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 mutexes, rather they rely on the
84 * hash table mutexes for the bulk of their protection (i.e. most
85 * fields in the arc_buf_hdr_t are protected by these mutexes).
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_memory_throttle_disable = 1;
193 int zfs_disable_dup_eviction = 0;
194 int zfs_arc_meta_prune = 0;
197 * Note that buffers can be in one of 6 states:
198 * ARC_anon - anonymous (discussed below)
199 * ARC_mru - recently used, currently cached
200 * ARC_mru_ghost - recentely used, no longer in cache
201 * ARC_mfu - frequently used, currently cached
202 * ARC_mfu_ghost - frequently used, no longer in cache
203 * ARC_l2c_only - exists in L2ARC but not other states
204 * When there are no active references to the buffer, they are
205 * are linked onto a list in one of these arc states. These are
206 * the only buffers that can be evicted or deleted. Within each
207 * state there are multiple lists, one for meta-data and one for
208 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
209 * etc.) is tracked separately so that it can be managed more
210 * explicitly: favored over data, limited explicitly.
212 * Anonymous buffers are buffers that are not associated with
213 * a DVA. These are buffers that hold dirty block copies
214 * before they are written to stable storage. By definition,
215 * they are "ref'd" and are considered part of arc_mru
216 * that cannot be freed. Generally, they will aquire a DVA
217 * as they are written and migrate onto the arc_mru list.
219 * The ARC_l2c_only state is for buffers that are in the second
220 * level ARC but no longer in any of the ARC_m* lists. The second
221 * level ARC itself may also contain buffers that are in any of
222 * the ARC_m* states - meaning that a buffer can exist in two
223 * places. The reason for the ARC_l2c_only state is to keep the
224 * buffer header in the hash table, so that reads that hit the
225 * second level ARC benefit from these fast lookups.
228 typedef struct arc_state {
229 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
230 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
231 uint64_t arcs_size; /* total amount of data in this state */
236 static arc_state_t ARC_anon;
237 static arc_state_t ARC_mru;
238 static arc_state_t ARC_mru_ghost;
239 static arc_state_t ARC_mfu;
240 static arc_state_t ARC_mfu_ghost;
241 static arc_state_t ARC_l2c_only;
243 typedef struct arc_stats {
244 kstat_named_t arcstat_hits;
245 kstat_named_t arcstat_misses;
246 kstat_named_t arcstat_demand_data_hits;
247 kstat_named_t arcstat_demand_data_misses;
248 kstat_named_t arcstat_demand_metadata_hits;
249 kstat_named_t arcstat_demand_metadata_misses;
250 kstat_named_t arcstat_prefetch_data_hits;
251 kstat_named_t arcstat_prefetch_data_misses;
252 kstat_named_t arcstat_prefetch_metadata_hits;
253 kstat_named_t arcstat_prefetch_metadata_misses;
254 kstat_named_t arcstat_mru_hits;
255 kstat_named_t arcstat_mru_ghost_hits;
256 kstat_named_t arcstat_mfu_hits;
257 kstat_named_t arcstat_mfu_ghost_hits;
258 kstat_named_t arcstat_deleted;
259 kstat_named_t arcstat_recycle_miss;
260 kstat_named_t arcstat_mutex_miss;
261 kstat_named_t arcstat_evict_skip;
262 kstat_named_t arcstat_evict_l2_cached;
263 kstat_named_t arcstat_evict_l2_eligible;
264 kstat_named_t arcstat_evict_l2_ineligible;
265 kstat_named_t arcstat_hash_elements;
266 kstat_named_t arcstat_hash_elements_max;
267 kstat_named_t arcstat_hash_collisions;
268 kstat_named_t arcstat_hash_chains;
269 kstat_named_t arcstat_hash_chain_max;
270 kstat_named_t arcstat_p;
271 kstat_named_t arcstat_c;
272 kstat_named_t arcstat_c_min;
273 kstat_named_t arcstat_c_max;
274 kstat_named_t arcstat_size;
275 kstat_named_t arcstat_hdr_size;
276 kstat_named_t arcstat_data_size;
277 kstat_named_t arcstat_other_size;
278 kstat_named_t arcstat_anon_size;
279 kstat_named_t arcstat_anon_evict_data;
280 kstat_named_t arcstat_anon_evict_metadata;
281 kstat_named_t arcstat_mru_size;
282 kstat_named_t arcstat_mru_evict_data;
283 kstat_named_t arcstat_mru_evict_metadata;
284 kstat_named_t arcstat_mru_ghost_size;
285 kstat_named_t arcstat_mru_ghost_evict_data;
286 kstat_named_t arcstat_mru_ghost_evict_metadata;
287 kstat_named_t arcstat_mfu_size;
288 kstat_named_t arcstat_mfu_evict_data;
289 kstat_named_t arcstat_mfu_evict_metadata;
290 kstat_named_t arcstat_mfu_ghost_size;
291 kstat_named_t arcstat_mfu_ghost_evict_data;
292 kstat_named_t arcstat_mfu_ghost_evict_metadata;
293 kstat_named_t arcstat_l2_hits;
294 kstat_named_t arcstat_l2_misses;
295 kstat_named_t arcstat_l2_feeds;
296 kstat_named_t arcstat_l2_rw_clash;
297 kstat_named_t arcstat_l2_read_bytes;
298 kstat_named_t arcstat_l2_write_bytes;
299 kstat_named_t arcstat_l2_writes_sent;
300 kstat_named_t arcstat_l2_writes_done;
301 kstat_named_t arcstat_l2_writes_error;
302 kstat_named_t arcstat_l2_writes_hdr_miss;
303 kstat_named_t arcstat_l2_evict_lock_retry;
304 kstat_named_t arcstat_l2_evict_reading;
305 kstat_named_t arcstat_l2_free_on_write;
306 kstat_named_t arcstat_l2_abort_lowmem;
307 kstat_named_t arcstat_l2_cksum_bad;
308 kstat_named_t arcstat_l2_io_error;
309 kstat_named_t arcstat_l2_size;
310 kstat_named_t arcstat_l2_hdr_size;
311 kstat_named_t arcstat_memory_throttle_count;
312 kstat_named_t arcstat_duplicate_buffers;
313 kstat_named_t arcstat_duplicate_buffers_size;
314 kstat_named_t arcstat_duplicate_reads;
315 kstat_named_t arcstat_memory_direct_count;
316 kstat_named_t arcstat_memory_indirect_count;
317 kstat_named_t arcstat_no_grow;
318 kstat_named_t arcstat_tempreserve;
319 kstat_named_t arcstat_loaned_bytes;
320 kstat_named_t arcstat_prune;
321 kstat_named_t arcstat_meta_used;
322 kstat_named_t arcstat_meta_limit;
323 kstat_named_t arcstat_meta_max;
326 static arc_stats_t arc_stats = {
327 { "hits", KSTAT_DATA_UINT64 },
328 { "misses", KSTAT_DATA_UINT64 },
329 { "demand_data_hits", KSTAT_DATA_UINT64 },
330 { "demand_data_misses", KSTAT_DATA_UINT64 },
331 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
332 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
333 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
334 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
335 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
336 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
337 { "mru_hits", KSTAT_DATA_UINT64 },
338 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
339 { "mfu_hits", KSTAT_DATA_UINT64 },
340 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
341 { "deleted", KSTAT_DATA_UINT64 },
342 { "recycle_miss", KSTAT_DATA_UINT64 },
343 { "mutex_miss", KSTAT_DATA_UINT64 },
344 { "evict_skip", KSTAT_DATA_UINT64 },
345 { "evict_l2_cached", KSTAT_DATA_UINT64 },
346 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
347 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
348 { "hash_elements", KSTAT_DATA_UINT64 },
349 { "hash_elements_max", KSTAT_DATA_UINT64 },
350 { "hash_collisions", KSTAT_DATA_UINT64 },
351 { "hash_chains", KSTAT_DATA_UINT64 },
352 { "hash_chain_max", KSTAT_DATA_UINT64 },
353 { "p", KSTAT_DATA_UINT64 },
354 { "c", KSTAT_DATA_UINT64 },
355 { "c_min", KSTAT_DATA_UINT64 },
356 { "c_max", KSTAT_DATA_UINT64 },
357 { "size", KSTAT_DATA_UINT64 },
358 { "hdr_size", KSTAT_DATA_UINT64 },
359 { "data_size", KSTAT_DATA_UINT64 },
360 { "other_size", KSTAT_DATA_UINT64 },
361 { "anon_size", KSTAT_DATA_UINT64 },
362 { "anon_evict_data", KSTAT_DATA_UINT64 },
363 { "anon_evict_metadata", KSTAT_DATA_UINT64 },
364 { "mru_size", KSTAT_DATA_UINT64 },
365 { "mru_evict_data", KSTAT_DATA_UINT64 },
366 { "mru_evict_metadata", KSTAT_DATA_UINT64 },
367 { "mru_ghost_size", KSTAT_DATA_UINT64 },
368 { "mru_ghost_evict_data", KSTAT_DATA_UINT64 },
369 { "mru_ghost_evict_metadata", KSTAT_DATA_UINT64 },
370 { "mfu_size", KSTAT_DATA_UINT64 },
371 { "mfu_evict_data", KSTAT_DATA_UINT64 },
372 { "mfu_evict_metadata", KSTAT_DATA_UINT64 },
373 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
374 { "mfu_ghost_evict_data", KSTAT_DATA_UINT64 },
375 { "mfu_ghost_evict_metadata", KSTAT_DATA_UINT64 },
376 { "l2_hits", KSTAT_DATA_UINT64 },
377 { "l2_misses", KSTAT_DATA_UINT64 },
378 { "l2_feeds", KSTAT_DATA_UINT64 },
379 { "l2_rw_clash", KSTAT_DATA_UINT64 },
380 { "l2_read_bytes", KSTAT_DATA_UINT64 },
381 { "l2_write_bytes", KSTAT_DATA_UINT64 },
382 { "l2_writes_sent", KSTAT_DATA_UINT64 },
383 { "l2_writes_done", KSTAT_DATA_UINT64 },
384 { "l2_writes_error", KSTAT_DATA_UINT64 },
385 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
386 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
387 { "l2_evict_reading", KSTAT_DATA_UINT64 },
388 { "l2_free_on_write", KSTAT_DATA_UINT64 },
389 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
390 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
391 { "l2_io_error", KSTAT_DATA_UINT64 },
392 { "l2_size", KSTAT_DATA_UINT64 },
393 { "l2_hdr_size", KSTAT_DATA_UINT64 },
394 { "memory_throttle_count", KSTAT_DATA_UINT64 },
395 { "duplicate_buffers", KSTAT_DATA_UINT64 },
396 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
397 { "duplicate_reads", KSTAT_DATA_UINT64 },
398 { "memory_direct_count", KSTAT_DATA_UINT64 },
399 { "memory_indirect_count", KSTAT_DATA_UINT64 },
400 { "arc_no_grow", KSTAT_DATA_UINT64 },
401 { "arc_tempreserve", KSTAT_DATA_UINT64 },
402 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
403 { "arc_prune", KSTAT_DATA_UINT64 },
404 { "arc_meta_used", KSTAT_DATA_UINT64 },
405 { "arc_meta_limit", KSTAT_DATA_UINT64 },
406 { "arc_meta_max", KSTAT_DATA_UINT64 },
409 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
411 #define ARCSTAT_INCR(stat, val) \
412 atomic_add_64(&arc_stats.stat.value.ui64, (val));
414 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
415 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
417 #define ARCSTAT_MAX(stat, val) { \
419 while ((val) > (m = arc_stats.stat.value.ui64) && \
420 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
424 #define ARCSTAT_MAXSTAT(stat) \
425 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
428 * We define a macro to allow ARC hits/misses to be easily broken down by
429 * two separate conditions, giving a total of four different subtypes for
430 * each of hits and misses (so eight statistics total).
432 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
435 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
437 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
441 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
443 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
448 static arc_state_t *arc_anon;
449 static arc_state_t *arc_mru;
450 static arc_state_t *arc_mru_ghost;
451 static arc_state_t *arc_mfu;
452 static arc_state_t *arc_mfu_ghost;
453 static arc_state_t *arc_l2c_only;
456 * There are several ARC variables that are critical to export as kstats --
457 * but we don't want to have to grovel around in the kstat whenever we wish to
458 * manipulate them. For these variables, we therefore define them to be in
459 * terms of the statistic variable. This assures that we are not introducing
460 * the possibility of inconsistency by having shadow copies of the variables,
461 * while still allowing the code to be readable.
463 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
464 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
465 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
466 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
467 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
468 #define arc_no_grow ARCSTAT(arcstat_no_grow)
469 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
470 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
471 #define arc_meta_used ARCSTAT(arcstat_meta_used)
472 #define arc_meta_limit ARCSTAT(arcstat_meta_limit)
473 #define arc_meta_max ARCSTAT(arcstat_meta_max)
475 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
477 typedef struct arc_callback arc_callback_t;
479 struct arc_callback {
481 arc_done_func_t *acb_done;
483 zio_t *acb_zio_dummy;
484 arc_callback_t *acb_next;
487 typedef struct arc_write_callback arc_write_callback_t;
489 struct arc_write_callback {
491 arc_done_func_t *awcb_ready;
492 arc_done_func_t *awcb_done;
497 /* protected by hash lock */
502 kmutex_t b_freeze_lock;
503 zio_cksum_t *b_freeze_cksum;
506 arc_buf_hdr_t *b_hash_next;
511 arc_callback_t *b_acb;
515 arc_buf_contents_t b_type;
519 /* protected by arc state mutex */
520 arc_state_t *b_state;
521 list_node_t b_arc_node;
523 /* updated atomically */
524 clock_t b_arc_access;
526 /* self protecting */
529 l2arc_buf_hdr_t *b_l2hdr;
530 list_node_t b_l2node;
533 static list_t arc_prune_list;
534 static kmutex_t arc_prune_mtx;
535 static arc_buf_t *arc_eviction_list;
536 static kmutex_t arc_eviction_mtx;
537 static arc_buf_hdr_t arc_eviction_hdr;
538 static void arc_get_data_buf(arc_buf_t *buf);
539 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
540 static int arc_evict_needed(arc_buf_contents_t type);
541 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
543 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
545 #define GHOST_STATE(state) \
546 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
547 (state) == arc_l2c_only)
550 * Private ARC flags. These flags are private ARC only flags that will show up
551 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
552 * be passed in as arc_flags in things like arc_read. However, these flags
553 * should never be passed and should only be set by ARC code. When adding new
554 * public flags, make sure not to smash the private ones.
