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) 2011 by Delphix. All rights reserved.
28 * DVA-based Adjustable Replacement Cache
30 * While much of the theory of operation used here is
31 * based on the self-tuning, low overhead replacement cache
32 * presented by Megiddo and Modha at FAST 2003, there are some
33 * significant differences:
35 * 1. The Megiddo and Modha model assumes any page is evictable.
36 * Pages in its cache cannot be "locked" into memory. This makes
37 * the eviction algorithm simple: evict the last page in the list.
38 * This also make the performance characteristics easy to reason
39 * about. Our cache is not so simple. At any given moment, some
40 * subset of the blocks in the cache are un-evictable because we
41 * have handed out a reference to them. Blocks are only evictable
42 * when there are no external references active. This makes
43 * eviction far more problematic: we choose to evict the evictable
44 * blocks that are the "lowest" in the list.
46 * There are times when it is not possible to evict the requested
47 * space. In these circumstances we are unable to adjust the cache
48 * size. To prevent the cache growing unbounded at these times we
49 * implement a "cache throttle" that slows the flow of new data
50 * into the cache until we can make space available.
52 * 2. The Megiddo and Modha model assumes a fixed cache size.
53 * Pages are evicted when the cache is full and there is a cache
54 * miss. Our model has a variable sized cache. It grows with
55 * high use, but also tries to react to memory pressure from the
56 * operating system: decreasing its size when system memory is
59 * 3. The Megiddo and Modha model assumes a fixed page size. All
60 * elements of the cache are therefor exactly the same size. So
61 * when adjusting the cache size following a cache miss, its simply
62 * a matter of choosing a single page to evict. In our model, we
63 * have variable sized cache blocks (rangeing from 512 bytes to
64 * 128K bytes). We therefor choose a set of blocks to evict to make
65 * space for a cache miss that approximates as closely as possible
66 * the space used by the new block.
68 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
69 * by N. Megiddo & D. Modha, FAST 2003
75 * A new reference to a cache buffer can be obtained in two
76 * ways: 1) via a hash table lookup using the DVA as a key,
77 * or 2) via one of the ARC lists. The arc_read() interface
78 * uses method 1, while the internal arc algorithms for
79 * adjusting the cache use method 2. We therefor provide two
80 * types of locks: 1) the hash table lock array, and 2) the
83 * Buffers do not have their own mutexs, rather they rely on the
84 * hash table mutexs for the bulk of their protection (i.e. most
85 * fields in the arc_buf_hdr_t are protected by these mutexs).
87 * buf_hash_find() returns the appropriate mutex (held) when it
88 * locates the requested buffer in the hash table. It returns
89 * NULL for the mutex if the buffer was not in the table.
91 * buf_hash_remove() expects the appropriate hash mutex to be
92 * already held before it is invoked.
94 * Each arc state also has a mutex which is used to protect the
95 * buffer list associated with the state. When attempting to
96 * obtain a hash table lock while holding an arc list lock you
97 * must use: mutex_tryenter() to avoid deadlock. Also note that
98 * the active state mutex must be held before the ghost state mutex.
100 * Arc buffers may have an associated eviction callback function.
101 * This function will be invoked prior to removing the buffer (e.g.
102 * in arc_do_user_evicts()). Note however that the data associated
103 * with the buffer may be evicted prior to the callback. The callback
104 * must be made with *no locks held* (to prevent deadlock). Additionally,
105 * the users of callbacks must ensure that their private data is
106 * protected from simultaneous callbacks from arc_buf_evict()
107 * and arc_do_user_evicts().
109 * It as also possible to register a callback which is run when the
110 * arc_meta_limit is reached and no buffers can be safely evicted. In
111 * this case the arc user should drop a reference on some arc buffers so
112 * they can be reclaimed and the arc_meta_limit honored. For example,
113 * when using the ZPL each dentry holds a references on a znode. These
114 * dentries must be pruned before the arc buffer holding the znode can
117 * Note that the majority of the performance stats are manipulated
118 * with atomic operations.
120 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
122 * - L2ARC buflist creation
123 * - L2ARC buflist eviction
124 * - L2ARC write completion, which walks L2ARC buflists
125 * - ARC header destruction, as it removes from L2ARC buflists
126 * - ARC header release, as it removes from L2ARC buflists
131 #include <sys/zfs_context.h>
133 #include <sys/vdev.h>
134 #include <sys/vdev_impl.h>
136 #include <sys/vmsystm.h>
138 #include <sys/fs/swapnode.h>
141 #include <sys/callb.h>
142 #include <sys/kstat.h>
143 #include <sys/dmu_tx.h>
144 #include <zfs_fletcher.h>
146 static kmutex_t arc_reclaim_thr_lock;
147 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
148 static uint8_t arc_thread_exit;
150 /* number of bytes to prune from caches when at arc_meta_limit is reached */
151 uint_t arc_meta_prune = 1048576;
153 typedef enum arc_reclaim_strategy {
154 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
155 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
156 } arc_reclaim_strategy_t;
158 /* number of seconds before growing cache again */
159 static int arc_grow_retry = 5;
161 /* expiration time for arc_no_grow */
162 static clock_t arc_grow_time = 0;
164 /* shift of arc_c for calculating both min and max arc_p */
165 static int arc_p_min_shift = 4;
167 /* log2(fraction of arc to reclaim) */
168 static int arc_shrink_shift = 5;
171 * minimum lifespan of a prefetch block in clock ticks
172 * (initialized in arc_init())
174 static int arc_min_prefetch_lifespan;
179 * The arc has filled available memory and has now warmed up.
181 static boolean_t arc_warm;
184 * These tunables are for performance analysis.
186 unsigned long zfs_arc_max = 0;
187 unsigned long zfs_arc_min = 0;
188 unsigned long zfs_arc_meta_limit = 0;
189 int zfs_arc_grow_retry = 0;
190 int zfs_arc_shrink_shift = 0;
191 int zfs_arc_p_min_shift = 0;
192 int zfs_arc_meta_prune = 0;
195 * Note that buffers can be in one of 6 states:
196 * ARC_anon - anonymous (discussed below)
197 * ARC_mru - recently used, currently cached
198 * ARC_mru_ghost - recentely used, no longer in cache
199 * ARC_mfu - frequently used, currently cached
200 * ARC_mfu_ghost - frequently used, no longer in cache
201 * ARC_l2c_only - exists in L2ARC but not other states
202 * When there are no active references to the buffer, they are
203 * are linked onto a list in one of these arc states. These are
204 * the only buffers that can be evicted or deleted. Within each
205 * state there are multiple lists, one for meta-data and one for
206 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
207 * etc.) is tracked separately so that it can be managed more
208 * explicitly: favored over data, limited explicitly.
210 * Anonymous buffers are buffers that are not associated with
211 * a DVA. These are buffers that hold dirty block copies
212 * before they are written to stable storage. By definition,
213 * they are "ref'd" and are considered part of arc_mru
214 * that cannot be freed. Generally, they will aquire a DVA
215 * as they are written and migrate onto the arc_mru list.
217 * The ARC_l2c_only state is for buffers that are in the second
218 * level ARC but no longer in any of the ARC_m* lists. The second
219 * level ARC itself may also contain buffers that are in any of
220 * the ARC_m* states - meaning that a buffer can exist in two
221 * places. The reason for the ARC_l2c_only state is to keep the
222 * buffer header in the hash table, so that reads that hit the
223 * second level ARC benefit from these fast lookups.
226 typedef struct arc_state {
227 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
228 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
229 uint64_t arcs_size; /* total amount of data in this state */
234 static arc_state_t ARC_anon;
235 static arc_state_t ARC_mru;
236 static arc_state_t ARC_mru_ghost;
237 static arc_state_t ARC_mfu;
238 static arc_state_t ARC_mfu_ghost;
239 static arc_state_t ARC_l2c_only;
241 typedef struct arc_stats {
242 kstat_named_t arcstat_hits;
243 kstat_named_t arcstat_misses;
244 kstat_named_t arcstat_demand_data_hits;
245 kstat_named_t arcstat_demand_data_misses;
246 kstat_named_t arcstat_demand_metadata_hits;
247 kstat_named_t arcstat_demand_metadata_misses;
248 kstat_named_t arcstat_prefetch_data_hits;
249 kstat_named_t arcstat_prefetch_data_misses;
250 kstat_named_t arcstat_prefetch_metadata_hits;
251 kstat_named_t arcstat_prefetch_metadata_misses;
252 kstat_named_t arcstat_mru_hits;
253 kstat_named_t arcstat_mru_ghost_hits;
254 kstat_named_t arcstat_mfu_hits;
255 kstat_named_t arcstat_mfu_ghost_hits;
256 kstat_named_t arcstat_deleted;
257 kstat_named_t arcstat_recycle_miss;
258 kstat_named_t arcstat_mutex_miss;
259 kstat_named_t arcstat_evict_skip;
260 kstat_named_t arcstat_evict_l2_cached;
261 kstat_named_t arcstat_evict_l2_eligible;
262 kstat_named_t arcstat_evict_l2_ineligible;
263 kstat_named_t arcstat_hash_elements;
264 kstat_named_t arcstat_hash_elements_max;
265 kstat_named_t arcstat_hash_collisions;
266 kstat_named_t arcstat_hash_chains;
267 kstat_named_t arcstat_hash_chain_max;
268 kstat_named_t arcstat_p;
269 kstat_named_t arcstat_c;
270 kstat_named_t arcstat_c_min;
271 kstat_named_t arcstat_c_max;
272 kstat_named_t arcstat_size;
273 kstat_named_t arcstat_hdr_size;
274 kstat_named_t arcstat_data_size;
275 kstat_named_t arcstat_other_size;
276 kstat_named_t arcstat_anon_size;
277 kstat_named_t arcstat_anon_evict_data;
278 kstat_named_t arcstat_anon_evict_metadata;
279 kstat_named_t arcstat_mru_size;
280 kstat_named_t arcstat_mru_evict_data;
281 kstat_named_t arcstat_mru_evict_metadata;
282 kstat_named_t arcstat_mru_ghost_size;
283 kstat_named_t arcstat_mru_ghost_evict_data;
284 kstat_named_t arcstat_mru_ghost_evict_metadata;
285 kstat_named_t arcstat_mfu_size;
286 kstat_named_t arcstat_mfu_evict_data;
287 kstat_named_t arcstat_mfu_evict_metadata;
288 kstat_named_t arcstat_mfu_ghost_size;
289 kstat_named_t arcstat_mfu_ghost_evict_data;
290 kstat_named_t arcstat_mfu_ghost_evict_metadata;
291 kstat_named_t arcstat_l2_hits;
292 kstat_named_t arcstat_l2_misses;
293 kstat_named_t arcstat_l2_feeds;
294 kstat_named_t arcstat_l2_rw_clash;
295 kstat_named_t arcstat_l2_read_bytes;
296 kstat_named_t arcstat_l2_write_bytes;
297 kstat_named_t arcstat_l2_writes_sent;
298 kstat_named_t arcstat_l2_writes_done;
299 kstat_named_t arcstat_l2_writes_error;
300 kstat_named_t arcstat_l2_writes_hdr_miss;
301 kstat_named_t arcstat_l2_evict_lock_retry;
302 kstat_named_t arcstat_l2_evict_reading;
303 kstat_named_t arcstat_l2_free_on_write;
304 kstat_named_t arcstat_l2_abort_lowmem;
305 kstat_named_t arcstat_l2_cksum_bad;
306 kstat_named_t arcstat_l2_io_error;
307 kstat_named_t arcstat_l2_size;
308 kstat_named_t arcstat_l2_hdr_size;
309 kstat_named_t arcstat_memory_throttle_count;
310 kstat_named_t arcstat_memory_direct_count;
311 kstat_named_t arcstat_memory_indirect_count;
312 kstat_named_t arcstat_no_grow;
313 kstat_named_t arcstat_tempreserve;
314 kstat_named_t arcstat_loaned_bytes;
315 kstat_named_t arcstat_prune;
316 kstat_named_t arcstat_meta_used;
317 kstat_named_t arcstat_meta_limit;
318 kstat_named_t arcstat_meta_max;
321 static arc_stats_t arc_stats = {
322 { "hits", KSTAT_DATA_UINT64 },
323 { "misses", KSTAT_DATA_UINT64 },
324 { "demand_data_hits", KSTAT_DATA_UINT64 },
325 { "demand_data_misses", KSTAT_DATA_UINT64 },
326 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
327 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
328 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
329 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
330 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
331 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
332 { "mru_hits", KSTAT_DATA_UINT64 },
333 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
334 { "mfu_hits", KSTAT_DATA_UINT64 },
335 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
336 { "deleted", KSTAT_DATA_UINT64 },
337 { "recycle_miss", KSTAT_DATA_UINT64 },
338 { "mutex_miss", KSTAT_DATA_UINT64 },
339 { "evict_skip", KSTAT_DATA_UINT64 },
340 { "evict_l2_cached", KSTAT_DATA_UINT64 },
341 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
342 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
343 { "hash_elements", KSTAT_DATA_UINT64 },
344 { "hash_elements_max", KSTAT_DATA_UINT64 },
345 { "hash_collisions", KSTAT_DATA_UINT64 },
346 { "hash_chains", KSTAT_DATA_UINT64 },
347 { "hash_chain_max", KSTAT_DATA_UINT64 },
348 { "p", KSTAT_DATA_UINT64 },
349 { "c", KSTAT_DATA_UINT64 },
350 { "c_min", KSTAT_DATA_UINT64 },
351 { "c_max", KSTAT_DATA_UINT64 },
352 { "size", KSTAT_DATA_UINT64 },
353 { "hdr_size", KSTAT_DATA_UINT64 },
354 { "data_size", KSTAT_DATA_UINT64 },
355 { "other_size", KSTAT_DATA_UINT64 },
356 { "anon_size", KSTAT_DATA_UINT64 },
357 { "anon_evict_data", KSTAT_DATA_UINT64 },
358 { "anon_evict_metadata", KSTAT_DATA_UINT64 },
359 { "mru_size", KSTAT_DATA_UINT64 },
360 { "mru_evict_data", KSTAT_DATA_UINT64 },
361 { "mru_evict_metadata", KSTAT_DATA_UINT64 },
362 { "mru_ghost_size", KSTAT_DATA_UINT64 },
363 { "mru_ghost_evict_data", KSTAT_DATA_UINT64 },
364 { "mru_ghost_evict_metadata", KSTAT_DATA_UINT64 },
365 { "mfu_size", KSTAT_DATA_UINT64 },
366 { "mfu_evict_data", KSTAT_DATA_UINT64 },
367 { "mfu_evict_metadata", KSTAT_DATA_UINT64 },
368 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
369 { "mfu_ghost_evict_data", KSTAT_DATA_UINT64 },
370 { "mfu_ghost_evict_metadata", KSTAT_DATA_UINT64 },
371 { "l2_hits", KSTAT_DATA_UINT64 },
372 { "l2_misses", KSTAT_DATA_UINT64 },
373 { "l2_feeds", KSTAT_DATA_UINT64 },
374 { "l2_rw_clash", KSTAT_DATA_UINT64 },
375 { "l2_read_bytes", KSTAT_DATA_UINT64 },
376 { "l2_write_bytes", KSTAT_DATA_UINT64 },
377 { "l2_writes_sent", KSTAT_DATA_UINT64 },
378 { "l2_writes_done", KSTAT_DATA_UINT64 },
379 { "l2_writes_error", KSTAT_DATA_UINT64 },
380 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
381 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
382 { "l2_evict_reading", KSTAT_DATA_UINT64 },
383 { "l2_free_on_write", KSTAT_DATA_UINT64 },
384 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
385 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
386 { "l2_io_error", KSTAT_DATA_UINT64 },
387 { "l2_size", KSTAT_DATA_UINT64 },
388 { "l2_hdr_size", KSTAT_DATA_UINT64 },
389 { "memory_throttle_count", KSTAT_DATA_UINT64 },
390 { "memory_direct_count", KSTAT_DATA_UINT64 },
391 { "memory_indirect_count", KSTAT_DATA_UINT64 },
392 { "arc_no_grow", KSTAT_DATA_UINT64 },
393 { "arc_tempreserve", KSTAT_DATA_UINT64 },
394 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
395 { "arc_prune", KSTAT_DATA_UINT64 },
396 { "arc_meta_used", KSTAT_DATA_UINT64 },
397 { "arc_meta_limit", KSTAT_DATA_UINT64 },
398 { "arc_meta_max", KSTAT_DATA_UINT64 },
401 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
403 #define ARCSTAT_INCR(stat, val) \
404 atomic_add_64(&arc_stats.stat.value.ui64, (val));
406 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
407 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
409 #define ARCSTAT_MAX(stat, val) { \
411 while ((val) > (m = arc_stats.stat.value.ui64) && \
412 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
416 #define ARCSTAT_MAXSTAT(stat) \
417 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
420 * We define a macro to allow ARC hits/misses to be easily broken down by
421 * two separate conditions, giving a total of four different subtypes for
422 * each of hits and misses (so eight statistics total).
