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 (c) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright 2014 Nexenta Systems, Inc. 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_clear_callback()
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 l2ad_mtx on each vdev 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>
138 #include <sys/multilist.h>
140 #include <sys/vmsystm.h>
142 #include <sys/fs/swapnode.h>
144 #include <linux/mm_compat.h>
146 #include <sys/callb.h>
147 #include <sys/kstat.h>
148 #include <sys/dmu_tx.h>
149 #include <zfs_fletcher.h>
150 #include <sys/arc_impl.h>
151 #include <sys/trace_arc.h>
154 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
155 boolean_t arc_watch
= B_FALSE
;
158 static kmutex_t arc_reclaim_lock
;
159 static kcondvar_t arc_reclaim_thread_cv
;
160 static boolean_t arc_reclaim_thread_exit
;
161 static kcondvar_t arc_reclaim_waiters_cv
;
163 static kmutex_t arc_user_evicts_lock
;
164 static kcondvar_t arc_user_evicts_cv
;
165 static boolean_t arc_user_evicts_thread_exit
;
167 /* number of objects to prune from caches when arc_meta_limit is reached */
168 int zfs_arc_meta_prune
= 10000;
170 /* The preferred strategy to employ when arc_meta_limit is reached */
171 int zfs_arc_meta_strategy
= ARC_STRATEGY_META_BALANCED
;
173 typedef enum arc_reclaim_strategy
{
174 ARC_RECLAIM_AGGR
, /* Aggressive reclaim strategy */
175 ARC_RECLAIM_CONS
/* Conservative reclaim strategy */
176 } arc_reclaim_strategy_t
;
179 * The number of headers to evict in arc_evict_state_impl() before
180 * dropping the sublist lock and evicting from another sublist. A lower
181 * value means we're more likely to evict the "correct" header (i.e. the
182 * oldest header in the arc state), but comes with higher overhead
183 * (i.e. more invocations of arc_evict_state_impl()).
185 int zfs_arc_evict_batch_limit
= 10;
188 * The number of sublists used for each of the arc state lists. If this
189 * is not set to a suitable value by the user, it will be configured to
190 * the number of CPUs on the system in arc_init().
192 int zfs_arc_num_sublists_per_state
= 0;
194 /* number of seconds before growing cache again */
195 int zfs_arc_grow_retry
= 5;
197 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
198 int zfs_arc_overflow_shift
= 8;
200 /* disable anon data aggressively growing arc_p */
201 int zfs_arc_p_aggressive_disable
= 1;
203 /* disable arc_p adapt dampener in arc_adapt */
204 int zfs_arc_p_dampener_disable
= 1;
206 /* log2(fraction of arc to reclaim) */
207 int zfs_arc_shrink_shift
= 5;
210 * minimum lifespan of a prefetch block in clock ticks
211 * (initialized in arc_init())
213 int zfs_arc_min_prefetch_lifespan
= HZ
;
215 /* disable arc proactive arc throttle due to low memory */
216 int zfs_arc_memory_throttle_disable
= 1;
218 /* disable duplicate buffer eviction */
219 int zfs_disable_dup_eviction
= 0;
221 /* average block used to size buf_hash_table */
222 int zfs_arc_average_blocksize
= 8 * 1024; /* 8KB */
225 * minimum lifespan of a prefetch block in clock ticks
226 * (initialized in arc_init())
228 static int arc_min_prefetch_lifespan
;
231 * If this percent of memory is free, don't throttle.
233 int arc_lotsfree_percent
= 10;
237 /* expiration time for arc_no_grow */
238 static clock_t arc_grow_time
= 0;
241 * The arc has filled available memory and has now warmed up.
243 static boolean_t arc_warm
;
246 * These tunables are for performance analysis.
248 unsigned long zfs_arc_max
= 0;
249 unsigned long zfs_arc_min
= 0;
250 unsigned long zfs_arc_meta_limit
= 0;
251 unsigned long zfs_arc_meta_min
= 0;
254 * Limit the number of restarts in arc_adjust_meta()
256 unsigned long zfs_arc_meta_adjust_restarts
= 4096;
259 static arc_state_t ARC_anon
;
260 static arc_state_t ARC_mru
;
261 static arc_state_t ARC_mru_ghost
;
262 static arc_state_t ARC_mfu
;
263 static arc_state_t ARC_mfu_ghost
;
264 static arc_state_t ARC_l2c_only
;
266 typedef struct arc_stats
{
267 kstat_named_t arcstat_hits
;
268 kstat_named_t arcstat_misses
;
269 kstat_named_t arcstat_demand_data_hits
;
270 kstat_named_t arcstat_demand_data_misses
;
271 kstat_named_t arcstat_demand_metadata_hits
;
272 kstat_named_t arcstat_demand_metadata_misses
;
273 kstat_named_t arcstat_prefetch_data_hits
;
274 kstat_named_t arcstat_prefetch_data_misses
;
275 kstat_named_t arcstat_prefetch_metadata_hits
;
276 kstat_named_t arcstat_prefetch_metadata_misses
;
277 kstat_named_t arcstat_mru_hits
;
278 kstat_named_t arcstat_mru_ghost_hits
;
279 kstat_named_t arcstat_mfu_hits
;
280 kstat_named_t arcstat_mfu_ghost_hits
;
281 kstat_named_t arcstat_deleted
;
283 * Number of buffers that could not be evicted because the hash lock
284 * was held by another thread. The lock may not necessarily be held
285 * by something using the same buffer, since hash locks are shared
286 * by multiple buffers.
288 kstat_named_t arcstat_mutex_miss
;
290 * Number of buffers skipped because they have I/O in progress, are
291 * indrect prefetch buffers that have not lived long enough, or are
292 * not from the spa we're trying to evict from.
294 kstat_named_t arcstat_evict_skip
;
296 * Number of times arc_evict_state() was unable to evict enough
297 * buffers to reach its target amount.
299 kstat_named_t arcstat_evict_not_enough
;
300 kstat_named_t arcstat_evict_l2_cached
;
301 kstat_named_t arcstat_evict_l2_eligible
;
302 kstat_named_t arcstat_evict_l2_ineligible
;
303 kstat_named_t arcstat_evict_l2_skip
;
304 kstat_named_t arcstat_hash_elements
;
305 kstat_named_t arcstat_hash_elements_max
;
306 kstat_named_t arcstat_hash_collisions
;
307 kstat_named_t arcstat_hash_chains
;
308 kstat_named_t arcstat_hash_chain_max
;
309 kstat_named_t arcstat_p
;
310 kstat_named_t arcstat_c
;
311 kstat_named_t arcstat_c_min
;
312 kstat_named_t arcstat_c_max
;
313 kstat_named_t arcstat_size
;
314 kstat_named_t arcstat_hdr_size
;
315 kstat_named_t arcstat_data_size
;
316 kstat_named_t arcstat_meta_size
;
317 kstat_named_t arcstat_other_size
;
318 kstat_named_t arcstat_anon_size
;
319 kstat_named_t arcstat_anon_evict_data
;
320 kstat_named_t arcstat_anon_evict_metadata
;
321 kstat_named_t arcstat_mru_size
;
322 kstat_named_t arcstat_mru_evict_data
;
323 kstat_named_t arcstat_mru_evict_metadata
;
324 kstat_named_t arcstat_mru_ghost_size
;
325 kstat_named_t arcstat_mru_ghost_evict_data
;
326 kstat_named_t arcstat_mru_ghost_evict_metadata
;
327 kstat_named_t arcstat_mfu_size
;
328 kstat_named_t arcstat_mfu_evict_data
;
329 kstat_named_t arcstat_mfu_evict_metadata
;
330 kstat_named_t arcstat_mfu_ghost_size
;
331 kstat_named_t arcstat_mfu_ghost_evict_data
;
332 kstat_named_t arcstat_mfu_ghost_evict_metadata
;
333 kstat_named_t arcstat_l2_hits
;
334 kstat_named_t arcstat_l2_misses
;
335 kstat_named_t arcstat_l2_feeds
;
336 kstat_named_t arcstat_l2_rw_clash
;
337 kstat_named_t arcstat_l2_read_bytes
;
338 kstat_named_t arcstat_l2_write_bytes
;
339 kstat_named_t arcstat_l2_writes_sent
;
340 kstat_named_t arcstat_l2_writes_done
;
341 kstat_named_t arcstat_l2_writes_error
;
342 kstat_named_t arcstat_l2_writes_lock_retry
;
343 kstat_named_t arcstat_l2_evict_lock_retry
;
344 kstat_named_t arcstat_l2_evict_reading
;
345 kstat_named_t arcstat_l2_evict_l1cached
;
346 kstat_named_t arcstat_l2_free_on_write
;
347 kstat_named_t arcstat_l2_cdata_free_on_write
;
348 kstat_named_t arcstat_l2_abort_lowmem
;
349 kstat_named_t arcstat_l2_cksum_bad
;
350 kstat_named_t arcstat_l2_io_error
;
351 kstat_named_t arcstat_l2_size
;
352 kstat_named_t arcstat_l2_asize
;
353 kstat_named_t arcstat_l2_hdr_size
;
354 kstat_named_t arcstat_l2_compress_successes
;
355 kstat_named_t arcstat_l2_compress_zeros
;
356 kstat_named_t arcstat_l2_compress_failures
;
357 kstat_named_t arcstat_memory_throttle_count
;
358 kstat_named_t arcstat_duplicate_buffers
;
359 kstat_named_t arcstat_duplicate_buffers_size
;
360 kstat_named_t arcstat_duplicate_reads
;
361 kstat_named_t arcstat_memory_direct_count
;
362 kstat_named_t arcstat_memory_indirect_count
;
363 kstat_named_t arcstat_no_grow
;
364 kstat_named_t arcstat_tempreserve
;
365 kstat_named_t arcstat_loaned_bytes
;
366 kstat_named_t arcstat_prune
;
367 kstat_named_t arcstat_meta_used
;
368 kstat_named_t arcstat_meta_limit
;
369 kstat_named_t arcstat_meta_max
;
370 kstat_named_t arcstat_meta_min
;
373 static arc_stats_t arc_stats
= {
374 { "hits", KSTAT_DATA_UINT64
},
375 { "misses", KSTAT_DATA_UINT64
},
376 { "demand_data_hits", KSTAT_DATA_UINT64
},
377 { "demand_data_misses", KSTAT_DATA_UINT64
},
378 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
379 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
380 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
381 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
382 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
383 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
384 { "mru_hits", KSTAT_DATA_UINT64
},
385 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
386 { "mfu_hits", KSTAT_DATA_UINT64
},
387 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
388 { "deleted", KSTAT_DATA_UINT64
},
389 { "mutex_miss", KSTAT_DATA_UINT64
},
390 { "evict_skip", KSTAT_DATA_UINT64
},
391 { "evict_not_enough", KSTAT_DATA_UINT64
},
392 { "evict_l2_cached", KSTAT_DATA_UINT64
},
393 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
394 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
395 { "evict_l2_skip", KSTAT_DATA_UINT64
},
396 { "hash_elements", KSTAT_DATA_UINT64
},
397 { "hash_elements_max", KSTAT_DATA_UINT64
},
398 { "hash_collisions", KSTAT_DATA_UINT64
},
399 { "hash_chains", KSTAT_DATA_UINT64
},
400 { "hash_chain_max", KSTAT_DATA_UINT64
},
401 { "p", KSTAT_DATA_UINT64
},
402 { "c", KSTAT_DATA_UINT64
},
403 { "c_min", KSTAT_DATA_UINT64
},
404 { "c_max", KSTAT_DATA_UINT64
},
405 { "size", KSTAT_DATA_UINT64
},
406 { "hdr_size", KSTAT_DATA_UINT64
},
407 { "data_size", KSTAT_DATA_UINT64
},
408 { "meta_size", KSTAT_DATA_UINT64
},
409 { "other_size", KSTAT_DATA_UINT64
},
410 { "anon_size", KSTAT_DATA_UINT64
},
411 { "anon_evict_data", KSTAT_DATA_UINT64
},
412 { "anon_evict_metadata", KSTAT_DATA_UINT64
},
413 { "mru_size", KSTAT_DATA_UINT64
},
414 { "mru_evict_data", KSTAT_DATA_UINT64
},
415 { "mru_evict_metadata", KSTAT_DATA_UINT64
},
416 { "mru_ghost_size", KSTAT_DATA_UINT64
},
417 { "mru_ghost_evict_data", KSTAT_DATA_UINT64
},
418 { "mru_ghost_evict_metadata", KSTAT_DATA_UINT64
},
419 { "mfu_size", KSTAT_DATA_UINT64
},
420 { "mfu_evict_data", KSTAT_DATA_UINT64
},
421 { "mfu_evict_metadata", KSTAT_DATA_UINT64
},
422 { "mfu_ghost_size", KSTAT_DATA_UINT64
},
423 { "mfu_ghost_evict_data", KSTAT_DATA_UINT64
},
424 { "mfu_ghost_evict_metadata", KSTAT_DATA_UINT64
},
425 { "l2_hits", KSTAT_DATA_UINT64
},
426 { "l2_misses", KSTAT_DATA_UINT64
},
427 { "l2_feeds", KSTAT_DATA_UINT64
},
428 { "l2_rw_clash", KSTAT_DATA_UINT64
},
429 { "l2_read_bytes", KSTAT_DATA_UINT64
},
430 { "l2_write_bytes", KSTAT_DATA_UINT64
},
431 { "l2_writes_sent", KSTAT_DATA_UINT64
},
432 { "l2_writes_done", KSTAT_DATA_UINT64
},
433 { "l2_writes_error", KSTAT_DATA_UINT64
},
434 { "l2_writes_lock_retry", KSTAT_DATA_UINT64
},
435 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
436 { "l2_evict_reading", KSTAT_DATA_UINT64
},
437 { "l2_evict_l1cached", KSTAT_DATA_UINT64
},
438 { "l2_free_on_write", KSTAT_DATA_UINT64
},
439 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64
},
440 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
441 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
442 { "l2_io_error", KSTAT_DATA_UINT64
},
443 { "l2_size", KSTAT_DATA_UINT64
},
444 { "l2_asize", KSTAT_DATA_UINT64
},
445 { "l2_hdr_size", KSTAT_DATA_UINT64
},
446 { "l2_compress_successes", KSTAT_DATA_UINT64
},
447 { "l2_compress_zeros", KSTAT_DATA_UINT64
},
448 { "l2_compress_failures", KSTAT_DATA_UINT64
},
449 { "memory_throttle_count", KSTAT_DATA_UINT64
},
450 { "duplicate_buffers", KSTAT_DATA_UINT64
},
451 { "duplicate_buffers_size", KSTAT_DATA_UINT64
},
452 { "duplicate_reads", KSTAT_DATA_UINT64
},
453 { "memory_direct_count", KSTAT_DATA_UINT64
},
454 { "memory_indirect_count", KSTAT_DATA_UINT64
},
455 { "arc_no_grow", KSTAT_DATA_UINT64
},
456 { "arc_tempreserve", KSTAT_DATA_UINT64
},
457 { "arc_loaned_bytes", KSTAT_DATA_UINT64
},
458 { "arc_prune", KSTAT_DATA_UINT64
},
459 { "arc_meta_used", KSTAT_DATA_UINT64
},
460 { "arc_meta_limit", KSTAT_DATA_UINT64
},
461 { "arc_meta_max", KSTAT_DATA_UINT64
},
462 { "arc_meta_min", KSTAT_DATA_UINT64
},
465 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
467 #define ARCSTAT_INCR(stat, val) \
468 atomic_add_64(&arc_stats.stat.value.ui64, (val))
470 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
471 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
473 #define ARCSTAT_MAX(stat, val) { \
475 while ((val) > (m = arc_stats.stat.value.ui64) && \
476 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
480 #define ARCSTAT_MAXSTAT(stat) \
481 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
484 * We define a macro to allow ARC hits/misses to be easily broken down by
485 * two separate conditions, giving a total of four different subtypes for
486 * each of hits and misses (so eight statistics total).
488 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
491 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
493 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
497 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
499 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
504 static arc_state_t
*arc_anon
;
505 static arc_state_t
*arc_mru
;
506 static arc_state_t
*arc_mru_ghost
;
507 static arc_state_t
*arc_mfu
;
508 static arc_state_t
*arc_mfu_ghost
;
509 static arc_state_t
*arc_l2c_only
;
512 * There are several ARC variables that are critical to export as kstats --
513 * but we don't want to have to grovel around in the kstat whenever we wish to
514 * manipulate them. For these variables, we therefore define them to be in
515 * terms of the statistic variable. This assures that we are not introducing
516 * the possibility of inconsistency by having shadow copies of the variables,
517 * while still allowing the code to be readable.
519 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
520 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
521 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
522 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
523 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
524 #define arc_no_grow ARCSTAT(arcstat_no_grow)
525 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
526 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
527 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
528 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
529 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
530 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
532 #define L2ARC_IS_VALID_COMPRESS(_c_) \
533 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
535 static list_t arc_prune_list
;
536 static kmutex_t arc_prune_mtx
;
537 static taskq_t
*arc_prune_taskq
;
538 static arc_buf_t
*arc_eviction_list
;
539 static arc_buf_hdr_t arc_eviction_hdr
;
541 #define GHOST_STATE(state) \
542 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
543 (state) == arc_l2c_only)
545 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
546 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
547 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
548 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
549 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
550 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
552 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
553 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
554 #define HDR_L2_READING(hdr) \
555 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
556 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
557 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
558 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
559 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
561 #define HDR_ISTYPE_METADATA(hdr) \
562 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
563 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
565 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
566 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
568 /* For storing compression mode in b_flags */
569 #define HDR_COMPRESS_OFFSET 24
570 #define HDR_COMPRESS_NBITS 7
572 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \
573 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS))
574 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \
575 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp))
581 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
582 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
585 * Hash table routines
588 #define HT_LOCK_ALIGN 64
589 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
594 unsigned char pad
[HT_LOCK_PAD
];
598 #define BUF_LOCKS 8192
599 typedef struct buf_hash_table
{
601 arc_buf_hdr_t
**ht_table
;
602 struct ht_lock ht_locks
[BUF_LOCKS
];
605 static buf_hash_table_t buf_hash_table
;
607 #define BUF_HASH_INDEX(spa, dva, birth) \
608 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
609 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
610 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
611 #define HDR_LOCK(hdr) \
612 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
614 uint64_t zfs_crc64_table
[256];
620 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
621 #define L2ARC_HEADROOM 2 /* num of writes */
623 * If we discover during ARC scan any buffers to be compressed, we boost
624 * our headroom for the next scanning cycle by this percentage multiple.
626 #define L2ARC_HEADROOM_BOOST 200
627 #define L2ARC_FEED_SECS 1 /* caching interval secs */
628 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
630 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
631 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
633 /* L2ARC Performance Tunables */
634 unsigned long l2arc_write_max
= L2ARC_WRITE_SIZE
; /* def max write size */
635 unsigned long l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra warmup write */
636 unsigned long l2arc_headroom
= L2ARC_HEADROOM
; /* # of dev writes */
637 unsigned long l2arc_headroom_boost
= L2ARC_HEADROOM_BOOST
;
638 unsigned long l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
639 unsigned long l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval msecs */
640 int l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
641 int l2arc_nocompress
= B_FALSE
; /* don't compress bufs */
642 int l2arc_feed_again
= B_TRUE
; /* turbo warmup */
643 int l2arc_norw
= B_FALSE
; /* no reads during writes */
648 static list_t L2ARC_dev_list
; /* device list */
649 static list_t
*l2arc_dev_list
; /* device list pointer */
650 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
651 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
652 static list_t L2ARC_free_on_write
; /* free after write buf list */
653 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
654 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
655 static uint64_t l2arc_ndev
; /* number of devices */
657 typedef struct l2arc_read_callback
{
658 arc_buf_t
*l2rcb_buf
; /* read buffer */
659 spa_t
*l2rcb_spa
; /* spa */
660 blkptr_t l2rcb_bp
; /* original blkptr */
661 zbookmark_phys_t l2rcb_zb
; /* original bookmark */
662 int l2rcb_flags
; /* original flags */
663 enum zio_compress l2rcb_compress
; /* applied compress */
664 } l2arc_read_callback_t
;
666 typedef struct l2arc_data_free
{
667 /* protected by l2arc_free_on_write_mtx */
670 void (*l2df_func
)(void *, size_t);
671 list_node_t l2df_list_node
;
674 static kmutex_t l2arc_feed_thr_lock
;
675 static kcondvar_t l2arc_feed_thr_cv
;
676 static uint8_t l2arc_thread_exit
;
678 static void arc_get_data_buf(arc_buf_t
*);
679 static void arc_access(arc_buf_hdr_t
*, kmutex_t
*);
680 static boolean_t
arc_is_overflowing(void);
681 static void arc_buf_watch(arc_buf_t
*);
683 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t
*);
684 static uint32_t arc_bufc_to_flags(arc_buf_contents_t
);
686 static boolean_t
l2arc_write_eligible(uint64_t, arc_buf_hdr_t
*);
687 static void l2arc_read_done(zio_t
*);
689 static boolean_t
l2arc_compress_buf(arc_buf_hdr_t
*);
690 static void l2arc_decompress_zio(zio_t
*, arc_buf_hdr_t
*, enum zio_compress
);
691 static void l2arc_release_cdata_buf(arc_buf_hdr_t
*);
694 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
696 uint8_t *vdva
= (uint8_t *)dva
;
697 uint64_t crc
= -1ULL;
700 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
702 for (i
= 0; i
< sizeof (dva_t
); i
++)
703 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
705 crc
^= (spa
>>8) ^ birth
;
710 #define BUF_EMPTY(buf) \
711 ((buf)->b_dva.dva_word[0] == 0 && \
712 (buf)->b_dva.dva_word[1] == 0)
714 #define BUF_EQUAL(spa, dva, birth, buf) \
715 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
716 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
717 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
720 buf_discard_identity(arc_buf_hdr_t
*hdr
)
722 hdr
->b_dva
.dva_word
[0] = 0;
723 hdr
->b_dva
.dva_word
[1] = 0;
727 static arc_buf_hdr_t
*
728 buf_hash_find(uint64_t spa
, const blkptr_t
*bp
, kmutex_t
**lockp
)
730 const dva_t
*dva
= BP_IDENTITY(bp
);
731 uint64_t birth
= BP_PHYSICAL_BIRTH(bp
);
732 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
733 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
736 mutex_enter(hash_lock
);
737 for (hdr
= buf_hash_table
.ht_table
[idx
]; hdr
!= NULL
;
738 hdr
= hdr
->b_hash_next
) {
739 if (BUF_EQUAL(spa
, dva
, birth
, hdr
)) {
744 mutex_exit(hash_lock
);
750 * Insert an entry into the hash table. If there is already an element
751 * equal to elem in the hash table, then the already existing element
752 * will be returned and the new element will not be inserted.
753 * Otherwise returns NULL.
754 * If lockp == NULL, the caller is assumed to already hold the hash lock.
756 static arc_buf_hdr_t
*
757 buf_hash_insert(arc_buf_hdr_t
*hdr
, kmutex_t
**lockp
)
759 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
760 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
764 ASSERT(!DVA_IS_EMPTY(&hdr
->b_dva
));
765 ASSERT(hdr
->b_birth
!= 0);
766 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
770 mutex_enter(hash_lock
);
772 ASSERT(MUTEX_HELD(hash_lock
));
775 for (fhdr
= buf_hash_table
.ht_table
[idx
], i
= 0; fhdr
!= NULL
;
776 fhdr
= fhdr
->b_hash_next
, i
++) {
777 if (BUF_EQUAL(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
, fhdr
))
781 hdr
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
782 buf_hash_table
.ht_table
[idx
] = hdr
;
783 hdr
->b_flags
|= ARC_FLAG_IN_HASH_TABLE
;
785 /* collect some hash table performance data */
787 ARCSTAT_BUMP(arcstat_hash_collisions
);
789 ARCSTAT_BUMP(arcstat_hash_chains
);
791 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
794 ARCSTAT_BUMP(arcstat_hash_elements
);
795 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
801 buf_hash_remove(arc_buf_hdr_t
*hdr
)
803 arc_buf_hdr_t
*fhdr
, **hdrp
;
804 uint64_t idx
= BUF_HASH_INDEX(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
);
806 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
807 ASSERT(HDR_IN_HASH_TABLE(hdr
));
809 hdrp
= &buf_hash_table
.ht_table
[idx
];
810 while ((fhdr
= *hdrp
) != hdr
) {
811 ASSERT(fhdr
!= NULL
);
812 hdrp
= &fhdr
->b_hash_next
;
814 *hdrp
= hdr
->b_hash_next
;
815 hdr
->b_hash_next
= NULL
;
816 hdr
->b_flags
&= ~ARC_FLAG_IN_HASH_TABLE
;
818 /* collect some hash table performance data */
819 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
821 if (buf_hash_table
.ht_table
[idx
] &&
822 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
823 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
827 * Global data structures and functions for the buf kmem cache.
829 static kmem_cache_t
*hdr_full_cache
;
830 static kmem_cache_t
*hdr_l2only_cache
;
831 static kmem_cache_t
*buf_cache
;
838 #if defined(_KERNEL) && defined(HAVE_SPL)
840 * Large allocations which do not require contiguous pages
841 * should be using vmem_free() in the linux kernel\
843 vmem_free(buf_hash_table
.ht_table
,
844 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
846 kmem_free(buf_hash_table
.ht_table
,
847 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
849 for (i
= 0; i
< BUF_LOCKS
; i
++)
850 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
851 kmem_cache_destroy(hdr_full_cache
);
852 kmem_cache_destroy(hdr_l2only_cache
);
853 kmem_cache_destroy(buf_cache
);
857 * Constructor callback - called when the cache is empty
858 * and a new buf is requested.
862 hdr_full_cons(void *vbuf
, void *unused
, int kmflag
)
864 arc_buf_hdr_t
*hdr
= vbuf
;
866 bzero(hdr
, HDR_FULL_SIZE
);
867 cv_init(&hdr
->b_l1hdr
.b_cv
, NULL
, CV_DEFAULT
, NULL
);
868 refcount_create(&hdr
->b_l1hdr
.b_refcnt
);
869 mutex_init(&hdr
->b_l1hdr
.b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
870 list_link_init(&hdr
->b_l1hdr
.b_arc_node
);
871 list_link_init(&hdr
->b_l2hdr
.b_l2node
);
872 multilist_link_init(&hdr
->b_l1hdr
.b_arc_node
);
873 arc_space_consume(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
880 hdr_l2only_cons(void *vbuf
, void *unused
, int kmflag
)
882 arc_buf_hdr_t
*hdr
= vbuf
;
884 bzero(hdr
, HDR_L2ONLY_SIZE
);
885 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
892 buf_cons(void *vbuf
, void *unused
, int kmflag
)
894 arc_buf_t
*buf
= vbuf
;
896 bzero(buf
, sizeof (arc_buf_t
));
897 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
898 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
904 * Destructor callback - called when a cached buf is
905 * no longer required.
