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.
26 * DVA-based Adjustable Replacement Cache
28 * While much of the theory of operation used here is
29 * based on the self-tuning, low overhead replacement cache
30 * presented by Megiddo and Modha at FAST 2003, there are some
31 * significant differences:
33 * 1. The Megiddo and Modha model assumes any page is evictable.
34 * Pages in its cache cannot be "locked" into memory. This makes
35 * the eviction algorithm simple: evict the last page in the list.
36 * This also make the performance characteristics easy to reason
37 * about. Our cache is not so simple. At any given moment, some
38 * subset of the blocks in the cache are un-evictable because we
39 * have handed out a reference to them. Blocks are only evictable
40 * when there are no external references active. This makes
41 * eviction far more problematic: we choose to evict the evictable
42 * blocks that are the "lowest" in the list.
44 * There are times when it is not possible to evict the requested
45 * space. In these circumstances we are unable to adjust the cache
46 * size. To prevent the cache growing unbounded at these times we
47 * implement a "cache throttle" that slows the flow of new data
48 * into the cache until we can make space available.
50 * 2. The Megiddo and Modha model assumes a fixed cache size.
51 * Pages are evicted when the cache is full and there is a cache
52 * miss. Our model has a variable sized cache. It grows with
53 * high use, but also tries to react to memory pressure from the
54 * operating system: decreasing its size when system memory is
57 * 3. The Megiddo and Modha model assumes a fixed page size. All
58 * elements of the cache are therefor exactly the same size. So
59 * when adjusting the cache size following a cache miss, its simply
60 * a matter of choosing a single page to evict. In our model, we
61 * have variable sized cache blocks (rangeing from 512 bytes to
62 * 128K bytes). We therefor choose a set of blocks to evict to make
63 * space for a cache miss that approximates as closely as possible
64 * the space used by the new block.
66 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
67 * by N. Megiddo & D. Modha, FAST 2003
73 * A new reference to a cache buffer can be obtained in two
74 * ways: 1) via a hash table lookup using the DVA as a key,
75 * or 2) via one of the ARC lists. The arc_read() interface
76 * uses method 1, while the internal arc algorithms for
77 * adjusting the cache use method 2. We therefor provide two
78 * types of locks: 1) the hash table lock array, and 2) the
81 * Buffers do not have their own mutexs, rather they rely on the
82 * hash table mutexs for the bulk of their protection (i.e. most
83 * fields in the arc_buf_hdr_t are protected by these mutexs).
85 * buf_hash_find() returns the appropriate mutex (held) when it
86 * locates the requested buffer in the hash table. It returns
87 * NULL for the mutex if the buffer was not in the table.
89 * buf_hash_remove() expects the appropriate hash mutex to be
90 * already held before it is invoked.
92 * Each arc state also has a mutex which is used to protect the
93 * buffer list associated with the state. When attempting to
94 * obtain a hash table lock while holding an arc list lock you
95 * must use: mutex_tryenter() to avoid deadlock. Also note that
96 * the active state mutex must be held before the ghost state mutex.
98 * Arc buffers may have an associated eviction callback function.
99 * This function will be invoked prior to removing the buffer (e.g.
100 * in arc_do_user_evicts()). Note however that the data associated
101 * with the buffer may be evicted prior to the callback. The callback
102 * must be made with *no locks held* (to prevent deadlock). Additionally,
103 * the users of callbacks must ensure that their private data is
104 * protected from simultaneous callbacks from arc_buf_evict()
105 * and arc_do_user_evicts().
107 * Note that the majority of the performance stats are manipulated
108 * with atomic operations.
110 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
112 * - L2ARC buflist creation
113 * - L2ARC buflist eviction
114 * - L2ARC write completion, which walks L2ARC buflists
115 * - ARC header destruction, as it removes from L2ARC buflists
116 * - ARC header release, as it removes from L2ARC buflists
121 #include <sys/zfs_context.h>
123 #include <sys/refcount.h>
124 #include <sys/vdev.h>
125 #include <sys/vdev_impl.h>
127 #include <sys/vmsystm.h>
129 #include <sys/fs/swapnode.h>
130 #include <sys/dnlc.h>
132 #include <sys/callb.h>
133 #include <sys/kstat.h>
134 #include <zfs_fletcher.h>
136 static kmutex_t arc_reclaim_thr_lock
;
137 static kcondvar_t arc_reclaim_thr_cv
; /* used to signal reclaim thr */
138 static uint8_t arc_thread_exit
;
140 extern int zfs_write_limit_shift
;
141 extern uint64_t zfs_write_limit_max
;
142 extern kmutex_t zfs_write_limit_lock
;
144 #define ARC_REDUCE_DNLC_PERCENT 3
145 uint_t arc_reduce_dnlc_percent
= ARC_REDUCE_DNLC_PERCENT
;
147 typedef enum arc_reclaim_strategy
{
148 ARC_RECLAIM_AGGR
, /* Aggressive reclaim strategy */
149 ARC_RECLAIM_CONS
/* Conservative reclaim strategy */
150 } arc_reclaim_strategy_t
;
152 /* number of seconds before growing cache again */
153 static int arc_grow_retry
= 60;
155 /* shift of arc_c for calculating both min and max arc_p */
156 static int arc_p_min_shift
= 4;
158 /* log2(fraction of arc to reclaim) */
159 static int arc_shrink_shift
= 5;
162 * minimum lifespan of a prefetch block in clock ticks
163 * (initialized in arc_init())
165 static int arc_min_prefetch_lifespan
;
170 * The arc has filled available memory and has now warmed up.
172 static boolean_t arc_warm
;
175 * These tunables are for performance analysis.
177 uint64_t zfs_arc_max
;
178 uint64_t zfs_arc_min
;
179 uint64_t zfs_arc_meta_limit
= 0;
180 int zfs_arc_grow_retry
= 0;
181 int zfs_arc_shrink_shift
= 0;
182 int zfs_arc_p_min_shift
= 0;
185 * Note that buffers can be in one of 6 states:
186 * ARC_anon - anonymous (discussed below)
187 * ARC_mru - recently used, currently cached
188 * ARC_mru_ghost - recentely used, no longer in cache
189 * ARC_mfu - frequently used, currently cached
190 * ARC_mfu_ghost - frequently used, no longer in cache
191 * ARC_l2c_only - exists in L2ARC but not other states
192 * When there are no active references to the buffer, they are
193 * are linked onto a list in one of these arc states. These are
194 * the only buffers that can be evicted or deleted. Within each
195 * state there are multiple lists, one for meta-data and one for
196 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
197 * etc.) is tracked separately so that it can be managed more
198 * explicitly: favored over data, limited explicitly.
200 * Anonymous buffers are buffers that are not associated with
201 * a DVA. These are buffers that hold dirty block copies
202 * before they are written to stable storage. By definition,
203 * they are "ref'd" and are considered part of arc_mru
204 * that cannot be freed. Generally, they will aquire a DVA
205 * as they are written and migrate onto the arc_mru list.
207 * The ARC_l2c_only state is for buffers that are in the second
208 * level ARC but no longer in any of the ARC_m* lists. The second
209 * level ARC itself may also contain buffers that are in any of
210 * the ARC_m* states - meaning that a buffer can exist in two
211 * places. The reason for the ARC_l2c_only state is to keep the
212 * buffer header in the hash table, so that reads that hit the
213 * second level ARC benefit from these fast lookups.
216 typedef struct arc_state
{
217 list_t arcs_list
[ARC_BUFC_NUMTYPES
]; /* list of evictable buffers */
218 uint64_t arcs_lsize
[ARC_BUFC_NUMTYPES
]; /* amount of evictable data */
219 uint64_t arcs_size
; /* total amount of data in this state */
224 static arc_state_t ARC_anon
;
225 static arc_state_t ARC_mru
;
226 static arc_state_t ARC_mru_ghost
;
227 static arc_state_t ARC_mfu
;
228 static arc_state_t ARC_mfu_ghost
;
229 static arc_state_t ARC_l2c_only
;
231 typedef struct arc_stats
{
232 kstat_named_t arcstat_hits
;
233 kstat_named_t arcstat_misses
;
234 kstat_named_t arcstat_demand_data_hits
;
235 kstat_named_t arcstat_demand_data_misses
;
236 kstat_named_t arcstat_demand_metadata_hits
;
237 kstat_named_t arcstat_demand_metadata_misses
;
238 kstat_named_t arcstat_prefetch_data_hits
;
239 kstat_named_t arcstat_prefetch_data_misses
;
240 kstat_named_t arcstat_prefetch_metadata_hits
;
241 kstat_named_t arcstat_prefetch_metadata_misses
;
242 kstat_named_t arcstat_mru_hits
;
243 kstat_named_t arcstat_mru_ghost_hits
;
244 kstat_named_t arcstat_mfu_hits
;
245 kstat_named_t arcstat_mfu_ghost_hits
;
246 kstat_named_t arcstat_deleted
;
247 kstat_named_t arcstat_recycle_miss
;
248 kstat_named_t arcstat_mutex_miss
;
249 kstat_named_t arcstat_evict_skip
;
250 kstat_named_t arcstat_evict_l2_cached
;
251 kstat_named_t arcstat_evict_l2_eligible
;
252 kstat_named_t arcstat_evict_l2_ineligible
;
253 kstat_named_t arcstat_hash_elements
;
254 kstat_named_t arcstat_hash_elements_max
;
255 kstat_named_t arcstat_hash_collisions
;
256 kstat_named_t arcstat_hash_chains
;
257 kstat_named_t arcstat_hash_chain_max
;
258 kstat_named_t arcstat_p
;
259 kstat_named_t arcstat_c
;
260 kstat_named_t arcstat_c_min
;
261 kstat_named_t arcstat_c_max
;
262 kstat_named_t arcstat_size
;
263 kstat_named_t arcstat_hdr_size
;
264 kstat_named_t arcstat_data_size
;
265 kstat_named_t arcstat_other_size
;
266 kstat_named_t arcstat_l2_hits
;
267 kstat_named_t arcstat_l2_misses
;
268 kstat_named_t arcstat_l2_feeds
;
269 kstat_named_t arcstat_l2_rw_clash
;
270 kstat_named_t arcstat_l2_read_bytes
;
271 kstat_named_t arcstat_l2_write_bytes
;
272 kstat_named_t arcstat_l2_writes_sent
;
273 kstat_named_t arcstat_l2_writes_done
;
274 kstat_named_t arcstat_l2_writes_error
;
275 kstat_named_t arcstat_l2_writes_hdr_miss
;
276 kstat_named_t arcstat_l2_evict_lock_retry
;
277 kstat_named_t arcstat_l2_evict_reading
;
278 kstat_named_t arcstat_l2_free_on_write
;
279 kstat_named_t arcstat_l2_abort_lowmem
;
280 kstat_named_t arcstat_l2_cksum_bad
;
281 kstat_named_t arcstat_l2_io_error
;
282 kstat_named_t arcstat_l2_size
;
283 kstat_named_t arcstat_l2_hdr_size
;
284 kstat_named_t arcstat_memory_throttle_count
;
287 static arc_stats_t arc_stats
= {
288 { "hits", KSTAT_DATA_UINT64
},
289 { "misses", KSTAT_DATA_UINT64
},
290 { "demand_data_hits", KSTAT_DATA_UINT64
},
291 { "demand_data_misses", KSTAT_DATA_UINT64
},
292 { "demand_metadata_hits", KSTAT_DATA_UINT64
},
293 { "demand_metadata_misses", KSTAT_DATA_UINT64
},
294 { "prefetch_data_hits", KSTAT_DATA_UINT64
},
295 { "prefetch_data_misses", KSTAT_DATA_UINT64
},
296 { "prefetch_metadata_hits", KSTAT_DATA_UINT64
},
297 { "prefetch_metadata_misses", KSTAT_DATA_UINT64
},
298 { "mru_hits", KSTAT_DATA_UINT64
},
299 { "mru_ghost_hits", KSTAT_DATA_UINT64
},
300 { "mfu_hits", KSTAT_DATA_UINT64
},
301 { "mfu_ghost_hits", KSTAT_DATA_UINT64
},
302 { "deleted", KSTAT_DATA_UINT64
},
303 { "recycle_miss", KSTAT_DATA_UINT64
},
304 { "mutex_miss", KSTAT_DATA_UINT64
},
305 { "evict_skip", KSTAT_DATA_UINT64
},
306 { "evict_l2_cached", KSTAT_DATA_UINT64
},
307 { "evict_l2_eligible", KSTAT_DATA_UINT64
},
308 { "evict_l2_ineligible", KSTAT_DATA_UINT64
},
309 { "hash_elements", KSTAT_DATA_UINT64
},
310 { "hash_elements_max", KSTAT_DATA_UINT64
},
311 { "hash_collisions", KSTAT_DATA_UINT64
},
312 { "hash_chains", KSTAT_DATA_UINT64
},
313 { "hash_chain_max", KSTAT_DATA_UINT64
},
314 { "p", KSTAT_DATA_UINT64
},
315 { "c", KSTAT_DATA_UINT64
},
316 { "c_min", KSTAT_DATA_UINT64
},
317 { "c_max", KSTAT_DATA_UINT64
},
318 { "size", KSTAT_DATA_UINT64
},
319 { "hdr_size", KSTAT_DATA_UINT64
},
320 { "data_size", KSTAT_DATA_UINT64
},
321 { "other_size", KSTAT_DATA_UINT64
},
322 { "l2_hits", KSTAT_DATA_UINT64
},
323 { "l2_misses", KSTAT_DATA_UINT64
},
324 { "l2_feeds", KSTAT_DATA_UINT64
},
325 { "l2_rw_clash", KSTAT_DATA_UINT64
},
326 { "l2_read_bytes", KSTAT_DATA_UINT64
},
327 { "l2_write_bytes", KSTAT_DATA_UINT64
},
328 { "l2_writes_sent", KSTAT_DATA_UINT64
},
329 { "l2_writes_done", KSTAT_DATA_UINT64
},
330 { "l2_writes_error", KSTAT_DATA_UINT64
},
331 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64
},
332 { "l2_evict_lock_retry", KSTAT_DATA_UINT64
},
333 { "l2_evict_reading", KSTAT_DATA_UINT64
},
334 { "l2_free_on_write", KSTAT_DATA_UINT64
},
335 { "l2_abort_lowmem", KSTAT_DATA_UINT64
},
336 { "l2_cksum_bad", KSTAT_DATA_UINT64
},
337 { "l2_io_error", KSTAT_DATA_UINT64
},
338 { "l2_size", KSTAT_DATA_UINT64
},
339 { "l2_hdr_size", KSTAT_DATA_UINT64
},
340 { "memory_throttle_count", KSTAT_DATA_UINT64
}
343 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
345 #define ARCSTAT_INCR(stat, val) \
346 atomic_add_64(&arc_stats.stat.value.ui64, (val));
348 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
349 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
351 #define ARCSTAT_MAX(stat, val) { \
353 while ((val) > (m = arc_stats.stat.value.ui64) && \
354 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
358 #define ARCSTAT_MAXSTAT(stat) \
359 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
362 * We define a macro to allow ARC hits/misses to be easily broken down by
363 * two separate conditions, giving a total of four different subtypes for
364 * each of hits and misses (so eight statistics total).
366 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
369 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
371 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
375 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
377 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
382 static arc_state_t
*arc_anon
;
383 static arc_state_t
*arc_mru
;
384 static arc_state_t
*arc_mru_ghost
;
385 static arc_state_t
*arc_mfu
;
386 static arc_state_t
*arc_mfu_ghost
;
387 static arc_state_t
*arc_l2c_only
;
390 * There are several ARC variables that are critical to export as kstats --
391 * but we don't want to have to grovel around in the kstat whenever we wish to
392 * manipulate them. For these variables, we therefore define them to be in
393 * terms of the statistic variable. This assures that we are not introducing
394 * the possibility of inconsistency by having shadow copies of the variables,
395 * while still allowing the code to be readable.
397 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
398 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
399 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
400 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
401 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
403 static int arc_no_grow
; /* Don't try to grow cache size */
404 static uint64_t arc_tempreserve
;
405 static uint64_t arc_loaned_bytes
;
406 static uint64_t arc_meta_used
;
407 static uint64_t arc_meta_limit
;
408 static uint64_t arc_meta_max
= 0;
410 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t
;
412 typedef struct arc_callback arc_callback_t
;
414 struct arc_callback
{
416 arc_done_func_t
*acb_done
;
418 zio_t
*acb_zio_dummy
;
419 arc_callback_t
*acb_next
;
422 typedef struct arc_write_callback arc_write_callback_t
;
424 struct arc_write_callback
{
426 arc_done_func_t
*awcb_ready
;
427 arc_done_func_t
*awcb_done
;
432 /* protected by hash lock */
437 kmutex_t b_freeze_lock
;
438 zio_cksum_t
*b_freeze_cksum
;
441 arc_buf_hdr_t
*b_hash_next
;
446 arc_callback_t
*b_acb
;
450 arc_buf_contents_t b_type
;
454 /* protected by arc state mutex */
455 arc_state_t
*b_state
;
456 list_node_t b_arc_node
;
458 /* updated atomically */
459 clock_t b_arc_access
;
461 /* self protecting */
464 l2arc_buf_hdr_t
*b_l2hdr
;
465 list_node_t b_l2node
;
468 static arc_buf_t
*arc_eviction_list
;
469 static kmutex_t arc_eviction_mtx
;
470 static arc_buf_hdr_t arc_eviction_hdr
;
471 static void arc_get_data_buf(arc_buf_t
*buf
);
472 static void arc_access(arc_buf_hdr_t
*buf
, kmutex_t
*hash_lock
);
473 static int arc_evict_needed(arc_buf_contents_t type
);
474 static void arc_evict_ghost(arc_state_t
*state
, uint64_t spa
, int64_t bytes
);
476 static boolean_t
l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*ab
);
478 #define GHOST_STATE(state) \
479 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
480 (state) == arc_l2c_only)
483 * Private ARC flags. These flags are private ARC only flags that will show up
484 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
485 * be passed in as arc_flags in things like arc_read. However, these flags
486 * should never be passed and should only be set by ARC code. When adding new
487 * public flags, make sure not to smash the private ones.