424 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
427 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
429 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
433 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
435 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
440 static arc_state_t *arc_anon;
441 static arc_state_t *arc_mru;
442 static arc_state_t *arc_mru_ghost;
443 static arc_state_t *arc_mfu;
444 static arc_state_t *arc_mfu_ghost;
445 static arc_state_t *arc_l2c_only;
448 * There are several ARC variables that are critical to export as kstats --
449 * but we don't want to have to grovel around in the kstat whenever we wish to
450 * manipulate them. For these variables, we therefore define them to be in
451 * terms of the statistic variable. This assures that we are not introducing
452 * the possibility of inconsistency by having shadow copies of the variables,
453 * while still allowing the code to be readable.
455 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
456 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
457 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
458 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
459 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
460 #define arc_no_grow ARCSTAT(arcstat_no_grow)
461 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
462 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
463 #define arc_meta_used ARCSTAT(arcstat_meta_used)
464 #define arc_meta_limit ARCSTAT(arcstat_meta_limit)
465 #define arc_meta_max ARCSTAT(arcstat_meta_max)
467 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
469 typedef struct arc_callback arc_callback_t;
471 struct arc_callback {
473 arc_done_func_t *acb_done;
475 zio_t *acb_zio_dummy;
476 arc_callback_t *acb_next;
479 typedef struct arc_write_callback arc_write_callback_t;
481 struct arc_write_callback {
483 arc_done_func_t *awcb_ready;
484 arc_done_func_t *awcb_done;
489 /* protected by hash lock */
494 kmutex_t b_freeze_lock;
495 zio_cksum_t *b_freeze_cksum;
498 arc_buf_hdr_t *b_hash_next;
503 arc_callback_t *b_acb;
507 arc_buf_contents_t b_type;
511 /* protected by arc state mutex */
512 arc_state_t *b_state;
513 list_node_t b_arc_node;
515 /* updated atomically */
516 clock_t b_arc_access;
518 /* self protecting */
521 l2arc_buf_hdr_t *b_l2hdr;
522 list_node_t b_l2node;
525 static list_t arc_prune_list;
526 static kmutex_t arc_prune_mtx;
527 static arc_buf_t *arc_eviction_list;
528 static kmutex_t arc_eviction_mtx;
529 static arc_buf_hdr_t arc_eviction_hdr;
530 static void arc_get_data_buf(arc_buf_t *buf);
531 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
532 static int arc_evict_needed(arc_buf_contents_t type);
533 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
535 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
537 #define GHOST_STATE(state) \
538 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
539 (state) == arc_l2c_only)
542 * Private ARC flags. These flags are private ARC only flags that will show up
543 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
544 * be passed in as arc_flags in things like arc_read. However, these flags
545 * should never be passed and should only be set by ARC code. When adding new
546 * public flags, make sure not to smash the private ones.
549 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
550 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
551 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
552 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
553 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
554 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
555 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
556 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
557 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
558 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
560 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
561 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
562 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
563 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
564 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
565 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
566 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
567 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
568 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
569 (hdr)->b_l2hdr != NULL)
570 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
571 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
572 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
578 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
579 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
582 * Hash table routines
585 #define HT_LOCK_ALIGN 64
586 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
591 unsigned char pad[HT_LOCK_PAD];
595 #define BUF_LOCKS 256
596 typedef struct buf_hash_table {
598 arc_buf_hdr_t **ht_table;
599 struct ht_lock ht_locks[BUF_LOCKS];
602 static buf_hash_table_t buf_hash_table;
604 #define BUF_HASH_INDEX(spa, dva, birth) \
605 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
606 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
607 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
608 #define HDR_LOCK(hdr) \
609 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
611 uint64_t zfs_crc64_table[256];
617 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
618 #define L2ARC_HEADROOM 2 /* num of writes */
619 #define L2ARC_FEED_SECS 1 /* caching interval secs */
620 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
622 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
623 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
626 * L2ARC Performance Tunables
628 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
629 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
630 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
631 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
632 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
633 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
634 int l2arc_feed_again = B_TRUE; /* turbo warmup */
635 int l2arc_norw = B_TRUE; /* no reads during writes */
640 typedef struct l2arc_dev {
641 vdev_t *l2ad_vdev; /* vdev */
642 spa_t *l2ad_spa; /* spa */
643 uint64_t l2ad_hand; /* next write location */
644 uint64_t l2ad_write; /* desired write size, bytes */
645 uint64_t l2ad_boost; /* warmup write boost, bytes */
646 uint64_t l2ad_start; /* first addr on device */
647 uint64_t l2ad_end; /* last addr on device */
648 uint64_t l2ad_evict; /* last addr eviction reached */
649 boolean_t l2ad_first; /* first sweep through */
650 boolean_t l2ad_writing; /* currently writing */
651 list_t *l2ad_buflist; /* buffer list */
652 list_node_t l2ad_node; /* device list node */
655 static list_t L2ARC_dev_list; /* device list */
656 static list_t *l2arc_dev_list; /* device list pointer */
657 static kmutex_t l2arc_dev_mtx; /* device list mutex */
658 static l2arc_dev_t *l2arc_dev_last; /* last device used */
659 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
660 static list_t L2ARC_free_on_write; /* free after write buf list */
661 static list_t *l2arc_free_on_write; /* free after write list ptr */
662 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
663 static uint64_t l2arc_ndev; /* number of devices */
665 typedef struct l2arc_read_callback {
666 arc_buf_t *l2rcb_buf; /* read buffer */
667 spa_t *l2rcb_spa; /* spa */
668 blkptr_t l2rcb_bp; /* original blkptr */
669 zbookmark_t l2rcb_zb; /* original bookmark */
670 int l2rcb_flags; /* original flags */
671 } l2arc_read_callback_t;
673 typedef struct l2arc_write_callback {
674 l2arc_dev_t *l2wcb_dev; /* device info */
675 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
676 } l2arc_write_callback_t;
678 struct l2arc_buf_hdr {
679 /* protected by arc_buf_hdr mutex */
680 l2arc_dev_t *b_dev; /* L2ARC device */
681 uint64_t b_daddr; /* disk address, offset byte */
684 typedef struct l2arc_data_free {
685 /* protected by l2arc_free_on_write_mtx */
688 void (*l2df_func)(void *, size_t);
689 list_node_t l2df_list_node;
692 static kmutex_t l2arc_feed_thr_lock;
693 static kcondvar_t l2arc_feed_thr_cv;
694 static uint8_t l2arc_thread_exit;
696 static void l2arc_read_done(zio_t *zio);
697 static void l2arc_hdr_stat_add(void);
698 static void l2arc_hdr_stat_remove(void);
701 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
703 uint8_t *vdva = (uint8_t *)dva;
704 uint64_t crc = -1ULL;
707 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
709 for (i = 0; i < sizeof (dva_t); i++)
710 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
712 crc ^= (spa>>8) ^ birth;
717 #define BUF_EMPTY(buf) \
718 ((buf)->b_dva.dva_word[0] == 0 && \
719 (buf)->b_dva.dva_word[1] == 0 && \
722 #define BUF_EQUAL(spa, dva, birth, buf) \
723 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
724 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
725 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
728 buf_discard_identity(arc_buf_hdr_t *hdr)
730 hdr->b_dva.dva_word[0] = 0;
731 hdr->b_dva.dva_word[1] = 0;
736 static arc_buf_hdr_t *
737 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
739 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
740 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
743 mutex_enter(hash_lock);
744 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
745 buf = buf->b_hash_next) {
746 if (BUF_EQUAL(spa, dva, birth, buf)) {
751 mutex_exit(hash_lock);
757 * Insert an entry into the hash table. If there is already an element
758 * equal to elem in the hash table, then the already existing element
759 * will be returned and the new element will not be inserted.
760 * Otherwise returns NULL.
762 static arc_buf_hdr_t *
763 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
765 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
766 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
770 ASSERT(!HDR_IN_HASH_TABLE(buf));
772 mutex_enter(hash_lock);
773 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
774 fbuf = fbuf->b_hash_next, i++) {
775 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
779 buf->b_hash_next = buf_hash_table.ht_table[idx];
780 buf_hash_table.ht_table[idx] = buf;
781 buf->b_flags |= ARC_IN_HASH_TABLE;
783 /* collect some hash table performance data */
785 ARCSTAT_BUMP(arcstat_hash_collisions);
787 ARCSTAT_BUMP(arcstat_hash_chains);
789 ARCSTAT_MAX(arcstat_hash_chain_max, i);
792 ARCSTAT_BUMP(arcstat_hash_elements);
793 ARCSTAT_MAXSTAT(arcstat_hash_elements);
799 buf_hash_remove(arc_buf_hdr_t *buf)
801 arc_buf_hdr_t *fbuf, **bufp;
802 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
804 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
805 ASSERT(HDR_IN_HASH_TABLE(buf));
807 bufp = &buf_hash_table.ht_table[idx];
808 while ((fbuf = *bufp) != buf) {
809 ASSERT(fbuf != NULL);
810 bufp = &fbuf->b_hash_next;
812 *bufp = buf->b_hash_next;
813 buf->b_hash_next = NULL;
814 buf->b_flags &= ~ARC_IN_HASH_TABLE;
816 /* collect some hash table performance data */
817 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
819 if (buf_hash_table.ht_table[idx] &&
820 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
821 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
825 * Global data structures and functions for the buf kmem cache.
827 static kmem_cache_t *hdr_cache;
828 static kmem_cache_t *buf_cache;
835 #if defined(_KERNEL) && defined(HAVE_SPL)
836 /* Large allocations which do not require contiguous pages
837 * should be using vmem_free() in the linux kernel */
838 vmem_free(buf_hash_table.ht_table,
839 (buf_hash_table.ht_mask + 1) * sizeof (void *));
841 kmem_free(buf_hash_table.ht_table,
842 (buf_hash_table.ht_mask + 1) * sizeof (void *));
844 for (i = 0; i < BUF_LOCKS; i++)
845 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
846 kmem_cache_destroy(hdr_cache);
847 kmem_cache_destroy(buf_cache);
851 * Constructor callback - called when the cache is empty
852 * and a new buf is requested.
856 hdr_cons(void *vbuf, void *unused, int kmflag)
858 arc_buf_hdr_t *buf = vbuf;
860 bzero(buf, sizeof (arc_buf_hdr_t));
861 refcount_create(&buf->b_refcnt);
862 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
863 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
864 list_link_init(&buf->b_arc_node);
865 list_link_init(&buf->b_l2node);
866 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
873 buf_cons(void *vbuf, void *unused, int kmflag)
875 arc_buf_t *buf = vbuf;
877 bzero(buf, sizeof (arc_buf_t));
878 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
879 rw_init(&buf->b_data_lock, NULL, RW_DEFAULT, NULL);
880 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
886 * Destructor callback - called when a cached buf is
887 * no longer required.
891 hdr_dest(void *vbuf, void *unused)
893 arc_buf_hdr_t *buf = vbuf;
895 ASSERT(BUF_EMPTY(buf));
896 refcount_destroy(&buf->b_refcnt);
897 cv_destroy(&buf->b_cv);
898 mutex_destroy(&buf->b_freeze_lock);
899 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
904 buf_dest(void *vbuf, void *unused)
906 arc_buf_t *buf = vbuf;
908 mutex_destroy(&buf->b_evict_lock);
909 rw_destroy(&buf->b_data_lock);
910 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
917 uint64_t hsize = 1ULL << 12;
921 * The hash table is big enough to fill all of physical memory
922 * with an average 64K block size. The table will take up
923 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
925 while (hsize * 65536 < physmem * PAGESIZE)
928 buf_hash_table.ht_mask = hsize - 1;
929 #if defined(_KERNEL) && defined(HAVE_SPL)
930 /* Large allocations which do not require contiguous pages
931 * should be using vmem_alloc() in the linux kernel */
932 buf_hash_table.ht_table =
933 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
935 buf_hash_table.ht_table =
936 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
938 if (buf_hash_table.ht_table == NULL) {
939 ASSERT(hsize > (1ULL << 8));
944 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
945 0, hdr_cons, hdr_dest, NULL, NULL, NULL, 0);
946 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
947 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
949 for (i = 0; i < 256; i++)
950 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
951 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
953 for (i = 0; i < BUF_LOCKS; i++) {
954 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
955 NULL, MUTEX_DEFAULT, NULL);
959 #define ARC_MINTIME (hz>>4) /* 62 ms */
962 arc_cksum_verify(arc_buf_t *buf)
966 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
969 mutex_enter(&buf->b_hdr->b_freeze_lock);
970 if (buf->b_hdr->b_freeze_cksum == NULL ||
971 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
972 mutex_exit(&buf->b_hdr->b_freeze_lock);
975 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
976 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
977 panic("buffer modified while frozen!");
978 mutex_exit(&buf->b_hdr->b_freeze_lock);
982 arc_cksum_equal(arc_buf_t *buf)
987 mutex_enter(&buf->b_hdr->b_freeze_lock);
988 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
989 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
990 mutex_exit(&buf->b_hdr->b_freeze_lock);
996 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
998 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1001 mutex_enter(&buf->b_hdr->b_freeze_lock);
1002 if (buf->b_hdr->b_freeze_cksum != NULL) {
1003 mutex_exit(&buf->b_hdr->b_freeze_lock);
1006 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1008 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1009 buf->b_hdr->b_freeze_cksum);
1010 mutex_exit(&buf->b_hdr->b_freeze_lock);
1014 arc_buf_thaw(arc_buf_t *buf)
1016 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1017 if (buf->b_hdr->b_state != arc_anon)
1018 panic("modifying non-anon buffer!");
1019 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1020 panic("modifying buffer while i/o in progress!");
1021 arc_cksum_verify(buf);
1024 mutex_enter(&buf->b_hdr->b_freeze_lock);
1025 if (buf->b_hdr->b_freeze_cksum != NULL) {
1026 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1027 buf->b_hdr->b_freeze_cksum = NULL;
1030 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1031 if (buf->b_hdr->b_thawed)
1032 kmem_free(buf->b_hdr->b_thawed, 1);
1033 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1036 mutex_exit(&buf->b_hdr->b_freeze_lock);
1040 arc_buf_freeze(arc_buf_t *buf)
1042 kmutex_t *hash_lock;
1044 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1047 hash_lock = HDR_LOCK(buf->b_hdr);
1048 mutex_enter(hash_lock);
1050 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1051 buf->b_hdr->b_state == arc_anon);
1052 arc_cksum_compute(buf, B_FALSE);
1053 mutex_exit(hash_lock);
1057 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1059 ASSERT(MUTEX_HELD(hash_lock));
1061 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1062 (ab->b_state != arc_anon)) {
1063 uint64_t delta = ab->b_size * ab->b_datacnt;
1064 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1065 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1067 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1068 mutex_enter(&ab->b_state->arcs_mtx);
1069 ASSERT(list_link_active(&ab->b_arc_node));
1070 list_remove(list, ab);
1071 if (GHOST_STATE(ab->b_state)) {
1072 ASSERT3U(ab->b_datacnt, ==, 0);
1073 ASSERT3P(ab->b_buf, ==, NULL);
1077 ASSERT3U(*size, >=, delta);
1078 atomic_add_64(size, -delta);
1079 mutex_exit(&ab->b_state->arcs_mtx);
1080 /* remove the prefetch flag if we get a reference */
1081 if (ab->b_flags & ARC_PREFETCH)
1082 ab->b_flags &= ~ARC_PREFETCH;
1087 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1090 arc_state_t *state = ab->b_state;
1092 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1093 ASSERT(!GHOST_STATE(state));
1095 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1096 (state != arc_anon)) {
1097 uint64_t *size = &state->arcs_lsize[ab->b_type];
1099 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1100 mutex_enter(&state->arcs_mtx);
1101 ASSERT(!list_link_active(&ab->b_arc_node));
1102 list_insert_head(&state->arcs_list[ab->b_type], ab);
1103 ASSERT(ab->b_datacnt > 0);
1104 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1105 mutex_exit(&state->arcs_mtx);
1111 * Move the supplied buffer to the indicated state. The mutex
1112 * for the buffer must be held by the caller.
1115 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1117 arc_state_t *old_state = ab->b_state;
1118 int64_t refcnt = refcount_count(&ab->b_refcnt);
1119 uint64_t from_delta, to_delta;
1121 ASSERT(MUTEX_HELD(hash_lock));
1122 ASSERT(new_state != old_state);
1123 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1124 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1125 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1127 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1130 * If this buffer is evictable, transfer it from the
1131 * old state list to the new state list.
