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) 2013 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
29 * DVA-based Adjustable Replacement Cache
31 * While much of the theory of operation used here is
32 * based on the self-tuning, low overhead replacement cache
33 * presented by Megiddo and Modha at FAST 2003, there are some
34 * significant differences:
36 * 1. The Megiddo and Modha model assumes any page is evictable.
37 * Pages in its cache cannot be "locked" into memory. This makes
38 * the eviction algorithm simple: evict the last page in the list.
39 * This also make the performance characteristics easy to reason
40 * about. Our cache is not so simple. At any given moment, some
41 * subset of the blocks in the cache are un-evictable because we
42 * have handed out a reference to them. Blocks are only evictable
43 * when there are no external references active. This makes
44 * eviction far more problematic: we choose to evict the evictable
45 * blocks that are the "lowest" in the list.
47 * There are times when it is not possible to evict the requested
48 * space. In these circumstances we are unable to adjust the cache
49 * size. To prevent the cache growing unbounded at these times we
50 * implement a "cache throttle" that slows the flow of new data
51 * into the cache until we can make space available.
53 * 2. The Megiddo and Modha model assumes a fixed cache size.
54 * Pages are evicted when the cache is full and there is a cache
55 * miss. Our model has a variable sized cache. It grows with
56 * high use, but also tries to react to memory pressure from the
57 * operating system: decreasing its size when system memory is
60 * 3. The Megiddo and Modha model assumes a fixed page size. All
61 * elements of the cache are therefore exactly the same size. So
62 * when adjusting the cache size following a cache miss, its simply
63 * a matter of choosing a single page to evict. In our model, we
64 * have variable sized cache blocks (rangeing from 512 bytes to
65 * 128K bytes). We therefore choose a set of blocks to evict to make
66 * space for a cache miss that approximates as closely as possible
67 * the space used by the new block.
69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
70 * by N. Megiddo & D. Modha, FAST 2003
76 * A new reference to a cache buffer can be obtained in two
77 * ways: 1) via a hash table lookup using the DVA as a key,
78 * or 2) via one of the ARC lists. The arc_read() interface
79 * uses method 1, while the internal arc algorithms for
80 * adjusting the cache use method 2. We therefore provide two
81 * types of locks: 1) the hash table lock array, and 2) the
84 * Buffers do not have their own mutexes, rather they rely on the
85 * hash table mutexes for the bulk of their protection (i.e. most
86 * fields in the arc_buf_hdr_t are protected by these mutexes).
88 * buf_hash_find() returns the appropriate mutex (held) when it
89 * locates the requested buffer in the hash table. It returns
90 * NULL for the mutex if the buffer was not in the table.
92 * buf_hash_remove() expects the appropriate hash mutex to be
93 * already held before it is invoked.
95 * Each arc state also has a mutex which is used to protect the
96 * buffer list associated with the state. When attempting to
97 * obtain a hash table lock while holding an arc list lock you
98 * must use: mutex_tryenter() to avoid deadlock. Also note that
99 * the active state mutex must be held before the ghost state mutex.
101 * Arc buffers may have an associated eviction callback function.
102 * This function will be invoked prior to removing the buffer (e.g.
103 * in arc_do_user_evicts()). Note however that the data associated
104 * with the buffer may be evicted prior to the callback. The callback
105 * must be made with *no locks held* (to prevent deadlock). Additionally,
106 * the users of callbacks must ensure that their private data is
107 * protected from simultaneous callbacks from arc_buf_evict()
108 * and arc_do_user_evicts().
110 * It as also possible to register a callback which is run when the
111 * arc_meta_limit is reached and no buffers can be safely evicted. In
112 * this case the arc user should drop a reference on some arc buffers so
113 * they can be reclaimed and the arc_meta_limit honored. For example,
114 * when using the ZPL each dentry holds a references on a znode. These
115 * dentries must be pruned before the arc buffer holding the znode can
118 * Note that the majority of the performance stats are manipulated
119 * with atomic operations.
121 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
123 * - L2ARC buflist creation
124 * - L2ARC buflist eviction
125 * - L2ARC write completion, which walks L2ARC buflists
126 * - ARC header destruction, as it removes from L2ARC buflists
127 * - ARC header release, as it removes from L2ARC buflists
132 #include <sys/zio_compress.h>
133 #include <sys/zfs_context.h>
135 #include <sys/vdev.h>
136 #include <sys/vdev_impl.h>
137 #include <sys/dsl_pool.h>
139 #include <sys/vmsystm.h>
141 #include <sys/fs/swapnode.h>
144 #include <sys/callb.h>
145 #include <sys/kstat.h>
146 #include <sys/dmu_tx.h>
147 #include <zfs_fletcher.h>
150 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
151 boolean_t arc_watch
= B_FALSE
;
154 static kmutex_t arc_reclaim_thr_lock
;
155 static kcondvar_t arc_reclaim_thr_cv
; /* used to signal reclaim thr */
156 static uint8_t arc_thread_exit
;
158 /* number of bytes to prune from caches when at arc_meta_limit is reached */
159 int zfs_arc_meta_prune
= 1048576;
161 typedef enum arc_reclaim_strategy
{
162 ARC_RECLAIM_AGGR
, /* Aggressive reclaim strategy */
163 ARC_RECLAIM_CONS
/* Conservative reclaim strategy */
164 } arc_reclaim_strategy_t
;
167 * The number of iterations through arc_evict_*() before we
168 * drop & reacquire the lock.
170 int arc_evict_iterations
= 100;
172 /* number of seconds before growing cache again */
173 int zfs_arc_grow_retry
= 5;
175 /* shift of arc_c for calculating both min and max arc_p */
176 int zfs_arc_p_min_shift
= 4;
178 /* log2(fraction of arc to reclaim) */
179 int zfs_arc_shrink_shift
= 5;
182 * minimum lifespan of a prefetch block in clock ticks
183 * (initialized in arc_init())
185 int zfs_arc_min_prefetch_lifespan
= HZ
;
187 /* disable arc proactive arc throttle due to low memory */
188 int zfs_arc_memory_throttle_disable
= 1;
190 /* disable duplicate buffer eviction */
191 int zfs_disable_dup_eviction
= 0;
194 * If this percent of memory is free, don't throttle.
196 int arc_lotsfree_percent
= 10;
200 /* expiration time for arc_no_grow */
201 static clock_t arc_grow_time
= 0;
204 * The arc has filled available memory and has now warmed up.
206 static boolean_t arc_warm
;
209 * These tunables are for performance analysis.
211 unsigned long zfs_arc_max
= 0;
212 unsigned long zfs_arc_min
= 0;
213 unsigned long zfs_arc_meta_limit
= 0;
216 * Note that buffers can be in one of 6 states:
217 * ARC_anon - anonymous (discussed below)
218 * ARC_mru - recently used, currently cached
219 * ARC_mru_ghost - recentely used, no longer in cache
220 * ARC_mfu - frequently used, currently cached
221 * ARC_mfu_ghost - frequently used, no longer in cache
222 * ARC_l2c_only - exists in L2ARC but not other states
223 * When there are no active references to the buffer, they are
224 * are linked onto a list in one of these arc states. These are
225 * the only buffers that can be evicted or deleted. Within each
226 * state there are multiple lists, one for meta-data and one for
227 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
228 * etc.) is tracked separately so that it can be managed more
229 * explicitly: favored over data, limited explicitly.
231 * Anonymous buffers are buffers that are not associated with
232 * a DVA. These are buffers that hold dirty block copies
233 * before they are written to stable storage. By definition,
234 * they are "ref'd" and are considered part of arc_mru
235 * that cannot be freed. Generally, they will aquire a DVA
236 * as they are written and migrate onto the arc_mru list.
238 * The ARC_l2c_only state is for buffers that are in the second
239 * level ARC but no longer in any of the ARC_m* lists. The second
240 * level ARC itself may also contain buffers that are in any of
241 * the ARC_m* states - meaning that a buffer can exist in two
242 * places. The reason for the ARC_l2c_only state is to keep the
243 * buffer header in the hash table, so that reads that hit the
244 * second level ARC benefit from these fast lookups.
247 typedef struct arc_state
{
248 list_t arcs_list
[ARC_BUFC_NUMTYPES
]; /* list of evictable buffers */
249 uint64_t arcs_lsize
[ARC_BUFC_NUMTYPES
]; /* amount of evictable data */
250 uint64_t arcs_size
; /* total amount of data in this state */
252 arc_state_type_t arcs_state
;
256 static arc_state_t ARC_anon
;
257 static arc_state_t ARC_mru
;
258 static arc_state_t ARC_mru_ghost
;
259 static arc_state_t ARC_mfu
;
260 static arc_state_t ARC_mfu_ghost
;
261 static arc_state_t ARC_l2c_only
;
263 typedef struct arc_stats
{
264 kstat_named_t arcstat_hits
;
265 kstat_named_t arcstat_misses
;
266 kstat_named_t arcstat_demand_data_hits
;
267 kstat_named_t arcstat_demand_data_misses
;
268 kstat_named_t arcstat_demand_metadata_hits
;
269 kstat_named_t arcstat_demand_metadata_misses
;
270 kstat_named_t arcstat_prefetch_data_hits
;
271 kstat_named_t arcstat_prefetch_data_misses
;
272 kstat_named_t arcstat_prefetch_metadata_hits
;
273 kstat_named_t arcstat_prefetch_metadata_misses
;
274 kstat_named_t arcstat_mru_hits
;
275 kstat_named_t arcstat_mru_ghost_hits
;
276 kstat_named_t arcstat_mfu_hits
;
277 kstat_named_t arcstat_mfu_ghost_hits
;
278 kstat_named_t arcstat_deleted
;
279 kstat_named_t arcstat_recycle_miss
;
281 * Number of buffers that could not be evicted because the hash lock
282 * was held by another thread. The lock may not necessarily be held
283 * by something using the same buffer, since hash locks are shared
284 * by multiple buffers.
286 kstat_named_t arcstat_mutex_miss
;
288 * Number of buffers skipped because they have I/O in progress, are
289 * indrect prefetch buffers that have not lived long enough, or are
290 * not from the spa we're trying to evict from.
292 kstat_named_t arcstat_evict_skip
;
293 kstat_named_t arcstat_evict_l2_cached
;
294 kstat_named_t arcstat_evict_l2_eligible
;
295 kstat_named_t arcstat_evict_l2_ineligible
;
296 kstat_named_t arcstat_hash_elements
;
297 kstat_named_t arcstat_hash_elements_max
;
298 kstat_named_t arcstat_hash_collisions
;
299 kstat_named_t arcstat_hash_chains
;
300 kstat_named_t arcstat_hash_chain_max
;
301 kstat_named_t arcstat_p
;
302 kstat_named_t arcstat_c
;
303 kstat_named_t arcstat_c_min
;
304 kstat_named_t arcstat_c_max
;
305 kstat_named_t arcstat_size
;
306 kstat_named_t arcstat_hdr_size
;
307 kstat_named_t arcstat_data_size
;
308 kstat_named_t arcstat_other_size
;
309 kstat_named_t arcstat_anon_size
;
310 kstat_named_t arcstat_anon_evict_data
;
311 kstat_named_t arcstat_anon_evict_metadata
;
312 kstat_named_t arcstat_mru_size
;
313 kstat_named_t arcstat_mru_evict_data
;
314 kstat_named_t arcstat_mru_evict_metadata
;
315 kstat_named_t arcstat_mru_ghost_size
;
316 kstat_named_t arcstat_mru_ghost_evict_data
;
317 kstat_named_t arcstat_mru_ghost_evict_metadata
;
318 kstat_named_t arcstat_mfu_size
;
319 kstat_named_t arcstat_mfu_evict_data
;
320 kstat_named_t arcstat_mfu_evict_metadata
;
321 kstat_named_t arcstat_mfu_ghost_size
;
322 kstat_named_t arcstat_mfu_ghost_evict_data
;
323 kstat_named_t arcstat_mfu_ghost_evict_metadata
;
324 kstat_named_t arcstat_l2_hits
;
325 kstat_named_t arcstat_l2_misses
;
326 kstat_named_t arcstat_l2_feeds
;
327 kstat_named_t arcstat_l2_rw_clash
;
328 kstat_named_t arcstat_l2_read_bytes
;
329 kstat_named_t arcstat_l2_write_bytes
;
330 kstat_named_t arcstat_l2_writes_sent
;
331 kstat_named_t arcstat_l2_writes_done
;
332 kstat_named_t arcstat_l2_writes_error
;
333 kstat_named_t arcstat_l2_writes_hdr_miss
;
334 kstat_named_t arcstat_l2_evict_lock_retry
;
335 kstat_named_t arcstat_l2_evict_reading
;
336 kstat_named_t arcstat_l2_free_on_write
;
337 kstat_named_t arcstat_l2_abort_lowmem
;
338 kstat_named_t arcstat_l2_cksum_bad
;
339 kstat_named_t arcstat_l2_io_error
;
340 kstat_named_t arcstat_l2_size
;
341 kstat_named_t arcstat_l2_asize
;
342 kstat_named_t arcstat_l2_hdr_size
;
343 kstat_named_t arcstat_l2_compress_successes
;
344 kstat_named_t arcstat_l2_compress_zeros
;
345 kstat_named_t arcstat_l2_compress_failures
;
346 kstat_named_t arcstat_memory_throttle_count
;
347 kstat_named_t arcstat_duplicate_buffers
;
348 kstat_named_t arcstat_duplicate_buffers_size
;
349 kstat_named_t arcstat_duplicate_reads
;
350 kstat_named_t arcstat_memory_direct_count
;
351 kstat_named_t arcstat_memory_indirect_count
;
352 kstat_named_t arcstat_no_grow
;
353 kstat_named_t arcstat_tempreserve
;
354 kstat_named_t arcstat_loaned_bytes
;
355 kstat_named_t arcstat_prune
;
356 kstat_named_t arcstat_meta_used
;
357 kstat_named_t arcstat_meta_limit
;
358 kstat_named_t arcstat_meta_max
;
361 static arc_stats_t arc_stats
= {
362 { "hits", KSTAT_DATA_UINT64
},
363 { "misses", KSTAT_DATA_UINT64
},
364 { "demand_data_hits", KSTAT_DATA_UINT64
},
365 { "demand_data_misses", KSTAT_DATA_UINT64
},
366 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
367 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
368 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
369 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
370 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
371 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
372 { "mru_hits", KSTAT_DATA_UINT64
},
373 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
374 { "mfu_hits", KSTAT_DATA_UINT64
},
375 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
376 { "deleted", KSTAT_DATA_UINT64
},
377 { "recycle_miss", KSTAT_DATA_UINT64
},
378 { "mutex_miss", KSTAT_DATA_UINT64
},
379 { "evict_skip", KSTAT_DATA_UINT64
},
380 { "evict_l2_cached", KSTAT_DATA_UINT64
},
381 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
382 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
383 { "hash_elements", KSTAT_DATA_UINT64
},
384 { "hash_elements_max", KSTAT_DATA_UINT64
},
385 { "hash_collisions", KSTAT_DATA_UINT64
},
386 { "hash_chains", KSTAT_DATA_UINT64
},
387 { "hash_chain_max", KSTAT_DATA_UINT64
},
388 { "p", KSTAT_DATA_UINT64
},
389 { "c", KSTAT_DATA_UINT64
},
390 { "c_min", KSTAT_DATA_UINT64
},
391 { "c_max", KSTAT_DATA_UINT64
},
392 { "size", KSTAT_DATA_UINT64
},
393 { "hdr_size", KSTAT_DATA_UINT64
},
394 { "data_size", KSTAT_DATA_UINT64
},
395 { "other_size", KSTAT_DATA_UINT64
},
396 { "anon_size", KSTAT_DATA_UINT64
},
397 { "anon_evict_data", KSTAT_DATA_UINT64
},
398 { "anon_evict_metadata", KSTAT_DATA_UINT64
},
399 { "mru_size", KSTAT_DATA_UINT64
},
400 { "mru_evict_data", KSTAT_DATA_UINT64
},
401 { "mru_evict_metadata", KSTAT_DATA_UINT64
},
402 { "mru_ghost_size", KSTAT_DATA_UINT64
},
403 { "mru_ghost_evict_data", KSTAT_DATA_UINT64
},
404 { "mru_ghost_evict_metadata", KSTAT_DATA_UINT64
},
405 { "mfu_size", KSTAT_DATA_UINT64
},
406 { "mfu_evict_data", KSTAT_DATA_UINT64
},
407 { "mfu_evict_metadata", KSTAT_DATA_UINT64
},
408 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
409 { "mfu_ghost_evict_data", KSTAT_DATA_UINT64
},
410 { "mfu_ghost_evict_metadata", KSTAT_DATA_UINT64
},
411 { "l2_hits", KSTAT_DATA_UINT64
},
412 { "l2_misses", KSTAT_DATA_UINT64
},
413 { "l2_feeds", KSTAT_DATA_UINT64
},
414 { "l2_rw_clash", KSTAT_DATA_UINT64
},
415 { "l2_read_bytes", KSTAT_DATA_UINT64
},
416 { "l2_write_bytes", KSTAT_DATA_UINT64
},
417 { "l2_writes_sent", KSTAT_DATA_UINT64
},
418 { "l2_writes_done", KSTAT_DATA_UINT64
},
419 { "l2_writes_error", KSTAT_DATA_UINT64
},
420 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64
},
421 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
422 { "l2_evict_reading", KSTAT_DATA_UINT64
},
423 { "l2_free_on_write", KSTAT_DATA_UINT64
},
424 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
425 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
426 { "l2_io_error", KSTAT_DATA_UINT64
},
427 { "l2_size", KSTAT_DATA_UINT64
},
428 { "l2_asize", KSTAT_DATA_UINT64
},
429 { "l2_hdr_size", KSTAT_DATA_UINT64
},
430 { "l2_compress_successes", KSTAT_DATA_UINT64
},
431 { "l2_compress_zeros", KSTAT_DATA_UINT64
},
432 { "l2_compress_failures", KSTAT_DATA_UINT64
},
433 { "memory_throttle_count", KSTAT_DATA_UINT64
},
434 { "duplicate_buffers", KSTAT_DATA_UINT64
},
435 { "duplicate_buffers_size", KSTAT_DATA_UINT64
},
436 { "duplicate_reads", KSTAT_DATA_UINT64
},
437 { "memory_direct_count", KSTAT_DATA_UINT64
},
438 { "memory_indirect_count", KSTAT_DATA_UINT64
},
439 { "arc_no_grow", KSTAT_DATA_UINT64
},
440 { "arc_tempreserve", KSTAT_DATA_UINT64
},
441 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
442 { "arc_prune", KSTAT_DATA_UINT64
},
443 { "arc_meta_used", KSTAT_DATA_UINT64
},
444 { "arc_meta_limit", KSTAT_DATA_UINT64
},
445 { "arc_meta_max", KSTAT_DATA_UINT64
},
448 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
450 #define ARCSTAT_INCR(stat, val) \
451 atomic_add_64(&arc_stats.stat.value.ui64, (val))
453 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
454 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
456 #define ARCSTAT_MAX(stat, val) { \
458 while ((val) > (m = arc_stats.stat.value.ui64) && \
459 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
463 #define ARCSTAT_MAXSTAT(stat) \
464 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
467 * We define a macro to allow ARC hits/misses to be easily broken down by
468 * two separate conditions, giving a total of four different subtypes for
469 * each of hits and misses (so eight statistics total).
471 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
474 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
476 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
480 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
482 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
487 static arc_state_t
*arc_anon
;
488 static arc_state_t
*arc_mru
;
489 static arc_state_t
*arc_mru_ghost
;
490 static arc_state_t
*arc_mfu
;
491 static arc_state_t
*arc_mfu_ghost
;
492 static arc_state_t
*arc_l2c_only
;
495 * There are several ARC variables that are critical to export as kstats --
496 * but we don't want to have to grovel around in the kstat whenever we wish to
497 * manipulate them. For these variables, we therefore define them to be in
498 * terms of the statistic variable. This assures that we are not introducing
499 * the possibility of inconsistency by having shadow copies of the variables,
500 * while still allowing the code to be readable.
502 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
503 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
504 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
505 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
506 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
507 #define arc_no_grow ARCSTAT(arcstat_no_grow)
508 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
509 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
510 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
511 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
512 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
514 #define L2ARC_IS_VALID_COMPRESS(_c_) \
515 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
517 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t
;
519 typedef struct arc_callback arc_callback_t
;
521 struct arc_callback
{
523 arc_done_func_t
*acb_done
;
525 zio_t
*acb_zio_dummy
;
526 arc_callback_t
*acb_next
;
529 typedef struct arc_write_callback arc_write_callback_t
;
531 struct arc_write_callback
{
533 arc_done_func_t
*awcb_ready
;
534 arc_done_func_t
*awcb_physdone
;
535 arc_done_func_t
*awcb_done
;
540 /* protected by hash lock */
545 kmutex_t b_freeze_lock
;
546 zio_cksum_t
*b_freeze_cksum
;
548 arc_buf_hdr_t
*b_hash_next
;
553 arc_callback_t
*b_acb
;
557 arc_buf_contents_t b_type
;
561 /* protected by arc state mutex */
562 arc_state_t
*b_state
;
563 list_node_t b_arc_node
;
565 /* updated atomically */
566 clock_t b_arc_access
;
568 uint32_t b_mru_ghost_hits
;
570 uint32_t b_mfu_ghost_hits
;
573 /* self protecting */
576 l2arc_buf_hdr_t
*b_l2hdr
;
577 list_node_t b_l2node
;
580 static list_t arc_prune_list
;
581 static kmutex_t arc_prune_mtx
;
582 static arc_buf_t
*arc_eviction_list
;
583 static kmutex_t arc_eviction_mtx
;
584 static arc_buf_hdr_t arc_eviction_hdr
;
585 static void arc_get_data_buf(arc_buf_t
*buf
);
586 static void arc_access(arc_buf_hdr_t
*buf
, kmutex_t
*hash_lock
);
587 static int arc_evict_needed(arc_buf_contents_t type
);
588 static void arc_evict_ghost(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
589 arc_buf_contents_t type
);
590 static void arc_buf_watch(arc_buf_t
*buf
);
592 static boolean_t
l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*ab
);
594 #define GHOST_STATE(state) \
595 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
596 (state) == arc_l2c_only)
599 * Private ARC flags. These flags are private ARC only flags that will show up
600 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
601 * be passed in as arc_flags in things like arc_read. However, these flags
602 * should never be passed and should only be set by ARC code. When adding new
603 * public flags, make sure not to smash the private ones.
