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 2015 Nexenta Systems, Inc. All rights reserved.
30 * DVA-based Adjustable Replacement Cache
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal ARC algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
96 * Each ARC state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an ARC list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
102 * Note that the majority of the performance stats are manipulated
103 * with atomic operations.
105 * The L2ARC uses the l2ad_mtx on each vdev for the following:
107 * - L2ARC buflist creation
108 * - L2ARC buflist eviction
109 * - L2ARC write completion, which walks L2ARC buflists
110 * - ARC header destruction, as it removes from L2ARC buflists
111 * - ARC header release, as it removes from L2ARC buflists
117 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
118 * This structure can point either to a block that is still in the cache or to
119 * one that is only accessible in an L2 ARC device, or it can provide
120 * information about a block that was recently evicted. If a block is
121 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
122 * information to retrieve it from the L2ARC device. This information is
123 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
124 * that is in this state cannot access the data directly.
126 * Blocks that are actively being referenced or have not been evicted
127 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
128 * the arc_buf_hdr_t that will point to the data block in memory. A block can
129 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
130 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
131 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
133 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
134 * ability to store the physical data (b_pabd) associated with the DVA of the
135 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
136 * it will match its on-disk compression characteristics. This behavior can be
137 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
138 * compressed ARC functionality is disabled, the b_pabd will point to an
139 * uncompressed version of the on-disk data.
141 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
142 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
143 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
144 * consumer. The ARC will provide references to this data and will keep it
145 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
146 * data block and will evict any arc_buf_t that is no longer referenced. The
147 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
148 * "overhead_size" kstat.
150 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
151 * compressed form. The typical case is that consumers will want uncompressed
152 * data, and when that happens a new data buffer is allocated where the data is
153 * decompressed for them to use. Currently the only consumer who wants
154 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
155 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
156 * with the arc_buf_hdr_t.
158 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
159 * first one is owned by a compressed send consumer (and therefore references
160 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
161 * used by any other consumer (and has its own uncompressed copy of the data
176 * | b_buf +------------>+-----------+ arc_buf_t
177 * | b_pabd +-+ |b_next +---->+-----------+
178 * +-----------+ | |-----------| |b_next +-->NULL
179 * | |b_comp = T | +-----------+
180 * | |b_data +-+ |b_comp = F |
181 * | +-----------+ | |b_data +-+
182 * +->+------+ | +-----------+ |
184 * data | |<--------------+ | uncompressed
185 * +------+ compressed, | data
186 * shared +-->+------+
191 * When a consumer reads a block, the ARC must first look to see if the
192 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
193 * arc_buf_t and either copies uncompressed data into a new data buffer from an
194 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
195 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
196 * hdr is compressed and the desired compression characteristics of the
197 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
198 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
199 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
200 * be anywhere in the hdr's list.
202 * The diagram below shows an example of an uncompressed ARC hdr that is
203 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
204 * the last element in the buf list):
216 * | | arc_buf_t (shared)
217 * | b_buf +------------>+---------+ arc_buf_t
218 * | | |b_next +---->+---------+
219 * | b_pabd +-+ |---------| |b_next +-->NULL
220 * +-----------+ | | | +---------+
222 * | +---------+ | |b_data +-+
223 * +->+------+ | +---------+ |
225 * uncompressed | | | |
228 * | uncompressed | | |
231 * +---------------------------------+
233 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
234 * since the physical block is about to be rewritten. The new data contents
235 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
236 * it may compress the data before writing it to disk. The ARC will be called
237 * with the transformed data and will bcopy the transformed on-disk block into
238 * a newly allocated b_pabd. Writes are always done into buffers which have
239 * either been loaned (and hence are new and don't have other readers) or
240 * buffers which have been released (and hence have their own hdr, if there
241 * were originally other readers of the buf's original hdr). This ensures that
242 * the ARC only needs to update a single buf and its hdr after a write occurs.
244 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
245 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
246 * that when compressed ARC is enabled that the L2ARC blocks are identical
247 * to the on-disk block in the main data pool. This provides a significant
248 * advantage since the ARC can leverage the bp's checksum when reading from the
249 * L2ARC to determine if the contents are valid. However, if the compressed
250 * ARC is disabled, then the L2ARC's block must be transformed to look
251 * like the physical block in the main data pool before comparing the
252 * checksum and determining its validity.
257 #include <sys/spa_impl.h>
258 #include <sys/zio_compress.h>
259 #include <sys/zio_checksum.h>
260 #include <sys/zfs_context.h>
262 #include <sys/refcount.h>
263 #include <sys/vdev.h>
264 #include <sys/vdev_impl.h>
265 #include <sys/dsl_pool.h>
266 #include <sys/zio_checksum.h>
267 #include <sys/multilist.h>
270 #include <sys/dnlc.h>
271 #include <sys/racct.h>
273 #include <sys/callb.h>
274 #include <sys/kstat.h>
275 #include <sys/trim_map.h>
276 #include <zfs_fletcher.h>
279 #include <machine/vmparam.h>
283 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
284 boolean_t arc_watch = B_FALSE;
289 static kmutex_t arc_reclaim_lock;
290 static kcondvar_t arc_reclaim_thread_cv;
291 static boolean_t arc_reclaim_thread_exit;
292 static kcondvar_t arc_reclaim_waiters_cv;
294 static kmutex_t arc_dnlc_evicts_lock;
295 static kcondvar_t arc_dnlc_evicts_cv;
296 static boolean_t arc_dnlc_evicts_thread_exit;
298 uint_t arc_reduce_dnlc_percent = 3;
301 * The number of headers to evict in arc_evict_state_impl() before
302 * dropping the sublist lock and evicting from another sublist. A lower
303 * value means we're more likely to evict the "correct" header (i.e. the
304 * oldest header in the arc state), but comes with higher overhead
305 * (i.e. more invocations of arc_evict_state_impl()).
307 int zfs_arc_evict_batch_limit = 10;
309 /* number of seconds before growing cache again */
310 static int arc_grow_retry = 60;
312 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
313 int zfs_arc_overflow_shift = 8;
315 /* shift of arc_c for calculating both min and max arc_p */
316 static int arc_p_min_shift = 4;
318 /* log2(fraction of arc to reclaim) */
319 static int arc_shrink_shift = 7;
322 * log2(fraction of ARC which must be free to allow growing).
323 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
324 * when reading a new block into the ARC, we will evict an equal-sized block
327 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
328 * we will still not allow it to grow.
330 int arc_no_grow_shift = 5;
334 * minimum lifespan of a prefetch block in clock ticks
335 * (initialized in arc_init())
337 static int arc_min_prefetch_lifespan;
340 * If this percent of memory is free, don't throttle.
342 int arc_lotsfree_percent = 10;
345 extern boolean_t zfs_prefetch_disable;
348 * The arc has filled available memory and has now warmed up.
350 static boolean_t arc_warm;
353 * These tunables are for performance analysis.
355 uint64_t zfs_arc_max;
356 uint64_t zfs_arc_min;
357 uint64_t zfs_arc_meta_limit = 0;
358 uint64_t zfs_arc_meta_min = 0;
359 int zfs_arc_grow_retry = 0;
360 int zfs_arc_shrink_shift = 0;
361 int zfs_arc_no_grow_shift = 0;
362 int zfs_arc_p_min_shift = 0;
363 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
364 u_int zfs_arc_free_target = 0;
366 /* Absolute min for arc min / max is 16MB. */
367 static uint64_t arc_abs_min = 16 << 20;
369 boolean_t zfs_compressed_arc_enabled = B_TRUE;
371 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
372 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
373 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
374 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
375 static int sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS);
377 #if defined(__FreeBSD__) && defined(_KERNEL)
379 arc_free_target_init(void *unused __unused)
382 zfs_arc_free_target = vm_pageout_wakeup_thresh;
384 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
385 arc_free_target_init, NULL);
387 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
388 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
389 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
390 TUNABLE_INT("vfs.zfs.arc_grow_retry", &zfs_arc_grow_retry);
391 TUNABLE_INT("vfs.zfs.arc_no_grow_shift", &zfs_arc_no_grow_shift);
392 SYSCTL_DECL(_vfs_zfs);
393 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
394 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
395 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
396 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
397 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_no_grow_shift, CTLTYPE_U32 | CTLFLAG_RWTUN,
398 0, sizeof(uint32_t), sysctl_vfs_zfs_arc_no_grow_shift, "U",
399 "log2(fraction of ARC which must be free to allow growing)");
400 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
401 &zfs_arc_average_blocksize, 0,
402 "ARC average blocksize");
403 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
404 &arc_shrink_shift, 0,
405 "log2(fraction of arc to reclaim)");
406 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_grow_retry, CTLFLAG_RW,
408 "Wait in seconds before considering growing ARC");
409 SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN,
410 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC");
413 * We don't have a tunable for arc_free_target due to the dependency on
414 * pagedaemon initialisation.
416 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
417 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
418 sysctl_vfs_zfs_arc_free_target, "IU",
419 "Desired number of free pages below which ARC triggers reclaim");
422 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
427 val = zfs_arc_free_target;
428 err = sysctl_handle_int(oidp, &val, 0, req);
429 if (err != 0 || req->newptr == NULL)
434 if (val > vm_cnt.v_page_count)
437 zfs_arc_free_target = val;
443 * Must be declared here, before the definition of corresponding kstat
444 * macro which uses the same names will confuse the compiler.
446 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
447 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
448 sysctl_vfs_zfs_arc_meta_limit, "QU",
449 "ARC metadata limit");
453 * Note that buffers can be in one of 6 states:
454 * ARC_anon - anonymous (discussed below)
455 * ARC_mru - recently used, currently cached
456 * ARC_mru_ghost - recentely used, no longer in cache
457 * ARC_mfu - frequently used, currently cached
458 * ARC_mfu_ghost - frequently used, no longer in cache
459 * ARC_l2c_only - exists in L2ARC but not other states
460 * When there are no active references to the buffer, they are
461 * are linked onto a list in one of these arc states. These are
462 * the only buffers that can be evicted or deleted. Within each
463 * state there are multiple lists, one for meta-data and one for
464 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
465 * etc.) is tracked separately so that it can be managed more
466 * explicitly: favored over data, limited explicitly.
468 * Anonymous buffers are buffers that are not associated with
469 * a DVA. These are buffers that hold dirty block copies
470 * before they are written to stable storage. By definition,
471 * they are "ref'd" and are considered part of arc_mru
472 * that cannot be freed. Generally, they will aquire a DVA
473 * as they are written and migrate onto the arc_mru list.
475 * The ARC_l2c_only state is for buffers that are in the second
476 * level ARC but no longer in any of the ARC_m* lists. The second
477 * level ARC itself may also contain buffers that are in any of
478 * the ARC_m* states - meaning that a buffer can exist in two
479 * places. The reason for the ARC_l2c_only state is to keep the
480 * buffer header in the hash table, so that reads that hit the
481 * second level ARC benefit from these fast lookups.
484 typedef struct arc_state {
486 * list of evictable buffers
488 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
490 * total amount of evictable data in this state
492 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
494 * total amount of data in this state; this includes: evictable,
495 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
497 refcount_t arcs_size;
501 static arc_state_t ARC_anon;
502 static arc_state_t ARC_mru;
503 static arc_state_t ARC_mru_ghost;
504 static arc_state_t ARC_mfu;
505 static arc_state_t ARC_mfu_ghost;
506 static arc_state_t ARC_l2c_only;
508 typedef struct arc_stats {
509 kstat_named_t arcstat_hits;
510 kstat_named_t arcstat_misses;
511 kstat_named_t arcstat_demand_data_hits;
512 kstat_named_t arcstat_demand_data_misses;
513 kstat_named_t arcstat_demand_metadata_hits;
514 kstat_named_t arcstat_demand_metadata_misses;
515 kstat_named_t arcstat_prefetch_data_hits;
516 kstat_named_t arcstat_prefetch_data_misses;
517 kstat_named_t arcstat_prefetch_metadata_hits;
518 kstat_named_t arcstat_prefetch_metadata_misses;
519 kstat_named_t arcstat_mru_hits;
520 kstat_named_t arcstat_mru_ghost_hits;
521 kstat_named_t arcstat_mfu_hits;
522 kstat_named_t arcstat_mfu_ghost_hits;
523 kstat_named_t arcstat_allocated;
524 kstat_named_t arcstat_deleted;
526 * Number of buffers that could not be evicted because the hash lock
527 * was held by another thread. The lock may not necessarily be held
528 * by something using the same buffer, since hash locks are shared
529 * by multiple buffers.
531 kstat_named_t arcstat_mutex_miss;
533 * Number of buffers skipped because they have I/O in progress, are
534 * indrect prefetch buffers that have not lived long enough, or are
535 * not from the spa we're trying to evict from.
537 kstat_named_t arcstat_evict_skip;
539 * Number of times arc_evict_state() was unable to evict enough
540 * buffers to reach it's target amount.
542 kstat_named_t arcstat_evict_not_enough;
543 kstat_named_t arcstat_evict_l2_cached;
544 kstat_named_t arcstat_evict_l2_eligible;
545 kstat_named_t arcstat_evict_l2_ineligible;
546 kstat_named_t arcstat_evict_l2_skip;
547 kstat_named_t arcstat_hash_elements;
548 kstat_named_t arcstat_hash_elements_max;
549 kstat_named_t arcstat_hash_collisions;
550 kstat_named_t arcstat_hash_chains;
551 kstat_named_t arcstat_hash_chain_max;
552 kstat_named_t arcstat_p;
553 kstat_named_t arcstat_c;
554 kstat_named_t arcstat_c_min;
555 kstat_named_t arcstat_c_max;
556 kstat_named_t arcstat_size;
558 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
559 * Note that the compressed bytes may match the uncompressed bytes
560 * if the block is either not compressed or compressed arc is disabled.
562 kstat_named_t arcstat_compressed_size;
564 * Uncompressed size of the data stored in b_pabd. If compressed
565 * arc is disabled then this value will be identical to the stat
568 kstat_named_t arcstat_uncompressed_size;
570 * Number of bytes stored in all the arc_buf_t's. This is classified
571 * as "overhead" since this data is typically short-lived and will
572 * be evicted from the arc when it becomes unreferenced unless the
573 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
574 * values have been set (see comment in dbuf.c for more information).
576 kstat_named_t arcstat_overhead_size;
578 * Number of bytes consumed by internal ARC structures necessary
579 * for tracking purposes; these structures are not actually
580 * backed by ARC buffers. This includes arc_buf_hdr_t structures
581 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
582 * caches), and arc_buf_t structures (allocated via arc_buf_t
585 kstat_named_t arcstat_hdr_size;
587 * Number of bytes consumed by ARC buffers of type equal to
588 * ARC_BUFC_DATA. This is generally consumed by buffers backing
589 * on disk user data (e.g. plain file contents).
591 kstat_named_t arcstat_data_size;
593 * Number of bytes consumed by ARC buffers of type equal to
594 * ARC_BUFC_METADATA. This is generally consumed by buffers
595 * backing on disk data that is used for internal ZFS
596 * structures (e.g. ZAP, dnode, indirect blocks, etc).
598 kstat_named_t arcstat_metadata_size;
600 * Number of bytes consumed by various buffers and structures
601 * not actually backed with ARC buffers. This includes bonus
602 * buffers (allocated directly via zio_buf_* functions),
603 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
604 * cache), and dnode_t structures (allocated via dnode_t cache).
606 kstat_named_t arcstat_other_size;
608 * Total number of bytes consumed by ARC buffers residing in the
609 * arc_anon state. This includes *all* buffers in the arc_anon
610 * state; e.g. data, metadata, evictable, and unevictable buffers
611 * are all included in this value.
613 kstat_named_t arcstat_anon_size;
615 * Number of bytes consumed by ARC buffers that meet the
616 * following criteria: backing buffers of type ARC_BUFC_DATA,
617 * residing in the arc_anon state, and are eligible for eviction
618 * (e.g. have no outstanding holds on the buffer).
620 kstat_named_t arcstat_anon_evictable_data;
622 * Number of bytes consumed by ARC buffers that meet the
623 * following criteria: backing buffers of type ARC_BUFC_METADATA,
624 * residing in the arc_anon state, and are eligible for eviction
625 * (e.g. have no outstanding holds on the buffer).
627 kstat_named_t arcstat_anon_evictable_metadata;
629 * Total number of bytes consumed by ARC buffers residing in the
630 * arc_mru state. This includes *all* buffers in the arc_mru
631 * state; e.g. data, metadata, evictable, and unevictable buffers
632 * are all included in this value.
634 kstat_named_t arcstat_mru_size;
636 * Number of bytes consumed by ARC buffers that meet the
637 * following criteria: backing buffers of type ARC_BUFC_DATA,
638 * residing in the arc_mru state, and are eligible for eviction
639 * (e.g. have no outstanding holds on the buffer).
641 kstat_named_t arcstat_mru_evictable_data;
643 * Number of bytes consumed by ARC buffers that meet the
644 * following criteria: backing buffers of type ARC_BUFC_METADATA,
645 * residing in the arc_mru state, and are eligible for eviction
646 * (e.g. have no outstanding holds on the buffer).
648 kstat_named_t arcstat_mru_evictable_metadata;
650 * Total number of bytes that *would have been* consumed by ARC
651 * buffers in the arc_mru_ghost state. The key thing to note
652 * here, is the fact that this size doesn't actually indicate
653 * RAM consumption. The ghost lists only consist of headers and
654 * don't actually have ARC buffers linked off of these headers.
655 * Thus, *if* the headers had associated ARC buffers, these
656 * buffers *would have* consumed this number of bytes.
658 kstat_named_t arcstat_mru_ghost_size;
660 * Number of bytes that *would have been* consumed by ARC
661 * buffers that are eligible for eviction, of type
662 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
664 kstat_named_t arcstat_mru_ghost_evictable_data;
666 * Number of bytes that *would have been* consumed by ARC
667 * buffers that are eligible for eviction, of type
668 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
670 kstat_named_t arcstat_mru_ghost_evictable_metadata;
672 * Total number of bytes consumed by ARC buffers residing in the
673 * arc_mfu state. This includes *all* buffers in the arc_mfu
674 * state; e.g. data, metadata, evictable, and unevictable buffers
675 * are all included in this value.
677 kstat_named_t arcstat_mfu_size;
679 * Number of bytes consumed by ARC buffers that are eligible for
680 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
683 kstat_named_t arcstat_mfu_evictable_data;
685 * Number of bytes consumed by ARC buffers that are eligible for
686 * eviction, of type ARC_BUFC_METADATA, and reside in the
689 kstat_named_t arcstat_mfu_evictable_metadata;
691 * Total number of bytes that *would have been* consumed by ARC
692 * buffers in the arc_mfu_ghost state. See the comment above
693 * arcstat_mru_ghost_size for more details.
695 kstat_named_t arcstat_mfu_ghost_size;
697 * Number of bytes that *would have been* consumed by ARC
698 * buffers that are eligible for eviction, of type
699 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
701 kstat_named_t arcstat_mfu_ghost_evictable_data;
703 * Number of bytes that *would have been* consumed by ARC
704 * buffers that are eligible for eviction, of type
705 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
707 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
708 kstat_named_t arcstat_l2_hits;
709 kstat_named_t arcstat_l2_misses;
710 kstat_named_t arcstat_l2_feeds;
711 kstat_named_t arcstat_l2_rw_clash;
712 kstat_named_t arcstat_l2_read_bytes;
713 kstat_named_t arcstat_l2_write_bytes;
714 kstat_named_t arcstat_l2_writes_sent;
715 kstat_named_t arcstat_l2_writes_done;
716 kstat_named_t arcstat_l2_writes_error;
717 kstat_named_t arcstat_l2_writes_lock_retry;
718 kstat_named_t arcstat_l2_evict_lock_retry;
719 kstat_named_t arcstat_l2_evict_reading;
720 kstat_named_t arcstat_l2_evict_l1cached;
721 kstat_named_t arcstat_l2_free_on_write;
722 kstat_named_t arcstat_l2_abort_lowmem;
723 kstat_named_t arcstat_l2_cksum_bad;
724 kstat_named_t arcstat_l2_io_error;
725 kstat_named_t arcstat_l2_lsize;
726 kstat_named_t arcstat_l2_psize;
727 kstat_named_t arcstat_l2_hdr_size;
728 kstat_named_t arcstat_l2_write_trylock_fail;
729 kstat_named_t arcstat_l2_write_passed_headroom;
730 kstat_named_t arcstat_l2_write_spa_mismatch;
731 kstat_named_t arcstat_l2_write_in_l2;
732 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
733 kstat_named_t arcstat_l2_write_not_cacheable;
734 kstat_named_t arcstat_l2_write_full;
735 kstat_named_t arcstat_l2_write_buffer_iter;
736 kstat_named_t arcstat_l2_write_pios;
737 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
738 kstat_named_t arcstat_l2_write_buffer_list_iter;
739 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
740 kstat_named_t arcstat_memory_throttle_count;
741 kstat_named_t arcstat_meta_used;
742 kstat_named_t arcstat_meta_limit;
743 kstat_named_t arcstat_meta_max;
744 kstat_named_t arcstat_meta_min;
745 kstat_named_t arcstat_sync_wait_for_async;
746 kstat_named_t arcstat_demand_hit_predictive_prefetch;
749 static arc_stats_t arc_stats = {
750 { "hits", KSTAT_DATA_UINT64 },
751 { "misses", KSTAT_DATA_UINT64 },
752 { "demand_data_hits", KSTAT_DATA_UINT64 },
753 { "demand_data_misses", KSTAT_DATA_UINT64 },
754 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
755 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
756 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
757 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
758 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
759 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
760 { "mru_hits", KSTAT_DATA_UINT64 },
761 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
762 { "mfu_hits", KSTAT_DATA_UINT64 },
763 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
764 { "allocated", KSTAT_DATA_UINT64 },
765 { "deleted", KSTAT_DATA_UINT64 },
766 { "mutex_miss", KSTAT_DATA_UINT64 },
767 { "evict_skip", KSTAT_DATA_UINT64 },
768 { "evict_not_enough", KSTAT_DATA_UINT64 },
769 { "evict_l2_cached", KSTAT_DATA_UINT64 },
770 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
771 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
772 { "evict_l2_skip", KSTAT_DATA_UINT64 },
773 { "hash_elements", KSTAT_DATA_UINT64 },
774 { "hash_elements_max", KSTAT_DATA_UINT64 },
775 { "hash_collisions", KSTAT_DATA_UINT64 },
776 { "hash_chains", KSTAT_DATA_UINT64 },
777 { "hash_chain_max", KSTAT_DATA_UINT64 },
778 { "p", KSTAT_DATA_UINT64 },
779 { "c", KSTAT_DATA_UINT64 },
780 { "c_min", KSTAT_DATA_UINT64 },
781 { "c_max", KSTAT_DATA_UINT64 },
782 { "size", KSTAT_DATA_UINT64 },
783 { "compressed_size", KSTAT_DATA_UINT64 },
784 { "uncompressed_size", KSTAT_DATA_UINT64 },
785 { "overhead_size", KSTAT_DATA_UINT64 },
786 { "hdr_size", KSTAT_DATA_UINT64 },
787 { "data_size", KSTAT_DATA_UINT64 },
788 { "metadata_size", KSTAT_DATA_UINT64 },
789 { "other_size", KSTAT_DATA_UINT64 },
790 { "anon_size", KSTAT_DATA_UINT64 },
791 { "anon_evictable_data", KSTAT_DATA_UINT64 },
792 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
793 { "mru_size", KSTAT_DATA_UINT64 },
794 { "mru_evictable_data", KSTAT_DATA_UINT64 },
795 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
796 { "mru_ghost_size", KSTAT_DATA_UINT64 },
797 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
798 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
799 { "mfu_size", KSTAT_DATA_UINT64 },
800 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
801 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
802 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
803 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
804 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
805 { "l2_hits", KSTAT_DATA_UINT64 },
806 { "l2_misses", KSTAT_DATA_UINT64 },
807 { "l2_feeds", KSTAT_DATA_UINT64 },
808 { "l2_rw_clash", KSTAT_DATA_UINT64 },
809 { "l2_read_bytes", KSTAT_DATA_UINT64 },
810 { "l2_write_bytes", KSTAT_DATA_UINT64 },
811 { "l2_writes_sent", KSTAT_DATA_UINT64 },
812 { "l2_writes_done", KSTAT_DATA_UINT64 },
813 { "l2_writes_error", KSTAT_DATA_UINT64 },
814 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
815 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
816 { "l2_evict_reading", KSTAT_DATA_UINT64 },
817 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
818 { "l2_free_on_write", KSTAT_DATA_UINT64 },
819 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
820 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
821 { "l2_io_error", KSTAT_DATA_UINT64 },
822 { "l2_size", KSTAT_DATA_UINT64 },
823 { "l2_asize", KSTAT_DATA_UINT64 },
824 { "l2_hdr_size", KSTAT_DATA_UINT64 },
825 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
826 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
827 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
828 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
829 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
830 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
831 { "l2_write_full", KSTAT_DATA_UINT64 },
832 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
833 { "l2_write_pios", KSTAT_DATA_UINT64 },
834 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
835 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
836 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
837 { "memory_throttle_count", KSTAT_DATA_UINT64 },
838 { "arc_meta_used", KSTAT_DATA_UINT64 },
839 { "arc_meta_limit", KSTAT_DATA_UINT64 },
840 { "arc_meta_max", KSTAT_DATA_UINT64 },
841 { "arc_meta_min", KSTAT_DATA_UINT64 },
842 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
843 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
846 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
848 #define ARCSTAT_INCR(stat, val) \
849 atomic_add_64(&arc_stats.stat.value.ui64, (val))
851 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
852 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
854 #define ARCSTAT_MAX(stat, val) { \
856 while ((val) > (m = arc_stats.stat.value.ui64) && \
857 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
861 #define ARCSTAT_MAXSTAT(stat) \
862 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
865 * We define a macro to allow ARC hits/misses to be easily broken down by
866 * two separate conditions, giving a total of four different subtypes for
867 * each of hits and misses (so eight statistics total).
