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 * log2 fraction of the zio arena to keep free.
355 int arc_zio_arena_free_shift = 2;
358 * These tunables are for performance analysis.
360 uint64_t zfs_arc_max;
361 uint64_t zfs_arc_min;
362 uint64_t zfs_arc_meta_limit = 0;
363 uint64_t zfs_arc_meta_min = 0;
364 int zfs_arc_grow_retry = 0;
365 int zfs_arc_shrink_shift = 0;
366 int zfs_arc_no_grow_shift = 0;
367 int zfs_arc_p_min_shift = 0;
368 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
369 u_int zfs_arc_free_target = 0;
371 /* Absolute min for arc min / max is 16MB. */
372 static uint64_t arc_abs_min = 16 << 20;
374 boolean_t zfs_compressed_arc_enabled = B_TRUE;
376 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
377 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
378 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
379 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
380 static int sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS);
382 #if defined(__FreeBSD__) && defined(_KERNEL)
384 arc_free_target_init(void *unused __unused)
387 zfs_arc_free_target = (vm_cnt.v_free_min / 10) * 11;
389 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
390 arc_free_target_init, NULL);
392 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
393 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
394 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
395 TUNABLE_INT("vfs.zfs.arc_grow_retry", &zfs_arc_grow_retry);
396 TUNABLE_INT("vfs.zfs.arc_no_grow_shift", &zfs_arc_no_grow_shift);
397 SYSCTL_DECL(_vfs_zfs);
398 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
399 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
400 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
401 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
402 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_no_grow_shift, CTLTYPE_U32 | CTLFLAG_RWTUN,
403 0, sizeof(uint32_t), sysctl_vfs_zfs_arc_no_grow_shift, "U",
404 "log2(fraction of ARC which must be free to allow growing)");
405 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
406 &zfs_arc_average_blocksize, 0,
407 "ARC average blocksize");
408 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
409 &arc_shrink_shift, 0,
410 "log2(fraction of arc to reclaim)");
411 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_grow_retry, CTLFLAG_RW,
413 "Wait in seconds before considering growing ARC");
414 SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN,
415 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC");
418 * We don't have a tunable for arc_free_target due to the dependency on
419 * pagedaemon initialisation.
421 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
422 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
423 sysctl_vfs_zfs_arc_free_target, "IU",
424 "Desired number of free pages below which ARC triggers reclaim");
427 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
432 val = zfs_arc_free_target;
433 err = sysctl_handle_int(oidp, &val, 0, req);
434 if (err != 0 || req->newptr == NULL)
439 if (val > vm_cnt.v_page_count)
442 zfs_arc_free_target = val;
448 * Must be declared here, before the definition of corresponding kstat
449 * macro which uses the same names will confuse the compiler.
451 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
452 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
453 sysctl_vfs_zfs_arc_meta_limit, "QU",
454 "ARC metadata limit");
458 * Note that buffers can be in one of 6 states:
459 * ARC_anon - anonymous (discussed below)
460 * ARC_mru - recently used, currently cached
461 * ARC_mru_ghost - recentely used, no longer in cache
462 * ARC_mfu - frequently used, currently cached
463 * ARC_mfu_ghost - frequently used, no longer in cache
464 * ARC_l2c_only - exists in L2ARC but not other states
465 * When there are no active references to the buffer, they are
466 * are linked onto a list in one of these arc states. These are
467 * the only buffers that can be evicted or deleted. Within each
468 * state there are multiple lists, one for meta-data and one for
469 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
470 * etc.) is tracked separately so that it can be managed more
471 * explicitly: favored over data, limited explicitly.
473 * Anonymous buffers are buffers that are not associated with
474 * a DVA. These are buffers that hold dirty block copies
475 * before they are written to stable storage. By definition,
476 * they are "ref'd" and are considered part of arc_mru
477 * that cannot be freed. Generally, they will aquire a DVA
478 * as they are written and migrate onto the arc_mru list.
480 * The ARC_l2c_only state is for buffers that are in the second
481 * level ARC but no longer in any of the ARC_m* lists. The second
482 * level ARC itself may also contain buffers that are in any of
483 * the ARC_m* states - meaning that a buffer can exist in two
484 * places. The reason for the ARC_l2c_only state is to keep the
485 * buffer header in the hash table, so that reads that hit the
486 * second level ARC benefit from these fast lookups.
489 typedef struct arc_state {
491 * list of evictable buffers
493 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
495 * total amount of evictable data in this state
497 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
499 * total amount of data in this state; this includes: evictable,
500 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
502 refcount_t arcs_size;
506 static arc_state_t ARC_anon;
507 static arc_state_t ARC_mru;
508 static arc_state_t ARC_mru_ghost;
509 static arc_state_t ARC_mfu;
510 static arc_state_t ARC_mfu_ghost;
511 static arc_state_t ARC_l2c_only;
513 typedef struct arc_stats {
514 kstat_named_t arcstat_hits;
515 kstat_named_t arcstat_misses;
516 kstat_named_t arcstat_demand_data_hits;
517 kstat_named_t arcstat_demand_data_misses;
518 kstat_named_t arcstat_demand_metadata_hits;
519 kstat_named_t arcstat_demand_metadata_misses;
520 kstat_named_t arcstat_prefetch_data_hits;
521 kstat_named_t arcstat_prefetch_data_misses;
522 kstat_named_t arcstat_prefetch_metadata_hits;
523 kstat_named_t arcstat_prefetch_metadata_misses;
524 kstat_named_t arcstat_mru_hits;
525 kstat_named_t arcstat_mru_ghost_hits;
526 kstat_named_t arcstat_mfu_hits;
527 kstat_named_t arcstat_mfu_ghost_hits;
528 kstat_named_t arcstat_allocated;
529 kstat_named_t arcstat_deleted;
531 * Number of buffers that could not be evicted because the hash lock
532 * was held by another thread. The lock may not necessarily be held
533 * by something using the same buffer, since hash locks are shared
534 * by multiple buffers.
536 kstat_named_t arcstat_mutex_miss;
538 * Number of buffers skipped because they have I/O in progress, are
539 * indrect prefetch buffers that have not lived long enough, or are
540 * not from the spa we're trying to evict from.
542 kstat_named_t arcstat_evict_skip;
544 * Number of times arc_evict_state() was unable to evict enough
545 * buffers to reach it's target amount.
547 kstat_named_t arcstat_evict_not_enough;
548 kstat_named_t arcstat_evict_l2_cached;
549 kstat_named_t arcstat_evict_l2_eligible;
550 kstat_named_t arcstat_evict_l2_ineligible;
551 kstat_named_t arcstat_evict_l2_skip;
552 kstat_named_t arcstat_hash_elements;
553 kstat_named_t arcstat_hash_elements_max;
554 kstat_named_t arcstat_hash_collisions;
555 kstat_named_t arcstat_hash_chains;
556 kstat_named_t arcstat_hash_chain_max;
557 kstat_named_t arcstat_p;
558 kstat_named_t arcstat_c;
559 kstat_named_t arcstat_c_min;
560 kstat_named_t arcstat_c_max;
561 kstat_named_t arcstat_size;
563 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
564 * Note that the compressed bytes may match the uncompressed bytes
565 * if the block is either not compressed or compressed arc is disabled.
567 kstat_named_t arcstat_compressed_size;
569 * Uncompressed size of the data stored in b_pabd. If compressed
570 * arc is disabled then this value will be identical to the stat
573 kstat_named_t arcstat_uncompressed_size;
575 * Number of bytes stored in all the arc_buf_t's. This is classified
576 * as "overhead" since this data is typically short-lived and will
577 * be evicted from the arc when it becomes unreferenced unless the
578 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
579 * values have been set (see comment in dbuf.c for more information).
581 kstat_named_t arcstat_overhead_size;
583 * Number of bytes consumed by internal ARC structures necessary
584 * for tracking purposes; these structures are not actually
585 * backed by ARC buffers. This includes arc_buf_hdr_t structures
586 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
587 * caches), and arc_buf_t structures (allocated via arc_buf_t
590 kstat_named_t arcstat_hdr_size;
592 * Number of bytes consumed by ARC buffers of type equal to
593 * ARC_BUFC_DATA. This is generally consumed by buffers backing
594 * on disk user data (e.g. plain file contents).
596 kstat_named_t arcstat_data_size;
598 * Number of bytes consumed by ARC buffers of type equal to
599 * ARC_BUFC_METADATA. This is generally consumed by buffers
600 * backing on disk data that is used for internal ZFS
601 * structures (e.g. ZAP, dnode, indirect blocks, etc).
603 kstat_named_t arcstat_metadata_size;
605 * Number of bytes consumed by various buffers and structures
606 * not actually backed with ARC buffers. This includes bonus
607 * buffers (allocated directly via zio_buf_* functions),
608 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
609 * cache), and dnode_t structures (allocated via dnode_t cache).
611 kstat_named_t arcstat_other_size;
613 * Total number of bytes consumed by ARC buffers residing in the
614 * arc_anon state. This includes *all* buffers in the arc_anon
615 * state; e.g. data, metadata, evictable, and unevictable buffers
616 * are all included in this value.
618 kstat_named_t arcstat_anon_size;
620 * Number of bytes consumed by ARC buffers that meet the
621 * following criteria: backing buffers of type ARC_BUFC_DATA,
622 * residing in the arc_anon state, and are eligible for eviction
623 * (e.g. have no outstanding holds on the buffer).
625 kstat_named_t arcstat_anon_evictable_data;
627 * Number of bytes consumed by ARC buffers that meet the
628 * following criteria: backing buffers of type ARC_BUFC_METADATA,
629 * residing in the arc_anon state, and are eligible for eviction
630 * (e.g. have no outstanding holds on the buffer).
632 kstat_named_t arcstat_anon_evictable_metadata;
634 * Total number of bytes consumed by ARC buffers residing in the
635 * arc_mru state. This includes *all* buffers in the arc_mru
636 * state; e.g. data, metadata, evictable, and unevictable buffers
637 * are all included in this value.
639 kstat_named_t arcstat_mru_size;
641 * Number of bytes consumed by ARC buffers that meet the
642 * following criteria: backing buffers of type ARC_BUFC_DATA,
643 * residing in the arc_mru state, and are eligible for eviction
644 * (e.g. have no outstanding holds on the buffer).
646 kstat_named_t arcstat_mru_evictable_data;
648 * Number of bytes consumed by ARC buffers that meet the
649 * following criteria: backing buffers of type ARC_BUFC_METADATA,
650 * residing in the arc_mru state, and are eligible for eviction
651 * (e.g. have no outstanding holds on the buffer).
653 kstat_named_t arcstat_mru_evictable_metadata;
655 * Total number of bytes that *would have been* consumed by ARC
656 * buffers in the arc_mru_ghost state. The key thing to note
657 * here, is the fact that this size doesn't actually indicate
658 * RAM consumption. The ghost lists only consist of headers and
659 * don't actually have ARC buffers linked off of these headers.
660 * Thus, *if* the headers had associated ARC buffers, these
661 * buffers *would have* consumed this number of bytes.
663 kstat_named_t arcstat_mru_ghost_size;
665 * Number of bytes that *would have been* consumed by ARC
666 * buffers that are eligible for eviction, of type
667 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
669 kstat_named_t arcstat_mru_ghost_evictable_data;
671 * Number of bytes that *would have been* consumed by ARC
672 * buffers that are eligible for eviction, of type
673 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
675 kstat_named_t arcstat_mru_ghost_evictable_metadata;
677 * Total number of bytes consumed by ARC buffers residing in the
678 * arc_mfu state. This includes *all* buffers in the arc_mfu
679 * state; e.g. data, metadata, evictable, and unevictable buffers
680 * are all included in this value.
682 kstat_named_t arcstat_mfu_size;
684 * Number of bytes consumed by ARC buffers that are eligible for
685 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
688 kstat_named_t arcstat_mfu_evictable_data;
690 * Number of bytes consumed by ARC buffers that are eligible for
691 * eviction, of type ARC_BUFC_METADATA, and reside in the
694 kstat_named_t arcstat_mfu_evictable_metadata;
696 * Total number of bytes that *would have been* consumed by ARC
697 * buffers in the arc_mfu_ghost state. See the comment above
698 * arcstat_mru_ghost_size for more details.
700 kstat_named_t arcstat_mfu_ghost_size;
702 * Number of bytes that *would have been* consumed by ARC
703 * buffers that are eligible for eviction, of type
704 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
706 kstat_named_t arcstat_mfu_ghost_evictable_data;
708 * Number of bytes that *would have been* consumed by ARC
709 * buffers that are eligible for eviction, of type
710 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
712 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
713 kstat_named_t arcstat_l2_hits;
714 kstat_named_t arcstat_l2_misses;
715 kstat_named_t arcstat_l2_feeds;
716 kstat_named_t arcstat_l2_rw_clash;
717 kstat_named_t arcstat_l2_read_bytes;
718 kstat_named_t arcstat_l2_write_bytes;
719 kstat_named_t arcstat_l2_writes_sent;
720 kstat_named_t arcstat_l2_writes_done;
721 kstat_named_t arcstat_l2_writes_error;
722 kstat_named_t arcstat_l2_writes_lock_retry;
723 kstat_named_t arcstat_l2_evict_lock_retry;
724 kstat_named_t arcstat_l2_evict_reading;
725 kstat_named_t arcstat_l2_evict_l1cached;
726 kstat_named_t arcstat_l2_free_on_write;
727 kstat_named_t arcstat_l2_abort_lowmem;
728 kstat_named_t arcstat_l2_cksum_bad;
729 kstat_named_t arcstat_l2_io_error;
730 kstat_named_t arcstat_l2_lsize;
731 kstat_named_t arcstat_l2_psize;
732 kstat_named_t arcstat_l2_hdr_size;
733 kstat_named_t arcstat_l2_write_trylock_fail;
734 kstat_named_t arcstat_l2_write_passed_headroom;
735 kstat_named_t arcstat_l2_write_spa_mismatch;
736 kstat_named_t arcstat_l2_write_in_l2;
737 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
738 kstat_named_t arcstat_l2_write_not_cacheable;
739 kstat_named_t arcstat_l2_write_full;
740 kstat_named_t arcstat_l2_write_buffer_iter;
741 kstat_named_t arcstat_l2_write_pios;
742 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
743 kstat_named_t arcstat_l2_write_buffer_list_iter;
744 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
745 kstat_named_t arcstat_memory_throttle_count;
746 kstat_named_t arcstat_meta_used;
747 kstat_named_t arcstat_meta_limit;
748 kstat_named_t arcstat_meta_max;
749 kstat_named_t arcstat_meta_min;
750 kstat_named_t arcstat_sync_wait_for_async;
751 kstat_named_t arcstat_demand_hit_predictive_prefetch;
754 static arc_stats_t arc_stats = {
755 { "hits", KSTAT_DATA_UINT64 },
756 { "misses", KSTAT_DATA_UINT64 },
757 { "demand_data_hits", KSTAT_DATA_UINT64 },
758 { "demand_data_misses", KSTAT_DATA_UINT64 },
759 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
760 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
761 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
762 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
763 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
764 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
765 { "mru_hits", KSTAT_DATA_UINT64 },
766 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
767 { "mfu_hits", KSTAT_DATA_UINT64 },
768 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
769 { "allocated", KSTAT_DATA_UINT64 },
770 { "deleted", KSTAT_DATA_UINT64 },
771 { "mutex_miss", KSTAT_DATA_UINT64 },
772 { "evict_skip", KSTAT_DATA_UINT64 },
773 { "evict_not_enough", KSTAT_DATA_UINT64 },
774 { "evict_l2_cached", KSTAT_DATA_UINT64 },
775 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
776 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
777 { "evict_l2_skip", KSTAT_DATA_UINT64 },
778 { "hash_elements", KSTAT_DATA_UINT64 },
779 { "hash_elements_max", KSTAT_DATA_UINT64 },
780 { "hash_collisions", KSTAT_DATA_UINT64 },
781 { "hash_chains", KSTAT_DATA_UINT64 },
782 { "hash_chain_max", KSTAT_DATA_UINT64 },
783 { "p", KSTAT_DATA_UINT64 },
784 { "c", KSTAT_DATA_UINT64 },
785 { "c_min", KSTAT_DATA_UINT64 },
786 { "c_max", KSTAT_DATA_UINT64 },
787 { "size", KSTAT_DATA_UINT64 },
788 { "compressed_size", KSTAT_DATA_UINT64 },
789 { "uncompressed_size", KSTAT_DATA_UINT64 },
790 { "overhead_size", KSTAT_DATA_UINT64 },
791 { "hdr_size", KSTAT_DATA_UINT64 },
792 { "data_size", KSTAT_DATA_UINT64 },
793 { "metadata_size", KSTAT_DATA_UINT64 },
794 { "other_size", KSTAT_DATA_UINT64 },
795 { "anon_size", KSTAT_DATA_UINT64 },
796 { "anon_evictable_data", KSTAT_DATA_UINT64 },
797 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
798 { "mru_size", KSTAT_DATA_UINT64 },
799 { "mru_evictable_data", KSTAT_DATA_UINT64 },
800 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
801 { "mru_ghost_size", KSTAT_DATA_UINT64 },
802 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
803 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
804 { "mfu_size", KSTAT_DATA_UINT64 },
805 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
806 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
807 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
808 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
809 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
810 { "l2_hits", KSTAT_DATA_UINT64 },
811 { "l2_misses", KSTAT_DATA_UINT64 },
812 { "l2_feeds", KSTAT_DATA_UINT64 },
813 { "l2_rw_clash", KSTAT_DATA_UINT64 },
814 { "l2_read_bytes", KSTAT_DATA_UINT64 },
815 { "l2_write_bytes", KSTAT_DATA_UINT64 },
816 { "l2_writes_sent", KSTAT_DATA_UINT64 },
817 { "l2_writes_done", KSTAT_DATA_UINT64 },
818 { "l2_writes_error", KSTAT_DATA_UINT64 },
819 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
820 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
821 { "l2_evict_reading", KSTAT_DATA_UINT64 },
822 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
823 { "l2_free_on_write", KSTAT_DATA_UINT64 },
824 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
825 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
826 { "l2_io_error", KSTAT_DATA_UINT64 },
827 { "l2_size", KSTAT_DATA_UINT64 },
828 { "l2_asize", KSTAT_DATA_UINT64 },
829 { "l2_hdr_size", KSTAT_DATA_UINT64 },
830 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
831 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
832 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
833 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
834 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
835 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
836 { "l2_write_full", KSTAT_DATA_UINT64 },
837 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
838 { "l2_write_pios", KSTAT_DATA_UINT64 },
839 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
840 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
841 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
842 { "memory_throttle_count", KSTAT_DATA_UINT64 },
843 { "arc_meta_used", KSTAT_DATA_UINT64 },
844 { "arc_meta_limit", KSTAT_DATA_UINT64 },
845 { "arc_meta_max", KSTAT_DATA_UINT64 },
846 { "arc_meta_min", KSTAT_DATA_UINT64 },
847 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
848 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
851 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
853 #define ARCSTAT_INCR(stat, val) \
854 atomic_add_64(&arc_stats.stat.value.ui64, (val))
856 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
857 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
859 #define ARCSTAT_MAX(stat, val) { \
861 while ((val) > (m = arc_stats.stat.value.ui64) && \
862 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
866 #define ARCSTAT_MAXSTAT(stat) \
867 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
870 * We define a macro to allow ARC hits/misses to be easily broken down by
871 * two separate conditions, giving a total of four different subtypes for
872 * each of hits and misses (so eight statistics total).
874 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
877 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
879 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
883 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
885 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
890 static arc_state_t *arc_anon;
891 static arc_state_t *arc_mru;
892 static arc_state_t *arc_mru_ghost;
893 static arc_state_t *arc_mfu;
894 static arc_state_t *arc_mfu_ghost;
895 static arc_state_t *arc_l2c_only;
898 * There are several ARC variables that are critical to export as kstats --
899 * but we don't want to have to grovel around in the kstat whenever we wish to
900 * manipulate them. For these variables, we therefore define them to be in
901 * terms of the statistic variable. This assures that we are not introducing
902 * the possibility of inconsistency by having shadow copies of the variables,
903 * while still allowing the code to be readable.
905 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
906 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
907 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
908 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
909 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
910 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
911 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
912 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
913 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
915 /* compressed size of entire arc */
916 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
917 /* uncompressed size of entire arc */
918 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
919 /* number of bytes in the arc from arc_buf_t's */
920 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
922 static int arc_no_grow; /* Don't try to grow cache size */
923 static uint64_t arc_tempreserve;
924 static uint64_t arc_loaned_bytes;
926 typedef struct arc_callback arc_callback_t;
928 struct arc_callback {
930 arc_done_func_t *acb_done;
932 boolean_t acb_compressed;
933 zio_t *acb_zio_dummy;
934 arc_callback_t *acb_next;
937 typedef struct arc_write_callback arc_write_callback_t;
939 struct arc_write_callback {
941 arc_done_func_t *awcb_ready;
942 arc_done_func_t *awcb_children_ready;
943 arc_done_func_t *awcb_physdone;
944 arc_done_func_t *awcb_done;
949 * ARC buffers are separated into multiple structs as a memory saving measure:
950 * - Common fields struct, always defined, and embedded within it:
951 * - L2-only fields, always allocated but undefined when not in L2ARC
952 * - L1-only fields, only allocated when in L1ARC
954 * Buffer in L1 Buffer only in L2
955 * +------------------------+ +------------------------+
956 * | arc_buf_hdr_t | | arc_buf_hdr_t |
960 * +------------------------+ +------------------------+
961 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
962 * | (undefined if L1-only) | | |
963 * +------------------------+ +------------------------+
964 * | l1arc_buf_hdr_t |
969 * +------------------------+
971 * Because it's possible for the L2ARC to become extremely large, we can wind
972 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
973 * is minimized by only allocating the fields necessary for an L1-cached buffer
974 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
975 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
976 * words in pointers. arc_hdr_realloc() is used to switch a header between
977 * these two allocation states.
979 typedef struct l1arc_buf_hdr {
980 kmutex_t b_freeze_lock;
981 zio_cksum_t *b_freeze_cksum;
984 * Used for debugging with kmem_flags - by allocating and freeing
985 * b_thawed when the buffer is thawed, we get a record of the stack
986 * trace that thawed it.
993 /* for waiting on writes to complete */
997 /* protected by arc state mutex */
998 arc_state_t *b_state;
999 multilist_node_t b_arc_node;
1001 /* updated atomically */
1002 clock_t b_arc_access;
1004 /* self protecting */
1005 refcount_t b_refcnt;
1007 arc_callback_t *b_acb;
1011 typedef struct l2arc_dev l2arc_dev_t;
1013 typedef struct l2arc_buf_hdr {
1014 /* protected by arc_buf_hdr mutex */
1015 l2arc_dev_t *b_dev; /* L2ARC device */
1016 uint64_t b_daddr; /* disk address, offset byte */
1018 list_node_t b_l2node;
1021 struct arc_buf_hdr {
1022 /* protected by hash lock */
1026 arc_buf_contents_t b_type;
1027 arc_buf_hdr_t *b_hash_next;
1028 arc_flags_t b_flags;
1031 * This field stores the size of the data buffer after
1032 * compression, and is set in the arc's zio completion handlers.
1033 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1035 * While the block pointers can store up to 32MB in their psize
1036 * field, we can only store up to 32MB minus 512B. This is due
1037 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1038 * a field of zeros represents 512B in the bp). We can't use a
1039 * bias of 1 since we need to reserve a psize of zero, here, to
1040 * represent holes and embedded blocks.
