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_cnt.v_free_target;
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 when updating the access state due to the
544 * header having already been released after acquiring the hash lock.
546 kstat_named_t arcstat_access_skip;
548 * Number of buffers skipped because they have I/O in progress, are
549 * indirect prefetch buffers that have not lived long enough, or are
550 * not from the spa we're trying to evict from.
552 kstat_named_t arcstat_evict_skip;
554 * Number of times arc_evict_state() was unable to evict enough
555 * buffers to reach it's target amount.
557 kstat_named_t arcstat_evict_not_enough;
558 kstat_named_t arcstat_evict_l2_cached;
559 kstat_named_t arcstat_evict_l2_eligible;
560 kstat_named_t arcstat_evict_l2_ineligible;
561 kstat_named_t arcstat_evict_l2_skip;
562 kstat_named_t arcstat_hash_elements;
563 kstat_named_t arcstat_hash_elements_max;
564 kstat_named_t arcstat_hash_collisions;
565 kstat_named_t arcstat_hash_chains;
566 kstat_named_t arcstat_hash_chain_max;
567 kstat_named_t arcstat_p;
568 kstat_named_t arcstat_c;
569 kstat_named_t arcstat_c_min;
570 kstat_named_t arcstat_c_max;
571 /* Not updated directly; only synced in arc_kstat_update. */
572 kstat_named_t arcstat_size;
574 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
575 * Note that the compressed bytes may match the uncompressed bytes
576 * if the block is either not compressed or compressed arc is disabled.
578 kstat_named_t arcstat_compressed_size;
580 * Uncompressed size of the data stored in b_pabd. If compressed
581 * arc is disabled then this value will be identical to the stat
584 kstat_named_t arcstat_uncompressed_size;
586 * Number of bytes stored in all the arc_buf_t's. This is classified
587 * as "overhead" since this data is typically short-lived and will
588 * be evicted from the arc when it becomes unreferenced unless the
589 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
590 * values have been set (see comment in dbuf.c for more information).
592 kstat_named_t arcstat_overhead_size;
594 * Number of bytes consumed by internal ARC structures necessary
595 * for tracking purposes; these structures are not actually
596 * backed by ARC buffers. This includes arc_buf_hdr_t structures
597 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
598 * caches), and arc_buf_t structures (allocated via arc_buf_t
600 * Not updated directly; only synced in arc_kstat_update.
602 kstat_named_t arcstat_hdr_size;
604 * Number of bytes consumed by ARC buffers of type equal to
605 * ARC_BUFC_DATA. This is generally consumed by buffers backing
606 * on disk user data (e.g. plain file contents).
607 * Not updated directly; only synced in arc_kstat_update.
609 kstat_named_t arcstat_data_size;
611 * Number of bytes consumed by ARC buffers of type equal to
612 * ARC_BUFC_METADATA. This is generally consumed by buffers
613 * backing on disk data that is used for internal ZFS
614 * structures (e.g. ZAP, dnode, indirect blocks, etc).
615 * Not updated directly; only synced in arc_kstat_update.
617 kstat_named_t arcstat_metadata_size;
619 * Number of bytes consumed by various buffers and structures
620 * not actually backed with ARC buffers. This includes bonus
621 * buffers (allocated directly via zio_buf_* functions),
622 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
623 * cache), and dnode_t structures (allocated via dnode_t cache).
624 * Not updated directly; only synced in arc_kstat_update.
626 kstat_named_t arcstat_other_size;
628 * Total number of bytes consumed by ARC buffers residing in the
629 * arc_anon state. This includes *all* buffers in the arc_anon
630 * state; e.g. data, metadata, evictable, and unevictable buffers
631 * are all included in this value.
632 * Not updated directly; only synced in arc_kstat_update.
634 kstat_named_t arcstat_anon_size;
636 * Number of bytes consumed by ARC buffers that meet the
637 * following criteria: backing buffers of type ARC_BUFC_DATA,
638 * residing in the arc_anon state, and are eligible for eviction
639 * (e.g. have no outstanding holds on the buffer).
640 * Not updated directly; only synced in arc_kstat_update.
642 kstat_named_t arcstat_anon_evictable_data;
644 * Number of bytes consumed by ARC buffers that meet the
645 * following criteria: backing buffers of type ARC_BUFC_METADATA,
646 * residing in the arc_anon state, and are eligible for eviction
647 * (e.g. have no outstanding holds on the buffer).
648 * Not updated directly; only synced in arc_kstat_update.
650 kstat_named_t arcstat_anon_evictable_metadata;
652 * Total number of bytes consumed by ARC buffers residing in the
653 * arc_mru state. This includes *all* buffers in the arc_mru
654 * state; e.g. data, metadata, evictable, and unevictable buffers
655 * are all included in this value.
656 * Not updated directly; only synced in arc_kstat_update.
658 kstat_named_t arcstat_mru_size;
660 * Number of bytes consumed by ARC buffers that meet the
661 * following criteria: backing buffers of type ARC_BUFC_DATA,
662 * residing in the arc_mru state, and are eligible for eviction
663 * (e.g. have no outstanding holds on the buffer).
664 * Not updated directly; only synced in arc_kstat_update.
666 kstat_named_t arcstat_mru_evictable_data;
668 * Number of bytes consumed by ARC buffers that meet the
669 * following criteria: backing buffers of type ARC_BUFC_METADATA,
670 * residing in the arc_mru state, and are eligible for eviction
671 * (e.g. have no outstanding holds on the buffer).
672 * Not updated directly; only synced in arc_kstat_update.
674 kstat_named_t arcstat_mru_evictable_metadata;
676 * Total number of bytes that *would have been* consumed by ARC
677 * buffers in the arc_mru_ghost state. The key thing to note
678 * here, is the fact that this size doesn't actually indicate
679 * RAM consumption. The ghost lists only consist of headers and
680 * don't actually have ARC buffers linked off of these headers.
681 * Thus, *if* the headers had associated ARC buffers, these
682 * buffers *would have* consumed this number of bytes.
683 * Not updated directly; only synced in arc_kstat_update.
685 kstat_named_t arcstat_mru_ghost_size;
687 * Number of bytes that *would have been* consumed by ARC
688 * buffers that are eligible for eviction, of type
689 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
690 * Not updated directly; only synced in arc_kstat_update.
692 kstat_named_t arcstat_mru_ghost_evictable_data;
694 * Number of bytes that *would have been* consumed by ARC
695 * buffers that are eligible for eviction, of type
696 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
697 * Not updated directly; only synced in arc_kstat_update.
699 kstat_named_t arcstat_mru_ghost_evictable_metadata;
701 * Total number of bytes consumed by ARC buffers residing in the
702 * arc_mfu state. This includes *all* buffers in the arc_mfu
703 * state; e.g. data, metadata, evictable, and unevictable buffers
704 * are all included in this value.
705 * Not updated directly; only synced in arc_kstat_update.
707 kstat_named_t arcstat_mfu_size;
709 * Number of bytes consumed by ARC buffers that are eligible for
710 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
712 * Not updated directly; only synced in arc_kstat_update.
714 kstat_named_t arcstat_mfu_evictable_data;
716 * Number of bytes consumed by ARC buffers that are eligible for
717 * eviction, of type ARC_BUFC_METADATA, and reside in the
719 * Not updated directly; only synced in arc_kstat_update.
721 kstat_named_t arcstat_mfu_evictable_metadata;
723 * Total number of bytes that *would have been* consumed by ARC
724 * buffers in the arc_mfu_ghost state. See the comment above
725 * arcstat_mru_ghost_size for more details.
726 * Not updated directly; only synced in arc_kstat_update.
728 kstat_named_t arcstat_mfu_ghost_size;
730 * Number of bytes that *would have been* consumed by ARC
731 * buffers that are eligible for eviction, of type
732 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
733 * Not updated directly; only synced in arc_kstat_update.
735 kstat_named_t arcstat_mfu_ghost_evictable_data;
737 * Number of bytes that *would have been* consumed by ARC
738 * buffers that are eligible for eviction, of type
739 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
740 * Not updated directly; only synced in arc_kstat_update.
742 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
743 kstat_named_t arcstat_l2_hits;
744 kstat_named_t arcstat_l2_misses;
745 kstat_named_t arcstat_l2_feeds;
746 kstat_named_t arcstat_l2_rw_clash;
747 kstat_named_t arcstat_l2_read_bytes;
748 kstat_named_t arcstat_l2_write_bytes;
749 kstat_named_t arcstat_l2_writes_sent;
750 kstat_named_t arcstat_l2_writes_done;
751 kstat_named_t arcstat_l2_writes_error;
752 kstat_named_t arcstat_l2_writes_lock_retry;
753 kstat_named_t arcstat_l2_evict_lock_retry;
754 kstat_named_t arcstat_l2_evict_reading;
755 kstat_named_t arcstat_l2_evict_l1cached;
756 kstat_named_t arcstat_l2_free_on_write;
757 kstat_named_t arcstat_l2_abort_lowmem;
758 kstat_named_t arcstat_l2_cksum_bad;
759 kstat_named_t arcstat_l2_io_error;
760 kstat_named_t arcstat_l2_lsize;
761 kstat_named_t arcstat_l2_psize;
762 /* Not updated directly; only synced in arc_kstat_update. */
763 kstat_named_t arcstat_l2_hdr_size;
764 kstat_named_t arcstat_l2_write_trylock_fail;
765 kstat_named_t arcstat_l2_write_passed_headroom;
766 kstat_named_t arcstat_l2_write_spa_mismatch;
767 kstat_named_t arcstat_l2_write_in_l2;
768 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
769 kstat_named_t arcstat_l2_write_not_cacheable;
770 kstat_named_t arcstat_l2_write_full;
771 kstat_named_t arcstat_l2_write_buffer_iter;
772 kstat_named_t arcstat_l2_write_pios;
773 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
774 kstat_named_t arcstat_l2_write_buffer_list_iter;
775 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
776 kstat_named_t arcstat_memory_throttle_count;
777 /* Not updated directly; only synced in arc_kstat_update. */
778 kstat_named_t arcstat_meta_used;
779 kstat_named_t arcstat_meta_limit;
780 kstat_named_t arcstat_meta_max;
781 kstat_named_t arcstat_meta_min;
782 kstat_named_t arcstat_sync_wait_for_async;
783 kstat_named_t arcstat_demand_hit_predictive_prefetch;
786 static arc_stats_t arc_stats = {
787 { "hits", KSTAT_DATA_UINT64 },
788 { "misses", KSTAT_DATA_UINT64 },
789 { "demand_data_hits", KSTAT_DATA_UINT64 },
790 { "demand_data_misses", KSTAT_DATA_UINT64 },
791 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
792 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
793 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
794 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
795 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
796 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
797 { "mru_hits", KSTAT_DATA_UINT64 },
798 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
799 { "mfu_hits", KSTAT_DATA_UINT64 },
800 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
801 { "allocated", KSTAT_DATA_UINT64 },
802 { "deleted", KSTAT_DATA_UINT64 },
803 { "mutex_miss", KSTAT_DATA_UINT64 },
804 { "access_skip", KSTAT_DATA_UINT64 },
805 { "evict_skip", KSTAT_DATA_UINT64 },
806 { "evict_not_enough", KSTAT_DATA_UINT64 },
807 { "evict_l2_cached", KSTAT_DATA_UINT64 },
808 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
809 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
810 { "evict_l2_skip", KSTAT_DATA_UINT64 },
811 { "hash_elements", KSTAT_DATA_UINT64 },
812 { "hash_elements_max", KSTAT_DATA_UINT64 },
813 { "hash_collisions", KSTAT_DATA_UINT64 },
814 { "hash_chains", KSTAT_DATA_UINT64 },
815 { "hash_chain_max", KSTAT_DATA_UINT64 },
816 { "p", KSTAT_DATA_UINT64 },
817 { "c", KSTAT_DATA_UINT64 },
818 { "c_min", KSTAT_DATA_UINT64 },
819 { "c_max", KSTAT_DATA_UINT64 },
820 { "size", KSTAT_DATA_UINT64 },
821 { "compressed_size", KSTAT_DATA_UINT64 },
822 { "uncompressed_size", KSTAT_DATA_UINT64 },
823 { "overhead_size", KSTAT_DATA_UINT64 },
824 { "hdr_size", KSTAT_DATA_UINT64 },
825 { "data_size", KSTAT_DATA_UINT64 },
826 { "metadata_size", KSTAT_DATA_UINT64 },
827 { "other_size", KSTAT_DATA_UINT64 },
828 { "anon_size", KSTAT_DATA_UINT64 },
829 { "anon_evictable_data", KSTAT_DATA_UINT64 },
830 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
831 { "mru_size", KSTAT_DATA_UINT64 },
832 { "mru_evictable_data", KSTAT_DATA_UINT64 },
833 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
834 { "mru_ghost_size", KSTAT_DATA_UINT64 },
835 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
836 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
837 { "mfu_size", KSTAT_DATA_UINT64 },
838 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
839 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
840 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
841 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
842 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
843 { "l2_hits", KSTAT_DATA_UINT64 },
844 { "l2_misses", KSTAT_DATA_UINT64 },
845 { "l2_feeds", KSTAT_DATA_UINT64 },
846 { "l2_rw_clash", KSTAT_DATA_UINT64 },
847 { "l2_read_bytes", KSTAT_DATA_UINT64 },
848 { "l2_write_bytes", KSTAT_DATA_UINT64 },
849 { "l2_writes_sent", KSTAT_DATA_UINT64 },
850 { "l2_writes_done", KSTAT_DATA_UINT64 },
851 { "l2_writes_error", KSTAT_DATA_UINT64 },
852 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
853 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
854 { "l2_evict_reading", KSTAT_DATA_UINT64 },
855 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
856 { "l2_free_on_write", KSTAT_DATA_UINT64 },
857 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
858 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
859 { "l2_io_error", KSTAT_DATA_UINT64 },
860 { "l2_size", KSTAT_DATA_UINT64 },
861 { "l2_asize", KSTAT_DATA_UINT64 },
862 { "l2_hdr_size", KSTAT_DATA_UINT64 },
863 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
864 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
865 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
866 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
867 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
868 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
869 { "l2_write_full", KSTAT_DATA_UINT64 },
870 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
871 { "l2_write_pios", KSTAT_DATA_UINT64 },
872 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
873 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
874 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
875 { "memory_throttle_count", KSTAT_DATA_UINT64 },
876 { "arc_meta_used", KSTAT_DATA_UINT64 },
877 { "arc_meta_limit", KSTAT_DATA_UINT64 },
878 { "arc_meta_max", KSTAT_DATA_UINT64 },
879 { "arc_meta_min", KSTAT_DATA_UINT64 },
880 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
881 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
884 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
886 #define ARCSTAT_INCR(stat, val) \
887 atomic_add_64(&arc_stats.stat.value.ui64, (val))
889 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
890 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
892 #define ARCSTAT_MAX(stat, val) { \
894 while ((val) > (m = arc_stats.stat.value.ui64) && \
895 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
899 #define ARCSTAT_MAXSTAT(stat) \
900 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
903 * We define a macro to allow ARC hits/misses to be easily broken down by
904 * two separate conditions, giving a total of four different subtypes for
905 * each of hits and misses (so eight statistics total).
907 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
910 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
912 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
916 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
918 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
923 static arc_state_t *arc_anon;
924 static arc_state_t *arc_mru;
925 static arc_state_t *arc_mru_ghost;
926 static arc_state_t *arc_mfu;
927 static arc_state_t *arc_mfu_ghost;
928 static arc_state_t *arc_l2c_only;
931 * There are several ARC variables that are critical to export as kstats --
932 * but we don't want to have to grovel around in the kstat whenever we wish to
933 * manipulate them. For these variables, we therefore define them to be in
934 * terms of the statistic variable. This assures that we are not introducing
935 * the possibility of inconsistency by having shadow copies of the variables,
936 * while still allowing the code to be readable.
938 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
939 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
940 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
941 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
942 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
943 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
944 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
946 /* compressed size of entire arc */
947 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
948 /* uncompressed size of entire arc */
949 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
950 /* number of bytes in the arc from arc_buf_t's */
951 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
954 * There are also some ARC variables that we want to export, but that are
955 * updated so often that having the canonical representation be the statistic
956 * variable causes a performance bottleneck. We want to use aggsum_t's for these
957 * instead, but still be able to export the kstat in the same way as before.
958 * The solution is to always use the aggsum version, except in the kstat update
962 aggsum_t arc_meta_used;
963 aggsum_t astat_data_size;
964 aggsum_t astat_metadata_size;
965 aggsum_t astat_hdr_size;
966 aggsum_t astat_other_size;
967 aggsum_t astat_l2_hdr_size;
969 static int arc_no_grow; /* Don't try to grow cache size */
970 static uint64_t arc_tempreserve;
971 static uint64_t arc_loaned_bytes;
973 typedef struct arc_callback arc_callback_t;
975 struct arc_callback {
977 arc_done_func_t *acb_done;
979 boolean_t acb_compressed;
980 zio_t *acb_zio_dummy;
981 arc_callback_t *acb_next;
984 typedef struct arc_write_callback arc_write_callback_t;
986 struct arc_write_callback {
988 arc_done_func_t *awcb_ready;
989 arc_done_func_t *awcb_children_ready;
990 arc_done_func_t *awcb_physdone;
991 arc_done_func_t *awcb_done;
996 * ARC buffers are separated into multiple structs as a memory saving measure:
997 * - Common fields struct, always defined, and embedded within it:
998 * - L2-only fields, always allocated but undefined when not in L2ARC
999 * - L1-only fields, only allocated when in L1ARC
1001 * Buffer in L1 Buffer only in L2
1002 * +------------------------+ +------------------------+
1003 * | arc_buf_hdr_t | | arc_buf_hdr_t |
1007 * +------------------------+ +------------------------+
1008 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
1009 * | (undefined if L1-only) | | |
1010 * +------------------------+ +------------------------+
1011 * | l1arc_buf_hdr_t |
1016 * +------------------------+
1018 * Because it's possible for the L2ARC to become extremely large, we can wind
1019 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
1020 * is minimized by only allocating the fields necessary for an L1-cached buffer
1021 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
1022 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
1023 * words in pointers. arc_hdr_realloc() is used to switch a header between
1024 * these two allocation states.
1026 typedef struct l1arc_buf_hdr {
1027 kmutex_t b_freeze_lock;
1028 zio_cksum_t *b_freeze_cksum;
1031 * Used for debugging with kmem_flags - by allocating and freeing
1032 * b_thawed when the buffer is thawed, we get a record of the stack
1033 * trace that thawed it.
1040 /* for waiting on writes to complete */
1044 /* protected by arc state mutex */
1045 arc_state_t *b_state;
1046 multilist_node_t b_arc_node;
1048 /* updated atomically */
1049 clock_t b_arc_access;
1051 /* self protecting */
1052 refcount_t b_refcnt;
1054 arc_callback_t *b_acb;
1058 typedef struct l2arc_dev l2arc_dev_t;
1060 typedef struct l2arc_buf_hdr {
1061 /* protected by arc_buf_hdr mutex */
1062 l2arc_dev_t *b_dev; /* L2ARC device */
1063 uint64_t b_daddr; /* disk address, offset byte */
1065 list_node_t b_l2node;
1068 struct arc_buf_hdr {
1069 /* protected by hash lock */
1073 arc_buf_contents_t b_type;
1074 arc_buf_hdr_t *b_hash_next;
1075 arc_flags_t b_flags;
1078 * This field stores the size of the data buffer after
1079 * compression, and is set in the arc's zio completion handlers.
1080 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1082 * While the block pointers can store up to 32MB in their psize
1083 * field, we can only store up to 32MB minus 512B. This is due
1084 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1085 * a field of zeros represents 512B in the bp). We can't use a
1086 * bias of 1 since we need to reserve a psize of zero, here, to
1087 * represent holes and embedded blocks.
1089 * This isn't a problem in practice, since the maximum size of a
1090 * buffer is limited to 16MB, so we never need to store 32MB in
1091 * this field. Even in the upstream illumos code base, the
1092 * maximum size of a buffer is limited to 16MB.
1097 * This field stores the size of the data buffer before
1098 * compression, and cannot change once set. It is in units
1099 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1101 uint16_t b_lsize; /* immutable */
1102 uint64_t b_spa; /* immutable */
1104 /* L2ARC fields. Undefined when not in L2ARC. */
1105 l2arc_buf_hdr_t b_l2hdr;
1106 /* L1ARC fields. Undefined when in l2arc_only state */
1107 l1arc_buf_hdr_t b_l1hdr;
1110 #if defined(__FreeBSD__) && defined(_KERNEL)
1112 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1117 val = arc_meta_limit;
1118 err = sysctl_handle_64(oidp, &val, 0, req);
1119 if (err != 0 || req->newptr == NULL)
1122 if (val <= 0 || val > arc_c_max)
1125 arc_meta_limit = val;
1130 sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS)
1135 val = arc_no_grow_shift;
1136 err = sysctl_handle_32(oidp, &val, 0, req);
1137 if (err != 0 || req->newptr == NULL)
1140 if (val >= arc_shrink_shift)
1143 arc_no_grow_shift = val;
1148 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1154 err = sysctl_handle_64(oidp, &val, 0, req);
1155 if (err != 0 || req->newptr == NULL)
1158 if (zfs_arc_max == 0) {
1159 /* Loader tunable so blindly set */
1164 if (val < arc_abs_min || val > kmem_size())
1166 if (val < arc_c_min)
1168 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1174 arc_p = (arc_c >> 1);
1176 if (zfs_arc_meta_limit == 0) {
1177 /* limit meta-data to 1/4 of the arc capacity */
1178 arc_meta_limit = arc_c_max / 4;
1181 /* if kmem_flags are set, lets try to use less memory */
1182 if (kmem_debugging())
1185 zfs_arc_max = arc_c;
1191 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1197 err = sysctl_handle_64(oidp, &val, 0, req);
1198 if (err != 0 || req->newptr == NULL)
1201 if (zfs_arc_min == 0) {
1202 /* Loader tunable so blindly set */
1207 if (val < arc_abs_min || val > arc_c_max)
1212 if (zfs_arc_meta_min == 0)
1213 arc_meta_min = arc_c_min / 2;
1215 if (arc_c < arc_c_min)
1218 zfs_arc_min = arc_c_min;
1224 #define GHOST_STATE(state) \
1225 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1226 (state) == arc_l2c_only)
1228 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1229 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1230 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1231 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1232 #define HDR_COMPRESSION_ENABLED(hdr) \
1233 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1235 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1236 #define HDR_L2_READING(hdr) \
1237 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1238 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1239 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1240 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1241 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1242 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1244 #define HDR_ISTYPE_METADATA(hdr) \
1245 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1246 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1248 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1249 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1251 /* For storing compression mode in b_flags */
1252 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1254 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1255 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1256 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1257 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1259 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1260 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1261 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1267 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1268 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1271 * Hash table routines
1274 #define HT_LOCK_PAD CACHE_LINE_SIZE
1279 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1283 #define BUF_LOCKS 256
1284 typedef struct buf_hash_table {
1286 arc_buf_hdr_t **ht_table;
1287 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1290 static buf_hash_table_t buf_hash_table;
1292 #define BUF_HASH_INDEX(spa, dva, birth) \
1293 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1294 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1295 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1296 #define HDR_LOCK(hdr) \
1297 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1299 uint64_t zfs_crc64_table[256];
1305 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1306 #define L2ARC_HEADROOM 2 /* num of writes */
1308 * If we discover during ARC scan any buffers to be compressed, we boost
1309 * our headroom for the next scanning cycle by this percentage multiple.
