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 unsigned long zfs_arc_sys_free
= 0;
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;
250 int zfs_arc_lotsfree_percent
= 10;
253 static arc_state_t ARC_anon
;
254 static arc_state_t ARC_mru
;
255 static arc_state_t ARC_mru_ghost
;
256 static arc_state_t ARC_mfu
;
257 static arc_state_t ARC_mfu_ghost
;
258 static arc_state_t ARC_l2c_only
;
260 typedef struct arc_stats
{
261 kstat_named_t arcstat_hits
;
262 kstat_named_t arcstat_misses
;
263 kstat_named_t arcstat_demand_data_hits
;
264 kstat_named_t arcstat_demand_data_misses
;
265 kstat_named_t arcstat_demand_metadata_hits
;
266 kstat_named_t arcstat_demand_metadata_misses
;
267 kstat_named_t arcstat_prefetch_data_hits
;
268 kstat_named_t arcstat_prefetch_data_misses
;
269 kstat_named_t arcstat_prefetch_metadata_hits
;
270 kstat_named_t arcstat_prefetch_metadata_misses
;
271 kstat_named_t arcstat_mru_hits
;
272 kstat_named_t arcstat_mru_ghost_hits
;
273 kstat_named_t arcstat_mfu_hits
;
274 kstat_named_t arcstat_mfu_ghost_hits
;
275 kstat_named_t arcstat_deleted
;
277 * Number of buffers that could not be evicted because the hash lock
278 * was held by another thread. The lock may not necessarily be held
279 * by something using the same buffer, since hash locks are shared
280 * by multiple buffers.
282 kstat_named_t arcstat_mutex_miss
;
284 * Number of buffers skipped because they have I/O in progress, are
285 * indrect prefetch buffers that have not lived long enough, or are
286 * not from the spa we're trying to evict from.
288 kstat_named_t arcstat_evict_skip
;
290 * Number of times arc_evict_state() was unable to evict enough
291 * buffers to reach its target amount.
293 kstat_named_t arcstat_evict_not_enough
;
294 kstat_named_t arcstat_evict_l2_cached
;
295 kstat_named_t arcstat_evict_l2_eligible
;
296 kstat_named_t arcstat_evict_l2_ineligible
;
297 kstat_named_t arcstat_evict_l2_skip
;
298 kstat_named_t arcstat_hash_elements
;
299 kstat_named_t arcstat_hash_elements_max
;
300 kstat_named_t arcstat_hash_collisions
;
301 kstat_named_t arcstat_hash_chains
;
302 kstat_named_t arcstat_hash_chain_max
;
303 kstat_named_t arcstat_p
;
304 kstat_named_t arcstat_c
;
305 kstat_named_t arcstat_c_min
;
306 kstat_named_t arcstat_c_max
;
307 kstat_named_t arcstat_size
;
309 * Number of bytes consumed by internal ARC structures necessary
310 * for tracking purposes; these structures are not actually
311 * backed by ARC buffers. This includes arc_buf_hdr_t structures
312 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
313 * caches), and arc_buf_t structures (allocated via arc_buf_t
316 kstat_named_t arcstat_hdr_size
;
318 * Number of bytes consumed by ARC buffers of type equal to
319 * ARC_BUFC_DATA. This is generally consumed by buffers backing
320 * on disk user data (e.g. plain file contents).
322 kstat_named_t arcstat_data_size
;
324 * Number of bytes consumed by ARC buffers of type equal to
325 * ARC_BUFC_METADATA. This is generally consumed by buffers
326 * backing on disk data that is used for internal ZFS
327 * structures (e.g. ZAP, dnode, indirect blocks, etc).
329 kstat_named_t arcstat_metadata_size
;
331 * Number of bytes consumed by various buffers and structures
332 * not actually backed with ARC buffers. This includes bonus
333 * buffers (allocated directly via zio_buf_* functions),
334 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
335 * cache), and dnode_t structures (allocated via dnode_t cache).
337 kstat_named_t arcstat_other_size
;
339 * Total number of bytes consumed by ARC buffers residing in the
340 * arc_anon state. This includes *all* buffers in the arc_anon
341 * state; e.g. data, metadata, evictable, and unevictable buffers
342 * are all included in this value.
344 kstat_named_t arcstat_anon_size
;
346 * Number of bytes consumed by ARC buffers that meet the
347 * following criteria: backing buffers of type ARC_BUFC_DATA,
348 * residing in the arc_anon state, and are eligible for eviction
349 * (e.g. have no outstanding holds on the buffer).
351 kstat_named_t arcstat_anon_evictable_data
;
353 * Number of bytes consumed by ARC buffers that meet the
354 * following criteria: backing buffers of type ARC_BUFC_METADATA,
355 * residing in the arc_anon state, and are eligible for eviction
356 * (e.g. have no outstanding holds on the buffer).
358 kstat_named_t arcstat_anon_evictable_metadata
;
360 * Total number of bytes consumed by ARC buffers residing in the
361 * arc_mru state. This includes *all* buffers in the arc_mru
362 * state; e.g. data, metadata, evictable, and unevictable buffers
363 * are all included in this value.
365 kstat_named_t arcstat_mru_size
;
367 * Number of bytes consumed by ARC buffers that meet the
368 * following criteria: backing buffers of type ARC_BUFC_DATA,
369 * residing in the arc_mru state, and are eligible for eviction
370 * (e.g. have no outstanding holds on the buffer).
372 kstat_named_t arcstat_mru_evictable_data
;
374 * Number of bytes consumed by ARC buffers that meet the
375 * following criteria: backing buffers of type ARC_BUFC_METADATA,
376 * residing in the arc_mru state, and are eligible for eviction
377 * (e.g. have no outstanding holds on the buffer).
379 kstat_named_t arcstat_mru_evictable_metadata
;
381 * Total number of bytes that *would have been* consumed by ARC
382 * buffers in the arc_mru_ghost state. The key thing to note
383 * here, is the fact that this size doesn't actually indicate
384 * RAM consumption. The ghost lists only consist of headers and
385 * don't actually have ARC buffers linked off of these headers.
386 * Thus, *if* the headers had associated ARC buffers, these
387 * buffers *would have* consumed this number of bytes.
389 kstat_named_t arcstat_mru_ghost_size
;
391 * Number of bytes that *would have been* consumed by ARC
392 * buffers that are eligible for eviction, of type
393 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
395 kstat_named_t arcstat_mru_ghost_evictable_data
;
397 * Number of bytes that *would have been* consumed by ARC
398 * buffers that are eligible for eviction, of type
399 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
401 kstat_named_t arcstat_mru_ghost_evictable_metadata
;
403 * Total number of bytes consumed by ARC buffers residing in the
404 * arc_mfu state. This includes *all* buffers in the arc_mfu
405 * state; e.g. data, metadata, evictable, and unevictable buffers
406 * are all included in this value.
408 kstat_named_t arcstat_mfu_size
;
410 * Number of bytes consumed by ARC buffers that are eligible for
411 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
414 kstat_named_t arcstat_mfu_evictable_data
;
416 * Number of bytes consumed by ARC buffers that are eligible for
417 * eviction, of type ARC_BUFC_METADATA, and reside in the
420 kstat_named_t arcstat_mfu_evictable_metadata
;
422 * Total number of bytes that *would have been* consumed by ARC
423 * buffers in the arc_mfu_ghost state. See the comment above
424 * arcstat_mru_ghost_size for more details.
426 kstat_named_t arcstat_mfu_ghost_size
;
428 * Number of bytes that *would have been* consumed by ARC
429 * buffers that are eligible for eviction, of type
430 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
432 kstat_named_t arcstat_mfu_ghost_evictable_data
;
434 * Number of bytes that *would have been* consumed by ARC
435 * buffers that are eligible for eviction, of type
436 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
438 kstat_named_t arcstat_mfu_ghost_evictable_metadata
;
439 kstat_named_t arcstat_l2_hits
;
440 kstat_named_t arcstat_l2_misses
;
441 kstat_named_t arcstat_l2_feeds
;
442 kstat_named_t arcstat_l2_rw_clash
;
443 kstat_named_t arcstat_l2_read_bytes
;
444 kstat_named_t arcstat_l2_write_bytes
;
445 kstat_named_t arcstat_l2_writes_sent
;
446 kstat_named_t arcstat_l2_writes_done
;
447 kstat_named_t arcstat_l2_writes_error
;
448 kstat_named_t arcstat_l2_writes_lock_retry
;
449 kstat_named_t arcstat_l2_evict_lock_retry
;
450 kstat_named_t arcstat_l2_evict_reading
;
451 kstat_named_t arcstat_l2_evict_l1cached
;
452 kstat_named_t arcstat_l2_free_on_write
;
453 kstat_named_t arcstat_l2_cdata_free_on_write
;
454 kstat_named_t arcstat_l2_abort_lowmem
;
455 kstat_named_t arcstat_l2_cksum_bad
;
456 kstat_named_t arcstat_l2_io_error
;
457 kstat_named_t arcstat_l2_size
;
458 kstat_named_t arcstat_l2_asize
;
459 kstat_named_t arcstat_l2_hdr_size
;
460 kstat_named_t arcstat_l2_compress_successes
;
461 kstat_named_t arcstat_l2_compress_zeros
;
462 kstat_named_t arcstat_l2_compress_failures
;
463 kstat_named_t arcstat_memory_throttle_count
;
464 kstat_named_t arcstat_duplicate_buffers
;
465 kstat_named_t arcstat_duplicate_buffers_size
;
466 kstat_named_t arcstat_duplicate_reads
;
467 kstat_named_t arcstat_memory_direct_count
;
468 kstat_named_t arcstat_memory_indirect_count
;
469 kstat_named_t arcstat_no_grow
;
470 kstat_named_t arcstat_tempreserve
;
471 kstat_named_t arcstat_loaned_bytes
;
472 kstat_named_t arcstat_prune
;
473 kstat_named_t arcstat_meta_used
;
474 kstat_named_t arcstat_meta_limit
;
475 kstat_named_t arcstat_meta_max
;
476 kstat_named_t arcstat_meta_min
;
477 kstat_named_t arcstat_need_free
;
478 kstat_named_t arcstat_sys_free
;
481 static arc_stats_t arc_stats
= {
482 { "hits", KSTAT_DATA_UINT64
},
483 { "misses", KSTAT_DATA_UINT64
},
484 { "demand_data_hits", KSTAT_DATA_UINT64
},
485 { "demand_data_misses", KSTAT_DATA_UINT64
},
486 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
487 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
488 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
489 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
490 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
491 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
492 { "mru_hits", KSTAT_DATA_UINT64
},
493 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
494 { "mfu_hits", KSTAT_DATA_UINT64
},
495 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
496 { "deleted", KSTAT_DATA_UINT64
},
497 { "mutex_miss", KSTAT_DATA_UINT64
},
498 { "evict_skip", KSTAT_DATA_UINT64
},
499 { "evict_not_enough", KSTAT_DATA_UINT64
},
500 { "evict_l2_cached", KSTAT_DATA_UINT64
},
501 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
502 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
503 { "evict_l2_skip", KSTAT_DATA_UINT64
},
504 { "hash_elements", KSTAT_DATA_UINT64
},
505 { "hash_elements_max", KSTAT_DATA_UINT64
},
506 { "hash_collisions", KSTAT_DATA_UINT64
},
507 { "hash_chains", KSTAT_DATA_UINT64
},
508 { "hash_chain_max", KSTAT_DATA_UINT64
},
509 { "p", KSTAT_DATA_UINT64
},
510 { "c", KSTAT_DATA_UINT64
},
511 { "c_min", KSTAT_DATA_UINT64
},
512 { "c_max", KSTAT_DATA_UINT64
},
513 { "size", KSTAT_DATA_UINT64
},
514 { "hdr_size", KSTAT_DATA_UINT64
},
515 { "data_size", KSTAT_DATA_UINT64
},
516 { "metadata_size", KSTAT_DATA_UINT64
},
517 { "other_size", KSTAT_DATA_UINT64
},
518 { "anon_size", KSTAT_DATA_UINT64
},
519 { "anon_evictable_data", KSTAT_DATA_UINT64
},
520 { "anon_evictable_metadata", KSTAT_DATA_UINT64
},
521 { "mru_size", KSTAT_DATA_UINT64
},
522 { "mru_evictable_data", KSTAT_DATA_UINT64
},
523 { "mru_evictable_metadata", KSTAT_DATA_UINT64
},
524 { "mru_ghost_size", KSTAT_DATA_UINT64
},
525 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64
},
526 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
527 { "mfu_size", KSTAT_DATA_UINT64
},
528 { "mfu_evictable_data", KSTAT_DATA_UINT64
},
529 { "mfu_evictable_metadata", KSTAT_DATA_UINT64
},
530 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
531 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64
},
532 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
533 { "l2_hits", KSTAT_DATA_UINT64
},
534 { "l2_misses", KSTAT_DATA_UINT64
},
535 { "l2_feeds", KSTAT_DATA_UINT64
},
536 { "l2_rw_clash", KSTAT_DATA_UINT64
},
537 { "l2_read_bytes", KSTAT_DATA_UINT64
},
538 { "l2_write_bytes", KSTAT_DATA_UINT64
},
539 { "l2_writes_sent", KSTAT_DATA_UINT64
},
540 { "l2_writes_done", KSTAT_DATA_UINT64
},
541 { "l2_writes_error", KSTAT_DATA_UINT64
},
542 { "l2_writes_lock_retry", KSTAT_DATA_UINT64
},
543 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
544 { "l2_evict_reading", KSTAT_DATA_UINT64
},
545 { "l2_evict_l1cached", KSTAT_DATA_UINT64
},
546 { "l2_free_on_write", KSTAT_DATA_UINT64
},
547 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64
},
548 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
549 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
550 { "l2_io_error", KSTAT_DATA_UINT64
},
551 { "l2_size", KSTAT_DATA_UINT64
},
552 { "l2_asize", KSTAT_DATA_UINT64
},
553 { "l2_hdr_size", KSTAT_DATA_UINT64
},
554 { "l2_compress_successes", KSTAT_DATA_UINT64
},
555 { "l2_compress_zeros", KSTAT_DATA_UINT64
},
556 { "l2_compress_failures", KSTAT_DATA_UINT64
},
557 { "memory_throttle_count", KSTAT_DATA_UINT64
},
558 { "duplicate_buffers", KSTAT_DATA_UINT64
},
559 { "duplicate_buffers_size", KSTAT_DATA_UINT64
},
560 { "duplicate_reads", KSTAT_DATA_UINT64
},
561 { "memory_direct_count", KSTAT_DATA_UINT64
},
562 { "memory_indirect_count", KSTAT_DATA_UINT64
},
563 { "arc_no_grow", KSTAT_DATA_UINT64
},
564 { "arc_tempreserve", KSTAT_DATA_UINT64
},
565 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
566 { "arc_prune", KSTAT_DATA_UINT64
},
567 { "arc_meta_used", KSTAT_DATA_UINT64
},
568 { "arc_meta_limit", KSTAT_DATA_UINT64
},
569 { "arc_meta_max", KSTAT_DATA_UINT64
},
570 { "arc_meta_min", KSTAT_DATA_UINT64
},
571 { "arc_need_free", KSTAT_DATA_UINT64
},
572 { "arc_sys_free", KSTAT_DATA_UINT64
}
575 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
577 #define ARCSTAT_INCR(stat, val) \
578 atomic_add_64(&arc_stats.stat.value.ui64, (val))
580 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
581 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
583 #define ARCSTAT_MAX(stat, val) { \
585 while ((val) > (m = arc_stats.stat.value.ui64) && \
586 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
590 #define ARCSTAT_MAXSTAT(stat) \
591 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
594 * We define a macro to allow ARC hits/misses to be easily broken down by
595 * two separate conditions, giving a total of four different subtypes for
596 * each of hits and misses (so eight statistics total).
598 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
601 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
603 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
607 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
609 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
614 static arc_state_t
*arc_anon
;
615 static arc_state_t
*arc_mru
;
616 static arc_state_t
*arc_mru_ghost
;
617 static arc_state_t
*arc_mfu
;
618 static arc_state_t
*arc_mfu_ghost
;
619 static arc_state_t
*arc_l2c_only
;
622 * There are several ARC variables that are critical to export as kstats --
623 * but we don't want to have to grovel around in the kstat whenever we wish to
624 * manipulate them. For these variables, we therefore define them to be in
625 * terms of the statistic variable. This assures that we are not introducing
626 * the possibility of inconsistency by having shadow copies of the variables,
627 * while still allowing the code to be readable.
629 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
630 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
631 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
632 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
633 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
634 #define arc_no_grow ARCSTAT(arcstat_no_grow)
635 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
636 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
637 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
638 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
639 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
640 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
641 #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */
642 #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */
644 #define L2ARC_IS_VALID_COMPRESS(_c_) \
645 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
647 static list_t arc_prune_list
;
648 static kmutex_t arc_prune_mtx
;
649 static taskq_t
*arc_prune_taskq
;
650 static arc_buf_t
*arc_eviction_list
;
651 static arc_buf_hdr_t arc_eviction_hdr
;
653 #define GHOST_STATE(state) \
654 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
655 (state) == arc_l2c_only)
657 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
658 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
659 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
660 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
661 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
662 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
664 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
665 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
666 #define HDR_L2_READING(hdr) \
667 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
668 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
669 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
670 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
671 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
673 #define HDR_ISTYPE_METADATA(hdr) \
674 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
675 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
677 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
678 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
684 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
685 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
688 * Hash table routines
691 #define HT_LOCK_ALIGN 64
692 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
697 unsigned char pad
[HT_LOCK_PAD
];
701 #define BUF_LOCKS 8192
702 typedef struct buf_hash_table
{
704 arc_buf_hdr_t
**ht_table
;
705 struct ht_lock ht_locks
[BUF_LOCKS
];
708 static buf_hash_table_t buf_hash_table
;
710 #define BUF_HASH_INDEX(spa, dva, birth) \
711 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
712 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
713 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
714 #define HDR_LOCK(hdr) \
715 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
717 uint64_t zfs_crc64_table
[256];
723 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
724 #define L2ARC_HEADROOM 2 /* num of writes */
726 * If we discover during ARC scan any buffers to be compressed, we boost
727 * our headroom for the next scanning cycle by this percentage multiple.
729 #define L2ARC_HEADROOM_BOOST 200
730 #define L2ARC_FEED_SECS 1 /* caching interval secs */
731 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
734 * Used to distinguish headers that are being process by
735 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
736 * address. This can happen when the header is added to the l2arc's list
737 * of buffers to write in the first stage of l2arc_write_buffers(), but
738 * has not yet been written out which happens in the second stage of
739 * l2arc_write_buffers().
741 #define L2ARC_ADDR_UNSET ((uint64_t)(-1))
743 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
744 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
746 /* L2ARC Performance Tunables */
747 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
748 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
749 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
750 unsigned long l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
751 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
752 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
753 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
754 int l2arc_nocompress
= B_FALSE
; /* don't compress bufs */
755 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
756 int l2arc_norw
= B_FALSE
; /* no reads during writes */
761 static list_t L2ARC_dev_list
; /* device list */
762 static list_t
*l2arc_dev_list
; /* device list pointer */
763 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
764 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
765 static list_t L2ARC_free_on_write
; /* free after write buf list */
766 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
767 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
768 static uint64_t l2arc_ndev
; /* number of devices */
770 typedef struct l2arc_read_callback
{
771 arc_buf_t
*l2rcb_buf
; /* read buffer */
772 spa_t
*l2rcb_spa
; /* spa */
773 blkptr_t l2rcb_bp
; /* original blkptr */
774 zbookmark_phys_t l2rcb_zb
; /* original bookmark */
775 int l2rcb_flags
; /* original flags */
776 enum zio_compress l2rcb_compress
; /* applied compress */
777 } l2arc_read_callback_t
;
779 typedef struct l2arc_data_free
{
780 /* protected by l2arc_free_on_write_mtx */
783 void (*l2df_func
)(void *, size_t);
784 list_node_t l2df_list_node
;
787 static kmutex_t l2arc_feed_thr_lock
;
788 static kcondvar_t l2arc_feed_thr_cv
;
789 static uint8_t l2arc_thread_exit
;
791 static void arc_get_data_buf(arc_buf_t
*);
792 static void arc_access(arc_buf_hdr_t
*, kmutex_t
*);
793 static boolean_t
arc_is_overflowing(void);
794 static void arc_buf_watch(arc_buf_t
*);
795 static void arc_tuning_update(void);
797 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t
*);
798 static uint32_t arc_bufc_to_flags(arc_buf_contents_t
);
800 static boolean_t
l2arc_write_eligible(uint64_t, arc_buf_hdr_t
*);
801 static void l2arc_read_done(zio_t
*);
803 static boolean_t
l2arc_compress_buf(arc_buf_hdr_t
*);
804 static void l2arc_decompress_zio(zio_t
*, arc_buf_hdr_t
*, enum zio_compress
);
805 static void l2arc_release_cdata_buf(arc_buf_hdr_t
*);
808 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
810 uint8_t *vdva
= (uint8_t *)dva
;
811 uint64_t crc
= -1ULL;
814 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
816 for (i
= 0; i
< sizeof (dva_t
); i
++)
817 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
819 crc
^= (spa
>>8) ^ birth
;
824 #define BUF_EMPTY(buf) \
825 ((buf)->b_dva.dva_word[0] == 0 && \
826 (buf)->b_dva.dva_word[1] == 0)
828 #define BUF_EQUAL(spa, dva, birth, buf) \
829 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
830 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
831 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
834 buf_discard_identity(arc_buf_hdr_t
*hdr
)
836 hdr
->b_dva
.dva_word
[0] = 0;
837 hdr
->b_dva
.dva_word
[1] = 0;
841 static arc_buf_hdr_t
*
842 buf_hash_find(uint64_t spa
, const blkptr_t
*bp
, kmutex_t
**lockp
)
844 const dva_t
*dva
= BP_IDENTITY(bp
);
845 uint64_t birth
= BP_PHYSICAL_BIRTH(bp
);
846 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
847 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
850 mutex_enter(hash_lock
);
851 for (hdr
= buf_hash_table
.ht_table
[idx
]; hdr
!= NULL
;
852 hdr
= hdr
->b_hash_next
) {
853 if (BUF_EQUAL(spa
, dva
, birth
, hdr
)) {
858 mutex_exit(hash_lock
);
864 * Insert an entry into the hash table. If there is already an element
865 * equal to elem in the hash table, then the already existing element
866 * will be returned and the new element will not be inserted.
867 * Otherwise returns NULL.
868 * If lockp == NULL, the caller is assumed to already hold the hash lock.
870 static arc_buf_hdr_t
*
871 buf_hash_insert(arc_buf_hdr_t
*hdr
, kmutex_t
**lockp
)
873 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
874 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
878 ASSERT(!DVA_IS_EMPTY(&hdr
->b_dva
));
879 ASSERT(hdr
->b_birth
!= 0);
880 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
884 mutex_enter(hash_lock
);
886 ASSERT(MUTEX_HELD(hash_lock
));
889 for (fhdr
= buf_hash_table
.ht_table
[idx
], i
= 0; fhdr
!= NULL
;
890 fhdr
= fhdr
->b_hash_next
, i
++) {
891 if (BUF_EQUAL(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
, fhdr
))
895 hdr
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
896 buf_hash_table
.ht_table
[idx
] = hdr
;
897 hdr
->b_flags
|= ARC_FLAG_IN_HASH_TABLE
;
899 /* collect some hash table performance data */
901 ARCSTAT_BUMP(arcstat_hash_collisions
);
903 ARCSTAT_BUMP(arcstat_hash_chains
);
905 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
908 ARCSTAT_BUMP(arcstat_hash_elements
);
909 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
915 buf_hash_remove(arc_buf_hdr_t
*hdr
)
917 arc_buf_hdr_t
*fhdr
, **hdrp
;
918 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
920 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
921 ASSERT(HDR_IN_HASH_TABLE(hdr
));
923 hdrp
= &buf_hash_table
.ht_table
[idx
];
924 while ((fhdr
= *hdrp
) != hdr
) {
925 ASSERT(fhdr
!= NULL
);
926 hdrp
= &fhdr
->b_hash_next
;
928 *hdrp
= hdr
->b_hash_next
;
929 hdr
->b_hash_next
= NULL
;
930 hdr
->b_flags
&= ~ARC_FLAG_IN_HASH_TABLE
;
932 /* collect some hash table performance data */
933 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
935 if (buf_hash_table
.ht_table
[idx
] &&
936 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
937 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
941 * Global data structures and functions for the buf kmem cache.
943 static kmem_cache_t
*hdr_full_cache
;
944 static kmem_cache_t
*hdr_l2only_cache
;
945 static kmem_cache_t
*buf_cache
;
952 #if defined(_KERNEL) && defined(HAVE_SPL)
954 * Large allocations which do not require contiguous pages
955 * should be using vmem_free() in the linux kernel\
957 vmem_free(buf_hash_table
.ht_table
,
958 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
960 kmem_free(buf_hash_table
.ht_table
,
961 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
963 for (i
= 0; i
< BUF_LOCKS
; i
++)
964 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
965 kmem_cache_destroy(hdr_full_cache
);
966 kmem_cache_destroy(hdr_l2only_cache
);
967 kmem_cache_destroy(buf_cache
);
971 * Constructor callback - called when the cache is empty
972 * and a new buf is requested.
976 hdr_full_cons(void *vbuf
, void *unused
, int kmflag
)
978 arc_buf_hdr_t
*hdr
= vbuf
;
980 bzero(hdr
, HDR_FULL_SIZE
);
981 cv_init(&hdr
->b_l1hdr
.b_cv
, NULL
, CV_DEFAULT
, NULL
);
982 refcount_create(&hdr
->b_l1hdr
.b_refcnt
);
983 mutex_init(&hdr
->b_l1hdr
.b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
984 list_link_init(&hdr
->b_l1hdr
.b_arc_node
);
985 list_link_init(&hdr
->b_l2hdr
.b_l2node
);
986 multilist_link_init(&hdr
->b_l1hdr
.b_arc_node
);
987 arc_space_consume(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
994 hdr_l2only_cons(void *vbuf
, void *unused
, int kmflag
)
996 arc_buf_hdr_t
*hdr
= vbuf
;
998 bzero(hdr
, HDR_L2ONLY_SIZE
);
999 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1006 buf_cons(void *vbuf
, void *unused
, int kmflag
)
1008 arc_buf_t
*buf
= vbuf
;
1010 bzero(buf
, sizeof (arc_buf_t
));
1011 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1012 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1018 * Destructor callback - called when a cached buf is
1019 * no longer required.
1023 hdr_full_dest(void *vbuf
, void *unused
)
1025 arc_buf_hdr_t
*hdr
= vbuf
;
1027 ASSERT(BUF_EMPTY(hdr
));
1028 cv_destroy(&hdr
->b_l1hdr
.b_cv
);
1029 refcount_destroy(&hdr
->b_l1hdr
.b_refcnt
);
1030 mutex_destroy(&hdr
->b_l1hdr
.b_freeze_lock
);
1031 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1032 arc_space_return(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1037 hdr_l2only_dest(void *vbuf
, void *unused
)
1039 ASSERTV(arc_buf_hdr_t
*hdr
= vbuf
);
1041 ASSERT(BUF_EMPTY(hdr
));
1042 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1047 buf_dest(void *vbuf
, void *unused
)
1049 arc_buf_t
*buf
= vbuf
;
1051 mutex_destroy(&buf
->b_evict_lock
);
1052 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1056 * Reclaim callback -- invoked when memory is low.
