4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
27 * DVA-based Adjustable Replacement Cache
29 * While much of the theory of operation used here is
30 * based on the self-tuning, low overhead replacement cache
31 * presented by Megiddo and Modha at FAST 2003, there are some
32 * significant differences:
34 * 1. The Megiddo and Modha model assumes any page is evictable.
35 * Pages in its cache cannot be "locked" into memory. This makes
36 * the eviction algorithm simple: evict the last page in the list.
37 * This also make the performance characteristics easy to reason
38 * about. Our cache is not so simple. At any given moment, some
39 * subset of the blocks in the cache are un-evictable because we
40 * have handed out a reference to them. Blocks are only evictable
41 * when there are no external references active. This makes
42 * eviction far more problematic: we choose to evict the evictable
43 * blocks that are the "lowest" in the list.
45 * There are times when it is not possible to evict the requested
46 * space. In these circumstances we are unable to adjust the cache
47 * size. To prevent the cache growing unbounded at these times we
48 * implement a "cache throttle" that slows the flow of new data
49 * into the cache until we can make space available.
51 * 2. The Megiddo and Modha model assumes a fixed cache size.
52 * Pages are evicted when the cache is full and there is a cache
53 * miss. Our model has a variable sized cache. It grows with
54 * high use, but also tries to react to memory pressure from the
55 * operating system: decreasing its size when system memory is
58 * 3. The Megiddo and Modha model assumes a fixed page size. All
59 * elements of the cache are therefor exactly the same size. So
60 * when adjusting the cache size following a cache miss, its simply
61 * a matter of choosing a single page to evict. In our model, we
62 * have variable sized cache blocks (rangeing from 512 bytes to
63 * 128K bytes). We therefor choose a set of blocks to evict to make
64 * space for a cache miss that approximates as closely as possible
65 * the space used by the new block.
67 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
68 * by N. Megiddo & D. Modha, FAST 2003
74 * A new reference to a cache buffer can be obtained in two
75 * ways: 1) via a hash table lookup using the DVA as a key,
76 * or 2) via one of the ARC lists. The arc_read() interface
77 * uses method 1, while the internal arc algorithms for
78 * adjusting the cache use method 2. We therefor provide two
79 * types of locks: 1) the hash table lock array, and 2) the
82 * Buffers do not have their own mutexs, rather they rely on the
83 * hash table mutexs for the bulk of their protection (i.e. most
84 * fields in the arc_buf_hdr_t are protected by these mutexs).
86 * buf_hash_find() returns the appropriate mutex (held) when it
87 * locates the requested buffer in the hash table. It returns
88 * NULL for the mutex if the buffer was not in the table.
90 * buf_hash_remove() expects the appropriate hash mutex to be
91 * already held before it is invoked.
93 * Each arc state also has a mutex which is used to protect the
94 * buffer list associated with the state. When attempting to
95 * obtain a hash table lock while holding an arc list lock you
96 * must use: mutex_tryenter() to avoid deadlock. Also note that
97 * the active state mutex must be held before the ghost state mutex.
99 * Arc buffers may have an associated eviction callback function.
100 * This function will be invoked prior to removing the buffer (e.g.
101 * in arc_do_user_evicts()). Note however that the data associated
102 * with the buffer may be evicted prior to the callback. The callback
103 * must be made with *no locks held* (to prevent deadlock). Additionally,
104 * the users of callbacks must ensure that their private data is
105 * protected from simultaneous callbacks from arc_buf_evict()
106 * and arc_do_user_evicts().
108 * Note that the majority of the performance stats are manipulated
109 * with atomic operations.
111 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
113 * - L2ARC buflist creation
114 * - L2ARC buflist eviction
115 * - L2ARC write completion, which walks L2ARC buflists
116 * - ARC header destruction, as it removes from L2ARC buflists
117 * - ARC header release, as it removes from L2ARC buflists
122 #include <sys/zio_checksum.h>
123 #include <sys/zfs_context.h>
125 #include <sys/refcount.h>
126 #include <sys/vdev.h>
128 #include <sys/dnlc.h>
130 #include <sys/callb.h>
131 #include <sys/kstat.h>
134 #include <vm/vm_pageout.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;
168 extern int zfs_prefetch_disable;
171 * The arc has filled available memory and has now warmed up.
173 static boolean_t arc_warm;
176 * These tunables are for performance analysis.
178 uint64_t zfs_arc_max;
179 uint64_t zfs_arc_min;
180 uint64_t zfs_arc_meta_limit = 0;
181 int zfs_mdcomp_disable = 0;
182 int zfs_arc_grow_retry = 0;
183 int zfs_arc_shrink_shift = 0;
184 int zfs_arc_p_min_shift = 0;
186 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
187 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
188 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
189 TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable);
190 SYSCTL_DECL(_vfs_zfs);
191 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
193 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
195 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RDTUN,
196 &zfs_mdcomp_disable, 0, "Disable metadata compression");
199 * Note that buffers can be in one of 6 states:
200 * ARC_anon - anonymous (discussed below)
201 * ARC_mru - recently used, currently cached
202 * ARC_mru_ghost - recentely used, no longer in cache
203 * ARC_mfu - frequently used, currently cached
204 * ARC_mfu_ghost - frequently used, no longer in cache
205 * ARC_l2c_only - exists in L2ARC but not other states
206 * When there are no active references to the buffer, they are
207 * are linked onto a list in one of these arc states. These are
208 * the only buffers that can be evicted or deleted. Within each
209 * state there are multiple lists, one for meta-data and one for
210 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
211 * etc.) is tracked separately so that it can be managed more
212 * explicitly: favored over data, limited explicitly.
214 * Anonymous buffers are buffers that are not associated with
215 * a DVA. These are buffers that hold dirty block copies
216 * before they are written to stable storage. By definition,
217 * they are "ref'd" and are considered part of arc_mru
218 * that cannot be freed. Generally, they will aquire a DVA
219 * as they are written and migrate onto the arc_mru list.
221 * The ARC_l2c_only state is for buffers that are in the second
222 * level ARC but no longer in any of the ARC_m* lists. The second
223 * level ARC itself may also contain buffers that are in any of
224 * the ARC_m* states - meaning that a buffer can exist in two
225 * places. The reason for the ARC_l2c_only state is to keep the
226 * buffer header in the hash table, so that reads that hit the
227 * second level ARC benefit from these fast lookups.
230 #define ARCS_LOCK_PAD CACHE_LINE_SIZE
234 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))];
239 * must be power of two for mask use to work
242 #define ARC_BUFC_NUMDATALISTS 16
243 #define ARC_BUFC_NUMMETADATALISTS 16
244 #define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS)
246 typedef struct arc_state {
247 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
248 uint64_t arcs_size; /* total amount of data in this state */
249 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */
250 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE);
253 #define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock))
256 static arc_state_t ARC_anon;
257 static arc_state_t ARC_mru;
258 static arc_state_t ARC_mru_ghost;
259 static arc_state_t ARC_mfu;
260 static arc_state_t ARC_mfu_ghost;
261 static arc_state_t ARC_l2c_only;
263 typedef struct arc_stats {
264 kstat_named_t arcstat_hits;
265 kstat_named_t arcstat_misses;
266 kstat_named_t arcstat_demand_data_hits;
267 kstat_named_t arcstat_demand_data_misses;
268 kstat_named_t arcstat_demand_metadata_hits;
269 kstat_named_t arcstat_demand_metadata_misses;
270 kstat_named_t arcstat_prefetch_data_hits;
271 kstat_named_t arcstat_prefetch_data_misses;
272 kstat_named_t arcstat_prefetch_metadata_hits;
273 kstat_named_t arcstat_prefetch_metadata_misses;
274 kstat_named_t arcstat_mru_hits;
275 kstat_named_t arcstat_mru_ghost_hits;
276 kstat_named_t arcstat_mfu_hits;
277 kstat_named_t arcstat_mfu_ghost_hits;
278 kstat_named_t arcstat_allocated;
279 kstat_named_t arcstat_deleted;
280 kstat_named_t arcstat_stolen;
281 kstat_named_t arcstat_recycle_miss;
282 kstat_named_t arcstat_mutex_miss;
283 kstat_named_t arcstat_evict_skip;
284 kstat_named_t arcstat_evict_l2_cached;
285 kstat_named_t arcstat_evict_l2_eligible;
286 kstat_named_t arcstat_evict_l2_ineligible;
287 kstat_named_t arcstat_hash_elements;
288 kstat_named_t arcstat_hash_elements_max;
289 kstat_named_t arcstat_hash_collisions;
290 kstat_named_t arcstat_hash_chains;
291 kstat_named_t arcstat_hash_chain_max;
292 kstat_named_t arcstat_p;
293 kstat_named_t arcstat_c;
294 kstat_named_t arcstat_c_min;
295 kstat_named_t arcstat_c_max;
296 kstat_named_t arcstat_size;
297 kstat_named_t arcstat_hdr_size;
298 kstat_named_t arcstat_data_size;
299 kstat_named_t arcstat_other_size;
300 kstat_named_t arcstat_l2_hits;
301 kstat_named_t arcstat_l2_misses;
302 kstat_named_t arcstat_l2_feeds;
303 kstat_named_t arcstat_l2_rw_clash;
304 kstat_named_t arcstat_l2_read_bytes;
305 kstat_named_t arcstat_l2_write_bytes;
306 kstat_named_t arcstat_l2_writes_sent;
307 kstat_named_t arcstat_l2_writes_done;
308 kstat_named_t arcstat_l2_writes_error;
309 kstat_named_t arcstat_l2_writes_hdr_miss;
310 kstat_named_t arcstat_l2_evict_lock_retry;
311 kstat_named_t arcstat_l2_evict_reading;
312 kstat_named_t arcstat_l2_free_on_write;
313 kstat_named_t arcstat_l2_abort_lowmem;
314 kstat_named_t arcstat_l2_cksum_bad;
315 kstat_named_t arcstat_l2_io_error;
316 kstat_named_t arcstat_l2_size;
317 kstat_named_t arcstat_l2_hdr_size;
318 kstat_named_t arcstat_memory_throttle_count;
319 kstat_named_t arcstat_l2_write_trylock_fail;
320 kstat_named_t arcstat_l2_write_passed_headroom;
321 kstat_named_t arcstat_l2_write_spa_mismatch;
322 kstat_named_t arcstat_l2_write_in_l2;
323 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
324 kstat_named_t arcstat_l2_write_not_cacheable;
325 kstat_named_t arcstat_l2_write_full;
326 kstat_named_t arcstat_l2_write_buffer_iter;
327 kstat_named_t arcstat_l2_write_pios;
328 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
329 kstat_named_t arcstat_l2_write_buffer_list_iter;
330 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
333 static arc_stats_t arc_stats = {
334 { "hits", KSTAT_DATA_UINT64 },
335 { "misses", KSTAT_DATA_UINT64 },
336 { "demand_data_hits", KSTAT_DATA_UINT64 },
337 { "demand_data_misses", KSTAT_DATA_UINT64 },
338 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
339 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
340 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
341 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
342 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
343 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
344 { "mru_hits", KSTAT_DATA_UINT64 },
345 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
346 { "mfu_hits", KSTAT_DATA_UINT64 },
347 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
348 { "allocated", KSTAT_DATA_UINT64 },
349 { "deleted", KSTAT_DATA_UINT64 },
350 { "stolen", KSTAT_DATA_UINT64 },
351 { "recycle_miss", KSTAT_DATA_UINT64 },
352 { "mutex_miss", KSTAT_DATA_UINT64 },
353 { "evict_skip", KSTAT_DATA_UINT64 },
354 { "evict_l2_cached", KSTAT_DATA_UINT64 },
355 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
356 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
357 { "hash_elements", KSTAT_DATA_UINT64 },
358 { "hash_elements_max", KSTAT_DATA_UINT64 },
359 { "hash_collisions", KSTAT_DATA_UINT64 },
360 { "hash_chains", KSTAT_DATA_UINT64 },
361 { "hash_chain_max", KSTAT_DATA_UINT64 },
362 { "p", KSTAT_DATA_UINT64 },
363 { "c", KSTAT_DATA_UINT64 },
364 { "c_min", KSTAT_DATA_UINT64 },
365 { "c_max", KSTAT_DATA_UINT64 },
366 { "size", KSTAT_DATA_UINT64 },
367 { "hdr_size", KSTAT_DATA_UINT64 },
368 { "data_size", KSTAT_DATA_UINT64 },
369 { "other_size", KSTAT_DATA_UINT64 },
370 { "l2_hits", KSTAT_DATA_UINT64 },
371 { "l2_misses", KSTAT_DATA_UINT64 },
372 { "l2_feeds", KSTAT_DATA_UINT64 },
373 { "l2_rw_clash", KSTAT_DATA_UINT64 },
374 { "l2_read_bytes", KSTAT_DATA_UINT64 },
375 { "l2_write_bytes", KSTAT_DATA_UINT64 },
376 { "l2_writes_sent", KSTAT_DATA_UINT64 },
377 { "l2_writes_done", KSTAT_DATA_UINT64 },
378 { "l2_writes_error", KSTAT_DATA_UINT64 },
379 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
380 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
381 { "l2_evict_reading", KSTAT_DATA_UINT64 },
382 { "l2_free_on_write", KSTAT_DATA_UINT64 },
383 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
384 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
385 { "l2_io_error", KSTAT_DATA_UINT64 },
386 { "l2_size", KSTAT_DATA_UINT64 },
387 { "l2_hdr_size", KSTAT_DATA_UINT64 },
388 { "memory_throttle_count", KSTAT_DATA_UINT64 },
389 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
390 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
391 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
392 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
393 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
394 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
395 { "l2_write_full", KSTAT_DATA_UINT64 },
396 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
397 { "l2_write_pios", KSTAT_DATA_UINT64 },
398 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
399 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
400 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }
403 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
405 #define ARCSTAT_INCR(stat, val) \
406 atomic_add_64(&arc_stats.stat.value.ui64, (val));
408 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
409 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
411 #define ARCSTAT_MAX(stat, val) { \
413 while ((val) > (m = arc_stats.stat.value.ui64) && \
414 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
418 #define ARCSTAT_MAXSTAT(stat) \
419 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
422 * We define a macro to allow ARC hits/misses to be easily broken down by
423 * two separate conditions, giving a total of four different subtypes for
424 * each of hits and misses (so eight statistics total).
