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 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
29 * DVA-based Adjustable Replacement Cache
31 * While much of the theory of operation used here is
32 * based on the self-tuning, low overhead replacement cache
33 * presented by Megiddo and Modha at FAST 2003, there are some
34 * significant differences:
36 * 1. The Megiddo and Modha model assumes any page is evictable.
37 * Pages in its cache cannot be "locked" into memory. This makes
38 * the eviction algorithm simple: evict the last page in the list.
39 * This also make the performance characteristics easy to reason
40 * about. Our cache is not so simple. At any given moment, some
41 * subset of the blocks in the cache are un-evictable because we
42 * have handed out a reference to them. Blocks are only evictable
43 * when there are no external references active. This makes
44 * eviction far more problematic: we choose to evict the evictable
45 * blocks that are the "lowest" in the list.
47 * There are times when it is not possible to evict the requested
48 * space. In these circumstances we are unable to adjust the cache
49 * size. To prevent the cache growing unbounded at these times we
50 * implement a "cache throttle" that slows the flow of new data
51 * into the cache until we can make space available.
53 * 2. The Megiddo and Modha model assumes a fixed cache size.
54 * Pages are evicted when the cache is full and there is a cache
55 * miss. Our model has a variable sized cache. It grows with
56 * high use, but also tries to react to memory pressure from the
57 * operating system: decreasing its size when system memory is
60 * 3. The Megiddo and Modha model assumes a fixed page size. All
61 * elements of the cache are therefore exactly the same size. So
62 * when adjusting the cache size following a cache miss, its simply
63 * a matter of choosing a single page to evict. In our model, we
64 * have variable sized cache blocks (rangeing from 512 bytes to
65 * 128K bytes). We therefore choose a set of blocks to evict to make
66 * space for a cache miss that approximates as closely as possible
67 * the space used by the new block.
69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
70 * by N. Megiddo & D. Modha, FAST 2003
76 * A new reference to a cache buffer can be obtained in two
77 * ways: 1) via a hash table lookup using the DVA as a key,
78 * or 2) via one of the ARC lists. The arc_read() interface
79 * uses method 1, while the internal arc algorithms for
80 * adjusting the cache use method 2. We therefore provide two
81 * types of locks: 1) the hash table lock array, and 2) the
84 * Buffers do not have their own mutexes, rather they rely on the
85 * hash table mutexes for the bulk of their protection (i.e. most
86 * fields in the arc_buf_hdr_t are protected by these mutexes).
88 * buf_hash_find() returns the appropriate mutex (held) when it
89 * locates the requested buffer in the hash table. It returns
90 * NULL for the mutex if the buffer was not in the table.
92 * buf_hash_remove() expects the appropriate hash mutex to be
93 * already held before it is invoked.
95 * Each arc state also has a mutex which is used to protect the
96 * buffer list associated with the state. When attempting to
97 * obtain a hash table lock while holding an arc list lock you
98 * must use: mutex_tryenter() to avoid deadlock. Also note that
99 * the active state mutex must be held before the ghost state mutex.
101 * Arc buffers may have an associated eviction callback function.
102 * This function will be invoked prior to removing the buffer (e.g.
103 * in arc_do_user_evicts()). Note however that the data associated
104 * with the buffer may be evicted prior to the callback. The callback
105 * must be made with *no locks held* (to prevent deadlock). Additionally,
106 * the users of callbacks must ensure that their private data is
107 * protected from simultaneous callbacks from arc_buf_evict()
108 * and arc_do_user_evicts().
110 * It as also possible to register a callback which is run when the
111 * arc_meta_limit is reached and no buffers can be safely evicted. In
112 * this case the arc user should drop a reference on some arc buffers so
113 * they can be reclaimed and the arc_meta_limit honored. For example,
114 * when using the ZPL each dentry holds a references on a znode. These
115 * dentries must be pruned before the arc buffer holding the znode can
118 * Note that the majority of the performance stats are manipulated
119 * with atomic operations.
121 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
123 * - L2ARC buflist creation
124 * - L2ARC buflist eviction
125 * - L2ARC write completion, which walks L2ARC buflists
126 * - ARC header destruction, as it removes from L2ARC buflists
127 * - ARC header release, as it removes from L2ARC buflists
132 #include <sys/zio_compress.h>
133 #include <sys/zfs_context.h>
135 #include <sys/vdev.h>
136 #include <sys/vdev_impl.h>
137 #include <sys/dsl_pool.h>
139 #include <sys/vmsystm.h>
141 #include <sys/fs/swapnode.h>
144 #include <sys/callb.h>
145 #include <sys/kstat.h>
146 #include <sys/dmu_tx.h>
147 #include <zfs_fletcher.h>
150 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
151 boolean_t arc_watch = B_FALSE;
154 static kmutex_t arc_reclaim_thr_lock;
155 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
156 static uint8_t arc_thread_exit;
158 /* number of bytes to prune from caches when at arc_meta_limit is reached */
159 int zfs_arc_meta_prune = 1048576;
161 typedef enum arc_reclaim_strategy {
162 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
163 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
164 } arc_reclaim_strategy_t;
167 * The number of iterations through arc_evict_*() before we
168 * drop & reacquire the lock.
170 int arc_evict_iterations = 100;
172 /* number of seconds before growing cache again */
173 int zfs_arc_grow_retry = 5;
175 /* disable anon data aggressively growing arc_p */
176 int zfs_arc_p_aggressive_disable = 1;
178 /* log2(fraction of arc to reclaim) */
179 int zfs_arc_shrink_shift = 5;
182 * minimum lifespan of a prefetch block in clock ticks
183 * (initialized in arc_init())
185 int zfs_arc_min_prefetch_lifespan = HZ;
187 /* disable arc proactive arc throttle due to low memory */
188 int zfs_arc_memory_throttle_disable = 1;
190 /* disable duplicate buffer eviction */
191 int zfs_disable_dup_eviction = 0;
194 * If this percent of memory is free, don't throttle.
196 int arc_lotsfree_percent = 10;
200 /* expiration time for arc_no_grow */
201 static clock_t arc_grow_time = 0;
204 * The arc has filled available memory and has now warmed up.
206 static boolean_t arc_warm;
209 * These tunables are for performance analysis.
211 unsigned long zfs_arc_max = 0;
212 unsigned long zfs_arc_min = 0;
213 unsigned long zfs_arc_meta_limit = 0;
216 * Note that buffers can be in one of 6 states:
217 * ARC_anon - anonymous (discussed below)
218 * ARC_mru - recently used, currently cached
219 * ARC_mru_ghost - recentely used, no longer in cache
220 * ARC_mfu - frequently used, currently cached
221 * ARC_mfu_ghost - frequently used, no longer in cache
222 * ARC_l2c_only - exists in L2ARC but not other states
223 * When there are no active references to the buffer, they are
224 * are linked onto a list in one of these arc states. These are
225 * the only buffers that can be evicted or deleted. Within each
226 * state there are multiple lists, one for meta-data and one for
227 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
228 * etc.) is tracked separately so that it can be managed more
229 * explicitly: favored over data, limited explicitly.
231 * Anonymous buffers are buffers that are not associated with
232 * a DVA. These are buffers that hold dirty block copies
233 * before they are written to stable storage. By definition,
234 * they are "ref'd" and are considered part of arc_mru
235 * that cannot be freed. Generally, they will aquire a DVA
236 * as they are written and migrate onto the arc_mru list.
238 * The ARC_l2c_only state is for buffers that are in the second
239 * level ARC but no longer in any of the ARC_m* lists. The second
240 * level ARC itself may also contain buffers that are in any of
241 * the ARC_m* states - meaning that a buffer can exist in two
242 * places. The reason for the ARC_l2c_only state is to keep the
243 * buffer header in the hash table, so that reads that hit the
244 * second level ARC benefit from these fast lookups.
247 typedef struct arc_state {
248 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
249 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
250 uint64_t arcs_size; /* total amount of data in this state */
252 arc_state_type_t arcs_state;
256 static arc_state_t ARC_anon;
257 static arc_state_t ARC_mru;
258 static arc_state_t ARC_mru_ghost;
259 static arc_state_t ARC_mfu;
260 static arc_state_t ARC_mfu_ghost;
261 static arc_state_t ARC_l2c_only;
263 typedef struct arc_stats {
264 kstat_named_t arcstat_hits;
265 kstat_named_t arcstat_misses;
266 kstat_named_t arcstat_demand_data_hits;
267 kstat_named_t arcstat_demand_data_misses;
268 kstat_named_t arcstat_demand_metadata_hits;
269 kstat_named_t arcstat_demand_metadata_misses;
270 kstat_named_t arcstat_prefetch_data_hits;
271 kstat_named_t arcstat_prefetch_data_misses;
272 kstat_named_t arcstat_prefetch_metadata_hits;
273 kstat_named_t arcstat_prefetch_metadata_misses;
274 kstat_named_t arcstat_mru_hits;
275 kstat_named_t arcstat_mru_ghost_hits;
276 kstat_named_t arcstat_mfu_hits;
277 kstat_named_t arcstat_mfu_ghost_hits;
278 kstat_named_t arcstat_deleted;
279 kstat_named_t arcstat_recycle_miss;
281 * Number of buffers that could not be evicted because the hash lock
282 * was held by another thread. The lock may not necessarily be held
283 * by something using the same buffer, since hash locks are shared
284 * by multiple buffers.
286 kstat_named_t arcstat_mutex_miss;
288 * Number of buffers skipped because they have I/O in progress, are
289 * indrect prefetch buffers that have not lived long enough, or are
290 * not from the spa we're trying to evict from.
292 kstat_named_t arcstat_evict_skip;
293 kstat_named_t arcstat_evict_l2_cached;
294 kstat_named_t arcstat_evict_l2_eligible;
295 kstat_named_t arcstat_evict_l2_ineligible;
296 kstat_named_t arcstat_hash_elements;
297 kstat_named_t arcstat_hash_elements_max;
298 kstat_named_t arcstat_hash_collisions;
299 kstat_named_t arcstat_hash_chains;
300 kstat_named_t arcstat_hash_chain_max;
301 kstat_named_t arcstat_p;
302 kstat_named_t arcstat_c;
303 kstat_named_t arcstat_c_min;
304 kstat_named_t arcstat_c_max;
305 kstat_named_t arcstat_size;
306 kstat_named_t arcstat_hdr_size;
307 kstat_named_t arcstat_data_size;
308 kstat_named_t arcstat_other_size;
309 kstat_named_t arcstat_anon_size;
310 kstat_named_t arcstat_anon_evict_data;
311 kstat_named_t arcstat_anon_evict_metadata;
312 kstat_named_t arcstat_mru_size;
313 kstat_named_t arcstat_mru_evict_data;
314 kstat_named_t arcstat_mru_evict_metadata;
315 kstat_named_t arcstat_mru_ghost_size;
316 kstat_named_t arcstat_mru_ghost_evict_data;
317 kstat_named_t arcstat_mru_ghost_evict_metadata;
318 kstat_named_t arcstat_mfu_size;
319 kstat_named_t arcstat_mfu_evict_data;
320 kstat_named_t arcstat_mfu_evict_metadata;
321 kstat_named_t arcstat_mfu_ghost_size;
322 kstat_named_t arcstat_mfu_ghost_evict_data;
323 kstat_named_t arcstat_mfu_ghost_evict_metadata;
324 kstat_named_t arcstat_l2_hits;
325 kstat_named_t arcstat_l2_misses;
326 kstat_named_t arcstat_l2_feeds;
327 kstat_named_t arcstat_l2_rw_clash;
328 kstat_named_t arcstat_l2_read_bytes;
329 kstat_named_t arcstat_l2_write_bytes;
330 kstat_named_t arcstat_l2_writes_sent;
331 kstat_named_t arcstat_l2_writes_done;
332 kstat_named_t arcstat_l2_writes_error;
333 kstat_named_t arcstat_l2_writes_hdr_miss;
334 kstat_named_t arcstat_l2_evict_lock_retry;
335 kstat_named_t arcstat_l2_evict_reading;
336 kstat_named_t arcstat_l2_free_on_write;
337 kstat_named_t arcstat_l2_abort_lowmem;
338 kstat_named_t arcstat_l2_cksum_bad;
339 kstat_named_t arcstat_l2_io_error;
340 kstat_named_t arcstat_l2_size;
341 kstat_named_t arcstat_l2_asize;
342 kstat_named_t arcstat_l2_hdr_size;
343 kstat_named_t arcstat_l2_compress_successes;
344 kstat_named_t arcstat_l2_compress_zeros;
345 kstat_named_t arcstat_l2_compress_failures;
346 kstat_named_t arcstat_memory_throttle_count;
347 kstat_named_t arcstat_duplicate_buffers;
348 kstat_named_t arcstat_duplicate_buffers_size;
349 kstat_named_t arcstat_duplicate_reads;
350 kstat_named_t arcstat_memory_direct_count;
351 kstat_named_t arcstat_memory_indirect_count;
352 kstat_named_t arcstat_no_grow;
353 kstat_named_t arcstat_tempreserve;
354 kstat_named_t arcstat_loaned_bytes;
355 kstat_named_t arcstat_prune;
356 kstat_named_t arcstat_meta_used;
357 kstat_named_t arcstat_meta_limit;
358 kstat_named_t arcstat_meta_max;
361 static arc_stats_t arc_stats = {
362 { "hits", KSTAT_DATA_UINT64 },
363 { "misses", KSTAT_DATA_UINT64 },
364 { "demand_data_hits", KSTAT_DATA_UINT64 },
365 { "demand_data_misses", KSTAT_DATA_UINT64 },
366 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
367 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
368 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
369 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
370 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
371 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
372 { "mru_hits", KSTAT_DATA_UINT64 },
373 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
374 { "mfu_hits", KSTAT_DATA_UINT64 },
375 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
376 { "deleted", KSTAT_DATA_UINT64 },
377 { "recycle_miss", KSTAT_DATA_UINT64 },
378 { "mutex_miss", KSTAT_DATA_UINT64 },
379 { "evict_skip", KSTAT_DATA_UINT64 },
380 { "evict_l2_cached", KSTAT_DATA_UINT64 },
381 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
382 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
383 { "hash_elements", KSTAT_DATA_UINT64 },
384 { "hash_elements_max", KSTAT_DATA_UINT64 },
385 { "hash_collisions", KSTAT_DATA_UINT64 },
386 { "hash_chains", KSTAT_DATA_UINT64 },
387 { "hash_chain_max", KSTAT_DATA_UINT64 },
388 { "p", KSTAT_DATA_UINT64 },
389 { "c", KSTAT_DATA_UINT64 },
390 { "c_min", KSTAT_DATA_UINT64 },
391 { "c_max", KSTAT_DATA_UINT64 },
392 { "size", KSTAT_DATA_UINT64 },
393 { "hdr_size", KSTAT_DATA_UINT64 },
394 { "data_size", KSTAT_DATA_UINT64 },
395 { "other_size", KSTAT_DATA_UINT64 },
396 { "anon_size", KSTAT_DATA_UINT64 },
397 { "anon_evict_data", KSTAT_DATA_UINT64 },
398 { "anon_evict_metadata", KSTAT_DATA_UINT64 },
399 { "mru_size", KSTAT_DATA_UINT64 },
400 { "mru_evict_data", KSTAT_DATA_UINT64 },
401 { "mru_evict_metadata", KSTAT_DATA_UINT64 },
402 { "mru_ghost_size", KSTAT_DATA_UINT64 },
403 { "mru_ghost_evict_data", KSTAT_DATA_UINT64 },
404 { "mru_ghost_evict_metadata", KSTAT_DATA_UINT64 },
405 { "mfu_size", KSTAT_DATA_UINT64 },
406 { "mfu_evict_data", KSTAT_DATA_UINT64 },
407 { "mfu_evict_metadata", KSTAT_DATA_UINT64 },
408 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
409 { "mfu_ghost_evict_data", KSTAT_DATA_UINT64 },
410 { "mfu_ghost_evict_metadata", KSTAT_DATA_UINT64 },
411 { "l2_hits", KSTAT_DATA_UINT64 },
412 { "l2_misses", KSTAT_DATA_UINT64 },
413 { "l2_feeds", KSTAT_DATA_UINT64 },
414 { "l2_rw_clash", KSTAT_DATA_UINT64 },
415 { "l2_read_bytes", KSTAT_DATA_UINT64 },
416 { "l2_write_bytes", KSTAT_DATA_UINT64 },
417 { "l2_writes_sent", KSTAT_DATA_UINT64 },
418 { "l2_writes_done", KSTAT_DATA_UINT64 },
419 { "l2_writes_error", KSTAT_DATA_UINT64 },
420 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
421 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
422 { "l2_evict_reading", KSTAT_DATA_UINT64 },
423 { "l2_free_on_write", KSTAT_DATA_UINT64 },
424 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
425 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
426 { "l2_io_error", KSTAT_DATA_UINT64 },
427 { "l2_size", KSTAT_DATA_UINT64 },
428 { "l2_asize", KSTAT_DATA_UINT64 },
429 { "l2_hdr_size", KSTAT_DATA_UINT64 },
430 { "l2_compress_successes", KSTAT_DATA_UINT64 },
431 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
432 { "l2_compress_failures", KSTAT_DATA_UINT64 },
433 { "memory_throttle_count", KSTAT_DATA_UINT64 },
434 { "duplicate_buffers", KSTAT_DATA_UINT64 },
435 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
436 { "duplicate_reads", KSTAT_DATA_UINT64 },
437 { "memory_direct_count", KSTAT_DATA_UINT64 },
438 { "memory_indirect_count", KSTAT_DATA_UINT64 },
439 { "arc_no_grow", KSTAT_DATA_UINT64 },
440 { "arc_tempreserve", KSTAT_DATA_UINT64 },
441 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
442 { "arc_prune", KSTAT_DATA_UINT64 },
443 { "arc_meta_used", KSTAT_DATA_UINT64 },
444 { "arc_meta_limit", KSTAT_DATA_UINT64 },
445 { "arc_meta_max", KSTAT_DATA_UINT64 },
448 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
450 #define ARCSTAT_INCR(stat, val) \
451 atomic_add_64(&arc_stats.stat.value.ui64, (val))
453 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
454 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
456 #define ARCSTAT_MAX(stat, val) { \
458 while ((val) > (m = arc_stats.stat.value.ui64) && \
459 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
463 #define ARCSTAT_MAXSTAT(stat) \
464 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
467 * We define a macro to allow ARC hits/misses to be easily broken down by
468 * two separate conditions, giving a total of four different subtypes for
469 * each of hits and misses (so eight statistics total).
471 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
474 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
476 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
480 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
482 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
487 static arc_state_t *arc_anon;
488 static arc_state_t *arc_mru;
489 static arc_state_t *arc_mru_ghost;
490 static arc_state_t *arc_mfu;
491 static arc_state_t *arc_mfu_ghost;
492 static arc_state_t *arc_l2c_only;
495 * There are several ARC variables that are critical to export as kstats --
496 * but we don't want to have to grovel around in the kstat whenever we wish to
497 * manipulate them. For these variables, we therefore define them to be in
498 * terms of the statistic variable. This assures that we are not introducing
499 * the possibility of inconsistency by having shadow copies of the variables,
500 * while still allowing the code to be readable.
