4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 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>
279 #include <machine/vmparam.h>
283 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
284 boolean_t arc_watch = B_FALSE;
289 static kmutex_t arc_reclaim_lock;
290 static kcondvar_t arc_reclaim_thread_cv;
291 static boolean_t arc_reclaim_thread_exit;
292 static kcondvar_t arc_reclaim_waiters_cv;
294 static kmutex_t arc_dnlc_evicts_lock;
295 static kcondvar_t arc_dnlc_evicts_cv;
296 static boolean_t arc_dnlc_evicts_thread_exit;
298 uint_t arc_reduce_dnlc_percent = 3;
301 * The number of headers to evict in arc_evict_state_impl() before
302 * dropping the sublist lock and evicting from another sublist. A lower
303 * value means we're more likely to evict the "correct" header (i.e. the
304 * oldest header in the arc state), but comes with higher overhead
305 * (i.e. more invocations of arc_evict_state_impl()).
307 int zfs_arc_evict_batch_limit = 10;
309 /* number of seconds before growing cache again */
310 static int arc_grow_retry = 60;
312 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
313 int zfs_arc_overflow_shift = 8;
315 /* shift of arc_c for calculating both min and max arc_p */
316 static int arc_p_min_shift = 4;
318 /* log2(fraction of arc to reclaim) */
319 static int arc_shrink_shift = 7;
322 * log2(fraction of ARC which must be free to allow growing).
323 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
324 * when reading a new block into the ARC, we will evict an equal-sized block
327 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
328 * we will still not allow it to grow.
330 int arc_no_grow_shift = 5;
334 * minimum lifespan of a prefetch block in clock ticks
335 * (initialized in arc_init())
337 static int arc_min_prefetch_lifespan;
340 * If this percent of memory is free, don't throttle.
342 int arc_lotsfree_percent = 10;
345 extern boolean_t zfs_prefetch_disable;
348 * The arc has filled available memory and has now warmed up.
350 static boolean_t arc_warm;
353 * These tunables are for performance analysis.
355 uint64_t zfs_arc_max;
356 uint64_t zfs_arc_min;
357 uint64_t zfs_arc_meta_limit = 0;
358 uint64_t zfs_arc_meta_min = 0;
359 int zfs_arc_grow_retry = 0;
360 int zfs_arc_shrink_shift = 0;
361 int zfs_arc_no_grow_shift = 0;
362 int zfs_arc_p_min_shift = 0;
363 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
364 u_int zfs_arc_free_target = 0;
366 /* Absolute min for arc min / max is 16MB. */
367 static uint64_t arc_abs_min = 16 << 20;
369 boolean_t zfs_compressed_arc_enabled = B_TRUE;
371 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
372 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
373 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
374 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
375 static int sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS);
377 #if defined(__FreeBSD__) && defined(_KERNEL)
379 arc_free_target_init(void *unused __unused)
382 zfs_arc_free_target = vm_pageout_wakeup_thresh;
384 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
385 arc_free_target_init, NULL);
387 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
388 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
389 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
390 TUNABLE_INT("vfs.zfs.arc_grow_retry", &zfs_arc_grow_retry);
391 TUNABLE_INT("vfs.zfs.arc_no_grow_shift", &zfs_arc_no_grow_shift);
392 SYSCTL_DECL(_vfs_zfs);
393 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
394 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
395 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
396 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
397 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_no_grow_shift, CTLTYPE_U32 | CTLFLAG_RWTUN,
398 0, sizeof(uint32_t), sysctl_vfs_zfs_arc_no_grow_shift, "U",
399 "log2(fraction of ARC which must be free to allow growing)");
400 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
401 &zfs_arc_average_blocksize, 0,
402 "ARC average blocksize");
403 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
404 &arc_shrink_shift, 0,
405 "log2(fraction of arc to reclaim)");
406 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_grow_retry, CTLFLAG_RW,
408 "Wait in seconds before considering growing ARC");
409 SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN,
410 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC");
413 * We don't have a tunable for arc_free_target due to the dependency on
414 * pagedaemon initialisation.
416 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
417 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
418 sysctl_vfs_zfs_arc_free_target, "IU",
419 "Desired number of free pages below which ARC triggers reclaim");
422 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
427 val = zfs_arc_free_target;
428 err = sysctl_handle_int(oidp, &val, 0, req);
429 if (err != 0 || req->newptr == NULL)
434 if (val > vm_cnt.v_page_count)
437 zfs_arc_free_target = val;
443 * Must be declared here, before the definition of corresponding kstat
444 * macro which uses the same names will confuse the compiler.
446 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
447 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
448 sysctl_vfs_zfs_arc_meta_limit, "QU",
449 "ARC metadata limit");
453 * Note that buffers can be in one of 6 states:
454 * ARC_anon - anonymous (discussed below)
455 * ARC_mru - recently used, currently cached
456 * ARC_mru_ghost - recentely used, no longer in cache
457 * ARC_mfu - frequently used, currently cached
458 * ARC_mfu_ghost - frequently used, no longer in cache
459 * ARC_l2c_only - exists in L2ARC but not other states
460 * When there are no active references to the buffer, they are
461 * are linked onto a list in one of these arc states. These are
462 * the only buffers that can be evicted or deleted. Within each
463 * state there are multiple lists, one for meta-data and one for
464 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
465 * etc.) is tracked separately so that it can be managed more
466 * explicitly: favored over data, limited explicitly.
468 * Anonymous buffers are buffers that are not associated with
469 * a DVA. These are buffers that hold dirty block copies
470 * before they are written to stable storage. By definition,
471 * they are "ref'd" and are considered part of arc_mru
472 * that cannot be freed. Generally, they will aquire a DVA
473 * as they are written and migrate onto the arc_mru list.
475 * The ARC_l2c_only state is for buffers that are in the second
476 * level ARC but no longer in any of the ARC_m* lists. The second
477 * level ARC itself may also contain buffers that are in any of
478 * the ARC_m* states - meaning that a buffer can exist in two
479 * places. The reason for the ARC_l2c_only state is to keep the
480 * buffer header in the hash table, so that reads that hit the
481 * second level ARC benefit from these fast lookups.
484 typedef struct arc_state {
486 * list of evictable buffers
488 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
490 * total amount of evictable data in this state
492 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
494 * total amount of data in this state; this includes: evictable,
495 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
497 refcount_t arcs_size;
501 static arc_state_t ARC_anon;
502 static arc_state_t ARC_mru;
503 static arc_state_t ARC_mru_ghost;
504 static arc_state_t ARC_mfu;
505 static arc_state_t ARC_mfu_ghost;
506 static arc_state_t ARC_l2c_only;
508 typedef struct arc_stats {
509 kstat_named_t arcstat_hits;
510 kstat_named_t arcstat_misses;
511 kstat_named_t arcstat_demand_data_hits;
512 kstat_named_t arcstat_demand_data_misses;
513 kstat_named_t arcstat_demand_metadata_hits;
514 kstat_named_t arcstat_demand_metadata_misses;
515 kstat_named_t arcstat_prefetch_data_hits;
516 kstat_named_t arcstat_prefetch_data_misses;
517 kstat_named_t arcstat_prefetch_metadata_hits;
518 kstat_named_t arcstat_prefetch_metadata_misses;
519 kstat_named_t arcstat_mru_hits;
520 kstat_named_t arcstat_mru_ghost_hits;
521 kstat_named_t arcstat_mfu_hits;
522 kstat_named_t arcstat_mfu_ghost_hits;
523 kstat_named_t arcstat_allocated;
524 kstat_named_t arcstat_deleted;
526 * Number of buffers that could not be evicted because the hash lock
527 * was held by another thread. The lock may not necessarily be held
528 * by something using the same buffer, since hash locks are shared
529 * by multiple buffers.
531 kstat_named_t arcstat_mutex_miss;
533 * Number of buffers skipped because they have I/O in progress, are
534 * indrect prefetch buffers that have not lived long enough, or are
535 * not from the spa we're trying to evict from.
537 kstat_named_t arcstat_evict_skip;
539 * Number of times arc_evict_state() was unable to evict enough
540 * buffers to reach it's target amount.
542 kstat_named_t arcstat_evict_not_enough;
543 kstat_named_t arcstat_evict_l2_cached;
544 kstat_named_t arcstat_evict_l2_eligible;
545 kstat_named_t arcstat_evict_l2_ineligible;
546 kstat_named_t arcstat_evict_l2_skip;
547 kstat_named_t arcstat_hash_elements;
548 kstat_named_t arcstat_hash_elements_max;
549 kstat_named_t arcstat_hash_collisions;
550 kstat_named_t arcstat_hash_chains;
551 kstat_named_t arcstat_hash_chain_max;
552 kstat_named_t arcstat_p;
553 kstat_named_t arcstat_c;
554 kstat_named_t arcstat_c_min;
555 kstat_named_t arcstat_c_max;
556 kstat_named_t arcstat_size;
558 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
559 * Note that the compressed bytes may match the uncompressed bytes
560 * if the block is either not compressed or compressed arc is disabled.
562 kstat_named_t arcstat_compressed_size;
564 * Uncompressed size of the data stored in b_pabd. If compressed
565 * arc is disabled then this value will be identical to the stat
568 kstat_named_t arcstat_uncompressed_size;
570 * Number of bytes stored in all the arc_buf_t's. This is classified
571 * as "overhead" since this data is typically short-lived and will
572 * be evicted from the arc when it becomes unreferenced unless the
573 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
574 * values have been set (see comment in dbuf.c for more information).
576 kstat_named_t arcstat_overhead_size;
578 * Number of bytes consumed by internal ARC structures necessary
579 * for tracking purposes; these structures are not actually
580 * backed by ARC buffers. This includes arc_buf_hdr_t structures
581 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
582 * caches), and arc_buf_t structures (allocated via arc_buf_t
585 kstat_named_t arcstat_hdr_size;
587 * Number of bytes consumed by ARC buffers of type equal to
588 * ARC_BUFC_DATA. This is generally consumed by buffers backing
589 * on disk user data (e.g. plain file contents).
591 kstat_named_t arcstat_data_size;
593 * Number of bytes consumed by ARC buffers of type equal to
594 * ARC_BUFC_METADATA. This is generally consumed by buffers
595 * backing on disk data that is used for internal ZFS
596 * structures (e.g. ZAP, dnode, indirect blocks, etc).
598 kstat_named_t arcstat_metadata_size;
600 * Number of bytes consumed by various buffers and structures
601 * not actually backed with ARC buffers. This includes bonus
602 * buffers (allocated directly via zio_buf_* functions),
603 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
604 * cache), and dnode_t structures (allocated via dnode_t cache).
606 kstat_named_t arcstat_other_size;
608 * Total number of bytes consumed by ARC buffers residing in the
609 * arc_anon state. This includes *all* buffers in the arc_anon
610 * state; e.g. data, metadata, evictable, and unevictable buffers
611 * are all included in this value.
613 kstat_named_t arcstat_anon_size;
615 * Number of bytes consumed by ARC buffers that meet the
616 * following criteria: backing buffers of type ARC_BUFC_DATA,
617 * residing in the arc_anon state, and are eligible for eviction
618 * (e.g. have no outstanding holds on the buffer).
620 kstat_named_t arcstat_anon_evictable_data;
622 * Number of bytes consumed by ARC buffers that meet the
623 * following criteria: backing buffers of type ARC_BUFC_METADATA,
624 * residing in the arc_anon state, and are eligible for eviction
625 * (e.g. have no outstanding holds on the buffer).
627 kstat_named_t arcstat_anon_evictable_metadata;
629 * Total number of bytes consumed by ARC buffers residing in the
630 * arc_mru state. This includes *all* buffers in the arc_mru
631 * state; e.g. data, metadata, evictable, and unevictable buffers
632 * are all included in this value.
634 kstat_named_t arcstat_mru_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_mru state, and are eligible for eviction
639 * (e.g. have no outstanding holds on the buffer).
641 kstat_named_t arcstat_mru_evictable_data;
643 * Number of bytes consumed by ARC buffers that meet the
644 * following criteria: backing buffers of type ARC_BUFC_METADATA,
645 * residing in the arc_mru state, and are eligible for eviction
646 * (e.g. have no outstanding holds on the buffer).
648 kstat_named_t arcstat_mru_evictable_metadata;
650 * Total number of bytes that *would have been* consumed by ARC
651 * buffers in the arc_mru_ghost state. The key thing to note
652 * here, is the fact that this size doesn't actually indicate
653 * RAM consumption. The ghost lists only consist of headers and
654 * don't actually have ARC buffers linked off of these headers.
655 * Thus, *if* the headers had associated ARC buffers, these
656 * buffers *would have* consumed this number of bytes.
658 kstat_named_t arcstat_mru_ghost_size;
660 * Number of bytes that *would have been* consumed by ARC
661 * buffers that are eligible for eviction, of type
662 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
664 kstat_named_t arcstat_mru_ghost_evictable_data;
666 * Number of bytes that *would have been* consumed by ARC
667 * buffers that are eligible for eviction, of type
668 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
670 kstat_named_t arcstat_mru_ghost_evictable_metadata;
672 * Total number of bytes consumed by ARC buffers residing in the
673 * arc_mfu state. This includes *all* buffers in the arc_mfu
674 * state; e.g. data, metadata, evictable, and unevictable buffers
675 * are all included in this value.
677 kstat_named_t arcstat_mfu_size;
679 * Number of bytes consumed by ARC buffers that are eligible for
680 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
683 kstat_named_t arcstat_mfu_evictable_data;
685 * Number of bytes consumed by ARC buffers that are eligible for
686 * eviction, of type ARC_BUFC_METADATA, and reside in the
689 kstat_named_t arcstat_mfu_evictable_metadata;
691 * Total number of bytes that *would have been* consumed by ARC
692 * buffers in the arc_mfu_ghost state. See the comment above
693 * arcstat_mru_ghost_size for more details.
695 kstat_named_t arcstat_mfu_ghost_size;
697 * Number of bytes that *would have been* consumed by ARC
698 * buffers that are eligible for eviction, of type
699 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
701 kstat_named_t arcstat_mfu_ghost_evictable_data;
703 * Number of bytes that *would have been* consumed by ARC
704 * buffers that are eligible for eviction, of type
705 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
707 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
708 kstat_named_t arcstat_l2_hits;
709 kstat_named_t arcstat_l2_misses;
710 kstat_named_t arcstat_l2_feeds;
711 kstat_named_t arcstat_l2_rw_clash;
712 kstat_named_t arcstat_l2_read_bytes;
713 kstat_named_t arcstat_l2_write_bytes;
714 kstat_named_t arcstat_l2_writes_sent;
715 kstat_named_t arcstat_l2_writes_done;
716 kstat_named_t arcstat_l2_writes_error;
717 kstat_named_t arcstat_l2_writes_lock_retry;
718 kstat_named_t arcstat_l2_evict_lock_retry;
719 kstat_named_t arcstat_l2_evict_reading;
720 kstat_named_t arcstat_l2_evict_l1cached;
721 kstat_named_t arcstat_l2_free_on_write;
722 kstat_named_t arcstat_l2_abort_lowmem;
723 kstat_named_t arcstat_l2_cksum_bad;
724 kstat_named_t arcstat_l2_io_error;
725 kstat_named_t arcstat_l2_lsize;
726 kstat_named_t arcstat_l2_psize;
727 kstat_named_t arcstat_l2_hdr_size;
728 kstat_named_t arcstat_l2_write_trylock_fail;
729 kstat_named_t arcstat_l2_write_passed_headroom;
730 kstat_named_t arcstat_l2_write_spa_mismatch;
731 kstat_named_t arcstat_l2_write_in_l2;
732 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
733 kstat_named_t arcstat_l2_write_not_cacheable;
734 kstat_named_t arcstat_l2_write_full;
735 kstat_named_t arcstat_l2_write_buffer_iter;
736 kstat_named_t arcstat_l2_write_pios;
737 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
738 kstat_named_t arcstat_l2_write_buffer_list_iter;
739 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
740 kstat_named_t arcstat_memory_throttle_count;
741 kstat_named_t arcstat_meta_used;
742 kstat_named_t arcstat_meta_limit;
743 kstat_named_t arcstat_meta_max;
744 kstat_named_t arcstat_meta_min;
745 kstat_named_t arcstat_sync_wait_for_async;
746 kstat_named_t arcstat_demand_hit_predictive_prefetch;
749 static arc_stats_t arc_stats = {
750 { "hits", KSTAT_DATA_UINT64 },
751 { "misses", KSTAT_DATA_UINT64 },
752 { "demand_data_hits", KSTAT_DATA_UINT64 },
753 { "demand_data_misses", KSTAT_DATA_UINT64 },
754 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
755 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
756 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
757 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
758 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
759 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
760 { "mru_hits", KSTAT_DATA_UINT64 },
761 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
762 { "mfu_hits", KSTAT_DATA_UINT64 },
763 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
764 { "allocated", KSTAT_DATA_UINT64 },
765 { "deleted", KSTAT_DATA_UINT64 },
766 { "mutex_miss", KSTAT_DATA_UINT64 },
767 { "evict_skip", KSTAT_DATA_UINT64 },
768 { "evict_not_enough", KSTAT_DATA_UINT64 },
769 { "evict_l2_cached", KSTAT_DATA_UINT64 },
770 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
771 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
772 { "evict_l2_skip", KSTAT_DATA_UINT64 },
773 { "hash_elements", KSTAT_DATA_UINT64 },
774 { "hash_elements_max", KSTAT_DATA_UINT64 },
775 { "hash_collisions", KSTAT_DATA_UINT64 },
776 { "hash_chains", KSTAT_DATA_UINT64 },
777 { "hash_chain_max", KSTAT_DATA_UINT64 },
778 { "p", KSTAT_DATA_UINT64 },
779 { "c", KSTAT_DATA_UINT64 },
780 { "c_min", KSTAT_DATA_UINT64 },
781 { "c_max", KSTAT_DATA_UINT64 },
782 { "size", KSTAT_DATA_UINT64 },
783 { "compressed_size", KSTAT_DATA_UINT64 },
784 { "uncompressed_size", KSTAT_DATA_UINT64 },
785 { "overhead_size", KSTAT_DATA_UINT64 },
786 { "hdr_size", KSTAT_DATA_UINT64 },
787 { "data_size", KSTAT_DATA_UINT64 },
788 { "metadata_size", KSTAT_DATA_UINT64 },
789 { "other_size", KSTAT_DATA_UINT64 },
790 { "anon_size", KSTAT_DATA_UINT64 },
791 { "anon_evictable_data", KSTAT_DATA_UINT64 },
792 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
793 { "mru_size", KSTAT_DATA_UINT64 },
794 { "mru_evictable_data", KSTAT_DATA_UINT64 },
795 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
796 { "mru_ghost_size", KSTAT_DATA_UINT64 },
797 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
798 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
799 { "mfu_size", KSTAT_DATA_UINT64 },
800 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
801 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
802 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
803 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
804 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
805 { "l2_hits", KSTAT_DATA_UINT64 },
806 { "l2_misses", KSTAT_DATA_UINT64 },
807 { "l2_feeds", KSTAT_DATA_UINT64 },
808 { "l2_rw_clash", KSTAT_DATA_UINT64 },
809 { "l2_read_bytes", KSTAT_DATA_UINT64 },
810 { "l2_write_bytes", KSTAT_DATA_UINT64 },
811 { "l2_writes_sent", KSTAT_DATA_UINT64 },
812 { "l2_writes_done", KSTAT_DATA_UINT64 },
813 { "l2_writes_error", KSTAT_DATA_UINT64 },
814 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
815 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
816 { "l2_evict_reading", KSTAT_DATA_UINT64 },
817 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
818 { "l2_free_on_write", KSTAT_DATA_UINT64 },
819 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
820 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
821 { "l2_io_error", KSTAT_DATA_UINT64 },
822 { "l2_size", KSTAT_DATA_UINT64 },
823 { "l2_asize", KSTAT_DATA_UINT64 },
824 { "l2_hdr_size", KSTAT_DATA_UINT64 },
825 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
826 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
827 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
828 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
829 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
830 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
831 { "l2_write_full", KSTAT_DATA_UINT64 },
832 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
833 { "l2_write_pios", KSTAT_DATA_UINT64 },
834 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
835 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
836 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
837 { "memory_throttle_count", KSTAT_DATA_UINT64 },
838 { "arc_meta_used", KSTAT_DATA_UINT64 },
839 { "arc_meta_limit", KSTAT_DATA_UINT64 },
840 { "arc_meta_max", KSTAT_DATA_UINT64 },
841 { "arc_meta_min", KSTAT_DATA_UINT64 },
842 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
843 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
846 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
848 #define ARCSTAT_INCR(stat, val) \
849 atomic_add_64(&arc_stats.stat.value.ui64, (val))
851 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
852 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
854 #define ARCSTAT_MAX(stat, val) { \
856 while ((val) > (m = arc_stats.stat.value.ui64) && \
857 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
861 #define ARCSTAT_MAXSTAT(stat) \
862 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
865 * We define a macro to allow ARC hits/misses to be easily broken down by
866 * two separate conditions, giving a total of four different subtypes for
867 * each of hits and misses (so eight statistics total).
869 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
872 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
874 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
878 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
880 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
885 static arc_state_t *arc_anon;
886 static arc_state_t *arc_mru;
887 static arc_state_t *arc_mru_ghost;
888 static arc_state_t *arc_mfu;
889 static arc_state_t *arc_mfu_ghost;
890 static arc_state_t *arc_l2c_only;
893 * There are several ARC variables that are critical to export as kstats --
894 * but we don't want to have to grovel around in the kstat whenever we wish to
895 * manipulate them. For these variables, we therefore define them to be in
896 * terms of the statistic variable. This assures that we are not introducing
897 * the possibility of inconsistency by having shadow copies of the variables,
898 * while still allowing the code to be readable.
900 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
901 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
902 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
903 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
904 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
905 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
906 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
907 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
908 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
910 /* compressed size of entire arc */
911 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
912 /* uncompressed size of entire arc */
913 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
914 /* number of bytes in the arc from arc_buf_t's */
915 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
917 static int arc_no_grow; /* Don't try to grow cache size */
918 static uint64_t arc_tempreserve;
919 static uint64_t arc_loaned_bytes;
921 typedef struct arc_callback arc_callback_t;
923 struct arc_callback {
925 arc_done_func_t *acb_done;
927 boolean_t acb_compressed;
928 zio_t *acb_zio_dummy;
929 arc_callback_t *acb_next;
932 typedef struct arc_write_callback arc_write_callback_t;
934 struct arc_write_callback {
936 arc_done_func_t *awcb_ready;
937 arc_done_func_t *awcb_children_ready;
938 arc_done_func_t *awcb_physdone;
939 arc_done_func_t *awcb_done;
944 * ARC buffers are separated into multiple structs as a memory saving measure:
945 * - Common fields struct, always defined, and embedded within it:
946 * - L2-only fields, always allocated but undefined when not in L2ARC
947 * - L1-only fields, only allocated when in L1ARC
949 * Buffer in L1 Buffer only in L2
950 * +------------------------+ +------------------------+
951 * | arc_buf_hdr_t | | arc_buf_hdr_t |
955 * +------------------------+ +------------------------+
956 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
957 * | (undefined if L1-only) | | |
958 * +------------------------+ +------------------------+
959 * | l1arc_buf_hdr_t |
964 * +------------------------+
966 * Because it's possible for the L2ARC to become extremely large, we can wind
967 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
968 * is minimized by only allocating the fields necessary for an L1-cached buffer
969 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
970 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
971 * words in pointers. arc_hdr_realloc() is used to switch a header between
972 * these two allocation states.
974 typedef struct l1arc_buf_hdr {
975 kmutex_t b_freeze_lock;
976 zio_cksum_t *b_freeze_cksum;
979 * Used for debugging with kmem_flags - by allocating and freeing
980 * b_thawed when the buffer is thawed, we get a record of the stack
981 * trace that thawed it.
