4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2018, Joyent, Inc.
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2017 Nexenta Systems, Inc. All rights reserved.
30 * DVA-based Adjustable Replacement Cache
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal ARC algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
96 * Each ARC state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an ARC list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
102 * Note that the majority of the performance stats are manipulated
103 * with atomic operations.
105 * The L2ARC uses the l2ad_mtx on each vdev for the following:
107 * - L2ARC buflist creation
108 * - L2ARC buflist eviction
109 * - L2ARC write completion, which walks L2ARC buflists
110 * - ARC header destruction, as it removes from L2ARC buflists
111 * - ARC header release, as it removes from L2ARC buflists
117 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
118 * This structure can point either to a block that is still in the cache or to
119 * one that is only accessible in an L2 ARC device, or it can provide
120 * information about a block that was recently evicted. If a block is
121 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
122 * information to retrieve it from the L2ARC device. This information is
123 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
124 * that is in this state cannot access the data directly.
126 * Blocks that are actively being referenced or have not been evicted
127 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
128 * the arc_buf_hdr_t that will point to the data block in memory. A block can
129 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
130 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
131 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
133 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
134 * ability to store the physical data (b_pabd) associated with the DVA of the
135 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
136 * it will match its on-disk compression characteristics. This behavior can be
137 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
138 * compressed ARC functionality is disabled, the b_pabd will point to an
139 * uncompressed version of the on-disk data.
141 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
142 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
143 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
144 * consumer. The ARC will provide references to this data and will keep it
145 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
146 * data block and will evict any arc_buf_t that is no longer referenced. The
147 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
148 * "overhead_size" kstat.
150 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
151 * compressed form. The typical case is that consumers will want uncompressed
152 * data, and when that happens a new data buffer is allocated where the data is
153 * decompressed for them to use. Currently the only consumer who wants
154 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
155 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
156 * with the arc_buf_hdr_t.
158 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
159 * first one is owned by a compressed send consumer (and therefore references
160 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
161 * used by any other consumer (and has its own uncompressed copy of the data
176 * | b_buf +------------>+-----------+ arc_buf_t
177 * | b_pabd +-+ |b_next +---->+-----------+
178 * +-----------+ | |-----------| |b_next +-->NULL
179 * | |b_comp = T | +-----------+
180 * | |b_data +-+ |b_comp = F |
181 * | +-----------+ | |b_data +-+
182 * +->+------+ | +-----------+ |
184 * data | |<--------------+ | uncompressed
185 * +------+ compressed, | data
186 * shared +-->+------+
191 * When a consumer reads a block, the ARC must first look to see if the
192 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
193 * arc_buf_t and either copies uncompressed data into a new data buffer from an
194 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
195 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
196 * hdr is compressed and the desired compression characteristics of the
197 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
198 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
199 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
200 * be anywhere in the hdr's list.
202 * The diagram below shows an example of an uncompressed ARC hdr that is
203 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
204 * the last element in the buf list):
216 * | | arc_buf_t (shared)
217 * | b_buf +------------>+---------+ arc_buf_t
218 * | | |b_next +---->+---------+
219 * | b_pabd +-+ |---------| |b_next +-->NULL
220 * +-----------+ | | | +---------+
222 * | +---------+ | |b_data +-+
223 * +->+------+ | +---------+ |
225 * uncompressed | | | |
228 * | uncompressed | | |
231 * +---------------------------------+
233 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
234 * since the physical block is about to be rewritten. The new data contents
235 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
236 * it may compress the data before writing it to disk. The ARC will be called
237 * with the transformed data and will bcopy the transformed on-disk block into
238 * a newly allocated b_pabd. Writes are always done into buffers which have
239 * either been loaned (and hence are new and don't have other readers) or
240 * buffers which have been released (and hence have their own hdr, if there
241 * were originally other readers of the buf's original hdr). This ensures that
242 * the ARC only needs to update a single buf and its hdr after a write occurs.
244 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
245 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
246 * that when compressed ARC is enabled that the L2ARC blocks are identical
247 * to the on-disk block in the main data pool. This provides a significant
248 * advantage since the ARC can leverage the bp's checksum when reading from the
249 * L2ARC to determine if the contents are valid. However, if the compressed
250 * ARC is disabled, then the L2ARC's block must be transformed to look
251 * like the physical block in the main data pool before comparing the
252 * checksum and determining its validity.
257 #include <sys/spa_impl.h>
258 #include <sys/zio_compress.h>
259 #include <sys/zio_checksum.h>
260 #include <sys/zfs_context.h>
262 #include <sys/refcount.h>
263 #include <sys/vdev.h>
264 #include <sys/vdev_impl.h>
265 #include <sys/dsl_pool.h>
266 #include <sys/zio_checksum.h>
267 #include <sys/multilist.h>
270 #include <sys/dnlc.h>
271 #include <sys/racct.h>
273 #include <sys/callb.h>
274 #include <sys/kstat.h>
275 #include <sys/trim_map.h>
276 #include <zfs_fletcher.h>
278 #include <sys/aggsum.h>
279 #include <sys/cityhash.h>
281 #include <machine/vmparam.h>
285 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
286 boolean_t arc_watch = B_FALSE;
291 static kmutex_t arc_reclaim_lock;
292 static kcondvar_t arc_reclaim_thread_cv;
293 static boolean_t arc_reclaim_thread_exit;
294 static kcondvar_t arc_reclaim_waiters_cv;
296 static kmutex_t arc_dnlc_evicts_lock;
297 static kcondvar_t arc_dnlc_evicts_cv;
298 static boolean_t arc_dnlc_evicts_thread_exit;
300 uint_t arc_reduce_dnlc_percent = 3;
303 * The number of headers to evict in arc_evict_state_impl() before
304 * dropping the sublist lock and evicting from another sublist. A lower
305 * value means we're more likely to evict the "correct" header (i.e. the
306 * oldest header in the arc state), but comes with higher overhead
307 * (i.e. more invocations of arc_evict_state_impl()).
309 int zfs_arc_evict_batch_limit = 10;
311 /* number of seconds before growing cache again */
312 static int arc_grow_retry = 60;
314 /* number of milliseconds before attempting a kmem-cache-reap */
315 static int arc_kmem_cache_reap_retry_ms = 0;
317 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
318 int zfs_arc_overflow_shift = 8;
320 /* shift of arc_c for calculating both min and max arc_p */
321 static int arc_p_min_shift = 4;
323 /* log2(fraction of arc to reclaim) */
324 static int arc_shrink_shift = 7;
327 * log2(fraction of ARC which must be free to allow growing).
328 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
329 * when reading a new block into the ARC, we will evict an equal-sized block
332 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
333 * we will still not allow it to grow.
335 int arc_no_grow_shift = 5;
339 * minimum lifespan of a prefetch block in clock ticks
340 * (initialized in arc_init())
342 static int zfs_arc_min_prefetch_ms = 1;
343 static int zfs_arc_min_prescient_prefetch_ms = 6;
346 * If this percent of memory is free, don't throttle.
348 int arc_lotsfree_percent = 10;
351 extern boolean_t zfs_prefetch_disable;
354 * The arc has filled available memory and has now warmed up.
356 static boolean_t arc_warm;
359 * log2 fraction of the zio arena to keep free.
361 int arc_zio_arena_free_shift = 2;
364 * These tunables are for performance analysis.
366 uint64_t zfs_arc_max;
367 uint64_t zfs_arc_min;
368 uint64_t zfs_arc_meta_limit = 0;
369 uint64_t zfs_arc_meta_min = 0;
370 int zfs_arc_grow_retry = 0;
371 int zfs_arc_shrink_shift = 0;
372 int zfs_arc_no_grow_shift = 0;
373 int zfs_arc_p_min_shift = 0;
374 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
375 u_int zfs_arc_free_target = 0;
377 /* Absolute min for arc min / max is 16MB. */
378 static uint64_t arc_abs_min = 16 << 20;
380 boolean_t zfs_compressed_arc_enabled = B_TRUE;
382 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
383 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
384 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
385 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
386 static int sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS);
388 #if defined(__FreeBSD__) && defined(_KERNEL)
390 arc_free_target_init(void *unused __unused)
393 zfs_arc_free_target = vm_pageout_wakeup_thresh;
395 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
396 arc_free_target_init, NULL);
398 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
399 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
400 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
401 TUNABLE_INT("vfs.zfs.arc_grow_retry", &zfs_arc_grow_retry);
402 TUNABLE_INT("vfs.zfs.arc_no_grow_shift", &zfs_arc_no_grow_shift);
403 SYSCTL_DECL(_vfs_zfs);
404 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
405 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
406 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
407 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
408 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_no_grow_shift, CTLTYPE_U32 | CTLFLAG_RWTUN,
409 0, sizeof(uint32_t), sysctl_vfs_zfs_arc_no_grow_shift, "U",
410 "log2(fraction of ARC which must be free to allow growing)");
411 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
412 &zfs_arc_average_blocksize, 0,
413 "ARC average blocksize");
414 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
415 &arc_shrink_shift, 0,
416 "log2(fraction of arc to reclaim)");
417 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_grow_retry, CTLFLAG_RW,
419 "Wait in seconds before considering growing ARC");
420 SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN,
421 &zfs_compressed_arc_enabled, 0,
422 "Enable compressed ARC");
423 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_kmem_cache_reap_retry_ms, CTLFLAG_RWTUN,
424 &arc_kmem_cache_reap_retry_ms, 0,
425 "Interval between ARC kmem_cache reapings");
428 * We don't have a tunable for arc_free_target due to the dependency on
429 * pagedaemon initialisation.
431 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
432 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
433 sysctl_vfs_zfs_arc_free_target, "IU",
434 "Desired number of free pages below which ARC triggers reclaim");
437 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
442 val = zfs_arc_free_target;
443 err = sysctl_handle_int(oidp, &val, 0, req);
444 if (err != 0 || req->newptr == NULL)
449 if (val > vm_cnt.v_page_count)
452 zfs_arc_free_target = val;
458 * Must be declared here, before the definition of corresponding kstat
459 * macro which uses the same names will confuse the compiler.
461 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
462 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
463 sysctl_vfs_zfs_arc_meta_limit, "QU",
464 "ARC metadata limit");
468 * Note that buffers can be in one of 6 states:
469 * ARC_anon - anonymous (discussed below)
470 * ARC_mru - recently used, currently cached
471 * ARC_mru_ghost - recentely used, no longer in cache
472 * ARC_mfu - frequently used, currently cached
473 * ARC_mfu_ghost - frequently used, no longer in cache
474 * ARC_l2c_only - exists in L2ARC but not other states
475 * When there are no active references to the buffer, they are
476 * are linked onto a list in one of these arc states. These are
477 * the only buffers that can be evicted or deleted. Within each
478 * state there are multiple lists, one for meta-data and one for
479 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
480 * etc.) is tracked separately so that it can be managed more
481 * explicitly: favored over data, limited explicitly.
483 * Anonymous buffers are buffers that are not associated with
484 * a DVA. These are buffers that hold dirty block copies
485 * before they are written to stable storage. By definition,
486 * they are "ref'd" and are considered part of arc_mru
487 * that cannot be freed. Generally, they will aquire a DVA
488 * as they are written and migrate onto the arc_mru list.
490 * The ARC_l2c_only state is for buffers that are in the second
491 * level ARC but no longer in any of the ARC_m* lists. The second
492 * level ARC itself may also contain buffers that are in any of
493 * the ARC_m* states - meaning that a buffer can exist in two
494 * places. The reason for the ARC_l2c_only state is to keep the
495 * buffer header in the hash table, so that reads that hit the
496 * second level ARC benefit from these fast lookups.
499 typedef struct arc_state {
501 * list of evictable buffers
503 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
505 * total amount of evictable data in this state
507 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
509 * total amount of data in this state; this includes: evictable,
510 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
512 refcount_t arcs_size;
516 static arc_state_t ARC_anon;
517 static arc_state_t ARC_mru;
518 static arc_state_t ARC_mru_ghost;
519 static arc_state_t ARC_mfu;
520 static arc_state_t ARC_mfu_ghost;
521 static arc_state_t ARC_l2c_only;
523 typedef struct arc_stats {
524 kstat_named_t arcstat_hits;
525 kstat_named_t arcstat_misses;
526 kstat_named_t arcstat_demand_data_hits;
527 kstat_named_t arcstat_demand_data_misses;
528 kstat_named_t arcstat_demand_metadata_hits;
529 kstat_named_t arcstat_demand_metadata_misses;
530 kstat_named_t arcstat_prefetch_data_hits;
531 kstat_named_t arcstat_prefetch_data_misses;
532 kstat_named_t arcstat_prefetch_metadata_hits;
533 kstat_named_t arcstat_prefetch_metadata_misses;
534 kstat_named_t arcstat_mru_hits;
535 kstat_named_t arcstat_mru_ghost_hits;
536 kstat_named_t arcstat_mfu_hits;
537 kstat_named_t arcstat_mfu_ghost_hits;
538 kstat_named_t arcstat_allocated;
539 kstat_named_t arcstat_deleted;
541 * Number of buffers that could not be evicted because the hash lock
542 * was held by another thread. The lock may not necessarily be held
543 * by something using the same buffer, since hash locks are shared
544 * by multiple buffers.
546 kstat_named_t arcstat_mutex_miss;
548 * Number of buffers skipped when updating the access state due to the
549 * header having already been released after acquiring the hash lock.
551 kstat_named_t arcstat_access_skip;
553 * Number of buffers skipped because they have I/O in progress, are
554 * indirect prefetch buffers that have not lived long enough, or are
555 * not from the spa we're trying to evict from.
557 kstat_named_t arcstat_evict_skip;
559 * Number of times arc_evict_state() was unable to evict enough
560 * buffers to reach it's target amount.
562 kstat_named_t arcstat_evict_not_enough;
563 kstat_named_t arcstat_evict_l2_cached;
564 kstat_named_t arcstat_evict_l2_eligible;
565 kstat_named_t arcstat_evict_l2_ineligible;
566 kstat_named_t arcstat_evict_l2_skip;
567 kstat_named_t arcstat_hash_elements;
568 kstat_named_t arcstat_hash_elements_max;
569 kstat_named_t arcstat_hash_collisions;
570 kstat_named_t arcstat_hash_chains;
571 kstat_named_t arcstat_hash_chain_max;
572 kstat_named_t arcstat_p;
573 kstat_named_t arcstat_c;
574 kstat_named_t arcstat_c_min;
575 kstat_named_t arcstat_c_max;
576 /* Not updated directly; only synced in arc_kstat_update. */
577 kstat_named_t arcstat_size;
579 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
580 * Note that the compressed bytes may match the uncompressed bytes
581 * if the block is either not compressed or compressed arc is disabled.
583 kstat_named_t arcstat_compressed_size;
585 * Uncompressed size of the data stored in b_pabd. If compressed
586 * arc is disabled then this value will be identical to the stat
589 kstat_named_t arcstat_uncompressed_size;
591 * Number of bytes stored in all the arc_buf_t's. This is classified
592 * as "overhead" since this data is typically short-lived and will
593 * be evicted from the arc when it becomes unreferenced unless the
594 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
595 * values have been set (see comment in dbuf.c for more information).
597 kstat_named_t arcstat_overhead_size;
599 * Number of bytes consumed by internal ARC structures necessary
600 * for tracking purposes; these structures are not actually
601 * backed by ARC buffers. This includes arc_buf_hdr_t structures
602 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
603 * caches), and arc_buf_t structures (allocated via arc_buf_t
605 * Not updated directly; only synced in arc_kstat_update.
607 kstat_named_t arcstat_hdr_size;
609 * Number of bytes consumed by ARC buffers of type equal to
610 * ARC_BUFC_DATA. This is generally consumed by buffers backing
611 * on disk user data (e.g. plain file contents).
612 * Not updated directly; only synced in arc_kstat_update.
614 kstat_named_t arcstat_data_size;
616 * Number of bytes consumed by ARC buffers of type equal to
617 * ARC_BUFC_METADATA. This is generally consumed by buffers
618 * backing on disk data that is used for internal ZFS
619 * structures (e.g. ZAP, dnode, indirect blocks, etc).
620 * Not updated directly; only synced in arc_kstat_update.
622 kstat_named_t arcstat_metadata_size;
624 * Number of bytes consumed by various buffers and structures
625 * not actually backed with ARC buffers. This includes bonus
626 * buffers (allocated directly via zio_buf_* functions),
627 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
628 * cache), and dnode_t structures (allocated via dnode_t cache).
629 * Not updated directly; only synced in arc_kstat_update.
631 kstat_named_t arcstat_other_size;
633 * Total number of bytes consumed by ARC buffers residing in the
634 * arc_anon state. This includes *all* buffers in the arc_anon
635 * state; e.g. data, metadata, evictable, and unevictable buffers
636 * are all included in this value.
637 * Not updated directly; only synced in arc_kstat_update.
639 kstat_named_t arcstat_anon_size;
641 * Number of bytes consumed by ARC buffers that meet the
642 * following criteria: backing buffers of type ARC_BUFC_DATA,
643 * residing in the arc_anon state, and are eligible for eviction
644 * (e.g. have no outstanding holds on the buffer).
645 * Not updated directly; only synced in arc_kstat_update.
647 kstat_named_t arcstat_anon_evictable_data;
649 * Number of bytes consumed by ARC buffers that meet the
650 * following criteria: backing buffers of type ARC_BUFC_METADATA,
651 * residing in the arc_anon state, and are eligible for eviction
652 * (e.g. have no outstanding holds on the buffer).
653 * Not updated directly; only synced in arc_kstat_update.
655 kstat_named_t arcstat_anon_evictable_metadata;
657 * Total number of bytes consumed by ARC buffers residing in the
658 * arc_mru state. This includes *all* buffers in the arc_mru
659 * state; e.g. data, metadata, evictable, and unevictable buffers
660 * are all included in this value.
661 * Not updated directly; only synced in arc_kstat_update.
663 kstat_named_t arcstat_mru_size;
665 * Number of bytes consumed by ARC buffers that meet the
666 * following criteria: backing buffers of type ARC_BUFC_DATA,
667 * residing in the arc_mru state, and are eligible for eviction
668 * (e.g. have no outstanding holds on the buffer).
669 * Not updated directly; only synced in arc_kstat_update.
671 kstat_named_t arcstat_mru_evictable_data;
673 * Number of bytes consumed by ARC buffers that meet the
674 * following criteria: backing buffers of type ARC_BUFC_METADATA,
675 * residing in the arc_mru state, and are eligible for eviction
676 * (e.g. have no outstanding holds on the buffer).
677 * Not updated directly; only synced in arc_kstat_update.
679 kstat_named_t arcstat_mru_evictable_metadata;
681 * Total number of bytes that *would have been* consumed by ARC
682 * buffers in the arc_mru_ghost state. The key thing to note
683 * here, is the fact that this size doesn't actually indicate
684 * RAM consumption. The ghost lists only consist of headers and
685 * don't actually have ARC buffers linked off of these headers.
686 * Thus, *if* the headers had associated ARC buffers, these
687 * buffers *would have* consumed this number of bytes.
688 * Not updated directly; only synced in arc_kstat_update.
690 kstat_named_t arcstat_mru_ghost_size;
692 * Number of bytes that *would have been* consumed by ARC
693 * buffers that are eligible for eviction, of type
694 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
695 * Not updated directly; only synced in arc_kstat_update.
697 kstat_named_t arcstat_mru_ghost_evictable_data;
699 * Number of bytes that *would have been* consumed by ARC
700 * buffers that are eligible for eviction, of type
701 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
702 * Not updated directly; only synced in arc_kstat_update.
704 kstat_named_t arcstat_mru_ghost_evictable_metadata;
706 * Total number of bytes consumed by ARC buffers residing in the
707 * arc_mfu state. This includes *all* buffers in the arc_mfu
708 * state; e.g. data, metadata, evictable, and unevictable buffers
709 * are all included in this value.
710 * Not updated directly; only synced in arc_kstat_update.
712 kstat_named_t arcstat_mfu_size;
714 * Number of bytes consumed by ARC buffers that are eligible for
715 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
717 * Not updated directly; only synced in arc_kstat_update.
719 kstat_named_t arcstat_mfu_evictable_data;
721 * Number of bytes consumed by ARC buffers that are eligible for
722 * eviction, of type ARC_BUFC_METADATA, and reside in the
724 * Not updated directly; only synced in arc_kstat_update.
726 kstat_named_t arcstat_mfu_evictable_metadata;
728 * Total number of bytes that *would have been* consumed by ARC
729 * buffers in the arc_mfu_ghost state. See the comment above
730 * arcstat_mru_ghost_size for more details.
731 * Not updated directly; only synced in arc_kstat_update.
733 kstat_named_t arcstat_mfu_ghost_size;
735 * Number of bytes that *would have been* consumed by ARC
736 * buffers that are eligible for eviction, of type
737 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
738 * Not updated directly; only synced in arc_kstat_update.
740 kstat_named_t arcstat_mfu_ghost_evictable_data;
742 * Number of bytes that *would have been* consumed by ARC
743 * buffers that are eligible for eviction, of type
744 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
745 * Not updated directly; only synced in arc_kstat_update.
747 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
748 kstat_named_t arcstat_l2_hits;
749 kstat_named_t arcstat_l2_misses;
750 kstat_named_t arcstat_l2_feeds;
751 kstat_named_t arcstat_l2_rw_clash;
752 kstat_named_t arcstat_l2_read_bytes;
753 kstat_named_t arcstat_l2_write_bytes;
754 kstat_named_t arcstat_l2_writes_sent;
755 kstat_named_t arcstat_l2_writes_done;
756 kstat_named_t arcstat_l2_writes_error;
757 kstat_named_t arcstat_l2_writes_lock_retry;
758 kstat_named_t arcstat_l2_evict_lock_retry;
759 kstat_named_t arcstat_l2_evict_reading;
760 kstat_named_t arcstat_l2_evict_l1cached;
761 kstat_named_t arcstat_l2_free_on_write;
762 kstat_named_t arcstat_l2_abort_lowmem;
763 kstat_named_t arcstat_l2_cksum_bad;
764 kstat_named_t arcstat_l2_io_error;
765 kstat_named_t arcstat_l2_lsize;
766 kstat_named_t arcstat_l2_psize;
767 /* Not updated directly; only synced in arc_kstat_update. */
768 kstat_named_t arcstat_l2_hdr_size;
769 kstat_named_t arcstat_l2_write_trylock_fail;
770 kstat_named_t arcstat_l2_write_passed_headroom;
771 kstat_named_t arcstat_l2_write_spa_mismatch;
772 kstat_named_t arcstat_l2_write_in_l2;
773 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
774 kstat_named_t arcstat_l2_write_not_cacheable;
775 kstat_named_t arcstat_l2_write_full;
776 kstat_named_t arcstat_l2_write_buffer_iter;
777 kstat_named_t arcstat_l2_write_pios;
778 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
779 kstat_named_t arcstat_l2_write_buffer_list_iter;
780 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
781 kstat_named_t arcstat_memory_throttle_count;
782 /* Not updated directly; only synced in arc_kstat_update. */
783 kstat_named_t arcstat_meta_used;
784 kstat_named_t arcstat_meta_limit;
785 kstat_named_t arcstat_meta_max;
786 kstat_named_t arcstat_meta_min;
787 kstat_named_t arcstat_async_upgrade_sync;
788 kstat_named_t arcstat_demand_hit_predictive_prefetch;
789 kstat_named_t arcstat_demand_hit_prescient_prefetch;
792 static arc_stats_t arc_stats = {
793 { "hits", KSTAT_DATA_UINT64 },
794 { "misses", KSTAT_DATA_UINT64 },
795 { "demand_data_hits", KSTAT_DATA_UINT64 },
796 { "demand_data_misses", KSTAT_DATA_UINT64 },
797 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
798 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
799 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
800 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
801 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
802 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
803 { "mru_hits", KSTAT_DATA_UINT64 },
804 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
805 { "mfu_hits", KSTAT_DATA_UINT64 },
806 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
807 { "allocated", KSTAT_DATA_UINT64 },
808 { "deleted", KSTAT_DATA_UINT64 },
809 { "mutex_miss", KSTAT_DATA_UINT64 },
810 { "access_skip", KSTAT_DATA_UINT64 },
811 { "evict_skip", KSTAT_DATA_UINT64 },
812 { "evict_not_enough", KSTAT_DATA_UINT64 },
813 { "evict_l2_cached", KSTAT_DATA_UINT64 },
814 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
815 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
816 { "evict_l2_skip", KSTAT_DATA_UINT64 },
817 { "hash_elements", KSTAT_DATA_UINT64 },
818 { "hash_elements_max", KSTAT_DATA_UINT64 },
819 { "hash_collisions", KSTAT_DATA_UINT64 },
820 { "hash_chains", KSTAT_DATA_UINT64 },
821 { "hash_chain_max", KSTAT_DATA_UINT64 },
822 { "p", KSTAT_DATA_UINT64 },
823 { "c", KSTAT_DATA_UINT64 },
824 { "c_min", KSTAT_DATA_UINT64 },
825 { "c_max", KSTAT_DATA_UINT64 },
826 { "size", KSTAT_DATA_UINT64 },
827 { "compressed_size", KSTAT_DATA_UINT64 },
828 { "uncompressed_size", KSTAT_DATA_UINT64 },
829 { "overhead_size", KSTAT_DATA_UINT64 },
830 { "hdr_size", KSTAT_DATA_UINT64 },
831 { "data_size", KSTAT_DATA_UINT64 },
832 { "metadata_size", KSTAT_DATA_UINT64 },
833 { "other_size", KSTAT_DATA_UINT64 },
834 { "anon_size", KSTAT_DATA_UINT64 },
835 { "anon_evictable_data", KSTAT_DATA_UINT64 },
836 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
837 { "mru_size", KSTAT_DATA_UINT64 },
838 { "mru_evictable_data", KSTAT_DATA_UINT64 },
839 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
840 { "mru_ghost_size", KSTAT_DATA_UINT64 },
841 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
842 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
843 { "mfu_size", KSTAT_DATA_UINT64 },
844 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
845 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
846 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
847 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
848 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
849 { "l2_hits", KSTAT_DATA_UINT64 },
850 { "l2_misses", KSTAT_DATA_UINT64 },
851 { "l2_feeds", KSTAT_DATA_UINT64 },
852 { "l2_rw_clash", KSTAT_DATA_UINT64 },
853 { "l2_read_bytes", KSTAT_DATA_UINT64 },
854 { "l2_write_bytes", KSTAT_DATA_UINT64 },
855 { "l2_writes_sent", KSTAT_DATA_UINT64 },
856 { "l2_writes_done", KSTAT_DATA_UINT64 },
857 { "l2_writes_error", KSTAT_DATA_UINT64 },
858 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
859 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
860 { "l2_evict_reading", KSTAT_DATA_UINT64 },
861 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
862 { "l2_free_on_write", KSTAT_DATA_UINT64 },
863 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
864 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
865 { "l2_io_error", KSTAT_DATA_UINT64 },
866 { "l2_size", KSTAT_DATA_UINT64 },
867 { "l2_asize", KSTAT_DATA_UINT64 },
868 { "l2_hdr_size", KSTAT_DATA_UINT64 },
869 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
870 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
871 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
872 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
873 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
874 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
875 { "l2_write_full", KSTAT_DATA_UINT64 },
876 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
877 { "l2_write_pios", KSTAT_DATA_UINT64 },
878 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
879 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
880 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
881 { "memory_throttle_count", KSTAT_DATA_UINT64 },
882 { "arc_meta_used", KSTAT_DATA_UINT64 },
883 { "arc_meta_limit", KSTAT_DATA_UINT64 },
884 { "arc_meta_max", KSTAT_DATA_UINT64 },
885 { "arc_meta_min", KSTAT_DATA_UINT64 },
886 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
887 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
888 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
891 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
893 #define ARCSTAT_INCR(stat, val) \
894 atomic_add_64(&arc_stats.stat.value.ui64, (val))
896 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
897 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
899 #define ARCSTAT_MAX(stat, val) { \
901 while ((val) > (m = arc_stats.stat.value.ui64) && \
902 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
906 #define ARCSTAT_MAXSTAT(stat) \
907 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
910 * We define a macro to allow ARC hits/misses to be easily broken down by
911 * two separate conditions, giving a total of four different subtypes for
912 * each of hits and misses (so eight statistics total).
