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.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
29 * DVA-based Adjustable Replacement Cache
31 * While much of the theory of operation used here is
32 * based on the self-tuning, low overhead replacement cache
33 * presented by Megiddo and Modha at FAST 2003, there are some
34 * significant differences:
36 * 1. The Megiddo and Modha model assumes any page is evictable.
37 * Pages in its cache cannot be "locked" into memory. This makes
38 * the eviction algorithm simple: evict the last page in the list.
39 * This also make the performance characteristics easy to reason
40 * about. Our cache is not so simple. At any given moment, some
41 * subset of the blocks in the cache are un-evictable because we
42 * have handed out a reference to them. Blocks are only evictable
43 * when there are no external references active. This makes
44 * eviction far more problematic: we choose to evict the evictable
45 * blocks that are the "lowest" in the list.
47 * There are times when it is not possible to evict the requested
48 * space. In these circumstances we are unable to adjust the cache
49 * size. To prevent the cache growing unbounded at these times we
50 * implement a "cache throttle" that slows the flow of new data
51 * into the cache until we can make space available.
53 * 2. The Megiddo and Modha model assumes a fixed cache size.
54 * Pages are evicted when the cache is full and there is a cache
55 * miss. Our model has a variable sized cache. It grows with
56 * high use, but also tries to react to memory pressure from the
57 * operating system: decreasing its size when system memory is
60 * 3. The Megiddo and Modha model assumes a fixed page size. All
61 * elements of the cache are therefor exactly the same size. So
62 * when adjusting the cache size following a cache miss, its simply
63 * a matter of choosing a single page to evict. In our model, we
64 * have variable sized cache blocks (rangeing from 512 bytes to
65 * 128K bytes). We therefor choose a set of blocks to evict to make
66 * space for a cache miss that approximates as closely as possible
67 * the space used by the new block.
69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
70 * by N. Megiddo & D. Modha, FAST 2003
76 * A new reference to a cache buffer can be obtained in two
77 * ways: 1) via a hash table lookup using the DVA as a key,
78 * or 2) via one of the ARC lists. The arc_read() interface
79 * uses method 1, while the internal arc algorithms for
80 * adjusting the cache use method 2. We therefor provide two
81 * types of locks: 1) the hash table lock array, and 2) the
84 * Buffers do not have their own mutexes, rather they rely on the
85 * hash table mutexes for the bulk of their protection (i.e. most
86 * fields in the arc_buf_hdr_t are protected by these mutexes).
88 * buf_hash_find() returns the appropriate mutex (held) when it
89 * locates the requested buffer in the hash table. It returns
90 * NULL for the mutex if the buffer was not in the table.
92 * buf_hash_remove() expects the appropriate hash mutex to be
93 * already held before it is invoked.
95 * Each arc state also has a mutex which is used to protect the
96 * buffer list associated with the state. When attempting to
97 * obtain a hash table lock while holding an arc list lock you
98 * must use: mutex_tryenter() to avoid deadlock. Also note that
99 * the active state mutex must be held before the ghost state mutex.
101 * Arc buffers may have an associated eviction callback function.
102 * This function will be invoked prior to removing the buffer (e.g.
103 * in arc_do_user_evicts()). Note however that the data associated
104 * with the buffer may be evicted prior to the callback. The callback
105 * must be made with *no locks held* (to prevent deadlock). Additionally,
106 * the users of callbacks must ensure that their private data is
107 * protected from simultaneous callbacks from arc_buf_evict()
108 * and arc_do_user_evicts().
110 * It as also possible to register a callback which is run when the
111 * arc_meta_limit is reached and no buffers can be safely evicted. In
112 * this case the arc user should drop a reference on some arc buffers so
113 * they can be reclaimed and the arc_meta_limit honored. For example,
114 * when using the ZPL each dentry holds a references on a znode. These
115 * dentries must be pruned before the arc buffer holding the znode can
118 * Note that the majority of the performance stats are manipulated
119 * with atomic operations.
121 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
123 * - L2ARC buflist creation
124 * - L2ARC buflist eviction
125 * - L2ARC write completion, which walks L2ARC buflists
126 * - ARC header destruction, as it removes from L2ARC buflists
127 * - ARC header release, as it removes from L2ARC buflists
132 #include <sys/zio_compress.h>
133 #include <sys/zfs_context.h>
135 #include <sys/vdev.h>
136 #include <sys/vdev_impl.h>
138 #include <sys/vmsystm.h>
140 #include <sys/fs/swapnode.h>
143 #include <sys/callb.h>
144 #include <sys/kstat.h>
145 #include <sys/dmu_tx.h>
146 #include <zfs_fletcher.h>
148 static kmutex_t arc_reclaim_thr_lock;
149 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
150 static uint8_t arc_thread_exit;
152 /* number of bytes to prune from caches when at arc_meta_limit is reached */
153 int zfs_arc_meta_prune = 1048576;
155 typedef enum arc_reclaim_strategy {
156 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
157 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
158 } arc_reclaim_strategy_t;
160 /* number of seconds before growing cache again */
161 int zfs_arc_grow_retry = 5;
163 /* shift of arc_c for calculating both min and max arc_p */
164 int zfs_arc_p_min_shift = 4;
166 /* log2(fraction of arc to reclaim) */
167 int zfs_arc_shrink_shift = 5;
170 * minimum lifespan of a prefetch block in clock ticks
171 * (initialized in arc_init())
173 int zfs_arc_min_prefetch_lifespan = HZ;
175 /* disable arc proactive arc throttle due to low memory */
176 int zfs_arc_memory_throttle_disable = 1;
178 /* disable duplicate buffer eviction */
179 int zfs_disable_dup_eviction = 0;
183 /* expiration time for arc_no_grow */
184 static clock_t arc_grow_time = 0;
187 * The arc has filled available memory and has now warmed up.
189 static boolean_t arc_warm;
192 * These tunables are for performance analysis.
194 unsigned long zfs_arc_max = 0;
195 unsigned long zfs_arc_min = 0;
196 unsigned long zfs_arc_meta_limit = 0;
199 * Note that buffers can be in one of 6 states:
200 * ARC_anon - anonymous (discussed below)
201 * ARC_mru - recently used, currently cached
202 * ARC_mru_ghost - recentely used, no longer in cache
203 * ARC_mfu - frequently used, currently cached
204 * ARC_mfu_ghost - frequently used, no longer in cache
205 * ARC_l2c_only - exists in L2ARC but not other states
206 * When there are no active references to the buffer, they are
207 * are linked onto a list in one of these arc states. These are
208 * the only buffers that can be evicted or deleted. Within each
209 * state there are multiple lists, one for meta-data and one for
210 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
211 * etc.) is tracked separately so that it can be managed more
212 * explicitly: favored over data, limited explicitly.
214 * Anonymous buffers are buffers that are not associated with
215 * a DVA. These are buffers that hold dirty block copies
216 * before they are written to stable storage. By definition,
217 * they are "ref'd" and are considered part of arc_mru
218 * that cannot be freed. Generally, they will aquire a DVA
219 * as they are written and migrate onto the arc_mru list.
221 * The ARC_l2c_only state is for buffers that are in the second
222 * level ARC but no longer in any of the ARC_m* lists. The second
223 * level ARC itself may also contain buffers that are in any of
224 * the ARC_m* states - meaning that a buffer can exist in two
225 * places. The reason for the ARC_l2c_only state is to keep the
226 * buffer header in the hash table, so that reads that hit the
227 * second level ARC benefit from these fast lookups.
230 typedef struct arc_state {
231 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
232 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
233 uint64_t arcs_size; /* total amount of data in this state */
238 static arc_state_t ARC_anon;
239 static arc_state_t ARC_mru;
240 static arc_state_t ARC_mru_ghost;
241 static arc_state_t ARC_mfu;
242 static arc_state_t ARC_mfu_ghost;
243 static arc_state_t ARC_l2c_only;
245 typedef struct arc_stats {
246 kstat_named_t arcstat_hits;
247 kstat_named_t arcstat_misses;
248 kstat_named_t arcstat_demand_data_hits;
249 kstat_named_t arcstat_demand_data_misses;
250 kstat_named_t arcstat_demand_metadata_hits;
251 kstat_named_t arcstat_demand_metadata_misses;
252 kstat_named_t arcstat_prefetch_data_hits;
253 kstat_named_t arcstat_prefetch_data_misses;
254 kstat_named_t arcstat_prefetch_metadata_hits;
255 kstat_named_t arcstat_prefetch_metadata_misses;
256 kstat_named_t arcstat_mru_hits;
257 kstat_named_t arcstat_mru_ghost_hits;
258 kstat_named_t arcstat_mfu_hits;
259 kstat_named_t arcstat_mfu_ghost_hits;
260 kstat_named_t arcstat_deleted;
261 kstat_named_t arcstat_recycle_miss;
262 kstat_named_t arcstat_mutex_miss;
263 kstat_named_t arcstat_evict_skip;
264 kstat_named_t arcstat_evict_l2_cached;
265 kstat_named_t arcstat_evict_l2_eligible;
266 kstat_named_t arcstat_evict_l2_ineligible;
267 kstat_named_t arcstat_hash_elements;
268 kstat_named_t arcstat_hash_elements_max;
269 kstat_named_t arcstat_hash_collisions;
270 kstat_named_t arcstat_hash_chains;
271 kstat_named_t arcstat_hash_chain_max;
272 kstat_named_t arcstat_p;
273 kstat_named_t arcstat_c;
274 kstat_named_t arcstat_c_min;
275 kstat_named_t arcstat_c_max;
276 kstat_named_t arcstat_size;
277 kstat_named_t arcstat_hdr_size;
278 kstat_named_t arcstat_data_size;
279 kstat_named_t arcstat_other_size;
280 kstat_named_t arcstat_anon_size;
281 kstat_named_t arcstat_anon_evict_data;
282 kstat_named_t arcstat_anon_evict_metadata;
283 kstat_named_t arcstat_mru_size;
284 kstat_named_t arcstat_mru_evict_data;
285 kstat_named_t arcstat_mru_evict_metadata;
286 kstat_named_t arcstat_mru_ghost_size;
287 kstat_named_t arcstat_mru_ghost_evict_data;
288 kstat_named_t arcstat_mru_ghost_evict_metadata;
289 kstat_named_t arcstat_mfu_size;
290 kstat_named_t arcstat_mfu_evict_data;
291 kstat_named_t arcstat_mfu_evict_metadata;
292 kstat_named_t arcstat_mfu_ghost_size;
293 kstat_named_t arcstat_mfu_ghost_evict_data;
294 kstat_named_t arcstat_mfu_ghost_evict_metadata;
295 kstat_named_t arcstat_l2_hits;
296 kstat_named_t arcstat_l2_misses;
297 kstat_named_t arcstat_l2_feeds;
298 kstat_named_t arcstat_l2_rw_clash;
299 kstat_named_t arcstat_l2_read_bytes;
300 kstat_named_t arcstat_l2_write_bytes;
301 kstat_named_t arcstat_l2_writes_sent;
302 kstat_named_t arcstat_l2_writes_done;
303 kstat_named_t arcstat_l2_writes_error;
304 kstat_named_t arcstat_l2_writes_hdr_miss;
305 kstat_named_t arcstat_l2_evict_lock_retry;
306 kstat_named_t arcstat_l2_evict_reading;
307 kstat_named_t arcstat_l2_free_on_write;
308 kstat_named_t arcstat_l2_abort_lowmem;
309 kstat_named_t arcstat_l2_cksum_bad;
310 kstat_named_t arcstat_l2_io_error;
311 kstat_named_t arcstat_l2_size;
312 kstat_named_t arcstat_l2_asize;
313 kstat_named_t arcstat_l2_hdr_size;
314 kstat_named_t arcstat_l2_compress_successes;
315 kstat_named_t arcstat_l2_compress_zeros;
316 kstat_named_t arcstat_l2_compress_failures;
317 kstat_named_t arcstat_memory_throttle_count;
318 kstat_named_t arcstat_duplicate_buffers;
319 kstat_named_t arcstat_duplicate_buffers_size;
320 kstat_named_t arcstat_duplicate_reads;
321 kstat_named_t arcstat_memory_direct_count;
322 kstat_named_t arcstat_memory_indirect_count;
323 kstat_named_t arcstat_no_grow;
324 kstat_named_t arcstat_tempreserve;
325 kstat_named_t arcstat_loaned_bytes;
326 kstat_named_t arcstat_prune;
327 kstat_named_t arcstat_meta_used;
328 kstat_named_t arcstat_meta_limit;
329 kstat_named_t arcstat_meta_max;
332 static arc_stats_t arc_stats = {
333 { "hits", KSTAT_DATA_UINT64 },
334 { "misses", KSTAT_DATA_UINT64 },
335 { "demand_data_hits", KSTAT_DATA_UINT64 },
336 { "demand_data_misses", KSTAT_DATA_UINT64 },
337 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
338 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
339 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
340 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
341 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
342 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
343 { "mru_hits", KSTAT_DATA_UINT64 },
344 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
345 { "mfu_hits", KSTAT_DATA_UINT64 },
346 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
347 { "deleted", KSTAT_DATA_UINT64 },
348 { "recycle_miss", KSTAT_DATA_UINT64 },
349 { "mutex_miss", KSTAT_DATA_UINT64 },
350 { "evict_skip", KSTAT_DATA_UINT64 },
351 { "evict_l2_cached", KSTAT_DATA_UINT64 },
352 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
353 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
354 { "hash_elements", KSTAT_DATA_UINT64 },
355 { "hash_elements_max", KSTAT_DATA_UINT64 },
356 { "hash_collisions", KSTAT_DATA_UINT64 },
357 { "hash_chains", KSTAT_DATA_UINT64 },
358 { "hash_chain_max", KSTAT_DATA_UINT64 },
359 { "p", KSTAT_DATA_UINT64 },
360 { "c", KSTAT_DATA_UINT64 },
361 { "c_min", KSTAT_DATA_UINT64 },
362 { "c_max", KSTAT_DATA_UINT64 },
363 { "size", KSTAT_DATA_UINT64 },
364 { "hdr_size", KSTAT_DATA_UINT64 },
365 { "data_size", KSTAT_DATA_UINT64 },
366 { "other_size", KSTAT_DATA_UINT64 },
367 { "anon_size", KSTAT_DATA_UINT64 },
368 { "anon_evict_data", KSTAT_DATA_UINT64 },
369 { "anon_evict_metadata", KSTAT_DATA_UINT64 },
370 { "mru_size", KSTAT_DATA_UINT64 },
371 { "mru_evict_data", KSTAT_DATA_UINT64 },
372 { "mru_evict_metadata", KSTAT_DATA_UINT64 },
373 { "mru_ghost_size", KSTAT_DATA_UINT64 },
374 { "mru_ghost_evict_data", KSTAT_DATA_UINT64 },
375 { "mru_ghost_evict_metadata", KSTAT_DATA_UINT64 },
376 { "mfu_size", KSTAT_DATA_UINT64 },
377 { "mfu_evict_data", KSTAT_DATA_UINT64 },
378 { "mfu_evict_metadata", KSTAT_DATA_UINT64 },
379 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
380 { "mfu_ghost_evict_data", KSTAT_DATA_UINT64 },
381 { "mfu_ghost_evict_metadata", KSTAT_DATA_UINT64 },
382 { "l2_hits", KSTAT_DATA_UINT64 },
383 { "l2_misses", KSTAT_DATA_UINT64 },
384 { "l2_feeds", KSTAT_DATA_UINT64 },
385 { "l2_rw_clash", KSTAT_DATA_UINT64 },
386 { "l2_read_bytes", KSTAT_DATA_UINT64 },
387 { "l2_write_bytes", KSTAT_DATA_UINT64 },
388 { "l2_writes_sent", KSTAT_DATA_UINT64 },
389 { "l2_writes_done", KSTAT_DATA_UINT64 },
390 { "l2_writes_error", KSTAT_DATA_UINT64 },
391 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
392 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
393 { "l2_evict_reading", KSTAT_DATA_UINT64 },
394 { "l2_free_on_write", KSTAT_DATA_UINT64 },
395 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
396 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
397 { "l2_io_error", KSTAT_DATA_UINT64 },
398 { "l2_size", KSTAT_DATA_UINT64 },
399 { "l2_asize", KSTAT_DATA_UINT64 },
400 { "l2_hdr_size", KSTAT_DATA_UINT64 },
401 { "l2_compress_successes", KSTAT_DATA_UINT64 },
402 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
403 { "l2_compress_failures", KSTAT_DATA_UINT64 },
404 { "memory_throttle_count", KSTAT_DATA_UINT64 },
405 { "duplicate_buffers", KSTAT_DATA_UINT64 },
406 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
407 { "duplicate_reads", KSTAT_DATA_UINT64 },
408 { "memory_direct_count", KSTAT_DATA_UINT64 },
409 { "memory_indirect_count", KSTAT_DATA_UINT64 },
410 { "arc_no_grow", KSTAT_DATA_UINT64 },
411 { "arc_tempreserve", KSTAT_DATA_UINT64 },
412 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
413 { "arc_prune", KSTAT_DATA_UINT64 },
414 { "arc_meta_used", KSTAT_DATA_UINT64 },
415 { "arc_meta_limit", KSTAT_DATA_UINT64 },
416 { "arc_meta_max", KSTAT_DATA_UINT64 },
419 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
421 #define ARCSTAT_INCR(stat, val) \
422 atomic_add_64(&arc_stats.stat.value.ui64, (val));
424 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
425 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
427 #define ARCSTAT_MAX(stat, val) { \
429 while ((val) > (m = arc_stats.stat.value.ui64) && \
430 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
434 #define ARCSTAT_MAXSTAT(stat) \
435 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
438 * We define a macro to allow ARC hits/misses to be easily broken down by
439 * two separate conditions, giving a total of four different subtypes for
440 * each of hits and misses (so eight statistics total).
442 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
445 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
447 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
451 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
453 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
458 static arc_state_t *arc_anon;
459 static arc_state_t *arc_mru;
460 static arc_state_t *arc_mru_ghost;
461 static arc_state_t *arc_mfu;
462 static arc_state_t *arc_mfu_ghost;
463 static arc_state_t *arc_l2c_only;
466 * There are several ARC variables that are critical to export as kstats --
467 * but we don't want to have to grovel around in the kstat whenever we wish to
468 * manipulate them. For these variables, we therefore define them to be in
469 * terms of the statistic variable. This assures that we are not introducing
470 * the possibility of inconsistency by having shadow copies of the variables,
471 * while still allowing the code to be readable.