988 /* for waiting on writes to complete */
992 /* protected by arc state mutex */
993 arc_state_t *b_state;
994 multilist_node_t b_arc_node;
996 /* updated atomically */
997 clock_t b_arc_access;
999 /* self protecting */
1000 refcount_t b_refcnt;
1002 arc_callback_t *b_acb;
1006 typedef struct l2arc_dev l2arc_dev_t;
1008 typedef struct l2arc_buf_hdr {
1009 /* protected by arc_buf_hdr mutex */
1010 l2arc_dev_t *b_dev; /* L2ARC device */
1011 uint64_t b_daddr; /* disk address, offset byte */
1013 list_node_t b_l2node;
1016 struct arc_buf_hdr {
1017 /* protected by hash lock */
1021 arc_buf_contents_t b_type;
1022 arc_buf_hdr_t *b_hash_next;
1023 arc_flags_t b_flags;
1026 * This field stores the size of the data buffer after
1027 * compression, and is set in the arc's zio completion handlers.
1028 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1030 * While the block pointers can store up to 32MB in their psize
1031 * field, we can only store up to 32MB minus 512B. This is due
1032 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1033 * a field of zeros represents 512B in the bp). We can't use a
1034 * bias of 1 since we need to reserve a psize of zero, here, to
1035 * represent holes and embedded blocks.
1037 * This isn't a problem in practice, since the maximum size of a
1038 * buffer is limited to 16MB, so we never need to store 32MB in
1039 * this field. Even in the upstream illumos code base, the
1040 * maximum size of a buffer is limited to 16MB.
1045 * This field stores the size of the data buffer before
1046 * compression, and cannot change once set. It is in units
1047 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1049 uint16_t b_lsize; /* immutable */
1050 uint64_t b_spa; /* immutable */
1052 /* L2ARC fields. Undefined when not in L2ARC. */
1053 l2arc_buf_hdr_t b_l2hdr;
1054 /* L1ARC fields. Undefined when in l2arc_only state */
1055 l1arc_buf_hdr_t b_l1hdr;
1058 #if defined(__FreeBSD__) && defined(_KERNEL)
1060 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1065 val = arc_meta_limit;
1066 err = sysctl_handle_64(oidp, &val, 0, req);
1067 if (err != 0 || req->newptr == NULL)
1070 if (val <= 0 || val > arc_c_max)
1073 arc_meta_limit = val;
1078 sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS)
1083 val = arc_no_grow_shift;
1084 err = sysctl_handle_32(oidp, &val, 0, req);
1085 if (err != 0 || req->newptr == NULL)
1088 if (val >= arc_shrink_shift)
1091 arc_no_grow_shift = val;
1096 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1102 err = sysctl_handle_64(oidp, &val, 0, req);
1103 if (err != 0 || req->newptr == NULL)
1106 if (zfs_arc_max == 0) {
1107 /* Loader tunable so blindly set */
1112 if (val < arc_abs_min || val > kmem_size())
1114 if (val < arc_c_min)
1116 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1122 arc_p = (arc_c >> 1);
1124 if (zfs_arc_meta_limit == 0) {
1125 /* limit meta-data to 1/4 of the arc capacity */
1126 arc_meta_limit = arc_c_max / 4;
1129 /* if kmem_flags are set, lets try to use less memory */
1130 if (kmem_debugging())
1133 zfs_arc_max = arc_c;
1139 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1145 err = sysctl_handle_64(oidp, &val, 0, req);
1146 if (err != 0 || req->newptr == NULL)
1149 if (zfs_arc_min == 0) {
1150 /* Loader tunable so blindly set */
1155 if (val < arc_abs_min || val > arc_c_max)
1160 if (zfs_arc_meta_min == 0)
1161 arc_meta_min = arc_c_min / 2;
1163 if (arc_c < arc_c_min)
1166 zfs_arc_min = arc_c_min;
1172 #define GHOST_STATE(state) \
1173 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1174 (state) == arc_l2c_only)
1176 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1177 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1178 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1179 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1180 #define HDR_COMPRESSION_ENABLED(hdr) \
1181 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1183 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1184 #define HDR_L2_READING(hdr) \
1185 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1186 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1187 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1188 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1189 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1190 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1192 #define HDR_ISTYPE_METADATA(hdr) \
1193 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1194 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1196 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1197 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1199 /* For storing compression mode in b_flags */
1200 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1202 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1203 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1204 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1205 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1207 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1208 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1209 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1215 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1216 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1219 * Hash table routines
1222 #define HT_LOCK_PAD CACHE_LINE_SIZE
1227 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1231 #define BUF_LOCKS 256
1232 typedef struct buf_hash_table {
1234 arc_buf_hdr_t **ht_table;
1235 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1238 static buf_hash_table_t buf_hash_table;
1240 #define BUF_HASH_INDEX(spa, dva, birth) \
1241 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1242 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1243 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1244 #define HDR_LOCK(hdr) \
1245 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1247 uint64_t zfs_crc64_table[256];
1253 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1254 #define L2ARC_HEADROOM 2 /* num of writes */
1256 * If we discover during ARC scan any buffers to be compressed, we boost
1257 * our headroom for the next scanning cycle by this percentage multiple.
1259 #define L2ARC_HEADROOM_BOOST 200
1260 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1261 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1263 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1264 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1266 /* L2ARC Performance Tunables */
1267 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1268 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1269 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1270 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1271 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1272 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1273 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1274 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1275 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1277 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1278 &l2arc_write_max, 0, "max write size");
1279 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1280 &l2arc_write_boost, 0, "extra write during warmup");
1281 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1282 &l2arc_headroom, 0, "number of dev writes");
1283 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1284 &l2arc_feed_secs, 0, "interval seconds");
1285 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1286 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1288 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1289 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1290 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1291 &l2arc_feed_again, 0, "turbo warmup");
1292 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1293 &l2arc_norw, 0, "no reads during writes");
1295 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1296 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1297 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1298 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1299 "size of anonymous state");
1300 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1301 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1302 "size of anonymous state");
1304 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1305 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1306 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1307 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1308 "size of metadata in mru state");
1309 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1310 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1311 "size of data in mru state");
1313 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1314 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1315 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1316 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1317 "size of metadata in mru ghost state");
1318 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1319 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1320 "size of data in mru ghost state");
1322 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1323 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1324 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1325 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1326 "size of metadata in mfu state");
1327 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1328 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1329 "size of data in mfu state");
1331 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1332 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1333 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1334 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1335 "size of metadata in mfu ghost state");
1336 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1337 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1338 "size of data in mfu ghost state");
1340 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1341 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1347 vdev_t *l2ad_vdev; /* vdev */
1348 spa_t *l2ad_spa; /* spa */
1349 uint64_t l2ad_hand; /* next write location */
1350 uint64_t l2ad_start; /* first addr on device */
1351 uint64_t l2ad_end; /* last addr on device */
1352 boolean_t l2ad_first; /* first sweep through */
1353 boolean_t l2ad_writing; /* currently writing */
1354 kmutex_t l2ad_mtx; /* lock for buffer list */
1355 list_t l2ad_buflist; /* buffer list */
1356 list_node_t l2ad_node; /* device list node */
1357 refcount_t l2ad_alloc; /* allocated bytes */
1360 static list_t L2ARC_dev_list; /* device list */
1361 static list_t *l2arc_dev_list; /* device list pointer */
1362 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1363 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1364 static list_t L2ARC_free_on_write; /* free after write buf list */
1365 static list_t *l2arc_free_on_write; /* free after write list ptr */
1366 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1367 static uint64_t l2arc_ndev; /* number of devices */
1369 typedef struct l2arc_read_callback {
1370 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1371 blkptr_t l2rcb_bp; /* original blkptr */
1372 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1373 int l2rcb_flags; /* original flags */
1374 abd_t *l2rcb_abd; /* temporary buffer */
1375 } l2arc_read_callback_t;
1377 typedef struct l2arc_write_callback {
1378 l2arc_dev_t *l2wcb_dev; /* device info */
1379 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1380 } l2arc_write_callback_t;
1382 typedef struct l2arc_data_free {
1383 /* protected by l2arc_free_on_write_mtx */
1386 arc_buf_contents_t l2df_type;
1387 list_node_t l2df_list_node;
1388 } l2arc_data_free_t;
1390 static kmutex_t l2arc_feed_thr_lock;
1391 static kcondvar_t l2arc_feed_thr_cv;
1392 static uint8_t l2arc_thread_exit;
1394 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1395 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1396 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1397 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1398 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1399 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1400 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1401 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1402 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1403 static boolean_t arc_is_overflowing();
1404 static void arc_buf_watch(arc_buf_t *);
1406 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1407 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1408 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1409 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1411 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1412 static void l2arc_read_done(zio_t *);
1415 l2arc_trim(const arc_buf_hdr_t *hdr)
1417 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1419 ASSERT(HDR_HAS_L2HDR(hdr));
1420 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1422 if (HDR_GET_PSIZE(hdr) != 0) {
1423 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1424 HDR_GET_PSIZE(hdr), 0);
1429 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1431 uint8_t *vdva = (uint8_t *)dva;
1432 uint64_t crc = -1ULL;
1435 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1437 for (i = 0; i < sizeof (dva_t); i++)
1438 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1440 crc ^= (spa>>8) ^ birth;
1445 #define HDR_EMPTY(hdr) \
1446 ((hdr)->b_dva.dva_word[0] == 0 && \
1447 (hdr)->b_dva.dva_word[1] == 0)
1449 #define HDR_EQUAL(spa, dva, birth, hdr) \
1450 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1451 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1452 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1455 buf_discard_identity(arc_buf_hdr_t *hdr)
1457 hdr->b_dva.dva_word[0] = 0;
1458 hdr->b_dva.dva_word[1] = 0;
1462 static arc_buf_hdr_t *
1463 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1465 const dva_t *dva = BP_IDENTITY(bp);
1466 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1467 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1468 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1471 mutex_enter(hash_lock);
1472 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1473 hdr = hdr->b_hash_next) {
1474 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1479 mutex_exit(hash_lock);
1485 * Insert an entry into the hash table. If there is already an element
1486 * equal to elem in the hash table, then the already existing element
1487 * will be returned and the new element will not be inserted.
1488 * Otherwise returns NULL.
1489 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1491 static arc_buf_hdr_t *
1492 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1494 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1495 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1496 arc_buf_hdr_t *fhdr;
1499 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1500 ASSERT(hdr->b_birth != 0);
1501 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1503 if (lockp != NULL) {
1505 mutex_enter(hash_lock);
1507 ASSERT(MUTEX_HELD(hash_lock));
1510 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1511 fhdr = fhdr->b_hash_next, i++) {
1512 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1516 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1517 buf_hash_table.ht_table[idx] = hdr;
1518 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1520 /* collect some hash table performance data */
1522 ARCSTAT_BUMP(arcstat_hash_collisions);
1524 ARCSTAT_BUMP(arcstat_hash_chains);
1526 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1529 ARCSTAT_BUMP(arcstat_hash_elements);
1530 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1536 buf_hash_remove(arc_buf_hdr_t *hdr)
1538 arc_buf_hdr_t *fhdr, **hdrp;
1539 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1541 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1542 ASSERT(HDR_IN_HASH_TABLE(hdr));
1544 hdrp = &buf_hash_table.ht_table[idx];
1545 while ((fhdr = *hdrp) != hdr) {
1546 ASSERT3P(fhdr, !=, NULL);
1547 hdrp = &fhdr->b_hash_next;
1549 *hdrp = hdr->b_hash_next;
1550 hdr->b_hash_next = NULL;
1551 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1553 /* collect some hash table performance data */
1554 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1556 if (buf_hash_table.ht_table[idx] &&
1557 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1558 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1562 * Global data structures and functions for the buf kmem cache.
1564 static kmem_cache_t *hdr_full_cache;
1565 static kmem_cache_t *hdr_l2only_cache;
1566 static kmem_cache_t *buf_cache;
1573 kmem_free(buf_hash_table.ht_table,
1574 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1575 for (i = 0; i < BUF_LOCKS; i++)
1576 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1577 kmem_cache_destroy(hdr_full_cache);
1578 kmem_cache_destroy(hdr_l2only_cache);
1579 kmem_cache_destroy(buf_cache);
1583 * Constructor callback - called when the cache is empty
1584 * and a new buf is requested.
1588 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1590 arc_buf_hdr_t *hdr = vbuf;
1592 bzero(hdr, HDR_FULL_SIZE);
1593 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1594 refcount_create(&hdr->b_l1hdr.b_refcnt);
1595 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1596 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1597 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1604 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1606 arc_buf_hdr_t *hdr = vbuf;
1608 bzero(hdr, HDR_L2ONLY_SIZE);
1609 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1616 buf_cons(void *vbuf, void *unused, int kmflag)
1618 arc_buf_t *buf = vbuf;
1620 bzero(buf, sizeof (arc_buf_t));
1621 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1622 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1628 * Destructor callback - called when a cached buf is
1629 * no longer required.
1633 hdr_full_dest(void *vbuf, void *unused)
1635 arc_buf_hdr_t *hdr = vbuf;
1637 ASSERT(HDR_EMPTY(hdr));
1638 cv_destroy(&hdr->b_l1hdr.b_cv);
1639 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1640 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1641 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1642 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1647 hdr_l2only_dest(void *vbuf, void *unused)
1649 arc_buf_hdr_t *hdr = vbuf;
1651 ASSERT(HDR_EMPTY(hdr));
1652 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1657 buf_dest(void *vbuf, void *unused)
1659 arc_buf_t *buf = vbuf;
1661 mutex_destroy(&buf->b_evict_lock);
1662 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1666 * Reclaim callback -- invoked when memory is low.
1670 hdr_recl(void *unused)
1672 dprintf("hdr_recl called\n");
1674 * umem calls the reclaim func when we destroy the buf cache,
1675 * which is after we do arc_fini().
1678 cv_signal(&arc_reclaim_thread_cv);
1685 uint64_t hsize = 1ULL << 12;
1689 * The hash table is big enough to fill all of physical memory
1690 * with an average block size of zfs_arc_average_blocksize (default 8K).
1691 * By default, the table will take up
1692 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1694 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1697 buf_hash_table.ht_mask = hsize - 1;
1698 buf_hash_table.ht_table =
1699 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1700 if (buf_hash_table.ht_table == NULL) {
1701 ASSERT(hsize > (1ULL << 8));
1706 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1707 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1708 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1709 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1711 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1712 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1714 for (i = 0; i < 256; i++)
1715 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1716 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1718 for (i = 0; i < BUF_LOCKS; i++) {
1719 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1720 NULL, MUTEX_DEFAULT, NULL);
1725 * This is the size that the buf occupies in memory. If the buf is compressed,
1726 * it will correspond to the compressed size. You should use this method of
1727 * getting the buf size unless you explicitly need the logical size.
1730 arc_buf_size(arc_buf_t *buf)
1732 return (ARC_BUF_COMPRESSED(buf) ?
1733 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1737 arc_buf_lsize(arc_buf_t *buf)
1739 return (HDR_GET_LSIZE(buf->b_hdr));
1743 arc_get_compression(arc_buf_t *buf)
1745 return (ARC_BUF_COMPRESSED(buf) ?
1746 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1749 #define ARC_MINTIME (hz>>4) /* 62 ms */
1751 static inline boolean_t
1752 arc_buf_is_shared(arc_buf_t *buf)
1754 boolean_t shared = (buf->b_data != NULL &&
1755 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1756 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1757 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1758 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1759 IMPLY(shared, ARC_BUF_SHARED(buf));
1760 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1763 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1764 * already being shared" requirement prevents us from doing that.
1771 * Free the checksum associated with this header. If there is no checksum, this
1775 arc_cksum_free(arc_buf_hdr_t *hdr)
1777 ASSERT(HDR_HAS_L1HDR(hdr));
1778 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1779 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1780 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1781 hdr->b_l1hdr.b_freeze_cksum = NULL;
1783 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1787 * Return true iff at least one of the bufs on hdr is not compressed.
1790 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1792 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1793 if (!ARC_BUF_COMPRESSED(b)) {
1801 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1802 * matches the checksum that is stored in the hdr. If there is no checksum,
1803 * or if the buf is compressed, this is a no-op.
1806 arc_cksum_verify(arc_buf_t *buf)
1808 arc_buf_hdr_t *hdr = buf->b_hdr;
1811 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1814 if (ARC_BUF_COMPRESSED(buf)) {
1815 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1816 arc_hdr_has_uncompressed_buf(hdr));
1820 ASSERT(HDR_HAS_L1HDR(hdr));
1822 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1823 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1824 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1828 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1829 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1830 panic("buffer modified while frozen!");
1831 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1835 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1837 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1838 boolean_t valid_cksum;
1840 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1841 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1844 * We rely on the blkptr's checksum to determine if the block
1845 * is valid or not. When compressed arc is enabled, the l2arc
1846 * writes the block to the l2arc just as it appears in the pool.
1847 * This allows us to use the blkptr's checksum to validate the
1848 * data that we just read off of the l2arc without having to store
1849 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1850 * arc is disabled, then the data written to the l2arc is always
1851 * uncompressed and won't match the block as it exists in the main
1852 * pool. When this is the case, we must first compress it if it is
1853 * compressed on the main pool before we can validate the checksum.
1855 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1856 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1857 uint64_t lsize = HDR_GET_LSIZE(hdr);
1860 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1861 csize = zio_compress_data(compress, zio->io_abd,
1862 abd_to_buf(cdata), lsize);
1864 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1865 if (csize < HDR_GET_PSIZE(hdr)) {
1867 * Compressed blocks are always a multiple of the
1868 * smallest ashift in the pool. Ideally, we would
1869 * like to round up the csize to the next
1870 * spa_min_ashift but that value may have changed
1871 * since the block was last written. Instead,
1872 * we rely on the fact that the hdr's psize
1873 * was set to the psize of the block when it was
1874 * last written. We set the csize to that value
1875 * and zero out any part that should not contain
1878 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1879 csize = HDR_GET_PSIZE(hdr);
1881 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1885 * Block pointers always store the checksum for the logical data.
1886 * If the block pointer has the gang bit set, then the checksum
1887 * it represents is for the reconstituted data and not for an
1888 * individual gang member. The zio pipeline, however, must be able to
1889 * determine the checksum of each of the gang constituents so it
1890 * treats the checksum comparison differently than what we need
1891 * for l2arc blocks. This prevents us from using the
1892 * zio_checksum_error() interface directly. Instead we must call the
1893 * zio_checksum_error_impl() so that we can ensure the checksum is
1894 * generated using the correct checksum algorithm and accounts for the
1895 * logical I/O size and not just a gang fragment.
1897 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1898 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1899 zio->io_offset, NULL) == 0);
1900 zio_pop_transforms(zio);
1901 return (valid_cksum);
1905 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1906 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1907 * isn't modified later on. If buf is compressed or there is already a checksum
1908 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1911 arc_cksum_compute(arc_buf_t *buf)
1913 arc_buf_hdr_t *hdr = buf->b_hdr;
1915 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1918 ASSERT(HDR_HAS_L1HDR(hdr));
1920 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1921 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1922 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1923 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1925 } else if (ARC_BUF_COMPRESSED(buf)) {
1926 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1930 ASSERT(!ARC_BUF_COMPRESSED(buf));
1931 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1933 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1934 hdr->b_l1hdr.b_freeze_cksum);
1935 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1943 typedef struct procctl {
1951 arc_buf_unwatch(arc_buf_t *buf)
1958 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1959 ctl.prwatch.pr_size = 0;
1960 ctl.prwatch.pr_wflags = 0;
1961 result = write(arc_procfd, &ctl, sizeof (ctl));
1962 ASSERT3U(result, ==, sizeof (ctl));
1969 arc_buf_watch(arc_buf_t *buf)
1976 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1977 ctl.prwatch.pr_size = arc_buf_size(buf);
1978 ctl.prwatch.pr_wflags = WA_WRITE;
1979 result = write(arc_procfd, &ctl, sizeof (ctl));
1980 ASSERT3U(result, ==, sizeof (ctl));
1984 #endif /* illumos */
1986 static arc_buf_contents_t
1987 arc_buf_type(arc_buf_hdr_t *hdr)
1989 arc_buf_contents_t type;
1990 if (HDR_ISTYPE_METADATA(hdr)) {
1991 type = ARC_BUFC_METADATA;
1993 type = ARC_BUFC_DATA;
1995 VERIFY3U(hdr->b_type, ==, type);
2000 arc_is_metadata(arc_buf_t *buf)
2002 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2006 arc_bufc_to_flags(arc_buf_contents_t type)
2010 /* metadata field is 0 if buffer contains normal data */
2012 case ARC_BUFC_METADATA:
2013 return (ARC_FLAG_BUFC_METADATA);
2017 panic("undefined ARC buffer type!");
2018 return ((uint32_t)-1);
2022 arc_buf_thaw(arc_buf_t *buf)
2024 arc_buf_hdr_t *hdr = buf->b_hdr;
2026 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2027 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2029 arc_cksum_verify(buf);
2032 * Compressed buffers do not manipulate the b_freeze_cksum or
2033 * allocate b_thawed.
2035 if (ARC_BUF_COMPRESSED(buf)) {
2036 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2037 arc_hdr_has_uncompressed_buf(hdr));
2041 ASSERT(HDR_HAS_L1HDR(hdr));
2042 arc_cksum_free(hdr);
2044 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2046 if (zfs_flags & ZFS_DEBUG_MODIFY) {
2047 if (hdr->b_l1hdr.b_thawed != NULL)
2048 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2049 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2053 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2056 arc_buf_unwatch(buf);
2061 arc_buf_freeze(arc_buf_t *buf)
2063 arc_buf_hdr_t *hdr = buf->b_hdr;
2064 kmutex_t *hash_lock;
2066 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2069 if (ARC_BUF_COMPRESSED(buf)) {
2070 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2071 arc_hdr_has_uncompressed_buf(hdr));
2075 hash_lock = HDR_LOCK(hdr);
2076 mutex_enter(hash_lock);
2078 ASSERT(HDR_HAS_L1HDR(hdr));
2079 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
2080 hdr->b_l1hdr.b_state == arc_anon);
2081 arc_cksum_compute(buf);
2082 mutex_exit(hash_lock);
2086 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2087 * the following functions should be used to ensure that the flags are
2088 * updated in a thread-safe way. When manipulating the flags either
2089 * the hash_lock must be held or the hdr must be undiscoverable. This
2090 * ensures that we're not racing with any other threads when updating
2094 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2096 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2097 hdr->b_flags |= flags;
2101 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2103 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2104 hdr->b_flags &= ~flags;
2108 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2109 * done in a special way since we have to clear and set bits
2110 * at the same time. Consumers that wish to set the compression bits
2111 * must use this function to ensure that the flags are updated in
2112 * thread-safe manner.
2115 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2117 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2120 * Holes and embedded blocks will always have a psize = 0 so
2121 * we ignore the compression of the blkptr and set the
2122 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2123 * Holes and embedded blocks remain anonymous so we don't
2124 * want to uncompress them. Mark them as uncompressed.
2126 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2127 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2128 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2129 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2130 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2132 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2133 HDR_SET_COMPRESS(hdr, cmp);
2134 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2135 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2140 * Looks for another buf on the same hdr which has the data decompressed, copies
2141 * from it, and returns true. If no such buf exists, returns false.
2144 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
2146 arc_buf_hdr_t *hdr = buf->b_hdr;
2147 boolean_t copied = B_FALSE;
2149 ASSERT(HDR_HAS_L1HDR(hdr));
2150 ASSERT3P(buf->b_data, !=, NULL);
2151 ASSERT(!ARC_BUF_COMPRESSED(buf));
2153 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2154 from = from->b_next) {
2155 /* can't use our own data buffer */
2160 if (!ARC_BUF_COMPRESSED(from)) {
2161 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2168 * There were no decompressed bufs, so there should not be a
2169 * checksum on the hdr either.
