_evicted += bytes;
4126 * If we couldn't evict our target number of bytes from
4127 * metadata, we try to get the rest from data.
4132 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4134 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4135 total_evicted += bytes;
4138 * If we couldn't evict our target number of bytes from
4139 * data, we try to get the rest from data.
4144 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4148 * Adjust ghost lists
4150 * In addition to the above, the ARC also defines target values
4151 * for the ghost lists. The sum of the mru list and mru ghost
4152 * list should never exceed the target size of the cache, and
4153 * the sum of the mru list, mfu list, mru ghost list, and mfu
4154 * ghost list should never exceed twice the target size of the
4155 * cache. The following logic enforces these limits on the ghost
4156 * caches, and evicts from them as needed.
4158 target = refcount_count(&arc_mru->arcs_size) +
4159 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4161 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4162 total_evicted += bytes;
4167 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4170 * We assume the sum of the mru list and mfu list is less than
4171 * or equal to arc_c (we enforced this above), which means we
4172 * can use the simpler of the two equations below:
4174 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4175 * mru ghost + mfu ghost <= arc_c
4177 target = refcount_count(&arc_mru_ghost->arcs_size) +
4178 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4180 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4181 total_evicted += bytes;
4186 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4188 return (total_evicted);
4192 arc_flush(spa_t *spa, boolean_t retry)
4197 * If retry is B_TRUE, a spa must not be specified since we have
4198 * no good way to determine if all of a spa's buffers have been
4199 * evicted from an arc state.
4201 ASSERT(!retry || spa == 0);
4204 guid = spa_load_guid(spa);
4206 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4207 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4209 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4210 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4212 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4213 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4215 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4216 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4220 arc_shrink(int64_t to_free)
4222 uint64_t asize = aggsum_value(&arc_size);
4223 if (arc_c > arc_c_min) {
4224 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
4225 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
4226 if (arc_c > arc_c_min + to_free)
4227 atomic_add_64(&arc_c, -to_free);
4231 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4233 arc_c = MAX(asize, arc_c_min);
4235 arc_p = (arc_c >> 1);
4237 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
4240 ASSERT(arc_c >= arc_c_min);
4241 ASSERT((int64_t)arc_p >= 0);
4244 if (asize > arc_c) {
4245 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, asize,
4247 (void) arc_adjust();
4251 typedef enum free_memory_reason_t {
4256 FMR_PAGES_PP_MAXIMUM,
4260 } free_memory_reason_t;
4262 int64_t last_free_memory;
4263 free_memory_reason_t last_free_reason;
4266 * Additional reserve of pages for pp_reserve.
4268 int64_t arc_pages_pp_reserve = 64;
4271 * Additional reserve of pages for swapfs.
4273 int64_t arc_swapfs_reserve = 64;
4276 * Return the amount of memory that can be consumed before reclaim will be
4277 * needed. Positive if there is sufficient free memory, negative indicates
4278 * the amount of memory that needs to be freed up.
4281 arc_available_memory(void)
4283 int64_t lowest = INT64_MAX;
4285 free_memory_reason_t r = FMR_UNKNOWN;
4290 * Cooperate with pagedaemon when it's time for it to scan
4291 * and reclaim some pages.
4293 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
4301 n = PAGESIZE * (-needfree);
4309 * check that we're out of range of the pageout scanner. It starts to
4310 * schedule paging if freemem is less than lotsfree and needfree.
4311 * lotsfree is the high-water mark for pageout, and needfree is the
4312 * number of needed free pages. We add extra pages here to make sure
4313 * the scanner doesn't start up while we're freeing memory.
4315 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4322 * check to make sure that swapfs has enough space so that anon
4323 * reservations can still succeed. anon_resvmem() checks that the
4324 * availrmem is greater than swapfs_minfree, and the number of reserved
4325 * swap pages. We also add a bit of extra here just to prevent
4326 * circumstances from getting really dire.
4328 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4329 desfree - arc_swapfs_reserve);
4332 r = FMR_SWAPFS_MINFREE;
4337 * Check that we have enough availrmem that memory locking (e.g., via
4338 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4339 * stores the number of pages that cannot be locked; when availrmem
4340 * drops below pages_pp_maximum, page locking mechanisms such as
4341 * page_pp_lock() will fail.)
4343 n = PAGESIZE * (availrmem - pages_pp_maximum -
4344 arc_pages_pp_reserve);
4347 r = FMR_PAGES_PP_MAXIMUM;
4350 #endif /* __FreeBSD__ */
4351 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4353 * If we're on an i386 platform, it's possible that we'll exhaust the
4354 * kernel heap space before we ever run out of available physical
4355 * memory. Most checks of the size of the heap_area compare against
4356 * tune.t_minarmem, which is the minimum available real memory that we
4357 * can have in the system. However, this is generally fixed at 25 pages
4358 * which is so low that it's useless. In this comparison, we seek to
4359 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4360 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4363 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
4364 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
4369 #define zio_arena NULL
4371 #define zio_arena heap_arena
4375 * If zio data pages are being allocated out of a separate heap segment,
4376 * then enforce that the size of available vmem for this arena remains
4377 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4379 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4380 * memory (in the zio_arena) free, which can avoid memory
4381 * fragmentation issues.
4383 if (zio_arena != NULL) {
4384 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4385 (vmem_size(zio_arena, VMEM_ALLOC) >>
4386 arc_zio_arena_free_shift);
4394 * Above limits know nothing about real level of KVA fragmentation.
4395 * Start aggressive reclamation if too little sequential KVA left.
4398 n = (vmem_size(heap_arena, VMEM_MAXFREE) < SPA_MAXBLOCKSIZE) ?
4399 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) :
4408 /* Every 100 calls, free a small amount */
4409 if (spa_get_random(100) == 0)
4411 #endif /* _KERNEL */
4413 last_free_memory = lowest;
4414 last_free_reason = r;
4415 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4421 * Determine if the system is under memory pressure and is asking
4422 * to reclaim memory. A return value of B_TRUE indicates that the system
4423 * is under memory pressure and that the arc should adjust accordingly.
4426 arc_reclaim_needed(void)
4428 return (arc_available_memory() < 0);
4431 extern kmem_cache_t *zio_buf_cache[];
4432 extern kmem_cache_t *zio_data_buf_cache[];
4433 extern kmem_cache_t *range_seg_cache;
4434 extern kmem_cache_t *abd_chunk_cache;
4436 static __noinline void
4437 arc_kmem_reap_now(void)
4440 kmem_cache_t *prev_cache = NULL;
4441 kmem_cache_t *prev_data_cache = NULL;
4443 DTRACE_PROBE(arc__kmem_reap_start);
4445 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4447 * We are exceeding our meta-data cache limit.
4448 * Purge some DNLC entries to release holds on meta-data.
4450 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4454 * Reclaim unused memory from all kmem caches.
4461 * If a kmem reap is already active, don't schedule more. We must
4462 * check for this because kmem_cache_reap_soon() won't actually
4463 * block on the cache being reaped (this is to prevent callers from
4464 * becoming implicitly blocked by a system-wide kmem reap -- which,
4465 * on a system with many, many full magazines, can take minutes).
4467 if (kmem_cache_reap_active())
4470 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4471 if (zio_buf_cache[i] != prev_cache) {
4472 prev_cache = zio_buf_cache[i];
4473 kmem_cache_reap_soon(zio_buf_cache[i]);
4475 if (zio_data_buf_cache[i] != prev_data_cache) {
4476 prev_data_cache = zio_data_buf_cache[i];
4477 kmem_cache_reap_soon(zio_data_buf_cache[i]);
4480 kmem_cache_reap_soon(abd_chunk_cache);
4481 kmem_cache_reap_soon(buf_cache);
4482 kmem_cache_reap_soon(hdr_full_cache);
4483 kmem_cache_reap_soon(hdr_l2only_cache);
4484 kmem_cache_reap_soon(range_seg_cache);
4487 if (zio_arena != NULL) {
4489 * Ask the vmem arena to reclaim unused memory from its
4492 vmem_qcache_reap(zio_arena);
4495 DTRACE_PROBE(arc__kmem_reap_end);
4499 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4500 * enough data and signal them to proceed. When this happens, the threads in
4501 * arc_get_data_impl() are sleeping while holding the hash lock for their
4502 * particular arc header. Thus, we must be careful to never sleep on a
4503 * hash lock in this thread. This is to prevent the following deadlock:
4505 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4506 * waiting for the reclaim thread to signal it.
4508 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4509 * fails, and goes to sleep forever.
4511 * This possible deadlock is avoided by always acquiring a hash lock
4512 * using mutex_tryenter() from arc_reclaim_thread().
4516 arc_reclaim_thread(void *unused __unused)
4518 hrtime_t growtime = 0;
4519 hrtime_t kmem_reap_time = 0;
4522 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4524 mutex_enter(&arc_reclaim_lock);
4525 while (!arc_reclaim_thread_exit) {
4526 uint64_t evicted = 0;
4529 * This is necessary in order for the mdb ::arc dcmd to
4530 * show up to date information. Since the ::arc command
4531 * does not call the kstat's update function, without
4532 * this call, the command may show stale stats for the
4533 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4534 * with this change, the data might be up to 1 second
4535 * out of date; but that should suffice. The arc_state_t
4536 * structures can be queried directly if more accurate
4537 * information is needed.
4539 if (arc_ksp != NULL)
4540 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4542 mutex_exit(&arc_reclaim_lock);
4545 * We call arc_adjust() before (possibly) calling
4546 * arc_kmem_reap_now(), so that we can wake up
4547 * arc_get_data_impl() sooner.
4549 evicted = arc_adjust();
4551 int64_t free_memory = arc_available_memory();
4552 if (free_memory < 0) {
4553 hrtime_t curtime = gethrtime();
4554 arc_no_grow = B_TRUE;
4558 * Wait at least zfs_grow_retry (default 60) seconds
4559 * before considering growing.
4561 growtime = curtime + SEC2NSEC(arc_grow_retry);
4564 * Wait at least arc_kmem_cache_reap_retry_ms
4565 * between arc_kmem_reap_now() calls. Without
4566 * this check it is possible to end up in a
4567 * situation where we spend lots of time
4568 * reaping caches, while we're near arc_c_min.
4570 if (curtime >= kmem_reap_time) {
4571 arc_kmem_reap_now();
4572 kmem_reap_time = gethrtime() +
4573 MSEC2NSEC(arc_kmem_cache_reap_retry_ms);
4577 * If we are still low on memory, shrink the ARC
4578 * so that we have arc_shrink_min free space.
4580 free_memory = arc_available_memory();
4583 (arc_c >> arc_shrink_shift) - free_memory;
4587 to_free = MAX(to_free, ptob(needfree));
4590 arc_shrink(to_free);
4592 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4593 arc_no_grow = B_TRUE;
4594 } else if (gethrtime() >= growtime) {
4595 arc_no_grow = B_FALSE;
4598 mutex_enter(&arc_reclaim_lock);
4601 * If evicted is zero, we couldn't evict anything via
4602 * arc_adjust(). This could be due to hash lock
4603 * collisions, but more likely due to the majority of
4604 * arc buffers being unevictable. Therefore, even if
4605 * arc_size is above arc_c, another pass is unlikely to
4606 * be helpful and could potentially cause us to enter an
4609 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) {
4611 * We're either no longer overflowing, or we
4612 * can't evict anything more, so we should wake
4613 * up any threads before we go to sleep.