1311 #define L2ARC_HEADROOM_BOOST 200
1312 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1313 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1315 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1316 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1318 /* L2ARC Performance Tunables */
1319 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1320 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1321 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1322 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1323 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1324 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1325 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1326 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1327 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1329 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1330 &l2arc_write_max, 0, "max write size");
1331 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1332 &l2arc_write_boost, 0, "extra write during warmup");
1333 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1334 &l2arc_headroom, 0, "number of dev writes");
1335 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1336 &l2arc_feed_secs, 0, "interval seconds");
1337 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1338 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1340 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1341 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1342 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1343 &l2arc_feed_again, 0, "turbo warmup");
1344 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1345 &l2arc_norw, 0, "no reads during writes");
1347 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1348 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1349 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1350 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1351 "size of anonymous state");
1352 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1353 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1354 "size of anonymous state");
1356 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1357 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1358 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1359 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1360 "size of metadata in mru state");
1361 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1362 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1363 "size of data in mru state");
1365 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1366 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1367 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1368 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1369 "size of metadata in mru ghost state");
1370 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1371 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1372 "size of data in mru ghost state");
1374 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1375 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1376 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1377 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1378 "size of metadata in mfu state");
1379 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1380 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1381 "size of data in mfu state");
1383 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1384 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1385 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1386 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1387 "size of metadata in mfu ghost state");
1388 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1389 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1390 "size of data in mfu ghost state");
1392 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1393 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1399 vdev_t *l2ad_vdev; /* vdev */
1400 spa_t *l2ad_spa; /* spa */
1401 uint64_t l2ad_hand; /* next write location */
1402 uint64_t l2ad_start; /* first addr on device */
1403 uint64_t l2ad_end; /* last addr on device */
1404 boolean_t l2ad_first; /* first sweep through */
1405 boolean_t l2ad_writing; /* currently writing */
1406 kmutex_t l2ad_mtx; /* lock for buffer list */
1407 list_t l2ad_buflist; /* buffer list */
1408 list_node_t l2ad_node; /* device list node */
1409 refcount_t l2ad_alloc; /* allocated bytes */
1412 static list_t L2ARC_dev_list; /* device list */
1413 static list_t *l2arc_dev_list; /* device list pointer */
1414 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1415 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1416 static list_t L2ARC_free_on_write; /* free after write buf list */
1417 static list_t *l2arc_free_on_write; /* free after write list ptr */
1418 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1419 static uint64_t l2arc_ndev; /* number of devices */
1421 typedef struct l2arc_read_callback {
1422 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1423 blkptr_t l2rcb_bp; /* original blkptr */
1424 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1425 int l2rcb_flags; /* original flags */
1426 abd_t *l2rcb_abd; /* temporary buffer */
1427 } l2arc_read_callback_t;
1429 typedef struct l2arc_write_callback {
1430 l2arc_dev_t *l2wcb_dev; /* device info */
1431 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1432 } l2arc_write_callback_t;
1434 typedef struct l2arc_data_free {
1435 /* protected by l2arc_free_on_write_mtx */
1438 arc_buf_contents_t l2df_type;
1439 list_node_t l2df_list_node;
1440 } l2arc_data_free_t;
1442 static kmutex_t l2arc_feed_thr_lock;
1443 static kcondvar_t l2arc_feed_thr_cv;
1444 static uint8_t l2arc_thread_exit;
1446 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1447 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1448 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1449 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1450 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1451 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1452 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1453 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1454 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1455 static boolean_t arc_is_overflowing();
1456 static void arc_buf_watch(arc_buf_t *);
1458 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1459 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1460 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1461 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1463 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1464 static void l2arc_read_done(zio_t *);
1467 l2arc_trim(const arc_buf_hdr_t *hdr)
1469 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1471 ASSERT(HDR_HAS_L2HDR(hdr));
1472 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1474 if (HDR_GET_PSIZE(hdr) != 0) {
1475 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1476 HDR_GET_PSIZE(hdr), 0);
1481 * We use Cityhash for this. It's fast, and has good hash properties without
1482 * requiring any large static buffers.
1485 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1487 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1490 #define HDR_EMPTY(hdr) \
1491 ((hdr)->b_dva.dva_word[0] == 0 && \
1492 (hdr)->b_dva.dva_word[1] == 0)
1494 #define HDR_EQUAL(spa, dva, birth, hdr) \
1495 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1496 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1497 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1500 buf_discard_identity(arc_buf_hdr_t *hdr)
1502 hdr->b_dva.dva_word[0] = 0;
1503 hdr->b_dva.dva_word[1] = 0;
1507 static arc_buf_hdr_t *
1508 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1510 const dva_t *dva = BP_IDENTITY(bp);
1511 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1512 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1513 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1516 mutex_enter(hash_lock);
1517 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1518 hdr = hdr->b_hash_next) {
1519 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1524 mutex_exit(hash_lock);
1530 * Insert an entry into the hash table. If there is already an element
1531 * equal to elem in the hash table, then the already existing element
1532 * will be returned and the new element will not be inserted.
1533 * Otherwise returns NULL.
1534 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1536 static arc_buf_hdr_t *
1537 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1539 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1540 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1541 arc_buf_hdr_t *fhdr;
1544 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1545 ASSERT(hdr->b_birth != 0);
1546 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1548 if (lockp != NULL) {
1550 mutex_enter(hash_lock);
1552 ASSERT(MUTEX_HELD(hash_lock));
1555 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1556 fhdr = fhdr->b_hash_next, i++) {
1557 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1561 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1562 buf_hash_table.ht_table[idx] = hdr;
1563 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1565 /* collect some hash table performance data */
1567 ARCSTAT_BUMP(arcstat_hash_collisions);
1569 ARCSTAT_BUMP(arcstat_hash_chains);
1571 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1574 ARCSTAT_BUMP(arcstat_hash_elements);
1575 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1581 buf_hash_remove(arc_buf_hdr_t *hdr)
1583 arc_buf_hdr_t *fhdr, **hdrp;
1584 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1586 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1587 ASSERT(HDR_IN_HASH_TABLE(hdr));
1589 hdrp = &buf_hash_table.ht_table[idx];
1590 while ((fhdr = *hdrp) != hdr) {
1591 ASSERT3P(fhdr, !=, NULL);
1592 hdrp = &fhdr->b_hash_next;
1594 *hdrp = hdr->b_hash_next;
1595 hdr->b_hash_next = NULL;
1596 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1598 /* collect some hash table performance data */
1599 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1601 if (buf_hash_table.ht_table[idx] &&
1602 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1603 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1607 * Global data structures and functions for the buf kmem cache.
1609 static kmem_cache_t *hdr_full_cache;
1610 static kmem_cache_t *hdr_l2only_cache;
1611 static kmem_cache_t *buf_cache;
1618 kmem_free(buf_hash_table.ht_table,
1619 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1620 for (i = 0; i < BUF_LOCKS; i++)
1621 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1622 kmem_cache_destroy(hdr_full_cache);
1623 kmem_cache_destroy(hdr_l2only_cache);
1624 kmem_cache_destroy(buf_cache);
1628 * Constructor callback - called when the cache is empty
1629 * and a new buf is requested.
1633 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1635 arc_buf_hdr_t *hdr = vbuf;
1637 bzero(hdr, HDR_FULL_SIZE);
1638 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1639 refcount_create(&hdr->b_l1hdr.b_refcnt);
1640 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1641 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1642 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1649 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1651 arc_buf_hdr_t *hdr = vbuf;
1653 bzero(hdr, HDR_L2ONLY_SIZE);
1654 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1661 buf_cons(void *vbuf, void *unused, int kmflag)
1663 arc_buf_t *buf = vbuf;
1665 bzero(buf, sizeof (arc_buf_t));
1666 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1667 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1673 * Destructor callback - called when a cached buf is
1674 * no longer required.
1678 hdr_full_dest(void *vbuf, void *unused)
1680 arc_buf_hdr_t *hdr = vbuf;
1682 ASSERT(HDR_EMPTY(hdr));
1683 cv_destroy(&hdr->b_l1hdr.b_cv);
1684 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1685 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1686 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1687 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1692 hdr_l2only_dest(void *vbuf, void *unused)
1694 arc_buf_hdr_t *hdr = vbuf;
1696 ASSERT(HDR_EMPTY(hdr));
1697 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1702 buf_dest(void *vbuf, void *unused)
1704 arc_buf_t *buf = vbuf;
1706 mutex_destroy(&buf->b_evict_lock);
1707 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1711 * Reclaim callback -- invoked when memory is low.
1715 hdr_recl(void *unused)
1717 dprintf("hdr_recl called\n");
1719 * umem calls the reclaim func when we destroy the buf cache,
1720 * which is after we do arc_fini().
1723 cv_signal(&arc_reclaim_thread_cv);
1730 uint64_t hsize = 1ULL << 12;
1734 * The hash table is big enough to fill all of physical memory
1735 * with an average block size of zfs_arc_average_blocksize (default 8K).
1736 * By default, the table will take up
1737 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1739 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1742 buf_hash_table.ht_mask = hsize - 1;
1743 buf_hash_table.ht_table =
1744 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1745 if (buf_hash_table.ht_table == NULL) {
1746 ASSERT(hsize > (1ULL << 8));
1751 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1752 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1753 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1754 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1756 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1757 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1759 for (i = 0; i < 256; i++)
1760 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1761 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1763 for (i = 0; i < BUF_LOCKS; i++) {
1764 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1765 NULL, MUTEX_DEFAULT, NULL);
1770 * This is the size that the buf occupies in memory. If the buf is compressed,
1771 * it will correspond to the compressed size. You should use this method of
1772 * getting the buf size unless you explicitly need the logical size.
1775 arc_buf_size(arc_buf_t *buf)
1777 return (ARC_BUF_COMPRESSED(buf) ?
1778 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1782 arc_buf_lsize(arc_buf_t *buf)
1784 return (HDR_GET_LSIZE(buf->b_hdr));
1788 arc_get_compression(arc_buf_t *buf)
1790 return (ARC_BUF_COMPRESSED(buf) ?
1791 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1794 #define ARC_MINTIME (hz>>4) /* 62 ms */
1796 static inline boolean_t
1797 arc_buf_is_shared(arc_buf_t *buf)
1799 boolean_t shared = (buf->b_data != NULL &&
1800 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1801 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1802 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1803 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1804 IMPLY(shared, ARC_BUF_SHARED(buf));
1805 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1808 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1809 * already being shared" requirement prevents us from doing that.
1816 * Free the checksum associated with this header. If there is no checksum, this
1820 arc_cksum_free(arc_buf_hdr_t *hdr)
1822 ASSERT(HDR_HAS_L1HDR(hdr));
1823 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1824 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1825 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1826 hdr->b_l1hdr.b_freeze_cksum = NULL;
1828 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1832 * Return true iff at least one of the bufs on hdr is not compressed.
1835 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1837 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1838 if (!ARC_BUF_COMPRESSED(b)) {
1846 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1847 * matches the checksum that is stored in the hdr. If there is no checksum,
1848 * or if the buf is compressed, this is a no-op.
1851 arc_cksum_verify(arc_buf_t *buf)
1853 arc_buf_hdr_t *hdr = buf->b_hdr;
1856 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1859 if (ARC_BUF_COMPRESSED(buf)) {
1860 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1861 arc_hdr_has_uncompressed_buf(hdr));
1865 ASSERT(HDR_HAS_L1HDR(hdr));
1867 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1868 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1869 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1873 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1874 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1875 panic("buffer modified while frozen!");
1876 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1880 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1882 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1883 boolean_t valid_cksum;
1885 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1886 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1889 * We rely on the blkptr's checksum to determine if the block
1890 * is valid or not. When compressed arc is enabled, the l2arc
1891 * writes the block to the l2arc just as it appears in the pool.
1892 * This allows us to use the blkptr's checksum to validate the
1893 * data that we just read off of the l2arc without having to store
1894 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1895 * arc is disabled, then the data written to the l2arc is always
1896 * uncompressed and won't match the block as it exists in the main
1897 * pool. When this is the case, we must first compress it if it is
1898 * compressed on the main pool before we can validate the checksum.
1900 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1901 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1902 uint64_t lsize = HDR_GET_LSIZE(hdr);
1905 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1906 csize = zio_compress_data(compress, zio->io_abd,
1907 abd_to_buf(cdata), lsize);
1909 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1910 if (csize < HDR_GET_PSIZE(hdr)) {
1912 * Compressed blocks are always a multiple of the
1913 * smallest ashift in the pool. Ideally, we would
1914 * like to round up the csize to the next
1915 * spa_min_ashift but that value may have changed
1916 * since the block was last written. Instead,
1917 * we rely on the fact that the hdr's psize
1918 * was set to the psize of the block when it was
1919 * last written. We set the csize to that value
1920 * and zero out any part that should not contain
1923 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1924 csize = HDR_GET_PSIZE(hdr);
1926 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1930 * Block pointers always store the checksum for the logical data.
1931 * If the block pointer has the gang bit set, then the checksum
1932 * it represents is for the reconstituted data and not for an
1933 * individual gang member. The zio pipeline, however, must be able to
1934 * determine the checksum of each of the gang constituents so it
1935 * treats the checksum comparison differently than what we need
1936 * for l2arc blocks. This prevents us from using the
1937 * zio_checksum_error() interface directly. Instead we must call the
1938 * zio_checksum_error_impl() so that we can ensure the checksum is
1939 * generated using the correct checksum algorithm and accounts for the
1940 * logical I/O size and not just a gang fragment.
1942 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1943 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1944 zio->io_offset, NULL) == 0);
1945 zio_pop_transforms(zio);
1946 return (valid_cksum);
1950 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1951 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1952 * isn't modified later on. If buf is compressed or there is already a checksum
1953 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1956 arc_cksum_compute(arc_buf_t *buf)
1958 arc_buf_hdr_t *hdr = buf->b_hdr;
1960 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1963 ASSERT(HDR_HAS_L1HDR(hdr));
1965 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1966 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1967 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1968 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1970 } else if (ARC_BUF_COMPRESSED(buf)) {
1971 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1975 ASSERT(!ARC_BUF_COMPRESSED(buf));
1976 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1978 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1979 hdr->b_l1hdr.b_freeze_cksum);
1980 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1988 typedef struct procctl {
1996 arc_buf_unwatch(arc_buf_t *buf)
2003 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2004 ctl.prwatch.pr_size = 0;
2005 ctl.prwatch.pr_wflags = 0;
2006 result = write(arc_procfd, &ctl, sizeof (ctl));
2007 ASSERT3U(result, ==, sizeof (ctl));
2014 arc_buf_watch(arc_buf_t *buf)
2021 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2022 ctl.prwatch.pr_size = arc_buf_size(buf);
2023 ctl.prwatch.pr_wflags = WA_WRITE;
2024 result = write(arc_procfd, &ctl, sizeof (ctl));
2025 ASSERT3U(result, ==, sizeof (ctl));
2029 #endif /* illumos */
2031 static arc_buf_contents_t
2032 arc_buf_type(arc_buf_hdr_t *hdr)
2034 arc_buf_contents_t type;
2035 if (HDR_ISTYPE_METADATA(hdr)) {
2036 type = ARC_BUFC_METADATA;
2038 type = ARC_BUFC_DATA;
2040 VERIFY3U(hdr->b_type, ==, type);
2045 arc_is_metadata(arc_buf_t *buf)
2047 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2051 arc_bufc_to_flags(arc_buf_contents_t type)
2055 /* metadata field is 0 if buffer contains normal data */
2057 case ARC_BUFC_METADATA:
2058 return (ARC_FLAG_BUFC_METADATA);
2062 panic("undefined ARC buffer type!");
2063 return ((uint32_t)-1);
2067 arc_buf_thaw(arc_buf_t *buf)
2069 arc_buf_hdr_t *hdr = buf->b_hdr;
2071 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2072 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2074 arc_cksum_verify(buf);
2077 * Compressed buffers do not manipulate the b_freeze_cksum or
2078 * allocate b_thawed.
2080 if (ARC_BUF_COMPRESSED(buf)) {
2081 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2082 arc_hdr_has_uncompressed_buf(hdr));
2086 ASSERT(HDR_HAS_L1HDR(hdr));
2087 arc_cksum_free(hdr);
2089 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2091 if (zfs_flags & ZFS_DEBUG_MODIFY) {
2092 if (hdr->b_l1hdr.b_thawed != NULL)
2093 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2094 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2098 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2101 arc_buf_unwatch(buf);
2106 arc_buf_freeze(arc_buf_t *buf)
2108 arc_buf_hdr_t *hdr = buf->b_hdr;
2109 kmutex_t *hash_lock;
2111 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2114 if (ARC_BUF_COMPRESSED(buf)) {
2115 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2116 arc_hdr_has_uncompressed_buf(hdr));
2120 hash_lock = HDR_LOCK(hdr);
2121 mutex_enter(hash_lock);
2123 ASSERT(HDR_HAS_L1HDR(hdr));
2124 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
2125 hdr->b_l1hdr.b_state == arc_anon);
2126 arc_cksum_compute(buf);
2127 mutex_exit(hash_lock);
2131 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2132 * the following functions should be used to ensure that the flags are
2133 * updated in a thread-safe way. When manipulating the flags either
2134 * the hash_lock must be held or the hdr must be undiscoverable. This
2135 * ensures that we're not racing with any other threads when updating
2139 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2141 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2142 hdr->b_flags |= flags;
2146 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2148 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2149 hdr->b_flags &= ~flags;
2153 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2154 * done in a special way since we have to clear and set bits
2155 * at the same time. Consumers that wish to set the compression bits
2156 * must use this function to ensure that the flags are updated in
2157 * thread-safe manner.
2160 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2162 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2165 * Holes and embedded blocks will always have a psize = 0 so
2166 * we ignore the compression of the blkptr and set the
2167 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2168 * Holes and embedded blocks remain anonymous so we don't
2169 * want to uncompress them. Mark them as uncompressed.
2171 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2172 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2173 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2174 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2175 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2177 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2178 HDR_SET_COMPRESS(hdr, cmp);
2179 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2180 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2185 * Looks for another buf on the same hdr which has the data decompressed, copies
2186 * from it, and returns true. If no such buf exists, returns false.
2189 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
2191 arc_buf_hdr_t *hdr = buf->b_hdr;
2192 boolean_t copied = B_FALSE;
2194 ASSERT(HDR_HAS_L1HDR(hdr));
2195 ASSERT3P(buf->b_data, !=, NULL);
2196 ASSERT(!ARC_BUF_COMPRESSED(buf));
2198 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2199 from = from->b_next) {
2200 /* can't use our own data buffer */
2205 if (!ARC_BUF_COMPRESSED(from)) {
2206 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2213 * There were no decompressed bufs, so there should not be a
2214 * checksum on the hdr either.
2216 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
2222 * Given a buf that has a data buffer attached to it, this function will
2223 * efficiently fill the buf with data of the specified compression setting from
2224 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2225 * are already sharing a data buf, no copy is performed.
2227 * If the buf is marked as compressed but uncompressed data was requested, this
2228 * will allocate a new data buffer for the buf, remove that flag, and fill the
2229 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2230 * uncompressed data, and (since we haven't added support for it yet) if you
2231 * want compressed data your buf must already be marked as compressed and have
2232 * the correct-sized data buffer.
2235 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2237 arc_buf_hdr_t *hdr = buf->b_hdr;
2238 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2239 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2241 ASSERT3P(buf->b_data, !=, NULL);
2242 IMPLY(compressed, hdr_compressed);
2243 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2245 if (hdr_compressed == compressed) {
2246 if (!arc_buf_is_shared(buf)) {
2247 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2251 ASSERT(hdr_compressed);
2252 ASSERT(!compressed);
2253 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2256 * If the buf is sharing its data with the hdr, unlink it and
2257 * allocate a new data buffer for the buf.
2259 if (arc_buf_is_shared(buf)) {
2260 ASSERT(ARC_BUF_COMPRESSED(buf));
2262 /* We need to give the buf it's own b_data */
2263 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2265 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2266 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2268 /* Previously overhead was 0; just add new overhead */
2269 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2270 } else if (ARC_BUF_COMPRESSED(buf)) {
2271 /* We need to reallocate the buf's b_data */
2272 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2275 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2277 /* We increased the size of b_data; update overhead */
2278 ARCSTAT_INCR(arcstat_overhead_size,
2279 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2283 * Regardless of the buf's previous compression settings, it
2284 * should not be compressed at the end of this function.
2286 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2289 * Try copying the data from another buf which already has a
2290 * decompressed version. If that's not possible, it's time to
2291 * bite the bullet and decompress the data from the hdr.
2293 if (arc_buf_try_copy_decompressed_data(buf)) {
2294 /* Skip byteswapping and checksumming (already done) */
2295 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2298 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2299 hdr->b_l1hdr.b_pabd, buf->b_data,
2300 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2303 * Absent hardware errors or software bugs, this should
2304 * be impossible, but log it anyway so we can debug it.
2308 "hdr %p, compress %d, psize %d, lsize %d",
2309 hdr, HDR_GET_COMPRESS(hdr),
2310 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2311 return (SET_ERROR(EIO));
2316 /* Byteswap the buf's data if necessary */
2317 if (bswap != DMU_BSWAP_NUMFUNCS) {
2318 ASSERT(!HDR_SHARED_DATA(hdr));
2319 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2320 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2323 /* Compute the hdr's checksum if necessary */
2324 arc_cksum_compute(buf);
2330 arc_decompress(arc_buf_t *buf)
2332 return (arc_buf_fill(buf, B_FALSE));
2336 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2339 arc_hdr_size(arc_buf_hdr_t *hdr)
2343 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2344 HDR_GET_PSIZE(hdr) > 0) {
2345 size = HDR_GET_PSIZE(hdr);
2347 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2348 size = HDR_GET_LSIZE(hdr);
2354 * Increment the amount of evictable space in the arc_state_t's refcount.
2355 * We account for the space used by the hdr and the arc buf individually
2356 * so that we can add and remove them from the refcount individually.
2359 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2361 arc_buf_contents_t type = arc_buf_type(hdr);
2363 ASSERT(HDR_HAS_L1HDR(hdr));
2365 if (GHOST_STATE(state)) {
2366 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2367 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2368 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2369 (void) refcount_add_many(&state->arcs_esize[type],
2370 HDR_GET_LSIZE(hdr), hdr);
2374 ASSERT(!GHOST_STATE(state));
2375 if (hdr->b_l1hdr.b_pabd != NULL) {
2376 (void) refcount_add_many(&state->arcs_esize[type],
2377 arc_hdr_size(hdr), hdr);
2379 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2380 buf = buf->b_next) {
2381 if (arc_buf_is_shared(buf))
2383 (void) refcount_add_many(&state->arcs_esize[type],
2384 arc_buf_size(buf), buf);
2389 * Decrement the amount of evictable space in the arc_state_t's refcount.
2390 * We account for the space used by the hdr and the arc buf individually
2391 * so that we can add and remove them from the refcount individually.
2394 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2396 arc_buf_contents_t type = arc_buf_type(hdr);
2398 ASSERT(HDR_HAS_L1HDR(hdr));
2400 if (GHOST_STATE(state)) {
2401 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2402 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2403 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2404 (void) refcount_remove_many(&state->arcs_esize[type],
2405 HDR_GET_LSIZE(hdr), hdr);
2409 ASSERT(!GHOST_STATE(state));
2410 if (hdr->b_l1hdr.b_pabd != NULL) {
2411 (void) refcount_remove_many(&state->arcs_esize[type],
2412 arc_hdr_size(hdr), hdr);
2414 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2415 buf = buf->b_next) {
2416 if (arc_buf_is_shared(buf))
2418 (void) refcount_remove_many(&state->arcs_esize[type],
2419 arc_buf_size(buf), buf);
2424 * Add a reference to this hdr indicating that someone is actively
2425 * referencing that memory. When the refcount transitions from 0 to 1,
2426 * we remove it from the respective arc_state_t list to indicate that
2427 * it is not evictable.
