4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2011 by Delphix. All rights reserved.
28 * DVA-based Adjustable Replacement Cache
30 * While much of the theory of operation used here is
31 * based on the self-tuning, low overhead replacement cache
32 * presented by Megiddo and Modha at FAST 2003, there are some
33 * significant differences:
35 * 1. The Megiddo and Modha model assumes any page is evictable.
36 * Pages in its cache cannot be "locked" into memory. This makes
37 * the eviction algorithm simple: evict the last page in the list.
38 * This also make the performance characteristics easy to reason
39 * about. Our cache is not so simple. At any given moment, some
40 * subset of the blocks in the cache are un-evictable because we
41 * have handed out a reference to them. Blocks are only evictable
42 * when there are no external references active. This makes
43 * eviction far more problematic: we choose to evict the evictable
44 * blocks that are the "lowest" in the list.
46 * There are times when it is not possible to evict the requested
47 * space. In these circumstances we are unable to adjust the cache
48 * size. To prevent the cache growing unbounded at these times we
49 * implement a "cache throttle" that slows the flow of new data
50 * into the cache until we can make space available.
52 * 2. The Megiddo and Modha model assumes a fixed cache size.
53 * Pages are evicted when the cache is full and there is a cache
54 * miss. Our model has a variable sized cache. It grows with
55 * high use, but also tries to react to memory pressure from the
56 * operating system: decreasing its size when system memory is
59 * 3. The Megiddo and Modha model assumes a fixed page size. All
60 * elements of the cache are therefor exactly the same size. So
61 * when adjusting the cache size following a cache miss, its simply
62 * a matter of choosing a single page to evict. In our model, we
63 * have variable sized cache blocks (rangeing from 512 bytes to
64 * 128K bytes). We therefor choose a set of blocks to evict to make
65 * space for a cache miss that approximates as closely as possible
66 * the space used by the new block.
68 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
69 * by N. Megiddo & D. Modha, FAST 2003
75 * A new reference to a cache buffer can be obtained in two
76 * ways: 1) via a hash table lookup using the DVA as a key,
77 * or 2) via one of the ARC lists. The arc_read() interface
78 * uses method 1, while the internal arc algorithms for
79 * adjusting the cache use method 2. We therefor provide two
80 * types of locks: 1) the hash table lock array, and 2) the
83 * Buffers do not have their own mutexes, rather they rely on the
84 * hash table mutexes for the bulk of their protection (i.e. most
85 * fields in the arc_buf_hdr_t are protected by these mutexes).
87 * buf_hash_find() returns the appropriate mutex (held) when it
88 * locates the requested buffer in the hash table. It returns
89 * NULL for the mutex if the buffer was not in the table.
91 * buf_hash_remove() expects the appropriate hash mutex to be
92 * already held before it is invoked.
94 * Each arc state also has a mutex which is used to protect the
95 * buffer list associated with the state. When attempting to
96 * obtain a hash table lock while holding an arc list lock you
97 * must use: mutex_tryenter() to avoid deadlock. Also note that
98 * the active state mutex must be held before the ghost state mutex.
100 * Arc buffers may have an associated eviction callback function.
101 * This function will be invoked prior to removing the buffer (e.g.
102 * in arc_do_user_evicts()). Note however that the data associated
103 * with the buffer may be evicted prior to the callback. The callback
104 * must be made with *no locks held* (to prevent deadlock). Additionally,
105 * the users of callbacks must ensure that their private data is
106 * protected from simultaneous callbacks from arc_buf_evict()
107 * and arc_do_user_evicts().
109 * It as also possible to register a callback which is run when the
110 * arc_meta_limit is reached and no buffers can be safely evicted. In
111 * this case the arc user should drop a reference on some arc buffers so
112 * they can be reclaimed and the arc_meta_limit honored. For example,
113 * when using the ZPL each dentry holds a references on a znode. These
114 * dentries must be pruned before the arc buffer holding the znode can
117 * Note that the majority of the performance stats are manipulated
118 * with atomic operations.
120 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
122 * - L2ARC buflist creation
123 * - L2ARC buflist eviction
124 * - L2ARC write completion, which walks L2ARC buflists
125 * - ARC header destruction, as it removes from L2ARC buflists
126 * - ARC header release, as it removes from L2ARC buflists
131 #include <sys/zfs_context.h>
133 #include <sys/vdev.h>
134 #include <sys/vdev_impl.h>
136 #include <sys/vmsystm.h>
138 #include <sys/fs/swapnode.h>
141 #include <sys/callb.h>
142 #include <sys/kstat.h>
143 #include <sys/dmu_tx.h>
144 #include <zfs_fletcher.h>
146 static kmutex_t arc_reclaim_thr_lock
;
147 static kcondvar_t arc_reclaim_thr_cv
; /* used to signal reclaim thr */
148 static uint8_t arc_thread_exit
;
150 /* number of bytes to prune from caches when at arc_meta_limit is reached */
151 uint_t arc_meta_prune
= 1048576;
153 typedef enum arc_reclaim_strategy
{
154 ARC_RECLAIM_AGGR
, /* Aggressive reclaim strategy */
155 ARC_RECLAIM_CONS
/* Conservative reclaim strategy */
156 } arc_reclaim_strategy_t
;
158 /* number of seconds before growing cache again */
159 static int arc_grow_retry
= 5;
161 /* expiration time for arc_no_grow */
162 static clock_t arc_grow_time
= 0;
164 /* shift of arc_c for calculating both min and max arc_p */
165 static int arc_p_min_shift
= 4;
167 /* log2(fraction of arc to reclaim) */
168 static int arc_shrink_shift
= 5;
171 * minimum lifespan of a prefetch block in clock ticks
172 * (initialized in arc_init())
174 static int arc_min_prefetch_lifespan
;
179 * The arc has filled available memory and has now warmed up.
181 static boolean_t arc_warm
;
184 * These tunables are for performance analysis.
186 unsigned long zfs_arc_max
= 0;
187 unsigned long zfs_arc_min
= 0;
188 unsigned long zfs_arc_meta_limit
= 0;
189 int zfs_arc_grow_retry
= 0;
190 int zfs_arc_shrink_shift
= 0;
191 int zfs_arc_p_min_shift
= 0;
192 int zfs_arc_memory_throttle_disable
= 1;
193 int zfs_disable_dup_eviction
= 0;
194 int zfs_arc_meta_prune
= 0;
197 * Note that buffers can be in one of 6 states:
198 * ARC_anon - anonymous (discussed below)
199 * ARC_mru - recently used, currently cached
200 * ARC_mru_ghost - recentely used, no longer in cache
201 * ARC_mfu - frequently used, currently cached
202 * ARC_mfu_ghost - frequently used, no longer in cache
203 * ARC_l2c_only - exists in L2ARC but not other states
204 * When there are no active references to the buffer, they are
205 * are linked onto a list in one of these arc states. These are
206 * the only buffers that can be evicted or deleted. Within each
207 * state there are multiple lists, one for meta-data and one for
208 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
209 * etc.) is tracked separately so that it can be managed more
210 * explicitly: favored over data, limited explicitly.
212 * Anonymous buffers are buffers that are not associated with
213 * a DVA. These are buffers that hold dirty block copies
214 * before they are written to stable storage. By definition,
215 * they are "ref'd" and are considered part of arc_mru
216 * that cannot be freed. Generally, they will aquire a DVA
217 * as they are written and migrate onto the arc_mru list.
219 * The ARC_l2c_only state is for buffers that are in the second
220 * level ARC but no longer in any of the ARC_m* lists. The second
221 * level ARC itself may also contain buffers that are in any of
222 * the ARC_m* states - meaning that a buffer can exist in two
223 * places. The reason for the ARC_l2c_only state is to keep the
224 * buffer header in the hash table, so that reads that hit the
225 * second level ARC benefit from these fast lookups.
228 typedef struct arc_state
{
229 list_t arcs_list
[ARC_BUFC_NUMTYPES
]; /* list of evictable buffers */
230 uint64_t arcs_lsize
[ARC_BUFC_NUMTYPES
]; /* amount of evictable data */
231 uint64_t arcs_size
; /* total amount of data in this state */
236 static arc_state_t ARC_anon
;
237 static arc_state_t ARC_mru
;
238 static arc_state_t ARC_mru_ghost
;
239 static arc_state_t ARC_mfu
;
240 static arc_state_t ARC_mfu_ghost
;
241 static arc_state_t ARC_l2c_only
;
243 typedef struct arc_stats
{
244 kstat_named_t arcstat_hits
;
245 kstat_named_t arcstat_misses
;
246 kstat_named_t arcstat_demand_data_hits
;
247 kstat_named_t arcstat_demand_data_misses
;
248 kstat_named_t arcstat_demand_metadata_hits
;
249 kstat_named_t arcstat_demand_metadata_misses
;
250 kstat_named_t arcstat_prefetch_data_hits
;
251 kstat_named_t arcstat_prefetch_data_misses
;
252 kstat_named_t arcstat_prefetch_metadata_hits
;
253 kstat_named_t arcstat_prefetch_metadata_misses
;
254 kstat_named_t arcstat_mru_hits
;
255 kstat_named_t arcstat_mru_ghost_hits
;
256 kstat_named_t arcstat_mfu_hits
;
257 kstat_named_t arcstat_mfu_ghost_hits
;
258 kstat_named_t arcstat_deleted
;
259 kstat_named_t arcstat_recycle_miss
;
260 kstat_named_t arcstat_mutex_miss
;
261 kstat_named_t arcstat_evict_skip
;
262 kstat_named_t arcstat_evict_l2_cached
;
263 kstat_named_t arcstat_evict_l2_eligible
;
264 kstat_named_t arcstat_evict_l2_ineligible
;
265 kstat_named_t arcstat_hash_elements
;
266 kstat_named_t arcstat_hash_elements_max
;
267 kstat_named_t arcstat_hash_collisions
;
268 kstat_named_t arcstat_hash_chains
;
269 kstat_named_t arcstat_hash_chain_max
;
270 kstat_named_t arcstat_p
;
271 kstat_named_t arcstat_c
;
272 kstat_named_t arcstat_c_min
;
273 kstat_named_t arcstat_c_max
;
274 kstat_named_t arcstat_size
;
275 kstat_named_t arcstat_hdr_size
;
276 kstat_named_t arcstat_data_size
;
277 kstat_named_t arcstat_other_size
;
278 kstat_named_t arcstat_anon_size
;
279 kstat_named_t arcstat_anon_evict_data
;
280 kstat_named_t arcstat_anon_evict_metadata
;
281 kstat_named_t arcstat_mru_size
;
282 kstat_named_t arcstat_mru_evict_data
;
283 kstat_named_t arcstat_mru_evict_metadata
;
284 kstat_named_t arcstat_mru_ghost_size
;
285 kstat_named_t arcstat_mru_ghost_evict_data
;
286 kstat_named_t arcstat_mru_ghost_evict_metadata
;
287 kstat_named_t arcstat_mfu_size
;
288 kstat_named_t arcstat_mfu_evict_data
;
289 kstat_named_t arcstat_mfu_evict_metadata
;
290 kstat_named_t arcstat_mfu_ghost_size
;
291 kstat_named_t arcstat_mfu_ghost_evict_data
;
292 kstat_named_t arcstat_mfu_ghost_evict_metadata
;
293 kstat_named_t arcstat_l2_hits
;
294 kstat_named_t arcstat_l2_misses
;
295 kstat_named_t arcstat_l2_feeds
;
296 kstat_named_t arcstat_l2_rw_clash
;
297 kstat_named_t arcstat_l2_read_bytes
;
298 kstat_named_t arcstat_l2_write_bytes
;
299 kstat_named_t arcstat_l2_writes_sent
;
300 kstat_named_t arcstat_l2_writes_done
;
301 kstat_named_t arcstat_l2_writes_error
;
302 kstat_named_t arcstat_l2_writes_hdr_miss
;
303 kstat_named_t arcstat_l2_evict_lock_retry
;
304 kstat_named_t arcstat_l2_evict_reading
;
305 kstat_named_t arcstat_l2_free_on_write
;
306 kstat_named_t arcstat_l2_abort_lowmem
;
307 kstat_named_t arcstat_l2_cksum_bad
;
308 kstat_named_t arcstat_l2_io_error
;
309 kstat_named_t arcstat_l2_size
;
310 kstat_named_t arcstat_l2_hdr_size
;
311 kstat_named_t arcstat_memory_throttle_count
;
312 kstat_named_t arcstat_duplicate_buffers
;
313 kstat_named_t arcstat_duplicate_buffers_size
;
314 kstat_named_t arcstat_duplicate_reads
;
315 kstat_named_t arcstat_memory_direct_count
;
316 kstat_named_t arcstat_memory_indirect_count
;
317 kstat_named_t arcstat_no_grow
;
318 kstat_named_t arcstat_tempreserve
;
319 kstat_named_t arcstat_loaned_bytes
;
320 kstat_named_t arcstat_prune
;
321 kstat_named_t arcstat_meta_used
;
322 kstat_named_t arcstat_meta_limit
;
323 kstat_named_t arcstat_meta_max
;
326 static arc_stats_t arc_stats
= {
327 { "hits", KSTAT_DATA_UINT64
},
328 { "misses", KSTAT_DATA_UINT64
},
329 { "demand_data_hits", KSTAT_DATA_UINT64
},
330 { "demand_data_misses", KSTAT_DATA_UINT64
},
331 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
332 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
333 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
334 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
335 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
336 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
337 { "mru_hits", KSTAT_DATA_UINT64
},
338 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
339 { "mfu_hits", KSTAT_DATA_UINT64
},
340 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
341 { "deleted", KSTAT_DATA_UINT64
},
342 { "recycle_miss", KSTAT_DATA_UINT64
},
343 { "mutex_miss", KSTAT_DATA_UINT64
},
344 { "evict_skip", KSTAT_DATA_UINT64
},
345 { "evict_l2_cached", KSTAT_DATA_UINT64
},
346 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
347 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
348 { "hash_elements", KSTAT_DATA_UINT64
},
349 { "hash_elements_max", KSTAT_DATA_UINT64
},
350 { "hash_collisions", KSTAT_DATA_UINT64
},
351 { "hash_chains", KSTAT_DATA_UINT64
},
352 { "hash_chain_max", KSTAT_DATA_UINT64
},
353 { "p", KSTAT_DATA_UINT64
},
354 { "c", KSTAT_DATA_UINT64
},
355 { "c_min", KSTAT_DATA_UINT64
},
356 { "c_max", KSTAT_DATA_UINT64
},
357 { "size", KSTAT_DATA_UINT64
},
358 { "hdr_size", KSTAT_DATA_UINT64
},
359 { "data_size", KSTAT_DATA_UINT64
},
360 { "other_size", KSTAT_DATA_UINT64
},
361 { "anon_size", KSTAT_DATA_UINT64
},
362 { "anon_evict_data", KSTAT_DATA_UINT64
},
363 { "anon_evict_metadata", KSTAT_DATA_UINT64
},
364 { "mru_size", KSTAT_DATA_UINT64
},
365 { "mru_evict_data", KSTAT_DATA_UINT64
},
366 { "mru_evict_metadata", KSTAT_DATA_UINT64
},
367 { "mru_ghost_size", KSTAT_DATA_UINT64
},
368 { "mru_ghost_evict_data", KSTAT_DATA_UINT64
},
369 { "mru_ghost_evict_metadata", KSTAT_DATA_UINT64
},
370 { "mfu_size", KSTAT_DATA_UINT64
},
371 { "mfu_evict_data", KSTAT_DATA_UINT64
},
372 { "mfu_evict_metadata", KSTAT_DATA_UINT64
},
373 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
374 { "mfu_ghost_evict_data", KSTAT_DATA_UINT64
},
375 { "mfu_ghost_evict_metadata", KSTAT_DATA_UINT64
},
376 { "l2_hits", KSTAT_DATA_UINT64
},
377 { "l2_misses", KSTAT_DATA_UINT64
},
378 { "l2_feeds", KSTAT_DATA_UINT64
},
379 { "l2_rw_clash", KSTAT_DATA_UINT64
},
380 { "l2_read_bytes", KSTAT_DATA_UINT64
},
381 { "l2_write_bytes", KSTAT_DATA_UINT64
},
382 { "l2_writes_sent", KSTAT_DATA_UINT64
},
383 { "l2_writes_done", KSTAT_DATA_UINT64
},
384 { "l2_writes_error", KSTAT_DATA_UINT64
},
385 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64
},
386 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
387 { "l2_evict_reading", KSTAT_DATA_UINT64
},
388 { "l2_free_on_write", KSTAT_DATA_UINT64
},
389 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
390 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
391 { "l2_io_error", KSTAT_DATA_UINT64
},
392 { "l2_size", KSTAT_DATA_UINT64
},
393 { "l2_hdr_size", KSTAT_DATA_UINT64
},
394 { "memory_throttle_count", KSTAT_DATA_UINT64
},
395 { "duplicate_buffers", KSTAT_DATA_UINT64
},
396 { "duplicate_buffers_size", KSTAT_DATA_UINT64
},
397 { "duplicate_reads", KSTAT_DATA_UINT64
},
398 { "memory_direct_count", KSTAT_DATA_UINT64
},
399 { "memory_indirect_count", KSTAT_DATA_UINT64
},
400 { "arc_no_grow", KSTAT_DATA_UINT64
},
401 { "arc_tempreserve", KSTAT_DATA_UINT64
},
402 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
403 { "arc_prune", KSTAT_DATA_UINT64
},
404 { "arc_meta_used", KSTAT_DATA_UINT64
},
405 { "arc_meta_limit", KSTAT_DATA_UINT64
},
406 { "arc_meta_max", KSTAT_DATA_UINT64
},
409 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
411 #define ARCSTAT_INCR(stat, val) \
412 atomic_add_64(&arc_stats.stat.value.ui64, (val));
414 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
415 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
417 #define ARCSTAT_MAX(stat, val) { \
419 while ((val) > (m = arc_stats.stat.value.ui64) && \
420 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
424 #define ARCSTAT_MAXSTAT(stat) \
425 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
428 * We define a macro to allow ARC hits/misses to be easily broken down by
429 * two separate conditions, giving a total of four different subtypes for
430 * each of hits and misses (so eight statistics total).
432 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
435 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
437 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
441 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
443 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
448 static arc_state_t
*arc_anon
;
449 static arc_state_t
*arc_mru
;
450 static arc_state_t
*arc_mru_ghost
;
451 static arc_state_t
*arc_mfu
;
452 static arc_state_t
*arc_mfu_ghost
;
453 static arc_state_t
*arc_l2c_only
;
456 * There are several ARC variables that are critical to export as kstats --
457 * but we don't want to have to grovel around in the kstat whenever we wish to
458 * manipulate them. For these variables, we therefore define them to be in
459 * terms of the statistic variable. This assures that we are not introducing
460 * the possibility of inconsistency by having shadow copies of the variables,
461 * while still allowing the code to be readable.
463 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
464 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
465 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
466 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
467 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
468 #define arc_no_grow ARCSTAT(arcstat_no_grow)
469 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
470 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
471 #define arc_meta_used ARCSTAT(arcstat_meta_used)
472 #define arc_meta_limit ARCSTAT(arcstat_meta_limit)
473 #define arc_meta_max ARCSTAT(arcstat_meta_max)
475 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t
;
477 typedef struct arc_callback arc_callback_t
;
479 struct arc_callback
{
481 arc_done_func_t
*acb_done
;
483 zio_t
*acb_zio_dummy
;
484 arc_callback_t
*acb_next
;
487 typedef struct arc_write_callback arc_write_callback_t
;
489 struct arc_write_callback
{
491 arc_done_func_t
*awcb_ready
;
492 arc_done_func_t
*awcb_done
;
497 /* protected by hash lock */
502 kmutex_t b_freeze_lock
;
503 zio_cksum_t
*b_freeze_cksum
;
506 arc_buf_hdr_t
*b_hash_next
;
511 arc_callback_t
*b_acb
;
515 arc_buf_contents_t b_type
;
519 /* protected by arc state mutex */
520 arc_state_t
*b_state
;
521 list_node_t b_arc_node
;
523 /* updated atomically */
524 clock_t b_arc_access
;
526 /* self protecting */
529 l2arc_buf_hdr_t
*b_l2hdr
;
530 list_node_t b_l2node
;
533 static list_t arc_prune_list
;
534 static kmutex_t arc_prune_mtx
;
535 static arc_buf_t
*arc_eviction_list
;
536 static kmutex_t arc_eviction_mtx
;
537 static arc_buf_hdr_t arc_eviction_hdr
;
538 static void arc_get_data_buf(arc_buf_t
*buf
);
539 static void arc_access(arc_buf_hdr_t
*buf
, kmutex_t
*hash_lock
);
540 static int arc_evict_needed(arc_buf_contents_t type
);
541 static void arc_evict_ghost(arc_state_t
*state
, uint64_t spa
, int64_t bytes
);
543 static boolean_t
l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*ab
);
545 #define GHOST_STATE(state) \
546 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
547 (state) == arc_l2c_only)
550 * Private ARC flags. These flags are private ARC only flags that will show up
551 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
552 * be passed in as arc_flags in things like arc_read. However, these flags
553 * should never be passed and should only be set by ARC code. When adding new
554 * public flags, make sure not to smash the private ones.