4615 cv_broadcast(&arc_reclaim_waiters_cv);
4618 * Block until signaled, or after one second (we
4619 * might need to perform arc_kmem_reap_now()
4620 * even if we aren't being signalled)
4622 CALLB_CPR_SAFE_BEGIN(&cpr);
4623 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4624 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4625 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4629 arc_reclaim_thread_exit = B_FALSE;
4630 cv_broadcast(&arc_reclaim_thread_cv);
4631 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4635 static u_int arc_dnlc_evicts_arg;
4636 extern struct vfsops zfs_vfsops;
4639 arc_dnlc_evicts_thread(void *dummy __unused)
4644 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4646 mutex_enter(&arc_dnlc_evicts_lock);
4647 while (!arc_dnlc_evicts_thread_exit) {
4648 CALLB_CPR_SAFE_BEGIN(&cpr);
4649 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4650 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4651 if (arc_dnlc_evicts_arg != 0) {
4652 percent = arc_dnlc_evicts_arg;
4653 mutex_exit(&arc_dnlc_evicts_lock);
4655 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4657 mutex_enter(&arc_dnlc_evicts_lock);
4659 * Clear our token only after vnlru_free()
4660 * pass is done, to avoid false queueing of
4663 arc_dnlc_evicts_arg = 0;
4666 arc_dnlc_evicts_thread_exit = FALSE;
4667 cv_broadcast(&arc_dnlc_evicts_cv);
4668 CALLB_CPR_EXIT(&cpr);
4673 dnlc_reduce_cache(void *arg)
4677 percent = (u_int)(uintptr_t)arg;
4678 mutex_enter(&arc_dnlc_evicts_lock);
4679 if (arc_dnlc_evicts_arg == 0) {
4680 arc_dnlc_evicts_arg = percent;
4681 cv_broadcast(&arc_dnlc_evicts_cv);
4683 mutex_exit(&arc_dnlc_evicts_lock);
4687 * Adapt arc info given the number of bytes we are trying to add and
4688 * the state that we are comming from. This function is only called
4689 * when we are adding new content to the cache.
4692 arc_adapt(int bytes, arc_state_t *state)
4695 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4696 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4697 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4699 if (state == arc_l2c_only)
4704 * Adapt the target size of the MRU list:
4705 * - if we just hit in the MRU ghost list, then increase
4706 * the target size of the MRU list.
4707 * - if we just hit in the MFU ghost list, then increase
4708 * the target size of the MFU list by decreasing the
4709 * target size of the MRU list.
4711 if (state == arc_mru_ghost) {
4712 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4713 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4715 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4716 } else if (state == arc_mfu_ghost) {
4719 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4720 mult = MIN(mult, 10);
4722 delta = MIN(bytes * mult, arc_p);
4723 arc_p = MAX(arc_p_min, arc_p - delta);
4725 ASSERT((int64_t)arc_p >= 0);
4727 if (arc_reclaim_needed()) {
4728 cv_signal(&arc_reclaim_thread_cv);
4735 if (arc_c >= arc_c_max)
4739 * If we're within (2 * maxblocksize) bytes of the target
4740 * cache size, increment the target cache size
4742 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
4744 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
4745 atomic_add_64(&arc_c, (int64_t)bytes);
4746 if (arc_c > arc_c_max)
4748 else if (state == arc_anon)
4749 atomic_add_64(&arc_p, (int64_t)bytes);
4753 ASSERT((int64_t)arc_p >= 0);
4757 * Check if arc_size has grown past our upper threshold, determined by
4758 * zfs_arc_overflow_shift.
4761 arc_is_overflowing(void)
4763 /* Always allow at least one block of overflow */
4764 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4765 arc_c >> zfs_arc_overflow_shift);
4768 * We just compare the lower bound here for performance reasons. Our
4769 * primary goals are to make sure that the arc never grows without
4770 * bound, and that it can reach its maximum size. This check
4771 * accomplishes both goals. The maximum amount we could run over by is
4772 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4773 * in the ARC. In practice, that's in the tens of MB, which is low
4774 * enough to be safe.
4776 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
4780 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4782 arc_buf_contents_t type = arc_buf_type(hdr);
4784 arc_get_data_impl(hdr, size, tag);
4785 if (type == ARC_BUFC_METADATA) {
4786 return (abd_alloc(size, B_TRUE));
4788 ASSERT(type == ARC_BUFC_DATA);
4789 return (abd_alloc(size, B_FALSE));
4794 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4796 arc_buf_contents_t type = arc_buf_type(hdr);
4798 arc_get_data_impl(hdr, size, tag);
4799 if (type == ARC_BUFC_METADATA) {
4800 return (zio_buf_alloc(size));
4802 ASSERT(type == ARC_BUFC_DATA);
4803 return (zio_data_buf_alloc(size));
4808 * Allocate a block and return it to the caller. If we are hitting the
4809 * hard limit for the cache size, we must sleep, waiting for the eviction
4810 * thread to catch up. If we're past the target size but below the hard
4811 * limit, we'll only signal the reclaim thread and continue on.
4814 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4816 arc_state_t *state = hdr->b_l1hdr.b_state;
4817 arc_buf_contents_t type = arc_buf_type(hdr);
4819 arc_adapt(size, state);
4822 * If arc_size is currently overflowing, and has grown past our
4823 * upper limit, we must be adding data faster than the evict
4824 * thread can evict. Thus, to ensure we don't compound the
4825 * problem by adding more data and forcing arc_size to grow even
4826 * further past it's target size, we halt and wait for the
4827 * eviction thread to catch up.
4829 * It's also possible that the reclaim thread is unable to evict
4830 * enough buffers to get arc_size below the overflow limit (e.g.
4831 * due to buffers being un-evictable, or hash lock collisions).
4832 * In this case, we want to proceed regardless if we're
4833 * overflowing; thus we don't use a while loop here.
4835 if (arc_is_overflowing()) {
4836 mutex_enter(&arc_reclaim_lock);
4839 * Now that we've acquired the lock, we may no longer be
4840 * over the overflow limit, lets check.
4842 * We're ignoring the case of spurious wake ups. If that
4843 * were to happen, it'd let this thread consume an ARC
4844 * buffer before it should have (i.e. before we're under
4845 * the overflow limit and were signalled by the reclaim
4846 * thread). As long as that is a rare occurrence, it
4847 * shouldn't cause any harm.
4849 if (arc_is_overflowing()) {
4850 cv_signal(&arc_reclaim_thread_cv);
4851 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4854 mutex_exit(&arc_reclaim_lock);
4857 VERIFY3U(hdr->b_type, ==, type);
4858 if (type == ARC_BUFC_METADATA) {
4859 arc_space_consume(size, ARC_SPACE_META);
4861 arc_space_consume(size, ARC_SPACE_DATA);
4865 * Update the state size. Note that ghost states have a
4866 * "ghost size" and so don't need to be updated.
4868 if (!GHOST_STATE(state)) {
4870 (void) refcount_add_many(&state->arcs_size, size, tag);
4873 * If this is reached via arc_read, the link is
4874 * protected by the hash lock. If reached via
4875 * arc_buf_alloc, the header should not be accessed by
4876 * any other thread. And, if reached via arc_read_done,
4877 * the hash lock will protect it if it's found in the
4878 * hash table; otherwise no other thread should be
4879 * trying to [add|remove]_reference it.
4881 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4882 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4883 (void) refcount_add_many(&state->arcs_esize[type],
4888 * If we are growing the cache, and we are adding anonymous
4889 * data, and we have outgrown arc_p, update arc_p
4891 if (aggsum_compare(&arc_size, arc_c) < 0 &&
4892 hdr->b_l1hdr.b_state == arc_anon &&
4893 (refcount_count(&arc_anon->arcs_size) +
4894 refcount_count(&arc_mru->arcs_size) > arc_p))
4895 arc_p = MIN(arc_c, arc_p + size);
4897 ARCSTAT_BUMP(arcstat_allocated);
4901 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4903 arc_free_data_impl(hdr, size, tag);
4908 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4910 arc_buf_contents_t type = arc_buf_type(hdr);
4912 arc_free_data_impl(hdr, size, tag);
4913 if (type == ARC_BUFC_METADATA) {
4914 zio_buf_free(buf, size);
4916 ASSERT(type == ARC_BUFC_DATA);
4917 zio_data_buf_free(buf, size);
4922 * Free the arc data buffer.
4925 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4927 arc_state_t *state = hdr->b_l1hdr.b_state;
4928 arc_buf_contents_t type = arc_buf_type(hdr);
4930 /* protected by hash lock, if in the hash table */
4931 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4932 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4933 ASSERT(state != arc_anon && state != arc_l2c_only);
4935 (void) refcount_remove_many(&state->arcs_esize[type],
4938 (void) refcount_remove_many(&state->arcs_size, size, tag);
4940 VERIFY3U(hdr->b_type, ==, type);
4941 if (type == ARC_BUFC_METADATA) {
4942 arc_space_return(size, ARC_SPACE_META);
4944 ASSERT(type == ARC_BUFC_DATA);
4945 arc_space_return(size, ARC_SPACE_DATA);
4950 * This routine is called whenever a buffer is accessed.
4951 * NOTE: the hash lock is dropped in this function.
4954 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4958 ASSERT(MUTEX_HELD(hash_lock));
4959 ASSERT(HDR_HAS_L1HDR(hdr));
4961 if (hdr->b_l1hdr.b_state == arc_anon) {
4963 * This buffer is not in the cache, and does not
4964 * appear in our "ghost" list. Add the new buffer
4968 ASSERT0(hdr->b_l1hdr.b_arc_access);
4969 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4970 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4971 arc_change_state(arc_mru, hdr, hash_lock);
4973 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4974 now = ddi_get_lbolt();
4977 * If this buffer is here because of a prefetch, then either:
4978 * - clear the flag if this is a "referencing" read
4979 * (any subsequent access will bump this into the MFU state).
4981 * - move the buffer to the head of the list if this is
4982 * another prefetch (to make it less likely to be evicted).
4984 if (HDR_PREFETCH(hdr)) {
4985 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4986 /* link protected by hash lock */
4987 ASSERT(multilist_link_active(
4988 &hdr->b_l1hdr.b_arc_node));
4990 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4991 ARCSTAT_BUMP(arcstat_mru_hits);
4993 hdr->b_l1hdr.b_arc_access = now;
4998 * This buffer has been "accessed" only once so far,
4999 * but it is still in the cache. Move it to the MFU
5002 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
5004 * More than 125ms have passed since we
5005 * instantiated this buffer. Move it to the
5006 * most frequently used state.
5008 hdr->b_l1hdr.b_arc_access = now;
5009 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5010 arc_change_state(arc_mfu, hdr, hash_lock);
5012 ARCSTAT_BUMP(arcstat_mru_hits);
5013 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5014 arc_state_t *new_state;
5016 * This buffer has been "accessed" recently, but
5017 * was evicted from the cache. Move it to the
5021 if (HDR_PREFETCH(hdr)) {
5022 new_state = arc_mru;
5023 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
5024 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
5025 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5027 new_state = arc_mfu;
5028 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5031 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5032 arc_change_state(new_state, hdr, hash_lock);
5034 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5035 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5037 * This buffer has been accessed more than once and is
5038 * still in the cache. Keep it in the MFU state.
5040 * NOTE: an add_reference() that occurred when we did
5041 * the arc_read() will have kicked this off the list.
5042 * If it was a prefetch, we will explicitly move it to
5043 * the head of the list now.
5045 if ((HDR_PREFETCH(hdr)) != 0) {
5046 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5047 /* link protected by hash_lock */
5048 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5050 ARCSTAT_BUMP(arcstat_mfu_hits);
5051 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5052 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5053 arc_state_t *new_state = arc_mfu;
5055 * This buffer has been accessed more than once but has
5056 * been evicted from the cache. Move it back to the
5060 if (HDR_PREFETCH(hdr)) {
5062 * This is a prefetch access...
5063 * move this block back to the MRU state.
5065 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
5066 new_state = arc_mru;
5069 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5070 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5071 arc_change_state(new_state, hdr, hash_lock);
5073 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5074 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5076 * This buffer is on the 2nd Level ARC.
5079 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5080 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5081 arc_change_state(arc_mfu, hdr, hash_lock);
5083 ASSERT(!"invalid arc state");
5087 /* a generic arc_done_func_t which you can use */
5090 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
5092 if (zio == NULL || zio->io_error == 0)
5093 bcopy(buf->b_data, arg, arc_buf_size(buf));
5094 arc_buf_destroy(buf, arg);
5097 /* a generic arc_done_func_t */
5099 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
5101 arc_buf_t **bufp = arg;
5102 if (zio && zio->io_error) {
5103 arc_buf_destroy(buf, arg);
5107 ASSERT(buf->b_data);
5112 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5114 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5115 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5116 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
5118 if (HDR_COMPRESSION_ENABLED(hdr)) {
5119 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
5120 BP_GET_COMPRESS(bp));
5122 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5123 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5128 arc_read_done(zio_t *zio)
5130 arc_buf_hdr_t *hdr = zio->io_private;
5131 kmutex_t *hash_lock = NULL;
5132 arc_callback_t *callback_list;
5133 arc_callback_t *acb;
5134 boolean_t freeable = B_FALSE;
5135 boolean_t no_zio_error = (zio->io_error == 0);
5138 * The hdr was inserted into hash-table and removed from lists
5139 * prior to starting I/O. We should find this header, since
5140 * it's in the hash table, and it should be legit since it's
5141 * not possible to evict it during the I/O. The only possible
5142 * reason for it not to be found is if we were freed during the
5145 if (HDR_IN_HASH_TABLE(hdr)) {
5146 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5147 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5148 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5149 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5150 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5152 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5155 ASSERT((found == hdr &&
5156 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5157 (found == hdr && HDR_L2_READING(hdr)));
5158 ASSERT3P(hash_lock, !=, NULL);
5162 /* byteswap if necessary */
5163 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5164 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5165 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5167 hdr->b_l1hdr.b_byteswap =
5168 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5171 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5175 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5176 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5177 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5179 callback_list = hdr->b_l1hdr.b_acb;
5180 ASSERT3P(callback_list, !=, NULL);
5182 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
5184 * Only call arc_access on anonymous buffers. This is because
5185 * if we've issued an I/O for an evicted buffer, we've already
5186 * called arc_access (to prevent any simultaneous readers from
5187 * getting confused).