2430 add_reference(arc_buf_hdr_t *hdr, void *tag)
2432 ASSERT(HDR_HAS_L1HDR(hdr));
2433 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2434 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2435 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2436 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2439 arc_state_t *state = hdr->b_l1hdr.b_state;
2441 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2442 (state != arc_anon)) {
2443 /* We don't use the L2-only state list. */
2444 if (state != arc_l2c_only) {
2445 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2447 arc_evictable_space_decrement(hdr, state);
2449 /* remove the prefetch flag if we get a reference */
2450 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2455 * Remove a reference from this hdr. When the reference transitions from
2456 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2457 * list making it eligible for eviction.
2460 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2463 arc_state_t *state = hdr->b_l1hdr.b_state;
2465 ASSERT(HDR_HAS_L1HDR(hdr));
2466 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2467 ASSERT(!GHOST_STATE(state));
2470 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2471 * check to prevent usage of the arc_l2c_only list.
2473 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2474 (state != arc_anon)) {
2475 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2476 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2477 arc_evictable_space_increment(hdr, state);
2483 * Move the supplied buffer to the indicated state. The hash lock
2484 * for the buffer must be held by the caller.
2487 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2488 kmutex_t *hash_lock)
2490 arc_state_t *old_state;
2493 boolean_t update_old, update_new;
2494 arc_buf_contents_t buftype = arc_buf_type(hdr);
2497 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2498 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2499 * L1 hdr doesn't always exist when we change state to arc_anon before
2500 * destroying a header, in which case reallocating to add the L1 hdr is
2503 if (HDR_HAS_L1HDR(hdr)) {
2504 old_state = hdr->b_l1hdr.b_state;
2505 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2506 bufcnt = hdr->b_l1hdr.b_bufcnt;
2507 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2509 old_state = arc_l2c_only;
2512 update_old = B_FALSE;
2514 update_new = update_old;
2516 ASSERT(MUTEX_HELD(hash_lock));
2517 ASSERT3P(new_state, !=, old_state);
2518 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2519 ASSERT(old_state != arc_anon || bufcnt <= 1);
2522 * If this buffer is evictable, transfer it from the
2523 * old state list to the new state list.
2526 if (old_state != arc_anon && old_state != arc_l2c_only) {
2527 ASSERT(HDR_HAS_L1HDR(hdr));
2528 multilist_remove(old_state->arcs_list[buftype], hdr);
2530 if (GHOST_STATE(old_state)) {
2532 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2533 update_old = B_TRUE;
2535 arc_evictable_space_decrement(hdr, old_state);
2537 if (new_state != arc_anon && new_state != arc_l2c_only) {
2540 * An L1 header always exists here, since if we're
2541 * moving to some L1-cached state (i.e. not l2c_only or
2542 * anonymous), we realloc the header to add an L1hdr
2545 ASSERT(HDR_HAS_L1HDR(hdr));
2546 multilist_insert(new_state->arcs_list[buftype], hdr);
2548 if (GHOST_STATE(new_state)) {
2550 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2551 update_new = B_TRUE;
2553 arc_evictable_space_increment(hdr, new_state);
2557 ASSERT(!HDR_EMPTY(hdr));
2558 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2559 buf_hash_remove(hdr);
2561 /* adjust state sizes (ignore arc_l2c_only) */
2563 if (update_new && new_state != arc_l2c_only) {
2564 ASSERT(HDR_HAS_L1HDR(hdr));
2565 if (GHOST_STATE(new_state)) {
2569 * When moving a header to a ghost state, we first
2570 * remove all arc buffers. Thus, we'll have a
2571 * bufcnt of zero, and no arc buffer to use for
2572 * the reference. As a result, we use the arc
2573 * header pointer for the reference.
2575 (void) refcount_add_many(&new_state->arcs_size,
2576 HDR_GET_LSIZE(hdr), hdr);
2577 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2579 uint32_t buffers = 0;
2582 * Each individual buffer holds a unique reference,
2583 * thus we must remove each of these references one
2586 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2587 buf = buf->b_next) {
2588 ASSERT3U(bufcnt, !=, 0);
2592 * When the arc_buf_t is sharing the data
2593 * block with the hdr, the owner of the
2594 * reference belongs to the hdr. Only
2595 * add to the refcount if the arc_buf_t is
2598 if (arc_buf_is_shared(buf))
2601 (void) refcount_add_many(&new_state->arcs_size,
2602 arc_buf_size(buf), buf);
2604 ASSERT3U(bufcnt, ==, buffers);
2606 if (hdr->b_l1hdr.b_pabd != NULL) {
2607 (void) refcount_add_many(&new_state->arcs_size,
2608 arc_hdr_size(hdr), hdr);
2610 ASSERT(GHOST_STATE(old_state));
2615 if (update_old && old_state != arc_l2c_only) {
2616 ASSERT(HDR_HAS_L1HDR(hdr));
2617 if (GHOST_STATE(old_state)) {
2619 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2622 * When moving a header off of a ghost state,
2623 * the header will not contain any arc buffers.
2624 * We use the arc header pointer for the reference
2625 * which is exactly what we did when we put the
2626 * header on the ghost state.
2629 (void) refcount_remove_many(&old_state->arcs_size,
2630 HDR_GET_LSIZE(hdr), hdr);
2632 uint32_t buffers = 0;
2635 * Each individual buffer holds a unique reference,
2636 * thus we must remove each of these references one
2639 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2640 buf = buf->b_next) {
2641 ASSERT3U(bufcnt, !=, 0);
2645 * When the arc_buf_t is sharing the data
2646 * block with the hdr, the owner of the
2647 * reference belongs to the hdr. Only
2648 * add to the refcount if the arc_buf_t is
2651 if (arc_buf_is_shared(buf))
2654 (void) refcount_remove_many(
2655 &old_state->arcs_size, arc_buf_size(buf),
2658 ASSERT3U(bufcnt, ==, buffers);
2659 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2660 (void) refcount_remove_many(
2661 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2665 if (HDR_HAS_L1HDR(hdr))
2666 hdr->b_l1hdr.b_state = new_state;
2669 * L2 headers should never be on the L2 state list since they don't
2670 * have L1 headers allocated.
2672 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2673 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2677 arc_space_consume(uint64_t space, arc_space_type_t type)
2679 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2682 case ARC_SPACE_DATA:
2683 aggsum_add(&astat_data_size, space);
2685 case ARC_SPACE_META:
2686 aggsum_add(&astat_metadata_size, space);
2688 case ARC_SPACE_OTHER:
2689 aggsum_add(&astat_other_size, space);
2691 case ARC_SPACE_HDRS:
2692 aggsum_add(&astat_hdr_size, space);
2694 case ARC_SPACE_L2HDRS:
2695 aggsum_add(&astat_l2_hdr_size, space);
2699 if (type != ARC_SPACE_DATA)
2700 aggsum_add(&arc_meta_used, space);
2702 aggsum_add(&arc_size, space);
2706 arc_space_return(uint64_t space, arc_space_type_t type)
2708 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2711 case ARC_SPACE_DATA:
2712 aggsum_add(&astat_data_size, -space);
2714 case ARC_SPACE_META:
2715 aggsum_add(&astat_metadata_size, -space);
2717 case ARC_SPACE_OTHER:
2718 aggsum_add(&astat_other_size, -space);
2720 case ARC_SPACE_HDRS:
2721 aggsum_add(&astat_hdr_size, -space);
2723 case ARC_SPACE_L2HDRS:
2724 aggsum_add(&astat_l2_hdr_size, -space);
2728 if (type != ARC_SPACE_DATA) {
2729 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2731 * We use the upper bound here rather than the precise value
2732 * because the arc_meta_max value doesn't need to be
2733 * precise. It's only consumed by humans via arcstats.
2735 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2736 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2737 aggsum_add(&arc_meta_used, -space);
2740 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2741 aggsum_add(&arc_size, -space);
2745 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2746 * with the hdr's b_pabd.
2749 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2752 * The criteria for sharing a hdr's data are:
2753 * 1. the hdr's compression matches the buf's compression
2754 * 2. the hdr doesn't need to be byteswapped
2755 * 3. the hdr isn't already being shared
2756 * 4. the buf is either compressed or it is the last buf in the hdr list
2758 * Criterion #4 maintains the invariant that shared uncompressed
2759 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2760 * might ask, "if a compressed buf is allocated first, won't that be the
2761 * last thing in the list?", but in that case it's impossible to create
2762 * a shared uncompressed buf anyway (because the hdr must be compressed
2763 * to have the compressed buf). You might also think that #3 is
2764 * sufficient to make this guarantee, however it's possible
2765 * (specifically in the rare L2ARC write race mentioned in
2766 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2767 * is sharable, but wasn't at the time of its allocation. Rather than
2768 * allow a new shared uncompressed buf to be created and then shuffle
2769 * the list around to make it the last element, this simply disallows
2770 * sharing if the new buf isn't the first to be added.
2772 ASSERT3P(buf->b_hdr, ==, hdr);
2773 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2774 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2775 return (buf_compressed == hdr_compressed &&
2776 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2777 !HDR_SHARED_DATA(hdr) &&
2778 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2782 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2783 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2784 * copy was made successfully, or an error code otherwise.
2787 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2788 boolean_t fill, arc_buf_t **ret)
2792 ASSERT(HDR_HAS_L1HDR(hdr));
2793 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2794 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2795 hdr->b_type == ARC_BUFC_METADATA);
2796 ASSERT3P(ret, !=, NULL);
2797 ASSERT3P(*ret, ==, NULL);
2799 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2802 buf->b_next = hdr->b_l1hdr.b_buf;
2805 add_reference(hdr, tag);
2808 * We're about to change the hdr's b_flags. We must either
2809 * hold the hash_lock or be undiscoverable.
2811 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2814 * Only honor requests for compressed bufs if the hdr is actually
2817 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2818 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2821 * If the hdr's data can be shared then we share the data buffer and
2822 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2823 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2824 * buffer to store the buf's data.
2826 * There are two additional restrictions here because we're sharing
2827 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2828 * actively involved in an L2ARC write, because if this buf is used by
2829 * an arc_write() then the hdr's data buffer will be released when the
2830 * write completes, even though the L2ARC write might still be using it.
2831 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2832 * need to be ABD-aware.
2834 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2835 abd_is_linear(hdr->b_l1hdr.b_pabd);
2837 /* Set up b_data and sharing */
2839 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2840 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2841 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2844 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2845 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2847 VERIFY3P(buf->b_data, !=, NULL);
2849 hdr->b_l1hdr.b_buf = buf;
2850 hdr->b_l1hdr.b_bufcnt += 1;
2853 * If the user wants the data from the hdr, we need to either copy or
2854 * decompress the data.
2857 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2863 static char *arc_onloan_tag = "onloan";
2866 arc_loaned_bytes_update(int64_t delta)
2868 atomic_add_64(&arc_loaned_bytes, delta);
2870 /* assert that it did not wrap around */
2871 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2875 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2876 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2877 * buffers must be returned to the arc before they can be used by the DMU or
2881 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2883 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2884 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2886 arc_loaned_bytes_update(arc_buf_size(buf));
2892 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2893 enum zio_compress compression_type)
2895 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2896 psize, lsize, compression_type);
2898 arc_loaned_bytes_update(arc_buf_size(buf));
2905 * Return a loaned arc buffer to the arc.
2908 arc_return_buf(arc_buf_t *buf, void *tag)
2910 arc_buf_hdr_t *hdr = buf->b_hdr;
2912 ASSERT3P(buf->b_data, !=, NULL);
2913 ASSERT(HDR_HAS_L1HDR(hdr));
2914 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2915 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2917 arc_loaned_bytes_update(-arc_buf_size(buf));
2920 /* Detach an arc_buf from a dbuf (tag) */
2922 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2924 arc_buf_hdr_t *hdr = buf->b_hdr;
2926 ASSERT3P(buf->b_data, !=, NULL);
2927 ASSERT(HDR_HAS_L1HDR(hdr));
2928 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2929 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2931 arc_loaned_bytes_update(arc_buf_size(buf));
2935 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2937 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2940 df->l2df_size = size;
2941 df->l2df_type = type;
2942 mutex_enter(&l2arc_free_on_write_mtx);
2943 list_insert_head(l2arc_free_on_write, df);
2944 mutex_exit(&l2arc_free_on_write_mtx);
2948 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2950 arc_state_t *state = hdr->b_l1hdr.b_state;
2951 arc_buf_contents_t type = arc_buf_type(hdr);
2952 uint64_t size = arc_hdr_size(hdr);
2954 /* protected by hash lock, if in the hash table */
2955 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2956 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2957 ASSERT(state != arc_anon && state != arc_l2c_only);
2959 (void) refcount_remove_many(&state->arcs_esize[type],
2962 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2963 if (type == ARC_BUFC_METADATA) {
2964 arc_space_return(size, ARC_SPACE_META);
2966 ASSERT(type == ARC_BUFC_DATA);
2967 arc_space_return(size, ARC_SPACE_DATA);
2970 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2974 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2975 * data buffer, we transfer the refcount ownership to the hdr and update
2976 * the appropriate kstats.
2979 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2981 arc_state_t *state = hdr->b_l1hdr.b_state;
2983 ASSERT(arc_can_share(hdr, buf));
2984 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2985 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2988 * Start sharing the data buffer. We transfer the
2989 * refcount ownership to the hdr since it always owns
2990 * the refcount whenever an arc_buf_t is shared.
2992 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2993 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2994 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2995 HDR_ISTYPE_METADATA(hdr));
2996 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2997 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3000 * Since we've transferred ownership to the hdr we need
3001 * to increment its compressed and uncompressed kstats and
3002 * decrement the overhead size.
3004 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3005 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3006 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3010 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3012 arc_state_t *state = hdr->b_l1hdr.b_state;
3014 ASSERT(arc_buf_is_shared(buf));
3015 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3016 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3019 * We are no longer sharing this buffer so we need
3020 * to transfer its ownership to the rightful owner.
3022 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
3023 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3024 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3025 abd_put(hdr->b_l1hdr.b_pabd);
3026 hdr->b_l1hdr.b_pabd = NULL;
3027 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3030 * Since the buffer is no longer shared between
3031 * the arc buf and the hdr, count it as overhead.
3033 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3034 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3035 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3039 * Remove an arc_buf_t from the hdr's buf list and return the last
3040 * arc_buf_t on the list. If no buffers remain on the list then return
3044 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3046 ASSERT(HDR_HAS_L1HDR(hdr));
3047 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3049 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3050 arc_buf_t *lastbuf = NULL;
3053 * Remove the buf from the hdr list and locate the last
3054 * remaining buffer on the list.
3056 while (*bufp != NULL) {
3058 *bufp = buf->b_next;
3061 * If we've removed a buffer in the middle of
3062 * the list then update the lastbuf and update
3065 if (*bufp != NULL) {
3067 bufp = &(*bufp)->b_next;
3071 ASSERT3P(lastbuf, !=, buf);
3072 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3073 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3074 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3080 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3084 arc_buf_destroy_impl(arc_buf_t *buf)
3086 arc_buf_hdr_t *hdr = buf->b_hdr;
3089 * Free up the data associated with the buf but only if we're not
3090 * sharing this with the hdr. If we are sharing it with the hdr, the
3091 * hdr is responsible for doing the free.
3093 if (buf->b_data != NULL) {
3095 * We're about to change the hdr's b_flags. We must either
3096 * hold the hash_lock or be undiscoverable.
3098 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3100 arc_cksum_verify(buf);
3102 arc_buf_unwatch(buf);
3105 if (arc_buf_is_shared(buf)) {
3106 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3108 uint64_t size = arc_buf_size(buf);
3109 arc_free_data_buf(hdr, buf->b_data, size, buf);
3110 ARCSTAT_INCR(arcstat_overhead_size, -size);
3114 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3115 hdr->b_l1hdr.b_bufcnt -= 1;
3118 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3120 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3122 * If the current arc_buf_t is sharing its data buffer with the
3123 * hdr, then reassign the hdr's b_pabd to share it with the new
3124 * buffer at the end of the list. The shared buffer is always
3125 * the last one on the hdr's buffer list.
3127 * There is an equivalent case for compressed bufs, but since
3128 * they aren't guaranteed to be the last buf in the list and
3129 * that is an exceedingly rare case, we just allow that space be
3130 * wasted temporarily.
3132 if (lastbuf != NULL) {
3133 /* Only one buf can be shared at once */
3134 VERIFY(!arc_buf_is_shared(lastbuf));
3135 /* hdr is uncompressed so can't have compressed buf */
3136 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3138 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3139 arc_hdr_free_pabd(hdr);
3142 * We must setup a new shared block between the
3143 * last buffer and the hdr. The data would have
3144 * been allocated by the arc buf so we need to transfer
3145 * ownership to the hdr since it's now being shared.
3147 arc_share_buf(hdr, lastbuf);
3149 } else if (HDR_SHARED_DATA(hdr)) {
3151 * Uncompressed shared buffers are always at the end
3152 * of the list. Compressed buffers don't have the
3153 * same requirements. This makes it hard to
3154 * simply assert that the lastbuf is shared so
3155 * we rely on the hdr's compression flags to determine
3156 * if we have a compressed, shared buffer.
3158 ASSERT3P(lastbuf, !=, NULL);
3159 ASSERT(arc_buf_is_shared(lastbuf) ||
3160 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
3164 * Free the checksum if we're removing the last uncompressed buf from
3167 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3168 arc_cksum_free(hdr);
3171 /* clean up the buf */
3173 kmem_cache_free(buf_cache, buf);
3177 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
3179 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3180 ASSERT(HDR_HAS_L1HDR(hdr));
3181 ASSERT(!HDR_SHARED_DATA(hdr));
3183 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3184 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
3185 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3186 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3188 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3189 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3193 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3195 ASSERT(HDR_HAS_L1HDR(hdr));
3196 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3199 * If the hdr is currently being written to the l2arc then
3200 * we defer freeing the data by adding it to the l2arc_free_on_write
3201 * list. The l2arc will free the data once it's finished
3202 * writing it to the l2arc device.
3204 if (HDR_L2_WRITING(hdr)) {
3205 arc_hdr_free_on_write(hdr);
3206 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3208 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3209 arc_hdr_size(hdr), hdr);
3211 hdr->b_l1hdr.b_pabd = NULL;
3212 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3214 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3215 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3218 static arc_buf_hdr_t *
3219 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3220 enum zio_compress compression_type, arc_buf_contents_t type)
3224 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3226 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3227 ASSERT(HDR_EMPTY(hdr));
3228 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3229 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3230 HDR_SET_PSIZE(hdr, psize);
3231 HDR_SET_LSIZE(hdr, lsize);
3235 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3236 arc_hdr_set_compress(hdr, compression_type);
3238 hdr->b_l1hdr.b_state = arc_anon;
3239 hdr->b_l1hdr.b_arc_access = 0;
3240 hdr->b_l1hdr.b_bufcnt = 0;
3241 hdr->b_l1hdr.b_buf = NULL;
3244 * Allocate the hdr's buffer. This will contain either
3245 * the compressed or uncompressed data depending on the block
3246 * it references and compressed arc enablement.
3248 arc_hdr_alloc_pabd(hdr);
3249 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3255 * Transition between the two allocation states for the arc_buf_hdr struct.
3256 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3257 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3258 * version is used when a cache buffer is only in the L2ARC in order to reduce
3261 static arc_buf_hdr_t *
3262 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3264 ASSERT(HDR_HAS_L2HDR(hdr));
3266 arc_buf_hdr_t *nhdr;
3267 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3269 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3270 (old == hdr_l2only_cache && new == hdr_full_cache));
3272 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3274 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3275 buf_hash_remove(hdr);
3277 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3279 if (new == hdr_full_cache) {
3280 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3282 * arc_access and arc_change_state need to be aware that a
3283 * header has just come out of L2ARC, so we set its state to
3284 * l2c_only even though it's about to change.
3286 nhdr->b_l1hdr.b_state = arc_l2c_only;
3288 /* Verify previous threads set to NULL before freeing */
3289 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3291 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3292 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3293 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3296 * If we've reached here, We must have been called from
3297 * arc_evict_hdr(), as such we should have already been
3298 * removed from any ghost list we were previously on
3299 * (which protects us from racing with arc_evict_state),
3300 * thus no locking is needed during this check.
3302 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3305 * A buffer must not be moved into the arc_l2c_only
3306 * state if it's not finished being written out to the
3307 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3308 * might try to be accessed, even though it was removed.
3310 VERIFY(!HDR_L2_WRITING(hdr));
3311 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3314 if (hdr->b_l1hdr.b_thawed != NULL) {
3315 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3316 hdr->b_l1hdr.b_thawed = NULL;
3320 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3323 * The header has been reallocated so we need to re-insert it into any
3326 (void) buf_hash_insert(nhdr, NULL);
3328 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3330 mutex_enter(&dev->l2ad_mtx);
3333 * We must place the realloc'ed header back into the list at
3334 * the same spot. Otherwise, if it's placed earlier in the list,
3335 * l2arc_write_buffers() could find it during the function's
3336 * write phase, and try to write it out to the l2arc.
3338 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3339 list_remove(&dev->l2ad_buflist, hdr);
3341 mutex_exit(&dev->l2ad_mtx);
3344 * Since we're using the pointer address as the tag when
3345 * incrementing and decrementing the l2ad_alloc refcount, we
3346 * must remove the old pointer (that we're about to destroy) and
3347 * add the new pointer to the refcount. Otherwise we'd remove
3348 * the wrong pointer address when calling arc_hdr_destroy() later.
3351 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3352 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3354 buf_discard_identity(hdr);
3355 kmem_cache_free(old, hdr);
3361 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3362 * The buf is returned thawed since we expect the consumer to modify it.
3365 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3367 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3368 ZIO_COMPRESS_OFF, type);
3369 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3371 arc_buf_t *buf = NULL;
3372 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3379 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3380 * for bufs containing metadata.
3383 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3384 enum zio_compress compression_type)
3386 ASSERT3U(lsize, >, 0);
3387 ASSERT3U(lsize, >=, psize);
3388 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3389 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3391 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3392 compression_type, ARC_BUFC_DATA);
3393 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3395 arc_buf_t *buf = NULL;
3396 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3398 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3400 if (!arc_buf_is_shared(buf)) {
3402 * To ensure that the hdr has the correct data in it if we call
3403 * arc_decompress() on this buf before it's been written to
3404 * disk, it's easiest if we just set up sharing between the
3407 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3408 arc_hdr_free_pabd(hdr);
3409 arc_share_buf(hdr, buf);
3416 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3418 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3419 l2arc_dev_t *dev = l2hdr->b_dev;
3420 uint64_t psize = arc_hdr_size(hdr);
3422 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3423 ASSERT(HDR_HAS_L2HDR(hdr));
3425 list_remove(&dev->l2ad_buflist, hdr);
3427 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3428 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3430 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3432 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3433 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3437 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3439 if (HDR_HAS_L1HDR(hdr)) {
3440 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3441 hdr->b_l1hdr.b_bufcnt > 0);
3442 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3443 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3445 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3446 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3448 if (!HDR_EMPTY(hdr))
3449 buf_discard_identity(hdr);
3451 if (HDR_HAS_L2HDR(hdr)) {
3452 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3453 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3456 mutex_enter(&dev->l2ad_mtx);
3459 * Even though we checked this conditional above, we
3460 * need to check this again now that we have the
3461 * l2ad_mtx. This is because we could be racing with
3462 * another thread calling l2arc_evict() which might have
3463 * destroyed this header's L2 portion as we were waiting
3464 * to acquire the l2ad_mtx. If that happens, we don't
3465 * want to re-destroy the header's L2 portion.
