4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2014 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 * Arc buffers may have an associated eviction callback function.
103 * This function will be invoked prior to removing the buffer (e.g.
104 * in arc_do_user_evicts()). Note however that the data associated
105 * with the buffer may be evicted prior to the callback. The callback
106 * must be made with *no locks held* (to prevent deadlock). Additionally,
107 * the users of callbacks must ensure that their private data is
108 * protected from simultaneous callbacks from arc_clear_callback()
109 * and arc_do_user_evicts().
111 * It as also possible to register a callback which is run when the
112 * arc_meta_limit is reached and no buffers can be safely evicted. In
113 * this case the arc user should drop a reference on some arc buffers so
114 * they can be reclaimed and the arc_meta_limit honored. For example,
115 * when using the ZPL each dentry holds a references on a znode. These
116 * dentries must be pruned before the arc buffer holding the znode can
119 * Note that the majority of the performance stats are manipulated
120 * with atomic operations.
122 * The L2ARC uses the l2ad_mtx on each vdev for the following:
124 * - L2ARC buflist creation
125 * - L2ARC buflist eviction
126 * - L2ARC write completion, which walks L2ARC buflists
127 * - ARC header destruction, as it removes from L2ARC buflists
128 * - ARC header release, as it removes from L2ARC buflists
133 #include <sys/zio_compress.h>
134 #include <sys/zfs_context.h>
136 #include <sys/refcount.h>
137 #include <sys/vdev.h>
138 #include <sys/vdev_impl.h>
139 #include <sys/dsl_pool.h>
140 #include <sys/multilist.h>
142 #include <sys/vmsystm.h>
144 #include <sys/fs/swapnode.h>
146 #include <linux/mm_compat.h>
148 #include <sys/callb.h>
149 #include <sys/kstat.h>
150 #include <sys/dmu_tx.h>
151 #include <zfs_fletcher.h>
152 #include <sys/arc_impl.h>
153 #include <sys/trace_arc.h>
156 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
157 boolean_t arc_watch
= B_FALSE
;
160 static kmutex_t arc_reclaim_lock
;
161 static kcondvar_t arc_reclaim_thread_cv
;
162 static boolean_t arc_reclaim_thread_exit
;
163 static kcondvar_t arc_reclaim_waiters_cv
;
165 static kmutex_t arc_user_evicts_lock
;
166 static kcondvar_t arc_user_evicts_cv
;
167 static boolean_t arc_user_evicts_thread_exit
;
170 * The number of headers to evict in arc_evict_state_impl() before
171 * dropping the sublist lock and evicting from another sublist. A lower
172 * value means we're more likely to evict the "correct" header (i.e. the
173 * oldest header in the arc state), but comes with higher overhead
174 * (i.e. more invocations of arc_evict_state_impl()).
176 int zfs_arc_evict_batch_limit
= 10;
179 * The number of sublists used for each of the arc state lists. If this
180 * is not set to a suitable value by the user, it will be configured to
181 * the number of CPUs on the system in arc_init().
183 int zfs_arc_num_sublists_per_state
= 0;
185 /* number of seconds before growing cache again */
186 static int arc_grow_retry
= 5;
188 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
189 int zfs_arc_overflow_shift
= 8;
191 /* shift of arc_c for calculating both min and max arc_p */
192 static int arc_p_min_shift
= 4;
194 /* log2(fraction of arc to reclaim) */
195 static int arc_shrink_shift
= 7;
198 * log2(fraction of ARC which must be free to allow growing).
199 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
200 * when reading a new block into the ARC, we will evict an equal-sized block
203 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
204 * we will still not allow it to grow.
206 int arc_no_grow_shift
= 5;
210 * minimum lifespan of a prefetch block in clock ticks
211 * (initialized in arc_init())
213 static int arc_min_prefetch_lifespan
;
216 * If this percent of memory is free, don't throttle.
218 int arc_lotsfree_percent
= 10;
223 * The arc has filled available memory and has now warmed up.
225 static boolean_t arc_warm
;
228 * These tunables are for performance analysis.
230 unsigned long zfs_arc_max
= 0;
231 unsigned long zfs_arc_min
= 0;
232 unsigned long zfs_arc_meta_limit
= 0;
233 unsigned long zfs_arc_meta_min
= 0;
234 int zfs_arc_grow_retry
= 0;
235 int zfs_arc_shrink_shift
= 0;
236 int zfs_arc_p_min_shift
= 0;
237 int zfs_disable_dup_eviction
= 0;
238 int zfs_arc_average_blocksize
= 8 * 1024; /* 8KB */
241 * These tunables are Linux specific
243 int zfs_arc_memory_throttle_disable
= 1;
244 int zfs_arc_min_prefetch_lifespan
= 0;
245 int zfs_arc_p_aggressive_disable
= 1;
246 int zfs_arc_p_dampener_disable
= 1;
247 int zfs_arc_meta_prune
= 10000;
248 int zfs_arc_meta_strategy
= ARC_STRATEGY_META_BALANCED
;
249 int zfs_arc_meta_adjust_restarts
= 4096;
252 static arc_state_t ARC_anon
;
253 static arc_state_t ARC_mru
;
254 static arc_state_t ARC_mru_ghost
;
255 static arc_state_t ARC_mfu
;
256 static arc_state_t ARC_mfu_ghost
;
257 static arc_state_t ARC_l2c_only
;
259 typedef struct arc_stats
{
260 kstat_named_t arcstat_hits
;
261 kstat_named_t arcstat_misses
;
262 kstat_named_t arcstat_demand_data_hits
;
263 kstat_named_t arcstat_demand_data_misses
;
264 kstat_named_t arcstat_demand_metadata_hits
;
265 kstat_named_t arcstat_demand_metadata_misses
;
266 kstat_named_t arcstat_prefetch_data_hits
;
267 kstat_named_t arcstat_prefetch_data_misses
;
268 kstat_named_t arcstat_prefetch_metadata_hits
;
269 kstat_named_t arcstat_prefetch_metadata_misses
;
270 kstat_named_t arcstat_mru_hits
;
271 kstat_named_t arcstat_mru_ghost_hits
;
272 kstat_named_t arcstat_mfu_hits
;
273 kstat_named_t arcstat_mfu_ghost_hits
;
274 kstat_named_t arcstat_deleted
;
276 * Number of buffers that could not be evicted because the hash lock
277 * was held by another thread. The lock may not necessarily be held
278 * by something using the same buffer, since hash locks are shared
279 * by multiple buffers.
281 kstat_named_t arcstat_mutex_miss
;
283 * Number of buffers skipped because they have I/O in progress, are
284 * indrect prefetch buffers that have not lived long enough, or are
285 * not from the spa we're trying to evict from.
287 kstat_named_t arcstat_evict_skip
;
289 * Number of times arc_evict_state() was unable to evict enough
290 * buffers to reach its target amount.
292 kstat_named_t arcstat_evict_not_enough
;
293 kstat_named_t arcstat_evict_l2_cached
;
294 kstat_named_t arcstat_evict_l2_eligible
;
295 kstat_named_t arcstat_evict_l2_ineligible
;
296 kstat_named_t arcstat_evict_l2_skip
;
297 kstat_named_t arcstat_hash_elements
;
298 kstat_named_t arcstat_hash_elements_max
;
299 kstat_named_t arcstat_hash_collisions
;
300 kstat_named_t arcstat_hash_chains
;
301 kstat_named_t arcstat_hash_chain_max
;
302 kstat_named_t arcstat_p
;
303 kstat_named_t arcstat_c
;
304 kstat_named_t arcstat_c_min
;
305 kstat_named_t arcstat_c_max
;
306 kstat_named_t arcstat_size
;
308 * Number of bytes consumed by internal ARC structures necessary
309 * for tracking purposes; these structures are not actually
310 * backed by ARC buffers. This includes arc_buf_hdr_t structures
311 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
312 * caches), and arc_buf_t structures (allocated via arc_buf_t
315 kstat_named_t arcstat_hdr_size
;
317 * Number of bytes consumed by ARC buffers of type equal to
318 * ARC_BUFC_DATA. This is generally consumed by buffers backing
319 * on disk user data (e.g. plain file contents).
321 kstat_named_t arcstat_data_size
;
323 * Number of bytes consumed by ARC buffers of type equal to
324 * ARC_BUFC_METADATA. This is generally consumed by buffers
325 * backing on disk data that is used for internal ZFS
326 * structures (e.g. ZAP, dnode, indirect blocks, etc).
328 kstat_named_t arcstat_metadata_size
;
330 * Number of bytes consumed by various buffers and structures
331 * not actually backed with ARC buffers. This includes bonus
332 * buffers (allocated directly via zio_buf_* functions),
333 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
334 * cache), and dnode_t structures (allocated via dnode_t cache).
336 kstat_named_t arcstat_other_size
;
338 * Total number of bytes consumed by ARC buffers residing in the
339 * arc_anon state. This includes *all* buffers in the arc_anon
340 * state; e.g. data, metadata, evictable, and unevictable buffers
341 * are all included in this value.
343 kstat_named_t arcstat_anon_size
;
345 * Number of bytes consumed by ARC buffers that meet the
346 * following criteria: backing buffers of type ARC_BUFC_DATA,
347 * residing in the arc_anon state, and are eligible for eviction
348 * (e.g. have no outstanding holds on the buffer).
350 kstat_named_t arcstat_anon_evictable_data
;
352 * Number of bytes consumed by ARC buffers that meet the
353 * following criteria: backing buffers of type ARC_BUFC_METADATA,
354 * residing in the arc_anon state, and are eligible for eviction
355 * (e.g. have no outstanding holds on the buffer).
357 kstat_named_t arcstat_anon_evictable_metadata
;
359 * Total number of bytes consumed by ARC buffers residing in the
360 * arc_mru state. This includes *all* buffers in the arc_mru
361 * state; e.g. data, metadata, evictable, and unevictable buffers
362 * are all included in this value.
364 kstat_named_t arcstat_mru_size
;
366 * Number of bytes consumed by ARC buffers that meet the
367 * following criteria: backing buffers of type ARC_BUFC_DATA,
368 * residing in the arc_mru state, and are eligible for eviction
369 * (e.g. have no outstanding holds on the buffer).
371 kstat_named_t arcstat_mru_evictable_data
;
373 * Number of bytes consumed by ARC buffers that meet the
374 * following criteria: backing buffers of type ARC_BUFC_METADATA,
375 * residing in the arc_mru state, and are eligible for eviction
376 * (e.g. have no outstanding holds on the buffer).
378 kstat_named_t arcstat_mru_evictable_metadata
;
380 * Total number of bytes that *would have been* consumed by ARC
381 * buffers in the arc_mru_ghost state. The key thing to note
382 * here, is the fact that this size doesn't actually indicate
383 * RAM consumption. The ghost lists only consist of headers and
384 * don't actually have ARC buffers linked off of these headers.
385 * Thus, *if* the headers had associated ARC buffers, these
386 * buffers *would have* consumed this number of bytes.
388 kstat_named_t arcstat_mru_ghost_size
;
390 * Number of bytes that *would have been* consumed by ARC
391 * buffers that are eligible for eviction, of type
392 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
394 kstat_named_t arcstat_mru_ghost_evictable_data
;
396 * Number of bytes that *would have been* consumed by ARC
397 * buffers that are eligible for eviction, of type
398 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
400 kstat_named_t arcstat_mru_ghost_evictable_metadata
;
402 * Total number of bytes consumed by ARC buffers residing in the
403 * arc_mfu state. This includes *all* buffers in the arc_mfu
404 * state; e.g. data, metadata, evictable, and unevictable buffers
405 * are all included in this value.
407 kstat_named_t arcstat_mfu_size
;
409 * Number of bytes consumed by ARC buffers that are eligible for
410 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
413 kstat_named_t arcstat_mfu_evictable_data
;
415 * Number of bytes consumed by ARC buffers that are eligible for
416 * eviction, of type ARC_BUFC_METADATA, and reside in the
419 kstat_named_t arcstat_mfu_evictable_metadata
;
421 * Total number of bytes that *would have been* consumed by ARC
422 * buffers in the arc_mfu_ghost state. See the comment above
423 * arcstat_mru_ghost_size for more details.
425 kstat_named_t arcstat_mfu_ghost_size
;
427 * Number of bytes that *would have been* consumed by ARC
428 * buffers that are eligible for eviction, of type
429 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
431 kstat_named_t arcstat_mfu_ghost_evictable_data
;
433 * Number of bytes that *would have been* consumed by ARC
434 * buffers that are eligible for eviction, of type
435 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
437 kstat_named_t arcstat_mfu_ghost_evictable_metadata
;
438 kstat_named_t arcstat_l2_hits
;
439 kstat_named_t arcstat_l2_misses
;
440 kstat_named_t arcstat_l2_feeds
;
441 kstat_named_t arcstat_l2_rw_clash
;
442 kstat_named_t arcstat_l2_read_bytes
;
443 kstat_named_t arcstat_l2_write_bytes
;
444 kstat_named_t arcstat_l2_writes_sent
;
445 kstat_named_t arcstat_l2_writes_done
;
446 kstat_named_t arcstat_l2_writes_error
;
447 kstat_named_t arcstat_l2_writes_lock_retry
;
448 kstat_named_t arcstat_l2_evict_lock_retry
;
449 kstat_named_t arcstat_l2_evict_reading
;
450 kstat_named_t arcstat_l2_evict_l1cached
;
451 kstat_named_t arcstat_l2_free_on_write
;
452 kstat_named_t arcstat_l2_cdata_free_on_write
;
453 kstat_named_t arcstat_l2_abort_lowmem
;
454 kstat_named_t arcstat_l2_cksum_bad
;
455 kstat_named_t arcstat_l2_io_error
;
456 kstat_named_t arcstat_l2_size
;
457 kstat_named_t arcstat_l2_asize
;
458 kstat_named_t arcstat_l2_hdr_size
;
459 kstat_named_t arcstat_l2_compress_successes
;
460 kstat_named_t arcstat_l2_compress_zeros
;
461 kstat_named_t arcstat_l2_compress_failures
;
462 kstat_named_t arcstat_memory_throttle_count
;
463 kstat_named_t arcstat_duplicate_buffers
;
464 kstat_named_t arcstat_duplicate_buffers_size
;
465 kstat_named_t arcstat_duplicate_reads
;
466 kstat_named_t arcstat_memory_direct_count
;
467 kstat_named_t arcstat_memory_indirect_count
;
468 kstat_named_t arcstat_no_grow
;
469 kstat_named_t arcstat_tempreserve
;
470 kstat_named_t arcstat_loaned_bytes
;
471 kstat_named_t arcstat_prune
;
472 kstat_named_t arcstat_meta_used
;
473 kstat_named_t arcstat_meta_limit
;
474 kstat_named_t arcstat_meta_max
;
475 kstat_named_t arcstat_meta_min
;
478 static arc_stats_t arc_stats
= {
479 { "hits", KSTAT_DATA_UINT64
},
480 { "misses", KSTAT_DATA_UINT64
},
481 { "demand_data_hits", KSTAT_DATA_UINT64
},
482 { "demand_data_misses", KSTAT_DATA_UINT64
},
483 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
484 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
485 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
486 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
487 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
488 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
489 { "mru_hits", KSTAT_DATA_UINT64
},
490 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
491 { "mfu_hits", KSTAT_DATA_UINT64
},
492 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
493 { "deleted", KSTAT_DATA_UINT64
},
494 { "mutex_miss", KSTAT_DATA_UINT64
},
495 { "evict_skip", KSTAT_DATA_UINT64
},
496 { "evict_not_enough", KSTAT_DATA_UINT64
},
497 { "evict_l2_cached", KSTAT_DATA_UINT64
},
498 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
499 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
500 { "evict_l2_skip", KSTAT_DATA_UINT64
},
501 { "hash_elements", KSTAT_DATA_UINT64
},
502 { "hash_elements_max", KSTAT_DATA_UINT64
},
503 { "hash_collisions", KSTAT_DATA_UINT64
},
504 { "hash_chains", KSTAT_DATA_UINT64
},
505 { "hash_chain_max", KSTAT_DATA_UINT64
},
506 { "p", KSTAT_DATA_UINT64
},
507 { "c", KSTAT_DATA_UINT64
},
508 { "c_min", KSTAT_DATA_UINT64
},
509 { "c_max", KSTAT_DATA_UINT64
},
510 { "size", KSTAT_DATA_UINT64
},
511 { "hdr_size", KSTAT_DATA_UINT64
},
512 { "data_size", KSTAT_DATA_UINT64
},
513 { "metadata_size", KSTAT_DATA_UINT64
},
514 { "other_size", KSTAT_DATA_UINT64
},
515 { "anon_size", KSTAT_DATA_UINT64
},
516 { "anon_evictable_data", KSTAT_DATA_UINT64
},
517 { "anon_evictable_metadata", KSTAT_DATA_UINT64
},
518 { "mru_size", KSTAT_DATA_UINT64
},
519 { "mru_evictable_data", KSTAT_DATA_UINT64
},
520 { "mru_evictable_metadata", KSTAT_DATA_UINT64
},
521 { "mru_ghost_size", KSTAT_DATA_UINT64
},
522 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64
},
523 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
524 { "mfu_size", KSTAT_DATA_UINT64
},
525 { "mfu_evictable_data", KSTAT_DATA_UINT64
},
526 { "mfu_evictable_metadata", KSTAT_DATA_UINT64
},
527 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
528 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64
},
529 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
530 { "l2_hits", KSTAT_DATA_UINT64
},
531 { "l2_misses", KSTAT_DATA_UINT64
},
532 { "l2_feeds", KSTAT_DATA_UINT64
},
533 { "l2_rw_clash", KSTAT_DATA_UINT64
},
534 { "l2_read_bytes", KSTAT_DATA_UINT64
},
535 { "l2_write_bytes", KSTAT_DATA_UINT64
},
536 { "l2_writes_sent", KSTAT_DATA_UINT64
},
537 { "l2_writes_done", KSTAT_DATA_UINT64
},
538 { "l2_writes_error", KSTAT_DATA_UINT64
},
539 { "l2_writes_lock_retry", KSTAT_DATA_UINT64
},
540 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
541 { "l2_evict_reading", KSTAT_DATA_UINT64
},
542 { "l2_evict_l1cached", KSTAT_DATA_UINT64
},
543 { "l2_free_on_write", KSTAT_DATA_UINT64
},
544 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64
},
545 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
546 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
547 { "l2_io_error", KSTAT_DATA_UINT64
},
548 { "l2_size", KSTAT_DATA_UINT64
},
549 { "l2_asize", KSTAT_DATA_UINT64
},
550 { "l2_hdr_size", KSTAT_DATA_UINT64
},
551 { "l2_compress_successes", KSTAT_DATA_UINT64
},
552 { "l2_compress_zeros", KSTAT_DATA_UINT64
},
553 { "l2_compress_failures", KSTAT_DATA_UINT64
},
554 { "memory_throttle_count", KSTAT_DATA_UINT64
},
555 { "duplicate_buffers", KSTAT_DATA_UINT64
},
556 { "duplicate_buffers_size", KSTAT_DATA_UINT64
},
557 { "duplicate_reads", KSTAT_DATA_UINT64
},
558 { "memory_direct_count", KSTAT_DATA_UINT64
},
559 { "memory_indirect_count", KSTAT_DATA_UINT64
},
560 { "arc_no_grow", KSTAT_DATA_UINT64
},
561 { "arc_tempreserve", KSTAT_DATA_UINT64
},
562 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
563 { "arc_prune", KSTAT_DATA_UINT64
},
564 { "arc_meta_used", KSTAT_DATA_UINT64
},
565 { "arc_meta_limit", KSTAT_DATA_UINT64
},
566 { "arc_meta_max", KSTAT_DATA_UINT64
},
567 { "arc_meta_min", KSTAT_DATA_UINT64
}
570 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
572 #define ARCSTAT_INCR(stat, val) \
573 atomic_add_64(&arc_stats.stat.value.ui64, (val))
575 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
576 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
578 #define ARCSTAT_MAX(stat, val) { \
580 while ((val) > (m = arc_stats.stat.value.ui64) && \
581 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
585 #define ARCSTAT_MAXSTAT(stat) \
586 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
589 * We define a macro to allow ARC hits/misses to be easily broken down by
590 * two separate conditions, giving a total of four different subtypes for
591 * each of hits and misses (so eight statistics total).
593 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
596 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
598 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
602 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
604 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
609 static arc_state_t
*arc_anon
;
610 static arc_state_t
*arc_mru
;
611 static arc_state_t
*arc_mru_ghost
;
612 static arc_state_t
*arc_mfu
;
613 static arc_state_t
*arc_mfu_ghost
;
614 static arc_state_t
*arc_l2c_only
;
617 * There are several ARC variables that are critical to export as kstats --
618 * but we don't want to have to grovel around in the kstat whenever we wish to
619 * manipulate them. For these variables, we therefore define them to be in
620 * terms of the statistic variable. This assures that we are not introducing
621 * the possibility of inconsistency by having shadow copies of the variables,
622 * while still allowing the code to be readable.
624 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
625 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
626 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
627 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
628 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
629 #define arc_no_grow ARCSTAT(arcstat_no_grow)
630 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
631 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
632 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
633 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
634 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
635 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
637 #define L2ARC_IS_VALID_COMPRESS(_c_) \
638 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
640 static list_t arc_prune_list
;
641 static kmutex_t arc_prune_mtx
;
642 static taskq_t
*arc_prune_taskq
;
643 static arc_buf_t
*arc_eviction_list
;
644 static arc_buf_hdr_t arc_eviction_hdr
;
646 #define GHOST_STATE(state) \
647 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
648 (state) == arc_l2c_only)
650 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
651 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
652 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
653 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
654 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
655 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
657 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
658 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
659 #define HDR_L2_READING(hdr) \
660 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
661 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
662 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
663 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
664 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
666 #define HDR_ISTYPE_METADATA(hdr) \
667 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
668 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
670 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
671 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
673 /* For storing compression mode in b_flags */
674 #define HDR_COMPRESS_OFFSET 24
675 #define HDR_COMPRESS_NBITS 7
677 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \
678 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS))
679 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \
680 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp))
686 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
687 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
690 * Hash table routines
693 #define HT_LOCK_ALIGN 64
694 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
699 unsigned char pad
[HT_LOCK_PAD
];
703 #define BUF_LOCKS 8192
704 typedef struct buf_hash_table
{
706 arc_buf_hdr_t
**ht_table
;
707 struct ht_lock ht_locks
[BUF_LOCKS
];
710 static buf_hash_table_t buf_hash_table
;
712 #define BUF_HASH_INDEX(spa, dva, birth) \
713 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
714 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
715 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
716 #define HDR_LOCK(hdr) \
717 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
719 uint64_t zfs_crc64_table
[256];
725 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
726 #define L2ARC_HEADROOM 2 /* num of writes */
728 * If we discover during ARC scan any buffers to be compressed, we boost
729 * our headroom for the next scanning cycle by this percentage multiple.
731 #define L2ARC_HEADROOM_BOOST 200
732 #define L2ARC_FEED_SECS 1 /* caching interval secs */
733 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
736 * Used to distinguish headers that are being process by
737 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
738 * address. This can happen when the header is added to the l2arc's list
739 * of buffers to write in the first stage of l2arc_write_buffers(), but
740 * has not yet been written out which happens in the second stage of
741 * l2arc_write_buffers().
743 #define L2ARC_ADDR_UNSET ((uint64_t)(-1))
745 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
746 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
748 /* L2ARC Performance Tunables */
749 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
750 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
751 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
752 unsigned long l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
753 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
754 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
755 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
756 int l2arc_nocompress
= B_FALSE
; /* don't compress bufs */
757 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
758 int l2arc_norw
= B_FALSE
; /* no reads during writes */
763 static list_t L2ARC_dev_list
; /* device list */
764 static list_t
*l2arc_dev_list
; /* device list pointer */
765 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
766 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
767 static list_t L2ARC_free_on_write
; /* free after write buf list */
768 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
769 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
770 static uint64_t l2arc_ndev
; /* number of devices */
772 typedef struct l2arc_read_callback
{
773 arc_buf_t
*l2rcb_buf
; /* read buffer */
774 spa_t
*l2rcb_spa
; /* spa */
775 blkptr_t l2rcb_bp
; /* original blkptr */
776 zbookmark_phys_t l2rcb_zb
; /* original bookmark */
777 int l2rcb_flags
; /* original flags */
778 enum zio_compress l2rcb_compress
; /* applied compress */
779 } l2arc_read_callback_t
;
781 typedef struct l2arc_data_free
{
782 /* protected by l2arc_free_on_write_mtx */
785 void (*l2df_func
)(void *, size_t);
786 list_node_t l2df_list_node
;
789 static kmutex_t l2arc_feed_thr_lock
;
790 static kcondvar_t l2arc_feed_thr_cv
;
791 static uint8_t l2arc_thread_exit
;
793 static void arc_get_data_buf(arc_buf_t
*);
794 static void arc_access(arc_buf_hdr_t
*, kmutex_t
*);
795 static boolean_t
arc_is_overflowing(void);
796 static void arc_buf_watch(arc_buf_t
*);
797 static void arc_tuning_update(void);
799 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t
*);
800 static uint32_t arc_bufc_to_flags(arc_buf_contents_t
);
802 static boolean_t
l2arc_write_eligible(uint64_t, arc_buf_hdr_t
*);
803 static void l2arc_read_done(zio_t
*);
805 static boolean_t
l2arc_compress_buf(arc_buf_hdr_t
*);
806 static void l2arc_decompress_zio(zio_t
*, arc_buf_hdr_t
*, enum zio_compress
);
807 static void l2arc_release_cdata_buf(arc_buf_hdr_t
*);
810 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
812 uint8_t *vdva
= (uint8_t *)dva
;
813 uint64_t crc
= -1ULL;
816 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
818 for (i
= 0; i
< sizeof (dva_t
); i
++)
819 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
821 crc
^= (spa
>>8) ^ birth
;
826 #define BUF_EMPTY(buf) \
827 ((buf)->b_dva.dva_word[0] == 0 && \
828 (buf)->b_dva.dva_word[1] == 0)
830 #define BUF_EQUAL(spa, dva, birth, buf) \
831 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
832 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
833 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
836 buf_discard_identity(arc_buf_hdr_t
*hdr
)
838 hdr
->b_dva
.dva_word
[0] = 0;
839 hdr
->b_dva
.dva_word
[1] = 0;
843 static arc_buf_hdr_t
*
844 buf_hash_find(uint64_t spa
, const blkptr_t
*bp
, kmutex_t
**lockp
)
846 const dva_t
*dva
= BP_IDENTITY(bp
);
847 uint64_t birth
= BP_PHYSICAL_BIRTH(bp
);
848 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
849 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
852 mutex_enter(hash_lock
);
853 for (hdr
= buf_hash_table
.ht_table
[idx
]; hdr
!= NULL
;
854 hdr
= hdr
->b_hash_next
) {
855 if (BUF_EQUAL(spa
, dva
, birth
, hdr
)) {
860 mutex_exit(hash_lock
);
866 * Insert an entry into the hash table. If there is already an element
867 * equal to elem in the hash table, then the already existing element
868 * will be returned and the new element will not be inserted.
869 * Otherwise returns NULL.
870 * If lockp == NULL, the caller is assumed to already hold the hash lock.
