4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
26 * DVA-based Adjustable Replacement Cache
28 * While much of the theory of operation used here is
29 * based on the self-tuning, low overhead replacement cache
30 * presented by Megiddo and Modha at FAST 2003, there are some
31 * significant differences:
33 * 1. The Megiddo and Modha model assumes any page is evictable.
34 * Pages in its cache cannot be "locked" into memory. This makes
35 * the eviction algorithm simple: evict the last page in the list.
36 * This also make the performance characteristics easy to reason
37 * about. Our cache is not so simple. At any given moment, some
38 * subset of the blocks in the cache are un-evictable because we
39 * have handed out a reference to them. Blocks are only evictable
40 * when there are no external references active. This makes
41 * eviction far more problematic: we choose to evict the evictable
42 * blocks that are the "lowest" in the list.
44 * There are times when it is not possible to evict the requested
45 * space. In these circumstances we are unable to adjust the cache
46 * size. To prevent the cache growing unbounded at these times we
47 * implement a "cache throttle" that slows the flow of new data
48 * into the cache until we can make space available.
50 * 2. The Megiddo and Modha model assumes a fixed cache size.
51 * Pages are evicted when the cache is full and there is a cache
52 * miss. Our model has a variable sized cache. It grows with
53 * high use, but also tries to react to memory pressure from the
54 * operating system: decreasing its size when system memory is
57 * 3. The Megiddo and Modha model assumes a fixed page size. All
58 * elements of the cache are therefor exactly the same size. So
59 * when adjusting the cache size following a cache miss, its simply
60 * a matter of choosing a single page to evict. In our model, we
61 * have variable sized cache blocks (rangeing from 512 bytes to
62 * 128K bytes). We therefor choose a set of blocks to evict to make
63 * space for a cache miss that approximates as closely as possible
64 * the space used by the new block.
66 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
67 * by N. Megiddo & D. Modha, FAST 2003
73 * A new reference to a cache buffer can be obtained in two
74 * ways: 1) via a hash table lookup using the DVA as a key,
75 * or 2) via one of the ARC lists. The arc_read() interface
76 * uses method 1, while the internal arc algorithms for
77 * adjusting the cache use method 2. We therefor provide two
78 * types of locks: 1) the hash table lock array, and 2) the
81 * Buffers do not have their own mutexs, rather they rely on the
82 * hash table mutexs for the bulk of their protection (i.e. most
83 * fields in the arc_buf_hdr_t are protected by these mutexs).
85 * buf_hash_find() returns the appropriate mutex (held) when it
86 * locates the requested buffer in the hash table. It returns
87 * NULL for the mutex if the buffer was not in the table.
89 * buf_hash_remove() expects the appropriate hash mutex to be
90 * already held before it is invoked.
92 * Each arc state also has a mutex which is used to protect the
93 * buffer list associated with the state. When attempting to
94 * obtain a hash table lock while holding an arc list lock you
95 * must use: mutex_tryenter() to avoid deadlock. Also note that
96 * the active state mutex must be held before the ghost state mutex.
98 * Arc buffers may have an associated eviction callback function.
99 * This function will be invoked prior to removing the buffer (e.g.
100 * in arc_do_user_evicts()). Note however that the data associated
101 * with the buffer may be evicted prior to the callback. The callback
102 * must be made with *no locks held* (to prevent deadlock). Additionally,
103 * the users of callbacks must ensure that their private data is
104 * protected from simultaneous callbacks from arc_buf_evict()
105 * and arc_do_user_evicts().
107 * Note that the majority of the performance stats are manipulated
108 * with atomic operations.
110 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
112 * - L2ARC buflist creation
113 * - L2ARC buflist eviction
114 * - L2ARC write completion, which walks L2ARC buflists
115 * - ARC header destruction, as it removes from L2ARC buflists
116 * - ARC header release, as it removes from L2ARC buflists
121 #include <sys/zfs_context.h>
123 #include <sys/refcount.h>
124 #include <sys/vdev.h>
125 #include <sys/vdev_impl.h>
127 #include <sys/vmsystm.h>
129 #include <sys/fs/swapnode.h>
130 #include <sys/dnlc.h>
132 #include <sys/callb.h>
133 #include <sys/kstat.h>
134 #include <zfs_fletcher.h>
136 static kmutex_t arc_reclaim_thr_lock;
137 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
138 static uint8_t arc_thread_exit;
140 extern int zfs_write_limit_shift;
141 extern uint64_t zfs_write_limit_max;
142 extern kmutex_t zfs_write_limit_lock;
144 #define ARC_REDUCE_DNLC_PERCENT 3
145 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
147 typedef enum arc_reclaim_strategy {
148 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
149 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
150 } arc_reclaim_strategy_t;
152 /* number of seconds before growing cache again */
153 static int arc_grow_retry = 60;
155 /* shift of arc_c for calculating both min and max arc_p */
156 static int arc_p_min_shift = 4;
158 /* log2(fraction of arc to reclaim) */
159 static int arc_shrink_shift = 5;
162 * minimum lifespan of a prefetch block in clock ticks
163 * (initialized in arc_init())
165 static int arc_min_prefetch_lifespan;
170 * The arc has filled available memory and has now warmed up.
172 static boolean_t arc_warm;
175 * These tunables are for performance analysis.
177 unsigned long zfs_arc_max = 0;
178 unsigned long zfs_arc_min = 0;
179 unsigned long zfs_arc_meta_limit = 0;
180 int zfs_arc_grow_retry = 0;
181 int zfs_arc_shrink_shift = 0;
182 int zfs_arc_p_min_shift = 0;
183 int zfs_arc_reduce_dnlc_percent = 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_evict_l2_cached;
252 kstat_named_t arcstat_evict_l2_eligible;
253 kstat_named_t arcstat_evict_l2_ineligible;
254 kstat_named_t arcstat_hash_elements;
255 kstat_named_t arcstat_hash_elements_max;
256 kstat_named_t arcstat_hash_collisions;
257 kstat_named_t arcstat_hash_chains;
258 kstat_named_t arcstat_hash_chain_max;
259 kstat_named_t arcstat_p;
260 kstat_named_t arcstat_c;
261 kstat_named_t arcstat_c_min;
262 kstat_named_t arcstat_c_max;
263 kstat_named_t arcstat_size;
264 kstat_named_t arcstat_hdr_size;
265 kstat_named_t arcstat_data_size;
266 kstat_named_t arcstat_other_size;
267 kstat_named_t arcstat_l2_hits;
268 kstat_named_t arcstat_l2_misses;
269 kstat_named_t arcstat_l2_feeds;
270 kstat_named_t arcstat_l2_rw_clash;
271 kstat_named_t arcstat_l2_read_bytes;
272 kstat_named_t arcstat_l2_write_bytes;
273 kstat_named_t arcstat_l2_writes_sent;
274 kstat_named_t arcstat_l2_writes_done;
275 kstat_named_t arcstat_l2_writes_error;
276 kstat_named_t arcstat_l2_writes_hdr_miss;
277 kstat_named_t arcstat_l2_evict_lock_retry;
278 kstat_named_t arcstat_l2_evict_reading;
279 kstat_named_t arcstat_l2_free_on_write;
280 kstat_named_t arcstat_l2_abort_lowmem;
281 kstat_named_t arcstat_l2_cksum_bad;
282 kstat_named_t arcstat_l2_io_error;
283 kstat_named_t arcstat_l2_size;
284 kstat_named_t arcstat_l2_hdr_size;
285 kstat_named_t arcstat_memory_throttle_count;
286 kstat_named_t arcstat_memory_direct_count;
287 kstat_named_t arcstat_memory_indirect_count;
288 kstat_named_t arcstat_no_grow;
289 kstat_named_t arcstat_tempreserve;
290 kstat_named_t arcstat_loaned_bytes;
291 kstat_named_t arcstat_meta_used;
292 kstat_named_t arcstat_meta_limit;
293 kstat_named_t arcstat_meta_max;
296 static arc_stats_t arc_stats = {
297 { "hits", KSTAT_DATA_UINT64 },
298 { "misses", KSTAT_DATA_UINT64 },
299 { "demand_data_hits", KSTAT_DATA_UINT64 },
300 { "demand_data_misses", KSTAT_DATA_UINT64 },
301 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
302 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
303 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
304 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
305 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
306 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
307 { "mru_hits", KSTAT_DATA_UINT64 },
308 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
309 { "mfu_hits", KSTAT_DATA_UINT64 },
310 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
311 { "deleted", KSTAT_DATA_UINT64 },
312 { "recycle_miss", KSTAT_DATA_UINT64 },
313 { "mutex_miss", KSTAT_DATA_UINT64 },
314 { "evict_skip", KSTAT_DATA_UINT64 },
315 { "evict_l2_cached", KSTAT_DATA_UINT64 },
316 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
317 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
318 { "hash_elements", KSTAT_DATA_UINT64 },
319 { "hash_elements_max", KSTAT_DATA_UINT64 },
320 { "hash_collisions", KSTAT_DATA_UINT64 },
321 { "hash_chains", KSTAT_DATA_UINT64 },
322 { "hash_chain_max", KSTAT_DATA_UINT64 },
323 { "p", KSTAT_DATA_UINT64 },
324 { "c", KSTAT_DATA_UINT64 },
325 { "c_min", KSTAT_DATA_UINT64 },
326 { "c_max", KSTAT_DATA_UINT64 },
327 { "size", KSTAT_DATA_UINT64 },
328 { "hdr_size", KSTAT_DATA_UINT64 },
329 { "data_size", KSTAT_DATA_UINT64 },
330 { "other_size", KSTAT_DATA_UINT64 },
331 { "l2_hits", KSTAT_DATA_UINT64 },
332 { "l2_misses", KSTAT_DATA_UINT64 },
333 { "l2_feeds", KSTAT_DATA_UINT64 },
334 { "l2_rw_clash", KSTAT_DATA_UINT64 },
335 { "l2_read_bytes", KSTAT_DATA_UINT64 },
336 { "l2_write_bytes", KSTAT_DATA_UINT64 },
337 { "l2_writes_sent", KSTAT_DATA_UINT64 },
338 { "l2_writes_done", KSTAT_DATA_UINT64 },
339 { "l2_writes_error", KSTAT_DATA_UINT64 },
340 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
341 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
342 { "l2_evict_reading", KSTAT_DATA_UINT64 },
343 { "l2_free_on_write", KSTAT_DATA_UINT64 },
344 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
345 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
346 { "l2_io_error", KSTAT_DATA_UINT64 },
347 { "l2_size", KSTAT_DATA_UINT64 },
348 { "l2_hdr_size", KSTAT_DATA_UINT64 },
349 { "memory_throttle_count", KSTAT_DATA_UINT64 },
350 { "memory_direct_count", KSTAT_DATA_UINT64 },
351 { "memory_indirect_count", KSTAT_DATA_UINT64 },
352 { "arc_no_grow", KSTAT_DATA_UINT64 },
353 { "arc_tempreserve", KSTAT_DATA_UINT64 },
354 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
355 { "arc_meta_used", KSTAT_DATA_UINT64 },
356 { "arc_meta_limit", KSTAT_DATA_UINT64 },
357 { "arc_meta_max", KSTAT_DATA_UINT64 },
360 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
362 #define ARCSTAT_INCR(stat, val) \
363 atomic_add_64(&arc_stats.stat.value.ui64, (val));
365 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
366 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
368 #define ARCSTAT_MAX(stat, val) { \
370 while ((val) > (m = arc_stats.stat.value.ui64) && \
371 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
375 #define ARCSTAT_MAXSTAT(stat) \
376 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
379 * We define a macro to allow ARC hits/misses to be easily broken down by
380 * two separate conditions, giving a total of four different subtypes for
381 * each of hits and misses (so eight statistics total).
383 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
386 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
388 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
392 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
394 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
399 static arc_state_t *arc_anon;
400 static arc_state_t *arc_mru;
401 static arc_state_t *arc_mru_ghost;
402 static arc_state_t *arc_mfu;
403 static arc_state_t *arc_mfu_ghost;
404 static arc_state_t *arc_l2c_only;
407 * There are several ARC variables that are critical to export as kstats --
408 * but we don't want to have to grovel around in the kstat whenever we wish to
409 * manipulate them. For these variables, we therefore define them to be in
410 * terms of the statistic variable. This assures that we are not introducing
411 * the possibility of inconsistency by having shadow copies of the variables,
412 * while still allowing the code to be readable.
414 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
415 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
416 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
417 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
418 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
419 #define arc_no_grow ARCSTAT(arcstat_no_grow)
420 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
421 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
422 #define arc_meta_used ARCSTAT(arcstat_meta_used)
423 #define arc_meta_limit ARCSTAT(arcstat_meta_limit)
424 #define arc_meta_max ARCSTAT(arcstat_meta_max)
426 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
428 typedef struct arc_callback arc_callback_t;
430 struct arc_callback {
432 arc_done_func_t *acb_done;
434 zio_t *acb_zio_dummy;
435 arc_callback_t *acb_next;
438 typedef struct arc_write_callback arc_write_callback_t;
440 struct arc_write_callback {
442 arc_done_func_t *awcb_ready;
443 arc_done_func_t *awcb_done;
448 /* protected by hash lock */
453 kmutex_t b_freeze_lock;
454 zio_cksum_t *b_freeze_cksum;
457 arc_buf_hdr_t *b_hash_next;
462 arc_callback_t *b_acb;
466 arc_buf_contents_t b_type;
470 /* protected by arc state mutex */
471 arc_state_t *b_state;
472 list_node_t b_arc_node;
474 /* updated atomically */
475 clock_t b_arc_access;
477 /* self protecting */
480 l2arc_buf_hdr_t *b_l2hdr;
481 list_node_t b_l2node;
484 static arc_buf_t *arc_eviction_list;
485 static kmutex_t arc_eviction_mtx;
486 static arc_buf_hdr_t arc_eviction_hdr;
487 static void arc_get_data_buf(arc_buf_t *buf);
488 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
489 static int arc_evict_needed(arc_buf_contents_t type);
490 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
492 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
494 #define GHOST_STATE(state) \
495 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
496 (state) == arc_l2c_only)
499 * Private ARC flags. These flags are private ARC only flags that will show up
500 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
501 * be passed in as arc_flags in things like arc_read. However, these flags
502 * should never be passed and should only be set by ARC code. When adding new
503 * public flags, make sure not to smash the private ones.