3467 if (HDR_HAS_L2HDR(hdr)) {
3469 arc_hdr_l2hdr_destroy(hdr);
3473 mutex_exit(&dev->l2ad_mtx);
3476 if (HDR_HAS_L1HDR(hdr)) {
3477 arc_cksum_free(hdr);
3479 while (hdr->b_l1hdr.b_buf != NULL)
3480 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3483 if (hdr->b_l1hdr.b_thawed != NULL) {
3484 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3485 hdr->b_l1hdr.b_thawed = NULL;
3489 if (hdr->b_l1hdr.b_pabd != NULL) {
3490 arc_hdr_free_pabd(hdr);
3494 ASSERT3P(hdr->b_hash_next, ==, NULL);
3495 if (HDR_HAS_L1HDR(hdr)) {
3496 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3497 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3498 kmem_cache_free(hdr_full_cache, hdr);
3500 kmem_cache_free(hdr_l2only_cache, hdr);
3505 arc_buf_destroy(arc_buf_t *buf, void* tag)
3507 arc_buf_hdr_t *hdr = buf->b_hdr;
3508 kmutex_t *hash_lock = HDR_LOCK(hdr);
3510 if (hdr->b_l1hdr.b_state == arc_anon) {
3511 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3512 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3513 VERIFY0(remove_reference(hdr, NULL, tag));
3514 arc_hdr_destroy(hdr);
3518 mutex_enter(hash_lock);
3519 ASSERT3P(hdr, ==, buf->b_hdr);
3520 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3521 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3522 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3523 ASSERT3P(buf->b_data, !=, NULL);
3525 (void) remove_reference(hdr, hash_lock, tag);
3526 arc_buf_destroy_impl(buf);
3527 mutex_exit(hash_lock);
3531 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3532 * state of the header is dependent on its state prior to entering this
3533 * function. The following transitions are possible:
3535 * - arc_mru -> arc_mru_ghost
3536 * - arc_mfu -> arc_mfu_ghost
3537 * - arc_mru_ghost -> arc_l2c_only
3538 * - arc_mru_ghost -> deleted
3539 * - arc_mfu_ghost -> arc_l2c_only
3540 * - arc_mfu_ghost -> deleted
3543 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3545 arc_state_t *evicted_state, *state;
3546 int64_t bytes_evicted = 0;
3548 ASSERT(MUTEX_HELD(hash_lock));
3549 ASSERT(HDR_HAS_L1HDR(hdr));
3551 state = hdr->b_l1hdr.b_state;
3552 if (GHOST_STATE(state)) {
3553 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3554 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3557 * l2arc_write_buffers() relies on a header's L1 portion
3558 * (i.e. its b_pabd field) during it's write phase.
3559 * Thus, we cannot push a header onto the arc_l2c_only
3560 * state (removing it's L1 piece) until the header is
3561 * done being written to the l2arc.
3563 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3564 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3565 return (bytes_evicted);
3568 ARCSTAT_BUMP(arcstat_deleted);
3569 bytes_evicted += HDR_GET_LSIZE(hdr);
3571 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3573 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3574 if (HDR_HAS_L2HDR(hdr)) {
3576 * This buffer is cached on the 2nd Level ARC;
3577 * don't destroy the header.
3579 arc_change_state(arc_l2c_only, hdr, hash_lock);
3581 * dropping from L1+L2 cached to L2-only,
3582 * realloc to remove the L1 header.
3584 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3587 arc_change_state(arc_anon, hdr, hash_lock);
3588 arc_hdr_destroy(hdr);
3590 return (bytes_evicted);
3593 ASSERT(state == arc_mru || state == arc_mfu);
3594 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3596 /* prefetch buffers have a minimum lifespan */
3597 if (HDR_IO_IN_PROGRESS(hdr) ||
3598 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3599 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3600 arc_min_prefetch_lifespan)) {
3601 ARCSTAT_BUMP(arcstat_evict_skip);
3602 return (bytes_evicted);
3605 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3606 while (hdr->b_l1hdr.b_buf) {
3607 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3608 if (!mutex_tryenter(&buf->b_evict_lock)) {
3609 ARCSTAT_BUMP(arcstat_mutex_miss);
3612 if (buf->b_data != NULL)
3613 bytes_evicted += HDR_GET_LSIZE(hdr);
3614 mutex_exit(&buf->b_evict_lock);
3615 arc_buf_destroy_impl(buf);
3618 if (HDR_HAS_L2HDR(hdr)) {
3619 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3621 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3622 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3623 HDR_GET_LSIZE(hdr));
3625 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3626 HDR_GET_LSIZE(hdr));
3630 if (hdr->b_l1hdr.b_bufcnt == 0) {
3631 arc_cksum_free(hdr);
3633 bytes_evicted += arc_hdr_size(hdr);
3636 * If this hdr is being evicted and has a compressed
3637 * buffer then we discard it here before we change states.
3638 * This ensures that the accounting is updated correctly
3639 * in arc_free_data_impl().
3641 arc_hdr_free_pabd(hdr);
3643 arc_change_state(evicted_state, hdr, hash_lock);
3644 ASSERT(HDR_IN_HASH_TABLE(hdr));
3645 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3646 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3649 return (bytes_evicted);
3653 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3654 uint64_t spa, int64_t bytes)
3656 multilist_sublist_t *mls;
3657 uint64_t bytes_evicted = 0;
3659 kmutex_t *hash_lock;
3660 int evict_count = 0;
3662 ASSERT3P(marker, !=, NULL);
3663 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3665 mls = multilist_sublist_lock(ml, idx);
3667 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3668 hdr = multilist_sublist_prev(mls, marker)) {
3669 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3670 (evict_count >= zfs_arc_evict_batch_limit))
3674 * To keep our iteration location, move the marker
3675 * forward. Since we're not holding hdr's hash lock, we
3676 * must be very careful and not remove 'hdr' from the
3677 * sublist. Otherwise, other consumers might mistake the
3678 * 'hdr' as not being on a sublist when they call the
3679 * multilist_link_active() function (they all rely on
3680 * the hash lock protecting concurrent insertions and
3681 * removals). multilist_sublist_move_forward() was
3682 * specifically implemented to ensure this is the case
3683 * (only 'marker' will be removed and re-inserted).
3685 multilist_sublist_move_forward(mls, marker);
3688 * The only case where the b_spa field should ever be
3689 * zero, is the marker headers inserted by
3690 * arc_evict_state(). It's possible for multiple threads
3691 * to be calling arc_evict_state() concurrently (e.g.
3692 * dsl_pool_close() and zio_inject_fault()), so we must
3693 * skip any markers we see from these other threads.
3695 if (hdr->b_spa == 0)
3698 /* we're only interested in evicting buffers of a certain spa */
3699 if (spa != 0 && hdr->b_spa != spa) {
3700 ARCSTAT_BUMP(arcstat_evict_skip);
3704 hash_lock = HDR_LOCK(hdr);
3707 * We aren't calling this function from any code path
3708 * that would already be holding a hash lock, so we're
3709 * asserting on this assumption to be defensive in case
3710 * this ever changes. Without this check, it would be
3711 * possible to incorrectly increment arcstat_mutex_miss
3712 * below (e.g. if the code changed such that we called
3713 * this function with a hash lock held).
3715 ASSERT(!MUTEX_HELD(hash_lock));
3717 if (mutex_tryenter(hash_lock)) {
3718 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3719 mutex_exit(hash_lock);
3721 bytes_evicted += evicted;
3724 * If evicted is zero, arc_evict_hdr() must have
3725 * decided to skip this header, don't increment
3726 * evict_count in this case.
3732 * If arc_size isn't overflowing, signal any
3733 * threads that might happen to be waiting.
3735 * For each header evicted, we wake up a single
3736 * thread. If we used cv_broadcast, we could
3737 * wake up "too many" threads causing arc_size
3738 * to significantly overflow arc_c; since
3739 * arc_get_data_impl() doesn't check for overflow
3740 * when it's woken up (it doesn't because it's
3741 * possible for the ARC to be overflowing while
3742 * full of un-evictable buffers, and the
3743 * function should proceed in this case).
3745 * If threads are left sleeping, due to not
3746 * using cv_broadcast, they will be woken up
3747 * just before arc_reclaim_thread() sleeps.
3749 mutex_enter(&arc_reclaim_lock);
3750 if (!arc_is_overflowing())
3751 cv_signal(&arc_reclaim_waiters_cv);
3752 mutex_exit(&arc_reclaim_lock);
3754 ARCSTAT_BUMP(arcstat_mutex_miss);
3758 multilist_sublist_unlock(mls);
3760 return (bytes_evicted);
3764 * Evict buffers from the given arc state, until we've removed the
3765 * specified number of bytes. Move the removed buffers to the
3766 * appropriate evict state.
3768 * This function makes a "best effort". It skips over any buffers
3769 * it can't get a hash_lock on, and so, may not catch all candidates.
3770 * It may also return without evicting as much space as requested.
3772 * If bytes is specified using the special value ARC_EVICT_ALL, this
3773 * will evict all available (i.e. unlocked and evictable) buffers from
3774 * the given arc state; which is used by arc_flush().
3777 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3778 arc_buf_contents_t type)
3780 uint64_t total_evicted = 0;
3781 multilist_t *ml = state->arcs_list[type];
3783 arc_buf_hdr_t **markers;
3785 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3787 num_sublists = multilist_get_num_sublists(ml);
3790 * If we've tried to evict from each sublist, made some
3791 * progress, but still have not hit the target number of bytes
3792 * to evict, we want to keep trying. The markers allow us to
3793 * pick up where we left off for each individual sublist, rather
3794 * than starting from the tail each time.
3796 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3797 for (int i = 0; i < num_sublists; i++) {
3798 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3801 * A b_spa of 0 is used to indicate that this header is
3802 * a marker. This fact is used in arc_adjust_type() and
3803 * arc_evict_state_impl().
3805 markers[i]->b_spa = 0;
3807 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3808 multilist_sublist_insert_tail(mls, markers[i]);
3809 multilist_sublist_unlock(mls);
3813 * While we haven't hit our target number of bytes to evict, or
3814 * we're evicting all available buffers.
3816 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3818 * Start eviction using a randomly selected sublist,
3819 * this is to try and evenly balance eviction across all
3820 * sublists. Always starting at the same sublist
3821 * (e.g. index 0) would cause evictions to favor certain
3822 * sublists over others.
3824 int sublist_idx = multilist_get_random_index(ml);
3825 uint64_t scan_evicted = 0;
3827 for (int i = 0; i < num_sublists; i++) {
3828 uint64_t bytes_remaining;
3829 uint64_t bytes_evicted;
3831 if (bytes == ARC_EVICT_ALL)
3832 bytes_remaining = ARC_EVICT_ALL;
3833 else if (total_evicted < bytes)
3834 bytes_remaining = bytes - total_evicted;
3838 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3839 markers[sublist_idx], spa, bytes_remaining);
3841 scan_evicted += bytes_evicted;
3842 total_evicted += bytes_evicted;
3844 /* we've reached the end, wrap to the beginning */
3845 if (++sublist_idx >= num_sublists)
3850 * If we didn't evict anything during this scan, we have
3851 * no reason to believe we'll evict more during another
3852 * scan, so break the loop.
3854 if (scan_evicted == 0) {
3855 /* This isn't possible, let's make that obvious */
3856 ASSERT3S(bytes, !=, 0);
3859 * When bytes is ARC_EVICT_ALL, the only way to
3860 * break the loop is when scan_evicted is zero.
3861 * In that case, we actually have evicted enough,
3862 * so we don't want to increment the kstat.
3864 if (bytes != ARC_EVICT_ALL) {
3865 ASSERT3S(total_evicted, <, bytes);
3866 ARCSTAT_BUMP(arcstat_evict_not_enough);
3873 for (int i = 0; i < num_sublists; i++) {
3874 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3875 multilist_sublist_remove(mls, markers[i]);
3876 multilist_sublist_unlock(mls);
3878 kmem_cache_free(hdr_full_cache, markers[i]);
3880 kmem_free(markers, sizeof (*markers) * num_sublists);
3882 return (total_evicted);
3886 * Flush all "evictable" data of the given type from the arc state
3887 * specified. This will not evict any "active" buffers (i.e. referenced).
3889 * When 'retry' is set to B_FALSE, the function will make a single pass
3890 * over the state and evict any buffers that it can. Since it doesn't
3891 * continually retry the eviction, it might end up leaving some buffers
3892 * in the ARC due to lock misses.
3894 * When 'retry' is set to B_TRUE, the function will continually retry the
3895 * eviction until *all* evictable buffers have been removed from the
3896 * state. As a result, if concurrent insertions into the state are
3897 * allowed (e.g. if the ARC isn't shutting down), this function might
3898 * wind up in an infinite loop, continually trying to evict buffers.
3901 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3904 uint64_t evicted = 0;
3906 while (refcount_count(&state->arcs_esize[type]) != 0) {
3907 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3917 * Evict the specified number of bytes from the state specified,
3918 * restricting eviction to the spa and type given. This function
3919 * prevents us from trying to evict more from a state's list than
3920 * is "evictable", and to skip evicting altogether when passed a
3921 * negative value for "bytes". In contrast, arc_evict_state() will
3922 * evict everything it can, when passed a negative value for "bytes".
3925 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3926 arc_buf_contents_t type)
3930 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3931 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3932 return (arc_evict_state(state, spa, delta, type));
3939 * Evict metadata buffers from the cache, such that arc_meta_used is
3940 * capped by the arc_meta_limit tunable.
3943 arc_adjust_meta(uint64_t meta_used)
3945 uint64_t total_evicted = 0;
3949 * If we're over the meta limit, we want to evict enough
3950 * metadata to get back under the meta limit. We don't want to
3951 * evict so much that we drop the MRU below arc_p, though. If
3952 * we're over the meta limit more than we're over arc_p, we
3953 * evict some from the MRU here, and some from the MFU below.
3955 target = MIN((int64_t)(meta_used - arc_meta_limit),
3956 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3957 refcount_count(&arc_mru->arcs_size) - arc_p));
3959 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3962 * Similar to the above, we want to evict enough bytes to get us
3963 * below the meta limit, but not so much as to drop us below the
3964 * space allotted to the MFU (which is defined as arc_c - arc_p).
3966 target = MIN((int64_t)(meta_used - arc_meta_limit),
3967 (int64_t)(refcount_count(&arc_mfu->arcs_size) -
3970 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3972 return (total_evicted);
3976 * Return the type of the oldest buffer in the given arc state
3978 * This function will select a random sublist of type ARC_BUFC_DATA and
3979 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3980 * is compared, and the type which contains the "older" buffer will be
3983 static arc_buf_contents_t
3984 arc_adjust_type(arc_state_t *state)
3986 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
3987 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
3988 int data_idx = multilist_get_random_index(data_ml);
3989 int meta_idx = multilist_get_random_index(meta_ml);
3990 multilist_sublist_t *data_mls;
3991 multilist_sublist_t *meta_mls;
3992 arc_buf_contents_t type;
3993 arc_buf_hdr_t *data_hdr;
3994 arc_buf_hdr_t *meta_hdr;
3997 * We keep the sublist lock until we're finished, to prevent
3998 * the headers from being destroyed via arc_evict_state().
4000 data_mls = multilist_sublist_lock(data_ml, data_idx);
4001 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4004 * These two loops are to ensure we skip any markers that
4005 * might be at the tail of the lists due to arc_evict_state().
4008 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4009 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4010 if (data_hdr->b_spa != 0)
4014 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4015 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4016 if (meta_hdr->b_spa != 0)
4020 if (data_hdr == NULL && meta_hdr == NULL) {
4021 type = ARC_BUFC_DATA;
4022 } else if (data_hdr == NULL) {
4023 ASSERT3P(meta_hdr, !=, NULL);
4024 type = ARC_BUFC_METADATA;
4025 } else if (meta_hdr == NULL) {
4026 ASSERT3P(data_hdr, !=, NULL);
4027 type = ARC_BUFC_DATA;
4029 ASSERT3P(data_hdr, !=, NULL);
4030 ASSERT3P(meta_hdr, !=, NULL);
4032 /* The headers can't be on the sublist without an L1 header */
4033 ASSERT(HDR_HAS_L1HDR(data_hdr));
4034 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4036 if (data_hdr->b_l1hdr.b_arc_access <
4037 meta_hdr->b_l1hdr.b_arc_access) {
4038 type = ARC_BUFC_DATA;
4040 type = ARC_BUFC_METADATA;
4044 multilist_sublist_unlock(meta_mls);
4045 multilist_sublist_unlock(data_mls);
4051 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4056 uint64_t total_evicted = 0;
4059 uint64_t asize = aggsum_value(&arc_size);
4060 uint64_t ameta = aggsum_value(&arc_meta_used);
4063 * If we're over arc_meta_limit, we want to correct that before
4064 * potentially evicting data buffers below.
4066 total_evicted += arc_adjust_meta(ameta);
4071 * If we're over the target cache size, we want to evict enough
4072 * from the list to get back to our target size. We don't want
4073 * to evict too much from the MRU, such that it drops below
4074 * arc_p. So, if we're over our target cache size more than
4075 * the MRU is over arc_p, we'll evict enough to get back to
4076 * arc_p here, and then evict more from the MFU below.
4078 target = MIN((int64_t)(asize - arc_c),
4079 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4080 refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4083 * If we're below arc_meta_min, always prefer to evict data.
4084 * Otherwise, try to satisfy the requested number of bytes to
4085 * evict from the type which contains older buffers; in an
4086 * effort to keep newer buffers in the cache regardless of their
4087 * type. If we cannot satisfy the number of bytes from this
4088 * type, spill over into the next type.
4090 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4091 ameta > arc_meta_min) {
4092 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4093 total_evicted += bytes;
4096 * If we couldn't evict our target number of bytes from
4097 * metadata, we try to get the rest from data.
4102 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4104 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4105 total_evicted += bytes;
4108 * If we couldn't evict our target number of bytes from
4109 * data, we try to get the rest from metadata.
4114 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4120 * Now that we've tried to evict enough from the MRU to get its
4121 * size back to arc_p, if we're still above the target cache
4122 * size, we evict the rest from the MFU.
4124 target = asize - arc_c;
4126 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4127 ameta > arc_meta_min) {
4128 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4129 total_evicted += bytes;
4132 * If we couldn't evict our target number of bytes from
4133 * metadata, we try to get the rest from data.
4138 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4140 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4141 total_evicted += bytes;
4144 * If we couldn't evict our target number of bytes from
4145 * data, we try to get the rest from data.
4150 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4154 * Adjust ghost lists
4156 * In addition to the above, the ARC also defines target values
4157 * for the ghost lists. The sum of the mru list and mru ghost
4158 * list should never exceed the target size of the cache, and
4159 * the sum of the mru list, mfu list, mru ghost list, and mfu
4160 * ghost list should never exceed twice the target size of the
4161 * cache. The following logic enforces these limits on the ghost
4162 * caches, and evicts from them as needed.
4164 target = refcount_count(&arc_mru->arcs_size) +
4165 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4167 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4168 total_evicted += bytes;
4173 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4176 * We assume the sum of the mru list and mfu list is less than
4177 * or equal to arc_c (we enforced this above), which means we
4178 * can use the simpler of the two equations below:
4180 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4181 * mru ghost + mfu ghost <= arc_c
4183 target = refcount_count(&arc_mru_ghost->arcs_size) +
4184 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4186 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4187 total_evicted += bytes;
4192 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4194 return (total_evicted);
4198 arc_flush(spa_t *spa, boolean_t retry)
4203 * If retry is B_TRUE, a spa must not be specified since we have
4204 * no good way to determine if all of a spa's buffers have been
4205 * evicted from an arc state.
4207 ASSERT(!retry || spa == 0);
4210 guid = spa_load_guid(spa);
4212 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4213 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4215 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4216 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4218 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4219 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4221 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4222 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4226 arc_shrink(int64_t to_free)
4228 uint64_t asize = aggsum_value(&arc_size);
4229 if (arc_c > arc_c_min) {
4230 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
4231 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
4232 if (arc_c > arc_c_min + to_free)
4233 atomic_add_64(&arc_c, -to_free);
4237 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4239 arc_c = MAX(asize, arc_c_min);
4241 arc_p = (arc_c >> 1);
4243 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
4246 ASSERT(arc_c >= arc_c_min);
4247 ASSERT((int64_t)arc_p >= 0);
4250 if (asize > arc_c) {
4251 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, asize,
4253 (void) arc_adjust();
4257 typedef enum free_memory_reason_t {
4262 FMR_PAGES_PP_MAXIMUM,
4265 } free_memory_reason_t;
4267 int64_t last_free_memory;
4268 free_memory_reason_t last_free_reason;
4271 * Additional reserve of pages for pp_reserve.
4273 int64_t arc_pages_pp_reserve = 64;
4276 * Additional reserve of pages for swapfs.
4278 int64_t arc_swapfs_reserve = 64;
4281 * Return the amount of memory that can be consumed before reclaim will be
4282 * needed. Positive if there is sufficient free memory, negative indicates
4283 * the amount of memory that needs to be freed up.
4286 arc_available_memory(void)
4288 int64_t lowest = INT64_MAX;
4290 free_memory_reason_t r = FMR_UNKNOWN;
4295 * Cooperate with pagedaemon when it's time for it to scan
4296 * and reclaim some pages.
4298 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
4306 n = PAGESIZE * (-needfree);
4314 * check that we're out of range of the pageout scanner. It starts to
4315 * schedule paging if freemem is less than lotsfree and needfree.
4316 * lotsfree is the high-water mark for pageout, and needfree is the
4317 * number of needed free pages. We add extra pages here to make sure
4318 * the scanner doesn't start up while we're freeing memory.
4320 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4327 * check to make sure that swapfs has enough space so that anon
4328 * reservations can still succeed. anon_resvmem() checks that the
4329 * availrmem is greater than swapfs_minfree, and the number of reserved
4330 * swap pages. We also add a bit of extra here just to prevent
4331 * circumstances from getting really dire.
4333 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4334 desfree - arc_swapfs_reserve);
4337 r = FMR_SWAPFS_MINFREE;
4342 * Check that we have enough availrmem that memory locking (e.g., via
4343 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4344 * stores the number of pages that cannot be locked; when availrmem
4345 * drops below pages_pp_maximum, page locking mechanisms such as
4346 * page_pp_lock() will fail.)
4348 n = PAGESIZE * (availrmem - pages_pp_maximum -
4349 arc_pages_pp_reserve);
4352 r = FMR_PAGES_PP_MAXIMUM;
4355 #endif /* __FreeBSD__ */
4356 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4358 * If we're on an i386 platform, it's possible that we'll exhaust the
4359 * kernel heap space before we ever run out of available physical
4360 * memory. Most checks of the size of the heap_area compare against
4361 * tune.t_minarmem, which is the minimum available real memory that we
4362 * can have in the system. However, this is generally fixed at 25 pages
4363 * which is so low that it's useless. In this comparison, we seek to
4364 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4365 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4368 n = uma_avail() - (long)(uma_limit() / 4);
4376 * If zio data pages are being allocated out of a separate heap segment,
4377 * then enforce that the size of available vmem for this arena remains
4378 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4380 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4381 * memory (in the zio_arena) free, which can avoid memory
4382 * fragmentation issues.