5189 arc_access(hdr, hash_lock);
5193 * If a read request has a callback (i.e. acb_done is not NULL), then we
5194 * make a buf containing the data according to the parameters which were
5195 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5196 * aren't needlessly decompressing the data multiple times.
5198 int callback_cnt = 0;
5199 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5203 /* This is a demand read since prefetches don't use callbacks */
5206 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
5207 acb->acb_compressed, no_zio_error, &acb->acb_buf);
5209 zio->io_error = error;
5212 hdr->b_l1hdr.b_acb = NULL;
5213 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5214 if (callback_cnt == 0) {
5215 ASSERT(HDR_PREFETCH(hdr));
5216 ASSERT0(hdr->b_l1hdr.b_bufcnt);
5217 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5220 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5221 callback_list != NULL);
5224 arc_hdr_verify(hdr, zio->io_bp);
5226 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5227 if (hdr->b_l1hdr.b_state != arc_anon)
5228 arc_change_state(arc_anon, hdr, hash_lock);
5229 if (HDR_IN_HASH_TABLE(hdr))
5230 buf_hash_remove(hdr);
5231 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5235 * Broadcast before we drop the hash_lock to avoid the possibility
5236 * that the hdr (and hence the cv) might be freed before we get to
5237 * the cv_broadcast().
5239 cv_broadcast(&hdr->b_l1hdr.b_cv);
5241 if (hash_lock != NULL) {
5242 mutex_exit(hash_lock);
5245 * This block was freed while we waited for the read to
5246 * complete. It has been removed from the hash table and
5247 * moved to the anonymous state (so that it won't show up
5250 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5251 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5254 /* execute each callback and free its structure */
5255 while ((acb = callback_list) != NULL) {
5257 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
5259 if (acb->acb_zio_dummy != NULL) {
5260 acb->acb_zio_dummy->io_error = zio->io_error;
5261 zio_nowait(acb->acb_zio_dummy);
5264 callback_list = acb->acb_next;
5265 kmem_free(acb, sizeof (arc_callback_t));
5269 arc_hdr_destroy(hdr);
5273 * "Read" the block at the specified DVA (in bp) via the
5274 * cache. If the block is found in the cache, invoke the provided
5275 * callback immediately and return. Note that the `zio' parameter
5276 * in the callback will be NULL in this case, since no IO was
5277 * required. If the block is not in the cache pass the read request
5278 * on to the spa with a substitute callback function, so that the
5279 * requested block will be added to the cache.
5281 * If a read request arrives for a block that has a read in-progress,
5282 * either wait for the in-progress read to complete (and return the
5283 * results); or, if this is a read with a "done" func, add a record
5284 * to the read to invoke the "done" func when the read completes,
5285 * and return; or just return.
5287 * arc_read_done() will invoke all the requested "done" functions
5288 * for readers of this block.
5291 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
5292 void *private, zio_priority_t priority, int zio_flags,
5293 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5295 arc_buf_hdr_t *hdr = NULL;
5296 kmutex_t *hash_lock = NULL;
5298 uint64_t guid = spa_load_guid(spa);
5299 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5301 ASSERT(!BP_IS_EMBEDDED(bp) ||
5302 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5305 if (!BP_IS_EMBEDDED(bp)) {
5307 * Embedded BP's have no DVA and require no I/O to "read".
5308 * Create an anonymous arc buf to back it.
5310 hdr = buf_hash_find(guid, bp, &hash_lock);
5313 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5314 arc_buf_t *buf = NULL;
5315 *arc_flags |= ARC_FLAG_CACHED;
5317 if (HDR_IO_IN_PROGRESS(hdr)) {
5319 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5320 priority == ZIO_PRIORITY_SYNC_READ) {
5322 * This sync read must wait for an
5323 * in-progress async read (e.g. a predictive
5324 * prefetch). Async reads are queued
5325 * separately at the vdev_queue layer, so
5326 * this is a form of priority inversion.
5327 * Ideally, we would "inherit" the demand
5328 * i/o's priority by moving the i/o from
5329 * the async queue to the synchronous queue,
5330 * but there is currently no mechanism to do
5331 * so. Track this so that we can evaluate
5332 * the magnitude of this potential performance
5335 * Note that if the prefetch i/o is already
5336 * active (has been issued to the device),
5337 * the prefetch improved performance, because
5338 * we issued it sooner than we would have
5339 * without the prefetch.
5341 DTRACE_PROBE1(arc__sync__wait__for__async,
5342 arc_buf_hdr_t *, hdr);
5343 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
5345 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5346 arc_hdr_clear_flags(hdr,
5347 ARC_FLAG_PREDICTIVE_PREFETCH);
5350 if (*arc_flags & ARC_FLAG_WAIT) {
5351 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5352 mutex_exit(hash_lock);
5355 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5358 arc_callback_t *acb = NULL;
5360 acb = kmem_zalloc(sizeof (arc_callback_t),
5362 acb->acb_done = done;
5363 acb->acb_private = private;
5364 acb->acb_compressed = compressed_read;
5366 acb->acb_zio_dummy = zio_null(pio,
5367 spa, NULL, NULL, NULL, zio_flags);
5369 ASSERT3P(acb->acb_done, !=, NULL);
5370 acb->acb_next = hdr->b_l1hdr.b_acb;
5371 hdr->b_l1hdr.b_acb = acb;
5372 mutex_exit(hash_lock);
5375 mutex_exit(hash_lock);
5379 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5380 hdr->b_l1hdr.b_state == arc_mfu);
5383 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5385 * This is a demand read which does not have to
5386 * wait for i/o because we did a predictive
5387 * prefetch i/o for it, which has completed.
5390 arc__demand__hit__predictive__prefetch,
5391 arc_buf_hdr_t *, hdr);
5393 arcstat_demand_hit_predictive_prefetch);
5394 arc_hdr_clear_flags(hdr,
5395 ARC_FLAG_PREDICTIVE_PREFETCH);
5397 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5399 /* Get a buf with the desired data in it. */
5400 VERIFY0(arc_buf_alloc_impl(hdr, private,
5401 compressed_read, B_TRUE, &buf));
5402 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5403 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5404 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5406 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5407 arc_access(hdr, hash_lock);
5408 if (*arc_flags & ARC_FLAG_L2CACHE)
5409 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5410 mutex_exit(hash_lock);
5411 ARCSTAT_BUMP(arcstat_hits);
5412 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5413 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5414 data, metadata, hits);
5417 done(NULL, buf, private);
5419 uint64_t lsize = BP_GET_LSIZE(bp);
5420 uint64_t psize = BP_GET_PSIZE(bp);
5421 arc_callback_t *acb;
5424 boolean_t devw = B_FALSE;
5428 /* this block is not in the cache */
5429 arc_buf_hdr_t *exists = NULL;
5430 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5431 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5432 BP_GET_COMPRESS(bp), type);
5434 if (!BP_IS_EMBEDDED(bp)) {
5435 hdr->b_dva = *BP_IDENTITY(bp);
5436 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5437 exists = buf_hash_insert(hdr, &hash_lock);
5439 if (exists != NULL) {
5440 /* somebody beat us to the hash insert */
5441 mutex_exit(hash_lock);
5442 buf_discard_identity(hdr);
5443 arc_hdr_destroy(hdr);
5444 goto top; /* restart the IO request */
5448 * This block is in the ghost cache. If it was L2-only
5449 * (and thus didn't have an L1 hdr), we realloc the
5450 * header to add an L1 hdr.
5452 if (!HDR_HAS_L1HDR(hdr)) {
5453 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5456 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5457 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5458 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5459 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5460 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5461 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5464 * This is a delicate dance that we play here.
5465 * This hdr is in the ghost list so we access it
5466 * to move it out of the ghost list before we
5467 * initiate the read. If it's a prefetch then
5468 * it won't have a callback so we'll remove the
5469 * reference that arc_buf_alloc_impl() created. We
5470 * do this after we've called arc_access() to
5471 * avoid hitting an assert in remove_reference().
5473 arc_access(hdr, hash_lock);
5474 arc_hdr_alloc_pabd(hdr);
5476 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5477 size = arc_hdr_size(hdr);
5480 * If compression is enabled on the hdr, then will do
5481 * RAW I/O and will store the compressed data in the hdr's
5482 * data block. Otherwise, the hdr's data block will contain
5483 * the uncompressed data.
5485 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5486 zio_flags |= ZIO_FLAG_RAW;
5489 if (*arc_flags & ARC_FLAG_PREFETCH)
5490 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5491 if (*arc_flags & ARC_FLAG_L2CACHE)
5492 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5493 if (BP_GET_LEVEL(bp) > 0)
5494 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5495 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5496 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5497 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5499 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5500 acb->acb_done = done;
5501 acb->acb_private = private;
5502 acb->acb_compressed = compressed_read;
5504 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5505 hdr->b_l1hdr.b_acb = acb;
5506 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5508 if (HDR_HAS_L2HDR(hdr) &&
5509 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5510 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5511 addr = hdr->b_l2hdr.b_daddr;
5513 * Lock out L2ARC device removal.
5515 if (vdev_is_dead(vd) ||
5516 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5520 if (priority == ZIO_PRIORITY_ASYNC_READ)
5521 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5523 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5525 if (hash_lock != NULL)
5526 mutex_exit(hash_lock);
5529 * At this point, we have a level 1 cache miss. Try again in
5530 * L2ARC if possible.
5532 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5534 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5535 uint64_t, lsize, zbookmark_phys_t *, zb);
5536 ARCSTAT_BUMP(arcstat_misses);
5537 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5538 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5539 data, metadata, misses);
5544 racct_add_force(curproc, RACCT_READBPS, size);
5545 racct_add_force(curproc, RACCT_READIOPS, 1);
5546 PROC_UNLOCK(curproc);
5549 curthread->td_ru.ru_inblock++;
5552 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5554 * Read from the L2ARC if the following are true:
5555 * 1. The L2ARC vdev was previously cached.
5556 * 2. This buffer still has L2ARC metadata.
5557 * 3. This buffer isn't currently writing to the L2ARC.
5558 * 4. The L2ARC entry wasn't evicted, which may
5559 * also have invalidated the vdev.
5560 * 5. This isn't prefetch and l2arc_noprefetch is set.
5562 if (HDR_HAS_L2HDR(hdr) &&
5563 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5564 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5565 l2arc_read_callback_t *cb;
5569 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5570 ARCSTAT_BUMP(arcstat_l2_hits);
5572 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5574 cb->l2rcb_hdr = hdr;
5577 cb->l2rcb_flags = zio_flags;
5579 asize = vdev_psize_to_asize(vd, size);
5580 if (asize != size) {
5581 abd = abd_alloc_for_io(asize,
5582 HDR_ISTYPE_METADATA(hdr));
5583 cb->l2rcb_abd = abd;
5585 abd = hdr->b_l1hdr.b_pabd;
5588 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5589 addr + asize <= vd->vdev_psize -
5590 VDEV_LABEL_END_SIZE);
5593 * l2arc read. The SCL_L2ARC lock will be
5594 * released by l2arc_read_done().
5595 * Issue a null zio if the underlying buffer
5596 * was squashed to zero size by compression.