1042 * This isn't a problem in practice, since the maximum size of a
1043 * buffer is limited to 16MB, so we never need to store 32MB in
1044 * this field. Even in the upstream illumos code base, the
1045 * maximum size of a buffer is limited to 16MB.
1050 * This field stores the size of the data buffer before
1051 * compression, and cannot change once set. It is in units
1052 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1054 uint16_t b_lsize; /* immutable */
1055 uint64_t b_spa; /* immutable */
1057 /* L2ARC fields. Undefined when not in L2ARC. */
1058 l2arc_buf_hdr_t b_l2hdr;
1059 /* L1ARC fields. Undefined when in l2arc_only state */
1060 l1arc_buf_hdr_t b_l1hdr;
1063 #if defined(__FreeBSD__) && defined(_KERNEL)
1065 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1070 val = arc_meta_limit;
1071 err = sysctl_handle_64(oidp, &val, 0, req);
1072 if (err != 0 || req->newptr == NULL)
1075 if (val <= 0 || val > arc_c_max)
1078 arc_meta_limit = val;
1083 sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS)
1088 val = arc_no_grow_shift;
1089 err = sysctl_handle_32(oidp, &val, 0, req);
1090 if (err != 0 || req->newptr == NULL)
1093 if (val >= arc_shrink_shift)
1096 arc_no_grow_shift = val;
1101 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1107 err = sysctl_handle_64(oidp, &val, 0, req);
1108 if (err != 0 || req->newptr == NULL)
1111 if (zfs_arc_max == 0) {
1112 /* Loader tunable so blindly set */
1117 if (val < arc_abs_min || val > kmem_size())
1119 if (val < arc_c_min)
1121 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1127 arc_p = (arc_c >> 1);
1129 if (zfs_arc_meta_limit == 0) {
1130 /* limit meta-data to 1/4 of the arc capacity */
1131 arc_meta_limit = arc_c_max / 4;
1134 /* if kmem_flags are set, lets try to use less memory */
1135 if (kmem_debugging())
1138 zfs_arc_max = arc_c;
1144 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1150 err = sysctl_handle_64(oidp, &val, 0, req);
1151 if (err != 0 || req->newptr == NULL)
1154 if (zfs_arc_min == 0) {
1155 /* Loader tunable so blindly set */
1160 if (val < arc_abs_min || val > arc_c_max)
1165 if (zfs_arc_meta_min == 0)
1166 arc_meta_min = arc_c_min / 2;
1168 if (arc_c < arc_c_min)
1171 zfs_arc_min = arc_c_min;
1177 #define GHOST_STATE(state) \
1178 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1179 (state) == arc_l2c_only)
1181 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1182 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1183 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1184 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1185 #define HDR_COMPRESSION_ENABLED(hdr) \
1186 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1188 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1189 #define HDR_L2_READING(hdr) \
1190 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1191 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1192 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1193 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1194 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1195 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1197 #define HDR_ISTYPE_METADATA(hdr) \
1198 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1199 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1201 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1202 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1204 /* For storing compression mode in b_flags */
1205 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1207 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1208 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1209 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1210 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1212 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1213 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1214 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1220 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1221 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1224 * Hash table routines
1227 #define HT_LOCK_PAD CACHE_LINE_SIZE
1232 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1236 #define BUF_LOCKS 256
1237 typedef struct buf_hash_table {
1239 arc_buf_hdr_t **ht_table;
1240 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1243 static buf_hash_table_t buf_hash_table;
1245 #define BUF_HASH_INDEX(spa, dva, birth) \
1246 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1247 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1248 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1249 #define HDR_LOCK(hdr) \
1250 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1252 uint64_t zfs_crc64_table[256];
1258 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1259 #define L2ARC_HEADROOM 2 /* num of writes */
1261 * If we discover during ARC scan any buffers to be compressed, we boost
1262 * our headroom for the next scanning cycle by this percentage multiple.
1264 #define L2ARC_HEADROOM_BOOST 200
1265 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1266 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1268 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1269 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1271 /* L2ARC Performance Tunables */
1272 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1273 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1274 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1275 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1276 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1277 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1278 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1279 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1280 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1282 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1283 &l2arc_write_max, 0, "max write size");
1284 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1285 &l2arc_write_boost, 0, "extra write during warmup");
1286 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1287 &l2arc_headroom, 0, "number of dev writes");
1288 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1289 &l2arc_feed_secs, 0, "interval seconds");
1290 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1291 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1293 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1294 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1295 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1296 &l2arc_feed_again, 0, "turbo warmup");
1297 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1298 &l2arc_norw, 0, "no reads during writes");
1300 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1301 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1302 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1303 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1304 "size of anonymous state");
1305 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1306 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1307 "size of anonymous state");
1309 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1310 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1311 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1312 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1313 "size of metadata in mru state");
1314 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1315 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1316 "size of data in mru state");
1318 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1319 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1320 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1321 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1322 "size of metadata in mru ghost state");
1323 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1324 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1325 "size of data in mru ghost state");
1327 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1328 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1329 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1330 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1331 "size of metadata in mfu state");
1332 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1333 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1334 "size of data in mfu state");
1336 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1337 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1338 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1339 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1340 "size of metadata in mfu ghost state");
1341 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1342 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1343 "size of data in mfu ghost state");
1345 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1346 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1352 vdev_t *l2ad_vdev; /* vdev */
1353 spa_t *l2ad_spa; /* spa */
1354 uint64_t l2ad_hand; /* next write location */
1355 uint64_t l2ad_start; /* first addr on device */
1356 uint64_t l2ad_end; /* last addr on device */
1357 boolean_t l2ad_first; /* first sweep through */
1358 boolean_t l2ad_writing; /* currently writing */
1359 kmutex_t l2ad_mtx; /* lock for buffer list */
1360 list_t l2ad_buflist; /* buffer list */
1361 list_node_t l2ad_node; /* device list node */
1362 refcount_t l2ad_alloc; /* allocated bytes */
1365 static list_t L2ARC_dev_list; /* device list */
1366 static list_t *l2arc_dev_list; /* device list pointer */
1367 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1368 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1369 static list_t L2ARC_free_on_write; /* free after write buf list */
1370 static list_t *l2arc_free_on_write; /* free after write list ptr */
1371 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1372 static uint64_t l2arc_ndev; /* number of devices */
1374 typedef struct l2arc_read_callback {
1375 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1376 blkptr_t l2rcb_bp; /* original blkptr */
1377 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1378 int l2rcb_flags; /* original flags */
1379 abd_t *l2rcb_abd; /* temporary buffer */
1380 } l2arc_read_callback_t;
1382 typedef struct l2arc_write_callback {
1383 l2arc_dev_t *l2wcb_dev; /* device info */
1384 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1385 } l2arc_write_callback_t;
1387 typedef struct l2arc_data_free {
1388 /* protected by l2arc_free_on_write_mtx */
1391 arc_buf_contents_t l2df_type;
1392 list_node_t l2df_list_node;
1393 } l2arc_data_free_t;
1395 static kmutex_t l2arc_feed_thr_lock;
1396 static kcondvar_t l2arc_feed_thr_cv;
1397 static uint8_t l2arc_thread_exit;
1399 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1400 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1401 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1402 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1403 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1404 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1405 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1406 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1407 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1408 static boolean_t arc_is_overflowing();
1409 static void arc_buf_watch(arc_buf_t *);
1411 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1412 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1413 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1414 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1416 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1417 static void l2arc_read_done(zio_t *);
1420 l2arc_trim(const arc_buf_hdr_t *hdr)
1422 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1424 ASSERT(HDR_HAS_L2HDR(hdr));
1425 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1427 if (HDR_GET_PSIZE(hdr) != 0) {
1428 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1429 HDR_GET_PSIZE(hdr), 0);
1434 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1436 uint8_t *vdva = (uint8_t *)dva;
1437 uint64_t crc = -1ULL;
1440 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1442 for (i = 0; i < sizeof (dva_t); i++)
1443 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1445 crc ^= (spa>>8) ^ birth;
1450 #define HDR_EMPTY(hdr) \
1451 ((hdr)->b_dva.dva_word[0] == 0 && \
1452 (hdr)->b_dva.dva_word[1] == 0)
1454 #define HDR_EQUAL(spa, dva, birth, hdr) \
1455 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1456 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1457 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1460 buf_discard_identity(arc_buf_hdr_t *hdr)
1462 hdr->b_dva.dva_word[0] = 0;
1463 hdr->b_dva.dva_word[1] = 0;
1467 static arc_buf_hdr_t *
1468 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1470 const dva_t *dva = BP_IDENTITY(bp);
1471 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1472 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1473 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1476 mutex_enter(hash_lock);
1477 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1478 hdr = hdr->b_hash_next) {
1479 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1484 mutex_exit(hash_lock);
1490 * Insert an entry into the hash table. If there is already an element
1491 * equal to elem in the hash table, then the already existing element
1492 * will be returned and the new element will not be inserted.
1493 * Otherwise returns NULL.
1494 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1496 static arc_buf_hdr_t *
1497 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1499 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1500 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1501 arc_buf_hdr_t *fhdr;
1504 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1505 ASSERT(hdr->b_birth != 0);
1506 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1508 if (lockp != NULL) {
1510 mutex_enter(hash_lock);
1512 ASSERT(MUTEX_HELD(hash_lock));
1515 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1516 fhdr = fhdr->b_hash_next, i++) {
1517 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1521 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1522 buf_hash_table.ht_table[idx] = hdr;
1523 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1525 /* collect some hash table performance data */
1527 ARCSTAT_BUMP(arcstat_hash_collisions);
1529 ARCSTAT_BUMP(arcstat_hash_chains);
1531 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1534 ARCSTAT_BUMP(arcstat_hash_elements);
1535 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1541 buf_hash_remove(arc_buf_hdr_t *hdr)
1543 arc_buf_hdr_t *fhdr, **hdrp;
1544 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1546 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1547 ASSERT(HDR_IN_HASH_TABLE(hdr));
1549 hdrp = &buf_hash_table.ht_table[idx];
1550 while ((fhdr = *hdrp) != hdr) {
1551 ASSERT3P(fhdr, !=, NULL);
1552 hdrp = &fhdr->b_hash_next;
1554 *hdrp = hdr->b_hash_next;
1555 hdr->b_hash_next = NULL;
1556 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1558 /* collect some hash table performance data */
1559 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1561 if (buf_hash_table.ht_table[idx] &&
1562 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1563 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1567 * Global data structures and functions for the buf kmem cache.
1569 static kmem_cache_t *hdr_full_cache;
1570 static kmem_cache_t *hdr_l2only_cache;
1571 static kmem_cache_t *buf_cache;
1578 kmem_free(buf_hash_table.ht_table,
1579 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1580 for (i = 0; i < BUF_LOCKS; i++)
1581 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1582 kmem_cache_destroy(hdr_full_cache);
1583 kmem_cache_destroy(hdr_l2only_cache);
1584 kmem_cache_destroy(buf_cache);
1588 * Constructor callback - called when the cache is empty
1589 * and a new buf is requested.
1593 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1595 arc_buf_hdr_t *hdr = vbuf;
1597 bzero(hdr, HDR_FULL_SIZE);
1598 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1599 refcount_create(&hdr->b_l1hdr.b_refcnt);
1600 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1601 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1602 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1609 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1611 arc_buf_hdr_t *hdr = vbuf;
1613 bzero(hdr, HDR_L2ONLY_SIZE);
1614 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1621 buf_cons(void *vbuf, void *unused, int kmflag)
1623 arc_buf_t *buf = vbuf;
1625 bzero(buf, sizeof (arc_buf_t));
1626 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1627 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1633 * Destructor callback - called when a cached buf is
1634 * no longer required.
1638 hdr_full_dest(void *vbuf, void *unused)
1640 arc_buf_hdr_t *hdr = vbuf;
1642 ASSERT(HDR_EMPTY(hdr));
1643 cv_destroy(&hdr->b_l1hdr.b_cv);
1644 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1645 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1646 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1647 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1652 hdr_l2only_dest(void *vbuf, void *unused)
1654 arc_buf_hdr_t *hdr = vbuf;
1656 ASSERT(HDR_EMPTY(hdr));
1657 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1662 buf_dest(void *vbuf, void *unused)
1664 arc_buf_t *buf = vbuf;
1666 mutex_destroy(&buf->b_evict_lock);
1667 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1671 * Reclaim callback -- invoked when memory is low.
1675 hdr_recl(void *unused)
1677 dprintf("hdr_recl called\n");
1679 * umem calls the reclaim func when we destroy the buf cache,
1680 * which is after we do arc_fini().
1683 cv_signal(&arc_reclaim_thread_cv);
1690 uint64_t hsize = 1ULL << 12;
1694 * The hash table is big enough to fill all of physical memory
1695 * with an average block size of zfs_arc_average_blocksize (default 8K).
1696 * By default, the table will take up
1697 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1699 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1702 buf_hash_table.ht_mask = hsize - 1;
1703 buf_hash_table.ht_table =
1704 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1705 if (buf_hash_table.ht_table == NULL) {
1706 ASSERT(hsize > (1ULL << 8));
1711 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1712 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1713 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1714 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1716 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1717 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1719 for (i = 0; i < 256; i++)
1720 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1721 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1723 for (i = 0; i < BUF_LOCKS; i++) {
1724 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1725 NULL, MUTEX_DEFAULT, NULL);
1730 * This is the size that the buf occupies in memory. If the buf is compressed,
1731 * it will correspond to the compressed size. You should use this method of
1732 * getting the buf size unless you explicitly need the logical size.
1735 arc_buf_size(arc_buf_t *buf)
1737 return (ARC_BUF_COMPRESSED(buf) ?
1738 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1742 arc_buf_lsize(arc_buf_t *buf)
1744 return (HDR_GET_LSIZE(buf->b_hdr));
1748 arc_get_compression(arc_buf_t *buf)
1750 return (ARC_BUF_COMPRESSED(buf) ?
1751 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1754 #define ARC_MINTIME (hz>>4) /* 62 ms */
1756 static inline boolean_t
1757 arc_buf_is_shared(arc_buf_t *buf)
1759 boolean_t shared = (buf->b_data != NULL &&
1760 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1761 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1762 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1763 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1764 IMPLY(shared, ARC_BUF_SHARED(buf));
1765 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1768 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1769 * already being shared" requirement prevents us from doing that.
1776 * Free the checksum associated with this header. If there is no checksum, this
1780 arc_cksum_free(arc_buf_hdr_t *hdr)
1782 ASSERT(HDR_HAS_L1HDR(hdr));
1783 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1784 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1785 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1786 hdr->b_l1hdr.b_freeze_cksum = NULL;
1788 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1792 * Return true iff at least one of the bufs on hdr is not compressed.
1795 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1797 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1798 if (!ARC_BUF_COMPRESSED(b)) {
1806 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1807 * matches the checksum that is stored in the hdr. If there is no checksum,
1808 * or if the buf is compressed, this is a no-op.
1811 arc_cksum_verify(arc_buf_t *buf)
1813 arc_buf_hdr_t *hdr = buf->b_hdr;
1816 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1819 if (ARC_BUF_COMPRESSED(buf)) {
1820 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1821 arc_hdr_has_uncompressed_buf(hdr));
1825 ASSERT(HDR_HAS_L1HDR(hdr));
1827 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1828 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1829 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1833 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1834 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1835 panic("buffer modified while frozen!");
1836 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1840 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1842 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1843 boolean_t valid_cksum;
1845 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1846 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1849 * We rely on the blkptr's checksum to determine if the block
1850 * is valid or not. When compressed arc is enabled, the l2arc
1851 * writes the block to the l2arc just as it appears in the pool.
1852 * This allows us to use the blkptr's checksum to validate the
1853 * data that we just read off of the l2arc without having to store
1854 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1855 * arc is disabled, then the data written to the l2arc is always
1856 * uncompressed and won't match the block as it exists in the main
1857 * pool. When this is the case, we must first compress it if it is
1858 * compressed on the main pool before we can validate the checksum.
1860 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1861 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1862 uint64_t lsize = HDR_GET_LSIZE(hdr);
1865 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1866 csize = zio_compress_data(compress, zio->io_abd,
1867 abd_to_buf(cdata), lsize);
1869 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1870 if (csize < HDR_GET_PSIZE(hdr)) {
1872 * Compressed blocks are always a multiple of the
1873 * smallest ashift in the pool. Ideally, we would
1874 * like to round up the csize to the next
1875 * spa_min_ashift but that value may have changed
1876 * since the block was last written. Instead,
1877 * we rely on the fact that the hdr's psize
1878 * was set to the psize of the block when it was
1879 * last written. We set the csize to that value
1880 * and zero out any part that should not contain
1883 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1884 csize = HDR_GET_PSIZE(hdr);
1886 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1890 * Block pointers always store the checksum for the logical data.
1891 * If the block pointer has the gang bit set, then the checksum
1892 * it represents is for the reconstituted data and not for an
1893 * individual gang member. The zio pipeline, however, must be able to
1894 * determine the checksum of each of the gang constituents so it
1895 * treats the checksum comparison differently than what we need
1896 * for l2arc blocks. This prevents us from using the
1897 * zio_checksum_error() interface directly. Instead we must call the
1898 * zio_checksum_error_impl() so that we can ensure the checksum is
1899 * generated using the correct checksum algorithm and accounts for the
1900 * logical I/O size and not just a gang fragment.
1902 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1903 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1904 zio->io_offset, NULL) == 0);
1905 zio_pop_transforms(zio);
1906 return (valid_cksum);
1910 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1911 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1912 * isn't modified later on. If buf is compressed or there is already a checksum
1913 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1916 arc_cksum_compute(arc_buf_t *buf)
1918 arc_buf_hdr_t *hdr = buf->b_hdr;
1920 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1923 ASSERT(HDR_HAS_L1HDR(hdr));
1925 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1926 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1927 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1928 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1930 } else if (ARC_BUF_COMPRESSED(buf)) {
1931 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1935 ASSERT(!ARC_BUF_COMPRESSED(buf));
1936 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1938 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1939 hdr->b_l1hdr.b_freeze_cksum);
1940 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1948 typedef struct procctl {
1956 arc_buf_unwatch(arc_buf_t *buf)
1963 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1964 ctl.prwatch.pr_size = 0;
1965 ctl.prwatch.pr_wflags = 0;
1966 result = write(arc_procfd, &ctl, sizeof (ctl));
1967 ASSERT3U(result, ==, sizeof (ctl));
1974 arc_buf_watch(arc_buf_t *buf)
1981 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1982 ctl.prwatch.pr_size = arc_buf_size(buf);
1983 ctl.prwatch.pr_wflags = WA_WRITE;
1984 result = write(arc_procfd, &ctl, sizeof (ctl));
1985 ASSERT3U(result, ==, sizeof (ctl));
1989 #endif /* illumos */
1991 static arc_buf_contents_t
1992 arc_buf_type(arc_buf_hdr_t *hdr)
1994 arc_buf_contents_t type;
1995 if (HDR_ISTYPE_METADATA(hdr)) {
1996 type = ARC_BUFC_METADATA;
1998 type = ARC_BUFC_DATA;
2000 VERIFY3U(hdr->b_type, ==, type);
2005 arc_is_metadata(arc_buf_t *buf)
2007 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2011 arc_bufc_to_flags(arc_buf_contents_t type)
2015 /* metadata field is 0 if buffer contains normal data */
2017 case ARC_BUFC_METADATA:
2018 return (ARC_FLAG_BUFC_METADATA);
2022 panic("undefined ARC buffer type!");
2023 return ((uint32_t)-1);
2027 arc_buf_thaw(arc_buf_t *buf)
2029 arc_buf_hdr_t *hdr = buf->b_hdr;
2031 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2032 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2034 arc_cksum_verify(buf);
2037 * Compressed buffers do not manipulate the b_freeze_cksum or
2038 * allocate b_thawed.
2040 if (ARC_BUF_COMPRESSED(buf)) {
2041 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2042 arc_hdr_has_uncompressed_buf(hdr));
2046 ASSERT(HDR_HAS_L1HDR(hdr));
2047 arc_cksum_free(hdr);
2049 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2051 if (zfs_flags & ZFS_DEBUG_MODIFY) {
2052 if (hdr->b_l1hdr.b_thawed != NULL)
2053 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2054 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2058 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2061 arc_buf_unwatch(buf);
2066 arc_buf_freeze(arc_buf_t *buf)
2068 arc_buf_hdr_t *hdr = buf->b_hdr;
2069 kmutex_t *hash_lock;
2071 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2074 if (ARC_BUF_COMPRESSED(buf)) {
2075 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2076 arc_hdr_has_uncompressed_buf(hdr));
2080 hash_lock = HDR_LOCK(hdr);
2081 mutex_enter(hash_lock);
2083 ASSERT(HDR_HAS_L1HDR(hdr));
2084 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
2085 hdr->b_l1hdr.b_state == arc_anon);
2086 arc_cksum_compute(buf);
2087 mutex_exit(hash_lock);
2091 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2092 * the following functions should be used to ensure that the flags are
2093 * updated in a thread-safe way. When manipulating the flags either
2094 * the hash_lock must be held or the hdr must be undiscoverable. This
2095 * ensures that we're not racing with any other threads when updating
2099 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2101 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2102 hdr->b_flags |= flags;
2106 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2108 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2109 hdr->b_flags &= ~flags;
2113 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2114 * done in a special way since we have to clear and set bits
2115 * at the same time. Consumers that wish to set the compression bits
2116 * must use this function to ensure that the flags are updated in
2117 * thread-safe manner.
2120 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2122 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2125 * Holes and embedded blocks will always have a psize = 0 so
2126 * we ignore the compression of the blkptr and set the
2127 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2128 * Holes and embedded blocks remain anonymous so we don't
2129 * want to uncompress them. Mark them as uncompressed.
2131 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2132 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2133 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2134 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2135 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2137 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2138 HDR_SET_COMPRESS(hdr, cmp);
2139 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2140 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2145 * Looks for another buf on the same hdr which has the data decompressed, copies
2146 * from it, and returns true. If no such buf exists, returns false.
2149 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
2151 arc_buf_hdr_t *hdr = buf->b_hdr;
2152 boolean_t copied = B_FALSE;
2154 ASSERT(HDR_HAS_L1HDR(hdr));
2155 ASSERT3P(buf->b_data, !=, NULL);
2156 ASSERT(!ARC_BUF_COMPRESSED(buf));
2158 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2159 from = from->b_next) {
2160 /* can't use our own data buffer */
2165 if (!ARC_BUF_COMPRESSED(from)) {
2166 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2173 * There were no decompressed bufs, so there should not be a
2174 * checksum on the hdr either.
2176 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
2182 * Given a buf that has a data buffer attached to it, this function will
2183 * efficiently fill the buf with data of the specified compression setting from
2184 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2185 * are already sharing a data buf, no copy is performed.
2187 * If the buf is marked as compressed but uncompressed data was requested, this
2188 * will allocate a new data buffer for the buf, remove that flag, and fill the
2189 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2190 * uncompressed data, and (since we haven't added support for it yet) if you
2191 * want compressed data your buf must already be marked as compressed and have
2192 * the correct-sized data buffer.
2195 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2197 arc_buf_hdr_t *hdr = buf->b_hdr;
2198 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2199 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2201 ASSERT3P(buf->b_data, !=, NULL);
2202 IMPLY(compressed, hdr_compressed);
2203 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2205 if (hdr_compressed == compressed) {
2206 if (!arc_buf_is_shared(buf)) {
2207 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2211 ASSERT(hdr_compressed);
2212 ASSERT(!compressed);
2213 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2216 * If the buf is sharing its data with the hdr, unlink it and
2217 * allocate a new data buffer for the buf.
