4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2011 by Delphix. All rights reserved.
28 * DVA-based Adjustable Replacement Cache
30 * While much of the theory of operation used here is
31 * based on the self-tuning, low overhead replacement cache
32 * presented by Megiddo and Modha at FAST 2003, there are some
33 * significant differences:
35 * 1. The Megiddo and Modha model assumes any page is evictable.
36 * Pages in its cache cannot be "locked" into memory. This makes
37 * the eviction algorithm simple: evict the last page in the list.
38 * This also make the performance characteristics easy to reason
39 * about. Our cache is not so simple. At any given moment, some
40 * subset of the blocks in the cache are un-evictable because we
41 * have handed out a reference to them. Blocks are only evictable
42 * when there are no external references active. This makes
43 * eviction far more problematic: we choose to evict the evictable
44 * blocks that are the "lowest" in the list.
46 * There are times when it is not possible to evict the requested
47 * space. In these circumstances we are unable to adjust the cache
48 * size. To prevent the cache growing unbounded at these times we
49 * implement a "cache throttle" that slows the flow of new data
50 * into the cache until we can make space available.
52 * 2. The Megiddo and Modha model assumes a fixed cache size.
53 * Pages are evicted when the cache is full and there is a cache
54 * miss. Our model has a variable sized cache. It grows with
55 * high use, but also tries to react to memory pressure from the
56 * operating system: decreasing its size when system memory is
59 * 3. The Megiddo and Modha model assumes a fixed page size. All
60 * elements of the cache are therefor exactly the same size. So
61 * when adjusting the cache size following a cache miss, its simply
62 * a matter of choosing a single page to evict. In our model, we
63 * have variable sized cache blocks (rangeing from 512 bytes to
64 * 128K bytes). We therefor choose a set of blocks to evict to make
65 * space for a cache miss that approximates as closely as possible
66 * the space used by the new block.
68 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
69 * by N. Megiddo & D. Modha, FAST 2003
75 * A new reference to a cache buffer can be obtained in two
76 * ways: 1) via a hash table lookup using the DVA as a key,
77 * or 2) via one of the ARC lists. The arc_read() interface
78 * uses method 1, while the internal arc algorithms for
79 * adjusting the cache use method 2. We therefor provide two
80 * types of locks: 1) the hash table lock array, and 2) the
83 * Buffers do not have their own mutexes, rather they rely on the
84 * hash table mutexes for the bulk of their protection (i.e. most
85 * fields in the arc_buf_hdr_t are protected by these mutexes).
87 * buf_hash_find() returns the appropriate mutex (held) when it
88 * locates the requested buffer in the hash table. It returns
89 * NULL for the mutex if the buffer was not in the table.
91 * buf_hash_remove() expects the appropriate hash mutex to be
92 * already held before it is invoked.
94 * Each arc state also has a mutex which is used to protect the
95 * buffer list associated with the state. When attempting to
96 * obtain a hash table lock while holding an arc list lock you
97 * must use: mutex_tryenter() to avoid deadlock. Also note that
98 * the active state mutex must be held before the ghost state mutex.
100 * Arc buffers may have an associated eviction callback function.
101 * This function will be invoked prior to removing the buffer (e.g.
102 * in arc_do_user_evicts()). Note however that the data associated
103 * with the buffer may be evicted prior to the callback. The callback
104 * must be made with *no locks held* (to prevent deadlock). Additionally,
105 * the users of callbacks must ensure that their private data is
106 * protected from simultaneous callbacks from arc_buf_evict()
107 * and arc_do_user_evicts().
109 * It as also possible to register a callback which is run when the
110 * arc_meta_limit is reached and no buffers can be safely evicted. In
111 * this case the arc user should drop a reference on some arc buffers so
112 * they can be reclaimed and the arc_meta_limit honored. For example,
113 * when using the ZPL each dentry holds a references on a znode. These
114 * dentries must be pruned before the arc buffer holding the znode can
117 * Note that the majority of the performance stats are manipulated
118 * with atomic operations.
120 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
122 * - L2ARC buflist creation
123 * - L2ARC buflist eviction
124 * - L2ARC write completion, which walks L2ARC buflists
125 * - ARC header destruction, as it removes from L2ARC buflists
126 * - ARC header release, as it removes from L2ARC buflists
131 #include <sys/zfs_context.h>
133 #include <sys/vdev.h>
134 #include <sys/vdev_impl.h>
136 #include <sys/vmsystm.h>
138 #include <sys/fs/swapnode.h>
141 #include <sys/callb.h>
142 #include <sys/kstat.h>
143 #include <sys/dmu_tx.h>
144 #include <zfs_fletcher.h>
146 static kmutex_t arc_reclaim_thr_lock;
147 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
148 static uint8_t arc_thread_exit;
150 /* number of bytes to prune from caches when at arc_meta_limit is reached */
151 int zfs_arc_meta_prune = 1048576;
153 typedef enum arc_reclaim_strategy {
154 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
155 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
156 } arc_reclaim_strategy_t;
158 /* number of seconds before growing cache again */
159 int zfs_arc_grow_retry = 5;
161 /* shift of arc_c for calculating both min and max arc_p */
162 int zfs_arc_p_min_shift = 4;
164 /* log2(fraction of arc to reclaim) */
165 int zfs_arc_shrink_shift = 5;
168 * minimum lifespan of a prefetch block in clock ticks
169 * (initialized in arc_init())
171 int zfs_arc_min_prefetch_lifespan = HZ;
173 /* disable arc proactive arc throttle due to low memory */
174 int zfs_arc_memory_throttle_disable = 1;
176 /* disable duplicate buffer eviction */
177 int zfs_disable_dup_eviction = 0;
181 /* expiration time for arc_no_grow */
182 static clock_t arc_grow_time = 0;
185 * The arc has filled available memory and has now warmed up.
187 static boolean_t arc_warm;
190 * These tunables are for performance analysis.
192 unsigned long zfs_arc_max = 0;
193 unsigned long zfs_arc_min = 0;
194 unsigned long zfs_arc_meta_limit = 0;
197 * Note that buffers can be in one of 6 states:
198 * ARC_anon - anonymous (discussed below)
199 * ARC_mru - recently used, currently cached
200 * ARC_mru_ghost - recentely used, no longer in cache
201 * ARC_mfu - frequently used, currently cached
202 * ARC_mfu_ghost - frequently used, no longer in cache
203 * ARC_l2c_only - exists in L2ARC but not other states
204 * When there are no active references to the buffer, they are
205 * are linked onto a list in one of these arc states. These are
206 * the only buffers that can be evicted or deleted. Within each
207 * state there are multiple lists, one for meta-data and one for
208 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
209 * etc.) is tracked separately so that it can be managed more
210 * explicitly: favored over data, limited explicitly.
212 * Anonymous buffers are buffers that are not associated with
213 * a DVA. These are buffers that hold dirty block copies
214 * before they are written to stable storage. By definition,
215 * they are "ref'd" and are considered part of arc_mru
216 * that cannot be freed. Generally, they will aquire a DVA
217 * as they are written and migrate onto the arc_mru list.
219 * The ARC_l2c_only state is for buffers that are in the second
220 * level ARC but no longer in any of the ARC_m* lists. The second
221 * level ARC itself may also contain buffers that are in any of
222 * the ARC_m* states - meaning that a buffer can exist in two
223 * places. The reason for the ARC_l2c_only state is to keep the
224 * buffer header in the hash table, so that reads that hit the
225 * second level ARC benefit from these fast lookups.
228 typedef struct arc_state {
229 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
230 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
231 uint64_t arcs_size; /* total amount of data in this state */
236 static arc_state_t ARC_anon;
237 static arc_state_t ARC_mru;
238 static arc_state_t ARC_mru_ghost;
239 static arc_state_t ARC_mfu;
240 static arc_state_t ARC_mfu_ghost;
241 static arc_state_t ARC_l2c_only;
243 typedef struct arc_stats {
244 kstat_named_t arcstat_hits;
245 kstat_named_t arcstat_misses;
246 kstat_named_t arcstat_demand_data_hits;
247 kstat_named_t arcstat_demand_data_misses;
248 kstat_named_t arcstat_demand_metadata_hits;
249 kstat_named_t arcstat_demand_metadata_misses;
250 kstat_named_t arcstat_prefetch_data_hits;
251 kstat_named_t arcstat_prefetch_data_misses;
252 kstat_named_t arcstat_prefetch_metadata_hits;
253 kstat_named_t arcstat_prefetch_metadata_misses;
254 kstat_named_t arcstat_mru_hits;
255 kstat_named_t arcstat_mru_ghost_hits;
256 kstat_named_t arcstat_mfu_hits;
257 kstat_named_t arcstat_mfu_ghost_hits;
258 kstat_named_t arcstat_deleted;
259 kstat_named_t arcstat_recycle_miss;
260 kstat_named_t arcstat_mutex_miss;
261 kstat_named_t arcstat_evict_skip;
262 kstat_named_t arcstat_evict_l2_cached;
263 kstat_named_t arcstat_evict_l2_eligible;
264 kstat_named_t arcstat_evict_l2_ineligible;
265 kstat_named_t arcstat_hash_elements;
266 kstat_named_t arcstat_hash_elements_max;
267 kstat_named_t arcstat_hash_collisions;
268 kstat_named_t arcstat_hash_chains;
269 kstat_named_t arcstat_hash_chain_max;
270 kstat_named_t arcstat_p;
271 kstat_named_t arcstat_c;
272 kstat_named_t arcstat_c_min;
273 kstat_named_t arcstat_c_max;
274 kstat_named_t arcstat_size;
275 kstat_named_t arcstat_hdr_size;
276 kstat_named_t arcstat_data_size;
277 kstat_named_t arcstat_other_size;
278 kstat_named_t arcstat_anon_size;
279 kstat_named_t arcstat_anon_evict_data;
280 kstat_named_t arcstat_anon_evict_metadata;
281 kstat_named_t arcstat_mru_size;
282 kstat_named_t arcstat_mru_evict_data;
283 kstat_named_t arcstat_mru_evict_metadata;
284 kstat_named_t arcstat_mru_ghost_size;
285 kstat_named_t arcstat_mru_ghost_evict_data;
286 kstat_named_t arcstat_mru_ghost_evict_metadata;
287 kstat_named_t arcstat_mfu_size;
288 kstat_named_t arcstat_mfu_evict_data;
289 kstat_named_t arcstat_mfu_evict_metadata;
290 kstat_named_t arcstat_mfu_ghost_size;
291 kstat_named_t arcstat_mfu_ghost_evict_data;
292 kstat_named_t arcstat_mfu_ghost_evict_metadata;
293 kstat_named_t arcstat_l2_hits;
294 kstat_named_t arcstat_l2_misses;
295 kstat_named_t arcstat_l2_feeds;
296 kstat_named_t arcstat_l2_rw_clash;
297 kstat_named_t arcstat_l2_read_bytes;
298 kstat_named_t arcstat_l2_write_bytes;
299 kstat_named_t arcstat_l2_writes_sent;
300 kstat_named_t arcstat_l2_writes_done;
301 kstat_named_t arcstat_l2_writes_error;
302 kstat_named_t arcstat_l2_writes_hdr_miss;
303 kstat_named_t arcstat_l2_evict_lock_retry;
304 kstat_named_t arcstat_l2_evict_reading;
305 kstat_named_t arcstat_l2_free_on_write;
306 kstat_named_t arcstat_l2_abort_lowmem;
307 kstat_named_t arcstat_l2_cksum_bad;
308 kstat_named_t arcstat_l2_io_error;
309 kstat_named_t arcstat_l2_size;
310 kstat_named_t arcstat_l2_hdr_size;
311 kstat_named_t arcstat_memory_throttle_count;
312 kstat_named_t arcstat_duplicate_buffers;
313 kstat_named_t arcstat_duplicate_buffers_size;
314 kstat_named_t arcstat_duplicate_reads;
315 kstat_named_t arcstat_memory_direct_count;
316 kstat_named_t arcstat_memory_indirect_count;
317 kstat_named_t arcstat_no_grow;
318 kstat_named_t arcstat_tempreserve;
319 kstat_named_t arcstat_loaned_bytes;
320 kstat_named_t arcstat_prune;
321 kstat_named_t arcstat_meta_used;
322 kstat_named_t arcstat_meta_limit;
323 kstat_named_t arcstat_meta_max;
326 static arc_stats_t arc_stats = {
327 { "hits", KSTAT_DATA_UINT64 },
328 { "misses", KSTAT_DATA_UINT64 },
329 { "demand_data_hits", KSTAT_DATA_UINT64 },
330 { "demand_data_misses", KSTAT_DATA_UINT64 },
331 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
332 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
333 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
334 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
335 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
336 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
337 { "mru_hits", KSTAT_DATA_UINT64 },
338 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
339 { "mfu_hits", KSTAT_DATA_UINT64 },
340 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
341 { "deleted", KSTAT_DATA_UINT64 },
342 { "recycle_miss", KSTAT_DATA_UINT64 },
343 { "mutex_miss", KSTAT_DATA_UINT64 },
344 { "evict_skip", KSTAT_DATA_UINT64 },
345 { "evict_l2_cached", KSTAT_DATA_UINT64 },
346 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
347 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
348 { "hash_elements", KSTAT_DATA_UINT64 },
349 { "hash_elements_max", KSTAT_DATA_UINT64 },
350 { "hash_collisions", KSTAT_DATA_UINT64 },
351 { "hash_chains", KSTAT_DATA_UINT64 },
352 { "hash_chain_max", KSTAT_DATA_UINT64 },
353 { "p", KSTAT_DATA_UINT64 },
354 { "c", KSTAT_DATA_UINT64 },
355 { "c_min", KSTAT_DATA_UINT64 },
356 { "c_max", KSTAT_DATA_UINT64 },
357 { "size", KSTAT_DATA_UINT64 },
358 { "hdr_size", KSTAT_DATA_UINT64 },
359 { "data_size", KSTAT_DATA_UINT64 },
360 { "other_size", KSTAT_DATA_UINT64 },
361 { "anon_size", KSTAT_DATA_UINT64 },
362 { "anon_evict_data", KSTAT_DATA_UINT64 },
363 { "anon_evict_metadata", KSTAT_DATA_UINT64 },
364 { "mru_size", KSTAT_DATA_UINT64 },
365 { "mru_evict_data", KSTAT_DATA_UINT64 },
366 { "mru_evict_metadata", KSTAT_DATA_UINT64 },
367 { "mru_ghost_size", KSTAT_DATA_UINT64 },
368 { "mru_ghost_evict_data", KSTAT_DATA_UINT64 },
369 { "mru_ghost_evict_metadata", KSTAT_DATA_UINT64 },
370 { "mfu_size", KSTAT_DATA_UINT64 },
371 { "mfu_evict_data", KSTAT_DATA_UINT64 },
372 { "mfu_evict_metadata", KSTAT_DATA_UINT64 },
373 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
374 { "mfu_ghost_evict_data", KSTAT_DATA_UINT64 },
375 { "mfu_ghost_evict_metadata", KSTAT_DATA_UINT64 },
376 { "l2_hits", KSTAT_DATA_UINT64 },
377 { "l2_misses", KSTAT_DATA_UINT64 },
378 { "l2_feeds", KSTAT_DATA_UINT64 },
379 { "l2_rw_clash", KSTAT_DATA_UINT64 },
380 { "l2_read_bytes", KSTAT_DATA_UINT64 },
381 { "l2_write_bytes", KSTAT_DATA_UINT64 },
382 { "l2_writes_sent", KSTAT_DATA_UINT64 },
383 { "l2_writes_done", KSTAT_DATA_UINT64 },
384 { "l2_writes_error", KSTAT_DATA_UINT64 },
385 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
386 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
387 { "l2_evict_reading", KSTAT_DATA_UINT64 },
388 { "l2_free_on_write", KSTAT_DATA_UINT64 },
389 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
390 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
391 { "l2_io_error", KSTAT_DATA_UINT64 },
392 { "l2_size", KSTAT_DATA_UINT64 },
393 { "l2_hdr_size", KSTAT_DATA_UINT64 },
394 { "memory_throttle_count", KSTAT_DATA_UINT64 },
395 { "duplicate_buffers", KSTAT_DATA_UINT64 },
396 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
397 { "duplicate_reads", KSTAT_DATA_UINT64 },
398 { "memory_direct_count", KSTAT_DATA_UINT64 },
399 { "memory_indirect_count", KSTAT_DATA_UINT64 },
400 { "arc_no_grow", KSTAT_DATA_UINT64 },
401 { "arc_tempreserve", KSTAT_DATA_UINT64 },
402 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
403 { "arc_prune", KSTAT_DATA_UINT64 },
404 { "arc_meta_used", KSTAT_DATA_UINT64 },
405 { "arc_meta_limit", KSTAT_DATA_UINT64 },
406 { "arc_meta_max", KSTAT_DATA_UINT64 },
409 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
411 #define ARCSTAT_INCR(stat, val) \
412 atomic_add_64(&arc_stats.stat.value.ui64, (val));
414 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
415 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
417 #define ARCSTAT_MAX(stat, val) { \
419 while ((val) > (m = arc_stats.stat.value.ui64) && \
420 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
424 #define ARCSTAT_MAXSTAT(stat) \
425 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
428 * We define a macro to allow ARC hits/misses to be easily broken down by
429 * two separate conditions, giving a total of four different subtypes for
430 * each of hits and misses (so eight statistics total).