1134 if (old_state != arc_anon) {
1135 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1136 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1139 mutex_enter(&old_state->arcs_mtx);
1141 ASSERT(list_link_active(&ab->b_arc_node));
1142 list_remove(&old_state->arcs_list[ab->b_type], ab);
1145 * If prefetching out of the ghost cache,
1146 * we will have a non-zero datacnt.
1148 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1149 /* ghost elements have a ghost size */
1150 ASSERT(ab->b_buf == NULL);
1151 from_delta = ab->b_size;
1153 ASSERT3U(*size, >=, from_delta);
1154 atomic_add_64(size, -from_delta);
1157 mutex_exit(&old_state->arcs_mtx);
1159 if (new_state != arc_anon) {
1160 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1161 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1164 mutex_enter(&new_state->arcs_mtx);
1166 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1168 /* ghost elements have a ghost size */
1169 if (GHOST_STATE(new_state)) {
1170 ASSERT(ab->b_datacnt == 0);
1171 ASSERT(ab->b_buf == NULL);
1172 to_delta = ab->b_size;
1174 atomic_add_64(size, to_delta);
1177 mutex_exit(&new_state->arcs_mtx);
1181 ASSERT(!BUF_EMPTY(ab));
1182 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1183 buf_hash_remove(ab);
1185 /* adjust state sizes */
1187 atomic_add_64(&new_state->arcs_size, to_delta);
1189 ASSERT3U(old_state->arcs_size, >=, from_delta);
1190 atomic_add_64(&old_state->arcs_size, -from_delta);
1192 ab->b_state = new_state;
1194 /* adjust l2arc hdr stats */
1195 if (new_state == arc_l2c_only)
1196 l2arc_hdr_stat_add();
1197 else if (old_state == arc_l2c_only)
1198 l2arc_hdr_stat_remove();
1202 arc_space_consume(uint64_t space, arc_space_type_t type)
1204 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1209 case ARC_SPACE_DATA:
1210 ARCSTAT_INCR(arcstat_data_size, space);
1212 case ARC_SPACE_OTHER:
1213 ARCSTAT_INCR(arcstat_other_size, space);
1215 case ARC_SPACE_HDRS:
1216 ARCSTAT_INCR(arcstat_hdr_size, space);
1218 case ARC_SPACE_L2HDRS:
1219 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1223 atomic_add_64(&arc_meta_used, space);
1224 atomic_add_64(&arc_size, space);
1228 arc_space_return(uint64_t space, arc_space_type_t type)
1230 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1235 case ARC_SPACE_DATA:
1236 ARCSTAT_INCR(arcstat_data_size, -space);
1238 case ARC_SPACE_OTHER:
1239 ARCSTAT_INCR(arcstat_other_size, -space);
1241 case ARC_SPACE_HDRS:
1242 ARCSTAT_INCR(arcstat_hdr_size, -space);
1244 case ARC_SPACE_L2HDRS:
1245 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1249 ASSERT(arc_meta_used >= space);
1250 if (arc_meta_max < arc_meta_used)
1251 arc_meta_max = arc_meta_used;
1252 atomic_add_64(&arc_meta_used, -space);
1253 ASSERT(arc_size >= space);
1254 atomic_add_64(&arc_size, -space);
1258 arc_data_buf_alloc(uint64_t size)
1260 if (arc_evict_needed(ARC_BUFC_DATA))
1261 cv_signal(&arc_reclaim_thr_cv);
1262 atomic_add_64(&arc_size, size);
1263 return (zio_data_buf_alloc(size));
1267 arc_data_buf_free(void *buf, uint64_t size)
1269 zio_data_buf_free(buf, size);
1270 ASSERT(arc_size >= size);
1271 atomic_add_64(&arc_size, -size);
1275 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1280 ASSERT3U(size, >, 0);
1281 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1282 ASSERT(BUF_EMPTY(hdr));
1285 hdr->b_spa = spa_load_guid(spa);
1286 hdr->b_state = arc_anon;
1287 hdr->b_arc_access = 0;
1288 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1291 buf->b_efunc = NULL;
1292 buf->b_private = NULL;
1295 arc_get_data_buf(buf);
1298 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1299 (void) refcount_add(&hdr->b_refcnt, tag);
1304 static char *arc_onloan_tag = "onloan";
1307 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1308 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1309 * buffers must be returned to the arc before they can be used by the DMU or
1313 arc_loan_buf(spa_t *spa, int size)
1317 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1319 atomic_add_64(&arc_loaned_bytes, size);
1324 * Return a loaned arc buffer to the arc.
1327 arc_return_buf(arc_buf_t *buf, void *tag)
1329 arc_buf_hdr_t *hdr = buf->b_hdr;
1331 ASSERT(buf->b_data != NULL);
1332 (void) refcount_add(&hdr->b_refcnt, tag);
1333 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1335 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1338 /* Detach an arc_buf from a dbuf (tag) */
1340 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1344 ASSERT(buf->b_data != NULL);
1346 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1347 (void) refcount_remove(&hdr->b_refcnt, tag);
1348 buf->b_efunc = NULL;
1349 buf->b_private = NULL;
1351 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1355 arc_buf_clone(arc_buf_t *from)
1358 arc_buf_hdr_t *hdr = from->b_hdr;
1359 uint64_t size = hdr->b_size;
1361 ASSERT(hdr->b_state != arc_anon);
1363 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1366 buf->b_efunc = NULL;
1367 buf->b_private = NULL;
1368 buf->b_next = hdr->b_buf;
1370 arc_get_data_buf(buf);
1371 bcopy(from->b_data, buf->b_data, size);
1372 hdr->b_datacnt += 1;
1377 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1380 kmutex_t *hash_lock;
1383 * Check to see if this buffer is evicted. Callers
1384 * must verify b_data != NULL to know if the add_ref
1387 mutex_enter(&buf->b_evict_lock);
1388 if (buf->b_data == NULL) {
1389 mutex_exit(&buf->b_evict_lock);
1392 hash_lock = HDR_LOCK(buf->b_hdr);
1393 mutex_enter(hash_lock);
1395 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1396 mutex_exit(&buf->b_evict_lock);
1398 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1399 add_reference(hdr, hash_lock, tag);
1400 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1401 arc_access(hdr, hash_lock);
1402 mutex_exit(hash_lock);
1403 ARCSTAT_BUMP(arcstat_hits);
1404 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1405 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1406 data, metadata, hits);
1410 * Free the arc data buffer. If it is an l2arc write in progress,
1411 * the buffer is placed on l2arc_free_on_write to be freed later.
1414 arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t),
1415 void *data, size_t size)
1417 if (HDR_L2_WRITING(hdr)) {
1418 l2arc_data_free_t *df;
1419 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_PUSHPAGE);
1420 df->l2df_data = data;
1421 df->l2df_size = size;
1422 df->l2df_func = free_func;
1423 mutex_enter(&l2arc_free_on_write_mtx);
1424 list_insert_head(l2arc_free_on_write, df);
1425 mutex_exit(&l2arc_free_on_write_mtx);
1426 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1428 free_func(data, size);
1433 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1437 /* free up data associated with the buf */
1439 arc_state_t *state = buf->b_hdr->b_state;
1440 uint64_t size = buf->b_hdr->b_size;
1441 arc_buf_contents_t type = buf->b_hdr->b_type;
1443 arc_cksum_verify(buf);
1446 if (type == ARC_BUFC_METADATA) {
1447 arc_buf_data_free(buf->b_hdr, zio_buf_free,
1449 arc_space_return(size, ARC_SPACE_DATA);
1451 ASSERT(type == ARC_BUFC_DATA);
1452 arc_buf_data_free(buf->b_hdr,
1453 zio_data_buf_free, buf->b_data, size);
1454 ARCSTAT_INCR(arcstat_data_size, -size);
1455 atomic_add_64(&arc_size, -size);
1458 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1459 uint64_t *cnt = &state->arcs_lsize[type];
1461 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1462 ASSERT(state != arc_anon);
1464 ASSERT3U(*cnt, >=, size);
1465 atomic_add_64(cnt, -size);
1467 ASSERT3U(state->arcs_size, >=, size);
1468 atomic_add_64(&state->arcs_size, -size);
1470 ASSERT(buf->b_hdr->b_datacnt > 0);
1471 buf->b_hdr->b_datacnt -= 1;
1474 /* only remove the buf if requested */
1478 /* remove the buf from the hdr list */
1479 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1481 *bufp = buf->b_next;
1484 ASSERT(buf->b_efunc == NULL);
1486 /* clean up the buf */
1488 kmem_cache_free(buf_cache, buf);
1492 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1494 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1496 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1497 ASSERT3P(hdr->b_state, ==, arc_anon);
1498 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1500 if (l2hdr != NULL) {
1501 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1503 * To prevent arc_free() and l2arc_evict() from
1504 * attempting to free the same buffer at the same time,
1505 * a FREE_IN_PROGRESS flag is given to arc_free() to
1506 * give it priority. l2arc_evict() can't destroy this
1507 * header while we are waiting on l2arc_buflist_mtx.
1509 * The hdr may be removed from l2ad_buflist before we
1510 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1512 if (!buflist_held) {
1513 mutex_enter(&l2arc_buflist_mtx);
1514 l2hdr = hdr->b_l2hdr;
1517 if (l2hdr != NULL) {
1518 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1519 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1520 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1521 if (hdr->b_state == arc_l2c_only)
1522 l2arc_hdr_stat_remove();
1523 hdr->b_l2hdr = NULL;
1527 mutex_exit(&l2arc_buflist_mtx);
1530 if (!BUF_EMPTY(hdr)) {
1531 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1532 buf_discard_identity(hdr);
1534 while (hdr->b_buf) {
1535 arc_buf_t *buf = hdr->b_buf;
1538 mutex_enter(&arc_eviction_mtx);
1539 mutex_enter(&buf->b_evict_lock);
1540 ASSERT(buf->b_hdr != NULL);
1541 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1542 hdr->b_buf = buf->b_next;
1543 buf->b_hdr = &arc_eviction_hdr;
1544 buf->b_next = arc_eviction_list;
1545 arc_eviction_list = buf;
1546 mutex_exit(&buf->b_evict_lock);
1547 mutex_exit(&arc_eviction_mtx);
1549 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1552 if (hdr->b_freeze_cksum != NULL) {
1553 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1554 hdr->b_freeze_cksum = NULL;
1556 if (hdr->b_thawed) {
1557 kmem_free(hdr->b_thawed, 1);
1558 hdr->b_thawed = NULL;
1561 ASSERT(!list_link_active(&hdr->b_arc_node));
1562 ASSERT3P(hdr->b_hash_next, ==, NULL);
1563 ASSERT3P(hdr->b_acb, ==, NULL);
1564 kmem_cache_free(hdr_cache, hdr);
1568 arc_buf_free(arc_buf_t *buf, void *tag)
1570 arc_buf_hdr_t *hdr = buf->b_hdr;
1571 int hashed = hdr->b_state != arc_anon;
1573 ASSERT(buf->b_efunc == NULL);
1574 ASSERT(buf->b_data != NULL);
1577 kmutex_t *hash_lock = HDR_LOCK(hdr);
1579 mutex_enter(hash_lock);
1581 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1583 (void) remove_reference(hdr, hash_lock, tag);
1584 if (hdr->b_datacnt > 1) {
1585 arc_buf_destroy(buf, FALSE, TRUE);
1587 ASSERT(buf == hdr->b_buf);
1588 ASSERT(buf->b_efunc == NULL);
1589 hdr->b_flags |= ARC_BUF_AVAILABLE;
1591 mutex_exit(hash_lock);
1592 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1595 * We are in the middle of an async write. Don't destroy
1596 * this buffer unless the write completes before we finish
1597 * decrementing the reference count.
1599 mutex_enter(&arc_eviction_mtx);
1600 (void) remove_reference(hdr, NULL, tag);
1601 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1602 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1603 mutex_exit(&arc_eviction_mtx);
1605 arc_hdr_destroy(hdr);
1607 if (remove_reference(hdr, NULL, tag) > 0)
1608 arc_buf_destroy(buf, FALSE, TRUE);
1610 arc_hdr_destroy(hdr);
1615 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1617 arc_buf_hdr_t *hdr = buf->b_hdr;
1618 kmutex_t *hash_lock = HDR_LOCK(hdr);
1619 int no_callback = (buf->b_efunc == NULL);
1621 if (hdr->b_state == arc_anon) {
1622 ASSERT(hdr->b_datacnt == 1);
1623 arc_buf_free(buf, tag);
1624 return (no_callback);
1627 mutex_enter(hash_lock);
1629 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1630 ASSERT(hdr->b_state != arc_anon);
1631 ASSERT(buf->b_data != NULL);
1633 (void) remove_reference(hdr, hash_lock, tag);
1634 if (hdr->b_datacnt > 1) {
1636 arc_buf_destroy(buf, FALSE, TRUE);
1637 } else if (no_callback) {
1638 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1639 ASSERT(buf->b_efunc == NULL);
1640 hdr->b_flags |= ARC_BUF_AVAILABLE;
1642 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1643 refcount_is_zero(&hdr->b_refcnt));
1644 mutex_exit(hash_lock);
1645 return (no_callback);
1649 arc_buf_size(arc_buf_t *buf)
1651 return (buf->b_hdr->b_size);
1655 * Evict buffers from list until we've removed the specified number of
1656 * bytes. Move the removed buffers to the appropriate evict state.
1657 * If the recycle flag is set, then attempt to "recycle" a buffer:
1658 * - look for a buffer to evict that is `bytes' long.
1659 * - return the data block from this buffer rather than freeing it.
1660 * This flag is used by callers that are trying to make space for a
1661 * new buffer in a full arc cache.
1663 * This function makes a "best effort". It skips over any buffers
1664 * it can't get a hash_lock on, and so may not catch all candidates.
1665 * It may also return without evicting as much space as requested.
1668 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1669 arc_buf_contents_t type)
1671 arc_state_t *evicted_state;
1672 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1673 arc_buf_hdr_t *ab, *ab_prev = NULL;
1674 list_t *list = &state->arcs_list[type];
1675 kmutex_t *hash_lock;
1676 boolean_t have_lock;
1677 void *stolen = NULL;
1679 ASSERT(state == arc_mru || state == arc_mfu);
1681 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1683 mutex_enter(&state->arcs_mtx);
1684 mutex_enter(&evicted_state->arcs_mtx);
1686 for (ab = list_tail(list); ab; ab = ab_prev) {
1687 ab_prev = list_prev(list, ab);
1688 /* prefetch buffers have a minimum lifespan */
1689 if (HDR_IO_IN_PROGRESS(ab) ||
1690 (spa && ab->b_spa != spa) ||
1691 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1692 ddi_get_lbolt() - ab->b_arc_access <
1693 arc_min_prefetch_lifespan)) {
1697 /* "lookahead" for better eviction candidate */
1698 if (recycle && ab->b_size != bytes &&
1699 ab_prev && ab_prev->b_size == bytes)
1701 hash_lock = HDR_LOCK(ab);
1702 have_lock = MUTEX_HELD(hash_lock);
1703 if (have_lock || mutex_tryenter(hash_lock)) {
1704 ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0);
1705 ASSERT(ab->b_datacnt > 0);
1707 arc_buf_t *buf = ab->b_buf;
1708 if (!mutex_tryenter(&buf->b_evict_lock)) {
1713 bytes_evicted += ab->b_size;
1714 if (recycle && ab->b_type == type &&
1715 ab->b_size == bytes &&
1716 !HDR_L2_WRITING(ab)) {
1717 stolen = buf->b_data;
1722 mutex_enter(&arc_eviction_mtx);
1723 arc_buf_destroy(buf,
1724 buf->b_data == stolen, FALSE);
1725 ab->b_buf = buf->b_next;
1726 buf->b_hdr = &arc_eviction_hdr;
1727 buf->b_next = arc_eviction_list;
1728 arc_eviction_list = buf;
1729 mutex_exit(&arc_eviction_mtx);
1730 mutex_exit(&buf->b_evict_lock);
1732 mutex_exit(&buf->b_evict_lock);
1733 arc_buf_destroy(buf,
1734 buf->b_data == stolen, TRUE);
1739 ARCSTAT_INCR(arcstat_evict_l2_cached,
1742 if (l2arc_write_eligible(ab->b_spa, ab)) {
1743 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1747 arcstat_evict_l2_ineligible,
1752 if (ab->b_datacnt == 0) {
1753 arc_change_state(evicted_state, ab, hash_lock);
1754 ASSERT(HDR_IN_HASH_TABLE(ab));
1755 ab->b_flags |= ARC_IN_HASH_TABLE;
1756 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1757 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1760 mutex_exit(hash_lock);
1761 if (bytes >= 0 && bytes_evicted >= bytes)
1768 mutex_exit(&evicted_state->arcs_mtx);
1769 mutex_exit(&state->arcs_mtx);
1771 if (bytes_evicted < bytes)
1772 dprintf("only evicted %lld bytes from %x\n",
1773 (longlong_t)bytes_evicted, state);
1776 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1779 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1782 * We have just evicted some date into the ghost state, make
1783 * sure we also adjust the ghost state size if necessary.