914 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
917 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
919 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
923 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
925 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
930 static arc_state_t *arc_anon;
931 static arc_state_t *arc_mru;
932 static arc_state_t *arc_mru_ghost;
933 static arc_state_t *arc_mfu;
934 static arc_state_t *arc_mfu_ghost;
935 static arc_state_t *arc_l2c_only;
938 * There are several ARC variables that are critical to export as kstats --
939 * but we don't want to have to grovel around in the kstat whenever we wish to
940 * manipulate them. For these variables, we therefore define them to be in
941 * terms of the statistic variable. This assures that we are not introducing
942 * the possibility of inconsistency by having shadow copies of the variables,
943 * while still allowing the code to be readable.
945 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
946 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
947 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
948 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
949 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
950 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
951 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
953 /* compressed size of entire arc */
954 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
955 /* uncompressed size of entire arc */
956 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
957 /* number of bytes in the arc from arc_buf_t's */
958 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
961 * There are also some ARC variables that we want to export, but that are
962 * updated so often that having the canonical representation be the statistic
963 * variable causes a performance bottleneck. We want to use aggsum_t's for these
964 * instead, but still be able to export the kstat in the same way as before.
965 * The solution is to always use the aggsum version, except in the kstat update
969 aggsum_t arc_meta_used;
970 aggsum_t astat_data_size;
971 aggsum_t astat_metadata_size;
972 aggsum_t astat_hdr_size;
973 aggsum_t astat_other_size;
974 aggsum_t astat_l2_hdr_size;
976 static int arc_no_grow; /* Don't try to grow cache size */
977 static uint64_t arc_tempreserve;
978 static uint64_t arc_loaned_bytes;
980 typedef struct arc_callback arc_callback_t;
982 struct arc_callback {
984 arc_read_done_func_t *acb_done;
986 boolean_t acb_compressed;
987 zio_t *acb_zio_dummy;
989 arc_callback_t *acb_next;
992 typedef struct arc_write_callback arc_write_callback_t;
994 struct arc_write_callback {
996 arc_write_done_func_t *awcb_ready;
997 arc_write_done_func_t *awcb_children_ready;
998 arc_write_done_func_t *awcb_physdone;
999 arc_write_done_func_t *awcb_done;
1000 arc_buf_t *awcb_buf;
1004 * ARC buffers are separated into multiple structs as a memory saving measure:
1005 * - Common fields struct, always defined, and embedded within it:
1006 * - L2-only fields, always allocated but undefined when not in L2ARC
1007 * - L1-only fields, only allocated when in L1ARC
1009 * Buffer in L1 Buffer only in L2
1010 * +------------------------+ +------------------------+
1011 * | arc_buf_hdr_t | | arc_buf_hdr_t |
1015 * +------------------------+ +------------------------+
1016 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
1017 * | (undefined if L1-only) | | |
1018 * +------------------------+ +------------------------+
1019 * | l1arc_buf_hdr_t |
1024 * +------------------------+
1026 * Because it's possible for the L2ARC to become extremely large, we can wind
1027 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
1028 * is minimized by only allocating the fields necessary for an L1-cached buffer
1029 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
1030 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
1031 * words in pointers. arc_hdr_realloc() is used to switch a header between
1032 * these two allocation states.
1034 typedef struct l1arc_buf_hdr {
1035 kmutex_t b_freeze_lock;
1036 zio_cksum_t *b_freeze_cksum;
1039 * Used for debugging with kmem_flags - by allocating and freeing
1040 * b_thawed when the buffer is thawed, we get a record of the stack
1041 * trace that thawed it.
1048 /* for waiting on writes to complete */
1052 /* protected by arc state mutex */
1053 arc_state_t *b_state;
1054 multilist_node_t b_arc_node;
1056 /* updated atomically */
1057 clock_t b_arc_access;
1059 /* self protecting */
1060 refcount_t b_refcnt;
1062 arc_callback_t *b_acb;
1066 typedef struct l2arc_dev l2arc_dev_t;
1068 typedef struct l2arc_buf_hdr {
1069 /* protected by arc_buf_hdr mutex */
1070 l2arc_dev_t *b_dev; /* L2ARC device */
1071 uint64_t b_daddr; /* disk address, offset byte */
1073 list_node_t b_l2node;
1076 struct arc_buf_hdr {
1077 /* protected by hash lock */
1081 arc_buf_contents_t b_type;
1082 arc_buf_hdr_t *b_hash_next;
1083 arc_flags_t b_flags;
1086 * This field stores the size of the data buffer after
1087 * compression, and is set in the arc's zio completion handlers.
1088 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1090 * While the block pointers can store up to 32MB in their psize
1091 * field, we can only store up to 32MB minus 512B. This is due
1092 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1093 * a field of zeros represents 512B in the bp). We can't use a
1094 * bias of 1 since we need to reserve a psize of zero, here, to
1095 * represent holes and embedded blocks.
1097 * This isn't a problem in practice, since the maximum size of a
1098 * buffer is limited to 16MB, so we never need to store 32MB in
1099 * this field. Even in the upstream illumos code base, the
1100 * maximum size of a buffer is limited to 16MB.
1105 * This field stores the size of the data buffer before
1106 * compression, and cannot change once set. It is in units
1107 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
1109 uint16_t b_lsize; /* immutable */
1110 uint64_t b_spa; /* immutable */
1112 /* L2ARC fields. Undefined when not in L2ARC. */
1113 l2arc_buf_hdr_t b_l2hdr;
1114 /* L1ARC fields. Undefined when in l2arc_only state */
1115 l1arc_buf_hdr_t b_l1hdr;
1118 #if defined(__FreeBSD__) && defined(_KERNEL)
1120 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
1125 val = arc_meta_limit;
1126 err = sysctl_handle_64(oidp, &val, 0, req);
1127 if (err != 0 || req->newptr == NULL)
1130 if (val <= 0 || val > arc_c_max)
1133 arc_meta_limit = val;
1138 sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS)
1143 val = arc_no_grow_shift;
1144 err = sysctl_handle_32(oidp, &val, 0, req);
1145 if (err != 0 || req->newptr == NULL)
1148 if (val >= arc_shrink_shift)
1151 arc_no_grow_shift = val;
1156 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
1162 err = sysctl_handle_64(oidp, &val, 0, req);
1163 if (err != 0 || req->newptr == NULL)
1166 if (zfs_arc_max == 0) {
1167 /* Loader tunable so blindly set */
1172 if (val < arc_abs_min || val > kmem_size())
1174 if (val < arc_c_min)
1176 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
1182 arc_p = (arc_c >> 1);
1184 if (zfs_arc_meta_limit == 0) {
1185 /* limit meta-data to 1/4 of the arc capacity */
1186 arc_meta_limit = arc_c_max / 4;
1189 /* if kmem_flags are set, lets try to use less memory */
1190 if (kmem_debugging())
1193 zfs_arc_max = arc_c;
1199 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
1205 err = sysctl_handle_64(oidp, &val, 0, req);
1206 if (err != 0 || req->newptr == NULL)
1209 if (zfs_arc_min == 0) {
1210 /* Loader tunable so blindly set */
1215 if (val < arc_abs_min || val > arc_c_max)
1220 if (zfs_arc_meta_min == 0)
1221 arc_meta_min = arc_c_min / 2;
1223 if (arc_c < arc_c_min)
1226 zfs_arc_min = arc_c_min;
1232 #define GHOST_STATE(state) \
1233 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1234 (state) == arc_l2c_only)
1236 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1237 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1238 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1239 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1240 #define HDR_PRESCIENT_PREFETCH(hdr) \
1241 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
1242 #define HDR_COMPRESSION_ENABLED(hdr) \
1243 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1245 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1246 #define HDR_L2_READING(hdr) \
1247 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1248 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1249 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1250 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1251 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1252 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1254 #define HDR_ISTYPE_METADATA(hdr) \
1255 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1256 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1258 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1259 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1261 /* For storing compression mode in b_flags */
1262 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1264 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1265 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1266 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1267 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1269 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1270 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1271 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1277 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1278 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1281 * Hash table routines
1284 #define HT_LOCK_PAD CACHE_LINE_SIZE
1289 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1293 #define BUF_LOCKS 256
1294 typedef struct buf_hash_table {
1296 arc_buf_hdr_t **ht_table;
1297 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1300 static buf_hash_table_t buf_hash_table;
1302 #define BUF_HASH_INDEX(spa, dva, birth) \
1303 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1304 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1305 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1306 #define HDR_LOCK(hdr) \
1307 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1309 uint64_t zfs_crc64_table[256];
1315 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1316 #define L2ARC_HEADROOM 2 /* num of writes */
1318 * If we discover during ARC scan any buffers to be compressed, we boost
1319 * our headroom for the next scanning cycle by this percentage multiple.
1321 #define L2ARC_HEADROOM_BOOST 200
1322 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1323 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1325 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1326 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1328 /* L2ARC Performance Tunables */
1329 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1330 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1331 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1332 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1333 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1334 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1335 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1336 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1337 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1339 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1340 &l2arc_write_max, 0, "max write size");
1341 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1342 &l2arc_write_boost, 0, "extra write during warmup");
1343 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1344 &l2arc_headroom, 0, "number of dev writes");
1345 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1346 &l2arc_feed_secs, 0, "interval seconds");
1347 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1348 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1350 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1351 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1352 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1353 &l2arc_feed_again, 0, "turbo warmup");
1354 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1355 &l2arc_norw, 0, "no reads during writes");
1357 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1358 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1359 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD,
1360 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1361 "size of anonymous state");
1362 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD,
1363 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1364 "size of anonymous state");
1366 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1367 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1368 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD,
1369 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1370 "size of metadata in mru state");
1371 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD,
1372 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1373 "size of data in mru state");
1375 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1376 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1377 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD,
1378 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1379 "size of metadata in mru ghost state");
1380 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD,
1381 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1382 "size of data in mru ghost state");
1384 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1385 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1386 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD,
1387 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1388 "size of metadata in mfu state");
1389 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD,
1390 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1391 "size of data in mfu state");
1393 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1394 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1395 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD,
1396 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0,
1397 "size of metadata in mfu ghost state");
1398 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD,
1399 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0,
1400 "size of data in mfu ghost state");
1402 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1403 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1405 SYSCTL_UINT(_vfs_zfs, OID_AUTO, arc_min_prefetch_ms, CTLFLAG_RW,
1406 &zfs_arc_min_prefetch_ms, 0, "Min life of prefetch block in ms");
1407 SYSCTL_UINT(_vfs_zfs, OID_AUTO, arc_min_prescient_prefetch_ms, CTLFLAG_RW,
1408 &zfs_arc_min_prescient_prefetch_ms, 0, "Min life of prescient prefetched block in ms");
1414 vdev_t *l2ad_vdev; /* vdev */
1415 spa_t *l2ad_spa; /* spa */
1416 uint64_t l2ad_hand; /* next write location */
1417 uint64_t l2ad_start; /* first addr on device */
1418 uint64_t l2ad_end; /* last addr on device */
1419 boolean_t l2ad_first; /* first sweep through */
1420 boolean_t l2ad_writing; /* currently writing */
1421 kmutex_t l2ad_mtx; /* lock for buffer list */
1422 list_t l2ad_buflist; /* buffer list */
1423 list_node_t l2ad_node; /* device list node */
1424 refcount_t l2ad_alloc; /* allocated bytes */
1427 static list_t L2ARC_dev_list; /* device list */
1428 static list_t *l2arc_dev_list; /* device list pointer */
1429 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1430 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1431 static list_t L2ARC_free_on_write; /* free after write buf list */
1432 static list_t *l2arc_free_on_write; /* free after write list ptr */
1433 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1434 static uint64_t l2arc_ndev; /* number of devices */
1436 typedef struct l2arc_read_callback {
1437 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1438 blkptr_t l2rcb_bp; /* original blkptr */
1439 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1440 int l2rcb_flags; /* original flags */
1441 abd_t *l2rcb_abd; /* temporary buffer */
1442 } l2arc_read_callback_t;
1444 typedef struct l2arc_write_callback {
1445 l2arc_dev_t *l2wcb_dev; /* device info */
1446 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1447 } l2arc_write_callback_t;
1449 typedef struct l2arc_data_free {
1450 /* protected by l2arc_free_on_write_mtx */
1453 arc_buf_contents_t l2df_type;
1454 list_node_t l2df_list_node;
1455 } l2arc_data_free_t;
1457 static kmutex_t l2arc_feed_thr_lock;
1458 static kcondvar_t l2arc_feed_thr_cv;
1459 static uint8_t l2arc_thread_exit;
1461 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1462 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1463 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1464 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1465 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1466 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1467 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1468 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1469 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1470 static boolean_t arc_is_overflowing();
1471 static void arc_buf_watch(arc_buf_t *);
1473 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1474 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1475 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1476 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1478 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1479 static void l2arc_read_done(zio_t *);
1482 l2arc_trim(const arc_buf_hdr_t *hdr)
1484 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1486 ASSERT(HDR_HAS_L2HDR(hdr));
1487 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1489 if (HDR_GET_PSIZE(hdr) != 0) {
1490 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1491 HDR_GET_PSIZE(hdr), 0);
1496 * We use Cityhash for this. It's fast, and has good hash properties without
1497 * requiring any large static buffers.
1500 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1502 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1505 #define HDR_EMPTY(hdr) \
1506 ((hdr)->b_dva.dva_word[0] == 0 && \
1507 (hdr)->b_dva.dva_word[1] == 0)
1509 #define HDR_EQUAL(spa, dva, birth, hdr) \
1510 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1511 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1512 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1515 buf_discard_identity(arc_buf_hdr_t *hdr)
1517 hdr->b_dva.dva_word[0] = 0;
1518 hdr->b_dva.dva_word[1] = 0;
1522 static arc_buf_hdr_t *
1523 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1525 const dva_t *dva = BP_IDENTITY(bp);
1526 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1527 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1528 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1531 mutex_enter(hash_lock);
1532 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1533 hdr = hdr->b_hash_next) {
1534 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1539 mutex_exit(hash_lock);
1545 * Insert an entry into the hash table. If there is already an element
1546 * equal to elem in the hash table, then the already existing element
1547 * will be returned and the new element will not be inserted.
1548 * Otherwise returns NULL.
1549 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1551 static arc_buf_hdr_t *
1552 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1554 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1555 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1556 arc_buf_hdr_t *fhdr;
1559 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1560 ASSERT(hdr->b_birth != 0);
1561 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1563 if (lockp != NULL) {
1565 mutex_enter(hash_lock);
1567 ASSERT(MUTEX_HELD(hash_lock));
1570 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1571 fhdr = fhdr->b_hash_next, i++) {
1572 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1576 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1577 buf_hash_table.ht_table[idx] = hdr;
1578 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1580 /* collect some hash table performance data */
1582 ARCSTAT_BUMP(arcstat_hash_collisions);
1584 ARCSTAT_BUMP(arcstat_hash_chains);
1586 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1589 ARCSTAT_BUMP(arcstat_hash_elements);
1590 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1596 buf_hash_remove(arc_buf_hdr_t *hdr)
1598 arc_buf_hdr_t *fhdr, **hdrp;
1599 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1601 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1602 ASSERT(HDR_IN_HASH_TABLE(hdr));
1604 hdrp = &buf_hash_table.ht_table[idx];
1605 while ((fhdr = *hdrp) != hdr) {
1606 ASSERT3P(fhdr, !=, NULL);
1607 hdrp = &fhdr->b_hash_next;
1609 *hdrp = hdr->b_hash_next;
1610 hdr->b_hash_next = NULL;
1611 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1613 /* collect some hash table performance data */
1614 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1616 if (buf_hash_table.ht_table[idx] &&
1617 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1618 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1622 * Global data structures and functions for the buf kmem cache.
1624 static kmem_cache_t *hdr_full_cache;
1625 static kmem_cache_t *hdr_l2only_cache;
1626 static kmem_cache_t *buf_cache;
1633 kmem_free(buf_hash_table.ht_table,
1634 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1635 for (i = 0; i < BUF_LOCKS; i++)
1636 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1637 kmem_cache_destroy(hdr_full_cache);
1638 kmem_cache_destroy(hdr_l2only_cache);
1639 kmem_cache_destroy(buf_cache);
1643 * Constructor callback - called when the cache is empty
1644 * and a new buf is requested.
1648 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1650 arc_buf_hdr_t *hdr = vbuf;
1652 bzero(hdr, HDR_FULL_SIZE);
1653 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1654 refcount_create(&hdr->b_l1hdr.b_refcnt);
1655 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1656 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1657 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1664 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1666 arc_buf_hdr_t *hdr = vbuf;
1668 bzero(hdr, HDR_L2ONLY_SIZE);
1669 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1676 buf_cons(void *vbuf, void *unused, int kmflag)
1678 arc_buf_t *buf = vbuf;
1680 bzero(buf, sizeof (arc_buf_t));
1681 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1682 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1688 * Destructor callback - called when a cached buf is
1689 * no longer required.
1693 hdr_full_dest(void *vbuf, void *unused)
1695 arc_buf_hdr_t *hdr = vbuf;
1697 ASSERT(HDR_EMPTY(hdr));
1698 cv_destroy(&hdr->b_l1hdr.b_cv);
1699 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1700 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1701 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1702 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1707 hdr_l2only_dest(void *vbuf, void *unused)
1709 arc_buf_hdr_t *hdr = vbuf;
1711 ASSERT(HDR_EMPTY(hdr));
1712 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1717 buf_dest(void *vbuf, void *unused)
1719 arc_buf_t *buf = vbuf;
1721 mutex_destroy(&buf->b_evict_lock);
1722 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1726 * Reclaim callback -- invoked when memory is low.
1730 hdr_recl(void *unused)
1732 dprintf("hdr_recl called\n");
1734 * umem calls the reclaim func when we destroy the buf cache,
1735 * which is after we do arc_fini().
1738 cv_signal(&arc_reclaim_thread_cv);
1745 uint64_t hsize = 1ULL << 12;
1749 * The hash table is big enough to fill all of physical memory
1750 * with an average block size of zfs_arc_average_blocksize (default 8K).
1751 * By default, the table will take up
1752 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1754 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1757 buf_hash_table.ht_mask = hsize - 1;
1758 buf_hash_table.ht_table =
1759 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1760 if (buf_hash_table.ht_table == NULL) {
1761 ASSERT(hsize > (1ULL << 8));
1766 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1767 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1768 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1769 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1771 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1772 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1774 for (i = 0; i < 256; i++)
1775 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1776 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1778 for (i = 0; i < BUF_LOCKS; i++) {
1779 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1780 NULL, MUTEX_DEFAULT, NULL);
1785 * This is the size that the buf occupies in memory. If the buf is compressed,
1786 * it will correspond to the compressed size. You should use this method of
1787 * getting the buf size unless you explicitly need the logical size.
1790 arc_buf_size(arc_buf_t *buf)
1792 return (ARC_BUF_COMPRESSED(buf) ?
1793 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1797 arc_buf_lsize(arc_buf_t *buf)
1799 return (HDR_GET_LSIZE(buf->b_hdr));
1803 arc_get_compression(arc_buf_t *buf)
1805 return (ARC_BUF_COMPRESSED(buf) ?
1806 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1809 #define ARC_MINTIME (hz>>4) /* 62 ms */
1811 static inline boolean_t
1812 arc_buf_is_shared(arc_buf_t *buf)
1814 boolean_t shared = (buf->b_data != NULL &&
1815 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1816 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1817 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1818 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1819 IMPLY(shared, ARC_BUF_SHARED(buf));
1820 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1823 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1824 * already being shared" requirement prevents us from doing that.
1831 * Free the checksum associated with this header. If there is no checksum, this
1835 arc_cksum_free(arc_buf_hdr_t *hdr)
1837 ASSERT(HDR_HAS_L1HDR(hdr));
1838 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1839 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1840 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1841 hdr->b_l1hdr.b_freeze_cksum = NULL;
1843 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1847 * Return true iff at least one of the bufs on hdr is not compressed.
1850 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1852 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1853 if (!ARC_BUF_COMPRESSED(b)) {
1861 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1862 * matches the checksum that is stored in the hdr. If there is no checksum,
1863 * or if the buf is compressed, this is a no-op.
1866 arc_cksum_verify(arc_buf_t *buf)
1868 arc_buf_hdr_t *hdr = buf->b_hdr;
1871 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1874 if (ARC_BUF_COMPRESSED(buf)) {
1875 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1876 arc_hdr_has_uncompressed_buf(hdr));
1880 ASSERT(HDR_HAS_L1HDR(hdr));
1882 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1883 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1884 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1888 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1889 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1890 panic("buffer modified while frozen!");
1891 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1895 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1897 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1898 boolean_t valid_cksum;
1900 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1901 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1904 * We rely on the blkptr's checksum to determine if the block
1905 * is valid or not. When compressed arc is enabled, the l2arc
1906 * writes the block to the l2arc just as it appears in the pool.
1907 * This allows us to use the blkptr's checksum to validate the
1908 * data that we just read off of the l2arc without having to store
1909 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1910 * arc is disabled, then the data written to the l2arc is always
1911 * uncompressed and won't match the block as it exists in the main
1912 * pool. When this is the case, we must first compress it if it is
1913 * compressed on the main pool before we can validate the checksum.
1915 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1916 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1917 uint64_t lsize = HDR_GET_LSIZE(hdr);
1920 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1921 csize = zio_compress_data(compress, zio->io_abd,
1922 abd_to_buf(cdata), lsize);
1924 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1925 if (csize < HDR_GET_PSIZE(hdr)) {
1927 * Compressed blocks are always a multiple of the
1928 * smallest ashift in the pool. Ideally, we would
1929 * like to round up the csize to the next
1930 * spa_min_ashift but that value may have changed
1931 * since the block was last written. Instead,
1932 * we rely on the fact that the hdr's psize
1933 * was set to the psize of the block when it was
1934 * last written. We set the csize to that value
1935 * and zero out any part that should not contain
1938 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1939 csize = HDR_GET_PSIZE(hdr);
1941 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1945 * Block pointers always store the checksum for the logical data.
1946 * If the block pointer has the gang bit set, then the checksum
1947 * it represents is for the reconstituted data and not for an
1948 * individual gang member. The zio pipeline, however, must be able to
1949 * determine the checksum of each of the gang constituents so it
1950 * treats the checksum comparison differently than what we need
1951 * for l2arc blocks. This prevents us from using the
1952 * zio_checksum_error() interface directly. Instead we must call the
1953 * zio_checksum_error_impl() so that we can ensure the checksum is
1954 * generated using the correct checksum algorithm and accounts for the
1955 * logical I/O size and not just a gang fragment.