426 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
429 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
431 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
435 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
437 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
442 static arc_state_t *arc_anon;
443 static arc_state_t *arc_mru;
444 static arc_state_t *arc_mru_ghost;
445 static arc_state_t *arc_mfu;
446 static arc_state_t *arc_mfu_ghost;
447 static arc_state_t *arc_l2c_only;
450 * There are several ARC variables that are critical to export as kstats --
451 * but we don't want to have to grovel around in the kstat whenever we wish to
452 * manipulate them. For these variables, we therefore define them to be in
453 * terms of the statistic variable. This assures that we are not introducing
454 * the possibility of inconsistency by having shadow copies of the variables,
455 * while still allowing the code to be readable.
457 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
458 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
459 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
460 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
461 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
463 static int arc_no_grow; /* Don't try to grow cache size */
464 static uint64_t arc_tempreserve;
465 static uint64_t arc_loaned_bytes;
466 static uint64_t arc_meta_used;
467 static uint64_t arc_meta_limit;
468 static uint64_t arc_meta_max = 0;
469 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RDTUN,
470 &arc_meta_used, 0, "ARC metadata used");
471 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RDTUN,
472 &arc_meta_limit, 0, "ARC metadata limit");
474 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
476 typedef struct arc_callback arc_callback_t;
478 struct arc_callback {
480 arc_done_func_t *acb_done;
482 zio_t *acb_zio_dummy;
483 arc_callback_t *acb_next;
486 typedef struct arc_write_callback arc_write_callback_t;
488 struct arc_write_callback {
490 arc_done_func_t *awcb_ready;
491 arc_done_func_t *awcb_done;
496 /* protected by hash lock */
501 kmutex_t b_freeze_lock;
502 zio_cksum_t *b_freeze_cksum;
504 arc_buf_hdr_t *b_hash_next;
509 arc_callback_t *b_acb;
513 arc_buf_contents_t b_type;
517 /* protected by arc state mutex */
518 arc_state_t *b_state;
519 list_node_t b_arc_node;
521 /* updated atomically */
522 clock_t b_arc_access;
524 /* self protecting */
527 l2arc_buf_hdr_t *b_l2hdr;
528 list_node_t b_l2node;
531 static arc_buf_t *arc_eviction_list;
532 static kmutex_t arc_eviction_mtx;
533 static arc_buf_hdr_t arc_eviction_hdr;
534 static void arc_get_data_buf(arc_buf_t *buf);
535 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
536 static int arc_evict_needed(arc_buf_contents_t type);
537 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
539 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
541 #define GHOST_STATE(state) \
542 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
543 (state) == arc_l2c_only)
546 * Private ARC flags. These flags are private ARC only flags that will show up
547 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
548 * be passed in as arc_flags in things like arc_read. However, these flags
549 * should never be passed and should only be set by ARC code. When adding new
550 * public flags, make sure not to smash the private ones.
553 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
554 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
555 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
556 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
557 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
558 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
559 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
560 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
561 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
562 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
563 #define ARC_STORED (1 << 19) /* has been store()d to */
565 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
566 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
567 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
568 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
569 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
570 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
571 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
572 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
573 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
574 (hdr)->b_l2hdr != NULL)
575 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
576 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
577 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
583 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
584 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
587 * Hash table routines
590 #define HT_LOCK_PAD CACHE_LINE_SIZE
595 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
599 #define BUF_LOCKS 256
600 typedef struct buf_hash_table {
602 arc_buf_hdr_t **ht_table;
603 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
606 static buf_hash_table_t buf_hash_table;
608 #define BUF_HASH_INDEX(spa, dva, birth) \
609 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
610 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
611 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
612 #define HDR_LOCK(buf) \
613 (BUF_HASH_LOCK(BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth)))
615 uint64_t zfs_crc64_table[256];
621 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
622 #define L2ARC_HEADROOM 2 /* num of writes */
623 #define L2ARC_FEED_SECS 1 /* caching interval secs */
624 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
626 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
627 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
630 * L2ARC Performance Tunables
632 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
633 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
634 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
635 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
636 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
637 boolean_t l2arc_noprefetch = B_FALSE; /* don't cache prefetch bufs */
638 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
639 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
641 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
642 &l2arc_write_max, 0, "max write size");
643 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
644 &l2arc_write_boost, 0, "extra write during warmup");
645 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
646 &l2arc_headroom, 0, "number of dev writes");
647 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
648 &l2arc_feed_secs, 0, "interval seconds");
649 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
650 &l2arc_feed_min_ms, 0, "min interval milliseconds");
652 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
653 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
654 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
655 &l2arc_feed_again, 0, "turbo warmup");
656 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
657 &l2arc_norw, 0, "no reads during writes");
659 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
660 &ARC_anon.arcs_size, 0, "size of anonymous state");
661 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
662 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
663 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
664 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
666 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
667 &ARC_mru.arcs_size, 0, "size of mru state");
668 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
669 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
670 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
671 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
673 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
674 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state");
675 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
676 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
677 "size of metadata in mru ghost state");
678 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
679 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
680 "size of data in mru ghost state");
682 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
683 &ARC_mfu.arcs_size, 0, "size of mfu state");
684 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
685 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
686 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
687 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
689 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
690 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state");
691 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
692 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
693 "size of metadata in mfu ghost state");
694 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
695 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
696 "size of data in mfu ghost state");
698 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
699 &ARC_l2c_only.arcs_size, 0, "size of mru state");
704 typedef struct l2arc_dev {
705 vdev_t *l2ad_vdev; /* vdev */
706 spa_t *l2ad_spa; /* spa */
707 uint64_t l2ad_hand; /* next write location */
708 uint64_t l2ad_write; /* desired write size, bytes */
709 uint64_t l2ad_boost; /* warmup write boost, bytes */
710 uint64_t l2ad_start; /* first addr on device */
711 uint64_t l2ad_end; /* last addr on device */
712 uint64_t l2ad_evict; /* last addr eviction reached */
713 boolean_t l2ad_first; /* first sweep through */
714 boolean_t l2ad_writing; /* currently writing */
715 list_t *l2ad_buflist; /* buffer list */
716 list_node_t l2ad_node; /* device list node */
719 static list_t L2ARC_dev_list; /* device list */
720 static list_t *l2arc_dev_list; /* device list pointer */
721 static kmutex_t l2arc_dev_mtx; /* device list mutex */
722 static l2arc_dev_t *l2arc_dev_last; /* last device used */
723 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
724 static list_t L2ARC_free_on_write; /* free after write buf list */
725 static list_t *l2arc_free_on_write; /* free after write list ptr */
726 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
727 static uint64_t l2arc_ndev; /* number of devices */
729 typedef struct l2arc_read_callback {
730 arc_buf_t *l2rcb_buf; /* read buffer */
731 spa_t *l2rcb_spa; /* spa */
732 blkptr_t l2rcb_bp; /* original blkptr */
733 zbookmark_t l2rcb_zb; /* original bookmark */
734 int l2rcb_flags; /* original flags */
735 } l2arc_read_callback_t;
737 typedef struct l2arc_write_callback {
738 l2arc_dev_t *l2wcb_dev; /* device info */
739 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
740 } l2arc_write_callback_t;
742 struct l2arc_buf_hdr {
743 /* protected by arc_buf_hdr mutex */
744 l2arc_dev_t *b_dev; /* L2ARC device */
745 uint64_t b_daddr; /* disk address, offset byte */
748 typedef struct l2arc_data_free {
749 /* protected by l2arc_free_on_write_mtx */
752 void (*l2df_func)(void *, size_t);
753 list_node_t l2df_list_node;
756 static kmutex_t l2arc_feed_thr_lock;
757 static kcondvar_t l2arc_feed_thr_cv;
758 static uint8_t l2arc_thread_exit;
760 static void l2arc_read_done(zio_t *zio);
761 static void l2arc_hdr_stat_add(void);
762 static void l2arc_hdr_stat_remove(void);
765 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
767 uint8_t *vdva = (uint8_t *)dva;
768 uint64_t crc = -1ULL;
771 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
773 for (i = 0; i < sizeof (dva_t); i++)
774 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
776 crc ^= (spa>>8) ^ birth;
781 #define BUF_EMPTY(buf) \
782 ((buf)->b_dva.dva_word[0] == 0 && \
783 (buf)->b_dva.dva_word[1] == 0 && \
786 #define BUF_EQUAL(spa, dva, birth, buf) \
787 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
788 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
789 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
791 static arc_buf_hdr_t *
792 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
794 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
795 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
798 mutex_enter(hash_lock);
799 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
800 buf = buf->b_hash_next) {
801 if (BUF_EQUAL(spa, dva, birth, buf)) {
806 mutex_exit(hash_lock);
812 * Insert an entry into the hash table. If there is already an element
813 * equal to elem in the hash table, then the already existing element
814 * will be returned and the new element will not be inserted.
815 * Otherwise returns NULL.
817 static arc_buf_hdr_t *
818 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
820 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
821 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
825 ASSERT(!HDR_IN_HASH_TABLE(buf));
827 mutex_enter(hash_lock);
828 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
829 fbuf = fbuf->b_hash_next, i++) {
830 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
834 buf->b_hash_next = buf_hash_table.ht_table[idx];
835 buf_hash_table.ht_table[idx] = buf;
836 buf->b_flags |= ARC_IN_HASH_TABLE;
838 /* collect some hash table performance data */
840 ARCSTAT_BUMP(arcstat_hash_collisions);
842 ARCSTAT_BUMP(arcstat_hash_chains);
844 ARCSTAT_MAX(arcstat_hash_chain_max, i);
847 ARCSTAT_BUMP(arcstat_hash_elements);
848 ARCSTAT_MAXSTAT(arcstat_hash_elements);
854 buf_hash_remove(arc_buf_hdr_t *buf)
856 arc_buf_hdr_t *fbuf, **bufp;
857 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
859 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
860 ASSERT(HDR_IN_HASH_TABLE(buf));
862 bufp = &buf_hash_table.ht_table[idx];
863 while ((fbuf = *bufp) != buf) {
864 ASSERT(fbuf != NULL);
865 bufp = &fbuf->b_hash_next;
867 *bufp = buf->b_hash_next;
868 buf->b_hash_next = NULL;
869 buf->b_flags &= ~ARC_IN_HASH_TABLE;
871 /* collect some hash table performance data */
872 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
874 if (buf_hash_table.ht_table[idx] &&
875 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
876 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
880 * Global data structures and functions for the buf kmem cache.
882 static kmem_cache_t *hdr_cache;
883 static kmem_cache_t *buf_cache;
890 kmem_free(buf_hash_table.ht_table,
891 (buf_hash_table.ht_mask + 1) * sizeof (void *));
892 for (i = 0; i < BUF_LOCKS; i++)
893 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
894 kmem_cache_destroy(hdr_cache);
895 kmem_cache_destroy(buf_cache);
899 * Constructor callback - called when the cache is empty
900 * and a new buf is requested.
904 hdr_cons(void *vbuf, void *unused, int kmflag)
906 arc_buf_hdr_t *buf = vbuf;
908 bzero(buf, sizeof (arc_buf_hdr_t));
909 refcount_create(&buf->b_refcnt);
910 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
911 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
912 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
919 buf_cons(void *vbuf, void *unused, int kmflag)
921 arc_buf_t *buf = vbuf;
923 bzero(buf, sizeof (arc_buf_t));
924 rw_init(&buf->b_lock, NULL, RW_DEFAULT, NULL);
925 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
931 * Destructor callback - called when a cached buf is
932 * no longer required.
936 hdr_dest(void *vbuf, void *unused)
938 arc_buf_hdr_t *buf = vbuf;
940 refcount_destroy(&buf->b_refcnt);
941 cv_destroy(&buf->b_cv);
942 mutex_destroy(&buf->b_freeze_lock);
943 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
948 buf_dest(void *vbuf, void *unused)
950 arc_buf_t *buf = vbuf;
952 rw_destroy(&buf->b_lock);
953 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
957 * Reclaim callback -- invoked when memory is low.
961 hdr_recl(void *unused)
963 dprintf("hdr_recl called\n");
965 * umem calls the reclaim func when we destroy the buf cache,
966 * which is after we do arc_fini().
969 cv_signal(&arc_reclaim_thr_cv);
976 uint64_t hsize = 1ULL << 12;
980 * The hash table is big enough to fill all of physical memory
981 * with an average 64K block size. The table will take up
982 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
984 while (hsize * 65536 < (uint64_t)physmem * PAGESIZE)
987 buf_hash_table.ht_mask = hsize - 1;
988 buf_hash_table.ht_table =
989 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
990 if (buf_hash_table.ht_table == NULL) {
991 ASSERT(hsize > (1ULL << 8));
996 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
997 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
998 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
999 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1001 for (i = 0; i < 256; i++)
1002 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1003 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1005 for (i = 0; i < BUF_LOCKS; i++) {
1006 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1007 NULL, MUTEX_DEFAULT, NULL);
1011 #define ARC_MINTIME (hz>>4) /* 62 ms */
1014 arc_cksum_verify(arc_buf_t *buf)
1018 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1021 mutex_enter(&buf->b_hdr->b_freeze_lock);
1022 if (buf->b_hdr->b_freeze_cksum == NULL ||
1023 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
1024 mutex_exit(&buf->b_hdr->b_freeze_lock);
1027 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1028 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1029 panic("buffer modified while frozen!");
1030 mutex_exit(&buf->b_hdr->b_freeze_lock);
1034 arc_cksum_equal(arc_buf_t *buf)
1039 mutex_enter(&buf->b_hdr->b_freeze_lock);
1040 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1041 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1042 mutex_exit(&buf->b_hdr->b_freeze_lock);
1048 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1050 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1053 mutex_enter(&buf->b_hdr->b_freeze_lock);
1054 if (buf->b_hdr->b_freeze_cksum != NULL) {
1055 mutex_exit(&buf->b_hdr->b_freeze_lock);
1058 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1059 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1060 buf->b_hdr->b_freeze_cksum);
1061 mutex_exit(&buf->b_hdr->b_freeze_lock);
1065 arc_buf_thaw(arc_buf_t *buf)
1067 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1068 if (buf->b_hdr->b_state != arc_anon)
1069 panic("modifying non-anon buffer!");
1070 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1071 panic("modifying buffer while i/o in progress!");
1072 arc_cksum_verify(buf);
1075 mutex_enter(&buf->b_hdr->b_freeze_lock);
1076 if (buf->b_hdr->b_freeze_cksum != NULL) {
1077 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1078 buf->b_hdr->b_freeze_cksum = NULL;
1080 mutex_exit(&buf->b_hdr->b_freeze_lock);
1084 arc_buf_freeze(arc_buf_t *buf)
1086 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1089 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1090 buf->b_hdr->b_state == arc_anon);
1091 arc_cksum_compute(buf, B_FALSE);
1095 get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock)
1097 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth);
1099 if (ab->b_type == ARC_BUFC_METADATA)
1100 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1);
1102 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1);
1103 buf_hashid += ARC_BUFC_NUMMETADATALISTS;
1106 *list = &state->arcs_lists[buf_hashid];
1107 *lock = ARCS_LOCK(state, buf_hashid);
1112 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1115 ASSERT(MUTEX_HELD(hash_lock));
1117 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1118 (ab->b_state != arc_anon)) {
1119 uint64_t delta = ab->b_size * ab->b_datacnt;
1120 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1124 get_buf_info(ab, ab->b_state, &list, &lock);
1125 ASSERT(!MUTEX_HELD(lock));
1127 ASSERT(list_link_active(&ab->b_arc_node));
1128 list_remove(list, ab);
1129 if (GHOST_STATE(ab->b_state)) {
1130 ASSERT3U(ab->b_datacnt, ==, 0);
1131 ASSERT3P(ab->b_buf, ==, NULL);
1135 ASSERT3U(*size, >=, delta);
1136 atomic_add_64(size, -delta);
1138 /* remove the prefetch flag if we get a reference */
1139 if (ab->b_flags & ARC_PREFETCH)
1140 ab->b_flags &= ~ARC_PREFETCH;
1145 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1148 arc_state_t *state = ab->b_state;
1150 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1151 ASSERT(!GHOST_STATE(state));
1153 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1154 (state != arc_anon)) {
1155 uint64_t *size = &state->arcs_lsize[ab->b_type];
1159 get_buf_info(ab, state, &list, &lock);
1160 ASSERT(!MUTEX_HELD(lock));
1162 ASSERT(!list_link_active(&ab->b_arc_node));
1163 list_insert_head(list, ab);
1164 ASSERT(ab->b_datacnt > 0);
1165 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1172 * Move the supplied buffer to the indicated state. The mutex
1173 * for the buffer must be held by the caller.