909 hdr_full_dest(void *vbuf
, void *unused
)
911 arc_buf_hdr_t
*hdr
= vbuf
;
913 ASSERT(BUF_EMPTY(hdr
));
914 cv_destroy(&hdr
->b_l1hdr
.b_cv
);
915 refcount_destroy(&hdr
->b_l1hdr
.b_refcnt
);
916 mutex_destroy(&hdr
->b_l1hdr
.b_freeze_lock
);
917 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
918 arc_space_return(HDR_FULL_SIZE
, ARC_SPACE_HDRS
);
923 hdr_l2only_dest(void *vbuf
, void *unused
)
925 ASSERTV(arc_buf_hdr_t
*hdr
= vbuf
);
927 ASSERT(BUF_EMPTY(hdr
));
928 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
933 buf_dest(void *vbuf
, void *unused
)
935 arc_buf_t
*buf
= vbuf
;
937 mutex_destroy(&buf
->b_evict_lock
);
938 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
945 uint64_t hsize
= 1ULL << 12;
949 * The hash table is big enough to fill all of physical memory
950 * with an average block size of zfs_arc_average_blocksize (default 8K).
951 * By default, the table will take up
952 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
954 while (hsize
* zfs_arc_average_blocksize
< physmem
* PAGESIZE
)
957 buf_hash_table
.ht_mask
= hsize
- 1;
958 #if defined(_KERNEL) && defined(HAVE_SPL)
960 * Large allocations which do not require contiguous pages
961 * should be using vmem_alloc() in the linux kernel
963 buf_hash_table
.ht_table
=
964 vmem_zalloc(hsize
* sizeof (void*), KM_SLEEP
);
966 buf_hash_table
.ht_table
=
967 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
969 if (buf_hash_table
.ht_table
== NULL
) {
970 ASSERT(hsize
> (1ULL << 8));
975 hdr_full_cache
= kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE
,
976 0, hdr_full_cons
, hdr_full_dest
, NULL
, NULL
, NULL
, 0);
977 hdr_l2only_cache
= kmem_cache_create("arc_buf_hdr_t_l2only",
978 HDR_L2ONLY_SIZE
, 0, hdr_l2only_cons
, hdr_l2only_dest
, NULL
,
980 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
981 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
983 for (i
= 0; i
< 256; i
++)
984 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
985 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
987 for (i
= 0; i
< BUF_LOCKS
; i
++) {
988 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
989 NULL
, MUTEX_DEFAULT
, NULL
);
994 * Transition between the two allocation states for the arc_buf_hdr struct.
995 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
996 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
997 * version is used when a cache buffer is only in the L2ARC in order to reduce
1000 static arc_buf_hdr_t
*
1001 arc_hdr_realloc(arc_buf_hdr_t
*hdr
, kmem_cache_t
*old
, kmem_cache_t
*new)
1003 arc_buf_hdr_t
*nhdr
;
1006 ASSERT(HDR_HAS_L2HDR(hdr
));
1007 ASSERT((old
== hdr_full_cache
&& new == hdr_l2only_cache
) ||
1008 (old
== hdr_l2only_cache
&& new == hdr_full_cache
));
1010 dev
= hdr
->b_l2hdr
.b_dev
;
1011 nhdr
= kmem_cache_alloc(new, KM_PUSHPAGE
);
1013 ASSERT(MUTEX_HELD(HDR_LOCK(hdr
)));
1014 buf_hash_remove(hdr
);
1016 bcopy(hdr
, nhdr
, HDR_L2ONLY_SIZE
);
1017 if (new == hdr_full_cache
) {
1018 nhdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
1020 * arc_access and arc_change_state need to be aware that a
1021 * header has just come out of L2ARC, so we set its state to
1022 * l2c_only even though it's about to change.
1024 nhdr
->b_l1hdr
.b_state
= arc_l2c_only
;
1026 /* Verify previous threads set to NULL before freeing */
1027 ASSERT3P(nhdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1029 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1030 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
1033 * If we've reached here, We must have been called from
1034 * arc_evict_hdr(), as such we should have already been
1035 * removed from any ghost list we were previously on
1036 * (which protects us from racing with arc_evict_state),
1037 * thus no locking is needed during this check.
1039 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1042 * A buffer must not be moved into the arc_l2c_only
1043 * state if it's not finished being written out to the
1044 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1045 * might try to be accessed, even though it was removed.
1047 VERIFY(!HDR_L2_WRITING(hdr
));
1048 VERIFY3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1050 nhdr
->b_flags
&= ~ARC_FLAG_HAS_L1HDR
;
1053 * The header has been reallocated so we need to re-insert it into any
1056 (void) buf_hash_insert(nhdr
, NULL
);
1058 ASSERT(list_link_active(&hdr
->b_l2hdr
.b_l2node
));
1060 mutex_enter(&dev
->l2ad_mtx
);
1063 * We must place the realloc'ed header back into the list at
1064 * the same spot. Otherwise, if it's placed earlier in the list,
1065 * l2arc_write_buffers() could find it during the function's
1066 * write phase, and try to write it out to the l2arc.
1068 list_insert_after(&dev
->l2ad_buflist
, hdr
, nhdr
);
1069 list_remove(&dev
->l2ad_buflist
, hdr
);
1071 mutex_exit(&dev
->l2ad_mtx
);
1073 buf_discard_identity(hdr
);
1074 hdr
->b_freeze_cksum
= NULL
;
1075 kmem_cache_free(old
, hdr
);
1081 #define ARC_MINTIME (hz>>4) /* 62 ms */
1084 arc_cksum_verify(arc_buf_t
*buf
)
1088 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1091 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1092 if (buf
->b_hdr
->b_freeze_cksum
== NULL
|| HDR_IO_ERROR(buf
->b_hdr
)) {
1093 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1096 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
1097 if (!ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
))
1098 panic("buffer modified while frozen!");
1099 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1103 arc_cksum_equal(arc_buf_t
*buf
)
1108 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1109 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
1110 equal
= ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
);
1111 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1117 arc_cksum_compute(arc_buf_t
*buf
, boolean_t force
)
1119 if (!force
&& !(zfs_flags
& ZFS_DEBUG_MODIFY
))
1122 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1123 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1124 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1127 buf
->b_hdr
->b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
),
1129 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
,
1130 buf
->b_hdr
->b_freeze_cksum
);
1131 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1137 arc_buf_sigsegv(int sig
, siginfo_t
*si
, void *unused
)
1139 panic("Got SIGSEGV at address: 0x%lx\n", (long) si
->si_addr
);
1145 arc_buf_unwatch(arc_buf_t
*buf
)
1149 ASSERT0(mprotect(buf
->b_data
, buf
->b_hdr
->b_size
,
1150 PROT_READ
| PROT_WRITE
));
1157 arc_buf_watch(arc_buf_t
*buf
)
1161 ASSERT0(mprotect(buf
->b_data
, buf
->b_hdr
->b_size
, PROT_READ
));
1165 static arc_buf_contents_t
1166 arc_buf_type(arc_buf_hdr_t
*hdr
)
1168 if (HDR_ISTYPE_METADATA(hdr
)) {
1169 return (ARC_BUFC_METADATA
);
1171 return (ARC_BUFC_DATA
);
1176 arc_bufc_to_flags(arc_buf_contents_t type
)
1180 /* metadata field is 0 if buffer contains normal data */
1182 case ARC_BUFC_METADATA
:
1183 return (ARC_FLAG_BUFC_METADATA
);
1187 panic("undefined ARC buffer type!");
1188 return ((uint32_t)-1);
1192 arc_buf_thaw(arc_buf_t
*buf
)
1194 if (zfs_flags
& ZFS_DEBUG_MODIFY
) {
1195 if (buf
->b_hdr
->b_l1hdr
.b_state
!= arc_anon
)
1196 panic("modifying non-anon buffer!");
1197 if (HDR_IO_IN_PROGRESS(buf
->b_hdr
))
1198 panic("modifying buffer while i/o in progress!");
1199 arc_cksum_verify(buf
);
1202 mutex_enter(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1203 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
1204 kmem_free(buf
->b_hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
1205 buf
->b_hdr
->b_freeze_cksum
= NULL
;
1208 mutex_exit(&buf
->b_hdr
->b_l1hdr
.b_freeze_lock
);
1210 arc_buf_unwatch(buf
);
1214 arc_buf_freeze(arc_buf_t
*buf
)
1216 kmutex_t
*hash_lock
;
1218 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
1221 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1222 mutex_enter(hash_lock
);
1224 ASSERT(buf
->b_hdr
->b_freeze_cksum
!= NULL
||
1225 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
1226 arc_cksum_compute(buf
, B_FALSE
);
1227 mutex_exit(hash_lock
);
1232 add_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1236 ASSERT(HDR_HAS_L1HDR(hdr
));
1237 ASSERT(MUTEX_HELD(hash_lock
));
1239 state
= hdr
->b_l1hdr
.b_state
;
1241 if ((refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
) == 1) &&
1242 (state
!= arc_anon
)) {
1243 /* We don't use the L2-only state list. */
1244 if (state
!= arc_l2c_only
) {
1245 arc_buf_contents_t type
= arc_buf_type(hdr
);
1246 uint64_t delta
= hdr
->b_size
* hdr
->b_l1hdr
.b_datacnt
;
1247 multilist_t
*list
= &state
->arcs_list
[type
];
1248 uint64_t *size
= &state
->arcs_lsize
[type
];
1250 multilist_remove(list
, hdr
);
1252 if (GHOST_STATE(state
)) {
1253 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
1254 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
1255 delta
= hdr
->b_size
;
1258 ASSERT3U(*size
, >=, delta
);
1259 atomic_add_64(size
, -delta
);
1261 /* remove the prefetch flag if we get a reference */
1262 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
1267 remove_reference(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
, void *tag
)
1270 arc_state_t
*state
= hdr
->b_l1hdr
.b_state
;
1272 ASSERT(HDR_HAS_L1HDR(hdr
));
1273 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
1274 ASSERT(!GHOST_STATE(state
));
1277 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1278 * check to prevent usage of the arc_l2c_only list.
1280 if (((cnt
= refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
)) == 0) &&
1281 (state
!= arc_anon
)) {
1282 arc_buf_contents_t type
= arc_buf_type(hdr
);
1283 multilist_t
*list
= &state
->arcs_list
[type
];
1284 uint64_t *size
= &state
->arcs_lsize
[type
];
1286 multilist_insert(list
, hdr
);
1288 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
1289 atomic_add_64(size
, hdr
->b_size
*
1290 hdr
->b_l1hdr
.b_datacnt
);
1296 * Returns detailed information about a specific arc buffer. When the
1297 * state_index argument is set the function will calculate the arc header
1298 * list position for its arc state. Since this requires a linear traversal
1299 * callers are strongly encourage not to do this. However, it can be helpful
1300 * for targeted analysis so the functionality is provided.
1303 arc_buf_info(arc_buf_t
*ab
, arc_buf_info_t
*abi
, int state_index
)
1305 arc_buf_hdr_t
*hdr
= ab
->b_hdr
;
1306 l1arc_buf_hdr_t
*l1hdr
= NULL
;
1307 l2arc_buf_hdr_t
*l2hdr
= NULL
;
1308 arc_state_t
*state
= NULL
;
1310 if (HDR_HAS_L1HDR(hdr
)) {
1311 l1hdr
= &hdr
->b_l1hdr
;
1312 state
= l1hdr
->b_state
;
1314 if (HDR_HAS_L2HDR(hdr
))
1315 l2hdr
= &hdr
->b_l2hdr
;
1317 memset(abi
, 0, sizeof (arc_buf_info_t
));
1318 abi
->abi_flags
= hdr
->b_flags
;
1321 abi
->abi_datacnt
= l1hdr
->b_datacnt
;
1322 abi
->abi_access
= l1hdr
->b_arc_access
;
1323 abi
->abi_mru_hits
= l1hdr
->b_mru_hits
;
1324 abi
->abi_mru_ghost_hits
= l1hdr
->b_mru_ghost_hits
;
1325 abi
->abi_mfu_hits
= l1hdr
->b_mfu_hits
;
1326 abi
->abi_mfu_ghost_hits
= l1hdr
->b_mfu_ghost_hits
;
1327 abi
->abi_holds
= refcount_count(&l1hdr
->b_refcnt
);
1331 abi
->abi_l2arc_dattr
= l2hdr
->b_daddr
;
1332 abi
->abi_l2arc_asize
= l2hdr
->b_asize
;
1333 abi
->abi_l2arc_compress
= HDR_GET_COMPRESS(hdr
);
1334 abi
->abi_l2arc_hits
= l2hdr
->b_hits
;
1337 abi
->abi_state_type
= state
? state
->arcs_state
: ARC_STATE_ANON
;
1338 abi
->abi_state_contents
= arc_buf_type(hdr
);
1339 abi
->abi_size
= hdr
->b_size
;
1343 * Move the supplied buffer to the indicated state. The hash lock
1344 * for the buffer must be held by the caller.
1347 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*hdr
,
1348 kmutex_t
*hash_lock
)
1350 arc_state_t
*old_state
;
1353 uint64_t from_delta
, to_delta
;
1354 arc_buf_contents_t buftype
= arc_buf_type(hdr
);
1357 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1358 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1359 * L1 hdr doesn't always exist when we change state to arc_anon before
1360 * destroying a header, in which case reallocating to add the L1 hdr is
1363 if (HDR_HAS_L1HDR(hdr
)) {
1364 old_state
= hdr
->b_l1hdr
.b_state
;
1365 refcnt
= refcount_count(&hdr
->b_l1hdr
.b_refcnt
);
1366 datacnt
= hdr
->b_l1hdr
.b_datacnt
;
1368 old_state
= arc_l2c_only
;
1373 ASSERT(MUTEX_HELD(hash_lock
));
1374 ASSERT3P(new_state
, !=, old_state
);
1375 ASSERT(refcnt
== 0 || datacnt
> 0);
1376 ASSERT(!GHOST_STATE(new_state
) || datacnt
== 0);
1377 ASSERT(old_state
!= arc_anon
|| datacnt
<= 1);
1379 from_delta
= to_delta
= datacnt
* hdr
->b_size
;
1382 * If this buffer is evictable, transfer it from the
1383 * old state list to the new state list.
1386 if (old_state
!= arc_anon
&& old_state
!= arc_l2c_only
) {
1387 uint64_t *size
= &old_state
->arcs_lsize
[buftype
];
1389 ASSERT(HDR_HAS_L1HDR(hdr
));
1390 multilist_remove(&old_state
->arcs_list
[buftype
], hdr
);
1393 * If prefetching out of the ghost cache,
1394 * we will have a non-zero datacnt.
1396 if (GHOST_STATE(old_state
) && datacnt
== 0) {
1397 /* ghost elements have a ghost size */
1398 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1399 from_delta
= hdr
->b_size
;
1401 ASSERT3U(*size
, >=, from_delta
);
1402 atomic_add_64(size
, -from_delta
);
1404 if (new_state
!= arc_anon
&& new_state
!= arc_l2c_only
) {
1405 uint64_t *size
= &new_state
->arcs_lsize
[buftype
];
1408 * An L1 header always exists here, since if we're
1409 * moving to some L1-cached state (i.e. not l2c_only or
1410 * anonymous), we realloc the header to add an L1hdr
1413 ASSERT(HDR_HAS_L1HDR(hdr
));
1414 multilist_insert(&new_state
->arcs_list
[buftype
], hdr
);
1416 /* ghost elements have a ghost size */
1417 if (GHOST_STATE(new_state
)) {
1419 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
1420 to_delta
= hdr
->b_size
;
1422 atomic_add_64(size
, to_delta
);
1426 ASSERT(!BUF_EMPTY(hdr
));
1427 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(hdr
))
1428 buf_hash_remove(hdr
);
1430 /* adjust state sizes (ignore arc_l2c_only) */
1431 if (to_delta
&& new_state
!= arc_l2c_only
)
1432 atomic_add_64(&new_state
->arcs_size
, to_delta
);
1433 if (from_delta
&& old_state
!= arc_l2c_only
) {
1434 ASSERT3U(old_state
->arcs_size
, >=, from_delta
);
1435 atomic_add_64(&old_state
->arcs_size
, -from_delta
);
1437 if (HDR_HAS_L1HDR(hdr
))
1438 hdr
->b_l1hdr
.b_state
= new_state
;
1441 * L2 headers should never be on the L2 state list since they don't
1442 * have L1 headers allocated.
1444 ASSERT(multilist_is_empty(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]) &&
1445 multilist_is_empty(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]));
1449 arc_space_consume(uint64_t space
, arc_space_type_t type
)
1451 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1456 case ARC_SPACE_DATA
:
1457 ARCSTAT_INCR(arcstat_data_size
, space
);
1459 case ARC_SPACE_META
:
1460 ARCSTAT_INCR(arcstat_meta_size
, space
);
1462 case ARC_SPACE_OTHER
:
1463 ARCSTAT_INCR(arcstat_other_size
, space
);
1465 case ARC_SPACE_HDRS
:
1466 ARCSTAT_INCR(arcstat_hdr_size
, space
);
1468 case ARC_SPACE_L2HDRS
:
1469 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
1473 if (type
!= ARC_SPACE_DATA
) {
1474 ARCSTAT_INCR(arcstat_meta_used
, space
);
1475 if (arc_meta_max
< arc_meta_used
)
1476 arc_meta_max
= arc_meta_used
;
1479 atomic_add_64(&arc_size
, space
);
1483 arc_space_return(uint64_t space
, arc_space_type_t type
)
1485 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1490 case ARC_SPACE_DATA
:
1491 ARCSTAT_INCR(arcstat_data_size
, -space
);
1493 case ARC_SPACE_META
:
1494 ARCSTAT_INCR(arcstat_meta_size
, -space
);
1496 case ARC_SPACE_OTHER
:
1497 ARCSTAT_INCR(arcstat_other_size
, -space
);
1499 case ARC_SPACE_HDRS
:
1500 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
1502 case ARC_SPACE_L2HDRS
:
1503 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
1507 if (type
!= ARC_SPACE_DATA
) {
1508 ASSERT(arc_meta_used
>= space
);
1509 ARCSTAT_INCR(arcstat_meta_used
, -space
);
1512 ASSERT(arc_size
>= space
);
1513 atomic_add_64(&arc_size
, -space
);
1517 arc_buf_alloc(spa_t
*spa
, uint64_t size
, void *tag
, arc_buf_contents_t type
)
1522 VERIFY3U(size
, <=, spa_maxblocksize(spa
));
1523 hdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
1524 ASSERT(BUF_EMPTY(hdr
));
1525 ASSERT3P(hdr
->b_freeze_cksum
, ==, NULL
);
1527 hdr
->b_spa
= spa_load_guid(spa
);
1528 hdr
->b_l1hdr
.b_mru_hits
= 0;
1529 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
1530 hdr
->b_l1hdr
.b_mfu_hits
= 0;
1531 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
1532 hdr
->b_l1hdr
.b_l2_hits
= 0;
1534 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1537 buf
->b_efunc
= NULL
;
1538 buf
->b_private
= NULL
;
1541 hdr
->b_flags
= arc_bufc_to_flags(type
);
1542 hdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
1544 hdr
->b_l1hdr
.b_buf
= buf
;
1545 hdr
->b_l1hdr
.b_state
= arc_anon
;
1546 hdr
->b_l1hdr
.b_arc_access
= 0;
1547 hdr
->b_l1hdr
.b_datacnt
= 1;
1548 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1550 arc_get_data_buf(buf
);
1552 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
1553 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1558 static char *arc_onloan_tag
= "onloan";
1561 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1562 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1563 * buffers must be returned to the arc before they can be used by the DMU or
1567 arc_loan_buf(spa_t
*spa
, uint64_t size
)
1571 buf
= arc_buf_alloc(spa
, size
, arc_onloan_tag
, ARC_BUFC_DATA
);
1573 atomic_add_64(&arc_loaned_bytes
, size
);
1578 * Return a loaned arc buffer to the arc.
1581 arc_return_buf(arc_buf_t
*buf
, void *tag
)
1583 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1585 ASSERT(buf
->b_data
!= NULL
);
1586 ASSERT(HDR_HAS_L1HDR(hdr
));
1587 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1588 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
1590 atomic_add_64(&arc_loaned_bytes
, -hdr
->b_size
);
1593 /* Detach an arc_buf from a dbuf (tag) */
1595 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
1597 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1599 ASSERT(buf
->b_data
!= NULL
);
1600 ASSERT(HDR_HAS_L1HDR(hdr
));
1601 (void) refcount_add(&hdr
->b_l1hdr
.b_refcnt
, arc_onloan_tag
);
1602 (void) refcount_remove(&hdr
->b_l1hdr
.b_refcnt
, tag
);
1603 buf
->b_efunc
= NULL
;
1604 buf
->b_private
= NULL
;
1606 atomic_add_64(&arc_loaned_bytes
, hdr
->b_size
);
1610 arc_buf_clone(arc_buf_t
*from
)
1613 arc_buf_hdr_t
*hdr
= from
->b_hdr
;
1614 uint64_t size
= hdr
->b_size
;
1616 ASSERT(HDR_HAS_L1HDR(hdr
));
1617 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_anon
);
1619 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1622 buf
->b_efunc
= NULL
;
1623 buf
->b_private
= NULL
;
1624 buf
->b_next
= hdr
->b_l1hdr
.b_buf
;
1625 hdr
->b_l1hdr
.b_buf
= buf
;
1626 arc_get_data_buf(buf
);
1627 bcopy(from
->b_data
, buf
->b_data
, size
);
1630 * This buffer already exists in the arc so create a duplicate
1631 * copy for the caller. If the buffer is associated with user data
1632 * then track the size and number of duplicates. These stats will be
1633 * updated as duplicate buffers are created and destroyed.
1635 if (HDR_ISTYPE_DATA(hdr
)) {
1636 ARCSTAT_BUMP(arcstat_duplicate_buffers
);
1637 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, size
);
1639 hdr
->b_l1hdr
.b_datacnt
+= 1;
1644 arc_buf_add_ref(arc_buf_t
*buf
, void* tag
)
1647 kmutex_t
*hash_lock
;
1650 * Check to see if this buffer is evicted. Callers
1651 * must verify b_data != NULL to know if the add_ref
1654 mutex_enter(&buf
->b_evict_lock
);
1655 if (buf
->b_data
== NULL
) {
1656 mutex_exit(&buf
->b_evict_lock
);
1659 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1660 mutex_enter(hash_lock
);
1662 ASSERT(HDR_HAS_L1HDR(hdr
));
1663 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1664 mutex_exit(&buf
->b_evict_lock
);
1666 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
1667 hdr
->b_l1hdr
.b_state
== arc_mfu
);
1669 add_reference(hdr
, hash_lock
, tag
);
1670 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
1671 arc_access(hdr
, hash_lock
);
1672 mutex_exit(hash_lock
);
1673 ARCSTAT_BUMP(arcstat_hits
);
1674 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
1675 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
1676 data
, metadata
, hits
);
1680 arc_buf_free_on_write(void *data
, size_t size
,
1681 void (*free_func
)(void *, size_t))
1683 l2arc_data_free_t
*df
;
1685 df
= kmem_alloc(sizeof (*df
), KM_SLEEP
);
1686 df
->l2df_data
= data
;
1687 df
->l2df_size
= size
;
1688 df
->l2df_func
= free_func
;
1689 mutex_enter(&l2arc_free_on_write_mtx
);
1690 list_insert_head(l2arc_free_on_write
, df
);
1691 mutex_exit(&l2arc_free_on_write_mtx
);
1695 * Free the arc data buffer. If it is an l2arc write in progress,
1696 * the buffer is placed on l2arc_free_on_write to be freed later.
1699 arc_buf_data_free(arc_buf_t
*buf
, void (*free_func
)(void *, size_t))
1701 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1703 if (HDR_L2_WRITING(hdr
)) {
1704 arc_buf_free_on_write(buf
->b_data
, hdr
->b_size
, free_func
);
1705 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
1707 free_func(buf
->b_data
, hdr
->b_size
);
1712 arc_buf_l2_cdata_free(arc_buf_hdr_t
*hdr
)
1714 ASSERT(HDR_HAS_L2HDR(hdr
));
1715 ASSERT(MUTEX_HELD(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
));
1718 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
1719 * that doesn't exist, the header is in the arc_l2c_only state,
1720 * and there isn't anything to free (it's already been freed).
1722 if (!HDR_HAS_L1HDR(hdr
))
1726 * The header isn't being written to the l2arc device, thus it
1727 * shouldn't have a b_tmp_cdata to free.
1729 if (!HDR_L2_WRITING(hdr
)) {
1730 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1735 * The header does not have compression enabled. This can be due
1736 * to the buffer not being compressible, or because we're
1737 * freeing the buffer before the second phase of
1738 * l2arc_write_buffer() has started (which does the compression
1739 * step). In either case, b_tmp_cdata does not point to a
1740 * separately compressed buffer, so there's nothing to free (it
1741 * points to the same buffer as the arc_buf_t's b_data field).
1743 if (HDR_GET_COMPRESS(hdr
) == ZIO_COMPRESS_OFF
) {
1744 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1749 * There's nothing to free since the buffer was all zero's and
1750 * compressed to a zero length buffer.