872 static arc_buf_hdr_t
*
873 buf_hash_insert(arc_buf_hdr_t
*hdr
, kmutex_t
**lockp
)
875 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
876 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
880 ASSERT(!DVA_IS_EMPTY(&hdr
->b_dva
));
881 ASSERT(hdr
->b_birth
!= 0);
882 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
886 mutex_enter(hash_lock
);
888 ASSERT(MUTEX_HELD(hash_lock
));
891 for (fhdr
= buf_hash_table
.ht_table
[idx
], i
= 0; fhdr
!= NULL
;
892 fhdr
= fhdr
->b_hash_next
, i
++) {
893 if (BUF_EQUAL(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
, fhdr
))
897 hdr
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
898 buf_hash_table
.ht_table
[idx
] = hdr
;
899 hdr
->b_flags
|= ARC_FLAG_IN_HASH_TABLE
;
901 /* collect some hash table performance data */
903 ARCSTAT_BUMP(arcstat_hash_collisions
);
905 ARCSTAT_BUMP(arcstat_hash_chains
);
907 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
910 ARCSTAT_BUMP(arcstat_hash_elements
);
911 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
917 buf_hash_remove(arc_buf_hdr_t
*hdr
)
919 arc_buf_hdr_t
*fhdr
, **hdrp
;
920 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
922 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
923 ASSERT(HDR_IN_HASH_TABLE(hdr
));
925 hdrp
= &buf_hash_table
.ht_table
[idx
];
926 while ((fhdr
= *hdrp
) != hdr
) {
927 ASSERT(fhdr
!= NULL
);
928 hdrp
= &fhdr
->b_hash_next
;
930 *hdrp
= hdr
->b_hash_next
;
931 hdr
->b_hash_next
= NULL
;
932 hdr
->b_flags
&= ~ARC_FLAG_IN_HASH_TABLE
;
934 /* collect some hash table performance data */
935 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
937 if (buf_hash_table
.ht_table
[idx
] &&
938 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
939 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
943 * Global data structures and functions for the buf kmem cache.
945 static kmem_cache_t
*hdr_full_cache
;
946 static kmem_cache_t
*hdr_l2only_cache
;
947 static kmem_cache_t
*buf_cache
;
954 #if defined(_KERNEL) && defined(HAVE_SPL)
956 * Large allocations which do not require contiguous pages
957 * should be using vmem_free() in the linux kernel\
959 vmem_free(buf_hash_table
.ht_table
,
960 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
962 kmem_free(buf_hash_table
.ht_table
,
963 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
965 for (i
= 0; i
< BUF_LOCKS
; i
++)
966 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
967 kmem_cache_destroy(hdr_full_cache
);
968 kmem_cache_destroy(hdr_l2only_cache
);
969 kmem_cache_destroy(buf_cache
);
973 * Constructor callback - called when the cache is empty
974 * and a new buf is requested.
978 hdr_full_cons(void *vbuf
, void *unused
, int kmflag
)
980 arc_buf_hdr_t
*hdr
= vbuf
;
982 bzero(hdr
, HDR_FULL_SIZE
);
983 cv_init(&hdr
->b_l1hdr
.b_cv
, NULL
, CV_DEFAULT
, NULL
);
984 refcount_create(&hdr
->b_l1hdr
.b_refcnt
);
985 mutex_init(&hdr
->b_l1hdr
.b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
986 list_link_init(&hdr
->b_l1hdr
.b_arc_node
);
987 list_link_init(&hdr
->b_l2hdr
.b_l2node
);
988 multilist_link_init(&hdr
->b_l1hdr
.b_arc_node
);
989 arc_space_consume(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
996 hdr_l2only_cons(void *vbuf
, void *unused
, int kmflag
)
998 arc_buf_hdr_t
*hdr
= vbuf
;
1000 bzero(hdr
, HDR_L2ONLY_SIZE
);
1001 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1008 buf_cons(void *vbuf
, void *unused
, int kmflag
)
1010 arc_buf_t
*buf
= vbuf
;
1012 bzero(buf
, sizeof (arc_buf_t
));
1013 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1014 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1020 * Destructor callback - called when a cached buf is
1021 * no longer required.
1025 hdr_full_dest(void *vbuf
, void *unused
)
1027 arc_buf_hdr_t
*hdr
= vbuf
;
1029 ASSERT(BUF_EMPTY(hdr
));
1030 cv_destroy(&hdr
->b_l1hdr
.b_cv
);
1031 refcount_destroy(&hdr
->b_l1hdr
.b_refcnt
);
1032 mutex_destroy(&hdr
->b_l1hdr
.b_freeze_lock
);
1033 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1034 arc_space_return(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1039 hdr_l2only_dest(void *vbuf
, void *unused
)
1041 ASSERTV(arc_buf_hdr_t
*hdr
= vbuf
);
1043 ASSERT(BUF_EMPTY(hdr
));
1044 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1049 buf_dest(void *vbuf
, void *unused
)
1051 arc_buf_t
*buf
= vbuf
;
1053 mutex_destroy(&buf
->b_evict_lock
);
1054 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1058 * Reclaim callback -- invoked when memory is low.
1062 hdr_recl(void *unused
)
1064 dprintf("hdr_recl called\n");
1066 * umem calls the reclaim func when we destroy the buf cache,
1067 * which is after we do arc_fini().
1070 cv_signal(&arc_reclaim_thread_cv
);
1077 uint64_t hsize
= 1ULL << 12;
1081 * The hash table is big enough to fill all of physical memory
1082 * with an average block size of zfs_arc_average_blocksize (default 8K).
1083 * By default, the table will take up
1084 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1086 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
1089 buf_hash_table
.ht_mask
= hsize
- 1;
1090 #if defined(_KERNEL) && defined(HAVE_SPL)
1092 * Large allocations which do not require contiguous pages
1093 * should be using vmem_alloc() in the linux kernel
1095 buf_hash_table
.ht_table
=
1096 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
1098 buf_hash_table
.ht_table
=
1099 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
1101 if (buf_hash_table
.ht_table
== NULL
) {
1102 ASSERT(hsize
> (1ULL << 8));
1107 hdr_full_cache
= kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE
,
1108 0, hdr_full_cons
, hdr_full_dest
, hdr_recl
, NULL
, NULL
, 0);
1109 hdr_l2only_cache
= kmem_cache_create("arc_buf_hdr_t_l2only",
1110 HDR_L2ONLY_SIZE
, 0, hdr_l2only_cons
, hdr_l2only_dest
, hdr_recl
,
1112 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
1113 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
1115 for (i
= 0; i
< 256; i
++)
1116 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
1117 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
1119 for (i
= 0; i
< BUF_LOCKS
; i
++) {
1120 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
1121 NULL
, MUTEX_DEFAULT
, NULL
);
1126 * Transition between the two allocation states for the arc_buf_hdr struct.
1127 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1128 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1129 * version is used when a cache buffer is only in the L2ARC in order to reduce
1132 static arc_buf_hdr_t
*
1133 arc_hdr_realloc(arc_buf_hdr_t
*hdr
, kmem_cache_t
*old
, kmem_cache_t
*new)
1135 arc_buf_hdr_t
*nhdr
;
1138 ASSERT(HDR_HAS_L2HDR(hdr
));
1139 ASSERT((old
== hdr_full_cache
&& new == hdr_l2only_cache
) ||
1140 (old
== hdr_l2only_cache
&& new == hdr_full_cache
));
1142 dev
= hdr
->b_l2hdr
.b_dev
;
1143 nhdr
= kmem_cache_alloc(new, KM_PUSHPAGE
);
1145 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
1146 buf_hash_remove(hdr
);
1148 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
1150 if (new == hdr_full_cache
) {
1151 nhdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
1153 * arc_access and arc_change_state need to be aware that a
1154 * header has just come out of L2ARC, so we set its state to
1155 * l2c_only even though it's about to change.
1157 nhdr
->b_l1hdr
.b_state
= arc_l2c_only
;
1159 /* Verify previous threads set to NULL before freeing */
1160 ASSERT3P(nhdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1162 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1163 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
1166 * If we've reached here, We must have been called from
1167 * arc_evict_hdr(), as such we should have already been
1168 * removed from any ghost list we were previously on
1169 * (which protects us from racing with arc_evict_state),
1170 * thus no locking is needed during this check.
1172 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1175 * A buffer must not be moved into the arc_l2c_only
1176 * state if it's not finished being written out to the
1177 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1178 * might try to be accessed, even though it was removed.
1180 VERIFY(!HDR_L2_WRITING(hdr
));
1181 VERIFY3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1183 nhdr
->b_flags
&= ~ARC_FLAG_HAS_L1HDR
;
1186 * The header has been reallocated so we need to re-insert it into any
1189 (void) buf_hash_insert(nhdr
, NULL
);
1191 ASSERT(list_link_active(&hdr
->b_l2hdr
.b_l2node
));
1193 mutex_enter(&dev
->l2ad_mtx
);
1196 * We must place the realloc'ed header back into the list at
1197 * the same spot. Otherwise, if it's placed earlier in the list,
1198 * l2arc_write_buffers() could find it during the function's
1199 * write phase, and try to write it out to the l2arc.
1201 list_insert_after(&dev
->l2ad_buflist
, hdr
, nhdr
);
1202 list_remove(&dev
->l2ad_buflist
, hdr
);
1204 mutex_exit(&dev
->l2ad_mtx
);
1207 * Since we're using the pointer address as the tag when
1208 * incrementing and decrementing the l2ad_alloc refcount, we
1209 * must remove the old pointer (that we're about to destroy) and
1210 * add the new pointer to the refcount. Otherwise we'd remove
1211 * the wrong pointer address when calling arc_hdr_destroy() later.
1214 (void) refcount_remove_many(&dev
->l2ad_alloc
,
1215 hdr
->b_l2hdr
.b_asize
, hdr
);
1217 (void) refcount_add_many(&dev
->l2ad_alloc
,
1218 nhdr
->b_l2hdr
.b_asize
, nhdr
);
1220 buf_discard_identity(hdr
);
1221 hdr
->b_freeze_cksum
= NULL
;
1222 kmem_cache_free(old
, hdr
);
1228 #define ARC_MINTIME (hz>>4) /* 62 ms */
1231 arc_cksum_verify(arc_buf_t
*buf
)
1235 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1238 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1239 if (buf
->b_hdr
->b_freeze_cksum
== NULL
|| HDR_IO_ERROR(buf
->b_hdr
)) {
1240 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1243 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
1244 if (!ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
))
1245 panic("buffer modified while frozen!");
1246 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1250 arc_cksum_equal(arc_buf_t
*buf
)
1255 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1256 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
1257 equal
= ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
);
1258 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1264 arc_cksum_compute(arc_buf_t
*buf
, boolean_t force
)
1266 if (!force
&& !(zfs_flags
& ZFS_DEBUG_MODIFY
))
1269 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1270 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1271 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1274 buf
->b_hdr
->b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
), KM_SLEEP
);
1275 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
,
1276 buf
->b_hdr
->b_freeze_cksum
);
1277 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1283 arc_buf_sigsegv(int sig
, siginfo_t
*si
, void *unused
)
1285 panic("Got SIGSEGV at address: 0x%lx\n", (long) si
->si_addr
);
1291 arc_buf_unwatch(arc_buf_t
*buf
)
1295 ASSERT0(mprotect(buf
->b_data
, buf
->b_hdr
->b_size
,
1296 PROT_READ
| PROT_WRITE
));
1303 arc_buf_watch(arc_buf_t
*buf
)
1307 ASSERT0(mprotect(buf
->b_data
, buf
->b_hdr
->b_size
, PROT_READ
));
1311 static arc_buf_contents_t
1312 arc_buf_type(arc_buf_hdr_t
*hdr
)
1314 if (HDR_ISTYPE_METADATA(hdr
)) {
1315 return (ARC_BUFC_METADATA
);
1317 return (ARC_BUFC_DATA
);
1322 arc_bufc_to_flags(arc_buf_contents_t type
)
1326 /* metadata field is 0 if buffer contains normal data */
1328 case ARC_BUFC_METADATA
:
1329 return (ARC_FLAG_BUFC_METADATA
);
1333 panic("undefined ARC buffer type!");
1334 return ((uint32_t)-1);
1338 arc_buf_thaw(arc_buf_t
*buf
)
1340 if (zfs_flags
& ZFS_DEBUG_MODIFY
) {
1341 if (buf
->b_hdr
->b_l1hdr
.b_state
!= arc_anon
)
1342 panic("modifying non-anon buffer!");
1343 if (HDR_IO_IN_PROGRESS(buf
->b_hdr
))
1344 panic("modifying buffer while i/o in progress!");
1345 arc_cksum_verify(buf
);
1348 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1349 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1350 kmem_free(buf
->b_hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
1351 buf
->b_hdr
->b_freeze_cksum
= NULL
;
1354 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1356 arc_buf_unwatch(buf
);
1360 arc_buf_freeze(arc_buf_t
*buf
)
1362 kmutex_t
*hash_lock
;
1364 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1367 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1368 mutex_enter(hash_lock
);
1370 ASSERT(buf
->b_hdr
->b_freeze_cksum
!= NULL
||
1371 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
1372 arc_cksum_compute(buf
, B_FALSE
);
1373 mutex_exit(hash_lock
);
1378 add_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1382 ASSERT(HDR_HAS_L1HDR(hdr
));
1383 ASSERT(MUTEX_HELD(hash_lock
));
1385 state
= hdr
->b_l1hdr
.b_state
;
1387 if ((refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
) == 1) &&
1388 (state
!= arc_anon
)) {
1389 /* We don't use the L2-only state list. */
1390 if (state
!= arc_l2c_only
) {
1391 arc_buf_contents_t type
= arc_buf_type(hdr
);
1392 uint64_t delta
= hdr
->b_size
* hdr
->b_l1hdr
.b_datacnt
;
1393 multilist_t
*list
= &state
->arcs_list
[type
];
1394 uint64_t *size
= &state
->arcs_lsize
[type
];
1396 multilist_remove(list
, hdr
);
1398 if (GHOST_STATE(state
)) {
1399 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
1400 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1401 delta
= hdr
->b_size
;
1404 ASSERT3U(*size
, >=, delta
);
1405 atomic_add_64(size
, -delta
);
1407 /* remove the prefetch flag if we get a reference */
1408 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
1413 remove_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1416 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
1418 ASSERT(HDR_HAS_L1HDR(hdr
));
1419 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
1420 ASSERT(!GHOST_STATE(state
));
1423 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1424 * check to prevent usage of the arc_l2c_only list.
1426 if (((cnt
= refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
)) == 0) &&
1427 (state
!= arc_anon
)) {
1428 arc_buf_contents_t type
= arc_buf_type(hdr
);
1429 multilist_t
*list
= &state
->arcs_list
[type
];
1430 uint64_t *size
= &state
->arcs_lsize
[type
];
1432 multilist_insert(list
, hdr
);
1434 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
1435 atomic_add_64(size
, hdr
->b_size
*
1436 hdr
->b_l1hdr
.b_datacnt
);
1442 * Returns detailed information about a specific arc buffer. When the
1443 * state_index argument is set the function will calculate the arc header
1444 * list position for its arc state. Since this requires a linear traversal
1445 * callers are strongly encourage not to do this. However, it can be helpful
1446 * for targeted analysis so the functionality is provided.
1449 arc_buf_info(arc_buf_t
*ab
, arc_buf_info_t
*abi
, int state_index
)
1451 arc_buf_hdr_t
*hdr
= ab
->b_hdr
;
1452 l1arc_buf_hdr_t
*l1hdr
= NULL
;
1453 l2arc_buf_hdr_t
*l2hdr
= NULL
;
1454 arc_state_t
*state
= NULL
;
1456 if (HDR_HAS_L1HDR(hdr
)) {
1457 l1hdr
= &hdr
->b_l1hdr
;
1458 state
= l1hdr
->b_state
;
1460 if (HDR_HAS_L2HDR(hdr
))
1461 l2hdr
= &hdr
->b_l2hdr
;
1463 memset(abi
, 0, sizeof (arc_buf_info_t
));
1464 abi
->abi_flags
= hdr
->b_flags
;
1467 abi
->abi_datacnt
= l1hdr
->b_datacnt
;
1468 abi
->abi_access
= l1hdr
->b_arc_access
;
1469 abi
->abi_mru_hits
= l1hdr
->b_mru_hits
;
1470 abi
->abi_mru_ghost_hits
= l1hdr
->b_mru_ghost_hits
;
1471 abi
->abi_mfu_hits
= l1hdr
->b_mfu_hits
;
1472 abi
->abi_mfu_ghost_hits
= l1hdr
->b_mfu_ghost_hits
;
1473 abi
->abi_holds
= refcount_count(&l1hdr
->b_refcnt
);
1477 abi
->abi_l2arc_dattr
= l2hdr
->b_daddr
;
1478 abi
->abi_l2arc_asize
= l2hdr
->b_asize
;
1479 abi
->abi_l2arc_compress
= HDR_GET_COMPRESS(hdr
);
1480 abi
->abi_l2arc_hits
= l2hdr
->b_hits
;
1483 abi
->abi_state_type
= state
? state
->arcs_state
: ARC_STATE_ANON
;
1484 abi
->abi_state_contents
= arc_buf_type(hdr
);
1485 abi
->abi_size
= hdr
->b_size
;
1489 * Move the supplied buffer to the indicated state. The hash lock
1490 * for the buffer must be held by the caller.
1493 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*hdr
,
1494 kmutex_t
*hash_lock
)
1496 arc_state_t
*old_state
;
1499 uint64_t from_delta
, to_delta
;
1500 arc_buf_contents_t buftype
= arc_buf_type(hdr
);
1503 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1504 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1505 * L1 hdr doesn't always exist when we change state to arc_anon before
1506 * destroying a header, in which case reallocating to add the L1 hdr is
1509 if (HDR_HAS_L1HDR(hdr
)) {
1510 old_state
= hdr
->b_l1hdr
.b_state
;
1511 refcnt
= refcount_count(&hdr
->b_l1hdr
.b_refcnt
);
1512 datacnt
= hdr
->b_l1hdr
.b_datacnt
;
1514 old_state
= arc_l2c_only
;
1519 ASSERT(MUTEX_HELD(hash_lock
));
1520 ASSERT3P(new_state
, !=, old_state
);
1521 ASSERT(refcnt
== 0 || datacnt
> 0);
1522 ASSERT(!GHOST_STATE(new_state
) || datacnt
== 0);
1523 ASSERT(old_state
!= arc_anon
|| datacnt
<= 1);
1525 from_delta
= to_delta
= datacnt
* hdr
->b_size
;
1528 * If this buffer is evictable, transfer it from the
1529 * old state list to the new state list.
1532 if (old_state
!= arc_anon
&& old_state
!= arc_l2c_only
) {
1533 uint64_t *size
= &old_state
->arcs_lsize
[buftype
];
1535 ASSERT(HDR_HAS_L1HDR(hdr
));
1536 multilist_remove(&old_state
->arcs_list
[buftype
], hdr
);
1539 * If prefetching out of the ghost cache,
1540 * we will have a non-zero datacnt.
1542 if (GHOST_STATE(old_state
) && datacnt
== 0) {
1543 /* ghost elements have a ghost size */
1544 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1545 from_delta
= hdr
->b_size
;
1547 ASSERT3U(*size
, >=, from_delta
);
1548 atomic_add_64(size
, -from_delta
);
1550 if (new_state
!= arc_anon
&& new_state
!= arc_l2c_only
) {
1551 uint64_t *size
= &new_state
->arcs_lsize
[buftype
];
1554 * An L1 header always exists here, since if we're
1555 * moving to some L1-cached state (i.e. not l2c_only or
1556 * anonymous), we realloc the header to add an L1hdr
1559 ASSERT(HDR_HAS_L1HDR(hdr
));
1560 multilist_insert(&new_state
->arcs_list
[buftype
], hdr
);
1562 /* ghost elements have a ghost size */
1563 if (GHOST_STATE(new_state
)) {
1565 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1566 to_delta
= hdr
->b_size
;
1568 atomic_add_64(size
, to_delta
);
1572 ASSERT(!BUF_EMPTY(hdr
));
1573 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(hdr
))
1574 buf_hash_remove(hdr
);
1576 /* adjust state sizes (ignore arc_l2c_only) */
1578 if (to_delta
&& new_state
!= arc_l2c_only
) {
1579 ASSERT(HDR_HAS_L1HDR(hdr
));
1580 if (GHOST_STATE(new_state
)) {
1584 * We moving a header to a ghost state, we first
1585 * remove all arc buffers. Thus, we'll have a
1586 * datacnt of zero, and no arc buffer to use for
1587 * the reference. As a result, we use the arc
1588 * header pointer for the reference.
1590 (void) refcount_add_many(&new_state
->arcs_size
,
1594 ASSERT3U(datacnt
, !=, 0);
1597 * Each individual buffer holds a unique reference,
1598 * thus we must remove each of these references one
1601 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
1602 buf
= buf
->b_next
) {
1603 (void) refcount_add_many(&new_state
->arcs_size
,
1609 if (from_delta
&& old_state
!= arc_l2c_only
) {
1610 ASSERT(HDR_HAS_L1HDR(hdr
));
1611 if (GHOST_STATE(old_state
)) {
1613 * When moving a header off of a ghost state,
1614 * there's the possibility for datacnt to be
1615 * non-zero. This is because we first add the
1616 * arc buffer to the header prior to changing
1617 * the header's state. Since we used the header
1618 * for the reference when putting the header on
1619 * the ghost state, we must balance that and use
1620 * the header when removing off the ghost state
1621 * (even though datacnt is non zero).
1624 IMPLY(datacnt
== 0, new_state
== arc_anon
||
1625 new_state
== arc_l2c_only
);
1627 (void) refcount_remove_many(&old_state
->arcs_size
,
1631 ASSERT3U(datacnt
, !=, 0);
1634 * Each individual buffer holds a unique reference,
1635 * thus we must remove each of these references one
1638 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
1639 buf
= buf
->b_next
) {
1640 (void) refcount_remove_many(
1641 &old_state
->arcs_size
, hdr
->b_size
, buf
);
1646 if (HDR_HAS_L1HDR(hdr
))
1647 hdr
->b_l1hdr
.b_state
= new_state
;
1650 * L2 headers should never be on the L2 state list since they don't
1651 * have L1 headers allocated.
1653 ASSERT(multilist_is_empty(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]) &&
1654 multilist_is_empty(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]));
1658 arc_space_consume(uint64_t space
, arc_space_type_t type
)
1660 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1665 case ARC_SPACE_DATA
:
1666 ARCSTAT_INCR(arcstat_data_size
, space
);
1668 case ARC_SPACE_META
:
1669 ARCSTAT_INCR(arcstat_metadata_size
, space
);
1671 case ARC_SPACE_OTHER
:
1672 ARCSTAT_INCR(arcstat_other_size
, space
);
1674 case ARC_SPACE_HDRS
:
1675 ARCSTAT_INCR(arcstat_hdr_size
, space
);
1677 case ARC_SPACE_L2HDRS
:
1678 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
1682 if (type
!= ARC_SPACE_DATA
)
1683 ARCSTAT_INCR(arcstat_meta_used
, space
);
1685 atomic_add_64(&arc_size
, space
);
1689 arc_space_return(uint64_t space
, arc_space_type_t type
)
1691 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1696 case ARC_SPACE_DATA
:
1697 ARCSTAT_INCR(arcstat_data_size
, -space
);
1699 case ARC_SPACE_META
:
1700 ARCSTAT_INCR(arcstat_metadata_size
, -space
);
1702 case ARC_SPACE_OTHER
:
1703 ARCSTAT_INCR(arcstat_other_size
, -space
);
1705 case ARC_SPACE_HDRS
:
1706 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
1708 case ARC_SPACE_L2HDRS
:
1709 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
1713 if (type
!= ARC_SPACE_DATA
) {
1714 ASSERT(arc_meta_used
>= space
);
1715 if (arc_meta_max
< arc_meta_used
)
1716 arc_meta_max
= arc_meta_used
;
1717 ARCSTAT_INCR(arcstat_meta_used
, -space
);
1720 ASSERT(arc_size
>= space
);
1721 atomic_add_64(&arc_size
, -space
);
1725 arc_buf_alloc(spa_t
*spa
, uint64_t size
, void *tag
, arc_buf_contents_t type
)
1730 VERIFY3U(size
, <=, spa_maxblocksize(spa
));
1731 hdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
1732 ASSERT(BUF_EMPTY(hdr
));
1733 ASSERT3P(hdr
->b_freeze_cksum
, ==, NULL
);
1735 hdr
->b_spa
= spa_load_guid(spa
);
1736 hdr
->b_l1hdr
.b_mru_hits
= 0;
1737 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
1738 hdr
->b_l1hdr
.b_mfu_hits
= 0;
1739 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
1740 hdr
->b_l1hdr
.b_l2_hits
= 0;
1742 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1745 buf
->b_efunc
= NULL
;
1746 buf
->b_private
= NULL
;
1749 hdr
->b_flags
= arc_bufc_to_flags(type
);
1750 hdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
1752 hdr
->b_l1hdr
.b_buf
= buf
;
1753 hdr
->b_l1hdr
.b_state
= arc_anon
;
1754 hdr
->b_l1hdr
.b_arc_access
= 0;
1755 hdr
->b_l1hdr
.b_datacnt
= 1;
1756 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1758 arc_get_data_buf(buf
);
1759 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
1760 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1765 static char *arc_onloan_tag
= "onloan";
1768 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1769 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1770 * buffers must be returned to the arc before they can be used by the DMU or
1774 arc_loan_buf(spa_t
*spa
, uint64_t size
)
1778 buf
= arc_buf_alloc(spa
, size
, arc_onloan_tag
, ARC_BUFC_DATA
);
1780 atomic_add_64(&arc_loaned_bytes
, size
);
1785 * Return a loaned arc buffer to the arc.
1788 arc_return_buf(arc_buf_t
*buf
, void *tag
)
1790 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1792 ASSERT(buf
->b_data
!= NULL
);
1793 ASSERT(HDR_HAS_L1HDR(hdr
));
1794 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1795 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
1797 atomic_add_64(&arc_loaned_bytes
, -hdr
->b_size
);
1800 /* Detach an arc_buf from a dbuf (tag) */
1802 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
1804 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1806 ASSERT(buf
->b_data
!= NULL
);
1807 ASSERT(HDR_HAS_L1HDR(hdr
));
1808 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
1809 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1810 buf
->b_efunc
= NULL
;
1811 buf
->b_private
= NULL
;
1813 atomic_add_64(&arc_loaned_bytes
, hdr
->b_size
);
1817 arc_buf_clone(arc_buf_t
*from
)
1820 arc_buf_hdr_t
*hdr
= from
->b_hdr
;
1821 uint64_t size
= hdr
->b_size
;
1823 ASSERT(HDR_HAS_L1HDR(hdr
));
1824 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_anon
);
1826 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1829 buf
->b_efunc
= NULL
;
1830 buf
->b_private
= NULL
;
1831 buf
->b_next
= hdr
->b_l1hdr
.b_buf
;
1832 hdr
->b_l1hdr
.b_buf
= buf
;
1833 arc_get_data_buf(buf
);
1834 bcopy(from
->b_data
, buf
->b_data
, size
);
1837 * This buffer already exists in the arc so create a duplicate
1838 * copy for the caller. If the buffer is associated with user data
1839 * then track the size and number of duplicates. These stats will be
1840 * updated as duplicate buffers are created and destroyed.