473 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
474 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
475 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
476 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
477 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
478 #define arc_no_grow ARCSTAT(arcstat_no_grow)
479 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
480 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
481 #define arc_meta_used ARCSTAT(arcstat_meta_used)
482 #define arc_meta_limit ARCSTAT(arcstat_meta_limit)
483 #define arc_meta_max ARCSTAT(arcstat_meta_max)
485 #define L2ARC_IS_VALID_COMPRESS(_c_) \
486 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
488 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
490 typedef struct arc_callback arc_callback_t;
492 struct arc_callback {
494 arc_done_func_t *acb_done;
496 zio_t *acb_zio_dummy;
497 arc_callback_t *acb_next;
500 typedef struct arc_write_callback arc_write_callback_t;
502 struct arc_write_callback {
504 arc_done_func_t *awcb_ready;
505 arc_done_func_t *awcb_done;
510 /* protected by hash lock */
515 kmutex_t b_freeze_lock;
516 zio_cksum_t *b_freeze_cksum;
518 arc_buf_hdr_t *b_hash_next;
523 arc_callback_t *b_acb;
527 arc_buf_contents_t b_type;
531 /* protected by arc state mutex */
532 arc_state_t *b_state;
533 list_node_t b_arc_node;
535 /* updated atomically */
536 clock_t b_arc_access;
538 /* self protecting */
541 l2arc_buf_hdr_t *b_l2hdr;
542 list_node_t b_l2node;
545 static list_t arc_prune_list;
546 static kmutex_t arc_prune_mtx;
547 static arc_buf_t *arc_eviction_list;
548 static kmutex_t arc_eviction_mtx;
549 static arc_buf_hdr_t arc_eviction_hdr;
550 static void arc_get_data_buf(arc_buf_t *buf);
551 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
552 static int arc_evict_needed(arc_buf_contents_t type);
553 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes,
554 arc_buf_contents_t type);
556 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
558 #define GHOST_STATE(state) \
559 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
560 (state) == arc_l2c_only)
563 * Private ARC flags. These flags are private ARC only flags that will show up
564 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
565 * be passed in as arc_flags in things like arc_read. However, these flags
566 * should never be passed and should only be set by ARC code. When adding new
567 * public flags, make sure not to smash the private ones.
570 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
571 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
572 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
573 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
574 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
575 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
576 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
577 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
578 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
579 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
581 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
582 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
583 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
584 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
585 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
586 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
587 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
588 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
589 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
590 (hdr)->b_l2hdr != NULL)
591 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
592 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
593 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
599 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
600 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
603 * Hash table routines
606 #define HT_LOCK_ALIGN 64
607 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
612 unsigned char pad[HT_LOCK_PAD];
616 #define BUF_LOCKS 256
617 typedef struct buf_hash_table {
619 arc_buf_hdr_t **ht_table;
620 struct ht_lock ht_locks[BUF_LOCKS];
623 static buf_hash_table_t buf_hash_table;
625 #define BUF_HASH_INDEX(spa, dva, birth) \
626 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
627 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
628 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
629 #define HDR_LOCK(hdr) \
630 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
632 uint64_t zfs_crc64_table[256];
638 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
639 #define L2ARC_HEADROOM 2 /* num of writes */
641 * If we discover during ARC scan any buffers to be compressed, we boost
642 * our headroom for the next scanning cycle by this percentage multiple.
644 #define L2ARC_HEADROOM_BOOST 200
645 #define L2ARC_FEED_SECS 1 /* caching interval secs */
646 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
648 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
649 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
652 * L2ARC Performance Tunables
654 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
655 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
656 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
657 unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
658 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
659 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
660 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
661 int l2arc_nocompress = B_FALSE; /* don't compress bufs */
662 int l2arc_feed_again = B_TRUE; /* turbo warmup */
663 int l2arc_norw = B_FALSE; /* no reads during writes */
668 typedef struct l2arc_dev {
669 vdev_t *l2ad_vdev; /* vdev */
670 spa_t *l2ad_spa; /* spa */
671 uint64_t l2ad_hand; /* next write location */
672 uint64_t l2ad_start; /* first addr on device */
673 uint64_t l2ad_end; /* last addr on device */
674 uint64_t l2ad_evict; /* last addr eviction reached */
675 boolean_t l2ad_first; /* first sweep through */
676 boolean_t l2ad_writing; /* currently writing */
677 list_t *l2ad_buflist; /* buffer list */
678 list_node_t l2ad_node; /* device list node */
681 static list_t L2ARC_dev_list; /* device list */
682 static list_t *l2arc_dev_list; /* device list pointer */
683 static kmutex_t l2arc_dev_mtx; /* device list mutex */
684 static l2arc_dev_t *l2arc_dev_last; /* last device used */
685 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
686 static list_t L2ARC_free_on_write; /* free after write buf list */
687 static list_t *l2arc_free_on_write; /* free after write list ptr */
688 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
689 static uint64_t l2arc_ndev; /* number of devices */
691 typedef struct l2arc_read_callback {
692 arc_buf_t *l2rcb_buf; /* read buffer */
693 spa_t *l2rcb_spa; /* spa */
694 blkptr_t l2rcb_bp; /* original blkptr */
695 zbookmark_t l2rcb_zb; /* original bookmark */
696 int l2rcb_flags; /* original flags */
697 enum zio_compress l2rcb_compress; /* applied compress */
698 } l2arc_read_callback_t;
700 typedef struct l2arc_write_callback {
701 l2arc_dev_t *l2wcb_dev; /* device info */
702 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
703 } l2arc_write_callback_t;
705 struct l2arc_buf_hdr {
706 /* protected by arc_buf_hdr mutex */
707 l2arc_dev_t *b_dev; /* L2ARC device */
708 uint64_t b_daddr; /* disk address, offset byte */
709 /* compression applied to buffer data */
710 enum zio_compress b_compress;
711 /* real alloc'd buffer size depending on b_compress applied */
713 /* temporary buffer holder for in-flight compressed data */
717 typedef struct l2arc_data_free {
718 /* protected by l2arc_free_on_write_mtx */
721 void (*l2df_func)(void *, size_t);
722 list_node_t l2df_list_node;
725 static kmutex_t l2arc_feed_thr_lock;
726 static kcondvar_t l2arc_feed_thr_cv;
727 static uint8_t l2arc_thread_exit;
729 static void l2arc_read_done(zio_t *zio);
730 static void l2arc_hdr_stat_add(void);
731 static void l2arc_hdr_stat_remove(void);
733 static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr);
734 static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr,
735 enum zio_compress c);
736 static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab);
739 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
741 uint8_t *vdva = (uint8_t *)dva;
742 uint64_t crc = -1ULL;
745 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
747 for (i = 0; i < sizeof (dva_t); i++)
748 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
750 crc ^= (spa>>8) ^ birth;
755 #define BUF_EMPTY(buf) \
756 ((buf)->b_dva.dva_word[0] == 0 && \
757 (buf)->b_dva.dva_word[1] == 0 && \
760 #define BUF_EQUAL(spa, dva, birth, buf) \
761 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
762 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
763 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
766 buf_discard_identity(arc_buf_hdr_t *hdr)
768 hdr->b_dva.dva_word[0] = 0;
769 hdr->b_dva.dva_word[1] = 0;
774 static arc_buf_hdr_t *
775 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
777 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
778 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
781 mutex_enter(hash_lock);
782 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
783 buf = buf->b_hash_next) {
784 if (BUF_EQUAL(spa, dva, birth, buf)) {
789 mutex_exit(hash_lock);
795 * Insert an entry into the hash table. If there is already an element
796 * equal to elem in the hash table, then the already existing element
797 * will be returned and the new element will not be inserted.
798 * Otherwise returns NULL.
800 static arc_buf_hdr_t *
801 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
803 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
804 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
808 ASSERT(!HDR_IN_HASH_TABLE(buf));
810 mutex_enter(hash_lock);
811 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
812 fbuf = fbuf->b_hash_next, i++) {
813 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
817 buf->b_hash_next = buf_hash_table.ht_table[idx];
818 buf_hash_table.ht_table[idx] = buf;
819 buf->b_flags |= ARC_IN_HASH_TABLE;
821 /* collect some hash table performance data */
823 ARCSTAT_BUMP(arcstat_hash_collisions);
825 ARCSTAT_BUMP(arcstat_hash_chains);
827 ARCSTAT_MAX(arcstat_hash_chain_max, i);
830 ARCSTAT_BUMP(arcstat_hash_elements);
831 ARCSTAT_MAXSTAT(arcstat_hash_elements);
837 buf_hash_remove(arc_buf_hdr_t *buf)
839 arc_buf_hdr_t *fbuf, **bufp;
840 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
842 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
843 ASSERT(HDR_IN_HASH_TABLE(buf));
845 bufp = &buf_hash_table.ht_table[idx];
846 while ((fbuf = *bufp) != buf) {
847 ASSERT(fbuf != NULL);
848 bufp = &fbuf->b_hash_next;
850 *bufp = buf->b_hash_next;
851 buf->b_hash_next = NULL;
852 buf->b_flags &= ~ARC_IN_HASH_TABLE;
854 /* collect some hash table performance data */
855 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
857 if (buf_hash_table.ht_table[idx] &&
858 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
859 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
863 * Global data structures and functions for the buf kmem cache.
865 static kmem_cache_t *hdr_cache;
866 static kmem_cache_t *buf_cache;
873 #if defined(_KERNEL) && defined(HAVE_SPL)
874 /* Large allocations which do not require contiguous pages
875 * should be using vmem_free() in the linux kernel */
876 vmem_free(buf_hash_table.ht_table,
877 (buf_hash_table.ht_mask + 1) * sizeof (void *));
879 kmem_free(buf_hash_table.ht_table,
880 (buf_hash_table.ht_mask + 1) * sizeof (void *));
882 for (i = 0; i < BUF_LOCKS; i++)
883 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
884 kmem_cache_destroy(hdr_cache);
885 kmem_cache_destroy(buf_cache);
889 * Constructor callback - called when the cache is empty
890 * and a new buf is requested.
894 hdr_cons(void *vbuf, void *unused, int kmflag)
896 arc_buf_hdr_t *buf = vbuf;
898 bzero(buf, sizeof (arc_buf_hdr_t));
899 refcount_create(&buf->b_refcnt);
900 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
901 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
902 list_link_init(&buf->b_arc_node);
903 list_link_init(&buf->b_l2node);
904 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
911 buf_cons(void *vbuf, void *unused, int kmflag)
913 arc_buf_t *buf = vbuf;
915 bzero(buf, sizeof (arc_buf_t));
916 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
917 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
923 * Destructor callback - called when a cached buf is
924 * no longer required.
928 hdr_dest(void *vbuf, void *unused)
930 arc_buf_hdr_t *buf = vbuf;
932 ASSERT(BUF_EMPTY(buf));
933 refcount_destroy(&buf->b_refcnt);
934 cv_destroy(&buf->b_cv);
935 mutex_destroy(&buf->b_freeze_lock);
936 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
941 buf_dest(void *vbuf, void *unused)
943 arc_buf_t *buf = vbuf;
945 mutex_destroy(&buf->b_evict_lock);
946 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
953 uint64_t hsize = 1ULL << 12;
957 * The hash table is big enough to fill all of physical memory
958 * with an average 64K block size. The table will take up
959 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
961 while (hsize * 65536 < physmem * PAGESIZE)
964 buf_hash_table.ht_mask = hsize - 1;
965 #if defined(_KERNEL) && defined(HAVE_SPL)
966 /* Large allocations which do not require contiguous pages
967 * should be using vmem_alloc() in the linux kernel */
968 buf_hash_table.ht_table =
969 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
971 buf_hash_table.ht_table =
972 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
974 if (buf_hash_table.ht_table == NULL) {
975 ASSERT(hsize > (1ULL << 8));
980 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
981 0, hdr_cons, hdr_dest, NULL, NULL, NULL, 0);
982 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
983 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
985 for (i = 0; i < 256; i++)
986 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
987 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
989 for (i = 0; i < BUF_LOCKS; i++) {
990 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
991 NULL, MUTEX_DEFAULT, NULL);
995 #define ARC_MINTIME (hz>>4) /* 62 ms */
998 arc_cksum_verify(arc_buf_t *buf)
1002 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1005 mutex_enter(&buf->b_hdr->b_freeze_lock);
1006 if (buf->b_hdr->b_freeze_cksum == NULL ||
1007 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
1008 mutex_exit(&buf->b_hdr->b_freeze_lock);
1011 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1012 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1013 panic("buffer modified while frozen!");
1014 mutex_exit(&buf->b_hdr->b_freeze_lock);
1018 arc_cksum_equal(arc_buf_t *buf)
1023 mutex_enter(&buf->b_hdr->b_freeze_lock);
1024 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1025 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1026 mutex_exit(&buf->b_hdr->b_freeze_lock);
1032 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1034 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1037 mutex_enter(&buf->b_hdr->b_freeze_lock);
1038 if (buf->b_hdr->b_freeze_cksum != NULL) {
1039 mutex_exit(&buf->b_hdr->b_freeze_lock);
1042 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1044 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1045 buf->b_hdr->b_freeze_cksum);
1046 mutex_exit(&buf->b_hdr->b_freeze_lock);
1050 arc_buf_thaw(arc_buf_t *buf)
1052 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1053 if (buf->b_hdr->b_state != arc_anon)
1054 panic("modifying non-anon buffer!");
1055 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1056 panic("modifying buffer while i/o in progress!");
1057 arc_cksum_verify(buf);
1060 mutex_enter(&buf->b_hdr->b_freeze_lock);
1061 if (buf->b_hdr->b_freeze_cksum != NULL) {
1062 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1063 buf->b_hdr->b_freeze_cksum = NULL;
1066 mutex_exit(&buf->b_hdr->b_freeze_lock);
1070 arc_buf_freeze(arc_buf_t *buf)
1072 kmutex_t *hash_lock;
1074 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1077 hash_lock = HDR_LOCK(buf->b_hdr);
1078 mutex_enter(hash_lock);
1080 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1081 buf->b_hdr->b_state == arc_anon);
1082 arc_cksum_compute(buf, B_FALSE);
1083 mutex_exit(hash_lock);
1087 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1089 ASSERT(MUTEX_HELD(hash_lock));
1091 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1092 (ab->b_state != arc_anon)) {
1093 uint64_t delta = ab->b_size * ab->b_datacnt;
1094 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1095 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1097 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1098 mutex_enter(&ab->b_state->arcs_mtx);
1099 ASSERT(list_link_active(&ab->b_arc_node));
1100 list_remove(list, ab);
1101 if (GHOST_STATE(ab->b_state)) {
1102 ASSERT0(ab->b_datacnt);
1103 ASSERT3P(ab->b_buf, ==, NULL);
1107 ASSERT3U(*size, >=, delta);
1108 atomic_add_64(size, -delta);
1109 mutex_exit(&ab->b_state->arcs_mtx);
1110 /* remove the prefetch flag if we get a reference */
1111 if (ab->b_flags & ARC_PREFETCH)
1112 ab->b_flags &= ~ARC_PREFETCH;
1117 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1120 arc_state_t *state = ab->b_state;
1122 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1123 ASSERT(!GHOST_STATE(state));
1125 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1126 (state != arc_anon)) {
1127 uint64_t *size = &state->arcs_lsize[ab->b_type];
1129 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1130 mutex_enter(&state->arcs_mtx);
1131 ASSERT(!list_link_active(&ab->b_arc_node));
1132 list_insert_head(&state->arcs_list[ab->b_type], ab);
1133 ASSERT(ab->b_datacnt > 0);
1134 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1135 mutex_exit(&state->arcs_mtx);
1141 * Move the supplied buffer to the indicated state. The mutex
1142 * for the buffer must be held by the caller.
1145 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1147 arc_state_t *old_state = ab->b_state;
1148 int64_t refcnt = refcount_count(&ab->b_refcnt);
1149 uint64_t from_delta, to_delta;
1151 ASSERT(MUTEX_HELD(hash_lock));
1152 ASSERT(new_state != old_state);
1153 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1154 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1155 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1157 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1160 * If this buffer is evictable, transfer it from the
1161 * old state list to the new state list.
1164 if (old_state != arc_anon) {
1165 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1166 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1169 mutex_enter(&old_state->arcs_mtx);
1171 ASSERT(list_link_active(&ab->b_arc_node));
1172 list_remove(&old_state->arcs_list[ab->b_type], ab);
1175 * If prefetching out of the ghost cache,
1176 * we will have a non-zero datacnt.