557 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
558 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
559 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
560 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
561 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
562 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
563 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
564 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
565 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
566 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
568 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
569 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
570 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
571 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
572 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
573 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
574 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
575 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
576 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
577 (hdr)->b_l2hdr != NULL)
578 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
579 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
580 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
586 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
587 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
590 * Hash table routines
593 #define HT_LOCK_ALIGN 64
594 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
599 unsigned char pad
[HT_LOCK_PAD
];
603 #define BUF_LOCKS 256
604 typedef struct buf_hash_table
{
606 arc_buf_hdr_t
**ht_table
;
607 struct ht_lock ht_locks
[BUF_LOCKS
];
610 static buf_hash_table_t buf_hash_table
;
612 #define BUF_HASH_INDEX(spa, dva, birth) \
613 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
614 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
615 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
616 #define HDR_LOCK(hdr) \
617 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
619 uint64_t zfs_crc64_table
[256];
625 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
626 #define L2ARC_HEADROOM 2 /* num of writes */
627 #define L2ARC_FEED_SECS 1 /* caching interval secs */
628 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
630 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
631 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
634 * L2ARC Performance Tunables
636 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
637 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
638 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
639 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
640 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
641 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
642 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
643 int l2arc_norw
= B_TRUE
; /* no reads during writes */
648 typedef struct l2arc_dev
{
649 vdev_t
*l2ad_vdev
; /* vdev */
650 spa_t
*l2ad_spa
; /* spa */
651 uint64_t l2ad_hand
; /* next write location */
652 uint64_t l2ad_write
; /* desired write size, bytes */
653 uint64_t l2ad_boost
; /* warmup write boost, bytes */
654 uint64_t l2ad_start
; /* first addr on device */
655 uint64_t l2ad_end
; /* last addr on device */
656 uint64_t l2ad_evict
; /* last addr eviction reached */
657 boolean_t l2ad_first
; /* first sweep through */
658 boolean_t l2ad_writing
; /* currently writing */
659 list_t
*l2ad_buflist
; /* buffer list */
660 list_node_t l2ad_node
; /* device list node */
663 static list_t L2ARC_dev_list
; /* device list */
664 static list_t
*l2arc_dev_list
; /* device list pointer */
665 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
666 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
667 static kmutex_t l2arc_buflist_mtx
; /* mutex for all buflists */
668 static list_t L2ARC_free_on_write
; /* free after write buf list */
669 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
670 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
671 static uint64_t l2arc_ndev
; /* number of devices */
673 typedef struct l2arc_read_callback
{
674 arc_buf_t
*l2rcb_buf
; /* read buffer */
675 spa_t
*l2rcb_spa
; /* spa */
676 blkptr_t l2rcb_bp
; /* original blkptr */
677 zbookmark_t l2rcb_zb
; /* original bookmark */
678 int l2rcb_flags
; /* original flags */
679 } l2arc_read_callback_t
;
681 typedef struct l2arc_write_callback
{
682 l2arc_dev_t
*l2wcb_dev
; /* device info */
683 arc_buf_hdr_t
*l2wcb_head
; /* head of write buflist */
684 } l2arc_write_callback_t
;
686 struct l2arc_buf_hdr
{
687 /* protected by arc_buf_hdr mutex */
688 l2arc_dev_t
*b_dev
; /* L2ARC device */
689 uint64_t b_daddr
; /* disk address, offset byte */
692 typedef struct l2arc_data_free
{
693 /* protected by l2arc_free_on_write_mtx */
696 void (*l2df_func
)(void *, size_t);
697 list_node_t l2df_list_node
;
700 static kmutex_t l2arc_feed_thr_lock
;
701 static kcondvar_t l2arc_feed_thr_cv
;
702 static uint8_t l2arc_thread_exit
;
704 static void l2arc_read_done(zio_t
*zio
);
705 static void l2arc_hdr_stat_add(void);
706 static void l2arc_hdr_stat_remove(void);
709 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
711 uint8_t *vdva
= (uint8_t *)dva
;
712 uint64_t crc
= -1ULL;
715 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
717 for (i
= 0; i
< sizeof (dva_t
); i
++)
718 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
720 crc
^= (spa
>>8) ^ birth
;
725 #define BUF_EMPTY(buf) \
726 ((buf)->b_dva.dva_word[0] == 0 && \
727 (buf)->b_dva.dva_word[1] == 0 && \
730 #define BUF_EQUAL(spa, dva, birth, buf) \
731 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
732 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
733 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
736 buf_discard_identity(arc_buf_hdr_t
*hdr
)
738 hdr
->b_dva
.dva_word
[0] = 0;
739 hdr
->b_dva
.dva_word
[1] = 0;
744 static arc_buf_hdr_t
*
745 buf_hash_find(uint64_t spa
, const dva_t
*dva
, uint64_t birth
, kmutex_t
**lockp
)
747 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
748 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
751 mutex_enter(hash_lock
);
752 for (buf
= buf_hash_table
.ht_table
[idx
]; buf
!= NULL
;
753 buf
= buf
->b_hash_next
) {
754 if (BUF_EQUAL(spa
, dva
, birth
, buf
)) {
759 mutex_exit(hash_lock
);
765 * Insert an entry into the hash table. If there is already an element
766 * equal to elem in the hash table, then the already existing element
767 * will be returned and the new element will not be inserted.
768 * Otherwise returns NULL.
770 static arc_buf_hdr_t
*
771 buf_hash_insert(arc_buf_hdr_t
*buf
, kmutex_t
**lockp
)
773 uint64_t idx
= BUF_HASH_INDEX(buf
->b_spa
, &buf
->b_dva
, buf
->b_birth
);
774 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
778 ASSERT(!HDR_IN_HASH_TABLE(buf
));
780 mutex_enter(hash_lock
);
781 for (fbuf
= buf_hash_table
.ht_table
[idx
], i
= 0; fbuf
!= NULL
;
782 fbuf
= fbuf
->b_hash_next
, i
++) {
783 if (BUF_EQUAL(buf
->b_spa
, &buf
->b_dva
, buf
->b_birth
, fbuf
))
787 buf
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
788 buf_hash_table
.ht_table
[idx
] = buf
;
789 buf
->b_flags
|= ARC_IN_HASH_TABLE
;
791 /* collect some hash table performance data */
793 ARCSTAT_BUMP(arcstat_hash_collisions
);
795 ARCSTAT_BUMP(arcstat_hash_chains
);
797 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
800 ARCSTAT_BUMP(arcstat_hash_elements
);
801 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
807 buf_hash_remove(arc_buf_hdr_t
*buf
)
809 arc_buf_hdr_t
*fbuf
, **bufp
;
810 uint64_t idx
= BUF_HASH_INDEX(buf
->b_spa
, &buf
->b_dva
, buf
->b_birth
);
812 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
813 ASSERT(HDR_IN_HASH_TABLE(buf
));
815 bufp
= &buf_hash_table
.ht_table
[idx
];
816 while ((fbuf
= *bufp
) != buf
) {
817 ASSERT(fbuf
!= NULL
);
818 bufp
= &fbuf
->b_hash_next
;
820 *bufp
= buf
->b_hash_next
;
821 buf
->b_hash_next
= NULL
;
822 buf
->b_flags
&= ~ARC_IN_HASH_TABLE
;
824 /* collect some hash table performance data */
825 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
827 if (buf_hash_table
.ht_table
[idx
] &&
828 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
829 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
833 * Global data structures and functions for the buf kmem cache.
835 static kmem_cache_t
*hdr_cache
;
836 static kmem_cache_t
*buf_cache
;
843 #if defined(_KERNEL) && defined(HAVE_SPL)
844 /* Large allocations which do not require contiguous pages
845 * should be using vmem_free() in the linux kernel */
846 vmem_free(buf_hash_table
.ht_table
,
847 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
849 kmem_free(buf_hash_table
.ht_table
,
850 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
852 for (i
= 0; i
< BUF_LOCKS
; i
++)
853 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
854 kmem_cache_destroy(hdr_cache
);
855 kmem_cache_destroy(buf_cache
);
859 * Constructor callback - called when the cache is empty
860 * and a new buf is requested.
864 hdr_cons(void *vbuf
, void *unused
, int kmflag
)
866 arc_buf_hdr_t
*buf
= vbuf
;
868 bzero(buf
, sizeof (arc_buf_hdr_t
));
869 refcount_create(&buf
->b_refcnt
);
870 cv_init(&buf
->b_cv
, NULL
, CV_DEFAULT
, NULL
);
871 mutex_init(&buf
->b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
872 list_link_init(&buf
->b_arc_node
);
873 list_link_init(&buf
->b_l2node
);
874 arc_space_consume(sizeof (arc_buf_hdr_t
), ARC_SPACE_HDRS
);
881 buf_cons(void *vbuf
, void *unused
, int kmflag
)
883 arc_buf_t
*buf
= vbuf
;
885 bzero(buf
, sizeof (arc_buf_t
));
886 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
887 rw_init(&buf
->b_data_lock
, NULL
, RW_DEFAULT
, NULL
);
888 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
894 * Destructor callback - called when a cached buf is
895 * no longer required.
899 hdr_dest(void *vbuf
, void *unused
)
901 arc_buf_hdr_t
*buf
= vbuf
;
903 ASSERT(BUF_EMPTY(buf
));
904 refcount_destroy(&buf
->b_refcnt
);
905 cv_destroy(&buf
->b_cv
);
906 mutex_destroy(&buf
->b_freeze_lock
);
907 arc_space_return(sizeof (arc_buf_hdr_t
), ARC_SPACE_HDRS
);
912 buf_dest(void *vbuf
, void *unused
)
914 arc_buf_t
*buf
= vbuf
;
916 mutex_destroy(&buf
->b_evict_lock
);
917 rw_destroy(&buf
->b_data_lock
);
918 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
925 uint64_t hsize
= 1ULL << 12;
929 * The hash table is big enough to fill all of physical memory
930 * with an average 64K block size. The table will take up
931 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
933 while (hsize
* 65536 < physmem
* PAGESIZE
)
936 buf_hash_table
.ht_mask
= hsize
- 1;
937 #if defined(_KERNEL) && defined(HAVE_SPL)
938 /* Large allocations which do not require contiguous pages
939 * should be using vmem_alloc() in the linux kernel */
940 buf_hash_table
.ht_table
=
941 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
943 buf_hash_table
.ht_table
=
944 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
946 if (buf_hash_table
.ht_table
== NULL
) {
947 ASSERT(hsize
> (1ULL << 8));
952 hdr_cache
= kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t
),
953 0, hdr_cons
, hdr_dest
, NULL
, NULL
, NULL
, 0);
954 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
955 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
957 for (i
= 0; i
< 256; i
++)
958 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
959 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
961 for (i
= 0; i
< BUF_LOCKS
; i
++) {
962 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
963 NULL
, MUTEX_DEFAULT
, NULL
);
967 #define ARC_MINTIME (hz>>4) /* 62 ms */
970 arc_cksum_verify(arc_buf_t
*buf
)
974 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
977 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
978 if (buf
->b_hdr
->b_freeze_cksum
== NULL
||
979 (buf
->b_hdr
->b_flags
& ARC_IO_ERROR
)) {
980 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
983 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
984 if (!ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
))
985 panic("buffer modified while frozen!");
986 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
990 arc_cksum_equal(arc_buf_t
*buf
)
995 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
996 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
997 equal
= ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
);
998 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
1004 arc_cksum_compute(arc_buf_t
*buf
, boolean_t force
)
1006 if (!force
&& !(zfs_flags
& ZFS_DEBUG_MODIFY
))
1009 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
1010 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1011 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
1014 buf
->b_hdr
->b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
),
1016 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
,
1017 buf
->b_hdr
->b_freeze_cksum
);
1018 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
1022 arc_buf_thaw(arc_buf_t
*buf
)
1024 if (zfs_flags
& ZFS_DEBUG_MODIFY
) {
1025 if (buf
->b_hdr
->b_state
!= arc_anon
)
1026 panic("modifying non-anon buffer!");
1027 if (buf
->b_hdr
->b_flags
& ARC_IO_IN_PROGRESS
)
1028 panic("modifying buffer while i/o in progress!");
1029 arc_cksum_verify(buf
);
1032 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
1033 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1034 kmem_free(buf
->b_hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
1035 buf
->b_hdr
->b_freeze_cksum
= NULL
;
1038 if (zfs_flags
& ZFS_DEBUG_MODIFY
) {
1039 if (buf
->b_hdr
->b_thawed
)
1040 kmem_free(buf
->b_hdr
->b_thawed
, 1);
1041 buf
->b_hdr
->b_thawed
= kmem_alloc(1, KM_SLEEP
);
1044 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
1048 arc_buf_freeze(arc_buf_t
*buf
)
1050 kmutex_t
*hash_lock
;
1052 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1055 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1056 mutex_enter(hash_lock
);
1058 ASSERT(buf
->b_hdr
->b_freeze_cksum
!= NULL
||
1059 buf
->b_hdr
->b_state
== arc_anon
);
1060 arc_cksum_compute(buf
, B_FALSE
);
1061 mutex_exit(hash_lock
);
1065 add_reference(arc_buf_hdr_t
*ab
, kmutex_t
*hash_lock
, void *tag
)
1067 ASSERT(MUTEX_HELD(hash_lock
));
1069 if ((refcount_add(&ab
->b_refcnt
, tag
) == 1) &&
1070 (ab
->b_state
!= arc_anon
)) {
1071 uint64_t delta
= ab
->b_size
* ab
->b_datacnt
;
1072 list_t
*list
= &ab
->b_state
->arcs_list
[ab
->b_type
];
1073 uint64_t *size
= &ab
->b_state
->arcs_lsize
[ab
->b_type
];
1075 ASSERT(!MUTEX_HELD(&ab
->b_state
->arcs_mtx
));
1076 mutex_enter(&ab
->b_state
->arcs_mtx
);
1077 ASSERT(list_link_active(&ab
->b_arc_node
));
1078 list_remove(list
, ab
);
1079 if (GHOST_STATE(ab
->b_state
)) {
1080 ASSERT0(ab
->b_datacnt
);
1081 ASSERT3P(ab
->b_buf
, ==, NULL
);
1085 ASSERT3U(*size
, >=, delta
);
1086 atomic_add_64(size
, -delta
);
1087 mutex_exit(&ab
->b_state
->arcs_mtx
);
1088 /* remove the prefetch flag if we get a reference */
1089 if (ab
->b_flags
& ARC_PREFETCH
)
1090 ab
->b_flags
&= ~ARC_PREFETCH
;
1095 remove_reference(arc_buf_hdr_t
*ab
, kmutex_t
*hash_lock
, void *tag
)
1098 arc_state_t
*state
= ab
->b_state
;
1100 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
1101 ASSERT(!GHOST_STATE(state
));
1103 if (((cnt
= refcount_remove(&ab
->b_refcnt
, tag
)) == 0) &&
1104 (state
!= arc_anon
)) {
1105 uint64_t *size
= &state
->arcs_lsize
[ab
->b_type
];
1107 ASSERT(!MUTEX_HELD(&state
->arcs_mtx
));
1108 mutex_enter(&state
->arcs_mtx
);
1109 ASSERT(!list_link_active(&ab
->b_arc_node
));
1110 list_insert_head(&state
->arcs_list
[ab
->b_type
], ab
);
1111 ASSERT(ab
->b_datacnt
> 0);
1112 atomic_add_64(size
, ab
->b_size
* ab
->b_datacnt
);
1113 mutex_exit(&state
->arcs_mtx
);
1119 * Move the supplied buffer to the indicated state. The mutex
1120 * for the buffer must be held by the caller.
1123 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*ab
, kmutex_t
*hash_lock
)
1125 arc_state_t
*old_state
= ab
->b_state
;
1126 int64_t refcnt
= refcount_count(&ab
->b_refcnt
);
1127 uint64_t from_delta
, to_delta
;
1129 ASSERT(MUTEX_HELD(hash_lock
));
1130 ASSERT(new_state
!= old_state
);
1131 ASSERT(refcnt
== 0 || ab
->b_datacnt
> 0);
1132 ASSERT(ab
->b_datacnt
== 0 || !GHOST_STATE(new_state
));
1133 ASSERT(ab
->b_datacnt
<= 1 || old_state
!= arc_anon
);
1135 from_delta
= to_delta
= ab
->b_datacnt
* ab
->b_size
;
1138 * If this buffer is evictable, transfer it from the
1139 * old state list to the new state list.
1142 if (old_state
!= arc_anon
) {
1143 int use_mutex
= !MUTEX_HELD(&old_state
->arcs_mtx
);
1144 uint64_t *size
= &old_state
->arcs_lsize
[ab
->b_type
];
1147 mutex_enter(&old_state
->arcs_mtx
);
1149 ASSERT(list_link_active(&ab
->b_arc_node
));
1150 list_remove(&old_state
->arcs_list
[ab
->b_type
], ab
);
1153 * If prefetching out of the ghost cache,
1154 * we will have a non-zero datacnt.