1176 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1178 arc_state_t *old_state = ab->b_state;
1179 int64_t refcnt = refcount_count(&ab->b_refcnt);
1180 uint64_t from_delta, to_delta;
1184 ASSERT(MUTEX_HELD(hash_lock));
1185 ASSERT(new_state != old_state);
1186 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1187 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1189 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1192 * If this buffer is evictable, transfer it from the
1193 * old state list to the new state list.
1196 if (old_state != arc_anon) {
1198 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1200 get_buf_info(ab, old_state, &list, &lock);
1201 use_mutex = !MUTEX_HELD(lock);
1205 ASSERT(list_link_active(&ab->b_arc_node));
1206 list_remove(list, ab);
1209 * If prefetching out of the ghost cache,
1210 * we will have a non-null datacnt.
1212 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1213 /* ghost elements have a ghost size */
1214 ASSERT(ab->b_buf == NULL);
1215 from_delta = ab->b_size;
1217 ASSERT3U(*size, >=, from_delta);
1218 atomic_add_64(size, -from_delta);
1223 if (new_state != arc_anon) {
1225 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1227 get_buf_info(ab, new_state, &list, &lock);
1228 use_mutex = !MUTEX_HELD(lock);
1232 list_insert_head(list, ab);
1234 /* ghost elements have a ghost size */
1235 if (GHOST_STATE(new_state)) {
1236 ASSERT(ab->b_datacnt == 0);
1237 ASSERT(ab->b_buf == NULL);
1238 to_delta = ab->b_size;
1240 atomic_add_64(size, to_delta);
1247 ASSERT(!BUF_EMPTY(ab));
1248 if (new_state == arc_anon) {
1249 buf_hash_remove(ab);
1252 /* adjust state sizes */
1254 atomic_add_64(&new_state->arcs_size, to_delta);
1256 ASSERT3U(old_state->arcs_size, >=, from_delta);
1257 atomic_add_64(&old_state->arcs_size, -from_delta);
1259 ab->b_state = new_state;
1261 /* adjust l2arc hdr stats */
1262 if (new_state == arc_l2c_only)
1263 l2arc_hdr_stat_add();
1264 else if (old_state == arc_l2c_only)
1265 l2arc_hdr_stat_remove();
1269 arc_space_consume(uint64_t space, arc_space_type_t type)
1271 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1274 case ARC_SPACE_DATA:
1275 ARCSTAT_INCR(arcstat_data_size, space);
1277 case ARC_SPACE_OTHER:
1278 ARCSTAT_INCR(arcstat_other_size, space);
1280 case ARC_SPACE_HDRS:
1281 ARCSTAT_INCR(arcstat_hdr_size, space);
1283 case ARC_SPACE_L2HDRS:
1284 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1288 atomic_add_64(&arc_meta_used, space);
1289 atomic_add_64(&arc_size, space);
1293 arc_space_return(uint64_t space, arc_space_type_t type)
1295 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1298 case ARC_SPACE_DATA:
1299 ARCSTAT_INCR(arcstat_data_size, -space);
1301 case ARC_SPACE_OTHER:
1302 ARCSTAT_INCR(arcstat_other_size, -space);
1304 case ARC_SPACE_HDRS:
1305 ARCSTAT_INCR(arcstat_hdr_size, -space);
1307 case ARC_SPACE_L2HDRS:
1308 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1312 ASSERT(arc_meta_used >= space);
1313 if (arc_meta_max < arc_meta_used)
1314 arc_meta_max = arc_meta_used;
1315 atomic_add_64(&arc_meta_used, -space);
1316 ASSERT(arc_size >= space);
1317 atomic_add_64(&arc_size, -space);
1321 arc_data_buf_alloc(uint64_t size)
1323 if (arc_evict_needed(ARC_BUFC_DATA))
1324 cv_signal(&arc_reclaim_thr_cv);
1325 atomic_add_64(&arc_size, size);
1326 return (zio_data_buf_alloc(size));
1330 arc_data_buf_free(void *buf, uint64_t size)
1332 zio_data_buf_free(buf, size);
1333 ASSERT(arc_size >= size);
1334 atomic_add_64(&arc_size, -size);
1338 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1343 ASSERT3U(size, >, 0);
1344 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1345 ASSERT(BUF_EMPTY(hdr));
1348 hdr->b_spa = spa_guid(spa);
1349 hdr->b_state = arc_anon;
1350 hdr->b_arc_access = 0;
1351 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1354 buf->b_efunc = NULL;
1355 buf->b_private = NULL;
1358 arc_get_data_buf(buf);
1361 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1362 (void) refcount_add(&hdr->b_refcnt, tag);
1367 static char *arc_onloan_tag = "onloan";
1370 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1371 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1372 * buffers must be returned to the arc before they can be used by the DMU or
1376 arc_loan_buf(spa_t *spa, int size)
1380 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1382 atomic_add_64(&arc_loaned_bytes, size);
1387 * Return a loaned arc buffer to the arc.
1390 arc_return_buf(arc_buf_t *buf, void *tag)
1392 arc_buf_hdr_t *hdr = buf->b_hdr;
1394 ASSERT(hdr->b_state == arc_anon);
1395 ASSERT(buf->b_data != NULL);
1396 VERIFY(refcount_remove(&hdr->b_refcnt, arc_onloan_tag) == 0);
1397 VERIFY(refcount_add(&hdr->b_refcnt, tag) == 1);
1399 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1403 arc_buf_clone(arc_buf_t *from)
1406 arc_buf_hdr_t *hdr = from->b_hdr;
1407 uint64_t size = hdr->b_size;
1409 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1412 buf->b_efunc = NULL;
1413 buf->b_private = NULL;
1414 buf->b_next = hdr->b_buf;
1416 arc_get_data_buf(buf);
1417 bcopy(from->b_data, buf->b_data, size);
1418 hdr->b_datacnt += 1;
1423 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1426 kmutex_t *hash_lock;
1429 * Check to see if this buffer is evicted. Callers
1430 * must verify b_data != NULL to know if the add_ref
1433 rw_enter(&buf->b_lock, RW_READER);
1434 if (buf->b_data == NULL) {
1435 rw_exit(&buf->b_lock);
1439 ASSERT(hdr != NULL);
1440 hash_lock = HDR_LOCK(hdr);
1441 mutex_enter(hash_lock);
1442 rw_exit(&buf->b_lock);
1444 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1445 add_reference(hdr, hash_lock, tag);
1446 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1447 arc_access(hdr, hash_lock);
1448 mutex_exit(hash_lock);
1449 ARCSTAT_BUMP(arcstat_hits);
1450 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1451 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1452 data, metadata, hits);
1456 * Free the arc data buffer. If it is an l2arc write in progress,
1457 * the buffer is placed on l2arc_free_on_write to be freed later.
1460 arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t),
1461 void *data, size_t size)
1463 if (HDR_L2_WRITING(hdr)) {
1464 l2arc_data_free_t *df;
1465 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1466 df->l2df_data = data;
1467 df->l2df_size = size;
1468 df->l2df_func = free_func;
1469 mutex_enter(&l2arc_free_on_write_mtx);
1470 list_insert_head(l2arc_free_on_write, df);
1471 mutex_exit(&l2arc_free_on_write_mtx);
1472 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1474 free_func(data, size);
1479 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1483 /* free up data associated with the buf */
1485 arc_state_t *state = buf->b_hdr->b_state;
1486 uint64_t size = buf->b_hdr->b_size;
1487 arc_buf_contents_t type = buf->b_hdr->b_type;
1489 arc_cksum_verify(buf);
1491 if (type == ARC_BUFC_METADATA) {
1492 arc_buf_data_free(buf->b_hdr, zio_buf_free,
1494 arc_space_return(size, ARC_SPACE_DATA);
1496 ASSERT(type == ARC_BUFC_DATA);
1497 arc_buf_data_free(buf->b_hdr,
1498 zio_data_buf_free, buf->b_data, size);
1499 ARCSTAT_INCR(arcstat_data_size, -size);
1500 atomic_add_64(&arc_size, -size);
1503 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1504 uint64_t *cnt = &state->arcs_lsize[type];
1506 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1507 ASSERT(state != arc_anon);
1509 ASSERT3U(*cnt, >=, size);
1510 atomic_add_64(cnt, -size);
1512 ASSERT3U(state->arcs_size, >=, size);
1513 atomic_add_64(&state->arcs_size, -size);
1515 ASSERT(buf->b_hdr->b_datacnt > 0);
1516 buf->b_hdr->b_datacnt -= 1;
1519 /* only remove the buf if requested */
1523 /* remove the buf from the hdr list */
1524 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1526 *bufp = buf->b_next;
1528 ASSERT(buf->b_efunc == NULL);
1530 /* clean up the buf */
1532 kmem_cache_free(buf_cache, buf);
1536 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1538 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1539 ASSERT3P(hdr->b_state, ==, arc_anon);
1540 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1541 ASSERT(!(hdr->b_flags & ARC_STORED));
1543 if (hdr->b_l2hdr != NULL) {
1544 if (!MUTEX_HELD(&l2arc_buflist_mtx)) {
1546 * To prevent arc_free() and l2arc_evict() from
1547 * attempting to free the same buffer at the same time,
1548 * a FREE_IN_PROGRESS flag is given to arc_free() to
1549 * give it priority. l2arc_evict() can't destroy this
1550 * header while we are waiting on l2arc_buflist_mtx.
1552 * The hdr may be removed from l2ad_buflist before we
1553 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1555 mutex_enter(&l2arc_buflist_mtx);
1556 if (hdr->b_l2hdr != NULL) {
1557 list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist,
1560 mutex_exit(&l2arc_buflist_mtx);
1562 list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist, hdr);
1564 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1565 kmem_free(hdr->b_l2hdr, sizeof (l2arc_buf_hdr_t));
1566 if (hdr->b_state == arc_l2c_only)
1567 l2arc_hdr_stat_remove();
1568 hdr->b_l2hdr = NULL;
1571 if (!BUF_EMPTY(hdr)) {
1572 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1573 bzero(&hdr->b_dva, sizeof (dva_t));
1577 while (hdr->b_buf) {
1578 arc_buf_t *buf = hdr->b_buf;
1581 mutex_enter(&arc_eviction_mtx);
1582 rw_enter(&buf->b_lock, RW_WRITER);
1583 ASSERT(buf->b_hdr != NULL);
1584 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1585 hdr->b_buf = buf->b_next;
1586 buf->b_hdr = &arc_eviction_hdr;
1587 buf->b_next = arc_eviction_list;
1588 arc_eviction_list = buf;
1589 rw_exit(&buf->b_lock);
1590 mutex_exit(&arc_eviction_mtx);
1592 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1595 if (hdr->b_freeze_cksum != NULL) {
1596 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1597 hdr->b_freeze_cksum = NULL;
1600 ASSERT(!list_link_active(&hdr->b_arc_node));
1601 ASSERT3P(hdr->b_hash_next, ==, NULL);
1602 ASSERT3P(hdr->b_acb, ==, NULL);
1603 kmem_cache_free(hdr_cache, hdr);
1607 arc_buf_free(arc_buf_t *buf, void *tag)
1609 arc_buf_hdr_t *hdr = buf->b_hdr;
1610 int hashed = hdr->b_state != arc_anon;
1612 ASSERT(buf->b_efunc == NULL);
1613 ASSERT(buf->b_data != NULL);
1616 kmutex_t *hash_lock = HDR_LOCK(hdr);
1618 mutex_enter(hash_lock);
1619 (void) remove_reference(hdr, hash_lock, tag);
1620 if (hdr->b_datacnt > 1)
1621 arc_buf_destroy(buf, FALSE, TRUE);
1623 hdr->b_flags |= ARC_BUF_AVAILABLE;
1624 mutex_exit(hash_lock);
1625 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1628 * We are in the middle of an async write. Don't destroy
1629 * this buffer unless the write completes before we finish
1630 * decrementing the reference count.
1632 mutex_enter(&arc_eviction_mtx);
1633 (void) remove_reference(hdr, NULL, tag);
1634 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1635 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1636 mutex_exit(&arc_eviction_mtx);
1638 arc_hdr_destroy(hdr);
1640 if (remove_reference(hdr, NULL, tag) > 0) {
1641 ASSERT(HDR_IO_ERROR(hdr));
1642 arc_buf_destroy(buf, FALSE, TRUE);
1644 arc_hdr_destroy(hdr);
1650 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1652 arc_buf_hdr_t *hdr = buf->b_hdr;
1653 kmutex_t *hash_lock = HDR_LOCK(hdr);
1654 int no_callback = (buf->b_efunc == NULL);
1656 if (hdr->b_state == arc_anon) {
1657 arc_buf_free(buf, tag);
1658 return (no_callback);
1661 mutex_enter(hash_lock);
1662 ASSERT(hdr->b_state != arc_anon);
1663 ASSERT(buf->b_data != NULL);
1665 (void) remove_reference(hdr, hash_lock, tag);
1666 if (hdr->b_datacnt > 1) {
1668 arc_buf_destroy(buf, FALSE, TRUE);
1669 } else if (no_callback) {
1670 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1671 hdr->b_flags |= ARC_BUF_AVAILABLE;
1673 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1674 refcount_is_zero(&hdr->b_refcnt));
1675 mutex_exit(hash_lock);
1676 return (no_callback);
1680 arc_buf_size(arc_buf_t *buf)
1682 return (buf->b_hdr->b_size);
1686 * Evict buffers from list until we've removed the specified number of
1687 * bytes. Move the removed buffers to the appropriate evict state.