432 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
435 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
437 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
441 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
443 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
448 static arc_state_t *arc_anon;
449 static arc_state_t *arc_mru;
450 static arc_state_t *arc_mru_ghost;
451 static arc_state_t *arc_mfu;
452 static arc_state_t *arc_mfu_ghost;
453 static arc_state_t *arc_l2c_only;
456 * There are several ARC variables that are critical to export as kstats --
457 * but we don't want to have to grovel around in the kstat whenever we wish to
458 * manipulate them. For these variables, we therefore define them to be in
459 * terms of the statistic variable. This assures that we are not introducing
460 * the possibility of inconsistency by having shadow copies of the variables,
461 * while still allowing the code to be readable.
463 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
464 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
465 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
466 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
467 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
468 #define arc_no_grow ARCSTAT(arcstat_no_grow)
469 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
470 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
471 #define arc_meta_used ARCSTAT(arcstat_meta_used)
472 #define arc_meta_limit ARCSTAT(arcstat_meta_limit)
473 #define arc_meta_max ARCSTAT(arcstat_meta_max)
475 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
477 typedef struct arc_callback arc_callback_t;
479 struct arc_callback {
481 arc_done_func_t *acb_done;
483 zio_t *acb_zio_dummy;
484 arc_callback_t *acb_next;
487 typedef struct arc_write_callback arc_write_callback_t;
489 struct arc_write_callback {
491 arc_done_func_t *awcb_ready;
492 arc_done_func_t *awcb_done;
497 /* protected by hash lock */
502 kmutex_t b_freeze_lock;
503 zio_cksum_t *b_freeze_cksum;
506 arc_buf_hdr_t *b_hash_next;
511 arc_callback_t *b_acb;
515 arc_buf_contents_t b_type;
519 /* protected by arc state mutex */
520 arc_state_t *b_state;
521 list_node_t b_arc_node;
523 /* updated atomically */
524 clock_t b_arc_access;
526 /* self protecting */
529 l2arc_buf_hdr_t *b_l2hdr;
530 list_node_t b_l2node;
533 static list_t arc_prune_list;
534 static kmutex_t arc_prune_mtx;
535 static arc_buf_t *arc_eviction_list;
536 static kmutex_t arc_eviction_mtx;
537 static arc_buf_hdr_t arc_eviction_hdr;
538 static void arc_get_data_buf(arc_buf_t *buf);
539 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
540 static int arc_evict_needed(arc_buf_contents_t type);
541 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
543 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
545 #define GHOST_STATE(state) \
546 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
547 (state) == arc_l2c_only)
550 * Private ARC flags. These flags are private ARC only flags that will show up
551 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
552 * be passed in as arc_flags in things like arc_read. However, these flags
553 * should never be passed and should only be set by ARC code. When adding new
554 * public flags, make sure not to smash the private ones.
557 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
558 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
559 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
560 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
561 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
562 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
563 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
564 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
565 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
566 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
568 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
569 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
570 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
571 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
572 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
573 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
574 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
575 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
576 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
577 (hdr)->b_l2hdr != NULL)
578 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
579 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
580 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
586 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
587 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
590 * Hash table routines
593 #define HT_LOCK_ALIGN 64
594 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
599 unsigned char pad[HT_LOCK_PAD];
603 #define BUF_LOCKS 256
604 typedef struct buf_hash_table {
606 arc_buf_hdr_t **ht_table;
607 struct ht_lock ht_locks[BUF_LOCKS];
610 static buf_hash_table_t buf_hash_table;
612 #define BUF_HASH_INDEX(spa, dva, birth) \
613 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
614 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
615 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
616 #define HDR_LOCK(hdr) \
617 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
619 uint64_t zfs_crc64_table[256];
625 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
626 #define L2ARC_HEADROOM 2 /* num of writes */
627 #define L2ARC_FEED_SECS 1 /* caching interval secs */
628 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
630 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
631 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
634 * L2ARC Performance Tunables
636 unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
637 unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
638 unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
639 unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
640 unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
641 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
642 int l2arc_feed_again = B_TRUE; /* turbo warmup */
643 int l2arc_norw = B_FALSE; /* no reads during writes */
648 typedef struct l2arc_dev {
649 vdev_t *l2ad_vdev; /* vdev */
650 spa_t *l2ad_spa; /* spa */
651 uint64_t l2ad_hand; /* next write location */
652 uint64_t l2ad_write; /* desired write size, bytes */
653 uint64_t l2ad_boost; /* warmup write boost, bytes */
654 uint64_t l2ad_start; /* first addr on device */
655 uint64_t l2ad_end; /* last addr on device */
656 uint64_t l2ad_evict; /* last addr eviction reached */
657 boolean_t l2ad_first; /* first sweep through */
658 boolean_t l2ad_writing; /* currently writing */
659 list_t *l2ad_buflist; /* buffer list */
660 list_node_t l2ad_node; /* device list node */
663 static list_t L2ARC_dev_list; /* device list */
664 static list_t *l2arc_dev_list; /* device list pointer */
665 static kmutex_t l2arc_dev_mtx; /* device list mutex */
666 static l2arc_dev_t *l2arc_dev_last; /* last device used */
667 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
668 static list_t L2ARC_free_on_write; /* free after write buf list */
669 static list_t *l2arc_free_on_write; /* free after write list ptr */
670 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
671 static uint64_t l2arc_ndev; /* number of devices */
673 typedef struct l2arc_read_callback {
674 arc_buf_t *l2rcb_buf; /* read buffer */
675 spa_t *l2rcb_spa; /* spa */
676 blkptr_t l2rcb_bp; /* original blkptr */
677 zbookmark_t l2rcb_zb; /* original bookmark */
678 int l2rcb_flags; /* original flags */
679 } l2arc_read_callback_t;
681 typedef struct l2arc_write_callback {
682 l2arc_dev_t *l2wcb_dev; /* device info */
683 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
684 } l2arc_write_callback_t;
686 struct l2arc_buf_hdr {
687 /* protected by arc_buf_hdr mutex */
688 l2arc_dev_t *b_dev; /* L2ARC device */
689 uint64_t b_daddr; /* disk address, offset byte */
692 typedef struct l2arc_data_free {
693 /* protected by l2arc_free_on_write_mtx */
696 void (*l2df_func)(void *, size_t);
697 list_node_t l2df_list_node;
700 static kmutex_t l2arc_feed_thr_lock;
701 static kcondvar_t l2arc_feed_thr_cv;
702 static uint8_t l2arc_thread_exit;
704 static void l2arc_read_done(zio_t *zio);
705 static void l2arc_hdr_stat_add(void);
706 static void l2arc_hdr_stat_remove(void);
709 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
711 uint8_t *vdva = (uint8_t *)dva;
712 uint64_t crc = -1ULL;
715 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
717 for (i = 0; i < sizeof (dva_t); i++)
718 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
720 crc ^= (spa>>8) ^ birth;
725 #define BUF_EMPTY(buf) \
726 ((buf)->b_dva.dva_word[0] == 0 && \
727 (buf)->b_dva.dva_word[1] == 0 && \
730 #define BUF_EQUAL(spa, dva, birth, buf) \
731 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
732 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
733 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
736 buf_discard_identity(arc_buf_hdr_t *hdr)
738 hdr->b_dva.dva_word[0] = 0;
739 hdr->b_dva.dva_word[1] = 0;
744 static arc_buf_hdr_t *
745 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
747 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
748 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
751 mutex_enter(hash_lock);
752 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
753 buf = buf->b_hash_next) {
754 if (BUF_EQUAL(spa, dva, birth, buf)) {
759 mutex_exit(hash_lock);
765 * Insert an entry into the hash table. If there is already an element
766 * equal to elem in the hash table, then the already existing element
767 * will be returned and the new element will not be inserted.
768 * Otherwise returns NULL.
770 static arc_buf_hdr_t *
771 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
773 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
774 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
778 ASSERT(!HDR_IN_HASH_TABLE(buf));
780 mutex_enter(hash_lock);
781 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
782 fbuf = fbuf->b_hash_next, i++) {
783 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
787 buf->b_hash_next = buf_hash_table.ht_table[idx];
788 buf_hash_table.ht_table[idx] = buf;
789 buf->b_flags |= ARC_IN_HASH_TABLE;
791 /* collect some hash table performance data */
793 ARCSTAT_BUMP(arcstat_hash_collisions);
795 ARCSTAT_BUMP(arcstat_hash_chains);
797 ARCSTAT_MAX(arcstat_hash_chain_max, i);
800 ARCSTAT_BUMP(arcstat_hash_elements);
801 ARCSTAT_MAXSTAT(arcstat_hash_elements);
807 buf_hash_remove(arc_buf_hdr_t *buf)
809 arc_buf_hdr_t *fbuf, **bufp;
810 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
812 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
813 ASSERT(HDR_IN_HASH_TABLE(buf));
815 bufp = &buf_hash_table.ht_table[idx];
816 while ((fbuf = *bufp) != buf) {
817 ASSERT(fbuf != NULL);
818 bufp = &fbuf->b_hash_next;
820 *bufp = buf->b_hash_next;
821 buf->b_hash_next = NULL;
822 buf->b_flags &= ~ARC_IN_HASH_TABLE;
824 /* collect some hash table performance data */
825 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
827 if (buf_hash_table.ht_table[idx] &&
828 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
829 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
833 * Global data structures and functions for the buf kmem cache.
835 static kmem_cache_t *hdr_cache;
836 static kmem_cache_t *buf_cache;
843 #if defined(_KERNEL) && defined(HAVE_SPL)
844 /* Large allocations which do not require contiguous pages
845 * should be using vmem_free() in the linux kernel */
846 vmem_free(buf_hash_table.ht_table,
847 (buf_hash_table.ht_mask + 1) * sizeof (void *));
849 kmem_free(buf_hash_table.ht_table,
850 (buf_hash_table.ht_mask + 1) * sizeof (void *));
852 for (i = 0; i < BUF_LOCKS; i++)
853 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
854 kmem_cache_destroy(hdr_cache);
855 kmem_cache_destroy(buf_cache);
859 * Constructor callback - called when the cache is empty
860 * and a new buf is requested.
864 hdr_cons(void *vbuf, void *unused, int kmflag)
866 arc_buf_hdr_t *buf = vbuf;
868 bzero(buf, sizeof (arc_buf_hdr_t));
869 refcount_create(&buf->b_refcnt);
870 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
871 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
872 list_link_init(&buf->b_arc_node);
873 list_link_init(&buf->b_l2node);
874 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
881 buf_cons(void *vbuf, void *unused, int kmflag)
883 arc_buf_t *buf = vbuf;
885 bzero(buf, sizeof (arc_buf_t));
886 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
887 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
893 * Destructor callback - called when a cached buf is
894 * no longer required.
898 hdr_dest(void *vbuf, void *unused)
900 arc_buf_hdr_t *buf = vbuf;
902 ASSERT(BUF_EMPTY(buf));
903 refcount_destroy(&buf->b_refcnt);
904 cv_destroy(&buf->b_cv);
905 mutex_destroy(&buf->b_freeze_lock);
906 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
911 buf_dest(void *vbuf, void *unused)
913 arc_buf_t *buf = vbuf;
915 mutex_destroy(&buf->b_evict_lock);
916 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
923 uint64_t hsize = 1ULL << 12;
927 * The hash table is big enough to fill all of physical memory
928 * with an average 64K block size. The table will take up
929 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
931 while (hsize * 65536 < physmem * PAGESIZE)
934 buf_hash_table.ht_mask = hsize - 1;
935 #if defined(_KERNEL) && defined(HAVE_SPL)
936 /* Large allocations which do not require contiguous pages
937 * should be using vmem_alloc() in the linux kernel */
938 buf_hash_table.ht_table =
939 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
941 buf_hash_table.ht_table =
942 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
944 if (buf_hash_table.ht_table == NULL) {
945 ASSERT(hsize > (1ULL << 8));
950 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
951 0, hdr_cons, hdr_dest, NULL, NULL, NULL, 0);
952 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
953 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
955 for (i = 0; i < 256; i++)
956 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
957 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
959 for (i = 0; i < BUF_LOCKS; i++) {
960 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
961 NULL, MUTEX_DEFAULT, NULL);
965 #define ARC_MINTIME (hz>>4) /* 62 ms */
968 arc_cksum_verify(arc_buf_t *buf)
972 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
975 mutex_enter(&buf->b_hdr->b_freeze_lock);
976 if (buf->b_hdr->b_freeze_cksum == NULL ||
977 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
978 mutex_exit(&buf->b_hdr->b_freeze_lock);
981 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
982 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
983 panic("buffer modified while frozen!");
984 mutex_exit(&buf->b_hdr->b_freeze_lock);
988 arc_cksum_equal(arc_buf_t *buf)
993 mutex_enter(&buf->b_hdr->b_freeze_lock);
994 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
995 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
996 mutex_exit(&buf->b_hdr->b_freeze_lock);
1002 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1004 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1007 mutex_enter(&buf->b_hdr->b_freeze_lock);
1008 if (buf->b_hdr->b_freeze_cksum != NULL) {
1009 mutex_exit(&buf->b_hdr->b_freeze_lock);
1012 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1014 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1015 buf->b_hdr->b_freeze_cksum);
1016 mutex_exit(&buf->b_hdr->b_freeze_lock);
1020 arc_buf_thaw(arc_buf_t *buf)
1022 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1023 if (buf->b_hdr->b_state != arc_anon)
1024 panic("modifying non-anon buffer!");
1025 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1026 panic("modifying buffer while i/o in progress!");
1027 arc_cksum_verify(buf);
1030 mutex_enter(&buf->b_hdr->b_freeze_lock);
1031 if (buf->b_hdr->b_freeze_cksum != NULL) {
1032 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1033 buf->b_hdr->b_freeze_cksum = NULL;
1036 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1037 if (buf->b_hdr->b_thawed)
1038 kmem_free(buf->b_hdr->b_thawed, 1);
1039 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1042 mutex_exit(&buf->b_hdr->b_freeze_lock);
1046 arc_buf_freeze(arc_buf_t *buf)
1048 kmutex_t *hash_lock;
1050 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1053 hash_lock = HDR_LOCK(buf->b_hdr);
1054 mutex_enter(hash_lock);
1056 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1057 buf->b_hdr->b_state == arc_anon);
1058 arc_cksum_compute(buf, B_FALSE);
1059 mutex_exit(hash_lock);
1063 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1065 ASSERT(MUTEX_HELD(hash_lock));
1067 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1068 (ab->b_state != arc_anon)) {
1069 uint64_t delta = ab->b_size * ab->b_datacnt;
1070 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1071 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1073 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1074 mutex_enter(&ab->b_state->arcs_mtx);
1075 ASSERT(list_link_active(&ab->b_arc_node));
1076 list_remove(list, ab);
1077 if (GHOST_STATE(ab->b_state)) {
1078 ASSERT0(ab->b_datacnt);
1079 ASSERT3P(ab->b_buf, ==, NULL);
1083 ASSERT3U(*size, >=, delta);
1084 atomic_add_64(size, -delta);
1085 mutex_exit(&ab->b_state->arcs_mtx);
1086 /* remove the prefetch flag if we get a reference */
1087 if (ab->b_flags & ARC_PREFETCH)
1088 ab->b_flags &= ~ARC_PREFETCH;
1093 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1096 arc_state_t *state = ab->b_state;
1098 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1099 ASSERT(!GHOST_STATE(state));
1101 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1102 (state != arc_anon)) {
1103 uint64_t *size = &state->arcs_lsize[ab->b_type];
1105 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1106 mutex_enter(&state->arcs_mtx);
1107 ASSERT(!list_link_active(&ab->b_arc_node));
1108 list_insert_head(&state->arcs_list[ab->b_type], ab);
1109 ASSERT(ab->b_datacnt > 0);
1110 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1111 mutex_exit(&state->arcs_mtx);
1117 * Move the supplied buffer to the indicated state. The mutex
1118 * for the buffer must be held by the caller.
1121 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1123 arc_state_t *old_state = ab->b_state;
1124 int64_t refcnt = refcount_count(&ab->b_refcnt);
1125 uint64_t from_delta, to_delta;
1127 ASSERT(MUTEX_HELD(hash_lock));
1128 ASSERT(new_state != old_state);
1129 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1130 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1131 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1133 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1136 * If this buffer is evictable, transfer it from the
1137 * old state list to the new state list.
1140 if (old_state != arc_anon) {
1141 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1142 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1145 mutex_enter(&old_state->arcs_mtx);
1147 ASSERT(list_link_active(&ab->b_arc_node));
1148 list_remove(&old_state->arcs_list[ab->b_type], ab);
1151 * If prefetching out of the ghost cache,
1152 * we will have a non-zero datacnt.
