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 int zfs_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 int zfs_arc_grow_retry = 5;
161 /* shift of arc_c for calculating both min and max arc_p */
162 int zfs_arc_p_min_shift = 4;
164 /* log2(fraction of arc to reclaim) */
165 int zfs_arc_shrink_shift = 5;
168 * minimum lifespan of a prefetch block in clock ticks
169 * (initialized in arc_init())
171 int zfs_arc_min_prefetch_lifespan = HZ;
173 /* disable arc proactive arc throttle due to low memory */
174 int zfs_arc_memory_throttle_disable = 1;
176 /* disable duplicate buffer eviction */
177 int zfs_disable_dup_eviction = 0;
181 /* expiration time for arc_no_grow */
182 static clock_t arc_grow_time = 0;
185 * The arc has filled available memory and has now warmed up.
187 static boolean_t arc_warm;
190 * These tunables are for performance analysis.
192 unsigned long zfs_arc_max = 0;
193 unsigned long zfs_arc_min = 0;
194 unsigned long zfs_arc_meta_limit = 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;
505 arc_buf_hdr_t *b_hash_next;
510 arc_callback_t *b_acb;
514 arc_buf_contents_t b_type;
518 /* protected by arc state mutex */
519 arc_state_t *b_state;
520 list_node_t b_arc_node;
522 /* updated atomically */
523 clock_t b_arc_access;
525 /* self protecting */
528 l2arc_buf_hdr_t *b_l2hdr;
529 list_node_t b_l2node;
532 static list_t arc_prune_list;
533 static kmutex_t arc_prune_mtx;
534 static arc_buf_t *arc_eviction_list;
535 static kmutex_t arc_eviction_mtx;
536 static arc_buf_hdr_t arc_eviction_hdr;
537 static void arc_get_data_buf(arc_buf_t *buf);
538 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
539 static int arc_evict_needed(arc_buf_contents_t type);
540 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
542 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
544 #define GHOST_STATE(state) \
545 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
546 (state) == arc_l2c_only)
549 * Private ARC flags. These flags are private ARC only flags that will show up
550 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
551 * be passed in as arc_flags in things like arc_read. However, these flags
552 * should never be passed and should only be set by ARC code. When adding new
553 * public flags, make sure not to smash the private ones.
556 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
557 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
558 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
559 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
560 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
561 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
562 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
563 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
564 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
565 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
567 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
568 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
569 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
570 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
571 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
572 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
573 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
574 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
575 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
576 (hdr)->b_l2hdr != NULL)
577 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
578 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
579 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
585 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
586 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
589 * Hash table routines
592 #define HT_LOCK_ALIGN 64
593 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
598 unsigned char pad[HT_LOCK_PAD];
602 #define BUF_LOCKS 256
603 typedef struct buf_hash_table {
605 arc_buf_hdr_t **ht_table;
606 struct ht_lock ht_locks[BUF_LOCKS];
609 static buf_hash_table_t buf_hash_table;
611 #define BUF_HASH_INDEX(spa, dva, birth) \
612 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
613 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
614 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
615 #define HDR_LOCK(hdr) \
616 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
618 uint64_t zfs_crc64_table[256];
624 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
625 #define L2ARC_HEADROOM 2 /* num of writes */
626 #define L2ARC_FEED_SECS 1 /* caching interval secs */
627 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
629 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
630 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
633 * L2ARC Performance Tunables
635 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
636 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
637 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
638 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
639 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
640 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
641 int l2arc_feed_again = B_TRUE; /* turbo warmup */
642 int l2arc_norw = B_FALSE; /* no reads during writes */
647 typedef struct l2arc_dev {
648 vdev_t *l2ad_vdev; /* vdev */
649 spa_t *l2ad_spa; /* spa */
650 uint64_t l2ad_hand; /* next write location */
651 uint64_t l2ad_write; /* desired write size, bytes */
652 uint64_t l2ad_boost; /* warmup write boost, bytes */
653 uint64_t l2ad_start; /* first addr on device */
654 uint64_t l2ad_end; /* last addr on device */
655 uint64_t l2ad_evict; /* last addr eviction reached */
656 boolean_t l2ad_first; /* first sweep through */
657 boolean_t l2ad_writing; /* currently writing */
658 list_t *l2ad_buflist; /* buffer list */
659 list_node_t l2ad_node; /* device list node */
662 static list_t L2ARC_dev_list; /* device list */
663 static list_t *l2arc_dev_list; /* device list pointer */
664 static kmutex_t l2arc_dev_mtx; /* device list mutex */
665 static l2arc_dev_t *l2arc_dev_last; /* last device used */
666 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
667 static list_t L2ARC_free_on_write; /* free after write buf list */
668 static list_t *l2arc_free_on_write; /* free after write list ptr */
669 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
670 static uint64_t l2arc_ndev; /* number of devices */
672 typedef struct l2arc_read_callback {
673 arc_buf_t *l2rcb_buf; /* read buffer */
674 spa_t *l2rcb_spa; /* spa */
675 blkptr_t l2rcb_bp; /* original blkptr */
676 zbookmark_t l2rcb_zb; /* original bookmark */
677 int l2rcb_flags; /* original flags */
678 } l2arc_read_callback_t;
680 typedef struct l2arc_write_callback {
681 l2arc_dev_t *l2wcb_dev; /* device info */
682 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
683 } l2arc_write_callback_t;
685 struct l2arc_buf_hdr {
686 /* protected by arc_buf_hdr mutex */
687 l2arc_dev_t *b_dev; /* L2ARC device */
688 uint64_t b_daddr; /* disk address, offset byte */
691 typedef struct l2arc_data_free {
692 /* protected by l2arc_free_on_write_mtx */
695 void (*l2df_func)(void *, size_t);
696 list_node_t l2df_list_node;
699 static kmutex_t l2arc_feed_thr_lock;
700 static kcondvar_t l2arc_feed_thr_cv;
701 static uint8_t l2arc_thread_exit;
703 static void l2arc_read_done(zio_t *zio);
704 static void l2arc_hdr_stat_add(void);
705 static void l2arc_hdr_stat_remove(void);
708 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
710 uint8_t *vdva = (uint8_t *)dva;
711 uint64_t crc = -1ULL;
714 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
716 for (i = 0; i < sizeof (dva_t); i++)
717 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
719 crc ^= (spa>>8) ^ birth;
724 #define BUF_EMPTY(buf) \
725 ((buf)->b_dva.dva_word[0] == 0 && \
726 (buf)->b_dva.dva_word[1] == 0 && \
729 #define BUF_EQUAL(spa, dva, birth, buf) \
730 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
731 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
732 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
735 buf_discard_identity(arc_buf_hdr_t *hdr)
737 hdr->b_dva.dva_word[0] = 0;
738 hdr->b_dva.dva_word[1] = 0;
743 static arc_buf_hdr_t *
744 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
746 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
747 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
750 mutex_enter(hash_lock);
751 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
752 buf = buf->b_hash_next) {
753 if (BUF_EQUAL(spa, dva, birth, buf)) {
758 mutex_exit(hash_lock);
764 * Insert an entry into the hash table. If there is already an element
765 * equal to elem in the hash table, then the already existing element
766 * will be returned and the new element will not be inserted.
767 * Otherwise returns NULL.
769 static arc_buf_hdr_t *
770 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
772 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
773 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
777 ASSERT(!HDR_IN_HASH_TABLE(buf));
779 mutex_enter(hash_lock);
780 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
781 fbuf = fbuf->b_hash_next, i++) {
782 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
786 buf->b_hash_next = buf_hash_table.ht_table[idx];
787 buf_hash_table.ht_table[idx] = buf;
788 buf->b_flags |= ARC_IN_HASH_TABLE;
790 /* collect some hash table performance data */
792 ARCSTAT_BUMP(arcstat_hash_collisions);
794 ARCSTAT_BUMP(arcstat_hash_chains);
796 ARCSTAT_MAX(arcstat_hash_chain_max, i);
799 ARCSTAT_BUMP(arcstat_hash_elements);
800 ARCSTAT_MAXSTAT(arcstat_hash_elements);
806 buf_hash_remove(arc_buf_hdr_t *buf)
808 arc_buf_hdr_t *fbuf, **bufp;
809 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
811 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
812 ASSERT(HDR_IN_HASH_TABLE(buf));
814 bufp = &buf_hash_table.ht_table[idx];
815 while ((fbuf = *bufp) != buf) {
816 ASSERT(fbuf != NULL);
817 bufp = &fbuf->b_hash_next;
819 *bufp = buf->b_hash_next;
820 buf->b_hash_next = NULL;
821 buf->b_flags &= ~ARC_IN_HASH_TABLE;
823 /* collect some hash table performance data */
824 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
826 if (buf_hash_table.ht_table[idx] &&
827 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
828 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
832 * Global data structures and functions for the buf kmem cache.
834 static kmem_cache_t *hdr_cache;
835 static kmem_cache_t *buf_cache;
842 #if defined(_KERNEL) && defined(HAVE_SPL)
843 /* Large allocations which do not require contiguous pages
844 * should be using vmem_free() in the linux kernel */
845 vmem_free(buf_hash_table.ht_table,
846 (buf_hash_table.ht_mask + 1) * sizeof (void *));
848 kmem_free(buf_hash_table.ht_table,
849 (buf_hash_table.ht_mask + 1) * sizeof (void *));
851 for (i = 0; i < BUF_LOCKS; i++)
852 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
853 kmem_cache_destroy(hdr_cache);
854 kmem_cache_destroy(buf_cache);
858 * Constructor callback - called when the cache is empty
859 * and a new buf is requested.
863 hdr_cons(void *vbuf, void *unused, int kmflag)
865 arc_buf_hdr_t *buf = vbuf;
867 bzero(buf, sizeof (arc_buf_hdr_t));
868 refcount_create(&buf->b_refcnt);
869 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
870 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
871 list_link_init(&buf->b_arc_node);
872 list_link_init(&buf->b_l2node);
873 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
880 buf_cons(void *vbuf, void *unused, int kmflag)
882 arc_buf_t *buf = vbuf;
884 bzero(buf, sizeof (arc_buf_t));
885 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
886 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
892 * Destructor callback - called when a cached buf is
893 * no longer required.
897 hdr_dest(void *vbuf, void *unused)
899 arc_buf_hdr_t *buf = vbuf;
901 ASSERT(BUF_EMPTY(buf));
902 refcount_destroy(&buf->b_refcnt);
903 cv_destroy(&buf->b_cv);
904 mutex_destroy(&buf->b_freeze_lock);
905 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
910 buf_dest(void *vbuf, void *unused)
912 arc_buf_t *buf = vbuf;
914 mutex_destroy(&buf->b_evict_lock);
915 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
922 uint64_t hsize = 1ULL << 12;
926 * The hash table is big enough to fill all of physical memory
927 * with an average 64K block size. The table will take up
928 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
930 while (hsize * 65536 < physmem * PAGESIZE)
933 buf_hash_table.ht_mask = hsize - 1;
934 #if defined(_KERNEL) && defined(HAVE_SPL)
935 /* Large allocations which do not require contiguous pages
936 * should be using vmem_alloc() in the linux kernel */
937 buf_hash_table.ht_table =
938 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
940 buf_hash_table.ht_table =
941 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
943 if (buf_hash_table.ht_table == NULL) {
944 ASSERT(hsize > (1ULL << 8));
949 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
950 0, hdr_cons, hdr_dest, NULL, NULL, NULL, 0);
951 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
952 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
954 for (i = 0; i < 256; i++)
955 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
956 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
958 for (i = 0; i < BUF_LOCKS; i++) {
959 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
960 NULL, MUTEX_DEFAULT, NULL);
964 #define ARC_MINTIME (hz>>4) /* 62 ms */
967 arc_cksum_verify(arc_buf_t *buf)
971 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
974 mutex_enter(&buf->b_hdr->b_freeze_lock);
975 if (buf->b_hdr->b_freeze_cksum == NULL ||
976 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
977 mutex_exit(&buf->b_hdr->b_freeze_lock);
980 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
981 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
982 panic("buffer modified while frozen!");
983 mutex_exit(&buf->b_hdr->b_freeze_lock);
987 arc_cksum_equal(arc_buf_t *buf)
992 mutex_enter(&buf->b_hdr->b_freeze_lock);
993 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
994 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
995 mutex_exit(&buf->b_hdr->b_freeze_lock);
1001 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1003 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1006 mutex_enter(&buf->b_hdr->b_freeze_lock);
1007 if (buf->b_hdr->b_freeze_cksum != NULL) {
1008 mutex_exit(&buf->b_hdr->b_freeze_lock);
1011 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1013 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1014 buf->b_hdr->b_freeze_cksum);
1015 mutex_exit(&buf->b_hdr->b_freeze_lock);
1019 arc_buf_thaw(arc_buf_t *buf)
1021 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1022 if (buf->b_hdr->b_state != arc_anon)
1023 panic("modifying non-anon buffer!");
1024 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1025 panic("modifying buffer while i/o in progress!");
1026 arc_cksum_verify(buf);
1029 mutex_enter(&buf->b_hdr->b_freeze_lock);
1030 if (buf->b_hdr->b_freeze_cksum != NULL) {
1031 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1032 buf->b_hdr->b_freeze_cksum = NULL;
1035 mutex_exit(&buf->b_hdr->b_freeze_lock);
1039 arc_buf_freeze(arc_buf_t *buf)
1041 kmutex_t *hash_lock;
1043 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1046 hash_lock = HDR_LOCK(buf->b_hdr);
1047 mutex_enter(hash_lock);
1049 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1050 buf->b_hdr->b_state == arc_anon);
1051 arc_cksum_compute(buf, B_FALSE);
1052 mutex_exit(hash_lock);
1056 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1058 ASSERT(MUTEX_HELD(hash_lock));
1060 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1061 (ab->b_state != arc_anon)) {
1062 uint64_t delta = ab->b_size * ab->b_datacnt;
1063 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1064 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1066 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1067 mutex_enter(&ab->b_state->arcs_mtx);
1068 ASSERT(list_link_active(&ab->b_arc_node));
1069 list_remove(list, ab);
1070 if (GHOST_STATE(ab->b_state)) {
1071 ASSERT0(ab->b_datacnt);
1072 ASSERT3P(ab->b_buf, ==, NULL);
1076 ASSERT3U(*size, >=, delta);
1077 atomic_add_64(size, -delta);
1078 mutex_exit(&ab->b_state->arcs_mtx);
1079 /* remove the prefetch flag if we get a reference */
1080 if (ab->b_flags & ARC_PREFETCH)
1081 ab->b_flags &= ~ARC_PREFETCH;
1086 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1089 arc_state_t *state = ab->b_state;
1091 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1092 ASSERT(!GHOST_STATE(state));
1094 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1095 (state != arc_anon)) {
1096 uint64_t *size = &state->arcs_lsize[ab->b_type];
1098 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1099 mutex_enter(&state->arcs_mtx);
1100 ASSERT(!list_link_active(&ab->b_arc_node));
1101 list_insert_head(&state->arcs_list[ab->b_type], ab);
1102 ASSERT(ab->b_datacnt > 0);
1103 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1104 mutex_exit(&state->arcs_mtx);
1110 * Move the supplied buffer to the indicated state. The mutex
1111 * for the buffer must be held by the caller.