1842 if (HDR_ISTYPE_DATA(hdr
)) {
1843 ARCSTAT_BUMP(arcstat_duplicate_buffers
);
1844 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, size
);
1846 hdr
->b_l1hdr
.b_datacnt
+= 1;
1851 arc_buf_add_ref(arc_buf_t
*buf
, void* tag
)
1854 kmutex_t
*hash_lock
;
1857 * Check to see if this buffer is evicted. Callers
1858 * must verify b_data != NULL to know if the add_ref
1861 mutex_enter(&buf
->b_evict_lock
);
1862 if (buf
->b_data
== NULL
) {
1863 mutex_exit(&buf
->b_evict_lock
);
1866 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1867 mutex_enter(hash_lock
);
1869 ASSERT(HDR_HAS_L1HDR(hdr
));
1870 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1871 mutex_exit(&buf
->b_evict_lock
);
1873 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
1874 hdr
->b_l1hdr
.b_state
== arc_mfu
);
1876 add_reference(hdr
, hash_lock
, tag
);
1877 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
1878 arc_access(hdr
, hash_lock
);
1879 mutex_exit(hash_lock
);
1880 ARCSTAT_BUMP(arcstat_hits
);
1881 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
1882 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
1883 data
, metadata
, hits
);
1887 arc_buf_free_on_write(void *data
, size_t size
,
1888 void (*free_func
)(void *, size_t))
1890 l2arc_data_free_t
*df
;
1892 df
= kmem_alloc(sizeof (*df
), KM_SLEEP
);
1893 df
->l2df_data
= data
;
1894 df
->l2df_size
= size
;
1895 df
->l2df_func
= free_func
;
1896 mutex_enter(&l2arc_free_on_write_mtx
);
1897 list_insert_head(l2arc_free_on_write
, df
);
1898 mutex_exit(&l2arc_free_on_write_mtx
);
1902 * Free the arc data buffer. If it is an l2arc write in progress,
1903 * the buffer is placed on l2arc_free_on_write to be freed later.
1906 arc_buf_data_free(arc_buf_t
*buf
, void (*free_func
)(void *, size_t))
1908 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1910 if (HDR_L2_WRITING(hdr
)) {
1911 arc_buf_free_on_write(buf
->b_data
, hdr
->b_size
, free_func
);
1912 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
1914 free_func(buf
->b_data
, hdr
->b_size
);
1919 arc_buf_l2_cdata_free(arc_buf_hdr_t
*hdr
)
1921 ASSERT(HDR_HAS_L2HDR(hdr
));
1922 ASSERT(MUTEX_HELD(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
));
1925 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
1926 * that doesn't exist, the header is in the arc_l2c_only state,
1927 * and there isn't anything to free (it's already been freed).
1929 if (!HDR_HAS_L1HDR(hdr
))
1933 * The header isn't being written to the l2arc device, thus it
1934 * shouldn't have a b_tmp_cdata to free.
1936 if (!HDR_L2_WRITING(hdr
)) {
1937 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1942 * The header does not have compression enabled. This can be due
1943 * to the buffer not being compressible, or because we're
1944 * freeing the buffer before the second phase of
1945 * l2arc_write_buffer() has started (which does the compression
1946 * step). In either case, b_tmp_cdata does not point to a
1947 * separately compressed buffer, so there's nothing to free (it
1948 * points to the same buffer as the arc_buf_t's b_data field).
1950 if (HDR_GET_COMPRESS(hdr
) == ZIO_COMPRESS_OFF
) {
1951 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1956 * There's nothing to free since the buffer was all zero's and
1957 * compressed to a zero length buffer.
1959 if (HDR_GET_COMPRESS(hdr
) == ZIO_COMPRESS_EMPTY
) {
1960 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1964 ASSERT(L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr
)));
1966 arc_buf_free_on_write(hdr
->b_l1hdr
.b_tmp_cdata
,
1967 hdr
->b_size
, zio_data_buf_free
);
1969 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write
);
1970 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1974 * Free up buf->b_data and if 'remove' is set, then pull the
1975 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
1978 arc_buf_destroy(arc_buf_t
*buf
, boolean_t remove
)
1982 /* free up data associated with the buf */
1983 if (buf
->b_data
!= NULL
) {
1984 arc_state_t
*state
= buf
->b_hdr
->b_l1hdr
.b_state
;
1985 uint64_t size
= buf
->b_hdr
->b_size
;
1986 arc_buf_contents_t type
= arc_buf_type(buf
->b_hdr
);
1988 arc_cksum_verify(buf
);
1989 arc_buf_unwatch(buf
);
1991 if (type
== ARC_BUFC_METADATA
) {
1992 arc_buf_data_free(buf
, zio_buf_free
);
1993 arc_space_return(size
, ARC_SPACE_META
);
1995 ASSERT(type
== ARC_BUFC_DATA
);
1996 arc_buf_data_free(buf
, zio_data_buf_free
);
1997 arc_space_return(size
, ARC_SPACE_DATA
);
2000 /* protected by hash lock, if in the hash table */
2001 if (multilist_link_active(&buf
->b_hdr
->b_l1hdr
.b_arc_node
)) {
2002 uint64_t *cnt
= &state
->arcs_lsize
[type
];
2004 ASSERT(refcount_is_zero(
2005 &buf
->b_hdr
->b_l1hdr
.b_refcnt
));
2006 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
2008 ASSERT3U(*cnt
, >=, size
);
2009 atomic_add_64(cnt
, -size
);
2012 (void) refcount_remove_many(&state
->arcs_size
, size
, buf
);
2016 * If we're destroying a duplicate buffer make sure
2017 * that the appropriate statistics are updated.
2019 if (buf
->b_hdr
->b_l1hdr
.b_datacnt
> 1 &&
2020 HDR_ISTYPE_DATA(buf
->b_hdr
)) {
2021 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
2022 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, -size
);
2024 ASSERT(buf
->b_hdr
->b_l1hdr
.b_datacnt
> 0);
2025 buf
->b_hdr
->b_l1hdr
.b_datacnt
-= 1;
2028 /* only remove the buf if requested */
2032 /* remove the buf from the hdr list */
2033 for (bufp
= &buf
->b_hdr
->b_l1hdr
.b_buf
; *bufp
!= buf
;
2034 bufp
= &(*bufp
)->b_next
)
2036 *bufp
= buf
->b_next
;
2039 ASSERT(buf
->b_efunc
== NULL
);
2041 /* clean up the buf */
2043 kmem_cache_free(buf_cache
, buf
);
2047 arc_hdr_l2hdr_destroy(arc_buf_hdr_t
*hdr
)
2049 l2arc_buf_hdr_t
*l2hdr
= &hdr
->b_l2hdr
;
2050 l2arc_dev_t
*dev
= l2hdr
->b_dev
;
2052 ASSERT(MUTEX_HELD(&dev
->l2ad_mtx
));
2053 ASSERT(HDR_HAS_L2HDR(hdr
));
2055 list_remove(&dev
->l2ad_buflist
, hdr
);
2058 * We don't want to leak the b_tmp_cdata buffer that was
2059 * allocated in l2arc_write_buffers()
2061 arc_buf_l2_cdata_free(hdr
);
2064 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2065 * this header is being processed by l2arc_write_buffers() (i.e.
2066 * it's in the first stage of l2arc_write_buffers()).
2067 * Re-affirming that truth here, just to serve as a reminder. If
2068 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2069 * may not have its HDR_L2_WRITING flag set. (the write may have
2070 * completed, in which case HDR_L2_WRITING will be false and the
2071 * b_daddr field will point to the address of the buffer on disk).
2073 IMPLY(l2hdr
->b_daddr
== L2ARC_ADDR_UNSET
, HDR_L2_WRITING(hdr
));
2076 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2077 * l2arc_write_buffers(). Since we've just removed this header
2078 * from the l2arc buffer list, this header will never reach the
2079 * second stage of l2arc_write_buffers(), which increments the
2080 * accounting stats for this header. Thus, we must be careful
2081 * not to decrement them for this header either.
2083 if (l2hdr
->b_daddr
!= L2ARC_ADDR_UNSET
) {
2084 ARCSTAT_INCR(arcstat_l2_asize
, -l2hdr
->b_asize
);
2085 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
2087 vdev_space_update(dev
->l2ad_vdev
,
2088 -l2hdr
->b_asize
, 0, 0);
2090 (void) refcount_remove_many(&dev
->l2ad_alloc
,
2091 l2hdr
->b_asize
, hdr
);
2094 hdr
->b_flags
&= ~ARC_FLAG_HAS_L2HDR
;
2098 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
2100 if (HDR_HAS_L1HDR(hdr
)) {
2101 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
||
2102 hdr
->b_l1hdr
.b_datacnt
> 0);
2103 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2104 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
2106 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
2107 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
2109 if (HDR_HAS_L2HDR(hdr
)) {
2110 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
2111 boolean_t buflist_held
= MUTEX_HELD(&dev
->l2ad_mtx
);
2114 mutex_enter(&dev
->l2ad_mtx
);
2117 * Even though we checked this conditional above, we
2118 * need to check this again now that we have the
2119 * l2ad_mtx. This is because we could be racing with
2120 * another thread calling l2arc_evict() which might have
2121 * destroyed this header's L2 portion as we were waiting
2122 * to acquire the l2ad_mtx. If that happens, we don't
2123 * want to re-destroy the header's L2 portion.
2125 if (HDR_HAS_L2HDR(hdr
))
2126 arc_hdr_l2hdr_destroy(hdr
);
2129 mutex_exit(&dev
->l2ad_mtx
);
2132 if (!BUF_EMPTY(hdr
))
2133 buf_discard_identity(hdr
);
2135 if (hdr
->b_freeze_cksum
!= NULL
) {
2136 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
2137 hdr
->b_freeze_cksum
= NULL
;
2140 if (HDR_HAS_L1HDR(hdr
)) {
2141 while (hdr
->b_l1hdr
.b_buf
) {
2142 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
2144 if (buf
->b_efunc
!= NULL
) {
2145 mutex_enter(&arc_user_evicts_lock
);
2146 mutex_enter(&buf
->b_evict_lock
);
2147 ASSERT(buf
->b_hdr
!= NULL
);
2148 arc_buf_destroy(hdr
->b_l1hdr
.b_buf
, FALSE
);
2149 hdr
->b_l1hdr
.b_buf
= buf
->b_next
;
2150 buf
->b_hdr
= &arc_eviction_hdr
;
2151 buf
->b_next
= arc_eviction_list
;
2152 arc_eviction_list
= buf
;
2153 mutex_exit(&buf
->b_evict_lock
);
2154 cv_signal(&arc_user_evicts_cv
);
2155 mutex_exit(&arc_user_evicts_lock
);
2157 arc_buf_destroy(hdr
->b_l1hdr
.b_buf
, TRUE
);
2162 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
2163 if (HDR_HAS_L1HDR(hdr
)) {
2164 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
2165 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
2166 kmem_cache_free(hdr_full_cache
, hdr
);
2168 kmem_cache_free(hdr_l2only_cache
, hdr
);
2173 arc_buf_free(arc_buf_t
*buf
, void *tag
)
2175 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2176 int hashed
= hdr
->b_l1hdr
.b_state
!= arc_anon
;
2178 ASSERT(buf
->b_efunc
== NULL
);
2179 ASSERT(buf
->b_data
!= NULL
);
2182 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
2184 mutex_enter(hash_lock
);
2186 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
2188 (void) remove_reference(hdr
, hash_lock
, tag
);
2189 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
2190 arc_buf_destroy(buf
, TRUE
);
2192 ASSERT(buf
== hdr
->b_l1hdr
.b_buf
);
2193 ASSERT(buf
->b_efunc
== NULL
);
2194 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
2196 mutex_exit(hash_lock
);
2197 } else if (HDR_IO_IN_PROGRESS(hdr
)) {
2200 * We are in the middle of an async write. Don't destroy
2201 * this buffer unless the write completes before we finish
2202 * decrementing the reference count.
2204 mutex_enter(&arc_user_evicts_lock
);
2205 (void) remove_reference(hdr
, NULL
, tag
);
2206 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2207 destroy_hdr
= !HDR_IO_IN_PROGRESS(hdr
);
2208 mutex_exit(&arc_user_evicts_lock
);
2210 arc_hdr_destroy(hdr
);
2212 if (remove_reference(hdr
, NULL
, tag
) > 0)
2213 arc_buf_destroy(buf
, TRUE
);
2215 arc_hdr_destroy(hdr
);
2220 arc_buf_remove_ref(arc_buf_t
*buf
, void* tag
)
2222 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2223 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
2224 boolean_t no_callback
= (buf
->b_efunc
== NULL
);
2226 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
2227 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
2228 arc_buf_free(buf
, tag
);
2229 return (no_callback
);
2232 mutex_enter(hash_lock
);
2234 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
2235 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
2236 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_anon
);
2237 ASSERT(buf
->b_data
!= NULL
);
2239 (void) remove_reference(hdr
, hash_lock
, tag
);
2240 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
2242 arc_buf_destroy(buf
, TRUE
);
2243 } else if (no_callback
) {
2244 ASSERT(hdr
->b_l1hdr
.b_buf
== buf
&& buf
->b_next
== NULL
);
2245 ASSERT(buf
->b_efunc
== NULL
);
2246 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
2248 ASSERT(no_callback
|| hdr
->b_l1hdr
.b_datacnt
> 1 ||
2249 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2250 mutex_exit(hash_lock
);
2251 return (no_callback
);
2255 arc_buf_size(arc_buf_t
*buf
)
2257 return (buf
->b_hdr
->b_size
);
2261 * Called from the DMU to determine if the current buffer should be
2262 * evicted. In order to ensure proper locking, the eviction must be initiated
2263 * from the DMU. Return true if the buffer is associated with user data and
2264 * duplicate buffers still exist.
2267 arc_buf_eviction_needed(arc_buf_t
*buf
)
2270 boolean_t evict_needed
= B_FALSE
;
2272 if (zfs_disable_dup_eviction
)
2275 mutex_enter(&buf
->b_evict_lock
);
2279 * We are in arc_do_user_evicts(); let that function
2280 * perform the eviction.
2282 ASSERT(buf
->b_data
== NULL
);
2283 mutex_exit(&buf
->b_evict_lock
);
2285 } else if (buf
->b_data
== NULL
) {
2287 * We have already been added to the arc eviction list;
2288 * recommend eviction.
2290 ASSERT3P(hdr
, ==, &arc_eviction_hdr
);
2291 mutex_exit(&buf
->b_evict_lock
);
2295 if (hdr
->b_l1hdr
.b_datacnt
> 1 && HDR_ISTYPE_DATA(hdr
))
2296 evict_needed
= B_TRUE
;
2298 mutex_exit(&buf
->b_evict_lock
);
2299 return (evict_needed
);
2303 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2304 * state of the header is dependent on its state prior to entering this
2305 * function. The following transitions are possible:
2307 * - arc_mru -> arc_mru_ghost
2308 * - arc_mfu -> arc_mfu_ghost
2309 * - arc_mru_ghost -> arc_l2c_only
2310 * - arc_mru_ghost -> deleted
2311 * - arc_mfu_ghost -> arc_l2c_only
2312 * - arc_mfu_ghost -> deleted
2315 arc_evict_hdr(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
2317 arc_state_t
*evicted_state
, *state
;
2318 int64_t bytes_evicted
= 0;
2320 ASSERT(MUTEX_HELD(hash_lock
));
2321 ASSERT(HDR_HAS_L1HDR(hdr
));
2323 state
= hdr
->b_l1hdr
.b_state
;
2324 if (GHOST_STATE(state
)) {
2325 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
2326 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
2329 * l2arc_write_buffers() relies on a header's L1 portion
2330 * (i.e. its b_tmp_cdata field) during its write phase.
2331 * Thus, we cannot push a header onto the arc_l2c_only
2332 * state (removing its L1 piece) until the header is
2333 * done being written to the l2arc.
2335 if (HDR_HAS_L2HDR(hdr
) && HDR_L2_WRITING(hdr
)) {
2336 ARCSTAT_BUMP(arcstat_evict_l2_skip
);
2337 return (bytes_evicted
);
2340 ARCSTAT_BUMP(arcstat_deleted
);
2341 bytes_evicted
+= hdr
->b_size
;
2343 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, hdr
);
2345 if (HDR_HAS_L2HDR(hdr
)) {
2347 * This buffer is cached on the 2nd Level ARC;
2348 * don't destroy the header.
2350 arc_change_state(arc_l2c_only
, hdr
, hash_lock
);
2352 * dropping from L1+L2 cached to L2-only,
2353 * realloc to remove the L1 header.
2355 hdr
= arc_hdr_realloc(hdr
, hdr_full_cache
,
2358 arc_change_state(arc_anon
, hdr
, hash_lock
);
2359 arc_hdr_destroy(hdr
);
2361 return (bytes_evicted
);
2364 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
2365 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
2367 /* prefetch buffers have a minimum lifespan */
2368 if (HDR_IO_IN_PROGRESS(hdr
) ||
2369 ((hdr
->b_flags
& (ARC_FLAG_PREFETCH
| ARC_FLAG_INDIRECT
)) &&
2370 ddi_get_lbolt() - hdr
->b_l1hdr
.b_arc_access
<
2371 arc_min_prefetch_lifespan
)) {
2372 ARCSTAT_BUMP(arcstat_evict_skip
);
2373 return (bytes_evicted
);
2376 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
2377 ASSERT3U(hdr
->b_l1hdr
.b_datacnt
, >, 0);
2378 while (hdr
->b_l1hdr
.b_buf
) {
2379 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
2380 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
2381 ARCSTAT_BUMP(arcstat_mutex_miss
);
2384 if (buf
->b_data
!= NULL
)
2385 bytes_evicted
+= hdr
->b_size
;
2386 if (buf
->b_efunc
!= NULL
) {
2387 mutex_enter(&arc_user_evicts_lock
);
2388 arc_buf_destroy(buf
, FALSE
);
2389 hdr
->b_l1hdr
.b_buf
= buf
->b_next
;
2390 buf
->b_hdr
= &arc_eviction_hdr
;
2391 buf
->b_next
= arc_eviction_list
;
2392 arc_eviction_list
= buf
;
2393 cv_signal(&arc_user_evicts_cv
);
2394 mutex_exit(&arc_user_evicts_lock
);
2395 mutex_exit(&buf
->b_evict_lock
);
2397 mutex_exit(&buf
->b_evict_lock
);
2398 arc_buf_destroy(buf
, TRUE
);
2402 if (HDR_HAS_L2HDR(hdr
)) {
2403 ARCSTAT_INCR(arcstat_evict_l2_cached
, hdr
->b_size
);
2405 if (l2arc_write_eligible(hdr
->b_spa
, hdr
))
2406 ARCSTAT_INCR(arcstat_evict_l2_eligible
, hdr
->b_size
);
2408 ARCSTAT_INCR(arcstat_evict_l2_ineligible
, hdr
->b_size
);
2411 if (hdr
->b_l1hdr
.b_datacnt
== 0) {
2412 arc_change_state(evicted_state
, hdr
, hash_lock
);
2413 ASSERT(HDR_IN_HASH_TABLE(hdr
));
2414 hdr
->b_flags
|= ARC_FLAG_IN_HASH_TABLE
;
2415 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
2416 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, hdr
);
2419 return (bytes_evicted
);
2423 arc_evict_state_impl(multilist_t
*ml
, int idx
, arc_buf_hdr_t
*marker
,
2424 uint64_t spa
, int64_t bytes
)
2426 multilist_sublist_t
*mls
;
2427 uint64_t bytes_evicted
= 0;
2429 kmutex_t
*hash_lock
;
2430 int evict_count
= 0;
2432 ASSERT3P(marker
, !=, NULL
);
2433 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
2435 mls
= multilist_sublist_lock(ml
, idx
);
2437 for (hdr
= multilist_sublist_prev(mls
, marker
); hdr
!= NULL
;
2438 hdr
= multilist_sublist_prev(mls
, marker
)) {
2439 if ((bytes
!= ARC_EVICT_ALL
&& bytes_evicted
>= bytes
) ||
2440 (evict_count
>= zfs_arc_evict_batch_limit
))
2444 * To keep our iteration location, move the marker
2445 * forward. Since we're not holding hdr's hash lock, we
2446 * must be very careful and not remove 'hdr' from the
2447 * sublist. Otherwise, other consumers might mistake the
2448 * 'hdr' as not being on a sublist when they call the
2449 * multilist_link_active() function (they all rely on
2450 * the hash lock protecting concurrent insertions and
2451 * removals). multilist_sublist_move_forward() was
2452 * specifically implemented to ensure this is the case
2453 * (only 'marker' will be removed and re-inserted).
2455 multilist_sublist_move_forward(mls
, marker
);
2458 * The only case where the b_spa field should ever be
2459 * zero, is the marker headers inserted by
2460 * arc_evict_state(). It's possible for multiple threads
2461 * to be calling arc_evict_state() concurrently (e.g.
2462 * dsl_pool_close() and zio_inject_fault()), so we must
2463 * skip any markers we see from these other threads.
2465 if (hdr
->b_spa
== 0)
2468 /* we're only interested in evicting buffers of a certain spa */
2469 if (spa
!= 0 && hdr
->b_spa
!= spa
) {
2470 ARCSTAT_BUMP(arcstat_evict_skip
);
2474 hash_lock
= HDR_LOCK(hdr
);
2477 * We aren't calling this function from any code path
2478 * that would already be holding a hash lock, so we're
2479 * asserting on this assumption to be defensive in case
2480 * this ever changes. Without this check, it would be
2481 * possible to incorrectly increment arcstat_mutex_miss
2482 * below (e.g. if the code changed such that we called
2483 * this function with a hash lock held).
2485 ASSERT(!MUTEX_HELD(hash_lock
));
2487 if (mutex_tryenter(hash_lock
)) {
2488 uint64_t evicted
= arc_evict_hdr(hdr
, hash_lock
);
2489 mutex_exit(hash_lock
);
2491 bytes_evicted
+= evicted
;
2494 * If evicted is zero, arc_evict_hdr() must have
2495 * decided to skip this header, don't increment
2496 * evict_count in this case.
2502 * If arc_size isn't overflowing, signal any
2503 * threads that might happen to be waiting.
2505 * For each header evicted, we wake up a single
2506 * thread. If we used cv_broadcast, we could
2507 * wake up "too many" threads causing arc_size
2508 * to significantly overflow arc_c; since
2509 * arc_get_data_buf() doesn't check for overflow
2510 * when it's woken up (it doesn't because it's
2511 * possible for the ARC to be overflowing while
2512 * full of un-evictable buffers, and the
2513 * function should proceed in this case).
2515 * If threads are left sleeping, due to not
2516 * using cv_broadcast, they will be woken up
2517 * just before arc_reclaim_thread() sleeps.
2519 mutex_enter(&arc_reclaim_lock
);
2520 if (!arc_is_overflowing())
2521 cv_signal(&arc_reclaim_waiters_cv
);
2522 mutex_exit(&arc_reclaim_lock
);
2524 ARCSTAT_BUMP(arcstat_mutex_miss
);
2528 multilist_sublist_unlock(mls
);
2530 return (bytes_evicted
);
2534 * Evict buffers from the given arc state, until we've removed the
2535 * specified number of bytes. Move the removed buffers to the
2536 * appropriate evict state.
2538 * This function makes a "best effort". It skips over any buffers
2539 * it can't get a hash_lock on, and so, may not catch all candidates.
2540 * It may also return without evicting as much space as requested.
2542 * If bytes is specified using the special value ARC_EVICT_ALL, this
2543 * will evict all available (i.e. unlocked and evictable) buffers from
2544 * the given arc state; which is used by arc_flush().
2547 arc_evict_state(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
2548 arc_buf_contents_t type
)
2550 uint64_t total_evicted
= 0;
2551 multilist_t
*ml
= &state
->arcs_list
[type
];
2553 arc_buf_hdr_t
**markers
;
2556 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
2558 num_sublists
= multilist_get_num_sublists(ml
);
2561 * If we've tried to evict from each sublist, made some
2562 * progress, but still have not hit the target number of bytes
2563 * to evict, we want to keep trying. The markers allow us to
2564 * pick up where we left off for each individual sublist, rather
2565 * than starting from the tail each time.
2567 markers
= kmem_zalloc(sizeof (*markers
) * num_sublists
, KM_SLEEP
);
2568 for (i
= 0; i
< num_sublists
; i
++) {
2569 multilist_sublist_t
*mls
;
2571 markers
[i
] = kmem_cache_alloc(hdr_full_cache
, KM_SLEEP
);
2574 * A b_spa of 0 is used to indicate that this header is
2575 * a marker. This fact is used in arc_adjust_type() and
2576 * arc_evict_state_impl().
2578 markers
[i
]->b_spa
= 0;
2580 mls
= multilist_sublist_lock(ml
, i
);
2581 multilist_sublist_insert_tail(mls
, markers
[i
]);
2582 multilist_sublist_unlock(mls
);
2586 * While we haven't hit our target number of bytes to evict, or
2587 * we're evicting all available buffers.
2589 while (total_evicted
< bytes
|| bytes
== ARC_EVICT_ALL
) {
2591 * Start eviction using a randomly selected sublist,
2592 * this is to try and evenly balance eviction across all
2593 * sublists. Always starting at the same sublist
2594 * (e.g. index 0) would cause evictions to favor certain
2595 * sublists over others.
2597 int sublist_idx
= multilist_get_random_index(ml
);
2598 uint64_t scan_evicted
= 0;
2600 for (i
= 0; i
< num_sublists
; i
++) {
2601 uint64_t bytes_remaining
;
2602 uint64_t bytes_evicted
;
2604 if (bytes
== ARC_EVICT_ALL
)
2605 bytes_remaining
= ARC_EVICT_ALL
;
2606 else if (total_evicted
< bytes
)
2607 bytes_remaining
= bytes
- total_evicted
;
2611 bytes_evicted
= arc_evict_state_impl(ml
, sublist_idx
,
2612 markers
[sublist_idx
], spa
, bytes_remaining
);
2614 scan_evicted
+= bytes_evicted
;
2615 total_evicted
+= bytes_evicted
;
2617 /* we've reached the end, wrap to the beginning */
2618 if (++sublist_idx
>= num_sublists
)
2623 * If we didn't evict anything during this scan, we have
2624 * no reason to believe we'll evict more during another
2625 * scan, so break the loop.
2627 if (scan_evicted
== 0) {
2628 /* This isn't possible, let's make that obvious */
2629 ASSERT3S(bytes
, !=, 0);
2632 * When bytes is ARC_EVICT_ALL, the only way to
2633 * break the loop is when scan_evicted is zero.
2634 * In that case, we actually have evicted enough,
2635 * so we don't want to increment the kstat.
2637 if (bytes
!= ARC_EVICT_ALL
) {
2638 ASSERT3S(total_evicted
, <, bytes
);
2639 ARCSTAT_BUMP(arcstat_evict_not_enough
);
2646 for (i
= 0; i
< num_sublists
; i
++) {
2647 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
2648 multilist_sublist_remove(mls
, markers
[i
]);
2649 multilist_sublist_unlock(mls
);
2651 kmem_cache_free(hdr_full_cache
, markers
[i
]);
2653 kmem_free(markers
, sizeof (*markers
) * num_sublists
);
2655 return (total_evicted
);
2659 * Flush all "evictable" data of the given type from the arc state
2660 * specified. This will not evict any "active" buffers (i.e. referenced).
2662 * When 'retry' is set to FALSE, the function will make a single pass
2663 * over the state and evict any buffers that it can. Since it doesn't
2664 * continually retry the eviction, it might end up leaving some buffers
2665 * in the ARC due to lock misses.
2667 * When 'retry' is set to TRUE, the function will continually retry the
2668 * eviction until *all* evictable buffers have been removed from the
2669 * state. As a result, if concurrent insertions into the state are
2670 * allowed (e.g. if the ARC isn't shutting down), this function might
2671 * wind up in an infinite loop, continually trying to evict buffers.