1688 * If the recycle flag is set, then attempt to "recycle" a buffer:
1689 * - look for a buffer to evict that is `bytes' long.
1690 * - return the data block from this buffer rather than freeing it.
1691 * This flag is used by callers that are trying to make space for a
1692 * new buffer in a full arc cache.
1694 * This function makes a "best effort". It skips over any buffers
1695 * it can't get a hash_lock on, and so may not catch all candidates.
1696 * It may also return without evicting as much space as requested.
1699 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1700 arc_buf_contents_t type)
1702 arc_state_t *evicted_state;
1703 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1704 int64_t bytes_remaining;
1705 arc_buf_hdr_t *ab, *ab_prev = NULL;
1706 list_t *evicted_list, *list, *evicted_list_start, *list_start;
1707 kmutex_t *lock, *evicted_lock;
1708 kmutex_t *hash_lock;
1709 boolean_t have_lock;
1710 void *stolen = NULL;
1711 static int evict_metadata_offset, evict_data_offset;
1712 int i, idx, offset, list_count, count;
1714 ASSERT(state == arc_mru || state == arc_mfu);
1716 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1718 if (type == ARC_BUFC_METADATA) {
1720 list_count = ARC_BUFC_NUMMETADATALISTS;
1721 list_start = &state->arcs_lists[0];
1722 evicted_list_start = &evicted_state->arcs_lists[0];
1723 idx = evict_metadata_offset;
1725 offset = ARC_BUFC_NUMMETADATALISTS;
1726 list_start = &state->arcs_lists[offset];
1727 evicted_list_start = &evicted_state->arcs_lists[offset];
1728 list_count = ARC_BUFC_NUMDATALISTS;
1729 idx = evict_data_offset;
1731 bytes_remaining = evicted_state->arcs_lsize[type];
1735 list = &list_start[idx];
1736 evicted_list = &evicted_list_start[idx];
1737 lock = ARCS_LOCK(state, (offset + idx));
1738 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx));
1741 mutex_enter(evicted_lock);
1743 for (ab = list_tail(list); ab; ab = ab_prev) {
1744 ab_prev = list_prev(list, ab);
1745 bytes_remaining -= (ab->b_size * ab->b_datacnt);
1746 /* prefetch buffers have a minimum lifespan */
1747 if (HDR_IO_IN_PROGRESS(ab) ||
1748 (spa && ab->b_spa != spa) ||
1749 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1750 LBOLT - ab->b_arc_access < arc_min_prefetch_lifespan)) {
1754 /* "lookahead" for better eviction candidate */
1755 if (recycle && ab->b_size != bytes &&
1756 ab_prev && ab_prev->b_size == bytes)
1758 hash_lock = HDR_LOCK(ab);
1759 have_lock = MUTEX_HELD(hash_lock);
1760 if (have_lock || mutex_tryenter(hash_lock)) {
1761 ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0);
1762 ASSERT(ab->b_datacnt > 0);
1764 arc_buf_t *buf = ab->b_buf;
1765 if (!rw_tryenter(&buf->b_lock, RW_WRITER)) {
1770 bytes_evicted += ab->b_size;
1771 if (recycle && ab->b_type == type &&
1772 ab->b_size == bytes &&
1773 !HDR_L2_WRITING(ab)) {
1774 stolen = buf->b_data;
1779 mutex_enter(&arc_eviction_mtx);
1780 arc_buf_destroy(buf,
1781 buf->b_data == stolen, FALSE);
1782 ab->b_buf = buf->b_next;
1783 buf->b_hdr = &arc_eviction_hdr;
1784 buf->b_next = arc_eviction_list;
1785 arc_eviction_list = buf;
1786 mutex_exit(&arc_eviction_mtx);
1787 rw_exit(&buf->b_lock);
1789 rw_exit(&buf->b_lock);
1790 arc_buf_destroy(buf,
1791 buf->b_data == stolen, TRUE);
1796 ARCSTAT_INCR(arcstat_evict_l2_cached,
1799 if (l2arc_write_eligible(ab->b_spa, ab)) {
1800 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1804 arcstat_evict_l2_ineligible,
1809 if (ab->b_datacnt == 0) {
1810 arc_change_state(evicted_state, ab, hash_lock);
1811 ASSERT(HDR_IN_HASH_TABLE(ab));
1812 ab->b_flags |= ARC_IN_HASH_TABLE;
1813 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1814 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1817 mutex_exit(hash_lock);
1818 if (bytes >= 0 && bytes_evicted >= bytes)
1820 if (bytes_remaining > 0) {
1821 mutex_exit(evicted_lock);
1823 idx = ((idx + 1) & (list_count - 1));
1832 mutex_exit(evicted_lock);
1835 idx = ((idx + 1) & (list_count - 1));
1838 if (bytes_evicted < bytes) {
1839 if (count < list_count)
1842 dprintf("only evicted %lld bytes from %x",
1843 (longlong_t)bytes_evicted, state);
1845 if (type == ARC_BUFC_METADATA)
1846 evict_metadata_offset = idx;
1848 evict_data_offset = idx;
1851 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1854 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1857 * We have just evicted some date into the ghost state, make
1858 * sure we also adjust the ghost state size if necessary.
1861 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1862 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1863 arc_mru_ghost->arcs_size - arc_c;
1865 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1867 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1868 arc_evict_ghost(arc_mru_ghost, 0, todelete);
1869 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1870 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1871 arc_mru_ghost->arcs_size +
1872 arc_mfu_ghost->arcs_size - arc_c);
1873 arc_evict_ghost(arc_mfu_ghost, 0, todelete);
1877 ARCSTAT_BUMP(arcstat_stolen);
1883 * Remove buffers from list until we've removed the specified number of
1884 * bytes. Destroy the buffers that are removed.
1887 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1889 arc_buf_hdr_t *ab, *ab_prev;
1890 list_t *list, *list_start;
1891 kmutex_t *hash_lock, *lock;
1892 uint64_t bytes_deleted = 0;
1893 uint64_t bufs_skipped = 0;
1894 static int evict_offset;
1895 int list_count, idx = evict_offset;
1896 int offset, count = 0;
1898 ASSERT(GHOST_STATE(state));
1901 * data lists come after metadata lists
1903 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS];
1904 list_count = ARC_BUFC_NUMDATALISTS;
1905 offset = ARC_BUFC_NUMMETADATALISTS;
1908 list = &list_start[idx];
1909 lock = ARCS_LOCK(state, idx + offset);
1912 for (ab = list_tail(list); ab; ab = ab_prev) {
1913 ab_prev = list_prev(list, ab);
1914 if (spa && ab->b_spa != spa)
1916 hash_lock = HDR_LOCK(ab);
1917 if (mutex_tryenter(hash_lock)) {
1918 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1919 ASSERT(ab->b_buf == NULL);
1920 ARCSTAT_BUMP(arcstat_deleted);
1921 bytes_deleted += ab->b_size;
1923 if (ab->b_l2hdr != NULL) {
1925 * This buffer is cached on the 2nd Level ARC;
1926 * don't destroy the header.
1928 arc_change_state(arc_l2c_only, ab, hash_lock);
1929 mutex_exit(hash_lock);
1931 arc_change_state(arc_anon, ab, hash_lock);
1932 mutex_exit(hash_lock);
1933 arc_hdr_destroy(ab);
1936 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1937 if (bytes >= 0 && bytes_deleted >= bytes)
1942 * we're draining the ARC, retry
1945 mutex_enter(hash_lock);
1946 mutex_exit(hash_lock);
1953 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1));
1956 if (count < list_count)
1960 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] &&
1961 (bytes < 0 || bytes_deleted < bytes)) {
1962 list_start = &state->arcs_lists[0];
1963 list_count = ARC_BUFC_NUMMETADATALISTS;
1969 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1973 if (bytes_deleted < bytes)
1974 dprintf("only deleted %lld bytes from %p",
1975 (longlong_t)bytes_deleted, state);
1981 int64_t adjustment, delta;
1987 adjustment = MIN((int64_t)(arc_size - arc_c),
1988 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
1991 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1992 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
1993 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
1994 adjustment -= delta;
1997 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1998 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
1999 (void) arc_evict(arc_mru, 0, delta, FALSE,
2007 adjustment = arc_size - arc_c;
2009 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2010 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2011 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2012 adjustment -= delta;
2015 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2016 int64_t delta = MIN(adjustment,
2017 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2018 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2023 * Adjust ghost lists
2026 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2028 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2029 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2030 arc_evict_ghost(arc_mru_ghost, 0, delta);
2034 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2036 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2037 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2038 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2043 arc_do_user_evicts(void)
2045 static arc_buf_t *tmp_arc_eviction_list;
2048 * Move list over to avoid LOR
2051 mutex_enter(&arc_eviction_mtx);
2052 tmp_arc_eviction_list = arc_eviction_list;
2053 arc_eviction_list = NULL;
2054 mutex_exit(&arc_eviction_mtx);
2056 while (tmp_arc_eviction_list != NULL) {
2057 arc_buf_t *buf = tmp_arc_eviction_list;
2058 tmp_arc_eviction_list = buf->b_next;
2059 rw_enter(&buf->b_lock, RW_WRITER);
2061 rw_exit(&buf->b_lock);
2063 if (buf->b_efunc != NULL)
2064 VERIFY(buf->b_efunc(buf) == 0);
2066 buf->b_efunc = NULL;
2067 buf->b_private = NULL;
2068 kmem_cache_free(buf_cache, buf);
2071 if (arc_eviction_list != NULL)
2076 * Flush all *evictable* data from the cache for the given spa.
2077 * NOTE: this will not touch "active" (i.e. referenced) data.
2080 arc_flush(spa_t *spa)
2085 guid = spa_guid(spa);
2087 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) {
2088 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2092 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) {
2093 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2097 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) {
2098 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2102 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) {
2103 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2108 arc_evict_ghost(arc_mru_ghost, guid, -1);
2109 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2111 mutex_enter(&arc_reclaim_thr_lock);
2112 arc_do_user_evicts();
2113 mutex_exit(&arc_reclaim_thr_lock);
2114 ASSERT(spa || arc_eviction_list == NULL);
2120 if (arc_c > arc_c_min) {
2124 to_free = arc_c >> arc_shrink_shift;
2126 to_free = arc_c >> arc_shrink_shift;
2128 if (arc_c > arc_c_min + to_free)
2129 atomic_add_64(&arc_c, -to_free);
2133 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2134 if (arc_c > arc_size)
2135 arc_c = MAX(arc_size, arc_c_min);
2137 arc_p = (arc_c >> 1);
2138 ASSERT(arc_c >= arc_c_min);
2139 ASSERT((int64_t)arc_p >= 0);
2142 if (arc_size > arc_c)
2146 static int needfree = 0;
2149 arc_reclaim_needed(void)
2158 if (arc_size > arc_c_max)
2160 if (arc_size <= arc_c_min)
2164 * If pages are needed or we're within 2048 pages
2165 * of needing to page need to reclaim
2167 if (vm_pages_needed || (vm_paging_target() > -2048))
2172 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2177 * check that we're out of range of the pageout scanner. It starts to
2178 * schedule paging if freemem is less than lotsfree and needfree.
2179 * lotsfree is the high-water mark for pageout, and needfree is the
2180 * number of needed free pages. We add extra pages here to make sure
2181 * the scanner doesn't start up while we're freeing memory.
2183 if (freemem < lotsfree + needfree + extra)
2187 * check to make sure that swapfs has enough space so that anon
2188 * reservations can still succeed. anon_resvmem() checks that the
2189 * availrmem is greater than swapfs_minfree, and the number of reserved
2190 * swap pages. We also add a bit of extra here just to prevent
2191 * circumstances from getting really dire.
2193 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2198 * If we're on an i386 platform, it's possible that we'll exhaust the
2199 * kernel heap space before we ever run out of available physical
2200 * memory. Most checks of the size of the heap_area compare against
2201 * tune.t_minarmem, which is the minimum available real memory that we
2202 * can have in the system. However, this is generally fixed at 25 pages
2203 * which is so low that it's useless. In this comparison, we seek to
2204 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2205 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2208 if (btop(vmem_size(heap_arena, VMEM_FREE)) <
2209 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
2213 if (kmem_used() > (kmem_size() * 3) / 4)
2218 if (spa_get_random(100) == 0)
2224 extern kmem_cache_t *zio_buf_cache[];
2225 extern kmem_cache_t *zio_data_buf_cache[];
2228 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2231 kmem_cache_t *prev_cache = NULL;
2232 kmem_cache_t *prev_data_cache = NULL;
2235 if (arc_meta_used >= arc_meta_limit) {
2237 * We are exceeding our meta-data cache limit.
2238 * Purge some DNLC entries to release holds on meta-data.
2240 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2244 * Reclaim unused memory from all kmem caches.
2251 * An aggressive reclamation will shrink the cache size as well as
2252 * reap free buffers from the arc kmem caches.
2254 if (strat == ARC_RECLAIM_AGGR)
2257 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2258 if (zio_buf_cache[i] != prev_cache) {
2259 prev_cache = zio_buf_cache[i];
2260 kmem_cache_reap_now(zio_buf_cache[i]);
2262 if (zio_data_buf_cache[i] != prev_data_cache) {
2263 prev_data_cache = zio_data_buf_cache[i];
2264 kmem_cache_reap_now(zio_data_buf_cache[i]);
2267 kmem_cache_reap_now(buf_cache);
2268 kmem_cache_reap_now(hdr_cache);
2272 arc_reclaim_thread(void *dummy __unused)
2274 clock_t growtime = 0;
2275 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2278 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2280 mutex_enter(&arc_reclaim_thr_lock);
2281 while (arc_thread_exit == 0) {
2282 if (arc_reclaim_needed()) {
2285 if (last_reclaim == ARC_RECLAIM_CONS) {
2286 last_reclaim = ARC_RECLAIM_AGGR;
2288 last_reclaim = ARC_RECLAIM_CONS;
2292 last_reclaim = ARC_RECLAIM_AGGR;
2296 /* reset the growth delay for every reclaim */
2297 growtime = LBOLT + (arc_grow_retry * hz);
2299 if (needfree && last_reclaim == ARC_RECLAIM_CONS) {
2301 * If needfree is TRUE our vm_lowmem hook
2302 * was called and in that case we must free some
2303 * memory, so switch to aggressive mode.