2219 if (arc_buf_is_shared(buf)) {
2220 ASSERT(ARC_BUF_COMPRESSED(buf));
2222 /* We need to give the buf it's own b_data */
2223 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2225 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2226 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2228 /* Previously overhead was 0; just add new overhead */
2229 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2230 } else if (ARC_BUF_COMPRESSED(buf)) {
2231 /* We need to reallocate the buf's b_data */
2232 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2235 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2237 /* We increased the size of b_data; update overhead */
2238 ARCSTAT_INCR(arcstat_overhead_size,
2239 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2243 * Regardless of the buf's previous compression settings, it
2244 * should not be compressed at the end of this function.
2246 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2249 * Try copying the data from another buf which already has a
2250 * decompressed version. If that's not possible, it's time to
2251 * bite the bullet and decompress the data from the hdr.
2253 if (arc_buf_try_copy_decompressed_data(buf)) {
2254 /* Skip byteswapping and checksumming (already done) */
2255 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2258 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2259 hdr->b_l1hdr.b_pabd, buf->b_data,
2260 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2263 * Absent hardware errors or software bugs, this should
2264 * be impossible, but log it anyway so we can debug it.
2268 "hdr %p, compress %d, psize %d, lsize %d",
2269 hdr, HDR_GET_COMPRESS(hdr),
2270 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2271 return (SET_ERROR(EIO));
2276 /* Byteswap the buf's data if necessary */
2277 if (bswap != DMU_BSWAP_NUMFUNCS) {
2278 ASSERT(!HDR_SHARED_DATA(hdr));
2279 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2280 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2283 /* Compute the hdr's checksum if necessary */
2284 arc_cksum_compute(buf);
2290 arc_decompress(arc_buf_t *buf)
2292 return (arc_buf_fill(buf, B_FALSE));
2296 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2299 arc_hdr_size(arc_buf_hdr_t *hdr)
2303 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2304 HDR_GET_PSIZE(hdr) > 0) {
2305 size = HDR_GET_PSIZE(hdr);
2307 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2308 size = HDR_GET_LSIZE(hdr);
2314 * Increment the amount of evictable space in the arc_state_t's refcount.
2315 * We account for the space used by the hdr and the arc buf individually
2316 * so that we can add and remove them from the refcount individually.
2319 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2321 arc_buf_contents_t type = arc_buf_type(hdr);
2323 ASSERT(HDR_HAS_L1HDR(hdr));
2325 if (GHOST_STATE(state)) {
2326 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2327 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2328 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2329 (void) refcount_add_many(&state->arcs_esize[type],
2330 HDR_GET_LSIZE(hdr), hdr);
2334 ASSERT(!GHOST_STATE(state));
2335 if (hdr->b_l1hdr.b_pabd != NULL) {
2336 (void) refcount_add_many(&state->arcs_esize[type],
2337 arc_hdr_size(hdr), hdr);
2339 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2340 buf = buf->b_next) {
2341 if (arc_buf_is_shared(buf))
2343 (void) refcount_add_many(&state->arcs_esize[type],
2344 arc_buf_size(buf), buf);
2349 * Decrement the amount of evictable space in the arc_state_t's refcount.
2350 * We account for the space used by the hdr and the arc buf individually
2351 * so that we can add and remove them from the refcount individually.
2354 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2356 arc_buf_contents_t type = arc_buf_type(hdr);
2358 ASSERT(HDR_HAS_L1HDR(hdr));
2360 if (GHOST_STATE(state)) {
2361 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2362 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2363 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2364 (void) refcount_remove_many(&state->arcs_esize[type],
2365 HDR_GET_LSIZE(hdr), hdr);
2369 ASSERT(!GHOST_STATE(state));
2370 if (hdr->b_l1hdr.b_pabd != NULL) {
2371 (void) refcount_remove_many(&state->arcs_esize[type],
2372 arc_hdr_size(hdr), hdr);
2374 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2375 buf = buf->b_next) {
2376 if (arc_buf_is_shared(buf))
2378 (void) refcount_remove_many(&state->arcs_esize[type],
2379 arc_buf_size(buf), buf);
2384 * Add a reference to this hdr indicating that someone is actively
2385 * referencing that memory. When the refcount transitions from 0 to 1,
2386 * we remove it from the respective arc_state_t list to indicate that
2387 * it is not evictable.
2390 add_reference(arc_buf_hdr_t *hdr, void *tag)
2392 ASSERT(HDR_HAS_L1HDR(hdr));
2393 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2394 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2395 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2396 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2399 arc_state_t *state = hdr->b_l1hdr.b_state;
2401 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2402 (state != arc_anon)) {
2403 /* We don't use the L2-only state list. */
2404 if (state != arc_l2c_only) {
2405 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2407 arc_evictable_space_decrement(hdr, state);
2409 /* remove the prefetch flag if we get a reference */
2410 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2415 * Remove a reference from this hdr. When the reference transitions from
2416 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2417 * list making it eligible for eviction.
2420 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2423 arc_state_t *state = hdr->b_l1hdr.b_state;
2425 ASSERT(HDR_HAS_L1HDR(hdr));
2426 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2427 ASSERT(!GHOST_STATE(state));
2430 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2431 * check to prevent usage of the arc_l2c_only list.
2433 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2434 (state != arc_anon)) {
2435 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2436 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2437 arc_evictable_space_increment(hdr, state);
2443 * Move the supplied buffer to the indicated state. The hash lock
2444 * for the buffer must be held by the caller.
2447 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2448 kmutex_t *hash_lock)
2450 arc_state_t *old_state;
2453 boolean_t update_old, update_new;
2454 arc_buf_contents_t buftype = arc_buf_type(hdr);
2457 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2458 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2459 * L1 hdr doesn't always exist when we change state to arc_anon before
2460 * destroying a header, in which case reallocating to add the L1 hdr is
2463 if (HDR_HAS_L1HDR(hdr)) {
2464 old_state = hdr->b_l1hdr.b_state;
2465 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2466 bufcnt = hdr->b_l1hdr.b_bufcnt;
2467 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2469 old_state = arc_l2c_only;
2472 update_old = B_FALSE;
2474 update_new = update_old;
2476 ASSERT(MUTEX_HELD(hash_lock));
2477 ASSERT3P(new_state, !=, old_state);
2478 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2479 ASSERT(old_state != arc_anon || bufcnt <= 1);
2482 * If this buffer is evictable, transfer it from the
2483 * old state list to the new state list.
2486 if (old_state != arc_anon && old_state != arc_l2c_only) {
2487 ASSERT(HDR_HAS_L1HDR(hdr));
2488 multilist_remove(old_state->arcs_list[buftype], hdr);
2490 if (GHOST_STATE(old_state)) {
2492 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2493 update_old = B_TRUE;
2495 arc_evictable_space_decrement(hdr, old_state);
2497 if (new_state != arc_anon && new_state != arc_l2c_only) {
2500 * An L1 header always exists here, since if we're
2501 * moving to some L1-cached state (i.e. not l2c_only or
2502 * anonymous), we realloc the header to add an L1hdr
2505 ASSERT(HDR_HAS_L1HDR(hdr));
2506 multilist_insert(new_state->arcs_list[buftype], hdr);
2508 if (GHOST_STATE(new_state)) {
2510 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2511 update_new = B_TRUE;
2513 arc_evictable_space_increment(hdr, new_state);
2517 ASSERT(!HDR_EMPTY(hdr));
2518 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2519 buf_hash_remove(hdr);
2521 /* adjust state sizes (ignore arc_l2c_only) */
2523 if (update_new && new_state != arc_l2c_only) {
2524 ASSERT(HDR_HAS_L1HDR(hdr));
2525 if (GHOST_STATE(new_state)) {
2529 * When moving a header to a ghost state, we first
2530 * remove all arc buffers. Thus, we'll have a
2531 * bufcnt of zero, and no arc buffer to use for
2532 * the reference. As a result, we use the arc
2533 * header pointer for the reference.
2535 (void) refcount_add_many(&new_state->arcs_size,
2536 HDR_GET_LSIZE(hdr), hdr);
2537 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2539 uint32_t buffers = 0;
2542 * Each individual buffer holds a unique reference,
2543 * thus we must remove each of these references one
2546 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2547 buf = buf->b_next) {
2548 ASSERT3U(bufcnt, !=, 0);
2552 * When the arc_buf_t is sharing the data
2553 * block with the hdr, the owner of the
2554 * reference belongs to the hdr. Only
2555 * add to the refcount if the arc_buf_t is
2558 if (arc_buf_is_shared(buf))
2561 (void) refcount_add_many(&new_state->arcs_size,
2562 arc_buf_size(buf), buf);
2564 ASSERT3U(bufcnt, ==, buffers);
2566 if (hdr->b_l1hdr.b_pabd != NULL) {
2567 (void) refcount_add_many(&new_state->arcs_size,
2568 arc_hdr_size(hdr), hdr);
2570 ASSERT(GHOST_STATE(old_state));
2575 if (update_old && old_state != arc_l2c_only) {
2576 ASSERT(HDR_HAS_L1HDR(hdr));
2577 if (GHOST_STATE(old_state)) {
2579 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2582 * When moving a header off of a ghost state,
2583 * the header will not contain any arc buffers.
2584 * We use the arc header pointer for the reference
2585 * which is exactly what we did when we put the
2586 * header on the ghost state.
2589 (void) refcount_remove_many(&old_state->arcs_size,
2590 HDR_GET_LSIZE(hdr), hdr);
2592 uint32_t buffers = 0;
2595 * Each individual buffer holds a unique reference,
2596 * thus we must remove each of these references one
2599 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2600 buf = buf->b_next) {
2601 ASSERT3U(bufcnt, !=, 0);
2605 * When the arc_buf_t is sharing the data
2606 * block with the hdr, the owner of the
2607 * reference belongs to the hdr. Only
2608 * add to the refcount if the arc_buf_t is
2611 if (arc_buf_is_shared(buf))
2614 (void) refcount_remove_many(
2615 &old_state->arcs_size, arc_buf_size(buf),
2618 ASSERT3U(bufcnt, ==, buffers);
2619 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2620 (void) refcount_remove_many(
2621 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2625 if (HDR_HAS_L1HDR(hdr))
2626 hdr->b_l1hdr.b_state = new_state;
2629 * L2 headers should never be on the L2 state list since they don't
2630 * have L1 headers allocated.
2632 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2633 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2637 arc_space_consume(uint64_t space, arc_space_type_t type)
2639 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2642 case ARC_SPACE_DATA:
2643 ARCSTAT_INCR(arcstat_data_size, space);
2645 case ARC_SPACE_META:
2646 ARCSTAT_INCR(arcstat_metadata_size, space);
2648 case ARC_SPACE_OTHER:
2649 ARCSTAT_INCR(arcstat_other_size, space);
2651 case ARC_SPACE_HDRS:
2652 ARCSTAT_INCR(arcstat_hdr_size, space);
2654 case ARC_SPACE_L2HDRS:
2655 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
2659 if (type != ARC_SPACE_DATA)
2660 ARCSTAT_INCR(arcstat_meta_used, space);
2662 atomic_add_64(&arc_size, space);
2666 arc_space_return(uint64_t space, arc_space_type_t type)
2668 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2671 case ARC_SPACE_DATA:
2672 ARCSTAT_INCR(arcstat_data_size, -space);
2674 case ARC_SPACE_META:
2675 ARCSTAT_INCR(arcstat_metadata_size, -space);
2677 case ARC_SPACE_OTHER:
2678 ARCSTAT_INCR(arcstat_other_size, -space);
2680 case ARC_SPACE_HDRS:
2681 ARCSTAT_INCR(arcstat_hdr_size, -space);
2683 case ARC_SPACE_L2HDRS:
2684 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
2688 if (type != ARC_SPACE_DATA) {
2689 ASSERT(arc_meta_used >= space);
2690 if (arc_meta_max < arc_meta_used)
2691 arc_meta_max = arc_meta_used;
2692 ARCSTAT_INCR(arcstat_meta_used, -space);
2695 ASSERT(arc_size >= space);
2696 atomic_add_64(&arc_size, -space);
2700 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2701 * with the hdr's b_pabd.
2704 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2707 * The criteria for sharing a hdr's data are:
2708 * 1. the hdr's compression matches the buf's compression
2709 * 2. the hdr doesn't need to be byteswapped
2710 * 3. the hdr isn't already being shared
2711 * 4. the buf is either compressed or it is the last buf in the hdr list
2713 * Criterion #4 maintains the invariant that shared uncompressed
2714 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2715 * might ask, "if a compressed buf is allocated first, won't that be the
2716 * last thing in the list?", but in that case it's impossible to create
2717 * a shared uncompressed buf anyway (because the hdr must be compressed
2718 * to have the compressed buf). You might also think that #3 is
2719 * sufficient to make this guarantee, however it's possible
2720 * (specifically in the rare L2ARC write race mentioned in
2721 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2722 * is sharable, but wasn't at the time of its allocation. Rather than
2723 * allow a new shared uncompressed buf to be created and then shuffle
2724 * the list around to make it the last element, this simply disallows
2725 * sharing if the new buf isn't the first to be added.
2727 ASSERT3P(buf->b_hdr, ==, hdr);
2728 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2729 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2730 return (buf_compressed == hdr_compressed &&
2731 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2732 !HDR_SHARED_DATA(hdr) &&
2733 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2737 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2738 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2739 * copy was made successfully, or an error code otherwise.
2742 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2743 boolean_t fill, arc_buf_t **ret)
2747 ASSERT(HDR_HAS_L1HDR(hdr));
2748 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2749 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2750 hdr->b_type == ARC_BUFC_METADATA);
2751 ASSERT3P(ret, !=, NULL);
2752 ASSERT3P(*ret, ==, NULL);
2754 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2757 buf->b_next = hdr->b_l1hdr.b_buf;
2760 add_reference(hdr, tag);
2763 * We're about to change the hdr's b_flags. We must either
2764 * hold the hash_lock or be undiscoverable.
2766 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2769 * Only honor requests for compressed bufs if the hdr is actually
2772 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2773 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2776 * If the hdr's data can be shared then we share the data buffer and
2777 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2778 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2779 * buffer to store the buf's data.
2781 * There are two additional restrictions here because we're sharing
2782 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2783 * actively involved in an L2ARC write, because if this buf is used by
2784 * an arc_write() then the hdr's data buffer will be released when the
2785 * write completes, even though the L2ARC write might still be using it.
2786 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2787 * need to be ABD-aware.
2789 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2790 abd_is_linear(hdr->b_l1hdr.b_pabd);
2792 /* Set up b_data and sharing */
2794 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2795 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2796 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2799 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2800 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2802 VERIFY3P(buf->b_data, !=, NULL);
2804 hdr->b_l1hdr.b_buf = buf;
2805 hdr->b_l1hdr.b_bufcnt += 1;
2808 * If the user wants the data from the hdr, we need to either copy or
2809 * decompress the data.
2812 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2818 static char *arc_onloan_tag = "onloan";
2821 arc_loaned_bytes_update(int64_t delta)
2823 atomic_add_64(&arc_loaned_bytes, delta);
2825 /* assert that it did not wrap around */
2826 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2830 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2831 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2832 * buffers must be returned to the arc before they can be used by the DMU or
2836 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2838 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2839 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2841 arc_loaned_bytes_update(size);
2847 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2848 enum zio_compress compression_type)
2850 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2851 psize, lsize, compression_type);
2853 arc_loaned_bytes_update(psize);
2860 * Return a loaned arc buffer to the arc.
2863 arc_return_buf(arc_buf_t *buf, void *tag)
2865 arc_buf_hdr_t *hdr = buf->b_hdr;
2867 ASSERT3P(buf->b_data, !=, NULL);
2868 ASSERT(HDR_HAS_L1HDR(hdr));
2869 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2870 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2872 arc_loaned_bytes_update(-arc_buf_size(buf));
2875 /* Detach an arc_buf from a dbuf (tag) */
2877 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2879 arc_buf_hdr_t *hdr = buf->b_hdr;
2881 ASSERT3P(buf->b_data, !=, NULL);
2882 ASSERT(HDR_HAS_L1HDR(hdr));
2883 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2884 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2886 arc_loaned_bytes_update(arc_buf_size(buf));
2890 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2892 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2895 df->l2df_size = size;
2896 df->l2df_type = type;
2897 mutex_enter(&l2arc_free_on_write_mtx);
2898 list_insert_head(l2arc_free_on_write, df);
2899 mutex_exit(&l2arc_free_on_write_mtx);
2903 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2905 arc_state_t *state = hdr->b_l1hdr.b_state;
2906 arc_buf_contents_t type = arc_buf_type(hdr);
2907 uint64_t size = arc_hdr_size(hdr);
2909 /* protected by hash lock, if in the hash table */
2910 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2911 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2912 ASSERT(state != arc_anon && state != arc_l2c_only);
2914 (void) refcount_remove_many(&state->arcs_esize[type],
2917 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2918 if (type == ARC_BUFC_METADATA) {
2919 arc_space_return(size, ARC_SPACE_META);
2921 ASSERT(type == ARC_BUFC_DATA);
2922 arc_space_return(size, ARC_SPACE_DATA);
2925 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2929 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2930 * data buffer, we transfer the refcount ownership to the hdr and update
2931 * the appropriate kstats.
2934 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2936 arc_state_t *state = hdr->b_l1hdr.b_state;
2938 ASSERT(arc_can_share(hdr, buf));
2939 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2940 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2943 * Start sharing the data buffer. We transfer the
2944 * refcount ownership to the hdr since it always owns
2945 * the refcount whenever an arc_buf_t is shared.
2947 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2948 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2949 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2950 HDR_ISTYPE_METADATA(hdr));
2951 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2952 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2955 * Since we've transferred ownership to the hdr we need
2956 * to increment its compressed and uncompressed kstats and
2957 * decrement the overhead size.
2959 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2960 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2961 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2965 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2967 arc_state_t *state = hdr->b_l1hdr.b_state;
2969 ASSERT(arc_buf_is_shared(buf));
2970 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2971 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2974 * We are no longer sharing this buffer so we need
2975 * to transfer its ownership to the rightful owner.
2977 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
2978 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2979 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
2980 abd_put(hdr->b_l1hdr.b_pabd);
2981 hdr->b_l1hdr.b_pabd = NULL;
2982 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2985 * Since the buffer is no longer shared between
2986 * the arc buf and the hdr, count it as overhead.
2988 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2989 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2990 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2994 * Remove an arc_buf_t from the hdr's buf list and return the last
2995 * arc_buf_t on the list. If no buffers remain on the list then return
2999 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3001 ASSERT(HDR_HAS_L1HDR(hdr));
3002 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3004 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3005 arc_buf_t *lastbuf = NULL;
3008 * Remove the buf from the hdr list and locate the last
3009 * remaining buffer on the list.
3011 while (*bufp != NULL) {
3013 *bufp = buf->b_next;
3016 * If we've removed a buffer in the middle of
3017 * the list then update the lastbuf and update
3020 if (*bufp != NULL) {
3022 bufp = &(*bufp)->b_next;
3026 ASSERT3P(lastbuf, !=, buf);
3027 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3028 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3029 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3035 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3039 arc_buf_destroy_impl(arc_buf_t *buf)
3041 arc_buf_hdr_t *hdr = buf->b_hdr;
3044 * Free up the data associated with the buf but only if we're not
3045 * sharing this with the hdr. If we are sharing it with the hdr, the
3046 * hdr is responsible for doing the free.
3048 if (buf->b_data != NULL) {
3050 * We're about to change the hdr's b_flags. We must either
3051 * hold the hash_lock or be undiscoverable.
3053 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3055 arc_cksum_verify(buf);
3057 arc_buf_unwatch(buf);
3060 if (arc_buf_is_shared(buf)) {
3061 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3063 uint64_t size = arc_buf_size(buf);
3064 arc_free_data_buf(hdr, buf->b_data, size, buf);
3065 ARCSTAT_INCR(arcstat_overhead_size, -size);
3069 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3070 hdr->b_l1hdr.b_bufcnt -= 1;
3073 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3075 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3077 * If the current arc_buf_t is sharing its data buffer with the
3078 * hdr, then reassign the hdr's b_pabd to share it with the new
3079 * buffer at the end of the list. The shared buffer is always
3080 * the last one on the hdr's buffer list.
3082 * There is an equivalent case for compressed bufs, but since
3083 * they aren't guaranteed to be the last buf in the list and
3084 * that is an exceedingly rare case, we just allow that space be
3085 * wasted temporarily.
3087 if (lastbuf != NULL) {
3088 /* Only one buf can be shared at once */
3089 VERIFY(!arc_buf_is_shared(lastbuf));
3090 /* hdr is uncompressed so can't have compressed buf */
3091 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3093 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3094 arc_hdr_free_pabd(hdr);
3097 * We must setup a new shared block between the
3098 * last buffer and the hdr. The data would have
3099 * been allocated by the arc buf so we need to transfer
3100 * ownership to the hdr since it's now being shared.
3102 arc_share_buf(hdr, lastbuf);
3104 } else if (HDR_SHARED_DATA(hdr)) {
3106 * Uncompressed shared buffers are always at the end
3107 * of the list. Compressed buffers don't have the
3108 * same requirements. This makes it hard to
3109 * simply assert that the lastbuf is shared so
3110 * we rely on the hdr's compression flags to determine
3111 * if we have a compressed, shared buffer.
3113 ASSERT3P(lastbuf, !=, NULL);
3114 ASSERT(arc_buf_is_shared(lastbuf) ||
3115 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
3119 * Free the checksum if we're removing the last uncompressed buf from
3122 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3123 arc_cksum_free(hdr);
3126 /* clean up the buf */
3128 kmem_cache_free(buf_cache, buf);
3132 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
3134 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3135 ASSERT(HDR_HAS_L1HDR(hdr));
3136 ASSERT(!HDR_SHARED_DATA(hdr));
3138 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3139 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
3140 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3141 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3143 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3144 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3148 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3150 ASSERT(HDR_HAS_L1HDR(hdr));
3151 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3154 * If the hdr is currently being written to the l2arc then
3155 * we defer freeing the data by adding it to the l2arc_free_on_write
3156 * list. The l2arc will free the data once it's finished
3157 * writing it to the l2arc device.
3159 if (HDR_L2_WRITING(hdr)) {
3160 arc_hdr_free_on_write(hdr);
3161 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3163 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3164 arc_hdr_size(hdr), hdr);
3166 hdr->b_l1hdr.b_pabd = NULL;
3167 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3169 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3170 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3173 static arc_buf_hdr_t *
3174 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3175 enum zio_compress compression_type, arc_buf_contents_t type)
3179 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3181 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3182 ASSERT(HDR_EMPTY(hdr));
3183 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3184 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3185 HDR_SET_PSIZE(hdr, psize);
3186 HDR_SET_LSIZE(hdr, lsize);
3190 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3191 arc_hdr_set_compress(hdr, compression_type);
3193 hdr->b_l1hdr.b_state = arc_anon;
3194 hdr->b_l1hdr.b_arc_access = 0;
3195 hdr->b_l1hdr.b_bufcnt = 0;
3196 hdr->b_l1hdr.b_buf = NULL;
3199 * Allocate the hdr's buffer. This will contain either
3200 * the compressed or uncompressed data depending on the block
3201 * it references and compressed arc enablement.
3203 arc_hdr_alloc_pabd(hdr);
3204 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3210 * Transition between the two allocation states for the arc_buf_hdr struct.
3211 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3212 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3213 * version is used when a cache buffer is only in the L2ARC in order to reduce
3216 static arc_buf_hdr_t *
3217 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3219 ASSERT(HDR_HAS_L2HDR(hdr));
3221 arc_buf_hdr_t *nhdr;
3222 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3224 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3225 (old == hdr_l2only_cache && new == hdr_full_cache));
3227 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3229 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3230 buf_hash_remove(hdr);
3232 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3234 if (new == hdr_full_cache) {
3235 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3237 * arc_access and arc_change_state need to be aware that a
3238 * header has just come out of L2ARC, so we set its state to
3239 * l2c_only even though it's about to change.