2674 arc_flush_state(arc_state_t
*state
, uint64_t spa
, arc_buf_contents_t type
,
2677 uint64_t evicted
= 0;
2679 while (state
->arcs_lsize
[type
] != 0) {
2680 evicted
+= arc_evict_state(state
, spa
, ARC_EVICT_ALL
, type
);
2690 * Helper function for arc_prune() it is responsible for safely handling
2691 * the execution of a registered arc_prune_func_t.
2694 arc_prune_task(void *ptr
)
2696 arc_prune_t
*ap
= (arc_prune_t
*)ptr
;
2697 arc_prune_func_t
*func
= ap
->p_pfunc
;
2700 func(ap
->p_adjust
, ap
->p_private
);
2702 /* Callback unregistered concurrently with execution */
2703 if (refcount_remove(&ap
->p_refcnt
, func
) == 0) {
2704 ASSERT(!list_link_active(&ap
->p_node
));
2705 refcount_destroy(&ap
->p_refcnt
);
2706 kmem_free(ap
, sizeof (*ap
));
2711 * Notify registered consumers they must drop holds on a portion of the ARC
2712 * buffered they reference. This provides a mechanism to ensure the ARC can
2713 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
2714 * is analogous to dnlc_reduce_cache() but more generic.
2716 * This operation is performed asyncronously so it may be safely called
2717 * in the context of the arc_reclaim_thread(). A reference is taken here
2718 * for each registered arc_prune_t and the arc_prune_task() is responsible
2719 * for releasing it once the registered arc_prune_func_t has completed.
2722 arc_prune_async(int64_t adjust
)
2726 mutex_enter(&arc_prune_mtx
);
2727 for (ap
= list_head(&arc_prune_list
); ap
!= NULL
;
2728 ap
= list_next(&arc_prune_list
, ap
)) {
2730 if (refcount_count(&ap
->p_refcnt
) >= 2)
2733 refcount_add(&ap
->p_refcnt
, ap
->p_pfunc
);
2734 ap
->p_adjust
= adjust
;
2735 taskq_dispatch(arc_prune_taskq
, arc_prune_task
, ap
, TQ_SLEEP
);
2736 ARCSTAT_BUMP(arcstat_prune
);
2738 mutex_exit(&arc_prune_mtx
);
2742 arc_prune(int64_t adjust
)
2744 arc_prune_async(adjust
);
2745 taskq_wait_outstanding(arc_prune_taskq
, 0);
2749 * Evict the specified number of bytes from the state specified,
2750 * restricting eviction to the spa and type given. This function
2751 * prevents us from trying to evict more from a state's list than
2752 * is "evictable", and to skip evicting altogether when passed a
2753 * negative value for "bytes". In contrast, arc_evict_state() will
2754 * evict everything it can, when passed a negative value for "bytes".
2757 arc_adjust_impl(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
2758 arc_buf_contents_t type
)
2762 if (bytes
> 0 && state
->arcs_lsize
[type
] > 0) {
2763 delta
= MIN(state
->arcs_lsize
[type
], bytes
);
2764 return (arc_evict_state(state
, spa
, delta
, type
));
2771 * The goal of this function is to evict enough meta data buffers from the
2772 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
2773 * more complicated than it appears because it is common for data buffers
2774 * to have holds on meta data buffers. In addition, dnode meta data buffers
2775 * will be held by the dnodes in the block preventing them from being freed.
2776 * This means we can't simply traverse the ARC and expect to always find
2777 * enough unheld meta data buffer to release.
2779 * Therefore, this function has been updated to make alternating passes
2780 * over the ARC releasing data buffers and then newly unheld meta data
2781 * buffers. This ensures forward progress is maintained and arc_meta_used
2782 * will decrease. Normally this is sufficient, but if required the ARC
2783 * will call the registered prune callbacks causing dentry and inodes to
2784 * be dropped from the VFS cache. This will make dnode meta data buffers
2785 * available for reclaim.
2788 arc_adjust_meta_balanced(void)
2790 int64_t adjustmnt
, delta
, prune
= 0;
2791 uint64_t total_evicted
= 0;
2792 arc_buf_contents_t type
= ARC_BUFC_DATA
;
2793 int restarts
= MAX(zfs_arc_meta_adjust_restarts
, 0);
2797 * This slightly differs than the way we evict from the mru in
2798 * arc_adjust because we don't have a "target" value (i.e. no
2799 * "meta" arc_p). As a result, I think we can completely
2800 * cannibalize the metadata in the MRU before we evict the
2801 * metadata from the MFU. I think we probably need to implement a
2802 * "metadata arc_p" value to do this properly.
2804 adjustmnt
= arc_meta_used
- arc_meta_limit
;
2806 if (adjustmnt
> 0 && arc_mru
->arcs_lsize
[type
] > 0) {
2807 delta
= MIN(arc_mru
->arcs_lsize
[type
], adjustmnt
);
2808 total_evicted
+= arc_adjust_impl(arc_mru
, 0, delta
, type
);
2813 * We can't afford to recalculate adjustmnt here. If we do,
2814 * new metadata buffers can sneak into the MRU or ANON lists,
2815 * thus penalize the MFU metadata. Although the fudge factor is
2816 * small, it has been empirically shown to be significant for
2817 * certain workloads (e.g. creating many empty directories). As
2818 * such, we use the original calculation for adjustmnt, and
2819 * simply decrement the amount of data evicted from the MRU.
2822 if (adjustmnt
> 0 && arc_mfu
->arcs_lsize
[type
] > 0) {
2823 delta
= MIN(arc_mfu
->arcs_lsize
[type
], adjustmnt
);
2824 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, delta
, type
);
2827 adjustmnt
= arc_meta_used
- arc_meta_limit
;
2829 if (adjustmnt
> 0 && arc_mru_ghost
->arcs_lsize
[type
] > 0) {
2830 delta
= MIN(adjustmnt
,
2831 arc_mru_ghost
->arcs_lsize
[type
]);
2832 total_evicted
+= arc_adjust_impl(arc_mru_ghost
, 0, delta
, type
);
2836 if (adjustmnt
> 0 && arc_mfu_ghost
->arcs_lsize
[type
] > 0) {
2837 delta
= MIN(adjustmnt
,
2838 arc_mfu_ghost
->arcs_lsize
[type
]);
2839 total_evicted
+= arc_adjust_impl(arc_mfu_ghost
, 0, delta
, type
);
2843 * If after attempting to make the requested adjustment to the ARC
2844 * the meta limit is still being exceeded then request that the
2845 * higher layers drop some cached objects which have holds on ARC
2846 * meta buffers. Requests to the upper layers will be made with
2847 * increasingly large scan sizes until the ARC is below the limit.
2849 if (arc_meta_used
> arc_meta_limit
) {
2850 if (type
== ARC_BUFC_DATA
) {
2851 type
= ARC_BUFC_METADATA
;
2853 type
= ARC_BUFC_DATA
;
2855 if (zfs_arc_meta_prune
) {
2856 prune
+= zfs_arc_meta_prune
;
2857 arc_prune_async(prune
);
2866 return (total_evicted
);
2870 * Evict metadata buffers from the cache, such that arc_meta_used is
2871 * capped by the arc_meta_limit tunable.
2874 arc_adjust_meta_only(void)
2876 uint64_t total_evicted
= 0;
2880 * If we're over the meta limit, we want to evict enough
2881 * metadata to get back under the meta limit. We don't want to
2882 * evict so much that we drop the MRU below arc_p, though. If
2883 * we're over the meta limit more than we're over arc_p, we
2884 * evict some from the MRU here, and some from the MFU below.
2886 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
2887 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
2888 refcount_count(&arc_mru
->arcs_size
) - arc_p
));
2890 total_evicted
+= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
2893 * Similar to the above, we want to evict enough bytes to get us
2894 * below the meta limit, but not so much as to drop us below the
2895 * space alloted to the MFU (which is defined as arc_c - arc_p).
2897 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
2898 (int64_t)(refcount_count(&arc_mfu
->arcs_size
) - (arc_c
- arc_p
)));
2900 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
2902 return (total_evicted
);
2906 arc_adjust_meta(void)
2908 if (zfs_arc_meta_strategy
== ARC_STRATEGY_META_ONLY
)
2909 return (arc_adjust_meta_only());
2911 return (arc_adjust_meta_balanced());
2915 * Return the type of the oldest buffer in the given arc state
2917 * This function will select a random sublist of type ARC_BUFC_DATA and
2918 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
2919 * is compared, and the type which contains the "older" buffer will be
2922 static arc_buf_contents_t
2923 arc_adjust_type(arc_state_t
*state
)
2925 multilist_t
*data_ml
= &state
->arcs_list
[ARC_BUFC_DATA
];
2926 multilist_t
*meta_ml
= &state
->arcs_list
[ARC_BUFC_METADATA
];
2927 int data_idx
= multilist_get_random_index(data_ml
);
2928 int meta_idx
= multilist_get_random_index(meta_ml
);
2929 multilist_sublist_t
*data_mls
;
2930 multilist_sublist_t
*meta_mls
;
2931 arc_buf_contents_t type
;
2932 arc_buf_hdr_t
*data_hdr
;
2933 arc_buf_hdr_t
*meta_hdr
;
2936 * We keep the sublist lock until we're finished, to prevent
2937 * the headers from being destroyed via arc_evict_state().
2939 data_mls
= multilist_sublist_lock(data_ml
, data_idx
);
2940 meta_mls
= multilist_sublist_lock(meta_ml
, meta_idx
);
2943 * These two loops are to ensure we skip any markers that
2944 * might be at the tail of the lists due to arc_evict_state().
2947 for (data_hdr
= multilist_sublist_tail(data_mls
); data_hdr
!= NULL
;
2948 data_hdr
= multilist_sublist_prev(data_mls
, data_hdr
)) {
2949 if (data_hdr
->b_spa
!= 0)
2953 for (meta_hdr
= multilist_sublist_tail(meta_mls
); meta_hdr
!= NULL
;
2954 meta_hdr
= multilist_sublist_prev(meta_mls
, meta_hdr
)) {
2955 if (meta_hdr
->b_spa
!= 0)
2959 if (data_hdr
== NULL
&& meta_hdr
== NULL
) {
2960 type
= ARC_BUFC_DATA
;
2961 } else if (data_hdr
== NULL
) {
2962 ASSERT3P(meta_hdr
, !=, NULL
);
2963 type
= ARC_BUFC_METADATA
;
2964 } else if (meta_hdr
== NULL
) {
2965 ASSERT3P(data_hdr
, !=, NULL
);
2966 type
= ARC_BUFC_DATA
;
2968 ASSERT3P(data_hdr
, !=, NULL
);
2969 ASSERT3P(meta_hdr
, !=, NULL
);
2971 /* The headers can't be on the sublist without an L1 header */
2972 ASSERT(HDR_HAS_L1HDR(data_hdr
));
2973 ASSERT(HDR_HAS_L1HDR(meta_hdr
));
2975 if (data_hdr
->b_l1hdr
.b_arc_access
<
2976 meta_hdr
->b_l1hdr
.b_arc_access
) {
2977 type
= ARC_BUFC_DATA
;
2979 type
= ARC_BUFC_METADATA
;
2983 multilist_sublist_unlock(meta_mls
);
2984 multilist_sublist_unlock(data_mls
);
2990 * Evict buffers from the cache, such that arc_size is capped by arc_c.
2995 uint64_t total_evicted
= 0;
3000 * If we're over arc_meta_limit, we want to correct that before
3001 * potentially evicting data buffers below.
3003 total_evicted
+= arc_adjust_meta();
3008 * If we're over the target cache size, we want to evict enough
3009 * from the list to get back to our target size. We don't want
3010 * to evict too much from the MRU, such that it drops below
3011 * arc_p. So, if we're over our target cache size more than
3012 * the MRU is over arc_p, we'll evict enough to get back to
3013 * arc_p here, and then evict more from the MFU below.
3015 target
= MIN((int64_t)(arc_size
- arc_c
),
3016 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
3017 refcount_count(&arc_mru
->arcs_size
) + arc_meta_used
- arc_p
));
3020 * If we're below arc_meta_min, always prefer to evict data.
3021 * Otherwise, try to satisfy the requested number of bytes to
3022 * evict from the type which contains older buffers; in an
3023 * effort to keep newer buffers in the cache regardless of their
3024 * type. If we cannot satisfy the number of bytes from this
3025 * type, spill over into the next type.
3027 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
3028 arc_meta_used
> arc_meta_min
) {
3029 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3030 total_evicted
+= bytes
;
3033 * If we couldn't evict our target number of bytes from
3034 * metadata, we try to get the rest from data.
3039 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3041 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3042 total_evicted
+= bytes
;
3045 * If we couldn't evict our target number of bytes from
3046 * data, we try to get the rest from metadata.
3051 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3057 * Now that we've tried to evict enough from the MRU to get its
3058 * size back to arc_p, if we're still above the target cache
3059 * size, we evict the rest from the MFU.
3061 target
= arc_size
- arc_c
;
3063 if (arc_adjust_type(arc_mfu
) == ARC_BUFC_METADATA
&&
3064 arc_meta_used
> arc_meta_min
) {
3065 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3066 total_evicted
+= bytes
;
3069 * If we couldn't evict our target number of bytes from
3070 * metadata, we try to get the rest from data.
3075 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3077 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3078 total_evicted
+= bytes
;
3081 * If we couldn't evict our target number of bytes from
3082 * data, we try to get the rest from data.
3087 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3091 * Adjust ghost lists
3093 * In addition to the above, the ARC also defines target values
3094 * for the ghost lists. The sum of the mru list and mru ghost
3095 * list should never exceed the target size of the cache, and
3096 * the sum of the mru list, mfu list, mru ghost list, and mfu
3097 * ghost list should never exceed twice the target size of the
3098 * cache. The following logic enforces these limits on the ghost
3099 * caches, and evicts from them as needed.
3101 target
= refcount_count(&arc_mru
->arcs_size
) +
3102 refcount_count(&arc_mru_ghost
->arcs_size
) - arc_c
;
3104 bytes
= arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_DATA
);
3105 total_evicted
+= bytes
;
3110 arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_METADATA
);
3113 * We assume the sum of the mru list and mfu list is less than
3114 * or equal to arc_c (we enforced this above), which means we
3115 * can use the simpler of the two equations below:
3117 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3118 * mru ghost + mfu ghost <= arc_c
3120 target
= refcount_count(&arc_mru_ghost
->arcs_size
) +
3121 refcount_count(&arc_mfu_ghost
->arcs_size
) - arc_c
;
3123 bytes
= arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_DATA
);
3124 total_evicted
+= bytes
;
3129 arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_METADATA
);
3131 return (total_evicted
);
3135 arc_do_user_evicts(void)
3137 mutex_enter(&arc_user_evicts_lock
);
3138 while (arc_eviction_list
!= NULL
) {
3139 arc_buf_t
*buf
= arc_eviction_list
;
3140 arc_eviction_list
= buf
->b_next
;
3141 mutex_enter(&buf
->b_evict_lock
);
3143 mutex_exit(&buf
->b_evict_lock
);
3144 mutex_exit(&arc_user_evicts_lock
);
3146 if (buf
->b_efunc
!= NULL
)
3147 VERIFY0(buf
->b_efunc(buf
->b_private
));
3149 buf
->b_efunc
= NULL
;
3150 buf
->b_private
= NULL
;
3151 kmem_cache_free(buf_cache
, buf
);
3152 mutex_enter(&arc_user_evicts_lock
);
3154 mutex_exit(&arc_user_evicts_lock
);
3158 arc_flush(spa_t
*spa
, boolean_t retry
)
3163 * If retry is TRUE, a spa must not be specified since we have
3164 * no good way to determine if all of a spa's buffers have been
3165 * evicted from an arc state.
3167 ASSERT(!retry
|| spa
== 0);
3170 guid
= spa_load_guid(spa
);
3172 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_DATA
, retry
);
3173 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_METADATA
, retry
);
3175 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_DATA
, retry
);
3176 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_METADATA
, retry
);
3178 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3179 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3181 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3182 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3184 arc_do_user_evicts();
3185 ASSERT(spa
|| arc_eviction_list
== NULL
);
3189 arc_shrink(int64_t to_free
)
3191 if (arc_c
> arc_c_min
) {
3193 if (arc_c
> arc_c_min
+ to_free
)
3194 atomic_add_64(&arc_c
, -to_free
);
3198 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
3199 if (arc_c
> arc_size
)
3200 arc_c
= MAX(arc_size
, arc_c_min
);
3202 arc_p
= (arc_c
>> 1);
3203 ASSERT(arc_c
>= arc_c_min
);
3204 ASSERT((int64_t)arc_p
>= 0);
3207 if (arc_size
> arc_c
)
3208 (void) arc_adjust();
3211 typedef enum free_memory_reason_t
{
3216 FMR_PAGES_PP_MAXIMUM
,
3219 } free_memory_reason_t
;
3221 int64_t last_free_memory
;
3222 free_memory_reason_t last_free_reason
;
3227 * expiration time for arc_no_grow set by direct memory reclaim.
3229 static clock_t arc_grow_time
= 0;
3232 * Additional reserve of pages for pp_reserve.
3234 int64_t arc_pages_pp_reserve
= 64;
3237 * Additional reserve of pages for swapfs.
3239 int64_t arc_swapfs_reserve
= 64;
3241 #endif /* _KERNEL */
3244 * Return the amount of memory that can be consumed before reclaim will be
3245 * needed. Positive if there is sufficient free memory, negative indicates
3246 * the amount of memory that needs to be freed up.
3249 arc_available_memory(void)
3251 int64_t lowest
= INT64_MAX
;
3252 free_memory_reason_t r
= FMR_UNKNOWN
;
3257 * Under Linux we are not allowed to directly interrogate the global
3258 * memory state. Instead rely on observing that direct reclaim has
3259 * recently occurred therefore the system must be low on memory. The
3260 * exact values returned are not critical but should be small.
3262 if (ddi_time_after_eq(ddi_get_lbolt(), arc_grow_time
))
3265 lowest
= -PAGE_SIZE
;
3270 * Platforms like illumos have greater visibility in to the memory
3271 * subsystem and can return a more detailed analysis of memory.
3274 n
= PAGESIZE
* (-needfree
);
3282 * check that we're out of range of the pageout scanner. It starts to
3283 * schedule paging if freemem is less than lotsfree and needfree.
3284 * lotsfree is the high-water mark for pageout, and needfree is the
3285 * number of needed free pages. We add extra pages here to make sure
3286 * the scanner doesn't start up while we're freeing memory.
3288 n
= PAGESIZE
* (freemem
- lotsfree
- needfree
- desfree
);
3295 * check to make sure that swapfs has enough space so that anon
3296 * reservations can still succeed. anon_resvmem() checks that the
3297 * availrmem is greater than swapfs_minfree, and the number of reserved
3298 * swap pages. We also add a bit of extra here just to prevent
3299 * circumstances from getting really dire.
3301 n
= PAGESIZE
* (availrmem
- swapfs_minfree
- swapfs_reserve
-
3302 desfree
- arc_swapfs_reserve
);
3305 r
= FMR_SWAPFS_MINFREE
;
3310 * Check that we have enough availrmem that memory locking (e.g., via
3311 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3312 * stores the number of pages that cannot be locked; when availrmem
3313 * drops below pages_pp_maximum, page locking mechanisms such as
3314 * page_pp_lock() will fail.)
3316 n
= PAGESIZE
* (availrmem
- pages_pp_maximum
-
3317 arc_pages_pp_reserve
);
3320 r
= FMR_PAGES_PP_MAXIMUM
;
3325 * If we're on an i386 platform, it's possible that we'll exhaust the
3326 * kernel heap space before we ever run out of available physical
3327 * memory. Most checks of the size of the heap_area compare against
3328 * tune.t_minarmem, which is the minimum available real memory that we
3329 * can have in the system. However, this is generally fixed at 25 pages
3330 * which is so low that it's useless. In this comparison, we seek to
3331 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3332 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3335 n
= vmem_size(heap_arena
, VMEM_FREE
) -
3336 (vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
) >> 2);
3344 * If zio data pages are being allocated out of a separate heap segment,
3345 * then enforce that the size of available vmem for this arena remains
3346 * above about 1/16th free.
3348 * Note: The 1/16th arena free requirement was put in place
3349 * to aggressively evict memory from the arc in order to avoid
3350 * memory fragmentation issues.
3352 if (zio_arena
!= NULL
) {
3353 n
= vmem_size(zio_arena
, VMEM_FREE
) -
3354 (vmem_size(zio_arena
, VMEM_ALLOC
) >> 4);
3360 #endif /* __linux__ */
3362 /* Every 100 calls, free a small amount */
3363 if (spa_get_random(100) == 0)
3367 last_free_memory
= lowest
;
3368 last_free_reason
= r
;
3374 * Determine if the system is under memory pressure and is asking
3375 * to reclaim memory. A return value of TRUE indicates that the system
3376 * is under memory pressure and that the arc should adjust accordingly.
3379 arc_reclaim_needed(void)
3381 return (arc_available_memory() < 0);
3385 arc_kmem_reap_now(void)
3388 kmem_cache_t
*prev_cache
= NULL
;
3389 kmem_cache_t
*prev_data_cache
= NULL
;
3390 extern kmem_cache_t
*zio_buf_cache
[];
3391 extern kmem_cache_t
*zio_data_buf_cache
[];
3392 extern kmem_cache_t
*range_seg_cache
;
3394 if ((arc_meta_used
>= arc_meta_limit
) && zfs_arc_meta_prune
) {
3396 * We are exceeding our meta-data cache limit.
3397 * Prune some entries to release holds on meta-data.
3399 arc_prune(zfs_arc_meta_prune
);
3402 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
3403 if (zio_buf_cache
[i
] != prev_cache
) {
3404 prev_cache
= zio_buf_cache
[i
];
3405 kmem_cache_reap_now(zio_buf_cache
[i
]);
3407 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
3408 prev_data_cache
= zio_data_buf_cache
[i
];
3409 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
3412 kmem_cache_reap_now(buf_cache
);
3413 kmem_cache_reap_now(hdr_full_cache
);
3414 kmem_cache_reap_now(hdr_l2only_cache
);
3415 kmem_cache_reap_now(range_seg_cache
);
3417 if (zio_arena
!= NULL
) {
3419 * Ask the vmem arena to reclaim unused memory from its
3422 vmem_qcache_reap(zio_arena
);
3427 * Threads can block in arc_get_data_buf() waiting for this thread to evict
3428 * enough data and signal them to proceed. When this happens, the threads in
3429 * arc_get_data_buf() are sleeping while holding the hash lock for their
3430 * particular arc header. Thus, we must be careful to never sleep on a
3431 * hash lock in this thread. This is to prevent the following deadlock:
3433 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3434 * waiting for the reclaim thread to signal it.
3436 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3437 * fails, and goes to sleep forever.
3439 * This possible deadlock is avoided by always acquiring a hash lock
3440 * using mutex_tryenter() from arc_reclaim_thread().
3443 arc_reclaim_thread(void)
3445 fstrans_cookie_t cookie
= spl_fstrans_mark();
3446 clock_t growtime
= 0;
3449 CALLB_CPR_INIT(&cpr
, &arc_reclaim_lock
, callb_generic_cpr
, FTAG
);
3451 mutex_enter(&arc_reclaim_lock
);
3452 while (!arc_reclaim_thread_exit
) {
3454 int64_t free_memory
= arc_available_memory();
3455 uint64_t evicted
= 0;
3457 arc_tuning_update();
3459 mutex_exit(&arc_reclaim_lock
);
3461 if (free_memory
< 0) {
3463 arc_no_grow
= B_TRUE
;
3467 * Wait at least zfs_grow_retry (default 5) seconds
3468 * before considering growing.
3470 growtime
= ddi_get_lbolt() + (arc_grow_retry
* hz
);
3472 arc_kmem_reap_now();
3475 * If we are still low on memory, shrink the ARC
3476 * so that we have arc_shrink_min free space.
3478 free_memory
= arc_available_memory();
3480 to_free
= (arc_c
>> arc_shrink_shift
) - free_memory
;
3483 to_free
= MAX(to_free
, ptob(needfree
));
3485 arc_shrink(to_free
);
3487 } else if (free_memory
< arc_c
>> arc_no_grow_shift
) {
3488 arc_no_grow
= B_TRUE
;
3489 } else if (ddi_get_lbolt() >= growtime
) {
3490 arc_no_grow
= B_FALSE
;
3493 evicted
= arc_adjust();
3495 mutex_enter(&arc_reclaim_lock
);
3498 * If evicted is zero, we couldn't evict anything via
3499 * arc_adjust(). This could be due to hash lock
3500 * collisions, but more likely due to the majority of
3501 * arc buffers being unevictable. Therefore, even if
3502 * arc_size is above arc_c, another pass is unlikely to
3503 * be helpful and could potentially cause us to enter an
3506 if (arc_size
<= arc_c
|| evicted
== 0) {
3508 * We're either no longer overflowing, or we
3509 * can't evict anything more, so we should wake
3510 * up any threads before we go to sleep.
3512 cv_broadcast(&arc_reclaim_waiters_cv
);
3515 * Block until signaled, or after one second (we
3516 * might need to perform arc_kmem_reap_now()
3517 * even if we aren't being signalled)
3519 CALLB_CPR_SAFE_BEGIN(&cpr
);
3520 (void) cv_timedwait_sig(&arc_reclaim_thread_cv
,
3521 &arc_reclaim_lock
, ddi_get_lbolt() + hz
);
3522 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_lock
);
3526 arc_reclaim_thread_exit
= FALSE
;
3527 cv_broadcast(&arc_reclaim_thread_cv
);
3528 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_lock */
3529 spl_fstrans_unmark(cookie
);
3534 arc_user_evicts_thread(void)
3536 fstrans_cookie_t cookie
= spl_fstrans_mark();
3539 CALLB_CPR_INIT(&cpr
, &arc_user_evicts_lock
, callb_generic_cpr
, FTAG
);
3541 mutex_enter(&arc_user_evicts_lock
);
3542 while (!arc_user_evicts_thread_exit
) {
3543 mutex_exit(&arc_user_evicts_lock
);
3545 arc_do_user_evicts();
3548 * This is necessary in order for the mdb ::arc dcmd to
3549 * show up to date information. Since the ::arc command
3550 * does not call the kstat's update function, without
3551 * this call, the command may show stale stats for the
3552 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3553 * with this change, the data might be up to 1 second
3554 * out of date; but that should suffice. The arc_state_t
3555 * structures can be queried directly if more accurate
3556 * information is needed.
3558 if (arc_ksp
!= NULL
)
3559 arc_ksp
->ks_update(arc_ksp
, KSTAT_READ
);
3561 mutex_enter(&arc_user_evicts_lock
);
3564 * Block until signaled, or after one second (we need to
3565 * call the arc's kstat update function regularly).
3567 CALLB_CPR_SAFE_BEGIN(&cpr
);
3568 (void) cv_timedwait_sig(&arc_user_evicts_cv
,
3569 &arc_user_evicts_lock
, ddi_get_lbolt() + hz
);
3570 CALLB_CPR_SAFE_END(&cpr
, &arc_user_evicts_lock
);
3573 arc_user_evicts_thread_exit
= FALSE
;
3574 cv_broadcast(&arc_user_evicts_cv
);
3575 CALLB_CPR_EXIT(&cpr
); /* drops arc_user_evicts_lock */
3576 spl_fstrans_unmark(cookie
);
3582 * Determine the amount of memory eligible for eviction contained in the
3583 * ARC. All clean data reported by the ghost lists can always be safely
3584 * evicted. Due to arc_c_min, the same does not hold for all clean data
3585 * contained by the regular mru and mfu lists.
