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_sync_wait_for_async
;
478 kstat_named_t arcstat_demand_hit_predictive_prefetch
;
479 kstat_named_t arcstat_need_free
;
480 kstat_named_t arcstat_sys_free
;
483 static arc_stats_t arc_stats
= {
484 { "hits", KSTAT_DATA_UINT64
},
485 { "misses", KSTAT_DATA_UINT64
},
486 { "demand_data_hits", KSTAT_DATA_UINT64
},
487 { "demand_data_misses", KSTAT_DATA_UINT64
},
488 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
489 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
490 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
491 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
492 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
493 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
494 { "mru_hits", KSTAT_DATA_UINT64
},
495 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
496 { "mfu_hits", KSTAT_DATA_UINT64
},
497 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
498 { "deleted", KSTAT_DATA_UINT64
},
499 { "mutex_miss", KSTAT_DATA_UINT64
},
500 { "evict_skip", KSTAT_DATA_UINT64
},
501 { "evict_not_enough", KSTAT_DATA_UINT64
},
502 { "evict_l2_cached", KSTAT_DATA_UINT64
},
503 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
504 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
505 { "evict_l2_skip", KSTAT_DATA_UINT64
},
506 { "hash_elements", KSTAT_DATA_UINT64
},
507 { "hash_elements_max", KSTAT_DATA_UINT64
},
508 { "hash_collisions", KSTAT_DATA_UINT64
},
509 { "hash_chains", KSTAT_DATA_UINT64
},
510 { "hash_chain_max", KSTAT_DATA_UINT64
},
511 { "p", KSTAT_DATA_UINT64
},
512 { "c", KSTAT_DATA_UINT64
},
513 { "c_min", KSTAT_DATA_UINT64
},
514 { "c_max", KSTAT_DATA_UINT64
},
515 { "size", KSTAT_DATA_UINT64
},
516 { "hdr_size", KSTAT_DATA_UINT64
},
517 { "data_size", KSTAT_DATA_UINT64
},
518 { "metadata_size", KSTAT_DATA_UINT64
},
519 { "other_size", KSTAT_DATA_UINT64
},
520 { "anon_size", KSTAT_DATA_UINT64
},
521 { "anon_evictable_data", KSTAT_DATA_UINT64
},
522 { "anon_evictable_metadata", KSTAT_DATA_UINT64
},
523 { "mru_size", KSTAT_DATA_UINT64
},
524 { "mru_evictable_data", KSTAT_DATA_UINT64
},
525 { "mru_evictable_metadata", KSTAT_DATA_UINT64
},
526 { "mru_ghost_size", KSTAT_DATA_UINT64
},
527 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64
},
528 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
529 { "mfu_size", KSTAT_DATA_UINT64
},
530 { "mfu_evictable_data", KSTAT_DATA_UINT64
},
531 { "mfu_evictable_metadata", KSTAT_DATA_UINT64
},
532 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
533 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64
},
534 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64
},
535 { "l2_hits", KSTAT_DATA_UINT64
},
536 { "l2_misses", KSTAT_DATA_UINT64
},
537 { "l2_feeds", KSTAT_DATA_UINT64
},
538 { "l2_rw_clash", KSTAT_DATA_UINT64
},
539 { "l2_read_bytes", KSTAT_DATA_UINT64
},
540 { "l2_write_bytes", KSTAT_DATA_UINT64
},
541 { "l2_writes_sent", KSTAT_DATA_UINT64
},
542 { "l2_writes_done", KSTAT_DATA_UINT64
},
543 { "l2_writes_error", KSTAT_DATA_UINT64
},
544 { "l2_writes_lock_retry", KSTAT_DATA_UINT64
},
545 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
546 { "l2_evict_reading", KSTAT_DATA_UINT64
},
547 { "l2_evict_l1cached", KSTAT_DATA_UINT64
},
548 { "l2_free_on_write", KSTAT_DATA_UINT64
},
549 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64
},
550 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
551 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
552 { "l2_io_error", KSTAT_DATA_UINT64
},
553 { "l2_size", KSTAT_DATA_UINT64
},
554 { "l2_asize", KSTAT_DATA_UINT64
},
555 { "l2_hdr_size", KSTAT_DATA_UINT64
},
556 { "l2_compress_successes", KSTAT_DATA_UINT64
},
557 { "l2_compress_zeros", KSTAT_DATA_UINT64
},
558 { "l2_compress_failures", KSTAT_DATA_UINT64
},
559 { "memory_throttle_count", KSTAT_DATA_UINT64
},
560 { "duplicate_buffers", KSTAT_DATA_UINT64
},
561 { "duplicate_buffers_size", KSTAT_DATA_UINT64
},
562 { "duplicate_reads", KSTAT_DATA_UINT64
},
563 { "memory_direct_count", KSTAT_DATA_UINT64
},
564 { "memory_indirect_count", KSTAT_DATA_UINT64
},
565 { "arc_no_grow", KSTAT_DATA_UINT64
},
566 { "arc_tempreserve", KSTAT_DATA_UINT64
},
567 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
568 { "arc_prune", KSTAT_DATA_UINT64
},
569 { "arc_meta_used", KSTAT_DATA_UINT64
},
570 { "arc_meta_limit", KSTAT_DATA_UINT64
},
571 { "arc_meta_max", KSTAT_DATA_UINT64
},
572 { "arc_meta_min", KSTAT_DATA_UINT64
},
573 { "sync_wait_for_async", KSTAT_DATA_UINT64
},
574 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64
},
575 { "arc_need_free", KSTAT_DATA_UINT64
},
576 { "arc_sys_free", KSTAT_DATA_UINT64
}
579 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
581 #define ARCSTAT_INCR(stat, val) \
582 atomic_add_64(&arc_stats.stat.value.ui64, (val))
584 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
585 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
587 #define ARCSTAT_MAX(stat, val) { \
589 while ((val) > (m = arc_stats.stat.value.ui64) && \
590 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
594 #define ARCSTAT_MAXSTAT(stat) \
595 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
598 * We define a macro to allow ARC hits/misses to be easily broken down by
599 * two separate conditions, giving a total of four different subtypes for
600 * each of hits and misses (so eight statistics total).
602 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
605 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
607 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
611 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
613 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
618 static arc_state_t
*arc_anon
;
619 static arc_state_t
*arc_mru
;
620 static arc_state_t
*arc_mru_ghost
;
621 static arc_state_t
*arc_mfu
;
622 static arc_state_t
*arc_mfu_ghost
;
623 static arc_state_t
*arc_l2c_only
;
626 * There are several ARC variables that are critical to export as kstats --
627 * but we don't want to have to grovel around in the kstat whenever we wish to
628 * manipulate them. For these variables, we therefore define them to be in
629 * terms of the statistic variable. This assures that we are not introducing
630 * the possibility of inconsistency by having shadow copies of the variables,
631 * while still allowing the code to be readable.
633 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
634 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
635 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
636 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
637 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
638 #define arc_no_grow ARCSTAT(arcstat_no_grow)
639 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
640 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
641 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
642 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
643 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
644 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
645 #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */
646 #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */
648 #define L2ARC_IS_VALID_COMPRESS(_c_) \
649 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
651 static list_t arc_prune_list
;
652 static kmutex_t arc_prune_mtx
;
653 static taskq_t
*arc_prune_taskq
;
654 static arc_buf_t
*arc_eviction_list
;
655 static arc_buf_hdr_t arc_eviction_hdr
;
657 #define GHOST_STATE(state) \
658 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
659 (state) == arc_l2c_only)
661 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
662 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
663 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
664 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
665 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
666 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
668 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
669 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
670 #define HDR_L2_READING(hdr) \
671 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
672 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
673 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
674 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
675 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
677 #define HDR_ISTYPE_METADATA(hdr) \
678 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
679 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
681 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
682 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
688 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
689 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
692 * Hash table routines
695 #define HT_LOCK_ALIGN 64
696 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
701 unsigned char pad
[HT_LOCK_PAD
];
705 #define BUF_LOCKS 8192
706 typedef struct buf_hash_table
{
708 arc_buf_hdr_t
**ht_table
;
709 struct ht_lock ht_locks
[BUF_LOCKS
];
712 static buf_hash_table_t buf_hash_table
;
714 #define BUF_HASH_INDEX(spa, dva, birth) \
715 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
716 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
717 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
718 #define HDR_LOCK(hdr) \
719 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
721 uint64_t zfs_crc64_table
[256];
727 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
728 #define L2ARC_HEADROOM 2 /* num of writes */
730 * If we discover during ARC scan any buffers to be compressed, we boost
731 * our headroom for the next scanning cycle by this percentage multiple.
733 #define L2ARC_HEADROOM_BOOST 200
734 #define L2ARC_FEED_SECS 1 /* caching interval secs */
735 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
738 * Used to distinguish headers that are being process by
739 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
740 * address. This can happen when the header is added to the l2arc's list
741 * of buffers to write in the first stage of l2arc_write_buffers(), but
742 * has not yet been written out which happens in the second stage of
743 * l2arc_write_buffers().
745 #define L2ARC_ADDR_UNSET ((uint64_t)(-1))
747 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
748 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
750 /* L2ARC Performance Tunables */
751 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
752 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
753 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
754 unsigned long l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
755 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
756 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
757 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
758 int l2arc_nocompress
= B_FALSE
; /* don't compress bufs */
759 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
760 int l2arc_norw
= B_FALSE
; /* no reads during writes */
765 static list_t L2ARC_dev_list
; /* device list */
766 static list_t
*l2arc_dev_list
; /* device list pointer */
767 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
768 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
769 static list_t L2ARC_free_on_write
; /* free after write buf list */
770 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
771 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
772 static uint64_t l2arc_ndev
; /* number of devices */
774 typedef struct l2arc_read_callback
{
775 arc_buf_t
*l2rcb_buf
; /* read buffer */
776 spa_t
*l2rcb_spa
; /* spa */
777 blkptr_t l2rcb_bp
; /* original blkptr */
778 zbookmark_phys_t l2rcb_zb
; /* original bookmark */
779 int l2rcb_flags
; /* original flags */
780 enum zio_compress l2rcb_compress
; /* applied compress */
781 } l2arc_read_callback_t
;
783 typedef struct l2arc_data_free
{
784 /* protected by l2arc_free_on_write_mtx */
787 void (*l2df_func
)(void *, size_t);
788 list_node_t l2df_list_node
;
791 static kmutex_t l2arc_feed_thr_lock
;
792 static kcondvar_t l2arc_feed_thr_cv
;
793 static uint8_t l2arc_thread_exit
;
795 static void arc_get_data_buf(arc_buf_t
*);
796 static void arc_access(arc_buf_hdr_t
*, kmutex_t
*);
797 static boolean_t
arc_is_overflowing(void);
798 static void arc_buf_watch(arc_buf_t
*);
799 static void arc_tuning_update(void);
801 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t
*);
802 static uint32_t arc_bufc_to_flags(arc_buf_contents_t
);
804 static boolean_t
l2arc_write_eligible(uint64_t, arc_buf_hdr_t
*);
805 static void l2arc_read_done(zio_t
*);
807 static boolean_t
l2arc_compress_buf(arc_buf_hdr_t
*);
808 static void l2arc_decompress_zio(zio_t
*, arc_buf_hdr_t
*, enum zio_compress
);
809 static void l2arc_release_cdata_buf(arc_buf_hdr_t
*);
812 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
814 uint8_t *vdva
= (uint8_t *)dva
;
815 uint64_t crc
= -1ULL;
818 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
820 for (i
= 0; i
< sizeof (dva_t
); i
++)
821 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
823 crc
^= (spa
>>8) ^ birth
;
828 #define BUF_EMPTY(buf) \
829 ((buf)->b_dva.dva_word[0] == 0 && \
830 (buf)->b_dva.dva_word[1] == 0)
832 #define BUF_EQUAL(spa, dva, birth, buf) \
833 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
834 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
835 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
838 buf_discard_identity(arc_buf_hdr_t
*hdr
)
840 hdr
->b_dva
.dva_word
[0] = 0;
841 hdr
->b_dva
.dva_word
[1] = 0;
845 static arc_buf_hdr_t
*
846 buf_hash_find(uint64_t spa
, const blkptr_t
*bp
, kmutex_t
**lockp
)
848 const dva_t
*dva
= BP_IDENTITY(bp
);
849 uint64_t birth
= BP_PHYSICAL_BIRTH(bp
);
850 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
851 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
854 mutex_enter(hash_lock
);
855 for (hdr
= buf_hash_table
.ht_table
[idx
]; hdr
!= NULL
;
856 hdr
= hdr
->b_hash_next
) {
857 if (BUF_EQUAL(spa
, dva
, birth
, hdr
)) {
862 mutex_exit(hash_lock
);
868 * Insert an entry into the hash table. If there is already an element
869 * equal to elem in the hash table, then the already existing element
870 * will be returned and the new element will not be inserted.
871 * Otherwise returns NULL.
872 * If lockp == NULL, the caller is assumed to already hold the hash lock.
874 static arc_buf_hdr_t
*
875 buf_hash_insert(arc_buf_hdr_t
*hdr
, kmutex_t
**lockp
)
877 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
878 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
882 ASSERT(!DVA_IS_EMPTY(&hdr
->b_dva
));
883 ASSERT(hdr
->b_birth
!= 0);
884 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
888 mutex_enter(hash_lock
);
890 ASSERT(MUTEX_HELD(hash_lock
));
893 for (fhdr
= buf_hash_table
.ht_table
[idx
], i
= 0; fhdr
!= NULL
;
894 fhdr
= fhdr
->b_hash_next
, i
++) {
895 if (BUF_EQUAL(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
, fhdr
))
899 hdr
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
900 buf_hash_table
.ht_table
[idx
] = hdr
;
901 hdr
->b_flags
|= ARC_FLAG_IN_HASH_TABLE
;
903 /* collect some hash table performance data */
905 ARCSTAT_BUMP(arcstat_hash_collisions
);
907 ARCSTAT_BUMP(arcstat_hash_chains
);
909 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
912 ARCSTAT_BUMP(arcstat_hash_elements
);
913 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
919 buf_hash_remove(arc_buf_hdr_t
*hdr
)
921 arc_buf_hdr_t
*fhdr
, **hdrp
;
922 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
924 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
925 ASSERT(HDR_IN_HASH_TABLE(hdr
));
927 hdrp
= &buf_hash_table
.ht_table
[idx
];
928 while ((fhdr
= *hdrp
) != hdr
) {
929 ASSERT(fhdr
!= NULL
);
930 hdrp
= &fhdr
->b_hash_next
;
932 *hdrp
= hdr
->b_hash_next
;
933 hdr
->b_hash_next
= NULL
;
934 hdr
->b_flags
&= ~ARC_FLAG_IN_HASH_TABLE
;
936 /* collect some hash table performance data */
937 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
939 if (buf_hash_table
.ht_table
[idx
] &&
940 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
941 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
945 * Global data structures and functions for the buf kmem cache.
947 static kmem_cache_t
*hdr_full_cache
;
948 static kmem_cache_t
*hdr_l2only_cache
;
949 static kmem_cache_t
*buf_cache
;
956 #if defined(_KERNEL) && defined(HAVE_SPL)
958 * Large allocations which do not require contiguous pages
959 * should be using vmem_free() in the linux kernel\
961 vmem_free(buf_hash_table
.ht_table
,
962 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
964 kmem_free(buf_hash_table
.ht_table
,
965 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
967 for (i
= 0; i
< BUF_LOCKS
; i
++)
968 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
969 kmem_cache_destroy(hdr_full_cache
);
970 kmem_cache_destroy(hdr_l2only_cache
);
971 kmem_cache_destroy(buf_cache
);
975 * Constructor callback - called when the cache is empty
976 * and a new buf is requested.
980 hdr_full_cons(void *vbuf
, void *unused
, int kmflag
)
982 arc_buf_hdr_t
*hdr
= vbuf
;
984 bzero(hdr
, HDR_FULL_SIZE
);
985 cv_init(&hdr
->b_l1hdr
.b_cv
, NULL
, CV_DEFAULT
, NULL
);
986 refcount_create(&hdr
->b_l1hdr
.b_refcnt
);
987 mutex_init(&hdr
->b_l1hdr
.b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
988 list_link_init(&hdr
->b_l1hdr
.b_arc_node
);
989 list_link_init(&hdr
->b_l2hdr
.b_l2node
);
990 multilist_link_init(&hdr
->b_l1hdr
.b_arc_node
);
991 arc_space_consume(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
998 hdr_l2only_cons(void *vbuf
, void *unused
, int kmflag
)
1000 arc_buf_hdr_t
*hdr
= vbuf
;
1002 bzero(hdr
, HDR_L2ONLY_SIZE
);
1003 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1010 buf_cons(void *vbuf
, void *unused
, int kmflag
)
1012 arc_buf_t
*buf
= vbuf
;
1014 bzero(buf
, sizeof (arc_buf_t
));
1015 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
1016 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1022 * Destructor callback - called when a cached buf is
1023 * no longer required.
1027 hdr_full_dest(void *vbuf
, void *unused
)
1029 arc_buf_hdr_t
*hdr
= vbuf
;
1031 ASSERT(BUF_EMPTY(hdr
));
1032 cv_destroy(&hdr
->b_l1hdr
.b_cv
);
1033 refcount_destroy(&hdr
->b_l1hdr
.b_refcnt
);
1034 mutex_destroy(&hdr
->b_l1hdr
.b_freeze_lock
);
1035 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1036 arc_space_return(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
1041 hdr_l2only_dest(void *vbuf
, void *unused
)
1043 ASSERTV(arc_buf_hdr_t
*hdr
= vbuf
);
1045 ASSERT(BUF_EMPTY(hdr
));
1046 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1051 buf_dest(void *vbuf
, void *unused
)
1053 arc_buf_t
*buf
= vbuf
;
1055 mutex_destroy(&buf
->b_evict_lock
);
1056 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
1060 * Reclaim callback -- invoked when memory is low.
1064 hdr_recl(void *unused
)
1066 dprintf("hdr_recl called\n");
1068 * umem calls the reclaim func when we destroy the buf cache,
1069 * which is after we do arc_fini().
1072 cv_signal(&arc_reclaim_thread_cv
);
1079 uint64_t hsize
= 1ULL << 12;
1083 * The hash table is big enough to fill all of physical memory
1084 * with an average block size of zfs_arc_average_blocksize (default 8K).
1085 * By default, the table will take up
1086 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1088 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
1091 buf_hash_table
.ht_mask
= hsize
- 1;
1092 #if defined(_KERNEL) && defined(HAVE_SPL)
1094 * Large allocations which do not require contiguous pages
1095 * should be using vmem_alloc() in the linux kernel
1097 buf_hash_table
.ht_table
=
1098 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
1100 buf_hash_table
.ht_table
=
1101 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
1103 if (buf_hash_table
.ht_table
== NULL
) {
1104 ASSERT(hsize
> (1ULL << 8));
1109 hdr_full_cache
= kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE
,
1110 0, hdr_full_cons
, hdr_full_dest
, hdr_recl
, NULL
, NULL
, 0);
1111 hdr_l2only_cache
= kmem_cache_create("arc_buf_hdr_t_l2only",
1112 HDR_L2ONLY_SIZE
, 0, hdr_l2only_cons
, hdr_l2only_dest
, hdr_recl
,
1114 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
1115 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
1117 for (i
= 0; i
< 256; i
++)
1118 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
1119 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
1121 for (i
= 0; i
< BUF_LOCKS
; i
++) {
1122 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
1123 NULL
, MUTEX_DEFAULT
, NULL
);
1128 * Transition between the two allocation states for the arc_buf_hdr struct.
1129 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1130 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1131 * version is used when a cache buffer is only in the L2ARC in order to reduce
1134 static arc_buf_hdr_t
*
1135 arc_hdr_realloc(arc_buf_hdr_t
*hdr
, kmem_cache_t
*old
, kmem_cache_t
*new)
1137 arc_buf_hdr_t
*nhdr
;
1140 ASSERT(HDR_HAS_L2HDR(hdr
));
1141 ASSERT((old
== hdr_full_cache
&& new == hdr_l2only_cache
) ||
1142 (old
== hdr_l2only_cache
&& new == hdr_full_cache
));
1144 dev
= hdr
->b_l2hdr
.b_dev
;
1145 nhdr
= kmem_cache_alloc(new, KM_PUSHPAGE
);
1147 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
1148 buf_hash_remove(hdr
);
1150 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
1152 if (new == hdr_full_cache
) {
1153 nhdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
1155 * arc_access and arc_change_state need to be aware that a
1156 * header has just come out of L2ARC, so we set its state to
1157 * l2c_only even though it's about to change.
1159 nhdr
->b_l1hdr
.b_state
= arc_l2c_only
;
1161 /* Verify previous threads set to NULL before freeing */
1162 ASSERT3P(nhdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1164 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1165 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
1168 * If we've reached here, We must have been called from
1169 * arc_evict_hdr(), as such we should have already been
1170 * removed from any ghost list we were previously on
1171 * (which protects us from racing with arc_evict_state),
1172 * thus no locking is needed during this check.
1174 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1177 * A buffer must not be moved into the arc_l2c_only
1178 * state if it's not finished being written out to the
1179 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1180 * might try to be accessed, even though it was removed.
1182 VERIFY(!HDR_L2_WRITING(hdr
));
1183 VERIFY3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1185 nhdr
->b_flags
&= ~ARC_FLAG_HAS_L1HDR
;
1188 * The header has been reallocated so we need to re-insert it into any
1191 (void) buf_hash_insert(nhdr
, NULL
);
1193 ASSERT(list_link_active(&hdr
->b_l2hdr
.b_l2node
));
1195 mutex_enter(&dev
->l2ad_mtx
);
1198 * We must place the realloc'ed header back into the list at
1199 * the same spot. Otherwise, if it's placed earlier in the list,
1200 * l2arc_write_buffers() could find it during the function's
1201 * write phase, and try to write it out to the l2arc.
1203 list_insert_after(&dev
->l2ad_buflist
, hdr
, nhdr
);
1204 list_remove(&dev
->l2ad_buflist
, hdr
);
1206 mutex_exit(&dev
->l2ad_mtx
);
1209 * Since we're using the pointer address as the tag when
1210 * incrementing and decrementing the l2ad_alloc refcount, we
1211 * must remove the old pointer (that we're about to destroy) and
1212 * add the new pointer to the refcount. Otherwise we'd remove
1213 * the wrong pointer address when calling arc_hdr_destroy() later.
1216 (void) refcount_remove_many(&dev
->l2ad_alloc
,
1217 hdr
->b_l2hdr
.b_asize
, hdr
);
1219 (void) refcount_add_many(&dev
->l2ad_alloc
,
1220 nhdr
->b_l2hdr
.b_asize
, nhdr
);
1222 buf_discard_identity(hdr
);
1223 hdr
->b_freeze_cksum
= NULL
;
1224 kmem_cache_free(old
, hdr
);
1230 #define ARC_MINTIME (hz>>4) /* 62 ms */
1233 arc_cksum_verify(arc_buf_t
*buf
)
1237 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1240 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1241 if (buf
->b_hdr
->b_freeze_cksum
== NULL
|| HDR_IO_ERROR(buf
->b_hdr
)) {
1242 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1245 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
1246 if (!ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
))
1247 panic("buffer modified while frozen!");
1248 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1252 arc_cksum_equal(arc_buf_t
*buf
)
1257 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1258 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
1259 equal
= ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
);
1260 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1266 arc_cksum_compute(arc_buf_t
*buf
, boolean_t force
)
1268 if (!force
&& !(zfs_flags
& ZFS_DEBUG_MODIFY
))
1271 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1272 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1273 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1276 buf
->b_hdr
->b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
), KM_SLEEP
);
1277 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
,
1278 buf
->b_hdr
->b_freeze_cksum
);
1279 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1285 arc_buf_sigsegv(int sig
, siginfo_t
*si
, void *unused
)
1287 panic("Got SIGSEGV at address: 0x%lx\n", (long) si
->si_addr
);
1293 arc_buf_unwatch(arc_buf_t
*buf
)
1297 ASSERT0(mprotect(buf
->b_data
, buf
->b_hdr
->b_size
,
1298 PROT_READ
| PROT_WRITE
));
1305 arc_buf_watch(arc_buf_t
*buf
)
1309 ASSERT0(mprotect(buf
->b_data
, buf
->b_hdr
->b_size
, PROT_READ
));
1313 static arc_buf_contents_t
1314 arc_buf_type(arc_buf_hdr_t
*hdr
)
1316 if (HDR_ISTYPE_METADATA(hdr
)) {
1317 return (ARC_BUFC_METADATA
);
1319 return (ARC_BUFC_DATA
);
1324 arc_bufc_to_flags(arc_buf_contents_t type
)
1328 /* metadata field is 0 if buffer contains normal data */
1330 case ARC_BUFC_METADATA
:
1331 return (ARC_FLAG_BUFC_METADATA
);
1335 panic("undefined ARC buffer type!");
1336 return ((uint32_t)-1);
1340 arc_buf_thaw(arc_buf_t
*buf
)
1342 if (zfs_flags
& ZFS_DEBUG_MODIFY
) {
1343 if (buf
->b_hdr
->b_l1hdr
.b_state
!= arc_anon
)
1344 panic("modifying non-anon buffer!");
1345 if (HDR_IO_IN_PROGRESS(buf
->b_hdr
))
1346 panic("modifying buffer while i/o in progress!");
1347 arc_cksum_verify(buf
);
1350 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1351 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1352 kmem_free(buf
->b_hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
1353 buf
->b_hdr
->b_freeze_cksum
= NULL
;
1356 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1358 arc_buf_unwatch(buf
);
1362 arc_buf_freeze(arc_buf_t
*buf
)
1364 kmutex_t
*hash_lock
;
1366 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1369 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1370 mutex_enter(hash_lock
);
1372 ASSERT(buf
->b_hdr
->b_freeze_cksum
!= NULL
||
1373 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
1374 arc_cksum_compute(buf
, B_FALSE
);
1375 mutex_exit(hash_lock
);
1380 add_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1384 ASSERT(HDR_HAS_L1HDR(hdr
));
1385 ASSERT(MUTEX_HELD(hash_lock
));
1387 state
= hdr
->b_l1hdr
.b_state
;
1389 if ((refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
) == 1) &&
1390 (state
!= arc_anon
)) {
1391 /* We don't use the L2-only state list. */
1392 if (state
!= arc_l2c_only
) {
1393 arc_buf_contents_t type
= arc_buf_type(hdr
);
1394 uint64_t delta
= hdr
->b_size
* hdr
->b_l1hdr
.b_datacnt
;
1395 multilist_t
*list
= &state
->arcs_list
[type
];
1396 uint64_t *size
= &state
->arcs_lsize
[type
];
1398 multilist_remove(list
, hdr
);
1400 if (GHOST_STATE(state
)) {
1401 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
1402 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1403 delta
= hdr
->b_size
;
1406 ASSERT3U(*size
, >=, delta
);
1407 atomic_add_64(size
, -delta
);
1409 /* remove the prefetch flag if we get a reference */
1410 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
1415 remove_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1418 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
1420 ASSERT(HDR_HAS_L1HDR(hdr
));
1421 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
1422 ASSERT(!GHOST_STATE(state
));
1425 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1426 * check to prevent usage of the arc_l2c_only list.