557 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
558 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
559 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
560 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
561 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
562 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
563 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
564 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
565 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
566 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
568 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
569 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
570 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
571 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
572 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
573 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
574 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
575 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
576 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
577 (hdr)->b_l2hdr != NULL)
578 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
579 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
580 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
586 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
587 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
590 * Hash table routines
593 #define HT_LOCK_ALIGN 64
594 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
599 unsigned char pad[HT_LOCK_PAD];
603 #define BUF_LOCKS 256
604 typedef struct buf_hash_table {
606 arc_buf_hdr_t **ht_table;
607 struct ht_lock ht_locks[BUF_LOCKS];
610 static buf_hash_table_t buf_hash_table;
612 #define BUF_HASH_INDEX(spa, dva, birth) \
613 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
614 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
615 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
616 #define HDR_LOCK(hdr) \
617 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
619 uint64_t zfs_crc64_table[256];
625 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
626 #define L2ARC_HEADROOM 2 /* num of writes */
627 #define L2ARC_FEED_SECS 1 /* caching interval secs */
628 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
630 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
631 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
634 * L2ARC Performance Tunables
636 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
637 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
638 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
639 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
640 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
641 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
642 int l2arc_feed_again = B_TRUE; /* turbo warmup */
643 int l2arc_norw = B_TRUE; /* no reads during writes */
648 typedef struct l2arc_dev {
649 vdev_t *l2ad_vdev; /* vdev */
650 spa_t *l2ad_spa; /* spa */
651 uint64_t l2ad_hand; /* next write location */
652 uint64_t l2ad_write; /* desired write size, bytes */
653 uint64_t l2ad_boost; /* warmup write boost, bytes */
654 uint64_t l2ad_start; /* first addr on device */
655 uint64_t l2ad_end; /* last addr on device */
656 uint64_t l2ad_evict; /* last addr eviction reached */
657 boolean_t l2ad_first; /* first sweep through */
658 boolean_t l2ad_writing; /* currently writing */
659 list_t *l2ad_buflist; /* buffer list */
660 list_node_t l2ad_node; /* device list node */
663 static list_t L2ARC_dev_list; /* device list */
664 static list_t *l2arc_dev_list; /* device list pointer */
665 static kmutex_t l2arc_dev_mtx; /* device list mutex */
666 static l2arc_dev_t *l2arc_dev_last; /* last device used */
667 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
668 static list_t L2ARC_free_on_write; /* free after write buf list */
669 static list_t *l2arc_free_on_write; /* free after write list ptr */
670 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
671 static uint64_t l2arc_ndev; /* number of devices */
673 typedef struct l2arc_read_callback {
674 arc_buf_t *l2rcb_buf; /* read buffer */
675 spa_t *l2rcb_spa; /* spa */
676 blkptr_t l2rcb_bp; /* original blkptr */
677 zbookmark_t l2rcb_zb; /* original bookmark */
678 int l2rcb_flags; /* original flags */
679 } l2arc_read_callback_t;
681 typedef struct l2arc_write_callback {
682 l2arc_dev_t *l2wcb_dev; /* device info */
683 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
684 } l2arc_write_callback_t;
686 struct l2arc_buf_hdr {
687 /* protected by arc_buf_hdr mutex */
688 l2arc_dev_t *b_dev; /* L2ARC device */
689 uint64_t b_daddr; /* disk address, offset byte */
692 typedef struct l2arc_data_free {
693 /* protected by l2arc_free_on_write_mtx */
696 void (*l2df_func)(void *, size_t);
697 list_node_t l2df_list_node;
700 static kmutex_t l2arc_feed_thr_lock;
701 static kcondvar_t l2arc_feed_thr_cv;
702 static uint8_t l2arc_thread_exit;
704 static void l2arc_read_done(zio_t *zio);
705 static void l2arc_hdr_stat_add(void);
706 static void l2arc_hdr_stat_remove(void);
709 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
711 uint8_t *vdva = (uint8_t *)dva;
712 uint64_t crc = -1ULL;
715 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
717 for (i = 0; i < sizeof (dva_t); i++)
718 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
720 crc ^= (spa>>8) ^ birth;
725 #define BUF_EMPTY(buf) \
726 ((buf)->b_dva.dva_word[0] == 0 && \
727 (buf)->b_dva.dva_word[1] == 0 && \
730 #define BUF_EQUAL(spa, dva, birth, buf) \
731 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
732 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
733 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
736 buf_discard_identity(arc_buf_hdr_t *hdr)
738 hdr->b_dva.dva_word[0] = 0;
739 hdr->b_dva.dva_word[1] = 0;
744 static arc_buf_hdr_t *
745 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
747 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
748 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
751 mutex_enter(hash_lock);
752 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
753 buf = buf->b_hash_next) {
754 if (BUF_EQUAL(spa, dva, birth, buf)) {
759 mutex_exit(hash_lock);
765 * Insert an entry into the hash table. If there is already an element
766 * equal to elem in the hash table, then the already existing element
767 * will be returned and the new element will not be inserted.
768 * Otherwise returns NULL.
770 static arc_buf_hdr_t *
771 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
773 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
774 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
778 ASSERT(!HDR_IN_HASH_TABLE(buf));
780 mutex_enter(hash_lock);
781 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
782 fbuf = fbuf->b_hash_next, i++) {
783 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
787 buf->b_hash_next = buf_hash_table.ht_table[idx];
788 buf_hash_table.ht_table[idx] = buf;
789 buf->b_flags |= ARC_IN_HASH_TABLE;
791 /* collect some hash table performance data */
793 ARCSTAT_BUMP(arcstat_hash_collisions);
795 ARCSTAT_BUMP(arcstat_hash_chains);
797 ARCSTAT_MAX(arcstat_hash_chain_max, i);
800 ARCSTAT_BUMP(arcstat_hash_elements);
801 ARCSTAT_MAXSTAT(arcstat_hash_elements);
807 buf_hash_remove(arc_buf_hdr_t *buf)
809 arc_buf_hdr_t *fbuf, **bufp;
810 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
812 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
813 ASSERT(HDR_IN_HASH_TABLE(buf));
815 bufp = &buf_hash_table.ht_table[idx];
816 while ((fbuf = *bufp) != buf) {
817 ASSERT(fbuf != NULL);
818 bufp = &fbuf->b_hash_next;
820 *bufp = buf->b_hash_next;
821 buf->b_hash_next = NULL;
822 buf->b_flags &= ~ARC_IN_HASH_TABLE;
824 /* collect some hash table performance data */
825 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
827 if (buf_hash_table.ht_table[idx] &&
828 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
829 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
833 * Global data structures and functions for the buf kmem cache.
835 static kmem_cache_t *hdr_cache;
836 static kmem_cache_t *buf_cache;
843 #if defined(_KERNEL) && defined(HAVE_SPL)
844 /* Large allocations which do not require contiguous pages
845 * should be using vmem_free() in the linux kernel */
846 vmem_free(buf_hash_table.ht_table,
847 (buf_hash_table.ht_mask + 1) * sizeof (void *));
849 kmem_free(buf_hash_table.ht_table,
850 (buf_hash_table.ht_mask + 1) * sizeof (void *));
852 for (i = 0; i < BUF_LOCKS; i++)
853 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
854 kmem_cache_destroy(hdr_cache);
855 kmem_cache_destroy(buf_cache);
859 * Constructor callback - called when the cache is empty
860 * and a new buf is requested.
864 hdr_cons(void *vbuf, void *unused, int kmflag)
866 arc_buf_hdr_t *buf = vbuf;
868 bzero(buf, sizeof (arc_buf_hdr_t));
869 refcount_create(&buf->b_refcnt);
870 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
871 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
872 list_link_init(&buf->b_arc_node);
873 list_link_init(&buf->b_l2node);
874 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
881 buf_cons(void *vbuf, void *unused, int kmflag)
883 arc_buf_t *buf = vbuf;
885 bzero(buf, sizeof (arc_buf_t));
886 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
887 rw_init(&buf->b_data_lock, NULL, RW_DEFAULT, NULL);
888 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
894 * Destructor callback - called when a cached buf is
895 * no longer required.
899 hdr_dest(void *vbuf, void *unused)
901 arc_buf_hdr_t *buf = vbuf;
903 ASSERT(BUF_EMPTY(buf));
904 refcount_destroy(&buf->b_refcnt);
905 cv_destroy(&buf->b_cv);
906 mutex_destroy(&buf->b_freeze_lock);
907 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
912 buf_dest(void *vbuf, void *unused)
914 arc_buf_t *buf = vbuf;
916 mutex_destroy(&buf->b_evict_lock);
917 rw_destroy(&buf->b_data_lock);
918 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
925 uint64_t hsize = 1ULL << 12;
929 * The hash table is big enough to fill all of physical memory
930 * with an average 64K block size. The table will take up
931 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
933 while (hsize * 65536 < physmem * PAGESIZE)
936 buf_hash_table.ht_mask = hsize - 1;
937 #if defined(_KERNEL) && defined(HAVE_SPL)
938 /* Large allocations which do not require contiguous pages
939 * should be using vmem_alloc() in the linux kernel */
940 buf_hash_table.ht_table =
941 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
943 buf_hash_table.ht_table =
944 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
946 if (buf_hash_table.ht_table == NULL) {
947 ASSERT(hsize > (1ULL << 8));
952 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
953 0, hdr_cons, hdr_dest, NULL, NULL, NULL, 0);
954 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
955 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
957 for (i = 0; i < 256; i++)
958 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
959 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
961 for (i = 0; i < BUF_LOCKS; i++) {
962 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
963 NULL, MUTEX_DEFAULT, NULL);
967 #define ARC_MINTIME (hz>>4) /* 62 ms */
970 arc_cksum_verify(arc_buf_t *buf)
974 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
977 mutex_enter(&buf->b_hdr->b_freeze_lock);
978 if (buf->b_hdr->b_freeze_cksum == NULL ||
979 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
980 mutex_exit(&buf->b_hdr->b_freeze_lock);
983 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
984 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
985 panic("buffer modified while frozen!");
986 mutex_exit(&buf->b_hdr->b_freeze_lock);
990 arc_cksum_equal(arc_buf_t *buf)
995 mutex_enter(&buf->b_hdr->b_freeze_lock);
996 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
997 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
998 mutex_exit(&buf->b_hdr->b_freeze_lock);
1004 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1006 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1009 mutex_enter(&buf->b_hdr->b_freeze_lock);
1010 if (buf->b_hdr->b_freeze_cksum != NULL) {
1011 mutex_exit(&buf->b_hdr->b_freeze_lock);
1014 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1016 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1017 buf->b_hdr->b_freeze_cksum);
1018 mutex_exit(&buf->b_hdr->b_freeze_lock);
1022 arc_buf_thaw(arc_buf_t *buf)
1024 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1025 if (buf->b_hdr->b_state != arc_anon)
1026 panic("modifying non-anon buffer!");
1027 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1028 panic("modifying buffer while i/o in progress!");
1029 arc_cksum_verify(buf);
1032 mutex_enter(&buf->b_hdr->b_freeze_lock);
1033 if (buf->b_hdr->b_freeze_cksum != NULL) {
1034 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1035 buf->b_hdr->b_freeze_cksum = NULL;
1038 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1039 if (buf->b_hdr->b_thawed)
1040 kmem_free(buf->b_hdr->b_thawed, 1);
1041 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1044 mutex_exit(&buf->b_hdr->b_freeze_lock);
1048 arc_buf_freeze(arc_buf_t *buf)
1050 kmutex_t *hash_lock;
1052 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1055 hash_lock = HDR_LOCK(buf->b_hdr);
1056 mutex_enter(hash_lock);
1058 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1059 buf->b_hdr->b_state == arc_anon);
1060 arc_cksum_compute(buf, B_FALSE);
1061 mutex_exit(hash_lock);
1065 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1067 ASSERT(MUTEX_HELD(hash_lock));
1069 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1070 (ab->b_state != arc_anon)) {
1071 uint64_t delta = ab->b_size * ab->b_datacnt;
1072 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1073 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1075 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1076 mutex_enter(&ab->b_state->arcs_mtx);
1077 ASSERT(list_link_active(&ab->b_arc_node));
1078 list_remove(list, ab);
1079 if (GHOST_STATE(ab->b_state)) {
1080 ASSERT0(ab->b_datacnt);
1081 ASSERT3P(ab->b_buf, ==, NULL);
1085 ASSERT3U(*size, >=, delta);
1086 atomic_add_64(size, -delta);
1087 mutex_exit(&ab->b_state->arcs_mtx);
1088 /* remove the prefetch flag if we get a reference */
1089 if (ab->b_flags & ARC_PREFETCH)
1090 ab->b_flags &= ~ARC_PREFETCH;
1095 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1098 arc_state_t *state = ab->b_state;
1100 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1101 ASSERT(!GHOST_STATE(state));
1103 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1104 (state != arc_anon)) {
1105 uint64_t *size = &state->arcs_lsize[ab->b_type];
1107 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1108 mutex_enter(&state->arcs_mtx);
1109 ASSERT(!list_link_active(&ab->b_arc_node));
1110 list_insert_head(&state->arcs_list[ab->b_type], ab);
1111 ASSERT(ab->b_datacnt > 0);
1112 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1113 mutex_exit(&state->arcs_mtx);
1119 * Move the supplied buffer to the indicated state. The mutex
1120 * for the buffer must be held by the caller.
1123 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1125 arc_state_t *old_state = ab->b_state;
1126 int64_t refcnt = refcount_count(&ab->b_refcnt);
1127 uint64_t from_delta, to_delta;
1129 ASSERT(MUTEX_HELD(hash_lock));
1130 ASSERT(new_state != old_state);
1131 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1132 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1133 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1135 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1138 * If this buffer is evictable, transfer it from the
1139 * old state list to the new state list.
1142 if (old_state != arc_anon) {
1143 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1144 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1147 mutex_enter(&old_state->arcs_mtx);
1149 ASSERT(list_link_active(&ab->b_arc_node));
1150 list_remove(&old_state->arcs_list[ab->b_type], ab);
1153 * If prefetching out of the ghost cache,
1154 * we will have a non-zero datacnt.
1156 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1157 /* ghost elements have a ghost size */
1158 ASSERT(ab->b_buf == NULL);
1159 from_delta = ab->b_size;
1161 ASSERT3U(*size, >=, from_delta);
1162 atomic_add_64(size, -from_delta);
1165 mutex_exit(&old_state->arcs_mtx);
1167 if (new_state != arc_anon) {
1168 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1169 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1172 mutex_enter(&new_state->arcs_mtx);
1174 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1176 /* ghost elements have a ghost size */
1177 if (GHOST_STATE(new_state)) {
1178 ASSERT(ab->b_datacnt == 0);
1179 ASSERT(ab->b_buf == NULL);
1180 to_delta = ab->b_size;
1182 atomic_add_64(size, to_delta);
1185 mutex_exit(&new_state->arcs_mtx);
1189 ASSERT(!BUF_EMPTY(ab));
1190 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1191 buf_hash_remove(ab);
1193 /* adjust state sizes */
1195 atomic_add_64(&new_state->arcs_size, to_delta);
1197 ASSERT3U(old_state->arcs_size, >=, from_delta);
1198 atomic_add_64(&old_state->arcs_size, -from_delta);
1200 ab->b_state = new_state;
1202 /* adjust l2arc hdr stats */
1203 if (new_state == arc_l2c_only)
1204 l2arc_hdr_stat_add();
1205 else if (old_state == arc_l2c_only)
1206 l2arc_hdr_stat_remove();
1210 arc_space_consume(uint64_t space, arc_space_type_t type)
1212 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1217 case ARC_SPACE_DATA:
1218 ARCSTAT_INCR(arcstat_data_size, space);
1220 case ARC_SPACE_OTHER:
1221 ARCSTAT_INCR(arcstat_other_size, space);
1223 case ARC_SPACE_HDRS:
1224 ARCSTAT_INCR(arcstat_hdr_size, space);
1226 case ARC_SPACE_L2HDRS:
1227 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1231 atomic_add_64(&arc_meta_used, space);
1232 atomic_add_64(&arc_size, space);
1236 arc_space_return(uint64_t space, arc_space_type_t type)
1238 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1243 case ARC_SPACE_DATA:
1244 ARCSTAT_INCR(arcstat_data_size, -space);
1246 case ARC_SPACE_OTHER:
1247 ARCSTAT_INCR(arcstat_other_size, -space);
1249 case ARC_SPACE_HDRS:
1250 ARCSTAT_INCR(arcstat_hdr_size, -space);
1252 case ARC_SPACE_L2HDRS:
1253 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1257 ASSERT(arc_meta_used >= space);
1258 if (arc_meta_max < arc_meta_used)
1259 arc_meta_max = arc_meta_used;
1260 atomic_add_64(&arc_meta_used, -space);
1261 ASSERT(arc_size >= space);
1262 atomic_add_64(&arc_size, -space);
1266 arc_data_buf_alloc(uint64_t size)
1268 if (arc_evict_needed(ARC_BUFC_DATA))
1269 cv_signal(&arc_reclaim_thr_cv);
1270 atomic_add_64(&arc_size, size);
1271 return (zio_data_buf_alloc(size));
1275 arc_data_buf_free(void *buf, uint64_t size)
1277 zio_data_buf_free(buf, size);
1278 ASSERT(arc_size >= size);
1279 atomic_add_64(&arc_size, -size);
1283 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1288 ASSERT3U(size, >, 0);
1289 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1290 ASSERT(BUF_EMPTY(hdr));
1293 hdr->b_spa = spa_load_guid(spa);
1294 hdr->b_state = arc_anon;
1295 hdr->b_arc_access = 0;
1296 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1299 buf->b_efunc = NULL;
1300 buf->b_private = NULL;
1303 arc_get_data_buf(buf);
1306 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1307 (void) refcount_add(&hdr->b_refcnt, tag);
1312 static char *arc_onloan_tag = "onloan";
1315 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1316 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1317 * buffers must be returned to the arc before they can be used by the DMU or
1321 arc_loan_buf(spa_t *spa, int size)
1325 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1327 atomic_add_64(&arc_loaned_bytes, size);
1332 * Return a loaned arc buffer to the arc.