1752 if (HDR_GET_COMPRESS(hdr
) == ZIO_COMPRESS_EMPTY
) {
1753 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
1757 ASSERT(L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr
)));
1759 arc_buf_free_on_write(hdr
->b_l1hdr
.b_tmp_cdata
,
1760 hdr
->b_size
, zio_data_buf_free
);
1762 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write
);
1763 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
1767 * Free up buf->b_data and if 'remove' is set, then pull the
1768 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
1771 arc_buf_destroy(arc_buf_t
*buf
, boolean_t remove
)
1775 /* free up data associated with the buf */
1776 if (buf
->b_data
!= NULL
) {
1777 arc_state_t
*state
= buf
->b_hdr
->b_l1hdr
.b_state
;
1778 uint64_t size
= buf
->b_hdr
->b_size
;
1779 arc_buf_contents_t type
= arc_buf_type(buf
->b_hdr
);
1781 arc_cksum_verify(buf
);
1782 arc_buf_unwatch(buf
);
1784 if (type
== ARC_BUFC_METADATA
) {
1785 arc_buf_data_free(buf
, zio_buf_free
);
1786 arc_space_return(size
, ARC_SPACE_META
);
1788 ASSERT(type
== ARC_BUFC_DATA
);
1789 arc_buf_data_free(buf
, zio_data_buf_free
);
1790 arc_space_return(size
, ARC_SPACE_DATA
);
1793 /* protected by hash lock, if in the hash table */
1794 if (multilist_link_active(&buf
->b_hdr
->b_l1hdr
.b_arc_node
)) {
1795 uint64_t *cnt
= &state
->arcs_lsize
[type
];
1797 ASSERT(refcount_is_zero(
1798 &buf
->b_hdr
->b_l1hdr
.b_refcnt
));
1799 ASSERT(state
!= arc_anon
&& state
!= arc_l2c_only
);
1801 ASSERT3U(*cnt
, >=, size
);
1802 atomic_add_64(cnt
, -size
);
1804 ASSERT3U(state
->arcs_size
, >=, size
);
1805 atomic_add_64(&state
->arcs_size
, -size
);
1809 * If we're destroying a duplicate buffer make sure
1810 * that the appropriate statistics are updated.
1812 if (buf
->b_hdr
->b_l1hdr
.b_datacnt
> 1 &&
1813 HDR_ISTYPE_DATA(buf
->b_hdr
)) {
1814 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
1815 ARCSTAT_INCR(arcstat_duplicate_buffers_size
, -size
);
1817 ASSERT(buf
->b_hdr
->b_l1hdr
.b_datacnt
> 0);
1818 buf
->b_hdr
->b_l1hdr
.b_datacnt
-= 1;
1821 /* only remove the buf if requested */
1825 /* remove the buf from the hdr list */
1826 for (bufp
= &buf
->b_hdr
->b_l1hdr
.b_buf
; *bufp
!= buf
;
1827 bufp
= &(*bufp
)->b_next
)
1829 *bufp
= buf
->b_next
;
1832 ASSERT(buf
->b_efunc
== NULL
);
1834 /* clean up the buf */
1836 kmem_cache_free(buf_cache
, buf
);
1840 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
1842 if (HDR_HAS_L1HDR(hdr
)) {
1843 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
||
1844 hdr
->b_l1hdr
.b_datacnt
> 0);
1845 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
1846 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
1848 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
1849 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
1851 if (HDR_HAS_L2HDR(hdr
)) {
1852 l2arc_buf_hdr_t
*l2hdr
= &hdr
->b_l2hdr
;
1853 boolean_t buflist_held
= MUTEX_HELD(&l2hdr
->b_dev
->l2ad_mtx
);
1855 if (!buflist_held
) {
1856 mutex_enter(&l2hdr
->b_dev
->l2ad_mtx
);
1857 l2hdr
= &hdr
->b_l2hdr
;
1860 list_remove(&l2hdr
->b_dev
->l2ad_buflist
, hdr
);
1863 * We don't want to leak the b_tmp_cdata buffer that was
1864 * allocated in l2arc_write_buffers()
1866 arc_buf_l2_cdata_free(hdr
);
1868 arc_space_return(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
1869 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
1870 ARCSTAT_INCR(arcstat_l2_asize
, -l2hdr
->b_asize
);
1873 mutex_exit(&l2hdr
->b_dev
->l2ad_mtx
);
1875 hdr
->b_flags
&= ~ARC_FLAG_HAS_L2HDR
;
1878 if (!BUF_EMPTY(hdr
))
1879 buf_discard_identity(hdr
);
1881 if (hdr
->b_freeze_cksum
!= NULL
) {
1882 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
1883 hdr
->b_freeze_cksum
= NULL
;
1886 if (HDR_HAS_L1HDR(hdr
)) {
1887 while (hdr
->b_l1hdr
.b_buf
) {
1888 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
1890 if (buf
->b_efunc
!= NULL
) {
1891 mutex_enter(&arc_user_evicts_lock
);
1892 mutex_enter(&buf
->b_evict_lock
);
1893 ASSERT(buf
->b_hdr
!= NULL
);
1894 arc_buf_destroy(hdr
->b_l1hdr
.b_buf
, FALSE
);
1895 hdr
->b_l1hdr
.b_buf
= buf
->b_next
;
1896 buf
->b_hdr
= &arc_eviction_hdr
;
1897 buf
->b_next
= arc_eviction_list
;
1898 arc_eviction_list
= buf
;
1899 mutex_exit(&buf
->b_evict_lock
);
1900 cv_signal(&arc_user_evicts_cv
);
1901 mutex_exit(&arc_user_evicts_lock
);
1903 arc_buf_destroy(hdr
->b_l1hdr
.b_buf
, TRUE
);
1908 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
1909 if (HDR_HAS_L1HDR(hdr
)) {
1910 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
1911 ASSERT3P(hdr
->b_l1hdr
.b_acb
, ==, NULL
);
1912 kmem_cache_free(hdr_full_cache
, hdr
);
1914 kmem_cache_free(hdr_l2only_cache
, hdr
);
1919 arc_buf_free(arc_buf_t
*buf
, void *tag
)
1921 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1922 int hashed
= hdr
->b_l1hdr
.b_state
!= arc_anon
;
1924 ASSERT(buf
->b_efunc
== NULL
);
1925 ASSERT(buf
->b_data
!= NULL
);
1928 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
1930 mutex_enter(hash_lock
);
1932 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1934 (void) remove_reference(hdr
, hash_lock
, tag
);
1935 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
1936 arc_buf_destroy(buf
, TRUE
);
1938 ASSERT(buf
== hdr
->b_l1hdr
.b_buf
);
1939 ASSERT(buf
->b_efunc
== NULL
);
1940 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
1942 mutex_exit(hash_lock
);
1943 } else if (HDR_IO_IN_PROGRESS(hdr
)) {
1946 * We are in the middle of an async write. Don't destroy
1947 * this buffer unless the write completes before we finish
1948 * decrementing the reference count.
1950 mutex_enter(&arc_user_evicts_lock
);
1951 (void) remove_reference(hdr
, NULL
, tag
);
1952 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
1953 destroy_hdr
= !HDR_IO_IN_PROGRESS(hdr
);
1954 mutex_exit(&arc_user_evicts_lock
);
1956 arc_hdr_destroy(hdr
);
1958 if (remove_reference(hdr
, NULL
, tag
) > 0)
1959 arc_buf_destroy(buf
, TRUE
);
1961 arc_hdr_destroy(hdr
);
1966 arc_buf_remove_ref(arc_buf_t
*buf
, void* tag
)
1968 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1969 kmutex_t
*hash_lock
= NULL
;
1970 boolean_t no_callback
= (buf
->b_efunc
== NULL
);
1972 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
1973 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
1974 arc_buf_free(buf
, tag
);
1975 return (no_callback
);
1978 hash_lock
= HDR_LOCK(hdr
);
1979 mutex_enter(hash_lock
);
1981 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
1982 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1983 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_anon
);
1984 ASSERT(buf
->b_data
!= NULL
);
1986 (void) remove_reference(hdr
, hash_lock
, tag
);
1987 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
1989 arc_buf_destroy(buf
, TRUE
);
1990 } else if (no_callback
) {
1991 ASSERT(hdr
->b_l1hdr
.b_buf
== buf
&& buf
->b_next
== NULL
);
1992 ASSERT(buf
->b_efunc
== NULL
);
1993 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
1995 ASSERT(no_callback
|| hdr
->b_l1hdr
.b_datacnt
> 1 ||
1996 refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
1997 mutex_exit(hash_lock
);
1998 return (no_callback
);
2002 arc_buf_size(arc_buf_t
*buf
)
2004 return (buf
->b_hdr
->b_size
);
2008 * Called from the DMU to determine if the current buffer should be
2009 * evicted. In order to ensure proper locking, the eviction must be initiated
2010 * from the DMU. Return true if the buffer is associated with user data and
2011 * duplicate buffers still exist.
2014 arc_buf_eviction_needed(arc_buf_t
*buf
)
2017 boolean_t evict_needed
= B_FALSE
;
2019 if (zfs_disable_dup_eviction
)
2022 mutex_enter(&buf
->b_evict_lock
);
2026 * We are in arc_do_user_evicts(); let that function
2027 * perform the eviction.
2029 ASSERT(buf
->b_data
== NULL
);
2030 mutex_exit(&buf
->b_evict_lock
);
2032 } else if (buf
->b_data
== NULL
) {
2034 * We have already been added to the arc eviction list;
2035 * recommend eviction.
2037 ASSERT3P(hdr
, ==, &arc_eviction_hdr
);
2038 mutex_exit(&buf
->b_evict_lock
);
2042 if (hdr
->b_l1hdr
.b_datacnt
> 1 && HDR_ISTYPE_DATA(hdr
))
2043 evict_needed
= B_TRUE
;
2045 mutex_exit(&buf
->b_evict_lock
);
2046 return (evict_needed
);
2050 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2051 * state of the header is dependent on its state prior to entering this
2052 * function. The following transitions are possible:
2054 * - arc_mru -> arc_mru_ghost
2055 * - arc_mfu -> arc_mfu_ghost
2056 * - arc_mru_ghost -> arc_l2c_only
2057 * - arc_mru_ghost -> deleted
2058 * - arc_mfu_ghost -> arc_l2c_only
2059 * - arc_mfu_ghost -> deleted
2062 arc_evict_hdr(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
2064 arc_state_t
*evicted_state
, *state
;
2065 int64_t bytes_evicted
= 0;
2067 ASSERT(MUTEX_HELD(hash_lock
));
2068 ASSERT(HDR_HAS_L1HDR(hdr
));
2070 state
= hdr
->b_l1hdr
.b_state
;
2071 if (GHOST_STATE(state
)) {
2072 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
2073 ASSERT(hdr
->b_l1hdr
.b_buf
== NULL
);
2076 * l2arc_write_buffers() relies on a header's L1 portion
2077 * (i.e. its b_tmp_cdata field) during its write phase.
2078 * Thus, we cannot push a header onto the arc_l2c_only
2079 * state (removing its L1 piece) until the header is
2080 * done being written to the l2arc.
2082 if (HDR_HAS_L2HDR(hdr
) && HDR_L2_WRITING(hdr
)) {
2083 ARCSTAT_BUMP(arcstat_evict_l2_skip
);
2084 return (bytes_evicted
);
2087 ARCSTAT_BUMP(arcstat_deleted
);
2088 bytes_evicted
+= hdr
->b_size
;
2090 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, hdr
);
2092 if (HDR_HAS_L2HDR(hdr
)) {
2094 * This buffer is cached on the 2nd Level ARC;
2095 * don't destroy the header.
2097 arc_change_state(arc_l2c_only
, hdr
, hash_lock
);
2099 * dropping from L1+L2 cached to L2-only,
2100 * realloc to remove the L1 header.
2102 hdr
= arc_hdr_realloc(hdr
, hdr_full_cache
,
2105 arc_change_state(arc_anon
, hdr
, hash_lock
);
2106 arc_hdr_destroy(hdr
);
2108 return (bytes_evicted
);
2111 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
2112 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
2114 /* prefetch buffers have a minimum lifespan */
2115 if (HDR_IO_IN_PROGRESS(hdr
) ||
2116 ((hdr
->b_flags
& (ARC_FLAG_PREFETCH
| ARC_FLAG_INDIRECT
)) &&
2117 ddi_get_lbolt() - hdr
->b_l1hdr
.b_arc_access
<
2118 arc_min_prefetch_lifespan
)) {
2119 ARCSTAT_BUMP(arcstat_evict_skip
);
2120 return (bytes_evicted
);
2123 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
2124 ASSERT3U(hdr
->b_l1hdr
.b_datacnt
, >, 0);
2125 while (hdr
->b_l1hdr
.b_buf
) {
2126 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
2127 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
2128 ARCSTAT_BUMP(arcstat_mutex_miss
);
2131 if (buf
->b_data
!= NULL
)
2132 bytes_evicted
+= hdr
->b_size
;
2133 if (buf
->b_efunc
!= NULL
) {
2134 mutex_enter(&arc_user_evicts_lock
);
2135 arc_buf_destroy(buf
, FALSE
);
2136 hdr
->b_l1hdr
.b_buf
= buf
->b_next
;
2137 buf
->b_hdr
= &arc_eviction_hdr
;
2138 buf
->b_next
= arc_eviction_list
;
2139 arc_eviction_list
= buf
;
2140 cv_signal(&arc_user_evicts_cv
);
2141 mutex_exit(&arc_user_evicts_lock
);
2142 mutex_exit(&buf
->b_evict_lock
);
2144 mutex_exit(&buf
->b_evict_lock
);
2145 arc_buf_destroy(buf
, TRUE
);
2149 if (HDR_HAS_L2HDR(hdr
)) {
2150 ARCSTAT_INCR(arcstat_evict_l2_cached
, hdr
->b_size
);
2152 if (l2arc_write_eligible(hdr
->b_spa
, hdr
))
2153 ARCSTAT_INCR(arcstat_evict_l2_eligible
, hdr
->b_size
);
2155 ARCSTAT_INCR(arcstat_evict_l2_ineligible
, hdr
->b_size
);
2158 if (hdr
->b_l1hdr
.b_datacnt
== 0) {
2159 arc_change_state(evicted_state
, hdr
, hash_lock
);
2160 ASSERT(HDR_IN_HASH_TABLE(hdr
));
2161 hdr
->b_flags
|= ARC_FLAG_IN_HASH_TABLE
;
2162 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
2163 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, hdr
);
2166 return (bytes_evicted
);
2170 arc_evict_state_impl(multilist_t
*ml
, int idx
, arc_buf_hdr_t
*marker
,
2171 uint64_t spa
, int64_t bytes
)
2173 multilist_sublist_t
*mls
;
2174 uint64_t bytes_evicted
= 0;
2176 kmutex_t
*hash_lock
;
2177 int evict_count
= 0;
2179 ASSERT3P(marker
, !=, NULL
);
2180 ASSERTV(if (bytes
< 0) ASSERT(bytes
== ARC_EVICT_ALL
));
2182 mls
= multilist_sublist_lock(ml
, idx
);
2184 for (hdr
= multilist_sublist_prev(mls
, marker
); hdr
!= NULL
;
2185 hdr
= multilist_sublist_prev(mls
, marker
)) {
2186 if ((bytes
!= ARC_EVICT_ALL
&& bytes_evicted
>= bytes
) ||
2187 (evict_count
>= zfs_arc_evict_batch_limit
))
2191 * To keep our iteration location, move the marker
2192 * forward. Since we're not holding hdr's hash lock, we
2193 * must be very careful and not remove 'hdr' from the
2194 * sublist. Otherwise, other consumers might mistake the
2195 * 'hdr' as not being on a sublist when they call the
2196 * multilist_link_active() function (they all rely on
2197 * the hash lock protecting concurrent insertions and
2198 * removals). multilist_sublist_move_forward() was
2199 * specifically implemented to ensure this is the case
2200 * (only 'marker' will be removed and re-inserted).
2202 multilist_sublist_move_forward(mls
, marker
);
2205 * The only case where the b_spa field should ever be
2206 * zero, is the marker headers inserted by
2207 * arc_evict_state(). It's possible for multiple threads
2208 * to be calling arc_evict_state() concurrently (e.g.
2209 * dsl_pool_close() and zio_inject_fault()), so we must
2210 * skip any markers we see from these other threads.
2212 if (hdr
->b_spa
== 0)
2215 /* we're only interested in evicting buffers of a certain spa */
2216 if (spa
!= 0 && hdr
->b_spa
!= spa
) {
2217 ARCSTAT_BUMP(arcstat_evict_skip
);
2221 hash_lock
= HDR_LOCK(hdr
);
2224 * We aren't calling this function from any code path
2225 * that would already be holding a hash lock, so we're
2226 * asserting on this assumption to be defensive in case
2227 * this ever changes. Without this check, it would be
2228 * possible to incorrectly increment arcstat_mutex_miss
2229 * below (e.g. if the code changed such that we called
2230 * this function with a hash lock held).
2232 ASSERT(!MUTEX_HELD(hash_lock
));
2234 if (mutex_tryenter(hash_lock
)) {
2235 uint64_t evicted
= arc_evict_hdr(hdr
, hash_lock
);
2236 mutex_exit(hash_lock
);
2238 bytes_evicted
+= evicted
;
2241 * If evicted is zero, arc_evict_hdr() must have
2242 * decided to skip this header, don't increment
2243 * evict_count in this case.
2249 * If arc_size isn't overflowing, signal any
2250 * threads that might happen to be waiting.
2252 * For each header evicted, we wake up a single
2253 * thread. If we used cv_broadcast, we could
2254 * wake up "too many" threads causing arc_size
2255 * to significantly overflow arc_c; since
2256 * arc_get_data_buf() doesn't check for overflow
2257 * when it's woken up (it doesn't because it's
2258 * possible for the ARC to be overflowing while
2259 * full of un-evictable buffers, and the
2260 * function should proceed in this case).
2262 * If threads are left sleeping, due to not
2263 * using cv_broadcast, they will be woken up
2264 * just before arc_reclaim_thread() sleeps.
2266 mutex_enter(&arc_reclaim_lock
);
2267 if (!arc_is_overflowing())
2268 cv_signal(&arc_reclaim_waiters_cv
);
2269 mutex_exit(&arc_reclaim_lock
);
2271 ARCSTAT_BUMP(arcstat_mutex_miss
);
2275 multilist_sublist_unlock(mls
);
2277 return (bytes_evicted
);
2281 * Evict buffers from the given arc state, until we've removed the
2282 * specified number of bytes. Move the removed buffers to the
2283 * appropriate evict state.
2285 * This function makes a "best effort". It skips over any buffers
2286 * it can't get a hash_lock on, and so, may not catch all candidates.
2287 * It may also return without evicting as much space as requested.
2289 * If bytes is specified using the special value ARC_EVICT_ALL, this
2290 * will evict all available (i.e. unlocked and evictable) buffers from
2291 * the given arc state; which is used by arc_flush().
2294 arc_evict_state(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
2295 arc_buf_contents_t type
)
2297 uint64_t total_evicted
= 0;
2298 multilist_t
*ml
= &state
->arcs_list
[type
];
2300 arc_buf_hdr_t
**markers
;
2303 ASSERTV(if (bytes
< 0) ASSERT(bytes
== ARC_EVICT_ALL
));
2305 num_sublists
= multilist_get_num_sublists(ml
);
2308 * If we've tried to evict from each sublist, made some
2309 * progress, but still have not hit the target number of bytes
2310 * to evict, we want to keep trying. The markers allow us to
2311 * pick up where we left off for each individual sublist, rather
2312 * than starting from the tail each time.
2314 markers
= kmem_zalloc(sizeof (*markers
) * num_sublists
, KM_SLEEP
);
2315 for (i
= 0; i
< num_sublists
; i
++) {
2316 multilist_sublist_t
*mls
;
2318 markers
[i
] = kmem_cache_alloc(hdr_full_cache
, KM_SLEEP
);
2321 * A b_spa of 0 is used to indicate that this header is
2322 * a marker. This fact is used in arc_adjust_type() and
2323 * arc_evict_state_impl().
2325 markers
[i
]->b_spa
= 0;
2327 mls
= multilist_sublist_lock(ml
, i
);
2328 multilist_sublist_insert_tail(mls
, markers
[i
]);
2329 multilist_sublist_unlock(mls
);
2333 * While we haven't hit our target number of bytes to evict, or
2334 * we're evicting all available buffers.
2336 while (total_evicted
< bytes
|| bytes
== ARC_EVICT_ALL
) {
2338 * Start eviction using a randomly selected sublist,
2339 * this is to try and evenly balance eviction across all
2340 * sublists. Always starting at the same sublist
2341 * (e.g. index 0) would cause evictions to favor certain
2342 * sublists over others.
2344 int sublist_idx
= multilist_get_random_index(ml
);
2345 uint64_t scan_evicted
= 0;
2347 for (i
= 0; i
< num_sublists
; i
++) {
2348 uint64_t bytes_remaining
;
2349 uint64_t bytes_evicted
;
2351 if (bytes
== ARC_EVICT_ALL
)
2352 bytes_remaining
= ARC_EVICT_ALL
;
2353 else if (total_evicted
< bytes
)
2354 bytes_remaining
= bytes
- total_evicted
;
2358 bytes_evicted
= arc_evict_state_impl(ml
, sublist_idx
,
2359 markers
[sublist_idx
], spa
, bytes_remaining
);
2361 scan_evicted
+= bytes_evicted
;
2362 total_evicted
+= bytes_evicted
;
2364 /* we've reached the end, wrap to the beginning */
2365 if (++sublist_idx
>= num_sublists
)
2370 * If we didn't evict anything during this scan, we have
2371 * no reason to believe we'll evict more during another
2372 * scan, so break the loop.
2374 if (scan_evicted
== 0) {
2375 /* This isn't possible, let's make that obvious */
2376 ASSERT3S(bytes
, !=, 0);
2379 * When bytes is ARC_EVICT_ALL, the only way to
2380 * break the loop is when scan_evicted is zero.
2381 * In that case, we actually have evicted enough,
2382 * so we don't want to increment the kstat.
2384 if (bytes
!= ARC_EVICT_ALL
) {
2385 ASSERT3S(total_evicted
, <, bytes
);
2386 ARCSTAT_BUMP(arcstat_evict_not_enough
);
2393 for (i
= 0; i
< num_sublists
; i
++) {
2394 multilist_sublist_t
*mls
= multilist_sublist_lock(ml
, i
);
2395 multilist_sublist_remove(mls
, markers
[i
]);
2396 multilist_sublist_unlock(mls
);
2398 kmem_cache_free(hdr_full_cache
, markers
[i
]);
2400 kmem_free(markers
, sizeof (*markers
) * num_sublists
);
2402 return (total_evicted
);
2406 * Flush all "evictable" data of the given type from the arc state
2407 * specified. This will not evict any "active" buffers (i.e. referenced).
2409 * When 'retry' is set to FALSE, the function will make a single pass
2410 * over the state and evict any buffers that it can. Since it doesn't
2411 * continually retry the eviction, it might end up leaving some buffers
2412 * in the ARC due to lock misses.
2414 * When 'retry' is set to TRUE, the function will continually retry the
2415 * eviction until *all* evictable buffers have been removed from the
2416 * state. As a result, if concurrent insertions into the state are
2417 * allowed (e.g. if the ARC isn't shutting down), this function might
2418 * wind up in an infinite loop, continually trying to evict buffers.
2421 arc_flush_state(arc_state_t
*state
, uint64_t spa
, arc_buf_contents_t type
,
2424 uint64_t evicted
= 0;
2426 while (state
->arcs_lsize
[type
] != 0) {
2427 evicted
+= arc_evict_state(state
, spa
, ARC_EVICT_ALL
, type
);
2437 * Helper function for arc_prune() it is responsible for safely handling
2438 * the execution of a registered arc_prune_func_t.
2441 arc_prune_task(void *ptr
)
2443 arc_prune_t
*ap
= (arc_prune_t
*)ptr
;
2444 arc_prune_func_t
*func
= ap
->p_pfunc
;
2447 func(ap
->p_adjust
, ap
->p_private
);
2449 /* Callback unregistered concurrently with execution */
2450 if (refcount_remove(&ap
->p_refcnt
, func
) == 0) {
2451 ASSERT(!list_link_active(&ap
->p_node
));
2452 refcount_destroy(&ap
->p_refcnt
);
2453 kmem_free(ap
, sizeof (*ap
));
2458 * Notify registered consumers they must drop holds on a portion of the ARC
2459 * buffered they reference. This provides a mechanism to ensure the ARC can
2460 * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This
2461 * is analogous to dnlc_reduce_cache() but more generic.
2463 * This operation is performed asyncronously so it may be safely called
2464 * in the context of the arc_adapt_thread(). A reference is taken here
2465 * for each registered arc_prune_t and the arc_prune_task() is responsible
2466 * for releasing it once the registered arc_prune_func_t has completed.
2469 arc_prune_async(int64_t adjust
)
2473 mutex_enter(&arc_prune_mtx
);
2474 for (ap
= list_head(&arc_prune_list
); ap
!= NULL
;
2475 ap
= list_next(&arc_prune_list
, ap
)) {
2477 if (refcount_count(&ap
->p_refcnt
) >= 2)
2480 refcount_add(&ap
->p_refcnt
, ap
->p_pfunc
);
2481 ap
->p_adjust
= adjust
;
2482 taskq_dispatch(arc_prune_taskq
, arc_prune_task
, ap
, TQ_SLEEP
);
2483 ARCSTAT_BUMP(arcstat_prune
);
2485 mutex_exit(&arc_prune_mtx
);
2489 arc_prune(int64_t adjust
)
2491 arc_prune_async(adjust
);
2492 taskq_wait_outstanding(arc_prune_taskq
, 0);
2496 * Evict the specified number of bytes from the state specified,
2497 * restricting eviction to the spa and type given. This function
2498 * prevents us from trying to evict more from a state's list than
2499 * is "evictable", and to skip evicting altogether when passed a
2500 * negative value for "bytes". In contrast, arc_evict_state() will
2501 * evict everything it can, when passed a negative value for "bytes".
2504 arc_adjust_impl(arc_state_t
*state
, uint64_t spa
, int64_t bytes
,
2505 arc_buf_contents_t type
)
2509 if (bytes
> 0 && state
->arcs_lsize
[type
] > 0) {
2510 delta
= MIN(state
->arcs_lsize
[type
], bytes
);
2511 return (arc_evict_state(state
, spa
, delta
, type
));
2518 * The goal of this function is to evict enough meta data buffers from the
2519 * ARC in order to enforce the arc_meta_limit. Achieving this is slightly
2520 * more complicated than it appears because it is common for data buffers
2521 * to have holds on meta data buffers. In addition, dnode meta data buffers
2522 * will be held by the dnodes in the block preventing them from being freed.
2523 * This means we can't simply traverse the ARC and expect to always find
2524 * enough unheld meta data buffer to release.
2526 * Therefore, this function has been updated to make alternating passes
2527 * over the ARC releasing data buffers and then newly unheld meta data
2528 * buffers. This ensures forward progress is maintained and arc_meta_used
2529 * will decrease. Normally this is sufficient, but if required the ARC
2530 * will call the registered prune callbacks causing dentry and inodes to
2531 * be dropped from the VFS cache. This will make dnode meta data buffers
2532 * available for reclaim.