869 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
872 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
874 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
878 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
880 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
885 static arc_state_t *arc_anon;
886 static arc_state_t *arc_mru;
887 static arc_state_t *arc_mru_ghost;
888 static arc_state_t *arc_mfu;
889 static arc_state_t *arc_mfu_ghost;
890 static arc_state_t *arc_l2c_only;
893 * There are several ARC variables that are critical to export as kstats --
894 * but we don't want to have to grovel around in the kstat whenever we wish to
895 * manipulate them. For these variables, we therefore define them to be in
896 * terms of the statistic variable. This assures that we are not introducing
897 * the possibility of inconsistency by having shadow copies of the variables,
898 * while still allowing the code to be readable.
900 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
901 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
902 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
903 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
904 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
905 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
906 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
907 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
908 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
910 /* compressed size of entire arc */
911 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
912 /* uncompressed size of entire arc */
913 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
914 /* number of bytes in the arc from arc_buf_t's */
915 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
917 static int arc_no_grow; /* Don't try to grow cache size */
918 static uint64_t arc_tempreserve;
919 static uint64_t arc_loaned_bytes;
921 typedef struct arc_callback arc_callback_t;
923 struct arc_callback {
925 arc_done_func_t *acb_done;
927 boolean_t acb_compressed;
928 zio_t *acb_zio_dummy;
929 arc_callback_t *acb_next;
932 typedef struct arc_write_callback arc_write_callback_t;
934 struct arc_write_callback {
936 arc_done_func_t *awcb_ready;
937 arc_done_func_t *awcb_children_ready;
938 arc_done_func_t *awcb_physdone;
939 arc_done_func_t *awcb_done;
944 * ARC buffers are separated into multiple structs as a memory saving measure:
945 * - Common fields struct, always defined, and embedded within it:
946 * - L2-only fields, always allocated but undefined when not in L2ARC
947 * - L1-only fields, only allocated when in L1ARC
949 * Buffer in L1 Buffer only in L2
950 * +------------------------+ +------------------------+
951 * | arc_buf_hdr_t | | arc_buf_hdr_t |
955 * +------------------------+ +------------------------+
956 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
957 * | (undefined if L1-only) | | |
958 * +------------------------+ +------------------------+
959 * | l1arc_buf_hdr_t |
964 * +------------------------+
966 * Because it's possible for the L2ARC to become extremely large, we can wind
967 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
968 * is minimized by only allocating the fields necessary for an L1-cached buffer
969 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
970 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
971 * words in pointers. arc_hdr_realloc() is used to switch a header between
972 * these two allocation states.
974 typedef struct l1arc_buf_hdr {
975 kmutex_t b_freeze_lock;
976 zio_cksum_t *b_freeze_cksum;
979 * Used for debugging with kmem_flags - by allocating and freeing
980 * b_thawed when the buffer is thawed, we get a record of the stack
981 * trace that thawed it.
988 /* for waiting on writes to complete */
992 /* protected by arc state mutex */
993 arc_state_t *b_state;
994 multilist_node_t b_arc_node;
996 /* updated atomically */
997 clock_t b_arc_access;
999 /* self protecting */
1000 refcount_t b_refcnt;
1002 arc_callback_t *b_acb;
1006 typedef struct l2arc_dev l2arc_dev_t;
1008 typedef struct l2arc_buf_hdr {
1009 /* protected by arc_buf_hdr mutex */
1010 l2arc_dev_t *b_dev; /* L2ARC device */
1011 uint64_t b_daddr; /* disk address, offset byte */
1013 list_node_t b_l2node;
1016 struct arc_buf_hdr {
1017 /* protected by hash lock */
1021 arc_buf_contents_t b_type;
1022 arc_buf_hdr_t *b_hash_next;
1023 arc_flags_t b_flags;
1026 * This field stores the size of the data buffer after
1027 * compression, and is set in the arc's zio completion handlers.
1028 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1030 * While the block pointers can store up to 32MB in their psize
1031 * field, we can only store up to 32MB minus 512B. This is due
1032 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1033 * a field of zeros represents 512B in the bp). We can't use a
1034 * bias of 1 since we need to reserve a psize of zero, here, to
1035 * represent holes and embedded blocks.
1037 * This isn't a problem in practice, since the maximum size of a
1038 * buffer is limited to 16MB, so we never need to store 32MB in
1039 * this field. Even in the upstream illumos code base, the
1040 * maximum size of a buffer is limited to 16MB.
1045 * This field stores the size of the data buffer before
1046 * compression, and cannot change once set. It is in units
1047 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1049 uint16_t b_lsize; /* immutable */
1050 uint64_t b_spa; /* immutable */
1052 /* L2ARC fields. Undefined when not in L2ARC. */
1053 l2arc_buf_hdr_t b_l2hdr;
1054 /* L1ARC fields. Undefined when in l2arc_only state */
1055 l1arc_buf_hdr_t b_l1hdr;
1058 #if defined(__FreeBSD__) && defined(_KERNEL)
1060 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1065 val = arc_meta_limit;
1066 err = sysctl_handle_64(oidp, &val, 0, req);
1067 if (err != 0 || req->newptr == NULL)
1070 if (val <= 0 || val > arc_c_max)
1073 arc_meta_limit = val;
1078 sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS)
1083 val = arc_no_grow_shift;
1084 err = sysctl_handle_32(oidp, &val, 0, req);
1085 if (err != 0 || req->newptr == NULL)
1088 if (val >= arc_shrink_shift)
1091 arc_no_grow_shift = val;
1096 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1102 err = sysctl_handle_64(oidp, &val, 0, req);
1103 if (err != 0 || req->newptr == NULL)
1106 if (zfs_arc_max == 0) {
1107 /* Loader tunable so blindly set */
1112 if (val < arc_abs_min || val > kmem_size())
1114 if (val < arc_c_min)
1116 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1122 arc_p = (arc_c >> 1);
1124 if (zfs_arc_meta_limit == 0) {
1125 /* limit meta-data to 1/4 of the arc capacity */
1126 arc_meta_limit = arc_c_max / 4;
1129 /* if kmem_flags are set, lets try to use less memory */
1130 if (kmem_debugging())
1133 zfs_arc_max = arc_c;
1139 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1145 err = sysctl_handle_64(oidp, &val, 0, req);
1146 if (err != 0 || req->newptr == NULL)
1149 if (zfs_arc_min == 0) {
1150 /* Loader tunable so blindly set */
1155 if (val < arc_abs_min || val > arc_c_max)
1160 if (zfs_arc_meta_min == 0)
1161 arc_meta_min = arc_c_min / 2;
1163 if (arc_c < arc_c_min)
1166 zfs_arc_min = arc_c_min;
1172 #define GHOST_STATE(state) \
1173 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1174 (state) == arc_l2c_only)
1176 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1177 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1178 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1179 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1180 #define HDR_COMPRESSION_ENABLED(hdr) \
1181 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1183 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1184 #define HDR_L2_READING(hdr) \
1185 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1186 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1187 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1188 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1189 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1190 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1192 #define HDR_ISTYPE_METADATA(hdr) \
1193 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1194 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1196 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1197 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1199 /* For storing compression mode in b_flags */
1200 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1202 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1203 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1204 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1205 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1207 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1208 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1209 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1215 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1216 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1219 * Hash table routines
1222 #define HT_LOCK_PAD CACHE_LINE_SIZE
1227 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1231 #define BUF_LOCKS 256
1232 typedef struct buf_hash_table {
1234 arc_buf_hdr_t **ht_table;
1235 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1238 static buf_hash_table_t buf_hash_table;
1240 #define BUF_HASH_INDEX(spa, dva, birth) \
1241 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1242 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1243 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1244 #define HDR_LOCK(hdr) \
1245 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1247 uint64_t zfs_crc64_table[256];
1253 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1254 #define L2ARC_HEADROOM 2 /* num of writes */
1256 * If we discover during ARC scan any buffers to be compressed, we boost
1257 * our headroom for the next scanning cycle by this percentage multiple.
1259 #define L2ARC_HEADROOM_BOOST 200
1260 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1261 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1263 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1264 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1266 /* L2ARC Performance Tunables */
1267 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1268 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1269 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1270 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1271 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1272 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1273 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1274 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1275 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1277 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1278 &l2arc_write_max, 0, "max write size");
1279 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1280 &l2arc_write_boost, 0, "extra write during warmup");
1281 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1282 &l2arc_headroom, 0, "number of dev writes");
1283 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1284 &l2arc_feed_secs, 0, "interval seconds");
1285 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1286 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1288 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1289 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1290 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1291 &l2arc_feed_again, 0, "turbo warmup");
1292 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1293 &l2arc_norw, 0, "no reads during writes");
1295 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1296 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1297 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1298 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1299 "size of anonymous state");
1300 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1301 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1302 "size of anonymous state");
1304 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1305 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1306 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1307 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1308 "size of metadata in mru state");
1309 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1310 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1311 "size of data in mru state");
1313 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1314 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1315 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1316 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1317 "size of metadata in mru ghost state");
1318 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1319 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1320 "size of data in mru ghost state");
1322 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1323 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1324 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1325 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1326 "size of metadata in mfu state");
1327 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1328 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1329 "size of data in mfu state");
1331 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1332 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1333 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1334 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1335 "size of metadata in mfu ghost state");
1336 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1337 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1338 "size of data in mfu ghost state");
1340 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1341 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1347 vdev_t *l2ad_vdev; /* vdev */
1348 spa_t *l2ad_spa; /* spa */
1349 uint64_t l2ad_hand; /* next write location */
1350 uint64_t l2ad_start; /* first addr on device */
1351 uint64_t l2ad_end; /* last addr on device */
1352 boolean_t l2ad_first; /* first sweep through */
1353 boolean_t l2ad_writing; /* currently writing */
1354 kmutex_t l2ad_mtx; /* lock for buffer list */
1355 list_t l2ad_buflist; /* buffer list */
1356 list_node_t l2ad_node; /* device list node */
1357 refcount_t l2ad_alloc; /* allocated bytes */
1360 static list_t L2ARC_dev_list; /* device list */
1361 static list_t *l2arc_dev_list; /* device list pointer */
1362 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1363 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1364 static list_t L2ARC_free_on_write; /* free after write buf list */
1365 static list_t *l2arc_free_on_write; /* free after write list ptr */
1366 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1367 static uint64_t l2arc_ndev; /* number of devices */
1369 typedef struct l2arc_read_callback {
1370 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1371 blkptr_t l2rcb_bp; /* original blkptr */
1372 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1373 int l2rcb_flags; /* original flags */
1374 abd_t *l2rcb_abd; /* temporary buffer */
1375 } l2arc_read_callback_t;
1377 typedef struct l2arc_write_callback {
1378 l2arc_dev_t *l2wcb_dev; /* device info */
1379 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1380 } l2arc_write_callback_t;
1382 typedef struct l2arc_data_free {
1383 /* protected by l2arc_free_on_write_mtx */
1386 arc_buf_contents_t l2df_type;
1387 list_node_t l2df_list_node;
1388 } l2arc_data_free_t;
1390 static kmutex_t l2arc_feed_thr_lock;
1391 static kcondvar_t l2arc_feed_thr_cv;
1392 static uint8_t l2arc_thread_exit;
1394 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1395 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1396 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1397 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1398 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1399 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1400 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1401 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1402 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1403 static boolean_t arc_is_overflowing();
1404 static void arc_buf_watch(arc_buf_t *);
1406 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1407 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1408 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1409 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1411 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1412 static void l2arc_read_done(zio_t *);
1415 l2arc_trim(const arc_buf_hdr_t *hdr)
1417 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1419 ASSERT(HDR_HAS_L2HDR(hdr));
1420 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1422 if (HDR_GET_PSIZE(hdr) != 0) {
1423 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1424 HDR_GET_PSIZE(hdr), 0);
1429 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1431 uint8_t *vdva = (uint8_t *)dva;
1432 uint64_t crc = -1ULL;
1435 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1437 for (i = 0; i < sizeof (dva_t); i++)
1438 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1440 crc ^= (spa>>8) ^ birth;
1445 #define HDR_EMPTY(hdr) \
1446 ((hdr)->b_dva.dva_word[0] == 0 && \
1447 (hdr)->b_dva.dva_word[1] == 0)
1449 #define HDR_EQUAL(spa, dva, birth, hdr) \
1450 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1451 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1452 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1455 buf_discard_identity(arc_buf_hdr_t *hdr)
1457 hdr->b_dva.dva_word[0] = 0;
1458 hdr->b_dva.dva_word[1] = 0;
1462 static arc_buf_hdr_t *
1463 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1465 const dva_t *dva = BP_IDENTITY(bp);
1466 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1467 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1468 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1471 mutex_enter(hash_lock);
1472 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1473 hdr = hdr->b_hash_next) {
1474 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1479 mutex_exit(hash_lock);
1485 * Insert an entry into the hash table. If there is already an element
1486 * equal to elem in the hash table, then the already existing element
1487 * will be returned and the new element will not be inserted.
1488 * Otherwise returns NULL.
1489 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1491 static arc_buf_hdr_t *
1492 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1494 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1495 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1496 arc_buf_hdr_t *fhdr;
1499 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1500 ASSERT(hdr->b_birth != 0);
1501 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1503 if (lockp != NULL) {
1505 mutex_enter(hash_lock);
1507 ASSERT(MUTEX_HELD(hash_lock));
1510 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1511 fhdr = fhdr->b_hash_next, i++) {
1512 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1516 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1517 buf_hash_table.ht_table[idx] = hdr;
1518 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1520 /* collect some hash table performance data */
1522 ARCSTAT_BUMP(arcstat_hash_collisions);
1524 ARCSTAT_BUMP(arcstat_hash_chains);
1526 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1529 ARCSTAT_BUMP(arcstat_hash_elements);
1530 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1536 buf_hash_remove(arc_buf_hdr_t *hdr)
1538 arc_buf_hdr_t *fhdr, **hdrp;
1539 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1541 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1542 ASSERT(HDR_IN_HASH_TABLE(hdr));
1544 hdrp = &buf_hash_table.ht_table[idx];
1545 while ((fhdr = *hdrp) != hdr) {
1546 ASSERT3P(fhdr, !=, NULL);
1547 hdrp = &fhdr->b_hash_next;
1549 *hdrp = hdr->b_hash_next;
1550 hdr->b_hash_next = NULL;
1551 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1553 /* collect some hash table performance data */
1554 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1556 if (buf_hash_table.ht_table[idx] &&
1557 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1558 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1562 * Global data structures and functions for the buf kmem cache.
1564 static kmem_cache_t *hdr_full_cache;
1565 static kmem_cache_t *hdr_l2only_cache;
1566 static kmem_cache_t *buf_cache;
1573 kmem_free(buf_hash_table.ht_table,
1574 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1575 for (i = 0; i < BUF_LOCKS; i++)
1576 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1577 kmem_cache_destroy(hdr_full_cache);
1578 kmem_cache_destroy(hdr_l2only_cache);
1579 kmem_cache_destroy(buf_cache);
1583 * Constructor callback - called when the cache is empty
1584 * and a new buf is requested.
1588 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1590 arc_buf_hdr_t *hdr = vbuf;
1592 bzero(hdr, HDR_FULL_SIZE);
1593 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1594 refcount_create(&hdr->b_l1hdr.b_refcnt);
1595 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1596 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1597 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1604 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1606 arc_buf_hdr_t *hdr = vbuf;
1608 bzero(hdr, HDR_L2ONLY_SIZE);
1609 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1616 buf_cons(void *vbuf, void *unused, int kmflag)
1618 arc_buf_t *buf = vbuf;
1620 bzero(buf, sizeof (arc_buf_t));
1621 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1622 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1628 * Destructor callback - called when a cached buf is
1629 * no longer required.
1633 hdr_full_dest(void *vbuf, void *unused)
1635 arc_buf_hdr_t *hdr = vbuf;
1637 ASSERT(HDR_EMPTY(hdr));
1638 cv_destroy(&hdr->b_l1hdr.b_cv);
1639 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1640 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1641 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1642 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1647 hdr_l2only_dest(void *vbuf, void *unused)
1649 arc_buf_hdr_t *hdr = vbuf;
1651 ASSERT(HDR_EMPTY(hdr));
1652 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1657 buf_dest(void *vbuf, void *unused)
1659 arc_buf_t *buf = vbuf;
1661 mutex_destroy(&buf->b_evict_lock);
1662 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1666 * Reclaim callback -- invoked when memory is low.
1670 hdr_recl(void *unused)
1672 dprintf("hdr_recl called\n");
1674 * umem calls the reclaim func when we destroy the buf cache,
1675 * which is after we do arc_fini().
1678 cv_signal(&arc_reclaim_thread_cv);
1685 uint64_t hsize = 1ULL << 12;
1689 * The hash table is big enough to fill all of physical memory
1690 * with an average block size of zfs_arc_average_blocksize (default 8K).
1691 * By default, the table will take up
1692 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1694 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1697 buf_hash_table.ht_mask = hsize - 1;
1698 buf_hash_table.ht_table =
1699 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1700 if (buf_hash_table.ht_table == NULL) {
1701 ASSERT(hsize > (1ULL << 8));
1706 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1707 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1708 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1709 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1711 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1712 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1714 for (i = 0; i < 256; i++)
1715 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1716 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1718 for (i = 0; i < BUF_LOCKS; i++) {
1719 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1720 NULL, MUTEX_DEFAULT, NULL);
1725 * This is the size that the buf occupies in memory. If the buf is compressed,
1726 * it will correspond to the compressed size. You should use this method of
1727 * getting the buf size unless you explicitly need the logical size.
1730 arc_buf_size(arc_buf_t *buf)
1732 return (ARC_BUF_COMPRESSED(buf) ?
1733 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1737 arc_buf_lsize(arc_buf_t *buf)
1739 return (HDR_GET_LSIZE(buf->b_hdr));
1743 arc_get_compression(arc_buf_t *buf)
1745 return (ARC_BUF_COMPRESSED(buf) ?
1746 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1749 #define ARC_MINTIME (hz>>4) /* 62 ms */
1751 static inline boolean_t
1752 arc_buf_is_shared(arc_buf_t *buf)
1754 boolean_t shared = (buf->b_data != NULL &&
1755 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1756 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1757 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1758 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1759 IMPLY(shared, ARC_BUF_SHARED(buf));
1760 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1763 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1764 * already being shared" requirement prevents us from doing that.
1771 * Free the checksum associated with this header. If there is no checksum, this
1775 arc_cksum_free(arc_buf_hdr_t *hdr)
1777 ASSERT(HDR_HAS_L1HDR(hdr));
1778 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1779 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1780 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1781 hdr->b_l1hdr.b_freeze_cksum = NULL;
1783 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1787 * Return true iff at least one of the bufs on hdr is not compressed.
1790 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1792 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1793 if (!ARC_BUF_COMPRESSED(b)) {
1801 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1802 * matches the checksum that is stored in the hdr. If there is no checksum,
1803 * or if the buf is compressed, this is a no-op.
1806 arc_cksum_verify(arc_buf_t *buf)
1808 arc_buf_hdr_t *hdr = buf->b_hdr;
1811 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1814 if (ARC_BUF_COMPRESSED(buf)) {
1815 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1816 arc_hdr_has_uncompressed_buf(hdr));
1820 ASSERT(HDR_HAS_L1HDR(hdr));
1822 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1823 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1824 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1828 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1829 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1830 panic("buffer modified while frozen!");
1831 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1835 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1837 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1838 boolean_t valid_cksum;
1840 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1841 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1844 * We rely on the blkptr's checksum to determine if the block
1845 * is valid or not. When compressed arc is enabled, the l2arc
1846 * writes the block to the l2arc just as it appears in the pool.
1847 * This allows us to use the blkptr's checksum to validate the
1848 * data that we just read off of the l2arc without having to store
1849 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1850 * arc is disabled, then the data written to the l2arc is always
1851 * uncompressed and won't match the block as it exists in the main
1852 * pool. When this is the case, we must first compress it if it is
1853 * compressed on the main pool before we can validate the checksum.
1855 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1856 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1857 uint64_t lsize = HDR_GET_LSIZE(hdr);
1860 void *cbuf = zio_buf_alloc(HDR_GET_PSIZE(hdr));
1861 csize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
1863 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1864 if (csize < HDR_GET_PSIZE(hdr)) {
1866 * Compressed blocks are always a multiple of the
1867 * smallest ashift in the pool. Ideally, we would
1868 * like to round up the csize to the next
1869 * spa_min_ashift but that value may have changed
1870 * since the block was last written. Instead,
1871 * we rely on the fact that the hdr's psize
1872 * was set to the psize of the block when it was
1873 * last written. We set the csize to that value
1874 * and zero out any part that should not contain
1877 bzero((char *)cbuf + csize, HDR_GET_PSIZE(hdr) - csize);
1878 csize = HDR_GET_PSIZE(hdr);
1880 zio_push_transform(zio, cbuf, csize, HDR_GET_PSIZE(hdr), NULL);
1884 * Block pointers always store the checksum for the logical data.
1885 * If the block pointer has the gang bit set, then the checksum
1886 * it represents is for the reconstituted data and not for an
1887 * individual gang member. The zio pipeline, however, must be able to
1888 * determine the checksum of each of the gang constituents so it
1889 * treats the checksum comparison differently than what we need
1890 * for l2arc blocks. This prevents us from using the
1891 * zio_checksum_error() interface directly. Instead we must call the
1892 * zio_checksum_error_impl() so that we can ensure the checksum is
1893 * generated using the correct checksum algorithm and accounts for the
1894 * logical I/O size and not just a gang fragment.
1896 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1897 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1898 zio->io_offset, NULL) == 0);
1899 zio_pop_transforms(zio);
1900 return (valid_cksum);
1904 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1905 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1906 * isn't modified later on. If buf is compressed or there is already a checksum
1907 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1910 arc_cksum_compute(arc_buf_t *buf)
1912 arc_buf_hdr_t *hdr = buf->b_hdr;
1914 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1917 ASSERT(HDR_HAS_L1HDR(hdr));
1919 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1920 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1921 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1922 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1924 } else if (ARC_BUF_COMPRESSED(buf)) {
1925 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1929 ASSERT(!ARC_BUF_COMPRESSED(buf));
1930 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1932 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1933 hdr->b_l1hdr.b_freeze_cksum);
1934 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1942 typedef struct procctl {
1950 arc_buf_unwatch(arc_buf_t *buf)
1957 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1958 ctl.prwatch.pr_size = 0;
1959 ctl.prwatch.pr_wflags = 0;
1960 result = write(arc_procfd, &ctl, sizeof (ctl));
1961 ASSERT3U(result, ==, sizeof (ctl));
1968 arc_buf_watch(arc_buf_t *buf)
1975 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1976 ctl.prwatch.pr_size = arc_buf_size(buf);
1977 ctl.prwatch.pr_wflags = WA_WRITE;
1978 result = write(arc_procfd, &ctl, sizeof (ctl));
1979 ASSERT3U(result, ==, sizeof (ctl));
1983 #endif /* illumos */
1985 static arc_buf_contents_t
1986 arc_buf_type(arc_buf_hdr_t *hdr)
1988 arc_buf_contents_t type;
1989 if (HDR_ISTYPE_METADATA(hdr)) {
1990 type = ARC_BUFC_METADATA;
1992 type = ARC_BUFC_DATA;
1994 VERIFY3U(hdr->b_type, ==, type);
1999 arc_is_metadata(arc_buf_t *buf)
2001 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2005 arc_bufc_to_flags(arc_buf_contents_t type)
2009 /* metadata field is 0 if buffer contains normal data */
2011 case ARC_BUFC_METADATA:
2012 return (ARC_FLAG_BUFC_METADATA);
2016 panic("undefined ARC buffer type!");
2017 return ((uint32_t)-1);
2021 arc_buf_thaw(arc_buf_t *buf)
2023 arc_buf_hdr_t *hdr = buf->b_hdr;
2025 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2026 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2028 arc_cksum_verify(buf);
2031 * Compressed buffers do not manipulate the b_freeze_cksum or
2032 * allocate b_thawed.