1156 if (GHOST_STATE(old_state
) && ab
->b_datacnt
== 0) {
1157 /* ghost elements have a ghost size */
1158 ASSERT(ab
->b_buf
== NULL
);
1159 from_delta
= ab
->b_size
;
1161 ASSERT3U(*size
, >=, from_delta
);
1162 atomic_add_64(size
, -from_delta
);
1165 mutex_exit(&old_state
->arcs_mtx
);
1167 if (new_state
!= arc_anon
) {
1168 int use_mutex
= !MUTEX_HELD(&new_state
->arcs_mtx
);
1169 uint64_t *size
= &new_state
->arcs_lsize
[ab
->b_type
];
1172 mutex_enter(&new_state
->arcs_mtx
);
1174 list_insert_head(&new_state
->arcs_list
[ab
->b_type
], ab
);
1176 /* ghost elements have a ghost size */
1177 if (GHOST_STATE(new_state
)) {
1178 ASSERT(ab
->b_datacnt
== 0);
1179 ASSERT(ab
->b_buf
== NULL
);
1180 to_delta
= ab
->b_size
;
1182 atomic_add_64(size
, to_delta
);
1185 mutex_exit(&new_state
->arcs_mtx
);
1189 ASSERT(!BUF_EMPTY(ab
));
1190 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(ab
))
1191 buf_hash_remove(ab
);
1193 /* adjust state sizes */
1195 atomic_add_64(&new_state
->arcs_size
, to_delta
);
1197 ASSERT3U(old_state
->arcs_size
, >=, from_delta
);
1198 atomic_add_64(&old_state
->arcs_size
, -from_delta
);
1200 ab
->b_state
= new_state
;
1202 /* adjust l2arc hdr stats */
1203 if (new_state
== arc_l2c_only
)
1204 l2arc_hdr_stat_add();
1205 else if (old_state
== arc_l2c_only
)
1206 l2arc_hdr_stat_remove();
1210 arc_space_consume(uint64_t space
, arc_space_type_t type
)
1212 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1217 case ARC_SPACE_DATA
:
1218 ARCSTAT_INCR(arcstat_data_size
, space
);
1220 case ARC_SPACE_OTHER
:
1221 ARCSTAT_INCR(arcstat_other_size
, space
);
1223 case ARC_SPACE_HDRS
:
1224 ARCSTAT_INCR(arcstat_hdr_size
, space
);
1226 case ARC_SPACE_L2HDRS
:
1227 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
1231 atomic_add_64(&arc_meta_used
, space
);
1232 atomic_add_64(&arc_size
, space
);
1236 arc_space_return(uint64_t space
, arc_space_type_t type
)
1238 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1243 case ARC_SPACE_DATA
:
1244 ARCSTAT_INCR(arcstat_data_size
, -space
);
1246 case ARC_SPACE_OTHER
:
1247 ARCSTAT_INCR(arcstat_other_size
, -space
);
1249 case ARC_SPACE_HDRS
:
1250 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
1252 case ARC_SPACE_L2HDRS
:
1253 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
1257 ASSERT(arc_meta_used
>= space
);
1258 if (arc_meta_max
< arc_meta_used
)
1259 arc_meta_max
= arc_meta_used
;
1260 atomic_add_64(&arc_meta_used
, -space
);
1261 ASSERT(arc_size
>= space
);
1262 atomic_add_64(&arc_size
, -space
);
1266 arc_data_buf_alloc(uint64_t size
)
1268 if (arc_evict_needed(ARC_BUFC_DATA
))
1269 cv_signal(&arc_reclaim_thr_cv
);
1270 atomic_add_64(&arc_size
, size
);
1271 return (zio_data_buf_alloc(size
));
1275 arc_data_buf_free(void *buf
, uint64_t size
)
1277 zio_data_buf_free(buf
, size
);
1278 ASSERT(arc_size
>= size
);
1279 atomic_add_64(&arc_size
, -size
);
1283 arc_buf_alloc(spa_t
*spa
, int size
, void *tag
, arc_buf_contents_t type
)
1288 ASSERT3U(size
, >, 0);
1289 hdr
= kmem_cache_alloc(hdr_cache
, KM_PUSHPAGE
);
1290 ASSERT(BUF_EMPTY(hdr
));
1293 hdr
->b_spa
= spa_load_guid(spa
);
1294 hdr
->b_state
= arc_anon
;
1295 hdr
->b_arc_access
= 0;
1296 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1299 buf
->b_efunc
= NULL
;
1300 buf
->b_private
= NULL
;
1303 arc_get_data_buf(buf
);
1306 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
1307 (void) refcount_add(&hdr
->b_refcnt
, tag
);
1312 static char *arc_onloan_tag
= "onloan";
1315 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1316 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1317 * buffers must be returned to the arc before they can be used by the DMU or
1321 arc_loan_buf(spa_t
*spa
, int size
)
1325 buf
= arc_buf_alloc(spa
, size
, arc_onloan_tag
, ARC_BUFC_DATA
);
1327 atomic_add_64(&arc_loaned_bytes
, size
);
1332 * Return a loaned arc buffer to the arc.
1335 arc_return_buf(arc_buf_t
*buf
, void *tag
)
1337 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1339 ASSERT(buf
->b_data
!= NULL
);
1340 (void) refcount_add(&hdr
->b_refcnt
, tag
);
1341 (void) refcount_remove(&hdr
->b_refcnt
, arc_onloan_tag
);
1343 atomic_add_64(&arc_loaned_bytes
, -hdr
->b_size
);
1346 /* Detach an arc_buf from a dbuf (tag) */
1348 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
1352 ASSERT(buf
->b_data
!= NULL
);
1354 (void) refcount_add(&hdr
->b_refcnt
, arc_onloan_tag
);
1355 (void) refcount_remove(&hdr
->b_refcnt
, tag
);
1356 buf
->b_efunc
= NULL
;
1357 buf
->b_private
= NULL
;
1359 atomic_add_64(&arc_loaned_bytes
, hdr
->b_size
);
1363 arc_buf_clone(arc_buf_t
*from
)
1366 arc_buf_hdr_t
*hdr
= from
->b_hdr
;
1367 uint64_t size
= hdr
->b_size
;
1369 ASSERT(hdr
->b_state
!= arc_anon
);
1371 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1374 buf
->b_efunc
= NULL
;
1375 buf
->b_private
= NULL
;
1376 buf
->b_next
= hdr
->b_buf
;
1378 arc_get_data_buf(buf
);
1379 bcopy(from
->b_data
, buf
->b_data
, size
);
1382 * This buffer already exists in the arc so create a duplicate
1383 * copy for the caller. If the buffer is associated with user data
1384 * then track the size and number of duplicates. These stats will be
1385 * updated as duplicate buffers are created and destroyed.
1387 if (hdr
->b_type
== ARC_BUFC_DATA
) {
1388 ARCSTAT_BUMP(arcstat_duplicate_buffers
);
1389 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, size
);
1391 hdr
->b_datacnt
+= 1;
1396 arc_buf_add_ref(arc_buf_t
*buf
, void* tag
)
1399 kmutex_t
*hash_lock
;
1402 * Check to see if this buffer is evicted. Callers
1403 * must verify b_data != NULL to know if the add_ref
1406 mutex_enter(&buf
->b_evict_lock
);
1407 if (buf
->b_data
== NULL
) {
1408 mutex_exit(&buf
->b_evict_lock
);
1411 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1412 mutex_enter(hash_lock
);
1414 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1415 mutex_exit(&buf
->b_evict_lock
);
1417 ASSERT(hdr
->b_state
== arc_mru
|| hdr
->b_state
== arc_mfu
);
1418 add_reference(hdr
, hash_lock
, tag
);
1419 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
1420 arc_access(hdr
, hash_lock
);
1421 mutex_exit(hash_lock
);
1422 ARCSTAT_BUMP(arcstat_hits
);
1423 ARCSTAT_CONDSTAT(!(hdr
->b_flags
& ARC_PREFETCH
),
1424 demand
, prefetch
, hdr
->b_type
!= ARC_BUFC_METADATA
,
1425 data
, metadata
, hits
);
1429 * Free the arc data buffer. If it is an l2arc write in progress,
1430 * the buffer is placed on l2arc_free_on_write to be freed later.
1433 arc_buf_data_free(arc_buf_hdr_t
*hdr
, void (*free_func
)(void *, size_t),
1434 void *data
, size_t size
)
1436 if (HDR_L2_WRITING(hdr
)) {
1437 l2arc_data_free_t
*df
;
1438 df
= kmem_alloc(sizeof (l2arc_data_free_t
), KM_PUSHPAGE
);
1439 df
->l2df_data
= data
;
1440 df
->l2df_size
= size
;
1441 df
->l2df_func
= free_func
;
1442 mutex_enter(&l2arc_free_on_write_mtx
);
1443 list_insert_head(l2arc_free_on_write
, df
);
1444 mutex_exit(&l2arc_free_on_write_mtx
);
1445 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
1447 free_func(data
, size
);
1452 arc_buf_destroy(arc_buf_t
*buf
, boolean_t recycle
, boolean_t all
)
1456 /* free up data associated with the buf */
1458 arc_state_t
*state
= buf
->b_hdr
->b_state
;
1459 uint64_t size
= buf
->b_hdr
->b_size
;
1460 arc_buf_contents_t type
= buf
->b_hdr
->b_type
;
1462 arc_cksum_verify(buf
);
1465 if (type
== ARC_BUFC_METADATA
) {
1466 arc_buf_data_free(buf
->b_hdr
, zio_buf_free
,
1468 arc_space_return(size
, ARC_SPACE_DATA
);
1470 ASSERT(type
== ARC_BUFC_DATA
);
1471 arc_buf_data_free(buf
->b_hdr
,
1472 zio_data_buf_free
, buf
->b_data
, size
);
1473 ARCSTAT_INCR(arcstat_data_size
, -size
);
1474 atomic_add_64(&arc_size
, -size
);
1477 if (list_link_active(&buf
->b_hdr
->b_arc_node
)) {
1478 uint64_t *cnt
= &state
->arcs_lsize
[type
];
1480 ASSERT(refcount_is_zero(&buf
->b_hdr
->b_refcnt
));
1481 ASSERT(state
!= arc_anon
);
1483 ASSERT3U(*cnt
, >=, size
);
1484 atomic_add_64(cnt
, -size
);
1486 ASSERT3U(state
->arcs_size
, >=, size
);
1487 atomic_add_64(&state
->arcs_size
, -size
);
1491 * If we're destroying a duplicate buffer make sure
1492 * that the appropriate statistics are updated.
1494 if (buf
->b_hdr
->b_datacnt
> 1 &&
1495 buf
->b_hdr
->b_type
== ARC_BUFC_DATA
) {
1496 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
1497 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, -size
);
1499 ASSERT(buf
->b_hdr
->b_datacnt
> 0);
1500 buf
->b_hdr
->b_datacnt
-= 1;
1503 /* only remove the buf if requested */
1507 /* remove the buf from the hdr list */
1508 for (bufp
= &buf
->b_hdr
->b_buf
; *bufp
!= buf
; bufp
= &(*bufp
)->b_next
)
1510 *bufp
= buf
->b_next
;
1513 ASSERT(buf
->b_efunc
== NULL
);
1515 /* clean up the buf */
1517 kmem_cache_free(buf_cache
, buf
);
1521 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
1523 l2arc_buf_hdr_t
*l2hdr
= hdr
->b_l2hdr
;
1525 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
1526 ASSERT3P(hdr
->b_state
, ==, arc_anon
);
1527 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
1529 if (l2hdr
!= NULL
) {
1530 boolean_t buflist_held
= MUTEX_HELD(&l2arc_buflist_mtx
);
1532 * To prevent arc_free() and l2arc_evict() from
1533 * attempting to free the same buffer at the same time,
1534 * a FREE_IN_PROGRESS flag is given to arc_free() to
1535 * give it priority. l2arc_evict() can't destroy this
1536 * header while we are waiting on l2arc_buflist_mtx.
1538 * The hdr may be removed from l2ad_buflist before we
1539 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1541 if (!buflist_held
) {
1542 mutex_enter(&l2arc_buflist_mtx
);
1543 l2hdr
= hdr
->b_l2hdr
;
1546 if (l2hdr
!= NULL
) {
1547 list_remove(l2hdr
->b_dev
->l2ad_buflist
, hdr
);
1548 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
1549 kmem_free(l2hdr
, sizeof (l2arc_buf_hdr_t
));
1550 if (hdr
->b_state
== arc_l2c_only
)
1551 l2arc_hdr_stat_remove();
1552 hdr
->b_l2hdr
= NULL
;
1556 mutex_exit(&l2arc_buflist_mtx
);
1559 if (!BUF_EMPTY(hdr
)) {
1560 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
1561 buf_discard_identity(hdr
);
1563 while (hdr
->b_buf
) {
1564 arc_buf_t
*buf
= hdr
->b_buf
;
1567 mutex_enter(&arc_eviction_mtx
);
1568 mutex_enter(&buf
->b_evict_lock
);
1569 ASSERT(buf
->b_hdr
!= NULL
);
1570 arc_buf_destroy(hdr
->b_buf
, FALSE
, FALSE
);
1571 hdr
->b_buf
= buf
->b_next
;
1572 buf
->b_hdr
= &arc_eviction_hdr
;
1573 buf
->b_next
= arc_eviction_list
;
1574 arc_eviction_list
= buf
;
1575 mutex_exit(&buf
->b_evict_lock
);
1576 mutex_exit(&arc_eviction_mtx
);
1578 arc_buf_destroy(hdr
->b_buf
, FALSE
, TRUE
);
1581 if (hdr
->b_freeze_cksum
!= NULL
) {
1582 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
1583 hdr
->b_freeze_cksum
= NULL
;
1585 if (hdr
->b_thawed
) {
1586 kmem_free(hdr
->b_thawed
, 1);
1587 hdr
->b_thawed
= NULL
;
1590 ASSERT(!list_link_active(&hdr
->b_arc_node
));
1591 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
1592 ASSERT3P(hdr
->b_acb
, ==, NULL
);
1593 kmem_cache_free(hdr_cache
, hdr
);
1597 arc_buf_free(arc_buf_t
*buf
, void *tag
)
1599 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1600 int hashed
= hdr
->b_state
!= arc_anon
;
1602 ASSERT(buf
->b_efunc
== NULL
);
1603 ASSERT(buf
->b_data
!= NULL
);
1606 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
1608 mutex_enter(hash_lock
);
1610 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1612 (void) remove_reference(hdr
, hash_lock
, tag
);
1613 if (hdr
->b_datacnt
> 1) {
1614 arc_buf_destroy(buf
, FALSE
, TRUE
);
1616 ASSERT(buf
== hdr
->b_buf
);
1617 ASSERT(buf
->b_efunc
== NULL
);
1618 hdr
->b_flags
|= ARC_BUF_AVAILABLE
;
1620 mutex_exit(hash_lock
);
1621 } else if (HDR_IO_IN_PROGRESS(hdr
)) {
1624 * We are in the middle of an async write. Don't destroy
1625 * this buffer unless the write completes before we finish
1626 * decrementing the reference count.
1628 mutex_enter(&arc_eviction_mtx
);
1629 (void) remove_reference(hdr
, NULL
, tag
);
1630 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
1631 destroy_hdr
= !HDR_IO_IN_PROGRESS(hdr
);
1632 mutex_exit(&arc_eviction_mtx
);
1634 arc_hdr_destroy(hdr
);
1636 if (remove_reference(hdr
, NULL
, tag
) > 0)
1637 arc_buf_destroy(buf
, FALSE
, TRUE
);
1639 arc_hdr_destroy(hdr
);
1644 arc_buf_remove_ref(arc_buf_t
*buf
, void* tag
)
1646 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1647 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
1648 int no_callback
= (buf
->b_efunc
== NULL
);
1650 if (hdr
->b_state
== arc_anon
) {
1651 ASSERT(hdr
->b_datacnt
== 1);
1652 arc_buf_free(buf
, tag
);
1653 return (no_callback
);
1656 mutex_enter(hash_lock
);
1658 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1659 ASSERT(hdr
->b_state
!= arc_anon
);
1660 ASSERT(buf
->b_data
!= NULL
);
1662 (void) remove_reference(hdr
, hash_lock
, tag
);
1663 if (hdr
->b_datacnt
> 1) {
1665 arc_buf_destroy(buf
, FALSE
, TRUE
);
1666 } else if (no_callback
) {
1667 ASSERT(hdr
->b_buf
== buf
&& buf
->b_next
== NULL
);
1668 ASSERT(buf
->b_efunc
== NULL
);
1669 hdr
->b_flags
|= ARC_BUF_AVAILABLE
;
1671 ASSERT(no_callback
|| hdr
->b_datacnt
> 1 ||
1672 refcount_is_zero(&hdr
->b_refcnt
));
1673 mutex_exit(hash_lock
);
1674 return (no_callback
);
1678 arc_buf_size(arc_buf_t
*buf
)
1680 return (buf
->b_hdr
->b_size
);
1684 * Called from the DMU to determine if the current buffer should be
1685 * evicted. In order to ensure proper locking, the eviction must be initiated
1686 * from the DMU. Return true if the buffer is associated with user data and
1687 * duplicate buffers still exist.
1690 arc_buf_eviction_needed(arc_buf_t
*buf
)
1693 boolean_t evict_needed
= B_FALSE
;
1695 if (zfs_disable_dup_eviction
)
1698 mutex_enter(&buf
->b_evict_lock
);
1702 * We are in arc_do_user_evicts(); let that function
1703 * perform the eviction.
1705 ASSERT(buf
->b_data
== NULL
);
1706 mutex_exit(&buf
->b_evict_lock
);
1708 } else if (buf
->b_data
== NULL
) {
1710 * We have already been added to the arc eviction list;
1711 * recommend eviction.
1713 ASSERT3P(hdr
, ==, &arc_eviction_hdr
);
1714 mutex_exit(&buf
->b_evict_lock
);
1718 if (hdr
->b_datacnt
> 1 && hdr
->b_type
== ARC_BUFC_DATA
)
1719 evict_needed
= B_TRUE
;
1721 mutex_exit(&buf
->b_evict_lock
);
1722 return (evict_needed
);
1726 * Evict buffers from list until we've removed the specified number of
1727 * bytes. Move the removed buffers to the appropriate evict state.
1728 * If the recycle flag is set, then attempt to "recycle" a buffer:
1729 * - look for a buffer to evict that is `bytes' long.
1730 * - return the data block from this buffer rather than freeing it.
1731 * This flag is used by callers that are trying to make space for a
1732 * new buffer in a full arc cache.
1734 * This function makes a "best effort". It skips over any buffers
1735 * it can't get a hash_lock on, and so may not catch all candidates.
1736 * It may also return without evicting as much space as requested.
1739 arc_evict(arc_state_t
*state
, uint64_t spa
, int64_t bytes
, boolean_t recycle
,
1740 arc_buf_contents_t type
)
1742 arc_state_t
*evicted_state
;
1743 uint64_t bytes_evicted
= 0, skipped
= 0, missed
= 0;
1744 arc_buf_hdr_t
*ab
, *ab_prev
= NULL
;
1745 list_t
*list
= &state
->arcs_list
[type
];
1746 kmutex_t
*hash_lock
;
1747 boolean_t have_lock
;
1748 void *stolen
= NULL
;
1750 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
1752 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
1754 mutex_enter(&state
->arcs_mtx
);
1755 mutex_enter(&evicted_state
->arcs_mtx
);
1757 for (ab
= list_tail(list
); ab
; ab
= ab_prev
) {
1758 ab_prev
= list_prev(list
, ab
);
1759 /* prefetch buffers have a minimum lifespan */
1760 if (HDR_IO_IN_PROGRESS(ab
) ||
1761 (spa
&& ab
->b_spa
!= spa
) ||
1762 (ab
->b_flags
& (ARC_PREFETCH
|ARC_INDIRECT
) &&
1763 ddi_get_lbolt() - ab
->b_arc_access
<
1764 arc_min_prefetch_lifespan
)) {
1768 /* "lookahead" for better eviction candidate */
1769 if (recycle
&& ab
->b_size
!= bytes
&&
1770 ab_prev
&& ab_prev
->b_size
== bytes
)
1772 hash_lock
= HDR_LOCK(ab
);
1773 have_lock
= MUTEX_HELD(hash_lock
);
1774 if (have_lock
|| mutex_tryenter(hash_lock
)) {
1775 ASSERT0(refcount_count(&ab
->b_refcnt
));
1776 ASSERT(ab
->b_datacnt
> 0);
1778 arc_buf_t
*buf
= ab
->b_buf
;
1779 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
1784 bytes_evicted
+= ab
->b_size
;
1785 if (recycle
&& ab
->b_type
== type
&&
1786 ab
->b_size
== bytes
&&
1787 !HDR_L2_WRITING(ab
)) {
1788 stolen
= buf
->b_data
;
1793 mutex_enter(&arc_eviction_mtx
);
1794 arc_buf_destroy(buf
,
1795 buf
->b_data
== stolen
, FALSE
);
1796 ab
->b_buf
= buf
->b_next
;
1797 buf
->b_hdr
= &arc_eviction_hdr
;
1798 buf
->b_next
= arc_eviction_list
;
1799 arc_eviction_list
= buf
;
1800 mutex_exit(&arc_eviction_mtx
);
1801 mutex_exit(&buf
->b_evict_lock
);
1803 mutex_exit(&buf
->b_evict_lock
);
1804 arc_buf_destroy(buf
,
1805 buf
->b_data
== stolen
, TRUE
);
1810 ARCSTAT_INCR(arcstat_evict_l2_cached
,
1813 if (l2arc_write_eligible(ab
->b_spa
, ab
)) {
1814 ARCSTAT_INCR(arcstat_evict_l2_eligible
,
1818 arcstat_evict_l2_ineligible
,
1823 if (ab
->b_datacnt
== 0) {
1824 arc_change_state(evicted_state
, ab
, hash_lock
);
1825 ASSERT(HDR_IN_HASH_TABLE(ab
));
1826 ab
->b_flags
|= ARC_IN_HASH_TABLE
;
1827 ab
->b_flags
&= ~ARC_BUF_AVAILABLE
;
1828 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, ab
);
1831 mutex_exit(hash_lock
);
1832 if (bytes
>= 0 && bytes_evicted
>= bytes
)
1839 mutex_exit(&evicted_state
->arcs_mtx
);
1840 mutex_exit(&state
->arcs_mtx
);
1842 if (bytes_evicted
< bytes
)
1843 dprintf("only evicted %lld bytes from %x\n",
1844 (longlong_t
)bytes_evicted
, state
);
1847 ARCSTAT_INCR(arcstat_evict_skip
, skipped
);
1850 ARCSTAT_INCR(arcstat_mutex_miss
, missed
);
1853 * We have just evicted some date into the ghost state, make
1854 * sure we also adjust the ghost state size if necessary.