1335 arc_return_buf(arc_buf_t *buf, void *tag)
1337 arc_buf_hdr_t *hdr = buf->b_hdr;
1339 ASSERT(buf->b_data != NULL);
1340 (void) refcount_add(&hdr->b_refcnt, tag);
1341 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1343 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1346 /* Detach an arc_buf from a dbuf (tag) */
1348 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1352 ASSERT(buf->b_data != NULL);
1354 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1355 (void) refcount_remove(&hdr->b_refcnt, tag);
1356 buf->b_efunc = NULL;
1357 buf->b_private = NULL;
1359 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1363 arc_buf_clone(arc_buf_t *from)
1366 arc_buf_hdr_t *hdr = from->b_hdr;
1367 uint64_t size = hdr->b_size;
1369 ASSERT(hdr->b_state != arc_anon);
1371 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1374 buf->b_efunc = NULL;
1375 buf->b_private = NULL;
1376 buf->b_next = hdr->b_buf;
1378 arc_get_data_buf(buf);
1379 bcopy(from->b_data, buf->b_data, size);
1382 * This buffer already exists in the arc so create a duplicate
1383 * copy for the caller. If the buffer is associated with user data
1384 * then track the size and number of duplicates. These stats will be
1385 * updated as duplicate buffers are created and destroyed.
1387 if (hdr->b_type == ARC_BUFC_DATA) {
1388 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1389 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1391 hdr->b_datacnt += 1;
1396 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1399 kmutex_t *hash_lock;
1402 * Check to see if this buffer is evicted. Callers
1403 * must verify b_data != NULL to know if the add_ref
1406 mutex_enter(&buf->b_evict_lock);
1407 if (buf->b_data == NULL) {
1408 mutex_exit(&buf->b_evict_lock);
1411 hash_lock = HDR_LOCK(buf->b_hdr);
1412 mutex_enter(hash_lock);
1414 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1415 mutex_exit(&buf->b_evict_lock);
1417 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1418 add_reference(hdr, hash_lock, tag);
1419 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1420 arc_access(hdr, hash_lock);
1421 mutex_exit(hash_lock);
1422 ARCSTAT_BUMP(arcstat_hits);
1423 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1424 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1425 data, metadata, hits);
1429 * Free the arc data buffer. If it is an l2arc write in progress,
1430 * the buffer is placed on l2arc_free_on_write to be freed later.
1433 arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t),
1434 void *data, size_t size)
1436 if (HDR_L2_WRITING(hdr)) {
1437 l2arc_data_free_t *df;
1438 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_PUSHPAGE);
1439 df->l2df_data = data;
1440 df->l2df_size = size;
1441 df->l2df_func = free_func;
1442 mutex_enter(&l2arc_free_on_write_mtx);
1443 list_insert_head(l2arc_free_on_write, df);
1444 mutex_exit(&l2arc_free_on_write_mtx);
1445 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1447 free_func(data, size);
1452 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1456 /* free up data associated with the buf */
1458 arc_state_t *state = buf->b_hdr->b_state;
1459 uint64_t size = buf->b_hdr->b_size;
1460 arc_buf_contents_t type = buf->b_hdr->b_type;
1462 arc_cksum_verify(buf);
1465 if (type == ARC_BUFC_METADATA) {
1466 arc_buf_data_free(buf->b_hdr, zio_buf_free,
1468 arc_space_return(size, ARC_SPACE_DATA);
1470 ASSERT(type == ARC_BUFC_DATA);
1471 arc_buf_data_free(buf->b_hdr,
1472 zio_data_buf_free, buf->b_data, size);
1473 ARCSTAT_INCR(arcstat_data_size, -size);
1474 atomic_add_64(&arc_size, -size);
1477 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1478 uint64_t *cnt = &state->arcs_lsize[type];
1480 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1481 ASSERT(state != arc_anon);
1483 ASSERT3U(*cnt, >=, size);
1484 atomic_add_64(cnt, -size);
1486 ASSERT3U(state->arcs_size, >=, size);
1487 atomic_add_64(&state->arcs_size, -size);
1491 * If we're destroying a duplicate buffer make sure
1492 * that the appropriate statistics are updated.
1494 if (buf->b_hdr->b_datacnt > 1 &&
1495 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1496 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1497 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1499 ASSERT(buf->b_hdr->b_datacnt > 0);
1500 buf->b_hdr->b_datacnt -= 1;
1503 /* only remove the buf if requested */
1507 /* remove the buf from the hdr list */
1508 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1510 *bufp = buf->b_next;
1513 ASSERT(buf->b_efunc == NULL);
1515 /* clean up the buf */
1517 kmem_cache_free(buf_cache, buf);
1521 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1523 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1525 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1526 ASSERT3P(hdr->b_state, ==, arc_anon);
1527 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1529 if (l2hdr != NULL) {
1530 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1532 * To prevent arc_free() and l2arc_evict() from
1533 * attempting to free the same buffer at the same time,
1534 * a FREE_IN_PROGRESS flag is given to arc_free() to
1535 * give it priority. l2arc_evict() can't destroy this
1536 * header while we are waiting on l2arc_buflist_mtx.
1538 * The hdr may be removed from l2ad_buflist before we
1539 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1541 if (!buflist_held) {
1542 mutex_enter(&l2arc_buflist_mtx);
1543 l2hdr = hdr->b_l2hdr;
1546 if (l2hdr != NULL) {
1547 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1548 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1549 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1550 if (hdr->b_state == arc_l2c_only)
1551 l2arc_hdr_stat_remove();
1552 hdr->b_l2hdr = NULL;
1556 mutex_exit(&l2arc_buflist_mtx);
1559 if (!BUF_EMPTY(hdr)) {
1560 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1561 buf_discard_identity(hdr);
1563 while (hdr->b_buf) {
1564 arc_buf_t *buf = hdr->b_buf;
1567 mutex_enter(&arc_eviction_mtx);
1568 mutex_enter(&buf->b_evict_lock);
1569 ASSERT(buf->b_hdr != NULL);
1570 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1571 hdr->b_buf = buf->b_next;
1572 buf->b_hdr = &arc_eviction_hdr;
1573 buf->b_next = arc_eviction_list;
1574 arc_eviction_list = buf;
1575 mutex_exit(&buf->b_evict_lock);
1576 mutex_exit(&arc_eviction_mtx);
1578 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1581 if (hdr->b_freeze_cksum != NULL) {
1582 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1583 hdr->b_freeze_cksum = NULL;
1585 if (hdr->b_thawed) {
1586 kmem_free(hdr->b_thawed, 1);
1587 hdr->b_thawed = NULL;
1590 ASSERT(!list_link_active(&hdr->b_arc_node));
1591 ASSERT3P(hdr->b_hash_next, ==, NULL);
1592 ASSERT3P(hdr->b_acb, ==, NULL);
1593 kmem_cache_free(hdr_cache, hdr);
1597 arc_buf_free(arc_buf_t *buf, void *tag)
1599 arc_buf_hdr_t *hdr = buf->b_hdr;
1600 int hashed = hdr->b_state != arc_anon;
1602 ASSERT(buf->b_efunc == NULL);
1603 ASSERT(buf->b_data != NULL);
1606 kmutex_t *hash_lock = HDR_LOCK(hdr);
1608 mutex_enter(hash_lock);
1610 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1612 (void) remove_reference(hdr, hash_lock, tag);
1613 if (hdr->b_datacnt > 1) {
1614 arc_buf_destroy(buf, FALSE, TRUE);
1616 ASSERT(buf == hdr->b_buf);
1617 ASSERT(buf->b_efunc == NULL);
1618 hdr->b_flags |= ARC_BUF_AVAILABLE;
1620 mutex_exit(hash_lock);
1621 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1624 * We are in the middle of an async write. Don't destroy
1625 * this buffer unless the write completes before we finish
1626 * decrementing the reference count.
1628 mutex_enter(&arc_eviction_mtx);
1629 (void) remove_reference(hdr, NULL, tag);
1630 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1631 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1632 mutex_exit(&arc_eviction_mtx);
1634 arc_hdr_destroy(hdr);
1636 if (remove_reference(hdr, NULL, tag) > 0)
1637 arc_buf_destroy(buf, FALSE, TRUE);
1639 arc_hdr_destroy(hdr);
1644 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1646 arc_buf_hdr_t *hdr = buf->b_hdr;
1647 kmutex_t *hash_lock = HDR_LOCK(hdr);
1648 int no_callback = (buf->b_efunc == NULL);
1650 if (hdr->b_state == arc_anon) {
1651 ASSERT(hdr->b_datacnt == 1);
1652 arc_buf_free(buf, tag);
1653 return (no_callback);
1656 mutex_enter(hash_lock);
1658 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1659 ASSERT(hdr->b_state != arc_anon);
1660 ASSERT(buf->b_data != NULL);
1662 (void) remove_reference(hdr, hash_lock, tag);
1663 if (hdr->b_datacnt > 1) {
1665 arc_buf_destroy(buf, FALSE, TRUE);
1666 } else if (no_callback) {
1667 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1668 ASSERT(buf->b_efunc == NULL);
1669 hdr->b_flags |= ARC_BUF_AVAILABLE;
1671 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1672 refcount_is_zero(&hdr->b_refcnt));
1673 mutex_exit(hash_lock);
1674 return (no_callback);
1678 arc_buf_size(arc_buf_t *buf)
1680 return (buf->b_hdr->b_size);
1684 * Called from the DMU to determine if the current buffer should be
1685 * evicted. In order to ensure proper locking, the eviction must be initiated
1686 * from the DMU. Return true if the buffer is associated with user data and
1687 * duplicate buffers still exist.
1690 arc_buf_eviction_needed(arc_buf_t *buf)
1693 boolean_t evict_needed = B_FALSE;
1695 if (zfs_disable_dup_eviction)
1698 mutex_enter(&buf->b_evict_lock);
1702 * We are in arc_do_user_evicts(); let that function
1703 * perform the eviction.
1705 ASSERT(buf->b_data == NULL);
1706 mutex_exit(&buf->b_evict_lock);
1708 } else if (buf->b_data == NULL) {
1710 * We have already been added to the arc eviction list;
1711 * recommend eviction.
1713 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1714 mutex_exit(&buf->b_evict_lock);
1718 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1719 evict_needed = B_TRUE;
1721 mutex_exit(&buf->b_evict_lock);
1722 return (evict_needed);
1726 * Evict buffers from list until we've removed the specified number of
1727 * bytes. Move the removed buffers to the appropriate evict state.
1728 * If the recycle flag is set, then attempt to "recycle" a buffer:
1729 * - look for a buffer to evict that is `bytes' long.
1730 * - return the data block from this buffer rather than freeing it.
1731 * This flag is used by callers that are trying to make space for a
1732 * new buffer in a full arc cache.
1734 * This function makes a "best effort". It skips over any buffers
1735 * it can't get a hash_lock on, and so may not catch all candidates.
1736 * It may also return without evicting as much space as requested.
1739 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1740 arc_buf_contents_t type)
1742 arc_state_t *evicted_state;
1743 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1744 arc_buf_hdr_t *ab, *ab_prev = NULL;
1745 list_t *list = &state->arcs_list[type];
1746 kmutex_t *hash_lock;
1747 boolean_t have_lock;
1748 void *stolen = NULL;
1750 ASSERT(state == arc_mru || state == arc_mfu);
1752 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1754 mutex_enter(&state->arcs_mtx);
1755 mutex_enter(&evicted_state->arcs_mtx);
1757 for (ab = list_tail(list); ab; ab = ab_prev) {
1758 ab_prev = list_prev(list, ab);
1759 /* prefetch buffers have a minimum lifespan */
1760 if (HDR_IO_IN_PROGRESS(ab) ||
1761 (spa && ab->b_spa != spa) ||
1762 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1763 ddi_get_lbolt() - ab->b_arc_access <
1764 arc_min_prefetch_lifespan)) {
1768 /* "lookahead" for better eviction candidate */
1769 if (recycle && ab->b_size != bytes &&
1770 ab_prev && ab_prev->b_size == bytes)
1772 hash_lock = HDR_LOCK(ab);
1773 have_lock = MUTEX_HELD(hash_lock);
1774 if (have_lock || mutex_tryenter(hash_lock)) {
1775 ASSERT0(refcount_count(&ab->b_refcnt));
1776 ASSERT(ab->b_datacnt > 0);
1778 arc_buf_t *buf = ab->b_buf;
1779 if (!mutex_tryenter(&buf->b_evict_lock)) {
1784 bytes_evicted += ab->b_size;
1785 if (recycle && ab->b_type == type &&
1786 ab->b_size == bytes &&
1787 !HDR_L2_WRITING(ab)) {
1788 stolen = buf->b_data;
1793 mutex_enter(&arc_eviction_mtx);
1794 arc_buf_destroy(buf,
1795 buf->b_data == stolen, FALSE);
1796 ab->b_buf = buf->b_next;
1797 buf->b_hdr = &arc_eviction_hdr;
1798 buf->b_next = arc_eviction_list;
1799 arc_eviction_list = buf;
1800 mutex_exit(&arc_eviction_mtx);
1801 mutex_exit(&buf->b_evict_lock);
1803 mutex_exit(&buf->b_evict_lock);
1804 arc_buf_destroy(buf,
1805 buf->b_data == stolen, TRUE);
1810 ARCSTAT_INCR(arcstat_evict_l2_cached,
1813 if (l2arc_write_eligible(ab->b_spa, ab)) {
1814 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1818 arcstat_evict_l2_ineligible,
1823 if (ab->b_datacnt == 0) {
1824 arc_change_state(evicted_state, ab, hash_lock);
1825 ASSERT(HDR_IN_HASH_TABLE(ab));
1826 ab->b_flags |= ARC_IN_HASH_TABLE;
1827 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1828 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1831 mutex_exit(hash_lock);
1832 if (bytes >= 0 && bytes_evicted >= bytes)
1839 mutex_exit(&evicted_state->arcs_mtx);
1840 mutex_exit(&state->arcs_mtx);
1842 if (bytes_evicted < bytes)
1843 dprintf("only evicted %lld bytes from %x\n",
1844 (longlong_t)bytes_evicted, state);
1847 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1850 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1853 * We have just evicted some date into the ghost state, make
1854 * sure we also adjust the ghost state size if necessary.
1857 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1858 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1859 arc_mru_ghost->arcs_size - arc_c;
1861 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1863 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1864 arc_evict_ghost(arc_mru_ghost, 0, todelete);
1865 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1866 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1867 arc_mru_ghost->arcs_size +
1868 arc_mfu_ghost->arcs_size - arc_c);
1869 arc_evict_ghost(arc_mfu_ghost, 0, todelete);
1877 * Remove buffers from list until we've removed the specified number of
1878 * bytes. Destroy the buffers that are removed.
1881 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1883 arc_buf_hdr_t *ab, *ab_prev;
1884 arc_buf_hdr_t marker;
1885 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1886 kmutex_t *hash_lock;
1887 uint64_t bytes_deleted = 0;
1888 uint64_t bufs_skipped = 0;
1890 ASSERT(GHOST_STATE(state));
1891 bzero(&marker, sizeof(marker));
1893 mutex_enter(&state->arcs_mtx);
1894 for (ab = list_tail(list); ab; ab = ab_prev) {
1895 ab_prev = list_prev(list, ab);
1896 if (spa && ab->b_spa != spa)
1899 /* ignore markers */
1903 hash_lock = HDR_LOCK(ab);
1904 /* caller may be trying to modify this buffer, skip it */
1905 if (MUTEX_HELD(hash_lock))
1907 if (mutex_tryenter(hash_lock)) {
1908 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1909 ASSERT(ab->b_buf == NULL);
1910 ARCSTAT_BUMP(arcstat_deleted);
1911 bytes_deleted += ab->b_size;
1913 if (ab->b_l2hdr != NULL) {
1915 * This buffer is cached on the 2nd Level ARC;
1916 * don't destroy the header.
1918 arc_change_state(arc_l2c_only, ab, hash_lock);
1919 mutex_exit(hash_lock);
1921 arc_change_state(arc_anon, ab, hash_lock);
1922 mutex_exit(hash_lock);
1923 arc_hdr_destroy(ab);
1926 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1927 if (bytes >= 0 && bytes_deleted >= bytes)
1929 } else if (bytes < 0) {
1931 * Insert a list marker and then wait for the
1932 * hash lock to become available. Once its
1933 * available, restart from where we left off.