1428 if (((cnt
= refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
)) == 0) &&
1429 (state
!= arc_anon
)) {
1430 arc_buf_contents_t type
= arc_buf_type(hdr
);
1431 multilist_t
*list
= &state
->arcs_list
[type
];
1432 uint64_t *size
= &state
->arcs_lsize
[type
];
1434 multilist_insert(list
, hdr
);
1436 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
1437 atomic_add_64(size
, hdr
->b_size
*
1438 hdr
->b_l1hdr
.b_datacnt
);
1444 * Returns detailed information about a specific arc buffer. When the
1445 * state_index argument is set the function will calculate the arc header
1446 * list position for its arc state. Since this requires a linear traversal
1447 * callers are strongly encourage not to do this. However, it can be helpful
1448 * for targeted analysis so the functionality is provided.
1451 arc_buf_info(arc_buf_t
*ab
, arc_buf_info_t
*abi
, int state_index
)
1453 arc_buf_hdr_t
*hdr
= ab
->b_hdr
;
1454 l1arc_buf_hdr_t
*l1hdr
= NULL
;
1455 l2arc_buf_hdr_t
*l2hdr
= NULL
;
1456 arc_state_t
*state
= NULL
;
1458 if (HDR_HAS_L1HDR(hdr
)) {
1459 l1hdr
= &hdr
->b_l1hdr
;
1460 state
= l1hdr
->b_state
;
1462 if (HDR_HAS_L2HDR(hdr
))
1463 l2hdr
= &hdr
->b_l2hdr
;
1465 memset(abi
, 0, sizeof (arc_buf_info_t
));
1466 abi
->abi_flags
= hdr
->b_flags
;
1469 abi
->abi_datacnt
= l1hdr
->b_datacnt
;
1470 abi
->abi_access
= l1hdr
->b_arc_access
;
1471 abi
->abi_mru_hits
= l1hdr
->b_mru_hits
;
1472 abi
->abi_mru_ghost_hits
= l1hdr
->b_mru_ghost_hits
;
1473 abi
->abi_mfu_hits
= l1hdr
->b_mfu_hits
;
1474 abi
->abi_mfu_ghost_hits
= l1hdr
->b_mfu_ghost_hits
;
1475 abi
->abi_holds
= refcount_count(&l1hdr
->b_refcnt
);
1479 abi
->abi_l2arc_dattr
= l2hdr
->b_daddr
;
1480 abi
->abi_l2arc_asize
= l2hdr
->b_asize
;
1481 abi
->abi_l2arc_compress
= l2hdr
->b_compress
;
1482 abi
->abi_l2arc_hits
= l2hdr
->b_hits
;
1485 abi
->abi_state_type
= state
? state
->arcs_state
: ARC_STATE_ANON
;
1486 abi
->abi_state_contents
= arc_buf_type(hdr
);
1487 abi
->abi_size
= hdr
->b_size
;
1491 * Move the supplied buffer to the indicated state. The hash lock
1492 * for the buffer must be held by the caller.
1495 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*hdr
,
1496 kmutex_t
*hash_lock
)
1498 arc_state_t
*old_state
;
1501 uint64_t from_delta
, to_delta
;
1502 arc_buf_contents_t buftype
= arc_buf_type(hdr
);
1505 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1506 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1507 * L1 hdr doesn't always exist when we change state to arc_anon before
1508 * destroying a header, in which case reallocating to add the L1 hdr is
1511 if (HDR_HAS_L1HDR(hdr
)) {
1512 old_state
= hdr
->b_l1hdr
.b_state
;
1513 refcnt
= refcount_count(&hdr
->b_l1hdr
.b_refcnt
);
1514 datacnt
= hdr
->b_l1hdr
.b_datacnt
;
1516 old_state
= arc_l2c_only
;
1521 ASSERT(MUTEX_HELD(hash_lock
));
1522 ASSERT3P(new_state
, !=, old_state
);
1523 ASSERT(refcnt
== 0 || datacnt
> 0);
1524 ASSERT(!GHOST_STATE(new_state
) || datacnt
== 0);
1525 ASSERT(old_state
!= arc_anon
|| datacnt
<= 1);
1527 from_delta
= to_delta
= datacnt
* hdr
->b_size
;
1530 * If this buffer is evictable, transfer it from the
1531 * old state list to the new state list.
1534 if (old_state
!= arc_anon
&& old_state
!= arc_l2c_only
) {
1535 uint64_t *size
= &old_state
->arcs_lsize
[buftype
];
1537 ASSERT(HDR_HAS_L1HDR(hdr
));
1538 multilist_remove(&old_state
->arcs_list
[buftype
], hdr
);
1541 * If prefetching out of the ghost cache,
1542 * we will have a non-zero datacnt.
1544 if (GHOST_STATE(old_state
) && datacnt
== 0) {
1545 /* ghost elements have a ghost size */
1546 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1547 from_delta
= hdr
->b_size
;
1549 ASSERT3U(*size
, >=, from_delta
);
1550 atomic_add_64(size
, -from_delta
);
1552 if (new_state
!= arc_anon
&& new_state
!= arc_l2c_only
) {
1553 uint64_t *size
= &new_state
->arcs_lsize
[buftype
];
1556 * An L1 header always exists here, since if we're
1557 * moving to some L1-cached state (i.e. not l2c_only or
1558 * anonymous), we realloc the header to add an L1hdr
1561 ASSERT(HDR_HAS_L1HDR(hdr
));
1562 multilist_insert(&new_state
->arcs_list
[buftype
], hdr
);
1564 /* ghost elements have a ghost size */
1565 if (GHOST_STATE(new_state
)) {
1567 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1568 to_delta
= hdr
->b_size
;
1570 atomic_add_64(size
, to_delta
);
1574 ASSERT(!BUF_EMPTY(hdr
));
1575 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(hdr
))
1576 buf_hash_remove(hdr
);
1578 /* adjust state sizes (ignore arc_l2c_only) */
1580 if (to_delta
&& new_state
!= arc_l2c_only
) {
1581 ASSERT(HDR_HAS_L1HDR(hdr
));
1582 if (GHOST_STATE(new_state
)) {
1586 * We moving a header to a ghost state, we first
1587 * remove all arc buffers. Thus, we'll have a
1588 * datacnt of zero, and no arc buffer to use for
1589 * the reference. As a result, we use the arc
1590 * header pointer for the reference.
1592 (void) refcount_add_many(&new_state
->arcs_size
,
1596 ASSERT3U(datacnt
, !=, 0);
1599 * Each individual buffer holds a unique reference,
1600 * thus we must remove each of these references one
1603 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
1604 buf
= buf
->b_next
) {
1605 (void) refcount_add_many(&new_state
->arcs_size
,
1611 if (from_delta
&& old_state
!= arc_l2c_only
) {
1612 ASSERT(HDR_HAS_L1HDR(hdr
));
1613 if (GHOST_STATE(old_state
)) {
1615 * When moving a header off of a ghost state,
1616 * there's the possibility for datacnt to be
1617 * non-zero. This is because we first add the
1618 * arc buffer to the header prior to changing
1619 * the header's state. Since we used the header
1620 * for the reference when putting the header on
1621 * the ghost state, we must balance that and use
1622 * the header when removing off the ghost state
1623 * (even though datacnt is non zero).
1626 IMPLY(datacnt
== 0, new_state
== arc_anon
||
1627 new_state
== arc_l2c_only
);
1629 (void) refcount_remove_many(&old_state
->arcs_size
,
1633 ASSERT3U(datacnt
, !=, 0);
1636 * Each individual buffer holds a unique reference,
1637 * thus we must remove each of these references one
1640 for (buf
= hdr
->b_l1hdr
.b_buf
; buf
!= NULL
;
1641 buf
= buf
->b_next
) {
1642 (void) refcount_remove_many(
1643 &old_state
->arcs_size
, hdr
->b_size
, buf
);
1648 if (HDR_HAS_L1HDR(hdr
))
1649 hdr
->b_l1hdr
.b_state
= new_state
;
1652 * L2 headers should never be on the L2 state list since they don't
1653 * have L1 headers allocated.
1655 ASSERT(multilist_is_empty(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]) &&
1656 multilist_is_empty(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]));
1660 arc_space_consume(uint64_t space
, arc_space_type_t type
)
1662 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1667 case ARC_SPACE_DATA
:
1668 ARCSTAT_INCR(arcstat_data_size
, space
);
1670 case ARC_SPACE_META
:
1671 ARCSTAT_INCR(arcstat_metadata_size
, space
);
1673 case ARC_SPACE_OTHER
:
1674 ARCSTAT_INCR(arcstat_other_size
, space
);
1676 case ARC_SPACE_HDRS
:
1677 ARCSTAT_INCR(arcstat_hdr_size
, space
);
1679 case ARC_SPACE_L2HDRS
:
1680 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
1684 if (type
!= ARC_SPACE_DATA
)
1685 ARCSTAT_INCR(arcstat_meta_used
, space
);
1687 atomic_add_64(&arc_size
, space
);
1691 arc_space_return(uint64_t space
, arc_space_type_t type
)
1693 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1698 case ARC_SPACE_DATA
:
1699 ARCSTAT_INCR(arcstat_data_size
, -space
);
1701 case ARC_SPACE_META
:
1702 ARCSTAT_INCR(arcstat_metadata_size
, -space
);
1704 case ARC_SPACE_OTHER
:
1705 ARCSTAT_INCR(arcstat_other_size
, -space
);
1707 case ARC_SPACE_HDRS
:
1708 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
1710 case ARC_SPACE_L2HDRS
:
1711 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
1715 if (type
!= ARC_SPACE_DATA
) {
1716 ASSERT(arc_meta_used
>= space
);
1717 if (arc_meta_max
< arc_meta_used
)
1718 arc_meta_max
= arc_meta_used
;
1719 ARCSTAT_INCR(arcstat_meta_used
, -space
);
1722 ASSERT(arc_size
>= space
);
1723 atomic_add_64(&arc_size
, -space
);
1727 arc_buf_alloc(spa_t
*spa
, uint64_t size
, void *tag
, arc_buf_contents_t type
)
1732 VERIFY3U(size
, <=, spa_maxblocksize(spa
));
1733 hdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
1734 ASSERT(BUF_EMPTY(hdr
));
1735 ASSERT3P(hdr
->b_freeze_cksum
, ==, NULL
);
1737 hdr
->b_spa
= spa_load_guid(spa
);
1738 hdr
->b_l1hdr
.b_mru_hits
= 0;
1739 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
1740 hdr
->b_l1hdr
.b_mfu_hits
= 0;
1741 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
1742 hdr
->b_l1hdr
.b_l2_hits
= 0;
1744 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1747 buf
->b_efunc
= NULL
;
1748 buf
->b_private
= NULL
;
1751 hdr
->b_flags
= arc_bufc_to_flags(type
);
1752 hdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
1754 hdr
->b_l1hdr
.b_buf
= buf
;
1755 hdr
->b_l1hdr
.b_state
= arc_anon
;
1756 hdr
->b_l1hdr
.b_arc_access
= 0;
1757 hdr
->b_l1hdr
.b_datacnt
= 1;
1758 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1760 arc_get_data_buf(buf
);
1761 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
1762 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1767 static char *arc_onloan_tag
= "onloan";
1770 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1771 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1772 * buffers must be returned to the arc before they can be used by the DMU or
1776 arc_loan_buf(spa_t
*spa
, uint64_t size
)
1780 buf
= arc_buf_alloc(spa
, size
, arc_onloan_tag
, ARC_BUFC_DATA
);
1782 atomic_add_64(&arc_loaned_bytes
, size
);
1787 * Return a loaned arc buffer to the arc.
1790 arc_return_buf(arc_buf_t
*buf
, void *tag
)
1792 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1794 ASSERT(buf
->b_data
!= NULL
);
1795 ASSERT(HDR_HAS_L1HDR(hdr
));
1796 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1797 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
1799 atomic_add_64(&arc_loaned_bytes
, -hdr
->b_size
);
1802 /* Detach an arc_buf from a dbuf (tag) */
1804 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
1806 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1808 ASSERT(buf
->b_data
!= NULL
);
1809 ASSERT(HDR_HAS_L1HDR(hdr
));
1810 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
1811 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1812 buf
->b_efunc
= NULL
;
1813 buf
->b_private
= NULL
;
1815 atomic_add_64(&arc_loaned_bytes
, hdr
->b_size
);
1819 arc_buf_clone(arc_buf_t
*from
)
1822 arc_buf_hdr_t
*hdr
= from
->b_hdr
;
1823 uint64_t size
= hdr
->b_size
;
1825 ASSERT(HDR_HAS_L1HDR(hdr
));
1826 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_anon
);
1828 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1831 buf
->b_efunc
= NULL
;
1832 buf
->b_private
= NULL
;
1833 buf
->b_next
= hdr
->b_l1hdr
.b_buf
;
1834 hdr
->b_l1hdr
.b_buf
= buf
;
1835 arc_get_data_buf(buf
);
1836 bcopy(from
->b_data
, buf
->b_data
, size
);
1839 * This buffer already exists in the arc so create a duplicate
1840 * copy for the caller. If the buffer is associated with user data
1841 * then track the size and number of duplicates. These stats will be
1842 * updated as duplicate buffers are created and destroyed.
1844 if (HDR_ISTYPE_DATA(hdr
)) {
1845 ARCSTAT_BUMP(arcstat_duplicate_buffers
);
1846 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, size
);
1848 hdr
->b_l1hdr
.b_datacnt
+= 1;
1853 arc_buf_add_ref(arc_buf_t
*buf
, void* tag
)
1856 kmutex_t
*hash_lock
;
1859 * Check to see if this buffer is evicted. Callers
1860 * must verify b_data != NULL to know if the add_ref
1863 mutex_enter(&buf
->b_evict_lock
);
1864 if (buf
->b_data
== NULL
) {
1865 mutex_exit(&buf
->b_evict_lock
);
1868 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1869 mutex_enter(hash_lock
);
1871 ASSERT(HDR_HAS_L1HDR(hdr
));
1872 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1873 mutex_exit(&buf
->b_evict_lock
);
1875 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
1876 hdr
->b_l1hdr
.b_state
== arc_mfu
);
1878 add_reference(hdr
, hash_lock
, tag
);
1879 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
1880 arc_access(hdr
, hash_lock
);
1881 mutex_exit(hash_lock
);
1882 ARCSTAT_BUMP(arcstat_hits
);
1883 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
1884 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
1885 data
, metadata
, hits
);
1889 arc_buf_free_on_write(void *data
, size_t size
,
1890 void (*free_func
)(void *, size_t))
1892 l2arc_data_free_t
*df
;
1894 df
= kmem_alloc(sizeof (*df
), KM_SLEEP
);
1895 df
->l2df_data
= data
;
1896 df
->l2df_size
= size
;
1897 df
->l2df_func
= free_func
;
1898 mutex_enter(&l2arc_free_on_write_mtx
);
1899 list_insert_head(l2arc_free_on_write
, df
);
1900 mutex_exit(&l2arc_free_on_write_mtx
);
1904 * Free the arc data buffer. If it is an l2arc write in progress,
1905 * the buffer is placed on l2arc_free_on_write to be freed later.
1908 arc_buf_data_free(arc_buf_t
*buf
, void (*free_func
)(void *, size_t))
1910 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1912 if (HDR_L2_WRITING(hdr
)) {
1913 arc_buf_free_on_write(buf
->b_data
, hdr
->b_size
, free_func
);
1914 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
1916 free_func(buf
->b_data
, hdr
->b_size
);
1921 arc_buf_l2_cdata_free(arc_buf_hdr_t
*hdr
)
1923 ASSERT(HDR_HAS_L2HDR(hdr
));
1924 ASSERT(MUTEX_HELD(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
));
1927 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
1928 * that doesn't exist, the header is in the arc_l2c_only state,
1929 * and there isn't anything to free (it's already been freed).
1931 if (!HDR_HAS_L1HDR(hdr
))
1935 * The header isn't being written to the l2arc device, thus it
1936 * shouldn't have a b_tmp_cdata to free.
1938 if (!HDR_L2_WRITING(hdr
)) {
1939 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1944 * The header does not have compression enabled. This can be due
1945 * to the buffer not being compressible, or because we're
1946 * freeing the buffer before the second phase of
1947 * l2arc_write_buffer() has started (which does the compression
1948 * step). In either case, b_tmp_cdata does not point to a
1949 * separately compressed buffer, so there's nothing to free (it
1950 * points to the same buffer as the arc_buf_t's b_data field).
1952 if (hdr
->b_l2hdr
.b_compress
== ZIO_COMPRESS_OFF
) {
1953 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1958 * There's nothing to free since the buffer was all zero's and
1959 * compressed to a zero length buffer.
1961 if (hdr
->b_l2hdr
.b_compress
== ZIO_COMPRESS_EMPTY
) {
1962 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1966 ASSERT(L2ARC_IS_VALID_COMPRESS(hdr
->b_l2hdr
.b_compress
));
1968 arc_buf_free_on_write(hdr
->b_l1hdr
.b_tmp_cdata
,
1969 hdr
->b_size
, zio_data_buf_free
);
1971 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write
);
1972 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1976 * Free up buf->b_data and if 'remove' is set, then pull the
1977 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
1980 arc_buf_destroy(arc_buf_t
*buf
, boolean_t remove
)
1984 /* free up data associated with the buf */
1985 if (buf
->b_data
!= NULL
) {
1986 arc_state_t
*state
= buf
->b_hdr
->b_l1hdr
.b_state
;
1987 uint64_t size
= buf
->b_hdr
->b_size
;
1988 arc_buf_contents_t type
= arc_buf_type(buf
->b_hdr
);
1990 arc_cksum_verify(buf
);
1991 arc_buf_unwatch(buf
);
1993 if (type
== ARC_BUFC_METADATA
) {
1994 arc_buf_data_free(buf
, zio_buf_free
);
1995 arc_space_return(size
, ARC_SPACE_META
);
1997 ASSERT(type
== ARC_BUFC_DATA
);
1998 arc_buf_data_free(buf
, zio_data_buf_free
);
1999 arc_space_return(size
, ARC_SPACE_DATA
);
2002 /* protected by hash lock, if in the hash table */
2003 if (multilist_link_active(&buf
->b_hdr
->b_l1hdr
.b_arc_node
)) {
2004 uint64_t *cnt
= &state
->arcs_lsize
[type
];
2006 ASSERT(refcount_is_zero(
2007 &buf
->b_hdr
->b_l1hdr
.b_refcnt
));
2008 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
2010 ASSERT3U(*cnt
, >=, size
);
2011 atomic_add_64(cnt
, -size
);
2014 (void) refcount_remove_many(&state
->arcs_size
, size
, buf
);
2018 * If we're destroying a duplicate buffer make sure
2019 * that the appropriate statistics are updated.
2021 if (buf
->b_hdr
->b_l1hdr
.b_datacnt
> 1 &&
2022 HDR_ISTYPE_DATA(buf
->b_hdr
)) {
2023 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
2024 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, -size
);
2026 ASSERT(buf
->b_hdr
->b_l1hdr
.b_datacnt
> 0);
2027 buf
->b_hdr
->b_l1hdr
.b_datacnt
-= 1;
2030 /* only remove the buf if requested */
2034 /* remove the buf from the hdr list */
2035 for (bufp
= &buf
->b_hdr
->b_l1hdr
.b_buf
; *bufp
!= buf
;
2036 bufp
= &(*bufp
)->b_next
)
2038 *bufp
= buf
->b_next
;
2041 ASSERT(buf
->b_efunc
== NULL
);
2043 /* clean up the buf */
2045 kmem_cache_free(buf_cache
, buf
);
2049 arc_hdr_l2hdr_destroy(arc_buf_hdr_t
*hdr
)
2051 l2arc_buf_hdr_t
*l2hdr
= &hdr
->b_l2hdr
;
2052 l2arc_dev_t
*dev
= l2hdr
->b_dev
;
2054 ASSERT(MUTEX_HELD(&dev
->l2ad_mtx
));
2055 ASSERT(HDR_HAS_L2HDR(hdr
));
2057 list_remove(&dev
->l2ad_buflist
, hdr
);
2060 * We don't want to leak the b_tmp_cdata buffer that was
2061 * allocated in l2arc_write_buffers()
2063 arc_buf_l2_cdata_free(hdr
);
2066 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2067 * this header is being processed by l2arc_write_buffers() (i.e.
2068 * it's in the first stage of l2arc_write_buffers()).
2069 * Re-affirming that truth here, just to serve as a reminder. If
2070 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2071 * may not have its HDR_L2_WRITING flag set. (the write may have
2072 * completed, in which case HDR_L2_WRITING will be false and the
2073 * b_daddr field will point to the address of the buffer on disk).
2075 IMPLY(l2hdr
->b_daddr
== L2ARC_ADDR_UNSET
, HDR_L2_WRITING(hdr
));
2078 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2079 * l2arc_write_buffers(). Since we've just removed this header
2080 * from the l2arc buffer list, this header will never reach the
2081 * second stage of l2arc_write_buffers(), which increments the
2082 * accounting stats for this header. Thus, we must be careful
2083 * not to decrement them for this header either.
2085 if (l2hdr
->b_daddr
!= L2ARC_ADDR_UNSET
) {
2086 ARCSTAT_INCR(arcstat_l2_asize
, -l2hdr
->b_asize
);
2087 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
2089 vdev_space_update(dev
->l2ad_vdev
,
2090 -l2hdr
->b_asize
, 0, 0);
2092 (void) refcount_remove_many(&dev
->l2ad_alloc
,
2093 l2hdr
->b_asize
, hdr
);
2096 hdr
->b_flags
&= ~ARC_FLAG_HAS_L2HDR
;
2100 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
2102 if (HDR_HAS_L1HDR(hdr
)) {
2103 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
||
2104 hdr
->b_l1hdr
.b_datacnt
> 0);
2105 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2106 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
2108 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
2109 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
2111 if (HDR_HAS_L2HDR(hdr
)) {
2112 l2arc_dev_t
*dev
= hdr
->b_l2hdr
.b_dev
;
2113 boolean_t buflist_held
= MUTEX_HELD(&dev
->l2ad_mtx
);
2116 mutex_enter(&dev
->l2ad_mtx
);
2119 * Even though we checked this conditional above, we
2120 * need to check this again now that we have the
2121 * l2ad_mtx. This is because we could be racing with
2122 * another thread calling l2arc_evict() which might have
2123 * destroyed this header's L2 portion as we were waiting
2124 * to acquire the l2ad_mtx. If that happens, we don't
2125 * want to re-destroy the header's L2 portion.
2127 if (HDR_HAS_L2HDR(hdr
))
2128 arc_hdr_l2hdr_destroy(hdr
);
2131 mutex_exit(&dev
->l2ad_mtx
);
2134 if (!BUF_EMPTY(hdr
))
2135 buf_discard_identity(hdr
);
2137 if (hdr
->b_freeze_cksum
!= NULL
) {
2138 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
2139 hdr
->b_freeze_cksum
= NULL
;
2142 if (HDR_HAS_L1HDR(hdr
)) {
2143 while (hdr
->b_l1hdr
.b_buf
) {
2144 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
2146 if (buf
->b_efunc
!= NULL
) {
2147 mutex_enter(&arc_user_evicts_lock
);
2148 mutex_enter(&buf
->b_evict_lock
);
2149 ASSERT(buf
->b_hdr
!= NULL
);
2150 arc_buf_destroy(hdr
->b_l1hdr
.b_buf
, FALSE
);
2151 hdr
->b_l1hdr
.b_buf
= buf
->b_next
;
2152 buf
->b_hdr
= &arc_eviction_hdr
;
2153 buf
->b_next
= arc_eviction_list
;
2154 arc_eviction_list
= buf
;
2155 mutex_exit(&buf
->b_evict_lock
);
2156 cv_signal(&arc_user_evicts_cv
);
2157 mutex_exit(&arc_user_evicts_lock
);
2159 arc_buf_destroy(hdr
->b_l1hdr
.b_buf
, TRUE
);
2164 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
2165 if (HDR_HAS_L1HDR(hdr
)) {
2166 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
2167 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
2168 kmem_cache_free(hdr_full_cache
, hdr
);
2170 kmem_cache_free(hdr_l2only_cache
, hdr
);
2175 arc_buf_free(arc_buf_t
*buf
, void *tag
)
2177 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2178 int hashed
= hdr
->b_l1hdr
.b_state
!= arc_anon
;
2180 ASSERT(buf
->b_efunc
== NULL
);
2181 ASSERT(buf
->b_data
!= NULL
);
2184 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
2186 mutex_enter(hash_lock
);
2188 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
2190 (void) remove_reference(hdr
, hash_lock
, tag
);
2191 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
2192 arc_buf_destroy(buf
, TRUE
);
2194 ASSERT(buf
== hdr
->b_l1hdr
.b_buf
);
2195 ASSERT(buf
->b_efunc
== NULL
);
2196 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
2198 mutex_exit(hash_lock
);
2199 } else if (HDR_IO_IN_PROGRESS(hdr
)) {
2202 * We are in the middle of an async write. Don't destroy
2203 * this buffer unless the write completes before we finish
2204 * decrementing the reference count.