1786 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1787 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1788 arc_mru_ghost->arcs_size - arc_c;
1790 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1792 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1793 arc_evict_ghost(arc_mru_ghost, 0, todelete);
1794 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1795 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1796 arc_mru_ghost->arcs_size +
1797 arc_mfu_ghost->arcs_size - arc_c);
1798 arc_evict_ghost(arc_mfu_ghost, 0, todelete);
1806 * Remove buffers from list until we've removed the specified number of
1807 * bytes. Destroy the buffers that are removed.
1810 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1812 arc_buf_hdr_t *ab, *ab_prev;
1813 arc_buf_hdr_t marker;
1814 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1815 kmutex_t *hash_lock;
1816 uint64_t bytes_deleted = 0;
1817 uint64_t bufs_skipped = 0;
1819 ASSERT(GHOST_STATE(state));
1820 bzero(&marker, sizeof(marker));
1822 mutex_enter(&state->arcs_mtx);
1823 for (ab = list_tail(list); ab; ab = ab_prev) {
1824 ab_prev = list_prev(list, ab);
1825 if (spa && ab->b_spa != spa)
1828 /* ignore markers */
1832 hash_lock = HDR_LOCK(ab);
1833 /* caller may be trying to modify this buffer, skip it */
1834 if (MUTEX_HELD(hash_lock))
1836 if (mutex_tryenter(hash_lock)) {
1837 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1838 ASSERT(ab->b_buf == NULL);
1839 ARCSTAT_BUMP(arcstat_deleted);
1840 bytes_deleted += ab->b_size;
1842 if (ab->b_l2hdr != NULL) {
1844 * This buffer is cached on the 2nd Level ARC;
1845 * don't destroy the header.
1847 arc_change_state(arc_l2c_only, ab, hash_lock);
1848 mutex_exit(hash_lock);
1850 arc_change_state(arc_anon, ab, hash_lock);
1851 mutex_exit(hash_lock);
1852 arc_hdr_destroy(ab);
1855 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1856 if (bytes >= 0 && bytes_deleted >= bytes)
1858 } else if (bytes < 0) {
1860 * Insert a list marker and then wait for the
1861 * hash lock to become available. Once its
1862 * available, restart from where we left off.
1864 list_insert_after(list, ab, &marker);
1865 mutex_exit(&state->arcs_mtx);
1866 mutex_enter(hash_lock);
1867 mutex_exit(hash_lock);
1868 mutex_enter(&state->arcs_mtx);
1869 ab_prev = list_prev(list, &marker);
1870 list_remove(list, &marker);
1874 mutex_exit(&state->arcs_mtx);
1876 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1877 (bytes < 0 || bytes_deleted < bytes)) {
1878 list = &state->arcs_list[ARC_BUFC_METADATA];
1883 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1887 if (bytes_deleted < bytes)
1888 dprintf("only deleted %lld bytes from %p\n",
1889 (longlong_t)bytes_deleted, state);
1895 int64_t adjustment, delta;
1901 adjustment = MIN((int64_t)(arc_size - arc_c),
1902 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
1905 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1906 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
1907 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
1908 adjustment -= delta;
1911 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1912 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
1913 (void) arc_evict(arc_mru, 0, delta, FALSE,
1921 adjustment = arc_size - arc_c;
1923 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1924 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
1925 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
1926 adjustment -= delta;
1929 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1930 int64_t delta = MIN(adjustment,
1931 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
1932 (void) arc_evict(arc_mfu, 0, delta, FALSE,
1937 * Adjust ghost lists
1940 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
1942 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
1943 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
1944 arc_evict_ghost(arc_mru_ghost, 0, delta);
1948 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
1950 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
1951 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
1952 arc_evict_ghost(arc_mfu_ghost, 0, delta);
1957 * Request that arc user drop references so that N bytes can be released
1958 * from the cache. This provides a mechanism to ensure the arc can honor
1959 * the arc_meta_limit and reclaim buffers which are pinned in the cache
1960 * by higher layers. (i.e. the zpl)
1963 arc_do_user_prune(int64_t adjustment)
1965 arc_prune_func_t *func;
1967 arc_prune_t *cp, *np;
1969 mutex_enter(&arc_prune_mtx);
1971 cp = list_head(&arc_prune_list);
1972 while (cp != NULL) {
1974 private = cp->p_private;
1975 np = list_next(&arc_prune_list, cp);
1976 refcount_add(&cp->p_refcnt, func);
1977 mutex_exit(&arc_prune_mtx);
1980 func(adjustment, private);
1982 mutex_enter(&arc_prune_mtx);
1984 /* User removed prune callback concurrently with execution */
1985 if (refcount_remove(&cp->p_refcnt, func) == 0) {
1986 ASSERT(!list_link_active(&cp->p_node));
1987 refcount_destroy(&cp->p_refcnt);
1988 kmem_free(cp, sizeof (*cp));
1994 ARCSTAT_BUMP(arcstat_prune);
1995 mutex_exit(&arc_prune_mtx);
1999 arc_do_user_evicts(void)
2001 mutex_enter(&arc_eviction_mtx);
2002 while (arc_eviction_list != NULL) {
2003 arc_buf_t *buf = arc_eviction_list;
2004 arc_eviction_list = buf->b_next;
2005 mutex_enter(&buf->b_evict_lock);
2007 mutex_exit(&buf->b_evict_lock);
2008 mutex_exit(&arc_eviction_mtx);
2010 if (buf->b_efunc != NULL)
2011 VERIFY(buf->b_efunc(buf) == 0);
2013 buf->b_efunc = NULL;
2014 buf->b_private = NULL;
2015 kmem_cache_free(buf_cache, buf);
2016 mutex_enter(&arc_eviction_mtx);
2018 mutex_exit(&arc_eviction_mtx);
2022 * Evict only meta data objects from the cache leaving the data objects.
2023 * This is only used to enforce the tunable arc_meta_limit, if we are
2024 * unable to evict enough buffers notify the user via the prune callback.
2027 arc_adjust_meta(int64_t adjustment, boolean_t may_prune)
2031 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2032 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2033 arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_METADATA);
2034 adjustment -= delta;
2037 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2038 delta = MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2039 arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_METADATA);
2040 adjustment -= delta;
2043 if (may_prune && (adjustment > 0) && (arc_meta_used > arc_meta_limit))
2044 arc_do_user_prune(arc_meta_prune);
2048 * Flush all *evictable* data from the cache for the given spa.
2049 * NOTE: this will not touch "active" (i.e. referenced) data.
2052 arc_flush(spa_t *spa)
2057 guid = spa_load_guid(spa);
2059 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2060 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2064 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2065 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2069 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2070 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2074 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2075 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2080 arc_evict_ghost(arc_mru_ghost, guid, -1);
2081 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2083 mutex_enter(&arc_reclaim_thr_lock);
2084 arc_do_user_evicts();
2085 mutex_exit(&arc_reclaim_thr_lock);
2086 ASSERT(spa || arc_eviction_list == NULL);
2090 arc_shrink(uint64_t bytes)
2092 if (arc_c > arc_c_min) {
2095 to_free = bytes ? bytes : arc_c >> arc_shrink_shift;
2097 if (arc_c > arc_c_min + to_free)
2098 atomic_add_64(&arc_c, -to_free);
2102 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2103 if (arc_c > arc_size)
2104 arc_c = MAX(arc_size, arc_c_min);
2106 arc_p = (arc_c >> 1);
2107 ASSERT(arc_c >= arc_c_min);
2108 ASSERT((int64_t)arc_p >= 0);
2111 if (arc_size > arc_c)
2116 arc_kmem_reap_now(arc_reclaim_strategy_t strat, uint64_t bytes)
2119 kmem_cache_t *prev_cache = NULL;
2120 kmem_cache_t *prev_data_cache = NULL;
2121 extern kmem_cache_t *zio_buf_cache[];
2122 extern kmem_cache_t *zio_data_buf_cache[];
2125 * An aggressive reclamation will shrink the cache size as well as
2126 * reap free buffers from the arc kmem caches.
2128 if (strat == ARC_RECLAIM_AGGR)
2131 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2132 if (zio_buf_cache[i] != prev_cache) {
2133 prev_cache = zio_buf_cache[i];
2134 kmem_cache_reap_now(zio_buf_cache[i]);
2136 if (zio_data_buf_cache[i] != prev_data_cache) {
2137 prev_data_cache = zio_data_buf_cache[i];
2138 kmem_cache_reap_now(zio_data_buf_cache[i]);
2142 kmem_cache_reap_now(buf_cache);
2143 kmem_cache_reap_now(hdr_cache);
2147 * Unlike other ZFS implementations this thread is only responsible for
2148 * adapting the target ARC size on Linux. The responsibility for memory
2149 * reclamation has been entirely delegated to the arc_shrinker_func()
2150 * which is registered with the VM. To reflect this change in behavior
2151 * the arc_reclaim thread has been renamed to arc_adapt.
2154 arc_adapt_thread(void)
2159 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2161 mutex_enter(&arc_reclaim_thr_lock);
2162 while (arc_thread_exit == 0) {
2164 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2166 if (spa_get_random(100) == 0) {
2169 if (last_reclaim == ARC_RECLAIM_CONS) {
2170 last_reclaim = ARC_RECLAIM_AGGR;
2172 last_reclaim = ARC_RECLAIM_CONS;
2176 last_reclaim = ARC_RECLAIM_AGGR;
2180 /* reset the growth delay for every reclaim */
2181 arc_grow_time = ddi_get_lbolt()+(arc_grow_retry * hz);
2183 arc_kmem_reap_now(last_reclaim, 0);
2186 #endif /* !_KERNEL */
2188 /* No recent memory pressure allow the ARC to grow. */
2189 if (arc_no_grow && ddi_get_lbolt() >= arc_grow_time)
2190 arc_no_grow = FALSE;
2193 * Keep meta data usage within limits, arc_shrink() is not
2194 * used to avoid collapsing the arc_c value when only the
2195 * arc_meta_limit is being exceeded.
2197 prune = (int64_t)arc_meta_used - (int64_t)arc_meta_limit;
2199 arc_adjust_meta(prune, B_TRUE);
2203 if (arc_eviction_list != NULL)
2204 arc_do_user_evicts();
2206 /* block until needed, or one second, whichever is shorter */
2207 CALLB_CPR_SAFE_BEGIN(&cpr);
2208 (void) cv_timedwait_interruptible(&arc_reclaim_thr_cv,
2209 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2210 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2213 arc_thread_exit = 0;
2214 cv_broadcast(&arc_reclaim_thr_cv);
2215 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2221 * Determine the amount of memory eligible for eviction contained in the
2222 * ARC. All clean data reported by the ghost lists can always be safely
2223 * evicted. Due to arc_c_min, the same does not hold for all clean data
2224 * contained by the regular mru and mfu lists.
2226 * In the case of the regular mru and mfu lists, we need to report as
2227 * much clean data as possible, such that evicting that same reported
2228 * data will not bring arc_size below arc_c_min. Thus, in certain
2229 * circumstances, the total amount of clean data in the mru and mfu
2230 * lists might not actually be evictable.
2232 * The following two distinct cases are accounted for:
2234 * 1. The sum of the amount of dirty data contained by both the mru and
2235 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2236 * is greater than or equal to arc_c_min.
2237 * (i.e. amount of dirty data >= arc_c_min)
2239 * This is the easy case; all clean data contained by the mru and mfu
2240 * lists is evictable. Evicting all clean data can only drop arc_size
2241 * to the amount of dirty data, which is greater than arc_c_min.
2243 * 2. The sum of the amount of dirty data contained by both the mru and
2244 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2245 * is less than arc_c_min.
2246 * (i.e. arc_c_min > amount of dirty data)
2248 * 2.1. arc_size is greater than or equal arc_c_min.
2249 * (i.e. arc_size >= arc_c_min > amount of dirty data)
2251 * In this case, not all clean data from the regular mru and mfu
2252 * lists is actually evictable; we must leave enough clean data
2253 * to keep arc_size above arc_c_min. Thus, the maximum amount of
2254 * evictable data from the two lists combined, is exactly the
2255 * difference between arc_size and arc_c_min.
2257 * 2.2. arc_size is less than arc_c_min
2258 * (i.e. arc_c_min > arc_size > amount of dirty data)
2260 * In this case, none of the data contained in the mru and mfu
2261 * lists is evictable, even if it's clean. Since arc_size is
2262 * already below arc_c_min, evicting any more would only
2263 * increase this negative difference.
2266 arc_evictable_memory(void) {
2267 uint64_t arc_clean =
2268 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
2269 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
2270 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
2271 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
2272 uint64_t ghost_clean =
2273 arc_mru_ghost->arcs_lsize[ARC_BUFC_DATA] +
2274 arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] +
2275 arc_mfu_ghost->arcs_lsize[ARC_BUFC_DATA] +
2276 arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA];
2277 uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0);
2279 if (arc_dirty >= arc_c_min)
2280 return (ghost_clean + arc_clean);
2282 return (ghost_clean + MAX((int64_t)arc_size - (int64_t)arc_c_min, 0));
2286 __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc)
2290 /* The arc is considered warm once reclaim has occurred */
2291 if (unlikely(arc_warm == B_FALSE))
2294 /* Return the potential number of reclaimable pages */
2295 pages = btop(arc_evictable_memory());
2296 if (sc->nr_to_scan == 0)
2299 /* Not allowed to perform filesystem reclaim */
2300 if (!(sc->gfp_mask & __GFP_FS))
2303 /* Reclaim in progress */
2304 if (mutex_tryenter(&arc_reclaim_thr_lock) == 0)
2308 * Evict the requested number of pages by shrinking arc_c the
2309 * requested amount. If there is nothing left to evict just
2310 * reap whatever we can from the various arc slabs.
2313 arc_kmem_reap_now(ARC_RECLAIM_AGGR, ptob(sc->nr_to_scan));
2314 pages = btop(arc_evictable_memory());
2316 arc_kmem_reap_now(ARC_RECLAIM_CONS, ptob(sc->nr_to_scan));
2321 * When direct reclaim is observed it usually indicates a rapid
2322 * increase in memory pressure. This occurs because the kswapd
2323 * threads were unable to asynchronously keep enough free memory
2324 * available. In this case set arc_no_grow to briefly pause arc
2325 * growth to avoid compounding the memory pressure.
2327 if (current_is_kswapd()) {
2328 ARCSTAT_BUMP(arcstat_memory_indirect_count);
2330 arc_no_grow = B_TRUE;
2331 arc_grow_time = ddi_get_lbolt() + (arc_grow_retry * hz);
2332 ARCSTAT_BUMP(arcstat_memory_direct_count);
2335 mutex_exit(&arc_reclaim_thr_lock);
2339 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func);
2341 SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS);
2342 #endif /* _KERNEL */
2345 * Adapt arc info given the number of bytes we are trying to add and
2346 * the state that we are comming from. This function is only called
2347 * when we are adding new content to the cache.
2350 arc_adapt(int bytes, arc_state_t *state)
2353 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2355 if (state == arc_l2c_only)
2360 * Adapt the target size of the MRU list:
2361 * - if we just hit in the MRU ghost list, then increase
2362 * the target size of the MRU list.
2363 * - if we just hit in the MFU ghost list, then increase
2364 * the target size of the MFU list by decreasing the
2365 * target size of the MRU list.
2367 if (state == arc_mru_ghost) {
2368 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2369 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2370 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2372 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2373 } else if (state == arc_mfu_ghost) {
2376 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2377 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2378 mult = MIN(mult, 10);
2380 delta = MIN(bytes * mult, arc_p);
2381 arc_p = MAX(arc_p_min, arc_p - delta);
2383 ASSERT((int64_t)arc_p >= 0);
2388 if (arc_c >= arc_c_max)
2392 * If we're within (2 * maxblocksize) bytes of the target
2393 * cache size, increment the target cache size
2395 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2396 atomic_add_64(&arc_c, (int64_t)bytes);
2397 if (arc_c > arc_c_max)
2399 else if (state == arc_anon)
2400 atomic_add_64(&arc_p, (int64_t)bytes);
2404 ASSERT((int64_t)arc_p >= 0);
2408 * Check if the cache has reached its limits and eviction is required
2412 arc_evict_needed(arc_buf_contents_t type)
2414 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2419 * If zio data pages are being allocated out of a separate heap segment,
2420 * then enforce that the size of available vmem for this area remains
2421 * above about 1/32nd free.