1178 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1179 /* ghost elements have a ghost size */
1180 ASSERT(ab->b_buf == NULL);
1181 from_delta = ab->b_size;
1183 ASSERT3U(*size, >=, from_delta);
1184 atomic_add_64(size, -from_delta);
1187 mutex_exit(&old_state->arcs_mtx);
1189 if (new_state != arc_anon) {
1190 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1191 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1194 mutex_enter(&new_state->arcs_mtx);
1196 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1198 /* ghost elements have a ghost size */
1199 if (GHOST_STATE(new_state)) {
1200 ASSERT(ab->b_datacnt == 0);
1201 ASSERT(ab->b_buf == NULL);
1202 to_delta = ab->b_size;
1204 atomic_add_64(size, to_delta);
1207 mutex_exit(&new_state->arcs_mtx);
1211 ASSERT(!BUF_EMPTY(ab));
1212 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1213 buf_hash_remove(ab);
1215 /* adjust state sizes */
1217 atomic_add_64(&new_state->arcs_size, to_delta);
1219 ASSERT3U(old_state->arcs_size, >=, from_delta);
1220 atomic_add_64(&old_state->arcs_size, -from_delta);
1222 ab->b_state = new_state;
1224 /* adjust l2arc hdr stats */
1225 if (new_state == arc_l2c_only)
1226 l2arc_hdr_stat_add();
1227 else if (old_state == arc_l2c_only)
1228 l2arc_hdr_stat_remove();
1232 arc_space_consume(uint64_t space, arc_space_type_t type)
1234 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1239 case ARC_SPACE_DATA:
1240 ARCSTAT_INCR(arcstat_data_size, space);
1242 case ARC_SPACE_OTHER:
1243 ARCSTAT_INCR(arcstat_other_size, space);
1245 case ARC_SPACE_HDRS:
1246 ARCSTAT_INCR(arcstat_hdr_size, space);
1248 case ARC_SPACE_L2HDRS:
1249 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1253 atomic_add_64(&arc_meta_used, space);
1254 atomic_add_64(&arc_size, space);
1258 arc_space_return(uint64_t space, arc_space_type_t type)
1260 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1265 case ARC_SPACE_DATA:
1266 ARCSTAT_INCR(arcstat_data_size, -space);
1268 case ARC_SPACE_OTHER:
1269 ARCSTAT_INCR(arcstat_other_size, -space);
1271 case ARC_SPACE_HDRS:
1272 ARCSTAT_INCR(arcstat_hdr_size, -space);
1274 case ARC_SPACE_L2HDRS:
1275 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1279 ASSERT(arc_meta_used >= space);
1280 if (arc_meta_max < arc_meta_used)
1281 arc_meta_max = arc_meta_used;
1282 atomic_add_64(&arc_meta_used, -space);
1283 ASSERT(arc_size >= space);
1284 atomic_add_64(&arc_size, -space);
1288 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1293 ASSERT3U(size, >, 0);
1294 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1295 ASSERT(BUF_EMPTY(hdr));
1298 hdr->b_spa = spa_load_guid(spa);
1299 hdr->b_state = arc_anon;
1300 hdr->b_arc_access = 0;
1301 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1304 buf->b_efunc = NULL;
1305 buf->b_private = NULL;
1308 arc_get_data_buf(buf);
1311 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1312 (void) refcount_add(&hdr->b_refcnt, tag);
1317 static char *arc_onloan_tag = "onloan";
1320 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1321 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1322 * buffers must be returned to the arc before they can be used by the DMU or
1326 arc_loan_buf(spa_t *spa, int size)
1330 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1332 atomic_add_64(&arc_loaned_bytes, size);
1337 * Return a loaned arc buffer to the arc.
1340 arc_return_buf(arc_buf_t *buf, void *tag)
1342 arc_buf_hdr_t *hdr = buf->b_hdr;
1344 ASSERT(buf->b_data != NULL);
1345 (void) refcount_add(&hdr->b_refcnt, tag);
1346 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1348 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1351 /* Detach an arc_buf from a dbuf (tag) */
1353 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1357 ASSERT(buf->b_data != NULL);
1359 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1360 (void) refcount_remove(&hdr->b_refcnt, tag);
1361 buf->b_efunc = NULL;
1362 buf->b_private = NULL;
1364 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1368 arc_buf_clone(arc_buf_t *from)
1371 arc_buf_hdr_t *hdr = from->b_hdr;
1372 uint64_t size = hdr->b_size;
1374 ASSERT(hdr->b_state != arc_anon);
1376 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1379 buf->b_efunc = NULL;
1380 buf->b_private = NULL;
1381 buf->b_next = hdr->b_buf;
1383 arc_get_data_buf(buf);
1384 bcopy(from->b_data, buf->b_data, size);
1387 * This buffer already exists in the arc so create a duplicate
1388 * copy for the caller. If the buffer is associated with user data
1389 * then track the size and number of duplicates. These stats will be
1390 * updated as duplicate buffers are created and destroyed.
1392 if (hdr->b_type == ARC_BUFC_DATA) {
1393 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1394 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1396 hdr->b_datacnt += 1;
1401 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1404 kmutex_t *hash_lock;
1407 * Check to see if this buffer is evicted. Callers
1408 * must verify b_data != NULL to know if the add_ref
1411 mutex_enter(&buf->b_evict_lock);
1412 if (buf->b_data == NULL) {
1413 mutex_exit(&buf->b_evict_lock);
1416 hash_lock = HDR_LOCK(buf->b_hdr);
1417 mutex_enter(hash_lock);
1419 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1420 mutex_exit(&buf->b_evict_lock);
1422 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1423 add_reference(hdr, hash_lock, tag);
1424 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1425 arc_access(hdr, hash_lock);
1426 mutex_exit(hash_lock);
1427 ARCSTAT_BUMP(arcstat_hits);
1428 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1429 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1430 data, metadata, hits);
1434 * Free the arc data buffer. If it is an l2arc write in progress,
1435 * the buffer is placed on l2arc_free_on_write to be freed later.
1438 arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t),
1439 void *data, size_t size)
1441 if (HDR_L2_WRITING(hdr)) {
1442 l2arc_data_free_t *df;
1443 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_PUSHPAGE);
1444 df->l2df_data = data;
1445 df->l2df_size = size;
1446 df->l2df_func = free_func;
1447 mutex_enter(&l2arc_free_on_write_mtx);
1448 list_insert_head(l2arc_free_on_write, df);
1449 mutex_exit(&l2arc_free_on_write_mtx);
1450 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1452 free_func(data, size);
1457 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1461 /* free up data associated with the buf */
1463 arc_state_t *state = buf->b_hdr->b_state;
1464 uint64_t size = buf->b_hdr->b_size;
1465 arc_buf_contents_t type = buf->b_hdr->b_type;
1467 arc_cksum_verify(buf);
1470 if (type == ARC_BUFC_METADATA) {
1471 arc_buf_data_free(buf->b_hdr, zio_buf_free,
1473 arc_space_return(size, ARC_SPACE_DATA);
1475 ASSERT(type == ARC_BUFC_DATA);
1476 arc_buf_data_free(buf->b_hdr,
1477 zio_data_buf_free, buf->b_data, size);
1478 ARCSTAT_INCR(arcstat_data_size, -size);
1479 atomic_add_64(&arc_size, -size);
1482 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1483 uint64_t *cnt = &state->arcs_lsize[type];
1485 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1486 ASSERT(state != arc_anon);
1488 ASSERT3U(*cnt, >=, size);
1489 atomic_add_64(cnt, -size);
1491 ASSERT3U(state->arcs_size, >=, size);
1492 atomic_add_64(&state->arcs_size, -size);
1496 * If we're destroying a duplicate buffer make sure
1497 * that the appropriate statistics are updated.
1499 if (buf->b_hdr->b_datacnt > 1 &&
1500 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1501 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1502 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1504 ASSERT(buf->b_hdr->b_datacnt > 0);
1505 buf->b_hdr->b_datacnt -= 1;
1508 /* only remove the buf if requested */
1512 /* remove the buf from the hdr list */
1513 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1515 *bufp = buf->b_next;
1518 ASSERT(buf->b_efunc == NULL);
1520 /* clean up the buf */
1522 kmem_cache_free(buf_cache, buf);
1526 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1528 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1530 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1531 ASSERT3P(hdr->b_state, ==, arc_anon);
1532 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1534 if (l2hdr != NULL) {
1535 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1537 * To prevent arc_free() and l2arc_evict() from
1538 * attempting to free the same buffer at the same time,
1539 * a FREE_IN_PROGRESS flag is given to arc_free() to
1540 * give it priority. l2arc_evict() can't destroy this
1541 * header while we are waiting on l2arc_buflist_mtx.
1543 * The hdr may be removed from l2ad_buflist before we
1544 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1546 if (!buflist_held) {
1547 mutex_enter(&l2arc_buflist_mtx);
1548 l2hdr = hdr->b_l2hdr;
1551 if (l2hdr != NULL) {
1552 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1553 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1554 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
1555 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1556 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
1557 if (hdr->b_state == arc_l2c_only)
1558 l2arc_hdr_stat_remove();
1559 hdr->b_l2hdr = NULL;
1563 mutex_exit(&l2arc_buflist_mtx);
1566 if (!BUF_EMPTY(hdr)) {
1567 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1568 buf_discard_identity(hdr);
1570 while (hdr->b_buf) {
1571 arc_buf_t *buf = hdr->b_buf;
1574 mutex_enter(&arc_eviction_mtx);
1575 mutex_enter(&buf->b_evict_lock);
1576 ASSERT(buf->b_hdr != NULL);
1577 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1578 hdr->b_buf = buf->b_next;
1579 buf->b_hdr = &arc_eviction_hdr;
1580 buf->b_next = arc_eviction_list;
1581 arc_eviction_list = buf;
1582 mutex_exit(&buf->b_evict_lock);
1583 mutex_exit(&arc_eviction_mtx);
1585 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1588 if (hdr->b_freeze_cksum != NULL) {
1589 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1590 hdr->b_freeze_cksum = NULL;
1593 ASSERT(!list_link_active(&hdr->b_arc_node));
1594 ASSERT3P(hdr->b_hash_next, ==, NULL);
1595 ASSERT3P(hdr->b_acb, ==, NULL);
1596 kmem_cache_free(hdr_cache, hdr);
1600 arc_buf_free(arc_buf_t *buf, void *tag)
1602 arc_buf_hdr_t *hdr = buf->b_hdr;
1603 int hashed = hdr->b_state != arc_anon;
1605 ASSERT(buf->b_efunc == NULL);
1606 ASSERT(buf->b_data != NULL);
1609 kmutex_t *hash_lock = HDR_LOCK(hdr);
1611 mutex_enter(hash_lock);
1613 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1615 (void) remove_reference(hdr, hash_lock, tag);
1616 if (hdr->b_datacnt > 1) {
1617 arc_buf_destroy(buf, FALSE, TRUE);
1619 ASSERT(buf == hdr->b_buf);
1620 ASSERT(buf->b_efunc == NULL);
1621 hdr->b_flags |= ARC_BUF_AVAILABLE;
1623 mutex_exit(hash_lock);
1624 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1627 * We are in the middle of an async write. Don't destroy
1628 * this buffer unless the write completes before we finish
1629 * decrementing the reference count.
1631 mutex_enter(&arc_eviction_mtx);
1632 (void) remove_reference(hdr, NULL, tag);
1633 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1634 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1635 mutex_exit(&arc_eviction_mtx);
1637 arc_hdr_destroy(hdr);
1639 if (remove_reference(hdr, NULL, tag) > 0)
1640 arc_buf_destroy(buf, FALSE, TRUE);
1642 arc_hdr_destroy(hdr);
1647 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1649 arc_buf_hdr_t *hdr = buf->b_hdr;
1650 kmutex_t *hash_lock = NULL;
1651 int no_callback = (buf->b_efunc == NULL);
1653 if (hdr->b_state == arc_anon) {
1654 ASSERT(hdr->b_datacnt == 1);
1655 arc_buf_free(buf, tag);
1656 return (no_callback);
1659 hash_lock = HDR_LOCK(hdr);
1660 mutex_enter(hash_lock);
1662 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1663 ASSERT(hdr->b_state != arc_anon);
1664 ASSERT(buf->b_data != NULL);
1666 (void) remove_reference(hdr, hash_lock, tag);
1667 if (hdr->b_datacnt > 1) {
1669 arc_buf_destroy(buf, FALSE, TRUE);
1670 } else if (no_callback) {
1671 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1672 ASSERT(buf->b_efunc == NULL);
1673 hdr->b_flags |= ARC_BUF_AVAILABLE;
1675 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1676 refcount_is_zero(&hdr->b_refcnt));
1677 mutex_exit(hash_lock);
1678 return (no_callback);
1682 arc_buf_size(arc_buf_t *buf)
1684 return (buf->b_hdr->b_size);
1688 * Called from the DMU to determine if the current buffer should be
1689 * evicted. In order to ensure proper locking, the eviction must be initiated
1690 * from the DMU. Return true if the buffer is associated with user data and
1691 * duplicate buffers still exist.
1694 arc_buf_eviction_needed(arc_buf_t *buf)
1697 boolean_t evict_needed = B_FALSE;
1699 if (zfs_disable_dup_eviction)
1702 mutex_enter(&buf->b_evict_lock);
1706 * We are in arc_do_user_evicts(); let that function
1707 * perform the eviction.
1709 ASSERT(buf->b_data == NULL);
1710 mutex_exit(&buf->b_evict_lock);
1712 } else if (buf->b_data == NULL) {
1714 * We have already been added to the arc eviction list;
1715 * recommend eviction.
1717 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1718 mutex_exit(&buf->b_evict_lock);
1722 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1723 evict_needed = B_TRUE;
1725 mutex_exit(&buf->b_evict_lock);
1726 return (evict_needed);
1730 * Evict buffers from list until we've removed the specified number of
1731 * bytes. Move the removed buffers to the appropriate evict state.
1732 * If the recycle flag is set, then attempt to "recycle" a buffer:
1733 * - look for a buffer to evict that is `bytes' long.
1734 * - return the data block from this buffer rather than freeing it.
1735 * This flag is used by callers that are trying to make space for a
1736 * new buffer in a full arc cache.
1738 * This function makes a "best effort". It skips over any buffers
1739 * it can't get a hash_lock on, and so may not catch all candidates.
1740 * It may also return without evicting as much space as requested.
1743 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1744 arc_buf_contents_t type)
1746 arc_state_t *evicted_state;
1747 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1748 arc_buf_hdr_t *ab, *ab_prev = NULL;
1749 list_t *list = &state->arcs_list[type];
1750 kmutex_t *hash_lock;
1751 boolean_t have_lock;
1752 void *stolen = NULL;
1754 ASSERT(state == arc_mru || state == arc_mfu);
1756 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1758 mutex_enter(&state->arcs_mtx);
1759 mutex_enter(&evicted_state->arcs_mtx);
1761 for (ab = list_tail(list); ab; ab = ab_prev) {
1762 ab_prev = list_prev(list, ab);
1763 /* prefetch buffers have a minimum lifespan */
1764 if (HDR_IO_IN_PROGRESS(ab) ||
1765 (spa && ab->b_spa != spa) ||
1766 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1767 ddi_get_lbolt() - ab->b_arc_access <
1768 zfs_arc_min_prefetch_lifespan)) {
1772 /* "lookahead" for better eviction candidate */
1773 if (recycle && ab->b_size != bytes &&
1774 ab_prev && ab_prev->b_size == bytes)
1776 hash_lock = HDR_LOCK(ab);
1777 have_lock = MUTEX_HELD(hash_lock);
1778 if (have_lock || mutex_tryenter(hash_lock)) {
1779 ASSERT0(refcount_count(&ab->b_refcnt));
1780 ASSERT(ab->b_datacnt > 0);
1782 arc_buf_t *buf = ab->b_buf;
1783 if (!mutex_tryenter(&buf->b_evict_lock)) {
1788 bytes_evicted += ab->b_size;
1789 if (recycle && ab->b_type == type &&
1790 ab->b_size == bytes &&
1791 !HDR_L2_WRITING(ab)) {
1792 stolen = buf->b_data;
1797 mutex_enter(&arc_eviction_mtx);
1798 arc_buf_destroy(buf,
1799 buf->b_data == stolen, FALSE);
1800 ab->b_buf = buf->b_next;
1801 buf->b_hdr = &arc_eviction_hdr;
1802 buf->b_next = arc_eviction_list;
1803 arc_eviction_list = buf;
1804 mutex_exit(&arc_eviction_mtx);
1805 mutex_exit(&buf->b_evict_lock);
1807 mutex_exit(&buf->b_evict_lock);
1808 arc_buf_destroy(buf,
1809 buf->b_data == stolen, TRUE);
1814 ARCSTAT_INCR(arcstat_evict_l2_cached,
1817 if (l2arc_write_eligible(ab->b_spa, ab)) {
1818 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1822 arcstat_evict_l2_ineligible,
1827 if (ab->b_datacnt == 0) {
1828 arc_change_state(evicted_state, ab, hash_lock);
1829 ASSERT(HDR_IN_HASH_TABLE(ab));
1830 ab->b_flags |= ARC_IN_HASH_TABLE;
1831 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1832 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1835 mutex_exit(hash_lock);
1836 if (bytes >= 0 && bytes_evicted >= bytes)
1843 mutex_exit(&evicted_state->arcs_mtx);
1844 mutex_exit(&state->arcs_mtx);
1846 if (bytes_evicted < bytes)
1847 dprintf("only evicted %lld bytes from %x\n",
1848 (longlong_t)bytes_evicted, state);
1851 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1854 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1857 * We have just evicted some date into the ghost state, make
1858 * sure we also adjust the ghost state size if necessary.
1861 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1862 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1863 arc_mru_ghost->arcs_size - arc_c;
1865 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1867 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1868 arc_evict_ghost(arc_mru_ghost, 0, todelete,
1870 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1871 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1872 arc_mru_ghost->arcs_size +
1873 arc_mfu_ghost->arcs_size - arc_c);
1874 arc_evict_ghost(arc_mfu_ghost, 0, todelete,
1883 * Remove buffers from list until we've removed the specified number of
1884 * bytes. Destroy the buffers that are removed.
1887 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes,
1888 arc_buf_contents_t type)
1890 arc_buf_hdr_t *ab, *ab_prev;
1891 arc_buf_hdr_t marker;
1892 list_t *list = &state->arcs_list[type];
1893 kmutex_t *hash_lock;
1894 uint64_t bytes_deleted = 0;
1895 uint64_t bufs_skipped = 0;
1897 ASSERT(GHOST_STATE(state));
1898 bzero(&marker, sizeof(marker));
1900 mutex_enter(&state->arcs_mtx);
1901 for (ab = list_tail(list); ab; ab = ab_prev) {
1902 ab_prev = list_prev(list, ab);
1903 if (spa && ab->b_spa != spa)
1906 /* ignore markers */
1910 hash_lock = HDR_LOCK(ab);
1911 /* caller may be trying to modify this buffer, skip it */
1912 if (MUTEX_HELD(hash_lock))
1914 if (mutex_tryenter(hash_lock)) {
1915 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1916 ASSERT(ab->b_buf == NULL);
1917 ARCSTAT_BUMP(arcstat_deleted);
1918 bytes_deleted += ab->b_size;
1920 if (ab->b_l2hdr != NULL) {
1922 * This buffer is cached on the 2nd Level ARC;
1923 * don't destroy the header.