2306 last_reclaim = ARC_RECLAIM_AGGR;
2308 arc_kmem_reap_now(last_reclaim);
2311 } else if (arc_no_grow && LBOLT >= growtime) {
2312 arc_no_grow = FALSE;
2317 if (arc_eviction_list != NULL)
2318 arc_do_user_evicts();
2327 /* block until needed, or one second, whichever is shorter */
2328 CALLB_CPR_SAFE_BEGIN(&cpr);
2329 (void) cv_timedwait(&arc_reclaim_thr_cv,
2330 &arc_reclaim_thr_lock, hz);
2331 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2334 arc_thread_exit = 0;
2335 cv_broadcast(&arc_reclaim_thr_cv);
2336 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2341 * Adapt arc info given the number of bytes we are trying to add and
2342 * the state that we are comming from. This function is only called
2343 * when we are adding new content to the cache.
2346 arc_adapt(int bytes, arc_state_t *state)
2349 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2351 if (state == arc_l2c_only)
2356 * Adapt the target size of the MRU list:
2357 * - if we just hit in the MRU ghost list, then increase
2358 * the target size of the MRU list.
2359 * - if we just hit in the MFU ghost list, then increase
2360 * the target size of the MFU list by decreasing the
2361 * target size of the MRU list.
2363 if (state == arc_mru_ghost) {
2364 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2365 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2366 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2368 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2369 } else if (state == arc_mfu_ghost) {
2372 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2373 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2374 mult = MIN(mult, 10);
2376 delta = MIN(bytes * mult, arc_p);
2377 arc_p = MAX(arc_p_min, arc_p - delta);
2379 ASSERT((int64_t)arc_p >= 0);
2381 if (arc_reclaim_needed()) {
2382 cv_signal(&arc_reclaim_thr_cv);
2389 if (arc_c >= arc_c_max)
2393 * If we're within (2 * maxblocksize) bytes of the target
2394 * cache size, increment the target cache size
2396 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2397 atomic_add_64(&arc_c, (int64_t)bytes);
2398 if (arc_c > arc_c_max)
2400 else if (state == arc_anon)
2401 atomic_add_64(&arc_p, (int64_t)bytes);
2405 ASSERT((int64_t)arc_p >= 0);
2409 * Check if the cache has reached its limits and eviction is required
2413 arc_evict_needed(arc_buf_contents_t type)
2415 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2421 * If zio data pages are being allocated out of a separate heap segment,
2422 * then enforce that the size of available vmem for this area remains
2423 * above about 1/32nd free.
2425 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2426 vmem_size(zio_arena, VMEM_FREE) <
2427 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2432 if (arc_reclaim_needed())
2435 return (arc_size > arc_c);
2439 * The buffer, supplied as the first argument, needs a data block.
2440 * So, if we are at cache max, determine which cache should be victimized.
2441 * We have the following cases:
2443 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2444 * In this situation if we're out of space, but the resident size of the MFU is
2445 * under the limit, victimize the MFU cache to satisfy this insertion request.
2447 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2448 * Here, we've used up all of the available space for the MRU, so we need to
2449 * evict from our own cache instead. Evict from the set of resident MRU
2452 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2453 * c minus p represents the MFU space in the cache, since p is the size of the
2454 * cache that is dedicated to the MRU. In this situation there's still space on
2455 * the MFU side, so the MRU side needs to be victimized.
2457 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2458 * MFU's resident set is consuming more space than it has been allotted. In
2459 * this situation, we must victimize our own cache, the MFU, for this insertion.
2462 arc_get_data_buf(arc_buf_t *buf)
2464 arc_state_t *state = buf->b_hdr->b_state;
2465 uint64_t size = buf->b_hdr->b_size;
2466 arc_buf_contents_t type = buf->b_hdr->b_type;
2468 arc_adapt(size, state);
2471 * We have not yet reached cache maximum size,
2472 * just allocate a new buffer.
2474 if (!arc_evict_needed(type)) {
2475 if (type == ARC_BUFC_METADATA) {
2476 buf->b_data = zio_buf_alloc(size);
2477 arc_space_consume(size, ARC_SPACE_DATA);
2479 ASSERT(type == ARC_BUFC_DATA);
2480 buf->b_data = zio_data_buf_alloc(size);
2481 ARCSTAT_INCR(arcstat_data_size, size);
2482 atomic_add_64(&arc_size, size);
2488 * If we are prefetching from the mfu ghost list, this buffer
2489 * will end up on the mru list; so steal space from there.
2491 if (state == arc_mfu_ghost)
2492 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2493 else if (state == arc_mru_ghost)
2496 if (state == arc_mru || state == arc_anon) {
2497 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2498 state = (arc_mfu->arcs_lsize[type] >= size &&
2499 arc_p > mru_used) ? arc_mfu : arc_mru;
2502 uint64_t mfu_space = arc_c - arc_p;
2503 state = (arc_mru->arcs_lsize[type] >= size &&
2504 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2506 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2507 if (type == ARC_BUFC_METADATA) {
2508 buf->b_data = zio_buf_alloc(size);
2509 arc_space_consume(size, ARC_SPACE_DATA);
2511 ASSERT(type == ARC_BUFC_DATA);
2512 buf->b_data = zio_data_buf_alloc(size);
2513 ARCSTAT_INCR(arcstat_data_size, size);
2514 atomic_add_64(&arc_size, size);
2516 ARCSTAT_BUMP(arcstat_recycle_miss);
2518 ASSERT(buf->b_data != NULL);
2521 * Update the state size. Note that ghost states have a
2522 * "ghost size" and so don't need to be updated.
2524 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2525 arc_buf_hdr_t *hdr = buf->b_hdr;
2527 atomic_add_64(&hdr->b_state->arcs_size, size);
2528 if (list_link_active(&hdr->b_arc_node)) {
2529 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2530 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2533 * If we are growing the cache, and we are adding anonymous
2534 * data, and we have outgrown arc_p, update arc_p
2536 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2537 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2538 arc_p = MIN(arc_c, arc_p + size);
2540 ARCSTAT_BUMP(arcstat_allocated);
2544 * This routine is called whenever a buffer is accessed.
2545 * NOTE: the hash lock is dropped in this function.
2548 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2550 ASSERT(MUTEX_HELD(hash_lock));
2552 if (buf->b_state == arc_anon) {
2554 * This buffer is not in the cache, and does not
2555 * appear in our "ghost" list. Add the new buffer
2559 ASSERT(buf->b_arc_access == 0);
2560 buf->b_arc_access = LBOLT;
2561 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2562 arc_change_state(arc_mru, buf, hash_lock);
2564 } else if (buf->b_state == arc_mru) {
2566 * If this buffer is here because of a prefetch, then either:
2567 * - clear the flag if this is a "referencing" read
2568 * (any subsequent access will bump this into the MFU state).
2570 * - move the buffer to the head of the list if this is
2571 * another prefetch (to make it less likely to be evicted).
2573 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2574 if (refcount_count(&buf->b_refcnt) == 0) {
2575 ASSERT(list_link_active(&buf->b_arc_node));
2577 buf->b_flags &= ~ARC_PREFETCH;
2578 ARCSTAT_BUMP(arcstat_mru_hits);
2580 buf->b_arc_access = LBOLT;
2585 * This buffer has been "accessed" only once so far,
2586 * but it is still in the cache. Move it to the MFU
2589 if (LBOLT > buf->b_arc_access + ARC_MINTIME) {
2591 * More than 125ms have passed since we
2592 * instantiated this buffer. Move it to the
2593 * most frequently used state.
2595 buf->b_arc_access = LBOLT;
2596 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2597 arc_change_state(arc_mfu, buf, hash_lock);
2599 ARCSTAT_BUMP(arcstat_mru_hits);
2600 } else if (buf->b_state == arc_mru_ghost) {
2601 arc_state_t *new_state;
2603 * This buffer has been "accessed" recently, but
2604 * was evicted from the cache. Move it to the
2608 if (buf->b_flags & ARC_PREFETCH) {
2609 new_state = arc_mru;
2610 if (refcount_count(&buf->b_refcnt) > 0)
2611 buf->b_flags &= ~ARC_PREFETCH;
2612 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2614 new_state = arc_mfu;
2615 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2618 buf->b_arc_access = LBOLT;
2619 arc_change_state(new_state, buf, hash_lock);
2621 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2622 } else if (buf->b_state == arc_mfu) {
2624 * This buffer has been accessed more than once and is
2625 * still in the cache. Keep it in the MFU state.
2627 * NOTE: an add_reference() that occurred when we did
2628 * the arc_read() will have kicked this off the list.
2629 * If it was a prefetch, we will explicitly move it to
2630 * the head of the list now.
2632 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2633 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2634 ASSERT(list_link_active(&buf->b_arc_node));
2636 ARCSTAT_BUMP(arcstat_mfu_hits);
2637 buf->b_arc_access = LBOLT;
2638 } else if (buf->b_state == arc_mfu_ghost) {
2639 arc_state_t *new_state = arc_mfu;
2641 * This buffer has been accessed more than once but has
2642 * been evicted from the cache. Move it back to the
2646 if (buf->b_flags & ARC_PREFETCH) {
2648 * This is a prefetch access...
2649 * move this block back to the MRU state.
2651 ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0);
2652 new_state = arc_mru;
2655 buf->b_arc_access = LBOLT;
2656 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2657 arc_change_state(new_state, buf, hash_lock);
2659 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2660 } else if (buf->b_state == arc_l2c_only) {
2662 * This buffer is on the 2nd Level ARC.
2665 buf->b_arc_access = LBOLT;
2666 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2667 arc_change_state(arc_mfu, buf, hash_lock);
2669 ASSERT(!"invalid arc state");
2673 /* a generic arc_done_func_t which you can use */
2676 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2678 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2679 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2682 /* a generic arc_done_func_t */
2684 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2686 arc_buf_t **bufp = arg;
2687 if (zio && zio->io_error) {
2688 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2696 arc_read_done(zio_t *zio)
2698 arc_buf_hdr_t *hdr, *found;
2700 arc_buf_t *abuf; /* buffer we're assigning to callback */
2701 kmutex_t *hash_lock;
2702 arc_callback_t *callback_list, *acb;
2703 int freeable = FALSE;
2705 buf = zio->io_private;
2709 * The hdr was inserted into hash-table and removed from lists
2710 * prior to starting I/O. We should find this header, since
2711 * it's in the hash table, and it should be legit since it's
2712 * not possible to evict it during the I/O. The only possible
2713 * reason for it not to be found is if we were freed during the
2716 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2719 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2720 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2721 (found == hdr && HDR_L2_READING(hdr)));
2723 hdr->b_flags &= ~ARC_L2_EVICTED;
2724 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2725 hdr->b_flags &= ~ARC_L2CACHE;
2727 /* byteswap if necessary */
2728 callback_list = hdr->b_acb;
2729 ASSERT(callback_list != NULL);
2730 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2731 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2732 byteswap_uint64_array :
2733 dmu_ot[BP_GET_TYPE(zio->io_bp)].ot_byteswap;
2734 func(buf->b_data, hdr->b_size);
2737 arc_cksum_compute(buf, B_FALSE);
2739 /* create copies of the data buffer for the callers */
2741 for (acb = callback_list; acb; acb = acb->acb_next) {
2742 if (acb->acb_done) {
2744 abuf = arc_buf_clone(buf);
2745 acb->acb_buf = abuf;
2750 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2751 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2753 hdr->b_flags |= ARC_BUF_AVAILABLE;
2755 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2757 if (zio->io_error != 0) {
2758 hdr->b_flags |= ARC_IO_ERROR;
2759 if (hdr->b_state != arc_anon)
2760 arc_change_state(arc_anon, hdr, hash_lock);
2761 if (HDR_IN_HASH_TABLE(hdr))
2762 buf_hash_remove(hdr);
2763 freeable = refcount_is_zero(&hdr->b_refcnt);
2767 * Broadcast before we drop the hash_lock to avoid the possibility
2768 * that the hdr (and hence the cv) might be freed before we get to
2769 * the cv_broadcast().
2771 cv_broadcast(&hdr->b_cv);
2775 * Only call arc_access on anonymous buffers. This is because
2776 * if we've issued an I/O for an evicted buffer, we've already
2777 * called arc_access (to prevent any simultaneous readers from
2778 * getting confused).
2780 if (zio->io_error == 0 && hdr->b_state == arc_anon)
2781 arc_access(hdr, hash_lock);
2782 mutex_exit(hash_lock);
2785 * This block was freed while we waited for the read to
2786 * complete. It has been removed from the hash table and
2787 * moved to the anonymous state (so that it won't show up
2790 ASSERT3P(hdr->b_state, ==, arc_anon);
2791 freeable = refcount_is_zero(&hdr->b_refcnt);
2794 /* execute each callback and free its structure */
2795 while ((acb = callback_list) != NULL) {
2797 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2799 if (acb->acb_zio_dummy != NULL) {
2800 acb->acb_zio_dummy->io_error = zio->io_error;
2801 zio_nowait(acb->acb_zio_dummy);
2804 callback_list = acb->acb_next;
2805 kmem_free(acb, sizeof (arc_callback_t));
2809 arc_hdr_destroy(hdr);
2813 * "Read" the block block at the specified DVA (in bp) via the
2814 * cache. If the block is found in the cache, invoke the provided
2815 * callback immediately and return. Note that the `zio' parameter
2816 * in the callback will be NULL in this case, since no IO was
2817 * required. If the block is not in the cache pass the read request
2818 * on to the spa with a substitute callback function, so that the
2819 * requested block will be added to the cache.
2821 * If a read request arrives for a block that has a read in-progress,
2822 * either wait for the in-progress read to complete (and return the
2823 * results); or, if this is a read with a "done" func, add a record
2824 * to the read to invoke the "done" func when the read completes,
2825 * and return; or just return.
2827 * arc_read_done() will invoke all the requested "done" functions
2828 * for readers of this block.