1857 arc_mru_ghost
->arcs_size
+ arc_mfu_ghost
->arcs_size
> arc_c
) {
1858 int64_t mru_over
= arc_anon
->arcs_size
+ arc_mru
->arcs_size
+
1859 arc_mru_ghost
->arcs_size
- arc_c
;
1861 if (mru_over
> 0 && arc_mru_ghost
->arcs_lsize
[type
] > 0) {
1863 MIN(arc_mru_ghost
->arcs_lsize
[type
], mru_over
);
1864 arc_evict_ghost(arc_mru_ghost
, 0, todelete
);
1865 } else if (arc_mfu_ghost
->arcs_lsize
[type
] > 0) {
1866 int64_t todelete
= MIN(arc_mfu_ghost
->arcs_lsize
[type
],
1867 arc_mru_ghost
->arcs_size
+
1868 arc_mfu_ghost
->arcs_size
- arc_c
);
1869 arc_evict_ghost(arc_mfu_ghost
, 0, todelete
);
1877 * Remove buffers from list until we've removed the specified number of
1878 * bytes. Destroy the buffers that are removed.
1881 arc_evict_ghost(arc_state_t
*state
, uint64_t spa
, int64_t bytes
)
1883 arc_buf_hdr_t
*ab
, *ab_prev
;
1884 arc_buf_hdr_t marker
;
1885 list_t
*list
= &state
->arcs_list
[ARC_BUFC_DATA
];
1886 kmutex_t
*hash_lock
;
1887 uint64_t bytes_deleted
= 0;
1888 uint64_t bufs_skipped
= 0;
1890 ASSERT(GHOST_STATE(state
));
1891 bzero(&marker
, sizeof(marker
));
1893 mutex_enter(&state
->arcs_mtx
);
1894 for (ab
= list_tail(list
); ab
; ab
= ab_prev
) {
1895 ab_prev
= list_prev(list
, ab
);
1896 if (spa
&& ab
->b_spa
!= spa
)
1899 /* ignore markers */
1903 hash_lock
= HDR_LOCK(ab
);
1904 /* caller may be trying to modify this buffer, skip it */
1905 if (MUTEX_HELD(hash_lock
))
1907 if (mutex_tryenter(hash_lock
)) {
1908 ASSERT(!HDR_IO_IN_PROGRESS(ab
));
1909 ASSERT(ab
->b_buf
== NULL
);
1910 ARCSTAT_BUMP(arcstat_deleted
);
1911 bytes_deleted
+= ab
->b_size
;
1913 if (ab
->b_l2hdr
!= NULL
) {
1915 * This buffer is cached on the 2nd Level ARC;
1916 * don't destroy the header.
1918 arc_change_state(arc_l2c_only
, ab
, hash_lock
);
1919 mutex_exit(hash_lock
);
1921 arc_change_state(arc_anon
, ab
, hash_lock
);
1922 mutex_exit(hash_lock
);
1923 arc_hdr_destroy(ab
);
1926 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, ab
);
1927 if (bytes
>= 0 && bytes_deleted
>= bytes
)
1929 } else if (bytes
< 0) {
1931 * Insert a list marker and then wait for the
1932 * hash lock to become available. Once its
1933 * available, restart from where we left off.
1935 list_insert_after(list
, ab
, &marker
);
1936 mutex_exit(&state
->arcs_mtx
);
1937 mutex_enter(hash_lock
);
1938 mutex_exit(hash_lock
);
1939 mutex_enter(&state
->arcs_mtx
);
1940 ab_prev
= list_prev(list
, &marker
);
1941 list_remove(list
, &marker
);
1945 mutex_exit(&state
->arcs_mtx
);
1947 if (list
== &state
->arcs_list
[ARC_BUFC_DATA
] &&
1948 (bytes
< 0 || bytes_deleted
< bytes
)) {
1949 list
= &state
->arcs_list
[ARC_BUFC_METADATA
];
1954 ARCSTAT_INCR(arcstat_mutex_miss
, bufs_skipped
);
1958 if (bytes_deleted
< bytes
)
1959 dprintf("only deleted %lld bytes from %p\n",
1960 (longlong_t
)bytes_deleted
, state
);
1966 int64_t adjustment
, delta
;
1972 adjustment
= MIN((int64_t)(arc_size
- arc_c
),
1973 (int64_t)(arc_anon
->arcs_size
+ arc_mru
->arcs_size
+ arc_meta_used
-
1976 if (adjustment
> 0 && arc_mru
->arcs_lsize
[ARC_BUFC_DATA
] > 0) {
1977 delta
= MIN(arc_mru
->arcs_lsize
[ARC_BUFC_DATA
], adjustment
);
1978 (void) arc_evict(arc_mru
, 0, delta
, FALSE
, ARC_BUFC_DATA
);
1979 adjustment
-= delta
;
1982 if (adjustment
> 0 && arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] > 0) {
1983 delta
= MIN(arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
], adjustment
);
1984 (void) arc_evict(arc_mru
, 0, delta
, FALSE
,
1992 adjustment
= arc_size
- arc_c
;
1994 if (adjustment
> 0 && arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
] > 0) {
1995 delta
= MIN(adjustment
, arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
]);
1996 (void) arc_evict(arc_mfu
, 0, delta
, FALSE
, ARC_BUFC_DATA
);
1997 adjustment
-= delta
;
2000 if (adjustment
> 0 && arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
] > 0) {
2001 int64_t delta
= MIN(adjustment
,
2002 arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
]);
2003 (void) arc_evict(arc_mfu
, 0, delta
, FALSE
,
2008 * Adjust ghost lists
2011 adjustment
= arc_mru
->arcs_size
+ arc_mru_ghost
->arcs_size
- arc_c
;
2013 if (adjustment
> 0 && arc_mru_ghost
->arcs_size
> 0) {
2014 delta
= MIN(arc_mru_ghost
->arcs_size
, adjustment
);
2015 arc_evict_ghost(arc_mru_ghost
, 0, delta
);
2019 arc_mru_ghost
->arcs_size
+ arc_mfu_ghost
->arcs_size
- arc_c
;
2021 if (adjustment
> 0 && arc_mfu_ghost
->arcs_size
> 0) {
2022 delta
= MIN(arc_mfu_ghost
->arcs_size
, adjustment
);
2023 arc_evict_ghost(arc_mfu_ghost
, 0, delta
);
2028 * Request that arc user drop references so that N bytes can be released
2029 * from the cache. This provides a mechanism to ensure the arc can honor
2030 * the arc_meta_limit and reclaim buffers which are pinned in the cache
2031 * by higher layers. (i.e. the zpl)
2034 arc_do_user_prune(int64_t adjustment
)
2036 arc_prune_func_t
*func
;
2038 arc_prune_t
*cp
, *np
;
2040 mutex_enter(&arc_prune_mtx
);
2042 cp
= list_head(&arc_prune_list
);
2043 while (cp
!= NULL
) {
2045 private = cp
->p_private
;
2046 np
= list_next(&arc_prune_list
, cp
);
2047 refcount_add(&cp
->p_refcnt
, func
);
2048 mutex_exit(&arc_prune_mtx
);
2051 func(adjustment
, private);
2053 mutex_enter(&arc_prune_mtx
);
2055 /* User removed prune callback concurrently with execution */
2056 if (refcount_remove(&cp
->p_refcnt
, func
) == 0) {
2057 ASSERT(!list_link_active(&cp
->p_node
));
2058 refcount_destroy(&cp
->p_refcnt
);
2059 kmem_free(cp
, sizeof (*cp
));
2065 ARCSTAT_BUMP(arcstat_prune
);
2066 mutex_exit(&arc_prune_mtx
);
2070 arc_do_user_evicts(void)
2072 mutex_enter(&arc_eviction_mtx
);
2073 while (arc_eviction_list
!= NULL
) {
2074 arc_buf_t
*buf
= arc_eviction_list
;
2075 arc_eviction_list
= buf
->b_next
;
2076 mutex_enter(&buf
->b_evict_lock
);
2078 mutex_exit(&buf
->b_evict_lock
);
2079 mutex_exit(&arc_eviction_mtx
);
2081 if (buf
->b_efunc
!= NULL
)
2082 VERIFY(buf
->b_efunc(buf
) == 0);
2084 buf
->b_efunc
= NULL
;
2085 buf
->b_private
= NULL
;
2086 kmem_cache_free(buf_cache
, buf
);
2087 mutex_enter(&arc_eviction_mtx
);
2089 mutex_exit(&arc_eviction_mtx
);
2093 * Evict only meta data objects from the cache leaving the data objects.
2094 * This is only used to enforce the tunable arc_meta_limit, if we are
2095 * unable to evict enough buffers notify the user via the prune callback.
2098 arc_adjust_meta(int64_t adjustment
, boolean_t may_prune
)
2102 if (adjustment
> 0 && arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] > 0) {
2103 delta
= MIN(arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
], adjustment
);
2104 arc_evict(arc_mru
, 0, delta
, FALSE
, ARC_BUFC_METADATA
);
2105 adjustment
-= delta
;
2108 if (adjustment
> 0 && arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
] > 0) {
2109 delta
= MIN(arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
], adjustment
);
2110 arc_evict(arc_mfu
, 0, delta
, FALSE
, ARC_BUFC_METADATA
);
2111 adjustment
-= delta
;
2114 if (may_prune
&& (adjustment
> 0) && (arc_meta_used
> arc_meta_limit
))
2115 arc_do_user_prune(arc_meta_prune
);
2119 * Flush all *evictable* data from the cache for the given spa.
2120 * NOTE: this will not touch "active" (i.e. referenced) data.
2123 arc_flush(spa_t
*spa
)
2128 guid
= spa_load_guid(spa
);
2130 while (list_head(&arc_mru
->arcs_list
[ARC_BUFC_DATA
])) {
2131 (void) arc_evict(arc_mru
, guid
, -1, FALSE
, ARC_BUFC_DATA
);
2135 while (list_head(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
])) {
2136 (void) arc_evict(arc_mru
, guid
, -1, FALSE
, ARC_BUFC_METADATA
);
2140 while (list_head(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
])) {
2141 (void) arc_evict(arc_mfu
, guid
, -1, FALSE
, ARC_BUFC_DATA
);
2145 while (list_head(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
])) {
2146 (void) arc_evict(arc_mfu
, guid
, -1, FALSE
, ARC_BUFC_METADATA
);
2151 arc_evict_ghost(arc_mru_ghost
, guid
, -1);
2152 arc_evict_ghost(arc_mfu_ghost
, guid
, -1);
2154 mutex_enter(&arc_reclaim_thr_lock
);
2155 arc_do_user_evicts();
2156 mutex_exit(&arc_reclaim_thr_lock
);
2157 ASSERT(spa
|| arc_eviction_list
== NULL
);
2161 arc_shrink(uint64_t bytes
)
2163 if (arc_c
> arc_c_min
) {
2166 to_free
= bytes
? bytes
: arc_c
>> arc_shrink_shift
;
2168 if (arc_c
> arc_c_min
+ to_free
)
2169 atomic_add_64(&arc_c
, -to_free
);
2173 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
2174 if (arc_c
> arc_size
)
2175 arc_c
= MAX(arc_size
, arc_c_min
);
2177 arc_p
= (arc_c
>> 1);
2178 ASSERT(arc_c
>= arc_c_min
);
2179 ASSERT((int64_t)arc_p
>= 0);
2182 if (arc_size
> arc_c
)
2187 arc_kmem_reap_now(arc_reclaim_strategy_t strat
, uint64_t bytes
)
2190 kmem_cache_t
*prev_cache
= NULL
;
2191 kmem_cache_t
*prev_data_cache
= NULL
;
2192 extern kmem_cache_t
*zio_buf_cache
[];
2193 extern kmem_cache_t
*zio_data_buf_cache
[];
2196 * An aggressive reclamation will shrink the cache size as well as
2197 * reap free buffers from the arc kmem caches.
2199 if (strat
== ARC_RECLAIM_AGGR
)
2202 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
2203 if (zio_buf_cache
[i
] != prev_cache
) {
2204 prev_cache
= zio_buf_cache
[i
];
2205 kmem_cache_reap_now(zio_buf_cache
[i
]);
2207 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
2208 prev_data_cache
= zio_data_buf_cache
[i
];
2209 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
2213 kmem_cache_reap_now(buf_cache
);
2214 kmem_cache_reap_now(hdr_cache
);
2218 * Unlike other ZFS implementations this thread is only responsible for
2219 * adapting the target ARC size on Linux. The responsibility for memory
2220 * reclamation has been entirely delegated to the arc_shrinker_func()
2221 * which is registered with the VM. To reflect this change in behavior
2222 * the arc_reclaim thread has been renamed to arc_adapt.
2225 arc_adapt_thread(void)
2230 CALLB_CPR_INIT(&cpr
, &arc_reclaim_thr_lock
, callb_generic_cpr
, FTAG
);
2232 mutex_enter(&arc_reclaim_thr_lock
);
2233 while (arc_thread_exit
== 0) {
2235 arc_reclaim_strategy_t last_reclaim
= ARC_RECLAIM_CONS
;
2237 if (spa_get_random(100) == 0) {
2240 if (last_reclaim
== ARC_RECLAIM_CONS
) {
2241 last_reclaim
= ARC_RECLAIM_AGGR
;
2243 last_reclaim
= ARC_RECLAIM_CONS
;
2247 last_reclaim
= ARC_RECLAIM_AGGR
;
2251 /* reset the growth delay for every reclaim */
2252 arc_grow_time
= ddi_get_lbolt()+(arc_grow_retry
* hz
);
2254 arc_kmem_reap_now(last_reclaim
, 0);
2257 #endif /* !_KERNEL */
2259 /* No recent memory pressure allow the ARC to grow. */
2260 if (arc_no_grow
&& ddi_get_lbolt() >= arc_grow_time
)
2261 arc_no_grow
= FALSE
;
2264 * Keep meta data usage within limits, arc_shrink() is not
2265 * used to avoid collapsing the arc_c value when only the
2266 * arc_meta_limit is being exceeded.
2268 prune
= (int64_t)arc_meta_used
- (int64_t)arc_meta_limit
;
2270 arc_adjust_meta(prune
, B_TRUE
);
2274 if (arc_eviction_list
!= NULL
)
2275 arc_do_user_evicts();
2277 /* block until needed, or one second, whichever is shorter */
2278 CALLB_CPR_SAFE_BEGIN(&cpr
);
2279 (void) cv_timedwait_interruptible(&arc_reclaim_thr_cv
,
2280 &arc_reclaim_thr_lock
, (ddi_get_lbolt() + hz
));
2281 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_thr_lock
);
2284 arc_thread_exit
= 0;
2285 cv_broadcast(&arc_reclaim_thr_cv
);
2286 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_thr_lock */
2292 * Determine the amount of memory eligible for eviction contained in the
2293 * ARC. All clean data reported by the ghost lists can always be safely
2294 * evicted. Due to arc_c_min, the same does not hold for all clean data
2295 * contained by the regular mru and mfu lists.
2297 * In the case of the regular mru and mfu lists, we need to report as
2298 * much clean data as possible, such that evicting that same reported
2299 * data will not bring arc_size below arc_c_min. Thus, in certain
2300 * circumstances, the total amount of clean data in the mru and mfu
2301 * lists might not actually be evictable.
2303 * The following two distinct cases are accounted for:
2305 * 1. The sum of the amount of dirty data contained by both the mru and
2306 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2307 * is greater than or equal to arc_c_min.
2308 * (i.e. amount of dirty data >= arc_c_min)
2310 * This is the easy case; all clean data contained by the mru and mfu
2311 * lists is evictable. Evicting all clean data can only drop arc_size
2312 * to the amount of dirty data, which is greater than arc_c_min.
2314 * 2. The sum of the amount of dirty data contained by both the mru and
2315 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2316 * is less than arc_c_min.
2317 * (i.e. arc_c_min > amount of dirty data)
2319 * 2.1. arc_size is greater than or equal arc_c_min.
2320 * (i.e. arc_size >= arc_c_min > amount of dirty data)
2322 * In this case, not all clean data from the regular mru and mfu
2323 * lists is actually evictable; we must leave enough clean data
2324 * to keep arc_size above arc_c_min. Thus, the maximum amount of
2325 * evictable data from the two lists combined, is exactly the
2326 * difference between arc_size and arc_c_min.
2328 * 2.2. arc_size is less than arc_c_min
2329 * (i.e. arc_c_min > arc_size > amount of dirty data)
2331 * In this case, none of the data contained in the mru and mfu
2332 * lists is evictable, even if it's clean. Since arc_size is
2333 * already below arc_c_min, evicting any more would only
2334 * increase this negative difference.
2337 arc_evictable_memory(void) {
2338 uint64_t arc_clean
=
2339 arc_mru
->arcs_lsize
[ARC_BUFC_DATA
] +
2340 arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] +
2341 arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
] +
2342 arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
];
2343 uint64_t ghost_clean
=
2344 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
2345 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_METADATA
] +
2346 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
2347 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_METADATA
];
2348 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
2350 if (arc_dirty
>= arc_c_min
)
2351 return (ghost_clean
+ arc_clean
);
2353 return (ghost_clean
+ MAX((int64_t)arc_size
- (int64_t)arc_c_min
, 0));
2357 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
2361 /* The arc is considered warm once reclaim has occurred */
2362 if (unlikely(arc_warm
== B_FALSE
))
2365 /* Return the potential number of reclaimable pages */
2366 pages
= btop(arc_evictable_memory());
2367 if (sc
->nr_to_scan
== 0)
2370 /* Not allowed to perform filesystem reclaim */
2371 if (!(sc
->gfp_mask
& __GFP_FS
))
2374 /* Reclaim in progress */
2375 if (mutex_tryenter(&arc_reclaim_thr_lock
) == 0)
2379 * Evict the requested number of pages by shrinking arc_c the
2380 * requested amount. If there is nothing left to evict just
2381 * reap whatever we can from the various arc slabs.
2384 arc_kmem_reap_now(ARC_RECLAIM_AGGR
, ptob(sc
->nr_to_scan
));
2385 pages
= btop(arc_evictable_memory());
2387 arc_kmem_reap_now(ARC_RECLAIM_CONS
, ptob(sc
->nr_to_scan
));
2392 * When direct reclaim is observed it usually indicates a rapid
2393 * increase in memory pressure. This occurs because the kswapd
2394 * threads were unable to asynchronously keep enough free memory
2395 * available. In this case set arc_no_grow to briefly pause arc
2396 * growth to avoid compounding the memory pressure.
2398 if (current_is_kswapd()) {
2399 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
2401 arc_no_grow
= B_TRUE
;
2402 arc_grow_time
= ddi_get_lbolt() + (arc_grow_retry
* hz
);
2403 ARCSTAT_BUMP(arcstat_memory_direct_count
);
2406 mutex_exit(&arc_reclaim_thr_lock
);
2410 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
2412 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
2413 #endif /* _KERNEL */
2416 * Adapt arc info given the number of bytes we are trying to add and
2417 * the state that we are comming from. This function is only called
2418 * when we are adding new content to the cache.
2421 arc_adapt(int bytes
, arc_state_t
*state
)
2424 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
2426 if (state
== arc_l2c_only
)
2431 * Adapt the target size of the MRU list:
2432 * - if we just hit in the MRU ghost list, then increase
2433 * the target size of the MRU list.
2434 * - if we just hit in the MFU ghost list, then increase
2435 * the target size of the MFU list by decreasing the
2436 * target size of the MRU list.
2438 if (state
== arc_mru_ghost
) {
2439 mult
= ((arc_mru_ghost
->arcs_size
>= arc_mfu_ghost
->arcs_size
) ?
2440 1 : (arc_mfu_ghost
->arcs_size
/arc_mru_ghost
->arcs_size
));
2441 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
2443 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
2444 } else if (state
== arc_mfu_ghost
) {
2447 mult
= ((arc_mfu_ghost
->arcs_size
>= arc_mru_ghost
->arcs_size
) ?