1114 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1116 arc_state_t *old_state = ab->b_state;
1117 int64_t refcnt = refcount_count(&ab->b_refcnt);
1118 uint64_t from_delta, to_delta;
1120 ASSERT(MUTEX_HELD(hash_lock));
1121 ASSERT(new_state != old_state);
1122 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1123 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1124 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1126 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1129 * If this buffer is evictable, transfer it from the
1130 * old state list to the new state list.
1133 if (old_state != arc_anon) {
1134 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1135 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1138 mutex_enter(&old_state->arcs_mtx);
1140 ASSERT(list_link_active(&ab->b_arc_node));
1141 list_remove(&old_state->arcs_list[ab->b_type], ab);
1144 * If prefetching out of the ghost cache,
1145 * we will have a non-zero datacnt.
1147 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1148 /* ghost elements have a ghost size */
1149 ASSERT(ab->b_buf == NULL);
1150 from_delta = ab->b_size;
1152 ASSERT3U(*size, >=, from_delta);
1153 atomic_add_64(size, -from_delta);
1156 mutex_exit(&old_state->arcs_mtx);
1158 if (new_state != arc_anon) {
1159 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1160 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1163 mutex_enter(&new_state->arcs_mtx);
1165 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1167 /* ghost elements have a ghost size */
1168 if (GHOST_STATE(new_state)) {
1169 ASSERT(ab->b_datacnt == 0);
1170 ASSERT(ab->b_buf == NULL);
1171 to_delta = ab->b_size;
1173 atomic_add_64(size, to_delta);
1176 mutex_exit(&new_state->arcs_mtx);
1180 ASSERT(!BUF_EMPTY(ab));
1181 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1182 buf_hash_remove(ab);
1184 /* adjust state sizes */
1186 atomic_add_64(&new_state->arcs_size, to_delta);
1188 ASSERT3U(old_state->arcs_size, >=, from_delta);
1189 atomic_add_64(&old_state->arcs_size, -from_delta);
1191 ab->b_state = new_state;
1193 /* adjust l2arc hdr stats */
1194 if (new_state == arc_l2c_only)
1195 l2arc_hdr_stat_add();
1196 else if (old_state == arc_l2c_only)
1197 l2arc_hdr_stat_remove();
1201 arc_space_consume(uint64_t space, arc_space_type_t type)
1203 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1208 case ARC_SPACE_DATA:
1209 ARCSTAT_INCR(arcstat_data_size, space);
1211 case ARC_SPACE_OTHER:
1212 ARCSTAT_INCR(arcstat_other_size, space);
1214 case ARC_SPACE_HDRS:
1215 ARCSTAT_INCR(arcstat_hdr_size, space);
1217 case ARC_SPACE_L2HDRS:
1218 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1222 atomic_add_64(&arc_meta_used, space);
1223 atomic_add_64(&arc_size, space);
1227 arc_space_return(uint64_t space, arc_space_type_t type)
1229 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1234 case ARC_SPACE_DATA:
1235 ARCSTAT_INCR(arcstat_data_size, -space);
1237 case ARC_SPACE_OTHER:
1238 ARCSTAT_INCR(arcstat_other_size, -space);
1240 case ARC_SPACE_HDRS:
1241 ARCSTAT_INCR(arcstat_hdr_size, -space);
1243 case ARC_SPACE_L2HDRS:
1244 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1248 ASSERT(arc_meta_used >= space);
1249 if (arc_meta_max < arc_meta_used)
1250 arc_meta_max = arc_meta_used;
1251 atomic_add_64(&arc_meta_used, -space);
1252 ASSERT(arc_size >= space);
1253 atomic_add_64(&arc_size, -space);
1257 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1262 ASSERT3U(size, >, 0);
1263 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1264 ASSERT(BUF_EMPTY(hdr));
1267 hdr->b_spa = spa_load_guid(spa);
1268 hdr->b_state = arc_anon;
1269 hdr->b_arc_access = 0;
1270 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1273 buf->b_efunc = NULL;
1274 buf->b_private = NULL;
1277 arc_get_data_buf(buf);
1280 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1281 (void) refcount_add(&hdr->b_refcnt, tag);
1286 static char *arc_onloan_tag = "onloan";
1289 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1290 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1291 * buffers must be returned to the arc before they can be used by the DMU or
1295 arc_loan_buf(spa_t *spa, int size)
1299 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1301 atomic_add_64(&arc_loaned_bytes, size);
1306 * Return a loaned arc buffer to the arc.
1309 arc_return_buf(arc_buf_t *buf, void *tag)
1311 arc_buf_hdr_t *hdr = buf->b_hdr;
1313 ASSERT(buf->b_data != NULL);
1314 (void) refcount_add(&hdr->b_refcnt, tag);
1315 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1317 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1320 /* Detach an arc_buf from a dbuf (tag) */
1322 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1326 ASSERT(buf->b_data != NULL);
1328 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1329 (void) refcount_remove(&hdr->b_refcnt, tag);
1330 buf->b_efunc = NULL;
1331 buf->b_private = NULL;
1333 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1337 arc_buf_clone(arc_buf_t *from)
1340 arc_buf_hdr_t *hdr = from->b_hdr;
1341 uint64_t size = hdr->b_size;
1343 ASSERT(hdr->b_state != arc_anon);
1345 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1348 buf->b_efunc = NULL;
1349 buf->b_private = NULL;
1350 buf->b_next = hdr->b_buf;
1352 arc_get_data_buf(buf);
1353 bcopy(from->b_data, buf->b_data, size);
1356 * This buffer already exists in the arc so create a duplicate
1357 * copy for the caller. If the buffer is associated with user data
1358 * then track the size and number of duplicates. These stats will be
1359 * updated as duplicate buffers are created and destroyed.
1361 if (hdr->b_type == ARC_BUFC_DATA) {
1362 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1363 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1365 hdr->b_datacnt += 1;
1370 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1373 kmutex_t *hash_lock;
1376 * Check to see if this buffer is evicted. Callers
1377 * must verify b_data != NULL to know if the add_ref
1380 mutex_enter(&buf->b_evict_lock);
1381 if (buf->b_data == NULL) {
1382 mutex_exit(&buf->b_evict_lock);
1385 hash_lock = HDR_LOCK(buf->b_hdr);
1386 mutex_enter(hash_lock);
1388 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1389 mutex_exit(&buf->b_evict_lock);
1391 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1392 add_reference(hdr, hash_lock, tag);
1393 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1394 arc_access(hdr, hash_lock);
1395 mutex_exit(hash_lock);
1396 ARCSTAT_BUMP(arcstat_hits);
1397 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1398 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1399 data, metadata, hits);
1403 * Free the arc data buffer. If it is an l2arc write in progress,
1404 * the buffer is placed on l2arc_free_on_write to be freed later.
1407 arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t),
1408 void *data, size_t size)
1410 if (HDR_L2_WRITING(hdr)) {
1411 l2arc_data_free_t *df;
1412 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_PUSHPAGE);
1413 df->l2df_data = data;
1414 df->l2df_size = size;
1415 df->l2df_func = free_func;
1416 mutex_enter(&l2arc_free_on_write_mtx);
1417 list_insert_head(l2arc_free_on_write, df);
1418 mutex_exit(&l2arc_free_on_write_mtx);
1419 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1421 free_func(data, size);
1426 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1430 /* free up data associated with the buf */
1432 arc_state_t *state = buf->b_hdr->b_state;
1433 uint64_t size = buf->b_hdr->b_size;
1434 arc_buf_contents_t type = buf->b_hdr->b_type;
1436 arc_cksum_verify(buf);
1439 if (type == ARC_BUFC_METADATA) {
1440 arc_buf_data_free(buf->b_hdr, zio_buf_free,
1442 arc_space_return(size, ARC_SPACE_DATA);
1444 ASSERT(type == ARC_BUFC_DATA);
1445 arc_buf_data_free(buf->b_hdr,
1446 zio_data_buf_free, buf->b_data, size);
1447 ARCSTAT_INCR(arcstat_data_size, -size);
1448 atomic_add_64(&arc_size, -size);
1451 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1452 uint64_t *cnt = &state->arcs_lsize[type];
1454 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1455 ASSERT(state != arc_anon);
1457 ASSERT3U(*cnt, >=, size);
1458 atomic_add_64(cnt, -size);
1460 ASSERT3U(state->arcs_size, >=, size);
1461 atomic_add_64(&state->arcs_size, -size);
1465 * If we're destroying a duplicate buffer make sure
1466 * that the appropriate statistics are updated.
1468 if (buf->b_hdr->b_datacnt > 1 &&
1469 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1470 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1471 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1473 ASSERT(buf->b_hdr->b_datacnt > 0);
1474 buf->b_hdr->b_datacnt -= 1;
1477 /* only remove the buf if requested */
1481 /* remove the buf from the hdr list */
1482 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1484 *bufp = buf->b_next;
1487 ASSERT(buf->b_efunc == NULL);
1489 /* clean up the buf */
1491 kmem_cache_free(buf_cache, buf);
1495 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1497 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1499 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1500 ASSERT3P(hdr->b_state, ==, arc_anon);
1501 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1503 if (l2hdr != NULL) {
1504 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1506 * To prevent arc_free() and l2arc_evict() from
1507 * attempting to free the same buffer at the same time,
1508 * a FREE_IN_PROGRESS flag is given to arc_free() to
1509 * give it priority. l2arc_evict() can't destroy this
1510 * header while we are waiting on l2arc_buflist_mtx.
1512 * The hdr may be removed from l2ad_buflist before we
1513 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1515 if (!buflist_held) {
1516 mutex_enter(&l2arc_buflist_mtx);
1517 l2hdr = hdr->b_l2hdr;
1520 if (l2hdr != NULL) {
1521 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1522 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1523 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1524 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
1525 if (hdr->b_state == arc_l2c_only)
1526 l2arc_hdr_stat_remove();
1527 hdr->b_l2hdr = NULL;
1531 mutex_exit(&l2arc_buflist_mtx);
1534 if (!BUF_EMPTY(hdr)) {
1535 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1536 buf_discard_identity(hdr);
1538 while (hdr->b_buf) {
1539 arc_buf_t *buf = hdr->b_buf;
1542 mutex_enter(&arc_eviction_mtx);
1543 mutex_enter(&buf->b_evict_lock);
1544 ASSERT(buf->b_hdr != NULL);
1545 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1546 hdr->b_buf = buf->b_next;
1547 buf->b_hdr = &arc_eviction_hdr;
1548 buf->b_next = arc_eviction_list;
1549 arc_eviction_list = buf;
1550 mutex_exit(&buf->b_evict_lock);
1551 mutex_exit(&arc_eviction_mtx);
1553 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1556 if (hdr->b_freeze_cksum != NULL) {
1557 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1558 hdr->b_freeze_cksum = NULL;
1561 ASSERT(!list_link_active(&hdr->b_arc_node));
1562 ASSERT3P(hdr->b_hash_next, ==, NULL);
1563 ASSERT3P(hdr->b_acb, ==, NULL);
1564 kmem_cache_free(hdr_cache, hdr);
1568 arc_buf_free(arc_buf_t *buf, void *tag)
1570 arc_buf_hdr_t *hdr = buf->b_hdr;
1571 int hashed = hdr->b_state != arc_anon;
1573 ASSERT(buf->b_efunc == NULL);
1574 ASSERT(buf->b_data != NULL);
1577 kmutex_t *hash_lock = HDR_LOCK(hdr);
1579 mutex_enter(hash_lock);
1581 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1583 (void) remove_reference(hdr, hash_lock, tag);
1584 if (hdr->b_datacnt > 1) {
1585 arc_buf_destroy(buf, FALSE, TRUE);
1587 ASSERT(buf == hdr->b_buf);
1588 ASSERT(buf->b_efunc == NULL);
1589 hdr->b_flags |= ARC_BUF_AVAILABLE;
1591 mutex_exit(hash_lock);
1592 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1595 * We are in the middle of an async write. Don't destroy
1596 * this buffer unless the write completes before we finish
1597 * decrementing the reference count.
1599 mutex_enter(&arc_eviction_mtx);
1600 (void) remove_reference(hdr, NULL, tag);
1601 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1602 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1603 mutex_exit(&arc_eviction_mtx);
1605 arc_hdr_destroy(hdr);
1607 if (remove_reference(hdr, NULL, tag) > 0)
1608 arc_buf_destroy(buf, FALSE, TRUE);
1610 arc_hdr_destroy(hdr);
1615 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1617 arc_buf_hdr_t *hdr = buf->b_hdr;
1618 kmutex_t *hash_lock = NULL;
1619 int no_callback = (buf->b_efunc == NULL);
1621 if (hdr->b_state == arc_anon) {
1622 ASSERT(hdr->b_datacnt == 1);
1623 arc_buf_free(buf, tag);
1624 return (no_callback);
1627 hash_lock = HDR_LOCK(hdr);
1628 mutex_enter(hash_lock);
1630 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1631 ASSERT(hdr->b_state != arc_anon);
1632 ASSERT(buf->b_data != NULL);
1634 (void) remove_reference(hdr, hash_lock, tag);
1635 if (hdr->b_datacnt > 1) {
1637 arc_buf_destroy(buf, FALSE, TRUE);
1638 } else if (no_callback) {
1639 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1640 ASSERT(buf->b_efunc == NULL);
1641 hdr->b_flags |= ARC_BUF_AVAILABLE;
1643 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1644 refcount_is_zero(&hdr->b_refcnt));
1645 mutex_exit(hash_lock);
1646 return (no_callback);
1650 arc_buf_size(arc_buf_t *buf)
1652 return (buf->b_hdr->b_size);
1656 * Called from the DMU to determine if the current buffer should be
1657 * evicted. In order to ensure proper locking, the eviction must be initiated
1658 * from the DMU. Return true if the buffer is associated with user data and
1659 * duplicate buffers still exist.
1662 arc_buf_eviction_needed(arc_buf_t *buf)
1665 boolean_t evict_needed = B_FALSE;
1667 if (zfs_disable_dup_eviction)
1670 mutex_enter(&buf->b_evict_lock);
1674 * We are in arc_do_user_evicts(); let that function
1675 * perform the eviction.