506 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
507 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
508 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
509 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
510 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
511 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
512 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
513 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
514 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
515 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
517 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
518 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
519 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
520 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
521 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
522 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
523 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
524 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
525 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
526 (hdr)->b_l2hdr != NULL)
527 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
528 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
529 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
535 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
536 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
539 * Hash table routines
542 #define HT_LOCK_ALIGN 64
543 #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN)))
548 unsigned char pad[HT_LOCK_PAD];
552 #define BUF_LOCKS 256
553 typedef struct buf_hash_table {
555 arc_buf_hdr_t **ht_table;
556 struct ht_lock ht_locks[BUF_LOCKS];
559 static buf_hash_table_t buf_hash_table;
561 #define BUF_HASH_INDEX(spa, dva, birth) \
562 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
563 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
564 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
565 #define HDR_LOCK(hdr) \
566 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
568 uint64_t zfs_crc64_table[256];
574 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
575 #define L2ARC_HEADROOM 2 /* num of writes */
576 #define L2ARC_FEED_SECS 1 /* caching interval secs */
577 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
579 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
580 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
583 * L2ARC Performance Tunables
585 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
586 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
587 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
588 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
589 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
590 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
591 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
592 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
597 typedef struct l2arc_dev {
598 vdev_t *l2ad_vdev; /* vdev */
599 spa_t *l2ad_spa; /* spa */
600 uint64_t l2ad_hand; /* next write location */
601 uint64_t l2ad_write; /* desired write size, bytes */
602 uint64_t l2ad_boost; /* warmup write boost, bytes */
603 uint64_t l2ad_start; /* first addr on device */
604 uint64_t l2ad_end; /* last addr on device */
605 uint64_t l2ad_evict; /* last addr eviction reached */
606 boolean_t l2ad_first; /* first sweep through */
607 boolean_t l2ad_writing; /* currently writing */
608 list_t *l2ad_buflist; /* buffer list */
609 list_node_t l2ad_node; /* device list node */
612 static list_t L2ARC_dev_list; /* device list */
613 static list_t *l2arc_dev_list; /* device list pointer */
614 static kmutex_t l2arc_dev_mtx; /* device list mutex */
615 static l2arc_dev_t *l2arc_dev_last; /* last device used */
616 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
617 static list_t L2ARC_free_on_write; /* free after write buf list */
618 static list_t *l2arc_free_on_write; /* free after write list ptr */
619 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
620 static uint64_t l2arc_ndev; /* number of devices */
622 typedef struct l2arc_read_callback {
623 arc_buf_t *l2rcb_buf; /* read buffer */
624 spa_t *l2rcb_spa; /* spa */
625 blkptr_t l2rcb_bp; /* original blkptr */
626 zbookmark_t l2rcb_zb; /* original bookmark */
627 int l2rcb_flags; /* original flags */
628 } l2arc_read_callback_t;
630 typedef struct l2arc_write_callback {
631 l2arc_dev_t *l2wcb_dev; /* device info */
632 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
633 } l2arc_write_callback_t;
635 struct l2arc_buf_hdr {
636 /* protected by arc_buf_hdr mutex */
637 l2arc_dev_t *b_dev; /* L2ARC device */
638 uint64_t b_daddr; /* disk address, offset byte */
641 typedef struct l2arc_data_free {
642 /* protected by l2arc_free_on_write_mtx */
645 void (*l2df_func)(void *, size_t);
646 list_node_t l2df_list_node;
649 static kmutex_t l2arc_feed_thr_lock;
650 static kcondvar_t l2arc_feed_thr_cv;
651 static uint8_t l2arc_thread_exit;
653 static void l2arc_read_done(zio_t *zio);
654 static void l2arc_hdr_stat_add(void);
655 static void l2arc_hdr_stat_remove(void);
658 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
660 uint8_t *vdva = (uint8_t *)dva;
661 uint64_t crc = -1ULL;
664 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
666 for (i = 0; i < sizeof (dva_t); i++)
667 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
669 crc ^= (spa>>8) ^ birth;
674 #define BUF_EMPTY(buf) \
675 ((buf)->b_dva.dva_word[0] == 0 && \
676 (buf)->b_dva.dva_word[1] == 0 && \
679 #define BUF_EQUAL(spa, dva, birth, buf) \
680 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
681 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
682 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
685 buf_discard_identity(arc_buf_hdr_t *hdr)
687 hdr->b_dva.dva_word[0] = 0;
688 hdr->b_dva.dva_word[1] = 0;
693 static arc_buf_hdr_t *
694 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
696 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
697 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
700 mutex_enter(hash_lock);
701 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
702 buf = buf->b_hash_next) {
703 if (BUF_EQUAL(spa, dva, birth, buf)) {
708 mutex_exit(hash_lock);
714 * Insert an entry into the hash table. If there is already an element
715 * equal to elem in the hash table, then the already existing element
716 * will be returned and the new element will not be inserted.
717 * Otherwise returns NULL.
719 static arc_buf_hdr_t *
720 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
722 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
723 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
727 ASSERT(!HDR_IN_HASH_TABLE(buf));
729 mutex_enter(hash_lock);
730 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
731 fbuf = fbuf->b_hash_next, i++) {
732 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
736 buf->b_hash_next = buf_hash_table.ht_table[idx];
737 buf_hash_table.ht_table[idx] = buf;
738 buf->b_flags |= ARC_IN_HASH_TABLE;
740 /* collect some hash table performance data */
742 ARCSTAT_BUMP(arcstat_hash_collisions);
744 ARCSTAT_BUMP(arcstat_hash_chains);
746 ARCSTAT_MAX(arcstat_hash_chain_max, i);
749 ARCSTAT_BUMP(arcstat_hash_elements);
750 ARCSTAT_MAXSTAT(arcstat_hash_elements);
756 buf_hash_remove(arc_buf_hdr_t *buf)
758 arc_buf_hdr_t *fbuf, **bufp;
759 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
761 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
762 ASSERT(HDR_IN_HASH_TABLE(buf));
764 bufp = &buf_hash_table.ht_table[idx];
765 while ((fbuf = *bufp) != buf) {
766 ASSERT(fbuf != NULL);
767 bufp = &fbuf->b_hash_next;
769 *bufp = buf->b_hash_next;
770 buf->b_hash_next = NULL;
771 buf->b_flags &= ~ARC_IN_HASH_TABLE;
773 /* collect some hash table performance data */
774 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
776 if (buf_hash_table.ht_table[idx] &&
777 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
778 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
782 * Global data structures and functions for the buf kmem cache.
784 static kmem_cache_t *hdr_cache;
785 static kmem_cache_t *buf_cache;
792 #if defined(_KERNEL) && defined(HAVE_SPL)
793 /* Large allocations which do not require contiguous pages
794 * should be using vmem_free() in the linux kernel */
795 vmem_free(buf_hash_table.ht_table,
796 (buf_hash_table.ht_mask + 1) * sizeof (void *));
798 kmem_free(buf_hash_table.ht_table,
799 (buf_hash_table.ht_mask + 1) * sizeof (void *));
801 for (i = 0; i < BUF_LOCKS; i++)
802 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
803 kmem_cache_destroy(hdr_cache);
804 kmem_cache_destroy(buf_cache);
808 * Constructor callback - called when the cache is empty
809 * and a new buf is requested.
813 hdr_cons(void *vbuf, void *unused, int kmflag)
815 arc_buf_hdr_t *buf = vbuf;
817 bzero(buf, sizeof (arc_buf_hdr_t));
818 refcount_create(&buf->b_refcnt);
819 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
820 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
821 list_link_init(&buf->b_arc_node);
822 list_link_init(&buf->b_l2node);
823 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
830 buf_cons(void *vbuf, void *unused, int kmflag)
832 arc_buf_t *buf = vbuf;
834 bzero(buf, sizeof (arc_buf_t));
835 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
836 rw_init(&buf->b_data_lock, NULL, RW_DEFAULT, NULL);
837 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
843 * Destructor callback - called when a cached buf is
844 * no longer required.
848 hdr_dest(void *vbuf, void *unused)
850 arc_buf_hdr_t *buf = vbuf;
852 ASSERT(BUF_EMPTY(buf));
853 refcount_destroy(&buf->b_refcnt);
854 cv_destroy(&buf->b_cv);
855 mutex_destroy(&buf->b_freeze_lock);
856 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
861 buf_dest(void *vbuf, void *unused)
863 arc_buf_t *buf = vbuf;
865 mutex_destroy(&buf->b_evict_lock);
866 rw_destroy(&buf->b_data_lock);
867 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
871 * Reclaim callback -- invoked when memory is low.
875 hdr_recl(void *unused)
877 dprintf("hdr_recl called\n");
879 * umem calls the reclaim func when we destroy the buf cache,
880 * which is after we do arc_fini().
883 cv_signal(&arc_reclaim_thr_cv);
890 uint64_t hsize = 1ULL << 12;
894 * The hash table is big enough to fill all of physical memory
895 * with an average 64K block size. The table will take up
896 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
898 while (hsize * 65536 < physmem * PAGESIZE)
901 buf_hash_table.ht_mask = hsize - 1;
902 #if defined(_KERNEL) && defined(HAVE_SPL)
903 /* Large allocations which do not require contiguous pages
904 * should be using vmem_alloc() in the linux kernel */
905 buf_hash_table.ht_table =
906 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
908 buf_hash_table.ht_table =
909 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
911 if (buf_hash_table.ht_table == NULL) {
912 ASSERT(hsize > (1ULL << 8));
917 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
918 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
919 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
920 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
922 for (i = 0; i < 256; i++)
923 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
924 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
926 for (i = 0; i < BUF_LOCKS; i++) {
927 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
928 NULL, MUTEX_DEFAULT, NULL);
932 #define ARC_MINTIME (hz>>4) /* 62 ms */
935 arc_cksum_verify(arc_buf_t *buf)
939 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
942 mutex_enter(&buf->b_hdr->b_freeze_lock);
943 if (buf->b_hdr->b_freeze_cksum == NULL ||
944 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
945 mutex_exit(&buf->b_hdr->b_freeze_lock);
948 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
949 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
950 panic("buffer modified while frozen!");
951 mutex_exit(&buf->b_hdr->b_freeze_lock);
955 arc_cksum_equal(arc_buf_t *buf)
960 mutex_enter(&buf->b_hdr->b_freeze_lock);
961 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
962 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
963 mutex_exit(&buf->b_hdr->b_freeze_lock);
969 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
971 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
974 mutex_enter(&buf->b_hdr->b_freeze_lock);
975 if (buf->b_hdr->b_freeze_cksum != NULL) {
976 mutex_exit(&buf->b_hdr->b_freeze_lock);
979 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
980 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
981 buf->b_hdr->b_freeze_cksum);
982 mutex_exit(&buf->b_hdr->b_freeze_lock);
986 arc_buf_thaw(arc_buf_t *buf)
988 if (zfs_flags & ZFS_DEBUG_MODIFY) {
989 if (buf->b_hdr->b_state != arc_anon)
990 panic("modifying non-anon buffer!");
991 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
992 panic("modifying buffer while i/o in progress!");
993 arc_cksum_verify(buf);
996 mutex_enter(&buf->b_hdr->b_freeze_lock);
997 if (buf->b_hdr->b_freeze_cksum != NULL) {
998 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
999 buf->b_hdr->b_freeze_cksum = NULL;
1002 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1003 if (buf->b_hdr->b_thawed)
1004 kmem_free(buf->b_hdr->b_thawed, 1);
1005 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1008 mutex_exit(&buf->b_hdr->b_freeze_lock);
1012 arc_buf_freeze(arc_buf_t *buf)
1014 kmutex_t *hash_lock;
1016 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1019 hash_lock = HDR_LOCK(buf->b_hdr);
1020 mutex_enter(hash_lock);
1022 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1023 buf->b_hdr->b_state == arc_anon);
1024 arc_cksum_compute(buf, B_FALSE);
1025 mutex_exit(hash_lock);
1029 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1031 ASSERT(MUTEX_HELD(hash_lock));
1033 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1034 (ab->b_state != arc_anon)) {
1035 uint64_t delta = ab->b_size * ab->b_datacnt;
1036 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1037 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1039 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1040 mutex_enter(&ab->b_state->arcs_mtx);
1041 ASSERT(list_link_active(&ab->b_arc_node));
1042 list_remove(list, ab);
1043 if (GHOST_STATE(ab->b_state)) {
1044 ASSERT3U(ab->b_datacnt, ==, 0);
1045 ASSERT3P(ab->b_buf, ==, NULL);
1049 ASSERT3U(*size, >=, delta);
1050 atomic_add_64(size, -delta);
1051 mutex_exit(&ab->b_state->arcs_mtx);
1052 /* remove the prefetch flag if we get a reference */
1053 if (ab->b_flags & ARC_PREFETCH)
1054 ab->b_flags &= ~ARC_PREFETCH;
1059 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1062 arc_state_t *state = ab->b_state;
1064 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1065 ASSERT(!GHOST_STATE(state));
1067 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1068 (state != arc_anon)) {
1069 uint64_t *size = &state->arcs_lsize[ab->b_type];
1071 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1072 mutex_enter(&state->arcs_mtx);
1073 ASSERT(!list_link_active(&ab->b_arc_node));
1074 list_insert_head(&state->arcs_list[ab->b_type], ab);
1075 ASSERT(ab->b_datacnt > 0);
1076 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1077 mutex_exit(&state->arcs_mtx);
1083 * Move the supplied buffer to the indicated state. The mutex
1084 * for the buffer must be held by the caller.
1087 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1089 arc_state_t *old_state = ab->b_state;
1090 int64_t refcnt = refcount_count(&ab->b_refcnt);
1091 uint64_t from_delta, to_delta;
1093 ASSERT(MUTEX_HELD(hash_lock));
1094 ASSERT(new_state != old_state);
1095 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1096 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1097 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1099 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1102 * If this buffer is evictable, transfer it from the
1103 * old state list to the new state list.
1106 if (old_state != arc_anon) {
1107 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1108 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1111 mutex_enter(&old_state->arcs_mtx);
1113 ASSERT(list_link_active(&ab->b_arc_node));
1114 list_remove(&old_state->arcs_list[ab->b_type], ab);
1117 * If prefetching out of the ghost cache,
1118 * we will have a non-zero datacnt.