2448 1 : (arc_mru_ghost
->arcs_size
/arc_mfu_ghost
->arcs_size
));
2449 mult
= MIN(mult
, 10);
2451 delta
= MIN(bytes
* mult
, arc_p
);
2452 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
2454 ASSERT((int64_t)arc_p
>= 0);
2459 if (arc_c
>= arc_c_max
)
2463 * If we're within (2 * maxblocksize) bytes of the target
2464 * cache size, increment the target cache size
2466 if (arc_size
> arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
2467 atomic_add_64(&arc_c
, (int64_t)bytes
);
2468 if (arc_c
> arc_c_max
)
2470 else if (state
== arc_anon
)
2471 atomic_add_64(&arc_p
, (int64_t)bytes
);
2475 ASSERT((int64_t)arc_p
>= 0);
2479 * Check if the cache has reached its limits and eviction is required
2483 arc_evict_needed(arc_buf_contents_t type
)
2485 if (type
== ARC_BUFC_METADATA
&& arc_meta_used
>= arc_meta_limit
)
2491 return (arc_size
> arc_c
);
2495 * The buffer, supplied as the first argument, needs a data block.
2496 * So, if we are at cache max, determine which cache should be victimized.
2497 * We have the following cases:
2499 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2500 * In this situation if we're out of space, but the resident size of the MFU is
2501 * under the limit, victimize the MFU cache to satisfy this insertion request.
2503 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2504 * Here, we've used up all of the available space for the MRU, so we need to
2505 * evict from our own cache instead. Evict from the set of resident MRU
2508 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2509 * c minus p represents the MFU space in the cache, since p is the size of the
2510 * cache that is dedicated to the MRU. In this situation there's still space on
2511 * the MFU side, so the MRU side needs to be victimized.
2513 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2514 * MFU's resident set is consuming more space than it has been allotted. In
2515 * this situation, we must victimize our own cache, the MFU, for this insertion.
2518 arc_get_data_buf(arc_buf_t
*buf
)
2520 arc_state_t
*state
= buf
->b_hdr
->b_state
;
2521 uint64_t size
= buf
->b_hdr
->b_size
;
2522 arc_buf_contents_t type
= buf
->b_hdr
->b_type
;
2524 arc_adapt(size
, state
);
2527 * We have not yet reached cache maximum size,
2528 * just allocate a new buffer.
2530 if (!arc_evict_needed(type
)) {
2531 if (type
== ARC_BUFC_METADATA
) {
2532 buf
->b_data
= zio_buf_alloc(size
);
2533 arc_space_consume(size
, ARC_SPACE_DATA
);
2535 ASSERT(type
== ARC_BUFC_DATA
);
2536 buf
->b_data
= zio_data_buf_alloc(size
);
2537 ARCSTAT_INCR(arcstat_data_size
, size
);
2538 atomic_add_64(&arc_size
, size
);
2544 * If we are prefetching from the mfu ghost list, this buffer
2545 * will end up on the mru list; so steal space from there.
2547 if (state
== arc_mfu_ghost
)
2548 state
= buf
->b_hdr
->b_flags
& ARC_PREFETCH
? arc_mru
: arc_mfu
;
2549 else if (state
== arc_mru_ghost
)
2552 if (state
== arc_mru
|| state
== arc_anon
) {
2553 uint64_t mru_used
= arc_anon
->arcs_size
+ arc_mru
->arcs_size
;
2554 state
= (arc_mfu
->arcs_lsize
[type
] >= size
&&
2555 arc_p
> mru_used
) ? arc_mfu
: arc_mru
;
2558 uint64_t mfu_space
= arc_c
- arc_p
;
2559 state
= (arc_mru
->arcs_lsize
[type
] >= size
&&
2560 mfu_space
> arc_mfu
->arcs_size
) ? arc_mru
: arc_mfu
;
2563 if ((buf
->b_data
= arc_evict(state
, 0, size
, TRUE
, type
)) == NULL
) {
2564 if (type
== ARC_BUFC_METADATA
) {
2565 buf
->b_data
= zio_buf_alloc(size
);
2566 arc_space_consume(size
, ARC_SPACE_DATA
);
2569 * If we are unable to recycle an existing meta buffer
2570 * signal the reclaim thread. It will notify users
2571 * via the prune callback to drop references. The
2572 * prune callback in run in the context of the reclaim
2573 * thread to avoid deadlocking on the hash_lock.
2575 cv_signal(&arc_reclaim_thr_cv
);
2577 ASSERT(type
== ARC_BUFC_DATA
);
2578 buf
->b_data
= zio_data_buf_alloc(size
);
2579 ARCSTAT_INCR(arcstat_data_size
, size
);
2580 atomic_add_64(&arc_size
, size
);
2583 ARCSTAT_BUMP(arcstat_recycle_miss
);
2585 ASSERT(buf
->b_data
!= NULL
);
2588 * Update the state size. Note that ghost states have a
2589 * "ghost size" and so don't need to be updated.
2591 if (!GHOST_STATE(buf
->b_hdr
->b_state
)) {
2592 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2594 atomic_add_64(&hdr
->b_state
->arcs_size
, size
);
2595 if (list_link_active(&hdr
->b_arc_node
)) {
2596 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
2597 atomic_add_64(&hdr
->b_state
->arcs_lsize
[type
], size
);
2600 * If we are growing the cache, and we are adding anonymous
2601 * data, and we have outgrown arc_p, update arc_p
2603 if (arc_size
< arc_c
&& hdr
->b_state
== arc_anon
&&
2604 arc_anon
->arcs_size
+ arc_mru
->arcs_size
> arc_p
)
2605 arc_p
= MIN(arc_c
, arc_p
+ size
);
2610 * This routine is called whenever a buffer is accessed.
2611 * NOTE: the hash lock is dropped in this function.
2614 arc_access(arc_buf_hdr_t
*buf
, kmutex_t
*hash_lock
)
2618 ASSERT(MUTEX_HELD(hash_lock
));
2620 if (buf
->b_state
== arc_anon
) {
2622 * This buffer is not in the cache, and does not
2623 * appear in our "ghost" list. Add the new buffer
2627 ASSERT(buf
->b_arc_access
== 0);
2628 buf
->b_arc_access
= ddi_get_lbolt();
2629 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, buf
);
2630 arc_change_state(arc_mru
, buf
, hash_lock
);
2632 } else if (buf
->b_state
== arc_mru
) {
2633 now
= ddi_get_lbolt();
2636 * If this buffer is here because of a prefetch, then either:
2637 * - clear the flag if this is a "referencing" read
2638 * (any subsequent access will bump this into the MFU state).
2640 * - move the buffer to the head of the list if this is
2641 * another prefetch (to make it less likely to be evicted).
2643 if ((buf
->b_flags
& ARC_PREFETCH
) != 0) {
2644 if (refcount_count(&buf
->b_refcnt
) == 0) {
2645 ASSERT(list_link_active(&buf
->b_arc_node
));
2647 buf
->b_flags
&= ~ARC_PREFETCH
;
2648 ARCSTAT_BUMP(arcstat_mru_hits
);
2650 buf
->b_arc_access
= now
;
2655 * This buffer has been "accessed" only once so far,
2656 * but it is still in the cache. Move it to the MFU
2659 if (now
> buf
->b_arc_access
+ ARC_MINTIME
) {
2661 * More than 125ms have passed since we
2662 * instantiated this buffer. Move it to the
2663 * most frequently used state.
2665 buf
->b_arc_access
= now
;
2666 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2667 arc_change_state(arc_mfu
, buf
, hash_lock
);
2669 ARCSTAT_BUMP(arcstat_mru_hits
);
2670 } else if (buf
->b_state
== arc_mru_ghost
) {
2671 arc_state_t
*new_state
;
2673 * This buffer has been "accessed" recently, but
2674 * was evicted from the cache. Move it to the
2678 if (buf
->b_flags
& ARC_PREFETCH
) {
2679 new_state
= arc_mru
;
2680 if (refcount_count(&buf
->b_refcnt
) > 0)
2681 buf
->b_flags
&= ~ARC_PREFETCH
;
2682 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, buf
);
2684 new_state
= arc_mfu
;
2685 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2688 buf
->b_arc_access
= ddi_get_lbolt();
2689 arc_change_state(new_state
, buf
, hash_lock
);
2691 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
2692 } else if (buf
->b_state
== arc_mfu
) {
2694 * This buffer has been accessed more than once and is
2695 * still in the cache. Keep it in the MFU state.
2697 * NOTE: an add_reference() that occurred when we did
2698 * the arc_read() will have kicked this off the list.
2699 * If it was a prefetch, we will explicitly move it to
2700 * the head of the list now.
2702 if ((buf
->b_flags
& ARC_PREFETCH
) != 0) {
2703 ASSERT(refcount_count(&buf
->b_refcnt
) == 0);
2704 ASSERT(list_link_active(&buf
->b_arc_node
));
2706 ARCSTAT_BUMP(arcstat_mfu_hits
);
2707 buf
->b_arc_access
= ddi_get_lbolt();
2708 } else if (buf
->b_state
== arc_mfu_ghost
) {
2709 arc_state_t
*new_state
= arc_mfu
;
2711 * This buffer has been accessed more than once but has
2712 * been evicted from the cache. Move it back to the
2716 if (buf
->b_flags
& ARC_PREFETCH
) {
2718 * This is a prefetch access...
2719 * move this block back to the MRU state.
2721 ASSERT0(refcount_count(&buf
->b_refcnt
));
2722 new_state
= arc_mru
;
2725 buf
->b_arc_access
= ddi_get_lbolt();
2726 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2727 arc_change_state(new_state
, buf
, hash_lock
);
2729 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
2730 } else if (buf
->b_state
== arc_l2c_only
) {
2732 * This buffer is on the 2nd Level ARC.
2735 buf
->b_arc_access
= ddi_get_lbolt();
2736 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2737 arc_change_state(arc_mfu
, buf
, hash_lock
);
2739 ASSERT(!"invalid arc state");
2743 /* a generic arc_done_func_t which you can use */
2746 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
2748 if (zio
== NULL
|| zio
->io_error
== 0)
2749 bcopy(buf
->b_data
, arg
, buf
->b_hdr
->b_size
);
2750 VERIFY(arc_buf_remove_ref(buf
, arg
) == 1);
2753 /* a generic arc_done_func_t */
2755 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
2757 arc_buf_t
**bufp
= arg
;
2758 if (zio
&& zio
->io_error
) {
2759 VERIFY(arc_buf_remove_ref(buf
, arg
) == 1);
2763 ASSERT(buf
->b_data
);
2768 arc_read_done(zio_t
*zio
)
2770 arc_buf_hdr_t
*hdr
, *found
;
2772 arc_buf_t
*abuf
; /* buffer we're assigning to callback */
2773 kmutex_t
*hash_lock
;
2774 arc_callback_t
*callback_list
, *acb
;
2775 int freeable
= FALSE
;
2777 buf
= zio
->io_private
;
2781 * The hdr was inserted into hash-table and removed from lists
2782 * prior to starting I/O. We should find this header, since
2783 * it's in the hash table, and it should be legit since it's
2784 * not possible to evict it during the I/O. The only possible
2785 * reason for it not to be found is if we were freed during the
2788 found
= buf_hash_find(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
,
2791 ASSERT((found
== NULL
&& HDR_FREED_IN_READ(hdr
) && hash_lock
== NULL
) ||
2792 (found
== hdr
&& DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
2793 (found
== hdr
&& HDR_L2_READING(hdr
)));
2795 hdr
->b_flags
&= ~ARC_L2_EVICTED
;
2796 if (l2arc_noprefetch
&& (hdr
->b_flags
& ARC_PREFETCH
))
2797 hdr
->b_flags
&= ~ARC_L2CACHE
;
2799 /* byteswap if necessary */
2800 callback_list
= hdr
->b_acb
;
2801 ASSERT(callback_list
!= NULL
);
2802 if (BP_SHOULD_BYTESWAP(zio
->io_bp
) && zio
->io_error
== 0) {
2803 dmu_object_byteswap_t bswap
=
2804 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
2805 if (BP_GET_LEVEL(zio
->io_bp
) > 0)
2806 byteswap_uint64_array(buf
->b_data
, hdr
->b_size
);
2808 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, hdr
->b_size
);
2811 arc_cksum_compute(buf
, B_FALSE
);
2813 if (hash_lock
&& zio
->io_error
== 0 && hdr
->b_state
== arc_anon
) {
2815 * Only call arc_access on anonymous buffers. This is because
2816 * if we've issued an I/O for an evicted buffer, we've already
2817 * called arc_access (to prevent any simultaneous readers from
2818 * getting confused).
2820 arc_access(hdr
, hash_lock
);
2823 /* create copies of the data buffer for the callers */
2825 for (acb
= callback_list
; acb
; acb
= acb
->acb_next
) {
2826 if (acb
->acb_done
) {
2828 ARCSTAT_BUMP(arcstat_duplicate_reads
);
2829 abuf
= arc_buf_clone(buf
);
2831 acb
->acb_buf
= abuf
;
2836 hdr
->b_flags
&= ~ARC_IO_IN_PROGRESS
;
2837 ASSERT(!HDR_BUF_AVAILABLE(hdr
));
2839 ASSERT(buf
->b_efunc
== NULL
);
2840 ASSERT(hdr
->b_datacnt
== 1);
2841 hdr
->b_flags
|= ARC_BUF_AVAILABLE
;
2844 ASSERT(refcount_is_zero(&hdr
->b_refcnt
) || callback_list
!= NULL
);
2846 if (zio
->io_error
!= 0) {
2847 hdr
->b_flags
|= ARC_IO_ERROR
;
2848 if (hdr
->b_state
!= arc_anon
)
2849 arc_change_state(arc_anon
, hdr
, hash_lock
);
2850 if (HDR_IN_HASH_TABLE(hdr
))
2851 buf_hash_remove(hdr
);
2852 freeable
= refcount_is_zero(&hdr
->b_refcnt
);
2856 * Broadcast before we drop the hash_lock to avoid the possibility
2857 * that the hdr (and hence the cv) might be freed before we get to
2858 * the cv_broadcast().
2860 cv_broadcast(&hdr
->b_cv
);
2863 mutex_exit(hash_lock
);
2866 * This block was freed while we waited for the read to
2867 * complete. It has been removed from the hash table and
2868 * moved to the anonymous state (so that it won't show up
2871 ASSERT3P(hdr
->b_state
, ==, arc_anon
);
2872 freeable
= refcount_is_zero(&hdr
->b_refcnt
);
2875 /* execute each callback and free its structure */
2876 while ((acb
= callback_list
) != NULL
) {
2878 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
2880 if (acb
->acb_zio_dummy
!= NULL
) {
2881 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
2882 zio_nowait(acb
->acb_zio_dummy
);
2885 callback_list
= acb
->acb_next
;
2886 kmem_free(acb
, sizeof (arc_callback_t
));
2890 arc_hdr_destroy(hdr
);
2894 * "Read" the block at the specified DVA (in bp) via the
2895 * cache. If the block is found in the cache, invoke the provided
2896 * callback immediately and return. Note that the `zio' parameter
2897 * in the callback will be NULL in this case, since no IO was
2898 * required. If the block is not in the cache pass the read request
2899 * on to the spa with a substitute callback function, so that the
2900 * requested block will be added to the cache.
2902 * If a read request arrives for a block that has a read in-progress,
2903 * either wait for the in-progress read to complete (and return the
2904 * results); or, if this is a read with a "done" func, add a record
2905 * to the read to invoke the "done" func when the read completes,
2906 * and return; or just return.
2908 * arc_read_done() will invoke all the requested "done" functions
2909 * for readers of this block.
2911 * Normal callers should use arc_read and pass the arc buffer and offset
2912 * for the bp. But if you know you don't need locking, you can use
2916 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_buf_t
*pbuf
,
2917 arc_done_func_t
*done
, void *private, int priority
, int zio_flags
,
2918 uint32_t *arc_flags
, const zbookmark_t
*zb
)
2924 * XXX This happens from traverse callback funcs, for
2925 * the objset_phys_t block.
2927 return (arc_read_nolock(pio
, spa
, bp
, done
, private, priority
,
2928 zio_flags
, arc_flags
, zb
));
2931 ASSERT(!refcount_is_zero(&pbuf
->b_hdr
->b_refcnt
));
2932 ASSERT3U((char *)bp
- (char *)pbuf
->b_data
, <, pbuf
->b_hdr
->b_size
);
2933 rw_enter(&pbuf
->b_data_lock
, RW_READER
);
2935 err
= arc_read_nolock(pio
, spa
, bp
, done
, private, priority
,
2936 zio_flags
, arc_flags
, zb
);
2937 rw_exit(&pbuf
->b_data_lock
);
2943 arc_read_nolock(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
2944 arc_done_func_t
*done
, void *private, int priority
, int zio_flags
,
2945 uint32_t *arc_flags
, const zbookmark_t
*zb
)
2948 arc_buf_t
*buf
= NULL
;
2949 kmutex_t
*hash_lock
;
2951 uint64_t guid
= spa_load_guid(spa
);
2954 hdr
= buf_hash_find(guid
, BP_IDENTITY(bp
), BP_PHYSICAL_BIRTH(bp
),
2956 if (hdr
&& hdr
->b_datacnt
> 0) {
2958 *arc_flags
|= ARC_CACHED
;
2960 if (HDR_IO_IN_PROGRESS(hdr
)) {
2962 if (*arc_flags
& ARC_WAIT
) {
2963 cv_wait(&hdr
->b_cv
, hash_lock
);
2964 mutex_exit(hash_lock
);
2967 ASSERT(*arc_flags
& ARC_NOWAIT
);
2970 arc_callback_t
*acb
= NULL
;
2972 acb
= kmem_zalloc(sizeof (arc_callback_t
),
2974 acb
->acb_done
= done
;
2975 acb
->acb_private
= private;
2977 acb
->acb_zio_dummy
= zio_null(pio
,
2978 spa
, NULL
, NULL
, NULL
, zio_flags
);
2980 ASSERT(acb
->acb_done
!= NULL
);
2981 acb
->acb_next
= hdr
->b_acb
;
2983 add_reference(hdr
, hash_lock
, private);
2984 mutex_exit(hash_lock
);
2987 mutex_exit(hash_lock
);
2991 ASSERT(hdr
->b_state
== arc_mru
|| hdr
->b_state
== arc_mfu
);
2994 add_reference(hdr
, hash_lock
, private);
2996 * If this block is already in use, create a new
2997 * copy of the data so that we will be guaranteed
2998 * that arc_release() will always succeed.
3002 ASSERT(buf
->b_data
);
3003 if (HDR_BUF_AVAILABLE(hdr
)) {
3004 ASSERT(buf
->b_efunc
== NULL
);
3005 hdr
->b_flags
&= ~ARC_BUF_AVAILABLE
;
3007 buf
= arc_buf_clone(buf
);
3010 } else if (*arc_flags
& ARC_PREFETCH
&&
3011 refcount_count(&hdr
->b_refcnt
) == 0) {
3012 hdr
->b_flags
|= ARC_PREFETCH
;
3014 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
3015 arc_access(hdr
, hash_lock
);
3016 if (*arc_flags
& ARC_L2CACHE
)
3017 hdr
->b_flags
|= ARC_L2CACHE
;
3018 mutex_exit(hash_lock
);
3019 ARCSTAT_BUMP(arcstat_hits
);
3020 ARCSTAT_CONDSTAT(!(hdr
->b_flags
& ARC_PREFETCH
),
3021 demand
, prefetch
, hdr
->b_type
!= ARC_BUFC_METADATA
,
3022 data
, metadata
, hits
);
3025 done(NULL
, buf
, private);
3027 uint64_t size
= BP_GET_LSIZE(bp
);
3028 arc_callback_t
*acb
;
3031 boolean_t devw
= B_FALSE
;
3034 /* this block is not in the cache */
3035 arc_buf_hdr_t
*exists
;
3036 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
3037 buf
= arc_buf_alloc(spa
, size
, private, type
);
3039 hdr
->b_dva
= *BP_IDENTITY(bp
);
3040 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
3041 hdr
->b_cksum0
= bp
->blk_cksum
.zc_word
[0];
3042 exists
= buf_hash_insert(hdr
, &hash_lock
);
3044 /* somebody beat us to the hash insert */
3045 mutex_exit(hash_lock
);
3046 buf_discard_identity(hdr
);
3047 (void) arc_buf_remove_ref(buf
, private);
3048 goto top
; /* restart the IO request */
3050 /* if this is a prefetch, we don't have a reference */
3051 if (*arc_flags
& ARC_PREFETCH
) {
3052 (void) remove_reference(hdr
, hash_lock
,
3054 hdr
->b_flags
|= ARC_PREFETCH
;
3056 if (*arc_flags
& ARC_L2CACHE
)
3057 hdr
->b_flags
|= ARC_L2CACHE
;
3058 if (BP_GET_LEVEL(bp
) > 0)
3059 hdr
->b_flags
|= ARC_INDIRECT
;
3061 /* this block is in the ghost cache */
3062 ASSERT(GHOST_STATE(hdr
->b_state
));
3063 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3064 ASSERT0(refcount_count(&hdr
->b_refcnt
));
3065 ASSERT(hdr
->b_buf
== NULL
);
3067 /* if this is a prefetch, we don't have a reference */
3068 if (*arc_flags
& ARC_PREFETCH
)
3069 hdr
->b_flags
|= ARC_PREFETCH
;
3071 add_reference(hdr
, hash_lock
, private);
3072 if (*arc_flags
& ARC_L2CACHE
)
3073 hdr
->b_flags
|= ARC_L2CACHE
;
3074 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
3077 buf
->b_efunc
= NULL
;
3078 buf
->b_private
= NULL
;
3081 ASSERT(hdr
->b_datacnt
== 0);
3083 arc_get_data_buf(buf
);
3084 arc_access(hdr
, hash_lock
);
3087 ASSERT(!GHOST_STATE(hdr
->b_state
));
3089 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_PUSHPAGE
);
3090 acb
->acb_done
= done
;
3091 acb
->acb_private
= private;
3093 ASSERT(hdr
->b_acb
== NULL
);
3095 hdr
->b_flags
|= ARC_IO_IN_PROGRESS
;
3097 if (HDR_L2CACHE(hdr
) && hdr
->b_l2hdr
!= NULL
&&
3098 (vd
= hdr
->b_l2hdr
->b_dev
->l2ad_vdev
) != NULL
) {
3099 devw
= hdr
->b_l2hdr
->b_dev
->l2ad_writing
;
3100 addr
= hdr
->b_l2hdr
->b_daddr
;
3102 * Lock out device removal.