606 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
607 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
608 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
609 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
610 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
611 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
612 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
613 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
614 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
615 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
617 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
618 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
619 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
620 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
621 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
622 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
623 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
624 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
625 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
626 (hdr)->b_l2hdr != NULL)
627 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
628 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
629 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
635 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
636 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
639 * Hash table routines
642 #define HT_LOCK_ALIGN 64
643 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
648 unsigned char pad
[HT_LOCK_PAD
];
652 #define BUF_LOCKS 256
653 typedef struct buf_hash_table
{
655 arc_buf_hdr_t
**ht_table
;
656 struct ht_lock ht_locks
[BUF_LOCKS
];
659 static buf_hash_table_t buf_hash_table
;
661 #define BUF_HASH_INDEX(spa, dva, birth) \
662 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
663 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
664 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
665 #define HDR_LOCK(hdr) \
666 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
668 uint64_t zfs_crc64_table
[256];
674 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
675 #define L2ARC_HEADROOM 2 /* num of writes */
677 * If we discover during ARC scan any buffers to be compressed, we boost
678 * our headroom for the next scanning cycle by this percentage multiple.
680 #define L2ARC_HEADROOM_BOOST 200
681 #define L2ARC_FEED_SECS 1 /* caching interval secs */
682 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
684 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
685 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
687 /* L2ARC Performance Tunables */
688 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
689 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
690 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
691 unsigned long l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
692 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
693 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
694 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
695 int l2arc_nocompress
= B_FALSE
; /* don't compress bufs */
696 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
697 int l2arc_norw
= B_FALSE
; /* no reads during writes */
702 typedef struct l2arc_dev
{
703 vdev_t
*l2ad_vdev
; /* vdev */
704 spa_t
*l2ad_spa
; /* spa */
705 uint64_t l2ad_hand
; /* next write location */
706 uint64_t l2ad_start
; /* first addr on device */
707 uint64_t l2ad_end
; /* last addr on device */
708 uint64_t l2ad_evict
; /* last addr eviction reached */
709 boolean_t l2ad_first
; /* first sweep through */
710 boolean_t l2ad_writing
; /* currently writing */
711 list_t
*l2ad_buflist
; /* buffer list */
712 list_node_t l2ad_node
; /* device list node */
715 static list_t L2ARC_dev_list
; /* device list */
716 static list_t
*l2arc_dev_list
; /* device list pointer */
717 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
718 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
719 static kmutex_t l2arc_buflist_mtx
; /* mutex for all buflists */
720 static list_t L2ARC_free_on_write
; /* free after write buf list */
721 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
722 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
723 static uint64_t l2arc_ndev
; /* number of devices */
725 typedef struct l2arc_read_callback
{
726 arc_buf_t
*l2rcb_buf
; /* read buffer */
727 spa_t
*l2rcb_spa
; /* spa */
728 blkptr_t l2rcb_bp
; /* original blkptr */
729 zbookmark_t l2rcb_zb
; /* original bookmark */
730 int l2rcb_flags
; /* original flags */
731 enum zio_compress l2rcb_compress
; /* applied compress */
732 } l2arc_read_callback_t
;
734 typedef struct l2arc_write_callback
{
735 l2arc_dev_t
*l2wcb_dev
; /* device info */
736 arc_buf_hdr_t
*l2wcb_head
; /* head of write buflist */
737 } l2arc_write_callback_t
;
739 struct l2arc_buf_hdr
{
740 /* protected by arc_buf_hdr mutex */
741 l2arc_dev_t
*b_dev
; /* L2ARC device */
742 uint64_t b_daddr
; /* disk address, offset byte */
743 /* compression applied to buffer data */
744 enum zio_compress b_compress
;
745 /* real alloc'd buffer size depending on b_compress applied */
748 /* temporary buffer holder for in-flight compressed data */
752 typedef struct l2arc_data_free
{
753 /* protected by l2arc_free_on_write_mtx */
756 void (*l2df_func
)(void *, size_t);
757 list_node_t l2df_list_node
;
760 static kmutex_t l2arc_feed_thr_lock
;
761 static kcondvar_t l2arc_feed_thr_cv
;
762 static uint8_t l2arc_thread_exit
;
764 static void l2arc_read_done(zio_t
*zio
);
765 static void l2arc_hdr_stat_add(void);
766 static void l2arc_hdr_stat_remove(void);
768 static boolean_t
l2arc_compress_buf(l2arc_buf_hdr_t
*l2hdr
);
769 static void l2arc_decompress_zio(zio_t
*zio
, arc_buf_hdr_t
*hdr
,
770 enum zio_compress c
);
771 static void l2arc_release_cdata_buf(arc_buf_hdr_t
*ab
);
774 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
776 uint8_t *vdva
= (uint8_t *)dva
;
777 uint64_t crc
= -1ULL;
780 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
782 for (i
= 0; i
< sizeof (dva_t
); i
++)
783 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
785 crc
^= (spa
>>8) ^ birth
;
790 #define BUF_EMPTY(buf) \
791 ((buf)->b_dva.dva_word[0] == 0 && \
792 (buf)->b_dva.dva_word[1] == 0 && \
795 #define BUF_EQUAL(spa, dva, birth, buf) \
796 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
797 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
798 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
801 buf_discard_identity(arc_buf_hdr_t
*hdr
)
803 hdr
->b_dva
.dva_word
[0] = 0;
804 hdr
->b_dva
.dva_word
[1] = 0;
809 static arc_buf_hdr_t
*
810 buf_hash_find(uint64_t spa
, const dva_t
*dva
, uint64_t birth
, kmutex_t
**lockp
)
812 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
813 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
816 mutex_enter(hash_lock
);
817 for (buf
= buf_hash_table
.ht_table
[idx
]; buf
!= NULL
;
818 buf
= buf
->b_hash_next
) {
819 if (BUF_EQUAL(spa
, dva
, birth
, buf
)) {
824 mutex_exit(hash_lock
);
830 * Insert an entry into the hash table. If there is already an element
831 * equal to elem in the hash table, then the already existing element
832 * will be returned and the new element will not be inserted.
833 * Otherwise returns NULL.
835 static arc_buf_hdr_t
*
836 buf_hash_insert(arc_buf_hdr_t
*buf
, kmutex_t
**lockp
)
838 uint64_t idx
= BUF_HASH_INDEX(buf
->b_spa
, &buf
->b_dva
, buf
->b_birth
);
839 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
843 ASSERT(!HDR_IN_HASH_TABLE(buf
));
845 mutex_enter(hash_lock
);
846 for (fbuf
= buf_hash_table
.ht_table
[idx
], i
= 0; fbuf
!= NULL
;
847 fbuf
= fbuf
->b_hash_next
, i
++) {
848 if (BUF_EQUAL(buf
->b_spa
, &buf
->b_dva
, buf
->b_birth
, fbuf
))
852 buf
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
853 buf_hash_table
.ht_table
[idx
] = buf
;
854 buf
->b_flags
|= ARC_IN_HASH_TABLE
;
856 /* collect some hash table performance data */
858 ARCSTAT_BUMP(arcstat_hash_collisions
);
860 ARCSTAT_BUMP(arcstat_hash_chains
);
862 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
865 ARCSTAT_BUMP(arcstat_hash_elements
);
866 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
872 buf_hash_remove(arc_buf_hdr_t
*buf
)
874 arc_buf_hdr_t
*fbuf
, **bufp
;
875 uint64_t idx
= BUF_HASH_INDEX(buf
->b_spa
, &buf
->b_dva
, buf
->b_birth
);
877 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
878 ASSERT(HDR_IN_HASH_TABLE(buf
));
880 bufp
= &buf_hash_table
.ht_table
[idx
];
881 while ((fbuf
= *bufp
) != buf
) {
882 ASSERT(fbuf
!= NULL
);
883 bufp
= &fbuf
->b_hash_next
;
885 *bufp
= buf
->b_hash_next
;
886 buf
->b_hash_next
= NULL
;
887 buf
->b_flags
&= ~ARC_IN_HASH_TABLE
;
889 /* collect some hash table performance data */
890 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
892 if (buf_hash_table
.ht_table
[idx
] &&
893 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
894 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
898 * Global data structures and functions for the buf kmem cache.
900 static kmem_cache_t
*hdr_cache
;
901 static kmem_cache_t
*buf_cache
;
902 static kmem_cache_t
*l2arc_hdr_cache
;
909 #if defined(_KERNEL) && defined(HAVE_SPL)
911 * Large allocations which do not require contiguous pages
912 * should be using vmem_free() in the linux kernel\
914 vmem_free(buf_hash_table
.ht_table
,
915 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
917 kmem_free(buf_hash_table
.ht_table
,
918 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
920 for (i
= 0; i
< BUF_LOCKS
; i
++)
921 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
922 kmem_cache_destroy(hdr_cache
);
923 kmem_cache_destroy(buf_cache
);
924 kmem_cache_destroy(l2arc_hdr_cache
);
928 * Constructor callback - called when the cache is empty
929 * and a new buf is requested.
933 hdr_cons(void *vbuf
, void *unused
, int kmflag
)
935 arc_buf_hdr_t
*buf
= vbuf
;
937 bzero(buf
, sizeof (arc_buf_hdr_t
));
938 refcount_create(&buf
->b_refcnt
);
939 cv_init(&buf
->b_cv
, NULL
, CV_DEFAULT
, NULL
);
940 mutex_init(&buf
->b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
941 list_link_init(&buf
->b_arc_node
);
942 list_link_init(&buf
->b_l2node
);
943 arc_space_consume(sizeof (arc_buf_hdr_t
), ARC_SPACE_HDRS
);
950 buf_cons(void *vbuf
, void *unused
, int kmflag
)
952 arc_buf_t
*buf
= vbuf
;
954 bzero(buf
, sizeof (arc_buf_t
));
955 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
956 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
962 * Destructor callback - called when a cached buf is
963 * no longer required.
967 hdr_dest(void *vbuf
, void *unused
)
969 arc_buf_hdr_t
*buf
= vbuf
;
971 ASSERT(BUF_EMPTY(buf
));
972 refcount_destroy(&buf
->b_refcnt
);
973 cv_destroy(&buf
->b_cv
);
974 mutex_destroy(&buf
->b_freeze_lock
);
975 arc_space_return(sizeof (arc_buf_hdr_t
), ARC_SPACE_HDRS
);
980 buf_dest(void *vbuf
, void *unused
)
982 arc_buf_t
*buf
= vbuf
;
984 mutex_destroy(&buf
->b_evict_lock
);
985 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
992 uint64_t hsize
= 1ULL << 12;
996 * The hash table is big enough to fill all of physical memory
997 * with an average 64K block size. The table will take up
998 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
1000 while (hsize
* 65536 < physmem
* PAGESIZE
)
1003 buf_hash_table
.ht_mask
= hsize
- 1;
1004 #if defined(_KERNEL) && defined(HAVE_SPL)
1006 * Large allocations which do not require contiguous pages
1007 * should be using vmem_alloc() in the linux kernel
1009 buf_hash_table
.ht_table
=
1010 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
1012 buf_hash_table
.ht_table
=
1013 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
1015 if (buf_hash_table
.ht_table
== NULL
) {
1016 ASSERT(hsize
> (1ULL << 8));
1021 hdr_cache
= kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t
),
1022 0, hdr_cons
, hdr_dest
, NULL
, NULL
, NULL
, 0);
1023 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
1024 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
1025 l2arc_hdr_cache
= kmem_cache_create("l2arc_buf_hdr_t", L2HDR_SIZE
,
1026 0, NULL
, NULL
, NULL
, NULL
, NULL
, 0);
1028 for (i
= 0; i
< 256; i
++)
1029 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
1030 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
1032 for (i
= 0; i
< BUF_LOCKS
; i
++) {
1033 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
1034 NULL
, MUTEX_DEFAULT
, NULL
);
1038 #define ARC_MINTIME (hz>>4) /* 62 ms */
1041 arc_cksum_verify(arc_buf_t
*buf
)
1045 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1048 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
1049 if (buf
->b_hdr
->b_freeze_cksum
== NULL
||
1050 (buf
->b_hdr
->b_flags
& ARC_IO_ERROR
)) {
1051 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
1054 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
1055 if (!ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
))
1056 panic("buffer modified while frozen!");
1057 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
1061 arc_cksum_equal(arc_buf_t
*buf
)
1066 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
1067 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
1068 equal
= ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
);
1069 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
1075 arc_cksum_compute(arc_buf_t
*buf
, boolean_t force
)
1077 if (!force
&& !(zfs_flags
& ZFS_DEBUG_MODIFY
))
1080 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
1081 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1082 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
1085 buf
->b_hdr
->b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
),
1087 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
,
1088 buf
->b_hdr
->b_freeze_cksum
);
1089 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
1095 arc_buf_sigsegv(int sig
, siginfo_t
*si
, void *unused
)
1097 panic("Got SIGSEGV at address: 0x%lx\n", (long) si
->si_addr
);
1103 arc_buf_unwatch(arc_buf_t
*buf
)
1107 ASSERT0(mprotect(buf
->b_data
, buf
->b_hdr
->b_size
,
1108 PROT_READ
| PROT_WRITE
));
1115 arc_buf_watch(arc_buf_t
*buf
)
1119 ASSERT0(mprotect(buf
->b_data
, buf
->b_hdr
->b_size
, PROT_READ
));
1124 arc_buf_thaw(arc_buf_t
*buf
)
1126 if (zfs_flags
& ZFS_DEBUG_MODIFY
) {
1127 if (buf
->b_hdr
->b_state
!= arc_anon
)
1128 panic("modifying non-anon buffer!");
1129 if (buf
->b_hdr
->b_flags
& ARC_IO_IN_PROGRESS
)
1130 panic("modifying buffer while i/o in progress!");
1131 arc_cksum_verify(buf
);
1134 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
1135 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1136 kmem_free(buf
->b_hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
1137 buf
->b_hdr
->b_freeze_cksum
= NULL
;
1140 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
1142 arc_buf_unwatch(buf
);
1146 arc_buf_freeze(arc_buf_t
*buf
)
1148 kmutex_t
*hash_lock
;
1150 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1153 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1154 mutex_enter(hash_lock
);
1156 ASSERT(buf
->b_hdr
->b_freeze_cksum
!= NULL
||
1157 buf
->b_hdr
->b_state
== arc_anon
);
1158 arc_cksum_compute(buf
, B_FALSE
);
1159 mutex_exit(hash_lock
);
1164 add_reference(arc_buf_hdr_t
*ab
, kmutex_t
*hash_lock
, void *tag
)
1166 ASSERT(MUTEX_HELD(hash_lock
));
1168 if ((refcount_add(&ab
->b_refcnt
, tag
) == 1) &&
1169 (ab
->b_state
!= arc_anon
)) {
1170 uint64_t delta
= ab
->b_size
* ab
->b_datacnt
;
1171 list_t
*list
= &ab
->b_state
->arcs_list
[ab
->b_type
];
1172 uint64_t *size
= &ab
->b_state
->arcs_lsize
[ab
->b_type
];
1174 ASSERT(!MUTEX_HELD(&ab
->b_state
->arcs_mtx
));
1175 mutex_enter(&ab
->b_state
->arcs_mtx
);
1176 ASSERT(list_link_active(&ab
->b_arc_node
));
1177 list_remove(list
, ab
);
1178 if (GHOST_STATE(ab
->b_state
)) {
1179 ASSERT0(ab
->b_datacnt
);
1180 ASSERT3P(ab
->b_buf
, ==, NULL
);
1184 ASSERT3U(*size
, >=, delta
);
1185 atomic_add_64(size
, -delta
);
1186 mutex_exit(&ab
->b_state
->arcs_mtx
);
1187 /* remove the prefetch flag if we get a reference */
1188 if (ab
->b_flags
& ARC_PREFETCH
)
1189 ab
->b_flags
&= ~ARC_PREFETCH
;
1194 remove_reference(arc_buf_hdr_t
*ab
, kmutex_t
*hash_lock
, void *tag
)
1197 arc_state_t
*state
= ab
->b_state
;
1199 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
1200 ASSERT(!GHOST_STATE(state
));
1202 if (((cnt
= refcount_remove(&ab
->b_refcnt
, tag
)) == 0) &&
1203 (state
!= arc_anon
)) {
1204 uint64_t *size
= &state
->arcs_lsize
[ab
->b_type
];
1206 ASSERT(!MUTEX_HELD(&state
->arcs_mtx
));
1207 mutex_enter(&state
->arcs_mtx
);
1208 ASSERT(!list_link_active(&ab
->b_arc_node
));
1209 list_insert_head(&state
->arcs_list
[ab
->b_type
], ab
);
1210 ASSERT(ab
->b_datacnt
> 0);
1211 atomic_add_64(size
, ab
->b_size
* ab
->b_datacnt
);
1212 mutex_exit(&state
->arcs_mtx
);
1218 * Returns detailed information about a specific arc buffer. When the
1219 * state_index argument is set the function will calculate the arc header
1220 * list position for its arc state. Since this requires a linear traversal
1221 * callers are strongly encourage not to do this. However, it can be helpful
1222 * for targeted analysis so the functionality is provided.
1225 arc_buf_info(arc_buf_t
*ab
, arc_buf_info_t
*abi
, int state_index
)
1227 arc_buf_hdr_t
*hdr
= ab
->b_hdr
;
1228 arc_state_t
*state
= hdr
->b_state
;
1230 memset(abi
, 0, sizeof (arc_buf_info_t
));
1231 abi
->abi_flags
= hdr
->b_flags
;
1232 abi
->abi_datacnt
= hdr
->b_datacnt
;
1233 abi
->abi_state_type
= state
? state
->arcs_state
: ARC_STATE_ANON
;
1234 abi
->abi_state_contents
= hdr
->b_type
;
1235 abi
->abi_state_index
= -1;
1236 abi
->abi_size
= hdr
->b_size
;
1237 abi
->abi_access
= hdr
->b_arc_access
;
1238 abi
->abi_mru_hits
= hdr
->b_mru_hits
;
1239 abi
->abi_mru_ghost_hits
= hdr
->b_mru_ghost_hits
;
1240 abi
->abi_mfu_hits
= hdr
->b_mfu_hits
;
1241 abi
->abi_mfu_ghost_hits
= hdr
->b_mfu_ghost_hits
;
1242 abi
->abi_holds
= refcount_count(&hdr
->b_refcnt
);
1245 abi
->abi_l2arc_dattr
= hdr
->b_l2hdr
->b_daddr
;
1246 abi
->abi_l2arc_asize
= hdr
->b_l2hdr
->b_asize
;
1247 abi
->abi_l2arc_compress
= hdr
->b_l2hdr
->b_compress
;
1248 abi
->abi_l2arc_hits
= hdr
->b_l2hdr
->b_hits
;
1251 if (state
&& state_index
&& list_link_active(&hdr
->b_arc_node
)) {
1252 list_t
*list
= &state
->arcs_list
[hdr
->b_type
];
1255 mutex_enter(&state
->arcs_mtx
);
1256 for (h
= list_head(list
); h
!= NULL
; h
= list_next(list
, h
)) {
1257 abi
->abi_state_index
++;
1261 mutex_exit(&state
->arcs_mtx
);
1266 * Move the supplied buffer to the indicated state. The mutex
1267 * for the buffer must be held by the caller.
1270 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*ab
, kmutex_t
*hash_lock
)
1272 arc_state_t
*old_state
= ab
->b_state
;
1273 int64_t refcnt
= refcount_count(&ab
->b_refcnt
);
1274 uint64_t from_delta
, to_delta
;
1276 ASSERT(MUTEX_HELD(hash_lock
));
1277 ASSERT3P(new_state
, !=, old_state
);
1278 ASSERT(refcnt
== 0 || ab
->b_datacnt
> 0);
1279 ASSERT(ab
->b_datacnt
== 0 || !GHOST_STATE(new_state
));
1280 ASSERT(ab
->b_datacnt
<= 1 || old_state
!= arc_anon
);
1282 from_delta
= to_delta
= ab
->b_datacnt
* ab
->b_size
;
1285 * If this buffer is evictable, transfer it from the
1286 * old state list to the new state list.
1289 if (old_state
!= arc_anon
) {
1290 int use_mutex
= !MUTEX_HELD(&old_state
->arcs_mtx
);
1291 uint64_t *size
= &old_state
->arcs_lsize
[ab
->b_type
];
1294 mutex_enter(&old_state
->arcs_mtx
);
1296 ASSERT(list_link_active(&ab
->b_arc_node
));
1297 list_remove(&old_state
->arcs_list
[ab
->b_type
], ab
);
1300 * If prefetching out of the ghost cache,
1301 * we will have a non-zero datacnt.
1303 if (GHOST_STATE(old_state
) && ab
->b_datacnt
== 0) {
1304 /* ghost elements have a ghost size */
1305 ASSERT(ab
->b_buf
== NULL
);
1306 from_delta
= ab
->b_size
;
1308 ASSERT3U(*size
, >=, from_delta
);
1309 atomic_add_64(size
, -from_delta
);
1312 mutex_exit(&old_state
->arcs_mtx
);
1314 if (new_state
!= arc_anon
) {
1315 int use_mutex
= !MUTEX_HELD(&new_state
->arcs_mtx
);
1316 uint64_t *size
= &new_state
->arcs_lsize
[ab
->b_type
];
1319 mutex_enter(&new_state
->arcs_mtx
);
1321 list_insert_head(&new_state
->arcs_list
[ab
->b_type
], ab
);
1323 /* ghost elements have a ghost size */
1324 if (GHOST_STATE(new_state
)) {
1325 ASSERT(ab
->b_datacnt
== 0);
1326 ASSERT(ab
->b_buf
== NULL
);
1327 to_delta
= ab
->b_size
;
1329 atomic_add_64(size
, to_delta
);
1332 mutex_exit(&new_state
->arcs_mtx
);
1336 ASSERT(!BUF_EMPTY(ab
));
1337 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(ab
))
1338 buf_hash_remove(ab
);
1340 /* adjust state sizes */
1342 atomic_add_64(&new_state
->arcs_size
, to_delta
);
1344 ASSERT3U(old_state
->arcs_size
, >=, from_delta
);
1345 atomic_add_64(&old_state
->arcs_size
, -from_delta
);
1347 ab
->b_state
= new_state
;
1349 /* adjust l2arc hdr stats */
1350 if (new_state
== arc_l2c_only
)
1351 l2arc_hdr_stat_add();
1352 else if (old_state
== arc_l2c_only
)
1353 l2arc_hdr_stat_remove();
1357 arc_space_consume(uint64_t space
, arc_space_type_t type
)
1359 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1364 case ARC_SPACE_DATA
:
1365 ARCSTAT_INCR(arcstat_data_size
, space
);
1367 case ARC_SPACE_OTHER
:
1368 ARCSTAT_INCR(arcstat_other_size
, space
);
1370 case ARC_SPACE_HDRS
:
1371 ARCSTAT_INCR(arcstat_hdr_size
, space
);
1373 case ARC_SPACE_L2HDRS
:
1374 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
1378 ARCSTAT_INCR(arcstat_meta_used
, space
);
1379 atomic_add_64(&arc_size
, space
);
1383 arc_space_return(uint64_t space
, arc_space_type_t type
)
1385 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1390 case ARC_SPACE_DATA
:
1391 ARCSTAT_INCR(arcstat_data_size
, -space
);
1393 case ARC_SPACE_OTHER
:
1394 ARCSTAT_INCR(arcstat_other_size
, -space
);
1396 case ARC_SPACE_HDRS
:
1397 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
1399 case ARC_SPACE_L2HDRS
:
1400 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
1404 ASSERT(arc_meta_used
>= space
);
1405 if (arc_meta_max
< arc_meta_used
)
1406 arc_meta_max
= arc_meta_used
;
1407 ARCSTAT_INCR(arcstat_meta_used
, -space
);
1408 ASSERT(arc_size
>= space
);
1409 atomic_add_64(&arc_size
, -space
);
1413 arc_buf_alloc(spa_t
*spa
, int size
, void *tag
, arc_buf_contents_t type
)
1418 ASSERT3U(size
, >, 0);
1419 hdr
= kmem_cache_alloc(hdr_cache
, KM_PUSHPAGE
);
1420 ASSERT(BUF_EMPTY(hdr
));
1423 hdr
->b_spa
= spa_load_guid(spa
);
1424 hdr
->b_state
= arc_anon
;
1425 hdr
->b_arc_access
= 0;
1426 hdr
->b_mru_hits
= 0;
1427 hdr
->b_mru_ghost_hits
= 0;
1428 hdr
->b_mfu_hits
= 0;
1429 hdr
->b_mfu_ghost_hits
= 0;
1431 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1434 buf
->b_efunc
= NULL
;
1435 buf
->b_private
= NULL
;
1438 arc_get_data_buf(buf
);
1441 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
1442 (void) refcount_add(&hdr
->b_refcnt
, tag
);
1447 static char *arc_onloan_tag
= "onloan";
1450 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1451 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1452 * buffers must be returned to the arc before they can be used by the DMU or
1456 arc_loan_buf(spa_t
*spa
, int size
)
1460 buf
= arc_buf_alloc(spa
, size
, arc_onloan_tag
, ARC_BUFC_DATA
);
1462 atomic_add_64(&arc_loaned_bytes
, size
);
1467 * Return a loaned arc buffer to the arc.