1120 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1121 /* ghost elements have a ghost size */
1122 ASSERT(ab->b_buf == NULL);
1123 from_delta = ab->b_size;
1125 ASSERT3U(*size, >=, from_delta);
1126 atomic_add_64(size, -from_delta);
1129 mutex_exit(&old_state->arcs_mtx);
1131 if (new_state != arc_anon) {
1132 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1133 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1136 mutex_enter(&new_state->arcs_mtx);
1138 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1140 /* ghost elements have a ghost size */
1141 if (GHOST_STATE(new_state)) {
1142 ASSERT(ab->b_datacnt == 0);
1143 ASSERT(ab->b_buf == NULL);
1144 to_delta = ab->b_size;
1146 atomic_add_64(size, to_delta);
1149 mutex_exit(&new_state->arcs_mtx);
1153 ASSERT(!BUF_EMPTY(ab));
1154 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1155 buf_hash_remove(ab);
1157 /* adjust state sizes */
1159 atomic_add_64(&new_state->arcs_size, to_delta);
1161 ASSERT3U(old_state->arcs_size, >=, from_delta);
1162 atomic_add_64(&old_state->arcs_size, -from_delta);
1164 ab->b_state = new_state;
1166 /* adjust l2arc hdr stats */
1167 if (new_state == arc_l2c_only)
1168 l2arc_hdr_stat_add();
1169 else if (old_state == arc_l2c_only)
1170 l2arc_hdr_stat_remove();
1174 arc_space_consume(uint64_t space, arc_space_type_t type)
1176 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1181 case ARC_SPACE_DATA:
1182 ARCSTAT_INCR(arcstat_data_size, space);
1184 case ARC_SPACE_OTHER:
1185 ARCSTAT_INCR(arcstat_other_size, space);
1187 case ARC_SPACE_HDRS:
1188 ARCSTAT_INCR(arcstat_hdr_size, space);
1190 case ARC_SPACE_L2HDRS:
1191 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1195 atomic_add_64(&arc_meta_used, space);
1196 atomic_add_64(&arc_size, space);
1200 arc_space_return(uint64_t space, arc_space_type_t type)
1202 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1207 case ARC_SPACE_DATA:
1208 ARCSTAT_INCR(arcstat_data_size, -space);
1210 case ARC_SPACE_OTHER:
1211 ARCSTAT_INCR(arcstat_other_size, -space);
1213 case ARC_SPACE_HDRS:
1214 ARCSTAT_INCR(arcstat_hdr_size, -space);
1216 case ARC_SPACE_L2HDRS:
1217 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1221 ASSERT(arc_meta_used >= space);
1222 if (arc_meta_max < arc_meta_used)
1223 arc_meta_max = arc_meta_used;
1224 atomic_add_64(&arc_meta_used, -space);
1225 ASSERT(arc_size >= space);
1226 atomic_add_64(&arc_size, -space);
1230 arc_data_buf_alloc(uint64_t size)
1232 if (arc_evict_needed(ARC_BUFC_DATA))
1233 cv_signal(&arc_reclaim_thr_cv);
1234 atomic_add_64(&arc_size, size);
1235 return (zio_data_buf_alloc(size));
1239 arc_data_buf_free(void *buf, uint64_t size)
1241 zio_data_buf_free(buf, size);
1242 ASSERT(arc_size >= size);
1243 atomic_add_64(&arc_size, -size);
1247 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1252 ASSERT3U(size, >, 0);
1253 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1254 ASSERT(BUF_EMPTY(hdr));
1257 hdr->b_spa = spa_guid(spa);
1258 hdr->b_state = arc_anon;
1259 hdr->b_arc_access = 0;
1260 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1263 buf->b_efunc = NULL;
1264 buf->b_private = NULL;
1267 arc_get_data_buf(buf);
1270 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1271 (void) refcount_add(&hdr->b_refcnt, tag);
1276 static char *arc_onloan_tag = "onloan";
1279 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1280 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1281 * buffers must be returned to the arc before they can be used by the DMU or
1285 arc_loan_buf(spa_t *spa, int size)
1289 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1291 atomic_add_64(&arc_loaned_bytes, size);
1296 * Return a loaned arc buffer to the arc.
1299 arc_return_buf(arc_buf_t *buf, void *tag)
1301 arc_buf_hdr_t *hdr = buf->b_hdr;
1303 ASSERT(buf->b_data != NULL);
1304 (void) refcount_add(&hdr->b_refcnt, tag);
1305 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1307 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1310 /* Detach an arc_buf from a dbuf (tag) */
1312 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1316 ASSERT(buf->b_data != NULL);
1318 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1319 (void) refcount_remove(&hdr->b_refcnt, tag);
1320 buf->b_efunc = NULL;
1321 buf->b_private = NULL;
1323 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1327 arc_buf_clone(arc_buf_t *from)
1330 arc_buf_hdr_t *hdr = from->b_hdr;
1331 uint64_t size = hdr->b_size;
1333 ASSERT(hdr->b_state != arc_anon);
1335 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1338 buf->b_efunc = NULL;
1339 buf->b_private = NULL;
1340 buf->b_next = hdr->b_buf;
1342 arc_get_data_buf(buf);
1343 bcopy(from->b_data, buf->b_data, size);
1344 hdr->b_datacnt += 1;
1349 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1352 kmutex_t *hash_lock;
1355 * Check to see if this buffer is evicted. Callers
1356 * must verify b_data != NULL to know if the add_ref
1359 mutex_enter(&buf->b_evict_lock);
1360 if (buf->b_data == NULL) {
1361 mutex_exit(&buf->b_evict_lock);
1364 hash_lock = HDR_LOCK(buf->b_hdr);
1365 mutex_enter(hash_lock);
1367 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1368 mutex_exit(&buf->b_evict_lock);
1370 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1371 add_reference(hdr, hash_lock, tag);
1372 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1373 arc_access(hdr, hash_lock);
1374 mutex_exit(hash_lock);
1375 ARCSTAT_BUMP(arcstat_hits);
1376 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1377 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1378 data, metadata, hits);
1382 * Free the arc data buffer. If it is an l2arc write in progress,
1383 * the buffer is placed on l2arc_free_on_write to be freed later.
1386 arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t),
1387 void *data, size_t size)
1389 if (HDR_L2_WRITING(hdr)) {
1390 l2arc_data_free_t *df;
1391 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1392 df->l2df_data = data;
1393 df->l2df_size = size;
1394 df->l2df_func = free_func;
1395 mutex_enter(&l2arc_free_on_write_mtx);
1396 list_insert_head(l2arc_free_on_write, df);
1397 mutex_exit(&l2arc_free_on_write_mtx);
1398 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1400 free_func(data, size);
1405 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1409 /* free up data associated with the buf */
1411 arc_state_t *state = buf->b_hdr->b_state;
1412 uint64_t size = buf->b_hdr->b_size;
1413 arc_buf_contents_t type = buf->b_hdr->b_type;
1415 arc_cksum_verify(buf);
1418 if (type == ARC_BUFC_METADATA) {
1419 arc_buf_data_free(buf->b_hdr, zio_buf_free,
1421 arc_space_return(size, ARC_SPACE_DATA);
1423 ASSERT(type == ARC_BUFC_DATA);
1424 arc_buf_data_free(buf->b_hdr,
1425 zio_data_buf_free, buf->b_data, size);
1426 ARCSTAT_INCR(arcstat_data_size, -size);
1427 atomic_add_64(&arc_size, -size);
1430 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1431 uint64_t *cnt = &state->arcs_lsize[type];
1433 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1434 ASSERT(state != arc_anon);
1436 ASSERT3U(*cnt, >=, size);
1437 atomic_add_64(cnt, -size);
1439 ASSERT3U(state->arcs_size, >=, size);
1440 atomic_add_64(&state->arcs_size, -size);
1442 ASSERT(buf->b_hdr->b_datacnt > 0);
1443 buf->b_hdr->b_datacnt -= 1;
1446 /* only remove the buf if requested */
1450 /* remove the buf from the hdr list */
1451 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1453 *bufp = buf->b_next;
1456 ASSERT(buf->b_efunc == NULL);
1458 /* clean up the buf */
1460 kmem_cache_free(buf_cache, buf);
1464 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1466 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1468 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1469 ASSERT3P(hdr->b_state, ==, arc_anon);
1470 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1472 if (l2hdr != NULL) {
1473 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1475 * To prevent arc_free() and l2arc_evict() from
1476 * attempting to free the same buffer at the same time,
1477 * a FREE_IN_PROGRESS flag is given to arc_free() to
1478 * give it priority. l2arc_evict() can't destroy this
1479 * header while we are waiting on l2arc_buflist_mtx.
1481 * The hdr may be removed from l2ad_buflist before we
1482 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1484 if (!buflist_held) {
1485 mutex_enter(&l2arc_buflist_mtx);
1486 l2hdr = hdr->b_l2hdr;
1489 if (l2hdr != NULL) {
1490 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1491 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1492 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1493 if (hdr->b_state == arc_l2c_only)
1494 l2arc_hdr_stat_remove();
1495 hdr->b_l2hdr = NULL;
1499 mutex_exit(&l2arc_buflist_mtx);
1502 if (!BUF_EMPTY(hdr)) {
1503 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1504 buf_discard_identity(hdr);
1506 while (hdr->b_buf) {
1507 arc_buf_t *buf = hdr->b_buf;
1510 mutex_enter(&arc_eviction_mtx);
1511 mutex_enter(&buf->b_evict_lock);
1512 ASSERT(buf->b_hdr != NULL);
1513 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1514 hdr->b_buf = buf->b_next;
1515 buf->b_hdr = &arc_eviction_hdr;
1516 buf->b_next = arc_eviction_list;
1517 arc_eviction_list = buf;
1518 mutex_exit(&buf->b_evict_lock);
1519 mutex_exit(&arc_eviction_mtx);
1521 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1524 if (hdr->b_freeze_cksum != NULL) {
1525 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1526 hdr->b_freeze_cksum = NULL;
1528 if (hdr->b_thawed) {
1529 kmem_free(hdr->b_thawed, 1);
1530 hdr->b_thawed = NULL;
1533 ASSERT(!list_link_active(&hdr->b_arc_node));
1534 ASSERT3P(hdr->b_hash_next, ==, NULL);
1535 ASSERT3P(hdr->b_acb, ==, NULL);
1536 kmem_cache_free(hdr_cache, hdr);
1540 arc_buf_free(arc_buf_t *buf, void *tag)
1542 arc_buf_hdr_t *hdr = buf->b_hdr;
1543 int hashed = hdr->b_state != arc_anon;
1545 ASSERT(buf->b_efunc == NULL);
1546 ASSERT(buf->b_data != NULL);
1549 kmutex_t *hash_lock = HDR_LOCK(hdr);
1551 mutex_enter(hash_lock);
1553 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1555 (void) remove_reference(hdr, hash_lock, tag);
1556 if (hdr->b_datacnt > 1) {
1557 arc_buf_destroy(buf, FALSE, TRUE);
1559 ASSERT(buf == hdr->b_buf);
1560 ASSERT(buf->b_efunc == NULL);
1561 hdr->b_flags |= ARC_BUF_AVAILABLE;
1563 mutex_exit(hash_lock);
1564 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1567 * We are in the middle of an async write. Don't destroy
1568 * this buffer unless the write completes before we finish
1569 * decrementing the reference count.
1571 mutex_enter(&arc_eviction_mtx);
1572 (void) remove_reference(hdr, NULL, tag);
1573 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1574 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1575 mutex_exit(&arc_eviction_mtx);
1577 arc_hdr_destroy(hdr);
1579 if (remove_reference(hdr, NULL, tag) > 0)
1580 arc_buf_destroy(buf, FALSE, TRUE);
1582 arc_hdr_destroy(hdr);
1587 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1589 arc_buf_hdr_t *hdr = buf->b_hdr;
1590 kmutex_t *hash_lock = HDR_LOCK(hdr);
1591 int no_callback = (buf->b_efunc == NULL);
1593 if (hdr->b_state == arc_anon) {
1594 ASSERT(hdr->b_datacnt == 1);
1595 arc_buf_free(buf, tag);
1596 return (no_callback);
1599 mutex_enter(hash_lock);
1601 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1602 ASSERT(hdr->b_state != arc_anon);
1603 ASSERT(buf->b_data != NULL);
1605 (void) remove_reference(hdr, hash_lock, tag);
1606 if (hdr->b_datacnt > 1) {
1608 arc_buf_destroy(buf, FALSE, TRUE);
1609 } else if (no_callback) {
1610 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1611 ASSERT(buf->b_efunc == NULL);
1612 hdr->b_flags |= ARC_BUF_AVAILABLE;
1614 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1615 refcount_is_zero(&hdr->b_refcnt));
1616 mutex_exit(hash_lock);
1617 return (no_callback);
1621 arc_buf_size(arc_buf_t *buf)
1623 return (buf->b_hdr->b_size);
1627 * Evict buffers from list until we've removed the specified number of
1628 * bytes. Move the removed buffers to the appropriate evict state.
1629 * If the recycle flag is set, then attempt to "recycle" a buffer:
1630 * - look for a buffer to evict that is `bytes' long.
1631 * - return the data block from this buffer rather than freeing it.
1632 * This flag is used by callers that are trying to make space for a
1633 * new buffer in a full arc cache.
1635 * This function makes a "best effort". It skips over any buffers
1636 * it can't get a hash_lock on, and so may not catch all candidates.
1637 * It may also return without evicting as much space as requested.
1640 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1641 arc_buf_contents_t type)
1643 arc_state_t *evicted_state;
1644 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1645 arc_buf_hdr_t *ab, *ab_prev = NULL;
1646 list_t *list = &state->arcs_list[type];
1647 kmutex_t *hash_lock;
1648 boolean_t have_lock;
1649 void *stolen = NULL;
1651 ASSERT(state == arc_mru || state == arc_mfu);
1653 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1655 mutex_enter(&state->arcs_mtx);
1656 mutex_enter(&evicted_state->arcs_mtx);
1658 for (ab = list_tail(list); ab; ab = ab_prev) {
1659 ab_prev = list_prev(list, ab);
1660 /* prefetch buffers have a minimum lifespan */
1661 if (HDR_IO_IN_PROGRESS(ab) ||
1662 (spa && ab->b_spa != spa) ||
1663 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1664 ddi_get_lbolt() - ab->b_arc_access <
1665 arc_min_prefetch_lifespan)) {
1669 /* "lookahead" for better eviction candidate */
1670 if (recycle && ab->b_size != bytes &&
1671 ab_prev && ab_prev->b_size == bytes)
1673 hash_lock = HDR_LOCK(ab);
1674 have_lock = MUTEX_HELD(hash_lock);
1675 if (have_lock || mutex_tryenter(hash_lock)) {
1676 ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0);
1677 ASSERT(ab->b_datacnt > 0);
1679 arc_buf_t *buf = ab->b_buf;
1680 if (!mutex_tryenter(&buf->b_evict_lock)) {
1685 bytes_evicted += ab->b_size;
1686 if (recycle && ab->b_type == type &&
1687 ab->b_size == bytes &&
1688 !HDR_L2_WRITING(ab)) {
1689 stolen = buf->b_data;
1694 mutex_enter(&arc_eviction_mtx);
1695 arc_buf_destroy(buf,
1696 buf->b_data == stolen, FALSE);
1697 ab->b_buf = buf->b_next;
1698 buf->b_hdr = &arc_eviction_hdr;
1699 buf->b_next = arc_eviction_list;
1700 arc_eviction_list = buf;
1701 mutex_exit(&arc_eviction_mtx);
1702 mutex_exit(&buf->b_evict_lock);
1704 mutex_exit(&buf->b_evict_lock);
1705 arc_buf_destroy(buf,
1706 buf->b_data == stolen, TRUE);
1711 ARCSTAT_INCR(arcstat_evict_l2_cached,
1714 if (l2arc_write_eligible(ab->b_spa, ab)) {
1715 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1719 arcstat_evict_l2_ineligible,
1724 if (ab->b_datacnt == 0) {
1725 arc_change_state(evicted_state, ab, hash_lock);
1726 ASSERT(HDR_IN_HASH_TABLE(ab));
1727 ab->b_flags |= ARC_IN_HASH_TABLE;
1728 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1729 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1732 mutex_exit(hash_lock);
1733 if (bytes >= 0 && bytes_evicted >= bytes)
1740 mutex_exit(&evicted_state->arcs_mtx);
1741 mutex_exit(&state->arcs_mtx);
1743 if (bytes_evicted < bytes)
1744 dprintf("only evicted %lld bytes from %x\n",
1745 (longlong_t)bytes_evicted, state);
1748 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1751 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1754 * We have just evicted some date into the ghost state, make
1755 * sure we also adjust the ghost state size if necessary.