1925 arc_change_state(arc_l2c_only, ab, hash_lock);
1926 mutex_exit(hash_lock);
1928 arc_change_state(arc_anon, ab, hash_lock);
1929 mutex_exit(hash_lock);
1930 arc_hdr_destroy(ab);
1933 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1934 if (bytes >= 0 && bytes_deleted >= bytes)
1936 } else if (bytes < 0) {
1938 * Insert a list marker and then wait for the
1939 * hash lock to become available. Once its
1940 * available, restart from where we left off.
1942 list_insert_after(list, ab, &marker);
1943 mutex_exit(&state->arcs_mtx);
1944 mutex_enter(hash_lock);
1945 mutex_exit(hash_lock);
1946 mutex_enter(&state->arcs_mtx);
1947 ab_prev = list_prev(list, &marker);
1948 list_remove(list, &marker);
1952 mutex_exit(&state->arcs_mtx);
1954 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1955 (bytes < 0 || bytes_deleted < bytes)) {
1956 list = &state->arcs_list[ARC_BUFC_METADATA];
1961 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1965 if (bytes_deleted < bytes)
1966 dprintf("only deleted %lld bytes from %p\n",
1967 (longlong_t)bytes_deleted, state);
1973 int64_t adjustment, delta;
1979 adjustment = MIN((int64_t)(arc_size - arc_c),
1980 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
1983 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1984 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
1985 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
1986 adjustment -= delta;
1989 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1990 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
1991 (void) arc_evict(arc_mru, 0, delta, FALSE,
1999 adjustment = arc_size - arc_c;
2001 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2002 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2003 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2004 adjustment -= delta;
2007 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2008 int64_t delta = MIN(adjustment,
2009 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2010 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2015 * Adjust ghost lists
2018 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2020 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2021 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2022 arc_evict_ghost(arc_mru_ghost, 0, delta, ARC_BUFC_DATA);
2026 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2028 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2029 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2030 arc_evict_ghost(arc_mfu_ghost, 0, delta, ARC_BUFC_DATA);
2035 * Request that arc user drop references so that N bytes can be released
2036 * from the cache. This provides a mechanism to ensure the arc can honor
2037 * the arc_meta_limit and reclaim buffers which are pinned in the cache
2038 * by higher layers. (i.e. the zpl)
2041 arc_do_user_prune(int64_t adjustment)
2043 arc_prune_func_t *func;
2045 arc_prune_t *cp, *np;
2047 mutex_enter(&arc_prune_mtx);
2049 cp = list_head(&arc_prune_list);
2050 while (cp != NULL) {
2052 private = cp->p_private;
2053 np = list_next(&arc_prune_list, cp);
2054 refcount_add(&cp->p_refcnt, func);
2055 mutex_exit(&arc_prune_mtx);
2058 func(adjustment, private);
2060 mutex_enter(&arc_prune_mtx);
2062 /* User removed prune callback concurrently with execution */
2063 if (refcount_remove(&cp->p_refcnt, func) == 0) {
2064 ASSERT(!list_link_active(&cp->p_node));
2065 refcount_destroy(&cp->p_refcnt);
2066 kmem_free(cp, sizeof (*cp));
2072 ARCSTAT_BUMP(arcstat_prune);
2073 mutex_exit(&arc_prune_mtx);
2077 arc_do_user_evicts(void)
2079 mutex_enter(&arc_eviction_mtx);
2080 while (arc_eviction_list != NULL) {
2081 arc_buf_t *buf = arc_eviction_list;
2082 arc_eviction_list = buf->b_next;
2083 mutex_enter(&buf->b_evict_lock);
2085 mutex_exit(&buf->b_evict_lock);
2086 mutex_exit(&arc_eviction_mtx);
2088 if (buf->b_efunc != NULL)
2089 VERIFY(buf->b_efunc(buf) == 0);
2091 buf->b_efunc = NULL;
2092 buf->b_private = NULL;
2093 kmem_cache_free(buf_cache, buf);
2094 mutex_enter(&arc_eviction_mtx);
2096 mutex_exit(&arc_eviction_mtx);
2100 * Evict only meta data objects from the cache leaving the data objects.
2101 * This is only used to enforce the tunable arc_meta_limit, if we are
2102 * unable to evict enough buffers notify the user via the prune callback.
2105 arc_adjust_meta(int64_t adjustment, boolean_t may_prune)
2109 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2110 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2111 arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_METADATA);
2112 adjustment -= delta;
2115 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2116 delta = MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2117 arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_METADATA);
2118 adjustment -= delta;
2121 if (may_prune && (adjustment > 0) && (arc_meta_used > arc_meta_limit))
2122 arc_do_user_prune(zfs_arc_meta_prune);
2126 * Flush all *evictable* data from the cache for the given spa.
2127 * NOTE: this will not touch "active" (i.e. referenced) data.
2130 arc_flush(spa_t *spa)
2135 guid = spa_load_guid(spa);
2137 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2138 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2142 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2143 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2147 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2148 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2152 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2153 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2158 arc_evict_ghost(arc_mru_ghost, guid, -1, ARC_BUFC_DATA);
2159 arc_evict_ghost(arc_mfu_ghost, guid, -1, ARC_BUFC_DATA);
2161 mutex_enter(&arc_reclaim_thr_lock);
2162 arc_do_user_evicts();
2163 mutex_exit(&arc_reclaim_thr_lock);
2164 ASSERT(spa || arc_eviction_list == NULL);
2168 arc_shrink(uint64_t bytes)
2170 if (arc_c > arc_c_min) {
2173 to_free = bytes ? bytes : arc_c >> zfs_arc_shrink_shift;
2175 if (arc_c > arc_c_min + to_free)
2176 atomic_add_64(&arc_c, -to_free);
2180 atomic_add_64(&arc_p, -(arc_p >> zfs_arc_shrink_shift));
2181 if (arc_c > arc_size)
2182 arc_c = MAX(arc_size, arc_c_min);
2184 arc_p = (arc_c >> 1);
2185 ASSERT(arc_c >= arc_c_min);
2186 ASSERT((int64_t)arc_p >= 0);
2189 if (arc_size > arc_c)
2194 arc_kmem_reap_now(arc_reclaim_strategy_t strat, uint64_t bytes)
2197 kmem_cache_t *prev_cache = NULL;
2198 kmem_cache_t *prev_data_cache = NULL;
2199 extern kmem_cache_t *zio_buf_cache[];
2200 extern kmem_cache_t *zio_data_buf_cache[];
2203 * An aggressive reclamation will shrink the cache size as well as
2204 * reap free buffers from the arc kmem caches.
2206 if (strat == ARC_RECLAIM_AGGR)
2209 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2210 if (zio_buf_cache[i] != prev_cache) {
2211 prev_cache = zio_buf_cache[i];
2212 kmem_cache_reap_now(zio_buf_cache[i]);
2214 if (zio_data_buf_cache[i] != prev_data_cache) {
2215 prev_data_cache = zio_data_buf_cache[i];
2216 kmem_cache_reap_now(zio_data_buf_cache[i]);
2220 kmem_cache_reap_now(buf_cache);
2221 kmem_cache_reap_now(hdr_cache);
2225 * Unlike other ZFS implementations this thread is only responsible for
2226 * adapting the target ARC size on Linux. The responsibility for memory
2227 * reclamation has been entirely delegated to the arc_shrinker_func()
2228 * which is registered with the VM. To reflect this change in behavior
2229 * the arc_reclaim thread has been renamed to arc_adapt.
2232 arc_adapt_thread(void)
2237 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2239 mutex_enter(&arc_reclaim_thr_lock);
2240 while (arc_thread_exit == 0) {
2242 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2244 if (spa_get_random(100) == 0) {
2247 if (last_reclaim == ARC_RECLAIM_CONS) {
2248 last_reclaim = ARC_RECLAIM_AGGR;
2250 last_reclaim = ARC_RECLAIM_CONS;
2254 last_reclaim = ARC_RECLAIM_AGGR;
2258 /* reset the growth delay for every reclaim */
2259 arc_grow_time = ddi_get_lbolt()+(zfs_arc_grow_retry * hz);
2261 arc_kmem_reap_now(last_reclaim, 0);
2264 #endif /* !_KERNEL */
2266 /* No recent memory pressure allow the ARC to grow. */
2267 if (arc_no_grow && ddi_get_lbolt() >= arc_grow_time)
2268 arc_no_grow = FALSE;
2271 * Keep meta data usage within limits, arc_shrink() is not
2272 * used to avoid collapsing the arc_c value when only the
2273 * arc_meta_limit is being exceeded.
2275 prune = (int64_t)arc_meta_used - (int64_t)arc_meta_limit;
2277 arc_adjust_meta(prune, B_TRUE);
2281 if (arc_eviction_list != NULL)
2282 arc_do_user_evicts();
2284 /* block until needed, or one second, whichever is shorter */
2285 CALLB_CPR_SAFE_BEGIN(&cpr);
2286 (void) cv_timedwait_interruptible(&arc_reclaim_thr_cv,
2287 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2288 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2291 /* Allow the module options to be changed */
2292 if (zfs_arc_max > 64 << 20 &&
2293 zfs_arc_max < physmem * PAGESIZE &&
2294 zfs_arc_max != arc_c_max)
2295 arc_c_max = zfs_arc_max;
2297 if (zfs_arc_min > 0 &&
2298 zfs_arc_min < arc_c_max &&
2299 zfs_arc_min != arc_c_min)
2300 arc_c_min = zfs_arc_min;
2302 if (zfs_arc_meta_limit > 0 &&
2303 zfs_arc_meta_limit <= arc_c_max &&
2304 zfs_arc_meta_limit != arc_meta_limit)
2305 arc_meta_limit = zfs_arc_meta_limit;
2311 arc_thread_exit = 0;
2312 cv_broadcast(&arc_reclaim_thr_cv);
2313 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2319 * Determine the amount of memory eligible for eviction contained in the
2320 * ARC. All clean data reported by the ghost lists can always be safely
2321 * evicted. Due to arc_c_min, the same does not hold for all clean data
2322 * contained by the regular mru and mfu lists.
2324 * In the case of the regular mru and mfu lists, we need to report as
2325 * much clean data as possible, such that evicting that same reported
2326 * data will not bring arc_size below arc_c_min. Thus, in certain
2327 * circumstances, the total amount of clean data in the mru and mfu
2328 * lists might not actually be evictable.
2330 * The following two distinct cases are accounted for:
2332 * 1. The sum of the amount of dirty data contained by both the mru and
2333 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2334 * is greater than or equal to arc_c_min.
2335 * (i.e. amount of dirty data >= arc_c_min)
2337 * This is the easy case; all clean data contained by the mru and mfu
2338 * lists is evictable. Evicting all clean data can only drop arc_size
2339 * to the amount of dirty data, which is greater than arc_c_min.
2341 * 2. The sum of the amount of dirty data contained by both the mru and
2342 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2343 * is less than arc_c_min.
2344 * (i.e. arc_c_min > amount of dirty data)
2346 * 2.1. arc_size is greater than or equal arc_c_min.
2347 * (i.e. arc_size >= arc_c_min > amount of dirty data)
2349 * In this case, not all clean data from the regular mru and mfu
2350 * lists is actually evictable; we must leave enough clean data
2351 * to keep arc_size above arc_c_min. Thus, the maximum amount of
2352 * evictable data from the two lists combined, is exactly the
2353 * difference between arc_size and arc_c_min.
2355 * 2.2. arc_size is less than arc_c_min
2356 * (i.e. arc_c_min > arc_size > amount of dirty data)
2358 * In this case, none of the data contained in the mru and mfu
2359 * lists is evictable, even if it's clean. Since arc_size is
2360 * already below arc_c_min, evicting any more would only
2361 * increase this negative difference.
2364 arc_evictable_memory(void) {
2365 uint64_t arc_clean =
2366 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
2367 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
2368 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
2369 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
2370 uint64_t ghost_clean =
2371 arc_mru_ghost->arcs_lsize[ARC_BUFC_DATA] +
2372 arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] +
2373 arc_mfu_ghost->arcs_lsize[ARC_BUFC_DATA] +
2374 arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA];
2375 uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0);
2377 if (arc_dirty >= arc_c_min)
2378 return (ghost_clean + arc_clean);
2380 return (ghost_clean + MAX((int64_t)arc_size - (int64_t)arc_c_min, 0));
2384 __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc)
2388 /* The arc is considered warm once reclaim has occurred */
2389 if (unlikely(arc_warm == B_FALSE))
2392 /* Return the potential number of reclaimable pages */
2393 pages = btop(arc_evictable_memory());
2394 if (sc->nr_to_scan == 0)
2397 /* Not allowed to perform filesystem reclaim */
2398 if (!(sc->gfp_mask & __GFP_FS))
2401 /* Reclaim in progress */
2402 if (mutex_tryenter(&arc_reclaim_thr_lock) == 0)
2406 * Evict the requested number of pages by shrinking arc_c the
2407 * requested amount. If there is nothing left to evict just
2408 * reap whatever we can from the various arc slabs.
2411 arc_kmem_reap_now(ARC_RECLAIM_AGGR, ptob(sc->nr_to_scan));
2413 arc_kmem_reap_now(ARC_RECLAIM_CONS, ptob(sc->nr_to_scan));
2417 * When direct reclaim is observed it usually indicates a rapid
2418 * increase in memory pressure. This occurs because the kswapd
2419 * threads were unable to asynchronously keep enough free memory
2420 * available. In this case set arc_no_grow to briefly pause arc
2421 * growth to avoid compounding the memory pressure.
2423 if (current_is_kswapd()) {
2424 ARCSTAT_BUMP(arcstat_memory_indirect_count);
2426 arc_no_grow = B_TRUE;
2427 arc_grow_time = ddi_get_lbolt() + (zfs_arc_grow_retry * hz);
2428 ARCSTAT_BUMP(arcstat_memory_direct_count);
2431 mutex_exit(&arc_reclaim_thr_lock);
2435 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func);
2437 SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS);
2438 #endif /* _KERNEL */
2441 * Adapt arc info given the number of bytes we are trying to add and
2442 * the state that we are comming from. This function is only called
2443 * when we are adding new content to the cache.
2446 arc_adapt(int bytes, arc_state_t *state)
2449 uint64_t arc_p_min = (arc_c >> zfs_arc_p_min_shift);
2451 if (state == arc_l2c_only)
2456 * Adapt the target size of the MRU list:
2457 * - if we just hit in the MRU ghost list, then increase
2458 * the target size of the MRU list.
2459 * - if we just hit in the MFU ghost list, then increase
2460 * the target size of the MFU list by decreasing the
2461 * target size of the MRU list.
2463 if (state == arc_mru_ghost) {
2464 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2465 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2466 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2468 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2469 } else if (state == arc_mfu_ghost) {
2472 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2473 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2474 mult = MIN(mult, 10);
2476 delta = MIN(bytes * mult, arc_p);
2477 arc_p = MAX(arc_p_min, arc_p - delta);
2479 ASSERT((int64_t)arc_p >= 0);
2484 if (arc_c >= arc_c_max)
2488 * If we're within (2 * maxblocksize) bytes of the target
2489 * cache size, increment the target cache size
2491 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2492 atomic_add_64(&arc_c, (int64_t)bytes);
2493 if (arc_c > arc_c_max)
2495 else if (state == arc_anon)
2496 atomic_add_64(&arc_p, (int64_t)bytes);
2500 ASSERT((int64_t)arc_p >= 0);
2504 * Check if the cache has reached its limits and eviction is required
2508 arc_evict_needed(arc_buf_contents_t type)
2510 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2516 return (arc_size > arc_c);
2520 * The buffer, supplied as the first argument, needs a data block.
2521 * So, if we are at cache max, determine which cache should be victimized.
2522 * We have the following cases:
2524 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2525 * In this situation if we're out of space, but the resident size of the MFU is
2526 * under the limit, victimize the MFU cache to satisfy this insertion request.
2528 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2529 * Here, we've used up all of the available space for the MRU, so we need to
2530 * evict from our own cache instead. Evict from the set of resident MRU
2533 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2534 * c minus p represents the MFU space in the cache, since p is the size of the
2535 * cache that is dedicated to the MRU. In this situation there's still space on
2536 * the MFU side, so the MRU side needs to be victimized.
2538 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2539 * MFU's resident set is consuming more space than it has been allotted. In
2540 * this situation, we must victimize our own cache, the MFU, for this insertion.
2543 arc_get_data_buf(arc_buf_t *buf)
2545 arc_state_t *state = buf->b_hdr->b_state;
2546 uint64_t size = buf->b_hdr->b_size;
2547 arc_buf_contents_t type = buf->b_hdr->b_type;
2549 arc_adapt(size, state);
2552 * We have not yet reached cache maximum size,
2553 * just allocate a new buffer.
2555 if (!arc_evict_needed(type)) {
2556 if (type == ARC_BUFC_METADATA) {
2557 buf->b_data = zio_buf_alloc(size);
2558 arc_space_consume(size, ARC_SPACE_DATA);
2560 ASSERT(type == ARC_BUFC_DATA);
2561 buf->b_data = zio_data_buf_alloc(size);
2562 ARCSTAT_INCR(arcstat_data_size, size);
2563 atomic_add_64(&arc_size, size);
2569 * If we are prefetching from the mfu ghost list, this buffer
2570 * will end up on the mru list; so steal space from there.
2572 if (state == arc_mfu_ghost)
2573 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2574 else if (state == arc_mru_ghost)
2577 if (state == arc_mru || state == arc_anon) {
2578 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2579 state = (arc_mfu->arcs_lsize[type] >= size &&
2580 arc_p > mru_used) ? arc_mfu : arc_mru;
2583 uint64_t mfu_space = arc_c - arc_p;
2584 state = (arc_mru->arcs_lsize[type] >= size &&
2585 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2588 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2589 if (type == ARC_BUFC_METADATA) {
2590 buf->b_data = zio_buf_alloc(size);
2591 arc_space_consume(size, ARC_SPACE_DATA);
2594 * If we are unable to recycle an existing meta buffer
2595 * signal the reclaim thread. It will notify users
2596 * via the prune callback to drop references. The
2597 * prune callback in run in the context of the reclaim
2598 * thread to avoid deadlocking on the hash_lock.