1935 list_insert_after(list, ab, &marker);
1936 mutex_exit(&state->arcs_mtx);
1937 mutex_enter(hash_lock);
1938 mutex_exit(hash_lock);
1939 mutex_enter(&state->arcs_mtx);
1940 ab_prev = list_prev(list, &marker);
1941 list_remove(list, &marker);
1945 mutex_exit(&state->arcs_mtx);
1947 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1948 (bytes < 0 || bytes_deleted < bytes)) {
1949 list = &state->arcs_list[ARC_BUFC_METADATA];
1954 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1958 if (bytes_deleted < bytes)
1959 dprintf("only deleted %lld bytes from %p\n",
1960 (longlong_t)bytes_deleted, state);
1966 int64_t adjustment, delta;
1972 adjustment = MIN((int64_t)(arc_size - arc_c),
1973 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
1976 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1977 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
1978 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
1979 adjustment -= delta;
1982 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1983 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
1984 (void) arc_evict(arc_mru, 0, delta, FALSE,
1992 adjustment = arc_size - arc_c;
1994 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1995 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
1996 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
1997 adjustment -= delta;
2000 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2001 int64_t delta = MIN(adjustment,
2002 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2003 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2008 * Adjust ghost lists
2011 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2013 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2014 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2015 arc_evict_ghost(arc_mru_ghost, 0, delta);
2019 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2021 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2022 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2023 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2028 * Request that arc user drop references so that N bytes can be released
2029 * from the cache. This provides a mechanism to ensure the arc can honor
2030 * the arc_meta_limit and reclaim buffers which are pinned in the cache
2031 * by higher layers. (i.e. the zpl)
2034 arc_do_user_prune(int64_t adjustment)
2036 arc_prune_func_t *func;
2038 arc_prune_t *cp, *np;
2040 mutex_enter(&arc_prune_mtx);
2042 cp = list_head(&arc_prune_list);
2043 while (cp != NULL) {
2045 private = cp->p_private;
2046 np = list_next(&arc_prune_list, cp);
2047 refcount_add(&cp->p_refcnt, func);
2048 mutex_exit(&arc_prune_mtx);
2051 func(adjustment, private);
2053 mutex_enter(&arc_prune_mtx);
2055 /* User removed prune callback concurrently with execution */
2056 if (refcount_remove(&cp->p_refcnt, func) == 0) {
2057 ASSERT(!list_link_active(&cp->p_node));
2058 refcount_destroy(&cp->p_refcnt);
2059 kmem_free(cp, sizeof (*cp));
2065 ARCSTAT_BUMP(arcstat_prune);
2066 mutex_exit(&arc_prune_mtx);
2070 arc_do_user_evicts(void)
2072 mutex_enter(&arc_eviction_mtx);
2073 while (arc_eviction_list != NULL) {
2074 arc_buf_t *buf = arc_eviction_list;
2075 arc_eviction_list = buf->b_next;
2076 mutex_enter(&buf->b_evict_lock);
2078 mutex_exit(&buf->b_evict_lock);
2079 mutex_exit(&arc_eviction_mtx);
2081 if (buf->b_efunc != NULL)
2082 VERIFY(buf->b_efunc(buf) == 0);
2084 buf->b_efunc = NULL;
2085 buf->b_private = NULL;
2086 kmem_cache_free(buf_cache, buf);
2087 mutex_enter(&arc_eviction_mtx);
2089 mutex_exit(&arc_eviction_mtx);
2093 * Evict only meta data objects from the cache leaving the data objects.
2094 * This is only used to enforce the tunable arc_meta_limit, if we are
2095 * unable to evict enough buffers notify the user via the prune callback.
2098 arc_adjust_meta(int64_t adjustment, boolean_t may_prune)
2102 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2103 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2104 arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_METADATA);
2105 adjustment -= delta;
2108 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2109 delta = MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2110 arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_METADATA);
2111 adjustment -= delta;
2114 if (may_prune && (adjustment > 0) && (arc_meta_used > arc_meta_limit))
2115 arc_do_user_prune(arc_meta_prune);
2119 * Flush all *evictable* data from the cache for the given spa.
2120 * NOTE: this will not touch "active" (i.e. referenced) data.
2123 arc_flush(spa_t *spa)
2128 guid = spa_load_guid(spa);
2130 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2131 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2135 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2136 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2140 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2141 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2145 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2146 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2151 arc_evict_ghost(arc_mru_ghost, guid, -1);
2152 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2154 mutex_enter(&arc_reclaim_thr_lock);
2155 arc_do_user_evicts();
2156 mutex_exit(&arc_reclaim_thr_lock);
2157 ASSERT(spa || arc_eviction_list == NULL);
2161 arc_shrink(uint64_t bytes)
2163 if (arc_c > arc_c_min) {
2166 to_free = bytes ? bytes : arc_c >> arc_shrink_shift;
2168 if (arc_c > arc_c_min + to_free)
2169 atomic_add_64(&arc_c, -to_free);
2173 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2174 if (arc_c > arc_size)
2175 arc_c = MAX(arc_size, arc_c_min);
2177 arc_p = (arc_c >> 1);
2178 ASSERT(arc_c >= arc_c_min);
2179 ASSERT((int64_t)arc_p >= 0);
2182 if (arc_size > arc_c)
2187 arc_kmem_reap_now(arc_reclaim_strategy_t strat, uint64_t bytes)
2190 kmem_cache_t *prev_cache = NULL;
2191 kmem_cache_t *prev_data_cache = NULL;
2192 extern kmem_cache_t *zio_buf_cache[];
2193 extern kmem_cache_t *zio_data_buf_cache[];
2196 * An aggressive reclamation will shrink the cache size as well as
2197 * reap free buffers from the arc kmem caches.
2199 if (strat == ARC_RECLAIM_AGGR)
2202 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2203 if (zio_buf_cache[i] != prev_cache) {
2204 prev_cache = zio_buf_cache[i];
2205 kmem_cache_reap_now(zio_buf_cache[i]);
2207 if (zio_data_buf_cache[i] != prev_data_cache) {
2208 prev_data_cache = zio_data_buf_cache[i];
2209 kmem_cache_reap_now(zio_data_buf_cache[i]);
2213 kmem_cache_reap_now(buf_cache);
2214 kmem_cache_reap_now(hdr_cache);
2218 * Unlike other ZFS implementations this thread is only responsible for
2219 * adapting the target ARC size on Linux. The responsibility for memory
2220 * reclamation has been entirely delegated to the arc_shrinker_func()
2221 * which is registered with the VM. To reflect this change in behavior
2222 * the arc_reclaim thread has been renamed to arc_adapt.
2225 arc_adapt_thread(void)
2230 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2232 mutex_enter(&arc_reclaim_thr_lock);
2233 while (arc_thread_exit == 0) {
2235 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2237 if (spa_get_random(100) == 0) {
2240 if (last_reclaim == ARC_RECLAIM_CONS) {
2241 last_reclaim = ARC_RECLAIM_AGGR;
2243 last_reclaim = ARC_RECLAIM_CONS;
2247 last_reclaim = ARC_RECLAIM_AGGR;
2251 /* reset the growth delay for every reclaim */
2252 arc_grow_time = ddi_get_lbolt()+(arc_grow_retry * hz);
2254 arc_kmem_reap_now(last_reclaim, 0);
2257 #endif /* !_KERNEL */
2259 /* No recent memory pressure allow the ARC to grow. */
2260 if (arc_no_grow && ddi_get_lbolt() >= arc_grow_time)
2261 arc_no_grow = FALSE;
2264 * Keep meta data usage within limits, arc_shrink() is not
2265 * used to avoid collapsing the arc_c value when only the
2266 * arc_meta_limit is being exceeded.
2268 prune = (int64_t)arc_meta_used - (int64_t)arc_meta_limit;
2270 arc_adjust_meta(prune, B_TRUE);
2274 if (arc_eviction_list != NULL)
2275 arc_do_user_evicts();
2277 /* block until needed, or one second, whichever is shorter */
2278 CALLB_CPR_SAFE_BEGIN(&cpr);
2279 (void) cv_timedwait_interruptible(&arc_reclaim_thr_cv,
2280 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2281 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2284 arc_thread_exit = 0;
2285 cv_broadcast(&arc_reclaim_thr_cv);
2286 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2292 * Determine the amount of memory eligible for eviction contained in the
2293 * ARC. All clean data reported by the ghost lists can always be safely
2294 * evicted. Due to arc_c_min, the same does not hold for all clean data
2295 * contained by the regular mru and mfu lists.
2297 * In the case of the regular mru and mfu lists, we need to report as
2298 * much clean data as possible, such that evicting that same reported
2299 * data will not bring arc_size below arc_c_min. Thus, in certain
2300 * circumstances, the total amount of clean data in the mru and mfu
2301 * lists might not actually be evictable.
2303 * The following two distinct cases are accounted for:
2305 * 1. The sum of the amount of dirty data contained by both the mru and
2306 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2307 * is greater than or equal to arc_c_min.
2308 * (i.e. amount of dirty data >= arc_c_min)
2310 * This is the easy case; all clean data contained by the mru and mfu
2311 * lists is evictable. Evicting all clean data can only drop arc_size
2312 * to the amount of dirty data, which is greater than arc_c_min.
2314 * 2. The sum of the amount of dirty data contained by both the mru and
2315 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2316 * is less than arc_c_min.
2317 * (i.e. arc_c_min > amount of dirty data)
2319 * 2.1. arc_size is greater than or equal arc_c_min.
2320 * (i.e. arc_size >= arc_c_min > amount of dirty data)
2322 * In this case, not all clean data from the regular mru and mfu
2323 * lists is actually evictable; we must leave enough clean data
2324 * to keep arc_size above arc_c_min. Thus, the maximum amount of
2325 * evictable data from the two lists combined, is exactly the
2326 * difference between arc_size and arc_c_min.
2328 * 2.2. arc_size is less than arc_c_min
2329 * (i.e. arc_c_min > arc_size > amount of dirty data)
2331 * In this case, none of the data contained in the mru and mfu
2332 * lists is evictable, even if it's clean. Since arc_size is
2333 * already below arc_c_min, evicting any more would only
2334 * increase this negative difference.
2337 arc_evictable_memory(void) {
2338 uint64_t arc_clean =
2339 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
2340 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
2341 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
2342 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
2343 uint64_t ghost_clean =
2344 arc_mru_ghost->arcs_lsize[ARC_BUFC_DATA] +
2345 arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] +
2346 arc_mfu_ghost->arcs_lsize[ARC_BUFC_DATA] +
2347 arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA];
2348 uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0);
2350 if (arc_dirty >= arc_c_min)
2351 return (ghost_clean + arc_clean);
2353 return (ghost_clean + MAX((int64_t)arc_size - (int64_t)arc_c_min, 0));
2357 __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc)
2361 /* The arc is considered warm once reclaim has occurred */
2362 if (unlikely(arc_warm == B_FALSE))
2365 /* Return the potential number of reclaimable pages */
2366 pages = btop(arc_evictable_memory());
2367 if (sc->nr_to_scan == 0)
2370 /* Not allowed to perform filesystem reclaim */
2371 if (!(sc->gfp_mask & __GFP_FS))
2374 /* Reclaim in progress */
2375 if (mutex_tryenter(&arc_reclaim_thr_lock) == 0)
2379 * Evict the requested number of pages by shrinking arc_c the
2380 * requested amount. If there is nothing left to evict just
2381 * reap whatever we can from the various arc slabs.
2384 arc_kmem_reap_now(ARC_RECLAIM_AGGR, ptob(sc->nr_to_scan));
2385 pages = btop(arc_evictable_memory());
2387 arc_kmem_reap_now(ARC_RECLAIM_CONS, ptob(sc->nr_to_scan));
2392 * When direct reclaim is observed it usually indicates a rapid
2393 * increase in memory pressure. This occurs because the kswapd
2394 * threads were unable to asynchronously keep enough free memory
2395 * available. In this case set arc_no_grow to briefly pause arc
2396 * growth to avoid compounding the memory pressure.
2398 if (current_is_kswapd()) {
2399 ARCSTAT_BUMP(arcstat_memory_indirect_count);
2401 arc_no_grow = B_TRUE;
2402 arc_grow_time = ddi_get_lbolt() + (arc_grow_retry * hz);
2403 ARCSTAT_BUMP(arcstat_memory_direct_count);
2406 mutex_exit(&arc_reclaim_thr_lock);
2410 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func);
2412 SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS);
2413 #endif /* _KERNEL */
2416 * Adapt arc info given the number of bytes we are trying to add and
2417 * the state that we are comming from. This function is only called
2418 * when we are adding new content to the cache.
2421 arc_adapt(int bytes, arc_state_t *state)
2424 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2426 if (state == arc_l2c_only)
2431 * Adapt the target size of the MRU list:
2432 * - if we just hit in the MRU ghost list, then increase
2433 * the target size of the MRU list.
2434 * - if we just hit in the MFU ghost list, then increase
2435 * the target size of the MFU list by decreasing the
2436 * target size of the MRU list.
2438 if (state == arc_mru_ghost) {
2439 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2440 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2441 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2443 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2444 } else if (state == arc_mfu_ghost) {
2447 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2448 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2449 mult = MIN(mult, 10);
2451 delta = MIN(bytes * mult, arc_p);
2452 arc_p = MAX(arc_p_min, arc_p - delta);
2454 ASSERT((int64_t)arc_p >= 0);
2459 if (arc_c >= arc_c_max)
2463 * If we're within (2 * maxblocksize) bytes of the target
2464 * cache size, increment the target cache size
2466 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2467 atomic_add_64(&arc_c, (int64_t)bytes);
2468 if (arc_c > arc_c_max)
2470 else if (state == arc_anon)
2471 atomic_add_64(&arc_p, (int64_t)bytes);
2475 ASSERT((int64_t)arc_p >= 0);
2479 * Check if the cache has reached its limits and eviction is required
2483 arc_evict_needed(arc_buf_contents_t type)
2485 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2491 return (arc_size > arc_c);
2495 * The buffer, supplied as the first argument, needs a data block.
2496 * So, if we are at cache max, determine which cache should be victimized.
2497 * We have the following cases:
2499 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2500 * In this situation if we're out of space, but the resident size of the MFU is
2501 * under the limit, victimize the MFU cache to satisfy this insertion request.
2503 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2504 * Here, we've used up all of the available space for the MRU, so we need to
2505 * evict from our own cache instead. Evict from the set of resident MRU
2508 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2509 * c minus p represents the MFU space in the cache, since p is the size of the
2510 * cache that is dedicated to the MRU. In this situation there's still space on
2511 * the MFU side, so the MRU side needs to be victimized.
2513 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2514 * MFU's resident set is consuming more space than it has been allotted. In
2515 * this situation, we must victimize our own cache, the MFU, for this insertion.
2518 arc_get_data_buf(arc_buf_t *buf)
2520 arc_state_t *state = buf->b_hdr->b_state;
2521 uint64_t size = buf->b_hdr->b_size;
2522 arc_buf_contents_t type = buf->b_hdr->b_type;
2524 arc_adapt(size, state);
2527 * We have not yet reached cache maximum size,
2528 * just allocate a new buffer.
2530 if (!arc_evict_needed(type)) {
2531 if (type == ARC_BUFC_METADATA) {
2532 buf->b_data = zio_buf_alloc(size);
2533 arc_space_consume(size, ARC_SPACE_DATA);
2535 ASSERT(type == ARC_BUFC_DATA);
2536 buf->b_data = zio_data_buf_alloc(size);
2537 ARCSTAT_INCR(arcstat_data_size, size);
2538 atomic_add_64(&arc_size, size);
2544 * If we are prefetching from the mfu ghost list, this buffer
2545 * will end up on the mru list; so steal space from there.
2547 if (state == arc_mfu_ghost)
2548 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2549 else if (state == arc_mru_ghost)
2552 if (state == arc_mru || state == arc_anon) {
2553 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2554 state = (arc_mfu->arcs_lsize[type] >= size &&
2555 arc_p > mru_used) ? arc_mfu : arc_mru;
2558 uint64_t mfu_space = arc_c - arc_p;
2559 state = (arc_mru->arcs_lsize[type] >= size &&
2560 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2563 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2564 if (type == ARC_BUFC_METADATA) {
2565 buf->b_data = zio_buf_alloc(size);
2566 arc_space_consume(size, ARC_SPACE_DATA);
2569 * If we are unable to recycle an existing meta buffer
2570 * signal the reclaim thread. It will notify users
2571 * via the prune callback to drop references. The
2572 * prune callback in run in the context of the reclaim
2573 * thread to avoid deadlocking on the hash_lock.