5598 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5599 ZIO_COMPRESS_EMPTY);
5600 rzio = zio_read_phys(pio, vd, addr,
5603 l2arc_read_done, cb, priority,
5604 zio_flags | ZIO_FLAG_DONT_CACHE |
5606 ZIO_FLAG_DONT_PROPAGATE |
5607 ZIO_FLAG_DONT_RETRY, B_FALSE);
5608 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5610 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5612 if (*arc_flags & ARC_FLAG_NOWAIT) {
5617 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5618 if (zio_wait(rzio) == 0)
5621 /* l2arc read error; goto zio_read() */
5623 DTRACE_PROBE1(l2arc__miss,
5624 arc_buf_hdr_t *, hdr);
5625 ARCSTAT_BUMP(arcstat_l2_misses);
5626 if (HDR_L2_WRITING(hdr))
5627 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5628 spa_config_exit(spa, SCL_L2ARC, vd);
5632 spa_config_exit(spa, SCL_L2ARC, vd);
5633 if (l2arc_ndev != 0) {
5634 DTRACE_PROBE1(l2arc__miss,
5635 arc_buf_hdr_t *, hdr);
5636 ARCSTAT_BUMP(arcstat_l2_misses);
5640 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
5641 arc_read_done, hdr, priority, zio_flags, zb);
5643 if (*arc_flags & ARC_FLAG_WAIT)
5644 return (zio_wait(rzio));
5646 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5653 * Notify the arc that a block was freed, and thus will never be used again.
5656 arc_freed(spa_t *spa, const blkptr_t *bp)
5659 kmutex_t *hash_lock;
5660 uint64_t guid = spa_load_guid(spa);
5662 ASSERT(!BP_IS_EMBEDDED(bp));
5664 hdr = buf_hash_find(guid, bp, &hash_lock);
5669 * We might be trying to free a block that is still doing I/O
5670 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5671 * dmu_sync-ed block). If this block is being prefetched, then it
5672 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5673 * until the I/O completes. A block may also have a reference if it is
5674 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5675 * have written the new block to its final resting place on disk but
5676 * without the dedup flag set. This would have left the hdr in the MRU
5677 * state and discoverable. When the txg finally syncs it detects that
5678 * the block was overridden in open context and issues an override I/O.
5679 * Since this is a dedup block, the override I/O will determine if the
5680 * block is already in the DDT. If so, then it will replace the io_bp
5681 * with the bp from the DDT and allow the I/O to finish. When the I/O
5682 * reaches the done callback, dbuf_write_override_done, it will
5683 * check to see if the io_bp and io_bp_override are identical.
5684 * If they are not, then it indicates that the bp was replaced with
5685 * the bp in the DDT and the override bp is freed. This allows
5686 * us to arrive here with a reference on a block that is being
5687 * freed. So if we have an I/O in progress, or a reference to
5688 * this hdr, then we don't destroy the hdr.
5690 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5691 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5692 arc_change_state(arc_anon, hdr, hash_lock);
5693 arc_hdr_destroy(hdr);
5694 mutex_exit(hash_lock);
5696 mutex_exit(hash_lock);
5702 * Release this buffer from the cache, making it an anonymous buffer. This
5703 * must be done after a read and prior to modifying the buffer contents.
5704 * If the buffer has more than one reference, we must make
5705 * a new hdr for the buffer.
5708 arc_release(arc_buf_t *buf, void *tag)
5710 arc_buf_hdr_t *hdr = buf->b_hdr;
5713 * It would be nice to assert that if it's DMU metadata (level >
5714 * 0 || it's the dnode file), then it must be syncing context.
5715 * But we don't know that information at this level.
5718 mutex_enter(&buf->b_evict_lock);
5720 ASSERT(HDR_HAS_L1HDR(hdr));
5723 * We don't grab the hash lock prior to this check, because if
5724 * the buffer's header is in the arc_anon state, it won't be
5725 * linked into the hash table.
5727 if (hdr->b_l1hdr.b_state == arc_anon) {
5728 mutex_exit(&buf->b_evict_lock);
5729 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5730 ASSERT(!HDR_IN_HASH_TABLE(hdr));
5731 ASSERT(!HDR_HAS_L2HDR(hdr));
5732 ASSERT(HDR_EMPTY(hdr));
5733 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5734 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5735 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5737 hdr->b_l1hdr.b_arc_access = 0;
5740 * If the buf is being overridden then it may already
5741 * have a hdr that is not empty.
5743 buf_discard_identity(hdr);
5749 kmutex_t *hash_lock = HDR_LOCK(hdr);
5750 mutex_enter(hash_lock);
5753 * This assignment is only valid as long as the hash_lock is
5754 * held, we must be careful not to reference state or the
5755 * b_state field after dropping the lock.
5757 arc_state_t *state = hdr->b_l1hdr.b_state;
5758 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5759 ASSERT3P(state, !=, arc_anon);
5761 /* this buffer is not on any list */
5762 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
5764 if (HDR_HAS_L2HDR(hdr)) {
5765 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5768 * We have to recheck this conditional again now that
5769 * we're holding the l2ad_mtx to prevent a race with
5770 * another thread which might be concurrently calling
5771 * l2arc_evict(). In that case, l2arc_evict() might have
5772 * destroyed the header's L2 portion as we were waiting
5773 * to acquire the l2ad_mtx.
5775 if (HDR_HAS_L2HDR(hdr)) {
5777 arc_hdr_l2hdr_destroy(hdr);
5780 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5784 * Do we have more than one buf?
5786 if (hdr->b_l1hdr.b_bufcnt > 1) {
5787 arc_buf_hdr_t *nhdr;
5788 uint64_t spa = hdr->b_spa;
5789 uint64_t psize = HDR_GET_PSIZE(hdr);
5790 uint64_t lsize = HDR_GET_LSIZE(hdr);
5791 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5792 arc_buf_contents_t type = arc_buf_type(hdr);
5793 VERIFY3U(hdr->b_type, ==, type);
5795 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5796 (void) remove_reference(hdr, hash_lock, tag);
5798 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
5799 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5800 ASSERT(ARC_BUF_LAST(buf));
5804 * Pull the data off of this hdr and attach it to
5805 * a new anonymous hdr. Also find the last buffer
5806 * in the hdr's buffer list.
5808 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
5809 ASSERT3P(lastbuf, !=, NULL);
5812 * If the current arc_buf_t and the hdr are sharing their data
5813 * buffer, then we must stop sharing that block.
5815 if (arc_buf_is_shared(buf)) {
5816 VERIFY(!arc_buf_is_shared(lastbuf));
5819 * First, sever the block sharing relationship between
5820 * buf and the arc_buf_hdr_t.
5822 arc_unshare_buf(hdr, buf);
5825 * Now we need to recreate the hdr's b_pabd. Since we
5826 * have lastbuf handy, we try to share with it, but if
5827 * we can't then we allocate a new b_pabd and copy the
5828 * data from buf into it.
5830 if (arc_can_share(hdr, lastbuf)) {
5831 arc_share_buf(hdr, lastbuf);
5833 arc_hdr_alloc_pabd(hdr);
5834 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
5835 buf->b_data, psize);
5837 VERIFY3P(lastbuf->b_data, !=, NULL);
5838 } else if (HDR_SHARED_DATA(hdr)) {
5840 * Uncompressed shared buffers are always at the end
5841 * of the list. Compressed buffers don't have the
5842 * same requirements. This makes it hard to
5843 * simply assert that the lastbuf is shared so
5844 * we rely on the hdr's compression flags to determine
5845 * if we have a compressed, shared buffer.
5847 ASSERT(arc_buf_is_shared(lastbuf) ||
5848 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
5849 ASSERT(!ARC_BUF_SHARED(buf));
5851 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5852 ASSERT3P(state, !=, arc_l2c_only);
5854 (void) refcount_remove_many(&state->arcs_size,
5855 arc_buf_size(buf), buf);
5857 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5858 ASSERT3P(state, !=, arc_l2c_only);
5859 (void) refcount_remove_many(&state->arcs_esize[type],
5860 arc_buf_size(buf), buf);
5863 hdr->b_l1hdr.b_bufcnt -= 1;
5864 arc_cksum_verify(buf);
5866 arc_buf_unwatch(buf);
5869 mutex_exit(hash_lock);
5872 * Allocate a new hdr. The new hdr will contain a b_pabd
5873 * buffer which will be freed in arc_write().
5875 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5876 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5877 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5878 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5879 VERIFY3U(nhdr->b_type, ==, type);
5880 ASSERT(!HDR_SHARED_DATA(nhdr));
5882 nhdr->b_l1hdr.b_buf = buf;
5883 nhdr->b_l1hdr.b_bufcnt = 1;
5884 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5887 mutex_exit(&buf->b_evict_lock);
5888 (void) refcount_add_many(&arc_anon->arcs_size,
5889 arc_buf_size(buf), buf);
5891 mutex_exit(&buf->b_evict_lock);
5892 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5893 /* protected by hash lock, or hdr is on arc_anon */
5894 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5895 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5896 arc_change_state(arc_anon, hdr, hash_lock);
5897 hdr->b_l1hdr.b_arc_access = 0;
5898 mutex_exit(hash_lock);
5900 buf_discard_identity(hdr);
5906 arc_released(arc_buf_t *buf)
5910 mutex_enter(&buf->b_evict_lock);
5911 released = (buf->b_data != NULL &&
5912 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5913 mutex_exit(&buf->b_evict_lock);
5919 arc_referenced(arc_buf_t *buf)
5923 mutex_enter(&buf->b_evict_lock);
5924 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5925 mutex_exit(&buf->b_evict_lock);
5926 return (referenced);
5931 arc_write_ready(zio_t *zio)
5933 arc_write_callback_t *callback = zio->io_private;
5934 arc_buf_t *buf = callback->awcb_buf;
5935 arc_buf_hdr_t *hdr = buf->b_hdr;
5936 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5938 ASSERT(HDR_HAS_L1HDR(hdr));
5939 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5940 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5943 * If we're reexecuting this zio because the pool suspended, then
5944 * cleanup any state that was previously set the first time the
5945 * callback was invoked.
5947 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5948 arc_cksum_free(hdr);
5950 arc_buf_unwatch(buf);
5952 if (hdr->b_l1hdr.b_pabd != NULL) {
5953 if (arc_buf_is_shared(buf)) {
5954 arc_unshare_buf(hdr, buf);
5956 arc_hdr_free_pabd(hdr);
5960 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5961 ASSERT(!HDR_SHARED_DATA(hdr));
5962 ASSERT(!arc_buf_is_shared(buf));
5964 callback->awcb_ready(zio, buf, callback->awcb_private);
5966 if (HDR_IO_IN_PROGRESS(hdr))
5967 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
5969 arc_cksum_compute(buf);
5970 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5972 enum zio_compress compress;
5973 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5974 compress = ZIO_COMPRESS_OFF;
5976 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
5977 compress = BP_GET_COMPRESS(zio->io_bp);
5979 HDR_SET_PSIZE(hdr, psize);
5980 arc_hdr_set_compress(hdr, compress);
5984 * Fill the hdr with data. If the hdr is compressed, the data we want
5985 * is available from the zio, otherwise we can take it from the buf.
5987 * We might be able to share the buf's data with the hdr here. However,
5988 * doing so would cause the ARC to be full of linear ABDs if we write a
5989 * lot of shareable data. As a compromise, we check whether scattered
5990 * ABDs are allowed, and assume that if they are then the user wants
5991 * the ARC to be primarily filled with them regardless of the data being
5992 * written. Therefore, if they're allowed then we allocate one and copy
5993 * the data into it; otherwise, we share the data directly if we can.
5995 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
5996 arc_hdr_alloc_pabd(hdr);
5999 * Ideally, we would always copy the io_abd into b_pabd, but the
6000 * user may have disabled compressed ARC, thus we must check the
6001 * hdr's compression setting rather than the io_bp's.
6003 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
6004 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
6006 ASSERT3U(psize, >, 0);
6008 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6010 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6012 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6016 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6017 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6018 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6020 arc_share_buf(hdr, buf);
6023 arc_hdr_verify(hdr, zio->io_bp);
6027 arc_write_children_ready(zio_t *zio)
6029 arc_write_callback_t *callback = zio->io_private;
6030 arc_buf_t *buf = callback->awcb_buf;
6032 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6036 * The SPA calls this callback for each physical write that happens on behalf
6037 * of a logical write. See the comment in dbuf_write_physdone() for details.