1060 hdr_recl(void *unused
)
1062 dprintf("hdr_recl called\n");
1064 * umem calls the reclaim func when we destroy the buf cache,
1065 * which is after we do arc_fini().
1068 cv_signal(&arc_reclaim_thread_cv
);
1075 uint64_t hsize
= 1ULL << 12;
1079 * The hash table is big enough to fill all of physical memory
1080 * with an average block size of zfs_arc_average_blocksize (default 8K).
1081 * By default, the table will take up
1082 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1084 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
1087 buf_hash_table
.ht_mask
= hsize
- 1;
1088 #if defined(_KERNEL) && defined(HAVE_SPL)
1090 * Large allocations which do not require contiguous pages
1091 * should be using vmem_alloc() in the linux kernel
1093 buf_hash_table
.ht_table
=
1094 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
1096 buf_hash_table
.ht_table
=
1097 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
1099 if (buf_hash_table
.ht_table
== NULL
) {
1100 ASSERT(hsize
> (1ULL << 8));
1105 hdr_full_cache
= kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE
,
1106 0, hdr_full_cons
, hdr_full_dest
, hdr_recl
, NULL
, NULL
, 0);
1107 hdr_l2only_cache
= kmem_cache_create("arc_buf_hdr_t_l2only",
1108 HDR_L2ONLY_SIZE
, 0, hdr_l2only_cons
, hdr_l2only_dest
, hdr_recl
,
1110 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
1111 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
1113 for (i
= 0; i
< 256; i
++)
1114 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
1115 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
1117 for (i
= 0; i
< BUF_LOCKS
; i
++) {
1118 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
1119 NULL
, MUTEX_DEFAULT
, NULL
);
1124 * Transition between the two allocation states for the arc_buf_hdr struct.
1125 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1126 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1127 * version is used when a cache buffer is only in the L2ARC in order to reduce
1130 static arc_buf_hdr_t
*
1131 arc_hdr_realloc(arc_buf_hdr_t
*hdr
, kmem_cache_t
*old
, kmem_cache_t
*new)
1133 arc_buf_hdr_t
*nhdr
;
1136 ASSERT(HDR_HAS_L2HDR(hdr
));
1137 ASSERT((old
== hdr_full_cache
&& new == hdr_l2only_cache
) ||
1138 (old
== hdr_l2only_cache
&& new == hdr_full_cache
));
1140 dev
= hdr
->b_l2hdr
.b_dev
;
1141 nhdr
= kmem_cache_alloc(new, KM_PUSHPAGE
);
1143 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
1144 buf_hash_remove(hdr
);
1146 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
1148 if (new == hdr_full_cache
) {
1149 nhdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
1151 * arc_access and arc_change_state need to be aware that a
1152 * header has just come out of L2ARC, so we set its state to
1153 * l2c_only even though it's about to change.
1155 nhdr
->b_l1hdr
.b_state
= arc_l2c_only
;
1157 /* Verify previous threads set to NULL before freeing */
1158 ASSERT3P(nhdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1160 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1161 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
1164 * If we've reached here, We must have been called from
1165 * arc_evict_hdr(), as such we should have already been
1166 * removed from any ghost list we were previously on
1167 * (which protects us from racing with arc_evict_state),
1168 * thus no locking is needed during this check.
1170 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1173 * A buffer must not be moved into the arc_l2c_only
1174 * state if it's not finished being written out to the
1175 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1176 * might try to be accessed, even though it was removed.
1178 VERIFY(!HDR_L2_WRITING(hdr
));
1179 VERIFY3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1181 nhdr
->b_flags
&= ~ARC_FLAG_HAS_L1HDR
;
1184 * The header has been reallocated so we need to re-insert it into any
1187 (void) buf_hash_insert(nhdr
, NULL
);
1189 ASSERT(list_link_active(&hdr
->b_l2hdr
.b_l2node
));
1191 mutex_enter(&dev
->l2ad_mtx
);
1194 * We must place the realloc'ed header back into the list at
1195 * the same spot. Otherwise, if it's placed earlier in the list,
1196 * l2arc_write_buffers() could find it during the function's
1197 * write phase, and try to write it out to the l2arc.
1199 list_insert_after(&dev
->l2ad_buflist
, hdr
, nhdr
);
1200 list_remove(&dev
->l2ad_buflist
, hdr
);
1202 mutex_exit(&dev
->l2ad_mtx
);
1205 * Since we're using the pointer address as the tag when
1206 * incrementing and decrementing the l2ad_alloc refcount, we
1207 * must remove the old pointer (that we're about to destroy) and
1208 * add the new pointer to the refcount. Otherwise we'd remove
1209 * the wrong pointer address when calling arc_hdr_destroy() later.
1212 (void) refcount_remove_many(&dev
->l2ad_alloc
,
1213 hdr
->b_l2hdr
.b_asize
, hdr
);
1215 (void) refcount_add_many(&dev
->l2ad_alloc
,
1216 nhdr
->b_l2hdr
.b_asize
, nhdr
);
1218 buf_discard_identity(hdr
);
1219 hdr
->b_freeze_cksum
= NULL
;
1220 kmem_cache_free(old
, hdr
);
1226 #define ARC_MINTIME (hz>>4) /* 62 ms */
1229 arc_cksum_verify(arc_buf_t
*buf
)
1233 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1236 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1237 if (buf
->b_hdr
->b_freeze_cksum
== NULL
|| HDR_IO_ERROR(buf
->b_hdr
)) {
1238 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1241 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
1242 if (!ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
))
1243 panic("buffer modified while frozen!");
1244 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1248 arc_cksum_equal(arc_buf_t
*buf
)
1253 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1254 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
1255 equal
= ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
);
1256 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1262 arc_cksum_compute(arc_buf_t
*buf
, boolean_t force
)
1264 if (!force
&& !(zfs_flags
& ZFS_DEBUG_MODIFY
))
1267 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1268 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1269 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1272 buf
->b_hdr
->b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
), KM_SLEEP
);
1273 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
,
1274 buf
->b_hdr
->b_freeze_cksum
);
1275 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1281 arc_buf_sigsegv(int sig
, siginfo_t
*si
, void *unused
)
1283 panic("Got SIGSEGV at address: 0x%lx\n", (long) si
->si_addr
);
1289 arc_buf_unwatch(arc_buf_t
*buf
)
1293 ASSERT0(mprotect(buf
->b_data
, buf
->b_hdr
->b_size
,
1294 PROT_READ
| PROT_WRITE
));
1301 arc_buf_watch(arc_buf_t
*buf
)
1305 ASSERT0(mprotect(buf
->b_data
, buf
->b_hdr
->b_size
, PROT_READ
));
1309 static arc_buf_contents_t
1310 arc_buf_type(arc_buf_hdr_t
*hdr
)
1312 if (HDR_ISTYPE_METADATA(hdr
)) {
1313 return (ARC_BUFC_METADATA
);
1315 return (ARC_BUFC_DATA
);
1320 arc_bufc_to_flags(arc_buf_contents_t type
)
1324 /* metadata field is 0 if buffer contains normal data */
1326 case ARC_BUFC_METADATA
:
1327 return (ARC_FLAG_BUFC_METADATA
);
1331 panic("undefined ARC buffer type!");
1332 return ((uint32_t)-1);
1336 arc_buf_thaw(arc_buf_t
*buf
)
1338 if (zfs_flags
& ZFS_DEBUG_MODIFY
) {
1339 if (buf
->b_hdr
->b_l1hdr
.b_state
!= arc_anon
)
1340 panic("modifying non-anon buffer!");
1341 if (HDR_IO_IN_PROGRESS(buf
->b_hdr
))
1342 panic("modifying buffer while i/o in progress!");
1343 arc_cksum_verify(buf
);
1346 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1347 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1348 kmem_free(buf
->b_hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
1349 buf
->b_hdr
->b_freeze_cksum
= NULL
;
1352 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1354 arc_buf_unwatch(buf
);
1358 arc_buf_freeze(arc_buf_t
*buf
)
1360 kmutex_t
*hash_lock
;
1362 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1365 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1366 mutex_enter(hash_lock
);
1368 ASSERT(buf
->b_hdr
->b_freeze_cksum
!= NULL
||
1369 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
1370 arc_cksum_compute(buf
, B_FALSE
);
1371 mutex_exit(hash_lock
);
1376 add_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1380 ASSERT(HDR_HAS_L1HDR(hdr
));
1381 ASSERT(MUTEX_HELD(hash_lock
));
1383 state
= hdr
->b_l1hdr
.b_state
;
1385 if ((refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
) == 1) &&
1386 (state
!= arc_anon
)) {
1387 /* We don't use the L2-only state list. */
1388 if (state
!= arc_l2c_only
) {
1389 arc_buf_contents_t type
= arc_buf_type(hdr
);
1390 uint64_t delta
= hdr
->b_size
* hdr
->b_l1hdr
.b_datacnt
;
1391 multilist_t
*list
= &state
->arcs_list
[type
];
1392 uint64_t *size
= &state
->arcs_lsize
[type
];
1394 multilist_remove(list
, hdr
);
1396 if (GHOST_STATE(state
)) {
1397 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
1398 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1399 delta
= hdr
->b_size
;
1402 ASSERT3U(*size
, >=, delta
);
1403 atomic_add_64(size
, -delta
);
1405 /* remove the prefetch flag if we get a reference */
1406 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
1411 remove_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1414 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
1416 ASSERT(HDR_HAS_L1HDR(hdr
));
1417 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
1418 ASSERT(!GHOST_STATE(state
));
1421 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1422 * check to prevent usage of the arc_l2c_only list.
1424 if (((cnt
= refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
)) == 0) &&
1425 (state
!= arc_anon
)) {
1426 arc_buf_contents_t type
= arc_buf_type(hdr
);
1427 multilist_t
*list
= &state
->arcs_list
[type
];
1428 uint64_t *size
= &state
->arcs_lsize
[type
];
1430 multilist_insert(list
, hdr
);
1432 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
1433 atomic_add_64(size
, hdr
->b_size
*
1434 hdr
->b_l1hdr
.b_datacnt
);
1440 * Returns detailed information about a specific arc buffer. When the
1441 * state_index argument is set the function will calculate the arc header
1442 * list position for its arc state. Since this requires a linear traversal
1443 * callers are strongly encourage not to do this. However, it can be helpful
1444 * for targeted analysis so the functionality is provided.
1447 arc_buf_info(arc_buf_t
*ab
, arc_buf_info_t
*abi
, int state_index
)
1449 arc_buf_hdr_t
*hdr
= ab
->b_hdr
;
1450 l1arc_buf_hdr_t
*l1hdr
= NULL
;
1451 l2arc_buf_hdr_t
*l2hdr
= NULL
;
1452 arc_state_t
*state
= NULL
;
1454 if (HDR_HAS_L1HDR(hdr
)) {
1455 l1hdr
= &hdr
->b_l1hdr
;
1456 state
= l1hdr
->b_state
;
1458 if (HDR_HAS_L2HDR(hdr
))
1459 l2hdr
= &hdr
->b_l2hdr
;
1461 memset(abi
, 0, sizeof (arc_buf_info_t
));
1462 abi
->abi_flags
= hdr
->b_flags
;
1465 abi
->abi_datacnt
= l1hdr
->b_datacnt
;
1466 abi
->abi_access
= l1hdr
->b_arc_access
;
1467 abi
->abi_mru_hits
= l1hdr
->b_mru_hits
;
1468 abi
->abi_mru_ghost_hits
= l1hdr
->b_mru_ghost_hits
;
1469 abi
->abi_mfu_hits
= l1hdr
->b_mfu_hits
;
1470 abi
->abi_mfu_ghost_hits
= l1hdr
->b_mfu_ghost_hits
;
1471 abi
->abi_holds
= refcount_count(&l1hdr
->b_refcnt
);
1475 abi
->abi_l2arc_dattr
= l2hdr
->b_daddr
;
1476 abi
->abi_l2arc_asize
= l2hdr
->b_asize
;
1477 abi
->abi_l2arc_compress
= l2hdr
->b_compress
;
1478 abi
->abi_l2arc_hits
= l2hdr
->b_hits
;
1481 abi
->abi_state_type
= state
? state
->arcs_state
: ARC_STATE_ANON
;
1482 abi
->abi_state_contents
= arc_buf_type(hdr
);
1483 abi
->abi_size
= hdr
->b_size
;
1487 * Move the supplied buffer to the indicated state. The hash lock
1488 * for the buffer must be held by the caller.
1491 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*hdr
,
1492 kmutex_t
*hash_lock
)
1494 arc_state_t
*old_state
;
1497 uint64_t from_delta
, to_delta
;
1498 arc_buf_contents_t buftype
= arc_buf_type(hdr
);
1501 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1502 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1503 * L1 hdr doesn't always exist when we change state to arc_anon before
1504 * destroying a header, in which case reallocating to add the L1 hdr is
1507 if (HDR_HAS_L1HDR(hdr
)) {
1508 old_state
= hdr
->b_l1hdr
.b_state
;
1509 refcnt
= refcount_count(&hdr
->b_l1hdr
.b_refcnt
);
1510 datacnt
= hdr
->b_l1hdr
.b_datacnt
;
1512 old_state
= arc_l2c_only
;
1517 ASSERT(MUTEX_HELD(hash_lock
));
1518 ASSERT3P(new_state
, !=, old_state
);
1519 ASSERT(refcnt
== 0 || datacnt
> 0);
1520 ASSERT(!GHOST_STATE(new_state
) || datacnt
== 0);
1521 ASSERT(old_state
!= arc_anon
|| datacnt
<= 1);
1523 from_delta
= to_delta
= datacnt
* hdr
->b_size
;
1526 * If this buffer is evictable, transfer it from the
1527 * old state list to the new state list.
1530 if (old_state
!= arc_anon
&& old_state
!= arc_l2c_only
) {
1531 uint64_t *size
= &old_state
->arcs_lsize
[buftype
];
1533 ASSERT(HDR_HAS_L1HDR(hdr
));
1534 multilist_remove(&old_state
->arcs_list
[buftype
], hdr
);
1537 * If prefetching out of the ghost cache,
1538 * we will have a non-zero datacnt.
1540 if (GHOST_STATE(old_state
) && datacnt
== 0) {
1541 /* ghost elements have a ghost size */
1542 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1543 from_delta
= hdr
->b_size
;
1545 ASSERT3U(*size
, >=, from_delta
);
1546 atomic_add_64(size
, -from_delta
);
1548 if (new_state
!= arc_anon
&& new_state
!= arc_l2c_only
) {
1549 uint64_t *size
= &new_state
->arcs_lsize
[buftype
];
1552 * An L1 header always exists here, since if we're
1553 * moving to some L1-cached state (i.e. not l2c_only or
1554 * anonymous), we realloc the header to add an L1hdr
1557 ASSERT(HDR_HAS_L1HDR(hdr
));
1558 multilist_insert(&new_state
->arcs_list
[buftype
], hdr
);
1560 /* ghost elements have a ghost size */
1561 if (GHOST_STATE(new_state
)) {
1563 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1564 to_delta
= hdr
->b_size
;
1566 atomic_add_64(size
, to_delta
);
1570 ASSERT(!BUF_EMPTY(hdr
));
1571 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(hdr
))
1572 buf_hash_remove(hdr
);
1574 /* adjust state sizes (ignore arc_l2c_only) */
1576 if (to_delta
&& new_state
!= arc_l2c_only
) {
1577 ASSERT(HDR_HAS_L1HDR(hdr
));
1578 if (GHOST_STATE(new_state
)) {
1582 * We moving a header to a ghost state, we first
1583 * remove all arc buffers. Thus, we'll have a
1584 * datacnt of zero, and no arc buffer to use for
1585 * the reference. As a result, we use the arc
1586 * header pointer for the reference.
1588 (void) refcount_add_many(&new_state
->arcs_size
,
1592 ASSERT3U(datacnt
, !=, 0);
1595 * Each individual buffer holds a unique reference,
1596 * thus we must remove each of these references one
1599 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
1600 buf
= buf
->b_next
) {
1601 (void) refcount_add_many(&new_state
->arcs_size
,
1607 if (from_delta
&& old_state
!= arc_l2c_only
) {
1608 ASSERT(HDR_HAS_L1HDR(hdr
));
1609 if (GHOST_STATE(old_state
)) {
1611 * When moving a header off of a ghost state,
1612 * there's the possibility for datacnt to be
1613 * non-zero. This is because we first add the
1614 * arc buffer to the header prior to changing
1615 * the header's state. Since we used the header
1616 * for the reference when putting the header on
1617 * the ghost state, we must balance that and use
1618 * the header when removing off the ghost state
1619 * (even though datacnt is non zero).
1622 IMPLY(datacnt
== 0, new_state
== arc_anon
||
1623 new_state
== arc_l2c_only
);
1625 (void) refcount_remove_many(&old_state
->arcs_size
,
1629 ASSERT3U(datacnt
, !=, 0);
1632 * Each individual buffer holds a unique reference,
1633 * thus we must remove each of these references one
1636 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
1637 buf
= buf
->b_next
) {
1638 (void) refcount_remove_many(
1639 &old_state
->arcs_size
, hdr
->b_size
, buf
);
1644 if (HDR_HAS_L1HDR(hdr
))
1645 hdr
->b_l1hdr
.b_state
= new_state
;
1648 * L2 headers should never be on the L2 state list since they don't
1649 * have L1 headers allocated.
1651 ASSERT(multilist_is_empty(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]) &&
1652 multilist_is_empty(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]));
1656 arc_space_consume(uint64_t space
, arc_space_type_t type
)
1658 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1663 case ARC_SPACE_DATA
:
1664 ARCSTAT_INCR(arcstat_data_size
, space
);
1666 case ARC_SPACE_META
:
1667 ARCSTAT_INCR(arcstat_metadata_size
, space
);
1669 case ARC_SPACE_OTHER
:
1670 ARCSTAT_INCR(arcstat_other_size
, space
);
1672 case ARC_SPACE_HDRS
:
1673 ARCSTAT_INCR(arcstat_hdr_size
, space
);
1675 case ARC_SPACE_L2HDRS
:
1676 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
1680 if (type
!= ARC_SPACE_DATA
)
1681 ARCSTAT_INCR(arcstat_meta_used
, space
);
1683 atomic_add_64(&arc_size
, space
);
1687 arc_space_return(uint64_t space
, arc_space_type_t type
)
1689 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1694 case ARC_SPACE_DATA
:
1695 ARCSTAT_INCR(arcstat_data_size
, -space
);
1697 case ARC_SPACE_META
:
1698 ARCSTAT_INCR(arcstat_metadata_size
, -space
);
1700 case ARC_SPACE_OTHER
:
1701 ARCSTAT_INCR(arcstat_other_size
, -space
);
1703 case ARC_SPACE_HDRS
:
1704 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
1706 case ARC_SPACE_L2HDRS
:
1707 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
1711 if (type
!= ARC_SPACE_DATA
) {
1712 ASSERT(arc_meta_used
>= space
);
1713 if (arc_meta_max
< arc_meta_used
)
1714 arc_meta_max
= arc_meta_used
;
1715 ARCSTAT_INCR(arcstat_meta_used
, -space
);
1718 ASSERT(arc_size
>= space
);
1719 atomic_add_64(&arc_size
, -space
);
1723 arc_buf_alloc(spa_t
*spa
, uint64_t size
, void *tag
, arc_buf_contents_t type
)
1728 VERIFY3U(size
, <=, spa_maxblocksize(spa
));
1729 hdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
1730 ASSERT(BUF_EMPTY(hdr
));
1731 ASSERT3P(hdr
->b_freeze_cksum
, ==, NULL
);
1733 hdr
->b_spa
= spa_load_guid(spa
);
1734 hdr
->b_l1hdr
.b_mru_hits
= 0;
1735 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
1736 hdr
->b_l1hdr
.b_mfu_hits
= 0;
1737 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
1738 hdr
->b_l1hdr
.b_l2_hits
= 0;
1740 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1743 buf
->b_efunc
= NULL
;
1744 buf
->b_private
= NULL
;
1747 hdr
->b_flags
= arc_bufc_to_flags(type
);
1748 hdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
1750 hdr
->b_l1hdr
.b_buf
= buf
;
1751 hdr
->b_l1hdr
.b_state
= arc_anon
;
1752 hdr
->b_l1hdr
.b_arc_access
= 0;
1753 hdr
->b_l1hdr
.b_datacnt
= 1;
1754 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1756 arc_get_data_buf(buf
);
1757 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
1758 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1763 static char *arc_onloan_tag
= "onloan";
1766 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1767 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1768 * buffers must be returned to the arc before they can be used by the DMU or
1772 arc_loan_buf(spa_t
*spa
, uint64_t size
)
1776 buf
= arc_buf_alloc(spa
, size
, arc_onloan_tag
, ARC_BUFC_DATA
);
1778 atomic_add_64(&arc_loaned_bytes
, size
);
1783 * Return a loaned arc buffer to the arc.
1786 arc_return_buf(arc_buf_t
*buf
, void *tag
)
1788 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1790 ASSERT(buf
->b_data
!= NULL
);
1791 ASSERT(HDR_HAS_L1HDR(hdr
));
1792 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1793 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
1795 atomic_add_64(&arc_loaned_bytes
, -hdr
->b_size
);
1798 /* Detach an arc_buf from a dbuf (tag) */
1800 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
1802 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1804 ASSERT(buf
->b_data
!= NULL
);
1805 ASSERT(HDR_HAS_L1HDR(hdr
));
1806 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
1807 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1808 buf
->b_efunc
= NULL
;
1809 buf
->b_private
= NULL
;
1811 atomic_add_64(&arc_loaned_bytes
, hdr
->b_size
);
1815 arc_buf_clone(arc_buf_t
*from
)
1818 arc_buf_hdr_t
*hdr
= from
->b_hdr
;
1819 uint64_t size
= hdr
->b_size
;
1821 ASSERT(HDR_HAS_L1HDR(hdr
));
1822 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_anon
);
1824 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1827 buf
->b_efunc
= NULL
;
1828 buf
->b_private
= NULL
;
1829 buf
->b_next
= hdr
->b_l1hdr
.b_buf
;
1830 hdr
->b_l1hdr
.b_buf
= buf
;
1831 arc_get_data_buf(buf
);
1832 bcopy(from
->b_data
, buf
->b_data
, size
);
1835 * This buffer already exists in the arc so create a duplicate
1836 * copy for the caller. If the buffer is associated with user data
1837 * then track the size and number of duplicates. These stats will be
1838 * updated as duplicate buffers are created and destroyed.
1840 if (HDR_ISTYPE_DATA(hdr
)) {
1841 ARCSTAT_BUMP(arcstat_duplicate_buffers
);
1842 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, size
);
1844 hdr
->b_l1hdr
.b_datacnt
+= 1;
1849 arc_buf_add_ref(arc_buf_t
*buf
, void* tag
)
1852 kmutex_t
*hash_lock
;
1855 * Check to see if this buffer is evicted. Callers
1856 * must verify b_data != NULL to know if the add_ref
1859 mutex_enter(&buf
->b_evict_lock
);
1860 if (buf
->b_data
== NULL
) {
1861 mutex_exit(&buf
->b_evict_lock
);
1864 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1865 mutex_enter(hash_lock
);
1867 ASSERT(HDR_HAS_L1HDR(hdr
));
1868 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1869 mutex_exit(&buf
->b_evict_lock
);
1871 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
1872 hdr
->b_l1hdr
.b_state
== arc_mfu
);
1874 add_reference(hdr
, hash_lock
, tag
);
1875 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
1876 arc_access(hdr
, hash_lock
);
1877 mutex_exit(hash_lock
);
1878 ARCSTAT_BUMP(arcstat_hits
);
1879 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
1880 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
1881 data
, metadata
, hits
);
1885 arc_buf_free_on_write(void *data
, size_t size
,
1886 void (*free_func
)(void *, size_t))
1888 l2arc_data_free_t
*df
;
1890 df
= kmem_alloc(sizeof (*df
), KM_SLEEP
);
1891 df
->l2df_data
= data
;
1892 df
->l2df_size
= size
;
1893 df
->l2df_func
= free_func
;
1894 mutex_enter(&l2arc_free_on_write_mtx
);
1895 list_insert_head(l2arc_free_on_write
, df
);
1896 mutex_exit(&l2arc_free_on_write_mtx
);
1900 * Free the arc data buffer. If it is an l2arc write in progress,
1901 * the buffer is placed on l2arc_free_on_write to be freed later.
1904 arc_buf_data_free(arc_buf_t
*buf
, void (*free_func
)(void *, size_t))
1906 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1908 if (HDR_L2_WRITING(hdr
)) {
1909 arc_buf_free_on_write(buf
->b_data
, hdr
->b_size
, free_func
);
1910 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
1912 free_func(buf
->b_data
, hdr
->b_size
);
1917 arc_buf_l2_cdata_free(arc_buf_hdr_t
*hdr
)
1919 ASSERT(HDR_HAS_L2HDR(hdr
));
1920 ASSERT(MUTEX_HELD(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
));
1923 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
1924 * that doesn't exist, the header is in the arc_l2c_only state,
1925 * and there isn't anything to free (it's already been freed).
1927 if (!HDR_HAS_L1HDR(hdr
))
1931 * The header isn't being written to the l2arc device, thus it
1932 * shouldn't have a b_tmp_cdata to free.
1934 if (!HDR_L2_WRITING(hdr
)) {
1935 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1940 * The header does not have compression enabled. This can be due
1941 * to the buffer not being compressible, or because we're
1942 * freeing the buffer before the second phase of
1943 * l2arc_write_buffer() has started (which does the compression
1944 * step). In either case, b_tmp_cdata does not point to a
1945 * separately compressed buffer, so there's nothing to free (it
1946 * points to the same buffer as the arc_buf_t's b_data field).
1948 if (hdr
->b_l2hdr
.b_compress
== ZIO_COMPRESS_OFF
) {
1949 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1954 * There's nothing to free since the buffer was all zero's and
1955 * compressed to a zero length buffer.