2206 mutex_enter(&arc_user_evicts_lock
);
2207 (void) remove_reference(hdr
, NULL
, tag
);
2208 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2209 destroy_hdr
= !HDR_IO_IN_PROGRESS(hdr
);
2210 mutex_exit(&arc_user_evicts_lock
);
2212 arc_hdr_destroy(hdr
);
2214 if (remove_reference(hdr
, NULL
, tag
) > 0)
2215 arc_buf_destroy(buf
, TRUE
);
2217 arc_hdr_destroy(hdr
);
2222 arc_buf_remove_ref(arc_buf_t
*buf
, void* tag
)
2224 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2225 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
2226 boolean_t no_callback
= (buf
->b_efunc
== NULL
);
2228 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
2229 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
2230 arc_buf_free(buf
, tag
);
2231 return (no_callback
);
2234 mutex_enter(hash_lock
);
2236 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
2237 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
2238 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_anon
);
2239 ASSERT(buf
->b_data
!= NULL
);
2241 (void) remove_reference(hdr
, hash_lock
, tag
);
2242 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
2244 arc_buf_destroy(buf
, TRUE
);
2245 } else if (no_callback
) {
2246 ASSERT(hdr
->b_l1hdr
.b_buf
== buf
&& buf
->b_next
== NULL
);
2247 ASSERT(buf
->b_efunc
== NULL
);
2248 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
2250 ASSERT(no_callback
|| hdr
->b_l1hdr
.b_datacnt
> 1 ||
2251 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
2252 mutex_exit(hash_lock
);
2253 return (no_callback
);
2257 arc_buf_size(arc_buf_t
*buf
)
2259 return (buf
->b_hdr
->b_size
);
2263 * Called from the DMU to determine if the current buffer should be
2264 * evicted. In order to ensure proper locking, the eviction must be initiated
2265 * from the DMU. Return true if the buffer is associated with user data and
2266 * duplicate buffers still exist.
2269 arc_buf_eviction_needed(arc_buf_t
*buf
)
2272 boolean_t evict_needed
= B_FALSE
;
2274 if (zfs_disable_dup_eviction
)
2277 mutex_enter(&buf
->b_evict_lock
);
2281 * We are in arc_do_user_evicts(); let that function
2282 * perform the eviction.
2284 ASSERT(buf
->b_data
== NULL
);
2285 mutex_exit(&buf
->b_evict_lock
);
2287 } else if (buf
->b_data
== NULL
) {
2289 * We have already been added to the arc eviction list;
2290 * recommend eviction.
2292 ASSERT3P(hdr
, ==, &arc_eviction_hdr
);
2293 mutex_exit(&buf
->b_evict_lock
);
2297 if (hdr
->b_l1hdr
.b_datacnt
> 1 && HDR_ISTYPE_DATA(hdr
))
2298 evict_needed
= B_TRUE
;
2300 mutex_exit(&buf
->b_evict_lock
);
2301 return (evict_needed
);
2305 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2306 * state of the header is dependent on its state prior to entering this
2307 * function. The following transitions are possible:
2309 * - arc_mru -> arc_mru_ghost
2310 * - arc_mfu -> arc_mfu_ghost
2311 * - arc_mru_ghost -> arc_l2c_only
2312 * - arc_mru_ghost -> deleted
2313 * - arc_mfu_ghost -> arc_l2c_only
2314 * - arc_mfu_ghost -> deleted
2317 arc_evict_hdr(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
2319 arc_state_t
*evicted_state
, *state
;
2320 int64_t bytes_evicted
= 0;
2322 ASSERT(MUTEX_HELD(hash_lock
));
2323 ASSERT(HDR_HAS_L1HDR(hdr
));
2325 state
= hdr
->b_l1hdr
.b_state
;
2326 if (GHOST_STATE(state
)) {
2327 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
2328 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
2331 * l2arc_write_buffers() relies on a header's L1 portion
2332 * (i.e. its b_tmp_cdata field) during its write phase.
2333 * Thus, we cannot push a header onto the arc_l2c_only
2334 * state (removing its L1 piece) until the header is
2335 * done being written to the l2arc.
2337 if (HDR_HAS_L2HDR(hdr
) && HDR_L2_WRITING(hdr
)) {
2338 ARCSTAT_BUMP(arcstat_evict_l2_skip
);
2339 return (bytes_evicted
);
2342 ARCSTAT_BUMP(arcstat_deleted
);
2343 bytes_evicted
+= hdr
->b_size
;
2345 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, hdr
);
2347 if (HDR_HAS_L2HDR(hdr
)) {
2349 * This buffer is cached on the 2nd Level ARC;
2350 * don't destroy the header.
2352 arc_change_state(arc_l2c_only
, hdr
, hash_lock
);
2354 * dropping from L1+L2 cached to L2-only,
2355 * realloc to remove the L1 header.
2357 hdr
= arc_hdr_realloc(hdr
, hdr_full_cache
,
2360 arc_change_state(arc_anon
, hdr
, hash_lock
);
2361 arc_hdr_destroy(hdr
);
2363 return (bytes_evicted
);
2366 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
2367 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
2369 /* prefetch buffers have a minimum lifespan */
2370 if (HDR_IO_IN_PROGRESS(hdr
) ||
2371 ((hdr
->b_flags
& (ARC_FLAG_PREFETCH
| ARC_FLAG_INDIRECT
)) &&
2372 ddi_get_lbolt() - hdr
->b_l1hdr
.b_arc_access
<
2373 arc_min_prefetch_lifespan
)) {
2374 ARCSTAT_BUMP(arcstat_evict_skip
);
2375 return (bytes_evicted
);
2378 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
2379 ASSERT3U(hdr
->b_l1hdr
.b_datacnt
, >, 0);
2380 while (hdr
->b_l1hdr
.b_buf
) {
2381 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
2382 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
2383 ARCSTAT_BUMP(arcstat_mutex_miss
);
2386 if (buf
->b_data
!= NULL
)
2387 bytes_evicted
+= hdr
->b_size
;
2388 if (buf
->b_efunc
!= NULL
) {
2389 mutex_enter(&arc_user_evicts_lock
);
2390 arc_buf_destroy(buf
, FALSE
);
2391 hdr
->b_l1hdr
.b_buf
= buf
->b_next
;
2392 buf
->b_hdr
= &arc_eviction_hdr
;
2393 buf
->b_next
= arc_eviction_list
;
2394 arc_eviction_list
= buf
;
2395 cv_signal(&arc_user_evicts_cv
);
2396 mutex_exit(&arc_user_evicts_lock
);
2397 mutex_exit(&buf
->b_evict_lock
);
2399 mutex_exit(&buf
->b_evict_lock
);
2400 arc_buf_destroy(buf
, TRUE
);
2404 if (HDR_HAS_L2HDR(hdr
)) {
2405 ARCSTAT_INCR(arcstat_evict_l2_cached
, hdr
->b_size
);
2407 if (l2arc_write_eligible(hdr
->b_spa
, hdr
))
2408 ARCSTAT_INCR(arcstat_evict_l2_eligible
, hdr
->b_size
);
2410 ARCSTAT_INCR(arcstat_evict_l2_ineligible
, hdr
->b_size
);
2413 if (hdr
->b_l1hdr
.b_datacnt
== 0) {
2414 arc_change_state(evicted_state
, hdr
, hash_lock
);
2415 ASSERT(HDR_IN_HASH_TABLE(hdr
));
2416 hdr
->b_flags
|= ARC_FLAG_IN_HASH_TABLE
;
2417 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
2418 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, hdr
);
2421 return (bytes_evicted
);
2425 arc_evict_state_impl(multilist_t
*ml
, int idx
, arc_buf_hdr_t
*marker
,
2426 uint64_t spa
, int64_t bytes
)
2428 multilist_sublist_t
*mls
;
2429 uint64_t bytes_evicted
= 0;
2431 kmutex_t
*hash_lock
;
2432 int evict_count
= 0;
2434 ASSERT3P(marker
, !=, NULL
);
2435 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
2437 mls
= multilist_sublist_lock(ml
, idx
);
2439 for (hdr
= multilist_sublist_prev(mls
, marker
); hdr
!= NULL
;
2440 hdr
= multilist_sublist_prev(mls
, marker
)) {
2441 if ((bytes
!= ARC_EVICT_ALL
&& bytes_evicted
>= bytes
) ||
2442 (evict_count
>= zfs_arc_evict_batch_limit
))
2446 * To keep our iteration location, move the marker
2447 * forward. Since we're not holding hdr's hash lock, we
2448 * must be very careful and not remove 'hdr' from the
2449 * sublist. Otherwise, other consumers might mistake the
2450 * 'hdr' as not being on a sublist when they call the
2451 * multilist_link_active() function (they all rely on
2452 * the hash lock protecting concurrent insertions and
2453 * removals). multilist_sublist_move_forward() was
2454 * specifically implemented to ensure this is the case
2455 * (only 'marker' will be removed and re-inserted).
2457 multilist_sublist_move_forward(mls
, marker
);
2460 * The only case where the b_spa field should ever be
2461 * zero, is the marker headers inserted by
2462 * arc_evict_state(). It's possible for multiple threads
2463 * to be calling arc_evict_state() concurrently (e.g.
2464 * dsl_pool_close() and zio_inject_fault()), so we must
2465 * skip any markers we see from these other threads.
2467 if (hdr
->b_spa
== 0)
2470 /* we're only interested in evicting buffers of a certain spa */
2471 if (spa
!= 0 && hdr
->b_spa
!= spa
) {
2472 ARCSTAT_BUMP(arcstat_evict_skip
);
2476 hash_lock
= HDR_LOCK(hdr
);
2479 * We aren't calling this function from any code path
2480 * that would already be holding a hash lock, so we're
2481 * asserting on this assumption to be defensive in case
2482 * this ever changes. Without this check, it would be
2483 * possible to incorrectly increment arcstat_mutex_miss
2484 * below (e.g. if the code changed such that we called
2485 * this function with a hash lock held).
2487 ASSERT(!MUTEX_HELD(hash_lock
));
2489 if (mutex_tryenter(hash_lock
)) {
2490 uint64_t evicted
= arc_evict_hdr(hdr
, hash_lock
);
2491 mutex_exit(hash_lock
);
2493 bytes_evicted
+= evicted
;
2496 * If evicted is zero, arc_evict_hdr() must have
2497 * decided to skip this header, don't increment
2498 * evict_count in this case.
2504 * If arc_size isn't overflowing, signal any
2505 * threads that might happen to be waiting.
2507 * For each header evicted, we wake up a single
2508 * thread. If we used cv_broadcast, we could
2509 * wake up "too many" threads causing arc_size
2510 * to significantly overflow arc_c; since
2511 * arc_get_data_buf() doesn't check for overflow
2512 * when it's woken up (it doesn't because it's
2513 * possible for the ARC to be overflowing while
2514 * full of un-evictable buffers, and the
2515 * function should proceed in this case).
2517 * If threads are left sleeping, due to not
2518 * using cv_broadcast, they will be woken up
2519 * just before arc_reclaim_thread() sleeps.
2521 mutex_enter(&arc_reclaim_lock
);
2522 if (!arc_is_overflowing())
2523 cv_signal(&arc_reclaim_waiters_cv
);
2524 mutex_exit(&arc_reclaim_lock
);
2526 ARCSTAT_BUMP(arcstat_mutex_miss
);
2530 multilist_sublist_unlock(mls
);
2532 return (bytes_evicted
);
2536 * Evict buffers from the given arc state, until we've removed the
2537 * specified number of bytes. Move the removed buffers to the
2538 * appropriate evict state.
2540 * This function makes a "best effort". It skips over any buffers
2541 * it can't get a hash_lock on, and so, may not catch all candidates.
2542 * It may also return without evicting as much space as requested.
2544 * If bytes is specified using the special value ARC_EVICT_ALL, this
2545 * will evict all available (i.e. unlocked and evictable) buffers from
2546 * the given arc state; which is used by arc_flush().
2549 arc_evict_state(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
2550 arc_buf_contents_t type
)
2552 uint64_t total_evicted
= 0;
2553 multilist_t
*ml
= &state
->arcs_list
[type
];
2555 arc_buf_hdr_t
**markers
;
2558 IMPLY(bytes
< 0, bytes
== ARC_EVICT_ALL
);
2560 num_sublists
= multilist_get_num_sublists(ml
);
2563 * If we've tried to evict from each sublist, made some
2564 * progress, but still have not hit the target number of bytes
2565 * to evict, we want to keep trying. The markers allow us to
2566 * pick up where we left off for each individual sublist, rather
2567 * than starting from the tail each time.
2569 markers
= kmem_zalloc(sizeof (*markers
) * num_sublists
, KM_SLEEP
);
2570 for (i
= 0; i
< num_sublists
; i
++) {
2571 multilist_sublist_t
*mls
;
2573 markers
[i
] = kmem_cache_alloc(hdr_full_cache
, KM_SLEEP
);
2576 * A b_spa of 0 is used to indicate that this header is
2577 * a marker. This fact is used in arc_adjust_type() and
2578 * arc_evict_state_impl().
2580 markers
[i
]->b_spa
= 0;
2582 mls
= multilist_sublist_lock(ml
, i
);
2583 multilist_sublist_insert_tail(mls
, markers
[i
]);
2584 multilist_sublist_unlock(mls
);
2588 * While we haven't hit our target number of bytes to evict, or
2589 * we're evicting all available buffers.
2591 while (total_evicted
< bytes
|| bytes
== ARC_EVICT_ALL
) {
2593 * Start eviction using a randomly selected sublist,
2594 * this is to try and evenly balance eviction across all
2595 * sublists. Always starting at the same sublist
2596 * (e.g. index 0) would cause evictions to favor certain
2597 * sublists over others.
2599 int sublist_idx
= multilist_get_random_index(ml
);
2600 uint64_t scan_evicted
= 0;
2602 for (i
= 0; i
< num_sublists
; i
++) {
2603 uint64_t bytes_remaining
;
2604 uint64_t bytes_evicted
;
2606 if (bytes
== ARC_EVICT_ALL
)
2607 bytes_remaining
= ARC_EVICT_ALL
;
2608 else if (total_evicted
< bytes
)
2609 bytes_remaining
= bytes
- total_evicted
;
2613 bytes_evicted
= arc_evict_state_impl(ml
, sublist_idx
,
2614 markers
[sublist_idx
], spa
, bytes_remaining
);
2616 scan_evicted
+= bytes_evicted
;
2617 total_evicted
+= bytes_evicted
;
2619 /* we've reached the end, wrap to the beginning */
2620 if (++sublist_idx
>= num_sublists
)
2625 * If we didn't evict anything during this scan, we have
2626 * no reason to believe we'll evict more during another
2627 * scan, so break the loop.
2629 if (scan_evicted
== 0) {
2630 /* This isn't possible, let's make that obvious */
2631 ASSERT3S(bytes
, !=, 0);
2634 * When bytes is ARC_EVICT_ALL, the only way to
2635 * break the loop is when scan_evicted is zero.
2636 * In that case, we actually have evicted enough,
2637 * so we don't want to increment the kstat.
2639 if (bytes
!= ARC_EVICT_ALL
) {
2640 ASSERT3S(total_evicted
, <, bytes
);
2641 ARCSTAT_BUMP(arcstat_evict_not_enough
);
2648 for (i
= 0; i
< num_sublists
; i
++) {
2649 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
2650 multilist_sublist_remove(mls
, markers
[i
]);
2651 multilist_sublist_unlock(mls
);
2653 kmem_cache_free(hdr_full_cache
, markers
[i
]);
2655 kmem_free(markers
, sizeof (*markers
) * num_sublists
);
2657 return (total_evicted
);
2661 * Flush all "evictable" data of the given type from the arc state
2662 * specified. This will not evict any "active" buffers (i.e. referenced).
2664 * When 'retry' is set to FALSE, the function will make a single pass
2665 * over the state and evict any buffers that it can. Since it doesn't
2666 * continually retry the eviction, it might end up leaving some buffers
2667 * in the ARC due to lock misses.
2669 * When 'retry' is set to TRUE, the function will continually retry the
2670 * eviction until *all* evictable buffers have been removed from the
2671 * state. As a result, if concurrent insertions into the state are
2672 * allowed (e.g. if the ARC isn't shutting down), this function might
2673 * wind up in an infinite loop, continually trying to evict buffers.
2676 arc_flush_state(arc_state_t
*state
, uint64_t spa
, arc_buf_contents_t type
,
2679 uint64_t evicted
= 0;
2681 while (state
->arcs_lsize
[type
] != 0) {
2682 evicted
+= arc_evict_state(state
, spa
, ARC_EVICT_ALL
, type
);
2692 * Helper function for arc_prune_async() it is responsible for safely
2693 * handling the execution of a registered arc_prune_func_t.
2696 arc_prune_task(void *ptr
)
2698 arc_prune_t
*ap
= (arc_prune_t
*)ptr
;
2699 arc_prune_func_t
*func
= ap
->p_pfunc
;
2702 func(ap
->p_adjust
, ap
->p_private
);
2704 /* Callback unregistered concurrently with execution */
2705 if (refcount_remove(&ap
->p_refcnt
, func
) == 0) {
2706 ASSERT(!list_link_active(&ap
->p_node
));
2707 refcount_destroy(&ap
->p_refcnt
);
2708 kmem_free(ap
, sizeof (*ap
));
2713 * Notify registered consumers they must drop holds on a portion of the ARC
2714 * buffered they reference. This provides a mechanism to ensure the ARC can
2715 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
2716 * is analogous to dnlc_reduce_cache() but more generic.
2718 * This operation is performed asynchronously so it may be safely called
2719 * in the context of the arc_reclaim_thread(). A reference is taken here
2720 * for each registered arc_prune_t and the arc_prune_task() is responsible
2721 * for releasing it once the registered arc_prune_func_t has completed.
2724 arc_prune_async(int64_t adjust
)
2728 mutex_enter(&arc_prune_mtx
);
2729 for (ap
= list_head(&arc_prune_list
); ap
!= NULL
;
2730 ap
= list_next(&arc_prune_list
, ap
)) {
2732 if (refcount_count(&ap
->p_refcnt
) >= 2)
2735 refcount_add(&ap
->p_refcnt
, ap
->p_pfunc
);
2736 ap
->p_adjust
= adjust
;
2737 taskq_dispatch(arc_prune_taskq
, arc_prune_task
, ap
, TQ_SLEEP
);
2738 ARCSTAT_BUMP(arcstat_prune
);
2740 mutex_exit(&arc_prune_mtx
);
2744 * Evict the specified number of bytes from the state specified,
2745 * restricting eviction to the spa and type given. This function
2746 * prevents us from trying to evict more from a state's list than
2747 * is "evictable", and to skip evicting altogether when passed a
2748 * negative value for "bytes". In contrast, arc_evict_state() will
2749 * evict everything it can, when passed a negative value for "bytes".
2752 arc_adjust_impl(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
2753 arc_buf_contents_t type
)
2757 if (bytes
> 0 && state
->arcs_lsize
[type
] > 0) {
2758 delta
= MIN(state
->arcs_lsize
[type
], bytes
);
2759 return (arc_evict_state(state
, spa
, delta
, type
));
2766 * The goal of this function is to evict enough meta data buffers from the
2767 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
2768 * more complicated than it appears because it is common for data buffers
2769 * to have holds on meta data buffers. In addition, dnode meta data buffers
2770 * will be held by the dnodes in the block preventing them from being freed.
2771 * This means we can't simply traverse the ARC and expect to always find
2772 * enough unheld meta data buffer to release.
2774 * Therefore, this function has been updated to make alternating passes
2775 * over the ARC releasing data buffers and then newly unheld meta data
2776 * buffers. This ensures forward progress is maintained and arc_meta_used
2777 * will decrease. Normally this is sufficient, but if required the ARC
2778 * will call the registered prune callbacks causing dentry and inodes to
2779 * be dropped from the VFS cache. This will make dnode meta data buffers
2780 * available for reclaim.
2783 arc_adjust_meta_balanced(void)
2785 int64_t adjustmnt
, delta
, prune
= 0;
2786 uint64_t total_evicted
= 0;
2787 arc_buf_contents_t type
= ARC_BUFC_DATA
;
2788 int restarts
= MAX(zfs_arc_meta_adjust_restarts
, 0);
2792 * This slightly differs than the way we evict from the mru in
2793 * arc_adjust because we don't have a "target" value (i.e. no
2794 * "meta" arc_p). As a result, I think we can completely
2795 * cannibalize the metadata in the MRU before we evict the
2796 * metadata from the MFU. I think we probably need to implement a
2797 * "metadata arc_p" value to do this properly.
2799 adjustmnt
= arc_meta_used
- arc_meta_limit
;
2801 if (adjustmnt
> 0 && arc_mru
->arcs_lsize
[type
] > 0) {
2802 delta
= MIN(arc_mru
->arcs_lsize
[type
], adjustmnt
);
2803 total_evicted
+= arc_adjust_impl(arc_mru
, 0, delta
, type
);
2808 * We can't afford to recalculate adjustmnt here. If we do,
2809 * new metadata buffers can sneak into the MRU or ANON lists,
2810 * thus penalize the MFU metadata. Although the fudge factor is
2811 * small, it has been empirically shown to be significant for
2812 * certain workloads (e.g. creating many empty directories). As
2813 * such, we use the original calculation for adjustmnt, and
2814 * simply decrement the amount of data evicted from the MRU.
2817 if (adjustmnt
> 0 && arc_mfu
->arcs_lsize
[type
] > 0) {
2818 delta
= MIN(arc_mfu
->arcs_lsize
[type
], adjustmnt
);
2819 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, delta
, type
);
2822 adjustmnt
= arc_meta_used
- arc_meta_limit
;
2824 if (adjustmnt
> 0 && arc_mru_ghost
->arcs_lsize
[type
] > 0) {
2825 delta
= MIN(adjustmnt
,
2826 arc_mru_ghost
->arcs_lsize
[type
]);
2827 total_evicted
+= arc_adjust_impl(arc_mru_ghost
, 0, delta
, type
);
2831 if (adjustmnt
> 0 && arc_mfu_ghost
->arcs_lsize
[type
] > 0) {
2832 delta
= MIN(adjustmnt
,
2833 arc_mfu_ghost
->arcs_lsize
[type
]);
2834 total_evicted
+= arc_adjust_impl(arc_mfu_ghost
, 0, delta
, type
);
2838 * If after attempting to make the requested adjustment to the ARC
2839 * the meta limit is still being exceeded then request that the
2840 * higher layers drop some cached objects which have holds on ARC
2841 * meta buffers. Requests to the upper layers will be made with
2842 * increasingly large scan sizes until the ARC is below the limit.
2844 if (arc_meta_used
> arc_meta_limit
) {
2845 if (type
== ARC_BUFC_DATA
) {
2846 type
= ARC_BUFC_METADATA
;
2848 type
= ARC_BUFC_DATA
;
2850 if (zfs_arc_meta_prune
) {
2851 prune
+= zfs_arc_meta_prune
;
2852 arc_prune_async(prune
);
2861 return (total_evicted
);
2865 * Evict metadata buffers from the cache, such that arc_meta_used is
2866 * capped by the arc_meta_limit tunable.
2869 arc_adjust_meta_only(void)
2871 uint64_t total_evicted
= 0;
2875 * If we're over the meta limit, we want to evict enough
2876 * metadata to get back under the meta limit. We don't want to
2877 * evict so much that we drop the MRU below arc_p, though. If
2878 * we're over the meta limit more than we're over arc_p, we
2879 * evict some from the MRU here, and some from the MFU below.
2881 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
2882 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
2883 refcount_count(&arc_mru
->arcs_size
) - arc_p
));
2885 total_evicted
+= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
2888 * Similar to the above, we want to evict enough bytes to get us
2889 * below the meta limit, but not so much as to drop us below the
2890 * space alloted to the MFU (which is defined as arc_c - arc_p).
2892 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
2893 (int64_t)(refcount_count(&arc_mfu
->arcs_size
) - (arc_c
- arc_p
)));
2895 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
2897 return (total_evicted
);
2901 arc_adjust_meta(void)
2903 if (zfs_arc_meta_strategy
== ARC_STRATEGY_META_ONLY
)
2904 return (arc_adjust_meta_only());
2906 return (arc_adjust_meta_balanced());
2910 * Return the type of the oldest buffer in the given arc state
2912 * This function will select a random sublist of type ARC_BUFC_DATA and
2913 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
2914 * is compared, and the type which contains the "older" buffer will be
2917 static arc_buf_contents_t
2918 arc_adjust_type(arc_state_t
*state
)
2920 multilist_t
*data_ml
= &state
->arcs_list
[ARC_BUFC_DATA
];
2921 multilist_t
*meta_ml
= &state
->arcs_list
[ARC_BUFC_METADATA
];
2922 int data_idx
= multilist_get_random_index(data_ml
);
2923 int meta_idx
= multilist_get_random_index(meta_ml
);
2924 multilist_sublist_t
*data_mls
;
2925 multilist_sublist_t
*meta_mls
;
2926 arc_buf_contents_t type
;
2927 arc_buf_hdr_t
*data_hdr
;
2928 arc_buf_hdr_t
*meta_hdr
;
2931 * We keep the sublist lock until we're finished, to prevent
2932 * the headers from being destroyed via arc_evict_state().
2934 data_mls
= multilist_sublist_lock(data_ml
, data_idx
);
2935 meta_mls
= multilist_sublist_lock(meta_ml
, meta_idx
);
2938 * These two loops are to ensure we skip any markers that
2939 * might be at the tail of the lists due to arc_evict_state().