502 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
503 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
504 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
505 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
506 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
507 #define arc_no_grow ARCSTAT(arcstat_no_grow)
508 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
509 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
510 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
511 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
512 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
514 #define L2ARC_IS_VALID_COMPRESS(_c_) \
515 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
517 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
519 typedef struct arc_callback arc_callback_t;
521 struct arc_callback {
523 arc_done_func_t *acb_done;
525 zio_t *acb_zio_dummy;
526 arc_callback_t *acb_next;
529 typedef struct arc_write_callback arc_write_callback_t;
531 struct arc_write_callback {
533 arc_done_func_t *awcb_ready;
534 arc_done_func_t *awcb_physdone;
535 arc_done_func_t *awcb_done;
540 /* protected by hash lock */
545 kmutex_t b_freeze_lock;
546 zio_cksum_t *b_freeze_cksum;
548 arc_buf_hdr_t *b_hash_next;
553 arc_callback_t *b_acb;
557 arc_buf_contents_t b_type;
561 /* protected by arc state mutex */
562 arc_state_t *b_state;
563 list_node_t b_arc_node;
565 /* updated atomically */
566 clock_t b_arc_access;
568 uint32_t b_mru_ghost_hits;
570 uint32_t b_mfu_ghost_hits;
573 /* self protecting */
576 l2arc_buf_hdr_t *b_l2hdr;
577 list_node_t b_l2node;
580 static list_t arc_prune_list;
581 static kmutex_t arc_prune_mtx;
582 static arc_buf_t *arc_eviction_list;
583 static kmutex_t arc_eviction_mtx;
584 static arc_buf_hdr_t arc_eviction_hdr;
585 static void arc_get_data_buf(arc_buf_t *buf);
586 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
587 static int arc_evict_needed(arc_buf_contents_t type);
588 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes,
589 arc_buf_contents_t type);
590 static void arc_buf_watch(arc_buf_t *buf);
592 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
594 #define GHOST_STATE(state) \
595 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
596 (state) == arc_l2c_only)
599 * Private ARC flags. These flags are private ARC only flags that will show up
600 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
601 * be passed in as arc_flags in things like arc_read. However, these flags
602 * should never be passed and should only be set by ARC code. When adding new
603 * public flags, make sure not to smash the private ones.
606 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
607 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
608 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
609 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
610 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
611 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
612 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
613 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
614 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
615 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
617 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
618 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
619 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
620 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
621 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
622 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
623 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
624 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
625 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
626 (hdr)->b_l2hdr != NULL)
627 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
628 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
629 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
635 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
636 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
639 * Hash table routines
642 #define HT_LOCK_ALIGN 64
643 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
648 unsigned char pad[HT_LOCK_PAD];
652 #define BUF_LOCKS 256
653 typedef struct buf_hash_table {
655 arc_buf_hdr_t **ht_table;
656 struct ht_lock ht_locks[BUF_LOCKS];
659 static buf_hash_table_t buf_hash_table;
661 #define BUF_HASH_INDEX(spa, dva, birth) \
662 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
663 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
664 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
665 #define HDR_LOCK(hdr) \
666 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
668 uint64_t zfs_crc64_table[256];
674 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
675 #define L2ARC_HEADROOM 2 /* num of writes */
677 * If we discover during ARC scan any buffers to be compressed, we boost
678 * our headroom for the next scanning cycle by this percentage multiple.
680 #define L2ARC_HEADROOM_BOOST 200
681 #define L2ARC_FEED_SECS 1 /* caching interval secs */
682 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
684 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
685 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
687 /* L2ARC Performance Tunables */
688 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
689 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
690 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
691 unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
692 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
693 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
694 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
695 int l2arc_nocompress = B_FALSE; /* don't compress bufs */
696 int l2arc_feed_again = B_TRUE; /* turbo warmup */
697 int l2arc_norw = B_FALSE; /* no reads during writes */
702 typedef struct l2arc_dev {
703 vdev_t *l2ad_vdev; /* vdev */
704 spa_t *l2ad_spa; /* spa */
705 uint64_t l2ad_hand; /* next write location */
706 uint64_t l2ad_start; /* first addr on device */
707 uint64_t l2ad_end; /* last addr on device */
708 uint64_t l2ad_evict; /* last addr eviction reached */
709 boolean_t l2ad_first; /* first sweep through */
710 boolean_t l2ad_writing; /* currently writing */
711 list_t *l2ad_buflist; /* buffer list */
712 list_node_t l2ad_node; /* device list node */
715 static list_t L2ARC_dev_list; /* device list */
716 static list_t *l2arc_dev_list; /* device list pointer */
717 static kmutex_t l2arc_dev_mtx; /* device list mutex */
718 static l2arc_dev_t *l2arc_dev_last; /* last device used */
719 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
720 static list_t L2ARC_free_on_write; /* free after write buf list */
721 static list_t *l2arc_free_on_write; /* free after write list ptr */
722 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
723 static uint64_t l2arc_ndev; /* number of devices */
725 typedef struct l2arc_read_callback {
726 arc_buf_t *l2rcb_buf; /* read buffer */
727 spa_t *l2rcb_spa; /* spa */
728 blkptr_t l2rcb_bp; /* original blkptr */
729 zbookmark_t l2rcb_zb; /* original bookmark */
730 int l2rcb_flags; /* original flags */
731 enum zio_compress l2rcb_compress; /* applied compress */
732 } l2arc_read_callback_t;
734 typedef struct l2arc_write_callback {
735 l2arc_dev_t *l2wcb_dev; /* device info */
736 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
737 } l2arc_write_callback_t;
739 struct l2arc_buf_hdr {
740 /* protected by arc_buf_hdr mutex */
741 l2arc_dev_t *b_dev; /* L2ARC device */
742 uint64_t b_daddr; /* disk address, offset byte */
743 /* compression applied to buffer data */
744 enum zio_compress b_compress;
745 /* real alloc'd buffer size depending on b_compress applied */
748 /* temporary buffer holder for in-flight compressed data */
752 typedef struct l2arc_data_free {
753 /* protected by l2arc_free_on_write_mtx */
756 void (*l2df_func)(void *, size_t);
757 list_node_t l2df_list_node;
760 static kmutex_t l2arc_feed_thr_lock;
761 static kcondvar_t l2arc_feed_thr_cv;
762 static uint8_t l2arc_thread_exit;
764 static void l2arc_read_done(zio_t *zio);
765 static void l2arc_hdr_stat_add(void);
766 static void l2arc_hdr_stat_remove(void);
768 static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr);
769 static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr,
770 enum zio_compress c);
771 static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab);
774 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
776 uint8_t *vdva = (uint8_t *)dva;
777 uint64_t crc = -1ULL;
780 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
782 for (i = 0; i < sizeof (dva_t); i++)
783 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
785 crc ^= (spa>>8) ^ birth;
790 #define BUF_EMPTY(buf) \
791 ((buf)->b_dva.dva_word[0] == 0 && \
792 (buf)->b_dva.dva_word[1] == 0 && \
795 #define BUF_EQUAL(spa, dva, birth, buf) \
796 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
797 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
798 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
801 buf_discard_identity(arc_buf_hdr_t *hdr)
803 hdr->b_dva.dva_word[0] = 0;
804 hdr->b_dva.dva_word[1] = 0;
809 static arc_buf_hdr_t *
810 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
812 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
813 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
816 mutex_enter(hash_lock);
817 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
818 buf = buf->b_hash_next) {
819 if (BUF_EQUAL(spa, dva, birth, buf)) {
824 mutex_exit(hash_lock);
830 * Insert an entry into the hash table. If there is already an element
831 * equal to elem in the hash table, then the already existing element
832 * will be returned and the new element will not be inserted.
833 * Otherwise returns NULL.
835 static arc_buf_hdr_t *
836 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
838 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
839 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
843 ASSERT(!HDR_IN_HASH_TABLE(buf));
845 mutex_enter(hash_lock);
846 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
847 fbuf = fbuf->b_hash_next, i++) {
848 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
852 buf->b_hash_next = buf_hash_table.ht_table[idx];
853 buf_hash_table.ht_table[idx] = buf;
854 buf->b_flags |= ARC_IN_HASH_TABLE;
856 /* collect some hash table performance data */
858 ARCSTAT_BUMP(arcstat_hash_collisions);
860 ARCSTAT_BUMP(arcstat_hash_chains);
862 ARCSTAT_MAX(arcstat_hash_chain_max, i);
865 ARCSTAT_BUMP(arcstat_hash_elements);
866 ARCSTAT_MAXSTAT(arcstat_hash_elements);
872 buf_hash_remove(arc_buf_hdr_t *buf)
874 arc_buf_hdr_t *fbuf, **bufp;
875 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
877 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
878 ASSERT(HDR_IN_HASH_TABLE(buf));
880 bufp = &buf_hash_table.ht_table[idx];
881 while ((fbuf = *bufp) != buf) {
882 ASSERT(fbuf != NULL);
883 bufp = &fbuf->b_hash_next;
885 *bufp = buf->b_hash_next;
886 buf->b_hash_next = NULL;
887 buf->b_flags &= ~ARC_IN_HASH_TABLE;
889 /* collect some hash table performance data */
890 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
892 if (buf_hash_table.ht_table[idx] &&
893 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
894 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
898 * Global data structures and functions for the buf kmem cache.
900 static kmem_cache_t *hdr_cache;
901 static kmem_cache_t *buf_cache;
902 static kmem_cache_t *l2arc_hdr_cache;
909 #if defined(_KERNEL) && defined(HAVE_SPL)
911 * Large allocations which do not require contiguous pages
912 * should be using vmem_free() in the linux kernel\
914 vmem_free(buf_hash_table.ht_table,
915 (buf_hash_table.ht_mask + 1) * sizeof (void *));
917 kmem_free(buf_hash_table.ht_table,
918 (buf_hash_table.ht_mask + 1) * sizeof (void *));
920 for (i = 0; i < BUF_LOCKS; i++)
921 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
922 kmem_cache_destroy(hdr_cache);
923 kmem_cache_destroy(buf_cache);
924 kmem_cache_destroy(l2arc_hdr_cache);
928 * Constructor callback - called when the cache is empty
929 * and a new buf is requested.
933 hdr_cons(void *vbuf, void *unused, int kmflag)
935 arc_buf_hdr_t *buf = vbuf;
937 bzero(buf, sizeof (arc_buf_hdr_t));
938 refcount_create(&buf->b_refcnt);
939 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
940 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
941 list_link_init(&buf->b_arc_node);
942 list_link_init(&buf->b_l2node);
943 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
950 buf_cons(void *vbuf, void *unused, int kmflag)
952 arc_buf_t *buf = vbuf;
954 bzero(buf, sizeof (arc_buf_t));
955 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
956 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
962 * Destructor callback - called when a cached buf is
963 * no longer required.
967 hdr_dest(void *vbuf, void *unused)
969 arc_buf_hdr_t *buf = vbuf;
971 ASSERT(BUF_EMPTY(buf));
972 refcount_destroy(&buf->b_refcnt);
973 cv_destroy(&buf->b_cv);
974 mutex_destroy(&buf->b_freeze_lock);
975 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
980 buf_dest(void *vbuf, void *unused)
982 arc_buf_t *buf = vbuf;
984 mutex_destroy(&buf->b_evict_lock);
985 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
992 uint64_t hsize = 1ULL << 12;
996 * The hash table is big enough to fill all of physical memory
997 * with an average 64K block size. The table will take up
998 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
1000 while (hsize * 65536 < physmem * PAGESIZE)
1003 buf_hash_table.ht_mask = hsize - 1;
1004 #if defined(_KERNEL) && defined(HAVE_SPL)
1006 * Large allocations which do not require contiguous pages
1007 * should be using vmem_alloc() in the linux kernel
1009 buf_hash_table.ht_table =
1010 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
1012 buf_hash_table.ht_table =
1013 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1015 if (buf_hash_table.ht_table == NULL) {
1016 ASSERT(hsize > (1ULL << 8));
1021 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
1022 0, hdr_cons, hdr_dest, NULL, NULL, NULL, 0);
1023 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1024 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1025 l2arc_hdr_cache = kmem_cache_create("l2arc_buf_hdr_t", L2HDR_SIZE,
1026 0, NULL, NULL, NULL, NULL, NULL, 0);
1028 for (i = 0; i < 256; i++)
1029 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1030 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1032 for (i = 0; i < BUF_LOCKS; i++) {
1033 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1034 NULL, MUTEX_DEFAULT, NULL);
1038 #define ARC_MINTIME (hz>>4) /* 62 ms */
1041 arc_cksum_verify(arc_buf_t *buf)
1045 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1048 mutex_enter(&buf->b_hdr->b_freeze_lock);
1049 if (buf->b_hdr->b_freeze_cksum == NULL ||
1050 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
1051 mutex_exit(&buf->b_hdr->b_freeze_lock);
1054 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1055 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1056 panic("buffer modified while frozen!");
1057 mutex_exit(&buf->b_hdr->b_freeze_lock);
1061 arc_cksum_equal(arc_buf_t *buf)
1066 mutex_enter(&buf->b_hdr->b_freeze_lock);
1067 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1068 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1069 mutex_exit(&buf->b_hdr->b_freeze_lock);
1075 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1077 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1080 mutex_enter(&buf->b_hdr->b_freeze_lock);
1081 if (buf->b_hdr->b_freeze_cksum != NULL) {
1082 mutex_exit(&buf->b_hdr->b_freeze_lock);
1085 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1087 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1088 buf->b_hdr->b_freeze_cksum);
1089 mutex_exit(&buf->b_hdr->b_freeze_lock);
1095 arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
1097 panic("Got SIGSEGV at address: 0x%lx\n", (long) si->si_addr);
1103 arc_buf_unwatch(arc_buf_t *buf)
1107 ASSERT0(mprotect(buf->b_data, buf->b_hdr->b_size,
1108 PROT_READ | PROT_WRITE));
1115 arc_buf_watch(arc_buf_t *buf)
1119 ASSERT0(mprotect(buf->b_data, buf->b_hdr->b_size, PROT_READ));
1124 arc_buf_thaw(arc_buf_t *buf)
1126 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1127 if (buf->b_hdr->b_state != arc_anon)
1128 panic("modifying non-anon buffer!");
1129 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1130 panic("modifying buffer while i/o in progress!");
1131 arc_cksum_verify(buf);
1134 mutex_enter(&buf->b_hdr->b_freeze_lock);
1135 if (buf->b_hdr->b_freeze_cksum != NULL) {
1136 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1137 buf->b_hdr->b_freeze_cksum = NULL;
1140 mutex_exit(&buf->b_hdr->b_freeze_lock);
1142 arc_buf_unwatch(buf);
1146 arc_buf_freeze(arc_buf_t *buf)
1148 kmutex_t *hash_lock;
1150 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1153 hash_lock = HDR_LOCK(buf->b_hdr);
1154 mutex_enter(hash_lock);
1156 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1157 buf->b_hdr->b_state == arc_anon);
1158 arc_cksum_compute(buf, B_FALSE);
1159 mutex_exit(hash_lock);
1164 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1166 ASSERT(MUTEX_HELD(hash_lock));
1168 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1169 (ab->b_state != arc_anon)) {
1170 uint64_t delta = ab->b_size * ab->b_datacnt;
1171 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1172 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1174 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1175 mutex_enter(&ab->b_state->arcs_mtx);
1176 ASSERT(list_link_active(&ab->b_arc_node));
1177 list_remove(list, ab);
1178 if (GHOST_STATE(ab->b_state)) {
1179 ASSERT0(ab->b_datacnt);
1180 ASSERT3P(ab->b_buf, ==, NULL);
1184 ASSERT3U(*size, >=, delta);
1185 atomic_add_64(size, -delta);
1186 mutex_exit(&ab->b_state->arcs_mtx);
1187 /* remove the prefetch flag if we get a reference */
1188 if (ab->b_flags & ARC_PREFETCH)
1189 ab->b_flags &= ~ARC_PREFETCH;
1194 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1197 arc_state_t *state = ab->b_state;
1199 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1200 ASSERT(!GHOST_STATE(state));
1202 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1203 (state != arc_anon)) {
1204 uint64_t *size = &state->arcs_lsize[ab->b_type];
1206 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1207 mutex_enter(&state->arcs_mtx);
1208 ASSERT(!list_link_active(&ab->b_arc_node));
1209 list_insert_head(&state->arcs_list[ab->b_type], ab);
1210 ASSERT(ab->b_datacnt > 0);
1211 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1212 mutex_exit(&state->arcs_mtx);
1218 * Returns detailed information about a specific arc buffer. When the
1219 * state_index argument is set the function will calculate the arc header
1220 * list position for its arc state. Since this requires a linear traversal
1221 * callers are strongly encourage not to do this. However, it can be helpful
1222 * for targeted analysis so the functionality is provided.
1225 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
1227 arc_buf_hdr_t *hdr = ab->b_hdr;
1228 arc_state_t *state = hdr->b_state;
1230 memset(abi, 0, sizeof (arc_buf_info_t));
1231 abi->abi_flags = hdr->b_flags;
1232 abi->abi_datacnt = hdr->b_datacnt;
1233 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
1234 abi->abi_state_contents = hdr->b_type;
1235 abi->abi_state_index = -1;
1236 abi->abi_size = hdr->b_size;
1237 abi->abi_access = hdr->b_arc_access;
1238 abi->abi_mru_hits = hdr->b_mru_hits;
1239 abi->abi_mru_ghost_hits = hdr->b_mru_ghost_hits;
1240 abi->abi_mfu_hits = hdr->b_mfu_hits;
1241 abi->abi_mfu_ghost_hits = hdr->b_mfu_ghost_hits;
1242 abi->abi_holds = refcount_count(&hdr->b_refcnt);
1245 abi->abi_l2arc_dattr = hdr->b_l2hdr->b_daddr;
1246 abi->abi_l2arc_asize = hdr->b_l2hdr->b_asize;
1247 abi->abi_l2arc_compress = hdr->b_l2hdr->b_compress;
1248 abi->abi_l2arc_hits = hdr->b_l2hdr->b_hits;
1251 if (state && state_index && list_link_active(&hdr->b_arc_node)) {
1252 list_t *list = &state->arcs_list[hdr->b_type];
1255 mutex_enter(&state->arcs_mtx);
1256 for (h = list_head(list); h != NULL; h = list_next(list, h)) {
1257 abi->abi_state_index++;
1261 mutex_exit(&state->arcs_mtx);
1266 * Move the supplied buffer to the indicated state. The mutex
1267 * for the buffer must be held by the caller.
1270 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1272 arc_state_t *old_state = ab->b_state;
1273 int64_t refcnt = refcount_count(&ab->b_refcnt);
1274 uint64_t from_delta, to_delta;
1276 ASSERT(MUTEX_HELD(hash_lock));
1277 ASSERT3P(new_state, !=, old_state);
1278 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1279 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1280 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1282 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1285 * If this buffer is evictable, transfer it from the
1286 * old state list to the new state list.
1289 if (old_state != arc_anon) {
1290 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1291 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1294 mutex_enter(&old_state->arcs_mtx);
1296 ASSERT(list_link_active(&ab->b_arc_node));
1297 list_remove(&old_state->arcs_list[ab->b_type], ab);
1300 * If prefetching out of the ghost cache,
1301 * we will have a non-zero datacnt.