1677 ASSERT(buf->b_data == NULL);
1678 mutex_exit(&buf->b_evict_lock);
1680 } else if (buf->b_data == NULL) {
1682 * We have already been added to the arc eviction list;
1683 * recommend eviction.
1685 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1686 mutex_exit(&buf->b_evict_lock);
1690 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1691 evict_needed = B_TRUE;
1693 mutex_exit(&buf->b_evict_lock);
1694 return (evict_needed);
1698 * Evict buffers from list until we've removed the specified number of
1699 * bytes. Move the removed buffers to the appropriate evict state.
1700 * If the recycle flag is set, then attempt to "recycle" a buffer:
1701 * - look for a buffer to evict that is `bytes' long.
1702 * - return the data block from this buffer rather than freeing it.
1703 * This flag is used by callers that are trying to make space for a
1704 * new buffer in a full arc cache.
1706 * This function makes a "best effort". It skips over any buffers
1707 * it can't get a hash_lock on, and so may not catch all candidates.
1708 * It may also return without evicting as much space as requested.
1711 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1712 arc_buf_contents_t type)
1714 arc_state_t *evicted_state;
1715 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1716 arc_buf_hdr_t *ab, *ab_prev = NULL;
1717 list_t *list = &state->arcs_list[type];
1718 kmutex_t *hash_lock;
1719 boolean_t have_lock;
1720 void *stolen = NULL;
1722 ASSERT(state == arc_mru || state == arc_mfu);
1724 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1726 mutex_enter(&state->arcs_mtx);
1727 mutex_enter(&evicted_state->arcs_mtx);
1729 for (ab = list_tail(list); ab; ab = ab_prev) {
1730 ab_prev = list_prev(list, ab);
1731 /* prefetch buffers have a minimum lifespan */
1732 if (HDR_IO_IN_PROGRESS(ab) ||
1733 (spa && ab->b_spa != spa) ||
1734 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1735 ddi_get_lbolt() - ab->b_arc_access <
1736 zfs_arc_min_prefetch_lifespan)) {
1740 /* "lookahead" for better eviction candidate */
1741 if (recycle && ab->b_size != bytes &&
1742 ab_prev && ab_prev->b_size == bytes)
1744 hash_lock = HDR_LOCK(ab);
1745 have_lock = MUTEX_HELD(hash_lock);
1746 if (have_lock || mutex_tryenter(hash_lock)) {
1747 ASSERT0(refcount_count(&ab->b_refcnt));
1748 ASSERT(ab->b_datacnt > 0);
1750 arc_buf_t *buf = ab->b_buf;
1751 if (!mutex_tryenter(&buf->b_evict_lock)) {
1756 bytes_evicted += ab->b_size;
1757 if (recycle && ab->b_type == type &&
1758 ab->b_size == bytes &&
1759 !HDR_L2_WRITING(ab)) {
1760 stolen = buf->b_data;
1765 mutex_enter(&arc_eviction_mtx);
1766 arc_buf_destroy(buf,
1767 buf->b_data == stolen, FALSE);
1768 ab->b_buf = buf->b_next;
1769 buf->b_hdr = &arc_eviction_hdr;
1770 buf->b_next = arc_eviction_list;
1771 arc_eviction_list = buf;
1772 mutex_exit(&arc_eviction_mtx);
1773 mutex_exit(&buf->b_evict_lock);
1775 mutex_exit(&buf->b_evict_lock);
1776 arc_buf_destroy(buf,
1777 buf->b_data == stolen, TRUE);
1782 ARCSTAT_INCR(arcstat_evict_l2_cached,
1785 if (l2arc_write_eligible(ab->b_spa, ab)) {
1786 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1790 arcstat_evict_l2_ineligible,
1795 if (ab->b_datacnt == 0) {
1796 arc_change_state(evicted_state, ab, hash_lock);
1797 ASSERT(HDR_IN_HASH_TABLE(ab));
1798 ab->b_flags |= ARC_IN_HASH_TABLE;
1799 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1800 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1803 mutex_exit(hash_lock);
1804 if (bytes >= 0 && bytes_evicted >= bytes)
1811 mutex_exit(&evicted_state->arcs_mtx);
1812 mutex_exit(&state->arcs_mtx);
1814 if (bytes_evicted < bytes)
1815 dprintf("only evicted %lld bytes from %x\n",
1816 (longlong_t)bytes_evicted, state);
1819 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1822 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1825 * We have just evicted some date into the ghost state, make
1826 * sure we also adjust the ghost state size if necessary.
1829 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1830 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1831 arc_mru_ghost->arcs_size - arc_c;
1833 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1835 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1836 arc_evict_ghost(arc_mru_ghost, 0, todelete);
1837 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1838 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1839 arc_mru_ghost->arcs_size +
1840 arc_mfu_ghost->arcs_size - arc_c);
1841 arc_evict_ghost(arc_mfu_ghost, 0, todelete);
1849 * Remove buffers from list until we've removed the specified number of
1850 * bytes. Destroy the buffers that are removed.
1853 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1855 arc_buf_hdr_t *ab, *ab_prev;
1856 arc_buf_hdr_t marker;
1857 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1858 kmutex_t *hash_lock;
1859 uint64_t bytes_deleted = 0;
1860 uint64_t bufs_skipped = 0;
1862 ASSERT(GHOST_STATE(state));
1863 bzero(&marker, sizeof(marker));
1865 mutex_enter(&state->arcs_mtx);
1866 for (ab = list_tail(list); ab; ab = ab_prev) {
1867 ab_prev = list_prev(list, ab);
1868 if (spa && ab->b_spa != spa)
1871 /* ignore markers */
1875 hash_lock = HDR_LOCK(ab);
1876 /* caller may be trying to modify this buffer, skip it */
1877 if (MUTEX_HELD(hash_lock))
1879 if (mutex_tryenter(hash_lock)) {
1880 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1881 ASSERT(ab->b_buf == NULL);
1882 ARCSTAT_BUMP(arcstat_deleted);
1883 bytes_deleted += ab->b_size;
1885 if (ab->b_l2hdr != NULL) {
1887 * This buffer is cached on the 2nd Level ARC;
1888 * don't destroy the header.
1890 arc_change_state(arc_l2c_only, ab, hash_lock);
1891 mutex_exit(hash_lock);
1893 arc_change_state(arc_anon, ab, hash_lock);
1894 mutex_exit(hash_lock);
1895 arc_hdr_destroy(ab);
1898 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1899 if (bytes >= 0 && bytes_deleted >= bytes)
1901 } else if (bytes < 0) {
1903 * Insert a list marker and then wait for the
1904 * hash lock to become available. Once its
1905 * available, restart from where we left off.
1907 list_insert_after(list, ab, &marker);
1908 mutex_exit(&state->arcs_mtx);
1909 mutex_enter(hash_lock);
1910 mutex_exit(hash_lock);
1911 mutex_enter(&state->arcs_mtx);
1912 ab_prev = list_prev(list, &marker);
1913 list_remove(list, &marker);
1917 mutex_exit(&state->arcs_mtx);
1919 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1920 (bytes < 0 || bytes_deleted < bytes)) {
1921 list = &state->arcs_list[ARC_BUFC_METADATA];
1926 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1930 if (bytes_deleted < bytes)
1931 dprintf("only deleted %lld bytes from %p\n",
1932 (longlong_t)bytes_deleted, state);
1938 int64_t adjustment, delta;
1944 adjustment = MIN((int64_t)(arc_size - arc_c),
1945 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
1948 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1949 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
1950 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
1951 adjustment -= delta;
1954 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1955 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
1956 (void) arc_evict(arc_mru, 0, delta, FALSE,
1964 adjustment = arc_size - arc_c;
1966 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1967 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
1968 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
1969 adjustment -= delta;
1972 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1973 int64_t delta = MIN(adjustment,
1974 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
1975 (void) arc_evict(arc_mfu, 0, delta, FALSE,
1980 * Adjust ghost lists
1983 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
1985 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
1986 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
1987 arc_evict_ghost(arc_mru_ghost, 0, delta);
1991 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
1993 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
1994 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
1995 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2000 * Request that arc user drop references so that N bytes can be released
2001 * from the cache. This provides a mechanism to ensure the arc can honor
2002 * the arc_meta_limit and reclaim buffers which are pinned in the cache
2003 * by higher layers. (i.e. the zpl)
2006 arc_do_user_prune(int64_t adjustment)
2008 arc_prune_func_t *func;
2010 arc_prune_t *cp, *np;
2012 mutex_enter(&arc_prune_mtx);
2014 cp = list_head(&arc_prune_list);
2015 while (cp != NULL) {
2017 private = cp->p_private;
2018 np = list_next(&arc_prune_list, cp);
2019 refcount_add(&cp->p_refcnt, func);
2020 mutex_exit(&arc_prune_mtx);
2023 func(adjustment, private);
2025 mutex_enter(&arc_prune_mtx);
2027 /* User removed prune callback concurrently with execution */
2028 if (refcount_remove(&cp->p_refcnt, func) == 0) {
2029 ASSERT(!list_link_active(&cp->p_node));
2030 refcount_destroy(&cp->p_refcnt);
2031 kmem_free(cp, sizeof (*cp));
2037 ARCSTAT_BUMP(arcstat_prune);
2038 mutex_exit(&arc_prune_mtx);
2042 arc_do_user_evicts(void)
2044 mutex_enter(&arc_eviction_mtx);
2045 while (arc_eviction_list != NULL) {
2046 arc_buf_t *buf = arc_eviction_list;
2047 arc_eviction_list = buf->b_next;
2048 mutex_enter(&buf->b_evict_lock);
2050 mutex_exit(&buf->b_evict_lock);
2051 mutex_exit(&arc_eviction_mtx);
2053 if (buf->b_efunc != NULL)
2054 VERIFY(buf->b_efunc(buf) == 0);
2056 buf->b_efunc = NULL;
2057 buf->b_private = NULL;
2058 kmem_cache_free(buf_cache, buf);
2059 mutex_enter(&arc_eviction_mtx);
2061 mutex_exit(&arc_eviction_mtx);
2065 * Evict only meta data objects from the cache leaving the data objects.
2066 * This is only used to enforce the tunable arc_meta_limit, if we are
2067 * unable to evict enough buffers notify the user via the prune callback.
2070 arc_adjust_meta(int64_t adjustment, boolean_t may_prune)
2074 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2075 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2076 arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_METADATA);
2077 adjustment -= delta;
2080 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2081 delta = MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2082 arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_METADATA);
2083 adjustment -= delta;
2086 if (may_prune && (adjustment > 0) && (arc_meta_used > arc_meta_limit))
2087 arc_do_user_prune(zfs_arc_meta_prune);
2091 * Flush all *evictable* data from the cache for the given spa.
2092 * NOTE: this will not touch "active" (i.e. referenced) data.
2095 arc_flush(spa_t *spa)
2100 guid = spa_load_guid(spa);
2102 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2103 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2107 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2108 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2112 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2113 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2117 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2118 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2123 arc_evict_ghost(arc_mru_ghost, guid, -1);
2124 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2126 mutex_enter(&arc_reclaim_thr_lock);
2127 arc_do_user_evicts();
2128 mutex_exit(&arc_reclaim_thr_lock);
2129 ASSERT(spa || arc_eviction_list == NULL);
2133 arc_shrink(uint64_t bytes)
2135 if (arc_c > arc_c_min) {
2138 to_free = bytes ? bytes : arc_c >> zfs_arc_shrink_shift;
2140 if (arc_c > arc_c_min + to_free)
2141 atomic_add_64(&arc_c, -to_free);
2145 atomic_add_64(&arc_p, -(arc_p >> zfs_arc_shrink_shift));
2146 if (arc_c > arc_size)
2147 arc_c = MAX(arc_size, arc_c_min);
2149 arc_p = (arc_c >> 1);
2150 ASSERT(arc_c >= arc_c_min);
2151 ASSERT((int64_t)arc_p >= 0);
2154 if (arc_size > arc_c)
2159 arc_kmem_reap_now(arc_reclaim_strategy_t strat, uint64_t bytes)
2162 kmem_cache_t *prev_cache = NULL;
2163 kmem_cache_t *prev_data_cache = NULL;
2164 extern kmem_cache_t *zio_buf_cache[];
2165 extern kmem_cache_t *zio_data_buf_cache[];
2168 * An aggressive reclamation will shrink the cache size as well as
2169 * reap free buffers from the arc kmem caches.
2171 if (strat == ARC_RECLAIM_AGGR)
2174 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2175 if (zio_buf_cache[i] != prev_cache) {
2176 prev_cache = zio_buf_cache[i];
2177 kmem_cache_reap_now(zio_buf_cache[i]);
2179 if (zio_data_buf_cache[i] != prev_data_cache) {
2180 prev_data_cache = zio_data_buf_cache[i];
2181 kmem_cache_reap_now(zio_data_buf_cache[i]);
2185 kmem_cache_reap_now(buf_cache);
2186 kmem_cache_reap_now(hdr_cache);
2190 * Unlike other ZFS implementations this thread is only responsible for
2191 * adapting the target ARC size on Linux. The responsibility for memory
2192 * reclamation has been entirely delegated to the arc_shrinker_func()
2193 * which is registered with the VM. To reflect this change in behavior
2194 * the arc_reclaim thread has been renamed to arc_adapt.
2197 arc_adapt_thread(void)
2202 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2204 mutex_enter(&arc_reclaim_thr_lock);
2205 while (arc_thread_exit == 0) {
2207 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2209 if (spa_get_random(100) == 0) {
2212 if (last_reclaim == ARC_RECLAIM_CONS) {
2213 last_reclaim = ARC_RECLAIM_AGGR;
2215 last_reclaim = ARC_RECLAIM_CONS;
2219 last_reclaim = ARC_RECLAIM_AGGR;
2223 /* reset the growth delay for every reclaim */
2224 arc_grow_time = ddi_get_lbolt()+(zfs_arc_grow_retry * hz);
2226 arc_kmem_reap_now(last_reclaim, 0);
2229 #endif /* !_KERNEL */
2231 /* No recent memory pressure allow the ARC to grow. */
2232 if (arc_no_grow && ddi_get_lbolt() >= arc_grow_time)
2233 arc_no_grow = FALSE;
2236 * Keep meta data usage within limits, arc_shrink() is not
2237 * used to avoid collapsing the arc_c value when only the
2238 * arc_meta_limit is being exceeded.