3241 nhdr->b_l1hdr.b_state = arc_l2c_only;
3243 /* Verify previous threads set to NULL before freeing */
3244 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3246 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3247 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3248 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3251 * If we've reached here, We must have been called from
3252 * arc_evict_hdr(), as such we should have already been
3253 * removed from any ghost list we were previously on
3254 * (which protects us from racing with arc_evict_state),
3255 * thus no locking is needed during this check.
3257 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3260 * A buffer must not be moved into the arc_l2c_only
3261 * state if it's not finished being written out to the
3262 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3263 * might try to be accessed, even though it was removed.
3265 VERIFY(!HDR_L2_WRITING(hdr));
3266 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3269 if (hdr->b_l1hdr.b_thawed != NULL) {
3270 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3271 hdr->b_l1hdr.b_thawed = NULL;
3275 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3278 * The header has been reallocated so we need to re-insert it into any
3281 (void) buf_hash_insert(nhdr, NULL);
3283 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3285 mutex_enter(&dev->l2ad_mtx);
3288 * We must place the realloc'ed header back into the list at
3289 * the same spot. Otherwise, if it's placed earlier in the list,
3290 * l2arc_write_buffers() could find it during the function's
3291 * write phase, and try to write it out to the l2arc.
3293 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3294 list_remove(&dev->l2ad_buflist, hdr);
3296 mutex_exit(&dev->l2ad_mtx);
3299 * Since we're using the pointer address as the tag when
3300 * incrementing and decrementing the l2ad_alloc refcount, we
3301 * must remove the old pointer (that we're about to destroy) and
3302 * add the new pointer to the refcount. Otherwise we'd remove
3303 * the wrong pointer address when calling arc_hdr_destroy() later.
3306 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3307 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3309 buf_discard_identity(hdr);
3310 kmem_cache_free(old, hdr);
3316 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3317 * The buf is returned thawed since we expect the consumer to modify it.
3320 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3322 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3323 ZIO_COMPRESS_OFF, type);
3324 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3326 arc_buf_t *buf = NULL;
3327 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3334 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3335 * for bufs containing metadata.
3338 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3339 enum zio_compress compression_type)
3341 ASSERT3U(lsize, >, 0);
3342 ASSERT3U(lsize, >=, psize);
3343 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3344 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3346 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3347 compression_type, ARC_BUFC_DATA);
3348 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3350 arc_buf_t *buf = NULL;
3351 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3353 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3355 if (!arc_buf_is_shared(buf)) {
3357 * To ensure that the hdr has the correct data in it if we call
3358 * arc_decompress() on this buf before it's been written to
3359 * disk, it's easiest if we just set up sharing between the
3362 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3363 arc_hdr_free_pabd(hdr);
3364 arc_share_buf(hdr, buf);
3371 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3373 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3374 l2arc_dev_t *dev = l2hdr->b_dev;
3375 uint64_t psize = arc_hdr_size(hdr);
3377 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3378 ASSERT(HDR_HAS_L2HDR(hdr));
3380 list_remove(&dev->l2ad_buflist, hdr);
3382 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3383 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3385 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3387 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3388 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3392 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3394 if (HDR_HAS_L1HDR(hdr)) {
3395 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3396 hdr->b_l1hdr.b_bufcnt > 0);
3397 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3398 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3400 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3401 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3403 if (!HDR_EMPTY(hdr))
3404 buf_discard_identity(hdr);
3406 if (HDR_HAS_L2HDR(hdr)) {
3407 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3408 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3411 mutex_enter(&dev->l2ad_mtx);
3414 * Even though we checked this conditional above, we
3415 * need to check this again now that we have the
3416 * l2ad_mtx. This is because we could be racing with
3417 * another thread calling l2arc_evict() which might have
3418 * destroyed this header's L2 portion as we were waiting
3419 * to acquire the l2ad_mtx. If that happens, we don't
3420 * want to re-destroy the header's L2 portion.
3422 if (HDR_HAS_L2HDR(hdr)) {
3424 arc_hdr_l2hdr_destroy(hdr);
3428 mutex_exit(&dev->l2ad_mtx);
3431 if (HDR_HAS_L1HDR(hdr)) {
3432 arc_cksum_free(hdr);
3434 while (hdr->b_l1hdr.b_buf != NULL)
3435 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3438 if (hdr->b_l1hdr.b_thawed != NULL) {
3439 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3440 hdr->b_l1hdr.b_thawed = NULL;
3444 if (hdr->b_l1hdr.b_pabd != NULL) {
3445 arc_hdr_free_pabd(hdr);
3449 ASSERT3P(hdr->b_hash_next, ==, NULL);
3450 if (HDR_HAS_L1HDR(hdr)) {
3451 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3452 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3453 kmem_cache_free(hdr_full_cache, hdr);
3455 kmem_cache_free(hdr_l2only_cache, hdr);
3460 arc_buf_destroy(arc_buf_t *buf, void* tag)
3462 arc_buf_hdr_t *hdr = buf->b_hdr;
3463 kmutex_t *hash_lock = HDR_LOCK(hdr);
3465 if (hdr->b_l1hdr.b_state == arc_anon) {
3466 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3467 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3468 VERIFY0(remove_reference(hdr, NULL, tag));
3469 arc_hdr_destroy(hdr);
3473 mutex_enter(hash_lock);
3474 ASSERT3P(hdr, ==, buf->b_hdr);
3475 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3476 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3477 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3478 ASSERT3P(buf->b_data, !=, NULL);
3480 (void) remove_reference(hdr, hash_lock, tag);
3481 arc_buf_destroy_impl(buf);
3482 mutex_exit(hash_lock);
3486 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3487 * state of the header is dependent on its state prior to entering this
3488 * function. The following transitions are possible:
3490 * - arc_mru -> arc_mru_ghost
3491 * - arc_mfu -> arc_mfu_ghost
3492 * - arc_mru_ghost -> arc_l2c_only
3493 * - arc_mru_ghost -> deleted
3494 * - arc_mfu_ghost -> arc_l2c_only
3495 * - arc_mfu_ghost -> deleted
3498 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3500 arc_state_t *evicted_state, *state;
3501 int64_t bytes_evicted = 0;
3503 ASSERT(MUTEX_HELD(hash_lock));
3504 ASSERT(HDR_HAS_L1HDR(hdr));
3506 state = hdr->b_l1hdr.b_state;
3507 if (GHOST_STATE(state)) {
3508 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3509 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3512 * l2arc_write_buffers() relies on a header's L1 portion
3513 * (i.e. its b_pabd field) during it's write phase.
3514 * Thus, we cannot push a header onto the arc_l2c_only
3515 * state (removing it's L1 piece) until the header is
3516 * done being written to the l2arc.
3518 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3519 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3520 return (bytes_evicted);
3523 ARCSTAT_BUMP(arcstat_deleted);
3524 bytes_evicted += HDR_GET_LSIZE(hdr);
3526 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3528 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3529 if (HDR_HAS_L2HDR(hdr)) {
3531 * This buffer is cached on the 2nd Level ARC;
3532 * don't destroy the header.
3534 arc_change_state(arc_l2c_only, hdr, hash_lock);
3536 * dropping from L1+L2 cached to L2-only,
3537 * realloc to remove the L1 header.
3539 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3542 arc_change_state(arc_anon, hdr, hash_lock);
3543 arc_hdr_destroy(hdr);
3545 return (bytes_evicted);
3548 ASSERT(state == arc_mru || state == arc_mfu);
3549 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3551 /* prefetch buffers have a minimum lifespan */
3552 if (HDR_IO_IN_PROGRESS(hdr) ||
3553 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3554 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3555 arc_min_prefetch_lifespan)) {
3556 ARCSTAT_BUMP(arcstat_evict_skip);
3557 return (bytes_evicted);
3560 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3561 while (hdr->b_l1hdr.b_buf) {
3562 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3563 if (!mutex_tryenter(&buf->b_evict_lock)) {
3564 ARCSTAT_BUMP(arcstat_mutex_miss);
3567 if (buf->b_data != NULL)
3568 bytes_evicted += HDR_GET_LSIZE(hdr);
3569 mutex_exit(&buf->b_evict_lock);
3570 arc_buf_destroy_impl(buf);
3573 if (HDR_HAS_L2HDR(hdr)) {
3574 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3576 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3577 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3578 HDR_GET_LSIZE(hdr));
3580 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3581 HDR_GET_LSIZE(hdr));
3585 if (hdr->b_l1hdr.b_bufcnt == 0) {
3586 arc_cksum_free(hdr);
3588 bytes_evicted += arc_hdr_size(hdr);
3591 * If this hdr is being evicted and has a compressed
3592 * buffer then we discard it here before we change states.
3593 * This ensures that the accounting is updated correctly
3594 * in arc_free_data_impl().
3596 arc_hdr_free_pabd(hdr);
3598 arc_change_state(evicted_state, hdr, hash_lock);
3599 ASSERT(HDR_IN_HASH_TABLE(hdr));
3600 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3601 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3604 return (bytes_evicted);
3608 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3609 uint64_t spa, int64_t bytes)
3611 multilist_sublist_t *mls;
3612 uint64_t bytes_evicted = 0;
3614 kmutex_t *hash_lock;
3615 int evict_count = 0;
3617 ASSERT3P(marker, !=, NULL);
3618 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3620 mls = multilist_sublist_lock(ml, idx);
3622 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3623 hdr = multilist_sublist_prev(mls, marker)) {
3624 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3625 (evict_count >= zfs_arc_evict_batch_limit))
3629 * To keep our iteration location, move the marker
3630 * forward. Since we're not holding hdr's hash lock, we
3631 * must be very careful and not remove 'hdr' from the
3632 * sublist. Otherwise, other consumers might mistake the
3633 * 'hdr' as not being on a sublist when they call the
3634 * multilist_link_active() function (they all rely on
3635 * the hash lock protecting concurrent insertions and
3636 * removals). multilist_sublist_move_forward() was
3637 * specifically implemented to ensure this is the case
3638 * (only 'marker' will be removed and re-inserted).
3640 multilist_sublist_move_forward(mls, marker);
3643 * The only case where the b_spa field should ever be
3644 * zero, is the marker headers inserted by
3645 * arc_evict_state(). It's possible for multiple threads
3646 * to be calling arc_evict_state() concurrently (e.g.
3647 * dsl_pool_close() and zio_inject_fault()), so we must
3648 * skip any markers we see from these other threads.
3650 if (hdr->b_spa == 0)
3653 /* we're only interested in evicting buffers of a certain spa */
3654 if (spa != 0 && hdr->b_spa != spa) {
3655 ARCSTAT_BUMP(arcstat_evict_skip);
3659 hash_lock = HDR_LOCK(hdr);
3662 * We aren't calling this function from any code path
3663 * that would already be holding a hash lock, so we're
3664 * asserting on this assumption to be defensive in case
3665 * this ever changes. Without this check, it would be
3666 * possible to incorrectly increment arcstat_mutex_miss
3667 * below (e.g. if the code changed such that we called
3668 * this function with a hash lock held).
3670 ASSERT(!MUTEX_HELD(hash_lock));
3672 if (mutex_tryenter(hash_lock)) {
3673 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3674 mutex_exit(hash_lock);
3676 bytes_evicted += evicted;
3679 * If evicted is zero, arc_evict_hdr() must have
3680 * decided to skip this header, don't increment
3681 * evict_count in this case.
3687 * If arc_size isn't overflowing, signal any
3688 * threads that might happen to be waiting.
3690 * For each header evicted, we wake up a single
3691 * thread. If we used cv_broadcast, we could
3692 * wake up "too many" threads causing arc_size
3693 * to significantly overflow arc_c; since
3694 * arc_get_data_impl() doesn't check for overflow
3695 * when it's woken up (it doesn't because it's
3696 * possible for the ARC to be overflowing while
3697 * full of un-evictable buffers, and the
3698 * function should proceed in this case).
3700 * If threads are left sleeping, due to not
3701 * using cv_broadcast, they will be woken up
3702 * just before arc_reclaim_thread() sleeps.
3704 mutex_enter(&arc_reclaim_lock);
3705 if (!arc_is_overflowing())
3706 cv_signal(&arc_reclaim_waiters_cv);
3707 mutex_exit(&arc_reclaim_lock);
3709 ARCSTAT_BUMP(arcstat_mutex_miss);
3713 multilist_sublist_unlock(mls);
3715 return (bytes_evicted);
3719 * Evict buffers from the given arc state, until we've removed the
3720 * specified number of bytes. Move the removed buffers to the
3721 * appropriate evict state.
3723 * This function makes a "best effort". It skips over any buffers
3724 * it can't get a hash_lock on, and so, may not catch all candidates.
3725 * It may also return without evicting as much space as requested.
3727 * If bytes is specified using the special value ARC_EVICT_ALL, this
3728 * will evict all available (i.e. unlocked and evictable) buffers from
3729 * the given arc state; which is used by arc_flush().
3732 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3733 arc_buf_contents_t type)
3735 uint64_t total_evicted = 0;
3736 multilist_t *ml = state->arcs_list[type];
3738 arc_buf_hdr_t **markers;
3740 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3742 num_sublists = multilist_get_num_sublists(ml);
3745 * If we've tried to evict from each sublist, made some
3746 * progress, but still have not hit the target number of bytes
3747 * to evict, we want to keep trying. The markers allow us to
3748 * pick up where we left off for each individual sublist, rather
3749 * than starting from the tail each time.
3751 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3752 for (int i = 0; i < num_sublists; i++) {
3753 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3756 * A b_spa of 0 is used to indicate that this header is
3757 * a marker. This fact is used in arc_adjust_type() and
3758 * arc_evict_state_impl().
3760 markers[i]->b_spa = 0;
3762 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3763 multilist_sublist_insert_tail(mls, markers[i]);
3764 multilist_sublist_unlock(mls);
3768 * While we haven't hit our target number of bytes to evict, or
3769 * we're evicting all available buffers.
3771 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3773 * Start eviction using a randomly selected sublist,
3774 * this is to try and evenly balance eviction across all
3775 * sublists. Always starting at the same sublist
3776 * (e.g. index 0) would cause evictions to favor certain
3777 * sublists over others.
3779 int sublist_idx = multilist_get_random_index(ml);
3780 uint64_t scan_evicted = 0;
3782 for (int i = 0; i < num_sublists; i++) {
3783 uint64_t bytes_remaining;
3784 uint64_t bytes_evicted;
3786 if (bytes == ARC_EVICT_ALL)
3787 bytes_remaining = ARC_EVICT_ALL;
3788 else if (total_evicted < bytes)
3789 bytes_remaining = bytes - total_evicted;
3793 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3794 markers[sublist_idx], spa, bytes_remaining);
3796 scan_evicted += bytes_evicted;
3797 total_evicted += bytes_evicted;
3799 /* we've reached the end, wrap to the beginning */
3800 if (++sublist_idx >= num_sublists)
3805 * If we didn't evict anything during this scan, we have
3806 * no reason to believe we'll evict more during another
3807 * scan, so break the loop.
3809 if (scan_evicted == 0) {
3810 /* This isn't possible, let's make that obvious */
3811 ASSERT3S(bytes, !=, 0);
3814 * When bytes is ARC_EVICT_ALL, the only way to
3815 * break the loop is when scan_evicted is zero.
3816 * In that case, we actually have evicted enough,
3817 * so we don't want to increment the kstat.
3819 if (bytes != ARC_EVICT_ALL) {
3820 ASSERT3S(total_evicted, <, bytes);
3821 ARCSTAT_BUMP(arcstat_evict_not_enough);
3828 for (int i = 0; i < num_sublists; i++) {
3829 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3830 multilist_sublist_remove(mls, markers[i]);
3831 multilist_sublist_unlock(mls);
3833 kmem_cache_free(hdr_full_cache, markers[i]);
3835 kmem_free(markers, sizeof (*markers) * num_sublists);
3837 return (total_evicted);
3841 * Flush all "evictable" data of the given type from the arc state
3842 * specified. This will not evict any "active" buffers (i.e. referenced).
3844 * When 'retry' is set to B_FALSE, the function will make a single pass
3845 * over the state and evict any buffers that it can. Since it doesn't
3846 * continually retry the eviction, it might end up leaving some buffers
3847 * in the ARC due to lock misses.
3849 * When 'retry' is set to B_TRUE, the function will continually retry the
3850 * eviction until *all* evictable buffers have been removed from the
3851 * state. As a result, if concurrent insertions into the state are
3852 * allowed (e.g. if the ARC isn't shutting down), this function might
3853 * wind up in an infinite loop, continually trying to evict buffers.
3856 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3859 uint64_t evicted = 0;
3861 while (refcount_count(&state->arcs_esize[type]) != 0) {
3862 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3872 * Evict the specified number of bytes from the state specified,
3873 * restricting eviction to the spa and type given. This function
3874 * prevents us from trying to evict more from a state's list than
3875 * is "evictable", and to skip evicting altogether when passed a
3876 * negative value for "bytes". In contrast, arc_evict_state() will
3877 * evict everything it can, when passed a negative value for "bytes".
3880 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3881 arc_buf_contents_t type)
3885 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3886 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3887 return (arc_evict_state(state, spa, delta, type));
3894 * Evict metadata buffers from the cache, such that arc_meta_used is
3895 * capped by the arc_meta_limit tunable.
3898 arc_adjust_meta(void)
3900 uint64_t total_evicted = 0;
3904 * If we're over the meta limit, we want to evict enough
3905 * metadata to get back under the meta limit. We don't want to
3906 * evict so much that we drop the MRU below arc_p, though. If
3907 * we're over the meta limit more than we're over arc_p, we
3908 * evict some from the MRU here, and some from the MFU below.
3910 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3911 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3912 refcount_count(&arc_mru->arcs_size) - arc_p));
3914 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3917 * Similar to the above, we want to evict enough bytes to get us
3918 * below the meta limit, but not so much as to drop us below the
3919 * space allotted to the MFU (which is defined as arc_c - arc_p).
3921 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3922 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
3924 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3926 return (total_evicted);
3930 * Return the type of the oldest buffer in the given arc state
3932 * This function will select a random sublist of type ARC_BUFC_DATA and
3933 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3934 * is compared, and the type which contains the "older" buffer will be
3937 static arc_buf_contents_t
3938 arc_adjust_type(arc_state_t *state)
3940 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
3941 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
3942 int data_idx = multilist_get_random_index(data_ml);
3943 int meta_idx = multilist_get_random_index(meta_ml);
3944 multilist_sublist_t *data_mls;
3945 multilist_sublist_t *meta_mls;
3946 arc_buf_contents_t type;
3947 arc_buf_hdr_t *data_hdr;
3948 arc_buf_hdr_t *meta_hdr;
3951 * We keep the sublist lock until we're finished, to prevent
3952 * the headers from being destroyed via arc_evict_state().
3954 data_mls = multilist_sublist_lock(data_ml, data_idx);
3955 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3958 * These two loops are to ensure we skip any markers that
3959 * might be at the tail of the lists due to arc_evict_state().
3962 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3963 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3964 if (data_hdr->b_spa != 0)
3968 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3969 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3970 if (meta_hdr->b_spa != 0)
3974 if (data_hdr == NULL && meta_hdr == NULL) {
3975 type = ARC_BUFC_DATA;
3976 } else if (data_hdr == NULL) {
3977 ASSERT3P(meta_hdr, !=, NULL);
3978 type = ARC_BUFC_METADATA;
3979 } else if (meta_hdr == NULL) {
3980 ASSERT3P(data_hdr, !=, NULL);
3981 type = ARC_BUFC_DATA;
3983 ASSERT3P(data_hdr, !=, NULL);
3984 ASSERT3P(meta_hdr, !=, NULL);
3986 /* The headers can't be on the sublist without an L1 header */
3987 ASSERT(HDR_HAS_L1HDR(data_hdr));
3988 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3990 if (data_hdr->b_l1hdr.b_arc_access <
3991 meta_hdr->b_l1hdr.b_arc_access) {
3992 type = ARC_BUFC_DATA;
3994 type = ARC_BUFC_METADATA;
3998 multilist_sublist_unlock(meta_mls);
3999 multilist_sublist_unlock(data_mls);
4005 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4010 uint64_t total_evicted = 0;
4015 * If we're over arc_meta_limit, we want to correct that before
4016 * potentially evicting data buffers below.
4018 total_evicted += arc_adjust_meta();
4023 * If we're over the target cache size, we want to evict enough
4024 * from the list to get back to our target size. We don't want
4025 * to evict too much from the MRU, such that it drops below
4026 * arc_p. So, if we're over our target cache size more than
4027 * the MRU is over arc_p, we'll evict enough to get back to
4028 * arc_p here, and then evict more from the MFU below.
4030 target = MIN((int64_t)(arc_size - arc_c),
4031 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4032 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
4035 * If we're below arc_meta_min, always prefer to evict data.
4036 * Otherwise, try to satisfy the requested number of bytes to
4037 * evict from the type which contains older buffers; in an
4038 * effort to keep newer buffers in the cache regardless of their
4039 * type. If we cannot satisfy the number of bytes from this
4040 * type, spill over into the next type.
4042 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4043 arc_meta_used > arc_meta_min) {
4044 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4045 total_evicted += bytes;
4048 * If we couldn't evict our target number of bytes from
4049 * metadata, we try to get the rest from data.
4054 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4056 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4057 total_evicted += bytes;
4060 * If we couldn't evict our target number of bytes from
4061 * data, we try to get the rest from metadata.
4066 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4072 * Now that we've tried to evict enough from the MRU to get its
4073 * size back to arc_p, if we're still above the target cache
4074 * size, we evict the rest from the MFU.
4076 target = arc_size - arc_c;
4078 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4079 arc_meta_used > arc_meta_min) {
4080 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4081 total_evicted += bytes;
4084 * If we couldn't evict our target number of bytes from
4085 * metadata, we try to get the rest from data.
4090 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4092 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4093 total_evicted += bytes;
4096 * If we couldn't evict our target number of bytes from
4097 * data, we try to get the rest from data.
4102 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4106 * Adjust ghost lists
4108 * In addition to the above, the ARC also defines target values
4109 * for the ghost lists. The sum of the mru list and mru ghost
4110 * list should never exceed the target size of the cache, and
4111 * the sum of the mru list, mfu list, mru ghost list, and mfu
4112 * ghost list should never exceed twice the target size of the
4113 * cache. The following logic enforces these limits on the ghost
4114 * caches, and evicts from them as needed.
4116 target = refcount_count(&arc_mru->arcs_size) +
4117 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4119 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4120 total_evicted += bytes;
4125 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4128 * We assume the sum of the mru list and mfu list is less than
4129 * or equal to arc_c (we enforced this above), which means we
4130 * can use the simpler of the two equations below:
4132 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4133 * mru ghost + mfu ghost <= arc_c
4135 target = refcount_count(&arc_mru_ghost->arcs_size) +
4136 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4138 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4139 total_evicted += bytes;
4144 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4146 return (total_evicted);
4150 arc_flush(spa_t *spa, boolean_t retry)
4155 * If retry is B_TRUE, a spa must not be specified since we have
4156 * no good way to determine if all of a spa's buffers have been
4157 * evicted from an arc state.