2034 if (ARC_BUF_COMPRESSED(buf)) {
2035 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2036 arc_hdr_has_uncompressed_buf(hdr));
2040 ASSERT(HDR_HAS_L1HDR(hdr));
2041 arc_cksum_free(hdr);
2043 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2045 if (zfs_flags & ZFS_DEBUG_MODIFY) {
2046 if (hdr->b_l1hdr.b_thawed != NULL)
2047 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2048 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2052 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2055 arc_buf_unwatch(buf);
2060 arc_buf_freeze(arc_buf_t *buf)
2062 arc_buf_hdr_t *hdr = buf->b_hdr;
2063 kmutex_t *hash_lock;
2065 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2068 if (ARC_BUF_COMPRESSED(buf)) {
2069 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2070 arc_hdr_has_uncompressed_buf(hdr));
2074 hash_lock = HDR_LOCK(hdr);
2075 mutex_enter(hash_lock);
2077 ASSERT(HDR_HAS_L1HDR(hdr));
2078 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
2079 hdr->b_l1hdr.b_state == arc_anon);
2080 arc_cksum_compute(buf);
2081 mutex_exit(hash_lock);
2085 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2086 * the following functions should be used to ensure that the flags are
2087 * updated in a thread-safe way. When manipulating the flags either
2088 * the hash_lock must be held or the hdr must be undiscoverable. This
2089 * ensures that we're not racing with any other threads when updating
2093 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2095 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2096 hdr->b_flags |= flags;
2100 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2102 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2103 hdr->b_flags &= ~flags;
2107 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2108 * done in a special way since we have to clear and set bits
2109 * at the same time. Consumers that wish to set the compression bits
2110 * must use this function to ensure that the flags are updated in
2111 * thread-safe manner.
2114 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2116 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2119 * Holes and embedded blocks will always have a psize = 0 so
2120 * we ignore the compression of the blkptr and set the
2121 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2122 * Holes and embedded blocks remain anonymous so we don't
2123 * want to uncompress them. Mark them as uncompressed.
2125 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2126 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2127 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2128 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2129 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2131 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2132 HDR_SET_COMPRESS(hdr, cmp);
2133 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2134 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2139 * Looks for another buf on the same hdr which has the data decompressed, copies
2140 * from it, and returns true. If no such buf exists, returns false.
2143 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
2145 arc_buf_hdr_t *hdr = buf->b_hdr;
2146 boolean_t copied = B_FALSE;
2148 ASSERT(HDR_HAS_L1HDR(hdr));
2149 ASSERT3P(buf->b_data, !=, NULL);
2150 ASSERT(!ARC_BUF_COMPRESSED(buf));
2152 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2153 from = from->b_next) {
2154 /* can't use our own data buffer */
2159 if (!ARC_BUF_COMPRESSED(from)) {
2160 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2167 * There were no decompressed bufs, so there should not be a
2168 * checksum on the hdr either.
2170 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
2176 * Given a buf that has a data buffer attached to it, this function will
2177 * efficiently fill the buf with data of the specified compression setting from
2178 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2179 * are already sharing a data buf, no copy is performed.
2181 * If the buf is marked as compressed but uncompressed data was requested, this
2182 * will allocate a new data buffer for the buf, remove that flag, and fill the
2183 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2184 * uncompressed data, and (since we haven't added support for it yet) if you
2185 * want compressed data your buf must already be marked as compressed and have
2186 * the correct-sized data buffer.
2189 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2191 arc_buf_hdr_t *hdr = buf->b_hdr;
2192 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2193 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2195 ASSERT3P(buf->b_data, !=, NULL);
2196 IMPLY(compressed, hdr_compressed);
2197 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2199 if (hdr_compressed == compressed) {
2200 if (!arc_buf_is_shared(buf)) {
2201 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2205 ASSERT(hdr_compressed);
2206 ASSERT(!compressed);
2207 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2210 * If the buf is sharing its data with the hdr, unlink it and
2211 * allocate a new data buffer for the buf.
2213 if (arc_buf_is_shared(buf)) {
2214 ASSERT(ARC_BUF_COMPRESSED(buf));
2216 /* We need to give the buf it's own b_data */
2217 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2219 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2220 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2222 /* Previously overhead was 0; just add new overhead */
2223 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2224 } else if (ARC_BUF_COMPRESSED(buf)) {
2225 /* We need to reallocate the buf's b_data */
2226 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2229 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2231 /* We increased the size of b_data; update overhead */
2232 ARCSTAT_INCR(arcstat_overhead_size,
2233 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2237 * Regardless of the buf's previous compression settings, it
2238 * should not be compressed at the end of this function.
2240 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2243 * Try copying the data from another buf which already has a
2244 * decompressed version. If that's not possible, it's time to
2245 * bite the bullet and decompress the data from the hdr.
2247 if (arc_buf_try_copy_decompressed_data(buf)) {
2248 /* Skip byteswapping and checksumming (already done) */
2249 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2252 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2253 hdr->b_l1hdr.b_pabd, buf->b_data,
2254 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2257 * Absent hardware errors or software bugs, this should
2258 * be impossible, but log it anyway so we can debug it.
2262 "hdr %p, compress %d, psize %d, lsize %d",
2263 hdr, HDR_GET_COMPRESS(hdr),
2264 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2265 return (SET_ERROR(EIO));
2270 /* Byteswap the buf's data if necessary */
2271 if (bswap != DMU_BSWAP_NUMFUNCS) {
2272 ASSERT(!HDR_SHARED_DATA(hdr));
2273 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2274 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2277 /* Compute the hdr's checksum if necessary */
2278 arc_cksum_compute(buf);
2284 arc_decompress(arc_buf_t *buf)
2286 return (arc_buf_fill(buf, B_FALSE));
2290 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2293 arc_hdr_size(arc_buf_hdr_t *hdr)
2297 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2298 HDR_GET_PSIZE(hdr) > 0) {
2299 size = HDR_GET_PSIZE(hdr);
2301 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2302 size = HDR_GET_LSIZE(hdr);
2308 * Increment the amount of evictable space in the arc_state_t's refcount.
2309 * We account for the space used by the hdr and the arc buf individually
2310 * so that we can add and remove them from the refcount individually.
2313 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2315 arc_buf_contents_t type = arc_buf_type(hdr);
2317 ASSERT(HDR_HAS_L1HDR(hdr));
2319 if (GHOST_STATE(state)) {
2320 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2321 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2322 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2323 (void) refcount_add_many(&state->arcs_esize[type],
2324 HDR_GET_LSIZE(hdr), hdr);
2328 ASSERT(!GHOST_STATE(state));
2329 if (hdr->b_l1hdr.b_pabd != NULL) {
2330 (void) refcount_add_many(&state->arcs_esize[type],
2331 arc_hdr_size(hdr), hdr);
2333 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2334 buf = buf->b_next) {
2335 if (arc_buf_is_shared(buf))
2337 (void) refcount_add_many(&state->arcs_esize[type],
2338 arc_buf_size(buf), buf);
2343 * Decrement the amount of evictable space in the arc_state_t's refcount.
2344 * We account for the space used by the hdr and the arc buf individually
2345 * so that we can add and remove them from the refcount individually.
2348 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2350 arc_buf_contents_t type = arc_buf_type(hdr);
2352 ASSERT(HDR_HAS_L1HDR(hdr));
2354 if (GHOST_STATE(state)) {
2355 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2356 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2357 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2358 (void) refcount_remove_many(&state->arcs_esize[type],
2359 HDR_GET_LSIZE(hdr), hdr);
2363 ASSERT(!GHOST_STATE(state));
2364 if (hdr->b_l1hdr.b_pabd != NULL) {
2365 (void) refcount_remove_many(&state->arcs_esize[type],
2366 arc_hdr_size(hdr), hdr);
2368 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2369 buf = buf->b_next) {
2370 if (arc_buf_is_shared(buf))
2372 (void) refcount_remove_many(&state->arcs_esize[type],
2373 arc_buf_size(buf), buf);
2378 * Add a reference to this hdr indicating that someone is actively
2379 * referencing that memory. When the refcount transitions from 0 to 1,
2380 * we remove it from the respective arc_state_t list to indicate that
2381 * it is not evictable.
2384 add_reference(arc_buf_hdr_t *hdr, void *tag)
2386 ASSERT(HDR_HAS_L1HDR(hdr));
2387 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2388 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2389 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2390 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2393 arc_state_t *state = hdr->b_l1hdr.b_state;
2395 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2396 (state != arc_anon)) {
2397 /* We don't use the L2-only state list. */
2398 if (state != arc_l2c_only) {
2399 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2401 arc_evictable_space_decrement(hdr, state);
2403 /* remove the prefetch flag if we get a reference */
2404 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2409 * Remove a reference from this hdr. When the reference transitions from
2410 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2411 * list making it eligible for eviction.
2414 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2417 arc_state_t *state = hdr->b_l1hdr.b_state;
2419 ASSERT(HDR_HAS_L1HDR(hdr));
2420 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2421 ASSERT(!GHOST_STATE(state));
2424 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2425 * check to prevent usage of the arc_l2c_only list.
2427 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2428 (state != arc_anon)) {
2429 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2430 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2431 arc_evictable_space_increment(hdr, state);
2437 * Move the supplied buffer to the indicated state. The hash lock
2438 * for the buffer must be held by the caller.
2441 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2442 kmutex_t *hash_lock)
2444 arc_state_t *old_state;
2447 boolean_t update_old, update_new;
2448 arc_buf_contents_t buftype = arc_buf_type(hdr);
2451 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2452 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2453 * L1 hdr doesn't always exist when we change state to arc_anon before
2454 * destroying a header, in which case reallocating to add the L1 hdr is
2457 if (HDR_HAS_L1HDR(hdr)) {
2458 old_state = hdr->b_l1hdr.b_state;
2459 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2460 bufcnt = hdr->b_l1hdr.b_bufcnt;
2461 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2463 old_state = arc_l2c_only;
2466 update_old = B_FALSE;
2468 update_new = update_old;
2470 ASSERT(MUTEX_HELD(hash_lock));
2471 ASSERT3P(new_state, !=, old_state);
2472 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2473 ASSERT(old_state != arc_anon || bufcnt <= 1);
2476 * If this buffer is evictable, transfer it from the
2477 * old state list to the new state list.
2480 if (old_state != arc_anon && old_state != arc_l2c_only) {
2481 ASSERT(HDR_HAS_L1HDR(hdr));
2482 multilist_remove(old_state->arcs_list[buftype], hdr);
2484 if (GHOST_STATE(old_state)) {
2486 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2487 update_old = B_TRUE;
2489 arc_evictable_space_decrement(hdr, old_state);
2491 if (new_state != arc_anon && new_state != arc_l2c_only) {
2494 * An L1 header always exists here, since if we're
2495 * moving to some L1-cached state (i.e. not l2c_only or
2496 * anonymous), we realloc the header to add an L1hdr
2499 ASSERT(HDR_HAS_L1HDR(hdr));
2500 multilist_insert(new_state->arcs_list[buftype], hdr);
2502 if (GHOST_STATE(new_state)) {
2504 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2505 update_new = B_TRUE;
2507 arc_evictable_space_increment(hdr, new_state);
2511 ASSERT(!HDR_EMPTY(hdr));
2512 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2513 buf_hash_remove(hdr);
2515 /* adjust state sizes (ignore arc_l2c_only) */
2517 if (update_new && new_state != arc_l2c_only) {
2518 ASSERT(HDR_HAS_L1HDR(hdr));
2519 if (GHOST_STATE(new_state)) {
2523 * When moving a header to a ghost state, we first
2524 * remove all arc buffers. Thus, we'll have a
2525 * bufcnt of zero, and no arc buffer to use for
2526 * the reference. As a result, we use the arc
2527 * header pointer for the reference.
2529 (void) refcount_add_many(&new_state->arcs_size,
2530 HDR_GET_LSIZE(hdr), hdr);
2531 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2533 uint32_t buffers = 0;
2536 * Each individual buffer holds a unique reference,
2537 * thus we must remove each of these references one
2540 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2541 buf = buf->b_next) {
2542 ASSERT3U(bufcnt, !=, 0);
2546 * When the arc_buf_t is sharing the data
2547 * block with the hdr, the owner of the
2548 * reference belongs to the hdr. Only
2549 * add to the refcount if the arc_buf_t is
2552 if (arc_buf_is_shared(buf))
2555 (void) refcount_add_many(&new_state->arcs_size,
2556 arc_buf_size(buf), buf);
2558 ASSERT3U(bufcnt, ==, buffers);
2560 if (hdr->b_l1hdr.b_pabd != NULL) {
2561 (void) refcount_add_many(&new_state->arcs_size,
2562 arc_hdr_size(hdr), hdr);
2564 ASSERT(GHOST_STATE(old_state));
2569 if (update_old && old_state != arc_l2c_only) {
2570 ASSERT(HDR_HAS_L1HDR(hdr));
2571 if (GHOST_STATE(old_state)) {
2573 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2576 * When moving a header off of a ghost state,
2577 * the header will not contain any arc buffers.
2578 * We use the arc header pointer for the reference
2579 * which is exactly what we did when we put the
2580 * header on the ghost state.
2583 (void) refcount_remove_many(&old_state->arcs_size,
2584 HDR_GET_LSIZE(hdr), hdr);
2586 uint32_t buffers = 0;
2589 * Each individual buffer holds a unique reference,
2590 * thus we must remove each of these references one
2593 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2594 buf = buf->b_next) {
2595 ASSERT3U(bufcnt, !=, 0);
2599 * When the arc_buf_t is sharing the data
2600 * block with the hdr, the owner of the
2601 * reference belongs to the hdr. Only
2602 * add to the refcount if the arc_buf_t is
2605 if (arc_buf_is_shared(buf))
2608 (void) refcount_remove_many(
2609 &old_state->arcs_size, arc_buf_size(buf),
2612 ASSERT3U(bufcnt, ==, buffers);
2613 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2614 (void) refcount_remove_many(
2615 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2619 if (HDR_HAS_L1HDR(hdr))
2620 hdr->b_l1hdr.b_state = new_state;
2623 * L2 headers should never be on the L2 state list since they don't
2624 * have L1 headers allocated.
2626 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2627 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2631 arc_space_consume(uint64_t space, arc_space_type_t type)
2633 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2636 case ARC_SPACE_DATA:
2637 ARCSTAT_INCR(arcstat_data_size, space);
2639 case ARC_SPACE_META:
2640 ARCSTAT_INCR(arcstat_metadata_size, space);
2642 case ARC_SPACE_OTHER:
2643 ARCSTAT_INCR(arcstat_other_size, space);
2645 case ARC_SPACE_HDRS:
2646 ARCSTAT_INCR(arcstat_hdr_size, space);
2648 case ARC_SPACE_L2HDRS:
2649 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
2653 if (type != ARC_SPACE_DATA)
2654 ARCSTAT_INCR(arcstat_meta_used, space);
2656 atomic_add_64(&arc_size, space);
2660 arc_space_return(uint64_t space, arc_space_type_t type)
2662 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2665 case ARC_SPACE_DATA:
2666 ARCSTAT_INCR(arcstat_data_size, -space);
2668 case ARC_SPACE_META:
2669 ARCSTAT_INCR(arcstat_metadata_size, -space);
2671 case ARC_SPACE_OTHER:
2672 ARCSTAT_INCR(arcstat_other_size, -space);
2674 case ARC_SPACE_HDRS:
2675 ARCSTAT_INCR(arcstat_hdr_size, -space);
2677 case ARC_SPACE_L2HDRS:
2678 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
2682 if (type != ARC_SPACE_DATA) {
2683 ASSERT(arc_meta_used >= space);
2684 if (arc_meta_max < arc_meta_used)
2685 arc_meta_max = arc_meta_used;
2686 ARCSTAT_INCR(arcstat_meta_used, -space);
2689 ASSERT(arc_size >= space);
2690 atomic_add_64(&arc_size, -space);
2694 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2695 * with the hdr's b_pabd.
2698 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2701 * The criteria for sharing a hdr's data are:
2702 * 1. the hdr's compression matches the buf's compression
2703 * 2. the hdr doesn't need to be byteswapped
2704 * 3. the hdr isn't already being shared
2705 * 4. the buf is either compressed or it is the last buf in the hdr list
2707 * Criterion #4 maintains the invariant that shared uncompressed
2708 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2709 * might ask, "if a compressed buf is allocated first, won't that be the
2710 * last thing in the list?", but in that case it's impossible to create
2711 * a shared uncompressed buf anyway (because the hdr must be compressed
2712 * to have the compressed buf). You might also think that #3 is
2713 * sufficient to make this guarantee, however it's possible
2714 * (specifically in the rare L2ARC write race mentioned in
2715 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2716 * is sharable, but wasn't at the time of its allocation. Rather than
2717 * allow a new shared uncompressed buf to be created and then shuffle
2718 * the list around to make it the last element, this simply disallows
2719 * sharing if the new buf isn't the first to be added.
2721 ASSERT3P(buf->b_hdr, ==, hdr);
2722 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2723 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2724 return (buf_compressed == hdr_compressed &&
2725 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2726 !HDR_SHARED_DATA(hdr) &&
2727 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2731 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2732 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2733 * copy was made successfully, or an error code otherwise.
2736 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2737 boolean_t fill, arc_buf_t **ret)
2741 ASSERT(HDR_HAS_L1HDR(hdr));
2742 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2743 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2744 hdr->b_type == ARC_BUFC_METADATA);
2745 ASSERT3P(ret, !=, NULL);
2746 ASSERT3P(*ret, ==, NULL);
2748 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2751 buf->b_next = hdr->b_l1hdr.b_buf;
2754 add_reference(hdr, tag);
2757 * We're about to change the hdr's b_flags. We must either
2758 * hold the hash_lock or be undiscoverable.
2760 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2763 * Only honor requests for compressed bufs if the hdr is actually
2766 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2767 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2770 * If the hdr's data can be shared then we share the data buffer and
2771 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2772 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2773 * buffer to store the buf's data.
2775 * There are two additional restrictions here because we're sharing
2776 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2777 * actively involved in an L2ARC write, because if this buf is used by
2778 * an arc_write() then the hdr's data buffer will be released when the
2779 * write completes, even though the L2ARC write might still be using it.
2780 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2781 * need to be ABD-aware.
2783 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2784 abd_is_linear(hdr->b_l1hdr.b_pabd);
2786 /* Set up b_data and sharing */
2788 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2789 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2790 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2793 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2794 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2796 VERIFY3P(buf->b_data, !=, NULL);
2798 hdr->b_l1hdr.b_buf = buf;
2799 hdr->b_l1hdr.b_bufcnt += 1;
2802 * If the user wants the data from the hdr, we need to either copy or
2803 * decompress the data.
2806 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2812 static char *arc_onloan_tag = "onloan";
2815 arc_loaned_bytes_update(int64_t delta)
2817 atomic_add_64(&arc_loaned_bytes, delta);
2819 /* assert that it did not wrap around */
2820 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2824 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2825 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2826 * buffers must be returned to the arc before they can be used by the DMU or
2830 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2832 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2833 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2835 arc_loaned_bytes_update(size);
2841 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2842 enum zio_compress compression_type)
2844 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2845 psize, lsize, compression_type);
2847 arc_loaned_bytes_update(psize);
2854 * Return a loaned arc buffer to the arc.
2857 arc_return_buf(arc_buf_t *buf, void *tag)
2859 arc_buf_hdr_t *hdr = buf->b_hdr;
2861 ASSERT3P(buf->b_data, !=, NULL);
2862 ASSERT(HDR_HAS_L1HDR(hdr));
2863 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2864 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2866 arc_loaned_bytes_update(-arc_buf_size(buf));
2869 /* Detach an arc_buf from a dbuf (tag) */
2871 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2873 arc_buf_hdr_t *hdr = buf->b_hdr;
2875 ASSERT3P(buf->b_data, !=, NULL);
2876 ASSERT(HDR_HAS_L1HDR(hdr));
2877 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2878 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2880 arc_loaned_bytes_update(arc_buf_size(buf));
2884 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2886 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2889 df->l2df_size = size;
2890 df->l2df_type = type;
2891 mutex_enter(&l2arc_free_on_write_mtx);
2892 list_insert_head(l2arc_free_on_write, df);
2893 mutex_exit(&l2arc_free_on_write_mtx);
2897 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2899 arc_state_t *state = hdr->b_l1hdr.b_state;
2900 arc_buf_contents_t type = arc_buf_type(hdr);
2901 uint64_t size = arc_hdr_size(hdr);
2903 /* protected by hash lock, if in the hash table */
2904 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2905 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2906 ASSERT(state != arc_anon && state != arc_l2c_only);
2908 (void) refcount_remove_many(&state->arcs_esize[type],
2911 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2912 if (type == ARC_BUFC_METADATA) {
2913 arc_space_return(size, ARC_SPACE_META);
2915 ASSERT(type == ARC_BUFC_DATA);
2916 arc_space_return(size, ARC_SPACE_DATA);
2919 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2923 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2924 * data buffer, we transfer the refcount ownership to the hdr and update
2925 * the appropriate kstats.
2928 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2930 arc_state_t *state = hdr->b_l1hdr.b_state;
2932 ASSERT(arc_can_share(hdr, buf));
2933 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2934 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2937 * Start sharing the data buffer. We transfer the
2938 * refcount ownership to the hdr since it always owns
2939 * the refcount whenever an arc_buf_t is shared.
2941 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2942 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2943 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2944 HDR_ISTYPE_METADATA(hdr));
2945 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2946 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2949 * Since we've transferred ownership to the hdr we need
2950 * to increment its compressed and uncompressed kstats and
2951 * decrement the overhead size.
2953 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2954 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2955 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2959 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2961 arc_state_t *state = hdr->b_l1hdr.b_state;
2963 ASSERT(arc_buf_is_shared(buf));
2964 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2965 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2968 * We are no longer sharing this buffer so we need
2969 * to transfer its ownership to the rightful owner.
2971 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
2972 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2973 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
2974 abd_put(hdr->b_l1hdr.b_pabd);
2975 hdr->b_l1hdr.b_pabd = NULL;
2976 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2979 * Since the buffer is no longer shared between
2980 * the arc buf and the hdr, count it as overhead.
2982 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2983 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2984 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2988 * Remove an arc_buf_t from the hdr's buf list and return the last
2989 * arc_buf_t on the list. If no buffers remain on the list then return
2993 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2995 ASSERT(HDR_HAS_L1HDR(hdr));
2996 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2998 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
2999 arc_buf_t *lastbuf = NULL;
3002 * Remove the buf from the hdr list and locate the last
3003 * remaining buffer on the list.
3005 while (*bufp != NULL) {
3007 *bufp = buf->b_next;
3010 * If we've removed a buffer in the middle of
3011 * the list then update the lastbuf and update
3014 if (*bufp != NULL) {
3016 bufp = &(*bufp)->b_next;
3020 ASSERT3P(lastbuf, !=, buf);
3021 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3022 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3023 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3029 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3033 arc_buf_destroy_impl(arc_buf_t *buf)
3035 arc_buf_hdr_t *hdr = buf->b_hdr;
3038 * Free up the data associated with the buf but only if we're not
3039 * sharing this with the hdr. If we are sharing it with the hdr, the
3040 * hdr is responsible for doing the free.
3042 if (buf->b_data != NULL) {
3044 * We're about to change the hdr's b_flags. We must either
3045 * hold the hash_lock or be undiscoverable.