1957 if (hdr
->b_l2hdr
.b_compress
== ZIO_COMPRESS_EMPTY
) {
1958 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1962 ASSERT(L2ARC_IS_VALID_COMPRESS(hdr
->b_l2hdr
.b_compress
));
1964 arc_buf_free_on_write(hdr
->b_l1hdr
.b_tmp_cdata
,
1965 hdr
->b_size
, zio_data_buf_free
);
1967 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write
);
1968 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1972 * Free up buf->b_data and if 'remove' is set, then pull the
1973 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
1976 arc_buf_destroy(arc_buf_t
*buf
, boolean_t remove
)
1980 /* free up data associated with the buf */
1981 if (buf
->b_data
!= NULL
) {
1982 arc_state_t
*state
= buf
->b_hdr
->b_l1hdr
.b_state
;
1983 uint64_t size
= buf
->b_hdr
->b_size
;
1984 arc_buf_contents_t type
= arc_buf_type(buf
->b_hdr
);
1986 arc_cksum_verify(buf
);
1987 arc_buf_unwatch(buf
);
1989 if (type
== ARC_BUFC_METADATA
) {
1990 arc_buf_data_free(buf
, zio_buf_free
);
1991 arc_space_return(size
, ARC_SPACE_META
);
1993 ASSERT(type
== ARC_BUFC_DATA
);
1994 arc_buf_data_free(buf
, zio_data_buf_free
);
1995 arc_space_return(size
, ARC_SPACE_DATA
);
1998 /* protected by hash lock, if in the hash table */
1999 if (multilist_link_active(&buf
->b_hdr
->b_l1hdr
.b_arc_node
)) {
2000 uint64_t *cnt
= &state
->arcs_lsize
[type
];
2002 ASSERT(refcount_is_zero(
2003 &buf
->b_hdr
->b_l1hdr
.b_refcnt
));
2004 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
2006 ASSERT3U(*cnt
, >=, size
);
2007 atomic_add_64(cnt
, -size
);
2010 (void) refcount_remove_many(&state
->arcs_size
, size
, buf
);
2014 * If we're destroying a duplicate buffer make sure
2015 * that the appropriate statistics are updated.
2017 if (buf
->b_hdr
->b_l1hdr
.b_datacnt
> 1 &&
2018 HDR_ISTYPE_DATA(buf
->b_hdr
)) {
2019 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
2020 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, -size
);
2022 ASSERT(buf
->b_hdr
->b_l1hdr
.b_datacnt
> 0);
2023 buf
->b_hdr
->b_l1hdr
.b_datacnt
-= 1;
2026 /* only remove the buf if requested */
2030 /* remove the buf from the hdr list */
2031 for (bufp
= &buf
->b_hdr
->b_l1hdr
.b_buf
; *bufp
!= buf
;
2032 bufp
= &(*bufp
)->b_next
)
2034 *bufp
= buf
->b_next
;
2037 ASSERT(buf
->b_efunc
== NULL
);
2039 /* clean up the buf */
2041 kmem_cache_free(buf_cache
, buf
);
2045 arc_hdr_l2hdr_destroy(arc_buf_hdr_t
*hdr
)
2047 l2arc_buf_hdr_t
*l2hdr
= &hdr
->b_l2hdr
;
2048 l2arc_dev_t
*dev
= l2hdr
->b_dev
;
2050 ASSERT(MUTEX_HELD(&dev
->l2ad_mtx
));
2051 ASSERT(HDR_HAS_L2HDR(hdr
));
2053 list_remove(&dev
->l2ad_buflist
, hdr
);
2056 * We don't want to leak the b_tmp_cdata buffer that was
2057 * allocated in l2arc_write_buffers()
2059 arc_buf_l2_cdata_free(hdr
);
2062 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2063 * this header is being processed by l2arc_write_buffers() (i.e.
2064 * it's in the first stage of l2arc_write_buffers()).
2065 * Re-affirming that truth here, just to serve as a reminder. If
2066 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2067 * may not have its HDR_L2_WRITING flag set. (the write may have
2068 * completed, in which case HDR_L2_WRITING will be false and the
2069 * b_daddr field will point to the address of the buffer on disk).
2071 IMPLY(l2hdr
->b_daddr
== L2ARC_ADDR_UNSET
, HDR_L2_WRITING(hdr
));
2074 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2075 * l2arc_write_buffers(). Since we've just removed this header
2076 * from the l2arc buffer list, this header will never reach the
2077 * second stage of l2arc_write_buffers(), which increments the
2078 * accounting stats for this header. Thus, we must be careful
2079 * not to decrement them for this header either.
2081 if (l2hdr
->b_daddr
!= L2ARC_ADDR_UNSET
) {
2082 ARCSTAT_INCR(arcstat_l2_asize
, -l2hdr
->b_asize
);
2083 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
2085 vdev_space_update(dev
->l2ad_vdev
,
2086 -l2hdr
->b_asize
, 0, 0);
2088 (void) refcount_remove_many(&dev
->l2ad_alloc
,
2089 l2hdr
->b_asize
, hdr
);
2092 hdr
->b_flags
&= ~ARC_FLAG_HAS_L2HDR
;
2096 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
2098 if (HDR_HAS_L1HDR(hdr
)) {
2099 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
||
2100 hdr
->b_l1hdr
.b_datacnt
> 0);
2101 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2102 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
2104 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
2105 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
2107 if (HDR_HAS_L2HDR(hdr
)) {
2108 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
2109 boolean_t buflist_held
= MUTEX_HELD(&dev
->l2ad_mtx
);
2112 mutex_enter(&dev
->l2ad_mtx
);
2115 * Even though we checked this conditional above, we
2116 * need to check this again now that we have the
2117 * l2ad_mtx. This is because we could be racing with
2118 * another thread calling l2arc_evict() which might have
2119 * destroyed this header's L2 portion as we were waiting
2120 * to acquire the l2ad_mtx. If that happens, we don't
2121 * want to re-destroy the header's L2 portion.
2123 if (HDR_HAS_L2HDR(hdr
))
2124 arc_hdr_l2hdr_destroy(hdr
);
2127 mutex_exit(&dev
->l2ad_mtx
);
2130 if (!BUF_EMPTY(hdr
))
2131 buf_discard_identity(hdr
);
2133 if (hdr
->b_freeze_cksum
!= NULL
) {
2134 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
2135 hdr
->b_freeze_cksum
= NULL
;
2138 if (HDR_HAS_L1HDR(hdr
)) {
2139 while (hdr
->b_l1hdr
.b_buf
) {
2140 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
2142 if (buf
->b_efunc
!= NULL
) {
2143 mutex_enter(&arc_user_evicts_lock
);
2144 mutex_enter(&buf
->b_evict_lock
);
2145 ASSERT(buf
->b_hdr
!= NULL
);
2146 arc_buf_destroy(hdr
->b_l1hdr
.b_buf
, FALSE
);
2147 hdr
->b_l1hdr
.b_buf
= buf
->b_next
;
2148 buf
->b_hdr
= &arc_eviction_hdr
;
2149 buf
->b_next
= arc_eviction_list
;
2150 arc_eviction_list
= buf
;
2151 mutex_exit(&buf
->b_evict_lock
);
2152 cv_signal(&arc_user_evicts_cv
);
2153 mutex_exit(&arc_user_evicts_lock
);
2155 arc_buf_destroy(hdr
->b_l1hdr
.b_buf
, TRUE
);
2160 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
2161 if (HDR_HAS_L1HDR(hdr
)) {
2162 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
2163 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
2164 kmem_cache_free(hdr_full_cache
, hdr
);
2166 kmem_cache_free(hdr_l2only_cache
, hdr
);
2171 arc_buf_free(arc_buf_t
*buf
, void *tag
)
2173 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2174 int hashed
= hdr
->b_l1hdr
.b_state
!= arc_anon
;
2176 ASSERT(buf
->b_efunc
== NULL
);
2177 ASSERT(buf
->b_data
!= NULL
);
2180 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
2182 mutex_enter(hash_lock
);
2184 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
2186 (void) remove_reference(hdr
, hash_lock
, tag
);
2187 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
2188 arc_buf_destroy(buf
, TRUE
);
2190 ASSERT(buf
== hdr
->b_l1hdr
.b_buf
);
2191 ASSERT(buf
->b_efunc
== NULL
);
2192 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
2194 mutex_exit(hash_lock
);
2195 } else if (HDR_IO_IN_PROGRESS(hdr
)) {
2198 * We are in the middle of an async write. Don't destroy
2199 * this buffer unless the write completes before we finish
2200 * decrementing the reference count.
2202 mutex_enter(&arc_user_evicts_lock
);
2203 (void) remove_reference(hdr
, NULL
, tag
);
2204 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2205 destroy_hdr
= !HDR_IO_IN_PROGRESS(hdr
);
2206 mutex_exit(&arc_user_evicts_lock
);
2208 arc_hdr_destroy(hdr
);
2210 if (remove_reference(hdr
, NULL
, tag
) > 0)
2211 arc_buf_destroy(buf
, TRUE
);
2213 arc_hdr_destroy(hdr
);
2218 arc_buf_remove_ref(arc_buf_t
*buf
, void* tag
)
2220 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2221 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
2222 boolean_t no_callback
= (buf
->b_efunc
== NULL
);
2224 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
2225 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
2226 arc_buf_free(buf
, tag
);
2227 return (no_callback
);
2230 mutex_enter(hash_lock
);
2232 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
2233 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
2234 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_anon
);
2235 ASSERT(buf
->b_data
!= NULL
);
2237 (void) remove_reference(hdr
, hash_lock
, tag
);
2238 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
2240 arc_buf_destroy(buf
, TRUE
);
2241 } else if (no_callback
) {
2242 ASSERT(hdr
->b_l1hdr
.b_buf
== buf
&& buf
->b_next
== NULL
);
2243 ASSERT(buf
->b_efunc
== NULL
);
2244 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
2246 ASSERT(no_callback
|| hdr
->b_l1hdr
.b_datacnt
> 1 ||
2247 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2248 mutex_exit(hash_lock
);
2249 return (no_callback
);
2253 arc_buf_size(arc_buf_t
*buf
)
2255 return (buf
->b_hdr
->b_size
);
2259 * Called from the DMU to determine if the current buffer should be
2260 * evicted. In order to ensure proper locking, the eviction must be initiated
2261 * from the DMU. Return true if the buffer is associated with user data and
2262 * duplicate buffers still exist.
2265 arc_buf_eviction_needed(arc_buf_t
*buf
)
2268 boolean_t evict_needed
= B_FALSE
;
2270 if (zfs_disable_dup_eviction
)
2273 mutex_enter(&buf
->b_evict_lock
);
2277 * We are in arc_do_user_evicts(); let that function
2278 * perform the eviction.
2280 ASSERT(buf
->b_data
== NULL
);
2281 mutex_exit(&buf
->b_evict_lock
);
2283 } else if (buf
->b_data
== NULL
) {
2285 * We have already been added to the arc eviction list;
2286 * recommend eviction.
2288 ASSERT3P(hdr
, ==, &arc_eviction_hdr
);
2289 mutex_exit(&buf
->b_evict_lock
);
2293 if (hdr
->b_l1hdr
.b_datacnt
> 1 && HDR_ISTYPE_DATA(hdr
))
2294 evict_needed
= B_TRUE
;
2296 mutex_exit(&buf
->b_evict_lock
);
2297 return (evict_needed
);
2301 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2302 * state of the header is dependent on its state prior to entering this
2303 * function. The following transitions are possible:
2305 * - arc_mru -> arc_mru_ghost
2306 * - arc_mfu -> arc_mfu_ghost
2307 * - arc_mru_ghost -> arc_l2c_only
2308 * - arc_mru_ghost -> deleted
2309 * - arc_mfu_ghost -> arc_l2c_only
2310 * - arc_mfu_ghost -> deleted
2313 arc_evict_hdr(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
2315 arc_state_t
*evicted_state
, *state
;
2316 int64_t bytes_evicted
= 0;
2318 ASSERT(MUTEX_HELD(hash_lock
));
2319 ASSERT(HDR_HAS_L1HDR(hdr
));
2321 state
= hdr
->b_l1hdr
.b_state
;
2322 if (GHOST_STATE(state
)) {
2323 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
2324 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
2327 * l2arc_write_buffers() relies on a header's L1 portion
2328 * (i.e. its b_tmp_cdata field) during its write phase.
2329 * Thus, we cannot push a header onto the arc_l2c_only
2330 * state (removing its L1 piece) until the header is
2331 * done being written to the l2arc.
2333 if (HDR_HAS_L2HDR(hdr
) && HDR_L2_WRITING(hdr
)) {
2334 ARCSTAT_BUMP(arcstat_evict_l2_skip
);
2335 return (bytes_evicted
);
2338 ARCSTAT_BUMP(arcstat_deleted
);
2339 bytes_evicted
+= hdr
->b_size
;
2341 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, hdr
);
2343 if (HDR_HAS_L2HDR(hdr
)) {
2345 * This buffer is cached on the 2nd Level ARC;
2346 * don't destroy the header.
2348 arc_change_state(arc_l2c_only
, hdr
, hash_lock
);
2350 * dropping from L1+L2 cached to L2-only,
2351 * realloc to remove the L1 header.
2353 hdr
= arc_hdr_realloc(hdr
, hdr_full_cache
,
2356 arc_change_state(arc_anon
, hdr
, hash_lock
);
2357 arc_hdr_destroy(hdr
);
2359 return (bytes_evicted
);
2362 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
2363 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
2365 /* prefetch buffers have a minimum lifespan */
2366 if (HDR_IO_IN_PROGRESS(hdr
) ||
2367 ((hdr
->b_flags
& (ARC_FLAG_PREFETCH
| ARC_FLAG_INDIRECT
)) &&
2368 ddi_get_lbolt() - hdr
->b_l1hdr
.b_arc_access
<
2369 arc_min_prefetch_lifespan
)) {
2370 ARCSTAT_BUMP(arcstat_evict_skip
);
2371 return (bytes_evicted
);
2374 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
2375 ASSERT3U(hdr
->b_l1hdr
.b_datacnt
, >, 0);
2376 while (hdr
->b_l1hdr
.b_buf
) {
2377 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
2378 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
2379 ARCSTAT_BUMP(arcstat_mutex_miss
);
2382 if (buf
->b_data
!= NULL
)
2383 bytes_evicted
+= hdr
->b_size
;
2384 if (buf
->b_efunc
!= NULL
) {
2385 mutex_enter(&arc_user_evicts_lock
);
2386 arc_buf_destroy(buf
, FALSE
);
2387 hdr
->b_l1hdr
.b_buf
= buf
->b_next
;
2388 buf
->b_hdr
= &arc_eviction_hdr
;
2389 buf
->b_next
= arc_eviction_list
;
2390 arc_eviction_list
= buf
;
2391 cv_signal(&arc_user_evicts_cv
);
2392 mutex_exit(&arc_user_evicts_lock
);
2393 mutex_exit(&buf
->b_evict_lock
);
2395 mutex_exit(&buf
->b_evict_lock
);
2396 arc_buf_destroy(buf
, TRUE
);
2400 if (HDR_HAS_L2HDR(hdr
)) {
2401 ARCSTAT_INCR(arcstat_evict_l2_cached
, hdr
->b_size
);
2403 if (l2arc_write_eligible(hdr
->b_spa
, hdr
))
2404 ARCSTAT_INCR(arcstat_evict_l2_eligible
, hdr
->b_size
);
2406 ARCSTAT_INCR(arcstat_evict_l2_ineligible
, hdr
->b_size
);
2409 if (hdr
->b_l1hdr
.b_datacnt
== 0) {
2410 arc_change_state(evicted_state
, hdr
, hash_lock
);
2411 ASSERT(HDR_IN_HASH_TABLE(hdr
));
2412 hdr
->b_flags
|= ARC_FLAG_IN_HASH_TABLE
;
2413 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
2414 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, hdr
);
2417 return (bytes_evicted
);
2421 arc_evict_state_impl(multilist_t
*ml
, int idx
, arc_buf_hdr_t
*marker
,
2422 uint64_t spa
, int64_t bytes
)
2424 multilist_sublist_t
*mls
;
2425 uint64_t bytes_evicted
= 0;
2427 kmutex_t
*hash_lock
;
2428 int evict_count
= 0;
2430 ASSERT3P(marker
, !=, NULL
);
2431 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
2433 mls
= multilist_sublist_lock(ml
, idx
);
2435 for (hdr
= multilist_sublist_prev(mls
, marker
); hdr
!= NULL
;
2436 hdr
= multilist_sublist_prev(mls
, marker
)) {
2437 if ((bytes
!= ARC_EVICT_ALL
&& bytes_evicted
>= bytes
) ||
2438 (evict_count
>= zfs_arc_evict_batch_limit
))
2442 * To keep our iteration location, move the marker
2443 * forward. Since we're not holding hdr's hash lock, we
2444 * must be very careful and not remove 'hdr' from the
2445 * sublist. Otherwise, other consumers might mistake the
2446 * 'hdr' as not being on a sublist when they call the
2447 * multilist_link_active() function (they all rely on
2448 * the hash lock protecting concurrent insertions and
2449 * removals). multilist_sublist_move_forward() was
2450 * specifically implemented to ensure this is the case
2451 * (only 'marker' will be removed and re-inserted).
2453 multilist_sublist_move_forward(mls
, marker
);
2456 * The only case where the b_spa field should ever be
2457 * zero, is the marker headers inserted by
2458 * arc_evict_state(). It's possible for multiple threads
2459 * to be calling arc_evict_state() concurrently (e.g.
2460 * dsl_pool_close() and zio_inject_fault()), so we must
2461 * skip any markers we see from these other threads.
2463 if (hdr
->b_spa
== 0)
2466 /* we're only interested in evicting buffers of a certain spa */
2467 if (spa
!= 0 && hdr
->b_spa
!= spa
) {
2468 ARCSTAT_BUMP(arcstat_evict_skip
);
2472 hash_lock
= HDR_LOCK(hdr
);
2475 * We aren't calling this function from any code path
2476 * that would already be holding a hash lock, so we're
2477 * asserting on this assumption to be defensive in case
2478 * this ever changes. Without this check, it would be
2479 * possible to incorrectly increment arcstat_mutex_miss
2480 * below (e.g. if the code changed such that we called
2481 * this function with a hash lock held).
2483 ASSERT(!MUTEX_HELD(hash_lock
));
2485 if (mutex_tryenter(hash_lock
)) {
2486 uint64_t evicted
= arc_evict_hdr(hdr
, hash_lock
);
2487 mutex_exit(hash_lock
);
2489 bytes_evicted
+= evicted
;
2492 * If evicted is zero, arc_evict_hdr() must have
2493 * decided to skip this header, don't increment
2494 * evict_count in this case.
2500 * If arc_size isn't overflowing, signal any
2501 * threads that might happen to be waiting.
2503 * For each header evicted, we wake up a single
2504 * thread. If we used cv_broadcast, we could
2505 * wake up "too many" threads causing arc_size
2506 * to significantly overflow arc_c; since
2507 * arc_get_data_buf() doesn't check for overflow
2508 * when it's woken up (it doesn't because it's
2509 * possible for the ARC to be overflowing while
2510 * full of un-evictable buffers, and the
2511 * function should proceed in this case).
2513 * If threads are left sleeping, due to not
2514 * using cv_broadcast, they will be woken up
2515 * just before arc_reclaim_thread() sleeps.
2517 mutex_enter(&arc_reclaim_lock
);
2518 if (!arc_is_overflowing())
2519 cv_signal(&arc_reclaim_waiters_cv
);
2520 mutex_exit(&arc_reclaim_lock
);
2522 ARCSTAT_BUMP(arcstat_mutex_miss
);
2526 multilist_sublist_unlock(mls
);
2528 return (bytes_evicted
);
2532 * Evict buffers from the given arc state, until we've removed the
2533 * specified number of bytes. Move the removed buffers to the
2534 * appropriate evict state.
2536 * This function makes a "best effort". It skips over any buffers
2537 * it can't get a hash_lock on, and so, may not catch all candidates.
2538 * It may also return without evicting as much space as requested.
2540 * If bytes is specified using the special value ARC_EVICT_ALL, this
2541 * will evict all available (i.e. unlocked and evictable) buffers from
2542 * the given arc state; which is used by arc_flush().
2545 arc_evict_state(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
2546 arc_buf_contents_t type
)
2548 uint64_t total_evicted
= 0;
2549 multilist_t
*ml
= &state
->arcs_list
[type
];
2551 arc_buf_hdr_t
**markers
;
2554 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
2556 num_sublists
= multilist_get_num_sublists(ml
);
2559 * If we've tried to evict from each sublist, made some
2560 * progress, but still have not hit the target number of bytes
2561 * to evict, we want to keep trying. The markers allow us to
2562 * pick up where we left off for each individual sublist, rather
2563 * than starting from the tail each time.
2565 markers
= kmem_zalloc(sizeof (*markers
) * num_sublists
, KM_SLEEP
);
2566 for (i
= 0; i
< num_sublists
; i
++) {
2567 multilist_sublist_t
*mls
;
2569 markers
[i
] = kmem_cache_alloc(hdr_full_cache
, KM_SLEEP
);
2572 * A b_spa of 0 is used to indicate that this header is
2573 * a marker. This fact is used in arc_adjust_type() and
2574 * arc_evict_state_impl().
2576 markers
[i
]->b_spa
= 0;
2578 mls
= multilist_sublist_lock(ml
, i
);
2579 multilist_sublist_insert_tail(mls
, markers
[i
]);
2580 multilist_sublist_unlock(mls
);
2584 * While we haven't hit our target number of bytes to evict, or
2585 * we're evicting all available buffers.
2587 while (total_evicted
< bytes
|| bytes
== ARC_EVICT_ALL
) {
2589 * Start eviction using a randomly selected sublist,
2590 * this is to try and evenly balance eviction across all
2591 * sublists. Always starting at the same sublist
2592 * (e.g. index 0) would cause evictions to favor certain
2593 * sublists over others.
2595 int sublist_idx
= multilist_get_random_index(ml
);
2596 uint64_t scan_evicted
= 0;
2598 for (i
= 0; i
< num_sublists
; i
++) {
2599 uint64_t bytes_remaining
;
2600 uint64_t bytes_evicted
;
2602 if (bytes
== ARC_EVICT_ALL
)
2603 bytes_remaining
= ARC_EVICT_ALL
;
2604 else if (total_evicted
< bytes
)
2605 bytes_remaining
= bytes
- total_evicted
;
2609 bytes_evicted
= arc_evict_state_impl(ml
, sublist_idx
,
2610 markers
[sublist_idx
], spa
, bytes_remaining
);
2612 scan_evicted
+= bytes_evicted
;
2613 total_evicted
+= bytes_evicted
;
2615 /* we've reached the end, wrap to the beginning */
2616 if (++sublist_idx
>= num_sublists
)
2621 * If we didn't evict anything during this scan, we have
2622 * no reason to believe we'll evict more during another
2623 * scan, so break the loop.
2625 if (scan_evicted
== 0) {
2626 /* This isn't possible, let's make that obvious */
2627 ASSERT3S(bytes
, !=, 0);
2630 * When bytes is ARC_EVICT_ALL, the only way to
2631 * break the loop is when scan_evicted is zero.
2632 * In that case, we actually have evicted enough,
2633 * so we don't want to increment the kstat.
2635 if (bytes
!= ARC_EVICT_ALL
) {
2636 ASSERT3S(total_evicted
, <, bytes
);
2637 ARCSTAT_BUMP(arcstat_evict_not_enough
);
2644 for (i
= 0; i
< num_sublists
; i
++) {
2645 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
2646 multilist_sublist_remove(mls
, markers
[i
]);
2647 multilist_sublist_unlock(mls
);
2649 kmem_cache_free(hdr_full_cache
, markers
[i
]);
2651 kmem_free(markers
, sizeof (*markers
) * num_sublists
);
2653 return (total_evicted
);
2657 * Flush all "evictable" data of the given type from the arc state
2658 * specified. This will not evict any "active" buffers (i.e. referenced).
2660 * When 'retry' is set to FALSE, the function will make a single pass
2661 * over the state and evict any buffers that it can. Since it doesn't
2662 * continually retry the eviction, it might end up leaving some buffers
2663 * in the ARC due to lock misses.
2665 * When 'retry' is set to TRUE, the function will continually retry the
2666 * eviction until *all* evictable buffers have been removed from the
2667 * state. As a result, if concurrent insertions into the state are
2668 * allowed (e.g. if the ARC isn't shutting down), this function might
2669 * wind up in an infinite loop, continually trying to evict buffers.
2672 arc_flush_state(arc_state_t
*state
, uint64_t spa
, arc_buf_contents_t type
,
2675 uint64_t evicted
= 0;
2677 while (state
->arcs_lsize
[type
] != 0) {
2678 evicted
+= arc_evict_state(state
, spa
, ARC_EVICT_ALL
, type
);
2688 * Helper function for arc_prune_async() it is responsible for safely
2689 * handling the execution of a registered arc_prune_func_t.
2692 arc_prune_task(void *ptr
)
2694 arc_prune_t
*ap
= (arc_prune_t
*)ptr
;
2695 arc_prune_func_t
*func
= ap
->p_pfunc
;
2698 func(ap
->p_adjust
, ap
->p_private
);
2700 /* Callback unregistered concurrently with execution */
2701 if (refcount_remove(&ap
->p_refcnt
, func
) == 0) {
2702 ASSERT(!list_link_active(&ap
->p_node
));
2703 refcount_destroy(&ap
->p_refcnt
);
2704 kmem_free(ap
, sizeof (*ap
));
2709 * Notify registered consumers they must drop holds on a portion of the ARC
2710 * buffered they reference. This provides a mechanism to ensure the ARC can
2711 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
2712 * is analogous to dnlc_reduce_cache() but more generic.
2714 * This operation is performed asynchronously so it may be safely called
2715 * in the context of the arc_reclaim_thread(). A reference is taken here
2716 * for each registered arc_prune_t and the arc_prune_task() is responsible
2717 * for releasing it once the registered arc_prune_func_t has completed.
2720 arc_prune_async(int64_t adjust
)
2724 mutex_enter(&arc_prune_mtx
);
2725 for (ap
= list_head(&arc_prune_list
); ap
!= NULL
;
2726 ap
= list_next(&arc_prune_list
, ap
)) {
2728 if (refcount_count(&ap
->p_refcnt
) >= 2)
2731 refcount_add(&ap
->p_refcnt
, ap
->p_pfunc
);
2732 ap
->p_adjust
= adjust
;
2733 taskq_dispatch(arc_prune_taskq
, arc_prune_task
, ap
, TQ_SLEEP
);
2734 ARCSTAT_BUMP(arcstat_prune
);
2736 mutex_exit(&arc_prune_mtx
);
2740 * Evict the specified number of bytes from the state specified,
2741 * restricting eviction to the spa and type given. This function
2742 * prevents us from trying to evict more from a state's list than
2743 * is "evictable", and to skip evicting altogether when passed a
2744 * negative value for "bytes". In contrast, arc_evict_state() will
2745 * evict everything it can, when passed a negative value for "bytes".
2748 arc_adjust_impl(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
2749 arc_buf_contents_t type
)
2753 if (bytes
> 0 && state
->arcs_lsize
[type
] > 0) {
2754 delta
= MIN(state
->arcs_lsize
[type
], bytes
);
2755 return (arc_evict_state(state
, spa
, delta
, type
));
2762 * The goal of this function is to evict enough meta data buffers from the
2763 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
2764 * more complicated than it appears because it is common for data buffers
2765 * to have holds on meta data buffers. In addition, dnode meta data buffers
2766 * will be held by the dnodes in the block preventing them from being freed.
2767 * This means we can't simply traverse the ARC and expect to always find
2768 * enough unheld meta data buffer to release.
