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 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
27 * DVA-based Adjustable Replacement Cache
29 * While much of the theory of operation used here is
30 * based on the self-tuning, low overhead replacement cache
31 * presented by Megiddo and Modha at FAST 2003, there are some
32 * significant differences:
34 * 1. The Megiddo and Modha model assumes any page is evictable.
35 * Pages in its cache cannot be "locked" into memory. This makes
36 * the eviction algorithm simple: evict the last page in the list.
37 * This also make the performance characteristics easy to reason
38 * about. Our cache is not so simple. At any given moment, some
39 * subset of the blocks in the cache are un-evictable because we
40 * have handed out a reference to them. Blocks are only evictable
41 * when there are no external references active. This makes
42 * eviction far more problematic: we choose to evict the evictable
43 * blocks that are the "lowest" in the list.
45 * There are times when it is not possible to evict the requested
46 * space. In these circumstances we are unable to adjust the cache
47 * size. To prevent the cache growing unbounded at these times we
48 * implement a "cache throttle" that slows the flow of new data
49 * into the cache until we can make space available.
51 * 2. The Megiddo and Modha model assumes a fixed cache size.
52 * Pages are evicted when the cache is full and there is a cache
53 * miss. Our model has a variable sized cache. It grows with
54 * high use, but also tries to react to memory pressure from the
55 * operating system: decreasing its size when system memory is
58 * 3. The Megiddo and Modha model assumes a fixed page size. All
59 * elements of the cache are therefor exactly the same size. So
60 * when adjusting the cache size following a cache miss, its simply
61 * a matter of choosing a single page to evict. In our model, we
62 * have variable sized cache blocks (rangeing from 512 bytes to
63 * 128K bytes). We therefor choose a set of blocks to evict to make
64 * space for a cache miss that approximates as closely as possible
65 * the space used by the new block.
67 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
68 * by N. Megiddo & D. Modha, FAST 2003
74 * A new reference to a cache buffer can be obtained in two
75 * ways: 1) via a hash table lookup using the DVA as a key,
76 * or 2) via one of the ARC lists. The arc_read() interface
77 * uses method 1, while the internal arc algorithms for
78 * adjusting the cache use method 2. We therefor provide two
79 * types of locks: 1) the hash table lock array, and 2) the
82 * Buffers do not have their own mutexs, rather they rely on the
83 * hash table mutexs for the bulk of their protection (i.e. most
84 * fields in the arc_buf_hdr_t are protected by these mutexs).
86 * buf_hash_find() returns the appropriate mutex (held) when it
87 * locates the requested buffer in the hash table. It returns
88 * NULL for the mutex if the buffer was not in the table.
90 * buf_hash_remove() expects the appropriate hash mutex to be
91 * already held before it is invoked.
93 * Each arc state also has a mutex which is used to protect the
94 * buffer list associated with the state. When attempting to
95 * obtain a hash table lock while holding an arc list lock you
96 * must use: mutex_tryenter() to avoid deadlock. Also note that
97 * the active state mutex must be held before the ghost state mutex.
99 * Arc buffers may have an associated eviction callback function.
100 * This function will be invoked prior to removing the buffer (e.g.
101 * in arc_do_user_evicts()). Note however that the data associated
102 * with the buffer may be evicted prior to the callback. The callback
103 * must be made with *no locks held* (to prevent deadlock). Additionally,
104 * the users of callbacks must ensure that their private data is
105 * protected from simultaneous callbacks from arc_buf_evict()
106 * and arc_do_user_evicts().
108 * Note that the majority of the performance stats are manipulated
109 * with atomic operations.
111 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
113 * - L2ARC buflist creation
114 * - L2ARC buflist eviction
115 * - L2ARC write completion, which walks L2ARC buflists
116 * - ARC header destruction, as it removes from L2ARC buflists
117 * - ARC header release, as it removes from L2ARC buflists
122 #include <sys/zio_checksum.h>
123 #include <sys/zfs_context.h>
125 #include <sys/refcount.h>
126 #include <sys/vdev.h>
128 #include <sys/vmsystm.h>
130 #include <sys/fs/swapnode.h>
131 #include <sys/dnlc.h>
133 #include <sys/callb.h>
134 #include <sys/kstat.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_mdcomp_disable = 0;
181 int zfs_arc_grow_retry = 0;
182 int zfs_arc_shrink_shift = 0;
183 int zfs_arc_p_min_shift = 0;
186 * Note that buffers can be in one of 6 states:
187 * ARC_anon - anonymous (discussed below)
188 * ARC_mru - recently used, currently cached
189 * ARC_mru_ghost - recentely used, no longer in cache
190 * ARC_mfu - frequently used, currently cached
191 * ARC_mfu_ghost - frequently used, no longer in cache
192 * ARC_l2c_only - exists in L2ARC but not other states
193 * When there are no active references to the buffer, they are
194 * are linked onto a list in one of these arc states. These are
195 * the only buffers that can be evicted or deleted. Within each
196 * state there are multiple lists, one for meta-data and one for
197 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
198 * etc.) is tracked separately so that it can be managed more
199 * explicitly: favored over data, limited explicitly.
201 * Anonymous buffers are buffers that are not associated with
202 * a DVA. These are buffers that hold dirty block copies
203 * before they are written to stable storage. By definition,
204 * they are "ref'd" and are considered part of arc_mru
205 * that cannot be freed. Generally, they will aquire a DVA
206 * as they are written and migrate onto the arc_mru list.
208 * The ARC_l2c_only state is for buffers that are in the second
209 * level ARC but no longer in any of the ARC_m* lists. The second
210 * level ARC itself may also contain buffers that are in any of
211 * the ARC_m* states - meaning that a buffer can exist in two
212 * places. The reason for the ARC_l2c_only state is to keep the
213 * buffer header in the hash table, so that reads that hit the
214 * second level ARC benefit from these fast lookups.
217 typedef struct arc_state {
218 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
219 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
220 uint64_t arcs_size; /* total amount of data in this state */
225 static arc_state_t ARC_anon;
226 static arc_state_t ARC_mru;
227 static arc_state_t ARC_mru_ghost;
228 static arc_state_t ARC_mfu;
229 static arc_state_t ARC_mfu_ghost;
230 static arc_state_t ARC_l2c_only;
232 typedef struct arc_stats {
233 kstat_named_t arcstat_hits;
234 kstat_named_t arcstat_misses;
235 kstat_named_t arcstat_demand_data_hits;
236 kstat_named_t arcstat_demand_data_misses;
237 kstat_named_t arcstat_demand_metadata_hits;
238 kstat_named_t arcstat_demand_metadata_misses;
239 kstat_named_t arcstat_prefetch_data_hits;
240 kstat_named_t arcstat_prefetch_data_misses;
241 kstat_named_t arcstat_prefetch_metadata_hits;
242 kstat_named_t arcstat_prefetch_metadata_misses;
243 kstat_named_t arcstat_mru_hits;
244 kstat_named_t arcstat_mru_ghost_hits;
245 kstat_named_t arcstat_mfu_hits;
246 kstat_named_t arcstat_mfu_ghost_hits;
247 kstat_named_t arcstat_deleted;
248 kstat_named_t arcstat_recycle_miss;
249 kstat_named_t arcstat_mutex_miss;
250 kstat_named_t arcstat_evict_skip;
251 kstat_named_t arcstat_hash_elements;
252 kstat_named_t arcstat_hash_elements_max;
253 kstat_named_t arcstat_hash_collisions;
254 kstat_named_t arcstat_hash_chains;
255 kstat_named_t arcstat_hash_chain_max;
256 kstat_named_t arcstat_p;
257 kstat_named_t arcstat_c;
258 kstat_named_t arcstat_c_min;
259 kstat_named_t arcstat_c_max;
260 kstat_named_t arcstat_size;
261 kstat_named_t arcstat_hdr_size;
262 kstat_named_t arcstat_data_size;
263 kstat_named_t arcstat_other_size;
264 kstat_named_t arcstat_l2_hits;
265 kstat_named_t arcstat_l2_misses;
266 kstat_named_t arcstat_l2_feeds;
267 kstat_named_t arcstat_l2_rw_clash;
268 kstat_named_t arcstat_l2_read_bytes;
269 kstat_named_t arcstat_l2_write_bytes;
270 kstat_named_t arcstat_l2_writes_sent;
271 kstat_named_t arcstat_l2_writes_done;
272 kstat_named_t arcstat_l2_writes_error;
273 kstat_named_t arcstat_l2_writes_hdr_miss;
274 kstat_named_t arcstat_l2_evict_lock_retry;
275 kstat_named_t arcstat_l2_evict_reading;
276 kstat_named_t arcstat_l2_free_on_write;
277 kstat_named_t arcstat_l2_abort_lowmem;
278 kstat_named_t arcstat_l2_cksum_bad;
279 kstat_named_t arcstat_l2_io_error;
280 kstat_named_t arcstat_l2_size;
281 kstat_named_t arcstat_l2_hdr_size;
282 kstat_named_t arcstat_memory_throttle_count;
285 static arc_stats_t arc_stats = {
286 { "hits", KSTAT_DATA_UINT64 },
287 { "misses", KSTAT_DATA_UINT64 },
288 { "demand_data_hits", KSTAT_DATA_UINT64 },
289 { "demand_data_misses", KSTAT_DATA_UINT64 },
290 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
291 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
292 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
293 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
294 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
295 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
296 { "mru_hits", KSTAT_DATA_UINT64 },
297 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
298 { "mfu_hits", KSTAT_DATA_UINT64 },
299 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
300 { "deleted", KSTAT_DATA_UINT64 },
301 { "recycle_miss", KSTAT_DATA_UINT64 },
302 { "mutex_miss", KSTAT_DATA_UINT64 },
303 { "evict_skip", KSTAT_DATA_UINT64 },
304 { "hash_elements", KSTAT_DATA_UINT64 },
305 { "hash_elements_max", KSTAT_DATA_UINT64 },
306 { "hash_collisions", KSTAT_DATA_UINT64 },
307 { "hash_chains", KSTAT_DATA_UINT64 },
308 { "hash_chain_max", KSTAT_DATA_UINT64 },
309 { "p", KSTAT_DATA_UINT64 },
310 { "c", KSTAT_DATA_UINT64 },
311 { "c_min", KSTAT_DATA_UINT64 },
312 { "c_max", KSTAT_DATA_UINT64 },
313 { "size", KSTAT_DATA_UINT64 },
314 { "hdr_size", KSTAT_DATA_UINT64 },
315 { "data_size", KSTAT_DATA_UINT64 },
316 { "other_size", KSTAT_DATA_UINT64 },
317 { "l2_hits", KSTAT_DATA_UINT64 },
318 { "l2_misses", KSTAT_DATA_UINT64 },
319 { "l2_feeds", KSTAT_DATA_UINT64 },
320 { "l2_rw_clash", KSTAT_DATA_UINT64 },
321 { "l2_read_bytes", KSTAT_DATA_UINT64 },
322 { "l2_write_bytes", KSTAT_DATA_UINT64 },
323 { "l2_writes_sent", KSTAT_DATA_UINT64 },
324 { "l2_writes_done", KSTAT_DATA_UINT64 },
325 { "l2_writes_error", KSTAT_DATA_UINT64 },
326 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
327 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
328 { "l2_evict_reading", KSTAT_DATA_UINT64 },
329 { "l2_free_on_write", KSTAT_DATA_UINT64 },
330 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
331 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
332 { "l2_io_error", KSTAT_DATA_UINT64 },
333 { "l2_size", KSTAT_DATA_UINT64 },
334 { "l2_hdr_size", KSTAT_DATA_UINT64 },
335 { "memory_throttle_count", KSTAT_DATA_UINT64 }
338 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
340 #define ARCSTAT_INCR(stat, val) \
341 atomic_add_64(&arc_stats.stat.value.ui64, (val));
343 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
344 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
346 #define ARCSTAT_MAX(stat, val) { \
348 while ((val) > (m = arc_stats.stat.value.ui64) && \
349 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
353 #define ARCSTAT_MAXSTAT(stat) \
354 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
357 * We define a macro to allow ARC hits/misses to be easily broken down by
358 * two separate conditions, giving a total of four different subtypes for
359 * each of hits and misses (so eight statistics total).
361 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
364 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
366 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
370 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
372 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
377 static arc_state_t *arc_anon;
378 static arc_state_t *arc_mru;
379 static arc_state_t *arc_mru_ghost;
380 static arc_state_t *arc_mfu;
381 static arc_state_t *arc_mfu_ghost;
382 static arc_state_t *arc_l2c_only;
385 * There are several ARC variables that are critical to export as kstats --
386 * but we don't want to have to grovel around in the kstat whenever we wish to
387 * manipulate them. For these variables, we therefore define them to be in
388 * terms of the statistic variable. This assures that we are not introducing
389 * the possibility of inconsistency by having shadow copies of the variables,
390 * while still allowing the code to be readable.
392 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
393 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
394 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
395 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
396 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
398 static int arc_no_grow; /* Don't try to grow cache size */
399 static uint64_t arc_tempreserve;
400 static uint64_t arc_meta_used;
401 static uint64_t arc_meta_limit;
402 static uint64_t arc_meta_max = 0;
404 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
406 typedef struct arc_callback arc_callback_t;
408 struct arc_callback {
410 arc_done_func_t *acb_done;
412 zio_t *acb_zio_dummy;
413 arc_callback_t *acb_next;
416 typedef struct arc_write_callback arc_write_callback_t;
418 struct arc_write_callback {
420 arc_done_func_t *awcb_ready;
421 arc_done_func_t *awcb_done;
426 /* protected by hash lock */
431 kmutex_t b_freeze_lock;
432 zio_cksum_t *b_freeze_cksum;
434 arc_buf_hdr_t *b_hash_next;
439 arc_callback_t *b_acb;
443 arc_buf_contents_t b_type;
447 /* protected by arc state mutex */
448 arc_state_t *b_state;
449 list_node_t b_arc_node;
451 /* updated atomically */
452 clock_t b_arc_access;
454 /* self protecting */
457 l2arc_buf_hdr_t *b_l2hdr;
458 list_node_t b_l2node;
461 static arc_buf_t *arc_eviction_list;
462 static kmutex_t arc_eviction_mtx;
463 static arc_buf_hdr_t arc_eviction_hdr;
464 static void arc_get_data_buf(arc_buf_t *buf);
465 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
466 static int arc_evict_needed(arc_buf_contents_t type);
467 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
469 #define GHOST_STATE(state) \
470 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
471 (state) == arc_l2c_only)
474 * Private ARC flags. These flags are private ARC only flags that will show up
475 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
476 * be passed in as arc_flags in things like arc_read. However, these flags
477 * should never be passed and should only be set by ARC code. When adding new
478 * public flags, make sure not to smash the private ones.