2535 arc_adjust_meta_balanced(void)
2537 int64_t adjustmnt
, delta
, prune
= 0;
2538 uint64_t total_evicted
= 0;
2539 arc_buf_contents_t type
= ARC_BUFC_DATA
;
2540 unsigned long restarts
= zfs_arc_meta_adjust_restarts
;
2544 * This slightly differs than the way we evict from the mru in
2545 * arc_adjust because we don't have a "target" value (i.e. no
2546 * "meta" arc_p). As a result, I think we can completely
2547 * cannibalize the metadata in the MRU before we evict the
2548 * metadata from the MFU. I think we probably need to implement a
2549 * "metadata arc_p" value to do this properly.
2551 adjustmnt
= arc_meta_used
- arc_meta_limit
;
2553 if (adjustmnt
> 0 && arc_mru
->arcs_lsize
[type
] > 0) {
2554 delta
= MIN(arc_mru
->arcs_lsize
[type
], adjustmnt
);
2555 total_evicted
+= arc_adjust_impl(arc_mru
, 0, delta
, type
);
2560 * We can't afford to recalculate adjustmnt here. If we do,
2561 * new metadata buffers can sneak into the MRU or ANON lists,
2562 * thus penalize the MFU metadata. Although the fudge factor is
2563 * small, it has been empirically shown to be significant for
2564 * certain workloads (e.g. creating many empty directories). As
2565 * such, we use the original calculation for adjustmnt, and
2566 * simply decrement the amount of data evicted from the MRU.
2569 if (adjustmnt
> 0 && arc_mfu
->arcs_lsize
[type
] > 0) {
2570 delta
= MIN(arc_mfu
->arcs_lsize
[type
], adjustmnt
);
2571 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, delta
, type
);
2574 adjustmnt
= arc_meta_used
- arc_meta_limit
;
2576 if (adjustmnt
> 0 && arc_mru_ghost
->arcs_lsize
[type
] > 0) {
2577 delta
= MIN(adjustmnt
,
2578 arc_mru_ghost
->arcs_lsize
[type
]);
2579 total_evicted
+= arc_adjust_impl(arc_mru_ghost
, 0, delta
, type
);
2583 if (adjustmnt
> 0 && arc_mfu_ghost
->arcs_lsize
[type
] > 0) {
2584 delta
= MIN(adjustmnt
,
2585 arc_mfu_ghost
->arcs_lsize
[type
]);
2586 total_evicted
+= arc_adjust_impl(arc_mfu_ghost
, 0, delta
, type
);
2590 * If after attempting to make the requested adjustment to the ARC
2591 * the meta limit is still being exceeded then request that the
2592 * higher layers drop some cached objects which have holds on ARC
2593 * meta buffers. Requests to the upper layers will be made with
2594 * increasingly large scan sizes until the ARC is below the limit.
2596 if (arc_meta_used
> arc_meta_limit
) {
2597 if (type
== ARC_BUFC_DATA
) {
2598 type
= ARC_BUFC_METADATA
;
2600 type
= ARC_BUFC_DATA
;
2602 if (zfs_arc_meta_prune
) {
2603 prune
+= zfs_arc_meta_prune
;
2604 arc_prune_async(prune
);
2613 return (total_evicted
);
2617 * Evict metadata buffers from the cache, such that arc_meta_used is
2618 * capped by the arc_meta_limit tunable.
2621 arc_adjust_meta_only(void)
2623 uint64_t total_evicted
= 0;
2627 * If we're over the meta limit, we want to evict enough
2628 * metadata to get back under the meta limit. We don't want to
2629 * evict so much that we drop the MRU below arc_p, though. If
2630 * we're over the meta limit more than we're over arc_p, we
2631 * evict some from the MRU here, and some from the MFU below.
2633 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
2634 (int64_t)(arc_anon
->arcs_size
+ arc_mru
->arcs_size
- arc_p
));
2636 total_evicted
+= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
2639 * Similar to the above, we want to evict enough bytes to get us
2640 * below the meta limit, but not so much as to drop us below the
2641 * space alloted to the MFU (which is defined as arc_c - arc_p).
2643 target
= MIN((int64_t)(arc_meta_used
- arc_meta_limit
),
2644 (int64_t)(arc_mfu
->arcs_size
- (arc_c
- arc_p
)));
2646 total_evicted
+= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
2648 return (total_evicted
);
2652 arc_adjust_meta(void)
2654 if (zfs_arc_meta_strategy
== ARC_STRATEGY_META_ONLY
)
2655 return (arc_adjust_meta_only());
2657 return (arc_adjust_meta_balanced());
2661 * Return the type of the oldest buffer in the given arc state
2663 * This function will select a random sublist of type ARC_BUFC_DATA and
2664 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
2665 * is compared, and the type which contains the "older" buffer will be
2668 static arc_buf_contents_t
2669 arc_adjust_type(arc_state_t
*state
)
2671 multilist_t
*data_ml
= &state
->arcs_list
[ARC_BUFC_DATA
];
2672 multilist_t
*meta_ml
= &state
->arcs_list
[ARC_BUFC_METADATA
];
2673 int data_idx
= multilist_get_random_index(data_ml
);
2674 int meta_idx
= multilist_get_random_index(meta_ml
);
2675 multilist_sublist_t
*data_mls
;
2676 multilist_sublist_t
*meta_mls
;
2677 arc_buf_contents_t type
;
2678 arc_buf_hdr_t
*data_hdr
;
2679 arc_buf_hdr_t
*meta_hdr
;
2682 * We keep the sublist lock until we're finished, to prevent
2683 * the headers from being destroyed via arc_evict_state().
2685 data_mls
= multilist_sublist_lock(data_ml
, data_idx
);
2686 meta_mls
= multilist_sublist_lock(meta_ml
, meta_idx
);
2689 * These two loops are to ensure we skip any markers that
2690 * might be at the tail of the lists due to arc_evict_state().
2693 for (data_hdr
= multilist_sublist_tail(data_mls
); data_hdr
!= NULL
;
2694 data_hdr
= multilist_sublist_prev(data_mls
, data_hdr
)) {
2695 if (data_hdr
->b_spa
!= 0)
2699 for (meta_hdr
= multilist_sublist_tail(meta_mls
); meta_hdr
!= NULL
;
2700 meta_hdr
= multilist_sublist_prev(meta_mls
, meta_hdr
)) {
2701 if (meta_hdr
->b_spa
!= 0)
2705 if (data_hdr
== NULL
&& meta_hdr
== NULL
) {
2706 type
= ARC_BUFC_DATA
;
2707 } else if (data_hdr
== NULL
) {
2708 ASSERT3P(meta_hdr
, !=, NULL
);
2709 type
= ARC_BUFC_METADATA
;
2710 } else if (meta_hdr
== NULL
) {
2711 ASSERT3P(data_hdr
, !=, NULL
);
2712 type
= ARC_BUFC_DATA
;
2714 ASSERT3P(data_hdr
, !=, NULL
);
2715 ASSERT3P(meta_hdr
, !=, NULL
);
2717 /* The headers can't be on the sublist without an L1 header */
2718 ASSERT(HDR_HAS_L1HDR(data_hdr
));
2719 ASSERT(HDR_HAS_L1HDR(meta_hdr
));
2721 if (data_hdr
->b_l1hdr
.b_arc_access
<
2722 meta_hdr
->b_l1hdr
.b_arc_access
) {
2723 type
= ARC_BUFC_DATA
;
2725 type
= ARC_BUFC_METADATA
;
2729 multilist_sublist_unlock(meta_mls
);
2730 multilist_sublist_unlock(data_mls
);
2736 * Evict buffers from the cache, such that arc_size is capped by arc_c.
2741 uint64_t total_evicted
= 0;
2746 * If we're over arc_meta_limit, we want to correct that before
2747 * potentially evicting data buffers below.
2749 total_evicted
+= arc_adjust_meta();
2754 * If we're over the target cache size, we want to evict enough
2755 * from the list to get back to our target size. We don't want
2756 * to evict too much from the MRU, such that it drops below
2757 * arc_p. So, if we're over our target cache size more than
2758 * the MRU is over arc_p, we'll evict enough to get back to
2759 * arc_p here, and then evict more from the MFU below.
2761 target
= MIN((int64_t)(arc_size
- arc_c
),
2762 (int64_t)(arc_anon
->arcs_size
+ arc_mru
->arcs_size
+ arc_meta_used
-
2766 * If we're below arc_meta_min, always prefer to evict data.
2767 * Otherwise, try to satisfy the requested number of bytes to
2768 * evict from the type which contains older buffers; in an
2769 * effort to keep newer buffers in the cache regardless of their
2770 * type. If we cannot satisfy the number of bytes from this
2771 * type, spill over into the next type.
2773 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
2774 arc_meta_used
> arc_meta_min
) {
2775 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
2776 total_evicted
+= bytes
;
2779 * If we couldn't evict our target number of bytes from
2780 * metadata, we try to get the rest from data.
2785 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
2787 bytes
= arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_DATA
);
2788 total_evicted
+= bytes
;
2791 * If we couldn't evict our target number of bytes from
2792 * data, we try to get the rest from metadata.
2797 arc_adjust_impl(arc_mru
, 0, target
, ARC_BUFC_METADATA
);
2803 * Now that we've tried to evict enough from the MRU to get its
2804 * size back to arc_p, if we're still above the target cache
2805 * size, we evict the rest from the MFU.
2807 target
= arc_size
- arc_c
;
2809 if (arc_adjust_type(arc_mru
) == ARC_BUFC_METADATA
&&
2810 arc_meta_used
> arc_meta_min
) {
2811 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
2812 total_evicted
+= bytes
;
2815 * If we couldn't evict our target number of bytes from
2816 * metadata, we try to get the rest from data.
2821 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
2823 bytes
= arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_DATA
);
2824 total_evicted
+= bytes
;
2827 * If we couldn't evict our target number of bytes from
2828 * data, we try to get the rest from data.
2833 arc_adjust_impl(arc_mfu
, 0, target
, ARC_BUFC_METADATA
);
2837 * Adjust ghost lists
2839 * In addition to the above, the ARC also defines target values
2840 * for the ghost lists. The sum of the mru list and mru ghost
2841 * list should never exceed the target size of the cache, and
2842 * the sum of the mru list, mfu list, mru ghost list, and mfu
2843 * ghost list should never exceed twice the target size of the
2844 * cache. The following logic enforces these limits on the ghost
2845 * caches, and evicts from them as needed.
2847 target
= arc_mru
->arcs_size
+ arc_mru_ghost
->arcs_size
- arc_c
;
2849 bytes
= arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_DATA
);
2850 total_evicted
+= bytes
;
2855 arc_adjust_impl(arc_mru_ghost
, 0, target
, ARC_BUFC_METADATA
);
2858 * We assume the sum of the mru list and mfu list is less than
2859 * or equal to arc_c (we enforced this above), which means we
2860 * can use the simpler of the two equations below:
2862 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
2863 * mru ghost + mfu ghost <= arc_c
2865 target
= arc_mru_ghost
->arcs_size
+ arc_mfu_ghost
->arcs_size
- arc_c
;
2867 bytes
= arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_DATA
);
2868 total_evicted
+= bytes
;
2873 arc_adjust_impl(arc_mfu_ghost
, 0, target
, ARC_BUFC_METADATA
);
2875 return (total_evicted
);
2879 arc_do_user_evicts(void)
2881 mutex_enter(&arc_user_evicts_lock
);
2882 while (arc_eviction_list
!= NULL
) {
2883 arc_buf_t
*buf
= arc_eviction_list
;
2884 arc_eviction_list
= buf
->b_next
;
2885 mutex_enter(&buf
->b_evict_lock
);
2887 mutex_exit(&buf
->b_evict_lock
);
2888 mutex_exit(&arc_user_evicts_lock
);
2890 if (buf
->b_efunc
!= NULL
)
2891 VERIFY0(buf
->b_efunc(buf
->b_private
));
2893 buf
->b_efunc
= NULL
;
2894 buf
->b_private
= NULL
;
2895 kmem_cache_free(buf_cache
, buf
);
2896 mutex_enter(&arc_user_evicts_lock
);
2898 mutex_exit(&arc_user_evicts_lock
);
2902 arc_flush(spa_t
*spa
, boolean_t retry
)
2907 * If retry is TRUE, a spa must not be specified since we have
2908 * no good way to determine if all of a spa's buffers have been
2909 * evicted from an arc state.
2911 ASSERT(!retry
|| spa
== 0);
2914 guid
= spa_load_guid(spa
);
2916 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_DATA
, retry
);
2917 (void) arc_flush_state(arc_mru
, guid
, ARC_BUFC_METADATA
, retry
);
2919 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_DATA
, retry
);
2920 (void) arc_flush_state(arc_mfu
, guid
, ARC_BUFC_METADATA
, retry
);
2922 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_DATA
, retry
);
2923 (void) arc_flush_state(arc_mru_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
2925 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_DATA
, retry
);
2926 (void) arc_flush_state(arc_mfu_ghost
, guid
, ARC_BUFC_METADATA
, retry
);
2928 arc_do_user_evicts();
2929 ASSERT(spa
|| arc_eviction_list
== NULL
);
2933 arc_shrink(uint64_t bytes
)
2935 if (arc_c
> arc_c_min
) {
2938 to_free
= bytes
? bytes
: arc_c
>> zfs_arc_shrink_shift
;
2940 if (arc_c
> arc_c_min
+ to_free
)
2941 atomic_add_64(&arc_c
, -to_free
);
2945 to_free
= bytes
? bytes
: arc_p
>> zfs_arc_shrink_shift
;
2947 if (arc_p
> to_free
)
2948 atomic_add_64(&arc_p
, -to_free
);
2952 if (arc_c
> arc_size
)
2953 arc_c
= MAX(arc_size
, arc_c_min
);
2955 arc_p
= (arc_c
>> 1);
2956 ASSERT(arc_c
>= arc_c_min
);
2957 ASSERT((int64_t)arc_p
>= 0);
2960 if (arc_size
> arc_c
)
2961 (void) arc_adjust();
2965 arc_kmem_reap_now(arc_reclaim_strategy_t strat
, uint64_t bytes
)
2968 kmem_cache_t
*prev_cache
= NULL
;
2969 kmem_cache_t
*prev_data_cache
= NULL
;
2970 extern kmem_cache_t
*zio_buf_cache
[];
2971 extern kmem_cache_t
*zio_data_buf_cache
[];
2973 if ((arc_meta_used
>= arc_meta_limit
) && zfs_arc_meta_prune
) {
2975 * We are exceeding our meta-data cache limit.
2976 * Prune some entries to release holds on meta-data.
2978 arc_prune(zfs_arc_meta_prune
);
2982 * An aggressive reclamation will shrink the cache size as well as
2983 * reap free buffers from the arc kmem caches.
2985 if (strat
== ARC_RECLAIM_AGGR
)
2988 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
2989 if (zio_buf_cache
[i
] != prev_cache
) {
2990 prev_cache
= zio_buf_cache
[i
];
2991 kmem_cache_reap_now(zio_buf_cache
[i
]);
2993 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
2994 prev_data_cache
= zio_data_buf_cache
[i
];
2995 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
2999 kmem_cache_reap_now(buf_cache
);
3000 kmem_cache_reap_now(hdr_full_cache
);
3001 kmem_cache_reap_now(hdr_l2only_cache
);
3005 * Threads can block in arc_get_data_buf() waiting for this thread to evict
3006 * enough data and signal them to proceed. When this happens, the threads in
3007 * arc_get_data_buf() are sleeping while holding the hash lock for their
3008 * particular arc header. Thus, we must be careful to never sleep on a
3009 * hash lock in this thread. This is to prevent the following deadlock:
3011 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3012 * waiting for the reclaim thread to signal it.
3014 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3015 * fails, and goes to sleep forever.
3017 * This possible deadlock is avoided by always acquiring a hash lock
3018 * using mutex_tryenter() from arc_reclaim_thread().
3021 arc_adapt_thread(void)
3024 fstrans_cookie_t cookie
;
3025 uint64_t arc_evicted
;
3027 CALLB_CPR_INIT(&cpr
, &arc_reclaim_lock
, callb_generic_cpr
, FTAG
);
3029 cookie
= spl_fstrans_mark();
3030 mutex_enter(&arc_reclaim_lock
);
3031 while (arc_reclaim_thread_exit
== 0) {
3033 arc_reclaim_strategy_t last_reclaim
= ARC_RECLAIM_CONS
;
3035 mutex_exit(&arc_reclaim_lock
);
3036 if (spa_get_random(100) == 0) {
3039 if (last_reclaim
== ARC_RECLAIM_CONS
) {
3040 last_reclaim
= ARC_RECLAIM_AGGR
;
3042 last_reclaim
= ARC_RECLAIM_CONS
;
3046 last_reclaim
= ARC_RECLAIM_AGGR
;
3050 /* reset the growth delay for every reclaim */
3051 arc_grow_time
= ddi_get_lbolt() +
3052 (zfs_arc_grow_retry
* hz
);
3054 arc_kmem_reap_now(last_reclaim
, 0);
3058 mutex_exit(&arc_reclaim_lock
);
3059 #endif /* !_KERNEL */
3061 /* No recent memory pressure allow the ARC to grow. */
3063 ddi_time_after_eq(ddi_get_lbolt(), arc_grow_time
))
3064 arc_no_grow
= FALSE
;
3066 arc_evicted
= arc_adjust();
3069 * We're either no longer overflowing, or we
3070 * can't evict anything more, so we should wake
3071 * up any threads before we go to sleep.
3073 if (arc_size
<= arc_c
|| arc_evicted
== 0)
3074 cv_broadcast(&arc_reclaim_waiters_cv
);
3076 mutex_enter(&arc_reclaim_lock
);
3078 /* block until needed, or one second, whichever is shorter */
3079 CALLB_CPR_SAFE_BEGIN(&cpr
);
3080 (void) cv_timedwait_interruptible(&arc_reclaim_thread_cv
,
3081 &arc_reclaim_lock
, (ddi_get_lbolt() + hz
));
3082 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_lock
);
3085 /* Allow the module options to be changed */
3086 if (zfs_arc_max
> 64 << 20 &&
3087 zfs_arc_max
< physmem
* PAGESIZE
&&
3088 zfs_arc_max
!= arc_c_max
)
3089 arc_c_max
= zfs_arc_max
;
3091 if (zfs_arc_min
> 0 &&
3092 zfs_arc_min
< arc_c_max
&&
3093 zfs_arc_min
!= arc_c_min
)
3094 arc_c_min
= zfs_arc_min
;
3096 if (zfs_arc_meta_limit
> 0 &&
3097 zfs_arc_meta_limit
<= arc_c_max
&&
3098 zfs_arc_meta_limit
!= arc_meta_limit
)
3099 arc_meta_limit
= zfs_arc_meta_limit
;
3102 arc_reclaim_thread_exit
= 0;
3103 cv_broadcast(&arc_reclaim_thread_cv
);
3104 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_lock */
3105 spl_fstrans_unmark(cookie
);
3110 arc_user_evicts_thread(void)
3113 fstrans_cookie_t cookie
;
3115 CALLB_CPR_INIT(&cpr
, &arc_user_evicts_lock
, callb_generic_cpr
, FTAG
);
3117 cookie
= spl_fstrans_mark();
3118 mutex_enter(&arc_user_evicts_lock
);
3119 while (!arc_user_evicts_thread_exit
) {
3120 mutex_exit(&arc_user_evicts_lock
);
3122 arc_do_user_evicts();
3125 * This is necessary in order for the mdb ::arc dcmd to
3126 * show up to date information. Since the ::arc command
3127 * does not call the kstat's update function, without
3128 * this call, the command may show stale stats for the
3129 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3130 * with this change, the data might be up to 1 second
3131 * out of date; but that should suffice. The arc_state_t
3132 * structures can be queried directly if more accurate
3133 * information is needed.
3135 if (arc_ksp
!= NULL
)
3136 arc_ksp
->ks_update(arc_ksp
, KSTAT_READ
);
3138 mutex_enter(&arc_user_evicts_lock
);
3141 * Block until signaled, or after one second (we need to
3142 * call the arc's kstat update function regularly).
3144 CALLB_CPR_SAFE_BEGIN(&cpr
);
3145 (void) cv_timedwait_interruptible(&arc_user_evicts_cv
,
3146 &arc_user_evicts_lock
, ddi_get_lbolt() + hz
);
3147 CALLB_CPR_SAFE_END(&cpr
, &arc_user_evicts_lock
);
3150 arc_user_evicts_thread_exit
= FALSE
;
3151 cv_broadcast(&arc_user_evicts_cv
);
3152 CALLB_CPR_EXIT(&cpr
); /* drops arc_user_evicts_lock */
3153 spl_fstrans_unmark(cookie
);
3159 * Determine the amount of memory eligible for eviction contained in the
3160 * ARC. All clean data reported by the ghost lists can always be safely
3161 * evicted. Due to arc_c_min, the same does not hold for all clean data
3162 * contained by the regular mru and mfu lists.
3164 * In the case of the regular mru and mfu lists, we need to report as
3165 * much clean data as possible, such that evicting that same reported
3166 * data will not bring arc_size below arc_c_min. Thus, in certain
3167 * circumstances, the total amount of clean data in the mru and mfu
3168 * lists might not actually be evictable.
3170 * The following two distinct cases are accounted for:
3172 * 1. The sum of the amount of dirty data contained by both the mru and
3173 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
3174 * is greater than or equal to arc_c_min.
3175 * (i.e. amount of dirty data >= arc_c_min)
3177 * This is the easy case; all clean data contained by the mru and mfu
3178 * lists is evictable. Evicting all clean data can only drop arc_size
3179 * to the amount of dirty data, which is greater than arc_c_min.
3181 * 2. The sum of the amount of dirty data contained by both the mru and
3182 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
3183 * is less than arc_c_min.
3184 * (i.e. arc_c_min > amount of dirty data)
3186 * 2.1. arc_size is greater than or equal arc_c_min.
3187 * (i.e. arc_size >= arc_c_min > amount of dirty data)
3189 * In this case, not all clean data from the regular mru and mfu
3190 * lists is actually evictable; we must leave enough clean data
3191 * to keep arc_size above arc_c_min. Thus, the maximum amount of
3192 * evictable data from the two lists combined, is exactly the
3193 * difference between arc_size and arc_c_min.
3195 * 2.2. arc_size is less than arc_c_min
3196 * (i.e. arc_c_min > arc_size > amount of dirty data)
3198 * In this case, none of the data contained in the mru and mfu
3199 * lists is evictable, even if it's clean. Since arc_size is
3200 * already below arc_c_min, evicting any more would only
3201 * increase this negative difference.
3204 arc_evictable_memory(void) {
3205 uint64_t arc_clean
=
3206 arc_mru
->arcs_lsize
[ARC_BUFC_DATA
] +
3207 arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] +
3208 arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
] +
3209 arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
];
3210 uint64_t ghost_clean
=
3211 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
3212 arc_mru_ghost
->arcs_lsize
[ARC_BUFC_METADATA
] +
3213 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_DATA
] +
3214 arc_mfu_ghost
->arcs_lsize
[ARC_BUFC_METADATA
];
3215 uint64_t arc_dirty
= MAX((int64_t)arc_size
- (int64_t)arc_clean
, 0);
3217 if (arc_dirty
>= arc_c_min
)
3218 return (ghost_clean
+ arc_clean
);
3220 return (ghost_clean
+ MAX((int64_t)arc_size
- (int64_t)arc_c_min
, 0));
3224 * If sc->nr_to_scan is zero, the caller is requesting a query of the
3225 * number of objects which can potentially be freed. If it is nonzero,
3226 * the request is to free that many objects.
3228 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
3229 * in struct shrinker and also require the shrinker to return the number
3232 * Older kernels require the shrinker to return the number of freeable
3233 * objects following the freeing of nr_to_free.
3235 static spl_shrinker_t
3236 __arc_shrinker_func(struct shrinker
*shrink
, struct shrink_control
*sc
)
3240 /* The arc is considered warm once reclaim has occurred */
3241 if (unlikely(arc_warm
== B_FALSE
))
3244 /* Return the potential number of reclaimable pages */
3245 pages
= btop((int64_t)arc_evictable_memory());
3246 if (sc
->nr_to_scan
== 0)
3249 /* Not allowed to perform filesystem reclaim */
3250 if (!(sc
->gfp_mask
& __GFP_FS
))
3251 return (SHRINK_STOP
);
3253 /* Reclaim in progress */
3254 if (mutex_tryenter(&arc_reclaim_lock
) == 0)
3255 return (SHRINK_STOP
);
3257 mutex_exit(&arc_reclaim_lock
);
3260 * Evict the requested number of pages by shrinking arc_c the
3261 * requested amount. If there is nothing left to evict just
3262 * reap whatever we can from the various arc slabs.
3265 arc_kmem_reap_now(ARC_RECLAIM_AGGR
, ptob(sc
->nr_to_scan
));
3267 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
3268 pages
= MAX(pages
- btop(arc_evictable_memory()), 0);
3270 pages
= btop(arc_evictable_memory());
3273 arc_kmem_reap_now(ARC_RECLAIM_CONS
, ptob(sc
->nr_to_scan
));
3274 pages
= SHRINK_STOP
;
3278 * We've reaped what we can, wake up threads.
3280 cv_broadcast(&arc_reclaim_waiters_cv
);
3283 * When direct reclaim is observed it usually indicates a rapid
3284 * increase in memory pressure. This occurs because the kswapd
3285 * threads were unable to asynchronously keep enough free memory
3286 * available. In this case set arc_no_grow to briefly pause arc
3287 * growth to avoid compounding the memory pressure.
3289 if (current_is_kswapd()) {
3290 ARCSTAT_BUMP(arcstat_memory_indirect_count
);
3292 arc_no_grow
= B_TRUE
;
3293 arc_grow_time
= ddi_get_lbolt() + (zfs_arc_grow_retry
* hz
);
3294 ARCSTAT_BUMP(arcstat_memory_direct_count
);
3299 SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func
);
3301 SPL_SHRINKER_DECLARE(arc_shrinker
, arc_shrinker_func
, DEFAULT_SEEKS
);
3302 #endif /* _KERNEL */
3305 * Adapt arc info given the number of bytes we are trying to add and
3306 * the state that we are comming from. This function is only called
3307 * when we are adding new content to the cache.
3310 arc_adapt(int bytes
, arc_state_t
*state
)
3314 if (state
== arc_l2c_only
)
3319 * Adapt the target size of the MRU list:
3320 * - if we just hit in the MRU ghost list, then increase
3321 * the target size of the MRU list.