1154 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1155 /* ghost elements have a ghost size */
1156 ASSERT(ab->b_buf == NULL);
1157 from_delta = ab->b_size;
1159 ASSERT3U(*size, >=, from_delta);
1160 atomic_add_64(size, -from_delta);
1163 mutex_exit(&old_state->arcs_mtx);
1165 if (new_state != arc_anon) {
1166 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1167 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1170 mutex_enter(&new_state->arcs_mtx);
1172 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1174 /* ghost elements have a ghost size */
1175 if (GHOST_STATE(new_state)) {
1176 ASSERT(ab->b_datacnt == 0);
1177 ASSERT(ab->b_buf == NULL);
1178 to_delta = ab->b_size;
1180 atomic_add_64(size, to_delta);
1183 mutex_exit(&new_state->arcs_mtx);
1187 ASSERT(!BUF_EMPTY(ab));
1188 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1189 buf_hash_remove(ab);
1191 /* adjust state sizes */
1193 atomic_add_64(&new_state->arcs_size, to_delta);
1195 ASSERT3U(old_state->arcs_size, >=, from_delta);
1196 atomic_add_64(&old_state->arcs_size, -from_delta);
1198 ab->b_state = new_state;
1200 /* adjust l2arc hdr stats */
1201 if (new_state == arc_l2c_only)
1202 l2arc_hdr_stat_add();
1203 else if (old_state == arc_l2c_only)
1204 l2arc_hdr_stat_remove();
1208 arc_space_consume(uint64_t space, arc_space_type_t type)
1210 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1215 case ARC_SPACE_DATA:
1216 ARCSTAT_INCR(arcstat_data_size, space);
1218 case ARC_SPACE_OTHER:
1219 ARCSTAT_INCR(arcstat_other_size, space);
1221 case ARC_SPACE_HDRS:
1222 ARCSTAT_INCR(arcstat_hdr_size, space);
1224 case ARC_SPACE_L2HDRS:
1225 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1229 atomic_add_64(&arc_meta_used, space);
1230 atomic_add_64(&arc_size, space);
1234 arc_space_return(uint64_t space, arc_space_type_t type)
1236 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1241 case ARC_SPACE_DATA:
1242 ARCSTAT_INCR(arcstat_data_size, -space);
1244 case ARC_SPACE_OTHER:
1245 ARCSTAT_INCR(arcstat_other_size, -space);
1247 case ARC_SPACE_HDRS:
1248 ARCSTAT_INCR(arcstat_hdr_size, -space);
1250 case ARC_SPACE_L2HDRS:
1251 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1255 ASSERT(arc_meta_used >= space);
1256 if (arc_meta_max < arc_meta_used)
1257 arc_meta_max = arc_meta_used;
1258 atomic_add_64(&arc_meta_used, -space);
1259 ASSERT(arc_size >= space);
1260 atomic_add_64(&arc_size, -space);
1264 arc_data_buf_alloc(uint64_t size)
1266 if (arc_evict_needed(ARC_BUFC_DATA))
1267 cv_signal(&arc_reclaim_thr_cv);
1268 atomic_add_64(&arc_size, size);
1269 return (zio_data_buf_alloc(size));
1273 arc_data_buf_free(void *buf, uint64_t size)
1275 zio_data_buf_free(buf, size);
1276 ASSERT(arc_size >= size);
1277 atomic_add_64(&arc_size, -size);
1281 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1286 ASSERT3U(size, >, 0);
1287 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1288 ASSERT(BUF_EMPTY(hdr));
1291 hdr->b_spa = spa_load_guid(spa);
1292 hdr->b_state = arc_anon;
1293 hdr->b_arc_access = 0;
1294 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1297 buf->b_efunc = NULL;
1298 buf->b_private = NULL;
1301 arc_get_data_buf(buf);
1304 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1305 (void) refcount_add(&hdr->b_refcnt, tag);
1310 static char *arc_onloan_tag = "onloan";
1313 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1314 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1315 * buffers must be returned to the arc before they can be used by the DMU or
1319 arc_loan_buf(spa_t *spa, int size)
1323 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1325 atomic_add_64(&arc_loaned_bytes, size);
1330 * Return a loaned arc buffer to the arc.
1333 arc_return_buf(arc_buf_t *buf, void *tag)
1335 arc_buf_hdr_t *hdr = buf->b_hdr;
1337 ASSERT(buf->b_data != NULL);
1338 (void) refcount_add(&hdr->b_refcnt, tag);
1339 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1341 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1344 /* Detach an arc_buf from a dbuf (tag) */
1346 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1350 ASSERT(buf->b_data != NULL);
1352 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1353 (void) refcount_remove(&hdr->b_refcnt, tag);
1354 buf->b_efunc = NULL;
1355 buf->b_private = NULL;
1357 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1361 arc_buf_clone(arc_buf_t *from)
1364 arc_buf_hdr_t *hdr = from->b_hdr;
1365 uint64_t size = hdr->b_size;
1367 ASSERT(hdr->b_state != arc_anon);
1369 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1372 buf->b_efunc = NULL;
1373 buf->b_private = NULL;
1374 buf->b_next = hdr->b_buf;
1376 arc_get_data_buf(buf);
1377 bcopy(from->b_data, buf->b_data, size);
1380 * This buffer already exists in the arc so create a duplicate
1381 * copy for the caller. If the buffer is associated with user data
1382 * then track the size and number of duplicates. These stats will be
1383 * updated as duplicate buffers are created and destroyed.
1385 if (hdr->b_type == ARC_BUFC_DATA) {
1386 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1387 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1389 hdr->b_datacnt += 1;
1394 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1397 kmutex_t *hash_lock;
1400 * Check to see if this buffer is evicted. Callers
1401 * must verify b_data != NULL to know if the add_ref
1404 mutex_enter(&buf->b_evict_lock);
1405 if (buf->b_data == NULL) {
1406 mutex_exit(&buf->b_evict_lock);
1409 hash_lock = HDR_LOCK(buf->b_hdr);
1410 mutex_enter(hash_lock);
1412 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1413 mutex_exit(&buf->b_evict_lock);
1415 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1416 add_reference(hdr, hash_lock, tag);
1417 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1418 arc_access(hdr, hash_lock);
1419 mutex_exit(hash_lock);
1420 ARCSTAT_BUMP(arcstat_hits);
1421 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1422 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1423 data, metadata, hits);
1427 * Free the arc data buffer. If it is an l2arc write in progress,
1428 * the buffer is placed on l2arc_free_on_write to be freed later.
1431 arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t),
1432 void *data, size_t size)
1434 if (HDR_L2_WRITING(hdr)) {
1435 l2arc_data_free_t *df;
1436 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_PUSHPAGE);
1437 df->l2df_data = data;
1438 df->l2df_size = size;
1439 df->l2df_func = free_func;
1440 mutex_enter(&l2arc_free_on_write_mtx);
1441 list_insert_head(l2arc_free_on_write, df);
1442 mutex_exit(&l2arc_free_on_write_mtx);
1443 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1445 free_func(data, size);
1450 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1454 /* free up data associated with the buf */
1456 arc_state_t *state = buf->b_hdr->b_state;
1457 uint64_t size = buf->b_hdr->b_size;
1458 arc_buf_contents_t type = buf->b_hdr->b_type;
1460 arc_cksum_verify(buf);
1463 if (type == ARC_BUFC_METADATA) {
1464 arc_buf_data_free(buf->b_hdr, zio_buf_free,
1466 arc_space_return(size, ARC_SPACE_DATA);
1468 ASSERT(type == ARC_BUFC_DATA);
1469 arc_buf_data_free(buf->b_hdr,
1470 zio_data_buf_free, buf->b_data, size);
1471 ARCSTAT_INCR(arcstat_data_size, -size);
1472 atomic_add_64(&arc_size, -size);
1475 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1476 uint64_t *cnt = &state->arcs_lsize[type];
1478 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1479 ASSERT(state != arc_anon);
1481 ASSERT3U(*cnt, >=, size);
1482 atomic_add_64(cnt, -size);
1484 ASSERT3U(state->arcs_size, >=, size);
1485 atomic_add_64(&state->arcs_size, -size);
1489 * If we're destroying a duplicate buffer make sure
1490 * that the appropriate statistics are updated.
1492 if (buf->b_hdr->b_datacnt > 1 &&
1493 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1494 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1495 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1497 ASSERT(buf->b_hdr->b_datacnt > 0);
1498 buf->b_hdr->b_datacnt -= 1;
1501 /* only remove the buf if requested */
1505 /* remove the buf from the hdr list */
1506 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1508 *bufp = buf->b_next;
1511 ASSERT(buf->b_efunc == NULL);
1513 /* clean up the buf */
1515 kmem_cache_free(buf_cache, buf);
1519 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1521 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1523 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1524 ASSERT3P(hdr->b_state, ==, arc_anon);
1525 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1527 if (l2hdr != NULL) {
1528 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1530 * To prevent arc_free() and l2arc_evict() from
1531 * attempting to free the same buffer at the same time,
1532 * a FREE_IN_PROGRESS flag is given to arc_free() to
1533 * give it priority. l2arc_evict() can't destroy this
1534 * header while we are waiting on l2arc_buflist_mtx.
1536 * The hdr may be removed from l2ad_buflist before we
1537 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1539 if (!buflist_held) {
1540 mutex_enter(&l2arc_buflist_mtx);
1541 l2hdr = hdr->b_l2hdr;
1544 if (l2hdr != NULL) {
1545 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1546 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1547 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1548 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
1549 if (hdr->b_state == arc_l2c_only)
1550 l2arc_hdr_stat_remove();
1551 hdr->b_l2hdr = NULL;
1555 mutex_exit(&l2arc_buflist_mtx);
1558 if (!BUF_EMPTY(hdr)) {
1559 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1560 buf_discard_identity(hdr);
1562 while (hdr->b_buf) {
1563 arc_buf_t *buf = hdr->b_buf;
1566 mutex_enter(&arc_eviction_mtx);
1567 mutex_enter(&buf->b_evict_lock);
1568 ASSERT(buf->b_hdr != NULL);
1569 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1570 hdr->b_buf = buf->b_next;
1571 buf->b_hdr = &arc_eviction_hdr;
1572 buf->b_next = arc_eviction_list;
1573 arc_eviction_list = buf;
1574 mutex_exit(&buf->b_evict_lock);
1575 mutex_exit(&arc_eviction_mtx);
1577 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1580 if (hdr->b_freeze_cksum != NULL) {
1581 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1582 hdr->b_freeze_cksum = NULL;
1584 if (hdr->b_thawed) {
1585 kmem_free(hdr->b_thawed, 1);
1586 hdr->b_thawed = NULL;
1589 ASSERT(!list_link_active(&hdr->b_arc_node));
1590 ASSERT3P(hdr->b_hash_next, ==, NULL);
1591 ASSERT3P(hdr->b_acb, ==, NULL);
1592 kmem_cache_free(hdr_cache, hdr);
1596 arc_buf_free(arc_buf_t *buf, void *tag)
1598 arc_buf_hdr_t *hdr = buf->b_hdr;
1599 int hashed = hdr->b_state != arc_anon;
1601 ASSERT(buf->b_efunc == NULL);
1602 ASSERT(buf->b_data != NULL);
1605 kmutex_t *hash_lock = HDR_LOCK(hdr);
1607 mutex_enter(hash_lock);
1609 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1611 (void) remove_reference(hdr, hash_lock, tag);
1612 if (hdr->b_datacnt > 1) {
1613 arc_buf_destroy(buf, FALSE, TRUE);
1615 ASSERT(buf == hdr->b_buf);
1616 ASSERT(buf->b_efunc == NULL);
1617 hdr->b_flags |= ARC_BUF_AVAILABLE;
1619 mutex_exit(hash_lock);
1620 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1623 * We are in the middle of an async write. Don't destroy
1624 * this buffer unless the write completes before we finish
1625 * decrementing the reference count.
1627 mutex_enter(&arc_eviction_mtx);
1628 (void) remove_reference(hdr, NULL, tag);
1629 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1630 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1631 mutex_exit(&arc_eviction_mtx);
1633 arc_hdr_destroy(hdr);
1635 if (remove_reference(hdr, NULL, tag) > 0)
1636 arc_buf_destroy(buf, FALSE, TRUE);
1638 arc_hdr_destroy(hdr);
1643 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1645 arc_buf_hdr_t *hdr = buf->b_hdr;
1646 kmutex_t *hash_lock = NULL;
1647 int no_callback = (buf->b_efunc == NULL);
1649 if (hdr->b_state == arc_anon) {
1650 ASSERT(hdr->b_datacnt == 1);
1651 arc_buf_free(buf, tag);
1652 return (no_callback);
1655 hash_lock = HDR_LOCK(hdr);
1656 mutex_enter(hash_lock);
1658 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1659 ASSERT(hdr->b_state != arc_anon);
1660 ASSERT(buf->b_data != NULL);
1662 (void) remove_reference(hdr, hash_lock, tag);
1663 if (hdr->b_datacnt > 1) {
1665 arc_buf_destroy(buf, FALSE, TRUE);
1666 } else if (no_callback) {
1667 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1668 ASSERT(buf->b_efunc == NULL);
1669 hdr->b_flags |= ARC_BUF_AVAILABLE;
1671 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1672 refcount_is_zero(&hdr->b_refcnt));
1673 mutex_exit(hash_lock);
1674 return (no_callback);
1678 arc_buf_size(arc_buf_t *buf)
1680 return (buf->b_hdr->b_size);
1684 * Called from the DMU to determine if the current buffer should be
1685 * evicted. In order to ensure proper locking, the eviction must be initiated
1686 * from the DMU. Return true if the buffer is associated with user data and
1687 * duplicate buffers still exist.
1690 arc_buf_eviction_needed(arc_buf_t *buf)
1693 boolean_t evict_needed = B_FALSE;
1695 if (zfs_disable_dup_eviction)
1698 mutex_enter(&buf->b_evict_lock);
1702 * We are in arc_do_user_evicts(); let that function
1703 * perform the eviction.
1705 ASSERT(buf->b_data == NULL);
1706 mutex_exit(&buf->b_evict_lock);
1708 } else if (buf->b_data == NULL) {
1710 * We have already been added to the arc eviction list;
1711 * recommend eviction.
1713 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1714 mutex_exit(&buf->b_evict_lock);
1718 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1719 evict_needed = B_TRUE;
1721 mutex_exit(&buf->b_evict_lock);
1722 return (evict_needed);
1726 * Evict buffers from list until we've removed the specified number of
1727 * bytes. Move the removed buffers to the appropriate evict state.
1728 * If the recycle flag is set, then attempt to "recycle" a buffer:
1729 * - look for a buffer to evict that is `bytes' long.
1730 * - return the data block from this buffer rather than freeing it.
1731 * This flag is used by callers that are trying to make space for a
1732 * new buffer in a full arc cache.
1734 * This function makes a "best effort". It skips over any buffers
1735 * it can't get a hash_lock on, and so may not catch all candidates.
1736 * It may also return without evicting as much space as requested.
1739 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1740 arc_buf_contents_t type)
1742 arc_state_t *evicted_state;
1743 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1744 arc_buf_hdr_t *ab, *ab_prev = NULL;
1745 list_t *list = &state->arcs_list[type];
1746 kmutex_t *hash_lock;
1747 boolean_t have_lock;
1748 void *stolen = NULL;
1750 ASSERT(state == arc_mru || state == arc_mfu);
1752 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1754 mutex_enter(&state->arcs_mtx);
1755 mutex_enter(&evicted_state->arcs_mtx);
1757 for (ab = list_tail(list); ab; ab = ab_prev) {
1758 ab_prev = list_prev(list, ab);
1759 /* prefetch buffers have a minimum lifespan */
1760 if (HDR_IO_IN_PROGRESS(ab) ||
1761 (spa && ab->b_spa != spa) ||
1762 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1763 ddi_get_lbolt() - ab->b_arc_access <
1764 zfs_arc_min_prefetch_lifespan)) {
1768 /* "lookahead" for better eviction candidate */
1769 if (recycle && ab->b_size != bytes &&
1770 ab_prev && ab_prev->b_size == bytes)
1772 hash_lock = HDR_LOCK(ab);
1773 have_lock = MUTEX_HELD(hash_lock);
1774 if (have_lock || mutex_tryenter(hash_lock)) {
1775 ASSERT0(refcount_count(&ab->b_refcnt));
1776 ASSERT(ab->b_datacnt > 0);
1778 arc_buf_t *buf = ab->b_buf;
1779 if (!mutex_tryenter(&buf->b_evict_lock)) {
1784 bytes_evicted += ab->b_size;
1785 if (recycle && ab->b_type == type &&
1786 ab->b_size == bytes &&
1787 !HDR_L2_WRITING(ab)) {
1788 stolen = buf->b_data;
1793 mutex_enter(&arc_eviction_mtx);
1794 arc_buf_destroy(buf,
1795 buf->b_data == stolen, FALSE);
1796 ab->b_buf = buf->b_next;
1797 buf->b_hdr = &arc_eviction_hdr;
1798 buf->b_next = arc_eviction_list;
1799 arc_eviction_list = buf;
1800 mutex_exit(&arc_eviction_mtx);
1801 mutex_exit(&buf->b_evict_lock);
1803 mutex_exit(&buf->b_evict_lock);
1804 arc_buf_destroy(buf,
1805 buf->b_data == stolen, TRUE);
1810 ARCSTAT_INCR(arcstat_evict_l2_cached,
1813 if (l2arc_write_eligible(ab->b_spa, ab)) {
1814 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1818 arcstat_evict_l2_ineligible,
1823 if (ab->b_datacnt == 0) {
1824 arc_change_state(evicted_state, ab, hash_lock);
1825 ASSERT(HDR_IN_HASH_TABLE(ab));
1826 ab->b_flags |= ARC_IN_HASH_TABLE;
1827 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1828 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1831 mutex_exit(hash_lock);
1832 if (bytes >= 0 && bytes_evicted >= bytes)
1839 mutex_exit(&evicted_state->arcs_mtx);
1840 mutex_exit(&state->arcs_mtx);
1842 if (bytes_evicted < bytes)
1843 dprintf("only evicted %lld bytes from %x\n",
1844 (longlong_t)bytes_evicted, state);
1847 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1850 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1853 * We have just evicted some date into the ghost state, make
1854 * sure we also adjust the ghost state size if necessary.