3047 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3049 arc_cksum_verify(buf);
3051 arc_buf_unwatch(buf);
3054 if (arc_buf_is_shared(buf)) {
3055 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3057 uint64_t size = arc_buf_size(buf);
3058 arc_free_data_buf(hdr, buf->b_data, size, buf);
3059 ARCSTAT_INCR(arcstat_overhead_size, -size);
3063 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3064 hdr->b_l1hdr.b_bufcnt -= 1;
3067 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3069 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3071 * If the current arc_buf_t is sharing its data buffer with the
3072 * hdr, then reassign the hdr's b_pabd to share it with the new
3073 * buffer at the end of the list. The shared buffer is always
3074 * the last one on the hdr's buffer list.
3076 * There is an equivalent case for compressed bufs, but since
3077 * they aren't guaranteed to be the last buf in the list and
3078 * that is an exceedingly rare case, we just allow that space be
3079 * wasted temporarily.
3081 if (lastbuf != NULL) {
3082 /* Only one buf can be shared at once */
3083 VERIFY(!arc_buf_is_shared(lastbuf));
3084 /* hdr is uncompressed so can't have compressed buf */
3085 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3087 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3088 arc_hdr_free_pabd(hdr);
3091 * We must setup a new shared block between the
3092 * last buffer and the hdr. The data would have
3093 * been allocated by the arc buf so we need to transfer
3094 * ownership to the hdr since it's now being shared.
3096 arc_share_buf(hdr, lastbuf);
3098 } else if (HDR_SHARED_DATA(hdr)) {
3100 * Uncompressed shared buffers are always at the end
3101 * of the list. Compressed buffers don't have the
3102 * same requirements. This makes it hard to
3103 * simply assert that the lastbuf is shared so
3104 * we rely on the hdr's compression flags to determine
3105 * if we have a compressed, shared buffer.
3107 ASSERT3P(lastbuf, !=, NULL);
3108 ASSERT(arc_buf_is_shared(lastbuf) ||
3109 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
3113 * Free the checksum if we're removing the last uncompressed buf from
3116 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3117 arc_cksum_free(hdr);
3120 /* clean up the buf */
3122 kmem_cache_free(buf_cache, buf);
3126 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
3128 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3129 ASSERT(HDR_HAS_L1HDR(hdr));
3130 ASSERT(!HDR_SHARED_DATA(hdr));
3132 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3133 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
3134 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3135 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3137 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3138 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3142 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3144 ASSERT(HDR_HAS_L1HDR(hdr));
3145 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3148 * If the hdr is currently being written to the l2arc then
3149 * we defer freeing the data by adding it to the l2arc_free_on_write
3150 * list. The l2arc will free the data once it's finished
3151 * writing it to the l2arc device.
3153 if (HDR_L2_WRITING(hdr)) {
3154 arc_hdr_free_on_write(hdr);
3155 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3157 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3158 arc_hdr_size(hdr), hdr);
3160 hdr->b_l1hdr.b_pabd = NULL;
3161 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3163 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3164 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3167 static arc_buf_hdr_t *
3168 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3169 enum zio_compress compression_type, arc_buf_contents_t type)
3173 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3175 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3176 ASSERT(HDR_EMPTY(hdr));
3177 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3178 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3179 HDR_SET_PSIZE(hdr, psize);
3180 HDR_SET_LSIZE(hdr, lsize);
3184 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3185 arc_hdr_set_compress(hdr, compression_type);
3187 hdr->b_l1hdr.b_state = arc_anon;
3188 hdr->b_l1hdr.b_arc_access = 0;
3189 hdr->b_l1hdr.b_bufcnt = 0;
3190 hdr->b_l1hdr.b_buf = NULL;
3193 * Allocate the hdr's buffer. This will contain either
3194 * the compressed or uncompressed data depending on the block
3195 * it references and compressed arc enablement.
3197 arc_hdr_alloc_pabd(hdr);
3198 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3204 * Transition between the two allocation states for the arc_buf_hdr struct.
3205 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3206 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3207 * version is used when a cache buffer is only in the L2ARC in order to reduce
3210 static arc_buf_hdr_t *
3211 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3213 ASSERT(HDR_HAS_L2HDR(hdr));
3215 arc_buf_hdr_t *nhdr;
3216 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3218 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3219 (old == hdr_l2only_cache && new == hdr_full_cache));
3221 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3223 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3224 buf_hash_remove(hdr);
3226 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3228 if (new == hdr_full_cache) {
3229 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3231 * arc_access and arc_change_state need to be aware that a
3232 * header has just come out of L2ARC, so we set its state to
3233 * l2c_only even though it's about to change.
3235 nhdr->b_l1hdr.b_state = arc_l2c_only;
3237 /* Verify previous threads set to NULL before freeing */
3238 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3240 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3241 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3242 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3245 * If we've reached here, We must have been called from
3246 * arc_evict_hdr(), as such we should have already been
3247 * removed from any ghost list we were previously on
3248 * (which protects us from racing with arc_evict_state),
3249 * thus no locking is needed during this check.
3251 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3254 * A buffer must not be moved into the arc_l2c_only
3255 * state if it's not finished being written out to the
3256 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3257 * might try to be accessed, even though it was removed.
3259 VERIFY(!HDR_L2_WRITING(hdr));
3260 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3263 if (hdr->b_l1hdr.b_thawed != NULL) {
3264 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3265 hdr->b_l1hdr.b_thawed = NULL;
3269 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3272 * The header has been reallocated so we need to re-insert it into any
3275 (void) buf_hash_insert(nhdr, NULL);
3277 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3279 mutex_enter(&dev->l2ad_mtx);
3282 * We must place the realloc'ed header back into the list at
3283 * the same spot. Otherwise, if it's placed earlier in the list,
3284 * l2arc_write_buffers() could find it during the function's
3285 * write phase, and try to write it out to the l2arc.
3287 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3288 list_remove(&dev->l2ad_buflist, hdr);
3290 mutex_exit(&dev->l2ad_mtx);
3293 * Since we're using the pointer address as the tag when
3294 * incrementing and decrementing the l2ad_alloc refcount, we
3295 * must remove the old pointer (that we're about to destroy) and
3296 * add the new pointer to the refcount. Otherwise we'd remove
3297 * the wrong pointer address when calling arc_hdr_destroy() later.
3300 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3301 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3303 buf_discard_identity(hdr);
3304 kmem_cache_free(old, hdr);
3310 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3311 * The buf is returned thawed since we expect the consumer to modify it.
3314 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3316 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3317 ZIO_COMPRESS_OFF, type);
3318 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3320 arc_buf_t *buf = NULL;
3321 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3328 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3329 * for bufs containing metadata.
3332 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3333 enum zio_compress compression_type)
3335 ASSERT3U(lsize, >, 0);
3336 ASSERT3U(lsize, >=, psize);
3337 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3338 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3340 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3341 compression_type, ARC_BUFC_DATA);
3342 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3344 arc_buf_t *buf = NULL;
3345 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3347 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3349 if (!arc_buf_is_shared(buf)) {
3351 * To ensure that the hdr has the correct data in it if we call
3352 * arc_decompress() on this buf before it's been written to
3353 * disk, it's easiest if we just set up sharing between the
3356 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3357 arc_hdr_free_pabd(hdr);
3358 arc_share_buf(hdr, buf);
3365 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3367 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3368 l2arc_dev_t *dev = l2hdr->b_dev;
3369 uint64_t psize = arc_hdr_size(hdr);
3371 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3372 ASSERT(HDR_HAS_L2HDR(hdr));
3374 list_remove(&dev->l2ad_buflist, hdr);
3376 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3377 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3379 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3381 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3382 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3386 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3388 if (HDR_HAS_L1HDR(hdr)) {
3389 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3390 hdr->b_l1hdr.b_bufcnt > 0);
3391 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3392 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3394 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3395 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3397 if (!HDR_EMPTY(hdr))
3398 buf_discard_identity(hdr);
3400 if (HDR_HAS_L2HDR(hdr)) {
3401 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3402 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3405 mutex_enter(&dev->l2ad_mtx);
3408 * Even though we checked this conditional above, we
3409 * need to check this again now that we have the
3410 * l2ad_mtx. This is because we could be racing with
3411 * another thread calling l2arc_evict() which might have
3412 * destroyed this header's L2 portion as we were waiting
3413 * to acquire the l2ad_mtx. If that happens, we don't
3414 * want to re-destroy the header's L2 portion.
3416 if (HDR_HAS_L2HDR(hdr)) {
3418 arc_hdr_l2hdr_destroy(hdr);
3422 mutex_exit(&dev->l2ad_mtx);
3425 if (HDR_HAS_L1HDR(hdr)) {
3426 arc_cksum_free(hdr);
3428 while (hdr->b_l1hdr.b_buf != NULL)
3429 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3432 if (hdr->b_l1hdr.b_thawed != NULL) {
3433 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3434 hdr->b_l1hdr.b_thawed = NULL;
3438 if (hdr->b_l1hdr.b_pabd != NULL) {
3439 arc_hdr_free_pabd(hdr);
3443 ASSERT3P(hdr->b_hash_next, ==, NULL);
3444 if (HDR_HAS_L1HDR(hdr)) {
3445 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3446 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3447 kmem_cache_free(hdr_full_cache, hdr);
3449 kmem_cache_free(hdr_l2only_cache, hdr);
3454 arc_buf_destroy(arc_buf_t *buf, void* tag)
3456 arc_buf_hdr_t *hdr = buf->b_hdr;
3457 kmutex_t *hash_lock = HDR_LOCK(hdr);
3459 if (hdr->b_l1hdr.b_state == arc_anon) {
3460 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3461 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3462 VERIFY0(remove_reference(hdr, NULL, tag));
3463 arc_hdr_destroy(hdr);
3467 mutex_enter(hash_lock);
3468 ASSERT3P(hdr, ==, buf->b_hdr);
3469 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3470 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3471 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3472 ASSERT3P(buf->b_data, !=, NULL);
3474 (void) remove_reference(hdr, hash_lock, tag);
3475 arc_buf_destroy_impl(buf);
3476 mutex_exit(hash_lock);
3480 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3481 * state of the header is dependent on its state prior to entering this
3482 * function. The following transitions are possible:
3484 * - arc_mru -> arc_mru_ghost
3485 * - arc_mfu -> arc_mfu_ghost
3486 * - arc_mru_ghost -> arc_l2c_only
3487 * - arc_mru_ghost -> deleted
3488 * - arc_mfu_ghost -> arc_l2c_only
3489 * - arc_mfu_ghost -> deleted
3492 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3494 arc_state_t *evicted_state, *state;
3495 int64_t bytes_evicted = 0;
3497 ASSERT(MUTEX_HELD(hash_lock));
3498 ASSERT(HDR_HAS_L1HDR(hdr));
3500 state = hdr->b_l1hdr.b_state;
3501 if (GHOST_STATE(state)) {
3502 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3503 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3506 * l2arc_write_buffers() relies on a header's L1 portion
3507 * (i.e. its b_pabd field) during it's write phase.
3508 * Thus, we cannot push a header onto the arc_l2c_only
3509 * state (removing it's L1 piece) until the header is
3510 * done being written to the l2arc.
3512 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3513 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3514 return (bytes_evicted);
3517 ARCSTAT_BUMP(arcstat_deleted);
3518 bytes_evicted += HDR_GET_LSIZE(hdr);
3520 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3522 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3523 if (HDR_HAS_L2HDR(hdr)) {
3525 * This buffer is cached on the 2nd Level ARC;
3526 * don't destroy the header.
3528 arc_change_state(arc_l2c_only, hdr, hash_lock);
3530 * dropping from L1+L2 cached to L2-only,
3531 * realloc to remove the L1 header.
3533 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3536 arc_change_state(arc_anon, hdr, hash_lock);
3537 arc_hdr_destroy(hdr);
3539 return (bytes_evicted);
3542 ASSERT(state == arc_mru || state == arc_mfu);
3543 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3545 /* prefetch buffers have a minimum lifespan */
3546 if (HDR_IO_IN_PROGRESS(hdr) ||
3547 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3548 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3549 arc_min_prefetch_lifespan)) {
3550 ARCSTAT_BUMP(arcstat_evict_skip);
3551 return (bytes_evicted);
3554 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3555 while (hdr->b_l1hdr.b_buf) {
3556 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3557 if (!mutex_tryenter(&buf->b_evict_lock)) {
3558 ARCSTAT_BUMP(arcstat_mutex_miss);
3561 if (buf->b_data != NULL)
3562 bytes_evicted += HDR_GET_LSIZE(hdr);
3563 mutex_exit(&buf->b_evict_lock);
3564 arc_buf_destroy_impl(buf);
3567 if (HDR_HAS_L2HDR(hdr)) {
3568 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3570 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3571 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3572 HDR_GET_LSIZE(hdr));
3574 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3575 HDR_GET_LSIZE(hdr));
3579 if (hdr->b_l1hdr.b_bufcnt == 0) {
3580 arc_cksum_free(hdr);
3582 bytes_evicted += arc_hdr_size(hdr);
3585 * If this hdr is being evicted and has a compressed
3586 * buffer then we discard it here before we change states.
3587 * This ensures that the accounting is updated correctly
3588 * in arc_free_data_impl().
3590 arc_hdr_free_pabd(hdr);
3592 arc_change_state(evicted_state, hdr, hash_lock);
3593 ASSERT(HDR_IN_HASH_TABLE(hdr));
3594 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3595 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3598 return (bytes_evicted);
3602 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3603 uint64_t spa, int64_t bytes)
3605 multilist_sublist_t *mls;
3606 uint64_t bytes_evicted = 0;
3608 kmutex_t *hash_lock;
3609 int evict_count = 0;
3611 ASSERT3P(marker, !=, NULL);
3612 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3614 mls = multilist_sublist_lock(ml, idx);
3616 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3617 hdr = multilist_sublist_prev(mls, marker)) {
3618 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3619 (evict_count >= zfs_arc_evict_batch_limit))
3623 * To keep our iteration location, move the marker
3624 * forward. Since we're not holding hdr's hash lock, we
3625 * must be very careful and not remove 'hdr' from the
3626 * sublist. Otherwise, other consumers might mistake the
3627 * 'hdr' as not being on a sublist when they call the
3628 * multilist_link_active() function (they all rely on
3629 * the hash lock protecting concurrent insertions and
3630 * removals). multilist_sublist_move_forward() was
3631 * specifically implemented to ensure this is the case
3632 * (only 'marker' will be removed and re-inserted).
3634 multilist_sublist_move_forward(mls, marker);
3637 * The only case where the b_spa field should ever be
3638 * zero, is the marker headers inserted by
3639 * arc_evict_state(). It's possible for multiple threads
3640 * to be calling arc_evict_state() concurrently (e.g.
3641 * dsl_pool_close() and zio_inject_fault()), so we must
3642 * skip any markers we see from these other threads.
3644 if (hdr->b_spa == 0)
3647 /* we're only interested in evicting buffers of a certain spa */
3648 if (spa != 0 && hdr->b_spa != spa) {
3649 ARCSTAT_BUMP(arcstat_evict_skip);
3653 hash_lock = HDR_LOCK(hdr);
3656 * We aren't calling this function from any code path
3657 * that would already be holding a hash lock, so we're
3658 * asserting on this assumption to be defensive in case
3659 * this ever changes. Without this check, it would be
3660 * possible to incorrectly increment arcstat_mutex_miss
3661 * below (e.g. if the code changed such that we called
3662 * this function with a hash lock held).
3664 ASSERT(!MUTEX_HELD(hash_lock));
3666 if (mutex_tryenter(hash_lock)) {
3667 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3668 mutex_exit(hash_lock);
3670 bytes_evicted += evicted;
3673 * If evicted is zero, arc_evict_hdr() must have
3674 * decided to skip this header, don't increment
3675 * evict_count in this case.
3681 * If arc_size isn't overflowing, signal any
3682 * threads that might happen to be waiting.
3684 * For each header evicted, we wake up a single
3685 * thread. If we used cv_broadcast, we could
3686 * wake up "too many" threads causing arc_size
3687 * to significantly overflow arc_c; since
3688 * arc_get_data_impl() doesn't check for overflow
3689 * when it's woken up (it doesn't because it's
3690 * possible for the ARC to be overflowing while
3691 * full of un-evictable buffers, and the
3692 * function should proceed in this case).
3694 * If threads are left sleeping, due to not
3695 * using cv_broadcast, they will be woken up
3696 * just before arc_reclaim_thread() sleeps.
3698 mutex_enter(&arc_reclaim_lock);
3699 if (!arc_is_overflowing())
3700 cv_signal(&arc_reclaim_waiters_cv);
3701 mutex_exit(&arc_reclaim_lock);
3703 ARCSTAT_BUMP(arcstat_mutex_miss);
3707 multilist_sublist_unlock(mls);
3709 return (bytes_evicted);
3713 * Evict buffers from the given arc state, until we've removed the
3714 * specified number of bytes. Move the removed buffers to the
3715 * appropriate evict state.
3717 * This function makes a "best effort". It skips over any buffers
3718 * it can't get a hash_lock on, and so, may not catch all candidates.
3719 * It may also return without evicting as much space as requested.
3721 * If bytes is specified using the special value ARC_EVICT_ALL, this
3722 * will evict all available (i.e. unlocked and evictable) buffers from
3723 * the given arc state; which is used by arc_flush().
3726 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3727 arc_buf_contents_t type)
3729 uint64_t total_evicted = 0;
3730 multilist_t *ml = state->arcs_list[type];
3732 arc_buf_hdr_t **markers;
3734 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3736 num_sublists = multilist_get_num_sublists(ml);
3739 * If we've tried to evict from each sublist, made some
3740 * progress, but still have not hit the target number of bytes
3741 * to evict, we want to keep trying. The markers allow us to
3742 * pick up where we left off for each individual sublist, rather
3743 * than starting from the tail each time.
3745 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3746 for (int i = 0; i < num_sublists; i++) {
3747 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3750 * A b_spa of 0 is used to indicate that this header is
3751 * a marker. This fact is used in arc_adjust_type() and
3752 * arc_evict_state_impl().
3754 markers[i]->b_spa = 0;
3756 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3757 multilist_sublist_insert_tail(mls, markers[i]);
3758 multilist_sublist_unlock(mls);
3762 * While we haven't hit our target number of bytes to evict, or
3763 * we're evicting all available buffers.
3765 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3767 * Start eviction using a randomly selected sublist,
3768 * this is to try and evenly balance eviction across all
3769 * sublists. Always starting at the same sublist
3770 * (e.g. index 0) would cause evictions to favor certain
3771 * sublists over others.
3773 int sublist_idx = multilist_get_random_index(ml);
3774 uint64_t scan_evicted = 0;
3776 for (int i = 0; i < num_sublists; i++) {
3777 uint64_t bytes_remaining;
3778 uint64_t bytes_evicted;
3780 if (bytes == ARC_EVICT_ALL)
3781 bytes_remaining = ARC_EVICT_ALL;
3782 else if (total_evicted < bytes)
3783 bytes_remaining = bytes - total_evicted;
3787 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3788 markers[sublist_idx], spa, bytes_remaining);
3790 scan_evicted += bytes_evicted;
3791 total_evicted += bytes_evicted;
3793 /* we've reached the end, wrap to the beginning */
3794 if (++sublist_idx >= num_sublists)
3799 * If we didn't evict anything during this scan, we have
3800 * no reason to believe we'll evict more during another
3801 * scan, so break the loop.
3803 if (scan_evicted == 0) {
3804 /* This isn't possible, let's make that obvious */
3805 ASSERT3S(bytes, !=, 0);
3808 * When bytes is ARC_EVICT_ALL, the only way to
3809 * break the loop is when scan_evicted is zero.
3810 * In that case, we actually have evicted enough,
3811 * so we don't want to increment the kstat.
3813 if (bytes != ARC_EVICT_ALL) {
3814 ASSERT3S(total_evicted, <, bytes);
3815 ARCSTAT_BUMP(arcstat_evict_not_enough);
3822 for (int i = 0; i < num_sublists; i++) {
3823 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3824 multilist_sublist_remove(mls, markers[i]);
3825 multilist_sublist_unlock(mls);
3827 kmem_cache_free(hdr_full_cache, markers[i]);
3829 kmem_free(markers, sizeof (*markers) * num_sublists);
3831 return (total_evicted);
3835 * Flush all "evictable" data of the given type from the arc state
3836 * specified. This will not evict any "active" buffers (i.e. referenced).
3838 * When 'retry' is set to B_FALSE, the function will make a single pass
3839 * over the state and evict any buffers that it can. Since it doesn't
3840 * continually retry the eviction, it might end up leaving some buffers
3841 * in the ARC due to lock misses.
3843 * When 'retry' is set to B_TRUE, the function will continually retry the
3844 * eviction until *all* evictable buffers have been removed from the
3845 * state. As a result, if concurrent insertions into the state are
3846 * allowed (e.g. if the ARC isn't shutting down), this function might
3847 * wind up in an infinite loop, continually trying to evict buffers.
3850 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3853 uint64_t evicted = 0;
3855 while (refcount_count(&state->arcs_esize[type]) != 0) {
3856 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3866 * Evict the specified number of bytes from the state specified,
3867 * restricting eviction to the spa and type given. This function
3868 * prevents us from trying to evict more from a state's list than
3869 * is "evictable", and to skip evicting altogether when passed a
3870 * negative value for "bytes". In contrast, arc_evict_state() will
3871 * evict everything it can, when passed a negative value for "bytes".
3874 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3875 arc_buf_contents_t type)
3879 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3880 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3881 return (arc_evict_state(state, spa, delta, type));
3888 * Evict metadata buffers from the cache, such that arc_meta_used is
3889 * capped by the arc_meta_limit tunable.
3892 arc_adjust_meta(void)
3894 uint64_t total_evicted = 0;
3898 * If we're over the meta limit, we want to evict enough
3899 * metadata to get back under the meta limit. We don't want to
3900 * evict so much that we drop the MRU below arc_p, though. If
3901 * we're over the meta limit more than we're over arc_p, we
3902 * evict some from the MRU here, and some from the MFU below.
3904 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3905 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3906 refcount_count(&arc_mru->arcs_size) - arc_p));
3908 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3911 * Similar to the above, we want to evict enough bytes to get us
3912 * below the meta limit, but not so much as to drop us below the
3913 * space allotted to the MFU (which is defined as arc_c - arc_p).
3915 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3916 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
3918 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3920 return (total_evicted);
3924 * Return the type of the oldest buffer in the given arc state
3926 * This function will select a random sublist of type ARC_BUFC_DATA and
3927 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3928 * is compared, and the type which contains the "older" buffer will be
3931 static arc_buf_contents_t
3932 arc_adjust_type(arc_state_t *state)
3934 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
3935 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
3936 int data_idx = multilist_get_random_index(data_ml);
3937 int meta_idx = multilist_get_random_index(meta_ml);
3938 multilist_sublist_t *data_mls;
3939 multilist_sublist_t *meta_mls;
3940 arc_buf_contents_t type;
3941 arc_buf_hdr_t *data_hdr;
3942 arc_buf_hdr_t *meta_hdr;
3945 * We keep the sublist lock until we're finished, to prevent
3946 * the headers from being destroyed via arc_evict_state().
3948 data_mls = multilist_sublist_lock(data_ml, data_idx);
3949 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3952 * These two loops are to ensure we skip any markers that
3953 * might be at the tail of the lists due to arc_evict_state().
3956 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3957 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3958 if (data_hdr->b_spa != 0)
3962 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3963 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3964 if (meta_hdr->b_spa != 0)
3968 if (data_hdr == NULL && meta_hdr == NULL) {
3969 type = ARC_BUFC_DATA;
3970 } else if (data_hdr == NULL) {
3971 ASSERT3P(meta_hdr, !=, NULL);
3972 type = ARC_BUFC_METADATA;
3973 } else if (meta_hdr == NULL) {
3974 ASSERT3P(data_hdr, !=, NULL);
3975 type = ARC_BUFC_DATA;
3977 ASSERT3P(data_hdr, !=, NULL);
3978 ASSERT3P(meta_hdr, !=, NULL);
3980 /* The headers can't be on the sublist without an L1 header */
3981 ASSERT(HDR_HAS_L1HDR(data_hdr));
3982 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3984 if (data_hdr->b_l1hdr.b_arc_access <
3985 meta_hdr->b_l1hdr.b_arc_access) {
3986 type = ARC_BUFC_DATA;
3988 type = ARC_BUFC_METADATA;
3992 multilist_sublist_unlock(meta_mls);
3993 multilist_sublist_unlock(data_mls);
3999 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4004 uint64_t total_evicted = 0;
4009 * If we're over arc_meta_limit, we want to correct that before
4010 * potentially evicting data buffers below.