3322 * - if we just hit in the MFU ghost list, then increase
3323 * the target size of the MFU list by decreasing the
3324 * target size of the MRU list.
3326 if (state
== arc_mru_ghost
) {
3327 mult
= ((arc_mru_ghost
->arcs_size
>= arc_mfu_ghost
->arcs_size
) ?
3328 1 : (arc_mfu_ghost
->arcs_size
/arc_mru_ghost
->arcs_size
));
3330 if (!zfs_arc_p_dampener_disable
)
3331 mult
= MIN(mult
, 10); /* avoid wild arc_p adjustment */
3333 arc_p
= MIN(arc_c
, arc_p
+ bytes
* mult
);
3334 } else if (state
== arc_mfu_ghost
) {
3337 mult
= ((arc_mfu_ghost
->arcs_size
>= arc_mru_ghost
->arcs_size
) ?
3338 1 : (arc_mru_ghost
->arcs_size
/arc_mfu_ghost
->arcs_size
));
3340 if (!zfs_arc_p_dampener_disable
)
3341 mult
= MIN(mult
, 10);
3343 delta
= MIN(bytes
* mult
, arc_p
);
3344 arc_p
= MAX(0, arc_p
- delta
);
3346 ASSERT((int64_t)arc_p
>= 0);
3351 if (arc_c
>= arc_c_max
)
3355 * If we're within (2 * maxblocksize) bytes of the target
3356 * cache size, increment the target cache size
3358 if (arc_size
> arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
3359 atomic_add_64(&arc_c
, (int64_t)bytes
);
3360 if (arc_c
> arc_c_max
)
3362 else if (state
== arc_anon
)
3363 atomic_add_64(&arc_p
, (int64_t)bytes
);
3367 ASSERT((int64_t)arc_p
>= 0);
3371 * Check if arc_size has grown past our upper threshold, determined by
3372 * zfs_arc_overflow_shift.
3375 arc_is_overflowing(void)
3377 /* Always allow at least one block of overflow */
3378 uint64_t overflow
= MAX(SPA_MAXBLOCKSIZE
,
3379 arc_c
>> zfs_arc_overflow_shift
);
3381 return (arc_size
>= arc_c
+ overflow
);
3385 * The buffer, supplied as the first argument, needs a data block. If we
3386 * are hitting the hard limit for the cache size, we must sleep, waiting
3387 * for the eviction thread to catch up. If we're past the target size
3388 * but below the hard limit, we'll only signal the reclaim thread and
3392 arc_get_data_buf(arc_buf_t
*buf
)
3394 arc_state_t
*state
= buf
->b_hdr
->b_l1hdr
.b_state
;
3395 uint64_t size
= buf
->b_hdr
->b_size
;
3396 arc_buf_contents_t type
= arc_buf_type(buf
->b_hdr
);
3398 arc_adapt(size
, state
);
3401 * If arc_size is currently overflowing, and has grown past our
3402 * upper limit, we must be adding data faster than the evict
3403 * thread can evict. Thus, to ensure we don't compound the
3404 * problem by adding more data and forcing arc_size to grow even
3405 * further past it's target size, we halt and wait for the
3406 * eviction thread to catch up.
3408 * It's also possible that the reclaim thread is unable to evict
3409 * enough buffers to get arc_size below the overflow limit (e.g.
3410 * due to buffers being un-evictable, or hash lock collisions).
3411 * In this case, we want to proceed regardless if we're
3412 * overflowing; thus we don't use a while loop here.
3414 if (arc_is_overflowing()) {
3415 mutex_enter(&arc_reclaim_lock
);
3418 * Now that we've acquired the lock, we may no longer be
3419 * over the overflow limit, lets check.
3421 * We're ignoring the case of spurious wake ups. If that
3422 * were to happen, it'd let this thread consume an ARC
3423 * buffer before it should have (i.e. before we're under
3424 * the overflow limit and were signalled by the reclaim
3425 * thread). As long as that is a rare occurrence, it
3426 * shouldn't cause any harm.
3428 if (arc_is_overflowing()) {
3429 cv_signal(&arc_reclaim_thread_cv
);
3430 cv_wait(&arc_reclaim_waiters_cv
, &arc_reclaim_lock
);
3433 mutex_exit(&arc_reclaim_lock
);
3436 if (type
== ARC_BUFC_METADATA
) {
3437 buf
->b_data
= zio_buf_alloc(size
);
3438 arc_space_consume(size
, ARC_SPACE_META
);
3440 ASSERT(type
== ARC_BUFC_DATA
);
3441 buf
->b_data
= zio_data_buf_alloc(size
);
3442 arc_space_consume(size
, ARC_SPACE_DATA
);
3446 * Update the state size. Note that ghost states have a
3447 * "ghost size" and so don't need to be updated.
3449 if (!GHOST_STATE(buf
->b_hdr
->b_l1hdr
.b_state
)) {
3450 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3452 atomic_add_64(&hdr
->b_l1hdr
.b_state
->arcs_size
, size
);
3455 * If this is reached via arc_read, the link is
3456 * protected by the hash lock. If reached via
3457 * arc_buf_alloc, the header should not be accessed by
3458 * any other thread. And, if reached via arc_read_done,
3459 * the hash lock will protect it if it's found in the
3460 * hash table; otherwise no other thread should be
3461 * trying to [add|remove]_reference it.
3463 if (multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
)) {
3464 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3465 atomic_add_64(&hdr
->b_l1hdr
.b_state
->arcs_lsize
[type
],
3469 * If we are growing the cache, and we are adding anonymous
3470 * data, and we have outgrown arc_p, update arc_p
3472 if (arc_size
< arc_c
&& hdr
->b_l1hdr
.b_state
== arc_anon
&&
3473 arc_anon
->arcs_size
+ arc_mru
->arcs_size
> arc_p
)
3474 arc_p
= MIN(arc_c
, arc_p
+ size
);
3479 * This routine is called whenever a buffer is accessed.
3480 * NOTE: the hash lock is dropped in this function.
3483 arc_access(arc_buf_hdr_t
*hdr
, kmutex_t
*hash_lock
)
3487 ASSERT(MUTEX_HELD(hash_lock
));
3488 ASSERT(HDR_HAS_L1HDR(hdr
));
3490 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
3492 * This buffer is not in the cache, and does not
3493 * appear in our "ghost" list. Add the new buffer
3497 ASSERT0(hdr
->b_l1hdr
.b_arc_access
);
3498 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3499 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
3500 arc_change_state(arc_mru
, hdr
, hash_lock
);
3502 } else if (hdr
->b_l1hdr
.b_state
== arc_mru
) {
3503 now
= ddi_get_lbolt();
3506 * If this buffer is here because of a prefetch, then either:
3507 * - clear the flag if this is a "referencing" read
3508 * (any subsequent access will bump this into the MFU state).
3510 * - move the buffer to the head of the list if this is
3511 * another prefetch (to make it less likely to be evicted).
3513 if (HDR_PREFETCH(hdr
)) {
3514 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
3515 /* link protected by hash lock */
3516 ASSERT(multilist_link_active(
3517 &hdr
->b_l1hdr
.b_arc_node
));
3519 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
3520 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
3521 ARCSTAT_BUMP(arcstat_mru_hits
);
3523 hdr
->b_l1hdr
.b_arc_access
= now
;
3528 * This buffer has been "accessed" only once so far,
3529 * but it is still in the cache. Move it to the MFU
3532 if (ddi_time_after(now
, hdr
->b_l1hdr
.b_arc_access
+
3535 * More than 125ms have passed since we
3536 * instantiated this buffer. Move it to the
3537 * most frequently used state.
3539 hdr
->b_l1hdr
.b_arc_access
= now
;
3540 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
3541 arc_change_state(arc_mfu
, hdr
, hash_lock
);
3543 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_hits
);
3544 ARCSTAT_BUMP(arcstat_mru_hits
);
3545 } else if (hdr
->b_l1hdr
.b_state
== arc_mru_ghost
) {
3546 arc_state_t
*new_state
;
3548 * This buffer has been "accessed" recently, but
3549 * was evicted from the cache. Move it to the
3553 if (HDR_PREFETCH(hdr
)) {
3554 new_state
= arc_mru
;
3555 if (refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0)
3556 hdr
->b_flags
&= ~ARC_FLAG_PREFETCH
;
3557 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, hdr
);
3559 new_state
= arc_mfu
;
3560 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
3563 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3564 arc_change_state(new_state
, hdr
, hash_lock
);
3566 atomic_inc_32(&hdr
->b_l1hdr
.b_mru_ghost_hits
);
3567 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
3568 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu
) {
3570 * This buffer has been accessed more than once and is
3571 * still in the cache. Keep it in the MFU state.
3573 * NOTE: an add_reference() that occurred when we did
3574 * the arc_read() will have kicked this off the list.
3575 * If it was a prefetch, we will explicitly move it to
3576 * the head of the list now.
3578 if ((HDR_PREFETCH(hdr
)) != 0) {
3579 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3580 /* link protected by hash_lock */
3581 ASSERT(multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
3583 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_hits
);
3584 ARCSTAT_BUMP(arcstat_mfu_hits
);
3585 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3586 } else if (hdr
->b_l1hdr
.b_state
== arc_mfu_ghost
) {
3587 arc_state_t
*new_state
= arc_mfu
;
3589 * This buffer has been accessed more than once but has
3590 * been evicted from the cache. Move it back to the
3594 if (HDR_PREFETCH(hdr
)) {
3596 * This is a prefetch access...
3597 * move this block back to the MRU state.
3599 ASSERT0(refcount_count(&hdr
->b_l1hdr
.b_refcnt
));
3600 new_state
= arc_mru
;
3603 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3604 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
3605 arc_change_state(new_state
, hdr
, hash_lock
);
3607 atomic_inc_32(&hdr
->b_l1hdr
.b_mfu_ghost_hits
);
3608 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
3609 } else if (hdr
->b_l1hdr
.b_state
== arc_l2c_only
) {
3611 * This buffer is on the 2nd Level ARC.
3614 hdr
->b_l1hdr
.b_arc_access
= ddi_get_lbolt();
3615 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, hdr
);
3616 arc_change_state(arc_mfu
, hdr
, hash_lock
);
3618 cmn_err(CE_PANIC
, "invalid arc state 0x%p",
3619 hdr
->b_l1hdr
.b_state
);
3623 /* a generic arc_done_func_t which you can use */
3626 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3628 if (zio
== NULL
|| zio
->io_error
== 0)
3629 bcopy(buf
->b_data
, arg
, buf
->b_hdr
->b_size
);
3630 VERIFY(arc_buf_remove_ref(buf
, arg
));
3633 /* a generic arc_done_func_t */
3635 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3637 arc_buf_t
**bufp
= arg
;
3638 if (zio
&& zio
->io_error
) {
3639 VERIFY(arc_buf_remove_ref(buf
, arg
));
3643 ASSERT(buf
->b_data
);
3648 arc_read_done(zio_t
*zio
)
3652 arc_buf_t
*abuf
; /* buffer we're assigning to callback */
3653 kmutex_t
*hash_lock
= NULL
;
3654 arc_callback_t
*callback_list
, *acb
;
3655 int freeable
= FALSE
;
3657 buf
= zio
->io_private
;
3661 * The hdr was inserted into hash-table and removed from lists
3662 * prior to starting I/O. We should find this header, since
3663 * it's in the hash table, and it should be legit since it's
3664 * not possible to evict it during the I/O. The only possible
3665 * reason for it not to be found is if we were freed during the
3668 if (HDR_IN_HASH_TABLE(hdr
)) {
3669 arc_buf_hdr_t
*found
;
3671 ASSERT3U(hdr
->b_birth
, ==, BP_PHYSICAL_BIRTH(zio
->io_bp
));
3672 ASSERT3U(hdr
->b_dva
.dva_word
[0], ==,
3673 BP_IDENTITY(zio
->io_bp
)->dva_word
[0]);
3674 ASSERT3U(hdr
->b_dva
.dva_word
[1], ==,
3675 BP_IDENTITY(zio
->io_bp
)->dva_word
[1]);
3677 found
= buf_hash_find(hdr
->b_spa
, zio
->io_bp
,
3680 ASSERT((found
== NULL
&& HDR_FREED_IN_READ(hdr
) &&
3681 hash_lock
== NULL
) ||
3683 DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
3684 (found
== hdr
&& HDR_L2_READING(hdr
)));
3687 hdr
->b_flags
&= ~ARC_FLAG_L2_EVICTED
;
3688 if (l2arc_noprefetch
&& HDR_PREFETCH(hdr
))
3689 hdr
->b_flags
&= ~ARC_FLAG_L2CACHE
;
3691 /* byteswap if necessary */
3692 callback_list
= hdr
->b_l1hdr
.b_acb
;
3693 ASSERT(callback_list
!= NULL
);
3694 if (BP_SHOULD_BYTESWAP(zio
->io_bp
) && zio
->io_error
== 0) {
3695 dmu_object_byteswap_t bswap
=
3696 DMU_OT_BYTESWAP(BP_GET_TYPE(zio
->io_bp
));
3697 if (BP_GET_LEVEL(zio
->io_bp
) > 0)
3698 byteswap_uint64_array(buf
->b_data
, hdr
->b_size
);
3700 dmu_ot_byteswap
[bswap
].ob_func(buf
->b_data
, hdr
->b_size
);
3703 arc_cksum_compute(buf
, B_FALSE
);
3706 if (hash_lock
&& zio
->io_error
== 0 &&
3707 hdr
->b_l1hdr
.b_state
== arc_anon
) {
3709 * Only call arc_access on anonymous buffers. This is because
3710 * if we've issued an I/O for an evicted buffer, we've already
3711 * called arc_access (to prevent any simultaneous readers from
3712 * getting confused).
3714 arc_access(hdr
, hash_lock
);
3717 /* create copies of the data buffer for the callers */
3719 for (acb
= callback_list
; acb
; acb
= acb
->acb_next
) {
3720 if (acb
->acb_done
) {
3722 ARCSTAT_BUMP(arcstat_duplicate_reads
);
3723 abuf
= arc_buf_clone(buf
);
3725 acb
->acb_buf
= abuf
;
3729 hdr
->b_l1hdr
.b_acb
= NULL
;
3730 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
3731 ASSERT(!HDR_BUF_AVAILABLE(hdr
));
3733 ASSERT(buf
->b_efunc
== NULL
);
3734 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
3735 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
3738 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
) ||
3739 callback_list
!= NULL
);
3741 if (zio
->io_error
!= 0) {
3742 hdr
->b_flags
|= ARC_FLAG_IO_ERROR
;
3743 if (hdr
->b_l1hdr
.b_state
!= arc_anon
)
3744 arc_change_state(arc_anon
, hdr
, hash_lock
);
3745 if (HDR_IN_HASH_TABLE(hdr
))
3746 buf_hash_remove(hdr
);
3747 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
3751 * Broadcast before we drop the hash_lock to avoid the possibility
3752 * that the hdr (and hence the cv) might be freed before we get to
3753 * the cv_broadcast().
3755 cv_broadcast(&hdr
->b_l1hdr
.b_cv
);
3757 if (hash_lock
!= NULL
) {
3758 mutex_exit(hash_lock
);
3761 * This block was freed while we waited for the read to
3762 * complete. It has been removed from the hash table and
3763 * moved to the anonymous state (so that it won't show up
3766 ASSERT3P(hdr
->b_l1hdr
.b_state
, ==, arc_anon
);
3767 freeable
= refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
);
3770 /* execute each callback and free its structure */
3771 while ((acb
= callback_list
) != NULL
) {
3773 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
3775 if (acb
->acb_zio_dummy
!= NULL
) {
3776 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
3777 zio_nowait(acb
->acb_zio_dummy
);
3780 callback_list
= acb
->acb_next
;
3781 kmem_free(acb
, sizeof (arc_callback_t
));
3785 arc_hdr_destroy(hdr
);
3789 * "Read" the block at the specified DVA (in bp) via the
3790 * cache. If the block is found in the cache, invoke the provided
3791 * callback immediately and return. Note that the `zio' parameter
3792 * in the callback will be NULL in this case, since no IO was
3793 * required. If the block is not in the cache pass the read request
3794 * on to the spa with a substitute callback function, so that the
3795 * requested block will be added to the cache.
3797 * If a read request arrives for a block that has a read in-progress,
3798 * either wait for the in-progress read to complete (and return the
3799 * results); or, if this is a read with a "done" func, add a record
3800 * to the read to invoke the "done" func when the read completes,
3801 * and return; or just return.
3803 * arc_read_done() will invoke all the requested "done" functions
3804 * for readers of this block.
3807 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_done_func_t
*done
,
3808 void *private, zio_priority_t priority
, int zio_flags
,
3809 arc_flags_t
*arc_flags
, const zbookmark_phys_t
*zb
)
3811 arc_buf_hdr_t
*hdr
= NULL
;
3812 arc_buf_t
*buf
= NULL
;
3813 kmutex_t
*hash_lock
= NULL
;
3815 uint64_t guid
= spa_load_guid(spa
);
3818 ASSERT(!BP_IS_EMBEDDED(bp
) ||
3819 BPE_GET_ETYPE(bp
) == BP_EMBEDDED_TYPE_DATA
);
3822 if (!BP_IS_EMBEDDED(bp
)) {
3824 * Embedded BP's have no DVA and require no I/O to "read".
3825 * Create an anonymous arc buf to back it.
3827 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
3830 if (hdr
!= NULL
&& HDR_HAS_L1HDR(hdr
) && hdr
->b_l1hdr
.b_datacnt
> 0) {
3832 *arc_flags
|= ARC_FLAG_CACHED
;
3834 if (HDR_IO_IN_PROGRESS(hdr
)) {
3836 if (*arc_flags
& ARC_FLAG_WAIT
) {
3837 cv_wait(&hdr
->b_l1hdr
.b_cv
, hash_lock
);
3838 mutex_exit(hash_lock
);
3841 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
3844 arc_callback_t
*acb
= NULL
;
3846 acb
= kmem_zalloc(sizeof (arc_callback_t
),
3848 acb
->acb_done
= done
;
3849 acb
->acb_private
= private;
3851 acb
->acb_zio_dummy
= zio_null(pio
,
3852 spa
, NULL
, NULL
, NULL
, zio_flags
);
3854 ASSERT(acb
->acb_done
!= NULL
);
3855 acb
->acb_next
= hdr
->b_l1hdr
.b_acb
;
3856 hdr
->b_l1hdr
.b_acb
= acb
;
3857 add_reference(hdr
, hash_lock
, private);
3858 mutex_exit(hash_lock
);
3861 mutex_exit(hash_lock
);
3865 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
3866 hdr
->b_l1hdr
.b_state
== arc_mfu
);
3869 add_reference(hdr
, hash_lock
, private);
3871 * If this block is already in use, create a new
3872 * copy of the data so that we will be guaranteed
3873 * that arc_release() will always succeed.
3875 buf
= hdr
->b_l1hdr
.b_buf
;
3877 ASSERT(buf
->b_data
);
3878 if (HDR_BUF_AVAILABLE(hdr
)) {
3879 ASSERT(buf
->b_efunc
== NULL
);
3880 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
3882 buf
= arc_buf_clone(buf
);
3885 } else if (*arc_flags
& ARC_FLAG_PREFETCH
&&
3886 refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 0) {
3887 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
3889 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
3890 arc_access(hdr
, hash_lock
);
3891 if (*arc_flags
& ARC_FLAG_L2CACHE
)
3892 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
3893 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
3894 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
3895 mutex_exit(hash_lock
);
3896 ARCSTAT_BUMP(arcstat_hits
);
3897 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
3898 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
3899 data
, metadata
, hits
);
3902 done(NULL
, buf
, private);
3904 uint64_t size
= BP_GET_LSIZE(bp
);
3905 arc_callback_t
*acb
;
3908 boolean_t devw
= B_FALSE
;
3909 enum zio_compress b_compress
= ZIO_COMPRESS_OFF
;
3910 int32_t b_asize
= 0;
3913 * Gracefully handle a damaged logical block size as a
3914 * checksum error by passing a dummy zio to the done callback.
3916 if (size
> spa_maxblocksize(spa
)) {
3918 rzio
= zio_null(pio
, spa
, NULL
,
3919 NULL
, NULL
, zio_flags
);
3920 rzio
->io_error
= ECKSUM
;
3921 done(rzio
, buf
, private);
3929 /* this block is not in the cache */
3930 arc_buf_hdr_t
*exists
= NULL
;
3931 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
3932 buf
= arc_buf_alloc(spa
, size
, private, type
);
3934 if (!BP_IS_EMBEDDED(bp
)) {
3935 hdr
->b_dva
= *BP_IDENTITY(bp
);
3936 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
3937 exists
= buf_hash_insert(hdr
, &hash_lock
);
3939 if (exists
!= NULL
) {
3940 /* somebody beat us to the hash insert */
3941 mutex_exit(hash_lock
);
3942 buf_discard_identity(hdr
);
3943 (void) arc_buf_remove_ref(buf
, private);
3944 goto top
; /* restart the IO request */
3947 /* if this is a prefetch, we don't have a reference */
3948 if (*arc_flags
& ARC_FLAG_PREFETCH
) {
3949 (void) remove_reference(hdr
, hash_lock
,
3951 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
3953 if (*arc_flags
& ARC_FLAG_L2CACHE
)
3954 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
3955 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
3956 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
3957 if (BP_GET_LEVEL(bp
) > 0)
3958 hdr
->b_flags
|= ARC_FLAG_INDIRECT
;
3961 * This block is in the ghost cache. If it was L2-only
3962 * (and thus didn't have an L1 hdr), we realloc the
3963 * header to add an L1 hdr.
3965 if (!HDR_HAS_L1HDR(hdr
)) {
3966 hdr
= arc_hdr_realloc(hdr
, hdr_l2only_cache
,
3970 ASSERT(GHOST_STATE(hdr
->b_l1hdr
.b_state
));
3971 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3972 ASSERT(refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
3973 ASSERT3P(hdr
->b_l1hdr
.b_buf
, ==, NULL
);
3975 /* if this is a prefetch, we don't have a reference */
3976 if (*arc_flags
& ARC_FLAG_PREFETCH
)
3977 hdr
->b_flags
|= ARC_FLAG_PREFETCH
;
3979 add_reference(hdr
, hash_lock
, private);
3980 if (*arc_flags
& ARC_FLAG_L2CACHE
)
3981 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
3982 if (*arc_flags
& ARC_FLAG_L2COMPRESS
)
3983 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
3984 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
3987 buf
->b_efunc
= NULL
;
3988 buf
->b_private
= NULL
;
3990 hdr
->b_l1hdr
.b_buf
= buf
;
3991 ASSERT0(hdr
->b_l1hdr
.b_datacnt
);
3992 hdr
->b_l1hdr
.b_datacnt
= 1;
3993 arc_get_data_buf(buf
);
3994 arc_access(hdr
, hash_lock
);
3997 ASSERT(!GHOST_STATE(hdr
->b_l1hdr
.b_state
));
3999 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
4000 acb
->acb_done
= done
;
4001 acb
->acb_private
= private;
4003 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
4004 hdr
->b_l1hdr
.b_acb
= acb
;
4005 hdr
->b_flags
|= ARC_FLAG_IO_IN_PROGRESS
;
4007 if (HDR_HAS_L2HDR(hdr
) &&
4008 (vd
= hdr
->b_l2hdr
.b_dev
->l2ad_vdev
) != NULL
) {
4009 devw
= hdr
->b_l2hdr
.b_dev
->l2ad_writing
;
4010 addr
= hdr
->b_l2hdr
.b_daddr
;
4011 b_compress
= HDR_GET_COMPRESS(hdr
);
4012 b_asize
= hdr
->b_l2hdr
.b_asize
;
4014 * Lock out device removal.
4016 if (vdev_is_dead(vd
) ||
4017 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
4021 if (hash_lock
!= NULL
)
4022 mutex_exit(hash_lock
);
4025 * At this point, we have a level 1 cache miss. Try again in
4026 * L2ARC if possible.
4028 ASSERT3U(hdr
->b_size
, ==, size
);
4029 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
4030 uint64_t, size
, zbookmark_phys_t
*, zb
);
4031 ARCSTAT_BUMP(arcstat_misses
);
4032 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr
),
4033 demand
, prefetch
, !HDR_ISTYPE_METADATA(hdr
),
4034 data
, metadata
, misses
);
4036 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
4038 * Read from the L2ARC if the following are true:
4039 * 1. The L2ARC vdev was previously cached.
4040 * 2. This buffer still has L2ARC metadata.
4041 * 3. This buffer isn't currently writing to the L2ARC.
4042 * 4. The L2ARC entry wasn't evicted, which may
4043 * also have invalidated the vdev.
4044 * 5. This isn't prefetch and l2arc_noprefetch is set.
4046 if (HDR_HAS_L2HDR(hdr
) &&
4047 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
4048 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
4049 l2arc_read_callback_t
*cb
;
4051 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
4052 ARCSTAT_BUMP(arcstat_l2_hits
);
4053 atomic_inc_32(&hdr
->b_l2hdr
.b_hits
);
4055 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
4057 cb
->l2rcb_buf
= buf
;
4058 cb
->l2rcb_spa
= spa
;
4061 cb
->l2rcb_flags
= zio_flags
;
4062 cb
->l2rcb_compress
= b_compress
;
4064 ASSERT(addr
>= VDEV_LABEL_START_SIZE
&&
4065 addr
+ size
< vd
->vdev_psize
-
4066 VDEV_LABEL_END_SIZE
);
4069 * l2arc read. The SCL_L2ARC lock will be
4070 * released by l2arc_read_done().
4071 * Issue a null zio if the underlying buffer
4072 * was squashed to zero size by compression.