4384 if (zio_arena != NULL) {
4385 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4386 (vmem_size(zio_arena, VMEM_ALLOC) >>
4387 arc_zio_arena_free_shift);
4395 /* Every 100 calls, free a small amount */
4396 if (spa_get_random(100) == 0)
4398 #endif /* _KERNEL */
4400 last_free_memory = lowest;
4401 last_free_reason = r;
4402 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4408 * Determine if the system is under memory pressure and is asking
4409 * to reclaim memory. A return value of B_TRUE indicates that the system
4410 * is under memory pressure and that the arc should adjust accordingly.
4413 arc_reclaim_needed(void)
4415 return (arc_available_memory() < 0);
4418 extern kmem_cache_t *zio_buf_cache[];
4419 extern kmem_cache_t *zio_data_buf_cache[];
4420 extern kmem_cache_t *range_seg_cache;
4421 extern kmem_cache_t *abd_chunk_cache;
4423 static __noinline void
4424 arc_kmem_reap_now(void)
4427 kmem_cache_t *prev_cache = NULL;
4428 kmem_cache_t *prev_data_cache = NULL;
4430 DTRACE_PROBE(arc__kmem_reap_start);
4432 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4434 * We are exceeding our meta-data cache limit.
4435 * Purge some DNLC entries to release holds on meta-data.
4437 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4441 * Reclaim unused memory from all kmem caches.
4448 * If a kmem reap is already active, don't schedule more. We must
4449 * check for this because kmem_cache_reap_soon() won't actually
4450 * block on the cache being reaped (this is to prevent callers from
4451 * becoming implicitly blocked by a system-wide kmem reap -- which,
4452 * on a system with many, many full magazines, can take minutes).
4454 if (kmem_cache_reap_active())
4457 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4458 if (zio_buf_cache[i] != prev_cache) {
4459 prev_cache = zio_buf_cache[i];
4460 kmem_cache_reap_soon(zio_buf_cache[i]);
4462 if (zio_data_buf_cache[i] != prev_data_cache) {
4463 prev_data_cache = zio_data_buf_cache[i];
4464 kmem_cache_reap_soon(zio_data_buf_cache[i]);
4467 kmem_cache_reap_soon(abd_chunk_cache);
4468 kmem_cache_reap_soon(buf_cache);
4469 kmem_cache_reap_soon(hdr_full_cache);
4470 kmem_cache_reap_soon(hdr_l2only_cache);
4471 kmem_cache_reap_soon(range_seg_cache);
4474 if (zio_arena != NULL) {
4476 * Ask the vmem arena to reclaim unused memory from its
4479 vmem_qcache_reap(zio_arena);
4482 DTRACE_PROBE(arc__kmem_reap_end);
4486 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4487 * enough data and signal them to proceed. When this happens, the threads in
4488 * arc_get_data_impl() are sleeping while holding the hash lock for their
4489 * particular arc header. Thus, we must be careful to never sleep on a
4490 * hash lock in this thread. This is to prevent the following deadlock:
4492 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4493 * waiting for the reclaim thread to signal it.
4495 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4496 * fails, and goes to sleep forever.
4498 * This possible deadlock is avoided by always acquiring a hash lock
4499 * using mutex_tryenter() from arc_reclaim_thread().
4503 arc_reclaim_thread(void *unused __unused)
4505 hrtime_t growtime = 0;
4506 hrtime_t kmem_reap_time = 0;
4509 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4511 mutex_enter(&arc_reclaim_lock);
4512 while (!arc_reclaim_thread_exit) {
4513 uint64_t evicted = 0;
4516 * This is necessary in order for the mdb ::arc dcmd to
4517 * show up to date information. Since the ::arc command
4518 * does not call the kstat's update function, without
4519 * this call, the command may show stale stats for the
4520 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4521 * with this change, the data might be up to 1 second
4522 * out of date; but that should suffice. The arc_state_t
4523 * structures can be queried directly if more accurate
4524 * information is needed.
4526 if (arc_ksp != NULL)
4527 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4529 mutex_exit(&arc_reclaim_lock);
4532 * We call arc_adjust() before (possibly) calling
4533 * arc_kmem_reap_now(), so that we can wake up
4534 * arc_get_data_impl() sooner.
4536 evicted = arc_adjust();
4538 int64_t free_memory = arc_available_memory();
4539 if (free_memory < 0) {
4540 hrtime_t curtime = gethrtime();
4541 arc_no_grow = B_TRUE;
4545 * Wait at least zfs_grow_retry (default 60) seconds
4546 * before considering growing.
4548 growtime = curtime + SEC2NSEC(arc_grow_retry);
4551 * Wait at least arc_kmem_cache_reap_retry_ms
4552 * between arc_kmem_reap_now() calls. Without
4553 * this check it is possible to end up in a
4554 * situation where we spend lots of time
4555 * reaping caches, while we're near arc_c_min.
4557 if (curtime >= kmem_reap_time) {
4558 arc_kmem_reap_now();
4559 kmem_reap_time = gethrtime() +
4560 MSEC2NSEC(arc_kmem_cache_reap_retry_ms);
4564 * If we are still low on memory, shrink the ARC
4565 * so that we have arc_shrink_min free space.
4567 free_memory = arc_available_memory();
4570 (arc_c >> arc_shrink_shift) - free_memory;
4574 to_free = MAX(to_free, ptob(needfree));
4577 arc_shrink(to_free);
4579 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4580 arc_no_grow = B_TRUE;
4581 } else if (gethrtime() >= growtime) {
4582 arc_no_grow = B_FALSE;
4585 mutex_enter(&arc_reclaim_lock);
4588 * If evicted is zero, we couldn't evict anything via
4589 * arc_adjust(). This could be due to hash lock
4590 * collisions, but more likely due to the majority of
4591 * arc buffers being unevictable. Therefore, even if
4592 * arc_size is above arc_c, another pass is unlikely to
4593 * be helpful and could potentially cause us to enter an
4596 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) {
4598 * We're either no longer overflowing, or we
4599 * can't evict anything more, so we should wake
4600 * up any threads before we go to sleep.
4602 cv_broadcast(&arc_reclaim_waiters_cv);
4605 * Block until signaled, or after one second (we
4606 * might need to perform arc_kmem_reap_now()
4607 * even if we aren't being signalled)
4609 CALLB_CPR_SAFE_BEGIN(&cpr);
4610 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4611 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4612 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4616 arc_reclaim_thread_exit = B_FALSE;
4617 cv_broadcast(&arc_reclaim_thread_cv);
4618 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4622 static u_int arc_dnlc_evicts_arg;
4623 extern struct vfsops zfs_vfsops;
4626 arc_dnlc_evicts_thread(void *dummy __unused)
4631 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4633 mutex_enter(&arc_dnlc_evicts_lock);
4634 while (!arc_dnlc_evicts_thread_exit) {
4635 CALLB_CPR_SAFE_BEGIN(&cpr);
4636 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4637 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4638 if (arc_dnlc_evicts_arg != 0) {
4639 percent = arc_dnlc_evicts_arg;
4640 mutex_exit(&arc_dnlc_evicts_lock);
4642 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4644 mutex_enter(&arc_dnlc_evicts_lock);
4646 * Clear our token only after vnlru_free()
4647 * pass is done, to avoid false queueing of
4650 arc_dnlc_evicts_arg = 0;
4653 arc_dnlc_evicts_thread_exit = FALSE;
4654 cv_broadcast(&arc_dnlc_evicts_cv);
4655 CALLB_CPR_EXIT(&cpr);
4660 dnlc_reduce_cache(void *arg)
4664 percent = (u_int)(uintptr_t)arg;
4665 mutex_enter(&arc_dnlc_evicts_lock);
4666 if (arc_dnlc_evicts_arg == 0) {
4667 arc_dnlc_evicts_arg = percent;
4668 cv_broadcast(&arc_dnlc_evicts_cv);
4670 mutex_exit(&arc_dnlc_evicts_lock);
4674 * Adapt arc info given the number of bytes we are trying to add and
4675 * the state that we are comming from. This function is only called
4676 * when we are adding new content to the cache.
4679 arc_adapt(int bytes, arc_state_t *state)
4682 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4683 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4684 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4686 if (state == arc_l2c_only)
4691 * Adapt the target size of the MRU list:
4692 * - if we just hit in the MRU ghost list, then increase
4693 * the target size of the MRU list.
4694 * - if we just hit in the MFU ghost list, then increase
4695 * the target size of the MFU list by decreasing the
4696 * target size of the MRU list.
4698 if (state == arc_mru_ghost) {
4699 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4700 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4702 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4703 } else if (state == arc_mfu_ghost) {
4706 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4707 mult = MIN(mult, 10);
4709 delta = MIN(bytes * mult, arc_p);
4710 arc_p = MAX(arc_p_min, arc_p - delta);
4712 ASSERT((int64_t)arc_p >= 0);
4714 if (arc_reclaim_needed()) {
4715 cv_signal(&arc_reclaim_thread_cv);
4722 if (arc_c >= arc_c_max)
4726 * If we're within (2 * maxblocksize) bytes of the target
4727 * cache size, increment the target cache size
4729 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
4731 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
4732 atomic_add_64(&arc_c, (int64_t)bytes);
4733 if (arc_c > arc_c_max)
4735 else if (state == arc_anon)
4736 atomic_add_64(&arc_p, (int64_t)bytes);
4740 ASSERT((int64_t)arc_p >= 0);
4744 * Check if arc_size has grown past our upper threshold, determined by
4745 * zfs_arc_overflow_shift.
4748 arc_is_overflowing(void)
4750 /* Always allow at least one block of overflow */
4751 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4752 arc_c >> zfs_arc_overflow_shift);
4755 * We just compare the lower bound here for performance reasons. Our
4756 * primary goals are to make sure that the arc never grows without
4757 * bound, and that it can reach its maximum size. This check
4758 * accomplishes both goals. The maximum amount we could run over by is
4759 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4760 * in the ARC. In practice, that's in the tens of MB, which is low
4761 * enough to be safe.
4763 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
4767 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4769 arc_buf_contents_t type = arc_buf_type(hdr);
4771 arc_get_data_impl(hdr, size, tag);
4772 if (type == ARC_BUFC_METADATA) {
4773 return (abd_alloc(size, B_TRUE));
4775 ASSERT(type == ARC_BUFC_DATA);
4776 return (abd_alloc(size, B_FALSE));
4781 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4783 arc_buf_contents_t type = arc_buf_type(hdr);
4785 arc_get_data_impl(hdr, size, tag);
4786 if (type == ARC_BUFC_METADATA) {
4787 return (zio_buf_alloc(size));
4789 ASSERT(type == ARC_BUFC_DATA);
4790 return (zio_data_buf_alloc(size));
4795 * Allocate a block and return it to the caller. If we are hitting the
4796 * hard limit for the cache size, we must sleep, waiting for the eviction
4797 * thread to catch up. If we're past the target size but below the hard
4798 * limit, we'll only signal the reclaim thread and continue on.
4801 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4803 arc_state_t *state = hdr->b_l1hdr.b_state;
4804 arc_buf_contents_t type = arc_buf_type(hdr);
4806 arc_adapt(size, state);
4809 * If arc_size is currently overflowing, and has grown past our
4810 * upper limit, we must be adding data faster than the evict
4811 * thread can evict. Thus, to ensure we don't compound the
4812 * problem by adding more data and forcing arc_size to grow even
4813 * further past it's target size, we halt and wait for the
4814 * eviction thread to catch up.
4816 * It's also possible that the reclaim thread is unable to evict
4817 * enough buffers to get arc_size below the overflow limit (e.g.
4818 * due to buffers being un-evictable, or hash lock collisions).
4819 * In this case, we want to proceed regardless if we're
4820 * overflowing; thus we don't use a while loop here.
4822 if (arc_is_overflowing()) {
4823 mutex_enter(&arc_reclaim_lock);
4826 * Now that we've acquired the lock, we may no longer be
4827 * over the overflow limit, lets check.
4829 * We're ignoring the case of spurious wake ups. If that
4830 * were to happen, it'd let this thread consume an ARC
4831 * buffer before it should have (i.e. before we're under
4832 * the overflow limit and were signalled by the reclaim
4833 * thread). As long as that is a rare occurrence, it
4834 * shouldn't cause any harm.
4836 if (arc_is_overflowing()) {
4837 cv_signal(&arc_reclaim_thread_cv);
4838 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4841 mutex_exit(&arc_reclaim_lock);
4844 VERIFY3U(hdr->b_type, ==, type);
4845 if (type == ARC_BUFC_METADATA) {
4846 arc_space_consume(size, ARC_SPACE_META);
4848 arc_space_consume(size, ARC_SPACE_DATA);
4852 * Update the state size. Note that ghost states have a
4853 * "ghost size" and so don't need to be updated.
4855 if (!GHOST_STATE(state)) {
4857 (void) refcount_add_many(&state->arcs_size, size, tag);
4860 * If this is reached via arc_read, the link is
4861 * protected by the hash lock. If reached via
4862 * arc_buf_alloc, the header should not be accessed by
4863 * any other thread. And, if reached via arc_read_done,
4864 * the hash lock will protect it if it's found in the
4865 * hash table; otherwise no other thread should be
4866 * trying to [add|remove]_reference it.
4868 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4869 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4870 (void) refcount_add_many(&state->arcs_esize[type],
4875 * If we are growing the cache, and we are adding anonymous
4876 * data, and we have outgrown arc_p, update arc_p
4878 if (aggsum_compare(&arc_size, arc_c) < 0 &&
4879 hdr->b_l1hdr.b_state == arc_anon &&
4880 (refcount_count(&arc_anon->arcs_size) +
4881 refcount_count(&arc_mru->arcs_size) > arc_p))
4882 arc_p = MIN(arc_c, arc_p + size);
4884 ARCSTAT_BUMP(arcstat_allocated);
4888 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4890 arc_free_data_impl(hdr, size, tag);
4895 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4897 arc_buf_contents_t type = arc_buf_type(hdr);
4899 arc_free_data_impl(hdr, size, tag);
4900 if (type == ARC_BUFC_METADATA) {
4901 zio_buf_free(buf, size);
4903 ASSERT(type == ARC_BUFC_DATA);
4904 zio_data_buf_free(buf, size);
4909 * Free the arc data buffer.
4912 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4914 arc_state_t *state = hdr->b_l1hdr.b_state;
4915 arc_buf_contents_t type = arc_buf_type(hdr);
4917 /* protected by hash lock, if in the hash table */
4918 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4919 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4920 ASSERT(state != arc_anon && state != arc_l2c_only);
4922 (void) refcount_remove_many(&state->arcs_esize[type],
4925 (void) refcount_remove_many(&state->arcs_size, size, tag);
4927 VERIFY3U(hdr->b_type, ==, type);
4928 if (type == ARC_BUFC_METADATA) {
4929 arc_space_return(size, ARC_SPACE_META);
4931 ASSERT(type == ARC_BUFC_DATA);
4932 arc_space_return(size, ARC_SPACE_DATA);
4937 * This routine is called whenever a buffer is accessed.
4938 * NOTE: the hash lock is dropped in this function.
4941 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4945 ASSERT(MUTEX_HELD(hash_lock));
4946 ASSERT(HDR_HAS_L1HDR(hdr));
4948 if (hdr->b_l1hdr.b_state == arc_anon) {
4950 * This buffer is not in the cache, and does not
4951 * appear in our "ghost" list. Add the new buffer
4955 ASSERT0(hdr->b_l1hdr.b_arc_access);
4956 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4957 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4958 arc_change_state(arc_mru, hdr, hash_lock);
4960 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4961 now = ddi_get_lbolt();
4964 * If this buffer is here because of a prefetch, then either:
4965 * - clear the flag if this is a "referencing" read
4966 * (any subsequent access will bump this into the MFU state).
4968 * - move the buffer to the head of the list if this is
4969 * another prefetch (to make it less likely to be evicted).
4971 if (HDR_PREFETCH(hdr)) {
4972 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4973 /* link protected by hash lock */
4974 ASSERT(multilist_link_active(
4975 &hdr->b_l1hdr.b_arc_node));
4977 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4978 ARCSTAT_BUMP(arcstat_mru_hits);
4980 hdr->b_l1hdr.b_arc_access = now;
4985 * This buffer has been "accessed" only once so far,
4986 * but it is still in the cache. Move it to the MFU
4989 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4991 * More than 125ms have passed since we
4992 * instantiated this buffer. Move it to the
4993 * most frequently used state.
4995 hdr->b_l1hdr.b_arc_access = now;
4996 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4997 arc_change_state(arc_mfu, hdr, hash_lock);
4999 ARCSTAT_BUMP(arcstat_mru_hits);
5000 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5001 arc_state_t *new_state;
5003 * This buffer has been "accessed" recently, but
5004 * was evicted from the cache. Move it to the
5008 if (HDR_PREFETCH(hdr)) {
5009 new_state = arc_mru;
5010 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
5011 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
5012 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5014 new_state = arc_mfu;
5015 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5018 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5019 arc_change_state(new_state, hdr, hash_lock);
5021 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5022 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5024 * This buffer has been accessed more than once and is
5025 * still in the cache. Keep it in the MFU state.
5027 * NOTE: an add_reference() that occurred when we did
5028 * the arc_read() will have kicked this off the list.
5029 * If it was a prefetch, we will explicitly move it to
5030 * the head of the list now.
5032 if ((HDR_PREFETCH(hdr)) != 0) {
5033 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5034 /* link protected by hash_lock */
5035 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5037 ARCSTAT_BUMP(arcstat_mfu_hits);
5038 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5039 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5040 arc_state_t *new_state = arc_mfu;
5042 * This buffer has been accessed more than once but has
5043 * been evicted from the cache. Move it back to the
5047 if (HDR_PREFETCH(hdr)) {
5049 * This is a prefetch access...
5050 * move this block back to the MRU state.
5052 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
5053 new_state = arc_mru;
5056 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5057 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5058 arc_change_state(new_state, hdr, hash_lock);
5060 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5061 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5063 * This buffer is on the 2nd Level ARC.
5066 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5067 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5068 arc_change_state(arc_mfu, hdr, hash_lock);
5070 ASSERT(!"invalid arc state");
5075 * This routine is called by dbuf_hold() to update the arc_access() state
5076 * which otherwise would be skipped for entries in the dbuf cache.
5079 arc_buf_access(arc_buf_t *buf)
5081 mutex_enter(&buf->b_evict_lock);
5082 arc_buf_hdr_t *hdr = buf->b_hdr;
5085 * Avoid taking the hash_lock when possible as an optimization.
5086 * The header must be checked again under the hash_lock in order
5087 * to handle the case where it is concurrently being released.
5089 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5090 mutex_exit(&buf->b_evict_lock);
5091 ARCSTAT_BUMP(arcstat_access_skip);
5095 kmutex_t *hash_lock = HDR_LOCK(hdr);
5096 mutex_enter(hash_lock);
5098 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5099 mutex_exit(hash_lock);
5100 mutex_exit(&buf->b_evict_lock);
5101 ARCSTAT_BUMP(arcstat_access_skip);
5105 mutex_exit(&buf->b_evict_lock);
5107 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5108 hdr->b_l1hdr.b_state == arc_mfu);
5110 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5111 arc_access(hdr, hash_lock);
5112 mutex_exit(hash_lock);
5114 ARCSTAT_BUMP(arcstat_hits);
5115 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5116 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5119 /* a generic arc_done_func_t which you can use */
5122 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
5124 if (zio == NULL || zio->io_error == 0)
5125 bcopy(buf->b_data, arg, arc_buf_size(buf));
5126 arc_buf_destroy(buf, arg);
5129 /* a generic arc_done_func_t */
5131 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
5133 arc_buf_t **bufp = arg;
5134 if (zio && zio->io_error) {
5135 arc_buf_destroy(buf, arg);
5139 ASSERT(buf->b_data);
5144 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5146 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5147 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5148 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
5150 if (HDR_COMPRESSION_ENABLED(hdr)) {
5151 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
5152 BP_GET_COMPRESS(bp));
5154 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5155 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5160 arc_read_done(zio_t *zio)
5162 arc_buf_hdr_t *hdr = zio->io_private;
5163 kmutex_t *hash_lock = NULL;
5164 arc_callback_t *callback_list;
5165 arc_callback_t *acb;
5166 boolean_t freeable = B_FALSE;
5167 boolean_t no_zio_error = (zio->io_error == 0);
5170 * The hdr was inserted into hash-table and removed from lists
5171 * prior to starting I/O. We should find this header, since
5172 * it's in the hash table, and it should be legit since it's
5173 * not possible to evict it during the I/O. The only possible
5174 * reason for it not to be found is if we were freed during the
5177 if (HDR_IN_HASH_TABLE(hdr)) {
5178 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5179 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5180 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5181 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5182 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5184 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5187 ASSERT((found == hdr &&
5188 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5189 (found == hdr && HDR_L2_READING(hdr)));
5190 ASSERT3P(hash_lock, !=, NULL);
5194 /* byteswap if necessary */
5195 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5196 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5197 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5199 hdr->b_l1hdr.b_byteswap =
5200 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5203 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5207 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5208 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5209 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5211 callback_list = hdr->b_l1hdr.b_acb;
5212 ASSERT3P(callback_list, !=, NULL);
5214 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
5216 * Only call arc_access on anonymous buffers. This is because
5217 * if we've issued an I/O for an evicted buffer, we've already
5218 * called arc_access (to prevent any simultaneous readers from
5219 * getting confused).
5221 arc_access(hdr, hash_lock);
5225 * If a read request has a callback (i.e. acb_done is not NULL), then we
5226 * make a buf containing the data according to the parameters which were
5227 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5228 * aren't needlessly decompressing the data multiple times.
5230 int callback_cnt = 0;
5231 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5235 /* This is a demand read since prefetches don't use callbacks */
5238 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
5239 acb->acb_compressed, no_zio_error, &acb->acb_buf);
5241 zio->io_error = error;
5244 hdr->b_l1hdr.b_acb = NULL;
5245 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5246 if (callback_cnt == 0) {
5247 ASSERT(HDR_PREFETCH(hdr));
5248 ASSERT0(hdr->b_l1hdr.b_bufcnt);
5249 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5252 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5253 callback_list != NULL);
5256 arc_hdr_verify(hdr, zio->io_bp);
5258 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5259 if (hdr->b_l1hdr.b_state != arc_anon)
5260 arc_change_state(arc_anon, hdr, hash_lock);
5261 if (HDR_IN_HASH_TABLE(hdr))
5262 buf_hash_remove(hdr);
5263 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5267 * Broadcast before we drop the hash_lock to avoid the possibility
5268 * that the hdr (and hence the cv) might be freed before we get to
5269 * the cv_broadcast().
5271 cv_broadcast(&hdr->b_l1hdr.b_cv);
5273 if (hash_lock != NULL) {
5274 mutex_exit(hash_lock);
5277 * This block was freed while we waited for the read to
5278 * complete. It has been removed from the hash table and
5279 * moved to the anonymous state (so that it won't show up
5282 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5283 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5286 /* execute each callback and free its structure */
5287 while ((acb = callback_list) != NULL) {
5289 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
5291 if (acb->acb_zio_dummy != NULL) {
5292 acb->acb_zio_dummy->io_error = zio->io_error;
5293 zio_nowait(acb->acb_zio_dummy);
5296 callback_list = acb->acb_next;
5297 kmem_free(acb, sizeof (arc_callback_t));
5301 arc_hdr_destroy(hdr);
5305 * "Read" the block at the specified DVA (in bp) via the
5306 * cache. If the block is found in the cache, invoke the provided
5307 * callback immediately and return. Note that the `zio' parameter
5308 * in the callback will be NULL in this case, since no IO was
5309 * required. If the block is not in the cache pass the read request
5310 * on to the spa with a substitute callback function, so that the
5311 * requested block will be added to the cache.