3587 * In the case of the regular mru and mfu lists, we need to report as
3588 * much clean data as possible, such that evicting that same reported
3589 * data will not bring arc_size below arc_c_min. Thus, in certain
3590 * circumstances, the total amount of clean data in the mru and mfu
3591 * lists might not actually be evictable.
3593 * The following two distinct cases are accounted for:
3595 * 1. The sum of the amount of dirty data contained by both the mru and
3596 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
3597 * is greater than or equal to arc_c_min.
3598 * (i.e. amount of dirty data >= arc_c_min)
3600 * This is the easy case; all clean data contained by the mru and mfu
3601 * lists is evictable. Evicting all clean data can only drop arc_size
3602 * to the amount of dirty data, which is greater than arc_c_min.
3604 * 2. The sum of the amount of dirty data contained by both the mru and
3605 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
3606 * is less than arc_c_min.
3607 * (i.e. arc_c_min > amount of dirty data)
3609 * 2.1. arc_size is greater than or equal arc_c_min.
3610 * (i.e. arc_size >= arc_c_min > amount of dirty data)
3612 * In this case, not all clean data from the regular mru and mfu
3613 * lists is actually evictable; we must leave enough clean data
3614 * to keep arc_size above arc_c_min. Thus, the maximum amount of
3615 * evictable data from the two lists combined, is exactly the
3616 * difference between arc_size and arc_c_min.
3618 * 2.2. arc_size is less than arc_c_min
3619 * (i.e. arc_c_min > arc_size > amount of dirty data)
3621 * In this case, none of the data contained in the mru and mfu
3622 * lists is evictable, even if it's clean. Since arc_size is
3623 * already below arc_c_min, evicting any more would only
3624 * increase this negative difference.
3627 arc_evictable_memory(void) {
3628 uint64_t arc_clean
=
3629 arc_mru
->arcs_lsize
[ARC_BUFC_DATA
] +
3630 arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] +
3631 arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
] +
3632 arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
];
3633 uint64_t ghost_clean
=
3634 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
3635 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_METADATA
] +
3636 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
3637 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_METADATA
];
3638 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
3640 if (arc_dirty
>= arc_c_min
)
3641 return (ghost_clean
+ arc_clean
);
3643 return (ghost_clean
+ MAX((int64_t)arc_size
- (int64_t)arc_c_min
, 0));
3647 * If sc->nr_to_scan is zero, the caller is requesting a query of the
3648 * number of objects which can potentially be freed. If it is nonzero,
3649 * the request is to free that many objects.
3651 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
3652 * in struct shrinker and also require the shrinker to return the number
3655 * Older kernels require the shrinker to return the number of freeable
3656 * objects following the freeing of nr_to_free.
3658 static spl_shrinker_t
3659 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
3663 /* The arc is considered warm once reclaim has occurred */
3664 if (unlikely(arc_warm
== B_FALSE
))
3667 /* Return the potential number of reclaimable pages */
3668 pages
= btop((int64_t)arc_evictable_memory());
3669 if (sc
->nr_to_scan
== 0)
3672 /* Not allowed to perform filesystem reclaim */
3673 if (!(sc
->gfp_mask
& __GFP_FS
))
3674 return (SHRINK_STOP
);
3676 /* Reclaim in progress */
3677 if (mutex_tryenter(&arc_reclaim_lock
) == 0)
3678 return (SHRINK_STOP
);
3680 mutex_exit(&arc_reclaim_lock
);
3683 * Evict the requested number of pages by shrinking arc_c the
3684 * requested amount. If there is nothing left to evict just
3685 * reap whatever we can from the various arc slabs.
3688 arc_shrink(ptob(sc
->nr_to_scan
));
3689 arc_kmem_reap_now();
3690 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
3691 pages
= MAX(pages
- btop(arc_evictable_memory()), 0);
3693 pages
= btop(arc_evictable_memory());
3696 arc_kmem_reap_now();
3697 pages
= SHRINK_STOP
;
3701 * We've reaped what we can, wake up threads.
3703 cv_broadcast(&arc_reclaim_waiters_cv
);
3706 * When direct reclaim is observed it usually indicates a rapid
3707 * increase in memory pressure. This occurs because the kswapd
3708 * threads were unable to asynchronously keep enough free memory
3709 * available. In this case set arc_no_grow to briefly pause arc
3710 * growth to avoid compounding the memory pressure.
3712 if (current_is_kswapd()) {
3713 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
3715 arc_no_grow
= B_TRUE
;
3716 arc_grow_time
= ddi_get_lbolt() + (zfs_arc_grow_retry
* hz
);
3717 ARCSTAT_BUMP(arcstat_memory_direct_count
);
3722 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
3724 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
3725 #endif /* _KERNEL */
3728 * Adapt arc info given the number of bytes we are trying to add and
3729 * the state that we are comming from. This function is only called
3730 * when we are adding new content to the cache.
3733 arc_adapt(int bytes
, arc_state_t
*state
)
3736 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
3737 int64_t mrug_size
= refcount_count(&arc_mru_ghost
->arcs_size
);
3738 int64_t mfug_size
= refcount_count(&arc_mfu_ghost
->arcs_size
);
3740 if (state
== arc_l2c_only
)
3745 * Adapt the target size of the MRU list:
3746 * - if we just hit in the MRU ghost list, then increase
3747 * the target size of the MRU list.
3748 * - if we just hit in the MFU ghost list, then increase
3749 * the target size of the MFU list by decreasing the
3750 * target size of the MRU list.
3752 if (state
== arc_mru_ghost
) {
3753 mult
= (mrug_size
>= mfug_size
) ? 1 : (mfug_size
/ mrug_size
);
3754 if (!zfs_arc_p_dampener_disable
)
3755 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
3757 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
3758 } else if (state
== arc_mfu_ghost
) {
3761 mult
= (mfug_size
>= mrug_size
) ? 1 : (mrug_size
/ mfug_size
);
3762 if (!zfs_arc_p_dampener_disable
)
3763 mult
= MIN(mult
, 10);
3765 delta
= MIN(bytes
* mult
, arc_p
);
3766 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
3768 ASSERT((int64_t)arc_p
>= 0);
3770 if (arc_reclaim_needed()) {
3771 cv_signal(&arc_reclaim_thread_cv
);
3778 if (arc_c
>= arc_c_max
)
3782 * If we're within (2 * maxblocksize) bytes of the target
3783 * cache size, increment the target cache size
3785 VERIFY3U(arc_c
, >=, 2ULL << SPA_MAXBLOCKSHIFT
);
3786 if (arc_size
>= arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
3787 atomic_add_64(&arc_c
, (int64_t)bytes
);
3788 if (arc_c
> arc_c_max
)
3790 else if (state
== arc_anon
)
3791 atomic_add_64(&arc_p
, (int64_t)bytes
);
3795 ASSERT((int64_t)arc_p
>= 0);
3799 * Check if arc_size has grown past our upper threshold, determined by
3800 * zfs_arc_overflow_shift.
3803 arc_is_overflowing(void)
3805 /* Always allow at least one block of overflow */
3806 uint64_t overflow
= MAX(SPA_MAXBLOCKSIZE
,
3807 arc_c
>> zfs_arc_overflow_shift
);
3809 return (arc_size
>= arc_c
+ overflow
);
3813 * The buffer, supplied as the first argument, needs a data block. If we
3814 * are hitting the hard limit for the cache size, we must sleep, waiting
3815 * for the eviction thread to catch up. If we're past the target size
3816 * but below the hard limit, we'll only signal the reclaim thread and
3820 arc_get_data_buf(arc_buf_t
*buf
)
3822 arc_state_t
*state
= buf
->b_hdr
->b_l1hdr
.b_state
;
3823 uint64_t size
= buf
->b_hdr
->b_size
;
3824 arc_buf_contents_t type
= arc_buf_type(buf
->b_hdr
);
3826 arc_adapt(size
, state
);
3829 * If arc_size is currently overflowing, and has grown past our
3830 * upper limit, we must be adding data faster than the evict
3831 * thread can evict. Thus, to ensure we don't compound the
3832 * problem by adding more data and forcing arc_size to grow even
3833 * further past it's target size, we halt and wait for the
3834 * eviction thread to catch up.
3836 * It's also possible that the reclaim thread is unable to evict
3837 * enough buffers to get arc_size below the overflow limit (e.g.
3838 * due to buffers being un-evictable, or hash lock collisions).
3839 * In this case, we want to proceed regardless if we're
3840 * overflowing; thus we don't use a while loop here.
3842 if (arc_is_overflowing()) {
3843 mutex_enter(&arc_reclaim_lock
);
3846 * Now that we've acquired the lock, we may no longer be
3847 * over the overflow limit, lets check.
3849 * We're ignoring the case of spurious wake ups. If that
3850 * were to happen, it'd let this thread consume an ARC
3851 * buffer before it should have (i.e. before we're under
3852 * the overflow limit and were signalled by the reclaim
3853 * thread). As long as that is a rare occurrence, it
3854 * shouldn't cause any harm.
3856 if (arc_is_overflowing()) {
3857 cv_signal(&arc_reclaim_thread_cv
);
3858 cv_wait(&arc_reclaim_waiters_cv
, &arc_reclaim_lock
);
3861 mutex_exit(&arc_reclaim_lock
);
3864 if (type
== ARC_BUFC_METADATA
) {
3865 buf
->b_data
= zio_buf_alloc(size
);
3866 arc_space_consume(size
, ARC_SPACE_META
);
3868 ASSERT(type
== ARC_BUFC_DATA
);
3869 buf
->b_data
= zio_data_buf_alloc(size
);
3870 arc_space_consume(size
, ARC_SPACE_DATA
);
3874 * Update the state size. Note that ghost states have a
3875 * "ghost size" and so don't need to be updated.
3877 if (!GHOST_STATE(buf
->b_hdr
->b_l1hdr
.b_state
)) {
3878 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3879 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
3881 (void) refcount_add_many(&state
->arcs_size
, size
, buf
);
3884 * If this is reached via arc_read, the link is
3885 * protected by the hash lock. If reached via
3886 * arc_buf_alloc, the header should not be accessed by
3887 * any other thread. And, if reached via arc_read_done,
3888 * the hash lock will protect it if it's found in the
3889 * hash table; otherwise no other thread should be
3890 * trying to [add|remove]_reference it.
3892 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
3893 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3894 atomic_add_64(&hdr
->b_l1hdr
.b_state
->arcs_lsize
[type
],
3898 * If we are growing the cache, and we are adding anonymous
3899 * data, and we have outgrown arc_p, update arc_p
3901 if (arc_size
< arc_c
&& hdr
->b_l1hdr
.b_state
== arc_anon
&&
3902 (refcount_count(&arc_anon
->arcs_size
) +
3903 refcount_count(&arc_mru
->arcs_size
) > arc_p
))
3904 arc_p
= MIN(arc_c
, arc_p
+ size
);
3909 * This routine is called whenever a buffer is accessed.
3910 * NOTE: the hash lock is dropped in this function.
3913 arc_access(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
3917 ASSERT(MUTEX_HELD(hash_lock
));
3918 ASSERT(HDR_HAS_L1HDR(hdr
));
3920 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
3922 * This buffer is not in the cache, and does not
3923 * appear in our "ghost" list. Add the new buffer
3927 ASSERT0(hdr
->b_l1hdr
.b_arc_access
);
3928 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3929 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
3930 arc_change_state(arc_mru
, hdr
, hash_lock
);
3932 } else if (hdr
->b_l1hdr
.b_state
== arc_mru
) {
3933 now
= ddi_get_lbolt();
3936 * If this buffer is here because of a prefetch, then either:
3937 * - clear the flag if this is a "referencing" read
3938 * (any subsequent access will bump this into the MFU state).
3940 * - move the buffer to the head of the list if this is
3941 * another prefetch (to make it less likely to be evicted).
3943 if (HDR_PREFETCH(hdr
)) {
3944 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
3945 /* link protected by hash lock */
3946 ASSERT(multilist_link_active(
3947 &hdr
->b_l1hdr
.b_arc_node
));
3949 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
3950 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
3951 ARCSTAT_BUMP(arcstat_mru_hits
);
3953 hdr
->b_l1hdr
.b_arc_access
= now
;
3958 * This buffer has been "accessed" only once so far,
3959 * but it is still in the cache. Move it to the MFU
3962 if (ddi_time_after(now
, hdr
->b_l1hdr
.b_arc_access
+
3965 * More than 125ms have passed since we
3966 * instantiated this buffer. Move it to the
3967 * most frequently used state.
3969 hdr
->b_l1hdr
.b_arc_access
= now
;
3970 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
3971 arc_change_state(arc_mfu
, hdr
, hash_lock
);
3973 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
3974 ARCSTAT_BUMP(arcstat_mru_hits
);
3975 } else if (hdr
->b_l1hdr
.b_state
== arc_mru_ghost
) {
3976 arc_state_t
*new_state
;
3978 * This buffer has been "accessed" recently, but
3979 * was evicted from the cache. Move it to the
3983 if (HDR_PREFETCH(hdr
)) {
3984 new_state
= arc_mru
;
3985 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0)
3986 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
3987 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
3989 new_state
= arc_mfu
;
3990 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
3993 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3994 arc_change_state(new_state
, hdr
, hash_lock
);
3996 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_ghost_hits
);
3997 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
3998 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu
) {
4000 * This buffer has been accessed more than once and is
4001 * still in the cache. Keep it in the MFU state.
4003 * NOTE: an add_reference() that occurred when we did
4004 * the arc_read() will have kicked this off the list.
4005 * If it was a prefetch, we will explicitly move it to
4006 * the head of the list now.
4008 if ((HDR_PREFETCH(hdr
)) != 0) {
4009 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4010 /* link protected by hash_lock */
4011 ASSERT(multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4013 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_hits
);
4014 ARCSTAT_BUMP(arcstat_mfu_hits
);
4015 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4016 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu_ghost
) {
4017 arc_state_t
*new_state
= arc_mfu
;
4019 * This buffer has been accessed more than once but has
4020 * been evicted from the cache. Move it back to the
4024 if (HDR_PREFETCH(hdr
)) {
4026 * This is a prefetch access...
4027 * move this block back to the MRU state.
4029 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
4030 new_state
= arc_mru
;
4033 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4034 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4035 arc_change_state(new_state
, hdr
, hash_lock
);
4037 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_ghost_hits
);
4038 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
4039 } else if (hdr
->b_l1hdr
.b_state
== arc_l2c_only
) {
4041 * This buffer is on the 2nd Level ARC.
4044 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4045 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4046 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4048 cmn_err(CE_PANIC
, "invalid arc state 0x%p",
4049 hdr
->b_l1hdr
.b_state
);
4053 /* a generic arc_done_func_t which you can use */
4056 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4058 if (zio
== NULL
|| zio
->io_error
== 0)
4059 bcopy(buf
->b_data
, arg
, buf
->b_hdr
->b_size
);
4060 VERIFY(arc_buf_remove_ref(buf
, arg
));
4063 /* a generic arc_done_func_t */
4065 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4067 arc_buf_t
**bufp
= arg
;
4068 if (zio
&& zio
->io_error
) {
4069 VERIFY(arc_buf_remove_ref(buf
, arg
));
4073 ASSERT(buf
->b_data
);
4078 arc_read_done(zio_t
*zio
)
4082 arc_buf_t
*abuf
; /* buffer we're assigning to callback */
4083 kmutex_t
*hash_lock
= NULL
;
4084 arc_callback_t
*callback_list
, *acb
;
4085 int freeable
= FALSE
;
4087 buf
= zio
->io_private
;
4091 * The hdr was inserted into hash-table and removed from lists
4092 * prior to starting I/O. We should find this header, since
4093 * it's in the hash table, and it should be legit since it's
4094 * not possible to evict it during the I/O. The only possible
4095 * reason for it not to be found is if we were freed during the
4098 if (HDR_IN_HASH_TABLE(hdr
)) {
4099 arc_buf_hdr_t
*found
;
4101 ASSERT3U(hdr
->b_birth
, ==, BP_PHYSICAL_BIRTH(zio
->io_bp
));
4102 ASSERT3U(hdr
->b_dva
.dva_word
[0], ==,
4103 BP_IDENTITY(zio
->io_bp
)->dva_word
[0]);
4104 ASSERT3U(hdr
->b_dva
.dva_word
[1], ==,
4105 BP_IDENTITY(zio
->io_bp
)->dva_word
[1]);
4107 found
= buf_hash_find(hdr
->b_spa
, zio
->io_bp
,
4110 ASSERT((found
== NULL
&& HDR_FREED_IN_READ(hdr
) &&
4111 hash_lock
== NULL
) ||
4113 DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
4114 (found
== hdr
&& HDR_L2_READING(hdr
)));
4117 hdr
->b_flags
&= ~ARC_FLAG_L2_EVICTED
;
4118 if (l2arc_noprefetch
&& HDR_PREFETCH(hdr
))
4119 hdr
->b_flags
&= ~ARC_FLAG_L2CACHE
;
4121 /* byteswap if necessary */
4122 callback_list
= hdr
->b_l1hdr
.b_acb
;
4123 ASSERT(callback_list
!= NULL
);
4124 if (BP_SHOULD_BYTESWAP(zio
->io_bp
) && zio
->io_error
== 0) {
4125 dmu_object_byteswap_t bswap
=
4126 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
4127 if (BP_GET_LEVEL(zio
->io_bp
) > 0)
4128 byteswap_uint64_array(buf
->b_data
, hdr
->b_size
);
4130 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, hdr
->b_size
);
4133 arc_cksum_compute(buf
, B_FALSE
);
4136 if (hash_lock
&& zio
->io_error
== 0 &&
4137 hdr
->b_l1hdr
.b_state
== arc_anon
) {
4139 * Only call arc_access on anonymous buffers. This is because
4140 * if we've issued an I/O for an evicted buffer, we've already
4141 * called arc_access (to prevent any simultaneous readers from
4142 * getting confused).
4144 arc_access(hdr
, hash_lock
);
4147 /* create copies of the data buffer for the callers */
4149 for (acb
= callback_list
; acb
; acb
= acb
->acb_next
) {
4150 if (acb
->acb_done
) {
4152 ARCSTAT_BUMP(arcstat_duplicate_reads
);
4153 abuf
= arc_buf_clone(buf
);
4155 acb
->acb_buf
= abuf
;
4159 hdr
->b_l1hdr
.b_acb
= NULL
;
4160 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
4161 ASSERT(!HDR_BUF_AVAILABLE(hdr
));
4163 ASSERT(buf
->b_efunc
== NULL
);
4164 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
4165 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
4168 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
) ||
4169 callback_list
!= NULL
);
4171 if (zio
->io_error
!= 0) {
4172 hdr
->b_flags
|= ARC_FLAG_IO_ERROR
;
4173 if (hdr
->b_l1hdr
.b_state
!= arc_anon
)
4174 arc_change_state(arc_anon
, hdr
, hash_lock
);
4175 if (HDR_IN_HASH_TABLE(hdr
))
4176 buf_hash_remove(hdr
);
4177 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
4181 * Broadcast before we drop the hash_lock to avoid the possibility
4182 * that the hdr (and hence the cv) might be freed before we get to
4183 * the cv_broadcast().
4185 cv_broadcast(&hdr
->b_l1hdr
.b_cv
);
4187 if (hash_lock
!= NULL
) {
4188 mutex_exit(hash_lock
);
4191 * This block was freed while we waited for the read to
4192 * complete. It has been removed from the hash table and
4193 * moved to the anonymous state (so that it won't show up
4196 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
4197 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
4200 /* execute each callback and free its structure */
4201 while ((acb
= callback_list
) != NULL
) {
4203 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
4205 if (acb
->acb_zio_dummy
!= NULL
) {
4206 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
4207 zio_nowait(acb
->acb_zio_dummy
);
4210 callback_list
= acb
->acb_next
;
4211 kmem_free(acb
, sizeof (arc_callback_t
));
4215 arc_hdr_destroy(hdr
);
4219 * "Read" the block at the specified DVA (in bp) via the
4220 * cache. If the block is found in the cache, invoke the provided
4221 * callback immediately and return. Note that the `zio' parameter
4222 * in the callback will be NULL in this case, since no IO was
4223 * required. If the block is not in the cache pass the read request
4224 * on to the spa with a substitute callback function, so that the
4225 * requested block will be added to the cache.
4227 * If a read request arrives for a block that has a read in-progress,
4228 * either wait for the in-progress read to complete (and return the
4229 * results); or, if this is a read with a "done" func, add a record
4230 * to the read to invoke the "done" func when the read completes,
4231 * and return; or just return.
4233 * arc_read_done() will invoke all the requested "done" functions
4234 * for readers of this block.
4237 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_done_func_t
*done
,
4238 void *private, zio_priority_t priority
, int zio_flags
,
4239 arc_flags_t
*arc_flags
, const zbookmark_phys_t
*zb
)
4241 arc_buf_hdr_t
*hdr
= NULL
;
4242 arc_buf_t
*buf
= NULL
;
4243 kmutex_t
*hash_lock
= NULL
;
4245 uint64_t guid
= spa_load_guid(spa
);
4248 ASSERT(!BP_IS_EMBEDDED(bp
) ||
4249 BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
);
4252 if (!BP_IS_EMBEDDED(bp
)) {
4254 * Embedded BP's have no DVA and require no I/O to "read".
4255 * Create an anonymous arc buf to back it.
4257 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
4260 if (hdr
!= NULL
&& HDR_HAS_L1HDR(hdr
) && hdr
->b_l1hdr
.b_datacnt
> 0) {
4262 *arc_flags
|= ARC_FLAG_CACHED
;
4264 if (HDR_IO_IN_PROGRESS(hdr
)) {
4266 if (*arc_flags
& ARC_FLAG_WAIT
) {
4267 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
4268 mutex_exit(hash_lock
);
4271 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
4274 arc_callback_t
*acb
= NULL
;
4276 acb
= kmem_zalloc(sizeof (arc_callback_t
),
4278 acb
->acb_done
= done
;
4279 acb
->acb_private
= private;
4281 acb
->acb_zio_dummy
= zio_null(pio
,
4282 spa
, NULL
, NULL
, NULL
, zio_flags
);
4284 ASSERT(acb
->acb_done
!= NULL
);
4285 acb
->acb_next
= hdr
->b_l1hdr
.b_acb
;
4286 hdr
->b_l1hdr
.b_acb
= acb
;
4287 add_reference(hdr
, hash_lock
, private);
4288 mutex_exit(hash_lock
);
4291 mutex_exit(hash_lock
);
4295 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
4296 hdr
->b_l1hdr
.b_state
== arc_mfu
);
4299 add_reference(hdr
, hash_lock
, private);
4301 * If this block is already in use, create a new
4302 * copy of the data so that we will be guaranteed
4303 * that arc_release() will always succeed.
4305 buf
= hdr
->b_l1hdr
.b_buf
;
4307 ASSERT(buf
->b_data
);
4308 if (HDR_BUF_AVAILABLE(hdr
)) {
4309 ASSERT(buf
->b_efunc
== NULL
);
4310 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
4312 buf
= arc_buf_clone(buf
);
4315 } else if (*arc_flags
& ARC_FLAG_PREFETCH
&&
4316 refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
4317 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
4319 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
4320 arc_access(hdr
, hash_lock
);
4321 if (*arc_flags
& ARC_FLAG_L2CACHE
)
4322 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
4323 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
4324 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
4325 mutex_exit(hash_lock
);
4326 ARCSTAT_BUMP(arcstat_hits
);
4327 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
4328 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
4329 data
, metadata
, hits
);
4332 done(NULL
, buf
, private);
4334 uint64_t size
= BP_GET_LSIZE(bp
);
4335 arc_callback_t
*acb
;
4338 boolean_t devw
= B_FALSE
;
4339 enum zio_compress b_compress
= ZIO_COMPRESS_OFF
;
4340 int32_t b_asize
= 0;
4343 * Gracefully handle a damaged logical block size as a
4344 * checksum error by passing a dummy zio to the done callback.
4346 if (size
> spa_maxblocksize(spa
)) {
4348 rzio
= zio_null(pio
, spa
, NULL
,
4349 NULL
, NULL
, zio_flags
);
4350 rzio
->io_error
= ECKSUM
;
4351 done(rzio
, buf
, private);
4359 /* this block is not in the cache */
4360 arc_buf_hdr_t
*exists
= NULL
;
4361 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
4362 buf
= arc_buf_alloc(spa
, size
, private, type
);
4364 if (!BP_IS_EMBEDDED(bp
)) {
4365 hdr
->b_dva
= *BP_IDENTITY(bp
);
4366 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
4367 exists
= buf_hash_insert(hdr
, &hash_lock
);
4369 if (exists
!= NULL
) {
4370 /* somebody beat us to the hash insert */
4371 mutex_exit(hash_lock
);
4372 buf_discard_identity(hdr
);
4373 (void) arc_buf_remove_ref(buf
, private);
4374 goto top
; /* restart the IO request */
4377 /* if this is a prefetch, we don't have a reference */
4378 if (*arc_flags
& ARC_FLAG_PREFETCH
) {
4379 (void) remove_reference(hdr
, hash_lock
,
4381 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
4383 if (*arc_flags
& ARC_FLAG_L2CACHE
)
4384 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
4385 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
4386 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
4387 if (BP_GET_LEVEL(bp
) > 0)
4388 hdr
->b_flags
|= ARC_FLAG_INDIRECT
;
4391 * This block is in the ghost cache. If it was L2-only
4392 * (and thus didn't have an L1 hdr), we realloc the
4393 * header to add an L1 hdr.
4395 if (!HDR_HAS_L1HDR(hdr
)) {
4396 hdr
= arc_hdr_realloc(hdr
, hdr_l2only_cache
,
4400 ASSERT(GHOST_STATE(hdr
->b_l1hdr
.b_state
));
4401 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4402 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4403 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
4405 /* if this is a prefetch, we don't have a reference */
4406 if (*arc_flags
& ARC_FLAG_PREFETCH
)
4407 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
4409 add_reference(hdr
, hash_lock
, private);
4410 if (*arc_flags
& ARC_FLAG_L2CACHE
)
4411 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
4412 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
4413 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
4414 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
4417 buf
->b_efunc
= NULL
;
4418 buf
->b_private
= NULL
;
4420 hdr
->b_l1hdr
.b_buf
= buf
;
4421 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
4422 hdr
->b_l1hdr
.b_datacnt
= 1;
4423 arc_get_data_buf(buf
);
4424 arc_access(hdr
, hash_lock
);
4427 ASSERT(!GHOST_STATE(hdr
->b_l1hdr
.b_state
));
4429 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
4430 acb
->acb_done
= done
;
4431 acb
->acb_private
= private;
4433 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
4434 hdr
->b_l1hdr
.b_acb
= acb
;
4435 hdr
->b_flags
|= ARC_FLAG_IO_IN_PROGRESS
;
4437 if (HDR_HAS_L2HDR(hdr
) &&
4438 (vd
= hdr
->b_l2hdr
.b_dev
->l2ad_vdev
) != NULL
) {
4439 devw
= hdr
->b_l2hdr
.b_dev
->l2ad_writing
;
4440 addr
= hdr
->b_l2hdr
.b_daddr
;
4441 b_compress
= HDR_GET_COMPRESS(hdr
);
4442 b_asize
= hdr
->b_l2hdr
.b_asize
;
4444 * Lock out device removal.