2942 for (data_hdr
= multilist_sublist_tail(data_mls
); data_hdr
!= NULL
;
2943 data_hdr
= multilist_sublist_prev(data_mls
, data_hdr
)) {
2944 if (data_hdr
->b_spa
!= 0)
2948 for (meta_hdr
= multilist_sublist_tail(meta_mls
); meta_hdr
!= NULL
;
2949 meta_hdr
= multilist_sublist_prev(meta_mls
, meta_hdr
)) {
2950 if (meta_hdr
->b_spa
!= 0)
2954 if (data_hdr
== NULL
&& meta_hdr
== NULL
) {
2955 type
= ARC_BUFC_DATA
;
2956 } else if (data_hdr
== NULL
) {
2957 ASSERT3P(meta_hdr
, !=, NULL
);
2958 type
= ARC_BUFC_METADATA
;
2959 } else if (meta_hdr
== NULL
) {
2960 ASSERT3P(data_hdr
, !=, NULL
);
2961 type
= ARC_BUFC_DATA
;
2963 ASSERT3P(data_hdr
, !=, NULL
);
2964 ASSERT3P(meta_hdr
, !=, NULL
);
2966 /* The headers can't be on the sublist without an L1 header */
2967 ASSERT(HDR_HAS_L1HDR(data_hdr
));
2968 ASSERT(HDR_HAS_L1HDR(meta_hdr
));
2970 if (data_hdr
->b_l1hdr
.b_arc_access
<
2971 meta_hdr
->b_l1hdr
.b_arc_access
) {
2972 type
= ARC_BUFC_DATA
;
2974 type
= ARC_BUFC_METADATA
;
2978 multilist_sublist_unlock(meta_mls
);
2979 multilist_sublist_unlock(data_mls
);
2985 * Evict buffers from the cache, such that arc_size is capped by arc_c.
2990 uint64_t total_evicted
= 0;
2995 * If we're over arc_meta_limit, we want to correct that before
2996 * potentially evicting data buffers below.
2998 total_evicted
+= arc_adjust_meta();
3003 * If we're over the target cache size, we want to evict enough
3004 * from the list to get back to our target size. We don't want
3005 * to evict too much from the MRU, such that it drops below
3006 * arc_p. So, if we're over our target cache size more than
3007 * the MRU is over arc_p, we'll evict enough to get back to
3008 * arc_p here, and then evict more from the MFU below.
3010 target
= MIN((int64_t)(arc_size
- arc_c
),
3011 (int64_t)(refcount_count(&arc_anon
->arcs_size
) +
3012 refcount_count(&arc_mru
->arcs_size
) + arc_meta_used
- arc_p
));
3015 * If we're below arc_meta_min, always prefer to evict data.
3016 * Otherwise, try to satisfy the requested number of bytes to
3017 * evict from the type which contains older buffers; in an
3018 * effort to keep newer buffers in the cache regardless of their
3019 * type. If we cannot satisfy the number of bytes from this
3020 * type, spill over into the next type.
3022 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
3023 arc_meta_used
> arc_meta_min
) {
3024 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3025 total_evicted
+= bytes
;
3028 * If we couldn't evict our target number of bytes from
3029 * metadata, we try to get the rest from data.
3034 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3036 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
3037 total_evicted
+= bytes
;
3040 * If we couldn't evict our target number of bytes from
3041 * data, we try to get the rest from metadata.
3046 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
3052 * Now that we've tried to evict enough from the MRU to get its
3053 * size back to arc_p, if we're still above the target cache
3054 * size, we evict the rest from the MFU.
3056 target
= arc_size
- arc_c
;
3058 if (arc_adjust_type(arc_mfu
) == ARC_BUFC_METADATA
&&
3059 arc_meta_used
> arc_meta_min
) {
3060 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3061 total_evicted
+= bytes
;
3064 * If we couldn't evict our target number of bytes from
3065 * metadata, we try to get the rest from data.
3070 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3072 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
3073 total_evicted
+= bytes
;
3076 * If we couldn't evict our target number of bytes from
3077 * data, we try to get the rest from data.
3082 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
3086 * Adjust ghost lists
3088 * In addition to the above, the ARC also defines target values
3089 * for the ghost lists. The sum of the mru list and mru ghost
3090 * list should never exceed the target size of the cache, and
3091 * the sum of the mru list, mfu list, mru ghost list, and mfu
3092 * ghost list should never exceed twice the target size of the
3093 * cache. The following logic enforces these limits on the ghost
3094 * caches, and evicts from them as needed.
3096 target
= refcount_count(&arc_mru
->arcs_size
) +
3097 refcount_count(&arc_mru_ghost
->arcs_size
) - arc_c
;
3099 bytes
= arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_DATA
);
3100 total_evicted
+= bytes
;
3105 arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_METADATA
);
3108 * We assume the sum of the mru list and mfu list is less than
3109 * or equal to arc_c (we enforced this above), which means we
3110 * can use the simpler of the two equations below:
3112 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3113 * mru ghost + mfu ghost <= arc_c
3115 target
= refcount_count(&arc_mru_ghost
->arcs_size
) +
3116 refcount_count(&arc_mfu_ghost
->arcs_size
) - arc_c
;
3118 bytes
= arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_DATA
);
3119 total_evicted
+= bytes
;
3124 arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_METADATA
);
3126 return (total_evicted
);
3130 arc_do_user_evicts(void)
3132 mutex_enter(&arc_user_evicts_lock
);
3133 while (arc_eviction_list
!= NULL
) {
3134 arc_buf_t
*buf
= arc_eviction_list
;
3135 arc_eviction_list
= buf
->b_next
;
3136 mutex_enter(&buf
->b_evict_lock
);
3138 mutex_exit(&buf
->b_evict_lock
);
3139 mutex_exit(&arc_user_evicts_lock
);
3141 if (buf
->b_efunc
!= NULL
)
3142 VERIFY0(buf
->b_efunc(buf
->b_private
));
3144 buf
->b_efunc
= NULL
;
3145 buf
->b_private
= NULL
;
3146 kmem_cache_free(buf_cache
, buf
);
3147 mutex_enter(&arc_user_evicts_lock
);
3149 mutex_exit(&arc_user_evicts_lock
);
3153 arc_flush(spa_t
*spa
, boolean_t retry
)
3158 * If retry is TRUE, a spa must not be specified since we have
3159 * no good way to determine if all of a spa's buffers have been
3160 * evicted from an arc state.
3162 ASSERT(!retry
|| spa
== 0);
3165 guid
= spa_load_guid(spa
);
3167 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_DATA
, retry
);
3168 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_METADATA
, retry
);
3170 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_DATA
, retry
);
3171 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_METADATA
, retry
);
3173 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3174 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3176 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_DATA
, retry
);
3177 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
3179 arc_do_user_evicts();
3180 ASSERT(spa
|| arc_eviction_list
== NULL
);
3184 arc_shrink(int64_t to_free
)
3186 if (arc_c
> arc_c_min
) {
3188 if (arc_c
> arc_c_min
+ to_free
)
3189 atomic_add_64(&arc_c
, -to_free
);
3193 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
3194 if (arc_c
> arc_size
)
3195 arc_c
= MAX(arc_size
, arc_c_min
);
3197 arc_p
= (arc_c
>> 1);
3198 ASSERT(arc_c
>= arc_c_min
);
3199 ASSERT((int64_t)arc_p
>= 0);
3202 if (arc_size
> arc_c
)
3203 (void) arc_adjust();
3206 typedef enum free_memory_reason_t
{
3211 FMR_PAGES_PP_MAXIMUM
,
3214 } free_memory_reason_t
;
3216 int64_t last_free_memory
;
3217 free_memory_reason_t last_free_reason
;
3221 * Additional reserve of pages for pp_reserve.
3223 int64_t arc_pages_pp_reserve
= 64;
3226 * Additional reserve of pages for swapfs.
3228 int64_t arc_swapfs_reserve
= 64;
3229 #endif /* _KERNEL */
3232 * Return the amount of memory that can be consumed before reclaim will be
3233 * needed. Positive if there is sufficient free memory, negative indicates
3234 * the amount of memory that needs to be freed up.
3237 arc_available_memory(void)
3239 int64_t lowest
= INT64_MAX
;
3240 free_memory_reason_t r
= FMR_UNKNOWN
;
3244 pgcnt_t needfree
= btop(arc_need_free
);
3245 pgcnt_t lotsfree
= btop(arc_sys_free
);
3246 pgcnt_t desfree
= 0;
3250 n
= PAGESIZE
* (-needfree
);
3258 * check that we're out of range of the pageout scanner. It starts to
3259 * schedule paging if freemem is less than lotsfree and needfree.
3260 * lotsfree is the high-water mark for pageout, and needfree is the
3261 * number of needed free pages. We add extra pages here to make sure
3262 * the scanner doesn't start up while we're freeing memory.
3264 n
= PAGESIZE
* (freemem
- lotsfree
- needfree
- desfree
);
3272 * check to make sure that swapfs has enough space so that anon
3273 * reservations can still succeed. anon_resvmem() checks that the
3274 * availrmem is greater than swapfs_minfree, and the number of reserved
3275 * swap pages. We also add a bit of extra here just to prevent
3276 * circumstances from getting really dire.
3278 n
= PAGESIZE
* (availrmem
- swapfs_minfree
- swapfs_reserve
-
3279 desfree
- arc_swapfs_reserve
);
3282 r
= FMR_SWAPFS_MINFREE
;
3287 * Check that we have enough availrmem that memory locking (e.g., via
3288 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3289 * stores the number of pages that cannot be locked; when availrmem
3290 * drops below pages_pp_maximum, page locking mechanisms such as
3291 * page_pp_lock() will fail.)
3293 n
= PAGESIZE
* (availrmem
- pages_pp_maximum
-
3294 arc_pages_pp_reserve
);
3297 r
= FMR_PAGES_PP_MAXIMUM
;
3303 * If we're on an i386 platform, it's possible that we'll exhaust the
3304 * kernel heap space before we ever run out of available physical
3305 * memory. Most checks of the size of the heap_area compare against
3306 * tune.t_minarmem, which is the minimum available real memory that we
3307 * can have in the system. However, this is generally fixed at 25 pages
3308 * which is so low that it's useless. In this comparison, we seek to
3309 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3310 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3313 n
= vmem_size(heap_arena
, VMEM_FREE
) -
3314 (vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
) >> 2);
3322 * If zio data pages are being allocated out of a separate heap segment,
3323 * then enforce that the size of available vmem for this arena remains
3324 * above about 1/16th free.
3326 * Note: The 1/16th arena free requirement was put in place
3327 * to aggressively evict memory from the arc in order to avoid
3328 * memory fragmentation issues.
3330 if (zio_arena
!= NULL
) {
3331 n
= vmem_size(zio_arena
, VMEM_FREE
) -
3332 (vmem_size(zio_arena
, VMEM_ALLOC
) >> 4);
3339 /* Every 100 calls, free a small amount */
3340 if (spa_get_random(100) == 0)
3342 #endif /* _KERNEL */
3344 last_free_memory
= lowest
;
3345 last_free_reason
= r
;
3351 * Determine if the system is under memory pressure and is asking
3352 * to reclaim memory. A return value of TRUE indicates that the system
3353 * is under memory pressure and that the arc should adjust accordingly.
3356 arc_reclaim_needed(void)
3358 return (arc_available_memory() < 0);
3362 arc_kmem_reap_now(void)
3365 kmem_cache_t
*prev_cache
= NULL
;
3366 kmem_cache_t
*prev_data_cache
= NULL
;
3367 extern kmem_cache_t
*zio_buf_cache
[];
3368 extern kmem_cache_t
*zio_data_buf_cache
[];
3369 extern kmem_cache_t
*range_seg_cache
;
3371 if ((arc_meta_used
>= arc_meta_limit
) && zfs_arc_meta_prune
) {
3373 * We are exceeding our meta-data cache limit.
3374 * Prune some entries to release holds on meta-data.
3376 arc_prune_async(zfs_arc_meta_prune
);
3379 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
3381 /* reach upper limit of cache size on 32-bit */
3382 if (zio_buf_cache
[i
] == NULL
)
3385 if (zio_buf_cache
[i
] != prev_cache
) {
3386 prev_cache
= zio_buf_cache
[i
];
3387 kmem_cache_reap_now(zio_buf_cache
[i
]);
3389 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
3390 prev_data_cache
= zio_data_buf_cache
[i
];
3391 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
3394 kmem_cache_reap_now(buf_cache
);
3395 kmem_cache_reap_now(hdr_full_cache
);
3396 kmem_cache_reap_now(hdr_l2only_cache
);
3397 kmem_cache_reap_now(range_seg_cache
);
3399 if (zio_arena
!= NULL
) {
3401 * Ask the vmem arena to reclaim unused memory from its
3404 vmem_qcache_reap(zio_arena
);
3409 * Threads can block in arc_get_data_buf() waiting for this thread to evict
3410 * enough data and signal them to proceed. When this happens, the threads in
3411 * arc_get_data_buf() are sleeping while holding the hash lock for their
3412 * particular arc header. Thus, we must be careful to never sleep on a
3413 * hash lock in this thread. This is to prevent the following deadlock:
3415 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3416 * waiting for the reclaim thread to signal it.
3418 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3419 * fails, and goes to sleep forever.
3421 * This possible deadlock is avoided by always acquiring a hash lock
3422 * using mutex_tryenter() from arc_reclaim_thread().
3425 arc_reclaim_thread(void)
3427 fstrans_cookie_t cookie
= spl_fstrans_mark();
3428 clock_t growtime
= 0;
3431 CALLB_CPR_INIT(&cpr
, &arc_reclaim_lock
, callb_generic_cpr
, FTAG
);
3433 mutex_enter(&arc_reclaim_lock
);
3434 while (!arc_reclaim_thread_exit
) {
3436 int64_t free_memory
= arc_available_memory();
3437 uint64_t evicted
= 0;
3439 arc_tuning_update();
3441 mutex_exit(&arc_reclaim_lock
);
3443 if (free_memory
< 0) {
3445 arc_no_grow
= B_TRUE
;
3449 * Wait at least zfs_grow_retry (default 5) seconds
3450 * before considering growing.
3452 growtime
= ddi_get_lbolt() + (arc_grow_retry
* hz
);
3454 arc_kmem_reap_now();
3457 * If we are still low on memory, shrink the ARC
3458 * so that we have arc_shrink_min free space.
3460 free_memory
= arc_available_memory();
3462 to_free
= (arc_c
>> arc_shrink_shift
) - free_memory
;
3465 to_free
= MAX(to_free
, arc_need_free
);
3467 arc_shrink(to_free
);
3469 } else if (free_memory
< arc_c
>> arc_no_grow_shift
) {
3470 arc_no_grow
= B_TRUE
;
3471 } else if (ddi_get_lbolt() >= growtime
) {
3472 arc_no_grow
= B_FALSE
;
3475 evicted
= arc_adjust();
3477 mutex_enter(&arc_reclaim_lock
);
3480 * If evicted is zero, we couldn't evict anything via
3481 * arc_adjust(). This could be due to hash lock
3482 * collisions, but more likely due to the majority of
3483 * arc buffers being unevictable. Therefore, even if
3484 * arc_size is above arc_c, another pass is unlikely to
3485 * be helpful and could potentially cause us to enter an
3488 if (arc_size
<= arc_c
|| evicted
== 0) {
3490 * We're either no longer overflowing, or we
3491 * can't evict anything more, so we should wake
3492 * up any threads before we go to sleep and clear
3493 * arc_need_free since nothing more can be done.
3495 cv_broadcast(&arc_reclaim_waiters_cv
);
3499 * Block until signaled, or after one second (we
3500 * might need to perform arc_kmem_reap_now()
3501 * even if we aren't being signalled)
3503 CALLB_CPR_SAFE_BEGIN(&cpr
);
3504 (void) cv_timedwait_sig(&arc_reclaim_thread_cv
,
3505 &arc_reclaim_lock
, ddi_get_lbolt() + hz
);
3506 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_lock
);
3510 arc_reclaim_thread_exit
= FALSE
;
3511 cv_broadcast(&arc_reclaim_thread_cv
);
3512 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_lock */
3513 spl_fstrans_unmark(cookie
);
3518 arc_user_evicts_thread(void)
3520 fstrans_cookie_t cookie
= spl_fstrans_mark();
3523 CALLB_CPR_INIT(&cpr
, &arc_user_evicts_lock
, callb_generic_cpr
, FTAG
);
3525 mutex_enter(&arc_user_evicts_lock
);
3526 while (!arc_user_evicts_thread_exit
) {
3527 mutex_exit(&arc_user_evicts_lock
);
3529 arc_do_user_evicts();
3532 * This is necessary in order for the mdb ::arc dcmd to
3533 * show up to date information. Since the ::arc command
3534 * does not call the kstat's update function, without
3535 * this call, the command may show stale stats for the
3536 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3537 * with this change, the data might be up to 1 second
3538 * out of date; but that should suffice. The arc_state_t
3539 * structures can be queried directly if more accurate
3540 * information is needed.
3542 if (arc_ksp
!= NULL
)
3543 arc_ksp
->ks_update(arc_ksp
, KSTAT_READ
);
3545 mutex_enter(&arc_user_evicts_lock
);
3548 * Block until signaled, or after one second (we need to
3549 * call the arc's kstat update function regularly).
3551 CALLB_CPR_SAFE_BEGIN(&cpr
);
3552 (void) cv_timedwait_sig(&arc_user_evicts_cv
,
3553 &arc_user_evicts_lock
, ddi_get_lbolt() + hz
);
3554 CALLB_CPR_SAFE_END(&cpr
, &arc_user_evicts_lock
);
3557 arc_user_evicts_thread_exit
= FALSE
;
3558 cv_broadcast(&arc_user_evicts_cv
);
3559 CALLB_CPR_EXIT(&cpr
); /* drops arc_user_evicts_lock */
3560 spl_fstrans_unmark(cookie
);
3566 * Determine the amount of memory eligible for eviction contained in the
3567 * ARC. All clean data reported by the ghost lists can always be safely
3568 * evicted. Due to arc_c_min, the same does not hold for all clean data
3569 * contained by the regular mru and mfu lists.
3571 * In the case of the regular mru and mfu lists, we need to report as
3572 * much clean data as possible, such that evicting that same reported
3573 * data will not bring arc_size below arc_c_min. Thus, in certain
3574 * circumstances, the total amount of clean data in the mru and mfu
3575 * lists might not actually be evictable.
3577 * The following two distinct cases are accounted for:
3579 * 1. 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 greater than or equal to arc_c_min.
3582 * (i.e. amount of dirty data >= arc_c_min)
3584 * This is the easy case; all clean data contained by the mru and mfu
3585 * lists is evictable. Evicting all clean data can only drop arc_size
3586 * to the amount of dirty data, which is greater than arc_c_min.
3588 * 2. The sum of the amount of dirty data contained by both the mru and
3589 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
3590 * is less than arc_c_min.
3591 * (i.e. arc_c_min > amount of dirty data)
3593 * 2.1. arc_size is greater than or equal arc_c_min.
3594 * (i.e. arc_size >= arc_c_min > amount of dirty data)
3596 * In this case, not all clean data from the regular mru and mfu
3597 * lists is actually evictable; we must leave enough clean data
3598 * to keep arc_size above arc_c_min. Thus, the maximum amount of
3599 * evictable data from the two lists combined, is exactly the
3600 * difference between arc_size and arc_c_min.
3602 * 2.2. arc_size is less than arc_c_min
3603 * (i.e. arc_c_min > arc_size > amount of dirty data)
3605 * In this case, none of the data contained in the mru and mfu
3606 * lists is evictable, even if it's clean. Since arc_size is
3607 * already below arc_c_min, evicting any more would only
3608 * increase this negative difference.
3611 arc_evictable_memory(void) {
3612 uint64_t arc_clean
=
3613 arc_mru
->arcs_lsize
[ARC_BUFC_DATA
] +
3614 arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] +
3615 arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
] +
3616 arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
];
3617 uint64_t ghost_clean
=
3618 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
3619 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_METADATA
] +
3620 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
3621 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_METADATA
];
3622 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
3624 if (arc_dirty
>= arc_c_min
)
3625 return (ghost_clean
+ arc_clean
);
3627 return (ghost_clean
+ MAX((int64_t)arc_size
- (int64_t)arc_c_min
, 0));
3631 * If sc->nr_to_scan is zero, the caller is requesting a query of the
3632 * number of objects which can potentially be freed. If it is nonzero,
3633 * the request is to free that many objects.
3635 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
3636 * in struct shrinker and also require the shrinker to return the number
3639 * Older kernels require the shrinker to return the number of freeable
3640 * objects following the freeing of nr_to_free.
3642 static spl_shrinker_t
3643 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
3647 /* The arc is considered warm once reclaim has occurred */
3648 if (unlikely(arc_warm
== B_FALSE
))
3651 /* Return the potential number of reclaimable pages */
3652 pages
= btop((int64_t)arc_evictable_memory());
3653 if (sc
->nr_to_scan
== 0)
3656 /* Not allowed to perform filesystem reclaim */
3657 if (!(sc
->gfp_mask
& __GFP_FS
))
3658 return (SHRINK_STOP
);
3660 /* Reclaim in progress */
3661 if (mutex_tryenter(&arc_reclaim_lock
) == 0)
3662 return (SHRINK_STOP
);
3664 mutex_exit(&arc_reclaim_lock
);
3667 * Evict the requested number of pages by shrinking arc_c the
3668 * requested amount. If there is nothing left to evict just
3669 * reap whatever we can from the various arc slabs.
3672 arc_shrink(ptob(sc
->nr_to_scan
));
3673 arc_kmem_reap_now();
3674 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
3675 pages
= MAX(pages
- btop(arc_evictable_memory()), 0);
3677 pages
= btop(arc_evictable_memory());
3680 arc_kmem_reap_now();
3681 pages
= SHRINK_STOP
;
3685 * We've reaped what we can, wake up threads.
3687 cv_broadcast(&arc_reclaim_waiters_cv
);
3690 * When direct reclaim is observed it usually indicates a rapid
3691 * increase in memory pressure. This occurs because the kswapd
3692 * threads were unable to asynchronously keep enough free memory
3693 * available. In this case set arc_no_grow to briefly pause arc
3694 * growth to avoid compounding the memory pressure.
3696 if (current_is_kswapd()) {
3697 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
3699 arc_no_grow
= B_TRUE
;
3700 arc_need_free
= ptob(sc
->nr_to_scan
);
3701 ARCSTAT_BUMP(arcstat_memory_direct_count
);
3706 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
3708 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
3709 #endif /* _KERNEL */
3712 * Adapt arc info given the number of bytes we are trying to add and
3713 * the state that we are comming from. This function is only called
3714 * when we are adding new content to the cache.
3717 arc_adapt(int bytes
, arc_state_t
*state
)
3720 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
3721 int64_t mrug_size
= refcount_count(&arc_mru_ghost
->arcs_size
);
3722 int64_t mfug_size
= refcount_count(&arc_mfu_ghost
->arcs_size
);
3724 if (state
== arc_l2c_only
)
3729 * Adapt the target size of the MRU list:
3730 * - if we just hit in the MRU ghost list, then increase
3731 * the target size of the MRU list.
3732 * - if we just hit in the MFU ghost list, then increase
3733 * the target size of the MFU list by decreasing the
3734 * target size of the MRU list.
3736 if (state
== arc_mru_ghost
) {
3737 mult
= (mrug_size
>= mfug_size
) ? 1 : (mfug_size
/ mrug_size
);
3738 if (!zfs_arc_p_dampener_disable
)
3739 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
3741 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
3742 } else if (state
== arc_mfu_ghost
) {
3745 mult
= (mfug_size
>= mrug_size
) ? 1 : (mrug_size
/ mfug_size
);
3746 if (!zfs_arc_p_dampener_disable
)
3747 mult
= MIN(mult
, 10);
3749 delta
= MIN(bytes
* mult
, arc_p
);
3750 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
3752 ASSERT((int64_t)arc_p
>= 0);
3754 if (arc_reclaim_needed()) {
3755 cv_signal(&arc_reclaim_thread_cv
);
3762 if (arc_c
>= arc_c_max
)
3766 * If we're within (2 * maxblocksize) bytes of the target
3767 * cache size, increment the target cache size
3769 ASSERT3U(arc_c
, >=, 2ULL << SPA_MAXBLOCKSHIFT
);
3770 arc_c
= MAX(arc_c
, 2ULL << SPA_MAXBLOCKSHIFT
);
3771 if (arc_size
>= arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
3772 atomic_add_64(&arc_c
, (int64_t)bytes
);
3773 if (arc_c
> arc_c_max
)
3775 else if (state
== arc_anon
)
3776 atomic_add_64(&arc_p
, (int64_t)bytes
);
3780 ASSERT((int64_t)arc_p
>= 0);
3784 * Check if arc_size has grown past our upper threshold, determined by
3785 * zfs_arc_overflow_shift.
3788 arc_is_overflowing(void)
3790 /* Always allow at least one block of overflow */
3791 uint64_t overflow
= MAX(SPA_MAXBLOCKSIZE
,
3792 arc_c
>> zfs_arc_overflow_shift
);
3794 return (arc_size
>= arc_c
+ overflow
);
3798 * The buffer, supplied as the first argument, needs a data block. If we
3799 * are hitting the hard limit for the cache size, we must sleep, waiting
3800 * for the eviction thread to catch up. If we're past the target size
3801 * but below the hard limit, we'll only signal the reclaim thread and
3805 arc_get_data_buf(arc_buf_t
*buf
)
3807 arc_state_t
*state
= buf
->b_hdr
->b_l1hdr
.b_state
;
3808 uint64_t size
= buf
->b_hdr
->b_size
;
3809 arc_buf_contents_t type
= arc_buf_type(buf
->b_hdr
);
3811 arc_adapt(size
, state
);
3814 * If arc_size is currently overflowing, and has grown past our
3815 * upper limit, we must be adding data faster than the evict
3816 * thread can evict. Thus, to ensure we don't compound the
3817 * problem by adding more data and forcing arc_size to grow even
3818 * further past it's target size, we halt and wait for the
3819 * eviction thread to catch up.
3821 * It's also possible that the reclaim thread is unable to evict
3822 * enough buffers to get arc_size below the overflow limit (e.g.
3823 * due to buffers being un-evictable, or hash lock collisions).
3824 * In this case, we want to proceed regardless if we're
3825 * overflowing; thus we don't use a while loop here.
3827 if (arc_is_overflowing()) {
3828 mutex_enter(&arc_reclaim_lock
);
3831 * Now that we've acquired the lock, we may no longer be
3832 * over the overflow limit, lets check.
3834 * We're ignoring the case of spurious wake ups. If that
3835 * were to happen, it'd let this thread consume an ARC
3836 * buffer before it should have (i.e. before we're under
3837 * the overflow limit and were signalled by the reclaim
3838 * thread). As long as that is a rare occurrence, it
3839 * shouldn't cause any harm.