2575 cv_signal(&arc_reclaim_thr_cv);
2577 ASSERT(type == ARC_BUFC_DATA);
2578 buf->b_data = zio_data_buf_alloc(size);
2579 ARCSTAT_INCR(arcstat_data_size, size);
2580 atomic_add_64(&arc_size, size);
2583 ARCSTAT_BUMP(arcstat_recycle_miss);
2585 ASSERT(buf->b_data != NULL);
2588 * Update the state size. Note that ghost states have a
2589 * "ghost size" and so don't need to be updated.
2591 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2592 arc_buf_hdr_t *hdr = buf->b_hdr;
2594 atomic_add_64(&hdr->b_state->arcs_size, size);
2595 if (list_link_active(&hdr->b_arc_node)) {
2596 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2597 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2600 * If we are growing the cache, and we are adding anonymous
2601 * data, and we have outgrown arc_p, update arc_p
2603 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2604 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2605 arc_p = MIN(arc_c, arc_p + size);
2610 * This routine is called whenever a buffer is accessed.
2611 * NOTE: the hash lock is dropped in this function.
2614 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2618 ASSERT(MUTEX_HELD(hash_lock));
2620 if (buf->b_state == arc_anon) {
2622 * This buffer is not in the cache, and does not
2623 * appear in our "ghost" list. Add the new buffer
2627 ASSERT(buf->b_arc_access == 0);
2628 buf->b_arc_access = ddi_get_lbolt();
2629 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2630 arc_change_state(arc_mru, buf, hash_lock);
2632 } else if (buf->b_state == arc_mru) {
2633 now = ddi_get_lbolt();
2636 * If this buffer is here because of a prefetch, then either:
2637 * - clear the flag if this is a "referencing" read
2638 * (any subsequent access will bump this into the MFU state).
2640 * - move the buffer to the head of the list if this is
2641 * another prefetch (to make it less likely to be evicted).
2643 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2644 if (refcount_count(&buf->b_refcnt) == 0) {
2645 ASSERT(list_link_active(&buf->b_arc_node));
2647 buf->b_flags &= ~ARC_PREFETCH;
2648 ARCSTAT_BUMP(arcstat_mru_hits);
2650 buf->b_arc_access = now;
2655 * This buffer has been "accessed" only once so far,
2656 * but it is still in the cache. Move it to the MFU
2659 if (now > buf->b_arc_access + ARC_MINTIME) {
2661 * More than 125ms have passed since we
2662 * instantiated this buffer. Move it to the
2663 * most frequently used state.
2665 buf->b_arc_access = now;
2666 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2667 arc_change_state(arc_mfu, buf, hash_lock);
2669 ARCSTAT_BUMP(arcstat_mru_hits);
2670 } else if (buf->b_state == arc_mru_ghost) {
2671 arc_state_t *new_state;
2673 * This buffer has been "accessed" recently, but
2674 * was evicted from the cache. Move it to the
2678 if (buf->b_flags & ARC_PREFETCH) {
2679 new_state = arc_mru;
2680 if (refcount_count(&buf->b_refcnt) > 0)
2681 buf->b_flags &= ~ARC_PREFETCH;
2682 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2684 new_state = arc_mfu;
2685 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2688 buf->b_arc_access = ddi_get_lbolt();
2689 arc_change_state(new_state, buf, hash_lock);
2691 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2692 } else if (buf->b_state == arc_mfu) {
2694 * This buffer has been accessed more than once and is
2695 * still in the cache. Keep it in the MFU state.
2697 * NOTE: an add_reference() that occurred when we did
2698 * the arc_read() will have kicked this off the list.
2699 * If it was a prefetch, we will explicitly move it to
2700 * the head of the list now.
2702 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2703 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2704 ASSERT(list_link_active(&buf->b_arc_node));
2706 ARCSTAT_BUMP(arcstat_mfu_hits);
2707 buf->b_arc_access = ddi_get_lbolt();
2708 } else if (buf->b_state == arc_mfu_ghost) {
2709 arc_state_t *new_state = arc_mfu;
2711 * This buffer has been accessed more than once but has
2712 * been evicted from the cache. Move it back to the
2716 if (buf->b_flags & ARC_PREFETCH) {
2718 * This is a prefetch access...
2719 * move this block back to the MRU state.
2721 ASSERT0(refcount_count(&buf->b_refcnt));
2722 new_state = arc_mru;
2725 buf->b_arc_access = ddi_get_lbolt();
2726 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2727 arc_change_state(new_state, buf, hash_lock);
2729 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2730 } else if (buf->b_state == arc_l2c_only) {
2732 * This buffer is on the 2nd Level ARC.
2735 buf->b_arc_access = ddi_get_lbolt();
2736 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2737 arc_change_state(arc_mfu, buf, hash_lock);
2739 ASSERT(!"invalid arc state");
2743 /* a generic arc_done_func_t which you can use */
2746 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2748 if (zio == NULL || zio->io_error == 0)
2749 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2750 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2753 /* a generic arc_done_func_t */
2755 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2757 arc_buf_t **bufp = arg;
2758 if (zio && zio->io_error) {
2759 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2763 ASSERT(buf->b_data);
2768 arc_read_done(zio_t *zio)
2770 arc_buf_hdr_t *hdr, *found;
2772 arc_buf_t *abuf; /* buffer we're assigning to callback */
2773 kmutex_t *hash_lock;
2774 arc_callback_t *callback_list, *acb;
2775 int freeable = FALSE;
2777 buf = zio->io_private;
2781 * The hdr was inserted into hash-table and removed from lists
2782 * prior to starting I/O. We should find this header, since
2783 * it's in the hash table, and it should be legit since it's
2784 * not possible to evict it during the I/O. The only possible
2785 * reason for it not to be found is if we were freed during the
2788 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2791 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2792 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2793 (found == hdr && HDR_L2_READING(hdr)));
2795 hdr->b_flags &= ~ARC_L2_EVICTED;
2796 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2797 hdr->b_flags &= ~ARC_L2CACHE;
2799 /* byteswap if necessary */
2800 callback_list = hdr->b_acb;
2801 ASSERT(callback_list != NULL);
2802 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2803 dmu_object_byteswap_t bswap =
2804 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2805 if (BP_GET_LEVEL(zio->io_bp) > 0)
2806 byteswap_uint64_array(buf->b_data, hdr->b_size);
2808 dmu_ot_byteswap[bswap].ob_func(buf->b_data, hdr->b_size);
2811 arc_cksum_compute(buf, B_FALSE);
2813 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2815 * Only call arc_access on anonymous buffers. This is because
2816 * if we've issued an I/O for an evicted buffer, we've already
2817 * called arc_access (to prevent any simultaneous readers from
2818 * getting confused).
2820 arc_access(hdr, hash_lock);
2823 /* create copies of the data buffer for the callers */
2825 for (acb = callback_list; acb; acb = acb->acb_next) {
2826 if (acb->acb_done) {
2828 ARCSTAT_BUMP(arcstat_duplicate_reads);
2829 abuf = arc_buf_clone(buf);
2831 acb->acb_buf = abuf;
2836 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2837 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2839 ASSERT(buf->b_efunc == NULL);
2840 ASSERT(hdr->b_datacnt == 1);
2841 hdr->b_flags |= ARC_BUF_AVAILABLE;
2844 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2846 if (zio->io_error != 0) {
2847 hdr->b_flags |= ARC_IO_ERROR;
2848 if (hdr->b_state != arc_anon)
2849 arc_change_state(arc_anon, hdr, hash_lock);
2850 if (HDR_IN_HASH_TABLE(hdr))
2851 buf_hash_remove(hdr);
2852 freeable = refcount_is_zero(&hdr->b_refcnt);
2856 * Broadcast before we drop the hash_lock to avoid the possibility
2857 * that the hdr (and hence the cv) might be freed before we get to
2858 * the cv_broadcast().
2860 cv_broadcast(&hdr->b_cv);
2863 mutex_exit(hash_lock);
2866 * This block was freed while we waited for the read to
2867 * complete. It has been removed from the hash table and
2868 * moved to the anonymous state (so that it won't show up
2871 ASSERT3P(hdr->b_state, ==, arc_anon);
2872 freeable = refcount_is_zero(&hdr->b_refcnt);
2875 /* execute each callback and free its structure */
2876 while ((acb = callback_list) != NULL) {
2878 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2880 if (acb->acb_zio_dummy != NULL) {
2881 acb->acb_zio_dummy->io_error = zio->io_error;
2882 zio_nowait(acb->acb_zio_dummy);
2885 callback_list = acb->acb_next;
2886 kmem_free(acb, sizeof (arc_callback_t));
2890 arc_hdr_destroy(hdr);
2894 * "Read" the block at the specified DVA (in bp) via the
2895 * cache. If the block is found in the cache, invoke the provided
2896 * callback immediately and return. Note that the `zio' parameter
2897 * in the callback will be NULL in this case, since no IO was
2898 * required. If the block is not in the cache pass the read request
2899 * on to the spa with a substitute callback function, so that the
2900 * requested block will be added to the cache.
2902 * If a read request arrives for a block that has a read in-progress,
2903 * either wait for the in-progress read to complete (and return the
2904 * results); or, if this is a read with a "done" func, add a record
2905 * to the read to invoke the "done" func when the read completes,
2906 * and return; or just return.
2908 * arc_read_done() will invoke all the requested "done" functions
2909 * for readers of this block.
2911 * Normal callers should use arc_read and pass the arc buffer and offset
2912 * for the bp. But if you know you don't need locking, you can use
2916 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_buf_t *pbuf,
2917 arc_done_func_t *done, void *private, int priority, int zio_flags,
2918 uint32_t *arc_flags, const zbookmark_t *zb)
2924 * XXX This happens from traverse callback funcs, for
2925 * the objset_phys_t block.
2927 return (arc_read_nolock(pio, spa, bp, done, private, priority,
2928 zio_flags, arc_flags, zb));
2931 ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt));
2932 ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size);
2933 rw_enter(&pbuf->b_data_lock, RW_READER);
2935 err = arc_read_nolock(pio, spa, bp, done, private, priority,
2936 zio_flags, arc_flags, zb);
2937 rw_exit(&pbuf->b_data_lock);
2943 arc_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bp,
2944 arc_done_func_t *done, void *private, int priority, int zio_flags,
2945 uint32_t *arc_flags, const zbookmark_t *zb)
2948 arc_buf_t *buf = NULL;
2949 kmutex_t *hash_lock;
2951 uint64_t guid = spa_load_guid(spa);
2954 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2956 if (hdr && hdr->b_datacnt > 0) {
2958 *arc_flags |= ARC_CACHED;
2960 if (HDR_IO_IN_PROGRESS(hdr)) {
2962 if (*arc_flags & ARC_WAIT) {
2963 cv_wait(&hdr->b_cv, hash_lock);
2964 mutex_exit(hash_lock);
2967 ASSERT(*arc_flags & ARC_NOWAIT);
2970 arc_callback_t *acb = NULL;
2972 acb = kmem_zalloc(sizeof (arc_callback_t),
2974 acb->acb_done = done;
2975 acb->acb_private = private;
2977 acb->acb_zio_dummy = zio_null(pio,
2978 spa, NULL, NULL, NULL, zio_flags);
2980 ASSERT(acb->acb_done != NULL);
2981 acb->acb_next = hdr->b_acb;
2983 add_reference(hdr, hash_lock, private);
2984 mutex_exit(hash_lock);
2987 mutex_exit(hash_lock);
2991 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2994 add_reference(hdr, hash_lock, private);
2996 * If this block is already in use, create a new
2997 * copy of the data so that we will be guaranteed
2998 * that arc_release() will always succeed.
3002 ASSERT(buf->b_data);
3003 if (HDR_BUF_AVAILABLE(hdr)) {
3004 ASSERT(buf->b_efunc == NULL);
3005 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3007 buf = arc_buf_clone(buf);
3010 } else if (*arc_flags & ARC_PREFETCH &&
3011 refcount_count(&hdr->b_refcnt) == 0) {
3012 hdr->b_flags |= ARC_PREFETCH;
3014 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
3015 arc_access(hdr, hash_lock);
3016 if (*arc_flags & ARC_L2CACHE)
3017 hdr->b_flags |= ARC_L2CACHE;
3018 mutex_exit(hash_lock);
3019 ARCSTAT_BUMP(arcstat_hits);
3020 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3021 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3022 data, metadata, hits);
3025 done(NULL, buf, private);
3027 uint64_t size = BP_GET_LSIZE(bp);
3028 arc_callback_t *acb;
3031 boolean_t devw = B_FALSE;
3034 /* this block is not in the cache */
3035 arc_buf_hdr_t *exists;
3036 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
3037 buf = arc_buf_alloc(spa, size, private, type);
3039 hdr->b_dva = *BP_IDENTITY(bp);
3040 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3041 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3042 exists = buf_hash_insert(hdr, &hash_lock);
3044 /* somebody beat us to the hash insert */
3045 mutex_exit(hash_lock);
3046 buf_discard_identity(hdr);
3047 (void) arc_buf_remove_ref(buf, private);
3048 goto top; /* restart the IO request */
3050 /* if this is a prefetch, we don't have a reference */
3051 if (*arc_flags & ARC_PREFETCH) {
3052 (void) remove_reference(hdr, hash_lock,
3054 hdr->b_flags |= ARC_PREFETCH;
3056 if (*arc_flags & ARC_L2CACHE)
3057 hdr->b_flags |= ARC_L2CACHE;
3058 if (BP_GET_LEVEL(bp) > 0)
3059 hdr->b_flags |= ARC_INDIRECT;
3061 /* this block is in the ghost cache */
3062 ASSERT(GHOST_STATE(hdr->b_state));
3063 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3064 ASSERT0(refcount_count(&hdr->b_refcnt));
3065 ASSERT(hdr->b_buf == NULL);
3067 /* if this is a prefetch, we don't have a reference */
3068 if (*arc_flags & ARC_PREFETCH)
3069 hdr->b_flags |= ARC_PREFETCH;
3071 add_reference(hdr, hash_lock, private);
3072 if (*arc_flags & ARC_L2CACHE)
3073 hdr->b_flags |= ARC_L2CACHE;
3074 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3077 buf->b_efunc = NULL;
3078 buf->b_private = NULL;
3081 ASSERT(hdr->b_datacnt == 0);
3083 arc_get_data_buf(buf);
3084 arc_access(hdr, hash_lock);
3087 ASSERT(!GHOST_STATE(hdr->b_state));
3089 acb = kmem_zalloc(sizeof (arc_callback_t), KM_PUSHPAGE);
3090 acb->acb_done = done;
3091 acb->acb_private = private;
3093 ASSERT(hdr->b_acb == NULL);
3095 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3097 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3098 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3099 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3100 addr = hdr->b_l2hdr->b_daddr;
3102 * Lock out device removal.
3104 if (vdev_is_dead(vd) ||
3105 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3109 mutex_exit(hash_lock);
3111 ASSERT3U(hdr->b_size, ==, size);
3112 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3113 uint64_t, size, zbookmark_t *, zb);
3114 ARCSTAT_BUMP(arcstat_misses);
3115 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3116 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3117 data, metadata, misses);
3119 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3121 * Read from the L2ARC if the following are true:
3122 * 1. The L2ARC vdev was previously cached.
3123 * 2. This buffer still has L2ARC metadata.
3124 * 3. This buffer isn't currently writing to the L2ARC.
3125 * 4. The L2ARC entry wasn't evicted, which may
3126 * also have invalidated the vdev.
3127 * 5. This isn't prefetch and l2arc_noprefetch is set.
3129 if (hdr->b_l2hdr != NULL &&
3130 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3131 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3132 l2arc_read_callback_t *cb;
3134 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3135 ARCSTAT_BUMP(arcstat_l2_hits);
3137 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3139 cb->l2rcb_buf = buf;
3140 cb->l2rcb_spa = spa;
3143 cb->l2rcb_flags = zio_flags;
3146 * l2arc read. The SCL_L2ARC lock will be
3147 * released by l2arc_read_done().