2423 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2424 vmem_size(zio_arena, VMEM_FREE) <
2425 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2432 return (arc_size > arc_c);
2436 * The buffer, supplied as the first argument, needs a data block.
2437 * So, if we are at cache max, determine which cache should be victimized.
2438 * We have the following cases:
2440 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2441 * In this situation if we're out of space, but the resident size of the MFU is
2442 * under the limit, victimize the MFU cache to satisfy this insertion request.
2444 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2445 * Here, we've used up all of the available space for the MRU, so we need to
2446 * evict from our own cache instead. Evict from the set of resident MRU
2449 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2450 * c minus p represents the MFU space in the cache, since p is the size of the
2451 * cache that is dedicated to the MRU. In this situation there's still space on
2452 * the MFU side, so the MRU side needs to be victimized.
2454 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2455 * MFU's resident set is consuming more space than it has been allotted. In
2456 * this situation, we must victimize our own cache, the MFU, for this insertion.
2459 arc_get_data_buf(arc_buf_t *buf)
2461 arc_state_t *state = buf->b_hdr->b_state;
2462 uint64_t size = buf->b_hdr->b_size;
2463 arc_buf_contents_t type = buf->b_hdr->b_type;
2465 arc_adapt(size, state);
2468 * We have not yet reached cache maximum size,
2469 * just allocate a new buffer.
2471 if (!arc_evict_needed(type)) {
2472 if (type == ARC_BUFC_METADATA) {
2473 buf->b_data = zio_buf_alloc(size);
2474 arc_space_consume(size, ARC_SPACE_DATA);
2476 ASSERT(type == ARC_BUFC_DATA);
2477 buf->b_data = zio_data_buf_alloc(size);
2478 ARCSTAT_INCR(arcstat_data_size, size);
2479 atomic_add_64(&arc_size, size);
2485 * If we are prefetching from the mfu ghost list, this buffer
2486 * will end up on the mru list; so steal space from there.
2488 if (state == arc_mfu_ghost)
2489 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2490 else if (state == arc_mru_ghost)
2493 if (state == arc_mru || state == arc_anon) {
2494 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2495 state = (arc_mfu->arcs_lsize[type] >= size &&
2496 arc_p > mru_used) ? arc_mfu : arc_mru;
2499 uint64_t mfu_space = arc_c - arc_p;
2500 state = (arc_mru->arcs_lsize[type] >= size &&
2501 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2504 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2505 if (type == ARC_BUFC_METADATA) {
2506 buf->b_data = zio_buf_alloc(size);
2507 arc_space_consume(size, ARC_SPACE_DATA);
2510 * If we are unable to recycle an existing meta buffer
2511 * signal the reclaim thread. It will notify users
2512 * via the prune callback to drop references. The
2513 * prune callback in run in the context of the reclaim
2514 * thread to avoid deadlocking on the hash_lock.
2516 cv_signal(&arc_reclaim_thr_cv);
2518 ASSERT(type == ARC_BUFC_DATA);
2519 buf->b_data = zio_data_buf_alloc(size);
2520 ARCSTAT_INCR(arcstat_data_size, size);
2521 atomic_add_64(&arc_size, size);
2524 ARCSTAT_BUMP(arcstat_recycle_miss);
2526 ASSERT(buf->b_data != NULL);
2529 * Update the state size. Note that ghost states have a
2530 * "ghost size" and so don't need to be updated.
2532 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2533 arc_buf_hdr_t *hdr = buf->b_hdr;
2535 atomic_add_64(&hdr->b_state->arcs_size, size);
2536 if (list_link_active(&hdr->b_arc_node)) {
2537 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2538 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2541 * If we are growing the cache, and we are adding anonymous
2542 * data, and we have outgrown arc_p, update arc_p
2544 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2545 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2546 arc_p = MIN(arc_c, arc_p + size);
2551 * This routine is called whenever a buffer is accessed.
2552 * NOTE: the hash lock is dropped in this function.
2555 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2559 ASSERT(MUTEX_HELD(hash_lock));
2561 if (buf->b_state == arc_anon) {
2563 * This buffer is not in the cache, and does not
2564 * appear in our "ghost" list. Add the new buffer
2568 ASSERT(buf->b_arc_access == 0);
2569 buf->b_arc_access = ddi_get_lbolt();
2570 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2571 arc_change_state(arc_mru, buf, hash_lock);
2573 } else if (buf->b_state == arc_mru) {
2574 now = ddi_get_lbolt();
2577 * If this buffer is here because of a prefetch, then either:
2578 * - clear the flag if this is a "referencing" read
2579 * (any subsequent access will bump this into the MFU state).
2581 * - move the buffer to the head of the list if this is
2582 * another prefetch (to make it less likely to be evicted).
2584 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2585 if (refcount_count(&buf->b_refcnt) == 0) {
2586 ASSERT(list_link_active(&buf->b_arc_node));
2588 buf->b_flags &= ~ARC_PREFETCH;
2589 ARCSTAT_BUMP(arcstat_mru_hits);
2591 buf->b_arc_access = now;
2596 * This buffer has been "accessed" only once so far,
2597 * but it is still in the cache. Move it to the MFU
2600 if (now > buf->b_arc_access + ARC_MINTIME) {
2602 * More than 125ms have passed since we
2603 * instantiated this buffer. Move it to the
2604 * most frequently used state.
2606 buf->b_arc_access = now;
2607 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2608 arc_change_state(arc_mfu, buf, hash_lock);
2610 ARCSTAT_BUMP(arcstat_mru_hits);
2611 } else if (buf->b_state == arc_mru_ghost) {
2612 arc_state_t *new_state;
2614 * This buffer has been "accessed" recently, but
2615 * was evicted from the cache. Move it to the
2619 if (buf->b_flags & ARC_PREFETCH) {
2620 new_state = arc_mru;
2621 if (refcount_count(&buf->b_refcnt) > 0)
2622 buf->b_flags &= ~ARC_PREFETCH;
2623 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2625 new_state = arc_mfu;
2626 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2629 buf->b_arc_access = ddi_get_lbolt();
2630 arc_change_state(new_state, buf, hash_lock);
2632 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2633 } else if (buf->b_state == arc_mfu) {
2635 * This buffer has been accessed more than once and is
2636 * still in the cache. Keep it in the MFU state.
2638 * NOTE: an add_reference() that occurred when we did
2639 * the arc_read() will have kicked this off the list.
2640 * If it was a prefetch, we will explicitly move it to
2641 * the head of the list now.
2643 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2644 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2645 ASSERT(list_link_active(&buf->b_arc_node));
2647 ARCSTAT_BUMP(arcstat_mfu_hits);
2648 buf->b_arc_access = ddi_get_lbolt();
2649 } else if (buf->b_state == arc_mfu_ghost) {
2650 arc_state_t *new_state = arc_mfu;
2652 * This buffer has been accessed more than once but has
2653 * been evicted from the cache. Move it back to the
2657 if (buf->b_flags & ARC_PREFETCH) {
2659 * This is a prefetch access...
2660 * move this block back to the MRU state.
2662 ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0);
2663 new_state = arc_mru;
2666 buf->b_arc_access = ddi_get_lbolt();
2667 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2668 arc_change_state(new_state, buf, hash_lock);
2670 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2671 } else if (buf->b_state == arc_l2c_only) {
2673 * This buffer is on the 2nd Level ARC.
2676 buf->b_arc_access = ddi_get_lbolt();
2677 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2678 arc_change_state(arc_mfu, buf, hash_lock);
2680 ASSERT(!"invalid arc state");
2684 /* a generic arc_done_func_t which you can use */
2687 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2689 if (zio == NULL || zio->io_error == 0)
2690 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2691 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2694 /* a generic arc_done_func_t */
2696 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2698 arc_buf_t **bufp = arg;
2699 if (zio && zio->io_error) {
2700 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2704 ASSERT(buf->b_data);
2709 arc_read_done(zio_t *zio)
2711 arc_buf_hdr_t *hdr, *found;
2713 arc_buf_t *abuf; /* buffer we're assigning to callback */
2714 kmutex_t *hash_lock;
2715 arc_callback_t *callback_list, *acb;
2716 int freeable = FALSE;
2718 buf = zio->io_private;
2722 * The hdr was inserted into hash-table and removed from lists
2723 * prior to starting I/O. We should find this header, since
2724 * it's in the hash table, and it should be legit since it's
2725 * not possible to evict it during the I/O. The only possible
2726 * reason for it not to be found is if we were freed during the
2729 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2732 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2733 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2734 (found == hdr && HDR_L2_READING(hdr)));
2736 hdr->b_flags &= ~ARC_L2_EVICTED;
2737 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2738 hdr->b_flags &= ~ARC_L2CACHE;
2740 /* byteswap if necessary */
2741 callback_list = hdr->b_acb;
2742 ASSERT(callback_list != NULL);
2743 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2744 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2745 byteswap_uint64_array :
2746 dmu_ot[BP_GET_TYPE(zio->io_bp)].ot_byteswap;
2747 func(buf->b_data, hdr->b_size);
2750 arc_cksum_compute(buf, B_FALSE);
2752 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2754 * Only call arc_access on anonymous buffers. This is because
2755 * if we've issued an I/O for an evicted buffer, we've already
2756 * called arc_access (to prevent any simultaneous readers from
2757 * getting confused).
2759 arc_access(hdr, hash_lock);
2762 /* create copies of the data buffer for the callers */
2764 for (acb = callback_list; acb; acb = acb->acb_next) {
2765 if (acb->acb_done) {
2767 abuf = arc_buf_clone(buf);
2768 acb->acb_buf = abuf;
2773 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2774 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2776 ASSERT(buf->b_efunc == NULL);
2777 ASSERT(hdr->b_datacnt == 1);
2778 hdr->b_flags |= ARC_BUF_AVAILABLE;
2781 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2783 if (zio->io_error != 0) {
2784 hdr->b_flags |= ARC_IO_ERROR;
2785 if (hdr->b_state != arc_anon)
2786 arc_change_state(arc_anon, hdr, hash_lock);
2787 if (HDR_IN_HASH_TABLE(hdr))
2788 buf_hash_remove(hdr);
2789 freeable = refcount_is_zero(&hdr->b_refcnt);
2793 * Broadcast before we drop the hash_lock to avoid the possibility
2794 * that the hdr (and hence the cv) might be freed before we get to
2795 * the cv_broadcast().
2797 cv_broadcast(&hdr->b_cv);
2800 mutex_exit(hash_lock);
2803 * This block was freed while we waited for the read to
2804 * complete. It has been removed from the hash table and
2805 * moved to the anonymous state (so that it won't show up
2808 ASSERT3P(hdr->b_state, ==, arc_anon);
2809 freeable = refcount_is_zero(&hdr->b_refcnt);
2812 /* execute each callback and free its structure */
2813 while ((acb = callback_list) != NULL) {
2815 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2817 if (acb->acb_zio_dummy != NULL) {
2818 acb->acb_zio_dummy->io_error = zio->io_error;
2819 zio_nowait(acb->acb_zio_dummy);
2822 callback_list = acb->acb_next;
2823 kmem_free(acb, sizeof (arc_callback_t));
2827 arc_hdr_destroy(hdr);
2831 * "Read" the block block at the specified DVA (in bp) via the
2832 * cache. If the block is found in the cache, invoke the provided
2833 * callback immediately and return. Note that the `zio' parameter
2834 * in the callback will be NULL in this case, since no IO was
2835 * required. If the block is not in the cache pass the read request
2836 * on to the spa with a substitute callback function, so that the
2837 * requested block will be added to the cache.
2839 * If a read request arrives for a block that has a read in-progress,
2840 * either wait for the in-progress read to complete (and return the
2841 * results); or, if this is a read with a "done" func, add a record
2842 * to the read to invoke the "done" func when the read completes,
2843 * and return; or just return.
2845 * arc_read_done() will invoke all the requested "done" functions
2846 * for readers of this block.
2848 * Normal callers should use arc_read and pass the arc buffer and offset
2849 * for the bp. But if you know you don't need locking, you can use
2853 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_buf_t *pbuf,
2854 arc_done_func_t *done, void *private, int priority, int zio_flags,
2855 uint32_t *arc_flags, const zbookmark_t *zb)
2861 * XXX This happens from traverse callback funcs, for
2862 * the objset_phys_t block.
2864 return (arc_read_nolock(pio, spa, bp, done, private, priority,
2865 zio_flags, arc_flags, zb));
2868 ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt));
2869 ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size);
2870 rw_enter(&pbuf->b_data_lock, RW_READER);
2872 err = arc_read_nolock(pio, spa, bp, done, private, priority,
2873 zio_flags, arc_flags, zb);
2874 rw_exit(&pbuf->b_data_lock);
2880 arc_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bp,
2881 arc_done_func_t *done, void *private, int priority, int zio_flags,
2882 uint32_t *arc_flags, const zbookmark_t *zb)
2885 arc_buf_t *buf = NULL;
2886 kmutex_t *hash_lock;
2888 uint64_t guid = spa_load_guid(spa);
2891 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2893 if (hdr && hdr->b_datacnt > 0) {
2895 *arc_flags |= ARC_CACHED;
2897 if (HDR_IO_IN_PROGRESS(hdr)) {
2899 if (*arc_flags & ARC_WAIT) {
2900 cv_wait(&hdr->b_cv, hash_lock);
2901 mutex_exit(hash_lock);
2904 ASSERT(*arc_flags & ARC_NOWAIT);
2907 arc_callback_t *acb = NULL;
2909 acb = kmem_zalloc(sizeof (arc_callback_t),
2911 acb->acb_done = done;
2912 acb->acb_private = private;
2914 acb->acb_zio_dummy = zio_null(pio,
2915 spa, NULL, NULL, NULL, zio_flags);
2917 ASSERT(acb->acb_done != NULL);
2918 acb->acb_next = hdr->b_acb;
2920 add_reference(hdr, hash_lock, private);
2921 mutex_exit(hash_lock);
2924 mutex_exit(hash_lock);
2928 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2931 add_reference(hdr, hash_lock, private);
2933 * If this block is already in use, create a new
2934 * copy of the data so that we will be guaranteed
2935 * that arc_release() will always succeed.
2939 ASSERT(buf->b_data);
2940 if (HDR_BUF_AVAILABLE(hdr)) {
2941 ASSERT(buf->b_efunc == NULL);
2942 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2944 buf = arc_buf_clone(buf);
2947 } else if (*arc_flags & ARC_PREFETCH &&
2948 refcount_count(&hdr->b_refcnt) == 0) {
2949 hdr->b_flags |= ARC_PREFETCH;
2951 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2952 arc_access(hdr, hash_lock);
2953 if (*arc_flags & ARC_L2CACHE)
2954 hdr->b_flags |= ARC_L2CACHE;
2955 mutex_exit(hash_lock);
2956 ARCSTAT_BUMP(arcstat_hits);
2957 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2958 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2959 data, metadata, hits);
2962 done(NULL, buf, private);
2964 uint64_t size = BP_GET_LSIZE(bp);
2965 arc_callback_t *acb;
2968 boolean_t devw = B_FALSE;
2971 /* this block is not in the cache */
2972 arc_buf_hdr_t *exists;
2973 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2974 buf = arc_buf_alloc(spa, size, private, type);
2976 hdr->b_dva = *BP_IDENTITY(bp);
2977 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
2978 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2979 exists = buf_hash_insert(hdr, &hash_lock);
2981 /* somebody beat us to the hash insert */
2982 mutex_exit(hash_lock);
2983 buf_discard_identity(hdr);
2984 (void) arc_buf_remove_ref(buf, private);
2985 goto top; /* restart the IO request */
2987 /* if this is a prefetch, we don't have a reference */
2988 if (*arc_flags & ARC_PREFETCH) {
2989 (void) remove_reference(hdr, hash_lock,
2991 hdr->b_flags |= ARC_PREFETCH;
2993 if (*arc_flags & ARC_L2CACHE)
2994 hdr->b_flags |= ARC_L2CACHE;
2995 if (BP_GET_LEVEL(bp) > 0)
2996 hdr->b_flags |= ARC_INDIRECT;
2998 /* this block is in the ghost cache */
2999 ASSERT(GHOST_STATE(hdr->b_state));
3000 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3001 ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0);
3002 ASSERT(hdr->b_buf == NULL);
3004 /* if this is a prefetch, we don't have a reference */
3005 if (*arc_flags & ARC_PREFETCH)
3006 hdr->b_flags |= ARC_PREFETCH;
3008 add_reference(hdr, hash_lock, private);
3009 if (*arc_flags & ARC_L2CACHE)
3010 hdr->b_flags |= ARC_L2CACHE;
3011 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3014 buf->b_efunc = NULL;
3015 buf->b_private = NULL;
3018 ASSERT(hdr->b_datacnt == 0);
3020 arc_get_data_buf(buf);
3021 arc_access(hdr, hash_lock);
3024 ASSERT(!GHOST_STATE(hdr->b_state));
3026 acb = kmem_zalloc(sizeof (arc_callback_t), KM_PUSHPAGE);
3027 acb->acb_done = done;
3028 acb->acb_private = private;
3030 ASSERT(hdr->b_acb == NULL);
3032 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3034 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3035 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3036 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3037 addr = hdr->b_l2hdr->b_daddr;
3039 * Lock out device removal.