481 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
482 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
483 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
484 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
485 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
486 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
487 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
488 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
489 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
490 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
491 #define ARC_STORED (1 << 19) /* has been store()d to */
493 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
494 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
495 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
496 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
497 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
498 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
499 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
500 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
501 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
502 (hdr)->b_l2hdr != NULL)
503 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
504 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
505 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
511 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
512 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
515 * Hash table routines
518 #define HT_LOCK_PAD 64
523 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
527 #define BUF_LOCKS 256
528 typedef struct buf_hash_table {
530 arc_buf_hdr_t **ht_table;
531 struct ht_lock ht_locks[BUF_LOCKS];
534 static buf_hash_table_t buf_hash_table;
536 #define BUF_HASH_INDEX(spa, dva, birth) \
537 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
538 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
539 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
540 #define HDR_LOCK(buf) \
541 (BUF_HASH_LOCK(BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth)))
543 uint64_t zfs_crc64_table[256];
549 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
550 #define L2ARC_HEADROOM 2 /* num of writes */
551 #define L2ARC_FEED_SECS 1 /* caching interval secs */
552 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
554 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
555 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
558 * L2ARC Performance Tunables
560 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
561 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
562 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
563 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
564 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
565 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
566 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
567 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
572 typedef struct l2arc_dev {
573 vdev_t *l2ad_vdev; /* vdev */
574 spa_t *l2ad_spa; /* spa */
575 uint64_t l2ad_hand; /* next write location */
576 uint64_t l2ad_write; /* desired write size, bytes */
577 uint64_t l2ad_boost; /* warmup write boost, bytes */
578 uint64_t l2ad_start; /* first addr on device */
579 uint64_t l2ad_end; /* last addr on device */
580 uint64_t l2ad_evict; /* last addr eviction reached */
581 boolean_t l2ad_first; /* first sweep through */
582 boolean_t l2ad_writing; /* currently writing */
583 list_t *l2ad_buflist; /* buffer list */
584 list_node_t l2ad_node; /* device list node */
587 static list_t L2ARC_dev_list; /* device list */
588 static list_t *l2arc_dev_list; /* device list pointer */
589 static kmutex_t l2arc_dev_mtx; /* device list mutex */
590 static l2arc_dev_t *l2arc_dev_last; /* last device used */
591 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
592 static list_t L2ARC_free_on_write; /* free after write buf list */
593 static list_t *l2arc_free_on_write; /* free after write list ptr */
594 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
595 static uint64_t l2arc_ndev; /* number of devices */
597 typedef struct l2arc_read_callback {
598 arc_buf_t *l2rcb_buf; /* read buffer */
599 spa_t *l2rcb_spa; /* spa */
600 blkptr_t l2rcb_bp; /* original blkptr */
601 zbookmark_t l2rcb_zb; /* original bookmark */
602 int l2rcb_flags; /* original flags */
603 } l2arc_read_callback_t;
605 typedef struct l2arc_write_callback {
606 l2arc_dev_t *l2wcb_dev; /* device info */
607 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
608 } l2arc_write_callback_t;
610 struct l2arc_buf_hdr {
611 /* protected by arc_buf_hdr mutex */
612 l2arc_dev_t *b_dev; /* L2ARC device */
613 daddr_t b_daddr; /* disk address, offset byte */
616 typedef struct l2arc_data_free {
617 /* protected by l2arc_free_on_write_mtx */
620 void (*l2df_func)(void *, size_t);
621 list_node_t l2df_list_node;
624 static kmutex_t l2arc_feed_thr_lock;
625 static kcondvar_t l2arc_feed_thr_cv;
626 static uint8_t l2arc_thread_exit;
628 static void l2arc_read_done(zio_t *zio);
629 static void l2arc_hdr_stat_add(void);
630 static void l2arc_hdr_stat_remove(void);
633 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
635 uint8_t *vdva = (uint8_t *)dva;
636 uint64_t crc = -1ULL;
639 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
641 for (i = 0; i < sizeof (dva_t); i++)
642 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
644 crc ^= (spa>>8) ^ birth;
649 #define BUF_EMPTY(buf) \
650 ((buf)->b_dva.dva_word[0] == 0 && \
651 (buf)->b_dva.dva_word[1] == 0 && \
654 #define BUF_EQUAL(spa, dva, birth, buf) \
655 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
656 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
657 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
659 static arc_buf_hdr_t *
660 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
662 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
663 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
666 mutex_enter(hash_lock);
667 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
668 buf = buf->b_hash_next) {
669 if (BUF_EQUAL(spa, dva, birth, buf)) {
674 mutex_exit(hash_lock);
680 * Insert an entry into the hash table. If there is already an element
681 * equal to elem in the hash table, then the already existing element
682 * will be returned and the new element will not be inserted.
683 * Otherwise returns NULL.
685 static arc_buf_hdr_t *
686 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
688 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
689 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
693 ASSERT(!HDR_IN_HASH_TABLE(buf));
695 mutex_enter(hash_lock);
696 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
697 fbuf = fbuf->b_hash_next, i++) {
698 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
702 buf->b_hash_next = buf_hash_table.ht_table[idx];
703 buf_hash_table.ht_table[idx] = buf;
704 buf->b_flags |= ARC_IN_HASH_TABLE;
706 /* collect some hash table performance data */
708 ARCSTAT_BUMP(arcstat_hash_collisions);
710 ARCSTAT_BUMP(arcstat_hash_chains);
712 ARCSTAT_MAX(arcstat_hash_chain_max, i);
715 ARCSTAT_BUMP(arcstat_hash_elements);
716 ARCSTAT_MAXSTAT(arcstat_hash_elements);
722 buf_hash_remove(arc_buf_hdr_t *buf)
724 arc_buf_hdr_t *fbuf, **bufp;
725 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
727 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
728 ASSERT(HDR_IN_HASH_TABLE(buf));
730 bufp = &buf_hash_table.ht_table[idx];
731 while ((fbuf = *bufp) != buf) {
732 ASSERT(fbuf != NULL);
733 bufp = &fbuf->b_hash_next;
735 *bufp = buf->b_hash_next;
736 buf->b_hash_next = NULL;
737 buf->b_flags &= ~ARC_IN_HASH_TABLE;
739 /* collect some hash table performance data */
740 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
742 if (buf_hash_table.ht_table[idx] &&
743 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
744 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
748 * Global data structures and functions for the buf kmem cache.
750 static kmem_cache_t *hdr_cache;
751 static kmem_cache_t *buf_cache;
758 kmem_free(buf_hash_table.ht_table,
759 (buf_hash_table.ht_mask + 1) * sizeof (void *));
760 for (i = 0; i < BUF_LOCKS; i++)
761 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
762 kmem_cache_destroy(hdr_cache);
763 kmem_cache_destroy(buf_cache);
767 * Constructor callback - called when the cache is empty
768 * and a new buf is requested.
772 hdr_cons(void *vbuf, void *unused, int kmflag)
774 arc_buf_hdr_t *buf = vbuf;
776 bzero(buf, sizeof (arc_buf_hdr_t));
777 refcount_create(&buf->b_refcnt);
778 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
779 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
780 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
787 buf_cons(void *vbuf, void *unused, int kmflag)
789 arc_buf_t *buf = vbuf;
791 bzero(buf, sizeof (arc_buf_t));
792 rw_init(&buf->b_lock, NULL, RW_DEFAULT, NULL);
793 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
799 * Destructor callback - called when a cached buf is
800 * no longer required.
804 hdr_dest(void *vbuf, void *unused)
806 arc_buf_hdr_t *buf = vbuf;
808 refcount_destroy(&buf->b_refcnt);
809 cv_destroy(&buf->b_cv);
810 mutex_destroy(&buf->b_freeze_lock);
811 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
816 buf_dest(void *vbuf, void *unused)
818 arc_buf_t *buf = vbuf;
820 rw_destroy(&buf->b_lock);
821 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
825 * Reclaim callback -- invoked when memory is low.
829 hdr_recl(void *unused)
831 dprintf("hdr_recl called\n");
833 * umem calls the reclaim func when we destroy the buf cache,
834 * which is after we do arc_fini().
837 cv_signal(&arc_reclaim_thr_cv);
844 uint64_t hsize = 1ULL << 12;
848 * The hash table is big enough to fill all of physical memory
849 * with an average 64K block size. The table will take up
850 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
852 while (hsize * 65536 < physmem * PAGESIZE)
855 buf_hash_table.ht_mask = hsize - 1;
856 buf_hash_table.ht_table =
857 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
858 if (buf_hash_table.ht_table == NULL) {
859 ASSERT(hsize > (1ULL << 8));
864 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
865 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
866 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
867 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
869 for (i = 0; i < 256; i++)
870 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
871 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
873 for (i = 0; i < BUF_LOCKS; i++) {
874 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
875 NULL, MUTEX_DEFAULT, NULL);
879 #define ARC_MINTIME (hz>>4) /* 62 ms */
882 arc_cksum_verify(arc_buf_t *buf)
886 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
889 mutex_enter(&buf->b_hdr->b_freeze_lock);
890 if (buf->b_hdr->b_freeze_cksum == NULL ||
891 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
892 mutex_exit(&buf->b_hdr->b_freeze_lock);
895 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
896 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
897 panic("buffer modified while frozen!");
898 mutex_exit(&buf->b_hdr->b_freeze_lock);
902 arc_cksum_equal(arc_buf_t *buf)
907 mutex_enter(&buf->b_hdr->b_freeze_lock);
908 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
909 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
910 mutex_exit(&buf->b_hdr->b_freeze_lock);
916 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
918 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
921 mutex_enter(&buf->b_hdr->b_freeze_lock);
922 if (buf->b_hdr->b_freeze_cksum != NULL) {
923 mutex_exit(&buf->b_hdr->b_freeze_lock);
926 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
927 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
928 buf->b_hdr->b_freeze_cksum);
929 mutex_exit(&buf->b_hdr->b_freeze_lock);
933 arc_buf_thaw(arc_buf_t *buf)
935 if (zfs_flags & ZFS_DEBUG_MODIFY) {
936 if (buf->b_hdr->b_state != arc_anon)
937 panic("modifying non-anon buffer!");
938 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
939 panic("modifying buffer while i/o in progress!");
940 arc_cksum_verify(buf);
943 mutex_enter(&buf->b_hdr->b_freeze_lock);
944 if (buf->b_hdr->b_freeze_cksum != NULL) {
945 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
946 buf->b_hdr->b_freeze_cksum = NULL;
948 mutex_exit(&buf->b_hdr->b_freeze_lock);
952 arc_buf_freeze(arc_buf_t *buf)
954 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
957 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
958 buf->b_hdr->b_state == arc_anon);
959 arc_cksum_compute(buf, B_FALSE);
963 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
965 ASSERT(MUTEX_HELD(hash_lock));
967 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
968 (ab->b_state != arc_anon)) {
969 uint64_t delta = ab->b_size * ab->b_datacnt;
970 list_t *list = &ab->b_state->arcs_list[ab->b_type];
971 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
973 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
974 mutex_enter(&ab->b_state->arcs_mtx);
975 ASSERT(list_link_active(&ab->b_arc_node));
976 list_remove(list, ab);
977 if (GHOST_STATE(ab->b_state)) {
978 ASSERT3U(ab->b_datacnt, ==, 0);
979 ASSERT3P(ab->b_buf, ==, NULL);
983 ASSERT3U(*size, >=, delta);
984 atomic_add_64(size, -delta);
985 mutex_exit(&ab->b_state->arcs_mtx);
986 /* remove the prefetch flag if we get a reference */
987 if (ab->b_flags & ARC_PREFETCH)
988 ab->b_flags &= ~ARC_PREFETCH;
993 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
996 arc_state_t *state = ab->b_state;
998 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
999 ASSERT(!GHOST_STATE(state));
1001 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1002 (state != arc_anon)) {
1003 uint64_t *size = &state->arcs_lsize[ab->b_type];
1005 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1006 mutex_enter(&state->arcs_mtx);
1007 ASSERT(!list_link_active(&ab->b_arc_node));
1008 list_insert_head(&state->arcs_list[ab->b_type], ab);
1009 ASSERT(ab->b_datacnt > 0);
1010 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1011 mutex_exit(&state->arcs_mtx);
1017 * Move the supplied buffer to the indicated state. The mutex
1018 * for the buffer must be held by the caller.
1021 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1023 arc_state_t *old_state = ab->b_state;
1024 int64_t refcnt = refcount_count(&ab->b_refcnt);
1025 uint64_t from_delta, to_delta;
1027 ASSERT(MUTEX_HELD(hash_lock));
1028 ASSERT(new_state != old_state);
1029 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1030 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1032 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1035 * If this buffer is evictable, transfer it from the
1036 * old state list to the new state list.
1039 if (old_state != arc_anon) {
1040 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1041 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1044 mutex_enter(&old_state->arcs_mtx);
1046 ASSERT(list_link_active(&ab->b_arc_node));
1047 list_remove(&old_state->arcs_list[ab->b_type], ab);
1050 * If prefetching out of the ghost cache,
1051 * we will have a non-null datacnt.