1857 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1858 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1859 arc_mru_ghost->arcs_size - arc_c;
1861 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1863 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1864 arc_evict_ghost(arc_mru_ghost, 0, todelete);
1865 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1866 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1867 arc_mru_ghost->arcs_size +
1868 arc_mfu_ghost->arcs_size - arc_c);
1869 arc_evict_ghost(arc_mfu_ghost, 0, todelete);
1877 * Remove buffers from list until we've removed the specified number of
1878 * bytes. Destroy the buffers that are removed.
1881 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1883 arc_buf_hdr_t *ab, *ab_prev;
1884 arc_buf_hdr_t marker;
1885 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1886 kmutex_t *hash_lock;
1887 uint64_t bytes_deleted = 0;
1888 uint64_t bufs_skipped = 0;
1890 ASSERT(GHOST_STATE(state));
1891 bzero(&marker, sizeof(marker));
1893 mutex_enter(&state->arcs_mtx);
1894 for (ab = list_tail(list); ab; ab = ab_prev) {
1895 ab_prev = list_prev(list, ab);
1896 if (spa && ab->b_spa != spa)
1899 /* ignore markers */
1903 hash_lock = HDR_LOCK(ab);
1904 /* caller may be trying to modify this buffer, skip it */
1905 if (MUTEX_HELD(hash_lock))
1907 if (mutex_tryenter(hash_lock)) {
1908 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1909 ASSERT(ab->b_buf == NULL);
1910 ARCSTAT_BUMP(arcstat_deleted);
1911 bytes_deleted += ab->b_size;
1913 if (ab->b_l2hdr != NULL) {
1915 * This buffer is cached on the 2nd Level ARC;
1916 * don't destroy the header.
1918 arc_change_state(arc_l2c_only, ab, hash_lock);
1919 mutex_exit(hash_lock);
1921 arc_change_state(arc_anon, ab, hash_lock);
1922 mutex_exit(hash_lock);
1923 arc_hdr_destroy(ab);
1926 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1927 if (bytes >= 0 && bytes_deleted >= bytes)
1929 } else if (bytes < 0) {
1931 * Insert a list marker and then wait for the
1932 * hash lock to become available. Once its
1933 * available, restart from where we left off.
1935 list_insert_after(list, ab, &marker);
1936 mutex_exit(&state->arcs_mtx);
1937 mutex_enter(hash_lock);
1938 mutex_exit(hash_lock);
1939 mutex_enter(&state->arcs_mtx);
1940 ab_prev = list_prev(list, &marker);
1941 list_remove(list, &marker);
1945 mutex_exit(&state->arcs_mtx);
1947 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1948 (bytes < 0 || bytes_deleted < bytes)) {
1949 list = &state->arcs_list[ARC_BUFC_METADATA];
1954 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1958 if (bytes_deleted < bytes)
1959 dprintf("only deleted %lld bytes from %p\n",
1960 (longlong_t)bytes_deleted, state);
1966 int64_t adjustment, delta;
1972 adjustment = MIN((int64_t)(arc_size - arc_c),
1973 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
1976 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1977 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
1978 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
1979 adjustment -= delta;
1982 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1983 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
1984 (void) arc_evict(arc_mru, 0, delta, FALSE,
1992 adjustment = arc_size - arc_c;
1994 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1995 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
1996 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
1997 adjustment -= delta;
2000 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2001 int64_t delta = MIN(adjustment,
2002 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2003 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2008 * Adjust ghost lists
2011 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2013 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2014 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2015 arc_evict_ghost(arc_mru_ghost, 0, delta);
2019 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2021 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2022 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2023 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2028 * Request that arc user drop references so that N bytes can be released
2029 * from the cache. This provides a mechanism to ensure the arc can honor
2030 * the arc_meta_limit and reclaim buffers which are pinned in the cache
2031 * by higher layers. (i.e. the zpl)
2034 arc_do_user_prune(int64_t adjustment)
2036 arc_prune_func_t *func;
2038 arc_prune_t *cp, *np;
2040 mutex_enter(&arc_prune_mtx);
2042 cp = list_head(&arc_prune_list);
2043 while (cp != NULL) {
2045 private = cp->p_private;
2046 np = list_next(&arc_prune_list, cp);
2047 refcount_add(&cp->p_refcnt, func);
2048 mutex_exit(&arc_prune_mtx);
2051 func(adjustment, private);
2053 mutex_enter(&arc_prune_mtx);
2055 /* User removed prune callback concurrently with execution */
2056 if (refcount_remove(&cp->p_refcnt, func) == 0) {
2057 ASSERT(!list_link_active(&cp->p_node));
2058 refcount_destroy(&cp->p_refcnt);
2059 kmem_free(cp, sizeof (*cp));
2065 ARCSTAT_BUMP(arcstat_prune);
2066 mutex_exit(&arc_prune_mtx);
2070 arc_do_user_evicts(void)
2072 mutex_enter(&arc_eviction_mtx);
2073 while (arc_eviction_list != NULL) {
2074 arc_buf_t *buf = arc_eviction_list;
2075 arc_eviction_list = buf->b_next;
2076 mutex_enter(&buf->b_evict_lock);
2078 mutex_exit(&buf->b_evict_lock);
2079 mutex_exit(&arc_eviction_mtx);
2081 if (buf->b_efunc != NULL)
2082 VERIFY(buf->b_efunc(buf) == 0);
2084 buf->b_efunc = NULL;
2085 buf->b_private = NULL;
2086 kmem_cache_free(buf_cache, buf);
2087 mutex_enter(&arc_eviction_mtx);
2089 mutex_exit(&arc_eviction_mtx);
2093 * Evict only meta data objects from the cache leaving the data objects.
2094 * This is only used to enforce the tunable arc_meta_limit, if we are
2095 * unable to evict enough buffers notify the user via the prune callback.
2098 arc_adjust_meta(int64_t adjustment, boolean_t may_prune)
2102 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2103 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2104 arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_METADATA);
2105 adjustment -= delta;
2108 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2109 delta = MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2110 arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_METADATA);
2111 adjustment -= delta;
2114 if (may_prune && (adjustment > 0) && (arc_meta_used > arc_meta_limit))
2115 arc_do_user_prune(zfs_arc_meta_prune);
2119 * Flush all *evictable* data from the cache for the given spa.
2120 * NOTE: this will not touch "active" (i.e. referenced) data.
2123 arc_flush(spa_t *spa)
2128 guid = spa_load_guid(spa);
2130 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2131 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2135 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2136 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2140 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2141 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2145 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2146 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2151 arc_evict_ghost(arc_mru_ghost, guid, -1);
2152 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2154 mutex_enter(&arc_reclaim_thr_lock);
2155 arc_do_user_evicts();
2156 mutex_exit(&arc_reclaim_thr_lock);
2157 ASSERT(spa || arc_eviction_list == NULL);
2161 arc_shrink(uint64_t bytes)
2163 if (arc_c > arc_c_min) {
2166 to_free = bytes ? bytes : arc_c >> zfs_arc_shrink_shift;
2168 if (arc_c > arc_c_min + to_free)
2169 atomic_add_64(&arc_c, -to_free);
2173 atomic_add_64(&arc_p, -(arc_p >> zfs_arc_shrink_shift));
2174 if (arc_c > arc_size)
2175 arc_c = MAX(arc_size, arc_c_min);
2177 arc_p = (arc_c >> 1);
2178 ASSERT(arc_c >= arc_c_min);
2179 ASSERT((int64_t)arc_p >= 0);
2182 if (arc_size > arc_c)
2187 arc_kmem_reap_now(arc_reclaim_strategy_t strat, uint64_t bytes)
2190 kmem_cache_t *prev_cache = NULL;
2191 kmem_cache_t *prev_data_cache = NULL;
2192 extern kmem_cache_t *zio_buf_cache[];
2193 extern kmem_cache_t *zio_data_buf_cache[];
2196 * An aggressive reclamation will shrink the cache size as well as
2197 * reap free buffers from the arc kmem caches.
2199 if (strat == ARC_RECLAIM_AGGR)
2202 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2203 if (zio_buf_cache[i] != prev_cache) {
2204 prev_cache = zio_buf_cache[i];
2205 kmem_cache_reap_now(zio_buf_cache[i]);
2207 if (zio_data_buf_cache[i] != prev_data_cache) {
2208 prev_data_cache = zio_data_buf_cache[i];
2209 kmem_cache_reap_now(zio_data_buf_cache[i]);
2213 kmem_cache_reap_now(buf_cache);
2214 kmem_cache_reap_now(hdr_cache);
2218 * Unlike other ZFS implementations this thread is only responsible for
2219 * adapting the target ARC size on Linux. The responsibility for memory
2220 * reclamation has been entirely delegated to the arc_shrinker_func()
2221 * which is registered with the VM. To reflect this change in behavior
2222 * the arc_reclaim thread has been renamed to arc_adapt.
2225 arc_adapt_thread(void)
2230 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2232 mutex_enter(&arc_reclaim_thr_lock);
2233 while (arc_thread_exit == 0) {
2235 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2237 if (spa_get_random(100) == 0) {
2240 if (last_reclaim == ARC_RECLAIM_CONS) {
2241 last_reclaim = ARC_RECLAIM_AGGR;
2243 last_reclaim = ARC_RECLAIM_CONS;
2247 last_reclaim = ARC_RECLAIM_AGGR;
2251 /* reset the growth delay for every reclaim */
2252 arc_grow_time = ddi_get_lbolt()+(zfs_arc_grow_retry * hz);
2254 arc_kmem_reap_now(last_reclaim, 0);
2257 #endif /* !_KERNEL */
2259 /* No recent memory pressure allow the ARC to grow. */
2260 if (arc_no_grow && ddi_get_lbolt() >= arc_grow_time)
2261 arc_no_grow = FALSE;
2264 * Keep meta data usage within limits, arc_shrink() is not
2265 * used to avoid collapsing the arc_c value when only the
2266 * arc_meta_limit is being exceeded.
2268 prune = (int64_t)arc_meta_used - (int64_t)arc_meta_limit;
2270 arc_adjust_meta(prune, B_TRUE);
2274 if (arc_eviction_list != NULL)
2275 arc_do_user_evicts();
2277 /* block until needed, or one second, whichever is shorter */
2278 CALLB_CPR_SAFE_BEGIN(&cpr);
2279 (void) cv_timedwait_interruptible(&arc_reclaim_thr_cv,
2280 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2281 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2284 /* Allow the module options to be changed */
2285 if (zfs_arc_max > 64 << 20 &&
2286 zfs_arc_max < physmem * PAGESIZE &&
2287 zfs_arc_max != arc_c_max)
2288 arc_c_max = zfs_arc_max;
2290 if (zfs_arc_min > 0 &&
2291 zfs_arc_min < arc_c_max &&
2292 zfs_arc_min != arc_c_min)
2293 arc_c_min = zfs_arc_min;
2295 if (zfs_arc_meta_limit > 0 &&
2296 zfs_arc_meta_limit <= arc_c_max &&
2297 zfs_arc_meta_limit != arc_meta_limit)
2298 arc_meta_limit = zfs_arc_meta_limit;
2304 arc_thread_exit = 0;
2305 cv_broadcast(&arc_reclaim_thr_cv);
2306 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2312 * Determine the amount of memory eligible for eviction contained in the
2313 * ARC. All clean data reported by the ghost lists can always be safely
2314 * evicted. Due to arc_c_min, the same does not hold for all clean data
2315 * contained by the regular mru and mfu lists.
2317 * In the case of the regular mru and mfu lists, we need to report as
2318 * much clean data as possible, such that evicting that same reported
2319 * data will not bring arc_size below arc_c_min. Thus, in certain
2320 * circumstances, the total amount of clean data in the mru and mfu
2321 * lists might not actually be evictable.
2323 * The following two distinct cases are accounted for:
2325 * 1. The sum of the amount of dirty data contained by both the mru and
2326 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2327 * is greater than or equal to arc_c_min.
2328 * (i.e. amount of dirty data >= arc_c_min)
2330 * This is the easy case; all clean data contained by the mru and mfu
2331 * lists is evictable. Evicting all clean data can only drop arc_size
2332 * to the amount of dirty data, which is greater than arc_c_min.
2334 * 2. The sum of the amount of dirty data contained by both the mru and
2335 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2336 * is less than arc_c_min.
2337 * (i.e. arc_c_min > amount of dirty data)
2339 * 2.1. arc_size is greater than or equal arc_c_min.
2340 * (i.e. arc_size >= arc_c_min > amount of dirty data)
2342 * In this case, not all clean data from the regular mru and mfu
2343 * lists is actually evictable; we must leave enough clean data
2344 * to keep arc_size above arc_c_min. Thus, the maximum amount of
2345 * evictable data from the two lists combined, is exactly the
2346 * difference between arc_size and arc_c_min.
2348 * 2.2. arc_size is less than arc_c_min
2349 * (i.e. arc_c_min > arc_size > amount of dirty data)
2351 * In this case, none of the data contained in the mru and mfu
2352 * lists is evictable, even if it's clean. Since arc_size is
2353 * already below arc_c_min, evicting any more would only
2354 * increase this negative difference.
2357 arc_evictable_memory(void) {
2358 uint64_t arc_clean =
2359 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
2360 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
2361 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
2362 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
2363 uint64_t ghost_clean =
2364 arc_mru_ghost->arcs_lsize[ARC_BUFC_DATA] +
2365 arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] +
2366 arc_mfu_ghost->arcs_lsize[ARC_BUFC_DATA] +
2367 arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA];
2368 uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0);
2370 if (arc_dirty >= arc_c_min)
2371 return (ghost_clean + arc_clean);
2373 return (ghost_clean + MAX((int64_t)arc_size - (int64_t)arc_c_min, 0));
2377 __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc)
2381 /* The arc is considered warm once reclaim has occurred */
2382 if (unlikely(arc_warm == B_FALSE))
2385 /* Return the potential number of reclaimable pages */
2386 pages = btop(arc_evictable_memory());
2387 if (sc->nr_to_scan == 0)
2390 /* Not allowed to perform filesystem reclaim */
2391 if (!(sc->gfp_mask & __GFP_FS))
2394 /* Reclaim in progress */
2395 if (mutex_tryenter(&arc_reclaim_thr_lock) == 0)
2399 * Evict the requested number of pages by shrinking arc_c the
2400 * requested amount. If there is nothing left to evict just
2401 * reap whatever we can from the various arc slabs.
2404 arc_kmem_reap_now(ARC_RECLAIM_AGGR, ptob(sc->nr_to_scan));
2405 pages = btop(arc_evictable_memory());
2407 arc_kmem_reap_now(ARC_RECLAIM_CONS, ptob(sc->nr_to_scan));
2412 * When direct reclaim is observed it usually indicates a rapid
2413 * increase in memory pressure. This occurs because the kswapd
2414 * threads were unable to asynchronously keep enough free memory
2415 * available. In this case set arc_no_grow to briefly pause arc
2416 * growth to avoid compounding the memory pressure.
2418 if (current_is_kswapd()) {
2419 ARCSTAT_BUMP(arcstat_memory_indirect_count);
2421 arc_no_grow = B_TRUE;
2422 arc_grow_time = ddi_get_lbolt() + (zfs_arc_grow_retry * hz);
2423 ARCSTAT_BUMP(arcstat_memory_direct_count);
2426 mutex_exit(&arc_reclaim_thr_lock);
2430 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func);
2432 SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS);
2433 #endif /* _KERNEL */
2436 * Adapt arc info given the number of bytes we are trying to add and
2437 * the state that we are comming from. This function is only called
2438 * when we are adding new content to the cache.