1957 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1958 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1959 zio->io_offset, NULL) == 0);
1960 zio_pop_transforms(zio);
1961 return (valid_cksum);
1965 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1966 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1967 * isn't modified later on. If buf is compressed or there is already a checksum
1968 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1971 arc_cksum_compute(arc_buf_t *buf)
1973 arc_buf_hdr_t *hdr = buf->b_hdr;
1975 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1978 ASSERT(HDR_HAS_L1HDR(hdr));
1980 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1981 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1982 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1983 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1985 } else if (ARC_BUF_COMPRESSED(buf)) {
1986 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1990 ASSERT(!ARC_BUF_COMPRESSED(buf));
1991 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1993 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1994 hdr->b_l1hdr.b_freeze_cksum);
1995 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2003 typedef struct procctl {
2011 arc_buf_unwatch(arc_buf_t *buf)
2018 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2019 ctl.prwatch.pr_size = 0;
2020 ctl.prwatch.pr_wflags = 0;
2021 result = write(arc_procfd, &ctl, sizeof (ctl));
2022 ASSERT3U(result, ==, sizeof (ctl));
2029 arc_buf_watch(arc_buf_t *buf)
2036 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2037 ctl.prwatch.pr_size = arc_buf_size(buf);
2038 ctl.prwatch.pr_wflags = WA_WRITE;
2039 result = write(arc_procfd, &ctl, sizeof (ctl));
2040 ASSERT3U(result, ==, sizeof (ctl));
2044 #endif /* illumos */
2046 static arc_buf_contents_t
2047 arc_buf_type(arc_buf_hdr_t *hdr)
2049 arc_buf_contents_t type;
2050 if (HDR_ISTYPE_METADATA(hdr)) {
2051 type = ARC_BUFC_METADATA;
2053 type = ARC_BUFC_DATA;
2055 VERIFY3U(hdr->b_type, ==, type);
2060 arc_is_metadata(arc_buf_t *buf)
2062 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2066 arc_bufc_to_flags(arc_buf_contents_t type)
2070 /* metadata field is 0 if buffer contains normal data */
2072 case ARC_BUFC_METADATA:
2073 return (ARC_FLAG_BUFC_METADATA);
2077 panic("undefined ARC buffer type!");
2078 return ((uint32_t)-1);
2082 arc_buf_thaw(arc_buf_t *buf)
2084 arc_buf_hdr_t *hdr = buf->b_hdr;
2086 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2087 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2089 arc_cksum_verify(buf);
2092 * Compressed buffers do not manipulate the b_freeze_cksum or
2093 * allocate b_thawed.
2095 if (ARC_BUF_COMPRESSED(buf)) {
2096 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2097 arc_hdr_has_uncompressed_buf(hdr));
2101 ASSERT(HDR_HAS_L1HDR(hdr));
2102 arc_cksum_free(hdr);
2104 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2106 if (zfs_flags & ZFS_DEBUG_MODIFY) {
2107 if (hdr->b_l1hdr.b_thawed != NULL)
2108 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2109 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2113 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2116 arc_buf_unwatch(buf);
2121 arc_buf_freeze(arc_buf_t *buf)
2123 arc_buf_hdr_t *hdr = buf->b_hdr;
2124 kmutex_t *hash_lock;
2126 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2129 if (ARC_BUF_COMPRESSED(buf)) {
2130 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
2131 arc_hdr_has_uncompressed_buf(hdr));
2135 hash_lock = HDR_LOCK(hdr);
2136 mutex_enter(hash_lock);
2138 ASSERT(HDR_HAS_L1HDR(hdr));
2139 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
2140 hdr->b_l1hdr.b_state == arc_anon);
2141 arc_cksum_compute(buf);
2142 mutex_exit(hash_lock);
2146 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2147 * the following functions should be used to ensure that the flags are
2148 * updated in a thread-safe way. When manipulating the flags either
2149 * the hash_lock must be held or the hdr must be undiscoverable. This
2150 * ensures that we're not racing with any other threads when updating
2154 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2156 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2157 hdr->b_flags |= flags;
2161 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2163 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2164 hdr->b_flags &= ~flags;
2168 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
2169 * done in a special way since we have to clear and set bits
2170 * at the same time. Consumers that wish to set the compression bits
2171 * must use this function to ensure that the flags are updated in
2172 * thread-safe manner.
2175 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
2177 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2180 * Holes and embedded blocks will always have a psize = 0 so
2181 * we ignore the compression of the blkptr and set the
2182 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
2183 * Holes and embedded blocks remain anonymous so we don't
2184 * want to uncompress them. Mark them as uncompressed.
2186 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
2187 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2188 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
2189 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
2190 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2192 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
2193 HDR_SET_COMPRESS(hdr, cmp);
2194 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
2195 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
2200 * Looks for another buf on the same hdr which has the data decompressed, copies
2201 * from it, and returns true. If no such buf exists, returns false.
2204 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
2206 arc_buf_hdr_t *hdr = buf->b_hdr;
2207 boolean_t copied = B_FALSE;
2209 ASSERT(HDR_HAS_L1HDR(hdr));
2210 ASSERT3P(buf->b_data, !=, NULL);
2211 ASSERT(!ARC_BUF_COMPRESSED(buf));
2213 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2214 from = from->b_next) {
2215 /* can't use our own data buffer */
2220 if (!ARC_BUF_COMPRESSED(from)) {
2221 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2228 * There were no decompressed bufs, so there should not be a
2229 * checksum on the hdr either.
2231 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
2237 * Given a buf that has a data buffer attached to it, this function will
2238 * efficiently fill the buf with data of the specified compression setting from
2239 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2240 * are already sharing a data buf, no copy is performed.
2242 * If the buf is marked as compressed but uncompressed data was requested, this
2243 * will allocate a new data buffer for the buf, remove that flag, and fill the
2244 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2245 * uncompressed data, and (since we haven't added support for it yet) if you
2246 * want compressed data your buf must already be marked as compressed and have
2247 * the correct-sized data buffer.
2250 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2252 arc_buf_hdr_t *hdr = buf->b_hdr;
2253 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2254 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
2256 ASSERT3P(buf->b_data, !=, NULL);
2257 IMPLY(compressed, hdr_compressed);
2258 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
2260 if (hdr_compressed == compressed) {
2261 if (!arc_buf_is_shared(buf)) {
2262 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2266 ASSERT(hdr_compressed);
2267 ASSERT(!compressed);
2268 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
2271 * If the buf is sharing its data with the hdr, unlink it and
2272 * allocate a new data buffer for the buf.
2274 if (arc_buf_is_shared(buf)) {
2275 ASSERT(ARC_BUF_COMPRESSED(buf));
2277 /* We need to give the buf it's own b_data */
2278 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2280 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2281 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2283 /* Previously overhead was 0; just add new overhead */
2284 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2285 } else if (ARC_BUF_COMPRESSED(buf)) {
2286 /* We need to reallocate the buf's b_data */
2287 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2290 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2292 /* We increased the size of b_data; update overhead */
2293 ARCSTAT_INCR(arcstat_overhead_size,
2294 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2298 * Regardless of the buf's previous compression settings, it
2299 * should not be compressed at the end of this function.
2301 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2304 * Try copying the data from another buf which already has a
2305 * decompressed version. If that's not possible, it's time to
2306 * bite the bullet and decompress the data from the hdr.
2308 if (arc_buf_try_copy_decompressed_data(buf)) {
2309 /* Skip byteswapping and checksumming (already done) */
2310 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2313 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2314 hdr->b_l1hdr.b_pabd, buf->b_data,
2315 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2318 * Absent hardware errors or software bugs, this should
2319 * be impossible, but log it anyway so we can debug it.
2323 "hdr %p, compress %d, psize %d, lsize %d",
2324 hdr, HDR_GET_COMPRESS(hdr),
2325 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2326 return (SET_ERROR(EIO));
2331 /* Byteswap the buf's data if necessary */
2332 if (bswap != DMU_BSWAP_NUMFUNCS) {
2333 ASSERT(!HDR_SHARED_DATA(hdr));
2334 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2335 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2338 /* Compute the hdr's checksum if necessary */
2339 arc_cksum_compute(buf);
2345 arc_decompress(arc_buf_t *buf)
2347 return (arc_buf_fill(buf, B_FALSE));
2351 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2354 arc_hdr_size(arc_buf_hdr_t *hdr)
2358 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2359 HDR_GET_PSIZE(hdr) > 0) {
2360 size = HDR_GET_PSIZE(hdr);
2362 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2363 size = HDR_GET_LSIZE(hdr);
2369 * Increment the amount of evictable space in the arc_state_t's refcount.
2370 * We account for the space used by the hdr and the arc buf individually
2371 * so that we can add and remove them from the refcount individually.
2374 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2376 arc_buf_contents_t type = arc_buf_type(hdr);
2378 ASSERT(HDR_HAS_L1HDR(hdr));
2380 if (GHOST_STATE(state)) {
2381 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2382 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2383 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2384 (void) refcount_add_many(&state->arcs_esize[type],
2385 HDR_GET_LSIZE(hdr), hdr);
2389 ASSERT(!GHOST_STATE(state));
2390 if (hdr->b_l1hdr.b_pabd != NULL) {
2391 (void) refcount_add_many(&state->arcs_esize[type],
2392 arc_hdr_size(hdr), hdr);
2394 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2395 buf = buf->b_next) {
2396 if (arc_buf_is_shared(buf))
2398 (void) refcount_add_many(&state->arcs_esize[type],
2399 arc_buf_size(buf), buf);
2404 * Decrement the amount of evictable space in the arc_state_t's refcount.
2405 * We account for the space used by the hdr and the arc buf individually
2406 * so that we can add and remove them from the refcount individually.
2409 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2411 arc_buf_contents_t type = arc_buf_type(hdr);
2413 ASSERT(HDR_HAS_L1HDR(hdr));
2415 if (GHOST_STATE(state)) {
2416 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2417 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2418 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2419 (void) refcount_remove_many(&state->arcs_esize[type],
2420 HDR_GET_LSIZE(hdr), hdr);
2424 ASSERT(!GHOST_STATE(state));
2425 if (hdr->b_l1hdr.b_pabd != NULL) {
2426 (void) refcount_remove_many(&state->arcs_esize[type],
2427 arc_hdr_size(hdr), hdr);
2429 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2430 buf = buf->b_next) {
2431 if (arc_buf_is_shared(buf))
2433 (void) refcount_remove_many(&state->arcs_esize[type],
2434 arc_buf_size(buf), buf);
2439 * Add a reference to this hdr indicating that someone is actively
2440 * referencing that memory. When the refcount transitions from 0 to 1,
2441 * we remove it from the respective arc_state_t list to indicate that
2442 * it is not evictable.
2445 add_reference(arc_buf_hdr_t *hdr, void *tag)
2447 ASSERT(HDR_HAS_L1HDR(hdr));
2448 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2449 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2450 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2451 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2454 arc_state_t *state = hdr->b_l1hdr.b_state;
2456 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2457 (state != arc_anon)) {
2458 /* We don't use the L2-only state list. */
2459 if (state != arc_l2c_only) {
2460 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2462 arc_evictable_space_decrement(hdr, state);
2464 /* remove the prefetch flag if we get a reference */
2465 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2470 * Remove a reference from this hdr. When the reference transitions from
2471 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2472 * list making it eligible for eviction.
2475 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2478 arc_state_t *state = hdr->b_l1hdr.b_state;
2480 ASSERT(HDR_HAS_L1HDR(hdr));
2481 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2482 ASSERT(!GHOST_STATE(state));
2485 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2486 * check to prevent usage of the arc_l2c_only list.
2488 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2489 (state != arc_anon)) {
2490 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2491 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2492 arc_evictable_space_increment(hdr, state);
2498 * Move the supplied buffer to the indicated state. The hash lock
2499 * for the buffer must be held by the caller.
2502 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2503 kmutex_t *hash_lock)
2505 arc_state_t *old_state;
2508 boolean_t update_old, update_new;
2509 arc_buf_contents_t buftype = arc_buf_type(hdr);
2512 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2513 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2514 * L1 hdr doesn't always exist when we change state to arc_anon before
2515 * destroying a header, in which case reallocating to add the L1 hdr is
2518 if (HDR_HAS_L1HDR(hdr)) {
2519 old_state = hdr->b_l1hdr.b_state;
2520 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2521 bufcnt = hdr->b_l1hdr.b_bufcnt;
2522 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2524 old_state = arc_l2c_only;
2527 update_old = B_FALSE;
2529 update_new = update_old;
2531 ASSERT(MUTEX_HELD(hash_lock));
2532 ASSERT3P(new_state, !=, old_state);
2533 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2534 ASSERT(old_state != arc_anon || bufcnt <= 1);
2537 * If this buffer is evictable, transfer it from the
2538 * old state list to the new state list.
2541 if (old_state != arc_anon && old_state != arc_l2c_only) {
2542 ASSERT(HDR_HAS_L1HDR(hdr));
2543 multilist_remove(old_state->arcs_list[buftype], hdr);
2545 if (GHOST_STATE(old_state)) {
2547 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2548 update_old = B_TRUE;
2550 arc_evictable_space_decrement(hdr, old_state);
2552 if (new_state != arc_anon && new_state != arc_l2c_only) {
2555 * An L1 header always exists here, since if we're
2556 * moving to some L1-cached state (i.e. not l2c_only or
2557 * anonymous), we realloc the header to add an L1hdr
2560 ASSERT(HDR_HAS_L1HDR(hdr));
2561 multilist_insert(new_state->arcs_list[buftype], hdr);
2563 if (GHOST_STATE(new_state)) {
2565 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2566 update_new = B_TRUE;
2568 arc_evictable_space_increment(hdr, new_state);
2572 ASSERT(!HDR_EMPTY(hdr));
2573 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2574 buf_hash_remove(hdr);
2576 /* adjust state sizes (ignore arc_l2c_only) */
2578 if (update_new && new_state != arc_l2c_only) {
2579 ASSERT(HDR_HAS_L1HDR(hdr));
2580 if (GHOST_STATE(new_state)) {
2584 * When moving a header to a ghost state, we first
2585 * remove all arc buffers. Thus, we'll have a
2586 * bufcnt of zero, and no arc buffer to use for
2587 * the reference. As a result, we use the arc
2588 * header pointer for the reference.
2590 (void) refcount_add_many(&new_state->arcs_size,
2591 HDR_GET_LSIZE(hdr), hdr);
2592 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2594 uint32_t buffers = 0;
2597 * Each individual buffer holds a unique reference,
2598 * thus we must remove each of these references one
2601 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2602 buf = buf->b_next) {
2603 ASSERT3U(bufcnt, !=, 0);
2607 * When the arc_buf_t is sharing the data
2608 * block with the hdr, the owner of the
2609 * reference belongs to the hdr. Only
2610 * add to the refcount if the arc_buf_t is
2613 if (arc_buf_is_shared(buf))
2616 (void) refcount_add_many(&new_state->arcs_size,
2617 arc_buf_size(buf), buf);
2619 ASSERT3U(bufcnt, ==, buffers);
2621 if (hdr->b_l1hdr.b_pabd != NULL) {
2622 (void) refcount_add_many(&new_state->arcs_size,
2623 arc_hdr_size(hdr), hdr);
2625 ASSERT(GHOST_STATE(old_state));
2630 if (update_old && old_state != arc_l2c_only) {
2631 ASSERT(HDR_HAS_L1HDR(hdr));
2632 if (GHOST_STATE(old_state)) {
2634 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2637 * When moving a header off of a ghost state,
2638 * the header will not contain any arc buffers.
2639 * We use the arc header pointer for the reference
2640 * which is exactly what we did when we put the
2641 * header on the ghost state.
2644 (void) refcount_remove_many(&old_state->arcs_size,
2645 HDR_GET_LSIZE(hdr), hdr);
2647 uint32_t buffers = 0;
2650 * Each individual buffer holds a unique reference,
2651 * thus we must remove each of these references one
2654 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2655 buf = buf->b_next) {
2656 ASSERT3U(bufcnt, !=, 0);
2660 * When the arc_buf_t is sharing the data
2661 * block with the hdr, the owner of the
2662 * reference belongs to the hdr. Only
2663 * add to the refcount if the arc_buf_t is
2666 if (arc_buf_is_shared(buf))
2669 (void) refcount_remove_many(
2670 &old_state->arcs_size, arc_buf_size(buf),
2673 ASSERT3U(bufcnt, ==, buffers);
2674 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2675 (void) refcount_remove_many(
2676 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2680 if (HDR_HAS_L1HDR(hdr))
2681 hdr->b_l1hdr.b_state = new_state;
2684 * L2 headers should never be on the L2 state list since they don't
2685 * have L1 headers allocated.
2687 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2688 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2692 arc_space_consume(uint64_t space, arc_space_type_t type)
2694 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2697 case ARC_SPACE_DATA:
2698 aggsum_add(&astat_data_size, space);
2700 case ARC_SPACE_META:
2701 aggsum_add(&astat_metadata_size, space);
2703 case ARC_SPACE_OTHER:
2704 aggsum_add(&astat_other_size, space);
2706 case ARC_SPACE_HDRS:
2707 aggsum_add(&astat_hdr_size, space);
2709 case ARC_SPACE_L2HDRS:
2710 aggsum_add(&astat_l2_hdr_size, space);
2714 if (type != ARC_SPACE_DATA)
2715 aggsum_add(&arc_meta_used, space);
2717 aggsum_add(&arc_size, space);
2721 arc_space_return(uint64_t space, arc_space_type_t type)
2723 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2726 case ARC_SPACE_DATA:
2727 aggsum_add(&astat_data_size, -space);
2729 case ARC_SPACE_META:
2730 aggsum_add(&astat_metadata_size, -space);
2732 case ARC_SPACE_OTHER:
2733 aggsum_add(&astat_other_size, -space);
2735 case ARC_SPACE_HDRS:
2736 aggsum_add(&astat_hdr_size, -space);
2738 case ARC_SPACE_L2HDRS:
2739 aggsum_add(&astat_l2_hdr_size, -space);
2743 if (type != ARC_SPACE_DATA) {
2744 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2746 * We use the upper bound here rather than the precise value
2747 * because the arc_meta_max value doesn't need to be
2748 * precise. It's only consumed by humans via arcstats.
2750 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2751 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2752 aggsum_add(&arc_meta_used, -space);
2755 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2756 aggsum_add(&arc_size, -space);
2760 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2761 * with the hdr's b_pabd.
2764 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2767 * The criteria for sharing a hdr's data are:
2768 * 1. the hdr's compression matches the buf's compression
2769 * 2. the hdr doesn't need to be byteswapped
2770 * 3. the hdr isn't already being shared
2771 * 4. the buf is either compressed or it is the last buf in the hdr list
2773 * Criterion #4 maintains the invariant that shared uncompressed
2774 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2775 * might ask, "if a compressed buf is allocated first, won't that be the
2776 * last thing in the list?", but in that case it's impossible to create
2777 * a shared uncompressed buf anyway (because the hdr must be compressed
2778 * to have the compressed buf). You might also think that #3 is
2779 * sufficient to make this guarantee, however it's possible
2780 * (specifically in the rare L2ARC write race mentioned in
2781 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2782 * is sharable, but wasn't at the time of its allocation. Rather than
2783 * allow a new shared uncompressed buf to be created and then shuffle
2784 * the list around to make it the last element, this simply disallows
2785 * sharing if the new buf isn't the first to be added.
2787 ASSERT3P(buf->b_hdr, ==, hdr);
2788 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2789 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2790 return (buf_compressed == hdr_compressed &&
2791 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2792 !HDR_SHARED_DATA(hdr) &&
2793 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2797 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2798 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2799 * copy was made successfully, or an error code otherwise.
2802 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2803 boolean_t fill, arc_buf_t **ret)
2807 ASSERT(HDR_HAS_L1HDR(hdr));
2808 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2809 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2810 hdr->b_type == ARC_BUFC_METADATA);
2811 ASSERT3P(ret, !=, NULL);
2812 ASSERT3P(*ret, ==, NULL);
2814 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2817 buf->b_next = hdr->b_l1hdr.b_buf;
2820 add_reference(hdr, tag);
2823 * We're about to change the hdr's b_flags. We must either
2824 * hold the hash_lock or be undiscoverable.
2826 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2829 * Only honor requests for compressed bufs if the hdr is actually
2832 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2833 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2836 * If the hdr's data can be shared then we share the data buffer and
2837 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2838 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2839 * buffer to store the buf's data.
2841 * There are two additional restrictions here because we're sharing
2842 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2843 * actively involved in an L2ARC write, because if this buf is used by
2844 * an arc_write() then the hdr's data buffer will be released when the
2845 * write completes, even though the L2ARC write might still be using it.
2846 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2847 * need to be ABD-aware.
2849 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2850 abd_is_linear(hdr->b_l1hdr.b_pabd);
2852 /* Set up b_data and sharing */
2854 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2855 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2856 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2859 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2860 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2862 VERIFY3P(buf->b_data, !=, NULL);
2864 hdr->b_l1hdr.b_buf = buf;
2865 hdr->b_l1hdr.b_bufcnt += 1;
2868 * If the user wants the data from the hdr, we need to either copy or
2869 * decompress the data.
2872 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2878 static char *arc_onloan_tag = "onloan";
2881 arc_loaned_bytes_update(int64_t delta)
2883 atomic_add_64(&arc_loaned_bytes, delta);
2885 /* assert that it did not wrap around */
2886 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2890 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2891 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2892 * buffers must be returned to the arc before they can be used by the DMU or
2896 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2898 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2899 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2901 arc_loaned_bytes_update(arc_buf_size(buf));
2907 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2908 enum zio_compress compression_type)
2910 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2911 psize, lsize, compression_type);
2913 arc_loaned_bytes_update(arc_buf_size(buf));
2920 * Return a loaned arc buffer to the arc.
2923 arc_return_buf(arc_buf_t *buf, void *tag)
2925 arc_buf_hdr_t *hdr = buf->b_hdr;
2927 ASSERT3P(buf->b_data, !=, NULL);
2928 ASSERT(HDR_HAS_L1HDR(hdr));
2929 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2930 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2932 arc_loaned_bytes_update(-arc_buf_size(buf));
2935 /* Detach an arc_buf from a dbuf (tag) */
2937 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2939 arc_buf_hdr_t *hdr = buf->b_hdr;
2941 ASSERT3P(buf->b_data, !=, NULL);
2942 ASSERT(HDR_HAS_L1HDR(hdr));
2943 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2944 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2946 arc_loaned_bytes_update(arc_buf_size(buf));
2950 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2952 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2955 df->l2df_size = size;
2956 df->l2df_type = type;
2957 mutex_enter(&l2arc_free_on_write_mtx);
2958 list_insert_head(l2arc_free_on_write, df);
2959 mutex_exit(&l2arc_free_on_write_mtx);
2963 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2965 arc_state_t *state = hdr->b_l1hdr.b_state;
2966 arc_buf_contents_t type = arc_buf_type(hdr);
2967 uint64_t size = arc_hdr_size(hdr);
2969 /* protected by hash lock, if in the hash table */
2970 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2971 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2972 ASSERT(state != arc_anon && state != arc_l2c_only);
2974 (void) refcount_remove_many(&state->arcs_esize[type],
2977 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2978 if (type == ARC_BUFC_METADATA) {
2979 arc_space_return(size, ARC_SPACE_META);
2981 ASSERT(type == ARC_BUFC_DATA);
2982 arc_space_return(size, ARC_SPACE_DATA);
2985 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2989 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2990 * data buffer, we transfer the refcount ownership to the hdr and update
2991 * the appropriate kstats.
2994 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2996 arc_state_t *state = hdr->b_l1hdr.b_state;
2998 ASSERT(arc_can_share(hdr, buf));
2999 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3000 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3003 * Start sharing the data buffer. We transfer the
3004 * refcount ownership to the hdr since it always owns
3005 * the refcount whenever an arc_buf_t is shared.
3007 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
3008 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
3009 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
3010 HDR_ISTYPE_METADATA(hdr));
3011 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3012 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3015 * Since we've transferred ownership to the hdr we need
3016 * to increment its compressed and uncompressed kstats and
3017 * decrement the overhead size.
3019 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3020 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3021 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3025 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3027 arc_state_t *state = hdr->b_l1hdr.b_state;
3029 ASSERT(arc_buf_is_shared(buf));
3030 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3031 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3034 * We are no longer sharing this buffer so we need
3035 * to transfer its ownership to the rightful owner.
3037 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
3038 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3039 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
3040 abd_put(hdr->b_l1hdr.b_pabd);
3041 hdr->b_l1hdr.b_pabd = NULL;
3042 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
3045 * Since the buffer is no longer shared between
3046 * the arc buf and the hdr, count it as overhead.
3048 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3049 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3050 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3054 * Remove an arc_buf_t from the hdr's buf list and return the last
3055 * arc_buf_t on the list. If no buffers remain on the list then return
3059 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3061 ASSERT(HDR_HAS_L1HDR(hdr));
3062 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3064 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3065 arc_buf_t *lastbuf = NULL;
3068 * Remove the buf from the hdr list and locate the last
3069 * remaining buffer on the list.