2171 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
2177 * Given a buf that has a data buffer attached to it, this function will
2178 * efficiently fill the buf with data of the specified compression setting from
2179 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2180 * are already sharing a data buf, no copy is performed.
2182 * If the buf is marked as compressed but uncompressed data was requested, this
2183 * will allocate a new data buffer for the buf, remove that flag, and fill the
2184 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2185 * uncompressed data, and (since we haven't added support for it yet) if you
2186 * want compressed data your buf must already be marked as compressed and have
2187 * the correct-sized data buffer.
2190 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2192 arc_buf_hdr_t *hdr = buf->b_hdr;
2193 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2194 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2196 ASSERT3P(buf->b_data, !=, NULL);
2197 IMPLY(compressed, hdr_compressed);
2198 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2200 if (hdr_compressed == compressed) {
2201 if (!arc_buf_is_shared(buf)) {
2202 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2206 ASSERT(hdr_compressed);
2207 ASSERT(!compressed);
2208 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2211 * If the buf is sharing its data with the hdr, unlink it and
2212 * allocate a new data buffer for the buf.
2214 if (arc_buf_is_shared(buf)) {
2215 ASSERT(ARC_BUF_COMPRESSED(buf));
2217 /* We need to give the buf it's own b_data */
2218 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2220 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2221 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2223 /* Previously overhead was 0; just add new overhead */
2224 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2225 } else if (ARC_BUF_COMPRESSED(buf)) {
2226 /* We need to reallocate the buf's b_data */
2227 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2230 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2232 /* We increased the size of b_data; update overhead */
2233 ARCSTAT_INCR(arcstat_overhead_size,
2234 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2238 * Regardless of the buf's previous compression settings, it
2239 * should not be compressed at the end of this function.
2241 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2244 * Try copying the data from another buf which already has a
2245 * decompressed version. If that's not possible, it's time to
2246 * bite the bullet and decompress the data from the hdr.
2248 if (arc_buf_try_copy_decompressed_data(buf)) {
2249 /* Skip byteswapping and checksumming (already done) */
2250 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2253 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2254 hdr->b_l1hdr.b_pabd, buf->b_data,
2255 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2258 * Absent hardware errors or software bugs, this should
2259 * be impossible, but log it anyway so we can debug it.
2263 "hdr %p, compress %d, psize %d, lsize %d",
2264 hdr, HDR_GET_COMPRESS(hdr),
2265 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2266 return (SET_ERROR(EIO));
2271 /* Byteswap the buf's data if necessary */
2272 if (bswap != DMU_BSWAP_NUMFUNCS) {
2273 ASSERT(!HDR_SHARED_DATA(hdr));
2274 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2275 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2278 /* Compute the hdr's checksum if necessary */
2279 arc_cksum_compute(buf);
2285 arc_decompress(arc_buf_t *buf)
2287 return (arc_buf_fill(buf, B_FALSE));
2291 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2294 arc_hdr_size(arc_buf_hdr_t *hdr)
2298 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2299 HDR_GET_PSIZE(hdr) > 0) {
2300 size = HDR_GET_PSIZE(hdr);
2302 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2303 size = HDR_GET_LSIZE(hdr);
2309 * Increment the amount of evictable space in the arc_state_t's refcount.
2310 * We account for the space used by the hdr and the arc buf individually
2311 * so that we can add and remove them from the refcount individually.
2314 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2316 arc_buf_contents_t type = arc_buf_type(hdr);
2318 ASSERT(HDR_HAS_L1HDR(hdr));
2320 if (GHOST_STATE(state)) {
2321 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2322 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2323 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2324 (void) refcount_add_many(&state->arcs_esize[type],
2325 HDR_GET_LSIZE(hdr), hdr);
2329 ASSERT(!GHOST_STATE(state));
2330 if (hdr->b_l1hdr.b_pabd != NULL) {
2331 (void) refcount_add_many(&state->arcs_esize[type],
2332 arc_hdr_size(hdr), hdr);
2334 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2335 buf = buf->b_next) {
2336 if (arc_buf_is_shared(buf))
2338 (void) refcount_add_many(&state->arcs_esize[type],
2339 arc_buf_size(buf), buf);
2344 * Decrement the amount of evictable space in the arc_state_t's refcount.
2345 * We account for the space used by the hdr and the arc buf individually
2346 * so that we can add and remove them from the refcount individually.
2349 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2351 arc_buf_contents_t type = arc_buf_type(hdr);
2353 ASSERT(HDR_HAS_L1HDR(hdr));
2355 if (GHOST_STATE(state)) {
2356 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2357 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2358 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2359 (void) refcount_remove_many(&state->arcs_esize[type],
2360 HDR_GET_LSIZE(hdr), hdr);
2364 ASSERT(!GHOST_STATE(state));
2365 if (hdr->b_l1hdr.b_pabd != NULL) {
2366 (void) refcount_remove_many(&state->arcs_esize[type],
2367 arc_hdr_size(hdr), hdr);
2369 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2370 buf = buf->b_next) {
2371 if (arc_buf_is_shared(buf))
2373 (void) refcount_remove_many(&state->arcs_esize[type],
2374 arc_buf_size(buf), buf);
2379 * Add a reference to this hdr indicating that someone is actively
2380 * referencing that memory. When the refcount transitions from 0 to 1,
2381 * we remove it from the respective arc_state_t list to indicate that
2382 * it is not evictable.
2385 add_reference(arc_buf_hdr_t *hdr, void *tag)
2387 ASSERT(HDR_HAS_L1HDR(hdr));
2388 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2389 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2390 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2391 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2394 arc_state_t *state = hdr->b_l1hdr.b_state;
2396 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2397 (state != arc_anon)) {
2398 /* We don't use the L2-only state list. */
2399 if (state != arc_l2c_only) {
2400 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2402 arc_evictable_space_decrement(hdr, state);
2404 /* remove the prefetch flag if we get a reference */
2405 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2410 * Remove a reference from this hdr. When the reference transitions from
2411 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2412 * list making it eligible for eviction.
2415 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2418 arc_state_t *state = hdr->b_l1hdr.b_state;
2420 ASSERT(HDR_HAS_L1HDR(hdr));
2421 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2422 ASSERT(!GHOST_STATE(state));
2425 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2426 * check to prevent usage of the arc_l2c_only list.
2428 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2429 (state != arc_anon)) {
2430 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2431 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2432 arc_evictable_space_increment(hdr, state);
2438 * Move the supplied buffer to the indicated state. The hash lock
2439 * for the buffer must be held by the caller.
2442 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2443 kmutex_t *hash_lock)
2445 arc_state_t *old_state;
2448 boolean_t update_old, update_new;
2449 arc_buf_contents_t buftype = arc_buf_type(hdr);
2452 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2453 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2454 * L1 hdr doesn't always exist when we change state to arc_anon before
2455 * destroying a header, in which case reallocating to add the L1 hdr is
2458 if (HDR_HAS_L1HDR(hdr)) {
2459 old_state = hdr->b_l1hdr.b_state;
2460 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2461 bufcnt = hdr->b_l1hdr.b_bufcnt;
2462 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2464 old_state = arc_l2c_only;
2467 update_old = B_FALSE;
2469 update_new = update_old;
2471 ASSERT(MUTEX_HELD(hash_lock));
2472 ASSERT3P(new_state, !=, old_state);
2473 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2474 ASSERT(old_state != arc_anon || bufcnt <= 1);
2477 * If this buffer is evictable, transfer it from the
2478 * old state list to the new state list.
2481 if (old_state != arc_anon && old_state != arc_l2c_only) {
2482 ASSERT(HDR_HAS_L1HDR(hdr));
2483 multilist_remove(old_state->arcs_list[buftype], hdr);
2485 if (GHOST_STATE(old_state)) {
2487 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2488 update_old = B_TRUE;
2490 arc_evictable_space_decrement(hdr, old_state);
2492 if (new_state != arc_anon && new_state != arc_l2c_only) {
2495 * An L1 header always exists here, since if we're
2496 * moving to some L1-cached state (i.e. not l2c_only or
2497 * anonymous), we realloc the header to add an L1hdr
2500 ASSERT(HDR_HAS_L1HDR(hdr));
2501 multilist_insert(new_state->arcs_list[buftype], hdr);
2503 if (GHOST_STATE(new_state)) {
2505 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2506 update_new = B_TRUE;
2508 arc_evictable_space_increment(hdr, new_state);
2512 ASSERT(!HDR_EMPTY(hdr));
2513 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2514 buf_hash_remove(hdr);
2516 /* adjust state sizes (ignore arc_l2c_only) */
2518 if (update_new && new_state != arc_l2c_only) {
2519 ASSERT(HDR_HAS_L1HDR(hdr));
2520 if (GHOST_STATE(new_state)) {
2524 * When moving a header to a ghost state, we first
2525 * remove all arc buffers. Thus, we'll have a
2526 * bufcnt of zero, and no arc buffer to use for
2527 * the reference. As a result, we use the arc
2528 * header pointer for the reference.
2530 (void) refcount_add_many(&new_state->arcs_size,
2531 HDR_GET_LSIZE(hdr), hdr);
2532 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2534 uint32_t buffers = 0;
2537 * Each individual buffer holds a unique reference,
2538 * thus we must remove each of these references one
2541 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2542 buf = buf->b_next) {
2543 ASSERT3U(bufcnt, !=, 0);
2547 * When the arc_buf_t is sharing the data
2548 * block with the hdr, the owner of the
2549 * reference belongs to the hdr. Only
2550 * add to the refcount if the arc_buf_t is
2553 if (arc_buf_is_shared(buf))
2556 (void) refcount_add_many(&new_state->arcs_size,
2557 arc_buf_size(buf), buf);
2559 ASSERT3U(bufcnt, ==, buffers);
2561 if (hdr->b_l1hdr.b_pabd != NULL) {
2562 (void) refcount_add_many(&new_state->arcs_size,
2563 arc_hdr_size(hdr), hdr);
2565 ASSERT(GHOST_STATE(old_state));
2570 if (update_old && old_state != arc_l2c_only) {
2571 ASSERT(HDR_HAS_L1HDR(hdr));
2572 if (GHOST_STATE(old_state)) {
2574 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2577 * When moving a header off of a ghost state,
2578 * the header will not contain any arc buffers.
2579 * We use the arc header pointer for the reference
2580 * which is exactly what we did when we put the
2581 * header on the ghost state.
2584 (void) refcount_remove_many(&old_state->arcs_size,
2585 HDR_GET_LSIZE(hdr), hdr);
2587 uint32_t buffers = 0;
2590 * Each individual buffer holds a unique reference,
2591 * thus we must remove each of these references one
2594 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2595 buf = buf->b_next) {
2596 ASSERT3U(bufcnt, !=, 0);
2600 * When the arc_buf_t is sharing the data
2601 * block with the hdr, the owner of the
2602 * reference belongs to the hdr. Only
2603 * add to the refcount if the arc_buf_t is
2606 if (arc_buf_is_shared(buf))
2609 (void) refcount_remove_many(
2610 &old_state->arcs_size, arc_buf_size(buf),
2613 ASSERT3U(bufcnt, ==, buffers);
2614 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2615 (void) refcount_remove_many(
2616 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2620 if (HDR_HAS_L1HDR(hdr))
2621 hdr->b_l1hdr.b_state = new_state;
2624 * L2 headers should never be on the L2 state list since they don't
2625 * have L1 headers allocated.
2627 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2628 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2632 arc_space_consume(uint64_t space, arc_space_type_t type)
2634 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2637 case ARC_SPACE_DATA:
2638 ARCSTAT_INCR(arcstat_data_size, space);
2640 case ARC_SPACE_META:
2641 ARCSTAT_INCR(arcstat_metadata_size, space);
2643 case ARC_SPACE_OTHER:
2644 ARCSTAT_INCR(arcstat_other_size, space);
2646 case ARC_SPACE_HDRS:
2647 ARCSTAT_INCR(arcstat_hdr_size, space);
2649 case ARC_SPACE_L2HDRS:
2650 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
2654 if (type != ARC_SPACE_DATA)
2655 ARCSTAT_INCR(arcstat_meta_used, space);
2657 atomic_add_64(&arc_size, space);
2661 arc_space_return(uint64_t space, arc_space_type_t type)
2663 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2666 case ARC_SPACE_DATA:
2667 ARCSTAT_INCR(arcstat_data_size, -space);
2669 case ARC_SPACE_META:
2670 ARCSTAT_INCR(arcstat_metadata_size, -space);
2672 case ARC_SPACE_OTHER:
2673 ARCSTAT_INCR(arcstat_other_size, -space);
2675 case ARC_SPACE_HDRS:
2676 ARCSTAT_INCR(arcstat_hdr_size, -space);
2678 case ARC_SPACE_L2HDRS:
2679 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
2683 if (type != ARC_SPACE_DATA) {
2684 ASSERT(arc_meta_used >= space);
2685 if (arc_meta_max < arc_meta_used)
2686 arc_meta_max = arc_meta_used;
2687 ARCSTAT_INCR(arcstat_meta_used, -space);
2690 ASSERT(arc_size >= space);
2691 atomic_add_64(&arc_size, -space);
2695 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2696 * with the hdr's b_pabd.
2699 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2702 * The criteria for sharing a hdr's data are:
2703 * 1. the hdr's compression matches the buf's compression
2704 * 2. the hdr doesn't need to be byteswapped
2705 * 3. the hdr isn't already being shared
2706 * 4. the buf is either compressed or it is the last buf in the hdr list
2708 * Criterion #4 maintains the invariant that shared uncompressed
2709 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2710 * might ask, "if a compressed buf is allocated first, won't that be the
2711 * last thing in the list?", but in that case it's impossible to create
2712 * a shared uncompressed buf anyway (because the hdr must be compressed
2713 * to have the compressed buf). You might also think that #3 is
2714 * sufficient to make this guarantee, however it's possible
2715 * (specifically in the rare L2ARC write race mentioned in
2716 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2717 * is sharable, but wasn't at the time of its allocation. Rather than
2718 * allow a new shared uncompressed buf to be created and then shuffle
2719 * the list around to make it the last element, this simply disallows
2720 * sharing if the new buf isn't the first to be added.
2722 ASSERT3P(buf->b_hdr, ==, hdr);
2723 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2724 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2725 return (buf_compressed == hdr_compressed &&
2726 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2727 !HDR_SHARED_DATA(hdr) &&
2728 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2732 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2733 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2734 * copy was made successfully, or an error code otherwise.
2737 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2738 boolean_t fill, arc_buf_t **ret)
2742 ASSERT(HDR_HAS_L1HDR(hdr));
2743 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2744 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2745 hdr->b_type == ARC_BUFC_METADATA);
2746 ASSERT3P(ret, !=, NULL);
2747 ASSERT3P(*ret, ==, NULL);
2749 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2752 buf->b_next = hdr->b_l1hdr.b_buf;
2755 add_reference(hdr, tag);
2758 * We're about to change the hdr's b_flags. We must either
2759 * hold the hash_lock or be undiscoverable.
2761 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2764 * Only honor requests for compressed bufs if the hdr is actually
2767 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2768 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2771 * If the hdr's data can be shared then we share the data buffer and
2772 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2773 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2774 * buffer to store the buf's data.
2776 * There are two additional restrictions here because we're sharing
2777 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2778 * actively involved in an L2ARC write, because if this buf is used by
2779 * an arc_write() then the hdr's data buffer will be released when the
2780 * write completes, even though the L2ARC write might still be using it.
2781 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2782 * need to be ABD-aware.
2784 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2785 abd_is_linear(hdr->b_l1hdr.b_pabd);
2787 /* Set up b_data and sharing */
2789 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2790 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2791 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2794 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2795 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2797 VERIFY3P(buf->b_data, !=, NULL);
2799 hdr->b_l1hdr.b_buf = buf;
2800 hdr->b_l1hdr.b_bufcnt += 1;
2803 * If the user wants the data from the hdr, we need to either copy or
2804 * decompress the data.
2807 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2813 static char *arc_onloan_tag = "onloan";
2816 arc_loaned_bytes_update(int64_t delta)
2818 atomic_add_64(&arc_loaned_bytes, delta);
2820 /* assert that it did not wrap around */
2821 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2825 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2826 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2827 * buffers must be returned to the arc before they can be used by the DMU or
2831 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2833 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2834 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2836 arc_loaned_bytes_update(size);
2842 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2843 enum zio_compress compression_type)
2845 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2846 psize, lsize, compression_type);
2848 arc_loaned_bytes_update(psize);
2855 * Return a loaned arc buffer to the arc.
2858 arc_return_buf(arc_buf_t *buf, void *tag)
2860 arc_buf_hdr_t *hdr = buf->b_hdr;
2862 ASSERT3P(buf->b_data, !=, NULL);
2863 ASSERT(HDR_HAS_L1HDR(hdr));
2864 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2865 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2867 arc_loaned_bytes_update(-arc_buf_size(buf));
2870 /* Detach an arc_buf from a dbuf (tag) */
2872 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2874 arc_buf_hdr_t *hdr = buf->b_hdr;
2876 ASSERT3P(buf->b_data, !=, NULL);
2877 ASSERT(HDR_HAS_L1HDR(hdr));
2878 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2879 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2881 arc_loaned_bytes_update(arc_buf_size(buf));
2885 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2887 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2890 df->l2df_size = size;
2891 df->l2df_type = type;
2892 mutex_enter(&l2arc_free_on_write_mtx);
2893 list_insert_head(l2arc_free_on_write, df);
2894 mutex_exit(&l2arc_free_on_write_mtx);
2898 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2900 arc_state_t *state = hdr->b_l1hdr.b_state;
2901 arc_buf_contents_t type = arc_buf_type(hdr);
2902 uint64_t size = arc_hdr_size(hdr);
2904 /* protected by hash lock, if in the hash table */
2905 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2906 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2907 ASSERT(state != arc_anon && state != arc_l2c_only);
2909 (void) refcount_remove_many(&state->arcs_esize[type],
2912 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2913 if (type == ARC_BUFC_METADATA) {
2914 arc_space_return(size, ARC_SPACE_META);
2916 ASSERT(type == ARC_BUFC_DATA);
2917 arc_space_return(size, ARC_SPACE_DATA);
2920 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2924 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2925 * data buffer, we transfer the refcount ownership to the hdr and update
2926 * the appropriate kstats.
2929 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2931 arc_state_t *state = hdr->b_l1hdr.b_state;
2933 ASSERT(arc_can_share(hdr, buf));
2934 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2935 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2938 * Start sharing the data buffer. We transfer the
2939 * refcount ownership to the hdr since it always owns
2940 * the refcount whenever an arc_buf_t is shared.
2942 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2943 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2944 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2945 HDR_ISTYPE_METADATA(hdr));
2946 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2947 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2950 * Since we've transferred ownership to the hdr we need
2951 * to increment its compressed and uncompressed kstats and
2952 * decrement the overhead size.
2954 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2955 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2956 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2960 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2962 arc_state_t *state = hdr->b_l1hdr.b_state;
2964 ASSERT(arc_buf_is_shared(buf));
2965 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2966 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2969 * We are no longer sharing this buffer so we need
2970 * to transfer its ownership to the rightful owner.
2972 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
2973 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2974 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
2975 abd_put(hdr->b_l1hdr.b_pabd);
2976 hdr->b_l1hdr.b_pabd = NULL;
2977 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2980 * Since the buffer is no longer shared between
2981 * the arc buf and the hdr, count it as overhead.
2983 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2984 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2985 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2989 * Remove an arc_buf_t from the hdr's buf list and return the last
2990 * arc_buf_t on the list. If no buffers remain on the list then return
2994 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2996 ASSERT(HDR_HAS_L1HDR(hdr));
2997 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2999 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3000 arc_buf_t *lastbuf = NULL;
3003 * Remove the buf from the hdr list and locate the last
3004 * remaining buffer on the list.
3006 while (*bufp != NULL) {
3008 *bufp = buf->b_next;
3011 * If we've removed a buffer in the middle of
3012 * the list then update the lastbuf and update
3015 if (*bufp != NULL) {
3017 bufp = &(*bufp)->b_next;
3021 ASSERT3P(lastbuf, !=, buf);
3022 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3023 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3024 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3030 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3034 arc_buf_destroy_impl(arc_buf_t *buf)
3036 arc_buf_hdr_t *hdr = buf->b_hdr;
3039 * Free up the data associated with the buf but only if we're not
3040 * sharing this with the hdr. If we are sharing it with the hdr, the
3041 * hdr is responsible for doing the free.
3043 if (buf->b_data != NULL) {
3045 * We're about to change the hdr's b_flags. We must either
3046 * hold the hash_lock or be undiscoverable.
3048 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3050 arc_cksum_verify(buf);
3052 arc_buf_unwatch(buf);
3055 if (arc_buf_is_shared(buf)) {
3056 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3058 uint64_t size = arc_buf_size(buf);
3059 arc_free_data_buf(hdr, buf->b_data, size, buf);
3060 ARCSTAT_INCR(arcstat_overhead_size, -size);
3064 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3065 hdr->b_l1hdr.b_bufcnt -= 1;
3068 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3070 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3072 * If the current arc_buf_t is sharing its data buffer with the
3073 * hdr, then reassign the hdr's b_pabd to share it with the new
3074 * buffer at the end of the list. The shared buffer is always
3075 * the last one on the hdr's buffer list.
3077 * There is an equivalent case for compressed bufs, but since
3078 * they aren't guaranteed to be the last buf in the list and
3079 * that is an exceedingly rare case, we just allow that space be
3080 * wasted temporarily.
3082 if (lastbuf != NULL) {
3083 /* Only one buf can be shared at once */
3084 VERIFY(!arc_buf_is_shared(lastbuf));
3085 /* hdr is uncompressed so can't have compressed buf */
3086 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3088 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3089 arc_hdr_free_pabd(hdr);
3092 * We must setup a new shared block between the
3093 * last buffer and the hdr. The data would have
3094 * been allocated by the arc buf so we need to transfer
3095 * ownership to the hdr since it's now being shared.
3097 arc_share_buf(hdr, lastbuf);
3099 } else if (HDR_SHARED_DATA(hdr)) {
3101 * Uncompressed shared buffers are always at the end
3102 * of the list. Compressed buffers don't have the
3103 * same requirements. This makes it hard to
3104 * simply assert that the lastbuf is shared so
3105 * we rely on the hdr's compression flags to determine
3106 * if we have a compressed, shared buffer.
3108 ASSERT3P(lastbuf, !=, NULL);
3109 ASSERT(arc_buf_is_shared(lastbuf) ||
3110 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
3114 * Free the checksum if we're removing the last uncompressed buf from
3117 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3118 arc_cksum_free(hdr);
3121 /* clean up the buf */
3123 kmem_cache_free(buf_cache, buf);
3127 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
3129 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3130 ASSERT(HDR_HAS_L1HDR(hdr));
3131 ASSERT(!HDR_SHARED_DATA(hdr));
3133 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3134 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
3135 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3136 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3138 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3139 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3143 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3145 ASSERT(HDR_HAS_L1HDR(hdr));
3146 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3149 * If the hdr is currently being written to the l2arc then
3150 * we defer freeing the data by adding it to the l2arc_free_on_write
3151 * list. The l2arc will free the data once it's finished
3152 * writing it to the l2arc device.