2770 * Therefore, this function has been updated to make alternating passes
2771 * over the ARC releasing data buffers and then newly unheld meta data
2772 * buffers. This ensures forward progress is maintained and arc_meta_used
2773 * will decrease. Normally this is sufficient, but if required the ARC
2774 * will call the registered prune callbacks causing dentry and inodes to
2775 * be dropped from the VFS cache. This will make dnode meta data buffers
2776 * available for reclaim.
2779 arc_adjust_meta_balanced(void)
2781 int64_t adjustmnt
, delta
, prune
= 0;
2782 uint64_t total_evicted
= 0;
2783 arc_buf_contents_t type
= ARC_BUFC_DATA
;
2784 int restarts
= MAX(zfs_arc_meta_adjust_restarts
, 0);
2788 * This slightly differs than the way we evict from the mru in
2789 * arc_adjust because we don't have a "target" value (i.e. no
2790 * "meta" arc_p). As a result, I think we can completely
2791 * cannibalize the metadata in the MRU before we evict the
2792 * metadata from the MFU. I think we probably need to implement a
2793 * "metadata arc_p" value to do this properly.
2795 adjustmnt
= arc_meta_used
- arc_meta_limit
;
2797 if (adjustmnt
> 0 && arc_mru
->arcs_lsize
[type
] > 0) {
2798 delta
= MIN(arc_mru
->arcs_lsize
[type
], adjustmnt
);
2799 total_evicted
+= arc_adjust_impl(arc_mru
, 0, delta
, type
);
2804 * We can't afford to recalculate adjustmnt here. If we do,
2805 * new metadata buffers can sneak into the MRU or ANON lists,
2806 * thus penalize the MFU metadata. Although the fudge factor is
2807 * small, it has been empirically shown to be significant for
2808 * certain workloads (e.g. creating many empty directories). As
2809 * such, we use the original calculation for adjustmnt, and
2810 * simply decrement the amount of data evicted from the MRU.
2813 if (adjustmnt
> 0 && arc_mfu
->arcs_lsize
[type
] > 0) {
2814 delta
= MIN(arc_mfu
->arcs_lsize
[type
], adjustmnt
);
2815 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, delta
, type
);
2818 adjustmnt
= arc_meta_used
- arc_meta_limit
;
2820 if (adjustmnt
> 0 && arc_mru_ghost
->arcs_lsize
[type
] > 0) {
2821 delta
= MIN(adjustmnt
,
2822 arc_mru_ghost
->arcs_lsize
[type
]);
2823 total_evicted
+= arc_adjust_impl(arc_mru_ghost
, 0, delta
, type
);
2827 if (adjustmnt
> 0 && arc_mfu_ghost
->arcs_lsize
[type
] > 0) {
2828 delta
= MIN(adjustmnt
,
2829 arc_mfu_ghost
->arcs_lsize
[type
]);
2830 total_evicted
+= arc_adjust_impl(arc_mfu_ghost
, 0, delta
, type
);
2834 * If after attempting to make the requested adjustment to the ARC
2835 * the meta limit is still being exceeded then request that the
2836 * higher layers drop some cached objects which have holds on ARC
2837 * meta buffers. Requests to the upper layers will be made with
2838 * increasingly large scan sizes until the ARC is below the limit.
2840 if (arc_meta_used
> arc_meta_limit
) {
2841 if (type
== ARC_BUFC_DATA
) {
2842 type
= ARC_BUFC_METADATA
;
2844 type
= ARC_BUFC_DATA
;
2846 if (zfs_arc_meta_prune
) {
2847 prune
+= zfs_arc_meta_prune
;
2848 arc_prune_async(prune
);
2857 return (total_evicted
);
2861 * Evict metadata buffers from the cache, such that arc_meta_used is
2862 * capped by the arc_meta_limit tunable.
2865 arc_adjust_meta_only(void)
2867 uint64_t total_evicted
= 0;
2871 * If we're over the meta limit, we want to evict enough
2872 * metadata to get back under the meta limit. We don't want to
2873 * evict so much that we drop the MRU below arc_p, though. If
2874 * we're over the meta limit more than we're over arc_p, we
2875 * evict some from the MRU here, and some from the MFU below.
2877 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
2878 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
2879 refcount_count(&arc_mru
->arcs_size
) - arc_p
));
2881 total_evicted
+= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
2884 * Similar to the above, we want to evict enough bytes to get us
2885 * below the meta limit, but not so much as to drop us below the
2886 * space alloted to the MFU (which is defined as arc_c - arc_p).
2888 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
2889 (int64_t)(refcount_count(&arc_mfu
->arcs_size
) - (arc_c
- arc_p
)));
2891 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
2893 return (total_evicted
);
2897 arc_adjust_meta(void)
2899 if (zfs_arc_meta_strategy
== ARC_STRATEGY_META_ONLY
)
2900 return (arc_adjust_meta_only());
2902 return (arc_adjust_meta_balanced());
2906 * Return the type of the oldest buffer in the given arc state
2908 * This function will select a random sublist of type ARC_BUFC_DATA and
2909 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
2910 * is compared, and the type which contains the "older" buffer will be
2913 static arc_buf_contents_t
2914 arc_adjust_type(arc_state_t
*state
)
2916 multilist_t
*data_ml
= &state
->arcs_list
[ARC_BUFC_DATA
];
2917 multilist_t
*meta_ml
= &state
->arcs_list
[ARC_BUFC_METADATA
];
2918 int data_idx
= multilist_get_random_index(data_ml
);
2919 int meta_idx
= multilist_get_random_index(meta_ml
);
2920 multilist_sublist_t
*data_mls
;
2921 multilist_sublist_t
*meta_mls
;
2922 arc_buf_contents_t type
;
2923 arc_buf_hdr_t
*data_hdr
;
2924 arc_buf_hdr_t
*meta_hdr
;
2927 * We keep the sublist lock until we're finished, to prevent
2928 * the headers from being destroyed via arc_evict_state().
2930 data_mls
= multilist_sublist_lock(data_ml
, data_idx
);
2931 meta_mls
= multilist_sublist_lock(meta_ml
, meta_idx
);
2934 * These two loops are to ensure we skip any markers that
2935 * might be at the tail of the lists due to arc_evict_state().
2938 for (data_hdr
= multilist_sublist_tail(data_mls
); data_hdr
!= NULL
;
2939 data_hdr
= multilist_sublist_prev(data_mls
, data_hdr
)) {
2940 if (data_hdr
->b_spa
!= 0)
2944 for (meta_hdr
= multilist_sublist_tail(meta_mls
); meta_hdr
!= NULL
;
2945 meta_hdr
= multilist_sublist_prev(meta_mls
, meta_hdr
)) {
2946 if (meta_hdr
->b_spa
!= 0)
2950 if (data_hdr
== NULL
&& meta_hdr
== NULL
) {
2951 type
= ARC_BUFC_DATA
;
2952 } else if (data_hdr
== NULL
) {
2953 ASSERT3P(meta_hdr
, !=, NULL
);
2954 type
= ARC_BUFC_METADATA
;
2955 } else if (meta_hdr
== NULL
) {
2956 ASSERT3P(data_hdr
, !=, NULL
);
2957 type
= ARC_BUFC_DATA
;
2959 ASSERT3P(data_hdr
, !=, NULL
);
2960 ASSERT3P(meta_hdr
, !=, NULL
);
2962 /* The headers can't be on the sublist without an L1 header */
2963 ASSERT(HDR_HAS_L1HDR(data_hdr
));
2964 ASSERT(HDR_HAS_L1HDR(meta_hdr
));
2966 if (data_hdr
->b_l1hdr
.b_arc_access
<
2967 meta_hdr
->b_l1hdr
.b_arc_access
) {
2968 type
= ARC_BUFC_DATA
;
2970 type
= ARC_BUFC_METADATA
;
2974 multilist_sublist_unlock(meta_mls
);
2975 multilist_sublist_unlock(data_mls
);
2981 * Evict buffers from the cache, such that arc_size is capped by arc_c.
2986 uint64_t total_evicted
= 0;
2991 * If we're over arc_meta_limit, we want to correct that before
2992 * potentially evicting data buffers below.
2994 total_evicted
+= arc_adjust_meta();
2999 * If we're over the target cache size, we want to evict enough
3000 * from the list to get back to our target size. We don't want
3001 * to evict too much from the MRU, such that it drops below
3002 * arc_p. So, if we're over our target cache size more than
3003 * the MRU is over arc_p, we'll evict enough to get back to
3004 * arc_p here, and then evict more from the MFU below.
3006 target
= MIN((int64_t)(arc_size
- arc_c
),
3007 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
3008 refcount_count(&arc_mru
->arcs_size
) + arc_meta_used
- arc_p
));
3011 * If we're below arc_meta_min, always prefer to evict data.
3012 * Otherwise, try to satisfy the requested number of bytes to
3013 * evict from the type which contains older buffers; in an
3014 * effort to keep newer buffers in the cache regardless of their
3015 * type. If we cannot satisfy the number of bytes from this
3016 * type, spill over into the next type.
3018 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
3019 arc_meta_used
> arc_meta_min
) {
3020 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3021 total_evicted
+= bytes
;
3024 * If we couldn't evict our target number of bytes from
3025 * metadata, we try to get the rest from data.
3030 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3032 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3033 total_evicted
+= bytes
;
3036 * If we couldn't evict our target number of bytes from
3037 * data, we try to get the rest from metadata.
3042 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3048 * Now that we've tried to evict enough from the MRU to get its
3049 * size back to arc_p, if we're still above the target cache
3050 * size, we evict the rest from the MFU.
3052 target
= arc_size
- arc_c
;
3054 if (arc_adjust_type(arc_mfu
) == ARC_BUFC_METADATA
&&
3055 arc_meta_used
> arc_meta_min
) {
3056 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3057 total_evicted
+= bytes
;
3060 * If we couldn't evict our target number of bytes from
3061 * metadata, we try to get the rest from data.
3066 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3068 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3069 total_evicted
+= bytes
;
3072 * If we couldn't evict our target number of bytes from
3073 * data, we try to get the rest from data.
3078 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3082 * Adjust ghost lists
3084 * In addition to the above, the ARC also defines target values
3085 * for the ghost lists. The sum of the mru list and mru ghost
3086 * list should never exceed the target size of the cache, and
3087 * the sum of the mru list, mfu list, mru ghost list, and mfu
3088 * ghost list should never exceed twice the target size of the
3089 * cache. The following logic enforces these limits on the ghost
3090 * caches, and evicts from them as needed.
3092 target
= refcount_count(&arc_mru
->arcs_size
) +
3093 refcount_count(&arc_mru_ghost
->arcs_size
) - arc_c
;
3095 bytes
= arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_DATA
);
3096 total_evicted
+= bytes
;
3101 arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_METADATA
);
3104 * We assume the sum of the mru list and mfu list is less than
3105 * or equal to arc_c (we enforced this above), which means we
3106 * can use the simpler of the two equations below:
3108 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3109 * mru ghost + mfu ghost <= arc_c
3111 target
= refcount_count(&arc_mru_ghost
->arcs_size
) +
3112 refcount_count(&arc_mfu_ghost
->arcs_size
) - arc_c
;
3114 bytes
= arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_DATA
);
3115 total_evicted
+= bytes
;
3120 arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_METADATA
);
3122 return (total_evicted
);
3126 arc_do_user_evicts(void)
3128 mutex_enter(&arc_user_evicts_lock
);
3129 while (arc_eviction_list
!= NULL
) {
3130 arc_buf_t
*buf
= arc_eviction_list
;
3131 arc_eviction_list
= buf
->b_next
;
3132 mutex_enter(&buf
->b_evict_lock
);
3134 mutex_exit(&buf
->b_evict_lock
);
3135 mutex_exit(&arc_user_evicts_lock
);
3137 if (buf
->b_efunc
!= NULL
)
3138 VERIFY0(buf
->b_efunc(buf
->b_private
));
3140 buf
->b_efunc
= NULL
;
3141 buf
->b_private
= NULL
;
3142 kmem_cache_free(buf_cache
, buf
);
3143 mutex_enter(&arc_user_evicts_lock
);
3145 mutex_exit(&arc_user_evicts_lock
);
3149 arc_flush(spa_t
*spa
, boolean_t retry
)
3154 * If retry is TRUE, a spa must not be specified since we have
3155 * no good way to determine if all of a spa's buffers have been
3156 * evicted from an arc state.
3158 ASSERT(!retry
|| spa
== 0);
3161 guid
= spa_load_guid(spa
);
3163 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_DATA
, retry
);
3164 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_METADATA
, retry
);
3166 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_DATA
, retry
);
3167 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_METADATA
, retry
);
3169 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3170 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3172 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3173 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3175 arc_do_user_evicts();
3176 ASSERT(spa
|| arc_eviction_list
== NULL
);
3180 arc_shrink(int64_t to_free
)
3182 if (arc_c
> arc_c_min
) {
3184 if (arc_c
> arc_c_min
+ to_free
)
3185 atomic_add_64(&arc_c
, -to_free
);
3189 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
3190 if (arc_c
> arc_size
)
3191 arc_c
= MAX(arc_size
, arc_c_min
);
3193 arc_p
= (arc_c
>> 1);
3194 ASSERT(arc_c
>= arc_c_min
);
3195 ASSERT((int64_t)arc_p
>= 0);
3198 if (arc_size
> arc_c
)
3199 (void) arc_adjust();
3202 typedef enum free_memory_reason_t
{
3207 FMR_PAGES_PP_MAXIMUM
,
3210 } free_memory_reason_t
;
3212 int64_t last_free_memory
;
3213 free_memory_reason_t last_free_reason
;
3217 * Additional reserve of pages for pp_reserve.
3219 int64_t arc_pages_pp_reserve
= 64;
3222 * Additional reserve of pages for swapfs.
3224 int64_t arc_swapfs_reserve
= 64;
3225 #endif /* _KERNEL */
3228 * Return the amount of memory that can be consumed before reclaim will be
3229 * needed. Positive if there is sufficient free memory, negative indicates
3230 * the amount of memory that needs to be freed up.
3233 arc_available_memory(void)
3235 int64_t lowest
= INT64_MAX
;
3236 free_memory_reason_t r
= FMR_UNKNOWN
;
3240 pgcnt_t needfree
= btop(arc_need_free
);
3241 pgcnt_t lotsfree
= btop(arc_sys_free
);
3242 pgcnt_t desfree
= 0;
3246 n
= PAGESIZE
* (-needfree
);
3254 * check that we're out of range of the pageout scanner. It starts to
3255 * schedule paging if freemem is less than lotsfree and needfree.
3256 * lotsfree is the high-water mark for pageout, and needfree is the
3257 * number of needed free pages. We add extra pages here to make sure
3258 * the scanner doesn't start up while we're freeing memory.
3260 n
= PAGESIZE
* (freemem
- lotsfree
- needfree
- desfree
);
3268 * check to make sure that swapfs has enough space so that anon
3269 * reservations can still succeed. anon_resvmem() checks that the
3270 * availrmem is greater than swapfs_minfree, and the number of reserved
3271 * swap pages. We also add a bit of extra here just to prevent
3272 * circumstances from getting really dire.
3274 n
= PAGESIZE
* (availrmem
- swapfs_minfree
- swapfs_reserve
-
3275 desfree
- arc_swapfs_reserve
);
3278 r
= FMR_SWAPFS_MINFREE
;
3283 * Check that we have enough availrmem that memory locking (e.g., via
3284 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3285 * stores the number of pages that cannot be locked; when availrmem
3286 * drops below pages_pp_maximum, page locking mechanisms such as
3287 * page_pp_lock() will fail.)
3289 n
= PAGESIZE
* (availrmem
- pages_pp_maximum
-
3290 arc_pages_pp_reserve
);
3293 r
= FMR_PAGES_PP_MAXIMUM
;
3299 * If we're on an i386 platform, it's possible that we'll exhaust the
3300 * kernel heap space before we ever run out of available physical
3301 * memory. Most checks of the size of the heap_area compare against
3302 * tune.t_minarmem, which is the minimum available real memory that we
3303 * can have in the system. However, this is generally fixed at 25 pages
3304 * which is so low that it's useless. In this comparison, we seek to
3305 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3306 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3309 n
= vmem_size(heap_arena
, VMEM_FREE
) -
3310 (vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
) >> 2);
3318 * If zio data pages are being allocated out of a separate heap segment,
3319 * then enforce that the size of available vmem for this arena remains
3320 * above about 1/16th free.
3322 * Note: The 1/16th arena free requirement was put in place
3323 * to aggressively evict memory from the arc in order to avoid
3324 * memory fragmentation issues.
3326 if (zio_arena
!= NULL
) {
3327 n
= vmem_size(zio_arena
, VMEM_FREE
) -
3328 (vmem_size(zio_arena
, VMEM_ALLOC
) >> 4);
3335 /* Every 100 calls, free a small amount */
3336 if (spa_get_random(100) == 0)
3338 #endif /* _KERNEL */
3340 last_free_memory
= lowest
;
3341 last_free_reason
= r
;
3347 * Determine if the system is under memory pressure and is asking
3348 * to reclaim memory. A return value of TRUE indicates that the system
3349 * is under memory pressure and that the arc should adjust accordingly.
3352 arc_reclaim_needed(void)
3354 return (arc_available_memory() < 0);
3358 arc_kmem_reap_now(void)
3361 kmem_cache_t
*prev_cache
= NULL
;
3362 kmem_cache_t
*prev_data_cache
= NULL
;
3363 extern kmem_cache_t
*zio_buf_cache
[];
3364 extern kmem_cache_t
*zio_data_buf_cache
[];
3365 extern kmem_cache_t
*range_seg_cache
;
3367 if ((arc_meta_used
>= arc_meta_limit
) && zfs_arc_meta_prune
) {
3369 * We are exceeding our meta-data cache limit.
3370 * Prune some entries to release holds on meta-data.
3372 arc_prune_async(zfs_arc_meta_prune
);
3375 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
3376 if (zio_buf_cache
[i
] != prev_cache
) {
3377 prev_cache
= zio_buf_cache
[i
];
3378 kmem_cache_reap_now(zio_buf_cache
[i
]);
3380 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
3381 prev_data_cache
= zio_data_buf_cache
[i
];
3382 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
3385 kmem_cache_reap_now(buf_cache
);
3386 kmem_cache_reap_now(hdr_full_cache
);
3387 kmem_cache_reap_now(hdr_l2only_cache
);
3388 kmem_cache_reap_now(range_seg_cache
);
3390 if (zio_arena
!= NULL
) {
3392 * Ask the vmem arena to reclaim unused memory from its
3395 vmem_qcache_reap(zio_arena
);
3400 * Threads can block in arc_get_data_buf() waiting for this thread to evict
3401 * enough data and signal them to proceed. When this happens, the threads in
3402 * arc_get_data_buf() are sleeping while holding the hash lock for their
3403 * particular arc header. Thus, we must be careful to never sleep on a
3404 * hash lock in this thread. This is to prevent the following deadlock:
3406 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3407 * waiting for the reclaim thread to signal it.
3409 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3410 * fails, and goes to sleep forever.
3412 * This possible deadlock is avoided by always acquiring a hash lock
3413 * using mutex_tryenter() from arc_reclaim_thread().
3416 arc_reclaim_thread(void)
3418 fstrans_cookie_t cookie
= spl_fstrans_mark();
3419 clock_t growtime
= 0;
3422 CALLB_CPR_INIT(&cpr
, &arc_reclaim_lock
, callb_generic_cpr
, FTAG
);
3424 mutex_enter(&arc_reclaim_lock
);
3425 while (!arc_reclaim_thread_exit
) {
3427 int64_t free_memory
= arc_available_memory();
3428 uint64_t evicted
= 0;
3430 arc_tuning_update();
3432 mutex_exit(&arc_reclaim_lock
);
3434 if (free_memory
< 0) {
3436 arc_no_grow
= B_TRUE
;
3440 * Wait at least zfs_grow_retry (default 5) seconds
3441 * before considering growing.
3443 growtime
= ddi_get_lbolt() + (arc_grow_retry
* hz
);
3445 arc_kmem_reap_now();
3448 * If we are still low on memory, shrink the ARC
3449 * so that we have arc_shrink_min free space.
3451 free_memory
= arc_available_memory();
3453 to_free
= (arc_c
>> arc_shrink_shift
) - free_memory
;
3456 to_free
= MAX(to_free
, arc_need_free
);
3458 arc_shrink(to_free
);
3460 } else if (free_memory
< arc_c
>> arc_no_grow_shift
) {
3461 arc_no_grow
= B_TRUE
;
3462 } else if (ddi_get_lbolt() >= growtime
) {
3463 arc_no_grow
= B_FALSE
;
3466 evicted
= arc_adjust();
3468 mutex_enter(&arc_reclaim_lock
);
3471 * If evicted is zero, we couldn't evict anything via
3472 * arc_adjust(). This could be due to hash lock
3473 * collisions, but more likely due to the majority of
3474 * arc buffers being unevictable. Therefore, even if
3475 * arc_size is above arc_c, another pass is unlikely to
3476 * be helpful and could potentially cause us to enter an
3479 if (arc_size
<= arc_c
|| evicted
== 0) {
3481 * We're either no longer overflowing, or we
3482 * can't evict anything more, so we should wake
3483 * up any threads before we go to sleep and clear
3484 * arc_need_free since nothing more can be done.
3486 cv_broadcast(&arc_reclaim_waiters_cv
);
3490 * Block until signaled, or after one second (we
3491 * might need to perform arc_kmem_reap_now()
3492 * even if we aren't being signalled)
3494 CALLB_CPR_SAFE_BEGIN(&cpr
);
3495 (void) cv_timedwait_sig(&arc_reclaim_thread_cv
,
3496 &arc_reclaim_lock
, ddi_get_lbolt() + hz
);
3497 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_lock
);
3501 arc_reclaim_thread_exit
= FALSE
;
3502 cv_broadcast(&arc_reclaim_thread_cv
);
3503 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_lock */
3504 spl_fstrans_unmark(cookie
);
3509 arc_user_evicts_thread(void)
3511 fstrans_cookie_t cookie
= spl_fstrans_mark();
3514 CALLB_CPR_INIT(&cpr
, &arc_user_evicts_lock
, callb_generic_cpr
, FTAG
);
3516 mutex_enter(&arc_user_evicts_lock
);
3517 while (!arc_user_evicts_thread_exit
) {
3518 mutex_exit(&arc_user_evicts_lock
);
3520 arc_do_user_evicts();
3523 * This is necessary in order for the mdb ::arc dcmd to
3524 * show up to date information. Since the ::arc command
3525 * does not call the kstat's update function, without
3526 * this call, the command may show stale stats for the
3527 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3528 * with this change, the data might be up to 1 second
3529 * out of date; but that should suffice. The arc_state_t
3530 * structures can be queried directly if more accurate
3531 * information is needed.
3533 if (arc_ksp
!= NULL
)
3534 arc_ksp
->ks_update(arc_ksp
, KSTAT_READ
);
3536 mutex_enter(&arc_user_evicts_lock
);
3539 * Block until signaled, or after one second (we need to
3540 * call the arc's kstat update function regularly).
3542 CALLB_CPR_SAFE_BEGIN(&cpr
);
3543 (void) cv_timedwait_sig(&arc_user_evicts_cv
,
3544 &arc_user_evicts_lock
, ddi_get_lbolt() + hz
);
3545 CALLB_CPR_SAFE_END(&cpr
, &arc_user_evicts_lock
);
3548 arc_user_evicts_thread_exit
= FALSE
;
3549 cv_broadcast(&arc_user_evicts_cv
);
3550 CALLB_CPR_EXIT(&cpr
); /* drops arc_user_evicts_lock */
3551 spl_fstrans_unmark(cookie
);
3557 * Determine the amount of memory eligible for eviction contained in the
3558 * ARC. All clean data reported by the ghost lists can always be safely
3559 * evicted. Due to arc_c_min, the same does not hold for all clean data
3560 * contained by the regular mru and mfu lists.
3562 * In the case of the regular mru and mfu lists, we need to report as
3563 * much clean data as possible, such that evicting that same reported
3564 * data will not bring arc_size below arc_c_min. Thus, in certain
3565 * circumstances, the total amount of clean data in the mru and mfu
3566 * lists might not actually be evictable.
3568 * The following two distinct cases are accounted for:
3570 * 1. The sum of the amount of dirty data contained by both the mru and
3571 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
3572 * is greater than or equal to arc_c_min.
3573 * (i.e. amount of dirty data >= arc_c_min)
3575 * This is the easy case; all clean data contained by the mru and mfu
3576 * lists is evictable. Evicting all clean data can only drop arc_size
3577 * to the amount of dirty data, which is greater than arc_c_min.
3579 * 2. The sum of the amount of dirty data contained by both the mru and
3580 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
3581 * is less than arc_c_min.
3582 * (i.e. arc_c_min > amount of dirty data)
3584 * 2.1. arc_size is greater than or equal arc_c_min.
3585 * (i.e. arc_size >= arc_c_min > amount of dirty data)
3587 * In this case, not all clean data from the regular mru and mfu
3588 * lists is actually evictable; we must leave enough clean data
3589 * to keep arc_size above arc_c_min. Thus, the maximum amount of
3590 * evictable data from the two lists combined, is exactly the
3591 * difference between arc_size and arc_c_min.
3593 * 2.2. arc_size is less than arc_c_min
3594 * (i.e. arc_c_min > arc_size > amount of dirty data)
3596 * In this case, none of the data contained in the mru and mfu
3597 * lists is evictable, even if it's clean. Since arc_size is
3598 * already below arc_c_min, evicting any more would only
3599 * increase this negative difference.
3602 arc_evictable_memory(void) {
3603 uint64_t arc_clean
=
3604 arc_mru
->arcs_lsize
[ARC_BUFC_DATA
] +
3605 arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] +
3606 arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
] +
3607 arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
];
3608 uint64_t ghost_clean
=
3609 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
3610 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_METADATA
] +
3611 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
3612 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_METADATA
];
3613 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
3615 if (arc_dirty
>= arc_c_min
)
3616 return (ghost_clean
+ arc_clean
);
3618 return (ghost_clean
+ MAX((int64_t)arc_size
- (int64_t)arc_c_min
, 0));
3622 * If sc->nr_to_scan is zero, the caller is requesting a query of the
3623 * number of objects which can potentially be freed. If it is nonzero,
3624 * the request is to free that many objects.
3626 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
3627 * in struct shrinker and also require the shrinker to return the number
3630 * Older kernels require the shrinker to return the number of freeable
3631 * objects following the freeing of nr_to_free.
3633 static spl_shrinker_t
3634 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
3638 /* The arc is considered warm once reclaim has occurred */
3639 if (unlikely(arc_warm
== B_FALSE
))
3642 /* Return the potential number of reclaimable pages */
3643 pages
= btop((int64_t)arc_evictable_memory());
3644 if (sc
->nr_to_scan
== 0)
3647 /* Not allowed to perform filesystem reclaim */
3648 if (!(sc
->gfp_mask
& __GFP_FS
))
3649 return (SHRINK_STOP
);
3651 /* Reclaim in progress */
3652 if (mutex_tryenter(&arc_reclaim_lock
) == 0)
3653 return (SHRINK_STOP
);
3655 mutex_exit(&arc_reclaim_lock
);
3658 * Evict the requested number of pages by shrinking arc_c the
3659 * requested amount. If there is nothing left to evict just
3660 * reap whatever we can from the various arc slabs.