6040 arc_write_physdone(zio_t *zio)
6042 arc_write_callback_t *cb = zio->io_private;
6043 if (cb->awcb_physdone != NULL)
6044 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6048 arc_write_done(zio_t *zio)
6050 arc_write_callback_t *callback = zio->io_private;
6051 arc_buf_t *buf = callback->awcb_buf;
6052 arc_buf_hdr_t *hdr = buf->b_hdr;
6054 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6056 if (zio->io_error == 0) {
6057 arc_hdr_verify(hdr, zio->io_bp);
6059 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6060 buf_discard_identity(hdr);
6062 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6063 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6066 ASSERT(HDR_EMPTY(hdr));
6070 * If the block to be written was all-zero or compressed enough to be
6071 * embedded in the BP, no write was performed so there will be no
6072 * dva/birth/checksum. The buffer must therefore remain anonymous
6075 if (!HDR_EMPTY(hdr)) {
6076 arc_buf_hdr_t *exists;
6077 kmutex_t *hash_lock;
6079 ASSERT3U(zio->io_error, ==, 0);
6081 arc_cksum_verify(buf);
6083 exists = buf_hash_insert(hdr, &hash_lock);
6084 if (exists != NULL) {
6086 * This can only happen if we overwrite for
6087 * sync-to-convergence, because we remove
6088 * buffers from the hash table when we arc_free().
6090 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6091 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6092 panic("bad overwrite, hdr=%p exists=%p",
6093 (void *)hdr, (void *)exists);
6094 ASSERT(refcount_is_zero(
6095 &exists->b_l1hdr.b_refcnt));
6096 arc_change_state(arc_anon, exists, hash_lock);
6097 mutex_exit(hash_lock);
6098 arc_hdr_destroy(exists);
6099 exists = buf_hash_insert(hdr, &hash_lock);
6100 ASSERT3P(exists, ==, NULL);
6101 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6103 ASSERT(zio->io_prop.zp_nopwrite);
6104 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6105 panic("bad nopwrite, hdr=%p exists=%p",
6106 (void *)hdr, (void *)exists);
6109 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6110 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6111 ASSERT(BP_GET_DEDUP(zio->io_bp));
6112 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6115 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6116 /* if it's not anon, we are doing a scrub */
6117 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6118 arc_access(hdr, hash_lock);
6119 mutex_exit(hash_lock);
6121 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6124 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6125 callback->awcb_done(zio, buf, callback->awcb_private);
6127 abd_put(zio->io_abd);
6128 kmem_free(callback, sizeof (arc_write_callback_t));
6132 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6133 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
6134 arc_done_func_t *children_ready, arc_done_func_t *physdone,
6135 arc_done_func_t *done, void *private, zio_priority_t priority,
6136 int zio_flags, const zbookmark_phys_t *zb)
6138 arc_buf_hdr_t *hdr = buf->b_hdr;
6139 arc_write_callback_t *callback;
6141 zio_prop_t localprop = *zp;
6143 ASSERT3P(ready, !=, NULL);
6144 ASSERT3P(done, !=, NULL);
6145 ASSERT(!HDR_IO_ERROR(hdr));
6146 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6147 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6148 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6150 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6151 if (ARC_BUF_COMPRESSED(buf)) {
6153 * We're writing a pre-compressed buffer. Make the
6154 * compression algorithm requested by the zio_prop_t match
6155 * the pre-compressed buffer's compression algorithm.
6157 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6159 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6160 zio_flags |= ZIO_FLAG_RAW;
6162 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6163 callback->awcb_ready = ready;
6164 callback->awcb_children_ready = children_ready;
6165 callback->awcb_physdone = physdone;
6166 callback->awcb_done = done;
6167 callback->awcb_private = private;
6168 callback->awcb_buf = buf;
6171 * The hdr's b_pabd is now stale, free it now. A new data block
6172 * will be allocated when the zio pipeline calls arc_write_ready().
6174 if (hdr->b_l1hdr.b_pabd != NULL) {
6176 * If the buf is currently sharing the data block with
6177 * the hdr then we need to break that relationship here.
6178 * The hdr will remain with a NULL data pointer and the
6179 * buf will take sole ownership of the block.
6181 if (arc_buf_is_shared(buf)) {
6182 arc_unshare_buf(hdr, buf);
6184 arc_hdr_free_pabd(hdr);
6186 VERIFY3P(buf->b_data, !=, NULL);
6187 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6189 ASSERT(!arc_buf_is_shared(buf));
6190 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6192 zio = zio_write(pio, spa, txg, bp,
6193 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6194 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6195 (children_ready != NULL) ? arc_write_children_ready : NULL,
6196 arc_write_physdone, arc_write_done, callback,
6197 priority, zio_flags, zb);
6203 arc_memory_throttle(uint64_t reserve, uint64_t txg)
6206 uint64_t available_memory = ptob(freemem);
6207 static uint64_t page_load = 0;
6208 static uint64_t last_txg = 0;
6210 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
6212 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE)));
6215 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
6218 if (txg > last_txg) {
6223 * If we are in pageout, we know that memory is already tight,
6224 * the arc is already going to be evicting, so we just want to
6225 * continue to let page writes occur as quickly as possible.
6227 if (curproc == pageproc) {
6228 if (page_load > MAX(ptob(minfree), available_memory) / 4)
6229 return (SET_ERROR(ERESTART));
6230 /* Note: reserve is inflated, so we deflate */
6231 page_load += reserve / 8;
6233 } else if (page_load > 0 && arc_reclaim_needed()) {
6234 /* memory is low, delay before restarting */
6235 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6236 return (SET_ERROR(EAGAIN));
6244 arc_tempreserve_clear(uint64_t reserve)
6246 atomic_add_64(&arc_tempreserve, -reserve);
6247 ASSERT((int64_t)arc_tempreserve >= 0);
6251 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
6256 if (reserve > arc_c/4 && !arc_no_grow) {
6257 arc_c = MIN(arc_c_max, reserve * 4);
6258 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
6260 if (reserve > arc_c)
6261 return (SET_ERROR(ENOMEM));
6264 * Don't count loaned bufs as in flight dirty data to prevent long
6265 * network delays from blocking transactions that are ready to be
6266 * assigned to a txg.
6269 /* assert that it has not wrapped around */
6270 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6272 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6273 arc_loaned_bytes), 0);
6276 * Writes will, almost always, require additional memory allocations
6277 * in order to compress/encrypt/etc the data. We therefore need to
6278 * make sure that there is sufficient available memory for this.
6280 error = arc_memory_throttle(reserve, txg);
6285 * Throttle writes when the amount of dirty data in the cache
6286 * gets too large. We try to keep the cache less than half full
6287 * of dirty blocks so that our sync times don't grow too large.
6288 * Note: if two requests come in concurrently, we might let them
6289 * both succeed, when one of them should fail. Not a huge deal.
6292 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
6293 anon_size > arc_c / 4) {
6294 uint64_t meta_esize =
6295 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6296 uint64_t data_esize =
6297 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6298 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6299 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6300 arc_tempreserve >> 10, meta_esize >> 10,
6301 data_esize >> 10, reserve >> 10, arc_c >> 10);
6302 return (SET_ERROR(ERESTART));
6304 atomic_add_64(&arc_tempreserve, reserve);
6309 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6310 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6312 size->value.ui64 = refcount_count(&state->arcs_size);
6313 evict_data->value.ui64 =
6314 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6315 evict_metadata->value.ui64 =
6316 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6320 arc_kstat_update(kstat_t *ksp, int rw)
6322 arc_stats_t *as = ksp->ks_data;
6324 if (rw == KSTAT_WRITE) {
6327 arc_kstat_update_state(arc_anon,
6328 &as->arcstat_anon_size,
6329 &as->arcstat_anon_evictable_data,
6330 &as->arcstat_anon_evictable_metadata);
6331 arc_kstat_update_state(arc_mru,
6332 &as->arcstat_mru_size,
6333 &as->arcstat_mru_evictable_data,
6334 &as->arcstat_mru_evictable_metadata);
6335 arc_kstat_update_state(arc_mru_ghost,
6336 &as->arcstat_mru_ghost_size,
6337 &as->arcstat_mru_ghost_evictable_data,
6338 &as->arcstat_mru_ghost_evictable_metadata);
6339 arc_kstat_update_state(arc_mfu,
6340 &as->arcstat_mfu_size,
6341 &as->arcstat_mfu_evictable_data,
6342 &as->arcstat_mfu_evictable_metadata);
6343 arc_kstat_update_state(arc_mfu_ghost,
6344 &as->arcstat_mfu_ghost_size,
6345 &as->arcstat_mfu_ghost_evictable_data,
6346 &as->arcstat_mfu_ghost_evictable_metadata);
6348 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
6349 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
6350 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
6351 ARCSTAT(arcstat_metadata_size) =
6352 aggsum_value(&astat_metadata_size);
6353 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
6354 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size);
6355 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
6362 * This function *must* return indices evenly distributed between all
6363 * sublists of the multilist. This is needed due to how the ARC eviction
6364 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6365 * distributed between all sublists and uses this assumption when
6366 * deciding which sublist to evict from and how much to evict from it.
6369 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6371 arc_buf_hdr_t *hdr = obj;
6374 * We rely on b_dva to generate evenly distributed index
6375 * numbers using buf_hash below. So, as an added precaution,
6376 * let's make sure we never add empty buffers to the arc lists.
6378 ASSERT(!HDR_EMPTY(hdr));
6381 * The assumption here, is the hash value for a given
6382 * arc_buf_hdr_t will remain constant throughout it's lifetime
6383 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6384 * Thus, we don't need to store the header's sublist index
6385 * on insertion, as this index can be recalculated on removal.
6387 * Also, the low order bits of the hash value are thought to be
6388 * distributed evenly. Otherwise, in the case that the multilist
6389 * has a power of two number of sublists, each sublists' usage
6390 * would not be evenly distributed.
6392 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6393 multilist_get_num_sublists(ml));
6397 static eventhandler_tag arc_event_lowmem = NULL;
6400 arc_lowmem(void *arg __unused, int howto __unused)
6403 mutex_enter(&arc_reclaim_lock);
6404 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE);
6405 cv_signal(&arc_reclaim_thread_cv);
6408 * It is unsafe to block here in arbitrary threads, because we can come
6409 * here from ARC itself and may hold ARC locks and thus risk a deadlock
6410 * with ARC reclaim thread.
6412 if (curproc == pageproc)
6413 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
6414 mutex_exit(&arc_reclaim_lock);
6419 arc_state_init(void)
6421 arc_anon = &ARC_anon;
6423 arc_mru_ghost = &ARC_mru_ghost;
6425 arc_mfu_ghost = &ARC_mfu_ghost;
6426 arc_l2c_only = &ARC_l2c_only;
6428 arc_mru->arcs_list[ARC_BUFC_METADATA] =
6429 multilist_create(sizeof (arc_buf_hdr_t),
6430 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6431 arc_state_multilist_index_func);
6432 arc_mru->arcs_list[ARC_BUFC_DATA] =
6433 multilist_create(sizeof (arc_buf_hdr_t),
6434 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6435 arc_state_multilist_index_func);
6436 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6437 multilist_create(sizeof (arc_buf_hdr_t),
6438 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6439 arc_state_multilist_index_func);
6440 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6441 multilist_create(sizeof (arc_buf_hdr_t),
6442 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6443 arc_state_multilist_index_func);
6444 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6445 multilist_create(sizeof (arc_buf_hdr_t),
6446 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6447 arc_state_multilist_index_func);
6448 arc_mfu->arcs_list[ARC_BUFC_DATA] =
6449 multilist_create(sizeof (arc_buf_hdr_t),
6450 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6451 arc_state_multilist_index_func);
6452 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6453 multilist_create(sizeof (arc_buf_hdr_t),
6454 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6455 arc_state_multilist_index_func);
6456 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6457 multilist_create(sizeof (arc_buf_hdr_t),
6458 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6459 arc_state_multilist_index_func);
6460 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6461 multilist_create(sizeof (arc_buf_hdr_t),
6462 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6463 arc_state_multilist_index_func);
6464 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6465 multilist_create(sizeof (arc_buf_hdr_t),
6466 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6467 arc_state_multilist_index_func);
6469 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6470 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6471 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6472 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6473 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6474 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6475 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6476 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6477 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6478 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6479 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6480 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6482 refcount_create(&arc_anon->arcs_size);
6483 refcount_create(&arc_mru->arcs_size);
6484 refcount_create(&arc_mru_ghost->arcs_size);
6485 refcount_create(&arc_mfu->arcs_size);
6486 refcount_create(&arc_mfu_ghost->arcs_size);
6487 refcount_create(&arc_l2c_only->arcs_size);
6489 aggsum_init(&arc_meta_used, 0);
6490 aggsum_init(&arc_size, 0);
6491 aggsum_init(&astat_data_size, 0);
6492 aggsum_init(&astat_metadata_size, 0);
6493 aggsum_init(&astat_hdr_size, 0);
6494 aggsum_init(&astat_other_size, 0);
6495 aggsum_init(&astat_l2_hdr_size, 0);
6499 arc_state_fini(void)
6501 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6502 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6503 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6504 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6505 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6506 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6507 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6508 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6509 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6510 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6511 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6512 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6514 refcount_destroy(&arc_anon->arcs_size);
6515 refcount_destroy(&arc_mru->arcs_size);
6516 refcount_destroy(&arc_mru_ghost->arcs_size);
6517 refcount_destroy(&arc_mfu->arcs_size);
6518 refcount_destroy(&arc_mfu_ghost->arcs_size);
6519 refcount_destroy(&arc_l2c_only->arcs_size);
6521 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6522 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6523 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6524 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6525 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6526 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6527 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6528 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6540 int i, prefetch_tunable_set = 0;
6543 * allmem is "all memory that we could possibly use".