1303 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1304 /* ghost elements have a ghost size */
1305 ASSERT(ab->b_buf == NULL);
1306 from_delta = ab->b_size;
1308 ASSERT3U(*size, >=, from_delta);
1309 atomic_add_64(size, -from_delta);
1312 mutex_exit(&old_state->arcs_mtx);
1314 if (new_state != arc_anon) {
1315 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1316 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1319 mutex_enter(&new_state->arcs_mtx);
1321 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1323 /* ghost elements have a ghost size */
1324 if (GHOST_STATE(new_state)) {
1325 ASSERT(ab->b_datacnt == 0);
1326 ASSERT(ab->b_buf == NULL);
1327 to_delta = ab->b_size;
1329 atomic_add_64(size, to_delta);
1332 mutex_exit(&new_state->arcs_mtx);
1336 ASSERT(!BUF_EMPTY(ab));
1337 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1338 buf_hash_remove(ab);
1340 /* adjust state sizes */
1342 atomic_add_64(&new_state->arcs_size, to_delta);
1344 ASSERT3U(old_state->arcs_size, >=, from_delta);
1345 atomic_add_64(&old_state->arcs_size, -from_delta);
1347 ab->b_state = new_state;
1349 /* adjust l2arc hdr stats */
1350 if (new_state == arc_l2c_only)
1351 l2arc_hdr_stat_add();
1352 else if (old_state == arc_l2c_only)
1353 l2arc_hdr_stat_remove();
1357 arc_space_consume(uint64_t space, arc_space_type_t type)
1359 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1364 case ARC_SPACE_DATA:
1365 ARCSTAT_INCR(arcstat_data_size, space);
1367 case ARC_SPACE_OTHER:
1368 ARCSTAT_INCR(arcstat_other_size, space);
1370 case ARC_SPACE_HDRS:
1371 ARCSTAT_INCR(arcstat_hdr_size, space);
1373 case ARC_SPACE_L2HDRS:
1374 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1378 ARCSTAT_INCR(arcstat_meta_used, space);
1379 atomic_add_64(&arc_size, space);
1383 arc_space_return(uint64_t space, arc_space_type_t type)
1385 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1390 case ARC_SPACE_DATA:
1391 ARCSTAT_INCR(arcstat_data_size, -space);
1393 case ARC_SPACE_OTHER:
1394 ARCSTAT_INCR(arcstat_other_size, -space);
1396 case ARC_SPACE_HDRS:
1397 ARCSTAT_INCR(arcstat_hdr_size, -space);
1399 case ARC_SPACE_L2HDRS:
1400 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1404 ASSERT(arc_meta_used >= space);
1405 if (arc_meta_max < arc_meta_used)
1406 arc_meta_max = arc_meta_used;
1407 ARCSTAT_INCR(arcstat_meta_used, -space);
1408 ASSERT(arc_size >= space);
1409 atomic_add_64(&arc_size, -space);
1413 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1418 ASSERT3U(size, >, 0);
1419 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1420 ASSERT(BUF_EMPTY(hdr));
1423 hdr->b_spa = spa_load_guid(spa);
1424 hdr->b_state = arc_anon;
1425 hdr->b_arc_access = 0;
1426 hdr->b_mru_hits = 0;
1427 hdr->b_mru_ghost_hits = 0;
1428 hdr->b_mfu_hits = 0;
1429 hdr->b_mfu_ghost_hits = 0;
1431 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1434 buf->b_efunc = NULL;
1435 buf->b_private = NULL;
1438 arc_get_data_buf(buf);
1441 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1442 (void) refcount_add(&hdr->b_refcnt, tag);
1447 static char *arc_onloan_tag = "onloan";
1450 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1451 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1452 * buffers must be returned to the arc before they can be used by the DMU or
1456 arc_loan_buf(spa_t *spa, int size)
1460 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1462 atomic_add_64(&arc_loaned_bytes, size);
1467 * Return a loaned arc buffer to the arc.
1470 arc_return_buf(arc_buf_t *buf, void *tag)
1472 arc_buf_hdr_t *hdr = buf->b_hdr;
1474 ASSERT(buf->b_data != NULL);
1475 (void) refcount_add(&hdr->b_refcnt, tag);
1476 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1478 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1481 /* Detach an arc_buf from a dbuf (tag) */
1483 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1487 ASSERT(buf->b_data != NULL);
1489 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1490 (void) refcount_remove(&hdr->b_refcnt, tag);
1491 buf->b_efunc = NULL;
1492 buf->b_private = NULL;
1494 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1498 arc_buf_clone(arc_buf_t *from)
1501 arc_buf_hdr_t *hdr = from->b_hdr;
1502 uint64_t size = hdr->b_size;
1504 ASSERT(hdr->b_state != arc_anon);
1506 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1509 buf->b_efunc = NULL;
1510 buf->b_private = NULL;
1511 buf->b_next = hdr->b_buf;
1513 arc_get_data_buf(buf);
1514 bcopy(from->b_data, buf->b_data, size);
1517 * This buffer already exists in the arc so create a duplicate
1518 * copy for the caller. If the buffer is associated with user data
1519 * then track the size and number of duplicates. These stats will be
1520 * updated as duplicate buffers are created and destroyed.
1522 if (hdr->b_type == ARC_BUFC_DATA) {
1523 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1524 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1526 hdr->b_datacnt += 1;
1531 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1534 kmutex_t *hash_lock;
1537 * Check to see if this buffer is evicted. Callers
1538 * must verify b_data != NULL to know if the add_ref
1541 mutex_enter(&buf->b_evict_lock);
1542 if (buf->b_data == NULL) {
1543 mutex_exit(&buf->b_evict_lock);
1546 hash_lock = HDR_LOCK(buf->b_hdr);
1547 mutex_enter(hash_lock);
1549 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1550 mutex_exit(&buf->b_evict_lock);
1552 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1553 add_reference(hdr, hash_lock, tag);
1554 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1555 arc_access(hdr, hash_lock);
1556 mutex_exit(hash_lock);
1557 ARCSTAT_BUMP(arcstat_hits);
1558 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1559 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1560 data, metadata, hits);
1564 * Free the arc data buffer. If it is an l2arc write in progress,
1565 * the buffer is placed on l2arc_free_on_write to be freed later.
1568 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1570 arc_buf_hdr_t *hdr = buf->b_hdr;
1572 if (HDR_L2_WRITING(hdr)) {
1573 l2arc_data_free_t *df;
1574 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_PUSHPAGE);
1575 df->l2df_data = buf->b_data;
1576 df->l2df_size = hdr->b_size;
1577 df->l2df_func = free_func;
1578 mutex_enter(&l2arc_free_on_write_mtx);
1579 list_insert_head(l2arc_free_on_write, df);
1580 mutex_exit(&l2arc_free_on_write_mtx);
1581 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1583 free_func(buf->b_data, hdr->b_size);
1588 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1592 /* free up data associated with the buf */
1594 arc_state_t *state = buf->b_hdr->b_state;
1595 uint64_t size = buf->b_hdr->b_size;
1596 arc_buf_contents_t type = buf->b_hdr->b_type;
1598 arc_cksum_verify(buf);
1599 arc_buf_unwatch(buf);
1602 if (type == ARC_BUFC_METADATA) {
1603 arc_buf_data_free(buf, zio_buf_free);
1604 arc_space_return(size, ARC_SPACE_DATA);
1606 ASSERT(type == ARC_BUFC_DATA);
1607 arc_buf_data_free(buf, zio_data_buf_free);
1608 ARCSTAT_INCR(arcstat_data_size, -size);
1609 atomic_add_64(&arc_size, -size);
1612 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1613 uint64_t *cnt = &state->arcs_lsize[type];
1615 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1616 ASSERT(state != arc_anon);
1618 ASSERT3U(*cnt, >=, size);
1619 atomic_add_64(cnt, -size);
1621 ASSERT3U(state->arcs_size, >=, size);
1622 atomic_add_64(&state->arcs_size, -size);
1626 * If we're destroying a duplicate buffer make sure
1627 * that the appropriate statistics are updated.
1629 if (buf->b_hdr->b_datacnt > 1 &&
1630 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1631 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1632 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1634 ASSERT(buf->b_hdr->b_datacnt > 0);
1635 buf->b_hdr->b_datacnt -= 1;
1638 /* only remove the buf if requested */
1642 /* remove the buf from the hdr list */
1643 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1645 *bufp = buf->b_next;
1648 ASSERT(buf->b_efunc == NULL);
1650 /* clean up the buf */
1652 kmem_cache_free(buf_cache, buf);
1656 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1658 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1660 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1661 ASSERT3P(hdr->b_state, ==, arc_anon);
1662 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1664 if (l2hdr != NULL) {
1665 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1667 * To prevent arc_free() and l2arc_evict() from
1668 * attempting to free the same buffer at the same time,
1669 * a FREE_IN_PROGRESS flag is given to arc_free() to
1670 * give it priority. l2arc_evict() can't destroy this
1671 * header while we are waiting on l2arc_buflist_mtx.
1673 * The hdr may be removed from l2ad_buflist before we
1674 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1676 if (!buflist_held) {
1677 mutex_enter(&l2arc_buflist_mtx);
1678 l2hdr = hdr->b_l2hdr;
1681 if (l2hdr != NULL) {
1682 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1683 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1684 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
1685 kmem_cache_free(l2arc_hdr_cache, l2hdr);
1686 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
1687 if (hdr->b_state == arc_l2c_only)
1688 l2arc_hdr_stat_remove();
1689 hdr->b_l2hdr = NULL;
1693 mutex_exit(&l2arc_buflist_mtx);
1696 if (!BUF_EMPTY(hdr)) {
1697 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1698 buf_discard_identity(hdr);
1700 while (hdr->b_buf) {
1701 arc_buf_t *buf = hdr->b_buf;
1704 mutex_enter(&arc_eviction_mtx);
1705 mutex_enter(&buf->b_evict_lock);
1706 ASSERT(buf->b_hdr != NULL);
1707 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1708 hdr->b_buf = buf->b_next;
1709 buf->b_hdr = &arc_eviction_hdr;
1710 buf->b_next = arc_eviction_list;
1711 arc_eviction_list = buf;
1712 mutex_exit(&buf->b_evict_lock);
1713 mutex_exit(&arc_eviction_mtx);
1715 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1718 if (hdr->b_freeze_cksum != NULL) {
1719 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1720 hdr->b_freeze_cksum = NULL;
1723 ASSERT(!list_link_active(&hdr->b_arc_node));
1724 ASSERT3P(hdr->b_hash_next, ==, NULL);
1725 ASSERT3P(hdr->b_acb, ==, NULL);
1726 kmem_cache_free(hdr_cache, hdr);
1730 arc_buf_free(arc_buf_t *buf, void *tag)
1732 arc_buf_hdr_t *hdr = buf->b_hdr;
1733 int hashed = hdr->b_state != arc_anon;
1735 ASSERT(buf->b_efunc == NULL);
1736 ASSERT(buf->b_data != NULL);
1739 kmutex_t *hash_lock = HDR_LOCK(hdr);
1741 mutex_enter(hash_lock);
1743 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1745 (void) remove_reference(hdr, hash_lock, tag);
1746 if (hdr->b_datacnt > 1) {
1747 arc_buf_destroy(buf, FALSE, TRUE);
1749 ASSERT(buf == hdr->b_buf);
1750 ASSERT(buf->b_efunc == NULL);
1751 hdr->b_flags |= ARC_BUF_AVAILABLE;
1753 mutex_exit(hash_lock);
1754 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1757 * We are in the middle of an async write. Don't destroy
1758 * this buffer unless the write completes before we finish
1759 * decrementing the reference count.
1761 mutex_enter(&arc_eviction_mtx);
1762 (void) remove_reference(hdr, NULL, tag);
1763 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1764 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1765 mutex_exit(&arc_eviction_mtx);
1767 arc_hdr_destroy(hdr);
1769 if (remove_reference(hdr, NULL, tag) > 0)
1770 arc_buf_destroy(buf, FALSE, TRUE);
1772 arc_hdr_destroy(hdr);
1777 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1779 arc_buf_hdr_t *hdr = buf->b_hdr;
1780 kmutex_t *hash_lock = NULL;
1781 boolean_t no_callback = (buf->b_efunc == NULL);
1783 if (hdr->b_state == arc_anon) {
1784 ASSERT(hdr->b_datacnt == 1);
1785 arc_buf_free(buf, tag);
1786 return (no_callback);
1789 hash_lock = HDR_LOCK(hdr);
1790 mutex_enter(hash_lock);
1792 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1793 ASSERT(hdr->b_state != arc_anon);
1794 ASSERT(buf->b_data != NULL);
1796 (void) remove_reference(hdr, hash_lock, tag);
1797 if (hdr->b_datacnt > 1) {
1799 arc_buf_destroy(buf, FALSE, TRUE);
1800 } else if (no_callback) {
1801 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1802 ASSERT(buf->b_efunc == NULL);
1803 hdr->b_flags |= ARC_BUF_AVAILABLE;
1805 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1806 refcount_is_zero(&hdr->b_refcnt));
1807 mutex_exit(hash_lock);
1808 return (no_callback);
1812 arc_buf_size(arc_buf_t *buf)
1814 return (buf->b_hdr->b_size);
1818 * Called from the DMU to determine if the current buffer should be
1819 * evicted. In order to ensure proper locking, the eviction must be initiated
1820 * from the DMU. Return true if the buffer is associated with user data and
1821 * duplicate buffers still exist.
1824 arc_buf_eviction_needed(arc_buf_t *buf)
1827 boolean_t evict_needed = B_FALSE;
1829 if (zfs_disable_dup_eviction)
1832 mutex_enter(&buf->b_evict_lock);
1836 * We are in arc_do_user_evicts(); let that function
1837 * perform the eviction.
1839 ASSERT(buf->b_data == NULL);
1840 mutex_exit(&buf->b_evict_lock);
1842 } else if (buf->b_data == NULL) {
1844 * We have already been added to the arc eviction list;
1845 * recommend eviction.
1847 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1848 mutex_exit(&buf->b_evict_lock);
1852 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1853 evict_needed = B_TRUE;
1855 mutex_exit(&buf->b_evict_lock);
1856 return (evict_needed);
1860 * Evict buffers from list until we've removed the specified number of
1861 * bytes. Move the removed buffers to the appropriate evict state.
1862 * If the recycle flag is set, then attempt to "recycle" a buffer:
1863 * - look for a buffer to evict that is `bytes' long.
1864 * - return the data block from this buffer rather than freeing it.
1865 * This flag is used by callers that are trying to make space for a
1866 * new buffer in a full arc cache.
1868 * This function makes a "best effort". It skips over any buffers
1869 * it can't get a hash_lock on, and so may not catch all candidates.
1870 * It may also return without evicting as much space as requested.
1873 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1874 arc_buf_contents_t type)
1876 arc_state_t *evicted_state;
1877 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1878 arc_buf_hdr_t *ab, *ab_prev = NULL;
1879 list_t *list = &state->arcs_list[type];
1880 kmutex_t *hash_lock;
1881 boolean_t have_lock;
1882 void *stolen = NULL;
1883 arc_buf_hdr_t marker = {{{ 0 }}};
1886 ASSERT(state == arc_mru || state == arc_mfu);
1888 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1890 mutex_enter(&state->arcs_mtx);
1891 mutex_enter(&evicted_state->arcs_mtx);
1893 for (ab = list_tail(list); ab; ab = ab_prev) {
1894 ab_prev = list_prev(list, ab);
1895 /* prefetch buffers have a minimum lifespan */
1896 if (HDR_IO_IN_PROGRESS(ab) ||
1897 (spa && ab->b_spa != spa) ||
1898 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1899 ddi_get_lbolt() - ab->b_arc_access <
1900 zfs_arc_min_prefetch_lifespan)) {
1904 /* "lookahead" for better eviction candidate */
1905 if (recycle && ab->b_size != bytes &&
1906 ab_prev && ab_prev->b_size == bytes)
1909 /* ignore markers */
1914 * It may take a long time to evict all the bufs requested.
1915 * To avoid blocking all arc activity, periodically drop
1916 * the arcs_mtx and give other threads a chance to run
1917 * before reacquiring the lock.
1919 * If we are looking for a buffer to recycle, we are in
1920 * the hot code path, so don't sleep.
1922 if (!recycle && count++ > arc_evict_iterations) {
1923 list_insert_after(list, ab, &marker);
1924 mutex_exit(&evicted_state->arcs_mtx);
1925 mutex_exit(&state->arcs_mtx);
1926 kpreempt(KPREEMPT_SYNC);
1927 mutex_enter(&state->arcs_mtx);
1928 mutex_enter(&evicted_state->arcs_mtx);
1929 ab_prev = list_prev(list, &marker);
1930 list_remove(list, &marker);
1935 hash_lock = HDR_LOCK(ab);
1936 have_lock = MUTEX_HELD(hash_lock);
1937 if (have_lock || mutex_tryenter(hash_lock)) {
1938 ASSERT0(refcount_count(&ab->b_refcnt));
1939 ASSERT(ab->b_datacnt > 0);
1941 arc_buf_t *buf = ab->b_buf;
1942 if (!mutex_tryenter(&buf->b_evict_lock)) {
1947 bytes_evicted += ab->b_size;
1948 if (recycle && ab->b_type == type &&
1949 ab->b_size == bytes &&
1950 !HDR_L2_WRITING(ab)) {
1951 stolen = buf->b_data;
1956 mutex_enter(&arc_eviction_mtx);
1957 arc_buf_destroy(buf,
1958 buf->b_data == stolen, FALSE);
1959 ab->b_buf = buf->b_next;
1960 buf->b_hdr = &arc_eviction_hdr;
1961 buf->b_next = arc_eviction_list;
1962 arc_eviction_list = buf;
1963 mutex_exit(&arc_eviction_mtx);
1964 mutex_exit(&buf->b_evict_lock);
1966 mutex_exit(&buf->b_evict_lock);
1967 arc_buf_destroy(buf,
1968 buf->b_data == stolen, TRUE);
1973 ARCSTAT_INCR(arcstat_evict_l2_cached,
1976 if (l2arc_write_eligible(ab->b_spa, ab)) {
1977 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1981 arcstat_evict_l2_ineligible,
1986 if (ab->b_datacnt == 0) {
1987 arc_change_state(evicted_state, ab, hash_lock);
1988 ASSERT(HDR_IN_HASH_TABLE(ab));
1989 ab->b_flags |= ARC_IN_HASH_TABLE;
1990 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1991 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1994 mutex_exit(hash_lock);
1995 if (bytes >= 0 && bytes_evicted >= bytes)
2002 mutex_exit(&evicted_state->arcs_mtx);
2003 mutex_exit(&state->arcs_mtx);
2005 if (bytes_evicted < bytes)
2006 dprintf("only evicted %lld bytes from %x\n",
2007 (longlong_t)bytes_evicted, state);
2010 ARCSTAT_INCR(arcstat_evict_skip, skipped);
2013 ARCSTAT_INCR(arcstat_mutex_miss, missed);
2016 * Note: we have just evicted some data into the ghost state,
2017 * potentially putting the ghost size over the desired size. Rather
2018 * that evicting from the ghost list in this hot code path, leave
2019 * this chore to the arc_reclaim_thread().
2026 * Remove buffers from list until we've removed the specified number of
2027 * bytes. Destroy the buffers that are removed.