4446 if (vdev_is_dead(vd
) ||
4447 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
4451 if (hash_lock
!= NULL
)
4452 mutex_exit(hash_lock
);
4455 * At this point, we have a level 1 cache miss. Try again in
4456 * L2ARC if possible.
4458 ASSERT3U(hdr
->b_size
, ==, size
);
4459 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
4460 uint64_t, size
, zbookmark_phys_t
*, zb
);
4461 ARCSTAT_BUMP(arcstat_misses
);
4462 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
4463 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
4464 data
, metadata
, misses
);
4466 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
4468 * Read from the L2ARC if the following are true:
4469 * 1. The L2ARC vdev was previously cached.
4470 * 2. This buffer still has L2ARC metadata.
4471 * 3. This buffer isn't currently writing to the L2ARC.
4472 * 4. The L2ARC entry wasn't evicted, which may
4473 * also have invalidated the vdev.
4474 * 5. This isn't prefetch and l2arc_noprefetch is set.
4476 if (HDR_HAS_L2HDR(hdr
) &&
4477 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
4478 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
4479 l2arc_read_callback_t
*cb
;
4481 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
4482 ARCSTAT_BUMP(arcstat_l2_hits
);
4483 atomic_inc_32(&hdr
->b_l2hdr
.b_hits
);
4485 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
4487 cb
->l2rcb_buf
= buf
;
4488 cb
->l2rcb_spa
= spa
;
4491 cb
->l2rcb_flags
= zio_flags
;
4492 cb
->l2rcb_compress
= b_compress
;
4494 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
4495 addr
+ size
< vd
->vdev_psize
-
4496 VDEV_LABEL_END_SIZE
);
4499 * l2arc read. The SCL_L2ARC lock will be
4500 * released by l2arc_read_done().
4501 * Issue a null zio if the underlying buffer
4502 * was squashed to zero size by compression.
4504 if (b_compress
== ZIO_COMPRESS_EMPTY
) {
4505 rzio
= zio_null(pio
, spa
, vd
,
4506 l2arc_read_done
, cb
,
4507 zio_flags
| ZIO_FLAG_DONT_CACHE
|
4509 ZIO_FLAG_DONT_PROPAGATE
|
4510 ZIO_FLAG_DONT_RETRY
);
4512 rzio
= zio_read_phys(pio
, vd
, addr
,
4513 b_asize
, buf
->b_data
,
4515 l2arc_read_done
, cb
, priority
,
4516 zio_flags
| ZIO_FLAG_DONT_CACHE
|
4518 ZIO_FLAG_DONT_PROPAGATE
|
4519 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
4521 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
4523 ARCSTAT_INCR(arcstat_l2_read_bytes
, b_asize
);
4525 if (*arc_flags
& ARC_FLAG_NOWAIT
) {
4530 ASSERT(*arc_flags
& ARC_FLAG_WAIT
);
4531 if (zio_wait(rzio
) == 0)
4534 /* l2arc read error; goto zio_read() */
4536 DTRACE_PROBE1(l2arc__miss
,
4537 arc_buf_hdr_t
*, hdr
);
4538 ARCSTAT_BUMP(arcstat_l2_misses
);
4539 if (HDR_L2_WRITING(hdr
))
4540 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
4541 spa_config_exit(spa
, SCL_L2ARC
, vd
);
4545 spa_config_exit(spa
, SCL_L2ARC
, vd
);
4546 if (l2arc_ndev
!= 0) {
4547 DTRACE_PROBE1(l2arc__miss
,
4548 arc_buf_hdr_t
*, hdr
);
4549 ARCSTAT_BUMP(arcstat_l2_misses
);
4553 rzio
= zio_read(pio
, spa
, bp
, buf
->b_data
, size
,
4554 arc_read_done
, buf
, priority
, zio_flags
, zb
);
4556 if (*arc_flags
& ARC_FLAG_WAIT
) {
4557 rc
= zio_wait(rzio
);
4561 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
4566 spa_read_history_add(spa
, zb
, *arc_flags
);
4571 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
4575 p
= kmem_alloc(sizeof (*p
), KM_SLEEP
);
4577 p
->p_private
= private;
4578 list_link_init(&p
->p_node
);
4579 refcount_create(&p
->p_refcnt
);
4581 mutex_enter(&arc_prune_mtx
);
4582 refcount_add(&p
->p_refcnt
, &arc_prune_list
);
4583 list_insert_head(&arc_prune_list
, p
);
4584 mutex_exit(&arc_prune_mtx
);
4590 arc_remove_prune_callback(arc_prune_t
*p
)
4592 mutex_enter(&arc_prune_mtx
);
4593 list_remove(&arc_prune_list
, p
);
4594 if (refcount_remove(&p
->p_refcnt
, &arc_prune_list
) == 0) {
4595 refcount_destroy(&p
->p_refcnt
);
4596 kmem_free(p
, sizeof (*p
));
4598 mutex_exit(&arc_prune_mtx
);
4602 arc_set_callback(arc_buf_t
*buf
, arc_evict_func_t
*func
, void *private)
4604 ASSERT(buf
->b_hdr
!= NULL
);
4605 ASSERT(buf
->b_hdr
->b_l1hdr
.b_state
!= arc_anon
);
4606 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
) ||
4608 ASSERT(buf
->b_efunc
== NULL
);
4609 ASSERT(!HDR_BUF_AVAILABLE(buf
->b_hdr
));
4611 buf
->b_efunc
= func
;
4612 buf
->b_private
= private;
4616 * Notify the arc that a block was freed, and thus will never be used again.
4619 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
4622 kmutex_t
*hash_lock
;
4623 uint64_t guid
= spa_load_guid(spa
);
4625 ASSERT(!BP_IS_EMBEDDED(bp
));
4627 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
4630 if (HDR_BUF_AVAILABLE(hdr
)) {
4631 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
4632 add_reference(hdr
, hash_lock
, FTAG
);
4633 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
4634 mutex_exit(hash_lock
);
4636 arc_release(buf
, FTAG
);
4637 (void) arc_buf_remove_ref(buf
, FTAG
);
4639 mutex_exit(hash_lock
);
4645 * Clear the user eviction callback set by arc_set_callback(), first calling
4646 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4647 * clearing the callback may result in the arc_buf being destroyed. However,
4648 * it will not result in the *last* arc_buf being destroyed, hence the data
4649 * will remain cached in the ARC. We make a copy of the arc buffer here so
4650 * that we can process the callback without holding any locks.
4652 * It's possible that the callback is already in the process of being cleared
4653 * by another thread. In this case we can not clear the callback.
4655 * Returns B_TRUE if the callback was successfully called and cleared.
4658 arc_clear_callback(arc_buf_t
*buf
)
4661 kmutex_t
*hash_lock
;
4662 arc_evict_func_t
*efunc
= buf
->b_efunc
;
4663 void *private = buf
->b_private
;
4665 mutex_enter(&buf
->b_evict_lock
);
4669 * We are in arc_do_user_evicts().
4671 ASSERT(buf
->b_data
== NULL
);
4672 mutex_exit(&buf
->b_evict_lock
);
4674 } else if (buf
->b_data
== NULL
) {
4676 * We are on the eviction list; process this buffer now
4677 * but let arc_do_user_evicts() do the reaping.
4679 buf
->b_efunc
= NULL
;
4680 mutex_exit(&buf
->b_evict_lock
);
4681 VERIFY0(efunc(private));
4684 hash_lock
= HDR_LOCK(hdr
);
4685 mutex_enter(hash_lock
);
4687 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
4689 ASSERT3U(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), <,
4690 hdr
->b_l1hdr
.b_datacnt
);
4691 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
4692 hdr
->b_l1hdr
.b_state
== arc_mfu
);
4694 buf
->b_efunc
= NULL
;
4695 buf
->b_private
= NULL
;
4697 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
4698 mutex_exit(&buf
->b_evict_lock
);
4699 arc_buf_destroy(buf
, TRUE
);
4701 ASSERT(buf
== hdr
->b_l1hdr
.b_buf
);
4702 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
4703 mutex_exit(&buf
->b_evict_lock
);
4706 mutex_exit(hash_lock
);
4707 VERIFY0(efunc(private));
4712 * Release this buffer from the cache, making it an anonymous buffer. This
4713 * must be done after a read and prior to modifying the buffer contents.
4714 * If the buffer has more than one reference, we must make
4715 * a new hdr for the buffer.
4718 arc_release(arc_buf_t
*buf
, void *tag
)
4720 kmutex_t
*hash_lock
;
4722 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4725 * It would be nice to assert that if its DMU metadata (level >
4726 * 0 || it's the dnode file), then it must be syncing context.
4727 * But we don't know that information at this level.
4730 mutex_enter(&buf
->b_evict_lock
);
4732 ASSERT(HDR_HAS_L1HDR(hdr
));
4735 * We don't grab the hash lock prior to this check, because if
4736 * the buffer's header is in the arc_anon state, it won't be
4737 * linked into the hash table.
4739 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
4740 mutex_exit(&buf
->b_evict_lock
);
4741 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4742 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
4743 ASSERT(!HDR_HAS_L2HDR(hdr
));
4744 ASSERT(BUF_EMPTY(hdr
));
4746 ASSERT3U(hdr
->b_l1hdr
.b_datacnt
, ==, 1);
4747 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), ==, 1);
4748 ASSERT(!list_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4750 ASSERT3P(buf
->b_efunc
, ==, NULL
);
4751 ASSERT3P(buf
->b_private
, ==, NULL
);
4753 hdr
->b_l1hdr
.b_arc_access
= 0;
4759 hash_lock
= HDR_LOCK(hdr
);
4760 mutex_enter(hash_lock
);
4763 * This assignment is only valid as long as the hash_lock is
4764 * held, we must be careful not to reference state or the
4765 * b_state field after dropping the lock.
4767 state
= hdr
->b_l1hdr
.b_state
;
4768 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
4769 ASSERT3P(state
, !=, arc_anon
);
4771 /* this buffer is not on any list */
4772 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0);
4774 if (HDR_HAS_L2HDR(hdr
)) {
4775 mutex_enter(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
4778 * We have to recheck this conditional again now that
4779 * we're holding the l2ad_mtx to prevent a race with
4780 * another thread which might be concurrently calling
4781 * l2arc_evict(). In that case, l2arc_evict() might have
4782 * destroyed the header's L2 portion as we were waiting
4783 * to acquire the l2ad_mtx.
4785 if (HDR_HAS_L2HDR(hdr
))
4786 arc_hdr_l2hdr_destroy(hdr
);
4788 mutex_exit(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
4792 * Do we have more than one buf?
4794 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
4795 arc_buf_hdr_t
*nhdr
;
4797 uint64_t blksz
= hdr
->b_size
;
4798 uint64_t spa
= hdr
->b_spa
;
4799 arc_buf_contents_t type
= arc_buf_type(hdr
);
4800 uint32_t flags
= hdr
->b_flags
;
4802 ASSERT(hdr
->b_l1hdr
.b_buf
!= buf
|| buf
->b_next
!= NULL
);
4804 * Pull the data off of this hdr and attach it to
4805 * a new anonymous hdr.
4807 (void) remove_reference(hdr
, hash_lock
, tag
);
4808 bufp
= &hdr
->b_l1hdr
.b_buf
;
4809 while (*bufp
!= buf
)
4810 bufp
= &(*bufp
)->b_next
;
4811 *bufp
= buf
->b_next
;
4814 ASSERT3P(state
, !=, arc_l2c_only
);
4816 (void) refcount_remove_many(
4817 &state
->arcs_size
, hdr
->b_size
, buf
);
4819 if (refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
)) {
4822 ASSERT3P(state
, !=, arc_l2c_only
);
4823 size
= &state
->arcs_lsize
[type
];
4824 ASSERT3U(*size
, >=, hdr
->b_size
);
4825 atomic_add_64(size
, -hdr
->b_size
);
4829 * We're releasing a duplicate user data buffer, update
4830 * our statistics accordingly.
4832 if (HDR_ISTYPE_DATA(hdr
)) {
4833 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
4834 ARCSTAT_INCR(arcstat_duplicate_buffers_size
,
4837 hdr
->b_l1hdr
.b_datacnt
-= 1;
4838 arc_cksum_verify(buf
);
4839 arc_buf_unwatch(buf
);
4841 mutex_exit(hash_lock
);
4843 nhdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
4844 nhdr
->b_size
= blksz
;
4847 nhdr
->b_l1hdr
.b_mru_hits
= 0;
4848 nhdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
4849 nhdr
->b_l1hdr
.b_mfu_hits
= 0;
4850 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
4851 nhdr
->b_l1hdr
.b_l2_hits
= 0;
4852 nhdr
->b_flags
= flags
& ARC_FLAG_L2_WRITING
;
4853 nhdr
->b_flags
|= arc_bufc_to_flags(type
);
4854 nhdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
4856 nhdr
->b_l1hdr
.b_buf
= buf
;
4857 nhdr
->b_l1hdr
.b_datacnt
= 1;
4858 nhdr
->b_l1hdr
.b_state
= arc_anon
;
4859 nhdr
->b_l1hdr
.b_arc_access
= 0;
4860 nhdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
4861 nhdr
->b_freeze_cksum
= NULL
;
4863 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, tag
);
4865 mutex_exit(&buf
->b_evict_lock
);
4866 (void) refcount_add_many(&arc_anon
->arcs_size
, blksz
, buf
);
4868 mutex_exit(&buf
->b_evict_lock
);
4869 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 1);
4870 /* protected by hash lock, or hdr is on arc_anon */
4871 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4872 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4873 hdr
->b_l1hdr
.b_mru_hits
= 0;
4874 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
4875 hdr
->b_l1hdr
.b_mfu_hits
= 0;
4876 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
4877 hdr
->b_l1hdr
.b_l2_hits
= 0;
4878 arc_change_state(arc_anon
, hdr
, hash_lock
);
4879 hdr
->b_l1hdr
.b_arc_access
= 0;
4880 mutex_exit(hash_lock
);
4882 buf_discard_identity(hdr
);
4885 buf
->b_efunc
= NULL
;
4886 buf
->b_private
= NULL
;
4890 arc_released(arc_buf_t
*buf
)
4894 mutex_enter(&buf
->b_evict_lock
);
4895 released
= (buf
->b_data
!= NULL
&&
4896 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
4897 mutex_exit(&buf
->b_evict_lock
);
4903 arc_referenced(arc_buf_t
*buf
)
4907 mutex_enter(&buf
->b_evict_lock
);
4908 referenced
= (refcount_count(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
4909 mutex_exit(&buf
->b_evict_lock
);
4910 return (referenced
);
4915 arc_write_ready(zio_t
*zio
)
4917 arc_write_callback_t
*callback
= zio
->io_private
;
4918 arc_buf_t
*buf
= callback
->awcb_buf
;
4919 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4921 ASSERT(HDR_HAS_L1HDR(hdr
));
4922 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
4923 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
4924 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
4927 * If the IO is already in progress, then this is a re-write
4928 * attempt, so we need to thaw and re-compute the cksum.
4929 * It is the responsibility of the callback to handle the
4930 * accounting for any re-write attempt.
4932 if (HDR_IO_IN_PROGRESS(hdr
)) {
4933 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
4934 if (hdr
->b_freeze_cksum
!= NULL
) {
4935 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
4936 hdr
->b_freeze_cksum
= NULL
;
4938 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
4940 arc_cksum_compute(buf
, B_FALSE
);
4941 hdr
->b_flags
|= ARC_FLAG_IO_IN_PROGRESS
;
4945 * The SPA calls this callback for each physical write that happens on behalf
4946 * of a logical write. See the comment in dbuf_write_physdone() for details.
4949 arc_write_physdone(zio_t
*zio
)
4951 arc_write_callback_t
*cb
= zio
->io_private
;
4952 if (cb
->awcb_physdone
!= NULL
)
4953 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
4957 arc_write_done(zio_t
*zio
)
4959 arc_write_callback_t
*callback
= zio
->io_private
;
4960 arc_buf_t
*buf
= callback
->awcb_buf
;
4961 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4963 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
4965 if (zio
->io_error
== 0) {
4966 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
4967 buf_discard_identity(hdr
);
4969 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
4970 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
4973 ASSERT(BUF_EMPTY(hdr
));
4977 * If the block to be written was all-zero or compressed enough to be
4978 * embedded in the BP, no write was performed so there will be no
4979 * dva/birth/checksum. The buffer must therefore remain anonymous
4982 if (!BUF_EMPTY(hdr
)) {
4983 arc_buf_hdr_t
*exists
;
4984 kmutex_t
*hash_lock
;
4986 ASSERT(zio
->io_error
== 0);
4988 arc_cksum_verify(buf
);
4990 exists
= buf_hash_insert(hdr
, &hash_lock
);
4991 if (exists
!= NULL
) {
4993 * This can only happen if we overwrite for
4994 * sync-to-convergence, because we remove
4995 * buffers from the hash table when we arc_free().
4997 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
4998 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
4999 panic("bad overwrite, hdr=%p exists=%p",
5000 (void *)hdr
, (void *)exists
);
5001 ASSERT(refcount_is_zero(
5002 &exists
->b_l1hdr
.b_refcnt
));
5003 arc_change_state(arc_anon
, exists
, hash_lock
);
5004 mutex_exit(hash_lock
);
5005 arc_hdr_destroy(exists
);
5006 exists
= buf_hash_insert(hdr
, &hash_lock
);
5007 ASSERT3P(exists
, ==, NULL
);
5008 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
5010 ASSERT(zio
->io_prop
.zp_nopwrite
);
5011 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
5012 panic("bad nopwrite, hdr=%p exists=%p",
5013 (void *)hdr
, (void *)exists
);
5016 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
5017 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
5018 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
5019 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
5022 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
5023 /* if it's not anon, we are doing a scrub */
5024 if (exists
== NULL
&& hdr
->b_l1hdr
.b_state
== arc_anon
)
5025 arc_access(hdr
, hash_lock
);
5026 mutex_exit(hash_lock
);
5028 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
5031 ASSERT(!refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5032 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
5034 kmem_free(callback
, sizeof (arc_write_callback_t
));
5038 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
5039 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
, boolean_t l2arc_compress
,
5040 const zio_prop_t
*zp
, arc_done_func_t
*ready
, arc_done_func_t
*physdone
,
5041 arc_done_func_t
*done
, void *private, zio_priority_t priority
,
5042 int zio_flags
, const zbookmark_phys_t
*zb
)
5044 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5045 arc_write_callback_t
*callback
;
5048 ASSERT(ready
!= NULL
);
5049 ASSERT(done
!= NULL
);
5050 ASSERT(!HDR_IO_ERROR(hdr
));
5051 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5052 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
5053 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
5055 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
5057 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
5058 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
5059 callback
->awcb_ready
= ready
;
5060 callback
->awcb_physdone
= physdone
;
5061 callback
->awcb_done
= done
;
5062 callback
->awcb_private
= private;
5063 callback
->awcb_buf
= buf
;
5065 zio
= zio_write(pio
, spa
, txg
, bp
, buf
->b_data
, hdr
->b_size
, zp
,
5066 arc_write_ready
, arc_write_physdone
, arc_write_done
, callback
,
5067 priority
, zio_flags
, zb
);
5073 arc_memory_throttle(uint64_t reserve
, uint64_t txg
)
5076 if (zfs_arc_memory_throttle_disable
)
5079 if (freemem
> physmem
* arc_lotsfree_percent
/ 100)
5082 if (arc_reclaim_needed()) {
5083 /* memory is low, delay before restarting */
5084 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
5085 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
5086 return (SET_ERROR(EAGAIN
));
5093 arc_tempreserve_clear(uint64_t reserve
)
5095 atomic_add_64(&arc_tempreserve
, -reserve
);
5096 ASSERT((int64_t)arc_tempreserve
>= 0);
5100 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
5105 if (reserve
> arc_c
/4 && !arc_no_grow
)
5106 arc_c
= MIN(arc_c_max
, reserve
* 4);
5109 * Throttle when the calculated memory footprint for the TXG
5110 * exceeds the target ARC size.
5112 if (reserve
> arc_c
) {
5113 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
5114 return (SET_ERROR(ERESTART
));
5118 * Don't count loaned bufs as in flight dirty data to prevent long
5119 * network delays from blocking transactions that are ready to be
5120 * assigned to a txg.
5122 anon_size
= MAX((int64_t)(refcount_count(&arc_anon
->arcs_size
) -
5123 arc_loaned_bytes
), 0);
5126 * Writes will, almost always, require additional memory allocations
5127 * in order to compress/encrypt/etc the data. We therefore need to
5128 * make sure that there is sufficient available memory for this.
5130 error
= arc_memory_throttle(reserve
, txg
);
5135 * Throttle writes when the amount of dirty data in the cache
5136 * gets too large. We try to keep the cache less than half full
5137 * of dirty blocks so that our sync times don't grow too large.
5138 * Note: if two requests come in concurrently, we might let them
5139 * both succeed, when one of them should fail. Not a huge deal.
5142 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
5143 anon_size
> arc_c
/ 4) {
5144 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5145 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5146 arc_tempreserve
>>10,
5147 arc_anon
->arcs_lsize
[ARC_BUFC_METADATA
]>>10,
5148 arc_anon
->arcs_lsize
[ARC_BUFC_DATA
]>>10,
5149 reserve
>>10, arc_c
>>10);
5150 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
5151 return (SET_ERROR(ERESTART
));
5153 atomic_add_64(&arc_tempreserve
, reserve
);
5158 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
5159 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
5161 size
->value
.ui64
= refcount_count(&state
->arcs_size
);
5162 evict_data
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_DATA
];
5163 evict_metadata
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_METADATA
];
5167 arc_kstat_update(kstat_t
*ksp
, int rw
)
5169 arc_stats_t
*as
= ksp
->ks_data
;
5171 if (rw
== KSTAT_WRITE
) {
5174 arc_kstat_update_state(arc_anon
,
5175 &as
->arcstat_anon_size
,
5176 &as
->arcstat_anon_evictable_data
,
5177 &as
->arcstat_anon_evictable_metadata
);
5178 arc_kstat_update_state(arc_mru
,
5179 &as
->arcstat_mru_size
,
5180 &as
->arcstat_mru_evictable_data
,
5181 &as
->arcstat_mru_evictable_metadata
);
5182 arc_kstat_update_state(arc_mru_ghost
,
5183 &as
->arcstat_mru_ghost_size
,
5184 &as
->arcstat_mru_ghost_evictable_data
,
5185 &as
->arcstat_mru_ghost_evictable_metadata
);
5186 arc_kstat_update_state(arc_mfu
,
5187 &as
->arcstat_mfu_size
,
5188 &as
->arcstat_mfu_evictable_data
,
5189 &as
->arcstat_mfu_evictable_metadata
);
5190 arc_kstat_update_state(arc_mfu_ghost
,
5191 &as
->arcstat_mfu_ghost_size
,
5192 &as
->arcstat_mfu_ghost_evictable_data
,
5193 &as
->arcstat_mfu_ghost_evictable_metadata
);
5200 * This function *must* return indices evenly distributed between all
5201 * sublists of the multilist. This is needed due to how the ARC eviction
5202 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5203 * distributed between all sublists and uses this assumption when
5204 * deciding which sublist to evict from and how much to evict from it.
5207 arc_state_multilist_index_func(multilist_t
*ml
, void *obj
)
5209 arc_buf_hdr_t
*hdr
= obj
;
5212 * We rely on b_dva to generate evenly distributed index
5213 * numbers using buf_hash below. So, as an added precaution,
5214 * let's make sure we never add empty buffers to the arc lists.
5216 ASSERT(!BUF_EMPTY(hdr
));
5219 * The assumption here, is the hash value for a given
5220 * arc_buf_hdr_t will remain constant throughout its lifetime
5221 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
5222 * Thus, we don't need to store the header's sublist index
5223 * on insertion, as this index can be recalculated on removal.
5225 * Also, the low order bits of the hash value are thought to be
5226 * distributed evenly. Otherwise, in the case that the multilist
5227 * has a power of two number of sublists, each sublists' usage
5228 * would not be evenly distributed.
5230 return (buf_hash(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
) %
5231 multilist_get_num_sublists(ml
));
5235 * Called during module initialization and periodically thereafter to
5236 * apply reasonable changes to the exposed performance tunings. Non-zero
5237 * zfs_* values which differ from the currently set values will be applied.
5240 arc_tuning_update(void)
5242 /* Valid range: 64M - <all physical memory> */
5243 if ((zfs_arc_max
) && (zfs_arc_max
!= arc_c_max
) &&
5244 (zfs_arc_max
> 64 << 20) && (zfs_arc_max
< ptob(physmem
)) &&
5245 (zfs_arc_max
> arc_c_min
)) {
5246 arc_c_max
= zfs_arc_max
;
5248 arc_p
= (arc_c
>> 1);
5249 arc_meta_limit
= MIN(arc_meta_limit
, arc_c_max
);
5252 /* Valid range: 32M - <arc_c_max> */
5253 if ((zfs_arc_min
) && (zfs_arc_min
!= arc_c_min
) &&
5254 (zfs_arc_min
>= 2ULL << SPA_MAXBLOCKSHIFT
) &&
5255 (zfs_arc_min
<= arc_c_max
)) {
5256 arc_c_min
= zfs_arc_min
;
5257 arc_c
= MAX(arc_c
, arc_c_min
);
5260 /* Valid range: 16M - <arc_c_max> */
5261 if ((zfs_arc_meta_min
) && (zfs_arc_meta_min
!= arc_meta_min
) &&
5262 (zfs_arc_meta_min
>= 1ULL << SPA_MAXBLOCKSHIFT
) &&
5263 (zfs_arc_meta_min
<= arc_c_max
)) {
5264 arc_meta_min
= zfs_arc_meta_min
;
5265 arc_meta_limit
= MAX(arc_meta_limit
, arc_meta_min
);
5268 /* Valid range: <arc_meta_min> - <arc_c_max> */
5269 if ((zfs_arc_meta_limit
) && (zfs_arc_meta_limit
!= arc_meta_limit
) &&
5270 (zfs_arc_meta_limit
>= zfs_arc_meta_min
) &&
5271 (zfs_arc_meta_limit
<= arc_c_max
))
5272 arc_meta_limit
= zfs_arc_meta_limit
;
5274 /* Valid range: 1 - N */
5275 if (zfs_arc_grow_retry
)
5276 arc_grow_retry
= zfs_arc_grow_retry
;
5278 /* Valid range: 1 - N */
5279 if (zfs_arc_shrink_shift
) {
5280 arc_shrink_shift
= zfs_arc_shrink_shift
;
5281 arc_no_grow_shift
= MIN(arc_no_grow_shift
, arc_shrink_shift
-1);
5284 /* Valid range: 1 - N */
5285 if (zfs_arc_p_min_shift
)
5286 arc_p_min_shift
= zfs_arc_p_min_shift
;
5288 /* Valid range: 1 - N ticks */
5289 if (zfs_arc_min_prefetch_lifespan
)
5290 arc_min_prefetch_lifespan
= zfs_arc_min_prefetch_lifespan
;
5297 * allmem is "all memory that we could possibly use".