1470 arc_return_buf(arc_buf_t
*buf
, void *tag
)
1472 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1474 ASSERT(buf
->b_data
!= NULL
);
1475 (void) refcount_add(&hdr
->b_refcnt
, tag
);
1476 (void) refcount_remove(&hdr
->b_refcnt
, arc_onloan_tag
);
1478 atomic_add_64(&arc_loaned_bytes
, -hdr
->b_size
);
1481 /* Detach an arc_buf from a dbuf (tag) */
1483 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
1487 ASSERT(buf
->b_data
!= NULL
);
1489 (void) refcount_add(&hdr
->b_refcnt
, arc_onloan_tag
);
1490 (void) refcount_remove(&hdr
->b_refcnt
, tag
);
1491 buf
->b_efunc
= NULL
;
1492 buf
->b_private
= NULL
;
1494 atomic_add_64(&arc_loaned_bytes
, hdr
->b_size
);
1498 arc_buf_clone(arc_buf_t
*from
)
1501 arc_buf_hdr_t
*hdr
= from
->b_hdr
;
1502 uint64_t size
= hdr
->b_size
;
1504 ASSERT(hdr
->b_state
!= arc_anon
);
1506 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1509 buf
->b_efunc
= NULL
;
1510 buf
->b_private
= NULL
;
1511 buf
->b_next
= hdr
->b_buf
;
1513 arc_get_data_buf(buf
);
1514 bcopy(from
->b_data
, buf
->b_data
, size
);
1517 * This buffer already exists in the arc so create a duplicate
1518 * copy for the caller. If the buffer is associated with user data
1519 * then track the size and number of duplicates. These stats will be
1520 * updated as duplicate buffers are created and destroyed.
1522 if (hdr
->b_type
== ARC_BUFC_DATA
) {
1523 ARCSTAT_BUMP(arcstat_duplicate_buffers
);
1524 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, size
);
1526 hdr
->b_datacnt
+= 1;
1531 arc_buf_add_ref(arc_buf_t
*buf
, void* tag
)
1534 kmutex_t
*hash_lock
;
1537 * Check to see if this buffer is evicted. Callers
1538 * must verify b_data != NULL to know if the add_ref
1541 mutex_enter(&buf
->b_evict_lock
);
1542 if (buf
->b_data
== NULL
) {
1543 mutex_exit(&buf
->b_evict_lock
);
1546 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1547 mutex_enter(hash_lock
);
1549 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1550 mutex_exit(&buf
->b_evict_lock
);
1552 ASSERT(hdr
->b_state
== arc_mru
|| hdr
->b_state
== arc_mfu
);
1553 add_reference(hdr
, hash_lock
, tag
);
1554 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
1555 arc_access(hdr
, hash_lock
);
1556 mutex_exit(hash_lock
);
1557 ARCSTAT_BUMP(arcstat_hits
);
1558 ARCSTAT_CONDSTAT(!(hdr
->b_flags
& ARC_PREFETCH
),
1559 demand
, prefetch
, hdr
->b_type
!= ARC_BUFC_METADATA
,
1560 data
, metadata
, hits
);
1564 * Free the arc data buffer. If it is an l2arc write in progress,
1565 * the buffer is placed on l2arc_free_on_write to be freed later.
1568 arc_buf_data_free(arc_buf_t
*buf
, void (*free_func
)(void *, size_t))
1570 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1572 if (HDR_L2_WRITING(hdr
)) {
1573 l2arc_data_free_t
*df
;
1574 df
= kmem_alloc(sizeof (l2arc_data_free_t
), KM_PUSHPAGE
);
1575 df
->l2df_data
= buf
->b_data
;
1576 df
->l2df_size
= hdr
->b_size
;
1577 df
->l2df_func
= free_func
;
1578 mutex_enter(&l2arc_free_on_write_mtx
);
1579 list_insert_head(l2arc_free_on_write
, df
);
1580 mutex_exit(&l2arc_free_on_write_mtx
);
1581 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
1583 free_func(buf
->b_data
, hdr
->b_size
);
1588 arc_buf_destroy(arc_buf_t
*buf
, boolean_t recycle
, boolean_t all
)
1592 /* free up data associated with the buf */
1594 arc_state_t
*state
= buf
->b_hdr
->b_state
;
1595 uint64_t size
= buf
->b_hdr
->b_size
;
1596 arc_buf_contents_t type
= buf
->b_hdr
->b_type
;
1598 arc_cksum_verify(buf
);
1599 arc_buf_unwatch(buf
);
1602 if (type
== ARC_BUFC_METADATA
) {
1603 arc_buf_data_free(buf
, zio_buf_free
);
1604 arc_space_return(size
, ARC_SPACE_DATA
);
1606 ASSERT(type
== ARC_BUFC_DATA
);
1607 arc_buf_data_free(buf
, zio_data_buf_free
);
1608 ARCSTAT_INCR(arcstat_data_size
, -size
);
1609 atomic_add_64(&arc_size
, -size
);
1612 if (list_link_active(&buf
->b_hdr
->b_arc_node
)) {
1613 uint64_t *cnt
= &state
->arcs_lsize
[type
];
1615 ASSERT(refcount_is_zero(&buf
->b_hdr
->b_refcnt
));
1616 ASSERT(state
!= arc_anon
);
1618 ASSERT3U(*cnt
, >=, size
);
1619 atomic_add_64(cnt
, -size
);
1621 ASSERT3U(state
->arcs_size
, >=, size
);
1622 atomic_add_64(&state
->arcs_size
, -size
);
1626 * If we're destroying a duplicate buffer make sure
1627 * that the appropriate statistics are updated.
1629 if (buf
->b_hdr
->b_datacnt
> 1 &&
1630 buf
->b_hdr
->b_type
== ARC_BUFC_DATA
) {
1631 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
1632 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, -size
);
1634 ASSERT(buf
->b_hdr
->b_datacnt
> 0);
1635 buf
->b_hdr
->b_datacnt
-= 1;
1638 /* only remove the buf if requested */
1642 /* remove the buf from the hdr list */
1643 for (bufp
= &buf
->b_hdr
->b_buf
; *bufp
!= buf
; bufp
= &(*bufp
)->b_next
)
1645 *bufp
= buf
->b_next
;
1648 ASSERT(buf
->b_efunc
== NULL
);
1650 /* clean up the buf */
1652 kmem_cache_free(buf_cache
, buf
);
1656 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
1658 l2arc_buf_hdr_t
*l2hdr
= hdr
->b_l2hdr
;
1660 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
1661 ASSERT3P(hdr
->b_state
, ==, arc_anon
);
1662 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
1664 if (l2hdr
!= NULL
) {
1665 boolean_t buflist_held
= MUTEX_HELD(&l2arc_buflist_mtx
);
1667 * To prevent arc_free() and l2arc_evict() from
1668 * attempting to free the same buffer at the same time,
1669 * a FREE_IN_PROGRESS flag is given to arc_free() to
1670 * give it priority. l2arc_evict() can't destroy this
1671 * header while we are waiting on l2arc_buflist_mtx.
1673 * The hdr may be removed from l2ad_buflist before we
1674 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1676 if (!buflist_held
) {
1677 mutex_enter(&l2arc_buflist_mtx
);
1678 l2hdr
= hdr
->b_l2hdr
;
1681 if (l2hdr
!= NULL
) {
1682 list_remove(l2hdr
->b_dev
->l2ad_buflist
, hdr
);
1683 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
1684 ARCSTAT_INCR(arcstat_l2_asize
, -l2hdr
->b_asize
);
1685 kmem_cache_free(l2arc_hdr_cache
, l2hdr
);
1686 arc_space_return(L2HDR_SIZE
, ARC_SPACE_L2HDRS
);
1687 if (hdr
->b_state
== arc_l2c_only
)
1688 l2arc_hdr_stat_remove();
1689 hdr
->b_l2hdr
= NULL
;
1693 mutex_exit(&l2arc_buflist_mtx
);
1696 if (!BUF_EMPTY(hdr
)) {
1697 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
1698 buf_discard_identity(hdr
);
1700 while (hdr
->b_buf
) {
1701 arc_buf_t
*buf
= hdr
->b_buf
;
1704 mutex_enter(&arc_eviction_mtx
);
1705 mutex_enter(&buf
->b_evict_lock
);
1706 ASSERT(buf
->b_hdr
!= NULL
);
1707 arc_buf_destroy(hdr
->b_buf
, FALSE
, FALSE
);
1708 hdr
->b_buf
= buf
->b_next
;
1709 buf
->b_hdr
= &arc_eviction_hdr
;
1710 buf
->b_next
= arc_eviction_list
;
1711 arc_eviction_list
= buf
;
1712 mutex_exit(&buf
->b_evict_lock
);
1713 mutex_exit(&arc_eviction_mtx
);
1715 arc_buf_destroy(hdr
->b_buf
, FALSE
, TRUE
);
1718 if (hdr
->b_freeze_cksum
!= NULL
) {
1719 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
1720 hdr
->b_freeze_cksum
= NULL
;
1723 ASSERT(!list_link_active(&hdr
->b_arc_node
));
1724 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
1725 ASSERT3P(hdr
->b_acb
, ==, NULL
);
1726 kmem_cache_free(hdr_cache
, hdr
);
1730 arc_buf_free(arc_buf_t
*buf
, void *tag
)
1732 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1733 int hashed
= hdr
->b_state
!= arc_anon
;
1735 ASSERT(buf
->b_efunc
== NULL
);
1736 ASSERT(buf
->b_data
!= NULL
);
1739 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
1741 mutex_enter(hash_lock
);
1743 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1745 (void) remove_reference(hdr
, hash_lock
, tag
);
1746 if (hdr
->b_datacnt
> 1) {
1747 arc_buf_destroy(buf
, FALSE
, TRUE
);
1749 ASSERT(buf
== hdr
->b_buf
);
1750 ASSERT(buf
->b_efunc
== NULL
);
1751 hdr
->b_flags
|= ARC_BUF_AVAILABLE
;
1753 mutex_exit(hash_lock
);
1754 } else if (HDR_IO_IN_PROGRESS(hdr
)) {
1757 * We are in the middle of an async write. Don't destroy
1758 * this buffer unless the write completes before we finish
1759 * decrementing the reference count.
1761 mutex_enter(&arc_eviction_mtx
);
1762 (void) remove_reference(hdr
, NULL
, tag
);
1763 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
1764 destroy_hdr
= !HDR_IO_IN_PROGRESS(hdr
);
1765 mutex_exit(&arc_eviction_mtx
);
1767 arc_hdr_destroy(hdr
);
1769 if (remove_reference(hdr
, NULL
, tag
) > 0)
1770 arc_buf_destroy(buf
, FALSE
, TRUE
);
1772 arc_hdr_destroy(hdr
);
1777 arc_buf_remove_ref(arc_buf_t
*buf
, void* tag
)
1779 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1780 kmutex_t
*hash_lock
= NULL
;
1781 boolean_t no_callback
= (buf
->b_efunc
== NULL
);
1783 if (hdr
->b_state
== arc_anon
) {
1784 ASSERT(hdr
->b_datacnt
== 1);
1785 arc_buf_free(buf
, tag
);
1786 return (no_callback
);
1789 hash_lock
= HDR_LOCK(hdr
);
1790 mutex_enter(hash_lock
);
1792 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1793 ASSERT(hdr
->b_state
!= arc_anon
);
1794 ASSERT(buf
->b_data
!= NULL
);
1796 (void) remove_reference(hdr
, hash_lock
, tag
);
1797 if (hdr
->b_datacnt
> 1) {
1799 arc_buf_destroy(buf
, FALSE
, TRUE
);
1800 } else if (no_callback
) {
1801 ASSERT(hdr
->b_buf
== buf
&& buf
->b_next
== NULL
);
1802 ASSERT(buf
->b_efunc
== NULL
);
1803 hdr
->b_flags
|= ARC_BUF_AVAILABLE
;
1805 ASSERT(no_callback
|| hdr
->b_datacnt
> 1 ||
1806 refcount_is_zero(&hdr
->b_refcnt
));
1807 mutex_exit(hash_lock
);
1808 return (no_callback
);
1812 arc_buf_size(arc_buf_t
*buf
)
1814 return (buf
->b_hdr
->b_size
);
1818 * Called from the DMU to determine if the current buffer should be
1819 * evicted. In order to ensure proper locking, the eviction must be initiated
1820 * from the DMU. Return true if the buffer is associated with user data and
1821 * duplicate buffers still exist.
1824 arc_buf_eviction_needed(arc_buf_t
*buf
)
1827 boolean_t evict_needed
= B_FALSE
;
1829 if (zfs_disable_dup_eviction
)
1832 mutex_enter(&buf
->b_evict_lock
);
1836 * We are in arc_do_user_evicts(); let that function
1837 * perform the eviction.
1839 ASSERT(buf
->b_data
== NULL
);
1840 mutex_exit(&buf
->b_evict_lock
);
1842 } else if (buf
->b_data
== NULL
) {
1844 * We have already been added to the arc eviction list;
1845 * recommend eviction.
1847 ASSERT3P(hdr
, ==, &arc_eviction_hdr
);
1848 mutex_exit(&buf
->b_evict_lock
);
1852 if (hdr
->b_datacnt
> 1 && hdr
->b_type
== ARC_BUFC_DATA
)
1853 evict_needed
= B_TRUE
;
1855 mutex_exit(&buf
->b_evict_lock
);
1856 return (evict_needed
);
1860 * Evict buffers from list until we've removed the specified number of
1861 * bytes. Move the removed buffers to the appropriate evict state.
1862 * If the recycle flag is set, then attempt to "recycle" a buffer:
1863 * - look for a buffer to evict that is `bytes' long.
1864 * - return the data block from this buffer rather than freeing it.
1865 * This flag is used by callers that are trying to make space for a
1866 * new buffer in a full arc cache.
1868 * This function makes a "best effort". It skips over any buffers
1869 * it can't get a hash_lock on, and so may not catch all candidates.
1870 * It may also return without evicting as much space as requested.
1873 arc_evict(arc_state_t
*state
, uint64_t spa
, int64_t bytes
, boolean_t recycle
,
1874 arc_buf_contents_t type
)
1876 arc_state_t
*evicted_state
;
1877 uint64_t bytes_evicted
= 0, skipped
= 0, missed
= 0;
1878 arc_buf_hdr_t
*ab
, *ab_prev
= NULL
;
1879 list_t
*list
= &state
->arcs_list
[type
];
1880 kmutex_t
*hash_lock
;
1881 boolean_t have_lock
;
1882 void *stolen
= NULL
;
1883 arc_buf_hdr_t marker
= {{{ 0 }}};
1886 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
1888 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
1890 mutex_enter(&state
->arcs_mtx
);
1891 mutex_enter(&evicted_state
->arcs_mtx
);
1893 for (ab
= list_tail(list
); ab
; ab
= ab_prev
) {
1894 ab_prev
= list_prev(list
, ab
);
1895 /* prefetch buffers have a minimum lifespan */
1896 if (HDR_IO_IN_PROGRESS(ab
) ||
1897 (spa
&& ab
->b_spa
!= spa
) ||
1898 (ab
->b_flags
& (ARC_PREFETCH
|ARC_INDIRECT
) &&
1899 ddi_get_lbolt() - ab
->b_arc_access
<
1900 zfs_arc_min_prefetch_lifespan
)) {
1904 /* "lookahead" for better eviction candidate */
1905 if (recycle
&& ab
->b_size
!= bytes
&&
1906 ab_prev
&& ab_prev
->b_size
== bytes
)
1909 /* ignore markers */
1914 * It may take a long time to evict all the bufs requested.
1915 * To avoid blocking all arc activity, periodically drop
1916 * the arcs_mtx and give other threads a chance to run
1917 * before reacquiring the lock.
1919 * If we are looking for a buffer to recycle, we are in
1920 * the hot code path, so don't sleep.
1922 if (!recycle
&& count
++ > arc_evict_iterations
) {
1923 list_insert_after(list
, ab
, &marker
);
1924 mutex_exit(&evicted_state
->arcs_mtx
);
1925 mutex_exit(&state
->arcs_mtx
);
1926 kpreempt(KPREEMPT_SYNC
);
1927 mutex_enter(&state
->arcs_mtx
);
1928 mutex_enter(&evicted_state
->arcs_mtx
);
1929 ab_prev
= list_prev(list
, &marker
);
1930 list_remove(list
, &marker
);
1935 hash_lock
= HDR_LOCK(ab
);
1936 have_lock
= MUTEX_HELD(hash_lock
);
1937 if (have_lock
|| mutex_tryenter(hash_lock
)) {
1938 ASSERT0(refcount_count(&ab
->b_refcnt
));
1939 ASSERT(ab
->b_datacnt
> 0);
1941 arc_buf_t
*buf
= ab
->b_buf
;
1942 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
1947 bytes_evicted
+= ab
->b_size
;
1948 if (recycle
&& ab
->b_type
== type
&&
1949 ab
->b_size
== bytes
&&
1950 !HDR_L2_WRITING(ab
)) {
1951 stolen
= buf
->b_data
;
1956 mutex_enter(&arc_eviction_mtx
);
1957 arc_buf_destroy(buf
,
1958 buf
->b_data
== stolen
, FALSE
);
1959 ab
->b_buf
= buf
->b_next
;
1960 buf
->b_hdr
= &arc_eviction_hdr
;
1961 buf
->b_next
= arc_eviction_list
;
1962 arc_eviction_list
= buf
;
1963 mutex_exit(&arc_eviction_mtx
);
1964 mutex_exit(&buf
->b_evict_lock
);
1966 mutex_exit(&buf
->b_evict_lock
);
1967 arc_buf_destroy(buf
,
1968 buf
->b_data
== stolen
, TRUE
);
1973 ARCSTAT_INCR(arcstat_evict_l2_cached
,
1976 if (l2arc_write_eligible(ab
->b_spa
, ab
)) {
1977 ARCSTAT_INCR(arcstat_evict_l2_eligible
,
1981 arcstat_evict_l2_ineligible
,
1986 if (ab
->b_datacnt
== 0) {
1987 arc_change_state(evicted_state
, ab
, hash_lock
);
1988 ASSERT(HDR_IN_HASH_TABLE(ab
));
1989 ab
->b_flags
|= ARC_IN_HASH_TABLE
;
1990 ab
->b_flags
&= ~ARC_BUF_AVAILABLE
;
1991 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, ab
);
1994 mutex_exit(hash_lock
);
1995 if (bytes
>= 0 && bytes_evicted
>= bytes
)
2002 mutex_exit(&evicted_state
->arcs_mtx
);
2003 mutex_exit(&state
->arcs_mtx
);
2005 if (bytes_evicted
< bytes
)
2006 dprintf("only evicted %lld bytes from %x\n",
2007 (longlong_t
)bytes_evicted
, state
);
2010 ARCSTAT_INCR(arcstat_evict_skip
, skipped
);
2013 ARCSTAT_INCR(arcstat_mutex_miss
, missed
);
2016 * Note: we have just evicted some data into the ghost state,
2017 * potentially putting the ghost size over the desired size. Rather
2018 * that evicting from the ghost list in this hot code path, leave
2019 * this chore to the arc_reclaim_thread().
2026 * Remove buffers from list until we've removed the specified number of
2027 * bytes. Destroy the buffers that are removed.
2030 arc_evict_ghost(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
2031 arc_buf_contents_t type
)
2033 arc_buf_hdr_t
*ab
, *ab_prev
;
2034 arc_buf_hdr_t marker
;
2035 list_t
*list
= &state
->arcs_list
[type
];
2036 kmutex_t
*hash_lock
;
2037 uint64_t bytes_deleted
= 0;
2038 uint64_t bufs_skipped
= 0;
2041 ASSERT(GHOST_STATE(state
));
2042 bzero(&marker
, sizeof (marker
));
2044 mutex_enter(&state
->arcs_mtx
);
2045 for (ab
= list_tail(list
); ab
; ab
= ab_prev
) {
2046 ab_prev
= list_prev(list
, ab
);
2047 if (ab
->b_type
> ARC_BUFC_NUMTYPES
)
2048 panic("invalid ab=%p", (void *)ab
);
2049 if (spa
&& ab
->b_spa
!= spa
)
2052 /* ignore markers */
2056 hash_lock
= HDR_LOCK(ab
);
2057 /* caller may be trying to modify this buffer, skip it */
2058 if (MUTEX_HELD(hash_lock
))
2062 * It may take a long time to evict all the bufs requested.
2063 * To avoid blocking all arc activity, periodically drop
2064 * the arcs_mtx and give other threads a chance to run
2065 * before reacquiring the lock.