2600 cv_signal(&arc_reclaim_thr_cv);
2602 ASSERT(type == ARC_BUFC_DATA);
2603 buf->b_data = zio_data_buf_alloc(size);
2604 ARCSTAT_INCR(arcstat_data_size, size);
2605 atomic_add_64(&arc_size, size);
2608 ARCSTAT_BUMP(arcstat_recycle_miss);
2610 ASSERT(buf->b_data != NULL);
2613 * Update the state size. Note that ghost states have a
2614 * "ghost size" and so don't need to be updated.
2616 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2617 arc_buf_hdr_t *hdr = buf->b_hdr;
2619 atomic_add_64(&hdr->b_state->arcs_size, size);
2620 if (list_link_active(&hdr->b_arc_node)) {
2621 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2622 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2625 * If we are growing the cache, and we are adding anonymous
2626 * data, and we have outgrown arc_p, update arc_p
2628 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2629 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2630 arc_p = MIN(arc_c, arc_p + size);
2635 * This routine is called whenever a buffer is accessed.
2636 * NOTE: the hash lock is dropped in this function.
2639 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2643 ASSERT(MUTEX_HELD(hash_lock));
2645 if (buf->b_state == arc_anon) {
2647 * This buffer is not in the cache, and does not
2648 * appear in our "ghost" list. Add the new buffer
2652 ASSERT(buf->b_arc_access == 0);
2653 buf->b_arc_access = ddi_get_lbolt();
2654 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2655 arc_change_state(arc_mru, buf, hash_lock);
2657 } else if (buf->b_state == arc_mru) {
2658 now = ddi_get_lbolt();
2661 * If this buffer is here because of a prefetch, then either:
2662 * - clear the flag if this is a "referencing" read
2663 * (any subsequent access will bump this into the MFU state).
2665 * - move the buffer to the head of the list if this is
2666 * another prefetch (to make it less likely to be evicted).
2668 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2669 if (refcount_count(&buf->b_refcnt) == 0) {
2670 ASSERT(list_link_active(&buf->b_arc_node));
2672 buf->b_flags &= ~ARC_PREFETCH;
2673 ARCSTAT_BUMP(arcstat_mru_hits);
2675 buf->b_arc_access = now;
2680 * This buffer has been "accessed" only once so far,
2681 * but it is still in the cache. Move it to the MFU
2684 if (now > buf->b_arc_access + ARC_MINTIME) {
2686 * More than 125ms have passed since we
2687 * instantiated this buffer. Move it to the
2688 * most frequently used state.
2690 buf->b_arc_access = now;
2691 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2692 arc_change_state(arc_mfu, buf, hash_lock);
2694 ARCSTAT_BUMP(arcstat_mru_hits);
2695 } else if (buf->b_state == arc_mru_ghost) {
2696 arc_state_t *new_state;
2698 * This buffer has been "accessed" recently, but
2699 * was evicted from the cache. Move it to the
2703 if (buf->b_flags & ARC_PREFETCH) {
2704 new_state = arc_mru;
2705 if (refcount_count(&buf->b_refcnt) > 0)
2706 buf->b_flags &= ~ARC_PREFETCH;
2707 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2709 new_state = arc_mfu;
2710 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2713 buf->b_arc_access = ddi_get_lbolt();
2714 arc_change_state(new_state, buf, hash_lock);
2716 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2717 } else if (buf->b_state == arc_mfu) {
2719 * This buffer has been accessed more than once and is
2720 * still in the cache. Keep it in the MFU state.
2722 * NOTE: an add_reference() that occurred when we did
2723 * the arc_read() will have kicked this off the list.
2724 * If it was a prefetch, we will explicitly move it to
2725 * the head of the list now.
2727 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2728 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2729 ASSERT(list_link_active(&buf->b_arc_node));
2731 ARCSTAT_BUMP(arcstat_mfu_hits);
2732 buf->b_arc_access = ddi_get_lbolt();
2733 } else if (buf->b_state == arc_mfu_ghost) {
2734 arc_state_t *new_state = arc_mfu;
2736 * This buffer has been accessed more than once but has
2737 * been evicted from the cache. Move it back to the
2741 if (buf->b_flags & ARC_PREFETCH) {
2743 * This is a prefetch access...
2744 * move this block back to the MRU state.
2746 ASSERT0(refcount_count(&buf->b_refcnt));
2747 new_state = arc_mru;
2750 buf->b_arc_access = ddi_get_lbolt();
2751 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2752 arc_change_state(new_state, buf, hash_lock);
2754 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2755 } else if (buf->b_state == arc_l2c_only) {
2757 * This buffer is on the 2nd Level ARC.
2760 buf->b_arc_access = ddi_get_lbolt();
2761 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2762 arc_change_state(arc_mfu, buf, hash_lock);
2764 ASSERT(!"invalid arc state");
2768 /* a generic arc_done_func_t which you can use */
2771 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2773 if (zio == NULL || zio->io_error == 0)
2774 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2775 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2778 /* a generic arc_done_func_t */
2780 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2782 arc_buf_t **bufp = arg;
2783 if (zio && zio->io_error) {
2784 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2788 ASSERT(buf->b_data);
2793 arc_read_done(zio_t *zio)
2795 arc_buf_hdr_t *hdr, *found;
2797 arc_buf_t *abuf; /* buffer we're assigning to callback */
2798 kmutex_t *hash_lock;
2799 arc_callback_t *callback_list, *acb;
2800 int freeable = FALSE;
2802 buf = zio->io_private;
2806 * The hdr was inserted into hash-table and removed from lists
2807 * prior to starting I/O. We should find this header, since
2808 * it's in the hash table, and it should be legit since it's
2809 * not possible to evict it during the I/O. The only possible
2810 * reason for it not to be found is if we were freed during the
2813 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2816 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2817 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2818 (found == hdr && HDR_L2_READING(hdr)));
2820 hdr->b_flags &= ~ARC_L2_EVICTED;
2821 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2822 hdr->b_flags &= ~ARC_L2CACHE;
2824 /* byteswap if necessary */
2825 callback_list = hdr->b_acb;
2826 ASSERT(callback_list != NULL);
2827 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2828 dmu_object_byteswap_t bswap =
2829 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2830 if (BP_GET_LEVEL(zio->io_bp) > 0)
2831 byteswap_uint64_array(buf->b_data, hdr->b_size);
2833 dmu_ot_byteswap[bswap].ob_func(buf->b_data, hdr->b_size);
2836 arc_cksum_compute(buf, B_FALSE);
2838 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2840 * Only call arc_access on anonymous buffers. This is because
2841 * if we've issued an I/O for an evicted buffer, we've already
2842 * called arc_access (to prevent any simultaneous readers from
2843 * getting confused).
2845 arc_access(hdr, hash_lock);
2848 /* create copies of the data buffer for the callers */
2850 for (acb = callback_list; acb; acb = acb->acb_next) {
2851 if (acb->acb_done) {
2853 ARCSTAT_BUMP(arcstat_duplicate_reads);
2854 abuf = arc_buf_clone(buf);
2856 acb->acb_buf = abuf;
2861 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2862 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2864 ASSERT(buf->b_efunc == NULL);
2865 ASSERT(hdr->b_datacnt == 1);
2866 hdr->b_flags |= ARC_BUF_AVAILABLE;
2869 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2871 if (zio->io_error != 0) {
2872 hdr->b_flags |= ARC_IO_ERROR;
2873 if (hdr->b_state != arc_anon)
2874 arc_change_state(arc_anon, hdr, hash_lock);
2875 if (HDR_IN_HASH_TABLE(hdr))
2876 buf_hash_remove(hdr);
2877 freeable = refcount_is_zero(&hdr->b_refcnt);
2881 * Broadcast before we drop the hash_lock to avoid the possibility
2882 * that the hdr (and hence the cv) might be freed before we get to
2883 * the cv_broadcast().
2885 cv_broadcast(&hdr->b_cv);
2888 mutex_exit(hash_lock);
2891 * This block was freed while we waited for the read to
2892 * complete. It has been removed from the hash table and
2893 * moved to the anonymous state (so that it won't show up
2896 ASSERT3P(hdr->b_state, ==, arc_anon);
2897 freeable = refcount_is_zero(&hdr->b_refcnt);
2900 /* execute each callback and free its structure */
2901 while ((acb = callback_list) != NULL) {
2903 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2905 if (acb->acb_zio_dummy != NULL) {
2906 acb->acb_zio_dummy->io_error = zio->io_error;
2907 zio_nowait(acb->acb_zio_dummy);
2910 callback_list = acb->acb_next;
2911 kmem_free(acb, sizeof (arc_callback_t));
2915 arc_hdr_destroy(hdr);
2919 * "Read" the block at the specified DVA (in bp) via the
2920 * cache. If the block is found in the cache, invoke the provided
2921 * callback immediately and return. Note that the `zio' parameter
2922 * in the callback will be NULL in this case, since no IO was
2923 * required. If the block is not in the cache pass the read request
2924 * on to the spa with a substitute callback function, so that the
2925 * requested block will be added to the cache.
2927 * If a read request arrives for a block that has a read in-progress,
2928 * either wait for the in-progress read to complete (and return the
2929 * results); or, if this is a read with a "done" func, add a record
2930 * to the read to invoke the "done" func when the read completes,
2931 * and return; or just return.
2933 * arc_read_done() will invoke all the requested "done" functions
2934 * for readers of this block.
2937 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
2938 void *private, int priority, int zio_flags, uint32_t *arc_flags,
2939 const zbookmark_t *zb)
2942 arc_buf_t *buf = NULL;
2943 kmutex_t *hash_lock;
2945 uint64_t guid = spa_load_guid(spa);
2948 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2950 if (hdr && hdr->b_datacnt > 0) {
2952 *arc_flags |= ARC_CACHED;
2954 if (HDR_IO_IN_PROGRESS(hdr)) {
2956 if (*arc_flags & ARC_WAIT) {
2957 cv_wait(&hdr->b_cv, hash_lock);
2958 mutex_exit(hash_lock);
2961 ASSERT(*arc_flags & ARC_NOWAIT);
2964 arc_callback_t *acb = NULL;
2966 acb = kmem_zalloc(sizeof (arc_callback_t),
2968 acb->acb_done = done;
2969 acb->acb_private = private;
2971 acb->acb_zio_dummy = zio_null(pio,
2972 spa, NULL, NULL, NULL, zio_flags);
2974 ASSERT(acb->acb_done != NULL);
2975 acb->acb_next = hdr->b_acb;
2977 add_reference(hdr, hash_lock, private);
2978 mutex_exit(hash_lock);
2981 mutex_exit(hash_lock);
2985 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2988 add_reference(hdr, hash_lock, private);
2990 * If this block is already in use, create a new
2991 * copy of the data so that we will be guaranteed
2992 * that arc_release() will always succeed.
2996 ASSERT(buf->b_data);
2997 if (HDR_BUF_AVAILABLE(hdr)) {
2998 ASSERT(buf->b_efunc == NULL);
2999 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3001 buf = arc_buf_clone(buf);
3004 } else if (*arc_flags & ARC_PREFETCH &&
3005 refcount_count(&hdr->b_refcnt) == 0) {
3006 hdr->b_flags |= ARC_PREFETCH;
3008 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
3009 arc_access(hdr, hash_lock);
3010 if (*arc_flags & ARC_L2CACHE)
3011 hdr->b_flags |= ARC_L2CACHE;
3012 if (*arc_flags & ARC_L2COMPRESS)
3013 hdr->b_flags |= ARC_L2COMPRESS;
3014 mutex_exit(hash_lock);
3015 ARCSTAT_BUMP(arcstat_hits);
3016 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3017 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3018 data, metadata, hits);
3021 done(NULL, buf, private);
3023 uint64_t size = BP_GET_LSIZE(bp);
3024 arc_callback_t *acb;
3027 boolean_t devw = B_FALSE;
3030 /* this block is not in the cache */
3031 arc_buf_hdr_t *exists;
3032 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
3033 buf = arc_buf_alloc(spa, size, private, type);
3035 hdr->b_dva = *BP_IDENTITY(bp);
3036 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3037 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3038 exists = buf_hash_insert(hdr, &hash_lock);
3040 /* somebody beat us to the hash insert */
3041 mutex_exit(hash_lock);
3042 buf_discard_identity(hdr);
3043 (void) arc_buf_remove_ref(buf, private);
3044 goto top; /* restart the IO request */
3046 /* if this is a prefetch, we don't have a reference */
3047 if (*arc_flags & ARC_PREFETCH) {
3048 (void) remove_reference(hdr, hash_lock,
3050 hdr->b_flags |= ARC_PREFETCH;
3052 if (*arc_flags & ARC_L2CACHE)
3053 hdr->b_flags |= ARC_L2CACHE;
3054 if (*arc_flags & ARC_L2COMPRESS)
3055 hdr->b_flags |= ARC_L2COMPRESS;
3056 if (BP_GET_LEVEL(bp) > 0)
3057 hdr->b_flags |= ARC_INDIRECT;
3059 /* this block is in the ghost cache */
3060 ASSERT(GHOST_STATE(hdr->b_state));
3061 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3062 ASSERT0(refcount_count(&hdr->b_refcnt));
3063 ASSERT(hdr->b_buf == NULL);
3065 /* if this is a prefetch, we don't have a reference */
3066 if (*arc_flags & ARC_PREFETCH)
3067 hdr->b_flags |= ARC_PREFETCH;
3069 add_reference(hdr, hash_lock, private);
3070 if (*arc_flags & ARC_L2CACHE)
3071 hdr->b_flags |= ARC_L2CACHE;
3072 if (*arc_flags & ARC_L2COMPRESS)
3073 hdr->b_flags |= ARC_L2COMPRESS;
3074 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3077 buf->b_efunc = NULL;
3078 buf->b_private = NULL;
3081 ASSERT(hdr->b_datacnt == 0);
3083 arc_get_data_buf(buf);
3084 arc_access(hdr, hash_lock);
3087 ASSERT(!GHOST_STATE(hdr->b_state));
3089 acb = kmem_zalloc(sizeof (arc_callback_t), KM_PUSHPAGE);
3090 acb->acb_done = done;
3091 acb->acb_private = private;
3093 ASSERT(hdr->b_acb == NULL);
3095 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3097 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3098 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3099 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3100 addr = hdr->b_l2hdr->b_daddr;
3102 * Lock out device removal.
3104 if (vdev_is_dead(vd) ||
3105 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3109 mutex_exit(hash_lock);
3111 ASSERT3U(hdr->b_size, ==, size);
3112 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3113 uint64_t, size, zbookmark_t *, zb);
3114 ARCSTAT_BUMP(arcstat_misses);
3115 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3116 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3117 data, metadata, misses);
3119 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3121 * Read from the L2ARC if the following are true:
3122 * 1. The L2ARC vdev was previously cached.
3123 * 2. This buffer still has L2ARC metadata.
3124 * 3. This buffer isn't currently writing to the L2ARC.
3125 * 4. The L2ARC entry wasn't evicted, which may
3126 * also have invalidated the vdev.
3127 * 5. This isn't prefetch and l2arc_noprefetch is set.
3129 if (hdr->b_l2hdr != NULL &&
3130 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3131 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3132 l2arc_read_callback_t *cb;
3134 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3135 ARCSTAT_BUMP(arcstat_l2_hits);
3137 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3139 cb->l2rcb_buf = buf;
3140 cb->l2rcb_spa = spa;
3143 cb->l2rcb_flags = zio_flags;
3144 cb->l2rcb_compress = hdr->b_l2hdr->b_compress;
3147 * l2arc read. The SCL_L2ARC lock will be
3148 * released by l2arc_read_done().
3149 * Issue a null zio if the underlying buffer
3150 * was squashed to zero size by compression.
3152 if (hdr->b_l2hdr->b_compress ==
3153 ZIO_COMPRESS_EMPTY) {
3154 rzio = zio_null(pio, spa, vd,
3155 l2arc_read_done, cb,
3156 zio_flags | ZIO_FLAG_DONT_CACHE |
3158 ZIO_FLAG_DONT_PROPAGATE |
3159 ZIO_FLAG_DONT_RETRY);
3161 rzio = zio_read_phys(pio, vd, addr,
3162 hdr->b_l2hdr->b_asize,
3163 buf->b_data, ZIO_CHECKSUM_OFF,
3164 l2arc_read_done, cb, priority,
3165 zio_flags | ZIO_FLAG_DONT_CACHE |
3167 ZIO_FLAG_DONT_PROPAGATE |
3168 ZIO_FLAG_DONT_RETRY, B_FALSE);
3170 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3172 ARCSTAT_INCR(arcstat_l2_read_bytes,
3173 hdr->b_l2hdr->b_asize);
3175 if (*arc_flags & ARC_NOWAIT) {
3180 ASSERT(*arc_flags & ARC_WAIT);
3181 if (zio_wait(rzio) == 0)
3184 /* l2arc read error; goto zio_read() */
3186 DTRACE_PROBE1(l2arc__miss,
3187 arc_buf_hdr_t *, hdr);
3188 ARCSTAT_BUMP(arcstat_l2_misses);
3189 if (HDR_L2_WRITING(hdr))
3190 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3191 spa_config_exit(spa, SCL_L2ARC, vd);
3195 spa_config_exit(spa, SCL_L2ARC, vd);
3196 if (l2arc_ndev != 0) {
3197 DTRACE_PROBE1(l2arc__miss,
3198 arc_buf_hdr_t *, hdr);
3199 ARCSTAT_BUMP(arcstat_l2_misses);
3203 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3204 arc_read_done, buf, priority, zio_flags, zb);
3206 if (*arc_flags & ARC_WAIT)
3207 return (zio_wait(rzio));
3209 ASSERT(*arc_flags & ARC_NOWAIT);
3216 arc_add_prune_callback(arc_prune_func_t *func, void *private)
3220 p = kmem_alloc(sizeof(*p), KM_SLEEP);
3222 p->p_private = private;
3223 list_link_init(&p->p_node);
3224 refcount_create(&p->p_refcnt);
3226 mutex_enter(&arc_prune_mtx);
3227 refcount_add(&p->p_refcnt, &arc_prune_list);
3228 list_insert_head(&arc_prune_list, p);
3229 mutex_exit(&arc_prune_mtx);
3235 arc_remove_prune_callback(arc_prune_t *p)
3237 mutex_enter(&arc_prune_mtx);
3238 list_remove(&arc_prune_list, p);
3239 if (refcount_remove(&p->p_refcnt, &arc_prune_list) == 0) {
3240 refcount_destroy(&p->p_refcnt);
3241 kmem_free(p, sizeof (*p));
3243 mutex_exit(&arc_prune_mtx);
3247 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3249 ASSERT(buf->b_hdr != NULL);
3250 ASSERT(buf->b_hdr->b_state != arc_anon);
3251 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3252 ASSERT(buf->b_efunc == NULL);
3253 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3255 buf->b_efunc = func;
3256 buf->b_private = private;
3260 * Notify the arc that a block was freed, and thus will never be used again.