3041 if (vdev_is_dead(vd) ||
3042 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3046 mutex_exit(hash_lock);
3048 ASSERT3U(hdr->b_size, ==, size);
3049 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3050 uint64_t, size, zbookmark_t *, zb);
3051 ARCSTAT_BUMP(arcstat_misses);
3052 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3053 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3054 data, metadata, misses);
3056 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3058 * Read from the L2ARC if the following are true:
3059 * 1. The L2ARC vdev was previously cached.
3060 * 2. This buffer still has L2ARC metadata.
3061 * 3. This buffer isn't currently writing to the L2ARC.
3062 * 4. The L2ARC entry wasn't evicted, which may
3063 * also have invalidated the vdev.
3064 * 5. This isn't prefetch and l2arc_noprefetch is set.
3066 if (hdr->b_l2hdr != NULL &&
3067 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3068 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3069 l2arc_read_callback_t *cb;
3071 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3072 ARCSTAT_BUMP(arcstat_l2_hits);
3074 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3076 cb->l2rcb_buf = buf;
3077 cb->l2rcb_spa = spa;
3080 cb->l2rcb_flags = zio_flags;
3083 * l2arc read. The SCL_L2ARC lock will be
3084 * released by l2arc_read_done().
3086 rzio = zio_read_phys(pio, vd, addr, size,
3087 buf->b_data, ZIO_CHECKSUM_OFF,
3088 l2arc_read_done, cb, priority, zio_flags |
3089 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
3090 ZIO_FLAG_DONT_PROPAGATE |
3091 ZIO_FLAG_DONT_RETRY, B_FALSE);
3092 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3094 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
3096 if (*arc_flags & ARC_NOWAIT) {
3101 ASSERT(*arc_flags & ARC_WAIT);
3102 if (zio_wait(rzio) == 0)
3105 /* l2arc read error; goto zio_read() */
3107 DTRACE_PROBE1(l2arc__miss,
3108 arc_buf_hdr_t *, hdr);
3109 ARCSTAT_BUMP(arcstat_l2_misses);
3110 if (HDR_L2_WRITING(hdr))
3111 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3112 spa_config_exit(spa, SCL_L2ARC, vd);
3116 spa_config_exit(spa, SCL_L2ARC, vd);
3117 if (l2arc_ndev != 0) {
3118 DTRACE_PROBE1(l2arc__miss,
3119 arc_buf_hdr_t *, hdr);
3120 ARCSTAT_BUMP(arcstat_l2_misses);
3124 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3125 arc_read_done, buf, priority, zio_flags, zb);
3127 if (*arc_flags & ARC_WAIT)
3128 return (zio_wait(rzio));
3130 ASSERT(*arc_flags & ARC_NOWAIT);
3137 arc_add_prune_callback(arc_prune_func_t *func, void *private)
3141 p = kmem_alloc(sizeof(*p), KM_SLEEP);
3143 p->p_private = private;
3144 list_link_init(&p->p_node);
3145 refcount_create(&p->p_refcnt);
3147 mutex_enter(&arc_prune_mtx);
3148 refcount_add(&p->p_refcnt, &arc_prune_list);
3149 list_insert_head(&arc_prune_list, p);
3150 mutex_exit(&arc_prune_mtx);
3156 arc_remove_prune_callback(arc_prune_t *p)
3158 mutex_enter(&arc_prune_mtx);
3159 list_remove(&arc_prune_list, p);
3160 if (refcount_remove(&p->p_refcnt, &arc_prune_list) == 0) {
3161 refcount_destroy(&p->p_refcnt);
3162 kmem_free(p, sizeof (*p));
3164 mutex_exit(&arc_prune_mtx);
3168 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3170 ASSERT(buf->b_hdr != NULL);
3171 ASSERT(buf->b_hdr->b_state != arc_anon);
3172 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3173 ASSERT(buf->b_efunc == NULL);
3174 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3176 buf->b_efunc = func;
3177 buf->b_private = private;
3181 * This is used by the DMU to let the ARC know that a buffer is
3182 * being evicted, so the ARC should clean up. If this arc buf
3183 * is not yet in the evicted state, it will be put there.
3186 arc_buf_evict(arc_buf_t *buf)
3189 kmutex_t *hash_lock;
3192 mutex_enter(&buf->b_evict_lock);
3196 * We are in arc_do_user_evicts().
3198 ASSERT(buf->b_data == NULL);
3199 mutex_exit(&buf->b_evict_lock);
3201 } else if (buf->b_data == NULL) {
3202 arc_buf_t copy = *buf; /* structure assignment */
3204 * We are on the eviction list; process this buffer now
3205 * but let arc_do_user_evicts() do the reaping.
3207 buf->b_efunc = NULL;
3208 mutex_exit(&buf->b_evict_lock);
3209 VERIFY(copy.b_efunc(©) == 0);
3212 hash_lock = HDR_LOCK(hdr);
3213 mutex_enter(hash_lock);
3215 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3217 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3218 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3221 * Pull this buffer off of the hdr
3224 while (*bufp != buf)
3225 bufp = &(*bufp)->b_next;
3226 *bufp = buf->b_next;
3228 ASSERT(buf->b_data != NULL);
3229 arc_buf_destroy(buf, FALSE, FALSE);
3231 if (hdr->b_datacnt == 0) {
3232 arc_state_t *old_state = hdr->b_state;
3233 arc_state_t *evicted_state;
3235 ASSERT(hdr->b_buf == NULL);
3236 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3239 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3241 mutex_enter(&old_state->arcs_mtx);
3242 mutex_enter(&evicted_state->arcs_mtx);
3244 arc_change_state(evicted_state, hdr, hash_lock);
3245 ASSERT(HDR_IN_HASH_TABLE(hdr));
3246 hdr->b_flags |= ARC_IN_HASH_TABLE;
3247 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3249 mutex_exit(&evicted_state->arcs_mtx);
3250 mutex_exit(&old_state->arcs_mtx);
3252 mutex_exit(hash_lock);
3253 mutex_exit(&buf->b_evict_lock);
3255 VERIFY(buf->b_efunc(buf) == 0);
3256 buf->b_efunc = NULL;
3257 buf->b_private = NULL;
3260 kmem_cache_free(buf_cache, buf);
3265 * Release this buffer from the cache. This must be done
3266 * after a read and prior to modifying the buffer contents.
3267 * If the buffer has more than one reference, we must make
3268 * a new hdr for the buffer.
3271 arc_release(arc_buf_t *buf, void *tag)
3274 kmutex_t *hash_lock = NULL;
3275 l2arc_buf_hdr_t *l2hdr;
3276 uint64_t buf_size = 0;
3279 * It would be nice to assert that if it's DMU metadata (level >
3280 * 0 || it's the dnode file), then it must be syncing context.
3281 * But we don't know that information at this level.
3284 mutex_enter(&buf->b_evict_lock);
3287 /* this buffer is not on any list */
3288 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3290 if (hdr->b_state == arc_anon) {
3291 /* this buffer is already released */
3292 ASSERT(buf->b_efunc == NULL);
3294 hash_lock = HDR_LOCK(hdr);
3295 mutex_enter(hash_lock);
3297 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3300 l2hdr = hdr->b_l2hdr;
3302 mutex_enter(&l2arc_buflist_mtx);
3303 hdr->b_l2hdr = NULL;
3304 buf_size = hdr->b_size;
3308 * Do we have more than one buf?
3310 if (hdr->b_datacnt > 1) {
3311 arc_buf_hdr_t *nhdr;
3313 uint64_t blksz = hdr->b_size;
3314 uint64_t spa = hdr->b_spa;
3315 arc_buf_contents_t type = hdr->b_type;
3316 uint32_t flags = hdr->b_flags;
3318 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3320 * Pull the data off of this hdr and attach it to
3321 * a new anonymous hdr.
3323 (void) remove_reference(hdr, hash_lock, tag);
3325 while (*bufp != buf)
3326 bufp = &(*bufp)->b_next;
3327 *bufp = buf->b_next;
3330 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3331 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3332 if (refcount_is_zero(&hdr->b_refcnt)) {
3333 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3334 ASSERT3U(*size, >=, hdr->b_size);
3335 atomic_add_64(size, -hdr->b_size);
3337 hdr->b_datacnt -= 1;
3338 arc_cksum_verify(buf);
3340 mutex_exit(hash_lock);
3342 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3343 nhdr->b_size = blksz;
3345 nhdr->b_type = type;
3347 nhdr->b_state = arc_anon;
3348 nhdr->b_arc_access = 0;
3349 nhdr->b_flags = flags & ARC_L2_WRITING;
3350 nhdr->b_l2hdr = NULL;
3351 nhdr->b_datacnt = 1;
3352 nhdr->b_freeze_cksum = NULL;
3353 (void) refcount_add(&nhdr->b_refcnt, tag);
3355 mutex_exit(&buf->b_evict_lock);
3356 atomic_add_64(&arc_anon->arcs_size, blksz);
3358 mutex_exit(&buf->b_evict_lock);
3359 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3360 ASSERT(!list_link_active(&hdr->b_arc_node));
3361 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3362 if (hdr->b_state != arc_anon)
3363 arc_change_state(arc_anon, hdr, hash_lock);
3364 hdr->b_arc_access = 0;
3366 mutex_exit(hash_lock);
3368 buf_discard_identity(hdr);
3371 buf->b_efunc = NULL;
3372 buf->b_private = NULL;
3375 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3376 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3377 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3378 mutex_exit(&l2arc_buflist_mtx);
3383 * Release this buffer. If it does not match the provided BP, fill it
3384 * with that block's contents.
3388 arc_release_bp(arc_buf_t *buf, void *tag, blkptr_t *bp, spa_t *spa,
3391 arc_release(buf, tag);
3396 arc_released(arc_buf_t *buf)
3400 mutex_enter(&buf->b_evict_lock);
3401 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3402 mutex_exit(&buf->b_evict_lock);
3407 arc_has_callback(arc_buf_t *buf)
3411 mutex_enter(&buf->b_evict_lock);
3412 callback = (buf->b_efunc != NULL);
3413 mutex_exit(&buf->b_evict_lock);
3419 arc_referenced(arc_buf_t *buf)
3423 mutex_enter(&buf->b_evict_lock);
3424 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3425 mutex_exit(&buf->b_evict_lock);
3426 return (referenced);
3431 arc_write_ready(zio_t *zio)
3433 arc_write_callback_t *callback = zio->io_private;
3434 arc_buf_t *buf = callback->awcb_buf;
3435 arc_buf_hdr_t *hdr = buf->b_hdr;
3437 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3438 callback->awcb_ready(zio, buf, callback->awcb_private);
3441 * If the IO is already in progress, then this is a re-write
3442 * attempt, so we need to thaw and re-compute the cksum.
3443 * It is the responsibility of the callback to handle the
3444 * accounting for any re-write attempt.
3446 if (HDR_IO_IN_PROGRESS(hdr)) {
3447 mutex_enter(&hdr->b_freeze_lock);
3448 if (hdr->b_freeze_cksum != NULL) {
3449 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3450 hdr->b_freeze_cksum = NULL;
3452 mutex_exit(&hdr->b_freeze_lock);
3454 arc_cksum_compute(buf, B_FALSE);
3455 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3459 arc_write_done(zio_t *zio)
3461 arc_write_callback_t *callback = zio->io_private;
3462 arc_buf_t *buf = callback->awcb_buf;
3463 arc_buf_hdr_t *hdr = buf->b_hdr;
3465 ASSERT(hdr->b_acb == NULL);
3467 if (zio->io_error == 0) {
3468 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3469 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3470 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3472 ASSERT(BUF_EMPTY(hdr));
3476 * If the block to be written was all-zero, we may have
3477 * compressed it away. In this case no write was performed
3478 * so there will be no dva/birth/checksum. The buffer must
3479 * therefore remain anonymous (and uncached).
3481 if (!BUF_EMPTY(hdr)) {
3482 arc_buf_hdr_t *exists;
3483 kmutex_t *hash_lock;
3485 ASSERT(zio->io_error == 0);
3487 arc_cksum_verify(buf);
3489 exists = buf_hash_insert(hdr, &hash_lock);
3492 * This can only happen if we overwrite for
3493 * sync-to-convergence, because we remove
3494 * buffers from the hash table when we arc_free().
3496 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3497 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3498 panic("bad overwrite, hdr=%p exists=%p",
3499 (void *)hdr, (void *)exists);
3500 ASSERT(refcount_is_zero(&exists->b_refcnt));
3501 arc_change_state(arc_anon, exists, hash_lock);
3502 mutex_exit(hash_lock);
3503 arc_hdr_destroy(exists);
3504 exists = buf_hash_insert(hdr, &hash_lock);
3505 ASSERT3P(exists, ==, NULL);
3508 ASSERT(hdr->b_datacnt == 1);
3509 ASSERT(hdr->b_state == arc_anon);
3510 ASSERT(BP_GET_DEDUP(zio->io_bp));
3511 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3514 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3515 /* if it's not anon, we are doing a scrub */
3516 if (!exists && hdr->b_state == arc_anon)
3517 arc_access(hdr, hash_lock);
3518 mutex_exit(hash_lock);
3520 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3523 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3524 callback->awcb_done(zio, buf, callback->awcb_private);
3526 kmem_free(callback, sizeof (arc_write_callback_t));
3530 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3531 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3532 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3533 int priority, int zio_flags, const zbookmark_t *zb)
3535 arc_buf_hdr_t *hdr = buf->b_hdr;
3536 arc_write_callback_t *callback;
3539 ASSERT(ready != NULL);
3540 ASSERT(done != NULL);
3541 ASSERT(!HDR_IO_ERROR(hdr));
3542 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3543 ASSERT(hdr->b_acb == NULL);
3545 hdr->b_flags |= ARC_L2CACHE;
3546 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_PUSHPAGE);
3547 callback->awcb_ready = ready;
3548 callback->awcb_done = done;
3549 callback->awcb_private = private;
3550 callback->awcb_buf = buf;
3552 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3553 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3559 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3562 uint64_t available_memory;
3564 /* Easily reclaimable memory (free + inactive + arc-evictable) */
3565 available_memory = ptob(spl_kmem_availrmem()) + arc_evictable_memory();
3568 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3571 if (available_memory <= zfs_write_limit_max) {
3572 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3573 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
3577 if (inflight_data > available_memory / 4) {
3578 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3579 DMU_TX_STAT_BUMP(dmu_tx_memory_inflight);
3587 arc_tempreserve_clear(uint64_t reserve)
3589 atomic_add_64(&arc_tempreserve, -reserve);
3590 ASSERT((int64_t)arc_tempreserve >= 0);
3594 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3601 * Once in a while, fail for no reason. Everything should cope.
3603 if (spa_get_random(10000) == 0) {
3604 dprintf("forcing random failure\n");
3608 if (reserve > arc_c/4 && !arc_no_grow)
3609 arc_c = MIN(arc_c_max, reserve * 4);
3610 if (reserve > arc_c) {
3611 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
3616 * Don't count loaned bufs as in flight dirty data to prevent long
3617 * network delays from blocking transactions that are ready to be
3618 * assigned to a txg.
3620 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3623 * Writes will, almost always, require additional memory allocations
3624 * in order to compress/encrypt/etc the data. We therefor need to
3625 * make sure that there is sufficient available memory for this.
3627 if ((error = arc_memory_throttle(reserve, anon_size, txg)))
3631 * Throttle writes when the amount of dirty data in the cache
3632 * gets too large. We try to keep the cache less than half full
3633 * of dirty blocks so that our sync times don't grow too large.
3634 * Note: if two requests come in concurrently, we might let them
3635 * both succeed, when one of them should fail. Not a huge deal.