2067 if (count
++ > arc_evict_iterations
) {
2068 list_insert_after(list
, ab
, &marker
);
2069 mutex_exit(&state
->arcs_mtx
);
2070 kpreempt(KPREEMPT_SYNC
);
2071 mutex_enter(&state
->arcs_mtx
);
2072 ab_prev
= list_prev(list
, &marker
);
2073 list_remove(list
, &marker
);
2077 if (mutex_tryenter(hash_lock
)) {
2078 ASSERT(!HDR_IO_IN_PROGRESS(ab
));
2079 ASSERT(ab
->b_buf
== NULL
);
2080 ARCSTAT_BUMP(arcstat_deleted
);
2081 bytes_deleted
+= ab
->b_size
;
2083 if (ab
->b_l2hdr
!= NULL
) {
2085 * This buffer is cached on the 2nd Level ARC;
2086 * don't destroy the header.
2088 arc_change_state(arc_l2c_only
, ab
, hash_lock
);
2089 mutex_exit(hash_lock
);
2091 arc_change_state(arc_anon
, ab
, hash_lock
);
2092 mutex_exit(hash_lock
);
2093 arc_hdr_destroy(ab
);
2096 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, ab
);
2097 if (bytes
>= 0 && bytes_deleted
>= bytes
)
2099 } else if (bytes
< 0) {
2101 * Insert a list marker and then wait for the
2102 * hash lock to become available. Once its
2103 * available, restart from where we left off.
2105 list_insert_after(list
, ab
, &marker
);
2106 mutex_exit(&state
->arcs_mtx
);
2107 mutex_enter(hash_lock
);
2108 mutex_exit(hash_lock
);
2109 mutex_enter(&state
->arcs_mtx
);
2110 ab_prev
= list_prev(list
, &marker
);
2111 list_remove(list
, &marker
);
2116 mutex_exit(&state
->arcs_mtx
);
2118 if (list
== &state
->arcs_list
[ARC_BUFC_DATA
] &&
2119 (bytes
< 0 || bytes_deleted
< bytes
)) {
2120 list
= &state
->arcs_list
[ARC_BUFC_METADATA
];
2125 ARCSTAT_INCR(arcstat_mutex_miss
, bufs_skipped
);
2129 if (bytes_deleted
< bytes
)
2130 dprintf("only deleted %lld bytes from %p\n",
2131 (longlong_t
)bytes_deleted
, state
);
2137 int64_t adjustment
, delta
;
2143 adjustment
= MIN((int64_t)(arc_size
- arc_c
),
2144 (int64_t)(arc_anon
->arcs_size
+ arc_mru
->arcs_size
+ arc_meta_used
-
2147 if (adjustment
> 0 && arc_mru
->arcs_lsize
[ARC_BUFC_DATA
] > 0) {
2148 delta
= MIN(arc_mru
->arcs_lsize
[ARC_BUFC_DATA
], adjustment
);
2149 (void) arc_evict(arc_mru
, 0, delta
, FALSE
, ARC_BUFC_DATA
);
2150 adjustment
-= delta
;
2153 if (adjustment
> 0 && arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] > 0) {
2154 delta
= MIN(arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
], adjustment
);
2155 (void) arc_evict(arc_mru
, 0, delta
, FALSE
,
2163 adjustment
= arc_size
- arc_c
;
2165 if (adjustment
> 0 && arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
] > 0) {
2166 delta
= MIN(adjustment
, arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
]);
2167 (void) arc_evict(arc_mfu
, 0, delta
, FALSE
, ARC_BUFC_DATA
);
2168 adjustment
-= delta
;
2171 if (adjustment
> 0 && arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
] > 0) {
2172 int64_t delta
= MIN(adjustment
,
2173 arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
]);
2174 (void) arc_evict(arc_mfu
, 0, delta
, FALSE
,
2179 * Adjust ghost lists
2182 adjustment
= arc_mru
->arcs_size
+ arc_mru_ghost
->arcs_size
- arc_c
;
2184 if (adjustment
> 0 && arc_mru_ghost
->arcs_size
> 0) {
2185 delta
= MIN(arc_mru_ghost
->arcs_size
, adjustment
);
2186 arc_evict_ghost(arc_mru_ghost
, 0, delta
, ARC_BUFC_DATA
);
2190 arc_mru_ghost
->arcs_size
+ arc_mfu_ghost
->arcs_size
- arc_c
;
2192 if (adjustment
> 0 && arc_mfu_ghost
->arcs_size
> 0) {
2193 delta
= MIN(arc_mfu_ghost
->arcs_size
, adjustment
);
2194 arc_evict_ghost(arc_mfu_ghost
, 0, delta
, ARC_BUFC_DATA
);
2199 * Request that arc user drop references so that N bytes can be released
2200 * from the cache. This provides a mechanism to ensure the arc can honor
2201 * the arc_meta_limit and reclaim buffers which are pinned in the cache
2202 * by higher layers. (i.e. the zpl)
2205 arc_do_user_prune(int64_t adjustment
)
2207 arc_prune_func_t
*func
;
2209 arc_prune_t
*cp
, *np
;
2211 mutex_enter(&arc_prune_mtx
);
2213 cp
= list_head(&arc_prune_list
);
2214 while (cp
!= NULL
) {
2216 private = cp
->p_private
;
2217 np
= list_next(&arc_prune_list
, cp
);
2218 refcount_add(&cp
->p_refcnt
, func
);
2219 mutex_exit(&arc_prune_mtx
);
2222 func(adjustment
, private);
2224 mutex_enter(&arc_prune_mtx
);
2226 /* User removed prune callback concurrently with execution */
2227 if (refcount_remove(&cp
->p_refcnt
, func
) == 0) {
2228 ASSERT(!list_link_active(&cp
->p_node
));
2229 refcount_destroy(&cp
->p_refcnt
);
2230 kmem_free(cp
, sizeof (*cp
));
2236 ARCSTAT_BUMP(arcstat_prune
);
2237 mutex_exit(&arc_prune_mtx
);
2241 arc_do_user_evicts(void)
2243 mutex_enter(&arc_eviction_mtx
);
2244 while (arc_eviction_list
!= NULL
) {
2245 arc_buf_t
*buf
= arc_eviction_list
;
2246 arc_eviction_list
= buf
->b_next
;
2247 mutex_enter(&buf
->b_evict_lock
);
2249 mutex_exit(&buf
->b_evict_lock
);
2250 mutex_exit(&arc_eviction_mtx
);
2252 if (buf
->b_efunc
!= NULL
)
2253 VERIFY(buf
->b_efunc(buf
) == 0);
2255 buf
->b_efunc
= NULL
;
2256 buf
->b_private
= NULL
;
2257 kmem_cache_free(buf_cache
, buf
);
2258 mutex_enter(&arc_eviction_mtx
);
2260 mutex_exit(&arc_eviction_mtx
);
2264 * Evict only meta data objects from the cache leaving the data objects.
2265 * This is only used to enforce the tunable arc_meta_limit, if we are
2266 * unable to evict enough buffers notify the user via the prune callback.
2269 arc_adjust_meta(int64_t adjustment
, boolean_t may_prune
)
2273 if (adjustment
> 0 && arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] > 0) {
2274 delta
= MIN(arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
], adjustment
);
2275 arc_evict(arc_mru
, 0, delta
, FALSE
, ARC_BUFC_METADATA
);
2276 adjustment
-= delta
;
2279 if (adjustment
> 0 && arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
] > 0) {
2280 delta
= MIN(arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
], adjustment
);
2281 arc_evict(arc_mfu
, 0, delta
, FALSE
, ARC_BUFC_METADATA
);
2282 adjustment
-= delta
;
2285 if (may_prune
&& (adjustment
> 0) && (arc_meta_used
> arc_meta_limit
))
2286 arc_do_user_prune(zfs_arc_meta_prune
);
2290 * Flush all *evictable* data from the cache for the given spa.
2291 * NOTE: this will not touch "active" (i.e. referenced) data.
2294 arc_flush(spa_t
*spa
)
2299 guid
= spa_load_guid(spa
);
2301 while (list_head(&arc_mru
->arcs_list
[ARC_BUFC_DATA
])) {
2302 (void) arc_evict(arc_mru
, guid
, -1, FALSE
, ARC_BUFC_DATA
);
2306 while (list_head(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
])) {
2307 (void) arc_evict(arc_mru
, guid
, -1, FALSE
, ARC_BUFC_METADATA
);
2311 while (list_head(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
])) {
2312 (void) arc_evict(arc_mfu
, guid
, -1, FALSE
, ARC_BUFC_DATA
);
2316 while (list_head(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
])) {
2317 (void) arc_evict(arc_mfu
, guid
, -1, FALSE
, ARC_BUFC_METADATA
);
2322 arc_evict_ghost(arc_mru_ghost
, guid
, -1, ARC_BUFC_DATA
);
2323 arc_evict_ghost(arc_mfu_ghost
, guid
, -1, ARC_BUFC_DATA
);
2325 mutex_enter(&arc_reclaim_thr_lock
);
2326 arc_do_user_evicts();
2327 mutex_exit(&arc_reclaim_thr_lock
);
2328 ASSERT(spa
|| arc_eviction_list
== NULL
);
2332 arc_shrink(uint64_t bytes
)
2334 if (arc_c
> arc_c_min
) {
2338 to_free
= bytes
? bytes
: arc_c
>> zfs_arc_shrink_shift
;
2340 if (arc_c
> arc_c_min
+ to_free
)
2341 atomic_add_64(&arc_c
, -to_free
);
2345 arc_p_min
= (arc_c
>> zfs_arc_p_min_shift
);
2346 to_free
= bytes
? bytes
: arc_p
>> zfs_arc_shrink_shift
;
2348 if (arc_p
> arc_p_min
+ to_free
)
2349 atomic_add_64(&arc_p
, -to_free
);
2353 if (arc_c
> arc_size
)
2354 arc_c
= MAX(arc_size
, arc_c_min
);
2356 arc_p
= (arc_c
>> 1);
2357 ASSERT(arc_c
>= arc_c_min
);
2358 ASSERT((int64_t)arc_p
>= 0);
2361 if (arc_size
> arc_c
)
2366 arc_kmem_reap_now(arc_reclaim_strategy_t strat
, uint64_t bytes
)
2369 kmem_cache_t
*prev_cache
= NULL
;
2370 kmem_cache_t
*prev_data_cache
= NULL
;
2371 extern kmem_cache_t
*zio_buf_cache
[];
2372 extern kmem_cache_t
*zio_data_buf_cache
[];
2375 * An aggressive reclamation will shrink the cache size as well as
2376 * reap free buffers from the arc kmem caches.
2378 if (strat
== ARC_RECLAIM_AGGR
)
2381 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
2382 if (zio_buf_cache
[i
] != prev_cache
) {
2383 prev_cache
= zio_buf_cache
[i
];
2384 kmem_cache_reap_now(zio_buf_cache
[i
]);
2386 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
2387 prev_data_cache
= zio_data_buf_cache
[i
];
2388 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
2392 kmem_cache_reap_now(buf_cache
);
2393 kmem_cache_reap_now(hdr_cache
);
2397 * Unlike other ZFS implementations this thread is only responsible for
2398 * adapting the target ARC size on Linux. The responsibility for memory
2399 * reclamation has been entirely delegated to the arc_shrinker_func()
2400 * which is registered with the VM. To reflect this change in behavior
2401 * the arc_reclaim thread has been renamed to arc_adapt.
2404 arc_adapt_thread(void)
2409 CALLB_CPR_INIT(&cpr
, &arc_reclaim_thr_lock
, callb_generic_cpr
, FTAG
);
2411 mutex_enter(&arc_reclaim_thr_lock
);
2412 while (arc_thread_exit
== 0) {
2414 arc_reclaim_strategy_t last_reclaim
= ARC_RECLAIM_CONS
;
2416 if (spa_get_random(100) == 0) {
2419 if (last_reclaim
== ARC_RECLAIM_CONS
) {
2420 last_reclaim
= ARC_RECLAIM_AGGR
;
2422 last_reclaim
= ARC_RECLAIM_CONS
;
2426 last_reclaim
= ARC_RECLAIM_AGGR
;
2430 /* reset the growth delay for every reclaim */
2431 arc_grow_time
= ddi_get_lbolt() +
2432 (zfs_arc_grow_retry
* hz
);
2434 arc_kmem_reap_now(last_reclaim
, 0);
2437 #endif /* !_KERNEL */
2439 /* No recent memory pressure allow the ARC to grow. */
2440 if (arc_no_grow
&& ddi_get_lbolt() >= arc_grow_time
)
2441 arc_no_grow
= FALSE
;
2444 * Keep meta data usage within limits, arc_shrink() is not
2445 * used to avoid collapsing the arc_c value when only the
2446 * arc_meta_limit is being exceeded.
2448 prune
= (int64_t)arc_meta_used
- (int64_t)arc_meta_limit
;
2450 arc_adjust_meta(prune
, B_TRUE
);
2454 if (arc_eviction_list
!= NULL
)
2455 arc_do_user_evicts();
2457 /* block until needed, or one second, whichever is shorter */
2458 CALLB_CPR_SAFE_BEGIN(&cpr
);
2459 (void) cv_timedwait_interruptible(&arc_reclaim_thr_cv
,
2460 &arc_reclaim_thr_lock
, (ddi_get_lbolt() + hz
));
2461 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_thr_lock
);
2464 /* Allow the module options to be changed */
2465 if (zfs_arc_max
> 64 << 20 &&
2466 zfs_arc_max
< physmem
* PAGESIZE
&&
2467 zfs_arc_max
!= arc_c_max
)
2468 arc_c_max
= zfs_arc_max
;
2470 if (zfs_arc_min
> 0 &&
2471 zfs_arc_min
< arc_c_max
&&
2472 zfs_arc_min
!= arc_c_min
)
2473 arc_c_min
= zfs_arc_min
;
2475 if (zfs_arc_meta_limit
> 0 &&
2476 zfs_arc_meta_limit
<= arc_c_max
&&
2477 zfs_arc_meta_limit
!= arc_meta_limit
)
2478 arc_meta_limit
= zfs_arc_meta_limit
;
2484 arc_thread_exit
= 0;
2485 cv_broadcast(&arc_reclaim_thr_cv
);
2486 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_thr_lock */
2492 * Determine the amount of memory eligible for eviction contained in the
2493 * ARC. All clean data reported by the ghost lists can always be safely
2494 * evicted. Due to arc_c_min, the same does not hold for all clean data
2495 * contained by the regular mru and mfu lists.
2497 * In the case of the regular mru and mfu lists, we need to report as
2498 * much clean data as possible, such that evicting that same reported
2499 * data will not bring arc_size below arc_c_min. Thus, in certain
2500 * circumstances, the total amount of clean data in the mru and mfu
2501 * lists might not actually be evictable.
2503 * The following two distinct cases are accounted for:
2505 * 1. The sum of the amount of dirty data contained by both the mru and
2506 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2507 * is greater than or equal to arc_c_min.
2508 * (i.e. amount of dirty data >= arc_c_min)
2510 * This is the easy case; all clean data contained by the mru and mfu
2511 * lists is evictable. Evicting all clean data can only drop arc_size
2512 * to the amount of dirty data, which is greater than arc_c_min.
2514 * 2. The sum of the amount of dirty data contained by both the mru and
2515 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
2516 * is less than arc_c_min.
2517 * (i.e. arc_c_min > amount of dirty data)
2519 * 2.1. arc_size is greater than or equal arc_c_min.
2520 * (i.e. arc_size >= arc_c_min > amount of dirty data)
2522 * In this case, not all clean data from the regular mru and mfu
2523 * lists is actually evictable; we must leave enough clean data
2524 * to keep arc_size above arc_c_min. Thus, the maximum amount of
2525 * evictable data from the two lists combined, is exactly the
2526 * difference between arc_size and arc_c_min.
2528 * 2.2. arc_size is less than arc_c_min
2529 * (i.e. arc_c_min > arc_size > amount of dirty data)
2531 * In this case, none of the data contained in the mru and mfu
2532 * lists is evictable, even if it's clean. Since arc_size is
2533 * already below arc_c_min, evicting any more would only
2534 * increase this negative difference.
2537 arc_evictable_memory(void) {
2538 uint64_t arc_clean
=
2539 arc_mru
->arcs_lsize
[ARC_BUFC_DATA
] +
2540 arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] +
2541 arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
] +
2542 arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
];
2543 uint64_t ghost_clean
=
2544 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
2545 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_METADATA
] +
2546 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
2547 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_METADATA
];
2548 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
2550 if (arc_dirty
>= arc_c_min
)
2551 return (ghost_clean
+ arc_clean
);
2553 return (ghost_clean
+ MAX((int64_t)arc_size
- (int64_t)arc_c_min
, 0));
2557 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
2561 /* The arc is considered warm once reclaim has occurred */
2562 if (unlikely(arc_warm
== B_FALSE
))
2565 /* Return the potential number of reclaimable pages */
2566 pages
= btop(arc_evictable_memory());
2567 if (sc
->nr_to_scan
== 0)
2570 /* Not allowed to perform filesystem reclaim */
2571 if (!(sc
->gfp_mask
& __GFP_FS
))
2574 /* Reclaim in progress */
2575 if (mutex_tryenter(&arc_reclaim_thr_lock
) == 0)
2579 * Evict the requested number of pages by shrinking arc_c the
2580 * requested amount. If there is nothing left to evict just
2581 * reap whatever we can from the various arc slabs.
2584 arc_kmem_reap_now(ARC_RECLAIM_AGGR
, ptob(sc
->nr_to_scan
));
2586 arc_kmem_reap_now(ARC_RECLAIM_CONS
, ptob(sc
->nr_to_scan
));
2590 * When direct reclaim is observed it usually indicates a rapid
2591 * increase in memory pressure. This occurs because the kswapd
2592 * threads were unable to asynchronously keep enough free memory
2593 * available. In this case set arc_no_grow to briefly pause arc
2594 * growth to avoid compounding the memory pressure.
2596 if (current_is_kswapd()) {
2597 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
2599 arc_no_grow
= B_TRUE
;
2600 arc_grow_time
= ddi_get_lbolt() + (zfs_arc_grow_retry
* hz
);
2601 ARCSTAT_BUMP(arcstat_memory_direct_count
);
2604 mutex_exit(&arc_reclaim_thr_lock
);
2608 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
2610 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
2611 #endif /* _KERNEL */
2614 * Adapt arc info given the number of bytes we are trying to add and
2615 * the state that we are comming from. This function is only called
2616 * when we are adding new content to the cache.
2619 arc_adapt(int bytes
, arc_state_t
*state
)
2622 uint64_t arc_p_min
= (arc_c
>> zfs_arc_p_min_shift
);
2624 if (state
== arc_l2c_only
)
2629 * Adapt the target size of the MRU list:
2630 * - if we just hit in the MRU ghost list, then increase
2631 * the target size of the MRU list.
2632 * - if we just hit in the MFU ghost list, then increase
2633 * the target size of the MFU list by decreasing the
2634 * target size of the MRU list.
2636 if (state
== arc_mru_ghost
) {
2637 mult
= ((arc_mru_ghost
->arcs_size
>= arc_mfu_ghost
->arcs_size
) ?
2638 1 : (arc_mfu_ghost
->arcs_size
/arc_mru_ghost
->arcs_size
));
2639 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
2641 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
2642 } else if (state
== arc_mfu_ghost
) {
2645 mult
= ((arc_mfu_ghost
->arcs_size
>= arc_mru_ghost
->arcs_size
) ?
2646 1 : (arc_mru_ghost
->arcs_size
/arc_mfu_ghost
->arcs_size
));
2647 mult
= MIN(mult
, 10);
2649 delta
= MIN(bytes
* mult
, arc_p
);
2650 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
2652 ASSERT((int64_t)arc_p
>= 0);
2657 if (arc_c
>= arc_c_max
)
2661 * If we're within (2 * maxblocksize) bytes of the target
2662 * cache size, increment the target cache size
2664 if (arc_size
> arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
2665 atomic_add_64(&arc_c
, (int64_t)bytes
);
2666 if (arc_c
> arc_c_max
)
2668 else if (state
== arc_anon
)
2669 atomic_add_64(&arc_p
, (int64_t)bytes
);
2673 ASSERT((int64_t)arc_p
>= 0);
2677 * Check if the cache has reached its limits and eviction is required
2681 arc_evict_needed(arc_buf_contents_t type
)
2683 if (type
== ARC_BUFC_METADATA
&& arc_meta_used
>= arc_meta_limit
)
2689 return (arc_size
> arc_c
);
2693 * The buffer, supplied as the first argument, needs a data block.
2694 * So, if we are at cache max, determine which cache should be victimized.
2695 * We have the following cases:
2697 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2698 * In this situation if we're out of space, but the resident size of the MFU is
2699 * under the limit, victimize the MFU cache to satisfy this insertion request.
2701 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2702 * Here, we've used up all of the available space for the MRU, so we need to
2703 * evict from our own cache instead. Evict from the set of resident MRU
2706 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2707 * c minus p represents the MFU space in the cache, since p is the size of the
2708 * cache that is dedicated to the MRU. In this situation there's still space on
2709 * the MFU side, so the MRU side needs to be victimized.
2711 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2712 * MFU's resident set is consuming more space than it has been allotted. In
2713 * this situation, we must victimize our own cache, the MFU, for this insertion.
2716 arc_get_data_buf(arc_buf_t
*buf
)
2718 arc_state_t
*state
= buf
->b_hdr
->b_state
;
2719 uint64_t size
= buf
->b_hdr
->b_size
;
2720 arc_buf_contents_t type
= buf
->b_hdr
->b_type
;
2722 arc_adapt(size
, state
);
2725 * We have not yet reached cache maximum size,
2726 * just allocate a new buffer.
2728 if (!arc_evict_needed(type
)) {
2729 if (type
== ARC_BUFC_METADATA
) {
2730 buf
->b_data
= zio_buf_alloc(size
);
2731 arc_space_consume(size
, ARC_SPACE_DATA
);
2733 ASSERT(type
== ARC_BUFC_DATA
);
2734 buf
->b_data
= zio_data_buf_alloc(size
);
2735 ARCSTAT_INCR(arcstat_data_size
, size
);
2736 atomic_add_64(&arc_size
, size
);
2742 * If we are prefetching from the mfu ghost list, this buffer
2743 * will end up on the mru list; so steal space from there.