2830 * Normal callers should use arc_read and pass the arc buffer and offset
2831 * for the bp. But if you know you don't need locking, you can use
2835 arc_read(zio_t *pio, spa_t *spa, blkptr_t *bp, arc_buf_t *pbuf,
2836 arc_done_func_t *done, void *private, int priority, int zio_flags,
2837 uint32_t *arc_flags, const zbookmark_t *zb)
2841 ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt));
2842 ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size);
2843 rw_enter(&pbuf->b_lock, RW_READER);
2845 err = arc_read_nolock(pio, spa, bp, done, private, priority,
2846 zio_flags, arc_flags, zb);
2847 rw_exit(&pbuf->b_lock);
2852 arc_read_nolock(zio_t *pio, spa_t *spa, blkptr_t *bp,
2853 arc_done_func_t *done, void *private, int priority, int zio_flags,
2854 uint32_t *arc_flags, const zbookmark_t *zb)
2858 kmutex_t *hash_lock;
2860 uint64_t guid = spa_guid(spa);
2863 hdr = buf_hash_find(guid, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
2864 if (hdr && hdr->b_datacnt > 0) {
2866 *arc_flags |= ARC_CACHED;
2868 if (HDR_IO_IN_PROGRESS(hdr)) {
2870 if (*arc_flags & ARC_WAIT) {
2871 cv_wait(&hdr->b_cv, hash_lock);
2872 mutex_exit(hash_lock);
2875 ASSERT(*arc_flags & ARC_NOWAIT);
2878 arc_callback_t *acb = NULL;
2880 acb = kmem_zalloc(sizeof (arc_callback_t),
2882 acb->acb_done = done;
2883 acb->acb_private = private;
2885 acb->acb_zio_dummy = zio_null(pio,
2886 spa, NULL, NULL, NULL, zio_flags);
2888 ASSERT(acb->acb_done != NULL);
2889 acb->acb_next = hdr->b_acb;
2891 add_reference(hdr, hash_lock, private);
2892 mutex_exit(hash_lock);
2895 mutex_exit(hash_lock);
2899 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2902 add_reference(hdr, hash_lock, private);
2904 * If this block is already in use, create a new
2905 * copy of the data so that we will be guaranteed
2906 * that arc_release() will always succeed.
2910 ASSERT(buf->b_data);
2911 if (HDR_BUF_AVAILABLE(hdr)) {
2912 ASSERT(buf->b_efunc == NULL);
2913 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2915 buf = arc_buf_clone(buf);
2917 } else if (*arc_flags & ARC_PREFETCH &&
2918 refcount_count(&hdr->b_refcnt) == 0) {
2919 hdr->b_flags |= ARC_PREFETCH;
2921 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2922 arc_access(hdr, hash_lock);
2923 if (*arc_flags & ARC_L2CACHE)
2924 hdr->b_flags |= ARC_L2CACHE;
2925 mutex_exit(hash_lock);
2926 ARCSTAT_BUMP(arcstat_hits);
2927 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2928 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2929 data, metadata, hits);
2932 done(NULL, buf, private);
2934 uint64_t size = BP_GET_LSIZE(bp);
2935 arc_callback_t *acb;
2938 boolean_t devw = B_FALSE;
2941 /* this block is not in the cache */
2942 arc_buf_hdr_t *exists;
2943 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2944 buf = arc_buf_alloc(spa, size, private, type);
2946 hdr->b_dva = *BP_IDENTITY(bp);
2947 hdr->b_birth = bp->blk_birth;
2948 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2949 exists = buf_hash_insert(hdr, &hash_lock);
2951 /* somebody beat us to the hash insert */
2952 mutex_exit(hash_lock);
2953 bzero(&hdr->b_dva, sizeof (dva_t));
2956 (void) arc_buf_remove_ref(buf, private);
2957 goto top; /* restart the IO request */
2959 /* if this is a prefetch, we don't have a reference */
2960 if (*arc_flags & ARC_PREFETCH) {
2961 (void) remove_reference(hdr, hash_lock,
2963 hdr->b_flags |= ARC_PREFETCH;
2965 if (*arc_flags & ARC_L2CACHE)
2966 hdr->b_flags |= ARC_L2CACHE;
2967 if (BP_GET_LEVEL(bp) > 0)
2968 hdr->b_flags |= ARC_INDIRECT;
2970 /* this block is in the ghost cache */
2971 ASSERT(GHOST_STATE(hdr->b_state));
2972 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2973 ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0);
2974 ASSERT(hdr->b_buf == NULL);
2976 /* if this is a prefetch, we don't have a reference */
2977 if (*arc_flags & ARC_PREFETCH)
2978 hdr->b_flags |= ARC_PREFETCH;
2980 add_reference(hdr, hash_lock, private);
2981 if (*arc_flags & ARC_L2CACHE)
2982 hdr->b_flags |= ARC_L2CACHE;
2983 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2986 buf->b_efunc = NULL;
2987 buf->b_private = NULL;
2990 arc_get_data_buf(buf);
2991 ASSERT(hdr->b_datacnt == 0);
2996 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2997 acb->acb_done = done;
2998 acb->acb_private = private;
3000 ASSERT(hdr->b_acb == NULL);
3002 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3005 * If the buffer has been evicted, migrate it to a present state
3006 * before issuing the I/O. Once we drop the hash-table lock,
3007 * the header will be marked as I/O in progress and have an
3008 * attached buffer. At this point, anybody who finds this
3009 * buffer ought to notice that it's legit but has a pending I/O.
3012 if (GHOST_STATE(hdr->b_state))
3013 arc_access(hdr, hash_lock);
3015 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3016 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3017 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3018 addr = hdr->b_l2hdr->b_daddr;
3020 * Lock out device removal.
3022 if (vdev_is_dead(vd) ||
3023 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3027 mutex_exit(hash_lock);
3029 ASSERT3U(hdr->b_size, ==, size);
3030 DTRACE_PROBE3(arc__miss, blkptr_t *, bp, uint64_t, size,
3032 ARCSTAT_BUMP(arcstat_misses);
3033 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3034 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3035 data, metadata, misses);
3037 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3039 * Read from the L2ARC if the following are true:
3040 * 1. The L2ARC vdev was previously cached.
3041 * 2. This buffer still has L2ARC metadata.
3042 * 3. This buffer isn't currently writing to the L2ARC.
3043 * 4. The L2ARC entry wasn't evicted, which may
3044 * also have invalidated the vdev.
3045 * 5. This isn't prefetch and l2arc_noprefetch is set.
3047 if (hdr->b_l2hdr != NULL &&
3048 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3049 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3050 l2arc_read_callback_t *cb;
3052 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3053 ARCSTAT_BUMP(arcstat_l2_hits);
3055 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3057 cb->l2rcb_buf = buf;
3058 cb->l2rcb_spa = spa;
3061 cb->l2rcb_flags = zio_flags;
3064 * l2arc read. The SCL_L2ARC lock will be
3065 * released by l2arc_read_done().
3067 rzio = zio_read_phys(pio, vd, addr, size,
3068 buf->b_data, ZIO_CHECKSUM_OFF,
3069 l2arc_read_done, cb, priority, zio_flags |
3070 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
3071 ZIO_FLAG_DONT_PROPAGATE |
3072 ZIO_FLAG_DONT_RETRY, B_FALSE);
3073 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3075 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
3077 if (*arc_flags & ARC_NOWAIT) {
3082 ASSERT(*arc_flags & ARC_WAIT);
3083 if (zio_wait(rzio) == 0)
3086 /* l2arc read error; goto zio_read() */
3088 DTRACE_PROBE1(l2arc__miss,
3089 arc_buf_hdr_t *, hdr);
3090 ARCSTAT_BUMP(arcstat_l2_misses);
3091 if (HDR_L2_WRITING(hdr))
3092 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3093 spa_config_exit(spa, SCL_L2ARC, vd);
3097 spa_config_exit(spa, SCL_L2ARC, vd);
3098 if (l2arc_ndev != 0) {
3099 DTRACE_PROBE1(l2arc__miss,
3100 arc_buf_hdr_t *, hdr);
3101 ARCSTAT_BUMP(arcstat_l2_misses);
3105 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3106 arc_read_done, buf, priority, zio_flags, zb);
3108 if (*arc_flags & ARC_WAIT)
3109 return (zio_wait(rzio));
3111 ASSERT(*arc_flags & ARC_NOWAIT);
3118 * arc_read() variant to support pool traversal. If the block is already
3119 * in the ARC, make a copy of it; otherwise, the caller will do the I/O.
3120 * The idea is that we don't want pool traversal filling up memory, but
3121 * if the ARC already has the data anyway, we shouldn't pay for the I/O.
3124 arc_tryread(spa_t *spa, blkptr_t *bp, void *data)
3128 uint64_t guid = spa_guid(spa);
3131 hdr = buf_hash_find(guid, BP_IDENTITY(bp), bp->blk_birth, &hash_mtx);
3133 if (hdr && hdr->b_datacnt > 0 && !HDR_IO_IN_PROGRESS(hdr)) {
3134 arc_buf_t *buf = hdr->b_buf;
3137 while (buf->b_data == NULL) {
3141 bcopy(buf->b_data, data, hdr->b_size);
3147 mutex_exit(hash_mtx);
3153 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3155 ASSERT(buf->b_hdr != NULL);
3156 ASSERT(buf->b_hdr->b_state != arc_anon);
3157 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3158 buf->b_efunc = func;
3159 buf->b_private = private;
3163 * This is used by the DMU to let the ARC know that a buffer is
3164 * being evicted, so the ARC should clean up. If this arc buf
3165 * is not yet in the evicted state, it will be put there.
3168 arc_buf_evict(arc_buf_t *buf)
3171 kmutex_t *hash_lock;
3173 list_t *list, *evicted_list;
3174 kmutex_t *lock, *evicted_lock;
3176 rw_enter(&buf->b_lock, RW_WRITER);
3180 * We are in arc_do_user_evicts().
3182 ASSERT(buf->b_data == NULL);
3183 rw_exit(&buf->b_lock);
3185 } else if (buf->b_data == NULL) {
3186 arc_buf_t copy = *buf; /* structure assignment */
3188 * We are on the eviction list; process this buffer now
3189 * but let arc_do_user_evicts() do the reaping.
3191 buf->b_efunc = NULL;
3192 rw_exit(&buf->b_lock);
3193 VERIFY(copy.b_efunc(©) == 0);
3196 hash_lock = HDR_LOCK(hdr);
3197 mutex_enter(hash_lock);
3199 ASSERT(buf->b_hdr == hdr);
3200 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3201 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3204 * Pull this buffer off of the hdr
3207 while (*bufp != buf)
3208 bufp = &(*bufp)->b_next;
3209 *bufp = buf->b_next;
3211 ASSERT(buf->b_data != NULL);
3212 arc_buf_destroy(buf, FALSE, FALSE);
3214 if (hdr->b_datacnt == 0) {
3215 arc_state_t *old_state = hdr->b_state;
3216 arc_state_t *evicted_state;
3218 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3221 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3223 get_buf_info(hdr, old_state, &list, &lock);
3224 get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock);
3226 mutex_enter(evicted_lock);
3228 arc_change_state(evicted_state, hdr, hash_lock);
3229 ASSERT(HDR_IN_HASH_TABLE(hdr));
3230 hdr->b_flags |= ARC_IN_HASH_TABLE;
3231 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3233 mutex_exit(evicted_lock);
3236 mutex_exit(hash_lock);
3237 rw_exit(&buf->b_lock);
3239 VERIFY(buf->b_efunc(buf) == 0);
3240 buf->b_efunc = NULL;
3241 buf->b_private = NULL;
3243 kmem_cache_free(buf_cache, buf);
3248 * Release this buffer from the cache. This must be done
3249 * after a read and prior to modifying the buffer contents.
3250 * If the buffer has more than one reference, we must make
3251 * a new hdr for the buffer.
3254 arc_release(arc_buf_t *buf, void *tag)
3257 kmutex_t *hash_lock;
3258 l2arc_buf_hdr_t *l2hdr;
3260 boolean_t released = B_FALSE;
3262 rw_enter(&buf->b_lock, RW_WRITER);
3265 /* this buffer is not on any list */
3266 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3267 ASSERT(!(hdr->b_flags & ARC_STORED));
3269 if (hdr->b_state == arc_anon) {
3270 /* this buffer is already released */
3271 ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 1);
3272 ASSERT(BUF_EMPTY(hdr));
3273 ASSERT(buf->b_efunc == NULL);
3275 rw_exit(&buf->b_lock);
3278 hash_lock = HDR_LOCK(hdr);
3279 mutex_enter(hash_lock);
3282 l2hdr = hdr->b_l2hdr;
3284 mutex_enter(&l2arc_buflist_mtx);
3285 hdr->b_l2hdr = NULL;
3286 buf_size = hdr->b_size;
3293 * Do we have more than one buf?
3295 if (hdr->b_datacnt > 1) {
3296 arc_buf_hdr_t *nhdr;
3298 uint64_t blksz = hdr->b_size;
3299 uint64_t spa = hdr->b_spa;
3300 arc_buf_contents_t type = hdr->b_type;
3301 uint32_t flags = hdr->b_flags;
3303 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3305 * Pull the data off of this buf and attach it to
3306 * a new anonymous buf.
3308 (void) remove_reference(hdr, hash_lock, tag);
3310 while (*bufp != buf)
3311 bufp = &(*bufp)->b_next;
3312 *bufp = (*bufp)->b_next;
3315 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3316 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3317 if (refcount_is_zero(&hdr->b_refcnt)) {
3318 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3319 ASSERT3U(*size, >=, hdr->b_size);
3320 atomic_add_64(size, -hdr->b_size);
3322 hdr->b_datacnt -= 1;
3323 arc_cksum_verify(buf);
3325 mutex_exit(hash_lock);
3327 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3328 nhdr->b_size = blksz;
3330 nhdr->b_type = type;
3332 nhdr->b_state = arc_anon;
3333 nhdr->b_arc_access = 0;
3334 nhdr->b_flags = flags & ARC_L2_WRITING;
3335 nhdr->b_l2hdr = NULL;
3336 nhdr->b_datacnt = 1;
3337 nhdr->b_freeze_cksum = NULL;
3338 (void) refcount_add(&nhdr->b_refcnt, tag);
3340 rw_exit(&buf->b_lock);
3341 atomic_add_64(&arc_anon->arcs_size, blksz);
3343 rw_exit(&buf->b_lock);
3344 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3345 ASSERT(!list_link_active(&hdr->b_arc_node));
3346 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3347 arc_change_state(arc_anon, hdr, hash_lock);
3348 hdr->b_arc_access = 0;
3349 mutex_exit(hash_lock);
3351 bzero(&hdr->b_dva, sizeof (dva_t));
3356 buf->b_efunc = NULL;
3357 buf->b_private = NULL;
3361 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3362 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3363 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3364 mutex_exit(&l2arc_buflist_mtx);
3369 arc_released(arc_buf_t *buf)
3373 rw_enter(&buf->b_lock, RW_READER);
3374 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3375 rw_exit(&buf->b_lock);
3380 arc_has_callback(arc_buf_t *buf)
3384 rw_enter(&buf->b_lock, RW_READER);
3385 callback = (buf->b_efunc != NULL);
3386 rw_exit(&buf->b_lock);
3392 arc_referenced(arc_buf_t *buf)
3396 rw_enter(&buf->b_lock, RW_READER);
3397 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3398 rw_exit(&buf->b_lock);
3399 return (referenced);
3404 arc_write_ready(zio_t *zio)
3406 arc_write_callback_t *callback = zio->io_private;
3407 arc_buf_t *buf = callback->awcb_buf;
3408 arc_buf_hdr_t *hdr = buf->b_hdr;
3410 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3411 callback->awcb_ready(zio, buf, callback->awcb_private);
3414 * If the IO is already in progress, then this is a re-write
3415 * attempt, so we need to thaw and re-compute the cksum.