2240 prune = (int64_t)arc_meta_used - (int64_t)arc_meta_limit;
2242 arc_adjust_meta(prune, B_TRUE);
2246 if (arc_eviction_list != NULL)
2247 arc_do_user_evicts();
2249 /* block until needed, or one second, whichever is shorter */
2250 CALLB_CPR_SAFE_BEGIN(&cpr);
2251 (void) cv_timedwait_interruptible(&arc_reclaim_thr_cv,
2252 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2253 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2256 /* Allow the module options to be changed */
2257 if (zfs_arc_max > 64 << 20 &&
2258 zfs_arc_max < physmem * PAGESIZE &&
2259 zfs_arc_max != arc_c_max)
2260 arc_c_max = zfs_arc_max;
2262 if (zfs_arc_min > 0 &&
2263 zfs_arc_min < arc_c_max &&
2264 zfs_arc_min != arc_c_min)
2265 arc_c_min = zfs_arc_min;
2267 if (zfs_arc_meta_limit > 0 &&
2268 zfs_arc_meta_limit <= arc_c_max &&
2269 zfs_arc_meta_limit != arc_meta_limit)
2270 arc_meta_limit = zfs_arc_meta_limit;
2276 arc_thread_exit = 0;
2277 cv_broadcast(&arc_reclaim_thr_cv);
2278 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2284 * Determine the amount of memory eligible for eviction contained in the
2285 * ARC. All clean data reported by the ghost lists can always be safely
2286 * evicted. Due to arc_c_min, the same does not hold for all clean data
2287 * contained by the regular mru and mfu lists.
2289 * In the case of the regular mru and mfu lists, we need to report as
2290 * much clean data as possible, such that evicting that same reported
2291 * data will not bring arc_size below arc_c_min. Thus, in certain
2292 * circumstances, the total amount of clean data in the mru and mfu
2293 * lists might not actually be evictable.
2295 * The following two distinct cases are accounted for:
2297 * 1. The sum of the amount of dirty data contained by both the mru and
2298 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2299 * is greater than or equal to arc_c_min.
2300 * (i.e. amount of dirty data >= arc_c_min)
2302 * This is the easy case; all clean data contained by the mru and mfu
2303 * lists is evictable. Evicting all clean data can only drop arc_size
2304 * to the amount of dirty data, which is greater than arc_c_min.
2306 * 2. The sum of the amount of dirty data contained by both the mru and
2307 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2308 * is less than arc_c_min.
2309 * (i.e. arc_c_min > amount of dirty data)
2311 * 2.1. arc_size is greater than or equal arc_c_min.
2312 * (i.e. arc_size >= arc_c_min > amount of dirty data)
2314 * In this case, not all clean data from the regular mru and mfu
2315 * lists is actually evictable; we must leave enough clean data
2316 * to keep arc_size above arc_c_min. Thus, the maximum amount of
2317 * evictable data from the two lists combined, is exactly the
2318 * difference between arc_size and arc_c_min.
2320 * 2.2. arc_size is less than arc_c_min
2321 * (i.e. arc_c_min > arc_size > amount of dirty data)
2323 * In this case, none of the data contained in the mru and mfu
2324 * lists is evictable, even if it's clean. Since arc_size is
2325 * already below arc_c_min, evicting any more would only
2326 * increase this negative difference.
2329 arc_evictable_memory(void) {
2330 uint64_t arc_clean =
2331 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
2332 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
2333 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
2334 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
2335 uint64_t ghost_clean =
2336 arc_mru_ghost->arcs_lsize[ARC_BUFC_DATA] +
2337 arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] +
2338 arc_mfu_ghost->arcs_lsize[ARC_BUFC_DATA] +
2339 arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA];
2340 uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0);
2342 if (arc_dirty >= arc_c_min)
2343 return (ghost_clean + arc_clean);
2345 return (ghost_clean + MAX((int64_t)arc_size - (int64_t)arc_c_min, 0));
2349 __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc)
2353 /* The arc is considered warm once reclaim has occurred */
2354 if (unlikely(arc_warm == B_FALSE))
2357 /* Return the potential number of reclaimable pages */
2358 pages = btop(arc_evictable_memory());
2359 if (sc->nr_to_scan == 0)
2362 /* Not allowed to perform filesystem reclaim */
2363 if (!(sc->gfp_mask & __GFP_FS))
2366 /* Reclaim in progress */
2367 if (mutex_tryenter(&arc_reclaim_thr_lock) == 0)
2371 * Evict the requested number of pages by shrinking arc_c the
2372 * requested amount. If there is nothing left to evict just
2373 * reap whatever we can from the various arc slabs.
2376 arc_kmem_reap_now(ARC_RECLAIM_AGGR, ptob(sc->nr_to_scan));
2377 pages = btop(arc_evictable_memory());
2379 arc_kmem_reap_now(ARC_RECLAIM_CONS, ptob(sc->nr_to_scan));
2384 * When direct reclaim is observed it usually indicates a rapid
2385 * increase in memory pressure. This occurs because the kswapd
2386 * threads were unable to asynchronously keep enough free memory
2387 * available. In this case set arc_no_grow to briefly pause arc
2388 * growth to avoid compounding the memory pressure.
2390 if (current_is_kswapd()) {
2391 ARCSTAT_BUMP(arcstat_memory_indirect_count);
2393 arc_no_grow = B_TRUE;
2394 arc_grow_time = ddi_get_lbolt() + (zfs_arc_grow_retry * hz);
2395 ARCSTAT_BUMP(arcstat_memory_direct_count);
2398 mutex_exit(&arc_reclaim_thr_lock);
2402 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func);
2404 SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS);
2405 #endif /* _KERNEL */
2408 * Adapt arc info given the number of bytes we are trying to add and
2409 * the state that we are comming from. This function is only called
2410 * when we are adding new content to the cache.
2413 arc_adapt(int bytes, arc_state_t *state)
2416 uint64_t arc_p_min = (arc_c >> zfs_arc_p_min_shift);
2418 if (state == arc_l2c_only)
2423 * Adapt the target size of the MRU list:
2424 * - if we just hit in the MRU ghost list, then increase
2425 * the target size of the MRU list.
2426 * - if we just hit in the MFU ghost list, then increase
2427 * the target size of the MFU list by decreasing the
2428 * target size of the MRU list.
2430 if (state == arc_mru_ghost) {
2431 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2432 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2433 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2435 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2436 } else if (state == arc_mfu_ghost) {
2439 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2440 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2441 mult = MIN(mult, 10);
2443 delta = MIN(bytes * mult, arc_p);
2444 arc_p = MAX(arc_p_min, arc_p - delta);
2446 ASSERT((int64_t)arc_p >= 0);
2451 if (arc_c >= arc_c_max)
2455 * If we're within (2 * maxblocksize) bytes of the target
2456 * cache size, increment the target cache size
2458 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2459 atomic_add_64(&arc_c, (int64_t)bytes);
2460 if (arc_c > arc_c_max)
2462 else if (state == arc_anon)
2463 atomic_add_64(&arc_p, (int64_t)bytes);
2467 ASSERT((int64_t)arc_p >= 0);
2471 * Check if the cache has reached its limits and eviction is required
2475 arc_evict_needed(arc_buf_contents_t type)
2477 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2483 return (arc_size > arc_c);
2487 * The buffer, supplied as the first argument, needs a data block.
2488 * So, if we are at cache max, determine which cache should be victimized.
2489 * We have the following cases:
2491 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2492 * In this situation if we're out of space, but the resident size of the MFU is
2493 * under the limit, victimize the MFU cache to satisfy this insertion request.
2495 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2496 * Here, we've used up all of the available space for the MRU, so we need to
2497 * evict from our own cache instead. Evict from the set of resident MRU
2500 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2501 * c minus p represents the MFU space in the cache, since p is the size of the
2502 * cache that is dedicated to the MRU. In this situation there's still space on
2503 * the MFU side, so the MRU side needs to be victimized.
2505 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2506 * MFU's resident set is consuming more space than it has been allotted. In
2507 * this situation, we must victimize our own cache, the MFU, for this insertion.
2510 arc_get_data_buf(arc_buf_t *buf)
2512 arc_state_t *state = buf->b_hdr->b_state;
2513 uint64_t size = buf->b_hdr->b_size;
2514 arc_buf_contents_t type = buf->b_hdr->b_type;
2516 arc_adapt(size, state);
2519 * We have not yet reached cache maximum size,
2520 * just allocate a new buffer.
2522 if (!arc_evict_needed(type)) {
2523 if (type == ARC_BUFC_METADATA) {
2524 buf->b_data = zio_buf_alloc(size);
2525 arc_space_consume(size, ARC_SPACE_DATA);
2527 ASSERT(type == ARC_BUFC_DATA);
2528 buf->b_data = zio_data_buf_alloc(size);
2529 ARCSTAT_INCR(arcstat_data_size, size);
2530 atomic_add_64(&arc_size, size);
2536 * If we are prefetching from the mfu ghost list, this buffer
2537 * will end up on the mru list; so steal space from there.
2539 if (state == arc_mfu_ghost)
2540 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2541 else if (state == arc_mru_ghost)
2544 if (state == arc_mru || state == arc_anon) {
2545 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2546 state = (arc_mfu->arcs_lsize[type] >= size &&
2547 arc_p > mru_used) ? arc_mfu : arc_mru;
2550 uint64_t mfu_space = arc_c - arc_p;
2551 state = (arc_mru->arcs_lsize[type] >= size &&
2552 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2555 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2556 if (type == ARC_BUFC_METADATA) {
2557 buf->b_data = zio_buf_alloc(size);
2558 arc_space_consume(size, ARC_SPACE_DATA);
2561 * If we are unable to recycle an existing meta buffer
2562 * signal the reclaim thread. It will notify users
2563 * via the prune callback to drop references. The
2564 * prune callback in run in the context of the reclaim
2565 * thread to avoid deadlocking on the hash_lock.
2567 cv_signal(&arc_reclaim_thr_cv);
2569 ASSERT(type == ARC_BUFC_DATA);
2570 buf->b_data = zio_data_buf_alloc(size);
2571 ARCSTAT_INCR(arcstat_data_size, size);
2572 atomic_add_64(&arc_size, size);
2575 ARCSTAT_BUMP(arcstat_recycle_miss);
2577 ASSERT(buf->b_data != NULL);
2580 * Update the state size. Note that ghost states have a
2581 * "ghost size" and so don't need to be updated.
2583 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2584 arc_buf_hdr_t *hdr = buf->b_hdr;
2586 atomic_add_64(&hdr->b_state->arcs_size, size);
2587 if (list_link_active(&hdr->b_arc_node)) {
2588 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2589 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2592 * If we are growing the cache, and we are adding anonymous
2593 * data, and we have outgrown arc_p, update arc_p
2595 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2596 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2597 arc_p = MIN(arc_c, arc_p + size);
2602 * This routine is called whenever a buffer is accessed.
2603 * NOTE: the hash lock is dropped in this function.
2606 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2610 ASSERT(MUTEX_HELD(hash_lock));
2612 if (buf->b_state == arc_anon) {
2614 * This buffer is not in the cache, and does not
2615 * appear in our "ghost" list. Add the new buffer
2619 ASSERT(buf->b_arc_access == 0);
2620 buf->b_arc_access = ddi_get_lbolt();
2621 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2622 arc_change_state(arc_mru, buf, hash_lock);
2624 } else if (buf->b_state == arc_mru) {
2625 now = ddi_get_lbolt();
2628 * If this buffer is here because of a prefetch, then either:
2629 * - clear the flag if this is a "referencing" read
2630 * (any subsequent access will bump this into the MFU state).
2632 * - move the buffer to the head of the list if this is
2633 * another prefetch (to make it less likely to be evicted).
2635 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2636 if (refcount_count(&buf->b_refcnt) == 0) {
2637 ASSERT(list_link_active(&buf->b_arc_node));
2639 buf->b_flags &= ~ARC_PREFETCH;
2640 ARCSTAT_BUMP(arcstat_mru_hits);
2642 buf->b_arc_access = now;
2647 * This buffer has been "accessed" only once so far,
2648 * but it is still in the cache. Move it to the MFU
2651 if (now > buf->b_arc_access + ARC_MINTIME) {
2653 * More than 125ms have passed since we
2654 * instantiated this buffer. Move it to the
2655 * most frequently used state.
2657 buf->b_arc_access = now;
2658 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2659 arc_change_state(arc_mfu, buf, hash_lock);
2661 ARCSTAT_BUMP(arcstat_mru_hits);
2662 } else if (buf->b_state == arc_mru_ghost) {
2663 arc_state_t *new_state;
2665 * This buffer has been "accessed" recently, but
2666 * was evicted from the cache. Move it to the
2670 if (buf->b_flags & ARC_PREFETCH) {
2671 new_state = arc_mru;
2672 if (refcount_count(&buf->b_refcnt) > 0)
2673 buf->b_flags &= ~ARC_PREFETCH;
2674 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2676 new_state = arc_mfu;
2677 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2680 buf->b_arc_access = ddi_get_lbolt();
2681 arc_change_state(new_state, buf, hash_lock);
2683 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2684 } else if (buf->b_state == arc_mfu) {
2686 * This buffer has been accessed more than once and is
2687 * still in the cache. Keep it in the MFU state.
2689 * NOTE: an add_reference() that occurred when we did
2690 * the arc_read() will have kicked this off the list.
2691 * If it was a prefetch, we will explicitly move it to
2692 * the head of the list now.
2694 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2695 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2696 ASSERT(list_link_active(&buf->b_arc_node));
2698 ARCSTAT_BUMP(arcstat_mfu_hits);
2699 buf->b_arc_access = ddi_get_lbolt();
2700 } else if (buf->b_state == arc_mfu_ghost) {
2701 arc_state_t *new_state = arc_mfu;
2703 * This buffer has been accessed more than once but has
2704 * been evicted from the cache. Move it back to the
2708 if (buf->b_flags & ARC_PREFETCH) {
2710 * This is a prefetch access...
2711 * move this block back to the MRU state.
2713 ASSERT0(refcount_count(&buf->b_refcnt));
2714 new_state = arc_mru;
2717 buf->b_arc_access = ddi_get_lbolt();
2718 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2719 arc_change_state(new_state, buf, hash_lock);
2721 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2722 } else if (buf->b_state == arc_l2c_only) {
2724 * This buffer is on the 2nd Level ARC.