1053 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1054 /* ghost elements have a ghost size */
1055 ASSERT(ab->b_buf == NULL);
1056 from_delta = ab->b_size;
1058 ASSERT3U(*size, >=, from_delta);
1059 atomic_add_64(size, -from_delta);
1062 mutex_exit(&old_state->arcs_mtx);
1064 if (new_state != arc_anon) {
1065 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1066 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1069 mutex_enter(&new_state->arcs_mtx);
1071 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1073 /* ghost elements have a ghost size */
1074 if (GHOST_STATE(new_state)) {
1075 ASSERT(ab->b_datacnt == 0);
1076 ASSERT(ab->b_buf == NULL);
1077 to_delta = ab->b_size;
1079 atomic_add_64(size, to_delta);
1082 mutex_exit(&new_state->arcs_mtx);
1086 ASSERT(!BUF_EMPTY(ab));
1087 if (new_state == arc_anon) {
1088 buf_hash_remove(ab);
1091 /* adjust state sizes */
1093 atomic_add_64(&new_state->arcs_size, to_delta);
1095 ASSERT3U(old_state->arcs_size, >=, from_delta);
1096 atomic_add_64(&old_state->arcs_size, -from_delta);
1098 ab->b_state = new_state;
1100 /* adjust l2arc hdr stats */
1101 if (new_state == arc_l2c_only)
1102 l2arc_hdr_stat_add();
1103 else if (old_state == arc_l2c_only)
1104 l2arc_hdr_stat_remove();
1108 arc_space_consume(uint64_t space, arc_space_type_t type)
1110 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1113 case ARC_SPACE_DATA:
1114 ARCSTAT_INCR(arcstat_data_size, space);
1116 case ARC_SPACE_OTHER:
1117 ARCSTAT_INCR(arcstat_other_size, space);
1119 case ARC_SPACE_HDRS:
1120 ARCSTAT_INCR(arcstat_hdr_size, space);
1122 case ARC_SPACE_L2HDRS:
1123 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1127 atomic_add_64(&arc_meta_used, space);
1128 atomic_add_64(&arc_size, space);
1132 arc_space_return(uint64_t space, arc_space_type_t type)
1134 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1137 case ARC_SPACE_DATA:
1138 ARCSTAT_INCR(arcstat_data_size, -space);
1140 case ARC_SPACE_OTHER:
1141 ARCSTAT_INCR(arcstat_other_size, -space);
1143 case ARC_SPACE_HDRS:
1144 ARCSTAT_INCR(arcstat_hdr_size, -space);
1146 case ARC_SPACE_L2HDRS:
1147 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1151 ASSERT(arc_meta_used >= space);
1152 if (arc_meta_max < arc_meta_used)
1153 arc_meta_max = arc_meta_used;
1154 atomic_add_64(&arc_meta_used, -space);
1155 ASSERT(arc_size >= space);
1156 atomic_add_64(&arc_size, -space);
1160 arc_data_buf_alloc(uint64_t size)
1162 if (arc_evict_needed(ARC_BUFC_DATA))
1163 cv_signal(&arc_reclaim_thr_cv);
1164 atomic_add_64(&arc_size, size);
1165 return (zio_data_buf_alloc(size));
1169 arc_data_buf_free(void *buf, uint64_t size)
1171 zio_data_buf_free(buf, size);
1172 ASSERT(arc_size >= size);
1173 atomic_add_64(&arc_size, -size);
1177 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1182 ASSERT3U(size, >, 0);
1183 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1184 ASSERT(BUF_EMPTY(hdr));
1187 hdr->b_spa = spa_guid(spa);
1188 hdr->b_state = arc_anon;
1189 hdr->b_arc_access = 0;
1190 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1193 buf->b_efunc = NULL;
1194 buf->b_private = NULL;
1197 arc_get_data_buf(buf);
1200 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1201 (void) refcount_add(&hdr->b_refcnt, tag);
1207 arc_buf_clone(arc_buf_t *from)
1210 arc_buf_hdr_t *hdr = from->b_hdr;
1211 uint64_t size = hdr->b_size;
1213 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1216 buf->b_efunc = NULL;
1217 buf->b_private = NULL;
1218 buf->b_next = hdr->b_buf;
1220 arc_get_data_buf(buf);
1221 bcopy(from->b_data, buf->b_data, size);
1222 hdr->b_datacnt += 1;
1227 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1230 kmutex_t *hash_lock;
1233 * Check to see if this buffer is evicted. Callers
1234 * must verify b_data != NULL to know if the add_ref
1237 rw_enter(&buf->b_lock, RW_READER);
1238 if (buf->b_data == NULL) {
1239 rw_exit(&buf->b_lock);
1243 ASSERT(hdr != NULL);
1244 hash_lock = HDR_LOCK(hdr);
1245 mutex_enter(hash_lock);
1246 rw_exit(&buf->b_lock);
1248 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1249 add_reference(hdr, hash_lock, tag);
1250 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1251 arc_access(hdr, hash_lock);
1252 mutex_exit(hash_lock);
1253 ARCSTAT_BUMP(arcstat_hits);
1254 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1255 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1256 data, metadata, hits);
1260 * Free the arc data buffer. If it is an l2arc write in progress,
1261 * the buffer is placed on l2arc_free_on_write to be freed later.
1264 arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t),
1265 void *data, size_t size)
1267 if (HDR_L2_WRITING(hdr)) {
1268 l2arc_data_free_t *df;
1269 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1270 df->l2df_data = data;
1271 df->l2df_size = size;
1272 df->l2df_func = free_func;
1273 mutex_enter(&l2arc_free_on_write_mtx);
1274 list_insert_head(l2arc_free_on_write, df);
1275 mutex_exit(&l2arc_free_on_write_mtx);
1276 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1278 free_func(data, size);
1283 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1287 /* free up data associated with the buf */
1289 arc_state_t *state = buf->b_hdr->b_state;
1290 uint64_t size = buf->b_hdr->b_size;
1291 arc_buf_contents_t type = buf->b_hdr->b_type;
1293 arc_cksum_verify(buf);
1295 if (type == ARC_BUFC_METADATA) {
1296 arc_buf_data_free(buf->b_hdr, zio_buf_free,
1298 arc_space_return(size, ARC_SPACE_DATA);
1300 ASSERT(type == ARC_BUFC_DATA);
1301 arc_buf_data_free(buf->b_hdr,
1302 zio_data_buf_free, buf->b_data, size);
1303 ARCSTAT_INCR(arcstat_data_size, -size);
1304 atomic_add_64(&arc_size, -size);
1307 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1308 uint64_t *cnt = &state->arcs_lsize[type];
1310 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1311 ASSERT(state != arc_anon);
1313 ASSERT3U(*cnt, >=, size);
1314 atomic_add_64(cnt, -size);
1316 ASSERT3U(state->arcs_size, >=, size);
1317 atomic_add_64(&state->arcs_size, -size);
1319 ASSERT(buf->b_hdr->b_datacnt > 0);
1320 buf->b_hdr->b_datacnt -= 1;
1323 /* only remove the buf if requested */
1327 /* remove the buf from the hdr list */
1328 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1330 *bufp = buf->b_next;
1332 ASSERT(buf->b_efunc == NULL);
1334 /* clean up the buf */
1336 kmem_cache_free(buf_cache, buf);
1340 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1342 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1343 ASSERT3P(hdr->b_state, ==, arc_anon);
1344 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1345 ASSERT(!(hdr->b_flags & ARC_STORED));
1347 if (hdr->b_l2hdr != NULL) {
1348 if (!MUTEX_HELD(&l2arc_buflist_mtx)) {
1350 * To prevent arc_free() and l2arc_evict() from
1351 * attempting to free the same buffer at the same time,
1352 * a FREE_IN_PROGRESS flag is given to arc_free() to
1353 * give it priority. l2arc_evict() can't destroy this
1354 * header while we are waiting on l2arc_buflist_mtx.
1356 * The hdr may be removed from l2ad_buflist before we
1357 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1359 mutex_enter(&l2arc_buflist_mtx);
1360 if (hdr->b_l2hdr != NULL) {
1361 list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist,
1364 mutex_exit(&l2arc_buflist_mtx);
1366 list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist, hdr);
1368 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1369 kmem_free(hdr->b_l2hdr, sizeof (l2arc_buf_hdr_t));
1370 if (hdr->b_state == arc_l2c_only)
1371 l2arc_hdr_stat_remove();
1372 hdr->b_l2hdr = NULL;
1375 if (!BUF_EMPTY(hdr)) {
1376 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1377 bzero(&hdr->b_dva, sizeof (dva_t));
1381 while (hdr->b_buf) {
1382 arc_buf_t *buf = hdr->b_buf;
1385 mutex_enter(&arc_eviction_mtx);
1386 rw_enter(&buf->b_lock, RW_WRITER);
1387 ASSERT(buf->b_hdr != NULL);
1388 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1389 hdr->b_buf = buf->b_next;
1390 buf->b_hdr = &arc_eviction_hdr;
1391 buf->b_next = arc_eviction_list;
1392 arc_eviction_list = buf;
1393 rw_exit(&buf->b_lock);
1394 mutex_exit(&arc_eviction_mtx);
1396 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1399 if (hdr->b_freeze_cksum != NULL) {
1400 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1401 hdr->b_freeze_cksum = NULL;
1404 ASSERT(!list_link_active(&hdr->b_arc_node));
1405 ASSERT3P(hdr->b_hash_next, ==, NULL);
1406 ASSERT3P(hdr->b_acb, ==, NULL);
1407 kmem_cache_free(hdr_cache, hdr);
1411 arc_buf_free(arc_buf_t *buf, void *tag)
1413 arc_buf_hdr_t *hdr = buf->b_hdr;
1414 int hashed = hdr->b_state != arc_anon;
1416 ASSERT(buf->b_efunc == NULL);
1417 ASSERT(buf->b_data != NULL);
1420 kmutex_t *hash_lock = HDR_LOCK(hdr);
1422 mutex_enter(hash_lock);
1423 (void) remove_reference(hdr, hash_lock, tag);
1424 if (hdr->b_datacnt > 1)
1425 arc_buf_destroy(buf, FALSE, TRUE);
1427 hdr->b_flags |= ARC_BUF_AVAILABLE;
1428 mutex_exit(hash_lock);
1429 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1432 * We are in the middle of an async write. Don't destroy
1433 * this buffer unless the write completes before we finish
1434 * decrementing the reference count.
1436 mutex_enter(&arc_eviction_mtx);
1437 (void) remove_reference(hdr, NULL, tag);
1438 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1439 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1440 mutex_exit(&arc_eviction_mtx);
1442 arc_hdr_destroy(hdr);
1444 if (remove_reference(hdr, NULL, tag) > 0) {
1445 ASSERT(HDR_IO_ERROR(hdr));
1446 arc_buf_destroy(buf, FALSE, TRUE);
1448 arc_hdr_destroy(hdr);
1454 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1456 arc_buf_hdr_t *hdr = buf->b_hdr;
1457 kmutex_t *hash_lock = HDR_LOCK(hdr);
1458 int no_callback = (buf->b_efunc == NULL);
1460 if (hdr->b_state == arc_anon) {
1461 arc_buf_free(buf, tag);
1462 return (no_callback);
1465 mutex_enter(hash_lock);
1466 ASSERT(hdr->b_state != arc_anon);
1467 ASSERT(buf->b_data != NULL);
1469 (void) remove_reference(hdr, hash_lock, tag);
1470 if (hdr->b_datacnt > 1) {
1472 arc_buf_destroy(buf, FALSE, TRUE);
1473 } else if (no_callback) {
1474 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1475 hdr->b_flags |= ARC_BUF_AVAILABLE;
1477 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1478 refcount_is_zero(&hdr->b_refcnt));
1479 mutex_exit(hash_lock);
1480 return (no_callback);
1484 arc_buf_size(arc_buf_t *buf)
1486 return (buf->b_hdr->b_size);
1490 * Evict buffers from list until we've removed the specified number of
1491 * bytes. Move the removed buffers to the appropriate evict state.
1492 * If the recycle flag is set, then attempt to "recycle" a buffer:
1493 * - look for a buffer to evict that is `bytes' long.
1494 * - return the data block from this buffer rather than freeing it.
1495 * This flag is used by callers that are trying to make space for a
1496 * new buffer in a full arc cache.
1498 * This function makes a "best effort". It skips over any buffers
1499 * it can't get a hash_lock on, and so may not catch all candidates.
1500 * It may also return without evicting as much space as requested.
1503 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1504 arc_buf_contents_t type)
1506 arc_state_t *evicted_state;
1507 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1508 arc_buf_hdr_t *ab, *ab_prev = NULL;
1509 list_t *list = &state->arcs_list[type];
1510 kmutex_t *hash_lock;
1511 boolean_t have_lock;
1512 void *stolen = NULL;
1514 ASSERT(state == arc_mru || state == arc_mfu);
1516 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1518 mutex_enter(&state->arcs_mtx);
1519 mutex_enter(&evicted_state->arcs_mtx);
1521 for (ab = list_tail(list); ab; ab = ab_prev) {
1522 ab_prev = list_prev(list, ab);
1523 /* prefetch buffers have a minimum lifespan */
1524 if (HDR_IO_IN_PROGRESS(ab) ||
1525 (spa && ab->b_spa != spa) ||
1526 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1527 lbolt - ab->b_arc_access < arc_min_prefetch_lifespan)) {
1531 /* "lookahead" for better eviction candidate */
1532 if (recycle && ab->b_size != bytes &&
1533 ab_prev && ab_prev->b_size == bytes)
1535 hash_lock = HDR_LOCK(ab);
1536 have_lock = MUTEX_HELD(hash_lock);
1537 if (have_lock || mutex_tryenter(hash_lock)) {
1538 ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0);
1539 ASSERT(ab->b_datacnt > 0);
1541 arc_buf_t *buf = ab->b_buf;
1542 if (!rw_tryenter(&buf->b_lock, RW_WRITER)) {
1547 bytes_evicted += ab->b_size;
1548 if (recycle && ab->b_type == type &&
1549 ab->b_size == bytes &&
1550 !HDR_L2_WRITING(ab)) {
1551 stolen = buf->b_data;
1556 mutex_enter(&arc_eviction_mtx);
1557 arc_buf_destroy(buf,
1558 buf->b_data == stolen, FALSE);
1559 ab->b_buf = buf->b_next;
1560 buf->b_hdr = &arc_eviction_hdr;
1561 buf->b_next = arc_eviction_list;
1562 arc_eviction_list = buf;
1563 mutex_exit(&arc_eviction_mtx);
1564 rw_exit(&buf->b_lock);
1566 rw_exit(&buf->b_lock);
1567 arc_buf_destroy(buf,
1568 buf->b_data == stolen, TRUE);
1571 if (ab->b_datacnt == 0) {
1572 arc_change_state(evicted_state, ab, hash_lock);
1573 ASSERT(HDR_IN_HASH_TABLE(ab));
1574 ab->b_flags |= ARC_IN_HASH_TABLE;
1575 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1576 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1579 mutex_exit(hash_lock);
1580 if (bytes >= 0 && bytes_evicted >= bytes)
1587 mutex_exit(&evicted_state->arcs_mtx);
1588 mutex_exit(&state->arcs_mtx);
1590 if (bytes_evicted < bytes)
1591 dprintf("only evicted %lld bytes from %x",
1592 (longlong_t)bytes_evicted, state);
1595 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1598 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1601 * We have just evicted some date into the ghost state, make
1602 * sure we also adjust the ghost state size if necessary.
1605 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1606 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1607 arc_mru_ghost->arcs_size - arc_c;
1609 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1611 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1612 arc_evict_ghost(arc_mru_ghost, NULL, todelete);
1613 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1614 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1615 arc_mru_ghost->arcs_size +
1616 arc_mfu_ghost->arcs_size - arc_c);
1617 arc_evict_ghost(arc_mfu_ghost, NULL, todelete);
1625 * Remove buffers from list until we've removed the specified number of
1626 * bytes. Destroy the buffers that are removed.
1629 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1631 arc_buf_hdr_t *ab, *ab_prev;
1632 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1633 kmutex_t *hash_lock;
1634 uint64_t bytes_deleted = 0;
1635 uint64_t bufs_skipped = 0;
1637 ASSERT(GHOST_STATE(state));
1639 mutex_enter(&state->arcs_mtx);
1640 for (ab = list_tail(list); ab; ab = ab_prev) {
1641 ab_prev = list_prev(list, ab);
1642 if (spa && ab->b_spa != spa)
1644 hash_lock = HDR_LOCK(ab);
1645 if (mutex_tryenter(hash_lock)) {
1646 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1647 ASSERT(ab->b_buf == NULL);
1648 ARCSTAT_BUMP(arcstat_deleted);
1649 bytes_deleted += ab->b_size;
1651 if (ab->b_l2hdr != NULL) {
1653 * This buffer is cached on the 2nd Level ARC;
1654 * don't destroy the header.