3154 if (HDR_L2_WRITING(hdr)) {
3155 arc_hdr_free_on_write(hdr);
3156 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3158 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3159 arc_hdr_size(hdr), hdr);
3161 hdr->b_l1hdr.b_pabd = NULL;
3162 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3164 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3165 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3168 static arc_buf_hdr_t *
3169 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3170 enum zio_compress compression_type, arc_buf_contents_t type)
3174 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3176 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3177 ASSERT(HDR_EMPTY(hdr));
3178 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3179 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3180 HDR_SET_PSIZE(hdr, psize);
3181 HDR_SET_LSIZE(hdr, lsize);
3185 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3186 arc_hdr_set_compress(hdr, compression_type);
3188 hdr->b_l1hdr.b_state = arc_anon;
3189 hdr->b_l1hdr.b_arc_access = 0;
3190 hdr->b_l1hdr.b_bufcnt = 0;
3191 hdr->b_l1hdr.b_buf = NULL;
3194 * Allocate the hdr's buffer. This will contain either
3195 * the compressed or uncompressed data depending on the block
3196 * it references and compressed arc enablement.
3198 arc_hdr_alloc_pabd(hdr);
3199 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3205 * Transition between the two allocation states for the arc_buf_hdr struct.
3206 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3207 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3208 * version is used when a cache buffer is only in the L2ARC in order to reduce
3211 static arc_buf_hdr_t *
3212 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3214 ASSERT(HDR_HAS_L2HDR(hdr));
3216 arc_buf_hdr_t *nhdr;
3217 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3219 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3220 (old == hdr_l2only_cache && new == hdr_full_cache));
3222 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3224 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3225 buf_hash_remove(hdr);
3227 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3229 if (new == hdr_full_cache) {
3230 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3232 * arc_access and arc_change_state need to be aware that a
3233 * header has just come out of L2ARC, so we set its state to
3234 * l2c_only even though it's about to change.
3236 nhdr->b_l1hdr.b_state = arc_l2c_only;
3238 /* Verify previous threads set to NULL before freeing */
3239 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3241 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3242 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3243 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3246 * If we've reached here, We must have been called from
3247 * arc_evict_hdr(), as such we should have already been
3248 * removed from any ghost list we were previously on
3249 * (which protects us from racing with arc_evict_state),
3250 * thus no locking is needed during this check.
3252 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3255 * A buffer must not be moved into the arc_l2c_only
3256 * state if it's not finished being written out to the
3257 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3258 * might try to be accessed, even though it was removed.
3260 VERIFY(!HDR_L2_WRITING(hdr));
3261 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3264 if (hdr->b_l1hdr.b_thawed != NULL) {
3265 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3266 hdr->b_l1hdr.b_thawed = NULL;
3270 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3273 * The header has been reallocated so we need to re-insert it into any
3276 (void) buf_hash_insert(nhdr, NULL);
3278 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3280 mutex_enter(&dev->l2ad_mtx);
3283 * We must place the realloc'ed header back into the list at
3284 * the same spot. Otherwise, if it's placed earlier in the list,
3285 * l2arc_write_buffers() could find it during the function's
3286 * write phase, and try to write it out to the l2arc.
3288 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3289 list_remove(&dev->l2ad_buflist, hdr);
3291 mutex_exit(&dev->l2ad_mtx);
3294 * Since we're using the pointer address as the tag when
3295 * incrementing and decrementing the l2ad_alloc refcount, we
3296 * must remove the old pointer (that we're about to destroy) and
3297 * add the new pointer to the refcount. Otherwise we'd remove
3298 * the wrong pointer address when calling arc_hdr_destroy() later.
3301 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3302 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3304 buf_discard_identity(hdr);
3305 kmem_cache_free(old, hdr);
3311 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3312 * The buf is returned thawed since we expect the consumer to modify it.
3315 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3317 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3318 ZIO_COMPRESS_OFF, type);
3319 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3321 arc_buf_t *buf = NULL;
3322 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3329 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3330 * for bufs containing metadata.
3333 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3334 enum zio_compress compression_type)
3336 ASSERT3U(lsize, >, 0);
3337 ASSERT3U(lsize, >=, psize);
3338 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3339 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3341 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3342 compression_type, ARC_BUFC_DATA);
3343 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3345 arc_buf_t *buf = NULL;
3346 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3348 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3350 if (!arc_buf_is_shared(buf)) {
3352 * To ensure that the hdr has the correct data in it if we call
3353 * arc_decompress() on this buf before it's been written to
3354 * disk, it's easiest if we just set up sharing between the
3357 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3358 arc_hdr_free_pabd(hdr);
3359 arc_share_buf(hdr, buf);
3366 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3368 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3369 l2arc_dev_t *dev = l2hdr->b_dev;
3370 uint64_t psize = arc_hdr_size(hdr);
3372 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3373 ASSERT(HDR_HAS_L2HDR(hdr));
3375 list_remove(&dev->l2ad_buflist, hdr);
3377 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3378 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3380 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3382 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3383 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3387 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3389 if (HDR_HAS_L1HDR(hdr)) {
3390 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3391 hdr->b_l1hdr.b_bufcnt > 0);
3392 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3393 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3395 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3396 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3398 if (!HDR_EMPTY(hdr))
3399 buf_discard_identity(hdr);
3401 if (HDR_HAS_L2HDR(hdr)) {
3402 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3403 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3406 mutex_enter(&dev->l2ad_mtx);
3409 * Even though we checked this conditional above, we
3410 * need to check this again now that we have the
3411 * l2ad_mtx. This is because we could be racing with
3412 * another thread calling l2arc_evict() which might have
3413 * destroyed this header's L2 portion as we were waiting
3414 * to acquire the l2ad_mtx. If that happens, we don't
3415 * want to re-destroy the header's L2 portion.
3417 if (HDR_HAS_L2HDR(hdr)) {
3419 arc_hdr_l2hdr_destroy(hdr);
3423 mutex_exit(&dev->l2ad_mtx);
3426 if (HDR_HAS_L1HDR(hdr)) {
3427 arc_cksum_free(hdr);
3429 while (hdr->b_l1hdr.b_buf != NULL)
3430 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3433 if (hdr->b_l1hdr.b_thawed != NULL) {
3434 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3435 hdr->b_l1hdr.b_thawed = NULL;
3439 if (hdr->b_l1hdr.b_pabd != NULL) {
3440 arc_hdr_free_pabd(hdr);
3444 ASSERT3P(hdr->b_hash_next, ==, NULL);
3445 if (HDR_HAS_L1HDR(hdr)) {
3446 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3447 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3448 kmem_cache_free(hdr_full_cache, hdr);
3450 kmem_cache_free(hdr_l2only_cache, hdr);
3455 arc_buf_destroy(arc_buf_t *buf, void* tag)
3457 arc_buf_hdr_t *hdr = buf->b_hdr;
3458 kmutex_t *hash_lock = HDR_LOCK(hdr);
3460 if (hdr->b_l1hdr.b_state == arc_anon) {
3461 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3462 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3463 VERIFY0(remove_reference(hdr, NULL, tag));
3464 arc_hdr_destroy(hdr);
3468 mutex_enter(hash_lock);
3469 ASSERT3P(hdr, ==, buf->b_hdr);
3470 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3471 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3472 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3473 ASSERT3P(buf->b_data, !=, NULL);
3475 (void) remove_reference(hdr, hash_lock, tag);
3476 arc_buf_destroy_impl(buf);
3477 mutex_exit(hash_lock);
3481 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3482 * state of the header is dependent on its state prior to entering this
3483 * function. The following transitions are possible:
3485 * - arc_mru -> arc_mru_ghost
3486 * - arc_mfu -> arc_mfu_ghost
3487 * - arc_mru_ghost -> arc_l2c_only
3488 * - arc_mru_ghost -> deleted
3489 * - arc_mfu_ghost -> arc_l2c_only
3490 * - arc_mfu_ghost -> deleted
3493 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3495 arc_state_t *evicted_state, *state;
3496 int64_t bytes_evicted = 0;
3498 ASSERT(MUTEX_HELD(hash_lock));
3499 ASSERT(HDR_HAS_L1HDR(hdr));
3501 state = hdr->b_l1hdr.b_state;
3502 if (GHOST_STATE(state)) {
3503 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3504 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3507 * l2arc_write_buffers() relies on a header's L1 portion
3508 * (i.e. its b_pabd field) during it's write phase.
3509 * Thus, we cannot push a header onto the arc_l2c_only
3510 * state (removing it's L1 piece) until the header is
3511 * done being written to the l2arc.
3513 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3514 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3515 return (bytes_evicted);
3518 ARCSTAT_BUMP(arcstat_deleted);
3519 bytes_evicted += HDR_GET_LSIZE(hdr);
3521 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3523 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3524 if (HDR_HAS_L2HDR(hdr)) {
3526 * This buffer is cached on the 2nd Level ARC;
3527 * don't destroy the header.
3529 arc_change_state(arc_l2c_only, hdr, hash_lock);
3531 * dropping from L1+L2 cached to L2-only,
3532 * realloc to remove the L1 header.
3534 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3537 arc_change_state(arc_anon, hdr, hash_lock);
3538 arc_hdr_destroy(hdr);
3540 return (bytes_evicted);
3543 ASSERT(state == arc_mru || state == arc_mfu);
3544 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3546 /* prefetch buffers have a minimum lifespan */
3547 if (HDR_IO_IN_PROGRESS(hdr) ||
3548 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3549 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3550 arc_min_prefetch_lifespan)) {
3551 ARCSTAT_BUMP(arcstat_evict_skip);
3552 return (bytes_evicted);
3555 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3556 while (hdr->b_l1hdr.b_buf) {
3557 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3558 if (!mutex_tryenter(&buf->b_evict_lock)) {
3559 ARCSTAT_BUMP(arcstat_mutex_miss);
3562 if (buf->b_data != NULL)
3563 bytes_evicted += HDR_GET_LSIZE(hdr);
3564 mutex_exit(&buf->b_evict_lock);
3565 arc_buf_destroy_impl(buf);
3568 if (HDR_HAS_L2HDR(hdr)) {
3569 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3571 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3572 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3573 HDR_GET_LSIZE(hdr));
3575 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3576 HDR_GET_LSIZE(hdr));
3580 if (hdr->b_l1hdr.b_bufcnt == 0) {
3581 arc_cksum_free(hdr);
3583 bytes_evicted += arc_hdr_size(hdr);
3586 * If this hdr is being evicted and has a compressed
3587 * buffer then we discard it here before we change states.
3588 * This ensures that the accounting is updated correctly
3589 * in arc_free_data_impl().
3591 arc_hdr_free_pabd(hdr);
3593 arc_change_state(evicted_state, hdr, hash_lock);
3594 ASSERT(HDR_IN_HASH_TABLE(hdr));
3595 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3596 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3599 return (bytes_evicted);
3603 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3604 uint64_t spa, int64_t bytes)
3606 multilist_sublist_t *mls;
3607 uint64_t bytes_evicted = 0;
3609 kmutex_t *hash_lock;
3610 int evict_count = 0;
3612 ASSERT3P(marker, !=, NULL);
3613 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3615 mls = multilist_sublist_lock(ml, idx);
3617 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3618 hdr = multilist_sublist_prev(mls, marker)) {
3619 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3620 (evict_count >= zfs_arc_evict_batch_limit))
3624 * To keep our iteration location, move the marker
3625 * forward. Since we're not holding hdr's hash lock, we
3626 * must be very careful and not remove 'hdr' from the
3627 * sublist. Otherwise, other consumers might mistake the
3628 * 'hdr' as not being on a sublist when they call the
3629 * multilist_link_active() function (they all rely on
3630 * the hash lock protecting concurrent insertions and
3631 * removals). multilist_sublist_move_forward() was
3632 * specifically implemented to ensure this is the case
3633 * (only 'marker' will be removed and re-inserted).
3635 multilist_sublist_move_forward(mls, marker);
3638 * The only case where the b_spa field should ever be
3639 * zero, is the marker headers inserted by
3640 * arc_evict_state(). It's possible for multiple threads
3641 * to be calling arc_evict_state() concurrently (e.g.
3642 * dsl_pool_close() and zio_inject_fault()), so we must
3643 * skip any markers we see from these other threads.
3645 if (hdr->b_spa == 0)
3648 /* we're only interested in evicting buffers of a certain spa */
3649 if (spa != 0 && hdr->b_spa != spa) {
3650 ARCSTAT_BUMP(arcstat_evict_skip);
3654 hash_lock = HDR_LOCK(hdr);
3657 * We aren't calling this function from any code path
3658 * that would already be holding a hash lock, so we're
3659 * asserting on this assumption to be defensive in case
3660 * this ever changes. Without this check, it would be
3661 * possible to incorrectly increment arcstat_mutex_miss
3662 * below (e.g. if the code changed such that we called
3663 * this function with a hash lock held).
3665 ASSERT(!MUTEX_HELD(hash_lock));
3667 if (mutex_tryenter(hash_lock)) {
3668 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3669 mutex_exit(hash_lock);
3671 bytes_evicted += evicted;
3674 * If evicted is zero, arc_evict_hdr() must have
3675 * decided to skip this header, don't increment
3676 * evict_count in this case.
3682 * If arc_size isn't overflowing, signal any
3683 * threads that might happen to be waiting.
3685 * For each header evicted, we wake up a single
3686 * thread. If we used cv_broadcast, we could
3687 * wake up "too many" threads causing arc_size
3688 * to significantly overflow arc_c; since
3689 * arc_get_data_impl() doesn't check for overflow
3690 * when it's woken up (it doesn't because it's
3691 * possible for the ARC to be overflowing while
3692 * full of un-evictable buffers, and the
3693 * function should proceed in this case).
3695 * If threads are left sleeping, due to not
3696 * using cv_broadcast, they will be woken up
3697 * just before arc_reclaim_thread() sleeps.
3699 mutex_enter(&arc_reclaim_lock);
3700 if (!arc_is_overflowing())
3701 cv_signal(&arc_reclaim_waiters_cv);
3702 mutex_exit(&arc_reclaim_lock);
3704 ARCSTAT_BUMP(arcstat_mutex_miss);
3708 multilist_sublist_unlock(mls);
3710 return (bytes_evicted);
3714 * Evict buffers from the given arc state, until we've removed the
3715 * specified number of bytes. Move the removed buffers to the
3716 * appropriate evict state.
3718 * This function makes a "best effort". It skips over any buffers
3719 * it can't get a hash_lock on, and so, may not catch all candidates.
3720 * It may also return without evicting as much space as requested.
3722 * If bytes is specified using the special value ARC_EVICT_ALL, this
3723 * will evict all available (i.e. unlocked and evictable) buffers from
3724 * the given arc state; which is used by arc_flush().
3727 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3728 arc_buf_contents_t type)
3730 uint64_t total_evicted = 0;
3731 multilist_t *ml = state->arcs_list[type];
3733 arc_buf_hdr_t **markers;
3735 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3737 num_sublists = multilist_get_num_sublists(ml);
3740 * If we've tried to evict from each sublist, made some
3741 * progress, but still have not hit the target number of bytes
3742 * to evict, we want to keep trying. The markers allow us to
3743 * pick up where we left off for each individual sublist, rather
3744 * than starting from the tail each time.
3746 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3747 for (int i = 0; i < num_sublists; i++) {
3748 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3751 * A b_spa of 0 is used to indicate that this header is
3752 * a marker. This fact is used in arc_adjust_type() and
3753 * arc_evict_state_impl().
3755 markers[i]->b_spa = 0;
3757 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3758 multilist_sublist_insert_tail(mls, markers[i]);
3759 multilist_sublist_unlock(mls);
3763 * While we haven't hit our target number of bytes to evict, or
3764 * we're evicting all available buffers.
3766 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3768 * Start eviction using a randomly selected sublist,
3769 * this is to try and evenly balance eviction across all
3770 * sublists. Always starting at the same sublist
3771 * (e.g. index 0) would cause evictions to favor certain
3772 * sublists over others.
3774 int sublist_idx = multilist_get_random_index(ml);
3775 uint64_t scan_evicted = 0;
3777 for (int i = 0; i < num_sublists; i++) {
3778 uint64_t bytes_remaining;
3779 uint64_t bytes_evicted;
3781 if (bytes == ARC_EVICT_ALL)
3782 bytes_remaining = ARC_EVICT_ALL;
3783 else if (total_evicted < bytes)
3784 bytes_remaining = bytes - total_evicted;
3788 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3789 markers[sublist_idx], spa, bytes_remaining);
3791 scan_evicted += bytes_evicted;
3792 total_evicted += bytes_evicted;
3794 /* we've reached the end, wrap to the beginning */
3795 if (++sublist_idx >= num_sublists)
3800 * If we didn't evict anything during this scan, we have
3801 * no reason to believe we'll evict more during another
3802 * scan, so break the loop.
3804 if (scan_evicted == 0) {
3805 /* This isn't possible, let's make that obvious */
3806 ASSERT3S(bytes, !=, 0);
3809 * When bytes is ARC_EVICT_ALL, the only way to
3810 * break the loop is when scan_evicted is zero.
3811 * In that case, we actually have evicted enough,
3812 * so we don't want to increment the kstat.
3814 if (bytes != ARC_EVICT_ALL) {
3815 ASSERT3S(total_evicted, <, bytes);
3816 ARCSTAT_BUMP(arcstat_evict_not_enough);
3823 for (int i = 0; i < num_sublists; i++) {
3824 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3825 multilist_sublist_remove(mls, markers[i]);
3826 multilist_sublist_unlock(mls);
3828 kmem_cache_free(hdr_full_cache, markers[i]);
3830 kmem_free(markers, sizeof (*markers) * num_sublists);
3832 return (total_evicted);
3836 * Flush all "evictable" data of the given type from the arc state
3837 * specified. This will not evict any "active" buffers (i.e. referenced).
3839 * When 'retry' is set to B_FALSE, the function will make a single pass
3840 * over the state and evict any buffers that it can. Since it doesn't
3841 * continually retry the eviction, it might end up leaving some buffers
3842 * in the ARC due to lock misses.
3844 * When 'retry' is set to B_TRUE, the function will continually retry the
3845 * eviction until *all* evictable buffers have been removed from the
3846 * state. As a result, if concurrent insertions into the state are
3847 * allowed (e.g. if the ARC isn't shutting down), this function might
3848 * wind up in an infinite loop, continually trying to evict buffers.
3851 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3854 uint64_t evicted = 0;
3856 while (refcount_count(&state->arcs_esize[type]) != 0) {
3857 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3867 * Evict the specified number of bytes from the state specified,
3868 * restricting eviction to the spa and type given. This function
3869 * prevents us from trying to evict more from a state's list than
3870 * is "evictable", and to skip evicting altogether when passed a
3871 * negative value for "bytes". In contrast, arc_evict_state() will
3872 * evict everything it can, when passed a negative value for "bytes".
3875 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3876 arc_buf_contents_t type)
3880 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3881 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3882 return (arc_evict_state(state, spa, delta, type));
3889 * Evict metadata buffers from the cache, such that arc_meta_used is
3890 * capped by the arc_meta_limit tunable.
3893 arc_adjust_meta(void)
3895 uint64_t total_evicted = 0;
3899 * If we're over the meta limit, we want to evict enough
3900 * metadata to get back under the meta limit. We don't want to
3901 * evict so much that we drop the MRU below arc_p, though. If
3902 * we're over the meta limit more than we're over arc_p, we
3903 * evict some from the MRU here, and some from the MFU below.
3905 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3906 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3907 refcount_count(&arc_mru->arcs_size) - arc_p));
3909 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3912 * Similar to the above, we want to evict enough bytes to get us
3913 * below the meta limit, but not so much as to drop us below the
3914 * space allotted to the MFU (which is defined as arc_c - arc_p).
3916 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3917 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
3919 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3921 return (total_evicted);
3925 * Return the type of the oldest buffer in the given arc state
3927 * This function will select a random sublist of type ARC_BUFC_DATA and
3928 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3929 * is compared, and the type which contains the "older" buffer will be
3932 static arc_buf_contents_t
3933 arc_adjust_type(arc_state_t *state)
3935 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
3936 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
3937 int data_idx = multilist_get_random_index(data_ml);
3938 int meta_idx = multilist_get_random_index(meta_ml);
3939 multilist_sublist_t *data_mls;
3940 multilist_sublist_t *meta_mls;
3941 arc_buf_contents_t type;
3942 arc_buf_hdr_t *data_hdr;
3943 arc_buf_hdr_t *meta_hdr;
3946 * We keep the sublist lock until we're finished, to prevent
3947 * the headers from being destroyed via arc_evict_state().
3949 data_mls = multilist_sublist_lock(data_ml, data_idx);
3950 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3953 * These two loops are to ensure we skip any markers that
3954 * might be at the tail of the lists due to arc_evict_state().
3957 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3958 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3959 if (data_hdr->b_spa != 0)
3963 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3964 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3965 if (meta_hdr->b_spa != 0)
3969 if (data_hdr == NULL && meta_hdr == NULL) {
3970 type = ARC_BUFC_DATA;
3971 } else if (data_hdr == NULL) {
3972 ASSERT3P(meta_hdr, !=, NULL);
3973 type = ARC_BUFC_METADATA;
3974 } else if (meta_hdr == NULL) {
3975 ASSERT3P(data_hdr, !=, NULL);
3976 type = ARC_BUFC_DATA;
3978 ASSERT3P(data_hdr, !=, NULL);
3979 ASSERT3P(meta_hdr, !=, NULL);
3981 /* The headers can't be on the sublist without an L1 header */
3982 ASSERT(HDR_HAS_L1HDR(data_hdr));
3983 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3985 if (data_hdr->b_l1hdr.b_arc_access <
3986 meta_hdr->b_l1hdr.b_arc_access) {
3987 type = ARC_BUFC_DATA;
3989 type = ARC_BUFC_METADATA;
3993 multilist_sublist_unlock(meta_mls);
3994 multilist_sublist_unlock(data_mls);
4000 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4005 uint64_t total_evicted = 0;
4010 * If we're over arc_meta_limit, we want to correct that before
4011 * potentially evicting data buffers below.
4013 total_evicted += arc_adjust_meta();
4018 * If we're over the target cache size, we want to evict enough
4019 * from the list to get back to our target size. We don't want
4020 * to evict too much from the MRU, such that it drops below
4021 * arc_p. So, if we're over our target cache size more than
4022 * the MRU is over arc_p, we'll evict enough to get back to
4023 * arc_p here, and then evict more from the MFU below.
4025 target = MIN((int64_t)(arc_size - arc_c),
4026 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4027 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
4030 * If we're below arc_meta_min, always prefer to evict data.
4031 * Otherwise, try to satisfy the requested number of bytes to
4032 * evict from the type which contains older buffers; in an
4033 * effort to keep newer buffers in the cache regardless of their
4034 * type. If we cannot satisfy the number of bytes from this
4035 * type, spill over into the next type.
4037 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4038 arc_meta_used > arc_meta_min) {
4039 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4040 total_evicted += bytes;
4043 * If we couldn't evict our target number of bytes from
4044 * metadata, we try to get the rest from data.
4049 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4051 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4052 total_evicted += bytes;
4055 * If we couldn't evict our target number of bytes from
4056 * data, we try to get the rest from metadata.
4061 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4067 * Now that we've tried to evict enough from the MRU to get its
4068 * size back to arc_p, if we're still above the target cache
4069 * size, we evict the rest from the MFU.
4071 target = arc_size - arc_c;
4073 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4074 arc_meta_used > arc_meta_min) {
4075 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4076 total_evicted += bytes;
4079 * If we couldn't evict our target number of bytes from
4080 * metadata, we try to get the rest from data.
4085 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4087 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4088 total_evicted += bytes;
4091 * If we couldn't evict our target number of bytes from
4092 * data, we try to get the rest from data.
4097 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4101 * Adjust ghost lists
4103 * In addition to the above, the ARC also defines target values
4104 * for the ghost lists. The sum of the mru list and mru ghost
4105 * list should never exceed the target size of the cache, and
4106 * the sum of the mru list, mfu list, mru ghost list, and mfu
4107 * ghost list should never exceed twice the target size of the
4108 * cache. The following logic enforces these limits on the ghost
4109 * caches, and evicts from them as needed.