1758 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1759 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1760 arc_mru_ghost->arcs_size - arc_c;
1762 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1764 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1765 arc_evict_ghost(arc_mru_ghost, 0, todelete);
1766 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1767 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1768 arc_mru_ghost->arcs_size +
1769 arc_mfu_ghost->arcs_size - arc_c);
1770 arc_evict_ghost(arc_mfu_ghost, 0, todelete);
1778 * Remove buffers from list until we've removed the specified number of
1779 * bytes. Destroy the buffers that are removed.
1782 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1784 arc_buf_hdr_t *ab, *ab_prev;
1785 arc_buf_hdr_t marker;
1786 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1787 kmutex_t *hash_lock;
1788 uint64_t bytes_deleted = 0;
1789 uint64_t bufs_skipped = 0;
1791 ASSERT(GHOST_STATE(state));
1792 bzero(&marker, sizeof(marker));
1794 mutex_enter(&state->arcs_mtx);
1795 for (ab = list_tail(list); ab; ab = ab_prev) {
1796 ab_prev = list_prev(list, ab);
1797 if (spa && ab->b_spa != spa)
1800 /* ignore markers */
1804 hash_lock = HDR_LOCK(ab);
1805 /* caller may be trying to modify this buffer, skip it */
1806 if (MUTEX_HELD(hash_lock))
1808 if (mutex_tryenter(hash_lock)) {
1809 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1810 ASSERT(ab->b_buf == NULL);
1811 ARCSTAT_BUMP(arcstat_deleted);
1812 bytes_deleted += ab->b_size;
1814 if (ab->b_l2hdr != NULL) {
1816 * This buffer is cached on the 2nd Level ARC;
1817 * don't destroy the header.
1819 arc_change_state(arc_l2c_only, ab, hash_lock);
1820 mutex_exit(hash_lock);
1822 arc_change_state(arc_anon, ab, hash_lock);
1823 mutex_exit(hash_lock);
1824 arc_hdr_destroy(ab);
1827 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1828 if (bytes >= 0 && bytes_deleted >= bytes)
1830 } else if (bytes < 0) {
1832 * Insert a list marker and then wait for the
1833 * hash lock to become available. Once its
1834 * available, restart from where we left off.
1836 list_insert_after(list, ab, &marker);
1837 mutex_exit(&state->arcs_mtx);
1838 mutex_enter(hash_lock);
1839 mutex_exit(hash_lock);
1840 mutex_enter(&state->arcs_mtx);
1841 ab_prev = list_prev(list, &marker);
1842 list_remove(list, &marker);
1846 mutex_exit(&state->arcs_mtx);
1848 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1849 (bytes < 0 || bytes_deleted < bytes)) {
1850 list = &state->arcs_list[ARC_BUFC_METADATA];
1855 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1859 if (bytes_deleted < bytes)
1860 dprintf("only deleted %lld bytes from %p\n",
1861 (longlong_t)bytes_deleted, state);
1867 int64_t adjustment, delta;
1873 adjustment = MIN((int64_t)(arc_size - arc_c),
1874 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
1877 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1878 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
1879 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
1880 adjustment -= delta;
1883 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1884 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
1885 (void) arc_evict(arc_mru, 0, delta, FALSE,
1893 adjustment = arc_size - arc_c;
1895 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1896 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
1897 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
1898 adjustment -= delta;
1901 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1902 int64_t delta = MIN(adjustment,
1903 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
1904 (void) arc_evict(arc_mfu, 0, delta, FALSE,
1909 * Adjust ghost lists
1912 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
1914 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
1915 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
1916 arc_evict_ghost(arc_mru_ghost, 0, delta);
1920 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
1922 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
1923 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
1924 arc_evict_ghost(arc_mfu_ghost, 0, delta);
1929 arc_do_user_evicts(void)
1931 mutex_enter(&arc_eviction_mtx);
1932 while (arc_eviction_list != NULL) {
1933 arc_buf_t *buf = arc_eviction_list;
1934 arc_eviction_list = buf->b_next;
1935 mutex_enter(&buf->b_evict_lock);
1937 mutex_exit(&buf->b_evict_lock);
1938 mutex_exit(&arc_eviction_mtx);
1940 if (buf->b_efunc != NULL)
1941 VERIFY(buf->b_efunc(buf) == 0);
1943 buf->b_efunc = NULL;
1944 buf->b_private = NULL;
1945 kmem_cache_free(buf_cache, buf);
1946 mutex_enter(&arc_eviction_mtx);
1948 mutex_exit(&arc_eviction_mtx);
1952 * Flush all *evictable* data from the cache for the given spa.
1953 * NOTE: this will not touch "active" (i.e. referenced) data.
1956 arc_flush(spa_t *spa)
1961 guid = spa_guid(spa);
1963 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
1964 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
1968 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
1969 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
1973 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
1974 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
1978 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
1979 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
1984 arc_evict_ghost(arc_mru_ghost, guid, -1);
1985 arc_evict_ghost(arc_mfu_ghost, guid, -1);
1987 mutex_enter(&arc_reclaim_thr_lock);
1988 arc_do_user_evicts();
1989 mutex_exit(&arc_reclaim_thr_lock);
1990 ASSERT(spa || arc_eviction_list == NULL);
1996 if (arc_c > arc_c_min) {
2000 to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
2002 to_free = arc_c >> arc_shrink_shift;
2004 if (arc_c > arc_c_min + to_free)
2005 atomic_add_64(&arc_c, -to_free);
2009 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2010 if (arc_c > arc_size)
2011 arc_c = MAX(arc_size, arc_c_min);
2013 arc_p = (arc_c >> 1);
2014 ASSERT(arc_c >= arc_c_min);
2015 ASSERT((int64_t)arc_p >= 0);
2018 if (arc_size > arc_c)
2023 arc_reclaim_needed(void)
2032 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2037 * check that we're out of range of the pageout scanner. It starts to
2038 * schedule paging if freemem is less than lotsfree and needfree.
2039 * lotsfree is the high-water mark for pageout, and needfree is the
2040 * number of needed free pages. We add extra pages here to make sure
2041 * the scanner doesn't start up while we're freeing memory.
2043 if (freemem < lotsfree + needfree + extra)
2047 * check to make sure that swapfs has enough space so that anon
2048 * reservations can still succeed. anon_resvmem() checks that the
2049 * availrmem is greater than swapfs_minfree, and the number of reserved
2050 * swap pages. We also add a bit of extra here just to prevent
2051 * circumstances from getting really dire.
2053 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2058 * If we're on an i386 platform, it's possible that we'll exhaust the
2059 * kernel heap space before we ever run out of available physical
2060 * memory. Most checks of the size of the heap_area compare against
2061 * tune.t_minarmem, which is the minimum available real memory that we
2062 * can have in the system. However, this is generally fixed at 25 pages
2063 * which is so low that it's useless. In this comparison, we seek to
2064 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2065 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2068 if (btop(vmem_size(heap_arena, VMEM_FREE)) <
2069 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
2074 if (spa_get_random(100) == 0)
2081 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2084 kmem_cache_t *prev_cache = NULL;
2085 kmem_cache_t *prev_data_cache = NULL;
2086 extern kmem_cache_t *zio_buf_cache[];
2087 extern kmem_cache_t *zio_data_buf_cache[];
2091 while ((arc_meta_used >= arc_meta_limit) && (retry < 10)) {
2093 * We are exceeding our meta-data cache limit.
2094 * Purge some DNLC entries to release holds on meta-data.
2096 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2101 * Reclaim unused memory from all kmem caches.
2108 * An aggressive reclamation will shrink the cache size as well as
2109 * reap free buffers from the arc kmem caches.
2111 if (strat == ARC_RECLAIM_AGGR)
2114 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2115 if (zio_buf_cache[i] != prev_cache) {
2116 prev_cache = zio_buf_cache[i];
2117 kmem_cache_reap_now(zio_buf_cache[i]);
2119 if (zio_data_buf_cache[i] != prev_data_cache) {
2120 prev_data_cache = zio_data_buf_cache[i];
2121 kmem_cache_reap_now(zio_data_buf_cache[i]);
2124 kmem_cache_reap_now(buf_cache);
2125 kmem_cache_reap_now(hdr_cache);
2129 arc_reclaim_thread(void)
2131 clock_t growtime = 0;
2132 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2135 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2137 mutex_enter(&arc_reclaim_thr_lock);
2138 while (arc_thread_exit == 0) {
2139 if (arc_reclaim_needed()) {
2142 if (last_reclaim == ARC_RECLAIM_CONS) {
2143 last_reclaim = ARC_RECLAIM_AGGR;
2145 last_reclaim = ARC_RECLAIM_CONS;
2149 last_reclaim = ARC_RECLAIM_AGGR;
2153 /* reset the growth delay for every reclaim */
2154 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2156 arc_kmem_reap_now(last_reclaim);
2159 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2160 arc_no_grow = FALSE;
2163 /* Keep meta data usage within limits */
2164 if (arc_meta_used >= arc_meta_limit)
2165 arc_kmem_reap_now(ARC_RECLAIM_CONS);
2169 if (arc_eviction_list != NULL)
2170 arc_do_user_evicts();
2172 /* block until needed, or one second, whichever is shorter */
2173 CALLB_CPR_SAFE_BEGIN(&cpr);
2174 (void) cv_timedwait_interruptible(&arc_reclaim_thr_cv,
2175 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2176 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2179 arc_thread_exit = 0;
2180 cv_broadcast(&arc_reclaim_thr_cv);
2181 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2187 * Under Linux the arc shrinker may be called for synchronous (direct)
2188 * reclaim, or asynchronous (indirect) reclaim. When called by kswapd
2189 * for indirect reclaim we take a conservative approach and just reap
2190 * free slabs from the ARC caches. If this proves to be insufficient
2191 * direct reclaim will be trigger. In direct reclaim a more aggressive
2192 * strategy is used, data is evicted from the ARC and free slabs reaped.
2194 SPL_SHRINKER_CALLBACK_PROTO(arc_shrinker_func, cb, nr_to_scan, gfp_mask)
2196 arc_reclaim_strategy_t strategy;
2199 /* Not allowed to perform filesystem reclaim */
2200 if (!(gfp_mask & __GFP_FS))
2203 /* Return number of reclaimable pages based on arc_shrink_shift */
2204 arc_reclaim = btop((arc_size - arc_c_min)) >> arc_shrink_shift;
2205 if (nr_to_scan == 0)
2206 return (arc_reclaim);
2208 /* Reclaim in progress */
2209 if (mutex_tryenter(&arc_reclaim_thr_lock) == 0)
2212 if (current_is_kswapd()) {
2213 strategy = ARC_RECLAIM_CONS;
2214 ARCSTAT_INCR(arcstat_memory_indirect_count, 1);
2216 strategy = ARC_RECLAIM_AGGR;
2217 ARCSTAT_INCR(arcstat_memory_direct_count, 1);
2220 arc_kmem_reap_now(strategy);
2221 arc_reclaim = btop((arc_size - arc_c_min)) >> arc_shrink_shift;
2222 mutex_exit(&arc_reclaim_thr_lock);
2224 return (arc_reclaim);
2227 SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS);
2228 #endif /* _KERNEL */
2231 * Adapt arc info given the number of bytes we are trying to add and
2232 * the state that we are comming from. This function is only called
2233 * when we are adding new content to the cache.
2236 arc_adapt(int bytes, arc_state_t *state)
2239 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2241 if (state == arc_l2c_only)
2246 * Adapt the target size of the MRU list:
2247 * - if we just hit in the MRU ghost list, then increase
2248 * the target size of the MRU list.
2249 * - if we just hit in the MFU ghost list, then increase
2250 * the target size of the MFU list by decreasing the
2251 * target size of the MRU list.
2253 if (state == arc_mru_ghost) {
2254 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2255 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2256 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2258 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2259 } else if (state == arc_mfu_ghost) {
2262 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2263 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2264 mult = MIN(mult, 10);
2266 delta = MIN(bytes * mult, arc_p);
2267 arc_p = MAX(arc_p_min, arc_p - delta);
2269 ASSERT((int64_t)arc_p >= 0);
2271 if (arc_reclaim_needed()) {
2272 cv_signal(&arc_reclaim_thr_cv);
2279 if (arc_c >= arc_c_max)
2283 * If we're within (2 * maxblocksize) bytes of the target
2284 * cache size, increment the target cache size
2286 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2287 atomic_add_64(&arc_c, (int64_t)bytes);
2288 if (arc_c > arc_c_max)
2290 else if (state == arc_anon)
2291 atomic_add_64(&arc_p, (int64_t)bytes);
2295 ASSERT((int64_t)arc_p >= 0);
2299 * Check if the cache has reached its limits and eviction is required
2303 arc_evict_needed(arc_buf_contents_t type)
2305 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2310 * If zio data pages are being allocated out of a separate heap segment,
2311 * then enforce that the size of available vmem for this area remains
2312 * above about 1/32nd free.
2314 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2315 vmem_size(zio_arena, VMEM_FREE) <
2316 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2320 if (arc_reclaim_needed())
2323 return (arc_size > arc_c);
2327 * The buffer, supplied as the first argument, needs a data block.
2328 * So, if we are at cache max, determine which cache should be victimized.
2329 * We have the following cases:
2331 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2332 * In this situation if we're out of space, but the resident size of the MFU is
2333 * under the limit, victimize the MFU cache to satisfy this insertion request.
2335 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2336 * Here, we've used up all of the available space for the MRU, so we need to
2337 * evict from our own cache instead. Evict from the set of resident MRU
2340 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2341 * c minus p represents the MFU space in the cache, since p is the size of the
2342 * cache that is dedicated to the MRU. In this situation there's still space on
2343 * the MFU side, so the MRU side needs to be victimized.
2345 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2346 * MFU's resident set is consuming more space than it has been allotted. In
2347 * this situation, we must victimize our own cache, the MFU, for this insertion.
2350 arc_get_data_buf(arc_buf_t *buf)
2352 arc_state_t *state = buf->b_hdr->b_state;
2353 uint64_t size = buf->b_hdr->b_size;
2354 arc_buf_contents_t type = buf->b_hdr->b_type;
2356 arc_adapt(size, state);
2359 * We have not yet reached cache maximum size,
2360 * just allocate a new buffer.
2362 if (!arc_evict_needed(type)) {
2363 if (type == ARC_BUFC_METADATA) {
2364 buf->b_data = zio_buf_alloc(size);
2365 arc_space_consume(size, ARC_SPACE_DATA);
2367 ASSERT(type == ARC_BUFC_DATA);
2368 buf->b_data = zio_data_buf_alloc(size);
2369 ARCSTAT_INCR(arcstat_data_size, size);
2370 atomic_add_64(&arc_size, size);
2376 * If we are prefetching from the mfu ghost list, this buffer
2377 * will end up on the mru list; so steal space from there.
2379 if (state == arc_mfu_ghost)
2380 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2381 else if (state == arc_mru_ghost)
2384 if (state == arc_mru || state == arc_anon) {
2385 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2386 state = (arc_mfu->arcs_lsize[type] >= size &&
2387 arc_p > mru_used) ? arc_mfu : arc_mru;
2390 uint64_t mfu_space = arc_c - arc_p;
2391 state = (arc_mru->arcs_lsize[type] >= size &&
2392 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2394 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2395 if (type == ARC_BUFC_METADATA) {
2396 buf->b_data = zio_buf_alloc(size);
2397 arc_space_consume(size, ARC_SPACE_DATA);
2399 ASSERT(type == ARC_BUFC_DATA);
2400 buf->b_data = zio_data_buf_alloc(size);
2401 ARCSTAT_INCR(arcstat_data_size, size);
2402 atomic_add_64(&arc_size, size);
2404 ARCSTAT_BUMP(arcstat_recycle_miss);
2406 ASSERT(buf->b_data != NULL);
2409 * Update the state size. Note that ghost states have a
2410 * "ghost size" and so don't need to be updated.