3663 arc_shrink(ptob(sc
->nr_to_scan
));
3664 arc_kmem_reap_now();
3665 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
3666 pages
= MAX(pages
- btop(arc_evictable_memory()), 0);
3668 pages
= btop(arc_evictable_memory());
3671 arc_kmem_reap_now();
3672 pages
= SHRINK_STOP
;
3676 * We've reaped what we can, wake up threads.
3678 cv_broadcast(&arc_reclaim_waiters_cv
);
3681 * When direct reclaim is observed it usually indicates a rapid
3682 * increase in memory pressure. This occurs because the kswapd
3683 * threads were unable to asynchronously keep enough free memory
3684 * available. In this case set arc_no_grow to briefly pause arc
3685 * growth to avoid compounding the memory pressure.
3687 if (current_is_kswapd()) {
3688 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
3690 arc_no_grow
= B_TRUE
;
3691 arc_need_free
= ptob(sc
->nr_to_scan
);
3692 ARCSTAT_BUMP(arcstat_memory_direct_count
);
3697 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
3699 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
3700 #endif /* _KERNEL */
3703 * Adapt arc info given the number of bytes we are trying to add and
3704 * the state that we are comming from. This function is only called
3705 * when we are adding new content to the cache.
3708 arc_adapt(int bytes
, arc_state_t
*state
)
3711 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
3712 int64_t mrug_size
= refcount_count(&arc_mru_ghost
->arcs_size
);
3713 int64_t mfug_size
= refcount_count(&arc_mfu_ghost
->arcs_size
);
3715 if (state
== arc_l2c_only
)
3720 * Adapt the target size of the MRU list:
3721 * - if we just hit in the MRU ghost list, then increase
3722 * the target size of the MRU list.
3723 * - if we just hit in the MFU ghost list, then increase
3724 * the target size of the MFU list by decreasing the
3725 * target size of the MRU list.
3727 if (state
== arc_mru_ghost
) {
3728 mult
= (mrug_size
>= mfug_size
) ? 1 : (mfug_size
/ mrug_size
);
3729 if (!zfs_arc_p_dampener_disable
)
3730 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
3732 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
3733 } else if (state
== arc_mfu_ghost
) {
3736 mult
= (mfug_size
>= mrug_size
) ? 1 : (mrug_size
/ mfug_size
);
3737 if (!zfs_arc_p_dampener_disable
)
3738 mult
= MIN(mult
, 10);
3740 delta
= MIN(bytes
* mult
, arc_p
);
3741 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
3743 ASSERT((int64_t)arc_p
>= 0);
3745 if (arc_reclaim_needed()) {
3746 cv_signal(&arc_reclaim_thread_cv
);
3753 if (arc_c
>= arc_c_max
)
3757 * If we're within (2 * maxblocksize) bytes of the target
3758 * cache size, increment the target cache size
3760 ASSERT3U(arc_c
, >=, 2ULL << SPA_MAXBLOCKSHIFT
);
3761 arc_c
= MAX(arc_c
, 2ULL << SPA_MAXBLOCKSHIFT
);
3762 if (arc_size
>= arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
3763 atomic_add_64(&arc_c
, (int64_t)bytes
);
3764 if (arc_c
> arc_c_max
)
3766 else if (state
== arc_anon
)
3767 atomic_add_64(&arc_p
, (int64_t)bytes
);
3771 ASSERT((int64_t)arc_p
>= 0);
3775 * Check if arc_size has grown past our upper threshold, determined by
3776 * zfs_arc_overflow_shift.
3779 arc_is_overflowing(void)
3781 /* Always allow at least one block of overflow */
3782 uint64_t overflow
= MAX(SPA_MAXBLOCKSIZE
,
3783 arc_c
>> zfs_arc_overflow_shift
);
3785 return (arc_size
>= arc_c
+ overflow
);
3789 * The buffer, supplied as the first argument, needs a data block. If we
3790 * are hitting the hard limit for the cache size, we must sleep, waiting
3791 * for the eviction thread to catch up. If we're past the target size
3792 * but below the hard limit, we'll only signal the reclaim thread and
3796 arc_get_data_buf(arc_buf_t
*buf
)
3798 arc_state_t
*state
= buf
->b_hdr
->b_l1hdr
.b_state
;
3799 uint64_t size
= buf
->b_hdr
->b_size
;
3800 arc_buf_contents_t type
= arc_buf_type(buf
->b_hdr
);
3802 arc_adapt(size
, state
);
3805 * If arc_size is currently overflowing, and has grown past our
3806 * upper limit, we must be adding data faster than the evict
3807 * thread can evict. Thus, to ensure we don't compound the
3808 * problem by adding more data and forcing arc_size to grow even
3809 * further past it's target size, we halt and wait for the
3810 * eviction thread to catch up.
3812 * It's also possible that the reclaim thread is unable to evict
3813 * enough buffers to get arc_size below the overflow limit (e.g.
3814 * due to buffers being un-evictable, or hash lock collisions).
3815 * In this case, we want to proceed regardless if we're
3816 * overflowing; thus we don't use a while loop here.
3818 if (arc_is_overflowing()) {
3819 mutex_enter(&arc_reclaim_lock
);
3822 * Now that we've acquired the lock, we may no longer be
3823 * over the overflow limit, lets check.
3825 * We're ignoring the case of spurious wake ups. If that
3826 * were to happen, it'd let this thread consume an ARC
3827 * buffer before it should have (i.e. before we're under
3828 * the overflow limit and were signalled by the reclaim
3829 * thread). As long as that is a rare occurrence, it
3830 * shouldn't cause any harm.
3832 if (arc_is_overflowing()) {
3833 cv_signal(&arc_reclaim_thread_cv
);
3834 cv_wait(&arc_reclaim_waiters_cv
, &arc_reclaim_lock
);
3837 mutex_exit(&arc_reclaim_lock
);
3840 if (type
== ARC_BUFC_METADATA
) {
3841 buf
->b_data
= zio_buf_alloc(size
);
3842 arc_space_consume(size
, ARC_SPACE_META
);
3844 ASSERT(type
== ARC_BUFC_DATA
);
3845 buf
->b_data
= zio_data_buf_alloc(size
);
3846 arc_space_consume(size
, ARC_SPACE_DATA
);
3850 * Update the state size. Note that ghost states have a
3851 * "ghost size" and so don't need to be updated.
3853 if (!GHOST_STATE(buf
->b_hdr
->b_l1hdr
.b_state
)) {
3854 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3855 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
3857 (void) refcount_add_many(&state
->arcs_size
, size
, buf
);
3860 * If this is reached via arc_read, the link is
3861 * protected by the hash lock. If reached via
3862 * arc_buf_alloc, the header should not be accessed by
3863 * any other thread. And, if reached via arc_read_done,
3864 * the hash lock will protect it if it's found in the
3865 * hash table; otherwise no other thread should be
3866 * trying to [add|remove]_reference it.
3868 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
3869 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3870 atomic_add_64(&hdr
->b_l1hdr
.b_state
->arcs_lsize
[type
],
3874 * If we are growing the cache, and we are adding anonymous
3875 * data, and we have outgrown arc_p, update arc_p
3877 if (arc_size
< arc_c
&& hdr
->b_l1hdr
.b_state
== arc_anon
&&
3878 (refcount_count(&arc_anon
->arcs_size
) +
3879 refcount_count(&arc_mru
->arcs_size
) > arc_p
))
3880 arc_p
= MIN(arc_c
, arc_p
+ size
);
3885 * This routine is called whenever a buffer is accessed.
3886 * NOTE: the hash lock is dropped in this function.
3889 arc_access(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
3893 ASSERT(MUTEX_HELD(hash_lock
));
3894 ASSERT(HDR_HAS_L1HDR(hdr
));
3896 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
3898 * This buffer is not in the cache, and does not
3899 * appear in our "ghost" list. Add the new buffer
3903 ASSERT0(hdr
->b_l1hdr
.b_arc_access
);
3904 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3905 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
3906 arc_change_state(arc_mru
, hdr
, hash_lock
);
3908 } else if (hdr
->b_l1hdr
.b_state
== arc_mru
) {
3909 now
= ddi_get_lbolt();
3912 * If this buffer is here because of a prefetch, then either:
3913 * - clear the flag if this is a "referencing" read
3914 * (any subsequent access will bump this into the MFU state).
3916 * - move the buffer to the head of the list if this is
3917 * another prefetch (to make it less likely to be evicted).
3919 if (HDR_PREFETCH(hdr
)) {
3920 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
3921 /* link protected by hash lock */
3922 ASSERT(multilist_link_active(
3923 &hdr
->b_l1hdr
.b_arc_node
));
3925 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
3926 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
3927 ARCSTAT_BUMP(arcstat_mru_hits
);
3929 hdr
->b_l1hdr
.b_arc_access
= now
;
3934 * This buffer has been "accessed" only once so far,
3935 * but it is still in the cache. Move it to the MFU
3938 if (ddi_time_after(now
, hdr
->b_l1hdr
.b_arc_access
+
3941 * More than 125ms have passed since we
3942 * instantiated this buffer. Move it to the
3943 * most frequently used state.
3945 hdr
->b_l1hdr
.b_arc_access
= now
;
3946 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
3947 arc_change_state(arc_mfu
, hdr
, hash_lock
);
3949 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
3950 ARCSTAT_BUMP(arcstat_mru_hits
);
3951 } else if (hdr
->b_l1hdr
.b_state
== arc_mru_ghost
) {
3952 arc_state_t
*new_state
;
3954 * This buffer has been "accessed" recently, but
3955 * was evicted from the cache. Move it to the
3959 if (HDR_PREFETCH(hdr
)) {
3960 new_state
= arc_mru
;
3961 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0)
3962 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
3963 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
3965 new_state
= arc_mfu
;
3966 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
3969 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3970 arc_change_state(new_state
, hdr
, hash_lock
);
3972 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_ghost_hits
);
3973 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
3974 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu
) {
3976 * This buffer has been accessed more than once and is
3977 * still in the cache. Keep it in the MFU state.
3979 * NOTE: an add_reference() that occurred when we did
3980 * the arc_read() will have kicked this off the list.
3981 * If it was a prefetch, we will explicitly move it to
3982 * the head of the list now.
3984 if ((HDR_PREFETCH(hdr
)) != 0) {
3985 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3986 /* link protected by hash_lock */
3987 ASSERT(multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3989 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_hits
);
3990 ARCSTAT_BUMP(arcstat_mfu_hits
);
3991 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3992 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu_ghost
) {
3993 arc_state_t
*new_state
= arc_mfu
;
3995 * This buffer has been accessed more than once but has
3996 * been evicted from the cache. Move it back to the
4000 if (HDR_PREFETCH(hdr
)) {
4002 * This is a prefetch access...
4003 * move this block back to the MRU state.
4005 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
4006 new_state
= arc_mru
;
4009 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4010 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4011 arc_change_state(new_state
, hdr
, hash_lock
);
4013 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_ghost_hits
);
4014 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
4015 } else if (hdr
->b_l1hdr
.b_state
== arc_l2c_only
) {
4017 * This buffer is on the 2nd Level ARC.
4020 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4021 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4022 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4024 cmn_err(CE_PANIC
, "invalid arc state 0x%p",
4025 hdr
->b_l1hdr
.b_state
);
4029 /* a generic arc_done_func_t which you can use */
4032 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4034 if (zio
== NULL
|| zio
->io_error
== 0)
4035 bcopy(buf
->b_data
, arg
, buf
->b_hdr
->b_size
);
4036 VERIFY(arc_buf_remove_ref(buf
, arg
));
4039 /* a generic arc_done_func_t */
4041 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4043 arc_buf_t
**bufp
= arg
;
4044 if (zio
&& zio
->io_error
) {
4045 VERIFY(arc_buf_remove_ref(buf
, arg
));
4049 ASSERT(buf
->b_data
);
4054 arc_read_done(zio_t
*zio
)
4058 arc_buf_t
*abuf
; /* buffer we're assigning to callback */
4059 kmutex_t
*hash_lock
= NULL
;
4060 arc_callback_t
*callback_list
, *acb
;
4061 int freeable
= FALSE
;
4063 buf
= zio
->io_private
;
4067 * The hdr was inserted into hash-table and removed from lists
4068 * prior to starting I/O. We should find this header, since
4069 * it's in the hash table, and it should be legit since it's
4070 * not possible to evict it during the I/O. The only possible
4071 * reason for it not to be found is if we were freed during the
4074 if (HDR_IN_HASH_TABLE(hdr
)) {
4075 arc_buf_hdr_t
*found
;
4077 ASSERT3U(hdr
->b_birth
, ==, BP_PHYSICAL_BIRTH(zio
->io_bp
));
4078 ASSERT3U(hdr
->b_dva
.dva_word
[0], ==,
4079 BP_IDENTITY(zio
->io_bp
)->dva_word
[0]);
4080 ASSERT3U(hdr
->b_dva
.dva_word
[1], ==,
4081 BP_IDENTITY(zio
->io_bp
)->dva_word
[1]);
4083 found
= buf_hash_find(hdr
->b_spa
, zio
->io_bp
,
4086 ASSERT((found
== NULL
&& HDR_FREED_IN_READ(hdr
) &&
4087 hash_lock
== NULL
) ||
4089 DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
4090 (found
== hdr
&& HDR_L2_READING(hdr
)));
4093 hdr
->b_flags
&= ~ARC_FLAG_L2_EVICTED
;
4094 if (l2arc_noprefetch
&& HDR_PREFETCH(hdr
))
4095 hdr
->b_flags
&= ~ARC_FLAG_L2CACHE
;
4097 /* byteswap if necessary */
4098 callback_list
= hdr
->b_l1hdr
.b_acb
;
4099 ASSERT(callback_list
!= NULL
);
4100 if (BP_SHOULD_BYTESWAP(zio
->io_bp
) && zio
->io_error
== 0) {
4101 dmu_object_byteswap_t bswap
=
4102 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
4103 if (BP_GET_LEVEL(zio
->io_bp
) > 0)
4104 byteswap_uint64_array(buf
->b_data
, hdr
->b_size
);
4106 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, hdr
->b_size
);
4109 arc_cksum_compute(buf
, B_FALSE
);
4112 if (hash_lock
&& zio
->io_error
== 0 &&
4113 hdr
->b_l1hdr
.b_state
== arc_anon
) {
4115 * Only call arc_access on anonymous buffers. This is because
4116 * if we've issued an I/O for an evicted buffer, we've already
4117 * called arc_access (to prevent any simultaneous readers from
4118 * getting confused).
4120 arc_access(hdr
, hash_lock
);
4123 /* create copies of the data buffer for the callers */
4125 for (acb
= callback_list
; acb
; acb
= acb
->acb_next
) {
4126 if (acb
->acb_done
) {
4128 ARCSTAT_BUMP(arcstat_duplicate_reads
);
4129 abuf
= arc_buf_clone(buf
);
4131 acb
->acb_buf
= abuf
;
4135 hdr
->b_l1hdr
.b_acb
= NULL
;
4136 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
4137 ASSERT(!HDR_BUF_AVAILABLE(hdr
));
4139 ASSERT(buf
->b_efunc
== NULL
);
4140 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
4141 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
4144 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
) ||
4145 callback_list
!= NULL
);
4147 if (zio
->io_error
!= 0) {
4148 hdr
->b_flags
|= ARC_FLAG_IO_ERROR
;
4149 if (hdr
->b_l1hdr
.b_state
!= arc_anon
)
4150 arc_change_state(arc_anon
, hdr
, hash_lock
);
4151 if (HDR_IN_HASH_TABLE(hdr
))
4152 buf_hash_remove(hdr
);
4153 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
4157 * Broadcast before we drop the hash_lock to avoid the possibility
4158 * that the hdr (and hence the cv) might be freed before we get to
4159 * the cv_broadcast().
4161 cv_broadcast(&hdr
->b_l1hdr
.b_cv
);
4163 if (hash_lock
!= NULL
) {
4164 mutex_exit(hash_lock
);
4167 * This block was freed while we waited for the read to
4168 * complete. It has been removed from the hash table and
4169 * moved to the anonymous state (so that it won't show up
4172 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
4173 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
4176 /* execute each callback and free its structure */
4177 while ((acb
= callback_list
) != NULL
) {
4179 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
4181 if (acb
->acb_zio_dummy
!= NULL
) {
4182 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
4183 zio_nowait(acb
->acb_zio_dummy
);
4186 callback_list
= acb
->acb_next
;
4187 kmem_free(acb
, sizeof (arc_callback_t
));
4191 arc_hdr_destroy(hdr
);
4195 * "Read" the block at the specified DVA (in bp) via the
4196 * cache. If the block is found in the cache, invoke the provided
4197 * callback immediately and return. Note that the `zio' parameter
4198 * in the callback will be NULL in this case, since no IO was
4199 * required. If the block is not in the cache pass the read request
4200 * on to the spa with a substitute callback function, so that the
4201 * requested block will be added to the cache.
4203 * If a read request arrives for a block that has a read in-progress,
4204 * either wait for the in-progress read to complete (and return the
4205 * results); or, if this is a read with a "done" func, add a record
4206 * to the read to invoke the "done" func when the read completes,
4207 * and return; or just return.
4209 * arc_read_done() will invoke all the requested "done" functions
4210 * for readers of this block.
4213 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_done_func_t
*done
,
4214 void *private, zio_priority_t priority
, int zio_flags
,
4215 arc_flags_t
*arc_flags
, const zbookmark_phys_t
*zb
)
4217 arc_buf_hdr_t
*hdr
= NULL
;
4218 arc_buf_t
*buf
= NULL
;
4219 kmutex_t
*hash_lock
= NULL
;
4221 uint64_t guid
= spa_load_guid(spa
);
4224 ASSERT(!BP_IS_EMBEDDED(bp
) ||
4225 BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
);
4228 if (!BP_IS_EMBEDDED(bp
)) {
4230 * Embedded BP's have no DVA and require no I/O to "read".
4231 * Create an anonymous arc buf to back it.
4233 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
4236 if (hdr
!= NULL
&& HDR_HAS_L1HDR(hdr
) && hdr
->b_l1hdr
.b_datacnt
> 0) {
4238 *arc_flags
|= ARC_FLAG_CACHED
;
4240 if (HDR_IO_IN_PROGRESS(hdr
)) {
4242 if (*arc_flags
& ARC_FLAG_WAIT
) {
4243 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
4244 mutex_exit(hash_lock
);
4247 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
4250 arc_callback_t
*acb
= NULL
;
4252 acb
= kmem_zalloc(sizeof (arc_callback_t
),
4254 acb
->acb_done
= done
;
4255 acb
->acb_private
= private;
4257 acb
->acb_zio_dummy
= zio_null(pio
,
4258 spa
, NULL
, NULL
, NULL
, zio_flags
);
4260 ASSERT(acb
->acb_done
!= NULL
);
4261 acb
->acb_next
= hdr
->b_l1hdr
.b_acb
;
4262 hdr
->b_l1hdr
.b_acb
= acb
;
4263 add_reference(hdr
, hash_lock
, private);
4264 mutex_exit(hash_lock
);
4267 mutex_exit(hash_lock
);
4271 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
4272 hdr
->b_l1hdr
.b_state
== arc_mfu
);
4275 add_reference(hdr
, hash_lock
, private);
4277 * If this block is already in use, create a new
4278 * copy of the data so that we will be guaranteed
4279 * that arc_release() will always succeed.
4281 buf
= hdr
->b_l1hdr
.b_buf
;
4283 ASSERT(buf
->b_data
);
4284 if (HDR_BUF_AVAILABLE(hdr
)) {
4285 ASSERT(buf
->b_efunc
== NULL
);
4286 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
4288 buf
= arc_buf_clone(buf
);
4291 } else if (*arc_flags
& ARC_FLAG_PREFETCH
&&
4292 refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
4293 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
4295 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
4296 arc_access(hdr
, hash_lock
);
4297 if (*arc_flags
& ARC_FLAG_L2CACHE
)
4298 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
4299 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
4300 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
4301 mutex_exit(hash_lock
);
4302 ARCSTAT_BUMP(arcstat_hits
);
4303 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
4304 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
4305 data
, metadata
, hits
);
4308 done(NULL
, buf
, private);
4310 uint64_t size
= BP_GET_LSIZE(bp
);
4311 arc_callback_t
*acb
;
4314 boolean_t devw
= B_FALSE
;
4315 enum zio_compress b_compress
= ZIO_COMPRESS_OFF
;
4316 int32_t b_asize
= 0;
4319 * Gracefully handle a damaged logical block size as a
4320 * checksum error by passing a dummy zio to the done callback.
4322 if (size
> spa_maxblocksize(spa
)) {
4324 rzio
= zio_null(pio
, spa
, NULL
,
4325 NULL
, NULL
, zio_flags
);
4326 rzio
->io_error
= ECKSUM
;
4327 done(rzio
, buf
, private);
4335 /* this block is not in the cache */
4336 arc_buf_hdr_t
*exists
= NULL
;
4337 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
4338 buf
= arc_buf_alloc(spa
, size
, private, type
);
4340 if (!BP_IS_EMBEDDED(bp
)) {
4341 hdr
->b_dva
= *BP_IDENTITY(bp
);
4342 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
4343 exists
= buf_hash_insert(hdr
, &hash_lock
);
4345 if (exists
!= NULL
) {
4346 /* somebody beat us to the hash insert */
4347 mutex_exit(hash_lock
);
4348 buf_discard_identity(hdr
);
4349 (void) arc_buf_remove_ref(buf
, private);
4350 goto top
; /* restart the IO request */
4353 /* if this is a prefetch, we don't have a reference */
4354 if (*arc_flags
& ARC_FLAG_PREFETCH
) {
4355 (void) remove_reference(hdr
, hash_lock
,
4357 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
4359 if (*arc_flags
& ARC_FLAG_L2CACHE
)
4360 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
4361 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
4362 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
4363 if (BP_GET_LEVEL(bp
) > 0)
4364 hdr
->b_flags
|= ARC_FLAG_INDIRECT
;
4367 * This block is in the ghost cache. If it was L2-only
4368 * (and thus didn't have an L1 hdr), we realloc the
4369 * header to add an L1 hdr.
4371 if (!HDR_HAS_L1HDR(hdr
)) {
4372 hdr
= arc_hdr_realloc(hdr
, hdr_l2only_cache
,
4376 ASSERT(GHOST_STATE(hdr
->b_l1hdr
.b_state
));
4377 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4378 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4379 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
4381 /* if this is a prefetch, we don't have a reference */
4382 if (*arc_flags
& ARC_FLAG_PREFETCH
)
4383 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
4385 add_reference(hdr
, hash_lock
, private);
4386 if (*arc_flags
& ARC_FLAG_L2CACHE
)
4387 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
4388 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
4389 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
4390 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
4393 buf
->b_efunc
= NULL
;
4394 buf
->b_private
= NULL
;
4396 hdr
->b_l1hdr
.b_buf
= buf
;
4397 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
4398 hdr
->b_l1hdr
.b_datacnt
= 1;
4399 arc_get_data_buf(buf
);
4400 arc_access(hdr
, hash_lock
);
4403 ASSERT(!GHOST_STATE(hdr
->b_l1hdr
.b_state
));
4405 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
4406 acb
->acb_done
= done
;
4407 acb
->acb_private
= private;
4409 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
4410 hdr
->b_l1hdr
.b_acb
= acb
;
4411 hdr
->b_flags
|= ARC_FLAG_IO_IN_PROGRESS
;
4413 if (HDR_HAS_L2HDR(hdr
) &&
4414 (vd
= hdr
->b_l2hdr
.b_dev
->l2ad_vdev
) != NULL
) {
4415 devw
= hdr
->b_l2hdr
.b_dev
->l2ad_writing
;
4416 addr
= hdr
->b_l2hdr
.b_daddr
;
4417 b_compress
= hdr
->b_l2hdr
.b_compress
;
4418 b_asize
= hdr
->b_l2hdr
.b_asize
;
4420 * Lock out device removal.
4422 if (vdev_is_dead(vd
) ||
4423 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
4427 if (hash_lock
!= NULL
)
4428 mutex_exit(hash_lock
);
4431 * At this point, we have a level 1 cache miss. Try again in
4432 * L2ARC if possible.
4434 ASSERT3U(hdr
->b_size
, ==, size
);
4435 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
4436 uint64_t, size
, zbookmark_phys_t
*, zb
);
4437 ARCSTAT_BUMP(arcstat_misses
);
4438 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
4439 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
4440 data
, metadata
, misses
);
4442 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
4444 * Read from the L2ARC if the following are true:
4445 * 1. The L2ARC vdev was previously cached.
4446 * 2. This buffer still has L2ARC metadata.
4447 * 3. This buffer isn't currently writing to the L2ARC.
4448 * 4. The L2ARC entry wasn't evicted, which may
4449 * also have invalidated the vdev.
4450 * 5. This isn't prefetch and l2arc_noprefetch is set.
4452 if (HDR_HAS_L2HDR(hdr
) &&
4453 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
4454 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
4455 l2arc_read_callback_t
*cb
;
4457 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
4458 ARCSTAT_BUMP(arcstat_l2_hits
);
4459 atomic_inc_32(&hdr
->b_l2hdr
.b_hits
);
4461 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
4463 cb
->l2rcb_buf
= buf
;
4464 cb
->l2rcb_spa
= spa
;
4467 cb
->l2rcb_flags
= zio_flags
;
4468 cb
->l2rcb_compress
= b_compress
;
4470 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
4471 addr
+ size
< vd
->vdev_psize
-
4472 VDEV_LABEL_END_SIZE
);
4475 * l2arc read. The SCL_L2ARC lock will be
4476 * released by l2arc_read_done().
4477 * Issue a null zio if the underlying buffer
4478 * was squashed to zero size by compression.