3071 while (*bufp != NULL) {
3073 *bufp = buf->b_next;
3076 * If we've removed a buffer in the middle of
3077 * the list then update the lastbuf and update
3080 if (*bufp != NULL) {
3082 bufp = &(*bufp)->b_next;
3086 ASSERT3P(lastbuf, !=, buf);
3087 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
3088 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
3089 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3095 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
3099 arc_buf_destroy_impl(arc_buf_t *buf)
3101 arc_buf_hdr_t *hdr = buf->b_hdr;
3104 * Free up the data associated with the buf but only if we're not
3105 * sharing this with the hdr. If we are sharing it with the hdr, the
3106 * hdr is responsible for doing the free.
3108 if (buf->b_data != NULL) {
3110 * We're about to change the hdr's b_flags. We must either
3111 * hold the hash_lock or be undiscoverable.
3113 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3115 arc_cksum_verify(buf);
3117 arc_buf_unwatch(buf);
3120 if (arc_buf_is_shared(buf)) {
3121 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3123 uint64_t size = arc_buf_size(buf);
3124 arc_free_data_buf(hdr, buf->b_data, size, buf);
3125 ARCSTAT_INCR(arcstat_overhead_size, -size);
3129 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3130 hdr->b_l1hdr.b_bufcnt -= 1;
3133 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3135 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3137 * If the current arc_buf_t is sharing its data buffer with the
3138 * hdr, then reassign the hdr's b_pabd to share it with the new
3139 * buffer at the end of the list. The shared buffer is always
3140 * the last one on the hdr's buffer list.
3142 * There is an equivalent case for compressed bufs, but since
3143 * they aren't guaranteed to be the last buf in the list and
3144 * that is an exceedingly rare case, we just allow that space be
3145 * wasted temporarily.
3147 if (lastbuf != NULL) {
3148 /* Only one buf can be shared at once */
3149 VERIFY(!arc_buf_is_shared(lastbuf));
3150 /* hdr is uncompressed so can't have compressed buf */
3151 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3153 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3154 arc_hdr_free_pabd(hdr);
3157 * We must setup a new shared block between the
3158 * last buffer and the hdr. The data would have
3159 * been allocated by the arc buf so we need to transfer
3160 * ownership to the hdr since it's now being shared.
3162 arc_share_buf(hdr, lastbuf);
3164 } else if (HDR_SHARED_DATA(hdr)) {
3166 * Uncompressed shared buffers are always at the end
3167 * of the list. Compressed buffers don't have the
3168 * same requirements. This makes it hard to
3169 * simply assert that the lastbuf is shared so
3170 * we rely on the hdr's compression flags to determine
3171 * if we have a compressed, shared buffer.
3173 ASSERT3P(lastbuf, !=, NULL);
3174 ASSERT(arc_buf_is_shared(lastbuf) ||
3175 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
3179 * Free the checksum if we're removing the last uncompressed buf from
3182 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3183 arc_cksum_free(hdr);
3186 /* clean up the buf */
3188 kmem_cache_free(buf_cache, buf);
3192 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
3194 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3195 ASSERT(HDR_HAS_L1HDR(hdr));
3196 ASSERT(!HDR_SHARED_DATA(hdr));
3198 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3199 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
3200 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3201 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3203 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3204 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3208 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3210 ASSERT(HDR_HAS_L1HDR(hdr));
3211 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3214 * If the hdr is currently being written to the l2arc then
3215 * we defer freeing the data by adding it to the l2arc_free_on_write
3216 * list. The l2arc will free the data once it's finished
3217 * writing it to the l2arc device.
3219 if (HDR_L2_WRITING(hdr)) {
3220 arc_hdr_free_on_write(hdr);
3221 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3223 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3224 arc_hdr_size(hdr), hdr);
3226 hdr->b_l1hdr.b_pabd = NULL;
3227 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3229 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3230 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3233 static arc_buf_hdr_t *
3234 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3235 enum zio_compress compression_type, arc_buf_contents_t type)
3239 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3241 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3242 ASSERT(HDR_EMPTY(hdr));
3243 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3244 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3245 HDR_SET_PSIZE(hdr, psize);
3246 HDR_SET_LSIZE(hdr, lsize);
3250 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3251 arc_hdr_set_compress(hdr, compression_type);
3253 hdr->b_l1hdr.b_state = arc_anon;
3254 hdr->b_l1hdr.b_arc_access = 0;
3255 hdr->b_l1hdr.b_bufcnt = 0;
3256 hdr->b_l1hdr.b_buf = NULL;
3259 * Allocate the hdr's buffer. This will contain either
3260 * the compressed or uncompressed data depending on the block
3261 * it references and compressed arc enablement.
3263 arc_hdr_alloc_pabd(hdr);
3264 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3270 * Transition between the two allocation states for the arc_buf_hdr struct.
3271 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3272 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3273 * version is used when a cache buffer is only in the L2ARC in order to reduce
3276 static arc_buf_hdr_t *
3277 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3279 ASSERT(HDR_HAS_L2HDR(hdr));
3281 arc_buf_hdr_t *nhdr;
3282 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3284 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3285 (old == hdr_l2only_cache && new == hdr_full_cache));
3287 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3289 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3290 buf_hash_remove(hdr);
3292 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3294 if (new == hdr_full_cache) {
3295 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3297 * arc_access and arc_change_state need to be aware that a
3298 * header has just come out of L2ARC, so we set its state to
3299 * l2c_only even though it's about to change.
3301 nhdr->b_l1hdr.b_state = arc_l2c_only;
3303 /* Verify previous threads set to NULL before freeing */
3304 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3306 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3307 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3308 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3311 * If we've reached here, We must have been called from
3312 * arc_evict_hdr(), as such we should have already been
3313 * removed from any ghost list we were previously on
3314 * (which protects us from racing with arc_evict_state),
3315 * thus no locking is needed during this check.
3317 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3320 * A buffer must not be moved into the arc_l2c_only
3321 * state if it's not finished being written out to the
3322 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3323 * might try to be accessed, even though it was removed.
3325 VERIFY(!HDR_L2_WRITING(hdr));
3326 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3329 if (hdr->b_l1hdr.b_thawed != NULL) {
3330 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3331 hdr->b_l1hdr.b_thawed = NULL;
3335 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3338 * The header has been reallocated so we need to re-insert it into any
3341 (void) buf_hash_insert(nhdr, NULL);
3343 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3345 mutex_enter(&dev->l2ad_mtx);
3348 * We must place the realloc'ed header back into the list at
3349 * the same spot. Otherwise, if it's placed earlier in the list,
3350 * l2arc_write_buffers() could find it during the function's
3351 * write phase, and try to write it out to the l2arc.
3353 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3354 list_remove(&dev->l2ad_buflist, hdr);
3356 mutex_exit(&dev->l2ad_mtx);
3359 * Since we're using the pointer address as the tag when
3360 * incrementing and decrementing the l2ad_alloc refcount, we
3361 * must remove the old pointer (that we're about to destroy) and
3362 * add the new pointer to the refcount. Otherwise we'd remove
3363 * the wrong pointer address when calling arc_hdr_destroy() later.
3366 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3367 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3369 buf_discard_identity(hdr);
3370 kmem_cache_free(old, hdr);
3376 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3377 * The buf is returned thawed since we expect the consumer to modify it.
3380 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3382 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3383 ZIO_COMPRESS_OFF, type);
3384 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3386 arc_buf_t *buf = NULL;
3387 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3394 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3395 * for bufs containing metadata.
3398 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3399 enum zio_compress compression_type)
3401 ASSERT3U(lsize, >, 0);
3402 ASSERT3U(lsize, >=, psize);
3403 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3404 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3406 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3407 compression_type, ARC_BUFC_DATA);
3408 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3410 arc_buf_t *buf = NULL;
3411 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3413 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3415 if (!arc_buf_is_shared(buf)) {
3417 * To ensure that the hdr has the correct data in it if we call
3418 * arc_decompress() on this buf before it's been written to
3419 * disk, it's easiest if we just set up sharing between the
3422 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3423 arc_hdr_free_pabd(hdr);
3424 arc_share_buf(hdr, buf);
3431 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3433 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3434 l2arc_dev_t *dev = l2hdr->b_dev;
3435 uint64_t psize = arc_hdr_size(hdr);
3437 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3438 ASSERT(HDR_HAS_L2HDR(hdr));
3440 list_remove(&dev->l2ad_buflist, hdr);
3442 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3443 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3445 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3447 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3448 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3452 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3454 if (HDR_HAS_L1HDR(hdr)) {
3455 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3456 hdr->b_l1hdr.b_bufcnt > 0);
3457 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3458 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3460 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3461 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3463 if (!HDR_EMPTY(hdr))
3464 buf_discard_identity(hdr);
3466 if (HDR_HAS_L2HDR(hdr)) {
3467 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3468 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3471 mutex_enter(&dev->l2ad_mtx);
3474 * Even though we checked this conditional above, we
3475 * need to check this again now that we have the
3476 * l2ad_mtx. This is because we could be racing with
3477 * another thread calling l2arc_evict() which might have
3478 * destroyed this header's L2 portion as we were waiting
3479 * to acquire the l2ad_mtx. If that happens, we don't
3480 * want to re-destroy the header's L2 portion.
3482 if (HDR_HAS_L2HDR(hdr)) {
3484 arc_hdr_l2hdr_destroy(hdr);
3488 mutex_exit(&dev->l2ad_mtx);
3491 if (HDR_HAS_L1HDR(hdr)) {
3492 arc_cksum_free(hdr);
3494 while (hdr->b_l1hdr.b_buf != NULL)
3495 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3498 if (hdr->b_l1hdr.b_thawed != NULL) {
3499 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3500 hdr->b_l1hdr.b_thawed = NULL;
3504 if (hdr->b_l1hdr.b_pabd != NULL) {
3505 arc_hdr_free_pabd(hdr);
3509 ASSERT3P(hdr->b_hash_next, ==, NULL);
3510 if (HDR_HAS_L1HDR(hdr)) {
3511 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3512 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3513 kmem_cache_free(hdr_full_cache, hdr);
3515 kmem_cache_free(hdr_l2only_cache, hdr);
3520 arc_buf_destroy(arc_buf_t *buf, void* tag)
3522 arc_buf_hdr_t *hdr = buf->b_hdr;
3523 kmutex_t *hash_lock = HDR_LOCK(hdr);
3525 if (hdr->b_l1hdr.b_state == arc_anon) {
3526 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3527 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3528 VERIFY0(remove_reference(hdr, NULL, tag));
3529 arc_hdr_destroy(hdr);
3533 mutex_enter(hash_lock);
3534 ASSERT3P(hdr, ==, buf->b_hdr);
3535 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3536 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3537 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3538 ASSERT3P(buf->b_data, !=, NULL);
3540 (void) remove_reference(hdr, hash_lock, tag);
3541 arc_buf_destroy_impl(buf);
3542 mutex_exit(hash_lock);
3546 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3547 * state of the header is dependent on it's state prior to entering this
3548 * function. The following transitions are possible:
3550 * - arc_mru -> arc_mru_ghost
3551 * - arc_mfu -> arc_mfu_ghost
3552 * - arc_mru_ghost -> arc_l2c_only
3553 * - arc_mru_ghost -> deleted
3554 * - arc_mfu_ghost -> arc_l2c_only
3555 * - arc_mfu_ghost -> deleted
3558 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3560 arc_state_t *evicted_state, *state;
3561 int64_t bytes_evicted = 0;
3562 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3563 zfs_arc_min_prescient_prefetch_ms : zfs_arc_min_prefetch_ms;
3565 ASSERT(MUTEX_HELD(hash_lock));
3566 ASSERT(HDR_HAS_L1HDR(hdr));
3568 state = hdr->b_l1hdr.b_state;
3569 if (GHOST_STATE(state)) {
3570 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3571 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3574 * l2arc_write_buffers() relies on a header's L1 portion
3575 * (i.e. its b_pabd field) during it's write phase.
3576 * Thus, we cannot push a header onto the arc_l2c_only
3577 * state (removing it's L1 piece) until the header is
3578 * done being written to the l2arc.
3580 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3581 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3582 return (bytes_evicted);
3585 ARCSTAT_BUMP(arcstat_deleted);
3586 bytes_evicted += HDR_GET_LSIZE(hdr);
3588 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3590 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3591 if (HDR_HAS_L2HDR(hdr)) {
3593 * This buffer is cached on the 2nd Level ARC;
3594 * don't destroy the header.
3596 arc_change_state(arc_l2c_only, hdr, hash_lock);
3598 * dropping from L1+L2 cached to L2-only,
3599 * realloc to remove the L1 header.
3601 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3604 arc_change_state(arc_anon, hdr, hash_lock);
3605 arc_hdr_destroy(hdr);
3607 return (bytes_evicted);
3610 ASSERT(state == arc_mru || state == arc_mfu);
3611 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3613 /* prefetch buffers have a minimum lifespan */
3614 if (HDR_IO_IN_PROGRESS(hdr) ||
3615 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3616 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < min_lifetime * hz)) {
3617 ARCSTAT_BUMP(arcstat_evict_skip);
3618 return (bytes_evicted);
3621 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3622 while (hdr->b_l1hdr.b_buf) {
3623 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3624 if (!mutex_tryenter(&buf->b_evict_lock)) {
3625 ARCSTAT_BUMP(arcstat_mutex_miss);
3628 if (buf->b_data != NULL)
3629 bytes_evicted += HDR_GET_LSIZE(hdr);
3630 mutex_exit(&buf->b_evict_lock);
3631 arc_buf_destroy_impl(buf);
3634 if (HDR_HAS_L2HDR(hdr)) {
3635 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3637 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3638 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3639 HDR_GET_LSIZE(hdr));
3641 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3642 HDR_GET_LSIZE(hdr));
3646 if (hdr->b_l1hdr.b_bufcnt == 0) {
3647 arc_cksum_free(hdr);
3649 bytes_evicted += arc_hdr_size(hdr);
3652 * If this hdr is being evicted and has a compressed
3653 * buffer then we discard it here before we change states.
3654 * This ensures that the accounting is updated correctly
3655 * in arc_free_data_impl().
3657 arc_hdr_free_pabd(hdr);
3659 arc_change_state(evicted_state, hdr, hash_lock);
3660 ASSERT(HDR_IN_HASH_TABLE(hdr));
3661 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3662 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3665 return (bytes_evicted);
3669 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3670 uint64_t spa, int64_t bytes)
3672 multilist_sublist_t *mls;
3673 uint64_t bytes_evicted = 0;
3675 kmutex_t *hash_lock;
3676 int evict_count = 0;
3678 ASSERT3P(marker, !=, NULL);
3679 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3681 mls = multilist_sublist_lock(ml, idx);
3683 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3684 hdr = multilist_sublist_prev(mls, marker)) {
3685 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3686 (evict_count >= zfs_arc_evict_batch_limit))
3690 * To keep our iteration location, move the marker
3691 * forward. Since we're not holding hdr's hash lock, we
3692 * must be very careful and not remove 'hdr' from the
3693 * sublist. Otherwise, other consumers might mistake the
3694 * 'hdr' as not being on a sublist when they call the
3695 * multilist_link_active() function (they all rely on
3696 * the hash lock protecting concurrent insertions and
3697 * removals). multilist_sublist_move_forward() was
3698 * specifically implemented to ensure this is the case
3699 * (only 'marker' will be removed and re-inserted).
3701 multilist_sublist_move_forward(mls, marker);
3704 * The only case where the b_spa field should ever be
3705 * zero, is the marker headers inserted by
3706 * arc_evict_state(). It's possible for multiple threads
3707 * to be calling arc_evict_state() concurrently (e.g.
3708 * dsl_pool_close() and zio_inject_fault()), so we must
3709 * skip any markers we see from these other threads.
3711 if (hdr->b_spa == 0)
3714 /* we're only interested in evicting buffers of a certain spa */
3715 if (spa != 0 && hdr->b_spa != spa) {
3716 ARCSTAT_BUMP(arcstat_evict_skip);
3720 hash_lock = HDR_LOCK(hdr);
3723 * We aren't calling this function from any code path
3724 * that would already be holding a hash lock, so we're
3725 * asserting on this assumption to be defensive in case
3726 * this ever changes. Without this check, it would be
3727 * possible to incorrectly increment arcstat_mutex_miss
3728 * below (e.g. if the code changed such that we called
3729 * this function with a hash lock held).
3731 ASSERT(!MUTEX_HELD(hash_lock));
3733 if (mutex_tryenter(hash_lock)) {
3734 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3735 mutex_exit(hash_lock);
3737 bytes_evicted += evicted;
3740 * If evicted is zero, arc_evict_hdr() must have
3741 * decided to skip this header, don't increment
3742 * evict_count in this case.
3748 * If arc_size isn't overflowing, signal any
3749 * threads that might happen to be waiting.
3751 * For each header evicted, we wake up a single
3752 * thread. If we used cv_broadcast, we could
3753 * wake up "too many" threads causing arc_size
3754 * to significantly overflow arc_c; since
3755 * arc_get_data_impl() doesn't check for overflow
3756 * when it's woken up (it doesn't because it's
3757 * possible for the ARC to be overflowing while
3758 * full of un-evictable buffers, and the
3759 * function should proceed in this case).
3761 * If threads are left sleeping, due to not
3762 * using cv_broadcast, they will be woken up
3763 * just before arc_reclaim_thread() sleeps.
3765 mutex_enter(&arc_reclaim_lock);
3766 if (!arc_is_overflowing())
3767 cv_signal(&arc_reclaim_waiters_cv);
3768 mutex_exit(&arc_reclaim_lock);
3770 ARCSTAT_BUMP(arcstat_mutex_miss);
3774 multilist_sublist_unlock(mls);
3776 return (bytes_evicted);
3780 * Evict buffers from the given arc state, until we've removed the
3781 * specified number of bytes. Move the removed buffers to the
3782 * appropriate evict state.
3784 * This function makes a "best effort". It skips over any buffers
3785 * it can't get a hash_lock on, and so, may not catch all candidates.
3786 * It may also return without evicting as much space as requested.
3788 * If bytes is specified using the special value ARC_EVICT_ALL, this
3789 * will evict all available (i.e. unlocked and evictable) buffers from
3790 * the given arc state; which is used by arc_flush().
3793 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3794 arc_buf_contents_t type)
3796 uint64_t total_evicted = 0;
3797 multilist_t *ml = state->arcs_list[type];
3799 arc_buf_hdr_t **markers;
3801 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3803 num_sublists = multilist_get_num_sublists(ml);
3806 * If we've tried to evict from each sublist, made some
3807 * progress, but still have not hit the target number of bytes
3808 * to evict, we want to keep trying. The markers allow us to
3809 * pick up where we left off for each individual sublist, rather
3810 * than starting from the tail each time.
3812 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3813 for (int i = 0; i < num_sublists; i++) {
3814 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3817 * A b_spa of 0 is used to indicate that this header is
3818 * a marker. This fact is used in arc_adjust_type() and
3819 * arc_evict_state_impl().
3821 markers[i]->b_spa = 0;
3823 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3824 multilist_sublist_insert_tail(mls, markers[i]);
3825 multilist_sublist_unlock(mls);
3829 * While we haven't hit our target number of bytes to evict, or
3830 * we're evicting all available buffers.
3832 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3834 * Start eviction using a randomly selected sublist,
3835 * this is to try and evenly balance eviction across all
3836 * sublists. Always starting at the same sublist
3837 * (e.g. index 0) would cause evictions to favor certain
3838 * sublists over others.
3840 int sublist_idx = multilist_get_random_index(ml);
3841 uint64_t scan_evicted = 0;
3843 for (int i = 0; i < num_sublists; i++) {
3844 uint64_t bytes_remaining;
3845 uint64_t bytes_evicted;
3847 if (bytes == ARC_EVICT_ALL)
3848 bytes_remaining = ARC_EVICT_ALL;
3849 else if (total_evicted < bytes)
3850 bytes_remaining = bytes - total_evicted;
3854 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3855 markers[sublist_idx], spa, bytes_remaining);
3857 scan_evicted += bytes_evicted;
3858 total_evicted += bytes_evicted;
3860 /* we've reached the end, wrap to the beginning */
3861 if (++sublist_idx >= num_sublists)
3866 * If we didn't evict anything during this scan, we have
3867 * no reason to believe we'll evict more during another
3868 * scan, so break the loop.
3870 if (scan_evicted == 0) {
3871 /* This isn't possible, let's make that obvious */
3872 ASSERT3S(bytes, !=, 0);
3875 * When bytes is ARC_EVICT_ALL, the only way to
3876 * break the loop is when scan_evicted is zero.
3877 * In that case, we actually have evicted enough,
3878 * so we don't want to increment the kstat.
3880 if (bytes != ARC_EVICT_ALL) {
3881 ASSERT3S(total_evicted, <, bytes);
3882 ARCSTAT_BUMP(arcstat_evict_not_enough);
3889 for (int i = 0; i < num_sublists; i++) {
3890 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3891 multilist_sublist_remove(mls, markers[i]);
3892 multilist_sublist_unlock(mls);
3894 kmem_cache_free(hdr_full_cache, markers[i]);
3896 kmem_free(markers, sizeof (*markers) * num_sublists);
3898 return (total_evicted);
3902 * Flush all "evictable" data of the given type from the arc state
3903 * specified. This will not evict any "active" buffers (i.e. referenced).
3905 * When 'retry' is set to B_FALSE, the function will make a single pass
3906 * over the state and evict any buffers that it can. Since it doesn't
3907 * continually retry the eviction, it might end up leaving some buffers
3908 * in the ARC due to lock misses.
3910 * When 'retry' is set to B_TRUE, the function will continually retry the
3911 * eviction until *all* evictable buffers have been removed from the
3912 * state. As a result, if concurrent insertions into the state are
3913 * allowed (e.g. if the ARC isn't shutting down), this function might
3914 * wind up in an infinite loop, continually trying to evict buffers.
3917 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3920 uint64_t evicted = 0;
3922 while (refcount_count(&state->arcs_esize[type]) != 0) {
3923 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3933 * Evict the specified number of bytes from the state specified,
3934 * restricting eviction to the spa and type given. This function
3935 * prevents us from trying to evict more from a state's list than
3936 * is "evictable", and to skip evicting altogether when passed a
3937 * negative value for "bytes". In contrast, arc_evict_state() will
3938 * evict everything it can, when passed a negative value for "bytes".
3941 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3942 arc_buf_contents_t type)
3946 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3947 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3948 return (arc_evict_state(state, spa, delta, type));
3955 * Evict metadata buffers from the cache, such that arc_meta_used is
3956 * capped by the arc_meta_limit tunable.
3959 arc_adjust_meta(uint64_t meta_used)
3961 uint64_t total_evicted = 0;
3965 * If we're over the meta limit, we want to evict enough
3966 * metadata to get back under the meta limit. We don't want to
3967 * evict so much that we drop the MRU below arc_p, though. If
3968 * we're over the meta limit more than we're over arc_p, we
3969 * evict some from the MRU here, and some from the MFU below.
3971 target = MIN((int64_t)(meta_used - arc_meta_limit),
3972 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3973 refcount_count(&arc_mru->arcs_size) - arc_p));
3975 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3978 * Similar to the above, we want to evict enough bytes to get us
3979 * below the meta limit, but not so much as to drop us below the
3980 * space allotted to the MFU (which is defined as arc_c - arc_p).
3982 target = MIN((int64_t)(meta_used - arc_meta_limit),
3983 (int64_t)(refcount_count(&arc_mfu->arcs_size) -
3986 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3988 return (total_evicted);
3992 * Return the type of the oldest buffer in the given arc state
3994 * This function will select a random sublist of type ARC_BUFC_DATA and
3995 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3996 * is compared, and the type which contains the "older" buffer will be
3999 static arc_buf_contents_t
4000 arc_adjust_type(arc_state_t *state)
4002 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4003 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4004 int data_idx = multilist_get_random_index(data_ml);
4005 int meta_idx = multilist_get_random_index(meta_ml);
4006 multilist_sublist_t *data_mls;
4007 multilist_sublist_t *meta_mls;
4008 arc_buf_contents_t type;
4009 arc_buf_hdr_t *data_hdr;
4010 arc_buf_hdr_t *meta_hdr;
4013 * We keep the sublist lock until we're finished, to prevent
4014 * the headers from being destroyed via arc_evict_state().
4016 data_mls = multilist_sublist_lock(data_ml, data_idx);
4017 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4020 * These two loops are to ensure we skip any markers that
4021 * might be at the tail of the lists due to arc_evict_state().
4024 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4025 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4026 if (data_hdr->b_spa != 0)
4030 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4031 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4032 if (meta_hdr->b_spa != 0)
4036 if (data_hdr == NULL && meta_hdr == NULL) {
4037 type = ARC_BUFC_DATA;
4038 } else if (data_hdr == NULL) {
4039 ASSERT3P(meta_hdr, !=, NULL);
4040 type = ARC_BUFC_METADATA;
4041 } else if (meta_hdr == NULL) {
4042 ASSERT3P(data_hdr, !=, NULL);
4043 type = ARC_BUFC_DATA;
4045 ASSERT3P(data_hdr, !=, NULL);
4046 ASSERT3P(meta_hdr, !=, NULL);
4048 /* The headers can't be on the sublist without an L1 header */
4049 ASSERT(HDR_HAS_L1HDR(data_hdr));
4050 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4052 if (data_hdr->b_l1hdr.b_arc_access <
4053 meta_hdr->b_l1hdr.b_arc_access) {
4054 type = ARC_BUFC_DATA;
4056 type = ARC_BUFC_METADATA;
4060 multilist_sublist_unlock(meta_mls);
4061 multilist_sublist_unlock(data_mls);
4067 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4072 uint64_t total_evicted = 0;
4075 uint64_t asize = aggsum_value(&arc_size);
4076 uint64_t ameta = aggsum_value(&arc_meta_used);
4079 * If we're over arc_meta_limit, we want to correct that before
4080 * potentially evicting data buffers below.