2412 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2413 arc_buf_hdr_t *hdr = buf->b_hdr;
2415 atomic_add_64(&hdr->b_state->arcs_size, size);
2416 if (list_link_active(&hdr->b_arc_node)) {
2417 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2418 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2421 * If we are growing the cache, and we are adding anonymous
2422 * data, and we have outgrown arc_p, update arc_p
2424 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2425 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2426 arc_p = MIN(arc_c, arc_p + size);
2431 * This routine is called whenever a buffer is accessed.
2432 * NOTE: the hash lock is dropped in this function.
2435 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2439 ASSERT(MUTEX_HELD(hash_lock));
2441 if (buf->b_state == arc_anon) {
2443 * This buffer is not in the cache, and does not
2444 * appear in our "ghost" list. Add the new buffer
2448 ASSERT(buf->b_arc_access == 0);
2449 buf->b_arc_access = ddi_get_lbolt();
2450 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2451 arc_change_state(arc_mru, buf, hash_lock);
2453 } else if (buf->b_state == arc_mru) {
2454 now = ddi_get_lbolt();
2457 * If this buffer is here because of a prefetch, then either:
2458 * - clear the flag if this is a "referencing" read
2459 * (any subsequent access will bump this into the MFU state).
2461 * - move the buffer to the head of the list if this is
2462 * another prefetch (to make it less likely to be evicted).
2464 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2465 if (refcount_count(&buf->b_refcnt) == 0) {
2466 ASSERT(list_link_active(&buf->b_arc_node));
2468 buf->b_flags &= ~ARC_PREFETCH;
2469 ARCSTAT_BUMP(arcstat_mru_hits);
2471 buf->b_arc_access = now;
2476 * This buffer has been "accessed" only once so far,
2477 * but it is still in the cache. Move it to the MFU
2480 if (now > buf->b_arc_access + ARC_MINTIME) {
2482 * More than 125ms have passed since we
2483 * instantiated this buffer. Move it to the
2484 * most frequently used state.
2486 buf->b_arc_access = now;
2487 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2488 arc_change_state(arc_mfu, buf, hash_lock);
2490 ARCSTAT_BUMP(arcstat_mru_hits);
2491 } else if (buf->b_state == arc_mru_ghost) {
2492 arc_state_t *new_state;
2494 * This buffer has been "accessed" recently, but
2495 * was evicted from the cache. Move it to the
2499 if (buf->b_flags & ARC_PREFETCH) {
2500 new_state = arc_mru;
2501 if (refcount_count(&buf->b_refcnt) > 0)
2502 buf->b_flags &= ~ARC_PREFETCH;
2503 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2505 new_state = arc_mfu;
2506 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2509 buf->b_arc_access = ddi_get_lbolt();
2510 arc_change_state(new_state, buf, hash_lock);
2512 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2513 } else if (buf->b_state == arc_mfu) {
2515 * This buffer has been accessed more than once and is
2516 * still in the cache. Keep it in the MFU state.
2518 * NOTE: an add_reference() that occurred when we did
2519 * the arc_read() will have kicked this off the list.
2520 * If it was a prefetch, we will explicitly move it to
2521 * the head of the list now.
2523 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2524 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2525 ASSERT(list_link_active(&buf->b_arc_node));
2527 ARCSTAT_BUMP(arcstat_mfu_hits);
2528 buf->b_arc_access = ddi_get_lbolt();
2529 } else if (buf->b_state == arc_mfu_ghost) {
2530 arc_state_t *new_state = arc_mfu;
2532 * This buffer has been accessed more than once but has
2533 * been evicted from the cache. Move it back to the
2537 if (buf->b_flags & ARC_PREFETCH) {
2539 * This is a prefetch access...
2540 * move this block back to the MRU state.
2542 ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0);
2543 new_state = arc_mru;
2546 buf->b_arc_access = ddi_get_lbolt();
2547 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2548 arc_change_state(new_state, buf, hash_lock);
2550 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2551 } else if (buf->b_state == arc_l2c_only) {
2553 * This buffer is on the 2nd Level ARC.
2556 buf->b_arc_access = ddi_get_lbolt();
2557 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2558 arc_change_state(arc_mfu, buf, hash_lock);
2560 ASSERT(!"invalid arc state");
2564 /* a generic arc_done_func_t which you can use */
2567 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2569 if (zio == NULL || zio->io_error == 0)
2570 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2571 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2574 /* a generic arc_done_func_t */
2576 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2578 arc_buf_t **bufp = arg;
2579 if (zio && zio->io_error) {
2580 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2584 ASSERT(buf->b_data);
2589 arc_read_done(zio_t *zio)
2591 arc_buf_hdr_t *hdr, *found;
2593 arc_buf_t *abuf; /* buffer we're assigning to callback */
2594 kmutex_t *hash_lock;
2595 arc_callback_t *callback_list, *acb;
2596 int freeable = FALSE;
2598 buf = zio->io_private;
2602 * The hdr was inserted into hash-table and removed from lists
2603 * prior to starting I/O. We should find this header, since
2604 * it's in the hash table, and it should be legit since it's
2605 * not possible to evict it during the I/O. The only possible
2606 * reason for it not to be found is if we were freed during the
2609 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2612 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2613 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2614 (found == hdr && HDR_L2_READING(hdr)));
2616 hdr->b_flags &= ~ARC_L2_EVICTED;
2617 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2618 hdr->b_flags &= ~ARC_L2CACHE;
2620 /* byteswap if necessary */
2621 callback_list = hdr->b_acb;
2622 ASSERT(callback_list != NULL);
2623 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2624 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2625 byteswap_uint64_array :
2626 dmu_ot[BP_GET_TYPE(zio->io_bp)].ot_byteswap;
2627 func(buf->b_data, hdr->b_size);
2630 arc_cksum_compute(buf, B_FALSE);
2632 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2634 * Only call arc_access on anonymous buffers. This is because
2635 * if we've issued an I/O for an evicted buffer, we've already
2636 * called arc_access (to prevent any simultaneous readers from
2637 * getting confused).
2639 arc_access(hdr, hash_lock);
2642 /* create copies of the data buffer for the callers */
2644 for (acb = callback_list; acb; acb = acb->acb_next) {
2645 if (acb->acb_done) {
2647 abuf = arc_buf_clone(buf);
2648 acb->acb_buf = abuf;
2653 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2654 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2656 ASSERT(buf->b_efunc == NULL);
2657 ASSERT(hdr->b_datacnt == 1);
2658 hdr->b_flags |= ARC_BUF_AVAILABLE;
2661 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2663 if (zio->io_error != 0) {
2664 hdr->b_flags |= ARC_IO_ERROR;
2665 if (hdr->b_state != arc_anon)
2666 arc_change_state(arc_anon, hdr, hash_lock);
2667 if (HDR_IN_HASH_TABLE(hdr))
2668 buf_hash_remove(hdr);
2669 freeable = refcount_is_zero(&hdr->b_refcnt);
2673 * Broadcast before we drop the hash_lock to avoid the possibility
2674 * that the hdr (and hence the cv) might be freed before we get to
2675 * the cv_broadcast().
2677 cv_broadcast(&hdr->b_cv);
2680 mutex_exit(hash_lock);
2683 * This block was freed while we waited for the read to
2684 * complete. It has been removed from the hash table and
2685 * moved to the anonymous state (so that it won't show up
2688 ASSERT3P(hdr->b_state, ==, arc_anon);
2689 freeable = refcount_is_zero(&hdr->b_refcnt);
2692 /* execute each callback and free its structure */
2693 while ((acb = callback_list) != NULL) {
2695 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2697 if (acb->acb_zio_dummy != NULL) {
2698 acb->acb_zio_dummy->io_error = zio->io_error;
2699 zio_nowait(acb->acb_zio_dummy);
2702 callback_list = acb->acb_next;
2703 kmem_free(acb, sizeof (arc_callback_t));
2707 arc_hdr_destroy(hdr);
2711 * "Read" the block block at the specified DVA (in bp) via the
2712 * cache. If the block is found in the cache, invoke the provided
2713 * callback immediately and return. Note that the `zio' parameter
2714 * in the callback will be NULL in this case, since no IO was
2715 * required. If the block is not in the cache pass the read request
2716 * on to the spa with a substitute callback function, so that the
2717 * requested block will be added to the cache.
2719 * If a read request arrives for a block that has a read in-progress,
2720 * either wait for the in-progress read to complete (and return the
2721 * results); or, if this is a read with a "done" func, add a record
2722 * to the read to invoke the "done" func when the read completes,
2723 * and return; or just return.
2725 * arc_read_done() will invoke all the requested "done" functions
2726 * for readers of this block.
2728 * Normal callers should use arc_read and pass the arc buffer and offset
2729 * for the bp. But if you know you don't need locking, you can use
2733 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_buf_t *pbuf,
2734 arc_done_func_t *done, void *private, int priority, int zio_flags,
2735 uint32_t *arc_flags, const zbookmark_t *zb)
2741 * XXX This happens from traverse callback funcs, for
2742 * the objset_phys_t block.
2744 return (arc_read_nolock(pio, spa, bp, done, private, priority,
2745 zio_flags, arc_flags, zb));
2748 ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt));
2749 ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size);
2750 rw_enter(&pbuf->b_data_lock, RW_READER);
2752 err = arc_read_nolock(pio, spa, bp, done, private, priority,
2753 zio_flags, arc_flags, zb);
2754 rw_exit(&pbuf->b_data_lock);
2760 arc_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bp,
2761 arc_done_func_t *done, void *private, int priority, int zio_flags,
2762 uint32_t *arc_flags, const zbookmark_t *zb)
2765 arc_buf_t *buf = NULL;
2766 kmutex_t *hash_lock;
2768 uint64_t guid = spa_guid(spa);
2771 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2773 if (hdr && hdr->b_datacnt > 0) {
2775 *arc_flags |= ARC_CACHED;
2777 if (HDR_IO_IN_PROGRESS(hdr)) {
2779 if (*arc_flags & ARC_WAIT) {
2780 cv_wait(&hdr->b_cv, hash_lock);
2781 mutex_exit(hash_lock);
2784 ASSERT(*arc_flags & ARC_NOWAIT);
2787 arc_callback_t *acb = NULL;
2789 acb = kmem_zalloc(sizeof (arc_callback_t),
2791 acb->acb_done = done;
2792 acb->acb_private = private;
2794 acb->acb_zio_dummy = zio_null(pio,
2795 spa, NULL, NULL, NULL, zio_flags);
2797 ASSERT(acb->acb_done != NULL);
2798 acb->acb_next = hdr->b_acb;
2800 add_reference(hdr, hash_lock, private);
2801 mutex_exit(hash_lock);
2804 mutex_exit(hash_lock);
2808 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2811 add_reference(hdr, hash_lock, private);
2813 * If this block is already in use, create a new
2814 * copy of the data so that we will be guaranteed
2815 * that arc_release() will always succeed.
2819 ASSERT(buf->b_data);
2820 if (HDR_BUF_AVAILABLE(hdr)) {
2821 ASSERT(buf->b_efunc == NULL);
2822 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2824 buf = arc_buf_clone(buf);
2827 } else if (*arc_flags & ARC_PREFETCH &&
2828 refcount_count(&hdr->b_refcnt) == 0) {
2829 hdr->b_flags |= ARC_PREFETCH;
2831 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2832 arc_access(hdr, hash_lock);
2833 if (*arc_flags & ARC_L2CACHE)
2834 hdr->b_flags |= ARC_L2CACHE;
2835 mutex_exit(hash_lock);
2836 ARCSTAT_BUMP(arcstat_hits);
2837 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2838 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2839 data, metadata, hits);
2842 done(NULL, buf, private);
2844 uint64_t size = BP_GET_LSIZE(bp);
2845 arc_callback_t *acb;
2848 boolean_t devw = B_FALSE;
2851 /* this block is not in the cache */
2852 arc_buf_hdr_t *exists;
2853 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2854 buf = arc_buf_alloc(spa, size, private, type);
2856 hdr->b_dva = *BP_IDENTITY(bp);
2857 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
2858 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2859 exists = buf_hash_insert(hdr, &hash_lock);
2861 /* somebody beat us to the hash insert */
2862 mutex_exit(hash_lock);
2863 buf_discard_identity(hdr);
2864 (void) arc_buf_remove_ref(buf, private);
2865 goto top; /* restart the IO request */
2867 /* if this is a prefetch, we don't have a reference */
2868 if (*arc_flags & ARC_PREFETCH) {
2869 (void) remove_reference(hdr, hash_lock,
2871 hdr->b_flags |= ARC_PREFETCH;
2873 if (*arc_flags & ARC_L2CACHE)
2874 hdr->b_flags |= ARC_L2CACHE;
2875 if (BP_GET_LEVEL(bp) > 0)
2876 hdr->b_flags |= ARC_INDIRECT;
2878 /* this block is in the ghost cache */
2879 ASSERT(GHOST_STATE(hdr->b_state));
2880 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2881 ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0);
2882 ASSERT(hdr->b_buf == NULL);
2884 /* if this is a prefetch, we don't have a reference */
2885 if (*arc_flags & ARC_PREFETCH)
2886 hdr->b_flags |= ARC_PREFETCH;
2888 add_reference(hdr, hash_lock, private);
2889 if (*arc_flags & ARC_L2CACHE)
2890 hdr->b_flags |= ARC_L2CACHE;
2891 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2894 buf->b_efunc = NULL;
2895 buf->b_private = NULL;
2898 ASSERT(hdr->b_datacnt == 0);
2900 arc_get_data_buf(buf);
2901 arc_access(hdr, hash_lock);
2904 ASSERT(!GHOST_STATE(hdr->b_state));
2906 acb = kmem_zalloc(sizeof (arc_callback_t), KM_PUSHPAGE);
2907 acb->acb_done = done;
2908 acb->acb_private = private;
2910 ASSERT(hdr->b_acb == NULL);
2912 hdr->b_flags |= ARC_IO_IN_PROGRESS;
2914 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
2915 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
2916 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
2917 addr = hdr->b_l2hdr->b_daddr;
2919 * Lock out device removal.
2921 if (vdev_is_dead(vd) ||
2922 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
2926 mutex_exit(hash_lock);
2928 ASSERT3U(hdr->b_size, ==, size);
2929 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
2930 uint64_t, size, zbookmark_t *, zb);
2931 ARCSTAT_BUMP(arcstat_misses);
2932 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2933 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2934 data, metadata, misses);
2936 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
2938 * Read from the L2ARC if the following are true:
2939 * 1. The L2ARC vdev was previously cached.
2940 * 2. This buffer still has L2ARC metadata.
2941 * 3. This buffer isn't currently writing to the L2ARC.