1656 arc_change_state(arc_l2c_only, ab, hash_lock);
1657 mutex_exit(hash_lock);
1659 arc_change_state(arc_anon, ab, hash_lock);
1660 mutex_exit(hash_lock);
1661 arc_hdr_destroy(ab);
1664 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1665 if (bytes >= 0 && bytes_deleted >= bytes)
1669 mutex_exit(&state->arcs_mtx);
1670 mutex_enter(hash_lock);
1671 mutex_exit(hash_lock);
1677 mutex_exit(&state->arcs_mtx);
1679 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1680 (bytes < 0 || bytes_deleted < bytes)) {
1681 list = &state->arcs_list[ARC_BUFC_METADATA];
1686 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1690 if (bytes_deleted < bytes)
1691 dprintf("only deleted %lld bytes from %p",
1692 (longlong_t)bytes_deleted, state);
1698 int64_t adjustment, delta;
1704 adjustment = MIN(arc_size - arc_c,
1705 arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - arc_p);
1707 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1708 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
1709 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA);
1710 adjustment -= delta;
1713 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1714 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
1715 (void) arc_evict(arc_mru, NULL, delta, FALSE,
1723 adjustment = arc_size - arc_c;
1725 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1726 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
1727 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA);
1728 adjustment -= delta;
1731 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1732 int64_t delta = MIN(adjustment,
1733 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
1734 (void) arc_evict(arc_mfu, NULL, delta, FALSE,
1739 * Adjust ghost lists
1742 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
1744 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
1745 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
1746 arc_evict_ghost(arc_mru_ghost, NULL, delta);
1750 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
1752 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
1753 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
1754 arc_evict_ghost(arc_mfu_ghost, NULL, delta);
1759 arc_do_user_evicts(void)
1761 mutex_enter(&arc_eviction_mtx);
1762 while (arc_eviction_list != NULL) {
1763 arc_buf_t *buf = arc_eviction_list;
1764 arc_eviction_list = buf->b_next;
1765 rw_enter(&buf->b_lock, RW_WRITER);
1767 rw_exit(&buf->b_lock);
1768 mutex_exit(&arc_eviction_mtx);
1770 if (buf->b_efunc != NULL)
1771 VERIFY(buf->b_efunc(buf) == 0);
1773 buf->b_efunc = NULL;
1774 buf->b_private = NULL;
1775 kmem_cache_free(buf_cache, buf);
1776 mutex_enter(&arc_eviction_mtx);
1778 mutex_exit(&arc_eviction_mtx);
1782 * Flush all *evictable* data from the cache for the given spa.
1783 * NOTE: this will not touch "active" (i.e. referenced) data.
1786 arc_flush(spa_t *spa)
1791 guid = spa_guid(spa);
1793 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
1794 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
1798 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
1799 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
1803 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
1804 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
1808 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
1809 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
1814 arc_evict_ghost(arc_mru_ghost, guid, -1);
1815 arc_evict_ghost(arc_mfu_ghost, guid, -1);
1817 mutex_enter(&arc_reclaim_thr_lock);
1818 arc_do_user_evicts();
1819 mutex_exit(&arc_reclaim_thr_lock);
1820 ASSERT(spa || arc_eviction_list == NULL);
1826 if (arc_c > arc_c_min) {
1830 to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
1832 to_free = arc_c >> arc_shrink_shift;
1834 if (arc_c > arc_c_min + to_free)
1835 atomic_add_64(&arc_c, -to_free);
1839 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
1840 if (arc_c > arc_size)
1841 arc_c = MAX(arc_size, arc_c_min);
1843 arc_p = (arc_c >> 1);
1844 ASSERT(arc_c >= arc_c_min);
1845 ASSERT((int64_t)arc_p >= 0);
1848 if (arc_size > arc_c)
1853 arc_reclaim_needed(void)
1863 * take 'desfree' extra pages, so we reclaim sooner, rather than later
1868 * check that we're out of range of the pageout scanner. It starts to
1869 * schedule paging if freemem is less than lotsfree and needfree.
1870 * lotsfree is the high-water mark for pageout, and needfree is the
1871 * number of needed free pages. We add extra pages here to make sure
1872 * the scanner doesn't start up while we're freeing memory.
1874 if (freemem < lotsfree + needfree + extra)
1878 * check to make sure that swapfs has enough space so that anon
1879 * reservations can still succeed. anon_resvmem() checks that the
1880 * availrmem is greater than swapfs_minfree, and the number of reserved
1881 * swap pages. We also add a bit of extra here just to prevent
1882 * circumstances from getting really dire.
1884 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
1889 * If we're on an i386 platform, it's possible that we'll exhaust the
1890 * kernel heap space before we ever run out of available physical
1891 * memory. Most checks of the size of the heap_area compare against
1892 * tune.t_minarmem, which is the minimum available real memory that we
1893 * can have in the system. However, this is generally fixed at 25 pages
1894 * which is so low that it's useless. In this comparison, we seek to
1895 * calculate the total heap-size, and reclaim if more than 3/4ths of the
1896 * heap is allocated. (Or, in the calculation, if less than 1/4th is
1899 if (btop(vmem_size(heap_arena, VMEM_FREE)) <
1900 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
1905 if (spa_get_random(100) == 0)
1912 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
1915 kmem_cache_t *prev_cache = NULL;
1916 kmem_cache_t *prev_data_cache = NULL;
1917 extern kmem_cache_t *zio_buf_cache[];
1918 extern kmem_cache_t *zio_data_buf_cache[];
1921 if (arc_meta_used >= arc_meta_limit) {
1923 * We are exceeding our meta-data cache limit.
1924 * Purge some DNLC entries to release holds on meta-data.
1926 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
1930 * Reclaim unused memory from all kmem caches.
1937 * An aggressive reclamation will shrink the cache size as well as
1938 * reap free buffers from the arc kmem caches.
1940 if (strat == ARC_RECLAIM_AGGR)
1943 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
1944 if (zio_buf_cache[i] != prev_cache) {
1945 prev_cache = zio_buf_cache[i];
1946 kmem_cache_reap_now(zio_buf_cache[i]);
1948 if (zio_data_buf_cache[i] != prev_data_cache) {
1949 prev_data_cache = zio_data_buf_cache[i];
1950 kmem_cache_reap_now(zio_data_buf_cache[i]);
1953 kmem_cache_reap_now(buf_cache);
1954 kmem_cache_reap_now(hdr_cache);
1958 arc_reclaim_thread(void)
1960 clock_t growtime = 0;
1961 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
1964 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
1966 mutex_enter(&arc_reclaim_thr_lock);
1967 while (arc_thread_exit == 0) {
1968 if (arc_reclaim_needed()) {
1971 if (last_reclaim == ARC_RECLAIM_CONS) {
1972 last_reclaim = ARC_RECLAIM_AGGR;
1974 last_reclaim = ARC_RECLAIM_CONS;
1978 last_reclaim = ARC_RECLAIM_AGGR;
1982 /* reset the growth delay for every reclaim */
1983 growtime = lbolt + (arc_grow_retry * hz);
1985 arc_kmem_reap_now(last_reclaim);
1988 } else if (arc_no_grow && lbolt >= growtime) {
1989 arc_no_grow = FALSE;
1992 if (2 * arc_c < arc_size +
1993 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size)
1996 if (arc_eviction_list != NULL)
1997 arc_do_user_evicts();
1999 /* block until needed, or one second, whichever is shorter */
2000 CALLB_CPR_SAFE_BEGIN(&cpr);
2001 (void) cv_timedwait(&arc_reclaim_thr_cv,
2002 &arc_reclaim_thr_lock, (lbolt + hz));
2003 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2006 arc_thread_exit = 0;
2007 cv_broadcast(&arc_reclaim_thr_cv);
2008 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2013 * Adapt arc info given the number of bytes we are trying to add and
2014 * the state that we are comming from. This function is only called
2015 * when we are adding new content to the cache.
2018 arc_adapt(int bytes, arc_state_t *state)
2021 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2023 if (state == arc_l2c_only)
2028 * Adapt the target size of the MRU list:
2029 * - if we just hit in the MRU ghost list, then increase
2030 * the target size of the MRU list.
2031 * - if we just hit in the MFU ghost list, then increase
2032 * the target size of the MFU list by decreasing the
2033 * target size of the MRU list.
2035 if (state == arc_mru_ghost) {
2036 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2037 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2039 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2040 } else if (state == arc_mfu_ghost) {
2043 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2044 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2046 delta = MIN(bytes * mult, arc_p);
2047 arc_p = MAX(arc_p_min, arc_p - delta);
2049 ASSERT((int64_t)arc_p >= 0);
2051 if (arc_reclaim_needed()) {
2052 cv_signal(&arc_reclaim_thr_cv);
2059 if (arc_c >= arc_c_max)
2063 * If we're within (2 * maxblocksize) bytes of the target
2064 * cache size, increment the target cache size
2066 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2067 atomic_add_64(&arc_c, (int64_t)bytes);
2068 if (arc_c > arc_c_max)
2070 else if (state == arc_anon)
2071 atomic_add_64(&arc_p, (int64_t)bytes);
2075 ASSERT((int64_t)arc_p >= 0);
2079 * Check if the cache has reached its limits and eviction is required
2083 arc_evict_needed(arc_buf_contents_t type)
2085 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2090 * If zio data pages are being allocated out of a separate heap segment,
2091 * then enforce that the size of available vmem for this area remains
2092 * above about 1/32nd free.
2094 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2095 vmem_size(zio_arena, VMEM_FREE) <
2096 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2100 if (arc_reclaim_needed())
2103 return (arc_size > arc_c);
2107 * The buffer, supplied as the first argument, needs a data block.
2108 * So, if we are at cache max, determine which cache should be victimized.
2109 * We have the following cases:
2111 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2112 * In this situation if we're out of space, but the resident size of the MFU is
2113 * under the limit, victimize the MFU cache to satisfy this insertion request.
2115 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2116 * Here, we've used up all of the available space for the MRU, so we need to
2117 * evict from our own cache instead. Evict from the set of resident MRU
2120 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2121 * c minus p represents the MFU space in the cache, since p is the size of the
2122 * cache that is dedicated to the MRU. In this situation there's still space on
2123 * the MFU side, so the MRU side needs to be victimized.
2125 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2126 * MFU's resident set is consuming more space than it has been allotted. In
2127 * this situation, we must victimize our own cache, the MFU, for this insertion.
2130 arc_get_data_buf(arc_buf_t *buf)
2132 arc_state_t *state = buf->b_hdr->b_state;
2133 uint64_t size = buf->b_hdr->b_size;
2134 arc_buf_contents_t type = buf->b_hdr->b_type;
2136 arc_adapt(size, state);
2139 * We have not yet reached cache maximum size,
2140 * just allocate a new buffer.
2142 if (!arc_evict_needed(type)) {
2143 if (type == ARC_BUFC_METADATA) {
2144 buf->b_data = zio_buf_alloc(size);
2145 arc_space_consume(size, ARC_SPACE_DATA);
2147 ASSERT(type == ARC_BUFC_DATA);
2148 buf->b_data = zio_data_buf_alloc(size);
2149 ARCSTAT_INCR(arcstat_data_size, size);
2150 atomic_add_64(&arc_size, size);
2156 * If we are prefetching from the mfu ghost list, this buffer
2157 * will end up on the mru list; so steal space from there.
2159 if (state == arc_mfu_ghost)
2160 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2161 else if (state == arc_mru_ghost)
2164 if (state == arc_mru || state == arc_anon) {
2165 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2166 state = (arc_mfu->arcs_lsize[type] >= size &&
2167 arc_p > mru_used) ? arc_mfu : arc_mru;
2170 uint64_t mfu_space = arc_c - arc_p;
2171 state = (arc_mru->arcs_lsize[type] >= size &&
2172 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2174 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) {
2175 if (type == ARC_BUFC_METADATA) {
2176 buf->b_data = zio_buf_alloc(size);
2177 arc_space_consume(size, ARC_SPACE_DATA);
2179 ASSERT(type == ARC_BUFC_DATA);
2180 buf->b_data = zio_data_buf_alloc(size);
2181 ARCSTAT_INCR(arcstat_data_size, size);
2182 atomic_add_64(&arc_size, size);
2184 ARCSTAT_BUMP(arcstat_recycle_miss);
2186 ASSERT(buf->b_data != NULL);
2189 * Update the state size. Note that ghost states have a
2190 * "ghost size" and so don't need to be updated.
2192 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2193 arc_buf_hdr_t *hdr = buf->b_hdr;
2195 atomic_add_64(&hdr->b_state->arcs_size, size);
2196 if (list_link_active(&hdr->b_arc_node)) {
2197 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2198 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2201 * If we are growing the cache, and we are adding anonymous
2202 * data, and we have outgrown arc_p, update arc_p
2204 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2205 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2206 arc_p = MIN(arc_c, arc_p + size);
2211 * This routine is called whenever a buffer is accessed.
2212 * NOTE: the hash lock is dropped in this function.
2215 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2217 ASSERT(MUTEX_HELD(hash_lock));
2219 if (buf->b_state == arc_anon) {
2221 * This buffer is not in the cache, and does not
2222 * appear in our "ghost" list. Add the new buffer
2226 ASSERT(buf->b_arc_access == 0);
2227 buf->b_arc_access = lbolt;
2228 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2229 arc_change_state(arc_mru, buf, hash_lock);
2231 } else if (buf->b_state == arc_mru) {
2233 * If this buffer is here because of a prefetch, then either:
2234 * - clear the flag if this is a "referencing" read
2235 * (any subsequent access will bump this into the MFU state).
2237 * - move the buffer to the head of the list if this is
2238 * another prefetch (to make it less likely to be evicted).
2240 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2241 if (refcount_count(&buf->b_refcnt) == 0) {
2242 ASSERT(list_link_active(&buf->b_arc_node));
2244 buf->b_flags &= ~ARC_PREFETCH;
2245 ARCSTAT_BUMP(arcstat_mru_hits);
2247 buf->b_arc_access = lbolt;
2252 * This buffer has been "accessed" only once so far,
2253 * but it is still in the cache. Move it to the MFU
2256 if (lbolt > buf->b_arc_access + ARC_MINTIME) {
2258 * More than 125ms have passed since we
2259 * instantiated this buffer. Move it to the
2260 * most frequently used state.
2262 buf->b_arc_access = lbolt;
2263 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2264 arc_change_state(arc_mfu, buf, hash_lock);
2266 ARCSTAT_BUMP(arcstat_mru_hits);
2267 } else if (buf->b_state == arc_mru_ghost) {
2268 arc_state_t *new_state;
2270 * This buffer has been "accessed" recently, but
2271 * was evicted from the cache. Move it to the
2275 if (buf->b_flags & ARC_PREFETCH) {
2276 new_state = arc_mru;
2277 if (refcount_count(&buf->b_refcnt) > 0)
2278 buf->b_flags &= ~ARC_PREFETCH;
2279 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2281 new_state = arc_mfu;
2282 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2285 buf->b_arc_access = lbolt;
2286 arc_change_state(new_state, buf, hash_lock);
2288 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2289 } else if (buf->b_state == arc_mfu) {
2291 * This buffer has been accessed more than once and is
2292 * still in the cache. Keep it in the MFU state.
2294 * NOTE: an add_reference() that occurred when we did
2295 * the arc_read() will have kicked this off the list.
2296 * If it was a prefetch, we will explicitly move it to
2297 * the head of the list now.
2299 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2300 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2301 ASSERT(list_link_active(&buf->b_arc_node));
2303 ARCSTAT_BUMP(arcstat_mfu_hits);
2304 buf->b_arc_access = lbolt;
2305 } else if (buf->b_state == arc_mfu_ghost) {
2306 arc_state_t *new_state = arc_mfu;
2308 * This buffer has been accessed more than once but has
2309 * been evicted from the cache. Move it back to the
2313 if (buf->b_flags & ARC_PREFETCH) {
2315 * This is a prefetch access...