3263 arc_freed(spa_t *spa, const blkptr_t *bp)
3266 kmutex_t *hash_lock;
3267 uint64_t guid = spa_load_guid(spa);
3269 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3273 if (HDR_BUF_AVAILABLE(hdr)) {
3274 arc_buf_t *buf = hdr->b_buf;
3275 add_reference(hdr, hash_lock, FTAG);
3276 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3277 mutex_exit(hash_lock);
3279 arc_release(buf, FTAG);
3280 (void) arc_buf_remove_ref(buf, FTAG);
3282 mutex_exit(hash_lock);
3288 * This is used by the DMU to let the ARC know that a buffer is
3289 * being evicted, so the ARC should clean up. If this arc buf
3290 * is not yet in the evicted state, it will be put there.
3293 arc_buf_evict(arc_buf_t *buf)
3296 kmutex_t *hash_lock;
3299 mutex_enter(&buf->b_evict_lock);
3303 * We are in arc_do_user_evicts().
3305 ASSERT(buf->b_data == NULL);
3306 mutex_exit(&buf->b_evict_lock);
3308 } else if (buf->b_data == NULL) {
3309 arc_buf_t copy = *buf; /* structure assignment */
3311 * We are on the eviction list; process this buffer now
3312 * but let arc_do_user_evicts() do the reaping.
3314 buf->b_efunc = NULL;
3315 mutex_exit(&buf->b_evict_lock);
3316 VERIFY(copy.b_efunc(©) == 0);
3319 hash_lock = HDR_LOCK(hdr);
3320 mutex_enter(hash_lock);
3322 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3324 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3325 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3328 * Pull this buffer off of the hdr
3331 while (*bufp != buf)
3332 bufp = &(*bufp)->b_next;
3333 *bufp = buf->b_next;
3335 ASSERT(buf->b_data != NULL);
3336 arc_buf_destroy(buf, FALSE, FALSE);
3338 if (hdr->b_datacnt == 0) {
3339 arc_state_t *old_state = hdr->b_state;
3340 arc_state_t *evicted_state;
3342 ASSERT(hdr->b_buf == NULL);
3343 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3346 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3348 mutex_enter(&old_state->arcs_mtx);
3349 mutex_enter(&evicted_state->arcs_mtx);
3351 arc_change_state(evicted_state, hdr, hash_lock);
3352 ASSERT(HDR_IN_HASH_TABLE(hdr));
3353 hdr->b_flags |= ARC_IN_HASH_TABLE;
3354 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3356 mutex_exit(&evicted_state->arcs_mtx);
3357 mutex_exit(&old_state->arcs_mtx);
3359 mutex_exit(hash_lock);
3360 mutex_exit(&buf->b_evict_lock);
3362 VERIFY(buf->b_efunc(buf) == 0);
3363 buf->b_efunc = NULL;
3364 buf->b_private = NULL;
3367 kmem_cache_free(buf_cache, buf);
3372 * Release this buffer from the cache. This must be done
3373 * after a read and prior to modifying the buffer contents.
3374 * If the buffer has more than one reference, we must make
3375 * a new hdr for the buffer.
3378 arc_release(arc_buf_t *buf, void *tag)
3381 kmutex_t *hash_lock = NULL;
3382 l2arc_buf_hdr_t *l2hdr;
3383 uint64_t buf_size = 0;
3386 * It would be nice to assert that if it's DMU metadata (level >
3387 * 0 || it's the dnode file), then it must be syncing context.
3388 * But we don't know that information at this level.
3391 mutex_enter(&buf->b_evict_lock);
3394 /* this buffer is not on any list */
3395 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3397 if (hdr->b_state == arc_anon) {
3398 /* this buffer is already released */
3399 ASSERT(buf->b_efunc == NULL);
3401 hash_lock = HDR_LOCK(hdr);
3402 mutex_enter(hash_lock);
3404 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3407 l2hdr = hdr->b_l2hdr;
3409 mutex_enter(&l2arc_buflist_mtx);
3410 hdr->b_l2hdr = NULL;
3411 buf_size = hdr->b_size;
3415 * Do we have more than one buf?
3417 if (hdr->b_datacnt > 1) {
3418 arc_buf_hdr_t *nhdr;
3420 uint64_t blksz = hdr->b_size;
3421 uint64_t spa = hdr->b_spa;
3422 arc_buf_contents_t type = hdr->b_type;
3423 uint32_t flags = hdr->b_flags;
3425 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3427 * Pull the data off of this hdr and attach it to
3428 * a new anonymous hdr.
3430 (void) remove_reference(hdr, hash_lock, tag);
3432 while (*bufp != buf)
3433 bufp = &(*bufp)->b_next;
3434 *bufp = buf->b_next;
3437 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3438 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3439 if (refcount_is_zero(&hdr->b_refcnt)) {
3440 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3441 ASSERT3U(*size, >=, hdr->b_size);
3442 atomic_add_64(size, -hdr->b_size);
3446 * We're releasing a duplicate user data buffer, update
3447 * our statistics accordingly.
3449 if (hdr->b_type == ARC_BUFC_DATA) {
3450 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3451 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3454 hdr->b_datacnt -= 1;
3455 arc_cksum_verify(buf);
3457 mutex_exit(hash_lock);
3459 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3460 nhdr->b_size = blksz;
3462 nhdr->b_type = type;
3464 nhdr->b_state = arc_anon;
3465 nhdr->b_arc_access = 0;
3466 nhdr->b_flags = flags & ARC_L2_WRITING;
3467 nhdr->b_l2hdr = NULL;
3468 nhdr->b_datacnt = 1;
3469 nhdr->b_freeze_cksum = NULL;
3470 (void) refcount_add(&nhdr->b_refcnt, tag);
3472 mutex_exit(&buf->b_evict_lock);
3473 atomic_add_64(&arc_anon->arcs_size, blksz);
3475 mutex_exit(&buf->b_evict_lock);
3476 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3477 ASSERT(!list_link_active(&hdr->b_arc_node));
3478 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3479 if (hdr->b_state != arc_anon)
3480 arc_change_state(arc_anon, hdr, hash_lock);
3481 hdr->b_arc_access = 0;
3483 mutex_exit(hash_lock);
3485 buf_discard_identity(hdr);
3488 buf->b_efunc = NULL;
3489 buf->b_private = NULL;
3492 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
3493 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3494 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3495 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
3496 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3497 mutex_exit(&l2arc_buflist_mtx);
3502 arc_released(arc_buf_t *buf)
3506 mutex_enter(&buf->b_evict_lock);
3507 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3508 mutex_exit(&buf->b_evict_lock);
3513 arc_has_callback(arc_buf_t *buf)
3517 mutex_enter(&buf->b_evict_lock);
3518 callback = (buf->b_efunc != NULL);
3519 mutex_exit(&buf->b_evict_lock);
3525 arc_referenced(arc_buf_t *buf)
3529 mutex_enter(&buf->b_evict_lock);
3530 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3531 mutex_exit(&buf->b_evict_lock);
3532 return (referenced);
3537 arc_write_ready(zio_t *zio)
3539 arc_write_callback_t *callback = zio->io_private;
3540 arc_buf_t *buf = callback->awcb_buf;
3541 arc_buf_hdr_t *hdr = buf->b_hdr;
3543 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3544 callback->awcb_ready(zio, buf, callback->awcb_private);
3547 * If the IO is already in progress, then this is a re-write
3548 * attempt, so we need to thaw and re-compute the cksum.
3549 * It is the responsibility of the callback to handle the
3550 * accounting for any re-write attempt.
3552 if (HDR_IO_IN_PROGRESS(hdr)) {
3553 mutex_enter(&hdr->b_freeze_lock);
3554 if (hdr->b_freeze_cksum != NULL) {
3555 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3556 hdr->b_freeze_cksum = NULL;
3558 mutex_exit(&hdr->b_freeze_lock);
3560 arc_cksum_compute(buf, B_FALSE);
3561 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3565 arc_write_done(zio_t *zio)
3567 arc_write_callback_t *callback = zio->io_private;
3568 arc_buf_t *buf = callback->awcb_buf;
3569 arc_buf_hdr_t *hdr = buf->b_hdr;
3571 ASSERT(hdr->b_acb == NULL);
3573 if (zio->io_error == 0) {
3574 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3575 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3576 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3578 ASSERT(BUF_EMPTY(hdr));
3582 * If the block to be written was all-zero, we may have
3583 * compressed it away. In this case no write was performed
3584 * so there will be no dva/birth/checksum. The buffer must
3585 * therefore remain anonymous (and uncached).
3587 if (!BUF_EMPTY(hdr)) {
3588 arc_buf_hdr_t *exists;
3589 kmutex_t *hash_lock;
3591 ASSERT(zio->io_error == 0);
3593 arc_cksum_verify(buf);
3595 exists = buf_hash_insert(hdr, &hash_lock);
3598 * This can only happen if we overwrite for
3599 * sync-to-convergence, because we remove
3600 * buffers from the hash table when we arc_free().
3602 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3603 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3604 panic("bad overwrite, hdr=%p exists=%p",
3605 (void *)hdr, (void *)exists);
3606 ASSERT(refcount_is_zero(&exists->b_refcnt));
3607 arc_change_state(arc_anon, exists, hash_lock);
3608 mutex_exit(hash_lock);
3609 arc_hdr_destroy(exists);
3610 exists = buf_hash_insert(hdr, &hash_lock);
3611 ASSERT3P(exists, ==, NULL);
3614 ASSERT(hdr->b_datacnt == 1);
3615 ASSERT(hdr->b_state == arc_anon);
3616 ASSERT(BP_GET_DEDUP(zio->io_bp));
3617 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3620 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3621 /* if it's not anon, we are doing a scrub */
3622 if (!exists && hdr->b_state == arc_anon)
3623 arc_access(hdr, hash_lock);
3624 mutex_exit(hash_lock);
3626 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3629 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3630 callback->awcb_done(zio, buf, callback->awcb_private);
3632 kmem_free(callback, sizeof (arc_write_callback_t));
3636 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3637 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
3638 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *done,
3639 void *private, int priority, int zio_flags, const zbookmark_t *zb)
3641 arc_buf_hdr_t *hdr = buf->b_hdr;
3642 arc_write_callback_t *callback;
3645 ASSERT(ready != NULL);
3646 ASSERT(done != NULL);
3647 ASSERT(!HDR_IO_ERROR(hdr));
3648 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3649 ASSERT(hdr->b_acb == NULL);
3651 hdr->b_flags |= ARC_L2CACHE;
3653 hdr->b_flags |= ARC_L2COMPRESS;
3654 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_PUSHPAGE);
3655 callback->awcb_ready = ready;
3656 callback->awcb_done = done;
3657 callback->awcb_private = private;
3658 callback->awcb_buf = buf;
3660 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3661 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3667 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3670 uint64_t available_memory;
3672 if (zfs_arc_memory_throttle_disable)
3675 /* Easily reclaimable memory (free + inactive + arc-evictable) */
3676 available_memory = ptob(spl_kmem_availrmem()) + arc_evictable_memory();
3678 if (available_memory <= zfs_write_limit_max) {
3679 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3680 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
3684 if (inflight_data > available_memory / 4) {
3685 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3686 DMU_TX_STAT_BUMP(dmu_tx_memory_inflight);
3694 arc_tempreserve_clear(uint64_t reserve)
3696 atomic_add_64(&arc_tempreserve, -reserve);
3697 ASSERT((int64_t)arc_tempreserve >= 0);
3701 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3708 * Once in a while, fail for no reason. Everything should cope.
3710 if (spa_get_random(10000) == 0) {
3711 dprintf("forcing random failure\n");
3715 if (reserve > arc_c/4 && !arc_no_grow)
3716 arc_c = MIN(arc_c_max, reserve * 4);
3717 if (reserve > arc_c) {
3718 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
3723 * Don't count loaned bufs as in flight dirty data to prevent long
3724 * network delays from blocking transactions that are ready to be
3725 * assigned to a txg.
3727 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3730 * Writes will, almost always, require additional memory allocations
3731 * in order to compress/encrypt/etc the data. We therefor need to
3732 * make sure that there is sufficient available memory for this.
3734 if ((error = arc_memory_throttle(reserve, anon_size, txg)))
3738 * Throttle writes when the amount of dirty data in the cache
3739 * gets too large. We try to keep the cache less than half full
3740 * of dirty blocks so that our sync times don't grow too large.
3741 * Note: if two requests come in concurrently, we might let them
3742 * both succeed, when one of them should fail. Not a huge deal.
3745 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3746 anon_size > arc_c / 4) {
3747 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3748 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3749 arc_tempreserve>>10,
3750 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3751 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3752 reserve>>10, arc_c>>10);
3753 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
3756 atomic_add_64(&arc_tempreserve, reserve);
3761 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
3762 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
3764 size->value.ui64 = state->arcs_size;
3765 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
3766 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
3770 arc_kstat_update(kstat_t *ksp, int rw)
3772 arc_stats_t *as = ksp->ks_data;
3774 if (rw == KSTAT_WRITE) {
3777 arc_kstat_update_state(arc_anon,
3778 &as->arcstat_anon_size,
3779 &as->arcstat_anon_evict_data,
3780 &as->arcstat_anon_evict_metadata);
3781 arc_kstat_update_state(arc_mru,
3782 &as->arcstat_mru_size,
3783 &as->arcstat_mru_evict_data,
3784 &as->arcstat_mru_evict_metadata);
3785 arc_kstat_update_state(arc_mru_ghost,
3786 &as->arcstat_mru_ghost_size,
3787 &as->arcstat_mru_ghost_evict_data,
3788 &as->arcstat_mru_ghost_evict_metadata);
3789 arc_kstat_update_state(arc_mfu,
3790 &as->arcstat_mfu_size,
3791 &as->arcstat_mfu_evict_data,
3792 &as->arcstat_mfu_evict_metadata);
3793 arc_kstat_update_state(arc_mfu_ghost,
3794 &as->arcstat_mfu_ghost_size,
3795 &as->arcstat_mfu_ghost_evict_data,
3796 &as->arcstat_mfu_ghost_evict_metadata);
3805 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3806 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3808 /* Convert seconds to clock ticks */
3809 zfs_arc_min_prefetch_lifespan = 1 * hz;
3811 /* Start out with 1/8 of all memory */
3812 arc_c = physmem * PAGESIZE / 8;
3816 * On architectures where the physical memory can be larger
3817 * than the addressable space (intel in 32-bit mode), we may
3818 * need to limit the cache to 1/8 of VM size.
3820 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3822 * Register a shrinker to support synchronous (direct) memory
3823 * reclaim from the arc. This is done to prevent kswapd from
3824 * swapping out pages when it is preferable to shrink the arc.
3826 spl_register_shrinker(&arc_shrinker);
3829 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3830 arc_c_min = MAX(arc_c / 4, 64<<20);
3831 /* set max to 1/2 of all memory */
3832 arc_c_max = MAX(arc_c * 4, arc_c_max);
3835 * Allow the tunables to override our calculations if they are
3836 * reasonable (ie. over 64MB)
3838 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3839 arc_c_max = zfs_arc_max;
3840 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3841 arc_c_min = zfs_arc_min;
3844 arc_p = (arc_c >> 1);
3846 /* limit meta-data to 1/4 of the arc capacity */
3847 arc_meta_limit = arc_c_max / 4;
3850 /* Allow the tunable to override if it is reasonable */
3851 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3852 arc_meta_limit = zfs_arc_meta_limit;
3854 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3855 arc_c_min = arc_meta_limit / 2;
3857 /* if kmem_flags are set, lets try to use less memory */
3858 if (kmem_debugging())
3860 if (arc_c < arc_c_min)
3863 arc_anon = &ARC_anon;
3865 arc_mru_ghost = &ARC_mru_ghost;
3867 arc_mfu_ghost = &ARC_mfu_ghost;
3868 arc_l2c_only = &ARC_l2c_only;
3871 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3872 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3873 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3874 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3875 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3876 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3878 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3879 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3880 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3881 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3882 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3883 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3884 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3885 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3886 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3887 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3888 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3889 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3890 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3891 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3892 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3893 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3894 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3895 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3896 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3897 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3901 arc_thread_exit = 0;
3902 list_create(&arc_prune_list, sizeof (arc_prune_t),
3903 offsetof(arc_prune_t, p_node));
3904 arc_eviction_list = NULL;
3905 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
3906 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3907 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3909 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3910 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3912 if (arc_ksp != NULL) {
3913 arc_ksp->ks_data = &arc_stats;
3914 arc_ksp->ks_update = arc_kstat_update;
3915 kstat_install(arc_ksp);
3918 (void) thread_create(NULL, 0, arc_adapt_thread, NULL, 0, &p0,
3919 TS_RUN, minclsyspri);
3924 if (zfs_write_limit_max == 0)
3925 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3927 zfs_write_limit_shift = 0;
3928 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3936 mutex_enter(&arc_reclaim_thr_lock);
3938 spl_unregister_shrinker(&arc_shrinker);
3939 #endif /* _KERNEL */
3941 arc_thread_exit = 1;
3942 while (arc_thread_exit != 0)
3943 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3944 mutex_exit(&arc_reclaim_thr_lock);
3950 if (arc_ksp != NULL) {
3951 kstat_delete(arc_ksp);
3955 mutex_enter(&arc_prune_mtx);
3956 while ((p = list_head(&arc_prune_list)) != NULL) {
3957 list_remove(&arc_prune_list, p);
3958 refcount_remove(&p->p_refcnt, &arc_prune_list);
3959 refcount_destroy(&p->p_refcnt);
3960 kmem_free(p, sizeof (*p));
3962 mutex_exit(&arc_prune_mtx);
3964 list_destroy(&arc_prune_list);
3965 mutex_destroy(&arc_prune_mtx);
3966 mutex_destroy(&arc_eviction_mtx);
3967 mutex_destroy(&arc_reclaim_thr_lock);
3968 cv_destroy(&arc_reclaim_thr_cv);
3970 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3971 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3972 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3973 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3974 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3975 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3976 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3977 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3979 mutex_destroy(&arc_anon->arcs_mtx);
3980 mutex_destroy(&arc_mru->arcs_mtx);
3981 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3982 mutex_destroy(&arc_mfu->arcs_mtx);
3983 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3984 mutex_destroy(&arc_l2c_only->arcs_mtx);
3986 mutex_destroy(&zfs_write_limit_lock);
3990 ASSERT(arc_loaned_bytes == 0);
3996 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3997 * It uses dedicated storage devices to hold cached data, which are populated
3998 * using large infrequent writes. The main role of this cache is to boost
3999 * the performance of random read workloads. The intended L2ARC devices
4000 * include short-stroked disks, solid state disks, and other media with
4001 * substantially faster read latency than disk.