4480 if (b_compress
== ZIO_COMPRESS_EMPTY
) {
4481 rzio
= zio_null(pio
, spa
, vd
,
4482 l2arc_read_done
, cb
,
4483 zio_flags
| ZIO_FLAG_DONT_CACHE
|
4485 ZIO_FLAG_DONT_PROPAGATE
|
4486 ZIO_FLAG_DONT_RETRY
);
4488 rzio
= zio_read_phys(pio
, vd
, addr
,
4489 b_asize
, buf
->b_data
,
4491 l2arc_read_done
, cb
, priority
,
4492 zio_flags
| ZIO_FLAG_DONT_CACHE
|
4494 ZIO_FLAG_DONT_PROPAGATE
|
4495 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
4497 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
4499 ARCSTAT_INCR(arcstat_l2_read_bytes
, b_asize
);
4501 if (*arc_flags
& ARC_FLAG_NOWAIT
) {
4506 ASSERT(*arc_flags
& ARC_FLAG_WAIT
);
4507 if (zio_wait(rzio
) == 0)
4510 /* l2arc read error; goto zio_read() */
4512 DTRACE_PROBE1(l2arc__miss
,
4513 arc_buf_hdr_t
*, hdr
);
4514 ARCSTAT_BUMP(arcstat_l2_misses
);
4515 if (HDR_L2_WRITING(hdr
))
4516 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
4517 spa_config_exit(spa
, SCL_L2ARC
, vd
);
4521 spa_config_exit(spa
, SCL_L2ARC
, vd
);
4522 if (l2arc_ndev
!= 0) {
4523 DTRACE_PROBE1(l2arc__miss
,
4524 arc_buf_hdr_t
*, hdr
);
4525 ARCSTAT_BUMP(arcstat_l2_misses
);
4529 rzio
= zio_read(pio
, spa
, bp
, buf
->b_data
, size
,
4530 arc_read_done
, buf
, priority
, zio_flags
, zb
);
4532 if (*arc_flags
& ARC_FLAG_WAIT
) {
4533 rc
= zio_wait(rzio
);
4537 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
4542 spa_read_history_add(spa
, zb
, *arc_flags
);
4547 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
4551 p
= kmem_alloc(sizeof (*p
), KM_SLEEP
);
4553 p
->p_private
= private;
4554 list_link_init(&p
->p_node
);
4555 refcount_create(&p
->p_refcnt
);
4557 mutex_enter(&arc_prune_mtx
);
4558 refcount_add(&p
->p_refcnt
, &arc_prune_list
);
4559 list_insert_head(&arc_prune_list
, p
);
4560 mutex_exit(&arc_prune_mtx
);
4566 arc_remove_prune_callback(arc_prune_t
*p
)
4568 mutex_enter(&arc_prune_mtx
);
4569 list_remove(&arc_prune_list
, p
);
4570 if (refcount_remove(&p
->p_refcnt
, &arc_prune_list
) == 0) {
4571 refcount_destroy(&p
->p_refcnt
);
4572 kmem_free(p
, sizeof (*p
));
4574 mutex_exit(&arc_prune_mtx
);
4578 arc_set_callback(arc_buf_t
*buf
, arc_evict_func_t
*func
, void *private)
4580 ASSERT(buf
->b_hdr
!= NULL
);
4581 ASSERT(buf
->b_hdr
->b_l1hdr
.b_state
!= arc_anon
);
4582 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
) ||
4584 ASSERT(buf
->b_efunc
== NULL
);
4585 ASSERT(!HDR_BUF_AVAILABLE(buf
->b_hdr
));
4587 buf
->b_efunc
= func
;
4588 buf
->b_private
= private;
4592 * Notify the arc that a block was freed, and thus will never be used again.
4595 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
4598 kmutex_t
*hash_lock
;
4599 uint64_t guid
= spa_load_guid(spa
);
4601 ASSERT(!BP_IS_EMBEDDED(bp
));
4603 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
4606 if (HDR_BUF_AVAILABLE(hdr
)) {
4607 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
4608 add_reference(hdr
, hash_lock
, FTAG
);
4609 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
4610 mutex_exit(hash_lock
);
4612 arc_release(buf
, FTAG
);
4613 (void) arc_buf_remove_ref(buf
, FTAG
);
4615 mutex_exit(hash_lock
);
4621 * Clear the user eviction callback set by arc_set_callback(), first calling
4622 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4623 * clearing the callback may result in the arc_buf being destroyed. However,
4624 * it will not result in the *last* arc_buf being destroyed, hence the data
4625 * will remain cached in the ARC. We make a copy of the arc buffer here so
4626 * that we can process the callback without holding any locks.
4628 * It's possible that the callback is already in the process of being cleared
4629 * by another thread. In this case we can not clear the callback.
4631 * Returns B_TRUE if the callback was successfully called and cleared.
4634 arc_clear_callback(arc_buf_t
*buf
)
4637 kmutex_t
*hash_lock
;
4638 arc_evict_func_t
*efunc
= buf
->b_efunc
;
4639 void *private = buf
->b_private
;
4641 mutex_enter(&buf
->b_evict_lock
);
4645 * We are in arc_do_user_evicts().
4647 ASSERT(buf
->b_data
== NULL
);
4648 mutex_exit(&buf
->b_evict_lock
);
4650 } else if (buf
->b_data
== NULL
) {
4652 * We are on the eviction list; process this buffer now
4653 * but let arc_do_user_evicts() do the reaping.
4655 buf
->b_efunc
= NULL
;
4656 mutex_exit(&buf
->b_evict_lock
);
4657 VERIFY0(efunc(private));
4660 hash_lock
= HDR_LOCK(hdr
);
4661 mutex_enter(hash_lock
);
4663 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
4665 ASSERT3U(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), <,
4666 hdr
->b_l1hdr
.b_datacnt
);
4667 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
4668 hdr
->b_l1hdr
.b_state
== arc_mfu
);
4670 buf
->b_efunc
= NULL
;
4671 buf
->b_private
= NULL
;
4673 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
4674 mutex_exit(&buf
->b_evict_lock
);
4675 arc_buf_destroy(buf
, TRUE
);
4677 ASSERT(buf
== hdr
->b_l1hdr
.b_buf
);
4678 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
4679 mutex_exit(&buf
->b_evict_lock
);
4682 mutex_exit(hash_lock
);
4683 VERIFY0(efunc(private));
4688 * Release this buffer from the cache, making it an anonymous buffer. This
4689 * must be done after a read and prior to modifying the buffer contents.
4690 * If the buffer has more than one reference, we must make
4691 * a new hdr for the buffer.
4694 arc_release(arc_buf_t
*buf
, void *tag
)
4696 kmutex_t
*hash_lock
;
4698 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4701 * It would be nice to assert that if its DMU metadata (level >
4702 * 0 || it's the dnode file), then it must be syncing context.
4703 * But we don't know that information at this level.
4706 mutex_enter(&buf
->b_evict_lock
);
4708 ASSERT(HDR_HAS_L1HDR(hdr
));
4711 * We don't grab the hash lock prior to this check, because if
4712 * the buffer's header is in the arc_anon state, it won't be
4713 * linked into the hash table.
4715 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
4716 mutex_exit(&buf
->b_evict_lock
);
4717 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4718 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
4719 ASSERT(!HDR_HAS_L2HDR(hdr
));
4720 ASSERT(BUF_EMPTY(hdr
));
4722 ASSERT3U(hdr
->b_l1hdr
.b_datacnt
, ==, 1);
4723 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), ==, 1);
4724 ASSERT(!list_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4726 ASSERT3P(buf
->b_efunc
, ==, NULL
);
4727 ASSERT3P(buf
->b_private
, ==, NULL
);
4729 hdr
->b_l1hdr
.b_arc_access
= 0;
4735 hash_lock
= HDR_LOCK(hdr
);
4736 mutex_enter(hash_lock
);
4739 * This assignment is only valid as long as the hash_lock is
4740 * held, we must be careful not to reference state or the
4741 * b_state field after dropping the lock.
4743 state
= hdr
->b_l1hdr
.b_state
;
4744 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
4745 ASSERT3P(state
, !=, arc_anon
);
4747 /* this buffer is not on any list */
4748 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0);
4750 if (HDR_HAS_L2HDR(hdr
)) {
4751 mutex_enter(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
4754 * We have to recheck this conditional again now that
4755 * we're holding the l2ad_mtx to prevent a race with
4756 * another thread which might be concurrently calling
4757 * l2arc_evict(). In that case, l2arc_evict() might have
4758 * destroyed the header's L2 portion as we were waiting
4759 * to acquire the l2ad_mtx.
4761 if (HDR_HAS_L2HDR(hdr
))
4762 arc_hdr_l2hdr_destroy(hdr
);
4764 mutex_exit(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
4768 * Do we have more than one buf?
4770 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
4771 arc_buf_hdr_t
*nhdr
;
4773 uint64_t blksz
= hdr
->b_size
;
4774 uint64_t spa
= hdr
->b_spa
;
4775 arc_buf_contents_t type
= arc_buf_type(hdr
);
4776 uint32_t flags
= hdr
->b_flags
;
4778 ASSERT(hdr
->b_l1hdr
.b_buf
!= buf
|| buf
->b_next
!= NULL
);
4780 * Pull the data off of this hdr and attach it to
4781 * a new anonymous hdr.
4783 (void) remove_reference(hdr
, hash_lock
, tag
);
4784 bufp
= &hdr
->b_l1hdr
.b_buf
;
4785 while (*bufp
!= buf
)
4786 bufp
= &(*bufp
)->b_next
;
4787 *bufp
= buf
->b_next
;
4790 ASSERT3P(state
, !=, arc_l2c_only
);
4792 (void) refcount_remove_many(
4793 &state
->arcs_size
, hdr
->b_size
, buf
);
4795 if (refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
)) {
4798 ASSERT3P(state
, !=, arc_l2c_only
);
4799 size
= &state
->arcs_lsize
[type
];
4800 ASSERT3U(*size
, >=, hdr
->b_size
);
4801 atomic_add_64(size
, -hdr
->b_size
);
4805 * We're releasing a duplicate user data buffer, update
4806 * our statistics accordingly.
4808 if (HDR_ISTYPE_DATA(hdr
)) {
4809 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
4810 ARCSTAT_INCR(arcstat_duplicate_buffers_size
,
4813 hdr
->b_l1hdr
.b_datacnt
-= 1;
4814 arc_cksum_verify(buf
);
4815 arc_buf_unwatch(buf
);
4817 mutex_exit(hash_lock
);
4819 nhdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
4820 nhdr
->b_size
= blksz
;
4823 nhdr
->b_l1hdr
.b_mru_hits
= 0;
4824 nhdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
4825 nhdr
->b_l1hdr
.b_mfu_hits
= 0;
4826 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
4827 nhdr
->b_l1hdr
.b_l2_hits
= 0;
4828 nhdr
->b_flags
= flags
& ARC_FLAG_L2_WRITING
;
4829 nhdr
->b_flags
|= arc_bufc_to_flags(type
);
4830 nhdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
4832 nhdr
->b_l1hdr
.b_buf
= buf
;
4833 nhdr
->b_l1hdr
.b_datacnt
= 1;
4834 nhdr
->b_l1hdr
.b_state
= arc_anon
;
4835 nhdr
->b_l1hdr
.b_arc_access
= 0;
4836 nhdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
4837 nhdr
->b_freeze_cksum
= NULL
;
4839 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, tag
);
4841 mutex_exit(&buf
->b_evict_lock
);
4842 (void) refcount_add_many(&arc_anon
->arcs_size
, blksz
, buf
);
4844 mutex_exit(&buf
->b_evict_lock
);
4845 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 1);
4846 /* protected by hash lock, or hdr is on arc_anon */
4847 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4848 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4849 hdr
->b_l1hdr
.b_mru_hits
= 0;
4850 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
4851 hdr
->b_l1hdr
.b_mfu_hits
= 0;
4852 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
4853 hdr
->b_l1hdr
.b_l2_hits
= 0;
4854 arc_change_state(arc_anon
, hdr
, hash_lock
);
4855 hdr
->b_l1hdr
.b_arc_access
= 0;
4856 mutex_exit(hash_lock
);
4858 buf_discard_identity(hdr
);
4861 buf
->b_efunc
= NULL
;
4862 buf
->b_private
= NULL
;
4866 arc_released(arc_buf_t
*buf
)
4870 mutex_enter(&buf
->b_evict_lock
);
4871 released
= (buf
->b_data
!= NULL
&&
4872 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
4873 mutex_exit(&buf
->b_evict_lock
);
4879 arc_referenced(arc_buf_t
*buf
)
4883 mutex_enter(&buf
->b_evict_lock
);
4884 referenced
= (refcount_count(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
4885 mutex_exit(&buf
->b_evict_lock
);
4886 return (referenced
);
4891 arc_write_ready(zio_t
*zio
)
4893 arc_write_callback_t
*callback
= zio
->io_private
;
4894 arc_buf_t
*buf
= callback
->awcb_buf
;
4895 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4897 ASSERT(HDR_HAS_L1HDR(hdr
));
4898 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
4899 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
4900 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
4903 * If the IO is already in progress, then this is a re-write
4904 * attempt, so we need to thaw and re-compute the cksum.
4905 * It is the responsibility of the callback to handle the
4906 * accounting for any re-write attempt.
4908 if (HDR_IO_IN_PROGRESS(hdr
)) {
4909 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
4910 if (hdr
->b_freeze_cksum
!= NULL
) {
4911 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
4912 hdr
->b_freeze_cksum
= NULL
;
4914 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
4916 arc_cksum_compute(buf
, B_FALSE
);
4917 hdr
->b_flags
|= ARC_FLAG_IO_IN_PROGRESS
;
4921 * The SPA calls this callback for each physical write that happens on behalf
4922 * of a logical write. See the comment in dbuf_write_physdone() for details.
4925 arc_write_physdone(zio_t
*zio
)
4927 arc_write_callback_t
*cb
= zio
->io_private
;
4928 if (cb
->awcb_physdone
!= NULL
)
4929 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
4933 arc_write_done(zio_t
*zio
)
4935 arc_write_callback_t
*callback
= zio
->io_private
;
4936 arc_buf_t
*buf
= callback
->awcb_buf
;
4937 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4939 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
4941 if (zio
->io_error
== 0) {
4942 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
4943 buf_discard_identity(hdr
);
4945 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
4946 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
4949 ASSERT(BUF_EMPTY(hdr
));
4953 * If the block to be written was all-zero or compressed enough to be
4954 * embedded in the BP, no write was performed so there will be no
4955 * dva/birth/checksum. The buffer must therefore remain anonymous
4958 if (!BUF_EMPTY(hdr
)) {
4959 arc_buf_hdr_t
*exists
;
4960 kmutex_t
*hash_lock
;
4962 ASSERT(zio
->io_error
== 0);
4964 arc_cksum_verify(buf
);
4966 exists
= buf_hash_insert(hdr
, &hash_lock
);
4967 if (exists
!= NULL
) {
4969 * This can only happen if we overwrite for
4970 * sync-to-convergence, because we remove
4971 * buffers from the hash table when we arc_free().
4973 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
4974 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
4975 panic("bad overwrite, hdr=%p exists=%p",
4976 (void *)hdr
, (void *)exists
);
4977 ASSERT(refcount_is_zero(
4978 &exists
->b_l1hdr
.b_refcnt
));
4979 arc_change_state(arc_anon
, exists
, hash_lock
);
4980 mutex_exit(hash_lock
);
4981 arc_hdr_destroy(exists
);
4982 exists
= buf_hash_insert(hdr
, &hash_lock
);
4983 ASSERT3P(exists
, ==, NULL
);
4984 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
4986 ASSERT(zio
->io_prop
.zp_nopwrite
);
4987 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
4988 panic("bad nopwrite, hdr=%p exists=%p",
4989 (void *)hdr
, (void *)exists
);
4992 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
4993 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
4994 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
4995 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
4998 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
4999 /* if it's not anon, we are doing a scrub */
5000 if (exists
== NULL
&& hdr
->b_l1hdr
.b_state
== arc_anon
)
5001 arc_access(hdr
, hash_lock
);
5002 mutex_exit(hash_lock
);
5004 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
5007 ASSERT(!refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5008 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
5010 kmem_free(callback
, sizeof (arc_write_callback_t
));
5014 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
5015 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
, boolean_t l2arc_compress
,
5016 const zio_prop_t
*zp
, arc_done_func_t
*ready
, arc_done_func_t
*physdone
,
5017 arc_done_func_t
*done
, void *private, zio_priority_t priority
,
5018 int zio_flags
, const zbookmark_phys_t
*zb
)
5020 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5021 arc_write_callback_t
*callback
;
5024 ASSERT(ready
!= NULL
);
5025 ASSERT(done
!= NULL
);
5026 ASSERT(!HDR_IO_ERROR(hdr
));
5027 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5028 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
5029 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
5031 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
5033 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
5034 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
5035 callback
->awcb_ready
= ready
;
5036 callback
->awcb_physdone
= physdone
;
5037 callback
->awcb_done
= done
;
5038 callback
->awcb_private
= private;
5039 callback
->awcb_buf
= buf
;
5041 zio
= zio_write(pio
, spa
, txg
, bp
, buf
->b_data
, hdr
->b_size
, zp
,
5042 arc_write_ready
, arc_write_physdone
, arc_write_done
, callback
,
5043 priority
, zio_flags
, zb
);
5049 arc_memory_throttle(uint64_t reserve
, uint64_t txg
)
5052 uint64_t available_memory
= ptob(freemem
);
5053 static uint64_t page_load
= 0;
5054 static uint64_t last_txg
= 0;
5056 pgcnt_t minfree
= btop(arc_sys_free
/ 4);
5059 if (freemem
> physmem
* arc_lotsfree_percent
/ 100)
5062 if (txg
> last_txg
) {
5068 * If we are in pageout, we know that memory is already tight,
5069 * the arc is already going to be evicting, so we just want to
5070 * continue to let page writes occur as quickly as possible.
5072 if (current_is_kswapd()) {
5073 if (page_load
> MAX(ptob(minfree
), available_memory
) / 4) {
5074 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
5075 return (SET_ERROR(ERESTART
));
5077 /* Note: reserve is inflated, so we deflate */
5078 page_load
+= reserve
/ 8;
5080 } else if (page_load
> 0 && arc_reclaim_needed()) {
5081 /* memory is low, delay before restarting */
5082 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
5083 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
5084 return (SET_ERROR(EAGAIN
));
5092 arc_tempreserve_clear(uint64_t reserve
)
5094 atomic_add_64(&arc_tempreserve
, -reserve
);
5095 ASSERT((int64_t)arc_tempreserve
>= 0);
5099 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
5104 if (reserve
> arc_c
/4 && !arc_no_grow
)
5105 arc_c
= MIN(arc_c_max
, reserve
* 4);
5108 * Throttle when the calculated memory footprint for the TXG
5109 * exceeds the target ARC size.
5111 if (reserve
> arc_c
) {
5112 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
5113 return (SET_ERROR(ERESTART
));
5117 * Don't count loaned bufs as in flight dirty data to prevent long
5118 * network delays from blocking transactions that are ready to be
5119 * assigned to a txg.
5121 anon_size
= MAX((int64_t)(refcount_count(&arc_anon
->arcs_size
) -
5122 arc_loaned_bytes
), 0);
5125 * Writes will, almost always, require additional memory allocations
5126 * in order to compress/encrypt/etc the data. We therefore need to
5127 * make sure that there is sufficient available memory for this.
5129 error
= arc_memory_throttle(reserve
, txg
);
5134 * Throttle writes when the amount of dirty data in the cache
5135 * gets too large. We try to keep the cache less than half full
5136 * of dirty blocks so that our sync times don't grow too large.
5137 * Note: if two requests come in concurrently, we might let them
5138 * both succeed, when one of them should fail. Not a huge deal.
5141 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
5142 anon_size
> arc_c
/ 4) {
5143 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5144 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5145 arc_tempreserve
>>10,
5146 arc_anon
->arcs_lsize
[ARC_BUFC_METADATA
]>>10,
5147 arc_anon
->arcs_lsize
[ARC_BUFC_DATA
]>>10,
5148 reserve
>>10, arc_c
>>10);
5149 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
5150 return (SET_ERROR(ERESTART
));
5152 atomic_add_64(&arc_tempreserve
, reserve
);
5157 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
5158 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
5160 size
->value
.ui64
= refcount_count(&state
->arcs_size
);
5161 evict_data
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_DATA
];
5162 evict_metadata
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_METADATA
];
5166 arc_kstat_update(kstat_t
*ksp
, int rw
)
5168 arc_stats_t
*as
= ksp
->ks_data
;
5170 if (rw
== KSTAT_WRITE
) {
5173 arc_kstat_update_state(arc_anon
,
5174 &as
->arcstat_anon_size
,
5175 &as
->arcstat_anon_evictable_data
,
5176 &as
->arcstat_anon_evictable_metadata
);
5177 arc_kstat_update_state(arc_mru
,
5178 &as
->arcstat_mru_size
,
5179 &as
->arcstat_mru_evictable_data
,
5180 &as
->arcstat_mru_evictable_metadata
);
5181 arc_kstat_update_state(arc_mru_ghost
,
5182 &as
->arcstat_mru_ghost_size
,
5183 &as
->arcstat_mru_ghost_evictable_data
,
5184 &as
->arcstat_mru_ghost_evictable_metadata
);
5185 arc_kstat_update_state(arc_mfu
,
5186 &as
->arcstat_mfu_size
,
5187 &as
->arcstat_mfu_evictable_data
,
5188 &as
->arcstat_mfu_evictable_metadata
);
5189 arc_kstat_update_state(arc_mfu_ghost
,
5190 &as
->arcstat_mfu_ghost_size
,
5191 &as
->arcstat_mfu_ghost_evictable_data
,
5192 &as
->arcstat_mfu_ghost_evictable_metadata
);
5199 * This function *must* return indices evenly distributed between all
5200 * sublists of the multilist. This is needed due to how the ARC eviction
5201 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5202 * distributed between all sublists and uses this assumption when
5203 * deciding which sublist to evict from and how much to evict from it.
5206 arc_state_multilist_index_func(multilist_t
*ml
, void *obj
)
5208 arc_buf_hdr_t
*hdr
= obj
;
5211 * We rely on b_dva to generate evenly distributed index
5212 * numbers using buf_hash below. So, as an added precaution,
5213 * let's make sure we never add empty buffers to the arc lists.
5215 ASSERT(!BUF_EMPTY(hdr
));
5218 * The assumption here, is the hash value for a given
5219 * arc_buf_hdr_t will remain constant throughout its lifetime
5220 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
5221 * Thus, we don't need to store the header's sublist index
5222 * on insertion, as this index can be recalculated on removal.
5224 * Also, the low order bits of the hash value are thought to be
5225 * distributed evenly. Otherwise, in the case that the multilist
5226 * has a power of two number of sublists, each sublists' usage
5227 * would not be evenly distributed.
5229 return (buf_hash(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
) %
5230 multilist_get_num_sublists(ml
));
5234 * Called during module initialization and periodically thereafter to
5235 * apply reasonable changes to the exposed performance tunings. Non-zero
5236 * zfs_* values which differ from the currently set values will be applied.
5239 arc_tuning_update(void)
5241 /* Valid range: 64M - <all physical memory> */
5242 if ((zfs_arc_max
) && (zfs_arc_max
!= arc_c_max
) &&
5243 (zfs_arc_max
> 64 << 20) && (zfs_arc_max
< ptob(physmem
)) &&
5244 (zfs_arc_max
> arc_c_min
)) {
5245 arc_c_max
= zfs_arc_max
;
5247 arc_p
= (arc_c
>> 1);
5248 arc_meta_limit
= MIN(arc_meta_limit
, (3 * arc_c_max
) / 4);
5251 /* Valid range: 32M - <arc_c_max> */
5252 if ((zfs_arc_min
) && (zfs_arc_min
!= arc_c_min
) &&
5253 (zfs_arc_min
>= 2ULL << SPA_MAXBLOCKSHIFT
) &&
5254 (zfs_arc_min
<= arc_c_max
)) {
5255 arc_c_min
= zfs_arc_min
;
5256 arc_c
= MAX(arc_c
, arc_c_min
);
5259 /* Valid range: 16M - <arc_c_max> */
5260 if ((zfs_arc_meta_min
) && (zfs_arc_meta_min
!= arc_meta_min
) &&
5261 (zfs_arc_meta_min
>= 1ULL << SPA_MAXBLOCKSHIFT
) &&
5262 (zfs_arc_meta_min
<= arc_c_max
)) {
5263 arc_meta_min
= zfs_arc_meta_min
;
5264 arc_meta_limit
= MAX(arc_meta_limit
, arc_meta_min
);
5267 /* Valid range: <arc_meta_min> - <arc_c_max> */
5268 if ((zfs_arc_meta_limit
) && (zfs_arc_meta_limit
!= arc_meta_limit
) &&
5269 (zfs_arc_meta_limit
>= zfs_arc_meta_min
) &&
5270 (zfs_arc_meta_limit
<= arc_c_max
))
5271 arc_meta_limit
= zfs_arc_meta_limit
;
5273 /* Valid range: 1 - N */
5274 if (zfs_arc_grow_retry
)
5275 arc_grow_retry
= zfs_arc_grow_retry
;
5277 /* Valid range: 1 - N */
5278 if (zfs_arc_shrink_shift
) {
5279 arc_shrink_shift
= zfs_arc_shrink_shift
;
5280 arc_no_grow_shift
= MIN(arc_no_grow_shift
, arc_shrink_shift
-1);
5283 /* Valid range: 1 - N */
5284 if (zfs_arc_p_min_shift
)
5285 arc_p_min_shift
= zfs_arc_p_min_shift
;
5287 /* Valid range: 1 - N ticks */
5288 if (zfs_arc_min_prefetch_lifespan
)
5289 arc_min_prefetch_lifespan
= zfs_arc_min_prefetch_lifespan
;
5291 /* Valid range: 0 - 100 */
5292 if ((zfs_arc_lotsfree_percent
>= 0) &&
5293 (zfs_arc_lotsfree_percent
<= 100))
5294 arc_lotsfree_percent
= zfs_arc_lotsfree_percent
;
5296 /* Valid range: 0 - <all physical memory> */
5297 if ((zfs_arc_sys_free
) && (zfs_arc_sys_free
!= arc_sys_free
))
5298 arc_sys_free
= MIN(MAX(zfs_arc_sys_free
, 0), ptob(physmem
));
5306 * allmem is "all memory that we could possibly use".
5309 uint64_t allmem
= ptob(physmem
);
5311 uint64_t allmem
= (physmem
* PAGESIZE
) / 2;
5314 mutex_init(&arc_reclaim_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
5315 cv_init(&arc_reclaim_thread_cv
, NULL
, CV_DEFAULT
, NULL
);
5316 cv_init(&arc_reclaim_waiters_cv
, NULL
, CV_DEFAULT
, NULL
);
5318 mutex_init(&arc_user_evicts_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
5319 cv_init(&arc_user_evicts_cv
, NULL
, CV_DEFAULT
, NULL
);
5321 /* Convert seconds to clock ticks */
5322 arc_min_prefetch_lifespan
= 1 * hz
;
5324 /* Start out with 1/8 of all memory */
5329 * On architectures where the physical memory can be larger
5330 * than the addressable space (intel in 32-bit mode), we may
5331 * need to limit the cache to 1/8 of VM size.
5333 arc_c
= MIN(arc_c
, vmem_size(heap_arena
, VMEM_ALLOC
| VMEM_FREE
) / 8);
5336 * Register a shrinker to support synchronous (direct) memory
5337 * reclaim from the arc. This is done to prevent kswapd from
5338 * swapping out pages when it is preferable to shrink the arc.