4082 total_evicted += arc_adjust_meta(ameta);
4087 * If we're over the target cache size, we want to evict enough
4088 * from the list to get back to our target size. We don't want
4089 * to evict too much from the MRU, such that it drops below
4090 * arc_p. So, if we're over our target cache size more than
4091 * the MRU is over arc_p, we'll evict enough to get back to
4092 * arc_p here, and then evict more from the MFU below.
4094 target = MIN((int64_t)(asize - arc_c),
4095 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4096 refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4099 * If we're below arc_meta_min, always prefer to evict data.
4100 * Otherwise, try to satisfy the requested number of bytes to
4101 * evict from the type which contains older buffers; in an
4102 * effort to keep newer buffers in the cache regardless of their
4103 * type. If we cannot satisfy the number of bytes from this
4104 * type, spill over into the next type.
4106 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4107 ameta > arc_meta_min) {
4108 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4109 total_evicted += bytes;
4112 * If we couldn't evict our target number of bytes from
4113 * metadata, we try to get the rest from data.
4118 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4120 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4121 total_evicted += bytes;
4124 * If we couldn't evict our target number of bytes from
4125 * data, we try to get the rest from metadata.
4130 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4134 * Re-sum ARC stats after the first round of evictions.
4136 asize = aggsum_value(&arc_size);
4137 ameta = aggsum_value(&arc_meta_used);
4142 * Now that we've tried to evict enough from the MRU to get its
4143 * size back to arc_p, if we're still above the target cache
4144 * size, we evict the rest from the MFU.
4146 target = asize - arc_c;
4148 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4149 ameta > arc_meta_min) {
4150 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4151 total_evicted += bytes;
4154 * If we couldn't evict our target number of bytes from
4155 * metadata, we try to get the rest from data.
4160 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4162 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4163 total_evicted += bytes;
4166 * If we couldn't evict our target number of bytes from
4167 * data, we try to get the rest from data.
4172 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4176 * Adjust ghost lists
4178 * In addition to the above, the ARC also defines target values
4179 * for the ghost lists. The sum of the mru list and mru ghost
4180 * list should never exceed the target size of the cache, and
4181 * the sum of the mru list, mfu list, mru ghost list, and mfu
4182 * ghost list should never exceed twice the target size of the
4183 * cache. The following logic enforces these limits on the ghost
4184 * caches, and evicts from them as needed.
4186 target = refcount_count(&arc_mru->arcs_size) +
4187 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4189 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4190 total_evicted += bytes;
4195 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4198 * We assume the sum of the mru list and mfu list is less than
4199 * or equal to arc_c (we enforced this above), which means we
4200 * can use the simpler of the two equations below:
4202 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4203 * mru ghost + mfu ghost <= arc_c
4205 target = refcount_count(&arc_mru_ghost->arcs_size) +
4206 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4208 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4209 total_evicted += bytes;
4214 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4216 return (total_evicted);
4220 arc_flush(spa_t *spa, boolean_t retry)
4225 * If retry is B_TRUE, a spa must not be specified since we have
4226 * no good way to determine if all of a spa's buffers have been
4227 * evicted from an arc state.
4229 ASSERT(!retry || spa == 0);
4232 guid = spa_load_guid(spa);
4234 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4235 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4237 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4238 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4240 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4241 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4243 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4244 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4248 arc_shrink(int64_t to_free)
4250 uint64_t asize = aggsum_value(&arc_size);
4251 if (arc_c > arc_c_min) {
4252 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
4253 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
4254 if (arc_c > arc_c_min + to_free)
4255 atomic_add_64(&arc_c, -to_free);
4259 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4261 arc_c = MAX(asize, arc_c_min);
4263 arc_p = (arc_c >> 1);
4265 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
4268 ASSERT(arc_c >= arc_c_min);
4269 ASSERT((int64_t)arc_p >= 0);
4272 if (asize > arc_c) {
4273 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, asize,
4275 (void) arc_adjust();
4279 typedef enum free_memory_reason_t {
4284 FMR_PAGES_PP_MAXIMUM,
4288 } free_memory_reason_t;
4290 int64_t last_free_memory;
4291 free_memory_reason_t last_free_reason;
4294 * Additional reserve of pages for pp_reserve.
4296 int64_t arc_pages_pp_reserve = 64;
4299 * Additional reserve of pages for swapfs.
4301 int64_t arc_swapfs_reserve = 64;
4304 * Return the amount of memory that can be consumed before reclaim will be
4305 * needed. Positive if there is sufficient free memory, negative indicates
4306 * the amount of memory that needs to be freed up.
4309 arc_available_memory(void)
4311 int64_t lowest = INT64_MAX;
4313 free_memory_reason_t r = FMR_UNKNOWN;
4318 * Cooperate with pagedaemon when it's time for it to scan
4319 * and reclaim some pages.
4321 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
4329 n = PAGESIZE * (-needfree);
4337 * check that we're out of range of the pageout scanner. It starts to
4338 * schedule paging if freemem is less than lotsfree and needfree.
4339 * lotsfree is the high-water mark for pageout, and needfree is the
4340 * number of needed free pages. We add extra pages here to make sure
4341 * the scanner doesn't start up while we're freeing memory.
4343 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4350 * check to make sure that swapfs has enough space so that anon
4351 * reservations can still succeed. anon_resvmem() checks that the
4352 * availrmem is greater than swapfs_minfree, and the number of reserved
4353 * swap pages. We also add a bit of extra here just to prevent
4354 * circumstances from getting really dire.
4356 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4357 desfree - arc_swapfs_reserve);
4360 r = FMR_SWAPFS_MINFREE;
4365 * Check that we have enough availrmem that memory locking (e.g., via
4366 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4367 * stores the number of pages that cannot be locked; when availrmem
4368 * drops below pages_pp_maximum, page locking mechanisms such as
4369 * page_pp_lock() will fail.)
4371 n = PAGESIZE * (availrmem - pages_pp_maximum -
4372 arc_pages_pp_reserve);
4375 r = FMR_PAGES_PP_MAXIMUM;
4378 #endif /* __FreeBSD__ */
4379 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4381 * If we're on an i386 platform, it's possible that we'll exhaust the
4382 * kernel heap space before we ever run out of available physical
4383 * memory. Most checks of the size of the heap_area compare against
4384 * tune.t_minarmem, which is the minimum available real memory that we
4385 * can have in the system. However, this is generally fixed at 25 pages
4386 * which is so low that it's useless. In this comparison, we seek to
4387 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4388 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4391 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
4392 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
4397 #define zio_arena NULL
4399 #define zio_arena heap_arena
4403 * If zio data pages are being allocated out of a separate heap segment,
4404 * then enforce that the size of available vmem for this arena remains
4405 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4407 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4408 * memory (in the zio_arena) free, which can avoid memory
4409 * fragmentation issues.
4411 if (zio_arena != NULL) {
4412 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4413 (vmem_size(zio_arena, VMEM_ALLOC) >>
4414 arc_zio_arena_free_shift);
4422 * Above limits know nothing about real level of KVA fragmentation.
4423 * Start aggressive reclamation if too little sequential KVA left.
4426 n = (vmem_size(heap_arena, VMEM_MAXFREE) < SPA_MAXBLOCKSIZE) ?
4427 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) :
4436 /* Every 100 calls, free a small amount */
4437 if (spa_get_random(100) == 0)
4439 #endif /* _KERNEL */
4441 last_free_memory = lowest;
4442 last_free_reason = r;
4443 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
4449 * Determine if the system is under memory pressure and is asking
4450 * to reclaim memory. A return value of B_TRUE indicates that the system
4451 * is under memory pressure and that the arc should adjust accordingly.
4454 arc_reclaim_needed(void)
4456 return (arc_available_memory() < 0);
4459 extern kmem_cache_t *zio_buf_cache[];
4460 extern kmem_cache_t *zio_data_buf_cache[];
4461 extern kmem_cache_t *range_seg_cache;
4462 extern kmem_cache_t *abd_chunk_cache;
4464 static __noinline void
4465 arc_kmem_reap_now(void)
4468 kmem_cache_t *prev_cache = NULL;
4469 kmem_cache_t *prev_data_cache = NULL;
4471 DTRACE_PROBE(arc__kmem_reap_start);
4473 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4475 * We are exceeding our meta-data cache limit.
4476 * Purge some DNLC entries to release holds on meta-data.
4478 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4482 * Reclaim unused memory from all kmem caches.
4489 * If a kmem reap is already active, don't schedule more. We must
4490 * check for this because kmem_cache_reap_soon() won't actually
4491 * block on the cache being reaped (this is to prevent callers from
4492 * becoming implicitly blocked by a system-wide kmem reap -- which,
4493 * on a system with many, many full magazines, can take minutes).
4495 if (kmem_cache_reap_active())
4498 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4499 if (zio_buf_cache[i] != prev_cache) {
4500 prev_cache = zio_buf_cache[i];
4501 kmem_cache_reap_soon(zio_buf_cache[i]);
4503 if (zio_data_buf_cache[i] != prev_data_cache) {
4504 prev_data_cache = zio_data_buf_cache[i];
4505 kmem_cache_reap_soon(zio_data_buf_cache[i]);
4508 kmem_cache_reap_soon(abd_chunk_cache);
4509 kmem_cache_reap_soon(buf_cache);
4510 kmem_cache_reap_soon(hdr_full_cache);
4511 kmem_cache_reap_soon(hdr_l2only_cache);
4512 kmem_cache_reap_soon(range_seg_cache);
4515 if (zio_arena != NULL) {
4517 * Ask the vmem arena to reclaim unused memory from its
4520 vmem_qcache_reap(zio_arena);
4523 DTRACE_PROBE(arc__kmem_reap_end);
4527 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4528 * enough data and signal them to proceed. When this happens, the threads in
4529 * arc_get_data_impl() are sleeping while holding the hash lock for their
4530 * particular arc header. Thus, we must be careful to never sleep on a
4531 * hash lock in this thread. This is to prevent the following deadlock:
4533 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4534 * waiting for the reclaim thread to signal it.
4536 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4537 * fails, and goes to sleep forever.
4539 * This possible deadlock is avoided by always acquiring a hash lock
4540 * using mutex_tryenter() from arc_reclaim_thread().
4544 arc_reclaim_thread(void *unused __unused)
4546 hrtime_t growtime = 0;
4547 hrtime_t kmem_reap_time = 0;
4550 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4552 mutex_enter(&arc_reclaim_lock);
4553 while (!arc_reclaim_thread_exit) {
4554 uint64_t evicted = 0;
4557 * This is necessary in order for the mdb ::arc dcmd to
4558 * show up to date information. Since the ::arc command
4559 * does not call the kstat's update function, without
4560 * this call, the command may show stale stats for the
4561 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4562 * with this change, the data might be up to 1 second
4563 * out of date; but that should suffice. The arc_state_t
4564 * structures can be queried directly if more accurate
4565 * information is needed.
4567 if (arc_ksp != NULL)
4568 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4570 mutex_exit(&arc_reclaim_lock);
4573 * We call arc_adjust() before (possibly) calling
4574 * arc_kmem_reap_now(), so that we can wake up
4575 * arc_get_data_impl() sooner.
4577 evicted = arc_adjust();
4579 int64_t free_memory = arc_available_memory();
4580 if (free_memory < 0) {
4581 hrtime_t curtime = gethrtime();
4582 arc_no_grow = B_TRUE;
4586 * Wait at least zfs_grow_retry (default 60) seconds
4587 * before considering growing.
4589 growtime = curtime + SEC2NSEC(arc_grow_retry);
4592 * Wait at least arc_kmem_cache_reap_retry_ms
4593 * between arc_kmem_reap_now() calls. Without
4594 * this check it is possible to end up in a
4595 * situation where we spend lots of time
4596 * reaping caches, while we're near arc_c_min.
4598 if (curtime >= kmem_reap_time) {
4599 arc_kmem_reap_now();
4600 kmem_reap_time = gethrtime() +
4601 MSEC2NSEC(arc_kmem_cache_reap_retry_ms);
4605 * If we are still low on memory, shrink the ARC
4606 * so that we have arc_shrink_min free space.
4608 free_memory = arc_available_memory();
4611 (arc_c >> arc_shrink_shift) - free_memory;
4615 to_free = MAX(to_free, ptob(needfree));
4618 arc_shrink(to_free);
4620 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4621 arc_no_grow = B_TRUE;
4622 } else if (gethrtime() >= growtime) {
4623 arc_no_grow = B_FALSE;
4626 mutex_enter(&arc_reclaim_lock);
4629 * If evicted is zero, we couldn't evict anything via
4630 * arc_adjust(). This could be due to hash lock
4631 * collisions, but more likely due to the majority of
4632 * arc buffers being unevictable. Therefore, even if
4633 * arc_size is above arc_c, another pass is unlikely to
4634 * be helpful and could potentially cause us to enter an
4637 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) {
4639 * We're either no longer overflowing, or we
4640 * can't evict anything more, so we should wake
4641 * up any threads before we go to sleep.
4643 cv_broadcast(&arc_reclaim_waiters_cv);
4646 * Block until signaled, or after one second (we
4647 * might need to perform arc_kmem_reap_now()
4648 * even if we aren't being signalled)
4650 CALLB_CPR_SAFE_BEGIN(&cpr);
4651 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4652 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4653 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4657 arc_reclaim_thread_exit = B_FALSE;
4658 cv_broadcast(&arc_reclaim_thread_cv);
4659 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4663 static u_int arc_dnlc_evicts_arg;
4664 extern struct vfsops zfs_vfsops;
4667 arc_dnlc_evicts_thread(void *dummy __unused)
4672 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
4674 mutex_enter(&arc_dnlc_evicts_lock);
4675 while (!arc_dnlc_evicts_thread_exit) {
4676 CALLB_CPR_SAFE_BEGIN(&cpr);
4677 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
4678 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
4679 if (arc_dnlc_evicts_arg != 0) {
4680 percent = arc_dnlc_evicts_arg;
4681 mutex_exit(&arc_dnlc_evicts_lock);
4683 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
4685 mutex_enter(&arc_dnlc_evicts_lock);
4687 * Clear our token only after vnlru_free()
4688 * pass is done, to avoid false queueing of
4691 arc_dnlc_evicts_arg = 0;
4694 arc_dnlc_evicts_thread_exit = FALSE;
4695 cv_broadcast(&arc_dnlc_evicts_cv);
4696 CALLB_CPR_EXIT(&cpr);
4701 dnlc_reduce_cache(void *arg)
4705 percent = (u_int)(uintptr_t)arg;
4706 mutex_enter(&arc_dnlc_evicts_lock);
4707 if (arc_dnlc_evicts_arg == 0) {
4708 arc_dnlc_evicts_arg = percent;
4709 cv_broadcast(&arc_dnlc_evicts_cv);
4711 mutex_exit(&arc_dnlc_evicts_lock);
4715 * Adapt arc info given the number of bytes we are trying to add and
4716 * the state that we are comming from. This function is only called
4717 * when we are adding new content to the cache.
4720 arc_adapt(int bytes, arc_state_t *state)
4723 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4724 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4725 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4727 if (state == arc_l2c_only)
4732 * Adapt the target size of the MRU list:
4733 * - if we just hit in the MRU ghost list, then increase
4734 * the target size of the MRU list.
4735 * - if we just hit in the MFU ghost list, then increase
4736 * the target size of the MFU list by decreasing the
4737 * target size of the MRU list.
4739 if (state == arc_mru_ghost) {
4740 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4741 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4743 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4744 } else if (state == arc_mfu_ghost) {
4747 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4748 mult = MIN(mult, 10);
4750 delta = MIN(bytes * mult, arc_p);
4751 arc_p = MAX(arc_p_min, arc_p - delta);
4753 ASSERT((int64_t)arc_p >= 0);
4755 if (arc_reclaim_needed()) {
4756 cv_signal(&arc_reclaim_thread_cv);
4763 if (arc_c >= arc_c_max)
4767 * If we're within (2 * maxblocksize) bytes of the target
4768 * cache size, increment the target cache size
4770 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
4772 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
4773 atomic_add_64(&arc_c, (int64_t)bytes);
4774 if (arc_c > arc_c_max)
4776 else if (state == arc_anon)
4777 atomic_add_64(&arc_p, (int64_t)bytes);
4781 ASSERT((int64_t)arc_p >= 0);
4785 * Check if arc_size has grown past our upper threshold, determined by
4786 * zfs_arc_overflow_shift.
4789 arc_is_overflowing(void)
4791 /* Always allow at least one block of overflow */
4792 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4793 arc_c >> zfs_arc_overflow_shift);
4796 * We just compare the lower bound here for performance reasons. Our
4797 * primary goals are to make sure that the arc never grows without
4798 * bound, and that it can reach its maximum size. This check
4799 * accomplishes both goals. The maximum amount we could run over by is
4800 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4801 * in the ARC. In practice, that's in the tens of MB, which is low
4802 * enough to be safe.
4804 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
4808 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4810 arc_buf_contents_t type = arc_buf_type(hdr);
4812 arc_get_data_impl(hdr, size, tag);
4813 if (type == ARC_BUFC_METADATA) {
4814 return (abd_alloc(size, B_TRUE));
4816 ASSERT(type == ARC_BUFC_DATA);
4817 return (abd_alloc(size, B_FALSE));
4822 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4824 arc_buf_contents_t type = arc_buf_type(hdr);
4826 arc_get_data_impl(hdr, size, tag);
4827 if (type == ARC_BUFC_METADATA) {
4828 return (zio_buf_alloc(size));
4830 ASSERT(type == ARC_BUFC_DATA);
4831 return (zio_data_buf_alloc(size));
4836 * Allocate a block and return it to the caller. If we are hitting the
4837 * hard limit for the cache size, we must sleep, waiting for the eviction
4838 * thread to catch up. If we're past the target size but below the hard
4839 * limit, we'll only signal the reclaim thread and continue on.
4842 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4844 arc_state_t *state = hdr->b_l1hdr.b_state;
4845 arc_buf_contents_t type = arc_buf_type(hdr);
4847 arc_adapt(size, state);
4850 * If arc_size is currently overflowing, and has grown past our
4851 * upper limit, we must be adding data faster than the evict
4852 * thread can evict. Thus, to ensure we don't compound the
4853 * problem by adding more data and forcing arc_size to grow even
4854 * further past it's target size, we halt and wait for the
4855 * eviction thread to catch up.
4857 * It's also possible that the reclaim thread is unable to evict
4858 * enough buffers to get arc_size below the overflow limit (e.g.
4859 * due to buffers being un-evictable, or hash lock collisions).
4860 * In this case, we want to proceed regardless if we're
4861 * overflowing; thus we don't use a while loop here.
4863 if (arc_is_overflowing()) {
4864 mutex_enter(&arc_reclaim_lock);
4867 * Now that we've acquired the lock, we may no longer be
4868 * over the overflow limit, lets check.
4870 * We're ignoring the case of spurious wake ups. If that
4871 * were to happen, it'd let this thread consume an ARC
4872 * buffer before it should have (i.e. before we're under
4873 * the overflow limit and were signalled by the reclaim
4874 * thread). As long as that is a rare occurrence, it
4875 * shouldn't cause any harm.
4877 if (arc_is_overflowing()) {
4878 cv_signal(&arc_reclaim_thread_cv);
4879 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4882 mutex_exit(&arc_reclaim_lock);
4885 VERIFY3U(hdr->b_type, ==, type);
4886 if (type == ARC_BUFC_METADATA) {
4887 arc_space_consume(size, ARC_SPACE_META);
4889 arc_space_consume(size, ARC_SPACE_DATA);
4893 * Update the state size. Note that ghost states have a
4894 * "ghost size" and so don't need to be updated.
4896 if (!GHOST_STATE(state)) {
4898 (void) refcount_add_many(&state->arcs_size, size, tag);
4901 * If this is reached via arc_read, the link is
4902 * protected by the hash lock. If reached via
4903 * arc_buf_alloc, the header should not be accessed by
4904 * any other thread. And, if reached via arc_read_done,
4905 * the hash lock will protect it if it's found in the
4906 * hash table; otherwise no other thread should be
4907 * trying to [add|remove]_reference it.
4909 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4910 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4911 (void) refcount_add_many(&state->arcs_esize[type],
4916 * If we are growing the cache, and we are adding anonymous
4917 * data, and we have outgrown arc_p, update arc_p
4919 if (aggsum_compare(&arc_size, arc_c) < 0 &&
4920 hdr->b_l1hdr.b_state == arc_anon &&
4921 (refcount_count(&arc_anon->arcs_size) +
4922 refcount_count(&arc_mru->arcs_size) > arc_p))
4923 arc_p = MIN(arc_c, arc_p + size);
4925 ARCSTAT_BUMP(arcstat_allocated);
4929 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4931 arc_free_data_impl(hdr, size, tag);
4936 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4938 arc_buf_contents_t type = arc_buf_type(hdr);
4940 arc_free_data_impl(hdr, size, tag);
4941 if (type == ARC_BUFC_METADATA) {
4942 zio_buf_free(buf, size);
4944 ASSERT(type == ARC_BUFC_DATA);
4945 zio_data_buf_free(buf, size);
4950 * Free the arc data buffer.
4953 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4955 arc_state_t *state = hdr->b_l1hdr.b_state;
4956 arc_buf_contents_t type = arc_buf_type(hdr);
4958 /* protected by hash lock, if in the hash table */
4959 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4960 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4961 ASSERT(state != arc_anon && state != arc_l2c_only);
4963 (void) refcount_remove_many(&state->arcs_esize[type],
4966 (void) refcount_remove_many(&state->arcs_size, size, tag);
4968 VERIFY3U(hdr->b_type, ==, type);
4969 if (type == ARC_BUFC_METADATA) {
4970 arc_space_return(size, ARC_SPACE_META);
4972 ASSERT(type == ARC_BUFC_DATA);
4973 arc_space_return(size, ARC_SPACE_DATA);
4978 * This routine is called whenever a buffer is accessed.
4979 * NOTE: the hash lock is dropped in this function.
4982 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4986 ASSERT(MUTEX_HELD(hash_lock));
4987 ASSERT(HDR_HAS_L1HDR(hdr));
4989 if (hdr->b_l1hdr.b_state == arc_anon) {
4991 * This buffer is not in the cache, and does not
4992 * appear in our "ghost" list. Add the new buffer
4996 ASSERT0(hdr->b_l1hdr.b_arc_access);
4997 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4998 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4999 arc_change_state(arc_mru, hdr, hash_lock);
5001 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5002 now = ddi_get_lbolt();
5005 * If this buffer is here because of a prefetch, then either:
5006 * - clear the flag if this is a "referencing" read
5007 * (any subsequent access will bump this into the MFU state).
5009 * - move the buffer to the head of the list if this is
5010 * another prefetch (to make it less likely to be evicted).
5012 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5013 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5014 /* link protected by hash lock */
5015 ASSERT(multilist_link_active(
5016 &hdr->b_l1hdr.b_arc_node));
5018 arc_hdr_clear_flags(hdr,
5020 ARC_FLAG_PRESCIENT_PREFETCH);
5021 ARCSTAT_BUMP(arcstat_mru_hits);
5023 hdr->b_l1hdr.b_arc_access = now;
5028 * This buffer has been "accessed" only once so far,
5029 * but it is still in the cache. Move it to the MFU
5032 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
5034 * More than 125ms have passed since we
5035 * instantiated this buffer. Move it to the
5036 * most frequently used state.
5038 hdr->b_l1hdr.b_arc_access = now;
5039 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5040 arc_change_state(arc_mfu, hdr, hash_lock);
5042 ARCSTAT_BUMP(arcstat_mru_hits);
5043 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5044 arc_state_t *new_state;
5046 * This buffer has been "accessed" recently, but
5047 * was evicted from the cache. Move it to the
5051 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5052 new_state = arc_mru;
5053 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5054 arc_hdr_clear_flags(hdr,
5056 ARC_FLAG_PRESCIENT_PREFETCH);
5058 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5060 new_state = arc_mfu;
5061 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5064 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5065 arc_change_state(new_state, hdr, hash_lock);
5067 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5068 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5070 * This buffer has been accessed more than once and is
5071 * still in the cache. Keep it in the MFU state.
5073 * NOTE: an add_reference() that occurred when we did
5074 * the arc_read() will have kicked this off the list.
5075 * If it was a prefetch, we will explicitly move it to
5076 * the head of the list now.
5079 ARCSTAT_BUMP(arcstat_mfu_hits);
5080 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5081 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5082 arc_state_t *new_state = arc_mfu;
5084 * This buffer has been accessed more than once but has
5085 * been evicted from the cache. Move it back to the
5089 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5091 * This is a prefetch access...
5092 * move this block back to the MRU state.
5094 new_state = arc_mru;
5097 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5098 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5099 arc_change_state(new_state, hdr, hash_lock);
5101 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5102 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5104 * This buffer is on the 2nd Level ARC.
5107 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5108 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5109 arc_change_state(arc_mfu, hdr, hash_lock);
5111 ASSERT(!"invalid arc state");
5116 * This routine is called by dbuf_hold() to update the arc_access() state
5117 * which otherwise would be skipped for entries in the dbuf cache.
5120 arc_buf_access(arc_buf_t *buf)
5122 mutex_enter(&buf->b_evict_lock);
5123 arc_buf_hdr_t *hdr = buf->b_hdr;
5126 * Avoid taking the hash_lock when possible as an optimization.