4111 target = refcount_count(&arc_mru->arcs_size) +
4112 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4114 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4115 total_evicted += bytes;
4120 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4123 * We assume the sum of the mru list and mfu list is less than
4124 * or equal to arc_c (we enforced this above), which means we
4125 * can use the simpler of the two equations below:
4127 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4128 * mru ghost + mfu ghost <= arc_c
4130 target = refcount_count(&arc_mru_ghost->arcs_size) +
4131 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4133 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4134 total_evicted += bytes;
4139 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4141 return (total_evicted);
4145 arc_flush(spa_t *spa, boolean_t retry)
4150 * If retry is B_TRUE, a spa must not be specified since we have
4151 * no good way to determine if all of a spa's buffers have been
4152 * evicted from an arc state.
4154 ASSERT(!retry || spa == 0);
4157 guid = spa_load_guid(spa);
4159 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4160 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4162 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4163 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4165 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4166 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4168 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4169 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4173 arc_shrink(int64_t to_free)
4175 if (arc_c > arc_c_min) {
4176 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
4177 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
4178 if (arc_c > arc_c_min + to_free)
4179 atomic_add_64(&arc_c, -to_free);
4183 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4184 if (arc_c > arc_size)
4185 arc_c = MAX(arc_size, arc_c_min);
4187 arc_p = (arc_c >> 1);
4189 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
4192 ASSERT(arc_c >= arc_c_min);
4193 ASSERT((int64_t)arc_p >= 0);
4196 if (arc_size > arc_c) {
4197 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size,
4199 (void) arc_adjust();
4203 typedef enum free_memory_reason_t {
4208 FMR_PAGES_PP_MAXIMUM,
4211 } free_memory_reason_t;
4213 int64_t last_free_memory;
4214 free_memory_reason_t last_free_reason;
4217 * Additional reserve of pages for pp_reserve.
4219 int64_t arc_pages_pp_reserve = 64;
4222 * Additional reserve of pages for swapfs.
4224 int64_t arc_swapfs_reserve = 64;
4227 * Return the amount of memory that can be consumed before reclaim will be
4228 * needed. Positive if there is sufficient free memory, negative indicates
4229 * the amount of memory that needs to be freed up.
4232 arc_available_memory(void)
4234 int64_t lowest = INT64_MAX;
4236 free_memory_reason_t r = FMR_UNKNOWN;
4240 * Cooperate with pagedaemon when it's time for it to scan
4241 * and reclaim some pages.
4243 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
4251 * check that we're out of range of the pageout scanner. It starts to
4252 * schedule paging if freemem is less than lotsfree and needfree.
4253 * lotsfree is the high-water mark for pageout, and needfree is the
4254 * number of needed free pages. We add extra pages here to make sure
4255 * the scanner doesn't start up while we're freeing memory.
4257 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4264 * check to make sure that swapfs has enough space so that anon
4265 * reservations can still succeed. anon_resvmem() checks that the
4266 * availrmem is greater than swapfs_minfree, and the number of reserved
4267 * swap pages. We also add a bit of extra here just to prevent
4268 * circumstances from getting really dire.
4270 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4271 desfree - arc_swapfs_reserve);
4274 r = FMR_SWAPFS_MINFREE;
4279 * Check that we have enough availrmem that memory locking (e.g., via
4280 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4281 * stores the number of pages that cannot be locked; when availrmem
4282 * drops below pages_pp_maximum, page locking mechanisms such as
4283 * page_pp_lock() will fail.)
4285 n = PAGESIZE * (availrmem - pages_pp_maximum -
4286 arc_pages_pp_reserve);
4289 r = FMR_PAGES_PP_MAXIMUM;
4292 #endif /* illumos */
4293 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4295 * If we're on an i386 platform, it's possible that we'll exhaust the
4296 * kernel heap space before we ever run out of available physical
4297 * memory. Most checks of the size of the heap_area compare against
4298 * tune.t_minarmem, which is the minimum available real memory that we
4299 * can have in the system. However, this is generally fixed at 25 pages
4300 * which is so low that it's useless. In this comparison, we seek to
4301 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4302 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4305 n = uma_avail() - (long)(uma_limit() / 4);
4313 * If zio data pages are being allocated out of a separate heap segment,
4314 * then enforce that the size of available vmem for this arena remains
4315 * above about 1/16th free.
4317 * Note: The 1/16th arena free requirement was put in place
4318 * to aggressively evict memory from the arc in order to avoid
4319 * memory fragmentation issues.
4321 if (zio_arena != NULL) {
4322 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4323 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
4331 /* Every 100 calls, free a small amount */
4332 if (spa_get_random(100) == 0)
4334 #endif /* _KERNEL */
4336 last_free_memory = lowest;
4337 last_free_reason = r;
4338 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4344 * Determine if the system is under memory pressure and is asking
4345 * to reclaim memory. A return value of B_TRUE indicates that the system
4346 * is under memory pressure and that the arc should adjust accordingly.
4349 arc_reclaim_needed(void)
4351 return (arc_available_memory() < 0);
4354 extern kmem_cache_t *zio_buf_cache[];
4355 extern kmem_cache_t *zio_data_buf_cache[];
4356 extern kmem_cache_t *range_seg_cache;
4357 extern kmem_cache_t *abd_chunk_cache;
4359 static __noinline void
4360 arc_kmem_reap_now(void)
4363 kmem_cache_t *prev_cache = NULL;
4364 kmem_cache_t *prev_data_cache = NULL;
4366 DTRACE_PROBE(arc__kmem_reap_start);
4368 if (arc_meta_used >= arc_meta_limit) {
4370 * We are exceeding our meta-data cache limit.
4371 * Purge some DNLC entries to release holds on meta-data.
4373 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4377 * Reclaim unused memory from all kmem caches.
4383 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4384 if (zio_buf_cache[i] != prev_cache) {
4385 prev_cache = zio_buf_cache[i];
4386 kmem_cache_reap_now(zio_buf_cache[i]);
4388 if (zio_data_buf_cache[i] != prev_data_cache) {
4389 prev_data_cache = zio_data_buf_cache[i];
4390 kmem_cache_reap_now(zio_data_buf_cache[i]);
4393 kmem_cache_reap_now(abd_chunk_cache);
4394 kmem_cache_reap_now(buf_cache);
4395 kmem_cache_reap_now(hdr_full_cache);
4396 kmem_cache_reap_now(hdr_l2only_cache);
4397 kmem_cache_reap_now(range_seg_cache);
4400 if (zio_arena != NULL) {
4402 * Ask the vmem arena to reclaim unused memory from its
4405 vmem_qcache_reap(zio_arena);
4408 DTRACE_PROBE(arc__kmem_reap_end);
4412 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4413 * enough data and signal them to proceed. When this happens, the threads in
4414 * arc_get_data_impl() are sleeping while holding the hash lock for their
4415 * particular arc header. Thus, we must be careful to never sleep on a
4416 * hash lock in this thread. This is to prevent the following deadlock:
4418 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4419 * waiting for the reclaim thread to signal it.
4421 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4422 * fails, and goes to sleep forever.
4424 * This possible deadlock is avoided by always acquiring a hash lock
4425 * using mutex_tryenter() from arc_reclaim_thread().
4428 arc_reclaim_thread(void *dummy __unused)
4430 hrtime_t growtime = 0;
4433 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4435 mutex_enter(&arc_reclaim_lock);
4436 while (!arc_reclaim_thread_exit) {
4437 uint64_t evicted = 0;
4440 * This is necessary in order for the mdb ::arc dcmd to
4441 * show up to date information. Since the ::arc command
4442 * does not call the kstat's update function, without
4443 * this call, the command may show stale stats for the
4444 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4445 * with this change, the data might be up to 1 second
4446 * out of date; but that should suffice. The arc_state_t
4447 * structures can be queried directly if more accurate
4448 * information is needed.
4450 if (arc_ksp != NULL)
4451 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4453 mutex_exit(&arc_reclaim_lock);
4456 * We call arc_adjust() before (possibly) calling
4457 * arc_kmem_reap_now(), so that we can wake up
4458 * arc_get_data_impl() sooner.
4460 evicted = arc_adjust();
4462 int64_t free_memory = arc_available_memory();
4463 if (free_memory < 0) {
4465 arc_no_grow = B_TRUE;
4469 * Wait at least zfs_grow_retry (default 60) seconds
4470 * before considering growing.
4472 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4474 arc_kmem_reap_now();
4477 * If we are still low on memory, shrink the ARC
4478 * so that we have arc_shrink_min free space.
4480 free_memory = arc_available_memory();
4483 (arc_c >> arc_shrink_shift) - free_memory;
4485 arc_shrink(to_free);
4487 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4488 arc_no_grow = B_TRUE;
4489 } else if (gethrtime() >= growtime) {
4490 arc_no_grow = B_FALSE;
4493 mutex_enter(&arc_reclaim_lock);
4496 * If evicted is zero, we couldn't evict anything via
4497 * arc_adjust(). This could be due to hash lock
4498 * collisions, but more likely due to the majority of
4499 * arc buffers being unevictable. Therefore, even if
4500 * arc_size is above arc_c, another pass is unlikely to
4501 * be helpful and could potentially cause us to enter an
4504 if (arc_size <= arc_c || evicted == 0) {
4506 * We're either no longer overflowing, or we
4507 * can't evict anything more, so we should wake
4508 * up any threads before we go to sleep.
4510 cv_broadcast(&arc_reclaim_waiters_cv);
4513 * Block until signaled, or after one second (we
4514 * might need to perform arc_kmem_reap_now()
4515 * even if we aren't being signalled)
4517 CALLB_CPR_SAFE_BEGIN(&cpr);
4518 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4519 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4520 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4524 arc_reclaim_thread_exit = B_FALSE;
4525 cv_broadcast(&arc_reclaim_thread_cv);
4526 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4530 static u_int arc_dnlc_evicts_arg;
4531 extern struct vfsops zfs_vfsops;
4534 arc_dnlc_evicts_thread(void *dummy __unused)
4539 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4541 mutex_enter(&arc_dnlc_evicts_lock);
4542 while (!arc_dnlc_evicts_thread_exit) {
4543 CALLB_CPR_SAFE_BEGIN(&cpr);
4544 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4545 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4546 if (arc_dnlc_evicts_arg != 0) {
4547 percent = arc_dnlc_evicts_arg;
4548 mutex_exit(&arc_dnlc_evicts_lock);
4550 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4552 mutex_enter(&arc_dnlc_evicts_lock);
4554 * Clear our token only after vnlru_free()
4555 * pass is done, to avoid false queueing of
4558 arc_dnlc_evicts_arg = 0;
4561 arc_dnlc_evicts_thread_exit = FALSE;
4562 cv_broadcast(&arc_dnlc_evicts_cv);
4563 CALLB_CPR_EXIT(&cpr);
4568 dnlc_reduce_cache(void *arg)
4572 percent = (u_int)(uintptr_t)arg;
4573 mutex_enter(&arc_dnlc_evicts_lock);
4574 if (arc_dnlc_evicts_arg == 0) {
4575 arc_dnlc_evicts_arg = percent;
4576 cv_broadcast(&arc_dnlc_evicts_cv);
4578 mutex_exit(&arc_dnlc_evicts_lock);
4582 * Adapt arc info given the number of bytes we are trying to add and
4583 * the state that we are comming from. This function is only called
4584 * when we are adding new content to the cache.
4587 arc_adapt(int bytes, arc_state_t *state)
4590 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4591 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4592 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4594 if (state == arc_l2c_only)
4599 * Adapt the target size of the MRU list:
4600 * - if we just hit in the MRU ghost list, then increase
4601 * the target size of the MRU list.
4602 * - if we just hit in the MFU ghost list, then increase
4603 * the target size of the MFU list by decreasing the
4604 * target size of the MRU list.
4606 if (state == arc_mru_ghost) {
4607 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4608 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4610 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4611 } else if (state == arc_mfu_ghost) {
4614 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4615 mult = MIN(mult, 10);
4617 delta = MIN(bytes * mult, arc_p);
4618 arc_p = MAX(arc_p_min, arc_p - delta);
4620 ASSERT((int64_t)arc_p >= 0);
4622 if (arc_reclaim_needed()) {
4623 cv_signal(&arc_reclaim_thread_cv);
4630 if (arc_c >= arc_c_max)
4634 * If we're within (2 * maxblocksize) bytes of the target
4635 * cache size, increment the target cache size
4637 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
4638 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
4639 atomic_add_64(&arc_c, (int64_t)bytes);
4640 if (arc_c > arc_c_max)
4642 else if (state == arc_anon)
4643 atomic_add_64(&arc_p, (int64_t)bytes);
4647 ASSERT((int64_t)arc_p >= 0);
4651 * Check if arc_size has grown past our upper threshold, determined by
4652 * zfs_arc_overflow_shift.
4655 arc_is_overflowing(void)
4657 /* Always allow at least one block of overflow */
4658 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4659 arc_c >> zfs_arc_overflow_shift);
4661 return (arc_size >= arc_c + overflow);
4665 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4667 arc_buf_contents_t type = arc_buf_type(hdr);
4669 arc_get_data_impl(hdr, size, tag);
4670 if (type == ARC_BUFC_METADATA) {
4671 return (abd_alloc(size, B_TRUE));
4673 ASSERT(type == ARC_BUFC_DATA);
4674 return (abd_alloc(size, B_FALSE));
4679 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4681 arc_buf_contents_t type = arc_buf_type(hdr);
4683 arc_get_data_impl(hdr, size, tag);
4684 if (type == ARC_BUFC_METADATA) {
4685 return (zio_buf_alloc(size));
4687 ASSERT(type == ARC_BUFC_DATA);
4688 return (zio_data_buf_alloc(size));
4693 * Allocate a block and return it to the caller. If we are hitting the
4694 * hard limit for the cache size, we must sleep, waiting for the eviction
4695 * thread to catch up. If we're past the target size but below the hard
4696 * limit, we'll only signal the reclaim thread and continue on.
4699 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4701 arc_state_t *state = hdr->b_l1hdr.b_state;
4702 arc_buf_contents_t type = arc_buf_type(hdr);
4704 arc_adapt(size, state);
4707 * If arc_size is currently overflowing, and has grown past our
4708 * upper limit, we must be adding data faster than the evict
4709 * thread can evict. Thus, to ensure we don't compound the
4710 * problem by adding more data and forcing arc_size to grow even
4711 * further past it's target size, we halt and wait for the
4712 * eviction thread to catch up.
4714 * It's also possible that the reclaim thread is unable to evict
4715 * enough buffers to get arc_size below the overflow limit (e.g.
4716 * due to buffers being un-evictable, or hash lock collisions).
4717 * In this case, we want to proceed regardless if we're
4718 * overflowing; thus we don't use a while loop here.
4720 if (arc_is_overflowing()) {
4721 mutex_enter(&arc_reclaim_lock);
4724 * Now that we've acquired the lock, we may no longer be
4725 * over the overflow limit, lets check.
4727 * We're ignoring the case of spurious wake ups. If that
4728 * were to happen, it'd let this thread consume an ARC
4729 * buffer before it should have (i.e. before we're under
4730 * the overflow limit and were signalled by the reclaim
4731 * thread). As long as that is a rare occurrence, it
4732 * shouldn't cause any harm.
4734 if (arc_is_overflowing()) {
4735 cv_signal(&arc_reclaim_thread_cv);
4736 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4739 mutex_exit(&arc_reclaim_lock);
4742 VERIFY3U(hdr->b_type, ==, type);
4743 if (type == ARC_BUFC_METADATA) {
4744 arc_space_consume(size, ARC_SPACE_META);
4746 arc_space_consume(size, ARC_SPACE_DATA);
4750 * Update the state size. Note that ghost states have a
4751 * "ghost size" and so don't need to be updated.
4753 if (!GHOST_STATE(state)) {
4755 (void) refcount_add_many(&state->arcs_size, size, tag);
4758 * If this is reached via arc_read, the link is
4759 * protected by the hash lock. If reached via
4760 * arc_buf_alloc, the header should not be accessed by
4761 * any other thread. And, if reached via arc_read_done,
4762 * the hash lock will protect it if it's found in the
4763 * hash table; otherwise no other thread should be
4764 * trying to [add|remove]_reference it.
4766 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4767 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4768 (void) refcount_add_many(&state->arcs_esize[type],
4773 * If we are growing the cache, and we are adding anonymous
4774 * data, and we have outgrown arc_p, update arc_p
4776 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
4777 (refcount_count(&arc_anon->arcs_size) +
4778 refcount_count(&arc_mru->arcs_size) > arc_p))
4779 arc_p = MIN(arc_c, arc_p + size);
4781 ARCSTAT_BUMP(arcstat_allocated);
4785 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4787 arc_free_data_impl(hdr, size, tag);
4792 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4794 arc_buf_contents_t type = arc_buf_type(hdr);
4796 arc_free_data_impl(hdr, size, tag);
4797 if (type == ARC_BUFC_METADATA) {
4798 zio_buf_free(buf, size);
4800 ASSERT(type == ARC_BUFC_DATA);
4801 zio_data_buf_free(buf, size);
4806 * Free the arc data buffer.
4809 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4811 arc_state_t *state = hdr->b_l1hdr.b_state;
4812 arc_buf_contents_t type = arc_buf_type(hdr);
4814 /* protected by hash lock, if in the hash table */
4815 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4816 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4817 ASSERT(state != arc_anon && state != arc_l2c_only);
4819 (void) refcount_remove_many(&state->arcs_esize[type],
4822 (void) refcount_remove_many(&state->arcs_size, size, tag);
4824 VERIFY3U(hdr->b_type, ==, type);
4825 if (type == ARC_BUFC_METADATA) {
4826 arc_space_return(size, ARC_SPACE_META);
4828 ASSERT(type == ARC_BUFC_DATA);
4829 arc_space_return(size, ARC_SPACE_DATA);
4834 * This routine is called whenever a buffer is accessed.
4835 * NOTE: the hash lock is dropped in this function.
4838 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4842 ASSERT(MUTEX_HELD(hash_lock));
4843 ASSERT(HDR_HAS_L1HDR(hdr));
4845 if (hdr->b_l1hdr.b_state == arc_anon) {
4847 * This buffer is not in the cache, and does not
4848 * appear in our "ghost" list. Add the new buffer
4852 ASSERT0(hdr->b_l1hdr.b_arc_access);
4853 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4854 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4855 arc_change_state(arc_mru, hdr, hash_lock);
4857 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4858 now = ddi_get_lbolt();
4861 * If this buffer is here because of a prefetch, then either:
4862 * - clear the flag if this is a "referencing" read
4863 * (any subsequent access will bump this into the MFU state).
4865 * - move the buffer to the head of the list if this is
4866 * another prefetch (to make it less likely to be evicted).
4868 if (HDR_PREFETCH(hdr)) {
4869 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4870 /* link protected by hash lock */
4871 ASSERT(multilist_link_active(
4872 &hdr->b_l1hdr.b_arc_node));
4874 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4875 ARCSTAT_BUMP(arcstat_mru_hits);
4877 hdr->b_l1hdr.b_arc_access = now;
4882 * This buffer has been "accessed" only once so far,
4883 * but it is still in the cache. Move it to the MFU
4886 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4888 * More than 125ms have passed since we
4889 * instantiated this buffer. Move it to the
4890 * most frequently used state.
4892 hdr->b_l1hdr.b_arc_access = now;
4893 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4894 arc_change_state(arc_mfu, hdr, hash_lock);
4896 ARCSTAT_BUMP(arcstat_mru_hits);
4897 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
4898 arc_state_t *new_state;
4900 * This buffer has been "accessed" recently, but
4901 * was evicted from the cache. Move it to the
4905 if (HDR_PREFETCH(hdr)) {
4906 new_state = arc_mru;
4907 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
4908 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4909 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4911 new_state = arc_mfu;
4912 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4915 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4916 arc_change_state(new_state, hdr, hash_lock);
4918 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
4919 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
4921 * This buffer has been accessed more than once and is
4922 * still in the cache. Keep it in the MFU state.
4924 * NOTE: an add_reference() that occurred when we did
4925 * the arc_read() will have kicked this off the list.
4926 * If it was a prefetch, we will explicitly move it to
4927 * the head of the list now.
4929 if ((HDR_PREFETCH(hdr)) != 0) {
4930 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4931 /* link protected by hash_lock */
4932 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4934 ARCSTAT_BUMP(arcstat_mfu_hits);
4935 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4936 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4937 arc_state_t *new_state = arc_mfu;
4939 * This buffer has been accessed more than once but has
4940 * been evicted from the cache. Move it back to the
4944 if (HDR_PREFETCH(hdr)) {
4946 * This is a prefetch access...
4947 * move this block back to the MRU state.
4949 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4950 new_state = arc_mru;
4953 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4954 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4955 arc_change_state(new_state, hdr, hash_lock);
4957 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4958 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4960 * This buffer is on the 2nd Level ARC.
4963 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4964 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4965 arc_change_state(arc_mfu, hdr, hash_lock);
4967 ASSERT(!"invalid arc state");
4971 /* a generic arc_done_func_t which you can use */
4974 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4976 if (zio == NULL || zio->io_error == 0)
4977 bcopy(buf->b_data, arg, arc_buf_size(buf));
4978 arc_buf_destroy(buf, arg);
4981 /* a generic arc_done_func_t */
4983 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4985 arc_buf_t **bufp = arg;
4986 if (zio && zio->io_error) {
4987 arc_buf_destroy(buf, arg);
4991 ASSERT(buf->b_data);
4996 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
4998 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
4999 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5000 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
5002 if (HDR_COMPRESSION_ENABLED(hdr)) {
5003 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
5004 BP_GET_COMPRESS(bp));
5006 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5007 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5012 arc_read_done(zio_t *zio)
5014 arc_buf_hdr_t *hdr = zio->io_private;
5015 kmutex_t *hash_lock = NULL;
5016 arc_callback_t *callback_list;
5017 arc_callback_t *acb;
5018 boolean_t freeable = B_FALSE;
5019 boolean_t no_zio_error = (zio->io_error == 0);
5022 * The hdr was inserted into hash-table and removed from lists
5023 * prior to starting I/O. We should find this header, since
5024 * it's in the hash table, and it should be legit since it's
5025 * not possible to evict it during the I/O. The only possible
5026 * reason for it not to be found is if we were freed during the
5029 if (HDR_IN_HASH_TABLE(hdr)) {
5030 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5031 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5032 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5033 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5034 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5036 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5039 ASSERT((found == hdr &&
5040 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5041 (found == hdr && HDR_L2_READING(hdr)));
5042 ASSERT3P(hash_lock, !=, NULL);
5046 /* byteswap if necessary */
5047 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5048 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5049 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5051 hdr->b_l1hdr.b_byteswap =
5052 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5055 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5059 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5060 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5061 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5063 callback_list = hdr->b_l1hdr.b_acb;
5064 ASSERT3P(callback_list, !=, NULL);
5066 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
5068 * Only call arc_access on anonymous buffers. This is because
5069 * if we've issued an I/O for an evicted buffer, we've already
5070 * called arc_access (to prevent any simultaneous readers from
5071 * getting confused).
5073 arc_access(hdr, hash_lock);
5077 * If a read request has a callback (i.e. acb_done is not NULL), then we
5078 * make a buf containing the data according to the parameters which were
5079 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5080 * aren't needlessly decompressing the data multiple times.
5082 int callback_cnt = 0;
5083 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5087 /* This is a demand read since prefetches don't use callbacks */
5090 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
5091 acb->acb_compressed, no_zio_error, &acb->acb_buf);
5093 zio->io_error = error;
5096 hdr->b_l1hdr.b_acb = NULL;
5097 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5098 if (callback_cnt == 0) {
5099 ASSERT(HDR_PREFETCH(hdr));
5100 ASSERT0(hdr->b_l1hdr.b_bufcnt);
5101 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5104 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5105 callback_list != NULL);
5108 arc_hdr_verify(hdr, zio->io_bp);
5110 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5111 if (hdr->b_l1hdr.b_state != arc_anon)
5112 arc_change_state(arc_anon, hdr, hash_lock);
5113 if (HDR_IN_HASH_TABLE(hdr))
5114 buf_hash_remove(hdr);
5115 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5119 * Broadcast before we drop the hash_lock to avoid the possibility
5120 * that the hdr (and hence the cv) might be freed before we get to
5121 * the cv_broadcast().