3416 * It is the responsibility of the callback to handle the
3417 * accounting for any re-write attempt.
3419 if (HDR_IO_IN_PROGRESS(hdr)) {
3420 mutex_enter(&hdr->b_freeze_lock);
3421 if (hdr->b_freeze_cksum != NULL) {
3422 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3423 hdr->b_freeze_cksum = NULL;
3425 mutex_exit(&hdr->b_freeze_lock);
3427 arc_cksum_compute(buf, B_FALSE);
3428 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3432 arc_write_done(zio_t *zio)
3434 arc_write_callback_t *callback = zio->io_private;
3435 arc_buf_t *buf = callback->awcb_buf;
3436 arc_buf_hdr_t *hdr = buf->b_hdr;
3440 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3441 hdr->b_birth = zio->io_bp->blk_birth;
3442 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3444 * If the block to be written was all-zero, we may have
3445 * compressed it away. In this case no write was performed
3446 * so there will be no dva/birth-date/checksum. The buffer
3447 * must therefor remain anonymous (and uncached).
3449 if (!BUF_EMPTY(hdr)) {
3450 arc_buf_hdr_t *exists;
3451 kmutex_t *hash_lock;
3453 arc_cksum_verify(buf);
3455 exists = buf_hash_insert(hdr, &hash_lock);
3458 * This can only happen if we overwrite for
3459 * sync-to-convergence, because we remove
3460 * buffers from the hash table when we arc_free().
3462 ASSERT(zio->io_flags & ZIO_FLAG_IO_REWRITE);
3463 ASSERT(DVA_EQUAL(BP_IDENTITY(&zio->io_bp_orig),
3464 BP_IDENTITY(zio->io_bp)));
3465 ASSERT3U(zio->io_bp_orig.blk_birth, ==,
3466 zio->io_bp->blk_birth);
3468 ASSERT(refcount_is_zero(&exists->b_refcnt));
3469 arc_change_state(arc_anon, exists, hash_lock);
3470 mutex_exit(hash_lock);
3471 arc_hdr_destroy(exists);
3472 exists = buf_hash_insert(hdr, &hash_lock);
3473 ASSERT3P(exists, ==, NULL);
3475 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3476 /* if it's not anon, we are doing a scrub */
3477 if (hdr->b_state == arc_anon)
3478 arc_access(hdr, hash_lock);
3479 mutex_exit(hash_lock);
3480 } else if (callback->awcb_done == NULL) {
3483 * This is an anonymous buffer with no user callback,
3484 * destroy it if there are no active references.
3486 mutex_enter(&arc_eviction_mtx);
3487 destroy_hdr = refcount_is_zero(&hdr->b_refcnt);
3488 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3489 mutex_exit(&arc_eviction_mtx);
3491 arc_hdr_destroy(hdr);
3493 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3495 hdr->b_flags &= ~ARC_STORED;
3497 if (callback->awcb_done) {
3498 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3499 callback->awcb_done(zio, buf, callback->awcb_private);
3502 kmem_free(callback, sizeof (arc_write_callback_t));
3506 write_policy(spa_t *spa, const writeprops_t *wp, zio_prop_t *zp)
3508 boolean_t ismd = (wp->wp_level > 0 || dmu_ot[wp->wp_type].ot_metadata);
3510 /* Determine checksum setting */
3513 * Metadata always gets checksummed. If the data
3514 * checksum is multi-bit correctable, and it's not a
3515 * ZBT-style checksum, then it's suitable for metadata
3516 * as well. Otherwise, the metadata checksum defaults
3519 if (zio_checksum_table[wp->wp_oschecksum].ci_correctable &&
3520 !zio_checksum_table[wp->wp_oschecksum].ci_zbt)
3521 zp->zp_checksum = wp->wp_oschecksum;
3523 zp->zp_checksum = ZIO_CHECKSUM_FLETCHER_4;
3525 zp->zp_checksum = zio_checksum_select(wp->wp_dnchecksum,
3529 /* Determine compression setting */
3532 * XXX -- we should design a compression algorithm
3533 * that specializes in arrays of bps.
3535 zp->zp_compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
3538 zp->zp_compress = zio_compress_select(wp->wp_dncompress,
3542 zp->zp_type = wp->wp_type;
3543 zp->zp_level = wp->wp_level;
3544 zp->zp_ndvas = MIN(wp->wp_copies + ismd, spa_max_replication(spa));
3548 arc_write(zio_t *pio, spa_t *spa, const writeprops_t *wp,
3549 boolean_t l2arc, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
3550 arc_done_func_t *ready, arc_done_func_t *done, void *private, int priority,
3551 int zio_flags, const zbookmark_t *zb)
3553 arc_buf_hdr_t *hdr = buf->b_hdr;
3554 arc_write_callback_t *callback;
3558 ASSERT(ready != NULL);
3559 ASSERT(!HDR_IO_ERROR(hdr));
3560 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3561 ASSERT(hdr->b_acb == 0);
3563 hdr->b_flags |= ARC_L2CACHE;
3564 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3565 callback->awcb_ready = ready;
3566 callback->awcb_done = done;
3567 callback->awcb_private = private;
3568 callback->awcb_buf = buf;
3570 write_policy(spa, wp, &zp);
3571 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, &zp,
3572 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3578 arc_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
3579 zio_done_func_t *done, void *private, uint32_t arc_flags)
3582 kmutex_t *hash_lock;
3584 uint64_t guid = spa_guid(spa);
3587 * If this buffer is in the cache, release it, so it
3590 ab = buf_hash_find(guid, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
3593 * The checksum of blocks to free is not always
3594 * preserved (eg. on the deadlist). However, if it is
3595 * nonzero, it should match what we have in the cache.
3597 ASSERT(bp->blk_cksum.zc_word[0] == 0 ||
3598 bp->blk_cksum.zc_word[0] == ab->b_cksum0 ||
3599 bp->blk_fill == BLK_FILL_ALREADY_FREED);
3601 if (ab->b_state != arc_anon)
3602 arc_change_state(arc_anon, ab, hash_lock);
3603 if (HDR_IO_IN_PROGRESS(ab)) {
3605 * This should only happen when we prefetch.
3607 ASSERT(ab->b_flags & ARC_PREFETCH);
3608 ASSERT3U(ab->b_datacnt, ==, 1);
3609 ab->b_flags |= ARC_FREED_IN_READ;
3610 if (HDR_IN_HASH_TABLE(ab))
3611 buf_hash_remove(ab);
3612 ab->b_arc_access = 0;
3613 bzero(&ab->b_dva, sizeof (dva_t));
3616 ab->b_buf->b_efunc = NULL;
3617 ab->b_buf->b_private = NULL;
3618 mutex_exit(hash_lock);
3619 } else if (refcount_is_zero(&ab->b_refcnt)) {
3620 ab->b_flags |= ARC_FREE_IN_PROGRESS;
3621 mutex_exit(hash_lock);
3622 arc_hdr_destroy(ab);
3623 ARCSTAT_BUMP(arcstat_deleted);
3626 * We still have an active reference on this
3627 * buffer. This can happen, e.g., from
3628 * dbuf_unoverride().
3630 ASSERT(!HDR_IN_HASH_TABLE(ab));
3631 ab->b_arc_access = 0;
3632 bzero(&ab->b_dva, sizeof (dva_t));
3635 ab->b_buf->b_efunc = NULL;
3636 ab->b_buf->b_private = NULL;
3637 mutex_exit(hash_lock);
3641 zio = zio_free(pio, spa, txg, bp, done, private, ZIO_FLAG_MUSTSUCCEED);
3643 if (arc_flags & ARC_WAIT)
3644 return (zio_wait(zio));
3646 ASSERT(arc_flags & ARC_NOWAIT);
3653 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3656 uint64_t available_memory = ptoa((uintmax_t)cnt.v_free_count
3657 + cnt.v_cache_count);
3658 static uint64_t page_load = 0;
3659 static uint64_t last_txg = 0;
3664 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3667 if (available_memory >= zfs_write_limit_max)
3670 if (txg > last_txg) {
3675 * If we are in pageout, we know that memory is already tight,
3676 * the arc is already going to be evicting, so we just want to
3677 * continue to let page writes occur as quickly as possible.
3679 if (curproc == pageproc) {
3680 if (page_load > available_memory / 4)
3682 /* Note: reserve is inflated, so we deflate */
3683 page_load += reserve / 8;
3685 } else if (page_load > 0 && arc_reclaim_needed()) {
3686 /* memory is low, delay before restarting */
3687 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3692 if (arc_size > arc_c_min) {
3693 uint64_t evictable_memory =
3694 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3695 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3696 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3697 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3698 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3701 if (inflight_data > available_memory / 4) {
3702 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3710 arc_tempreserve_clear(uint64_t reserve)
3712 atomic_add_64(&arc_tempreserve, -reserve);
3713 ASSERT((int64_t)arc_tempreserve >= 0);
3717 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3724 * Once in a while, fail for no reason. Everything should cope.
3726 if (spa_get_random(10000) == 0) {
3727 dprintf("forcing random failure\n");
3731 if (reserve > arc_c/4 && !arc_no_grow)
3732 arc_c = MIN(arc_c_max, reserve * 4);
3733 if (reserve > arc_c)
3737 * Don't count loaned bufs as in flight dirty data to prevent long
3738 * network delays from blocking transactions that are ready to be
3739 * assigned to a txg.
3741 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3744 * Writes will, almost always, require additional memory allocations
3745 * in order to compress/encrypt/etc the data. We therefor need to
3746 * make sure that there is sufficient available memory for this.
3748 if (error = arc_memory_throttle(reserve, anon_size, txg))
3752 * Throttle writes when the amount of dirty data in the cache
3753 * gets too large. We try to keep the cache less than half full
3754 * of dirty blocks so that our sync times don't grow too large.
3755 * Note: if two requests come in concurrently, we might let them
3756 * both succeed, when one of them should fail. Not a huge deal.
3759 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3760 anon_size > arc_c / 4) {
3761 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3762 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3763 arc_tempreserve>>10,
3764 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3765 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3766 reserve>>10, arc_c>>10);
3769 atomic_add_64(&arc_tempreserve, reserve);
3773 static kmutex_t arc_lowmem_lock;
3775 static eventhandler_tag arc_event_lowmem = NULL;
3778 arc_lowmem(void *arg __unused, int howto __unused)
3781 /* Serialize access via arc_lowmem_lock. */
3782 mutex_enter(&arc_lowmem_lock);
3784 cv_signal(&arc_reclaim_thr_cv);
3786 tsleep(&needfree, 0, "zfs:lowmem", hz / 5);
3787 mutex_exit(&arc_lowmem_lock);
3794 int prefetch_tunable_set = 0;
3797 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3798 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3799 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL);
3801 /* Convert seconds to clock ticks */
3802 arc_min_prefetch_lifespan = 1 * hz;
3804 /* Start out with 1/8 of all memory */
3805 arc_c = kmem_size() / 8;
3809 * On architectures where the physical memory can be larger
3810 * than the addressable space (intel in 32-bit mode), we may
3811 * need to limit the cache to 1/8 of VM size.