490 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
491 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
492 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
493 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
494 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
495 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
496 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
497 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
498 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
499 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
501 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
502 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
503 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
504 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
505 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
506 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
507 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
508 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
509 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
510 (hdr)->b_l2hdr != NULL)
511 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
512 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
513 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
519 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
520 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
523 * Hash table routines
526 #define HT_LOCK_PAD 64
531 unsigned char pad
[(HT_LOCK_PAD
- sizeof (kmutex_t
))];
535 #define BUF_LOCKS 256
536 typedef struct buf_hash_table
{
538 arc_buf_hdr_t
**ht_table
;
539 struct ht_lock ht_locks
[BUF_LOCKS
];
542 static buf_hash_table_t buf_hash_table
;
544 #define BUF_HASH_INDEX(spa, dva, birth) \
545 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
546 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
547 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
548 #define HDR_LOCK(hdr) \
549 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
551 uint64_t zfs_crc64_table
[256];
557 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
558 #define L2ARC_HEADROOM 2 /* num of writes */
559 #define L2ARC_FEED_SECS 1 /* caching interval secs */
560 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
562 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
563 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
566 * L2ARC Performance Tunables
568 uint64_t l2arc_write_max
= L2ARC_WRITE_SIZE
; /* default max write size */
569 uint64_t l2arc_write_boost
= L2ARC_WRITE_SIZE
; /* extra write during warmup */
570 uint64_t l2arc_headroom
= L2ARC_HEADROOM
; /* number of dev writes */
571 uint64_t l2arc_feed_secs
= L2ARC_FEED_SECS
; /* interval seconds */
572 uint64_t l2arc_feed_min_ms
= L2ARC_FEED_MIN_MS
; /* min interval milliseconds */
573 boolean_t l2arc_noprefetch
= B_TRUE
; /* don't cache prefetch bufs */
574 boolean_t l2arc_feed_again
= B_TRUE
; /* turbo warmup */
575 boolean_t l2arc_norw
= B_TRUE
; /* no reads during writes */
580 typedef struct l2arc_dev
{
581 vdev_t
*l2ad_vdev
; /* vdev */
582 spa_t
*l2ad_spa
; /* spa */
583 uint64_t l2ad_hand
; /* next write location */
584 uint64_t l2ad_write
; /* desired write size, bytes */
585 uint64_t l2ad_boost
; /* warmup write boost, bytes */
586 uint64_t l2ad_start
; /* first addr on device */
587 uint64_t l2ad_end
; /* last addr on device */
588 uint64_t l2ad_evict
; /* last addr eviction reached */
589 boolean_t l2ad_first
; /* first sweep through */
590 boolean_t l2ad_writing
; /* currently writing */
591 list_t
*l2ad_buflist
; /* buffer list */
592 list_node_t l2ad_node
; /* device list node */
595 static list_t L2ARC_dev_list
; /* device list */
596 static list_t
*l2arc_dev_list
; /* device list pointer */
597 static kmutex_t l2arc_dev_mtx
; /* device list mutex */
598 static l2arc_dev_t
*l2arc_dev_last
; /* last device used */
599 static kmutex_t l2arc_buflist_mtx
; /* mutex for all buflists */
600 static list_t L2ARC_free_on_write
; /* free after write buf list */
601 static list_t
*l2arc_free_on_write
; /* free after write list ptr */
602 static kmutex_t l2arc_free_on_write_mtx
; /* mutex for list */
603 static uint64_t l2arc_ndev
; /* number of devices */
605 typedef struct l2arc_read_callback
{
606 arc_buf_t
*l2rcb_buf
; /* read buffer */
607 spa_t
*l2rcb_spa
; /* spa */
608 blkptr_t l2rcb_bp
; /* original blkptr */
609 zbookmark_t l2rcb_zb
; /* original bookmark */
610 int l2rcb_flags
; /* original flags */
611 } l2arc_read_callback_t
;
613 typedef struct l2arc_write_callback
{
614 l2arc_dev_t
*l2wcb_dev
; /* device info */
615 arc_buf_hdr_t
*l2wcb_head
; /* head of write buflist */
616 } l2arc_write_callback_t
;
618 struct l2arc_buf_hdr
{
619 /* protected by arc_buf_hdr mutex */
620 l2arc_dev_t
*b_dev
; /* L2ARC device */
621 uint64_t b_daddr
; /* disk address, offset byte */
624 typedef struct l2arc_data_free
{
625 /* protected by l2arc_free_on_write_mtx */
628 void (*l2df_func
)(void *, size_t);
629 list_node_t l2df_list_node
;
632 static kmutex_t l2arc_feed_thr_lock
;
633 static kcondvar_t l2arc_feed_thr_cv
;
634 static uint8_t l2arc_thread_exit
;
636 static void l2arc_read_done(zio_t
*zio
);
637 static void l2arc_hdr_stat_add(void);
638 static void l2arc_hdr_stat_remove(void);
641 buf_hash(uint64_t spa
, const dva_t
*dva
, uint64_t birth
)
643 uint8_t *vdva
= (uint8_t *)dva
;
644 uint64_t crc
= -1ULL;
647 ASSERT(zfs_crc64_table
[128] == ZFS_CRC64_POLY
);
649 for (i
= 0; i
< sizeof (dva_t
); i
++)
650 crc
= (crc
>> 8) ^ zfs_crc64_table
[(crc
^ vdva
[i
]) & 0xFF];
652 crc
^= (spa
>>8) ^ birth
;
657 #define BUF_EMPTY(buf) \
658 ((buf)->b_dva.dva_word[0] == 0 && \
659 (buf)->b_dva.dva_word[1] == 0 && \
662 #define BUF_EQUAL(spa, dva, birth, buf) \
663 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
664 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
665 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
668 buf_discard_identity(arc_buf_hdr_t
*hdr
)
670 hdr
->b_dva
.dva_word
[0] = 0;
671 hdr
->b_dva
.dva_word
[1] = 0;
676 static arc_buf_hdr_t
*
677 buf_hash_find(uint64_t spa
, const dva_t
*dva
, uint64_t birth
, kmutex_t
**lockp
)
679 uint64_t idx
= BUF_HASH_INDEX(spa
, dva
, birth
);
680 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
683 mutex_enter(hash_lock
);
684 for (buf
= buf_hash_table
.ht_table
[idx
]; buf
!= NULL
;
685 buf
= buf
->b_hash_next
) {
686 if (BUF_EQUAL(spa
, dva
, birth
, buf
)) {
691 mutex_exit(hash_lock
);
697 * Insert an entry into the hash table. If there is already an element
698 * equal to elem in the hash table, then the already existing element
699 * will be returned and the new element will not be inserted.
700 * Otherwise returns NULL.
702 static arc_buf_hdr_t
*
703 buf_hash_insert(arc_buf_hdr_t
*buf
, kmutex_t
**lockp
)
705 uint64_t idx
= BUF_HASH_INDEX(buf
->b_spa
, &buf
->b_dva
, buf
->b_birth
);
706 kmutex_t
*hash_lock
= BUF_HASH_LOCK(idx
);
710 ASSERT(!HDR_IN_HASH_TABLE(buf
));
712 mutex_enter(hash_lock
);
713 for (fbuf
= buf_hash_table
.ht_table
[idx
], i
= 0; fbuf
!= NULL
;
714 fbuf
= fbuf
->b_hash_next
, i
++) {
715 if (BUF_EQUAL(buf
->b_spa
, &buf
->b_dva
, buf
->b_birth
, fbuf
))
719 buf
->b_hash_next
= buf_hash_table
.ht_table
[idx
];
720 buf_hash_table
.ht_table
[idx
] = buf
;
721 buf
->b_flags
|= ARC_IN_HASH_TABLE
;
723 /* collect some hash table performance data */
725 ARCSTAT_BUMP(arcstat_hash_collisions
);
727 ARCSTAT_BUMP(arcstat_hash_chains
);
729 ARCSTAT_MAX(arcstat_hash_chain_max
, i
);
732 ARCSTAT_BUMP(arcstat_hash_elements
);
733 ARCSTAT_MAXSTAT(arcstat_hash_elements
);
739 buf_hash_remove(arc_buf_hdr_t
*buf
)
741 arc_buf_hdr_t
*fbuf
, **bufp
;
742 uint64_t idx
= BUF_HASH_INDEX(buf
->b_spa
, &buf
->b_dva
, buf
->b_birth
);
744 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx
)));
745 ASSERT(HDR_IN_HASH_TABLE(buf
));
747 bufp
= &buf_hash_table
.ht_table
[idx
];
748 while ((fbuf
= *bufp
) != buf
) {
749 ASSERT(fbuf
!= NULL
);
750 bufp
= &fbuf
->b_hash_next
;
752 *bufp
= buf
->b_hash_next
;
753 buf
->b_hash_next
= NULL
;
754 buf
->b_flags
&= ~ARC_IN_HASH_TABLE
;
756 /* collect some hash table performance data */
757 ARCSTAT_BUMPDOWN(arcstat_hash_elements
);
759 if (buf_hash_table
.ht_table
[idx
] &&
760 buf_hash_table
.ht_table
[idx
]->b_hash_next
== NULL
)
761 ARCSTAT_BUMPDOWN(arcstat_hash_chains
);
765 * Global data structures and functions for the buf kmem cache.
767 static kmem_cache_t
*hdr_cache
;
768 static kmem_cache_t
*buf_cache
;
775 kmem_free(buf_hash_table
.ht_table
,
776 (buf_hash_table
.ht_mask
+ 1) * sizeof (void *));
777 for (i
= 0; i
< BUF_LOCKS
; i
++)
778 mutex_destroy(&buf_hash_table
.ht_locks
[i
].ht_lock
);
779 kmem_cache_destroy(hdr_cache
);
780 kmem_cache_destroy(buf_cache
);
784 * Constructor callback - called when the cache is empty
785 * and a new buf is requested.
789 hdr_cons(void *vbuf
, void *unused
, int kmflag
)
791 arc_buf_hdr_t
*buf
= vbuf
;
793 bzero(buf
, sizeof (arc_buf_hdr_t
));
794 refcount_create(&buf
->b_refcnt
);
795 cv_init(&buf
->b_cv
, NULL
, CV_DEFAULT
, NULL
);
796 mutex_init(&buf
->b_freeze_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
797 arc_space_consume(sizeof (arc_buf_hdr_t
), ARC_SPACE_HDRS
);
804 buf_cons(void *vbuf
, void *unused
, int kmflag
)
806 arc_buf_t
*buf
= vbuf
;
808 bzero(buf
, sizeof (arc_buf_t
));
809 mutex_init(&buf
->b_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
810 rw_init(&buf
->b_data_lock
, NULL
, RW_DEFAULT
, NULL
);
811 arc_space_consume(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
817 * Destructor callback - called when a cached buf is
818 * no longer required.
822 hdr_dest(void *vbuf
, void *unused
)
824 arc_buf_hdr_t
*buf
= vbuf
;
826 ASSERT(BUF_EMPTY(buf
));
827 refcount_destroy(&buf
->b_refcnt
);
828 cv_destroy(&buf
->b_cv
);
829 mutex_destroy(&buf
->b_freeze_lock
);
830 arc_space_return(sizeof (arc_buf_hdr_t
), ARC_SPACE_HDRS
);
835 buf_dest(void *vbuf
, void *unused
)
837 arc_buf_t
*buf
= vbuf
;
839 mutex_destroy(&buf
->b_evict_lock
);
840 rw_destroy(&buf
->b_data_lock
);
841 arc_space_return(sizeof (arc_buf_t
), ARC_SPACE_HDRS
);
845 * Reclaim callback -- invoked when memory is low.
849 hdr_recl(void *unused
)
851 dprintf("hdr_recl called\n");
853 * umem calls the reclaim func when we destroy the buf cache,
854 * which is after we do arc_fini().
857 cv_signal(&arc_reclaim_thr_cv
);
864 uint64_t hsize
= 1ULL << 12;
868 * The hash table is big enough to fill all of physical memory
869 * with an average 64K block size. The table will take up
870 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
872 while (hsize
* 65536 < physmem
* PAGESIZE
)
875 buf_hash_table
.ht_mask
= hsize
- 1;
876 buf_hash_table
.ht_table
=
877 kmem_zalloc(hsize
* sizeof (void*), KM_NOSLEEP
);
878 if (buf_hash_table
.ht_table
== NULL
) {
879 ASSERT(hsize
> (1ULL << 8));
884 hdr_cache
= kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t
),
885 0, hdr_cons
, hdr_dest
, hdr_recl
, NULL
, NULL
, 0);
886 buf_cache
= kmem_cache_create("arc_buf_t", sizeof (arc_buf_t
),
887 0, buf_cons
, buf_dest
, NULL
, NULL
, NULL
, 0);
889 for (i
= 0; i
< 256; i
++)
890 for (ct
= zfs_crc64_table
+ i
, *ct
= i
, j
= 8; j
> 0; j
--)
891 *ct
= (*ct
>> 1) ^ (-(*ct
& 1) & ZFS_CRC64_POLY
);
893 for (i
= 0; i
< BUF_LOCKS
; i
++) {
894 mutex_init(&buf_hash_table
.ht_locks
[i
].ht_lock
,
895 NULL
, MUTEX_DEFAULT
, NULL
);
899 #define ARC_MINTIME (hz>>4) /* 62 ms */
902 arc_cksum_verify(arc_buf_t
*buf
)
906 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
909 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
910 if (buf
->b_hdr
->b_freeze_cksum
== NULL
||
911 (buf
->b_hdr
->b_flags
& ARC_IO_ERROR
)) {
912 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
915 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
916 if (!ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
))
917 panic("buffer modified while frozen!");
918 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
922 arc_cksum_equal(arc_buf_t
*buf
)
927 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
928 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
, &zc
);
929 equal
= ZIO_CHECKSUM_EQUAL(*buf
->b_hdr
->b_freeze_cksum
, zc
);
930 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
936 arc_cksum_compute(arc_buf_t
*buf
, boolean_t force
)
938 if (!force
&& !(zfs_flags
& ZFS_DEBUG_MODIFY
))
941 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
942 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
943 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
946 buf
->b_hdr
->b_freeze_cksum
= kmem_alloc(sizeof (zio_cksum_t
), KM_SLEEP
);
947 fletcher_2_native(buf
->b_data
, buf
->b_hdr
->b_size
,
948 buf
->b_hdr
->b_freeze_cksum
);
949 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
953 arc_buf_thaw(arc_buf_t
*buf
)
957 hash_lock
= HDR_LOCK(buf
->b_hdr
);
958 mutex_enter(hash_lock
);
960 if (zfs_flags
& ZFS_DEBUG_MODIFY
) {
961 if (buf
->b_hdr
->b_state
!= arc_anon
)
962 panic("modifying non-anon buffer!");
963 if (buf
->b_hdr
->b_flags
& ARC_IO_IN_PROGRESS
)
964 panic("modifying buffer while i/o in progress!");
965 arc_cksum_verify(buf
);
968 mutex_enter(&buf
->b_hdr
->b_freeze_lock
);
969 if (buf
->b_hdr
->b_freeze_cksum
!= NULL
) {
970 kmem_free(buf
->b_hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
971 buf
->b_hdr
->b_freeze_cksum
= NULL
;
974 if (zfs_flags
& ZFS_DEBUG_MODIFY
) {
975 if (buf
->b_hdr
->b_thawed
)
976 kmem_free(buf
->b_hdr
->b_thawed
, 1);
977 buf
->b_hdr
->b_thawed
= kmem_alloc(1, KM_SLEEP
);
980 mutex_exit(&buf
->b_hdr
->b_freeze_lock
);
981 mutex_exit(hash_lock
);
985 arc_buf_freeze(arc_buf_t
*buf
)
989 if (!(zfs_flags
& ZFS_DEBUG_MODIFY
))
992 hash_lock
= HDR_LOCK(buf
->b_hdr
);
993 mutex_enter(hash_lock
);
995 ASSERT(buf
->b_hdr
->b_freeze_cksum
!= NULL
||
996 buf
->b_hdr
->b_state
== arc_anon
);
997 arc_cksum_compute(buf
, B_FALSE
);
998 mutex_exit(hash_lock
);
1002 add_reference(arc_buf_hdr_t
*ab
, kmutex_t
*hash_lock
, void *tag
)
1004 ASSERT(MUTEX_HELD(hash_lock
));
1006 if ((refcount_add(&ab
->b_refcnt
, tag
) == 1) &&
1007 (ab
->b_state
!= arc_anon
)) {
1008 uint64_t delta
= ab
->b_size
* ab
->b_datacnt
;
1009 list_t
*list
= &ab
->b_state
->arcs_list
[ab
->b_type
];
1010 uint64_t *size
= &ab
->b_state
->arcs_lsize
[ab
->b_type
];
1012 ASSERT(!MUTEX_HELD(&ab
->b_state
->arcs_mtx
));
1013 mutex_enter(&ab
->b_state
->arcs_mtx
);
1014 ASSERT(list_link_active(&ab
->b_arc_node
));
1015 list_remove(list
, ab
);
1016 if (GHOST_STATE(ab
->b_state
)) {
1017 ASSERT3U(ab
->b_datacnt
, ==, 0);
1018 ASSERT3P(ab
->b_buf
, ==, NULL
);
1022 ASSERT3U(*size
, >=, delta
);
1023 atomic_add_64(size
, -delta
);
1024 mutex_exit(&ab
->b_state
->arcs_mtx
);
1025 /* remove the prefetch flag if we get a reference */
1026 if (ab
->b_flags
& ARC_PREFETCH
)
1027 ab
->b_flags
&= ~ARC_PREFETCH
;
1032 remove_reference(arc_buf_hdr_t
*ab
, kmutex_t
*hash_lock
, void *tag
)
1035 arc_state_t
*state
= ab
->b_state
;
1037 ASSERT(state
== arc_anon
|| MUTEX_HELD(hash_lock
));
1038 ASSERT(!GHOST_STATE(state
));
1040 if (((cnt
= refcount_remove(&ab
->b_refcnt
, tag
)) == 0) &&
1041 (state
!= arc_anon
)) {
1042 uint64_t *size
= &state
->arcs_lsize
[ab
->b_type
];
1044 ASSERT(!MUTEX_HELD(&state
->arcs_mtx
));
1045 mutex_enter(&state
->arcs_mtx
);
1046 ASSERT(!list_link_active(&ab
->b_arc_node
));
1047 list_insert_head(&state
->arcs_list
[ab
->b_type
], ab
);
1048 ASSERT(ab
->b_datacnt
> 0);
1049 atomic_add_64(size
, ab
->b_size
* ab
->b_datacnt
);
1050 mutex_exit(&state
->arcs_mtx
);
1056 * Move the supplied buffer to the indicated state. The mutex
1057 * for the buffer must be held by the caller.
1060 arc_change_state(arc_state_t
*new_state
, arc_buf_hdr_t
*ab
, kmutex_t
*hash_lock
)
1062 arc_state_t
*old_state
= ab
->b_state
;
1063 int64_t refcnt
= refcount_count(&ab
->b_refcnt
);
1064 uint64_t from_delta
, to_delta
;
1066 ASSERT(MUTEX_HELD(hash_lock
));
1067 ASSERT(new_state
!= old_state
);
1068 ASSERT(refcnt
== 0 || ab
->b_datacnt
> 0);
1069 ASSERT(ab
->b_datacnt
== 0 || !GHOST_STATE(new_state
));
1070 ASSERT(ab
->b_datacnt
<= 1 || old_state
!= arc_anon
);
1072 from_delta
= to_delta
= ab
->b_datacnt
* ab
->b_size
;
1075 * If this buffer is evictable, transfer it from the
1076 * old state list to the new state list.
1079 if (old_state
!= arc_anon
) {
1080 int use_mutex
= !MUTEX_HELD(&old_state
->arcs_mtx
);
1081 uint64_t *size
= &old_state
->arcs_lsize
[ab
->b_type
];
1084 mutex_enter(&old_state
->arcs_mtx
);
1086 ASSERT(list_link_active(&ab
->b_arc_node
));
1087 list_remove(&old_state
->arcs_list
[ab
->b_type
], ab
);
1090 * If prefetching out of the ghost cache,
1091 * we will have a non-zero datacnt.