6547 uint64_t allmem = ptob(physmem - swapfs_minfree);
6549 uint64_t allmem = (physmem * PAGESIZE) / 2;
6552 uint64_t allmem = kmem_size();
6556 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
6557 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
6558 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
6560 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
6561 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
6563 /* Convert seconds to clock ticks */
6564 arc_min_prefetch_lifespan = 1 * hz;
6566 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
6567 arc_c_min = MAX(allmem / 32, arc_abs_min);
6568 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */
6569 if (allmem >= 1 << 30)
6570 arc_c_max = allmem - (1 << 30);
6572 arc_c_max = arc_c_min;
6573 arc_c_max = MAX(allmem * 5 / 8, arc_c_max);
6576 * In userland, there's only the memory pressure that we artificially
6577 * create (see arc_available_memory()). Don't let arc_c get too
6578 * small, because it can cause transactions to be larger than
6579 * arc_c, causing arc_tempreserve_space() to fail.
6582 arc_c_min = arc_c_max / 2;
6587 * Allow the tunables to override our calculations if they are
6590 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) {
6591 arc_c_max = zfs_arc_max;
6592 arc_c_min = MIN(arc_c_min, arc_c_max);
6594 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
6595 arc_c_min = zfs_arc_min;
6599 arc_p = (arc_c >> 1);
6601 /* limit meta-data to 1/4 of the arc capacity */
6602 arc_meta_limit = arc_c_max / 4;
6606 * Metadata is stored in the kernel's heap. Don't let us
6607 * use more than half the heap for the ARC.
6609 arc_meta_limit = MIN(arc_meta_limit,
6610 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6613 /* Allow the tunable to override if it is reasonable */
6614 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6615 arc_meta_limit = zfs_arc_meta_limit;
6617 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6618 arc_c_min = arc_meta_limit / 2;
6620 if (zfs_arc_meta_min > 0) {
6621 arc_meta_min = zfs_arc_meta_min;
6623 arc_meta_min = arc_c_min / 2;
6626 if (zfs_arc_grow_retry > 0)
6627 arc_grow_retry = zfs_arc_grow_retry;
6629 if (zfs_arc_shrink_shift > 0)
6630 arc_shrink_shift = zfs_arc_shrink_shift;
6632 if (zfs_arc_no_grow_shift > 0)
6633 arc_no_grow_shift = zfs_arc_no_grow_shift;
6635 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6637 if (arc_no_grow_shift >= arc_shrink_shift)
6638 arc_no_grow_shift = arc_shrink_shift - 1;
6640 if (zfs_arc_p_min_shift > 0)
6641 arc_p_min_shift = zfs_arc_p_min_shift;
6643 /* if kmem_flags are set, lets try to use less memory */
6644 if (kmem_debugging())
6646 if (arc_c < arc_c_min)
6649 zfs_arc_min = arc_c_min;
6650 zfs_arc_max = arc_c_max;
6655 arc_reclaim_thread_exit = B_FALSE;
6656 arc_dnlc_evicts_thread_exit = FALSE;
6658 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6659 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6661 if (arc_ksp != NULL) {
6662 arc_ksp->ks_data = &arc_stats;
6663 arc_ksp->ks_update = arc_kstat_update;
6664 kstat_install(arc_ksp);
6667 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6668 TS_RUN, minclsyspri);
6671 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
6672 EVENTHANDLER_PRI_FIRST);
6675 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
6676 TS_RUN, minclsyspri);
6682 * Calculate maximum amount of dirty data per pool.
6684 * If it has been set by /etc/system, take that.
6685 * Otherwise, use a percentage of physical memory defined by
6686 * zfs_dirty_data_max_percent (default 10%) with a cap at
6687 * zfs_dirty_data_max_max (default 4GB).
6689 if (zfs_dirty_data_max == 0) {
6690 zfs_dirty_data_max = ptob(physmem) *
6691 zfs_dirty_data_max_percent / 100;
6692 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6693 zfs_dirty_data_max_max);
6697 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
6698 prefetch_tunable_set = 1;
6701 if (prefetch_tunable_set == 0) {
6702 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
6704 printf(" add \"vfs.zfs.prefetch_disable=0\" "
6705 "to /boot/loader.conf.\n");
6706 zfs_prefetch_disable = 1;
6709 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
6710 prefetch_tunable_set == 0) {
6711 printf("ZFS NOTICE: Prefetch is disabled by default if less "
6712 "than 4GB of RAM is present;\n"
6713 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
6714 "to /boot/loader.conf.\n");
6715 zfs_prefetch_disable = 1;
6718 /* Warn about ZFS memory and address space requirements. */
6719 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
6720 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
6721 "expect unstable behavior.\n");
6723 if (allmem < 512 * (1 << 20)) {
6724 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
6725 "expect unstable behavior.\n");
6726 printf(" Consider tuning vm.kmem_size and "
6727 "vm.kmem_size_max\n");
6728 printf(" in /boot/loader.conf.\n");
6737 if (arc_event_lowmem != NULL)
6738 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
6741 mutex_enter(&arc_reclaim_lock);
6742 arc_reclaim_thread_exit = B_TRUE;
6744 * The reclaim thread will set arc_reclaim_thread_exit back to
6745 * B_FALSE when it is finished exiting; we're waiting for that.
6747 while (arc_reclaim_thread_exit) {
6748 cv_signal(&arc_reclaim_thread_cv);
6749 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6751 mutex_exit(&arc_reclaim_lock);
6753 /* Use B_TRUE to ensure *all* buffers are evicted */
6754 arc_flush(NULL, B_TRUE);
6756 mutex_enter(&arc_dnlc_evicts_lock);
6757 arc_dnlc_evicts_thread_exit = TRUE;
6759 * The user evicts thread will set arc_user_evicts_thread_exit
6760 * to FALSE when it is finished exiting; we're waiting for that.
6762 while (arc_dnlc_evicts_thread_exit) {
6763 cv_signal(&arc_dnlc_evicts_cv);
6764 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
6766 mutex_exit(&arc_dnlc_evicts_lock);
6770 if (arc_ksp != NULL) {
6771 kstat_delete(arc_ksp);
6775 mutex_destroy(&arc_reclaim_lock);
6776 cv_destroy(&arc_reclaim_thread_cv);
6777 cv_destroy(&arc_reclaim_waiters_cv);
6779 mutex_destroy(&arc_dnlc_evicts_lock);
6780 cv_destroy(&arc_dnlc_evicts_cv);
6785 ASSERT0(arc_loaned_bytes);
6791 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6792 * It uses dedicated storage devices to hold cached data, which are populated
6793 * using large infrequent writes. The main role of this cache is to boost
6794 * the performance of random read workloads. The intended L2ARC devices
6795 * include short-stroked disks, solid state disks, and other media with
6796 * substantially faster read latency than disk.
6798 * +-----------------------+
6800 * +-----------------------+
6803 * l2arc_feed_thread() arc_read()
6807 * +---------------+ |
6809 * +---------------+ |
6814 * +-------+ +-------+
6816 * | cache | | cache |
6817 * +-------+ +-------+
6818 * +=========+ .-----.
6819 * : L2ARC : |-_____-|
6820 * : devices : | Disks |
6821 * +=========+ `-_____-'
6823 * Read requests are satisfied from the following sources, in order:
6826 * 2) vdev cache of L2ARC devices
6828 * 4) vdev cache of disks
6831 * Some L2ARC device types exhibit extremely slow write performance.
6832 * To accommodate for this there are some significant differences between
6833 * the L2ARC and traditional cache design:
6835 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6836 * the ARC behave as usual, freeing buffers and placing headers on ghost
6837 * lists. The ARC does not send buffers to the L2ARC during eviction as
6838 * this would add inflated write latencies for all ARC memory pressure.
6840 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6841 * It does this by periodically scanning buffers from the eviction-end of
6842 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6843 * not already there. It scans until a headroom of buffers is satisfied,
6844 * which itself is a buffer for ARC eviction. If a compressible buffer is
6845 * found during scanning and selected for writing to an L2ARC device, we
6846 * temporarily boost scanning headroom during the next scan cycle to make
6847 * sure we adapt to compression effects (which might significantly reduce
6848 * the data volume we write to L2ARC). The thread that does this is
6849 * l2arc_feed_thread(), illustrated below; example sizes are included to
6850 * provide a better sense of ratio than this diagram:
6853 * +---------------------+----------+
6854 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6855 * +---------------------+----------+ | o L2ARC eligible
6856 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6857 * +---------------------+----------+ |
6858 * 15.9 Gbytes ^ 32 Mbytes |
6860 * l2arc_feed_thread()
6862 * l2arc write hand <--[oooo]--'
6866 * +==============================+
6867 * L2ARC dev |####|#|###|###| |####| ... |
6868 * +==============================+
6871 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6872 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6873 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6874 * safe to say that this is an uncommon case, since buffers at the end of
6875 * the ARC lists have moved there due to inactivity.
6877 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6878 * then the L2ARC simply misses copying some buffers. This serves as a
6879 * pressure valve to prevent heavy read workloads from both stalling the ARC
6880 * with waits and clogging the L2ARC with writes. This also helps prevent
6881 * the potential for the L2ARC to churn if it attempts to cache content too
6882 * quickly, such as during backups of the entire pool.
6884 * 5. After system boot and before the ARC has filled main memory, there are
6885 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6886 * lists can remain mostly static. Instead of searching from tail of these
6887 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6888 * for eligible buffers, greatly increasing its chance of finding them.
6890 * The L2ARC device write speed is also boosted during this time so that
6891 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6892 * there are no L2ARC reads, and no fear of degrading read performance
6893 * through increased writes.
6895 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6896 * the vdev queue can aggregate them into larger and fewer writes. Each
6897 * device is written to in a rotor fashion, sweeping writes through
6898 * available space then repeating.
6900 * 7. The L2ARC does not store dirty content. It never needs to flush
6901 * write buffers back to disk based storage.
6903 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6904 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6906 * The performance of the L2ARC can be tweaked by a number of tunables, which
6907 * may be necessary for different workloads:
6909 * l2arc_write_max max write bytes per interval
6910 * l2arc_write_boost extra write bytes during device warmup
6911 * l2arc_noprefetch skip caching prefetched buffers
6912 * l2arc_headroom number of max device writes to precache
6913 * l2arc_headroom_boost when we find compressed buffers during ARC
6914 * scanning, we multiply headroom by this
6915 * percentage factor for the next scan cycle,
6916 * since more compressed buffers are likely to
6918 * l2arc_feed_secs seconds between L2ARC writing
6920 * Tunables may be removed or added as future performance improvements are
6921 * integrated, and also may become zpool properties.
6923 * There are three key functions that control how the L2ARC warms up:
6925 * l2arc_write_eligible() check if a buffer is eligible to cache
6926 * l2arc_write_size() calculate how much to write
6927 * l2arc_write_interval() calculate sleep delay between writes
6929 * These three functions determine what to write, how much, and how quickly
6934 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6937 * A buffer is *not* eligible for the L2ARC if it:
6938 * 1. belongs to a different spa.
6939 * 2. is already cached on the L2ARC.
6940 * 3. has an I/O in progress (it may be an incomplete read).
6941 * 4. is flagged not eligible (zfs property).