3638 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3639 anon_size > arc_c / 4) {
3640 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3641 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3642 arc_tempreserve>>10,
3643 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3644 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3645 reserve>>10, arc_c>>10);
3646 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
3649 atomic_add_64(&arc_tempreserve, reserve);
3654 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
3655 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
3657 size->value.ui64 = state->arcs_size;
3658 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
3659 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
3663 arc_kstat_update(kstat_t *ksp, int rw)
3665 arc_stats_t *as = ksp->ks_data;
3667 if (rw == KSTAT_WRITE) {
3670 arc_kstat_update_state(arc_anon,
3671 &as->arcstat_anon_size,
3672 &as->arcstat_anon_evict_data,
3673 &as->arcstat_anon_evict_metadata);
3674 arc_kstat_update_state(arc_mru,
3675 &as->arcstat_mru_size,
3676 &as->arcstat_mru_evict_data,
3677 &as->arcstat_mru_evict_metadata);
3678 arc_kstat_update_state(arc_mru_ghost,
3679 &as->arcstat_mru_ghost_size,
3680 &as->arcstat_mru_ghost_evict_data,
3681 &as->arcstat_mru_ghost_evict_metadata);
3682 arc_kstat_update_state(arc_mfu,
3683 &as->arcstat_mfu_size,
3684 &as->arcstat_mfu_evict_data,
3685 &as->arcstat_mfu_evict_metadata);
3686 arc_kstat_update_state(arc_mfu_ghost,
3687 &as->arcstat_mfu_ghost_size,
3688 &as->arcstat_mfu_ghost_evict_data,
3689 &as->arcstat_mfu_ghost_evict_metadata);
3698 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3699 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3701 /* Convert seconds to clock ticks */
3702 arc_min_prefetch_lifespan = 1 * hz;
3704 /* Start out with 1/8 of all memory */
3705 arc_c = physmem * PAGESIZE / 8;
3709 * On architectures where the physical memory can be larger
3710 * than the addressable space (intel in 32-bit mode), we may
3711 * need to limit the cache to 1/8 of VM size.
3713 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3715 * Register a shrinker to support synchronous (direct) memory
3716 * reclaim from the arc. This is done to prevent kswapd from
3717 * swapping out pages when it is preferable to shrink the arc.
3719 spl_register_shrinker(&arc_shrinker);
3722 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3723 arc_c_min = MAX(arc_c / 4, 64<<20);
3724 /* set max to 1/2 of all memory */
3725 arc_c_max = MAX(arc_c * 4, arc_c_max);
3728 * Allow the tunables to override our calculations if they are
3729 * reasonable (ie. over 64MB)
3731 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3732 arc_c_max = zfs_arc_max;
3733 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3734 arc_c_min = zfs_arc_min;
3737 arc_p = (arc_c >> 1);
3739 /* limit meta-data to 1/4 of the arc capacity */
3740 arc_meta_limit = arc_c_max / 4;
3743 /* Allow the tunable to override if it is reasonable */
3744 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3745 arc_meta_limit = zfs_arc_meta_limit;
3747 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3748 arc_c_min = arc_meta_limit / 2;
3750 if (zfs_arc_grow_retry > 0)
3751 arc_grow_retry = zfs_arc_grow_retry;
3753 if (zfs_arc_shrink_shift > 0)
3754 arc_shrink_shift = zfs_arc_shrink_shift;
3756 if (zfs_arc_p_min_shift > 0)
3757 arc_p_min_shift = zfs_arc_p_min_shift;
3759 if (zfs_arc_meta_prune > 0)
3760 arc_meta_prune = zfs_arc_meta_prune;
3762 /* if kmem_flags are set, lets try to use less memory */
3763 if (kmem_debugging())
3765 if (arc_c < arc_c_min)
3768 arc_anon = &ARC_anon;
3770 arc_mru_ghost = &ARC_mru_ghost;
3772 arc_mfu_ghost = &ARC_mfu_ghost;
3773 arc_l2c_only = &ARC_l2c_only;
3776 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3777 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3778 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3779 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3780 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3781 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3783 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3784 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3785 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3786 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3787 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3788 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3789 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3790 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3791 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3792 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3793 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3794 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3795 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3796 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3797 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3798 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3799 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3800 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3801 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3802 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3806 arc_thread_exit = 0;
3807 list_create(&arc_prune_list, sizeof (arc_prune_t),
3808 offsetof(arc_prune_t, p_node));
3809 arc_eviction_list = NULL;
3810 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
3811 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3812 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3814 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3815 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3817 if (arc_ksp != NULL) {
3818 arc_ksp->ks_data = &arc_stats;
3819 arc_ksp->ks_update = arc_kstat_update;
3820 kstat_install(arc_ksp);
3823 (void) thread_create(NULL, 0, arc_adapt_thread, NULL, 0, &p0,
3824 TS_RUN, minclsyspri);
3829 if (zfs_write_limit_max == 0)
3830 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3832 zfs_write_limit_shift = 0;
3833 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3841 mutex_enter(&arc_reclaim_thr_lock);
3843 spl_unregister_shrinker(&arc_shrinker);
3844 #endif /* _KERNEL */
3846 arc_thread_exit = 1;
3847 while (arc_thread_exit != 0)
3848 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3849 mutex_exit(&arc_reclaim_thr_lock);
3855 if (arc_ksp != NULL) {
3856 kstat_delete(arc_ksp);
3860 mutex_enter(&arc_prune_mtx);
3861 while ((p = list_head(&arc_prune_list)) != NULL) {
3862 list_remove(&arc_prune_list, p);
3863 refcount_remove(&p->p_refcnt, &arc_prune_list);
3864 refcount_destroy(&p->p_refcnt);
3865 kmem_free(p, sizeof (*p));
3867 mutex_exit(&arc_prune_mtx);
3869 list_destroy(&arc_prune_list);
3870 mutex_destroy(&arc_prune_mtx);
3871 mutex_destroy(&arc_eviction_mtx);
3872 mutex_destroy(&arc_reclaim_thr_lock);
3873 cv_destroy(&arc_reclaim_thr_cv);
3875 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3876 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3877 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3878 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3879 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3880 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3881 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3882 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3884 mutex_destroy(&arc_anon->arcs_mtx);
3885 mutex_destroy(&arc_mru->arcs_mtx);
3886 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3887 mutex_destroy(&arc_mfu->arcs_mtx);
3888 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3889 mutex_destroy(&arc_l2c_only->arcs_mtx);
3891 mutex_destroy(&zfs_write_limit_lock);
3895 ASSERT(arc_loaned_bytes == 0);
3901 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3902 * It uses dedicated storage devices to hold cached data, which are populated
3903 * using large infrequent writes. The main role of this cache is to boost
3904 * the performance of random read workloads. The intended L2ARC devices
3905 * include short-stroked disks, solid state disks, and other media with
3906 * substantially faster read latency than disk.
3908 * +-----------------------+
3910 * +-----------------------+
3913 * l2arc_feed_thread() arc_read()
3917 * +---------------+ |
3919 * +---------------+ |
3924 * +-------+ +-------+
3926 * | cache | | cache |
3927 * +-------+ +-------+
3928 * +=========+ .-----.
3929 * : L2ARC : |-_____-|
3930 * : devices : | Disks |
3931 * +=========+ `-_____-'
3933 * Read requests are satisfied from the following sources, in order:
3936 * 2) vdev cache of L2ARC devices
3938 * 4) vdev cache of disks
3941 * Some L2ARC device types exhibit extremely slow write performance.
3942 * To accommodate for this there are some significant differences between
3943 * the L2ARC and traditional cache design:
3945 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3946 * the ARC behave as usual, freeing buffers and placing headers on ghost
3947 * lists. The ARC does not send buffers to the L2ARC during eviction as
3948 * this would add inflated write latencies for all ARC memory pressure.
3950 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3951 * It does this by periodically scanning buffers from the eviction-end of
3952 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3953 * not already there. It scans until a headroom of buffers is satisfied,
3954 * which itself is a buffer for ARC eviction. The thread that does this is
3955 * l2arc_feed_thread(), illustrated below; example sizes are included to
3956 * provide a better sense of ratio than this diagram:
3959 * +---------------------+----------+
3960 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3961 * +---------------------+----------+ | o L2ARC eligible
3962 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
3963 * +---------------------+----------+ |
3964 * 15.9 Gbytes ^ 32 Mbytes |
3966 * l2arc_feed_thread()
3968 * l2arc write hand <--[oooo]--'
3972 * +==============================+
3973 * L2ARC dev |####|#|###|###| |####| ... |
3974 * +==============================+
3977 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
3978 * evicted, then the L2ARC has cached a buffer much sooner than it probably
3979 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
3980 * safe to say that this is an uncommon case, since buffers at the end of
3981 * the ARC lists have moved there due to inactivity.
3983 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
3984 * then the L2ARC simply misses copying some buffers. This serves as a
3985 * pressure valve to prevent heavy read workloads from both stalling the ARC
3986 * with waits and clogging the L2ARC with writes. This also helps prevent
3987 * the potential for the L2ARC to churn if it attempts to cache content too
3988 * quickly, such as during backups of the entire pool.
3990 * 5. After system boot and before the ARC has filled main memory, there are
3991 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
3992 * lists can remain mostly static. Instead of searching from tail of these
3993 * lists as pictured, the l2arc_feed_thread() will search from the list heads
3994 * for eligible buffers, greatly increasing its chance of finding them.
3996 * The L2ARC device write speed is also boosted during this time so that
3997 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
3998 * there are no L2ARC reads, and no fear of degrading read performance
3999 * through increased writes.
4001 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4002 * the vdev queue can aggregate them into larger and fewer writes. Each
4003 * device is written to in a rotor fashion, sweeping writes through
4004 * available space then repeating.
4006 * 7. The L2ARC does not store dirty content. It never needs to flush
4007 * write buffers back to disk based storage.
4009 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4010 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4012 * The performance of the L2ARC can be tweaked by a number of tunables, which
4013 * may be necessary for different workloads:
4015 * l2arc_write_max max write bytes per interval
4016 * l2arc_write_boost extra write bytes during device warmup
4017 * l2arc_noprefetch skip caching prefetched buffers
4018 * l2arc_headroom number of max device writes to precache
4019 * l2arc_feed_secs seconds between L2ARC writing
4021 * Tunables may be removed or added as future performance improvements are
4022 * integrated, and also may become zpool properties.
4024 * There are three key functions that control how the L2ARC warms up:
4026 * l2arc_write_eligible() check if a buffer is eligible to cache
4027 * l2arc_write_size() calculate how much to write
4028 * l2arc_write_interval() calculate sleep delay between writes
4030 * These three functions determine what to write, how much, and how quickly
4035 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4038 * A buffer is *not* eligible for the L2ARC if it:
4039 * 1. belongs to a different spa.
4040 * 2. is already cached on the L2ARC.
4041 * 3. has an I/O in progress (it may be an incomplete read).
4042 * 4. is flagged not eligible (zfs property).
4044 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
4045 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
4052 l2arc_write_size(l2arc_dev_t *dev)
4056 size = dev->l2ad_write;
4058 if (arc_warm == B_FALSE)
4059 size += dev->l2ad_boost;
4066 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4068 clock_t interval, next, now;
4071 * If the ARC lists are busy, increase our write rate; if the
4072 * lists are stale, idle back. This is achieved by checking
4073 * how much we previously wrote - if it was more than half of
4074 * what we wanted, schedule the next write much sooner.
4076 if (l2arc_feed_again && wrote > (wanted / 2))
4077 interval = (hz * l2arc_feed_min_ms) / 1000;
4079 interval = hz * l2arc_feed_secs;
4081 now = ddi_get_lbolt();
4082 next = MAX(now, MIN(now + interval, began + interval));
4088 l2arc_hdr_stat_add(void)
4090 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4091 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4095 l2arc_hdr_stat_remove(void)
4097 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4098 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4102 * Cycle through L2ARC devices. This is how L2ARC load balances.
4103 * If a device is returned, this also returns holding the spa config lock.
4105 static l2arc_dev_t *
4106 l2arc_dev_get_next(void)
4108 l2arc_dev_t *first, *next = NULL;
4111 * Lock out the removal of spas (spa_namespace_lock), then removal
4112 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4113 * both locks will be dropped and a spa config lock held instead.
4115 mutex_enter(&spa_namespace_lock);
4116 mutex_enter(&l2arc_dev_mtx);
4118 /* if there are no vdevs, there is nothing to do */
4119 if (l2arc_ndev == 0)
4123 next = l2arc_dev_last;
4125 /* loop around the list looking for a non-faulted vdev */
4127 next = list_head(l2arc_dev_list);
4129 next = list_next(l2arc_dev_list, next);
4131 next = list_head(l2arc_dev_list);
4134 /* if we have come back to the start, bail out */
4137 else if (next == first)
4140 } while (vdev_is_dead(next->l2ad_vdev));
4142 /* if we were unable to find any usable vdevs, return NULL */
4143 if (vdev_is_dead(next->l2ad_vdev))
4146 l2arc_dev_last = next;
4149 mutex_exit(&l2arc_dev_mtx);
4152 * Grab the config lock to prevent the 'next' device from being
4153 * removed while we are writing to it.
4156 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4157 mutex_exit(&spa_namespace_lock);
4163 * Free buffers that were tagged for destruction.
4166 l2arc_do_free_on_write(void)
4169 l2arc_data_free_t *df, *df_prev;
4171 mutex_enter(&l2arc_free_on_write_mtx);
4172 buflist = l2arc_free_on_write;
4174 for (df = list_tail(buflist); df; df = df_prev) {
4175 df_prev = list_prev(buflist, df);
4176 ASSERT(df->l2df_data != NULL);
4177 ASSERT(df->l2df_func != NULL);
4178 df->l2df_func(df->l2df_data, df->l2df_size);
4179 list_remove(buflist, df);
4180 kmem_free(df, sizeof (l2arc_data_free_t));
4183 mutex_exit(&l2arc_free_on_write_mtx);
4187 * A write to a cache device has completed. Update all headers to allow
4188 * reads from these buffers to begin.
4191 l2arc_write_done(zio_t *zio)
4193 l2arc_write_callback_t *cb;
4196 arc_buf_hdr_t *head, *ab, *ab_prev;
4197 l2arc_buf_hdr_t *abl2;
4198 kmutex_t *hash_lock;
4200 cb = zio->io_private;
4202 dev = cb->l2wcb_dev;
4203 ASSERT(dev != NULL);
4204 head = cb->l2wcb_head;
4205 ASSERT(head != NULL);
4206 buflist = dev->l2ad_buflist;
4207 ASSERT(buflist != NULL);
4208 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4209 l2arc_write_callback_t *, cb);
4211 if (zio->io_error != 0)
4212 ARCSTAT_BUMP(arcstat_l2_writes_error);
4214 mutex_enter(&l2arc_buflist_mtx);
4217 * All writes completed, or an error was hit.
4219 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4220 ab_prev = list_prev(buflist, ab);
4222 hash_lock = HDR_LOCK(ab);
4223 if (!mutex_tryenter(hash_lock)) {
4225 * This buffer misses out. It may be in a stage
4226 * of eviction. Its ARC_L2_WRITING flag will be
4227 * left set, denying reads to this buffer.
4229 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4233 if (zio->io_error != 0) {
4235 * Error - drop L2ARC entry.
4237 list_remove(buflist, ab);
4240 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4241 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4245 * Allow ARC to begin reads to this L2ARC entry.
4247 ab->b_flags &= ~ARC_L2_WRITING;
4249 mutex_exit(hash_lock);
4252 atomic_inc_64(&l2arc_writes_done);
4253 list_remove(buflist, head);
4254 kmem_cache_free(hdr_cache, head);
4255 mutex_exit(&l2arc_buflist_mtx);
4257 l2arc_do_free_on_write();
4259 kmem_free(cb, sizeof (l2arc_write_callback_t));
4263 * A read to a cache device completed. Validate buffer contents before
4264 * handing over to the regular ARC routines.
4267 l2arc_read_done(zio_t *zio)
4269 l2arc_read_callback_t *cb;
4272 kmutex_t *hash_lock;
4275 ASSERT(zio->io_vd != NULL);
4276 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4278 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4280 cb = zio->io_private;
4282 buf = cb->l2rcb_buf;
4283 ASSERT(buf != NULL);
4285 hash_lock = HDR_LOCK(buf->b_hdr);
4286 mutex_enter(hash_lock);
4288 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4291 * Check this survived the L2ARC journey.
4293 equal = arc_cksum_equal(buf);
4294 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4295 mutex_exit(hash_lock);
4296 zio->io_private = buf;
4297 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4298 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4301 mutex_exit(hash_lock);
4303 * Buffer didn't survive caching. Increment stats and
4304 * reissue to the original storage device.
4306 if (zio->io_error != 0) {
4307 ARCSTAT_BUMP(arcstat_l2_io_error);
4309 zio->io_error = EIO;
4312 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4315 * If there's no waiter, issue an async i/o to the primary
4316 * storage now. If there *is* a waiter, the caller must
4317 * issue the i/o in a context where it's OK to block.