2316 * move this block back to the MRU state.
2318 ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0);
2319 new_state = arc_mru;
2322 buf->b_arc_access = lbolt;
2323 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2324 arc_change_state(new_state, buf, hash_lock);
2326 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2327 } else if (buf->b_state == arc_l2c_only) {
2329 * This buffer is on the 2nd Level ARC.
2332 buf->b_arc_access = lbolt;
2333 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2334 arc_change_state(arc_mfu, buf, hash_lock);
2336 ASSERT(!"invalid arc state");
2340 /* a generic arc_done_func_t which you can use */
2343 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2345 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2346 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2349 /* a generic arc_done_func_t */
2351 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2353 arc_buf_t **bufp = arg;
2354 if (zio && zio->io_error) {
2355 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2363 arc_read_done(zio_t *zio)
2365 arc_buf_hdr_t *hdr, *found;
2367 arc_buf_t *abuf; /* buffer we're assigning to callback */
2368 kmutex_t *hash_lock;
2369 arc_callback_t *callback_list, *acb;
2370 int freeable = FALSE;
2372 buf = zio->io_private;
2376 * The hdr was inserted into hash-table and removed from lists
2377 * prior to starting I/O. We should find this header, since
2378 * it's in the hash table, and it should be legit since it's
2379 * not possible to evict it during the I/O. The only possible
2380 * reason for it not to be found is if we were freed during the
2383 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2386 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2387 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2388 (found == hdr && HDR_L2_READING(hdr)));
2390 hdr->b_flags &= ~ARC_L2_EVICTED;
2391 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2392 hdr->b_flags &= ~ARC_L2CACHE;
2394 /* byteswap if necessary */
2395 callback_list = hdr->b_acb;
2396 ASSERT(callback_list != NULL);
2397 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
2398 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2399 byteswap_uint64_array :
2400 dmu_ot[BP_GET_TYPE(zio->io_bp)].ot_byteswap;
2401 func(buf->b_data, hdr->b_size);
2404 arc_cksum_compute(buf, B_FALSE);
2406 /* create copies of the data buffer for the callers */
2408 for (acb = callback_list; acb; acb = acb->acb_next) {
2409 if (acb->acb_done) {
2411 abuf = arc_buf_clone(buf);
2412 acb->acb_buf = abuf;
2417 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2418 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2420 hdr->b_flags |= ARC_BUF_AVAILABLE;
2422 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2424 if (zio->io_error != 0) {
2425 hdr->b_flags |= ARC_IO_ERROR;
2426 if (hdr->b_state != arc_anon)
2427 arc_change_state(arc_anon, hdr, hash_lock);
2428 if (HDR_IN_HASH_TABLE(hdr))
2429 buf_hash_remove(hdr);
2430 freeable = refcount_is_zero(&hdr->b_refcnt);
2434 * Broadcast before we drop the hash_lock to avoid the possibility
2435 * that the hdr (and hence the cv) might be freed before we get to
2436 * the cv_broadcast().
2438 cv_broadcast(&hdr->b_cv);
2442 * Only call arc_access on anonymous buffers. This is because
2443 * if we've issued an I/O for an evicted buffer, we've already
2444 * called arc_access (to prevent any simultaneous readers from
2445 * getting confused).
2447 if (zio->io_error == 0 && hdr->b_state == arc_anon)
2448 arc_access(hdr, hash_lock);
2449 mutex_exit(hash_lock);
2452 * This block was freed while we waited for the read to
2453 * complete. It has been removed from the hash table and
2454 * moved to the anonymous state (so that it won't show up
2457 ASSERT3P(hdr->b_state, ==, arc_anon);
2458 freeable = refcount_is_zero(&hdr->b_refcnt);
2461 /* execute each callback and free its structure */
2462 while ((acb = callback_list) != NULL) {
2464 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2466 if (acb->acb_zio_dummy != NULL) {
2467 acb->acb_zio_dummy->io_error = zio->io_error;
2468 zio_nowait(acb->acb_zio_dummy);
2471 callback_list = acb->acb_next;
2472 kmem_free(acb, sizeof (arc_callback_t));
2476 arc_hdr_destroy(hdr);
2480 * "Read" the block block at the specified DVA (in bp) via the
2481 * cache. If the block is found in the cache, invoke the provided
2482 * callback immediately and return. Note that the `zio' parameter
2483 * in the callback will be NULL in this case, since no IO was
2484 * required. If the block is not in the cache pass the read request
2485 * on to the spa with a substitute callback function, so that the
2486 * requested block will be added to the cache.
2488 * If a read request arrives for a block that has a read in-progress,
2489 * either wait for the in-progress read to complete (and return the
2490 * results); or, if this is a read with a "done" func, add a record
2491 * to the read to invoke the "done" func when the read completes,
2492 * and return; or just return.
2494 * arc_read_done() will invoke all the requested "done" functions
2495 * for readers of this block.
2497 * Normal callers should use arc_read and pass the arc buffer and offset
2498 * for the bp. But if you know you don't need locking, you can use
2502 arc_read(zio_t *pio, spa_t *spa, blkptr_t *bp, arc_buf_t *pbuf,
2503 arc_done_func_t *done, void *private, int priority, int zio_flags,
2504 uint32_t *arc_flags, const zbookmark_t *zb)
2507 arc_buf_hdr_t *hdr = pbuf->b_hdr;
2509 ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt));
2510 ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size);
2511 rw_enter(&pbuf->b_lock, RW_READER);
2513 err = arc_read_nolock(pio, spa, bp, done, private, priority,
2514 zio_flags, arc_flags, zb);
2516 ASSERT3P(hdr, ==, pbuf->b_hdr);
2517 rw_exit(&pbuf->b_lock);
2522 arc_read_nolock(zio_t *pio, spa_t *spa, blkptr_t *bp,
2523 arc_done_func_t *done, void *private, int priority, int zio_flags,
2524 uint32_t *arc_flags, const zbookmark_t *zb)
2528 kmutex_t *hash_lock;
2530 uint64_t guid = spa_guid(spa);
2533 hdr = buf_hash_find(guid, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
2534 if (hdr && hdr->b_datacnt > 0) {
2536 *arc_flags |= ARC_CACHED;
2538 if (HDR_IO_IN_PROGRESS(hdr)) {
2540 if (*arc_flags & ARC_WAIT) {
2541 cv_wait(&hdr->b_cv, hash_lock);
2542 mutex_exit(hash_lock);
2545 ASSERT(*arc_flags & ARC_NOWAIT);
2548 arc_callback_t *acb = NULL;
2550 acb = kmem_zalloc(sizeof (arc_callback_t),
2552 acb->acb_done = done;
2553 acb->acb_private = private;
2555 acb->acb_zio_dummy = zio_null(pio,
2556 spa, NULL, NULL, NULL, zio_flags);
2558 ASSERT(acb->acb_done != NULL);
2559 acb->acb_next = hdr->b_acb;
2561 add_reference(hdr, hash_lock, private);
2562 mutex_exit(hash_lock);
2565 mutex_exit(hash_lock);
2569 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2572 add_reference(hdr, hash_lock, private);
2574 * If this block is already in use, create a new
2575 * copy of the data so that we will be guaranteed
2576 * that arc_release() will always succeed.
2580 ASSERT(buf->b_data);
2581 if (HDR_BUF_AVAILABLE(hdr)) {
2582 ASSERT(buf->b_efunc == NULL);
2583 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2585 buf = arc_buf_clone(buf);
2587 } else if (*arc_flags & ARC_PREFETCH &&
2588 refcount_count(&hdr->b_refcnt) == 0) {
2589 hdr->b_flags |= ARC_PREFETCH;
2591 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2592 arc_access(hdr, hash_lock);
2593 if (*arc_flags & ARC_L2CACHE)
2594 hdr->b_flags |= ARC_L2CACHE;
2595 mutex_exit(hash_lock);
2596 ARCSTAT_BUMP(arcstat_hits);
2597 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2598 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2599 data, metadata, hits);
2602 done(NULL, buf, private);
2604 uint64_t size = BP_GET_LSIZE(bp);
2605 arc_callback_t *acb;
2608 boolean_t devw = B_FALSE;
2611 /* this block is not in the cache */
2612 arc_buf_hdr_t *exists;
2613 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2614 buf = arc_buf_alloc(spa, size, private, type);
2616 hdr->b_dva = *BP_IDENTITY(bp);
2617 hdr->b_birth = bp->blk_birth;
2618 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2619 exists = buf_hash_insert(hdr, &hash_lock);
2621 /* somebody beat us to the hash insert */
2622 mutex_exit(hash_lock);
2623 bzero(&hdr->b_dva, sizeof (dva_t));
2626 (void) arc_buf_remove_ref(buf, private);
2627 goto top; /* restart the IO request */
2629 /* if this is a prefetch, we don't have a reference */
2630 if (*arc_flags & ARC_PREFETCH) {
2631 (void) remove_reference(hdr, hash_lock,
2633 hdr->b_flags |= ARC_PREFETCH;
2635 if (*arc_flags & ARC_L2CACHE)
2636 hdr->b_flags |= ARC_L2CACHE;
2637 if (BP_GET_LEVEL(bp) > 0)
2638 hdr->b_flags |= ARC_INDIRECT;
2640 /* this block is in the ghost cache */
2641 ASSERT(GHOST_STATE(hdr->b_state));
2642 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2643 ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0);
2644 ASSERT(hdr->b_buf == NULL);
2646 /* if this is a prefetch, we don't have a reference */
2647 if (*arc_flags & ARC_PREFETCH)
2648 hdr->b_flags |= ARC_PREFETCH;
2650 add_reference(hdr, hash_lock, private);
2651 if (*arc_flags & ARC_L2CACHE)
2652 hdr->b_flags |= ARC_L2CACHE;
2653 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2656 buf->b_efunc = NULL;
2657 buf->b_private = NULL;
2660 arc_get_data_buf(buf);
2661 ASSERT(hdr->b_datacnt == 0);
2666 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2667 acb->acb_done = done;
2668 acb->acb_private = private;
2670 ASSERT(hdr->b_acb == NULL);
2672 hdr->b_flags |= ARC_IO_IN_PROGRESS;
2675 * If the buffer has been evicted, migrate it to a present state
2676 * before issuing the I/O. Once we drop the hash-table lock,
2677 * the header will be marked as I/O in progress and have an
2678 * attached buffer. At this point, anybody who finds this
2679 * buffer ought to notice that it's legit but has a pending I/O.
2682 if (GHOST_STATE(hdr->b_state))
2683 arc_access(hdr, hash_lock);
2685 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
2686 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
2687 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
2688 addr = hdr->b_l2hdr->b_daddr;
2690 * Lock out device removal.
2692 if (vdev_is_dead(vd) ||
2693 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
2697 mutex_exit(hash_lock);
2699 ASSERT3U(hdr->b_size, ==, size);
2700 DTRACE_PROBE3(arc__miss, blkptr_t *, bp, uint64_t, size,
2702 ARCSTAT_BUMP(arcstat_misses);
2703 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2704 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2705 data, metadata, misses);
2707 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
2709 * Read from the L2ARC if the following are true:
2710 * 1. The L2ARC vdev was previously cached.
2711 * 2. This buffer still has L2ARC metadata.
2712 * 3. This buffer isn't currently writing to the L2ARC.
2713 * 4. The L2ARC entry wasn't evicted, which may
2714 * also have invalidated the vdev.
2715 * 5. This isn't prefetch and l2arc_noprefetch is set.
2717 if (hdr->b_l2hdr != NULL &&
2718 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
2719 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
2720 l2arc_read_callback_t *cb;
2722 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
2723 ARCSTAT_BUMP(arcstat_l2_hits);
2725 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
2727 cb->l2rcb_buf = buf;
2728 cb->l2rcb_spa = spa;
2731 cb->l2rcb_flags = zio_flags;
2734 * l2arc read. The SCL_L2ARC lock will be
2735 * released by l2arc_read_done().
2737 rzio = zio_read_phys(pio, vd, addr, size,
2738 buf->b_data, ZIO_CHECKSUM_OFF,
2739 l2arc_read_done, cb, priority, zio_flags |
2740 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
2741 ZIO_FLAG_DONT_PROPAGATE |
2742 ZIO_FLAG_DONT_RETRY, B_FALSE);
2743 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
2745 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
2747 if (*arc_flags & ARC_NOWAIT) {
2752 ASSERT(*arc_flags & ARC_WAIT);
2753 if (zio_wait(rzio) == 0)
2756 /* l2arc read error; goto zio_read() */
2758 DTRACE_PROBE1(l2arc__miss,
2759 arc_buf_hdr_t *, hdr);
2760 ARCSTAT_BUMP(arcstat_l2_misses);
2761 if (HDR_L2_WRITING(hdr))
2762 ARCSTAT_BUMP(arcstat_l2_rw_clash);
2763 spa_config_exit(spa, SCL_L2ARC, vd);
2767 spa_config_exit(spa, SCL_L2ARC, vd);
2768 if (l2arc_ndev != 0) {
2769 DTRACE_PROBE1(l2arc__miss,
2770 arc_buf_hdr_t *, hdr);
2771 ARCSTAT_BUMP(arcstat_l2_misses);
2775 rzio = zio_read(pio, spa, bp, buf->b_data, size,
2776 arc_read_done, buf, priority, zio_flags, zb);
2778 if (*arc_flags & ARC_WAIT)
2779 return (zio_wait(rzio));
2781 ASSERT(*arc_flags & ARC_NOWAIT);
2788 * arc_read() variant to support pool traversal. If the block is already
2789 * in the ARC, make a copy of it; otherwise, the caller will do the I/O.
2790 * The idea is that we don't want pool traversal filling up memory, but
2791 * if the ARC already has the data anyway, we shouldn't pay for the I/O.
2794 arc_tryread(spa_t *spa, blkptr_t *bp, void *data)
2798 uint64_t guid = spa_guid(spa);
2801 hdr = buf_hash_find(guid, BP_IDENTITY(bp), bp->blk_birth, &hash_mtx);
2803 if (hdr && hdr->b_datacnt > 0 && !HDR_IO_IN_PROGRESS(hdr)) {
2804 arc_buf_t *buf = hdr->b_buf;
2807 while (buf->b_data == NULL) {
2811 bcopy(buf->b_data, data, hdr->b_size);
2817 mutex_exit(hash_mtx);
2823 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
2825 ASSERT(buf->b_hdr != NULL);
2826 ASSERT(buf->b_hdr->b_state != arc_anon);
2827 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
2828 buf->b_efunc = func;
2829 buf->b_private = private;
2833 * This is used by the DMU to let the ARC know that a buffer is
2834 * being evicted, so the ARC should clean up. If this arc buf
2835 * is not yet in the evicted state, it will be put there.
2838 arc_buf_evict(arc_buf_t *buf)
2841 kmutex_t *hash_lock;
2844 rw_enter(&buf->b_lock, RW_WRITER);
2848 * We are in arc_do_user_evicts().
2850 ASSERT(buf->b_data == NULL);
2851 rw_exit(&buf->b_lock);
2853 } else if (buf->b_data == NULL) {
2854 arc_buf_t copy = *buf; /* structure assignment */
2856 * We are on the eviction list; process this buffer now
2857 * but let arc_do_user_evicts() do the reaping.