5340 spl_register_shrinker(&arc_shrinker
);
5342 /* Set to 1/64 of all memory or a minimum of 512K */
5343 arc_sys_free
= MAX(ptob(physmem
/ 64), (512 * 1024));
5347 /* Set min cache to allow safe operation of arc_adapt() */
5348 arc_c_min
= 2ULL << SPA_MAXBLOCKSHIFT
;
5349 /* Set max to 1/2 of all memory */
5350 arc_c_max
= allmem
/ 2;
5353 arc_p
= (arc_c
>> 1);
5355 /* Set min to 1/2 of arc_c_min */
5356 arc_meta_min
= 1ULL << SPA_MAXBLOCKSHIFT
;
5357 /* Initialize maximum observed usage to zero */
5359 /* Set limit to 3/4 of arc_c_max with a floor of arc_meta_min */
5360 arc_meta_limit
= MAX((3 * arc_c_max
) / 4, arc_meta_min
);
5362 /* Apply user specified tunings */
5363 arc_tuning_update();
5365 if (zfs_arc_num_sublists_per_state
< 1)
5366 zfs_arc_num_sublists_per_state
= MAX(boot_ncpus
, 1);
5368 /* if kmem_flags are set, lets try to use less memory */
5369 if (kmem_debugging())
5371 if (arc_c
< arc_c_min
)
5374 arc_anon
= &ARC_anon
;
5376 arc_mru_ghost
= &ARC_mru_ghost
;
5378 arc_mfu_ghost
= &ARC_mfu_ghost
;
5379 arc_l2c_only
= &ARC_l2c_only
;
5382 multilist_create(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
],
5383 sizeof (arc_buf_hdr_t
),
5384 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5385 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5386 multilist_create(&arc_mru
->arcs_list
[ARC_BUFC_DATA
],
5387 sizeof (arc_buf_hdr_t
),
5388 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5389 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5390 multilist_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
],
5391 sizeof (arc_buf_hdr_t
),
5392 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5393 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5394 multilist_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
],
5395 sizeof (arc_buf_hdr_t
),
5396 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5397 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5398 multilist_create(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
],
5399 sizeof (arc_buf_hdr_t
),
5400 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5401 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5402 multilist_create(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
],
5403 sizeof (arc_buf_hdr_t
),
5404 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5405 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5406 multilist_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
],
5407 sizeof (arc_buf_hdr_t
),
5408 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5409 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5410 multilist_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
],
5411 sizeof (arc_buf_hdr_t
),
5412 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5413 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5414 multilist_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
],
5415 sizeof (arc_buf_hdr_t
),
5416 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5417 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5418 multilist_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
],
5419 sizeof (arc_buf_hdr_t
),
5420 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5421 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5423 arc_anon
->arcs_state
= ARC_STATE_ANON
;
5424 arc_mru
->arcs_state
= ARC_STATE_MRU
;
5425 arc_mru_ghost
->arcs_state
= ARC_STATE_MRU_GHOST
;
5426 arc_mfu
->arcs_state
= ARC_STATE_MFU
;
5427 arc_mfu_ghost
->arcs_state
= ARC_STATE_MFU_GHOST
;
5428 arc_l2c_only
->arcs_state
= ARC_STATE_L2C_ONLY
;
5430 refcount_create(&arc_anon
->arcs_size
);
5431 refcount_create(&arc_mru
->arcs_size
);
5432 refcount_create(&arc_mru_ghost
->arcs_size
);
5433 refcount_create(&arc_mfu
->arcs_size
);
5434 refcount_create(&arc_mfu_ghost
->arcs_size
);
5435 refcount_create(&arc_l2c_only
->arcs_size
);
5439 arc_reclaim_thread_exit
= FALSE
;
5440 arc_user_evicts_thread_exit
= FALSE
;
5441 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
5442 offsetof(arc_prune_t
, p_node
));
5443 arc_eviction_list
= NULL
;
5444 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
5445 bzero(&arc_eviction_hdr
, sizeof (arc_buf_hdr_t
));
5447 arc_prune_taskq
= taskq_create("arc_prune", max_ncpus
, defclsyspri
,
5448 max_ncpus
, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
5450 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
5451 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
5453 if (arc_ksp
!= NULL
) {
5454 arc_ksp
->ks_data
= &arc_stats
;
5455 arc_ksp
->ks_update
= arc_kstat_update
;
5456 kstat_install(arc_ksp
);
5459 (void) thread_create(NULL
, 0, arc_reclaim_thread
, NULL
, 0, &p0
,
5460 TS_RUN
, defclsyspri
);
5462 (void) thread_create(NULL
, 0, arc_user_evicts_thread
, NULL
, 0, &p0
,
5463 TS_RUN
, defclsyspri
);
5469 * Calculate maximum amount of dirty data per pool.
5471 * If it has been set by a module parameter, take that.
5472 * Otherwise, use a percentage of physical memory defined by
5473 * zfs_dirty_data_max_percent (default 10%) with a cap at
5474 * zfs_dirty_data_max_max (default 25% of physical memory).
5476 if (zfs_dirty_data_max_max
== 0)
5477 zfs_dirty_data_max_max
= physmem
* PAGESIZE
*
5478 zfs_dirty_data_max_max_percent
/ 100;
5480 if (zfs_dirty_data_max
== 0) {
5481 zfs_dirty_data_max
= physmem
* PAGESIZE
*
5482 zfs_dirty_data_max_percent
/ 100;
5483 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
5484 zfs_dirty_data_max_max
);
5494 spl_unregister_shrinker(&arc_shrinker
);
5495 #endif /* _KERNEL */
5497 mutex_enter(&arc_reclaim_lock
);
5498 arc_reclaim_thread_exit
= TRUE
;
5500 * The reclaim thread will set arc_reclaim_thread_exit back to
5501 * FALSE when it is finished exiting; we're waiting for that.
5503 while (arc_reclaim_thread_exit
) {
5504 cv_signal(&arc_reclaim_thread_cv
);
5505 cv_wait(&arc_reclaim_thread_cv
, &arc_reclaim_lock
);
5507 mutex_exit(&arc_reclaim_lock
);
5509 mutex_enter(&arc_user_evicts_lock
);
5510 arc_user_evicts_thread_exit
= TRUE
;
5512 * The user evicts thread will set arc_user_evicts_thread_exit
5513 * to FALSE when it is finished exiting; we're waiting for that.
5515 while (arc_user_evicts_thread_exit
) {
5516 cv_signal(&arc_user_evicts_cv
);
5517 cv_wait(&arc_user_evicts_cv
, &arc_user_evicts_lock
);
5519 mutex_exit(&arc_user_evicts_lock
);
5521 /* Use TRUE to ensure *all* buffers are evicted */
5522 arc_flush(NULL
, TRUE
);
5526 if (arc_ksp
!= NULL
) {
5527 kstat_delete(arc_ksp
);
5531 taskq_wait(arc_prune_taskq
);
5532 taskq_destroy(arc_prune_taskq
);
5534 mutex_enter(&arc_prune_mtx
);
5535 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
5536 list_remove(&arc_prune_list
, p
);
5537 refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
5538 refcount_destroy(&p
->p_refcnt
);
5539 kmem_free(p
, sizeof (*p
));
5541 mutex_exit(&arc_prune_mtx
);
5543 list_destroy(&arc_prune_list
);
5544 mutex_destroy(&arc_prune_mtx
);
5545 mutex_destroy(&arc_reclaim_lock
);
5546 cv_destroy(&arc_reclaim_thread_cv
);
5547 cv_destroy(&arc_reclaim_waiters_cv
);
5549 mutex_destroy(&arc_user_evicts_lock
);
5550 cv_destroy(&arc_user_evicts_cv
);
5552 refcount_destroy(&arc_anon
->arcs_size
);
5553 refcount_destroy(&arc_mru
->arcs_size
);
5554 refcount_destroy(&arc_mru_ghost
->arcs_size
);
5555 refcount_destroy(&arc_mfu
->arcs_size
);
5556 refcount_destroy(&arc_mfu_ghost
->arcs_size
);
5557 refcount_destroy(&arc_l2c_only
->arcs_size
);
5559 multilist_destroy(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
5560 multilist_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
5561 multilist_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
5562 multilist_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
5563 multilist_destroy(&arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
5564 multilist_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
5565 multilist_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
5566 multilist_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
5567 multilist_destroy(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]);
5568 multilist_destroy(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]);
5572 ASSERT0(arc_loaned_bytes
);
5578 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5579 * It uses dedicated storage devices to hold cached data, which are populated
5580 * using large infrequent writes. The main role of this cache is to boost
5581 * the performance of random read workloads. The intended L2ARC devices
5582 * include short-stroked disks, solid state disks, and other media with
5583 * substantially faster read latency than disk.
5585 * +-----------------------+
5587 * +-----------------------+
5590 * l2arc_feed_thread() arc_read()
5594 * +---------------+ |
5596 * +---------------+ |
5601 * +-------+ +-------+
5603 * | cache | | cache |
5604 * +-------+ +-------+
5605 * +=========+ .-----.
5606 * : L2ARC : |-_____-|
5607 * : devices : | Disks |
5608 * +=========+ `-_____-'
5610 * Read requests are satisfied from the following sources, in order:
5613 * 2) vdev cache of L2ARC devices
5615 * 4) vdev cache of disks
5618 * Some L2ARC device types exhibit extremely slow write performance.
5619 * To accommodate for this there are some significant differences between
5620 * the L2ARC and traditional cache design:
5622 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5623 * the ARC behave as usual, freeing buffers and placing headers on ghost
5624 * lists. The ARC does not send buffers to the L2ARC during eviction as
5625 * this would add inflated write latencies for all ARC memory pressure.
5627 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5628 * It does this by periodically scanning buffers from the eviction-end of
5629 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5630 * not already there. It scans until a headroom of buffers is satisfied,
5631 * which itself is a buffer for ARC eviction. If a compressible buffer is
5632 * found during scanning and selected for writing to an L2ARC device, we
5633 * temporarily boost scanning headroom during the next scan cycle to make
5634 * sure we adapt to compression effects (which might significantly reduce
5635 * the data volume we write to L2ARC). The thread that does this is
5636 * l2arc_feed_thread(), illustrated below; example sizes are included to
5637 * provide a better sense of ratio than this diagram:
5640 * +---------------------+----------+
5641 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5642 * +---------------------+----------+ | o L2ARC eligible
5643 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5644 * +---------------------+----------+ |
5645 * 15.9 Gbytes ^ 32 Mbytes |
5647 * l2arc_feed_thread()
5649 * l2arc write hand <--[oooo]--'
5653 * +==============================+
5654 * L2ARC dev |####|#|###|###| |####| ... |
5655 * +==============================+
5658 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5659 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5660 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5661 * safe to say that this is an uncommon case, since buffers at the end of
5662 * the ARC lists have moved there due to inactivity.
5664 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5665 * then the L2ARC simply misses copying some buffers. This serves as a
5666 * pressure valve to prevent heavy read workloads from both stalling the ARC
5667 * with waits and clogging the L2ARC with writes. This also helps prevent
5668 * the potential for the L2ARC to churn if it attempts to cache content too
5669 * quickly, such as during backups of the entire pool.
5671 * 5. After system boot and before the ARC has filled main memory, there are
5672 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5673 * lists can remain mostly static. Instead of searching from tail of these
5674 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5675 * for eligible buffers, greatly increasing its chance of finding them.
5677 * The L2ARC device write speed is also boosted during this time so that
5678 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5679 * there are no L2ARC reads, and no fear of degrading read performance
5680 * through increased writes.
5682 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5683 * the vdev queue can aggregate them into larger and fewer writes. Each
5684 * device is written to in a rotor fashion, sweeping writes through
5685 * available space then repeating.
5687 * 7. The L2ARC does not store dirty content. It never needs to flush
5688 * write buffers back to disk based storage.
5690 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5691 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5693 * The performance of the L2ARC can be tweaked by a number of tunables, which
5694 * may be necessary for different workloads:
5696 * l2arc_write_max max write bytes per interval
5697 * l2arc_write_boost extra write bytes during device warmup
5698 * l2arc_noprefetch skip caching prefetched buffers
5699 * l2arc_nocompress skip compressing buffers
5700 * l2arc_headroom number of max device writes to precache
5701 * l2arc_headroom_boost when we find compressed buffers during ARC
5702 * scanning, we multiply headroom by this
5703 * percentage factor for the next scan cycle,
5704 * since more compressed buffers are likely to
5706 * l2arc_feed_secs seconds between L2ARC writing
5708 * Tunables may be removed or added as future performance improvements are
5709 * integrated, and also may become zpool properties.
5711 * There are three key functions that control how the L2ARC warms up:
5713 * l2arc_write_eligible() check if a buffer is eligible to cache
5714 * l2arc_write_size() calculate how much to write
5715 * l2arc_write_interval() calculate sleep delay between writes
5717 * These three functions determine what to write, how much, and how quickly
5722 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*hdr
)
5725 * A buffer is *not* eligible for the L2ARC if it:
5726 * 1. belongs to a different spa.
5727 * 2. is already cached on the L2ARC.
5728 * 3. has an I/O in progress (it may be an incomplete read).
5729 * 4. is flagged not eligible (zfs property).
5731 if (hdr
->b_spa
!= spa_guid
|| HDR_HAS_L2HDR(hdr
) ||
5732 HDR_IO_IN_PROGRESS(hdr
) || !HDR_L2CACHE(hdr
))
5739 l2arc_write_size(void)
5744 * Make sure our globals have meaningful values in case the user
5747 size
= l2arc_write_max
;
5749 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
5750 "be greater than zero, resetting it to the default (%d)",
5752 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
5755 if (arc_warm
== B_FALSE
)
5756 size
+= l2arc_write_boost
;
5763 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
5765 clock_t interval
, next
, now
;
5768 * If the ARC lists are busy, increase our write rate; if the
5769 * lists are stale, idle back. This is achieved by checking
5770 * how much we previously wrote - if it was more than half of
5771 * what we wanted, schedule the next write much sooner.
5773 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
5774 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
5776 interval
= hz
* l2arc_feed_secs
;
5778 now
= ddi_get_lbolt();
5779 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
5785 * Cycle through L2ARC devices. This is how L2ARC load balances.
5786 * If a device is returned, this also returns holding the spa config lock.
5788 static l2arc_dev_t
*
5789 l2arc_dev_get_next(void)
5791 l2arc_dev_t
*first
, *next
= NULL
;
5794 * Lock out the removal of spas (spa_namespace_lock), then removal
5795 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
5796 * both locks will be dropped and a spa config lock held instead.
5798 mutex_enter(&spa_namespace_lock
);
5799 mutex_enter(&l2arc_dev_mtx
);
5801 /* if there are no vdevs, there is nothing to do */
5802 if (l2arc_ndev
== 0)
5806 next
= l2arc_dev_last
;
5808 /* loop around the list looking for a non-faulted vdev */
5810 next
= list_head(l2arc_dev_list
);
5812 next
= list_next(l2arc_dev_list
, next
);
5814 next
= list_head(l2arc_dev_list
);
5817 /* if we have come back to the start, bail out */
5820 else if (next
== first
)
5823 } while (vdev_is_dead(next
->l2ad_vdev
));
5825 /* if we were unable to find any usable vdevs, return NULL */
5826 if (vdev_is_dead(next
->l2ad_vdev
))
5829 l2arc_dev_last
= next
;
5832 mutex_exit(&l2arc_dev_mtx
);
5835 * Grab the config lock to prevent the 'next' device from being
5836 * removed while we are writing to it.
5839 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
5840 mutex_exit(&spa_namespace_lock
);
5846 * Free buffers that were tagged for destruction.
5849 l2arc_do_free_on_write(void)
5852 l2arc_data_free_t
*df
, *df_prev
;
5854 mutex_enter(&l2arc_free_on_write_mtx
);
5855 buflist
= l2arc_free_on_write
;
5857 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
5858 df_prev
= list_prev(buflist
, df
);
5859 ASSERT(df
->l2df_data
!= NULL
);
5860 ASSERT(df
->l2df_func
!= NULL
);
5861 df
->l2df_func(df
->l2df_data
, df
->l2df_size
);
5862 list_remove(buflist
, df
);
5863 kmem_free(df
, sizeof (l2arc_data_free_t
));
5866 mutex_exit(&l2arc_free_on_write_mtx
);
5870 * A write to a cache device has completed. Update all headers to allow
5871 * reads from these buffers to begin.
5874 l2arc_write_done(zio_t
*zio
)
5876 l2arc_write_callback_t
*cb
;
5879 arc_buf_hdr_t
*head
, *hdr
, *hdr_prev
;
5880 kmutex_t
*hash_lock
;
5881 int64_t bytes_dropped
= 0;
5883 cb
= zio
->io_private
;
5885 dev
= cb
->l2wcb_dev
;
5886 ASSERT(dev
!= NULL
);
5887 head
= cb
->l2wcb_head
;
5888 ASSERT(head
!= NULL
);
5889 buflist
= &dev
->l2ad_buflist
;
5890 ASSERT(buflist
!= NULL
);
5891 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
5892 l2arc_write_callback_t
*, cb
);
5894 if (zio
->io_error
!= 0)
5895 ARCSTAT_BUMP(arcstat_l2_writes_error
);
5898 * All writes completed, or an error was hit.
5901 mutex_enter(&dev
->l2ad_mtx
);
5902 for (hdr
= list_prev(buflist
, head
); hdr
; hdr
= hdr_prev
) {
5903 hdr_prev
= list_prev(buflist
, hdr
);
5905 hash_lock
= HDR_LOCK(hdr
);
5908 * We cannot use mutex_enter or else we can deadlock
5909 * with l2arc_write_buffers (due to swapping the order
5910 * the hash lock and l2ad_mtx are taken).
5912 if (!mutex_tryenter(hash_lock
)) {
5914 * Missed the hash lock. We must retry so we
5915 * don't leave the ARC_FLAG_L2_WRITING bit set.
5917 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry
);
5920 * We don't want to rescan the headers we've
5921 * already marked as having been written out, so
5922 * we reinsert the head node so we can pick up
5923 * where we left off.
5925 list_remove(buflist
, head
);
5926 list_insert_after(buflist
, hdr
, head
);
5928 mutex_exit(&dev
->l2ad_mtx
);
5931 * We wait for the hash lock to become available
5932 * to try and prevent busy waiting, and increase
5933 * the chance we'll be able to acquire the lock
5934 * the next time around.
5936 mutex_enter(hash_lock
);
5937 mutex_exit(hash_lock
);
5942 * We could not have been moved into the arc_l2c_only
5943 * state while in-flight due to our ARC_FLAG_L2_WRITING
5944 * bit being set. Let's just ensure that's being enforced.
5946 ASSERT(HDR_HAS_L1HDR(hdr
));
5949 * We may have allocated a buffer for L2ARC compression,
5950 * we must release it to avoid leaking this data.
5952 l2arc_release_cdata_buf(hdr
);
5954 if (zio
->io_error
!= 0) {
5956 * Error - drop L2ARC entry.
5958 list_remove(buflist
, hdr
);
5959 hdr
->b_flags
&= ~ARC_FLAG_HAS_L2HDR
;
5961 ARCSTAT_INCR(arcstat_l2_asize
, -hdr
->b_l2hdr
.b_asize
);
5962 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
5964 bytes_dropped
+= hdr
->b_l2hdr
.b_asize
;
5965 (void) refcount_remove_many(&dev
->l2ad_alloc
,
5966 hdr
->b_l2hdr
.b_asize
, hdr
);
5970 * Allow ARC to begin reads and ghost list evictions to
5973 hdr
->b_flags
&= ~ARC_FLAG_L2_WRITING
;
5975 mutex_exit(hash_lock
);
5978 atomic_inc_64(&l2arc_writes_done
);
5979 list_remove(buflist
, head
);
5980 ASSERT(!HDR_HAS_L1HDR(head
));
5981 kmem_cache_free(hdr_l2only_cache
, head
);
5982 mutex_exit(&dev
->l2ad_mtx
);
5984 vdev_space_update(dev
->l2ad_vdev
, -bytes_dropped
, 0, 0);
5986 l2arc_do_free_on_write();
5988 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
5992 * A read to a cache device completed. Validate buffer contents before
5993 * handing over to the regular ARC routines.
5996 l2arc_read_done(zio_t
*zio
)
5998 l2arc_read_callback_t
*cb
;
6001 kmutex_t
*hash_lock
;
6004 ASSERT(zio
->io_vd
!= NULL
);
6005 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
6007 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
6009 cb
= zio
->io_private
;
6011 buf
= cb
->l2rcb_buf
;
6012 ASSERT(buf
!= NULL
);
6014 hash_lock
= HDR_LOCK(buf
->b_hdr
);
6015 mutex_enter(hash_lock
);
6017 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
6020 * If the buffer was compressed, decompress it first.
6022 if (cb
->l2rcb_compress
!= ZIO_COMPRESS_OFF
)
6023 l2arc_decompress_zio(zio
, hdr
, cb
->l2rcb_compress
);
6024 ASSERT(zio
->io_data
!= NULL
);
6025 ASSERT3U(zio
->io_size
, ==, hdr
->b_size
);
6026 ASSERT3U(BP_GET_LSIZE(&cb
->l2rcb_bp
), ==, hdr
->b_size
);
6029 * Check this survived the L2ARC journey.
6031 equal
= arc_cksum_equal(buf
);
6032 if (equal
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
6033 mutex_exit(hash_lock
);
6034 zio
->io_private
= buf
;
6035 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
6036 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
6039 mutex_exit(hash_lock
);
6041 * Buffer didn't survive caching. Increment stats and
6042 * reissue to the original storage device.
6044 if (zio
->io_error
!= 0) {
6045 ARCSTAT_BUMP(arcstat_l2_io_error
);
6047 zio
->io_error
= SET_ERROR(EIO
);
6050 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
6053 * If there's no waiter, issue an async i/o to the primary
6054 * storage now. If there *is* a waiter, the caller must
6055 * issue the i/o in a context where it's OK to block.
6057 if (zio
->io_waiter
== NULL
) {
6058 zio_t
*pio
= zio_unique_parent(zio
);
6060 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
6062 zio_nowait(zio_read(pio
, cb
->l2rcb_spa
, &cb
->l2rcb_bp
,
6063 buf
->b_data
, hdr
->b_size
, arc_read_done
, buf
,
6064 zio
->io_priority
, cb
->l2rcb_flags
, &cb
->l2rcb_zb
));
6068 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
6072 * This is the list priority from which the L2ARC will search for pages to
6073 * cache. This is used within loops (0..3) to cycle through lists in the
6074 * desired order. This order can have a significant effect on cache
6077 * Currently the metadata lists are hit first, MFU then MRU, followed by
6078 * the data lists. This function returns a locked list, and also returns
6081 static multilist_sublist_t
*
6082 l2arc_sublist_lock(int list_num
)
6084 multilist_t
*ml
= NULL
;
6087 ASSERT(list_num
>= 0 && list_num
<= 3);
6091 ml
= &arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
6094 ml
= &arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
6097 ml
= &arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
6100 ml
= &arc_mru
->arcs_list
[ARC_BUFC_DATA
];
6105 * Return a randomly-selected sublist. This is acceptable
6106 * because the caller feeds only a little bit of data for each
6107 * call (8MB). Subsequent calls will result in different
6108 * sublists being selected.
6110 idx
= multilist_get_random_index(ml
);
6111 return (multilist_sublist_lock(ml
, idx
));
6115 * Evict buffers from the device write hand to the distance specified in
6116 * bytes. This distance may span populated buffers, it may span nothing.
6117 * This is clearing a region on the L2ARC device ready for writing.
6118 * If the 'all' boolean is set, every buffer is evicted.
6121 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
6124 arc_buf_hdr_t
*hdr
, *hdr_prev
;
6125 kmutex_t
*hash_lock
;
6128 buflist
= &dev
->l2ad_buflist
;
6130 if (!all
&& dev
->l2ad_first
) {
6132 * This is the first sweep through the device. There is
6138 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
6140 * When nearing the end of the device, evict to the end
6141 * before the device write hand jumps to the start.
6143 taddr
= dev
->l2ad_end
;
6145 taddr
= dev
->l2ad_hand
+ distance
;
6147 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
6148 uint64_t, taddr
, boolean_t
, all
);
6151 mutex_enter(&dev
->l2ad_mtx
);
6152 for (hdr
= list_tail(buflist
); hdr
; hdr
= hdr_prev
) {
6153 hdr_prev
= list_prev(buflist
, hdr
);
6155 hash_lock
= HDR_LOCK(hdr
);
6158 * We cannot use mutex_enter or else we can deadlock
6159 * with l2arc_write_buffers (due to swapping the order
6160 * the hash lock and l2ad_mtx are taken).
6162 if (!mutex_tryenter(hash_lock
)) {
6164 * Missed the hash lock. Retry.
6166 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
6167 mutex_exit(&dev
->l2ad_mtx
);
6168 mutex_enter(hash_lock
);
6169 mutex_exit(hash_lock
);
6173 if (HDR_L2_WRITE_HEAD(hdr
)) {
6175 * We hit a write head node. Leave it for
6176 * l2arc_write_done().
6178 list_remove(buflist
, hdr
);
6179 mutex_exit(hash_lock
);
6183 if (!all
&& HDR_HAS_L2HDR(hdr
) &&
6184 (hdr
->b_l2hdr
.b_daddr
> taddr
||
6185 hdr
->b_l2hdr
.b_daddr
< dev
->l2ad_hand
)) {
6187 * We've evicted to the target address,
6188 * or the end of the device.
6190 mutex_exit(hash_lock
);
6194 ASSERT(HDR_HAS_L2HDR(hdr
));
6195 if (!HDR_HAS_L1HDR(hdr
)) {
6196 ASSERT(!HDR_L2_READING(hdr
));
6198 * This doesn't exist in the ARC. Destroy.
6199 * arc_hdr_destroy() will call list_remove()
6200 * and decrement arcstat_l2_size.
6202 arc_change_state(arc_anon
, hdr
, hash_lock
);
6203 arc_hdr_destroy(hdr
);
6205 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_l2c_only
);
6206 ARCSTAT_BUMP(arcstat_l2_evict_l1cached
);
6208 * Invalidate issued or about to be issued
6209 * reads, since we may be about to write
6210 * over this location.
6212 if (HDR_L2_READING(hdr
)) {
6213 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
6214 hdr
->b_flags
|= ARC_FLAG_L2_EVICTED
;
6217 /* Ensure this header has finished being written */
6218 ASSERT(!HDR_L2_WRITING(hdr
));
6219 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
6221 arc_hdr_l2hdr_destroy(hdr
);
6223 mutex_exit(hash_lock
);
6225 mutex_exit(&dev
->l2ad_mtx
);
6229 * Find and write ARC buffers to the L2ARC device.
6231 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6232 * for reading until they have completed writing.
6233 * The headroom_boost is an in-out parameter used to maintain headroom boost
6234 * state between calls to this function.
6236 * Returns the number of bytes actually written (which may be smaller than
6237 * the delta by which the device hand has changed due to alignment).
6240 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
,
6241 boolean_t
*headroom_boost
)
6243 arc_buf_hdr_t
*hdr
, *hdr_prev
, *head
;
6244 uint64_t write_asize
, write_sz
, headroom
, buf_compress_minsz
,
6248 l2arc_write_callback_t
*cb
;
6250 uint64_t guid
= spa_load_guid(spa
);
6252 const boolean_t do_headroom_boost
= *headroom_boost
;
6254 ASSERT(dev
->l2ad_vdev
!= NULL
);
6256 /* Lower the flag now, we might want to raise it again later. */
6257 *headroom_boost
= B_FALSE
;
6260 write_sz
= write_asize
= 0;
6262 head
= kmem_cache_alloc(hdr_l2only_cache
, KM_PUSHPAGE
);
6263 head
->b_flags
|= ARC_FLAG_L2_WRITE_HEAD
;
6264 head
->b_flags
|= ARC_FLAG_HAS_L2HDR
;
6267 * We will want to try to compress buffers that are at least 2x the
6268 * device sector size.