3104 if (vdev_is_dead(vd
) ||
3105 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
3109 mutex_exit(hash_lock
);
3111 ASSERT3U(hdr
->b_size
, ==, size
);
3112 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
3113 uint64_t, size
, zbookmark_t
*, zb
);
3114 ARCSTAT_BUMP(arcstat_misses
);
3115 ARCSTAT_CONDSTAT(!(hdr
->b_flags
& ARC_PREFETCH
),
3116 demand
, prefetch
, hdr
->b_type
!= ARC_BUFC_METADATA
,
3117 data
, metadata
, misses
);
3119 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
3121 * Read from the L2ARC if the following are true:
3122 * 1. The L2ARC vdev was previously cached.
3123 * 2. This buffer still has L2ARC metadata.
3124 * 3. This buffer isn't currently writing to the L2ARC.
3125 * 4. The L2ARC entry wasn't evicted, which may
3126 * also have invalidated the vdev.
3127 * 5. This isn't prefetch and l2arc_noprefetch is set.
3129 if (hdr
->b_l2hdr
!= NULL
&&
3130 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
3131 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
3132 l2arc_read_callback_t
*cb
;
3134 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
3135 ARCSTAT_BUMP(arcstat_l2_hits
);
3137 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
3139 cb
->l2rcb_buf
= buf
;
3140 cb
->l2rcb_spa
= spa
;
3143 cb
->l2rcb_flags
= zio_flags
;
3146 * l2arc read. The SCL_L2ARC lock will be
3147 * released by l2arc_read_done().
3149 rzio
= zio_read_phys(pio
, vd
, addr
, size
,
3150 buf
->b_data
, ZIO_CHECKSUM_OFF
,
3151 l2arc_read_done
, cb
, priority
, zio_flags
|
3152 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_CANFAIL
|
3153 ZIO_FLAG_DONT_PROPAGATE
|
3154 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
3155 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
3157 ARCSTAT_INCR(arcstat_l2_read_bytes
, size
);
3159 if (*arc_flags
& ARC_NOWAIT
) {
3164 ASSERT(*arc_flags
& ARC_WAIT
);
3165 if (zio_wait(rzio
) == 0)
3168 /* l2arc read error; goto zio_read() */
3170 DTRACE_PROBE1(l2arc__miss
,
3171 arc_buf_hdr_t
*, hdr
);
3172 ARCSTAT_BUMP(arcstat_l2_misses
);
3173 if (HDR_L2_WRITING(hdr
))
3174 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
3175 spa_config_exit(spa
, SCL_L2ARC
, vd
);
3179 spa_config_exit(spa
, SCL_L2ARC
, vd
);
3180 if (l2arc_ndev
!= 0) {
3181 DTRACE_PROBE1(l2arc__miss
,
3182 arc_buf_hdr_t
*, hdr
);
3183 ARCSTAT_BUMP(arcstat_l2_misses
);
3187 rzio
= zio_read(pio
, spa
, bp
, buf
->b_data
, size
,
3188 arc_read_done
, buf
, priority
, zio_flags
, zb
);
3190 if (*arc_flags
& ARC_WAIT
)
3191 return (zio_wait(rzio
));
3193 ASSERT(*arc_flags
& ARC_NOWAIT
);
3200 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
3204 p
= kmem_alloc(sizeof(*p
), KM_SLEEP
);
3206 p
->p_private
= private;
3207 list_link_init(&p
->p_node
);
3208 refcount_create(&p
->p_refcnt
);
3210 mutex_enter(&arc_prune_mtx
);
3211 refcount_add(&p
->p_refcnt
, &arc_prune_list
);
3212 list_insert_head(&arc_prune_list
, p
);
3213 mutex_exit(&arc_prune_mtx
);
3219 arc_remove_prune_callback(arc_prune_t
*p
)
3221 mutex_enter(&arc_prune_mtx
);
3222 list_remove(&arc_prune_list
, p
);
3223 if (refcount_remove(&p
->p_refcnt
, &arc_prune_list
) == 0) {
3224 refcount_destroy(&p
->p_refcnt
);
3225 kmem_free(p
, sizeof (*p
));
3227 mutex_exit(&arc_prune_mtx
);
3231 arc_set_callback(arc_buf_t
*buf
, arc_evict_func_t
*func
, void *private)
3233 ASSERT(buf
->b_hdr
!= NULL
);
3234 ASSERT(buf
->b_hdr
->b_state
!= arc_anon
);
3235 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_refcnt
) || func
== NULL
);
3236 ASSERT(buf
->b_efunc
== NULL
);
3237 ASSERT(!HDR_BUF_AVAILABLE(buf
->b_hdr
));
3239 buf
->b_efunc
= func
;
3240 buf
->b_private
= private;
3244 * This is used by the DMU to let the ARC know that a buffer is
3245 * being evicted, so the ARC should clean up. If this arc buf
3246 * is not yet in the evicted state, it will be put there.
3249 arc_buf_evict(arc_buf_t
*buf
)
3252 kmutex_t
*hash_lock
;
3255 mutex_enter(&buf
->b_evict_lock
);
3259 * We are in arc_do_user_evicts().
3261 ASSERT(buf
->b_data
== NULL
);
3262 mutex_exit(&buf
->b_evict_lock
);
3264 } else if (buf
->b_data
== NULL
) {
3265 arc_buf_t copy
= *buf
; /* structure assignment */
3267 * We are on the eviction list; process this buffer now
3268 * but let arc_do_user_evicts() do the reaping.
3270 buf
->b_efunc
= NULL
;
3271 mutex_exit(&buf
->b_evict_lock
);
3272 VERIFY(copy
.b_efunc(©
) == 0);
3275 hash_lock
= HDR_LOCK(hdr
);
3276 mutex_enter(hash_lock
);
3278 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3280 ASSERT3U(refcount_count(&hdr
->b_refcnt
), <, hdr
->b_datacnt
);
3281 ASSERT(hdr
->b_state
== arc_mru
|| hdr
->b_state
== arc_mfu
);
3284 * Pull this buffer off of the hdr
3287 while (*bufp
!= buf
)
3288 bufp
= &(*bufp
)->b_next
;
3289 *bufp
= buf
->b_next
;
3291 ASSERT(buf
->b_data
!= NULL
);
3292 arc_buf_destroy(buf
, FALSE
, FALSE
);
3294 if (hdr
->b_datacnt
== 0) {
3295 arc_state_t
*old_state
= hdr
->b_state
;
3296 arc_state_t
*evicted_state
;
3298 ASSERT(hdr
->b_buf
== NULL
);
3299 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
3302 (old_state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
3304 mutex_enter(&old_state
->arcs_mtx
);
3305 mutex_enter(&evicted_state
->arcs_mtx
);
3307 arc_change_state(evicted_state
, hdr
, hash_lock
);
3308 ASSERT(HDR_IN_HASH_TABLE(hdr
));
3309 hdr
->b_flags
|= ARC_IN_HASH_TABLE
;
3310 hdr
->b_flags
&= ~ARC_BUF_AVAILABLE
;
3312 mutex_exit(&evicted_state
->arcs_mtx
);
3313 mutex_exit(&old_state
->arcs_mtx
);
3315 mutex_exit(hash_lock
);
3316 mutex_exit(&buf
->b_evict_lock
);
3318 VERIFY(buf
->b_efunc(buf
) == 0);
3319 buf
->b_efunc
= NULL
;
3320 buf
->b_private
= NULL
;
3323 kmem_cache_free(buf_cache
, buf
);
3328 * Release this buffer from the cache. This must be done
3329 * after a read and prior to modifying the buffer contents.
3330 * If the buffer has more than one reference, we must make
3331 * a new hdr for the buffer.
3334 arc_release(arc_buf_t
*buf
, void *tag
)
3337 kmutex_t
*hash_lock
= NULL
;
3338 l2arc_buf_hdr_t
*l2hdr
;
3339 uint64_t buf_size
= 0;
3342 * It would be nice to assert that if it's DMU metadata (level >
3343 * 0 || it's the dnode file), then it must be syncing context.
3344 * But we don't know that information at this level.
3347 mutex_enter(&buf
->b_evict_lock
);
3350 /* this buffer is not on any list */
3351 ASSERT(refcount_count(&hdr
->b_refcnt
) > 0);
3353 if (hdr
->b_state
== arc_anon
) {
3354 /* this buffer is already released */
3355 ASSERT(buf
->b_efunc
== NULL
);
3357 hash_lock
= HDR_LOCK(hdr
);
3358 mutex_enter(hash_lock
);
3360 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3363 l2hdr
= hdr
->b_l2hdr
;
3365 mutex_enter(&l2arc_buflist_mtx
);
3366 hdr
->b_l2hdr
= NULL
;
3367 buf_size
= hdr
->b_size
;
3371 * Do we have more than one buf?
3373 if (hdr
->b_datacnt
> 1) {
3374 arc_buf_hdr_t
*nhdr
;
3376 uint64_t blksz
= hdr
->b_size
;
3377 uint64_t spa
= hdr
->b_spa
;
3378 arc_buf_contents_t type
= hdr
->b_type
;
3379 uint32_t flags
= hdr
->b_flags
;
3381 ASSERT(hdr
->b_buf
!= buf
|| buf
->b_next
!= NULL
);
3383 * Pull the data off of this hdr and attach it to
3384 * a new anonymous hdr.
3386 (void) remove_reference(hdr
, hash_lock
, tag
);
3388 while (*bufp
!= buf
)
3389 bufp
= &(*bufp
)->b_next
;
3390 *bufp
= buf
->b_next
;
3393 ASSERT3U(hdr
->b_state
->arcs_size
, >=, hdr
->b_size
);
3394 atomic_add_64(&hdr
->b_state
->arcs_size
, -hdr
->b_size
);
3395 if (refcount_is_zero(&hdr
->b_refcnt
)) {
3396 uint64_t *size
= &hdr
->b_state
->arcs_lsize
[hdr
->b_type
];
3397 ASSERT3U(*size
, >=, hdr
->b_size
);
3398 atomic_add_64(size
, -hdr
->b_size
);
3402 * We're releasing a duplicate user data buffer, update
3403 * our statistics accordingly.
3405 if (hdr
->b_type
== ARC_BUFC_DATA
) {
3406 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
3407 ARCSTAT_INCR(arcstat_duplicate_buffers_size
,
3410 hdr
->b_datacnt
-= 1;
3411 arc_cksum_verify(buf
);
3413 mutex_exit(hash_lock
);
3415 nhdr
= kmem_cache_alloc(hdr_cache
, KM_PUSHPAGE
);
3416 nhdr
->b_size
= blksz
;
3418 nhdr
->b_type
= type
;
3420 nhdr
->b_state
= arc_anon
;
3421 nhdr
->b_arc_access
= 0;
3422 nhdr
->b_flags
= flags
& ARC_L2_WRITING
;
3423 nhdr
->b_l2hdr
= NULL
;
3424 nhdr
->b_datacnt
= 1;
3425 nhdr
->b_freeze_cksum
= NULL
;
3426 (void) refcount_add(&nhdr
->b_refcnt
, tag
);
3428 mutex_exit(&buf
->b_evict_lock
);
3429 atomic_add_64(&arc_anon
->arcs_size
, blksz
);
3431 mutex_exit(&buf
->b_evict_lock
);
3432 ASSERT(refcount_count(&hdr
->b_refcnt
) == 1);
3433 ASSERT(!list_link_active(&hdr
->b_arc_node
));
3434 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3435 if (hdr
->b_state
!= arc_anon
)
3436 arc_change_state(arc_anon
, hdr
, hash_lock
);
3437 hdr
->b_arc_access
= 0;
3439 mutex_exit(hash_lock
);
3441 buf_discard_identity(hdr
);
3444 buf
->b_efunc
= NULL
;
3445 buf
->b_private
= NULL
;
3448 list_remove(l2hdr
->b_dev
->l2ad_buflist
, hdr
);
3449 kmem_free(l2hdr
, sizeof (l2arc_buf_hdr_t
));
3450 ARCSTAT_INCR(arcstat_l2_size
, -buf_size
);
3451 mutex_exit(&l2arc_buflist_mtx
);
3456 * Release this buffer. If it does not match the provided BP, fill it
3457 * with that block's contents.
3461 arc_release_bp(arc_buf_t
*buf
, void *tag
, blkptr_t
*bp
, spa_t
*spa
,
3464 arc_release(buf
, tag
);
3469 arc_released(arc_buf_t
*buf
)
3473 mutex_enter(&buf
->b_evict_lock
);
3474 released
= (buf
->b_data
!= NULL
&& buf
->b_hdr
->b_state
== arc_anon
);
3475 mutex_exit(&buf
->b_evict_lock
);
3480 arc_has_callback(arc_buf_t
*buf
)
3484 mutex_enter(&buf
->b_evict_lock
);
3485 callback
= (buf
->b_efunc
!= NULL
);
3486 mutex_exit(&buf
->b_evict_lock
);
3492 arc_referenced(arc_buf_t
*buf
)
3496 mutex_enter(&buf
->b_evict_lock
);
3497 referenced
= (refcount_count(&buf
->b_hdr
->b_refcnt
));
3498 mutex_exit(&buf
->b_evict_lock
);
3499 return (referenced
);
3504 arc_write_ready(zio_t
*zio
)
3506 arc_write_callback_t
*callback
= zio
->io_private
;
3507 arc_buf_t
*buf
= callback
->awcb_buf
;
3508 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3510 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_refcnt
));
3511 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
3514 * If the IO is already in progress, then this is a re-write
3515 * attempt, so we need to thaw and re-compute the cksum.
3516 * It is the responsibility of the callback to handle the
3517 * accounting for any re-write attempt.
3519 if (HDR_IO_IN_PROGRESS(hdr
)) {
3520 mutex_enter(&hdr
->b_freeze_lock
);
3521 if (hdr
->b_freeze_cksum
!= NULL
) {
3522 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
3523 hdr
->b_freeze_cksum
= NULL
;
3525 mutex_exit(&hdr
->b_freeze_lock
);
3527 arc_cksum_compute(buf
, B_FALSE
);
3528 hdr
->b_flags
|= ARC_IO_IN_PROGRESS
;
3532 arc_write_done(zio_t
*zio
)
3534 arc_write_callback_t
*callback
= zio
->io_private
;
3535 arc_buf_t
*buf
= callback
->awcb_buf
;
3536 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3538 ASSERT(hdr
->b_acb
== NULL
);
3540 if (zio
->io_error
== 0) {
3541 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
3542 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
3543 hdr
->b_cksum0
= zio
->io_bp
->blk_cksum
.zc_word
[0];
3545 ASSERT(BUF_EMPTY(hdr
));
3549 * If the block to be written was all-zero, we may have
3550 * compressed it away. In this case no write was performed
3551 * so there will be no dva/birth/checksum. The buffer must
3552 * therefore remain anonymous (and uncached).
3554 if (!BUF_EMPTY(hdr
)) {
3555 arc_buf_hdr_t
*exists
;
3556 kmutex_t
*hash_lock
;
3558 ASSERT(zio
->io_error
== 0);
3560 arc_cksum_verify(buf
);
3562 exists
= buf_hash_insert(hdr
, &hash_lock
);
3565 * This can only happen if we overwrite for
3566 * sync-to-convergence, because we remove
3567 * buffers from the hash table when we arc_free().
3569 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
3570 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
3571 panic("bad overwrite, hdr=%p exists=%p",
3572 (void *)hdr
, (void *)exists
);
3573 ASSERT(refcount_is_zero(&exists
->b_refcnt
));
3574 arc_change_state(arc_anon
, exists
, hash_lock
);
3575 mutex_exit(hash_lock
);
3576 arc_hdr_destroy(exists
);
3577 exists
= buf_hash_insert(hdr
, &hash_lock
);
3578 ASSERT3P(exists
, ==, NULL
);
3581 ASSERT(hdr
->b_datacnt
== 1);
3582 ASSERT(hdr
->b_state
== arc_anon
);
3583 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
3584 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
3587 hdr
->b_flags
&= ~ARC_IO_IN_PROGRESS
;
3588 /* if it's not anon, we are doing a scrub */
3589 if (!exists
&& hdr
->b_state
== arc_anon
)
3590 arc_access(hdr
, hash_lock
);
3591 mutex_exit(hash_lock
);
3593 hdr
->b_flags
&= ~ARC_IO_IN_PROGRESS
;
3596 ASSERT(!refcount_is_zero(&hdr
->b_refcnt
));
3597 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
3599 kmem_free(callback
, sizeof (arc_write_callback_t
));
3603 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
3604 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
, const zio_prop_t
*zp
,
3605 arc_done_func_t
*ready
, arc_done_func_t
*done
, void *private,
3606 int priority
, int zio_flags
, const zbookmark_t
*zb
)
3608 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3609 arc_write_callback_t
*callback
;
3612 ASSERT(ready
!= NULL
);
3613 ASSERT(done
!= NULL
);
3614 ASSERT(!HDR_IO_ERROR(hdr
));
3615 ASSERT((hdr
->b_flags
& ARC_IO_IN_PROGRESS
) == 0);
3616 ASSERT(hdr
->b_acb
== NULL
);
3618 hdr
->b_flags
|= ARC_L2CACHE
;
3619 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_PUSHPAGE
);
3620 callback
->awcb_ready
= ready
;
3621 callback
->awcb_done
= done
;
3622 callback
->awcb_private
= private;
3623 callback
->awcb_buf
= buf
;
3625 zio
= zio_write(pio
, spa
, txg
, bp
, buf
->b_data
, hdr
->b_size
, zp
,
3626 arc_write_ready
, arc_write_done
, callback
, priority
, zio_flags
, zb
);
3632 arc_memory_throttle(uint64_t reserve
, uint64_t inflight_data
, uint64_t txg
)
3635 uint64_t available_memory
;
3637 if (zfs_arc_memory_throttle_disable
)
3640 /* Easily reclaimable memory (free + inactive + arc-evictable) */
3641 available_memory
= ptob(spl_kmem_availrmem()) + arc_evictable_memory();
3643 if (available_memory
<= zfs_write_limit_max
) {
3644 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
3645 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
3649 if (inflight_data
> available_memory
/ 4) {
3650 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
3651 DMU_TX_STAT_BUMP(dmu_tx_memory_inflight
);
3659 arc_tempreserve_clear(uint64_t reserve
)
3661 atomic_add_64(&arc_tempreserve
, -reserve
);
3662 ASSERT((int64_t)arc_tempreserve
>= 0);
3666 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
3673 * Once in a while, fail for no reason. Everything should cope.