1093 if (GHOST_STATE(old_state
) && ab
->b_datacnt
== 0) {
1094 /* ghost elements have a ghost size */
1095 ASSERT(ab
->b_buf
== NULL
);
1096 from_delta
= ab
->b_size
;
1098 ASSERT3U(*size
, >=, from_delta
);
1099 atomic_add_64(size
, -from_delta
);
1102 mutex_exit(&old_state
->arcs_mtx
);
1104 if (new_state
!= arc_anon
) {
1105 int use_mutex
= !MUTEX_HELD(&new_state
->arcs_mtx
);
1106 uint64_t *size
= &new_state
->arcs_lsize
[ab
->b_type
];
1109 mutex_enter(&new_state
->arcs_mtx
);
1111 list_insert_head(&new_state
->arcs_list
[ab
->b_type
], ab
);
1113 /* ghost elements have a ghost size */
1114 if (GHOST_STATE(new_state
)) {
1115 ASSERT(ab
->b_datacnt
== 0);
1116 ASSERT(ab
->b_buf
== NULL
);
1117 to_delta
= ab
->b_size
;
1119 atomic_add_64(size
, to_delta
);
1122 mutex_exit(&new_state
->arcs_mtx
);
1126 ASSERT(!BUF_EMPTY(ab
));
1127 if (new_state
== arc_anon
&& HDR_IN_HASH_TABLE(ab
))
1128 buf_hash_remove(ab
);
1130 /* adjust state sizes */
1132 atomic_add_64(&new_state
->arcs_size
, to_delta
);
1134 ASSERT3U(old_state
->arcs_size
, >=, from_delta
);
1135 atomic_add_64(&old_state
->arcs_size
, -from_delta
);
1137 ab
->b_state
= new_state
;
1139 /* adjust l2arc hdr stats */
1140 if (new_state
== arc_l2c_only
)
1141 l2arc_hdr_stat_add();
1142 else if (old_state
== arc_l2c_only
)
1143 l2arc_hdr_stat_remove();
1147 arc_space_consume(uint64_t space
, arc_space_type_t type
)
1149 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1152 case ARC_SPACE_DATA
:
1153 ARCSTAT_INCR(arcstat_data_size
, space
);
1155 case ARC_SPACE_OTHER
:
1156 ARCSTAT_INCR(arcstat_other_size
, space
);
1158 case ARC_SPACE_HDRS
:
1159 ARCSTAT_INCR(arcstat_hdr_size
, space
);
1161 case ARC_SPACE_L2HDRS
:
1162 ARCSTAT_INCR(arcstat_l2_hdr_size
, space
);
1166 atomic_add_64(&arc_meta_used
, space
);
1167 atomic_add_64(&arc_size
, space
);
1171 arc_space_return(uint64_t space
, arc_space_type_t type
)
1173 ASSERT(type
>= 0 && type
< ARC_SPACE_NUMTYPES
);
1176 case ARC_SPACE_DATA
:
1177 ARCSTAT_INCR(arcstat_data_size
, -space
);
1179 case ARC_SPACE_OTHER
:
1180 ARCSTAT_INCR(arcstat_other_size
, -space
);
1182 case ARC_SPACE_HDRS
:
1183 ARCSTAT_INCR(arcstat_hdr_size
, -space
);
1185 case ARC_SPACE_L2HDRS
:
1186 ARCSTAT_INCR(arcstat_l2_hdr_size
, -space
);
1190 ASSERT(arc_meta_used
>= space
);
1191 if (arc_meta_max
< arc_meta_used
)
1192 arc_meta_max
= arc_meta_used
;
1193 atomic_add_64(&arc_meta_used
, -space
);
1194 ASSERT(arc_size
>= space
);
1195 atomic_add_64(&arc_size
, -space
);
1199 arc_data_buf_alloc(uint64_t size
)
1201 if (arc_evict_needed(ARC_BUFC_DATA
))
1202 cv_signal(&arc_reclaim_thr_cv
);
1203 atomic_add_64(&arc_size
, size
);
1204 return (zio_data_buf_alloc(size
));
1208 arc_data_buf_free(void *buf
, uint64_t size
)
1210 zio_data_buf_free(buf
, size
);
1211 ASSERT(arc_size
>= size
);
1212 atomic_add_64(&arc_size
, -size
);
1216 arc_buf_alloc(spa_t
*spa
, int size
, void *tag
, arc_buf_contents_t type
)
1221 ASSERT3U(size
, >, 0);
1222 hdr
= kmem_cache_alloc(hdr_cache
, KM_PUSHPAGE
);
1223 ASSERT(BUF_EMPTY(hdr
));
1226 hdr
->b_spa
= spa_guid(spa
);
1227 hdr
->b_state
= arc_anon
;
1228 hdr
->b_arc_access
= 0;
1229 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1232 buf
->b_efunc
= NULL
;
1233 buf
->b_private
= NULL
;
1236 arc_get_data_buf(buf
);
1239 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
1240 (void) refcount_add(&hdr
->b_refcnt
, tag
);
1245 static char *arc_onloan_tag
= "onloan";
1248 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1249 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1250 * buffers must be returned to the arc before they can be used by the DMU or
1254 arc_loan_buf(spa_t
*spa
, int size
)
1258 buf
= arc_buf_alloc(spa
, size
, arc_onloan_tag
, ARC_BUFC_DATA
);
1260 atomic_add_64(&arc_loaned_bytes
, size
);
1265 * Return a loaned arc buffer to the arc.
1268 arc_return_buf(arc_buf_t
*buf
, void *tag
)
1270 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1272 ASSERT(buf
->b_data
!= NULL
);
1273 (void) refcount_add(&hdr
->b_refcnt
, tag
);
1274 (void) refcount_remove(&hdr
->b_refcnt
, arc_onloan_tag
);
1276 atomic_add_64(&arc_loaned_bytes
, -hdr
->b_size
);
1279 /* Detach an arc_buf from a dbuf (tag) */
1281 arc_loan_inuse_buf(arc_buf_t
*buf
, void *tag
)
1285 ASSERT(buf
->b_data
!= NULL
);
1287 (void) refcount_add(&hdr
->b_refcnt
, arc_onloan_tag
);
1288 (void) refcount_remove(&hdr
->b_refcnt
, tag
);
1289 buf
->b_efunc
= NULL
;
1290 buf
->b_private
= NULL
;
1292 atomic_add_64(&arc_loaned_bytes
, hdr
->b_size
);
1296 arc_buf_clone(arc_buf_t
*from
)
1299 arc_buf_hdr_t
*hdr
= from
->b_hdr
;
1300 uint64_t size
= hdr
->b_size
;
1302 ASSERT(hdr
->b_state
!= arc_anon
);
1304 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
1307 buf
->b_efunc
= NULL
;
1308 buf
->b_private
= NULL
;
1309 buf
->b_next
= hdr
->b_buf
;
1311 arc_get_data_buf(buf
);
1312 bcopy(from
->b_data
, buf
->b_data
, size
);
1313 hdr
->b_datacnt
+= 1;
1318 arc_buf_add_ref(arc_buf_t
*buf
, void* tag
)
1321 kmutex_t
*hash_lock
;
1324 * Check to see if this buffer is evicted. Callers
1325 * must verify b_data != NULL to know if the add_ref
1328 mutex_enter(&buf
->b_evict_lock
);
1329 if (buf
->b_data
== NULL
) {
1330 mutex_exit(&buf
->b_evict_lock
);
1333 hash_lock
= HDR_LOCK(buf
->b_hdr
);
1334 mutex_enter(hash_lock
);
1336 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1337 mutex_exit(&buf
->b_evict_lock
);
1339 ASSERT(hdr
->b_state
== arc_mru
|| hdr
->b_state
== arc_mfu
);
1340 add_reference(hdr
, hash_lock
, tag
);
1341 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
1342 arc_access(hdr
, hash_lock
);
1343 mutex_exit(hash_lock
);
1344 ARCSTAT_BUMP(arcstat_hits
);
1345 ARCSTAT_CONDSTAT(!(hdr
->b_flags
& ARC_PREFETCH
),
1346 demand
, prefetch
, hdr
->b_type
!= ARC_BUFC_METADATA
,
1347 data
, metadata
, hits
);
1351 * Free the arc data buffer. If it is an l2arc write in progress,
1352 * the buffer is placed on l2arc_free_on_write to be freed later.
1355 arc_buf_data_free(arc_buf_hdr_t
*hdr
, void (*free_func
)(void *, size_t),
1356 void *data
, size_t size
)
1358 if (HDR_L2_WRITING(hdr
)) {
1359 l2arc_data_free_t
*df
;
1360 df
= kmem_alloc(sizeof (l2arc_data_free_t
), KM_SLEEP
);
1361 df
->l2df_data
= data
;
1362 df
->l2df_size
= size
;
1363 df
->l2df_func
= free_func
;
1364 mutex_enter(&l2arc_free_on_write_mtx
);
1365 list_insert_head(l2arc_free_on_write
, df
);
1366 mutex_exit(&l2arc_free_on_write_mtx
);
1367 ARCSTAT_BUMP(arcstat_l2_free_on_write
);
1369 free_func(data
, size
);
1374 arc_buf_destroy(arc_buf_t
*buf
, boolean_t recycle
, boolean_t all
)
1378 /* free up data associated with the buf */
1380 arc_state_t
*state
= buf
->b_hdr
->b_state
;
1381 uint64_t size
= buf
->b_hdr
->b_size
;
1382 arc_buf_contents_t type
= buf
->b_hdr
->b_type
;
1384 arc_cksum_verify(buf
);
1387 if (type
== ARC_BUFC_METADATA
) {
1388 arc_buf_data_free(buf
->b_hdr
, zio_buf_free
,
1390 arc_space_return(size
, ARC_SPACE_DATA
);
1392 ASSERT(type
== ARC_BUFC_DATA
);
1393 arc_buf_data_free(buf
->b_hdr
,
1394 zio_data_buf_free
, buf
->b_data
, size
);
1395 ARCSTAT_INCR(arcstat_data_size
, -size
);
1396 atomic_add_64(&arc_size
, -size
);
1399 if (list_link_active(&buf
->b_hdr
->b_arc_node
)) {
1400 uint64_t *cnt
= &state
->arcs_lsize
[type
];
1402 ASSERT(refcount_is_zero(&buf
->b_hdr
->b_refcnt
));
1403 ASSERT(state
!= arc_anon
);
1405 ASSERT3U(*cnt
, >=, size
);
1406 atomic_add_64(cnt
, -size
);
1408 ASSERT3U(state
->arcs_size
, >=, size
);
1409 atomic_add_64(&state
->arcs_size
, -size
);
1411 ASSERT(buf
->b_hdr
->b_datacnt
> 0);
1412 buf
->b_hdr
->b_datacnt
-= 1;
1415 /* only remove the buf if requested */
1419 /* remove the buf from the hdr list */
1420 for (bufp
= &buf
->b_hdr
->b_buf
; *bufp
!= buf
; bufp
= &(*bufp
)->b_next
)
1422 *bufp
= buf
->b_next
;
1425 ASSERT(buf
->b_efunc
== NULL
);
1427 /* clean up the buf */
1429 kmem_cache_free(buf_cache
, buf
);
1433 arc_hdr_destroy(arc_buf_hdr_t
*hdr
)
1435 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
1436 ASSERT3P(hdr
->b_state
, ==, arc_anon
);
1437 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
1438 l2arc_buf_hdr_t
*l2hdr
= hdr
->b_l2hdr
;
1440 if (l2hdr
!= NULL
) {
1441 boolean_t buflist_held
= MUTEX_HELD(&l2arc_buflist_mtx
);
1443 * To prevent arc_free() and l2arc_evict() from
1444 * attempting to free the same buffer at the same time,
1445 * a FREE_IN_PROGRESS flag is given to arc_free() to
1446 * give it priority. l2arc_evict() can't destroy this
1447 * header while we are waiting on l2arc_buflist_mtx.
1449 * The hdr may be removed from l2ad_buflist before we
1450 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1452 if (!buflist_held
) {
1453 mutex_enter(&l2arc_buflist_mtx
);
1454 l2hdr
= hdr
->b_l2hdr
;
1457 if (l2hdr
!= NULL
) {
1458 list_remove(l2hdr
->b_dev
->l2ad_buflist
, hdr
);
1459 ARCSTAT_INCR(arcstat_l2_size
, -hdr
->b_size
);
1460 kmem_free(l2hdr
, sizeof (l2arc_buf_hdr_t
));
1461 if (hdr
->b_state
== arc_l2c_only
)
1462 l2arc_hdr_stat_remove();
1463 hdr
->b_l2hdr
= NULL
;
1467 mutex_exit(&l2arc_buflist_mtx
);
1470 if (!BUF_EMPTY(hdr
)) {
1471 ASSERT(!HDR_IN_HASH_TABLE(hdr
));
1472 buf_discard_identity(hdr
);
1474 while (hdr
->b_buf
) {
1475 arc_buf_t
*buf
= hdr
->b_buf
;
1478 mutex_enter(&arc_eviction_mtx
);
1479 mutex_enter(&buf
->b_evict_lock
);
1480 ASSERT(buf
->b_hdr
!= NULL
);
1481 arc_buf_destroy(hdr
->b_buf
, FALSE
, FALSE
);
1482 hdr
->b_buf
= buf
->b_next
;
1483 buf
->b_hdr
= &arc_eviction_hdr
;
1484 buf
->b_next
= arc_eviction_list
;
1485 arc_eviction_list
= buf
;
1486 mutex_exit(&buf
->b_evict_lock
);
1487 mutex_exit(&arc_eviction_mtx
);
1489 arc_buf_destroy(hdr
->b_buf
, FALSE
, TRUE
);
1492 if (hdr
->b_freeze_cksum
!= NULL
) {
1493 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
1494 hdr
->b_freeze_cksum
= NULL
;
1496 if (hdr
->b_thawed
) {
1497 kmem_free(hdr
->b_thawed
, 1);
1498 hdr
->b_thawed
= NULL
;
1501 ASSERT(!list_link_active(&hdr
->b_arc_node
));
1502 ASSERT3P(hdr
->b_hash_next
, ==, NULL
);
1503 ASSERT3P(hdr
->b_acb
, ==, NULL
);
1504 kmem_cache_free(hdr_cache
, hdr
);
1508 arc_buf_free(arc_buf_t
*buf
, void *tag
)
1510 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1511 int hashed
= hdr
->b_state
!= arc_anon
;
1513 ASSERT(buf
->b_efunc
== NULL
);
1514 ASSERT(buf
->b_data
!= NULL
);
1517 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
1519 mutex_enter(hash_lock
);
1521 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1523 (void) remove_reference(hdr
, hash_lock
, tag
);
1524 if (hdr
->b_datacnt
> 1) {
1525 arc_buf_destroy(buf
, FALSE
, TRUE
);
1527 ASSERT(buf
== hdr
->b_buf
);
1528 ASSERT(buf
->b_efunc
== NULL
);
1529 hdr
->b_flags
|= ARC_BUF_AVAILABLE
;
1531 mutex_exit(hash_lock
);
1532 } else if (HDR_IO_IN_PROGRESS(hdr
)) {
1535 * We are in the middle of an async write. Don't destroy
1536 * this buffer unless the write completes before we finish
1537 * decrementing the reference count.
1539 mutex_enter(&arc_eviction_mtx
);
1540 (void) remove_reference(hdr
, NULL
, tag
);
1541 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
1542 destroy_hdr
= !HDR_IO_IN_PROGRESS(hdr
);
1543 mutex_exit(&arc_eviction_mtx
);
1545 arc_hdr_destroy(hdr
);
1547 if (remove_reference(hdr
, NULL
, tag
) > 0)
1548 arc_buf_destroy(buf
, FALSE
, TRUE
);
1550 arc_hdr_destroy(hdr
);
1555 arc_buf_remove_ref(arc_buf_t
*buf
, void* tag
)
1557 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
1558 kmutex_t
*hash_lock
= HDR_LOCK(hdr
);
1559 int no_callback
= (buf
->b_efunc
== NULL
);
1561 if (hdr
->b_state
== arc_anon
) {
1562 ASSERT(hdr
->b_datacnt
== 1);
1563 arc_buf_free(buf
, tag
);
1564 return (no_callback
);
1567 mutex_enter(hash_lock
);
1569 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
1570 ASSERT(hdr
->b_state
!= arc_anon
);
1571 ASSERT(buf
->b_data
!= NULL
);
1573 (void) remove_reference(hdr
, hash_lock
, tag
);
1574 if (hdr
->b_datacnt
> 1) {
1576 arc_buf_destroy(buf
, FALSE
, TRUE
);
1577 } else if (no_callback
) {
1578 ASSERT(hdr
->b_buf
== buf
&& buf
->b_next
== NULL
);
1579 ASSERT(buf
->b_efunc
== NULL
);
1580 hdr
->b_flags
|= ARC_BUF_AVAILABLE
;
1582 ASSERT(no_callback
|| hdr
->b_datacnt
> 1 ||
1583 refcount_is_zero(&hdr
->b_refcnt
));
1584 mutex_exit(hash_lock
);
1585 return (no_callback
);
1589 arc_buf_size(arc_buf_t
*buf
)
1591 return (buf
->b_hdr
->b_size
);
1595 * Evict buffers from list until we've removed the specified number of
1596 * bytes. Move the removed buffers to the appropriate evict state.
1597 * If the recycle flag is set, then attempt to "recycle" a buffer:
1598 * - look for a buffer to evict that is `bytes' long.
1599 * - return the data block from this buffer rather than freeing it.
1600 * This flag is used by callers that are trying to make space for a
1601 * new buffer in a full arc cache.
1603 * This function makes a "best effort". It skips over any buffers
1604 * it can't get a hash_lock on, and so may not catch all candidates.
1605 * It may also return without evicting as much space as requested.
1608 arc_evict(arc_state_t
*state
, uint64_t spa
, int64_t bytes
, boolean_t recycle
,
1609 arc_buf_contents_t type
)
1611 arc_state_t
*evicted_state
;
1612 uint64_t bytes_evicted
= 0, skipped
= 0, missed
= 0;
1613 arc_buf_hdr_t
*ab
, *ab_prev
= NULL
;
1614 list_t
*list
= &state
->arcs_list
[type
];
1615 kmutex_t
*hash_lock
;
1616 boolean_t have_lock
;
1617 void *stolen
= NULL
;
1619 ASSERT(state
== arc_mru
|| state
== arc_mfu
);
1621 evicted_state
= (state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
1623 mutex_enter(&state
->arcs_mtx
);
1624 mutex_enter(&evicted_state
->arcs_mtx
);
1626 for (ab
= list_tail(list
); ab
; ab
= ab_prev
) {
1627 ab_prev
= list_prev(list
, ab
);
1628 /* prefetch buffers have a minimum lifespan */
1629 if (HDR_IO_IN_PROGRESS(ab
) ||
1630 (spa
&& ab
->b_spa
!= spa
) ||
1631 (ab
->b_flags
& (ARC_PREFETCH
|ARC_INDIRECT
) &&
1632 ddi_get_lbolt() - ab
->b_arc_access
<
1633 arc_min_prefetch_lifespan
)) {
1637 /* "lookahead" for better eviction candidate */
1638 if (recycle
&& ab
->b_size
!= bytes
&&
1639 ab_prev
&& ab_prev
->b_size
== bytes
)
1641 hash_lock
= HDR_LOCK(ab
);
1642 have_lock
= MUTEX_HELD(hash_lock
);
1643 if (have_lock
|| mutex_tryenter(hash_lock
)) {
1644 ASSERT3U(refcount_count(&ab
->b_refcnt
), ==, 0);
1645 ASSERT(ab
->b_datacnt
> 0);
1647 arc_buf_t
*buf
= ab
->b_buf
;
1648 if (!mutex_tryenter(&buf
->b_evict_lock
)) {
1653 bytes_evicted
+= ab
->b_size
;
1654 if (recycle
&& ab
->b_type
== type
&&
1655 ab
->b_size
== bytes
&&
1656 !HDR_L2_WRITING(ab
)) {
1657 stolen
= buf
->b_data
;
1662 mutex_enter(&arc_eviction_mtx
);
1663 arc_buf_destroy(buf
,
1664 buf
->b_data
== stolen
, FALSE
);
1665 ab
->b_buf
= buf
->b_next
;
1666 buf
->b_hdr
= &arc_eviction_hdr
;
1667 buf
->b_next
= arc_eviction_list
;
1668 arc_eviction_list
= buf
;
1669 mutex_exit(&arc_eviction_mtx
);
1670 mutex_exit(&buf
->b_evict_lock
);
1672 mutex_exit(&buf
->b_evict_lock
);
1673 arc_buf_destroy(buf
,
1674 buf
->b_data
== stolen
, TRUE
);
1679 ARCSTAT_INCR(arcstat_evict_l2_cached
,
1682 if (l2arc_write_eligible(ab
->b_spa
, ab
)) {
1683 ARCSTAT_INCR(arcstat_evict_l2_eligible
,
1687 arcstat_evict_l2_ineligible
,
1692 if (ab
->b_datacnt
== 0) {
1693 arc_change_state(evicted_state
, ab
, hash_lock
);
1694 ASSERT(HDR_IN_HASH_TABLE(ab
));
1695 ab
->b_flags
|= ARC_IN_HASH_TABLE
;
1696 ab
->b_flags
&= ~ARC_BUF_AVAILABLE
;
1697 DTRACE_PROBE1(arc__evict
, arc_buf_hdr_t
*, ab
);
1700 mutex_exit(hash_lock
);
1701 if (bytes
>= 0 && bytes_evicted
>= bytes
)
1708 mutex_exit(&evicted_state
->arcs_mtx
);
1709 mutex_exit(&state
->arcs_mtx
);
1711 if (bytes_evicted
< bytes
)
1712 dprintf("only evicted %lld bytes from %x",
1713 (longlong_t
)bytes_evicted
, state
);
1716 ARCSTAT_INCR(arcstat_evict_skip
, skipped
);
1719 ARCSTAT_INCR(arcstat_mutex_miss
, missed
);
1722 * We have just evicted some date into the ghost state, make
1723 * sure we also adjust the ghost state size if necessary.