5300 uint64_t allmem
= ptob(physmem
);
5302 uint64_t allmem
= (physmem
* PAGESIZE
) / 2;
5305 mutex_init(&arc_reclaim_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
5306 cv_init(&arc_reclaim_thread_cv
, NULL
, CV_DEFAULT
, NULL
);
5307 cv_init(&arc_reclaim_waiters_cv
, NULL
, CV_DEFAULT
, NULL
);
5309 mutex_init(&arc_user_evicts_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
5310 cv_init(&arc_user_evicts_cv
, NULL
, CV_DEFAULT
, NULL
);
5312 /* Convert seconds to clock ticks */
5313 arc_min_prefetch_lifespan
= 1 * hz
;
5315 /* Start out with 1/8 of all memory */
5320 * On architectures where the physical memory can be larger
5321 * than the addressable space (intel in 32-bit mode), we may
5322 * need to limit the cache to 1/8 of VM size.
5324 arc_c
= MIN(arc_c
, vmem_size(heap_arena
, VMEM_ALLOC
| VMEM_FREE
) / 8);
5327 * Register a shrinker to support synchronous (direct) memory
5328 * reclaim from the arc. This is done to prevent kswapd from
5329 * swapping out pages when it is preferable to shrink the arc.
5331 spl_register_shrinker(&arc_shrinker
);
5334 /* Set min cache to allow safe operation of arc_adapt() */
5335 arc_c_min
= 2ULL << SPA_MAXBLOCKSHIFT
;
5336 /* Set max to 1/2 of all memory */
5337 arc_c_max
= allmem
/ 2;
5340 arc_p
= (arc_c
>> 1);
5342 /* Set min to 1/2 of arc_c_min */
5343 arc_meta_min
= 1ULL << SPA_MAXBLOCKSHIFT
;
5344 /* Initialize maximum observed usage to zero */
5346 /* Set limit to 3/4 of arc_c_max with a floor of arc_meta_min */
5347 arc_meta_limit
= MAX((3 * arc_c_max
) / 4, arc_meta_min
);
5349 /* Apply user specified tunings */
5350 arc_tuning_update();
5352 if (zfs_arc_num_sublists_per_state
< 1)
5353 zfs_arc_num_sublists_per_state
= MAX(boot_ncpus
, 1);
5355 /* if kmem_flags are set, lets try to use less memory */
5356 if (kmem_debugging())
5358 if (arc_c
< arc_c_min
)
5361 arc_anon
= &ARC_anon
;
5363 arc_mru_ghost
= &ARC_mru_ghost
;
5365 arc_mfu_ghost
= &ARC_mfu_ghost
;
5366 arc_l2c_only
= &ARC_l2c_only
;
5369 multilist_create(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
],
5370 sizeof (arc_buf_hdr_t
),
5371 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5372 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5373 multilist_create(&arc_mru
->arcs_list
[ARC_BUFC_DATA
],
5374 sizeof (arc_buf_hdr_t
),
5375 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5376 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5377 multilist_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
],
5378 sizeof (arc_buf_hdr_t
),
5379 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5380 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5381 multilist_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
],
5382 sizeof (arc_buf_hdr_t
),
5383 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5384 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5385 multilist_create(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
],
5386 sizeof (arc_buf_hdr_t
),
5387 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5388 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5389 multilist_create(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
],
5390 sizeof (arc_buf_hdr_t
),
5391 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5392 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5393 multilist_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
],
5394 sizeof (arc_buf_hdr_t
),
5395 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5396 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5397 multilist_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
],
5398 sizeof (arc_buf_hdr_t
),
5399 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5400 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5401 multilist_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
],
5402 sizeof (arc_buf_hdr_t
),
5403 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5404 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5405 multilist_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
],
5406 sizeof (arc_buf_hdr_t
),
5407 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5408 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5410 arc_anon
->arcs_state
= ARC_STATE_ANON
;
5411 arc_mru
->arcs_state
= ARC_STATE_MRU
;
5412 arc_mru_ghost
->arcs_state
= ARC_STATE_MRU_GHOST
;
5413 arc_mfu
->arcs_state
= ARC_STATE_MFU
;
5414 arc_mfu_ghost
->arcs_state
= ARC_STATE_MFU_GHOST
;
5415 arc_l2c_only
->arcs_state
= ARC_STATE_L2C_ONLY
;
5417 refcount_create(&arc_anon
->arcs_size
);
5418 refcount_create(&arc_mru
->arcs_size
);
5419 refcount_create(&arc_mru_ghost
->arcs_size
);
5420 refcount_create(&arc_mfu
->arcs_size
);
5421 refcount_create(&arc_mfu_ghost
->arcs_size
);
5422 refcount_create(&arc_l2c_only
->arcs_size
);
5426 arc_reclaim_thread_exit
= FALSE
;
5427 arc_user_evicts_thread_exit
= FALSE
;
5428 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
5429 offsetof(arc_prune_t
, p_node
));
5430 arc_eviction_list
= NULL
;
5431 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
5432 bzero(&arc_eviction_hdr
, sizeof (arc_buf_hdr_t
));
5434 arc_prune_taskq
= taskq_create("arc_prune", max_ncpus
, minclsyspri
,
5435 max_ncpus
, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
5437 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
5438 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
5440 if (arc_ksp
!= NULL
) {
5441 arc_ksp
->ks_data
= &arc_stats
;
5442 arc_ksp
->ks_update
= arc_kstat_update
;
5443 kstat_install(arc_ksp
);
5446 (void) thread_create(NULL
, 0, arc_reclaim_thread
, NULL
, 0, &p0
,
5447 TS_RUN
, minclsyspri
);
5449 (void) thread_create(NULL
, 0, arc_user_evicts_thread
, NULL
, 0, &p0
,
5450 TS_RUN
, minclsyspri
);
5456 * Calculate maximum amount of dirty data per pool.
5458 * If it has been set by a module parameter, take that.
5459 * Otherwise, use a percentage of physical memory defined by
5460 * zfs_dirty_data_max_percent (default 10%) with a cap at
5461 * zfs_dirty_data_max_max (default 25% of physical memory).
5463 if (zfs_dirty_data_max_max
== 0)
5464 zfs_dirty_data_max_max
= physmem
* PAGESIZE
*
5465 zfs_dirty_data_max_max_percent
/ 100;
5467 if (zfs_dirty_data_max
== 0) {
5468 zfs_dirty_data_max
= physmem
* PAGESIZE
*
5469 zfs_dirty_data_max_percent
/ 100;
5470 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
5471 zfs_dirty_data_max_max
);
5481 spl_unregister_shrinker(&arc_shrinker
);
5482 #endif /* _KERNEL */
5484 mutex_enter(&arc_reclaim_lock
);
5485 arc_reclaim_thread_exit
= TRUE
;
5487 * The reclaim thread will set arc_reclaim_thread_exit back to
5488 * FALSE when it is finished exiting; we're waiting for that.
5490 while (arc_reclaim_thread_exit
) {
5491 cv_signal(&arc_reclaim_thread_cv
);
5492 cv_wait(&arc_reclaim_thread_cv
, &arc_reclaim_lock
);
5494 mutex_exit(&arc_reclaim_lock
);
5496 mutex_enter(&arc_user_evicts_lock
);
5497 arc_user_evicts_thread_exit
= TRUE
;
5499 * The user evicts thread will set arc_user_evicts_thread_exit
5500 * to FALSE when it is finished exiting; we're waiting for that.
5502 while (arc_user_evicts_thread_exit
) {
5503 cv_signal(&arc_user_evicts_cv
);
5504 cv_wait(&arc_user_evicts_cv
, &arc_user_evicts_lock
);
5506 mutex_exit(&arc_user_evicts_lock
);
5508 /* Use TRUE to ensure *all* buffers are evicted */
5509 arc_flush(NULL
, TRUE
);
5513 if (arc_ksp
!= NULL
) {
5514 kstat_delete(arc_ksp
);
5518 taskq_wait(arc_prune_taskq
);
5519 taskq_destroy(arc_prune_taskq
);
5521 mutex_enter(&arc_prune_mtx
);
5522 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
5523 list_remove(&arc_prune_list
, p
);
5524 refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
5525 refcount_destroy(&p
->p_refcnt
);
5526 kmem_free(p
, sizeof (*p
));
5528 mutex_exit(&arc_prune_mtx
);
5530 list_destroy(&arc_prune_list
);
5531 mutex_destroy(&arc_prune_mtx
);
5532 mutex_destroy(&arc_reclaim_lock
);
5533 cv_destroy(&arc_reclaim_thread_cv
);
5534 cv_destroy(&arc_reclaim_waiters_cv
);
5536 mutex_destroy(&arc_user_evicts_lock
);
5537 cv_destroy(&arc_user_evicts_cv
);
5539 refcount_destroy(&arc_anon
->arcs_size
);
5540 refcount_destroy(&arc_mru
->arcs_size
);
5541 refcount_destroy(&arc_mru_ghost
->arcs_size
);
5542 refcount_destroy(&arc_mfu
->arcs_size
);
5543 refcount_destroy(&arc_mfu_ghost
->arcs_size
);
5544 refcount_destroy(&arc_l2c_only
->arcs_size
);
5546 multilist_destroy(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
5547 multilist_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
5548 multilist_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
5549 multilist_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
5550 multilist_destroy(&arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
5551 multilist_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
5552 multilist_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
5553 multilist_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
5554 multilist_destroy(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]);
5555 multilist_destroy(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]);
5559 ASSERT0(arc_loaned_bytes
);
5565 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5566 * It uses dedicated storage devices to hold cached data, which are populated
5567 * using large infrequent writes. The main role of this cache is to boost
5568 * the performance of random read workloads. The intended L2ARC devices
5569 * include short-stroked disks, solid state disks, and other media with
5570 * substantially faster read latency than disk.
5572 * +-----------------------+
5574 * +-----------------------+
5577 * l2arc_feed_thread() arc_read()
5581 * +---------------+ |
5583 * +---------------+ |
5588 * +-------+ +-------+
5590 * | cache | | cache |
5591 * +-------+ +-------+
5592 * +=========+ .-----.
5593 * : L2ARC : |-_____-|
5594 * : devices : | Disks |
5595 * +=========+ `-_____-'
5597 * Read requests are satisfied from the following sources, in order:
5600 * 2) vdev cache of L2ARC devices
5602 * 4) vdev cache of disks
5605 * Some L2ARC device types exhibit extremely slow write performance.
5606 * To accommodate for this there are some significant differences between
5607 * the L2ARC and traditional cache design:
5609 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5610 * the ARC behave as usual, freeing buffers and placing headers on ghost
5611 * lists. The ARC does not send buffers to the L2ARC during eviction as
5612 * this would add inflated write latencies for all ARC memory pressure.
5614 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5615 * It does this by periodically scanning buffers from the eviction-end of
5616 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5617 * not already there. It scans until a headroom of buffers is satisfied,
5618 * which itself is a buffer for ARC eviction. If a compressible buffer is
5619 * found during scanning and selected for writing to an L2ARC device, we
5620 * temporarily boost scanning headroom during the next scan cycle to make
5621 * sure we adapt to compression effects (which might significantly reduce
5622 * the data volume we write to L2ARC). The thread that does this is
5623 * l2arc_feed_thread(), illustrated below; example sizes are included to
5624 * provide a better sense of ratio than this diagram:
5627 * +---------------------+----------+
5628 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5629 * +---------------------+----------+ | o L2ARC eligible
5630 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5631 * +---------------------+----------+ |
5632 * 15.9 Gbytes ^ 32 Mbytes |
5634 * l2arc_feed_thread()
5636 * l2arc write hand <--[oooo]--'
5640 * +==============================+
5641 * L2ARC dev |####|#|###|###| |####| ... |
5642 * +==============================+
5645 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5646 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5647 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5648 * safe to say that this is an uncommon case, since buffers at the end of
5649 * the ARC lists have moved there due to inactivity.
5651 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5652 * then the L2ARC simply misses copying some buffers. This serves as a
5653 * pressure valve to prevent heavy read workloads from both stalling the ARC
5654 * with waits and clogging the L2ARC with writes. This also helps prevent
5655 * the potential for the L2ARC to churn if it attempts to cache content too
5656 * quickly, such as during backups of the entire pool.
5658 * 5. After system boot and before the ARC has filled main memory, there are
5659 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5660 * lists can remain mostly static. Instead of searching from tail of these
5661 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5662 * for eligible buffers, greatly increasing its chance of finding them.
5664 * The L2ARC device write speed is also boosted during this time so that
5665 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5666 * there are no L2ARC reads, and no fear of degrading read performance
5667 * through increased writes.
5669 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5670 * the vdev queue can aggregate them into larger and fewer writes. Each
5671 * device is written to in a rotor fashion, sweeping writes through
5672 * available space then repeating.
5674 * 7. The L2ARC does not store dirty content. It never needs to flush
5675 * write buffers back to disk based storage.
5677 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5678 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5680 * The performance of the L2ARC can be tweaked by a number of tunables, which
5681 * may be necessary for different workloads:
5683 * l2arc_write_max max write bytes per interval
5684 * l2arc_write_boost extra write bytes during device warmup
5685 * l2arc_noprefetch skip caching prefetched buffers
5686 * l2arc_nocompress skip compressing buffers
5687 * l2arc_headroom number of max device writes to precache
5688 * l2arc_headroom_boost when we find compressed buffers during ARC
5689 * scanning, we multiply headroom by this
5690 * percentage factor for the next scan cycle,
5691 * since more compressed buffers are likely to
5693 * l2arc_feed_secs seconds between L2ARC writing
5695 * Tunables may be removed or added as future performance improvements are
5696 * integrated, and also may become zpool properties.
5698 * There are three key functions that control how the L2ARC warms up:
5700 * l2arc_write_eligible() check if a buffer is eligible to cache
5701 * l2arc_write_size() calculate how much to write
5702 * l2arc_write_interval() calculate sleep delay between writes
5704 * These three functions determine what to write, how much, and how quickly
5709 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*hdr
)
5712 * A buffer is *not* eligible for the L2ARC if it:
5713 * 1. belongs to a different spa.
5714 * 2. is already cached on the L2ARC.
5715 * 3. has an I/O in progress (it may be an incomplete read).
5716 * 4. is flagged not eligible (zfs property).
5718 if (hdr
->b_spa
!= spa_guid
|| HDR_HAS_L2HDR(hdr
) ||
5719 HDR_IO_IN_PROGRESS(hdr
) || !HDR_L2CACHE(hdr
))
5726 l2arc_write_size(void)
5731 * Make sure our globals have meaningful values in case the user
5734 size
= l2arc_write_max
;
5736 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
5737 "be greater than zero, resetting it to the default (%d)",
5739 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
5742 if (arc_warm
== B_FALSE
)
5743 size
+= l2arc_write_boost
;
5750 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
5752 clock_t interval
, next
, now
;
5755 * If the ARC lists are busy, increase our write rate; if the
5756 * lists are stale, idle back. This is achieved by checking
5757 * how much we previously wrote - if it was more than half of
5758 * what we wanted, schedule the next write much sooner.
5760 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
5761 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
5763 interval
= hz
* l2arc_feed_secs
;
5765 now
= ddi_get_lbolt();
5766 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
5772 * Cycle through L2ARC devices. This is how L2ARC load balances.
5773 * If a device is returned, this also returns holding the spa config lock.
5775 static l2arc_dev_t
*
5776 l2arc_dev_get_next(void)
5778 l2arc_dev_t
*first
, *next
= NULL
;
5781 * Lock out the removal of spas (spa_namespace_lock), then removal
5782 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
5783 * both locks will be dropped and a spa config lock held instead.
5785 mutex_enter(&spa_namespace_lock
);
5786 mutex_enter(&l2arc_dev_mtx
);
5788 /* if there are no vdevs, there is nothing to do */
5789 if (l2arc_ndev
== 0)
5793 next
= l2arc_dev_last
;
5795 /* loop around the list looking for a non-faulted vdev */
5797 next
= list_head(l2arc_dev_list
);
5799 next
= list_next(l2arc_dev_list
, next
);
5801 next
= list_head(l2arc_dev_list
);
5804 /* if we have come back to the start, bail out */
5807 else if (next
== first
)
5810 } while (vdev_is_dead(next
->l2ad_vdev
));
5812 /* if we were unable to find any usable vdevs, return NULL */
5813 if (vdev_is_dead(next
->l2ad_vdev
))
5816 l2arc_dev_last
= next
;
5819 mutex_exit(&l2arc_dev_mtx
);
5822 * Grab the config lock to prevent the 'next' device from being
5823 * removed while we are writing to it.
5826 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
5827 mutex_exit(&spa_namespace_lock
);
5833 * Free buffers that were tagged for destruction.
5836 l2arc_do_free_on_write(void)
5839 l2arc_data_free_t
*df
, *df_prev
;
5841 mutex_enter(&l2arc_free_on_write_mtx
);
5842 buflist
= l2arc_free_on_write
;
5844 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
5845 df_prev
= list_prev(buflist
, df
);
5846 ASSERT(df
->l2df_data
!= NULL
);
5847 ASSERT(df
->l2df_func
!= NULL
);
5848 df
->l2df_func(df
->l2df_data
, df
->l2df_size
);
5849 list_remove(buflist
, df
);
5850 kmem_free(df
, sizeof (l2arc_data_free_t
));
5853 mutex_exit(&l2arc_free_on_write_mtx
);
5857 * A write to a cache device has completed. Update all headers to allow
5858 * reads from these buffers to begin.
5861 l2arc_write_done(zio_t
*zio
)
5863 l2arc_write_callback_t
*cb
;
5866 arc_buf_hdr_t
*head
, *hdr
, *hdr_prev
;
5867 kmutex_t
*hash_lock
;
5868 int64_t bytes_dropped
= 0;
5870 cb
= zio
->io_private
;
5872 dev
= cb
->l2wcb_dev
;
5873 ASSERT(dev
!= NULL
);
5874 head
= cb
->l2wcb_head
;
5875 ASSERT(head
!= NULL
);
5876 buflist
= &dev
->l2ad_buflist
;
5877 ASSERT(buflist
!= NULL
);
5878 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
5879 l2arc_write_callback_t
*, cb
);
5881 if (zio
->io_error
!= 0)
5882 ARCSTAT_BUMP(arcstat_l2_writes_error
);
5885 * All writes completed, or an error was hit.
5888 mutex_enter(&dev
->l2ad_mtx
);
5889 for (hdr
= list_prev(buflist
, head
); hdr
; hdr
= hdr_prev
) {
5890 hdr_prev
= list_prev(buflist
, hdr
);
5892 hash_lock
= HDR_LOCK(hdr
);
5895 * We cannot use mutex_enter or else we can deadlock
5896 * with l2arc_write_buffers (due to swapping the order
5897 * the hash lock and l2ad_mtx are taken).
5899 if (!mutex_tryenter(hash_lock
)) {
5901 * Missed the hash lock. We must retry so we
5902 * don't leave the ARC_FLAG_L2_WRITING bit set.
5904 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry
);
5907 * We don't want to rescan the headers we've
5908 * already marked as having been written out, so
5909 * we reinsert the head node so we can pick up
5910 * where we left off.
5912 list_remove(buflist
, head
);
5913 list_insert_after(buflist
, hdr
, head
);
5915 mutex_exit(&dev
->l2ad_mtx
);
5918 * We wait for the hash lock to become available
5919 * to try and prevent busy waiting, and increase
5920 * the chance we'll be able to acquire the lock
5921 * the next time around.
5923 mutex_enter(hash_lock
);
5924 mutex_exit(hash_lock
);
5929 * We could not have been moved into the arc_l2c_only
5930 * state while in-flight due to our ARC_FLAG_L2_WRITING
5931 * bit being set. Let's just ensure that's being enforced.
5933 ASSERT(HDR_HAS_L1HDR(hdr
));
5936 * We may have allocated a buffer for L2ARC compression,
5937 * we must release it to avoid leaking this data.
5939 l2arc_release_cdata_buf(hdr
);
5941 if (zio
->io_error
!= 0) {
5943 * Error - drop L2ARC entry.
5945 list_remove(buflist
, hdr
);
5946 hdr
->b_flags
&= ~ARC_FLAG_HAS_L2HDR
;
5948 ARCSTAT_INCR(arcstat_l2_asize
, -hdr
->b_l2hdr
.b_asize
);
5949 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
5951 bytes_dropped
+= hdr
->b_l2hdr
.b_asize
;
5952 (void) refcount_remove_many(&dev
->l2ad_alloc
,
5953 hdr
->b_l2hdr
.b_asize
, hdr
);
5957 * Allow ARC to begin reads and ghost list evictions to
5960 hdr
->b_flags
&= ~ARC_FLAG_L2_WRITING
;
5962 mutex_exit(hash_lock
);
5965 atomic_inc_64(&l2arc_writes_done
);
5966 list_remove(buflist
, head
);
5967 ASSERT(!HDR_HAS_L1HDR(head
));
5968 kmem_cache_free(hdr_l2only_cache
, head
);
5969 mutex_exit(&dev
->l2ad_mtx
);
5971 vdev_space_update(dev
->l2ad_vdev
, -bytes_dropped
, 0, 0);
5973 l2arc_do_free_on_write();
5975 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
5979 * A read to a cache device completed. Validate buffer contents before
5980 * handing over to the regular ARC routines.
5983 l2arc_read_done(zio_t
*zio
)
5985 l2arc_read_callback_t
*cb
;
5988 kmutex_t
*hash_lock
;
5991 ASSERT(zio
->io_vd
!= NULL
);
5992 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
5994 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
5996 cb
= zio
->io_private
;
5998 buf
= cb
->l2rcb_buf
;
5999 ASSERT(buf
!= NULL
);
6001 hash_lock
= HDR_LOCK(buf
->b_hdr
);
6002 mutex_enter(hash_lock
);
6004 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
6007 * If the buffer was compressed, decompress it first.
6009 if (cb
->l2rcb_compress
!= ZIO_COMPRESS_OFF
)
6010 l2arc_decompress_zio(zio
, hdr
, cb
->l2rcb_compress
);
6011 ASSERT(zio
->io_data
!= NULL
);
6014 * Check this survived the L2ARC journey.
6016 equal
= arc_cksum_equal(buf
);
6017 if (equal
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
6018 mutex_exit(hash_lock
);
6019 zio
->io_private
= buf
;
6020 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
6021 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
6024 mutex_exit(hash_lock
);
6026 * Buffer didn't survive caching. Increment stats and
6027 * reissue to the original storage device.
6029 if (zio
->io_error
!= 0) {
6030 ARCSTAT_BUMP(arcstat_l2_io_error
);
6032 zio
->io_error
= SET_ERROR(EIO
);
6035 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
6038 * If there's no waiter, issue an async i/o to the primary
6039 * storage now. If there *is* a waiter, the caller must
6040 * issue the i/o in a context where it's OK to block.
6042 if (zio
->io_waiter
== NULL
) {
6043 zio_t
*pio
= zio_unique_parent(zio
);
6045 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
6047 zio_nowait(zio_read(pio
, cb
->l2rcb_spa
, &cb
->l2rcb_bp
,
6048 buf
->b_data
, zio
->io_size
, arc_read_done
, buf
,
6049 zio
->io_priority
, cb
->l2rcb_flags
, &cb
->l2rcb_zb
));
6053 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
6057 * This is the list priority from which the L2ARC will search for pages to
6058 * cache. This is used within loops (0..3) to cycle through lists in the
6059 * desired order. This order can have a significant effect on cache
6062 * Currently the metadata lists are hit first, MFU then MRU, followed by
6063 * the data lists. This function returns a locked list, and also returns
6066 static multilist_sublist_t
*
6067 l2arc_sublist_lock(int list_num
)
6069 multilist_t
*ml
= NULL
;
6072 ASSERT(list_num
>= 0 && list_num
<= 3);
6076 ml
= &arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
6079 ml
= &arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
6082 ml
= &arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
6085 ml
= &arc_mru
->arcs_list
[ARC_BUFC_DATA
];
6090 * Return a randomly-selected sublist. This is acceptable
6091 * because the caller feeds only a little bit of data for each
6092 * call (8MB). Subsequent calls will result in different
6093 * sublists being selected.
6095 idx
= multilist_get_random_index(ml
);
6096 return (multilist_sublist_lock(ml
, idx
));
6100 * Evict buffers from the device write hand to the distance specified in
6101 * bytes. This distance may span populated buffers, it may span nothing.
6102 * This is clearing a region on the L2ARC device ready for writing.
6103 * If the 'all' boolean is set, every buffer is evicted.
6106 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
6109 arc_buf_hdr_t
*hdr
, *hdr_prev
;
6110 kmutex_t
*hash_lock
;
6113 buflist
= &dev
->l2ad_buflist
;
6115 if (!all
&& dev
->l2ad_first
) {
6117 * This is the first sweep through the device. There is
6123 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
6125 * When nearing the end of the device, evict to the end
6126 * before the device write hand jumps to the start.
6128 taddr
= dev
->l2ad_end
;
6130 taddr
= dev
->l2ad_hand
+ distance
;
6132 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
6133 uint64_t, taddr
, boolean_t
, all
);
6136 mutex_enter(&dev
->l2ad_mtx
);
6137 for (hdr
= list_tail(buflist
); hdr
; hdr
= hdr_prev
) {
6138 hdr_prev
= list_prev(buflist
, hdr
);
6140 hash_lock
= HDR_LOCK(hdr
);
6143 * We cannot use mutex_enter or else we can deadlock
6144 * with l2arc_write_buffers (due to swapping the order
6145 * the hash lock and l2ad_mtx are taken).
6147 if (!mutex_tryenter(hash_lock
)) {
6149 * Missed the hash lock. Retry.
6151 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
6152 mutex_exit(&dev
->l2ad_mtx
);
6153 mutex_enter(hash_lock
);
6154 mutex_exit(hash_lock
);
6158 if (HDR_L2_WRITE_HEAD(hdr
)) {
6160 * We hit a write head node. Leave it for
6161 * l2arc_write_done().
6163 list_remove(buflist
, hdr
);
6164 mutex_exit(hash_lock
);
6168 if (!all
&& HDR_HAS_L2HDR(hdr
) &&
6169 (hdr
->b_l2hdr
.b_daddr
> taddr
||
6170 hdr
->b_l2hdr
.b_daddr
< dev
->l2ad_hand
)) {
6172 * We've evicted to the target address,
6173 * or the end of the device.
6175 mutex_exit(hash_lock
);
6179 ASSERT(HDR_HAS_L2HDR(hdr
));
6180 if (!HDR_HAS_L1HDR(hdr
)) {
6181 ASSERT(!HDR_L2_READING(hdr
));
6183 * This doesn't exist in the ARC. Destroy.
6184 * arc_hdr_destroy() will call list_remove()
6185 * and decrement arcstat_l2_size.
6187 arc_change_state(arc_anon
, hdr
, hash_lock
);
6188 arc_hdr_destroy(hdr
);
6190 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_l2c_only
);
6191 ARCSTAT_BUMP(arcstat_l2_evict_l1cached
);
6193 * Invalidate issued or about to be issued
6194 * reads, since we may be about to write
6195 * over this location.
6197 if (HDR_L2_READING(hdr
)) {
6198 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
6199 hdr
->b_flags
|= ARC_FLAG_L2_EVICTED
;
6202 /* Ensure this header has finished being written */
6203 ASSERT(!HDR_L2_WRITING(hdr
));
6204 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
6206 arc_hdr_l2hdr_destroy(hdr
);
6208 mutex_exit(hash_lock
);
6210 mutex_exit(&dev
->l2ad_mtx
);
6214 * Find and write ARC buffers to the L2ARC device.
6216 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6217 * for reading until they have completed writing.
6218 * The headroom_boost is an in-out parameter used to maintain headroom boost
6219 * state between calls to this function.
6221 * Returns the number of bytes actually written (which may be smaller than
6222 * the delta by which the device hand has changed due to alignment).