4074 if (b_compress
== ZIO_COMPRESS_EMPTY
) {
4075 rzio
= zio_null(pio
, spa
, vd
,
4076 l2arc_read_done
, cb
,
4077 zio_flags
| ZIO_FLAG_DONT_CACHE
|
4079 ZIO_FLAG_DONT_PROPAGATE
|
4080 ZIO_FLAG_DONT_RETRY
);
4082 rzio
= zio_read_phys(pio
, vd
, addr
,
4083 b_asize
, buf
->b_data
,
4085 l2arc_read_done
, cb
, priority
,
4086 zio_flags
| ZIO_FLAG_DONT_CACHE
|
4088 ZIO_FLAG_DONT_PROPAGATE
|
4089 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
4091 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
4093 ARCSTAT_INCR(arcstat_l2_read_bytes
, b_asize
);
4095 if (*arc_flags
& ARC_FLAG_NOWAIT
) {
4100 ASSERT(*arc_flags
& ARC_FLAG_WAIT
);
4101 if (zio_wait(rzio
) == 0)
4104 /* l2arc read error; goto zio_read() */
4106 DTRACE_PROBE1(l2arc__miss
,
4107 arc_buf_hdr_t
*, hdr
);
4108 ARCSTAT_BUMP(arcstat_l2_misses
);
4109 if (HDR_L2_WRITING(hdr
))
4110 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
4111 spa_config_exit(spa
, SCL_L2ARC
, vd
);
4115 spa_config_exit(spa
, SCL_L2ARC
, vd
);
4116 if (l2arc_ndev
!= 0) {
4117 DTRACE_PROBE1(l2arc__miss
,
4118 arc_buf_hdr_t
*, hdr
);
4119 ARCSTAT_BUMP(arcstat_l2_misses
);
4123 rzio
= zio_read(pio
, spa
, bp
, buf
->b_data
, size
,
4124 arc_read_done
, buf
, priority
, zio_flags
, zb
);
4126 if (*arc_flags
& ARC_FLAG_WAIT
) {
4127 rc
= zio_wait(rzio
);
4131 ASSERT(*arc_flags
& ARC_FLAG_NOWAIT
);
4136 spa_read_history_add(spa
, zb
, *arc_flags
);
4141 arc_add_prune_callback(arc_prune_func_t
*func
, void *private)
4145 p
= kmem_alloc(sizeof (*p
), KM_SLEEP
);
4147 p
->p_private
= private;
4148 list_link_init(&p
->p_node
);
4149 refcount_create(&p
->p_refcnt
);
4151 mutex_enter(&arc_prune_mtx
);
4152 refcount_add(&p
->p_refcnt
, &arc_prune_list
);
4153 list_insert_head(&arc_prune_list
, p
);
4154 mutex_exit(&arc_prune_mtx
);
4160 arc_remove_prune_callback(arc_prune_t
*p
)
4162 mutex_enter(&arc_prune_mtx
);
4163 list_remove(&arc_prune_list
, p
);
4164 if (refcount_remove(&p
->p_refcnt
, &arc_prune_list
) == 0) {
4165 refcount_destroy(&p
->p_refcnt
);
4166 kmem_free(p
, sizeof (*p
));
4168 mutex_exit(&arc_prune_mtx
);
4172 arc_set_callback(arc_buf_t
*buf
, arc_evict_func_t
*func
, void *private)
4174 ASSERT(buf
->b_hdr
!= NULL
);
4175 ASSERT(buf
->b_hdr
->b_l1hdr
.b_state
!= arc_anon
);
4176 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
) ||
4178 ASSERT(buf
->b_efunc
== NULL
);
4179 ASSERT(!HDR_BUF_AVAILABLE(buf
->b_hdr
));
4181 buf
->b_efunc
= func
;
4182 buf
->b_private
= private;
4186 * Notify the arc that a block was freed, and thus will never be used again.
4189 arc_freed(spa_t
*spa
, const blkptr_t
*bp
)
4192 kmutex_t
*hash_lock
;
4193 uint64_t guid
= spa_load_guid(spa
);
4195 ASSERT(!BP_IS_EMBEDDED(bp
));
4197 hdr
= buf_hash_find(guid
, bp
, &hash_lock
);
4200 if (HDR_BUF_AVAILABLE(hdr
)) {
4201 arc_buf_t
*buf
= hdr
->b_l1hdr
.b_buf
;
4202 add_reference(hdr
, hash_lock
, FTAG
);
4203 hdr
->b_flags
&= ~ARC_FLAG_BUF_AVAILABLE
;
4204 mutex_exit(hash_lock
);
4206 arc_release(buf
, FTAG
);
4207 (void) arc_buf_remove_ref(buf
, FTAG
);
4209 mutex_exit(hash_lock
);
4215 * Clear the user eviction callback set by arc_set_callback(), first calling
4216 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4217 * clearing the callback may result in the arc_buf being destroyed. However,
4218 * it will not result in the *last* arc_buf being destroyed, hence the data
4219 * will remain cached in the ARC. We make a copy of the arc buffer here so
4220 * that we can process the callback without holding any locks.
4222 * It's possible that the callback is already in the process of being cleared
4223 * by another thread. In this case we can not clear the callback.
4225 * Returns B_TRUE if the callback was successfully called and cleared.
4228 arc_clear_callback(arc_buf_t
*buf
)
4231 kmutex_t
*hash_lock
;
4232 arc_evict_func_t
*efunc
= buf
->b_efunc
;
4233 void *private = buf
->b_private
;
4235 mutex_enter(&buf
->b_evict_lock
);
4239 * We are in arc_do_user_evicts().
4241 ASSERT(buf
->b_data
== NULL
);
4242 mutex_exit(&buf
->b_evict_lock
);
4244 } else if (buf
->b_data
== NULL
) {
4246 * We are on the eviction list; process this buffer now
4247 * but let arc_do_user_evicts() do the reaping.
4249 buf
->b_efunc
= NULL
;
4250 mutex_exit(&buf
->b_evict_lock
);
4251 VERIFY0(efunc(private));
4254 hash_lock
= HDR_LOCK(hdr
);
4255 mutex_enter(hash_lock
);
4257 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
4259 ASSERT3U(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), <,
4260 hdr
->b_l1hdr
.b_datacnt
);
4261 ASSERT(hdr
->b_l1hdr
.b_state
== arc_mru
||
4262 hdr
->b_l1hdr
.b_state
== arc_mfu
);
4264 buf
->b_efunc
= NULL
;
4265 buf
->b_private
= NULL
;
4267 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
4268 mutex_exit(&buf
->b_evict_lock
);
4269 arc_buf_destroy(buf
, TRUE
);
4271 ASSERT(buf
== hdr
->b_l1hdr
.b_buf
);
4272 hdr
->b_flags
|= ARC_FLAG_BUF_AVAILABLE
;
4273 mutex_exit(&buf
->b_evict_lock
);
4276 mutex_exit(hash_lock
);
4277 VERIFY0(efunc(private));
4282 * Release this buffer from the cache, making it an anonymous buffer. This
4283 * must be done after a read and prior to modifying the buffer contents.
4284 * If the buffer has more than one reference, we must make
4285 * a new hdr for the buffer.
4288 arc_release(arc_buf_t
*buf
, void *tag
)
4290 kmutex_t
*hash_lock
;
4292 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4295 * It would be nice to assert that if its DMU metadata (level >
4296 * 0 || it's the dnode file), then it must be syncing context.
4297 * But we don't know that information at this level.
4300 mutex_enter(&buf
->b_evict_lock
);
4302 ASSERT(HDR_HAS_L1HDR(hdr
));
4305 * We don't grab the hash lock prior to this check, because if
4306 * the buffer's header is in the arc_anon state, it won't be
4307 * linked into the hash table.
4309 if (hdr
->b_l1hdr
.b_state
== arc_anon
) {
4310 mutex_exit(&buf
->b_evict_lock
);
4311 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4312 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
4313 ASSERT(!HDR_HAS_L2HDR(hdr
));
4314 ASSERT(BUF_EMPTY(hdr
));
4316 ASSERT3U(hdr
->b_l1hdr
.b_datacnt
, ==, 1);
4317 ASSERT3S(refcount_count(&hdr
->b_l1hdr
.b_refcnt
), ==, 1);
4318 ASSERT(!list_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4320 ASSERT3P(buf
->b_efunc
, ==, NULL
);
4321 ASSERT3P(buf
->b_private
, ==, NULL
);
4323 hdr
->b_l1hdr
.b_arc_access
= 0;
4329 hash_lock
= HDR_LOCK(hdr
);
4330 mutex_enter(hash_lock
);
4333 * This assignment is only valid as long as the hash_lock is
4334 * held, we must be careful not to reference state or the
4335 * b_state field after dropping the lock.
4337 state
= hdr
->b_l1hdr
.b_state
;
4338 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
4339 ASSERT3P(state
, !=, arc_anon
);
4341 /* this buffer is not on any list */
4342 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) > 0);
4344 if (HDR_HAS_L2HDR(hdr
)) {
4345 ARCSTAT_INCR(arcstat_l2_asize
, -hdr
->b_l2hdr
.b_asize
);
4346 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
4348 mutex_enter(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
4349 list_remove(&hdr
->b_l2hdr
.b_dev
->l2ad_buflist
, hdr
);
4352 * We don't want to leak the b_tmp_cdata buffer that was
4353 * allocated in l2arc_write_buffers()
4355 arc_buf_l2_cdata_free(hdr
);
4357 mutex_exit(&hdr
->b_l2hdr
.b_dev
->l2ad_mtx
);
4359 hdr
->b_flags
&= ~ARC_FLAG_HAS_L2HDR
;
4363 * Do we have more than one buf?
4365 if (hdr
->b_l1hdr
.b_datacnt
> 1) {
4366 arc_buf_hdr_t
*nhdr
;
4368 uint64_t blksz
= hdr
->b_size
;
4369 uint64_t spa
= hdr
->b_spa
;
4370 arc_buf_contents_t type
= arc_buf_type(hdr
);
4371 uint32_t flags
= hdr
->b_flags
;
4373 ASSERT(hdr
->b_l1hdr
.b_buf
!= buf
|| buf
->b_next
!= NULL
);
4375 * Pull the data off of this hdr and attach it to
4376 * a new anonymous hdr.
4378 (void) remove_reference(hdr
, hash_lock
, tag
);
4379 bufp
= &hdr
->b_l1hdr
.b_buf
;
4380 while (*bufp
!= buf
)
4381 bufp
= &(*bufp
)->b_next
;
4382 *bufp
= buf
->b_next
;
4385 ASSERT3P(state
, !=, arc_l2c_only
);
4386 ASSERT3U(state
->arcs_size
, >=, hdr
->b_size
);
4387 atomic_add_64(&state
->arcs_size
, -hdr
->b_size
);
4388 if (refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
)) {
4391 ASSERT3P(state
, !=, arc_l2c_only
);
4392 size
= &state
->arcs_lsize
[type
];
4393 ASSERT3U(*size
, >=, hdr
->b_size
);
4394 atomic_add_64(size
, -hdr
->b_size
);
4398 * We're releasing a duplicate user data buffer, update
4399 * our statistics accordingly.
4401 if (HDR_ISTYPE_DATA(hdr
)) {
4402 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers
);
4403 ARCSTAT_INCR(arcstat_duplicate_buffers_size
,
4406 hdr
->b_l1hdr
.b_datacnt
-= 1;
4407 arc_cksum_verify(buf
);
4408 arc_buf_unwatch(buf
);
4410 mutex_exit(hash_lock
);
4412 nhdr
= kmem_cache_alloc(hdr_full_cache
, KM_PUSHPAGE
);
4413 nhdr
->b_size
= blksz
;
4416 nhdr
->b_l1hdr
.b_mru_hits
= 0;
4417 nhdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
4418 nhdr
->b_l1hdr
.b_mfu_hits
= 0;
4419 nhdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
4420 nhdr
->b_l1hdr
.b_l2_hits
= 0;
4421 nhdr
->b_flags
= flags
& ARC_FLAG_L2_WRITING
;
4422 nhdr
->b_flags
|= arc_bufc_to_flags(type
);
4423 nhdr
->b_flags
|= ARC_FLAG_HAS_L1HDR
;
4425 nhdr
->b_l1hdr
.b_buf
= buf
;
4426 nhdr
->b_l1hdr
.b_datacnt
= 1;
4427 nhdr
->b_l1hdr
.b_state
= arc_anon
;
4428 nhdr
->b_l1hdr
.b_arc_access
= 0;
4429 nhdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
4430 nhdr
->b_freeze_cksum
= NULL
;
4432 (void) refcount_add(&nhdr
->b_l1hdr
.b_refcnt
, tag
);
4434 mutex_exit(&buf
->b_evict_lock
);
4435 atomic_add_64(&arc_anon
->arcs_size
, blksz
);
4437 mutex_exit(&buf
->b_evict_lock
);
4438 ASSERT(refcount_count(&hdr
->b_l1hdr
.b_refcnt
) == 1);
4439 /* protected by hash lock, or hdr is on arc_anon */
4440 ASSERT(!multilist_link_active(&hdr
->b_l1hdr
.b_arc_node
));
4441 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4442 hdr
->b_l1hdr
.b_mru_hits
= 0;
4443 hdr
->b_l1hdr
.b_mru_ghost_hits
= 0;
4444 hdr
->b_l1hdr
.b_mfu_hits
= 0;
4445 hdr
->b_l1hdr
.b_mfu_ghost_hits
= 0;
4446 hdr
->b_l1hdr
.b_l2_hits
= 0;
4447 arc_change_state(arc_anon
, hdr
, hash_lock
);
4448 hdr
->b_l1hdr
.b_arc_access
= 0;
4449 mutex_exit(hash_lock
);
4451 buf_discard_identity(hdr
);
4454 buf
->b_efunc
= NULL
;
4455 buf
->b_private
= NULL
;
4459 arc_released(arc_buf_t
*buf
)
4463 mutex_enter(&buf
->b_evict_lock
);
4464 released
= (buf
->b_data
!= NULL
&&
4465 buf
->b_hdr
->b_l1hdr
.b_state
== arc_anon
);
4466 mutex_exit(&buf
->b_evict_lock
);
4472 arc_referenced(arc_buf_t
*buf
)
4476 mutex_enter(&buf
->b_evict_lock
);
4477 referenced
= (refcount_count(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
4478 mutex_exit(&buf
->b_evict_lock
);
4479 return (referenced
);
4484 arc_write_ready(zio_t
*zio
)
4486 arc_write_callback_t
*callback
= zio
->io_private
;
4487 arc_buf_t
*buf
= callback
->awcb_buf
;
4488 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4490 ASSERT(HDR_HAS_L1HDR(hdr
));
4491 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_l1hdr
.b_refcnt
));
4492 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
4493 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
4496 * If the IO is already in progress, then this is a re-write
4497 * attempt, so we need to thaw and re-compute the cksum.
4498 * It is the responsibility of the callback to handle the
4499 * accounting for any re-write attempt.
4501 if (HDR_IO_IN_PROGRESS(hdr
)) {
4502 mutex_enter(&hdr
->b_l1hdr
.b_freeze_lock
);
4503 if (hdr
->b_freeze_cksum
!= NULL
) {
4504 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
4505 hdr
->b_freeze_cksum
= NULL
;
4507 mutex_exit(&hdr
->b_l1hdr
.b_freeze_lock
);
4509 arc_cksum_compute(buf
, B_FALSE
);
4510 hdr
->b_flags
|= ARC_FLAG_IO_IN_PROGRESS
;
4514 * The SPA calls this callback for each physical write that happens on behalf
4515 * of a logical write. See the comment in dbuf_write_physdone() for details.
4518 arc_write_physdone(zio_t
*zio
)
4520 arc_write_callback_t
*cb
= zio
->io_private
;
4521 if (cb
->awcb_physdone
!= NULL
)
4522 cb
->awcb_physdone(zio
, cb
->awcb_buf
, cb
->awcb_private
);
4526 arc_write_done(zio_t
*zio
)
4528 arc_write_callback_t
*callback
= zio
->io_private
;
4529 arc_buf_t
*buf
= callback
->awcb_buf
;
4530 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4532 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
4534 if (zio
->io_error
== 0) {
4535 if (BP_IS_HOLE(zio
->io_bp
) || BP_IS_EMBEDDED(zio
->io_bp
)) {
4536 buf_discard_identity(hdr
);
4538 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
4539 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
4542 ASSERT(BUF_EMPTY(hdr
));
4546 * If the block to be written was all-zero or compressed enough to be
4547 * embedded in the BP, no write was performed so there will be no
4548 * dva/birth/checksum. The buffer must therefore remain anonymous
4551 if (!BUF_EMPTY(hdr
)) {
4552 arc_buf_hdr_t
*exists
;
4553 kmutex_t
*hash_lock
;
4555 ASSERT(zio
->io_error
== 0);
4557 arc_cksum_verify(buf
);
4559 exists
= buf_hash_insert(hdr
, &hash_lock
);
4560 if (exists
!= NULL
) {
4562 * This can only happen if we overwrite for
4563 * sync-to-convergence, because we remove
4564 * buffers from the hash table when we arc_free().
4566 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
4567 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
4568 panic("bad overwrite, hdr=%p exists=%p",
4569 (void *)hdr
, (void *)exists
);
4570 ASSERT(refcount_is_zero(
4571 &exists
->b_l1hdr
.b_refcnt
));
4572 arc_change_state(arc_anon
, exists
, hash_lock
);
4573 mutex_exit(hash_lock
);
4574 arc_hdr_destroy(exists
);
4575 exists
= buf_hash_insert(hdr
, &hash_lock
);
4576 ASSERT3P(exists
, ==, NULL
);
4577 } else if (zio
->io_flags
& ZIO_FLAG_NOPWRITE
) {
4579 ASSERT(zio
->io_prop
.zp_nopwrite
);
4580 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
4581 panic("bad nopwrite, hdr=%p exists=%p",
4582 (void *)hdr
, (void *)exists
);
4585 ASSERT(hdr
->b_l1hdr
.b_datacnt
== 1);
4586 ASSERT(hdr
->b_l1hdr
.b_state
== arc_anon
);
4587 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
4588 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
4591 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
4592 /* if it's not anon, we are doing a scrub */
4593 if (exists
== NULL
&& hdr
->b_l1hdr
.b_state
== arc_anon
)
4594 arc_access(hdr
, hash_lock
);
4595 mutex_exit(hash_lock
);
4597 hdr
->b_flags
&= ~ARC_FLAG_IO_IN_PROGRESS
;
4600 ASSERT(!refcount_is_zero(&hdr
->b_l1hdr
.b_refcnt
));
4601 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
4603 kmem_free(callback
, sizeof (arc_write_callback_t
));
4607 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
4608 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
, boolean_t l2arc_compress
,
4609 const zio_prop_t
*zp
, arc_done_func_t
*ready
, arc_done_func_t
*physdone
,
4610 arc_done_func_t
*done
, void *private, zio_priority_t priority
,
4611 int zio_flags
, const zbookmark_phys_t
*zb
)
4613 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
4614 arc_write_callback_t
*callback
;
4617 ASSERT(ready
!= NULL
);
4618 ASSERT(done
!= NULL
);
4619 ASSERT(!HDR_IO_ERROR(hdr
));
4620 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
4621 ASSERT(hdr
->b_l1hdr
.b_acb
== NULL
);
4622 ASSERT(hdr
->b_l1hdr
.b_datacnt
> 0);
4624 hdr
->b_flags
|= ARC_FLAG_L2CACHE
;
4626 hdr
->b_flags
|= ARC_FLAG_L2COMPRESS
;
4627 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
4628 callback
->awcb_ready
= ready
;
4629 callback
->awcb_physdone
= physdone
;
4630 callback
->awcb_done
= done
;
4631 callback
->awcb_private
= private;
4632 callback
->awcb_buf
= buf
;
4634 zio
= zio_write(pio
, spa
, txg
, bp
, buf
->b_data
, hdr
->b_size
, zp
,
4635 arc_write_ready
, arc_write_physdone
, arc_write_done
, callback
,
4636 priority
, zio_flags
, zb
);
4642 arc_memory_throttle(uint64_t reserve
, uint64_t txg
)
4645 if (zfs_arc_memory_throttle_disable
)
4648 if (freemem
<= physmem
* arc_lotsfree_percent
/ 100) {
4649 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
4650 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim
);
4651 return (SET_ERROR(EAGAIN
));
4658 arc_tempreserve_clear(uint64_t reserve
)
4660 atomic_add_64(&arc_tempreserve
, -reserve
);
4661 ASSERT((int64_t)arc_tempreserve
>= 0);
4665 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
4670 if (reserve
> arc_c
/4 && !arc_no_grow
)
4671 arc_c
= MIN(arc_c_max
, reserve
* 4);
4674 * Throttle when the calculated memory footprint for the TXG
4675 * exceeds the target ARC size.
4677 if (reserve
> arc_c
) {
4678 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve
);
4679 return (SET_ERROR(ERESTART
));
4683 * Don't count loaned bufs as in flight dirty data to prevent long
4684 * network delays from blocking transactions that are ready to be
4685 * assigned to a txg.
4687 anon_size
= MAX((int64_t)(arc_anon
->arcs_size
- arc_loaned_bytes
), 0);
4690 * Writes will, almost always, require additional memory allocations
4691 * in order to compress/encrypt/etc the data. We therefore need to
4692 * make sure that there is sufficient available memory for this.
4694 error
= arc_memory_throttle(reserve
, txg
);
4699 * Throttle writes when the amount of dirty data in the cache
4700 * gets too large. We try to keep the cache less than half full
4701 * of dirty blocks so that our sync times don't grow too large.
4702 * Note: if two requests come in concurrently, we might let them
4703 * both succeed, when one of them should fail. Not a huge deal.
4706 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
4707 anon_size
> arc_c
/ 4) {
4708 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
4709 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
4710 arc_tempreserve
>>10,
4711 arc_anon
->arcs_lsize
[ARC_BUFC_METADATA
]>>10,
4712 arc_anon
->arcs_lsize
[ARC_BUFC_DATA
]>>10,
4713 reserve
>>10, arc_c
>>10);
4714 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle
);
4715 return (SET_ERROR(ERESTART
));
4717 atomic_add_64(&arc_tempreserve
, reserve
);
4722 arc_kstat_update_state(arc_state_t
*state
, kstat_named_t
*size
,
4723 kstat_named_t
*evict_data
, kstat_named_t
*evict_metadata
)
4725 size
->value
.ui64
= state
->arcs_size
;
4726 evict_data
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_DATA
];
4727 evict_metadata
->value
.ui64
= state
->arcs_lsize
[ARC_BUFC_METADATA
];
4731 arc_kstat_update(kstat_t
*ksp
, int rw
)
4733 arc_stats_t
*as
= ksp
->ks_data
;
4735 if (rw
== KSTAT_WRITE
) {
4736 return (SET_ERROR(EACCES
));
4738 arc_kstat_update_state(arc_anon
,
4739 &as
->arcstat_anon_size
,
4740 &as
->arcstat_anon_evict_data
,
4741 &as
->arcstat_anon_evict_metadata
);
4742 arc_kstat_update_state(arc_mru
,
4743 &as
->arcstat_mru_size
,
4744 &as
->arcstat_mru_evict_data
,
4745 &as
->arcstat_mru_evict_metadata
);
4746 arc_kstat_update_state(arc_mru_ghost
,
4747 &as
->arcstat_mru_ghost_size
,
4748 &as
->arcstat_mru_ghost_evict_data
,
4749 &as
->arcstat_mru_ghost_evict_metadata
);
4750 arc_kstat_update_state(arc_mfu
,
4751 &as
->arcstat_mfu_size
,
4752 &as
->arcstat_mfu_evict_data
,
4753 &as
->arcstat_mfu_evict_metadata
);
4754 arc_kstat_update_state(arc_mfu_ghost
,
4755 &as
->arcstat_mfu_ghost_size
,
4756 &as
->arcstat_mfu_ghost_evict_data
,
4757 &as
->arcstat_mfu_ghost_evict_metadata
);
4764 * This function *must* return indices evenly distributed between all
4765 * sublists of the multilist. This is needed due to how the ARC eviction
4766 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
4767 * distributed between all sublists and uses this assumption when
4768 * deciding which sublist to evict from and how much to evict from it.
4771 arc_state_multilist_index_func(multilist_t
*ml
, void *obj
)
4773 arc_buf_hdr_t
*hdr
= obj
;
4776 * We rely on b_dva to generate evenly distributed index
4777 * numbers using buf_hash below. So, as an added precaution,
4778 * let's make sure we never add empty buffers to the arc lists.
4780 ASSERT(!BUF_EMPTY(hdr
));
4783 * The assumption here, is the hash value for a given
4784 * arc_buf_hdr_t will remain constant throughout its lifetime
4785 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
4786 * Thus, we don't need to store the header's sublist index
4787 * on insertion, as this index can be recalculated on removal.
4789 * Also, the low order bits of the hash value are thought to be
4790 * distributed evenly. Otherwise, in the case that the multilist
4791 * has a power of two number of sublists, each sublists' usage
4792 * would not be evenly distributed.
4794 return (buf_hash(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
) %
4795 multilist_get_num_sublists(ml
));
4801 mutex_init(&arc_reclaim_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
4802 cv_init(&arc_reclaim_thread_cv
, NULL
, CV_DEFAULT
, NULL
);
4803 cv_init(&arc_reclaim_waiters_cv
, NULL
, CV_DEFAULT
, NULL
);
4805 mutex_init(&arc_user_evicts_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
4806 cv_init(&arc_user_evicts_cv
, NULL
, CV_DEFAULT
, NULL
);
4808 /* Convert seconds to clock ticks */
4809 zfs_arc_min_prefetch_lifespan
= 1 * hz
;
4811 /* Start out with 1/8 of all memory */
4812 arc_c
= physmem
* PAGESIZE
/ 8;
4816 * On architectures where the physical memory can be larger
4817 * than the addressable space (intel in 32-bit mode), we may
4818 * need to limit the cache to 1/8 of VM size.
4820 arc_c
= MIN(arc_c
, vmem_size(heap_arena
, VMEM_ALLOC
| VMEM_FREE
) / 8);
4822 * Register a shrinker to support synchronous (direct) memory
4823 * reclaim from the arc. This is done to prevent kswapd from
4824 * swapping out pages when it is preferable to shrink the arc.