4012 total_evicted += arc_adjust_meta();
4017 * If we're over the target cache size, we want to evict enough
4018 * from the list to get back to our target size. We don't want
4019 * to evict too much from the MRU, such that it drops below
4020 * arc_p. So, if we're over our target cache size more than
4021 * the MRU is over arc_p, we'll evict enough to get back to
4022 * arc_p here, and then evict more from the MFU below.
4024 target = MIN((int64_t)(arc_size - arc_c),
4025 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4026 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
4029 * If we're below arc_meta_min, always prefer to evict data.
4030 * Otherwise, try to satisfy the requested number of bytes to
4031 * evict from the type which contains older buffers; in an
4032 * effort to keep newer buffers in the cache regardless of their
4033 * type. If we cannot satisfy the number of bytes from this
4034 * type, spill over into the next type.
4036 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4037 arc_meta_used > arc_meta_min) {
4038 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4039 total_evicted += bytes;
4042 * If we couldn't evict our target number of bytes from
4043 * metadata, we try to get the rest from data.
4048 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4050 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4051 total_evicted += bytes;
4054 * If we couldn't evict our target number of bytes from
4055 * data, we try to get the rest from metadata.
4060 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4066 * Now that we've tried to evict enough from the MRU to get its
4067 * size back to arc_p, if we're still above the target cache
4068 * size, we evict the rest from the MFU.
4070 target = arc_size - arc_c;
4072 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4073 arc_meta_used > arc_meta_min) {
4074 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4075 total_evicted += bytes;
4078 * If we couldn't evict our target number of bytes from
4079 * metadata, we try to get the rest from data.
4084 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4086 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4087 total_evicted += bytes;
4090 * If we couldn't evict our target number of bytes from
4091 * data, we try to get the rest from data.
4096 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4100 * Adjust ghost lists
4102 * In addition to the above, the ARC also defines target values
4103 * for the ghost lists. The sum of the mru list and mru ghost
4104 * list should never exceed the target size of the cache, and
4105 * the sum of the mru list, mfu list, mru ghost list, and mfu
4106 * ghost list should never exceed twice the target size of the
4107 * cache. The following logic enforces these limits on the ghost
4108 * caches, and evicts from them as needed.
4110 target = refcount_count(&arc_mru->arcs_size) +
4111 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4113 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4114 total_evicted += bytes;
4119 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4122 * We assume the sum of the mru list and mfu list is less than
4123 * or equal to arc_c (we enforced this above), which means we
4124 * can use the simpler of the two equations below:
4126 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4127 * mru ghost + mfu ghost <= arc_c
4129 target = refcount_count(&arc_mru_ghost->arcs_size) +
4130 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4132 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4133 total_evicted += bytes;
4138 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4140 return (total_evicted);
4144 arc_flush(spa_t *spa, boolean_t retry)
4149 * If retry is B_TRUE, a spa must not be specified since we have
4150 * no good way to determine if all of a spa's buffers have been
4151 * evicted from an arc state.
4153 ASSERT(!retry || spa == 0);
4156 guid = spa_load_guid(spa);
4158 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4159 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4161 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4162 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4164 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4165 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4167 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4168 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4172 arc_shrink(int64_t to_free)
4174 if (arc_c > arc_c_min) {
4175 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
4176 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
4177 if (arc_c > arc_c_min + to_free)
4178 atomic_add_64(&arc_c, -to_free);
4182 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4183 if (arc_c > arc_size)
4184 arc_c = MAX(arc_size, arc_c_min);
4186 arc_p = (arc_c >> 1);
4188 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
4191 ASSERT(arc_c >= arc_c_min);
4192 ASSERT((int64_t)arc_p >= 0);
4195 if (arc_size > arc_c) {
4196 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size,
4198 (void) arc_adjust();
4202 typedef enum free_memory_reason_t {
4207 FMR_PAGES_PP_MAXIMUM,
4211 } free_memory_reason_t;
4213 int64_t last_free_memory;
4214 free_memory_reason_t last_free_reason;
4217 * Additional reserve of pages for pp_reserve.
4219 int64_t arc_pages_pp_reserve = 64;
4222 * Additional reserve of pages for swapfs.
4224 int64_t arc_swapfs_reserve = 64;
4227 * Return the amount of memory that can be consumed before reclaim will be
4228 * needed. Positive if there is sufficient free memory, negative indicates
4229 * the amount of memory that needs to be freed up.
4232 arc_available_memory(void)
4234 int64_t lowest = INT64_MAX;
4236 free_memory_reason_t r = FMR_UNKNOWN;
4240 * Cooperate with pagedaemon when it's time for it to scan
4241 * and reclaim some pages.
4243 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
4251 * check that we're out of range of the pageout scanner. It starts to
4252 * schedule paging if freemem is less than lotsfree and needfree.
4253 * lotsfree is the high-water mark for pageout, and needfree is the
4254 * number of needed free pages. We add extra pages here to make sure
4255 * the scanner doesn't start up while we're freeing memory.
4257 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4264 * check to make sure that swapfs has enough space so that anon
4265 * reservations can still succeed. anon_resvmem() checks that the
4266 * availrmem is greater than swapfs_minfree, and the number of reserved
4267 * swap pages. We also add a bit of extra here just to prevent
4268 * circumstances from getting really dire.
4270 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4271 desfree - arc_swapfs_reserve);
4274 r = FMR_SWAPFS_MINFREE;
4279 * Check that we have enough availrmem that memory locking (e.g., via
4280 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4281 * stores the number of pages that cannot be locked; when availrmem
4282 * drops below pages_pp_maximum, page locking mechanisms such as
4283 * page_pp_lock() will fail.)
4285 n = PAGESIZE * (availrmem - pages_pp_maximum -
4286 arc_pages_pp_reserve);
4289 r = FMR_PAGES_PP_MAXIMUM;
4292 #endif /* illumos */
4293 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4295 * If we're on an i386 platform, it's possible that we'll exhaust the
4296 * kernel heap space before we ever run out of available physical
4297 * memory. Most checks of the size of the heap_area compare against
4298 * tune.t_minarmem, which is the minimum available real memory that we
4299 * can have in the system. However, this is generally fixed at 25 pages
4300 * which is so low that it's useless. In this comparison, we seek to
4301 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4302 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4305 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
4306 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
4311 #define zio_arena NULL
4313 #define zio_arena heap_arena
4317 * If zio data pages are being allocated out of a separate heap segment,
4318 * then enforce that the size of available vmem for this arena remains
4319 * above about 1/16th free.
4321 * Note: The 1/16th arena free requirement was put in place
4322 * to aggressively evict memory from the arc in order to avoid
4323 * memory fragmentation issues.
4325 if (zio_arena != NULL) {
4326 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4327 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
4335 * Above limits know nothing about real level of KVA fragmentation.
4336 * Start aggressive reclamation if too little sequential KVA left.
4339 n = (vmem_size(heap_arena, VMEM_MAXFREE) < SPA_MAXBLOCKSIZE) ?
4340 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) :
4349 /* Every 100 calls, free a small amount */
4350 if (spa_get_random(100) == 0)
4352 #endif /* _KERNEL */
4354 last_free_memory = lowest;
4355 last_free_reason = r;
4356 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4362 * Determine if the system is under memory pressure and is asking
4363 * to reclaim memory. A return value of B_TRUE indicates that the system
4364 * is under memory pressure and that the arc should adjust accordingly.
4367 arc_reclaim_needed(void)
4369 return (arc_available_memory() < 0);
4372 extern kmem_cache_t *zio_buf_cache[];
4373 extern kmem_cache_t *zio_data_buf_cache[];
4374 extern kmem_cache_t *range_seg_cache;
4375 extern kmem_cache_t *abd_chunk_cache;
4377 static __noinline void
4378 arc_kmem_reap_now(void)
4381 kmem_cache_t *prev_cache = NULL;
4382 kmem_cache_t *prev_data_cache = NULL;
4384 DTRACE_PROBE(arc__kmem_reap_start);
4386 if (arc_meta_used >= arc_meta_limit) {
4388 * We are exceeding our meta-data cache limit.
4389 * Purge some DNLC entries to release holds on meta-data.
4391 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4395 * Reclaim unused memory from all kmem caches.
4401 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4402 if (zio_buf_cache[i] != prev_cache) {
4403 prev_cache = zio_buf_cache[i];
4404 kmem_cache_reap_now(zio_buf_cache[i]);
4406 if (zio_data_buf_cache[i] != prev_data_cache) {
4407 prev_data_cache = zio_data_buf_cache[i];
4408 kmem_cache_reap_now(zio_data_buf_cache[i]);
4411 kmem_cache_reap_now(abd_chunk_cache);
4412 kmem_cache_reap_now(buf_cache);
4413 kmem_cache_reap_now(hdr_full_cache);
4414 kmem_cache_reap_now(hdr_l2only_cache);
4415 kmem_cache_reap_now(range_seg_cache);
4418 if (zio_arena != NULL) {
4420 * Ask the vmem arena to reclaim unused memory from its
4423 vmem_qcache_reap(zio_arena);
4426 DTRACE_PROBE(arc__kmem_reap_end);
4430 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4431 * enough data and signal them to proceed. When this happens, the threads in
4432 * arc_get_data_impl() are sleeping while holding the hash lock for their
4433 * particular arc header. Thus, we must be careful to never sleep on a
4434 * hash lock in this thread. This is to prevent the following deadlock:
4436 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4437 * waiting for the reclaim thread to signal it.
4439 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4440 * fails, and goes to sleep forever.
4442 * This possible deadlock is avoided by always acquiring a hash lock
4443 * using mutex_tryenter() from arc_reclaim_thread().
4446 arc_reclaim_thread(void *dummy __unused)
4448 hrtime_t growtime = 0;
4451 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4453 mutex_enter(&arc_reclaim_lock);
4454 while (!arc_reclaim_thread_exit) {
4455 uint64_t evicted = 0;
4458 * This is necessary in order for the mdb ::arc dcmd to
4459 * show up to date information. Since the ::arc command
4460 * does not call the kstat's update function, without
4461 * this call, the command may show stale stats for the
4462 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4463 * with this change, the data might be up to 1 second
4464 * out of date; but that should suffice. The arc_state_t
4465 * structures can be queried directly if more accurate
4466 * information is needed.
4468 if (arc_ksp != NULL)
4469 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4471 mutex_exit(&arc_reclaim_lock);
4474 * We call arc_adjust() before (possibly) calling
4475 * arc_kmem_reap_now(), so that we can wake up
4476 * arc_get_data_impl() sooner.
4478 evicted = arc_adjust();
4480 int64_t free_memory = arc_available_memory();
4481 if (free_memory < 0) {
4483 arc_no_grow = B_TRUE;
4487 * Wait at least zfs_grow_retry (default 60) seconds
4488 * before considering growing.
4490 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4492 arc_kmem_reap_now();
4495 * If we are still low on memory, shrink the ARC
4496 * so that we have arc_shrink_min free space.
4498 free_memory = arc_available_memory();
4501 (arc_c >> arc_shrink_shift) - free_memory;
4503 arc_shrink(to_free);
4505 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4506 arc_no_grow = B_TRUE;
4507 } else if (gethrtime() >= growtime) {
4508 arc_no_grow = B_FALSE;
4511 mutex_enter(&arc_reclaim_lock);
4514 * If evicted is zero, we couldn't evict anything via
4515 * arc_adjust(). This could be due to hash lock
4516 * collisions, but more likely due to the majority of
4517 * arc buffers being unevictable. Therefore, even if
4518 * arc_size is above arc_c, another pass is unlikely to
4519 * be helpful and could potentially cause us to enter an
4522 if (arc_size <= arc_c || evicted == 0) {
4524 * We're either no longer overflowing, or we
4525 * can't evict anything more, so we should wake
4526 * up any threads before we go to sleep.
4528 cv_broadcast(&arc_reclaim_waiters_cv);
4531 * Block until signaled, or after one second (we
4532 * might need to perform arc_kmem_reap_now()
4533 * even if we aren't being signalled)
4535 CALLB_CPR_SAFE_BEGIN(&cpr);
4536 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4537 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4538 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4542 arc_reclaim_thread_exit = B_FALSE;
4543 cv_broadcast(&arc_reclaim_thread_cv);
4544 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4548 static u_int arc_dnlc_evicts_arg;
4549 extern struct vfsops zfs_vfsops;
4552 arc_dnlc_evicts_thread(void *dummy __unused)
4557 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4559 mutex_enter(&arc_dnlc_evicts_lock);
4560 while (!arc_dnlc_evicts_thread_exit) {
4561 CALLB_CPR_SAFE_BEGIN(&cpr);
4562 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4563 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4564 if (arc_dnlc_evicts_arg != 0) {
4565 percent = arc_dnlc_evicts_arg;
4566 mutex_exit(&arc_dnlc_evicts_lock);
4568 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4570 mutex_enter(&arc_dnlc_evicts_lock);
4572 * Clear our token only after vnlru_free()
4573 * pass is done, to avoid false queueing of
4576 arc_dnlc_evicts_arg = 0;
4579 arc_dnlc_evicts_thread_exit = FALSE;
4580 cv_broadcast(&arc_dnlc_evicts_cv);
4581 CALLB_CPR_EXIT(&cpr);
4586 dnlc_reduce_cache(void *arg)
4590 percent = (u_int)(uintptr_t)arg;
4591 mutex_enter(&arc_dnlc_evicts_lock);
4592 if (arc_dnlc_evicts_arg == 0) {
4593 arc_dnlc_evicts_arg = percent;
4594 cv_broadcast(&arc_dnlc_evicts_cv);
4596 mutex_exit(&arc_dnlc_evicts_lock);
4600 * Adapt arc info given the number of bytes we are trying to add and
4601 * the state that we are comming from. This function is only called
4602 * when we are adding new content to the cache.
4605 arc_adapt(int bytes, arc_state_t *state)
4608 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4609 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4610 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4612 if (state == arc_l2c_only)
4617 * Adapt the target size of the MRU list:
4618 * - if we just hit in the MRU ghost list, then increase
4619 * the target size of the MRU list.
4620 * - if we just hit in the MFU ghost list, then increase
4621 * the target size of the MFU list by decreasing the
4622 * target size of the MRU list.
4624 if (state == arc_mru_ghost) {
4625 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4626 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4628 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4629 } else if (state == arc_mfu_ghost) {
4632 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4633 mult = MIN(mult, 10);
4635 delta = MIN(bytes * mult, arc_p);
4636 arc_p = MAX(arc_p_min, arc_p - delta);
4638 ASSERT((int64_t)arc_p >= 0);
4640 if (arc_reclaim_needed()) {
4641 cv_signal(&arc_reclaim_thread_cv);
4648 if (arc_c >= arc_c_max)
4652 * If we're within (2 * maxblocksize) bytes of the target
4653 * cache size, increment the target cache size
4655 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
4656 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
4657 atomic_add_64(&arc_c, (int64_t)bytes);
4658 if (arc_c > arc_c_max)
4660 else if (state == arc_anon)
4661 atomic_add_64(&arc_p, (int64_t)bytes);
4665 ASSERT((int64_t)arc_p >= 0);
4669 * Check if arc_size has grown past our upper threshold, determined by
4670 * zfs_arc_overflow_shift.
4673 arc_is_overflowing(void)
4675 /* Always allow at least one block of overflow */
4676 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4677 arc_c >> zfs_arc_overflow_shift);
4679 return (arc_size >= arc_c + overflow);
4683 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4685 arc_buf_contents_t type = arc_buf_type(hdr);
4687 arc_get_data_impl(hdr, size, tag);
4688 if (type == ARC_BUFC_METADATA) {
4689 return (abd_alloc(size, B_TRUE));
4691 ASSERT(type == ARC_BUFC_DATA);
4692 return (abd_alloc(size, B_FALSE));
4697 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4699 arc_buf_contents_t type = arc_buf_type(hdr);
4701 arc_get_data_impl(hdr, size, tag);
4702 if (type == ARC_BUFC_METADATA) {
4703 return (zio_buf_alloc(size));
4705 ASSERT(type == ARC_BUFC_DATA);
4706 return (zio_data_buf_alloc(size));
4711 * Allocate a block and return it to the caller. If we are hitting the
4712 * hard limit for the cache size, we must sleep, waiting for the eviction
4713 * thread to catch up. If we're past the target size but below the hard
4714 * limit, we'll only signal the reclaim thread and continue on.
4717 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4719 arc_state_t *state = hdr->b_l1hdr.b_state;
4720 arc_buf_contents_t type = arc_buf_type(hdr);
4722 arc_adapt(size, state);
4725 * If arc_size is currently overflowing, and has grown past our
4726 * upper limit, we must be adding data faster than the evict
4727 * thread can evict. Thus, to ensure we don't compound the
4728 * problem by adding more data and forcing arc_size to grow even
4729 * further past it's target size, we halt and wait for the
4730 * eviction thread to catch up.
4732 * It's also possible that the reclaim thread is unable to evict
4733 * enough buffers to get arc_size below the overflow limit (e.g.
4734 * due to buffers being un-evictable, or hash lock collisions).
4735 * In this case, we want to proceed regardless if we're
4736 * overflowing; thus we don't use a while loop here.
4738 if (arc_is_overflowing()) {
4739 mutex_enter(&arc_reclaim_lock);
4742 * Now that we've acquired the lock, we may no longer be
4743 * over the overflow limit, lets check.
4745 * We're ignoring the case of spurious wake ups. If that
4746 * were to happen, it'd let this thread consume an ARC
4747 * buffer before it should have (i.e. before we're under
4748 * the overflow limit and were signalled by the reclaim
4749 * thread). As long as that is a rare occurrence, it
4750 * shouldn't cause any harm.
4752 if (arc_is_overflowing()) {
4753 cv_signal(&arc_reclaim_thread_cv);
4754 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4757 mutex_exit(&arc_reclaim_lock);
4760 VERIFY3U(hdr->b_type, ==, type);
4761 if (type == ARC_BUFC_METADATA) {
4762 arc_space_consume(size, ARC_SPACE_META);
4764 arc_space_consume(size, ARC_SPACE_DATA);
4768 * Update the state size. Note that ghost states have a
4769 * "ghost size" and so don't need to be updated.
4771 if (!GHOST_STATE(state)) {
4773 (void) refcount_add_many(&state->arcs_size, size, tag);
4776 * If this is reached via arc_read, the link is
4777 * protected by the hash lock. If reached via
4778 * arc_buf_alloc, the header should not be accessed by
4779 * any other thread. And, if reached via arc_read_done,
4780 * the hash lock will protect it if it's found in the
4781 * hash table; otherwise no other thread should be
4782 * trying to [add|remove]_reference it.
4784 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4785 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4786 (void) refcount_add_many(&state->arcs_esize[type],
4791 * If we are growing the cache, and we are adding anonymous
4792 * data, and we have outgrown arc_p, update arc_p
4794 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
4795 (refcount_count(&arc_anon->arcs_size) +
4796 refcount_count(&arc_mru->arcs_size) > arc_p))
4797 arc_p = MIN(arc_c, arc_p + size);
4799 ARCSTAT_BUMP(arcstat_allocated);
4803 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4805 arc_free_data_impl(hdr, size, tag);
4810 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4812 arc_buf_contents_t type = arc_buf_type(hdr);
4814 arc_free_data_impl(hdr, size, tag);
4815 if (type == ARC_BUFC_METADATA) {
4816 zio_buf_free(buf, size);
4818 ASSERT(type == ARC_BUFC_DATA);
4819 zio_data_buf_free(buf, size);
4824 * Free the arc data buffer.
4827 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4829 arc_state_t *state = hdr->b_l1hdr.b_state;
4830 arc_buf_contents_t type = arc_buf_type(hdr);
4832 /* protected by hash lock, if in the hash table */
4833 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4834 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4835 ASSERT(state != arc_anon && state != arc_l2c_only);
4837 (void) refcount_remove_many(&state->arcs_esize[type],
4840 (void) refcount_remove_many(&state->arcs_size, size, tag);
4842 VERIFY3U(hdr->b_type, ==, type);
4843 if (type == ARC_BUFC_METADATA) {
4844 arc_space_return(size, ARC_SPACE_META);
4846 ASSERT(type == ARC_BUFC_DATA);
4847 arc_space_return(size, ARC_SPACE_DATA);
4852 * This routine is called whenever a buffer is accessed.
4853 * NOTE: the hash lock is dropped in this function.
4856 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4860 ASSERT(MUTEX_HELD(hash_lock));
4861 ASSERT(HDR_HAS_L1HDR(hdr));
4863 if (hdr->b_l1hdr.b_state == arc_anon) {
4865 * This buffer is not in the cache, and does not
4866 * appear in our "ghost" list. Add the new buffer
4870 ASSERT0(hdr->b_l1hdr.b_arc_access);
4871 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4872 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4873 arc_change_state(arc_mru, hdr, hash_lock);
4875 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4876 now = ddi_get_lbolt();
4879 * If this buffer is here because of a prefetch, then either:
4880 * - clear the flag if this is a "referencing" read
4881 * (any subsequent access will bump this into the MFU state).
4883 * - move the buffer to the head of the list if this is
4884 * another prefetch (to make it less likely to be evicted).
4886 if (HDR_PREFETCH(hdr)) {
4887 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4888 /* link protected by hash lock */
4889 ASSERT(multilist_link_active(
4890 &hdr->b_l1hdr.b_arc_node));
4892 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4893 ARCSTAT_BUMP(arcstat_mru_hits);
4895 hdr->b_l1hdr.b_arc_access = now;
4900 * This buffer has been "accessed" only once so far,
4901 * but it is still in the cache. Move it to the MFU
4904 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4906 * More than 125ms have passed since we
4907 * instantiated this buffer. Move it to the
4908 * most frequently used state.
4910 hdr->b_l1hdr.b_arc_access = now;
4911 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4912 arc_change_state(arc_mfu, hdr, hash_lock);
4914 ARCSTAT_BUMP(arcstat_mru_hits);
4915 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
4916 arc_state_t *new_state;
4918 * This buffer has been "accessed" recently, but
4919 * was evicted from the cache. Move it to the
4923 if (HDR_PREFETCH(hdr)) {
4924 new_state = arc_mru;
4925 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
4926 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4927 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4929 new_state = arc_mfu;
4930 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4933 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4934 arc_change_state(new_state, hdr, hash_lock);
4936 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
4937 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
4939 * This buffer has been accessed more than once and is
4940 * still in the cache. Keep it in the MFU state.
4942 * NOTE: an add_reference() that occurred when we did
4943 * the arc_read() will have kicked this off the list.
4944 * If it was a prefetch, we will explicitly move it to
4945 * the head of the list now.
4947 if ((HDR_PREFETCH(hdr)) != 0) {
4948 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4949 /* link protected by hash_lock */
4950 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4952 ARCSTAT_BUMP(arcstat_mfu_hits);
4953 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4954 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4955 arc_state_t *new_state = arc_mfu;
4957 * This buffer has been accessed more than once but has
4958 * been evicted from the cache. Move it back to the
4962 if (HDR_PREFETCH(hdr)) {
4964 * This is a prefetch access...
4965 * move this block back to the MRU state.
4967 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4968 new_state = arc_mru;
4971 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4972 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4973 arc_change_state(new_state, hdr, hash_lock);
4975 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4976 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4978 * This buffer is on the 2nd Level ARC.