2942 * 4. The L2ARC entry wasn't evicted, which may
2943 * also have invalidated the vdev.
2944 * 5. This isn't prefetch and l2arc_noprefetch is set.
2946 if (hdr->b_l2hdr != NULL &&
2947 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
2948 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
2949 l2arc_read_callback_t *cb;
2951 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
2952 ARCSTAT_BUMP(arcstat_l2_hits);
2954 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
2956 cb->l2rcb_buf = buf;
2957 cb->l2rcb_spa = spa;
2960 cb->l2rcb_flags = zio_flags;
2963 * l2arc read. The SCL_L2ARC lock will be
2964 * released by l2arc_read_done().
2966 rzio = zio_read_phys(pio, vd, addr, size,
2967 buf->b_data, ZIO_CHECKSUM_OFF,
2968 l2arc_read_done, cb, priority, zio_flags |
2969 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
2970 ZIO_FLAG_DONT_PROPAGATE |
2971 ZIO_FLAG_DONT_RETRY, B_FALSE);
2972 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
2974 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
2976 if (*arc_flags & ARC_NOWAIT) {
2981 ASSERT(*arc_flags & ARC_WAIT);
2982 if (zio_wait(rzio) == 0)
2985 /* l2arc read error; goto zio_read() */
2987 DTRACE_PROBE1(l2arc__miss,
2988 arc_buf_hdr_t *, hdr);
2989 ARCSTAT_BUMP(arcstat_l2_misses);
2990 if (HDR_L2_WRITING(hdr))
2991 ARCSTAT_BUMP(arcstat_l2_rw_clash);
2992 spa_config_exit(spa, SCL_L2ARC, vd);
2996 spa_config_exit(spa, SCL_L2ARC, vd);
2997 if (l2arc_ndev != 0) {
2998 DTRACE_PROBE1(l2arc__miss,
2999 arc_buf_hdr_t *, hdr);
3000 ARCSTAT_BUMP(arcstat_l2_misses);
3004 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3005 arc_read_done, buf, priority, zio_flags, zb);
3007 if (*arc_flags & ARC_WAIT)
3008 return (zio_wait(rzio));
3010 ASSERT(*arc_flags & ARC_NOWAIT);
3017 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3019 ASSERT(buf->b_hdr != NULL);
3020 ASSERT(buf->b_hdr->b_state != arc_anon);
3021 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3022 ASSERT(buf->b_efunc == NULL);
3023 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3025 buf->b_efunc = func;
3026 buf->b_private = private;
3030 * This is used by the DMU to let the ARC know that a buffer is
3031 * being evicted, so the ARC should clean up. If this arc buf
3032 * is not yet in the evicted state, it will be put there.
3035 arc_buf_evict(arc_buf_t *buf)
3038 kmutex_t *hash_lock;
3041 mutex_enter(&buf->b_evict_lock);
3045 * We are in arc_do_user_evicts().
3047 ASSERT(buf->b_data == NULL);
3048 mutex_exit(&buf->b_evict_lock);
3050 } else if (buf->b_data == NULL) {
3051 arc_buf_t copy = *buf; /* structure assignment */
3053 * We are on the eviction list; process this buffer now
3054 * but let arc_do_user_evicts() do the reaping.
3056 buf->b_efunc = NULL;
3057 mutex_exit(&buf->b_evict_lock);
3058 VERIFY(copy.b_efunc(©) == 0);
3061 hash_lock = HDR_LOCK(hdr);
3062 mutex_enter(hash_lock);
3064 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3066 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3067 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3070 * Pull this buffer off of the hdr
3073 while (*bufp != buf)
3074 bufp = &(*bufp)->b_next;
3075 *bufp = buf->b_next;
3077 ASSERT(buf->b_data != NULL);
3078 arc_buf_destroy(buf, FALSE, FALSE);
3080 if (hdr->b_datacnt == 0) {
3081 arc_state_t *old_state = hdr->b_state;
3082 arc_state_t *evicted_state;
3084 ASSERT(hdr->b_buf == NULL);
3085 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3088 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3090 mutex_enter(&old_state->arcs_mtx);
3091 mutex_enter(&evicted_state->arcs_mtx);
3093 arc_change_state(evicted_state, hdr, hash_lock);
3094 ASSERT(HDR_IN_HASH_TABLE(hdr));
3095 hdr->b_flags |= ARC_IN_HASH_TABLE;
3096 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3098 mutex_exit(&evicted_state->arcs_mtx);
3099 mutex_exit(&old_state->arcs_mtx);
3101 mutex_exit(hash_lock);
3102 mutex_exit(&buf->b_evict_lock);
3104 VERIFY(buf->b_efunc(buf) == 0);
3105 buf->b_efunc = NULL;
3106 buf->b_private = NULL;
3109 kmem_cache_free(buf_cache, buf);
3114 * Release this buffer from the cache. This must be done
3115 * after a read and prior to modifying the buffer contents.
3116 * If the buffer has more than one reference, we must make
3117 * a new hdr for the buffer.
3120 arc_release(arc_buf_t *buf, void *tag)
3123 kmutex_t *hash_lock = NULL;
3124 l2arc_buf_hdr_t *l2hdr;
3125 uint64_t buf_size = 0;
3128 * It would be nice to assert that if it's DMU metadata (level >
3129 * 0 || it's the dnode file), then it must be syncing context.
3130 * But we don't know that information at this level.
3133 mutex_enter(&buf->b_evict_lock);
3136 /* this buffer is not on any list */
3137 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3139 if (hdr->b_state == arc_anon) {
3140 /* this buffer is already released */
3141 ASSERT(buf->b_efunc == NULL);
3143 hash_lock = HDR_LOCK(hdr);
3144 mutex_enter(hash_lock);
3146 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3149 l2hdr = hdr->b_l2hdr;
3151 mutex_enter(&l2arc_buflist_mtx);
3152 hdr->b_l2hdr = NULL;
3153 buf_size = hdr->b_size;
3157 * Do we have more than one buf?
3159 if (hdr->b_datacnt > 1) {
3160 arc_buf_hdr_t *nhdr;
3162 uint64_t blksz = hdr->b_size;
3163 uint64_t spa = hdr->b_spa;
3164 arc_buf_contents_t type = hdr->b_type;
3165 uint32_t flags = hdr->b_flags;
3167 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3169 * Pull the data off of this hdr and attach it to
3170 * a new anonymous hdr.
3172 (void) remove_reference(hdr, hash_lock, tag);
3174 while (*bufp != buf)
3175 bufp = &(*bufp)->b_next;
3176 *bufp = buf->b_next;
3179 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3180 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3181 if (refcount_is_zero(&hdr->b_refcnt)) {
3182 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3183 ASSERT3U(*size, >=, hdr->b_size);
3184 atomic_add_64(size, -hdr->b_size);
3186 hdr->b_datacnt -= 1;
3187 arc_cksum_verify(buf);
3189 mutex_exit(hash_lock);
3191 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3192 nhdr->b_size = blksz;
3194 nhdr->b_type = type;
3196 nhdr->b_state = arc_anon;
3197 nhdr->b_arc_access = 0;
3198 nhdr->b_flags = flags & ARC_L2_WRITING;
3199 nhdr->b_l2hdr = NULL;
3200 nhdr->b_datacnt = 1;
3201 nhdr->b_freeze_cksum = NULL;
3202 (void) refcount_add(&nhdr->b_refcnt, tag);
3204 mutex_exit(&buf->b_evict_lock);
3205 atomic_add_64(&arc_anon->arcs_size, blksz);
3207 mutex_exit(&buf->b_evict_lock);
3208 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3209 ASSERT(!list_link_active(&hdr->b_arc_node));
3210 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3211 if (hdr->b_state != arc_anon)
3212 arc_change_state(arc_anon, hdr, hash_lock);
3213 hdr->b_arc_access = 0;
3215 mutex_exit(hash_lock);
3217 buf_discard_identity(hdr);
3220 buf->b_efunc = NULL;
3221 buf->b_private = NULL;
3224 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3225 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3226 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3227 mutex_exit(&l2arc_buflist_mtx);
3232 * Release this buffer. If it does not match the provided BP, fill it
3233 * with that block's contents.
3237 arc_release_bp(arc_buf_t *buf, void *tag, blkptr_t *bp, spa_t *spa,
3240 arc_release(buf, tag);
3245 arc_released(arc_buf_t *buf)
3249 mutex_enter(&buf->b_evict_lock);
3250 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3251 mutex_exit(&buf->b_evict_lock);
3256 arc_has_callback(arc_buf_t *buf)
3260 mutex_enter(&buf->b_evict_lock);
3261 callback = (buf->b_efunc != NULL);
3262 mutex_exit(&buf->b_evict_lock);
3268 arc_referenced(arc_buf_t *buf)
3272 mutex_enter(&buf->b_evict_lock);
3273 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3274 mutex_exit(&buf->b_evict_lock);
3275 return (referenced);
3280 arc_write_ready(zio_t *zio)
3282 arc_write_callback_t *callback = zio->io_private;
3283 arc_buf_t *buf = callback->awcb_buf;
3284 arc_buf_hdr_t *hdr = buf->b_hdr;
3286 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3287 callback->awcb_ready(zio, buf, callback->awcb_private);
3290 * If the IO is already in progress, then this is a re-write
3291 * attempt, so we need to thaw and re-compute the cksum.
3292 * It is the responsibility of the callback to handle the
3293 * accounting for any re-write attempt.
3295 if (HDR_IO_IN_PROGRESS(hdr)) {
3296 mutex_enter(&hdr->b_freeze_lock);
3297 if (hdr->b_freeze_cksum != NULL) {
3298 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3299 hdr->b_freeze_cksum = NULL;
3301 mutex_exit(&hdr->b_freeze_lock);
3303 arc_cksum_compute(buf, B_FALSE);
3304 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3308 arc_write_done(zio_t *zio)
3310 arc_write_callback_t *callback = zio->io_private;
3311 arc_buf_t *buf = callback->awcb_buf;
3312 arc_buf_hdr_t *hdr = buf->b_hdr;
3314 ASSERT(hdr->b_acb == NULL);
3316 if (zio->io_error == 0) {
3317 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3318 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3319 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3321 ASSERT(BUF_EMPTY(hdr));
3325 * If the block to be written was all-zero, we may have
3326 * compressed it away. In this case no write was performed
3327 * so there will be no dva/birth/checksum. The buffer must
3328 * therefore remain anonymous (and uncached).
3330 if (!BUF_EMPTY(hdr)) {
3331 arc_buf_hdr_t *exists;
3332 kmutex_t *hash_lock;
3334 ASSERT(zio->io_error == 0);
3336 arc_cksum_verify(buf);
3338 exists = buf_hash_insert(hdr, &hash_lock);
3341 * This can only happen if we overwrite for
3342 * sync-to-convergence, because we remove
3343 * buffers from the hash table when we arc_free().
3345 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3346 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3347 panic("bad overwrite, hdr=%p exists=%p",
3348 (void *)hdr, (void *)exists);
3349 ASSERT(refcount_is_zero(&exists->b_refcnt));
3350 arc_change_state(arc_anon, exists, hash_lock);
3351 mutex_exit(hash_lock);
3352 arc_hdr_destroy(exists);
3353 exists = buf_hash_insert(hdr, &hash_lock);
3354 ASSERT3P(exists, ==, NULL);
3357 ASSERT(hdr->b_datacnt == 1);
3358 ASSERT(hdr->b_state == arc_anon);
3359 ASSERT(BP_GET_DEDUP(zio->io_bp));
3360 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3363 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3364 /* if it's not anon, we are doing a scrub */
3365 if (!exists && hdr->b_state == arc_anon)
3366 arc_access(hdr, hash_lock);
3367 mutex_exit(hash_lock);
3369 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3372 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3373 callback->awcb_done(zio, buf, callback->awcb_private);
3375 kmem_free(callback, sizeof (arc_write_callback_t));
3379 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3380 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3381 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3382 int priority, int zio_flags, const zbookmark_t *zb)
3384 arc_buf_hdr_t *hdr = buf->b_hdr;
3385 arc_write_callback_t *callback;
3388 ASSERT(ready != NULL);
3389 ASSERT(done != NULL);
3390 ASSERT(!HDR_IO_ERROR(hdr));
3391 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3392 ASSERT(hdr->b_acb == NULL);
3394 hdr->b_flags |= ARC_L2CACHE;
3395 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3396 callback->awcb_ready = ready;
3397 callback->awcb_done = done;
3398 callback->awcb_private = private;
3399 callback->awcb_buf = buf;
3401 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3402 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3408 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3411 uint64_t available_memory = ptob(freemem);
3412 static uint64_t page_load = 0;
3413 static uint64_t last_txg = 0;
3417 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3419 if (available_memory >= zfs_write_limit_max)
3422 if (txg > last_txg) {
3427 * If we are in pageout, we know that memory is already tight,
3428 * the arc is already going to be evicting, so we just want to
3429 * continue to let page writes occur as quickly as possible.
3431 if (curproc == proc_pageout) {
3432 if (page_load > MAX(ptob(minfree), available_memory) / 4)
3434 /* Note: reserve is inflated, so we deflate */
3435 page_load += reserve / 8;
3437 } else if (page_load > 0 && arc_reclaim_needed()) {
3438 /* memory is low, delay before restarting */
3439 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3444 if (arc_size > arc_c_min) {
3445 uint64_t evictable_memory =
3446 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3447 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3448 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3449 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3450 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3453 if (inflight_data > available_memory / 4) {
3454 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3462 arc_tempreserve_clear(uint64_t reserve)
3464 atomic_add_64(&arc_tempreserve, -reserve);
3465 ASSERT((int64_t)arc_tempreserve >= 0);
3469 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3476 * Once in a while, fail for no reason. Everything should cope.
3478 if (spa_get_random(10000) == 0) {
3479 dprintf("forcing random failure\n");
3483 if (reserve > arc_c/4 && !arc_no_grow)
3484 arc_c = MIN(arc_c_max, reserve * 4);
3485 if (reserve > arc_c)
3489 * Don't count loaned bufs as in flight dirty data to prevent long
3490 * network delays from blocking transactions that are ready to be
3491 * assigned to a txg.
3493 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3496 * Writes will, almost always, require additional memory allocations
3497 * in order to compress/encrypt/etc the data. We therefor need to
3498 * make sure that there is sufficient available memory for this.
3500 if ((error = arc_memory_throttle(reserve, anon_size, txg)))
3504 * Throttle writes when the amount of dirty data in the cache
3505 * gets too large. We try to keep the cache less than half full
3506 * of dirty blocks so that our sync times don't grow too large.
3507 * Note: if two requests come in concurrently, we might let them
3508 * both succeed, when one of them should fail. Not a huge deal.
3511 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3512 anon_size > arc_c / 4) {
3513 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3514 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3515 arc_tempreserve>>10,
3516 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3517 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3518 reserve>>10, arc_c>>10);
3521 atomic_add_64(&arc_tempreserve, reserve);
3528 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3529 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3531 /* Convert seconds to clock ticks */
3532 arc_min_prefetch_lifespan = 1 * hz;
3534 /* Start out with 1/8 of all memory */
3535 arc_c = physmem * PAGESIZE / 8;
3539 * On architectures where the physical memory can be larger
3540 * than the addressable space (intel in 32-bit mode), we may
3541 * need to limit the cache to 1/8 of VM size.