3149 rzio = zio_read_phys(pio, vd, addr, size,
3150 buf->b_data, ZIO_CHECKSUM_OFF,
3151 l2arc_read_done, cb, priority, zio_flags |
3152 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
3153 ZIO_FLAG_DONT_PROPAGATE |
3154 ZIO_FLAG_DONT_RETRY, B_FALSE);
3155 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3157 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
3159 if (*arc_flags & ARC_NOWAIT) {
3164 ASSERT(*arc_flags & ARC_WAIT);
3165 if (zio_wait(rzio) == 0)
3168 /* l2arc read error; goto zio_read() */
3170 DTRACE_PROBE1(l2arc__miss,
3171 arc_buf_hdr_t *, hdr);
3172 ARCSTAT_BUMP(arcstat_l2_misses);
3173 if (HDR_L2_WRITING(hdr))
3174 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3175 spa_config_exit(spa, SCL_L2ARC, vd);
3179 spa_config_exit(spa, SCL_L2ARC, vd);
3180 if (l2arc_ndev != 0) {
3181 DTRACE_PROBE1(l2arc__miss,
3182 arc_buf_hdr_t *, hdr);
3183 ARCSTAT_BUMP(arcstat_l2_misses);
3187 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3188 arc_read_done, buf, priority, zio_flags, zb);
3190 if (*arc_flags & ARC_WAIT)
3191 return (zio_wait(rzio));
3193 ASSERT(*arc_flags & ARC_NOWAIT);
3200 arc_add_prune_callback(arc_prune_func_t *func, void *private)
3204 p = kmem_alloc(sizeof(*p), KM_SLEEP);
3206 p->p_private = private;
3207 list_link_init(&p->p_node);
3208 refcount_create(&p->p_refcnt);
3210 mutex_enter(&arc_prune_mtx);
3211 refcount_add(&p->p_refcnt, &arc_prune_list);
3212 list_insert_head(&arc_prune_list, p);
3213 mutex_exit(&arc_prune_mtx);
3219 arc_remove_prune_callback(arc_prune_t *p)
3221 mutex_enter(&arc_prune_mtx);
3222 list_remove(&arc_prune_list, p);
3223 if (refcount_remove(&p->p_refcnt, &arc_prune_list) == 0) {
3224 refcount_destroy(&p->p_refcnt);
3225 kmem_free(p, sizeof (*p));
3227 mutex_exit(&arc_prune_mtx);
3231 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3233 ASSERT(buf->b_hdr != NULL);
3234 ASSERT(buf->b_hdr->b_state != arc_anon);
3235 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3236 ASSERT(buf->b_efunc == NULL);
3237 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3239 buf->b_efunc = func;
3240 buf->b_private = private;
3244 * Notify the arc that a block was freed, and thus will never be used again.
3247 arc_freed(spa_t *spa, const blkptr_t *bp)
3250 kmutex_t *hash_lock;
3251 uint64_t guid = spa_load_guid(spa);
3253 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3257 if (HDR_BUF_AVAILABLE(hdr)) {
3258 arc_buf_t *buf = hdr->b_buf;
3259 add_reference(hdr, hash_lock, FTAG);
3260 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3261 mutex_exit(hash_lock);
3263 arc_release(buf, FTAG);
3264 (void) arc_buf_remove_ref(buf, FTAG);
3266 mutex_exit(hash_lock);
3272 * This is used by the DMU to let the ARC know that a buffer is
3273 * being evicted, so the ARC should clean up. If this arc buf
3274 * is not yet in the evicted state, it will be put there.
3277 arc_buf_evict(arc_buf_t *buf)
3280 kmutex_t *hash_lock;
3283 mutex_enter(&buf->b_evict_lock);
3287 * We are in arc_do_user_evicts().
3289 ASSERT(buf->b_data == NULL);
3290 mutex_exit(&buf->b_evict_lock);
3292 } else if (buf->b_data == NULL) {
3293 arc_buf_t copy = *buf; /* structure assignment */
3295 * We are on the eviction list; process this buffer now
3296 * but let arc_do_user_evicts() do the reaping.
3298 buf->b_efunc = NULL;
3299 mutex_exit(&buf->b_evict_lock);
3300 VERIFY(copy.b_efunc(©) == 0);
3303 hash_lock = HDR_LOCK(hdr);
3304 mutex_enter(hash_lock);
3306 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3308 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3309 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3312 * Pull this buffer off of the hdr
3315 while (*bufp != buf)
3316 bufp = &(*bufp)->b_next;
3317 *bufp = buf->b_next;
3319 ASSERT(buf->b_data != NULL);
3320 arc_buf_destroy(buf, FALSE, FALSE);
3322 if (hdr->b_datacnt == 0) {
3323 arc_state_t *old_state = hdr->b_state;
3324 arc_state_t *evicted_state;
3326 ASSERT(hdr->b_buf == NULL);
3327 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3330 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3332 mutex_enter(&old_state->arcs_mtx);
3333 mutex_enter(&evicted_state->arcs_mtx);
3335 arc_change_state(evicted_state, hdr, hash_lock);
3336 ASSERT(HDR_IN_HASH_TABLE(hdr));
3337 hdr->b_flags |= ARC_IN_HASH_TABLE;
3338 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3340 mutex_exit(&evicted_state->arcs_mtx);
3341 mutex_exit(&old_state->arcs_mtx);
3343 mutex_exit(hash_lock);
3344 mutex_exit(&buf->b_evict_lock);
3346 VERIFY(buf->b_efunc(buf) == 0);
3347 buf->b_efunc = NULL;
3348 buf->b_private = NULL;
3351 kmem_cache_free(buf_cache, buf);
3356 * Release this buffer from the cache. This must be done
3357 * after a read and prior to modifying the buffer contents.
3358 * If the buffer has more than one reference, we must make
3359 * a new hdr for the buffer.
3362 arc_release(arc_buf_t *buf, void *tag)
3365 kmutex_t *hash_lock = NULL;
3366 l2arc_buf_hdr_t *l2hdr;
3367 uint64_t buf_size = 0;
3370 * It would be nice to assert that if it's DMU metadata (level >
3371 * 0 || it's the dnode file), then it must be syncing context.
3372 * But we don't know that information at this level.
3375 mutex_enter(&buf->b_evict_lock);
3378 /* this buffer is not on any list */
3379 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3381 if (hdr->b_state == arc_anon) {
3382 /* this buffer is already released */
3383 ASSERT(buf->b_efunc == NULL);
3385 hash_lock = HDR_LOCK(hdr);
3386 mutex_enter(hash_lock);
3388 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3391 l2hdr = hdr->b_l2hdr;
3393 mutex_enter(&l2arc_buflist_mtx);
3394 hdr->b_l2hdr = NULL;
3395 buf_size = hdr->b_size;
3399 * Do we have more than one buf?
3401 if (hdr->b_datacnt > 1) {
3402 arc_buf_hdr_t *nhdr;
3404 uint64_t blksz = hdr->b_size;
3405 uint64_t spa = hdr->b_spa;
3406 arc_buf_contents_t type = hdr->b_type;
3407 uint32_t flags = hdr->b_flags;
3409 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3411 * Pull the data off of this hdr and attach it to
3412 * a new anonymous hdr.
3414 (void) remove_reference(hdr, hash_lock, tag);
3416 while (*bufp != buf)
3417 bufp = &(*bufp)->b_next;
3418 *bufp = buf->b_next;
3421 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3422 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3423 if (refcount_is_zero(&hdr->b_refcnt)) {
3424 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3425 ASSERT3U(*size, >=, hdr->b_size);
3426 atomic_add_64(size, -hdr->b_size);
3430 * We're releasing a duplicate user data buffer, update
3431 * our statistics accordingly.
3433 if (hdr->b_type == ARC_BUFC_DATA) {
3434 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3435 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3438 hdr->b_datacnt -= 1;
3439 arc_cksum_verify(buf);
3441 mutex_exit(hash_lock);
3443 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3444 nhdr->b_size = blksz;
3446 nhdr->b_type = type;
3448 nhdr->b_state = arc_anon;
3449 nhdr->b_arc_access = 0;
3450 nhdr->b_flags = flags & ARC_L2_WRITING;
3451 nhdr->b_l2hdr = NULL;
3452 nhdr->b_datacnt = 1;
3453 nhdr->b_freeze_cksum = NULL;
3454 (void) refcount_add(&nhdr->b_refcnt, tag);
3456 mutex_exit(&buf->b_evict_lock);
3457 atomic_add_64(&arc_anon->arcs_size, blksz);
3459 mutex_exit(&buf->b_evict_lock);
3460 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3461 ASSERT(!list_link_active(&hdr->b_arc_node));
3462 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3463 if (hdr->b_state != arc_anon)
3464 arc_change_state(arc_anon, hdr, hash_lock);
3465 hdr->b_arc_access = 0;
3467 mutex_exit(hash_lock);
3469 buf_discard_identity(hdr);
3472 buf->b_efunc = NULL;
3473 buf->b_private = NULL;
3476 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3477 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3478 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3479 mutex_exit(&l2arc_buflist_mtx);
3484 * Release this buffer. If it does not match the provided BP, fill it
3485 * with that block's contents.
3489 arc_release_bp(arc_buf_t *buf, void *tag, blkptr_t *bp, spa_t *spa,
3492 arc_release(buf, tag);
3497 arc_released(arc_buf_t *buf)
3501 mutex_enter(&buf->b_evict_lock);
3502 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3503 mutex_exit(&buf->b_evict_lock);
3508 arc_has_callback(arc_buf_t *buf)
3512 mutex_enter(&buf->b_evict_lock);
3513 callback = (buf->b_efunc != NULL);
3514 mutex_exit(&buf->b_evict_lock);
3520 arc_referenced(arc_buf_t *buf)
3524 mutex_enter(&buf->b_evict_lock);
3525 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3526 mutex_exit(&buf->b_evict_lock);
3527 return (referenced);
3532 arc_write_ready(zio_t *zio)
3534 arc_write_callback_t *callback = zio->io_private;
3535 arc_buf_t *buf = callback->awcb_buf;
3536 arc_buf_hdr_t *hdr = buf->b_hdr;
3538 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3539 callback->awcb_ready(zio, buf, callback->awcb_private);
3542 * If the IO is already in progress, then this is a re-write
3543 * attempt, so we need to thaw and re-compute the cksum.
3544 * It is the responsibility of the callback to handle the
3545 * accounting for any re-write attempt.
3547 if (HDR_IO_IN_PROGRESS(hdr)) {
3548 mutex_enter(&hdr->b_freeze_lock);
3549 if (hdr->b_freeze_cksum != NULL) {
3550 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3551 hdr->b_freeze_cksum = NULL;
3553 mutex_exit(&hdr->b_freeze_lock);
3555 arc_cksum_compute(buf, B_FALSE);
3556 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3560 arc_write_done(zio_t *zio)
3562 arc_write_callback_t *callback = zio->io_private;
3563 arc_buf_t *buf = callback->awcb_buf;
3564 arc_buf_hdr_t *hdr = buf->b_hdr;
3566 ASSERT(hdr->b_acb == NULL);
3568 if (zio->io_error == 0) {
3569 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3570 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3571 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3573 ASSERT(BUF_EMPTY(hdr));
3577 * If the block to be written was all-zero, we may have
3578 * compressed it away. In this case no write was performed
3579 * so there will be no dva/birth/checksum. The buffer must
3580 * therefore remain anonymous (and uncached).
3582 if (!BUF_EMPTY(hdr)) {
3583 arc_buf_hdr_t *exists;
3584 kmutex_t *hash_lock;
3586 ASSERT(zio->io_error == 0);
3588 arc_cksum_verify(buf);
3590 exists = buf_hash_insert(hdr, &hash_lock);
3593 * This can only happen if we overwrite for
3594 * sync-to-convergence, because we remove
3595 * buffers from the hash table when we arc_free().
3597 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3598 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3599 panic("bad overwrite, hdr=%p exists=%p",
3600 (void *)hdr, (void *)exists);
3601 ASSERT(refcount_is_zero(&exists->b_refcnt));
3602 arc_change_state(arc_anon, exists, hash_lock);
3603 mutex_exit(hash_lock);
3604 arc_hdr_destroy(exists);
3605 exists = buf_hash_insert(hdr, &hash_lock);
3606 ASSERT3P(exists, ==, NULL);
3609 ASSERT(hdr->b_datacnt == 1);
3610 ASSERT(hdr->b_state == arc_anon);
3611 ASSERT(BP_GET_DEDUP(zio->io_bp));
3612 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3615 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3616 /* if it's not anon, we are doing a scrub */
3617 if (!exists && hdr->b_state == arc_anon)
3618 arc_access(hdr, hash_lock);
3619 mutex_exit(hash_lock);
3621 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3624 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3625 callback->awcb_done(zio, buf, callback->awcb_private);
3627 kmem_free(callback, sizeof (arc_write_callback_t));
3631 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3632 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3633 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3634 int priority, int zio_flags, const zbookmark_t *zb)
3636 arc_buf_hdr_t *hdr = buf->b_hdr;
3637 arc_write_callback_t *callback;
3640 ASSERT(ready != NULL);
3641 ASSERT(done != NULL);
3642 ASSERT(!HDR_IO_ERROR(hdr));
3643 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3644 ASSERT(hdr->b_acb == NULL);
3646 hdr->b_flags |= ARC_L2CACHE;
3647 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_PUSHPAGE);
3648 callback->awcb_ready = ready;
3649 callback->awcb_done = done;
3650 callback->awcb_private = private;
3651 callback->awcb_buf = buf;
3653 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3654 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3660 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3663 uint64_t available_memory;
3665 if (zfs_arc_memory_throttle_disable)
3668 /* Easily reclaimable memory (free + inactive + arc-evictable) */
3669 available_memory = ptob(spl_kmem_availrmem()) + arc_evictable_memory();
3671 if (available_memory <= zfs_write_limit_max) {
3672 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3673 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
3677 if (inflight_data > available_memory / 4) {
3678 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3679 DMU_TX_STAT_BUMP(dmu_tx_memory_inflight);
3687 arc_tempreserve_clear(uint64_t reserve)
3689 atomic_add_64(&arc_tempreserve, -reserve);
3690 ASSERT((int64_t)arc_tempreserve >= 0);
3694 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3701 * Once in a while, fail for no reason. Everything should cope.
3703 if (spa_get_random(10000) == 0) {
3704 dprintf("forcing random failure\n");
3708 if (reserve > arc_c/4 && !arc_no_grow)
3709 arc_c = MIN(arc_c_max, reserve * 4);
3710 if (reserve > arc_c) {
3711 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
3716 * Don't count loaned bufs as in flight dirty data to prevent long
3717 * network delays from blocking transactions that are ready to be
3718 * assigned to a txg.
3720 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3723 * Writes will, almost always, require additional memory allocations
3724 * in order to compress/encrypt/etc the data. We therefor need to
3725 * make sure that there is sufficient available memory for this.
3727 if ((error = arc_memory_throttle(reserve, anon_size, txg)))
3731 * Throttle writes when the amount of dirty data in the cache
3732 * gets too large. We try to keep the cache less than half full
3733 * of dirty blocks so that our sync times don't grow too large.
3734 * Note: if two requests come in concurrently, we might let them
3735 * both succeed, when one of them should fail. Not a huge deal.
3738 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3739 anon_size > arc_c / 4) {
3740 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3741 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3742 arc_tempreserve>>10,
3743 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3744 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3745 reserve>>10, arc_c>>10);
3746 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
3749 atomic_add_64(&arc_tempreserve, reserve);
3754 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
3755 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
3757 size->value.ui64 = state->arcs_size;
3758 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
3759 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
3763 arc_kstat_update(kstat_t *ksp, int rw)
3765 arc_stats_t *as = ksp->ks_data;
3767 if (rw == KSTAT_WRITE) {
3770 arc_kstat_update_state(arc_anon,
3771 &as->arcstat_anon_size,
3772 &as->arcstat_anon_evict_data,
3773 &as->arcstat_anon_evict_metadata);
3774 arc_kstat_update_state(arc_mru,
3775 &as->arcstat_mru_size,
3776 &as->arcstat_mru_evict_data,
3777 &as->arcstat_mru_evict_metadata);
3778 arc_kstat_update_state(arc_mru_ghost,
3779 &as->arcstat_mru_ghost_size,
3780 &as->arcstat_mru_ghost_evict_data,
3781 &as->arcstat_mru_ghost_evict_metadata);
3782 arc_kstat_update_state(arc_mfu,
3783 &as->arcstat_mfu_size,
3784 &as->arcstat_mfu_evict_data,
3785 &as->arcstat_mfu_evict_metadata);
3786 arc_kstat_update_state(arc_mfu_ghost,
3787 &as->arcstat_mfu_ghost_size,
3788 &as->arcstat_mfu_ghost_evict_data,
3789 &as->arcstat_mfu_ghost_evict_metadata);
3798 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3799 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3801 /* Convert seconds to clock ticks */
3802 arc_min_prefetch_lifespan = 1 * hz;
3804 /* Start out with 1/8 of all memory */
3805 arc_c = physmem * PAGESIZE / 8;
3809 * On architectures where the physical memory can be larger
3810 * than the addressable space (intel in 32-bit mode), we may
3811 * need to limit the cache to 1/8 of VM size.