2441 arc_adapt(int bytes, arc_state_t *state)
2444 uint64_t arc_p_min = (arc_c >> zfs_arc_p_min_shift);
2446 if (state == arc_l2c_only)
2451 * Adapt the target size of the MRU list:
2452 * - if we just hit in the MRU ghost list, then increase
2453 * the target size of the MRU list.
2454 * - if we just hit in the MFU ghost list, then increase
2455 * the target size of the MFU list by decreasing the
2456 * target size of the MRU list.
2458 if (state == arc_mru_ghost) {
2459 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2460 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2461 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2463 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2464 } else if (state == arc_mfu_ghost) {
2467 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2468 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2469 mult = MIN(mult, 10);
2471 delta = MIN(bytes * mult, arc_p);
2472 arc_p = MAX(arc_p_min, arc_p - delta);
2474 ASSERT((int64_t)arc_p >= 0);
2479 if (arc_c >= arc_c_max)
2483 * If we're within (2 * maxblocksize) bytes of the target
2484 * cache size, increment the target cache size
2486 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2487 atomic_add_64(&arc_c, (int64_t)bytes);
2488 if (arc_c > arc_c_max)
2490 else if (state == arc_anon)
2491 atomic_add_64(&arc_p, (int64_t)bytes);
2495 ASSERT((int64_t)arc_p >= 0);
2499 * Check if the cache has reached its limits and eviction is required
2503 arc_evict_needed(arc_buf_contents_t type)
2505 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2511 return (arc_size > arc_c);
2515 * The buffer, supplied as the first argument, needs a data block.
2516 * So, if we are at cache max, determine which cache should be victimized.
2517 * We have the following cases:
2519 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2520 * In this situation if we're out of space, but the resident size of the MFU is
2521 * under the limit, victimize the MFU cache to satisfy this insertion request.
2523 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2524 * Here, we've used up all of the available space for the MRU, so we need to
2525 * evict from our own cache instead. Evict from the set of resident MRU
2528 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2529 * c minus p represents the MFU space in the cache, since p is the size of the
2530 * cache that is dedicated to the MRU. In this situation there's still space on
2531 * the MFU side, so the MRU side needs to be victimized.
2533 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2534 * MFU's resident set is consuming more space than it has been allotted. In
2535 * this situation, we must victimize our own cache, the MFU, for this insertion.
2538 arc_get_data_buf(arc_buf_t *buf)
2540 arc_state_t *state = buf->b_hdr->b_state;
2541 uint64_t size = buf->b_hdr->b_size;
2542 arc_buf_contents_t type = buf->b_hdr->b_type;
2544 arc_adapt(size, state);
2547 * We have not yet reached cache maximum size,
2548 * just allocate a new buffer.
2550 if (!arc_evict_needed(type)) {
2551 if (type == ARC_BUFC_METADATA) {
2552 buf->b_data = zio_buf_alloc(size);
2553 arc_space_consume(size, ARC_SPACE_DATA);
2555 ASSERT(type == ARC_BUFC_DATA);
2556 buf->b_data = zio_data_buf_alloc(size);
2557 ARCSTAT_INCR(arcstat_data_size, size);
2558 atomic_add_64(&arc_size, size);
2564 * If we are prefetching from the mfu ghost list, this buffer
2565 * will end up on the mru list; so steal space from there.
2567 if (state == arc_mfu_ghost)
2568 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2569 else if (state == arc_mru_ghost)
2572 if (state == arc_mru || state == arc_anon) {
2573 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2574 state = (arc_mfu->arcs_lsize[type] >= size &&
2575 arc_p > mru_used) ? arc_mfu : arc_mru;
2578 uint64_t mfu_space = arc_c - arc_p;
2579 state = (arc_mru->arcs_lsize[type] >= size &&
2580 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2583 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2584 if (type == ARC_BUFC_METADATA) {
2585 buf->b_data = zio_buf_alloc(size);
2586 arc_space_consume(size, ARC_SPACE_DATA);
2589 * If we are unable to recycle an existing meta buffer
2590 * signal the reclaim thread. It will notify users
2591 * via the prune callback to drop references. The
2592 * prune callback in run in the context of the reclaim
2593 * thread to avoid deadlocking on the hash_lock.
2595 cv_signal(&arc_reclaim_thr_cv);
2597 ASSERT(type == ARC_BUFC_DATA);
2598 buf->b_data = zio_data_buf_alloc(size);
2599 ARCSTAT_INCR(arcstat_data_size, size);
2600 atomic_add_64(&arc_size, size);
2603 ARCSTAT_BUMP(arcstat_recycle_miss);
2605 ASSERT(buf->b_data != NULL);
2608 * Update the state size. Note that ghost states have a
2609 * "ghost size" and so don't need to be updated.
2611 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2612 arc_buf_hdr_t *hdr = buf->b_hdr;
2614 atomic_add_64(&hdr->b_state->arcs_size, size);
2615 if (list_link_active(&hdr->b_arc_node)) {
2616 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2617 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2620 * If we are growing the cache, and we are adding anonymous
2621 * data, and we have outgrown arc_p, update arc_p
2623 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2624 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2625 arc_p = MIN(arc_c, arc_p + size);
2630 * This routine is called whenever a buffer is accessed.
2631 * NOTE: the hash lock is dropped in this function.
2634 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2638 ASSERT(MUTEX_HELD(hash_lock));
2640 if (buf->b_state == arc_anon) {
2642 * This buffer is not in the cache, and does not
2643 * appear in our "ghost" list. Add the new buffer
2647 ASSERT(buf->b_arc_access == 0);
2648 buf->b_arc_access = ddi_get_lbolt();
2649 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2650 arc_change_state(arc_mru, buf, hash_lock);
2652 } else if (buf->b_state == arc_mru) {
2653 now = ddi_get_lbolt();
2656 * If this buffer is here because of a prefetch, then either:
2657 * - clear the flag if this is a "referencing" read
2658 * (any subsequent access will bump this into the MFU state).
2660 * - move the buffer to the head of the list if this is
2661 * another prefetch (to make it less likely to be evicted).
2663 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2664 if (refcount_count(&buf->b_refcnt) == 0) {
2665 ASSERT(list_link_active(&buf->b_arc_node));
2667 buf->b_flags &= ~ARC_PREFETCH;
2668 ARCSTAT_BUMP(arcstat_mru_hits);
2670 buf->b_arc_access = now;
2675 * This buffer has been "accessed" only once so far,
2676 * but it is still in the cache. Move it to the MFU
2679 if (now > buf->b_arc_access + ARC_MINTIME) {
2681 * More than 125ms have passed since we
2682 * instantiated this buffer. Move it to the
2683 * most frequently used state.
2685 buf->b_arc_access = now;
2686 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2687 arc_change_state(arc_mfu, buf, hash_lock);
2689 ARCSTAT_BUMP(arcstat_mru_hits);
2690 } else if (buf->b_state == arc_mru_ghost) {
2691 arc_state_t *new_state;
2693 * This buffer has been "accessed" recently, but
2694 * was evicted from the cache. Move it to the
2698 if (buf->b_flags & ARC_PREFETCH) {
2699 new_state = arc_mru;
2700 if (refcount_count(&buf->b_refcnt) > 0)
2701 buf->b_flags &= ~ARC_PREFETCH;
2702 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2704 new_state = arc_mfu;
2705 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2708 buf->b_arc_access = ddi_get_lbolt();
2709 arc_change_state(new_state, buf, hash_lock);
2711 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2712 } else if (buf->b_state == arc_mfu) {
2714 * This buffer has been accessed more than once and is
2715 * still in the cache. Keep it in the MFU state.
2717 * NOTE: an add_reference() that occurred when we did
2718 * the arc_read() will have kicked this off the list.
2719 * If it was a prefetch, we will explicitly move it to
2720 * the head of the list now.
2722 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2723 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2724 ASSERT(list_link_active(&buf->b_arc_node));
2726 ARCSTAT_BUMP(arcstat_mfu_hits);
2727 buf->b_arc_access = ddi_get_lbolt();
2728 } else if (buf->b_state == arc_mfu_ghost) {
2729 arc_state_t *new_state = arc_mfu;
2731 * This buffer has been accessed more than once but has
2732 * been evicted from the cache. Move it back to the
2736 if (buf->b_flags & ARC_PREFETCH) {
2738 * This is a prefetch access...
2739 * move this block back to the MRU state.
2741 ASSERT0(refcount_count(&buf->b_refcnt));
2742 new_state = arc_mru;
2745 buf->b_arc_access = ddi_get_lbolt();
2746 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2747 arc_change_state(new_state, buf, hash_lock);
2749 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2750 } else if (buf->b_state == arc_l2c_only) {
2752 * This buffer is on the 2nd Level ARC.
2755 buf->b_arc_access = ddi_get_lbolt();
2756 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2757 arc_change_state(arc_mfu, buf, hash_lock);
2759 ASSERT(!"invalid arc state");
2763 /* a generic arc_done_func_t which you can use */
2766 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2768 if (zio == NULL || zio->io_error == 0)
2769 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2770 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2773 /* a generic arc_done_func_t */
2775 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2777 arc_buf_t **bufp = arg;
2778 if (zio && zio->io_error) {
2779 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2783 ASSERT(buf->b_data);
2788 arc_read_done(zio_t *zio)
2790 arc_buf_hdr_t *hdr, *found;
2792 arc_buf_t *abuf; /* buffer we're assigning to callback */
2793 kmutex_t *hash_lock;
2794 arc_callback_t *callback_list, *acb;
2795 int freeable = FALSE;
2797 buf = zio->io_private;
2801 * The hdr was inserted into hash-table and removed from lists
2802 * prior to starting I/O. We should find this header, since
2803 * it's in the hash table, and it should be legit since it's
2804 * not possible to evict it during the I/O. The only possible
2805 * reason for it not to be found is if we were freed during the
2808 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2811 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2812 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2813 (found == hdr && HDR_L2_READING(hdr)));
2815 hdr->b_flags &= ~ARC_L2_EVICTED;
2816 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2817 hdr->b_flags &= ~ARC_L2CACHE;
2819 /* byteswap if necessary */
2820 callback_list = hdr->b_acb;
2821 ASSERT(callback_list != NULL);
2822 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2823 dmu_object_byteswap_t bswap =
2824 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2825 if (BP_GET_LEVEL(zio->io_bp) > 0)
2826 byteswap_uint64_array(buf->b_data, hdr->b_size);
2828 dmu_ot_byteswap[bswap].ob_func(buf->b_data, hdr->b_size);
2831 arc_cksum_compute(buf, B_FALSE);
2833 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2835 * Only call arc_access on anonymous buffers. This is because
2836 * if we've issued an I/O for an evicted buffer, we've already
2837 * called arc_access (to prevent any simultaneous readers from
2838 * getting confused).
2840 arc_access(hdr, hash_lock);
2843 /* create copies of the data buffer for the callers */
2845 for (acb = callback_list; acb; acb = acb->acb_next) {
2846 if (acb->acb_done) {
2848 ARCSTAT_BUMP(arcstat_duplicate_reads);
2849 abuf = arc_buf_clone(buf);
2851 acb->acb_buf = abuf;
2856 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2857 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2859 ASSERT(buf->b_efunc == NULL);
2860 ASSERT(hdr->b_datacnt == 1);
2861 hdr->b_flags |= ARC_BUF_AVAILABLE;
2864 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2866 if (zio->io_error != 0) {
2867 hdr->b_flags |= ARC_IO_ERROR;
2868 if (hdr->b_state != arc_anon)
2869 arc_change_state(arc_anon, hdr, hash_lock);
2870 if (HDR_IN_HASH_TABLE(hdr))
2871 buf_hash_remove(hdr);
2872 freeable = refcount_is_zero(&hdr->b_refcnt);
2876 * Broadcast before we drop the hash_lock to avoid the possibility
2877 * that the hdr (and hence the cv) might be freed before we get to
2878 * the cv_broadcast().
2880 cv_broadcast(&hdr->b_cv);
2883 mutex_exit(hash_lock);
2886 * This block was freed while we waited for the read to
2887 * complete. It has been removed from the hash table and
2888 * moved to the anonymous state (so that it won't show up
2891 ASSERT3P(hdr->b_state, ==, arc_anon);
2892 freeable = refcount_is_zero(&hdr->b_refcnt);
2895 /* execute each callback and free its structure */
2896 while ((acb = callback_list) != NULL) {
2898 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2900 if (acb->acb_zio_dummy != NULL) {
2901 acb->acb_zio_dummy->io_error = zio->io_error;
2902 zio_nowait(acb->acb_zio_dummy);
2905 callback_list = acb->acb_next;
2906 kmem_free(acb, sizeof (arc_callback_t));
2910 arc_hdr_destroy(hdr);
2914 * "Read" the block at the specified DVA (in bp) via the
2915 * cache. If the block is found in the cache, invoke the provided
2916 * callback immediately and return. Note that the `zio' parameter
2917 * in the callback will be NULL in this case, since no IO was
2918 * required. If the block is not in the cache pass the read request
2919 * on to the spa with a substitute callback function, so that the
2920 * requested block will be added to the cache.
2922 * If a read request arrives for a block that has a read in-progress,
2923 * either wait for the in-progress read to complete (and return the
2924 * results); or, if this is a read with a "done" func, add a record
2925 * to the read to invoke the "done" func when the read completes,
2926 * and return; or just return.
2928 * arc_read_done() will invoke all the requested "done" functions
2929 * for readers of this block.
2932 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
2933 void *private, int priority, int zio_flags, uint32_t *arc_flags,
2934 const zbookmark_t *zb)
2937 arc_buf_t *buf = NULL;
2938 kmutex_t *hash_lock;
2940 uint64_t guid = spa_load_guid(spa);
2943 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2945 if (hdr && hdr->b_datacnt > 0) {
2947 *arc_flags |= ARC_CACHED;
2949 if (HDR_IO_IN_PROGRESS(hdr)) {
2951 if (*arc_flags & ARC_WAIT) {
2952 cv_wait(&hdr->b_cv, hash_lock);
2953 mutex_exit(hash_lock);
2956 ASSERT(*arc_flags & ARC_NOWAIT);
2959 arc_callback_t *acb = NULL;
2961 acb = kmem_zalloc(sizeof (arc_callback_t),
2963 acb->acb_done = done;
2964 acb->acb_private = private;
2966 acb->acb_zio_dummy = zio_null(pio,
2967 spa, NULL, NULL, NULL, zio_flags);
2969 ASSERT(acb->acb_done != NULL);
2970 acb->acb_next = hdr->b_acb;
2972 add_reference(hdr, hash_lock, private);
2973 mutex_exit(hash_lock);
2976 mutex_exit(hash_lock);
2980 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2983 add_reference(hdr, hash_lock, private);
2985 * If this block is already in use, create a new
2986 * copy of the data so that we will be guaranteed
2987 * that arc_release() will always succeed.
2991 ASSERT(buf->b_data);
2992 if (HDR_BUF_AVAILABLE(hdr)) {
2993 ASSERT(buf->b_efunc == NULL);
2994 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2996 buf = arc_buf_clone(buf);
2999 } else if (*arc_flags & ARC_PREFETCH &&
3000 refcount_count(&hdr->b_refcnt) == 0) {
3001 hdr->b_flags |= ARC_PREFETCH;
3003 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
3004 arc_access(hdr, hash_lock);
3005 if (*arc_flags & ARC_L2CACHE)
3006 hdr->b_flags |= ARC_L2CACHE;
3007 mutex_exit(hash_lock);
3008 ARCSTAT_BUMP(arcstat_hits);
3009 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3010 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3011 data, metadata, hits);
3014 done(NULL, buf, private);
3016 uint64_t size = BP_GET_LSIZE(bp);
3017 arc_callback_t *acb;
3020 boolean_t devw = B_FALSE;
3023 /* this block is not in the cache */
3024 arc_buf_hdr_t *exists;
3025 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
3026 buf = arc_buf_alloc(spa, size, private, type);
3028 hdr->b_dva = *BP_IDENTITY(bp);
3029 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3030 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3031 exists = buf_hash_insert(hdr, &hash_lock);
3033 /* somebody beat us to the hash insert */
3034 mutex_exit(hash_lock);
3035 buf_discard_identity(hdr);
3036 (void) arc_buf_remove_ref(buf, private);
3037 goto top; /* restart the IO request */
3039 /* if this is a prefetch, we don't have a reference */
3040 if (*arc_flags & ARC_PREFETCH) {
3041 (void) remove_reference(hdr, hash_lock,
3043 hdr->b_flags |= ARC_PREFETCH;
3045 if (*arc_flags & ARC_L2CACHE)
3046 hdr->b_flags |= ARC_L2CACHE;
3047 if (BP_GET_LEVEL(bp) > 0)
3048 hdr->b_flags |= ARC_INDIRECT;
3050 /* this block is in the ghost cache */
3051 ASSERT(GHOST_STATE(hdr->b_state));
3052 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3053 ASSERT0(refcount_count(&hdr->b_refcnt));
3054 ASSERT(hdr->b_buf == NULL);
3056 /* if this is a prefetch, we don't have a reference */
3057 if (*arc_flags & ARC_PREFETCH)
3058 hdr->b_flags |= ARC_PREFETCH;
3060 add_reference(hdr, hash_lock, private);
3061 if (*arc_flags & ARC_L2CACHE)
3062 hdr->b_flags |= ARC_L2CACHE;
3063 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3066 buf->b_efunc = NULL;
3067 buf->b_private = NULL;
3070 ASSERT(hdr->b_datacnt == 0);
3072 arc_get_data_buf(buf);
3073 arc_access(hdr, hash_lock);
3076 ASSERT(!GHOST_STATE(hdr->b_state));
3078 acb = kmem_zalloc(sizeof (arc_callback_t), KM_PUSHPAGE);
3079 acb->acb_done = done;
3080 acb->acb_private = private;
3082 ASSERT(hdr->b_acb == NULL);
3084 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3086 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3087 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3088 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3089 addr = hdr->b_l2hdr->b_daddr;
3091 * Lock out device removal.