3543 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3545 * Register a shrinker to support synchronous (direct) memory
3546 * reclaim from the arc. This is done to prevent kswapd from
3547 * swapping out pages when it is preferable to shrink the arc.
3549 spl_register_shrinker(&arc_shrinker);
3552 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3553 arc_c_min = MAX(arc_c / 4, 64<<20);
3554 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3555 if (arc_c * 8 >= 1<<30)
3556 arc_c_max = (arc_c * 8) - (1<<30);
3558 arc_c_max = arc_c_min;
3559 arc_c_max = MAX(arc_c * 6, arc_c_max);
3562 * Allow the tunables to override our calculations if they are
3563 * reasonable (ie. over 64MB)
3565 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3566 arc_c_max = zfs_arc_max;
3567 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3568 arc_c_min = zfs_arc_min;
3571 arc_p = (arc_c >> 1);
3573 /* limit meta-data to 1/4 of the arc capacity */
3574 arc_meta_limit = arc_c_max / 4;
3577 /* Allow the tunable to override if it is reasonable */
3578 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3579 arc_meta_limit = zfs_arc_meta_limit;
3581 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3582 arc_c_min = arc_meta_limit / 2;
3584 if (zfs_arc_grow_retry > 0)
3585 arc_grow_retry = zfs_arc_grow_retry;
3587 if (zfs_arc_shrink_shift > 0)
3588 arc_shrink_shift = zfs_arc_shrink_shift;
3590 if (zfs_arc_p_min_shift > 0)
3591 arc_p_min_shift = zfs_arc_p_min_shift;
3593 if (zfs_arc_reduce_dnlc_percent > 0)
3594 arc_reduce_dnlc_percent = zfs_arc_reduce_dnlc_percent;
3596 /* if kmem_flags are set, lets try to use less memory */
3597 if (kmem_debugging())
3599 if (arc_c < arc_c_min)
3602 arc_anon = &ARC_anon;
3604 arc_mru_ghost = &ARC_mru_ghost;
3606 arc_mfu_ghost = &ARC_mfu_ghost;
3607 arc_l2c_only = &ARC_l2c_only;
3610 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3611 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3612 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3613 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3614 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3615 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3617 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3618 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3619 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3620 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3621 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3622 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3623 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3624 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3625 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3626 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3627 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3628 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3629 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3630 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3631 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3632 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3633 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3634 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3635 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3636 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3640 arc_thread_exit = 0;
3641 arc_eviction_list = NULL;
3642 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3643 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3645 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3646 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3648 if (arc_ksp != NULL) {
3649 arc_ksp->ks_data = &arc_stats;
3650 kstat_install(arc_ksp);
3653 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3654 TS_RUN, minclsyspri);
3659 if (zfs_write_limit_max == 0)
3660 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3662 zfs_write_limit_shift = 0;
3663 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3669 mutex_enter(&arc_reclaim_thr_lock);
3671 spl_unregister_shrinker(&arc_shrinker);
3672 #endif /* _KERNEL */
3674 arc_thread_exit = 1;
3675 while (arc_thread_exit != 0)
3676 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3677 mutex_exit(&arc_reclaim_thr_lock);
3683 if (arc_ksp != NULL) {
3684 kstat_delete(arc_ksp);
3688 mutex_destroy(&arc_eviction_mtx);
3689 mutex_destroy(&arc_reclaim_thr_lock);
3690 cv_destroy(&arc_reclaim_thr_cv);
3692 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3693 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3694 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3695 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3696 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3697 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3698 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3699 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3701 mutex_destroy(&arc_anon->arcs_mtx);
3702 mutex_destroy(&arc_mru->arcs_mtx);
3703 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3704 mutex_destroy(&arc_mfu->arcs_mtx);
3705 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3706 mutex_destroy(&arc_l2c_only->arcs_mtx);
3708 mutex_destroy(&zfs_write_limit_lock);
3712 ASSERT(arc_loaned_bytes == 0);
3718 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3719 * It uses dedicated storage devices to hold cached data, which are populated
3720 * using large infrequent writes. The main role of this cache is to boost
3721 * the performance of random read workloads. The intended L2ARC devices
3722 * include short-stroked disks, solid state disks, and other media with
3723 * substantially faster read latency than disk.
3725 * +-----------------------+
3727 * +-----------------------+
3730 * l2arc_feed_thread() arc_read()
3734 * +---------------+ |
3736 * +---------------+ |
3741 * +-------+ +-------+
3743 * | cache | | cache |
3744 * +-------+ +-------+
3745 * +=========+ .-----.
3746 * : L2ARC : |-_____-|
3747 * : devices : | Disks |
3748 * +=========+ `-_____-'
3750 * Read requests are satisfied from the following sources, in order:
3753 * 2) vdev cache of L2ARC devices
3755 * 4) vdev cache of disks
3758 * Some L2ARC device types exhibit extremely slow write performance.
3759 * To accommodate for this there are some significant differences between
3760 * the L2ARC and traditional cache design:
3762 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3763 * the ARC behave as usual, freeing buffers and placing headers on ghost
3764 * lists. The ARC does not send buffers to the L2ARC during eviction as
3765 * this would add inflated write latencies for all ARC memory pressure.
3767 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3768 * It does this by periodically scanning buffers from the eviction-end of
3769 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3770 * not already there. It scans until a headroom of buffers is satisfied,
3771 * which itself is a buffer for ARC eviction. The thread that does this is
3772 * l2arc_feed_thread(), illustrated below; example sizes are included to
3773 * provide a better sense of ratio than this diagram:
3776 * +---------------------+----------+
3777 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3778 * +---------------------+----------+ | o L2ARC eligible
3779 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
3780 * +---------------------+----------+ |
3781 * 15.9 Gbytes ^ 32 Mbytes |
3783 * l2arc_feed_thread()
3785 * l2arc write hand <--[oooo]--'
3789 * +==============================+
3790 * L2ARC dev |####|#|###|###| |####| ... |
3791 * +==============================+
3794 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
3795 * evicted, then the L2ARC has cached a buffer much sooner than it probably
3796 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
3797 * safe to say that this is an uncommon case, since buffers at the end of
3798 * the ARC lists have moved there due to inactivity.
3800 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
3801 * then the L2ARC simply misses copying some buffers. This serves as a
3802 * pressure valve to prevent heavy read workloads from both stalling the ARC
3803 * with waits and clogging the L2ARC with writes. This also helps prevent
3804 * the potential for the L2ARC to churn if it attempts to cache content too
3805 * quickly, such as during backups of the entire pool.
3807 * 5. After system boot and before the ARC has filled main memory, there are
3808 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
3809 * lists can remain mostly static. Instead of searching from tail of these
3810 * lists as pictured, the l2arc_feed_thread() will search from the list heads
3811 * for eligible buffers, greatly increasing its chance of finding them.
3813 * The L2ARC device write speed is also boosted during this time so that
3814 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
3815 * there are no L2ARC reads, and no fear of degrading read performance
3816 * through increased writes.
3818 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
3819 * the vdev queue can aggregate them into larger and fewer writes. Each
3820 * device is written to in a rotor fashion, sweeping writes through
3821 * available space then repeating.
3823 * 7. The L2ARC does not store dirty content. It never needs to flush
3824 * write buffers back to disk based storage.
3826 * 8. If an ARC buffer is written (and dirtied) which also exists in the
3827 * L2ARC, the now stale L2ARC buffer is immediately dropped.
3829 * The performance of the L2ARC can be tweaked by a number of tunables, which
3830 * may be necessary for different workloads:
3832 * l2arc_write_max max write bytes per interval
3833 * l2arc_write_boost extra write bytes during device warmup
3834 * l2arc_noprefetch skip caching prefetched buffers
3835 * l2arc_headroom number of max device writes to precache
3836 * l2arc_feed_secs seconds between L2ARC writing
3838 * Tunables may be removed or added as future performance improvements are
3839 * integrated, and also may become zpool properties.
3841 * There are three key functions that control how the L2ARC warms up:
3843 * l2arc_write_eligible() check if a buffer is eligible to cache
3844 * l2arc_write_size() calculate how much to write
3845 * l2arc_write_interval() calculate sleep delay between writes
3847 * These three functions determine what to write, how much, and how quickly
3852 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
3855 * A buffer is *not* eligible for the L2ARC if it:
3856 * 1. belongs to a different spa.
3857 * 2. is already cached on the L2ARC.
3858 * 3. has an I/O in progress (it may be an incomplete read).
3859 * 4. is flagged not eligible (zfs property).
3861 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
3862 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
3869 l2arc_write_size(l2arc_dev_t *dev)
3873 size = dev->l2ad_write;
3875 if (arc_warm == B_FALSE)
3876 size += dev->l2ad_boost;
3883 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
3885 clock_t interval, next, now;
3888 * If the ARC lists are busy, increase our write rate; if the
3889 * lists are stale, idle back. This is achieved by checking
3890 * how much we previously wrote - if it was more than half of
3891 * what we wanted, schedule the next write much sooner.
3893 if (l2arc_feed_again && wrote > (wanted / 2))
3894 interval = (hz * l2arc_feed_min_ms) / 1000;
3896 interval = hz * l2arc_feed_secs;
3898 now = ddi_get_lbolt();
3899 next = MAX(now, MIN(now + interval, began + interval));
3905 l2arc_hdr_stat_add(void)
3907 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
3908 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
3912 l2arc_hdr_stat_remove(void)
3914 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
3915 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
3919 * Cycle through L2ARC devices. This is how L2ARC load balances.
3920 * If a device is returned, this also returns holding the spa config lock.
3922 static l2arc_dev_t *
3923 l2arc_dev_get_next(void)
3925 l2arc_dev_t *first, *next = NULL;
3928 * Lock out the removal of spas (spa_namespace_lock), then removal
3929 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
3930 * both locks will be dropped and a spa config lock held instead.
3932 mutex_enter(&spa_namespace_lock);
3933 mutex_enter(&l2arc_dev_mtx);
3935 /* if there are no vdevs, there is nothing to do */
3936 if (l2arc_ndev == 0)
3940 next = l2arc_dev_last;
3942 /* loop around the list looking for a non-faulted vdev */
3944 next = list_head(l2arc_dev_list);
3946 next = list_next(l2arc_dev_list, next);
3948 next = list_head(l2arc_dev_list);
3951 /* if we have come back to the start, bail out */
3954 else if (next == first)
3957 } while (vdev_is_dead(next->l2ad_vdev));
3959 /* if we were unable to find any usable vdevs, return NULL */
3960 if (vdev_is_dead(next->l2ad_vdev))
3963 l2arc_dev_last = next;
3966 mutex_exit(&l2arc_dev_mtx);
3969 * Grab the config lock to prevent the 'next' device from being
3970 * removed while we are writing to it.
3973 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
3974 mutex_exit(&spa_namespace_lock);
3980 * Free buffers that were tagged for destruction.
3983 l2arc_do_free_on_write(void)
3986 l2arc_data_free_t *df, *df_prev;
3988 mutex_enter(&l2arc_free_on_write_mtx);
3989 buflist = l2arc_free_on_write;
3991 for (df = list_tail(buflist); df; df = df_prev) {
3992 df_prev = list_prev(buflist, df);
3993 ASSERT(df->l2df_data != NULL);
3994 ASSERT(df->l2df_func != NULL);
3995 df->l2df_func(df->l2df_data, df->l2df_size);
3996 list_remove(buflist, df);
3997 kmem_free(df, sizeof (l2arc_data_free_t));
4000 mutex_exit(&l2arc_free_on_write_mtx);
4004 * A write to a cache device has completed. Update all headers to allow
4005 * reads from these buffers to begin.
4008 l2arc_write_done(zio_t *zio)
4010 l2arc_write_callback_t *cb;
4013 arc_buf_hdr_t *head, *ab, *ab_prev;
4014 l2arc_buf_hdr_t *abl2;
4015 kmutex_t *hash_lock;
4017 cb = zio->io_private;
4019 dev = cb->l2wcb_dev;
4020 ASSERT(dev != NULL);
4021 head = cb->l2wcb_head;
4022 ASSERT(head != NULL);
4023 buflist = dev->l2ad_buflist;
4024 ASSERT(buflist != NULL);
4025 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4026 l2arc_write_callback_t *, cb);
4028 if (zio->io_error != 0)
4029 ARCSTAT_BUMP(arcstat_l2_writes_error);
4031 mutex_enter(&l2arc_buflist_mtx);
4034 * All writes completed, or an error was hit.
4036 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4037 ab_prev = list_prev(buflist, ab);
4039 hash_lock = HDR_LOCK(ab);
4040 if (!mutex_tryenter(hash_lock)) {
4042 * This buffer misses out. It may be in a stage
4043 * of eviction. Its ARC_L2_WRITING flag will be
4044 * left set, denying reads to this buffer.
4046 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4050 if (zio->io_error != 0) {
4052 * Error - drop L2ARC entry.
4054 list_remove(buflist, ab);
4057 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4058 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4062 * Allow ARC to begin reads to this L2ARC entry.
4064 ab->b_flags &= ~ARC_L2_WRITING;
4066 mutex_exit(hash_lock);
4069 atomic_inc_64(&l2arc_writes_done);
4070 list_remove(buflist, head);
4071 kmem_cache_free(hdr_cache, head);
4072 mutex_exit(&l2arc_buflist_mtx);
4074 l2arc_do_free_on_write();
4076 kmem_free(cb, sizeof (l2arc_write_callback_t));
4080 * A read to a cache device completed. Validate buffer contents before
4081 * handing over to the regular ARC routines.
4084 l2arc_read_done(zio_t *zio)
4086 l2arc_read_callback_t *cb;
4089 kmutex_t *hash_lock;
4092 ASSERT(zio->io_vd != NULL);
4093 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4095 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4097 cb = zio->io_private;
4099 buf = cb->l2rcb_buf;
4100 ASSERT(buf != NULL);
4102 hash_lock = HDR_LOCK(buf->b_hdr);
4103 mutex_enter(hash_lock);
4105 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4108 * Check this survived the L2ARC journey.
4110 equal = arc_cksum_equal(buf);
4111 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4112 mutex_exit(hash_lock);
4113 zio->io_private = buf;
4114 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4115 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4118 mutex_exit(hash_lock);
4120 * Buffer didn't survive caching. Increment stats and
4121 * reissue to the original storage device.
4123 if (zio->io_error != 0) {
4124 ARCSTAT_BUMP(arcstat_l2_io_error);
4126 zio->io_error = EIO;
4129 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4132 * If there's no waiter, issue an async i/o to the primary
4133 * storage now. If there *is* a waiter, the caller must
4134 * issue the i/o in a context where it's OK to block.