5313 * If a read request arrives for a block that has a read in-progress,
5314 * either wait for the in-progress read to complete (and return the
5315 * results); or, if this is a read with a "done" func, add a record
5316 * to the read to invoke the "done" func when the read completes,
5317 * and return; or just return.
5319 * arc_read_done() will invoke all the requested "done" functions
5320 * for readers of this block.
5323 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
5324 void *private, zio_priority_t priority, int zio_flags,
5325 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5327 arc_buf_hdr_t *hdr = NULL;
5328 kmutex_t *hash_lock = NULL;
5330 uint64_t guid = spa_load_guid(spa);
5331 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5333 ASSERT(!BP_IS_EMBEDDED(bp) ||
5334 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5337 if (!BP_IS_EMBEDDED(bp)) {
5339 * Embedded BP's have no DVA and require no I/O to "read".
5340 * Create an anonymous arc buf to back it.
5342 hdr = buf_hash_find(guid, bp, &hash_lock);
5345 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5346 arc_buf_t *buf = NULL;
5347 *arc_flags |= ARC_FLAG_CACHED;
5349 if (HDR_IO_IN_PROGRESS(hdr)) {
5351 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5352 priority == ZIO_PRIORITY_SYNC_READ) {
5354 * This sync read must wait for an
5355 * in-progress async read (e.g. a predictive
5356 * prefetch). Async reads are queued
5357 * separately at the vdev_queue layer, so
5358 * this is a form of priority inversion.
5359 * Ideally, we would "inherit" the demand
5360 * i/o's priority by moving the i/o from
5361 * the async queue to the synchronous queue,
5362 * but there is currently no mechanism to do
5363 * so. Track this so that we can evaluate
5364 * the magnitude of this potential performance
5367 * Note that if the prefetch i/o is already
5368 * active (has been issued to the device),
5369 * the prefetch improved performance, because
5370 * we issued it sooner than we would have
5371 * without the prefetch.
5373 DTRACE_PROBE1(arc__sync__wait__for__async,
5374 arc_buf_hdr_t *, hdr);
5375 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
5377 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5378 arc_hdr_clear_flags(hdr,
5379 ARC_FLAG_PREDICTIVE_PREFETCH);
5382 if (*arc_flags & ARC_FLAG_WAIT) {
5383 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5384 mutex_exit(hash_lock);
5387 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5390 arc_callback_t *acb = NULL;
5392 acb = kmem_zalloc(sizeof (arc_callback_t),
5394 acb->acb_done = done;
5395 acb->acb_private = private;
5396 acb->acb_compressed = compressed_read;
5398 acb->acb_zio_dummy = zio_null(pio,
5399 spa, NULL, NULL, NULL, zio_flags);
5401 ASSERT3P(acb->acb_done, !=, NULL);
5402 acb->acb_next = hdr->b_l1hdr.b_acb;
5403 hdr->b_l1hdr.b_acb = acb;
5404 mutex_exit(hash_lock);
5407 mutex_exit(hash_lock);
5411 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5412 hdr->b_l1hdr.b_state == arc_mfu);
5415 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5417 * This is a demand read which does not have to
5418 * wait for i/o because we did a predictive
5419 * prefetch i/o for it, which has completed.
5422 arc__demand__hit__predictive__prefetch,
5423 arc_buf_hdr_t *, hdr);
5425 arcstat_demand_hit_predictive_prefetch);
5426 arc_hdr_clear_flags(hdr,
5427 ARC_FLAG_PREDICTIVE_PREFETCH);
5429 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5431 /* Get a buf with the desired data in it. */
5432 VERIFY0(arc_buf_alloc_impl(hdr, private,
5433 compressed_read, B_TRUE, &buf));
5434 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5435 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5436 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5438 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5439 arc_access(hdr, hash_lock);
5440 if (*arc_flags & ARC_FLAG_L2CACHE)
5441 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5442 mutex_exit(hash_lock);
5443 ARCSTAT_BUMP(arcstat_hits);
5444 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5445 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5446 data, metadata, hits);
5449 done(NULL, buf, private);
5451 uint64_t lsize = BP_GET_LSIZE(bp);
5452 uint64_t psize = BP_GET_PSIZE(bp);
5453 arc_callback_t *acb;
5456 boolean_t devw = B_FALSE;
5460 /* this block is not in the cache */
5461 arc_buf_hdr_t *exists = NULL;
5462 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5463 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5464 BP_GET_COMPRESS(bp), type);
5466 if (!BP_IS_EMBEDDED(bp)) {
5467 hdr->b_dva = *BP_IDENTITY(bp);
5468 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5469 exists = buf_hash_insert(hdr, &hash_lock);
5471 if (exists != NULL) {
5472 /* somebody beat us to the hash insert */
5473 mutex_exit(hash_lock);
5474 buf_discard_identity(hdr);
5475 arc_hdr_destroy(hdr);
5476 goto top; /* restart the IO request */
5480 * This block is in the ghost cache. If it was L2-only
5481 * (and thus didn't have an L1 hdr), we realloc the
5482 * header to add an L1 hdr.
5484 if (!HDR_HAS_L1HDR(hdr)) {
5485 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5488 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5489 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5490 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5491 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5492 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5493 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5496 * This is a delicate dance that we play here.
5497 * This hdr is in the ghost list so we access it
5498 * to move it out of the ghost list before we
5499 * initiate the read. If it's a prefetch then
5500 * it won't have a callback so we'll remove the
5501 * reference that arc_buf_alloc_impl() created. We
5502 * do this after we've called arc_access() to
5503 * avoid hitting an assert in remove_reference().
5505 arc_access(hdr, hash_lock);
5506 arc_hdr_alloc_pabd(hdr);
5508 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5509 size = arc_hdr_size(hdr);
5512 * If compression is enabled on the hdr, then will do
5513 * RAW I/O and will store the compressed data in the hdr's
5514 * data block. Otherwise, the hdr's data block will contain
5515 * the uncompressed data.
5517 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5518 zio_flags |= ZIO_FLAG_RAW;
5521 if (*arc_flags & ARC_FLAG_PREFETCH)
5522 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5523 if (*arc_flags & ARC_FLAG_L2CACHE)
5524 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5525 if (BP_GET_LEVEL(bp) > 0)
5526 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5527 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5528 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5529 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5531 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5532 acb->acb_done = done;
5533 acb->acb_private = private;
5534 acb->acb_compressed = compressed_read;
5536 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5537 hdr->b_l1hdr.b_acb = acb;
5538 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5540 if (HDR_HAS_L2HDR(hdr) &&
5541 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5542 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5543 addr = hdr->b_l2hdr.b_daddr;
5545 * Lock out L2ARC device removal.
5547 if (vdev_is_dead(vd) ||
5548 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5552 if (priority == ZIO_PRIORITY_ASYNC_READ)
5553 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5555 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5557 if (hash_lock != NULL)
5558 mutex_exit(hash_lock);
5561 * At this point, we have a level 1 cache miss. Try again in
5562 * L2ARC if possible.
5564 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5566 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5567 uint64_t, lsize, zbookmark_phys_t *, zb);
5568 ARCSTAT_BUMP(arcstat_misses);
5569 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5570 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5571 data, metadata, misses);
5576 racct_add_force(curproc, RACCT_READBPS, size);
5577 racct_add_force(curproc, RACCT_READIOPS, 1);
5578 PROC_UNLOCK(curproc);
5581 curthread->td_ru.ru_inblock++;
5584 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5586 * Read from the L2ARC if the following are true:
5587 * 1. The L2ARC vdev was previously cached.
5588 * 2. This buffer still has L2ARC metadata.
5589 * 3. This buffer isn't currently writing to the L2ARC.
5590 * 4. The L2ARC entry wasn't evicted, which may
5591 * also have invalidated the vdev.
5592 * 5. This isn't prefetch and l2arc_noprefetch is set.
5594 if (HDR_HAS_L2HDR(hdr) &&
5595 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5596 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5597 l2arc_read_callback_t *cb;
5601 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5602 ARCSTAT_BUMP(arcstat_l2_hits);
5604 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5606 cb->l2rcb_hdr = hdr;
5609 cb->l2rcb_flags = zio_flags;
5611 asize = vdev_psize_to_asize(vd, size);
5612 if (asize != size) {
5613 abd = abd_alloc_for_io(asize,
5614 HDR_ISTYPE_METADATA(hdr));
5615 cb->l2rcb_abd = abd;
5617 abd = hdr->b_l1hdr.b_pabd;
5620 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5621 addr + asize <= vd->vdev_psize -
5622 VDEV_LABEL_END_SIZE);
5625 * l2arc read. The SCL_L2ARC lock will be
5626 * released by l2arc_read_done().
5627 * Issue a null zio if the underlying buffer
5628 * was squashed to zero size by compression.
5630 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5631 ZIO_COMPRESS_EMPTY);
5632 rzio = zio_read_phys(pio, vd, addr,
5635 l2arc_read_done, cb, priority,
5636 zio_flags | ZIO_FLAG_DONT_CACHE |
5638 ZIO_FLAG_DONT_PROPAGATE |
5639 ZIO_FLAG_DONT_RETRY, B_FALSE);
5640 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5642 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5644 if (*arc_flags & ARC_FLAG_NOWAIT) {
5649 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5650 if (zio_wait(rzio) == 0)
5653 /* l2arc read error; goto zio_read() */
5655 DTRACE_PROBE1(l2arc__miss,
5656 arc_buf_hdr_t *, hdr);
5657 ARCSTAT_BUMP(arcstat_l2_misses);
5658 if (HDR_L2_WRITING(hdr))
5659 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5660 spa_config_exit(spa, SCL_L2ARC, vd);
5664 spa_config_exit(spa, SCL_L2ARC, vd);
5665 if (l2arc_ndev != 0) {
5666 DTRACE_PROBE1(l2arc__miss,
5667 arc_buf_hdr_t *, hdr);
5668 ARCSTAT_BUMP(arcstat_l2_misses);
5672 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
5673 arc_read_done, hdr, priority, zio_flags, zb);
5675 if (*arc_flags & ARC_FLAG_WAIT)
5676 return (zio_wait(rzio));
5678 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5685 * Notify the arc that a block was freed, and thus will never be used again.
5688 arc_freed(spa_t *spa, const blkptr_t *bp)
5691 kmutex_t *hash_lock;
5692 uint64_t guid = spa_load_guid(spa);
5694 ASSERT(!BP_IS_EMBEDDED(bp));
5696 hdr = buf_hash_find(guid, bp, &hash_lock);
5701 * We might be trying to free a block that is still doing I/O
5702 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5703 * dmu_sync-ed block). If this block is being prefetched, then it
5704 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5705 * until the I/O completes. A block may also have a reference if it is
5706 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5707 * have written the new block to its final resting place on disk but
5708 * without the dedup flag set. This would have left the hdr in the MRU
5709 * state and discoverable. When the txg finally syncs it detects that
5710 * the block was overridden in open context and issues an override I/O.
5711 * Since this is a dedup block, the override I/O will determine if the
5712 * block is already in the DDT. If so, then it will replace the io_bp
5713 * with the bp from the DDT and allow the I/O to finish. When the I/O
5714 * reaches the done callback, dbuf_write_override_done, it will
5715 * check to see if the io_bp and io_bp_override are identical.
5716 * If they are not, then it indicates that the bp was replaced with
5717 * the bp in the DDT and the override bp is freed. This allows
5718 * us to arrive here with a reference on a block that is being
5719 * freed. So if we have an I/O in progress, or a reference to
5720 * this hdr, then we don't destroy the hdr.
5722 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5723 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5724 arc_change_state(arc_anon, hdr, hash_lock);
5725 arc_hdr_destroy(hdr);
5726 mutex_exit(hash_lock);
5728 mutex_exit(hash_lock);
5734 * Release this buffer from the cache, making it an anonymous buffer. This
5735 * must be done after a read and prior to modifying the buffer contents.
5736 * If the buffer has more than one reference, we must make
5737 * a new hdr for the buffer.
5740 arc_release(arc_buf_t *buf, void *tag)
5742 arc_buf_hdr_t *hdr = buf->b_hdr;
5745 * It would be nice to assert that if it's DMU metadata (level >
5746 * 0 || it's the dnode file), then it must be syncing context.
5747 * But we don't know that information at this level.
5750 mutex_enter(&buf->b_evict_lock);
5752 ASSERT(HDR_HAS_L1HDR(hdr));
5755 * We don't grab the hash lock prior to this check, because if
5756 * the buffer's header is in the arc_anon state, it won't be
5757 * linked into the hash table.
5759 if (hdr->b_l1hdr.b_state == arc_anon) {
5760 mutex_exit(&buf->b_evict_lock);
5761 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5762 ASSERT(!HDR_IN_HASH_TABLE(hdr));
5763 ASSERT(!HDR_HAS_L2HDR(hdr));
5764 ASSERT(HDR_EMPTY(hdr));
5765 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5766 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5767 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5769 hdr->b_l1hdr.b_arc_access = 0;
5772 * If the buf is being overridden then it may already
5773 * have a hdr that is not empty.
5775 buf_discard_identity(hdr);
5781 kmutex_t *hash_lock = HDR_LOCK(hdr);
5782 mutex_enter(hash_lock);
5785 * This assignment is only valid as long as the hash_lock is
5786 * held, we must be careful not to reference state or the
5787 * b_state field after dropping the lock.
5789 arc_state_t *state = hdr->b_l1hdr.b_state;
5790 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5791 ASSERT3P(state, !=, arc_anon);
5793 /* this buffer is not on any list */
5794 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
5796 if (HDR_HAS_L2HDR(hdr)) {
5797 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5800 * We have to recheck this conditional again now that
5801 * we're holding the l2ad_mtx to prevent a race with
5802 * another thread which might be concurrently calling
5803 * l2arc_evict(). In that case, l2arc_evict() might have
5804 * destroyed the header's L2 portion as we were waiting
5805 * to acquire the l2ad_mtx.
5807 if (HDR_HAS_L2HDR(hdr)) {
5809 arc_hdr_l2hdr_destroy(hdr);
5812 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5816 * Do we have more than one buf?
5818 if (hdr->b_l1hdr.b_bufcnt > 1) {
5819 arc_buf_hdr_t *nhdr;
5820 uint64_t spa = hdr->b_spa;
5821 uint64_t psize = HDR_GET_PSIZE(hdr);
5822 uint64_t lsize = HDR_GET_LSIZE(hdr);
5823 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5824 arc_buf_contents_t type = arc_buf_type(hdr);
5825 VERIFY3U(hdr->b_type, ==, type);
5827 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5828 (void) remove_reference(hdr, hash_lock, tag);
5830 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
5831 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5832 ASSERT(ARC_BUF_LAST(buf));
5836 * Pull the data off of this hdr and attach it to
5837 * a new anonymous hdr. Also find the last buffer
5838 * in the hdr's buffer list.
5840 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
5841 ASSERT3P(lastbuf, !=, NULL);
5844 * If the current arc_buf_t and the hdr are sharing their data
5845 * buffer, then we must stop sharing that block.
5847 if (arc_buf_is_shared(buf)) {
5848 VERIFY(!arc_buf_is_shared(lastbuf));
5851 * First, sever the block sharing relationship between
5852 * buf and the arc_buf_hdr_t.
5854 arc_unshare_buf(hdr, buf);
5857 * Now we need to recreate the hdr's b_pabd. Since we
5858 * have lastbuf handy, we try to share with it, but if
5859 * we can't then we allocate a new b_pabd and copy the
5860 * data from buf into it.
5862 if (arc_can_share(hdr, lastbuf)) {
5863 arc_share_buf(hdr, lastbuf);
5865 arc_hdr_alloc_pabd(hdr);
5866 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
5867 buf->b_data, psize);
5869 VERIFY3P(lastbuf->b_data, !=, NULL);
5870 } else if (HDR_SHARED_DATA(hdr)) {
5872 * Uncompressed shared buffers are always at the end
5873 * of the list. Compressed buffers don't have the
5874 * same requirements. This makes it hard to
5875 * simply assert that the lastbuf is shared so
5876 * we rely on the hdr's compression flags to determine
5877 * if we have a compressed, shared buffer.
5879 ASSERT(arc_buf_is_shared(lastbuf) ||
5880 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
5881 ASSERT(!ARC_BUF_SHARED(buf));
5883 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5884 ASSERT3P(state, !=, arc_l2c_only);
5886 (void) refcount_remove_many(&state->arcs_size,
5887 arc_buf_size(buf), buf);
5889 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5890 ASSERT3P(state, !=, arc_l2c_only);
5891 (void) refcount_remove_many(&state->arcs_esize[type],
5892 arc_buf_size(buf), buf);
5895 hdr->b_l1hdr.b_bufcnt -= 1;
5896 arc_cksum_verify(buf);
5898 arc_buf_unwatch(buf);
5901 mutex_exit(hash_lock);
5904 * Allocate a new hdr. The new hdr will contain a b_pabd
5905 * buffer which will be freed in arc_write().
5907 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5908 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5909 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5910 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5911 VERIFY3U(nhdr->b_type, ==, type);
5912 ASSERT(!HDR_SHARED_DATA(nhdr));
5914 nhdr->b_l1hdr.b_buf = buf;
5915 nhdr->b_l1hdr.b_bufcnt = 1;
5916 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5919 mutex_exit(&buf->b_evict_lock);
5920 (void) refcount_add_many(&arc_anon->arcs_size,
5921 arc_buf_size(buf), buf);
5923 mutex_exit(&buf->b_evict_lock);
5924 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5925 /* protected by hash lock, or hdr is on arc_anon */
5926 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5927 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5928 arc_change_state(arc_anon, hdr, hash_lock);
5929 hdr->b_l1hdr.b_arc_access = 0;
5930 mutex_exit(hash_lock);
5932 buf_discard_identity(hdr);
5938 arc_released(arc_buf_t *buf)
5942 mutex_enter(&buf->b_evict_lock);
5943 released = (buf->b_data != NULL &&
5944 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5945 mutex_exit(&buf->b_evict_lock);
5951 arc_referenced(arc_buf_t *buf)
5955 mutex_enter(&buf->b_evict_lock);
5956 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5957 mutex_exit(&buf->b_evict_lock);
5958 return (referenced);
5963 arc_write_ready(zio_t *zio)
5965 arc_write_callback_t *callback = zio->io_private;
5966 arc_buf_t *buf = callback->awcb_buf;
5967 arc_buf_hdr_t *hdr = buf->b_hdr;
5968 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5970 ASSERT(HDR_HAS_L1HDR(hdr));
5971 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5972 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5975 * If we're reexecuting this zio because the pool suspended, then
5976 * cleanup any state that was previously set the first time the
5977 * callback was invoked.
5979 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5980 arc_cksum_free(hdr);
5982 arc_buf_unwatch(buf);
5984 if (hdr->b_l1hdr.b_pabd != NULL) {
5985 if (arc_buf_is_shared(buf)) {
5986 arc_unshare_buf(hdr, buf);
5988 arc_hdr_free_pabd(hdr);
5992 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5993 ASSERT(!HDR_SHARED_DATA(hdr));
5994 ASSERT(!arc_buf_is_shared(buf));
5996 callback->awcb_ready(zio, buf, callback->awcb_private);
5998 if (HDR_IO_IN_PROGRESS(hdr))
5999 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6001 arc_cksum_compute(buf);
6002 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6004 enum zio_compress compress;
6005 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6006 compress = ZIO_COMPRESS_OFF;
6008 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
6009 compress = BP_GET_COMPRESS(zio->io_bp);
6011 HDR_SET_PSIZE(hdr, psize);
6012 arc_hdr_set_compress(hdr, compress);
6016 * Fill the hdr with data. If the hdr is compressed, the data we want
6017 * is available from the zio, otherwise we can take it from the buf.
6019 * We might be able to share the buf's data with the hdr here. However,
6020 * doing so would cause the ARC to be full of linear ABDs if we write a
6021 * lot of shareable data. As a compromise, we check whether scattered
6022 * ABDs are allowed, and assume that if they are then the user wants
6023 * the ARC to be primarily filled with them regardless of the data being
6024 * written. Therefore, if they're allowed then we allocate one and copy
6025 * the data into it; otherwise, we share the data directly if we can.
6027 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6028 arc_hdr_alloc_pabd(hdr);
6031 * Ideally, we would always copy the io_abd into b_pabd, but the
6032 * user may have disabled compressed ARC, thus we must check the
6033 * hdr's compression setting rather than the io_bp's.
6035 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
6036 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
6038 ASSERT3U(psize, >, 0);
6040 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6042 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6044 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6048 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6049 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6050 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6052 arc_share_buf(hdr, buf);
6055 arc_hdr_verify(hdr, zio->io_bp);
6059 arc_write_children_ready(zio_t *zio)
6061 arc_write_callback_t *callback = zio->io_private;
6062 arc_buf_t *buf = callback->awcb_buf;
6064 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6068 * The SPA calls this callback for each physical write that happens on behalf
6069 * of a logical write. See the comment in dbuf_write_physdone() for details.
6072 arc_write_physdone(zio_t *zio)
6074 arc_write_callback_t *cb = zio->io_private;
6075 if (cb->awcb_physdone != NULL)
6076 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6080 arc_write_done(zio_t *zio)
6082 arc_write_callback_t *callback = zio->io_private;
6083 arc_buf_t *buf = callback->awcb_buf;
6084 arc_buf_hdr_t *hdr = buf->b_hdr;
6086 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6088 if (zio->io_error == 0) {
6089 arc_hdr_verify(hdr, zio->io_bp);
6091 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6092 buf_discard_identity(hdr);
6094 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6095 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6098 ASSERT(HDR_EMPTY(hdr));
6102 * If the block to be written was all-zero or compressed enough to be
6103 * embedded in the BP, no write was performed so there will be no
6104 * dva/birth/checksum. The buffer must therefore remain anonymous
6107 if (!HDR_EMPTY(hdr)) {
6108 arc_buf_hdr_t *exists;
6109 kmutex_t *hash_lock;
6111 ASSERT3U(zio->io_error, ==, 0);
6113 arc_cksum_verify(buf);
6115 exists = buf_hash_insert(hdr, &hash_lock);
6116 if (exists != NULL) {
6118 * This can only happen if we overwrite for
6119 * sync-to-convergence, because we remove
6120 * buffers from the hash table when we arc_free().
6122 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6123 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6124 panic("bad overwrite, hdr=%p exists=%p",
6125 (void *)hdr, (void *)exists);
6126 ASSERT(refcount_is_zero(
6127 &exists->b_l1hdr.b_refcnt));
6128 arc_change_state(arc_anon, exists, hash_lock);
6129 mutex_exit(hash_lock);
6130 arc_hdr_destroy(exists);
6131 exists = buf_hash_insert(hdr, &hash_lock);
6132 ASSERT3P(exists, ==, NULL);
6133 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6135 ASSERT(zio->io_prop.zp_nopwrite);
6136 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6137 panic("bad nopwrite, hdr=%p exists=%p",
6138 (void *)hdr, (void *)exists);
6141 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6142 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6143 ASSERT(BP_GET_DEDUP(zio->io_bp));
6144 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6147 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6148 /* if it's not anon, we are doing a scrub */
6149 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6150 arc_access(hdr, hash_lock);
6151 mutex_exit(hash_lock);
6153 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6156 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6157 callback->awcb_done(zio, buf, callback->awcb_private);
6159 abd_put(zio->io_abd);
6160 kmem_free(callback, sizeof (arc_write_callback_t));
6164 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6165 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
6166 arc_done_func_t *children_ready, arc_done_func_t *physdone,
6167 arc_done_func_t *done, void *private, zio_priority_t priority,
6168 int zio_flags, const zbookmark_phys_t *zb)
6170 arc_buf_hdr_t *hdr = buf->b_hdr;
6171 arc_write_callback_t *callback;
6173 zio_prop_t localprop = *zp;
6175 ASSERT3P(ready, !=, NULL);
6176 ASSERT3P(done, !=, NULL);
6177 ASSERT(!HDR_IO_ERROR(hdr));
6178 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6179 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6180 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6182 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6183 if (ARC_BUF_COMPRESSED(buf)) {
6185 * We're writing a pre-compressed buffer. Make the
6186 * compression algorithm requested by the zio_prop_t match
6187 * the pre-compressed buffer's compression algorithm.