2030 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes,
2031 arc_buf_contents_t type)
2033 arc_buf_hdr_t *ab, *ab_prev;
2034 arc_buf_hdr_t marker;
2035 list_t *list = &state->arcs_list[type];
2036 kmutex_t *hash_lock;
2037 uint64_t bytes_deleted = 0;
2038 uint64_t bufs_skipped = 0;
2041 ASSERT(GHOST_STATE(state));
2042 bzero(&marker, sizeof (marker));
2044 mutex_enter(&state->arcs_mtx);
2045 for (ab = list_tail(list); ab; ab = ab_prev) {
2046 ab_prev = list_prev(list, ab);
2047 if (ab->b_type > ARC_BUFC_NUMTYPES)
2048 panic("invalid ab=%p", (void *)ab);
2049 if (spa && ab->b_spa != spa)
2052 /* ignore markers */
2056 hash_lock = HDR_LOCK(ab);
2057 /* caller may be trying to modify this buffer, skip it */
2058 if (MUTEX_HELD(hash_lock))
2062 * It may take a long time to evict all the bufs requested.
2063 * To avoid blocking all arc activity, periodically drop
2064 * the arcs_mtx and give other threads a chance to run
2065 * before reacquiring the lock.
2067 if (count++ > arc_evict_iterations) {
2068 list_insert_after(list, ab, &marker);
2069 mutex_exit(&state->arcs_mtx);
2070 kpreempt(KPREEMPT_SYNC);
2071 mutex_enter(&state->arcs_mtx);
2072 ab_prev = list_prev(list, &marker);
2073 list_remove(list, &marker);
2077 if (mutex_tryenter(hash_lock)) {
2078 ASSERT(!HDR_IO_IN_PROGRESS(ab));
2079 ASSERT(ab->b_buf == NULL);
2080 ARCSTAT_BUMP(arcstat_deleted);
2081 bytes_deleted += ab->b_size;
2083 if (ab->b_l2hdr != NULL) {
2085 * This buffer is cached on the 2nd Level ARC;
2086 * don't destroy the header.
2088 arc_change_state(arc_l2c_only, ab, hash_lock);
2089 mutex_exit(hash_lock);
2091 arc_change_state(arc_anon, ab, hash_lock);
2092 mutex_exit(hash_lock);
2093 arc_hdr_destroy(ab);
2096 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
2097 if (bytes >= 0 && bytes_deleted >= bytes)
2099 } else if (bytes < 0) {
2101 * Insert a list marker and then wait for the
2102 * hash lock to become available. Once its
2103 * available, restart from where we left off.
2105 list_insert_after(list, ab, &marker);
2106 mutex_exit(&state->arcs_mtx);
2107 mutex_enter(hash_lock);
2108 mutex_exit(hash_lock);
2109 mutex_enter(&state->arcs_mtx);
2110 ab_prev = list_prev(list, &marker);
2111 list_remove(list, &marker);
2116 mutex_exit(&state->arcs_mtx);
2118 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
2119 (bytes < 0 || bytes_deleted < bytes)) {
2120 list = &state->arcs_list[ARC_BUFC_METADATA];
2125 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2129 if (bytes_deleted < bytes)
2130 dprintf("only deleted %lld bytes from %p\n",
2131 (longlong_t)bytes_deleted, state);
2137 int64_t adjustment, delta;
2143 adjustment = MIN((int64_t)(arc_size - arc_c),
2144 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2147 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2148 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2149 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
2150 adjustment -= delta;
2153 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2154 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2155 (void) arc_evict(arc_mru, 0, delta, FALSE,
2163 adjustment = arc_size - arc_c;
2165 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2166 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2167 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2168 adjustment -= delta;
2171 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2172 int64_t delta = MIN(adjustment,
2173 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2174 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2179 * Adjust ghost lists
2182 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2184 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2185 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2186 arc_evict_ghost(arc_mru_ghost, 0, delta, ARC_BUFC_DATA);
2190 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2192 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2193 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2194 arc_evict_ghost(arc_mfu_ghost, 0, delta, ARC_BUFC_DATA);
2199 * Request that arc user drop references so that N bytes can be released
2200 * from the cache. This provides a mechanism to ensure the arc can honor
2201 * the arc_meta_limit and reclaim buffers which are pinned in the cache
2202 * by higher layers. (i.e. the zpl)
2205 arc_do_user_prune(int64_t adjustment)
2207 arc_prune_func_t *func;
2209 arc_prune_t *cp, *np;
2211 mutex_enter(&arc_prune_mtx);
2213 cp = list_head(&arc_prune_list);
2214 while (cp != NULL) {
2216 private = cp->p_private;
2217 np = list_next(&arc_prune_list, cp);
2218 refcount_add(&cp->p_refcnt, func);
2219 mutex_exit(&arc_prune_mtx);
2222 func(adjustment, private);
2224 mutex_enter(&arc_prune_mtx);
2226 /* User removed prune callback concurrently with execution */
2227 if (refcount_remove(&cp->p_refcnt, func) == 0) {
2228 ASSERT(!list_link_active(&cp->p_node));
2229 refcount_destroy(&cp->p_refcnt);
2230 kmem_free(cp, sizeof (*cp));
2236 ARCSTAT_BUMP(arcstat_prune);
2237 mutex_exit(&arc_prune_mtx);
2241 arc_do_user_evicts(void)
2243 mutex_enter(&arc_eviction_mtx);
2244 while (arc_eviction_list != NULL) {
2245 arc_buf_t *buf = arc_eviction_list;
2246 arc_eviction_list = buf->b_next;
2247 mutex_enter(&buf->b_evict_lock);
2249 mutex_exit(&buf->b_evict_lock);
2250 mutex_exit(&arc_eviction_mtx);
2252 if (buf->b_efunc != NULL)
2253 VERIFY(buf->b_efunc(buf) == 0);
2255 buf->b_efunc = NULL;
2256 buf->b_private = NULL;
2257 kmem_cache_free(buf_cache, buf);
2258 mutex_enter(&arc_eviction_mtx);
2260 mutex_exit(&arc_eviction_mtx);
2264 * Evict only meta data objects from the cache leaving the data objects.
2265 * This is only used to enforce the tunable arc_meta_limit, if we are
2266 * unable to evict enough buffers notify the user via the prune callback.
2269 arc_adjust_meta(int64_t adjustment, boolean_t may_prune)
2273 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2274 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2275 arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_METADATA);
2276 adjustment -= delta;
2279 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2280 delta = MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2281 arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_METADATA);
2282 adjustment -= delta;
2285 if (may_prune && (adjustment > 0) && (arc_meta_used > arc_meta_limit))
2286 arc_do_user_prune(zfs_arc_meta_prune);
2290 * Flush all *evictable* data from the cache for the given spa.
2291 * NOTE: this will not touch "active" (i.e. referenced) data.
2294 arc_flush(spa_t *spa)
2299 guid = spa_load_guid(spa);
2301 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2302 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2306 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2307 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2311 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2312 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2316 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2317 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2322 arc_evict_ghost(arc_mru_ghost, guid, -1, ARC_BUFC_DATA);
2323 arc_evict_ghost(arc_mfu_ghost, guid, -1, ARC_BUFC_DATA);
2325 mutex_enter(&arc_reclaim_thr_lock);
2326 arc_do_user_evicts();
2327 mutex_exit(&arc_reclaim_thr_lock);
2328 ASSERT(spa || arc_eviction_list == NULL);
2332 arc_shrink(uint64_t bytes)
2334 if (arc_c > arc_c_min) {
2337 to_free = bytes ? bytes : arc_c >> zfs_arc_shrink_shift;
2339 if (arc_c > arc_c_min + to_free)
2340 atomic_add_64(&arc_c, -to_free);
2344 to_free = bytes ? bytes : arc_p >> zfs_arc_shrink_shift;
2346 if (arc_p > to_free)
2347 atomic_add_64(&arc_p, -to_free);
2351 if (arc_c > arc_size)
2352 arc_c = MAX(arc_size, arc_c_min);
2354 arc_p = (arc_c >> 1);
2355 ASSERT(arc_c >= arc_c_min);
2356 ASSERT((int64_t)arc_p >= 0);
2359 if (arc_size > arc_c)
2364 arc_kmem_reap_now(arc_reclaim_strategy_t strat, uint64_t bytes)
2367 kmem_cache_t *prev_cache = NULL;
2368 kmem_cache_t *prev_data_cache = NULL;
2369 extern kmem_cache_t *zio_buf_cache[];
2370 extern kmem_cache_t *zio_data_buf_cache[];
2373 * An aggressive reclamation will shrink the cache size as well as
2374 * reap free buffers from the arc kmem caches.
2376 if (strat == ARC_RECLAIM_AGGR)
2379 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2380 if (zio_buf_cache[i] != prev_cache) {
2381 prev_cache = zio_buf_cache[i];
2382 kmem_cache_reap_now(zio_buf_cache[i]);
2384 if (zio_data_buf_cache[i] != prev_data_cache) {
2385 prev_data_cache = zio_data_buf_cache[i];
2386 kmem_cache_reap_now(zio_data_buf_cache[i]);
2390 kmem_cache_reap_now(buf_cache);
2391 kmem_cache_reap_now(hdr_cache);
2395 * Unlike other ZFS implementations this thread is only responsible for
2396 * adapting the target ARC size on Linux. The responsibility for memory
2397 * reclamation has been entirely delegated to the arc_shrinker_func()
2398 * which is registered with the VM. To reflect this change in behavior
2399 * the arc_reclaim thread has been renamed to arc_adapt.
2402 arc_adapt_thread(void)
2407 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2409 mutex_enter(&arc_reclaim_thr_lock);
2410 while (arc_thread_exit == 0) {
2412 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2414 if (spa_get_random(100) == 0) {
2417 if (last_reclaim == ARC_RECLAIM_CONS) {
2418 last_reclaim = ARC_RECLAIM_AGGR;
2420 last_reclaim = ARC_RECLAIM_CONS;
2424 last_reclaim = ARC_RECLAIM_AGGR;
2428 /* reset the growth delay for every reclaim */
2429 arc_grow_time = ddi_get_lbolt() +
2430 (zfs_arc_grow_retry * hz);
2432 arc_kmem_reap_now(last_reclaim, 0);
2435 #endif /* !_KERNEL */
2437 /* No recent memory pressure allow the ARC to grow. */
2438 if (arc_no_grow && ddi_get_lbolt() >= arc_grow_time)
2439 arc_no_grow = FALSE;
2442 * Keep meta data usage within limits, arc_shrink() is not
2443 * used to avoid collapsing the arc_c value when only the
2444 * arc_meta_limit is being exceeded.
2446 prune = (int64_t)arc_meta_used - (int64_t)arc_meta_limit;
2448 arc_adjust_meta(prune, B_TRUE);
2452 if (arc_eviction_list != NULL)
2453 arc_do_user_evicts();
2455 /* block until needed, or one second, whichever is shorter */
2456 CALLB_CPR_SAFE_BEGIN(&cpr);
2457 (void) cv_timedwait_interruptible(&arc_reclaim_thr_cv,
2458 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2459 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2462 /* Allow the module options to be changed */
2463 if (zfs_arc_max > 64 << 20 &&
2464 zfs_arc_max < physmem * PAGESIZE &&
2465 zfs_arc_max != arc_c_max)
2466 arc_c_max = zfs_arc_max;
2468 if (zfs_arc_min > 0 &&
2469 zfs_arc_min < arc_c_max &&
2470 zfs_arc_min != arc_c_min)
2471 arc_c_min = zfs_arc_min;
2473 if (zfs_arc_meta_limit > 0 &&
2474 zfs_arc_meta_limit <= arc_c_max &&
2475 zfs_arc_meta_limit != arc_meta_limit)
2476 arc_meta_limit = zfs_arc_meta_limit;
2482 arc_thread_exit = 0;
2483 cv_broadcast(&arc_reclaim_thr_cv);
2484 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2490 * Determine the amount of memory eligible for eviction contained in the
2491 * ARC. All clean data reported by the ghost lists can always be safely
2492 * evicted. Due to arc_c_min, the same does not hold for all clean data
2493 * contained by the regular mru and mfu lists.
2495 * In the case of the regular mru and mfu lists, we need to report as
2496 * much clean data as possible, such that evicting that same reported
2497 * data will not bring arc_size below arc_c_min. Thus, in certain
2498 * circumstances, the total amount of clean data in the mru and mfu
2499 * lists might not actually be evictable.
2501 * The following two distinct cases are accounted for:
2503 * 1. The sum of the amount of dirty data contained by both the mru and
2504 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2505 * is greater than or equal to arc_c_min.
2506 * (i.e. amount of dirty data >= arc_c_min)
2508 * This is the easy case; all clean data contained by the mru and mfu
2509 * lists is evictable. Evicting all clean data can only drop arc_size
2510 * to the amount of dirty data, which is greater than arc_c_min.
2512 * 2. The sum of the amount of dirty data contained by both the mru and
2513 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2514 * is less than arc_c_min.
2515 * (i.e. arc_c_min > amount of dirty data)
2517 * 2.1. arc_size is greater than or equal arc_c_min.
2518 * (i.e. arc_size >= arc_c_min > amount of dirty data)
2520 * In this case, not all clean data from the regular mru and mfu
2521 * lists is actually evictable; we must leave enough clean data
2522 * to keep arc_size above arc_c_min. Thus, the maximum amount of
2523 * evictable data from the two lists combined, is exactly the
2524 * difference between arc_size and arc_c_min.
2526 * 2.2. arc_size is less than arc_c_min
2527 * (i.e. arc_c_min > arc_size > amount of dirty data)
2529 * In this case, none of the data contained in the mru and mfu
2530 * lists is evictable, even if it's clean. Since arc_size is
2531 * already below arc_c_min, evicting any more would only
2532 * increase this negative difference.
2535 arc_evictable_memory(void) {
2536 uint64_t arc_clean =
2537 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
2538 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
2539 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
2540 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
2541 uint64_t ghost_clean =
2542 arc_mru_ghost->arcs_lsize[ARC_BUFC_DATA] +
2543 arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] +
2544 arc_mfu_ghost->arcs_lsize[ARC_BUFC_DATA] +
2545 arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA];
2546 uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0);
2548 if (arc_dirty >= arc_c_min)
2549 return (ghost_clean + arc_clean);
2551 return (ghost_clean + MAX((int64_t)arc_size - (int64_t)arc_c_min, 0));
2555 __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc)
2559 /* The arc is considered warm once reclaim has occurred */
2560 if (unlikely(arc_warm == B_FALSE))
2563 /* Return the potential number of reclaimable pages */
2564 pages = btop(arc_evictable_memory());
2565 if (sc->nr_to_scan == 0)
2568 /* Not allowed to perform filesystem reclaim */
2569 if (!(sc->gfp_mask & __GFP_FS))
2572 /* Reclaim in progress */
2573 if (mutex_tryenter(&arc_reclaim_thr_lock) == 0)
2577 * Evict the requested number of pages by shrinking arc_c the
2578 * requested amount. If there is nothing left to evict just
2579 * reap whatever we can from the various arc slabs.
2582 arc_kmem_reap_now(ARC_RECLAIM_AGGR, ptob(sc->nr_to_scan));
2584 arc_kmem_reap_now(ARC_RECLAIM_CONS, ptob(sc->nr_to_scan));
2588 * When direct reclaim is observed it usually indicates a rapid
2589 * increase in memory pressure. This occurs because the kswapd
2590 * threads were unable to asynchronously keep enough free memory
2591 * available. In this case set arc_no_grow to briefly pause arc
2592 * growth to avoid compounding the memory pressure.
2594 if (current_is_kswapd()) {
2595 ARCSTAT_BUMP(arcstat_memory_indirect_count);
2597 arc_no_grow = B_TRUE;
2598 arc_grow_time = ddi_get_lbolt() + (zfs_arc_grow_retry * hz);
2599 ARCSTAT_BUMP(arcstat_memory_direct_count);
2602 mutex_exit(&arc_reclaim_thr_lock);
2606 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func);
2608 SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS);
2609 #endif /* _KERNEL */
2612 * Adapt arc info given the number of bytes we are trying to add and
2613 * the state that we are comming from. This function is only called
2614 * when we are adding new content to the cache.
2617 arc_adapt(int bytes, arc_state_t *state)
2621 if (state == arc_l2c_only)
2626 * Adapt the target size of the MRU list:
2627 * - if we just hit in the MRU ghost list, then increase
2628 * the target size of the MRU list.
2629 * - if we just hit in the MFU ghost list, then increase
2630 * the target size of the MFU list by decreasing the
2631 * target size of the MRU list.
2633 if (state == arc_mru_ghost) {
2634 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2635 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2636 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2638 arc_p = MIN(arc_c, arc_p + bytes * mult);
2639 } else if (state == arc_mfu_ghost) {
2642 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2643 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2644 mult = MIN(mult, 10);
2646 delta = MIN(bytes * mult, arc_p);
2647 arc_p = MAX(0, arc_p - delta);
2649 ASSERT((int64_t)arc_p >= 0);
2654 if (arc_c >= arc_c_max)
2658 * If we're within (2 * maxblocksize) bytes of the target
2659 * cache size, increment the target cache size
2661 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2662 atomic_add_64(&arc_c, (int64_t)bytes);
2663 if (arc_c > arc_c_max)
2665 else if (state == arc_anon)
2666 atomic_add_64(&arc_p, (int64_t)bytes);
2670 ASSERT((int64_t)arc_p >= 0);
2674 * Check if the cache has reached its limits and eviction is required
2678 arc_evict_needed(arc_buf_contents_t type)
2680 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2686 return (arc_size > arc_c);
2690 * The buffer, supplied as the first argument, needs a data block.
2691 * So, if we are at cache max, determine which cache should be victimized.
2692 * We have the following cases:
2694 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2695 * In this situation if we're out of space, but the resident size of the MFU is
2696 * under the limit, victimize the MFU cache to satisfy this insertion request.
2698 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2699 * Here, we've used up all of the available space for the MRU, so we need to
2700 * evict from our own cache instead. Evict from the set of resident MRU
2703 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2704 * c minus p represents the MFU space in the cache, since p is the size of the
2705 * cache that is dedicated to the MRU. In this situation there's still space on
2706 * the MFU side, so the MRU side needs to be victimized.
2708 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2709 * MFU's resident set is consuming more space than it has been allotted. In
2710 * this situation, we must victimize our own cache, the MFU, for this insertion.
2713 arc_get_data_buf(arc_buf_t *buf)
2715 arc_state_t *state = buf->b_hdr->b_state;
2716 uint64_t size = buf->b_hdr->b_size;
2717 arc_buf_contents_t type = buf->b_hdr->b_type;
2719 arc_adapt(size, state);
2722 * We have not yet reached cache maximum size,
2723 * just allocate a new buffer.
2725 if (!arc_evict_needed(type)) {
2726 if (type == ARC_BUFC_METADATA) {
2727 buf->b_data = zio_buf_alloc(size);
2728 arc_space_consume(size, ARC_SPACE_DATA);
2730 ASSERT(type == ARC_BUFC_DATA);
2731 buf->b_data = zio_data_buf_alloc(size);
2732 ARCSTAT_INCR(arcstat_data_size, size);
2733 atomic_add_64(&arc_size, size);
2739 * If we are prefetching from the mfu ghost list, this buffer
2740 * will end up on the mru list; so steal space from there.