6225 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
,
6226 boolean_t
*headroom_boost
)
6228 arc_buf_hdr_t
*hdr
, *hdr_prev
, *head
;
6229 uint64_t write_asize
, write_sz
, headroom
, buf_compress_minsz
,
6233 l2arc_write_callback_t
*cb
;
6235 uint64_t guid
= spa_load_guid(spa
);
6237 const boolean_t do_headroom_boost
= *headroom_boost
;
6239 ASSERT(dev
->l2ad_vdev
!= NULL
);
6241 /* Lower the flag now, we might want to raise it again later. */
6242 *headroom_boost
= B_FALSE
;
6245 write_sz
= write_asize
= 0;
6247 head
= kmem_cache_alloc(hdr_l2only_cache
, KM_PUSHPAGE
);
6248 head
->b_flags
|= ARC_FLAG_L2_WRITE_HEAD
;
6249 head
->b_flags
|= ARC_FLAG_HAS_L2HDR
;
6252 * We will want to try to compress buffers that are at least 2x the
6253 * device sector size.
6255 buf_compress_minsz
= 2 << dev
->l2ad_vdev
->vdev_ashift
;
6258 * Copy buffers for L2ARC writing.
6260 for (try = 0; try <= 3; try++) {
6261 multilist_sublist_t
*mls
= l2arc_sublist_lock(try);
6262 uint64_t passed_sz
= 0;
6265 * L2ARC fast warmup.
6267 * Until the ARC is warm and starts to evict, read from the
6268 * head of the ARC lists rather than the tail.
6270 if (arc_warm
== B_FALSE
)
6271 hdr
= multilist_sublist_head(mls
);
6273 hdr
= multilist_sublist_tail(mls
);
6275 headroom
= target_sz
* l2arc_headroom
;
6276 if (do_headroom_boost
)
6277 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
6279 for (; hdr
; hdr
= hdr_prev
) {
6280 kmutex_t
*hash_lock
;
6284 if (arc_warm
== B_FALSE
)
6285 hdr_prev
= multilist_sublist_next(mls
, hdr
);
6287 hdr_prev
= multilist_sublist_prev(mls
, hdr
);
6289 hash_lock
= HDR_LOCK(hdr
);
6290 if (!mutex_tryenter(hash_lock
)) {
6292 * Skip this buffer rather than waiting.
6297 passed_sz
+= hdr
->b_size
;
6298 if (passed_sz
> headroom
) {
6302 mutex_exit(hash_lock
);
6306 if (!l2arc_write_eligible(guid
, hdr
)) {
6307 mutex_exit(hash_lock
);
6312 * Assume that the buffer is not going to be compressed
6313 * and could take more space on disk because of a larger
6316 buf_sz
= hdr
->b_size
;
6317 buf_a_sz
= vdev_psize_to_asize(dev
->l2ad_vdev
, buf_sz
);
6319 if ((write_asize
+ buf_a_sz
) > target_sz
) {
6321 mutex_exit(hash_lock
);
6327 * Insert a dummy header on the buflist so
6328 * l2arc_write_done() can find where the
6329 * write buffers begin without searching.
6331 mutex_enter(&dev
->l2ad_mtx
);
6332 list_insert_head(&dev
->l2ad_buflist
, head
);
6333 mutex_exit(&dev
->l2ad_mtx
);
6336 sizeof (l2arc_write_callback_t
), KM_SLEEP
);
6337 cb
->l2wcb_dev
= dev
;
6338 cb
->l2wcb_head
= head
;
6339 pio
= zio_root(spa
, l2arc_write_done
, cb
,
6344 * Create and add a new L2ARC header.
6346 hdr
->b_l2hdr
.b_dev
= dev
;
6347 hdr
->b_flags
|= ARC_FLAG_L2_WRITING
;
6349 * Temporarily stash the data buffer in b_tmp_cdata.
6350 * The subsequent write step will pick it up from
6351 * there. This is because can't access b_l1hdr.b_buf
6352 * without holding the hash_lock, which we in turn
6353 * can't access without holding the ARC list locks
6354 * (which we want to avoid during compression/writing)
6356 HDR_SET_COMPRESS(hdr
, ZIO_COMPRESS_OFF
);
6357 hdr
->b_l2hdr
.b_asize
= hdr
->b_size
;
6358 hdr
->b_l2hdr
.b_hits
= 0;
6359 hdr
->b_l1hdr
.b_tmp_cdata
= hdr
->b_l1hdr
.b_buf
->b_data
;
6362 * Explicitly set the b_daddr field to a known
6363 * value which means "invalid address". This
6364 * enables us to differentiate which stage of
6365 * l2arc_write_buffers() the particular header
6366 * is in (e.g. this loop, or the one below).
6367 * ARC_FLAG_L2_WRITING is not enough to make
6368 * this distinction, and we need to know in
6369 * order to do proper l2arc vdev accounting in
6370 * arc_release() and arc_hdr_destroy().
6372 * Note, we can't use a new flag to distinguish
6373 * the two stages because we don't hold the
6374 * header's hash_lock below, in the second stage
6375 * of this function. Thus, we can't simply
6376 * change the b_flags field to denote that the
6377 * IO has been sent. We can change the b_daddr
6378 * field of the L2 portion, though, since we'll
6379 * be holding the l2ad_mtx; which is why we're
6380 * using it to denote the header's state change.
6382 hdr
->b_l2hdr
.b_daddr
= L2ARC_ADDR_UNSET
;
6383 hdr
->b_flags
|= ARC_FLAG_HAS_L2HDR
;
6385 mutex_enter(&dev
->l2ad_mtx
);
6386 list_insert_head(&dev
->l2ad_buflist
, hdr
);
6387 mutex_exit(&dev
->l2ad_mtx
);
6390 * Compute and store the buffer cksum before
6391 * writing. On debug the cksum is verified first.
6393 arc_cksum_verify(hdr
->b_l1hdr
.b_buf
);
6394 arc_cksum_compute(hdr
->b_l1hdr
.b_buf
, B_TRUE
);
6396 mutex_exit(hash_lock
);
6399 write_asize
+= buf_a_sz
;
6402 multilist_sublist_unlock(mls
);
6408 /* No buffers selected for writing? */
6411 ASSERT(!HDR_HAS_L1HDR(head
));
6412 kmem_cache_free(hdr_l2only_cache
, head
);
6416 mutex_enter(&dev
->l2ad_mtx
);
6419 * Note that elsewhere in this file arcstat_l2_asize
6420 * and the used space on l2ad_vdev are updated using b_asize,
6421 * which is not necessarily rounded up to the device block size.
6422 * Too keep accounting consistent we do the same here as well:
6423 * stats_size accumulates the sum of b_asize of the written buffers,
6424 * while write_asize accumulates the sum of b_asize rounded up
6425 * to the device block size.
6426 * The latter sum is used only to validate the corectness of the code.
6432 * Now start writing the buffers. We're starting at the write head
6433 * and work backwards, retracing the course of the buffer selector
6436 for (hdr
= list_prev(&dev
->l2ad_buflist
, head
); hdr
;
6437 hdr
= list_prev(&dev
->l2ad_buflist
, hdr
)) {
6441 * We rely on the L1 portion of the header below, so
6442 * it's invalid for this header to have been evicted out
6443 * of the ghost cache, prior to being written out. The
6444 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6446 ASSERT(HDR_HAS_L1HDR(hdr
));
6449 * We shouldn't need to lock the buffer here, since we flagged
6450 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6451 * take care to only access its L2 cache parameters. In
6452 * particular, hdr->l1hdr.b_buf may be invalid by now due to
6455 hdr
->b_l2hdr
.b_daddr
= dev
->l2ad_hand
;
6457 if ((!l2arc_nocompress
&& HDR_L2COMPRESS(hdr
)) &&
6458 hdr
->b_l2hdr
.b_asize
>= buf_compress_minsz
) {
6459 if (l2arc_compress_buf(hdr
)) {
6461 * If compression succeeded, enable headroom
6462 * boost on the next scan cycle.
6464 *headroom_boost
= B_TRUE
;
6469 * Pick up the buffer data we had previously stashed away
6470 * (and now potentially also compressed).
6472 buf_data
= hdr
->b_l1hdr
.b_tmp_cdata
;
6473 buf_sz
= hdr
->b_l2hdr
.b_asize
;
6476 * We need to do this regardless if buf_sz is zero or
6477 * not, otherwise, when this l2hdr is evicted we'll
6478 * remove a reference that was never added.
6480 (void) refcount_add_many(&dev
->l2ad_alloc
, buf_sz
, hdr
);
6482 /* Compression may have squashed the buffer to zero length. */
6486 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
6487 dev
->l2ad_hand
, buf_sz
, buf_data
, ZIO_CHECKSUM_OFF
,
6488 NULL
, NULL
, ZIO_PRIORITY_ASYNC_WRITE
,
6489 ZIO_FLAG_CANFAIL
, B_FALSE
);
6491 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
6493 (void) zio_nowait(wzio
);
6495 stats_size
+= buf_sz
;
6498 * Keep the clock hand suitably device-aligned.
6500 buf_a_sz
= vdev_psize_to_asize(dev
->l2ad_vdev
, buf_sz
);
6501 write_asize
+= buf_a_sz
;
6502 dev
->l2ad_hand
+= buf_a_sz
;
6506 mutex_exit(&dev
->l2ad_mtx
);
6508 ASSERT3U(write_asize
, <=, target_sz
);
6509 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
6510 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_asize
);
6511 ARCSTAT_INCR(arcstat_l2_size
, write_sz
);
6512 ARCSTAT_INCR(arcstat_l2_asize
, stats_size
);
6513 vdev_space_update(dev
->l2ad_vdev
, stats_size
, 0, 0);
6516 * Bump device hand to the device start if it is approaching the end.
6517 * l2arc_evict() will already have evicted ahead for this case.
6519 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
6520 dev
->l2ad_hand
= dev
->l2ad_start
;
6521 dev
->l2ad_first
= B_FALSE
;
6524 dev
->l2ad_writing
= B_TRUE
;
6525 (void) zio_wait(pio
);
6526 dev
->l2ad_writing
= B_FALSE
;
6528 return (write_asize
);
6532 * Compresses an L2ARC buffer.
6533 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6534 * size in l2hdr->b_asize. This routine tries to compress the data and
6535 * depending on the compression result there are three possible outcomes:
6536 * *) The buffer was incompressible. The original l2hdr contents were left
6537 * untouched and are ready for writing to an L2 device.
6538 * *) The buffer was all-zeros, so there is no need to write it to an L2
6539 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6540 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6541 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6542 * data buffer which holds the compressed data to be written, and b_asize
6543 * tells us how much data there is. b_compress is set to the appropriate
6544 * compression algorithm. Once writing is done, invoke
6545 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6547 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6548 * buffer was incompressible).
6551 l2arc_compress_buf(arc_buf_hdr_t
*hdr
)
6554 size_t csize
, len
, rounded
;
6555 l2arc_buf_hdr_t
*l2hdr
;
6557 ASSERT(HDR_HAS_L2HDR(hdr
));
6559 l2hdr
= &hdr
->b_l2hdr
;
6561 ASSERT(HDR_HAS_L1HDR(hdr
));
6562 ASSERT(HDR_GET_COMPRESS(hdr
) == ZIO_COMPRESS_OFF
);
6563 ASSERT(hdr
->b_l1hdr
.b_tmp_cdata
!= NULL
);
6565 len
= l2hdr
->b_asize
;
6566 cdata
= zio_data_buf_alloc(len
);
6567 ASSERT3P(cdata
, !=, NULL
);
6568 csize
= zio_compress_data(ZIO_COMPRESS_LZ4
, hdr
->b_l1hdr
.b_tmp_cdata
,
6569 cdata
, l2hdr
->b_asize
);
6571 rounded
= P2ROUNDUP(csize
, (size_t)SPA_MINBLOCKSIZE
);
6572 if (rounded
> csize
) {
6573 bzero((char *)cdata
+ csize
, rounded
- csize
);
6578 /* zero block, indicate that there's nothing to write */
6579 zio_data_buf_free(cdata
, len
);
6580 HDR_SET_COMPRESS(hdr
, ZIO_COMPRESS_EMPTY
);
6582 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6583 ARCSTAT_BUMP(arcstat_l2_compress_zeros
);
6585 } else if (csize
> 0 && csize
< len
) {
6587 * Compression succeeded, we'll keep the cdata around for
6588 * writing and release it afterwards.
6590 HDR_SET_COMPRESS(hdr
, ZIO_COMPRESS_LZ4
);
6591 l2hdr
->b_asize
= csize
;
6592 hdr
->b_l1hdr
.b_tmp_cdata
= cdata
;
6593 ARCSTAT_BUMP(arcstat_l2_compress_successes
);
6597 * Compression failed, release the compressed buffer.
6598 * l2hdr will be left unmodified.
6600 zio_data_buf_free(cdata
, len
);
6601 ARCSTAT_BUMP(arcstat_l2_compress_failures
);
6607 * Decompresses a zio read back from an l2arc device. On success, the
6608 * underlying zio's io_data buffer is overwritten by the uncompressed
6609 * version. On decompression error (corrupt compressed stream), the
6610 * zio->io_error value is set to signal an I/O error.
6612 * Please note that the compressed data stream is not checksummed, so
6613 * if the underlying device is experiencing data corruption, we may feed
6614 * corrupt data to the decompressor, so the decompressor needs to be
6615 * able to handle this situation (LZ4 does).
6618 l2arc_decompress_zio(zio_t
*zio
, arc_buf_hdr_t
*hdr
, enum zio_compress c
)
6623 ASSERT(L2ARC_IS_VALID_COMPRESS(c
));
6625 if (zio
->io_error
!= 0) {
6627 * An io error has occured, just restore the original io
6628 * size in preparation for a main pool read.
6630 zio
->io_orig_size
= zio
->io_size
= hdr
->b_size
;
6634 if (c
== ZIO_COMPRESS_EMPTY
) {
6636 * An empty buffer results in a null zio, which means we
6637 * need to fill its io_data after we're done restoring the
6638 * buffer's contents.
6640 ASSERT(hdr
->b_l1hdr
.b_buf
!= NULL
);
6641 bzero(hdr
->b_l1hdr
.b_buf
->b_data
, hdr
->b_size
);
6642 zio
->io_data
= zio
->io_orig_data
= hdr
->b_l1hdr
.b_buf
->b_data
;
6644 ASSERT(zio
->io_data
!= NULL
);
6646 * We copy the compressed data from the start of the arc buffer
6647 * (the zio_read will have pulled in only what we need, the
6648 * rest is garbage which we will overwrite at decompression)
6649 * and then decompress back to the ARC data buffer. This way we
6650 * can minimize copying by simply decompressing back over the
6651 * original compressed data (rather than decompressing to an
6652 * aux buffer and then copying back the uncompressed buffer,
6653 * which is likely to be much larger).
6655 csize
= zio
->io_size
;
6656 cdata
= zio_data_buf_alloc(csize
);
6657 bcopy(zio
->io_data
, cdata
, csize
);
6658 if (zio_decompress_data(c
, cdata
, zio
->io_data
, csize
,
6660 zio
->io_error
= EIO
;
6661 zio_data_buf_free(cdata
, csize
);
6664 /* Restore the expected uncompressed IO size. */
6665 zio
->io_orig_size
= zio
->io_size
= hdr
->b_size
;
6669 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6670 * This buffer serves as a temporary holder of compressed data while
6671 * the buffer entry is being written to an l2arc device. Once that is
6672 * done, we can dispose of it.
6675 l2arc_release_cdata_buf(arc_buf_hdr_t
*hdr
)
6677 enum zio_compress comp
= HDR_GET_COMPRESS(hdr
);
6679 ASSERT(HDR_HAS_L1HDR(hdr
));
6680 ASSERT(comp
== ZIO_COMPRESS_OFF
|| L2ARC_IS_VALID_COMPRESS(comp
));
6682 if (comp
== ZIO_COMPRESS_OFF
) {
6684 * In this case, b_tmp_cdata points to the same buffer
6685 * as the arc_buf_t's b_data field. We don't want to
6686 * free it, since the arc_buf_t will handle that.
6688 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6689 } else if (comp
== ZIO_COMPRESS_EMPTY
) {
6691 * In this case, b_tmp_cdata was compressed to an empty
6692 * buffer, thus there's nothing to free and b_tmp_cdata
6693 * should have been set to NULL in l2arc_write_buffers().
6695 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
6698 * If the data was compressed, then we've allocated a
6699 * temporary buffer for it, so now we need to release it.
6701 ASSERT(hdr
->b_l1hdr
.b_tmp_cdata
!= NULL
);
6702 zio_data_buf_free(hdr
->b_l1hdr
.b_tmp_cdata
,
6704 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6710 * This thread feeds the L2ARC at regular intervals. This is the beating
6711 * heart of the L2ARC.
6714 l2arc_feed_thread(void)
6719 uint64_t size
, wrote
;
6720 clock_t begin
, next
= ddi_get_lbolt();
6721 boolean_t headroom_boost
= B_FALSE
;
6722 fstrans_cookie_t cookie
;
6724 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
6726 mutex_enter(&l2arc_feed_thr_lock
);
6728 cookie
= spl_fstrans_mark();
6729 while (l2arc_thread_exit
== 0) {
6730 CALLB_CPR_SAFE_BEGIN(&cpr
);
6731 (void) cv_timedwait_sig(&l2arc_feed_thr_cv
,
6732 &l2arc_feed_thr_lock
, next
);
6733 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
6734 next
= ddi_get_lbolt() + hz
;
6737 * Quick check for L2ARC devices.
6739 mutex_enter(&l2arc_dev_mtx
);
6740 if (l2arc_ndev
== 0) {
6741 mutex_exit(&l2arc_dev_mtx
);
6744 mutex_exit(&l2arc_dev_mtx
);
6745 begin
= ddi_get_lbolt();
6748 * This selects the next l2arc device to write to, and in
6749 * doing so the next spa to feed from: dev->l2ad_spa. This
6750 * will return NULL if there are now no l2arc devices or if
6751 * they are all faulted.
6753 * If a device is returned, its spa's config lock is also
6754 * held to prevent device removal. l2arc_dev_get_next()
6755 * will grab and release l2arc_dev_mtx.
6757 if ((dev
= l2arc_dev_get_next()) == NULL
)
6760 spa
= dev
->l2ad_spa
;
6761 ASSERT(spa
!= NULL
);
6764 * If the pool is read-only then force the feed thread to
6765 * sleep a little longer.
6767 if (!spa_writeable(spa
)) {
6768 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
6769 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6774 * Avoid contributing to memory pressure.
6776 if (arc_reclaim_needed()) {
6777 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
6778 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6782 ARCSTAT_BUMP(arcstat_l2_feeds
);
6784 size
= l2arc_write_size();
6787 * Evict L2ARC buffers that will be overwritten.
6789 l2arc_evict(dev
, size
, B_FALSE
);
6792 * Write ARC buffers.
6794 wrote
= l2arc_write_buffers(spa
, dev
, size
, &headroom_boost
);
6797 * Calculate interval between writes.
6799 next
= l2arc_write_interval(begin
, size
, wrote
);
6800 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6802 spl_fstrans_unmark(cookie
);
6804 l2arc_thread_exit
= 0;
6805 cv_broadcast(&l2arc_feed_thr_cv
);
6806 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
6811 l2arc_vdev_present(vdev_t
*vd
)
6815 mutex_enter(&l2arc_dev_mtx
);
6816 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
6817 dev
= list_next(l2arc_dev_list
, dev
)) {
6818 if (dev
->l2ad_vdev
== vd
)
6821 mutex_exit(&l2arc_dev_mtx
);
6823 return (dev
!= NULL
);
6827 * Add a vdev for use by the L2ARC. By this point the spa has already
6828 * validated the vdev and opened it.
6831 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
6833 l2arc_dev_t
*adddev
;
6835 ASSERT(!l2arc_vdev_present(vd
));
6838 * Create a new l2arc device entry.
6840 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
6841 adddev
->l2ad_spa
= spa
;
6842 adddev
->l2ad_vdev
= vd
;
6843 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
6844 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
6845 adddev
->l2ad_hand
= adddev
->l2ad_start
;
6846 adddev
->l2ad_first
= B_TRUE
;
6847 adddev
->l2ad_writing
= B_FALSE
;
6848 list_link_init(&adddev
->l2ad_node
);
6850 mutex_init(&adddev
->l2ad_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6852 * This is a list of all ARC buffers that are still valid on the
6855 list_create(&adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
6856 offsetof(arc_buf_hdr_t
, b_l2hdr
.b_l2node
));
6858 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
6859 refcount_create(&adddev
->l2ad_alloc
);
6862 * Add device to global list
6864 mutex_enter(&l2arc_dev_mtx
);
6865 list_insert_head(l2arc_dev_list
, adddev
);
6866 atomic_inc_64(&l2arc_ndev
);
6867 mutex_exit(&l2arc_dev_mtx
);
6871 * Remove a vdev from the L2ARC.
6874 l2arc_remove_vdev(vdev_t
*vd
)
6876 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
6879 * Find the device by vdev
6881 mutex_enter(&l2arc_dev_mtx
);
6882 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
6883 nextdev
= list_next(l2arc_dev_list
, dev
);
6884 if (vd
== dev
->l2ad_vdev
) {
6889 ASSERT(remdev
!= NULL
);
6892 * Remove device from global list
6894 list_remove(l2arc_dev_list
, remdev
);
6895 l2arc_dev_last
= NULL
; /* may have been invalidated */
6896 atomic_dec_64(&l2arc_ndev
);
6897 mutex_exit(&l2arc_dev_mtx
);
6900 * Clear all buflists and ARC references. L2ARC device flush.
6902 l2arc_evict(remdev
, 0, B_TRUE
);
6903 list_destroy(&remdev
->l2ad_buflist
);
6904 mutex_destroy(&remdev
->l2ad_mtx
);
6905 refcount_destroy(&remdev
->l2ad_alloc
);
6906 kmem_free(remdev
, sizeof (l2arc_dev_t
));
6912 l2arc_thread_exit
= 0;
6914 l2arc_writes_sent
= 0;
6915 l2arc_writes_done
= 0;
6917 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
6918 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
6919 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6920 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6922 l2arc_dev_list
= &L2ARC_dev_list
;
6923 l2arc_free_on_write
= &L2ARC_free_on_write
;
6924 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
6925 offsetof(l2arc_dev_t
, l2ad_node
));
6926 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
6927 offsetof(l2arc_data_free_t
, l2df_list_node
));
6934 * This is called from dmu_fini(), which is called from spa_fini();
6935 * Because of this, we can assume that all l2arc devices have
6936 * already been removed when the pools themselves were removed.
6939 l2arc_do_free_on_write();
6941 mutex_destroy(&l2arc_feed_thr_lock
);
6942 cv_destroy(&l2arc_feed_thr_cv
);
6943 mutex_destroy(&l2arc_dev_mtx
);
6944 mutex_destroy(&l2arc_free_on_write_mtx
);
6946 list_destroy(l2arc_dev_list
);
6947 list_destroy(l2arc_free_on_write
);
6953 if (!(spa_mode_global
& FWRITE
))
6956 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
6957 TS_RUN
, minclsyspri
);
6963 if (!(spa_mode_global
& FWRITE
))
6966 mutex_enter(&l2arc_feed_thr_lock
);
6967 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
6968 l2arc_thread_exit
= 1;
6969 while (l2arc_thread_exit
!= 0)
6970 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
6971 mutex_exit(&l2arc_feed_thr_lock
);
6974 #if defined(_KERNEL) && defined(HAVE_SPL)
6975 EXPORT_SYMBOL(arc_buf_size
);
6976 EXPORT_SYMBOL(arc_write
);
6977 EXPORT_SYMBOL(arc_read
);
6978 EXPORT_SYMBOL(arc_buf_remove_ref
);
6979 EXPORT_SYMBOL(arc_buf_info
);
6980 EXPORT_SYMBOL(arc_getbuf_func
);
6981 EXPORT_SYMBOL(arc_add_prune_callback
);
6982 EXPORT_SYMBOL(arc_remove_prune_callback
);
6984 module_param(zfs_arc_min
, ulong
, 0644);
6985 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
6987 module_param(zfs_arc_max
, ulong
, 0644);
6988 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
6990 module_param(zfs_arc_meta_limit
, ulong
, 0644);
6991 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
6993 module_param(zfs_arc_meta_min
, ulong
, 0644);
6994 MODULE_PARM_DESC(zfs_arc_meta_min
, "Min arc metadata");
6996 module_param(zfs_arc_meta_prune
, int, 0644);
6997 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Meta objects to scan for prune");
6999 module_param(zfs_arc_meta_adjust_restarts
, int, 0644);
7000 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts
,
7001 "Limit number of restarts in arc_adjust_meta");
7003 module_param(zfs_arc_meta_strategy
, int, 0644);
7004 MODULE_PARM_DESC(zfs_arc_meta_strategy
, "Meta reclaim strategy");
7006 module_param(zfs_arc_grow_retry
, int, 0644);
7007 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
7009 module_param(zfs_arc_p_aggressive_disable
, int, 0644);
7010 MODULE_PARM_DESC(zfs_arc_p_aggressive_disable
, "disable aggressive arc_p grow");
7012 module_param(zfs_arc_p_dampener_disable
, int, 0644);
7013 MODULE_PARM_DESC(zfs_arc_p_dampener_disable
, "disable arc_p adapt dampener");
7015 module_param(zfs_arc_shrink_shift
, int, 0644);
7016 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
7018 module_param(zfs_arc_p_min_shift
, int, 0644);
7019 MODULE_PARM_DESC(zfs_arc_p_min_shift
, "arc_c shift to calc min/max arc_p");
7021 module_param(zfs_disable_dup_eviction
, int, 0644);
7022 MODULE_PARM_DESC(zfs_disable_dup_eviction
, "disable duplicate buffer eviction");
7024 module_param(zfs_arc_average_blocksize
, int, 0444);
7025 MODULE_PARM_DESC(zfs_arc_average_blocksize
, "Target average block size");
7027 module_param(zfs_arc_memory_throttle_disable
, int, 0644);
7028 MODULE_PARM_DESC(zfs_arc_memory_throttle_disable
, "disable memory throttle");
7030 module_param(zfs_arc_min_prefetch_lifespan
, int, 0644);
7031 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan
, "Min life of prefetch block");
7033 module_param(zfs_arc_num_sublists_per_state
, int, 0644);
7034 MODULE_PARM_DESC(zfs_arc_num_sublists_per_state
,
7035 "Number of sublists used in each of the ARC state lists");
7037 module_param(l2arc_write_max
, ulong
, 0644);
7038 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
7040 module_param(l2arc_write_boost
, ulong
, 0644);
7041 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
7043 module_param(l2arc_headroom
, ulong
, 0644);
7044 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
7046 module_param(l2arc_headroom_boost
, ulong
, 0644);
7047 MODULE_PARM_DESC(l2arc_headroom_boost
, "Compressed l2arc_headroom multiplier");
7049 module_param(l2arc_feed_secs
, ulong
, 0644);
7050 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
7052 module_param(l2arc_feed_min_ms
, ulong
, 0644);
7053 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
7055 module_param(l2arc_noprefetch
, int, 0644);
7056 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
7058 module_param(l2arc_nocompress
, int, 0644);
7059 MODULE_PARM_DESC(l2arc_nocompress
, "Skip compressing L2ARC buffers");
7061 module_param(l2arc_feed_again
, int, 0644);
7062 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
7064 module_param(l2arc_norw
, int, 0644);
7065 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");