6189 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6191 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6192 zio_flags |= ZIO_FLAG_RAW;
6194 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6195 callback->awcb_ready = ready;
6196 callback->awcb_children_ready = children_ready;
6197 callback->awcb_physdone = physdone;
6198 callback->awcb_done = done;
6199 callback->awcb_private = private;
6200 callback->awcb_buf = buf;
6203 * The hdr's b_pabd is now stale, free it now. A new data block
6204 * will be allocated when the zio pipeline calls arc_write_ready().
6206 if (hdr->b_l1hdr.b_pabd != NULL) {
6208 * If the buf is currently sharing the data block with
6209 * the hdr then we need to break that relationship here.
6210 * The hdr will remain with a NULL data pointer and the
6211 * buf will take sole ownership of the block.
6213 if (arc_buf_is_shared(buf)) {
6214 arc_unshare_buf(hdr, buf);
6216 arc_hdr_free_pabd(hdr);
6218 VERIFY3P(buf->b_data, !=, NULL);
6219 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6221 ASSERT(!arc_buf_is_shared(buf));
6222 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6224 zio = zio_write(pio, spa, txg, bp,
6225 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6226 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6227 (children_ready != NULL) ? arc_write_children_ready : NULL,
6228 arc_write_physdone, arc_write_done, callback,
6229 priority, zio_flags, zb);
6235 arc_memory_throttle(uint64_t reserve, uint64_t txg)
6238 uint64_t available_memory = ptob(freemem);
6239 static uint64_t page_load = 0;
6240 static uint64_t last_txg = 0;
6242 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
6243 available_memory = MIN(available_memory, uma_avail());
6246 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
6249 if (txg > last_txg) {
6254 * If we are in pageout, we know that memory is already tight,
6255 * the arc is already going to be evicting, so we just want to
6256 * continue to let page writes occur as quickly as possible.
6258 if (curproc == pageproc) {
6259 if (page_load > MAX(ptob(minfree), available_memory) / 4)
6260 return (SET_ERROR(ERESTART));
6261 /* Note: reserve is inflated, so we deflate */
6262 page_load += reserve / 8;
6264 } else if (page_load > 0 && arc_reclaim_needed()) {
6265 /* memory is low, delay before restarting */
6266 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6267 return (SET_ERROR(EAGAIN));
6275 arc_tempreserve_clear(uint64_t reserve)
6277 atomic_add_64(&arc_tempreserve, -reserve);
6278 ASSERT((int64_t)arc_tempreserve >= 0);
6282 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
6287 if (reserve > arc_c/4 && !arc_no_grow) {
6288 arc_c = MIN(arc_c_max, reserve * 4);
6289 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
6291 if (reserve > arc_c)
6292 return (SET_ERROR(ENOMEM));
6295 * Don't count loaned bufs as in flight dirty data to prevent long
6296 * network delays from blocking transactions that are ready to be
6297 * assigned to a txg.
6300 /* assert that it has not wrapped around */
6301 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6303 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6304 arc_loaned_bytes), 0);
6307 * Writes will, almost always, require additional memory allocations
6308 * in order to compress/encrypt/etc the data. We therefore need to
6309 * make sure that there is sufficient available memory for this.
6311 error = arc_memory_throttle(reserve, txg);
6316 * Throttle writes when the amount of dirty data in the cache
6317 * gets too large. We try to keep the cache less than half full
6318 * of dirty blocks so that our sync times don't grow too large.
6319 * Note: if two requests come in concurrently, we might let them
6320 * both succeed, when one of them should fail. Not a huge deal.
6323 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
6324 anon_size > arc_c / 4) {
6325 uint64_t meta_esize =
6326 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6327 uint64_t data_esize =
6328 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6329 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6330 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6331 arc_tempreserve >> 10, meta_esize >> 10,
6332 data_esize >> 10, reserve >> 10, arc_c >> 10);
6333 return (SET_ERROR(ERESTART));
6335 atomic_add_64(&arc_tempreserve, reserve);
6340 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6341 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6343 size->value.ui64 = refcount_count(&state->arcs_size);
6344 evict_data->value.ui64 =
6345 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6346 evict_metadata->value.ui64 =
6347 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6351 arc_kstat_update(kstat_t *ksp, int rw)
6353 arc_stats_t *as = ksp->ks_data;
6355 if (rw == KSTAT_WRITE) {
6358 arc_kstat_update_state(arc_anon,
6359 &as->arcstat_anon_size,
6360 &as->arcstat_anon_evictable_data,
6361 &as->arcstat_anon_evictable_metadata);
6362 arc_kstat_update_state(arc_mru,
6363 &as->arcstat_mru_size,
6364 &as->arcstat_mru_evictable_data,
6365 &as->arcstat_mru_evictable_metadata);
6366 arc_kstat_update_state(arc_mru_ghost,
6367 &as->arcstat_mru_ghost_size,
6368 &as->arcstat_mru_ghost_evictable_data,
6369 &as->arcstat_mru_ghost_evictable_metadata);
6370 arc_kstat_update_state(arc_mfu,
6371 &as->arcstat_mfu_size,
6372 &as->arcstat_mfu_evictable_data,
6373 &as->arcstat_mfu_evictable_metadata);
6374 arc_kstat_update_state(arc_mfu_ghost,
6375 &as->arcstat_mfu_ghost_size,
6376 &as->arcstat_mfu_ghost_evictable_data,
6377 &as->arcstat_mfu_ghost_evictable_metadata);
6379 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
6380 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
6381 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
6382 ARCSTAT(arcstat_metadata_size) =
6383 aggsum_value(&astat_metadata_size);
6384 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
6385 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size);
6386 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
6393 * This function *must* return indices evenly distributed between all
6394 * sublists of the multilist. This is needed due to how the ARC eviction
6395 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6396 * distributed between all sublists and uses this assumption when
6397 * deciding which sublist to evict from and how much to evict from it.
6400 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6402 arc_buf_hdr_t *hdr = obj;
6405 * We rely on b_dva to generate evenly distributed index
6406 * numbers using buf_hash below. So, as an added precaution,
6407 * let's make sure we never add empty buffers to the arc lists.
6409 ASSERT(!HDR_EMPTY(hdr));
6412 * The assumption here, is the hash value for a given
6413 * arc_buf_hdr_t will remain constant throughout it's lifetime
6414 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6415 * Thus, we don't need to store the header's sublist index
6416 * on insertion, as this index can be recalculated on removal.
6418 * Also, the low order bits of the hash value are thought to be
6419 * distributed evenly. Otherwise, in the case that the multilist
6420 * has a power of two number of sublists, each sublists' usage
6421 * would not be evenly distributed.
6423 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6424 multilist_get_num_sublists(ml));
6428 static eventhandler_tag arc_event_lowmem = NULL;
6431 arc_lowmem(void *arg __unused, int howto __unused)
6434 mutex_enter(&arc_reclaim_lock);
6435 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE);
6436 cv_signal(&arc_reclaim_thread_cv);
6439 * It is unsafe to block here in arbitrary threads, because we can come
6440 * here from ARC itself and may hold ARC locks and thus risk a deadlock
6441 * with ARC reclaim thread.
6443 if (curproc == pageproc)
6444 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
6445 mutex_exit(&arc_reclaim_lock);
6450 arc_state_init(void)
6452 arc_anon = &ARC_anon;
6454 arc_mru_ghost = &ARC_mru_ghost;
6456 arc_mfu_ghost = &ARC_mfu_ghost;
6457 arc_l2c_only = &ARC_l2c_only;
6459 arc_mru->arcs_list[ARC_BUFC_METADATA] =
6460 multilist_create(sizeof (arc_buf_hdr_t),
6461 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6462 arc_state_multilist_index_func);
6463 arc_mru->arcs_list[ARC_BUFC_DATA] =
6464 multilist_create(sizeof (arc_buf_hdr_t),
6465 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6466 arc_state_multilist_index_func);
6467 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6468 multilist_create(sizeof (arc_buf_hdr_t),
6469 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6470 arc_state_multilist_index_func);
6471 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6472 multilist_create(sizeof (arc_buf_hdr_t),
6473 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6474 arc_state_multilist_index_func);
6475 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6476 multilist_create(sizeof (arc_buf_hdr_t),
6477 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6478 arc_state_multilist_index_func);
6479 arc_mfu->arcs_list[ARC_BUFC_DATA] =
6480 multilist_create(sizeof (arc_buf_hdr_t),
6481 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6482 arc_state_multilist_index_func);
6483 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6484 multilist_create(sizeof (arc_buf_hdr_t),
6485 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6486 arc_state_multilist_index_func);
6487 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6488 multilist_create(sizeof (arc_buf_hdr_t),
6489 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6490 arc_state_multilist_index_func);
6491 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6492 multilist_create(sizeof (arc_buf_hdr_t),
6493 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6494 arc_state_multilist_index_func);
6495 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6496 multilist_create(sizeof (arc_buf_hdr_t),
6497 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6498 arc_state_multilist_index_func);
6500 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6501 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6502 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6503 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6504 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6505 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6506 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6507 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6508 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6509 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6510 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6511 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6513 refcount_create(&arc_anon->arcs_size);
6514 refcount_create(&arc_mru->arcs_size);
6515 refcount_create(&arc_mru_ghost->arcs_size);
6516 refcount_create(&arc_mfu->arcs_size);
6517 refcount_create(&arc_mfu_ghost->arcs_size);
6518 refcount_create(&arc_l2c_only->arcs_size);
6520 aggsum_init(&arc_meta_used, 0);
6521 aggsum_init(&arc_size, 0);
6522 aggsum_init(&astat_data_size, 0);
6523 aggsum_init(&astat_metadata_size, 0);
6524 aggsum_init(&astat_hdr_size, 0);
6525 aggsum_init(&astat_other_size, 0);
6526 aggsum_init(&astat_l2_hdr_size, 0);
6530 arc_state_fini(void)
6532 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6533 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6534 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6535 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6536 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6537 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6538 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6539 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6540 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6541 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6542 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6543 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6545 refcount_destroy(&arc_anon->arcs_size);
6546 refcount_destroy(&arc_mru->arcs_size);
6547 refcount_destroy(&arc_mru_ghost->arcs_size);
6548 refcount_destroy(&arc_mfu->arcs_size);
6549 refcount_destroy(&arc_mfu_ghost->arcs_size);
6550 refcount_destroy(&arc_l2c_only->arcs_size);
6552 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6553 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6554 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6555 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6556 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6557 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6558 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6559 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6571 int i, prefetch_tunable_set = 0;
6574 * allmem is "all memory that we could possibly use".
6578 uint64_t allmem = ptob(physmem - swapfs_minfree);
6580 uint64_t allmem = (physmem * PAGESIZE) / 2;
6583 uint64_t allmem = kmem_size();
6587 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
6588 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
6589 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
6591 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
6592 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
6594 /* Convert seconds to clock ticks */
6595 arc_min_prefetch_lifespan = 1 * hz;
6597 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
6598 arc_c_min = MAX(allmem / 32, arc_abs_min);
6599 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */
6600 if (allmem >= 1 << 30)
6601 arc_c_max = allmem - (1 << 30);
6603 arc_c_max = arc_c_min;
6604 arc_c_max = MAX(allmem * 5 / 8, arc_c_max);
6607 * In userland, there's only the memory pressure that we artificially
6608 * create (see arc_available_memory()). Don't let arc_c get too
6609 * small, because it can cause transactions to be larger than
6610 * arc_c, causing arc_tempreserve_space() to fail.
6613 arc_c_min = arc_c_max / 2;
6618 * Allow the tunables to override our calculations if they are
6621 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) {
6622 arc_c_max = zfs_arc_max;
6623 arc_c_min = MIN(arc_c_min, arc_c_max);
6625 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
6626 arc_c_min = zfs_arc_min;
6630 arc_p = (arc_c >> 1);
6632 /* limit meta-data to 1/4 of the arc capacity */
6633 arc_meta_limit = arc_c_max / 4;
6637 * Metadata is stored in the kernel's heap. Don't let us
6638 * use more than half the heap for the ARC.
6641 arc_meta_limit = MIN(arc_meta_limit, uma_limit() / 2);
6643 arc_meta_limit = MIN(arc_meta_limit,
6644 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6648 /* Allow the tunable to override if it is reasonable */
6649 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6650 arc_meta_limit = zfs_arc_meta_limit;
6652 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6653 arc_c_min = arc_meta_limit / 2;
6655 if (zfs_arc_meta_min > 0) {
6656 arc_meta_min = zfs_arc_meta_min;
6658 arc_meta_min = arc_c_min / 2;
6661 if (zfs_arc_grow_retry > 0)
6662 arc_grow_retry = zfs_arc_grow_retry;
6664 if (zfs_arc_shrink_shift > 0)
6665 arc_shrink_shift = zfs_arc_shrink_shift;
6667 if (zfs_arc_no_grow_shift > 0)
6668 arc_no_grow_shift = zfs_arc_no_grow_shift;
6670 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6672 if (arc_no_grow_shift >= arc_shrink_shift)
6673 arc_no_grow_shift = arc_shrink_shift - 1;
6675 if (zfs_arc_p_min_shift > 0)
6676 arc_p_min_shift = zfs_arc_p_min_shift;
6678 /* if kmem_flags are set, lets try to use less memory */
6679 if (kmem_debugging())
6681 if (arc_c < arc_c_min)
6684 zfs_arc_min = arc_c_min;
6685 zfs_arc_max = arc_c_max;
6690 arc_reclaim_thread_exit = B_FALSE;
6691 arc_dnlc_evicts_thread_exit = FALSE;
6693 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6694 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6696 if (arc_ksp != NULL) {
6697 arc_ksp->ks_data = &arc_stats;
6698 arc_ksp->ks_update = arc_kstat_update;
6699 kstat_install(arc_ksp);
6702 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6703 TS_RUN, minclsyspri);
6706 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
6707 EVENTHANDLER_PRI_FIRST);
6710 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
6711 TS_RUN, minclsyspri);
6717 * Calculate maximum amount of dirty data per pool.
6719 * If it has been set by /etc/system, take that.
6720 * Otherwise, use a percentage of physical memory defined by
6721 * zfs_dirty_data_max_percent (default 10%) with a cap at
6722 * zfs_dirty_data_max_max (default 4GB).
6724 if (zfs_dirty_data_max == 0) {
6725 zfs_dirty_data_max = ptob(physmem) *
6726 zfs_dirty_data_max_percent / 100;
6727 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6728 zfs_dirty_data_max_max);
6732 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
6733 prefetch_tunable_set = 1;
6736 if (prefetch_tunable_set == 0) {
6737 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
6739 printf(" add \"vfs.zfs.prefetch_disable=0\" "
6740 "to /boot/loader.conf.\n");
6741 zfs_prefetch_disable = 1;
6744 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
6745 prefetch_tunable_set == 0) {
6746 printf("ZFS NOTICE: Prefetch is disabled by default if less "
6747 "than 4GB of RAM is present;\n"
6748 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
6749 "to /boot/loader.conf.\n");
6750 zfs_prefetch_disable = 1;
6753 /* Warn about ZFS memory and address space requirements. */
6754 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
6755 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
6756 "expect unstable behavior.\n");
6758 if (allmem < 512 * (1 << 20)) {
6759 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
6760 "expect unstable behavior.\n");
6761 printf(" Consider tuning vm.kmem_size and "
6762 "vm.kmem_size_max\n");
6763 printf(" in /boot/loader.conf.\n");
6772 if (arc_event_lowmem != NULL)
6773 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
6776 mutex_enter(&arc_reclaim_lock);
6777 arc_reclaim_thread_exit = B_TRUE;
6779 * The reclaim thread will set arc_reclaim_thread_exit back to
6780 * B_FALSE when it is finished exiting; we're waiting for that.
6782 while (arc_reclaim_thread_exit) {
6783 cv_signal(&arc_reclaim_thread_cv);
6784 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6786 mutex_exit(&arc_reclaim_lock);
6788 /* Use B_TRUE to ensure *all* buffers are evicted */
6789 arc_flush(NULL, B_TRUE);
6791 mutex_enter(&arc_dnlc_evicts_lock);
6792 arc_dnlc_evicts_thread_exit = TRUE;
6794 * The user evicts thread will set arc_user_evicts_thread_exit
6795 * to FALSE when it is finished exiting; we're waiting for that.
6797 while (arc_dnlc_evicts_thread_exit) {
6798 cv_signal(&arc_dnlc_evicts_cv);
6799 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
6801 mutex_exit(&arc_dnlc_evicts_lock);
6805 if (arc_ksp != NULL) {
6806 kstat_delete(arc_ksp);
6810 mutex_destroy(&arc_reclaim_lock);
6811 cv_destroy(&arc_reclaim_thread_cv);
6812 cv_destroy(&arc_reclaim_waiters_cv);
6814 mutex_destroy(&arc_dnlc_evicts_lock);
6815 cv_destroy(&arc_dnlc_evicts_cv);
6820 ASSERT0(arc_loaned_bytes);
6826 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6827 * It uses dedicated storage devices to hold cached data, which are populated
6828 * using large infrequent writes. The main role of this cache is to boost
6829 * the performance of random read workloads. The intended L2ARC devices
6830 * include short-stroked disks, solid state disks, and other media with
6831 * substantially faster read latency than disk.
6833 * +-----------------------+
6835 * +-----------------------+
6838 * l2arc_feed_thread() arc_read()
6842 * +---------------+ |
6844 * +---------------+ |
6849 * +-------+ +-------+
6851 * | cache | | cache |
6852 * +-------+ +-------+
6853 * +=========+ .-----.
6854 * : L2ARC : |-_____-|
6855 * : devices : | Disks |
6856 * +=========+ `-_____-'
6858 * Read requests are satisfied from the following sources, in order:
6861 * 2) vdev cache of L2ARC devices
6863 * 4) vdev cache of disks
6866 * Some L2ARC device types exhibit extremely slow write performance.
6867 * To accommodate for this there are some significant differences between
6868 * the L2ARC and traditional cache design:
6870 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6871 * the ARC behave as usual, freeing buffers and placing headers on ghost
6872 * lists. The ARC does not send buffers to the L2ARC during eviction as
6873 * this would add inflated write latencies for all ARC memory pressure.
6875 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6876 * It does this by periodically scanning buffers from the eviction-end of
6877 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6878 * not already there. It scans until a headroom of buffers is satisfied,
6879 * which itself is a buffer for ARC eviction. If a compressible buffer is
6880 * found during scanning and selected for writing to an L2ARC device, we
6881 * temporarily boost scanning headroom during the next scan cycle to make
6882 * sure we adapt to compression effects (which might significantly reduce
6883 * the data volume we write to L2ARC). The thread that does this is
6884 * l2arc_feed_thread(), illustrated below; example sizes are included to
6885 * provide a better sense of ratio than this diagram:
6888 * +---------------------+----------+
6889 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6890 * +---------------------+----------+ | o L2ARC eligible
6891 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6892 * +---------------------+----------+ |
6893 * 15.9 Gbytes ^ 32 Mbytes |
6895 * l2arc_feed_thread()
6897 * l2arc write hand <--[oooo]--'
6901 * +==============================+
6902 * L2ARC dev |####|#|###|###| |####| ... |
6903 * +==============================+
6906 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6907 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6908 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6909 * safe to say that this is an uncommon case, since buffers at the end of
6910 * the ARC lists have moved there due to inactivity.
6912 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6913 * then the L2ARC simply misses copying some buffers. This serves as a
6914 * pressure valve to prevent heavy read workloads from both stalling the ARC
6915 * with waits and clogging the L2ARC with writes. This also helps prevent
6916 * the potential for the L2ARC to churn if it attempts to cache content too
6917 * quickly, such as during backups of the entire pool.
6919 * 5. After system boot and before the ARC has filled main memory, there are
6920 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6921 * lists can remain mostly static. Instead of searching from tail of these
6922 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6923 * for eligible buffers, greatly increasing its chance of finding them.
6925 * The L2ARC device write speed is also boosted during this time so that
6926 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6927 * there are no L2ARC reads, and no fear of degrading read performance
6928 * through increased writes.
6930 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6931 * the vdev queue can aggregate them into larger and fewer writes. Each
6932 * device is written to in a rotor fashion, sweeping writes through
6933 * available space then repeating.
6935 * 7. The L2ARC does not store dirty content. It never needs to flush
6936 * write buffers back to disk based storage.
6938 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6939 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6941 * The performance of the L2ARC can be tweaked by a number of tunables, which
6942 * may be necessary for different workloads:
6944 * l2arc_write_max max write bytes per interval
6945 * l2arc_write_boost extra write bytes during device warmup
6946 * l2arc_noprefetch skip caching prefetched buffers
6947 * l2arc_headroom number of max device writes to precache
6948 * l2arc_headroom_boost when we find compressed buffers during ARC
6949 * scanning, we multiply headroom by this
6950 * percentage factor for the next scan cycle,
6951 * since more compressed buffers are likely to
6953 * l2arc_feed_secs seconds between L2ARC writing
6955 * Tunables may be removed or added as future performance improvements are
6956 * integrated, and also may become zpool properties.
6958 * There are three key functions that control how the L2ARC warms up:
6960 * l2arc_write_eligible() check if a buffer is eligible to cache
6961 * l2arc_write_size() calculate how much to write
6962 * l2arc_write_interval() calculate sleep delay between writes
6964 * These three functions determine what to write, how much, and how quickly
6969 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6972 * A buffer is *not* eligible for the L2ARC if it:
6973 * 1. belongs to a different spa.
6974 * 2. is already cached on the L2ARC.
6975 * 3. has an I/O in progress (it may be an incomplete read).
6976 * 4. is flagged not eligible (zfs property).
6978 if (hdr->b_spa != spa_guid) {
6979 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
6982 if (HDR_HAS_L2HDR(hdr)) {
6983 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
6986 if (HDR_IO_IN_PROGRESS(hdr)) {
6987 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
6990 if (!HDR_L2CACHE(hdr)) {
6991 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
6999 l2arc_write_size(void)
7004 * Make sure our globals have meaningful values in case the user
7007 size = l2arc_write_max;
7009 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
7010 "be greater than zero, resetting it to the default (%d)",
7012 size = l2arc_write_max = L2ARC_WRITE_SIZE;
7015 if (arc_warm == B_FALSE)
7016 size += l2arc_write_boost;
7023 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
7025 clock_t interval, next, now;
7028 * If the ARC lists are busy, increase our write rate; if the
7029 * lists are stale, idle back. This is achieved by checking
7030 * how much we previously wrote - if it was more than half of
7031 * what we wanted, schedule the next write much sooner.