5127 * The header must be checked again under the hash_lock in order
5128 * to handle the case where it is concurrently being released.
5130 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5131 mutex_exit(&buf->b_evict_lock);
5132 ARCSTAT_BUMP(arcstat_access_skip);
5136 kmutex_t *hash_lock = HDR_LOCK(hdr);
5137 mutex_enter(hash_lock);
5139 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5140 mutex_exit(hash_lock);
5141 mutex_exit(&buf->b_evict_lock);
5142 ARCSTAT_BUMP(arcstat_access_skip);
5146 mutex_exit(&buf->b_evict_lock);
5148 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5149 hdr->b_l1hdr.b_state == arc_mfu);
5151 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5152 arc_access(hdr, hash_lock);
5153 mutex_exit(hash_lock);
5155 ARCSTAT_BUMP(arcstat_hits);
5156 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5157 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5160 /* a generic arc_read_done_func_t which you can use */
5163 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5164 arc_buf_t *buf, void *arg)
5169 bcopy(buf->b_data, arg, arc_buf_size(buf));
5170 arc_buf_destroy(buf, arg);
5173 /* a generic arc_read_done_func_t */
5176 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5177 arc_buf_t *buf, void *arg)
5179 arc_buf_t **bufp = arg;
5185 ASSERT(buf->b_data);
5190 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5192 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5193 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5194 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
5196 if (HDR_COMPRESSION_ENABLED(hdr)) {
5197 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
5198 BP_GET_COMPRESS(bp));
5200 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5201 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5206 arc_read_done(zio_t *zio)
5208 arc_buf_hdr_t *hdr = zio->io_private;
5209 kmutex_t *hash_lock = NULL;
5210 arc_callback_t *callback_list;
5211 arc_callback_t *acb;
5212 boolean_t freeable = B_FALSE;
5215 * The hdr was inserted into hash-table and removed from lists
5216 * prior to starting I/O. We should find this header, since
5217 * it's in the hash table, and it should be legit since it's
5218 * not possible to evict it during the I/O. The only possible
5219 * reason for it not to be found is if we were freed during the
5222 if (HDR_IN_HASH_TABLE(hdr)) {
5223 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5224 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5225 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5226 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5227 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5229 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5232 ASSERT((found == hdr &&
5233 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5234 (found == hdr && HDR_L2_READING(hdr)));
5235 ASSERT3P(hash_lock, !=, NULL);
5238 if (zio->io_error == 0) {
5239 /* byteswap if necessary */
5240 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5241 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5242 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5244 hdr->b_l1hdr.b_byteswap =
5245 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5248 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5252 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5253 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5254 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5256 callback_list = hdr->b_l1hdr.b_acb;
5257 ASSERT3P(callback_list, !=, NULL);
5259 if (hash_lock && zio->io_error == 0 &&
5260 hdr->b_l1hdr.b_state == arc_anon) {
5262 * Only call arc_access on anonymous buffers. This is because
5263 * if we've issued an I/O for an evicted buffer, we've already
5264 * called arc_access (to prevent any simultaneous readers from
5265 * getting confused).
5267 arc_access(hdr, hash_lock);
5271 * If a read request has a callback (i.e. acb_done is not NULL), then we
5272 * make a buf containing the data according to the parameters which were
5273 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5274 * aren't needlessly decompressing the data multiple times.
5276 int callback_cnt = 0;
5277 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5283 if (zio->io_error != 0)
5286 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
5287 acb->acb_compressed,
5288 B_TRUE, &acb->acb_buf);
5290 arc_buf_destroy(acb->acb_buf, acb->acb_private);
5291 acb->acb_buf = NULL;
5294 if (zio->io_error == 0)
5295 zio->io_error = error;
5297 hdr->b_l1hdr.b_acb = NULL;
5298 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5299 if (callback_cnt == 0) {
5300 ASSERT(HDR_PREFETCH(hdr));
5301 ASSERT0(hdr->b_l1hdr.b_bufcnt);
5302 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5305 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5306 callback_list != NULL);
5308 if (zio->io_error == 0) {
5309 arc_hdr_verify(hdr, zio->io_bp);
5311 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5312 if (hdr->b_l1hdr.b_state != arc_anon)
5313 arc_change_state(arc_anon, hdr, hash_lock);
5314 if (HDR_IN_HASH_TABLE(hdr))
5315 buf_hash_remove(hdr);
5316 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5320 * Broadcast before we drop the hash_lock to avoid the possibility
5321 * that the hdr (and hence the cv) might be freed before we get to
5322 * the cv_broadcast().
5324 cv_broadcast(&hdr->b_l1hdr.b_cv);
5326 if (hash_lock != NULL) {
5327 mutex_exit(hash_lock);
5330 * This block was freed while we waited for the read to
5331 * complete. It has been removed from the hash table and
5332 * moved to the anonymous state (so that it won't show up
5335 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5336 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5339 /* execute each callback and free its structure */
5340 while ((acb = callback_list) != NULL) {
5341 if (acb->acb_done) {
5342 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5343 acb->acb_buf, acb->acb_private);
5346 if (acb->acb_zio_dummy != NULL) {
5347 acb->acb_zio_dummy->io_error = zio->io_error;
5348 zio_nowait(acb->acb_zio_dummy);
5351 callback_list = acb->acb_next;
5352 kmem_free(acb, sizeof (arc_callback_t));
5356 arc_hdr_destroy(hdr);
5360 * "Read" the block at the specified DVA (in bp) via the
5361 * cache. If the block is found in the cache, invoke the provided
5362 * callback immediately and return. Note that the `zio' parameter
5363 * in the callback will be NULL in this case, since no IO was
5364 * required. If the block is not in the cache pass the read request
5365 * on to the spa with a substitute callback function, so that the
5366 * requested block will be added to the cache.
5368 * If a read request arrives for a block that has a read in-progress,
5369 * either wait for the in-progress read to complete (and return the
5370 * results); or, if this is a read with a "done" func, add a record
5371 * to the read to invoke the "done" func when the read completes,
5372 * and return; or just return.
5374 * arc_read_done() will invoke all the requested "done" functions
5375 * for readers of this block.
5378 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_read_done_func_t *done,
5379 void *private, zio_priority_t priority, int zio_flags,
5380 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5382 arc_buf_hdr_t *hdr = NULL;
5383 kmutex_t *hash_lock = NULL;
5385 uint64_t guid = spa_load_guid(spa);
5386 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
5389 ASSERT(!BP_IS_EMBEDDED(bp) ||
5390 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5393 if (!BP_IS_EMBEDDED(bp)) {
5395 * Embedded BP's have no DVA and require no I/O to "read".
5396 * Create an anonymous arc buf to back it.
5398 hdr = buf_hash_find(guid, bp, &hash_lock);
5401 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
5402 arc_buf_t *buf = NULL;
5403 *arc_flags |= ARC_FLAG_CACHED;
5405 if (HDR_IO_IN_PROGRESS(hdr)) {
5406 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5408 ASSERT3P(head_zio, !=, NULL);
5409 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5410 priority == ZIO_PRIORITY_SYNC_READ) {
5412 * This is a sync read that needs to wait for
5413 * an in-flight async read. Request that the
5414 * zio have its priority upgraded.
5416 zio_change_priority(head_zio, priority);
5417 DTRACE_PROBE1(arc__async__upgrade__sync,
5418 arc_buf_hdr_t *, hdr);
5419 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5421 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5422 arc_hdr_clear_flags(hdr,
5423 ARC_FLAG_PREDICTIVE_PREFETCH);
5426 if (*arc_flags & ARC_FLAG_WAIT) {
5427 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5428 mutex_exit(hash_lock);
5431 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5434 arc_callback_t *acb = NULL;
5436 acb = kmem_zalloc(sizeof (arc_callback_t),
5438 acb->acb_done = done;
5439 acb->acb_private = private;
5440 acb->acb_compressed = compressed_read;
5442 acb->acb_zio_dummy = zio_null(pio,
5443 spa, NULL, NULL, NULL, zio_flags);
5445 ASSERT3P(acb->acb_done, !=, NULL);
5446 acb->acb_zio_head = head_zio;
5447 acb->acb_next = hdr->b_l1hdr.b_acb;
5448 hdr->b_l1hdr.b_acb = acb;
5449 mutex_exit(hash_lock);
5452 mutex_exit(hash_lock);
5456 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5457 hdr->b_l1hdr.b_state == arc_mfu);
5460 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5462 * This is a demand read which does not have to
5463 * wait for i/o because we did a predictive
5464 * prefetch i/o for it, which has completed.
5467 arc__demand__hit__predictive__prefetch,
5468 arc_buf_hdr_t *, hdr);
5470 arcstat_demand_hit_predictive_prefetch);
5471 arc_hdr_clear_flags(hdr,
5472 ARC_FLAG_PREDICTIVE_PREFETCH);
5475 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5477 arcstat_demand_hit_prescient_prefetch);
5478 arc_hdr_clear_flags(hdr,
5479 ARC_FLAG_PRESCIENT_PREFETCH);
5482 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5483 /* Get a buf with the desired data in it. */
5484 rc = arc_buf_alloc_impl(hdr, private,
5485 compressed_read, B_TRUE, &buf);
5487 arc_buf_destroy(buf, private);
5490 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5491 rc == 0 || rc != ENOENT);
5492 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5493 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5494 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5496 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5497 arc_access(hdr, hash_lock);
5498 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5499 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5500 if (*arc_flags & ARC_FLAG_L2CACHE)
5501 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5502 mutex_exit(hash_lock);
5503 ARCSTAT_BUMP(arcstat_hits);
5504 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5505 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5506 data, metadata, hits);
5509 done(NULL, zb, bp, buf, private);
5511 uint64_t lsize = BP_GET_LSIZE(bp);
5512 uint64_t psize = BP_GET_PSIZE(bp);
5513 arc_callback_t *acb;
5516 boolean_t devw = B_FALSE;
5520 /* this block is not in the cache */
5521 arc_buf_hdr_t *exists = NULL;
5522 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5523 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5524 BP_GET_COMPRESS(bp), type);
5526 if (!BP_IS_EMBEDDED(bp)) {
5527 hdr->b_dva = *BP_IDENTITY(bp);
5528 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5529 exists = buf_hash_insert(hdr, &hash_lock);
5531 if (exists != NULL) {
5532 /* somebody beat us to the hash insert */
5533 mutex_exit(hash_lock);
5534 buf_discard_identity(hdr);
5535 arc_hdr_destroy(hdr);
5536 goto top; /* restart the IO request */
5540 * This block is in the ghost cache. If it was L2-only
5541 * (and thus didn't have an L1 hdr), we realloc the
5542 * header to add an L1 hdr.
5544 if (!HDR_HAS_L1HDR(hdr)) {
5545 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5548 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5549 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5550 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5551 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5552 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5553 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5556 * This is a delicate dance that we play here.
5557 * This hdr is in the ghost list so we access it
5558 * to move it out of the ghost list before we
5559 * initiate the read. If it's a prefetch then
5560 * it won't have a callback so we'll remove the
5561 * reference that arc_buf_alloc_impl() created. We
5562 * do this after we've called arc_access() to
5563 * avoid hitting an assert in remove_reference().
5565 arc_access(hdr, hash_lock);
5566 arc_hdr_alloc_pabd(hdr);
5568 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5569 size = arc_hdr_size(hdr);
5572 * If compression is enabled on the hdr, then will do
5573 * RAW I/O and will store the compressed data in the hdr's
5574 * data block. Otherwise, the hdr's data block will contain
5575 * the uncompressed data.
5577 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5578 zio_flags |= ZIO_FLAG_RAW;
5581 if (*arc_flags & ARC_FLAG_PREFETCH)
5582 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5583 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5584 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5586 if (*arc_flags & ARC_FLAG_L2CACHE)
5587 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5588 if (BP_GET_LEVEL(bp) > 0)
5589 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5590 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5591 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5592 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5594 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5595 acb->acb_done = done;
5596 acb->acb_private = private;
5597 acb->acb_compressed = compressed_read;
5599 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5600 hdr->b_l1hdr.b_acb = acb;
5601 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5603 if (HDR_HAS_L2HDR(hdr) &&
5604 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5605 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5606 addr = hdr->b_l2hdr.b_daddr;
5608 * Lock out L2ARC device removal.
5610 if (vdev_is_dead(vd) ||
5611 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5616 * We count both async reads and scrub IOs as asynchronous so
5617 * that both can be upgraded in the event of a cache hit while
5618 * the read IO is still in-flight.
5620 if (priority == ZIO_PRIORITY_ASYNC_READ ||
5621 priority == ZIO_PRIORITY_SCRUB)
5622 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5624 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5627 * At this point, we have a level 1 cache miss. Try again in
5628 * L2ARC if possible.
5630 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5632 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5633 uint64_t, lsize, zbookmark_phys_t *, zb);
5634 ARCSTAT_BUMP(arcstat_misses);
5635 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5636 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5637 data, metadata, misses);
5642 racct_add_force(curproc, RACCT_READBPS, size);
5643 racct_add_force(curproc, RACCT_READIOPS, 1);
5644 PROC_UNLOCK(curproc);
5647 curthread->td_ru.ru_inblock++;
5650 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5652 * Read from the L2ARC if the following are true:
5653 * 1. The L2ARC vdev was previously cached.
5654 * 2. This buffer still has L2ARC metadata.
5655 * 3. This buffer isn't currently writing to the L2ARC.
5656 * 4. The L2ARC entry wasn't evicted, which may
5657 * also have invalidated the vdev.
5658 * 5. This isn't prefetch and l2arc_noprefetch is set.
5660 if (HDR_HAS_L2HDR(hdr) &&
5661 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5662 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5663 l2arc_read_callback_t *cb;
5667 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5668 ARCSTAT_BUMP(arcstat_l2_hits);
5670 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5672 cb->l2rcb_hdr = hdr;
5675 cb->l2rcb_flags = zio_flags;
5677 asize = vdev_psize_to_asize(vd, size);
5678 if (asize != size) {
5679 abd = abd_alloc_for_io(asize,
5680 HDR_ISTYPE_METADATA(hdr));
5681 cb->l2rcb_abd = abd;
5683 abd = hdr->b_l1hdr.b_pabd;
5686 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5687 addr + asize <= vd->vdev_psize -
5688 VDEV_LABEL_END_SIZE);
5691 * l2arc read. The SCL_L2ARC lock will be
5692 * released by l2arc_read_done().
5693 * Issue a null zio if the underlying buffer
5694 * was squashed to zero size by compression.
5696 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5697 ZIO_COMPRESS_EMPTY);
5698 rzio = zio_read_phys(pio, vd, addr,
5701 l2arc_read_done, cb, priority,
5702 zio_flags | ZIO_FLAG_DONT_CACHE |
5704 ZIO_FLAG_DONT_PROPAGATE |
5705 ZIO_FLAG_DONT_RETRY, B_FALSE);
5706 acb->acb_zio_head = rzio;
5708 if (hash_lock != NULL)
5709 mutex_exit(hash_lock);
5711 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5713 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5715 if (*arc_flags & ARC_FLAG_NOWAIT) {
5720 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5721 if (zio_wait(rzio) == 0)
5724 /* l2arc read error; goto zio_read() */
5725 if (hash_lock != NULL)
5726 mutex_enter(hash_lock);
5728 DTRACE_PROBE1(l2arc__miss,
5729 arc_buf_hdr_t *, hdr);
5730 ARCSTAT_BUMP(arcstat_l2_misses);
5731 if (HDR_L2_WRITING(hdr))
5732 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5733 spa_config_exit(spa, SCL_L2ARC, vd);
5737 spa_config_exit(spa, SCL_L2ARC, vd);
5738 if (l2arc_ndev != 0) {
5739 DTRACE_PROBE1(l2arc__miss,
5740 arc_buf_hdr_t *, hdr);
5741 ARCSTAT_BUMP(arcstat_l2_misses);
5745 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
5746 arc_read_done, hdr, priority, zio_flags, zb);
5747 acb->acb_zio_head = rzio;
5749 if (hash_lock != NULL)
5750 mutex_exit(hash_lock);
5752 if (*arc_flags & ARC_FLAG_WAIT)
5753 return (zio_wait(rzio));
5755 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5762 * Notify the arc that a block was freed, and thus will never be used again.
5765 arc_freed(spa_t *spa, const blkptr_t *bp)
5768 kmutex_t *hash_lock;
5769 uint64_t guid = spa_load_guid(spa);
5771 ASSERT(!BP_IS_EMBEDDED(bp));
5773 hdr = buf_hash_find(guid, bp, &hash_lock);
5778 * We might be trying to free a block that is still doing I/O
5779 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5780 * dmu_sync-ed block). If this block is being prefetched, then it
5781 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5782 * until the I/O completes. A block may also have a reference if it is
5783 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5784 * have written the new block to its final resting place on disk but
5785 * without the dedup flag set. This would have left the hdr in the MRU
5786 * state and discoverable. When the txg finally syncs it detects that
5787 * the block was overridden in open context and issues an override I/O.
5788 * Since this is a dedup block, the override I/O will determine if the
5789 * block is already in the DDT. If so, then it will replace the io_bp
5790 * with the bp from the DDT and allow the I/O to finish. When the I/O
5791 * reaches the done callback, dbuf_write_override_done, it will
5792 * check to see if the io_bp and io_bp_override are identical.
5793 * If they are not, then it indicates that the bp was replaced with
5794 * the bp in the DDT and the override bp is freed. This allows
5795 * us to arrive here with a reference on a block that is being
5796 * freed. So if we have an I/O in progress, or a reference to
5797 * this hdr, then we don't destroy the hdr.
5799 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5800 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5801 arc_change_state(arc_anon, hdr, hash_lock);
5802 arc_hdr_destroy(hdr);
5803 mutex_exit(hash_lock);
5805 mutex_exit(hash_lock);
5811 * Release this buffer from the cache, making it an anonymous buffer. This
5812 * must be done after a read and prior to modifying the buffer contents.
5813 * If the buffer has more than one reference, we must make
5814 * a new hdr for the buffer.
5817 arc_release(arc_buf_t *buf, void *tag)
5819 arc_buf_hdr_t *hdr = buf->b_hdr;
5822 * It would be nice to assert that if it's DMU metadata (level >
5823 * 0 || it's the dnode file), then it must be syncing context.
5824 * But we don't know that information at this level.
5827 mutex_enter(&buf->b_evict_lock);
5829 ASSERT(HDR_HAS_L1HDR(hdr));
5832 * We don't grab the hash lock prior to this check, because if
5833 * the buffer's header is in the arc_anon state, it won't be
5834 * linked into the hash table.
5836 if (hdr->b_l1hdr.b_state == arc_anon) {
5837 mutex_exit(&buf->b_evict_lock);
5838 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5839 ASSERT(!HDR_IN_HASH_TABLE(hdr));
5840 ASSERT(!HDR_HAS_L2HDR(hdr));
5841 ASSERT(HDR_EMPTY(hdr));
5842 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5843 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5844 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5846 hdr->b_l1hdr.b_arc_access = 0;
5849 * If the buf is being overridden then it may already
5850 * have a hdr that is not empty.
5852 buf_discard_identity(hdr);
5858 kmutex_t *hash_lock = HDR_LOCK(hdr);
5859 mutex_enter(hash_lock);
5862 * This assignment is only valid as long as the hash_lock is
5863 * held, we must be careful not to reference state or the
5864 * b_state field after dropping the lock.
5866 arc_state_t *state = hdr->b_l1hdr.b_state;
5867 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5868 ASSERT3P(state, !=, arc_anon);
5870 /* this buffer is not on any list */
5871 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
5873 if (HDR_HAS_L2HDR(hdr)) {
5874 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5877 * We have to recheck this conditional again now that
5878 * we're holding the l2ad_mtx to prevent a race with
5879 * another thread which might be concurrently calling
5880 * l2arc_evict(). In that case, l2arc_evict() might have
5881 * destroyed the header's L2 portion as we were waiting
5882 * to acquire the l2ad_mtx.
5884 if (HDR_HAS_L2HDR(hdr)) {
5886 arc_hdr_l2hdr_destroy(hdr);
5889 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5893 * Do we have more than one buf?
5895 if (hdr->b_l1hdr.b_bufcnt > 1) {
5896 arc_buf_hdr_t *nhdr;
5897 uint64_t spa = hdr->b_spa;
5898 uint64_t psize = HDR_GET_PSIZE(hdr);
5899 uint64_t lsize = HDR_GET_LSIZE(hdr);
5900 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5901 arc_buf_contents_t type = arc_buf_type(hdr);
5902 VERIFY3U(hdr->b_type, ==, type);
5904 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5905 (void) remove_reference(hdr, hash_lock, tag);
5907 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
5908 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5909 ASSERT(ARC_BUF_LAST(buf));
5913 * Pull the data off of this hdr and attach it to
5914 * a new anonymous hdr. Also find the last buffer
5915 * in the hdr's buffer list.
5917 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
5918 ASSERT3P(lastbuf, !=, NULL);
5921 * If the current arc_buf_t and the hdr are sharing their data
5922 * buffer, then we must stop sharing that block.
5924 if (arc_buf_is_shared(buf)) {
5925 VERIFY(!arc_buf_is_shared(lastbuf));
5928 * First, sever the block sharing relationship between
5929 * buf and the arc_buf_hdr_t.
5931 arc_unshare_buf(hdr, buf);
5934 * Now we need to recreate the hdr's b_pabd. Since we
5935 * have lastbuf handy, we try to share with it, but if
5936 * we can't then we allocate a new b_pabd and copy the
5937 * data from buf into it.
5939 if (arc_can_share(hdr, lastbuf)) {
5940 arc_share_buf(hdr, lastbuf);
5942 arc_hdr_alloc_pabd(hdr);
5943 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
5944 buf->b_data, psize);
5946 VERIFY3P(lastbuf->b_data, !=, NULL);
5947 } else if (HDR_SHARED_DATA(hdr)) {
5949 * Uncompressed shared buffers are always at the end
5950 * of the list. Compressed buffers don't have the
5951 * same requirements. This makes it hard to
5952 * simply assert that the lastbuf is shared so
5953 * we rely on the hdr's compression flags to determine
5954 * if we have a compressed, shared buffer.
5956 ASSERT(arc_buf_is_shared(lastbuf) ||
5957 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
5958 ASSERT(!ARC_BUF_SHARED(buf));
5960 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5961 ASSERT3P(state, !=, arc_l2c_only);
5963 (void) refcount_remove_many(&state->arcs_size,
5964 arc_buf_size(buf), buf);
5966 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5967 ASSERT3P(state, !=, arc_l2c_only);
5968 (void) refcount_remove_many(&state->arcs_esize[type],
5969 arc_buf_size(buf), buf);
5972 hdr->b_l1hdr.b_bufcnt -= 1;
5973 arc_cksum_verify(buf);
5975 arc_buf_unwatch(buf);
5978 mutex_exit(hash_lock);
5981 * Allocate a new hdr. The new hdr will contain a b_pabd
5982 * buffer which will be freed in arc_write().
5984 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5985 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5986 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5987 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5988 VERIFY3U(nhdr->b_type, ==, type);
5989 ASSERT(!HDR_SHARED_DATA(nhdr));
5991 nhdr->b_l1hdr.b_buf = buf;
5992 nhdr->b_l1hdr.b_bufcnt = 1;
5993 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5996 mutex_exit(&buf->b_evict_lock);
5997 (void) refcount_add_many(&arc_anon->arcs_size,
5998 arc_buf_size(buf), buf);
6000 mutex_exit(&buf->b_evict_lock);
6001 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6002 /* protected by hash lock, or hdr is on arc_anon */
6003 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6004 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6005 arc_change_state(arc_anon, hdr, hash_lock);
6006 hdr->b_l1hdr.b_arc_access = 0;
6007 mutex_exit(hash_lock);
6009 buf_discard_identity(hdr);
6015 arc_released(arc_buf_t *buf)
6019 mutex_enter(&buf->b_evict_lock);
6020 released = (buf->b_data != NULL &&
6021 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6022 mutex_exit(&buf->b_evict_lock);
6028 arc_referenced(arc_buf_t *buf)
6032 mutex_enter(&buf->b_evict_lock);
6033 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6034 mutex_exit(&buf->b_evict_lock);
6035 return (referenced);
6040 arc_write_ready(zio_t *zio)
6042 arc_write_callback_t *callback = zio->io_private;
6043 arc_buf_t *buf = callback->awcb_buf;
6044 arc_buf_hdr_t *hdr = buf->b_hdr;
6045 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
6047 ASSERT(HDR_HAS_L1HDR(hdr));
6048 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6049 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6052 * If we're reexecuting this zio because the pool suspended, then
6053 * cleanup any state that was previously set the first time the
6054 * callback was invoked.
6056 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6057 arc_cksum_free(hdr);
6059 arc_buf_unwatch(buf);
6061 if (hdr->b_l1hdr.b_pabd != NULL) {
6062 if (arc_buf_is_shared(buf)) {
6063 arc_unshare_buf(hdr, buf);
6065 arc_hdr_free_pabd(hdr);
6069 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6070 ASSERT(!HDR_SHARED_DATA(hdr));
6071 ASSERT(!arc_buf_is_shared(buf));
6073 callback->awcb_ready(zio, buf, callback->awcb_private);
6075 if (HDR_IO_IN_PROGRESS(hdr))
6076 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6078 arc_cksum_compute(buf);
6079 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6081 enum zio_compress compress;
6082 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6083 compress = ZIO_COMPRESS_OFF;
6085 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
6086 compress = BP_GET_COMPRESS(zio->io_bp);
6088 HDR_SET_PSIZE(hdr, psize);
6089 arc_hdr_set_compress(hdr, compress);
6093 * Fill the hdr with data. If the hdr is compressed, the data we want
6094 * is available from the zio, otherwise we can take it from the buf.