1726 arc_mru_ghost
->arcs_size
+ arc_mfu_ghost
->arcs_size
> arc_c
) {
1727 int64_t mru_over
= arc_anon
->arcs_size
+ arc_mru
->arcs_size
+
1728 arc_mru_ghost
->arcs_size
- arc_c
;
1730 if (mru_over
> 0 && arc_mru_ghost
->arcs_lsize
[type
] > 0) {
1732 MIN(arc_mru_ghost
->arcs_lsize
[type
], mru_over
);
1733 arc_evict_ghost(arc_mru_ghost
, NULL
, todelete
);
1734 } else if (arc_mfu_ghost
->arcs_lsize
[type
] > 0) {
1735 int64_t todelete
= MIN(arc_mfu_ghost
->arcs_lsize
[type
],
1736 arc_mru_ghost
->arcs_size
+
1737 arc_mfu_ghost
->arcs_size
- arc_c
);
1738 arc_evict_ghost(arc_mfu_ghost
, NULL
, todelete
);
1746 * Remove buffers from list until we've removed the specified number of
1747 * bytes. Destroy the buffers that are removed.
1750 arc_evict_ghost(arc_state_t
*state
, uint64_t spa
, int64_t bytes
)
1752 arc_buf_hdr_t
*ab
, *ab_prev
;
1753 list_t
*list
= &state
->arcs_list
[ARC_BUFC_DATA
];
1754 kmutex_t
*hash_lock
;
1755 uint64_t bytes_deleted
= 0;
1756 uint64_t bufs_skipped
= 0;
1758 ASSERT(GHOST_STATE(state
));
1760 mutex_enter(&state
->arcs_mtx
);
1761 for (ab
= list_tail(list
); ab
; ab
= ab_prev
) {
1762 ab_prev
= list_prev(list
, ab
);
1763 if (spa
&& ab
->b_spa
!= spa
)
1765 hash_lock
= HDR_LOCK(ab
);
1766 /* caller may be trying to modify this buffer, skip it */
1767 if (MUTEX_HELD(hash_lock
))
1769 if (mutex_tryenter(hash_lock
)) {
1770 ASSERT(!HDR_IO_IN_PROGRESS(ab
));
1771 ASSERT(ab
->b_buf
== NULL
);
1772 ARCSTAT_BUMP(arcstat_deleted
);
1773 bytes_deleted
+= ab
->b_size
;
1775 if (ab
->b_l2hdr
!= NULL
) {
1777 * This buffer is cached on the 2nd Level ARC;
1778 * don't destroy the header.
1780 arc_change_state(arc_l2c_only
, ab
, hash_lock
);
1781 mutex_exit(hash_lock
);
1783 arc_change_state(arc_anon
, ab
, hash_lock
);
1784 mutex_exit(hash_lock
);
1785 arc_hdr_destroy(ab
);
1788 DTRACE_PROBE1(arc__delete
, arc_buf_hdr_t
*, ab
);
1789 if (bytes
>= 0 && bytes_deleted
>= bytes
)
1793 mutex_exit(&state
->arcs_mtx
);
1794 mutex_enter(hash_lock
);
1795 mutex_exit(hash_lock
);
1801 mutex_exit(&state
->arcs_mtx
);
1803 if (list
== &state
->arcs_list
[ARC_BUFC_DATA
] &&
1804 (bytes
< 0 || bytes_deleted
< bytes
)) {
1805 list
= &state
->arcs_list
[ARC_BUFC_METADATA
];
1810 ARCSTAT_INCR(arcstat_mutex_miss
, bufs_skipped
);
1814 if (bytes_deleted
< bytes
)
1815 dprintf("only deleted %lld bytes from %p",
1816 (longlong_t
)bytes_deleted
, state
);
1822 int64_t adjustment
, delta
;
1828 adjustment
= MIN(arc_size
- arc_c
,
1829 arc_anon
->arcs_size
+ arc_mru
->arcs_size
+ arc_meta_used
- arc_p
);
1831 if (adjustment
> 0 && arc_mru
->arcs_lsize
[ARC_BUFC_DATA
] > 0) {
1832 delta
= MIN(arc_mru
->arcs_lsize
[ARC_BUFC_DATA
], adjustment
);
1833 (void) arc_evict(arc_mru
, NULL
, delta
, FALSE
, ARC_BUFC_DATA
);
1834 adjustment
-= delta
;
1837 if (adjustment
> 0 && arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] > 0) {
1838 delta
= MIN(arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
], adjustment
);
1839 (void) arc_evict(arc_mru
, NULL
, delta
, FALSE
,
1847 adjustment
= arc_size
- arc_c
;
1849 if (adjustment
> 0 && arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
] > 0) {
1850 delta
= MIN(adjustment
, arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
]);
1851 (void) arc_evict(arc_mfu
, NULL
, delta
, FALSE
, ARC_BUFC_DATA
);
1852 adjustment
-= delta
;
1855 if (adjustment
> 0 && arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
] > 0) {
1856 int64_t delta
= MIN(adjustment
,
1857 arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
]);
1858 (void) arc_evict(arc_mfu
, NULL
, delta
, FALSE
,
1863 * Adjust ghost lists
1866 adjustment
= arc_mru
->arcs_size
+ arc_mru_ghost
->arcs_size
- arc_c
;
1868 if (adjustment
> 0 && arc_mru_ghost
->arcs_size
> 0) {
1869 delta
= MIN(arc_mru_ghost
->arcs_size
, adjustment
);
1870 arc_evict_ghost(arc_mru_ghost
, NULL
, delta
);
1874 arc_mru_ghost
->arcs_size
+ arc_mfu_ghost
->arcs_size
- arc_c
;
1876 if (adjustment
> 0 && arc_mfu_ghost
->arcs_size
> 0) {
1877 delta
= MIN(arc_mfu_ghost
->arcs_size
, adjustment
);
1878 arc_evict_ghost(arc_mfu_ghost
, NULL
, delta
);
1883 arc_do_user_evicts(void)
1885 mutex_enter(&arc_eviction_mtx
);
1886 while (arc_eviction_list
!= NULL
) {
1887 arc_buf_t
*buf
= arc_eviction_list
;
1888 arc_eviction_list
= buf
->b_next
;
1889 mutex_enter(&buf
->b_evict_lock
);
1891 mutex_exit(&buf
->b_evict_lock
);
1892 mutex_exit(&arc_eviction_mtx
);
1894 if (buf
->b_efunc
!= NULL
)
1895 VERIFY(buf
->b_efunc(buf
) == 0);
1897 buf
->b_efunc
= NULL
;
1898 buf
->b_private
= NULL
;
1899 kmem_cache_free(buf_cache
, buf
);
1900 mutex_enter(&arc_eviction_mtx
);
1902 mutex_exit(&arc_eviction_mtx
);
1906 * Flush all *evictable* data from the cache for the given spa.
1907 * NOTE: this will not touch "active" (i.e. referenced) data.
1910 arc_flush(spa_t
*spa
)
1915 guid
= spa_guid(spa
);
1917 while (list_head(&arc_mru
->arcs_list
[ARC_BUFC_DATA
])) {
1918 (void) arc_evict(arc_mru
, guid
, -1, FALSE
, ARC_BUFC_DATA
);
1922 while (list_head(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
])) {
1923 (void) arc_evict(arc_mru
, guid
, -1, FALSE
, ARC_BUFC_METADATA
);
1927 while (list_head(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
])) {
1928 (void) arc_evict(arc_mfu
, guid
, -1, FALSE
, ARC_BUFC_DATA
);
1932 while (list_head(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
])) {
1933 (void) arc_evict(arc_mfu
, guid
, -1, FALSE
, ARC_BUFC_METADATA
);
1938 arc_evict_ghost(arc_mru_ghost
, guid
, -1);
1939 arc_evict_ghost(arc_mfu_ghost
, guid
, -1);
1941 mutex_enter(&arc_reclaim_thr_lock
);
1942 arc_do_user_evicts();
1943 mutex_exit(&arc_reclaim_thr_lock
);
1944 ASSERT(spa
|| arc_eviction_list
== NULL
);
1950 if (arc_c
> arc_c_min
) {
1954 to_free
= MAX(arc_c
>> arc_shrink_shift
, ptob(needfree
));
1956 to_free
= arc_c
>> arc_shrink_shift
;
1958 if (arc_c
> arc_c_min
+ to_free
)
1959 atomic_add_64(&arc_c
, -to_free
);
1963 atomic_add_64(&arc_p
, -(arc_p
>> arc_shrink_shift
));
1964 if (arc_c
> arc_size
)
1965 arc_c
= MAX(arc_size
, arc_c_min
);
1967 arc_p
= (arc_c
>> 1);
1968 ASSERT(arc_c
>= arc_c_min
);
1969 ASSERT((int64_t)arc_p
>= 0);
1972 if (arc_size
> arc_c
)
1977 arc_reclaim_needed(void)
1987 * take 'desfree' extra pages, so we reclaim sooner, rather than later
1992 * check that we're out of range of the pageout scanner. It starts to
1993 * schedule paging if freemem is less than lotsfree and needfree.
1994 * lotsfree is the high-water mark for pageout, and needfree is the
1995 * number of needed free pages. We add extra pages here to make sure
1996 * the scanner doesn't start up while we're freeing memory.
1998 if (freemem
< lotsfree
+ needfree
+ extra
)
2002 * check to make sure that swapfs has enough space so that anon
2003 * reservations can still succeed. anon_resvmem() checks that the
2004 * availrmem is greater than swapfs_minfree, and the number of reserved
2005 * swap pages. We also add a bit of extra here just to prevent
2006 * circumstances from getting really dire.
2008 if (availrmem
< swapfs_minfree
+ swapfs_reserve
+ extra
)
2013 * If we're on an i386 platform, it's possible that we'll exhaust the
2014 * kernel heap space before we ever run out of available physical
2015 * memory. Most checks of the size of the heap_area compare against
2016 * tune.t_minarmem, which is the minimum available real memory that we
2017 * can have in the system. However, this is generally fixed at 25 pages
2018 * which is so low that it's useless. In this comparison, we seek to
2019 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2020 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2023 if (btop(vmem_size(heap_arena
, VMEM_FREE
)) <
2024 (btop(vmem_size(heap_arena
, VMEM_FREE
| VMEM_ALLOC
)) >> 2))
2029 if (spa_get_random(100) == 0)
2036 arc_kmem_reap_now(arc_reclaim_strategy_t strat
)
2039 kmem_cache_t
*prev_cache
= NULL
;
2040 kmem_cache_t
*prev_data_cache
= NULL
;
2041 extern kmem_cache_t
*zio_buf_cache
[];
2042 extern kmem_cache_t
*zio_data_buf_cache
[];
2045 if (arc_meta_used
>= arc_meta_limit
) {
2047 * We are exceeding our meta-data cache limit.
2048 * Purge some DNLC entries to release holds on meta-data.
2050 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent
);
2054 * Reclaim unused memory from all kmem caches.
2061 * An aggressive reclamation will shrink the cache size as well as
2062 * reap free buffers from the arc kmem caches.
2064 if (strat
== ARC_RECLAIM_AGGR
)
2067 for (i
= 0; i
< SPA_MAXBLOCKSIZE
>> SPA_MINBLOCKSHIFT
; i
++) {
2068 if (zio_buf_cache
[i
] != prev_cache
) {
2069 prev_cache
= zio_buf_cache
[i
];
2070 kmem_cache_reap_now(zio_buf_cache
[i
]);
2072 if (zio_data_buf_cache
[i
] != prev_data_cache
) {
2073 prev_data_cache
= zio_data_buf_cache
[i
];
2074 kmem_cache_reap_now(zio_data_buf_cache
[i
]);
2077 kmem_cache_reap_now(buf_cache
);
2078 kmem_cache_reap_now(hdr_cache
);
2082 arc_reclaim_thread(void)
2084 clock_t growtime
= 0;
2085 arc_reclaim_strategy_t last_reclaim
= ARC_RECLAIM_CONS
;
2088 CALLB_CPR_INIT(&cpr
, &arc_reclaim_thr_lock
, callb_generic_cpr
, FTAG
);
2090 mutex_enter(&arc_reclaim_thr_lock
);
2091 while (arc_thread_exit
== 0) {
2092 if (arc_reclaim_needed()) {
2095 if (last_reclaim
== ARC_RECLAIM_CONS
) {
2096 last_reclaim
= ARC_RECLAIM_AGGR
;
2098 last_reclaim
= ARC_RECLAIM_CONS
;
2102 last_reclaim
= ARC_RECLAIM_AGGR
;
2106 /* reset the growth delay for every reclaim */
2107 growtime
= ddi_get_lbolt() + (arc_grow_retry
* hz
);
2109 arc_kmem_reap_now(last_reclaim
);
2112 } else if (arc_no_grow
&& ddi_get_lbolt() >= growtime
) {
2113 arc_no_grow
= FALSE
;
2116 if (2 * arc_c
< arc_size
+
2117 arc_mru_ghost
->arcs_size
+ arc_mfu_ghost
->arcs_size
)
2120 if (arc_eviction_list
!= NULL
)
2121 arc_do_user_evicts();
2123 /* block until needed, or one second, whichever is shorter */
2124 CALLB_CPR_SAFE_BEGIN(&cpr
);
2125 (void) cv_timedwait(&arc_reclaim_thr_cv
,
2126 &arc_reclaim_thr_lock
, (ddi_get_lbolt() + hz
));
2127 CALLB_CPR_SAFE_END(&cpr
, &arc_reclaim_thr_lock
);
2130 arc_thread_exit
= 0;
2131 cv_broadcast(&arc_reclaim_thr_cv
);
2132 CALLB_CPR_EXIT(&cpr
); /* drops arc_reclaim_thr_lock */
2137 * Adapt arc info given the number of bytes we are trying to add and
2138 * the state that we are comming from. This function is only called
2139 * when we are adding new content to the cache.
2142 arc_adapt(int bytes
, arc_state_t
*state
)
2145 uint64_t arc_p_min
= (arc_c
>> arc_p_min_shift
);
2147 if (state
== arc_l2c_only
)
2152 * Adapt the target size of the MRU list:
2153 * - if we just hit in the MRU ghost list, then increase
2154 * the target size of the MRU list.
2155 * - if we just hit in the MFU ghost list, then increase
2156 * the target size of the MFU list by decreasing the
2157 * target size of the MRU list.
2159 if (state
== arc_mru_ghost
) {
2160 mult
= ((arc_mru_ghost
->arcs_size
>= arc_mfu_ghost
->arcs_size
) ?
2161 1 : (arc_mfu_ghost
->arcs_size
/arc_mru_ghost
->arcs_size
));
2163 arc_p
= MIN(arc_c
- arc_p_min
, arc_p
+ bytes
* mult
);
2164 } else if (state
== arc_mfu_ghost
) {
2167 mult
= ((arc_mfu_ghost
->arcs_size
>= arc_mru_ghost
->arcs_size
) ?
2168 1 : (arc_mru_ghost
->arcs_size
/arc_mfu_ghost
->arcs_size
));
2170 delta
= MIN(bytes
* mult
, arc_p
);
2171 arc_p
= MAX(arc_p_min
, arc_p
- delta
);
2173 ASSERT((int64_t)arc_p
>= 0);
2175 if (arc_reclaim_needed()) {
2176 cv_signal(&arc_reclaim_thr_cv
);
2183 if (arc_c
>= arc_c_max
)
2187 * If we're within (2 * maxblocksize) bytes of the target
2188 * cache size, increment the target cache size
2190 if (arc_size
> arc_c
- (2ULL << SPA_MAXBLOCKSHIFT
)) {
2191 atomic_add_64(&arc_c
, (int64_t)bytes
);
2192 if (arc_c
> arc_c_max
)
2194 else if (state
== arc_anon
)
2195 atomic_add_64(&arc_p
, (int64_t)bytes
);
2199 ASSERT((int64_t)arc_p
>= 0);
2203 * Check if the cache has reached its limits and eviction is required
2207 arc_evict_needed(arc_buf_contents_t type
)
2209 if (type
== ARC_BUFC_METADATA
&& arc_meta_used
>= arc_meta_limit
)
2214 * If zio data pages are being allocated out of a separate heap segment,
2215 * then enforce that the size of available vmem for this area remains
2216 * above about 1/32nd free.
2218 if (type
== ARC_BUFC_DATA
&& zio_arena
!= NULL
&&
2219 vmem_size(zio_arena
, VMEM_FREE
) <
2220 (vmem_size(zio_arena
, VMEM_ALLOC
) >> 5))
2224 if (arc_reclaim_needed())
2227 return (arc_size
> arc_c
);
2231 * The buffer, supplied as the first argument, needs a data block.
2232 * So, if we are at cache max, determine which cache should be victimized.
2233 * We have the following cases:
2235 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2236 * In this situation if we're out of space, but the resident size of the MFU is
2237 * under the limit, victimize the MFU cache to satisfy this insertion request.
2239 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2240 * Here, we've used up all of the available space for the MRU, so we need to
2241 * evict from our own cache instead. Evict from the set of resident MRU
2244 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2245 * c minus p represents the MFU space in the cache, since p is the size of the
2246 * cache that is dedicated to the MRU. In this situation there's still space on
2247 * the MFU side, so the MRU side needs to be victimized.
2249 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2250 * MFU's resident set is consuming more space than it has been allotted. In
2251 * this situation, we must victimize our own cache, the MFU, for this insertion.
2254 arc_get_data_buf(arc_buf_t
*buf
)
2256 arc_state_t
*state
= buf
->b_hdr
->b_state
;
2257 uint64_t size
= buf
->b_hdr
->b_size
;
2258 arc_buf_contents_t type
= buf
->b_hdr
->b_type
;
2260 arc_adapt(size
, state
);
2263 * We have not yet reached cache maximum size,
2264 * just allocate a new buffer.
2266 if (!arc_evict_needed(type
)) {
2267 if (type
== ARC_BUFC_METADATA
) {
2268 buf
->b_data
= zio_buf_alloc(size
);
2269 arc_space_consume(size
, ARC_SPACE_DATA
);
2271 ASSERT(type
== ARC_BUFC_DATA
);
2272 buf
->b_data
= zio_data_buf_alloc(size
);
2273 ARCSTAT_INCR(arcstat_data_size
, size
);
2274 atomic_add_64(&arc_size
, size
);
2280 * If we are prefetching from the mfu ghost list, this buffer
2281 * will end up on the mru list; so steal space from there.
2283 if (state
== arc_mfu_ghost
)
2284 state
= buf
->b_hdr
->b_flags
& ARC_PREFETCH
? arc_mru
: arc_mfu
;
2285 else if (state
== arc_mru_ghost
)
2288 if (state
== arc_mru
|| state
== arc_anon
) {
2289 uint64_t mru_used
= arc_anon
->arcs_size
+ arc_mru
->arcs_size
;
2290 state
= (arc_mfu
->arcs_lsize
[type
] >= size
&&
2291 arc_p
> mru_used
) ? arc_mfu
: arc_mru
;
2294 uint64_t mfu_space
= arc_c
- arc_p
;
2295 state
= (arc_mru
->arcs_lsize
[type
] >= size
&&
2296 mfu_space
> arc_mfu
->arcs_size
) ? arc_mru
: arc_mfu
;
2298 if ((buf
->b_data
= arc_evict(state
, NULL
, size
, TRUE
, type
)) == NULL
) {
2299 if (type
== ARC_BUFC_METADATA
) {
2300 buf
->b_data
= zio_buf_alloc(size
);
2301 arc_space_consume(size
, ARC_SPACE_DATA
);
2303 ASSERT(type
== ARC_BUFC_DATA
);
2304 buf
->b_data
= zio_data_buf_alloc(size
);
2305 ARCSTAT_INCR(arcstat_data_size
, size
);
2306 atomic_add_64(&arc_size
, size
);
2308 ARCSTAT_BUMP(arcstat_recycle_miss
);
2310 ASSERT(buf
->b_data
!= NULL
);
2313 * Update the state size. Note that ghost states have a
2314 * "ghost size" and so don't need to be updated.