4003 * +-----------------------+
4005 * +-----------------------+
4008 * l2arc_feed_thread() arc_read()
4012 * +---------------+ |
4014 * +---------------+ |
4019 * +-------+ +-------+
4021 * | cache | | cache |
4022 * +-------+ +-------+
4023 * +=========+ .-----.
4024 * : L2ARC : |-_____-|
4025 * : devices : | Disks |
4026 * +=========+ `-_____-'
4028 * Read requests are satisfied from the following sources, in order:
4031 * 2) vdev cache of L2ARC devices
4033 * 4) vdev cache of disks
4036 * Some L2ARC device types exhibit extremely slow write performance.
4037 * To accommodate for this there are some significant differences between
4038 * the L2ARC and traditional cache design:
4040 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4041 * the ARC behave as usual, freeing buffers and placing headers on ghost
4042 * lists. The ARC does not send buffers to the L2ARC during eviction as
4043 * this would add inflated write latencies for all ARC memory pressure.
4045 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4046 * It does this by periodically scanning buffers from the eviction-end of
4047 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4048 * not already there. It scans until a headroom of buffers is satisfied,
4049 * which itself is a buffer for ARC eviction. If a compressible buffer is
4050 * found during scanning and selected for writing to an L2ARC device, we
4051 * temporarily boost scanning headroom during the next scan cycle to make
4052 * sure we adapt to compression effects (which might significantly reduce
4053 * the data volume we write to L2ARC). The thread that does this is
4054 * l2arc_feed_thread(), illustrated below; example sizes are included to
4055 * provide a better sense of ratio than this diagram:
4058 * +---------------------+----------+
4059 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4060 * +---------------------+----------+ | o L2ARC eligible
4061 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4062 * +---------------------+----------+ |
4063 * 15.9 Gbytes ^ 32 Mbytes |
4065 * l2arc_feed_thread()
4067 * l2arc write hand <--[oooo]--'
4071 * +==============================+
4072 * L2ARC dev |####|#|###|###| |####| ... |
4073 * +==============================+
4076 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4077 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4078 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4079 * safe to say that this is an uncommon case, since buffers at the end of
4080 * the ARC lists have moved there due to inactivity.
4082 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4083 * then the L2ARC simply misses copying some buffers. This serves as a
4084 * pressure valve to prevent heavy read workloads from both stalling the ARC
4085 * with waits and clogging the L2ARC with writes. This also helps prevent
4086 * the potential for the L2ARC to churn if it attempts to cache content too
4087 * quickly, such as during backups of the entire pool.
4089 * 5. After system boot and before the ARC has filled main memory, there are
4090 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4091 * lists can remain mostly static. Instead of searching from tail of these
4092 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4093 * for eligible buffers, greatly increasing its chance of finding them.
4095 * The L2ARC device write speed is also boosted during this time so that
4096 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4097 * there are no L2ARC reads, and no fear of degrading read performance
4098 * through increased writes.
4100 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4101 * the vdev queue can aggregate them into larger and fewer writes. Each
4102 * device is written to in a rotor fashion, sweeping writes through
4103 * available space then repeating.
4105 * 7. The L2ARC does not store dirty content. It never needs to flush
4106 * write buffers back to disk based storage.
4108 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4109 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4111 * The performance of the L2ARC can be tweaked by a number of tunables, which
4112 * may be necessary for different workloads:
4114 * l2arc_write_max max write bytes per interval
4115 * l2arc_write_boost extra write bytes during device warmup
4116 * l2arc_noprefetch skip caching prefetched buffers
4117 * l2arc_nocompress skip compressing buffers
4118 * l2arc_headroom number of max device writes to precache
4119 * l2arc_headroom_boost when we find compressed buffers during ARC
4120 * scanning, we multiply headroom by this
4121 * percentage factor for the next scan cycle,
4122 * since more compressed buffers are likely to
4124 * l2arc_feed_secs seconds between L2ARC writing
4126 * Tunables may be removed or added as future performance improvements are
4127 * integrated, and also may become zpool properties.
4129 * There are three key functions that control how the L2ARC warms up:
4131 * l2arc_write_eligible() check if a buffer is eligible to cache
4132 * l2arc_write_size() calculate how much to write
4133 * l2arc_write_interval() calculate sleep delay between writes
4135 * These three functions determine what to write, how much, and how quickly
4140 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4143 * A buffer is *not* eligible for the L2ARC if it:
4144 * 1. belongs to a different spa.
4145 * 2. is already cached on the L2ARC.
4146 * 3. has an I/O in progress (it may be an incomplete read).
4147 * 4. is flagged not eligible (zfs property).
4149 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
4150 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
4157 l2arc_write_size(void)
4162 * Make sure our globals have meaningful values in case the user
4165 size = l2arc_write_max;
4167 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
4168 "be greater than zero, resetting it to the default (%d)",
4170 size = l2arc_write_max = L2ARC_WRITE_SIZE;
4173 if (arc_warm == B_FALSE)
4174 size += l2arc_write_boost;
4181 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4183 clock_t interval, next, now;
4186 * If the ARC lists are busy, increase our write rate; if the
4187 * lists are stale, idle back. This is achieved by checking
4188 * how much we previously wrote - if it was more than half of
4189 * what we wanted, schedule the next write much sooner.
4191 if (l2arc_feed_again && wrote > (wanted / 2))
4192 interval = (hz * l2arc_feed_min_ms) / 1000;
4194 interval = hz * l2arc_feed_secs;
4196 now = ddi_get_lbolt();
4197 next = MAX(now, MIN(now + interval, began + interval));
4203 l2arc_hdr_stat_add(void)
4205 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE);
4206 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4210 l2arc_hdr_stat_remove(void)
4212 ARCSTAT_INCR(arcstat_l2_hdr_size, -HDR_SIZE);
4213 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4217 * Cycle through L2ARC devices. This is how L2ARC load balances.
4218 * If a device is returned, this also returns holding the spa config lock.
4220 static l2arc_dev_t *
4221 l2arc_dev_get_next(void)
4223 l2arc_dev_t *first, *next = NULL;
4226 * Lock out the removal of spas (spa_namespace_lock), then removal
4227 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4228 * both locks will be dropped and a spa config lock held instead.
4230 mutex_enter(&spa_namespace_lock);
4231 mutex_enter(&l2arc_dev_mtx);
4233 /* if there are no vdevs, there is nothing to do */
4234 if (l2arc_ndev == 0)
4238 next = l2arc_dev_last;
4240 /* loop around the list looking for a non-faulted vdev */
4242 next = list_head(l2arc_dev_list);
4244 next = list_next(l2arc_dev_list, next);
4246 next = list_head(l2arc_dev_list);
4249 /* if we have come back to the start, bail out */
4252 else if (next == first)
4255 } while (vdev_is_dead(next->l2ad_vdev));
4257 /* if we were unable to find any usable vdevs, return NULL */
4258 if (vdev_is_dead(next->l2ad_vdev))
4261 l2arc_dev_last = next;
4264 mutex_exit(&l2arc_dev_mtx);
4267 * Grab the config lock to prevent the 'next' device from being
4268 * removed while we are writing to it.
4271 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4272 mutex_exit(&spa_namespace_lock);
4278 * Free buffers that were tagged for destruction.
4281 l2arc_do_free_on_write(void)
4284 l2arc_data_free_t *df, *df_prev;
4286 mutex_enter(&l2arc_free_on_write_mtx);
4287 buflist = l2arc_free_on_write;
4289 for (df = list_tail(buflist); df; df = df_prev) {
4290 df_prev = list_prev(buflist, df);
4291 ASSERT(df->l2df_data != NULL);
4292 ASSERT(df->l2df_func != NULL);
4293 df->l2df_func(df->l2df_data, df->l2df_size);
4294 list_remove(buflist, df);
4295 kmem_free(df, sizeof (l2arc_data_free_t));
4298 mutex_exit(&l2arc_free_on_write_mtx);
4302 * A write to a cache device has completed. Update all headers to allow
4303 * reads from these buffers to begin.
4306 l2arc_write_done(zio_t *zio)
4308 l2arc_write_callback_t *cb;
4311 arc_buf_hdr_t *head, *ab, *ab_prev;
4312 l2arc_buf_hdr_t *abl2;
4313 kmutex_t *hash_lock;
4315 cb = zio->io_private;
4317 dev = cb->l2wcb_dev;
4318 ASSERT(dev != NULL);
4319 head = cb->l2wcb_head;
4320 ASSERT(head != NULL);
4321 buflist = dev->l2ad_buflist;
4322 ASSERT(buflist != NULL);
4323 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4324 l2arc_write_callback_t *, cb);
4326 if (zio->io_error != 0)
4327 ARCSTAT_BUMP(arcstat_l2_writes_error);
4329 mutex_enter(&l2arc_buflist_mtx);
4332 * All writes completed, or an error was hit.
4334 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4335 ab_prev = list_prev(buflist, ab);
4337 hash_lock = HDR_LOCK(ab);
4338 if (!mutex_tryenter(hash_lock)) {
4340 * This buffer misses out. It may be in a stage
4341 * of eviction. Its ARC_L2_WRITING flag will be
4342 * left set, denying reads to this buffer.
4344 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4351 * Release the temporary compressed buffer as soon as possible.
4353 if (abl2->b_compress != ZIO_COMPRESS_OFF)
4354 l2arc_release_cdata_buf(ab);
4356 if (zio->io_error != 0) {
4358 * Error - drop L2ARC entry.
4360 list_remove(buflist, ab);
4361 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4363 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4364 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4365 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4369 * Allow ARC to begin reads to this L2ARC entry.
4371 ab->b_flags &= ~ARC_L2_WRITING;
4373 mutex_exit(hash_lock);
4376 atomic_inc_64(&l2arc_writes_done);
4377 list_remove(buflist, head);
4378 kmem_cache_free(hdr_cache, head);
4379 mutex_exit(&l2arc_buflist_mtx);
4381 l2arc_do_free_on_write();
4383 kmem_free(cb, sizeof (l2arc_write_callback_t));
4387 * A read to a cache device completed. Validate buffer contents before
4388 * handing over to the regular ARC routines.
4391 l2arc_read_done(zio_t *zio)
4393 l2arc_read_callback_t *cb;
4396 kmutex_t *hash_lock;
4399 ASSERT(zio->io_vd != NULL);
4400 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4402 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4404 cb = zio->io_private;
4406 buf = cb->l2rcb_buf;
4407 ASSERT(buf != NULL);
4409 hash_lock = HDR_LOCK(buf->b_hdr);
4410 mutex_enter(hash_lock);
4412 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4415 * If the buffer was compressed, decompress it first.
4417 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
4418 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
4419 ASSERT(zio->io_data != NULL);
4422 * Check this survived the L2ARC journey.
4424 equal = arc_cksum_equal(buf);
4425 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4426 mutex_exit(hash_lock);
4427 zio->io_private = buf;
4428 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4429 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4432 mutex_exit(hash_lock);
4434 * Buffer didn't survive caching. Increment stats and
4435 * reissue to the original storage device.
4437 if (zio->io_error != 0) {
4438 ARCSTAT_BUMP(arcstat_l2_io_error);
4440 zio->io_error = EIO;
4443 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4446 * If there's no waiter, issue an async i/o to the primary
4447 * storage now. If there *is* a waiter, the caller must
4448 * issue the i/o in a context where it's OK to block.
4450 if (zio->io_waiter == NULL) {
4451 zio_t *pio = zio_unique_parent(zio);
4453 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4455 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4456 buf->b_data, zio->io_size, arc_read_done, buf,
4457 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4461 kmem_free(cb, sizeof (l2arc_read_callback_t));
4465 * This is the list priority from which the L2ARC will search for pages to
4466 * cache. This is used within loops (0..3) to cycle through lists in the
4467 * desired order. This order can have a significant effect on cache
4470 * Currently the metadata lists are hit first, MFU then MRU, followed by
4471 * the data lists. This function returns a locked list, and also returns
4475 l2arc_list_locked(int list_num, kmutex_t **lock)
4477 list_t *list = NULL;
4479 ASSERT(list_num >= 0 && list_num <= 3);
4483 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4484 *lock = &arc_mfu->arcs_mtx;
4487 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4488 *lock = &arc_mru->arcs_mtx;
4491 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4492 *lock = &arc_mfu->arcs_mtx;
4495 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4496 *lock = &arc_mru->arcs_mtx;
4500 ASSERT(!(MUTEX_HELD(*lock)));
4506 * Evict buffers from the device write hand to the distance specified in
4507 * bytes. This distance may span populated buffers, it may span nothing.
4508 * This is clearing a region on the L2ARC device ready for writing.
4509 * If the 'all' boolean is set, every buffer is evicted.
4512 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4515 l2arc_buf_hdr_t *abl2;
4516 arc_buf_hdr_t *ab, *ab_prev;
4517 kmutex_t *hash_lock;
4520 buflist = dev->l2ad_buflist;
4522 if (buflist == NULL)
4525 if (!all && dev->l2ad_first) {
4527 * This is the first sweep through the device. There is
4533 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4535 * When nearing the end of the device, evict to the end
4536 * before the device write hand jumps to the start.
4538 taddr = dev->l2ad_end;
4540 taddr = dev->l2ad_hand + distance;
4542 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4543 uint64_t, taddr, boolean_t, all);
4546 mutex_enter(&l2arc_buflist_mtx);
4547 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4548 ab_prev = list_prev(buflist, ab);
4550 hash_lock = HDR_LOCK(ab);
4551 if (!mutex_tryenter(hash_lock)) {
4553 * Missed the hash lock. Retry.
4555 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4556 mutex_exit(&l2arc_buflist_mtx);
4557 mutex_enter(hash_lock);
4558 mutex_exit(hash_lock);
4562 if (HDR_L2_WRITE_HEAD(ab)) {
4564 * We hit a write head node. Leave it for
4565 * l2arc_write_done().
4567 list_remove(buflist, ab);
4568 mutex_exit(hash_lock);
4572 if (!all && ab->b_l2hdr != NULL &&
4573 (ab->b_l2hdr->b_daddr > taddr ||
4574 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4576 * We've evicted to the target address,
4577 * or the end of the device.
4579 mutex_exit(hash_lock);
4583 if (HDR_FREE_IN_PROGRESS(ab)) {
4585 * Already on the path to destruction.
4587 mutex_exit(hash_lock);
4591 if (ab->b_state == arc_l2c_only) {
4592 ASSERT(!HDR_L2_READING(ab));
4594 * This doesn't exist in the ARC. Destroy.
4595 * arc_hdr_destroy() will call list_remove()
4596 * and decrement arcstat_l2_size.
4598 arc_change_state(arc_anon, ab, hash_lock);
4599 arc_hdr_destroy(ab);
4602 * Invalidate issued or about to be issued
4603 * reads, since we may be about to write
4604 * over this location.
4606 if (HDR_L2_READING(ab)) {
4607 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4608 ab->b_flags |= ARC_L2_EVICTED;
4612 * Tell ARC this no longer exists in L2ARC.
4614 if (ab->b_l2hdr != NULL) {
4616 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4618 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4619 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4620 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4622 list_remove(buflist, ab);
4625 * This may have been leftover after a
4628 ab->b_flags &= ~ARC_L2_WRITING;
4630 mutex_exit(hash_lock);
4632 mutex_exit(&l2arc_buflist_mtx);
4634 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4635 dev->l2ad_evict = taddr;
4639 * Find and write ARC buffers to the L2ARC device.
4641 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4642 * for reading until they have completed writing.
4643 * The headroom_boost is an in-out parameter used to maintain headroom boost
4644 * state between calls to this function.
4646 * Returns the number of bytes actually written (which may be smaller than
4647 * the delta by which the device hand has changed due to alignment).
4650 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
4651 boolean_t *headroom_boost)
4653 arc_buf_hdr_t *ab, *ab_prev, *head;
4655 uint64_t write_asize, write_psize, write_sz, headroom,
4658 kmutex_t *list_lock = NULL;
4660 l2arc_write_callback_t *cb;
4662 uint64_t guid = spa_load_guid(spa);
4664 const boolean_t do_headroom_boost = *headroom_boost;
4666 ASSERT(dev->l2ad_vdev != NULL);
4668 /* Lower the flag now, we might want to raise it again later. */
4669 *headroom_boost = B_FALSE;
4672 write_sz = write_asize = write_psize = 0;
4674 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4675 head->b_flags |= ARC_L2_WRITE_HEAD;
4678 * We will want to try to compress buffers that are at least 2x the
4679 * device sector size.
4681 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
4684 * Copy buffers for L2ARC writing.
4686 mutex_enter(&l2arc_buflist_mtx);
4687 for (try = 0; try <= 3; try++) {
4688 uint64_t passed_sz = 0;
4690 list = l2arc_list_locked(try, &list_lock);
4693 * L2ARC fast warmup.
4695 * Until the ARC is warm and starts to evict, read from the
4696 * head of the ARC lists rather than the tail.