4136 if (zio->io_waiter == NULL) {
4137 zio_t *pio = zio_unique_parent(zio);
4139 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4141 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4142 buf->b_data, zio->io_size, arc_read_done, buf,
4143 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4147 kmem_free(cb, sizeof (l2arc_read_callback_t));
4151 * This is the list priority from which the L2ARC will search for pages to
4152 * cache. This is used within loops (0..3) to cycle through lists in the
4153 * desired order. This order can have a significant effect on cache
4156 * Currently the metadata lists are hit first, MFU then MRU, followed by
4157 * the data lists. This function returns a locked list, and also returns
4161 l2arc_list_locked(int list_num, kmutex_t **lock)
4163 list_t *list = NULL;
4165 ASSERT(list_num >= 0 && list_num <= 3);
4169 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4170 *lock = &arc_mfu->arcs_mtx;
4173 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4174 *lock = &arc_mru->arcs_mtx;
4177 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4178 *lock = &arc_mfu->arcs_mtx;
4181 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4182 *lock = &arc_mru->arcs_mtx;
4186 ASSERT(!(MUTEX_HELD(*lock)));
4192 * Evict buffers from the device write hand to the distance specified in
4193 * bytes. This distance may span populated buffers, it may span nothing.
4194 * This is clearing a region on the L2ARC device ready for writing.
4195 * If the 'all' boolean is set, every buffer is evicted.
4198 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4201 l2arc_buf_hdr_t *abl2;
4202 arc_buf_hdr_t *ab, *ab_prev;
4203 kmutex_t *hash_lock;
4206 buflist = dev->l2ad_buflist;
4208 if (buflist == NULL)
4211 if (!all && dev->l2ad_first) {
4213 * This is the first sweep through the device. There is
4219 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4221 * When nearing the end of the device, evict to the end
4222 * before the device write hand jumps to the start.
4224 taddr = dev->l2ad_end;
4226 taddr = dev->l2ad_hand + distance;
4228 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4229 uint64_t, taddr, boolean_t, all);
4232 mutex_enter(&l2arc_buflist_mtx);
4233 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4234 ab_prev = list_prev(buflist, ab);
4236 hash_lock = HDR_LOCK(ab);
4237 if (!mutex_tryenter(hash_lock)) {
4239 * Missed the hash lock. Retry.
4241 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4242 mutex_exit(&l2arc_buflist_mtx);
4243 mutex_enter(hash_lock);
4244 mutex_exit(hash_lock);
4248 if (HDR_L2_WRITE_HEAD(ab)) {
4250 * We hit a write head node. Leave it for
4251 * l2arc_write_done().
4253 list_remove(buflist, ab);
4254 mutex_exit(hash_lock);
4258 if (!all && ab->b_l2hdr != NULL &&
4259 (ab->b_l2hdr->b_daddr > taddr ||
4260 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4262 * We've evicted to the target address,
4263 * or the end of the device.
4265 mutex_exit(hash_lock);
4269 if (HDR_FREE_IN_PROGRESS(ab)) {
4271 * Already on the path to destruction.
4273 mutex_exit(hash_lock);
4277 if (ab->b_state == arc_l2c_only) {
4278 ASSERT(!HDR_L2_READING(ab));
4280 * This doesn't exist in the ARC. Destroy.
4281 * arc_hdr_destroy() will call list_remove()
4282 * and decrement arcstat_l2_size.
4284 arc_change_state(arc_anon, ab, hash_lock);
4285 arc_hdr_destroy(ab);
4288 * Invalidate issued or about to be issued
4289 * reads, since we may be about to write
4290 * over this location.
4292 if (HDR_L2_READING(ab)) {
4293 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4294 ab->b_flags |= ARC_L2_EVICTED;
4298 * Tell ARC this no longer exists in L2ARC.
4300 if (ab->b_l2hdr != NULL) {
4303 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4304 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4306 list_remove(buflist, ab);
4309 * This may have been leftover after a
4312 ab->b_flags &= ~ARC_L2_WRITING;
4314 mutex_exit(hash_lock);
4316 mutex_exit(&l2arc_buflist_mtx);
4318 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4319 dev->l2ad_evict = taddr;
4323 * Find and write ARC buffers to the L2ARC device.
4325 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4326 * for reading until they have completed writing.
4329 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4331 arc_buf_hdr_t *ab, *ab_prev, *head;
4332 l2arc_buf_hdr_t *hdrl2;
4334 uint64_t passed_sz, write_sz, buf_sz, headroom;
4336 kmutex_t *hash_lock, *list_lock = NULL;
4337 boolean_t have_lock, full;
4338 l2arc_write_callback_t *cb;
4340 uint64_t guid = spa_guid(spa);
4343 ASSERT(dev->l2ad_vdev != NULL);
4348 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4349 head->b_flags |= ARC_L2_WRITE_HEAD;
4352 * Copy buffers for L2ARC writing.
4354 mutex_enter(&l2arc_buflist_mtx);
4355 for (try = 0; try <= 3; try++) {
4356 list = l2arc_list_locked(try, &list_lock);
4360 * L2ARC fast warmup.
4362 * Until the ARC is warm and starts to evict, read from the
4363 * head of the ARC lists rather than the tail.
4365 headroom = target_sz * l2arc_headroom;
4366 if (arc_warm == B_FALSE)
4367 ab = list_head(list);
4369 ab = list_tail(list);
4371 for (; ab; ab = ab_prev) {
4372 if (arc_warm == B_FALSE)
4373 ab_prev = list_next(list, ab);
4375 ab_prev = list_prev(list, ab);
4377 hash_lock = HDR_LOCK(ab);
4378 have_lock = MUTEX_HELD(hash_lock);
4379 if (!have_lock && !mutex_tryenter(hash_lock)) {
4381 * Skip this buffer rather than waiting.
4386 passed_sz += ab->b_size;
4387 if (passed_sz > headroom) {
4391 mutex_exit(hash_lock);
4395 if (!l2arc_write_eligible(guid, ab)) {
4396 mutex_exit(hash_lock);
4400 if ((write_sz + ab->b_size) > target_sz) {
4402 mutex_exit(hash_lock);
4408 * Insert a dummy header on the buflist so
4409 * l2arc_write_done() can find where the
4410 * write buffers begin without searching.
4412 list_insert_head(dev->l2ad_buflist, head);
4415 sizeof (l2arc_write_callback_t), KM_SLEEP);
4416 cb->l2wcb_dev = dev;
4417 cb->l2wcb_head = head;
4418 pio = zio_root(spa, l2arc_write_done, cb,
4423 * Create and add a new L2ARC header.
4425 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4427 hdrl2->b_daddr = dev->l2ad_hand;
4429 ab->b_flags |= ARC_L2_WRITING;
4430 ab->b_l2hdr = hdrl2;
4431 list_insert_head(dev->l2ad_buflist, ab);
4432 buf_data = ab->b_buf->b_data;
4433 buf_sz = ab->b_size;
4436 * Compute and store the buffer cksum before
4437 * writing. On debug the cksum is verified first.
4439 arc_cksum_verify(ab->b_buf);
4440 arc_cksum_compute(ab->b_buf, B_TRUE);
4442 mutex_exit(hash_lock);
4444 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4445 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4446 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4447 ZIO_FLAG_CANFAIL, B_FALSE);
4449 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4451 (void) zio_nowait(wzio);
4454 * Keep the clock hand suitably device-aligned.
4456 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4459 dev->l2ad_hand += buf_sz;
4462 mutex_exit(list_lock);
4467 mutex_exit(&l2arc_buflist_mtx);
4470 ASSERT3U(write_sz, ==, 0);
4471 kmem_cache_free(hdr_cache, head);
4475 ASSERT3U(write_sz, <=, target_sz);
4476 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4477 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4478 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4479 vdev_space_update(dev->l2ad_vdev, write_sz, 0, 0);
4482 * Bump device hand to the device start if it is approaching the end.
4483 * l2arc_evict() will already have evicted ahead for this case.
4485 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4486 vdev_space_update(dev->l2ad_vdev,
4487 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4488 dev->l2ad_hand = dev->l2ad_start;
4489 dev->l2ad_evict = dev->l2ad_start;
4490 dev->l2ad_first = B_FALSE;
4493 dev->l2ad_writing = B_TRUE;
4494 (void) zio_wait(pio);
4495 dev->l2ad_writing = B_FALSE;
4501 * This thread feeds the L2ARC at regular intervals. This is the beating
4502 * heart of the L2ARC.
4505 l2arc_feed_thread(void)
4510 uint64_t size, wrote;
4511 clock_t begin, next = ddi_get_lbolt();
4513 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4515 mutex_enter(&l2arc_feed_thr_lock);
4517 while (l2arc_thread_exit == 0) {
4518 CALLB_CPR_SAFE_BEGIN(&cpr);
4519 (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv,
4520 &l2arc_feed_thr_lock, next);
4521 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4522 next = ddi_get_lbolt() + hz;
4525 * Quick check for L2ARC devices.
4527 mutex_enter(&l2arc_dev_mtx);
4528 if (l2arc_ndev == 0) {
4529 mutex_exit(&l2arc_dev_mtx);
4532 mutex_exit(&l2arc_dev_mtx);
4533 begin = ddi_get_lbolt();
4536 * This selects the next l2arc device to write to, and in
4537 * doing so the next spa to feed from: dev->l2ad_spa. This
4538 * will return NULL if there are now no l2arc devices or if
4539 * they are all faulted.
4541 * If a device is returned, its spa's config lock is also
4542 * held to prevent device removal. l2arc_dev_get_next()
4543 * will grab and release l2arc_dev_mtx.
4545 if ((dev = l2arc_dev_get_next()) == NULL)
4548 spa = dev->l2ad_spa;
4549 ASSERT(spa != NULL);
4552 * If the pool is read-only then force the feed thread to
4553 * sleep a little longer.
4555 if (!spa_writeable(spa)) {
4556 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4557 spa_config_exit(spa, SCL_L2ARC, dev);
4562 * Avoid contributing to memory pressure.
4564 if (arc_reclaim_needed()) {
4565 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4566 spa_config_exit(spa, SCL_L2ARC, dev);
4570 ARCSTAT_BUMP(arcstat_l2_feeds);
4572 size = l2arc_write_size(dev);
4575 * Evict L2ARC buffers that will be overwritten.
4577 l2arc_evict(dev, size, B_FALSE);
4580 * Write ARC buffers.
4582 wrote = l2arc_write_buffers(spa, dev, size);
4585 * Calculate interval between writes.
4587 next = l2arc_write_interval(begin, size, wrote);
4588 spa_config_exit(spa, SCL_L2ARC, dev);
4591 l2arc_thread_exit = 0;
4592 cv_broadcast(&l2arc_feed_thr_cv);
4593 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4598 l2arc_vdev_present(vdev_t *vd)
4602 mutex_enter(&l2arc_dev_mtx);
4603 for (dev = list_head(l2arc_dev_list); dev != NULL;
4604 dev = list_next(l2arc_dev_list, dev)) {
4605 if (dev->l2ad_vdev == vd)
4608 mutex_exit(&l2arc_dev_mtx);
4610 return (dev != NULL);
4614 * Add a vdev for use by the L2ARC. By this point the spa has already
4615 * validated the vdev and opened it.
4618 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
4620 l2arc_dev_t *adddev;
4622 ASSERT(!l2arc_vdev_present(vd));
4625 * Create a new l2arc device entry.
4627 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4628 adddev->l2ad_spa = spa;
4629 adddev->l2ad_vdev = vd;
4630 adddev->l2ad_write = l2arc_write_max;
4631 adddev->l2ad_boost = l2arc_write_boost;
4632 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
4633 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
4634 adddev->l2ad_hand = adddev->l2ad_start;
4635 adddev->l2ad_evict = adddev->l2ad_start;
4636 adddev->l2ad_first = B_TRUE;
4637 adddev->l2ad_writing = B_FALSE;
4638 list_link_init(&adddev->l2ad_node);
4639 ASSERT3U(adddev->l2ad_write, >, 0);
4642 * This is a list of all ARC buffers that are still valid on the
4645 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4646 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4647 offsetof(arc_buf_hdr_t, b_l2node));
4649 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
4652 * Add device to global list
4654 mutex_enter(&l2arc_dev_mtx);
4655 list_insert_head(l2arc_dev_list, adddev);
4656 atomic_inc_64(&l2arc_ndev);
4657 mutex_exit(&l2arc_dev_mtx);
4661 * Remove a vdev from the L2ARC.
4664 l2arc_remove_vdev(vdev_t *vd)
4666 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4669 * Find the device by vdev
4671 mutex_enter(&l2arc_dev_mtx);
4672 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4673 nextdev = list_next(l2arc_dev_list, dev);
4674 if (vd == dev->l2ad_vdev) {
4679 ASSERT(remdev != NULL);
4682 * Remove device from global list
4684 list_remove(l2arc_dev_list, remdev);
4685 l2arc_dev_last = NULL; /* may have been invalidated */
4686 atomic_dec_64(&l2arc_ndev);
4687 mutex_exit(&l2arc_dev_mtx);
4690 * Clear all buflists and ARC references. L2ARC device flush.
4692 l2arc_evict(remdev, 0, B_TRUE);
4693 list_destroy(remdev->l2ad_buflist);
4694 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
4695 kmem_free(remdev, sizeof (l2arc_dev_t));
4701 l2arc_thread_exit = 0;
4703 l2arc_writes_sent = 0;
4704 l2arc_writes_done = 0;
4706 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4707 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
4708 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
4709 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
4710 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
4712 l2arc_dev_list = &L2ARC_dev_list;
4713 l2arc_free_on_write = &L2ARC_free_on_write;
4714 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
4715 offsetof(l2arc_dev_t, l2ad_node));
4716 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
4717 offsetof(l2arc_data_free_t, l2df_list_node));
4724 * This is called from dmu_fini(), which is called from spa_fini();
4725 * Because of this, we can assume that all l2arc devices have
4726 * already been removed when the pools themselves were removed.
4729 l2arc_do_free_on_write();
4731 mutex_destroy(&l2arc_feed_thr_lock);
4732 cv_destroy(&l2arc_feed_thr_cv);
4733 mutex_destroy(&l2arc_dev_mtx);
4734 mutex_destroy(&l2arc_buflist_mtx);
4735 mutex_destroy(&l2arc_free_on_write_mtx);
4737 list_destroy(l2arc_dev_list);
4738 list_destroy(l2arc_free_on_write);
4744 if (!(spa_mode_global & FWRITE))
4747 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
4748 TS_RUN, minclsyspri);
4754 if (!(spa_mode_global & FWRITE))
4757 mutex_enter(&l2arc_feed_thr_lock);
4758 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
4759 l2arc_thread_exit = 1;
4760 while (l2arc_thread_exit != 0)
4761 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
4762 mutex_exit(&l2arc_feed_thr_lock);
4765 #if defined(_KERNEL) && defined(HAVE_SPL)
4766 EXPORT_SYMBOL(arc_read);
4767 EXPORT_SYMBOL(arc_buf_remove_ref);
4768 EXPORT_SYMBOL(arc_getbuf_func);
4770 module_param(zfs_arc_min, ulong, 0644);
4771 MODULE_PARM_DESC(zfs_arc_min, "Minimum arc size");
4773 module_param(zfs_arc_max, ulong, 0644);
4774 MODULE_PARM_DESC(zfs_arc_max, "Maximum arc size");
4776 module_param(zfs_arc_meta_limit, ulong, 0644);
4777 MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size");
4779 module_param(arc_reduce_dnlc_percent, uint, 0644);
4780 MODULE_PARM_DESC(arc_reduce_dnlc_percent, "Meta reclaim percentage");