2316 if (!GHOST_STATE(buf
->b_hdr
->b_state
)) {
2317 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
2319 atomic_add_64(&hdr
->b_state
->arcs_size
, size
);
2320 if (list_link_active(&hdr
->b_arc_node
)) {
2321 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
2322 atomic_add_64(&hdr
->b_state
->arcs_lsize
[type
], size
);
2325 * If we are growing the cache, and we are adding anonymous
2326 * data, and we have outgrown arc_p, update arc_p
2328 if (arc_size
< arc_c
&& hdr
->b_state
== arc_anon
&&
2329 arc_anon
->arcs_size
+ arc_mru
->arcs_size
> arc_p
)
2330 arc_p
= MIN(arc_c
, arc_p
+ size
);
2335 * This routine is called whenever a buffer is accessed.
2336 * NOTE: the hash lock is dropped in this function.
2339 arc_access(arc_buf_hdr_t
*buf
, kmutex_t
*hash_lock
)
2343 ASSERT(MUTEX_HELD(hash_lock
));
2345 if (buf
->b_state
== arc_anon
) {
2347 * This buffer is not in the cache, and does not
2348 * appear in our "ghost" list. Add the new buffer
2352 ASSERT(buf
->b_arc_access
== 0);
2353 buf
->b_arc_access
= ddi_get_lbolt();
2354 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, buf
);
2355 arc_change_state(arc_mru
, buf
, hash_lock
);
2357 } else if (buf
->b_state
== arc_mru
) {
2358 now
= ddi_get_lbolt();
2361 * If this buffer is here because of a prefetch, then either:
2362 * - clear the flag if this is a "referencing" read
2363 * (any subsequent access will bump this into the MFU state).
2365 * - move the buffer to the head of the list if this is
2366 * another prefetch (to make it less likely to be evicted).
2368 if ((buf
->b_flags
& ARC_PREFETCH
) != 0) {
2369 if (refcount_count(&buf
->b_refcnt
) == 0) {
2370 ASSERT(list_link_active(&buf
->b_arc_node
));
2372 buf
->b_flags
&= ~ARC_PREFETCH
;
2373 ARCSTAT_BUMP(arcstat_mru_hits
);
2375 buf
->b_arc_access
= now
;
2380 * This buffer has been "accessed" only once so far,
2381 * but it is still in the cache. Move it to the MFU
2384 if (now
> buf
->b_arc_access
+ ARC_MINTIME
) {
2386 * More than 125ms have passed since we
2387 * instantiated this buffer. Move it to the
2388 * most frequently used state.
2390 buf
->b_arc_access
= now
;
2391 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2392 arc_change_state(arc_mfu
, buf
, hash_lock
);
2394 ARCSTAT_BUMP(arcstat_mru_hits
);
2395 } else if (buf
->b_state
== arc_mru_ghost
) {
2396 arc_state_t
*new_state
;
2398 * This buffer has been "accessed" recently, but
2399 * was evicted from the cache. Move it to the
2403 if (buf
->b_flags
& ARC_PREFETCH
) {
2404 new_state
= arc_mru
;
2405 if (refcount_count(&buf
->b_refcnt
) > 0)
2406 buf
->b_flags
&= ~ARC_PREFETCH
;
2407 DTRACE_PROBE1(new_state__mru
, arc_buf_hdr_t
*, buf
);
2409 new_state
= arc_mfu
;
2410 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2413 buf
->b_arc_access
= ddi_get_lbolt();
2414 arc_change_state(new_state
, buf
, hash_lock
);
2416 ARCSTAT_BUMP(arcstat_mru_ghost_hits
);
2417 } else if (buf
->b_state
== arc_mfu
) {
2419 * This buffer has been accessed more than once and is
2420 * still in the cache. Keep it in the MFU state.
2422 * NOTE: an add_reference() that occurred when we did
2423 * the arc_read() will have kicked this off the list.
2424 * If it was a prefetch, we will explicitly move it to
2425 * the head of the list now.
2427 if ((buf
->b_flags
& ARC_PREFETCH
) != 0) {
2428 ASSERT(refcount_count(&buf
->b_refcnt
) == 0);
2429 ASSERT(list_link_active(&buf
->b_arc_node
));
2431 ARCSTAT_BUMP(arcstat_mfu_hits
);
2432 buf
->b_arc_access
= ddi_get_lbolt();
2433 } else if (buf
->b_state
== arc_mfu_ghost
) {
2434 arc_state_t
*new_state
= arc_mfu
;
2436 * This buffer has been accessed more than once but has
2437 * been evicted from the cache. Move it back to the
2441 if (buf
->b_flags
& ARC_PREFETCH
) {
2443 * This is a prefetch access...
2444 * move this block back to the MRU state.
2446 ASSERT3U(refcount_count(&buf
->b_refcnt
), ==, 0);
2447 new_state
= arc_mru
;
2450 buf
->b_arc_access
= ddi_get_lbolt();
2451 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2452 arc_change_state(new_state
, buf
, hash_lock
);
2454 ARCSTAT_BUMP(arcstat_mfu_ghost_hits
);
2455 } else if (buf
->b_state
== arc_l2c_only
) {
2457 * This buffer is on the 2nd Level ARC.
2460 buf
->b_arc_access
= ddi_get_lbolt();
2461 DTRACE_PROBE1(new_state__mfu
, arc_buf_hdr_t
*, buf
);
2462 arc_change_state(arc_mfu
, buf
, hash_lock
);
2464 ASSERT(!"invalid arc state");
2468 /* a generic arc_done_func_t which you can use */
2471 arc_bcopy_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
2473 if (zio
== NULL
|| zio
->io_error
== 0)
2474 bcopy(buf
->b_data
, arg
, buf
->b_hdr
->b_size
);
2475 VERIFY(arc_buf_remove_ref(buf
, arg
) == 1);
2478 /* a generic arc_done_func_t */
2480 arc_getbuf_func(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
2482 arc_buf_t
**bufp
= arg
;
2483 if (zio
&& zio
->io_error
) {
2484 VERIFY(arc_buf_remove_ref(buf
, arg
) == 1);
2488 ASSERT(buf
->b_data
);
2493 arc_read_done(zio_t
*zio
)
2495 arc_buf_hdr_t
*hdr
, *found
;
2497 arc_buf_t
*abuf
; /* buffer we're assigning to callback */
2498 kmutex_t
*hash_lock
;
2499 arc_callback_t
*callback_list
, *acb
;
2500 int freeable
= FALSE
;
2502 buf
= zio
->io_private
;
2506 * The hdr was inserted into hash-table and removed from lists
2507 * prior to starting I/O. We should find this header, since
2508 * it's in the hash table, and it should be legit since it's
2509 * not possible to evict it during the I/O. The only possible
2510 * reason for it not to be found is if we were freed during the
2513 found
= buf_hash_find(hdr
->b_spa
, &hdr
->b_dva
, hdr
->b_birth
,
2516 ASSERT((found
== NULL
&& HDR_FREED_IN_READ(hdr
) && hash_lock
== NULL
) ||
2517 (found
== hdr
&& DVA_EQUAL(&hdr
->b_dva
, BP_IDENTITY(zio
->io_bp
))) ||
2518 (found
== hdr
&& HDR_L2_READING(hdr
)));
2520 hdr
->b_flags
&= ~ARC_L2_EVICTED
;
2521 if (l2arc_noprefetch
&& (hdr
->b_flags
& ARC_PREFETCH
))
2522 hdr
->b_flags
&= ~ARC_L2CACHE
;
2524 /* byteswap if necessary */
2525 callback_list
= hdr
->b_acb
;
2526 ASSERT(callback_list
!= NULL
);
2527 if (BP_SHOULD_BYTESWAP(zio
->io_bp
) && zio
->io_error
== 0) {
2528 arc_byteswap_func_t
*func
= BP_GET_LEVEL(zio
->io_bp
) > 0 ?
2529 byteswap_uint64_array
:
2530 dmu_ot
[BP_GET_TYPE(zio
->io_bp
)].ot_byteswap
;
2531 func(buf
->b_data
, hdr
->b_size
);
2534 arc_cksum_compute(buf
, B_FALSE
);
2536 if (hash_lock
&& zio
->io_error
== 0 && hdr
->b_state
== arc_anon
) {
2538 * Only call arc_access on anonymous buffers. This is because
2539 * if we've issued an I/O for an evicted buffer, we've already
2540 * called arc_access (to prevent any simultaneous readers from
2541 * getting confused).
2543 arc_access(hdr
, hash_lock
);
2546 /* create copies of the data buffer for the callers */
2548 for (acb
= callback_list
; acb
; acb
= acb
->acb_next
) {
2549 if (acb
->acb_done
) {
2551 abuf
= arc_buf_clone(buf
);
2552 acb
->acb_buf
= abuf
;
2557 hdr
->b_flags
&= ~ARC_IO_IN_PROGRESS
;
2558 ASSERT(!HDR_BUF_AVAILABLE(hdr
));
2560 ASSERT(buf
->b_efunc
== NULL
);
2561 ASSERT(hdr
->b_datacnt
== 1);
2562 hdr
->b_flags
|= ARC_BUF_AVAILABLE
;
2565 ASSERT(refcount_is_zero(&hdr
->b_refcnt
) || callback_list
!= NULL
);
2567 if (zio
->io_error
!= 0) {
2568 hdr
->b_flags
|= ARC_IO_ERROR
;
2569 if (hdr
->b_state
!= arc_anon
)
2570 arc_change_state(arc_anon
, hdr
, hash_lock
);
2571 if (HDR_IN_HASH_TABLE(hdr
))
2572 buf_hash_remove(hdr
);
2573 freeable
= refcount_is_zero(&hdr
->b_refcnt
);
2577 * Broadcast before we drop the hash_lock to avoid the possibility
2578 * that the hdr (and hence the cv) might be freed before we get to
2579 * the cv_broadcast().
2581 cv_broadcast(&hdr
->b_cv
);
2584 mutex_exit(hash_lock
);
2587 * This block was freed while we waited for the read to
2588 * complete. It has been removed from the hash table and
2589 * moved to the anonymous state (so that it won't show up
2592 ASSERT3P(hdr
->b_state
, ==, arc_anon
);
2593 freeable
= refcount_is_zero(&hdr
->b_refcnt
);
2596 /* execute each callback and free its structure */
2597 while ((acb
= callback_list
) != NULL
) {
2599 acb
->acb_done(zio
, acb
->acb_buf
, acb
->acb_private
);
2601 if (acb
->acb_zio_dummy
!= NULL
) {
2602 acb
->acb_zio_dummy
->io_error
= zio
->io_error
;
2603 zio_nowait(acb
->acb_zio_dummy
);
2606 callback_list
= acb
->acb_next
;
2607 kmem_free(acb
, sizeof (arc_callback_t
));
2611 arc_hdr_destroy(hdr
);
2615 * "Read" the block block at the specified DVA (in bp) via the
2616 * cache. If the block is found in the cache, invoke the provided
2617 * callback immediately and return. Note that the `zio' parameter
2618 * in the callback will be NULL in this case, since no IO was
2619 * required. If the block is not in the cache pass the read request
2620 * on to the spa with a substitute callback function, so that the
2621 * requested block will be added to the cache.
2623 * If a read request arrives for a block that has a read in-progress,
2624 * either wait for the in-progress read to complete (and return the
2625 * results); or, if this is a read with a "done" func, add a record
2626 * to the read to invoke the "done" func when the read completes,
2627 * and return; or just return.
2629 * arc_read_done() will invoke all the requested "done" functions
2630 * for readers of this block.
2632 * Normal callers should use arc_read and pass the arc buffer and offset
2633 * for the bp. But if you know you don't need locking, you can use
2637 arc_read(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
, arc_buf_t
*pbuf
,
2638 arc_done_func_t
*done
, void *private, int priority
, int zio_flags
,
2639 uint32_t *arc_flags
, const zbookmark_t
*zb
)
2645 * XXX This happens from traverse callback funcs, for
2646 * the objset_phys_t block.
2648 return (arc_read_nolock(pio
, spa
, bp
, done
, private, priority
,
2649 zio_flags
, arc_flags
, zb
));
2652 ASSERT(!refcount_is_zero(&pbuf
->b_hdr
->b_refcnt
));
2653 ASSERT3U((char *)bp
- (char *)pbuf
->b_data
, <, pbuf
->b_hdr
->b_size
);
2654 rw_enter(&pbuf
->b_data_lock
, RW_READER
);
2656 err
= arc_read_nolock(pio
, spa
, bp
, done
, private, priority
,
2657 zio_flags
, arc_flags
, zb
);
2658 rw_exit(&pbuf
->b_data_lock
);
2664 arc_read_nolock(zio_t
*pio
, spa_t
*spa
, const blkptr_t
*bp
,
2665 arc_done_func_t
*done
, void *private, int priority
, int zio_flags
,
2666 uint32_t *arc_flags
, const zbookmark_t
*zb
)
2670 kmutex_t
*hash_lock
;
2672 uint64_t guid
= spa_guid(spa
);
2675 hdr
= buf_hash_find(guid
, BP_IDENTITY(bp
), BP_PHYSICAL_BIRTH(bp
),
2677 if (hdr
&& hdr
->b_datacnt
> 0) {
2679 *arc_flags
|= ARC_CACHED
;
2681 if (HDR_IO_IN_PROGRESS(hdr
)) {
2683 if (*arc_flags
& ARC_WAIT
) {
2684 cv_wait(&hdr
->b_cv
, hash_lock
);
2685 mutex_exit(hash_lock
);
2688 ASSERT(*arc_flags
& ARC_NOWAIT
);
2691 arc_callback_t
*acb
= NULL
;
2693 acb
= kmem_zalloc(sizeof (arc_callback_t
),
2695 acb
->acb_done
= done
;
2696 acb
->acb_private
= private;
2698 acb
->acb_zio_dummy
= zio_null(pio
,
2699 spa
, NULL
, NULL
, NULL
, zio_flags
);
2701 ASSERT(acb
->acb_done
!= NULL
);
2702 acb
->acb_next
= hdr
->b_acb
;
2704 add_reference(hdr
, hash_lock
, private);
2705 mutex_exit(hash_lock
);
2708 mutex_exit(hash_lock
);
2712 ASSERT(hdr
->b_state
== arc_mru
|| hdr
->b_state
== arc_mfu
);
2715 add_reference(hdr
, hash_lock
, private);
2717 * If this block is already in use, create a new
2718 * copy of the data so that we will be guaranteed
2719 * that arc_release() will always succeed.
2723 ASSERT(buf
->b_data
);
2724 if (HDR_BUF_AVAILABLE(hdr
)) {
2725 ASSERT(buf
->b_efunc
== NULL
);
2726 hdr
->b_flags
&= ~ARC_BUF_AVAILABLE
;
2728 buf
= arc_buf_clone(buf
);
2731 } else if (*arc_flags
& ARC_PREFETCH
&&
2732 refcount_count(&hdr
->b_refcnt
) == 0) {
2733 hdr
->b_flags
|= ARC_PREFETCH
;
2735 DTRACE_PROBE1(arc__hit
, arc_buf_hdr_t
*, hdr
);
2736 arc_access(hdr
, hash_lock
);
2737 if (*arc_flags
& ARC_L2CACHE
)
2738 hdr
->b_flags
|= ARC_L2CACHE
;
2739 mutex_exit(hash_lock
);
2740 ARCSTAT_BUMP(arcstat_hits
);
2741 ARCSTAT_CONDSTAT(!(hdr
->b_flags
& ARC_PREFETCH
),
2742 demand
, prefetch
, hdr
->b_type
!= ARC_BUFC_METADATA
,
2743 data
, metadata
, hits
);
2746 done(NULL
, buf
, private);
2748 uint64_t size
= BP_GET_LSIZE(bp
);
2749 arc_callback_t
*acb
;
2752 boolean_t devw
= B_FALSE
;
2755 /* this block is not in the cache */
2756 arc_buf_hdr_t
*exists
;
2757 arc_buf_contents_t type
= BP_GET_BUFC_TYPE(bp
);
2758 buf
= arc_buf_alloc(spa
, size
, private, type
);
2760 hdr
->b_dva
= *BP_IDENTITY(bp
);
2761 hdr
->b_birth
= BP_PHYSICAL_BIRTH(bp
);
2762 hdr
->b_cksum0
= bp
->blk_cksum
.zc_word
[0];
2763 exists
= buf_hash_insert(hdr
, &hash_lock
);
2765 /* somebody beat us to the hash insert */
2766 mutex_exit(hash_lock
);
2767 buf_discard_identity(hdr
);
2768 (void) arc_buf_remove_ref(buf
, private);
2769 goto top
; /* restart the IO request */
2771 /* if this is a prefetch, we don't have a reference */
2772 if (*arc_flags
& ARC_PREFETCH
) {
2773 (void) remove_reference(hdr
, hash_lock
,
2775 hdr
->b_flags
|= ARC_PREFETCH
;
2777 if (*arc_flags
& ARC_L2CACHE
)
2778 hdr
->b_flags
|= ARC_L2CACHE
;
2779 if (BP_GET_LEVEL(bp
) > 0)
2780 hdr
->b_flags
|= ARC_INDIRECT
;
2782 /* this block is in the ghost cache */
2783 ASSERT(GHOST_STATE(hdr
->b_state
));
2784 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
2785 ASSERT3U(refcount_count(&hdr
->b_refcnt
), ==, 0);
2786 ASSERT(hdr
->b_buf
== NULL
);
2788 /* if this is a prefetch, we don't have a reference */
2789 if (*arc_flags
& ARC_PREFETCH
)
2790 hdr
->b_flags
|= ARC_PREFETCH
;
2792 add_reference(hdr
, hash_lock
, private);
2793 if (*arc_flags
& ARC_L2CACHE
)
2794 hdr
->b_flags
|= ARC_L2CACHE
;
2795 buf
= kmem_cache_alloc(buf_cache
, KM_PUSHPAGE
);
2798 buf
->b_efunc
= NULL
;
2799 buf
->b_private
= NULL
;
2802 ASSERT(hdr
->b_datacnt
== 0);
2804 arc_get_data_buf(buf
);
2805 arc_access(hdr
, hash_lock
);
2808 ASSERT(!GHOST_STATE(hdr
->b_state
));
2810 acb
= kmem_zalloc(sizeof (arc_callback_t
), KM_SLEEP
);
2811 acb
->acb_done
= done
;
2812 acb
->acb_private
= private;
2814 ASSERT(hdr
->b_acb
== NULL
);
2816 hdr
->b_flags
|= ARC_IO_IN_PROGRESS
;
2818 if (HDR_L2CACHE(hdr
) && hdr
->b_l2hdr
!= NULL
&&
2819 (vd
= hdr
->b_l2hdr
->b_dev
->l2ad_vdev
) != NULL
) {
2820 devw
= hdr
->b_l2hdr
->b_dev
->l2ad_writing
;
2821 addr
= hdr
->b_l2hdr
->b_daddr
;
2823 * Lock out device removal.
2825 if (vdev_is_dead(vd
) ||
2826 !spa_config_tryenter(spa
, SCL_L2ARC
, vd
, RW_READER
))
2830 mutex_exit(hash_lock
);
2832 ASSERT3U(hdr
->b_size
, ==, size
);
2833 DTRACE_PROBE4(arc__miss
, arc_buf_hdr_t
*, hdr
, blkptr_t
*, bp
,
2834 uint64_t, size
, zbookmark_t
*, zb
);
2835 ARCSTAT_BUMP(arcstat_misses
);
2836 ARCSTAT_CONDSTAT(!(hdr
->b_flags
& ARC_PREFETCH
),
2837 demand
, prefetch
, hdr
->b_type
!= ARC_BUFC_METADATA
,
2838 data
, metadata
, misses
);
2840 if (vd
!= NULL
&& l2arc_ndev
!= 0 && !(l2arc_norw
&& devw
)) {
2842 * Read from the L2ARC if the following are true:
2843 * 1. The L2ARC vdev was previously cached.
2844 * 2. This buffer still has L2ARC metadata.
2845 * 3. This buffer isn't currently writing to the L2ARC.
2846 * 4. The L2ARC entry wasn't evicted, which may
2847 * also have invalidated the vdev.
2848 * 5. This isn't prefetch and l2arc_noprefetch is set.