2727 buf->b_arc_access = ddi_get_lbolt();
2728 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2729 arc_change_state(arc_mfu, buf, hash_lock);
2731 ASSERT(!"invalid arc state");
2735 /* a generic arc_done_func_t which you can use */
2738 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2740 if (zio == NULL || zio->io_error == 0)
2741 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2742 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2745 /* a generic arc_done_func_t */
2747 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2749 arc_buf_t **bufp = arg;
2750 if (zio && zio->io_error) {
2751 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2755 ASSERT(buf->b_data);
2760 arc_read_done(zio_t *zio)
2762 arc_buf_hdr_t *hdr, *found;
2764 arc_buf_t *abuf; /* buffer we're assigning to callback */
2765 kmutex_t *hash_lock;
2766 arc_callback_t *callback_list, *acb;
2767 int freeable = FALSE;
2769 buf = zio->io_private;
2773 * The hdr was inserted into hash-table and removed from lists
2774 * prior to starting I/O. We should find this header, since
2775 * it's in the hash table, and it should be legit since it's
2776 * not possible to evict it during the I/O. The only possible
2777 * reason for it not to be found is if we were freed during the
2780 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2783 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2784 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2785 (found == hdr && HDR_L2_READING(hdr)));
2787 hdr->b_flags &= ~ARC_L2_EVICTED;
2788 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2789 hdr->b_flags &= ~ARC_L2CACHE;
2791 /* byteswap if necessary */
2792 callback_list = hdr->b_acb;
2793 ASSERT(callback_list != NULL);
2794 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2795 dmu_object_byteswap_t bswap =
2796 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2797 if (BP_GET_LEVEL(zio->io_bp) > 0)
2798 byteswap_uint64_array(buf->b_data, hdr->b_size);
2800 dmu_ot_byteswap[bswap].ob_func(buf->b_data, hdr->b_size);
2803 arc_cksum_compute(buf, B_FALSE);
2805 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2807 * Only call arc_access on anonymous buffers. This is because
2808 * if we've issued an I/O for an evicted buffer, we've already
2809 * called arc_access (to prevent any simultaneous readers from
2810 * getting confused).
2812 arc_access(hdr, hash_lock);
2815 /* create copies of the data buffer for the callers */
2817 for (acb = callback_list; acb; acb = acb->acb_next) {
2818 if (acb->acb_done) {
2820 ARCSTAT_BUMP(arcstat_duplicate_reads);
2821 abuf = arc_buf_clone(buf);
2823 acb->acb_buf = abuf;
2828 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2829 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2831 ASSERT(buf->b_efunc == NULL);
2832 ASSERT(hdr->b_datacnt == 1);
2833 hdr->b_flags |= ARC_BUF_AVAILABLE;
2836 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2838 if (zio->io_error != 0) {
2839 hdr->b_flags |= ARC_IO_ERROR;
2840 if (hdr->b_state != arc_anon)
2841 arc_change_state(arc_anon, hdr, hash_lock);
2842 if (HDR_IN_HASH_TABLE(hdr))
2843 buf_hash_remove(hdr);
2844 freeable = refcount_is_zero(&hdr->b_refcnt);
2848 * Broadcast before we drop the hash_lock to avoid the possibility
2849 * that the hdr (and hence the cv) might be freed before we get to
2850 * the cv_broadcast().
2852 cv_broadcast(&hdr->b_cv);
2855 mutex_exit(hash_lock);
2858 * This block was freed while we waited for the read to
2859 * complete. It has been removed from the hash table and
2860 * moved to the anonymous state (so that it won't show up
2863 ASSERT3P(hdr->b_state, ==, arc_anon);
2864 freeable = refcount_is_zero(&hdr->b_refcnt);
2867 /* execute each callback and free its structure */
2868 while ((acb = callback_list) != NULL) {
2870 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2872 if (acb->acb_zio_dummy != NULL) {
2873 acb->acb_zio_dummy->io_error = zio->io_error;
2874 zio_nowait(acb->acb_zio_dummy);
2877 callback_list = acb->acb_next;
2878 kmem_free(acb, sizeof (arc_callback_t));
2882 arc_hdr_destroy(hdr);
2886 * "Read" the block at the specified DVA (in bp) via the
2887 * cache. If the block is found in the cache, invoke the provided
2888 * callback immediately and return. Note that the `zio' parameter
2889 * in the callback will be NULL in this case, since no IO was
2890 * required. If the block is not in the cache pass the read request
2891 * on to the spa with a substitute callback function, so that the
2892 * requested block will be added to the cache.
2894 * If a read request arrives for a block that has a read in-progress,
2895 * either wait for the in-progress read to complete (and return the
2896 * results); or, if this is a read with a "done" func, add a record
2897 * to the read to invoke the "done" func when the read completes,
2898 * and return; or just return.
2900 * arc_read_done() will invoke all the requested "done" functions
2901 * for readers of this block.
2904 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
2905 void *private, int priority, int zio_flags, uint32_t *arc_flags,
2906 const zbookmark_t *zb)
2909 arc_buf_t *buf = NULL;
2910 kmutex_t *hash_lock;
2912 uint64_t guid = spa_load_guid(spa);
2915 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2917 if (hdr && hdr->b_datacnt > 0) {
2919 *arc_flags |= ARC_CACHED;
2921 if (HDR_IO_IN_PROGRESS(hdr)) {
2923 if (*arc_flags & ARC_WAIT) {
2924 cv_wait(&hdr->b_cv, hash_lock);
2925 mutex_exit(hash_lock);
2928 ASSERT(*arc_flags & ARC_NOWAIT);
2931 arc_callback_t *acb = NULL;
2933 acb = kmem_zalloc(sizeof (arc_callback_t),
2935 acb->acb_done = done;
2936 acb->acb_private = private;
2938 acb->acb_zio_dummy = zio_null(pio,
2939 spa, NULL, NULL, NULL, zio_flags);
2941 ASSERT(acb->acb_done != NULL);
2942 acb->acb_next = hdr->b_acb;
2944 add_reference(hdr, hash_lock, private);
2945 mutex_exit(hash_lock);
2948 mutex_exit(hash_lock);
2952 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2955 add_reference(hdr, hash_lock, private);
2957 * If this block is already in use, create a new
2958 * copy of the data so that we will be guaranteed
2959 * that arc_release() will always succeed.
2963 ASSERT(buf->b_data);
2964 if (HDR_BUF_AVAILABLE(hdr)) {
2965 ASSERT(buf->b_efunc == NULL);
2966 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2968 buf = arc_buf_clone(buf);
2971 } else if (*arc_flags & ARC_PREFETCH &&
2972 refcount_count(&hdr->b_refcnt) == 0) {
2973 hdr->b_flags |= ARC_PREFETCH;
2975 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2976 arc_access(hdr, hash_lock);
2977 if (*arc_flags & ARC_L2CACHE)
2978 hdr->b_flags |= ARC_L2CACHE;
2979 mutex_exit(hash_lock);
2980 ARCSTAT_BUMP(arcstat_hits);
2981 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2982 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2983 data, metadata, hits);
2986 done(NULL, buf, private);
2988 uint64_t size = BP_GET_LSIZE(bp);
2989 arc_callback_t *acb;
2992 boolean_t devw = B_FALSE;
2995 /* this block is not in the cache */
2996 arc_buf_hdr_t *exists;
2997 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2998 buf = arc_buf_alloc(spa, size, private, type);
3000 hdr->b_dva = *BP_IDENTITY(bp);
3001 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3002 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3003 exists = buf_hash_insert(hdr, &hash_lock);
3005 /* somebody beat us to the hash insert */
3006 mutex_exit(hash_lock);
3007 buf_discard_identity(hdr);
3008 (void) arc_buf_remove_ref(buf, private);
3009 goto top; /* restart the IO request */
3011 /* if this is a prefetch, we don't have a reference */
3012 if (*arc_flags & ARC_PREFETCH) {
3013 (void) remove_reference(hdr, hash_lock,
3015 hdr->b_flags |= ARC_PREFETCH;
3017 if (*arc_flags & ARC_L2CACHE)
3018 hdr->b_flags |= ARC_L2CACHE;
3019 if (BP_GET_LEVEL(bp) > 0)
3020 hdr->b_flags |= ARC_INDIRECT;
3022 /* this block is in the ghost cache */
3023 ASSERT(GHOST_STATE(hdr->b_state));
3024 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3025 ASSERT0(refcount_count(&hdr->b_refcnt));
3026 ASSERT(hdr->b_buf == NULL);
3028 /* if this is a prefetch, we don't have a reference */
3029 if (*arc_flags & ARC_PREFETCH)
3030 hdr->b_flags |= ARC_PREFETCH;
3032 add_reference(hdr, hash_lock, private);
3033 if (*arc_flags & ARC_L2CACHE)
3034 hdr->b_flags |= ARC_L2CACHE;
3035 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3038 buf->b_efunc = NULL;
3039 buf->b_private = NULL;
3042 ASSERT(hdr->b_datacnt == 0);
3044 arc_get_data_buf(buf);
3045 arc_access(hdr, hash_lock);
3048 ASSERT(!GHOST_STATE(hdr->b_state));
3050 acb = kmem_zalloc(sizeof (arc_callback_t), KM_PUSHPAGE);
3051 acb->acb_done = done;
3052 acb->acb_private = private;
3054 ASSERT(hdr->b_acb == NULL);
3056 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3058 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3059 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3060 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3061 addr = hdr->b_l2hdr->b_daddr;
3063 * Lock out device removal.
3065 if (vdev_is_dead(vd) ||
3066 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3070 mutex_exit(hash_lock);
3072 ASSERT3U(hdr->b_size, ==, size);
3073 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3074 uint64_t, size, zbookmark_t *, zb);
3075 ARCSTAT_BUMP(arcstat_misses);
3076 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3077 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3078 data, metadata, misses);
3080 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3082 * Read from the L2ARC if the following are true:
3083 * 1. The L2ARC vdev was previously cached.
3084 * 2. This buffer still has L2ARC metadata.
3085 * 3. This buffer isn't currently writing to the L2ARC.
3086 * 4. The L2ARC entry wasn't evicted, which may
3087 * also have invalidated the vdev.
3088 * 5. This isn't prefetch and l2arc_noprefetch is set.
3090 if (hdr->b_l2hdr != NULL &&
3091 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3092 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3093 l2arc_read_callback_t *cb;
3095 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3096 ARCSTAT_BUMP(arcstat_l2_hits);
3098 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3100 cb->l2rcb_buf = buf;
3101 cb->l2rcb_spa = spa;
3104 cb->l2rcb_flags = zio_flags;
3107 * l2arc read. The SCL_L2ARC lock will be
3108 * released by l2arc_read_done().
3110 rzio = zio_read_phys(pio, vd, addr, size,
3111 buf->b_data, ZIO_CHECKSUM_OFF,
3112 l2arc_read_done, cb, priority, zio_flags |
3113 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
3114 ZIO_FLAG_DONT_PROPAGATE |
3115 ZIO_FLAG_DONT_RETRY, B_FALSE);
3116 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3118 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
3120 if (*arc_flags & ARC_NOWAIT) {
3125 ASSERT(*arc_flags & ARC_WAIT);
3126 if (zio_wait(rzio) == 0)
3129 /* l2arc read error; goto zio_read() */
3131 DTRACE_PROBE1(l2arc__miss,
3132 arc_buf_hdr_t *, hdr);
3133 ARCSTAT_BUMP(arcstat_l2_misses);
3134 if (HDR_L2_WRITING(hdr))
3135 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3136 spa_config_exit(spa, SCL_L2ARC, vd);
3140 spa_config_exit(spa, SCL_L2ARC, vd);
3141 if (l2arc_ndev != 0) {
3142 DTRACE_PROBE1(l2arc__miss,
3143 arc_buf_hdr_t *, hdr);
3144 ARCSTAT_BUMP(arcstat_l2_misses);
3148 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3149 arc_read_done, buf, priority, zio_flags, zb);
3151 if (*arc_flags & ARC_WAIT)
3152 return (zio_wait(rzio));
3154 ASSERT(*arc_flags & ARC_NOWAIT);
3161 arc_add_prune_callback(arc_prune_func_t *func, void *private)
3165 p = kmem_alloc(sizeof(*p), KM_SLEEP);
3167 p->p_private = private;
3168 list_link_init(&p->p_node);
3169 refcount_create(&p->p_refcnt);
3171 mutex_enter(&arc_prune_mtx);
3172 refcount_add(&p->p_refcnt, &arc_prune_list);
3173 list_insert_head(&arc_prune_list, p);
3174 mutex_exit(&arc_prune_mtx);
3180 arc_remove_prune_callback(arc_prune_t *p)
3182 mutex_enter(&arc_prune_mtx);
3183 list_remove(&arc_prune_list, p);
3184 if (refcount_remove(&p->p_refcnt, &arc_prune_list) == 0) {
3185 refcount_destroy(&p->p_refcnt);
3186 kmem_free(p, sizeof (*p));
3188 mutex_exit(&arc_prune_mtx);
3192 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3194 ASSERT(buf->b_hdr != NULL);
3195 ASSERT(buf->b_hdr->b_state != arc_anon);
3196 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3197 ASSERT(buf->b_efunc == NULL);
3198 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3200 buf->b_efunc = func;
3201 buf->b_private = private;
3205 * Notify the arc that a block was freed, and thus will never be used again.
3208 arc_freed(spa_t *spa, const blkptr_t *bp)
3211 kmutex_t *hash_lock;
3212 uint64_t guid = spa_load_guid(spa);
3214 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3218 if (HDR_BUF_AVAILABLE(hdr)) {
3219 arc_buf_t *buf = hdr->b_buf;
3220 add_reference(hdr, hash_lock, FTAG);
3221 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3222 mutex_exit(hash_lock);
3224 arc_release(buf, FTAG);
3225 (void) arc_buf_remove_ref(buf, FTAG);
3227 mutex_exit(hash_lock);
3233 * This is used by the DMU to let the ARC know that a buffer is
3234 * being evicted, so the ARC should clean up. If this arc buf
3235 * is not yet in the evicted state, it will be put there.
3238 arc_buf_evict(arc_buf_t *buf)
3241 kmutex_t *hash_lock;
3244 mutex_enter(&buf->b_evict_lock);
3248 * We are in arc_do_user_evicts().
3250 ASSERT(buf->b_data == NULL);
3251 mutex_exit(&buf->b_evict_lock);
3253 } else if (buf->b_data == NULL) {
3254 arc_buf_t copy = *buf; /* structure assignment */
3256 * We are on the eviction list; process this buffer now
3257 * but let arc_do_user_evicts() do the reaping.