4698 if (arc_warm == B_FALSE)
4699 ab = list_head(list);
4701 ab = list_tail(list);
4703 headroom = target_sz * l2arc_headroom;
4704 if (do_headroom_boost)
4705 headroom = (headroom * l2arc_headroom_boost) / 100;
4707 for (; ab; ab = ab_prev) {
4708 l2arc_buf_hdr_t *l2hdr;
4709 kmutex_t *hash_lock;
4712 if (arc_warm == B_FALSE)
4713 ab_prev = list_next(list, ab);
4715 ab_prev = list_prev(list, ab);
4717 hash_lock = HDR_LOCK(ab);
4718 if (!mutex_tryenter(hash_lock)) {
4720 * Skip this buffer rather than waiting.
4725 passed_sz += ab->b_size;
4726 if (passed_sz > headroom) {
4730 mutex_exit(hash_lock);
4734 if (!l2arc_write_eligible(guid, ab)) {
4735 mutex_exit(hash_lock);
4739 if ((write_sz + ab->b_size) > target_sz) {
4741 mutex_exit(hash_lock);
4747 * Insert a dummy header on the buflist so
4748 * l2arc_write_done() can find where the
4749 * write buffers begin without searching.
4751 list_insert_head(dev->l2ad_buflist, head);
4753 cb = kmem_alloc(sizeof (l2arc_write_callback_t),
4755 cb->l2wcb_dev = dev;
4756 cb->l2wcb_head = head;
4757 pio = zio_root(spa, l2arc_write_done, cb,
4762 * Create and add a new L2ARC header.
4764 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t),
4767 arc_space_consume(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4769 ab->b_flags |= ARC_L2_WRITING;
4772 * Temporarily stash the data buffer in b_tmp_cdata.
4773 * The subsequent write step will pick it up from
4774 * there. This is because can't access ab->b_buf
4775 * without holding the hash_lock, which we in turn
4776 * can't access without holding the ARC list locks
4777 * (which we want to avoid during compression/writing)
4779 l2hdr->b_compress = ZIO_COMPRESS_OFF;
4780 l2hdr->b_asize = ab->b_size;
4781 l2hdr->b_tmp_cdata = ab->b_buf->b_data;
4783 buf_sz = ab->b_size;
4784 ab->b_l2hdr = l2hdr;
4786 list_insert_head(dev->l2ad_buflist, ab);
4789 * Compute and store the buffer cksum before
4790 * writing. On debug the cksum is verified first.
4792 arc_cksum_verify(ab->b_buf);
4793 arc_cksum_compute(ab->b_buf, B_TRUE);
4795 mutex_exit(hash_lock);
4800 mutex_exit(list_lock);
4806 /* No buffers selected for writing? */
4809 mutex_exit(&l2arc_buflist_mtx);
4810 kmem_cache_free(hdr_cache, head);
4815 * Now start writing the buffers. We're starting at the write head
4816 * and work backwards, retracing the course of the buffer selector
4819 for (ab = list_prev(dev->l2ad_buflist, head); ab;
4820 ab = list_prev(dev->l2ad_buflist, ab)) {
4821 l2arc_buf_hdr_t *l2hdr;
4825 * We shouldn't need to lock the buffer here, since we flagged
4826 * it as ARC_L2_WRITING in the previous step, but we must take
4827 * care to only access its L2 cache parameters. In particular,
4828 * ab->b_buf may be invalid by now due to ARC eviction.
4830 l2hdr = ab->b_l2hdr;
4831 l2hdr->b_daddr = dev->l2ad_hand;
4833 if (!l2arc_nocompress && (ab->b_flags & ARC_L2COMPRESS) &&
4834 l2hdr->b_asize >= buf_compress_minsz) {
4835 if (l2arc_compress_buf(l2hdr)) {
4837 * If compression succeeded, enable headroom
4838 * boost on the next scan cycle.
4840 *headroom_boost = B_TRUE;
4845 * Pick up the buffer data we had previously stashed away
4846 * (and now potentially also compressed).
4848 buf_data = l2hdr->b_tmp_cdata;
4849 buf_sz = l2hdr->b_asize;
4851 /* Compression may have squashed the buffer to zero length. */
4855 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4856 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4857 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4858 ZIO_FLAG_CANFAIL, B_FALSE);
4860 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4862 (void) zio_nowait(wzio);
4864 write_asize += buf_sz;
4866 * Keep the clock hand suitably device-aligned.
4868 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4869 write_psize += buf_p_sz;
4870 dev->l2ad_hand += buf_p_sz;
4874 mutex_exit(&l2arc_buflist_mtx);
4876 ASSERT3U(write_asize, <=, target_sz);
4877 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4878 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
4879 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4880 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
4881 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
4884 * Bump device hand to the device start if it is approaching the end.
4885 * l2arc_evict() will already have evicted ahead for this case.
4887 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4888 vdev_space_update(dev->l2ad_vdev,
4889 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4890 dev->l2ad_hand = dev->l2ad_start;
4891 dev->l2ad_evict = dev->l2ad_start;
4892 dev->l2ad_first = B_FALSE;
4895 dev->l2ad_writing = B_TRUE;
4896 (void) zio_wait(pio);
4897 dev->l2ad_writing = B_FALSE;
4899 return (write_asize);
4903 * Compresses an L2ARC buffer.
4904 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its
4905 * size in l2hdr->b_asize. This routine tries to compress the data and
4906 * depending on the compression result there are three possible outcomes:
4907 * *) The buffer was incompressible. The original l2hdr contents were left
4908 * untouched and are ready for writing to an L2 device.
4909 * *) The buffer was all-zeros, so there is no need to write it to an L2
4910 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
4911 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
4912 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
4913 * data buffer which holds the compressed data to be written, and b_asize
4914 * tells us how much data there is. b_compress is set to the appropriate
4915 * compression algorithm. Once writing is done, invoke
4916 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
4918 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
4919 * buffer was incompressible).
4922 l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr)
4927 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
4928 ASSERT(l2hdr->b_tmp_cdata != NULL);
4930 len = l2hdr->b_asize;
4931 cdata = zio_data_buf_alloc(len);
4932 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata,
4933 cdata, l2hdr->b_asize);
4936 /* zero block, indicate that there's nothing to write */
4937 zio_data_buf_free(cdata, len);
4938 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
4940 l2hdr->b_tmp_cdata = NULL;
4941 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
4943 } else if (csize > 0 && csize < len) {
4945 * Compression succeeded, we'll keep the cdata around for
4946 * writing and release it afterwards.
4948 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
4949 l2hdr->b_asize = csize;
4950 l2hdr->b_tmp_cdata = cdata;
4951 ARCSTAT_BUMP(arcstat_l2_compress_successes);
4955 * Compression failed, release the compressed buffer.
4956 * l2hdr will be left unmodified.
4958 zio_data_buf_free(cdata, len);
4959 ARCSTAT_BUMP(arcstat_l2_compress_failures);
4965 * Decompresses a zio read back from an l2arc device. On success, the
4966 * underlying zio's io_data buffer is overwritten by the uncompressed
4967 * version. On decompression error (corrupt compressed stream), the
4968 * zio->io_error value is set to signal an I/O error.
4970 * Please note that the compressed data stream is not checksummed, so
4971 * if the underlying device is experiencing data corruption, we may feed
4972 * corrupt data to the decompressor, so the decompressor needs to be
4973 * able to handle this situation (LZ4 does).
4976 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
4981 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
4983 if (zio->io_error != 0) {
4985 * An io error has occured, just restore the original io
4986 * size in preparation for a main pool read.
4988 zio->io_orig_size = zio->io_size = hdr->b_size;
4992 if (c == ZIO_COMPRESS_EMPTY) {
4994 * An empty buffer results in a null zio, which means we
4995 * need to fill its io_data after we're done restoring the
4996 * buffer's contents.
4998 ASSERT(hdr->b_buf != NULL);
4999 bzero(hdr->b_buf->b_data, hdr->b_size);
5000 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data;
5002 ASSERT(zio->io_data != NULL);
5004 * We copy the compressed data from the start of the arc buffer
5005 * (the zio_read will have pulled in only what we need, the
5006 * rest is garbage which we will overwrite at decompression)
5007 * and then decompress back to the ARC data buffer. This way we
5008 * can minimize copying by simply decompressing back over the
5009 * original compressed data (rather than decompressing to an
5010 * aux buffer and then copying back the uncompressed buffer,
5011 * which is likely to be much larger).
5013 csize = zio->io_size;
5014 cdata = zio_data_buf_alloc(csize);
5015 bcopy(zio->io_data, cdata, csize);
5016 if (zio_decompress_data(c, cdata, zio->io_data, csize,
5018 zio->io_error = EIO;
5019 zio_data_buf_free(cdata, csize);
5022 /* Restore the expected uncompressed IO size. */
5023 zio->io_orig_size = zio->io_size = hdr->b_size;
5027 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
5028 * This buffer serves as a temporary holder of compressed data while
5029 * the buffer entry is being written to an l2arc device. Once that is
5030 * done, we can dispose of it.
5033 l2arc_release_cdata_buf(arc_buf_hdr_t *ab)
5035 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr;
5037 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) {
5039 * If the data was compressed, then we've allocated a
5040 * temporary buffer for it, so now we need to release it.
5042 ASSERT(l2hdr->b_tmp_cdata != NULL);
5043 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size);
5045 l2hdr->b_tmp_cdata = NULL;
5049 * This thread feeds the L2ARC at regular intervals. This is the beating
5050 * heart of the L2ARC.
5053 l2arc_feed_thread(void)
5058 uint64_t size, wrote;
5059 clock_t begin, next = ddi_get_lbolt();
5060 boolean_t headroom_boost = B_FALSE;
5062 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
5064 mutex_enter(&l2arc_feed_thr_lock);
5066 while (l2arc_thread_exit == 0) {
5067 CALLB_CPR_SAFE_BEGIN(&cpr);
5068 (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv,
5069 &l2arc_feed_thr_lock, next);
5070 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
5071 next = ddi_get_lbolt() + hz;
5074 * Quick check for L2ARC devices.
5076 mutex_enter(&l2arc_dev_mtx);
5077 if (l2arc_ndev == 0) {
5078 mutex_exit(&l2arc_dev_mtx);
5081 mutex_exit(&l2arc_dev_mtx);
5082 begin = ddi_get_lbolt();
5085 * This selects the next l2arc device to write to, and in
5086 * doing so the next spa to feed from: dev->l2ad_spa. This
5087 * will return NULL if there are now no l2arc devices or if
5088 * they are all faulted.
5090 * If a device is returned, its spa's config lock is also
5091 * held to prevent device removal. l2arc_dev_get_next()
5092 * will grab and release l2arc_dev_mtx.
5094 if ((dev = l2arc_dev_get_next()) == NULL)
5097 spa = dev->l2ad_spa;
5098 ASSERT(spa != NULL);
5101 * If the pool is read-only then force the feed thread to
5102 * sleep a little longer.
5104 if (!spa_writeable(spa)) {
5105 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
5106 spa_config_exit(spa, SCL_L2ARC, dev);
5111 * Avoid contributing to memory pressure.
5114 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
5115 spa_config_exit(spa, SCL_L2ARC, dev);
5119 ARCSTAT_BUMP(arcstat_l2_feeds);
5121 size = l2arc_write_size();
5124 * Evict L2ARC buffers that will be overwritten.
5126 l2arc_evict(dev, size, B_FALSE);
5129 * Write ARC buffers.
5131 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
5134 * Calculate interval between writes.
5136 next = l2arc_write_interval(begin, size, wrote);
5137 spa_config_exit(spa, SCL_L2ARC, dev);
5140 l2arc_thread_exit = 0;
5141 cv_broadcast(&l2arc_feed_thr_cv);
5142 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
5147 l2arc_vdev_present(vdev_t *vd)
5151 mutex_enter(&l2arc_dev_mtx);
5152 for (dev = list_head(l2arc_dev_list); dev != NULL;
5153 dev = list_next(l2arc_dev_list, dev)) {
5154 if (dev->l2ad_vdev == vd)
5157 mutex_exit(&l2arc_dev_mtx);
5159 return (dev != NULL);
5163 * Add a vdev for use by the L2ARC. By this point the spa has already
5164 * validated the vdev and opened it.
5167 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5169 l2arc_dev_t *adddev;
5171 ASSERT(!l2arc_vdev_present(vd));
5174 * Create a new l2arc device entry.
5176 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5177 adddev->l2ad_spa = spa;
5178 adddev->l2ad_vdev = vd;
5179 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5180 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5181 adddev->l2ad_hand = adddev->l2ad_start;
5182 adddev->l2ad_evict = adddev->l2ad_start;
5183 adddev->l2ad_first = B_TRUE;
5184 adddev->l2ad_writing = B_FALSE;
5185 list_link_init(&adddev->l2ad_node);
5188 * This is a list of all ARC buffers that are still valid on the
5191 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5192 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5193 offsetof(arc_buf_hdr_t, b_l2node));
5195 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5198 * Add device to global list
5200 mutex_enter(&l2arc_dev_mtx);
5201 list_insert_head(l2arc_dev_list, adddev);
5202 atomic_inc_64(&l2arc_ndev);
5203 mutex_exit(&l2arc_dev_mtx);
5207 * Remove a vdev from the L2ARC.
5210 l2arc_remove_vdev(vdev_t *vd)
5212 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5215 * Find the device by vdev
5217 mutex_enter(&l2arc_dev_mtx);
5218 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5219 nextdev = list_next(l2arc_dev_list, dev);
5220 if (vd == dev->l2ad_vdev) {
5225 ASSERT(remdev != NULL);
5228 * Remove device from global list
5230 list_remove(l2arc_dev_list, remdev);
5231 l2arc_dev_last = NULL; /* may have been invalidated */
5232 atomic_dec_64(&l2arc_ndev);
5233 mutex_exit(&l2arc_dev_mtx);
5236 * Clear all buflists and ARC references. L2ARC device flush.
5238 l2arc_evict(remdev, 0, B_TRUE);
5239 list_destroy(remdev->l2ad_buflist);
5240 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5241 kmem_free(remdev, sizeof (l2arc_dev_t));
5247 l2arc_thread_exit = 0;
5249 l2arc_writes_sent = 0;
5250 l2arc_writes_done = 0;
5252 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5253 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5254 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5255 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5256 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5258 l2arc_dev_list = &L2ARC_dev_list;
5259 l2arc_free_on_write = &L2ARC_free_on_write;
5260 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5261 offsetof(l2arc_dev_t, l2ad_node));
5262 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5263 offsetof(l2arc_data_free_t, l2df_list_node));
5270 * This is called from dmu_fini(), which is called from spa_fini();
5271 * Because of this, we can assume that all l2arc devices have
5272 * already been removed when the pools themselves were removed.
5275 l2arc_do_free_on_write();
5277 mutex_destroy(&l2arc_feed_thr_lock);
5278 cv_destroy(&l2arc_feed_thr_cv);
5279 mutex_destroy(&l2arc_dev_mtx);
5280 mutex_destroy(&l2arc_buflist_mtx);
5281 mutex_destroy(&l2arc_free_on_write_mtx);
5283 list_destroy(l2arc_dev_list);
5284 list_destroy(l2arc_free_on_write);
5290 if (!(spa_mode_global & FWRITE))
5293 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5294 TS_RUN, minclsyspri);
5300 if (!(spa_mode_global & FWRITE))
5303 mutex_enter(&l2arc_feed_thr_lock);
5304 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5305 l2arc_thread_exit = 1;
5306 while (l2arc_thread_exit != 0)
5307 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5308 mutex_exit(&l2arc_feed_thr_lock);
5311 #if defined(_KERNEL) && defined(HAVE_SPL)
5312 EXPORT_SYMBOL(arc_read);
5313 EXPORT_SYMBOL(arc_buf_remove_ref);
5314 EXPORT_SYMBOL(arc_getbuf_func);
5315 EXPORT_SYMBOL(arc_add_prune_callback);
5316 EXPORT_SYMBOL(arc_remove_prune_callback);
5318 module_param(zfs_arc_min, ulong, 0644);
5319 MODULE_PARM_DESC(zfs_arc_min, "Min arc size");
5321 module_param(zfs_arc_max, ulong, 0644);
5322 MODULE_PARM_DESC(zfs_arc_max, "Max arc size");
5324 module_param(zfs_arc_meta_limit, ulong, 0644);
5325 MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size");
5327 module_param(zfs_arc_meta_prune, int, 0644);
5328 MODULE_PARM_DESC(zfs_arc_meta_prune, "Bytes of meta data to prune");
5330 module_param(zfs_arc_grow_retry, int, 0644);
5331 MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size");
5333 module_param(zfs_arc_shrink_shift, int, 0644);
5334 MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)");
5336 module_param(zfs_arc_p_min_shift, int, 0644);
5337 MODULE_PARM_DESC(zfs_arc_p_min_shift, "arc_c shift to calc min/max arc_p");
5339 module_param(zfs_disable_dup_eviction, int, 0644);
5340 MODULE_PARM_DESC(zfs_disable_dup_eviction, "disable duplicate buffer eviction");
5342 module_param(zfs_arc_memory_throttle_disable, int, 0644);
5343 MODULE_PARM_DESC(zfs_arc_memory_throttle_disable, "disable memory throttle");
5345 module_param(zfs_arc_min_prefetch_lifespan, int, 0644);
5346 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan, "Min life of prefetch block");
5348 module_param(l2arc_write_max, ulong, 0644);
5349 MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval");
5351 module_param(l2arc_write_boost, ulong, 0644);
5352 MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup");
5354 module_param(l2arc_headroom, ulong, 0644);
5355 MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache");
5357 module_param(l2arc_headroom_boost, ulong, 0644);
5358 MODULE_PARM_DESC(l2arc_headroom_boost, "Compressed l2arc_headroom multiplier");
5360 module_param(l2arc_feed_secs, ulong, 0644);
5361 MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing");
5363 module_param(l2arc_feed_min_ms, ulong, 0644);
5364 MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds");
5366 module_param(l2arc_noprefetch, int, 0644);
5367 MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers");
5369 module_param(l2arc_nocompress, int, 0644);
5370 MODULE_PARM_DESC(l2arc_nocompress, "Skip compressing L2ARC buffers");
5372 module_param(l2arc_feed_again, int, 0644);
5373 MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup");
5375 module_param(l2arc_norw, int, 0644);
5376 MODULE_PARM_DESC(l2arc_norw, "No reads during writes");