2850 if (hdr
->b_l2hdr
!= NULL
&&
2851 !HDR_L2_WRITING(hdr
) && !HDR_L2_EVICTED(hdr
) &&
2852 !(l2arc_noprefetch
&& HDR_PREFETCH(hdr
))) {
2853 l2arc_read_callback_t
*cb
;
2855 DTRACE_PROBE1(l2arc__hit
, arc_buf_hdr_t
*, hdr
);
2856 ARCSTAT_BUMP(arcstat_l2_hits
);
2858 cb
= kmem_zalloc(sizeof (l2arc_read_callback_t
),
2860 cb
->l2rcb_buf
= buf
;
2861 cb
->l2rcb_spa
= spa
;
2864 cb
->l2rcb_flags
= zio_flags
;
2867 * l2arc read. The SCL_L2ARC lock will be
2868 * released by l2arc_read_done().
2870 rzio
= zio_read_phys(pio
, vd
, addr
, size
,
2871 buf
->b_data
, ZIO_CHECKSUM_OFF
,
2872 l2arc_read_done
, cb
, priority
, zio_flags
|
2873 ZIO_FLAG_DONT_CACHE
| ZIO_FLAG_CANFAIL
|
2874 ZIO_FLAG_DONT_PROPAGATE
|
2875 ZIO_FLAG_DONT_RETRY
, B_FALSE
);
2876 DTRACE_PROBE2(l2arc__read
, vdev_t
*, vd
,
2878 ARCSTAT_INCR(arcstat_l2_read_bytes
, size
);
2880 if (*arc_flags
& ARC_NOWAIT
) {
2885 ASSERT(*arc_flags
& ARC_WAIT
);
2886 if (zio_wait(rzio
) == 0)
2889 /* l2arc read error; goto zio_read() */
2891 DTRACE_PROBE1(l2arc__miss
,
2892 arc_buf_hdr_t
*, hdr
);
2893 ARCSTAT_BUMP(arcstat_l2_misses
);
2894 if (HDR_L2_WRITING(hdr
))
2895 ARCSTAT_BUMP(arcstat_l2_rw_clash
);
2896 spa_config_exit(spa
, SCL_L2ARC
, vd
);
2900 spa_config_exit(spa
, SCL_L2ARC
, vd
);
2901 if (l2arc_ndev
!= 0) {
2902 DTRACE_PROBE1(l2arc__miss
,
2903 arc_buf_hdr_t
*, hdr
);
2904 ARCSTAT_BUMP(arcstat_l2_misses
);
2908 rzio
= zio_read(pio
, spa
, bp
, buf
->b_data
, size
,
2909 arc_read_done
, buf
, priority
, zio_flags
, zb
);
2911 if (*arc_flags
& ARC_WAIT
)
2912 return (zio_wait(rzio
));
2914 ASSERT(*arc_flags
& ARC_NOWAIT
);
2921 arc_set_callback(arc_buf_t
*buf
, arc_evict_func_t
*func
, void *private)
2923 ASSERT(buf
->b_hdr
!= NULL
);
2924 ASSERT(buf
->b_hdr
->b_state
!= arc_anon
);
2925 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_refcnt
) || func
== NULL
);
2926 ASSERT(buf
->b_efunc
== NULL
);
2927 ASSERT(!HDR_BUF_AVAILABLE(buf
->b_hdr
));
2929 buf
->b_efunc
= func
;
2930 buf
->b_private
= private;
2934 * This is used by the DMU to let the ARC know that a buffer is
2935 * being evicted, so the ARC should clean up. If this arc buf
2936 * is not yet in the evicted state, it will be put there.
2939 arc_buf_evict(arc_buf_t
*buf
)
2942 kmutex_t
*hash_lock
;
2945 mutex_enter(&buf
->b_evict_lock
);
2949 * We are in arc_do_user_evicts().
2951 ASSERT(buf
->b_data
== NULL
);
2952 mutex_exit(&buf
->b_evict_lock
);
2954 } else if (buf
->b_data
== NULL
) {
2955 arc_buf_t copy
= *buf
; /* structure assignment */
2957 * We are on the eviction list; process this buffer now
2958 * but let arc_do_user_evicts() do the reaping.
2960 buf
->b_efunc
= NULL
;
2961 mutex_exit(&buf
->b_evict_lock
);
2962 VERIFY(copy
.b_efunc(©
) == 0);
2965 hash_lock
= HDR_LOCK(hdr
);
2966 mutex_enter(hash_lock
);
2968 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
2970 ASSERT3U(refcount_count(&hdr
->b_refcnt
), <, hdr
->b_datacnt
);
2971 ASSERT(hdr
->b_state
== arc_mru
|| hdr
->b_state
== arc_mfu
);
2974 * Pull this buffer off of the hdr
2977 while (*bufp
!= buf
)
2978 bufp
= &(*bufp
)->b_next
;
2979 *bufp
= buf
->b_next
;
2981 ASSERT(buf
->b_data
!= NULL
);
2982 arc_buf_destroy(buf
, FALSE
, FALSE
);
2984 if (hdr
->b_datacnt
== 0) {
2985 arc_state_t
*old_state
= hdr
->b_state
;
2986 arc_state_t
*evicted_state
;
2988 ASSERT(hdr
->b_buf
== NULL
);
2989 ASSERT(refcount_is_zero(&hdr
->b_refcnt
));
2992 (old_state
== arc_mru
) ? arc_mru_ghost
: arc_mfu_ghost
;
2994 mutex_enter(&old_state
->arcs_mtx
);
2995 mutex_enter(&evicted_state
->arcs_mtx
);
2997 arc_change_state(evicted_state
, hdr
, hash_lock
);
2998 ASSERT(HDR_IN_HASH_TABLE(hdr
));
2999 hdr
->b_flags
|= ARC_IN_HASH_TABLE
;
3000 hdr
->b_flags
&= ~ARC_BUF_AVAILABLE
;
3002 mutex_exit(&evicted_state
->arcs_mtx
);
3003 mutex_exit(&old_state
->arcs_mtx
);
3005 mutex_exit(hash_lock
);
3006 mutex_exit(&buf
->b_evict_lock
);
3008 VERIFY(buf
->b_efunc(buf
) == 0);
3009 buf
->b_efunc
= NULL
;
3010 buf
->b_private
= NULL
;
3013 kmem_cache_free(buf_cache
, buf
);
3018 * Release this buffer from the cache. This must be done
3019 * after a read and prior to modifying the buffer contents.
3020 * If the buffer has more than one reference, we must make
3021 * a new hdr for the buffer.
3024 arc_release(arc_buf_t
*buf
, void *tag
)
3027 kmutex_t
*hash_lock
= NULL
;
3028 l2arc_buf_hdr_t
*l2hdr
;
3032 * It would be nice to assert that if it's DMU metadata (level >
3033 * 0 || it's the dnode file), then it must be syncing context.
3034 * But we don't know that information at this level.
3037 mutex_enter(&buf
->b_evict_lock
);
3040 /* this buffer is not on any list */
3041 ASSERT(refcount_count(&hdr
->b_refcnt
) > 0);
3043 if (hdr
->b_state
== arc_anon
) {
3044 /* this buffer is already released */
3045 ASSERT(buf
->b_efunc
== NULL
);
3047 hash_lock
= HDR_LOCK(hdr
);
3048 mutex_enter(hash_lock
);
3050 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3053 l2hdr
= hdr
->b_l2hdr
;
3055 mutex_enter(&l2arc_buflist_mtx
);
3056 hdr
->b_l2hdr
= NULL
;
3057 buf_size
= hdr
->b_size
;
3061 * Do we have more than one buf?
3063 if (hdr
->b_datacnt
> 1) {
3064 arc_buf_hdr_t
*nhdr
;
3066 uint64_t blksz
= hdr
->b_size
;
3067 uint64_t spa
= hdr
->b_spa
;
3068 arc_buf_contents_t type
= hdr
->b_type
;
3069 uint32_t flags
= hdr
->b_flags
;
3071 ASSERT(hdr
->b_buf
!= buf
|| buf
->b_next
!= NULL
);
3073 * Pull the data off of this hdr and attach it to
3074 * a new anonymous hdr.
3076 (void) remove_reference(hdr
, hash_lock
, tag
);
3078 while (*bufp
!= buf
)
3079 bufp
= &(*bufp
)->b_next
;
3080 *bufp
= buf
->b_next
;
3083 ASSERT3U(hdr
->b_state
->arcs_size
, >=, hdr
->b_size
);
3084 atomic_add_64(&hdr
->b_state
->arcs_size
, -hdr
->b_size
);
3085 if (refcount_is_zero(&hdr
->b_refcnt
)) {
3086 uint64_t *size
= &hdr
->b_state
->arcs_lsize
[hdr
->b_type
];
3087 ASSERT3U(*size
, >=, hdr
->b_size
);
3088 atomic_add_64(size
, -hdr
->b_size
);
3090 hdr
->b_datacnt
-= 1;
3091 arc_cksum_verify(buf
);
3093 mutex_exit(hash_lock
);
3095 nhdr
= kmem_cache_alloc(hdr_cache
, KM_PUSHPAGE
);
3096 nhdr
->b_size
= blksz
;
3098 nhdr
->b_type
= type
;
3100 nhdr
->b_state
= arc_anon
;
3101 nhdr
->b_arc_access
= 0;
3102 nhdr
->b_flags
= flags
& ARC_L2_WRITING
;
3103 nhdr
->b_l2hdr
= NULL
;
3104 nhdr
->b_datacnt
= 1;
3105 nhdr
->b_freeze_cksum
= NULL
;
3106 (void) refcount_add(&nhdr
->b_refcnt
, tag
);
3108 mutex_exit(&buf
->b_evict_lock
);
3109 atomic_add_64(&arc_anon
->arcs_size
, blksz
);
3111 mutex_exit(&buf
->b_evict_lock
);
3112 ASSERT(refcount_count(&hdr
->b_refcnt
) == 1);
3113 ASSERT(!list_link_active(&hdr
->b_arc_node
));
3114 ASSERT(!HDR_IO_IN_PROGRESS(hdr
));
3115 if (hdr
->b_state
!= arc_anon
)
3116 arc_change_state(arc_anon
, hdr
, hash_lock
);
3117 hdr
->b_arc_access
= 0;
3119 mutex_exit(hash_lock
);
3121 buf_discard_identity(hdr
);
3124 buf
->b_efunc
= NULL
;
3125 buf
->b_private
= NULL
;
3128 list_remove(l2hdr
->b_dev
->l2ad_buflist
, hdr
);
3129 kmem_free(l2hdr
, sizeof (l2arc_buf_hdr_t
));
3130 ARCSTAT_INCR(arcstat_l2_size
, -buf_size
);
3131 mutex_exit(&l2arc_buflist_mtx
);
3136 * Release this buffer. If it does not match the provided BP, fill it
3137 * with that block's contents.
3141 arc_release_bp(arc_buf_t
*buf
, void *tag
, blkptr_t
*bp
, spa_t
*spa
,
3144 arc_release(buf
, tag
);
3149 arc_released(arc_buf_t
*buf
)
3153 mutex_enter(&buf
->b_evict_lock
);
3154 released
= (buf
->b_data
!= NULL
&& buf
->b_hdr
->b_state
== arc_anon
);
3155 mutex_exit(&buf
->b_evict_lock
);
3160 arc_has_callback(arc_buf_t
*buf
)
3164 mutex_enter(&buf
->b_evict_lock
);
3165 callback
= (buf
->b_efunc
!= NULL
);
3166 mutex_exit(&buf
->b_evict_lock
);
3172 arc_referenced(arc_buf_t
*buf
)
3176 mutex_enter(&buf
->b_evict_lock
);
3177 referenced
= (refcount_count(&buf
->b_hdr
->b_refcnt
));
3178 mutex_exit(&buf
->b_evict_lock
);
3179 return (referenced
);
3184 arc_write_ready(zio_t
*zio
)
3186 arc_write_callback_t
*callback
= zio
->io_private
;
3187 arc_buf_t
*buf
= callback
->awcb_buf
;
3188 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3190 ASSERT(!refcount_is_zero(&buf
->b_hdr
->b_refcnt
));
3191 callback
->awcb_ready(zio
, buf
, callback
->awcb_private
);
3194 * If the IO is already in progress, then this is a re-write
3195 * attempt, so we need to thaw and re-compute the cksum.
3196 * It is the responsibility of the callback to handle the
3197 * accounting for any re-write attempt.
3199 if (HDR_IO_IN_PROGRESS(hdr
)) {
3200 mutex_enter(&hdr
->b_freeze_lock
);
3201 if (hdr
->b_freeze_cksum
!= NULL
) {
3202 kmem_free(hdr
->b_freeze_cksum
, sizeof (zio_cksum_t
));
3203 hdr
->b_freeze_cksum
= NULL
;
3205 mutex_exit(&hdr
->b_freeze_lock
);
3207 arc_cksum_compute(buf
, B_FALSE
);
3208 hdr
->b_flags
|= ARC_IO_IN_PROGRESS
;
3212 arc_write_done(zio_t
*zio
)
3214 arc_write_callback_t
*callback
= zio
->io_private
;
3215 arc_buf_t
*buf
= callback
->awcb_buf
;
3216 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3218 ASSERT(hdr
->b_acb
== NULL
);
3220 if (zio
->io_error
== 0) {
3221 hdr
->b_dva
= *BP_IDENTITY(zio
->io_bp
);
3222 hdr
->b_birth
= BP_PHYSICAL_BIRTH(zio
->io_bp
);
3223 hdr
->b_cksum0
= zio
->io_bp
->blk_cksum
.zc_word
[0];
3225 ASSERT(BUF_EMPTY(hdr
));
3229 * If the block to be written was all-zero, we may have
3230 * compressed it away. In this case no write was performed
3231 * so there will be no dva/birth/checksum. The buffer must
3232 * therefore remain anonymous (and uncached).
3234 if (!BUF_EMPTY(hdr
)) {
3235 arc_buf_hdr_t
*exists
;
3236 kmutex_t
*hash_lock
;
3238 ASSERT(zio
->io_error
== 0);
3240 arc_cksum_verify(buf
);
3242 exists
= buf_hash_insert(hdr
, &hash_lock
);
3245 * This can only happen if we overwrite for
3246 * sync-to-convergence, because we remove
3247 * buffers from the hash table when we arc_free().
3249 if (zio
->io_flags
& ZIO_FLAG_IO_REWRITE
) {
3250 if (!BP_EQUAL(&zio
->io_bp_orig
, zio
->io_bp
))
3251 panic("bad overwrite, hdr=%p exists=%p",
3252 (void *)hdr
, (void *)exists
);
3253 ASSERT(refcount_is_zero(&exists
->b_refcnt
));
3254 arc_change_state(arc_anon
, exists
, hash_lock
);
3255 mutex_exit(hash_lock
);
3256 arc_hdr_destroy(exists
);
3257 exists
= buf_hash_insert(hdr
, &hash_lock
);
3258 ASSERT3P(exists
, ==, NULL
);
3261 ASSERT(hdr
->b_datacnt
== 1);
3262 ASSERT(hdr
->b_state
== arc_anon
);
3263 ASSERT(BP_GET_DEDUP(zio
->io_bp
));
3264 ASSERT(BP_GET_LEVEL(zio
->io_bp
) == 0);
3267 hdr
->b_flags
&= ~ARC_IO_IN_PROGRESS
;
3268 /* if it's not anon, we are doing a scrub */
3269 if (!exists
&& hdr
->b_state
== arc_anon
)
3270 arc_access(hdr
, hash_lock
);
3271 mutex_exit(hash_lock
);
3273 hdr
->b_flags
&= ~ARC_IO_IN_PROGRESS
;
3276 ASSERT(!refcount_is_zero(&hdr
->b_refcnt
));
3277 callback
->awcb_done(zio
, buf
, callback
->awcb_private
);
3279 kmem_free(callback
, sizeof (arc_write_callback_t
));
3283 arc_write(zio_t
*pio
, spa_t
*spa
, uint64_t txg
,
3284 blkptr_t
*bp
, arc_buf_t
*buf
, boolean_t l2arc
, const zio_prop_t
*zp
,
3285 arc_done_func_t
*ready
, arc_done_func_t
*done
, void *private,
3286 int priority
, int zio_flags
, const zbookmark_t
*zb
)
3288 arc_buf_hdr_t
*hdr
= buf
->b_hdr
;
3289 arc_write_callback_t
*callback
;
3292 ASSERT(ready
!= NULL
);
3293 ASSERT(done
!= NULL
);
3294 ASSERT(!HDR_IO_ERROR(hdr
));
3295 ASSERT((hdr
->b_flags
& ARC_IO_IN_PROGRESS
) == 0);
3296 ASSERT(hdr
->b_acb
== NULL
);
3298 hdr
->b_flags
|= ARC_L2CACHE
;
3299 callback
= kmem_zalloc(sizeof (arc_write_callback_t
), KM_SLEEP
);
3300 callback
->awcb_ready
= ready
;
3301 callback
->awcb_done
= done
;
3302 callback
->awcb_private
= private;
3303 callback
->awcb_buf
= buf
;
3305 zio
= zio_write(pio
, spa
, txg
, bp
, buf
->b_data
, hdr
->b_size
, zp
,
3306 arc_write_ready
, arc_write_done
, callback
, priority
, zio_flags
, zb
);
3312 arc_memory_throttle(uint64_t reserve
, uint64_t inflight_data
, uint64_t txg
)
3315 uint64_t available_memory
= ptob(freemem
);
3316 static uint64_t page_load
= 0;
3317 static uint64_t last_txg
= 0;
3321 MIN(available_memory
, vmem_size(heap_arena
, VMEM_FREE
));
3323 if (available_memory
>= zfs_write_limit_max
)
3326 if (txg
> last_txg
) {
3331 * If we are in pageout, we know that memory is already tight,
3332 * the arc is already going to be evicting, so we just want to
3333 * continue to let page writes occur as quickly as possible.
3335 if (curproc
== proc_pageout
) {
3336 if (page_load
> MAX(ptob(minfree
), available_memory
) / 4)
3338 /* Note: reserve is inflated, so we deflate */
3339 page_load
+= reserve
/ 8;
3341 } else if (page_load
> 0 && arc_reclaim_needed()) {
3342 /* memory is low, delay before restarting */
3343 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
3348 if (arc_size
> arc_c_min
) {
3349 uint64_t evictable_memory
=
3350 arc_mru
->arcs_lsize
[ARC_BUFC_DATA
] +
3351 arc_mru
->arcs_lsize
[ARC_BUFC_METADATA
] +
3352 arc_mfu
->arcs_lsize
[ARC_BUFC_DATA
] +
3353 arc_mfu
->arcs_lsize
[ARC_BUFC_METADATA
];
3354 available_memory
+= MIN(evictable_memory
, arc_size
- arc_c_min
);
3357 if (inflight_data
> available_memory
/ 4) {
3358 ARCSTAT_INCR(arcstat_memory_throttle_count
, 1);
3366 arc_tempreserve_clear(uint64_t reserve
)
3368 atomic_add_64(&arc_tempreserve
, -reserve
);
3369 ASSERT((int64_t)arc_tempreserve
>= 0);
3373 arc_tempreserve_space(uint64_t reserve
, uint64_t txg
)
3380 * Once in a while, fail for no reason. Everything should cope.
3382 if (spa_get_random(10000) == 0) {
3383 dprintf("forcing random failure\n");
3387 if (reserve
> arc_c
/4 && !arc_no_grow
)
3388 arc_c
= MIN(arc_c_max
, reserve
* 4);
3389 if (reserve
> arc_c
)
3393 * Don't count loaned bufs as in flight dirty data to prevent long
3394 * network delays from blocking transactions that are ready to be
3395 * assigned to a txg.
3397 anon_size
= MAX((int64_t)(arc_anon
->arcs_size
- arc_loaned_bytes
), 0);
3400 * Writes will, almost always, require additional memory allocations
3401 * in order to compress/encrypt/etc the data. We therefor need to
3402 * make sure that there is sufficient available memory for this.