4826 spl_register_shrinker(&arc_shrinker
);
4829 /* set min cache to zero */
4831 /* set max to 1/2 of all memory */
4832 arc_c_max
= arc_c
* 4;
4835 * Allow the tunables to override our calculations if they are
4836 * reasonable (ie. over 64MB)
4838 if (zfs_arc_max
> 64<<20 && zfs_arc_max
< physmem
* PAGESIZE
)
4839 arc_c_max
= zfs_arc_max
;
4840 if (zfs_arc_min
> 0 && zfs_arc_min
<= arc_c_max
)
4841 arc_c_min
= zfs_arc_min
;
4844 arc_p
= (arc_c
>> 1);
4846 /* limit meta-data to 3/4 of the arc capacity */
4847 arc_meta_limit
= (3 * arc_c_max
) / 4;
4850 /* Allow the tunable to override if it is reasonable */
4851 if (zfs_arc_meta_limit
> 0 && zfs_arc_meta_limit
<= arc_c_max
)
4852 arc_meta_limit
= zfs_arc_meta_limit
;
4854 if (zfs_arc_num_sublists_per_state
< 1)
4855 zfs_arc_num_sublists_per_state
= num_online_cpus();
4857 /* if kmem_flags are set, lets try to use less memory */
4858 if (kmem_debugging())
4860 if (arc_c
< arc_c_min
)
4863 arc_anon
= &ARC_anon
;
4865 arc_mru_ghost
= &ARC_mru_ghost
;
4867 arc_mfu_ghost
= &ARC_mfu_ghost
;
4868 arc_l2c_only
= &ARC_l2c_only
;
4871 multilist_create(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
],
4872 sizeof (arc_buf_hdr_t
),
4873 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
4874 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
4875 multilist_create(&arc_mru
->arcs_list
[ARC_BUFC_DATA
],
4876 sizeof (arc_buf_hdr_t
),
4877 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
4878 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
4879 multilist_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
],
4880 sizeof (arc_buf_hdr_t
),
4881 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
4882 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
4883 multilist_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
],
4884 sizeof (arc_buf_hdr_t
),
4885 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
4886 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
4887 multilist_create(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
],
4888 sizeof (arc_buf_hdr_t
),
4889 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
4890 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
4891 multilist_create(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
],
4892 sizeof (arc_buf_hdr_t
),
4893 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
4894 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
4895 multilist_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
],
4896 sizeof (arc_buf_hdr_t
),
4897 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
4898 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
4899 multilist_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
],
4900 sizeof (arc_buf_hdr_t
),
4901 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
4902 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
4903 multilist_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
],
4904 sizeof (arc_buf_hdr_t
),
4905 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
4906 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
4907 multilist_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
],
4908 sizeof (arc_buf_hdr_t
),
4909 offsetof(arc_buf_hdr_t
, b_l1hdr
.b_arc_node
),
4910 zfs_arc_num_sublists_per_state
, arc_state_multilist_index_func
);
4912 arc_anon
->arcs_state
= ARC_STATE_ANON
;
4913 arc_mru
->arcs_state
= ARC_STATE_MRU
;
4914 arc_mru_ghost
->arcs_state
= ARC_STATE_MRU_GHOST
;
4915 arc_mfu
->arcs_state
= ARC_STATE_MFU
;
4916 arc_mfu_ghost
->arcs_state
= ARC_STATE_MFU_GHOST
;
4917 arc_l2c_only
->arcs_state
= ARC_STATE_L2C_ONLY
;
4921 arc_reclaim_thread_exit
= FALSE
;
4922 arc_user_evicts_thread_exit
= FALSE
;
4923 list_create(&arc_prune_list
, sizeof (arc_prune_t
),
4924 offsetof(arc_prune_t
, p_node
));
4925 arc_eviction_list
= NULL
;
4926 mutex_init(&arc_prune_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4927 bzero(&arc_eviction_hdr
, sizeof (arc_buf_hdr_t
));
4929 arc_prune_taskq
= taskq_create("arc_prune", max_ncpus
, minclsyspri
,
4930 max_ncpus
, INT_MAX
, TASKQ_PREPOPULATE
);
4932 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
4933 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
4935 if (arc_ksp
!= NULL
) {
4936 arc_ksp
->ks_data
= &arc_stats
;
4937 arc_ksp
->ks_update
= arc_kstat_update
;
4938 kstat_install(arc_ksp
);
4941 (void) thread_create(NULL
, 0, arc_adapt_thread
, NULL
, 0, &p0
,
4942 TS_RUN
, minclsyspri
);
4944 (void) thread_create(NULL
, 0, arc_user_evicts_thread
, NULL
, 0, &p0
,
4945 TS_RUN
, minclsyspri
);
4951 * Calculate maximum amount of dirty data per pool.
4953 * If it has been set by a module parameter, take that.
4954 * Otherwise, use a percentage of physical memory defined by
4955 * zfs_dirty_data_max_percent (default 10%) with a cap at
4956 * zfs_dirty_data_max_max (default 25% of physical memory).
4958 if (zfs_dirty_data_max_max
== 0)
4959 zfs_dirty_data_max_max
= physmem
* PAGESIZE
*
4960 zfs_dirty_data_max_max_percent
/ 100;
4962 if (zfs_dirty_data_max
== 0) {
4963 zfs_dirty_data_max
= physmem
* PAGESIZE
*
4964 zfs_dirty_data_max_percent
/ 100;
4965 zfs_dirty_data_max
= MIN(zfs_dirty_data_max
,
4966 zfs_dirty_data_max_max
);
4976 spl_unregister_shrinker(&arc_shrinker
);
4977 #endif /* _KERNEL */
4979 mutex_enter(&arc_reclaim_lock
);
4980 arc_reclaim_thread_exit
= TRUE
;
4982 * The reclaim thread will set arc_reclaim_thread_exit back to
4983 * FALSE when it is finished exiting; we're waiting for that.
4985 while (arc_reclaim_thread_exit
) {
4986 cv_signal(&arc_reclaim_thread_cv
);
4987 cv_wait(&arc_reclaim_thread_cv
, &arc_reclaim_lock
);
4989 mutex_exit(&arc_reclaim_lock
);
4991 mutex_enter(&arc_user_evicts_lock
);
4992 arc_user_evicts_thread_exit
= TRUE
;
4994 * The user evicts thread will set arc_user_evicts_thread_exit
4995 * to FALSE when it is finished exiting; we're waiting for that.
4997 while (arc_user_evicts_thread_exit
) {
4998 cv_signal(&arc_user_evicts_cv
);
4999 cv_wait(&arc_user_evicts_cv
, &arc_user_evicts_lock
);
5001 mutex_exit(&arc_user_evicts_lock
);
5003 /* Use TRUE to ensure *all* buffers are evicted */
5004 arc_flush(NULL
, TRUE
);
5008 if (arc_ksp
!= NULL
) {
5009 kstat_delete(arc_ksp
);
5013 taskq_wait(arc_prune_taskq
);
5014 taskq_destroy(arc_prune_taskq
);
5016 mutex_enter(&arc_prune_mtx
);
5017 while ((p
= list_head(&arc_prune_list
)) != NULL
) {
5018 list_remove(&arc_prune_list
, p
);
5019 refcount_remove(&p
->p_refcnt
, &arc_prune_list
);
5020 refcount_destroy(&p
->p_refcnt
);
5021 kmem_free(p
, sizeof (*p
));
5023 mutex_exit(&arc_prune_mtx
);
5025 list_destroy(&arc_prune_list
);
5026 mutex_destroy(&arc_prune_mtx
);
5027 mutex_destroy(&arc_reclaim_lock
);
5028 cv_destroy(&arc_reclaim_thread_cv
);
5029 cv_destroy(&arc_reclaim_waiters_cv
);
5031 mutex_destroy(&arc_user_evicts_lock
);
5032 cv_destroy(&arc_user_evicts_cv
);
5034 multilist_destroy(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
5035 multilist_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
5036 multilist_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
5037 multilist_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
5038 multilist_destroy(&arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
5039 multilist_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
5040 multilist_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
5041 multilist_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
5042 multilist_destroy(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
]);
5043 multilist_destroy(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
]);
5047 ASSERT0(arc_loaned_bytes
);
5053 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5054 * It uses dedicated storage devices to hold cached data, which are populated
5055 * using large infrequent writes. The main role of this cache is to boost
5056 * the performance of random read workloads. The intended L2ARC devices
5057 * include short-stroked disks, solid state disks, and other media with
5058 * substantially faster read latency than disk.
5060 * +-----------------------+
5062 * +-----------------------+
5065 * l2arc_feed_thread() arc_read()
5069 * +---------------+ |
5071 * +---------------+ |
5076 * +-------+ +-------+
5078 * | cache | | cache |
5079 * +-------+ +-------+
5080 * +=========+ .-----.
5081 * : L2ARC : |-_____-|
5082 * : devices : | Disks |
5083 * +=========+ `-_____-'
5085 * Read requests are satisfied from the following sources, in order:
5088 * 2) vdev cache of L2ARC devices
5090 * 4) vdev cache of disks
5093 * Some L2ARC device types exhibit extremely slow write performance.
5094 * To accommodate for this there are some significant differences between
5095 * the L2ARC and traditional cache design:
5097 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5098 * the ARC behave as usual, freeing buffers and placing headers on ghost
5099 * lists. The ARC does not send buffers to the L2ARC during eviction as
5100 * this would add inflated write latencies for all ARC memory pressure.
5102 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5103 * It does this by periodically scanning buffers from the eviction-end of
5104 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5105 * not already there. It scans until a headroom of buffers is satisfied,
5106 * which itself is a buffer for ARC eviction. If a compressible buffer is
5107 * found during scanning and selected for writing to an L2ARC device, we
5108 * temporarily boost scanning headroom during the next scan cycle to make
5109 * sure we adapt to compression effects (which might significantly reduce
5110 * the data volume we write to L2ARC). The thread that does this is
5111 * l2arc_feed_thread(), illustrated below; example sizes are included to
5112 * provide a better sense of ratio than this diagram:
5115 * +---------------------+----------+
5116 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5117 * +---------------------+----------+ | o L2ARC eligible
5118 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5119 * +---------------------+----------+ |
5120 * 15.9 Gbytes ^ 32 Mbytes |
5122 * l2arc_feed_thread()
5124 * l2arc write hand <--[oooo]--'
5128 * +==============================+
5129 * L2ARC dev |####|#|###|###| |####| ... |
5130 * +==============================+
5133 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5134 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5135 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5136 * safe to say that this is an uncommon case, since buffers at the end of
5137 * the ARC lists have moved there due to inactivity.
5139 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5140 * then the L2ARC simply misses copying some buffers. This serves as a
5141 * pressure valve to prevent heavy read workloads from both stalling the ARC
5142 * with waits and clogging the L2ARC with writes. This also helps prevent
5143 * the potential for the L2ARC to churn if it attempts to cache content too
5144 * quickly, such as during backups of the entire pool.
5146 * 5. After system boot and before the ARC has filled main memory, there are
5147 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5148 * lists can remain mostly static. Instead of searching from tail of these
5149 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5150 * for eligible buffers, greatly increasing its chance of finding them.
5152 * The L2ARC device write speed is also boosted during this time so that
5153 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5154 * there are no L2ARC reads, and no fear of degrading read performance
5155 * through increased writes.
5157 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5158 * the vdev queue can aggregate them into larger and fewer writes. Each
5159 * device is written to in a rotor fashion, sweeping writes through
5160 * available space then repeating.
5162 * 7. The L2ARC does not store dirty content. It never needs to flush
5163 * write buffers back to disk based storage.
5165 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5166 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5168 * The performance of the L2ARC can be tweaked by a number of tunables, which
5169 * may be necessary for different workloads:
5171 * l2arc_write_max max write bytes per interval
5172 * l2arc_write_boost extra write bytes during device warmup
5173 * l2arc_noprefetch skip caching prefetched buffers
5174 * l2arc_nocompress skip compressing buffers
5175 * l2arc_headroom number of max device writes to precache
5176 * l2arc_headroom_boost when we find compressed buffers during ARC
5177 * scanning, we multiply headroom by this
5178 * percentage factor for the next scan cycle,
5179 * since more compressed buffers are likely to
5181 * l2arc_feed_secs seconds between L2ARC writing
5183 * Tunables may be removed or added as future performance improvements are
5184 * integrated, and also may become zpool properties.
5186 * There are three key functions that control how the L2ARC warms up:
5188 * l2arc_write_eligible() check if a buffer is eligible to cache
5189 * l2arc_write_size() calculate how much to write
5190 * l2arc_write_interval() calculate sleep delay between writes
5192 * These three functions determine what to write, how much, and how quickly
5197 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*hdr
)
5200 * A buffer is *not* eligible for the L2ARC if it:
5201 * 1. belongs to a different spa.
5202 * 2. is already cached on the L2ARC.
5203 * 3. has an I/O in progress (it may be an incomplete read).
5204 * 4. is flagged not eligible (zfs property).
5206 if (hdr
->b_spa
!= spa_guid
|| HDR_HAS_L2HDR(hdr
) ||
5207 HDR_IO_IN_PROGRESS(hdr
) || !HDR_L2CACHE(hdr
))
5214 l2arc_write_size(void)
5219 * Make sure our globals have meaningful values in case the user
5222 size
= l2arc_write_max
;
5224 cmn_err(CE_NOTE
, "Bad value for l2arc_write_max, value must "
5225 "be greater than zero, resetting it to the default (%d)",
5227 size
= l2arc_write_max
= L2ARC_WRITE_SIZE
;
5230 if (arc_warm
== B_FALSE
)
5231 size
+= l2arc_write_boost
;
5238 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
5240 clock_t interval
, next
, now
;
5243 * If the ARC lists are busy, increase our write rate; if the
5244 * lists are stale, idle back. This is achieved by checking
5245 * how much we previously wrote - if it was more than half of
5246 * what we wanted, schedule the next write much sooner.
5248 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
5249 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
5251 interval
= hz
* l2arc_feed_secs
;
5253 now
= ddi_get_lbolt();
5254 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
5260 * Cycle through L2ARC devices. This is how L2ARC load balances.
5261 * If a device is returned, this also returns holding the spa config lock.
5263 static l2arc_dev_t
*
5264 l2arc_dev_get_next(void)
5266 l2arc_dev_t
*first
, *next
= NULL
;
5269 * Lock out the removal of spas (spa_namespace_lock), then removal
5270 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
5271 * both locks will be dropped and a spa config lock held instead.
5273 mutex_enter(&spa_namespace_lock
);
5274 mutex_enter(&l2arc_dev_mtx
);
5276 /* if there are no vdevs, there is nothing to do */
5277 if (l2arc_ndev
== 0)
5281 next
= l2arc_dev_last
;
5283 /* loop around the list looking for a non-faulted vdev */
5285 next
= list_head(l2arc_dev_list
);
5287 next
= list_next(l2arc_dev_list
, next
);
5289 next
= list_head(l2arc_dev_list
);
5292 /* if we have come back to the start, bail out */
5295 else if (next
== first
)
5298 } while (vdev_is_dead(next
->l2ad_vdev
));
5300 /* if we were unable to find any usable vdevs, return NULL */
5301 if (vdev_is_dead(next
->l2ad_vdev
))
5304 l2arc_dev_last
= next
;
5307 mutex_exit(&l2arc_dev_mtx
);
5310 * Grab the config lock to prevent the 'next' device from being
5311 * removed while we are writing to it.
5314 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
5315 mutex_exit(&spa_namespace_lock
);
5321 * Free buffers that were tagged for destruction.
5324 l2arc_do_free_on_write(void)
5327 l2arc_data_free_t
*df
, *df_prev
;
5329 mutex_enter(&l2arc_free_on_write_mtx
);
5330 buflist
= l2arc_free_on_write
;
5332 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
5333 df_prev
= list_prev(buflist
, df
);
5334 ASSERT(df
->l2df_data
!= NULL
);
5335 ASSERT(df
->l2df_func
!= NULL
);
5336 df
->l2df_func(df
->l2df_data
, df
->l2df_size
);
5337 list_remove(buflist
, df
);
5338 kmem_free(df
, sizeof (l2arc_data_free_t
));
5341 mutex_exit(&l2arc_free_on_write_mtx
);
5345 * A write to a cache device has completed. Update all headers to allow
5346 * reads from these buffers to begin.
5349 l2arc_write_done(zio_t
*zio
)
5351 l2arc_write_callback_t
*cb
;
5354 arc_buf_hdr_t
*head
, *hdr
, *hdr_prev
;
5355 kmutex_t
*hash_lock
;
5356 int64_t bytes_dropped
= 0;
5358 cb
= zio
->io_private
;
5360 dev
= cb
->l2wcb_dev
;
5361 ASSERT(dev
!= NULL
);
5362 head
= cb
->l2wcb_head
;
5363 ASSERT(head
!= NULL
);
5364 buflist
= &dev
->l2ad_buflist
;
5365 ASSERT(buflist
!= NULL
);
5366 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
5367 l2arc_write_callback_t
*, cb
);
5369 if (zio
->io_error
!= 0)
5370 ARCSTAT_BUMP(arcstat_l2_writes_error
);
5373 * All writes completed, or an error was hit.
5376 mutex_enter(&dev
->l2ad_mtx
);
5377 for (hdr
= list_prev(buflist
, head
); hdr
; hdr
= hdr_prev
) {
5378 hdr_prev
= list_prev(buflist
, hdr
);
5380 hash_lock
= HDR_LOCK(hdr
);
5383 * We cannot use mutex_enter or else we can deadlock
5384 * with l2arc_write_buffers (due to swapping the order
5385 * the hash lock and l2ad_mtx are taken).
5387 if (!mutex_tryenter(hash_lock
)) {
5389 * Missed the hash lock. We must retry so we
5390 * don't leave the ARC_FLAG_L2_WRITING bit set.
5392 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry
);
5395 * We don't want to rescan the headers we've
5396 * already marked as having been written out, so
5397 * we reinsert the head node so we can pick up
5398 * where we left off.
5400 list_remove(buflist
, head
);
5401 list_insert_after(buflist
, hdr
, head
);
5403 mutex_exit(&dev
->l2ad_mtx
);
5406 * We wait for the hash lock to become available
5407 * to try and prevent busy waiting, and increase
5408 * the chance we'll be able to acquire the lock
5409 * the next time around.
5411 mutex_enter(hash_lock
);
5412 mutex_exit(hash_lock
);
5417 * We could not have been moved into the arc_l2c_only
5418 * state while in-flight due to our ARC_FLAG_L2_WRITING
5419 * bit being set. Let's just ensure that's being enforced.
5421 ASSERT(HDR_HAS_L1HDR(hdr
));
5424 * We may have allocated a buffer for L2ARC compression,
5425 * we must release it to avoid leaking this data.
5427 l2arc_release_cdata_buf(hdr
);
5429 if (zio
->io_error
!= 0) {
5431 * Error - drop L2ARC entry.
5433 list_remove(buflist
, hdr
);
5434 hdr
->b_flags
&= ~ARC_FLAG_HAS_L2HDR
;
5436 ARCSTAT_INCR(arcstat_l2_asize
, -hdr
->b_l2hdr
.b_asize
);
5437 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
5441 * Allow ARC to begin reads and ghost list evictions to
5444 hdr
->b_flags
&= ~ARC_FLAG_L2_WRITING
;
5446 mutex_exit(hash_lock
);
5449 atomic_inc_64(&l2arc_writes_done
);
5450 list_remove(buflist
, head
);
5451 ASSERT(!HDR_HAS_L1HDR(head
));
5452 kmem_cache_free(hdr_l2only_cache
, head
);
5453 mutex_exit(&dev
->l2ad_mtx
);
5455 vdev_space_update(dev
->l2ad_vdev
, -bytes_dropped
, 0, 0);
5457 l2arc_do_free_on_write();
5459 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
5463 * A read to a cache device completed. Validate buffer contents before
5464 * handing over to the regular ARC routines.
5467 l2arc_read_done(zio_t
*zio
)
5469 l2arc_read_callback_t
*cb
;
5472 kmutex_t
*hash_lock
;
5475 ASSERT(zio
->io_vd
!= NULL
);
5476 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
5478 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
5480 cb
= zio
->io_private
;
5482 buf
= cb
->l2rcb_buf
;
5483 ASSERT(buf
!= NULL
);
5485 hash_lock
= HDR_LOCK(buf
->b_hdr
);
5486 mutex_enter(hash_lock
);
5488 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
5491 * If the buffer was compressed, decompress it first.
5493 if (cb
->l2rcb_compress
!= ZIO_COMPRESS_OFF
)
5494 l2arc_decompress_zio(zio
, hdr
, cb
->l2rcb_compress
);
5495 ASSERT(zio
->io_data
!= NULL
);
5498 * Check this survived the L2ARC journey.
5500 equal
= arc_cksum_equal(buf
);
5501 if (equal
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
5502 mutex_exit(hash_lock
);
5503 zio
->io_private
= buf
;
5504 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
5505 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
5508 mutex_exit(hash_lock
);
5510 * Buffer didn't survive caching. Increment stats and
5511 * reissue to the original storage device.
5513 if (zio
->io_error
!= 0) {
5514 ARCSTAT_BUMP(arcstat_l2_io_error
);
5516 zio
->io_error
= SET_ERROR(EIO
);
5519 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
5522 * If there's no waiter, issue an async i/o to the primary
5523 * storage now. If there *is* a waiter, the caller must
5524 * issue the i/o in a context where it's OK to block.
5526 if (zio
->io_waiter
== NULL
) {
5527 zio_t
*pio
= zio_unique_parent(zio
);
5529 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
5531 zio_nowait(zio_read(pio
, cb
->l2rcb_spa
, &cb
->l2rcb_bp
,
5532 buf
->b_data
, zio
->io_size
, arc_read_done
, buf
,
5533 zio
->io_priority
, cb
->l2rcb_flags
, &cb
->l2rcb_zb
));
5537 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
5541 * This is the list priority from which the L2ARC will search for pages to
5542 * cache. This is used within loops (0..3) to cycle through lists in the
5543 * desired order. This order can have a significant effect on cache
5546 * Currently the metadata lists are hit first, MFU then MRU, followed by
5547 * the data lists. This function returns a locked list, and also returns
5550 static multilist_sublist_t
*
5551 l2arc_sublist_lock(int list_num
)
5553 multilist_t
*ml
= NULL
;
5556 ASSERT(list_num
>= 0 && list_num
<= 3);
5560 ml
= &arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
5563 ml
= &arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
5566 ml
= &arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
5569 ml
= &arc_mru
->arcs_list
[ARC_BUFC_DATA
];
5574 * Return a randomly-selected sublist. This is acceptable
5575 * because the caller feeds only a little bit of data for each
5576 * call (8MB). Subsequent calls will result in different
5577 * sublists being selected.
5579 idx
= multilist_get_random_index(ml
);
5580 return (multilist_sublist_lock(ml
, idx
));
5584 * Evict buffers from the device write hand to the distance specified in
5585 * bytes. This distance may span populated buffers, it may span nothing.
5586 * This is clearing a region on the L2ARC device ready for writing.
5587 * If the 'all' boolean is set, every buffer is evicted.
5590 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
5593 arc_buf_hdr_t
*hdr
, *hdr_prev
;
5594 kmutex_t
*hash_lock
;
5596 int64_t bytes_evicted
= 0;
5598 buflist
= &dev
->l2ad_buflist
;
5600 if (!all
&& dev
->l2ad_first
) {
5602 * This is the first sweep through the device. There is
5608 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
5610 * When nearing the end of the device, evict to the end
5611 * before the device write hand jumps to the start.
5613 taddr
= dev
->l2ad_end
;
5615 taddr
= dev
->l2ad_hand
+ distance
;
5617 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
5618 uint64_t, taddr
, boolean_t
, all
);
5621 mutex_enter(&dev
->l2ad_mtx
);
5622 for (hdr
= list_tail(buflist
); hdr
; hdr
= hdr_prev
) {
5623 hdr_prev
= list_prev(buflist
, hdr
);
5625 hash_lock
= HDR_LOCK(hdr
);
5628 * We cannot use mutex_enter or else we can deadlock
5629 * with l2arc_write_buffers (due to swapping the order
5630 * the hash lock and l2ad_mtx are taken).
5632 if (!mutex_tryenter(hash_lock
)) {
5634 * Missed the hash lock. Retry.
5636 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
5637 mutex_exit(&dev
->l2ad_mtx
);
5638 mutex_enter(hash_lock
);
5639 mutex_exit(hash_lock
);
5643 if (HDR_L2_WRITE_HEAD(hdr
)) {
5645 * We hit a write head node. Leave it for
5646 * l2arc_write_done().
5648 list_remove(buflist
, hdr
);
5649 mutex_exit(hash_lock
);
5653 if (!all
&& HDR_HAS_L2HDR(hdr
) &&
5654 (hdr
->b_l2hdr
.b_daddr
> taddr
||
5655 hdr
->b_l2hdr
.b_daddr
< dev
->l2ad_hand
)) {
5657 * We've evicted to the target address,
5658 * or the end of the device.
5660 mutex_exit(hash_lock
);
5664 ASSERT(HDR_HAS_L2HDR(hdr
));
5665 if (!HDR_HAS_L1HDR(hdr
)) {
5666 ASSERT(!HDR_L2_READING(hdr
));
5668 * This doesn't exist in the ARC. Destroy.
5669 * arc_hdr_destroy() will call list_remove()
5670 * and decrement arcstat_l2_size.
5672 arc_change_state(arc_anon
, hdr
, hash_lock
);
5673 arc_hdr_destroy(hdr
);
5675 ASSERT(hdr
->b_l1hdr
.b_state
!= arc_l2c_only
);
5676 ARCSTAT_BUMP(arcstat_l2_evict_l1cached
);
5678 * Invalidate issued or about to be issued
5679 * reads, since we may be about to write
5680 * over this location.
5682 if (HDR_L2_READING(hdr
)) {
5683 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
5684 hdr
->b_flags
|= ARC_FLAG_L2_EVICTED
;
5688 * Tell ARC this no longer exists in L2ARC.
5690 /* Tell ARC this no longer exists in L2ARC. */
5691 ARCSTAT_INCR(arcstat_l2_asize
, -hdr
->b_l2hdr
.b_asize
);
5692 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
5693 hdr
->b_flags
&= ~ARC_FLAG_HAS_L2HDR
;
5694 list_remove(buflist
, hdr
);
5696 /* Ensure this header has finished being written */
5697 ASSERT(!HDR_L2_WRITING(hdr
));
5698 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
5700 mutex_exit(hash_lock
);
5702 mutex_exit(&dev
->l2ad_mtx
);
5704 vdev_space_update(dev
->l2ad_vdev
, -bytes_evicted
, 0, 0);
5705 dev
->l2ad_evict
= taddr
;
5709 * Find and write ARC buffers to the L2ARC device.
5711 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
5712 * for reading until they have completed writing.
5713 * The headroom_boost is an in-out parameter used to maintain headroom boost
5714 * state between calls to this function.
5716 * Returns the number of bytes actually written (which may be smaller than
5717 * the delta by which the device hand has changed due to alignment).
5720 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
,
5721 boolean_t
*headroom_boost
)
5723 arc_buf_hdr_t
*hdr
, *hdr_prev
, *head
;
5724 uint64_t write_asize
, write_psize
, write_sz
, headroom
,
5728 l2arc_write_callback_t
*cb
;
5730 uint64_t guid
= spa_load_guid(spa
);
5732 const boolean_t do_headroom_boost
= *headroom_boost
;
5734 ASSERT(dev
->l2ad_vdev
!= NULL
);
5736 /* Lower the flag now, we might want to raise it again later. */
5737 *headroom_boost
= B_FALSE
;
5740 write_sz
= write_asize
= write_psize
= 0;
5742 head
= kmem_cache_alloc(hdr_l2only_cache
, KM_PUSHPAGE
);
5743 head
->b_flags
|= ARC_FLAG_L2_WRITE_HEAD
;
5744 head
->b_flags
|= ARC_FLAG_HAS_L2HDR
;
5747 * We will want to try to compress buffers that are at least 2x the
5748 * device sector size.