2745 if (state
== arc_mfu_ghost
)
2746 state
= buf
->b_hdr
->b_flags
& ARC_PREFETCH
? arc_mru
: arc_mfu
;
2747 else if (state
== arc_mru_ghost
)
2750 if (state
== arc_mru
|| state
== arc_anon
) {
2751 uint64_t mru_used
= arc_anon
->arcs_size
+ arc_mru
->arcs_size
;
2752 state
= (arc_mfu
->arcs_lsize
[type
] >= size
&&
2753 arc_p
> mru_used
) ? arc_mfu
: arc_mru
;
2756 uint64_t mfu_space
= arc_c
- arc_p
;
2757 state
= (arc_mru
->arcs_lsize
[type
] >= size
&&
2758 mfu_space
> arc_mfu
->arcs_size
) ? arc_mru
: arc_mfu
;
2761 if ((buf
->b_data
= arc_evict(state
, 0, size
, TRUE
, type
)) == NULL
) {
2762 if (type
== ARC_BUFC_METADATA
) {
2763 buf
->b_data
= zio_buf_alloc(size
);
2764 arc_space_consume(size
, ARC_SPACE_DATA
);
2767 * If we are unable to recycle an existing meta buffer
2768 * signal the reclaim thread. It will notify users
2769 * via the prune callback to drop references. The
2770 * prune callback in run in the context of the reclaim
2771 * thread to avoid deadlocking on the hash_lock.
2773 cv_signal(&arc_reclaim_thr_cv
);
2775 ASSERT(type
== ARC_BUFC_DATA
);
2776 buf
->b_data
= zio_data_buf_alloc(size
);
2777 ARCSTAT_INCR(arcstat_data_size
, size
);
2778 atomic_add_64(&arc_size
, size
);
2781 ARCSTAT_BUMP(arcstat_recycle_miss
);
2783 ASSERT(buf
->b_data
!= NULL
);
2786 * Update the state size. Note that ghost states have a
2787 * "ghost size" and so don't need to be updated.
2789 if (!GHOST_STATE(buf
->b_hdr
->b_state
)) {
2790 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2792 atomic_add_64(&hdr
->b_state
->arcs_size
, size
);
2793 if (list_link_active(&hdr
->b_arc_node
)) {
2794 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
2795 atomic_add_64(&hdr
->b_state
->arcs_lsize
[type
], size
);
2798 * If we are growing the cache, and we are adding anonymous
2799 * data, and we have outgrown arc_p, update arc_p
2801 if (arc_size
< arc_c
&& hdr
->b_state
== arc_anon
&&
2802 arc_anon
->arcs_size
+ arc_mru
->arcs_size
> arc_p
)
2803 arc_p
= MIN(arc_c
, arc_p
+ size
);
2808 * This routine is called whenever a buffer is accessed.
2809 * NOTE: the hash lock is dropped in this function.
2812 arc_access(arc_buf_hdr_t
*buf
, kmutex_t
*hash_lock
)
2816 ASSERT(MUTEX_HELD(hash_lock
));
2818 if (buf
->b_state
== arc_anon
) {
2820 * This buffer is not in the cache, and does not
2821 * appear in our "ghost" list. Add the new buffer
2825 ASSERT(buf
->b_arc_access
== 0);
2826 buf
->b_arc_access
= ddi_get_lbolt();
2827 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, buf
);
2828 arc_change_state(arc_mru
, buf
, hash_lock
);
2830 } else if (buf
->b_state
== arc_mru
) {
2831 now
= ddi_get_lbolt();
2834 * If this buffer is here because of a prefetch, then either:
2835 * - clear the flag if this is a "referencing" read
2836 * (any subsequent access will bump this into the MFU state).
2838 * - move the buffer to the head of the list if this is
2839 * another prefetch (to make it less likely to be evicted).
2841 if ((buf
->b_flags
& ARC_PREFETCH
) != 0) {
2842 if (refcount_count(&buf
->b_refcnt
) == 0) {
2843 ASSERT(list_link_active(&buf
->b_arc_node
));
2845 buf
->b_flags
&= ~ARC_PREFETCH
;
2846 atomic_inc_32(&buf
->b_mru_hits
);
2847 ARCSTAT_BUMP(arcstat_mru_hits
);
2849 buf
->b_arc_access
= now
;
2854 * This buffer has been "accessed" only once so far,
2855 * but it is still in the cache. Move it to the MFU
2858 if (now
> buf
->b_arc_access
+ ARC_MINTIME
) {
2860 * More than 125ms have passed since we
2861 * instantiated this buffer. Move it to the
2862 * most frequently used state.
2864 buf
->b_arc_access
= now
;
2865 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2866 arc_change_state(arc_mfu
, buf
, hash_lock
);
2868 atomic_inc_32(&buf
->b_mru_hits
);
2869 ARCSTAT_BUMP(arcstat_mru_hits
);
2870 } else if (buf
->b_state
== arc_mru_ghost
) {
2871 arc_state_t
*new_state
;
2873 * This buffer has been "accessed" recently, but
2874 * was evicted from the cache. Move it to the
2878 if (buf
->b_flags
& ARC_PREFETCH
) {
2879 new_state
= arc_mru
;
2880 if (refcount_count(&buf
->b_refcnt
) > 0)
2881 buf
->b_flags
&= ~ARC_PREFETCH
;
2882 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, buf
);
2884 new_state
= arc_mfu
;
2885 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2888 buf
->b_arc_access
= ddi_get_lbolt();
2889 arc_change_state(new_state
, buf
, hash_lock
);
2891 atomic_inc_32(&buf
->b_mru_ghost_hits
);
2892 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
2893 } else if (buf
->b_state
== arc_mfu
) {
2895 * This buffer has been accessed more than once and is
2896 * still in the cache. Keep it in the MFU state.
2898 * NOTE: an add_reference() that occurred when we did
2899 * the arc_read() will have kicked this off the list.
2900 * If it was a prefetch, we will explicitly move it to
2901 * the head of the list now.
2903 if ((buf
->b_flags
& ARC_PREFETCH
) != 0) {
2904 ASSERT(refcount_count(&buf
->b_refcnt
) == 0);
2905 ASSERT(list_link_active(&buf
->b_arc_node
));
2907 atomic_inc_32(&buf
->b_mfu_hits
);
2908 ARCSTAT_BUMP(arcstat_mfu_hits
);
2909 buf
->b_arc_access
= ddi_get_lbolt();
2910 } else if (buf
->b_state
== arc_mfu_ghost
) {
2911 arc_state_t
*new_state
= arc_mfu
;
2913 * This buffer has been accessed more than once but has
2914 * been evicted from the cache. Move it back to the
2918 if (buf
->b_flags
& ARC_PREFETCH
) {
2920 * This is a prefetch access...
2921 * move this block back to the MRU state.
2923 ASSERT0(refcount_count(&buf
->b_refcnt
));
2924 new_state
= arc_mru
;
2927 buf
->b_arc_access
= ddi_get_lbolt();
2928 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2929 arc_change_state(new_state
, buf
, hash_lock
);
2931 atomic_inc_32(&buf
->b_mfu_ghost_hits
);
2932 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
2933 } else if (buf
->b_state
== arc_l2c_only
) {
2935 * This buffer is on the 2nd Level ARC.
2938 buf
->b_arc_access
= ddi_get_lbolt();
2939 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2940 arc_change_state(arc_mfu
, buf
, hash_lock
);
2942 ASSERT(!"invalid arc state");
2946 /* a generic arc_done_func_t which you can use */
2949 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
2951 if (zio
== NULL
|| zio
->io_error
== 0)
2952 bcopy(buf
->b_data
, arg
, buf
->b_hdr
->b_size
);
2953 VERIFY(arc_buf_remove_ref(buf
, arg
));
2956 /* a generic arc_done_func_t */
2958 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
2960 arc_buf_t
**bufp
= arg
;
2961 if (zio
&& zio
->io_error
) {
2962 VERIFY(arc_buf_remove_ref(buf
, arg
));
2966 ASSERT(buf
->b_data
);
2971 arc_read_done(zio_t
*zio
)
2973 arc_buf_hdr_t
*hdr
, *found
;
2975 arc_buf_t
*abuf
; /* buffer we're assigning to callback */
2976 kmutex_t
*hash_lock
;
2977 arc_callback_t
*callback_list
, *acb
;
2978 int freeable
= FALSE
;
2980 buf
= zio
->io_private
;
2984 * The hdr was inserted into hash-table and removed from lists
2985 * prior to starting I/O. We should find this header, since
2986 * it's in the hash table, and it should be legit since it's
2987 * not possible to evict it during the I/O. The only possible
2988 * reason for it not to be found is if we were freed during the
2991 found
= buf_hash_find(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
,
2994 ASSERT((found
== NULL
&& HDR_FREED_IN_READ(hdr
) && hash_lock
== NULL
) ||
2995 (found
== hdr
&& DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
2996 (found
== hdr
&& HDR_L2_READING(hdr
)));
2998 hdr
->b_flags
&= ~ARC_L2_EVICTED
;
2999 if (l2arc_noprefetch
&& (hdr
->b_flags
& ARC_PREFETCH
))
3000 hdr
->b_flags
&= ~ARC_L2CACHE
;
3002 /* byteswap if necessary */
3003 callback_list
= hdr
->b_acb
;
3004 ASSERT(callback_list
!= NULL
);
3005 if (BP_SHOULD_BYTESWAP(zio
->io_bp
) && zio
->io_error
== 0) {
3006 dmu_object_byteswap_t bswap
=
3007 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
3008 if (BP_GET_LEVEL(zio
->io_bp
) > 0)
3009 byteswap_uint64_array(buf
->b_data
, hdr
->b_size
);
3011 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, hdr
->b_size
);
3014 arc_cksum_compute(buf
, B_FALSE
);
3017 if (hash_lock
&& zio
->io_error
== 0 && hdr
->b_state
== arc_anon
) {
3019 * Only call arc_access on anonymous buffers. This is because
3020 * if we've issued an I/O for an evicted buffer, we've already
3021 * called arc_access (to prevent any simultaneous readers from
3022 * getting confused).
3024 arc_access(hdr
, hash_lock
);
3027 /* create copies of the data buffer for the callers */
3029 for (acb
= callback_list
; acb
; acb
= acb
->acb_next
) {
3030 if (acb
->acb_done
) {
3032 ARCSTAT_BUMP(arcstat_duplicate_reads
);
3033 abuf
= arc_buf_clone(buf
);
3035 acb
->acb_buf
= abuf
;
3040 hdr
->b_flags
&= ~ARC_IO_IN_PROGRESS
;
3041 ASSERT(!HDR_BUF_AVAILABLE(hdr
));
3043 ASSERT(buf
->b_efunc
== NULL
);
3044 ASSERT(hdr
->b_datacnt
== 1);
3045 hdr
->b_flags
|= ARC_BUF_AVAILABLE
;
3048 ASSERT(refcount_is_zero(&hdr
->b_refcnt
) || callback_list
!= NULL
);
3050 if (zio
->io_error
!= 0) {
3051 hdr
->b_flags
|= ARC_IO_ERROR
;
3052 if (hdr
->b_state
!= arc_anon
)
3053 arc_change_state(arc_anon
, hdr
, hash_lock
);
3054 if (HDR_IN_HASH_TABLE(hdr
))
3055 buf_hash_remove(hdr
);
3056 freeable
= refcount_is_zero(&hdr
->b_refcnt
);
3060 * Broadcast before we drop the hash_lock to avoid the possibility
3061 * that the hdr (and hence the cv) might be freed before we get to
3062 * the cv_broadcast().
3064 cv_broadcast(&hdr
->b_cv
);
3067 mutex_exit(hash_lock
);
3070 * This block was freed while we waited for the read to
3071 * complete. It has been removed from the hash table and
3072 * moved to the anonymous state (so that it won't show up
3075 ASSERT3P(hdr
->b_state
, ==, arc_anon
);
3076 freeable
= refcount_is_zero(&hdr
->b_refcnt
);
3079 /* execute each callback and free its structure */
3080 while ((acb
= callback_list
) != NULL
) {
3082 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
3084 if (acb
->acb_zio_dummy
!= NULL
) {
3085 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
3086 zio_nowait(acb
->acb_zio_dummy
);
3089 callback_list
= acb
->acb_next
;
3090 kmem_free(acb
, sizeof (arc_callback_t
));
3094 arc_hdr_destroy(hdr
);
3098 * "Read" the block at the specified DVA (in bp) via the
3099 * cache. If the block is found in the cache, invoke the provided
3100 * callback immediately and return. Note that the `zio' parameter
3101 * in the callback will be NULL in this case, since no IO was
3102 * required. If the block is not in the cache pass the read request
3103 * on to the spa with a substitute callback function, so that the
3104 * requested block will be added to the cache.
3106 * If a read request arrives for a block that has a read in-progress,
3107 * either wait for the in-progress read to complete (and return the
3108 * results); or, if this is a read with a "done" func, add a record
3109 * to the read to invoke the "done" func when the read completes,
3110 * and return; or just return.
3112 * arc_read_done() will invoke all the requested "done" functions
3113 * for readers of this block.
3116 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_done_func_t
*done
,
3117 void *private, zio_priority_t priority
, int zio_flags
, uint32_t *arc_flags
,
3118 const zbookmark_t
*zb
)
3121 arc_buf_t
*buf
= NULL
;
3122 kmutex_t
*hash_lock
;
3124 uint64_t guid
= spa_load_guid(spa
);
3128 hdr
= buf_hash_find(guid
, BP_IDENTITY(bp
), BP_PHYSICAL_BIRTH(bp
),
3130 if (hdr
&& hdr
->b_datacnt
> 0) {
3132 *arc_flags
|= ARC_CACHED
;
3134 if (HDR_IO_IN_PROGRESS(hdr
)) {
3136 if (*arc_flags
& ARC_WAIT
) {
3137 cv_wait(&hdr
->b_cv
, hash_lock
);
3138 mutex_exit(hash_lock
);
3141 ASSERT(*arc_flags
& ARC_NOWAIT
);
3144 arc_callback_t
*acb
= NULL
;
3146 acb
= kmem_zalloc(sizeof (arc_callback_t
),
3148 acb
->acb_done
= done
;
3149 acb
->acb_private
= private;
3151 acb
->acb_zio_dummy
= zio_null(pio
,
3152 spa
, NULL
, NULL
, NULL
, zio_flags
);
3154 ASSERT(acb
->acb_done
!= NULL
);
3155 acb
->acb_next
= hdr
->b_acb
;
3157 add_reference(hdr
, hash_lock
, private);
3158 mutex_exit(hash_lock
);
3161 mutex_exit(hash_lock
);
3165 ASSERT(hdr
->b_state
== arc_mru
|| hdr
->b_state
== arc_mfu
);
3168 add_reference(hdr
, hash_lock
, private);
3170 * If this block is already in use, create a new
3171 * copy of the data so that we will be guaranteed
3172 * that arc_release() will always succeed.
3176 ASSERT(buf
->b_data
);
3177 if (HDR_BUF_AVAILABLE(hdr
)) {
3178 ASSERT(buf
->b_efunc
== NULL
);
3179 hdr
->b_flags
&= ~ARC_BUF_AVAILABLE
;
3181 buf
= arc_buf_clone(buf
);
3184 } else if (*arc_flags
& ARC_PREFETCH
&&
3185 refcount_count(&hdr
->b_refcnt
) == 0) {
3186 hdr
->b_flags
|= ARC_PREFETCH
;
3188 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
3189 arc_access(hdr
, hash_lock
);
3190 if (*arc_flags
& ARC_L2CACHE
)
3191 hdr
->b_flags
|= ARC_L2CACHE
;
3192 if (*arc_flags
& ARC_L2COMPRESS
)
3193 hdr
->b_flags
|= ARC_L2COMPRESS
;
3194 mutex_exit(hash_lock
);
3195 ARCSTAT_BUMP(arcstat_hits
);
3196 ARCSTAT_CONDSTAT(!(hdr
->b_flags
& ARC_PREFETCH
),
3197 demand
, prefetch
, hdr
->b_type
!= ARC_BUFC_METADATA
,
3198 data
, metadata
, hits
);
3201 done(NULL
, buf
, private);
3203 uint64_t size
= BP_GET_LSIZE(bp
);
3204 arc_callback_t
*acb
;
3207 boolean_t devw
= B_FALSE
;
3210 /* this block is not in the cache */
3211 arc_buf_hdr_t
*exists
;
3212 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
3213 buf
= arc_buf_alloc(spa
, size
, private, type
);
3215 hdr
->b_dva
= *BP_IDENTITY(bp
);
3216 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
3217 hdr
->b_cksum0
= bp
->blk_cksum
.zc_word
[0];
3218 exists
= buf_hash_insert(hdr
, &hash_lock
);
3220 /* somebody beat us to the hash insert */
3221 mutex_exit(hash_lock
);
3222 buf_discard_identity(hdr
);
3223 (void) arc_buf_remove_ref(buf
, private);
3224 goto top
; /* restart the IO request */
3226 /* if this is a prefetch, we don't have a reference */
3227 if (*arc_flags
& ARC_PREFETCH
) {
3228 (void) remove_reference(hdr
, hash_lock
,
3230 hdr
->b_flags
|= ARC_PREFETCH
;
3232 if (*arc_flags
& ARC_L2CACHE
)
3233 hdr
->b_flags
|= ARC_L2CACHE
;
3234 if (*arc_flags
& ARC_L2COMPRESS
)
3235 hdr
->b_flags
|= ARC_L2COMPRESS
;
3236 if (BP_GET_LEVEL(bp
) > 0)
3237 hdr
->b_flags
|= ARC_INDIRECT
;
3239 /* this block is in the ghost cache */
3240 ASSERT(GHOST_STATE(hdr
->b_state
));
3241 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3242 ASSERT0(refcount_count(&hdr
->b_refcnt
));
3243 ASSERT(hdr
->b_buf
== NULL
);
3245 /* if this is a prefetch, we don't have a reference */
3246 if (*arc_flags
& ARC_PREFETCH
)
3247 hdr
->b_flags
|= ARC_PREFETCH
;
3249 add_reference(hdr
, hash_lock
, private);
3250 if (*arc_flags
& ARC_L2CACHE
)
3251 hdr
->b_flags
|= ARC_L2CACHE
;
3252 if (*arc_flags
& ARC_L2COMPRESS
)
3253 hdr
->b_flags
|= ARC_L2COMPRESS
;
3254 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
3257 buf
->b_efunc
= NULL
;
3258 buf
->b_private
= NULL
;
3261 ASSERT(hdr
->b_datacnt
== 0);
3263 arc_get_data_buf(buf
);
3264 arc_access(hdr
, hash_lock
);
3267 ASSERT(!GHOST_STATE(hdr
->b_state
));
3269 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_PUSHPAGE
);
3270 acb
->acb_done
= done
;
3271 acb
->acb_private
= private;
3273 ASSERT(hdr
->b_acb
== NULL
);
3275 hdr
->b_flags
|= ARC_IO_IN_PROGRESS
;
3277 if (HDR_L2CACHE(hdr
) && hdr
->b_l2hdr
!= NULL
&&
3278 (vd
= hdr
->b_l2hdr
->b_dev
->l2ad_vdev
) != NULL
) {
3279 devw
= hdr
->b_l2hdr
->b_dev
->l2ad_writing
;
3280 addr
= hdr
->b_l2hdr
->b_daddr
;
3282 * Lock out device removal.
3284 if (vdev_is_dead(vd
) ||
3285 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
3289 mutex_exit(hash_lock
);
3292 * At this point, we have a level 1 cache miss. Try again in
3293 * L2ARC if possible.
3295 ASSERT3U(hdr
->b_size
, ==, size
);
3296 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
3297 uint64_t, size
, zbookmark_t
*, zb
);
3298 ARCSTAT_BUMP(arcstat_misses
);
3299 ARCSTAT_CONDSTAT(!(hdr
->b_flags
& ARC_PREFETCH
),
3300 demand
, prefetch
, hdr
->b_type
!= ARC_BUFC_METADATA
,
3301 data
, metadata
, misses
);
3303 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
3305 * Read from the L2ARC if the following are true:
3306 * 1. The L2ARC vdev was previously cached.
3307 * 2. This buffer still has L2ARC metadata.
3308 * 3. This buffer isn't currently writing to the L2ARC.
3309 * 4. The L2ARC entry wasn't evicted, which may
3310 * also have invalidated the vdev.
3311 * 5. This isn't prefetch and l2arc_noprefetch is set.
3313 if (hdr
->b_l2hdr
!= NULL
&&
3314 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
3315 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
3316 l2arc_read_callback_t
*cb
;
3318 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
3319 ARCSTAT_BUMP(arcstat_l2_hits
);
3320 atomic_inc_32(&hdr
->b_l2hdr
->b_hits
);
3322 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
3324 cb
->l2rcb_buf
= buf
;
3325 cb
->l2rcb_spa
= spa
;
3328 cb
->l2rcb_flags
= zio_flags
;
3329 cb
->l2rcb_compress
= hdr
->b_l2hdr
->b_compress
;
3331 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
3332 addr
+ size
< vd
->vdev_psize
-
3333 VDEV_LABEL_END_SIZE
);
3336 * l2arc read. The SCL_L2ARC lock will be
3337 * released by l2arc_read_done().
3338 * Issue a null zio if the underlying buffer
3339 * was squashed to zero size by compression.