4319 if (zio->io_waiter == NULL) {
4320 zio_t *pio = zio_unique_parent(zio);
4322 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4324 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4325 buf->b_data, zio->io_size, arc_read_done, buf,
4326 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4330 kmem_free(cb, sizeof (l2arc_read_callback_t));
4334 * This is the list priority from which the L2ARC will search for pages to
4335 * cache. This is used within loops (0..3) to cycle through lists in the
4336 * desired order. This order can have a significant effect on cache
4339 * Currently the metadata lists are hit first, MFU then MRU, followed by
4340 * the data lists. This function returns a locked list, and also returns
4344 l2arc_list_locked(int list_num, kmutex_t **lock)
4346 list_t *list = NULL;
4348 ASSERT(list_num >= 0 && list_num <= 3);
4352 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4353 *lock = &arc_mfu->arcs_mtx;
4356 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4357 *lock = &arc_mru->arcs_mtx;
4360 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4361 *lock = &arc_mfu->arcs_mtx;
4364 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4365 *lock = &arc_mru->arcs_mtx;
4369 ASSERT(!(MUTEX_HELD(*lock)));
4375 * Evict buffers from the device write hand to the distance specified in
4376 * bytes. This distance may span populated buffers, it may span nothing.
4377 * This is clearing a region on the L2ARC device ready for writing.
4378 * If the 'all' boolean is set, every buffer is evicted.
4381 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4384 l2arc_buf_hdr_t *abl2;
4385 arc_buf_hdr_t *ab, *ab_prev;
4386 kmutex_t *hash_lock;
4389 buflist = dev->l2ad_buflist;
4391 if (buflist == NULL)
4394 if (!all && dev->l2ad_first) {
4396 * This is the first sweep through the device. There is
4402 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4404 * When nearing the end of the device, evict to the end
4405 * before the device write hand jumps to the start.
4407 taddr = dev->l2ad_end;
4409 taddr = dev->l2ad_hand + distance;
4411 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4412 uint64_t, taddr, boolean_t, all);
4415 mutex_enter(&l2arc_buflist_mtx);
4416 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4417 ab_prev = list_prev(buflist, ab);
4419 hash_lock = HDR_LOCK(ab);
4420 if (!mutex_tryenter(hash_lock)) {
4422 * Missed the hash lock. Retry.
4424 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4425 mutex_exit(&l2arc_buflist_mtx);
4426 mutex_enter(hash_lock);
4427 mutex_exit(hash_lock);
4431 if (HDR_L2_WRITE_HEAD(ab)) {
4433 * We hit a write head node. Leave it for
4434 * l2arc_write_done().
4436 list_remove(buflist, ab);
4437 mutex_exit(hash_lock);
4441 if (!all && ab->b_l2hdr != NULL &&
4442 (ab->b_l2hdr->b_daddr > taddr ||
4443 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4445 * We've evicted to the target address,
4446 * or the end of the device.
4448 mutex_exit(hash_lock);
4452 if (HDR_FREE_IN_PROGRESS(ab)) {
4454 * Already on the path to destruction.
4456 mutex_exit(hash_lock);
4460 if (ab->b_state == arc_l2c_only) {
4461 ASSERT(!HDR_L2_READING(ab));
4463 * This doesn't exist in the ARC. Destroy.
4464 * arc_hdr_destroy() will call list_remove()
4465 * and decrement arcstat_l2_size.
4467 arc_change_state(arc_anon, ab, hash_lock);
4468 arc_hdr_destroy(ab);
4471 * Invalidate issued or about to be issued
4472 * reads, since we may be about to write
4473 * over this location.
4475 if (HDR_L2_READING(ab)) {
4476 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4477 ab->b_flags |= ARC_L2_EVICTED;
4481 * Tell ARC this no longer exists in L2ARC.
4483 if (ab->b_l2hdr != NULL) {
4486 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4487 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4489 list_remove(buflist, ab);
4492 * This may have been leftover after a
4495 ab->b_flags &= ~ARC_L2_WRITING;
4497 mutex_exit(hash_lock);
4499 mutex_exit(&l2arc_buflist_mtx);
4501 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4502 dev->l2ad_evict = taddr;
4506 * Find and write ARC buffers to the L2ARC device.
4508 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4509 * for reading until they have completed writing.
4512 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4514 arc_buf_hdr_t *ab, *ab_prev, *head;
4515 l2arc_buf_hdr_t *hdrl2;
4517 uint64_t passed_sz, write_sz, buf_sz, headroom;
4519 kmutex_t *hash_lock, *list_lock = NULL;
4520 boolean_t have_lock, full;
4521 l2arc_write_callback_t *cb;
4523 uint64_t guid = spa_load_guid(spa);
4526 ASSERT(dev->l2ad_vdev != NULL);
4531 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4532 head->b_flags |= ARC_L2_WRITE_HEAD;
4535 * Copy buffers for L2ARC writing.
4537 mutex_enter(&l2arc_buflist_mtx);
4538 for (try = 0; try <= 3; try++) {
4539 list = l2arc_list_locked(try, &list_lock);
4543 * L2ARC fast warmup.
4545 * Until the ARC is warm and starts to evict, read from the
4546 * head of the ARC lists rather than the tail.
4548 headroom = target_sz * l2arc_headroom;
4549 if (arc_warm == B_FALSE)
4550 ab = list_head(list);
4552 ab = list_tail(list);
4554 for (; ab; ab = ab_prev) {
4555 if (arc_warm == B_FALSE)
4556 ab_prev = list_next(list, ab);
4558 ab_prev = list_prev(list, ab);
4560 hash_lock = HDR_LOCK(ab);
4561 have_lock = MUTEX_HELD(hash_lock);
4562 if (!have_lock && !mutex_tryenter(hash_lock)) {
4564 * Skip this buffer rather than waiting.
4569 passed_sz += ab->b_size;
4570 if (passed_sz > headroom) {
4574 mutex_exit(hash_lock);
4578 if (!l2arc_write_eligible(guid, ab)) {
4579 mutex_exit(hash_lock);
4583 if ((write_sz + ab->b_size) > target_sz) {
4585 mutex_exit(hash_lock);
4591 * Insert a dummy header on the buflist so
4592 * l2arc_write_done() can find where the
4593 * write buffers begin without searching.
4595 list_insert_head(dev->l2ad_buflist, head);
4597 cb = kmem_alloc(sizeof (l2arc_write_callback_t),
4599 cb->l2wcb_dev = dev;
4600 cb->l2wcb_head = head;
4601 pio = zio_root(spa, l2arc_write_done, cb,
4606 * Create and add a new L2ARC header.
4608 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t),
4611 hdrl2->b_daddr = dev->l2ad_hand;
4613 ab->b_flags |= ARC_L2_WRITING;
4614 ab->b_l2hdr = hdrl2;
4615 list_insert_head(dev->l2ad_buflist, ab);
4616 buf_data = ab->b_buf->b_data;
4617 buf_sz = ab->b_size;
4620 * Compute and store the buffer cksum before
4621 * writing. On debug the cksum is verified first.
4623 arc_cksum_verify(ab->b_buf);
4624 arc_cksum_compute(ab->b_buf, B_TRUE);
4626 mutex_exit(hash_lock);
4628 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4629 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4630 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4631 ZIO_FLAG_CANFAIL, B_FALSE);
4633 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4635 (void) zio_nowait(wzio);
4638 * Keep the clock hand suitably device-aligned.
4640 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4643 dev->l2ad_hand += buf_sz;
4646 mutex_exit(list_lock);
4651 mutex_exit(&l2arc_buflist_mtx);
4654 ASSERT3U(write_sz, ==, 0);
4655 kmem_cache_free(hdr_cache, head);
4659 ASSERT3U(write_sz, <=, target_sz);
4660 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4661 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4662 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4663 vdev_space_update(dev->l2ad_vdev, write_sz, 0, 0);
4666 * Bump device hand to the device start if it is approaching the end.
4667 * l2arc_evict() will already have evicted ahead for this case.
4669 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4670 vdev_space_update(dev->l2ad_vdev,
4671 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4672 dev->l2ad_hand = dev->l2ad_start;
4673 dev->l2ad_evict = dev->l2ad_start;
4674 dev->l2ad_first = B_FALSE;
4677 dev->l2ad_writing = B_TRUE;
4678 (void) zio_wait(pio);
4679 dev->l2ad_writing = B_FALSE;
4685 * This thread feeds the L2ARC at regular intervals. This is the beating
4686 * heart of the L2ARC.
4689 l2arc_feed_thread(void)
4694 uint64_t size, wrote;
4695 clock_t begin, next = ddi_get_lbolt();
4697 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4699 mutex_enter(&l2arc_feed_thr_lock);
4701 while (l2arc_thread_exit == 0) {
4702 CALLB_CPR_SAFE_BEGIN(&cpr);
4703 (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv,
4704 &l2arc_feed_thr_lock, next);
4705 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4706 next = ddi_get_lbolt() + hz;
4709 * Quick check for L2ARC devices.
4711 mutex_enter(&l2arc_dev_mtx);
4712 if (l2arc_ndev == 0) {
4713 mutex_exit(&l2arc_dev_mtx);
4716 mutex_exit(&l2arc_dev_mtx);
4717 begin = ddi_get_lbolt();
4720 * This selects the next l2arc device to write to, and in
4721 * doing so the next spa to feed from: dev->l2ad_spa. This
4722 * will return NULL if there are now no l2arc devices or if
4723 * they are all faulted.
4725 * If a device is returned, its spa's config lock is also
4726 * held to prevent device removal. l2arc_dev_get_next()
4727 * will grab and release l2arc_dev_mtx.
4729 if ((dev = l2arc_dev_get_next()) == NULL)
4732 spa = dev->l2ad_spa;
4733 ASSERT(spa != NULL);
4736 * If the pool is read-only then force the feed thread to
4737 * sleep a little longer.
4739 if (!spa_writeable(spa)) {
4740 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4741 spa_config_exit(spa, SCL_L2ARC, dev);
4746 * Avoid contributing to memory pressure.
4749 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4750 spa_config_exit(spa, SCL_L2ARC, dev);
4754 ARCSTAT_BUMP(arcstat_l2_feeds);
4756 size = l2arc_write_size(dev);
4759 * Evict L2ARC buffers that will be overwritten.
4761 l2arc_evict(dev, size, B_FALSE);
4764 * Write ARC buffers.
4766 wrote = l2arc_write_buffers(spa, dev, size);
4769 * Calculate interval between writes.
4771 next = l2arc_write_interval(begin, size, wrote);
4772 spa_config_exit(spa, SCL_L2ARC, dev);
4775 l2arc_thread_exit = 0;
4776 cv_broadcast(&l2arc_feed_thr_cv);
4777 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4782 l2arc_vdev_present(vdev_t *vd)
4786 mutex_enter(&l2arc_dev_mtx);
4787 for (dev = list_head(l2arc_dev_list); dev != NULL;
4788 dev = list_next(l2arc_dev_list, dev)) {
4789 if (dev->l2ad_vdev == vd)
4792 mutex_exit(&l2arc_dev_mtx);
4794 return (dev != NULL);
4798 * Add a vdev for use by the L2ARC. By this point the spa has already
4799 * validated the vdev and opened it.
4802 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
4804 l2arc_dev_t *adddev;
4806 ASSERT(!l2arc_vdev_present(vd));
4809 * Create a new l2arc device entry.
4811 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4812 adddev->l2ad_spa = spa;
4813 adddev->l2ad_vdev = vd;
4814 adddev->l2ad_write = l2arc_write_max;
4815 adddev->l2ad_boost = l2arc_write_boost;
4816 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
4817 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
4818 adddev->l2ad_hand = adddev->l2ad_start;
4819 adddev->l2ad_evict = adddev->l2ad_start;
4820 adddev->l2ad_first = B_TRUE;
4821 adddev->l2ad_writing = B_FALSE;
4822 list_link_init(&adddev->l2ad_node);
4823 ASSERT3U(adddev->l2ad_write, >, 0);
4826 * This is a list of all ARC buffers that are still valid on the
4829 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4830 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4831 offsetof(arc_buf_hdr_t, b_l2node));
4833 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
4836 * Add device to global list
4838 mutex_enter(&l2arc_dev_mtx);
4839 list_insert_head(l2arc_dev_list, adddev);
4840 atomic_inc_64(&l2arc_ndev);
4841 mutex_exit(&l2arc_dev_mtx);
4845 * Remove a vdev from the L2ARC.
4848 l2arc_remove_vdev(vdev_t *vd)
4850 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4853 * Find the device by vdev
4855 mutex_enter(&l2arc_dev_mtx);
4856 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4857 nextdev = list_next(l2arc_dev_list, dev);
4858 if (vd == dev->l2ad_vdev) {
4863 ASSERT(remdev != NULL);
4866 * Remove device from global list
4868 list_remove(l2arc_dev_list, remdev);
4869 l2arc_dev_last = NULL; /* may have been invalidated */
4870 atomic_dec_64(&l2arc_ndev);
4871 mutex_exit(&l2arc_dev_mtx);
4874 * Clear all buflists and ARC references. L2ARC device flush.
4876 l2arc_evict(remdev, 0, B_TRUE);
4877 list_destroy(remdev->l2ad_buflist);
4878 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
4879 kmem_free(remdev, sizeof (l2arc_dev_t));
4885 l2arc_thread_exit = 0;
4887 l2arc_writes_sent = 0;
4888 l2arc_writes_done = 0;
4890 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4891 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
4892 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
4893 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
4894 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
4896 l2arc_dev_list = &L2ARC_dev_list;
4897 l2arc_free_on_write = &L2ARC_free_on_write;
4898 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
4899 offsetof(l2arc_dev_t, l2ad_node));
4900 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
4901 offsetof(l2arc_data_free_t, l2df_list_node));
4908 * This is called from dmu_fini(), which is called from spa_fini();
4909 * Because of this, we can assume that all l2arc devices have
4910 * already been removed when the pools themselves were removed.
4913 l2arc_do_free_on_write();
4915 mutex_destroy(&l2arc_feed_thr_lock);
4916 cv_destroy(&l2arc_feed_thr_cv);
4917 mutex_destroy(&l2arc_dev_mtx);
4918 mutex_destroy(&l2arc_buflist_mtx);
4919 mutex_destroy(&l2arc_free_on_write_mtx);
4921 list_destroy(l2arc_dev_list);
4922 list_destroy(l2arc_free_on_write);
4928 if (!(spa_mode_global & FWRITE))
4931 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
4932 TS_RUN, minclsyspri);
4938 if (!(spa_mode_global & FWRITE))
4941 mutex_enter(&l2arc_feed_thr_lock);
4942 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
4943 l2arc_thread_exit = 1;
4944 while (l2arc_thread_exit != 0)
4945 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
4946 mutex_exit(&l2arc_feed_thr_lock);
4949 #if defined(_KERNEL) && defined(HAVE_SPL)
4950 EXPORT_SYMBOL(arc_read);
4951 EXPORT_SYMBOL(arc_buf_remove_ref);
4952 EXPORT_SYMBOL(arc_getbuf_func);
4953 EXPORT_SYMBOL(arc_add_prune_callback);
4954 EXPORT_SYMBOL(arc_remove_prune_callback);
4956 module_param(zfs_arc_min, ulong, 0444);
4957 MODULE_PARM_DESC(zfs_arc_min, "Min arc size");
4959 module_param(zfs_arc_max, ulong, 0444);
4960 MODULE_PARM_DESC(zfs_arc_max, "Max arc size");
4962 module_param(zfs_arc_meta_limit, ulong, 0444);
4963 MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size");
4965 module_param(zfs_arc_meta_prune, int, 0444);
4966 MODULE_PARM_DESC(zfs_arc_meta_prune, "Bytes of meta data to prune");
4968 module_param(zfs_arc_grow_retry, int, 0444);
4969 MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size");
4971 module_param(zfs_arc_shrink_shift, int, 0444);
4972 MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)");
4974 module_param(zfs_arc_p_min_shift, int, 0444);
4975 MODULE_PARM_DESC(zfs_arc_p_min_shift, "arc_c shift to calc min/max arc_p");
4977 module_param(l2arc_write_max, ulong, 0444);
4978 MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval");
4980 module_param(l2arc_write_boost, ulong, 0444);
4981 MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup");
4983 module_param(l2arc_headroom, ulong, 0444);
4984 MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache");
4986 module_param(l2arc_feed_secs, ulong, 0444);
4987 MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing");
4989 module_param(l2arc_feed_min_ms, ulong, 0444);
4990 MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds");
4992 module_param(l2arc_noprefetch, int, 0444);
4993 MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers");
4995 module_param(l2arc_feed_again, int, 0444);
4996 MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup");
4998 module_param(l2arc_norw, int, 0444);
4999 MODULE_PARM_DESC(l2arc_norw, "No reads during writes");