4159 ASSERT(!retry || spa == 0);
4162 guid = spa_load_guid(spa);
4164 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4165 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4167 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4168 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4170 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4171 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4173 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4174 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4178 arc_shrink(int64_t to_free)
4180 if (arc_c > arc_c_min) {
4181 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
4182 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
4183 if (arc_c > arc_c_min + to_free)
4184 atomic_add_64(&arc_c, -to_free);
4188 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4189 if (arc_c > arc_size)
4190 arc_c = MAX(arc_size, arc_c_min);
4192 arc_p = (arc_c >> 1);
4194 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
4197 ASSERT(arc_c >= arc_c_min);
4198 ASSERT((int64_t)arc_p >= 0);
4201 if (arc_size > arc_c) {
4202 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size,
4204 (void) arc_adjust();
4208 typedef enum free_memory_reason_t {
4213 FMR_PAGES_PP_MAXIMUM,
4216 } free_memory_reason_t;
4218 int64_t last_free_memory;
4219 free_memory_reason_t last_free_reason;
4222 * Additional reserve of pages for pp_reserve.
4224 int64_t arc_pages_pp_reserve = 64;
4227 * Additional reserve of pages for swapfs.
4229 int64_t arc_swapfs_reserve = 64;
4232 * Return the amount of memory that can be consumed before reclaim will be
4233 * needed. Positive if there is sufficient free memory, negative indicates
4234 * the amount of memory that needs to be freed up.
4237 arc_available_memory(void)
4239 int64_t lowest = INT64_MAX;
4241 free_memory_reason_t r = FMR_UNKNOWN;
4246 * Cooperate with pagedaemon when it's time for it to scan
4247 * and reclaim some pages.
4249 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
4257 n = PAGESIZE * (-needfree);
4265 * check that we're out of range of the pageout scanner. It starts to
4266 * schedule paging if freemem is less than lotsfree and needfree.
4267 * lotsfree is the high-water mark for pageout, and needfree is the
4268 * number of needed free pages. We add extra pages here to make sure
4269 * the scanner doesn't start up while we're freeing memory.
4271 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4278 * check to make sure that swapfs has enough space so that anon
4279 * reservations can still succeed. anon_resvmem() checks that the
4280 * availrmem is greater than swapfs_minfree, and the number of reserved
4281 * swap pages. We also add a bit of extra here just to prevent
4282 * circumstances from getting really dire.
4284 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4285 desfree - arc_swapfs_reserve);
4288 r = FMR_SWAPFS_MINFREE;
4293 * Check that we have enough availrmem that memory locking (e.g., via
4294 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4295 * stores the number of pages that cannot be locked; when availrmem
4296 * drops below pages_pp_maximum, page locking mechanisms such as
4297 * page_pp_lock() will fail.)
4299 n = PAGESIZE * (availrmem - pages_pp_maximum -
4300 arc_pages_pp_reserve);
4303 r = FMR_PAGES_PP_MAXIMUM;
4306 #endif /* __FreeBSD__ */
4307 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4309 * If we're on an i386 platform, it's possible that we'll exhaust the
4310 * kernel heap space before we ever run out of available physical
4311 * memory. Most checks of the size of the heap_area compare against
4312 * tune.t_minarmem, which is the minimum available real memory that we
4313 * can have in the system. However, this is generally fixed at 25 pages
4314 * which is so low that it's useless. In this comparison, we seek to
4315 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4316 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4319 n = uma_avail() - (long)(uma_limit() / 4);
4327 * If zio data pages are being allocated out of a separate heap segment,
4328 * then enforce that the size of available vmem for this arena remains
4329 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4331 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4332 * memory (in the zio_arena) free, which can avoid memory
4333 * fragmentation issues.
4335 if (zio_arena != NULL) {
4336 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4337 (vmem_size(zio_arena, VMEM_ALLOC) >>
4338 arc_zio_arena_free_shift);
4346 /* Every 100 calls, free a small amount */
4347 if (spa_get_random(100) == 0)
4349 #endif /* _KERNEL */
4351 last_free_memory = lowest;
4352 last_free_reason = r;
4353 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4359 * Determine if the system is under memory pressure and is asking
4360 * to reclaim memory. A return value of B_TRUE indicates that the system
4361 * is under memory pressure and that the arc should adjust accordingly.
4364 arc_reclaim_needed(void)
4366 return (arc_available_memory() < 0);
4369 extern kmem_cache_t *zio_buf_cache[];
4370 extern kmem_cache_t *zio_data_buf_cache[];
4371 extern kmem_cache_t *range_seg_cache;
4372 extern kmem_cache_t *abd_chunk_cache;
4374 static __noinline void
4375 arc_kmem_reap_now(void)
4378 kmem_cache_t *prev_cache = NULL;
4379 kmem_cache_t *prev_data_cache = NULL;
4381 DTRACE_PROBE(arc__kmem_reap_start);
4383 if (arc_meta_used >= arc_meta_limit) {
4385 * We are exceeding our meta-data cache limit.
4386 * Purge some DNLC entries to release holds on meta-data.
4388 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4392 * Reclaim unused memory from all kmem caches.
4398 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4399 if (zio_buf_cache[i] != prev_cache) {
4400 prev_cache = zio_buf_cache[i];
4401 kmem_cache_reap_now(zio_buf_cache[i]);
4403 if (zio_data_buf_cache[i] != prev_data_cache) {
4404 prev_data_cache = zio_data_buf_cache[i];
4405 kmem_cache_reap_now(zio_data_buf_cache[i]);
4408 kmem_cache_reap_now(abd_chunk_cache);
4409 kmem_cache_reap_now(buf_cache);
4410 kmem_cache_reap_now(hdr_full_cache);
4411 kmem_cache_reap_now(hdr_l2only_cache);
4412 kmem_cache_reap_now(range_seg_cache);
4415 if (zio_arena != NULL) {
4417 * Ask the vmem arena to reclaim unused memory from its
4420 vmem_qcache_reap(zio_arena);
4423 DTRACE_PROBE(arc__kmem_reap_end);
4427 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4428 * enough data and signal them to proceed. When this happens, the threads in
4429 * arc_get_data_impl() are sleeping while holding the hash lock for their
4430 * particular arc header. Thus, we must be careful to never sleep on a
4431 * hash lock in this thread. This is to prevent the following deadlock:
4433 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4434 * waiting for the reclaim thread to signal it.
4436 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4437 * fails, and goes to sleep forever.
4439 * This possible deadlock is avoided by always acquiring a hash lock
4440 * using mutex_tryenter() from arc_reclaim_thread().
4444 arc_reclaim_thread(void *unused __unused)
4446 hrtime_t growtime = 0;
4449 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4451 mutex_enter(&arc_reclaim_lock);
4452 while (!arc_reclaim_thread_exit) {
4453 uint64_t evicted = 0;
4456 * This is necessary in order for the mdb ::arc dcmd to
4457 * show up to date information. Since the ::arc command
4458 * does not call the kstat's update function, without
4459 * this call, the command may show stale stats for the
4460 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4461 * with this change, the data might be up to 1 second
4462 * out of date; but that should suffice. The arc_state_t
4463 * structures can be queried directly if more accurate
4464 * information is needed.
4466 if (arc_ksp != NULL)
4467 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4469 mutex_exit(&arc_reclaim_lock);
4472 * We call arc_adjust() before (possibly) calling
4473 * arc_kmem_reap_now(), so that we can wake up
4474 * arc_get_data_impl() sooner.
4476 evicted = arc_adjust();
4478 int64_t free_memory = arc_available_memory();
4479 if (free_memory < 0) {
4481 arc_no_grow = B_TRUE;
4485 * Wait at least zfs_grow_retry (default 60) seconds
4486 * before considering growing.
4488 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4490 arc_kmem_reap_now();
4493 * If we are still low on memory, shrink the ARC
4494 * so that we have arc_shrink_min free space.
4496 free_memory = arc_available_memory();
4499 (arc_c >> arc_shrink_shift) - free_memory;
4503 to_free = MAX(to_free, ptob(needfree));
4506 arc_shrink(to_free);
4508 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4509 arc_no_grow = B_TRUE;
4510 } else if (gethrtime() >= growtime) {
4511 arc_no_grow = B_FALSE;
4514 mutex_enter(&arc_reclaim_lock);
4517 * If evicted is zero, we couldn't evict anything via
4518 * arc_adjust(). This could be due to hash lock
4519 * collisions, but more likely due to the majority of
4520 * arc buffers being unevictable. Therefore, even if
4521 * arc_size is above arc_c, another pass is unlikely to
4522 * be helpful and could potentially cause us to enter an
4525 if (arc_size <= arc_c || evicted == 0) {
4527 * We're either no longer overflowing, or we
4528 * can't evict anything more, so we should wake
4529 * up any threads before we go to sleep.
4531 cv_broadcast(&arc_reclaim_waiters_cv);
4534 * Block until signaled, or after one second (we
4535 * might need to perform arc_kmem_reap_now()
4536 * even if we aren't being signalled)
4538 CALLB_CPR_SAFE_BEGIN(&cpr);
4539 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4540 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4541 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4545 arc_reclaim_thread_exit = B_FALSE;
4546 cv_broadcast(&arc_reclaim_thread_cv);
4547 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4551 static u_int arc_dnlc_evicts_arg;
4552 extern struct vfsops zfs_vfsops;
4555 arc_dnlc_evicts_thread(void *dummy __unused)
4560 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4562 mutex_enter(&arc_dnlc_evicts_lock);
4563 while (!arc_dnlc_evicts_thread_exit) {
4564 CALLB_CPR_SAFE_BEGIN(&cpr);
4565 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4566 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4567 if (arc_dnlc_evicts_arg != 0) {
4568 percent = arc_dnlc_evicts_arg;
4569 mutex_exit(&arc_dnlc_evicts_lock);
4571 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4573 mutex_enter(&arc_dnlc_evicts_lock);
4575 * Clear our token only after vnlru_free()
4576 * pass is done, to avoid false queueing of
4579 arc_dnlc_evicts_arg = 0;
4582 arc_dnlc_evicts_thread_exit = FALSE;
4583 cv_broadcast(&arc_dnlc_evicts_cv);
4584 CALLB_CPR_EXIT(&cpr);
4589 dnlc_reduce_cache(void *arg)
4593 percent = (u_int)(uintptr_t)arg;
4594 mutex_enter(&arc_dnlc_evicts_lock);
4595 if (arc_dnlc_evicts_arg == 0) {
4596 arc_dnlc_evicts_arg = percent;
4597 cv_broadcast(&arc_dnlc_evicts_cv);
4599 mutex_exit(&arc_dnlc_evicts_lock);
4603 * Adapt arc info given the number of bytes we are trying to add and
4604 * the state that we are comming from. This function is only called
4605 * when we are adding new content to the cache.
4608 arc_adapt(int bytes, arc_state_t *state)
4611 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4612 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4613 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4615 if (state == arc_l2c_only)
4620 * Adapt the target size of the MRU list:
4621 * - if we just hit in the MRU ghost list, then increase
4622 * the target size of the MRU list.
4623 * - if we just hit in the MFU ghost list, then increase
4624 * the target size of the MFU list by decreasing the
4625 * target size of the MRU list.
4627 if (state == arc_mru_ghost) {
4628 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4629 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4631 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4632 } else if (state == arc_mfu_ghost) {
4635 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4636 mult = MIN(mult, 10);
4638 delta = MIN(bytes * mult, arc_p);
4639 arc_p = MAX(arc_p_min, arc_p - delta);
4641 ASSERT((int64_t)arc_p >= 0);
4643 if (arc_reclaim_needed()) {
4644 cv_signal(&arc_reclaim_thread_cv);
4651 if (arc_c >= arc_c_max)
4655 * If we're within (2 * maxblocksize) bytes of the target
4656 * cache size, increment the target cache size
4658 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
4659 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
4660 atomic_add_64(&arc_c, (int64_t)bytes);
4661 if (arc_c > arc_c_max)
4663 else if (state == arc_anon)
4664 atomic_add_64(&arc_p, (int64_t)bytes);
4668 ASSERT((int64_t)arc_p >= 0);
4672 * Check if arc_size has grown past our upper threshold, determined by
4673 * zfs_arc_overflow_shift.
4676 arc_is_overflowing(void)
4678 /* Always allow at least one block of overflow */
4679 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4680 arc_c >> zfs_arc_overflow_shift);
4682 return (arc_size >= arc_c + overflow);
4686 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4688 arc_buf_contents_t type = arc_buf_type(hdr);
4690 arc_get_data_impl(hdr, size, tag);
4691 if (type == ARC_BUFC_METADATA) {
4692 return (abd_alloc(size, B_TRUE));
4694 ASSERT(type == ARC_BUFC_DATA);
4695 return (abd_alloc(size, B_FALSE));
4700 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4702 arc_buf_contents_t type = arc_buf_type(hdr);
4704 arc_get_data_impl(hdr, size, tag);
4705 if (type == ARC_BUFC_METADATA) {
4706 return (zio_buf_alloc(size));
4708 ASSERT(type == ARC_BUFC_DATA);
4709 return (zio_data_buf_alloc(size));
4714 * Allocate a block and return it to the caller. If we are hitting the
4715 * hard limit for the cache size, we must sleep, waiting for the eviction
4716 * thread to catch up. If we're past the target size but below the hard
4717 * limit, we'll only signal the reclaim thread and continue on.
4720 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4722 arc_state_t *state = hdr->b_l1hdr.b_state;
4723 arc_buf_contents_t type = arc_buf_type(hdr);
4725 arc_adapt(size, state);
4728 * If arc_size is currently overflowing, and has grown past our
4729 * upper limit, we must be adding data faster than the evict
4730 * thread can evict. Thus, to ensure we don't compound the
4731 * problem by adding more data and forcing arc_size to grow even
4732 * further past it's target size, we halt and wait for the
4733 * eviction thread to catch up.
4735 * It's also possible that the reclaim thread is unable to evict
4736 * enough buffers to get arc_size below the overflow limit (e.g.
4737 * due to buffers being un-evictable, or hash lock collisions).
4738 * In this case, we want to proceed regardless if we're
4739 * overflowing; thus we don't use a while loop here.
4741 if (arc_is_overflowing()) {
4742 mutex_enter(&arc_reclaim_lock);
4745 * Now that we've acquired the lock, we may no longer be
4746 * over the overflow limit, lets check.
4748 * We're ignoring the case of spurious wake ups. If that
4749 * were to happen, it'd let this thread consume an ARC
4750 * buffer before it should have (i.e. before we're under
4751 * the overflow limit and were signalled by the reclaim
4752 * thread). As long as that is a rare occurrence, it
4753 * shouldn't cause any harm.
4755 if (arc_is_overflowing()) {
4756 cv_signal(&arc_reclaim_thread_cv);
4757 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4760 mutex_exit(&arc_reclaim_lock);
4763 VERIFY3U(hdr->b_type, ==, type);
4764 if (type == ARC_BUFC_METADATA) {
4765 arc_space_consume(size, ARC_SPACE_META);
4767 arc_space_consume(size, ARC_SPACE_DATA);
4771 * Update the state size. Note that ghost states have a
4772 * "ghost size" and so don't need to be updated.
4774 if (!GHOST_STATE(state)) {
4776 (void) refcount_add_many(&state->arcs_size, size, tag);
4779 * If this is reached via arc_read, the link is
4780 * protected by the hash lock. If reached via
4781 * arc_buf_alloc, the header should not be accessed by
4782 * any other thread. And, if reached via arc_read_done,
4783 * the hash lock will protect it if it's found in the
4784 * hash table; otherwise no other thread should be
4785 * trying to [add|remove]_reference it.
4787 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4788 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4789 (void) refcount_add_many(&state->arcs_esize[type],
4794 * If we are growing the cache, and we are adding anonymous
4795 * data, and we have outgrown arc_p, update arc_p
4797 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
4798 (refcount_count(&arc_anon->arcs_size) +
4799 refcount_count(&arc_mru->arcs_size) > arc_p))
4800 arc_p = MIN(arc_c, arc_p + size);
4802 ARCSTAT_BUMP(arcstat_allocated);
4806 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4808 arc_free_data_impl(hdr, size, tag);
4813 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4815 arc_buf_contents_t type = arc_buf_type(hdr);
4817 arc_free_data_impl(hdr, size, tag);
4818 if (type == ARC_BUFC_METADATA) {
4819 zio_buf_free(buf, size);
4821 ASSERT(type == ARC_BUFC_DATA);
4822 zio_data_buf_free(buf, size);
4827 * Free the arc data buffer.
4830 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4832 arc_state_t *state = hdr->b_l1hdr.b_state;
4833 arc_buf_contents_t type = arc_buf_type(hdr);
4835 /* protected by hash lock, if in the hash table */
4836 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4837 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4838 ASSERT(state != arc_anon && state != arc_l2c_only);
4840 (void) refcount_remove_many(&state->arcs_esize[type],
4843 (void) refcount_remove_many(&state->arcs_size, size, tag);
4845 VERIFY3U(hdr->b_type, ==, type);
4846 if (type == ARC_BUFC_METADATA) {
4847 arc_space_return(size, ARC_SPACE_META);
4849 ASSERT(type == ARC_BUFC_DATA);
4850 arc_space_return(size, ARC_SPACE_DATA);
4855 * This routine is called whenever a buffer is accessed.
4856 * NOTE: the hash lock is dropped in this function.
4859 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4863 ASSERT(MUTEX_HELD(hash_lock));
4864 ASSERT(HDR_HAS_L1HDR(hdr));
4866 if (hdr->b_l1hdr.b_state == arc_anon) {
4868 * This buffer is not in the cache, and does not
4869 * appear in our "ghost" list. Add the new buffer
4873 ASSERT0(hdr->b_l1hdr.b_arc_access);
4874 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4875 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4876 arc_change_state(arc_mru, hdr, hash_lock);
4878 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4879 now = ddi_get_lbolt();
4882 * If this buffer is here because of a prefetch, then either:
4883 * - clear the flag if this is a "referencing" read
4884 * (any subsequent access will bump this into the MFU state).
4886 * - move the buffer to the head of the list if this is
4887 * another prefetch (to make it less likely to be evicted).
4889 if (HDR_PREFETCH(hdr)) {
4890 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4891 /* link protected by hash lock */
4892 ASSERT(multilist_link_active(
4893 &hdr->b_l1hdr.b_arc_node));
4895 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4896 ARCSTAT_BUMP(arcstat_mru_hits);
4898 hdr->b_l1hdr.b_arc_access = now;
4903 * This buffer has been "accessed" only once so far,
4904 * but it is still in the cache. Move it to the MFU
4907 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4909 * More than 125ms have passed since we
4910 * instantiated this buffer. Move it to the
4911 * most frequently used state.
4913 hdr->b_l1hdr.b_arc_access = now;
4914 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4915 arc_change_state(arc_mfu, hdr, hash_lock);
4917 ARCSTAT_BUMP(arcstat_mru_hits);
4918 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
4919 arc_state_t *new_state;
4921 * This buffer has been "accessed" recently, but
4922 * was evicted from the cache. Move it to the
4926 if (HDR_PREFETCH(hdr)) {
4927 new_state = arc_mru;
4928 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
4929 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4930 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4932 new_state = arc_mfu;
4933 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4936 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4937 arc_change_state(new_state, hdr, hash_lock);
4939 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
4940 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
4942 * This buffer has been accessed more than once and is
4943 * still in the cache. Keep it in the MFU state.
4945 * NOTE: an add_reference() that occurred when we did
4946 * the arc_read() will have kicked this off the list.
4947 * If it was a prefetch, we will explicitly move it to
4948 * the head of the list now.
4950 if ((HDR_PREFETCH(hdr)) != 0) {
4951 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4952 /* link protected by hash_lock */
4953 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4955 ARCSTAT_BUMP(arcstat_mfu_hits);
4956 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4957 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4958 arc_state_t *new_state = arc_mfu;
4960 * This buffer has been accessed more than once but has
4961 * been evicted from the cache. Move it back to the
4965 if (HDR_PREFETCH(hdr)) {
4967 * This is a prefetch access...
4968 * move this block back to the MRU state.
4970 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4971 new_state = arc_mru;
4974 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4975 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4976 arc_change_state(new_state, hdr, hash_lock);
4978 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4979 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4981 * This buffer is on the 2nd Level ARC.
4984 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4985 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4986 arc_change_state(arc_mfu, hdr, hash_lock);
4988 ASSERT(!"invalid arc state");
4992 /* a generic arc_done_func_t which you can use */
4995 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4997 if (zio == NULL || zio->io_error == 0)
4998 bcopy(buf->b_data, arg, arc_buf_size(buf));
4999 arc_buf_destroy(buf, arg);
5002 /* a generic arc_done_func_t */
5004 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
5006 arc_buf_t **bufp = arg;
5007 if (zio && zio->io_error) {
5008 arc_buf_destroy(buf, arg);
5012 ASSERT(buf->b_data);
5017 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5019 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5020 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5021 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
5023 if (HDR_COMPRESSION_ENABLED(hdr)) {
5024 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
5025 BP_GET_COMPRESS(bp));
5027 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5028 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5033 arc_read_done(zio_t *zio)
5035 arc_buf_hdr_t *hdr = zio->io_private;
5036 kmutex_t *hash_lock = NULL;
5037 arc_callback_t *callback_list;
5038 arc_callback_t *acb;
5039 boolean_t freeable = B_FALSE;
5040 boolean_t no_zio_error = (zio->io_error == 0);
5043 * The hdr was inserted into hash-table and removed from lists
5044 * prior to starting I/O. We should find this header, since
5045 * it's in the hash table, and it should be legit since it's
5046 * not possible to evict it during the I/O. The only possible
5047 * reason for it not to be found is if we were freed during the
5050 if (HDR_IN_HASH_TABLE(hdr)) {
5051 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5052 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5053 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5054 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5055 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5057 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5060 ASSERT((found == hdr &&
5061 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5062 (found == hdr && HDR_L2_READING(hdr)));
5063 ASSERT3P(hash_lock, !=, NULL);
5067 /* byteswap if necessary */
5068 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5069 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5070 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5072 hdr->b_l1hdr.b_byteswap =
5073 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5076 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5080 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5081 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5082 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5084 callback_list = hdr->b_l1hdr.b_acb;
5085 ASSERT3P(callback_list, !=, NULL);
5087 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
5089 * Only call arc_access on anonymous buffers. This is because
5090 * if we've issued an I/O for an evicted buffer, we've already
5091 * called arc_access (to prevent any simultaneous readers from
5092 * getting confused).
5094 arc_access(hdr, hash_lock);
5098 * If a read request has a callback (i.e. acb_done is not NULL), then we
5099 * make a buf containing the data according to the parameters which were
5100 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5101 * aren't needlessly decompressing the data multiple times.
5103 int callback_cnt = 0;
5104 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5108 /* This is a demand read since prefetches don't use callbacks */
5111 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
5112 acb->acb_compressed, no_zio_error, &acb->acb_buf);
5114 zio->io_error = error;
5117 hdr->b_l1hdr.b_acb = NULL;
5118 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5119 if (callback_cnt == 0) {
5120 ASSERT(HDR_PREFETCH(hdr));
5121 ASSERT0(hdr->b_l1hdr.b_bufcnt);
5122 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5125 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5126 callback_list != NULL);
5129 arc_hdr_verify(hdr, zio->io_bp);
5131 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5132 if (hdr->b_l1hdr.b_state != arc_anon)
5133 arc_change_state(arc_anon, hdr, hash_lock);
5134 if (HDR_IN_HASH_TABLE(hdr))
5135 buf_hash_remove(hdr);
5136 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5140 * Broadcast before we drop the hash_lock to avoid the possibility
5141 * that the hdr (and hence the cv) might be freed before we get to
5142 * the cv_broadcast().