6943 if (hdr->b_spa != spa_guid) {
6944 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
6947 if (HDR_HAS_L2HDR(hdr)) {
6948 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
6951 if (HDR_IO_IN_PROGRESS(hdr)) {
6952 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
6955 if (!HDR_L2CACHE(hdr)) {
6956 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
6964 l2arc_write_size(void)
6969 * Make sure our globals have meaningful values in case the user
6972 size = l2arc_write_max;
6974 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
6975 "be greater than zero, resetting it to the default (%d)",
6977 size = l2arc_write_max = L2ARC_WRITE_SIZE;
6980 if (arc_warm == B_FALSE)
6981 size += l2arc_write_boost;
6988 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
6990 clock_t interval, next, now;
6993 * If the ARC lists are busy, increase our write rate; if the
6994 * lists are stale, idle back. This is achieved by checking
6995 * how much we previously wrote - if it was more than half of
6996 * what we wanted, schedule the next write much sooner.
6998 if (l2arc_feed_again && wrote > (wanted / 2))
6999 interval = (hz * l2arc_feed_min_ms) / 1000;
7001 interval = hz * l2arc_feed_secs;
7003 now = ddi_get_lbolt();
7004 next = MAX(now, MIN(now + interval, began + interval));
7010 * Cycle through L2ARC devices. This is how L2ARC load balances.
7011 * If a device is returned, this also returns holding the spa config lock.
7013 static l2arc_dev_t *
7014 l2arc_dev_get_next(void)
7016 l2arc_dev_t *first, *next = NULL;
7019 * Lock out the removal of spas (spa_namespace_lock), then removal
7020 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7021 * both locks will be dropped and a spa config lock held instead.
7023 mutex_enter(&spa_namespace_lock);
7024 mutex_enter(&l2arc_dev_mtx);
7026 /* if there are no vdevs, there is nothing to do */
7027 if (l2arc_ndev == 0)
7031 next = l2arc_dev_last;
7033 /* loop around the list looking for a non-faulted vdev */
7035 next = list_head(l2arc_dev_list);
7037 next = list_next(l2arc_dev_list, next);
7039 next = list_head(l2arc_dev_list);
7042 /* if we have come back to the start, bail out */
7045 else if (next == first)
7048 } while (vdev_is_dead(next->l2ad_vdev));
7050 /* if we were unable to find any usable vdevs, return NULL */
7051 if (vdev_is_dead(next->l2ad_vdev))
7054 l2arc_dev_last = next;
7057 mutex_exit(&l2arc_dev_mtx);
7060 * Grab the config lock to prevent the 'next' device from being
7061 * removed while we are writing to it.
7064 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7065 mutex_exit(&spa_namespace_lock);
7071 * Free buffers that were tagged for destruction.
7074 l2arc_do_free_on_write()
7077 l2arc_data_free_t *df, *df_prev;
7079 mutex_enter(&l2arc_free_on_write_mtx);
7080 buflist = l2arc_free_on_write;
7082 for (df = list_tail(buflist); df; df = df_prev) {
7083 df_prev = list_prev(buflist, df);
7084 ASSERT3P(df->l2df_abd, !=, NULL);
7085 abd_free(df->l2df_abd);
7086 list_remove(buflist, df);
7087 kmem_free(df, sizeof (l2arc_data_free_t));
7090 mutex_exit(&l2arc_free_on_write_mtx);
7094 * A write to a cache device has completed. Update all headers to allow
7095 * reads from these buffers to begin.
7098 l2arc_write_done(zio_t *zio)
7100 l2arc_write_callback_t *cb;
7103 arc_buf_hdr_t *head, *hdr, *hdr_prev;
7104 kmutex_t *hash_lock;
7105 int64_t bytes_dropped = 0;
7107 cb = zio->io_private;
7108 ASSERT3P(cb, !=, NULL);
7109 dev = cb->l2wcb_dev;
7110 ASSERT3P(dev, !=, NULL);
7111 head = cb->l2wcb_head;
7112 ASSERT3P(head, !=, NULL);
7113 buflist = &dev->l2ad_buflist;
7114 ASSERT3P(buflist, !=, NULL);
7115 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7116 l2arc_write_callback_t *, cb);
7118 if (zio->io_error != 0)
7119 ARCSTAT_BUMP(arcstat_l2_writes_error);
7122 * All writes completed, or an error was hit.
7125 mutex_enter(&dev->l2ad_mtx);
7126 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7127 hdr_prev = list_prev(buflist, hdr);
7129 hash_lock = HDR_LOCK(hdr);
7132 * We cannot use mutex_enter or else we can deadlock
7133 * with l2arc_write_buffers (due to swapping the order
7134 * the hash lock and l2ad_mtx are taken).
7136 if (!mutex_tryenter(hash_lock)) {
7138 * Missed the hash lock. We must retry so we
7139 * don't leave the ARC_FLAG_L2_WRITING bit set.
7141 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7144 * We don't want to rescan the headers we've
7145 * already marked as having been written out, so
7146 * we reinsert the head node so we can pick up
7147 * where we left off.
7149 list_remove(buflist, head);
7150 list_insert_after(buflist, hdr, head);
7152 mutex_exit(&dev->l2ad_mtx);
7155 * We wait for the hash lock to become available
7156 * to try and prevent busy waiting, and increase
7157 * the chance we'll be able to acquire the lock
7158 * the next time around.
7160 mutex_enter(hash_lock);
7161 mutex_exit(hash_lock);
7166 * We could not have been moved into the arc_l2c_only
7167 * state while in-flight due to our ARC_FLAG_L2_WRITING
7168 * bit being set. Let's just ensure that's being enforced.
7170 ASSERT(HDR_HAS_L1HDR(hdr));
7172 if (zio->io_error != 0) {
7174 * Error - drop L2ARC entry.
7176 list_remove(buflist, hdr);
7178 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7180 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
7181 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
7183 bytes_dropped += arc_hdr_size(hdr);
7184 (void) refcount_remove_many(&dev->l2ad_alloc,
7185 arc_hdr_size(hdr), hdr);
7189 * Allow ARC to begin reads and ghost list evictions to
7192 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7194 mutex_exit(hash_lock);
7197 atomic_inc_64(&l2arc_writes_done);
7198 list_remove(buflist, head);
7199 ASSERT(!HDR_HAS_L1HDR(head));
7200 kmem_cache_free(hdr_l2only_cache, head);
7201 mutex_exit(&dev->l2ad_mtx);
7203 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7205 l2arc_do_free_on_write();
7207 kmem_free(cb, sizeof (l2arc_write_callback_t));
7211 * A read to a cache device completed. Validate buffer contents before
7212 * handing over to the regular ARC routines.
7215 l2arc_read_done(zio_t *zio)
7217 l2arc_read_callback_t *cb;
7219 kmutex_t *hash_lock;
7220 boolean_t valid_cksum;
7222 ASSERT3P(zio->io_vd, !=, NULL);
7223 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7225 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7227 cb = zio->io_private;
7228 ASSERT3P(cb, !=, NULL);
7229 hdr = cb->l2rcb_hdr;
7230 ASSERT3P(hdr, !=, NULL);
7232 hash_lock = HDR_LOCK(hdr);
7233 mutex_enter(hash_lock);
7234 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7237 * If the data was read into a temporary buffer,
7238 * move it and free the buffer.
7240 if (cb->l2rcb_abd != NULL) {
7241 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7242 if (zio->io_error == 0) {
7243 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
7248 * The following must be done regardless of whether
7249 * there was an error:
7250 * - free the temporary buffer
7251 * - point zio to the real ARC buffer
7252 * - set zio size accordingly
7253 * These are required because zio is either re-used for
7254 * an I/O of the block in the case of the error
7255 * or the zio is passed to arc_read_done() and it
7258 abd_free(cb->l2rcb_abd);
7259 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7260 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
7263 ASSERT3P(zio->io_abd, !=, NULL);
7266 * Check this survived the L2ARC journey.
7268 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
7269 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
7270 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
7272 valid_cksum = arc_cksum_is_equal(hdr, zio);
7273 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
7274 mutex_exit(hash_lock);
7275 zio->io_private = hdr;
7278 mutex_exit(hash_lock);
7280 * Buffer didn't survive caching. Increment stats and
7281 * reissue to the original storage device.
7283 if (zio->io_error != 0) {
7284 ARCSTAT_BUMP(arcstat_l2_io_error);
7286 zio->io_error = SET_ERROR(EIO);
7289 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
7292 * If there's no waiter, issue an async i/o to the primary
7293 * storage now. If there *is* a waiter, the caller must
7294 * issue the i/o in a context where it's OK to block.
7296 if (zio->io_waiter == NULL) {
7297 zio_t *pio = zio_unique_parent(zio);
7299 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7301 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
7302 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
7303 hdr, zio->io_priority, cb->l2rcb_flags,
7308 kmem_free(cb, sizeof (l2arc_read_callback_t));
7312 * This is the list priority from which the L2ARC will search for pages to
7313 * cache. This is used within loops (0..3) to cycle through lists in the
7314 * desired order. This order can have a significant effect on cache
7317 * Currently the metadata lists are hit first, MFU then MRU, followed by
7318 * the data lists. This function returns a locked list, and also returns
7321 static multilist_sublist_t *
7322 l2arc_sublist_lock(int list_num)
7324 multilist_t *ml = NULL;
7327 ASSERT(list_num >= 0 && list_num <= 3);
7331 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7334 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7337 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7340 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7345 * Return a randomly-selected sublist. This is acceptable
7346 * because the caller feeds only a little bit of data for each
7347 * call (8MB). Subsequent calls will result in different
7348 * sublists being selected.
7350 idx = multilist_get_random_index(ml);
7351 return (multilist_sublist_lock(ml, idx));
7355 * Evict buffers from the device write hand to the distance specified in
7356 * bytes. This distance may span populated buffers, it may span nothing.
7357 * This is clearing a region on the L2ARC device ready for writing.
7358 * If the 'all' boolean is set, every buffer is evicted.
7361 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7364 arc_buf_hdr_t *hdr, *hdr_prev;
7365 kmutex_t *hash_lock;
7368 buflist = &dev->l2ad_buflist;
7370 if (!all && dev->l2ad_first) {
7372 * This is the first sweep through the device. There is
7378 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
7380 * When nearing the end of the device, evict to the end
7381 * before the device write hand jumps to the start.
7383 taddr = dev->l2ad_end;
7385 taddr = dev->l2ad_hand + distance;
7387 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
7388 uint64_t, taddr, boolean_t, all);
7391 mutex_enter(&dev->l2ad_mtx);
7392 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
7393 hdr_prev = list_prev(buflist, hdr);
7395 hash_lock = HDR_LOCK(hdr);
7398 * We cannot use mutex_enter or else we can deadlock
7399 * with l2arc_write_buffers (due to swapping the order
7400 * the hash lock and l2ad_mtx are taken).
7402 if (!mutex_tryenter(hash_lock)) {
7404 * Missed the hash lock. Retry.
7406 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
7407 mutex_exit(&dev->l2ad_mtx);
7408 mutex_enter(hash_lock);
7409 mutex_exit(hash_lock);
7414 * A header can't be on this list if it doesn't have L2 header.
7416 ASSERT(HDR_HAS_L2HDR(hdr));
7418 /* Ensure this header has finished being written. */
7419 ASSERT(!HDR_L2_WRITING(hdr));
7420 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
7422 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
7423 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
7425 * We've evicted to the target address,
7426 * or the end of the device.
7428 mutex_exit(hash_lock);
7432 if (!HDR_HAS_L1HDR(hdr)) {
7433 ASSERT(!HDR_L2_READING(hdr));
7435 * This doesn't exist in the ARC. Destroy.
7436 * arc_hdr_destroy() will call list_remove()
7437 * and decrement arcstat_l2_lsize.
7439 arc_change_state(arc_anon, hdr, hash_lock);
7440 arc_hdr_destroy(hdr);
7442 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
7443 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
7445 * Invalidate issued or about to be issued
7446 * reads, since we may be about to write
7447 * over this location.
7449 if (HDR_L2_READING(hdr)) {
7450 ARCSTAT_BUMP(arcstat_l2_evict_reading);
7451 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
7454 arc_hdr_l2hdr_destroy(hdr);
7456 mutex_exit(hash_lock);
7458 mutex_exit(&dev->l2ad_mtx);
7462 * Find and write ARC buffers to the L2ARC device.