2742 if (state == arc_mfu_ghost)
2743 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2744 else if (state == arc_mru_ghost)
2747 if (state == arc_mru || state == arc_anon) {
2748 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2749 state = (arc_mfu->arcs_lsize[type] >= size &&
2750 arc_p > mru_used) ? arc_mfu : arc_mru;
2753 uint64_t mfu_space = arc_c - arc_p;
2754 state = (arc_mru->arcs_lsize[type] >= size &&
2755 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2758 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2759 if (type == ARC_BUFC_METADATA) {
2760 buf->b_data = zio_buf_alloc(size);
2761 arc_space_consume(size, ARC_SPACE_DATA);
2764 * If we are unable to recycle an existing meta buffer
2765 * signal the reclaim thread. It will notify users
2766 * via the prune callback to drop references. The
2767 * prune callback in run in the context of the reclaim
2768 * thread to avoid deadlocking on the hash_lock.
2770 cv_signal(&arc_reclaim_thr_cv);
2772 ASSERT(type == ARC_BUFC_DATA);
2773 buf->b_data = zio_data_buf_alloc(size);
2774 ARCSTAT_INCR(arcstat_data_size, size);
2775 atomic_add_64(&arc_size, size);
2778 ARCSTAT_BUMP(arcstat_recycle_miss);
2780 ASSERT(buf->b_data != NULL);
2783 * Update the state size. Note that ghost states have a
2784 * "ghost size" and so don't need to be updated.
2786 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2787 arc_buf_hdr_t *hdr = buf->b_hdr;
2789 atomic_add_64(&hdr->b_state->arcs_size, size);
2790 if (list_link_active(&hdr->b_arc_node)) {
2791 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2792 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2795 * If we are growing the cache, and we are adding anonymous
2796 * data, and we have outgrown arc_p, update arc_p
2798 if (!zfs_arc_p_aggressive_disable &&
2799 arc_size < arc_c && hdr->b_state == arc_anon &&
2800 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2801 arc_p = MIN(arc_c, arc_p + size);
2806 * This routine is called whenever a buffer is accessed.
2807 * NOTE: the hash lock is dropped in this function.
2810 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2814 ASSERT(MUTEX_HELD(hash_lock));
2816 if (buf->b_state == arc_anon) {
2818 * This buffer is not in the cache, and does not
2819 * appear in our "ghost" list. Add the new buffer
2823 ASSERT(buf->b_arc_access == 0);
2824 buf->b_arc_access = ddi_get_lbolt();
2825 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2826 arc_change_state(arc_mru, buf, hash_lock);
2828 } else if (buf->b_state == arc_mru) {
2829 now = ddi_get_lbolt();
2832 * If this buffer is here because of a prefetch, then either:
2833 * - clear the flag if this is a "referencing" read
2834 * (any subsequent access will bump this into the MFU state).
2836 * - move the buffer to the head of the list if this is
2837 * another prefetch (to make it less likely to be evicted).
2839 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2840 if (refcount_count(&buf->b_refcnt) == 0) {
2841 ASSERT(list_link_active(&buf->b_arc_node));
2843 buf->b_flags &= ~ARC_PREFETCH;
2844 atomic_inc_32(&buf->b_mru_hits);
2845 ARCSTAT_BUMP(arcstat_mru_hits);
2847 buf->b_arc_access = now;
2852 * This buffer has been "accessed" only once so far,
2853 * but it is still in the cache. Move it to the MFU
2856 if (now > buf->b_arc_access + ARC_MINTIME) {
2858 * More than 125ms have passed since we
2859 * instantiated this buffer. Move it to the
2860 * most frequently used state.
2862 buf->b_arc_access = now;
2863 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2864 arc_change_state(arc_mfu, buf, hash_lock);
2866 atomic_inc_32(&buf->b_mru_hits);
2867 ARCSTAT_BUMP(arcstat_mru_hits);
2868 } else if (buf->b_state == arc_mru_ghost) {
2869 arc_state_t *new_state;
2871 * This buffer has been "accessed" recently, but
2872 * was evicted from the cache. Move it to the
2876 if (buf->b_flags & ARC_PREFETCH) {
2877 new_state = arc_mru;
2878 if (refcount_count(&buf->b_refcnt) > 0)
2879 buf->b_flags &= ~ARC_PREFETCH;
2880 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2882 new_state = arc_mfu;
2883 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2886 buf->b_arc_access = ddi_get_lbolt();
2887 arc_change_state(new_state, buf, hash_lock);
2889 atomic_inc_32(&buf->b_mru_ghost_hits);
2890 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2891 } else if (buf->b_state == arc_mfu) {
2893 * This buffer has been accessed more than once and is
2894 * still in the cache. Keep it in the MFU state.
2896 * NOTE: an add_reference() that occurred when we did
2897 * the arc_read() will have kicked this off the list.
2898 * If it was a prefetch, we will explicitly move it to
2899 * the head of the list now.
2901 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2902 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2903 ASSERT(list_link_active(&buf->b_arc_node));
2905 atomic_inc_32(&buf->b_mfu_hits);
2906 ARCSTAT_BUMP(arcstat_mfu_hits);
2907 buf->b_arc_access = ddi_get_lbolt();
2908 } else if (buf->b_state == arc_mfu_ghost) {
2909 arc_state_t *new_state = arc_mfu;
2911 * This buffer has been accessed more than once but has
2912 * been evicted from the cache. Move it back to the
2916 if (buf->b_flags & ARC_PREFETCH) {
2918 * This is a prefetch access...
2919 * move this block back to the MRU state.
2921 ASSERT0(refcount_count(&buf->b_refcnt));
2922 new_state = arc_mru;
2925 buf->b_arc_access = ddi_get_lbolt();
2926 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2927 arc_change_state(new_state, buf, hash_lock);
2929 atomic_inc_32(&buf->b_mfu_ghost_hits);
2930 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2931 } else if (buf->b_state == arc_l2c_only) {
2933 * This buffer is on the 2nd Level ARC.
2936 buf->b_arc_access = ddi_get_lbolt();
2937 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2938 arc_change_state(arc_mfu, buf, hash_lock);
2940 ASSERT(!"invalid arc state");
2944 /* a generic arc_done_func_t which you can use */
2947 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2949 if (zio == NULL || zio->io_error == 0)
2950 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2951 VERIFY(arc_buf_remove_ref(buf, arg));
2954 /* a generic arc_done_func_t */
2956 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2958 arc_buf_t **bufp = arg;
2959 if (zio && zio->io_error) {
2960 VERIFY(arc_buf_remove_ref(buf, arg));
2964 ASSERT(buf->b_data);
2969 arc_read_done(zio_t *zio)
2971 arc_buf_hdr_t *hdr, *found;
2973 arc_buf_t *abuf; /* buffer we're assigning to callback */
2974 kmutex_t *hash_lock;
2975 arc_callback_t *callback_list, *acb;
2976 int freeable = FALSE;
2978 buf = zio->io_private;
2982 * The hdr was inserted into hash-table and removed from lists
2983 * prior to starting I/O. We should find this header, since
2984 * it's in the hash table, and it should be legit since it's
2985 * not possible to evict it during the I/O. The only possible
2986 * reason for it not to be found is if we were freed during the
2989 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2992 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2993 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2994 (found == hdr && HDR_L2_READING(hdr)));
2996 hdr->b_flags &= ~ARC_L2_EVICTED;
2997 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2998 hdr->b_flags &= ~ARC_L2CACHE;
3000 /* byteswap if necessary */
3001 callback_list = hdr->b_acb;
3002 ASSERT(callback_list != NULL);
3003 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
3004 dmu_object_byteswap_t bswap =
3005 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
3006 if (BP_GET_LEVEL(zio->io_bp) > 0)
3007 byteswap_uint64_array(buf->b_data, hdr->b_size);
3009 dmu_ot_byteswap[bswap].ob_func(buf->b_data, hdr->b_size);
3012 arc_cksum_compute(buf, B_FALSE);
3015 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
3017 * Only call arc_access on anonymous buffers. This is because
3018 * if we've issued an I/O for an evicted buffer, we've already
3019 * called arc_access (to prevent any simultaneous readers from
3020 * getting confused).
3022 arc_access(hdr, hash_lock);
3025 /* create copies of the data buffer for the callers */
3027 for (acb = callback_list; acb; acb = acb->acb_next) {
3028 if (acb->acb_done) {
3030 ARCSTAT_BUMP(arcstat_duplicate_reads);
3031 abuf = arc_buf_clone(buf);
3033 acb->acb_buf = abuf;
3038 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3039 ASSERT(!HDR_BUF_AVAILABLE(hdr));
3041 ASSERT(buf->b_efunc == NULL);
3042 ASSERT(hdr->b_datacnt == 1);
3043 hdr->b_flags |= ARC_BUF_AVAILABLE;
3046 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
3048 if (zio->io_error != 0) {
3049 hdr->b_flags |= ARC_IO_ERROR;
3050 if (hdr->b_state != arc_anon)
3051 arc_change_state(arc_anon, hdr, hash_lock);
3052 if (HDR_IN_HASH_TABLE(hdr))
3053 buf_hash_remove(hdr);
3054 freeable = refcount_is_zero(&hdr->b_refcnt);
3058 * Broadcast before we drop the hash_lock to avoid the possibility
3059 * that the hdr (and hence the cv) might be freed before we get to
3060 * the cv_broadcast().
3062 cv_broadcast(&hdr->b_cv);
3065 mutex_exit(hash_lock);
3068 * This block was freed while we waited for the read to
3069 * complete. It has been removed from the hash table and
3070 * moved to the anonymous state (so that it won't show up
3073 ASSERT3P(hdr->b_state, ==, arc_anon);
3074 freeable = refcount_is_zero(&hdr->b_refcnt);
3077 /* execute each callback and free its structure */
3078 while ((acb = callback_list) != NULL) {
3080 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3082 if (acb->acb_zio_dummy != NULL) {
3083 acb->acb_zio_dummy->io_error = zio->io_error;
3084 zio_nowait(acb->acb_zio_dummy);
3087 callback_list = acb->acb_next;
3088 kmem_free(acb, sizeof (arc_callback_t));
3092 arc_hdr_destroy(hdr);
3096 * "Read" the block at the specified DVA (in bp) via the
3097 * cache. If the block is found in the cache, invoke the provided
3098 * callback immediately and return. Note that the `zio' parameter
3099 * in the callback will be NULL in this case, since no IO was
3100 * required. If the block is not in the cache pass the read request
3101 * on to the spa with a substitute callback function, so that the
3102 * requested block will be added to the cache.
3104 * If a read request arrives for a block that has a read in-progress,
3105 * either wait for the in-progress read to complete (and return the
3106 * results); or, if this is a read with a "done" func, add a record
3107 * to the read to invoke the "done" func when the read completes,
3108 * and return; or just return.
3110 * arc_read_done() will invoke all the requested "done" functions
3111 * for readers of this block.
3114 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3115 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags,
3116 const zbookmark_t *zb)
3119 arc_buf_t *buf = NULL;
3120 kmutex_t *hash_lock;
3122 uint64_t guid = spa_load_guid(spa);
3126 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3128 if (hdr && hdr->b_datacnt > 0) {
3130 *arc_flags |= ARC_CACHED;
3132 if (HDR_IO_IN_PROGRESS(hdr)) {
3134 if (*arc_flags & ARC_WAIT) {
3135 cv_wait(&hdr->b_cv, hash_lock);
3136 mutex_exit(hash_lock);
3139 ASSERT(*arc_flags & ARC_NOWAIT);
3142 arc_callback_t *acb = NULL;
3144 acb = kmem_zalloc(sizeof (arc_callback_t),
3146 acb->acb_done = done;
3147 acb->acb_private = private;
3149 acb->acb_zio_dummy = zio_null(pio,
3150 spa, NULL, NULL, NULL, zio_flags);
3152 ASSERT(acb->acb_done != NULL);
3153 acb->acb_next = hdr->b_acb;
3155 add_reference(hdr, hash_lock, private);
3156 mutex_exit(hash_lock);
3159 mutex_exit(hash_lock);
3163 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3166 add_reference(hdr, hash_lock, private);
3168 * If this block is already in use, create a new
3169 * copy of the data so that we will be guaranteed
3170 * that arc_release() will always succeed.
3174 ASSERT(buf->b_data);
3175 if (HDR_BUF_AVAILABLE(hdr)) {
3176 ASSERT(buf->b_efunc == NULL);
3177 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3179 buf = arc_buf_clone(buf);
3182 } else if (*arc_flags & ARC_PREFETCH &&
3183 refcount_count(&hdr->b_refcnt) == 0) {
3184 hdr->b_flags |= ARC_PREFETCH;
3186 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
3187 arc_access(hdr, hash_lock);
3188 if (*arc_flags & ARC_L2CACHE)
3189 hdr->b_flags |= ARC_L2CACHE;
3190 if (*arc_flags & ARC_L2COMPRESS)
3191 hdr->b_flags |= ARC_L2COMPRESS;
3192 mutex_exit(hash_lock);
3193 ARCSTAT_BUMP(arcstat_hits);
3194 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3195 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3196 data, metadata, hits);
3199 done(NULL, buf, private);
3201 uint64_t size = BP_GET_LSIZE(bp);
3202 arc_callback_t *acb;
3205 boolean_t devw = B_FALSE;
3208 /* this block is not in the cache */
3209 arc_buf_hdr_t *exists;
3210 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
3211 buf = arc_buf_alloc(spa, size, private, type);
3213 hdr->b_dva = *BP_IDENTITY(bp);
3214 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3215 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3216 exists = buf_hash_insert(hdr, &hash_lock);
3218 /* somebody beat us to the hash insert */
3219 mutex_exit(hash_lock);
3220 buf_discard_identity(hdr);
3221 (void) arc_buf_remove_ref(buf, private);
3222 goto top; /* restart the IO request */
3224 /* if this is a prefetch, we don't have a reference */
3225 if (*arc_flags & ARC_PREFETCH) {
3226 (void) remove_reference(hdr, hash_lock,
3228 hdr->b_flags |= ARC_PREFETCH;
3230 if (*arc_flags & ARC_L2CACHE)
3231 hdr->b_flags |= ARC_L2CACHE;
3232 if (*arc_flags & ARC_L2COMPRESS)
3233 hdr->b_flags |= ARC_L2COMPRESS;
3234 if (BP_GET_LEVEL(bp) > 0)
3235 hdr->b_flags |= ARC_INDIRECT;
3237 /* this block is in the ghost cache */
3238 ASSERT(GHOST_STATE(hdr->b_state));
3239 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3240 ASSERT0(refcount_count(&hdr->b_refcnt));
3241 ASSERT(hdr->b_buf == NULL);
3243 /* if this is a prefetch, we don't have a reference */
3244 if (*arc_flags & ARC_PREFETCH)
3245 hdr->b_flags |= ARC_PREFETCH;
3247 add_reference(hdr, hash_lock, private);
3248 if (*arc_flags & ARC_L2CACHE)
3249 hdr->b_flags |= ARC_L2CACHE;
3250 if (*arc_flags & ARC_L2COMPRESS)
3251 hdr->b_flags |= ARC_L2COMPRESS;
3252 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3255 buf->b_efunc = NULL;
3256 buf->b_private = NULL;
3259 ASSERT(hdr->b_datacnt == 0);
3261 arc_get_data_buf(buf);
3262 arc_access(hdr, hash_lock);
3265 ASSERT(!GHOST_STATE(hdr->b_state));
3267 acb = kmem_zalloc(sizeof (arc_callback_t), KM_PUSHPAGE);
3268 acb->acb_done = done;
3269 acb->acb_private = private;
3271 ASSERT(hdr->b_acb == NULL);
3273 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3275 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3276 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3277 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3278 addr = hdr->b_l2hdr->b_daddr;
3280 * Lock out device removal.
3282 if (vdev_is_dead(vd) ||
3283 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3287 mutex_exit(hash_lock);
3290 * At this point, we have a level 1 cache miss. Try again in
3291 * L2ARC if possible.
3293 ASSERT3U(hdr->b_size, ==, size);
3294 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3295 uint64_t, size, zbookmark_t *, zb);
3296 ARCSTAT_BUMP(arcstat_misses);
3297 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3298 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3299 data, metadata, misses);
3301 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3303 * Read from the L2ARC if the following are true:
3304 * 1. The L2ARC vdev was previously cached.
3305 * 2. This buffer still has L2ARC metadata.
3306 * 3. This buffer isn't currently writing to the L2ARC.
3307 * 4. The L2ARC entry wasn't evicted, which may
3308 * also have invalidated the vdev.
3309 * 5. This isn't prefetch and l2arc_noprefetch is set.
3311 if (hdr->b_l2hdr != NULL &&
3312 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3313 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3314 l2arc_read_callback_t *cb;
3316 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3317 ARCSTAT_BUMP(arcstat_l2_hits);
3318 atomic_inc_32(&hdr->b_l2hdr->b_hits);
3320 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3322 cb->l2rcb_buf = buf;
3323 cb->l2rcb_spa = spa;
3326 cb->l2rcb_flags = zio_flags;
3327 cb->l2rcb_compress = hdr->b_l2hdr->b_compress;
3329 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
3330 addr + size < vd->vdev_psize -
3331 VDEV_LABEL_END_SIZE);
3334 * l2arc read. The SCL_L2ARC lock will be
3335 * released by l2arc_read_done().
3336 * Issue a null zio if the underlying buffer
3337 * was squashed to zero size by compression.
3339 if (hdr->b_l2hdr->b_compress ==
3340 ZIO_COMPRESS_EMPTY) {
3341 rzio = zio_null(pio, spa, vd,
3342 l2arc_read_done, cb,
3343 zio_flags | ZIO_FLAG_DONT_CACHE |
3345 ZIO_FLAG_DONT_PROPAGATE |
3346 ZIO_FLAG_DONT_RETRY);
3348 rzio = zio_read_phys(pio, vd, addr,
3349 hdr->b_l2hdr->b_asize,
3350 buf->b_data, ZIO_CHECKSUM_OFF,
3351 l2arc_read_done, cb, priority,
3352 zio_flags | ZIO_FLAG_DONT_CACHE |
3354 ZIO_FLAG_DONT_PROPAGATE |
3355 ZIO_FLAG_DONT_RETRY, B_FALSE);
3357 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3359 ARCSTAT_INCR(arcstat_l2_read_bytes,
3360 hdr->b_l2hdr->b_asize);
3362 if (*arc_flags & ARC_NOWAIT) {
3367 ASSERT(*arc_flags & ARC_WAIT);
3368 if (zio_wait(rzio) == 0)
3371 /* l2arc read error; goto zio_read() */
3373 DTRACE_PROBE1(l2arc__miss,
3374 arc_buf_hdr_t *, hdr);
3375 ARCSTAT_BUMP(arcstat_l2_misses);
3376 if (HDR_L2_WRITING(hdr))
3377 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3378 spa_config_exit(spa, SCL_L2ARC, vd);
3382 spa_config_exit(spa, SCL_L2ARC, vd);
3383 if (l2arc_ndev != 0) {
3384 DTRACE_PROBE1(l2arc__miss,
3385 arc_buf_hdr_t *, hdr);
3386 ARCSTAT_BUMP(arcstat_l2_misses);
3390 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3391 arc_read_done, buf, priority, zio_flags, zb);
3393 if (*arc_flags & ARC_WAIT) {
3394 rc = zio_wait(rzio);
3398 ASSERT(*arc_flags & ARC_NOWAIT);
3403 spa_read_history_add(spa, zb, *arc_flags);
3408 arc_add_prune_callback(arc_prune_func_t *func, void *private)
3412 p = kmem_alloc(sizeof (*p), KM_SLEEP);
3414 p->p_private = private;
3415 list_link_init(&p->p_node);
3416 refcount_create(&p->p_refcnt);
3418 mutex_enter(&arc_prune_mtx);
3419 refcount_add(&p->p_refcnt, &arc_prune_list);
3420 list_insert_head(&arc_prune_list, p);
3421 mutex_exit(&arc_prune_mtx);
3427 arc_remove_prune_callback(arc_prune_t *p)
3429 mutex_enter(&arc_prune_mtx);
3430 list_remove(&arc_prune_list, p);
3431 if (refcount_remove(&p->p_refcnt, &arc_prune_list) == 0) {
3432 refcount_destroy(&p->p_refcnt);
3433 kmem_free(p, sizeof (*p));
3435 mutex_exit(&arc_prune_mtx);
3439 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3441 ASSERT(buf->b_hdr != NULL);
3442 ASSERT(buf->b_hdr->b_state != arc_anon);
3443 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3444 ASSERT(buf->b_efunc == NULL);
3445 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3447 buf->b_efunc = func;
3448 buf->b_private = private;
3452 * Notify the arc that a block was freed, and thus will never be used again.