3813 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3815 * Register a shrinker to support synchronous (direct) memory
3816 * reclaim from the arc. This is done to prevent kswapd from
3817 * swapping out pages when it is preferable to shrink the arc.
3819 spl_register_shrinker(&arc_shrinker);
3822 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3823 arc_c_min = MAX(arc_c / 4, 64<<20);
3824 /* set max to 1/2 of all memory */
3825 arc_c_max = MAX(arc_c * 4, arc_c_max);
3828 * Allow the tunables to override our calculations if they are
3829 * reasonable (ie. over 64MB)
3831 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3832 arc_c_max = zfs_arc_max;
3833 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3834 arc_c_min = zfs_arc_min;
3837 arc_p = (arc_c >> 1);
3839 /* limit meta-data to 1/4 of the arc capacity */
3840 arc_meta_limit = arc_c_max / 4;
3843 /* Allow the tunable to override if it is reasonable */
3844 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3845 arc_meta_limit = zfs_arc_meta_limit;
3847 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3848 arc_c_min = arc_meta_limit / 2;
3850 if (zfs_arc_grow_retry > 0)
3851 arc_grow_retry = zfs_arc_grow_retry;
3853 if (zfs_arc_shrink_shift > 0)
3854 arc_shrink_shift = zfs_arc_shrink_shift;
3856 if (zfs_arc_p_min_shift > 0)
3857 arc_p_min_shift = zfs_arc_p_min_shift;
3859 if (zfs_arc_meta_prune > 0)
3860 arc_meta_prune = zfs_arc_meta_prune;
3862 /* if kmem_flags are set, lets try to use less memory */
3863 if (kmem_debugging())
3865 if (arc_c < arc_c_min)
3868 arc_anon = &ARC_anon;
3870 arc_mru_ghost = &ARC_mru_ghost;
3872 arc_mfu_ghost = &ARC_mfu_ghost;
3873 arc_l2c_only = &ARC_l2c_only;
3876 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3877 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3878 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3879 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3880 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3881 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3883 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3884 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3885 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3886 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3887 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3888 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3889 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3890 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3891 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3892 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3893 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3894 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3895 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3896 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3897 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3898 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3899 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3900 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3901 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3902 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3906 arc_thread_exit = 0;
3907 list_create(&arc_prune_list, sizeof (arc_prune_t),
3908 offsetof(arc_prune_t, p_node));
3909 arc_eviction_list = NULL;
3910 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
3911 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3912 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3914 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3915 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3917 if (arc_ksp != NULL) {
3918 arc_ksp->ks_data = &arc_stats;
3919 arc_ksp->ks_update = arc_kstat_update;
3920 kstat_install(arc_ksp);
3923 (void) thread_create(NULL, 0, arc_adapt_thread, NULL, 0, &p0,
3924 TS_RUN, minclsyspri);
3929 if (zfs_write_limit_max == 0)
3930 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3932 zfs_write_limit_shift = 0;
3933 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3941 mutex_enter(&arc_reclaim_thr_lock);
3943 spl_unregister_shrinker(&arc_shrinker);
3944 #endif /* _KERNEL */
3946 arc_thread_exit = 1;
3947 while (arc_thread_exit != 0)
3948 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3949 mutex_exit(&arc_reclaim_thr_lock);
3955 if (arc_ksp != NULL) {
3956 kstat_delete(arc_ksp);
3960 mutex_enter(&arc_prune_mtx);
3961 while ((p = list_head(&arc_prune_list)) != NULL) {
3962 list_remove(&arc_prune_list, p);
3963 refcount_remove(&p->p_refcnt, &arc_prune_list);
3964 refcount_destroy(&p->p_refcnt);
3965 kmem_free(p, sizeof (*p));
3967 mutex_exit(&arc_prune_mtx);
3969 list_destroy(&arc_prune_list);
3970 mutex_destroy(&arc_prune_mtx);
3971 mutex_destroy(&arc_eviction_mtx);
3972 mutex_destroy(&arc_reclaim_thr_lock);
3973 cv_destroy(&arc_reclaim_thr_cv);
3975 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3976 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3977 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3978 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3979 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3980 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3981 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3982 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3984 mutex_destroy(&arc_anon->arcs_mtx);
3985 mutex_destroy(&arc_mru->arcs_mtx);
3986 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3987 mutex_destroy(&arc_mfu->arcs_mtx);
3988 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3989 mutex_destroy(&arc_l2c_only->arcs_mtx);
3991 mutex_destroy(&zfs_write_limit_lock);
3995 ASSERT(arc_loaned_bytes == 0);
4001 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
4002 * It uses dedicated storage devices to hold cached data, which are populated
4003 * using large infrequent writes. The main role of this cache is to boost
4004 * the performance of random read workloads. The intended L2ARC devices
4005 * include short-stroked disks, solid state disks, and other media with
4006 * substantially faster read latency than disk.
4008 * +-----------------------+
4010 * +-----------------------+
4013 * l2arc_feed_thread() arc_read()
4017 * +---------------+ |
4019 * +---------------+ |
4024 * +-------+ +-------+
4026 * | cache | | cache |
4027 * +-------+ +-------+
4028 * +=========+ .-----.
4029 * : L2ARC : |-_____-|
4030 * : devices : | Disks |
4031 * +=========+ `-_____-'
4033 * Read requests are satisfied from the following sources, in order:
4036 * 2) vdev cache of L2ARC devices
4038 * 4) vdev cache of disks
4041 * Some L2ARC device types exhibit extremely slow write performance.
4042 * To accommodate for this there are some significant differences between
4043 * the L2ARC and traditional cache design:
4045 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4046 * the ARC behave as usual, freeing buffers and placing headers on ghost
4047 * lists. The ARC does not send buffers to the L2ARC during eviction as
4048 * this would add inflated write latencies for all ARC memory pressure.
4050 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4051 * It does this by periodically scanning buffers from the eviction-end of
4052 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4053 * not already there. It scans until a headroom of buffers is satisfied,
4054 * which itself is a buffer for ARC eviction. The thread that does this is
4055 * l2arc_feed_thread(), illustrated below; example sizes are included to
4056 * provide a better sense of ratio than this diagram:
4059 * +---------------------+----------+
4060 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4061 * +---------------------+----------+ | o L2ARC eligible
4062 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4063 * +---------------------+----------+ |
4064 * 15.9 Gbytes ^ 32 Mbytes |
4066 * l2arc_feed_thread()
4068 * l2arc write hand <--[oooo]--'
4072 * +==============================+
4073 * L2ARC dev |####|#|###|###| |####| ... |
4074 * +==============================+
4077 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4078 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4079 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4080 * safe to say that this is an uncommon case, since buffers at the end of
4081 * the ARC lists have moved there due to inactivity.
4083 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4084 * then the L2ARC simply misses copying some buffers. This serves as a
4085 * pressure valve to prevent heavy read workloads from both stalling the ARC
4086 * with waits and clogging the L2ARC with writes. This also helps prevent
4087 * the potential for the L2ARC to churn if it attempts to cache content too
4088 * quickly, such as during backups of the entire pool.
4090 * 5. After system boot and before the ARC has filled main memory, there are
4091 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4092 * lists can remain mostly static. Instead of searching from tail of these
4093 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4094 * for eligible buffers, greatly increasing its chance of finding them.
4096 * The L2ARC device write speed is also boosted during this time so that
4097 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4098 * there are no L2ARC reads, and no fear of degrading read performance
4099 * through increased writes.
4101 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4102 * the vdev queue can aggregate them into larger and fewer writes. Each
4103 * device is written to in a rotor fashion, sweeping writes through
4104 * available space then repeating.
4106 * 7. The L2ARC does not store dirty content. It never needs to flush
4107 * write buffers back to disk based storage.
4109 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4110 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4112 * The performance of the L2ARC can be tweaked by a number of tunables, which
4113 * may be necessary for different workloads:
4115 * l2arc_write_max max write bytes per interval
4116 * l2arc_write_boost extra write bytes during device warmup
4117 * l2arc_noprefetch skip caching prefetched buffers
4118 * l2arc_headroom number of max device writes to precache
4119 * l2arc_feed_secs seconds between L2ARC writing
4121 * Tunables may be removed or added as future performance improvements are
4122 * integrated, and also may become zpool properties.
4124 * There are three key functions that control how the L2ARC warms up:
4126 * l2arc_write_eligible() check if a buffer is eligible to cache
4127 * l2arc_write_size() calculate how much to write
4128 * l2arc_write_interval() calculate sleep delay between writes
4130 * These three functions determine what to write, how much, and how quickly
4135 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4138 * A buffer is *not* eligible for the L2ARC if it:
4139 * 1. belongs to a different spa.
4140 * 2. is already cached on the L2ARC.
4141 * 3. has an I/O in progress (it may be an incomplete read).
4142 * 4. is flagged not eligible (zfs property).
4144 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
4145 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
4152 l2arc_write_size(l2arc_dev_t *dev)
4156 size = dev->l2ad_write;
4158 if (arc_warm == B_FALSE)
4159 size += dev->l2ad_boost;
4166 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4168 clock_t interval, next, now;
4171 * If the ARC lists are busy, increase our write rate; if the
4172 * lists are stale, idle back. This is achieved by checking
4173 * how much we previously wrote - if it was more than half of
4174 * what we wanted, schedule the next write much sooner.
4176 if (l2arc_feed_again && wrote > (wanted / 2))
4177 interval = (hz * l2arc_feed_min_ms) / 1000;
4179 interval = hz * l2arc_feed_secs;
4181 now = ddi_get_lbolt();
4182 next = MAX(now, MIN(now + interval, began + interval));
4188 l2arc_hdr_stat_add(void)
4190 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4191 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4195 l2arc_hdr_stat_remove(void)
4197 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4198 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4202 * Cycle through L2ARC devices. This is how L2ARC load balances.
4203 * If a device is returned, this also returns holding the spa config lock.
4205 static l2arc_dev_t *
4206 l2arc_dev_get_next(void)
4208 l2arc_dev_t *first, *next = NULL;
4211 * Lock out the removal of spas (spa_namespace_lock), then removal
4212 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4213 * both locks will be dropped and a spa config lock held instead.
4215 mutex_enter(&spa_namespace_lock);
4216 mutex_enter(&l2arc_dev_mtx);
4218 /* if there are no vdevs, there is nothing to do */
4219 if (l2arc_ndev == 0)
4223 next = l2arc_dev_last;
4225 /* loop around the list looking for a non-faulted vdev */
4227 next = list_head(l2arc_dev_list);
4229 next = list_next(l2arc_dev_list, next);
4231 next = list_head(l2arc_dev_list);
4234 /* if we have come back to the start, bail out */
4237 else if (next == first)
4240 } while (vdev_is_dead(next->l2ad_vdev));
4242 /* if we were unable to find any usable vdevs, return NULL */
4243 if (vdev_is_dead(next->l2ad_vdev))
4246 l2arc_dev_last = next;
4249 mutex_exit(&l2arc_dev_mtx);
4252 * Grab the config lock to prevent the 'next' device from being
4253 * removed while we are writing to it.
4256 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4257 mutex_exit(&spa_namespace_lock);
4263 * Free buffers that were tagged for destruction.
4266 l2arc_do_free_on_write(void)
4269 l2arc_data_free_t *df, *df_prev;
4271 mutex_enter(&l2arc_free_on_write_mtx);
4272 buflist = l2arc_free_on_write;
4274 for (df = list_tail(buflist); df; df = df_prev) {
4275 df_prev = list_prev(buflist, df);
4276 ASSERT(df->l2df_data != NULL);
4277 ASSERT(df->l2df_func != NULL);
4278 df->l2df_func(df->l2df_data, df->l2df_size);
4279 list_remove(buflist, df);
4280 kmem_free(df, sizeof (l2arc_data_free_t));
4283 mutex_exit(&l2arc_free_on_write_mtx);
4287 * A write to a cache device has completed. Update all headers to allow
4288 * reads from these buffers to begin.
4291 l2arc_write_done(zio_t *zio)
4293 l2arc_write_callback_t *cb;
4296 arc_buf_hdr_t *head, *ab, *ab_prev;
4297 l2arc_buf_hdr_t *abl2;
4298 kmutex_t *hash_lock;
4300 cb = zio->io_private;
4302 dev = cb->l2wcb_dev;
4303 ASSERT(dev != NULL);
4304 head = cb->l2wcb_head;
4305 ASSERT(head != NULL);
4306 buflist = dev->l2ad_buflist;
4307 ASSERT(buflist != NULL);
4308 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4309 l2arc_write_callback_t *, cb);
4311 if (zio->io_error != 0)
4312 ARCSTAT_BUMP(arcstat_l2_writes_error);
4314 mutex_enter(&l2arc_buflist_mtx);
4317 * All writes completed, or an error was hit.
4319 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4320 ab_prev = list_prev(buflist, ab);
4322 hash_lock = HDR_LOCK(ab);
4323 if (!mutex_tryenter(hash_lock)) {
4325 * This buffer misses out. It may be in a stage
4326 * of eviction. Its ARC_L2_WRITING flag will be
4327 * left set, denying reads to this buffer.
4329 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4333 if (zio->io_error != 0) {
4335 * Error - drop L2ARC entry.
4337 list_remove(buflist, ab);
4340 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4341 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4345 * Allow ARC to begin reads to this L2ARC entry.
4347 ab->b_flags &= ~ARC_L2_WRITING;
4349 mutex_exit(hash_lock);
4352 atomic_inc_64(&l2arc_writes_done);
4353 list_remove(buflist, head);
4354 kmem_cache_free(hdr_cache, head);
4355 mutex_exit(&l2arc_buflist_mtx);
4357 l2arc_do_free_on_write();
4359 kmem_free(cb, sizeof (l2arc_write_callback_t));
4363 * A read to a cache device completed. Validate buffer contents before
4364 * handing over to the regular ARC routines.
4367 l2arc_read_done(zio_t *zio)
4369 l2arc_read_callback_t *cb;
4372 kmutex_t *hash_lock;
4375 ASSERT(zio->io_vd != NULL);
4376 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4378 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4380 cb = zio->io_private;
4382 buf = cb->l2rcb_buf;
4383 ASSERT(buf != NULL);
4385 hash_lock = HDR_LOCK(buf->b_hdr);
4386 mutex_enter(hash_lock);
4388 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4391 * Check this survived the L2ARC journey.
4393 equal = arc_cksum_equal(buf);
4394 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4395 mutex_exit(hash_lock);
4396 zio->io_private = buf;
4397 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4398 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4401 mutex_exit(hash_lock);
4403 * Buffer didn't survive caching. Increment stats and
4404 * reissue to the original storage device.
4406 if (zio->io_error != 0) {
4407 ARCSTAT_BUMP(arcstat_l2_io_error);
4409 zio->io_error = EIO;
4412 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4415 * If there's no waiter, issue an async i/o to the primary
4416 * storage now. If there *is* a waiter, the caller must
4417 * issue the i/o in a context where it's OK to block.
4419 if (zio->io_waiter == NULL) {
4420 zio_t *pio = zio_unique_parent(zio);
4422 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4424 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4425 buf->b_data, zio->io_size, arc_read_done, buf,
4426 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4430 kmem_free(cb, sizeof (l2arc_read_callback_t));
4434 * This is the list priority from which the L2ARC will search for pages to
4435 * cache. This is used within loops (0..3) to cycle through lists in the
4436 * desired order. This order can have a significant effect on cache
4439 * Currently the metadata lists are hit first, MFU then MRU, followed by
4440 * the data lists. This function returns a locked list, and also returns
4444 l2arc_list_locked(int list_num, kmutex_t **lock)
4446 list_t *list = NULL;
4448 ASSERT(list_num >= 0 && list_num <= 3);
4452 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4453 *lock = &arc_mfu->arcs_mtx;
4456 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4457 *lock = &arc_mru->arcs_mtx;
4460 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4461 *lock = &arc_mfu->arcs_mtx;
4464 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4465 *lock = &arc_mru->arcs_mtx;
4469 ASSERT(!(MUTEX_HELD(*lock)));
4475 * Evict buffers from the device write hand to the distance specified in
4476 * bytes. This distance may span populated buffers, it may span nothing.