2859 buf->b_efunc = NULL;
2860 rw_exit(&buf->b_lock);
2861 VERIFY(copy.b_efunc(©) == 0);
2864 hash_lock = HDR_LOCK(hdr);
2865 mutex_enter(hash_lock);
2867 ASSERT(buf->b_hdr == hdr);
2868 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
2869 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2872 * Pull this buffer off of the hdr
2875 while (*bufp != buf)
2876 bufp = &(*bufp)->b_next;
2877 *bufp = buf->b_next;
2879 ASSERT(buf->b_data != NULL);
2880 arc_buf_destroy(buf, FALSE, FALSE);
2882 if (hdr->b_datacnt == 0) {
2883 arc_state_t *old_state = hdr->b_state;
2884 arc_state_t *evicted_state;
2886 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2889 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2891 mutex_enter(&old_state->arcs_mtx);
2892 mutex_enter(&evicted_state->arcs_mtx);
2894 arc_change_state(evicted_state, hdr, hash_lock);
2895 ASSERT(HDR_IN_HASH_TABLE(hdr));
2896 hdr->b_flags |= ARC_IN_HASH_TABLE;
2897 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2899 mutex_exit(&evicted_state->arcs_mtx);
2900 mutex_exit(&old_state->arcs_mtx);
2902 mutex_exit(hash_lock);
2903 rw_exit(&buf->b_lock);
2905 VERIFY(buf->b_efunc(buf) == 0);
2906 buf->b_efunc = NULL;
2907 buf->b_private = NULL;
2909 kmem_cache_free(buf_cache, buf);
2914 * Release this buffer from the cache. This must be done
2915 * after a read and prior to modifying the buffer contents.
2916 * If the buffer has more than one reference, we must make
2917 * a new hdr for the buffer.
2920 arc_release(arc_buf_t *buf, void *tag)
2923 kmutex_t *hash_lock;
2924 l2arc_buf_hdr_t *l2hdr;
2927 rw_enter(&buf->b_lock, RW_WRITER);
2930 /* this buffer is not on any list */
2931 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
2932 ASSERT(!(hdr->b_flags & ARC_STORED));
2934 if (hdr->b_state == arc_anon) {
2935 /* this buffer is already released */
2936 ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 1);
2937 ASSERT(BUF_EMPTY(hdr));
2938 ASSERT(buf->b_efunc == NULL);
2940 rw_exit(&buf->b_lock);
2944 hash_lock = HDR_LOCK(hdr);
2945 mutex_enter(hash_lock);
2947 l2hdr = hdr->b_l2hdr;
2949 mutex_enter(&l2arc_buflist_mtx);
2950 hdr->b_l2hdr = NULL;
2951 buf_size = hdr->b_size;
2955 * Do we have more than one buf?
2957 if (hdr->b_datacnt > 1) {
2958 arc_buf_hdr_t *nhdr;
2960 uint64_t blksz = hdr->b_size;
2961 uint64_t spa = hdr->b_spa;
2962 arc_buf_contents_t type = hdr->b_type;
2963 uint32_t flags = hdr->b_flags;
2965 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
2967 * Pull the data off of this buf and attach it to
2968 * a new anonymous buf.
2970 (void) remove_reference(hdr, hash_lock, tag);
2972 while (*bufp != buf)
2973 bufp = &(*bufp)->b_next;
2974 *bufp = (*bufp)->b_next;
2977 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
2978 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
2979 if (refcount_is_zero(&hdr->b_refcnt)) {
2980 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
2981 ASSERT3U(*size, >=, hdr->b_size);
2982 atomic_add_64(size, -hdr->b_size);
2984 hdr->b_datacnt -= 1;
2985 arc_cksum_verify(buf);
2987 mutex_exit(hash_lock);
2989 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
2990 nhdr->b_size = blksz;
2992 nhdr->b_type = type;
2994 nhdr->b_state = arc_anon;
2995 nhdr->b_arc_access = 0;
2996 nhdr->b_flags = flags & ARC_L2_WRITING;
2997 nhdr->b_l2hdr = NULL;
2998 nhdr->b_datacnt = 1;
2999 nhdr->b_freeze_cksum = NULL;
3000 (void) refcount_add(&nhdr->b_refcnt, tag);
3002 rw_exit(&buf->b_lock);
3003 atomic_add_64(&arc_anon->arcs_size, blksz);
3005 rw_exit(&buf->b_lock);
3006 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3007 ASSERT(!list_link_active(&hdr->b_arc_node));
3008 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3009 arc_change_state(arc_anon, hdr, hash_lock);
3010 hdr->b_arc_access = 0;
3011 mutex_exit(hash_lock);
3013 bzero(&hdr->b_dva, sizeof (dva_t));
3018 buf->b_efunc = NULL;
3019 buf->b_private = NULL;
3022 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3023 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3024 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3025 mutex_exit(&l2arc_buflist_mtx);
3030 arc_released(arc_buf_t *buf)
3034 rw_enter(&buf->b_lock, RW_READER);
3035 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3036 rw_exit(&buf->b_lock);
3041 arc_has_callback(arc_buf_t *buf)
3045 rw_enter(&buf->b_lock, RW_READER);
3046 callback = (buf->b_efunc != NULL);
3047 rw_exit(&buf->b_lock);
3053 arc_referenced(arc_buf_t *buf)
3057 rw_enter(&buf->b_lock, RW_READER);
3058 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3059 rw_exit(&buf->b_lock);
3060 return (referenced);
3065 arc_write_ready(zio_t *zio)
3067 arc_write_callback_t *callback = zio->io_private;
3068 arc_buf_t *buf = callback->awcb_buf;
3069 arc_buf_hdr_t *hdr = buf->b_hdr;
3071 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3072 callback->awcb_ready(zio, buf, callback->awcb_private);
3075 * If the IO is already in progress, then this is a re-write
3076 * attempt, so we need to thaw and re-compute the cksum.
3077 * It is the responsibility of the callback to handle the
3078 * accounting for any re-write attempt.
3080 if (HDR_IO_IN_PROGRESS(hdr)) {
3081 mutex_enter(&hdr->b_freeze_lock);
3082 if (hdr->b_freeze_cksum != NULL) {
3083 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3084 hdr->b_freeze_cksum = NULL;
3086 mutex_exit(&hdr->b_freeze_lock);
3088 arc_cksum_compute(buf, B_FALSE);
3089 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3093 arc_write_done(zio_t *zio)
3095 arc_write_callback_t *callback = zio->io_private;
3096 arc_buf_t *buf = callback->awcb_buf;
3097 arc_buf_hdr_t *hdr = buf->b_hdr;
3101 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3102 hdr->b_birth = zio->io_bp->blk_birth;
3103 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3105 * If the block to be written was all-zero, we may have
3106 * compressed it away. In this case no write was performed
3107 * so there will be no dva/birth-date/checksum. The buffer
3108 * must therefor remain anonymous (and uncached).
3110 if (!BUF_EMPTY(hdr)) {
3111 arc_buf_hdr_t *exists;
3112 kmutex_t *hash_lock;
3114 arc_cksum_verify(buf);
3116 exists = buf_hash_insert(hdr, &hash_lock);
3119 * This can only happen if we overwrite for
3120 * sync-to-convergence, because we remove
3121 * buffers from the hash table when we arc_free().
3123 ASSERT(zio->io_flags & ZIO_FLAG_IO_REWRITE);
3124 ASSERT(DVA_EQUAL(BP_IDENTITY(&zio->io_bp_orig),
3125 BP_IDENTITY(zio->io_bp)));
3126 ASSERT3U(zio->io_bp_orig.blk_birth, ==,
3127 zio->io_bp->blk_birth);
3129 ASSERT(refcount_is_zero(&exists->b_refcnt));
3130 arc_change_state(arc_anon, exists, hash_lock);
3131 mutex_exit(hash_lock);
3132 arc_hdr_destroy(exists);
3133 exists = buf_hash_insert(hdr, &hash_lock);
3134 ASSERT3P(exists, ==, NULL);
3136 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3137 /* if it's not anon, we are doing a scrub */
3138 if (hdr->b_state == arc_anon)
3139 arc_access(hdr, hash_lock);
3140 mutex_exit(hash_lock);
3141 } else if (callback->awcb_done == NULL) {
3144 * This is an anonymous buffer with no user callback,
3145 * destroy it if there are no active references.
3147 mutex_enter(&arc_eviction_mtx);
3148 destroy_hdr = refcount_is_zero(&hdr->b_refcnt);
3149 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3150 mutex_exit(&arc_eviction_mtx);
3152 arc_hdr_destroy(hdr);
3154 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3156 hdr->b_flags &= ~ARC_STORED;
3158 if (callback->awcb_done) {
3159 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3160 callback->awcb_done(zio, buf, callback->awcb_private);
3163 kmem_free(callback, sizeof (arc_write_callback_t));
3167 write_policy(spa_t *spa, const writeprops_t *wp, zio_prop_t *zp)
3169 boolean_t ismd = (wp->wp_level > 0 || dmu_ot[wp->wp_type].ot_metadata);
3171 /* Determine checksum setting */
3174 * Metadata always gets checksummed. If the data
3175 * checksum is multi-bit correctable, and it's not a
3176 * ZBT-style checksum, then it's suitable for metadata
3177 * as well. Otherwise, the metadata checksum defaults
3180 if (zio_checksum_table[wp->wp_oschecksum].ci_correctable &&
3181 !zio_checksum_table[wp->wp_oschecksum].ci_zbt)
3182 zp->zp_checksum = wp->wp_oschecksum;
3184 zp->zp_checksum = ZIO_CHECKSUM_FLETCHER_4;
3186 zp->zp_checksum = zio_checksum_select(wp->wp_dnchecksum,
3190 /* Determine compression setting */
3193 * XXX -- we should design a compression algorithm
3194 * that specializes in arrays of bps.
3196 zp->zp_compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
3199 zp->zp_compress = zio_compress_select(wp->wp_dncompress,
3203 zp->zp_type = wp->wp_type;
3204 zp->zp_level = wp->wp_level;
3205 zp->zp_ndvas = MIN(wp->wp_copies + ismd, spa_max_replication(spa));
3209 arc_write(zio_t *pio, spa_t *spa, const writeprops_t *wp,
3210 boolean_t l2arc, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
3211 arc_done_func_t *ready, arc_done_func_t *done, void *private, int priority,
3212 int zio_flags, const zbookmark_t *zb)
3214 arc_buf_hdr_t *hdr = buf->b_hdr;
3215 arc_write_callback_t *callback;
3219 ASSERT(ready != NULL);
3220 ASSERT(!HDR_IO_ERROR(hdr));
3221 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3222 ASSERT(hdr->b_acb == 0);
3224 hdr->b_flags |= ARC_L2CACHE;
3225 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3226 callback->awcb_ready = ready;
3227 callback->awcb_done = done;
3228 callback->awcb_private = private;
3229 callback->awcb_buf = buf;
3231 write_policy(spa, wp, &zp);
3232 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, &zp,
3233 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3239 arc_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
3240 zio_done_func_t *done, void *private, uint32_t arc_flags)
3243 kmutex_t *hash_lock;
3245 uint64_t guid = spa_guid(spa);
3248 * If this buffer is in the cache, release it, so it
3251 ab = buf_hash_find(guid, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
3254 * The checksum of blocks to free is not always
3255 * preserved (eg. on the deadlist). However, if it is
3256 * nonzero, it should match what we have in the cache.
3258 ASSERT(bp->blk_cksum.zc_word[0] == 0 ||
3259 bp->blk_cksum.zc_word[0] == ab->b_cksum0 ||
3260 bp->blk_fill == BLK_FILL_ALREADY_FREED);
3262 if (ab->b_state != arc_anon)
3263 arc_change_state(arc_anon, ab, hash_lock);
3264 if (HDR_IO_IN_PROGRESS(ab)) {
3266 * This should only happen when we prefetch.
3268 ASSERT(ab->b_flags & ARC_PREFETCH);
3269 ASSERT3U(ab->b_datacnt, ==, 1);
3270 ab->b_flags |= ARC_FREED_IN_READ;
3271 if (HDR_IN_HASH_TABLE(ab))
3272 buf_hash_remove(ab);
3273 ab->b_arc_access = 0;
3274 bzero(&ab->b_dva, sizeof (dva_t));
3277 ab->b_buf->b_efunc = NULL;
3278 ab->b_buf->b_private = NULL;
3279 mutex_exit(hash_lock);
3280 } else if (refcount_is_zero(&ab->b_refcnt)) {
3281 ab->b_flags |= ARC_FREE_IN_PROGRESS;
3282 mutex_exit(hash_lock);
3283 arc_hdr_destroy(ab);
3284 ARCSTAT_BUMP(arcstat_deleted);
3287 * We still have an active reference on this
3288 * buffer. This can happen, e.g., from
3289 * dbuf_unoverride().
3291 ASSERT(!HDR_IN_HASH_TABLE(ab));
3292 ab->b_arc_access = 0;
3293 bzero(&ab->b_dva, sizeof (dva_t));
3296 ab->b_buf->b_efunc = NULL;
3297 ab->b_buf->b_private = NULL;
3298 mutex_exit(hash_lock);
3302 zio = zio_free(pio, spa, txg, bp, done, private, ZIO_FLAG_MUSTSUCCEED);
3304 if (arc_flags & ARC_WAIT)
3305 return (zio_wait(zio));
3307 ASSERT(arc_flags & ARC_NOWAIT);
3314 arc_memory_throttle(uint64_t reserve, uint64_t txg)
3317 uint64_t inflight_data = arc_anon->arcs_size;
3318 uint64_t available_memory = ptob(freemem);
3319 static uint64_t page_load = 0;
3320 static uint64_t last_txg = 0;
3324 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3326 if (available_memory >= zfs_write_limit_max)
3329 if (txg > last_txg) {
3334 * If we are in pageout, we know that memory is already tight,
3335 * the arc is already going to be evicting, so we just want to
3336 * continue to let page writes occur as quickly as possible.
3338 if (curproc == proc_pageout) {
3339 if (page_load > MAX(ptob(minfree), available_memory) / 4)
3341 /* Note: reserve is inflated, so we deflate */
3342 page_load += reserve / 8;
3344 } else if (page_load > 0 && arc_reclaim_needed()) {
3345 /* memory is low, delay before restarting */
3346 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3351 if (arc_size > arc_c_min) {
3352 uint64_t evictable_memory =
3353 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3354 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3355 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3356 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3357 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3360 if (inflight_data > available_memory / 4) {
3361 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3369 arc_tempreserve_clear(uint64_t reserve)
3371 atomic_add_64(&arc_tempreserve, -reserve);
3372 ASSERT((int64_t)arc_tempreserve >= 0);
3376 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3382 * Once in a while, fail for no reason. Everything should cope.
3384 if (spa_get_random(10000) == 0) {
3385 dprintf("forcing random failure\n");
3389 if (reserve > arc_c/4 && !arc_no_grow)
3390 arc_c = MIN(arc_c_max, reserve * 4);
3391 if (reserve > arc_c)
3395 * Writes will, almost always, require additional memory allocations
3396 * in order to compress/encrypt/etc the data. We therefor need to
3397 * make sure that there is sufficient available memory for this.