3675 if (spa_get_random(10000) == 0) {
3676 dprintf("forcing random failure\n");
3680 if (reserve
> arc_c
/4 && !arc_no_grow
)
3681 arc_c
= MIN(arc_c_max
, reserve
* 4);
3682 if (reserve
> arc_c
) {
3683 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
3688 * Don't count loaned bufs as in flight dirty data to prevent long
3689 * network delays from blocking transactions that are ready to be
3690 * assigned to a txg.
3692 anon_size
= MAX((int64_t)(arc_anon
->arcs_size
- arc_loaned_bytes
), 0);
3695 * Writes will, almost always, require additional memory allocations
3696 * in order to compress/encrypt/etc the data. We therefor need to
3697 * make sure that there is sufficient available memory for this.
3699 if ((error
= arc_memory_throttle(reserve
, anon_size
, txg
)))
3703 * Throttle writes when the amount of dirty data in the cache
3704 * gets too large. We try to keep the cache less than half full
3705 * of dirty blocks so that our sync times don't grow too large.
3706 * Note: if two requests come in concurrently, we might let them
3707 * both succeed, when one of them should fail. Not a huge deal.
3710 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
3711 anon_size
> arc_c
/ 4) {
3712 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3713 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3714 arc_tempreserve
>>10,
3715 arc_anon
->arcs_lsize
[ARC_BUFC_METADATA
]>>10,
3716 arc_anon
->arcs_lsize
[ARC_BUFC_DATA
]>>10,
3717 reserve
>>10, arc_c
>>10);
3718 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
3721 atomic_add_64(&arc_tempreserve
, reserve
);
3726 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
3727 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
3729 size
->value
.ui64
= state
->arcs_size
;
3730 evict_data
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_DATA
];
3731 evict_metadata
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_METADATA
];
3735 arc_kstat_update(kstat_t
*ksp
, int rw
)
3737 arc_stats_t
*as
= ksp
->ks_data
;
3739 if (rw
== KSTAT_WRITE
) {
3742 arc_kstat_update_state(arc_anon
,
3743 &as
->arcstat_anon_size
,
3744 &as
->arcstat_anon_evict_data
,
3745 &as
->arcstat_anon_evict_metadata
);
3746 arc_kstat_update_state(arc_mru
,
3747 &as
->arcstat_mru_size
,
3748 &as
->arcstat_mru_evict_data
,
3749 &as
->arcstat_mru_evict_metadata
);
3750 arc_kstat_update_state(arc_mru_ghost
,
3751 &as
->arcstat_mru_ghost_size
,
3752 &as
->arcstat_mru_ghost_evict_data
,
3753 &as
->arcstat_mru_ghost_evict_metadata
);
3754 arc_kstat_update_state(arc_mfu
,
3755 &as
->arcstat_mfu_size
,
3756 &as
->arcstat_mfu_evict_data
,
3757 &as
->arcstat_mfu_evict_metadata
);
3758 arc_kstat_update_state(arc_mfu_ghost
,
3759 &as
->arcstat_mfu_ghost_size
,
3760 &as
->arcstat_mfu_ghost_evict_data
,
3761 &as
->arcstat_mfu_ghost_evict_metadata
);
3770 mutex_init(&arc_reclaim_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3771 cv_init(&arc_reclaim_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
3773 /* Convert seconds to clock ticks */
3774 arc_min_prefetch_lifespan
= 1 * hz
;
3776 /* Start out with 1/8 of all memory */
3777 arc_c
= physmem
* PAGESIZE
/ 8;
3781 * On architectures where the physical memory can be larger
3782 * than the addressable space (intel in 32-bit mode), we may
3783 * need to limit the cache to 1/8 of VM size.
3785 arc_c
= MIN(arc_c
, vmem_size(heap_arena
, VMEM_ALLOC
| VMEM_FREE
) / 8);
3787 * Register a shrinker to support synchronous (direct) memory
3788 * reclaim from the arc. This is done to prevent kswapd from
3789 * swapping out pages when it is preferable to shrink the arc.
3791 spl_register_shrinker(&arc_shrinker
);
3794 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3795 arc_c_min
= MAX(arc_c
/ 4, 64<<20);
3796 /* set max to 1/2 of all memory */
3797 arc_c_max
= MAX(arc_c
* 4, arc_c_max
);
3800 * Allow the tunables to override our calculations if they are
3801 * reasonable (ie. over 64MB)
3803 if (zfs_arc_max
> 64<<20 && zfs_arc_max
< physmem
* PAGESIZE
)
3804 arc_c_max
= zfs_arc_max
;
3805 if (zfs_arc_min
> 64<<20 && zfs_arc_min
<= arc_c_max
)
3806 arc_c_min
= zfs_arc_min
;
3809 arc_p
= (arc_c
>> 1);
3811 /* limit meta-data to 1/4 of the arc capacity */
3812 arc_meta_limit
= arc_c_max
/ 4;
3815 /* Allow the tunable to override if it is reasonable */
3816 if (zfs_arc_meta_limit
> 0 && zfs_arc_meta_limit
<= arc_c_max
)
3817 arc_meta_limit
= zfs_arc_meta_limit
;
3819 if (arc_c_min
< arc_meta_limit
/ 2 && zfs_arc_min
== 0)
3820 arc_c_min
= arc_meta_limit
/ 2;
3822 if (zfs_arc_grow_retry
> 0)
3823 arc_grow_retry
= zfs_arc_grow_retry
;
3825 if (zfs_arc_shrink_shift
> 0)
3826 arc_shrink_shift
= zfs_arc_shrink_shift
;
3828 if (zfs_arc_p_min_shift
> 0)
3829 arc_p_min_shift
= zfs_arc_p_min_shift
;
3831 if (zfs_arc_meta_prune
> 0)
3832 arc_meta_prune
= zfs_arc_meta_prune
;
3834 /* if kmem_flags are set, lets try to use less memory */
3835 if (kmem_debugging())
3837 if (arc_c
< arc_c_min
)
3840 arc_anon
= &ARC_anon
;
3842 arc_mru_ghost
= &ARC_mru_ghost
;
3844 arc_mfu_ghost
= &ARC_mfu_ghost
;
3845 arc_l2c_only
= &ARC_l2c_only
;
3848 mutex_init(&arc_anon
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3849 mutex_init(&arc_mru
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3850 mutex_init(&arc_mru_ghost
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3851 mutex_init(&arc_mfu
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3852 mutex_init(&arc_mfu_ghost
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3853 mutex_init(&arc_l2c_only
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3855 list_create(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
],
3856 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3857 list_create(&arc_mru
->arcs_list
[ARC_BUFC_DATA
],
3858 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3859 list_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
],
3860 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3861 list_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
],
3862 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3863 list_create(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
],
3864 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3865 list_create(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
],
3866 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3867 list_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
],
3868 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3869 list_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
],
3870 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3871 list_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
],
3872 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3873 list_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
],
3874 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3878 arc_thread_exit
= 0;
3879 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
3880 offsetof(arc_prune_t
, p_node
));
3881 arc_eviction_list
= NULL
;
3882 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3883 mutex_init(&arc_eviction_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3884 bzero(&arc_eviction_hdr
, sizeof (arc_buf_hdr_t
));
3886 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
3887 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
3889 if (arc_ksp
!= NULL
) {
3890 arc_ksp
->ks_data
= &arc_stats
;
3891 arc_ksp
->ks_update
= arc_kstat_update
;
3892 kstat_install(arc_ksp
);
3895 (void) thread_create(NULL
, 0, arc_adapt_thread
, NULL
, 0, &p0
,
3896 TS_RUN
, minclsyspri
);
3901 if (zfs_write_limit_max
== 0)
3902 zfs_write_limit_max
= ptob(physmem
) >> zfs_write_limit_shift
;
3904 zfs_write_limit_shift
= 0;
3905 mutex_init(&zfs_write_limit_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3913 mutex_enter(&arc_reclaim_thr_lock
);
3915 spl_unregister_shrinker(&arc_shrinker
);
3916 #endif /* _KERNEL */
3918 arc_thread_exit
= 1;
3919 while (arc_thread_exit
!= 0)
3920 cv_wait(&arc_reclaim_thr_cv
, &arc_reclaim_thr_lock
);
3921 mutex_exit(&arc_reclaim_thr_lock
);
3927 if (arc_ksp
!= NULL
) {
3928 kstat_delete(arc_ksp
);
3932 mutex_enter(&arc_prune_mtx
);
3933 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
3934 list_remove(&arc_prune_list
, p
);
3935 refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
3936 refcount_destroy(&p
->p_refcnt
);
3937 kmem_free(p
, sizeof (*p
));
3939 mutex_exit(&arc_prune_mtx
);
3941 list_destroy(&arc_prune_list
);
3942 mutex_destroy(&arc_prune_mtx
);
3943 mutex_destroy(&arc_eviction_mtx
);
3944 mutex_destroy(&arc_reclaim_thr_lock
);
3945 cv_destroy(&arc_reclaim_thr_cv
);
3947 list_destroy(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
3948 list_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
3949 list_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
3950 list_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
3951 list_destroy(&arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
3952 list_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
3953 list_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
3954 list_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
3956 mutex_destroy(&arc_anon
->arcs_mtx
);
3957 mutex_destroy(&arc_mru
->arcs_mtx
);
3958 mutex_destroy(&arc_mru_ghost
->arcs_mtx
);
3959 mutex_destroy(&arc_mfu
->arcs_mtx
);
3960 mutex_destroy(&arc_mfu_ghost
->arcs_mtx
);
3961 mutex_destroy(&arc_l2c_only
->arcs_mtx
);
3963 mutex_destroy(&zfs_write_limit_lock
);
3967 ASSERT(arc_loaned_bytes
== 0);
3973 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3974 * It uses dedicated storage devices to hold cached data, which are populated
3975 * using large infrequent writes. The main role of this cache is to boost
3976 * the performance of random read workloads. The intended L2ARC devices
3977 * include short-stroked disks, solid state disks, and other media with
3978 * substantially faster read latency than disk.
3980 * +-----------------------+
3982 * +-----------------------+
3985 * l2arc_feed_thread() arc_read()
3989 * +---------------+ |
3991 * +---------------+ |
3996 * +-------+ +-------+
3998 * | cache | | cache |
3999 * +-------+ +-------+
4000 * +=========+ .-----.
4001 * : L2ARC : |-_____-|
4002 * : devices : | Disks |
4003 * +=========+ `-_____-'
4005 * Read requests are satisfied from the following sources, in order:
4008 * 2) vdev cache of L2ARC devices
4010 * 4) vdev cache of disks
4013 * Some L2ARC device types exhibit extremely slow write performance.
4014 * To accommodate for this there are some significant differences between
4015 * the L2ARC and traditional cache design:
4017 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4018 * the ARC behave as usual, freeing buffers and placing headers on ghost
4019 * lists. The ARC does not send buffers to the L2ARC during eviction as
4020 * this would add inflated write latencies for all ARC memory pressure.
4022 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4023 * It does this by periodically scanning buffers from the eviction-end of
4024 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4025 * not already there. It scans until a headroom of buffers is satisfied,
4026 * which itself is a buffer for ARC eviction. The thread that does this is
4027 * l2arc_feed_thread(), illustrated below; example sizes are included to
4028 * provide a better sense of ratio than this diagram:
4031 * +---------------------+----------+
4032 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4033 * +---------------------+----------+ | o L2ARC eligible
4034 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4035 * +---------------------+----------+ |
4036 * 15.9 Gbytes ^ 32 Mbytes |
4038 * l2arc_feed_thread()
4040 * l2arc write hand <--[oooo]--'
4044 * +==============================+
4045 * L2ARC dev |####|#|###|###| |####| ... |
4046 * +==============================+
4049 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4050 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4051 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4052 * safe to say that this is an uncommon case, since buffers at the end of
4053 * the ARC lists have moved there due to inactivity.
4055 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4056 * then the L2ARC simply misses copying some buffers. This serves as a
4057 * pressure valve to prevent heavy read workloads from both stalling the ARC
4058 * with waits and clogging the L2ARC with writes. This also helps prevent
4059 * the potential for the L2ARC to churn if it attempts to cache content too
4060 * quickly, such as during backups of the entire pool.
4062 * 5. After system boot and before the ARC has filled main memory, there are
4063 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4064 * lists can remain mostly static. Instead of searching from tail of these
4065 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4066 * for eligible buffers, greatly increasing its chance of finding them.
4068 * The L2ARC device write speed is also boosted during this time so that
4069 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4070 * there are no L2ARC reads, and no fear of degrading read performance
4071 * through increased writes.
4073 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4074 * the vdev queue can aggregate them into larger and fewer writes. Each
4075 * device is written to in a rotor fashion, sweeping writes through
4076 * available space then repeating.
4078 * 7. The L2ARC does not store dirty content. It never needs to flush
4079 * write buffers back to disk based storage.
4081 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4082 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4084 * The performance of the L2ARC can be tweaked by a number of tunables, which
4085 * may be necessary for different workloads:
4087 * l2arc_write_max max write bytes per interval
4088 * l2arc_write_boost extra write bytes during device warmup
4089 * l2arc_noprefetch skip caching prefetched buffers
4090 * l2arc_headroom number of max device writes to precache
4091 * l2arc_feed_secs seconds between L2ARC writing
4093 * Tunables may be removed or added as future performance improvements are
4094 * integrated, and also may become zpool properties.
4096 * There are three key functions that control how the L2ARC warms up:
4098 * l2arc_write_eligible() check if a buffer is eligible to cache
4099 * l2arc_write_size() calculate how much to write
4100 * l2arc_write_interval() calculate sleep delay between writes
4102 * These three functions determine what to write, how much, and how quickly
4107 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*ab
)
4110 * A buffer is *not* eligible for the L2ARC if it:
4111 * 1. belongs to a different spa.
4112 * 2. is already cached on the L2ARC.
4113 * 3. has an I/O in progress (it may be an incomplete read).
4114 * 4. is flagged not eligible (zfs property).
4116 if (ab
->b_spa
!= spa_guid
|| ab
->b_l2hdr
!= NULL
||
4117 HDR_IO_IN_PROGRESS(ab
) || !HDR_L2CACHE(ab
))
4124 l2arc_write_size(l2arc_dev_t
*dev
)
4128 size
= dev
->l2ad_write
;
4130 if (arc_warm
== B_FALSE
)
4131 size
+= dev
->l2ad_boost
;
4138 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
4140 clock_t interval
, next
, now
;
4143 * If the ARC lists are busy, increase our write rate; if the
4144 * lists are stale, idle back. This is achieved by checking
4145 * how much we previously wrote - if it was more than half of
4146 * what we wanted, schedule the next write much sooner.
4148 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
4149 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
4151 interval
= hz
* l2arc_feed_secs
;
4153 now
= ddi_get_lbolt();
4154 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
4160 l2arc_hdr_stat_add(void)
4162 ARCSTAT_INCR(arcstat_l2_hdr_size
, HDR_SIZE
+ L2HDR_SIZE
);
4163 ARCSTAT_INCR(arcstat_hdr_size
, -HDR_SIZE
);
4167 l2arc_hdr_stat_remove(void)
4169 ARCSTAT_INCR(arcstat_l2_hdr_size
, -(HDR_SIZE
+ L2HDR_SIZE
));
4170 ARCSTAT_INCR(arcstat_hdr_size
, HDR_SIZE
);
4174 * Cycle through L2ARC devices. This is how L2ARC load balances.
4175 * If a device is returned, this also returns holding the spa config lock.
4177 static l2arc_dev_t
*
4178 l2arc_dev_get_next(void)
4180 l2arc_dev_t
*first
, *next
= NULL
;
4183 * Lock out the removal of spas (spa_namespace_lock), then removal
4184 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4185 * both locks will be dropped and a spa config lock held instead.
4187 mutex_enter(&spa_namespace_lock
);
4188 mutex_enter(&l2arc_dev_mtx
);
4190 /* if there are no vdevs, there is nothing to do */
4191 if (l2arc_ndev
== 0)
4195 next
= l2arc_dev_last
;
4197 /* loop around the list looking for a non-faulted vdev */
4199 next
= list_head(l2arc_dev_list
);
4201 next
= list_next(l2arc_dev_list
, next
);
4203 next
= list_head(l2arc_dev_list
);
4206 /* if we have come back to the start, bail out */
4209 else if (next
== first
)
4212 } while (vdev_is_dead(next
->l2ad_vdev
));
4214 /* if we were unable to find any usable vdevs, return NULL */
4215 if (vdev_is_dead(next
->l2ad_vdev
))
4218 l2arc_dev_last
= next
;
4221 mutex_exit(&l2arc_dev_mtx
);
4224 * Grab the config lock to prevent the 'next' device from being
4225 * removed while we are writing to it.
4228 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
4229 mutex_exit(&spa_namespace_lock
);
4235 * Free buffers that were tagged for destruction.
4238 l2arc_do_free_on_write(void)
4241 l2arc_data_free_t
*df
, *df_prev
;
4243 mutex_enter(&l2arc_free_on_write_mtx
);
4244 buflist
= l2arc_free_on_write
;
4246 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
4247 df_prev
= list_prev(buflist
, df
);
4248 ASSERT(df
->l2df_data
!= NULL
);
4249 ASSERT(df
->l2df_func
!= NULL
);
4250 df
->l2df_func(df
->l2df_data
, df
->l2df_size
);
4251 list_remove(buflist
, df
);
4252 kmem_free(df
, sizeof (l2arc_data_free_t
));
4255 mutex_exit(&l2arc_free_on_write_mtx
);
4259 * A write to a cache device has completed. Update all headers to allow
4260 * reads from these buffers to begin.
4263 l2arc_write_done(zio_t
*zio
)
4265 l2arc_write_callback_t
*cb
;
4268 arc_buf_hdr_t
*head
, *ab
, *ab_prev
;
4269 l2arc_buf_hdr_t
*abl2
;
4270 kmutex_t
*hash_lock
;
4272 cb
= zio
->io_private
;
4274 dev
= cb
->l2wcb_dev
;
4275 ASSERT(dev
!= NULL
);
4276 head
= cb
->l2wcb_head
;
4277 ASSERT(head
!= NULL
);
4278 buflist
= dev
->l2ad_buflist
;
4279 ASSERT(buflist
!= NULL
);
4280 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
4281 l2arc_write_callback_t
*, cb
);
4283 if (zio
->io_error
!= 0)
4284 ARCSTAT_BUMP(arcstat_l2_writes_error
);
4286 mutex_enter(&l2arc_buflist_mtx
);
4289 * All writes completed, or an error was hit.
4291 for (ab
= list_prev(buflist
, head
); ab
; ab
= ab_prev
) {
4292 ab_prev
= list_prev(buflist
, ab
);
4294 hash_lock
= HDR_LOCK(ab
);
4295 if (!mutex_tryenter(hash_lock
)) {
4297 * This buffer misses out. It may be in a stage
4298 * of eviction. Its ARC_L2_WRITING flag will be
4299 * left set, denying reads to this buffer.
4301 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss
);
4305 if (zio
->io_error
!= 0) {
4307 * Error - drop L2ARC entry.
4309 list_remove(buflist
, ab
);
4312 kmem_free(abl2
, sizeof (l2arc_buf_hdr_t
));
4313 ARCSTAT_INCR(arcstat_l2_size
, -ab
->b_size
);
4317 * Allow ARC to begin reads to this L2ARC entry.
4319 ab
->b_flags
&= ~ARC_L2_WRITING
;
4321 mutex_exit(hash_lock
);
4324 atomic_inc_64(&l2arc_writes_done
);
4325 list_remove(buflist
, head
);
4326 kmem_cache_free(hdr_cache
, head
);
4327 mutex_exit(&l2arc_buflist_mtx
);
4329 l2arc_do_free_on_write();
4331 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
4335 * A read to a cache device completed. Validate buffer contents before
4336 * handing over to the regular ARC routines.