3841 if (arc_is_overflowing()) {
3842 cv_signal(&arc_reclaim_thread_cv
);
3843 cv_wait(&arc_reclaim_waiters_cv
, &arc_reclaim_lock
);
3846 mutex_exit(&arc_reclaim_lock
);
3849 if (type
== ARC_BUFC_METADATA
) {
3850 buf
->b_data
= zio_buf_alloc(size
);
3851 arc_space_consume(size
, ARC_SPACE_META
);
3853 ASSERT(type
== ARC_BUFC_DATA
);
3854 buf
->b_data
= zio_data_buf_alloc(size
);
3855 arc_space_consume(size
, ARC_SPACE_DATA
);
3859 * Update the state size. Note that ghost states have a
3860 * "ghost size" and so don't need to be updated.
3862 if (!GHOST_STATE(buf
->b_hdr
->b_l1hdr
.b_state
)) {
3863 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3864 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
3866 (void) refcount_add_many(&state
->arcs_size
, size
, buf
);
3869 * If this is reached via arc_read, the link is
3870 * protected by the hash lock. If reached via
3871 * arc_buf_alloc, the header should not be accessed by
3872 * any other thread. And, if reached via arc_read_done,
3873 * the hash lock will protect it if it's found in the
3874 * hash table; otherwise no other thread should be
3875 * trying to [add|remove]_reference it.
3877 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
3878 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3879 atomic_add_64(&hdr
->b_l1hdr
.b_state
->arcs_lsize
[type
],
3883 * If we are growing the cache, and we are adding anonymous
3884 * data, and we have outgrown arc_p, update arc_p
3886 if (arc_size
< arc_c
&& hdr
->b_l1hdr
.b_state
== arc_anon
&&
3887 (refcount_count(&arc_anon
->arcs_size
) +
3888 refcount_count(&arc_mru
->arcs_size
) > arc_p
))
3889 arc_p
= MIN(arc_c
, arc_p
+ size
);
3894 * This routine is called whenever a buffer is accessed.
3895 * NOTE: the hash lock is dropped in this function.
3898 arc_access(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
3902 ASSERT(MUTEX_HELD(hash_lock
));
3903 ASSERT(HDR_HAS_L1HDR(hdr
));
3905 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
3907 * This buffer is not in the cache, and does not
3908 * appear in our "ghost" list. Add the new buffer
3912 ASSERT0(hdr
->b_l1hdr
.b_arc_access
);
3913 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3914 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
3915 arc_change_state(arc_mru
, hdr
, hash_lock
);
3917 } else if (hdr
->b_l1hdr
.b_state
== arc_mru
) {
3918 now
= ddi_get_lbolt();
3921 * If this buffer is here because of a prefetch, then either:
3922 * - clear the flag if this is a "referencing" read
3923 * (any subsequent access will bump this into the MFU state).
3925 * - move the buffer to the head of the list if this is
3926 * another prefetch (to make it less likely to be evicted).
3928 if (HDR_PREFETCH(hdr
)) {
3929 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
3930 /* link protected by hash lock */
3931 ASSERT(multilist_link_active(
3932 &hdr
->b_l1hdr
.b_arc_node
));
3934 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
3935 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
3936 ARCSTAT_BUMP(arcstat_mru_hits
);
3938 hdr
->b_l1hdr
.b_arc_access
= now
;
3943 * This buffer has been "accessed" only once so far,
3944 * but it is still in the cache. Move it to the MFU
3947 if (ddi_time_after(now
, hdr
->b_l1hdr
.b_arc_access
+
3950 * More than 125ms have passed since we
3951 * instantiated this buffer. Move it to the
3952 * most frequently used state.
3954 hdr
->b_l1hdr
.b_arc_access
= now
;
3955 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
3956 arc_change_state(arc_mfu
, hdr
, hash_lock
);
3958 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
3959 ARCSTAT_BUMP(arcstat_mru_hits
);
3960 } else if (hdr
->b_l1hdr
.b_state
== arc_mru_ghost
) {
3961 arc_state_t
*new_state
;
3963 * This buffer has been "accessed" recently, but
3964 * was evicted from the cache. Move it to the
3968 if (HDR_PREFETCH(hdr
)) {
3969 new_state
= arc_mru
;
3970 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0)
3971 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
3972 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
3974 new_state
= arc_mfu
;
3975 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
3978 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3979 arc_change_state(new_state
, hdr
, hash_lock
);
3981 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_ghost_hits
);
3982 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
3983 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu
) {
3985 * This buffer has been accessed more than once and is
3986 * still in the cache. Keep it in the MFU state.
3988 * NOTE: an add_reference() that occurred when we did
3989 * the arc_read() will have kicked this off the list.
3990 * If it was a prefetch, we will explicitly move it to
3991 * the head of the list now.
3993 if ((HDR_PREFETCH(hdr
)) != 0) {
3994 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3995 /* link protected by hash_lock */
3996 ASSERT(multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3998 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_hits
);
3999 ARCSTAT_BUMP(arcstat_mfu_hits
);
4000 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4001 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu_ghost
) {
4002 arc_state_t
*new_state
= arc_mfu
;
4004 * This buffer has been accessed more than once but has
4005 * been evicted from the cache. Move it back to the
4009 if (HDR_PREFETCH(hdr
)) {
4011 * This is a prefetch access...
4012 * move this block back to the MRU state.
4014 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
4015 new_state
= arc_mru
;
4018 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4019 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4020 arc_change_state(new_state
, hdr
, hash_lock
);
4022 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_ghost_hits
);
4023 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
4024 } else if (hdr
->b_l1hdr
.b_state
== arc_l2c_only
) {
4026 * This buffer is on the 2nd Level ARC.
4029 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
4030 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
4031 arc_change_state(arc_mfu
, hdr
, hash_lock
);
4033 cmn_err(CE_PANIC
, "invalid arc state 0x%p",
4034 hdr
->b_l1hdr
.b_state
);
4038 /* a generic arc_done_func_t which you can use */
4041 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4043 if (zio
== NULL
|| zio
->io_error
== 0)
4044 bcopy(buf
->b_data
, arg
, buf
->b_hdr
->b_size
);
4045 VERIFY(arc_buf_remove_ref(buf
, arg
));
4048 /* a generic arc_done_func_t */
4050 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
4052 arc_buf_t
**bufp
= arg
;
4053 if (zio
&& zio
->io_error
) {
4054 VERIFY(arc_buf_remove_ref(buf
, arg
));
4058 ASSERT(buf
->b_data
);
4063 arc_read_done(zio_t
*zio
)
4067 arc_buf_t
*abuf
; /* buffer we're assigning to callback */
4068 kmutex_t
*hash_lock
= NULL
;
4069 arc_callback_t
*callback_list
, *acb
;
4070 int freeable
= FALSE
;
4072 buf
= zio
->io_private
;
4076 * The hdr was inserted into hash-table and removed from lists
4077 * prior to starting I/O. We should find this header, since
4078 * it's in the hash table, and it should be legit since it's
4079 * not possible to evict it during the I/O. The only possible
4080 * reason for it not to be found is if we were freed during the
4083 if (HDR_IN_HASH_TABLE(hdr
)) {
4084 arc_buf_hdr_t
*found
;
4086 ASSERT3U(hdr
->b_birth
, ==, BP_PHYSICAL_BIRTH(zio
->io_bp
));
4087 ASSERT3U(hdr
->b_dva
.dva_word
[0], ==,
4088 BP_IDENTITY(zio
->io_bp
)->dva_word
[0]);
4089 ASSERT3U(hdr
->b_dva
.dva_word
[1], ==,
4090 BP_IDENTITY(zio
->io_bp
)->dva_word
[1]);
4092 found
= buf_hash_find(hdr
->b_spa
, zio
->io_bp
,
4095 ASSERT((found
== NULL
&& HDR_FREED_IN_READ(hdr
) &&
4096 hash_lock
== NULL
) ||
4098 DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
4099 (found
== hdr
&& HDR_L2_READING(hdr
)));
4102 hdr
->b_flags
&= ~ARC_FLAG_L2_EVICTED
;
4103 if (l2arc_noprefetch
&& HDR_PREFETCH(hdr
))
4104 hdr
->b_flags
&= ~ARC_FLAG_L2CACHE
;
4106 /* byteswap if necessary */
4107 callback_list
= hdr
->b_l1hdr
.b_acb
;
4108 ASSERT(callback_list
!= NULL
);
4109 if (BP_SHOULD_BYTESWAP(zio
->io_bp
) && zio
->io_error
== 0) {
4110 dmu_object_byteswap_t bswap
=
4111 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
4112 if (BP_GET_LEVEL(zio
->io_bp
) > 0)
4113 byteswap_uint64_array(buf
->b_data
, hdr
->b_size
);
4115 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, hdr
->b_size
);
4118 arc_cksum_compute(buf
, B_FALSE
);
4121 if (hash_lock
&& zio
->io_error
== 0 &&
4122 hdr
->b_l1hdr
.b_state
== arc_anon
) {
4124 * Only call arc_access on anonymous buffers. This is because
4125 * if we've issued an I/O for an evicted buffer, we've already
4126 * called arc_access (to prevent any simultaneous readers from
4127 * getting confused).
4129 arc_access(hdr
, hash_lock
);
4132 /* create copies of the data buffer for the callers */
4134 for (acb
= callback_list
; acb
; acb
= acb
->acb_next
) {
4135 if (acb
->acb_done
) {
4137 ARCSTAT_BUMP(arcstat_duplicate_reads
);
4138 abuf
= arc_buf_clone(buf
);
4140 acb
->acb_buf
= abuf
;
4144 hdr
->b_l1hdr
.b_acb
= NULL
;
4145 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
4146 ASSERT(!HDR_BUF_AVAILABLE(hdr
));
4148 ASSERT(buf
->b_efunc
== NULL
);
4149 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
4150 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
4153 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
) ||
4154 callback_list
!= NULL
);
4156 if (zio
->io_error
!= 0) {
4157 hdr
->b_flags
|= ARC_FLAG_IO_ERROR
;
4158 if (hdr
->b_l1hdr
.b_state
!= arc_anon
)
4159 arc_change_state(arc_anon
, hdr
, hash_lock
);
4160 if (HDR_IN_HASH_TABLE(hdr
))
4161 buf_hash_remove(hdr
);
4162 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
4166 * Broadcast before we drop the hash_lock to avoid the possibility
4167 * that the hdr (and hence the cv) might be freed before we get to
4168 * the cv_broadcast().
4170 cv_broadcast(&hdr
->b_l1hdr
.b_cv
);
4172 if (hash_lock
!= NULL
) {
4173 mutex_exit(hash_lock
);
4176 * This block was freed while we waited for the read to
4177 * complete. It has been removed from the hash table and
4178 * moved to the anonymous state (so that it won't show up
4181 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
4182 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
4185 /* execute each callback and free its structure */
4186 while ((acb
= callback_list
) != NULL
) {
4188 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
4190 if (acb
->acb_zio_dummy
!= NULL
) {
4191 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
4192 zio_nowait(acb
->acb_zio_dummy
);
4195 callback_list
= acb
->acb_next
;
4196 kmem_free(acb
, sizeof (arc_callback_t
));
4200 arc_hdr_destroy(hdr
);
4204 * "Read" the block at the specified DVA (in bp) via the
4205 * cache. If the block is found in the cache, invoke the provided
4206 * callback immediately and return. Note that the `zio' parameter
4207 * in the callback will be NULL in this case, since no IO was
4208 * required. If the block is not in the cache pass the read request
4209 * on to the spa with a substitute callback function, so that the
4210 * requested block will be added to the cache.
4212 * If a read request arrives for a block that has a read in-progress,
4213 * either wait for the in-progress read to complete (and return the
4214 * results); or, if this is a read with a "done" func, add a record
4215 * to the read to invoke the "done" func when the read completes,
4216 * and return; or just return.
4218 * arc_read_done() will invoke all the requested "done" functions
4219 * for readers of this block.
4222 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_done_func_t
*done
,
4223 void *private, zio_priority_t priority
, int zio_flags
,
4224 arc_flags_t
*arc_flags
, const zbookmark_phys_t
*zb
)
4226 arc_buf_hdr_t
*hdr
= NULL
;
4227 arc_buf_t
*buf
= NULL
;
4228 kmutex_t
*hash_lock
= NULL
;
4230 uint64_t guid
= spa_load_guid(spa
);
4233 ASSERT(!BP_IS_EMBEDDED(bp
) ||
4234 BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
);
4237 if (!BP_IS_EMBEDDED(bp
)) {
4239 * Embedded BP's have no DVA and require no I/O to "read".
4240 * Create an anonymous arc buf to back it.
4242 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
4245 if (hdr
!= NULL
&& HDR_HAS_L1HDR(hdr
) && hdr
->b_l1hdr
.b_datacnt
> 0) {
4247 *arc_flags
|= ARC_FLAG_CACHED
;
4249 if (HDR_IO_IN_PROGRESS(hdr
)) {
4251 if ((hdr
->b_flags
& ARC_FLAG_PRIO_ASYNC_READ
) &&
4252 priority
== ZIO_PRIORITY_SYNC_READ
) {
4254 * This sync read must wait for an
4255 * in-progress async read (e.g. a predictive
4256 * prefetch). Async reads are queued
4257 * separately at the vdev_queue layer, so
4258 * this is a form of priority inversion.
4259 * Ideally, we would "inherit" the demand
4260 * i/o's priority by moving the i/o from
4261 * the async queue to the synchronous queue,
4262 * but there is currently no mechanism to do
4263 * so. Track this so that we can evaluate
4264 * the magnitude of this potential performance
4267 * Note that if the prefetch i/o is already
4268 * active (has been issued to the device),
4269 * the prefetch improved performance, because
4270 * we issued it sooner than we would have
4271 * without the prefetch.
4273 DTRACE_PROBE1(arc__sync__wait__for__async
,
4274 arc_buf_hdr_t
*, hdr
);
4275 ARCSTAT_BUMP(arcstat_sync_wait_for_async
);
4277 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
4278 hdr
->b_flags
&= ~ARC_FLAG_PREDICTIVE_PREFETCH
;
4281 if (*arc_flags
& ARC_FLAG_WAIT
) {
4282 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
4283 mutex_exit(hash_lock
);
4286 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
4289 arc_callback_t
*acb
= NULL
;
4291 acb
= kmem_zalloc(sizeof (arc_callback_t
),
4293 acb
->acb_done
= done
;
4294 acb
->acb_private
= private;
4296 acb
->acb_zio_dummy
= zio_null(pio
,
4297 spa
, NULL
, NULL
, NULL
, zio_flags
);
4299 ASSERT(acb
->acb_done
!= NULL
);
4300 acb
->acb_next
= hdr
->b_l1hdr
.b_acb
;
4301 hdr
->b_l1hdr
.b_acb
= acb
;
4302 add_reference(hdr
, hash_lock
, private);
4303 mutex_exit(hash_lock
);
4306 mutex_exit(hash_lock
);
4310 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
4311 hdr
->b_l1hdr
.b_state
== arc_mfu
);
4314 if (hdr
->b_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
) {
4316 * This is a demand read which does not have to
4317 * wait for i/o because we did a predictive
4318 * prefetch i/o for it, which has completed.
4321 arc__demand__hit__predictive__prefetch
,
4322 arc_buf_hdr_t
*, hdr
);
4324 arcstat_demand_hit_predictive_prefetch
);
4325 hdr
->b_flags
&= ~ARC_FLAG_PREDICTIVE_PREFETCH
;
4327 add_reference(hdr
, hash_lock
, private);
4329 * If this block is already in use, create a new
4330 * copy of the data so that we will be guaranteed
4331 * that arc_release() will always succeed.
4333 buf
= hdr
->b_l1hdr
.b_buf
;
4335 ASSERT(buf
->b_data
);
4336 if (HDR_BUF_AVAILABLE(hdr
)) {
4337 ASSERT(buf
->b_efunc
== NULL
);
4338 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
4340 buf
= arc_buf_clone(buf
);
4343 } else if (*arc_flags
& ARC_FLAG_PREFETCH
&&
4344 refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
4345 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
4347 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
4348 arc_access(hdr
, hash_lock
);
4349 if (*arc_flags
& ARC_FLAG_L2CACHE
)
4350 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
4351 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
4352 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
4353 mutex_exit(hash_lock
);
4354 ARCSTAT_BUMP(arcstat_hits
);
4355 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
4356 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
4357 data
, metadata
, hits
);
4360 done(NULL
, buf
, private);
4362 uint64_t size
= BP_GET_LSIZE(bp
);
4363 arc_callback_t
*acb
;
4366 boolean_t devw
= B_FALSE
;
4367 enum zio_compress b_compress
= ZIO_COMPRESS_OFF
;
4368 int32_t b_asize
= 0;
4371 * Gracefully handle a damaged logical block size as a
4374 if (size
> spa_maxblocksize(spa
)) {
4375 ASSERT3P(buf
, ==, NULL
);
4376 rc
= SET_ERROR(ECKSUM
);
4381 /* this block is not in the cache */
4382 arc_buf_hdr_t
*exists
= NULL
;
4383 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
4384 buf
= arc_buf_alloc(spa
, size
, private, type
);
4386 if (!BP_IS_EMBEDDED(bp
)) {
4387 hdr
->b_dva
= *BP_IDENTITY(bp
);
4388 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
4389 exists
= buf_hash_insert(hdr
, &hash_lock
);
4391 if (exists
!= NULL
) {
4392 /* somebody beat us to the hash insert */
4393 mutex_exit(hash_lock
);
4394 buf_discard_identity(hdr
);
4395 (void) arc_buf_remove_ref(buf
, private);
4396 goto top
; /* restart the IO request */
4400 * If there is a callback, we pass our reference to
4401 * it; otherwise we remove our reference.
4404 (void) remove_reference(hdr
, hash_lock
,
4407 if (*arc_flags
& ARC_FLAG_PREFETCH
)
4408 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
4409 if (*arc_flags
& ARC_FLAG_L2CACHE
)
4410 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
4411 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
4412 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
4413 if (BP_GET_LEVEL(bp
) > 0)
4414 hdr
->b_flags
|= ARC_FLAG_INDIRECT
;
4417 * This block is in the ghost cache. If it was L2-only
4418 * (and thus didn't have an L1 hdr), we realloc the
4419 * header to add an L1 hdr.
4421 if (!HDR_HAS_L1HDR(hdr
)) {
4422 hdr
= arc_hdr_realloc(hdr
, hdr_l2only_cache
,
4426 ASSERT(GHOST_STATE(hdr
->b_l1hdr
.b_state
));
4427 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4428 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4429 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
4432 * If there is a callback, we pass a reference to it.
4435 add_reference(hdr
, hash_lock
, private);
4436 if (*arc_flags
& ARC_FLAG_PREFETCH
)
4437 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
4438 if (*arc_flags
& ARC_FLAG_L2CACHE
)
4439 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
4440 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
4441 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
4442 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
4445 buf
->b_efunc
= NULL
;
4446 buf
->b_private
= NULL
;
4448 hdr
->b_l1hdr
.b_buf
= buf
;
4449 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
4450 hdr
->b_l1hdr
.b_datacnt
= 1;
4451 arc_get_data_buf(buf
);
4452 arc_access(hdr
, hash_lock
);
4455 if (*arc_flags
& ARC_FLAG_PREDICTIVE_PREFETCH
)
4456 hdr
->b_flags
|= ARC_FLAG_PREDICTIVE_PREFETCH
;
4457 ASSERT(!GHOST_STATE(hdr
->b_l1hdr
.b_state
));
4459 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
4460 acb
->acb_done
= done
;
4461 acb
->acb_private
= private;
4463 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
4464 hdr
->b_l1hdr
.b_acb
= acb
;
4465 hdr
->b_flags
|= ARC_FLAG_IO_IN_PROGRESS
;
4467 if (HDR_HAS_L2HDR(hdr
) &&
4468 (vd
= hdr
->b_l2hdr
.b_dev
->l2ad_vdev
) != NULL
) {
4469 devw
= hdr
->b_l2hdr
.b_dev
->l2ad_writing
;
4470 addr
= hdr
->b_l2hdr
.b_daddr
;
4471 b_compress
= hdr
->b_l2hdr
.b_compress
;
4472 b_asize
= hdr
->b_l2hdr
.b_asize
;
4474 * Lock out device removal.
4476 if (vdev_is_dead(vd
) ||
4477 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
4481 if (hash_lock
!= NULL
)
4482 mutex_exit(hash_lock
);
4485 * At this point, we have a level 1 cache miss. Try again in
4486 * L2ARC if possible.
4488 ASSERT3U(hdr
->b_size
, ==, size
);
4489 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
4490 uint64_t, size
, zbookmark_phys_t
*, zb
);
4491 ARCSTAT_BUMP(arcstat_misses
);
4492 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
4493 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
4494 data
, metadata
, misses
);
4496 if (priority
== ZIO_PRIORITY_ASYNC_READ
)
4497 hdr
->b_flags
|= ARC_FLAG_PRIO_ASYNC_READ
;
4499 hdr
->b_flags
&= ~ARC_FLAG_PRIO_ASYNC_READ
;
4501 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
4503 * Read from the L2ARC if the following are true:
4504 * 1. The L2ARC vdev was previously cached.
4505 * 2. This buffer still has L2ARC metadata.
4506 * 3. This buffer isn't currently writing to the L2ARC.
4507 * 4. The L2ARC entry wasn't evicted, which may
4508 * also have invalidated the vdev.
4509 * 5. This isn't prefetch and l2arc_noprefetch is set.
4511 if (HDR_HAS_L2HDR(hdr
) &&
4512 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
4513 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
4514 l2arc_read_callback_t
*cb
;
4516 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
4517 ARCSTAT_BUMP(arcstat_l2_hits
);
4518 atomic_inc_32(&hdr
->b_l2hdr
.b_hits
);
4520 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
4522 cb
->l2rcb_buf
= buf
;
4523 cb
->l2rcb_spa
= spa
;
4526 cb
->l2rcb_flags
= zio_flags
;
4527 cb
->l2rcb_compress
= b_compress
;
4529 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
4530 addr
+ size
< vd
->vdev_psize
-
4531 VDEV_LABEL_END_SIZE
);
4534 * l2arc read. The SCL_L2ARC lock will be
4535 * released by l2arc_read_done().
4536 * Issue a null zio if the underlying buffer
4537 * was squashed to zero size by compression.
4539 if (b_compress
== ZIO_COMPRESS_EMPTY
) {
4540 rzio
= zio_null(pio
, spa
, vd
,
4541 l2arc_read_done
, cb
,
4542 zio_flags
| ZIO_FLAG_DONT_CACHE
|
4544 ZIO_FLAG_DONT_PROPAGATE
|
4545 ZIO_FLAG_DONT_RETRY
);
4547 rzio
= zio_read_phys(pio
, vd
, addr
,
4548 b_asize
, buf
->b_data
,
4550 l2arc_read_done
, cb
, priority
,
4551 zio_flags
| ZIO_FLAG_DONT_CACHE
|
4553 ZIO_FLAG_DONT_PROPAGATE
|
4554 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
4556 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
4558 ARCSTAT_INCR(arcstat_l2_read_bytes
, b_asize
);
4560 if (*arc_flags
& ARC_FLAG_NOWAIT
) {
4565 ASSERT(*arc_flags
& ARC_FLAG_WAIT
);
4566 if (zio_wait(rzio
) == 0)
4569 /* l2arc read error; goto zio_read() */
4571 DTRACE_PROBE1(l2arc__miss
,
4572 arc_buf_hdr_t
*, hdr
);
4573 ARCSTAT_BUMP(arcstat_l2_misses
);
4574 if (HDR_L2_WRITING(hdr
))
4575 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
4576 spa_config_exit(spa
, SCL_L2ARC
, vd
);
4580 spa_config_exit(spa
, SCL_L2ARC
, vd
);
4581 if (l2arc_ndev
!= 0) {
4582 DTRACE_PROBE1(l2arc__miss
,
4583 arc_buf_hdr_t
*, hdr
);
4584 ARCSTAT_BUMP(arcstat_l2_misses
);
4588 rzio
= zio_read(pio
, spa
, bp
, buf
->b_data
, size
,
4589 arc_read_done
, buf
, priority
, zio_flags
, zb
);
4591 if (*arc_flags
& ARC_FLAG_WAIT
) {
4592 rc
= zio_wait(rzio
);
4596 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
4601 spa_read_history_add(spa
, zb
, *arc_flags
);
4606 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
4610 p
= kmem_alloc(sizeof (*p
), KM_SLEEP
);
4612 p
->p_private
= private;
4613 list_link_init(&p
->p_node
);
4614 refcount_create(&p
->p_refcnt
);
4616 mutex_enter(&arc_prune_mtx
);
4617 refcount_add(&p
->p_refcnt
, &arc_prune_list
);
4618 list_insert_head(&arc_prune_list
, p
);
4619 mutex_exit(&arc_prune_mtx
);
4625 arc_remove_prune_callback(arc_prune_t
*p
)
4627 mutex_enter(&arc_prune_mtx
);
4628 list_remove(&arc_prune_list
, p
);
4629 if (refcount_remove(&p
->p_refcnt
, &arc_prune_list
) == 0) {
4630 refcount_destroy(&p
->p_refcnt
);
4631 kmem_free(p
, sizeof (*p
));
4633 mutex_exit(&arc_prune_mtx
);
4637 arc_set_callback(arc_buf_t
*buf
, arc_evict_func_t
*func
, void *private)
4639 ASSERT(buf
->b_hdr
!= NULL
);
4640 ASSERT(buf
->b_hdr
->b_l1hdr
.b_state
!= arc_anon
);
4641 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
) ||
4643 ASSERT(buf
->b_efunc
== NULL
);
4644 ASSERT(!HDR_BUF_AVAILABLE(buf
->b_hdr
));
4646 buf
->b_efunc
= func
;
4647 buf
->b_private
= private;
4651 * Notify the arc that a block was freed, and thus will never be used again.
4654 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
4657 kmutex_t
*hash_lock
;
4658 uint64_t guid
= spa_load_guid(spa
);
4660 ASSERT(!BP_IS_EMBEDDED(bp
));
4662 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
4665 if (HDR_BUF_AVAILABLE(hdr
)) {
4666 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
4667 add_reference(hdr
, hash_lock
, FTAG
);
4668 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
4669 mutex_exit(hash_lock
);
4671 arc_release(buf
, FTAG
);
4672 (void) arc_buf_remove_ref(buf
, FTAG
);
4674 mutex_exit(hash_lock
);
4680 * Clear the user eviction callback set by arc_set_callback(), first calling
4681 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4682 * clearing the callback may result in the arc_buf being destroyed. However,
4683 * it will not result in the *last* arc_buf being destroyed, hence the data
4684 * will remain cached in the ARC. We make a copy of the arc buffer here so
4685 * that we can process the callback without holding any locks.