4981 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4982 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4983 arc_change_state(arc_mfu, hdr, hash_lock);
4985 ASSERT(!"invalid arc state");
4989 /* a generic arc_done_func_t which you can use */
4992 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4994 if (zio == NULL || zio->io_error == 0)
4995 bcopy(buf->b_data, arg, arc_buf_size(buf));
4996 arc_buf_destroy(buf, arg);
4999 /* a generic arc_done_func_t */
5001 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
5003 arc_buf_t **bufp = arg;
5004 if (zio && zio->io_error) {
5005 arc_buf_destroy(buf, arg);
5009 ASSERT(buf->b_data);
5014 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5016 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5017 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5018 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
5020 if (HDR_COMPRESSION_ENABLED(hdr)) {
5021 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
5022 BP_GET_COMPRESS(bp));
5024 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5025 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5030 arc_read_done(zio_t *zio)
5032 arc_buf_hdr_t *hdr = zio->io_private;
5033 kmutex_t *hash_lock = NULL;
5034 arc_callback_t *callback_list;
5035 arc_callback_t *acb;
5036 boolean_t freeable = B_FALSE;
5037 boolean_t no_zio_error = (zio->io_error == 0);
5040 * The hdr was inserted into hash-table and removed from lists
5041 * prior to starting I/O. We should find this header, since
5042 * it's in the hash table, and it should be legit since it's
5043 * not possible to evict it during the I/O. The only possible
5044 * reason for it not to be found is if we were freed during the
5047 if (HDR_IN_HASH_TABLE(hdr)) {
5048 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5049 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5050 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5051 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5052 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5054 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5057 ASSERT((found == hdr &&
5058 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5059 (found == hdr && HDR_L2_READING(hdr)));
5060 ASSERT3P(hash_lock, !=, NULL);
5064 /* byteswap if necessary */
5065 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5066 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5067 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5069 hdr->b_l1hdr.b_byteswap =
5070 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5073 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5077 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5078 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5079 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5081 callback_list = hdr->b_l1hdr.b_acb;
5082 ASSERT3P(callback_list, !=, NULL);
5084 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
5086 * Only call arc_access on anonymous buffers. This is because
5087 * if we've issued an I/O for an evicted buffer, we've already
5088 * called arc_access (to prevent any simultaneous readers from
5089 * getting confused).
5091 arc_access(hdr, hash_lock);
5095 * If a read request has a callback (i.e. acb_done is not NULL), then we
5096 * make a buf containing the data according to the parameters which were
5097 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5098 * aren't needlessly decompressing the data multiple times.
5100 int callback_cnt = 0;
5101 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5105 /* This is a demand read since prefetches don't use callbacks */
5108 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
5109 acb->acb_compressed, no_zio_error, &acb->acb_buf);
5111 zio->io_error = error;
5114 hdr->b_l1hdr.b_acb = NULL;
5115 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5116 if (callback_cnt == 0) {
5117 ASSERT(HDR_PREFETCH(hdr));
5118 ASSERT0(hdr->b_l1hdr.b_bufcnt);
5119 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5122 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5123 callback_list != NULL);
5126 arc_hdr_verify(hdr, zio->io_bp);
5128 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5129 if (hdr->b_l1hdr.b_state != arc_anon)
5130 arc_change_state(arc_anon, hdr, hash_lock);
5131 if (HDR_IN_HASH_TABLE(hdr))
5132 buf_hash_remove(hdr);
5133 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5137 * Broadcast before we drop the hash_lock to avoid the possibility
5138 * that the hdr (and hence the cv) might be freed before we get to
5139 * the cv_broadcast().
5141 cv_broadcast(&hdr->b_l1hdr.b_cv);
5143 if (hash_lock != NULL) {
5144 mutex_exit(hash_lock);
5147 * This block was freed while we waited for the read to
5148 * complete. It has been removed from the hash table and
5149 * moved to the anonymous state (so that it won't show up
5152 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5153 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5156 /* execute each callback and free its structure */
5157 while ((acb = callback_list) != NULL) {
5159 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
5161 if (acb->acb_zio_dummy != NULL) {
5162 acb->acb_zio_dummy->io_error = zio->io_error;
5163 zio_nowait(acb->acb_zio_dummy);
5166 callback_list = acb->acb_next;
5167 kmem_free(acb, sizeof (arc_callback_t));
5171 arc_hdr_destroy(hdr);
5175 * "Read" the block at the specified DVA (in bp) via the
5176 * cache. If the block is found in the cache, invoke the provided
5177 * callback immediately and return. Note that the `zio' parameter
5178 * in the callback will be NULL in this case, since no IO was
5179 * required. If the block is not in the cache pass the read request
5180 * on to the spa with a substitute callback function, so that the
5181 * requested block will be added to the cache.
5183 * If a read request arrives for a block that has a read in-progress,
5184 * either wait for the in-progress read to complete (and return the
5185 * results); or, if this is a read with a "done" func, add a record
5186 * to the read to invoke the "done" func when the read completes,
5187 * and return; or just return.
5189 * arc_read_done() will invoke all the requested "done" functions
5190 * for readers of this block.
5193 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
5194 void *private, zio_priority_t priority, int zio_flags,
5195 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5197 arc_buf_hdr_t *hdr = NULL;
5198 kmutex_t *hash_lock = NULL;
5200 uint64_t guid = spa_load_guid(spa);
5201 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5203 ASSERT(!BP_IS_EMBEDDED(bp) ||
5204 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5207 if (!BP_IS_EMBEDDED(bp)) {
5209 * Embedded BP's have no DVA and require no I/O to "read".
5210 * Create an anonymous arc buf to back it.
5212 hdr = buf_hash_find(guid, bp, &hash_lock);
5215 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5216 arc_buf_t *buf = NULL;
5217 *arc_flags |= ARC_FLAG_CACHED;
5219 if (HDR_IO_IN_PROGRESS(hdr)) {
5221 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5222 priority == ZIO_PRIORITY_SYNC_READ) {
5224 * This sync read must wait for an
5225 * in-progress async read (e.g. a predictive
5226 * prefetch). Async reads are queued
5227 * separately at the vdev_queue layer, so
5228 * this is a form of priority inversion.
5229 * Ideally, we would "inherit" the demand
5230 * i/o's priority by moving the i/o from
5231 * the async queue to the synchronous queue,
5232 * but there is currently no mechanism to do
5233 * so. Track this so that we can evaluate
5234 * the magnitude of this potential performance
5237 * Note that if the prefetch i/o is already
5238 * active (has been issued to the device),
5239 * the prefetch improved performance, because
5240 * we issued it sooner than we would have
5241 * without the prefetch.
5243 DTRACE_PROBE1(arc__sync__wait__for__async,
5244 arc_buf_hdr_t *, hdr);
5245 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
5247 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5248 arc_hdr_clear_flags(hdr,
5249 ARC_FLAG_PREDICTIVE_PREFETCH);
5252 if (*arc_flags & ARC_FLAG_WAIT) {
5253 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5254 mutex_exit(hash_lock);
5257 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5260 arc_callback_t *acb = NULL;
5262 acb = kmem_zalloc(sizeof (arc_callback_t),
5264 acb->acb_done = done;
5265 acb->acb_private = private;
5266 acb->acb_compressed = compressed_read;
5268 acb->acb_zio_dummy = zio_null(pio,
5269 spa, NULL, NULL, NULL, zio_flags);
5271 ASSERT3P(acb->acb_done, !=, NULL);
5272 acb->acb_next = hdr->b_l1hdr.b_acb;
5273 hdr->b_l1hdr.b_acb = acb;
5274 mutex_exit(hash_lock);
5277 mutex_exit(hash_lock);
5281 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5282 hdr->b_l1hdr.b_state == arc_mfu);
5285 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5287 * This is a demand read which does not have to
5288 * wait for i/o because we did a predictive
5289 * prefetch i/o for it, which has completed.
5292 arc__demand__hit__predictive__prefetch,
5293 arc_buf_hdr_t *, hdr);
5295 arcstat_demand_hit_predictive_prefetch);
5296 arc_hdr_clear_flags(hdr,
5297 ARC_FLAG_PREDICTIVE_PREFETCH);
5299 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5301 /* Get a buf with the desired data in it. */
5302 VERIFY0(arc_buf_alloc_impl(hdr, private,
5303 compressed_read, B_TRUE, &buf));
5304 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5305 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5306 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5308 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5309 arc_access(hdr, hash_lock);
5310 if (*arc_flags & ARC_FLAG_L2CACHE)
5311 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5312 mutex_exit(hash_lock);
5313 ARCSTAT_BUMP(arcstat_hits);
5314 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5315 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5316 data, metadata, hits);
5319 done(NULL, buf, private);
5321 uint64_t lsize = BP_GET_LSIZE(bp);
5322 uint64_t psize = BP_GET_PSIZE(bp);
5323 arc_callback_t *acb;
5326 boolean_t devw = B_FALSE;
5330 /* this block is not in the cache */
5331 arc_buf_hdr_t *exists = NULL;
5332 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5333 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5334 BP_GET_COMPRESS(bp), type);
5336 if (!BP_IS_EMBEDDED(bp)) {
5337 hdr->b_dva = *BP_IDENTITY(bp);
5338 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5339 exists = buf_hash_insert(hdr, &hash_lock);
5341 if (exists != NULL) {
5342 /* somebody beat us to the hash insert */
5343 mutex_exit(hash_lock);
5344 buf_discard_identity(hdr);
5345 arc_hdr_destroy(hdr);
5346 goto top; /* restart the IO request */
5350 * This block is in the ghost cache. If it was L2-only
5351 * (and thus didn't have an L1 hdr), we realloc the
5352 * header to add an L1 hdr.
5354 if (!HDR_HAS_L1HDR(hdr)) {
5355 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5358 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5359 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5360 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5361 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5362 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5363 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5366 * This is a delicate dance that we play here.
5367 * This hdr is in the ghost list so we access it
5368 * to move it out of the ghost list before we
5369 * initiate the read. If it's a prefetch then
5370 * it won't have a callback so we'll remove the
5371 * reference that arc_buf_alloc_impl() created. We
5372 * do this after we've called arc_access() to
5373 * avoid hitting an assert in remove_reference().
5375 arc_access(hdr, hash_lock);
5376 arc_hdr_alloc_pabd(hdr);
5378 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5379 size = arc_hdr_size(hdr);
5382 * If compression is enabled on the hdr, then will do
5383 * RAW I/O and will store the compressed data in the hdr's
5384 * data block. Otherwise, the hdr's data block will contain
5385 * the uncompressed data.
5387 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5388 zio_flags |= ZIO_FLAG_RAW;
5391 if (*arc_flags & ARC_FLAG_PREFETCH)
5392 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5393 if (*arc_flags & ARC_FLAG_L2CACHE)
5394 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5395 if (BP_GET_LEVEL(bp) > 0)
5396 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5397 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5398 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5399 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5401 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5402 acb->acb_done = done;
5403 acb->acb_private = private;
5404 acb->acb_compressed = compressed_read;
5406 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5407 hdr->b_l1hdr.b_acb = acb;
5408 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5410 if (HDR_HAS_L2HDR(hdr) &&
5411 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5412 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5413 addr = hdr->b_l2hdr.b_daddr;
5415 * Lock out device removal.
5417 if (vdev_is_dead(vd) ||
5418 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5422 if (priority == ZIO_PRIORITY_ASYNC_READ)
5423 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5425 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5427 if (hash_lock != NULL)
5428 mutex_exit(hash_lock);
5431 * At this point, we have a level 1 cache miss. Try again in
5432 * L2ARC if possible.
5434 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5436 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5437 uint64_t, lsize, zbookmark_phys_t *, zb);
5438 ARCSTAT_BUMP(arcstat_misses);
5439 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5440 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5441 data, metadata, misses);
5446 racct_add_force(curproc, RACCT_READBPS, size);
5447 racct_add_force(curproc, RACCT_READIOPS, 1);
5448 PROC_UNLOCK(curproc);
5451 curthread->td_ru.ru_inblock++;
5454 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5456 * Read from the L2ARC if the following are true:
5457 * 1. The L2ARC vdev was previously cached.
5458 * 2. This buffer still has L2ARC metadata.
5459 * 3. This buffer isn't currently writing to the L2ARC.
5460 * 4. The L2ARC entry wasn't evicted, which may
5461 * also have invalidated the vdev.
5462 * 5. This isn't prefetch and l2arc_noprefetch is set.
5464 if (HDR_HAS_L2HDR(hdr) &&
5465 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5466 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5467 l2arc_read_callback_t *cb;
5471 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5472 ARCSTAT_BUMP(arcstat_l2_hits);
5474 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5476 cb->l2rcb_hdr = hdr;
5479 cb->l2rcb_flags = zio_flags;
5481 asize = vdev_psize_to_asize(vd, size);
5482 if (asize != size) {
5483 abd = abd_alloc_for_io(asize,
5484 HDR_ISTYPE_METADATA(hdr));
5485 cb->l2rcb_abd = abd;
5487 abd = hdr->b_l1hdr.b_pabd;
5490 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5491 addr + asize <= vd->vdev_psize -
5492 VDEV_LABEL_END_SIZE);
5495 * l2arc read. The SCL_L2ARC lock will be
5496 * released by l2arc_read_done().
5497 * Issue a null zio if the underlying buffer
5498 * was squashed to zero size by compression.
5500 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5501 ZIO_COMPRESS_EMPTY);
5502 rzio = zio_read_phys(pio, vd, addr,
5505 l2arc_read_done, cb, priority,
5506 zio_flags | ZIO_FLAG_DONT_CACHE |
5508 ZIO_FLAG_DONT_PROPAGATE |
5509 ZIO_FLAG_DONT_RETRY, B_FALSE);
5510 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5512 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5514 if (*arc_flags & ARC_FLAG_NOWAIT) {
5519 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5520 if (zio_wait(rzio) == 0)
5523 /* l2arc read error; goto zio_read() */
5525 DTRACE_PROBE1(l2arc__miss,
5526 arc_buf_hdr_t *, hdr);
5527 ARCSTAT_BUMP(arcstat_l2_misses);
5528 if (HDR_L2_WRITING(hdr))
5529 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5530 spa_config_exit(spa, SCL_L2ARC, vd);
5534 spa_config_exit(spa, SCL_L2ARC, vd);
5535 if (l2arc_ndev != 0) {
5536 DTRACE_PROBE1(l2arc__miss,
5537 arc_buf_hdr_t *, hdr);
5538 ARCSTAT_BUMP(arcstat_l2_misses);
5542 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
5543 arc_read_done, hdr, priority, zio_flags, zb);
5545 if (*arc_flags & ARC_FLAG_WAIT)
5546 return (zio_wait(rzio));
5548 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5555 * Notify the arc that a block was freed, and thus will never be used again.
5558 arc_freed(spa_t *spa, const blkptr_t *bp)
5561 kmutex_t *hash_lock;
5562 uint64_t guid = spa_load_guid(spa);
5564 ASSERT(!BP_IS_EMBEDDED(bp));
5566 hdr = buf_hash_find(guid, bp, &hash_lock);
5571 * We might be trying to free a block that is still doing I/O
5572 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5573 * dmu_sync-ed block). If this block is being prefetched, then it
5574 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5575 * until the I/O completes. A block may also have a reference if it is
5576 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5577 * have written the new block to its final resting place on disk but
5578 * without the dedup flag set. This would have left the hdr in the MRU
5579 * state and discoverable. When the txg finally syncs it detects that
5580 * the block was overridden in open context and issues an override I/O.
5581 * Since this is a dedup block, the override I/O will determine if the
5582 * block is already in the DDT. If so, then it will replace the io_bp
5583 * with the bp from the DDT and allow the I/O to finish. When the I/O
5584 * reaches the done callback, dbuf_write_override_done, it will
5585 * check to see if the io_bp and io_bp_override are identical.
5586 * If they are not, then it indicates that the bp was replaced with
5587 * the bp in the DDT and the override bp is freed. This allows
5588 * us to arrive here with a reference on a block that is being
5589 * freed. So if we have an I/O in progress, or a reference to
5590 * this hdr, then we don't destroy the hdr.
5592 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5593 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5594 arc_change_state(arc_anon, hdr, hash_lock);
5595 arc_hdr_destroy(hdr);
5596 mutex_exit(hash_lock);
5598 mutex_exit(hash_lock);
5604 * Release this buffer from the cache, making it an anonymous buffer. This
5605 * must be done after a read and prior to modifying the buffer contents.
5606 * If the buffer has more than one reference, we must make
5607 * a new hdr for the buffer.
5610 arc_release(arc_buf_t *buf, void *tag)
5612 arc_buf_hdr_t *hdr = buf->b_hdr;
5615 * It would be nice to assert that if it's DMU metadata (level >
5616 * 0 || it's the dnode file), then it must be syncing context.
5617 * But we don't know that information at this level.
5620 mutex_enter(&buf->b_evict_lock);
5622 ASSERT(HDR_HAS_L1HDR(hdr));
5625 * We don't grab the hash lock prior to this check, because if
5626 * the buffer's header is in the arc_anon state, it won't be
5627 * linked into the hash table.
5629 if (hdr->b_l1hdr.b_state == arc_anon) {
5630 mutex_exit(&buf->b_evict_lock);
5631 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5632 ASSERT(!HDR_IN_HASH_TABLE(hdr));
5633 ASSERT(!HDR_HAS_L2HDR(hdr));
5634 ASSERT(HDR_EMPTY(hdr));
5635 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5636 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5637 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5639 hdr->b_l1hdr.b_arc_access = 0;
5642 * If the buf is being overridden then it may already
5643 * have a hdr that is not empty.
5645 buf_discard_identity(hdr);
5651 kmutex_t *hash_lock = HDR_LOCK(hdr);
5652 mutex_enter(hash_lock);
5655 * This assignment is only valid as long as the hash_lock is
5656 * held, we must be careful not to reference state or the
5657 * b_state field after dropping the lock.
5659 arc_state_t *state = hdr->b_l1hdr.b_state;
5660 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5661 ASSERT3P(state, !=, arc_anon);
5663 /* this buffer is not on any list */
5664 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
5666 if (HDR_HAS_L2HDR(hdr)) {
5667 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5670 * We have to recheck this conditional again now that
5671 * we're holding the l2ad_mtx to prevent a race with
5672 * another thread which might be concurrently calling
5673 * l2arc_evict(). In that case, l2arc_evict() might have
5674 * destroyed the header's L2 portion as we were waiting
5675 * to acquire the l2ad_mtx.
5677 if (HDR_HAS_L2HDR(hdr)) {
5679 arc_hdr_l2hdr_destroy(hdr);
5682 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5686 * Do we have more than one buf?
5688 if (hdr->b_l1hdr.b_bufcnt > 1) {
5689 arc_buf_hdr_t *nhdr;
5690 uint64_t spa = hdr->b_spa;
5691 uint64_t psize = HDR_GET_PSIZE(hdr);
5692 uint64_t lsize = HDR_GET_LSIZE(hdr);
5693 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5694 arc_buf_contents_t type = arc_buf_type(hdr);
5695 VERIFY3U(hdr->b_type, ==, type);
5697 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5698 (void) remove_reference(hdr, hash_lock, tag);
5700 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
5701 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5702 ASSERT(ARC_BUF_LAST(buf));
5706 * Pull the data off of this hdr and attach it to
5707 * a new anonymous hdr. Also find the last buffer
5708 * in the hdr's buffer list.
5710 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
5711 ASSERT3P(lastbuf, !=, NULL);
5714 * If the current arc_buf_t and the hdr are sharing their data
5715 * buffer, then we must stop sharing that block.
5717 if (arc_buf_is_shared(buf)) {
5718 VERIFY(!arc_buf_is_shared(lastbuf));
5721 * First, sever the block sharing relationship between
5722 * buf and the arc_buf_hdr_t.
5724 arc_unshare_buf(hdr, buf);
5727 * Now we need to recreate the hdr's b_pabd. Since we
5728 * have lastbuf handy, we try to share with it, but if
5729 * we can't then we allocate a new b_pabd and copy the
5730 * data from buf into it.
5732 if (arc_can_share(hdr, lastbuf)) {
5733 arc_share_buf(hdr, lastbuf);
5735 arc_hdr_alloc_pabd(hdr);
5736 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
5737 buf->b_data, psize);
5739 VERIFY3P(lastbuf->b_data, !=, NULL);
5740 } else if (HDR_SHARED_DATA(hdr)) {
5742 * Uncompressed shared buffers are always at the end
5743 * of the list. Compressed buffers don't have the
5744 * same requirements. This makes it hard to
5745 * simply assert that the lastbuf is shared so
5746 * we rely on the hdr's compression flags to determine
5747 * if we have a compressed, shared buffer.
5749 ASSERT(arc_buf_is_shared(lastbuf) ||
5750 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
5751 ASSERT(!ARC_BUF_SHARED(buf));
5753 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5754 ASSERT3P(state, !=, arc_l2c_only);
5756 (void) refcount_remove_many(&state->arcs_size,
5757 arc_buf_size(buf), buf);
5759 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5760 ASSERT3P(state, !=, arc_l2c_only);
5761 (void) refcount_remove_many(&state->arcs_esize[type],
5762 arc_buf_size(buf), buf);
5765 hdr->b_l1hdr.b_bufcnt -= 1;
5766 arc_cksum_verify(buf);
5768 arc_buf_unwatch(buf);
5771 mutex_exit(hash_lock);
5774 * Allocate a new hdr. The new hdr will contain a b_pabd
5775 * buffer which will be freed in arc_write().
5777 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5778 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5779 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5780 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5781 VERIFY3U(nhdr->b_type, ==, type);
5782 ASSERT(!HDR_SHARED_DATA(nhdr));
5784 nhdr->b_l1hdr.b_buf = buf;
5785 nhdr->b_l1hdr.b_bufcnt = 1;
5786 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5789 mutex_exit(&buf->b_evict_lock);
5790 (void) refcount_add_many(&arc_anon->arcs_size,
5791 arc_buf_size(buf), buf);
5793 mutex_exit(&buf->b_evict_lock);
5794 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5795 /* protected by hash lock, or hdr is on arc_anon */
5796 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5797 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5798 arc_change_state(arc_anon, hdr, hash_lock);
5799 hdr->b_l1hdr.b_arc_access = 0;
5800 mutex_exit(hash_lock);
5802 buf_discard_identity(hdr);
5808 arc_released(arc_buf_t *buf)
5812 mutex_enter(&buf->b_evict_lock);
5813 released = (buf->b_data != NULL &&
5814 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5815 mutex_exit(&buf->b_evict_lock);
5821 arc_referenced(arc_buf_t *buf)
5825 mutex_enter(&buf->b_evict_lock);
5826 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5827 mutex_exit(&buf->b_evict_lock);
5828 return (referenced);
5833 arc_write_ready(zio_t *zio)
5835 arc_write_callback_t *callback = zio->io_private;
5836 arc_buf_t *buf = callback->awcb_buf;
5837 arc_buf_hdr_t *hdr = buf->b_hdr;
5838 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5840 ASSERT(HDR_HAS_L1HDR(hdr));
5841 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5842 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5845 * If we're reexecuting this zio because the pool suspended, then
5846 * cleanup any state that was previously set the first time the
5847 * callback was invoked.
5849 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5850 arc_cksum_free(hdr);
5852 arc_buf_unwatch(buf);
5854 if (hdr->b_l1hdr.b_pabd != NULL) {
5855 if (arc_buf_is_shared(buf)) {
5856 arc_unshare_buf(hdr, buf);
5858 arc_hdr_free_pabd(hdr);
5862 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5863 ASSERT(!HDR_SHARED_DATA(hdr));
5864 ASSERT(!arc_buf_is_shared(buf));
5866 callback->awcb_ready(zio, buf, callback->awcb_private);
5868 if (HDR_IO_IN_PROGRESS(hdr))
5869 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
5871 arc_cksum_compute(buf);
5872 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5874 enum zio_compress compress;
5875 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5876 compress = ZIO_COMPRESS_OFF;
5878 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
5879 compress = BP_GET_COMPRESS(zio->io_bp);
5881 HDR_SET_PSIZE(hdr, psize);
5882 arc_hdr_set_compress(hdr, compress);
5886 * Fill the hdr with data. If the hdr is compressed, the data we want
5887 * is available from the zio, otherwise we can take it from the buf.
5889 * We might be able to share the buf's data with the hdr here. However,
5890 * doing so would cause the ARC to be full of linear ABDs if we write a
5891 * lot of shareable data. As a compromise, we check whether scattered
5892 * ABDs are allowed, and assume that if they are then the user wants
5893 * the ARC to be primarily filled with them regardless of the data being
5894 * written. Therefore, if they're allowed then we allocate one and copy
5895 * the data into it; otherwise, we share the data directly if we can.
5897 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
5898 arc_hdr_alloc_pabd(hdr);
5901 * Ideally, we would always copy the io_abd into b_pabd, but the
5902 * user may have disabled compressed ARC, thus we must check the
5903 * hdr's compression setting rather than the io_bp's.
5905 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5906 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
5908 ASSERT3U(psize, >, 0);
5910 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
5912 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
5914 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
5918 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
5919 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
5920 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5922 arc_share_buf(hdr, buf);
5925 arc_hdr_verify(hdr, zio->io_bp);
5929 arc_write_children_ready(zio_t *zio)
5931 arc_write_callback_t *callback = zio->io_private;
5932 arc_buf_t *buf = callback->awcb_buf;
5934 callback->awcb_children_ready(zio, buf, callback->awcb_private);
5938 * The SPA calls this callback for each physical write that happens on behalf
5939 * of a logical write. See the comment in dbuf_write_physdone() for details.