4339 l2arc_read_done(zio_t
*zio
)
4341 l2arc_read_callback_t
*cb
;
4344 kmutex_t
*hash_lock
;
4347 ASSERT(zio
->io_vd
!= NULL
);
4348 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
4350 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
4352 cb
= zio
->io_private
;
4354 buf
= cb
->l2rcb_buf
;
4355 ASSERT(buf
!= NULL
);
4357 hash_lock
= HDR_LOCK(buf
->b_hdr
);
4358 mutex_enter(hash_lock
);
4360 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
4363 * Check this survived the L2ARC journey.
4365 equal
= arc_cksum_equal(buf
);
4366 if (equal
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
4367 mutex_exit(hash_lock
);
4368 zio
->io_private
= buf
;
4369 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
4370 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
4373 mutex_exit(hash_lock
);
4375 * Buffer didn't survive caching. Increment stats and
4376 * reissue to the original storage device.
4378 if (zio
->io_error
!= 0) {
4379 ARCSTAT_BUMP(arcstat_l2_io_error
);
4381 zio
->io_error
= EIO
;
4384 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
4387 * If there's no waiter, issue an async i/o to the primary
4388 * storage now. If there *is* a waiter, the caller must
4389 * issue the i/o in a context where it's OK to block.
4391 if (zio
->io_waiter
== NULL
) {
4392 zio_t
*pio
= zio_unique_parent(zio
);
4394 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4396 zio_nowait(zio_read(pio
, cb
->l2rcb_spa
, &cb
->l2rcb_bp
,
4397 buf
->b_data
, zio
->io_size
, arc_read_done
, buf
,
4398 zio
->io_priority
, cb
->l2rcb_flags
, &cb
->l2rcb_zb
));
4402 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
4406 * This is the list priority from which the L2ARC will search for pages to
4407 * cache. This is used within loops (0..3) to cycle through lists in the
4408 * desired order. This order can have a significant effect on cache
4411 * Currently the metadata lists are hit first, MFU then MRU, followed by
4412 * the data lists. This function returns a locked list, and also returns
4416 l2arc_list_locked(int list_num
, kmutex_t
**lock
)
4418 list_t
*list
= NULL
;
4420 ASSERT(list_num
>= 0 && list_num
<= 3);
4424 list
= &arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
4425 *lock
= &arc_mfu
->arcs_mtx
;
4428 list
= &arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
4429 *lock
= &arc_mru
->arcs_mtx
;
4432 list
= &arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
4433 *lock
= &arc_mfu
->arcs_mtx
;
4436 list
= &arc_mru
->arcs_list
[ARC_BUFC_DATA
];
4437 *lock
= &arc_mru
->arcs_mtx
;
4441 ASSERT(!(MUTEX_HELD(*lock
)));
4447 * Evict buffers from the device write hand to the distance specified in
4448 * bytes. This distance may span populated buffers, it may span nothing.
4449 * This is clearing a region on the L2ARC device ready for writing.
4450 * If the 'all' boolean is set, every buffer is evicted.
4453 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
4456 l2arc_buf_hdr_t
*abl2
;
4457 arc_buf_hdr_t
*ab
, *ab_prev
;
4458 kmutex_t
*hash_lock
;
4461 buflist
= dev
->l2ad_buflist
;
4463 if (buflist
== NULL
)
4466 if (!all
&& dev
->l2ad_first
) {
4468 * This is the first sweep through the device. There is
4474 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
4476 * When nearing the end of the device, evict to the end
4477 * before the device write hand jumps to the start.
4479 taddr
= dev
->l2ad_end
;
4481 taddr
= dev
->l2ad_hand
+ distance
;
4483 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
4484 uint64_t, taddr
, boolean_t
, all
);
4487 mutex_enter(&l2arc_buflist_mtx
);
4488 for (ab
= list_tail(buflist
); ab
; ab
= ab_prev
) {
4489 ab_prev
= list_prev(buflist
, ab
);
4491 hash_lock
= HDR_LOCK(ab
);
4492 if (!mutex_tryenter(hash_lock
)) {
4494 * Missed the hash lock. Retry.
4496 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
4497 mutex_exit(&l2arc_buflist_mtx
);
4498 mutex_enter(hash_lock
);
4499 mutex_exit(hash_lock
);
4503 if (HDR_L2_WRITE_HEAD(ab
)) {
4505 * We hit a write head node. Leave it for
4506 * l2arc_write_done().
4508 list_remove(buflist
, ab
);
4509 mutex_exit(hash_lock
);
4513 if (!all
&& ab
->b_l2hdr
!= NULL
&&
4514 (ab
->b_l2hdr
->b_daddr
> taddr
||
4515 ab
->b_l2hdr
->b_daddr
< dev
->l2ad_hand
)) {
4517 * We've evicted to the target address,
4518 * or the end of the device.
4520 mutex_exit(hash_lock
);
4524 if (HDR_FREE_IN_PROGRESS(ab
)) {
4526 * Already on the path to destruction.
4528 mutex_exit(hash_lock
);
4532 if (ab
->b_state
== arc_l2c_only
) {
4533 ASSERT(!HDR_L2_READING(ab
));
4535 * This doesn't exist in the ARC. Destroy.
4536 * arc_hdr_destroy() will call list_remove()
4537 * and decrement arcstat_l2_size.
4539 arc_change_state(arc_anon
, ab
, hash_lock
);
4540 arc_hdr_destroy(ab
);
4543 * Invalidate issued or about to be issued
4544 * reads, since we may be about to write
4545 * over this location.
4547 if (HDR_L2_READING(ab
)) {
4548 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
4549 ab
->b_flags
|= ARC_L2_EVICTED
;
4553 * Tell ARC this no longer exists in L2ARC.
4555 if (ab
->b_l2hdr
!= NULL
) {
4558 kmem_free(abl2
, sizeof (l2arc_buf_hdr_t
));
4559 ARCSTAT_INCR(arcstat_l2_size
, -ab
->b_size
);
4561 list_remove(buflist
, ab
);
4564 * This may have been leftover after a
4567 ab
->b_flags
&= ~ARC_L2_WRITING
;
4569 mutex_exit(hash_lock
);
4571 mutex_exit(&l2arc_buflist_mtx
);
4573 vdev_space_update(dev
->l2ad_vdev
, -(taddr
- dev
->l2ad_evict
), 0, 0);
4574 dev
->l2ad_evict
= taddr
;
4578 * Find and write ARC buffers to the L2ARC device.
4580 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4581 * for reading until they have completed writing.
4584 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
)
4586 arc_buf_hdr_t
*ab
, *ab_prev
, *head
;
4587 l2arc_buf_hdr_t
*hdrl2
;
4589 uint64_t passed_sz
, write_sz
, buf_sz
, headroom
;
4591 kmutex_t
*hash_lock
, *list_lock
= NULL
;
4592 boolean_t have_lock
, full
;
4593 l2arc_write_callback_t
*cb
;
4595 uint64_t guid
= spa_load_guid(spa
);
4598 ASSERT(dev
->l2ad_vdev
!= NULL
);
4603 head
= kmem_cache_alloc(hdr_cache
, KM_PUSHPAGE
);
4604 head
->b_flags
|= ARC_L2_WRITE_HEAD
;
4607 * Copy buffers for L2ARC writing.
4609 mutex_enter(&l2arc_buflist_mtx
);
4610 for (try = 0; try <= 3; try++) {
4611 list
= l2arc_list_locked(try, &list_lock
);
4615 * L2ARC fast warmup.
4617 * Until the ARC is warm and starts to evict, read from the
4618 * head of the ARC lists rather than the tail.
4620 headroom
= target_sz
* l2arc_headroom
;
4621 if (arc_warm
== B_FALSE
)
4622 ab
= list_head(list
);
4624 ab
= list_tail(list
);
4626 for (; ab
; ab
= ab_prev
) {
4627 if (arc_warm
== B_FALSE
)
4628 ab_prev
= list_next(list
, ab
);
4630 ab_prev
= list_prev(list
, ab
);
4632 hash_lock
= HDR_LOCK(ab
);
4633 have_lock
= MUTEX_HELD(hash_lock
);
4634 if (!have_lock
&& !mutex_tryenter(hash_lock
)) {
4636 * Skip this buffer rather than waiting.
4641 passed_sz
+= ab
->b_size
;
4642 if (passed_sz
> headroom
) {
4646 mutex_exit(hash_lock
);
4650 if (!l2arc_write_eligible(guid
, ab
)) {
4651 mutex_exit(hash_lock
);
4655 if ((write_sz
+ ab
->b_size
) > target_sz
) {
4657 mutex_exit(hash_lock
);
4663 * Insert a dummy header on the buflist so
4664 * l2arc_write_done() can find where the
4665 * write buffers begin without searching.
4667 list_insert_head(dev
->l2ad_buflist
, head
);
4669 cb
= kmem_alloc(sizeof (l2arc_write_callback_t
),
4671 cb
->l2wcb_dev
= dev
;
4672 cb
->l2wcb_head
= head
;
4673 pio
= zio_root(spa
, l2arc_write_done
, cb
,
4678 * Create and add a new L2ARC header.
4680 hdrl2
= kmem_zalloc(sizeof (l2arc_buf_hdr_t
),
4683 hdrl2
->b_daddr
= dev
->l2ad_hand
;
4685 ab
->b_flags
|= ARC_L2_WRITING
;
4686 ab
->b_l2hdr
= hdrl2
;
4687 list_insert_head(dev
->l2ad_buflist
, ab
);
4688 buf_data
= ab
->b_buf
->b_data
;
4689 buf_sz
= ab
->b_size
;
4692 * Compute and store the buffer cksum before
4693 * writing. On debug the cksum is verified first.
4695 arc_cksum_verify(ab
->b_buf
);
4696 arc_cksum_compute(ab
->b_buf
, B_TRUE
);
4698 mutex_exit(hash_lock
);
4700 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
4701 dev
->l2ad_hand
, buf_sz
, buf_data
, ZIO_CHECKSUM_OFF
,
4702 NULL
, NULL
, ZIO_PRIORITY_ASYNC_WRITE
,
4703 ZIO_FLAG_CANFAIL
, B_FALSE
);
4705 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
4707 (void) zio_nowait(wzio
);
4710 * Keep the clock hand suitably device-aligned.
4712 buf_sz
= vdev_psize_to_asize(dev
->l2ad_vdev
, buf_sz
);
4715 dev
->l2ad_hand
+= buf_sz
;
4718 mutex_exit(list_lock
);
4723 mutex_exit(&l2arc_buflist_mtx
);
4727 kmem_cache_free(hdr_cache
, head
);
4731 ASSERT3U(write_sz
, <=, target_sz
);
4732 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
4733 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_sz
);
4734 ARCSTAT_INCR(arcstat_l2_size
, write_sz
);
4735 vdev_space_update(dev
->l2ad_vdev
, write_sz
, 0, 0);
4738 * Bump device hand to the device start if it is approaching the end.
4739 * l2arc_evict() will already have evicted ahead for this case.
4741 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
4742 vdev_space_update(dev
->l2ad_vdev
,
4743 dev
->l2ad_end
- dev
->l2ad_hand
, 0, 0);
4744 dev
->l2ad_hand
= dev
->l2ad_start
;
4745 dev
->l2ad_evict
= dev
->l2ad_start
;
4746 dev
->l2ad_first
= B_FALSE
;
4749 dev
->l2ad_writing
= B_TRUE
;
4750 (void) zio_wait(pio
);
4751 dev
->l2ad_writing
= B_FALSE
;
4757 * This thread feeds the L2ARC at regular intervals. This is the beating
4758 * heart of the L2ARC.
4761 l2arc_feed_thread(void)
4766 uint64_t size
, wrote
;
4767 clock_t begin
, next
= ddi_get_lbolt();
4769 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
4771 mutex_enter(&l2arc_feed_thr_lock
);
4773 while (l2arc_thread_exit
== 0) {
4774 CALLB_CPR_SAFE_BEGIN(&cpr
);
4775 (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv
,
4776 &l2arc_feed_thr_lock
, next
);
4777 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
4778 next
= ddi_get_lbolt() + hz
;
4781 * Quick check for L2ARC devices.
4783 mutex_enter(&l2arc_dev_mtx
);
4784 if (l2arc_ndev
== 0) {
4785 mutex_exit(&l2arc_dev_mtx
);
4788 mutex_exit(&l2arc_dev_mtx
);
4789 begin
= ddi_get_lbolt();
4792 * This selects the next l2arc device to write to, and in
4793 * doing so the next spa to feed from: dev->l2ad_spa. This
4794 * will return NULL if there are now no l2arc devices or if
4795 * they are all faulted.
4797 * If a device is returned, its spa's config lock is also
4798 * held to prevent device removal. l2arc_dev_get_next()
4799 * will grab and release l2arc_dev_mtx.
4801 if ((dev
= l2arc_dev_get_next()) == NULL
)
4804 spa
= dev
->l2ad_spa
;
4805 ASSERT(spa
!= NULL
);
4808 * If the pool is read-only then force the feed thread to
4809 * sleep a little longer.
4811 if (!spa_writeable(spa
)) {
4812 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
4813 spa_config_exit(spa
, SCL_L2ARC
, dev
);
4818 * Avoid contributing to memory pressure.
4821 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
4822 spa_config_exit(spa
, SCL_L2ARC
, dev
);
4826 ARCSTAT_BUMP(arcstat_l2_feeds
);
4828 size
= l2arc_write_size(dev
);
4831 * Evict L2ARC buffers that will be overwritten.
4833 l2arc_evict(dev
, size
, B_FALSE
);
4836 * Write ARC buffers.
4838 wrote
= l2arc_write_buffers(spa
, dev
, size
);
4841 * Calculate interval between writes.
4843 next
= l2arc_write_interval(begin
, size
, wrote
);
4844 spa_config_exit(spa
, SCL_L2ARC
, dev
);
4847 l2arc_thread_exit
= 0;
4848 cv_broadcast(&l2arc_feed_thr_cv
);
4849 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
4854 l2arc_vdev_present(vdev_t
*vd
)
4858 mutex_enter(&l2arc_dev_mtx
);
4859 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
4860 dev
= list_next(l2arc_dev_list
, dev
)) {
4861 if (dev
->l2ad_vdev
== vd
)
4864 mutex_exit(&l2arc_dev_mtx
);
4866 return (dev
!= NULL
);
4870 * Add a vdev for use by the L2ARC. By this point the spa has already
4871 * validated the vdev and opened it.
4874 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
4876 l2arc_dev_t
*adddev
;
4878 ASSERT(!l2arc_vdev_present(vd
));
4881 * Create a new l2arc device entry.
4883 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
4884 adddev
->l2ad_spa
= spa
;
4885 adddev
->l2ad_vdev
= vd
;
4886 adddev
->l2ad_write
= l2arc_write_max
;
4887 adddev
->l2ad_boost
= l2arc_write_boost
;
4888 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
4889 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
4890 adddev
->l2ad_hand
= adddev
->l2ad_start
;
4891 adddev
->l2ad_evict
= adddev
->l2ad_start
;
4892 adddev
->l2ad_first
= B_TRUE
;
4893 adddev
->l2ad_writing
= B_FALSE
;
4894 list_link_init(&adddev
->l2ad_node
);
4895 ASSERT3U(adddev
->l2ad_write
, >, 0);
4898 * This is a list of all ARC buffers that are still valid on the
4901 adddev
->l2ad_buflist
= kmem_zalloc(sizeof (list_t
), KM_SLEEP
);
4902 list_create(adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
4903 offsetof(arc_buf_hdr_t
, b_l2node
));
4905 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
4908 * Add device to global list
4910 mutex_enter(&l2arc_dev_mtx
);
4911 list_insert_head(l2arc_dev_list
, adddev
);
4912 atomic_inc_64(&l2arc_ndev
);
4913 mutex_exit(&l2arc_dev_mtx
);
4917 * Remove a vdev from the L2ARC.
4920 l2arc_remove_vdev(vdev_t
*vd
)
4922 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
4925 * Find the device by vdev
4927 mutex_enter(&l2arc_dev_mtx
);
4928 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
4929 nextdev
= list_next(l2arc_dev_list
, dev
);
4930 if (vd
== dev
->l2ad_vdev
) {
4935 ASSERT(remdev
!= NULL
);
4938 * Remove device from global list
4940 list_remove(l2arc_dev_list
, remdev
);
4941 l2arc_dev_last
= NULL
; /* may have been invalidated */
4942 atomic_dec_64(&l2arc_ndev
);
4943 mutex_exit(&l2arc_dev_mtx
);
4946 * Clear all buflists and ARC references. L2ARC device flush.
4948 l2arc_evict(remdev
, 0, B_TRUE
);
4949 list_destroy(remdev
->l2ad_buflist
);
4950 kmem_free(remdev
->l2ad_buflist
, sizeof (list_t
));
4951 kmem_free(remdev
, sizeof (l2arc_dev_t
));
4957 l2arc_thread_exit
= 0;
4959 l2arc_writes_sent
= 0;
4960 l2arc_writes_done
= 0;
4962 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
4963 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
4964 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4965 mutex_init(&l2arc_buflist_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4966 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4968 l2arc_dev_list
= &L2ARC_dev_list
;
4969 l2arc_free_on_write
= &L2ARC_free_on_write
;
4970 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
4971 offsetof(l2arc_dev_t
, l2ad_node
));
4972 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
4973 offsetof(l2arc_data_free_t
, l2df_list_node
));
4980 * This is called from dmu_fini(), which is called from spa_fini();
4981 * Because of this, we can assume that all l2arc devices have
4982 * already been removed when the pools themselves were removed.
4985 l2arc_do_free_on_write();
4987 mutex_destroy(&l2arc_feed_thr_lock
);
4988 cv_destroy(&l2arc_feed_thr_cv
);
4989 mutex_destroy(&l2arc_dev_mtx
);
4990 mutex_destroy(&l2arc_buflist_mtx
);
4991 mutex_destroy(&l2arc_free_on_write_mtx
);
4993 list_destroy(l2arc_dev_list
);
4994 list_destroy(l2arc_free_on_write
);
5000 if (!(spa_mode_global
& FWRITE
))
5003 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
5004 TS_RUN
, minclsyspri
);
5010 if (!(spa_mode_global
& FWRITE
))
5013 mutex_enter(&l2arc_feed_thr_lock
);
5014 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
5015 l2arc_thread_exit
= 1;
5016 while (l2arc_thread_exit
!= 0)
5017 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
5018 mutex_exit(&l2arc_feed_thr_lock
);
5021 #if defined(_KERNEL) && defined(HAVE_SPL)
5022 EXPORT_SYMBOL(arc_read
);
5023 EXPORT_SYMBOL(arc_buf_remove_ref
);
5024 EXPORT_SYMBOL(arc_getbuf_func
);
5025 EXPORT_SYMBOL(arc_add_prune_callback
);
5026 EXPORT_SYMBOL(arc_remove_prune_callback
);
5028 module_param(zfs_arc_min
, ulong
, 0444);
5029 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
5031 module_param(zfs_arc_max
, ulong
, 0444);
5032 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
5034 module_param(zfs_arc_meta_limit
, ulong
, 0444);
5035 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
5037 module_param(zfs_arc_meta_prune
, int, 0444);
5038 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Bytes of meta data to prune");
5040 module_param(zfs_arc_grow_retry
, int, 0444);
5041 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
5043 module_param(zfs_arc_shrink_shift
, int, 0444);
5044 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
5046 module_param(zfs_arc_p_min_shift
, int, 0444);
5047 MODULE_PARM_DESC(zfs_arc_p_min_shift
, "arc_c shift to calc min/max arc_p");
5049 module_param(zfs_disable_dup_eviction
, int, 0644);
5050 MODULE_PARM_DESC(zfs_disable_dup_eviction
, "disable duplicate buffer eviction");
5052 module_param(zfs_arc_memory_throttle_disable
, int, 0644);
5053 MODULE_PARM_DESC(zfs_arc_memory_throttle_disable
, "disable memory throttle");
5055 module_param(l2arc_write_max
, ulong
, 0444);
5056 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
5058 module_param(l2arc_write_boost
, ulong
, 0444);
5059 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
5061 module_param(l2arc_headroom
, ulong
, 0444);
5062 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
5064 module_param(l2arc_feed_secs
, ulong
, 0444);
5065 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
5067 module_param(l2arc_feed_min_ms
, ulong
, 0444);
5068 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
5070 module_param(l2arc_noprefetch
, int, 0444);
5071 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
5073 module_param(l2arc_feed_again
, int, 0444);
5074 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
5076 module_param(l2arc_norw
, int, 0444);
5077 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");