7033 if (l2arc_feed_again && wrote > (wanted / 2))
7034 interval = (hz * l2arc_feed_min_ms) / 1000;
7036 interval = hz * l2arc_feed_secs;
7038 now = ddi_get_lbolt();
7039 next = MAX(now, MIN(now + interval, began + interval));
7045 * Cycle through L2ARC devices. This is how L2ARC load balances.
7046 * If a device is returned, this also returns holding the spa config lock.
7048 static l2arc_dev_t *
7049 l2arc_dev_get_next(void)
7051 l2arc_dev_t *first, *next = NULL;
7054 * Lock out the removal of spas (spa_namespace_lock), then removal
7055 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7056 * both locks will be dropped and a spa config lock held instead.
7058 mutex_enter(&spa_namespace_lock);
7059 mutex_enter(&l2arc_dev_mtx);
7061 /* if there are no vdevs, there is nothing to do */
7062 if (l2arc_ndev == 0)
7066 next = l2arc_dev_last;
7068 /* loop around the list looking for a non-faulted vdev */
7070 next = list_head(l2arc_dev_list);
7072 next = list_next(l2arc_dev_list, next);
7074 next = list_head(l2arc_dev_list);
7077 /* if we have come back to the start, bail out */
7080 else if (next == first)
7083 } while (vdev_is_dead(next->l2ad_vdev));
7085 /* if we were unable to find any usable vdevs, return NULL */
7086 if (vdev_is_dead(next->l2ad_vdev))
7089 l2arc_dev_last = next;
7092 mutex_exit(&l2arc_dev_mtx);
7095 * Grab the config lock to prevent the 'next' device from being
7096 * removed while we are writing to it.
7099 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7100 mutex_exit(&spa_namespace_lock);
7106 * Free buffers that were tagged for destruction.
7109 l2arc_do_free_on_write()
7112 l2arc_data_free_t *df, *df_prev;
7114 mutex_enter(&l2arc_free_on_write_mtx);
7115 buflist = l2arc_free_on_write;
7117 for (df = list_tail(buflist); df; df = df_prev) {
7118 df_prev = list_prev(buflist, df);
7119 ASSERT3P(df->l2df_abd, !=, NULL);
7120 abd_free(df->l2df_abd);
7121 list_remove(buflist, df);
7122 kmem_free(df, sizeof (l2arc_data_free_t));
7125 mutex_exit(&l2arc_free_on_write_mtx);
7129 * A write to a cache device has completed. Update all headers to allow
7130 * reads from these buffers to begin.
7133 l2arc_write_done(zio_t *zio)
7135 l2arc_write_callback_t *cb;
7138 arc_buf_hdr_t *head, *hdr, *hdr_prev;
7139 kmutex_t *hash_lock;
7140 int64_t bytes_dropped = 0;
7142 cb = zio->io_private;
7143 ASSERT3P(cb, !=, NULL);
7144 dev = cb->l2wcb_dev;
7145 ASSERT3P(dev, !=, NULL);
7146 head = cb->l2wcb_head;
7147 ASSERT3P(head, !=, NULL);
7148 buflist = &dev->l2ad_buflist;
7149 ASSERT3P(buflist, !=, NULL);
7150 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7151 l2arc_write_callback_t *, cb);
7153 if (zio->io_error != 0)
7154 ARCSTAT_BUMP(arcstat_l2_writes_error);
7157 * All writes completed, or an error was hit.
7160 mutex_enter(&dev->l2ad_mtx);
7161 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7162 hdr_prev = list_prev(buflist, hdr);
7164 hash_lock = HDR_LOCK(hdr);
7167 * We cannot use mutex_enter or else we can deadlock
7168 * with l2arc_write_buffers (due to swapping the order
7169 * the hash lock and l2ad_mtx are taken).
7171 if (!mutex_tryenter(hash_lock)) {
7173 * Missed the hash lock. We must retry so we
7174 * don't leave the ARC_FLAG_L2_WRITING bit set.
7176 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7179 * We don't want to rescan the headers we've
7180 * already marked as having been written out, so
7181 * we reinsert the head node so we can pick up
7182 * where we left off.
7184 list_remove(buflist, head);
7185 list_insert_after(buflist, hdr, head);
7187 mutex_exit(&dev->l2ad_mtx);
7190 * We wait for the hash lock to become available
7191 * to try and prevent busy waiting, and increase
7192 * the chance we'll be able to acquire the lock
7193 * the next time around.
7195 mutex_enter(hash_lock);
7196 mutex_exit(hash_lock);
7201 * We could not have been moved into the arc_l2c_only
7202 * state while in-flight due to our ARC_FLAG_L2_WRITING
7203 * bit being set. Let's just ensure that's being enforced.
7205 ASSERT(HDR_HAS_L1HDR(hdr));
7207 if (zio->io_error != 0) {
7209 * Error - drop L2ARC entry.
7211 list_remove(buflist, hdr);
7213 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7215 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
7216 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
7218 bytes_dropped += arc_hdr_size(hdr);
7219 (void) refcount_remove_many(&dev->l2ad_alloc,
7220 arc_hdr_size(hdr), hdr);
7224 * Allow ARC to begin reads and ghost list evictions to
7227 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7229 mutex_exit(hash_lock);
7232 atomic_inc_64(&l2arc_writes_done);
7233 list_remove(buflist, head);
7234 ASSERT(!HDR_HAS_L1HDR(head));
7235 kmem_cache_free(hdr_l2only_cache, head);
7236 mutex_exit(&dev->l2ad_mtx);
7238 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7240 l2arc_do_free_on_write();
7242 kmem_free(cb, sizeof (l2arc_write_callback_t));
7246 * A read to a cache device completed. Validate buffer contents before
7247 * handing over to the regular ARC routines.
7250 l2arc_read_done(zio_t *zio)
7252 l2arc_read_callback_t *cb;
7254 kmutex_t *hash_lock;
7255 boolean_t valid_cksum;
7257 ASSERT3P(zio->io_vd, !=, NULL);
7258 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7260 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7262 cb = zio->io_private;
7263 ASSERT3P(cb, !=, NULL);
7264 hdr = cb->l2rcb_hdr;
7265 ASSERT3P(hdr, !=, NULL);
7267 hash_lock = HDR_LOCK(hdr);
7268 mutex_enter(hash_lock);
7269 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7272 * If the data was read into a temporary buffer,
7273 * move it and free the buffer.
7275 if (cb->l2rcb_abd != NULL) {
7276 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7277 if (zio->io_error == 0) {
7278 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
7283 * The following must be done regardless of whether
7284 * there was an error:
7285 * - free the temporary buffer
7286 * - point zio to the real ARC buffer
7287 * - set zio size accordingly
7288 * These are required because zio is either re-used for
7289 * an I/O of the block in the case of the error
7290 * or the zio is passed to arc_read_done() and it
7293 abd_free(cb->l2rcb_abd);
7294 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7295 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
7298 ASSERT3P(zio->io_abd, !=, NULL);
7301 * Check this survived the L2ARC journey.
7303 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
7304 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
7305 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
7307 valid_cksum = arc_cksum_is_equal(hdr, zio);
7308 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
7309 mutex_exit(hash_lock);
7310 zio->io_private = hdr;
7313 mutex_exit(hash_lock);
7315 * Buffer didn't survive caching. Increment stats and
7316 * reissue to the original storage device.
7318 if (zio->io_error != 0) {
7319 ARCSTAT_BUMP(arcstat_l2_io_error);
7321 zio->io_error = SET_ERROR(EIO);
7324 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
7327 * If there's no waiter, issue an async i/o to the primary
7328 * storage now. If there *is* a waiter, the caller must
7329 * issue the i/o in a context where it's OK to block.
7331 if (zio->io_waiter == NULL) {
7332 zio_t *pio = zio_unique_parent(zio);
7334 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7336 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
7337 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
7338 hdr, zio->io_priority, cb->l2rcb_flags,
7343 kmem_free(cb, sizeof (l2arc_read_callback_t));
7347 * This is the list priority from which the L2ARC will search for pages to
7348 * cache. This is used within loops (0..3) to cycle through lists in the
7349 * desired order. This order can have a significant effect on cache
7352 * Currently the metadata lists are hit first, MFU then MRU, followed by
7353 * the data lists. This function returns a locked list, and also returns
7356 static multilist_sublist_t *
7357 l2arc_sublist_lock(int list_num)
7359 multilist_t *ml = NULL;
7362 ASSERT(list_num >= 0 && list_num <= 3);
7366 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7369 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7372 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7375 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7380 * Return a randomly-selected sublist. This is acceptable
7381 * because the caller feeds only a little bit of data for each
7382 * call (8MB). Subsequent calls will result in different
7383 * sublists being selected.
7385 idx = multilist_get_random_index(ml);
7386 return (multilist_sublist_lock(ml, idx));
7390 * Evict buffers from the device write hand to the distance specified in
7391 * bytes. This distance may span populated buffers, it may span nothing.
7392 * This is clearing a region on the L2ARC device ready for writing.
7393 * If the 'all' boolean is set, every buffer is evicted.
7396 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7399 arc_buf_hdr_t *hdr, *hdr_prev;
7400 kmutex_t *hash_lock;
7403 buflist = &dev->l2ad_buflist;
7405 if (!all && dev->l2ad_first) {
7407 * This is the first sweep through the device. There is
7413 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
7415 * When nearing the end of the device, evict to the end
7416 * before the device write hand jumps to the start.
7418 taddr = dev->l2ad_end;
7420 taddr = dev->l2ad_hand + distance;
7422 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
7423 uint64_t, taddr, boolean_t, all);
7426 mutex_enter(&dev->l2ad_mtx);
7427 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
7428 hdr_prev = list_prev(buflist, hdr);
7430 hash_lock = HDR_LOCK(hdr);
7433 * We cannot use mutex_enter or else we can deadlock
7434 * with l2arc_write_buffers (due to swapping the order
7435 * the hash lock and l2ad_mtx are taken).
7437 if (!mutex_tryenter(hash_lock)) {
7439 * Missed the hash lock. Retry.
7441 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
7442 mutex_exit(&dev->l2ad_mtx);
7443 mutex_enter(hash_lock);
7444 mutex_exit(hash_lock);
7449 * A header can't be on this list if it doesn't have L2 header.
7451 ASSERT(HDR_HAS_L2HDR(hdr));
7453 /* Ensure this header has finished being written. */
7454 ASSERT(!HDR_L2_WRITING(hdr));
7455 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
7457 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
7458 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
7460 * We've evicted to the target address,
7461 * or the end of the device.
7463 mutex_exit(hash_lock);
7467 if (!HDR_HAS_L1HDR(hdr)) {
7468 ASSERT(!HDR_L2_READING(hdr));
7470 * This doesn't exist in the ARC. Destroy.
7471 * arc_hdr_destroy() will call list_remove()
7472 * and decrement arcstat_l2_lsize.
7474 arc_change_state(arc_anon, hdr, hash_lock);
7475 arc_hdr_destroy(hdr);
7477 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
7478 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
7480 * Invalidate issued or about to be issued
7481 * reads, since we may be about to write
7482 * over this location.
7484 if (HDR_L2_READING(hdr)) {
7485 ARCSTAT_BUMP(arcstat_l2_evict_reading);
7486 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
7489 arc_hdr_l2hdr_destroy(hdr);
7491 mutex_exit(hash_lock);
7493 mutex_exit(&dev->l2ad_mtx);
7497 * Find and write ARC buffers to the L2ARC device.
7499 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7500 * for reading until they have completed writing.
7501 * The headroom_boost is an in-out parameter used to maintain headroom boost
7502 * state between calls to this function.
7504 * Returns the number of bytes actually written (which may be smaller than
7505 * the delta by which the device hand has changed due to alignment).
7508 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
7510 arc_buf_hdr_t *hdr, *hdr_prev, *head;
7511 uint64_t write_asize, write_psize, write_lsize, headroom;
7513 l2arc_write_callback_t *cb;
7515 uint64_t guid = spa_load_guid(spa);
7518 ASSERT3P(dev->l2ad_vdev, !=, NULL);
7521 write_lsize = write_asize = write_psize = 0;
7523 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
7524 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
7526 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
7528 * Copy buffers for L2ARC writing.
7530 for (try = 0; try <= 3; try++) {
7531 multilist_sublist_t *mls = l2arc_sublist_lock(try);
7532 uint64_t passed_sz = 0;
7534 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
7537 * L2ARC fast warmup.
7539 * Until the ARC is warm and starts to evict, read from the
7540 * head of the ARC lists rather than the tail.
7542 if (arc_warm == B_FALSE)
7543 hdr = multilist_sublist_head(mls);
7545 hdr = multilist_sublist_tail(mls);
7547 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
7549 headroom = target_sz * l2arc_headroom;
7550 if (zfs_compressed_arc_enabled)
7551 headroom = (headroom * l2arc_headroom_boost) / 100;
7553 for (; hdr; hdr = hdr_prev) {
7554 kmutex_t *hash_lock;
7556 if (arc_warm == B_FALSE)
7557 hdr_prev = multilist_sublist_next(mls, hdr);
7559 hdr_prev = multilist_sublist_prev(mls, hdr);
7560 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
7561 HDR_GET_LSIZE(hdr));
7563 hash_lock = HDR_LOCK(hdr);
7564 if (!mutex_tryenter(hash_lock)) {
7565 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
7567 * Skip this buffer rather than waiting.
7572 passed_sz += HDR_GET_LSIZE(hdr);
7573 if (passed_sz > headroom) {
7577 mutex_exit(hash_lock);
7578 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
7582 if (!l2arc_write_eligible(guid, hdr)) {
7583 mutex_exit(hash_lock);
7588 * We rely on the L1 portion of the header below, so
7589 * it's invalid for this header to have been evicted out
7590 * of the ghost cache, prior to being written out. The
7591 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7593 ASSERT(HDR_HAS_L1HDR(hdr));
7595 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
7596 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7597 ASSERT3U(arc_hdr_size(hdr), >, 0);
7598 uint64_t psize = arc_hdr_size(hdr);
7599 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
7602 if ((write_asize + asize) > target_sz) {
7604 mutex_exit(hash_lock);
7605 ARCSTAT_BUMP(arcstat_l2_write_full);
7611 * Insert a dummy header on the buflist so
7612 * l2arc_write_done() can find where the
7613 * write buffers begin without searching.
7615 mutex_enter(&dev->l2ad_mtx);
7616 list_insert_head(&dev->l2ad_buflist, head);
7617 mutex_exit(&dev->l2ad_mtx);
7620 sizeof (l2arc_write_callback_t), KM_SLEEP);
7621 cb->l2wcb_dev = dev;
7622 cb->l2wcb_head = head;
7623 pio = zio_root(spa, l2arc_write_done, cb,
7625 ARCSTAT_BUMP(arcstat_l2_write_pios);
7628 hdr->b_l2hdr.b_dev = dev;
7629 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7630 arc_hdr_set_flags(hdr,
7631 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7633 mutex_enter(&dev->l2ad_mtx);
7634 list_insert_head(&dev->l2ad_buflist, hdr);
7635 mutex_exit(&dev->l2ad_mtx);
7637 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
7640 * Normally the L2ARC can use the hdr's data, but if
7641 * we're sharing data between the hdr and one of its
7642 * bufs, L2ARC needs its own copy of the data so that
7643 * the ZIO below can't race with the buf consumer.
7644 * Another case where we need to create a copy of the
7645 * data is when the buffer size is not device-aligned
7646 * and we need to pad the block to make it such.
7647 * That also keeps the clock hand suitably aligned.
7649 * To ensure that the copy will be available for the
7650 * lifetime of the ZIO and be cleaned up afterwards, we
7651 * add it to the l2arc_free_on_write queue.
7654 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
7655 to_write = hdr->b_l1hdr.b_pabd;
7657 to_write = abd_alloc_for_io(asize,
7658 HDR_ISTYPE_METADATA(hdr));
7659 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
7660 if (asize != psize) {
7661 abd_zero_off(to_write, psize,
7664 l2arc_free_abd_on_write(to_write, asize,
7667 wzio = zio_write_phys(pio, dev->l2ad_vdev,
7668 hdr->b_l2hdr.b_daddr, asize, to_write,
7669 ZIO_CHECKSUM_OFF, NULL, hdr,
7670 ZIO_PRIORITY_ASYNC_WRITE,
7671 ZIO_FLAG_CANFAIL, B_FALSE);
7673 write_lsize += HDR_GET_LSIZE(hdr);
7674 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7677 write_psize += psize;
7678 write_asize += asize;
7679 dev->l2ad_hand += asize;
7681 mutex_exit(hash_lock);
7683 (void) zio_nowait(wzio);
7686 multilist_sublist_unlock(mls);
7692 /* No buffers selected for writing? */
7694 ASSERT0(write_lsize);
7695 ASSERT(!HDR_HAS_L1HDR(head));
7696 kmem_cache_free(hdr_l2only_cache, head);
7700 ASSERT3U(write_psize, <=, target_sz);
7701 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7702 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
7703 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
7704 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
7705 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
7708 * Bump device hand to the device start if it is approaching the end.
7709 * l2arc_evict() will already have evicted ahead for this case.
7711 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7712 dev->l2ad_hand = dev->l2ad_start;
7713 dev->l2ad_first = B_FALSE;
7716 dev->l2ad_writing = B_TRUE;
7717 (void) zio_wait(pio);
7718 dev->l2ad_writing = B_FALSE;
7720 return (write_asize);
7724 * This thread feeds the L2ARC at regular intervals. This is the beating
7725 * heart of the L2ARC.
7729 l2arc_feed_thread(void *unused __unused)
7734 uint64_t size, wrote;
7735 clock_t begin, next = ddi_get_lbolt();
7737 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7739 mutex_enter(&l2arc_feed_thr_lock);
7741 while (l2arc_thread_exit == 0) {
7742 CALLB_CPR_SAFE_BEGIN(&cpr);
7743 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7744 next - ddi_get_lbolt());
7745 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7746 next = ddi_get_lbolt() + hz;
7749 * Quick check for L2ARC devices.
7751 mutex_enter(&l2arc_dev_mtx);
7752 if (l2arc_ndev == 0) {
7753 mutex_exit(&l2arc_dev_mtx);
7756 mutex_exit(&l2arc_dev_mtx);
7757 begin = ddi_get_lbolt();
7760 * This selects the next l2arc device to write to, and in
7761 * doing so the next spa to feed from: dev->l2ad_spa. This
7762 * will return NULL if there are now no l2arc devices or if
7763 * they are all faulted.
7765 * If a device is returned, its spa's config lock is also
7766 * held to prevent device removal. l2arc_dev_get_next()
7767 * will grab and release l2arc_dev_mtx.
7769 if ((dev = l2arc_dev_get_next()) == NULL)
7772 spa = dev->l2ad_spa;
7773 ASSERT3P(spa, !=, NULL);
7776 * If the pool is read-only then force the feed thread to
7777 * sleep a little longer.
7779 if (!spa_writeable(spa)) {
7780 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7781 spa_config_exit(spa, SCL_L2ARC, dev);
7786 * Avoid contributing to memory pressure.
7788 if (arc_reclaim_needed()) {
7789 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7790 spa_config_exit(spa, SCL_L2ARC, dev);
7794 ARCSTAT_BUMP(arcstat_l2_feeds);
7796 size = l2arc_write_size();
7799 * Evict L2ARC buffers that will be overwritten.
7801 l2arc_evict(dev, size, B_FALSE);
7804 * Write ARC buffers.
7806 wrote = l2arc_write_buffers(spa, dev, size);
7809 * Calculate interval between writes.
7811 next = l2arc_write_interval(begin, size, wrote);
7812 spa_config_exit(spa, SCL_L2ARC, dev);
7815 l2arc_thread_exit = 0;
7816 cv_broadcast(&l2arc_feed_thr_cv);
7817 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7822 l2arc_vdev_present(vdev_t *vd)
7826 mutex_enter(&l2arc_dev_mtx);
7827 for (dev = list_head(l2arc_dev_list); dev != NULL;
7828 dev = list_next(l2arc_dev_list, dev)) {
7829 if (dev->l2ad_vdev == vd)
7832 mutex_exit(&l2arc_dev_mtx);
7834 return (dev != NULL);
7838 * Add a vdev for use by the L2ARC. By this point the spa has already
7839 * validated the vdev and opened it.
7842 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7844 l2arc_dev_t *adddev;
7846 ASSERT(!l2arc_vdev_present(vd));
7848 vdev_ashift_optimize(vd);
7851 * Create a new l2arc device entry.
7853 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7854 adddev->l2ad_spa = spa;
7855 adddev->l2ad_vdev = vd;
7856 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7857 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7858 adddev->l2ad_hand = adddev->l2ad_start;
7859 adddev->l2ad_first = B_TRUE;
7860 adddev->l2ad_writing = B_FALSE;
7862 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7864 * This is a list of all ARC buffers that are still valid on the
7867 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7868 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7870 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7871 refcount_create(&adddev->l2ad_alloc);
7874 * Add device to global list
7876 mutex_enter(&l2arc_dev_mtx);
7877 list_insert_head(l2arc_dev_list, adddev);
7878 atomic_inc_64(&l2arc_ndev);
7879 mutex_exit(&l2arc_dev_mtx);
7883 * Remove a vdev from the L2ARC.
7886 l2arc_remove_vdev(vdev_t *vd)
7888 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7891 * Find the device by vdev
7893 mutex_enter(&l2arc_dev_mtx);
7894 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7895 nextdev = list_next(l2arc_dev_list, dev);
7896 if (vd == dev->l2ad_vdev) {
7901 ASSERT3P(remdev, !=, NULL);
7904 * Remove device from global list
7906 list_remove(l2arc_dev_list, remdev);
7907 l2arc_dev_last = NULL; /* may have been invalidated */
7908 atomic_dec_64(&l2arc_ndev);
7909 mutex_exit(&l2arc_dev_mtx);
7912 * Clear all buflists and ARC references. L2ARC device flush.
7914 l2arc_evict(remdev, 0, B_TRUE);
7915 list_destroy(&remdev->l2ad_buflist);
7916 mutex_destroy(&remdev->l2ad_mtx);
7917 refcount_destroy(&remdev->l2ad_alloc);
7918 kmem_free(remdev, sizeof (l2arc_dev_t));
7924 l2arc_thread_exit = 0;
7926 l2arc_writes_sent = 0;
7927 l2arc_writes_done = 0;
7929 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7930 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7931 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7932 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7934 l2arc_dev_list = &L2ARC_dev_list;
7935 l2arc_free_on_write = &L2ARC_free_on_write;
7936 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7937 offsetof(l2arc_dev_t, l2ad_node));
7938 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7939 offsetof(l2arc_data_free_t, l2df_list_node));
7946 * This is called from dmu_fini(), which is called from spa_fini();
7947 * Because of this, we can assume that all l2arc devices have
7948 * already been removed when the pools themselves were removed.
7951 l2arc_do_free_on_write();
7953 mutex_destroy(&l2arc_feed_thr_lock);
7954 cv_destroy(&l2arc_feed_thr_cv);
7955 mutex_destroy(&l2arc_dev_mtx);
7956 mutex_destroy(&l2arc_free_on_write_mtx);
7958 list_destroy(l2arc_dev_list);
7959 list_destroy(l2arc_free_on_write);
7965 if (!(spa_mode_global & FWRITE))
7968 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7969 TS_RUN, minclsyspri);
7975 if (!(spa_mode_global & FWRITE))
7978 mutex_enter(&l2arc_feed_thr_lock);
7979 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7980 l2arc_thread_exit = 1;
7981 while (l2arc_thread_exit != 0)
7982 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7983 mutex_exit(&l2arc_feed_thr_lock);