6096 * We might be able to share the buf's data with the hdr here. However,
6097 * doing so would cause the ARC to be full of linear ABDs if we write a
6098 * lot of shareable data. As a compromise, we check whether scattered
6099 * ABDs are allowed, and assume that if they are then the user wants
6100 * the ARC to be primarily filled with them regardless of the data being
6101 * written. Therefore, if they're allowed then we allocate one and copy
6102 * the data into it; otherwise, we share the data directly if we can.
6104 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6105 arc_hdr_alloc_pabd(hdr);
6108 * Ideally, we would always copy the io_abd into b_pabd, but the
6109 * user may have disabled compressed ARC, thus we must check the
6110 * hdr's compression setting rather than the io_bp's.
6112 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
6113 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
6115 ASSERT3U(psize, >, 0);
6117 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6119 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6121 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6125 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6126 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6127 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6129 arc_share_buf(hdr, buf);
6132 arc_hdr_verify(hdr, zio->io_bp);
6136 arc_write_children_ready(zio_t *zio)
6138 arc_write_callback_t *callback = zio->io_private;
6139 arc_buf_t *buf = callback->awcb_buf;
6141 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6145 * The SPA calls this callback for each physical write that happens on behalf
6146 * of a logical write. See the comment in dbuf_write_physdone() for details.
6149 arc_write_physdone(zio_t *zio)
6151 arc_write_callback_t *cb = zio->io_private;
6152 if (cb->awcb_physdone != NULL)
6153 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6157 arc_write_done(zio_t *zio)
6159 arc_write_callback_t *callback = zio->io_private;
6160 arc_buf_t *buf = callback->awcb_buf;
6161 arc_buf_hdr_t *hdr = buf->b_hdr;
6163 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6165 if (zio->io_error == 0) {
6166 arc_hdr_verify(hdr, zio->io_bp);
6168 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6169 buf_discard_identity(hdr);
6171 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6172 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6175 ASSERT(HDR_EMPTY(hdr));
6179 * If the block to be written was all-zero or compressed enough to be
6180 * embedded in the BP, no write was performed so there will be no
6181 * dva/birth/checksum. The buffer must therefore remain anonymous
6184 if (!HDR_EMPTY(hdr)) {
6185 arc_buf_hdr_t *exists;
6186 kmutex_t *hash_lock;
6188 ASSERT3U(zio->io_error, ==, 0);
6190 arc_cksum_verify(buf);
6192 exists = buf_hash_insert(hdr, &hash_lock);
6193 if (exists != NULL) {
6195 * This can only happen if we overwrite for
6196 * sync-to-convergence, because we remove
6197 * buffers from the hash table when we arc_free().
6199 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6200 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6201 panic("bad overwrite, hdr=%p exists=%p",
6202 (void *)hdr, (void *)exists);
6203 ASSERT(refcount_is_zero(
6204 &exists->b_l1hdr.b_refcnt));
6205 arc_change_state(arc_anon, exists, hash_lock);
6206 mutex_exit(hash_lock);
6207 arc_hdr_destroy(exists);
6208 exists = buf_hash_insert(hdr, &hash_lock);
6209 ASSERT3P(exists, ==, NULL);
6210 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6212 ASSERT(zio->io_prop.zp_nopwrite);
6213 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6214 panic("bad nopwrite, hdr=%p exists=%p",
6215 (void *)hdr, (void *)exists);
6218 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6219 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6220 ASSERT(BP_GET_DEDUP(zio->io_bp));
6221 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6224 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6225 /* if it's not anon, we are doing a scrub */
6226 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6227 arc_access(hdr, hash_lock);
6228 mutex_exit(hash_lock);
6230 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6233 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6234 callback->awcb_done(zio, buf, callback->awcb_private);
6236 abd_put(zio->io_abd);
6237 kmem_free(callback, sizeof (arc_write_callback_t));
6241 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6242 boolean_t l2arc, const zio_prop_t *zp, arc_write_done_func_t *ready,
6243 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
6244 arc_write_done_func_t *done, void *private, zio_priority_t priority,
6245 int zio_flags, const zbookmark_phys_t *zb)
6247 arc_buf_hdr_t *hdr = buf->b_hdr;
6248 arc_write_callback_t *callback;
6250 zio_prop_t localprop = *zp;
6252 ASSERT3P(ready, !=, NULL);
6253 ASSERT3P(done, !=, NULL);
6254 ASSERT(!HDR_IO_ERROR(hdr));
6255 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6256 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6257 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6259 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6260 if (ARC_BUF_COMPRESSED(buf)) {
6262 * We're writing a pre-compressed buffer. Make the
6263 * compression algorithm requested by the zio_prop_t match
6264 * the pre-compressed buffer's compression algorithm.
6266 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6268 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6269 zio_flags |= ZIO_FLAG_RAW;
6271 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6272 callback->awcb_ready = ready;
6273 callback->awcb_children_ready = children_ready;
6274 callback->awcb_physdone = physdone;
6275 callback->awcb_done = done;
6276 callback->awcb_private = private;
6277 callback->awcb_buf = buf;
6280 * The hdr's b_pabd is now stale, free it now. A new data block
6281 * will be allocated when the zio pipeline calls arc_write_ready().
6283 if (hdr->b_l1hdr.b_pabd != NULL) {
6285 * If the buf is currently sharing the data block with
6286 * the hdr then we need to break that relationship here.
6287 * The hdr will remain with a NULL data pointer and the
6288 * buf will take sole ownership of the block.
6290 if (arc_buf_is_shared(buf)) {
6291 arc_unshare_buf(hdr, buf);
6293 arc_hdr_free_pabd(hdr);
6295 VERIFY3P(buf->b_data, !=, NULL);
6296 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6298 ASSERT(!arc_buf_is_shared(buf));
6299 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6301 zio = zio_write(pio, spa, txg, bp,
6302 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6303 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6304 (children_ready != NULL) ? arc_write_children_ready : NULL,
6305 arc_write_physdone, arc_write_done, callback,
6306 priority, zio_flags, zb);
6312 arc_memory_throttle(uint64_t reserve, uint64_t txg)
6315 uint64_t available_memory = ptob(freemem);
6316 static uint64_t page_load = 0;
6317 static uint64_t last_txg = 0;
6319 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
6321 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE)));
6324 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
6327 if (txg > last_txg) {
6332 * If we are in pageout, we know that memory is already tight,
6333 * the arc is already going to be evicting, so we just want to
6334 * continue to let page writes occur as quickly as possible.
6336 if (curproc == pageproc) {
6337 if (page_load > MAX(ptob(minfree), available_memory) / 4)
6338 return (SET_ERROR(ERESTART));
6339 /* Note: reserve is inflated, so we deflate */
6340 page_load += reserve / 8;
6342 } else if (page_load > 0 && arc_reclaim_needed()) {
6343 /* memory is low, delay before restarting */
6344 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6345 return (SET_ERROR(EAGAIN));
6353 arc_tempreserve_clear(uint64_t reserve)
6355 atomic_add_64(&arc_tempreserve, -reserve);
6356 ASSERT((int64_t)arc_tempreserve >= 0);
6360 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
6365 if (reserve > arc_c/4 && !arc_no_grow) {
6366 arc_c = MIN(arc_c_max, reserve * 4);
6367 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
6369 if (reserve > arc_c)
6370 return (SET_ERROR(ENOMEM));
6373 * Don't count loaned bufs as in flight dirty data to prevent long
6374 * network delays from blocking transactions that are ready to be
6375 * assigned to a txg.
6378 /* assert that it has not wrapped around */
6379 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6381 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6382 arc_loaned_bytes), 0);
6385 * Writes will, almost always, require additional memory allocations
6386 * in order to compress/encrypt/etc the data. We therefore need to
6387 * make sure that there is sufficient available memory for this.
6389 error = arc_memory_throttle(reserve, txg);
6394 * Throttle writes when the amount of dirty data in the cache
6395 * gets too large. We try to keep the cache less than half full
6396 * of dirty blocks so that our sync times don't grow too large.
6397 * Note: if two requests come in concurrently, we might let them
6398 * both succeed, when one of them should fail. Not a huge deal.
6401 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
6402 anon_size > arc_c / 4) {
6403 uint64_t meta_esize =
6404 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6405 uint64_t data_esize =
6406 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6407 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6408 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6409 arc_tempreserve >> 10, meta_esize >> 10,
6410 data_esize >> 10, reserve >> 10, arc_c >> 10);
6411 return (SET_ERROR(ERESTART));
6413 atomic_add_64(&arc_tempreserve, reserve);
6418 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6419 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6421 size->value.ui64 = refcount_count(&state->arcs_size);
6422 evict_data->value.ui64 =
6423 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6424 evict_metadata->value.ui64 =
6425 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6429 arc_kstat_update(kstat_t *ksp, int rw)
6431 arc_stats_t *as = ksp->ks_data;
6433 if (rw == KSTAT_WRITE) {
6436 arc_kstat_update_state(arc_anon,
6437 &as->arcstat_anon_size,
6438 &as->arcstat_anon_evictable_data,
6439 &as->arcstat_anon_evictable_metadata);
6440 arc_kstat_update_state(arc_mru,
6441 &as->arcstat_mru_size,
6442 &as->arcstat_mru_evictable_data,
6443 &as->arcstat_mru_evictable_metadata);
6444 arc_kstat_update_state(arc_mru_ghost,
6445 &as->arcstat_mru_ghost_size,
6446 &as->arcstat_mru_ghost_evictable_data,
6447 &as->arcstat_mru_ghost_evictable_metadata);
6448 arc_kstat_update_state(arc_mfu,
6449 &as->arcstat_mfu_size,
6450 &as->arcstat_mfu_evictable_data,
6451 &as->arcstat_mfu_evictable_metadata);
6452 arc_kstat_update_state(arc_mfu_ghost,
6453 &as->arcstat_mfu_ghost_size,
6454 &as->arcstat_mfu_ghost_evictable_data,
6455 &as->arcstat_mfu_ghost_evictable_metadata);
6457 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
6458 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
6459 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
6460 ARCSTAT(arcstat_metadata_size) =
6461 aggsum_value(&astat_metadata_size);
6462 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
6463 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size);
6464 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
6471 * This function *must* return indices evenly distributed between all
6472 * sublists of the multilist. This is needed due to how the ARC eviction
6473 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6474 * distributed between all sublists and uses this assumption when
6475 * deciding which sublist to evict from and how much to evict from it.
6478 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6480 arc_buf_hdr_t *hdr = obj;
6483 * We rely on b_dva to generate evenly distributed index
6484 * numbers using buf_hash below. So, as an added precaution,
6485 * let's make sure we never add empty buffers to the arc lists.
6487 ASSERT(!HDR_EMPTY(hdr));
6490 * The assumption here, is the hash value for a given
6491 * arc_buf_hdr_t will remain constant throughout it's lifetime
6492 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
6493 * Thus, we don't need to store the header's sublist index
6494 * on insertion, as this index can be recalculated on removal.
6496 * Also, the low order bits of the hash value are thought to be
6497 * distributed evenly. Otherwise, in the case that the multilist
6498 * has a power of two number of sublists, each sublists' usage
6499 * would not be evenly distributed.
6501 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6502 multilist_get_num_sublists(ml));
6506 static eventhandler_tag arc_event_lowmem = NULL;
6509 arc_lowmem(void *arg __unused, int howto __unused)
6512 mutex_enter(&arc_reclaim_lock);
6513 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE);
6514 cv_signal(&arc_reclaim_thread_cv);
6517 * It is unsafe to block here in arbitrary threads, because we can come
6518 * here from ARC itself and may hold ARC locks and thus risk a deadlock
6519 * with ARC reclaim thread.
6521 if (curproc == pageproc)
6522 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
6523 mutex_exit(&arc_reclaim_lock);
6528 arc_state_init(void)
6530 arc_anon = &ARC_anon;
6532 arc_mru_ghost = &ARC_mru_ghost;
6534 arc_mfu_ghost = &ARC_mfu_ghost;
6535 arc_l2c_only = &ARC_l2c_only;
6537 arc_mru->arcs_list[ARC_BUFC_METADATA] =
6538 multilist_create(sizeof (arc_buf_hdr_t),
6539 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6540 arc_state_multilist_index_func);
6541 arc_mru->arcs_list[ARC_BUFC_DATA] =
6542 multilist_create(sizeof (arc_buf_hdr_t),
6543 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6544 arc_state_multilist_index_func);
6545 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6546 multilist_create(sizeof (arc_buf_hdr_t),
6547 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6548 arc_state_multilist_index_func);
6549 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6550 multilist_create(sizeof (arc_buf_hdr_t),
6551 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6552 arc_state_multilist_index_func);
6553 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6554 multilist_create(sizeof (arc_buf_hdr_t),
6555 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6556 arc_state_multilist_index_func);
6557 arc_mfu->arcs_list[ARC_BUFC_DATA] =
6558 multilist_create(sizeof (arc_buf_hdr_t),
6559 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6560 arc_state_multilist_index_func);
6561 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6562 multilist_create(sizeof (arc_buf_hdr_t),
6563 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6564 arc_state_multilist_index_func);
6565 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6566 multilist_create(sizeof (arc_buf_hdr_t),
6567 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6568 arc_state_multilist_index_func);
6569 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6570 multilist_create(sizeof (arc_buf_hdr_t),
6571 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6572 arc_state_multilist_index_func);
6573 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6574 multilist_create(sizeof (arc_buf_hdr_t),
6575 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6576 arc_state_multilist_index_func);
6578 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6579 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6580 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6581 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6582 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6583 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6584 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6585 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6586 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6587 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6588 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6589 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6591 refcount_create(&arc_anon->arcs_size);
6592 refcount_create(&arc_mru->arcs_size);
6593 refcount_create(&arc_mru_ghost->arcs_size);
6594 refcount_create(&arc_mfu->arcs_size);
6595 refcount_create(&arc_mfu_ghost->arcs_size);
6596 refcount_create(&arc_l2c_only->arcs_size);
6598 aggsum_init(&arc_meta_used, 0);
6599 aggsum_init(&arc_size, 0);
6600 aggsum_init(&astat_data_size, 0);
6601 aggsum_init(&astat_metadata_size, 0);
6602 aggsum_init(&astat_hdr_size, 0);
6603 aggsum_init(&astat_other_size, 0);
6604 aggsum_init(&astat_l2_hdr_size, 0);
6608 arc_state_fini(void)
6610 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6611 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6612 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6613 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6614 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6615 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6616 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6617 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6618 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6619 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6620 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6621 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6623 refcount_destroy(&arc_anon->arcs_size);
6624 refcount_destroy(&arc_mru->arcs_size);
6625 refcount_destroy(&arc_mru_ghost->arcs_size);
6626 refcount_destroy(&arc_mfu->arcs_size);
6627 refcount_destroy(&arc_mfu_ghost->arcs_size);
6628 refcount_destroy(&arc_l2c_only->arcs_size);
6630 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6631 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6632 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6633 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6634 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6635 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6636 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6637 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6649 int i, prefetch_tunable_set = 0;
6652 * allmem is "all memory that we could possibly use".
6656 uint64_t allmem = ptob(physmem - swapfs_minfree);
6658 uint64_t allmem = (physmem * PAGESIZE) / 2;
6661 uint64_t allmem = kmem_size();
6665 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
6666 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
6667 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
6669 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
6670 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
6672 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
6673 arc_c_min = MAX(allmem / 32, arc_abs_min);
6674 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */
6675 if (allmem >= 1 << 30)
6676 arc_c_max = allmem - (1 << 30);
6678 arc_c_max = arc_c_min;
6679 arc_c_max = MAX(allmem * 5 / 8, arc_c_max);
6682 * In userland, there's only the memory pressure that we artificially
6683 * create (see arc_available_memory()). Don't let arc_c get too
6684 * small, because it can cause transactions to be larger than
6685 * arc_c, causing arc_tempreserve_space() to fail.
6688 arc_c_min = arc_c_max / 2;
6693 * Allow the tunables to override our calculations if they are
6696 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) {
6697 arc_c_max = zfs_arc_max;
6698 arc_c_min = MIN(arc_c_min, arc_c_max);
6700 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
6701 arc_c_min = zfs_arc_min;
6705 arc_p = (arc_c >> 1);
6707 /* limit meta-data to 1/4 of the arc capacity */
6708 arc_meta_limit = arc_c_max / 4;
6712 * Metadata is stored in the kernel's heap. Don't let us
6713 * use more than half the heap for the ARC.
6715 arc_meta_limit = MIN(arc_meta_limit,
6716 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6719 /* Allow the tunable to override if it is reasonable */
6720 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6721 arc_meta_limit = zfs_arc_meta_limit;
6723 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6724 arc_c_min = arc_meta_limit / 2;
6726 if (zfs_arc_meta_min > 0) {
6727 arc_meta_min = zfs_arc_meta_min;
6729 arc_meta_min = arc_c_min / 2;
6732 if (zfs_arc_grow_retry > 0)
6733 arc_grow_retry = zfs_arc_grow_retry;
6735 if (zfs_arc_shrink_shift > 0)
6736 arc_shrink_shift = zfs_arc_shrink_shift;
6738 if (zfs_arc_no_grow_shift > 0)
6739 arc_no_grow_shift = zfs_arc_no_grow_shift;
6741 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6743 if (arc_no_grow_shift >= arc_shrink_shift)
6744 arc_no_grow_shift = arc_shrink_shift - 1;
6746 if (zfs_arc_p_min_shift > 0)
6747 arc_p_min_shift = zfs_arc_p_min_shift;
6749 /* if kmem_flags are set, lets try to use less memory */
6750 if (kmem_debugging())
6752 if (arc_c < arc_c_min)
6755 zfs_arc_min = arc_c_min;
6756 zfs_arc_max = arc_c_max;
6761 arc_reclaim_thread_exit = B_FALSE;
6762 arc_dnlc_evicts_thread_exit = FALSE;
6764 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6765 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6767 if (arc_ksp != NULL) {
6768 arc_ksp->ks_data = &arc_stats;
6769 arc_ksp->ks_update = arc_kstat_update;
6770 kstat_install(arc_ksp);
6773 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6774 TS_RUN, minclsyspri);
6777 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
6778 EVENTHANDLER_PRI_FIRST);
6781 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
6782 TS_RUN, minclsyspri);
6788 * Calculate maximum amount of dirty data per pool.
6790 * If it has been set by /etc/system, take that.
6791 * Otherwise, use a percentage of physical memory defined by
6792 * zfs_dirty_data_max_percent (default 10%) with a cap at
6793 * zfs_dirty_data_max_max (default 4GB).
6795 if (zfs_dirty_data_max == 0) {
6796 zfs_dirty_data_max = ptob(physmem) *
6797 zfs_dirty_data_max_percent / 100;
6798 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6799 zfs_dirty_data_max_max);
6803 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
6804 prefetch_tunable_set = 1;
6807 if (prefetch_tunable_set == 0) {
6808 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
6810 printf(" add \"vfs.zfs.prefetch_disable=0\" "
6811 "to /boot/loader.conf.\n");
6812 zfs_prefetch_disable = 1;
6815 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
6816 prefetch_tunable_set == 0) {
6817 printf("ZFS NOTICE: Prefetch is disabled by default if less "
6818 "than 4GB of RAM is present;\n"
6819 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
6820 "to /boot/loader.conf.\n");
6821 zfs_prefetch_disable = 1;
6824 /* Warn about ZFS memory and address space requirements. */
6825 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
6826 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
6827 "expect unstable behavior.\n");
6829 if (allmem < 512 * (1 << 20)) {
6830 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
6831 "expect unstable behavior.\n");
6832 printf(" Consider tuning vm.kmem_size and "
6833 "vm.kmem_size_max\n");
6834 printf(" in /boot/loader.conf.\n");
6843 if (arc_event_lowmem != NULL)
6844 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
6847 mutex_enter(&arc_reclaim_lock);
6848 arc_reclaim_thread_exit = B_TRUE;
6850 * The reclaim thread will set arc_reclaim_thread_exit back to
6851 * B_FALSE when it is finished exiting; we're waiting for that.
6853 while (arc_reclaim_thread_exit) {
6854 cv_signal(&arc_reclaim_thread_cv);
6855 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6857 mutex_exit(&arc_reclaim_lock);
6859 /* Use B_TRUE to ensure *all* buffers are evicted */
6860 arc_flush(NULL, B_TRUE);
6862 mutex_enter(&arc_dnlc_evicts_lock);
6863 arc_dnlc_evicts_thread_exit = TRUE;
6865 * The user evicts thread will set arc_user_evicts_thread_exit
6866 * to FALSE when it is finished exiting; we're waiting for that.
6868 while (arc_dnlc_evicts_thread_exit) {
6869 cv_signal(&arc_dnlc_evicts_cv);
6870 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
6872 mutex_exit(&arc_dnlc_evicts_lock);
6876 if (arc_ksp != NULL) {
6877 kstat_delete(arc_ksp);
6881 mutex_destroy(&arc_reclaim_lock);
6882 cv_destroy(&arc_reclaim_thread_cv);
6883 cv_destroy(&arc_reclaim_waiters_cv);
6885 mutex_destroy(&arc_dnlc_evicts_lock);
6886 cv_destroy(&arc_dnlc_evicts_cv);
6891 ASSERT0(arc_loaned_bytes);
6897 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6898 * It uses dedicated storage devices to hold cached data, which are populated
6899 * using large infrequent writes. The main role of this cache is to boost
6900 * the performance of random read workloads. The intended L2ARC devices
6901 * include short-stroked disks, solid state disks, and other media with
6902 * substantially faster read latency than disk.
6904 * +-----------------------+
6906 * +-----------------------+
6909 * l2arc_feed_thread() arc_read()
6913 * +---------------+ |
6915 * +---------------+ |
6920 * +-------+ +-------+
6922 * | cache | | cache |
6923 * +-------+ +-------+
6924 * +=========+ .-----.
6925 * : L2ARC : |-_____-|
6926 * : devices : | Disks |
6927 * +=========+ `-_____-'
6929 * Read requests are satisfied from the following sources, in order:
6932 * 2) vdev cache of L2ARC devices
6934 * 4) vdev cache of disks
6937 * Some L2ARC device types exhibit extremely slow write performance.
6938 * To accommodate for this there are some significant differences between
6939 * the L2ARC and traditional cache design:
6941 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6942 * the ARC behave as usual, freeing buffers and placing headers on ghost
6943 * lists. The ARC does not send buffers to the L2ARC during eviction as
6944 * this would add inflated write latencies for all ARC memory pressure.
6946 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6947 * It does this by periodically scanning buffers from the eviction-end of
6948 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6949 * not already there. It scans until a headroom of buffers is satisfied,
6950 * which itself is a buffer for ARC eviction. If a compressible buffer is
6951 * found during scanning and selected for writing to an L2ARC device, we
6952 * temporarily boost scanning headroom during the next scan cycle to make
6953 * sure we adapt to compression effects (which might significantly reduce
6954 * the data volume we write to L2ARC). The thread that does this is
6955 * l2arc_feed_thread(), illustrated below; example sizes are included to
6956 * provide a better sense of ratio than this diagram:
6959 * +---------------------+----------+
6960 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6961 * +---------------------+----------+ | o L2ARC eligible
6962 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6963 * +---------------------+----------+ |
6964 * 15.9 Gbytes ^ 32 Mbytes |
6966 * l2arc_feed_thread()
6968 * l2arc write hand <--[oooo]--'
6972 * +==============================+
6973 * L2ARC dev |####|#|###|###| |####| ... |
6974 * +==============================+
6977 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6978 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6979 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6980 * safe to say that this is an uncommon case, since buffers at the end of
6981 * the ARC lists have moved there due to inactivity.
6983 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6984 * then the L2ARC simply misses copying some buffers. This serves as a
6985 * pressure valve to prevent heavy read workloads from both stalling the ARC
6986 * with waits and clogging the L2ARC with writes. This also helps prevent
6987 * the potential for the L2ARC to churn if it attempts to cache content too
6988 * quickly, such as during backups of the entire pool.
6990 * 5. After system boot and before the ARC has filled main memory, there are
6991 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6992 * lists can remain mostly static. Instead of searching from tail of these
6993 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6994 * for eligible buffers, greatly increasing its chance of finding them.
6996 * The L2ARC device write speed is also boosted during this time so that
6997 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6998 * there are no L2ARC reads, and no fear of degrading read performance
6999 * through increased writes.
7001 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7002 * the vdev queue can aggregate them into larger and fewer writes. Each
7003 * device is written to in a rotor fashion, sweeping writes through
7004 * available space then repeating.
7006 * 7. The L2ARC does not store dirty content. It never needs to flush
7007 * write buffers back to disk based storage.
7009 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7010 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7012 * The performance of the L2ARC can be tweaked by a number of tunables, which
7013 * may be necessary for different workloads:
7015 * l2arc_write_max max write bytes per interval
7016 * l2arc_write_boost extra write bytes during device warmup
7017 * l2arc_noprefetch skip caching prefetched buffers
7018 * l2arc_headroom number of max device writes to precache
7019 * l2arc_headroom_boost when we find compressed buffers during ARC
7020 * scanning, we multiply headroom by this
7021 * percentage factor for the next scan cycle,
7022 * since more compressed buffers are likely to
7024 * l2arc_feed_secs seconds between L2ARC writing
7026 * Tunables may be removed or added as future performance improvements are
7027 * integrated, and also may become zpool properties.