3404 if (error
= arc_memory_throttle(reserve
, anon_size
, txg
))
3408 * Throttle writes when the amount of dirty data in the cache
3409 * gets too large. We try to keep the cache less than half full
3410 * of dirty blocks so that our sync times don't grow too large.
3411 * Note: if two requests come in concurrently, we might let them
3412 * both succeed, when one of them should fail. Not a huge deal.
3415 if (reserve
+ arc_tempreserve
+ anon_size
> arc_c
/ 2 &&
3416 anon_size
> arc_c
/ 4) {
3417 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3418 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3419 arc_tempreserve
>>10,
3420 arc_anon
->arcs_lsize
[ARC_BUFC_METADATA
]>>10,
3421 arc_anon
->arcs_lsize
[ARC_BUFC_DATA
]>>10,
3422 reserve
>>10, arc_c
>>10);
3425 atomic_add_64(&arc_tempreserve
, reserve
);
3432 mutex_init(&arc_reclaim_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3433 cv_init(&arc_reclaim_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
3435 /* Convert seconds to clock ticks */
3436 arc_min_prefetch_lifespan
= 1 * hz
;
3438 /* Start out with 1/8 of all memory */
3439 arc_c
= physmem
* PAGESIZE
/ 8;
3443 * On architectures where the physical memory can be larger
3444 * than the addressable space (intel in 32-bit mode), we may
3445 * need to limit the cache to 1/8 of VM size.
3447 arc_c
= MIN(arc_c
, vmem_size(heap_arena
, VMEM_ALLOC
| VMEM_FREE
) / 8);
3450 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3451 arc_c_min
= MAX(arc_c
/ 4, 64<<20);
3452 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3453 if (arc_c
* 8 >= 1<<30)
3454 arc_c_max
= (arc_c
* 8) - (1<<30);
3456 arc_c_max
= arc_c_min
;
3457 arc_c_max
= MAX(arc_c
* 6, arc_c_max
);
3460 * Allow the tunables to override our calculations if they are
3461 * reasonable (ie. over 64MB)
3463 if (zfs_arc_max
> 64<<20 && zfs_arc_max
< physmem
* PAGESIZE
)
3464 arc_c_max
= zfs_arc_max
;
3465 if (zfs_arc_min
> 64<<20 && zfs_arc_min
<= arc_c_max
)
3466 arc_c_min
= zfs_arc_min
;
3469 arc_p
= (arc_c
>> 1);
3471 /* limit meta-data to 1/4 of the arc capacity */
3472 arc_meta_limit
= arc_c_max
/ 4;
3474 /* Allow the tunable to override if it is reasonable */
3475 if (zfs_arc_meta_limit
> 0 && zfs_arc_meta_limit
<= arc_c_max
)
3476 arc_meta_limit
= zfs_arc_meta_limit
;
3478 if (arc_c_min
< arc_meta_limit
/ 2 && zfs_arc_min
== 0)
3479 arc_c_min
= arc_meta_limit
/ 2;
3481 if (zfs_arc_grow_retry
> 0)
3482 arc_grow_retry
= zfs_arc_grow_retry
;
3484 if (zfs_arc_shrink_shift
> 0)
3485 arc_shrink_shift
= zfs_arc_shrink_shift
;
3487 if (zfs_arc_p_min_shift
> 0)
3488 arc_p_min_shift
= zfs_arc_p_min_shift
;
3490 /* if kmem_flags are set, lets try to use less memory */
3491 if (kmem_debugging())
3493 if (arc_c
< arc_c_min
)
3496 arc_anon
= &ARC_anon
;
3498 arc_mru_ghost
= &ARC_mru_ghost
;
3500 arc_mfu_ghost
= &ARC_mfu_ghost
;
3501 arc_l2c_only
= &ARC_l2c_only
;
3504 mutex_init(&arc_anon
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3505 mutex_init(&arc_mru
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3506 mutex_init(&arc_mru_ghost
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3507 mutex_init(&arc_mfu
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3508 mutex_init(&arc_mfu_ghost
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3509 mutex_init(&arc_l2c_only
->arcs_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3511 list_create(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
],
3512 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3513 list_create(&arc_mru
->arcs_list
[ARC_BUFC_DATA
],
3514 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3515 list_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
],
3516 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3517 list_create(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
],
3518 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3519 list_create(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
],
3520 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3521 list_create(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
],
3522 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3523 list_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
],
3524 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3525 list_create(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
],
3526 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3527 list_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_METADATA
],
3528 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3529 list_create(&arc_l2c_only
->arcs_list
[ARC_BUFC_DATA
],
3530 sizeof (arc_buf_hdr_t
), offsetof(arc_buf_hdr_t
, b_arc_node
));
3534 arc_thread_exit
= 0;
3535 arc_eviction_list
= NULL
;
3536 mutex_init(&arc_eviction_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
3537 bzero(&arc_eviction_hdr
, sizeof (arc_buf_hdr_t
));
3539 arc_ksp
= kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED
,
3540 sizeof (arc_stats
) / sizeof (kstat_named_t
), KSTAT_FLAG_VIRTUAL
);
3542 if (arc_ksp
!= NULL
) {
3543 arc_ksp
->ks_data
= &arc_stats
;
3544 kstat_install(arc_ksp
);
3547 (void) thread_create(NULL
, 0, arc_reclaim_thread
, NULL
, 0, &p0
,
3548 TS_RUN
, minclsyspri
);
3553 if (zfs_write_limit_max
== 0)
3554 zfs_write_limit_max
= ptob(physmem
) >> zfs_write_limit_shift
;
3556 zfs_write_limit_shift
= 0;
3557 mutex_init(&zfs_write_limit_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
3563 mutex_enter(&arc_reclaim_thr_lock
);
3564 arc_thread_exit
= 1;
3565 while (arc_thread_exit
!= 0)
3566 cv_wait(&arc_reclaim_thr_cv
, &arc_reclaim_thr_lock
);
3567 mutex_exit(&arc_reclaim_thr_lock
);
3573 if (arc_ksp
!= NULL
) {
3574 kstat_delete(arc_ksp
);
3578 mutex_destroy(&arc_eviction_mtx
);
3579 mutex_destroy(&arc_reclaim_thr_lock
);
3580 cv_destroy(&arc_reclaim_thr_cv
);
3582 list_destroy(&arc_mru
->arcs_list
[ARC_BUFC_METADATA
]);
3583 list_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
3584 list_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_METADATA
]);
3585 list_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_METADATA
]);
3586 list_destroy(&arc_mru
->arcs_list
[ARC_BUFC_DATA
]);
3587 list_destroy(&arc_mru_ghost
->arcs_list
[ARC_BUFC_DATA
]);
3588 list_destroy(&arc_mfu
->arcs_list
[ARC_BUFC_DATA
]);
3589 list_destroy(&arc_mfu_ghost
->arcs_list
[ARC_BUFC_DATA
]);
3591 mutex_destroy(&arc_anon
->arcs_mtx
);
3592 mutex_destroy(&arc_mru
->arcs_mtx
);
3593 mutex_destroy(&arc_mru_ghost
->arcs_mtx
);
3594 mutex_destroy(&arc_mfu
->arcs_mtx
);
3595 mutex_destroy(&arc_mfu_ghost
->arcs_mtx
);
3596 mutex_destroy(&arc_l2c_only
->arcs_mtx
);
3598 mutex_destroy(&zfs_write_limit_lock
);
3602 ASSERT(arc_loaned_bytes
== 0);
3608 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3609 * It uses dedicated storage devices to hold cached data, which are populated
3610 * using large infrequent writes. The main role of this cache is to boost
3611 * the performance of random read workloads. The intended L2ARC devices
3612 * include short-stroked disks, solid state disks, and other media with
3613 * substantially faster read latency than disk.
3615 * +-----------------------+
3617 * +-----------------------+
3620 * l2arc_feed_thread() arc_read()
3624 * +---------------+ |
3626 * +---------------+ |
3631 * +-------+ +-------+
3633 * | cache | | cache |
3634 * +-------+ +-------+
3635 * +=========+ .-----.
3636 * : L2ARC : |-_____-|
3637 * : devices : | Disks |
3638 * +=========+ `-_____-'
3640 * Read requests are satisfied from the following sources, in order:
3643 * 2) vdev cache of L2ARC devices
3645 * 4) vdev cache of disks
3648 * Some L2ARC device types exhibit extremely slow write performance.
3649 * To accommodate for this there are some significant differences between
3650 * the L2ARC and traditional cache design:
3652 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3653 * the ARC behave as usual, freeing buffers and placing headers on ghost
3654 * lists. The ARC does not send buffers to the L2ARC during eviction as
3655 * this would add inflated write latencies for all ARC memory pressure.
3657 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3658 * It does this by periodically scanning buffers from the eviction-end of
3659 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3660 * not already there. It scans until a headroom of buffers is satisfied,
3661 * which itself is a buffer for ARC eviction. The thread that does this is
3662 * l2arc_feed_thread(), illustrated below; example sizes are included to
3663 * provide a better sense of ratio than this diagram:
3666 * +---------------------+----------+
3667 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3668 * +---------------------+----------+ | o L2ARC eligible
3669 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
3670 * +---------------------+----------+ |
3671 * 15.9 Gbytes ^ 32 Mbytes |
3673 * l2arc_feed_thread()
3675 * l2arc write hand <--[oooo]--'
3679 * +==============================+
3680 * L2ARC dev |####|#|###|###| |####| ... |
3681 * +==============================+
3684 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
3685 * evicted, then the L2ARC has cached a buffer much sooner than it probably
3686 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
3687 * safe to say that this is an uncommon case, since buffers at the end of
3688 * the ARC lists have moved there due to inactivity.
3690 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
3691 * then the L2ARC simply misses copying some buffers. This serves as a
3692 * pressure valve to prevent heavy read workloads from both stalling the ARC
3693 * with waits and clogging the L2ARC with writes. This also helps prevent
3694 * the potential for the L2ARC to churn if it attempts to cache content too
3695 * quickly, such as during backups of the entire pool.
3697 * 5. After system boot and before the ARC has filled main memory, there are
3698 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
3699 * lists can remain mostly static. Instead of searching from tail of these
3700 * lists as pictured, the l2arc_feed_thread() will search from the list heads
3701 * for eligible buffers, greatly increasing its chance of finding them.
3703 * The L2ARC device write speed is also boosted during this time so that
3704 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
3705 * there are no L2ARC reads, and no fear of degrading read performance
3706 * through increased writes.
3708 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
3709 * the vdev queue can aggregate them into larger and fewer writes. Each
3710 * device is written to in a rotor fashion, sweeping writes through
3711 * available space then repeating.
3713 * 7. The L2ARC does not store dirty content. It never needs to flush
3714 * write buffers back to disk based storage.
3716 * 8. If an ARC buffer is written (and dirtied) which also exists in the
3717 * L2ARC, the now stale L2ARC buffer is immediately dropped.
3719 * The performance of the L2ARC can be tweaked by a number of tunables, which
3720 * may be necessary for different workloads:
3722 * l2arc_write_max max write bytes per interval
3723 * l2arc_write_boost extra write bytes during device warmup
3724 * l2arc_noprefetch skip caching prefetched buffers
3725 * l2arc_headroom number of max device writes to precache
3726 * l2arc_feed_secs seconds between L2ARC writing
3728 * Tunables may be removed or added as future performance improvements are
3729 * integrated, and also may become zpool properties.
3731 * There are three key functions that control how the L2ARC warms up:
3733 * l2arc_write_eligible() check if a buffer is eligible to cache
3734 * l2arc_write_size() calculate how much to write
3735 * l2arc_write_interval() calculate sleep delay between writes
3737 * These three functions determine what to write, how much, and how quickly
3742 l2arc_write_eligible(uint64_t spa_guid
, arc_buf_hdr_t
*ab
)
3745 * A buffer is *not* eligible for the L2ARC if it:
3746 * 1. belongs to a different spa.
3747 * 2. is already cached on the L2ARC.
3748 * 3. has an I/O in progress (it may be an incomplete read).
3749 * 4. is flagged not eligible (zfs property).
3751 if (ab
->b_spa
!= spa_guid
|| ab
->b_l2hdr
!= NULL
||
3752 HDR_IO_IN_PROGRESS(ab
) || !HDR_L2CACHE(ab
))
3759 l2arc_write_size(l2arc_dev_t
*dev
)
3763 size
= dev
->l2ad_write
;
3765 if (arc_warm
== B_FALSE
)
3766 size
+= dev
->l2ad_boost
;
3773 l2arc_write_interval(clock_t began
, uint64_t wanted
, uint64_t wrote
)
3775 clock_t interval
, next
, now
;
3778 * If the ARC lists are busy, increase our write rate; if the
3779 * lists are stale, idle back. This is achieved by checking
3780 * how much we previously wrote - if it was more than half of
3781 * what we wanted, schedule the next write much sooner.
3783 if (l2arc_feed_again
&& wrote
> (wanted
/ 2))
3784 interval
= (hz
* l2arc_feed_min_ms
) / 1000;
3786 interval
= hz
* l2arc_feed_secs
;
3788 now
= ddi_get_lbolt();
3789 next
= MAX(now
, MIN(now
+ interval
, began
+ interval
));
3795 l2arc_hdr_stat_add(void)
3797 ARCSTAT_INCR(arcstat_l2_hdr_size
, HDR_SIZE
+ L2HDR_SIZE
);
3798 ARCSTAT_INCR(arcstat_hdr_size
, -HDR_SIZE
);
3802 l2arc_hdr_stat_remove(void)
3804 ARCSTAT_INCR(arcstat_l2_hdr_size
, -(HDR_SIZE
+ L2HDR_SIZE
));
3805 ARCSTAT_INCR(arcstat_hdr_size
, HDR_SIZE
);
3809 * Cycle through L2ARC devices. This is how L2ARC load balances.
3810 * If a device is returned, this also returns holding the spa config lock.
3812 static l2arc_dev_t
*
3813 l2arc_dev_get_next(void)
3815 l2arc_dev_t
*first
, *next
= NULL
;
3818 * Lock out the removal of spas (spa_namespace_lock), then removal
3819 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
3820 * both locks will be dropped and a spa config lock held instead.
3822 mutex_enter(&spa_namespace_lock
);
3823 mutex_enter(&l2arc_dev_mtx
);
3825 /* if there are no vdevs, there is nothing to do */
3826 if (l2arc_ndev
== 0)
3830 next
= l2arc_dev_last
;
3832 /* loop around the list looking for a non-faulted vdev */
3834 next
= list_head(l2arc_dev_list
);
3836 next
= list_next(l2arc_dev_list
, next
);
3838 next
= list_head(l2arc_dev_list
);
3841 /* if we have come back to the start, bail out */
3844 else if (next
== first
)
3847 } while (vdev_is_dead(next
->l2ad_vdev
));
3849 /* if we were unable to find any usable vdevs, return NULL */
3850 if (vdev_is_dead(next
->l2ad_vdev
))
3853 l2arc_dev_last
= next
;
3856 mutex_exit(&l2arc_dev_mtx
);
3859 * Grab the config lock to prevent the 'next' device from being
3860 * removed while we are writing to it.
3863 spa_config_enter(next
->l2ad_spa
, SCL_L2ARC
, next
, RW_READER
);
3864 mutex_exit(&spa_namespace_lock
);
3870 * Free buffers that were tagged for destruction.
3873 l2arc_do_free_on_write()
3876 l2arc_data_free_t
*df
, *df_prev
;
3878 mutex_enter(&l2arc_free_on_write_mtx
);
3879 buflist
= l2arc_free_on_write
;
3881 for (df
= list_tail(buflist
); df
; df
= df_prev
) {
3882 df_prev
= list_prev(buflist
, df
);
3883 ASSERT(df
->l2df_data
!= NULL
);
3884 ASSERT(df
->l2df_func
!= NULL
);
3885 df
->l2df_func(df
->l2df_data
, df
->l2df_size
);
3886 list_remove(buflist
, df
);
3887 kmem_free(df
, sizeof (l2arc_data_free_t
));
3890 mutex_exit(&l2arc_free_on_write_mtx
);
3894 * A write to a cache device has completed. Update all headers to allow
3895 * reads from these buffers to begin.
3898 l2arc_write_done(zio_t
*zio
)
3900 l2arc_write_callback_t
*cb
;
3903 arc_buf_hdr_t
*head
, *ab
, *ab_prev
;
3904 l2arc_buf_hdr_t
*abl2
;
3905 kmutex_t
*hash_lock
;
3907 cb
= zio
->io_private
;
3909 dev
= cb
->l2wcb_dev
;
3910 ASSERT(dev
!= NULL
);
3911 head
= cb
->l2wcb_head
;
3912 ASSERT(head
!= NULL
);
3913 buflist
= dev
->l2ad_buflist
;
3914 ASSERT(buflist
!= NULL
);
3915 DTRACE_PROBE2(l2arc__iodone
, zio_t
*, zio
,
3916 l2arc_write_callback_t
*, cb
);
3918 if (zio
->io_error
!= 0)
3919 ARCSTAT_BUMP(arcstat_l2_writes_error
);
3921 mutex_enter(&l2arc_buflist_mtx
);
3924 * All writes completed, or an error was hit.
3926 for (ab
= list_prev(buflist
, head
); ab
; ab
= ab_prev
) {
3927 ab_prev
= list_prev(buflist
, ab
);
3929 hash_lock
= HDR_LOCK(ab
);
3930 if (!mutex_tryenter(hash_lock
)) {
3932 * This buffer misses out. It may be in a stage
3933 * of eviction. Its ARC_L2_WRITING flag will be
3934 * left set, denying reads to this buffer.
3936 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss
);
3940 if (zio
->io_error
!= 0) {
3942 * Error - drop L2ARC entry.
3944 list_remove(buflist
, ab
);
3947 kmem_free(abl2
, sizeof (l2arc_buf_hdr_t
));
3948 ARCSTAT_INCR(arcstat_l2_size
, -ab
->b_size
);
3952 * Allow ARC to begin reads to this L2ARC entry.
3954 ab
->b_flags
&= ~ARC_L2_WRITING
;
3956 mutex_exit(hash_lock
);
3959 atomic_inc_64(&l2arc_writes_done
);
3960 list_remove(buflist
, head
);
3961 kmem_cache_free(hdr_cache
, head
);
3962 mutex_exit(&l2arc_buflist_mtx
);
3964 l2arc_do_free_on_write();
3966 kmem_free(cb
, sizeof (l2arc_write_callback_t
));
3970 * A read to a cache device completed. Validate buffer contents before
3971 * handing over to the regular ARC routines.
3974 l2arc_read_done(zio_t
*zio
)
3976 l2arc_read_callback_t
*cb
;
3979 kmutex_t
*hash_lock
;
3982 ASSERT(zio
->io_vd
!= NULL
);
3983 ASSERT(zio
->io_flags
& ZIO_FLAG_DONT_PROPAGATE
);
3985 spa_config_exit(zio
->io_spa
, SCL_L2ARC
, zio
->io_vd
);
3987 cb
= zio
->io_private
;
3989 buf
= cb
->l2rcb_buf
;
3990 ASSERT(buf
!= NULL
);
3992 hash_lock
= HDR_LOCK(buf
->b_hdr
);
3993 mutex_enter(hash_lock
);
3995 ASSERT3P(hash_lock
, ==, HDR_LOCK(hdr
));
3998 * Check this survived the L2ARC journey.
4000 equal
= arc_cksum_equal(buf
);
4001 if (equal
&& zio
->io_error
== 0 && !HDR_L2_EVICTED(hdr
)) {
4002 mutex_exit(hash_lock
);
4003 zio
->io_private
= buf
;
4004 zio
->io_bp_copy
= cb
->l2rcb_bp
; /* XXX fix in L2ARC 2.0 */
4005 zio
->io_bp
= &zio
->io_bp_copy
; /* XXX fix in L2ARC 2.0 */
4008 mutex_exit(hash_lock
);
4010 * Buffer didn't survive caching. Increment stats and
4011 * reissue to the original storage device.
4013 if (zio
->io_error
!= 0) {
4014 ARCSTAT_BUMP(arcstat_l2_io_error
);
4016 zio
->io_error
= EIO
;
4019 ARCSTAT_BUMP(arcstat_l2_cksum_bad
);
4022 * If there's no waiter, issue an async i/o to the primary
4023 * storage now. If there *is* a waiter, the caller must
4024 * issue the i/o in a context where it's OK to block.