3093 if (vdev_is_dead(vd) ||
3094 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3098 mutex_exit(hash_lock);
3100 ASSERT3U(hdr->b_size, ==, size);
3101 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3102 uint64_t, size, zbookmark_t *, zb);
3103 ARCSTAT_BUMP(arcstat_misses);
3104 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3105 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3106 data, metadata, misses);
3108 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3110 * Read from the L2ARC if the following are true:
3111 * 1. The L2ARC vdev was previously cached.
3112 * 2. This buffer still has L2ARC metadata.
3113 * 3. This buffer isn't currently writing to the L2ARC.
3114 * 4. The L2ARC entry wasn't evicted, which may
3115 * also have invalidated the vdev.
3116 * 5. This isn't prefetch and l2arc_noprefetch is set.
3118 if (hdr->b_l2hdr != NULL &&
3119 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3120 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3121 l2arc_read_callback_t *cb;
3123 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3124 ARCSTAT_BUMP(arcstat_l2_hits);
3126 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3128 cb->l2rcb_buf = buf;
3129 cb->l2rcb_spa = spa;
3132 cb->l2rcb_flags = zio_flags;
3135 * l2arc read. The SCL_L2ARC lock will be
3136 * released by l2arc_read_done().
3138 rzio = zio_read_phys(pio, vd, addr, size,
3139 buf->b_data, ZIO_CHECKSUM_OFF,
3140 l2arc_read_done, cb, priority, zio_flags |
3141 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
3142 ZIO_FLAG_DONT_PROPAGATE |
3143 ZIO_FLAG_DONT_RETRY, B_FALSE);
3144 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3146 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
3148 if (*arc_flags & ARC_NOWAIT) {
3153 ASSERT(*arc_flags & ARC_WAIT);
3154 if (zio_wait(rzio) == 0)
3157 /* l2arc read error; goto zio_read() */
3159 DTRACE_PROBE1(l2arc__miss,
3160 arc_buf_hdr_t *, hdr);
3161 ARCSTAT_BUMP(arcstat_l2_misses);
3162 if (HDR_L2_WRITING(hdr))
3163 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3164 spa_config_exit(spa, SCL_L2ARC, vd);
3168 spa_config_exit(spa, SCL_L2ARC, vd);
3169 if (l2arc_ndev != 0) {
3170 DTRACE_PROBE1(l2arc__miss,
3171 arc_buf_hdr_t *, hdr);
3172 ARCSTAT_BUMP(arcstat_l2_misses);
3176 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3177 arc_read_done, buf, priority, zio_flags, zb);
3179 if (*arc_flags & ARC_WAIT)
3180 return (zio_wait(rzio));
3182 ASSERT(*arc_flags & ARC_NOWAIT);
3189 arc_add_prune_callback(arc_prune_func_t *func, void *private)
3193 p = kmem_alloc(sizeof(*p), KM_SLEEP);
3195 p->p_private = private;
3196 list_link_init(&p->p_node);
3197 refcount_create(&p->p_refcnt);
3199 mutex_enter(&arc_prune_mtx);
3200 refcount_add(&p->p_refcnt, &arc_prune_list);
3201 list_insert_head(&arc_prune_list, p);
3202 mutex_exit(&arc_prune_mtx);
3208 arc_remove_prune_callback(arc_prune_t *p)
3210 mutex_enter(&arc_prune_mtx);
3211 list_remove(&arc_prune_list, p);
3212 if (refcount_remove(&p->p_refcnt, &arc_prune_list) == 0) {
3213 refcount_destroy(&p->p_refcnt);
3214 kmem_free(p, sizeof (*p));
3216 mutex_exit(&arc_prune_mtx);
3220 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3222 ASSERT(buf->b_hdr != NULL);
3223 ASSERT(buf->b_hdr->b_state != arc_anon);
3224 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3225 ASSERT(buf->b_efunc == NULL);
3226 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3228 buf->b_efunc = func;
3229 buf->b_private = private;
3233 * Notify the arc that a block was freed, and thus will never be used again.
3236 arc_freed(spa_t *spa, const blkptr_t *bp)
3239 kmutex_t *hash_lock;
3240 uint64_t guid = spa_load_guid(spa);
3242 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3246 if (HDR_BUF_AVAILABLE(hdr)) {
3247 arc_buf_t *buf = hdr->b_buf;
3248 add_reference(hdr, hash_lock, FTAG);
3249 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3250 mutex_exit(hash_lock);
3252 arc_release(buf, FTAG);
3253 (void) arc_buf_remove_ref(buf, FTAG);
3255 mutex_exit(hash_lock);
3261 * This is used by the DMU to let the ARC know that a buffer is
3262 * being evicted, so the ARC should clean up. If this arc buf
3263 * is not yet in the evicted state, it will be put there.
3266 arc_buf_evict(arc_buf_t *buf)
3269 kmutex_t *hash_lock;
3272 mutex_enter(&buf->b_evict_lock);
3276 * We are in arc_do_user_evicts().
3278 ASSERT(buf->b_data == NULL);
3279 mutex_exit(&buf->b_evict_lock);
3281 } else if (buf->b_data == NULL) {
3282 arc_buf_t copy = *buf; /* structure assignment */
3284 * We are on the eviction list; process this buffer now
3285 * but let arc_do_user_evicts() do the reaping.
3287 buf->b_efunc = NULL;
3288 mutex_exit(&buf->b_evict_lock);
3289 VERIFY(copy.b_efunc(©) == 0);
3292 hash_lock = HDR_LOCK(hdr);
3293 mutex_enter(hash_lock);
3295 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3297 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3298 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3301 * Pull this buffer off of the hdr
3304 while (*bufp != buf)
3305 bufp = &(*bufp)->b_next;
3306 *bufp = buf->b_next;
3308 ASSERT(buf->b_data != NULL);
3309 arc_buf_destroy(buf, FALSE, FALSE);
3311 if (hdr->b_datacnt == 0) {
3312 arc_state_t *old_state = hdr->b_state;
3313 arc_state_t *evicted_state;
3315 ASSERT(hdr->b_buf == NULL);
3316 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3319 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3321 mutex_enter(&old_state->arcs_mtx);
3322 mutex_enter(&evicted_state->arcs_mtx);
3324 arc_change_state(evicted_state, hdr, hash_lock);
3325 ASSERT(HDR_IN_HASH_TABLE(hdr));
3326 hdr->b_flags |= ARC_IN_HASH_TABLE;
3327 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3329 mutex_exit(&evicted_state->arcs_mtx);
3330 mutex_exit(&old_state->arcs_mtx);
3332 mutex_exit(hash_lock);
3333 mutex_exit(&buf->b_evict_lock);
3335 VERIFY(buf->b_efunc(buf) == 0);
3336 buf->b_efunc = NULL;
3337 buf->b_private = NULL;
3340 kmem_cache_free(buf_cache, buf);
3345 * Release this buffer from the cache. This must be done
3346 * after a read and prior to modifying the buffer contents.
3347 * If the buffer has more than one reference, we must make
3348 * a new hdr for the buffer.
3351 arc_release(arc_buf_t *buf, void *tag)
3354 kmutex_t *hash_lock = NULL;
3355 l2arc_buf_hdr_t *l2hdr;
3356 uint64_t buf_size = 0;
3359 * It would be nice to assert that if it's DMU metadata (level >
3360 * 0 || it's the dnode file), then it must be syncing context.
3361 * But we don't know that information at this level.
3364 mutex_enter(&buf->b_evict_lock);
3367 /* this buffer is not on any list */
3368 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3370 if (hdr->b_state == arc_anon) {
3371 /* this buffer is already released */
3372 ASSERT(buf->b_efunc == NULL);
3374 hash_lock = HDR_LOCK(hdr);
3375 mutex_enter(hash_lock);
3377 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3380 l2hdr = hdr->b_l2hdr;
3382 mutex_enter(&l2arc_buflist_mtx);
3383 hdr->b_l2hdr = NULL;
3384 buf_size = hdr->b_size;
3388 * Do we have more than one buf?
3390 if (hdr->b_datacnt > 1) {
3391 arc_buf_hdr_t *nhdr;
3393 uint64_t blksz = hdr->b_size;
3394 uint64_t spa = hdr->b_spa;
3395 arc_buf_contents_t type = hdr->b_type;
3396 uint32_t flags = hdr->b_flags;
3398 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3400 * Pull the data off of this hdr and attach it to
3401 * a new anonymous hdr.
3403 (void) remove_reference(hdr, hash_lock, tag);
3405 while (*bufp != buf)
3406 bufp = &(*bufp)->b_next;
3407 *bufp = buf->b_next;
3410 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3411 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3412 if (refcount_is_zero(&hdr->b_refcnt)) {
3413 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3414 ASSERT3U(*size, >=, hdr->b_size);
3415 atomic_add_64(size, -hdr->b_size);
3419 * We're releasing a duplicate user data buffer, update
3420 * our statistics accordingly.
3422 if (hdr->b_type == ARC_BUFC_DATA) {
3423 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3424 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3427 hdr->b_datacnt -= 1;
3428 arc_cksum_verify(buf);
3430 mutex_exit(hash_lock);
3432 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3433 nhdr->b_size = blksz;
3435 nhdr->b_type = type;
3437 nhdr->b_state = arc_anon;
3438 nhdr->b_arc_access = 0;
3439 nhdr->b_flags = flags & ARC_L2_WRITING;
3440 nhdr->b_l2hdr = NULL;
3441 nhdr->b_datacnt = 1;
3442 nhdr->b_freeze_cksum = NULL;
3443 (void) refcount_add(&nhdr->b_refcnt, tag);
3445 mutex_exit(&buf->b_evict_lock);
3446 atomic_add_64(&arc_anon->arcs_size, blksz);
3448 mutex_exit(&buf->b_evict_lock);
3449 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3450 ASSERT(!list_link_active(&hdr->b_arc_node));
3451 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3452 if (hdr->b_state != arc_anon)
3453 arc_change_state(arc_anon, hdr, hash_lock);
3454 hdr->b_arc_access = 0;
3456 mutex_exit(hash_lock);
3458 buf_discard_identity(hdr);
3461 buf->b_efunc = NULL;
3462 buf->b_private = NULL;
3465 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3466 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3467 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
3468 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3469 mutex_exit(&l2arc_buflist_mtx);
3474 arc_released(arc_buf_t *buf)
3478 mutex_enter(&buf->b_evict_lock);
3479 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3480 mutex_exit(&buf->b_evict_lock);
3485 arc_has_callback(arc_buf_t *buf)
3489 mutex_enter(&buf->b_evict_lock);
3490 callback = (buf->b_efunc != NULL);
3491 mutex_exit(&buf->b_evict_lock);
3497 arc_referenced(arc_buf_t *buf)
3501 mutex_enter(&buf->b_evict_lock);
3502 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3503 mutex_exit(&buf->b_evict_lock);
3504 return (referenced);
3509 arc_write_ready(zio_t *zio)
3511 arc_write_callback_t *callback = zio->io_private;
3512 arc_buf_t *buf = callback->awcb_buf;
3513 arc_buf_hdr_t *hdr = buf->b_hdr;
3515 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3516 callback->awcb_ready(zio, buf, callback->awcb_private);
3519 * If the IO is already in progress, then this is a re-write
3520 * attempt, so we need to thaw and re-compute the cksum.
3521 * It is the responsibility of the callback to handle the
3522 * accounting for any re-write attempt.
3524 if (HDR_IO_IN_PROGRESS(hdr)) {
3525 mutex_enter(&hdr->b_freeze_lock);
3526 if (hdr->b_freeze_cksum != NULL) {
3527 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3528 hdr->b_freeze_cksum = NULL;
3530 mutex_exit(&hdr->b_freeze_lock);
3532 arc_cksum_compute(buf, B_FALSE);
3533 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3537 arc_write_done(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(hdr->b_acb == NULL);
3545 if (zio->io_error == 0) {
3546 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3547 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3548 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3550 ASSERT(BUF_EMPTY(hdr));
3554 * If the block to be written was all-zero, we may have
3555 * compressed it away. In this case no write was performed
3556 * so there will be no dva/birth/checksum. The buffer must
3557 * therefore remain anonymous (and uncached).
3559 if (!BUF_EMPTY(hdr)) {
3560 arc_buf_hdr_t *exists;
3561 kmutex_t *hash_lock;
3563 ASSERT(zio->io_error == 0);
3565 arc_cksum_verify(buf);
3567 exists = buf_hash_insert(hdr, &hash_lock);
3570 * This can only happen if we overwrite for
3571 * sync-to-convergence, because we remove
3572 * buffers from the hash table when we arc_free().
3574 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3575 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3576 panic("bad overwrite, hdr=%p exists=%p",
3577 (void *)hdr, (void *)exists);
3578 ASSERT(refcount_is_zero(&exists->b_refcnt));
3579 arc_change_state(arc_anon, exists, hash_lock);
3580 mutex_exit(hash_lock);
3581 arc_hdr_destroy(exists);
3582 exists = buf_hash_insert(hdr, &hash_lock);
3583 ASSERT3P(exists, ==, NULL);
3586 ASSERT(hdr->b_datacnt == 1);
3587 ASSERT(hdr->b_state == arc_anon);
3588 ASSERT(BP_GET_DEDUP(zio->io_bp));
3589 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3592 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3593 /* if it's not anon, we are doing a scrub */
3594 if (!exists && hdr->b_state == arc_anon)
3595 arc_access(hdr, hash_lock);
3596 mutex_exit(hash_lock);
3598 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3601 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3602 callback->awcb_done(zio, buf, callback->awcb_private);
3604 kmem_free(callback, sizeof (arc_write_callback_t));
3608 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3609 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3610 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3611 int priority, int zio_flags, const zbookmark_t *zb)
3613 arc_buf_hdr_t *hdr = buf->b_hdr;
3614 arc_write_callback_t *callback;
3617 ASSERT(ready != NULL);
3618 ASSERT(done != NULL);
3619 ASSERT(!HDR_IO_ERROR(hdr));
3620 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3621 ASSERT(hdr->b_acb == NULL);
3623 hdr->b_flags |= ARC_L2CACHE;
3624 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_PUSHPAGE);
3625 callback->awcb_ready = ready;
3626 callback->awcb_done = done;
3627 callback->awcb_private = private;
3628 callback->awcb_buf = buf;
3630 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3631 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3637 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3640 uint64_t available_memory;
3642 if (zfs_arc_memory_throttle_disable)
3645 /* Easily reclaimable memory (free + inactive + arc-evictable) */
3646 available_memory = ptob(spl_kmem_availrmem()) + arc_evictable_memory();
3648 if (available_memory <= zfs_write_limit_max) {
3649 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3650 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
3654 if (inflight_data > available_memory / 4) {
3655 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3656 DMU_TX_STAT_BUMP(dmu_tx_memory_inflight);
3664 arc_tempreserve_clear(uint64_t reserve)
3666 atomic_add_64(&arc_tempreserve, -reserve);
3667 ASSERT((int64_t)arc_tempreserve >= 0);
3671 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3678 * Once in a while, fail for no reason. Everything should cope.
3680 if (spa_get_random(10000) == 0) {
3681 dprintf("forcing random failure\n");
3685 if (reserve > arc_c/4 && !arc_no_grow)
3686 arc_c = MIN(arc_c_max, reserve * 4);
3687 if (reserve > arc_c) {
3688 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
3693 * Don't count loaned bufs as in flight dirty data to prevent long
3694 * network delays from blocking transactions that are ready to be
3695 * assigned to a txg.
3697 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3700 * Writes will, almost always, require additional memory allocations
3701 * in order to compress/encrypt/etc the data. We therefor need to
3702 * make sure that there is sufficient available memory for this.
3704 if ((error = arc_memory_throttle(reserve, anon_size, txg)))
3708 * Throttle writes when the amount of dirty data in the cache
3709 * gets too large. We try to keep the cache less than half full
3710 * of dirty blocks so that our sync times don't grow too large.