5144 cv_broadcast(&hdr->b_l1hdr.b_cv);
5146 if (hash_lock != NULL) {
5147 mutex_exit(hash_lock);
5150 * This block was freed while we waited for the read to
5151 * complete. It has been removed from the hash table and
5152 * moved to the anonymous state (so that it won't show up
5155 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5156 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5159 /* execute each callback and free its structure */
5160 while ((acb = callback_list) != NULL) {
5162 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
5164 if (acb->acb_zio_dummy != NULL) {
5165 acb->acb_zio_dummy->io_error = zio->io_error;
5166 zio_nowait(acb->acb_zio_dummy);
5169 callback_list = acb->acb_next;
5170 kmem_free(acb, sizeof (arc_callback_t));
5174 arc_hdr_destroy(hdr);
5178 * "Read" the block at the specified DVA (in bp) via the
5179 * cache. If the block is found in the cache, invoke the provided
5180 * callback immediately and return. Note that the `zio' parameter
5181 * in the callback will be NULL in this case, since no IO was
5182 * required. If the block is not in the cache pass the read request
5183 * on to the spa with a substitute callback function, so that the
5184 * requested block will be added to the cache.
5186 * If a read request arrives for a block that has a read in-progress,
5187 * either wait for the in-progress read to complete (and return the
5188 * results); or, if this is a read with a "done" func, add a record
5189 * to the read to invoke the "done" func when the read completes,
5190 * and return; or just return.
5192 * arc_read_done() will invoke all the requested "done" functions
5193 * for readers of this block.
5196 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
5197 void *private, zio_priority_t priority, int zio_flags,
5198 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5200 arc_buf_hdr_t *hdr = NULL;
5201 kmutex_t *hash_lock = NULL;
5203 uint64_t guid = spa_load_guid(spa);
5204 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5206 ASSERT(!BP_IS_EMBEDDED(bp) ||
5207 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5210 if (!BP_IS_EMBEDDED(bp)) {
5212 * Embedded BP's have no DVA and require no I/O to "read".
5213 * Create an anonymous arc buf to back it.
5215 hdr = buf_hash_find(guid, bp, &hash_lock);
5218 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5219 arc_buf_t *buf = NULL;
5220 *arc_flags |= ARC_FLAG_CACHED;
5222 if (HDR_IO_IN_PROGRESS(hdr)) {
5224 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5225 priority == ZIO_PRIORITY_SYNC_READ) {
5227 * This sync read must wait for an
5228 * in-progress async read (e.g. a predictive
5229 * prefetch). Async reads are queued
5230 * separately at the vdev_queue layer, so
5231 * this is a form of priority inversion.
5232 * Ideally, we would "inherit" the demand
5233 * i/o's priority by moving the i/o from
5234 * the async queue to the synchronous queue,
5235 * but there is currently no mechanism to do
5236 * so. Track this so that we can evaluate
5237 * the magnitude of this potential performance
5240 * Note that if the prefetch i/o is already
5241 * active (has been issued to the device),
5242 * the prefetch improved performance, because
5243 * we issued it sooner than we would have
5244 * without the prefetch.
5246 DTRACE_PROBE1(arc__sync__wait__for__async,
5247 arc_buf_hdr_t *, hdr);
5248 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
5250 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5251 arc_hdr_clear_flags(hdr,
5252 ARC_FLAG_PREDICTIVE_PREFETCH);
5255 if (*arc_flags & ARC_FLAG_WAIT) {
5256 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5257 mutex_exit(hash_lock);
5260 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5263 arc_callback_t *acb = NULL;
5265 acb = kmem_zalloc(sizeof (arc_callback_t),
5267 acb->acb_done = done;
5268 acb->acb_private = private;
5269 acb->acb_compressed = compressed_read;
5271 acb->acb_zio_dummy = zio_null(pio,
5272 spa, NULL, NULL, NULL, zio_flags);
5274 ASSERT3P(acb->acb_done, !=, NULL);
5275 acb->acb_next = hdr->b_l1hdr.b_acb;
5276 hdr->b_l1hdr.b_acb = acb;
5277 mutex_exit(hash_lock);
5280 mutex_exit(hash_lock);
5284 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5285 hdr->b_l1hdr.b_state == arc_mfu);
5288 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5290 * This is a demand read which does not have to
5291 * wait for i/o because we did a predictive
5292 * prefetch i/o for it, which has completed.
5295 arc__demand__hit__predictive__prefetch,
5296 arc_buf_hdr_t *, hdr);
5298 arcstat_demand_hit_predictive_prefetch);
5299 arc_hdr_clear_flags(hdr,
5300 ARC_FLAG_PREDICTIVE_PREFETCH);
5302 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5304 /* Get a buf with the desired data in it. */
5305 VERIFY0(arc_buf_alloc_impl(hdr, private,
5306 compressed_read, B_TRUE, &buf));
5307 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5308 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5309 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5311 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5312 arc_access(hdr, hash_lock);
5313 if (*arc_flags & ARC_FLAG_L2CACHE)
5314 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5315 mutex_exit(hash_lock);
5316 ARCSTAT_BUMP(arcstat_hits);
5317 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5318 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5319 data, metadata, hits);
5322 done(NULL, buf, private);
5324 uint64_t lsize = BP_GET_LSIZE(bp);
5325 uint64_t psize = BP_GET_PSIZE(bp);
5326 arc_callback_t *acb;
5329 boolean_t devw = B_FALSE;
5333 /* this block is not in the cache */
5334 arc_buf_hdr_t *exists = NULL;
5335 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5336 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5337 BP_GET_COMPRESS(bp), type);
5339 if (!BP_IS_EMBEDDED(bp)) {
5340 hdr->b_dva = *BP_IDENTITY(bp);
5341 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5342 exists = buf_hash_insert(hdr, &hash_lock);
5344 if (exists != NULL) {
5345 /* somebody beat us to the hash insert */
5346 mutex_exit(hash_lock);
5347 buf_discard_identity(hdr);
5348 arc_hdr_destroy(hdr);
5349 goto top; /* restart the IO request */
5353 * This block is in the ghost cache. If it was L2-only
5354 * (and thus didn't have an L1 hdr), we realloc the
5355 * header to add an L1 hdr.
5357 if (!HDR_HAS_L1HDR(hdr)) {
5358 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5361 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5362 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5363 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5364 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5365 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5366 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5369 * This is a delicate dance that we play here.
5370 * This hdr is in the ghost list so we access it
5371 * to move it out of the ghost list before we
5372 * initiate the read. If it's a prefetch then
5373 * it won't have a callback so we'll remove the
5374 * reference that arc_buf_alloc_impl() created. We
5375 * do this after we've called arc_access() to
5376 * avoid hitting an assert in remove_reference().
5378 arc_access(hdr, hash_lock);
5379 arc_hdr_alloc_pabd(hdr);
5381 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5382 size = arc_hdr_size(hdr);
5385 * If compression is enabled on the hdr, then will do
5386 * RAW I/O and will store the compressed data in the hdr's
5387 * data block. Otherwise, the hdr's data block will contain
5388 * the uncompressed data.
5390 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5391 zio_flags |= ZIO_FLAG_RAW;
5394 if (*arc_flags & ARC_FLAG_PREFETCH)
5395 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5396 if (*arc_flags & ARC_FLAG_L2CACHE)
5397 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5398 if (BP_GET_LEVEL(bp) > 0)
5399 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5400 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5401 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5402 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5404 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5405 acb->acb_done = done;
5406 acb->acb_private = private;
5407 acb->acb_compressed = compressed_read;
5409 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5410 hdr->b_l1hdr.b_acb = acb;
5411 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5413 if (HDR_HAS_L2HDR(hdr) &&
5414 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5415 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5416 addr = hdr->b_l2hdr.b_daddr;
5418 * Lock out L2ARC device removal.
5420 if (vdev_is_dead(vd) ||
5421 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5425 if (priority == ZIO_PRIORITY_ASYNC_READ)
5426 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5428 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5430 if (hash_lock != NULL)
5431 mutex_exit(hash_lock);
5434 * At this point, we have a level 1 cache miss. Try again in
5435 * L2ARC if possible.
5437 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5439 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5440 uint64_t, lsize, zbookmark_phys_t *, zb);
5441 ARCSTAT_BUMP(arcstat_misses);
5442 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5443 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5444 data, metadata, misses);
5449 racct_add_force(curproc, RACCT_READBPS, size);
5450 racct_add_force(curproc, RACCT_READIOPS, 1);
5451 PROC_UNLOCK(curproc);
5454 curthread->td_ru.ru_inblock++;
5457 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5459 * Read from the L2ARC if the following are true:
5460 * 1. The L2ARC vdev was previously cached.
5461 * 2. This buffer still has L2ARC metadata.
5462 * 3. This buffer isn't currently writing to the L2ARC.
5463 * 4. The L2ARC entry wasn't evicted, which may
5464 * also have invalidated the vdev.
5465 * 5. This isn't prefetch and l2arc_noprefetch is set.
5467 if (HDR_HAS_L2HDR(hdr) &&
5468 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5469 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5470 l2arc_read_callback_t *cb;
5474 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5475 ARCSTAT_BUMP(arcstat_l2_hits);
5477 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5479 cb->l2rcb_hdr = hdr;
5482 cb->l2rcb_flags = zio_flags;
5484 asize = vdev_psize_to_asize(vd, size);
5485 if (asize != size) {
5486 abd = abd_alloc_for_io(asize,
5487 HDR_ISTYPE_METADATA(hdr));
5488 cb->l2rcb_abd = abd;
5490 abd = hdr->b_l1hdr.b_pabd;
5493 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5494 addr + asize <= vd->vdev_psize -
5495 VDEV_LABEL_END_SIZE);
5498 * l2arc read. The SCL_L2ARC lock will be
5499 * released by l2arc_read_done().
5500 * Issue a null zio if the underlying buffer
5501 * was squashed to zero size by compression.
5503 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5504 ZIO_COMPRESS_EMPTY);
5505 rzio = zio_read_phys(pio, vd, addr,
5508 l2arc_read_done, cb, priority,
5509 zio_flags | ZIO_FLAG_DONT_CACHE |
5511 ZIO_FLAG_DONT_PROPAGATE |
5512 ZIO_FLAG_DONT_RETRY, B_FALSE);
5513 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5515 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5517 if (*arc_flags & ARC_FLAG_NOWAIT) {
5522 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5523 if (zio_wait(rzio) == 0)
5526 /* l2arc read error; goto zio_read() */
5528 DTRACE_PROBE1(l2arc__miss,
5529 arc_buf_hdr_t *, hdr);
5530 ARCSTAT_BUMP(arcstat_l2_misses);
5531 if (HDR_L2_WRITING(hdr))
5532 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5533 spa_config_exit(spa, SCL_L2ARC, vd);
5537 spa_config_exit(spa, SCL_L2ARC, vd);
5538 if (l2arc_ndev != 0) {
5539 DTRACE_PROBE1(l2arc__miss,
5540 arc_buf_hdr_t *, hdr);
5541 ARCSTAT_BUMP(arcstat_l2_misses);
5545 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
5546 arc_read_done, hdr, priority, zio_flags, zb);
5548 if (*arc_flags & ARC_FLAG_WAIT)
5549 return (zio_wait(rzio));
5551 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5558 * Notify the arc that a block was freed, and thus will never be used again.
5561 arc_freed(spa_t *spa, const blkptr_t *bp)
5564 kmutex_t *hash_lock;
5565 uint64_t guid = spa_load_guid(spa);
5567 ASSERT(!BP_IS_EMBEDDED(bp));
5569 hdr = buf_hash_find(guid, bp, &hash_lock);
5574 * We might be trying to free a block that is still doing I/O
5575 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5576 * dmu_sync-ed block). If this block is being prefetched, then it
5577 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5578 * until the I/O completes. A block may also have a reference if it is
5579 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5580 * have written the new block to its final resting place on disk but
5581 * without the dedup flag set. This would have left the hdr in the MRU
5582 * state and discoverable. When the txg finally syncs it detects that
5583 * the block was overridden in open context and issues an override I/O.
5584 * Since this is a dedup block, the override I/O will determine if the
5585 * block is already in the DDT. If so, then it will replace the io_bp
5586 * with the bp from the DDT and allow the I/O to finish. When the I/O
5587 * reaches the done callback, dbuf_write_override_done, it will
5588 * check to see if the io_bp and io_bp_override are identical.
5589 * If they are not, then it indicates that the bp was replaced with
5590 * the bp in the DDT and the override bp is freed. This allows
5591 * us to arrive here with a reference on a block that is being
5592 * freed. So if we have an I/O in progress, or a reference to
5593 * this hdr, then we don't destroy the hdr.
5595 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5596 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5597 arc_change_state(arc_anon, hdr, hash_lock);
5598 arc_hdr_destroy(hdr);
5599 mutex_exit(hash_lock);
5601 mutex_exit(hash_lock);
5607 * Release this buffer from the cache, making it an anonymous buffer. This
5608 * must be done after a read and prior to modifying the buffer contents.
5609 * If the buffer has more than one reference, we must make
5610 * a new hdr for the buffer.
5613 arc_release(arc_buf_t *buf, void *tag)
5615 arc_buf_hdr_t *hdr = buf->b_hdr;
5618 * It would be nice to assert that if it's DMU metadata (level >
5619 * 0 || it's the dnode file), then it must be syncing context.
5620 * But we don't know that information at this level.
5623 mutex_enter(&buf->b_evict_lock);
5625 ASSERT(HDR_HAS_L1HDR(hdr));
5628 * We don't grab the hash lock prior to this check, because if
5629 * the buffer's header is in the arc_anon state, it won't be
5630 * linked into the hash table.
5632 if (hdr->b_l1hdr.b_state == arc_anon) {
5633 mutex_exit(&buf->b_evict_lock);
5634 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5635 ASSERT(!HDR_IN_HASH_TABLE(hdr));
5636 ASSERT(!HDR_HAS_L2HDR(hdr));
5637 ASSERT(HDR_EMPTY(hdr));
5638 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5639 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5640 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5642 hdr->b_l1hdr.b_arc_access = 0;
5645 * If the buf is being overridden then it may already
5646 * have a hdr that is not empty.
5648 buf_discard_identity(hdr);
5654 kmutex_t *hash_lock = HDR_LOCK(hdr);
5655 mutex_enter(hash_lock);
5658 * This assignment is only valid as long as the hash_lock is
5659 * held, we must be careful not to reference state or the
5660 * b_state field after dropping the lock.
5662 arc_state_t *state = hdr->b_l1hdr.b_state;
5663 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5664 ASSERT3P(state, !=, arc_anon);
5666 /* this buffer is not on any list */
5667 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
5669 if (HDR_HAS_L2HDR(hdr)) {
5670 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5673 * We have to recheck this conditional again now that
5674 * we're holding the l2ad_mtx to prevent a race with
5675 * another thread which might be concurrently calling
5676 * l2arc_evict(). In that case, l2arc_evict() might have
5677 * destroyed the header's L2 portion as we were waiting
5678 * to acquire the l2ad_mtx.
5680 if (HDR_HAS_L2HDR(hdr)) {
5682 arc_hdr_l2hdr_destroy(hdr);
5685 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5689 * Do we have more than one buf?
5691 if (hdr->b_l1hdr.b_bufcnt > 1) {
5692 arc_buf_hdr_t *nhdr;
5693 uint64_t spa = hdr->b_spa;
5694 uint64_t psize = HDR_GET_PSIZE(hdr);
5695 uint64_t lsize = HDR_GET_LSIZE(hdr);
5696 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5697 arc_buf_contents_t type = arc_buf_type(hdr);
5698 VERIFY3U(hdr->b_type, ==, type);
5700 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5701 (void) remove_reference(hdr, hash_lock, tag);
5703 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
5704 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5705 ASSERT(ARC_BUF_LAST(buf));
5709 * Pull the data off of this hdr and attach it to
5710 * a new anonymous hdr. Also find the last buffer
5711 * in the hdr's buffer list.
5713 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
5714 ASSERT3P(lastbuf, !=, NULL);
5717 * If the current arc_buf_t and the hdr are sharing their data
5718 * buffer, then we must stop sharing that block.
5720 if (arc_buf_is_shared(buf)) {
5721 VERIFY(!arc_buf_is_shared(lastbuf));
5724 * First, sever the block sharing relationship between
5725 * buf and the arc_buf_hdr_t.
5727 arc_unshare_buf(hdr, buf);
5730 * Now we need to recreate the hdr's b_pabd. Since we
5731 * have lastbuf handy, we try to share with it, but if
5732 * we can't then we allocate a new b_pabd and copy the
5733 * data from buf into it.
5735 if (arc_can_share(hdr, lastbuf)) {
5736 arc_share_buf(hdr, lastbuf);
5738 arc_hdr_alloc_pabd(hdr);
5739 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
5740 buf->b_data, psize);
5742 VERIFY3P(lastbuf->b_data, !=, NULL);
5743 } else if (HDR_SHARED_DATA(hdr)) {
5745 * Uncompressed shared buffers are always at the end
5746 * of the list. Compressed buffers don't have the
5747 * same requirements. This makes it hard to
5748 * simply assert that the lastbuf is shared so
5749 * we rely on the hdr's compression flags to determine
5750 * if we have a compressed, shared buffer.
5752 ASSERT(arc_buf_is_shared(lastbuf) ||
5753 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
5754 ASSERT(!ARC_BUF_SHARED(buf));
5756 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5757 ASSERT3P(state, !=, arc_l2c_only);
5759 (void) refcount_remove_many(&state->arcs_size,
5760 arc_buf_size(buf), buf);
5762 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5763 ASSERT3P(state, !=, arc_l2c_only);
5764 (void) refcount_remove_many(&state->arcs_esize[type],
5765 arc_buf_size(buf), buf);
5768 hdr->b_l1hdr.b_bufcnt -= 1;
5769 arc_cksum_verify(buf);
5771 arc_buf_unwatch(buf);
5774 mutex_exit(hash_lock);
5777 * Allocate a new hdr. The new hdr will contain a b_pabd
5778 * buffer which will be freed in arc_write().
5780 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5781 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5782 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5783 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5784 VERIFY3U(nhdr->b_type, ==, type);
5785 ASSERT(!HDR_SHARED_DATA(nhdr));
5787 nhdr->b_l1hdr.b_buf = buf;
5788 nhdr->b_l1hdr.b_bufcnt = 1;
5789 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5792 mutex_exit(&buf->b_evict_lock);
5793 (void) refcount_add_many(&arc_anon->arcs_size,
5794 arc_buf_size(buf), buf);
5796 mutex_exit(&buf->b_evict_lock);
5797 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5798 /* protected by hash lock, or hdr is on arc_anon */
5799 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5800 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5801 arc_change_state(arc_anon, hdr, hash_lock);
5802 hdr->b_l1hdr.b_arc_access = 0;
5803 mutex_exit(hash_lock);
5805 buf_discard_identity(hdr);
5811 arc_released(arc_buf_t *buf)
5815 mutex_enter(&buf->b_evict_lock);
5816 released = (buf->b_data != NULL &&
5817 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5818 mutex_exit(&buf->b_evict_lock);
5824 arc_referenced(arc_buf_t *buf)
5828 mutex_enter(&buf->b_evict_lock);
5829 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5830 mutex_exit(&buf->b_evict_lock);
5831 return (referenced);
5836 arc_write_ready(zio_t *zio)
5838 arc_write_callback_t *callback = zio->io_private;
5839 arc_buf_t *buf = callback->awcb_buf;
5840 arc_buf_hdr_t *hdr = buf->b_hdr;
5841 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5843 ASSERT(HDR_HAS_L1HDR(hdr));
5844 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5845 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5848 * If we're reexecuting this zio because the pool suspended, then
5849 * cleanup any state that was previously set the first time the
5850 * callback was invoked.
5852 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5853 arc_cksum_free(hdr);
5855 arc_buf_unwatch(buf);
5857 if (hdr->b_l1hdr.b_pabd != NULL) {
5858 if (arc_buf_is_shared(buf)) {
5859 arc_unshare_buf(hdr, buf);
5861 arc_hdr_free_pabd(hdr);
5865 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5866 ASSERT(!HDR_SHARED_DATA(hdr));
5867 ASSERT(!arc_buf_is_shared(buf));
5869 callback->awcb_ready(zio, buf, callback->awcb_private);
5871 if (HDR_IO_IN_PROGRESS(hdr))
5872 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
5874 arc_cksum_compute(buf);
5875 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5877 enum zio_compress compress;
5878 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5879 compress = ZIO_COMPRESS_OFF;
5881 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
5882 compress = BP_GET_COMPRESS(zio->io_bp);
5884 HDR_SET_PSIZE(hdr, psize);
5885 arc_hdr_set_compress(hdr, compress);
5889 * Fill the hdr with data. If the hdr is compressed, the data we want
5890 * is available from the zio, otherwise we can take it from the buf.
5892 * We might be able to share the buf's data with the hdr here. However,
5893 * doing so would cause the ARC to be full of linear ABDs if we write a
5894 * lot of shareable data. As a compromise, we check whether scattered
5895 * ABDs are allowed, and assume that if they are then the user wants
5896 * the ARC to be primarily filled with them regardless of the data being
5897 * written. Therefore, if they're allowed then we allocate one and copy
5898 * the data into it; otherwise, we share the data directly if we can.
5900 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
5901 arc_hdr_alloc_pabd(hdr);
5904 * Ideally, we would always copy the io_abd into b_pabd, but the
5905 * user may have disabled compressed ARC, thus we must check the
5906 * hdr's compression setting rather than the io_bp's.
5908 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5909 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
5911 ASSERT3U(psize, >, 0);
5913 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
5915 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
5917 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
5921 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
5922 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
5923 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5925 arc_share_buf(hdr, buf);
5928 arc_hdr_verify(hdr, zio->io_bp);
5932 arc_write_children_ready(zio_t *zio)
5934 arc_write_callback_t *callback = zio->io_private;
5935 arc_buf_t *buf = callback->awcb_buf;
5937 callback->awcb_children_ready(zio, buf, callback->awcb_private);
5941 * The SPA calls this callback for each physical write that happens on behalf
5942 * of a logical write. See the comment in dbuf_write_physdone() for details.
5945 arc_write_physdone(zio_t *zio)
5947 arc_write_callback_t *cb = zio->io_private;
5948 if (cb->awcb_physdone != NULL)
5949 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
5953 arc_write_done(zio_t *zio)
5955 arc_write_callback_t *callback = zio->io_private;
5956 arc_buf_t *buf = callback->awcb_buf;
5957 arc_buf_hdr_t *hdr = buf->b_hdr;
5959 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5961 if (zio->io_error == 0) {
5962 arc_hdr_verify(hdr, zio->io_bp);
5964 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5965 buf_discard_identity(hdr);
5967 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
5968 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
5971 ASSERT(HDR_EMPTY(hdr));
5975 * If the block to be written was all-zero or compressed enough to be
5976 * embedded in the BP, no write was performed so there will be no
5977 * dva/birth/checksum. The buffer must therefore remain anonymous
5980 if (!HDR_EMPTY(hdr)) {
5981 arc_buf_hdr_t *exists;
5982 kmutex_t *hash_lock;
5984 ASSERT3U(zio->io_error, ==, 0);
5986 arc_cksum_verify(buf);
5988 exists = buf_hash_insert(hdr, &hash_lock);
5989 if (exists != NULL) {
5991 * This can only happen if we overwrite for
5992 * sync-to-convergence, because we remove
5993 * buffers from the hash table when we arc_free().