3259 buf->b_efunc = NULL;
3260 mutex_exit(&buf->b_evict_lock);
3261 VERIFY(copy.b_efunc(©) == 0);
3264 hash_lock = HDR_LOCK(hdr);
3265 mutex_enter(hash_lock);
3267 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3269 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3270 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3273 * Pull this buffer off of the hdr
3276 while (*bufp != buf)
3277 bufp = &(*bufp)->b_next;
3278 *bufp = buf->b_next;
3280 ASSERT(buf->b_data != NULL);
3281 arc_buf_destroy(buf, FALSE, FALSE);
3283 if (hdr->b_datacnt == 0) {
3284 arc_state_t *old_state = hdr->b_state;
3285 arc_state_t *evicted_state;
3287 ASSERT(hdr->b_buf == NULL);
3288 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3291 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3293 mutex_enter(&old_state->arcs_mtx);
3294 mutex_enter(&evicted_state->arcs_mtx);
3296 arc_change_state(evicted_state, hdr, hash_lock);
3297 ASSERT(HDR_IN_HASH_TABLE(hdr));
3298 hdr->b_flags |= ARC_IN_HASH_TABLE;
3299 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3301 mutex_exit(&evicted_state->arcs_mtx);
3302 mutex_exit(&old_state->arcs_mtx);
3304 mutex_exit(hash_lock);
3305 mutex_exit(&buf->b_evict_lock);
3307 VERIFY(buf->b_efunc(buf) == 0);
3308 buf->b_efunc = NULL;
3309 buf->b_private = NULL;
3312 kmem_cache_free(buf_cache, buf);
3317 * Release this buffer from the cache. This must be done
3318 * after a read and prior to modifying the buffer contents.
3319 * If the buffer has more than one reference, we must make
3320 * a new hdr for the buffer.
3323 arc_release(arc_buf_t *buf, void *tag)
3326 kmutex_t *hash_lock = NULL;
3327 l2arc_buf_hdr_t *l2hdr;
3328 uint64_t buf_size = 0;
3331 * It would be nice to assert that if it's DMU metadata (level >
3332 * 0 || it's the dnode file), then it must be syncing context.
3333 * But we don't know that information at this level.
3336 mutex_enter(&buf->b_evict_lock);
3339 /* this buffer is not on any list */
3340 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3342 if (hdr->b_state == arc_anon) {
3343 /* this buffer is already released */
3344 ASSERT(buf->b_efunc == NULL);
3346 hash_lock = HDR_LOCK(hdr);
3347 mutex_enter(hash_lock);
3349 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3352 l2hdr = hdr->b_l2hdr;
3354 mutex_enter(&l2arc_buflist_mtx);
3355 hdr->b_l2hdr = NULL;
3356 buf_size = hdr->b_size;
3360 * Do we have more than one buf?
3362 if (hdr->b_datacnt > 1) {
3363 arc_buf_hdr_t *nhdr;
3365 uint64_t blksz = hdr->b_size;
3366 uint64_t spa = hdr->b_spa;
3367 arc_buf_contents_t type = hdr->b_type;
3368 uint32_t flags = hdr->b_flags;
3370 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3372 * Pull the data off of this hdr and attach it to
3373 * a new anonymous hdr.
3375 (void) remove_reference(hdr, hash_lock, tag);
3377 while (*bufp != buf)
3378 bufp = &(*bufp)->b_next;
3379 *bufp = buf->b_next;
3382 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3383 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3384 if (refcount_is_zero(&hdr->b_refcnt)) {
3385 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3386 ASSERT3U(*size, >=, hdr->b_size);
3387 atomic_add_64(size, -hdr->b_size);
3391 * We're releasing a duplicate user data buffer, update
3392 * our statistics accordingly.
3394 if (hdr->b_type == ARC_BUFC_DATA) {
3395 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3396 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3399 hdr->b_datacnt -= 1;
3400 arc_cksum_verify(buf);
3402 mutex_exit(hash_lock);
3404 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3405 nhdr->b_size = blksz;
3407 nhdr->b_type = type;
3409 nhdr->b_state = arc_anon;
3410 nhdr->b_arc_access = 0;
3411 nhdr->b_flags = flags & ARC_L2_WRITING;
3412 nhdr->b_l2hdr = NULL;
3413 nhdr->b_datacnt = 1;
3414 nhdr->b_freeze_cksum = NULL;
3415 (void) refcount_add(&nhdr->b_refcnt, tag);
3417 mutex_exit(&buf->b_evict_lock);
3418 atomic_add_64(&arc_anon->arcs_size, blksz);
3420 mutex_exit(&buf->b_evict_lock);
3421 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3422 ASSERT(!list_link_active(&hdr->b_arc_node));
3423 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3424 if (hdr->b_state != arc_anon)
3425 arc_change_state(arc_anon, hdr, hash_lock);
3426 hdr->b_arc_access = 0;
3428 mutex_exit(hash_lock);
3430 buf_discard_identity(hdr);
3433 buf->b_efunc = NULL;
3434 buf->b_private = NULL;
3437 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3438 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3439 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
3440 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3441 mutex_exit(&l2arc_buflist_mtx);
3446 arc_released(arc_buf_t *buf)
3450 mutex_enter(&buf->b_evict_lock);
3451 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3452 mutex_exit(&buf->b_evict_lock);
3457 arc_has_callback(arc_buf_t *buf)
3461 mutex_enter(&buf->b_evict_lock);
3462 callback = (buf->b_efunc != NULL);
3463 mutex_exit(&buf->b_evict_lock);
3469 arc_referenced(arc_buf_t *buf)
3473 mutex_enter(&buf->b_evict_lock);
3474 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3475 mutex_exit(&buf->b_evict_lock);
3476 return (referenced);
3481 arc_write_ready(zio_t *zio)
3483 arc_write_callback_t *callback = zio->io_private;
3484 arc_buf_t *buf = callback->awcb_buf;
3485 arc_buf_hdr_t *hdr = buf->b_hdr;
3487 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3488 callback->awcb_ready(zio, buf, callback->awcb_private);
3491 * If the IO is already in progress, then this is a re-write
3492 * attempt, so we need to thaw and re-compute the cksum.
3493 * It is the responsibility of the callback to handle the
3494 * accounting for any re-write attempt.
3496 if (HDR_IO_IN_PROGRESS(hdr)) {
3497 mutex_enter(&hdr->b_freeze_lock);
3498 if (hdr->b_freeze_cksum != NULL) {
3499 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3500 hdr->b_freeze_cksum = NULL;
3502 mutex_exit(&hdr->b_freeze_lock);
3504 arc_cksum_compute(buf, B_FALSE);
3505 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3509 arc_write_done(zio_t *zio)
3511 arc_write_callback_t *callback = zio->io_private;
3512 arc_buf_t *buf = callback->awcb_buf;
3513 arc_buf_hdr_t *hdr = buf->b_hdr;
3515 ASSERT(hdr->b_acb == NULL);
3517 if (zio->io_error == 0) {
3518 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3519 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3520 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3522 ASSERT(BUF_EMPTY(hdr));
3526 * If the block to be written was all-zero, we may have
3527 * compressed it away. In this case no write was performed
3528 * so there will be no dva/birth/checksum. The buffer must
3529 * therefore remain anonymous (and uncached).
3531 if (!BUF_EMPTY(hdr)) {
3532 arc_buf_hdr_t *exists;
3533 kmutex_t *hash_lock;
3535 ASSERT(zio->io_error == 0);
3537 arc_cksum_verify(buf);
3539 exists = buf_hash_insert(hdr, &hash_lock);
3542 * This can only happen if we overwrite for
3543 * sync-to-convergence, because we remove
3544 * buffers from the hash table when we arc_free().
3546 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3547 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3548 panic("bad overwrite, hdr=%p exists=%p",
3549 (void *)hdr, (void *)exists);
3550 ASSERT(refcount_is_zero(&exists->b_refcnt));
3551 arc_change_state(arc_anon, exists, hash_lock);
3552 mutex_exit(hash_lock);
3553 arc_hdr_destroy(exists);
3554 exists = buf_hash_insert(hdr, &hash_lock);
3555 ASSERT3P(exists, ==, NULL);
3558 ASSERT(hdr->b_datacnt == 1);
3559 ASSERT(hdr->b_state == arc_anon);
3560 ASSERT(BP_GET_DEDUP(zio->io_bp));
3561 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3564 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3565 /* if it's not anon, we are doing a scrub */
3566 if (!exists && hdr->b_state == arc_anon)
3567 arc_access(hdr, hash_lock);
3568 mutex_exit(hash_lock);
3570 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3573 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3574 callback->awcb_done(zio, buf, callback->awcb_private);
3576 kmem_free(callback, sizeof (arc_write_callback_t));
3580 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3581 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3582 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3583 int priority, int zio_flags, const zbookmark_t *zb)
3585 arc_buf_hdr_t *hdr = buf->b_hdr;
3586 arc_write_callback_t *callback;
3589 ASSERT(ready != NULL);
3590 ASSERT(done != NULL);
3591 ASSERT(!HDR_IO_ERROR(hdr));
3592 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3593 ASSERT(hdr->b_acb == NULL);
3595 hdr->b_flags |= ARC_L2CACHE;
3596 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_PUSHPAGE);
3597 callback->awcb_ready = ready;
3598 callback->awcb_done = done;
3599 callback->awcb_private = private;
3600 callback->awcb_buf = buf;
3602 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3603 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3609 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3612 uint64_t available_memory;
3614 if (zfs_arc_memory_throttle_disable)
3617 /* Easily reclaimable memory (free + inactive + arc-evictable) */
3618 available_memory = ptob(spl_kmem_availrmem()) + arc_evictable_memory();
3620 if (available_memory <= zfs_write_limit_max) {
3621 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3622 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
3626 if (inflight_data > available_memory / 4) {
3627 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3628 DMU_TX_STAT_BUMP(dmu_tx_memory_inflight);
3636 arc_tempreserve_clear(uint64_t reserve)
3638 atomic_add_64(&arc_tempreserve, -reserve);
3639 ASSERT((int64_t)arc_tempreserve >= 0);
3643 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3650 * Once in a while, fail for no reason. Everything should cope.
3652 if (spa_get_random(10000) == 0) {
3653 dprintf("forcing random failure\n");
3657 if (reserve > arc_c/4 && !arc_no_grow)
3658 arc_c = MIN(arc_c_max, reserve * 4);
3659 if (reserve > arc_c) {
3660 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
3665 * Don't count loaned bufs as in flight dirty data to prevent long
3666 * network delays from blocking transactions that are ready to be
3667 * assigned to a txg.
3669 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3672 * Writes will, almost always, require additional memory allocations
3673 * in order to compress/encrypt/etc the data. We therefor need to
3674 * make sure that there is sufficient available memory for this.
3676 if ((error = arc_memory_throttle(reserve, anon_size, txg)))
3680 * Throttle writes when the amount of dirty data in the cache
3681 * gets too large. We try to keep the cache less than half full
3682 * of dirty blocks so that our sync times don't grow too large.
3683 * Note: if two requests come in concurrently, we might let them
3684 * both succeed, when one of them should fail. Not a huge deal.
3687 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3688 anon_size > arc_c / 4) {
3689 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3690 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3691 arc_tempreserve>>10,
3692 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3693 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3694 reserve>>10, arc_c>>10);
3695 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
3698 atomic_add_64(&arc_tempreserve, reserve);
3703 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
3704 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
3706 size->value.ui64 = state->arcs_size;
3707 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
3708 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
3712 arc_kstat_update(kstat_t *ksp, int rw)
3714 arc_stats_t *as = ksp->ks_data;
3716 if (rw == KSTAT_WRITE) {
3719 arc_kstat_update_state(arc_anon,
3720 &as->arcstat_anon_size,
3721 &as->arcstat_anon_evict_data,
3722 &as->arcstat_anon_evict_metadata);
3723 arc_kstat_update_state(arc_mru,
3724 &as->arcstat_mru_size,
3725 &as->arcstat_mru_evict_data,
3726 &as->arcstat_mru_evict_metadata);
3727 arc_kstat_update_state(arc_mru_ghost,
3728 &as->arcstat_mru_ghost_size,
3729 &as->arcstat_mru_ghost_evict_data,
3730 &as->arcstat_mru_ghost_evict_metadata);
3731 arc_kstat_update_state(arc_mfu,
3732 &as->arcstat_mfu_size,
3733 &as->arcstat_mfu_evict_data,
3734 &as->arcstat_mfu_evict_metadata);
3735 arc_kstat_update_state(arc_mfu_ghost,
3736 &as->arcstat_mfu_ghost_size,
3737 &as->arcstat_mfu_ghost_evict_data,
3738 &as->arcstat_mfu_ghost_evict_metadata);
3747 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3748 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3750 /* Convert seconds to clock ticks */
3751 zfs_arc_min_prefetch_lifespan = 1 * hz;
3753 /* Start out with 1/8 of all memory */
3754 arc_c = physmem * PAGESIZE / 8;
3758 * On architectures where the physical memory can be larger
3759 * than the addressable space (intel in 32-bit mode), we may
3760 * need to limit the cache to 1/8 of VM size.
3762 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3764 * Register a shrinker to support synchronous (direct) memory
3765 * reclaim from the arc. This is done to prevent kswapd from
3766 * swapping out pages when it is preferable to shrink the arc.
3768 spl_register_shrinker(&arc_shrinker);
3771 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3772 arc_c_min = MAX(arc_c / 4, 64<<20);
3773 /* set max to 1/2 of all memory */
3774 arc_c_max = MAX(arc_c * 4, arc_c_max);
3777 * Allow the tunables to override our calculations if they are
3778 * reasonable (ie. over 64MB)
3780 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3781 arc_c_max = zfs_arc_max;
3782 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3783 arc_c_min = zfs_arc_min;
3786 arc_p = (arc_c >> 1);
3788 /* limit meta-data to 1/4 of the arc capacity */
3789 arc_meta_limit = arc_c_max / 4;
3792 /* Allow the tunable to override if it is reasonable */
3793 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3794 arc_meta_limit = zfs_arc_meta_limit;
3796 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3797 arc_c_min = arc_meta_limit / 2;
3799 /* if kmem_flags are set, lets try to use less memory */
3800 if (kmem_debugging())
3802 if (arc_c < arc_c_min)
3805 arc_anon = &ARC_anon;
3807 arc_mru_ghost = &ARC_mru_ghost;
3809 arc_mfu_ghost = &ARC_mfu_ghost;
3810 arc_l2c_only = &ARC_l2c_only;
3813 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3814 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3815 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3816 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3817 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3818 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3820 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3821 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3822 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3823 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3824 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3825 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3826 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3827 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3828 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3829 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3830 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3831 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3832 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3833 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3834 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3835 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3836 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3837 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3838 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3839 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3843 arc_thread_exit = 0;
3844 list_create(&arc_prune_list, sizeof (arc_prune_t),
3845 offsetof(arc_prune_t, p_node));
3846 arc_eviction_list = NULL;
3847 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
3848 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3849 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3851 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3852 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3854 if (arc_ksp != NULL) {
3855 arc_ksp->ks_data = &arc_stats;
3856 arc_ksp->ks_update = arc_kstat_update;
3857 kstat_install(arc_ksp);
3860 (void) thread_create(NULL, 0, arc_adapt_thread, NULL, 0, &p0,
3861 TS_RUN, minclsyspri);
3866 if (zfs_write_limit_max == 0)
3867 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3869 zfs_write_limit_shift = 0;
3870 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3878 mutex_enter(&arc_reclaim_thr_lock);
3880 spl_unregister_shrinker(&arc_shrinker);
3881 #endif /* _KERNEL */
3883 arc_thread_exit = 1;
3884 while (arc_thread_exit != 0)
3885 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3886 mutex_exit(&arc_reclaim_thr_lock);
3892 if (arc_ksp != NULL) {
3893 kstat_delete(arc_ksp);
3897 mutex_enter(&arc_prune_mtx);
3898 while ((p = list_head(&arc_prune_list)) != NULL) {
3899 list_remove(&arc_prune_list, p);
3900 refcount_remove(&p->p_refcnt, &arc_prune_list);
3901 refcount_destroy(&p->p_refcnt);
3902 kmem_free(p, sizeof (*p));
3904 mutex_exit(&arc_prune_mtx);
3906 list_destroy(&arc_prune_list);
3907 mutex_destroy(&arc_prune_mtx);
3908 mutex_destroy(&arc_eviction_mtx);
3909 mutex_destroy(&arc_reclaim_thr_lock);
3910 cv_destroy(&arc_reclaim_thr_cv);
3912 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3913 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3914 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3915 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3916 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3917 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3918 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3919 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3921 mutex_destroy(&arc_anon->arcs_mtx);
3922 mutex_destroy(&arc_mru->arcs_mtx);
3923 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3924 mutex_destroy(&arc_mfu->arcs_mtx);
3925 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3926 mutex_destroy(&arc_l2c_only->arcs_mtx);
3928 mutex_destroy(&zfs_write_limit_lock);
3932 ASSERT(arc_loaned_bytes == 0);
3938 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3939 * It uses dedicated storage devices to hold cached data, which are populated
3940 * using large infrequent writes. The main role of this cache is to boost
3941 * the performance of random read workloads. The intended L2ARC devices
3942 * include short-stroked disks, solid state disks, and other media with
3943 * substantially faster read latency than disk.