3711 * Note: if two requests come in concurrently, we might let them
3712 * both succeed, when one of them should fail. Not a huge deal.
3715 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3716 anon_size > arc_c / 4) {
3717 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3718 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3719 arc_tempreserve>>10,
3720 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3721 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3722 reserve>>10, arc_c>>10);
3723 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
3726 atomic_add_64(&arc_tempreserve, reserve);
3731 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
3732 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
3734 size->value.ui64 = state->arcs_size;
3735 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
3736 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
3740 arc_kstat_update(kstat_t *ksp, int rw)
3742 arc_stats_t *as = ksp->ks_data;
3744 if (rw == KSTAT_WRITE) {
3747 arc_kstat_update_state(arc_anon,
3748 &as->arcstat_anon_size,
3749 &as->arcstat_anon_evict_data,
3750 &as->arcstat_anon_evict_metadata);
3751 arc_kstat_update_state(arc_mru,
3752 &as->arcstat_mru_size,
3753 &as->arcstat_mru_evict_data,
3754 &as->arcstat_mru_evict_metadata);
3755 arc_kstat_update_state(arc_mru_ghost,
3756 &as->arcstat_mru_ghost_size,
3757 &as->arcstat_mru_ghost_evict_data,
3758 &as->arcstat_mru_ghost_evict_metadata);
3759 arc_kstat_update_state(arc_mfu,
3760 &as->arcstat_mfu_size,
3761 &as->arcstat_mfu_evict_data,
3762 &as->arcstat_mfu_evict_metadata);
3763 arc_kstat_update_state(arc_mfu_ghost,
3764 &as->arcstat_mfu_ghost_size,
3765 &as->arcstat_mfu_ghost_evict_data,
3766 &as->arcstat_mfu_ghost_evict_metadata);
3775 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3776 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3778 /* Convert seconds to clock ticks */
3779 zfs_arc_min_prefetch_lifespan = 1 * hz;
3781 /* Start out with 1/8 of all memory */
3782 arc_c = physmem * PAGESIZE / 8;
3786 * On architectures where the physical memory can be larger
3787 * than the addressable space (intel in 32-bit mode), we may
3788 * need to limit the cache to 1/8 of VM size.
3790 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3792 * Register a shrinker to support synchronous (direct) memory
3793 * reclaim from the arc. This is done to prevent kswapd from
3794 * swapping out pages when it is preferable to shrink the arc.
3796 spl_register_shrinker(&arc_shrinker);
3799 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3800 arc_c_min = MAX(arc_c / 4, 64<<20);
3801 /* set max to 1/2 of all memory */
3802 arc_c_max = MAX(arc_c * 4, arc_c_max);
3805 * Allow the tunables to override our calculations if they are
3806 * reasonable (ie. over 64MB)
3808 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3809 arc_c_max = zfs_arc_max;
3810 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3811 arc_c_min = zfs_arc_min;
3814 arc_p = (arc_c >> 1);
3816 /* limit meta-data to 1/4 of the arc capacity */
3817 arc_meta_limit = arc_c_max / 4;
3820 /* Allow the tunable to override if it is reasonable */
3821 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3822 arc_meta_limit = zfs_arc_meta_limit;
3824 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3825 arc_c_min = arc_meta_limit / 2;
3827 /* if kmem_flags are set, lets try to use less memory */
3828 if (kmem_debugging())
3830 if (arc_c < arc_c_min)
3833 arc_anon = &ARC_anon;
3835 arc_mru_ghost = &ARC_mru_ghost;
3837 arc_mfu_ghost = &ARC_mfu_ghost;
3838 arc_l2c_only = &ARC_l2c_only;
3841 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3842 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3843 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3844 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3845 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3846 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3848 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3849 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3850 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3851 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3852 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3853 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3854 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3855 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3856 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3857 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3858 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3859 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3860 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3861 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3862 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3863 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3864 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3865 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3866 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3867 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3871 arc_thread_exit = 0;
3872 list_create(&arc_prune_list, sizeof (arc_prune_t),
3873 offsetof(arc_prune_t, p_node));
3874 arc_eviction_list = NULL;
3875 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
3876 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3877 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3879 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3880 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3882 if (arc_ksp != NULL) {
3883 arc_ksp->ks_data = &arc_stats;
3884 arc_ksp->ks_update = arc_kstat_update;
3885 kstat_install(arc_ksp);
3888 (void) thread_create(NULL, 0, arc_adapt_thread, NULL, 0, &p0,
3889 TS_RUN, minclsyspri);
3894 if (zfs_write_limit_max == 0)
3895 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3897 zfs_write_limit_shift = 0;
3898 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3906 mutex_enter(&arc_reclaim_thr_lock);
3908 spl_unregister_shrinker(&arc_shrinker);
3909 #endif /* _KERNEL */
3911 arc_thread_exit = 1;
3912 while (arc_thread_exit != 0)
3913 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3914 mutex_exit(&arc_reclaim_thr_lock);
3920 if (arc_ksp != NULL) {
3921 kstat_delete(arc_ksp);
3925 mutex_enter(&arc_prune_mtx);
3926 while ((p = list_head(&arc_prune_list)) != NULL) {
3927 list_remove(&arc_prune_list, p);
3928 refcount_remove(&p->p_refcnt, &arc_prune_list);
3929 refcount_destroy(&p->p_refcnt);
3930 kmem_free(p, sizeof (*p));
3932 mutex_exit(&arc_prune_mtx);
3934 list_destroy(&arc_prune_list);
3935 mutex_destroy(&arc_prune_mtx);
3936 mutex_destroy(&arc_eviction_mtx);
3937 mutex_destroy(&arc_reclaim_thr_lock);
3938 cv_destroy(&arc_reclaim_thr_cv);
3940 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3941 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3942 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3943 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3944 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3945 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3946 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3947 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3949 mutex_destroy(&arc_anon->arcs_mtx);
3950 mutex_destroy(&arc_mru->arcs_mtx);
3951 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3952 mutex_destroy(&arc_mfu->arcs_mtx);
3953 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3954 mutex_destroy(&arc_l2c_only->arcs_mtx);
3956 mutex_destroy(&zfs_write_limit_lock);
3960 ASSERT(arc_loaned_bytes == 0);
3966 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3967 * It uses dedicated storage devices to hold cached data, which are populated
3968 * using large infrequent writes. The main role of this cache is to boost
3969 * the performance of random read workloads. The intended L2ARC devices
3970 * include short-stroked disks, solid state disks, and other media with
3971 * substantially faster read latency than disk.
3973 * +-----------------------+
3975 * +-----------------------+
3978 * l2arc_feed_thread() arc_read()
3982 * +---------------+ |
3984 * +---------------+ |
3989 * +-------+ +-------+
3991 * | cache | | cache |
3992 * +-------+ +-------+
3993 * +=========+ .-----.
3994 * : L2ARC : |-_____-|
3995 * : devices : | Disks |
3996 * +=========+ `-_____-'
3998 * Read requests are satisfied from the following sources, in order:
4001 * 2) vdev cache of L2ARC devices
4003 * 4) vdev cache of disks
4006 * Some L2ARC device types exhibit extremely slow write performance.
4007 * To accommodate for this there are some significant differences between
4008 * the L2ARC and traditional cache design:
4010 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4011 * the ARC behave as usual, freeing buffers and placing headers on ghost
4012 * lists. The ARC does not send buffers to the L2ARC during eviction as
4013 * this would add inflated write latencies for all ARC memory pressure.
4015 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4016 * It does this by periodically scanning buffers from the eviction-end of
4017 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4018 * not already there. It scans until a headroom of buffers is satisfied,
4019 * which itself is a buffer for ARC eviction. The thread that does this is
4020 * l2arc_feed_thread(), illustrated below; example sizes are included to
4021 * provide a better sense of ratio than this diagram:
4024 * +---------------------+----------+
4025 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4026 * +---------------------+----------+ | o L2ARC eligible
4027 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4028 * +---------------------+----------+ |
4029 * 15.9 Gbytes ^ 32 Mbytes |
4031 * l2arc_feed_thread()
4033 * l2arc write hand <--[oooo]--'
4037 * +==============================+
4038 * L2ARC dev |####|#|###|###| |####| ... |
4039 * +==============================+
4042 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4043 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4044 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4045 * safe to say that this is an uncommon case, since buffers at the end of
4046 * the ARC lists have moved there due to inactivity.
4048 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4049 * then the L2ARC simply misses copying some buffers. This serves as a
4050 * pressure valve to prevent heavy read workloads from both stalling the ARC
4051 * with waits and clogging the L2ARC with writes. This also helps prevent
4052 * the potential for the L2ARC to churn if it attempts to cache content too
4053 * quickly, such as during backups of the entire pool.
4055 * 5. After system boot and before the ARC has filled main memory, there are
4056 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4057 * lists can remain mostly static. Instead of searching from tail of these
4058 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4059 * for eligible buffers, greatly increasing its chance of finding them.
4061 * The L2ARC device write speed is also boosted during this time so that
4062 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4063 * there are no L2ARC reads, and no fear of degrading read performance
4064 * through increased writes.
4066 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4067 * the vdev queue can aggregate them into larger and fewer writes. Each
4068 * device is written to in a rotor fashion, sweeping writes through
4069 * available space then repeating.
4071 * 7. The L2ARC does not store dirty content. It never needs to flush
4072 * write buffers back to disk based storage.
4074 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4075 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4077 * The performance of the L2ARC can be tweaked by a number of tunables, which
4078 * may be necessary for different workloads:
4080 * l2arc_write_max max write bytes per interval
4081 * l2arc_write_boost extra write bytes during device warmup
4082 * l2arc_noprefetch skip caching prefetched buffers
4083 * l2arc_headroom number of max device writes to precache
4084 * l2arc_feed_secs seconds between L2ARC writing
4086 * Tunables may be removed or added as future performance improvements are
4087 * integrated, and also may become zpool properties.
4089 * There are three key functions that control how the L2ARC warms up:
4091 * l2arc_write_eligible() check if a buffer is eligible to cache
4092 * l2arc_write_size() calculate how much to write
4093 * l2arc_write_interval() calculate sleep delay between writes
4095 * These three functions determine what to write, how much, and how quickly
4100 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4103 * A buffer is *not* eligible for the L2ARC if it:
4104 * 1. belongs to a different spa.
4105 * 2. is already cached on the L2ARC.
4106 * 3. has an I/O in progress (it may be an incomplete read).
4107 * 4. is flagged not eligible (zfs property).
4109 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
4110 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
4117 l2arc_write_size(l2arc_dev_t *dev)
4121 size = dev->l2ad_write;
4123 if (arc_warm == B_FALSE)
4124 size += dev->l2ad_boost;
4131 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4133 clock_t interval, next, now;
4136 * If the ARC lists are busy, increase our write rate; if the
4137 * lists are stale, idle back. This is achieved by checking
4138 * how much we previously wrote - if it was more than half of
4139 * what we wanted, schedule the next write much sooner.
4141 if (l2arc_feed_again && wrote > (wanted / 2))
4142 interval = (hz * l2arc_feed_min_ms) / 1000;
4144 interval = hz * l2arc_feed_secs;
4146 now = ddi_get_lbolt();
4147 next = MAX(now, MIN(now + interval, began + interval));
4153 l2arc_hdr_stat_add(void)
4155 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE);
4156 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4160 l2arc_hdr_stat_remove(void)
4162 ARCSTAT_INCR(arcstat_l2_hdr_size, -HDR_SIZE);
4163 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4167 * Cycle through L2ARC devices. This is how L2ARC load balances.
4168 * If a device is returned, this also returns holding the spa config lock.
4170 static l2arc_dev_t *
4171 l2arc_dev_get_next(void)
4173 l2arc_dev_t *first, *next = NULL;
4176 * Lock out the removal of spas (spa_namespace_lock), then removal
4177 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4178 * both locks will be dropped and a spa config lock held instead.
4180 mutex_enter(&spa_namespace_lock);
4181 mutex_enter(&l2arc_dev_mtx);
4183 /* if there are no vdevs, there is nothing to do */
4184 if (l2arc_ndev == 0)
4188 next = l2arc_dev_last;
4190 /* loop around the list looking for a non-faulted vdev */
4192 next = list_head(l2arc_dev_list);
4194 next = list_next(l2arc_dev_list, next);
4196 next = list_head(l2arc_dev_list);
4199 /* if we have come back to the start, bail out */
4202 else if (next == first)
4205 } while (vdev_is_dead(next->l2ad_vdev));
4207 /* if we were unable to find any usable vdevs, return NULL */
4208 if (vdev_is_dead(next->l2ad_vdev))
4211 l2arc_dev_last = next;
4214 mutex_exit(&l2arc_dev_mtx);
4217 * Grab the config lock to prevent the 'next' device from being
4218 * removed while we are writing to it.
4221 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4222 mutex_exit(&spa_namespace_lock);
4228 * Free buffers that were tagged for destruction.
4231 l2arc_do_free_on_write(void)
4234 l2arc_data_free_t *df, *df_prev;
4236 mutex_enter(&l2arc_free_on_write_mtx);
4237 buflist = l2arc_free_on_write;
4239 for (df = list_tail(buflist); df; df = df_prev) {
4240 df_prev = list_prev(buflist, df);
4241 ASSERT(df->l2df_data != NULL);
4242 ASSERT(df->l2df_func != NULL);
4243 df->l2df_func(df->l2df_data, df->l2df_size);
4244 list_remove(buflist, df);
4245 kmem_free(df, sizeof (l2arc_data_free_t));
4248 mutex_exit(&l2arc_free_on_write_mtx);
4252 * A write to a cache device has completed. Update all headers to allow
4253 * reads from these buffers to begin.
4256 l2arc_write_done(zio_t *zio)
4258 l2arc_write_callback_t *cb;
4261 arc_buf_hdr_t *head, *ab, *ab_prev;
4262 l2arc_buf_hdr_t *abl2;
4263 kmutex_t *hash_lock;
4265 cb = zio->io_private;
4267 dev = cb->l2wcb_dev;
4268 ASSERT(dev != NULL);
4269 head = cb->l2wcb_head;
4270 ASSERT(head != NULL);
4271 buflist = dev->l2ad_buflist;
4272 ASSERT(buflist != NULL);
4273 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4274 l2arc_write_callback_t *, cb);
4276 if (zio->io_error != 0)
4277 ARCSTAT_BUMP(arcstat_l2_writes_error);
4279 mutex_enter(&l2arc_buflist_mtx);
4282 * All writes completed, or an error was hit.
4284 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4285 ab_prev = list_prev(buflist, ab);
4287 hash_lock = HDR_LOCK(ab);
4288 if (!mutex_tryenter(hash_lock)) {
4290 * This buffer misses out. It may be in a stage
4291 * of eviction. Its ARC_L2_WRITING flag will be
4292 * left set, denying reads to this buffer.
4294 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4298 if (zio->io_error != 0) {
4300 * Error - drop L2ARC entry.
4302 list_remove(buflist, ab);
4305 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4306 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4307 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4311 * Allow ARC to begin reads to this L2ARC entry.
4313 ab->b_flags &= ~ARC_L2_WRITING;
4315 mutex_exit(hash_lock);
4318 atomic_inc_64(&l2arc_writes_done);
4319 list_remove(buflist, head);
4320 kmem_cache_free(hdr_cache, head);
4321 mutex_exit(&l2arc_buflist_mtx);
4323 l2arc_do_free_on_write();
4325 kmem_free(cb, sizeof (l2arc_write_callback_t));
4329 * A read to a cache device completed. Validate buffer contents before
4330 * handing over to the regular ARC routines.
4333 l2arc_read_done(zio_t *zio)
4335 l2arc_read_callback_t *cb;
4338 kmutex_t *hash_lock;
4341 ASSERT(zio->io_vd != NULL);
4342 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4344 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4346 cb = zio->io_private;
4348 buf = cb->l2rcb_buf;
4349 ASSERT(buf != NULL);
4351 hash_lock = HDR_LOCK(buf->b_hdr);
4352 mutex_enter(hash_lock);
4354 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4357 * Check this survived the L2ARC journey.
4359 equal = arc_cksum_equal(buf);
4360 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4361 mutex_exit(hash_lock);
4362 zio->io_private = buf;
4363 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4364 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4367 mutex_exit(hash_lock);
4369 * Buffer didn't survive caching. Increment stats and
4370 * reissue to the original storage device.
4372 if (zio->io_error != 0) {
4373 ARCSTAT_BUMP(arcstat_l2_io_error);
4375 zio->io_error = EIO;
4378 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4381 * If there's no waiter, issue an async i/o to the primary
4382 * storage now. If there *is* a waiter, the caller must
4383 * issue the i/o in a context where it's OK to block.