4687 * It's possible that the callback is already in the process of being cleared
4688 * by another thread. In this case we can not clear the callback.
4690 * Returns B_TRUE if the callback was successfully called and cleared.
4693 arc_clear_callback(arc_buf_t
*buf
)
4696 kmutex_t
*hash_lock
;
4697 arc_evict_func_t
*efunc
= buf
->b_efunc
;
4698 void *private = buf
->b_private
;
4700 mutex_enter(&buf
->b_evict_lock
);
4704 * We are in arc_do_user_evicts().
4706 ASSERT(buf
->b_data
== NULL
);
4707 mutex_exit(&buf
->b_evict_lock
);
4709 } else if (buf
->b_data
== NULL
) {
4711 * We are on the eviction list; process this buffer now
4712 * but let arc_do_user_evicts() do the reaping.
4714 buf
->b_efunc
= NULL
;
4715 mutex_exit(&buf
->b_evict_lock
);
4716 VERIFY0(efunc(private));
4719 hash_lock
= HDR_LOCK(hdr
);
4720 mutex_enter(hash_lock
);
4722 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
4724 ASSERT3U(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), <,
4725 hdr
->b_l1hdr
.b_datacnt
);
4726 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
4727 hdr
->b_l1hdr
.b_state
== arc_mfu
);
4729 buf
->b_efunc
= NULL
;
4730 buf
->b_private
= NULL
;
4732 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
4733 mutex_exit(&buf
->b_evict_lock
);
4734 arc_buf_destroy(buf
, TRUE
);
4736 ASSERT(buf
== hdr
->b_l1hdr
.b_buf
);
4737 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
4738 mutex_exit(&buf
->b_evict_lock
);
4741 mutex_exit(hash_lock
);
4742 VERIFY0(efunc(private));
4747 * Release this buffer from the cache, making it an anonymous buffer. This
4748 * must be done after a read and prior to modifying the buffer contents.
4749 * If the buffer has more than one reference, we must make
4750 * a new hdr for the buffer.
4753 arc_release(arc_buf_t
*buf
, void *tag
)
4755 kmutex_t
*hash_lock
;
4757 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4760 * It would be nice to assert that if its DMU metadata (level >
4761 * 0 || it's the dnode file), then it must be syncing context.
4762 * But we don't know that information at this level.
4765 mutex_enter(&buf
->b_evict_lock
);
4767 ASSERT(HDR_HAS_L1HDR(hdr
));
4770 * We don't grab the hash lock prior to this check, because if
4771 * the buffer's header is in the arc_anon state, it won't be
4772 * linked into the hash table.
4774 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
4775 mutex_exit(&buf
->b_evict_lock
);
4776 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4777 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
4778 ASSERT(!HDR_HAS_L2HDR(hdr
));
4779 ASSERT(BUF_EMPTY(hdr
));
4781 ASSERT3U(hdr
->b_l1hdr
.b_datacnt
, ==, 1);
4782 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), ==, 1);
4783 ASSERT(!list_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4785 ASSERT3P(buf
->b_efunc
, ==, NULL
);
4786 ASSERT3P(buf
->b_private
, ==, NULL
);
4788 hdr
->b_l1hdr
.b_arc_access
= 0;
4794 hash_lock
= HDR_LOCK(hdr
);
4795 mutex_enter(hash_lock
);
4798 * This assignment is only valid as long as the hash_lock is
4799 * held, we must be careful not to reference state or the
4800 * b_state field after dropping the lock.
4802 state
= hdr
->b_l1hdr
.b_state
;
4803 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
4804 ASSERT3P(state
, !=, arc_anon
);
4806 /* this buffer is not on any list */
4807 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0);
4809 if (HDR_HAS_L2HDR(hdr
)) {
4810 mutex_enter(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
4813 * We have to recheck this conditional again now that
4814 * we're holding the l2ad_mtx to prevent a race with
4815 * another thread which might be concurrently calling
4816 * l2arc_evict(). In that case, l2arc_evict() might have
4817 * destroyed the header's L2 portion as we were waiting
4818 * to acquire the l2ad_mtx.
4820 if (HDR_HAS_L2HDR(hdr
))
4821 arc_hdr_l2hdr_destroy(hdr
);
4823 mutex_exit(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
4827 * Do we have more than one buf?
4829 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
4830 arc_buf_hdr_t
*nhdr
;
4832 uint64_t blksz
= hdr
->b_size
;
4833 uint64_t spa
= hdr
->b_spa
;
4834 arc_buf_contents_t type
= arc_buf_type(hdr
);
4835 uint32_t flags
= hdr
->b_flags
;
4837 ASSERT(hdr
->b_l1hdr
.b_buf
!= buf
|| buf
->b_next
!= NULL
);
4839 * Pull the data off of this hdr and attach it to
4840 * a new anonymous hdr.
4842 (void) remove_reference(hdr
, hash_lock
, tag
);
4843 bufp
= &hdr
->b_l1hdr
.b_buf
;
4844 while (*bufp
!= buf
)
4845 bufp
= &(*bufp
)->b_next
;
4846 *bufp
= buf
->b_next
;
4849 ASSERT3P(state
, !=, arc_l2c_only
);
4851 (void) refcount_remove_many(
4852 &state
->arcs_size
, hdr
->b_size
, buf
);
4854 if (refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
)) {
4857 ASSERT3P(state
, !=, arc_l2c_only
);
4858 size
= &state
->arcs_lsize
[type
];
4859 ASSERT3U(*size
, >=, hdr
->b_size
);
4860 atomic_add_64(size
, -hdr
->b_size
);
4864 * We're releasing a duplicate user data buffer, update
4865 * our statistics accordingly.
4867 if (HDR_ISTYPE_DATA(hdr
)) {
4868 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
4869 ARCSTAT_INCR(arcstat_duplicate_buffers_size
,
4872 hdr
->b_l1hdr
.b_datacnt
-= 1;
4873 arc_cksum_verify(buf
);
4874 arc_buf_unwatch(buf
);
4876 mutex_exit(hash_lock
);
4878 nhdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
4879 nhdr
->b_size
= blksz
;
4882 nhdr
->b_l1hdr
.b_mru_hits
= 0;
4883 nhdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
4884 nhdr
->b_l1hdr
.b_mfu_hits
= 0;
4885 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
4886 nhdr
->b_l1hdr
.b_l2_hits
= 0;
4887 nhdr
->b_flags
= flags
& ARC_FLAG_L2_WRITING
;
4888 nhdr
->b_flags
|= arc_bufc_to_flags(type
);
4889 nhdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
4891 nhdr
->b_l1hdr
.b_buf
= buf
;
4892 nhdr
->b_l1hdr
.b_datacnt
= 1;
4893 nhdr
->b_l1hdr
.b_state
= arc_anon
;
4894 nhdr
->b_l1hdr
.b_arc_access
= 0;
4895 nhdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
4896 nhdr
->b_freeze_cksum
= NULL
;
4898 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, tag
);
4900 mutex_exit(&buf
->b_evict_lock
);
4901 (void) refcount_add_many(&arc_anon
->arcs_size
, blksz
, buf
);
4903 mutex_exit(&buf
->b_evict_lock
);
4904 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 1);
4905 /* protected by hash lock, or hdr is on arc_anon */
4906 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4907 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4908 hdr
->b_l1hdr
.b_mru_hits
= 0;
4909 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
4910 hdr
->b_l1hdr
.b_mfu_hits
= 0;
4911 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
4912 hdr
->b_l1hdr
.b_l2_hits
= 0;
4913 arc_change_state(arc_anon
, hdr
, hash_lock
);
4914 hdr
->b_l1hdr
.b_arc_access
= 0;
4915 mutex_exit(hash_lock
);
4917 buf_discard_identity(hdr
);
4920 buf
->b_efunc
= NULL
;
4921 buf
->b_private
= NULL
;
4925 arc_released(arc_buf_t
*buf
)
4929 mutex_enter(&buf
->b_evict_lock
);
4930 released
= (buf
->b_data
!= NULL
&&
4931 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
4932 mutex_exit(&buf
->b_evict_lock
);
4938 arc_referenced(arc_buf_t
*buf
)
4942 mutex_enter(&buf
->b_evict_lock
);
4943 referenced
= (refcount_count(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
4944 mutex_exit(&buf
->b_evict_lock
);
4945 return (referenced
);
4950 arc_write_ready(zio_t
*zio
)
4952 arc_write_callback_t
*callback
= zio
->io_private
;
4953 arc_buf_t
*buf
= callback
->awcb_buf
;
4954 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4956 ASSERT(HDR_HAS_L1HDR(hdr
));
4957 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
4958 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
4959 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
4962 * If the IO is already in progress, then this is a re-write
4963 * attempt, so we need to thaw and re-compute the cksum.
4964 * It is the responsibility of the callback to handle the
4965 * accounting for any re-write attempt.
4967 if (HDR_IO_IN_PROGRESS(hdr
)) {
4968 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
4969 if (hdr
->b_freeze_cksum
!= NULL
) {
4970 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
4971 hdr
->b_freeze_cksum
= NULL
;
4973 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
4975 arc_cksum_compute(buf
, B_FALSE
);
4976 hdr
->b_flags
|= ARC_FLAG_IO_IN_PROGRESS
;
4980 * The SPA calls this callback for each physical write that happens on behalf
4981 * of a logical write. See the comment in dbuf_write_physdone() for details.
4984 arc_write_physdone(zio_t
*zio
)
4986 arc_write_callback_t
*cb
= zio
->io_private
;
4987 if (cb
->awcb_physdone
!= NULL
)
4988 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
4992 arc_write_done(zio_t
*zio
)
4994 arc_write_callback_t
*callback
= zio
->io_private
;
4995 arc_buf_t
*buf
= callback
->awcb_buf
;
4996 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4998 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
5000 if (zio
->io_error
== 0) {
5001 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
5002 buf_discard_identity(hdr
);
5004 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
5005 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
5008 ASSERT(BUF_EMPTY(hdr
));
5012 * If the block to be written was all-zero or compressed enough to be
5013 * embedded in the BP, no write was performed so there will be no
5014 * dva/birth/checksum. The buffer must therefore remain anonymous
5017 if (!BUF_EMPTY(hdr
)) {
5018 arc_buf_hdr_t
*exists
;
5019 kmutex_t
*hash_lock
;
5021 ASSERT(zio
->io_error
== 0);
5023 arc_cksum_verify(buf
);
5025 exists
= buf_hash_insert(hdr
, &hash_lock
);
5026 if (exists
!= NULL
) {
5028 * This can only happen if we overwrite for
5029 * sync-to-convergence, because we remove
5030 * buffers from the hash table when we arc_free().
5032 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
5033 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
5034 panic("bad overwrite, hdr=%p exists=%p",
5035 (void *)hdr
, (void *)exists
);
5036 ASSERT(refcount_is_zero(
5037 &exists
->b_l1hdr
.b_refcnt
));
5038 arc_change_state(arc_anon
, exists
, hash_lock
);
5039 mutex_exit(hash_lock
);
5040 arc_hdr_destroy(exists
);
5041 exists
= buf_hash_insert(hdr
, &hash_lock
);
5042 ASSERT3P(exists
, ==, NULL
);
5043 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
5045 ASSERT(zio
->io_prop
.zp_nopwrite
);
5046 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
5047 panic("bad nopwrite, hdr=%p exists=%p",
5048 (void *)hdr
, (void *)exists
);
5051 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
5052 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
5053 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
5054 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
5057 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
5058 /* if it's not anon, we are doing a scrub */
5059 if (exists
== NULL
&& hdr
->b_l1hdr
.b_state
== arc_anon
)
5060 arc_access(hdr
, hash_lock
);
5061 mutex_exit(hash_lock
);
5063 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
5066 ASSERT(!refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
5067 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
5069 kmem_free(callback
, sizeof (arc_write_callback_t
));
5073 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
5074 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
, boolean_t l2arc_compress
,
5075 const zio_prop_t
*zp
, arc_done_func_t
*ready
, arc_done_func_t
*physdone
,
5076 arc_done_func_t
*done
, void *private, zio_priority_t priority
,
5077 int zio_flags
, const zbookmark_phys_t
*zb
)
5079 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
5080 arc_write_callback_t
*callback
;
5083 ASSERT(ready
!= NULL
);
5084 ASSERT(done
!= NULL
);
5085 ASSERT(!HDR_IO_ERROR(hdr
));
5086 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
5087 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
5088 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
5090 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
5092 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
5093 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
5094 callback
->awcb_ready
= ready
;
5095 callback
->awcb_physdone
= physdone
;
5096 callback
->awcb_done
= done
;
5097 callback
->awcb_private
= private;
5098 callback
->awcb_buf
= buf
;
5100 zio
= zio_write(pio
, spa
, txg
, bp
, buf
->b_data
, hdr
->b_size
, zp
,
5101 arc_write_ready
, arc_write_physdone
, arc_write_done
, callback
,
5102 priority
, zio_flags
, zb
);
5108 arc_memory_throttle(uint64_t reserve
, uint64_t txg
)
5111 uint64_t available_memory
= ptob(freemem
);
5112 static uint64_t page_load
= 0;
5113 static uint64_t last_txg
= 0;
5115 pgcnt_t minfree
= btop(arc_sys_free
/ 4);
5118 if (freemem
> physmem
* arc_lotsfree_percent
/ 100)
5121 if (txg
> last_txg
) {
5127 * If we are in pageout, we know that memory is already tight,
5128 * the arc is already going to be evicting, so we just want to
5129 * continue to let page writes occur as quickly as possible.
5131 if (current_is_kswapd()) {
5132 if (page_load
> MAX(ptob(minfree
), available_memory
) / 4) {
5133 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
5134 return (SET_ERROR(ERESTART
));
5136 /* Note: reserve is inflated, so we deflate */
5137 page_load
+= reserve
/ 8;
5139 } else if (page_load
> 0 && arc_reclaim_needed()) {
5140 /* memory is low, delay before restarting */
5141 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
5142 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
5143 return (SET_ERROR(EAGAIN
));
5151 arc_tempreserve_clear(uint64_t reserve
)
5153 atomic_add_64(&arc_tempreserve
, -reserve
);
5154 ASSERT((int64_t)arc_tempreserve
>= 0);
5158 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
5163 if (reserve
> arc_c
/4 && !arc_no_grow
)
5164 arc_c
= MIN(arc_c_max
, reserve
* 4);
5167 * Throttle when the calculated memory footprint for the TXG
5168 * exceeds the target ARC size.
5170 if (reserve
> arc_c
) {
5171 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
5172 return (SET_ERROR(ERESTART
));
5176 * Don't count loaned bufs as in flight dirty data to prevent long
5177 * network delays from blocking transactions that are ready to be
5178 * assigned to a txg.
5180 anon_size
= MAX((int64_t)(refcount_count(&arc_anon
->arcs_size
) -
5181 arc_loaned_bytes
), 0);
5184 * Writes will, almost always, require additional memory allocations
5185 * in order to compress/encrypt/etc the data. We therefore need to
5186 * make sure that there is sufficient available memory for this.
5188 error
= arc_memory_throttle(reserve
, txg
);
5193 * Throttle writes when the amount of dirty data in the cache
5194 * gets too large. We try to keep the cache less than half full
5195 * of dirty blocks so that our sync times don't grow too large.
5196 * Note: if two requests come in concurrently, we might let them
5197 * both succeed, when one of them should fail. Not a huge deal.
5200 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
5201 anon_size
> arc_c
/ 4) {
5202 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5203 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5204 arc_tempreserve
>>10,
5205 arc_anon
->arcs_lsize
[ARC_BUFC_METADATA
]>>10,
5206 arc_anon
->arcs_lsize
[ARC_BUFC_DATA
]>>10,
5207 reserve
>>10, arc_c
>>10);
5208 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
5209 return (SET_ERROR(ERESTART
));
5211 atomic_add_64(&arc_tempreserve
, reserve
);
5216 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
5217 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
5219 size
->value
.ui64
= refcount_count(&state
->arcs_size
);
5220 evict_data
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_DATA
];
5221 evict_metadata
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_METADATA
];
5225 arc_kstat_update(kstat_t
*ksp
, int rw
)
5227 arc_stats_t
*as
= ksp
->ks_data
;
5229 if (rw
== KSTAT_WRITE
) {
5232 arc_kstat_update_state(arc_anon
,
5233 &as
->arcstat_anon_size
,
5234 &as
->arcstat_anon_evictable_data
,
5235 &as
->arcstat_anon_evictable_metadata
);
5236 arc_kstat_update_state(arc_mru
,
5237 &as
->arcstat_mru_size
,
5238 &as
->arcstat_mru_evictable_data
,
5239 &as
->arcstat_mru_evictable_metadata
);
5240 arc_kstat_update_state(arc_mru_ghost
,
5241 &as
->arcstat_mru_ghost_size
,
5242 &as
->arcstat_mru_ghost_evictable_data
,
5243 &as
->arcstat_mru_ghost_evictable_metadata
);
5244 arc_kstat_update_state(arc_mfu
,
5245 &as
->arcstat_mfu_size
,
5246 &as
->arcstat_mfu_evictable_data
,
5247 &as
->arcstat_mfu_evictable_metadata
);
5248 arc_kstat_update_state(arc_mfu_ghost
,
5249 &as
->arcstat_mfu_ghost_size
,
5250 &as
->arcstat_mfu_ghost_evictable_data
,
5251 &as
->arcstat_mfu_ghost_evictable_metadata
);
5258 * This function *must* return indices evenly distributed between all
5259 * sublists of the multilist. This is needed due to how the ARC eviction
5260 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5261 * distributed between all sublists and uses this assumption when
5262 * deciding which sublist to evict from and how much to evict from it.
5265 arc_state_multilist_index_func(multilist_t
*ml
, void *obj
)
5267 arc_buf_hdr_t
*hdr
= obj
;
5270 * We rely on b_dva to generate evenly distributed index
5271 * numbers using buf_hash below. So, as an added precaution,
5272 * let's make sure we never add empty buffers to the arc lists.
5274 ASSERT(!BUF_EMPTY(hdr
));
5277 * The assumption here, is the hash value for a given
5278 * arc_buf_hdr_t will remain constant throughout its lifetime
5279 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
5280 * Thus, we don't need to store the header's sublist index
5281 * on insertion, as this index can be recalculated on removal.
5283 * Also, the low order bits of the hash value are thought to be
5284 * distributed evenly. Otherwise, in the case that the multilist
5285 * has a power of two number of sublists, each sublists' usage
5286 * would not be evenly distributed.
5288 return (buf_hash(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
) %
5289 multilist_get_num_sublists(ml
));
5293 * Called during module initialization and periodically thereafter to
5294 * apply reasonable changes to the exposed performance tunings. Non-zero
5295 * zfs_* values which differ from the currently set values will be applied.
5298 arc_tuning_update(void)
5300 /* Valid range: 64M - <all physical memory> */
5301 if ((zfs_arc_max
) && (zfs_arc_max
!= arc_c_max
) &&
5302 (zfs_arc_max
> 64 << 20) && (zfs_arc_max
< ptob(physmem
)) &&
5303 (zfs_arc_max
> arc_c_min
)) {
5304 arc_c_max
= zfs_arc_max
;
5306 arc_p
= (arc_c
>> 1);
5307 arc_meta_limit
= MIN(arc_meta_limit
, (3 * arc_c_max
) / 4);
5310 /* Valid range: 32M - <arc_c_max> */
5311 if ((zfs_arc_min
) && (zfs_arc_min
!= arc_c_min
) &&
5312 (zfs_arc_min
>= 2ULL << SPA_MAXBLOCKSHIFT
) &&
5313 (zfs_arc_min
<= arc_c_max
)) {
5314 arc_c_min
= zfs_arc_min
;
5315 arc_c
= MAX(arc_c
, arc_c_min
);
5318 /* Valid range: 16M - <arc_c_max> */
5319 if ((zfs_arc_meta_min
) && (zfs_arc_meta_min
!= arc_meta_min
) &&
5320 (zfs_arc_meta_min
>= 1ULL << SPA_MAXBLOCKSHIFT
) &&
5321 (zfs_arc_meta_min
<= arc_c_max
)) {
5322 arc_meta_min
= zfs_arc_meta_min
;
5323 arc_meta_limit
= MAX(arc_meta_limit
, arc_meta_min
);
5326 /* Valid range: <arc_meta_min> - <arc_c_max> */
5327 if ((zfs_arc_meta_limit
) && (zfs_arc_meta_limit
!= arc_meta_limit
) &&
5328 (zfs_arc_meta_limit
>= zfs_arc_meta_min
) &&
5329 (zfs_arc_meta_limit
<= arc_c_max
))
5330 arc_meta_limit
= zfs_arc_meta_limit
;
5332 /* Valid range: 1 - N */
5333 if (zfs_arc_grow_retry
)
5334 arc_grow_retry
= zfs_arc_grow_retry
;
5336 /* Valid range: 1 - N */
5337 if (zfs_arc_shrink_shift
) {
5338 arc_shrink_shift
= zfs_arc_shrink_shift
;
5339 arc_no_grow_shift
= MIN(arc_no_grow_shift
, arc_shrink_shift
-1);
5342 /* Valid range: 1 - N */
5343 if (zfs_arc_p_min_shift
)
5344 arc_p_min_shift
= zfs_arc_p_min_shift
;
5346 /* Valid range: 1 - N ticks */
5347 if (zfs_arc_min_prefetch_lifespan
)
5348 arc_min_prefetch_lifespan
= zfs_arc_min_prefetch_lifespan
;
5350 /* Valid range: 0 - 100 */
5351 if ((zfs_arc_lotsfree_percent
>= 0) &&
5352 (zfs_arc_lotsfree_percent
<= 100))
5353 arc_lotsfree_percent
= zfs_arc_lotsfree_percent
;
5355 /* Valid range: 0 - <all physical memory> */
5356 if ((zfs_arc_sys_free
) && (zfs_arc_sys_free
!= arc_sys_free
))
5357 arc_sys_free
= MIN(MAX(zfs_arc_sys_free
, 0), ptob(physmem
));
5365 * allmem is "all memory that we could possibly use".
5368 uint64_t allmem
= ptob(physmem
);
5370 uint64_t allmem
= (physmem
* PAGESIZE
) / 2;
5373 mutex_init(&arc_reclaim_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
5374 cv_init(&arc_reclaim_thread_cv
, NULL
, CV_DEFAULT
, NULL
);
5375 cv_init(&arc_reclaim_waiters_cv
, NULL
, CV_DEFAULT
, NULL
);
5377 mutex_init(&arc_user_evicts_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
5378 cv_init(&arc_user_evicts_cv
, NULL
, CV_DEFAULT
, NULL
);
5380 /* Convert seconds to clock ticks */
5381 arc_min_prefetch_lifespan
= 1 * hz
;
5383 /* Start out with 1/8 of all memory */
5388 * On architectures where the physical memory can be larger
5389 * than the addressable space (intel in 32-bit mode), we may
5390 * need to limit the cache to 1/8 of VM size.
5392 arc_c
= MIN(arc_c
, vmem_size(heap_arena
, VMEM_ALLOC
| VMEM_FREE
) / 8);
5395 * Register a shrinker to support synchronous (direct) memory
5396 * reclaim from the arc. This is done to prevent kswapd from
5397 * swapping out pages when it is preferable to shrink the arc.
5399 spl_register_shrinker(&arc_shrinker
);
5401 /* Set to 1/64 of all memory or a minimum of 512K */
5402 arc_sys_free
= MAX(ptob(physmem
/ 64), (512 * 1024));
5407 * In userland, there's only the memory pressure that we artificially
5408 * create (see arc_available_memory()). Don't let arc_c get too
5409 * small, because it can cause transactions to be larger than
5410 * arc_c, causing arc_tempreserve_space() to fail.