3945 * +-----------------------+
3947 * +-----------------------+
3950 * l2arc_feed_thread() arc_read()
3954 * +---------------+ |
3956 * +---------------+ |
3961 * +-------+ +-------+
3963 * | cache | | cache |
3964 * +-------+ +-------+
3965 * +=========+ .-----.
3966 * : L2ARC : |-_____-|
3967 * : devices : | Disks |
3968 * +=========+ `-_____-'
3970 * Read requests are satisfied from the following sources, in order:
3973 * 2) vdev cache of L2ARC devices
3975 * 4) vdev cache of disks
3978 * Some L2ARC device types exhibit extremely slow write performance.
3979 * To accommodate for this there are some significant differences between
3980 * the L2ARC and traditional cache design:
3982 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3983 * the ARC behave as usual, freeing buffers and placing headers on ghost
3984 * lists. The ARC does not send buffers to the L2ARC during eviction as
3985 * this would add inflated write latencies for all ARC memory pressure.
3987 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3988 * It does this by periodically scanning buffers from the eviction-end of
3989 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3990 * not already there. It scans until a headroom of buffers is satisfied,
3991 * which itself is a buffer for ARC eviction. The thread that does this is
3992 * l2arc_feed_thread(), illustrated below; example sizes are included to
3993 * provide a better sense of ratio than this diagram:
3996 * +---------------------+----------+
3997 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3998 * +---------------------+----------+ | o L2ARC eligible
3999 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4000 * +---------------------+----------+ |
4001 * 15.9 Gbytes ^ 32 Mbytes |
4003 * l2arc_feed_thread()
4005 * l2arc write hand <--[oooo]--'
4009 * +==============================+
4010 * L2ARC dev |####|#|###|###| |####| ... |
4011 * +==============================+
4014 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4015 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4016 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4017 * safe to say that this is an uncommon case, since buffers at the end of
4018 * the ARC lists have moved there due to inactivity.
4020 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4021 * then the L2ARC simply misses copying some buffers. This serves as a
4022 * pressure valve to prevent heavy read workloads from both stalling the ARC
4023 * with waits and clogging the L2ARC with writes. This also helps prevent
4024 * the potential for the L2ARC to churn if it attempts to cache content too
4025 * quickly, such as during backups of the entire pool.
4027 * 5. After system boot and before the ARC has filled main memory, there are
4028 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4029 * lists can remain mostly static. Instead of searching from tail of these
4030 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4031 * for eligible buffers, greatly increasing its chance of finding them.
4033 * The L2ARC device write speed is also boosted during this time so that
4034 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4035 * there are no L2ARC reads, and no fear of degrading read performance
4036 * through increased writes.
4038 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4039 * the vdev queue can aggregate them into larger and fewer writes. Each
4040 * device is written to in a rotor fashion, sweeping writes through
4041 * available space then repeating.
4043 * 7. The L2ARC does not store dirty content. It never needs to flush
4044 * write buffers back to disk based storage.
4046 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4047 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4049 * The performance of the L2ARC can be tweaked by a number of tunables, which
4050 * may be necessary for different workloads:
4052 * l2arc_write_max max write bytes per interval
4053 * l2arc_write_boost extra write bytes during device warmup
4054 * l2arc_noprefetch skip caching prefetched buffers
4055 * l2arc_headroom number of max device writes to precache
4056 * l2arc_feed_secs seconds between L2ARC writing
4058 * Tunables may be removed or added as future performance improvements are
4059 * integrated, and also may become zpool properties.
4061 * There are three key functions that control how the L2ARC warms up:
4063 * l2arc_write_eligible() check if a buffer is eligible to cache
4064 * l2arc_write_size() calculate how much to write
4065 * l2arc_write_interval() calculate sleep delay between writes
4067 * These three functions determine what to write, how much, and how quickly
4072 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4075 * A buffer is *not* eligible for the L2ARC if it:
4076 * 1. belongs to a different spa.
4077 * 2. is already cached on the L2ARC.
4078 * 3. has an I/O in progress (it may be an incomplete read).
4079 * 4. is flagged not eligible (zfs property).
4081 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
4082 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
4089 l2arc_write_size(l2arc_dev_t *dev)
4093 size = dev->l2ad_write;
4095 if (arc_warm == B_FALSE)
4096 size += dev->l2ad_boost;
4103 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4105 clock_t interval, next, now;
4108 * If the ARC lists are busy, increase our write rate; if the
4109 * lists are stale, idle back. This is achieved by checking
4110 * how much we previously wrote - if it was more than half of
4111 * what we wanted, schedule the next write much sooner.
4113 if (l2arc_feed_again && wrote > (wanted / 2))
4114 interval = (hz * l2arc_feed_min_ms) / 1000;
4116 interval = hz * l2arc_feed_secs;
4118 now = ddi_get_lbolt();
4119 next = MAX(now, MIN(now + interval, began + interval));
4125 l2arc_hdr_stat_add(void)
4127 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE);
4128 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4132 l2arc_hdr_stat_remove(void)
4134 ARCSTAT_INCR(arcstat_l2_hdr_size, -HDR_SIZE);
4135 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4139 * Cycle through L2ARC devices. This is how L2ARC load balances.
4140 * If a device is returned, this also returns holding the spa config lock.
4142 static l2arc_dev_t *
4143 l2arc_dev_get_next(void)
4145 l2arc_dev_t *first, *next = NULL;
4148 * Lock out the removal of spas (spa_namespace_lock), then removal
4149 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4150 * both locks will be dropped and a spa config lock held instead.
4152 mutex_enter(&spa_namespace_lock);
4153 mutex_enter(&l2arc_dev_mtx);
4155 /* if there are no vdevs, there is nothing to do */
4156 if (l2arc_ndev == 0)
4160 next = l2arc_dev_last;
4162 /* loop around the list looking for a non-faulted vdev */
4164 next = list_head(l2arc_dev_list);
4166 next = list_next(l2arc_dev_list, next);
4168 next = list_head(l2arc_dev_list);
4171 /* if we have come back to the start, bail out */
4174 else if (next == first)
4177 } while (vdev_is_dead(next->l2ad_vdev));
4179 /* if we were unable to find any usable vdevs, return NULL */
4180 if (vdev_is_dead(next->l2ad_vdev))
4183 l2arc_dev_last = next;
4186 mutex_exit(&l2arc_dev_mtx);
4189 * Grab the config lock to prevent the 'next' device from being
4190 * removed while we are writing to it.
4193 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4194 mutex_exit(&spa_namespace_lock);
4200 * Free buffers that were tagged for destruction.
4203 l2arc_do_free_on_write(void)
4206 l2arc_data_free_t *df, *df_prev;
4208 mutex_enter(&l2arc_free_on_write_mtx);
4209 buflist = l2arc_free_on_write;
4211 for (df = list_tail(buflist); df; df = df_prev) {
4212 df_prev = list_prev(buflist, df);
4213 ASSERT(df->l2df_data != NULL);
4214 ASSERT(df->l2df_func != NULL);
4215 df->l2df_func(df->l2df_data, df->l2df_size);
4216 list_remove(buflist, df);
4217 kmem_free(df, sizeof (l2arc_data_free_t));
4220 mutex_exit(&l2arc_free_on_write_mtx);
4224 * A write to a cache device has completed. Update all headers to allow
4225 * reads from these buffers to begin.
4228 l2arc_write_done(zio_t *zio)
4230 l2arc_write_callback_t *cb;
4233 arc_buf_hdr_t *head, *ab, *ab_prev;
4234 l2arc_buf_hdr_t *abl2;
4235 kmutex_t *hash_lock;
4237 cb = zio->io_private;
4239 dev = cb->l2wcb_dev;
4240 ASSERT(dev != NULL);
4241 head = cb->l2wcb_head;
4242 ASSERT(head != NULL);
4243 buflist = dev->l2ad_buflist;
4244 ASSERT(buflist != NULL);
4245 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4246 l2arc_write_callback_t *, cb);
4248 if (zio->io_error != 0)
4249 ARCSTAT_BUMP(arcstat_l2_writes_error);
4251 mutex_enter(&l2arc_buflist_mtx);
4254 * All writes completed, or an error was hit.
4256 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4257 ab_prev = list_prev(buflist, ab);
4259 hash_lock = HDR_LOCK(ab);
4260 if (!mutex_tryenter(hash_lock)) {
4262 * This buffer misses out. It may be in a stage
4263 * of eviction. Its ARC_L2_WRITING flag will be
4264 * left set, denying reads to this buffer.
4266 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4270 if (zio->io_error != 0) {
4272 * Error - drop L2ARC entry.
4274 list_remove(buflist, ab);
4277 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4278 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4279 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4283 * Allow ARC to begin reads to this L2ARC entry.
4285 ab->b_flags &= ~ARC_L2_WRITING;
4287 mutex_exit(hash_lock);
4290 atomic_inc_64(&l2arc_writes_done);
4291 list_remove(buflist, head);
4292 kmem_cache_free(hdr_cache, head);
4293 mutex_exit(&l2arc_buflist_mtx);
4295 l2arc_do_free_on_write();
4297 kmem_free(cb, sizeof (l2arc_write_callback_t));
4301 * A read to a cache device completed. Validate buffer contents before
4302 * handing over to the regular ARC routines.
4305 l2arc_read_done(zio_t *zio)
4307 l2arc_read_callback_t *cb;
4310 kmutex_t *hash_lock;
4313 ASSERT(zio->io_vd != NULL);
4314 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4316 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4318 cb = zio->io_private;
4320 buf = cb->l2rcb_buf;
4321 ASSERT(buf != NULL);
4323 hash_lock = HDR_LOCK(buf->b_hdr);
4324 mutex_enter(hash_lock);
4326 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4329 * Check this survived the L2ARC journey.
4331 equal = arc_cksum_equal(buf);
4332 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4333 mutex_exit(hash_lock);
4334 zio->io_private = buf;
4335 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4336 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4339 mutex_exit(hash_lock);
4341 * Buffer didn't survive caching. Increment stats and
4342 * reissue to the original storage device.
4344 if (zio->io_error != 0) {
4345 ARCSTAT_BUMP(arcstat_l2_io_error);
4347 zio->io_error = EIO;
4350 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4353 * If there's no waiter, issue an async i/o to the primary
4354 * storage now. If there *is* a waiter, the caller must
4355 * issue the i/o in a context where it's OK to block.
4357 if (zio->io_waiter == NULL) {
4358 zio_t *pio = zio_unique_parent(zio);
4360 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4362 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4363 buf->b_data, zio->io_size, arc_read_done, buf,
4364 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4368 kmem_free(cb, sizeof (l2arc_read_callback_t));
4372 * This is the list priority from which the L2ARC will search for pages to
4373 * cache. This is used within loops (0..3) to cycle through lists in the
4374 * desired order. This order can have a significant effect on cache
4377 * Currently the metadata lists are hit first, MFU then MRU, followed by
4378 * the data lists. This function returns a locked list, and also returns
4382 l2arc_list_locked(int list_num, kmutex_t **lock)
4384 list_t *list = NULL;
4386 ASSERT(list_num >= 0 && list_num <= 3);
4390 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4391 *lock = &arc_mfu->arcs_mtx;
4394 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4395 *lock = &arc_mru->arcs_mtx;
4398 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4399 *lock = &arc_mfu->arcs_mtx;
4402 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4403 *lock = &arc_mru->arcs_mtx;
4407 ASSERT(!(MUTEX_HELD(*lock)));
4413 * Evict buffers from the device write hand to the distance specified in
4414 * bytes. This distance may span populated buffers, it may span nothing.
4415 * This is clearing a region on the L2ARC device ready for writing.
4416 * If the 'all' boolean is set, every buffer is evicted.
4419 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4422 l2arc_buf_hdr_t *abl2;
4423 arc_buf_hdr_t *ab, *ab_prev;
4424 kmutex_t *hash_lock;
4427 buflist = dev->l2ad_buflist;
4429 if (buflist == NULL)
4432 if (!all && dev->l2ad_first) {
4434 * This is the first sweep through the device. There is
4440 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4442 * When nearing the end of the device, evict to the end
4443 * before the device write hand jumps to the start.
4445 taddr = dev->l2ad_end;
4447 taddr = dev->l2ad_hand + distance;
4449 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4450 uint64_t, taddr, boolean_t, all);
4453 mutex_enter(&l2arc_buflist_mtx);
4454 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4455 ab_prev = list_prev(buflist, ab);
4457 hash_lock = HDR_LOCK(ab);
4458 if (!mutex_tryenter(hash_lock)) {
4460 * Missed the hash lock. Retry.
4462 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4463 mutex_exit(&l2arc_buflist_mtx);
4464 mutex_enter(hash_lock);
4465 mutex_exit(hash_lock);
4469 if (HDR_L2_WRITE_HEAD(ab)) {
4471 * We hit a write head node. Leave it for
4472 * l2arc_write_done().