4385 if (zio->io_waiter == NULL) {
4386 zio_t *pio = zio_unique_parent(zio);
4388 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4390 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4391 buf->b_data, zio->io_size, arc_read_done, buf,
4392 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4396 kmem_free(cb, sizeof (l2arc_read_callback_t));
4400 * This is the list priority from which the L2ARC will search for pages to
4401 * cache. This is used within loops (0..3) to cycle through lists in the
4402 * desired order. This order can have a significant effect on cache
4405 * Currently the metadata lists are hit first, MFU then MRU, followed by
4406 * the data lists. This function returns a locked list, and also returns
4410 l2arc_list_locked(int list_num, kmutex_t **lock)
4412 list_t *list = NULL;
4414 ASSERT(list_num >= 0 && list_num <= 3);
4418 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4419 *lock = &arc_mfu->arcs_mtx;
4422 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4423 *lock = &arc_mru->arcs_mtx;
4426 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4427 *lock = &arc_mfu->arcs_mtx;
4430 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4431 *lock = &arc_mru->arcs_mtx;
4435 ASSERT(!(MUTEX_HELD(*lock)));
4441 * Evict buffers from the device write hand to the distance specified in
4442 * bytes. This distance may span populated buffers, it may span nothing.
4443 * This is clearing a region on the L2ARC device ready for writing.
4444 * If the 'all' boolean is set, every buffer is evicted.
4447 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4450 l2arc_buf_hdr_t *abl2;
4451 arc_buf_hdr_t *ab, *ab_prev;
4452 kmutex_t *hash_lock;
4455 buflist = dev->l2ad_buflist;
4457 if (buflist == NULL)
4460 if (!all && dev->l2ad_first) {
4462 * This is the first sweep through the device. There is
4468 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4470 * When nearing the end of the device, evict to the end
4471 * before the device write hand jumps to the start.
4473 taddr = dev->l2ad_end;
4475 taddr = dev->l2ad_hand + distance;
4477 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4478 uint64_t, taddr, boolean_t, all);
4481 mutex_enter(&l2arc_buflist_mtx);
4482 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4483 ab_prev = list_prev(buflist, ab);
4485 hash_lock = HDR_LOCK(ab);
4486 if (!mutex_tryenter(hash_lock)) {
4488 * Missed the hash lock. Retry.
4490 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4491 mutex_exit(&l2arc_buflist_mtx);
4492 mutex_enter(hash_lock);
4493 mutex_exit(hash_lock);
4497 if (HDR_L2_WRITE_HEAD(ab)) {
4499 * We hit a write head node. Leave it for
4500 * l2arc_write_done().
4502 list_remove(buflist, ab);
4503 mutex_exit(hash_lock);
4507 if (!all && ab->b_l2hdr != NULL &&
4508 (ab->b_l2hdr->b_daddr > taddr ||
4509 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4511 * We've evicted to the target address,
4512 * or the end of the device.
4514 mutex_exit(hash_lock);
4518 if (HDR_FREE_IN_PROGRESS(ab)) {
4520 * Already on the path to destruction.
4522 mutex_exit(hash_lock);
4526 if (ab->b_state == arc_l2c_only) {
4527 ASSERT(!HDR_L2_READING(ab));
4529 * This doesn't exist in the ARC. Destroy.
4530 * arc_hdr_destroy() will call list_remove()
4531 * and decrement arcstat_l2_size.
4533 arc_change_state(arc_anon, ab, hash_lock);
4534 arc_hdr_destroy(ab);
4537 * Invalidate issued or about to be issued
4538 * reads, since we may be about to write
4539 * over this location.
4541 if (HDR_L2_READING(ab)) {
4542 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4543 ab->b_flags |= ARC_L2_EVICTED;
4547 * Tell ARC this no longer exists in L2ARC.
4549 if (ab->b_l2hdr != NULL) {
4552 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4553 arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4554 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4556 list_remove(buflist, ab);
4559 * This may have been leftover after a
4562 ab->b_flags &= ~ARC_L2_WRITING;
4564 mutex_exit(hash_lock);
4566 mutex_exit(&l2arc_buflist_mtx);
4568 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4569 dev->l2ad_evict = taddr;
4573 * Find and write ARC buffers to the L2ARC device.
4575 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4576 * for reading until they have completed writing.
4579 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4581 arc_buf_hdr_t *ab, *ab_prev, *head;
4582 l2arc_buf_hdr_t *hdrl2;
4584 uint64_t passed_sz, write_sz, buf_sz, headroom;
4586 kmutex_t *hash_lock, *list_lock = NULL;
4587 boolean_t have_lock, full;
4588 l2arc_write_callback_t *cb;
4590 uint64_t guid = spa_load_guid(spa);
4593 ASSERT(dev->l2ad_vdev != NULL);
4598 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4599 head->b_flags |= ARC_L2_WRITE_HEAD;
4602 * Copy buffers for L2ARC writing.
4604 mutex_enter(&l2arc_buflist_mtx);
4605 for (try = 0; try <= 3; try++) {
4606 list = l2arc_list_locked(try, &list_lock);
4610 * L2ARC fast warmup.
4612 * Until the ARC is warm and starts to evict, read from the
4613 * head of the ARC lists rather than the tail.
4615 headroom = target_sz * l2arc_headroom;
4616 if (arc_warm == B_FALSE)
4617 ab = list_head(list);
4619 ab = list_tail(list);
4621 for (; ab; ab = ab_prev) {
4622 if (arc_warm == B_FALSE)
4623 ab_prev = list_next(list, ab);
4625 ab_prev = list_prev(list, ab);
4627 hash_lock = HDR_LOCK(ab);
4628 have_lock = MUTEX_HELD(hash_lock);
4629 if (!have_lock && !mutex_tryenter(hash_lock)) {
4631 * Skip this buffer rather than waiting.
4636 passed_sz += ab->b_size;
4637 if (passed_sz > headroom) {
4641 mutex_exit(hash_lock);
4645 if (!l2arc_write_eligible(guid, ab)) {
4646 mutex_exit(hash_lock);
4650 if ((write_sz + ab->b_size) > target_sz) {
4652 mutex_exit(hash_lock);
4658 * Insert a dummy header on the buflist so
4659 * l2arc_write_done() can find where the
4660 * write buffers begin without searching.
4662 list_insert_head(dev->l2ad_buflist, head);
4664 cb = kmem_alloc(sizeof (l2arc_write_callback_t),
4666 cb->l2wcb_dev = dev;
4667 cb->l2wcb_head = head;
4668 pio = zio_root(spa, l2arc_write_done, cb,
4673 * Create and add a new L2ARC header.
4675 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t),
4678 hdrl2->b_daddr = dev->l2ad_hand;
4679 arc_space_consume(L2HDR_SIZE, ARC_SPACE_L2HDRS);
4681 ab->b_flags |= ARC_L2_WRITING;
4682 ab->b_l2hdr = hdrl2;
4683 list_insert_head(dev->l2ad_buflist, ab);
4684 buf_data = ab->b_buf->b_data;
4685 buf_sz = ab->b_size;
4688 * Compute and store the buffer cksum before
4689 * writing. On debug the cksum is verified first.
4691 arc_cksum_verify(ab->b_buf);
4692 arc_cksum_compute(ab->b_buf, B_TRUE);
4694 mutex_exit(hash_lock);
4696 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4697 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4698 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4699 ZIO_FLAG_CANFAIL, B_FALSE);
4701 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4703 (void) zio_nowait(wzio);
4706 * Keep the clock hand suitably device-aligned.
4708 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4711 dev->l2ad_hand += buf_sz;
4714 mutex_exit(list_lock);
4719 mutex_exit(&l2arc_buflist_mtx);
4723 kmem_cache_free(hdr_cache, head);
4727 ASSERT3U(write_sz, <=, target_sz);
4728 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4729 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4730 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4731 vdev_space_update(dev->l2ad_vdev, write_sz, 0, 0);
4734 * Bump device hand to the device start if it is approaching the end.
4735 * l2arc_evict() will already have evicted ahead for this case.
4737 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4738 vdev_space_update(dev->l2ad_vdev,
4739 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4740 dev->l2ad_hand = dev->l2ad_start;
4741 dev->l2ad_evict = dev->l2ad_start;
4742 dev->l2ad_first = B_FALSE;
4745 dev->l2ad_writing = B_TRUE;
4746 (void) zio_wait(pio);
4747 dev->l2ad_writing = B_FALSE;
4753 * This thread feeds the L2ARC at regular intervals. This is the beating
4754 * heart of the L2ARC.
4757 l2arc_feed_thread(void)
4762 uint64_t size, wrote;
4763 clock_t begin, next = ddi_get_lbolt();
4765 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4767 mutex_enter(&l2arc_feed_thr_lock);
4769 while (l2arc_thread_exit == 0) {
4770 CALLB_CPR_SAFE_BEGIN(&cpr);
4771 (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv,
4772 &l2arc_feed_thr_lock, next);
4773 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4774 next = ddi_get_lbolt() + hz;
4777 * Quick check for L2ARC devices.
4779 mutex_enter(&l2arc_dev_mtx);
4780 if (l2arc_ndev == 0) {
4781 mutex_exit(&l2arc_dev_mtx);
4784 mutex_exit(&l2arc_dev_mtx);
4785 begin = ddi_get_lbolt();
4788 * This selects the next l2arc device to write to, and in
4789 * doing so the next spa to feed from: dev->l2ad_spa. This
4790 * will return NULL if there are now no l2arc devices or if
4791 * they are all faulted.
4793 * If a device is returned, its spa's config lock is also
4794 * held to prevent device removal. l2arc_dev_get_next()
4795 * will grab and release l2arc_dev_mtx.
4797 if ((dev = l2arc_dev_get_next()) == NULL)
4800 spa = dev->l2ad_spa;
4801 ASSERT(spa != NULL);
4804 * If the pool is read-only then force the feed thread to
4805 * sleep a little longer.
4807 if (!spa_writeable(spa)) {
4808 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4809 spa_config_exit(spa, SCL_L2ARC, dev);
4814 * Avoid contributing to memory pressure.
4817 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4818 spa_config_exit(spa, SCL_L2ARC, dev);
4822 ARCSTAT_BUMP(arcstat_l2_feeds);
4824 size = l2arc_write_size(dev);
4827 * Evict L2ARC buffers that will be overwritten.
4829 l2arc_evict(dev, size, B_FALSE);
4832 * Write ARC buffers.
4834 wrote = l2arc_write_buffers(spa, dev, size);
4837 * Calculate interval between writes.
4839 next = l2arc_write_interval(begin, size, wrote);
4840 spa_config_exit(spa, SCL_L2ARC, dev);
4843 l2arc_thread_exit = 0;
4844 cv_broadcast(&l2arc_feed_thr_cv);
4845 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4850 l2arc_vdev_present(vdev_t *vd)
4854 mutex_enter(&l2arc_dev_mtx);
4855 for (dev = list_head(l2arc_dev_list); dev != NULL;
4856 dev = list_next(l2arc_dev_list, dev)) {
4857 if (dev->l2ad_vdev == vd)
4860 mutex_exit(&l2arc_dev_mtx);
4862 return (dev != NULL);
4866 * Add a vdev for use by the L2ARC. By this point the spa has already
4867 * validated the vdev and opened it.
4870 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
4872 l2arc_dev_t *adddev;
4874 ASSERT(!l2arc_vdev_present(vd));
4877 * Create a new l2arc device entry.
4879 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4880 adddev->l2ad_spa = spa;
4881 adddev->l2ad_vdev = vd;
4882 adddev->l2ad_write = l2arc_write_max;
4883 adddev->l2ad_boost = l2arc_write_boost;
4884 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
4885 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
4886 adddev->l2ad_hand = adddev->l2ad_start;
4887 adddev->l2ad_evict = adddev->l2ad_start;
4888 adddev->l2ad_first = B_TRUE;
4889 adddev->l2ad_writing = B_FALSE;
4890 list_link_init(&adddev->l2ad_node);
4891 ASSERT3U(adddev->l2ad_write, >, 0);
4894 * This is a list of all ARC buffers that are still valid on the
4897 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4898 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4899 offsetof(arc_buf_hdr_t, b_l2node));
4901 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
4904 * Add device to global list
4906 mutex_enter(&l2arc_dev_mtx);
4907 list_insert_head(l2arc_dev_list, adddev);
4908 atomic_inc_64(&l2arc_ndev);
4909 mutex_exit(&l2arc_dev_mtx);
4913 * Remove a vdev from the L2ARC.
4916 l2arc_remove_vdev(vdev_t *vd)
4918 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4921 * Find the device by vdev
4923 mutex_enter(&l2arc_dev_mtx);
4924 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4925 nextdev = list_next(l2arc_dev_list, dev);
4926 if (vd == dev->l2ad_vdev) {
4931 ASSERT(remdev != NULL);
4934 * Remove device from global list
4936 list_remove(l2arc_dev_list, remdev);
4937 l2arc_dev_last = NULL; /* may have been invalidated */
4938 atomic_dec_64(&l2arc_ndev);
4939 mutex_exit(&l2arc_dev_mtx);
4942 * Clear all buflists and ARC references. L2ARC device flush.
4944 l2arc_evict(remdev, 0, B_TRUE);
4945 list_destroy(remdev->l2ad_buflist);
4946 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
4947 kmem_free(remdev, sizeof (l2arc_dev_t));
4953 l2arc_thread_exit = 0;
4955 l2arc_writes_sent = 0;
4956 l2arc_writes_done = 0;
4958 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4959 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
4960 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
4961 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
4962 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
4964 l2arc_dev_list = &L2ARC_dev_list;
4965 l2arc_free_on_write = &L2ARC_free_on_write;
4966 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
4967 offsetof(l2arc_dev_t, l2ad_node));
4968 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
4969 offsetof(l2arc_data_free_t, l2df_list_node));
4976 * This is called from dmu_fini(), which is called from spa_fini();
4977 * Because of this, we can assume that all l2arc devices have
4978 * already been removed when the pools themselves were removed.
4981 l2arc_do_free_on_write();
4983 mutex_destroy(&l2arc_feed_thr_lock);
4984 cv_destroy(&l2arc_feed_thr_cv);
4985 mutex_destroy(&l2arc_dev_mtx);
4986 mutex_destroy(&l2arc_buflist_mtx);
4987 mutex_destroy(&l2arc_free_on_write_mtx);
4989 list_destroy(l2arc_dev_list);
4990 list_destroy(l2arc_free_on_write);
4996 if (!(spa_mode_global & FWRITE))
4999 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5000 TS_RUN, minclsyspri);
5006 if (!(spa_mode_global & FWRITE))
5009 mutex_enter(&l2arc_feed_thr_lock);
5010 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5011 l2arc_thread_exit = 1;
5012 while (l2arc_thread_exit != 0)
5013 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5014 mutex_exit(&l2arc_feed_thr_lock);
5017 #if defined(_KERNEL) && defined(HAVE_SPL)
5018 EXPORT_SYMBOL(arc_read);
5019 EXPORT_SYMBOL(arc_buf_remove_ref);
5020 EXPORT_SYMBOL(arc_getbuf_func);
5021 EXPORT_SYMBOL(arc_add_prune_callback);
5022 EXPORT_SYMBOL(arc_remove_prune_callback);
5024 module_param(zfs_arc_min, ulong, 0644);
5025 MODULE_PARM_DESC(zfs_arc_min, "Min arc size");
5027 module_param(zfs_arc_max, ulong, 0644);
5028 MODULE_PARM_DESC(zfs_arc_max, "Max arc size");
5030 module_param(zfs_arc_meta_limit, ulong, 0644);
5031 MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size");
5033 module_param(zfs_arc_meta_prune, int, 0644);
5034 MODULE_PARM_DESC(zfs_arc_meta_prune, "Bytes of meta data to prune");
5036 module_param(zfs_arc_grow_retry, int, 0644);
5037 MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size");
5039 module_param(zfs_arc_shrink_shift, int, 0644);
5040 MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)");
5042 module_param(zfs_arc_p_min_shift, int, 0644);
5043 MODULE_PARM_DESC(zfs_arc_p_min_shift, "arc_c shift to calc min/max arc_p");
5045 module_param(zfs_disable_dup_eviction, int, 0644);
5046 MODULE_PARM_DESC(zfs_disable_dup_eviction, "disable duplicate buffer eviction");
5048 module_param(zfs_arc_memory_throttle_disable, int, 0644);
5049 MODULE_PARM_DESC(zfs_arc_memory_throttle_disable, "disable memory throttle");
5051 module_param(zfs_arc_min_prefetch_lifespan, int, 0644);
5052 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan, "Min life of prefetch block");
5054 module_param(l2arc_write_max, ulong, 0644);
5055 MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval");
5057 module_param(l2arc_write_boost, ulong, 0644);
5058 MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup");
5060 module_param(l2arc_headroom, ulong, 0644);
5061 MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache");
5063 module_param(l2arc_feed_secs, ulong, 0644);
5064 MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing");
5066 module_param(l2arc_feed_min_ms, ulong, 0644);
5067 MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds");
5069 module_param(l2arc_noprefetch, int, 0644);
5070 MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers");
5072 module_param(l2arc_feed_again, int, 0644);
5073 MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup");
5075 module_param(l2arc_norw, int, 0644);
5076 MODULE_PARM_DESC(l2arc_norw, "No reads during writes");