5942 arc_write_physdone(zio_t *zio)
5944 arc_write_callback_t *cb = zio->io_private;
5945 if (cb->awcb_physdone != NULL)
5946 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
5950 arc_write_done(zio_t *zio)
5952 arc_write_callback_t *callback = zio->io_private;
5953 arc_buf_t *buf = callback->awcb_buf;
5954 arc_buf_hdr_t *hdr = buf->b_hdr;
5956 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5958 if (zio->io_error == 0) {
5959 arc_hdr_verify(hdr, zio->io_bp);
5961 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5962 buf_discard_identity(hdr);
5964 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
5965 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
5968 ASSERT(HDR_EMPTY(hdr));
5972 * If the block to be written was all-zero or compressed enough to be
5973 * embedded in the BP, no write was performed so there will be no
5974 * dva/birth/checksum. The buffer must therefore remain anonymous
5977 if (!HDR_EMPTY(hdr)) {
5978 arc_buf_hdr_t *exists;
5979 kmutex_t *hash_lock;
5981 ASSERT3U(zio->io_error, ==, 0);
5983 arc_cksum_verify(buf);
5985 exists = buf_hash_insert(hdr, &hash_lock);
5986 if (exists != NULL) {
5988 * This can only happen if we overwrite for
5989 * sync-to-convergence, because we remove
5990 * buffers from the hash table when we arc_free().
5992 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5993 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5994 panic("bad overwrite, hdr=%p exists=%p",
5995 (void *)hdr, (void *)exists);
5996 ASSERT(refcount_is_zero(
5997 &exists->b_l1hdr.b_refcnt));
5998 arc_change_state(arc_anon, exists, hash_lock);
5999 mutex_exit(hash_lock);
6000 arc_hdr_destroy(exists);
6001 exists = buf_hash_insert(hdr, &hash_lock);
6002 ASSERT3P(exists, ==, NULL);
6003 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6005 ASSERT(zio->io_prop.zp_nopwrite);
6006 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6007 panic("bad nopwrite, hdr=%p exists=%p",
6008 (void *)hdr, (void *)exists);
6011 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6012 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6013 ASSERT(BP_GET_DEDUP(zio->io_bp));
6014 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6017 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6018 /* if it's not anon, we are doing a scrub */
6019 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6020 arc_access(hdr, hash_lock);
6021 mutex_exit(hash_lock);
6023 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6026 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6027 callback->awcb_done(zio, buf, callback->awcb_private);
6029 abd_put(zio->io_abd);
6030 kmem_free(callback, sizeof (arc_write_callback_t));
6034 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6035 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
6036 arc_done_func_t *children_ready, arc_done_func_t *physdone,
6037 arc_done_func_t *done, void *private, zio_priority_t priority,
6038 int zio_flags, const zbookmark_phys_t *zb)
6040 arc_buf_hdr_t *hdr = buf->b_hdr;
6041 arc_write_callback_t *callback;
6043 zio_prop_t localprop = *zp;
6045 ASSERT3P(ready, !=, NULL);
6046 ASSERT3P(done, !=, NULL);
6047 ASSERT(!HDR_IO_ERROR(hdr));
6048 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6049 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6050 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6052 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6053 if (ARC_BUF_COMPRESSED(buf)) {
6055 * We're writing a pre-compressed buffer. Make the
6056 * compression algorithm requested by the zio_prop_t match
6057 * the pre-compressed buffer's compression algorithm.
6059 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6061 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6062 zio_flags |= ZIO_FLAG_RAW;
6064 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6065 callback->awcb_ready = ready;
6066 callback->awcb_children_ready = children_ready;
6067 callback->awcb_physdone = physdone;
6068 callback->awcb_done = done;
6069 callback->awcb_private = private;
6070 callback->awcb_buf = buf;
6073 * The hdr's b_pabd is now stale, free it now. A new data block
6074 * will be allocated when the zio pipeline calls arc_write_ready().
6076 if (hdr->b_l1hdr.b_pabd != NULL) {
6078 * If the buf is currently sharing the data block with
6079 * the hdr then we need to break that relationship here.
6080 * The hdr will remain with a NULL data pointer and the
6081 * buf will take sole ownership of the block.
6083 if (arc_buf_is_shared(buf)) {
6084 arc_unshare_buf(hdr, buf);
6086 arc_hdr_free_pabd(hdr);
6088 VERIFY3P(buf->b_data, !=, NULL);
6089 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6091 ASSERT(!arc_buf_is_shared(buf));
6092 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6094 zio = zio_write(pio, spa, txg, bp,
6095 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6096 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6097 (children_ready != NULL) ? arc_write_children_ready : NULL,
6098 arc_write_physdone, arc_write_done, callback,
6099 priority, zio_flags, zb);
6105 arc_memory_throttle(uint64_t reserve, uint64_t txg)
6108 uint64_t available_memory = ptob(freemem);
6109 static uint64_t page_load = 0;
6110 static uint64_t last_txg = 0;
6112 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
6114 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE)));
6117 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
6120 if (txg > last_txg) {
6125 * If we are in pageout, we know that memory is already tight,
6126 * the arc is already going to be evicting, so we just want to
6127 * continue to let page writes occur as quickly as possible.
6129 if (curproc == pageproc) {
6130 if (page_load > MAX(ptob(minfree), available_memory) / 4)
6131 return (SET_ERROR(ERESTART));
6132 /* Note: reserve is inflated, so we deflate */
6133 page_load += reserve / 8;
6135 } else if (page_load > 0 && arc_reclaim_needed()) {
6136 /* memory is low, delay before restarting */
6137 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6138 return (SET_ERROR(EAGAIN));
6146 arc_tempreserve_clear(uint64_t reserve)
6148 atomic_add_64(&arc_tempreserve, -reserve);
6149 ASSERT((int64_t)arc_tempreserve >= 0);
6153 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
6158 if (reserve > arc_c/4 && !arc_no_grow) {
6159 arc_c = MIN(arc_c_max, reserve * 4);
6160 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
6162 if (reserve > arc_c)
6163 return (SET_ERROR(ENOMEM));
6166 * Don't count loaned bufs as in flight dirty data to prevent long
6167 * network delays from blocking transactions that are ready to be
6168 * assigned to a txg.
6171 /* assert that it has not wrapped around */
6172 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6174 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6175 arc_loaned_bytes), 0);
6178 * Writes will, almost always, require additional memory allocations
6179 * in order to compress/encrypt/etc the data. We therefore need to
6180 * make sure that there is sufficient available memory for this.
6182 error = arc_memory_throttle(reserve, txg);
6187 * Throttle writes when the amount of dirty data in the cache
6188 * gets too large. We try to keep the cache less than half full
6189 * of dirty blocks so that our sync times don't grow too large.
6190 * Note: if two requests come in concurrently, we might let them
6191 * both succeed, when one of them should fail. Not a huge deal.
6194 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
6195 anon_size > arc_c / 4) {
6196 uint64_t meta_esize =
6197 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6198 uint64_t data_esize =
6199 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6200 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6201 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6202 arc_tempreserve >> 10, meta_esize >> 10,
6203 data_esize >> 10, reserve >> 10, arc_c >> 10);
6204 return (SET_ERROR(ERESTART));
6206 atomic_add_64(&arc_tempreserve, reserve);
6211 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6212 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6214 size->value.ui64 = refcount_count(&state->arcs_size);
6215 evict_data->value.ui64 =
6216 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6217 evict_metadata->value.ui64 =
6218 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6222 arc_kstat_update(kstat_t *ksp, int rw)
6224 arc_stats_t *as = ksp->ks_data;
6226 if (rw == KSTAT_WRITE) {
6229 arc_kstat_update_state(arc_anon,
6230 &as->arcstat_anon_size,
6231 &as->arcstat_anon_evictable_data,
6232 &as->arcstat_anon_evictable_metadata);
6233 arc_kstat_update_state(arc_mru,
6234 &as->arcstat_mru_size,
6235 &as->arcstat_mru_evictable_data,
6236 &as->arcstat_mru_evictable_metadata);
6237 arc_kstat_update_state(arc_mru_ghost,
6238 &as->arcstat_mru_ghost_size,
6239 &as->arcstat_mru_ghost_evictable_data,
6240 &as->arcstat_mru_ghost_evictable_metadata);
6241 arc_kstat_update_state(arc_mfu,
6242 &as->arcstat_mfu_size,
6243 &as->arcstat_mfu_evictable_data,
6244 &as->arcstat_mfu_evictable_metadata);
6245 arc_kstat_update_state(arc_mfu_ghost,
6246 &as->arcstat_mfu_ghost_size,
6247 &as->arcstat_mfu_ghost_evictable_data,
6248 &as->arcstat_mfu_ghost_evictable_metadata);
6255 * This function *must* return indices evenly distributed between all
6256 * sublists of the multilist. This is needed due to how the ARC eviction
6257 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6258 * distributed between all sublists and uses this assumption when
6259 * deciding which sublist to evict from and how much to evict from it.
6262 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6264 arc_buf_hdr_t *hdr = obj;
6267 * We rely on b_dva to generate evenly distributed index
6268 * numbers using buf_hash below. So, as an added precaution,
6269 * let's make sure we never add empty buffers to the arc lists.
6271 ASSERT(!HDR_EMPTY(hdr));
6274 * The assumption here, is the hash value for a given
6275 * arc_buf_hdr_t will remain constant throughout it's lifetime
6276 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6277 * Thus, we don't need to store the header's sublist index
6278 * on insertion, as this index can be recalculated on removal.
6280 * Also, the low order bits of the hash value are thought to be
6281 * distributed evenly. Otherwise, in the case that the multilist
6282 * has a power of two number of sublists, each sublists' usage
6283 * would not be evenly distributed.
6285 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6286 multilist_get_num_sublists(ml));
6290 static eventhandler_tag arc_event_lowmem = NULL;
6293 arc_lowmem(void *arg __unused, int howto __unused)
6296 mutex_enter(&arc_reclaim_lock);
6297 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE);
6298 cv_signal(&arc_reclaim_thread_cv);
6301 * It is unsafe to block here in arbitrary threads, because we can come
6302 * here from ARC itself and may hold ARC locks and thus risk a deadlock
6303 * with ARC reclaim thread.
6305 if (curproc == pageproc)
6306 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
6307 mutex_exit(&arc_reclaim_lock);
6312 arc_state_init(void)
6314 arc_anon = &ARC_anon;
6316 arc_mru_ghost = &ARC_mru_ghost;
6318 arc_mfu_ghost = &ARC_mfu_ghost;
6319 arc_l2c_only = &ARC_l2c_only;
6321 arc_mru->arcs_list[ARC_BUFC_METADATA] =
6322 multilist_create(sizeof (arc_buf_hdr_t),
6323 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6324 arc_state_multilist_index_func);
6325 arc_mru->arcs_list[ARC_BUFC_DATA] =
6326 multilist_create(sizeof (arc_buf_hdr_t),
6327 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6328 arc_state_multilist_index_func);
6329 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6330 multilist_create(sizeof (arc_buf_hdr_t),
6331 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6332 arc_state_multilist_index_func);
6333 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6334 multilist_create(sizeof (arc_buf_hdr_t),
6335 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6336 arc_state_multilist_index_func);
6337 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6338 multilist_create(sizeof (arc_buf_hdr_t),
6339 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6340 arc_state_multilist_index_func);
6341 arc_mfu->arcs_list[ARC_BUFC_DATA] =
6342 multilist_create(sizeof (arc_buf_hdr_t),
6343 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6344 arc_state_multilist_index_func);
6345 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6346 multilist_create(sizeof (arc_buf_hdr_t),
6347 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6348 arc_state_multilist_index_func);
6349 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6350 multilist_create(sizeof (arc_buf_hdr_t),
6351 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6352 arc_state_multilist_index_func);
6353 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6354 multilist_create(sizeof (arc_buf_hdr_t),
6355 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6356 arc_state_multilist_index_func);
6357 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6358 multilist_create(sizeof (arc_buf_hdr_t),
6359 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6360 arc_state_multilist_index_func);
6362 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6363 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6364 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6365 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6366 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6367 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6368 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6369 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6370 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6371 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6372 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6373 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6375 refcount_create(&arc_anon->arcs_size);
6376 refcount_create(&arc_mru->arcs_size);
6377 refcount_create(&arc_mru_ghost->arcs_size);
6378 refcount_create(&arc_mfu->arcs_size);
6379 refcount_create(&arc_mfu_ghost->arcs_size);
6380 refcount_create(&arc_l2c_only->arcs_size);
6384 arc_state_fini(void)
6386 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6387 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6388 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6389 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6390 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6391 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6392 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6393 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6394 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6395 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6396 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6397 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6399 refcount_destroy(&arc_anon->arcs_size);
6400 refcount_destroy(&arc_mru->arcs_size);
6401 refcount_destroy(&arc_mru_ghost->arcs_size);
6402 refcount_destroy(&arc_mfu->arcs_size);
6403 refcount_destroy(&arc_mfu_ghost->arcs_size);
6404 refcount_destroy(&arc_l2c_only->arcs_size);
6406 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6407 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6408 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6409 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6410 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6411 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6412 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6413 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6425 int i, prefetch_tunable_set = 0;
6428 * allmem is "all memory that we could possibly use".
6432 uint64_t allmem = ptob(physmem - swapfs_minfree);
6434 uint64_t allmem = (physmem * PAGESIZE) / 2;
6437 uint64_t allmem = kmem_size();
6441 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
6442 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
6443 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
6445 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
6446 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
6448 /* Convert seconds to clock ticks */
6449 arc_min_prefetch_lifespan = 1 * hz;
6451 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
6452 arc_c_min = MAX(allmem / 32, arc_abs_min);
6453 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */
6454 if (allmem >= 1 << 30)
6455 arc_c_max = allmem - (1 << 30);
6457 arc_c_max = arc_c_min;
6458 arc_c_max = MAX(allmem * 5 / 8, arc_c_max);
6461 * In userland, there's only the memory pressure that we artificially
6462 * create (see arc_available_memory()). Don't let arc_c get too
6463 * small, because it can cause transactions to be larger than
6464 * arc_c, causing arc_tempreserve_space() to fail.
6467 arc_c_min = arc_c_max / 2;
6472 * Allow the tunables to override our calculations if they are
6475 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) {
6476 arc_c_max = zfs_arc_max;
6477 arc_c_min = MIN(arc_c_min, arc_c_max);
6479 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
6480 arc_c_min = zfs_arc_min;
6484 arc_p = (arc_c >> 1);
6487 /* limit meta-data to 1/4 of the arc capacity */
6488 arc_meta_limit = arc_c_max / 4;
6492 * Metadata is stored in the kernel's heap. Don't let us
6493 * use more than half the heap for the ARC.
6495 arc_meta_limit = MIN(arc_meta_limit,
6496 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6499 /* Allow the tunable to override if it is reasonable */
6500 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6501 arc_meta_limit = zfs_arc_meta_limit;
6503 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6504 arc_c_min = arc_meta_limit / 2;
6506 if (zfs_arc_meta_min > 0) {
6507 arc_meta_min = zfs_arc_meta_min;
6509 arc_meta_min = arc_c_min / 2;
6512 if (zfs_arc_grow_retry > 0)
6513 arc_grow_retry = zfs_arc_grow_retry;
6515 if (zfs_arc_shrink_shift > 0)
6516 arc_shrink_shift = zfs_arc_shrink_shift;
6518 if (zfs_arc_no_grow_shift > 0)
6519 arc_no_grow_shift = zfs_arc_no_grow_shift;
6521 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6523 if (arc_no_grow_shift >= arc_shrink_shift)
6524 arc_no_grow_shift = arc_shrink_shift - 1;
6526 if (zfs_arc_p_min_shift > 0)
6527 arc_p_min_shift = zfs_arc_p_min_shift;
6529 /* if kmem_flags are set, lets try to use less memory */
6530 if (kmem_debugging())
6532 if (arc_c < arc_c_min)
6535 zfs_arc_min = arc_c_min;
6536 zfs_arc_max = arc_c_max;
6541 arc_reclaim_thread_exit = B_FALSE;
6542 arc_dnlc_evicts_thread_exit = FALSE;
6544 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6545 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6547 if (arc_ksp != NULL) {
6548 arc_ksp->ks_data = &arc_stats;
6549 arc_ksp->ks_update = arc_kstat_update;
6550 kstat_install(arc_ksp);
6553 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6554 TS_RUN, minclsyspri);
6557 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
6558 EVENTHANDLER_PRI_FIRST);
6561 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
6562 TS_RUN, minclsyspri);
6568 * Calculate maximum amount of dirty data per pool.
6570 * If it has been set by /etc/system, take that.
6571 * Otherwise, use a percentage of physical memory defined by
6572 * zfs_dirty_data_max_percent (default 10%) with a cap at
6573 * zfs_dirty_data_max_max (default 4GB).
6575 if (zfs_dirty_data_max == 0) {
6576 zfs_dirty_data_max = ptob(physmem) *
6577 zfs_dirty_data_max_percent / 100;
6578 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6579 zfs_dirty_data_max_max);
6583 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
6584 prefetch_tunable_set = 1;
6587 if (prefetch_tunable_set == 0) {
6588 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
6590 printf(" add \"vfs.zfs.prefetch_disable=0\" "
6591 "to /boot/loader.conf.\n");
6592 zfs_prefetch_disable = 1;
6595 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
6596 prefetch_tunable_set == 0) {
6597 printf("ZFS NOTICE: Prefetch is disabled by default if less "
6598 "than 4GB of RAM is present;\n"
6599 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
6600 "to /boot/loader.conf.\n");
6601 zfs_prefetch_disable = 1;
6604 /* Warn about ZFS memory and address space requirements. */
6605 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
6606 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
6607 "expect unstable behavior.\n");
6609 if (allmem < 512 * (1 << 20)) {
6610 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
6611 "expect unstable behavior.\n");
6612 printf(" Consider tuning vm.kmem_size and "
6613 "vm.kmem_size_max\n");
6614 printf(" in /boot/loader.conf.\n");
6622 mutex_enter(&arc_reclaim_lock);
6623 arc_reclaim_thread_exit = B_TRUE;
6625 * The reclaim thread will set arc_reclaim_thread_exit back to
6626 * B_FALSE when it is finished exiting; we're waiting for that.
6628 while (arc_reclaim_thread_exit) {
6629 cv_signal(&arc_reclaim_thread_cv);
6630 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6632 mutex_exit(&arc_reclaim_lock);
6634 /* Use B_TRUE to ensure *all* buffers are evicted */
6635 arc_flush(NULL, B_TRUE);
6637 mutex_enter(&arc_dnlc_evicts_lock);
6638 arc_dnlc_evicts_thread_exit = TRUE;
6640 * The user evicts thread will set arc_user_evicts_thread_exit
6641 * to FALSE when it is finished exiting; we're waiting for that.
6643 while (arc_dnlc_evicts_thread_exit) {
6644 cv_signal(&arc_dnlc_evicts_cv);
6645 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
6647 mutex_exit(&arc_dnlc_evicts_lock);
6651 if (arc_ksp != NULL) {
6652 kstat_delete(arc_ksp);
6656 mutex_destroy(&arc_reclaim_lock);
6657 cv_destroy(&arc_reclaim_thread_cv);
6658 cv_destroy(&arc_reclaim_waiters_cv);
6660 mutex_destroy(&arc_dnlc_evicts_lock);
6661 cv_destroy(&arc_dnlc_evicts_cv);
6666 ASSERT0(arc_loaned_bytes);
6669 if (arc_event_lowmem != NULL)
6670 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
6677 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6678 * It uses dedicated storage devices to hold cached data, which are populated
6679 * using large infrequent writes. The main role of this cache is to boost
6680 * the performance of random read workloads. The intended L2ARC devices
6681 * include short-stroked disks, solid state disks, and other media with
6682 * substantially faster read latency than disk.
6684 * +-----------------------+
6686 * +-----------------------+
6689 * l2arc_feed_thread() arc_read()
6693 * +---------------+ |
6695 * +---------------+ |
6700 * +-------+ +-------+
6702 * | cache | | cache |
6703 * +-------+ +-------+
6704 * +=========+ .-----.
6705 * : L2ARC : |-_____-|
6706 * : devices : | Disks |
6707 * +=========+ `-_____-'
6709 * Read requests are satisfied from the following sources, in order:
6712 * 2) vdev cache of L2ARC devices
6714 * 4) vdev cache of disks
6717 * Some L2ARC device types exhibit extremely slow write performance.
6718 * To accommodate for this there are some significant differences between
6719 * the L2ARC and traditional cache design:
6721 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6722 * the ARC behave as usual, freeing buffers and placing headers on ghost
6723 * lists. The ARC does not send buffers to the L2ARC during eviction as
6724 * this would add inflated write latencies for all ARC memory pressure.
6726 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6727 * It does this by periodically scanning buffers from the eviction-end of
6728 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6729 * not already there. It scans until a headroom of buffers is satisfied,
6730 * which itself is a buffer for ARC eviction. If a compressible buffer is
6731 * found during scanning and selected for writing to an L2ARC device, we
6732 * temporarily boost scanning headroom during the next scan cycle to make
6733 * sure we adapt to compression effects (which might significantly reduce
6734 * the data volume we write to L2ARC). The thread that does this is
6735 * l2arc_feed_thread(), illustrated below; example sizes are included to
6736 * provide a better sense of ratio than this diagram:
6739 * +---------------------+----------+
6740 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6741 * +---------------------+----------+ | o L2ARC eligible
6742 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6743 * +---------------------+----------+ |
6744 * 15.9 Gbytes ^ 32 Mbytes |
6746 * l2arc_feed_thread()
6748 * l2arc write hand <--[oooo]--'
6752 * +==============================+
6753 * L2ARC dev |####|#|###|###| |####| ... |
6754 * +==============================+
6757 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6758 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6759 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6760 * safe to say that this is an uncommon case, since buffers at the end of
6761 * the ARC lists have moved there due to inactivity.
6763 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6764 * then the L2ARC simply misses copying some buffers. This serves as a
6765 * pressure valve to prevent heavy read workloads from both stalling the ARC
6766 * with waits and clogging the L2ARC with writes. This also helps prevent
6767 * the potential for the L2ARC to churn if it attempts to cache content too
6768 * quickly, such as during backups of the entire pool.
6770 * 5. After system boot and before the ARC has filled main memory, there are
6771 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6772 * lists can remain mostly static. Instead of searching from tail of these
6773 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6774 * for eligible buffers, greatly increasing its chance of finding them.
6776 * The L2ARC device write speed is also boosted during this time so that
6777 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6778 * there are no L2ARC reads, and no fear of degrading read performance
6779 * through increased writes.
6781 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6782 * the vdev queue can aggregate them into larger and fewer writes. Each
6783 * device is written to in a rotor fashion, sweeping writes through
6784 * available space then repeating.
6786 * 7. The L2ARC does not store dirty content. It never needs to flush
6787 * write buffers back to disk based storage.
6789 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6790 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6792 * The performance of the L2ARC can be tweaked by a number of tunables, which
6793 * may be necessary for different workloads:
6795 * l2arc_write_max max write bytes per interval
6796 * l2arc_write_boost extra write bytes during device warmup
6797 * l2arc_noprefetch skip caching prefetched buffers
6798 * l2arc_headroom number of max device writes to precache
6799 * l2arc_headroom_boost when we find compressed buffers during ARC
6800 * scanning, we multiply headroom by this
6801 * percentage factor for the next scan cycle,
6802 * since more compressed buffers are likely to
6804 * l2arc_feed_secs seconds between L2ARC writing
6806 * Tunables may be removed or added as future performance improvements are
6807 * integrated, and also may become zpool properties.
6809 * There are three key functions that control how the L2ARC warms up:
6811 * l2arc_write_eligible() check if a buffer is eligible to cache
6812 * l2arc_write_size() calculate how much to write
6813 * l2arc_write_interval() calculate sleep delay between writes
6815 * These three functions determine what to write, how much, and how quickly
6820 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6823 * A buffer is *not* eligible for the L2ARC if it:
6824 * 1. belongs to a different spa.
6825 * 2. is already cached on the L2ARC.
6826 * 3. has an I/O in progress (it may be an incomplete read).
6827 * 4. is flagged not eligible (zfs property).