3341 if (hdr
->b_l2hdr
->b_compress
==
3342 ZIO_COMPRESS_EMPTY
) {
3343 rzio
= zio_null(pio
, spa
, vd
,
3344 l2arc_read_done
, cb
,
3345 zio_flags
| ZIO_FLAG_DONT_CACHE
|
3347 ZIO_FLAG_DONT_PROPAGATE
|
3348 ZIO_FLAG_DONT_RETRY
);
3350 rzio
= zio_read_phys(pio
, vd
, addr
,
3351 hdr
->b_l2hdr
->b_asize
,
3352 buf
->b_data
, ZIO_CHECKSUM_OFF
,
3353 l2arc_read_done
, cb
, priority
,
3354 zio_flags
| ZIO_FLAG_DONT_CACHE
|
3356 ZIO_FLAG_DONT_PROPAGATE
|
3357 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
3359 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
3361 ARCSTAT_INCR(arcstat_l2_read_bytes
,
3362 hdr
->b_l2hdr
->b_asize
);
3364 if (*arc_flags
& ARC_NOWAIT
) {
3369 ASSERT(*arc_flags
& ARC_WAIT
);
3370 if (zio_wait(rzio
) == 0)
3373 /* l2arc read error; goto zio_read() */
3375 DTRACE_PROBE1(l2arc__miss
,
3376 arc_buf_hdr_t
*, hdr
);
3377 ARCSTAT_BUMP(arcstat_l2_misses
);
3378 if (HDR_L2_WRITING(hdr
))
3379 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
3380 spa_config_exit(spa
, SCL_L2ARC
, vd
);
3384 spa_config_exit(spa
, SCL_L2ARC
, vd
);
3385 if (l2arc_ndev
!= 0) {
3386 DTRACE_PROBE1(l2arc__miss
,
3387 arc_buf_hdr_t
*, hdr
);
3388 ARCSTAT_BUMP(arcstat_l2_misses
);
3392 rzio
= zio_read(pio
, spa
, bp
, buf
->b_data
, size
,
3393 arc_read_done
, buf
, priority
, zio_flags
, zb
);
3395 if (*arc_flags
& ARC_WAIT
) {
3396 rc
= zio_wait(rzio
);
3400 ASSERT(*arc_flags
& ARC_NOWAIT
);
3405 spa_read_history_add(spa
, zb
, *arc_flags
);
3410 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
3414 p
= kmem_alloc(sizeof (*p
), KM_SLEEP
);
3416 p
->p_private
= private;
3417 list_link_init(&p
->p_node
);
3418 refcount_create(&p
->p_refcnt
);
3420 mutex_enter(&arc_prune_mtx
);
3421 refcount_add(&p
->p_refcnt
, &arc_prune_list
);
3422 list_insert_head(&arc_prune_list
, p
);
3423 mutex_exit(&arc_prune_mtx
);
3429 arc_remove_prune_callback(arc_prune_t
*p
)
3431 mutex_enter(&arc_prune_mtx
);
3432 list_remove(&arc_prune_list
, p
);
3433 if (refcount_remove(&p
->p_refcnt
, &arc_prune_list
) == 0) {
3434 refcount_destroy(&p
->p_refcnt
);
3435 kmem_free(p
, sizeof (*p
));
3437 mutex_exit(&arc_prune_mtx
);
3441 arc_set_callback(arc_buf_t
*buf
, arc_evict_func_t
*func
, void *private)
3443 ASSERT(buf
->b_hdr
!= NULL
);
3444 ASSERT(buf
->b_hdr
->b_state
!= arc_anon
);
3445 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_refcnt
) || func
== NULL
);
3446 ASSERT(buf
->b_efunc
== NULL
);
3447 ASSERT(!HDR_BUF_AVAILABLE(buf
->b_hdr
));
3449 buf
->b_efunc
= func
;
3450 buf
->b_private
= private;
3454 * Notify the arc that a block was freed, and thus will never be used again.
3457 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
3460 kmutex_t
*hash_lock
;
3461 uint64_t guid
= spa_load_guid(spa
);
3463 hdr
= buf_hash_find(guid
, BP_IDENTITY(bp
), BP_PHYSICAL_BIRTH(bp
),
3467 if (HDR_BUF_AVAILABLE(hdr
)) {
3468 arc_buf_t
*buf
= hdr
->b_buf
;
3469 add_reference(hdr
, hash_lock
, FTAG
);
3470 hdr
->b_flags
&= ~ARC_BUF_AVAILABLE
;
3471 mutex_exit(hash_lock
);
3473 arc_release(buf
, FTAG
);
3474 (void) arc_buf_remove_ref(buf
, FTAG
);
3476 mutex_exit(hash_lock
);
3482 * This is used by the DMU to let the ARC know that a buffer is
3483 * being evicted, so the ARC should clean up. If this arc buf
3484 * is not yet in the evicted state, it will be put there.
3487 arc_buf_evict(arc_buf_t
*buf
)
3490 kmutex_t
*hash_lock
;
3493 mutex_enter(&buf
->b_evict_lock
);
3497 * We are in arc_do_user_evicts().
3499 ASSERT(buf
->b_data
== NULL
);
3500 mutex_exit(&buf
->b_evict_lock
);
3502 } else if (buf
->b_data
== NULL
) {
3503 arc_buf_t copy
= *buf
; /* structure assignment */
3505 * We are on the eviction list; process this buffer now
3506 * but let arc_do_user_evicts() do the reaping.
3508 buf
->b_efunc
= NULL
;
3509 mutex_exit(&buf
->b_evict_lock
);
3510 VERIFY(copy
.b_efunc(©
) == 0);
3513 hash_lock
= HDR_LOCK(hdr
);
3514 mutex_enter(hash_lock
);
3516 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3518 ASSERT3U(refcount_count(&hdr
->b_refcnt
), <, hdr
->b_datacnt
);
3519 ASSERT(hdr
->b_state
== arc_mru
|| hdr
->b_state
== arc_mfu
);
3522 * Pull this buffer off of the hdr
3525 while (*bufp
!= buf
)
3526 bufp
= &(*bufp
)->b_next
;
3527 *bufp
= buf
->b_next
;
3529 ASSERT(buf
->b_data
!= NULL
);
3530 arc_buf_destroy(buf
, FALSE
, FALSE
);
3532 if (hdr
->b_datacnt
== 0) {
3533 arc_state_t
*old_state
= hdr
->b_state
;
3534 arc_state_t
*evicted_state
;
3536 ASSERT(hdr
->b_buf
== NULL
);
3537 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
3540 (old_state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
3542 mutex_enter(&old_state
->arcs_mtx
);
3543 mutex_enter(&evicted_state
->arcs_mtx
);
3545 arc_change_state(evicted_state
, hdr
, hash_lock
);
3546 ASSERT(HDR_IN_HASH_TABLE(hdr
));
3547 hdr
->b_flags
|= ARC_IN_HASH_TABLE
;
3548 hdr
->b_flags
&= ~ARC_BUF_AVAILABLE
;
3550 mutex_exit(&evicted_state
->arcs_mtx
);
3551 mutex_exit(&old_state
->arcs_mtx
);
3553 mutex_exit(hash_lock
);
3554 mutex_exit(&buf
->b_evict_lock
);
3556 VERIFY(buf
->b_efunc(buf
) == 0);
3557 buf
->b_efunc
= NULL
;
3558 buf
->b_private
= NULL
;
3561 kmem_cache_free(buf_cache
, buf
);
3566 * Release this buffer from the cache, making it an anonymous buffer. This
3567 * must be done after a read and prior to modifying the buffer contents.
3568 * If the buffer has more than one reference, we must make
3569 * a new hdr for the buffer.
3572 arc_release(arc_buf_t
*buf
, void *tag
)
3575 kmutex_t
*hash_lock
= NULL
;
3576 l2arc_buf_hdr_t
*l2hdr
;
3577 uint64_t buf_size
= 0;
3580 * It would be nice to assert that if it's DMU metadata (level >
3581 * 0 || it's the dnode file), then it must be syncing context.
3582 * But we don't know that information at this level.
3585 mutex_enter(&buf
->b_evict_lock
);
3588 /* this buffer is not on any list */
3589 ASSERT(refcount_count(&hdr
->b_refcnt
) > 0);
3591 if (hdr
->b_state
== arc_anon
) {
3592 /* this buffer is already released */
3593 ASSERT(buf
->b_efunc
== NULL
);
3595 hash_lock
= HDR_LOCK(hdr
);
3596 mutex_enter(hash_lock
);
3598 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3601 l2hdr
= hdr
->b_l2hdr
;
3603 mutex_enter(&l2arc_buflist_mtx
);
3604 hdr
->b_l2hdr
= NULL
;
3606 buf_size
= hdr
->b_size
;
3609 * Do we have more than one buf?
3611 if (hdr
->b_datacnt
> 1) {
3612 arc_buf_hdr_t
*nhdr
;
3614 uint64_t blksz
= hdr
->b_size
;
3615 uint64_t spa
= hdr
->b_spa
;
3616 arc_buf_contents_t type
= hdr
->b_type
;
3617 uint32_t flags
= hdr
->b_flags
;
3619 ASSERT(hdr
->b_buf
!= buf
|| buf
->b_next
!= NULL
);
3621 * Pull the data off of this hdr and attach it to
3622 * a new anonymous hdr.
3624 (void) remove_reference(hdr
, hash_lock
, tag
);
3626 while (*bufp
!= buf
)
3627 bufp
= &(*bufp
)->b_next
;
3628 *bufp
= buf
->b_next
;
3631 ASSERT3U(hdr
->b_state
->arcs_size
, >=, hdr
->b_size
);
3632 atomic_add_64(&hdr
->b_state
->arcs_size
, -hdr
->b_size
);
3633 if (refcount_is_zero(&hdr
->b_refcnt
)) {
3634 uint64_t *size
= &hdr
->b_state
->arcs_lsize
[hdr
->b_type
];
3635 ASSERT3U(*size
, >=, hdr
->b_size
);
3636 atomic_add_64(size
, -hdr
->b_size
);
3640 * We're releasing a duplicate user data buffer, update
3641 * our statistics accordingly.
3643 if (hdr
->b_type
== ARC_BUFC_DATA
) {
3644 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
3645 ARCSTAT_INCR(arcstat_duplicate_buffers_size
,
3648 hdr
->b_datacnt
-= 1;
3649 arc_cksum_verify(buf
);
3650 arc_buf_unwatch(buf
);
3652 mutex_exit(hash_lock
);
3654 nhdr
= kmem_cache_alloc(hdr_cache
, KM_PUSHPAGE
);
3655 nhdr
->b_size
= blksz
;
3657 nhdr
->b_type
= type
;
3659 nhdr
->b_state
= arc_anon
;
3660 nhdr
->b_arc_access
= 0;
3661 nhdr
->b_mru_hits
= 0;
3662 nhdr
->b_mru_ghost_hits
= 0;
3663 nhdr
->b_mfu_hits
= 0;
3664 nhdr
->b_mfu_ghost_hits
= 0;
3665 nhdr
->b_l2_hits
= 0;
3666 nhdr
->b_flags
= flags
& ARC_L2_WRITING
;
3667 nhdr
->b_l2hdr
= NULL
;
3668 nhdr
->b_datacnt
= 1;
3669 nhdr
->b_freeze_cksum
= NULL
;
3670 (void) refcount_add(&nhdr
->b_refcnt
, tag
);
3672 mutex_exit(&buf
->b_evict_lock
);
3673 atomic_add_64(&arc_anon
->arcs_size
, blksz
);
3675 mutex_exit(&buf
->b_evict_lock
);
3676 ASSERT(refcount_count(&hdr
->b_refcnt
) == 1);
3677 ASSERT(!list_link_active(&hdr
->b_arc_node
));
3678 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3679 if (hdr
->b_state
!= arc_anon
)
3680 arc_change_state(arc_anon
, hdr
, hash_lock
);
3681 hdr
->b_arc_access
= 0;
3682 hdr
->b_mru_hits
= 0;
3683 hdr
->b_mru_ghost_hits
= 0;
3684 hdr
->b_mfu_hits
= 0;
3685 hdr
->b_mfu_ghost_hits
= 0;
3688 mutex_exit(hash_lock
);
3690 buf_discard_identity(hdr
);
3693 buf
->b_efunc
= NULL
;
3694 buf
->b_private
= NULL
;
3697 ARCSTAT_INCR(arcstat_l2_asize
, -l2hdr
->b_asize
);
3698 list_remove(l2hdr
->b_dev
->l2ad_buflist
, hdr
);
3699 kmem_cache_free(l2arc_hdr_cache
, l2hdr
);
3700 arc_space_return(L2HDR_SIZE
, ARC_SPACE_L2HDRS
);
3701 ARCSTAT_INCR(arcstat_l2_size
, -buf_size
);
3702 mutex_exit(&l2arc_buflist_mtx
);
3707 arc_released(arc_buf_t
*buf
)
3711 mutex_enter(&buf
->b_evict_lock
);
3712 released
= (buf
->b_data
!= NULL
&& buf
->b_hdr
->b_state
== arc_anon
);
3713 mutex_exit(&buf
->b_evict_lock
);
3718 arc_has_callback(arc_buf_t
*buf
)
3722 mutex_enter(&buf
->b_evict_lock
);
3723 callback
= (buf
->b_efunc
!= NULL
);
3724 mutex_exit(&buf
->b_evict_lock
);
3730 arc_referenced(arc_buf_t
*buf
)
3734 mutex_enter(&buf
->b_evict_lock
);
3735 referenced
= (refcount_count(&buf
->b_hdr
->b_refcnt
));
3736 mutex_exit(&buf
->b_evict_lock
);
3737 return (referenced
);
3742 arc_write_ready(zio_t
*zio
)
3744 arc_write_callback_t
*callback
= zio
->io_private
;
3745 arc_buf_t
*buf
= callback
->awcb_buf
;
3746 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3748 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_refcnt
));
3749 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
3752 * If the IO is already in progress, then this is a re-write
3753 * attempt, so we need to thaw and re-compute the cksum.
3754 * It is the responsibility of the callback to handle the
3755 * accounting for any re-write attempt.
3757 if (HDR_IO_IN_PROGRESS(hdr
)) {
3758 mutex_enter(&hdr
->b_freeze_lock
);
3759 if (hdr
->b_freeze_cksum
!= NULL
) {
3760 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
3761 hdr
->b_freeze_cksum
= NULL
;
3763 mutex_exit(&hdr
->b_freeze_lock
);
3765 arc_cksum_compute(buf
, B_FALSE
);
3766 hdr
->b_flags
|= ARC_IO_IN_PROGRESS
;
3770 * The SPA calls this callback for each physical write that happens on behalf
3771 * of a logical write. See the comment in dbuf_write_physdone() for details.
3774 arc_write_physdone(zio_t
*zio
)
3776 arc_write_callback_t
*cb
= zio
->io_private
;
3777 if (cb
->awcb_physdone
!= NULL
)
3778 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
3782 arc_write_done(zio_t
*zio
)
3784 arc_write_callback_t
*callback
= zio
->io_private
;
3785 arc_buf_t
*buf
= callback
->awcb_buf
;
3786 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3788 ASSERT(hdr
->b_acb
== NULL
);
3790 if (zio
->io_error
== 0) {
3791 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
3792 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
3793 hdr
->b_cksum0
= zio
->io_bp
->blk_cksum
.zc_word
[0];
3795 ASSERT(BUF_EMPTY(hdr
));
3799 * If the block to be written was all-zero, we may have
3800 * compressed it away. In this case no write was performed
3801 * so there will be no dva/birth/checksum. The buffer must
3802 * therefore remain anonymous (and uncached).
3804 if (!BUF_EMPTY(hdr
)) {
3805 arc_buf_hdr_t
*exists
;
3806 kmutex_t
*hash_lock
;
3808 ASSERT(zio
->io_error
== 0);
3810 arc_cksum_verify(buf
);
3812 exists
= buf_hash_insert(hdr
, &hash_lock
);
3815 * This can only happen if we overwrite for
3816 * sync-to-convergence, because we remove
3817 * buffers from the hash table when we arc_free().
3819 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
3820 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
3821 panic("bad overwrite, hdr=%p exists=%p",
3822 (void *)hdr
, (void *)exists
);
3823 ASSERT(refcount_is_zero(&exists
->b_refcnt
));
3824 arc_change_state(arc_anon
, exists
, hash_lock
);
3825 mutex_exit(hash_lock
);
3826 arc_hdr_destroy(exists
);
3827 exists
= buf_hash_insert(hdr
, &hash_lock
);
3828 ASSERT3P(exists
, ==, NULL
);
3829 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
3831 ASSERT(zio
->io_prop
.zp_nopwrite
);
3832 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
3833 panic("bad nopwrite, hdr=%p exists=%p",
3834 (void *)hdr
, (void *)exists
);
3837 ASSERT(hdr
->b_datacnt
== 1);
3838 ASSERT(hdr
->b_state
== arc_anon
);
3839 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
3840 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
3843 hdr
->b_flags
&= ~ARC_IO_IN_PROGRESS
;
3844 /* if it's not anon, we are doing a scrub */
3845 if (!exists
&& hdr
->b_state
== arc_anon
)
3846 arc_access(hdr
, hash_lock
);
3847 mutex_exit(hash_lock
);
3849 hdr
->b_flags
&= ~ARC_IO_IN_PROGRESS
;
3852 ASSERT(!refcount_is_zero(&hdr
->b_refcnt
));
3853 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
3855 kmem_free(callback
, sizeof (arc_write_callback_t
));
3859 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
3860 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
, boolean_t l2arc_compress
,
3861 const zio_prop_t
*zp
, arc_done_func_t
*ready
, arc_done_func_t
*physdone
,
3862 arc_done_func_t
*done
, void *private, zio_priority_t priority
,
3863 int zio_flags
, const zbookmark_t
*zb
)
3865 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3866 arc_write_callback_t
*callback
;
3869 ASSERT(ready
!= NULL
);
3870 ASSERT(done
!= NULL
);
3871 ASSERT(!HDR_IO_ERROR(hdr
));
3872 ASSERT((hdr
->b_flags
& ARC_IO_IN_PROGRESS
) == 0);
3873 ASSERT(hdr
->b_acb
== NULL
);
3875 hdr
->b_flags
|= ARC_L2CACHE
;
3877 hdr
->b_flags
|= ARC_L2COMPRESS
;
3878 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_PUSHPAGE
);
3879 callback
->awcb_ready
= ready
;
3880 callback
->awcb_physdone
= physdone
;
3881 callback
->awcb_done
= done
;
3882 callback
->awcb_private
= private;
3883 callback
->awcb_buf
= buf
;
3885 zio
= zio_write(pio
, spa
, txg
, bp
, buf
->b_data
, hdr
->b_size
, zp
,
3886 arc_write_ready
, arc_write_physdone
, arc_write_done
, callback
,
3887 priority
, zio_flags
, zb
);
3893 arc_memory_throttle(uint64_t reserve
, uint64_t txg
)
3896 if (zfs_arc_memory_throttle_disable
)
3899 if (freemem
<= physmem
* arc_lotsfree_percent
/ 100) {
3900 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
3901 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
3902 return (SET_ERROR(EAGAIN
));
3909 arc_tempreserve_clear(uint64_t reserve
)
3911 atomic_add_64(&arc_tempreserve
, -reserve
);
3912 ASSERT((int64_t)arc_tempreserve
>= 0);
3916 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
3921 if (reserve
> arc_c
/4 && !arc_no_grow
)
3922 arc_c
= MIN(arc_c_max
, reserve
* 4);
3923 if (reserve
> arc_c
) {
3924 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
3925 return (SET_ERROR(ENOMEM
));
3929 * Don't count loaned bufs as in flight dirty data to prevent long
3930 * network delays from blocking transactions that are ready to be
3931 * assigned to a txg.
3933 anon_size
= MAX((int64_t)(arc_anon
->arcs_size
- arc_loaned_bytes
), 0);
3936 * Writes will, almost always, require additional memory allocations
3937 * in order to compress/encrypt/etc the data. We therefore need to
3938 * make sure that there is sufficient available memory for this.
3940 error
= arc_memory_throttle(reserve
, txg
);
3945 * Throttle writes when the amount of dirty data in the cache
3946 * gets too large. We try to keep the cache less than half full
3947 * of dirty blocks so that our sync times don't grow too large.
3948 * Note: if two requests come in concurrently, we might let them
3949 * both succeed, when one of them should fail. Not a huge deal.
3952 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
3953 anon_size
> arc_c
/ 4) {
3954 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3955 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3956 arc_tempreserve
>>10,
3957 arc_anon
->arcs_lsize
[ARC_BUFC_METADATA
]>>10,
3958 arc_anon
->arcs_lsize
[ARC_BUFC_DATA
]>>10,
3959 reserve
>>10, arc_c
>>10);
3960 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
3961 return (SET_ERROR(ERESTART
));
3963 atomic_add_64(&arc_tempreserve
, reserve
);
3968 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
3969 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
3971 size
->value
.ui64
= state
->arcs_size
;
3972 evict_data
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_DATA
];
3973 evict_metadata
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_METADATA
];
3977 arc_kstat_update(kstat_t
*ksp
, int rw
)
3979 arc_stats_t
*as
= ksp
->ks_data
;
3981 if (rw
== KSTAT_WRITE
) {
3982 return (SET_ERROR(EACCES
));
3984 arc_kstat_update_state(arc_anon
,
3985 &as
->arcstat_anon_size
,
3986 &as
->arcstat_anon_evict_data
,
3987 &as
->arcstat_anon_evict_metadata
);
3988 arc_kstat_update_state(arc_mru
,
3989 &as
->arcstat_mru_size
,
3990 &as
->arcstat_mru_evict_data
,
3991 &as
->arcstat_mru_evict_metadata
);
3992 arc_kstat_update_state(arc_mru_ghost
,
3993 &as
->arcstat_mru_ghost_size
,
3994 &as
->arcstat_mru_ghost_evict_data
,
3995 &as
->arcstat_mru_ghost_evict_metadata
);
3996 arc_kstat_update_state(arc_mfu
,
3997 &as
->arcstat_mfu_size
,
3998 &as
->arcstat_mfu_evict_data
,
3999 &as
->arcstat_mfu_evict_metadata
);
4000 arc_kstat_update_state(arc_mfu_ghost
,
4001 &as
->arcstat_mfu_ghost_size
,
4002 &as
->arcstat_mfu_ghost_evict_data
,
4003 &as
->arcstat_mfu_ghost_evict_metadata
);
4012 mutex_init(&arc_reclaim_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
4013 cv_init(&arc_reclaim_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
4015 /* Convert seconds to clock ticks */
4016 zfs_arc_min_prefetch_lifespan
= 1 * hz
;
4018 /* Start out with 1/8 of all memory */
4019 arc_c
= physmem
* PAGESIZE
/ 8;
4023 * On architectures where the physical memory can be larger
4024 * than the addressable space (intel in 32-bit mode), we may
4025 * need to limit the cache to 1/8 of VM size.
4027 arc_c
= MIN(arc_c
, vmem_size(heap_arena
, VMEM_ALLOC
| VMEM_FREE
) / 8);
4029 * Register a shrinker to support synchronous (direct) memory
4030 * reclaim from the arc. This is done to prevent kswapd from
4031 * swapping out pages when it is preferable to shrink the arc.