5750 buf_compress_minsz
= 2 << dev
->l2ad_vdev
->vdev_ashift
;
5753 * Copy buffers for L2ARC writing.
5755 for (try = 0; try <= 3; try++) {
5756 multilist_sublist_t
*mls
= l2arc_sublist_lock(try);
5757 uint64_t passed_sz
= 0;
5760 * L2ARC fast warmup.
5762 * Until the ARC is warm and starts to evict, read from the
5763 * head of the ARC lists rather than the tail.
5765 if (arc_warm
== B_FALSE
)
5766 hdr
= multilist_sublist_head(mls
);
5768 hdr
= multilist_sublist_tail(mls
);
5770 headroom
= target_sz
* l2arc_headroom
;
5771 if (do_headroom_boost
)
5772 headroom
= (headroom
* l2arc_headroom_boost
) / 100;
5774 for (; hdr
; hdr
= hdr_prev
) {
5775 kmutex_t
*hash_lock
;
5778 if (arc_warm
== B_FALSE
)
5779 hdr_prev
= multilist_sublist_next(mls
, hdr
);
5781 hdr_prev
= multilist_sublist_prev(mls
, hdr
);
5783 hash_lock
= HDR_LOCK(hdr
);
5784 if (!mutex_tryenter(hash_lock
)) {
5786 * Skip this buffer rather than waiting.
5791 passed_sz
+= hdr
->b_size
;
5792 if (passed_sz
> headroom
) {
5796 mutex_exit(hash_lock
);
5800 if (!l2arc_write_eligible(guid
, hdr
)) {
5801 mutex_exit(hash_lock
);
5805 if ((write_sz
+ hdr
->b_size
) > target_sz
) {
5807 mutex_exit(hash_lock
);
5813 * Insert a dummy header on the buflist so
5814 * l2arc_write_done() can find where the
5815 * write buffers begin without searching.
5817 mutex_enter(&dev
->l2ad_mtx
);
5818 list_insert_head(&dev
->l2ad_buflist
, head
);
5819 mutex_exit(&dev
->l2ad_mtx
);
5821 cb
= kmem_alloc(sizeof (l2arc_write_callback_t
),
5823 cb
->l2wcb_dev
= dev
;
5824 cb
->l2wcb_head
= head
;
5825 pio
= zio_root(spa
, l2arc_write_done
, cb
,
5830 * Create and add a new L2ARC header.
5832 hdr
->b_l2hdr
.b_dev
= dev
;
5833 arc_space_consume(HDR_L2ONLY_SIZE
, ARC_SPACE_L2HDRS
);
5834 hdr
->b_flags
|= ARC_FLAG_L2_WRITING
;
5836 * Temporarily stash the data buffer in b_tmp_cdata.
5837 * The subsequent write step will pick it up from
5838 * there. This is because can't access b_l1hdr.b_buf
5839 * without holding the hash_lock, which we in turn
5840 * can't access without holding the ARC list locks
5841 * (which we want to avoid during compression/writing)
5843 HDR_SET_COMPRESS(hdr
, ZIO_COMPRESS_OFF
);
5844 hdr
->b_l2hdr
.b_asize
= hdr
->b_size
;
5845 hdr
->b_l2hdr
.b_hits
= 0;
5846 hdr
->b_l1hdr
.b_tmp_cdata
= hdr
->b_l1hdr
.b_buf
->b_data
;
5848 buf_sz
= hdr
->b_size
;
5849 hdr
->b_flags
|= ARC_FLAG_HAS_L2HDR
;
5851 mutex_enter(&dev
->l2ad_mtx
);
5852 list_insert_head(&dev
->l2ad_buflist
, hdr
);
5853 mutex_exit(&dev
->l2ad_mtx
);
5856 * Compute and store the buffer cksum before
5857 * writing. On debug the cksum is verified first.
5859 arc_cksum_verify(hdr
->b_l1hdr
.b_buf
);
5860 arc_cksum_compute(hdr
->b_l1hdr
.b_buf
, B_TRUE
);
5862 mutex_exit(hash_lock
);
5867 multilist_sublist_unlock(mls
);
5873 /* No buffers selected for writing? */
5876 ASSERT(!HDR_HAS_L1HDR(head
));
5877 kmem_cache_free(hdr_l2only_cache
, head
);
5881 mutex_enter(&dev
->l2ad_mtx
);
5884 * Now start writing the buffers. We're starting at the write head
5885 * and work backwards, retracing the course of the buffer selector
5888 for (hdr
= list_prev(&dev
->l2ad_buflist
, head
); hdr
;
5889 hdr
= list_prev(&dev
->l2ad_buflist
, hdr
)) {
5893 * We rely on the L1 portion of the header below, so
5894 * it's invalid for this header to have been evicted out
5895 * of the ghost cache, prior to being written out. The
5896 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
5898 ASSERT(HDR_HAS_L1HDR(hdr
));
5901 * We shouldn't need to lock the buffer here, since we flagged
5902 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
5903 * take care to only access its L2 cache parameters. In
5904 * particular, hdr->l1hdr.b_buf may be invalid by now due to
5907 hdr
->b_l2hdr
.b_daddr
= dev
->l2ad_hand
;
5909 if ((!l2arc_nocompress
&& HDR_L2COMPRESS(hdr
)) &&
5910 hdr
->b_l2hdr
.b_asize
>= buf_compress_minsz
) {
5911 if (l2arc_compress_buf(hdr
)) {
5913 * If compression succeeded, enable headroom
5914 * boost on the next scan cycle.
5916 *headroom_boost
= B_TRUE
;
5921 * Pick up the buffer data we had previously stashed away
5922 * (and now potentially also compressed).
5924 buf_data
= hdr
->b_l1hdr
.b_tmp_cdata
;
5925 buf_sz
= hdr
->b_l2hdr
.b_asize
;
5927 /* Compression may have squashed the buffer to zero length. */
5931 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
5932 dev
->l2ad_hand
, buf_sz
, buf_data
, ZIO_CHECKSUM_OFF
,
5933 NULL
, NULL
, ZIO_PRIORITY_ASYNC_WRITE
,
5934 ZIO_FLAG_CANFAIL
, B_FALSE
);
5936 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
5938 (void) zio_nowait(wzio
);
5940 write_asize
+= buf_sz
;
5942 * Keep the clock hand suitably device-aligned.
5944 buf_p_sz
= vdev_psize_to_asize(dev
->l2ad_vdev
, buf_sz
);
5945 write_psize
+= buf_p_sz
;
5946 dev
->l2ad_hand
+= buf_p_sz
;
5950 mutex_exit(&dev
->l2ad_mtx
);
5952 ASSERT3U(write_asize
, <=, target_sz
);
5953 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
5954 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_asize
);
5955 ARCSTAT_INCR(arcstat_l2_size
, write_sz
);
5956 ARCSTAT_INCR(arcstat_l2_asize
, write_asize
);
5957 vdev_space_update(dev
->l2ad_vdev
, write_asize
, 0, 0);
5960 * Bump device hand to the device start if it is approaching the end.
5961 * l2arc_evict() will already have evicted ahead for this case.
5963 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
5964 dev
->l2ad_hand
= dev
->l2ad_start
;
5965 dev
->l2ad_evict
= dev
->l2ad_start
;
5966 dev
->l2ad_first
= B_FALSE
;
5969 dev
->l2ad_writing
= B_TRUE
;
5970 (void) zio_wait(pio
);
5971 dev
->l2ad_writing
= B_FALSE
;
5973 return (write_asize
);
5977 * Compresses an L2ARC buffer.
5978 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
5979 * size in l2hdr->b_asize. This routine tries to compress the data and
5980 * depending on the compression result there are three possible outcomes:
5981 * *) The buffer was incompressible. The original l2hdr contents were left
5982 * untouched and are ready for writing to an L2 device.
5983 * *) The buffer was all-zeros, so there is no need to write it to an L2
5984 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
5985 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
5986 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
5987 * data buffer which holds the compressed data to be written, and b_asize
5988 * tells us how much data there is. b_compress is set to the appropriate
5989 * compression algorithm. Once writing is done, invoke
5990 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
5992 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
5993 * buffer was incompressible).
5996 l2arc_compress_buf(arc_buf_hdr_t
*hdr
)
5999 size_t csize
, len
, rounded
;
6000 l2arc_buf_hdr_t
*l2hdr
;
6002 ASSERT(HDR_HAS_L2HDR(hdr
));
6004 l2hdr
= &hdr
->b_l2hdr
;
6006 ASSERT(HDR_HAS_L1HDR(hdr
));
6007 ASSERT(HDR_GET_COMPRESS(hdr
) == ZIO_COMPRESS_OFF
);
6008 ASSERT(hdr
->b_l1hdr
.b_tmp_cdata
!= NULL
);
6010 len
= l2hdr
->b_asize
;
6011 cdata
= zio_data_buf_alloc(len
);
6012 ASSERT3P(cdata
, !=, NULL
);
6013 csize
= zio_compress_data(ZIO_COMPRESS_LZ4
, hdr
->b_l1hdr
.b_tmp_cdata
,
6014 cdata
, l2hdr
->b_asize
);
6016 rounded
= P2ROUNDUP(csize
, (size_t)SPA_MINBLOCKSIZE
);
6017 if (rounded
> csize
) {
6018 bzero((char *)cdata
+ csize
, rounded
- csize
);
6023 /* zero block, indicate that there's nothing to write */
6024 zio_data_buf_free(cdata
, len
);
6025 HDR_SET_COMPRESS(hdr
, ZIO_COMPRESS_EMPTY
);
6027 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6028 ARCSTAT_BUMP(arcstat_l2_compress_zeros
);
6030 } else if (csize
> 0 && csize
< len
) {
6032 * Compression succeeded, we'll keep the cdata around for
6033 * writing and release it afterwards.
6035 HDR_SET_COMPRESS(hdr
, ZIO_COMPRESS_LZ4
);
6036 l2hdr
->b_asize
= csize
;
6037 hdr
->b_l1hdr
.b_tmp_cdata
= cdata
;
6038 ARCSTAT_BUMP(arcstat_l2_compress_successes
);
6042 * Compression failed, release the compressed buffer.
6043 * l2hdr will be left unmodified.
6045 zio_data_buf_free(cdata
, len
);
6046 ARCSTAT_BUMP(arcstat_l2_compress_failures
);
6052 * Decompresses a zio read back from an l2arc device. On success, the
6053 * underlying zio's io_data buffer is overwritten by the uncompressed
6054 * version. On decompression error (corrupt compressed stream), the
6055 * zio->io_error value is set to signal an I/O error.
6057 * Please note that the compressed data stream is not checksummed, so
6058 * if the underlying device is experiencing data corruption, we may feed
6059 * corrupt data to the decompressor, so the decompressor needs to be
6060 * able to handle this situation (LZ4 does).
6063 l2arc_decompress_zio(zio_t
*zio
, arc_buf_hdr_t
*hdr
, enum zio_compress c
)
6068 ASSERT(L2ARC_IS_VALID_COMPRESS(c
));
6070 if (zio
->io_error
!= 0) {
6072 * An io error has occured, just restore the original io
6073 * size in preparation for a main pool read.
6075 zio
->io_orig_size
= zio
->io_size
= hdr
->b_size
;
6079 if (c
== ZIO_COMPRESS_EMPTY
) {
6081 * An empty buffer results in a null zio, which means we
6082 * need to fill its io_data after we're done restoring the
6083 * buffer's contents.
6085 ASSERT(hdr
->b_l1hdr
.b_buf
!= NULL
);
6086 bzero(hdr
->b_l1hdr
.b_buf
->b_data
, hdr
->b_size
);
6087 zio
->io_data
= zio
->io_orig_data
= hdr
->b_l1hdr
.b_buf
->b_data
;
6089 ASSERT(zio
->io_data
!= NULL
);
6091 * We copy the compressed data from the start of the arc buffer
6092 * (the zio_read will have pulled in only what we need, the
6093 * rest is garbage which we will overwrite at decompression)
6094 * and then decompress back to the ARC data buffer. This way we
6095 * can minimize copying by simply decompressing back over the
6096 * original compressed data (rather than decompressing to an
6097 * aux buffer and then copying back the uncompressed buffer,
6098 * which is likely to be much larger).
6100 csize
= zio
->io_size
;
6101 cdata
= zio_data_buf_alloc(csize
);
6102 bcopy(zio
->io_data
, cdata
, csize
);
6103 if (zio_decompress_data(c
, cdata
, zio
->io_data
, csize
,
6105 zio
->io_error
= SET_ERROR(EIO
);
6106 zio_data_buf_free(cdata
, csize
);
6109 /* Restore the expected uncompressed IO size. */
6110 zio
->io_orig_size
= zio
->io_size
= hdr
->b_size
;
6114 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6115 * This buffer serves as a temporary holder of compressed data while
6116 * the buffer entry is being written to an l2arc device. Once that is
6117 * done, we can dispose of it.
6120 l2arc_release_cdata_buf(arc_buf_hdr_t
*hdr
)
6122 enum zio_compress comp
= HDR_GET_COMPRESS(hdr
);
6124 ASSERT(HDR_HAS_L1HDR(hdr
));
6125 ASSERT(comp
== ZIO_COMPRESS_OFF
|| L2ARC_IS_VALID_COMPRESS(comp
));
6127 if (comp
== ZIO_COMPRESS_OFF
) {
6129 * In this case, b_tmp_cdata points to the same buffer
6130 * as the arc_buf_t's b_data field. We don't want to
6131 * free it, since the arc_buf_t will handle that.
6133 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6134 } else if (comp
== ZIO_COMPRESS_EMPTY
) {
6136 * In this case, b_tmp_cdata was compressed to an empty
6137 * buffer, thus there's nothing to free and b_tmp_cdata
6138 * should have been set to NULL in l2arc_write_buffers().
6140 ASSERT3P(hdr
->b_l1hdr
.b_tmp_cdata
, ==, NULL
);
6143 * If the data was compressed, then we've allocated a
6144 * temporary buffer for it, so now we need to release it.
6146 ASSERT(hdr
->b_l1hdr
.b_tmp_cdata
!= NULL
);
6147 zio_data_buf_free(hdr
->b_l1hdr
.b_tmp_cdata
,
6149 hdr
->b_l1hdr
.b_tmp_cdata
= NULL
;
6155 * This thread feeds the L2ARC at regular intervals. This is the beating
6156 * heart of the L2ARC.
6159 l2arc_feed_thread(void)
6164 uint64_t size
, wrote
;
6165 clock_t begin
, next
= ddi_get_lbolt();
6166 boolean_t headroom_boost
= B_FALSE
;
6167 fstrans_cookie_t cookie
;
6169 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
6171 mutex_enter(&l2arc_feed_thr_lock
);
6173 cookie
= spl_fstrans_mark();
6174 while (l2arc_thread_exit
== 0) {
6175 CALLB_CPR_SAFE_BEGIN(&cpr
);
6176 (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv
,
6177 &l2arc_feed_thr_lock
, next
);
6178 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
6179 next
= ddi_get_lbolt() + hz
;
6182 * Quick check for L2ARC devices.
6184 mutex_enter(&l2arc_dev_mtx
);
6185 if (l2arc_ndev
== 0) {
6186 mutex_exit(&l2arc_dev_mtx
);
6189 mutex_exit(&l2arc_dev_mtx
);
6190 begin
= ddi_get_lbolt();
6193 * This selects the next l2arc device to write to, and in
6194 * doing so the next spa to feed from: dev->l2ad_spa. This
6195 * will return NULL if there are now no l2arc devices or if
6196 * they are all faulted.
6198 * If a device is returned, its spa's config lock is also
6199 * held to prevent device removal. l2arc_dev_get_next()
6200 * will grab and release l2arc_dev_mtx.
6202 if ((dev
= l2arc_dev_get_next()) == NULL
)
6205 spa
= dev
->l2ad_spa
;
6206 ASSERT(spa
!= NULL
);
6209 * If the pool is read-only then force the feed thread to
6210 * sleep a little longer.
6212 if (!spa_writeable(spa
)) {
6213 next
= ddi_get_lbolt() + 5 * l2arc_feed_secs
* hz
;
6214 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6219 * Avoid contributing to memory pressure.
6222 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
6223 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6227 ARCSTAT_BUMP(arcstat_l2_feeds
);
6229 size
= l2arc_write_size();
6232 * Evict L2ARC buffers that will be overwritten.
6234 l2arc_evict(dev
, size
, B_FALSE
);
6237 * Write ARC buffers.
6239 wrote
= l2arc_write_buffers(spa
, dev
, size
, &headroom_boost
);
6242 * Calculate interval between writes.
6244 next
= l2arc_write_interval(begin
, size
, wrote
);
6245 spa_config_exit(spa
, SCL_L2ARC
, dev
);
6247 spl_fstrans_unmark(cookie
);
6249 l2arc_thread_exit
= 0;
6250 cv_broadcast(&l2arc_feed_thr_cv
);
6251 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
6256 l2arc_vdev_present(vdev_t
*vd
)
6260 mutex_enter(&l2arc_dev_mtx
);
6261 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
6262 dev
= list_next(l2arc_dev_list
, dev
)) {
6263 if (dev
->l2ad_vdev
== vd
)
6266 mutex_exit(&l2arc_dev_mtx
);
6268 return (dev
!= NULL
);
6272 * Add a vdev for use by the L2ARC. By this point the spa has already
6273 * validated the vdev and opened it.
6276 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
6278 l2arc_dev_t
*adddev
;
6280 ASSERT(!l2arc_vdev_present(vd
));
6283 * Create a new l2arc device entry.
6285 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
6286 adddev
->l2ad_spa
= spa
;
6287 adddev
->l2ad_vdev
= vd
;
6288 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
6289 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
6290 adddev
->l2ad_hand
= adddev
->l2ad_start
;
6291 adddev
->l2ad_evict
= adddev
->l2ad_start
;
6292 adddev
->l2ad_first
= B_TRUE
;
6293 adddev
->l2ad_writing
= B_FALSE
;
6294 list_link_init(&adddev
->l2ad_node
);
6296 mutex_init(&adddev
->l2ad_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6298 * This is a list of all ARC buffers that are still valid on the
6301 list_create(&adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
6302 offsetof(arc_buf_hdr_t
, b_l2hdr
.b_l2node
));
6304 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
6307 * Add device to global list
6309 mutex_enter(&l2arc_dev_mtx
);
6310 list_insert_head(l2arc_dev_list
, adddev
);
6311 atomic_inc_64(&l2arc_ndev
);
6312 mutex_exit(&l2arc_dev_mtx
);
6316 * Remove a vdev from the L2ARC.
6319 l2arc_remove_vdev(vdev_t
*vd
)
6321 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
6324 * Find the device by vdev
6326 mutex_enter(&l2arc_dev_mtx
);
6327 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
6328 nextdev
= list_next(l2arc_dev_list
, dev
);
6329 if (vd
== dev
->l2ad_vdev
) {
6334 ASSERT(remdev
!= NULL
);
6337 * Remove device from global list
6339 list_remove(l2arc_dev_list
, remdev
);
6340 l2arc_dev_last
= NULL
; /* may have been invalidated */
6341 atomic_dec_64(&l2arc_ndev
);
6342 mutex_exit(&l2arc_dev_mtx
);
6345 * Clear all buflists and ARC references. L2ARC device flush.
6347 l2arc_evict(remdev
, 0, B_TRUE
);
6348 list_destroy(&remdev
->l2ad_buflist
);
6349 mutex_destroy(&remdev
->l2ad_mtx
);
6350 kmem_free(remdev
, sizeof (l2arc_dev_t
));
6356 l2arc_thread_exit
= 0;
6358 l2arc_writes_sent
= 0;
6359 l2arc_writes_done
= 0;
6361 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
6362 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
6363 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6364 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
6366 l2arc_dev_list
= &L2ARC_dev_list
;
6367 l2arc_free_on_write
= &L2ARC_free_on_write
;
6368 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
6369 offsetof(l2arc_dev_t
, l2ad_node
));
6370 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
6371 offsetof(l2arc_data_free_t
, l2df_list_node
));
6378 * This is called from dmu_fini(), which is called from spa_fini();
6379 * Because of this, we can assume that all l2arc devices have
6380 * already been removed when the pools themselves were removed.
6383 l2arc_do_free_on_write();
6385 mutex_destroy(&l2arc_feed_thr_lock
);
6386 cv_destroy(&l2arc_feed_thr_cv
);
6387 mutex_destroy(&l2arc_dev_mtx
);
6388 mutex_destroy(&l2arc_free_on_write_mtx
);
6390 list_destroy(l2arc_dev_list
);
6391 list_destroy(l2arc_free_on_write
);
6397 if (!(spa_mode_global
& FWRITE
))
6400 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
6401 TS_RUN
, minclsyspri
);
6407 if (!(spa_mode_global
& FWRITE
))
6410 mutex_enter(&l2arc_feed_thr_lock
);
6411 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
6412 l2arc_thread_exit
= 1;
6413 while (l2arc_thread_exit
!= 0)
6414 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
6415 mutex_exit(&l2arc_feed_thr_lock
);
6418 #if defined(_KERNEL) && defined(HAVE_SPL)
6419 EXPORT_SYMBOL(arc_buf_size
);
6420 EXPORT_SYMBOL(arc_write
);
6421 EXPORT_SYMBOL(arc_read
);
6422 EXPORT_SYMBOL(arc_buf_remove_ref
);
6423 EXPORT_SYMBOL(arc_buf_info
);
6424 EXPORT_SYMBOL(arc_getbuf_func
);
6425 EXPORT_SYMBOL(arc_add_prune_callback
);
6426 EXPORT_SYMBOL(arc_remove_prune_callback
);
6428 module_param(zfs_arc_min
, ulong
, 0644);
6429 MODULE_PARM_DESC(zfs_arc_min
, "Min arc size");
6431 module_param(zfs_arc_max
, ulong
, 0644);
6432 MODULE_PARM_DESC(zfs_arc_max
, "Max arc size");
6434 module_param(zfs_arc_meta_limit
, ulong
, 0644);
6435 MODULE_PARM_DESC(zfs_arc_meta_limit
, "Meta limit for arc size");
6437 module_param(zfs_arc_meta_min
, ulong
, 0644);
6438 MODULE_PARM_DESC(zfs_arc_meta_min
, "Min arc metadata");
6440 module_param(zfs_arc_meta_prune
, int, 0644);
6441 MODULE_PARM_DESC(zfs_arc_meta_prune
, "Meta objects to scan for prune");
6443 module_param(zfs_arc_meta_adjust_restarts
, ulong
, 0644);
6444 MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts
,
6445 "Limit number of restarts in arc_adjust_meta");
6447 module_param(zfs_arc_meta_strategy
, int, 0644);
6448 MODULE_PARM_DESC(zfs_arc_meta_strategy
, "Meta reclaim strategy");
6450 module_param(zfs_arc_grow_retry
, int, 0644);
6451 MODULE_PARM_DESC(zfs_arc_grow_retry
, "Seconds before growing arc size");
6453 module_param(zfs_arc_p_aggressive_disable
, int, 0644);
6454 MODULE_PARM_DESC(zfs_arc_p_aggressive_disable
, "disable aggressive arc_p grow");
6456 module_param(zfs_arc_p_dampener_disable
, int, 0644);
6457 MODULE_PARM_DESC(zfs_arc_p_dampener_disable
, "disable arc_p adapt dampener");
6459 module_param(zfs_arc_shrink_shift
, int, 0644);
6460 MODULE_PARM_DESC(zfs_arc_shrink_shift
, "log2(fraction of arc to reclaim)");
6462 module_param(zfs_disable_dup_eviction
, int, 0644);
6463 MODULE_PARM_DESC(zfs_disable_dup_eviction
, "disable duplicate buffer eviction");
6465 module_param(zfs_arc_average_blocksize
, int, 0444);
6466 MODULE_PARM_DESC(zfs_arc_average_blocksize
, "Target average block size");
6468 module_param(zfs_arc_memory_throttle_disable
, int, 0644);
6469 MODULE_PARM_DESC(zfs_arc_memory_throttle_disable
, "disable memory throttle");
6471 module_param(zfs_arc_min_prefetch_lifespan
, int, 0644);
6472 MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan
, "Min life of prefetch block");
6474 module_param(zfs_arc_num_sublists_per_state
, int, 0644);
6475 MODULE_PARM_DESC(zfs_arc_num_sublists_per_state
,
6476 "Number of sublists used in each of the ARC state lists");
6478 module_param(l2arc_write_max
, ulong
, 0644);
6479 MODULE_PARM_DESC(l2arc_write_max
, "Max write bytes per interval");
6481 module_param(l2arc_write_boost
, ulong
, 0644);
6482 MODULE_PARM_DESC(l2arc_write_boost
, "Extra write bytes during device warmup");
6484 module_param(l2arc_headroom
, ulong
, 0644);
6485 MODULE_PARM_DESC(l2arc_headroom
, "Number of max device writes to precache");
6487 module_param(l2arc_headroom_boost
, ulong
, 0644);
6488 MODULE_PARM_DESC(l2arc_headroom_boost
, "Compressed l2arc_headroom multiplier");
6490 module_param(l2arc_feed_secs
, ulong
, 0644);
6491 MODULE_PARM_DESC(l2arc_feed_secs
, "Seconds between L2ARC writing");
6493 module_param(l2arc_feed_min_ms
, ulong
, 0644);
6494 MODULE_PARM_DESC(l2arc_feed_min_ms
, "Min feed interval in milliseconds");
6496 module_param(l2arc_noprefetch
, int, 0644);
6497 MODULE_PARM_DESC(l2arc_noprefetch
, "Skip caching prefetched buffers");
6499 module_param(l2arc_nocompress
, int, 0644);
6500 MODULE_PARM_DESC(l2arc_nocompress
, "Skip compressing L2ARC buffers");
6502 module_param(l2arc_feed_again
, int, 0644);
6503 MODULE_PARM_DESC(l2arc_feed_again
, "Turbo L2ARC warmup");
6505 module_param(l2arc_norw
, int, 0644);
6506 MODULE_PARM_DESC(l2arc_norw
, "No reads during writes");