5413 arc_c_min
= arc_c_max
/ 2;
5415 arc_c_min
= 2ULL << SPA_MAXBLOCKSHIFT
;
5418 /* Set max to 1/2 of all memory */
5419 arc_c_max
= allmem
/ 2;
5422 arc_p
= (arc_c
>> 1);
5424 /* Set min to 1/2 of arc_c_min */
5425 arc_meta_min
= 1ULL << SPA_MAXBLOCKSHIFT
;
5426 /* Initialize maximum observed usage to zero */
5428 /* Set limit to 3/4 of arc_c_max with a floor of arc_meta_min */
5429 arc_meta_limit
= MAX((3 * arc_c_max
) / 4, arc_meta_min
);
5431 /* Apply user specified tunings */
5432 arc_tuning_update();
5434 if (zfs_arc_num_sublists_per_state
< 1)
5435 zfs_arc_num_sublists_per_state
= MAX(boot_ncpus
, 1);
5437 /* if kmem_flags are set, lets try to use less memory */
5438 if (kmem_debugging())
5440 if (arc_c
< arc_c_min
)
5443 arc_anon
= &ARC_anon
;
5445 arc_mru_ghost
= &ARC_mru_ghost
;
5447 arc_mfu_ghost
= &ARC_mfu_ghost
;
5448 arc_l2c_only
= &ARC_l2c_only
;
5451 multilist_create(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
],
5452 sizeof (arc_buf_hdr_t
),
5453 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5454 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5455 multilist_create(&arc_mru
->arcs_list
[ARC_BUFC_DATA
],
5456 sizeof (arc_buf_hdr_t
),
5457 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5458 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5459 multilist_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
],
5460 sizeof (arc_buf_hdr_t
),
5461 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5462 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5463 multilist_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
],
5464 sizeof (arc_buf_hdr_t
),
5465 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5466 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5467 multilist_create(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
],
5468 sizeof (arc_buf_hdr_t
),
5469 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5470 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5471 multilist_create(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
],
5472 sizeof (arc_buf_hdr_t
),
5473 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5474 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5475 multilist_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
],
5476 sizeof (arc_buf_hdr_t
),
5477 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5478 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5479 multilist_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
],
5480 sizeof (arc_buf_hdr_t
),
5481 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5482 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5483 multilist_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
],
5484 sizeof (arc_buf_hdr_t
),
5485 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5486 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5487 multilist_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
],
5488 sizeof (arc_buf_hdr_t
),
5489 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
5490 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
5492 arc_anon
->arcs_state
= ARC_STATE_ANON
;
5493 arc_mru
->arcs_state
= ARC_STATE_MRU
;
5494 arc_mru_ghost
->arcs_state
= ARC_STATE_MRU_GHOST
;
5495 arc_mfu
->arcs_state
= ARC_STATE_MFU
;
5496 arc_mfu_ghost
->arcs_state
= ARC_STATE_MFU_GHOST
;
5497 arc_l2c_only
->arcs_state
= ARC_STATE_L2C_ONLY
;
5499 refcount_create(&arc_anon
->arcs_size
);
5500 refcount_create(&arc_mru
->arcs_size
);
5501 refcount_create(&arc_mru_ghost
->arcs_size
);
5502 refcount_create(&arc_mfu
->arcs_size
);
5503 refcount_create(&arc_mfu_ghost
->arcs_size
);
5504 refcount_create(&arc_l2c_only
->arcs_size
);
5508 arc_reclaim_thread_exit
= FALSE
;
5509 arc_user_evicts_thread_exit
= FALSE
;
5510 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
5511 offsetof(arc_prune_t
, p_node
));
5512 arc_eviction_list
= NULL
;
5513 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
5514 bzero(&arc_eviction_hdr
, sizeof (arc_buf_hdr_t
));
5516 arc_prune_taskq
= taskq_create("arc_prune", max_ncpus
, defclsyspri
,
5517 max_ncpus
, INT_MAX
, TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
5519 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
5520 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
5522 if (arc_ksp
!= NULL
) {
5523 arc_ksp
->ks_data
= &arc_stats
;
5524 arc_ksp
->ks_update
= arc_kstat_update
;
5525 kstat_install(arc_ksp
);
5528 (void) thread_create(NULL
, 0, arc_reclaim_thread
, NULL
, 0, &p0
,
5529 TS_RUN
, defclsyspri
);
5531 (void) thread_create(NULL
, 0, arc_user_evicts_thread
, NULL
, 0, &p0
,
5532 TS_RUN
, defclsyspri
);
5538 * Calculate maximum amount of dirty data per pool.
5540 * If it has been set by a module parameter, take that.
5541 * Otherwise, use a percentage of physical memory defined by
5542 * zfs_dirty_data_max_percent (default 10%) with a cap at
5543 * zfs_dirty_data_max_max (default 25% of physical memory).
5545 if (zfs_dirty_data_max_max
== 0)
5546 zfs_dirty_data_max_max
= (uint64_t)physmem
* PAGESIZE
*
5547 zfs_dirty_data_max_max_percent
/ 100;
5549 if (zfs_dirty_data_max
== 0) {
5550 zfs_dirty_data_max
= (uint64_t)physmem
* PAGESIZE
*
5551 zfs_dirty_data_max_percent
/ 100;
5552 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
5553 zfs_dirty_data_max_max
);
5563 spl_unregister_shrinker(&arc_shrinker
);
5564 #endif /* _KERNEL */
5566 mutex_enter(&arc_reclaim_lock
);
5567 arc_reclaim_thread_exit
= TRUE
;
5569 * The reclaim thread will set arc_reclaim_thread_exit back to
5570 * FALSE when it is finished exiting; we're waiting for that.
5572 while (arc_reclaim_thread_exit
) {
5573 cv_signal(&arc_reclaim_thread_cv
);
5574 cv_wait(&arc_reclaim_thread_cv
, &arc_reclaim_lock
);
5576 mutex_exit(&arc_reclaim_lock
);
5578 mutex_enter(&arc_user_evicts_lock
);
5579 arc_user_evicts_thread_exit
= TRUE
;
5581 * The user evicts thread will set arc_user_evicts_thread_exit
5582 * to FALSE when it is finished exiting; we're waiting for that.
5584 while (arc_user_evicts_thread_exit
) {
5585 cv_signal(&arc_user_evicts_cv
);
5586 cv_wait(&arc_user_evicts_cv
, &arc_user_evicts_lock
);
5588 mutex_exit(&arc_user_evicts_lock
);
5590 /* Use TRUE to ensure *all* buffers are evicted */
5591 arc_flush(NULL
, TRUE
);
5595 if (arc_ksp
!= NULL
) {
5596 kstat_delete(arc_ksp
);
5600 taskq_wait(arc_prune_taskq
);
5601 taskq_destroy(arc_prune_taskq
);
5603 mutex_enter(&arc_prune_mtx
);
5604 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
5605 list_remove(&arc_prune_list
, p
);
5606 refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
5607 refcount_destroy(&p
->p_refcnt
);
5608 kmem_free(p
, sizeof (*p
));
5610 mutex_exit(&arc_prune_mtx
);
5612 list_destroy(&arc_prune_list
);
5613 mutex_destroy(&arc_prune_mtx
);
5614 mutex_destroy(&arc_reclaim_lock
);
5615 cv_destroy(&arc_reclaim_thread_cv
);
5616 cv_destroy(&arc_reclaim_waiters_cv
);
5618 mutex_destroy(&arc_user_evicts_lock
);
5619 cv_destroy(&arc_user_evicts_cv
);
5621 refcount_destroy(&arc_anon
->arcs_size
);
5622 refcount_destroy(&arc_mru
->arcs_size
);
5623 refcount_destroy(&arc_mru_ghost
->arcs_size
);
5624 refcount_destroy(&arc_mfu
->arcs_size
);
5625 refcount_destroy(&arc_mfu_ghost
->arcs_size
);
5626 refcount_destroy(&arc_l2c_only
->arcs_size
);
5628 multilist_destroy(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
5629 multilist_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
5630 multilist_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
5631 multilist_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
5632 multilist_destroy(&arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
5633 multilist_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
5634 multilist_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
5635 multilist_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
5636 multilist_destroy(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]);
5637 multilist_destroy(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]);
5641 ASSERT0(arc_loaned_bytes
);
5647 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5648 * It uses dedicated storage devices to hold cached data, which are populated
5649 * using large infrequent writes. The main role of this cache is to boost
5650 * the performance of random read workloads. The intended L2ARC devices
5651 * include short-stroked disks, solid state disks, and other media with
5652 * substantially faster read latency than disk.
5654 * +-----------------------+
5656 * +-----------------------+
5659 * l2arc_feed_thread() arc_read()
5663 * +---------------+ |
5665 * +---------------+ |
5670 * +-------+ +-------+
5672 * | cache | | cache |
5673 * +-------+ +-------+
5674 * +=========+ .-----.
5675 * : L2ARC : |-_____-|
5676 * : devices : | Disks |
5677 * +=========+ `-_____-'
5679 * Read requests are satisfied from the following sources, in order:
5682 * 2) vdev cache of L2ARC devices
5684 * 4) vdev cache of disks
5687 * Some L2ARC device types exhibit extremely slow write performance.
5688 * To accommodate for this there are some significant differences between
5689 * the L2ARC and traditional cache design:
5691 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5692 * the ARC behave as usual, freeing buffers and placing headers on ghost
5693 * lists. The ARC does not send buffers to the L2ARC during eviction as
5694 * this would add inflated write latencies for all ARC memory pressure.
5696 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5697 * It does this by periodically scanning buffers from the eviction-end of
5698 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5699 * not already there. It scans until a headroom of buffers is satisfied,
5700 * which itself is a buffer for ARC eviction. If a compressible buffer is
5701 * found during scanning and selected for writing to an L2ARC device, we
5702 * temporarily boost scanning headroom during the next scan cycle to make
5703 * sure we adapt to compression effects (which might significantly reduce
5704 * the data volume we write to L2ARC). The thread that does this is
5705 * l2arc_feed_thread(), illustrated below; example sizes are included to
5706 * provide a better sense of ratio than this diagram:
5709 * +---------------------+----------+
5710 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5711 * +---------------------+----------+ | o L2ARC eligible
5712 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5713 * +---------------------+----------+ |
5714 * 15.9 Gbytes ^ 32 Mbytes |
5716 * l2arc_feed_thread()
5718 * l2arc write hand <--[oooo]--'
5722 * +==============================+
5723 * L2ARC dev |####|#|###|###| |####| ... |
5724 * +==============================+
5727 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5728 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5729 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5730 * safe to say that this is an uncommon case, since buffers at the end of
5731 * the ARC lists have moved there due to inactivity.
5733 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5734 * then the L2ARC simply misses copying some buffers. This serves as a
5735 * pressure valve to prevent heavy read workloads from both stalling the ARC
5736 * with waits and clogging the L2ARC with writes. This also helps prevent
5737 * the potential for the L2ARC to churn if it attempts to cache content too
5738 * quickly, such as during backups of the entire pool.
5740 * 5. After system boot and before the ARC has filled main memory, there are
5741 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5742 * lists can remain mostly static. Instead of searching from tail of these
5743 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5744 * for eligible buffers, greatly increasing its chance of finding them.
5746 * The L2ARC device write speed is also boosted during this time so that
5747 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5748 * there are no L2ARC reads, and no fear of degrading read performance
5749 * through increased writes.
5751 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5752 * the vdev queue can aggregate them into larger and fewer writes. Each
5753 * device is written to in a rotor fashion, sweeping writes through
5754 * available space then repeating.
5756 * 7. The L2ARC does not store dirty content. It never needs to flush
5757 * write buffers back to disk based storage.
5759 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5760 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5762 * The performance of the L2ARC can be tweaked by a number of tunables, which
5763 * may be necessary for different workloads:
5765 * l2arc_write_max max write bytes per interval
5766 * l2arc_write_boost extra write bytes during device warmup
5767 * l2arc_noprefetch skip caching prefetched buffers
5768 * l2arc_nocompress skip compressing buffers
5769 * l2arc_headroom number of max device writes to precache
5770 * l2arc_headroom_boost when we find compressed buffers during ARC
5771 * scanning, we multiply headroom by this
5772 * percentage factor for the next scan cycle,
5773 * since more compressed buffers are likely to
5775 * l2arc_feed_secs seconds between L2ARC writing
5777 * Tunables may be removed or added as future performance improvements are
5778 * integrated, and also may become zpool properties.
5780 * There are three key functions that control how the L2ARC warms up:
5782 * l2arc_write_eligible() check if a buffer is eligible to cache
5783 * l2arc_write_size() calculate how much to write
5784 * l2arc_write_interval() calculate sleep delay between writes
5786 * These three functions determine what to write, how much, and how quickly
5791 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*hdr
)
5794 * A buffer is *not* eligible for the L2ARC if it:
5795 * 1. belongs to a different spa.
5796 * 2. is already cached on the L2ARC.
5797 * 3. has an I/O in progress (it may be an incomplete read).
5798 * 4. is flagged not eligible (zfs property).
5800 if (hdr
->b_spa
!= spa_guid
|| HDR_HAS_L2HDR(hdr
) ||
5801 HDR_IO_IN_PROGRESS(hdr
) || !HDR_L2CACHE(hdr
))
5808 l2arc_write_size(void)
5813 * Make sure our globals have meaningful values in case the user
5816 size
= l2arc_write_max
;
5818 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
5819 "be greater than zero, resetting it to the default (%d)",
5821 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
5824 if (arc_warm
== B_FALSE
)
5825 size
+= l2arc_write_boost
;
5832 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
5834 clock_t interval
, next
, now
;
5837 * If the ARC lists are busy, increase our write rate; if the
5838 * lists are stale, idle back. This is achieved by checking
5839 * how much we previously wrote - if it was more than half of
5840 * what we wanted, schedule the next write much sooner.
5842 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
5843 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
5845 interval
= hz
* l2arc_feed_secs
;
5847 now
= ddi_get_lbolt();
5848 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
5854 * Cycle through L2ARC devices. This is how L2ARC load balances.
5855 * If a device is returned, this also returns holding the spa config lock.
5857 static l2arc_dev_t
*
5858 l2arc_dev_get_next(void)
5860 l2arc_dev_t
*first
, *next
= NULL
;
5863 * Lock out the removal of spas (spa_namespace_lock), then removal
5864 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
5865 * both locks will be dropped and a spa config lock held instead.
5867 mutex_enter(&spa_namespace_lock
);
5868 mutex_enter(&l2arc_dev_mtx
);
5870 /* if there are no vdevs, there is nothing to do */
5871 if (l2arc_ndev
== 0)
5875 next
= l2arc_dev_last
;
5877 /* loop around the list looking for a non-faulted vdev */
5879 next
= list_head(l2arc_dev_list
);
5881 next
= list_next(l2arc_dev_list
, next
);
5883 next
= list_head(l2arc_dev_list
);
5886 /* if we have come back to the start, bail out */
5889 else if (next
== first
)
5892 } while (vdev_is_dead(next
->l2ad_vdev
));
5894 /* if we were unable to find any usable vdevs, return NULL */
5895 if (vdev_is_dead(next
->l2ad_vdev
))
5898 l2arc_dev_last
= next
;
5901 mutex_exit(&l2arc_dev_mtx
);
5904 * Grab the config lock to prevent the 'next' device from being
5905 * removed while we are writing to it.
5908 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
5909 mutex_exit(&spa_namespace_lock
);
5915 * Free buffers that were tagged for destruction.
5918 l2arc_do_free_on_write(void)
5921 l2arc_data_free_t
*df
, *df_prev
;
5923 mutex_enter(&l2arc_free_on_write_mtx
);
5924 buflist
= l2arc_free_on_write
;
5926 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
5927 df_prev
= list_prev(buflist
, df
);
5928 ASSERT(df
->l2df_data
!= NULL
);
5929 ASSERT(df
->l2df_func
!= NULL
);
5930 df
->l2df_func(df
->l2df_data
, df
->l2df_size
);
5931 list_remove(buflist
, df
);
5932 kmem_free(df
, sizeof (l2arc_data_free_t
));
5935 mutex_exit(&l2arc_free_on_write_mtx
);
5939 * A write to a cache device has completed. Update all headers to allow
5940 * reads from these buffers to begin.
5943 l2arc_write_done(zio_t
*zio
)
5945 l2arc_write_callback_t
*cb
;
5948 arc_buf_hdr_t
*head
, *hdr
, *hdr_prev
;
5949 kmutex_t
*hash_lock
;
5950 int64_t bytes_dropped
= 0;
5952 cb
= zio
->io_private
;
5954 dev
= cb
->l2wcb_dev
;
5955 ASSERT(dev
!= NULL
);
5956 head
= cb
->l2wcb_head
;
5957 ASSERT(head
!= NULL
);
5958 buflist
= &dev
->l2ad_buflist
;
5959 ASSERT(buflist
!= NULL
);
5960 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
5961 l2arc_write_callback_t
*, cb
);
5963 if (zio
->io_error
!= 0)
5964 ARCSTAT_BUMP(arcstat_l2_writes_error
);
5967 * All writes completed, or an error was hit.
5970 mutex_enter(&dev
->l2ad_mtx
);
5971 for (hdr
= list_prev(buflist
, head
); hdr
; hdr
= hdr_prev
) {
5972 hdr_prev
= list_prev(buflist
, hdr
);
5974 hash_lock
= HDR_LOCK(hdr
);
5977 * We cannot use mutex_enter or else we can deadlock
5978 * with l2arc_write_buffers (due to swapping the order
5979 * the hash lock and l2ad_mtx are taken).
5981 if (!mutex_tryenter(hash_lock
)) {
5983 * Missed the hash lock. We must retry so we
5984 * don't leave the ARC_FLAG_L2_WRITING bit set.
5986 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry
);
5989 * We don't want to rescan the headers we've
5990 * already marked as having been written out, so
5991 * we reinsert the head node so we can pick up
5992 * where we left off.
5994 list_remove(buflist
, head
);
5995 list_insert_after(buflist
, hdr
, head
);
5997 mutex_exit(&dev
->l2ad_mtx
);
6000 * We wait for the hash lock to become available
6001 * to try and prevent busy waiting, and increase
6002 * the chance we'll be able to acquire the lock
6003 * the next time around.
6005 mutex_enter(hash_lock
);
6006 mutex_exit(hash_lock
);
6011 * We could not have been moved into the arc_l2c_only
6012 * state while in-flight due to our ARC_FLAG_L2_WRITING
6013 * bit being set. Let's just ensure that's being enforced.
6015 ASSERT(HDR_HAS_L1HDR(hdr
));
6018 * We may have allocated a buffer for L2ARC compression,
6019 * we must release it to avoid leaking this data.
6021 l2arc_release_cdata_buf(hdr
);
6023 if (zio
->io_error
!= 0) {
6025 * Error - drop L2ARC entry.
6027 list_remove(buflist
, hdr
);
6028 hdr
->b_flags
&= ~ARC_FLAG_HAS_L2HDR
;
6030 ARCSTAT_INCR(arcstat_l2_asize
, -hdr
->b_l2hdr
.b_asize
);
6031 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
6033 bytes_dropped
+= hdr
->b_l2hdr
.b_asize
;
6034 (void) refcount_remove_many(&dev
->l2ad_alloc
,
6035 hdr
->b_l2hdr
.b_asize
, hdr
);
6039 * Allow ARC to begin reads and ghost list evictions to
6042 hdr
->b_flags
&= ~ARC_FLAG_L2_WRITING
;
6044 mutex_exit(hash_lock
);
6047 atomic_inc_64(&l2arc_writes_done
);
6048 list_remove(buflist
, head
);
6049 ASSERT(!HDR_HAS_L1HDR(head
));
6050 kmem_cache_free(hdr_l2only_cache
, head
);
6051 mutex_exit(&dev
->l2ad_mtx
);
6053 vdev_space_update(dev
->l2ad_vdev
, -bytes_dropped
, 0, 0);
6055 l2arc_do_free_on_write();
6057 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
6061 * A read to a cache device completed. Validate buffer contents before
6062 * handing over to the regular ARC routines.
6065 l2arc_read_done(zio_t
*zio
)
6067 l2arc_read_callback_t
*cb
;
6070 kmutex_t
*hash_lock
;
6073 ASSERT(zio
->io_vd
!= NULL
);
6074 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
6076 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
6078 cb
= zio
->io_private
;
6080 buf
= cb
->l2rcb_buf
;
6081 ASSERT(buf
!= NULL
);
6083 hash_lock
= HDR_LOCK(buf
->b_hdr
);
6084 mutex_enter(hash_lock
);
6086 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
6089 * If the buffer was compressed, decompress it first.
6091 if (cb
->l2rcb_compress
!= ZIO_COMPRESS_OFF
)
6092 l2arc_decompress_zio(zio
, hdr
, cb
->l2rcb_compress
);
6093 ASSERT(zio
->io_data
!= NULL
);
6094 ASSERT3U(zio
->io_size
, ==, hdr
->b_size
);
6095 ASSERT3U(BP_GET_LSIZE(&cb
->l2rcb_bp
), ==, hdr
->b_size
);
6098 * Check this survived the L2ARC journey.
6100 equal
= arc_cksum_equal(buf
);
6101 if (equal
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
6102 mutex_exit(hash_lock
);
6103 zio
->io_private
= buf
;
6104 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
6105 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
6108 mutex_exit(hash_lock
);
6110 * Buffer didn't survive caching. Increment stats and
6111 * reissue to the original storage device.
6113 if (zio
->io_error
!= 0) {
6114 ARCSTAT_BUMP(arcstat_l2_io_error
);
6116 zio
->io_error
= SET_ERROR(EIO
);
6119 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
6122 * If there's no waiter, issue an async i/o to the primary
6123 * storage now. If there *is* a waiter, the caller must
6124 * issue the i/o in a context where it's OK to block.
6126 if (zio
->io_waiter
== NULL
) {
6127 zio_t
*pio
= zio_unique_parent(zio
);
6129 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
6131 zio_nowait(zio_read(pio
, cb
->l2rcb_spa
, &cb
->l2rcb_bp
,
6132 buf
->b_data
, hdr
->b_size
, arc_read_done
, buf
,
6133 zio
->io_priority
, cb
->l2rcb_flags
, &cb
->l2rcb_zb
));
6137 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
6141 * This is the list priority from which the L2ARC will search for pages to
6142 * cache. This is used within loops (0..3) to cycle through lists in the
6143 * desired order. This order can have a significant effect on cache
6146 * Currently the metadata lists are hit first, MFU then MRU, followed by
6147 * the data lists. This function returns a locked list, and also returns
6150 static multilist_sublist_t
*
6151 l2arc_sublist_lock(int list_num
)
6153 multilist_t
*ml
= NULL
;
6156 ASSERT(list_num
>= 0 && list_num
<= 3);
6160 ml
= &arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
6163 ml
= &arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
6166 ml
= &arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
6169 ml
= &arc_mru
->arcs_list
[ARC_BUFC_DATA
];
6174 * Return a randomly-selected sublist. This is acceptable
6175 * because the caller feeds only a little bit of data for each
6176 * call (8MB). Subsequent calls will result in different
6177 * sublists being selected.
6179 idx
= multilist_get_random_index(ml
);
6180 return (multilist_sublist_lock(ml
, idx
));
6184 * Evict buffers from the device write hand to the distance specified in
6185 * bytes. This distance may span populated buffers, it may span nothing.
6186 * This is clearing a region on the L2ARC device ready for writing.
6187 * If the 'all' boolean is set, every buffer is evicted.
6190 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
6193 arc_buf_hdr_t
*hdr
, *hdr_prev
;
6194 kmutex_t
*hash_lock
;
6197 buflist
= &dev
->l2ad_buflist
;
6199 if (!all
&& dev
->l2ad_first
) {
6201 * This is the first sweep through the device. There is
6207 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
6209 * When nearing the end of the device, evict to the end
6210 * before the device write hand jumps to the start.
6212 taddr
= dev
->l2ad_end
;
6214 taddr
= dev
->l2ad_hand
+ distance
;
6216 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
6217 uint64_t, taddr
, boolean_t
, all
);
6220 mutex_enter(&dev
->l2ad_mtx
);
6221 for (hdr
= list_tail(buflist
); hdr
; hdr
= hdr_prev
) {
6222 hdr_prev
= list_prev(buflist
, hdr
);
6224 hash_lock
= HDR_LOCK(hdr
);
6227 * We cannot use mutex_enter or else we can deadlock
6228 * with l2arc_write_buffers (due to swapping the order
6229 * the hash lock and l2ad_mtx are taken).
6231 if (!mutex_tryenter(hash_lock
)) {
6233 * Missed the hash lock. Retry.
6235 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
6236 mutex_exit(&dev
->l2ad_mtx
);
6237 mutex_enter(hash_lock
);
6238 mutex_exit(hash_lock
);
6242 if (HDR_L2_WRITE_HEAD(hdr
)) {
6244 * We hit a write head node. Leave it for
6245 * l2arc_write_done().
6247 list_remove(buflist
, hdr
);
6248 mutex_exit(hash_lock
);
6252 if (!all
&& HDR_HAS_L2HDR(hdr
) &&
6253 (hdr
->b_l2hdr
.b_daddr
> taddr
||
6254 hdr
->b_l2hdr
.b_daddr
< dev
->l2ad_hand
)) {
6256 * We've evicted to the target address,
6257 * or the end of the device.
6259 mutex_exit(hash_lock
);
6263 ASSERT(HDR_HAS_L2HDR(hdr
));
6264 if (!HDR_HAS_L1HDR(hdr
)) {
6265 ASSERT(!HDR_L2_READING(hdr
));
6267 * This doesn't exist in the ARC. Destroy.
6268 * arc_hdr_destroy() will call list_remove()
6269 * and decrement arcstat_l2_size.
6271 arc_change_state(arc_anon
, hdr
, hash_lock
);
6272 arc_hdr_destroy(hdr
);
6274 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_l2c_only
);
6275 ARCSTAT_BUMP(arcstat_l2_evict_l1cached
);
6277 * Invalidate issued or about to be issued
6278 * reads, since we may be about to write
6279 * over this location.
6281 if (HDR_L2_READING(hdr
)) {
6282 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
6283 hdr
->b_flags
|= ARC_FLAG_L2_EVICTED
;
6286 /* Ensure this header has finished being written */
6287 ASSERT(!HDR_L2_WRITING(hdr
));
6288 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
6290 arc_hdr_l2hdr_destroy(hdr
);
6292 mutex_exit(hash_lock
);
6294 mutex_exit(&dev
->l2ad_mtx
);
6298 * Find and write ARC buffers to the L2ARC device.
6300 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6301 * for reading until they have completed writing.
6302 * The headroom_boost is an in-out parameter used to maintain headroom boost
6303 * state between calls to this function.
6305 * Returns the number of bytes actually written (which may be smaller than
6306 * the delta by which the device hand has changed due to alignment).
6309 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
,
6310 boolean_t
*headroom_boost
)
6312 arc_buf_hdr_t
*hdr
, *hdr_prev
, *head
;
6313 uint64_t write_asize
, write_sz
, headroom
, buf_compress_minsz
,
6317 l2arc_write_callback_t
*cb
;
6319 uint64_t guid
= spa_load_guid(spa
);
6321 const boolean_t do_headroom_boost
= *headroom_boost
;
6323 ASSERT(dev
->l2ad_vdev
!= NULL
);
6325 /* Lower the flag now, we might want to raise it again later. */
6326 *headroom_boost
= B_FALSE
;
6329 write_sz
= write_asize
= 0;
6331 head
= kmem_cache_alloc(hdr_l2only_cache
, KM_PUSHPAGE
);
6332 head
->b_flags
|= ARC_FLAG_L2_WRITE_HEAD
;
6333 head
->b_flags
|= ARC_FLAG_HAS_L2HDR
;
6336 * We will want to try to compress buffers that are at least 2x the
6337 * device sector size.
6339 buf_compress_minsz
= 2 << dev
->l2ad_vdev
->vdev_ashift
;
6342 * Copy buffers for L2ARC writing.
6344 for (try = 0; try <= 3; try++) {
6345 multilist_sublist_t
*mls
= l2arc_sublist_lock(try);
6346 uint64_t passed_sz
= 0;
6349 * L2ARC fast warmup.
6351 * Until the ARC is warm and starts to evict, read from the
6352 * head of the ARC lists rather than the tail.