5995 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5996 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5997 panic("bad overwrite, hdr=%p exists=%p",
5998 (void *)hdr, (void *)exists);
5999 ASSERT(refcount_is_zero(
6000 &exists->b_l1hdr.b_refcnt));
6001 arc_change_state(arc_anon, exists, hash_lock);
6002 mutex_exit(hash_lock);
6003 arc_hdr_destroy(exists);
6004 exists = buf_hash_insert(hdr, &hash_lock);
6005 ASSERT3P(exists, ==, NULL);
6006 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6008 ASSERT(zio->io_prop.zp_nopwrite);
6009 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6010 panic("bad nopwrite, hdr=%p exists=%p",
6011 (void *)hdr, (void *)exists);
6014 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6015 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6016 ASSERT(BP_GET_DEDUP(zio->io_bp));
6017 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6020 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6021 /* if it's not anon, we are doing a scrub */
6022 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6023 arc_access(hdr, hash_lock);
6024 mutex_exit(hash_lock);
6026 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6029 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6030 callback->awcb_done(zio, buf, callback->awcb_private);
6032 abd_put(zio->io_abd);
6033 kmem_free(callback, sizeof (arc_write_callback_t));
6037 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6038 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
6039 arc_done_func_t *children_ready, arc_done_func_t *physdone,
6040 arc_done_func_t *done, void *private, zio_priority_t priority,
6041 int zio_flags, const zbookmark_phys_t *zb)
6043 arc_buf_hdr_t *hdr = buf->b_hdr;
6044 arc_write_callback_t *callback;
6046 zio_prop_t localprop = *zp;
6048 ASSERT3P(ready, !=, NULL);
6049 ASSERT3P(done, !=, NULL);
6050 ASSERT(!HDR_IO_ERROR(hdr));
6051 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6052 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6053 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6055 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6056 if (ARC_BUF_COMPRESSED(buf)) {
6058 * We're writing a pre-compressed buffer. Make the
6059 * compression algorithm requested by the zio_prop_t match
6060 * the pre-compressed buffer's compression algorithm.
6062 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6064 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6065 zio_flags |= ZIO_FLAG_RAW;
6067 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6068 callback->awcb_ready = ready;
6069 callback->awcb_children_ready = children_ready;
6070 callback->awcb_physdone = physdone;
6071 callback->awcb_done = done;
6072 callback->awcb_private = private;
6073 callback->awcb_buf = buf;
6076 * The hdr's b_pabd is now stale, free it now. A new data block
6077 * will be allocated when the zio pipeline calls arc_write_ready().
6079 if (hdr->b_l1hdr.b_pabd != NULL) {
6081 * If the buf is currently sharing the data block with
6082 * the hdr then we need to break that relationship here.
6083 * The hdr will remain with a NULL data pointer and the
6084 * buf will take sole ownership of the block.
6086 if (arc_buf_is_shared(buf)) {
6087 arc_unshare_buf(hdr, buf);
6089 arc_hdr_free_pabd(hdr);
6091 VERIFY3P(buf->b_data, !=, NULL);
6092 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6094 ASSERT(!arc_buf_is_shared(buf));
6095 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6097 zio = zio_write(pio, spa, txg, bp,
6098 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6099 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6100 (children_ready != NULL) ? arc_write_children_ready : NULL,
6101 arc_write_physdone, arc_write_done, callback,
6102 priority, zio_flags, zb);
6108 arc_memory_throttle(uint64_t reserve, uint64_t txg)
6111 uint64_t available_memory = ptob(freemem);
6112 static uint64_t page_load = 0;
6113 static uint64_t last_txg = 0;
6115 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
6116 available_memory = MIN(available_memory, uma_avail());
6119 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
6122 if (txg > last_txg) {
6127 * If we are in pageout, we know that memory is already tight,
6128 * the arc is already going to be evicting, so we just want to
6129 * continue to let page writes occur as quickly as possible.
6131 if (curproc == pageproc) {
6132 if (page_load > MAX(ptob(minfree), available_memory) / 4)
6133 return (SET_ERROR(ERESTART));
6134 /* Note: reserve is inflated, so we deflate */
6135 page_load += reserve / 8;
6137 } else if (page_load > 0 && arc_reclaim_needed()) {
6138 /* memory is low, delay before restarting */
6139 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6140 return (SET_ERROR(EAGAIN));
6148 arc_tempreserve_clear(uint64_t reserve)
6150 atomic_add_64(&arc_tempreserve, -reserve);
6151 ASSERT((int64_t)arc_tempreserve >= 0);
6155 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
6160 if (reserve > arc_c/4 && !arc_no_grow) {
6161 arc_c = MIN(arc_c_max, reserve * 4);
6162 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
6164 if (reserve > arc_c)
6165 return (SET_ERROR(ENOMEM));
6168 * Don't count loaned bufs as in flight dirty data to prevent long
6169 * network delays from blocking transactions that are ready to be
6170 * assigned to a txg.
6173 /* assert that it has not wrapped around */
6174 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6176 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6177 arc_loaned_bytes), 0);
6180 * Writes will, almost always, require additional memory allocations
6181 * in order to compress/encrypt/etc the data. We therefore need to
6182 * make sure that there is sufficient available memory for this.
6184 error = arc_memory_throttle(reserve, txg);
6189 * Throttle writes when the amount of dirty data in the cache
6190 * gets too large. We try to keep the cache less than half full
6191 * of dirty blocks so that our sync times don't grow too large.
6192 * Note: if two requests come in concurrently, we might let them
6193 * both succeed, when one of them should fail. Not a huge deal.
6196 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
6197 anon_size > arc_c / 4) {
6198 uint64_t meta_esize =
6199 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6200 uint64_t data_esize =
6201 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6202 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6203 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6204 arc_tempreserve >> 10, meta_esize >> 10,
6205 data_esize >> 10, reserve >> 10, arc_c >> 10);
6206 return (SET_ERROR(ERESTART));
6208 atomic_add_64(&arc_tempreserve, reserve);
6213 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6214 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6216 size->value.ui64 = refcount_count(&state->arcs_size);
6217 evict_data->value.ui64 =
6218 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6219 evict_metadata->value.ui64 =
6220 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6224 arc_kstat_update(kstat_t *ksp, int rw)
6226 arc_stats_t *as = ksp->ks_data;
6228 if (rw == KSTAT_WRITE) {
6231 arc_kstat_update_state(arc_anon,
6232 &as->arcstat_anon_size,
6233 &as->arcstat_anon_evictable_data,
6234 &as->arcstat_anon_evictable_metadata);
6235 arc_kstat_update_state(arc_mru,
6236 &as->arcstat_mru_size,
6237 &as->arcstat_mru_evictable_data,
6238 &as->arcstat_mru_evictable_metadata);
6239 arc_kstat_update_state(arc_mru_ghost,
6240 &as->arcstat_mru_ghost_size,
6241 &as->arcstat_mru_ghost_evictable_data,
6242 &as->arcstat_mru_ghost_evictable_metadata);
6243 arc_kstat_update_state(arc_mfu,
6244 &as->arcstat_mfu_size,
6245 &as->arcstat_mfu_evictable_data,
6246 &as->arcstat_mfu_evictable_metadata);
6247 arc_kstat_update_state(arc_mfu_ghost,
6248 &as->arcstat_mfu_ghost_size,
6249 &as->arcstat_mfu_ghost_evictable_data,
6250 &as->arcstat_mfu_ghost_evictable_metadata);
6257 * This function *must* return indices evenly distributed between all
6258 * sublists of the multilist. This is needed due to how the ARC eviction
6259 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6260 * distributed between all sublists and uses this assumption when
6261 * deciding which sublist to evict from and how much to evict from it.
6264 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6266 arc_buf_hdr_t *hdr = obj;
6269 * We rely on b_dva to generate evenly distributed index
6270 * numbers using buf_hash below. So, as an added precaution,
6271 * let's make sure we never add empty buffers to the arc lists.
6273 ASSERT(!HDR_EMPTY(hdr));
6276 * The assumption here, is the hash value for a given
6277 * arc_buf_hdr_t will remain constant throughout it's lifetime
6278 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6279 * Thus, we don't need to store the header's sublist index
6280 * on insertion, as this index can be recalculated on removal.
6282 * Also, the low order bits of the hash value are thought to be
6283 * distributed evenly. Otherwise, in the case that the multilist
6284 * has a power of two number of sublists, each sublists' usage
6285 * would not be evenly distributed.
6287 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6288 multilist_get_num_sublists(ml));
6292 static eventhandler_tag arc_event_lowmem = NULL;
6295 arc_lowmem(void *arg __unused, int howto __unused)
6298 mutex_enter(&arc_reclaim_lock);
6299 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE);
6300 cv_signal(&arc_reclaim_thread_cv);
6303 * It is unsafe to block here in arbitrary threads, because we can come
6304 * here from ARC itself and may hold ARC locks and thus risk a deadlock
6305 * with ARC reclaim thread.
6307 if (curproc == pageproc)
6308 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
6309 mutex_exit(&arc_reclaim_lock);
6314 arc_state_init(void)
6316 arc_anon = &ARC_anon;
6318 arc_mru_ghost = &ARC_mru_ghost;
6320 arc_mfu_ghost = &ARC_mfu_ghost;
6321 arc_l2c_only = &ARC_l2c_only;
6323 arc_mru->arcs_list[ARC_BUFC_METADATA] =
6324 multilist_create(sizeof (arc_buf_hdr_t),
6325 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6326 arc_state_multilist_index_func);
6327 arc_mru->arcs_list[ARC_BUFC_DATA] =
6328 multilist_create(sizeof (arc_buf_hdr_t),
6329 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6330 arc_state_multilist_index_func);
6331 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6332 multilist_create(sizeof (arc_buf_hdr_t),
6333 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6334 arc_state_multilist_index_func);
6335 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6336 multilist_create(sizeof (arc_buf_hdr_t),
6337 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6338 arc_state_multilist_index_func);
6339 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6340 multilist_create(sizeof (arc_buf_hdr_t),
6341 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6342 arc_state_multilist_index_func);
6343 arc_mfu->arcs_list[ARC_BUFC_DATA] =
6344 multilist_create(sizeof (arc_buf_hdr_t),
6345 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6346 arc_state_multilist_index_func);
6347 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6348 multilist_create(sizeof (arc_buf_hdr_t),
6349 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6350 arc_state_multilist_index_func);
6351 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6352 multilist_create(sizeof (arc_buf_hdr_t),
6353 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6354 arc_state_multilist_index_func);
6355 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6356 multilist_create(sizeof (arc_buf_hdr_t),
6357 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6358 arc_state_multilist_index_func);
6359 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6360 multilist_create(sizeof (arc_buf_hdr_t),
6361 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6362 arc_state_multilist_index_func);
6364 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6365 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6366 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6367 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6368 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6369 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6370 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6371 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6372 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6373 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6374 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6375 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6377 refcount_create(&arc_anon->arcs_size);
6378 refcount_create(&arc_mru->arcs_size);
6379 refcount_create(&arc_mru_ghost->arcs_size);
6380 refcount_create(&arc_mfu->arcs_size);
6381 refcount_create(&arc_mfu_ghost->arcs_size);
6382 refcount_create(&arc_l2c_only->arcs_size);
6386 arc_state_fini(void)
6388 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6389 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6390 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6391 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6392 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6393 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6394 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6395 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6396 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6397 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6398 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6399 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6401 refcount_destroy(&arc_anon->arcs_size);
6402 refcount_destroy(&arc_mru->arcs_size);
6403 refcount_destroy(&arc_mru_ghost->arcs_size);
6404 refcount_destroy(&arc_mfu->arcs_size);
6405 refcount_destroy(&arc_mfu_ghost->arcs_size);
6406 refcount_destroy(&arc_l2c_only->arcs_size);
6408 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6409 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6410 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6411 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6412 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6413 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6414 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6415 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6427 int i, prefetch_tunable_set = 0;
6430 * allmem is "all memory that we could possibly use".
6434 uint64_t allmem = ptob(physmem - swapfs_minfree);
6436 uint64_t allmem = (physmem * PAGESIZE) / 2;
6439 uint64_t allmem = kmem_size();
6443 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
6444 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
6445 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
6447 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
6448 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
6450 /* Convert seconds to clock ticks */
6451 arc_min_prefetch_lifespan = 1 * hz;
6453 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
6454 arc_c_min = MAX(allmem / 32, arc_abs_min);
6455 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */
6456 if (allmem >= 1 << 30)
6457 arc_c_max = allmem - (1 << 30);
6459 arc_c_max = arc_c_min;
6460 arc_c_max = MAX(allmem * 5 / 8, arc_c_max);
6463 * In userland, there's only the memory pressure that we artificially
6464 * create (see arc_available_memory()). Don't let arc_c get too
6465 * small, because it can cause transactions to be larger than
6466 * arc_c, causing arc_tempreserve_space() to fail.
6469 arc_c_min = arc_c_max / 2;
6474 * Allow the tunables to override our calculations if they are
6477 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) {
6478 arc_c_max = zfs_arc_max;
6479 arc_c_min = MIN(arc_c_min, arc_c_max);
6481 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
6482 arc_c_min = zfs_arc_min;
6486 arc_p = (arc_c >> 1);
6489 /* limit meta-data to 1/4 of the arc capacity */
6490 arc_meta_limit = arc_c_max / 4;
6494 * Metadata is stored in the kernel's heap. Don't let us
6495 * use more than half the heap for the ARC.
6498 arc_meta_limit = MIN(arc_meta_limit, uma_limit() / 2);
6500 arc_meta_limit = MIN(arc_meta_limit,
6501 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6505 /* Allow the tunable to override if it is reasonable */
6506 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6507 arc_meta_limit = zfs_arc_meta_limit;
6509 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6510 arc_c_min = arc_meta_limit / 2;
6512 if (zfs_arc_meta_min > 0) {
6513 arc_meta_min = zfs_arc_meta_min;
6515 arc_meta_min = arc_c_min / 2;
6518 if (zfs_arc_grow_retry > 0)
6519 arc_grow_retry = zfs_arc_grow_retry;
6521 if (zfs_arc_shrink_shift > 0)
6522 arc_shrink_shift = zfs_arc_shrink_shift;
6524 if (zfs_arc_no_grow_shift > 0)
6525 arc_no_grow_shift = zfs_arc_no_grow_shift;
6527 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6529 if (arc_no_grow_shift >= arc_shrink_shift)
6530 arc_no_grow_shift = arc_shrink_shift - 1;
6532 if (zfs_arc_p_min_shift > 0)
6533 arc_p_min_shift = zfs_arc_p_min_shift;
6535 /* if kmem_flags are set, lets try to use less memory */
6536 if (kmem_debugging())
6538 if (arc_c < arc_c_min)
6541 zfs_arc_min = arc_c_min;
6542 zfs_arc_max = arc_c_max;
6547 arc_reclaim_thread_exit = B_FALSE;
6548 arc_dnlc_evicts_thread_exit = FALSE;
6550 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6551 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6553 if (arc_ksp != NULL) {
6554 arc_ksp->ks_data = &arc_stats;
6555 arc_ksp->ks_update = arc_kstat_update;
6556 kstat_install(arc_ksp);
6559 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6560 TS_RUN, minclsyspri);
6563 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
6564 EVENTHANDLER_PRI_FIRST);
6567 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
6568 TS_RUN, minclsyspri);
6574 * Calculate maximum amount of dirty data per pool.
6576 * If it has been set by /etc/system, take that.
6577 * Otherwise, use a percentage of physical memory defined by
6578 * zfs_dirty_data_max_percent (default 10%) with a cap at
6579 * zfs_dirty_data_max_max (default 4GB).
6581 if (zfs_dirty_data_max == 0) {
6582 zfs_dirty_data_max = ptob(physmem) *
6583 zfs_dirty_data_max_percent / 100;
6584 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6585 zfs_dirty_data_max_max);
6589 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
6590 prefetch_tunable_set = 1;
6593 if (prefetch_tunable_set == 0) {
6594 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
6596 printf(" add \"vfs.zfs.prefetch_disable=0\" "
6597 "to /boot/loader.conf.\n");
6598 zfs_prefetch_disable = 1;
6601 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
6602 prefetch_tunable_set == 0) {
6603 printf("ZFS NOTICE: Prefetch is disabled by default if less "
6604 "than 4GB of RAM is present;\n"
6605 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
6606 "to /boot/loader.conf.\n");
6607 zfs_prefetch_disable = 1;
6610 /* Warn about ZFS memory and address space requirements. */
6611 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
6612 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
6613 "expect unstable behavior.\n");
6615 if (allmem < 512 * (1 << 20)) {
6616 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
6617 "expect unstable behavior.\n");
6618 printf(" Consider tuning vm.kmem_size and "
6619 "vm.kmem_size_max\n");
6620 printf(" in /boot/loader.conf.\n");
6629 if (arc_event_lowmem != NULL)
6630 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
6633 mutex_enter(&arc_reclaim_lock);
6634 arc_reclaim_thread_exit = B_TRUE;
6636 * The reclaim thread will set arc_reclaim_thread_exit back to
6637 * B_FALSE when it is finished exiting; we're waiting for that.
6639 while (arc_reclaim_thread_exit) {
6640 cv_signal(&arc_reclaim_thread_cv);
6641 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6643 mutex_exit(&arc_reclaim_lock);
6645 /* Use B_TRUE to ensure *all* buffers are evicted */
6646 arc_flush(NULL, B_TRUE);
6648 mutex_enter(&arc_dnlc_evicts_lock);
6649 arc_dnlc_evicts_thread_exit = TRUE;
6651 * The user evicts thread will set arc_user_evicts_thread_exit
6652 * to FALSE when it is finished exiting; we're waiting for that.
6654 while (arc_dnlc_evicts_thread_exit) {
6655 cv_signal(&arc_dnlc_evicts_cv);
6656 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
6658 mutex_exit(&arc_dnlc_evicts_lock);
6662 if (arc_ksp != NULL) {
6663 kstat_delete(arc_ksp);
6667 mutex_destroy(&arc_reclaim_lock);
6668 cv_destroy(&arc_reclaim_thread_cv);
6669 cv_destroy(&arc_reclaim_waiters_cv);
6671 mutex_destroy(&arc_dnlc_evicts_lock);
6672 cv_destroy(&arc_dnlc_evicts_cv);
6677 ASSERT0(arc_loaned_bytes);
6683 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6684 * It uses dedicated storage devices to hold cached data, which are populated
6685 * using large infrequent writes. The main role of this cache is to boost
6686 * the performance of random read workloads. The intended L2ARC devices
6687 * include short-stroked disks, solid state disks, and other media with
6688 * substantially faster read latency than disk.
6690 * +-----------------------+
6692 * +-----------------------+
6695 * l2arc_feed_thread() arc_read()
6699 * +---------------+ |
6701 * +---------------+ |
6706 * +-------+ +-------+
6708 * | cache | | cache |
6709 * +-------+ +-------+
6710 * +=========+ .-----.
6711 * : L2ARC : |-_____-|
6712 * : devices : | Disks |
6713 * +=========+ `-_____-'
6715 * Read requests are satisfied from the following sources, in order:
6718 * 2) vdev cache of L2ARC devices
6720 * 4) vdev cache of disks
6723 * Some L2ARC device types exhibit extremely slow write performance.
6724 * To accommodate for this there are some significant differences between
6725 * the L2ARC and traditional cache design:
6727 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6728 * the ARC behave as usual, freeing buffers and placing headers on ghost
6729 * lists. The ARC does not send buffers to the L2ARC during eviction as
6730 * this would add inflated write latencies for all ARC memory pressure.
6732 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6733 * It does this by periodically scanning buffers from the eviction-end of
6734 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6735 * not already there. It scans until a headroom of buffers is satisfied,
6736 * which itself is a buffer for ARC eviction. If a compressible buffer is
6737 * found during scanning and selected for writing to an L2ARC device, we
6738 * temporarily boost scanning headroom during the next scan cycle to make
6739 * sure we adapt to compression effects (which might significantly reduce
6740 * the data volume we write to L2ARC). The thread that does this is
6741 * l2arc_feed_thread(), illustrated below; example sizes are included to
6742 * provide a better sense of ratio than this diagram:
6745 * +---------------------+----------+
6746 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6747 * +---------------------+----------+ | o L2ARC eligible
6748 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6749 * +---------------------+----------+ |
6750 * 15.9 Gbytes ^ 32 Mbytes |
6752 * l2arc_feed_thread()
6754 * l2arc write hand <--[oooo]--'
6758 * +==============================+
6759 * L2ARC dev |####|#|###|###| |####| ... |
6760 * +==============================+
6763 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6764 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6765 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6766 * safe to say that this is an uncommon case, since buffers at the end of
6767 * the ARC lists have moved there due to inactivity.
6769 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6770 * then the L2ARC simply misses copying some buffers. This serves as a
6771 * pressure valve to prevent heavy read workloads from both stalling the ARC
6772 * with waits and clogging the L2ARC with writes. This also helps prevent
6773 * the potential for the L2ARC to churn if it attempts to cache content too
6774 * quickly, such as during backups of the entire pool.
6776 * 5. After system boot and before the ARC has filled main memory, there are
6777 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6778 * lists can remain mostly static. Instead of searching from tail of these
6779 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6780 * for eligible buffers, greatly increasing its chance of finding them.
6782 * The L2ARC device write speed is also boosted during this time so that
6783 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6784 * there are no L2ARC reads, and no fear of degrading read performance
6785 * through increased writes.
6787 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6788 * the vdev queue can aggregate them into larger and fewer writes. Each
6789 * device is written to in a rotor fashion, sweeping writes through
6790 * available space then repeating.
6792 * 7. The L2ARC does not store dirty content. It never needs to flush
6793 * write buffers back to disk based storage.
6795 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6796 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6798 * The performance of the L2ARC can be tweaked by a number of tunables, which
6799 * may be necessary for different workloads:
6801 * l2arc_write_max max write bytes per interval
6802 * l2arc_write_boost extra write bytes during device warmup
6803 * l2arc_noprefetch skip caching prefetched buffers
6804 * l2arc_headroom number of max device writes to precache
6805 * l2arc_headroom_boost when we find compressed buffers during ARC
6806 * scanning, we multiply headroom by this
6807 * percentage factor for the next scan cycle,
6808 * since more compressed buffers are likely to
6810 * l2arc_feed_secs seconds between L2ARC writing
6812 * Tunables may be removed or added as future performance improvements are
6813 * integrated, and also may become zpool properties.
6815 * There are three key functions that control how the L2ARC warms up:
6817 * l2arc_write_eligible() check if a buffer is eligible to cache
6818 * l2arc_write_size() calculate how much to write
6819 * l2arc_write_interval() calculate sleep delay between writes
6821 * These three functions determine what to write, how much, and how quickly
6826 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6829 * A buffer is *not* eligible for the L2ARC if it:
6830 * 1. belongs to a different spa.
6831 * 2. is already cached on the L2ARC.
6832 * 3. has an I/O in progress (it may be an incomplete read).
6833 * 4. is flagged not eligible (zfs property).
6835 if (hdr->b_spa != spa_guid) {
6836 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
6839 if (HDR_HAS_L2HDR(hdr)) {
6840 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
6843 if (HDR_IO_IN_PROGRESS(hdr)) {
6844 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
6847 if (!HDR_L2CACHE(hdr)) {
6848 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
6856 l2arc_write_size(void)
6861 * Make sure our globals have meaningful values in case the user
6864 size = l2arc_write_max;
6866 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
6867 "be greater than zero, resetting it to the default (%d)",
6869 size = l2arc_write_max = L2ARC_WRITE_SIZE;
6872 if (arc_warm == B_FALSE)
6873 size += l2arc_write_boost;
6880 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
6882 clock_t interval, next, now;
6885 * If the ARC lists are busy, increase our write rate; if the
6886 * lists are stale, idle back. This is achieved by checking
6887 * how much we previously wrote - if it was more than half of
6888 * what we wanted, schedule the next write much sooner.