4033 spl_register_shrinker(&arc_shrinker
);
4036 /* set min cache to zero */
4038 /* set max to 1/2 of all memory */
4039 arc_c_max
= arc_c
* 4;
4042 * Allow the tunables to override our calculations if they are
4043 * reasonable (ie. over 64MB)
4045 if (zfs_arc_max
> 64<<20 && zfs_arc_max
< physmem
* PAGESIZE
)
4046 arc_c_max
= zfs_arc_max
;
4047 if (zfs_arc_min
> 0 && zfs_arc_min
<= arc_c_max
)
4048 arc_c_min
= zfs_arc_min
;
4051 arc_p
= (arc_c
>> 1);
4053 /* limit meta-data to 1/4 of the arc capacity */
4054 arc_meta_limit
= arc_c_max
/ 4;
4057 /* Allow the tunable to override if it is reasonable */
4058 if (zfs_arc_meta_limit
> 0 && zfs_arc_meta_limit
<= arc_c_max
)
4059 arc_meta_limit
= zfs_arc_meta_limit
;
4061 /* if kmem_flags are set, lets try to use less memory */
4062 if (kmem_debugging())
4064 if (arc_c
< arc_c_min
)
4067 arc_anon
= &ARC_anon
;
4069 arc_mru_ghost
= &ARC_mru_ghost
;
4071 arc_mfu_ghost
= &ARC_mfu_ghost
;
4072 arc_l2c_only
= &ARC_l2c_only
;
4075 mutex_init(&arc_anon
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4076 mutex_init(&arc_mru
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4077 mutex_init(&arc_mru_ghost
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4078 mutex_init(&arc_mfu
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4079 mutex_init(&arc_mfu_ghost
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4080 mutex_init(&arc_l2c_only
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4082 list_create(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
],
4083 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
4084 list_create(&arc_mru
->arcs_list
[ARC_BUFC_DATA
],
4085 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
4086 list_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
],
4087 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
4088 list_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
],
4089 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
4090 list_create(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
],
4091 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
4092 list_create(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
],
4093 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
4094 list_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
],
4095 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
4096 list_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
],
4097 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
4098 list_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
],
4099 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
4100 list_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
],
4101 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
4103 arc_anon
->arcs_state
= ARC_STATE_ANON
;
4104 arc_mru
->arcs_state
= ARC_STATE_MRU
;
4105 arc_mru_ghost
->arcs_state
= ARC_STATE_MRU_GHOST
;
4106 arc_mfu
->arcs_state
= ARC_STATE_MFU
;
4107 arc_mfu_ghost
->arcs_state
= ARC_STATE_MFU_GHOST
;
4108 arc_l2c_only
->arcs_state
= ARC_STATE_L2C_ONLY
;
4112 arc_thread_exit
= 0;
4113 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
4114 offsetof(arc_prune_t
, p_node
));
4115 arc_eviction_list
= NULL
;
4116 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4117 mutex_init(&arc_eviction_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4118 bzero(&arc_eviction_hdr
, sizeof (arc_buf_hdr_t
));
4120 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
4121 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
4123 if (arc_ksp
!= NULL
) {
4124 arc_ksp
->ks_data
= &arc_stats
;
4125 arc_ksp
->ks_update
= arc_kstat_update
;
4126 kstat_install(arc_ksp
);
4129 (void) thread_create(NULL
, 0, arc_adapt_thread
, NULL
, 0, &p0
,
4130 TS_RUN
, minclsyspri
);
4136 * Calculate maximum amount of dirty data per pool.
4138 * If it has been set by a module parameter, take that.
4139 * Otherwise, use a percentage of physical memory defined by
4140 * zfs_dirty_data_max_percent (default 10%) with a cap at
4141 * zfs_dirty_data_max_max (default 25% of physical memory).
4143 if (zfs_dirty_data_max_max
== 0)
4144 zfs_dirty_data_max_max
= physmem
* PAGESIZE
*
4145 zfs_dirty_data_max_max_percent
/ 100;
4147 if (zfs_dirty_data_max
== 0) {
4148 zfs_dirty_data_max
= physmem
* PAGESIZE
*
4149 zfs_dirty_data_max_percent
/ 100;
4150 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
4151 zfs_dirty_data_max_max
);
4160 mutex_enter(&arc_reclaim_thr_lock
);
4162 spl_unregister_shrinker(&arc_shrinker
);
4163 #endif /* _KERNEL */
4165 arc_thread_exit
= 1;
4166 while (arc_thread_exit
!= 0)
4167 cv_wait(&arc_reclaim_thr_cv
, &arc_reclaim_thr_lock
);
4168 mutex_exit(&arc_reclaim_thr_lock
);
4174 if (arc_ksp
!= NULL
) {
4175 kstat_delete(arc_ksp
);
4179 mutex_enter(&arc_prune_mtx
);
4180 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
4181 list_remove(&arc_prune_list
, p
);
4182 refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
4183 refcount_destroy(&p
->p_refcnt
);
4184 kmem_free(p
, sizeof (*p
));
4186 mutex_exit(&arc_prune_mtx
);
4188 list_destroy(&arc_prune_list
);
4189 mutex_destroy(&arc_prune_mtx
);
4190 mutex_destroy(&arc_eviction_mtx
);
4191 mutex_destroy(&arc_reclaim_thr_lock
);
4192 cv_destroy(&arc_reclaim_thr_cv
);
4194 list_destroy(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
4195 list_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
4196 list_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
4197 list_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
4198 list_destroy(&arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
4199 list_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
4200 list_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
4201 list_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
4203 mutex_destroy(&arc_anon
->arcs_mtx
);
4204 mutex_destroy(&arc_mru
->arcs_mtx
);
4205 mutex_destroy(&arc_mru_ghost
->arcs_mtx
);
4206 mutex_destroy(&arc_mfu
->arcs_mtx
);
4207 mutex_destroy(&arc_mfu_ghost
->arcs_mtx
);
4208 mutex_destroy(&arc_l2c_only
->arcs_mtx
);
4212 ASSERT(arc_loaned_bytes
== 0);
4218 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
4219 * It uses dedicated storage devices to hold cached data, which are populated
4220 * using large infrequent writes. The main role of this cache is to boost
4221 * the performance of random read workloads. The intended L2ARC devices
4222 * include short-stroked disks, solid state disks, and other media with
4223 * substantially faster read latency than disk.
4225 * +-----------------------+
4227 * +-----------------------+
4230 * l2arc_feed_thread() arc_read()
4234 * +---------------+ |
4236 * +---------------+ |
4241 * +-------+ +-------+
4243 * | cache | | cache |
4244 * +-------+ +-------+
4245 * +=========+ .-----.
4246 * : L2ARC : |-_____-|
4247 * : devices : | Disks |
4248 * +=========+ `-_____-'
4250 * Read requests are satisfied from the following sources, in order:
4253 * 2) vdev cache of L2ARC devices
4255 * 4) vdev cache of disks
4258 * Some L2ARC device types exhibit extremely slow write performance.
4259 * To accommodate for this there are some significant differences between
4260 * the L2ARC and traditional cache design:
4262 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4263 * the ARC behave as usual, freeing buffers and placing headers on ghost
4264 * lists. The ARC does not send buffers to the L2ARC during eviction as
4265 * this would add inflated write latencies for all ARC memory pressure.
4267 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4268 * It does this by periodically scanning buffers from the eviction-end of
4269 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4270 * not already there. It scans until a headroom of buffers is satisfied,
4271 * which itself is a buffer for ARC eviction. If a compressible buffer is
4272 * found during scanning and selected for writing to an L2ARC device, we
4273 * temporarily boost scanning headroom during the next scan cycle to make
4274 * sure we adapt to compression effects (which might significantly reduce
4275 * the data volume we write to L2ARC). The thread that does this is
4276 * l2arc_feed_thread(), illustrated below; example sizes are included to
4277 * provide a better sense of ratio than this diagram:
4280 * +---------------------+----------+
4281 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4282 * +---------------------+----------+ | o L2ARC eligible
4283 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4284 * +---------------------+----------+ |
4285 * 15.9 Gbytes ^ 32 Mbytes |
4287 * l2arc_feed_thread()
4289 * l2arc write hand <--[oooo]--'
4293 * +==============================+
4294 * L2ARC dev |####|#|###|###| |####| ... |
4295 * +==============================+
4298 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4299 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4300 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4301 * safe to say that this is an uncommon case, since buffers at the end of
4302 * the ARC lists have moved there due to inactivity.
4304 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4305 * then the L2ARC simply misses copying some buffers. This serves as a
4306 * pressure valve to prevent heavy read workloads from both stalling the ARC
4307 * with waits and clogging the L2ARC with writes. This also helps prevent
4308 * the potential for the L2ARC to churn if it attempts to cache content too
4309 * quickly, such as during backups of the entire pool.
4311 * 5. After system boot and before the ARC has filled main memory, there are
4312 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4313 * lists can remain mostly static. Instead of searching from tail of these
4314 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4315 * for eligible buffers, greatly increasing its chance of finding them.
4317 * The L2ARC device write speed is also boosted during this time so that
4318 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4319 * there are no L2ARC reads, and no fear of degrading read performance
4320 * through increased writes.
4322 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4323 * the vdev queue can aggregate them into larger and fewer writes. Each
4324 * device is written to in a rotor fashion, sweeping writes through
4325 * available space then repeating.
4327 * 7. The L2ARC does not store dirty content. It never needs to flush
4328 * write buffers back to disk based storage.
4330 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4331 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4333 * The performance of the L2ARC can be tweaked by a number of tunables, which
4334 * may be necessary for different workloads:
4336 * l2arc_write_max max write bytes per interval
4337 * l2arc_write_boost extra write bytes during device warmup
4338 * l2arc_noprefetch skip caching prefetched buffers
4339 * l2arc_nocompress skip compressing buffers
4340 * l2arc_headroom number of max device writes to precache
4341 * l2arc_headroom_boost when we find compressed buffers during ARC
4342 * scanning, we multiply headroom by this
4343 * percentage factor for the next scan cycle,
4344 * since more compressed buffers are likely to
4346 * l2arc_feed_secs seconds between L2ARC writing
4348 * Tunables may be removed or added as future performance improvements are
4349 * integrated, and also may become zpool properties.
4351 * There are three key functions that control how the L2ARC warms up:
4353 * l2arc_write_eligible() check if a buffer is eligible to cache
4354 * l2arc_write_size() calculate how much to write
4355 * l2arc_write_interval() calculate sleep delay between writes
4357 * These three functions determine what to write, how much, and how quickly
4362 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*ab
)
4365 * A buffer is *not* eligible for the L2ARC if it:
4366 * 1. belongs to a different spa.
4367 * 2. is already cached on the L2ARC.
4368 * 3. has an I/O in progress (it may be an incomplete read).
4369 * 4. is flagged not eligible (zfs property).
4371 if (ab
->b_spa
!= spa_guid
|| ab
->b_l2hdr
!= NULL
||
4372 HDR_IO_IN_PROGRESS(ab
) || !HDR_L2CACHE(ab
))
4379 l2arc_write_size(void)
4384 * Make sure our globals have meaningful values in case the user
4387 size
= l2arc_write_max
;
4389 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
4390 "be greater than zero, resetting it to the default (%d)",
4392 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
4395 if (arc_warm
== B_FALSE
)
4396 size
+= l2arc_write_boost
;
4403 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
4405 clock_t interval
, next
, now
;
4408 * If the ARC lists are busy, increase our write rate; if the
4409 * lists are stale, idle back. This is achieved by checking
4410 * how much we previously wrote - if it was more than half of
4411 * what we wanted, schedule the next write much sooner.
4413 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
4414 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
4416 interval
= hz
* l2arc_feed_secs
;
4418 now
= ddi_get_lbolt();
4419 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
4425 l2arc_hdr_stat_add(void)
4427 ARCSTAT_INCR(arcstat_l2_hdr_size
, HDR_SIZE
);
4428 ARCSTAT_INCR(arcstat_hdr_size
, -HDR_SIZE
);
4432 l2arc_hdr_stat_remove(void)
4434 ARCSTAT_INCR(arcstat_l2_hdr_size
, -HDR_SIZE
);
4435 ARCSTAT_INCR(arcstat_hdr_size
, HDR_SIZE
);
4439 * Cycle through L2ARC devices. This is how L2ARC load balances.
4440 * If a device is returned, this also returns holding the spa config lock.
4442 static l2arc_dev_t
*
4443 l2arc_dev_get_next(void)
4445 l2arc_dev_t
*first
, *next
= NULL
;
4448 * Lock out the removal of spas (spa_namespace_lock), then removal
4449 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4450 * both locks will be dropped and a spa config lock held instead.
4452 mutex_enter(&spa_namespace_lock
);
4453 mutex_enter(&l2arc_dev_mtx
);
4455 /* if there are no vdevs, there is nothing to do */
4456 if (l2arc_ndev
== 0)
4460 next
= l2arc_dev_last
;
4462 /* loop around the list looking for a non-faulted vdev */
4464 next
= list_head(l2arc_dev_list
);
4466 next
= list_next(l2arc_dev_list
, next
);
4468 next
= list_head(l2arc_dev_list
);
4471 /* if we have come back to the start, bail out */
4474 else if (next
== first
)
4477 } while (vdev_is_dead(next
->l2ad_vdev
));
4479 /* if we were unable to find any usable vdevs, return NULL */
4480 if (vdev_is_dead(next
->l2ad_vdev
))
4483 l2arc_dev_last
= next
;
4486 mutex_exit(&l2arc_dev_mtx
);
4489 * Grab the config lock to prevent the 'next' device from being
4490 * removed while we are writing to it.
4493 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
4494 mutex_exit(&spa_namespace_lock
);
4500 * Free buffers that were tagged for destruction.
4503 l2arc_do_free_on_write(void)
4506 l2arc_data_free_t
*df
, *df_prev
;
4508 mutex_enter(&l2arc_free_on_write_mtx
);
4509 buflist
= l2arc_free_on_write
;
4511 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
4512 df_prev
= list_prev(buflist
, df
);
4513 ASSERT(df
->l2df_data
!= NULL
);
4514 ASSERT(df
->l2df_func
!= NULL
);
4515 df
->l2df_func(df
->l2df_data
, df
->l2df_size
);
4516 list_remove(buflist
, df
);
4517 kmem_free(df
, sizeof (l2arc_data_free_t
));
4520 mutex_exit(&l2arc_free_on_write_mtx
);
4524 * A write to a cache device has completed. Update all headers to allow
4525 * reads from these buffers to begin.
4528 l2arc_write_done(zio_t
*zio
)
4530 l2arc_write_callback_t
*cb
;
4533 arc_buf_hdr_t
*head
, *ab
, *ab_prev
;
4534 l2arc_buf_hdr_t
*abl2
;
4535 kmutex_t
*hash_lock
;
4537 cb
= zio
->io_private
;
4539 dev
= cb
->l2wcb_dev
;
4540 ASSERT(dev
!= NULL
);
4541 head
= cb
->l2wcb_head
;
4542 ASSERT(head
!= NULL
);
4543 buflist
= dev
->l2ad_buflist
;
4544 ASSERT(buflist
!= NULL
);
4545 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
4546 l2arc_write_callback_t
*, cb
);
4548 if (zio
->io_error
!= 0)
4549 ARCSTAT_BUMP(arcstat_l2_writes_error
);
4551 mutex_enter(&l2arc_buflist_mtx
);
4554 * All writes completed, or an error was hit.
4556 for (ab
= list_prev(buflist
, head
); ab
; ab
= ab_prev
) {
4557 ab_prev
= list_prev(buflist
, ab
);
4561 * Release the temporary compressed buffer as soon as possible.
4563 if (abl2
->b_compress
!= ZIO_COMPRESS_OFF
)
4564 l2arc_release_cdata_buf(ab
);
4566 hash_lock
= HDR_LOCK(ab
);
4567 if (!mutex_tryenter(hash_lock
)) {
4569 * This buffer misses out. It may be in a stage
4570 * of eviction. Its ARC_L2_WRITING flag will be
4571 * left set, denying reads to this buffer.
4573 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss
);
4577 if (zio
->io_error
!= 0) {
4579 * Error - drop L2ARC entry.
4581 list_remove(buflist
, ab
);
4582 ARCSTAT_INCR(arcstat_l2_asize
, -abl2
->b_asize
);
4584 kmem_cache_free(l2arc_hdr_cache
, abl2
);
4585 arc_space_return(L2HDR_SIZE
, ARC_SPACE_L2HDRS
);
4586 ARCSTAT_INCR(arcstat_l2_size
, -ab
->b_size
);
4590 * Allow ARC to begin reads to this L2ARC entry.
4592 ab
->b_flags
&= ~ARC_L2_WRITING
;
4594 mutex_exit(hash_lock
);
4597 atomic_inc_64(&l2arc_writes_done
);
4598 list_remove(buflist
, head
);
4599 kmem_cache_free(hdr_cache
, head
);
4600 mutex_exit(&l2arc_buflist_mtx
);
4602 l2arc_do_free_on_write();
4604 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
4608 * A read to a cache device completed. Validate buffer contents before
4609 * handing over to the regular ARC routines.
4612 l2arc_read_done(zio_t
*zio
)
4614 l2arc_read_callback_t
*cb
;
4617 kmutex_t
*hash_lock
;
4620 ASSERT(zio
->io_vd
!= NULL
);
4621 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
4623 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
4625 cb
= zio
->io_private
;
4627 buf
= cb
->l2rcb_buf
;
4628 ASSERT(buf
!= NULL
);
4630 hash_lock
= HDR_LOCK(buf
->b_hdr
);
4631 mutex_enter(hash_lock
);
4633 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
4636 * If the buffer was compressed, decompress it first.
4638 if (cb
->l2rcb_compress
!= ZIO_COMPRESS_OFF
)
4639 l2arc_decompress_zio(zio
, hdr
, cb
->l2rcb_compress
);
4640 ASSERT(zio
->io_data
!= NULL
);
4643 * Check this survived the L2ARC journey.
4645 equal
= arc_cksum_equal(buf
);
4646 if (equal
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
4647 mutex_exit(hash_lock
);
4648 zio
->io_private
= buf
;
4649 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
4650 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
4653 mutex_exit(hash_lock
);
4655 * Buffer didn't survive caching. Increment stats and
4656 * reissue to the original storage device.
4658 if (zio
->io_error
!= 0) {
4659 ARCSTAT_BUMP(arcstat_l2_io_error
);
4661 zio
->io_error
= SET_ERROR(EIO
);
4664 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
4667 * If there's no waiter, issue an async i/o to the primary
4668 * storage now. If there *is* a waiter, the caller must
4669 * issue the i/o in a context where it's OK to block.
4671 if (zio
->io_waiter
== NULL
) {
4672 zio_t
*pio
= zio_unique_parent(zio
);
4674 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4676 zio_nowait(zio_read(pio
, cb
->l2rcb_spa
, &cb
->l2rcb_bp
,
4677 buf
->b_data
, zio
->io_size
, arc_read_done
, buf
,
4678 zio
->io_priority
, cb
->l2rcb_flags
, &cb
->l2rcb_zb
));
4682 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
4686 * This is the list priority from which the L2ARC will search for pages to
4687 * cache. This is used within loops (0..3) to cycle through lists in the
4688 * desired order. This order can have a significant effect on cache
4691 * Currently the metadata lists are hit first, MFU then MRU, followed by
4692 * the data lists. This function returns a locked list, and also returns
4696 l2arc_list_locked(int list_num
, kmutex_t
**lock
)
4698 list_t
*list
= NULL
;
4700 ASSERT(list_num
>= 0 && list_num
<= 3);
4704 list
= &arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
4705 *lock
= &arc_mfu
->arcs_mtx
;
4708 list
= &arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
4709 *lock
= &arc_mru
->arcs_mtx
;
4712 list
= &arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
4713 *lock
= &arc_mfu
->arcs_mtx
;
4716 list
= &arc_mru
->arcs_list
[ARC_BUFC_DATA
];
4717 *lock
= &arc_mru
->arcs_mtx
;
4721 ASSERT(!(MUTEX_HELD(*lock
)));
4727 * Evict buffers from the device write hand to the distance specified in
4728 * bytes. This distance may span populated buffers, it may span nothing.
4729 * This is clearing a region on the L2ARC device ready for writing.
4730 * If the 'all' boolean is set, every buffer is evicted.
4733 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
4736 l2arc_buf_hdr_t
*abl2
;
4737 arc_buf_hdr_t
*ab
, *ab_prev
;
4738 kmutex_t
*hash_lock
;
4741 buflist
= dev
->l2ad_buflist
;
4743 if (buflist
== NULL
)
4746 if (!all
&& dev
->l2ad_first
) {
4748 * This is the first sweep through the device. There is
4754 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
4756 * When nearing the end of the device, evict to the end
4757 * before the device write hand jumps to the start.
4759 taddr
= dev
->l2ad_end
;
4761 taddr
= dev
->l2ad_hand
+ distance
;
4763 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
4764 uint64_t, taddr
, boolean_t
, all
);
4767 mutex_enter(&l2arc_buflist_mtx
);
4768 for (ab
= list_tail(buflist
); ab
; ab
= ab_prev
) {
4769 ab_prev
= list_prev(buflist
, ab
);
4771 hash_lock
= HDR_LOCK(ab
);
4772 if (!mutex_tryenter(hash_lock
)) {
4774 * Missed the hash lock. Retry.
4776 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
4777 mutex_exit(&l2arc_buflist_mtx
);
4778 mutex_enter(hash_lock
);
4779 mutex_exit(hash_lock
);
4783 if (HDR_L2_WRITE_HEAD(ab
)) {
4785 * We hit a write head node. Leave it for
4786 * l2arc_write_done().
4788 list_remove(buflist
, ab
);
4789 mutex_exit(hash_lock
);
4793 if (!all
&& ab
->b_l2hdr
!= NULL
&&
4794 (ab
->b_l2hdr
->b_daddr
> taddr
||
4795 ab
->b_l2hdr
->b_daddr
< dev
->l2ad_hand
)) {
4797 * We've evicted to the target address,
4798 * or the end of the device.
4800 mutex_exit(hash_lock
);
4804 if (HDR_FREE_IN_PROGRESS(ab
)) {
4806 * Already on the path to destruction.
4808 mutex_exit(hash_lock
);
4812 if (ab
->b_state
== arc_l2c_only
) {
4813 ASSERT(!HDR_L2_READING(ab
));
4815 * This doesn't exist in the ARC. Destroy.
4816 * arc_hdr_destroy() will call list_remove()
4817 * and decrement arcstat_l2_size.
4819 arc_change_state(arc_anon
, ab
, hash_lock
);
4820 arc_hdr_destroy(ab
);
4823 * Invalidate issued or about to be issued
4824 * reads, since we may be about to write
4825 * over this location.
4827 if (HDR_L2_READING(ab
)) {
4828 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
4829 ab
->b_flags
|= ARC_L2_EVICTED
;
4833 * Tell ARC this no longer exists in L2ARC.
4835 if (ab
->b_l2hdr
!= NULL
) {
4837 ARCSTAT_INCR(arcstat_l2_asize
, -abl2
->b_asize
);
4839 kmem_cache_free(l2arc_hdr_cache
, abl2
);
4840 arc_space_return(L2HDR_SIZE
, ARC_SPACE_L2HDRS
);
4841 ARCSTAT_INCR(arcstat_l2_size
, -ab
->b_size
);
4843 list_remove(buflist
, ab
);
4846 * This may have been leftover after a
4849 ab
->b_flags
&= ~ARC_L2_WRITING
;
4851 mutex_exit(hash_lock
);
4853 mutex_exit(&l2arc_buflist_mtx
);
4855 vdev_space_update(dev
->l2ad_vdev
, -(taddr
- dev
->l2ad_evict
), 0, 0);
4856 dev
->l2ad_evict
= taddr
;
4860 * Find and write ARC buffers to the L2ARC device.
4862 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4863 * for reading until they have completed writing.
4864 * The headroom_boost is an in-out parameter used to maintain headroom boost
4865 * state between calls to this function.