3399 if (error = arc_memory_throttle(reserve, txg))
3403 * Throttle writes when the amount of dirty data in the cache
3404 * gets too large. We try to keep the cache less than half full
3405 * of dirty blocks so that our sync times don't grow too large.
3406 * Note: if two requests come in concurrently, we might let them
3407 * both succeed, when one of them should fail. Not a huge deal.
3409 if (reserve + arc_tempreserve + arc_anon->arcs_size > arc_c / 2 &&
3410 arc_anon->arcs_size > arc_c / 4) {
3411 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3412 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3413 arc_tempreserve>>10,
3414 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3415 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3416 reserve>>10, arc_c>>10);
3419 atomic_add_64(&arc_tempreserve, reserve);
3426 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3427 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3429 /* Convert seconds to clock ticks */
3430 arc_min_prefetch_lifespan = 1 * hz;
3432 /* Start out with 1/8 of all memory */
3433 arc_c = physmem * PAGESIZE / 8;
3437 * On architectures where the physical memory can be larger
3438 * than the addressable space (intel in 32-bit mode), we may
3439 * need to limit the cache to 1/8 of VM size.
3441 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3444 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3445 arc_c_min = MAX(arc_c / 4, 64<<20);
3446 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3447 if (arc_c * 8 >= 1<<30)
3448 arc_c_max = (arc_c * 8) - (1<<30);
3450 arc_c_max = arc_c_min;
3451 arc_c_max = MAX(arc_c * 6, arc_c_max);
3454 * Allow the tunables to override our calculations if they are
3455 * reasonable (ie. over 64MB)
3457 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3458 arc_c_max = zfs_arc_max;
3459 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3460 arc_c_min = zfs_arc_min;
3463 arc_p = (arc_c >> 1);
3465 /* limit meta-data to 1/4 of the arc capacity */
3466 arc_meta_limit = arc_c_max / 4;
3468 /* Allow the tunable to override if it is reasonable */
3469 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3470 arc_meta_limit = zfs_arc_meta_limit;
3472 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3473 arc_c_min = arc_meta_limit / 2;
3475 if (zfs_arc_grow_retry > 0)
3476 arc_grow_retry = zfs_arc_grow_retry;
3478 if (zfs_arc_shrink_shift > 0)
3479 arc_shrink_shift = zfs_arc_shrink_shift;
3481 if (zfs_arc_p_min_shift > 0)
3482 arc_p_min_shift = zfs_arc_p_min_shift;
3484 /* if kmem_flags are set, lets try to use less memory */
3485 if (kmem_debugging())
3487 if (arc_c < arc_c_min)
3490 arc_anon = &ARC_anon;
3492 arc_mru_ghost = &ARC_mru_ghost;
3494 arc_mfu_ghost = &ARC_mfu_ghost;
3495 arc_l2c_only = &ARC_l2c_only;
3498 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3499 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3500 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3501 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3502 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3503 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3505 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3506 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3507 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3508 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3509 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3510 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3511 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3512 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3513 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3514 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3515 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3516 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3517 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3518 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3519 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3520 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3521 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3522 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3523 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3524 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3528 arc_thread_exit = 0;
3529 arc_eviction_list = NULL;
3530 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3531 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3533 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3534 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3536 if (arc_ksp != NULL) {
3537 arc_ksp->ks_data = &arc_stats;
3538 kstat_install(arc_ksp);
3541 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3542 TS_RUN, minclsyspri);
3547 if (zfs_write_limit_max == 0)
3548 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3550 zfs_write_limit_shift = 0;
3551 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3557 mutex_enter(&arc_reclaim_thr_lock);
3558 arc_thread_exit = 1;
3559 while (arc_thread_exit != 0)
3560 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3561 mutex_exit(&arc_reclaim_thr_lock);
3567 if (arc_ksp != NULL) {
3568 kstat_delete(arc_ksp);
3572 mutex_destroy(&arc_eviction_mtx);
3573 mutex_destroy(&arc_reclaim_thr_lock);
3574 cv_destroy(&arc_reclaim_thr_cv);
3576 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3577 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3578 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3579 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3580 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3581 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3582 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3583 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3585 mutex_destroy(&arc_anon->arcs_mtx);
3586 mutex_destroy(&arc_mru->arcs_mtx);
3587 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3588 mutex_destroy(&arc_mfu->arcs_mtx);
3589 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3590 mutex_destroy(&arc_l2c_only->arcs_mtx);
3592 mutex_destroy(&zfs_write_limit_lock);
3600 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3601 * It uses dedicated storage devices to hold cached data, which are populated
3602 * using large infrequent writes. The main role of this cache is to boost
3603 * the performance of random read workloads. The intended L2ARC devices
3604 * include short-stroked disks, solid state disks, and other media with
3605 * substantially faster read latency than disk.
3607 * +-----------------------+
3609 * +-----------------------+
3612 * l2arc_feed_thread() arc_read()
3616 * +---------------+ |
3618 * +---------------+ |
3623 * +-------+ +-------+
3625 * | cache | | cache |
3626 * +-------+ +-------+
3627 * +=========+ .-----.
3628 * : L2ARC : |-_____-|
3629 * : devices : | Disks |
3630 * +=========+ `-_____-'
3632 * Read requests are satisfied from the following sources, in order:
3635 * 2) vdev cache of L2ARC devices
3637 * 4) vdev cache of disks
3640 * Some L2ARC device types exhibit extremely slow write performance.
3641 * To accommodate for this there are some significant differences between
3642 * the L2ARC and traditional cache design:
3644 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3645 * the ARC behave as usual, freeing buffers and placing headers on ghost
3646 * lists. The ARC does not send buffers to the L2ARC during eviction as
3647 * this would add inflated write latencies for all ARC memory pressure.
3649 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3650 * It does this by periodically scanning buffers from the eviction-end of
3651 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3652 * not already there. It scans until a headroom of buffers is satisfied,
3653 * which itself is a buffer for ARC eviction. The thread that does this is
3654 * l2arc_feed_thread(), illustrated below; example sizes are included to
3655 * provide a better sense of ratio than this diagram:
3658 * +---------------------+----------+
3659 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3660 * +---------------------+----------+ | o L2ARC eligible
3661 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
3662 * +---------------------+----------+ |
3663 * 15.9 Gbytes ^ 32 Mbytes |
3665 * l2arc_feed_thread()
3667 * l2arc write hand <--[oooo]--'
3671 * +==============================+
3672 * L2ARC dev |####|#|###|###| |####| ... |
3673 * +==============================+
3676 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
3677 * evicted, then the L2ARC has cached a buffer much sooner than it probably
3678 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
3679 * safe to say that this is an uncommon case, since buffers at the end of
3680 * the ARC lists have moved there due to inactivity.
3682 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
3683 * then the L2ARC simply misses copying some buffers. This serves as a
3684 * pressure valve to prevent heavy read workloads from both stalling the ARC
3685 * with waits and clogging the L2ARC with writes. This also helps prevent
3686 * the potential for the L2ARC to churn if it attempts to cache content too
3687 * quickly, such as during backups of the entire pool.
3689 * 5. After system boot and before the ARC has filled main memory, there are
3690 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
3691 * lists can remain mostly static. Instead of searching from tail of these
3692 * lists as pictured, the l2arc_feed_thread() will search from the list heads
3693 * for eligible buffers, greatly increasing its chance of finding them.
3695 * The L2ARC device write speed is also boosted during this time so that
3696 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
3697 * there are no L2ARC reads, and no fear of degrading read performance
3698 * through increased writes.
3700 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
3701 * the vdev queue can aggregate them into larger and fewer writes. Each
3702 * device is written to in a rotor fashion, sweeping writes through
3703 * available space then repeating.
3705 * 7. The L2ARC does not store dirty content. It never needs to flush
3706 * write buffers back to disk based storage.
3708 * 8. If an ARC buffer is written (and dirtied) which also exists in the
3709 * L2ARC, the now stale L2ARC buffer is immediately dropped.
3711 * The performance of the L2ARC can be tweaked by a number of tunables, which
3712 * may be necessary for different workloads:
3714 * l2arc_write_max max write bytes per interval
3715 * l2arc_write_boost extra write bytes during device warmup
3716 * l2arc_noprefetch skip caching prefetched buffers
3717 * l2arc_headroom number of max device writes to precache
3718 * l2arc_feed_secs seconds between L2ARC writing
3720 * Tunables may be removed or added as future performance improvements are
3721 * integrated, and also may become zpool properties.
3723 * There are three key functions that control how the L2ARC warms up:
3725 * l2arc_write_eligible() check if a buffer is eligible to cache
3726 * l2arc_write_size() calculate how much to write
3727 * l2arc_write_interval() calculate sleep delay between writes
3729 * These three functions determine what to write, how much, and how quickly
3734 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
3737 * A buffer is *not* eligible for the L2ARC if it:
3738 * 1. belongs to a different spa.
3739 * 2. has no attached buffer.
3740 * 3. is already cached on the L2ARC.
3741 * 4. has an I/O in progress (it may be an incomplete read).
3742 * 5. is flagged not eligible (zfs property).
3744 if (ab->b_spa != spa_guid || ab->b_buf == NULL || ab->b_l2hdr != NULL ||
3745 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
3752 l2arc_write_size(l2arc_dev_t *dev)
3756 size = dev->l2ad_write;
3758 if (arc_warm == B_FALSE)
3759 size += dev->l2ad_boost;
3766 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
3768 clock_t interval, next;
3771 * If the ARC lists are busy, increase our write rate; if the
3772 * lists are stale, idle back. This is achieved by checking
3773 * how much we previously wrote - if it was more than half of
3774 * what we wanted, schedule the next write much sooner.
3776 if (l2arc_feed_again && wrote > (wanted / 2))
3777 interval = (hz * l2arc_feed_min_ms) / 1000;
3779 interval = hz * l2arc_feed_secs;
3781 next = MAX(lbolt, MIN(lbolt + interval, began + interval));
3787 l2arc_hdr_stat_add(void)
3789 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
3790 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
3794 l2arc_hdr_stat_remove(void)
3796 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
3797 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
3801 * Cycle through L2ARC devices. This is how L2ARC load balances.
3802 * If a device is returned, this also returns holding the spa config lock.
3804 static l2arc_dev_t *
3805 l2arc_dev_get_next(void)
3807 l2arc_dev_t *first, *next = NULL;
3810 * Lock out the removal of spas (spa_namespace_lock), then removal
3811 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
3812 * both locks will be dropped and a spa config lock held instead.
3814 mutex_enter(&spa_namespace_lock);
3815 mutex_enter(&l2arc_dev_mtx);
3817 /* if there are no vdevs, there is nothing to do */
3818 if (l2arc_ndev == 0)
3822 next = l2arc_dev_last;
3824 /* loop around the list looking for a non-faulted vdev */
3826 next = list_head(l2arc_dev_list);
3828 next = list_next(l2arc_dev_list, next);
3830 next = list_head(l2arc_dev_list);
3833 /* if we have come back to the start, bail out */
3836 else if (next == first)
3839 } while (vdev_is_dead(next->l2ad_vdev));
3841 /* if we were unable to find any usable vdevs, return NULL */
3842 if (vdev_is_dead(next->l2ad_vdev))
3845 l2arc_dev_last = next;
3848 mutex_exit(&l2arc_dev_mtx);
3851 * Grab the config lock to prevent the 'next' device from being
3852 * removed while we are writing to it.
3855 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
3856 mutex_exit(&spa_namespace_lock);
3862 * Free buffers that were tagged for destruction.
3865 l2arc_do_free_on_write()
3868 l2arc_data_free_t *df, *df_prev;
3870 mutex_enter(&l2arc_free_on_write_mtx);
3871 buflist = l2arc_free_on_write;
3873 for (df = list_tail(buflist); df; df = df_prev) {
3874 df_prev = list_prev(buflist, df);
3875 ASSERT(df->l2df_data != NULL);
3876 ASSERT(df->l2df_func != NULL);
3877 df->l2df_func(df->l2df_data, df->l2df_size);
3878 list_remove(buflist, df);
3879 kmem_free(df, sizeof (l2arc_data_free_t));
3882 mutex_exit(&l2arc_free_on_write_mtx);
3886 * A write to a cache device has completed. Update all headers to allow
3887 * reads from these buffers to begin.
3890 l2arc_write_done(zio_t *zio)
3892 l2arc_write_callback_t *cb;
3895 arc_buf_hdr_t *head, *ab, *ab_prev;
3896 l2arc_buf_hdr_t *abl2;
3897 kmutex_t *hash_lock;
3899 cb = zio->io_private;
3901 dev = cb->l2wcb_dev;
3902 ASSERT(dev != NULL);
3903 head = cb->l2wcb_head;
3904 ASSERT(head != NULL);
3905 buflist = dev->l2ad_buflist;
3906 ASSERT(buflist != NULL);
3907 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
3908 l2arc_write_callback_t *, cb);
3910 if (zio->io_error != 0)
3911 ARCSTAT_BUMP(arcstat_l2_writes_error);
3913 mutex_enter(&l2arc_buflist_mtx);
3916 * All writes completed, or an error was hit.
3918 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
3919 ab_prev = list_prev(buflist, ab);
3921 hash_lock = HDR_LOCK(ab);
3922 if (!mutex_tryenter(hash_lock)) {
3924 * This buffer misses out. It may be in a stage
3925 * of eviction. Its ARC_L2_WRITING flag will be
3926 * left set, denying reads to this buffer.
3928 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
3932 if (zio->io_error != 0) {
3934 * Error - drop L2ARC entry.
3936 list_remove(buflist, ab);
3939 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
3940 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
3944 * Allow ARC to begin reads to this L2ARC entry.
3946 ab->b_flags &= ~ARC_L2_WRITING;
3948 mutex_exit(hash_lock);
3951 atomic_inc_64(&l2arc_writes_done);
3952 list_remove(buflist, head);
3953 kmem_cache_free(hdr_cache, head);
3954 mutex_exit(&l2arc_buflist_mtx);
3956 l2arc_do_free_on_write();
3958 kmem_free(cb, sizeof (l2arc_write_callback_t));
3962 * A read to a cache device completed. Validate buffer contents before
3963 * handing over to the regular ARC routines.
3966 l2arc_read_done(zio_t *zio)
3968 l2arc_read_callback_t *cb;
3971 kmutex_t *hash_lock;
3974 ASSERT(zio->io_vd != NULL);
3975 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
3977 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
3979 cb = zio->io_private;
3981 buf = cb->l2rcb_buf;
3982 ASSERT(buf != NULL);
3984 ASSERT(hdr != NULL);
3986 hash_lock = HDR_LOCK(hdr);
3987 mutex_enter(hash_lock);
3990 * Check this survived the L2ARC journey.
3992 equal = arc_cksum_equal(buf);
3993 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
3994 mutex_exit(hash_lock);
3995 zio->io_private = buf;
3996 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
3997 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4000 mutex_exit(hash_lock);
4002 * Buffer didn't survive caching. Increment stats and
4003 * reissue to the original storage device.