5123 cv_broadcast(&hdr->b_l1hdr.b_cv);
5125 if (hash_lock != NULL) {
5126 mutex_exit(hash_lock);
5129 * This block was freed while we waited for the read to
5130 * complete. It has been removed from the hash table and
5131 * moved to the anonymous state (so that it won't show up
5134 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5135 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5138 /* execute each callback and free its structure */
5139 while ((acb = callback_list) != NULL) {
5141 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
5143 if (acb->acb_zio_dummy != NULL) {
5144 acb->acb_zio_dummy->io_error = zio->io_error;
5145 zio_nowait(acb->acb_zio_dummy);
5148 callback_list = acb->acb_next;
5149 kmem_free(acb, sizeof (arc_callback_t));
5153 arc_hdr_destroy(hdr);
5157 * "Read" the block at the specified DVA (in bp) via the
5158 * cache. If the block is found in the cache, invoke the provided
5159 * callback immediately and return. Note that the `zio' parameter
5160 * in the callback will be NULL in this case, since no IO was
5161 * required. If the block is not in the cache pass the read request
5162 * on to the spa with a substitute callback function, so that the
5163 * requested block will be added to the cache.
5165 * If a read request arrives for a block that has a read in-progress,
5166 * either wait for the in-progress read to complete (and return the
5167 * results); or, if this is a read with a "done" func, add a record
5168 * to the read to invoke the "done" func when the read completes,
5169 * and return; or just return.
5171 * arc_read_done() will invoke all the requested "done" functions
5172 * for readers of this block.
5175 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
5176 void *private, zio_priority_t priority, int zio_flags,
5177 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5179 arc_buf_hdr_t *hdr = NULL;
5180 kmutex_t *hash_lock = NULL;
5182 uint64_t guid = spa_load_guid(spa);
5183 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5185 ASSERT(!BP_IS_EMBEDDED(bp) ||
5186 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5189 if (!BP_IS_EMBEDDED(bp)) {
5191 * Embedded BP's have no DVA and require no I/O to "read".
5192 * Create an anonymous arc buf to back it.
5194 hdr = buf_hash_find(guid, bp, &hash_lock);
5197 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5198 arc_buf_t *buf = NULL;
5199 *arc_flags |= ARC_FLAG_CACHED;
5201 if (HDR_IO_IN_PROGRESS(hdr)) {
5203 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5204 priority == ZIO_PRIORITY_SYNC_READ) {
5206 * This sync read must wait for an
5207 * in-progress async read (e.g. a predictive
5208 * prefetch). Async reads are queued
5209 * separately at the vdev_queue layer, so
5210 * this is a form of priority inversion.
5211 * Ideally, we would "inherit" the demand
5212 * i/o's priority by moving the i/o from
5213 * the async queue to the synchronous queue,
5214 * but there is currently no mechanism to do
5215 * so. Track this so that we can evaluate
5216 * the magnitude of this potential performance
5219 * Note that if the prefetch i/o is already
5220 * active (has been issued to the device),
5221 * the prefetch improved performance, because
5222 * we issued it sooner than we would have
5223 * without the prefetch.
5225 DTRACE_PROBE1(arc__sync__wait__for__async,
5226 arc_buf_hdr_t *, hdr);
5227 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
5229 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5230 arc_hdr_clear_flags(hdr,
5231 ARC_FLAG_PREDICTIVE_PREFETCH);
5234 if (*arc_flags & ARC_FLAG_WAIT) {
5235 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5236 mutex_exit(hash_lock);
5239 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5242 arc_callback_t *acb = NULL;
5244 acb = kmem_zalloc(sizeof (arc_callback_t),
5246 acb->acb_done = done;
5247 acb->acb_private = private;
5248 acb->acb_compressed = compressed_read;
5250 acb->acb_zio_dummy = zio_null(pio,
5251 spa, NULL, NULL, NULL, zio_flags);
5253 ASSERT3P(acb->acb_done, !=, NULL);
5254 acb->acb_next = hdr->b_l1hdr.b_acb;
5255 hdr->b_l1hdr.b_acb = acb;
5256 mutex_exit(hash_lock);
5259 mutex_exit(hash_lock);
5263 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5264 hdr->b_l1hdr.b_state == arc_mfu);
5267 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5269 * This is a demand read which does not have to
5270 * wait for i/o because we did a predictive
5271 * prefetch i/o for it, which has completed.
5274 arc__demand__hit__predictive__prefetch,
5275 arc_buf_hdr_t *, hdr);
5277 arcstat_demand_hit_predictive_prefetch);
5278 arc_hdr_clear_flags(hdr,
5279 ARC_FLAG_PREDICTIVE_PREFETCH);
5281 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5283 /* Get a buf with the desired data in it. */
5284 VERIFY0(arc_buf_alloc_impl(hdr, private,
5285 compressed_read, B_TRUE, &buf));
5286 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5287 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5288 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5290 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5291 arc_access(hdr, hash_lock);
5292 if (*arc_flags & ARC_FLAG_L2CACHE)
5293 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5294 mutex_exit(hash_lock);
5295 ARCSTAT_BUMP(arcstat_hits);
5296 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5297 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5298 data, metadata, hits);
5301 done(NULL, buf, private);
5303 uint64_t lsize = BP_GET_LSIZE(bp);
5304 uint64_t psize = BP_GET_PSIZE(bp);
5305 arc_callback_t *acb;
5308 boolean_t devw = B_FALSE;
5312 /* this block is not in the cache */
5313 arc_buf_hdr_t *exists = NULL;
5314 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5315 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5316 BP_GET_COMPRESS(bp), type);
5318 if (!BP_IS_EMBEDDED(bp)) {
5319 hdr->b_dva = *BP_IDENTITY(bp);
5320 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5321 exists = buf_hash_insert(hdr, &hash_lock);
5323 if (exists != NULL) {
5324 /* somebody beat us to the hash insert */
5325 mutex_exit(hash_lock);
5326 buf_discard_identity(hdr);
5327 arc_hdr_destroy(hdr);
5328 goto top; /* restart the IO request */
5332 * This block is in the ghost cache. If it was L2-only
5333 * (and thus didn't have an L1 hdr), we realloc the
5334 * header to add an L1 hdr.
5336 if (!HDR_HAS_L1HDR(hdr)) {
5337 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5340 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5341 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5342 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5343 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5344 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5345 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5348 * This is a delicate dance that we play here.
5349 * This hdr is in the ghost list so we access it
5350 * to move it out of the ghost list before we
5351 * initiate the read. If it's a prefetch then
5352 * it won't have a callback so we'll remove the
5353 * reference that arc_buf_alloc_impl() created. We
5354 * do this after we've called arc_access() to
5355 * avoid hitting an assert in remove_reference().
5357 arc_access(hdr, hash_lock);
5358 arc_hdr_alloc_pabd(hdr);
5360 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5361 size = arc_hdr_size(hdr);
5364 * If compression is enabled on the hdr, then will do
5365 * RAW I/O and will store the compressed data in the hdr's
5366 * data block. Otherwise, the hdr's data block will contain
5367 * the uncompressed data.
5369 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5370 zio_flags |= ZIO_FLAG_RAW;
5373 if (*arc_flags & ARC_FLAG_PREFETCH)
5374 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5375 if (*arc_flags & ARC_FLAG_L2CACHE)
5376 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5377 if (BP_GET_LEVEL(bp) > 0)
5378 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5379 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5380 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5381 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5383 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5384 acb->acb_done = done;
5385 acb->acb_private = private;
5386 acb->acb_compressed = compressed_read;
5388 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5389 hdr->b_l1hdr.b_acb = acb;
5390 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5392 if (HDR_HAS_L2HDR(hdr) &&
5393 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5394 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5395 addr = hdr->b_l2hdr.b_daddr;
5397 * Lock out device removal.
5399 if (vdev_is_dead(vd) ||
5400 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5404 if (priority == ZIO_PRIORITY_ASYNC_READ)
5405 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5407 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5409 if (hash_lock != NULL)
5410 mutex_exit(hash_lock);
5413 * At this point, we have a level 1 cache miss. Try again in
5414 * L2ARC if possible.
5416 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5418 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5419 uint64_t, lsize, zbookmark_phys_t *, zb);
5420 ARCSTAT_BUMP(arcstat_misses);
5421 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5422 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5423 data, metadata, misses);
5428 racct_add_force(curproc, RACCT_READBPS, size);
5429 racct_add_force(curproc, RACCT_READIOPS, 1);
5430 PROC_UNLOCK(curproc);
5433 curthread->td_ru.ru_inblock++;
5436 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5438 * Read from the L2ARC if the following are true:
5439 * 1. The L2ARC vdev was previously cached.
5440 * 2. This buffer still has L2ARC metadata.
5441 * 3. This buffer isn't currently writing to the L2ARC.
5442 * 4. The L2ARC entry wasn't evicted, which may
5443 * also have invalidated the vdev.
5444 * 5. This isn't prefetch and l2arc_noprefetch is set.
5446 if (HDR_HAS_L2HDR(hdr) &&
5447 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5448 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5449 l2arc_read_callback_t *cb;
5453 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5454 ARCSTAT_BUMP(arcstat_l2_hits);
5456 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5458 cb->l2rcb_hdr = hdr;
5461 cb->l2rcb_flags = zio_flags;
5463 asize = vdev_psize_to_asize(vd, size);
5464 if (asize != size) {
5465 abd = abd_alloc_for_io(asize,
5466 HDR_ISTYPE_METADATA(hdr));
5467 cb->l2rcb_abd = abd;
5469 abd = hdr->b_l1hdr.b_pabd;
5472 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5473 addr + asize <= vd->vdev_psize -
5474 VDEV_LABEL_END_SIZE);
5477 * l2arc read. The SCL_L2ARC lock will be
5478 * released by l2arc_read_done().
5479 * Issue a null zio if the underlying buffer
5480 * was squashed to zero size by compression.
5482 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5483 ZIO_COMPRESS_EMPTY);
5484 rzio = zio_read_phys(pio, vd, addr,
5487 l2arc_read_done, cb, priority,
5488 zio_flags | ZIO_FLAG_DONT_CACHE |
5490 ZIO_FLAG_DONT_PROPAGATE |
5491 ZIO_FLAG_DONT_RETRY, B_FALSE);
5492 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5494 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5496 if (*arc_flags & ARC_FLAG_NOWAIT) {
5501 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5502 if (zio_wait(rzio) == 0)
5505 /* l2arc read error; goto zio_read() */
5507 DTRACE_PROBE1(l2arc__miss,
5508 arc_buf_hdr_t *, hdr);
5509 ARCSTAT_BUMP(arcstat_l2_misses);
5510 if (HDR_L2_WRITING(hdr))
5511 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5512 spa_config_exit(spa, SCL_L2ARC, vd);
5516 spa_config_exit(spa, SCL_L2ARC, vd);
5517 if (l2arc_ndev != 0) {
5518 DTRACE_PROBE1(l2arc__miss,
5519 arc_buf_hdr_t *, hdr);
5520 ARCSTAT_BUMP(arcstat_l2_misses);
5524 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
5525 arc_read_done, hdr, priority, zio_flags, zb);
5527 if (*arc_flags & ARC_FLAG_WAIT)
5528 return (zio_wait(rzio));
5530 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5537 * Notify the arc that a block was freed, and thus will never be used again.
5540 arc_freed(spa_t *spa, const blkptr_t *bp)
5543 kmutex_t *hash_lock;
5544 uint64_t guid = spa_load_guid(spa);
5546 ASSERT(!BP_IS_EMBEDDED(bp));
5548 hdr = buf_hash_find(guid, bp, &hash_lock);
5553 * We might be trying to free a block that is still doing I/O
5554 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5555 * dmu_sync-ed block). If this block is being prefetched, then it
5556 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5557 * until the I/O completes. A block may also have a reference if it is
5558 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5559 * have written the new block to its final resting place on disk but
5560 * without the dedup flag set. This would have left the hdr in the MRU
5561 * state and discoverable. When the txg finally syncs it detects that
5562 * the block was overridden in open context and issues an override I/O.
5563 * Since this is a dedup block, the override I/O will determine if the
5564 * block is already in the DDT. If so, then it will replace the io_bp
5565 * with the bp from the DDT and allow the I/O to finish. When the I/O
5566 * reaches the done callback, dbuf_write_override_done, it will
5567 * check to see if the io_bp and io_bp_override are identical.
5568 * If they are not, then it indicates that the bp was replaced with
5569 * the bp in the DDT and the override bp is freed. This allows
5570 * us to arrive here with a reference on a block that is being
5571 * freed. So if we have an I/O in progress, or a reference to
5572 * this hdr, then we don't destroy the hdr.
5574 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5575 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5576 arc_change_state(arc_anon, hdr, hash_lock);
5577 arc_hdr_destroy(hdr);
5578 mutex_exit(hash_lock);
5580 mutex_exit(hash_lock);
5586 * Release this buffer from the cache, making it an anonymous buffer. This
5587 * must be done after a read and prior to modifying the buffer contents.
5588 * If the buffer has more than one reference, we must make
5589 * a new hdr for the buffer.
5592 arc_release(arc_buf_t *buf, void *tag)
5594 arc_buf_hdr_t *hdr = buf->b_hdr;
5597 * It would be nice to assert that if it's DMU metadata (level >
5598 * 0 || it's the dnode file), then it must be syncing context.
5599 * But we don't know that information at this level.
5602 mutex_enter(&buf->b_evict_lock);
5604 ASSERT(HDR_HAS_L1HDR(hdr));
5607 * We don't grab the hash lock prior to this check, because if
5608 * the buffer's header is in the arc_anon state, it won't be
5609 * linked into the hash table.
5611 if (hdr->b_l1hdr.b_state == arc_anon) {
5612 mutex_exit(&buf->b_evict_lock);
5613 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5614 ASSERT(!HDR_IN_HASH_TABLE(hdr));
5615 ASSERT(!HDR_HAS_L2HDR(hdr));
5616 ASSERT(HDR_EMPTY(hdr));
5617 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5618 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5619 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5621 hdr->b_l1hdr.b_arc_access = 0;
5624 * If the buf is being overridden then it may already
5625 * have a hdr that is not empty.
5627 buf_discard_identity(hdr);
5633 kmutex_t *hash_lock = HDR_LOCK(hdr);
5634 mutex_enter(hash_lock);
5637 * This assignment is only valid as long as the hash_lock is
5638 * held, we must be careful not to reference state or the
5639 * b_state field after dropping the lock.
5641 arc_state_t *state = hdr->b_l1hdr.b_state;
5642 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5643 ASSERT3P(state, !=, arc_anon);
5645 /* this buffer is not on any list */
5646 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
5648 if (HDR_HAS_L2HDR(hdr)) {
5649 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5652 * We have to recheck this conditional again now that
5653 * we're holding the l2ad_mtx to prevent a race with
5654 * another thread which might be concurrently calling
5655 * l2arc_evict(). In that case, l2arc_evict() might have
5656 * destroyed the header's L2 portion as we were waiting
5657 * to acquire the l2ad_mtx.
5659 if (HDR_HAS_L2HDR(hdr)) {
5661 arc_hdr_l2hdr_destroy(hdr);
5664 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5668 * Do we have more than one buf?
5670 if (hdr->b_l1hdr.b_bufcnt > 1) {
5671 arc_buf_hdr_t *nhdr;
5672 uint64_t spa = hdr->b_spa;
5673 uint64_t psize = HDR_GET_PSIZE(hdr);
5674 uint64_t lsize = HDR_GET_LSIZE(hdr);
5675 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5676 arc_buf_contents_t type = arc_buf_type(hdr);
5677 VERIFY3U(hdr->b_type, ==, type);
5679 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5680 (void) remove_reference(hdr, hash_lock, tag);
5682 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
5683 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5684 ASSERT(ARC_BUF_LAST(buf));
5688 * Pull the data off of this hdr and attach it to
5689 * a new anonymous hdr. Also find the last buffer
5690 * in the hdr's buffer list.
5692 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
5693 ASSERT3P(lastbuf, !=, NULL);
5696 * If the current arc_buf_t and the hdr are sharing their data
5697 * buffer, then we must stop sharing that block.
5699 if (arc_buf_is_shared(buf)) {
5700 VERIFY(!arc_buf_is_shared(lastbuf));
5703 * First, sever the block sharing relationship between
5704 * buf and the arc_buf_hdr_t.
5706 arc_unshare_buf(hdr, buf);
5709 * Now we need to recreate the hdr's b_pabd. Since we
5710 * have lastbuf handy, we try to share with it, but if
5711 * we can't then we allocate a new b_pabd and copy the
5712 * data from buf into it.
5714 if (arc_can_share(hdr, lastbuf)) {
5715 arc_share_buf(hdr, lastbuf);
5717 arc_hdr_alloc_pabd(hdr);
5718 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
5719 buf->b_data, psize);
5721 VERIFY3P(lastbuf->b_data, !=, NULL);
5722 } else if (HDR_SHARED_DATA(hdr)) {
5724 * Uncompressed shared buffers are always at the end
5725 * of the list. Compressed buffers don't have the
5726 * same requirements. This makes it hard to
5727 * simply assert that the lastbuf is shared so
5728 * we rely on the hdr's compression flags to determine
5729 * if we have a compressed, shared buffer.
5731 ASSERT(arc_buf_is_shared(lastbuf) ||
5732 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
5733 ASSERT(!ARC_BUF_SHARED(buf));
5735 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5736 ASSERT3P(state, !=, arc_l2c_only);
5738 (void) refcount_remove_many(&state->arcs_size,
5739 arc_buf_size(buf), buf);
5741 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5742 ASSERT3P(state, !=, arc_l2c_only);
5743 (void) refcount_remove_many(&state->arcs_esize[type],
5744 arc_buf_size(buf), buf);
5747 hdr->b_l1hdr.b_bufcnt -= 1;
5748 arc_cksum_verify(buf);
5750 arc_buf_unwatch(buf);
5753 mutex_exit(hash_lock);
5756 * Allocate a new hdr. The new hdr will contain a b_pabd
5757 * buffer which will be freed in arc_write().
5759 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5760 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5761 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5762 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5763 VERIFY3U(nhdr->b_type, ==, type);
5764 ASSERT(!HDR_SHARED_DATA(nhdr));
5766 nhdr->b_l1hdr.b_buf = buf;
5767 nhdr->b_l1hdr.b_bufcnt = 1;
5768 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5771 mutex_exit(&buf->b_evict_lock);
5772 (void) refcount_add_many(&arc_anon->arcs_size,
5773 arc_buf_size(buf), buf);
5775 mutex_exit(&buf->b_evict_lock);
5776 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5777 /* protected by hash lock, or hdr is on arc_anon */
5778 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5779 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5780 arc_change_state(arc_anon, hdr, hash_lock);
5781 hdr->b_l1hdr.b_arc_access = 0;
5782 mutex_exit(hash_lock);
5784 buf_discard_identity(hdr);
5790 arc_released(arc_buf_t *buf)
5794 mutex_enter(&buf->b_evict_lock);
5795 released = (buf->b_data != NULL &&
5796 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5797 mutex_exit(&buf->b_evict_lock);
5803 arc_referenced(arc_buf_t *buf)
5807 mutex_enter(&buf->b_evict_lock);
5808 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5809 mutex_exit(&buf->b_evict_lock);
5810 return (referenced);
5815 arc_write_ready(zio_t *zio)
5817 arc_write_callback_t *callback = zio->io_private;
5818 arc_buf_t *buf = callback->awcb_buf;
5819 arc_buf_hdr_t *hdr = buf->b_hdr;
5820 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5822 ASSERT(HDR_HAS_L1HDR(hdr));
5823 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5824 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5827 * If we're reexecuting this zio because the pool suspended, then
5828 * cleanup any state that was previously set the first time the
5829 * callback was invoked.
5831 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5832 arc_cksum_free(hdr);
5834 arc_buf_unwatch(buf);
5836 if (hdr->b_l1hdr.b_pabd != NULL) {
5837 if (arc_buf_is_shared(buf)) {
5838 arc_unshare_buf(hdr, buf);
5840 arc_hdr_free_pabd(hdr);
5844 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5845 ASSERT(!HDR_SHARED_DATA(hdr));
5846 ASSERT(!arc_buf_is_shared(buf));
5848 callback->awcb_ready(zio, buf, callback->awcb_private);
5850 if (HDR_IO_IN_PROGRESS(hdr))
5851 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
5853 arc_cksum_compute(buf);
5854 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5856 enum zio_compress compress;
5857 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5858 compress = ZIO_COMPRESS_OFF;
5860 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
5861 compress = BP_GET_COMPRESS(zio->io_bp);
5863 HDR_SET_PSIZE(hdr, psize);
5864 arc_hdr_set_compress(hdr, compress);
5868 * Fill the hdr with data. If the hdr is compressed, the data we want
5869 * is available from the zio, otherwise we can take it from the buf.
5871 * We might be able to share the buf's data with the hdr here. However,
5872 * doing so would cause the ARC to be full of linear ABDs if we write a
5873 * lot of shareable data. As a compromise, we check whether scattered
5874 * ABDs are allowed, and assume that if they are then the user wants
5875 * the ARC to be primarily filled with them regardless of the data being
5876 * written. Therefore, if they're allowed then we allocate one and copy
5877 * the data into it; otherwise, we share the data directly if we can.
5879 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
5880 arc_hdr_alloc_pabd(hdr);
5883 * Ideally, we would always copy the io_abd into b_pabd, but the
5884 * user may have disabled compressed ARC, thus we must check the
5885 * hdr's compression setting rather than the io_bp's.
5887 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5888 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
5890 ASSERT3U(psize, >, 0);
5892 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
5894 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
5896 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
5900 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
5901 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
5902 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5904 arc_share_buf(hdr, buf);
5907 arc_hdr_verify(hdr, zio->io_bp);
5911 arc_write_children_ready(zio_t *zio)
5913 arc_write_callback_t *callback = zio->io_private;
5914 arc_buf_t *buf = callback->awcb_buf;
5916 callback->awcb_children_ready(zio, buf, callback->awcb_private);
5920 * The SPA calls this callback for each physical write that happens on behalf
5921 * of a logical write. See the comment in dbuf_write_physdone() for details.
5924 arc_write_physdone(zio_t *zio)
5926 arc_write_callback_t *cb = zio->io_private;
5927 if (cb->awcb_physdone != NULL)
5928 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
5932 arc_write_done(zio_t *zio)
5934 arc_write_callback_t *callback = zio->io_private;
5935 arc_buf_t *buf = callback->awcb_buf;
5936 arc_buf_hdr_t *hdr = buf->b_hdr;
5938 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5940 if (zio->io_error == 0) {
5941 arc_hdr_verify(hdr, zio->io_bp);
5943 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5944 buf_discard_identity(hdr);
5946 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
5947 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
5950 ASSERT(HDR_EMPTY(hdr));
5954 * If the block to be written was all-zero or compressed enough to be
5955 * embedded in the BP, no write was performed so there will be no
5956 * dva/birth/checksum. The buffer must therefore remain anonymous
5959 if (!HDR_EMPTY(hdr)) {
5960 arc_buf_hdr_t *exists;
5961 kmutex_t *hash_lock;
5963 ASSERT3U(zio->io_error, ==, 0);
5965 arc_cksum_verify(buf);
5967 exists = buf_hash_insert(hdr, &hash_lock);
5968 if (exists != NULL) {
5970 * This can only happen if we overwrite for
5971 * sync-to-convergence, because we remove
5972 * buffers from the hash table when we arc_free().