4026 if (zio
->io_waiter
== NULL
) {
4027 zio_t
*pio
= zio_unique_parent(zio
);
4029 ASSERT(!pio
|| pio
->io_child_type
== ZIO_CHILD_LOGICAL
);
4031 zio_nowait(zio_read(pio
, cb
->l2rcb_spa
, &cb
->l2rcb_bp
,
4032 buf
->b_data
, zio
->io_size
, arc_read_done
, buf
,
4033 zio
->io_priority
, cb
->l2rcb_flags
, &cb
->l2rcb_zb
));
4037 kmem_free(cb
, sizeof (l2arc_read_callback_t
));
4041 * This is the list priority from which the L2ARC will search for pages to
4042 * cache. This is used within loops (0..3) to cycle through lists in the
4043 * desired order. This order can have a significant effect on cache
4046 * Currently the metadata lists are hit first, MFU then MRU, followed by
4047 * the data lists. This function returns a locked list, and also returns
4051 l2arc_list_locked(int list_num
, kmutex_t
**lock
)
4055 ASSERT(list_num
>= 0 && list_num
<= 3);
4059 list
= &arc_mfu
->arcs_list
[ARC_BUFC_METADATA
];
4060 *lock
= &arc_mfu
->arcs_mtx
;
4063 list
= &arc_mru
->arcs_list
[ARC_BUFC_METADATA
];
4064 *lock
= &arc_mru
->arcs_mtx
;
4067 list
= &arc_mfu
->arcs_list
[ARC_BUFC_DATA
];
4068 *lock
= &arc_mfu
->arcs_mtx
;
4071 list
= &arc_mru
->arcs_list
[ARC_BUFC_DATA
];
4072 *lock
= &arc_mru
->arcs_mtx
;
4076 ASSERT(!(MUTEX_HELD(*lock
)));
4082 * Evict buffers from the device write hand to the distance specified in
4083 * bytes. This distance may span populated buffers, it may span nothing.
4084 * This is clearing a region on the L2ARC device ready for writing.
4085 * If the 'all' boolean is set, every buffer is evicted.
4088 l2arc_evict(l2arc_dev_t
*dev
, uint64_t distance
, boolean_t all
)
4091 l2arc_buf_hdr_t
*abl2
;
4092 arc_buf_hdr_t
*ab
, *ab_prev
;
4093 kmutex_t
*hash_lock
;
4096 buflist
= dev
->l2ad_buflist
;
4098 if (buflist
== NULL
)
4101 if (!all
&& dev
->l2ad_first
) {
4103 * This is the first sweep through the device. There is
4109 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- (2 * distance
))) {
4111 * When nearing the end of the device, evict to the end
4112 * before the device write hand jumps to the start.
4114 taddr
= dev
->l2ad_end
;
4116 taddr
= dev
->l2ad_hand
+ distance
;
4118 DTRACE_PROBE4(l2arc__evict
, l2arc_dev_t
*, dev
, list_t
*, buflist
,
4119 uint64_t, taddr
, boolean_t
, all
);
4122 mutex_enter(&l2arc_buflist_mtx
);
4123 for (ab
= list_tail(buflist
); ab
; ab
= ab_prev
) {
4124 ab_prev
= list_prev(buflist
, ab
);
4126 hash_lock
= HDR_LOCK(ab
);
4127 if (!mutex_tryenter(hash_lock
)) {
4129 * Missed the hash lock. Retry.
4131 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry
);
4132 mutex_exit(&l2arc_buflist_mtx
);
4133 mutex_enter(hash_lock
);
4134 mutex_exit(hash_lock
);
4138 if (HDR_L2_WRITE_HEAD(ab
)) {
4140 * We hit a write head node. Leave it for
4141 * l2arc_write_done().
4143 list_remove(buflist
, ab
);
4144 mutex_exit(hash_lock
);
4148 if (!all
&& ab
->b_l2hdr
!= NULL
&&
4149 (ab
->b_l2hdr
->b_daddr
> taddr
||
4150 ab
->b_l2hdr
->b_daddr
< dev
->l2ad_hand
)) {
4152 * We've evicted to the target address,
4153 * or the end of the device.
4155 mutex_exit(hash_lock
);
4159 if (HDR_FREE_IN_PROGRESS(ab
)) {
4161 * Already on the path to destruction.
4163 mutex_exit(hash_lock
);
4167 if (ab
->b_state
== arc_l2c_only
) {
4168 ASSERT(!HDR_L2_READING(ab
));
4170 * This doesn't exist in the ARC. Destroy.
4171 * arc_hdr_destroy() will call list_remove()
4172 * and decrement arcstat_l2_size.
4174 arc_change_state(arc_anon
, ab
, hash_lock
);
4175 arc_hdr_destroy(ab
);
4178 * Invalidate issued or about to be issued
4179 * reads, since we may be about to write
4180 * over this location.
4182 if (HDR_L2_READING(ab
)) {
4183 ARCSTAT_BUMP(arcstat_l2_evict_reading
);
4184 ab
->b_flags
|= ARC_L2_EVICTED
;
4188 * Tell ARC this no longer exists in L2ARC.
4190 if (ab
->b_l2hdr
!= NULL
) {
4193 kmem_free(abl2
, sizeof (l2arc_buf_hdr_t
));
4194 ARCSTAT_INCR(arcstat_l2_size
, -ab
->b_size
);
4196 list_remove(buflist
, ab
);
4199 * This may have been leftover after a
4202 ab
->b_flags
&= ~ARC_L2_WRITING
;
4204 mutex_exit(hash_lock
);
4206 mutex_exit(&l2arc_buflist_mtx
);
4208 vdev_space_update(dev
->l2ad_vdev
, -(taddr
- dev
->l2ad_evict
), 0, 0);
4209 dev
->l2ad_evict
= taddr
;
4213 * Find and write ARC buffers to the L2ARC device.
4215 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4216 * for reading until they have completed writing.
4219 l2arc_write_buffers(spa_t
*spa
, l2arc_dev_t
*dev
, uint64_t target_sz
)
4221 arc_buf_hdr_t
*ab
, *ab_prev
, *head
;
4222 l2arc_buf_hdr_t
*hdrl2
;
4224 uint64_t passed_sz
, write_sz
, buf_sz
, headroom
;
4226 kmutex_t
*hash_lock
, *list_lock
;
4227 boolean_t have_lock
, full
;
4228 l2arc_write_callback_t
*cb
;
4230 uint64_t guid
= spa_guid(spa
);
4232 ASSERT(dev
->l2ad_vdev
!= NULL
);
4237 head
= kmem_cache_alloc(hdr_cache
, KM_PUSHPAGE
);
4238 head
->b_flags
|= ARC_L2_WRITE_HEAD
;
4241 * Copy buffers for L2ARC writing.
4243 mutex_enter(&l2arc_buflist_mtx
);
4244 for (int try = 0; try <= 3; try++) {
4245 list
= l2arc_list_locked(try, &list_lock
);
4249 * L2ARC fast warmup.
4251 * Until the ARC is warm and starts to evict, read from the
4252 * head of the ARC lists rather than the tail.
4254 headroom
= target_sz
* l2arc_headroom
;
4255 if (arc_warm
== B_FALSE
)
4256 ab
= list_head(list
);
4258 ab
= list_tail(list
);
4260 for (; ab
; ab
= ab_prev
) {
4261 if (arc_warm
== B_FALSE
)
4262 ab_prev
= list_next(list
, ab
);
4264 ab_prev
= list_prev(list
, ab
);
4266 hash_lock
= HDR_LOCK(ab
);
4267 have_lock
= MUTEX_HELD(hash_lock
);
4268 if (!have_lock
&& !mutex_tryenter(hash_lock
)) {
4270 * Skip this buffer rather than waiting.
4275 passed_sz
+= ab
->b_size
;
4276 if (passed_sz
> headroom
) {
4280 mutex_exit(hash_lock
);
4284 if (!l2arc_write_eligible(guid
, ab
)) {
4285 mutex_exit(hash_lock
);
4289 if ((write_sz
+ ab
->b_size
) > target_sz
) {
4291 mutex_exit(hash_lock
);
4297 * Insert a dummy header on the buflist so
4298 * l2arc_write_done() can find where the
4299 * write buffers begin without searching.
4301 list_insert_head(dev
->l2ad_buflist
, head
);
4304 sizeof (l2arc_write_callback_t
), KM_SLEEP
);
4305 cb
->l2wcb_dev
= dev
;
4306 cb
->l2wcb_head
= head
;
4307 pio
= zio_root(spa
, l2arc_write_done
, cb
,
4312 * Create and add a new L2ARC header.
4314 hdrl2
= kmem_zalloc(sizeof (l2arc_buf_hdr_t
), KM_SLEEP
);
4316 hdrl2
->b_daddr
= dev
->l2ad_hand
;
4318 ab
->b_flags
|= ARC_L2_WRITING
;
4319 ab
->b_l2hdr
= hdrl2
;
4320 list_insert_head(dev
->l2ad_buflist
, ab
);
4321 buf_data
= ab
->b_buf
->b_data
;
4322 buf_sz
= ab
->b_size
;
4325 * Compute and store the buffer cksum before
4326 * writing. On debug the cksum is verified first.
4328 arc_cksum_verify(ab
->b_buf
);
4329 arc_cksum_compute(ab
->b_buf
, B_TRUE
);
4331 mutex_exit(hash_lock
);
4333 wzio
= zio_write_phys(pio
, dev
->l2ad_vdev
,
4334 dev
->l2ad_hand
, buf_sz
, buf_data
, ZIO_CHECKSUM_OFF
,
4335 NULL
, NULL
, ZIO_PRIORITY_ASYNC_WRITE
,
4336 ZIO_FLAG_CANFAIL
, B_FALSE
);
4338 DTRACE_PROBE2(l2arc__write
, vdev_t
*, dev
->l2ad_vdev
,
4340 (void) zio_nowait(wzio
);
4343 * Keep the clock hand suitably device-aligned.
4345 buf_sz
= vdev_psize_to_asize(dev
->l2ad_vdev
, buf_sz
);
4348 dev
->l2ad_hand
+= buf_sz
;
4351 mutex_exit(list_lock
);
4356 mutex_exit(&l2arc_buflist_mtx
);
4359 ASSERT3U(write_sz
, ==, 0);
4360 kmem_cache_free(hdr_cache
, head
);
4364 ASSERT3U(write_sz
, <=, target_sz
);
4365 ARCSTAT_BUMP(arcstat_l2_writes_sent
);
4366 ARCSTAT_INCR(arcstat_l2_write_bytes
, write_sz
);
4367 ARCSTAT_INCR(arcstat_l2_size
, write_sz
);
4368 vdev_space_update(dev
->l2ad_vdev
, write_sz
, 0, 0);
4371 * Bump device hand to the device start if it is approaching the end.
4372 * l2arc_evict() will already have evicted ahead for this case.
4374 if (dev
->l2ad_hand
>= (dev
->l2ad_end
- target_sz
)) {
4375 vdev_space_update(dev
->l2ad_vdev
,
4376 dev
->l2ad_end
- dev
->l2ad_hand
, 0, 0);
4377 dev
->l2ad_hand
= dev
->l2ad_start
;
4378 dev
->l2ad_evict
= dev
->l2ad_start
;
4379 dev
->l2ad_first
= B_FALSE
;
4382 dev
->l2ad_writing
= B_TRUE
;
4383 (void) zio_wait(pio
);
4384 dev
->l2ad_writing
= B_FALSE
;
4390 * This thread feeds the L2ARC at regular intervals. This is the beating
4391 * heart of the L2ARC.
4394 l2arc_feed_thread(void)
4399 uint64_t size
, wrote
;
4400 clock_t begin
, next
= ddi_get_lbolt();
4402 CALLB_CPR_INIT(&cpr
, &l2arc_feed_thr_lock
, callb_generic_cpr
, FTAG
);
4404 mutex_enter(&l2arc_feed_thr_lock
);
4406 while (l2arc_thread_exit
== 0) {
4407 CALLB_CPR_SAFE_BEGIN(&cpr
);
4408 (void) cv_timedwait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
,
4410 CALLB_CPR_SAFE_END(&cpr
, &l2arc_feed_thr_lock
);
4411 next
= ddi_get_lbolt() + hz
;
4414 * Quick check for L2ARC devices.
4416 mutex_enter(&l2arc_dev_mtx
);
4417 if (l2arc_ndev
== 0) {
4418 mutex_exit(&l2arc_dev_mtx
);
4421 mutex_exit(&l2arc_dev_mtx
);
4422 begin
= ddi_get_lbolt();
4425 * This selects the next l2arc device to write to, and in
4426 * doing so the next spa to feed from: dev->l2ad_spa. This
4427 * will return NULL if there are now no l2arc devices or if
4428 * they are all faulted.
4430 * If a device is returned, its spa's config lock is also
4431 * held to prevent device removal. l2arc_dev_get_next()
4432 * will grab and release l2arc_dev_mtx.
4434 if ((dev
= l2arc_dev_get_next()) == NULL
)
4437 spa
= dev
->l2ad_spa
;
4438 ASSERT(spa
!= NULL
);
4441 * Avoid contributing to memory pressure.
4443 if (arc_reclaim_needed()) {
4444 ARCSTAT_BUMP(arcstat_l2_abort_lowmem
);
4445 spa_config_exit(spa
, SCL_L2ARC
, dev
);
4449 ARCSTAT_BUMP(arcstat_l2_feeds
);
4451 size
= l2arc_write_size(dev
);
4454 * Evict L2ARC buffers that will be overwritten.
4456 l2arc_evict(dev
, size
, B_FALSE
);
4459 * Write ARC buffers.
4461 wrote
= l2arc_write_buffers(spa
, dev
, size
);
4464 * Calculate interval between writes.
4466 next
= l2arc_write_interval(begin
, size
, wrote
);
4467 spa_config_exit(spa
, SCL_L2ARC
, dev
);
4470 l2arc_thread_exit
= 0;
4471 cv_broadcast(&l2arc_feed_thr_cv
);
4472 CALLB_CPR_EXIT(&cpr
); /* drops l2arc_feed_thr_lock */
4477 l2arc_vdev_present(vdev_t
*vd
)
4481 mutex_enter(&l2arc_dev_mtx
);
4482 for (dev
= list_head(l2arc_dev_list
); dev
!= NULL
;
4483 dev
= list_next(l2arc_dev_list
, dev
)) {
4484 if (dev
->l2ad_vdev
== vd
)
4487 mutex_exit(&l2arc_dev_mtx
);
4489 return (dev
!= NULL
);
4493 * Add a vdev for use by the L2ARC. By this point the spa has already
4494 * validated the vdev and opened it.
4497 l2arc_add_vdev(spa_t
*spa
, vdev_t
*vd
)
4499 l2arc_dev_t
*adddev
;
4501 ASSERT(!l2arc_vdev_present(vd
));
4504 * Create a new l2arc device entry.
4506 adddev
= kmem_zalloc(sizeof (l2arc_dev_t
), KM_SLEEP
);
4507 adddev
->l2ad_spa
= spa
;
4508 adddev
->l2ad_vdev
= vd
;
4509 adddev
->l2ad_write
= l2arc_write_max
;
4510 adddev
->l2ad_boost
= l2arc_write_boost
;
4511 adddev
->l2ad_start
= VDEV_LABEL_START_SIZE
;
4512 adddev
->l2ad_end
= VDEV_LABEL_START_SIZE
+ vdev_get_min_asize(vd
);
4513 adddev
->l2ad_hand
= adddev
->l2ad_start
;
4514 adddev
->l2ad_evict
= adddev
->l2ad_start
;
4515 adddev
->l2ad_first
= B_TRUE
;
4516 adddev
->l2ad_writing
= B_FALSE
;
4517 ASSERT3U(adddev
->l2ad_write
, >, 0);
4520 * This is a list of all ARC buffers that are still valid on the
4523 adddev
->l2ad_buflist
= kmem_zalloc(sizeof (list_t
), KM_SLEEP
);
4524 list_create(adddev
->l2ad_buflist
, sizeof (arc_buf_hdr_t
),
4525 offsetof(arc_buf_hdr_t
, b_l2node
));
4527 vdev_space_update(vd
, 0, 0, adddev
->l2ad_end
- adddev
->l2ad_hand
);
4530 * Add device to global list
4532 mutex_enter(&l2arc_dev_mtx
);
4533 list_insert_head(l2arc_dev_list
, adddev
);
4534 atomic_inc_64(&l2arc_ndev
);
4535 mutex_exit(&l2arc_dev_mtx
);
4539 * Remove a vdev from the L2ARC.
4542 l2arc_remove_vdev(vdev_t
*vd
)
4544 l2arc_dev_t
*dev
, *nextdev
, *remdev
= NULL
;
4547 * Find the device by vdev
4549 mutex_enter(&l2arc_dev_mtx
);
4550 for (dev
= list_head(l2arc_dev_list
); dev
; dev
= nextdev
) {
4551 nextdev
= list_next(l2arc_dev_list
, dev
);
4552 if (vd
== dev
->l2ad_vdev
) {
4557 ASSERT(remdev
!= NULL
);
4560 * Remove device from global list
4562 list_remove(l2arc_dev_list
, remdev
);
4563 l2arc_dev_last
= NULL
; /* may have been invalidated */
4564 atomic_dec_64(&l2arc_ndev
);
4565 mutex_exit(&l2arc_dev_mtx
);
4568 * Clear all buflists and ARC references. L2ARC device flush.
4570 l2arc_evict(remdev
, 0, B_TRUE
);
4571 list_destroy(remdev
->l2ad_buflist
);
4572 kmem_free(remdev
->l2ad_buflist
, sizeof (list_t
));
4573 kmem_free(remdev
, sizeof (l2arc_dev_t
));
4579 l2arc_thread_exit
= 0;
4581 l2arc_writes_sent
= 0;
4582 l2arc_writes_done
= 0;
4584 mutex_init(&l2arc_feed_thr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
4585 cv_init(&l2arc_feed_thr_cv
, NULL
, CV_DEFAULT
, NULL
);
4586 mutex_init(&l2arc_dev_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4587 mutex_init(&l2arc_buflist_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4588 mutex_init(&l2arc_free_on_write_mtx
, NULL
, MUTEX_DEFAULT
, NULL
);
4590 l2arc_dev_list
= &L2ARC_dev_list
;
4591 l2arc_free_on_write
= &L2ARC_free_on_write
;
4592 list_create(l2arc_dev_list
, sizeof (l2arc_dev_t
),
4593 offsetof(l2arc_dev_t
, l2ad_node
));
4594 list_create(l2arc_free_on_write
, sizeof (l2arc_data_free_t
),
4595 offsetof(l2arc_data_free_t
, l2df_list_node
));
4602 * This is called from dmu_fini(), which is called from spa_fini();
4603 * Because of this, we can assume that all l2arc devices have
4604 * already been removed when the pools themselves were removed.
4607 l2arc_do_free_on_write();
4609 mutex_destroy(&l2arc_feed_thr_lock
);
4610 cv_destroy(&l2arc_feed_thr_cv
);
4611 mutex_destroy(&l2arc_dev_mtx
);
4612 mutex_destroy(&l2arc_buflist_mtx
);
4613 mutex_destroy(&l2arc_free_on_write_mtx
);
4615 list_destroy(l2arc_dev_list
);
4616 list_destroy(l2arc_free_on_write
);
4622 if (!(spa_mode_global
& FWRITE
))
4625 (void) thread_create(NULL
, 0, l2arc_feed_thread
, NULL
, 0, &p0
,
4626 TS_RUN
, minclsyspri
);
4632 if (!(spa_mode_global
& FWRITE
))
4635 mutex_enter(&l2arc_feed_thr_lock
);
4636 cv_signal(&l2arc_feed_thr_cv
); /* kick thread out of startup */
4637 l2arc_thread_exit
= 1;
4638 while (l2arc_thread_exit
!= 0)
4639 cv_wait(&l2arc_feed_thr_cv
, &l2arc_feed_thr_lock
);
4640 mutex_exit(&l2arc_feed_thr_lock
);