4474 list_remove(buflist, ab);
4475 mutex_exit(hash_lock);
4479 if (!all && ab->b_l2hdr != NULL &&
4480 (ab->b_l2hdr->b_daddr > taddr ||
4481 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4483 * We've evicted to the target address,
4484 * or the end of the device.
4486 mutex_exit(hash_lock);
4490 if (HDR_FREE_IN_PROGRESS(ab)) {
4492 * Already on the path to destruction.
4494 mutex_exit(hash_lock);
4498 if (ab->b_state == arc_l2c_only) {
4499 ASSERT(!HDR_L2_READING(ab));
4501 * This doesn't exist in the ARC. Destroy.
4502 * arc_hdr_destroy() will call list_remove()
4503 * and decrement arcstat_l2_size.
4505 arc_change_state(arc_anon, ab, hash_lock);
4506 arc_hdr_destroy(ab);
4509 * Invalidate issued or about to be issued
4510 * reads, since we may be about to write
4511 * over this location.
4513 if (HDR_L2_READING(ab)) {
4514 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4515 ab->b_flags |= ARC_L2_EVICTED;
4519 * Tell ARC this no longer exists in L2ARC.
4521 if (ab->b_l2hdr != NULL) {
4524 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4525 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4526 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4528 list_remove(buflist, ab);
4531 * This may have been leftover after a
4534 ab->b_flags &= ~ARC_L2_WRITING;
4536 mutex_exit(hash_lock);
4538 mutex_exit(&l2arc_buflist_mtx);
4540 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4541 dev->l2ad_evict = taddr;
4545 * Find and write ARC buffers to the L2ARC device.
4547 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4548 * for reading until they have completed writing.
4551 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4553 arc_buf_hdr_t *ab, *ab_prev, *head;
4554 l2arc_buf_hdr_t *hdrl2;
4556 uint64_t passed_sz, write_sz, buf_sz, headroom;
4558 kmutex_t *hash_lock, *list_lock = NULL;
4559 boolean_t have_lock, full;
4560 l2arc_write_callback_t *cb;
4562 uint64_t guid = spa_load_guid(spa);
4565 ASSERT(dev->l2ad_vdev != NULL);
4570 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4571 head->b_flags |= ARC_L2_WRITE_HEAD;
4574 * Copy buffers for L2ARC writing.
4576 mutex_enter(&l2arc_buflist_mtx);
4577 for (try = 0; try <= 3; try++) {
4578 list = l2arc_list_locked(try, &list_lock);
4582 * L2ARC fast warmup.
4584 * Until the ARC is warm and starts to evict, read from the
4585 * head of the ARC lists rather than the tail.
4587 headroom = target_sz * l2arc_headroom;
4588 if (arc_warm == B_FALSE)
4589 ab = list_head(list);
4591 ab = list_tail(list);
4593 for (; ab; ab = ab_prev) {
4594 if (arc_warm == B_FALSE)
4595 ab_prev = list_next(list, ab);
4597 ab_prev = list_prev(list, ab);
4599 hash_lock = HDR_LOCK(ab);
4600 have_lock = MUTEX_HELD(hash_lock);
4601 if (!have_lock && !mutex_tryenter(hash_lock)) {
4603 * Skip this buffer rather than waiting.
4608 passed_sz += ab->b_size;
4609 if (passed_sz > headroom) {
4613 mutex_exit(hash_lock);
4617 if (!l2arc_write_eligible(guid, ab)) {
4618 mutex_exit(hash_lock);
4622 if ((write_sz + ab->b_size) > target_sz) {
4624 mutex_exit(hash_lock);
4630 * Insert a dummy header on the buflist so
4631 * l2arc_write_done() can find where the
4632 * write buffers begin without searching.
4634 list_insert_head(dev->l2ad_buflist, head);
4636 cb = kmem_alloc(sizeof (l2arc_write_callback_t),
4638 cb->l2wcb_dev = dev;
4639 cb->l2wcb_head = head;
4640 pio = zio_root(spa, l2arc_write_done, cb,
4645 * Create and add a new L2ARC header.
4647 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t),
4650 hdrl2->b_daddr = dev->l2ad_hand;
4651 arc_space_consume(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4653 ab->b_flags |= ARC_L2_WRITING;
4654 ab->b_l2hdr = hdrl2;
4655 list_insert_head(dev->l2ad_buflist, ab);
4656 buf_data = ab->b_buf->b_data;
4657 buf_sz = ab->b_size;
4660 * Compute and store the buffer cksum before
4661 * writing. On debug the cksum is verified first.
4663 arc_cksum_verify(ab->b_buf);
4664 arc_cksum_compute(ab->b_buf, B_TRUE);
4666 mutex_exit(hash_lock);
4668 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4669 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4670 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4671 ZIO_FLAG_CANFAIL, B_FALSE);
4673 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4675 (void) zio_nowait(wzio);
4678 * Keep the clock hand suitably device-aligned.
4680 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4683 dev->l2ad_hand += buf_sz;
4686 mutex_exit(list_lock);
4691 mutex_exit(&l2arc_buflist_mtx);
4695 kmem_cache_free(hdr_cache, head);
4699 ASSERT3U(write_sz, <=, target_sz);
4700 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4701 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4702 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4703 vdev_space_update(dev->l2ad_vdev, write_sz, 0, 0);
4706 * Bump device hand to the device start if it is approaching the end.
4707 * l2arc_evict() will already have evicted ahead for this case.
4709 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4710 vdev_space_update(dev->l2ad_vdev,
4711 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4712 dev->l2ad_hand = dev->l2ad_start;
4713 dev->l2ad_evict = dev->l2ad_start;
4714 dev->l2ad_first = B_FALSE;
4717 dev->l2ad_writing = B_TRUE;
4718 (void) zio_wait(pio);
4719 dev->l2ad_writing = B_FALSE;
4725 * This thread feeds the L2ARC at regular intervals. This is the beating
4726 * heart of the L2ARC.
4729 l2arc_feed_thread(void)
4734 uint64_t size, wrote;
4735 clock_t begin, next = ddi_get_lbolt();
4737 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4739 mutex_enter(&l2arc_feed_thr_lock);
4741 while (l2arc_thread_exit == 0) {
4742 CALLB_CPR_SAFE_BEGIN(&cpr);
4743 (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv,
4744 &l2arc_feed_thr_lock, next);
4745 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4746 next = ddi_get_lbolt() + hz;
4749 * Quick check for L2ARC devices.
4751 mutex_enter(&l2arc_dev_mtx);
4752 if (l2arc_ndev == 0) {
4753 mutex_exit(&l2arc_dev_mtx);
4756 mutex_exit(&l2arc_dev_mtx);
4757 begin = ddi_get_lbolt();
4760 * This selects the next l2arc device to write to, and in
4761 * doing so the next spa to feed from: dev->l2ad_spa. This
4762 * will return NULL if there are now no l2arc devices or if
4763 * they are all faulted.
4765 * If a device is returned, its spa's config lock is also
4766 * held to prevent device removal. l2arc_dev_get_next()
4767 * will grab and release l2arc_dev_mtx.
4769 if ((dev = l2arc_dev_get_next()) == NULL)
4772 spa = dev->l2ad_spa;
4773 ASSERT(spa != NULL);
4776 * If the pool is read-only then force the feed thread to
4777 * sleep a little longer.
4779 if (!spa_writeable(spa)) {
4780 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4781 spa_config_exit(spa, SCL_L2ARC, dev);
4786 * Avoid contributing to memory pressure.
4789 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4790 spa_config_exit(spa, SCL_L2ARC, dev);
4794 ARCSTAT_BUMP(arcstat_l2_feeds);
4796 size = l2arc_write_size(dev);
4799 * Evict L2ARC buffers that will be overwritten.
4801 l2arc_evict(dev, size, B_FALSE);
4804 * Write ARC buffers.
4806 wrote = l2arc_write_buffers(spa, dev, size);
4809 * Calculate interval between writes.
4811 next = l2arc_write_interval(begin, size, wrote);
4812 spa_config_exit(spa, SCL_L2ARC, dev);
4815 l2arc_thread_exit = 0;
4816 cv_broadcast(&l2arc_feed_thr_cv);
4817 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4822 l2arc_vdev_present(vdev_t *vd)
4826 mutex_enter(&l2arc_dev_mtx);
4827 for (dev = list_head(l2arc_dev_list); dev != NULL;
4828 dev = list_next(l2arc_dev_list, dev)) {
4829 if (dev->l2ad_vdev == vd)
4832 mutex_exit(&l2arc_dev_mtx);
4834 return (dev != NULL);
4838 * Add a vdev for use by the L2ARC. By this point the spa has already
4839 * validated the vdev and opened it.
4842 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
4844 l2arc_dev_t *adddev;
4846 ASSERT(!l2arc_vdev_present(vd));
4849 * Create a new l2arc device entry.
4851 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4852 adddev->l2ad_spa = spa;
4853 adddev->l2ad_vdev = vd;
4854 adddev->l2ad_write = l2arc_write_max;
4855 adddev->l2ad_boost = l2arc_write_boost;
4856 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
4857 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
4858 adddev->l2ad_hand = adddev->l2ad_start;
4859 adddev->l2ad_evict = adddev->l2ad_start;
4860 adddev->l2ad_first = B_TRUE;
4861 adddev->l2ad_writing = B_FALSE;
4862 list_link_init(&adddev->l2ad_node);
4863 ASSERT3U(adddev->l2ad_write, >, 0);
4866 * This is a list of all ARC buffers that are still valid on the
4869 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4870 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4871 offsetof(arc_buf_hdr_t, b_l2node));
4873 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
4876 * Add device to global list
4878 mutex_enter(&l2arc_dev_mtx);
4879 list_insert_head(l2arc_dev_list, adddev);
4880 atomic_inc_64(&l2arc_ndev);
4881 mutex_exit(&l2arc_dev_mtx);
4885 * Remove a vdev from the L2ARC.
4888 l2arc_remove_vdev(vdev_t *vd)
4890 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4893 * Find the device by vdev
4895 mutex_enter(&l2arc_dev_mtx);
4896 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4897 nextdev = list_next(l2arc_dev_list, dev);
4898 if (vd == dev->l2ad_vdev) {
4903 ASSERT(remdev != NULL);
4906 * Remove device from global list
4908 list_remove(l2arc_dev_list, remdev);
4909 l2arc_dev_last = NULL; /* may have been invalidated */
4910 atomic_dec_64(&l2arc_ndev);
4911 mutex_exit(&l2arc_dev_mtx);
4914 * Clear all buflists and ARC references. L2ARC device flush.
4916 l2arc_evict(remdev, 0, B_TRUE);
4917 list_destroy(remdev->l2ad_buflist);
4918 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
4919 kmem_free(remdev, sizeof (l2arc_dev_t));
4925 l2arc_thread_exit = 0;
4927 l2arc_writes_sent = 0;
4928 l2arc_writes_done = 0;
4930 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4931 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
4932 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
4933 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
4934 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
4936 l2arc_dev_list = &L2ARC_dev_list;
4937 l2arc_free_on_write = &L2ARC_free_on_write;
4938 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
4939 offsetof(l2arc_dev_t, l2ad_node));
4940 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
4941 offsetof(l2arc_data_free_t, l2df_list_node));
4948 * This is called from dmu_fini(), which is called from spa_fini();
4949 * Because of this, we can assume that all l2arc devices have
4950 * already been removed when the pools themselves were removed.
4953 l2arc_do_free_on_write();
4955 mutex_destroy(&l2arc_feed_thr_lock);
4956 cv_destroy(&l2arc_feed_thr_cv);
4957 mutex_destroy(&l2arc_dev_mtx);
4958 mutex_destroy(&l2arc_buflist_mtx);
4959 mutex_destroy(&l2arc_free_on_write_mtx);
4961 list_destroy(l2arc_dev_list);
4962 list_destroy(l2arc_free_on_write);
4968 if (!(spa_mode_global & FWRITE))
4971 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
4972 TS_RUN, minclsyspri);
4978 if (!(spa_mode_global & FWRITE))
4981 mutex_enter(&l2arc_feed_thr_lock);
4982 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
4983 l2arc_thread_exit = 1;
4984 while (l2arc_thread_exit != 0)
4985 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
4986 mutex_exit(&l2arc_feed_thr_lock);
4989 #if defined(_KERNEL) && defined(HAVE_SPL)
4990 EXPORT_SYMBOL(arc_read);
4991 EXPORT_SYMBOL(arc_buf_remove_ref);
4992 EXPORT_SYMBOL(arc_getbuf_func);
4993 EXPORT_SYMBOL(arc_add_prune_callback);
4994 EXPORT_SYMBOL(arc_remove_prune_callback);
4996 module_param(zfs_arc_min, ulong, 0644);
4997 MODULE_PARM_DESC(zfs_arc_min, "Min arc size");
4999 module_param(zfs_arc_max, ulong, 0644);
5000 MODULE_PARM_DESC(zfs_arc_max, "Max arc size");
5002 module_param(zfs_arc_meta_limit, ulong, 0644);
5003 MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size");
5005 module_param(zfs_arc_meta_prune, int, 0644);
5006 MODULE_PARM_DESC(zfs_arc_meta_prune, "Bytes of meta data to prune");
5008 module_param(zfs_arc_grow_retry, int, 0644);
5009 MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size");
5011 module_param(zfs_arc_shrink_shift, int, 0644);
5012 MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)");
5014 module_param(zfs_arc_p_min_shift, int, 0644);
5015 MODULE_PARM_DESC(zfs_arc_p_min_shift, "arc_c shift to calc min/max arc_p");
5017 module_param(zfs_disable_dup_eviction, int, 0644);
5018 MODULE_PARM_DESC(zfs_disable_dup_eviction, "disable duplicate buffer eviction");
5020 module_param(zfs_arc_memory_throttle_disable, int, 0644);
5021 MODULE_PARM_DESC(zfs_arc_memory_throttle_disable, "disable memory throttle");
5023 module_param(zfs_arc_min_prefetch_lifespan, int, 0644);
5024 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan, "Min life of prefetch block");
5026 module_param(l2arc_write_max, ulong, 0644);
5027 MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval");
5029 module_param(l2arc_write_boost, ulong, 0644);
5030 MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup");
5032 module_param(l2arc_headroom, ulong, 0644);
5033 MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache");
5035 module_param(l2arc_feed_secs, ulong, 0644);
5036 MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing");
5038 module_param(l2arc_feed_min_ms, ulong, 0644);
5039 MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds");
5041 module_param(l2arc_noprefetch, int, 0644);
5042 MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers");
5044 module_param(l2arc_feed_again, int, 0644);
5045 MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup");
5047 module_param(l2arc_norw, int, 0644);
5048 MODULE_PARM_DESC(l2arc_norw, "No reads during writes");