5974 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5975 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5976 panic("bad overwrite, hdr=%p exists=%p",
5977 (void *)hdr, (void *)exists);
5978 ASSERT(refcount_is_zero(
5979 &exists->b_l1hdr.b_refcnt));
5980 arc_change_state(arc_anon, exists, hash_lock);
5981 mutex_exit(hash_lock);
5982 arc_hdr_destroy(exists);
5983 exists = buf_hash_insert(hdr, &hash_lock);
5984 ASSERT3P(exists, ==, NULL);
5985 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
5987 ASSERT(zio->io_prop.zp_nopwrite);
5988 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5989 panic("bad nopwrite, hdr=%p exists=%p",
5990 (void *)hdr, (void *)exists);
5993 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
5994 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
5995 ASSERT(BP_GET_DEDUP(zio->io_bp));
5996 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
5999 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6000 /* if it's not anon, we are doing a scrub */
6001 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6002 arc_access(hdr, hash_lock);
6003 mutex_exit(hash_lock);
6005 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6008 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6009 callback->awcb_done(zio, buf, callback->awcb_private);
6011 abd_put(zio->io_abd);
6012 kmem_free(callback, sizeof (arc_write_callback_t));
6016 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6017 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
6018 arc_done_func_t *children_ready, arc_done_func_t *physdone,
6019 arc_done_func_t *done, void *private, zio_priority_t priority,
6020 int zio_flags, const zbookmark_phys_t *zb)
6022 arc_buf_hdr_t *hdr = buf->b_hdr;
6023 arc_write_callback_t *callback;
6025 zio_prop_t localprop = *zp;
6027 ASSERT3P(ready, !=, NULL);
6028 ASSERT3P(done, !=, NULL);
6029 ASSERT(!HDR_IO_ERROR(hdr));
6030 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6031 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6032 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6034 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6035 if (ARC_BUF_COMPRESSED(buf)) {
6037 * We're writing a pre-compressed buffer. Make the
6038 * compression algorithm requested by the zio_prop_t match
6039 * the pre-compressed buffer's compression algorithm.
6041 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6043 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6044 zio_flags |= ZIO_FLAG_RAW;
6046 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6047 callback->awcb_ready = ready;
6048 callback->awcb_children_ready = children_ready;
6049 callback->awcb_physdone = physdone;
6050 callback->awcb_done = done;
6051 callback->awcb_private = private;
6052 callback->awcb_buf = buf;
6055 * The hdr's b_pabd is now stale, free it now. A new data block
6056 * will be allocated when the zio pipeline calls arc_write_ready().
6058 if (hdr->b_l1hdr.b_pabd != NULL) {
6060 * If the buf is currently sharing the data block with
6061 * the hdr then we need to break that relationship here.
6062 * The hdr will remain with a NULL data pointer and the
6063 * buf will take sole ownership of the block.
6065 if (arc_buf_is_shared(buf)) {
6066 arc_unshare_buf(hdr, buf);
6068 arc_hdr_free_pabd(hdr);
6070 VERIFY3P(buf->b_data, !=, NULL);
6071 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6073 ASSERT(!arc_buf_is_shared(buf));
6074 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6076 zio = zio_write(pio, spa, txg, bp,
6077 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6078 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6079 (children_ready != NULL) ? arc_write_children_ready : NULL,
6080 arc_write_physdone, arc_write_done, callback,
6081 priority, zio_flags, zb);
6087 arc_memory_throttle(uint64_t reserve, uint64_t txg)
6090 uint64_t available_memory = ptob(freemem);
6091 static uint64_t page_load = 0;
6092 static uint64_t last_txg = 0;
6094 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
6095 available_memory = MIN(available_memory, uma_avail());
6098 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
6101 if (txg > last_txg) {
6106 * If we are in pageout, we know that memory is already tight,
6107 * the arc is already going to be evicting, so we just want to
6108 * continue to let page writes occur as quickly as possible.
6110 if (curproc == pageproc) {
6111 if (page_load > MAX(ptob(minfree), available_memory) / 4)
6112 return (SET_ERROR(ERESTART));
6113 /* Note: reserve is inflated, so we deflate */
6114 page_load += reserve / 8;
6116 } else if (page_load > 0 && arc_reclaim_needed()) {
6117 /* memory is low, delay before restarting */
6118 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6119 return (SET_ERROR(EAGAIN));
6127 arc_tempreserve_clear(uint64_t reserve)
6129 atomic_add_64(&arc_tempreserve, -reserve);
6130 ASSERT((int64_t)arc_tempreserve >= 0);
6134 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
6139 if (reserve > arc_c/4 && !arc_no_grow) {
6140 arc_c = MIN(arc_c_max, reserve * 4);
6141 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
6143 if (reserve > arc_c)
6144 return (SET_ERROR(ENOMEM));
6147 * Don't count loaned bufs as in flight dirty data to prevent long
6148 * network delays from blocking transactions that are ready to be
6149 * assigned to a txg.
6152 /* assert that it has not wrapped around */
6153 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6155 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6156 arc_loaned_bytes), 0);
6159 * Writes will, almost always, require additional memory allocations
6160 * in order to compress/encrypt/etc the data. We therefore need to
6161 * make sure that there is sufficient available memory for this.
6163 error = arc_memory_throttle(reserve, txg);
6168 * Throttle writes when the amount of dirty data in the cache
6169 * gets too large. We try to keep the cache less than half full
6170 * of dirty blocks so that our sync times don't grow too large.
6171 * Note: if two requests come in concurrently, we might let them
6172 * both succeed, when one of them should fail. Not a huge deal.
6175 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
6176 anon_size > arc_c / 4) {
6177 uint64_t meta_esize =
6178 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6179 uint64_t data_esize =
6180 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6181 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6182 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6183 arc_tempreserve >> 10, meta_esize >> 10,
6184 data_esize >> 10, reserve >> 10, arc_c >> 10);
6185 return (SET_ERROR(ERESTART));
6187 atomic_add_64(&arc_tempreserve, reserve);
6192 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6193 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6195 size->value.ui64 = refcount_count(&state->arcs_size);
6196 evict_data->value.ui64 =
6197 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6198 evict_metadata->value.ui64 =
6199 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6203 arc_kstat_update(kstat_t *ksp, int rw)
6205 arc_stats_t *as = ksp->ks_data;
6207 if (rw == KSTAT_WRITE) {
6210 arc_kstat_update_state(arc_anon,
6211 &as->arcstat_anon_size,
6212 &as->arcstat_anon_evictable_data,
6213 &as->arcstat_anon_evictable_metadata);
6214 arc_kstat_update_state(arc_mru,
6215 &as->arcstat_mru_size,
6216 &as->arcstat_mru_evictable_data,
6217 &as->arcstat_mru_evictable_metadata);
6218 arc_kstat_update_state(arc_mru_ghost,
6219 &as->arcstat_mru_ghost_size,
6220 &as->arcstat_mru_ghost_evictable_data,
6221 &as->arcstat_mru_ghost_evictable_metadata);
6222 arc_kstat_update_state(arc_mfu,
6223 &as->arcstat_mfu_size,
6224 &as->arcstat_mfu_evictable_data,
6225 &as->arcstat_mfu_evictable_metadata);
6226 arc_kstat_update_state(arc_mfu_ghost,
6227 &as->arcstat_mfu_ghost_size,
6228 &as->arcstat_mfu_ghost_evictable_data,
6229 &as->arcstat_mfu_ghost_evictable_metadata);
6236 * This function *must* return indices evenly distributed between all
6237 * sublists of the multilist. This is needed due to how the ARC eviction
6238 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6239 * distributed between all sublists and uses this assumption when
6240 * deciding which sublist to evict from and how much to evict from it.
6243 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6245 arc_buf_hdr_t *hdr = obj;
6248 * We rely on b_dva to generate evenly distributed index
6249 * numbers using buf_hash below. So, as an added precaution,
6250 * let's make sure we never add empty buffers to the arc lists.
6252 ASSERT(!HDR_EMPTY(hdr));
6255 * The assumption here, is the hash value for a given
6256 * arc_buf_hdr_t will remain constant throughout it's lifetime
6257 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6258 * Thus, we don't need to store the header's sublist index
6259 * on insertion, as this index can be recalculated on removal.
6261 * Also, the low order bits of the hash value are thought to be
6262 * distributed evenly. Otherwise, in the case that the multilist
6263 * has a power of two number of sublists, each sublists' usage
6264 * would not be evenly distributed.
6266 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6267 multilist_get_num_sublists(ml));
6271 static eventhandler_tag arc_event_lowmem = NULL;
6274 arc_lowmem(void *arg __unused, int howto __unused)
6277 mutex_enter(&arc_reclaim_lock);
6278 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE);
6279 cv_signal(&arc_reclaim_thread_cv);
6282 * It is unsafe to block here in arbitrary threads, because we can come
6283 * here from ARC itself and may hold ARC locks and thus risk a deadlock
6284 * with ARC reclaim thread.
6286 if (curproc == pageproc)
6287 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
6288 mutex_exit(&arc_reclaim_lock);
6293 arc_state_init(void)
6295 arc_anon = &ARC_anon;
6297 arc_mru_ghost = &ARC_mru_ghost;
6299 arc_mfu_ghost = &ARC_mfu_ghost;
6300 arc_l2c_only = &ARC_l2c_only;
6302 arc_mru->arcs_list[ARC_BUFC_METADATA] =
6303 multilist_create(sizeof (arc_buf_hdr_t),
6304 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6305 arc_state_multilist_index_func);
6306 arc_mru->arcs_list[ARC_BUFC_DATA] =
6307 multilist_create(sizeof (arc_buf_hdr_t),
6308 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6309 arc_state_multilist_index_func);
6310 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6311 multilist_create(sizeof (arc_buf_hdr_t),
6312 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6313 arc_state_multilist_index_func);
6314 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6315 multilist_create(sizeof (arc_buf_hdr_t),
6316 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6317 arc_state_multilist_index_func);
6318 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6319 multilist_create(sizeof (arc_buf_hdr_t),
6320 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6321 arc_state_multilist_index_func);
6322 arc_mfu->arcs_list[ARC_BUFC_DATA] =
6323 multilist_create(sizeof (arc_buf_hdr_t),
6324 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6325 arc_state_multilist_index_func);
6326 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6327 multilist_create(sizeof (arc_buf_hdr_t),
6328 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6329 arc_state_multilist_index_func);
6330 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6331 multilist_create(sizeof (arc_buf_hdr_t),
6332 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6333 arc_state_multilist_index_func);
6334 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6335 multilist_create(sizeof (arc_buf_hdr_t),
6336 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6337 arc_state_multilist_index_func);
6338 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6339 multilist_create(sizeof (arc_buf_hdr_t),
6340 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6341 arc_state_multilist_index_func);
6343 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6344 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6345 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6346 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6347 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6348 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6349 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6350 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6351 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6352 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6353 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6354 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6356 refcount_create(&arc_anon->arcs_size);
6357 refcount_create(&arc_mru->arcs_size);
6358 refcount_create(&arc_mru_ghost->arcs_size);
6359 refcount_create(&arc_mfu->arcs_size);
6360 refcount_create(&arc_mfu_ghost->arcs_size);
6361 refcount_create(&arc_l2c_only->arcs_size);
6365 arc_state_fini(void)
6367 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6368 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6369 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6370 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6371 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6372 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6373 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6374 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6375 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6376 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6377 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6378 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6380 refcount_destroy(&arc_anon->arcs_size);
6381 refcount_destroy(&arc_mru->arcs_size);
6382 refcount_destroy(&arc_mru_ghost->arcs_size);
6383 refcount_destroy(&arc_mfu->arcs_size);
6384 refcount_destroy(&arc_mfu_ghost->arcs_size);
6385 refcount_destroy(&arc_l2c_only->arcs_size);
6387 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6388 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6389 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6390 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6391 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6392 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6393 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6394 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6406 int i, prefetch_tunable_set = 0;
6409 * allmem is "all memory that we could possibly use".
6413 uint64_t allmem = ptob(physmem - swapfs_minfree);
6415 uint64_t allmem = (physmem * PAGESIZE) / 2;
6418 uint64_t allmem = kmem_size();
6422 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
6423 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
6424 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
6426 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
6427 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
6429 /* Convert seconds to clock ticks */
6430 arc_min_prefetch_lifespan = 1 * hz;
6432 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
6433 arc_c_min = MAX(allmem / 32, arc_abs_min);
6434 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */
6435 if (allmem >= 1 << 30)
6436 arc_c_max = allmem - (1 << 30);
6438 arc_c_max = arc_c_min;
6439 arc_c_max = MAX(allmem * 5 / 8, arc_c_max);
6442 * In userland, there's only the memory pressure that we artificially
6443 * create (see arc_available_memory()). Don't let arc_c get too
6444 * small, because it can cause transactions to be larger than
6445 * arc_c, causing arc_tempreserve_space() to fail.
6448 arc_c_min = arc_c_max / 2;
6453 * Allow the tunables to override our calculations if they are
6456 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) {
6457 arc_c_max = zfs_arc_max;
6458 arc_c_min = MIN(arc_c_min, arc_c_max);
6460 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
6461 arc_c_min = zfs_arc_min;
6465 arc_p = (arc_c >> 1);
6468 /* limit meta-data to 1/4 of the arc capacity */
6469 arc_meta_limit = arc_c_max / 4;
6473 * Metadata is stored in the kernel's heap. Don't let us
6474 * use more than half the heap for the ARC.
6477 arc_meta_limit = MIN(arc_meta_limit, uma_limit() / 2);
6479 arc_meta_limit = MIN(arc_meta_limit,
6480 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6484 /* Allow the tunable to override if it is reasonable */
6485 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6486 arc_meta_limit = zfs_arc_meta_limit;
6488 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6489 arc_c_min = arc_meta_limit / 2;
6491 if (zfs_arc_meta_min > 0) {
6492 arc_meta_min = zfs_arc_meta_min;
6494 arc_meta_min = arc_c_min / 2;
6497 if (zfs_arc_grow_retry > 0)
6498 arc_grow_retry = zfs_arc_grow_retry;
6500 if (zfs_arc_shrink_shift > 0)
6501 arc_shrink_shift = zfs_arc_shrink_shift;
6503 if (zfs_arc_no_grow_shift > 0)
6504 arc_no_grow_shift = zfs_arc_no_grow_shift;
6506 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6508 if (arc_no_grow_shift >= arc_shrink_shift)
6509 arc_no_grow_shift = arc_shrink_shift - 1;
6511 if (zfs_arc_p_min_shift > 0)
6512 arc_p_min_shift = zfs_arc_p_min_shift;
6514 /* if kmem_flags are set, lets try to use less memory */
6515 if (kmem_debugging())
6517 if (arc_c < arc_c_min)
6520 zfs_arc_min = arc_c_min;
6521 zfs_arc_max = arc_c_max;
6526 arc_reclaim_thread_exit = B_FALSE;
6527 arc_dnlc_evicts_thread_exit = FALSE;
6529 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6530 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6532 if (arc_ksp != NULL) {
6533 arc_ksp->ks_data = &arc_stats;
6534 arc_ksp->ks_update = arc_kstat_update;
6535 kstat_install(arc_ksp);
6538 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6539 TS_RUN, minclsyspri);
6542 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
6543 EVENTHANDLER_PRI_FIRST);
6546 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
6547 TS_RUN, minclsyspri);
6553 * Calculate maximum amount of dirty data per pool.
6555 * If it has been set by /etc/system, take that.
6556 * Otherwise, use a percentage of physical memory defined by
6557 * zfs_dirty_data_max_percent (default 10%) with a cap at
6558 * zfs_dirty_data_max_max (default 4GB).
6560 if (zfs_dirty_data_max == 0) {
6561 zfs_dirty_data_max = ptob(physmem) *
6562 zfs_dirty_data_max_percent / 100;
6563 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6564 zfs_dirty_data_max_max);
6568 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
6569 prefetch_tunable_set = 1;
6572 if (prefetch_tunable_set == 0) {
6573 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
6575 printf(" add \"vfs.zfs.prefetch_disable=0\" "
6576 "to /boot/loader.conf.\n");
6577 zfs_prefetch_disable = 1;
6580 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
6581 prefetch_tunable_set == 0) {
6582 printf("ZFS NOTICE: Prefetch is disabled by default if less "
6583 "than 4GB of RAM is present;\n"
6584 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
6585 "to /boot/loader.conf.\n");
6586 zfs_prefetch_disable = 1;
6589 /* Warn about ZFS memory and address space requirements. */
6590 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
6591 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
6592 "expect unstable behavior.\n");
6594 if (allmem < 512 * (1 << 20)) {
6595 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
6596 "expect unstable behavior.\n");
6597 printf(" Consider tuning vm.kmem_size and "
6598 "vm.kmem_size_max\n");
6599 printf(" in /boot/loader.conf.\n");
6608 if (arc_event_lowmem != NULL)
6609 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
6612 mutex_enter(&arc_reclaim_lock);
6613 arc_reclaim_thread_exit = B_TRUE;
6615 * The reclaim thread will set arc_reclaim_thread_exit back to
6616 * B_FALSE when it is finished exiting; we're waiting for that.
6618 while (arc_reclaim_thread_exit) {
6619 cv_signal(&arc_reclaim_thread_cv);
6620 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6622 mutex_exit(&arc_reclaim_lock);
6624 /* Use B_TRUE to ensure *all* buffers are evicted */
6625 arc_flush(NULL, B_TRUE);
6627 mutex_enter(&arc_dnlc_evicts_lock);
6628 arc_dnlc_evicts_thread_exit = TRUE;
6630 * The user evicts thread will set arc_user_evicts_thread_exit
6631 * to FALSE when it is finished exiting; we're waiting for that.
6633 while (arc_dnlc_evicts_thread_exit) {
6634 cv_signal(&arc_dnlc_evicts_cv);
6635 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
6637 mutex_exit(&arc_dnlc_evicts_lock);
6641 if (arc_ksp != NULL) {
6642 kstat_delete(arc_ksp);
6646 mutex_destroy(&arc_reclaim_lock);
6647 cv_destroy(&arc_reclaim_thread_cv);
6648 cv_destroy(&arc_reclaim_waiters_cv);
6650 mutex_destroy(&arc_dnlc_evicts_lock);
6651 cv_destroy(&arc_dnlc_evicts_cv);
6656 ASSERT0(arc_loaned_bytes);
6662 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6663 * It uses dedicated storage devices to hold cached data, which are populated
6664 * using large infrequent writes. The main role of this cache is to boost
6665 * the performance of random read workloads. The intended L2ARC devices
6666 * include short-stroked disks, solid state disks, and other media with
6667 * substantially faster read latency than disk.
6669 * +-----------------------+
6671 * +-----------------------+
6674 * l2arc_feed_thread() arc_read()
6678 * +---------------+ |
6680 * +---------------+ |
6685 * +-------+ +-------+
6687 * | cache | | cache |
6688 * +-------+ +-------+
6689 * +=========+ .-----.
6690 * : L2ARC : |-_____-|
6691 * : devices : | Disks |
6692 * +=========+ `-_____-'
6694 * Read requests are satisfied from the following sources, in order:
6697 * 2) vdev cache of L2ARC devices
6699 * 4) vdev cache of disks
6702 * Some L2ARC device types exhibit extremely slow write performance.
6703 * To accommodate for this there are some significant differences between
6704 * the L2ARC and traditional cache design:
6706 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6707 * the ARC behave as usual, freeing buffers and placing headers on ghost
6708 * lists. The ARC does not send buffers to the L2ARC during eviction as
6709 * this would add inflated write latencies for all ARC memory pressure.
6711 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6712 * It does this by periodically scanning buffers from the eviction-end of
6713 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6714 * not already there. It scans until a headroom of buffers is satisfied,
6715 * which itself is a buffer for ARC eviction. If a compressible buffer is
6716 * found during scanning and selected for writing to an L2ARC device, we
6717 * temporarily boost scanning headroom during the next scan cycle to make
6718 * sure we adapt to compression effects (which might significantly reduce
6719 * the data volume we write to L2ARC). The thread that does this is
6720 * l2arc_feed_thread(), illustrated below; example sizes are included to
6721 * provide a better sense of ratio than this diagram:
6724 * +---------------------+----------+
6725 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6726 * +---------------------+----------+ | o L2ARC eligible
6727 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6728 * +---------------------+----------+ |
6729 * 15.9 Gbytes ^ 32 Mbytes |
6731 * l2arc_feed_thread()
6733 * l2arc write hand <--[oooo]--'
6737 * +==============================+
6738 * L2ARC dev |####|#|###|###| |####| ... |
6739 * +==============================+
6742 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6743 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6744 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6745 * safe to say that this is an uncommon case, since buffers at the end of
6746 * the ARC lists have moved there due to inactivity.
6748 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6749 * then the L2ARC simply misses copying some buffers. This serves as a
6750 * pressure valve to prevent heavy read workloads from both stalling the ARC
6751 * with waits and clogging the L2ARC with writes. This also helps prevent
6752 * the potential for the L2ARC to churn if it attempts to cache content too
6753 * quickly, such as during backups of the entire pool.
6755 * 5. After system boot and before the ARC has filled main memory, there are
6756 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6757 * lists can remain mostly static. Instead of searching from tail of these
6758 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6759 * for eligible buffers, greatly increasing its chance of finding them.
6761 * The L2ARC device write speed is also boosted during this time so that
6762 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6763 * there are no L2ARC reads, and no fear of degrading read performance
6764 * through increased writes.
6766 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6767 * the vdev queue can aggregate them into larger and fewer writes. Each
6768 * device is written to in a rotor fashion, sweeping writes through
6769 * available space then repeating.
6771 * 7. The L2ARC does not store dirty content. It never needs to flush
6772 * write buffers back to disk based storage.
6774 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6775 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6777 * The performance of the L2ARC can be tweaked by a number of tunables, which
6778 * may be necessary for different workloads:
6780 * l2arc_write_max max write bytes per interval
6781 * l2arc_write_boost extra write bytes during device warmup
6782 * l2arc_noprefetch skip caching prefetched buffers
6783 * l2arc_headroom number of max device writes to precache
6784 * l2arc_headroom_boost when we find compressed buffers during ARC
6785 * scanning, we multiply headroom by this
6786 * percentage factor for the next scan cycle,
6787 * since more compressed buffers are likely to
6789 * l2arc_feed_secs seconds between L2ARC writing
6791 * Tunables may be removed or added as future performance improvements are
6792 * integrated, and also may become zpool properties.
6794 * There are three key functions that control how the L2ARC warms up:
6796 * l2arc_write_eligible() check if a buffer is eligible to cache
6797 * l2arc_write_size() calculate how much to write
6798 * l2arc_write_interval() calculate sleep delay between writes
6800 * These three functions determine what to write, how much, and how quickly
6805 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6808 * A buffer is *not* eligible for the L2ARC if it:
6809 * 1. belongs to a different spa.
6810 * 2. is already cached on the L2ARC.
6811 * 3. has an I/O in progress (it may be an incomplete read).
6812 * 4. is flagged not eligible (zfs property).
6814 if (hdr->b_spa != spa_guid) {
6815 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
6818 if (HDR_HAS_L2HDR(hdr)) {
6819 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
6822 if (HDR_IO_IN_PROGRESS(hdr)) {
6823 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
6826 if (!HDR_L2CACHE(hdr)) {
6827 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
6835 l2arc_write_size(void)
6840 * Make sure our globals have meaningful values in case the user
6843 size = l2arc_write_max;
6845 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
6846 "be greater than zero, resetting it to the default (%d)",
6848 size = l2arc_write_max = L2ARC_WRITE_SIZE;
6851 if (arc_warm == B_FALSE)
6852 size += l2arc_write_boost;
6859 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
6861 clock_t interval, next, now;
6864 * If the ARC lists are busy, increase our write rate; if the
6865 * lists are stale, idle back. This is achieved by checking
6866 * how much we previously wrote - if it was more than half of
6867 * what we wanted, schedule the next write much sooner.