4867 * Returns the number of bytes actually written (which may be smaller than
4868 * the delta by which the device hand has changed due to alignment).
4871 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
,
4872 boolean_t
*headroom_boost
)
4874 arc_buf_hdr_t
*ab
, *ab_prev
, *head
;
4876 uint64_t write_asize
, write_psize
, write_sz
, headroom
,
4879 kmutex_t
*list_lock
= NULL
;
4881 l2arc_write_callback_t
*cb
;
4883 uint64_t guid
= spa_load_guid(spa
);
4885 const boolean_t do_headroom_boost
= *headroom_boost
;
4887 ASSERT(dev
->l2ad_vdev
!= NULL
);
4889 /* Lower the flag now, we might want to raise it again later. */
4890 *headroom_boost
= B_FALSE
;
4893 write_sz
= write_asize
= write_psize
= 0;
4895 head
= kmem_cache_alloc(hdr_cache
, KM_PUSHPAGE
);
4896 head
->b_flags
|= ARC_L2_WRITE_HEAD
;
4899 * We will want to try to compress buffers that are at least 2x the
4900 * device sector size.
4902 buf_compress_minsz
= 2 << dev
->l2ad_vdev
->vdev_ashift
;
4905 * Copy buffers for L2ARC writing.
4907 mutex_enter(&l2arc_buflist_mtx
);
4908 for (try = 0; try <= 3; try++) {
4909 uint64_t passed_sz
= 0;
4911 list
= l2arc_list_locked(try, &list_lock
);
4914 * L2ARC fast warmup.
4916 * Until the ARC is warm and starts to evict, read from the
4917 * head of the ARC lists rather than the tail.
4919 if (arc_warm
== B_FALSE
)
4920 ab
= list_head(list
);
4922 ab
= list_tail(list
);
4924 headroom
= target_sz
* l2arc_headroom
;
4925 if (do_headroom_boost
)
4926 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
4928 for (; ab
; ab
= ab_prev
) {
4929 l2arc_buf_hdr_t
*l2hdr
;
4930 kmutex_t
*hash_lock
;
4933 if (arc_warm
== B_FALSE
)
4934 ab_prev
= list_next(list
, ab
);
4936 ab_prev
= list_prev(list
, ab
);
4938 hash_lock
= HDR_LOCK(ab
);
4939 if (!mutex_tryenter(hash_lock
)) {
4941 * Skip this buffer rather than waiting.
4946 passed_sz
+= ab
->b_size
;
4947 if (passed_sz
> headroom
) {
4951 mutex_exit(hash_lock
);
4955 if (!l2arc_write_eligible(guid
, ab
)) {
4956 mutex_exit(hash_lock
);
4960 if ((write_sz
+ ab
->b_size
) > target_sz
) {
4962 mutex_exit(hash_lock
);
4968 * Insert a dummy header on the buflist so
4969 * l2arc_write_done() can find where the
4970 * write buffers begin without searching.
4972 list_insert_head(dev
->l2ad_buflist
, head
);
4974 cb
= kmem_alloc(sizeof (l2arc_write_callback_t
),
4976 cb
->l2wcb_dev
= dev
;
4977 cb
->l2wcb_head
= head
;
4978 pio
= zio_root(spa
, l2arc_write_done
, cb
,
4983 * Create and add a new L2ARC header.
4985 l2hdr
= kmem_cache_alloc(l2arc_hdr_cache
, KM_PUSHPAGE
);
4988 arc_space_consume(L2HDR_SIZE
, ARC_SPACE_L2HDRS
);
4990 ab
->b_flags
|= ARC_L2_WRITING
;
4993 * Temporarily stash the data buffer in b_tmp_cdata.
4994 * The subsequent write step will pick it up from
4995 * there. This is because can't access ab->b_buf
4996 * without holding the hash_lock, which we in turn
4997 * can't access without holding the ARC list locks
4998 * (which we want to avoid during compression/writing)
5000 l2hdr
->b_compress
= ZIO_COMPRESS_OFF
;
5001 l2hdr
->b_asize
= ab
->b_size
;
5002 l2hdr
->b_tmp_cdata
= ab
->b_buf
->b_data
;
5005 buf_sz
= ab
->b_size
;
5006 ab
->b_l2hdr
= l2hdr
;
5008 list_insert_head(dev
->l2ad_buflist
, ab
);
5011 * Compute and store the buffer cksum before
5012 * writing. On debug the cksum is verified first.
5014 arc_cksum_verify(ab
->b_buf
);
5015 arc_cksum_compute(ab
->b_buf
, B_TRUE
);
5017 mutex_exit(hash_lock
);
5022 mutex_exit(list_lock
);
5028 /* No buffers selected for writing? */
5031 mutex_exit(&l2arc_buflist_mtx
);
5032 kmem_cache_free(hdr_cache
, head
);
5037 * Now start writing the buffers. We're starting at the write head
5038 * and work backwards, retracing the course of the buffer selector
5041 for (ab
= list_prev(dev
->l2ad_buflist
, head
); ab
;
5042 ab
= list_prev(dev
->l2ad_buflist
, ab
)) {
5043 l2arc_buf_hdr_t
*l2hdr
;
5047 * We shouldn't need to lock the buffer here, since we flagged
5048 * it as ARC_L2_WRITING in the previous step, but we must take
5049 * care to only access its L2 cache parameters. In particular,
5050 * ab->b_buf may be invalid by now due to ARC eviction.
5052 l2hdr
= ab
->b_l2hdr
;
5053 l2hdr
->b_daddr
= dev
->l2ad_hand
;
5055 if (!l2arc_nocompress
&& (ab
->b_flags
& ARC_L2COMPRESS
) &&
5056 l2hdr
->b_asize
>= buf_compress_minsz
) {
5057 if (l2arc_compress_buf(l2hdr
)) {
5059 * If compression succeeded, enable headroom
5060 * boost on the next scan cycle.
5062 *headroom_boost
= B_TRUE
;
5067 * Pick up the buffer data we had previously stashed away
5068 * (and now potentially also compressed).
5070 buf_data
= l2hdr
->b_tmp_cdata
;
5071 buf_sz
= l2hdr
->b_asize
;
5073 /* Compression may have squashed the buffer to zero length. */
5077 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
5078 dev
->l2ad_hand
, buf_sz
, buf_data
, ZIO_CHECKSUM_OFF
,
5079 NULL
, NULL
, ZIO_PRIORITY_ASYNC_WRITE
,
5080 ZIO_FLAG_CANFAIL
, B_FALSE
);
5082 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
5084 (void) zio_nowait(wzio
);
5086 write_asize
+= buf_sz
;
5088 * Keep the clock hand suitably device-aligned.
5090 buf_p_sz
= vdev_psize_to_asize(dev
->l2ad_vdev
, buf_sz
);
5091 write_psize
+= buf_p_sz
;
5092 dev
->l2ad_hand
+= buf_p_sz
;
5096 mutex_exit(&l2arc_buflist_mtx
);
5098 ASSERT3U(write_asize
, <=, target_sz
);
5099 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
5100 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_asize
);
5101 ARCSTAT_INCR(arcstat_l2_size
, write_sz
);
5102 ARCSTAT_INCR(arcstat_l2_asize
, write_asize
);
5103 vdev_space_update(dev
->l2ad_vdev
, write_psize
, 0, 0);
5106 * Bump device hand to the device start if it is approaching the end.
5107 * l2arc_evict() will already have evicted ahead for this case.
5109 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
5110 vdev_space_update(dev
->l2ad_vdev
,
5111 dev
->l2ad_end
- dev
->l2ad_hand
, 0, 0);
5112 dev
->l2ad_hand
= dev
->l2ad_start
;
5113 dev
->l2ad_evict
= dev
->l2ad_start
;
5114 dev
->l2ad_first
= B_FALSE
;
5117 dev
->l2ad_writing
= B_TRUE
;
5118 (void) zio_wait(pio
);
5119 dev
->l2ad_writing
= B_FALSE
;
5121 return (write_asize
);
5125 * Compresses an L2ARC buffer.
5126 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its
5127 * size in l2hdr->b_asize. This routine tries to compress the data and
5128 * depending on the compression result there are three possible outcomes:
5129 * *) The buffer was incompressible. The original l2hdr contents were left
5130 * untouched and are ready for writing to an L2 device.
5131 * *) The buffer was all-zeros, so there is no need to write it to an L2
5132 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
5133 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
5134 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
5135 * data buffer which holds the compressed data to be written, and b_asize
5136 * tells us how much data there is. b_compress is set to the appropriate
5137 * compression algorithm. Once writing is done, invoke
5138 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
5140 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
5141 * buffer was incompressible).
5144 l2arc_compress_buf(l2arc_buf_hdr_t
*l2hdr
)
5149 ASSERT(l2hdr
->b_compress
== ZIO_COMPRESS_OFF
);
5150 ASSERT(l2hdr
->b_tmp_cdata
!= NULL
);
5152 len
= l2hdr
->b_asize
;
5153 cdata
= zio_data_buf_alloc(len
);
5154 csize
= zio_compress_data(ZIO_COMPRESS_LZ4
, l2hdr
->b_tmp_cdata
,
5155 cdata
, l2hdr
->b_asize
);
5158 /* zero block, indicate that there's nothing to write */
5159 zio_data_buf_free(cdata
, len
);
5160 l2hdr
->b_compress
= ZIO_COMPRESS_EMPTY
;
5162 l2hdr
->b_tmp_cdata
= NULL
;
5163 ARCSTAT_BUMP(arcstat_l2_compress_zeros
);
5165 } else if (csize
> 0 && csize
< len
) {
5167 * Compression succeeded, we'll keep the cdata around for
5168 * writing and release it afterwards.
5170 l2hdr
->b_compress
= ZIO_COMPRESS_LZ4
;
5171 l2hdr
->b_asize
= csize
;
5172 l2hdr
->b_tmp_cdata
= cdata
;
5173 ARCSTAT_BUMP(arcstat_l2_compress_successes
);
5177 * Compression failed, release the compressed buffer.
5178 * l2hdr will be left unmodified.
5180 zio_data_buf_free(cdata
, len
);
5181 ARCSTAT_BUMP(arcstat_l2_compress_failures
);
5187 * Decompresses a zio read back from an l2arc device. On success, the
5188 * underlying zio's io_data buffer is overwritten by the uncompressed
5189 * version. On decompression error (corrupt compressed stream), the
5190 * zio->io_error value is set to signal an I/O error.
5192 * Please note that the compressed data stream is not checksummed, so
5193 * if the underlying device is experiencing data corruption, we may feed
5194 * corrupt data to the decompressor, so the decompressor needs to be
5195 * able to handle this situation (LZ4 does).
5198 l2arc_decompress_zio(zio_t
*zio
, arc_buf_hdr_t
*hdr
, enum zio_compress c
)
5203 ASSERT(L2ARC_IS_VALID_COMPRESS(c
));
5205 if (zio
->io_error
!= 0) {
5207 * An io error has occured, just restore the original io
5208 * size in preparation for a main pool read.
5210 zio
->io_orig_size
= zio
->io_size
= hdr
->b_size
;
5214 if (c
== ZIO_COMPRESS_EMPTY
) {
5216 * An empty buffer results in a null zio, which means we
5217 * need to fill its io_data after we're done restoring the
5218 * buffer's contents.
5220 ASSERT(hdr
->b_buf
!= NULL
);
5221 bzero(hdr
->b_buf
->b_data
, hdr
->b_size
);
5222 zio
->io_data
= zio
->io_orig_data
= hdr
->b_buf
->b_data
;
5224 ASSERT(zio
->io_data
!= NULL
);
5226 * We copy the compressed data from the start of the arc buffer
5227 * (the zio_read will have pulled in only what we need, the
5228 * rest is garbage which we will overwrite at decompression)
5229 * and then decompress back to the ARC data buffer. This way we
5230 * can minimize copying by simply decompressing back over the
5231 * original compressed data (rather than decompressing to an
5232 * aux buffer and then copying back the uncompressed buffer,
5233 * which is likely to be much larger).
5235 csize
= zio
->io_size
;
5236 cdata
= zio_data_buf_alloc(csize
);
5237 bcopy(zio
->io_data
, cdata
, csize
);
5238 if (zio_decompress_data(c
, cdata
, zio
->io_data
, csize
,
5240 zio
->io_error
= SET_ERROR(EIO
);
5241 zio_data_buf_free(cdata
, csize
);
5244 /* Restore the expected uncompressed IO size. */
5245 zio
->io_orig_size
= zio
->io_size
= hdr
->b_size
;
5249 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
5250 * This buffer serves as a temporary holder of compressed data while
5251 * the buffer entry is being written to an l2arc device. Once that is
5252 * done, we can dispose of it.
5255 l2arc_release_cdata_buf(arc_buf_hdr_t
*ab
)
5257 l2arc_buf_hdr_t
*l2hdr
= ab
->b_l2hdr
;
5259 if (l2hdr
->b_compress
== ZIO_COMPRESS_LZ4
) {
5261 * If the data was compressed, then we've allocated a
5262 * temporary buffer for it, so now we need to release it.
5264 ASSERT(l2hdr
->b_tmp_cdata
!= NULL
);
5265 zio_data_buf_free(l2hdr
->b_tmp_cdata
, ab
->b_size
);
5267 l2hdr
->b_tmp_cdata
= NULL
;
5271 * This thread feeds the L2ARC at regular intervals. This is the beating
5272 * heart of the L2ARC.
5275 l2arc_feed_thread(void)
5280 uint64_t size
, wrote
;
5281 clock_t begin
, next
= ddi_get_lbolt();
5282 boolean_t headroom_boost
= B_FALSE
;
5284 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
5286 mutex_enter(&l2arc_feed_thr_lock
);
5288 while (l2arc_thread_exit
== 0) {
5289 CALLB_CPR_SAFE_BEGIN(&cpr
);
5290 (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv
,
5291 &l2arc_feed_thr_lock
, next
);
5292 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
5293 next
= ddi_get_lbolt() + hz
;
5296 * Quick check for L2ARC devices.
5298 mutex_enter(&l2arc_dev_mtx
);
5299 if (l2arc_ndev
== 0) {
5300 mutex_exit(&l2arc_dev_mtx
);
5303 mutex_exit(&l2arc_dev_mtx
);
5304 begin
= ddi_get_lbolt();
5307 * This selects the next l2arc device to write to, and in
5308 * doing so the next spa to feed from: dev->l2ad_spa. This
5309 * will return NULL if there are now no l2arc devices or if
5310 * they are all faulted.
5312 * If a device is returned, its spa's config lock is also
5313 * held to prevent device removal. l2arc_dev_get_next()
5314 * will grab and release l2arc_dev_mtx.
5316 if ((dev
= l2arc_dev_get_next()) == NULL
)
5319 spa
= dev
->l2ad_spa
;
5320 ASSERT(spa
!= NULL
);
5323 * If the pool is read-only then force the feed thread to
5324 * sleep a little longer.
5326 if (!spa_writeable(spa
)) {
5327 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
5328 spa_config_exit(spa
, SCL_L2ARC
, dev
);
5333 * Avoid contributing to memory pressure.
5336 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
5337 spa_config_exit(spa
, SCL_L2ARC
, dev
);
5341 ARCSTAT_BUMP(arcstat_l2_feeds
);
5343 size
= l2arc_write_size();
5346 * Evict L2ARC buffers that will be overwritten.
5348 l2arc_evict(dev
, size
, B_FALSE
);
5351 * Write ARC buffers.
5353 wrote
= l2arc_write_buffers(spa
, dev
, size
, &headroom_boost
);
5356 * Calculate interval between writes.
5358 next
= l2arc_write_interval(begin
, size
, wrote
);
5359 spa_config_exit(spa
, SCL_L2ARC
, dev
);
5362 l2arc_thread_exit
= 0;
5363 cv_broadcast(&l2arc_feed_thr_cv
);
5364 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
5369 l2arc_vdev_present(vdev_t
*vd
)
5373 mutex_enter(&l2arc_dev_mtx
);
5374 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
5375 dev
= list_next(l2arc_dev_list
, dev
)) {
5376 if (dev
->l2ad_vdev
== vd
)
5379 mutex_exit(&l2arc_dev_mtx
);
5381 return (dev
!= NULL
);
5385 * Add a vdev for use by the L2ARC. By this point the spa has already
5386 * validated the vdev and opened it.
5389 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
5391 l2arc_dev_t
*adddev
;
5393 ASSERT(!l2arc_vdev_present(vd
));
5396 * Create a new l2arc device entry.
5398 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
5399 adddev
->l2ad_spa
= spa
;
5400 adddev
->l2ad_vdev
= vd
;
5401 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
5402 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
5403 adddev
->l2ad_hand
= adddev
->l2ad_start
;
5404 adddev
->l2ad_evict
= adddev
->l2ad_start
;
5405 adddev
->l2ad_first
= B_TRUE
;
5406 adddev
->l2ad_writing
= B_FALSE
;
5407 list_link_init(&adddev
->l2ad_node
);
5410 * This is a list of all ARC buffers that are still valid on the
5413 adddev
->l2ad_buflist
= kmem_zalloc(sizeof (list_t
), KM_SLEEP
);
5414 list_create(adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
5415 offsetof(arc_buf_hdr_t
, b_l2node
));
5417 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
5420 * Add device to global list
5422 mutex_enter(&l2arc_dev_mtx
);
5423 list_insert_head(l2arc_dev_list
, adddev
);
5424 atomic_inc_64(&l2arc_ndev
);
5425 mutex_exit(&l2arc_dev_mtx
);
5429 * Remove a vdev from the L2ARC.
5432 l2arc_remove_vdev(vdev_t
*vd
)
5434 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
5437 * Find the device by vdev
5439 mutex_enter(&l2arc_dev_mtx
);
5440 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
5441 nextdev
= list_next(l2arc_dev_list
, dev
);
5442 if (vd
== dev
->l2ad_vdev
) {
5447 ASSERT(remdev
!= NULL
);
5450 * Remove device from global list
5452 list_remove(l2arc_dev_list
, remdev
);
5453 l2arc_dev_last
= NULL
; /* may have been invalidated */
5454 atomic_dec_64(&l2arc_ndev
);
5455 mutex_exit(&l2arc_dev_mtx
);
5458 * Clear all buflists and ARC references. L2ARC device flush.
5460 l2arc_evict(remdev
, 0, B_TRUE
);
5461 list_destroy(remdev
->l2ad_buflist
);
5462 kmem_free(remdev
->l2ad_buflist
, sizeof (list_t
));
5463 kmem_free(remdev
, sizeof (l2arc_dev_t
));
5469 l2arc_thread_exit
= 0;
5471 l2arc_writes_sent
= 0;
5472 l2arc_writes_done
= 0;
5474 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
5475 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
5476 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
5477 mutex_init(&l2arc_buflist_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
5478 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
5480 l2arc_dev_list
= &L2ARC_dev_list
;
5481 l2arc_free_on_write
= &L2ARC_free_on_write
;
5482 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
5483 offsetof(l2arc_dev_t
, l2ad_node
));
5484 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
5485 offsetof(l2arc_data_free_t
, l2df_list_node
));
5492 * This is called from dmu_fini(), which is called from spa_fini();
5493 * Because of this, we can assume that all l2arc devices have
5494 * already been removed when the pools themselves were removed.
5497 l2arc_do_free_on_write();
5499 mutex_destroy(&l2arc_feed_thr_lock
);
5500 cv_destroy(&l2arc_feed_thr_cv
);
5501 mutex_destroy(&l2arc_dev_mtx
);
5502 mutex_destroy(&l2arc_buflist_mtx
);
5503 mutex_destroy(&l2arc_free_on_write_mtx
);
5505 list_destroy(l2arc_dev_list
);
5506 list_destroy(l2arc_free_on_write
);
5512 if (!(spa_mode_global
& FWRITE
))
5515 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
5516 TS_RUN
, minclsyspri
);
5522 if (!(spa_mode_global
& FWRITE
))
5525 mutex_enter(&l2arc_feed_thr_lock
);
5526 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
5527 l2arc_thread_exit
= 1;
5528 while (l2arc_thread_exit
!= 0)
5529 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
5530 mutex_exit(&l2arc_feed_thr_lock
);
5533 #if defined(_KERNEL) && defined(HAVE_SPL)
5534 EXPORT_SYMBOL(arc_read
);
5535 EXPORT_SYMBOL(arc_buf_remove_ref
);
5536 EXPORT_SYMBOL(arc_buf_info
);
5537 EXPORT_SYMBOL(arc_getbuf_func
);
5538 EXPORT_SYMBOL(arc_add_prune_callback
);
5539 EXPORT_SYMBOL(arc_remove_prune_callback
);
5541 module_param(zfs_arc_min
, ulong
, 0644);
5542 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
5544 module_param(zfs_arc_max
, ulong
, 0644);
5545 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
5547 module_param(zfs_arc_meta_limit
, ulong
, 0644);
5548 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
5550 module_param(zfs_arc_meta_prune
, int, 0644);
5551 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Bytes of meta data to prune");
5553 module_param(zfs_arc_grow_retry
, int, 0644);
5554 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
5556 module_param(zfs_arc_shrink_shift
, int, 0644);
5557 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
5559 module_param(zfs_arc_p_min_shift
, int, 0644);
5560 MODULE_PARM_DESC(zfs_arc_p_min_shift
, "arc_c shift to calc min/max arc_p");
5562 module_param(zfs_disable_dup_eviction
, int, 0644);
5563 MODULE_PARM_DESC(zfs_disable_dup_eviction
, "disable duplicate buffer eviction");
5565 module_param(zfs_arc_memory_throttle_disable
, int, 0644);
5566 MODULE_PARM_DESC(zfs_arc_memory_throttle_disable
, "disable memory throttle");
5568 module_param(zfs_arc_min_prefetch_lifespan
, int, 0644);
5569 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan
, "Min life of prefetch block");
5571 module_param(l2arc_write_max
, ulong
, 0644);
5572 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
5574 module_param(l2arc_write_boost
, ulong
, 0644);
5575 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
5577 module_param(l2arc_headroom
, ulong
, 0644);
5578 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
5580 module_param(l2arc_headroom_boost
, ulong
, 0644);
5581 MODULE_PARM_DESC(l2arc_headroom_boost
, "Compressed l2arc_headroom multiplier");
5583 module_param(l2arc_feed_secs
, ulong
, 0644);
5584 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
5586 module_param(l2arc_feed_min_ms
, ulong
, 0644);
5587 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
5589 module_param(l2arc_noprefetch
, int, 0644);
5590 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
5592 module_param(l2arc_nocompress
, int, 0644);
5593 MODULE_PARM_DESC(l2arc_nocompress
, "Skip compressing L2ARC buffers");
5595 module_param(l2arc_feed_again
, int, 0644);
5596 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
5598 module_param(l2arc_norw
, int, 0644);
5599 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");