3455 arc_freed(spa_t *spa, const blkptr_t *bp)
3458 kmutex_t *hash_lock;
3459 uint64_t guid = spa_load_guid(spa);
3461 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3465 if (HDR_BUF_AVAILABLE(hdr)) {
3466 arc_buf_t *buf = hdr->b_buf;
3467 add_reference(hdr, hash_lock, FTAG);
3468 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3469 mutex_exit(hash_lock);
3471 arc_release(buf, FTAG);
3472 (void) arc_buf_remove_ref(buf, FTAG);
3474 mutex_exit(hash_lock);
3480 * This is used by the DMU to let the ARC know that a buffer is
3481 * being evicted, so the ARC should clean up. If this arc buf
3482 * is not yet in the evicted state, it will be put there.
3485 arc_buf_evict(arc_buf_t *buf)
3488 kmutex_t *hash_lock;
3491 mutex_enter(&buf->b_evict_lock);
3495 * We are in arc_do_user_evicts().
3497 ASSERT(buf->b_data == NULL);
3498 mutex_exit(&buf->b_evict_lock);
3500 } else if (buf->b_data == NULL) {
3501 arc_buf_t copy = *buf; /* structure assignment */
3503 * We are on the eviction list; process this buffer now
3504 * but let arc_do_user_evicts() do the reaping.
3506 buf->b_efunc = NULL;
3507 mutex_exit(&buf->b_evict_lock);
3508 VERIFY(copy.b_efunc(©) == 0);
3511 hash_lock = HDR_LOCK(hdr);
3512 mutex_enter(hash_lock);
3514 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3516 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3517 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3520 * Pull this buffer off of the hdr
3523 while (*bufp != buf)
3524 bufp = &(*bufp)->b_next;
3525 *bufp = buf->b_next;
3527 ASSERT(buf->b_data != NULL);
3528 arc_buf_destroy(buf, FALSE, FALSE);
3530 if (hdr->b_datacnt == 0) {
3531 arc_state_t *old_state = hdr->b_state;
3532 arc_state_t *evicted_state;
3534 ASSERT(hdr->b_buf == NULL);
3535 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3538 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3540 mutex_enter(&old_state->arcs_mtx);
3541 mutex_enter(&evicted_state->arcs_mtx);
3543 arc_change_state(evicted_state, hdr, hash_lock);
3544 ASSERT(HDR_IN_HASH_TABLE(hdr));
3545 hdr->b_flags |= ARC_IN_HASH_TABLE;
3546 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3548 mutex_exit(&evicted_state->arcs_mtx);
3549 mutex_exit(&old_state->arcs_mtx);
3551 mutex_exit(hash_lock);
3552 mutex_exit(&buf->b_evict_lock);
3554 VERIFY(buf->b_efunc(buf) == 0);
3555 buf->b_efunc = NULL;
3556 buf->b_private = NULL;
3559 kmem_cache_free(buf_cache, buf);
3564 * Release this buffer from the cache, making it an anonymous buffer. This
3565 * must be done after a read and prior to modifying the buffer contents.
3566 * If the buffer has more than one reference, we must make
3567 * a new hdr for the buffer.
3570 arc_release(arc_buf_t *buf, void *tag)
3573 kmutex_t *hash_lock = NULL;
3574 l2arc_buf_hdr_t *l2hdr;
3575 uint64_t buf_size = 0;
3578 * It would be nice to assert that if it's DMU metadata (level >
3579 * 0 || it's the dnode file), then it must be syncing context.
3580 * But we don't know that information at this level.
3583 mutex_enter(&buf->b_evict_lock);
3586 /* this buffer is not on any list */
3587 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3589 if (hdr->b_state == arc_anon) {
3590 /* this buffer is already released */
3591 ASSERT(buf->b_efunc == NULL);
3593 hash_lock = HDR_LOCK(hdr);
3594 mutex_enter(hash_lock);
3596 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3599 l2hdr = hdr->b_l2hdr;
3601 mutex_enter(&l2arc_buflist_mtx);
3602 hdr->b_l2hdr = NULL;
3604 buf_size = hdr->b_size;
3607 * Do we have more than one buf?
3609 if (hdr->b_datacnt > 1) {
3610 arc_buf_hdr_t *nhdr;
3612 uint64_t blksz = hdr->b_size;
3613 uint64_t spa = hdr->b_spa;
3614 arc_buf_contents_t type = hdr->b_type;
3615 uint32_t flags = hdr->b_flags;
3617 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3619 * Pull the data off of this hdr and attach it to
3620 * a new anonymous hdr.
3622 (void) remove_reference(hdr, hash_lock, tag);
3624 while (*bufp != buf)
3625 bufp = &(*bufp)->b_next;
3626 *bufp = buf->b_next;
3629 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3630 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3631 if (refcount_is_zero(&hdr->b_refcnt)) {
3632 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3633 ASSERT3U(*size, >=, hdr->b_size);
3634 atomic_add_64(size, -hdr->b_size);
3638 * We're releasing a duplicate user data buffer, update
3639 * our statistics accordingly.
3641 if (hdr->b_type == ARC_BUFC_DATA) {
3642 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3643 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3646 hdr->b_datacnt -= 1;
3647 arc_cksum_verify(buf);
3648 arc_buf_unwatch(buf);
3650 mutex_exit(hash_lock);
3652 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3653 nhdr->b_size = blksz;
3655 nhdr->b_type = type;
3657 nhdr->b_state = arc_anon;
3658 nhdr->b_arc_access = 0;
3659 nhdr->b_mru_hits = 0;
3660 nhdr->b_mru_ghost_hits = 0;
3661 nhdr->b_mfu_hits = 0;
3662 nhdr->b_mfu_ghost_hits = 0;
3663 nhdr->b_l2_hits = 0;
3664 nhdr->b_flags = flags & ARC_L2_WRITING;
3665 nhdr->b_l2hdr = NULL;
3666 nhdr->b_datacnt = 1;
3667 nhdr->b_freeze_cksum = NULL;
3668 (void) refcount_add(&nhdr->b_refcnt, tag);
3670 mutex_exit(&buf->b_evict_lock);
3671 atomic_add_64(&arc_anon->arcs_size, blksz);
3673 mutex_exit(&buf->b_evict_lock);
3674 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3675 ASSERT(!list_link_active(&hdr->b_arc_node));
3676 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3677 if (hdr->b_state != arc_anon)
3678 arc_change_state(arc_anon, hdr, hash_lock);
3679 hdr->b_arc_access = 0;
3680 hdr->b_mru_hits = 0;
3681 hdr->b_mru_ghost_hits = 0;
3682 hdr->b_mfu_hits = 0;
3683 hdr->b_mfu_ghost_hits = 0;
3686 mutex_exit(hash_lock);
3688 buf_discard_identity(hdr);
3691 buf->b_efunc = NULL;
3692 buf->b_private = NULL;
3695 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
3696 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3697 kmem_cache_free(l2arc_hdr_cache, l2hdr);
3698 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
3699 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3700 mutex_exit(&l2arc_buflist_mtx);
3705 arc_released(arc_buf_t *buf)
3709 mutex_enter(&buf->b_evict_lock);
3710 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3711 mutex_exit(&buf->b_evict_lock);
3716 arc_has_callback(arc_buf_t *buf)
3720 mutex_enter(&buf->b_evict_lock);
3721 callback = (buf->b_efunc != NULL);
3722 mutex_exit(&buf->b_evict_lock);
3728 arc_referenced(arc_buf_t *buf)
3732 mutex_enter(&buf->b_evict_lock);
3733 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3734 mutex_exit(&buf->b_evict_lock);
3735 return (referenced);
3740 arc_write_ready(zio_t *zio)
3742 arc_write_callback_t *callback = zio->io_private;
3743 arc_buf_t *buf = callback->awcb_buf;
3744 arc_buf_hdr_t *hdr = buf->b_hdr;
3746 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3747 callback->awcb_ready(zio, buf, callback->awcb_private);
3750 * If the IO is already in progress, then this is a re-write
3751 * attempt, so we need to thaw and re-compute the cksum.
3752 * It is the responsibility of the callback to handle the
3753 * accounting for any re-write attempt.
3755 if (HDR_IO_IN_PROGRESS(hdr)) {
3756 mutex_enter(&hdr->b_freeze_lock);
3757 if (hdr->b_freeze_cksum != NULL) {
3758 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3759 hdr->b_freeze_cksum = NULL;
3761 mutex_exit(&hdr->b_freeze_lock);
3763 arc_cksum_compute(buf, B_FALSE);
3764 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3768 * The SPA calls this callback for each physical write that happens on behalf
3769 * of a logical write. See the comment in dbuf_write_physdone() for details.
3772 arc_write_physdone(zio_t *zio)
3774 arc_write_callback_t *cb = zio->io_private;
3775 if (cb->awcb_physdone != NULL)
3776 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
3780 arc_write_done(zio_t *zio)
3782 arc_write_callback_t *callback = zio->io_private;
3783 arc_buf_t *buf = callback->awcb_buf;
3784 arc_buf_hdr_t *hdr = buf->b_hdr;
3786 ASSERT(hdr->b_acb == NULL);
3788 if (zio->io_error == 0) {
3789 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3790 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3791 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3793 ASSERT(BUF_EMPTY(hdr));
3797 * If the block to be written was all-zero, we may have
3798 * compressed it away. In this case no write was performed
3799 * so there will be no dva/birth/checksum. The buffer must
3800 * therefore remain anonymous (and uncached).
3802 if (!BUF_EMPTY(hdr)) {
3803 arc_buf_hdr_t *exists;
3804 kmutex_t *hash_lock;
3806 ASSERT(zio->io_error == 0);
3808 arc_cksum_verify(buf);
3810 exists = buf_hash_insert(hdr, &hash_lock);
3813 * This can only happen if we overwrite for
3814 * sync-to-convergence, because we remove
3815 * buffers from the hash table when we arc_free().
3817 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3818 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3819 panic("bad overwrite, hdr=%p exists=%p",
3820 (void *)hdr, (void *)exists);
3821 ASSERT(refcount_is_zero(&exists->b_refcnt));
3822 arc_change_state(arc_anon, exists, hash_lock);
3823 mutex_exit(hash_lock);
3824 arc_hdr_destroy(exists);
3825 exists = buf_hash_insert(hdr, &hash_lock);
3826 ASSERT3P(exists, ==, NULL);
3827 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3829 ASSERT(zio->io_prop.zp_nopwrite);
3830 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3831 panic("bad nopwrite, hdr=%p exists=%p",
3832 (void *)hdr, (void *)exists);
3835 ASSERT(hdr->b_datacnt == 1);
3836 ASSERT(hdr->b_state == arc_anon);
3837 ASSERT(BP_GET_DEDUP(zio->io_bp));
3838 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3841 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3842 /* if it's not anon, we are doing a scrub */
3843 if (!exists && hdr->b_state == arc_anon)
3844 arc_access(hdr, hash_lock);
3845 mutex_exit(hash_lock);
3847 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3850 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3851 callback->awcb_done(zio, buf, callback->awcb_private);
3853 kmem_free(callback, sizeof (arc_write_callback_t));
3857 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3858 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
3859 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
3860 arc_done_func_t *done, void *private, zio_priority_t priority,
3861 int zio_flags, const zbookmark_t *zb)
3863 arc_buf_hdr_t *hdr = buf->b_hdr;
3864 arc_write_callback_t *callback;
3867 ASSERT(ready != NULL);
3868 ASSERT(done != NULL);
3869 ASSERT(!HDR_IO_ERROR(hdr));
3870 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3871 ASSERT(hdr->b_acb == NULL);
3873 hdr->b_flags |= ARC_L2CACHE;
3875 hdr->b_flags |= ARC_L2COMPRESS;
3876 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_PUSHPAGE);
3877 callback->awcb_ready = ready;
3878 callback->awcb_physdone = physdone;
3879 callback->awcb_done = done;
3880 callback->awcb_private = private;
3881 callback->awcb_buf = buf;
3883 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3884 arc_write_ready, arc_write_physdone, arc_write_done, callback,
3885 priority, zio_flags, zb);
3891 arc_memory_throttle(uint64_t reserve, uint64_t txg)
3894 if (zfs_arc_memory_throttle_disable)
3897 if (freemem <= physmem * arc_lotsfree_percent / 100) {
3898 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3899 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
3900 return (SET_ERROR(EAGAIN));
3907 arc_tempreserve_clear(uint64_t reserve)
3909 atomic_add_64(&arc_tempreserve, -reserve);
3910 ASSERT((int64_t)arc_tempreserve >= 0);
3914 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3919 if (reserve > arc_c/4 && !arc_no_grow)
3920 arc_c = MIN(arc_c_max, reserve * 4);
3921 if (reserve > arc_c) {
3922 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
3923 return (SET_ERROR(ENOMEM));
3927 * Don't count loaned bufs as in flight dirty data to prevent long
3928 * network delays from blocking transactions that are ready to be
3929 * assigned to a txg.
3931 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3934 * Writes will, almost always, require additional memory allocations
3935 * in order to compress/encrypt/etc the data. We therefore need to
3936 * make sure that there is sufficient available memory for this.
3938 error = arc_memory_throttle(reserve, txg);
3943 * Throttle writes when the amount of dirty data in the cache
3944 * gets too large. We try to keep the cache less than half full
3945 * of dirty blocks so that our sync times don't grow too large.
3946 * Note: if two requests come in concurrently, we might let them
3947 * both succeed, when one of them should fail. Not a huge deal.
3950 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3951 anon_size > arc_c / 4) {
3952 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3953 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3954 arc_tempreserve>>10,
3955 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3956 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3957 reserve>>10, arc_c>>10);
3958 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
3959 return (SET_ERROR(ERESTART));
3961 atomic_add_64(&arc_tempreserve, reserve);
3966 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
3967 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
3969 size->value.ui64 = state->arcs_size;
3970 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
3971 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
3975 arc_kstat_update(kstat_t *ksp, int rw)
3977 arc_stats_t *as = ksp->ks_data;
3979 if (rw == KSTAT_WRITE) {
3980 return (SET_ERROR(EACCES));
3982 arc_kstat_update_state(arc_anon,
3983 &as->arcstat_anon_size,
3984 &as->arcstat_anon_evict_data,
3985 &as->arcstat_anon_evict_metadata);
3986 arc_kstat_update_state(arc_mru,
3987 &as->arcstat_mru_size,
3988 &as->arcstat_mru_evict_data,
3989 &as->arcstat_mru_evict_metadata);
3990 arc_kstat_update_state(arc_mru_ghost,
3991 &as->arcstat_mru_ghost_size,
3992 &as->arcstat_mru_ghost_evict_data,
3993 &as->arcstat_mru_ghost_evict_metadata);
3994 arc_kstat_update_state(arc_mfu,
3995 &as->arcstat_mfu_size,
3996 &as->arcstat_mfu_evict_data,
3997 &as->arcstat_mfu_evict_metadata);
3998 arc_kstat_update_state(arc_mfu_ghost,
3999 &as->arcstat_mfu_ghost_size,
4000 &as->arcstat_mfu_ghost_evict_data,
4001 &as->arcstat_mfu_ghost_evict_metadata);
4010 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4011 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
4013 /* Convert seconds to clock ticks */
4014 zfs_arc_min_prefetch_lifespan = 1 * hz;
4016 /* Start out with 1/8 of all memory */
4017 arc_c = physmem * PAGESIZE / 8;
4021 * On architectures where the physical memory can be larger
4022 * than the addressable space (intel in 32-bit mode), we may
4023 * need to limit the cache to 1/8 of VM size.
4025 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
4027 * Register a shrinker to support synchronous (direct) memory
4028 * reclaim from the arc. This is done to prevent kswapd from
4029 * swapping out pages when it is preferable to shrink the arc.
4031 spl_register_shrinker(&arc_shrinker);
4034 /* set min cache to zero */
4036 /* set max to 1/2 of all memory */
4037 arc_c_max = arc_c * 4;
4040 * Allow the tunables to override our calculations if they are
4041 * reasonable (ie. over 64MB)
4043 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
4044 arc_c_max = zfs_arc_max;
4045 if (zfs_arc_min > 0 && zfs_arc_min <= arc_c_max)
4046 arc_c_min = zfs_arc_min;
4049 arc_p = (arc_c >> 1);
4051 /* limit meta-data to 1/4 of the arc capacity */
4052 arc_meta_limit = arc_c_max / 4;
4055 /* Allow the tunable to override if it is reasonable */
4056 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
4057 arc_meta_limit = zfs_arc_meta_limit;
4059 /* if kmem_flags are set, lets try to use less memory */
4060 if (kmem_debugging())
4062 if (arc_c < arc_c_min)
4065 arc_anon = &ARC_anon;
4067 arc_mru_ghost = &ARC_mru_ghost;
4069 arc_mfu_ghost = &ARC_mfu_ghost;
4070 arc_l2c_only = &ARC_l2c_only;
4073 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
4074 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
4075 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
4076 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
4077 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
4078 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
4080 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
4081 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4082 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
4083 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4084 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
4085 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4086 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
4087 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4088 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
4089 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4090 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
4091 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4092 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
4093 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4094 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
4095 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4096 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
4097 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4098 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
4099 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4101 arc_anon->arcs_state = ARC_STATE_ANON;
4102 arc_mru->arcs_state = ARC_STATE_MRU;
4103 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
4104 arc_mfu->arcs_state = ARC_STATE_MFU;
4105 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
4106 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
4110 arc_thread_exit = 0;
4111 list_create(&arc_prune_list, sizeof (arc_prune_t),
4112 offsetof(arc_prune_t, p_node));
4113 arc_eviction_list = NULL;
4114 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
4115 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
4116 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
4118 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
4119 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
4121 if (arc_ksp != NULL) {
4122 arc_ksp->ks_data = &arc_stats;
4123 arc_ksp->ks_update = arc_kstat_update;
4124 kstat_install(arc_ksp);
4127 (void) thread_create(NULL, 0, arc_adapt_thread, NULL, 0, &p0,
4128 TS_RUN, minclsyspri);
4134 * Calculate maximum amount of dirty data per pool.