6890 if (l2arc_feed_again && wrote > (wanted / 2))
6891 interval = (hz * l2arc_feed_min_ms) / 1000;
6893 interval = hz * l2arc_feed_secs;
6895 now = ddi_get_lbolt();
6896 next = MAX(now, MIN(now + interval, began + interval));
6902 * Cycle through L2ARC devices. This is how L2ARC load balances.
6903 * If a device is returned, this also returns holding the spa config lock.
6905 static l2arc_dev_t *
6906 l2arc_dev_get_next(void)
6908 l2arc_dev_t *first, *next = NULL;
6911 * Lock out the removal of spas (spa_namespace_lock), then removal
6912 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6913 * both locks will be dropped and a spa config lock held instead.
6915 mutex_enter(&spa_namespace_lock);
6916 mutex_enter(&l2arc_dev_mtx);
6918 /* if there are no vdevs, there is nothing to do */
6919 if (l2arc_ndev == 0)
6923 next = l2arc_dev_last;
6925 /* loop around the list looking for a non-faulted vdev */
6927 next = list_head(l2arc_dev_list);
6929 next = list_next(l2arc_dev_list, next);
6931 next = list_head(l2arc_dev_list);
6934 /* if we have come back to the start, bail out */
6937 else if (next == first)
6940 } while (vdev_is_dead(next->l2ad_vdev));
6942 /* if we were unable to find any usable vdevs, return NULL */
6943 if (vdev_is_dead(next->l2ad_vdev))
6946 l2arc_dev_last = next;
6949 mutex_exit(&l2arc_dev_mtx);
6952 * Grab the config lock to prevent the 'next' device from being
6953 * removed while we are writing to it.
6956 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
6957 mutex_exit(&spa_namespace_lock);
6963 * Free buffers that were tagged for destruction.
6966 l2arc_do_free_on_write()
6969 l2arc_data_free_t *df, *df_prev;
6971 mutex_enter(&l2arc_free_on_write_mtx);
6972 buflist = l2arc_free_on_write;
6974 for (df = list_tail(buflist); df; df = df_prev) {
6975 df_prev = list_prev(buflist, df);
6976 ASSERT3P(df->l2df_abd, !=, NULL);
6977 abd_free(df->l2df_abd);
6978 list_remove(buflist, df);
6979 kmem_free(df, sizeof (l2arc_data_free_t));
6982 mutex_exit(&l2arc_free_on_write_mtx);
6986 * A write to a cache device has completed. Update all headers to allow
6987 * reads from these buffers to begin.
6990 l2arc_write_done(zio_t *zio)
6992 l2arc_write_callback_t *cb;
6995 arc_buf_hdr_t *head, *hdr, *hdr_prev;
6996 kmutex_t *hash_lock;
6997 int64_t bytes_dropped = 0;
6999 cb = zio->io_private;
7000 ASSERT3P(cb, !=, NULL);
7001 dev = cb->l2wcb_dev;
7002 ASSERT3P(dev, !=, NULL);
7003 head = cb->l2wcb_head;
7004 ASSERT3P(head, !=, NULL);
7005 buflist = &dev->l2ad_buflist;
7006 ASSERT3P(buflist, !=, NULL);
7007 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7008 l2arc_write_callback_t *, cb);
7010 if (zio->io_error != 0)
7011 ARCSTAT_BUMP(arcstat_l2_writes_error);
7014 * All writes completed, or an error was hit.
7017 mutex_enter(&dev->l2ad_mtx);
7018 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7019 hdr_prev = list_prev(buflist, hdr);
7021 hash_lock = HDR_LOCK(hdr);
7024 * We cannot use mutex_enter or else we can deadlock
7025 * with l2arc_write_buffers (due to swapping the order
7026 * the hash lock and l2ad_mtx are taken).
7028 if (!mutex_tryenter(hash_lock)) {
7030 * Missed the hash lock. We must retry so we
7031 * don't leave the ARC_FLAG_L2_WRITING bit set.
7033 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7036 * We don't want to rescan the headers we've
7037 * already marked as having been written out, so
7038 * we reinsert the head node so we can pick up
7039 * where we left off.
7041 list_remove(buflist, head);
7042 list_insert_after(buflist, hdr, head);
7044 mutex_exit(&dev->l2ad_mtx);
7047 * We wait for the hash lock to become available
7048 * to try and prevent busy waiting, and increase
7049 * the chance we'll be able to acquire the lock
7050 * the next time around.
7052 mutex_enter(hash_lock);
7053 mutex_exit(hash_lock);
7058 * We could not have been moved into the arc_l2c_only
7059 * state while in-flight due to our ARC_FLAG_L2_WRITING
7060 * bit being set. Let's just ensure that's being enforced.
7062 ASSERT(HDR_HAS_L1HDR(hdr));
7064 if (zio->io_error != 0) {
7066 * Error - drop L2ARC entry.
7068 list_remove(buflist, hdr);
7070 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7072 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
7073 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
7075 bytes_dropped += arc_hdr_size(hdr);
7076 (void) refcount_remove_many(&dev->l2ad_alloc,
7077 arc_hdr_size(hdr), hdr);
7081 * Allow ARC to begin reads and ghost list evictions to
7084 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7086 mutex_exit(hash_lock);
7089 atomic_inc_64(&l2arc_writes_done);
7090 list_remove(buflist, head);
7091 ASSERT(!HDR_HAS_L1HDR(head));
7092 kmem_cache_free(hdr_l2only_cache, head);
7093 mutex_exit(&dev->l2ad_mtx);
7095 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7097 l2arc_do_free_on_write();
7099 kmem_free(cb, sizeof (l2arc_write_callback_t));
7103 * A read to a cache device completed. Validate buffer contents before
7104 * handing over to the regular ARC routines.
7107 l2arc_read_done(zio_t *zio)
7109 l2arc_read_callback_t *cb;
7111 kmutex_t *hash_lock;
7112 boolean_t valid_cksum;
7114 ASSERT3P(zio->io_vd, !=, NULL);
7115 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7117 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7119 cb = zio->io_private;
7120 ASSERT3P(cb, !=, NULL);
7121 hdr = cb->l2rcb_hdr;
7122 ASSERT3P(hdr, !=, NULL);
7124 hash_lock = HDR_LOCK(hdr);
7125 mutex_enter(hash_lock);
7126 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7129 * If the data was read into a temporary buffer,
7130 * move it and free the buffer.
7132 if (cb->l2rcb_abd != NULL) {
7133 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7134 if (zio->io_error == 0) {
7135 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
7140 * The following must be done regardless of whether
7141 * there was an error:
7142 * - free the temporary buffer
7143 * - point zio to the real ARC buffer
7144 * - set zio size accordingly
7145 * These are required because zio is either re-used for
7146 * an I/O of the block in the case of the error
7147 * or the zio is passed to arc_read_done() and it
7150 abd_free(cb->l2rcb_abd);
7151 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7152 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
7155 ASSERT3P(zio->io_abd, !=, NULL);
7158 * Check this survived the L2ARC journey.
7160 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
7161 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
7162 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
7164 valid_cksum = arc_cksum_is_equal(hdr, zio);
7165 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
7166 mutex_exit(hash_lock);
7167 zio->io_private = hdr;
7170 mutex_exit(hash_lock);
7172 * Buffer didn't survive caching. Increment stats and
7173 * reissue to the original storage device.
7175 if (zio->io_error != 0) {
7176 ARCSTAT_BUMP(arcstat_l2_io_error);
7178 zio->io_error = SET_ERROR(EIO);
7181 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
7184 * If there's no waiter, issue an async i/o to the primary
7185 * storage now. If there *is* a waiter, the caller must
7186 * issue the i/o in a context where it's OK to block.
7188 if (zio->io_waiter == NULL) {
7189 zio_t *pio = zio_unique_parent(zio);
7191 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7193 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
7194 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
7195 hdr, zio->io_priority, cb->l2rcb_flags,
7200 kmem_free(cb, sizeof (l2arc_read_callback_t));
7204 * This is the list priority from which the L2ARC will search for pages to
7205 * cache. This is used within loops (0..3) to cycle through lists in the
7206 * desired order. This order can have a significant effect on cache
7209 * Currently the metadata lists are hit first, MFU then MRU, followed by
7210 * the data lists. This function returns a locked list, and also returns
7213 static multilist_sublist_t *
7214 l2arc_sublist_lock(int list_num)
7216 multilist_t *ml = NULL;
7219 ASSERT(list_num >= 0 && list_num <= 3);
7223 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7226 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7229 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7232 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7237 * Return a randomly-selected sublist. This is acceptable
7238 * because the caller feeds only a little bit of data for each
7239 * call (8MB). Subsequent calls will result in different
7240 * sublists being selected.
7242 idx = multilist_get_random_index(ml);
7243 return (multilist_sublist_lock(ml, idx));
7247 * Evict buffers from the device write hand to the distance specified in
7248 * bytes. This distance may span populated buffers, it may span nothing.
7249 * This is clearing a region on the L2ARC device ready for writing.
7250 * If the 'all' boolean is set, every buffer is evicted.
7253 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7256 arc_buf_hdr_t *hdr, *hdr_prev;
7257 kmutex_t *hash_lock;
7260 buflist = &dev->l2ad_buflist;
7262 if (!all && dev->l2ad_first) {
7264 * This is the first sweep through the device. There is
7270 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
7272 * When nearing the end of the device, evict to the end
7273 * before the device write hand jumps to the start.
7275 taddr = dev->l2ad_end;
7277 taddr = dev->l2ad_hand + distance;
7279 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
7280 uint64_t, taddr, boolean_t, all);
7283 mutex_enter(&dev->l2ad_mtx);
7284 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
7285 hdr_prev = list_prev(buflist, hdr);
7287 hash_lock = HDR_LOCK(hdr);
7290 * We cannot use mutex_enter or else we can deadlock
7291 * with l2arc_write_buffers (due to swapping the order
7292 * the hash lock and l2ad_mtx are taken).
7294 if (!mutex_tryenter(hash_lock)) {
7296 * Missed the hash lock. Retry.
7298 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
7299 mutex_exit(&dev->l2ad_mtx);
7300 mutex_enter(hash_lock);
7301 mutex_exit(hash_lock);
7306 * A header can't be on this list if it doesn't have L2 header.
7308 ASSERT(HDR_HAS_L2HDR(hdr));
7310 /* Ensure this header has finished being written. */
7311 ASSERT(!HDR_L2_WRITING(hdr));
7312 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
7314 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
7315 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
7317 * We've evicted to the target address,
7318 * or the end of the device.
7320 mutex_exit(hash_lock);
7324 if (!HDR_HAS_L1HDR(hdr)) {
7325 ASSERT(!HDR_L2_READING(hdr));
7327 * This doesn't exist in the ARC. Destroy.
7328 * arc_hdr_destroy() will call list_remove()
7329 * and decrement arcstat_l2_lsize.
7331 arc_change_state(arc_anon, hdr, hash_lock);
7332 arc_hdr_destroy(hdr);
7334 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
7335 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
7337 * Invalidate issued or about to be issued
7338 * reads, since we may be about to write
7339 * over this location.
7341 if (HDR_L2_READING(hdr)) {
7342 ARCSTAT_BUMP(arcstat_l2_evict_reading);
7343 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
7346 arc_hdr_l2hdr_destroy(hdr);
7348 mutex_exit(hash_lock);
7350 mutex_exit(&dev->l2ad_mtx);
7354 * Find and write ARC buffers to the L2ARC device.
7356 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7357 * for reading until they have completed writing.
7358 * The headroom_boost is an in-out parameter used to maintain headroom boost
7359 * state between calls to this function.
7361 * Returns the number of bytes actually written (which may be smaller than
7362 * the delta by which the device hand has changed due to alignment).
7365 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
7367 arc_buf_hdr_t *hdr, *hdr_prev, *head;
7368 uint64_t write_asize, write_psize, write_lsize, headroom;
7370 l2arc_write_callback_t *cb;
7372 uint64_t guid = spa_load_guid(spa);
7375 ASSERT3P(dev->l2ad_vdev, !=, NULL);
7378 write_lsize = write_asize = write_psize = 0;
7380 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
7381 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
7383 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
7385 * Copy buffers for L2ARC writing.
7387 for (try = 0; try <= 3; try++) {
7388 multilist_sublist_t *mls = l2arc_sublist_lock(try);
7389 uint64_t passed_sz = 0;
7391 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
7394 * L2ARC fast warmup.
7396 * Until the ARC is warm and starts to evict, read from the
7397 * head of the ARC lists rather than the tail.
7399 if (arc_warm == B_FALSE)
7400 hdr = multilist_sublist_head(mls);
7402 hdr = multilist_sublist_tail(mls);
7404 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
7406 headroom = target_sz * l2arc_headroom;
7407 if (zfs_compressed_arc_enabled)
7408 headroom = (headroom * l2arc_headroom_boost) / 100;
7410 for (; hdr; hdr = hdr_prev) {
7411 kmutex_t *hash_lock;
7413 if (arc_warm == B_FALSE)
7414 hdr_prev = multilist_sublist_next(mls, hdr);
7416 hdr_prev = multilist_sublist_prev(mls, hdr);
7417 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
7418 HDR_GET_LSIZE(hdr));
7420 hash_lock = HDR_LOCK(hdr);
7421 if (!mutex_tryenter(hash_lock)) {
7422 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
7424 * Skip this buffer rather than waiting.
7429 passed_sz += HDR_GET_LSIZE(hdr);
7430 if (passed_sz > headroom) {
7434 mutex_exit(hash_lock);
7435 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
7439 if (!l2arc_write_eligible(guid, hdr)) {
7440 mutex_exit(hash_lock);
7445 * We rely on the L1 portion of the header below, so
7446 * it's invalid for this header to have been evicted out
7447 * of the ghost cache, prior to being written out. The
7448 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7450 ASSERT(HDR_HAS_L1HDR(hdr));
7452 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
7453 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7454 ASSERT3U(arc_hdr_size(hdr), >, 0);
7455 uint64_t psize = arc_hdr_size(hdr);
7456 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
7459 if ((write_asize + asize) > target_sz) {
7461 mutex_exit(hash_lock);
7462 ARCSTAT_BUMP(arcstat_l2_write_full);
7468 * Insert a dummy header on the buflist so
7469 * l2arc_write_done() can find where the
7470 * write buffers begin without searching.
7472 mutex_enter(&dev->l2ad_mtx);
7473 list_insert_head(&dev->l2ad_buflist, head);
7474 mutex_exit(&dev->l2ad_mtx);
7477 sizeof (l2arc_write_callback_t), KM_SLEEP);
7478 cb->l2wcb_dev = dev;
7479 cb->l2wcb_head = head;
7480 pio = zio_root(spa, l2arc_write_done, cb,
7482 ARCSTAT_BUMP(arcstat_l2_write_pios);
7485 hdr->b_l2hdr.b_dev = dev;
7486 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7487 arc_hdr_set_flags(hdr,
7488 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7490 mutex_enter(&dev->l2ad_mtx);
7491 list_insert_head(&dev->l2ad_buflist, hdr);
7492 mutex_exit(&dev->l2ad_mtx);
7494 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
7497 * Normally the L2ARC can use the hdr's data, but if
7498 * we're sharing data between the hdr and one of its
7499 * bufs, L2ARC needs its own copy of the data so that
7500 * the ZIO below can't race with the buf consumer.
7501 * Another case where we need to create a copy of the
7502 * data is when the buffer size is not device-aligned
7503 * and we need to pad the block to make it such.
7504 * That also keeps the clock hand suitably aligned.
7506 * To ensure that the copy will be available for the
7507 * lifetime of the ZIO and be cleaned up afterwards, we
7508 * add it to the l2arc_free_on_write queue.
7511 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
7512 to_write = hdr->b_l1hdr.b_pabd;
7514 to_write = abd_alloc_for_io(asize,
7515 HDR_ISTYPE_METADATA(hdr));
7516 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
7517 if (asize != psize) {
7518 abd_zero_off(to_write, psize,
7521 l2arc_free_abd_on_write(to_write, asize,
7524 wzio = zio_write_phys(pio, dev->l2ad_vdev,
7525 hdr->b_l2hdr.b_daddr, asize, to_write,
7526 ZIO_CHECKSUM_OFF, NULL, hdr,
7527 ZIO_PRIORITY_ASYNC_WRITE,
7528 ZIO_FLAG_CANFAIL, B_FALSE);
7530 write_lsize += HDR_GET_LSIZE(hdr);
7531 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7534 write_psize += psize;
7535 write_asize += asize;
7536 dev->l2ad_hand += asize;
7538 mutex_exit(hash_lock);
7540 (void) zio_nowait(wzio);
7543 multilist_sublist_unlock(mls);
7549 /* No buffers selected for writing? */
7551 ASSERT0(write_lsize);
7552 ASSERT(!HDR_HAS_L1HDR(head));
7553 kmem_cache_free(hdr_l2only_cache, head);
7557 ASSERT3U(write_psize, <=, target_sz);
7558 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7559 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
7560 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
7561 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
7562 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
7565 * Bump device hand to the device start if it is approaching the end.
7566 * l2arc_evict() will already have evicted ahead for this case.
7568 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7569 dev->l2ad_hand = dev->l2ad_start;
7570 dev->l2ad_first = B_FALSE;
7573 dev->l2ad_writing = B_TRUE;
7574 (void) zio_wait(pio);
7575 dev->l2ad_writing = B_FALSE;
7577 return (write_asize);
7581 * This thread feeds the L2ARC at regular intervals. This is the beating
7582 * heart of the L2ARC.
7586 l2arc_feed_thread(void *unused __unused)
7591 uint64_t size, wrote;
7592 clock_t begin, next = ddi_get_lbolt();
7594 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7596 mutex_enter(&l2arc_feed_thr_lock);
7598 while (l2arc_thread_exit == 0) {
7599 CALLB_CPR_SAFE_BEGIN(&cpr);
7600 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7601 next - ddi_get_lbolt());
7602 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7603 next = ddi_get_lbolt() + hz;
7606 * Quick check for L2ARC devices.
7608 mutex_enter(&l2arc_dev_mtx);
7609 if (l2arc_ndev == 0) {
7610 mutex_exit(&l2arc_dev_mtx);
7613 mutex_exit(&l2arc_dev_mtx);
7614 begin = ddi_get_lbolt();
7617 * This selects the next l2arc device to write to, and in
7618 * doing so the next spa to feed from: dev->l2ad_spa. This
7619 * will return NULL if there are now no l2arc devices or if
7620 * they are all faulted.
7622 * If a device is returned, its spa's config lock is also
7623 * held to prevent device removal. l2arc_dev_get_next()
7624 * will grab and release l2arc_dev_mtx.
7626 if ((dev = l2arc_dev_get_next()) == NULL)
7629 spa = dev->l2ad_spa;
7630 ASSERT3P(spa, !=, NULL);
7633 * If the pool is read-only then force the feed thread to
7634 * sleep a little longer.
7636 if (!spa_writeable(spa)) {
7637 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7638 spa_config_exit(spa, SCL_L2ARC, dev);
7643 * Avoid contributing to memory pressure.
7645 if (arc_reclaim_needed()) {
7646 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7647 spa_config_exit(spa, SCL_L2ARC, dev);
7651 ARCSTAT_BUMP(arcstat_l2_feeds);
7653 size = l2arc_write_size();
7656 * Evict L2ARC buffers that will be overwritten.
7658 l2arc_evict(dev, size, B_FALSE);
7661 * Write ARC buffers.
7663 wrote = l2arc_write_buffers(spa, dev, size);
7666 * Calculate interval between writes.
7668 next = l2arc_write_interval(begin, size, wrote);
7669 spa_config_exit(spa, SCL_L2ARC, dev);
7672 l2arc_thread_exit = 0;
7673 cv_broadcast(&l2arc_feed_thr_cv);
7674 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7679 l2arc_vdev_present(vdev_t *vd)
7683 mutex_enter(&l2arc_dev_mtx);
7684 for (dev = list_head(l2arc_dev_list); dev != NULL;
7685 dev = list_next(l2arc_dev_list, dev)) {
7686 if (dev->l2ad_vdev == vd)
7689 mutex_exit(&l2arc_dev_mtx);
7691 return (dev != NULL);
7695 * Add a vdev for use by the L2ARC. By this point the spa has already
7696 * validated the vdev and opened it.
7699 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7701 l2arc_dev_t *adddev;
7703 ASSERT(!l2arc_vdev_present(vd));
7705 vdev_ashift_optimize(vd);
7708 * Create a new l2arc device entry.
7710 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7711 adddev->l2ad_spa = spa;
7712 adddev->l2ad_vdev = vd;
7713 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7714 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7715 adddev->l2ad_hand = adddev->l2ad_start;
7716 adddev->l2ad_first = B_TRUE;
7717 adddev->l2ad_writing = B_FALSE;
7719 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7721 * This is a list of all ARC buffers that are still valid on the
7724 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7725 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7727 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7728 refcount_create(&adddev->l2ad_alloc);
7731 * Add device to global list
7733 mutex_enter(&l2arc_dev_mtx);
7734 list_insert_head(l2arc_dev_list, adddev);
7735 atomic_inc_64(&l2arc_ndev);
7736 mutex_exit(&l2arc_dev_mtx);
7740 * Remove a vdev from the L2ARC.
7743 l2arc_remove_vdev(vdev_t *vd)
7745 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7748 * Find the device by vdev
7750 mutex_enter(&l2arc_dev_mtx);
7751 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7752 nextdev = list_next(l2arc_dev_list, dev);
7753 if (vd == dev->l2ad_vdev) {
7758 ASSERT3P(remdev, !=, NULL);
7761 * Remove device from global list
7763 list_remove(l2arc_dev_list, remdev);
7764 l2arc_dev_last = NULL; /* may have been invalidated */
7765 atomic_dec_64(&l2arc_ndev);
7766 mutex_exit(&l2arc_dev_mtx);
7769 * Clear all buflists and ARC references. L2ARC device flush.
7771 l2arc_evict(remdev, 0, B_TRUE);
7772 list_destroy(&remdev->l2ad_buflist);
7773 mutex_destroy(&remdev->l2ad_mtx);
7774 refcount_destroy(&remdev->l2ad_alloc);
7775 kmem_free(remdev, sizeof (l2arc_dev_t));
7781 l2arc_thread_exit = 0;
7783 l2arc_writes_sent = 0;
7784 l2arc_writes_done = 0;
7786 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7787 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7788 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7789 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7791 l2arc_dev_list = &L2ARC_dev_list;
7792 l2arc_free_on_write = &L2ARC_free_on_write;
7793 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7794 offsetof(l2arc_dev_t, l2ad_node));
7795 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7796 offsetof(l2arc_data_free_t, l2df_list_node));
7803 * This is called from dmu_fini(), which is called from spa_fini();
7804 * Because of this, we can assume that all l2arc devices have
7805 * already been removed when the pools themselves were removed.
7808 l2arc_do_free_on_write();
7810 mutex_destroy(&l2arc_feed_thr_lock);
7811 cv_destroy(&l2arc_feed_thr_cv);
7812 mutex_destroy(&l2arc_dev_mtx);
7813 mutex_destroy(&l2arc_free_on_write_mtx);
7815 list_destroy(l2arc_dev_list);
7816 list_destroy(l2arc_free_on_write);
7822 if (!(spa_mode_global & FWRITE))
7825 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7826 TS_RUN, minclsyspri);
7832 if (!(spa_mode_global & FWRITE))
7835 mutex_enter(&l2arc_feed_thr_lock);
7836 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7837 l2arc_thread_exit = 1;
7838 while (l2arc_thread_exit != 0)
7839 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7840 mutex_exit(&l2arc_feed_thr_lock);