6829 if (hdr->b_spa != spa_guid) {
6830 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
6833 if (HDR_HAS_L2HDR(hdr)) {
6834 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
6837 if (HDR_IO_IN_PROGRESS(hdr)) {
6838 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
6841 if (!HDR_L2CACHE(hdr)) {
6842 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
6850 l2arc_write_size(void)
6855 * Make sure our globals have meaningful values in case the user
6858 size = l2arc_write_max;
6860 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
6861 "be greater than zero, resetting it to the default (%d)",
6863 size = l2arc_write_max = L2ARC_WRITE_SIZE;
6866 if (arc_warm == B_FALSE)
6867 size += l2arc_write_boost;
6874 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
6876 clock_t interval, next, now;
6879 * If the ARC lists are busy, increase our write rate; if the
6880 * lists are stale, idle back. This is achieved by checking
6881 * how much we previously wrote - if it was more than half of
6882 * what we wanted, schedule the next write much sooner.
6884 if (l2arc_feed_again && wrote > (wanted / 2))
6885 interval = (hz * l2arc_feed_min_ms) / 1000;
6887 interval = hz * l2arc_feed_secs;
6889 now = ddi_get_lbolt();
6890 next = MAX(now, MIN(now + interval, began + interval));
6896 * Cycle through L2ARC devices. This is how L2ARC load balances.
6897 * If a device is returned, this also returns holding the spa config lock.
6899 static l2arc_dev_t *
6900 l2arc_dev_get_next(void)
6902 l2arc_dev_t *first, *next = NULL;
6905 * Lock out the removal of spas (spa_namespace_lock), then removal
6906 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6907 * both locks will be dropped and a spa config lock held instead.
6909 mutex_enter(&spa_namespace_lock);
6910 mutex_enter(&l2arc_dev_mtx);
6912 /* if there are no vdevs, there is nothing to do */
6913 if (l2arc_ndev == 0)
6917 next = l2arc_dev_last;
6919 /* loop around the list looking for a non-faulted vdev */
6921 next = list_head(l2arc_dev_list);
6923 next = list_next(l2arc_dev_list, next);
6925 next = list_head(l2arc_dev_list);
6928 /* if we have come back to the start, bail out */
6931 else if (next == first)
6934 } while (vdev_is_dead(next->l2ad_vdev));
6936 /* if we were unable to find any usable vdevs, return NULL */
6937 if (vdev_is_dead(next->l2ad_vdev))
6940 l2arc_dev_last = next;
6943 mutex_exit(&l2arc_dev_mtx);
6946 * Grab the config lock to prevent the 'next' device from being
6947 * removed while we are writing to it.
6950 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
6951 mutex_exit(&spa_namespace_lock);
6957 * Free buffers that were tagged for destruction.
6960 l2arc_do_free_on_write()
6963 l2arc_data_free_t *df, *df_prev;
6965 mutex_enter(&l2arc_free_on_write_mtx);
6966 buflist = l2arc_free_on_write;
6968 for (df = list_tail(buflist); df; df = df_prev) {
6969 df_prev = list_prev(buflist, df);
6970 ASSERT3P(df->l2df_abd, !=, NULL);
6971 abd_free(df->l2df_abd);
6972 list_remove(buflist, df);
6973 kmem_free(df, sizeof (l2arc_data_free_t));
6976 mutex_exit(&l2arc_free_on_write_mtx);
6980 * A write to a cache device has completed. Update all headers to allow
6981 * reads from these buffers to begin.
6984 l2arc_write_done(zio_t *zio)
6986 l2arc_write_callback_t *cb;
6989 arc_buf_hdr_t *head, *hdr, *hdr_prev;
6990 kmutex_t *hash_lock;
6991 int64_t bytes_dropped = 0;
6993 cb = zio->io_private;
6994 ASSERT3P(cb, !=, NULL);
6995 dev = cb->l2wcb_dev;
6996 ASSERT3P(dev, !=, NULL);
6997 head = cb->l2wcb_head;
6998 ASSERT3P(head, !=, NULL);
6999 buflist = &dev->l2ad_buflist;
7000 ASSERT3P(buflist, !=, NULL);
7001 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7002 l2arc_write_callback_t *, cb);
7004 if (zio->io_error != 0)
7005 ARCSTAT_BUMP(arcstat_l2_writes_error);
7008 * All writes completed, or an error was hit.
7011 mutex_enter(&dev->l2ad_mtx);
7012 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7013 hdr_prev = list_prev(buflist, hdr);
7015 hash_lock = HDR_LOCK(hdr);
7018 * We cannot use mutex_enter or else we can deadlock
7019 * with l2arc_write_buffers (due to swapping the order
7020 * the hash lock and l2ad_mtx are taken).
7022 if (!mutex_tryenter(hash_lock)) {
7024 * Missed the hash lock. We must retry so we
7025 * don't leave the ARC_FLAG_L2_WRITING bit set.
7027 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7030 * We don't want to rescan the headers we've
7031 * already marked as having been written out, so
7032 * we reinsert the head node so we can pick up
7033 * where we left off.
7035 list_remove(buflist, head);
7036 list_insert_after(buflist, hdr, head);
7038 mutex_exit(&dev->l2ad_mtx);
7041 * We wait for the hash lock to become available
7042 * to try and prevent busy waiting, and increase
7043 * the chance we'll be able to acquire the lock
7044 * the next time around.
7046 mutex_enter(hash_lock);
7047 mutex_exit(hash_lock);
7052 * We could not have been moved into the arc_l2c_only
7053 * state while in-flight due to our ARC_FLAG_L2_WRITING
7054 * bit being set. Let's just ensure that's being enforced.
7056 ASSERT(HDR_HAS_L1HDR(hdr));
7058 if (zio->io_error != 0) {
7060 * Error - drop L2ARC entry.
7062 list_remove(buflist, hdr);
7064 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7066 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
7067 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
7069 bytes_dropped += arc_hdr_size(hdr);
7070 (void) refcount_remove_many(&dev->l2ad_alloc,
7071 arc_hdr_size(hdr), hdr);
7075 * Allow ARC to begin reads and ghost list evictions to
7078 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7080 mutex_exit(hash_lock);
7083 atomic_inc_64(&l2arc_writes_done);
7084 list_remove(buflist, head);
7085 ASSERT(!HDR_HAS_L1HDR(head));
7086 kmem_cache_free(hdr_l2only_cache, head);
7087 mutex_exit(&dev->l2ad_mtx);
7089 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7091 l2arc_do_free_on_write();
7093 kmem_free(cb, sizeof (l2arc_write_callback_t));
7097 * A read to a cache device completed. Validate buffer contents before
7098 * handing over to the regular ARC routines.
7101 l2arc_read_done(zio_t *zio)
7103 l2arc_read_callback_t *cb;
7105 kmutex_t *hash_lock;
7106 boolean_t valid_cksum;
7108 ASSERT3P(zio->io_vd, !=, NULL);
7109 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7111 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7113 cb = zio->io_private;
7114 ASSERT3P(cb, !=, NULL);
7115 hdr = cb->l2rcb_hdr;
7116 ASSERT3P(hdr, !=, NULL);
7118 hash_lock = HDR_LOCK(hdr);
7119 mutex_enter(hash_lock);
7120 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7123 * If the data was read into a temporary buffer,
7124 * move it and free the buffer.
7126 if (cb->l2rcb_abd != NULL) {
7127 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7128 if (zio->io_error == 0) {
7129 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
7134 * The following must be done regardless of whether
7135 * there was an error:
7136 * - free the temporary buffer
7137 * - point zio to the real ARC buffer
7138 * - set zio size accordingly
7139 * These are required because zio is either re-used for
7140 * an I/O of the block in the case of the error
7141 * or the zio is passed to arc_read_done() and it
7144 abd_free(cb->l2rcb_abd);
7145 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7146 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
7149 ASSERT3P(zio->io_abd, !=, NULL);
7152 * Check this survived the L2ARC journey.
7154 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
7155 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
7156 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
7158 valid_cksum = arc_cksum_is_equal(hdr, zio);
7159 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
7160 mutex_exit(hash_lock);
7161 zio->io_private = hdr;
7164 mutex_exit(hash_lock);
7166 * Buffer didn't survive caching. Increment stats and
7167 * reissue to the original storage device.
7169 if (zio->io_error != 0) {
7170 ARCSTAT_BUMP(arcstat_l2_io_error);
7172 zio->io_error = SET_ERROR(EIO);
7175 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
7178 * If there's no waiter, issue an async i/o to the primary
7179 * storage now. If there *is* a waiter, the caller must
7180 * issue the i/o in a context where it's OK to block.
7182 if (zio->io_waiter == NULL) {
7183 zio_t *pio = zio_unique_parent(zio);
7185 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7187 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
7188 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
7189 hdr, zio->io_priority, cb->l2rcb_flags,
7194 kmem_free(cb, sizeof (l2arc_read_callback_t));
7198 * This is the list priority from which the L2ARC will search for pages to
7199 * cache. This is used within loops (0..3) to cycle through lists in the
7200 * desired order. This order can have a significant effect on cache
7203 * Currently the metadata lists are hit first, MFU then MRU, followed by
7204 * the data lists. This function returns a locked list, and also returns
7207 static multilist_sublist_t *
7208 l2arc_sublist_lock(int list_num)
7210 multilist_t *ml = NULL;
7213 ASSERT(list_num >= 0 && list_num <= 3);
7217 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7220 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7223 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7226 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7231 * Return a randomly-selected sublist. This is acceptable
7232 * because the caller feeds only a little bit of data for each
7233 * call (8MB). Subsequent calls will result in different
7234 * sublists being selected.
7236 idx = multilist_get_random_index(ml);
7237 return (multilist_sublist_lock(ml, idx));
7241 * Evict buffers from the device write hand to the distance specified in
7242 * bytes. This distance may span populated buffers, it may span nothing.
7243 * This is clearing a region on the L2ARC device ready for writing.
7244 * If the 'all' boolean is set, every buffer is evicted.
7247 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7250 arc_buf_hdr_t *hdr, *hdr_prev;
7251 kmutex_t *hash_lock;
7254 buflist = &dev->l2ad_buflist;
7256 if (!all && dev->l2ad_first) {
7258 * This is the first sweep through the device. There is
7264 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
7266 * When nearing the end of the device, evict to the end
7267 * before the device write hand jumps to the start.
7269 taddr = dev->l2ad_end;
7271 taddr = dev->l2ad_hand + distance;
7273 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
7274 uint64_t, taddr, boolean_t, all);
7277 mutex_enter(&dev->l2ad_mtx);
7278 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
7279 hdr_prev = list_prev(buflist, hdr);
7281 hash_lock = HDR_LOCK(hdr);
7284 * We cannot use mutex_enter or else we can deadlock
7285 * with l2arc_write_buffers (due to swapping the order
7286 * the hash lock and l2ad_mtx are taken).
7288 if (!mutex_tryenter(hash_lock)) {
7290 * Missed the hash lock. Retry.
7292 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
7293 mutex_exit(&dev->l2ad_mtx);
7294 mutex_enter(hash_lock);
7295 mutex_exit(hash_lock);
7300 * A header can't be on this list if it doesn't have L2 header.
7302 ASSERT(HDR_HAS_L2HDR(hdr));
7304 /* Ensure this header has finished being written. */
7305 ASSERT(!HDR_L2_WRITING(hdr));
7306 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
7308 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
7309 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
7311 * We've evicted to the target address,
7312 * or the end of the device.
7314 mutex_exit(hash_lock);
7318 if (!HDR_HAS_L1HDR(hdr)) {
7319 ASSERT(!HDR_L2_READING(hdr));
7321 * This doesn't exist in the ARC. Destroy.
7322 * arc_hdr_destroy() will call list_remove()
7323 * and decrement arcstat_l2_lsize.
7325 arc_change_state(arc_anon, hdr, hash_lock);
7326 arc_hdr_destroy(hdr);
7328 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
7329 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
7331 * Invalidate issued or about to be issued
7332 * reads, since we may be about to write
7333 * over this location.
7335 if (HDR_L2_READING(hdr)) {
7336 ARCSTAT_BUMP(arcstat_l2_evict_reading);
7337 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
7340 arc_hdr_l2hdr_destroy(hdr);
7342 mutex_exit(hash_lock);
7344 mutex_exit(&dev->l2ad_mtx);
7348 * Find and write ARC buffers to the L2ARC device.
7350 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7351 * for reading until they have completed writing.
7352 * The headroom_boost is an in-out parameter used to maintain headroom boost
7353 * state between calls to this function.
7355 * Returns the number of bytes actually written (which may be smaller than
7356 * the delta by which the device hand has changed due to alignment).
7359 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
7361 arc_buf_hdr_t *hdr, *hdr_prev, *head;
7362 uint64_t write_asize, write_psize, write_lsize, headroom;
7364 l2arc_write_callback_t *cb;
7366 uint64_t guid = spa_load_guid(spa);
7369 ASSERT3P(dev->l2ad_vdev, !=, NULL);
7372 write_lsize = write_asize = write_psize = 0;
7374 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
7375 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
7377 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
7379 * Copy buffers for L2ARC writing.
7381 for (try = 0; try <= 3; try++) {
7382 multilist_sublist_t *mls = l2arc_sublist_lock(try);
7383 uint64_t passed_sz = 0;
7385 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
7388 * L2ARC fast warmup.
7390 * Until the ARC is warm and starts to evict, read from the
7391 * head of the ARC lists rather than the tail.
7393 if (arc_warm == B_FALSE)
7394 hdr = multilist_sublist_head(mls);
7396 hdr = multilist_sublist_tail(mls);
7398 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
7400 headroom = target_sz * l2arc_headroom;
7401 if (zfs_compressed_arc_enabled)
7402 headroom = (headroom * l2arc_headroom_boost) / 100;
7404 for (; hdr; hdr = hdr_prev) {
7405 kmutex_t *hash_lock;
7407 if (arc_warm == B_FALSE)
7408 hdr_prev = multilist_sublist_next(mls, hdr);
7410 hdr_prev = multilist_sublist_prev(mls, hdr);
7411 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
7412 HDR_GET_LSIZE(hdr));
7414 hash_lock = HDR_LOCK(hdr);
7415 if (!mutex_tryenter(hash_lock)) {
7416 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
7418 * Skip this buffer rather than waiting.
7423 passed_sz += HDR_GET_LSIZE(hdr);
7424 if (passed_sz > headroom) {
7428 mutex_exit(hash_lock);
7429 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
7433 if (!l2arc_write_eligible(guid, hdr)) {
7434 mutex_exit(hash_lock);
7439 * We rely on the L1 portion of the header below, so
7440 * it's invalid for this header to have been evicted out
7441 * of the ghost cache, prior to being written out. The
7442 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7444 ASSERT(HDR_HAS_L1HDR(hdr));
7446 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
7447 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7448 ASSERT3U(arc_hdr_size(hdr), >, 0);
7449 uint64_t psize = arc_hdr_size(hdr);
7450 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
7453 if ((write_asize + asize) > target_sz) {
7455 mutex_exit(hash_lock);
7456 ARCSTAT_BUMP(arcstat_l2_write_full);
7462 * Insert a dummy header on the buflist so
7463 * l2arc_write_done() can find where the
7464 * write buffers begin without searching.
7466 mutex_enter(&dev->l2ad_mtx);
7467 list_insert_head(&dev->l2ad_buflist, head);
7468 mutex_exit(&dev->l2ad_mtx);
7471 sizeof (l2arc_write_callback_t), KM_SLEEP);
7472 cb->l2wcb_dev = dev;
7473 cb->l2wcb_head = head;
7474 pio = zio_root(spa, l2arc_write_done, cb,
7476 ARCSTAT_BUMP(arcstat_l2_write_pios);
7479 hdr->b_l2hdr.b_dev = dev;
7480 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7481 arc_hdr_set_flags(hdr,
7482 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7484 mutex_enter(&dev->l2ad_mtx);
7485 list_insert_head(&dev->l2ad_buflist, hdr);
7486 mutex_exit(&dev->l2ad_mtx);
7488 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
7491 * Normally the L2ARC can use the hdr's data, but if
7492 * we're sharing data between the hdr and one of its
7493 * bufs, L2ARC needs its own copy of the data so that
7494 * the ZIO below can't race with the buf consumer.
7495 * Another case where we need to create a copy of the
7496 * data is when the buffer size is not device-aligned
7497 * and we need to pad the block to make it such.
7498 * That also keeps the clock hand suitably aligned.
7500 * To ensure that the copy will be available for the
7501 * lifetime of the ZIO and be cleaned up afterwards, we
7502 * add it to the l2arc_free_on_write queue.
7505 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
7506 to_write = hdr->b_l1hdr.b_pabd;
7508 to_write = abd_alloc_for_io(asize,
7509 HDR_ISTYPE_METADATA(hdr));
7510 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
7511 if (asize != psize) {
7512 abd_zero_off(to_write, psize,
7515 l2arc_free_abd_on_write(to_write, asize,
7518 wzio = zio_write_phys(pio, dev->l2ad_vdev,
7519 hdr->b_l2hdr.b_daddr, asize, to_write,
7520 ZIO_CHECKSUM_OFF, NULL, hdr,
7521 ZIO_PRIORITY_ASYNC_WRITE,
7522 ZIO_FLAG_CANFAIL, B_FALSE);
7524 write_lsize += HDR_GET_LSIZE(hdr);
7525 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7528 write_psize += psize;
7529 write_asize += asize;
7530 dev->l2ad_hand += asize;
7532 mutex_exit(hash_lock);
7534 (void) zio_nowait(wzio);
7537 multilist_sublist_unlock(mls);
7543 /* No buffers selected for writing? */
7545 ASSERT0(write_lsize);
7546 ASSERT(!HDR_HAS_L1HDR(head));
7547 kmem_cache_free(hdr_l2only_cache, head);
7551 ASSERT3U(write_psize, <=, target_sz);
7552 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7553 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
7554 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
7555 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
7556 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
7559 * Bump device hand to the device start if it is approaching the end.
7560 * l2arc_evict() will already have evicted ahead for this case.
7562 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7563 dev->l2ad_hand = dev->l2ad_start;
7564 dev->l2ad_first = B_FALSE;
7567 dev->l2ad_writing = B_TRUE;
7568 (void) zio_wait(pio);
7569 dev->l2ad_writing = B_FALSE;
7571 return (write_asize);
7575 * This thread feeds the L2ARC at regular intervals. This is the beating
7576 * heart of the L2ARC.
7579 l2arc_feed_thread(void *dummy __unused)
7584 uint64_t size, wrote;
7585 clock_t begin, next = ddi_get_lbolt();
7587 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7589 mutex_enter(&l2arc_feed_thr_lock);
7591 while (l2arc_thread_exit == 0) {
7592 CALLB_CPR_SAFE_BEGIN(&cpr);
7593 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7594 next - ddi_get_lbolt());
7595 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7596 next = ddi_get_lbolt() + hz;
7599 * Quick check for L2ARC devices.
7601 mutex_enter(&l2arc_dev_mtx);
7602 if (l2arc_ndev == 0) {
7603 mutex_exit(&l2arc_dev_mtx);
7606 mutex_exit(&l2arc_dev_mtx);
7607 begin = ddi_get_lbolt();
7610 * This selects the next l2arc device to write to, and in
7611 * doing so the next spa to feed from: dev->l2ad_spa. This
7612 * will return NULL if there are now no l2arc devices or if
7613 * they are all faulted.
7615 * If a device is returned, its spa's config lock is also
7616 * held to prevent device removal. l2arc_dev_get_next()
7617 * will grab and release l2arc_dev_mtx.
7619 if ((dev = l2arc_dev_get_next()) == NULL)
7622 spa = dev->l2ad_spa;
7623 ASSERT3P(spa, !=, NULL);
7626 * If the pool is read-only then force the feed thread to
7627 * sleep a little longer.
7629 if (!spa_writeable(spa)) {
7630 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7631 spa_config_exit(spa, SCL_L2ARC, dev);
7636 * Avoid contributing to memory pressure.
7638 if (arc_reclaim_needed()) {
7639 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7640 spa_config_exit(spa, SCL_L2ARC, dev);
7644 ARCSTAT_BUMP(arcstat_l2_feeds);
7646 size = l2arc_write_size();
7649 * Evict L2ARC buffers that will be overwritten.
7651 l2arc_evict(dev, size, B_FALSE);
7654 * Write ARC buffers.
7656 wrote = l2arc_write_buffers(spa, dev, size);
7659 * Calculate interval between writes.
7661 next = l2arc_write_interval(begin, size, wrote);
7662 spa_config_exit(spa, SCL_L2ARC, dev);
7665 l2arc_thread_exit = 0;
7666 cv_broadcast(&l2arc_feed_thr_cv);
7667 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7672 l2arc_vdev_present(vdev_t *vd)
7676 mutex_enter(&l2arc_dev_mtx);
7677 for (dev = list_head(l2arc_dev_list); dev != NULL;
7678 dev = list_next(l2arc_dev_list, dev)) {
7679 if (dev->l2ad_vdev == vd)
7682 mutex_exit(&l2arc_dev_mtx);
7684 return (dev != NULL);
7688 * Add a vdev for use by the L2ARC. By this point the spa has already
7689 * validated the vdev and opened it.
7692 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7694 l2arc_dev_t *adddev;
7696 ASSERT(!l2arc_vdev_present(vd));
7698 vdev_ashift_optimize(vd);
7701 * Create a new l2arc device entry.
7703 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7704 adddev->l2ad_spa = spa;
7705 adddev->l2ad_vdev = vd;
7706 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7707 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7708 adddev->l2ad_hand = adddev->l2ad_start;
7709 adddev->l2ad_first = B_TRUE;
7710 adddev->l2ad_writing = B_FALSE;
7712 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7714 * This is a list of all ARC buffers that are still valid on the
7717 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7718 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7720 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7721 refcount_create(&adddev->l2ad_alloc);
7724 * Add device to global list
7726 mutex_enter(&l2arc_dev_mtx);
7727 list_insert_head(l2arc_dev_list, adddev);
7728 atomic_inc_64(&l2arc_ndev);
7729 mutex_exit(&l2arc_dev_mtx);
7733 * Remove a vdev from the L2ARC.
7736 l2arc_remove_vdev(vdev_t *vd)
7738 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7741 * Find the device by vdev
7743 mutex_enter(&l2arc_dev_mtx);
7744 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7745 nextdev = list_next(l2arc_dev_list, dev);
7746 if (vd == dev->l2ad_vdev) {
7751 ASSERT3P(remdev, !=, NULL);
7754 * Remove device from global list
7756 list_remove(l2arc_dev_list, remdev);
7757 l2arc_dev_last = NULL; /* may have been invalidated */
7758 atomic_dec_64(&l2arc_ndev);
7759 mutex_exit(&l2arc_dev_mtx);
7762 * Clear all buflists and ARC references. L2ARC device flush.
7764 l2arc_evict(remdev, 0, B_TRUE);
7765 list_destroy(&remdev->l2ad_buflist);
7766 mutex_destroy(&remdev->l2ad_mtx);
7767 refcount_destroy(&remdev->l2ad_alloc);
7768 kmem_free(remdev, sizeof (l2arc_dev_t));
7774 l2arc_thread_exit = 0;
7776 l2arc_writes_sent = 0;
7777 l2arc_writes_done = 0;
7779 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7780 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7781 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7782 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7784 l2arc_dev_list = &L2ARC_dev_list;
7785 l2arc_free_on_write = &L2ARC_free_on_write;
7786 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7787 offsetof(l2arc_dev_t, l2ad_node));
7788 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7789 offsetof(l2arc_data_free_t, l2df_list_node));
7796 * This is called from dmu_fini(), which is called from spa_fini();
7797 * Because of this, we can assume that all l2arc devices have
7798 * already been removed when the pools themselves were removed.
7801 l2arc_do_free_on_write();
7803 mutex_destroy(&l2arc_feed_thr_lock);
7804 cv_destroy(&l2arc_feed_thr_cv);
7805 mutex_destroy(&l2arc_dev_mtx);
7806 mutex_destroy(&l2arc_free_on_write_mtx);
7808 list_destroy(l2arc_dev_list);
7809 list_destroy(l2arc_free_on_write);
7815 if (!(spa_mode_global & FWRITE))
7818 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7819 TS_RUN, minclsyspri);
7825 if (!(spa_mode_global & FWRITE))
7828 mutex_enter(&l2arc_feed_thr_lock);
7829 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7830 l2arc_thread_exit = 1;
7831 while (l2arc_thread_exit != 0)
7832 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7833 mutex_exit(&l2arc_feed_thr_lock);