4136 * If it has been set by a module parameter, take that.
4137 * Otherwise, use a percentage of physical memory defined by
4138 * zfs_dirty_data_max_percent (default 10%) with a cap at
4139 * zfs_dirty_data_max_max (default 25% of physical memory).
4141 if (zfs_dirty_data_max_max == 0)
4142 zfs_dirty_data_max_max = physmem * PAGESIZE *
4143 zfs_dirty_data_max_max_percent / 100;
4145 if (zfs_dirty_data_max == 0) {
4146 zfs_dirty_data_max = physmem * PAGESIZE *
4147 zfs_dirty_data_max_percent / 100;
4148 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
4149 zfs_dirty_data_max_max);
4158 mutex_enter(&arc_reclaim_thr_lock);
4160 spl_unregister_shrinker(&arc_shrinker);
4161 #endif /* _KERNEL */
4163 arc_thread_exit = 1;
4164 while (arc_thread_exit != 0)
4165 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
4166 mutex_exit(&arc_reclaim_thr_lock);
4172 if (arc_ksp != NULL) {
4173 kstat_delete(arc_ksp);
4177 mutex_enter(&arc_prune_mtx);
4178 while ((p = list_head(&arc_prune_list)) != NULL) {
4179 list_remove(&arc_prune_list, p);
4180 refcount_remove(&p->p_refcnt, &arc_prune_list);
4181 refcount_destroy(&p->p_refcnt);
4182 kmem_free(p, sizeof (*p));
4184 mutex_exit(&arc_prune_mtx);
4186 list_destroy(&arc_prune_list);
4187 mutex_destroy(&arc_prune_mtx);
4188 mutex_destroy(&arc_eviction_mtx);
4189 mutex_destroy(&arc_reclaim_thr_lock);
4190 cv_destroy(&arc_reclaim_thr_cv);
4192 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
4193 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
4194 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
4195 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
4196 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
4197 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
4198 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
4199 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
4201 mutex_destroy(&arc_anon->arcs_mtx);
4202 mutex_destroy(&arc_mru->arcs_mtx);
4203 mutex_destroy(&arc_mru_ghost->arcs_mtx);
4204 mutex_destroy(&arc_mfu->arcs_mtx);
4205 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
4206 mutex_destroy(&arc_l2c_only->arcs_mtx);
4210 ASSERT(arc_loaned_bytes == 0);
4216 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
4217 * It uses dedicated storage devices to hold cached data, which are populated
4218 * using large infrequent writes. The main role of this cache is to boost
4219 * the performance of random read workloads. The intended L2ARC devices
4220 * include short-stroked disks, solid state disks, and other media with
4221 * substantially faster read latency than disk.
4223 * +-----------------------+
4225 * +-----------------------+
4228 * l2arc_feed_thread() arc_read()
4232 * +---------------+ |
4234 * +---------------+ |
4239 * +-------+ +-------+
4241 * | cache | | cache |
4242 * +-------+ +-------+
4243 * +=========+ .-----.
4244 * : L2ARC : |-_____-|
4245 * : devices : | Disks |
4246 * +=========+ `-_____-'
4248 * Read requests are satisfied from the following sources, in order:
4251 * 2) vdev cache of L2ARC devices
4253 * 4) vdev cache of disks
4256 * Some L2ARC device types exhibit extremely slow write performance.
4257 * To accommodate for this there are some significant differences between
4258 * the L2ARC and traditional cache design:
4260 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4261 * the ARC behave as usual, freeing buffers and placing headers on ghost
4262 * lists. The ARC does not send buffers to the L2ARC during eviction as
4263 * this would add inflated write latencies for all ARC memory pressure.
4265 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4266 * It does this by periodically scanning buffers from the eviction-end of
4267 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4268 * not already there. It scans until a headroom of buffers is satisfied,
4269 * which itself is a buffer for ARC eviction. If a compressible buffer is
4270 * found during scanning and selected for writing to an L2ARC device, we
4271 * temporarily boost scanning headroom during the next scan cycle to make
4272 * sure we adapt to compression effects (which might significantly reduce
4273 * the data volume we write to L2ARC). The thread that does this is
4274 * l2arc_feed_thread(), illustrated below; example sizes are included to
4275 * provide a better sense of ratio than this diagram:
4278 * +---------------------+----------+
4279 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4280 * +---------------------+----------+ | o L2ARC eligible
4281 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4282 * +---------------------+----------+ |
4283 * 15.9 Gbytes ^ 32 Mbytes |
4285 * l2arc_feed_thread()
4287 * l2arc write hand <--[oooo]--'
4291 * +==============================+
4292 * L2ARC dev |####|#|###|###| |####| ... |
4293 * +==============================+
4296 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4297 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4298 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4299 * safe to say that this is an uncommon case, since buffers at the end of
4300 * the ARC lists have moved there due to inactivity.
4302 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4303 * then the L2ARC simply misses copying some buffers. This serves as a
4304 * pressure valve to prevent heavy read workloads from both stalling the ARC
4305 * with waits and clogging the L2ARC with writes. This also helps prevent
4306 * the potential for the L2ARC to churn if it attempts to cache content too
4307 * quickly, such as during backups of the entire pool.
4309 * 5. After system boot and before the ARC has filled main memory, there are
4310 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4311 * lists can remain mostly static. Instead of searching from tail of these
4312 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4313 * for eligible buffers, greatly increasing its chance of finding them.
4315 * The L2ARC device write speed is also boosted during this time so that
4316 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4317 * there are no L2ARC reads, and no fear of degrading read performance
4318 * through increased writes.
4320 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4321 * the vdev queue can aggregate them into larger and fewer writes. Each
4322 * device is written to in a rotor fashion, sweeping writes through
4323 * available space then repeating.
4325 * 7. The L2ARC does not store dirty content. It never needs to flush
4326 * write buffers back to disk based storage.
4328 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4329 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4331 * The performance of the L2ARC can be tweaked by a number of tunables, which
4332 * may be necessary for different workloads:
4334 * l2arc_write_max max write bytes per interval
4335 * l2arc_write_boost extra write bytes during device warmup
4336 * l2arc_noprefetch skip caching prefetched buffers
4337 * l2arc_nocompress skip compressing buffers
4338 * l2arc_headroom number of max device writes to precache
4339 * l2arc_headroom_boost when we find compressed buffers during ARC
4340 * scanning, we multiply headroom by this
4341 * percentage factor for the next scan cycle,
4342 * since more compressed buffers are likely to
4344 * l2arc_feed_secs seconds between L2ARC writing
4346 * Tunables may be removed or added as future performance improvements are
4347 * integrated, and also may become zpool properties.
4349 * There are three key functions that control how the L2ARC warms up:
4351 * l2arc_write_eligible() check if a buffer is eligible to cache
4352 * l2arc_write_size() calculate how much to write
4353 * l2arc_write_interval() calculate sleep delay between writes
4355 * These three functions determine what to write, how much, and how quickly
4360 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4363 * A buffer is *not* eligible for the L2ARC if it:
4364 * 1. belongs to a different spa.
4365 * 2. is already cached on the L2ARC.
4366 * 3. has an I/O in progress (it may be an incomplete read).
4367 * 4. is flagged not eligible (zfs property).
4369 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
4370 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
4377 l2arc_write_size(void)
4382 * Make sure our globals have meaningful values in case the user
4385 size = l2arc_write_max;
4387 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
4388 "be greater than zero, resetting it to the default (%d)",
4390 size = l2arc_write_max = L2ARC_WRITE_SIZE;
4393 if (arc_warm == B_FALSE)
4394 size += l2arc_write_boost;
4401 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4403 clock_t interval, next, now;
4406 * If the ARC lists are busy, increase our write rate; if the
4407 * lists are stale, idle back. This is achieved by checking
4408 * how much we previously wrote - if it was more than half of
4409 * what we wanted, schedule the next write much sooner.
4411 if (l2arc_feed_again && wrote > (wanted / 2))
4412 interval = (hz * l2arc_feed_min_ms) / 1000;
4414 interval = hz * l2arc_feed_secs;
4416 now = ddi_get_lbolt();
4417 next = MAX(now, MIN(now + interval, began + interval));
4423 l2arc_hdr_stat_add(void)
4425 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE);
4426 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4430 l2arc_hdr_stat_remove(void)
4432 ARCSTAT_INCR(arcstat_l2_hdr_size, -HDR_SIZE);
4433 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4437 * Cycle through L2ARC devices. This is how L2ARC load balances.
4438 * If a device is returned, this also returns holding the spa config lock.
4440 static l2arc_dev_t *
4441 l2arc_dev_get_next(void)
4443 l2arc_dev_t *first, *next = NULL;
4446 * Lock out the removal of spas (spa_namespace_lock), then removal
4447 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4448 * both locks will be dropped and a spa config lock held instead.
4450 mutex_enter(&spa_namespace_lock);
4451 mutex_enter(&l2arc_dev_mtx);
4453 /* if there are no vdevs, there is nothing to do */
4454 if (l2arc_ndev == 0)
4458 next = l2arc_dev_last;
4460 /* loop around the list looking for a non-faulted vdev */
4462 next = list_head(l2arc_dev_list);
4464 next = list_next(l2arc_dev_list, next);
4466 next = list_head(l2arc_dev_list);
4469 /* if we have come back to the start, bail out */
4472 else if (next == first)
4475 } while (vdev_is_dead(next->l2ad_vdev));
4477 /* if we were unable to find any usable vdevs, return NULL */
4478 if (vdev_is_dead(next->l2ad_vdev))
4481 l2arc_dev_last = next;
4484 mutex_exit(&l2arc_dev_mtx);
4487 * Grab the config lock to prevent the 'next' device from being
4488 * removed while we are writing to it.
4491 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4492 mutex_exit(&spa_namespace_lock);
4498 * Free buffers that were tagged for destruction.
4501 l2arc_do_free_on_write(void)
4504 l2arc_data_free_t *df, *df_prev;
4506 mutex_enter(&l2arc_free_on_write_mtx);
4507 buflist = l2arc_free_on_write;
4509 for (df = list_tail(buflist); df; df = df_prev) {
4510 df_prev = list_prev(buflist, df);
4511 ASSERT(df->l2df_data != NULL);
4512 ASSERT(df->l2df_func != NULL);
4513 df->l2df_func(df->l2df_data, df->l2df_size);
4514 list_remove(buflist, df);
4515 kmem_free(df, sizeof (l2arc_data_free_t));
4518 mutex_exit(&l2arc_free_on_write_mtx);
4522 * A write to a cache device has completed. Update all headers to allow
4523 * reads from these buffers to begin.
4526 l2arc_write_done(zio_t *zio)
4528 l2arc_write_callback_t *cb;
4531 arc_buf_hdr_t *head, *ab, *ab_prev;
4532 l2arc_buf_hdr_t *abl2;
4533 kmutex_t *hash_lock;
4535 cb = zio->io_private;
4537 dev = cb->l2wcb_dev;
4538 ASSERT(dev != NULL);
4539 head = cb->l2wcb_head;
4540 ASSERT(head != NULL);
4541 buflist = dev->l2ad_buflist;
4542 ASSERT(buflist != NULL);
4543 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4544 l2arc_write_callback_t *, cb);
4546 if (zio->io_error != 0)
4547 ARCSTAT_BUMP(arcstat_l2_writes_error);
4549 mutex_enter(&l2arc_buflist_mtx);
4552 * All writes completed, or an error was hit.
4554 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4555 ab_prev = list_prev(buflist, ab);
4559 * Release the temporary compressed buffer as soon as possible.
4561 if (abl2->b_compress != ZIO_COMPRESS_OFF)
4562 l2arc_release_cdata_buf(ab);
4564 hash_lock = HDR_LOCK(ab);
4565 if (!mutex_tryenter(hash_lock)) {
4567 * This buffer misses out. It may be in a stage
4568 * of eviction. Its ARC_L2_WRITING flag will be
4569 * left set, denying reads to this buffer.
4571 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4575 if (zio->io_error != 0) {
4577 * Error - drop L2ARC entry.
4579 list_remove(buflist, ab);
4580 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4582 kmem_cache_free(l2arc_hdr_cache, abl2);
4583 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4584 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4588 * Allow ARC to begin reads to this L2ARC entry.
4590 ab->b_flags &= ~ARC_L2_WRITING;
4592 mutex_exit(hash_lock);
4595 atomic_inc_64(&l2arc_writes_done);
4596 list_remove(buflist, head);
4597 kmem_cache_free(hdr_cache, head);
4598 mutex_exit(&l2arc_buflist_mtx);
4600 l2arc_do_free_on_write();
4602 kmem_free(cb, sizeof (l2arc_write_callback_t));
4606 * A read to a cache device completed. Validate buffer contents before
4607 * handing over to the regular ARC routines.
4610 l2arc_read_done(zio_t *zio)
4612 l2arc_read_callback_t *cb;
4615 kmutex_t *hash_lock;
4618 ASSERT(zio->io_vd != NULL);
4619 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4621 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4623 cb = zio->io_private;
4625 buf = cb->l2rcb_buf;
4626 ASSERT(buf != NULL);
4628 hash_lock = HDR_LOCK(buf->b_hdr);
4629 mutex_enter(hash_lock);
4631 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4634 * If the buffer was compressed, decompress it first.
4636 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
4637 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
4638 ASSERT(zio->io_data != NULL);
4641 * Check this survived the L2ARC journey.
4643 equal = arc_cksum_equal(buf);
4644 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4645 mutex_exit(hash_lock);
4646 zio->io_private = buf;
4647 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4648 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4651 mutex_exit(hash_lock);
4653 * Buffer didn't survive caching. Increment stats and
4654 * reissue to the original storage device.
4656 if (zio->io_error != 0) {
4657 ARCSTAT_BUMP(arcstat_l2_io_error);
4659 zio->io_error = SET_ERROR(EIO);
4662 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4665 * If there's no waiter, issue an async i/o to the primary
4666 * storage now. If there *is* a waiter, the caller must
4667 * issue the i/o in a context where it's OK to block.
4669 if (zio->io_waiter == NULL) {
4670 zio_t *pio = zio_unique_parent(zio);
4672 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4674 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4675 buf->b_data, zio->io_size, arc_read_done, buf,
4676 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4680 kmem_free(cb, sizeof (l2arc_read_callback_t));
4684 * This is the list priority from which the L2ARC will search for pages to
4685 * cache. This is used within loops (0..3) to cycle through lists in the
4686 * desired order. This order can have a significant effect on cache
4689 * Currently the metadata lists are hit first, MFU then MRU, followed by
4690 * the data lists. This function returns a locked list, and also returns
4694 l2arc_list_locked(int list_num, kmutex_t **lock)
4696 list_t *list = NULL;
4698 ASSERT(list_num >= 0 && list_num <= 3);
4702 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4703 *lock = &arc_mfu->arcs_mtx;
4706 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4707 *lock = &arc_mru->arcs_mtx;
4710 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4711 *lock = &arc_mfu->arcs_mtx;
4714 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4715 *lock = &arc_mru->arcs_mtx;
4719 ASSERT(!(MUTEX_HELD(*lock)));
4725 * Evict buffers from the device write hand to the distance specified in
4726 * bytes. This distance may span populated buffers, it may span nothing.
4727 * This is clearing a region on the L2ARC device ready for writing.
4728 * If the 'all' boolean is set, every buffer is evicted.
4731 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4734 l2arc_buf_hdr_t *abl2;
4735 arc_buf_hdr_t *ab, *ab_prev;
4736 kmutex_t *hash_lock;
4739 buflist = dev->l2ad_buflist;
4741 if (buflist == NULL)
4744 if (!all && dev->l2ad_first) {
4746 * This is the first sweep through the device. There is
4752 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4754 * When nearing the end of the device, evict to the end
4755 * before the device write hand jumps to the start.
4757 taddr = dev->l2ad_end;
4759 taddr = dev->l2ad_hand + distance;
4761 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4762 uint64_t, taddr, boolean_t, all);
4765 mutex_enter(&l2arc_buflist_mtx);
4766 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4767 ab_prev = list_prev(buflist, ab);
4769 hash_lock = HDR_LOCK(ab);
4770 if (!mutex_tryenter(hash_lock)) {
4772 * Missed the hash lock. Retry.
4774 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4775 mutex_exit(&l2arc_buflist_mtx);
4776 mutex_enter(hash_lock);
4777 mutex_exit(hash_lock);
4781 if (HDR_L2_WRITE_HEAD(ab)) {
4783 * We hit a write head node. Leave it for
4784 * l2arc_write_done().
4786 list_remove(buflist, ab);
4787 mutex_exit(hash_lock);
4791 if (!all && ab->b_l2hdr != NULL &&
4792 (ab->b_l2hdr->b_daddr > taddr ||
4793 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4795 * We've evicted to the target address,
4796 * or the end of the device.
4798 mutex_exit(hash_lock);
4802 if (HDR_FREE_IN_PROGRESS(ab)) {
4804 * Already on the path to destruction.
4806 mutex_exit(hash_lock);
4810 if (ab->b_state == arc_l2c_only) {
4811 ASSERT(!HDR_L2_READING(ab));
4813 * This doesn't exist in the ARC. Destroy.
4814 * arc_hdr_destroy() will call list_remove()
4815 * and decrement arcstat_l2_size.
4817 arc_change_state(arc_anon, ab, hash_lock);
4818 arc_hdr_destroy(ab);
4821 * Invalidate issued or about to be issued
4822 * reads, since we may be about to write
4823 * over this location.
4825 if (HDR_L2_READING(ab)) {
4826 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4827 ab->b_flags |= ARC_L2_EVICTED;
4831 * Tell ARC this no longer exists in L2ARC.
4833 if (ab->b_l2hdr != NULL) {
4835 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4837 kmem_cache_free(l2arc_hdr_cache, abl2);
4838 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4839 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4841 list_remove(buflist, ab);
4844 * This may have been leftover after a
4847 ab->b_flags &= ~ARC_L2_WRITING;
4849 mutex_exit(hash_lock);
4851 mutex_exit(&l2arc_buflist_mtx);
4853 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4854 dev->l2ad_evict = taddr;
4858 * Find and write ARC buffers to the L2ARC device.
4860 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4861 * for reading until they have completed writing.
4862 * The headroom_boost is an in-out parameter used to maintain headroom boost
4863 * state between calls to this function.
4865 * Returns the number of bytes actually written (which may be smaller than
4866 * the delta by which the device hand has changed due to alignment).
4869 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
4870 boolean_t *headroom_boost)
4872 arc_buf_hdr_t *ab, *ab_prev, *head;
4874 uint64_t write_asize, write_psize, write_sz, headroom,
4877 kmutex_t *list_lock = NULL;
4879 l2arc_write_callback_t *cb;
4881 uint64_t guid = spa_load_guid(spa);
4883 const boolean_t do_headroom_boost = *headroom_boost;
4885 ASSERT(dev->l2ad_vdev != NULL);
4887 /* Lower the flag now, we might want to raise it again later. */
4888 *headroom_boost = B_FALSE;
4891 write_sz = write_asize = write_psize = 0;
4893 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4894 head->b_flags |= ARC_L2_WRITE_HEAD;
4897 * We will want to try to compress buffers that are at least 2x the
4898 * device sector size.