7029 * There are three key functions that control how the L2ARC warms up:
7031 * l2arc_write_eligible() check if a buffer is eligible to cache
7032 * l2arc_write_size() calculate how much to write
7033 * l2arc_write_interval() calculate sleep delay between writes
7035 * These three functions determine what to write, how much, and how quickly
7040 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
7043 * A buffer is *not* eligible for the L2ARC if it:
7044 * 1. belongs to a different spa.
7045 * 2. is already cached on the L2ARC.
7046 * 3. has an I/O in progress (it may be an incomplete read).
7047 * 4. is flagged not eligible (zfs property).
7049 if (hdr->b_spa != spa_guid) {
7050 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
7053 if (HDR_HAS_L2HDR(hdr)) {
7054 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
7057 if (HDR_IO_IN_PROGRESS(hdr)) {
7058 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
7061 if (!HDR_L2CACHE(hdr)) {
7062 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
7070 l2arc_write_size(void)
7075 * Make sure our globals have meaningful values in case the user
7078 size = l2arc_write_max;
7080 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
7081 "be greater than zero, resetting it to the default (%d)",
7083 size = l2arc_write_max = L2ARC_WRITE_SIZE;
7086 if (arc_warm == B_FALSE)
7087 size += l2arc_write_boost;
7094 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
7096 clock_t interval, next, now;
7099 * If the ARC lists are busy, increase our write rate; if the
7100 * lists are stale, idle back. This is achieved by checking
7101 * how much we previously wrote - if it was more than half of
7102 * what we wanted, schedule the next write much sooner.
7104 if (l2arc_feed_again && wrote > (wanted / 2))
7105 interval = (hz * l2arc_feed_min_ms) / 1000;
7107 interval = hz * l2arc_feed_secs;
7109 now = ddi_get_lbolt();
7110 next = MAX(now, MIN(now + interval, began + interval));
7116 * Cycle through L2ARC devices. This is how L2ARC load balances.
7117 * If a device is returned, this also returns holding the spa config lock.
7119 static l2arc_dev_t *
7120 l2arc_dev_get_next(void)
7122 l2arc_dev_t *first, *next = NULL;
7125 * Lock out the removal of spas (spa_namespace_lock), then removal
7126 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7127 * both locks will be dropped and a spa config lock held instead.
7129 mutex_enter(&spa_namespace_lock);
7130 mutex_enter(&l2arc_dev_mtx);
7132 /* if there are no vdevs, there is nothing to do */
7133 if (l2arc_ndev == 0)
7137 next = l2arc_dev_last;
7139 /* loop around the list looking for a non-faulted vdev */
7141 next = list_head(l2arc_dev_list);
7143 next = list_next(l2arc_dev_list, next);
7145 next = list_head(l2arc_dev_list);
7148 /* if we have come back to the start, bail out */
7151 else if (next == first)
7154 } while (vdev_is_dead(next->l2ad_vdev));
7156 /* if we were unable to find any usable vdevs, return NULL */
7157 if (vdev_is_dead(next->l2ad_vdev))
7160 l2arc_dev_last = next;
7163 mutex_exit(&l2arc_dev_mtx);
7166 * Grab the config lock to prevent the 'next' device from being
7167 * removed while we are writing to it.
7170 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7171 mutex_exit(&spa_namespace_lock);
7177 * Free buffers that were tagged for destruction.
7180 l2arc_do_free_on_write()
7183 l2arc_data_free_t *df, *df_prev;
7185 mutex_enter(&l2arc_free_on_write_mtx);
7186 buflist = l2arc_free_on_write;
7188 for (df = list_tail(buflist); df; df = df_prev) {
7189 df_prev = list_prev(buflist, df);
7190 ASSERT3P(df->l2df_abd, !=, NULL);
7191 abd_free(df->l2df_abd);
7192 list_remove(buflist, df);
7193 kmem_free(df, sizeof (l2arc_data_free_t));
7196 mutex_exit(&l2arc_free_on_write_mtx);
7200 * A write to a cache device has completed. Update all headers to allow
7201 * reads from these buffers to begin.
7204 l2arc_write_done(zio_t *zio)
7206 l2arc_write_callback_t *cb;
7209 arc_buf_hdr_t *head, *hdr, *hdr_prev;
7210 kmutex_t *hash_lock;
7211 int64_t bytes_dropped = 0;
7213 cb = zio->io_private;
7214 ASSERT3P(cb, !=, NULL);
7215 dev = cb->l2wcb_dev;
7216 ASSERT3P(dev, !=, NULL);
7217 head = cb->l2wcb_head;
7218 ASSERT3P(head, !=, NULL);
7219 buflist = &dev->l2ad_buflist;
7220 ASSERT3P(buflist, !=, NULL);
7221 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7222 l2arc_write_callback_t *, cb);
7224 if (zio->io_error != 0)
7225 ARCSTAT_BUMP(arcstat_l2_writes_error);
7228 * All writes completed, or an error was hit.
7231 mutex_enter(&dev->l2ad_mtx);
7232 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7233 hdr_prev = list_prev(buflist, hdr);
7235 hash_lock = HDR_LOCK(hdr);
7238 * We cannot use mutex_enter or else we can deadlock
7239 * with l2arc_write_buffers (due to swapping the order
7240 * the hash lock and l2ad_mtx are taken).
7242 if (!mutex_tryenter(hash_lock)) {
7244 * Missed the hash lock. We must retry so we
7245 * don't leave the ARC_FLAG_L2_WRITING bit set.
7247 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7250 * We don't want to rescan the headers we've
7251 * already marked as having been written out, so
7252 * we reinsert the head node so we can pick up
7253 * where we left off.
7255 list_remove(buflist, head);
7256 list_insert_after(buflist, hdr, head);
7258 mutex_exit(&dev->l2ad_mtx);
7261 * We wait for the hash lock to become available
7262 * to try and prevent busy waiting, and increase
7263 * the chance we'll be able to acquire the lock
7264 * the next time around.
7266 mutex_enter(hash_lock);
7267 mutex_exit(hash_lock);
7272 * We could not have been moved into the arc_l2c_only
7273 * state while in-flight due to our ARC_FLAG_L2_WRITING
7274 * bit being set. Let's just ensure that's being enforced.
7276 ASSERT(HDR_HAS_L1HDR(hdr));
7278 if (zio->io_error != 0) {
7280 * Error - drop L2ARC entry.
7282 list_remove(buflist, hdr);
7284 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7286 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
7287 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
7289 bytes_dropped += arc_hdr_size(hdr);
7290 (void) refcount_remove_many(&dev->l2ad_alloc,
7291 arc_hdr_size(hdr), hdr);
7295 * Allow ARC to begin reads and ghost list evictions to
7298 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7300 mutex_exit(hash_lock);
7303 atomic_inc_64(&l2arc_writes_done);
7304 list_remove(buflist, head);
7305 ASSERT(!HDR_HAS_L1HDR(head));
7306 kmem_cache_free(hdr_l2only_cache, head);
7307 mutex_exit(&dev->l2ad_mtx);
7309 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7311 l2arc_do_free_on_write();
7313 kmem_free(cb, sizeof (l2arc_write_callback_t));
7317 * A read to a cache device completed. Validate buffer contents before
7318 * handing over to the regular ARC routines.
7321 l2arc_read_done(zio_t *zio)
7323 l2arc_read_callback_t *cb;
7325 kmutex_t *hash_lock;
7326 boolean_t valid_cksum;
7328 ASSERT3P(zio->io_vd, !=, NULL);
7329 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7331 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7333 cb = zio->io_private;
7334 ASSERT3P(cb, !=, NULL);
7335 hdr = cb->l2rcb_hdr;
7336 ASSERT3P(hdr, !=, NULL);
7338 hash_lock = HDR_LOCK(hdr);
7339 mutex_enter(hash_lock);
7340 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7343 * If the data was read into a temporary buffer,
7344 * move it and free the buffer.
7346 if (cb->l2rcb_abd != NULL) {
7347 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7348 if (zio->io_error == 0) {
7349 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
7354 * The following must be done regardless of whether
7355 * there was an error:
7356 * - free the temporary buffer
7357 * - point zio to the real ARC buffer
7358 * - set zio size accordingly
7359 * These are required because zio is either re-used for
7360 * an I/O of the block in the case of the error
7361 * or the zio is passed to arc_read_done() and it
7364 abd_free(cb->l2rcb_abd);
7365 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7366 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
7369 ASSERT3P(zio->io_abd, !=, NULL);
7372 * Check this survived the L2ARC journey.
7374 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
7375 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
7376 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
7378 valid_cksum = arc_cksum_is_equal(hdr, zio);
7379 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
7380 mutex_exit(hash_lock);
7381 zio->io_private = hdr;
7384 mutex_exit(hash_lock);
7386 * Buffer didn't survive caching. Increment stats and
7387 * reissue to the original storage device.
7389 if (zio->io_error != 0) {
7390 ARCSTAT_BUMP(arcstat_l2_io_error);
7392 zio->io_error = SET_ERROR(EIO);
7395 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
7398 * If there's no waiter, issue an async i/o to the primary
7399 * storage now. If there *is* a waiter, the caller must
7400 * issue the i/o in a context where it's OK to block.
7402 if (zio->io_waiter == NULL) {
7403 zio_t *pio = zio_unique_parent(zio);
7405 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7407 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
7408 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
7409 hdr, zio->io_priority, cb->l2rcb_flags,
7414 kmem_free(cb, sizeof (l2arc_read_callback_t));
7418 * This is the list priority from which the L2ARC will search for pages to
7419 * cache. This is used within loops (0..3) to cycle through lists in the
7420 * desired order. This order can have a significant effect on cache
7423 * Currently the metadata lists are hit first, MFU then MRU, followed by
7424 * the data lists. This function returns a locked list, and also returns
7427 static multilist_sublist_t *
7428 l2arc_sublist_lock(int list_num)
7430 multilist_t *ml = NULL;
7433 ASSERT(list_num >= 0 && list_num <= 3);
7437 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7440 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7443 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7446 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7451 * Return a randomly-selected sublist. This is acceptable
7452 * because the caller feeds only a little bit of data for each
7453 * call (8MB). Subsequent calls will result in different
7454 * sublists being selected.
7456 idx = multilist_get_random_index(ml);
7457 return (multilist_sublist_lock(ml, idx));
7461 * Evict buffers from the device write hand to the distance specified in
7462 * bytes. This distance may span populated buffers, it may span nothing.
7463 * This is clearing a region on the L2ARC device ready for writing.
7464 * If the 'all' boolean is set, every buffer is evicted.
7467 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7470 arc_buf_hdr_t *hdr, *hdr_prev;
7471 kmutex_t *hash_lock;
7474 buflist = &dev->l2ad_buflist;
7476 if (!all && dev->l2ad_first) {
7478 * This is the first sweep through the device. There is
7484 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
7486 * When nearing the end of the device, evict to the end
7487 * before the device write hand jumps to the start.
7489 taddr = dev->l2ad_end;
7491 taddr = dev->l2ad_hand + distance;
7493 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
7494 uint64_t, taddr, boolean_t, all);
7497 mutex_enter(&dev->l2ad_mtx);
7498 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
7499 hdr_prev = list_prev(buflist, hdr);
7501 hash_lock = HDR_LOCK(hdr);
7504 * We cannot use mutex_enter or else we can deadlock
7505 * with l2arc_write_buffers (due to swapping the order
7506 * the hash lock and l2ad_mtx are taken).
7508 if (!mutex_tryenter(hash_lock)) {
7510 * Missed the hash lock. Retry.
7512 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
7513 mutex_exit(&dev->l2ad_mtx);
7514 mutex_enter(hash_lock);
7515 mutex_exit(hash_lock);
7520 * A header can't be on this list if it doesn't have L2 header.
7522 ASSERT(HDR_HAS_L2HDR(hdr));
7524 /* Ensure this header has finished being written. */
7525 ASSERT(!HDR_L2_WRITING(hdr));
7526 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
7528 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
7529 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
7531 * We've evicted to the target address,
7532 * or the end of the device.
7534 mutex_exit(hash_lock);
7538 if (!HDR_HAS_L1HDR(hdr)) {
7539 ASSERT(!HDR_L2_READING(hdr));
7541 * This doesn't exist in the ARC. Destroy.
7542 * arc_hdr_destroy() will call list_remove()
7543 * and decrement arcstat_l2_lsize.
7545 arc_change_state(arc_anon, hdr, hash_lock);
7546 arc_hdr_destroy(hdr);
7548 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
7549 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
7551 * Invalidate issued or about to be issued
7552 * reads, since we may be about to write
7553 * over this location.
7555 if (HDR_L2_READING(hdr)) {
7556 ARCSTAT_BUMP(arcstat_l2_evict_reading);
7557 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
7560 arc_hdr_l2hdr_destroy(hdr);
7562 mutex_exit(hash_lock);
7564 mutex_exit(&dev->l2ad_mtx);
7568 * Find and write ARC buffers to the L2ARC device.
7570 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7571 * for reading until they have completed writing.
7572 * The headroom_boost is an in-out parameter used to maintain headroom boost
7573 * state between calls to this function.
7575 * Returns the number of bytes actually written (which may be smaller than
7576 * the delta by which the device hand has changed due to alignment).
7579 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
7581 arc_buf_hdr_t *hdr, *hdr_prev, *head;
7582 uint64_t write_asize, write_psize, write_lsize, headroom;
7584 l2arc_write_callback_t *cb;
7586 uint64_t guid = spa_load_guid(spa);
7589 ASSERT3P(dev->l2ad_vdev, !=, NULL);
7592 write_lsize = write_asize = write_psize = 0;
7594 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
7595 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
7597 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
7599 * Copy buffers for L2ARC writing.
7601 for (try = 0; try <= 3; try++) {
7602 multilist_sublist_t *mls = l2arc_sublist_lock(try);
7603 uint64_t passed_sz = 0;
7605 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
7608 * L2ARC fast warmup.
7610 * Until the ARC is warm and starts to evict, read from the
7611 * head of the ARC lists rather than the tail.
7613 if (arc_warm == B_FALSE)
7614 hdr = multilist_sublist_head(mls);
7616 hdr = multilist_sublist_tail(mls);
7618 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
7620 headroom = target_sz * l2arc_headroom;
7621 if (zfs_compressed_arc_enabled)
7622 headroom = (headroom * l2arc_headroom_boost) / 100;
7624 for (; hdr; hdr = hdr_prev) {
7625 kmutex_t *hash_lock;
7627 if (arc_warm == B_FALSE)
7628 hdr_prev = multilist_sublist_next(mls, hdr);
7630 hdr_prev = multilist_sublist_prev(mls, hdr);
7631 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned,
7632 HDR_GET_LSIZE(hdr));
7634 hash_lock = HDR_LOCK(hdr);
7635 if (!mutex_tryenter(hash_lock)) {
7636 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
7638 * Skip this buffer rather than waiting.
7643 passed_sz += HDR_GET_LSIZE(hdr);
7644 if (passed_sz > headroom) {
7648 mutex_exit(hash_lock);
7649 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
7653 if (!l2arc_write_eligible(guid, hdr)) {
7654 mutex_exit(hash_lock);
7659 * We rely on the L1 portion of the header below, so
7660 * it's invalid for this header to have been evicted out
7661 * of the ghost cache, prior to being written out. The
7662 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7664 ASSERT(HDR_HAS_L1HDR(hdr));
7666 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
7667 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7668 ASSERT3U(arc_hdr_size(hdr), >, 0);
7669 uint64_t psize = arc_hdr_size(hdr);
7670 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
7673 if ((write_asize + asize) > target_sz) {
7675 mutex_exit(hash_lock);
7676 ARCSTAT_BUMP(arcstat_l2_write_full);
7682 * Insert a dummy header on the buflist so
7683 * l2arc_write_done() can find where the
7684 * write buffers begin without searching.
7686 mutex_enter(&dev->l2ad_mtx);
7687 list_insert_head(&dev->l2ad_buflist, head);
7688 mutex_exit(&dev->l2ad_mtx);
7691 sizeof (l2arc_write_callback_t), KM_SLEEP);
7692 cb->l2wcb_dev = dev;
7693 cb->l2wcb_head = head;
7694 pio = zio_root(spa, l2arc_write_done, cb,
7696 ARCSTAT_BUMP(arcstat_l2_write_pios);
7699 hdr->b_l2hdr.b_dev = dev;
7700 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7701 arc_hdr_set_flags(hdr,
7702 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7704 mutex_enter(&dev->l2ad_mtx);
7705 list_insert_head(&dev->l2ad_buflist, hdr);
7706 mutex_exit(&dev->l2ad_mtx);
7708 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
7711 * Normally the L2ARC can use the hdr's data, but if
7712 * we're sharing data between the hdr and one of its
7713 * bufs, L2ARC needs its own copy of the data so that
7714 * the ZIO below can't race with the buf consumer.
7715 * Another case where we need to create a copy of the
7716 * data is when the buffer size is not device-aligned
7717 * and we need to pad the block to make it such.
7718 * That also keeps the clock hand suitably aligned.
7720 * To ensure that the copy will be available for the
7721 * lifetime of the ZIO and be cleaned up afterwards, we
7722 * add it to the l2arc_free_on_write queue.
7725 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
7726 to_write = hdr->b_l1hdr.b_pabd;
7728 to_write = abd_alloc_for_io(asize,
7729 HDR_ISTYPE_METADATA(hdr));
7730 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
7731 if (asize != psize) {
7732 abd_zero_off(to_write, psize,
7735 l2arc_free_abd_on_write(to_write, asize,
7738 wzio = zio_write_phys(pio, dev->l2ad_vdev,
7739 hdr->b_l2hdr.b_daddr, asize, to_write,
7740 ZIO_CHECKSUM_OFF, NULL, hdr,
7741 ZIO_PRIORITY_ASYNC_WRITE,
7742 ZIO_FLAG_CANFAIL, B_FALSE);
7744 write_lsize += HDR_GET_LSIZE(hdr);
7745 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7748 write_psize += psize;
7749 write_asize += asize;
7750 dev->l2ad_hand += asize;
7752 mutex_exit(hash_lock);
7754 (void) zio_nowait(wzio);
7757 multilist_sublist_unlock(mls);
7763 /* No buffers selected for writing? */
7765 ASSERT0(write_lsize);
7766 ASSERT(!HDR_HAS_L1HDR(head));
7767 kmem_cache_free(hdr_l2only_cache, head);
7771 ASSERT3U(write_psize, <=, target_sz);
7772 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7773 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
7774 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
7775 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
7776 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
7779 * Bump device hand to the device start if it is approaching the end.
7780 * l2arc_evict() will already have evicted ahead for this case.
7782 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7783 dev->l2ad_hand = dev->l2ad_start;
7784 dev->l2ad_first = B_FALSE;
7787 dev->l2ad_writing = B_TRUE;
7788 (void) zio_wait(pio);
7789 dev->l2ad_writing = B_FALSE;
7791 return (write_asize);
7795 * This thread feeds the L2ARC at regular intervals. This is the beating
7796 * heart of the L2ARC.
7800 l2arc_feed_thread(void *unused __unused)
7805 uint64_t size, wrote;
7806 clock_t begin, next = ddi_get_lbolt();
7808 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7810 mutex_enter(&l2arc_feed_thr_lock);
7812 while (l2arc_thread_exit == 0) {
7813 CALLB_CPR_SAFE_BEGIN(&cpr);
7814 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7815 next - ddi_get_lbolt());
7816 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7817 next = ddi_get_lbolt() + hz;
7820 * Quick check for L2ARC devices.
7822 mutex_enter(&l2arc_dev_mtx);
7823 if (l2arc_ndev == 0) {
7824 mutex_exit(&l2arc_dev_mtx);
7827 mutex_exit(&l2arc_dev_mtx);
7828 begin = ddi_get_lbolt();
7831 * This selects the next l2arc device to write to, and in
7832 * doing so the next spa to feed from: dev->l2ad_spa. This
7833 * will return NULL if there are now no l2arc devices or if
7834 * they are all faulted.
7836 * If a device is returned, its spa's config lock is also
7837 * held to prevent device removal. l2arc_dev_get_next()
7838 * will grab and release l2arc_dev_mtx.
7840 if ((dev = l2arc_dev_get_next()) == NULL)
7843 spa = dev->l2ad_spa;
7844 ASSERT3P(spa, !=, NULL);
7847 * If the pool is read-only then force the feed thread to
7848 * sleep a little longer.
7850 if (!spa_writeable(spa)) {
7851 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7852 spa_config_exit(spa, SCL_L2ARC, dev);
7857 * Avoid contributing to memory pressure.
7859 if (arc_reclaim_needed()) {
7860 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7861 spa_config_exit(spa, SCL_L2ARC, dev);
7865 ARCSTAT_BUMP(arcstat_l2_feeds);
7867 size = l2arc_write_size();
7870 * Evict L2ARC buffers that will be overwritten.
7872 l2arc_evict(dev, size, B_FALSE);
7875 * Write ARC buffers.
7877 wrote = l2arc_write_buffers(spa, dev, size);
7880 * Calculate interval between writes.
7882 next = l2arc_write_interval(begin, size, wrote);
7883 spa_config_exit(spa, SCL_L2ARC, dev);
7886 l2arc_thread_exit = 0;
7887 cv_broadcast(&l2arc_feed_thr_cv);
7888 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7893 l2arc_vdev_present(vdev_t *vd)
7897 mutex_enter(&l2arc_dev_mtx);
7898 for (dev = list_head(l2arc_dev_list); dev != NULL;
7899 dev = list_next(l2arc_dev_list, dev)) {
7900 if (dev->l2ad_vdev == vd)
7903 mutex_exit(&l2arc_dev_mtx);
7905 return (dev != NULL);
7909 * Add a vdev for use by the L2ARC. By this point the spa has already
7910 * validated the vdev and opened it.
7913 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7915 l2arc_dev_t *adddev;
7917 ASSERT(!l2arc_vdev_present(vd));
7919 vdev_ashift_optimize(vd);
7922 * Create a new l2arc device entry.
7924 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7925 adddev->l2ad_spa = spa;
7926 adddev->l2ad_vdev = vd;
7927 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7928 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7929 adddev->l2ad_hand = adddev->l2ad_start;
7930 adddev->l2ad_first = B_TRUE;
7931 adddev->l2ad_writing = B_FALSE;
7933 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7935 * This is a list of all ARC buffers that are still valid on the
7938 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7939 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7941 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7942 refcount_create(&adddev->l2ad_alloc);
7945 * Add device to global list
7947 mutex_enter(&l2arc_dev_mtx);
7948 list_insert_head(l2arc_dev_list, adddev);
7949 atomic_inc_64(&l2arc_ndev);
7950 mutex_exit(&l2arc_dev_mtx);
7954 * Remove a vdev from the L2ARC.
7957 l2arc_remove_vdev(vdev_t *vd)
7959 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7962 * Find the device by vdev
7964 mutex_enter(&l2arc_dev_mtx);
7965 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7966 nextdev = list_next(l2arc_dev_list, dev);
7967 if (vd == dev->l2ad_vdev) {
7972 ASSERT3P(remdev, !=, NULL);
7975 * Remove device from global list
7977 list_remove(l2arc_dev_list, remdev);
7978 l2arc_dev_last = NULL; /* may have been invalidated */
7979 atomic_dec_64(&l2arc_ndev);
7980 mutex_exit(&l2arc_dev_mtx);
7983 * Clear all buflists and ARC references. L2ARC device flush.
7985 l2arc_evict(remdev, 0, B_TRUE);
7986 list_destroy(&remdev->l2ad_buflist);
7987 mutex_destroy(&remdev->l2ad_mtx);
7988 refcount_destroy(&remdev->l2ad_alloc);
7989 kmem_free(remdev, sizeof (l2arc_dev_t));
7995 l2arc_thread_exit = 0;
7997 l2arc_writes_sent = 0;
7998 l2arc_writes_done = 0;
8000 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
8001 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
8002 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
8003 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
8005 l2arc_dev_list = &L2ARC_dev_list;
8006 l2arc_free_on_write = &L2ARC_free_on_write;
8007 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
8008 offsetof(l2arc_dev_t, l2ad_node));
8009 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
8010 offsetof(l2arc_data_free_t, l2df_list_node));
8017 * This is called from dmu_fini(), which is called from spa_fini();
8018 * Because of this, we can assume that all l2arc devices have
8019 * already been removed when the pools themselves were removed.
8022 l2arc_do_free_on_write();
8024 mutex_destroy(&l2arc_feed_thr_lock);
8025 cv_destroy(&l2arc_feed_thr_cv);
8026 mutex_destroy(&l2arc_dev_mtx);
8027 mutex_destroy(&l2arc_free_on_write_mtx);
8029 list_destroy(l2arc_dev_list);
8030 list_destroy(l2arc_free_on_write);
8036 if (!(spa_mode_global & FWRITE))
8039 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
8040 TS_RUN, minclsyspri);
8046 if (!(spa_mode_global & FWRITE))
8049 mutex_enter(&l2arc_feed_thr_lock);
8050 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
8051 l2arc_thread_exit = 1;
8052 while (l2arc_thread_exit != 0)
8053 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
8054 mutex_exit(&l2arc_feed_thr_lock);