4477 * This is clearing a region on the L2ARC device ready for writing.
4478 * If the 'all' boolean is set, every buffer is evicted.
4481 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4484 l2arc_buf_hdr_t *abl2;
4485 arc_buf_hdr_t *ab, *ab_prev;
4486 kmutex_t *hash_lock;
4489 buflist = dev->l2ad_buflist;
4491 if (buflist == NULL)
4494 if (!all && dev->l2ad_first) {
4496 * This is the first sweep through the device. There is
4502 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4504 * When nearing the end of the device, evict to the end
4505 * before the device write hand jumps to the start.
4507 taddr = dev->l2ad_end;
4509 taddr = dev->l2ad_hand + distance;
4511 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4512 uint64_t, taddr, boolean_t, all);
4515 mutex_enter(&l2arc_buflist_mtx);
4516 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4517 ab_prev = list_prev(buflist, ab);
4519 hash_lock = HDR_LOCK(ab);
4520 if (!mutex_tryenter(hash_lock)) {
4522 * Missed the hash lock. Retry.
4524 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4525 mutex_exit(&l2arc_buflist_mtx);
4526 mutex_enter(hash_lock);
4527 mutex_exit(hash_lock);
4531 if (HDR_L2_WRITE_HEAD(ab)) {
4533 * We hit a write head node. Leave it for
4534 * l2arc_write_done().
4536 list_remove(buflist, ab);
4537 mutex_exit(hash_lock);
4541 if (!all && ab->b_l2hdr != NULL &&
4542 (ab->b_l2hdr->b_daddr > taddr ||
4543 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4545 * We've evicted to the target address,
4546 * or the end of the device.
4548 mutex_exit(hash_lock);
4552 if (HDR_FREE_IN_PROGRESS(ab)) {
4554 * Already on the path to destruction.
4556 mutex_exit(hash_lock);
4560 if (ab->b_state == arc_l2c_only) {
4561 ASSERT(!HDR_L2_READING(ab));
4563 * This doesn't exist in the ARC. Destroy.
4564 * arc_hdr_destroy() will call list_remove()
4565 * and decrement arcstat_l2_size.
4567 arc_change_state(arc_anon, ab, hash_lock);
4568 arc_hdr_destroy(ab);
4571 * Invalidate issued or about to be issued
4572 * reads, since we may be about to write
4573 * over this location.
4575 if (HDR_L2_READING(ab)) {
4576 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4577 ab->b_flags |= ARC_L2_EVICTED;
4581 * Tell ARC this no longer exists in L2ARC.
4583 if (ab->b_l2hdr != NULL) {
4586 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4587 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4589 list_remove(buflist, ab);
4592 * This may have been leftover after a
4595 ab->b_flags &= ~ARC_L2_WRITING;
4597 mutex_exit(hash_lock);
4599 mutex_exit(&l2arc_buflist_mtx);
4601 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4602 dev->l2ad_evict = taddr;
4606 * Find and write ARC buffers to the L2ARC device.
4608 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4609 * for reading until they have completed writing.
4612 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4614 arc_buf_hdr_t *ab, *ab_prev, *head;
4615 l2arc_buf_hdr_t *hdrl2;
4617 uint64_t passed_sz, write_sz, buf_sz, headroom;
4619 kmutex_t *hash_lock, *list_lock = NULL;
4620 boolean_t have_lock, full;
4621 l2arc_write_callback_t *cb;
4623 uint64_t guid = spa_load_guid(spa);
4626 ASSERT(dev->l2ad_vdev != NULL);
4631 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4632 head->b_flags |= ARC_L2_WRITE_HEAD;
4635 * Copy buffers for L2ARC writing.
4637 mutex_enter(&l2arc_buflist_mtx);
4638 for (try = 0; try <= 3; try++) {
4639 list = l2arc_list_locked(try, &list_lock);
4643 * L2ARC fast warmup.
4645 * Until the ARC is warm and starts to evict, read from the
4646 * head of the ARC lists rather than the tail.
4648 headroom = target_sz * l2arc_headroom;
4649 if (arc_warm == B_FALSE)
4650 ab = list_head(list);
4652 ab = list_tail(list);
4654 for (; ab; ab = ab_prev) {
4655 if (arc_warm == B_FALSE)
4656 ab_prev = list_next(list, ab);
4658 ab_prev = list_prev(list, ab);
4660 hash_lock = HDR_LOCK(ab);
4661 have_lock = MUTEX_HELD(hash_lock);
4662 if (!have_lock && !mutex_tryenter(hash_lock)) {
4664 * Skip this buffer rather than waiting.
4669 passed_sz += ab->b_size;
4670 if (passed_sz > headroom) {
4674 mutex_exit(hash_lock);
4678 if (!l2arc_write_eligible(guid, ab)) {
4679 mutex_exit(hash_lock);
4683 if ((write_sz + ab->b_size) > target_sz) {
4685 mutex_exit(hash_lock);
4691 * Insert a dummy header on the buflist so
4692 * l2arc_write_done() can find where the
4693 * write buffers begin without searching.
4695 list_insert_head(dev->l2ad_buflist, head);
4697 cb = kmem_alloc(sizeof (l2arc_write_callback_t),
4699 cb->l2wcb_dev = dev;
4700 cb->l2wcb_head = head;
4701 pio = zio_root(spa, l2arc_write_done, cb,
4706 * Create and add a new L2ARC header.
4708 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t),
4711 hdrl2->b_daddr = dev->l2ad_hand;
4713 ab->b_flags |= ARC_L2_WRITING;
4714 ab->b_l2hdr = hdrl2;
4715 list_insert_head(dev->l2ad_buflist, ab);
4716 buf_data = ab->b_buf->b_data;
4717 buf_sz = ab->b_size;
4720 * Compute and store the buffer cksum before
4721 * writing. On debug the cksum is verified first.
4723 arc_cksum_verify(ab->b_buf);
4724 arc_cksum_compute(ab->b_buf, B_TRUE);
4726 mutex_exit(hash_lock);
4728 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4729 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4730 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4731 ZIO_FLAG_CANFAIL, B_FALSE);
4733 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4735 (void) zio_nowait(wzio);
4738 * Keep the clock hand suitably device-aligned.
4740 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4743 dev->l2ad_hand += buf_sz;
4746 mutex_exit(list_lock);
4751 mutex_exit(&l2arc_buflist_mtx);
4755 kmem_cache_free(hdr_cache, head);
4759 ASSERT3U(write_sz, <=, target_sz);
4760 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4761 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4762 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4763 vdev_space_update(dev->l2ad_vdev, write_sz, 0, 0);
4766 * Bump device hand to the device start if it is approaching the end.
4767 * l2arc_evict() will already have evicted ahead for this case.
4769 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4770 vdev_space_update(dev->l2ad_vdev,
4771 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4772 dev->l2ad_hand = dev->l2ad_start;
4773 dev->l2ad_evict = dev->l2ad_start;
4774 dev->l2ad_first = B_FALSE;
4777 dev->l2ad_writing = B_TRUE;
4778 (void) zio_wait(pio);
4779 dev->l2ad_writing = B_FALSE;
4785 * This thread feeds the L2ARC at regular intervals. This is the beating
4786 * heart of the L2ARC.
4789 l2arc_feed_thread(void)
4794 uint64_t size, wrote;
4795 clock_t begin, next = ddi_get_lbolt();
4797 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4799 mutex_enter(&l2arc_feed_thr_lock);
4801 while (l2arc_thread_exit == 0) {
4802 CALLB_CPR_SAFE_BEGIN(&cpr);
4803 (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv,
4804 &l2arc_feed_thr_lock, next);
4805 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4806 next = ddi_get_lbolt() + hz;
4809 * Quick check for L2ARC devices.
4811 mutex_enter(&l2arc_dev_mtx);
4812 if (l2arc_ndev == 0) {
4813 mutex_exit(&l2arc_dev_mtx);
4816 mutex_exit(&l2arc_dev_mtx);
4817 begin = ddi_get_lbolt();
4820 * This selects the next l2arc device to write to, and in
4821 * doing so the next spa to feed from: dev->l2ad_spa. This
4822 * will return NULL if there are now no l2arc devices or if
4823 * they are all faulted.
4825 * If a device is returned, its spa's config lock is also
4826 * held to prevent device removal. l2arc_dev_get_next()
4827 * will grab and release l2arc_dev_mtx.
4829 if ((dev = l2arc_dev_get_next()) == NULL)
4832 spa = dev->l2ad_spa;
4833 ASSERT(spa != NULL);
4836 * If the pool is read-only then force the feed thread to
4837 * sleep a little longer.
4839 if (!spa_writeable(spa)) {
4840 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4841 spa_config_exit(spa, SCL_L2ARC, dev);
4846 * Avoid contributing to memory pressure.
4849 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4850 spa_config_exit(spa, SCL_L2ARC, dev);
4854 ARCSTAT_BUMP(arcstat_l2_feeds);
4856 size = l2arc_write_size(dev);
4859 * Evict L2ARC buffers that will be overwritten.
4861 l2arc_evict(dev, size, B_FALSE);
4864 * Write ARC buffers.
4866 wrote = l2arc_write_buffers(spa, dev, size);
4869 * Calculate interval between writes.
4871 next = l2arc_write_interval(begin, size, wrote);
4872 spa_config_exit(spa, SCL_L2ARC, dev);
4875 l2arc_thread_exit = 0;
4876 cv_broadcast(&l2arc_feed_thr_cv);
4877 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4882 l2arc_vdev_present(vdev_t *vd)
4886 mutex_enter(&l2arc_dev_mtx);
4887 for (dev = list_head(l2arc_dev_list); dev != NULL;
4888 dev = list_next(l2arc_dev_list, dev)) {
4889 if (dev->l2ad_vdev == vd)
4892 mutex_exit(&l2arc_dev_mtx);
4894 return (dev != NULL);
4898 * Add a vdev for use by the L2ARC. By this point the spa has already
4899 * validated the vdev and opened it.
4902 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
4904 l2arc_dev_t *adddev;
4906 ASSERT(!l2arc_vdev_present(vd));
4909 * Create a new l2arc device entry.
4911 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4912 adddev->l2ad_spa = spa;
4913 adddev->l2ad_vdev = vd;
4914 adddev->l2ad_write = l2arc_write_max;
4915 adddev->l2ad_boost = l2arc_write_boost;
4916 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
4917 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
4918 adddev->l2ad_hand = adddev->l2ad_start;
4919 adddev->l2ad_evict = adddev->l2ad_start;
4920 adddev->l2ad_first = B_TRUE;
4921 adddev->l2ad_writing = B_FALSE;
4922 list_link_init(&adddev->l2ad_node);
4923 ASSERT3U(adddev->l2ad_write, >, 0);
4926 * This is a list of all ARC buffers that are still valid on the
4929 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4930 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4931 offsetof(arc_buf_hdr_t, b_l2node));
4933 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
4936 * Add device to global list
4938 mutex_enter(&l2arc_dev_mtx);
4939 list_insert_head(l2arc_dev_list, adddev);
4940 atomic_inc_64(&l2arc_ndev);
4941 mutex_exit(&l2arc_dev_mtx);
4945 * Remove a vdev from the L2ARC.
4948 l2arc_remove_vdev(vdev_t *vd)
4950 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4953 * Find the device by vdev
4955 mutex_enter(&l2arc_dev_mtx);
4956 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4957 nextdev = list_next(l2arc_dev_list, dev);
4958 if (vd == dev->l2ad_vdev) {
4963 ASSERT(remdev != NULL);
4966 * Remove device from global list
4968 list_remove(l2arc_dev_list, remdev);
4969 l2arc_dev_last = NULL; /* may have been invalidated */
4970 atomic_dec_64(&l2arc_ndev);
4971 mutex_exit(&l2arc_dev_mtx);
4974 * Clear all buflists and ARC references. L2ARC device flush.
4976 l2arc_evict(remdev, 0, B_TRUE);
4977 list_destroy(remdev->l2ad_buflist);
4978 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
4979 kmem_free(remdev, sizeof (l2arc_dev_t));
4985 l2arc_thread_exit = 0;
4987 l2arc_writes_sent = 0;
4988 l2arc_writes_done = 0;
4990 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4991 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
4992 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
4993 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
4994 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
4996 l2arc_dev_list = &L2ARC_dev_list;
4997 l2arc_free_on_write = &L2ARC_free_on_write;
4998 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
4999 offsetof(l2arc_dev_t, l2ad_node));
5000 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5001 offsetof(l2arc_data_free_t, l2df_list_node));
5008 * This is called from dmu_fini(), which is called from spa_fini();
5009 * Because of this, we can assume that all l2arc devices have
5010 * already been removed when the pools themselves were removed.
5013 l2arc_do_free_on_write();
5015 mutex_destroy(&l2arc_feed_thr_lock);
5016 cv_destroy(&l2arc_feed_thr_cv);
5017 mutex_destroy(&l2arc_dev_mtx);
5018 mutex_destroy(&l2arc_buflist_mtx);
5019 mutex_destroy(&l2arc_free_on_write_mtx);
5021 list_destroy(l2arc_dev_list);
5022 list_destroy(l2arc_free_on_write);
5028 if (!(spa_mode_global & FWRITE))
5031 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5032 TS_RUN, minclsyspri);
5038 if (!(spa_mode_global & FWRITE))
5041 mutex_enter(&l2arc_feed_thr_lock);
5042 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5043 l2arc_thread_exit = 1;
5044 while (l2arc_thread_exit != 0)
5045 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5046 mutex_exit(&l2arc_feed_thr_lock);
5049 #if defined(_KERNEL) && defined(HAVE_SPL)
5050 EXPORT_SYMBOL(arc_read);
5051 EXPORT_SYMBOL(arc_buf_remove_ref);
5052 EXPORT_SYMBOL(arc_getbuf_func);
5053 EXPORT_SYMBOL(arc_add_prune_callback);
5054 EXPORT_SYMBOL(arc_remove_prune_callback);
5056 module_param(zfs_arc_min, ulong, 0444);
5057 MODULE_PARM_DESC(zfs_arc_min, "Min arc size");
5059 module_param(zfs_arc_max, ulong, 0444);
5060 MODULE_PARM_DESC(zfs_arc_max, "Max arc size");
5062 module_param(zfs_arc_meta_limit, ulong, 0444);
5063 MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size");
5065 module_param(zfs_arc_meta_prune, int, 0444);
5066 MODULE_PARM_DESC(zfs_arc_meta_prune, "Bytes of meta data to prune");
5068 module_param(zfs_arc_grow_retry, int, 0444);
5069 MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size");
5071 module_param(zfs_arc_shrink_shift, int, 0444);
5072 MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)");
5074 module_param(zfs_arc_p_min_shift, int, 0444);
5075 MODULE_PARM_DESC(zfs_arc_p_min_shift, "arc_c shift to calc min/max arc_p");
5077 module_param(zfs_disable_dup_eviction, int, 0644);
5078 MODULE_PARM_DESC(zfs_disable_dup_eviction, "disable duplicate buffer eviction");
5080 module_param(zfs_arc_memory_throttle_disable, int, 0644);
5081 MODULE_PARM_DESC(zfs_arc_memory_throttle_disable, "disable memory throttle");
5083 module_param(l2arc_write_max, ulong, 0444);
5084 MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval");
5086 module_param(l2arc_write_boost, ulong, 0444);
5087 MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup");
5089 module_param(l2arc_headroom, ulong, 0444);
5090 MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache");
5092 module_param(l2arc_feed_secs, ulong, 0444);
5093 MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing");
5095 module_param(l2arc_feed_min_ms, ulong, 0444);
5096 MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds");
5098 module_param(l2arc_noprefetch, int, 0444);
5099 MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers");
5101 module_param(l2arc_feed_again, int, 0444);
5102 MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup");
5104 module_param(l2arc_norw, int, 0444);
5105 MODULE_PARM_DESC(l2arc_norw, "No reads during writes");