6354 if (arc_warm
== B_FALSE
)
6355 hdr
= multilist_sublist_head(mls
);
6357 hdr
= multilist_sublist_tail(mls
);
6359 headroom
= target_sz
* l2arc_headroom
;
6360 if (do_headroom_boost
)
6361 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
6363 for (; hdr
; hdr
= hdr_prev
) {
6364 kmutex_t
*hash_lock
;
6368 if (arc_warm
== B_FALSE
)
6369 hdr_prev
= multilist_sublist_next(mls
, hdr
);
6371 hdr_prev
= multilist_sublist_prev(mls
, hdr
);
6373 hash_lock
= HDR_LOCK(hdr
);
6374 if (!mutex_tryenter(hash_lock
)) {
6376 * Skip this buffer rather than waiting.
6381 passed_sz
+= hdr
->b_size
;
6382 if (passed_sz
> headroom
) {
6386 mutex_exit(hash_lock
);
6390 if (!l2arc_write_eligible(guid
, hdr
)) {
6391 mutex_exit(hash_lock
);
6396 * Assume that the buffer is not going to be compressed
6397 * and could take more space on disk because of a larger
6400 buf_sz
= hdr
->b_size
;
6401 buf_a_sz
= vdev_psize_to_asize(dev
->l2ad_vdev
, buf_sz
);
6403 if ((write_asize
+ buf_a_sz
) > target_sz
) {
6405 mutex_exit(hash_lock
);
6411 * Insert a dummy header on the buflist so
6412 * l2arc_write_done() can find where the
6413 * write buffers begin without searching.
6415 mutex_enter(&dev
->l2ad_mtx
);
6416 list_insert_head(&dev
->l2ad_buflist
, head
);
6417 mutex_exit(&dev
->l2ad_mtx
);
6420 sizeof (l2arc_write_callback_t
), KM_SLEEP
);
6421 cb
->l2wcb_dev
= dev
;
6422 cb
->l2wcb_head
= head
;
6423 pio
= zio_root(spa
, l2arc_write_done
, cb
,
6428 * Create and add a new L2ARC header.
6430 hdr
->b_l2hdr
.b_dev
= dev
;
6431 hdr
->b_flags
|= ARC_FLAG_L2_WRITING
;
6433 * Temporarily stash the data buffer in b_tmp_cdata.
6434 * The subsequent write step will pick it up from
6435 * there. This is because can't access b_l1hdr.b_buf
6436 * without holding the hash_lock, which we in turn
6437 * can't access without holding the ARC list locks
6438 * (which we want to avoid during compression/writing)
6440 hdr
->b_l2hdr
.b_compress
= ZIO_COMPRESS_OFF
;
6441 hdr
->b_l2hdr
.b_asize
= hdr
->b_size
;
6442 hdr
->b_l2hdr
.b_hits
= 0;
6443 hdr
->b_l1hdr
.b_tmp_cdata
= hdr
->b_l1hdr
.b_buf
->b_data
;
6446 * Explicitly set the b_daddr field to a known
6447 * value which means "invalid address". This
6448 * enables us to differentiate which stage of
6449 * l2arc_write_buffers() the particular header
6450 * is in (e.g. this loop, or the one below).
6451 * ARC_FLAG_L2_WRITING is not enough to make
6452 * this distinction, and we need to know in
6453 * order to do proper l2arc vdev accounting in
6454 * arc_release() and arc_hdr_destroy().
6456 * Note, we can't use a new flag to distinguish
6457 * the two stages because we don't hold the
6458 * header's hash_lock below, in the second stage
6459 * of this function. Thus, we can't simply
6460 * change the b_flags field to denote that the
6461 * IO has been sent. We can change the b_daddr
6462 * field of the L2 portion, though, since we'll
6463 * be holding the l2ad_mtx; which is why we're
6464 * using it to denote the header's state change.
6466 hdr
->b_l2hdr
.b_daddr
= L2ARC_ADDR_UNSET
;
6467 hdr
->b_flags
|= ARC_FLAG_HAS_L2HDR
;
6469 mutex_enter(&dev
->l2ad_mtx
);
6470 list_insert_head(&dev
->l2ad_buflist
, hdr
);
6471 mutex_exit(&dev
->l2ad_mtx
);
6474 * Compute and store the buffer cksum before
6475 * writing. On debug the cksum is verified first.
6477 arc_cksum_verify(hdr
->b_l1hdr
.b_buf
);
6478 arc_cksum_compute(hdr
->b_l1hdr
.b_buf
, B_TRUE
);
6480 mutex_exit(hash_lock
);
6483 write_asize
+= buf_a_sz
;
6486 multilist_sublist_unlock(mls
);
6492 /* No buffers selected for writing? */
6495 ASSERT(!HDR_HAS_L1HDR(head
));
6496 kmem_cache_free(hdr_l2only_cache
, head
);
6500 mutex_enter(&dev
->l2ad_mtx
);
6503 * Note that elsewhere in this file arcstat_l2_asize
6504 * and the used space on l2ad_vdev are updated using b_asize,
6505 * which is not necessarily rounded up to the device block size.
6506 * Too keep accounting consistent we do the same here as well:
6507 * stats_size accumulates the sum of b_asize of the written buffers,
6508 * while write_asize accumulates the sum of b_asize rounded up
6509 * to the device block size.
6510 * The latter sum is used only to validate the corectness of the code.
6516 * Now start writing the buffers. We're starting at the write head
6517 * and work backwards, retracing the course of the buffer selector
6520 for (hdr
= list_prev(&dev
->l2ad_buflist
, head
); hdr
;
6521 hdr
= list_prev(&dev
->l2ad_buflist
, hdr
)) {
6525 * We rely on the L1 portion of the header below, so
6526 * it's invalid for this header to have been evicted out
6527 * of the ghost cache, prior to being written out. The
6528 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6530 ASSERT(HDR_HAS_L1HDR(hdr
));
6533 * We shouldn't need to lock the buffer here, since we flagged
6534 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6535 * take care to only access its L2 cache parameters. In
6536 * particular, hdr->l1hdr.b_buf may be invalid by now due to
6539 hdr
->b_l2hdr
.b_daddr
= dev
->l2ad_hand
;
6541 if ((!l2arc_nocompress
&& HDR_L2COMPRESS(hdr
)) &&
6542 hdr
->b_l2hdr
.b_asize
>= buf_compress_minsz
) {
6543 if (l2arc_compress_buf(hdr
)) {
6545 * If compression succeeded, enable headroom
6546 * boost on the next scan cycle.
6548 *headroom_boost
= B_TRUE
;
6553 * Pick up the buffer data we had previously stashed away
6554 * (and now potentially also compressed).
6556 buf_data
= hdr
->b_l1hdr
.b_tmp_cdata
;
6557 buf_sz
= hdr
->b_l2hdr
.b_asize
;
6560 * We need to do this regardless if buf_sz is zero or
6561 * not, otherwise, when this l2hdr is evicted we'll
6562 * remove a reference that was never added.
6564 (void) refcount_add_many(&dev
->l2ad_alloc
, buf_sz
, hdr
);
6566 /* Compression may have squashed the buffer to zero length. */
6570 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
6571 dev
->l2ad_hand
, buf_sz
, buf_data
, ZIO_CHECKSUM_OFF
,
6572 NULL
, NULL
, ZIO_PRIORITY_ASYNC_WRITE
,
6573 ZIO_FLAG_CANFAIL
, B_FALSE
);
6575 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
6577 (void) zio_nowait(wzio
);
6579 stats_size
+= buf_sz
;
6582 * Keep the clock hand suitably device-aligned.
6584 buf_a_sz
= vdev_psize_to_asize(dev
->l2ad_vdev
, buf_sz
);
6585 write_asize
+= buf_a_sz
;
6586 dev
->l2ad_hand
+= buf_a_sz
;
6590 mutex_exit(&dev
->l2ad_mtx
);
6592 ASSERT3U(write_asize
, <=, target_sz
);
6593 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
6594 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_asize
);
6595 ARCSTAT_INCR(arcstat_l2_size
, write_sz
);
6596 ARCSTAT_INCR(arcstat_l2_asize
, stats_size
);
6597 vdev_space_update(dev
->l2ad_vdev
, stats_size
, 0, 0);
6600 * Bump device hand to the device start if it is approaching the end.
6601 * l2arc_evict() will already have evicted ahead for this case.
6603 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
6604 dev
->l2ad_hand
= dev
->l2ad_start
;
6605 dev
->l2ad_first
= B_FALSE
;
6608 dev
->l2ad_writing
= B_TRUE
;
6609 (void) zio_wait(pio
);
6610 dev
->l2ad_writing
= B_FALSE
;
6612 return (write_asize
);
6616 * Compresses an L2ARC buffer.
6617 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6618 * size in l2hdr->b_asize. This routine tries to compress the data and
6619 * depending on the compression result there are three possible outcomes:
6620 * *) The buffer was incompressible. The original l2hdr contents were left
6621 * untouched and are ready for writing to an L2 device.
6622 * *) The buffer was all-zeros, so there is no need to write it to an L2
6623 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6624 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6625 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6626 * data buffer which holds the compressed data to be written, and b_asize
6627 * tells us how much data there is. b_compress is set to the appropriate
6628 * compression algorithm. Once writing is done, invoke
6629 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6631 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6632 * buffer was incompressible).
6635 l2arc_compress_buf(arc_buf_hdr_t
*hdr
)
6638 size_t csize
, len
, rounded
;
6639 l2arc_buf_hdr_t
*l2hdr
;
6641 ASSERT(HDR_HAS_L2HDR(hdr
));
6643 l2hdr
= &hdr
->b_l2hdr
;
6645 ASSERT(HDR_HAS_L1HDR(hdr
));
6646 ASSERT3U(l2hdr
->b_compress
, ==, ZIO_COMPRESS_OFF
);
6647 ASSERT(hdr
->b_l1hdr
.b_tmp_cdata
!= NULL
);
6649 len
= l2hdr
->b_asize
;
6650 cdata
= zio_data_buf_alloc(len
);
6651 ASSERT3P(cdata
, !=, NULL
);
6652 csize
= zio_compress_data(ZIO_COMPRESS_LZ4
, hdr
->b_l1hdr
.b_tmp_cdata
,
6653 cdata
, l2hdr
->b_asize
);
6655 rounded
= P2ROUNDUP(csize
, (size_t)SPA_MINBLOCKSIZE
);
6656 if (rounded
> csize
) {
6657 bzero((char *)cdata
+ csize
, rounded
- csize
);
6662 /* zero block, indicate that there's nothing to write */
6663 zio_data_buf_free(cdata
, len
);
6664 l2hdr
->b_compress
= ZIO_COMPRESS_EMPTY
;
6666 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6667 ARCSTAT_BUMP(arcstat_l2_compress_zeros
);
6669 } else if (csize
> 0 && csize
< len
) {
6671 * Compression succeeded, we'll keep the cdata around for
6672 * writing and release it afterwards.
6674 l2hdr
->b_compress
= ZIO_COMPRESS_LZ4
;
6675 l2hdr
->b_asize
= csize
;
6676 hdr
->b_l1hdr
.b_tmp_cdata
= cdata
;
6677 ARCSTAT_BUMP(arcstat_l2_compress_successes
);
6681 * Compression failed, release the compressed buffer.
6682 * l2hdr will be left unmodified.
6684 zio_data_buf_free(cdata
, len
);
6685 ARCSTAT_BUMP(arcstat_l2_compress_failures
);
6691 * Decompresses a zio read back from an l2arc device. On success, the
6692 * underlying zio's io_data buffer is overwritten by the uncompressed
6693 * version. On decompression error (corrupt compressed stream), the
6694 * zio->io_error value is set to signal an I/O error.
6696 * Please note that the compressed data stream is not checksummed, so
6697 * if the underlying device is experiencing data corruption, we may feed
6698 * corrupt data to the decompressor, so the decompressor needs to be
6699 * able to handle this situation (LZ4 does).
6702 l2arc_decompress_zio(zio_t
*zio
, arc_buf_hdr_t
*hdr
, enum zio_compress c
)
6707 ASSERT(L2ARC_IS_VALID_COMPRESS(c
));
6709 if (zio
->io_error
!= 0) {
6711 * An io error has occured, just restore the original io
6712 * size in preparation for a main pool read.
6714 zio
->io_orig_size
= zio
->io_size
= hdr
->b_size
;
6718 if (c
== ZIO_COMPRESS_EMPTY
) {
6720 * An empty buffer results in a null zio, which means we
6721 * need to fill its io_data after we're done restoring the
6722 * buffer's contents.
6724 ASSERT(hdr
->b_l1hdr
.b_buf
!= NULL
);
6725 bzero(hdr
->b_l1hdr
.b_buf
->b_data
, hdr
->b_size
);
6726 zio
->io_data
= zio
->io_orig_data
= hdr
->b_l1hdr
.b_buf
->b_data
;
6728 ASSERT(zio
->io_data
!= NULL
);
6730 * We copy the compressed data from the start of the arc buffer
6731 * (the zio_read will have pulled in only what we need, the
6732 * rest is garbage which we will overwrite at decompression)
6733 * and then decompress back to the ARC data buffer. This way we
6734 * can minimize copying by simply decompressing back over the
6735 * original compressed data (rather than decompressing to an
6736 * aux buffer and then copying back the uncompressed buffer,
6737 * which is likely to be much larger).
6739 csize
= zio
->io_size
;
6740 cdata
= zio_data_buf_alloc(csize
);
6741 bcopy(zio
->io_data
, cdata
, csize
);
6742 if (zio_decompress_data(c
, cdata
, zio
->io_data
, csize
,
6744 zio
->io_error
= EIO
;
6745 zio_data_buf_free(cdata
, csize
);
6748 /* Restore the expected uncompressed IO size. */
6749 zio
->io_orig_size
= zio
->io_size
= hdr
->b_size
;
6753 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6754 * This buffer serves as a temporary holder of compressed data while
6755 * the buffer entry is being written to an l2arc device. Once that is
6756 * done, we can dispose of it.
6759 l2arc_release_cdata_buf(arc_buf_hdr_t
*hdr
)
6761 enum zio_compress comp
;
6763 ASSERT(HDR_HAS_L1HDR(hdr
));
6764 ASSERT(HDR_HAS_L2HDR(hdr
));
6765 comp
= hdr
->b_l2hdr
.b_compress
;
6766 ASSERT(comp
== ZIO_COMPRESS_OFF
|| L2ARC_IS_VALID_COMPRESS(comp
));
6768 if (comp
== ZIO_COMPRESS_OFF
) {
6770 * In this case, b_tmp_cdata points to the same buffer
6771 * as the arc_buf_t's b_data field. We don't want to
6772 * free it, since the arc_buf_t will handle that.
6774 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6775 } else if (comp
== ZIO_COMPRESS_EMPTY
) {
6777 * In this case, b_tmp_cdata was compressed to an empty
6778 * buffer, thus there's nothing to free and b_tmp_cdata
6779 * should have been set to NULL in l2arc_write_buffers().
6781 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
6784 * If the data was compressed, then we've allocated a
6785 * temporary buffer for it, so now we need to release it.
6787 ASSERT(hdr
->b_l1hdr
.b_tmp_cdata
!= NULL
);
6788 zio_data_buf_free(hdr
->b_l1hdr
.b_tmp_cdata
,
6790 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6796 * This thread feeds the L2ARC at regular intervals. This is the beating
6797 * heart of the L2ARC.
6800 l2arc_feed_thread(void)
6805 uint64_t size
, wrote
;
6806 clock_t begin
, next
= ddi_get_lbolt();
6807 boolean_t headroom_boost
= B_FALSE
;
6808 fstrans_cookie_t cookie
;
6810 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
6812 mutex_enter(&l2arc_feed_thr_lock
);
6814 cookie
= spl_fstrans_mark();
6815 while (l2arc_thread_exit
== 0) {
6816 CALLB_CPR_SAFE_BEGIN(&cpr
);
6817 (void) cv_timedwait_sig(&l2arc_feed_thr_cv
,
6818 &l2arc_feed_thr_lock
, next
);
6819 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
6820 next
= ddi_get_lbolt() + hz
;
6823 * Quick check for L2ARC devices.
6825 mutex_enter(&l2arc_dev_mtx
);
6826 if (l2arc_ndev
== 0) {
6827 mutex_exit(&l2arc_dev_mtx
);
6830 mutex_exit(&l2arc_dev_mtx
);
6831 begin
= ddi_get_lbolt();
6834 * This selects the next l2arc device to write to, and in
6835 * doing so the next spa to feed from: dev->l2ad_spa. This
6836 * will return NULL if there are now no l2arc devices or if
6837 * they are all faulted.
6839 * If a device is returned, its spa's config lock is also
6840 * held to prevent device removal. l2arc_dev_get_next()
6841 * will grab and release l2arc_dev_mtx.
6843 if ((dev
= l2arc_dev_get_next()) == NULL
)
6846 spa
= dev
->l2ad_spa
;
6847 ASSERT(spa
!= NULL
);
6850 * If the pool is read-only then force the feed thread to
6851 * sleep a little longer.
6853 if (!spa_writeable(spa
)) {
6854 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
6855 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6860 * Avoid contributing to memory pressure.
6862 if (arc_reclaim_needed()) {
6863 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
6864 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6868 ARCSTAT_BUMP(arcstat_l2_feeds
);
6870 size
= l2arc_write_size();
6873 * Evict L2ARC buffers that will be overwritten.
6875 l2arc_evict(dev
, size
, B_FALSE
);
6878 * Write ARC buffers.
6880 wrote
= l2arc_write_buffers(spa
, dev
, size
, &headroom_boost
);
6883 * Calculate interval between writes.
6885 next
= l2arc_write_interval(begin
, size
, wrote
);
6886 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6888 spl_fstrans_unmark(cookie
);
6890 l2arc_thread_exit
= 0;
6891 cv_broadcast(&l2arc_feed_thr_cv
);
6892 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
6897 l2arc_vdev_present(vdev_t
*vd
)
6901 mutex_enter(&l2arc_dev_mtx
);
6902 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
6903 dev
= list_next(l2arc_dev_list
, dev
)) {
6904 if (dev
->l2ad_vdev
== vd
)
6907 mutex_exit(&l2arc_dev_mtx
);
6909 return (dev
!= NULL
);
6913 * Add a vdev for use by the L2ARC. By this point the spa has already
6914 * validated the vdev and opened it.
6917 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
6919 l2arc_dev_t
*adddev
;
6921 ASSERT(!l2arc_vdev_present(vd
));
6924 * Create a new l2arc device entry.
6926 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
6927 adddev
->l2ad_spa
= spa
;
6928 adddev
->l2ad_vdev
= vd
;
6929 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
6930 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
6931 adddev
->l2ad_hand
= adddev
->l2ad_start
;
6932 adddev
->l2ad_first
= B_TRUE
;
6933 adddev
->l2ad_writing
= B_FALSE
;
6934 list_link_init(&adddev
->l2ad_node
);
6936 mutex_init(&adddev
->l2ad_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6938 * This is a list of all ARC buffers that are still valid on the
6941 list_create(&adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
6942 offsetof(arc_buf_hdr_t
, b_l2hdr
.b_l2node
));
6944 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
6945 refcount_create(&adddev
->l2ad_alloc
);
6948 * Add device to global list
6950 mutex_enter(&l2arc_dev_mtx
);
6951 list_insert_head(l2arc_dev_list
, adddev
);
6952 atomic_inc_64(&l2arc_ndev
);
6953 mutex_exit(&l2arc_dev_mtx
);
6957 * Remove a vdev from the L2ARC.
6960 l2arc_remove_vdev(vdev_t
*vd
)
6962 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
6965 * Find the device by vdev
6967 mutex_enter(&l2arc_dev_mtx
);
6968 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
6969 nextdev
= list_next(l2arc_dev_list
, dev
);
6970 if (vd
== dev
->l2ad_vdev
) {
6975 ASSERT(remdev
!= NULL
);
6978 * Remove device from global list
6980 list_remove(l2arc_dev_list
, remdev
);
6981 l2arc_dev_last
= NULL
; /* may have been invalidated */
6982 atomic_dec_64(&l2arc_ndev
);
6983 mutex_exit(&l2arc_dev_mtx
);
6986 * Clear all buflists and ARC references. L2ARC device flush.
6988 l2arc_evict(remdev
, 0, B_TRUE
);
6989 list_destroy(&remdev
->l2ad_buflist
);
6990 mutex_destroy(&remdev
->l2ad_mtx
);
6991 refcount_destroy(&remdev
->l2ad_alloc
);
6992 kmem_free(remdev
, sizeof (l2arc_dev_t
));
6998 l2arc_thread_exit
= 0;
7000 l2arc_writes_sent
= 0;
7001 l2arc_writes_done
= 0;
7003 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
7004 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
7005 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7006 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
7008 l2arc_dev_list
= &L2ARC_dev_list
;
7009 l2arc_free_on_write
= &L2ARC_free_on_write
;
7010 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
7011 offsetof(l2arc_dev_t
, l2ad_node
));
7012 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
7013 offsetof(l2arc_data_free_t
, l2df_list_node
));
7020 * This is called from dmu_fini(), which is called from spa_fini();
7021 * Because of this, we can assume that all l2arc devices have
7022 * already been removed when the pools themselves were removed.
7025 l2arc_do_free_on_write();
7027 mutex_destroy(&l2arc_feed_thr_lock
);
7028 cv_destroy(&l2arc_feed_thr_cv
);
7029 mutex_destroy(&l2arc_dev_mtx
);
7030 mutex_destroy(&l2arc_free_on_write_mtx
);
7032 list_destroy(l2arc_dev_list
);
7033 list_destroy(l2arc_free_on_write
);
7039 if (!(spa_mode_global
& FWRITE
))
7042 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
7043 TS_RUN
, defclsyspri
);
7049 if (!(spa_mode_global
& FWRITE
))
7052 mutex_enter(&l2arc_feed_thr_lock
);
7053 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
7054 l2arc_thread_exit
= 1;
7055 while (l2arc_thread_exit
!= 0)
7056 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
7057 mutex_exit(&l2arc_feed_thr_lock
);
7060 #if defined(_KERNEL) && defined(HAVE_SPL)
7061 EXPORT_SYMBOL(arc_buf_size
);
7062 EXPORT_SYMBOL(arc_write
);
7063 EXPORT_SYMBOL(arc_read
);
7064 EXPORT_SYMBOL(arc_buf_remove_ref
);
7065 EXPORT_SYMBOL(arc_buf_info
);
7066 EXPORT_SYMBOL(arc_getbuf_func
);
7067 EXPORT_SYMBOL(arc_add_prune_callback
);
7068 EXPORT_SYMBOL(arc_remove_prune_callback
);
7070 module_param(zfs_arc_min
, ulong
, 0644);
7071 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
7073 module_param(zfs_arc_max
, ulong
, 0644);
7074 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
7076 module_param(zfs_arc_meta_limit
, ulong
, 0644);
7077 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
7079 module_param(zfs_arc_meta_min
, ulong
, 0644);
7080 MODULE_PARM_DESC(zfs_arc_meta_min
, "Min arc metadata");
7082 module_param(zfs_arc_meta_prune
, int, 0644);
7083 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Meta objects to scan for prune");
7085 module_param(zfs_arc_meta_adjust_restarts
, int, 0644);
7086 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts
,
7087 "Limit number of restarts in arc_adjust_meta");
7089 module_param(zfs_arc_meta_strategy
, int, 0644);
7090 MODULE_PARM_DESC(zfs_arc_meta_strategy
, "Meta reclaim strategy");
7092 module_param(zfs_arc_grow_retry
, int, 0644);
7093 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
7095 module_param(zfs_arc_p_aggressive_disable
, int, 0644);
7096 MODULE_PARM_DESC(zfs_arc_p_aggressive_disable
, "disable aggressive arc_p grow");
7098 module_param(zfs_arc_p_dampener_disable
, int, 0644);
7099 MODULE_PARM_DESC(zfs_arc_p_dampener_disable
, "disable arc_p adapt dampener");
7101 module_param(zfs_arc_shrink_shift
, int, 0644);
7102 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
7104 module_param(zfs_arc_p_min_shift
, int, 0644);
7105 MODULE_PARM_DESC(zfs_arc_p_min_shift
, "arc_c shift to calc min/max arc_p");
7107 module_param(zfs_disable_dup_eviction
, int, 0644);
7108 MODULE_PARM_DESC(zfs_disable_dup_eviction
, "disable duplicate buffer eviction");
7110 module_param(zfs_arc_average_blocksize
, int, 0444);
7111 MODULE_PARM_DESC(zfs_arc_average_blocksize
, "Target average block size");
7113 module_param(zfs_arc_min_prefetch_lifespan
, int, 0644);
7114 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan
, "Min life of prefetch block");
7116 module_param(zfs_arc_num_sublists_per_state
, int, 0644);
7117 MODULE_PARM_DESC(zfs_arc_num_sublists_per_state
,
7118 "Number of sublists used in each of the ARC state lists");
7120 module_param(l2arc_write_max
, ulong
, 0644);
7121 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
7123 module_param(l2arc_write_boost
, ulong
, 0644);
7124 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
7126 module_param(l2arc_headroom
, ulong
, 0644);
7127 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
7129 module_param(l2arc_headroom_boost
, ulong
, 0644);
7130 MODULE_PARM_DESC(l2arc_headroom_boost
, "Compressed l2arc_headroom multiplier");
7132 module_param(l2arc_feed_secs
, ulong
, 0644);
7133 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
7135 module_param(l2arc_feed_min_ms
, ulong
, 0644);
7136 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
7138 module_param(l2arc_noprefetch
, int, 0644);
7139 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
7141 module_param(l2arc_nocompress
, int, 0644);
7142 MODULE_PARM_DESC(l2arc_nocompress
, "Skip compressing L2ARC buffers");
7144 module_param(l2arc_feed_again
, int, 0644);
7145 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
7147 module_param(l2arc_norw
, int, 0644);
7148 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");
7150 module_param(zfs_arc_lotsfree_percent
, int, 0644);
7151 MODULE_PARM_DESC(zfs_arc_lotsfree_percent
,
7152 "System free memory I/O throttle in bytes");
7154 module_param(zfs_arc_sys_free
, ulong
, 0644);
7155 MODULE_PARM_DESC(zfs_arc_sys_free
, "System free memory target size in bytes");