7464 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7465 * for reading until they have completed writing.
7466 * The headroom_boost is an in-out parameter used to maintain headroom boost
7467 * state between calls to this function.
7469 * Returns the number of bytes actually written (which may be smaller than
7470 * the delta by which the device hand has changed due to alignment).
7473 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
7475 arc_buf_hdr_t *hdr, *hdr_prev, *head;
7476 uint64_t write_asize, write_psize, write_lsize, headroom;
7478 l2arc_write_callback_t *cb;
7480 uint64_t guid = spa_load_guid(spa);
7483 ASSERT3P(dev->l2ad_vdev, !=, NULL);
7486 write_lsize = write_asize = write_psize = 0;
7488 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
7489 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
7491 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
7493 * Copy buffers for L2ARC writing.
7495 for (try = 0; try <= 3; try++) {
7496 multilist_sublist_t *mls = l2arc_sublist_lock(try);
7497 uint64_t passed_sz = 0;
7499 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
7502 * L2ARC fast warmup.
7504 * Until the ARC is warm and starts to evict, read from the
7505 * head of the ARC lists rather than the tail.
7507 if (arc_warm == B_FALSE)
7508 hdr = multilist_sublist_head(mls);
7510 hdr = multilist_sublist_tail(mls);
7512 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
7514 headroom = target_sz * l2arc_headroom;
7515 if (zfs_compressed_arc_enabled)
7516 headroom = (headroom * l2arc_headroom_boost) / 100;
7518 for (; hdr; hdr = hdr_prev) {
7519 kmutex_t *hash_lock;
7521 if (arc_warm == B_FALSE)
7522 hdr_prev = multilist_sublist_next(mls, hdr);
7524 hdr_prev = multilist_sublist_prev(mls, hdr);
7525 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
7526 HDR_GET_LSIZE(hdr));
7528 hash_lock = HDR_LOCK(hdr);
7529 if (!mutex_tryenter(hash_lock)) {
7530 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
7532 * Skip this buffer rather than waiting.
7537 passed_sz += HDR_GET_LSIZE(hdr);
7538 if (passed_sz > headroom) {
7542 mutex_exit(hash_lock);
7543 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
7547 if (!l2arc_write_eligible(guid, hdr)) {
7548 mutex_exit(hash_lock);
7553 * We rely on the L1 portion of the header below, so
7554 * it's invalid for this header to have been evicted out
7555 * of the ghost cache, prior to being written out. The
7556 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7558 ASSERT(HDR_HAS_L1HDR(hdr));
7560 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
7561 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7562 ASSERT3U(arc_hdr_size(hdr), >, 0);
7563 uint64_t psize = arc_hdr_size(hdr);
7564 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
7567 if ((write_asize + asize) > target_sz) {
7569 mutex_exit(hash_lock);
7570 ARCSTAT_BUMP(arcstat_l2_write_full);
7576 * Insert a dummy header on the buflist so
7577 * l2arc_write_done() can find where the
7578 * write buffers begin without searching.
7580 mutex_enter(&dev->l2ad_mtx);
7581 list_insert_head(&dev->l2ad_buflist, head);
7582 mutex_exit(&dev->l2ad_mtx);
7585 sizeof (l2arc_write_callback_t), KM_SLEEP);
7586 cb->l2wcb_dev = dev;
7587 cb->l2wcb_head = head;
7588 pio = zio_root(spa, l2arc_write_done, cb,
7590 ARCSTAT_BUMP(arcstat_l2_write_pios);
7593 hdr->b_l2hdr.b_dev = dev;
7594 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7595 arc_hdr_set_flags(hdr,
7596 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7598 mutex_enter(&dev->l2ad_mtx);
7599 list_insert_head(&dev->l2ad_buflist, hdr);
7600 mutex_exit(&dev->l2ad_mtx);
7602 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
7605 * Normally the L2ARC can use the hdr's data, but if
7606 * we're sharing data between the hdr and one of its
7607 * bufs, L2ARC needs its own copy of the data so that
7608 * the ZIO below can't race with the buf consumer.
7609 * Another case where we need to create a copy of the
7610 * data is when the buffer size is not device-aligned
7611 * and we need to pad the block to make it such.
7612 * That also keeps the clock hand suitably aligned.
7614 * To ensure that the copy will be available for the
7615 * lifetime of the ZIO and be cleaned up afterwards, we
7616 * add it to the l2arc_free_on_write queue.
7619 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
7620 to_write = hdr->b_l1hdr.b_pabd;
7622 to_write = abd_alloc_for_io(asize,
7623 HDR_ISTYPE_METADATA(hdr));
7624 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
7625 if (asize != psize) {
7626 abd_zero_off(to_write, psize,
7629 l2arc_free_abd_on_write(to_write, asize,
7632 wzio = zio_write_phys(pio, dev->l2ad_vdev,
7633 hdr->b_l2hdr.b_daddr, asize, to_write,
7634 ZIO_CHECKSUM_OFF, NULL, hdr,
7635 ZIO_PRIORITY_ASYNC_WRITE,
7636 ZIO_FLAG_CANFAIL, B_FALSE);
7638 write_lsize += HDR_GET_LSIZE(hdr);
7639 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7642 write_psize += psize;
7643 write_asize += asize;
7644 dev->l2ad_hand += asize;
7646 mutex_exit(hash_lock);
7648 (void) zio_nowait(wzio);
7651 multilist_sublist_unlock(mls);
7657 /* No buffers selected for writing? */
7659 ASSERT0(write_lsize);
7660 ASSERT(!HDR_HAS_L1HDR(head));
7661 kmem_cache_free(hdr_l2only_cache, head);
7665 ASSERT3U(write_psize, <=, target_sz);
7666 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7667 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
7668 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
7669 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
7670 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
7673 * Bump device hand to the device start if it is approaching the end.
7674 * l2arc_evict() will already have evicted ahead for this case.
7676 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7677 dev->l2ad_hand = dev->l2ad_start;
7678 dev->l2ad_first = B_FALSE;
7681 dev->l2ad_writing = B_TRUE;
7682 (void) zio_wait(pio);
7683 dev->l2ad_writing = B_FALSE;
7685 return (write_asize);
7689 * This thread feeds the L2ARC at regular intervals. This is the beating
7690 * heart of the L2ARC.
7694 l2arc_feed_thread(void *unused __unused)
7699 uint64_t size, wrote;
7700 clock_t begin, next = ddi_get_lbolt();
7702 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7704 mutex_enter(&l2arc_feed_thr_lock);
7706 while (l2arc_thread_exit == 0) {
7707 CALLB_CPR_SAFE_BEGIN(&cpr);
7708 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7709 next - ddi_get_lbolt());
7710 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7711 next = ddi_get_lbolt() + hz;
7714 * Quick check for L2ARC devices.
7716 mutex_enter(&l2arc_dev_mtx);
7717 if (l2arc_ndev == 0) {
7718 mutex_exit(&l2arc_dev_mtx);
7721 mutex_exit(&l2arc_dev_mtx);
7722 begin = ddi_get_lbolt();
7725 * This selects the next l2arc device to write to, and in
7726 * doing so the next spa to feed from: dev->l2ad_spa. This
7727 * will return NULL if there are now no l2arc devices or if
7728 * they are all faulted.
7730 * If a device is returned, its spa's config lock is also
7731 * held to prevent device removal. l2arc_dev_get_next()
7732 * will grab and release l2arc_dev_mtx.
7734 if ((dev = l2arc_dev_get_next()) == NULL)
7737 spa = dev->l2ad_spa;
7738 ASSERT3P(spa, !=, NULL);
7741 * If the pool is read-only then force the feed thread to
7742 * sleep a little longer.
7744 if (!spa_writeable(spa)) {
7745 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7746 spa_config_exit(spa, SCL_L2ARC, dev);
7751 * Avoid contributing to memory pressure.
7753 if (arc_reclaim_needed()) {
7754 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7755 spa_config_exit(spa, SCL_L2ARC, dev);
7759 ARCSTAT_BUMP(arcstat_l2_feeds);
7761 size = l2arc_write_size();
7764 * Evict L2ARC buffers that will be overwritten.
7766 l2arc_evict(dev, size, B_FALSE);
7769 * Write ARC buffers.
7771 wrote = l2arc_write_buffers(spa, dev, size);
7774 * Calculate interval between writes.
7776 next = l2arc_write_interval(begin, size, wrote);
7777 spa_config_exit(spa, SCL_L2ARC, dev);
7780 l2arc_thread_exit = 0;
7781 cv_broadcast(&l2arc_feed_thr_cv);
7782 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7787 l2arc_vdev_present(vdev_t *vd)
7791 mutex_enter(&l2arc_dev_mtx);
7792 for (dev = list_head(l2arc_dev_list); dev != NULL;
7793 dev = list_next(l2arc_dev_list, dev)) {
7794 if (dev->l2ad_vdev == vd)
7797 mutex_exit(&l2arc_dev_mtx);
7799 return (dev != NULL);
7803 * Add a vdev for use by the L2ARC. By this point the spa has already
7804 * validated the vdev and opened it.
7807 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7809 l2arc_dev_t *adddev;
7811 ASSERT(!l2arc_vdev_present(vd));
7813 vdev_ashift_optimize(vd);
7816 * Create a new l2arc device entry.
7818 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7819 adddev->l2ad_spa = spa;
7820 adddev->l2ad_vdev = vd;
7821 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7822 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7823 adddev->l2ad_hand = adddev->l2ad_start;
7824 adddev->l2ad_first = B_TRUE;
7825 adddev->l2ad_writing = B_FALSE;
7827 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7829 * This is a list of all ARC buffers that are still valid on the
7832 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7833 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7835 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7836 refcount_create(&adddev->l2ad_alloc);
7839 * Add device to global list
7841 mutex_enter(&l2arc_dev_mtx);
7842 list_insert_head(l2arc_dev_list, adddev);
7843 atomic_inc_64(&l2arc_ndev);
7844 mutex_exit(&l2arc_dev_mtx);
7848 * Remove a vdev from the L2ARC.
7851 l2arc_remove_vdev(vdev_t *vd)
7853 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7856 * Find the device by vdev
7858 mutex_enter(&l2arc_dev_mtx);
7859 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7860 nextdev = list_next(l2arc_dev_list, dev);
7861 if (vd == dev->l2ad_vdev) {
7866 ASSERT3P(remdev, !=, NULL);
7869 * Remove device from global list
7871 list_remove(l2arc_dev_list, remdev);
7872 l2arc_dev_last = NULL; /* may have been invalidated */
7873 atomic_dec_64(&l2arc_ndev);
7874 mutex_exit(&l2arc_dev_mtx);
7877 * Clear all buflists and ARC references. L2ARC device flush.
7879 l2arc_evict(remdev, 0, B_TRUE);
7880 list_destroy(&remdev->l2ad_buflist);
7881 mutex_destroy(&remdev->l2ad_mtx);
7882 refcount_destroy(&remdev->l2ad_alloc);
7883 kmem_free(remdev, sizeof (l2arc_dev_t));
7889 l2arc_thread_exit = 0;
7891 l2arc_writes_sent = 0;
7892 l2arc_writes_done = 0;
7894 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7895 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7896 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7897 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7899 l2arc_dev_list = &L2ARC_dev_list;
7900 l2arc_free_on_write = &L2ARC_free_on_write;
7901 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7902 offsetof(l2arc_dev_t, l2ad_node));
7903 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7904 offsetof(l2arc_data_free_t, l2df_list_node));
7911 * This is called from dmu_fini(), which is called from spa_fini();
7912 * Because of this, we can assume that all l2arc devices have
7913 * already been removed when the pools themselves were removed.
7916 l2arc_do_free_on_write();
7918 mutex_destroy(&l2arc_feed_thr_lock);
7919 cv_destroy(&l2arc_feed_thr_cv);
7920 mutex_destroy(&l2arc_dev_mtx);
7921 mutex_destroy(&l2arc_free_on_write_mtx);
7923 list_destroy(l2arc_dev_list);
7924 list_destroy(l2arc_free_on_write);
7930 if (!(spa_mode_global & FWRITE))
7933 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7934 TS_RUN, minclsyspri);
7940 if (!(spa_mode_global & FWRITE))
7943 mutex_enter(&l2arc_feed_thr_lock);
7944 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7945 l2arc_thread_exit = 1;
7946 while (l2arc_thread_exit != 0)
7947 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7948 mutex_exit(&l2arc_feed_thr_lock);