4900 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
4903 * Copy buffers for L2ARC writing.
4905 mutex_enter(&l2arc_buflist_mtx);
4906 for (try = 0; try <= 3; try++) {
4907 uint64_t passed_sz = 0;
4909 list = l2arc_list_locked(try, &list_lock);
4912 * L2ARC fast warmup.
4914 * Until the ARC is warm and starts to evict, read from the
4915 * head of the ARC lists rather than the tail.
4917 if (arc_warm == B_FALSE)
4918 ab = list_head(list);
4920 ab = list_tail(list);
4922 headroom = target_sz * l2arc_headroom;
4923 if (do_headroom_boost)
4924 headroom = (headroom * l2arc_headroom_boost) / 100;
4926 for (; ab; ab = ab_prev) {
4927 l2arc_buf_hdr_t *l2hdr;
4928 kmutex_t *hash_lock;
4931 if (arc_warm == B_FALSE)
4932 ab_prev = list_next(list, ab);
4934 ab_prev = list_prev(list, ab);
4936 hash_lock = HDR_LOCK(ab);
4937 if (!mutex_tryenter(hash_lock)) {
4939 * Skip this buffer rather than waiting.
4944 passed_sz += ab->b_size;
4945 if (passed_sz > headroom) {
4949 mutex_exit(hash_lock);
4953 if (!l2arc_write_eligible(guid, ab)) {
4954 mutex_exit(hash_lock);
4958 if ((write_sz + ab->b_size) > target_sz) {
4960 mutex_exit(hash_lock);
4966 * Insert a dummy header on the buflist so
4967 * l2arc_write_done() can find where the
4968 * write buffers begin without searching.
4970 list_insert_head(dev->l2ad_buflist, head);
4972 cb = kmem_alloc(sizeof (l2arc_write_callback_t),
4974 cb->l2wcb_dev = dev;
4975 cb->l2wcb_head = head;
4976 pio = zio_root(spa, l2arc_write_done, cb,
4981 * Create and add a new L2ARC header.
4983 l2hdr = kmem_cache_alloc(l2arc_hdr_cache, KM_PUSHPAGE);
4986 arc_space_consume(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4988 ab->b_flags |= ARC_L2_WRITING;
4991 * Temporarily stash the data buffer in b_tmp_cdata.
4992 * The subsequent write step will pick it up from
4993 * there. This is because can't access ab->b_buf
4994 * without holding the hash_lock, which we in turn
4995 * can't access without holding the ARC list locks
4996 * (which we want to avoid during compression/writing)
4998 l2hdr->b_compress = ZIO_COMPRESS_OFF;
4999 l2hdr->b_asize = ab->b_size;
5000 l2hdr->b_tmp_cdata = ab->b_buf->b_data;
5003 buf_sz = ab->b_size;
5004 ab->b_l2hdr = l2hdr;
5006 list_insert_head(dev->l2ad_buflist, ab);
5009 * Compute and store the buffer cksum before
5010 * writing. On debug the cksum is verified first.
5012 arc_cksum_verify(ab->b_buf);
5013 arc_cksum_compute(ab->b_buf, B_TRUE);
5015 mutex_exit(hash_lock);
5020 mutex_exit(list_lock);
5026 /* No buffers selected for writing? */
5029 mutex_exit(&l2arc_buflist_mtx);
5030 kmem_cache_free(hdr_cache, head);
5035 * Now start writing the buffers. We're starting at the write head
5036 * and work backwards, retracing the course of the buffer selector
5039 for (ab = list_prev(dev->l2ad_buflist, head); ab;
5040 ab = list_prev(dev->l2ad_buflist, ab)) {
5041 l2arc_buf_hdr_t *l2hdr;
5045 * We shouldn't need to lock the buffer here, since we flagged
5046 * it as ARC_L2_WRITING in the previous step, but we must take
5047 * care to only access its L2 cache parameters. In particular,
5048 * ab->b_buf may be invalid by now due to ARC eviction.
5050 l2hdr = ab->b_l2hdr;
5051 l2hdr->b_daddr = dev->l2ad_hand;
5053 if (!l2arc_nocompress && (ab->b_flags & ARC_L2COMPRESS) &&
5054 l2hdr->b_asize >= buf_compress_minsz) {
5055 if (l2arc_compress_buf(l2hdr)) {
5057 * If compression succeeded, enable headroom
5058 * boost on the next scan cycle.
5060 *headroom_boost = B_TRUE;
5065 * Pick up the buffer data we had previously stashed away
5066 * (and now potentially also compressed).
5068 buf_data = l2hdr->b_tmp_cdata;
5069 buf_sz = l2hdr->b_asize;
5071 /* Compression may have squashed the buffer to zero length. */
5075 wzio = zio_write_phys(pio, dev->l2ad_vdev,
5076 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
5077 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
5078 ZIO_FLAG_CANFAIL, B_FALSE);
5080 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
5082 (void) zio_nowait(wzio);
5084 write_asize += buf_sz;
5086 * Keep the clock hand suitably device-aligned.
5088 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
5089 write_psize += buf_p_sz;
5090 dev->l2ad_hand += buf_p_sz;
5094 mutex_exit(&l2arc_buflist_mtx);
5096 ASSERT3U(write_asize, <=, target_sz);
5097 ARCSTAT_BUMP(arcstat_l2_writes_sent);
5098 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
5099 ARCSTAT_INCR(arcstat_l2_size, write_sz);
5100 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
5101 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
5104 * Bump device hand to the device start if it is approaching the end.
5105 * l2arc_evict() will already have evicted ahead for this case.
5107 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
5108 vdev_space_update(dev->l2ad_vdev,
5109 dev->l2ad_end - dev->l2ad_hand, 0, 0);
5110 dev->l2ad_hand = dev->l2ad_start;
5111 dev->l2ad_evict = dev->l2ad_start;
5112 dev->l2ad_first = B_FALSE;
5115 dev->l2ad_writing = B_TRUE;
5116 (void) zio_wait(pio);
5117 dev->l2ad_writing = B_FALSE;
5119 return (write_asize);
5123 * Compresses an L2ARC buffer.
5124 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its
5125 * size in l2hdr->b_asize. This routine tries to compress the data and
5126 * depending on the compression result there are three possible outcomes:
5127 * *) The buffer was incompressible. The original l2hdr contents were left
5128 * untouched and are ready for writing to an L2 device.
5129 * *) The buffer was all-zeros, so there is no need to write it to an L2
5130 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
5131 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
5132 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
5133 * data buffer which holds the compressed data to be written, and b_asize
5134 * tells us how much data there is. b_compress is set to the appropriate
5135 * compression algorithm. Once writing is done, invoke
5136 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
5138 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
5139 * buffer was incompressible).
5142 l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr)
5147 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
5148 ASSERT(l2hdr->b_tmp_cdata != NULL);
5150 len = l2hdr->b_asize;
5151 cdata = zio_data_buf_alloc(len);
5152 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata,
5153 cdata, l2hdr->b_asize);
5156 /* zero block, indicate that there's nothing to write */
5157 zio_data_buf_free(cdata, len);
5158 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
5160 l2hdr->b_tmp_cdata = NULL;
5161 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
5163 } else if (csize > 0 && csize < len) {
5165 * Compression succeeded, we'll keep the cdata around for
5166 * writing and release it afterwards.
5168 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
5169 l2hdr->b_asize = csize;
5170 l2hdr->b_tmp_cdata = cdata;
5171 ARCSTAT_BUMP(arcstat_l2_compress_successes);
5175 * Compression failed, release the compressed buffer.
5176 * l2hdr will be left unmodified.
5178 zio_data_buf_free(cdata, len);
5179 ARCSTAT_BUMP(arcstat_l2_compress_failures);
5185 * Decompresses a zio read back from an l2arc device. On success, the
5186 * underlying zio's io_data buffer is overwritten by the uncompressed
5187 * version. On decompression error (corrupt compressed stream), the
5188 * zio->io_error value is set to signal an I/O error.
5190 * Please note that the compressed data stream is not checksummed, so
5191 * if the underlying device is experiencing data corruption, we may feed
5192 * corrupt data to the decompressor, so the decompressor needs to be
5193 * able to handle this situation (LZ4 does).
5196 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
5201 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
5203 if (zio->io_error != 0) {
5205 * An io error has occured, just restore the original io
5206 * size in preparation for a main pool read.
5208 zio->io_orig_size = zio->io_size = hdr->b_size;
5212 if (c == ZIO_COMPRESS_EMPTY) {
5214 * An empty buffer results in a null zio, which means we
5215 * need to fill its io_data after we're done restoring the
5216 * buffer's contents.
5218 ASSERT(hdr->b_buf != NULL);
5219 bzero(hdr->b_buf->b_data, hdr->b_size);
5220 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data;
5222 ASSERT(zio->io_data != NULL);
5224 * We copy the compressed data from the start of the arc buffer
5225 * (the zio_read will have pulled in only what we need, the
5226 * rest is garbage which we will overwrite at decompression)
5227 * and then decompress back to the ARC data buffer. This way we
5228 * can minimize copying by simply decompressing back over the
5229 * original compressed data (rather than decompressing to an
5230 * aux buffer and then copying back the uncompressed buffer,
5231 * which is likely to be much larger).
5233 csize = zio->io_size;
5234 cdata = zio_data_buf_alloc(csize);
5235 bcopy(zio->io_data, cdata, csize);
5236 if (zio_decompress_data(c, cdata, zio->io_data, csize,
5238 zio->io_error = SET_ERROR(EIO);
5239 zio_data_buf_free(cdata, csize);
5242 /* Restore the expected uncompressed IO size. */
5243 zio->io_orig_size = zio->io_size = hdr->b_size;
5247 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
5248 * This buffer serves as a temporary holder of compressed data while
5249 * the buffer entry is being written to an l2arc device. Once that is
5250 * done, we can dispose of it.
5253 l2arc_release_cdata_buf(arc_buf_hdr_t *ab)
5255 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr;
5257 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) {
5259 * If the data was compressed, then we've allocated a
5260 * temporary buffer for it, so now we need to release it.
5262 ASSERT(l2hdr->b_tmp_cdata != NULL);
5263 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size);
5265 l2hdr->b_tmp_cdata = NULL;
5269 * This thread feeds the L2ARC at regular intervals. This is the beating
5270 * heart of the L2ARC.
5273 l2arc_feed_thread(void)
5278 uint64_t size, wrote;
5279 clock_t begin, next = ddi_get_lbolt();
5280 boolean_t headroom_boost = B_FALSE;
5282 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
5284 mutex_enter(&l2arc_feed_thr_lock);
5286 while (l2arc_thread_exit == 0) {
5287 CALLB_CPR_SAFE_BEGIN(&cpr);
5288 (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv,
5289 &l2arc_feed_thr_lock, next);
5290 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
5291 next = ddi_get_lbolt() + hz;
5294 * Quick check for L2ARC devices.
5296 mutex_enter(&l2arc_dev_mtx);
5297 if (l2arc_ndev == 0) {
5298 mutex_exit(&l2arc_dev_mtx);
5301 mutex_exit(&l2arc_dev_mtx);
5302 begin = ddi_get_lbolt();
5305 * This selects the next l2arc device to write to, and in
5306 * doing so the next spa to feed from: dev->l2ad_spa. This
5307 * will return NULL if there are now no l2arc devices or if
5308 * they are all faulted.
5310 * If a device is returned, its spa's config lock is also
5311 * held to prevent device removal. l2arc_dev_get_next()
5312 * will grab and release l2arc_dev_mtx.
5314 if ((dev = l2arc_dev_get_next()) == NULL)
5317 spa = dev->l2ad_spa;
5318 ASSERT(spa != NULL);
5321 * If the pool is read-only then force the feed thread to
5322 * sleep a little longer.
5324 if (!spa_writeable(spa)) {
5325 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
5326 spa_config_exit(spa, SCL_L2ARC, dev);
5331 * Avoid contributing to memory pressure.
5334 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
5335 spa_config_exit(spa, SCL_L2ARC, dev);
5339 ARCSTAT_BUMP(arcstat_l2_feeds);
5341 size = l2arc_write_size();
5344 * Evict L2ARC buffers that will be overwritten.
5346 l2arc_evict(dev, size, B_FALSE);
5349 * Write ARC buffers.
5351 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
5354 * Calculate interval between writes.
5356 next = l2arc_write_interval(begin, size, wrote);
5357 spa_config_exit(spa, SCL_L2ARC, dev);
5360 l2arc_thread_exit = 0;
5361 cv_broadcast(&l2arc_feed_thr_cv);
5362 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
5367 l2arc_vdev_present(vdev_t *vd)
5371 mutex_enter(&l2arc_dev_mtx);
5372 for (dev = list_head(l2arc_dev_list); dev != NULL;
5373 dev = list_next(l2arc_dev_list, dev)) {
5374 if (dev->l2ad_vdev == vd)
5377 mutex_exit(&l2arc_dev_mtx);
5379 return (dev != NULL);
5383 * Add a vdev for use by the L2ARC. By this point the spa has already
5384 * validated the vdev and opened it.
5387 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5389 l2arc_dev_t *adddev;
5391 ASSERT(!l2arc_vdev_present(vd));
5394 * Create a new l2arc device entry.
5396 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5397 adddev->l2ad_spa = spa;
5398 adddev->l2ad_vdev = vd;
5399 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5400 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5401 adddev->l2ad_hand = adddev->l2ad_start;
5402 adddev->l2ad_evict = adddev->l2ad_start;
5403 adddev->l2ad_first = B_TRUE;
5404 adddev->l2ad_writing = B_FALSE;
5405 list_link_init(&adddev->l2ad_node);
5408 * This is a list of all ARC buffers that are still valid on the
5411 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5412 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5413 offsetof(arc_buf_hdr_t, b_l2node));
5415 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5418 * Add device to global list
5420 mutex_enter(&l2arc_dev_mtx);
5421 list_insert_head(l2arc_dev_list, adddev);
5422 atomic_inc_64(&l2arc_ndev);
5423 mutex_exit(&l2arc_dev_mtx);
5427 * Remove a vdev from the L2ARC.
5430 l2arc_remove_vdev(vdev_t *vd)
5432 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5435 * Find the device by vdev
5437 mutex_enter(&l2arc_dev_mtx);
5438 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5439 nextdev = list_next(l2arc_dev_list, dev);
5440 if (vd == dev->l2ad_vdev) {
5445 ASSERT(remdev != NULL);
5448 * Remove device from global list
5450 list_remove(l2arc_dev_list, remdev);
5451 l2arc_dev_last = NULL; /* may have been invalidated */
5452 atomic_dec_64(&l2arc_ndev);
5453 mutex_exit(&l2arc_dev_mtx);
5456 * Clear all buflists and ARC references. L2ARC device flush.
5458 l2arc_evict(remdev, 0, B_TRUE);
5459 list_destroy(remdev->l2ad_buflist);
5460 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5461 kmem_free(remdev, sizeof (l2arc_dev_t));
5467 l2arc_thread_exit = 0;
5469 l2arc_writes_sent = 0;
5470 l2arc_writes_done = 0;
5472 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5473 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5474 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5475 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5476 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5478 l2arc_dev_list = &L2ARC_dev_list;
5479 l2arc_free_on_write = &L2ARC_free_on_write;
5480 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5481 offsetof(l2arc_dev_t, l2ad_node));
5482 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5483 offsetof(l2arc_data_free_t, l2df_list_node));
5490 * This is called from dmu_fini(), which is called from spa_fini();
5491 * Because of this, we can assume that all l2arc devices have
5492 * already been removed when the pools themselves were removed.
5495 l2arc_do_free_on_write();
5497 mutex_destroy(&l2arc_feed_thr_lock);
5498 cv_destroy(&l2arc_feed_thr_cv);
5499 mutex_destroy(&l2arc_dev_mtx);
5500 mutex_destroy(&l2arc_buflist_mtx);
5501 mutex_destroy(&l2arc_free_on_write_mtx);
5503 list_destroy(l2arc_dev_list);
5504 list_destroy(l2arc_free_on_write);
5510 if (!(spa_mode_global & FWRITE))
5513 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5514 TS_RUN, minclsyspri);
5520 if (!(spa_mode_global & FWRITE))
5523 mutex_enter(&l2arc_feed_thr_lock);
5524 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5525 l2arc_thread_exit = 1;
5526 while (l2arc_thread_exit != 0)
5527 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5528 mutex_exit(&l2arc_feed_thr_lock);
5531 #if defined(_KERNEL) && defined(HAVE_SPL)
5532 EXPORT_SYMBOL(arc_read);
5533 EXPORT_SYMBOL(arc_buf_remove_ref);
5534 EXPORT_SYMBOL(arc_buf_info);
5535 EXPORT_SYMBOL(arc_getbuf_func);
5536 EXPORT_SYMBOL(arc_add_prune_callback);
5537 EXPORT_SYMBOL(arc_remove_prune_callback);
5539 module_param(zfs_arc_min, ulong, 0644);
5540 MODULE_PARM_DESC(zfs_arc_min, "Min arc size");
5542 module_param(zfs_arc_max, ulong, 0644);
5543 MODULE_PARM_DESC(zfs_arc_max, "Max arc size");
5545 module_param(zfs_arc_meta_limit, ulong, 0644);
5546 MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size");
5548 module_param(zfs_arc_meta_prune, int, 0644);
5549 MODULE_PARM_DESC(zfs_arc_meta_prune, "Bytes of meta data to prune");
5551 module_param(zfs_arc_grow_retry, int, 0644);
5552 MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size");
5554 module_param(zfs_arc_p_aggressive_disable, int, 0644);
5555 MODULE_PARM_DESC(zfs_arc_p_aggressive_disable, "disable aggressive arc_p grow");
5557 module_param(zfs_arc_shrink_shift, int, 0644);
5558 MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)");
5560 module_param(zfs_disable_dup_eviction, int, 0644);
5561 MODULE_PARM_DESC(zfs_disable_dup_eviction, "disable duplicate buffer eviction");
5563 module_param(zfs_arc_memory_throttle_disable, int, 0644);
5564 MODULE_PARM_DESC(zfs_arc_memory_throttle_disable, "disable memory throttle");
5566 module_param(zfs_arc_min_prefetch_lifespan, int, 0644);
5567 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan, "Min life of prefetch block");
5569 module_param(l2arc_write_max, ulong, 0644);
5570 MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval");
5572 module_param(l2arc_write_boost, ulong, 0644);
5573 MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup");
5575 module_param(l2arc_headroom, ulong, 0644);
5576 MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache");
5578 module_param(l2arc_headroom_boost, ulong, 0644);
5579 MODULE_PARM_DESC(l2arc_headroom_boost, "Compressed l2arc_headroom multiplier");
5581 module_param(l2arc_feed_secs, ulong, 0644);
5582 MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing");
5584 module_param(l2arc_feed_min_ms, ulong, 0644);
5585 MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds");
5587 module_param(l2arc_noprefetch, int, 0644);
5588 MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers");
5590 module_param(l2arc_nocompress, int, 0644);
5591 MODULE_PARM_DESC(l2arc_nocompress, "Skip compressing L2ARC buffers");
5593 module_param(l2arc_feed_again, int, 0644);
5594 MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup");
5596 module_param(l2arc_norw, int, 0644);
5597 MODULE_PARM_DESC(l2arc_norw, "No reads during writes");