6869 if (l2arc_feed_again && wrote > (wanted / 2))
6870 interval = (hz * l2arc_feed_min_ms) / 1000;
6872 interval = hz * l2arc_feed_secs;
6874 now = ddi_get_lbolt();
6875 next = MAX(now, MIN(now + interval, began + interval));
6881 * Cycle through L2ARC devices. This is how L2ARC load balances.
6882 * If a device is returned, this also returns holding the spa config lock.
6884 static l2arc_dev_t *
6885 l2arc_dev_get_next(void)
6887 l2arc_dev_t *first, *next = NULL;
6890 * Lock out the removal of spas (spa_namespace_lock), then removal
6891 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6892 * both locks will be dropped and a spa config lock held instead.
6894 mutex_enter(&spa_namespace_lock);
6895 mutex_enter(&l2arc_dev_mtx);
6897 /* if there are no vdevs, there is nothing to do */
6898 if (l2arc_ndev == 0)
6902 next = l2arc_dev_last;
6904 /* loop around the list looking for a non-faulted vdev */
6906 next = list_head(l2arc_dev_list);
6908 next = list_next(l2arc_dev_list, next);
6910 next = list_head(l2arc_dev_list);
6913 /* if we have come back to the start, bail out */
6916 else if (next == first)
6919 } while (vdev_is_dead(next->l2ad_vdev));
6921 /* if we were unable to find any usable vdevs, return NULL */
6922 if (vdev_is_dead(next->l2ad_vdev))
6925 l2arc_dev_last = next;
6928 mutex_exit(&l2arc_dev_mtx);
6931 * Grab the config lock to prevent the 'next' device from being
6932 * removed while we are writing to it.
6935 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
6936 mutex_exit(&spa_namespace_lock);
6942 * Free buffers that were tagged for destruction.
6945 l2arc_do_free_on_write()
6948 l2arc_data_free_t *df, *df_prev;
6950 mutex_enter(&l2arc_free_on_write_mtx);
6951 buflist = l2arc_free_on_write;
6953 for (df = list_tail(buflist); df; df = df_prev) {
6954 df_prev = list_prev(buflist, df);
6955 ASSERT3P(df->l2df_abd, !=, NULL);
6956 abd_free(df->l2df_abd);
6957 list_remove(buflist, df);
6958 kmem_free(df, sizeof (l2arc_data_free_t));
6961 mutex_exit(&l2arc_free_on_write_mtx);
6965 * A write to a cache device has completed. Update all headers to allow
6966 * reads from these buffers to begin.
6969 l2arc_write_done(zio_t *zio)
6971 l2arc_write_callback_t *cb;
6974 arc_buf_hdr_t *head, *hdr, *hdr_prev;
6975 kmutex_t *hash_lock;
6976 int64_t bytes_dropped = 0;
6978 cb = zio->io_private;
6979 ASSERT3P(cb, !=, NULL);
6980 dev = cb->l2wcb_dev;
6981 ASSERT3P(dev, !=, NULL);
6982 head = cb->l2wcb_head;
6983 ASSERT3P(head, !=, NULL);
6984 buflist = &dev->l2ad_buflist;
6985 ASSERT3P(buflist, !=, NULL);
6986 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
6987 l2arc_write_callback_t *, cb);
6989 if (zio->io_error != 0)
6990 ARCSTAT_BUMP(arcstat_l2_writes_error);
6993 * All writes completed, or an error was hit.
6996 mutex_enter(&dev->l2ad_mtx);
6997 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
6998 hdr_prev = list_prev(buflist, hdr);
7000 hash_lock = HDR_LOCK(hdr);
7003 * We cannot use mutex_enter or else we can deadlock
7004 * with l2arc_write_buffers (due to swapping the order
7005 * the hash lock and l2ad_mtx are taken).
7007 if (!mutex_tryenter(hash_lock)) {
7009 * Missed the hash lock. We must retry so we
7010 * don't leave the ARC_FLAG_L2_WRITING bit set.
7012 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7015 * We don't want to rescan the headers we've
7016 * already marked as having been written out, so
7017 * we reinsert the head node so we can pick up
7018 * where we left off.
7020 list_remove(buflist, head);
7021 list_insert_after(buflist, hdr, head);
7023 mutex_exit(&dev->l2ad_mtx);
7026 * We wait for the hash lock to become available
7027 * to try and prevent busy waiting, and increase
7028 * the chance we'll be able to acquire the lock
7029 * the next time around.
7031 mutex_enter(hash_lock);
7032 mutex_exit(hash_lock);
7037 * We could not have been moved into the arc_l2c_only
7038 * state while in-flight due to our ARC_FLAG_L2_WRITING
7039 * bit being set. Let's just ensure that's being enforced.
7041 ASSERT(HDR_HAS_L1HDR(hdr));
7043 if (zio->io_error != 0) {
7045 * Error - drop L2ARC entry.
7047 list_remove(buflist, hdr);
7049 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7051 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
7052 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
7054 bytes_dropped += arc_hdr_size(hdr);
7055 (void) refcount_remove_many(&dev->l2ad_alloc,
7056 arc_hdr_size(hdr), hdr);
7060 * Allow ARC to begin reads and ghost list evictions to
7063 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7065 mutex_exit(hash_lock);
7068 atomic_inc_64(&l2arc_writes_done);
7069 list_remove(buflist, head);
7070 ASSERT(!HDR_HAS_L1HDR(head));
7071 kmem_cache_free(hdr_l2only_cache, head);
7072 mutex_exit(&dev->l2ad_mtx);
7074 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7076 l2arc_do_free_on_write();
7078 kmem_free(cb, sizeof (l2arc_write_callback_t));
7082 * A read to a cache device completed. Validate buffer contents before
7083 * handing over to the regular ARC routines.
7086 l2arc_read_done(zio_t *zio)
7088 l2arc_read_callback_t *cb;
7090 kmutex_t *hash_lock;
7091 boolean_t valid_cksum;
7093 ASSERT3P(zio->io_vd, !=, NULL);
7094 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7096 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7098 cb = zio->io_private;
7099 ASSERT3P(cb, !=, NULL);
7100 hdr = cb->l2rcb_hdr;
7101 ASSERT3P(hdr, !=, NULL);
7103 hash_lock = HDR_LOCK(hdr);
7104 mutex_enter(hash_lock);
7105 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7108 * If the data was read into a temporary buffer,
7109 * move it and free the buffer.
7111 if (cb->l2rcb_abd != NULL) {
7112 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7113 if (zio->io_error == 0) {
7114 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
7119 * The following must be done regardless of whether
7120 * there was an error:
7121 * - free the temporary buffer
7122 * - point zio to the real ARC buffer
7123 * - set zio size accordingly
7124 * These are required because zio is either re-used for
7125 * an I/O of the block in the case of the error
7126 * or the zio is passed to arc_read_done() and it
7129 abd_free(cb->l2rcb_abd);
7130 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7131 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
7134 ASSERT3P(zio->io_abd, !=, NULL);
7137 * Check this survived the L2ARC journey.
7139 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
7140 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
7141 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
7143 valid_cksum = arc_cksum_is_equal(hdr, zio);
7144 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
7145 mutex_exit(hash_lock);
7146 zio->io_private = hdr;
7149 mutex_exit(hash_lock);
7151 * Buffer didn't survive caching. Increment stats and
7152 * reissue to the original storage device.
7154 if (zio->io_error != 0) {
7155 ARCSTAT_BUMP(arcstat_l2_io_error);
7157 zio->io_error = SET_ERROR(EIO);
7160 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
7163 * If there's no waiter, issue an async i/o to the primary
7164 * storage now. If there *is* a waiter, the caller must
7165 * issue the i/o in a context where it's OK to block.
7167 if (zio->io_waiter == NULL) {
7168 zio_t *pio = zio_unique_parent(zio);
7170 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7172 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
7173 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
7174 hdr, zio->io_priority, cb->l2rcb_flags,
7179 kmem_free(cb, sizeof (l2arc_read_callback_t));
7183 * This is the list priority from which the L2ARC will search for pages to
7184 * cache. This is used within loops (0..3) to cycle through lists in the
7185 * desired order. This order can have a significant effect on cache
7188 * Currently the metadata lists are hit first, MFU then MRU, followed by
7189 * the data lists. This function returns a locked list, and also returns
7192 static multilist_sublist_t *
7193 l2arc_sublist_lock(int list_num)
7195 multilist_t *ml = NULL;
7198 ASSERT(list_num >= 0 && list_num <= 3);
7202 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7205 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7208 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7211 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7216 * Return a randomly-selected sublist. This is acceptable
7217 * because the caller feeds only a little bit of data for each
7218 * call (8MB). Subsequent calls will result in different
7219 * sublists being selected.
7221 idx = multilist_get_random_index(ml);
7222 return (multilist_sublist_lock(ml, idx));
7226 * Evict buffers from the device write hand to the distance specified in
7227 * bytes. This distance may span populated buffers, it may span nothing.
7228 * This is clearing a region on the L2ARC device ready for writing.
7229 * If the 'all' boolean is set, every buffer is evicted.
7232 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7235 arc_buf_hdr_t *hdr, *hdr_prev;
7236 kmutex_t *hash_lock;
7239 buflist = &dev->l2ad_buflist;
7241 if (!all && dev->l2ad_first) {
7243 * This is the first sweep through the device. There is
7249 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
7251 * When nearing the end of the device, evict to the end
7252 * before the device write hand jumps to the start.
7254 taddr = dev->l2ad_end;
7256 taddr = dev->l2ad_hand + distance;
7258 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
7259 uint64_t, taddr, boolean_t, all);
7262 mutex_enter(&dev->l2ad_mtx);
7263 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
7264 hdr_prev = list_prev(buflist, hdr);
7266 hash_lock = HDR_LOCK(hdr);
7269 * We cannot use mutex_enter or else we can deadlock
7270 * with l2arc_write_buffers (due to swapping the order
7271 * the hash lock and l2ad_mtx are taken).
7273 if (!mutex_tryenter(hash_lock)) {
7275 * Missed the hash lock. Retry.
7277 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
7278 mutex_exit(&dev->l2ad_mtx);
7279 mutex_enter(hash_lock);
7280 mutex_exit(hash_lock);
7285 * A header can't be on this list if it doesn't have L2 header.
7287 ASSERT(HDR_HAS_L2HDR(hdr));
7289 /* Ensure this header has finished being written. */
7290 ASSERT(!HDR_L2_WRITING(hdr));
7291 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
7293 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
7294 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
7296 * We've evicted to the target address,
7297 * or the end of the device.
7299 mutex_exit(hash_lock);
7303 if (!HDR_HAS_L1HDR(hdr)) {
7304 ASSERT(!HDR_L2_READING(hdr));
7306 * This doesn't exist in the ARC. Destroy.
7307 * arc_hdr_destroy() will call list_remove()
7308 * and decrement arcstat_l2_lsize.
7310 arc_change_state(arc_anon, hdr, hash_lock);
7311 arc_hdr_destroy(hdr);
7313 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
7314 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
7316 * Invalidate issued or about to be issued
7317 * reads, since we may be about to write
7318 * over this location.
7320 if (HDR_L2_READING(hdr)) {
7321 ARCSTAT_BUMP(arcstat_l2_evict_reading);
7322 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
7325 arc_hdr_l2hdr_destroy(hdr);
7327 mutex_exit(hash_lock);
7329 mutex_exit(&dev->l2ad_mtx);
7333 * Find and write ARC buffers to the L2ARC device.
7335 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7336 * for reading until they have completed writing.
7337 * The headroom_boost is an in-out parameter used to maintain headroom boost
7338 * state between calls to this function.
7340 * Returns the number of bytes actually written (which may be smaller than
7341 * the delta by which the device hand has changed due to alignment).
7344 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
7346 arc_buf_hdr_t *hdr, *hdr_prev, *head;
7347 uint64_t write_asize, write_psize, write_lsize, headroom;
7349 l2arc_write_callback_t *cb;
7351 uint64_t guid = spa_load_guid(spa);
7354 ASSERT3P(dev->l2ad_vdev, !=, NULL);
7357 write_lsize = write_asize = write_psize = 0;
7359 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
7360 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
7362 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
7364 * Copy buffers for L2ARC writing.
7366 for (try = 0; try <= 3; try++) {
7367 multilist_sublist_t *mls = l2arc_sublist_lock(try);
7368 uint64_t passed_sz = 0;
7370 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
7373 * L2ARC fast warmup.
7375 * Until the ARC is warm and starts to evict, read from the
7376 * head of the ARC lists rather than the tail.
7378 if (arc_warm == B_FALSE)
7379 hdr = multilist_sublist_head(mls);
7381 hdr = multilist_sublist_tail(mls);
7383 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
7385 headroom = target_sz * l2arc_headroom;
7386 if (zfs_compressed_arc_enabled)
7387 headroom = (headroom * l2arc_headroom_boost) / 100;
7389 for (; hdr; hdr = hdr_prev) {
7390 kmutex_t *hash_lock;
7392 if (arc_warm == B_FALSE)
7393 hdr_prev = multilist_sublist_next(mls, hdr);
7395 hdr_prev = multilist_sublist_prev(mls, hdr);
7396 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
7397 HDR_GET_LSIZE(hdr));
7399 hash_lock = HDR_LOCK(hdr);
7400 if (!mutex_tryenter(hash_lock)) {
7401 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
7403 * Skip this buffer rather than waiting.
7408 passed_sz += HDR_GET_LSIZE(hdr);
7409 if (passed_sz > headroom) {
7413 mutex_exit(hash_lock);
7414 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
7418 if (!l2arc_write_eligible(guid, hdr)) {
7419 mutex_exit(hash_lock);
7424 * We rely on the L1 portion of the header below, so
7425 * it's invalid for this header to have been evicted out
7426 * of the ghost cache, prior to being written out. The
7427 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7429 ASSERT(HDR_HAS_L1HDR(hdr));
7431 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
7432 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7433 ASSERT3U(arc_hdr_size(hdr), >, 0);
7434 uint64_t psize = arc_hdr_size(hdr);
7435 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
7438 if ((write_asize + asize) > target_sz) {
7440 mutex_exit(hash_lock);
7441 ARCSTAT_BUMP(arcstat_l2_write_full);
7447 * Insert a dummy header on the buflist so
7448 * l2arc_write_done() can find where the
7449 * write buffers begin without searching.
7451 mutex_enter(&dev->l2ad_mtx);
7452 list_insert_head(&dev->l2ad_buflist, head);
7453 mutex_exit(&dev->l2ad_mtx);
7456 sizeof (l2arc_write_callback_t), KM_SLEEP);
7457 cb->l2wcb_dev = dev;
7458 cb->l2wcb_head = head;
7459 pio = zio_root(spa, l2arc_write_done, cb,
7461 ARCSTAT_BUMP(arcstat_l2_write_pios);
7464 hdr->b_l2hdr.b_dev = dev;
7465 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7466 arc_hdr_set_flags(hdr,
7467 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7469 mutex_enter(&dev->l2ad_mtx);
7470 list_insert_head(&dev->l2ad_buflist, hdr);
7471 mutex_exit(&dev->l2ad_mtx);
7473 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
7476 * Normally the L2ARC can use the hdr's data, but if
7477 * we're sharing data between the hdr and one of its
7478 * bufs, L2ARC needs its own copy of the data so that
7479 * the ZIO below can't race with the buf consumer.
7480 * Another case where we need to create a copy of the
7481 * data is when the buffer size is not device-aligned
7482 * and we need to pad the block to make it such.
7483 * That also keeps the clock hand suitably aligned.
7485 * To ensure that the copy will be available for the
7486 * lifetime of the ZIO and be cleaned up afterwards, we
7487 * add it to the l2arc_free_on_write queue.
7490 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
7491 to_write = hdr->b_l1hdr.b_pabd;
7493 to_write = abd_alloc_for_io(asize,
7494 HDR_ISTYPE_METADATA(hdr));
7495 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
7496 if (asize != psize) {
7497 abd_zero_off(to_write, psize,
7500 l2arc_free_abd_on_write(to_write, asize,
7503 wzio = zio_write_phys(pio, dev->l2ad_vdev,
7504 hdr->b_l2hdr.b_daddr, asize, to_write,
7505 ZIO_CHECKSUM_OFF, NULL, hdr,
7506 ZIO_PRIORITY_ASYNC_WRITE,
7507 ZIO_FLAG_CANFAIL, B_FALSE);
7509 write_lsize += HDR_GET_LSIZE(hdr);
7510 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7513 write_psize += psize;
7514 write_asize += asize;
7515 dev->l2ad_hand += asize;
7517 mutex_exit(hash_lock);
7519 (void) zio_nowait(wzio);
7522 multilist_sublist_unlock(mls);
7528 /* No buffers selected for writing? */
7530 ASSERT0(write_lsize);
7531 ASSERT(!HDR_HAS_L1HDR(head));
7532 kmem_cache_free(hdr_l2only_cache, head);
7536 ASSERT3U(write_psize, <=, target_sz);
7537 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7538 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
7539 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
7540 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
7541 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
7544 * Bump device hand to the device start if it is approaching the end.
7545 * l2arc_evict() will already have evicted ahead for this case.
7547 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7548 dev->l2ad_hand = dev->l2ad_start;
7549 dev->l2ad_first = B_FALSE;
7552 dev->l2ad_writing = B_TRUE;
7553 (void) zio_wait(pio);
7554 dev->l2ad_writing = B_FALSE;
7556 return (write_asize);
7560 * This thread feeds the L2ARC at regular intervals. This is the beating
7561 * heart of the L2ARC.
7564 l2arc_feed_thread(void *dummy __unused)
7569 uint64_t size, wrote;
7570 clock_t begin, next = ddi_get_lbolt();
7572 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7574 mutex_enter(&l2arc_feed_thr_lock);
7576 while (l2arc_thread_exit == 0) {
7577 CALLB_CPR_SAFE_BEGIN(&cpr);
7578 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7579 next - ddi_get_lbolt());
7580 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7581 next = ddi_get_lbolt() + hz;
7584 * Quick check for L2ARC devices.
7586 mutex_enter(&l2arc_dev_mtx);
7587 if (l2arc_ndev == 0) {
7588 mutex_exit(&l2arc_dev_mtx);
7591 mutex_exit(&l2arc_dev_mtx);
7592 begin = ddi_get_lbolt();
7595 * This selects the next l2arc device to write to, and in
7596 * doing so the next spa to feed from: dev->l2ad_spa. This
7597 * will return NULL if there are now no l2arc devices or if
7598 * they are all faulted.
7600 * If a device is returned, its spa's config lock is also
7601 * held to prevent device removal. l2arc_dev_get_next()
7602 * will grab and release l2arc_dev_mtx.
7604 if ((dev = l2arc_dev_get_next()) == NULL)
7607 spa = dev->l2ad_spa;
7608 ASSERT3P(spa, !=, NULL);
7611 * If the pool is read-only then force the feed thread to
7612 * sleep a little longer.
7614 if (!spa_writeable(spa)) {
7615 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7616 spa_config_exit(spa, SCL_L2ARC, dev);
7621 * Avoid contributing to memory pressure.
7623 if (arc_reclaim_needed()) {
7624 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7625 spa_config_exit(spa, SCL_L2ARC, dev);
7629 ARCSTAT_BUMP(arcstat_l2_feeds);
7631 size = l2arc_write_size();
7634 * Evict L2ARC buffers that will be overwritten.
7636 l2arc_evict(dev, size, B_FALSE);
7639 * Write ARC buffers.
7641 wrote = l2arc_write_buffers(spa, dev, size);
7644 * Calculate interval between writes.
7646 next = l2arc_write_interval(begin, size, wrote);
7647 spa_config_exit(spa, SCL_L2ARC, dev);
7650 l2arc_thread_exit = 0;
7651 cv_broadcast(&l2arc_feed_thr_cv);
7652 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7657 l2arc_vdev_present(vdev_t *vd)
7661 mutex_enter(&l2arc_dev_mtx);
7662 for (dev = list_head(l2arc_dev_list); dev != NULL;
7663 dev = list_next(l2arc_dev_list, dev)) {
7664 if (dev->l2ad_vdev == vd)
7667 mutex_exit(&l2arc_dev_mtx);
7669 return (dev != NULL);
7673 * Add a vdev for use by the L2ARC. By this point the spa has already
7674 * validated the vdev and opened it.
7677 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7679 l2arc_dev_t *adddev;
7681 ASSERT(!l2arc_vdev_present(vd));
7683 vdev_ashift_optimize(vd);
7686 * Create a new l2arc device entry.
7688 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7689 adddev->l2ad_spa = spa;
7690 adddev->l2ad_vdev = vd;
7691 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7692 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7693 adddev->l2ad_hand = adddev->l2ad_start;
7694 adddev->l2ad_first = B_TRUE;
7695 adddev->l2ad_writing = B_FALSE;
7697 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7699 * This is a list of all ARC buffers that are still valid on the
7702 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7703 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7705 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7706 refcount_create(&adddev->l2ad_alloc);
7709 * Add device to global list
7711 mutex_enter(&l2arc_dev_mtx);
7712 list_insert_head(l2arc_dev_list, adddev);
7713 atomic_inc_64(&l2arc_ndev);
7714 mutex_exit(&l2arc_dev_mtx);
7718 * Remove a vdev from the L2ARC.
7721 l2arc_remove_vdev(vdev_t *vd)
7723 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7726 * Find the device by vdev
7728 mutex_enter(&l2arc_dev_mtx);
7729 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7730 nextdev = list_next(l2arc_dev_list, dev);
7731 if (vd == dev->l2ad_vdev) {
7736 ASSERT3P(remdev, !=, NULL);
7739 * Remove device from global list
7741 list_remove(l2arc_dev_list, remdev);
7742 l2arc_dev_last = NULL; /* may have been invalidated */
7743 atomic_dec_64(&l2arc_ndev);
7744 mutex_exit(&l2arc_dev_mtx);
7747 * Clear all buflists and ARC references. L2ARC device flush.
7749 l2arc_evict(remdev, 0, B_TRUE);
7750 list_destroy(&remdev->l2ad_buflist);
7751 mutex_destroy(&remdev->l2ad_mtx);
7752 refcount_destroy(&remdev->l2ad_alloc);
7753 kmem_free(remdev, sizeof (l2arc_dev_t));
7759 l2arc_thread_exit = 0;
7761 l2arc_writes_sent = 0;
7762 l2arc_writes_done = 0;
7764 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7765 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7766 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7767 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7769 l2arc_dev_list = &L2ARC_dev_list;
7770 l2arc_free_on_write = &L2ARC_free_on_write;
7771 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7772 offsetof(l2arc_dev_t, l2ad_node));
7773 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7774 offsetof(l2arc_data_free_t, l2df_list_node));
7781 * This is called from dmu_fini(), which is called from spa_fini();
7782 * Because of this, we can assume that all l2arc devices have
7783 * already been removed when the pools themselves were removed.
7786 l2arc_do_free_on_write();
7788 mutex_destroy(&l2arc_feed_thr_lock);
7789 cv_destroy(&l2arc_feed_thr_cv);
7790 mutex_destroy(&l2arc_dev_mtx);
7791 mutex_destroy(&l2arc_free_on_write_mtx);
7793 list_destroy(l2arc_dev_list);
7794 list_destroy(l2arc_free_on_write);
7800 if (!(spa_mode_global & FWRITE))
7803 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7804 TS_RUN, minclsyspri);
7810 if (!(spa_mode_global & FWRITE))
7813 mutex_enter(&l2arc_feed_thr_lock);
7814 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7815 l2arc_thread_exit = 1;
7816 while (l2arc_thread_exit != 0)
7817 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7818 mutex_exit(&l2arc_feed_thr_lock);