6270 buf_compress_minsz
= 2 << dev
->l2ad_vdev
->vdev_ashift
;
6273 * Copy buffers for L2ARC writing.
6275 for (try = 0; try <= 3; try++) {
6276 multilist_sublist_t
*mls
= l2arc_sublist_lock(try);
6277 uint64_t passed_sz
= 0;
6280 * L2ARC fast warmup.
6282 * Until the ARC is warm and starts to evict, read from the
6283 * head of the ARC lists rather than the tail.
6285 if (arc_warm
== B_FALSE
)
6286 hdr
= multilist_sublist_head(mls
);
6288 hdr
= multilist_sublist_tail(mls
);
6290 headroom
= target_sz
* l2arc_headroom
;
6291 if (do_headroom_boost
)
6292 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
6294 for (; hdr
; hdr
= hdr_prev
) {
6295 kmutex_t
*hash_lock
;
6299 if (arc_warm
== B_FALSE
)
6300 hdr_prev
= multilist_sublist_next(mls
, hdr
);
6302 hdr_prev
= multilist_sublist_prev(mls
, hdr
);
6304 hash_lock
= HDR_LOCK(hdr
);
6305 if (!mutex_tryenter(hash_lock
)) {
6307 * Skip this buffer rather than waiting.
6312 passed_sz
+= hdr
->b_size
;
6313 if (passed_sz
> headroom
) {
6317 mutex_exit(hash_lock
);
6321 if (!l2arc_write_eligible(guid
, hdr
)) {
6322 mutex_exit(hash_lock
);
6327 * Assume that the buffer is not going to be compressed
6328 * and could take more space on disk because of a larger
6331 buf_sz
= hdr
->b_size
;
6332 buf_a_sz
= vdev_psize_to_asize(dev
->l2ad_vdev
, buf_sz
);
6334 if ((write_asize
+ buf_a_sz
) > target_sz
) {
6336 mutex_exit(hash_lock
);
6342 * Insert a dummy header on the buflist so
6343 * l2arc_write_done() can find where the
6344 * write buffers begin without searching.
6346 mutex_enter(&dev
->l2ad_mtx
);
6347 list_insert_head(&dev
->l2ad_buflist
, head
);
6348 mutex_exit(&dev
->l2ad_mtx
);
6351 sizeof (l2arc_write_callback_t
), KM_SLEEP
);
6352 cb
->l2wcb_dev
= dev
;
6353 cb
->l2wcb_head
= head
;
6354 pio
= zio_root(spa
, l2arc_write_done
, cb
,
6359 * Create and add a new L2ARC header.
6361 hdr
->b_l2hdr
.b_dev
= dev
;
6362 hdr
->b_flags
|= ARC_FLAG_L2_WRITING
;
6364 * Temporarily stash the data buffer in b_tmp_cdata.
6365 * The subsequent write step will pick it up from
6366 * there. This is because can't access b_l1hdr.b_buf
6367 * without holding the hash_lock, which we in turn
6368 * can't access without holding the ARC list locks
6369 * (which we want to avoid during compression/writing)
6371 hdr
->b_l2hdr
.b_compress
= ZIO_COMPRESS_OFF
;
6372 hdr
->b_l2hdr
.b_asize
= hdr
->b_size
;
6373 hdr
->b_l2hdr
.b_hits
= 0;
6374 hdr
->b_l1hdr
.b_tmp_cdata
= hdr
->b_l1hdr
.b_buf
->b_data
;
6377 * Explicitly set the b_daddr field to a known
6378 * value which means "invalid address". This
6379 * enables us to differentiate which stage of
6380 * l2arc_write_buffers() the particular header
6381 * is in (e.g. this loop, or the one below).
6382 * ARC_FLAG_L2_WRITING is not enough to make
6383 * this distinction, and we need to know in
6384 * order to do proper l2arc vdev accounting in
6385 * arc_release() and arc_hdr_destroy().
6387 * Note, we can't use a new flag to distinguish
6388 * the two stages because we don't hold the
6389 * header's hash_lock below, in the second stage
6390 * of this function. Thus, we can't simply
6391 * change the b_flags field to denote that the
6392 * IO has been sent. We can change the b_daddr
6393 * field of the L2 portion, though, since we'll
6394 * be holding the l2ad_mtx; which is why we're
6395 * using it to denote the header's state change.
6397 hdr
->b_l2hdr
.b_daddr
= L2ARC_ADDR_UNSET
;
6398 hdr
->b_flags
|= ARC_FLAG_HAS_L2HDR
;
6400 mutex_enter(&dev
->l2ad_mtx
);
6401 list_insert_head(&dev
->l2ad_buflist
, hdr
);
6402 mutex_exit(&dev
->l2ad_mtx
);
6405 * Compute and store the buffer cksum before
6406 * writing. On debug the cksum is verified first.
6408 arc_cksum_verify(hdr
->b_l1hdr
.b_buf
);
6409 arc_cksum_compute(hdr
->b_l1hdr
.b_buf
, B_TRUE
);
6411 mutex_exit(hash_lock
);
6414 write_asize
+= buf_a_sz
;
6417 multilist_sublist_unlock(mls
);
6423 /* No buffers selected for writing? */
6426 ASSERT(!HDR_HAS_L1HDR(head
));
6427 kmem_cache_free(hdr_l2only_cache
, head
);
6431 mutex_enter(&dev
->l2ad_mtx
);
6434 * Note that elsewhere in this file arcstat_l2_asize
6435 * and the used space on l2ad_vdev are updated using b_asize,
6436 * which is not necessarily rounded up to the device block size.
6437 * Too keep accounting consistent we do the same here as well:
6438 * stats_size accumulates the sum of b_asize of the written buffers,
6439 * while write_asize accumulates the sum of b_asize rounded up
6440 * to the device block size.
6441 * The latter sum is used only to validate the corectness of the code.
6447 * Now start writing the buffers. We're starting at the write head
6448 * and work backwards, retracing the course of the buffer selector
6451 for (hdr
= list_prev(&dev
->l2ad_buflist
, head
); hdr
;
6452 hdr
= list_prev(&dev
->l2ad_buflist
, hdr
)) {
6456 * We rely on the L1 portion of the header below, so
6457 * it's invalid for this header to have been evicted out
6458 * of the ghost cache, prior to being written out. The
6459 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6461 ASSERT(HDR_HAS_L1HDR(hdr
));
6464 * We shouldn't need to lock the buffer here, since we flagged
6465 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6466 * take care to only access its L2 cache parameters. In
6467 * particular, hdr->l1hdr.b_buf may be invalid by now due to
6470 hdr
->b_l2hdr
.b_daddr
= dev
->l2ad_hand
;
6472 if ((!l2arc_nocompress
&& HDR_L2COMPRESS(hdr
)) &&
6473 hdr
->b_l2hdr
.b_asize
>= buf_compress_minsz
) {
6474 if (l2arc_compress_buf(hdr
)) {
6476 * If compression succeeded, enable headroom
6477 * boost on the next scan cycle.
6479 *headroom_boost
= B_TRUE
;
6484 * Pick up the buffer data we had previously stashed away
6485 * (and now potentially also compressed).
6487 buf_data
= hdr
->b_l1hdr
.b_tmp_cdata
;
6488 buf_sz
= hdr
->b_l2hdr
.b_asize
;
6491 * We need to do this regardless if buf_sz is zero or
6492 * not, otherwise, when this l2hdr is evicted we'll
6493 * remove a reference that was never added.
6495 (void) refcount_add_many(&dev
->l2ad_alloc
, buf_sz
, hdr
);
6497 /* Compression may have squashed the buffer to zero length. */
6501 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
6502 dev
->l2ad_hand
, buf_sz
, buf_data
, ZIO_CHECKSUM_OFF
,
6503 NULL
, NULL
, ZIO_PRIORITY_ASYNC_WRITE
,
6504 ZIO_FLAG_CANFAIL
, B_FALSE
);
6506 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
6508 (void) zio_nowait(wzio
);
6510 stats_size
+= buf_sz
;
6513 * Keep the clock hand suitably device-aligned.
6515 buf_a_sz
= vdev_psize_to_asize(dev
->l2ad_vdev
, buf_sz
);
6516 write_asize
+= buf_a_sz
;
6517 dev
->l2ad_hand
+= buf_a_sz
;
6521 mutex_exit(&dev
->l2ad_mtx
);
6523 ASSERT3U(write_asize
, <=, target_sz
);
6524 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
6525 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_asize
);
6526 ARCSTAT_INCR(arcstat_l2_size
, write_sz
);
6527 ARCSTAT_INCR(arcstat_l2_asize
, stats_size
);
6528 vdev_space_update(dev
->l2ad_vdev
, stats_size
, 0, 0);
6531 * Bump device hand to the device start if it is approaching the end.
6532 * l2arc_evict() will already have evicted ahead for this case.
6534 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
6535 dev
->l2ad_hand
= dev
->l2ad_start
;
6536 dev
->l2ad_first
= B_FALSE
;
6539 dev
->l2ad_writing
= B_TRUE
;
6540 (void) zio_wait(pio
);
6541 dev
->l2ad_writing
= B_FALSE
;
6543 return (write_asize
);
6547 * Compresses an L2ARC buffer.
6548 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6549 * size in l2hdr->b_asize. This routine tries to compress the data and
6550 * depending on the compression result there are three possible outcomes:
6551 * *) The buffer was incompressible. The original l2hdr contents were left
6552 * untouched and are ready for writing to an L2 device.
6553 * *) The buffer was all-zeros, so there is no need to write it to an L2
6554 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6555 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6556 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6557 * data buffer which holds the compressed data to be written, and b_asize
6558 * tells us how much data there is. b_compress is set to the appropriate
6559 * compression algorithm. Once writing is done, invoke
6560 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6562 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6563 * buffer was incompressible).
6566 l2arc_compress_buf(arc_buf_hdr_t
*hdr
)
6569 size_t csize
, len
, rounded
;
6570 l2arc_buf_hdr_t
*l2hdr
;
6572 ASSERT(HDR_HAS_L2HDR(hdr
));
6574 l2hdr
= &hdr
->b_l2hdr
;
6576 ASSERT(HDR_HAS_L1HDR(hdr
));
6577 ASSERT3U(l2hdr
->b_compress
, ==, ZIO_COMPRESS_OFF
);
6578 ASSERT(hdr
->b_l1hdr
.b_tmp_cdata
!= NULL
);
6580 len
= l2hdr
->b_asize
;
6581 cdata
= zio_data_buf_alloc(len
);
6582 ASSERT3P(cdata
, !=, NULL
);
6583 csize
= zio_compress_data(ZIO_COMPRESS_LZ4
, hdr
->b_l1hdr
.b_tmp_cdata
,
6584 cdata
, l2hdr
->b_asize
);
6586 rounded
= P2ROUNDUP(csize
, (size_t)SPA_MINBLOCKSIZE
);
6587 if (rounded
> csize
) {
6588 bzero((char *)cdata
+ csize
, rounded
- csize
);
6593 /* zero block, indicate that there's nothing to write */
6594 zio_data_buf_free(cdata
, len
);
6595 l2hdr
->b_compress
= ZIO_COMPRESS_EMPTY
;
6597 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6598 ARCSTAT_BUMP(arcstat_l2_compress_zeros
);
6600 } else if (csize
> 0 && csize
< len
) {
6602 * Compression succeeded, we'll keep the cdata around for
6603 * writing and release it afterwards.
6605 l2hdr
->b_compress
= ZIO_COMPRESS_LZ4
;
6606 l2hdr
->b_asize
= csize
;
6607 hdr
->b_l1hdr
.b_tmp_cdata
= cdata
;
6608 ARCSTAT_BUMP(arcstat_l2_compress_successes
);
6612 * Compression failed, release the compressed buffer.
6613 * l2hdr will be left unmodified.
6615 zio_data_buf_free(cdata
, len
);
6616 ARCSTAT_BUMP(arcstat_l2_compress_failures
);
6622 * Decompresses a zio read back from an l2arc device. On success, the
6623 * underlying zio's io_data buffer is overwritten by the uncompressed
6624 * version. On decompression error (corrupt compressed stream), the
6625 * zio->io_error value is set to signal an I/O error.
6627 * Please note that the compressed data stream is not checksummed, so
6628 * if the underlying device is experiencing data corruption, we may feed
6629 * corrupt data to the decompressor, so the decompressor needs to be
6630 * able to handle this situation (LZ4 does).
6633 l2arc_decompress_zio(zio_t
*zio
, arc_buf_hdr_t
*hdr
, enum zio_compress c
)
6638 ASSERT(L2ARC_IS_VALID_COMPRESS(c
));
6640 if (zio
->io_error
!= 0) {
6642 * An io error has occured, just restore the original io
6643 * size in preparation for a main pool read.
6645 zio
->io_orig_size
= zio
->io_size
= hdr
->b_size
;
6649 if (c
== ZIO_COMPRESS_EMPTY
) {
6651 * An empty buffer results in a null zio, which means we
6652 * need to fill its io_data after we're done restoring the
6653 * buffer's contents.
6655 ASSERT(hdr
->b_l1hdr
.b_buf
!= NULL
);
6656 bzero(hdr
->b_l1hdr
.b_buf
->b_data
, hdr
->b_size
);
6657 zio
->io_data
= zio
->io_orig_data
= hdr
->b_l1hdr
.b_buf
->b_data
;
6659 ASSERT(zio
->io_data
!= NULL
);
6661 * We copy the compressed data from the start of the arc buffer
6662 * (the zio_read will have pulled in only what we need, the
6663 * rest is garbage which we will overwrite at decompression)
6664 * and then decompress back to the ARC data buffer. This way we
6665 * can minimize copying by simply decompressing back over the
6666 * original compressed data (rather than decompressing to an
6667 * aux buffer and then copying back the uncompressed buffer,
6668 * which is likely to be much larger).
6670 csize
= zio
->io_size
;
6671 cdata
= zio_data_buf_alloc(csize
);
6672 bcopy(zio
->io_data
, cdata
, csize
);
6673 if (zio_decompress_data(c
, cdata
, zio
->io_data
, csize
,
6675 zio
->io_error
= EIO
;
6676 zio_data_buf_free(cdata
, csize
);
6679 /* Restore the expected uncompressed IO size. */
6680 zio
->io_orig_size
= zio
->io_size
= hdr
->b_size
;
6684 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6685 * This buffer serves as a temporary holder of compressed data while
6686 * the buffer entry is being written to an l2arc device. Once that is
6687 * done, we can dispose of it.
6690 l2arc_release_cdata_buf(arc_buf_hdr_t
*hdr
)
6692 enum zio_compress comp
;
6694 ASSERT(HDR_HAS_L1HDR(hdr
));
6695 ASSERT(HDR_HAS_L2HDR(hdr
));
6696 comp
= hdr
->b_l2hdr
.b_compress
;
6697 ASSERT(comp
== ZIO_COMPRESS_OFF
|| L2ARC_IS_VALID_COMPRESS(comp
));
6699 if (comp
== ZIO_COMPRESS_OFF
) {
6701 * In this case, b_tmp_cdata points to the same buffer
6702 * as the arc_buf_t's b_data field. We don't want to
6703 * free it, since the arc_buf_t will handle that.
6705 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6706 } else if (comp
== ZIO_COMPRESS_EMPTY
) {
6708 * In this case, b_tmp_cdata was compressed to an empty
6709 * buffer, thus there's nothing to free and b_tmp_cdata
6710 * should have been set to NULL in l2arc_write_buffers().
6712 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
6715 * If the data was compressed, then we've allocated a
6716 * temporary buffer for it, so now we need to release it.
6718 ASSERT(hdr
->b_l1hdr
.b_tmp_cdata
!= NULL
);
6719 zio_data_buf_free(hdr
->b_l1hdr
.b_tmp_cdata
,
6721 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6727 * This thread feeds the L2ARC at regular intervals. This is the beating
6728 * heart of the L2ARC.
6731 l2arc_feed_thread(void)
6736 uint64_t size
, wrote
;
6737 clock_t begin
, next
= ddi_get_lbolt();
6738 boolean_t headroom_boost
= B_FALSE
;
6739 fstrans_cookie_t cookie
;
6741 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
6743 mutex_enter(&l2arc_feed_thr_lock
);
6745 cookie
= spl_fstrans_mark();
6746 while (l2arc_thread_exit
== 0) {
6747 CALLB_CPR_SAFE_BEGIN(&cpr
);
6748 (void) cv_timedwait_sig(&l2arc_feed_thr_cv
,
6749 &l2arc_feed_thr_lock
, next
);
6750 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
6751 next
= ddi_get_lbolt() + hz
;
6754 * Quick check for L2ARC devices.
6756 mutex_enter(&l2arc_dev_mtx
);
6757 if (l2arc_ndev
== 0) {
6758 mutex_exit(&l2arc_dev_mtx
);
6761 mutex_exit(&l2arc_dev_mtx
);
6762 begin
= ddi_get_lbolt();
6765 * This selects the next l2arc device to write to, and in
6766 * doing so the next spa to feed from: dev->l2ad_spa. This
6767 * will return NULL if there are now no l2arc devices or if
6768 * they are all faulted.
6770 * If a device is returned, its spa's config lock is also
6771 * held to prevent device removal. l2arc_dev_get_next()
6772 * will grab and release l2arc_dev_mtx.
6774 if ((dev
= l2arc_dev_get_next()) == NULL
)
6777 spa
= dev
->l2ad_spa
;
6778 ASSERT(spa
!= NULL
);
6781 * If the pool is read-only then force the feed thread to
6782 * sleep a little longer.
6784 if (!spa_writeable(spa
)) {
6785 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
6786 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6791 * Avoid contributing to memory pressure.
6793 if (arc_reclaim_needed()) {
6794 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
6795 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6799 ARCSTAT_BUMP(arcstat_l2_feeds
);
6801 size
= l2arc_write_size();
6804 * Evict L2ARC buffers that will be overwritten.
6806 l2arc_evict(dev
, size
, B_FALSE
);
6809 * Write ARC buffers.
6811 wrote
= l2arc_write_buffers(spa
, dev
, size
, &headroom_boost
);
6814 * Calculate interval between writes.
6816 next
= l2arc_write_interval(begin
, size
, wrote
);
6817 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6819 spl_fstrans_unmark(cookie
);
6821 l2arc_thread_exit
= 0;
6822 cv_broadcast(&l2arc_feed_thr_cv
);
6823 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
6828 l2arc_vdev_present(vdev_t
*vd
)
6832 mutex_enter(&l2arc_dev_mtx
);
6833 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
6834 dev
= list_next(l2arc_dev_list
, dev
)) {
6835 if (dev
->l2ad_vdev
== vd
)
6838 mutex_exit(&l2arc_dev_mtx
);
6840 return (dev
!= NULL
);
6844 * Add a vdev for use by the L2ARC. By this point the spa has already
6845 * validated the vdev and opened it.
6848 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
6850 l2arc_dev_t
*adddev
;
6852 ASSERT(!l2arc_vdev_present(vd
));
6855 * Create a new l2arc device entry.
6857 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
6858 adddev
->l2ad_spa
= spa
;
6859 adddev
->l2ad_vdev
= vd
;
6860 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
6861 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
6862 adddev
->l2ad_hand
= adddev
->l2ad_start
;
6863 adddev
->l2ad_first
= B_TRUE
;
6864 adddev
->l2ad_writing
= B_FALSE
;
6865 list_link_init(&adddev
->l2ad_node
);
6867 mutex_init(&adddev
->l2ad_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6869 * This is a list of all ARC buffers that are still valid on the
6872 list_create(&adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
6873 offsetof(arc_buf_hdr_t
, b_l2hdr
.b_l2node
));
6875 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
6876 refcount_create(&adddev
->l2ad_alloc
);
6879 * Add device to global list
6881 mutex_enter(&l2arc_dev_mtx
);
6882 list_insert_head(l2arc_dev_list
, adddev
);
6883 atomic_inc_64(&l2arc_ndev
);
6884 mutex_exit(&l2arc_dev_mtx
);
6888 * Remove a vdev from the L2ARC.
6891 l2arc_remove_vdev(vdev_t
*vd
)
6893 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
6896 * Find the device by vdev
6898 mutex_enter(&l2arc_dev_mtx
);
6899 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
6900 nextdev
= list_next(l2arc_dev_list
, dev
);
6901 if (vd
== dev
->l2ad_vdev
) {
6906 ASSERT(remdev
!= NULL
);
6909 * Remove device from global list
6911 list_remove(l2arc_dev_list
, remdev
);
6912 l2arc_dev_last
= NULL
; /* may have been invalidated */
6913 atomic_dec_64(&l2arc_ndev
);
6914 mutex_exit(&l2arc_dev_mtx
);
6917 * Clear all buflists and ARC references. L2ARC device flush.
6919 l2arc_evict(remdev
, 0, B_TRUE
);
6920 list_destroy(&remdev
->l2ad_buflist
);
6921 mutex_destroy(&remdev
->l2ad_mtx
);
6922 refcount_destroy(&remdev
->l2ad_alloc
);
6923 kmem_free(remdev
, sizeof (l2arc_dev_t
));
6929 l2arc_thread_exit
= 0;
6931 l2arc_writes_sent
= 0;
6932 l2arc_writes_done
= 0;
6934 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
6935 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
6936 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6937 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6939 l2arc_dev_list
= &L2ARC_dev_list
;
6940 l2arc_free_on_write
= &L2ARC_free_on_write
;
6941 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
6942 offsetof(l2arc_dev_t
, l2ad_node
));
6943 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
6944 offsetof(l2arc_data_free_t
, l2df_list_node
));
6951 * This is called from dmu_fini(), which is called from spa_fini();
6952 * Because of this, we can assume that all l2arc devices have
6953 * already been removed when the pools themselves were removed.
6956 l2arc_do_free_on_write();
6958 mutex_destroy(&l2arc_feed_thr_lock
);
6959 cv_destroy(&l2arc_feed_thr_cv
);
6960 mutex_destroy(&l2arc_dev_mtx
);
6961 mutex_destroy(&l2arc_free_on_write_mtx
);
6963 list_destroy(l2arc_dev_list
);
6964 list_destroy(l2arc_free_on_write
);
6970 if (!(spa_mode_global
& FWRITE
))
6973 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
6974 TS_RUN
, defclsyspri
);
6980 if (!(spa_mode_global
& FWRITE
))
6983 mutex_enter(&l2arc_feed_thr_lock
);
6984 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
6985 l2arc_thread_exit
= 1;
6986 while (l2arc_thread_exit
!= 0)
6987 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
6988 mutex_exit(&l2arc_feed_thr_lock
);
6991 #if defined(_KERNEL) && defined(HAVE_SPL)
6992 EXPORT_SYMBOL(arc_buf_size
);
6993 EXPORT_SYMBOL(arc_write
);
6994 EXPORT_SYMBOL(arc_read
);
6995 EXPORT_SYMBOL(arc_buf_remove_ref
);
6996 EXPORT_SYMBOL(arc_buf_info
);
6997 EXPORT_SYMBOL(arc_getbuf_func
);
6998 EXPORT_SYMBOL(arc_add_prune_callback
);
6999 EXPORT_SYMBOL(arc_remove_prune_callback
);
7001 module_param(zfs_arc_min
, ulong
, 0644);
7002 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
7004 module_param(zfs_arc_max
, ulong
, 0644);
7005 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
7007 module_param(zfs_arc_meta_limit
, ulong
, 0644);
7008 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
7010 module_param(zfs_arc_meta_min
, ulong
, 0644);
7011 MODULE_PARM_DESC(zfs_arc_meta_min
, "Min arc metadata");
7013 module_param(zfs_arc_meta_prune
, int, 0644);
7014 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Meta objects to scan for prune");
7016 module_param(zfs_arc_meta_adjust_restarts
, int, 0644);
7017 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts
,
7018 "Limit number of restarts in arc_adjust_meta");
7020 module_param(zfs_arc_meta_strategy
, int, 0644);
7021 MODULE_PARM_DESC(zfs_arc_meta_strategy
, "Meta reclaim strategy");
7023 module_param(zfs_arc_grow_retry
, int, 0644);
7024 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
7026 module_param(zfs_arc_p_aggressive_disable
, int, 0644);
7027 MODULE_PARM_DESC(zfs_arc_p_aggressive_disable
, "disable aggressive arc_p grow");
7029 module_param(zfs_arc_p_dampener_disable
, int, 0644);
7030 MODULE_PARM_DESC(zfs_arc_p_dampener_disable
, "disable arc_p adapt dampener");
7032 module_param(zfs_arc_shrink_shift
, int, 0644);
7033 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
7035 module_param(zfs_arc_p_min_shift
, int, 0644);
7036 MODULE_PARM_DESC(zfs_arc_p_min_shift
, "arc_c shift to calc min/max arc_p");
7038 module_param(zfs_disable_dup_eviction
, int, 0644);
7039 MODULE_PARM_DESC(zfs_disable_dup_eviction
, "disable duplicate buffer eviction");
7041 module_param(zfs_arc_average_blocksize
, int, 0444);
7042 MODULE_PARM_DESC(zfs_arc_average_blocksize
, "Target average block size");
7044 module_param(zfs_arc_min_prefetch_lifespan
, int, 0644);
7045 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan
, "Min life of prefetch block");
7047 module_param(zfs_arc_num_sublists_per_state
, int, 0644);
7048 MODULE_PARM_DESC(zfs_arc_num_sublists_per_state
,
7049 "Number of sublists used in each of the ARC state lists");
7051 module_param(l2arc_write_max
, ulong
, 0644);
7052 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
7054 module_param(l2arc_write_boost
, ulong
, 0644);
7055 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
7057 module_param(l2arc_headroom
, ulong
, 0644);
7058 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
7060 module_param(l2arc_headroom_boost
, ulong
, 0644);
7061 MODULE_PARM_DESC(l2arc_headroom_boost
, "Compressed l2arc_headroom multiplier");
7063 module_param(l2arc_feed_secs
, ulong
, 0644);
7064 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
7066 module_param(l2arc_feed_min_ms
, ulong
, 0644);
7067 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
7069 module_param(l2arc_noprefetch
, int, 0644);
7070 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
7072 module_param(l2arc_nocompress
, int, 0644);
7073 MODULE_PARM_DESC(l2arc_nocompress
, "Skip compressing L2ARC buffers");
7075 module_param(l2arc_feed_again
, int, 0644);
7076 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
7078 module_param(l2arc_norw
, int, 0644);
7079 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");
7081 module_param(zfs_arc_lotsfree_percent
, int, 0644);
7082 MODULE_PARM_DESC(zfs_arc_lotsfree_percent
,
7083 "System free memory I/O throttle in bytes");
7085 module_param(zfs_arc_sys_free
, ulong
, 0644);
7086 MODULE_PARM_DESC(zfs_arc_sys_free
, "System free memory target size in bytes");