4005 if (zio->io_error != 0) {
4006 ARCSTAT_BUMP(arcstat_l2_io_error);
4008 zio->io_error = EIO;
4011 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4014 * If there's no waiter, issue an async i/o to the primary
4015 * storage now. If there *is* a waiter, the caller must
4016 * issue the i/o in a context where it's OK to block.
4018 if (zio->io_waiter == NULL) {
4019 zio_t *pio = zio_unique_parent(zio);
4021 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4023 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4024 buf->b_data, zio->io_size, arc_read_done, buf,
4025 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4029 kmem_free(cb, sizeof (l2arc_read_callback_t));
4033 * This is the list priority from which the L2ARC will search for pages to
4034 * cache. This is used within loops (0..3) to cycle through lists in the
4035 * desired order. This order can have a significant effect on cache
4038 * Currently the metadata lists are hit first, MFU then MRU, followed by
4039 * the data lists. This function returns a locked list, and also returns
4043 l2arc_list_locked(int list_num, kmutex_t **lock)
4047 ASSERT(list_num >= 0 && list_num <= 3);
4051 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4052 *lock = &arc_mfu->arcs_mtx;
4055 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4056 *lock = &arc_mru->arcs_mtx;
4059 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4060 *lock = &arc_mfu->arcs_mtx;
4063 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4064 *lock = &arc_mru->arcs_mtx;
4068 ASSERT(!(MUTEX_HELD(*lock)));
4074 * Evict buffers from the device write hand to the distance specified in
4075 * bytes. This distance may span populated buffers, it may span nothing.
4076 * This is clearing a region on the L2ARC device ready for writing.
4077 * If the 'all' boolean is set, every buffer is evicted.
4080 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4083 l2arc_buf_hdr_t *abl2;
4084 arc_buf_hdr_t *ab, *ab_prev;
4085 kmutex_t *hash_lock;
4088 buflist = dev->l2ad_buflist;
4090 if (buflist == NULL)
4093 if (!all && dev->l2ad_first) {
4095 * This is the first sweep through the device. There is
4101 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4103 * When nearing the end of the device, evict to the end
4104 * before the device write hand jumps to the start.
4106 taddr = dev->l2ad_end;
4108 taddr = dev->l2ad_hand + distance;
4110 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4111 uint64_t, taddr, boolean_t, all);
4114 mutex_enter(&l2arc_buflist_mtx);
4115 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4116 ab_prev = list_prev(buflist, ab);
4118 hash_lock = HDR_LOCK(ab);
4119 if (!mutex_tryenter(hash_lock)) {
4121 * Missed the hash lock. Retry.
4123 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4124 mutex_exit(&l2arc_buflist_mtx);
4125 mutex_enter(hash_lock);
4126 mutex_exit(hash_lock);
4130 if (HDR_L2_WRITE_HEAD(ab)) {
4132 * We hit a write head node. Leave it for
4133 * l2arc_write_done().
4135 list_remove(buflist, ab);
4136 mutex_exit(hash_lock);
4140 if (!all && ab->b_l2hdr != NULL &&
4141 (ab->b_l2hdr->b_daddr > taddr ||
4142 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4144 * We've evicted to the target address,
4145 * or the end of the device.
4147 mutex_exit(hash_lock);
4151 if (HDR_FREE_IN_PROGRESS(ab)) {
4153 * Already on the path to destruction.
4155 mutex_exit(hash_lock);
4159 if (ab->b_state == arc_l2c_only) {
4160 ASSERT(!HDR_L2_READING(ab));
4162 * This doesn't exist in the ARC. Destroy.
4163 * arc_hdr_destroy() will call list_remove()
4164 * and decrement arcstat_l2_size.
4166 arc_change_state(arc_anon, ab, hash_lock);
4167 arc_hdr_destroy(ab);
4170 * Invalidate issued or about to be issued
4171 * reads, since we may be about to write
4172 * over this location.
4174 if (HDR_L2_READING(ab)) {
4175 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4176 ab->b_flags |= ARC_L2_EVICTED;
4180 * Tell ARC this no longer exists in L2ARC.
4182 if (ab->b_l2hdr != NULL) {
4185 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4186 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4188 list_remove(buflist, ab);
4191 * This may have been leftover after a
4194 ab->b_flags &= ~ARC_L2_WRITING;
4196 mutex_exit(hash_lock);
4198 mutex_exit(&l2arc_buflist_mtx);
4200 spa_l2cache_space_update(dev->l2ad_vdev, 0, -(taddr - dev->l2ad_evict));
4201 dev->l2ad_evict = taddr;
4205 * Find and write ARC buffers to the L2ARC device.
4207 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4208 * for reading until they have completed writing.
4211 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4213 arc_buf_hdr_t *ab, *ab_prev, *head;
4214 l2arc_buf_hdr_t *hdrl2;
4216 uint64_t passed_sz, write_sz, buf_sz, headroom;
4218 kmutex_t *hash_lock, *list_lock;
4219 boolean_t have_lock, full;
4220 l2arc_write_callback_t *cb;
4222 uint64_t guid = spa_guid(spa);
4224 ASSERT(dev->l2ad_vdev != NULL);
4229 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4230 head->b_flags |= ARC_L2_WRITE_HEAD;
4233 * Copy buffers for L2ARC writing.
4235 mutex_enter(&l2arc_buflist_mtx);
4236 for (int try = 0; try <= 3; try++) {
4237 list = l2arc_list_locked(try, &list_lock);
4241 * L2ARC fast warmup.
4243 * Until the ARC is warm and starts to evict, read from the
4244 * head of the ARC lists rather than the tail.
4246 headroom = target_sz * l2arc_headroom;
4247 if (arc_warm == B_FALSE)
4248 ab = list_head(list);
4250 ab = list_tail(list);
4252 for (; ab; ab = ab_prev) {
4253 if (arc_warm == B_FALSE)
4254 ab_prev = list_next(list, ab);
4256 ab_prev = list_prev(list, ab);
4258 hash_lock = HDR_LOCK(ab);
4259 have_lock = MUTEX_HELD(hash_lock);
4260 if (!have_lock && !mutex_tryenter(hash_lock)) {
4262 * Skip this buffer rather than waiting.
4267 passed_sz += ab->b_size;
4268 if (passed_sz > headroom) {
4272 mutex_exit(hash_lock);
4276 if (!l2arc_write_eligible(guid, ab)) {
4277 mutex_exit(hash_lock);
4281 if ((write_sz + ab->b_size) > target_sz) {
4283 mutex_exit(hash_lock);
4289 * Insert a dummy header on the buflist so
4290 * l2arc_write_done() can find where the
4291 * write buffers begin without searching.
4293 list_insert_head(dev->l2ad_buflist, head);
4296 sizeof (l2arc_write_callback_t), KM_SLEEP);
4297 cb->l2wcb_dev = dev;
4298 cb->l2wcb_head = head;
4299 pio = zio_root(spa, l2arc_write_done, cb,
4304 * Create and add a new L2ARC header.
4306 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4308 hdrl2->b_daddr = dev->l2ad_hand;
4310 ab->b_flags |= ARC_L2_WRITING;
4311 ab->b_l2hdr = hdrl2;
4312 list_insert_head(dev->l2ad_buflist, ab);
4313 buf_data = ab->b_buf->b_data;
4314 buf_sz = ab->b_size;
4317 * Compute and store the buffer cksum before
4318 * writing. On debug the cksum is verified first.
4320 arc_cksum_verify(ab->b_buf);
4321 arc_cksum_compute(ab->b_buf, B_TRUE);
4323 mutex_exit(hash_lock);
4325 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4326 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4327 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4328 ZIO_FLAG_CANFAIL, B_FALSE);
4330 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4332 (void) zio_nowait(wzio);
4335 * Keep the clock hand suitably device-aligned.
4337 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4340 dev->l2ad_hand += buf_sz;
4343 mutex_exit(list_lock);
4348 mutex_exit(&l2arc_buflist_mtx);
4351 ASSERT3U(write_sz, ==, 0);
4352 kmem_cache_free(hdr_cache, head);
4356 ASSERT3U(write_sz, <=, target_sz);
4357 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4358 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4359 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4360 spa_l2cache_space_update(dev->l2ad_vdev, 0, write_sz);
4363 * Bump device hand to the device start if it is approaching the end.
4364 * l2arc_evict() will already have evicted ahead for this case.
4366 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4367 spa_l2cache_space_update(dev->l2ad_vdev, 0,
4368 dev->l2ad_end - dev->l2ad_hand);
4369 dev->l2ad_hand = dev->l2ad_start;
4370 dev->l2ad_evict = dev->l2ad_start;
4371 dev->l2ad_first = B_FALSE;
4374 dev->l2ad_writing = B_TRUE;
4375 (void) zio_wait(pio);
4376 dev->l2ad_writing = B_FALSE;
4382 * This thread feeds the L2ARC at regular intervals. This is the beating
4383 * heart of the L2ARC.
4386 l2arc_feed_thread(void)
4391 uint64_t size, wrote;
4392 clock_t begin, next = lbolt;
4394 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4396 mutex_enter(&l2arc_feed_thr_lock);
4398 while (l2arc_thread_exit == 0) {
4399 CALLB_CPR_SAFE_BEGIN(&cpr);
4400 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4402 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4406 * Quick check for L2ARC devices.
4408 mutex_enter(&l2arc_dev_mtx);
4409 if (l2arc_ndev == 0) {
4410 mutex_exit(&l2arc_dev_mtx);
4413 mutex_exit(&l2arc_dev_mtx);
4417 * This selects the next l2arc device to write to, and in
4418 * doing so the next spa to feed from: dev->l2ad_spa. This
4419 * will return NULL if there are now no l2arc devices or if
4420 * they are all faulted.
4422 * If a device is returned, its spa's config lock is also
4423 * held to prevent device removal. l2arc_dev_get_next()
4424 * will grab and release l2arc_dev_mtx.
4426 if ((dev = l2arc_dev_get_next()) == NULL)
4429 spa = dev->l2ad_spa;
4430 ASSERT(spa != NULL);
4433 * Avoid contributing to memory pressure.
4435 if (arc_reclaim_needed()) {
4436 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4437 spa_config_exit(spa, SCL_L2ARC, dev);
4441 ARCSTAT_BUMP(arcstat_l2_feeds);
4443 size = l2arc_write_size(dev);
4446 * Evict L2ARC buffers that will be overwritten.
4448 l2arc_evict(dev, size, B_FALSE);
4451 * Write ARC buffers.
4453 wrote = l2arc_write_buffers(spa, dev, size);
4456 * Calculate interval between writes.
4458 next = l2arc_write_interval(begin, size, wrote);
4459 spa_config_exit(spa, SCL_L2ARC, dev);
4462 l2arc_thread_exit = 0;
4463 cv_broadcast(&l2arc_feed_thr_cv);
4464 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4469 l2arc_vdev_present(vdev_t *vd)
4473 mutex_enter(&l2arc_dev_mtx);
4474 for (dev = list_head(l2arc_dev_list); dev != NULL;
4475 dev = list_next(l2arc_dev_list, dev)) {
4476 if (dev->l2ad_vdev == vd)
4479 mutex_exit(&l2arc_dev_mtx);
4481 return (dev != NULL);
4485 * Add a vdev for use by the L2ARC. By this point the spa has already
4486 * validated the vdev and opened it.
4489 l2arc_add_vdev(spa_t *spa, vdev_t *vd, uint64_t start, uint64_t end)
4491 l2arc_dev_t *adddev;
4493 ASSERT(!l2arc_vdev_present(vd));
4496 * Create a new l2arc device entry.
4498 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4499 adddev->l2ad_spa = spa;
4500 adddev->l2ad_vdev = vd;
4501 adddev->l2ad_write = l2arc_write_max;
4502 adddev->l2ad_boost = l2arc_write_boost;
4503 adddev->l2ad_start = start;
4504 adddev->l2ad_end = end;
4505 adddev->l2ad_hand = adddev->l2ad_start;
4506 adddev->l2ad_evict = adddev->l2ad_start;
4507 adddev->l2ad_first = B_TRUE;
4508 adddev->l2ad_writing = B_FALSE;
4509 ASSERT3U(adddev->l2ad_write, >, 0);
4512 * This is a list of all ARC buffers that are still valid on the
4515 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4516 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4517 offsetof(arc_buf_hdr_t, b_l2node));
4519 spa_l2cache_space_update(vd, adddev->l2ad_end - adddev->l2ad_hand, 0);
4522 * Add device to global list
4524 mutex_enter(&l2arc_dev_mtx);
4525 list_insert_head(l2arc_dev_list, adddev);
4526 atomic_inc_64(&l2arc_ndev);
4527 mutex_exit(&l2arc_dev_mtx);
4531 * Remove a vdev from the L2ARC.
4534 l2arc_remove_vdev(vdev_t *vd)
4536 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4539 * Find the device by vdev
4541 mutex_enter(&l2arc_dev_mtx);
4542 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4543 nextdev = list_next(l2arc_dev_list, dev);
4544 if (vd == dev->l2ad_vdev) {
4549 ASSERT(remdev != NULL);
4552 * Remove device from global list
4554 list_remove(l2arc_dev_list, remdev);
4555 l2arc_dev_last = NULL; /* may have been invalidated */
4556 atomic_dec_64(&l2arc_ndev);
4557 mutex_exit(&l2arc_dev_mtx);
4560 * Clear all buflists and ARC references. L2ARC device flush.
4562 l2arc_evict(remdev, 0, B_TRUE);
4563 list_destroy(remdev->l2ad_buflist);
4564 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
4565 kmem_free(remdev, sizeof (l2arc_dev_t));
4571 l2arc_thread_exit = 0;
4573 l2arc_writes_sent = 0;
4574 l2arc_writes_done = 0;
4576 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4577 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
4578 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
4579 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
4580 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
4582 l2arc_dev_list = &L2ARC_dev_list;
4583 l2arc_free_on_write = &L2ARC_free_on_write;
4584 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
4585 offsetof(l2arc_dev_t, l2ad_node));
4586 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
4587 offsetof(l2arc_data_free_t, l2df_list_node));
4594 * This is called from dmu_fini(), which is called from spa_fini();
4595 * Because of this, we can assume that all l2arc devices have
4596 * already been removed when the pools themselves were removed.
4599 l2arc_do_free_on_write();
4601 mutex_destroy(&l2arc_feed_thr_lock);
4602 cv_destroy(&l2arc_feed_thr_cv);
4603 mutex_destroy(&l2arc_dev_mtx);
4604 mutex_destroy(&l2arc_buflist_mtx);
4605 mutex_destroy(&l2arc_free_on_write_mtx);
4607 list_destroy(l2arc_dev_list);
4608 list_destroy(l2arc_free_on_write);
4614 if (!(spa_mode_global & FWRITE))
4617 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
4618 TS_RUN, minclsyspri);
4624 if (!(spa_mode_global & FWRITE))
4627 mutex_enter(&l2arc_feed_thr_lock);
4628 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
4629 l2arc_thread_exit = 1;
4630 while (l2arc_thread_exit != 0)
4631 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
4632 mutex_exit(&l2arc_feed_thr_lock);