3813 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3816 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
3817 arc_c_min = MAX(arc_c / 4, 64<<18);
3818 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
3819 if (arc_c * 8 >= 1<<30)
3820 arc_c_max = (arc_c * 8) - (1<<30);
3822 arc_c_max = arc_c_min;
3823 arc_c_max = MAX(arc_c * 5, arc_c_max);
3826 * Allow the tunables to override our calculations if they are
3827 * reasonable (ie. over 16MB)
3829 if (zfs_arc_max >= 64<<18 && zfs_arc_max < kmem_size())
3830 arc_c_max = zfs_arc_max;
3831 if (zfs_arc_min >= 64<<18 && zfs_arc_min <= arc_c_max)
3832 arc_c_min = zfs_arc_min;
3835 arc_p = (arc_c >> 1);
3837 /* limit meta-data to 1/4 of the arc capacity */
3838 arc_meta_limit = arc_c_max / 4;
3840 /* Allow the tunable to override if it is reasonable */
3841 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3842 arc_meta_limit = zfs_arc_meta_limit;
3844 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3845 arc_c_min = arc_meta_limit / 2;
3847 if (zfs_arc_grow_retry > 0)
3848 arc_grow_retry = zfs_arc_grow_retry;
3850 if (zfs_arc_shrink_shift > 0)
3851 arc_shrink_shift = zfs_arc_shrink_shift;
3853 if (zfs_arc_p_min_shift > 0)
3854 arc_p_min_shift = zfs_arc_p_min_shift;
3856 /* if kmem_flags are set, lets try to use less memory */
3857 if (kmem_debugging())
3859 if (arc_c < arc_c_min)
3862 zfs_arc_min = arc_c_min;
3863 zfs_arc_max = arc_c_max;
3865 arc_anon = &ARC_anon;
3867 arc_mru_ghost = &ARC_mru_ghost;
3869 arc_mfu_ghost = &ARC_mfu_ghost;
3870 arc_l2c_only = &ARC_l2c_only;
3873 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
3874 mutex_init(&arc_anon->arcs_locks[i].arcs_lock,
3875 NULL, MUTEX_DEFAULT, NULL);
3876 mutex_init(&arc_mru->arcs_locks[i].arcs_lock,
3877 NULL, MUTEX_DEFAULT, NULL);
3878 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock,
3879 NULL, MUTEX_DEFAULT, NULL);
3880 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock,
3881 NULL, MUTEX_DEFAULT, NULL);
3882 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock,
3883 NULL, MUTEX_DEFAULT, NULL);
3884 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock,
3885 NULL, MUTEX_DEFAULT, NULL);
3887 list_create(&arc_mru->arcs_lists[i],
3888 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3889 list_create(&arc_mru_ghost->arcs_lists[i],
3890 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3891 list_create(&arc_mfu->arcs_lists[i],
3892 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3893 list_create(&arc_mfu_ghost->arcs_lists[i],
3894 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3895 list_create(&arc_mfu_ghost->arcs_lists[i],
3896 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3897 list_create(&arc_l2c_only->arcs_lists[i],
3898 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3903 arc_thread_exit = 0;
3904 arc_eviction_list = NULL;
3905 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3906 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3908 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3909 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3911 if (arc_ksp != NULL) {
3912 arc_ksp->ks_data = &arc_stats;
3913 kstat_install(arc_ksp);
3916 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3917 TS_RUN, minclsyspri);
3920 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
3921 EVENTHANDLER_PRI_FIRST);
3927 if (zfs_write_limit_max == 0)
3928 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3930 zfs_write_limit_shift = 0;
3931 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3934 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
3935 prefetch_tunable_set = 1;
3938 if (prefetch_tunable_set == 0) {
3939 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
3941 printf(" add \"vfs.zfs.prefetch_disable=0\" "
3942 "to /boot/loader.conf.\n");
3943 zfs_prefetch_disable=1;
3946 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
3947 prefetch_tunable_set == 0) {
3948 printf("ZFS NOTICE: Prefetch is disabled by default if less "
3949 "than 4GB of RAM is present;\n"
3950 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
3951 "to /boot/loader.conf.\n");
3952 zfs_prefetch_disable=1;
3955 /* Warn about ZFS memory and address space requirements. */
3956 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
3957 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
3958 "expect unstable behavior.\n");
3960 if (kmem_size() < 512 * (1 << 20)) {
3961 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
3962 "expect unstable behavior.\n");
3963 printf(" Consider tuning vm.kmem_size and "
3964 "vm.kmem_size_max\n");
3965 printf(" in /boot/loader.conf.\n");
3975 mutex_enter(&arc_reclaim_thr_lock);
3976 arc_thread_exit = 1;
3977 cv_signal(&arc_reclaim_thr_cv);
3978 while (arc_thread_exit != 0)
3979 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3980 mutex_exit(&arc_reclaim_thr_lock);
3986 if (arc_ksp != NULL) {
3987 kstat_delete(arc_ksp);
3991 mutex_destroy(&arc_eviction_mtx);
3992 mutex_destroy(&arc_reclaim_thr_lock);
3993 cv_destroy(&arc_reclaim_thr_cv);
3995 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
3996 list_destroy(&arc_mru->arcs_lists[i]);
3997 list_destroy(&arc_mru_ghost->arcs_lists[i]);
3998 list_destroy(&arc_mfu->arcs_lists[i]);
3999 list_destroy(&arc_mfu_ghost->arcs_lists[i]);
4000 list_destroy(&arc_l2c_only->arcs_lists[i]);
4002 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock);
4003 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock);
4004 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock);
4005 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock);
4006 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock);
4007 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock);
4010 mutex_destroy(&zfs_write_limit_lock);
4014 ASSERT(arc_loaned_bytes == 0);
4016 mutex_destroy(&arc_lowmem_lock);
4018 if (arc_event_lowmem != NULL)
4019 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
4026 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
4027 * It uses dedicated storage devices to hold cached data, which are populated
4028 * using large infrequent writes. The main role of this cache is to boost
4029 * the performance of random read workloads. The intended L2ARC devices
4030 * include short-stroked disks, solid state disks, and other media with
4031 * substantially faster read latency than disk.
4033 * +-----------------------+
4035 * +-----------------------+
4038 * l2arc_feed_thread() arc_read()
4042 * +---------------+ |
4044 * +---------------+ |
4049 * +-------+ +-------+
4051 * | cache | | cache |
4052 * +-------+ +-------+
4053 * +=========+ .-----.
4054 * : L2ARC : |-_____-|
4055 * : devices : | Disks |
4056 * +=========+ `-_____-'
4058 * Read requests are satisfied from the following sources, in order:
4061 * 2) vdev cache of L2ARC devices
4063 * 4) vdev cache of disks
4066 * Some L2ARC device types exhibit extremely slow write performance.
4067 * To accommodate for this there are some significant differences between
4068 * the L2ARC and traditional cache design:
4070 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4071 * the ARC behave as usual, freeing buffers and placing headers on ghost
4072 * lists. The ARC does not send buffers to the L2ARC during eviction as
4073 * this would add inflated write latencies for all ARC memory pressure.
4075 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4076 * It does this by periodically scanning buffers from the eviction-end of
4077 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4078 * not already there. It scans until a headroom of buffers is satisfied,
4079 * which itself is a buffer for ARC eviction. The thread that does this is
4080 * l2arc_feed_thread(), illustrated below; example sizes are included to
4081 * provide a better sense of ratio than this diagram:
4084 * +---------------------+----------+
4085 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4086 * +---------------------+----------+ | o L2ARC eligible
4087 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4088 * +---------------------+----------+ |
4089 * 15.9 Gbytes ^ 32 Mbytes |
4091 * l2arc_feed_thread()
4093 * l2arc write hand <--[oooo]--'
4097 * +==============================+
4098 * L2ARC dev |####|#|###|###| |####| ... |
4099 * +==============================+
4102 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4103 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4104 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4105 * safe to say that this is an uncommon case, since buffers at the end of
4106 * the ARC lists have moved there due to inactivity.
4108 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4109 * then the L2ARC simply misses copying some buffers. This serves as a
4110 * pressure valve to prevent heavy read workloads from both stalling the ARC
4111 * with waits and clogging the L2ARC with writes. This also helps prevent
4112 * the potential for the L2ARC to churn if it attempts to cache content too
4113 * quickly, such as during backups of the entire pool.
4115 * 5. After system boot and before the ARC has filled main memory, there are
4116 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4117 * lists can remain mostly static. Instead of searching from tail of these
4118 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4119 * for eligible buffers, greatly increasing its chance of finding them.
4121 * The L2ARC device write speed is also boosted during this time so that
4122 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4123 * there are no L2ARC reads, and no fear of degrading read performance
4124 * through increased writes.
4126 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4127 * the vdev queue can aggregate them into larger and fewer writes. Each
4128 * device is written to in a rotor fashion, sweeping writes through
4129 * available space then repeating.
4131 * 7. The L2ARC does not store dirty content. It never needs to flush
4132 * write buffers back to disk based storage.
4134 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4135 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4137 * The performance of the L2ARC can be tweaked by a number of tunables, which
4138 * may be necessary for different workloads:
4140 * l2arc_write_max max write bytes per interval
4141 * l2arc_write_boost extra write bytes during device warmup
4142 * l2arc_noprefetch skip caching prefetched buffers
4143 * l2arc_headroom number of max device writes to precache
4144 * l2arc_feed_secs seconds between L2ARC writing
4146 * Tunables may be removed or added as future performance improvements are
4147 * integrated, and also may become zpool properties.
4149 * There are three key functions that control how the L2ARC warms up:
4151 * l2arc_write_eligible() check if a buffer is eligible to cache
4152 * l2arc_write_size() calculate how much to write
4153 * l2arc_write_interval() calculate sleep delay between writes
4155 * These three functions determine what to write, how much, and how quickly
4160 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4163 * A buffer is *not* eligible for the L2ARC if it:
4164 * 1. belongs to a different spa.
4165 * 2. is already cached on the L2ARC.
4166 * 3. has an I/O in progress (it may be an incomplete read).
4167 * 4. is flagged not eligible (zfs property).
4169 if (ab->b_spa != spa_guid) {
4170 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
4173 if (ab->b_l2hdr != NULL) {
4174 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
4177 if (HDR_IO_IN_PROGRESS(ab)) {
4178 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
4181 if (!HDR_L2CACHE(ab)) {
4182 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
4190 l2arc_write_size(l2arc_dev_t *dev)
4194 size = dev->l2ad_write;
4196 if (arc_warm == B_FALSE)
4197 size += dev->l2ad_boost;
4204 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4206 clock_t interval, next;
4209 * If the ARC lists are busy, increase our write rate; if the
4210 * lists are stale, idle back. This is achieved by checking
4211 * how much we previously wrote - if it was more than half of
4212 * what we wanted, schedule the next write much sooner.
4214 if (l2arc_feed_again && wrote > (wanted / 2))
4215 interval = (hz * l2arc_feed_min_ms) / 1000;
4217 interval = hz * l2arc_feed_secs;
4219 next = MAX(LBOLT, MIN(LBOLT + interval, began + interval));
4225 l2arc_hdr_stat_add(void)
4227 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4228 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4232 l2arc_hdr_stat_remove(void)
4234 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4235 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4239 * Cycle through L2ARC devices. This is how L2ARC load balances.
4240 * If a device is returned, this also returns holding the spa config lock.
4242 static l2arc_dev_t *
4243 l2arc_dev_get_next(void)
4245 l2arc_dev_t *first, *next = NULL;
4248 * Lock out the removal of spas (spa_namespace_lock), then removal
4249 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4250 * both locks will be dropped and a spa config lock held instead.
4252 mutex_enter(&spa_namespace_lock);
4253 mutex_enter(&l2arc_dev_mtx);
4255 /* if there are no vdevs, there is nothing to do */
4256 if (l2arc_ndev == 0)
4260 next = l2arc_dev_last;
4262 /* loop around the list looking for a non-faulted vdev */
4264 next = list_head(l2arc_dev_list);
4266 next = list_next(l2arc_dev_list, next);
4268 next = list_head(l2arc_dev_list);
4271 /* if we have come back to the start, bail out */
4274 else if (next == first)
4277 } while (vdev_is_dead(next->l2ad_vdev));
4279 /* if we were unable to find any usable vdevs, return NULL */
4280 if (vdev_is_dead(next->l2ad_vdev))
4283 l2arc_dev_last = next;
4286 mutex_exit(&l2arc_dev_mtx);
4289 * Grab the config lock to prevent the 'next' device from being
4290 * removed while we are writing to it.
4293 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4294 mutex_exit(&spa_namespace_lock);
4300 * Free buffers that were tagged for destruction.
4303 l2arc_do_free_on_write()
4306 l2arc_data_free_t *df, *df_prev;
4308 mutex_enter(&l2arc_free_on_write_mtx);
4309 buflist = l2arc_free_on_write;
4311 for (df = list_tail(buflist); df; df = df_prev) {
4312 df_prev = list_prev(buflist, df);
4313 ASSERT(df->l2df_data != NULL);
4314 ASSERT(df->l2df_func != NULL);
4315 df->l2df_func(df->l2df_data, df->l2df_size);
4316 list_remove(buflist, df);
4317 kmem_free(df, sizeof (l2arc_data_free_t));
4320 mutex_exit(&l2arc_free_on_write_mtx);
4324 * A write to a cache device has completed. Update all headers to allow
4325 * reads from these buffers to begin.
4328 l2arc_write_done(zio_t *zio)
4330 l2arc_write_callback_t *cb;
4333 arc_buf_hdr_t *head, *ab, *ab_prev;
4334 l2arc_buf_hdr_t *abl2;
4335 kmutex_t *hash_lock;
4337 cb = zio->io_private;
4339 dev = cb->l2wcb_dev;
4340 ASSERT(dev != NULL);
4341 head = cb->l2wcb_head;
4342 ASSERT(head != NULL);
4343 buflist = dev->l2ad_buflist;
4344 ASSERT(buflist != NULL);
4345 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4346 l2arc_write_callback_t *, cb);
4348 if (zio->io_error != 0)
4349 ARCSTAT_BUMP(arcstat_l2_writes_error);
4351 mutex_enter(&l2arc_buflist_mtx);
4354 * All writes completed, or an error was hit.
4356 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4357 ab_prev = list_prev(buflist, ab);
4359 hash_lock = HDR_LOCK(ab);
4360 if (!mutex_tryenter(hash_lock)) {
4362 * This buffer misses out. It may be in a stage
4363 * of eviction. Its ARC_L2_WRITING flag will be
4364 * left set, denying reads to this buffer.
4366 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4370 if (zio->io_error != 0) {
4372 * Error - drop L2ARC entry.
4374 list_remove(buflist, ab);
4377 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4378 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4382 * Allow ARC to begin reads to this L2ARC entry.
4384 ab->b_flags &= ~ARC_L2_WRITING;
4386 mutex_exit(hash_lock);
4389 atomic_inc_64(&l2arc_writes_done);
4390 list_remove(buflist, head);
4391 kmem_cache_free(hdr_cache, head);
4392 mutex_exit(&l2arc_buflist_mtx);
4394 l2arc_do_free_on_write();
4396 kmem_free(cb, sizeof (l2arc_write_callback_t));
4400 * A read to a cache device completed. Validate buffer contents before
4401 * handing over to the regular ARC routines.
4404 l2arc_read_done(zio_t *zio)
4406 l2arc_read_callback_t *cb;
4409 kmutex_t *hash_lock;
4412 ASSERT(zio->io_vd != NULL);
4413 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4415 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4417 cb = zio->io_private;
4419 buf = cb->l2rcb_buf;
4420 ASSERT(buf != NULL);
4422 ASSERT(hdr != NULL);
4424 hash_lock = HDR_LOCK(hdr);
4425 mutex_enter(hash_lock);
4428 * Check this survived the L2ARC journey.
4430 equal = arc_cksum_equal(buf);
4431 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4432 mutex_exit(hash_lock);
4433 zio->io_private = buf;
4434 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4435 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4438 mutex_exit(hash_lock);
4440 * Buffer didn't survive caching. Increment stats and
4441 * reissue to the original storage device.
4443 if (zio->io_error != 0) {
4444 ARCSTAT_BUMP(arcstat_l2_io_error);
4446 zio->io_error = EIO;
4449 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4452 * If there's no waiter, issue an async i/o to the primary
4453 * storage now. If there *is* a waiter, the caller must
4454 * issue the i/o in a context where it's OK to block.
4456 if (zio->io_waiter == NULL) {
4457 zio_t *pio = zio_unique_parent(zio);
4459 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4461 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4462 buf->b_data, zio->io_size, arc_read_done, buf,
4463 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4467 kmem_free(cb, sizeof (l2arc_read_callback_t));
4471 * This is the list priority from which the L2ARC will search for pages to
4472 * cache. This is used within loops (0..3) to cycle through lists in the
4473 * desired order. This order can have a significant effect on cache
4476 * Currently the metadata lists are hit first, MFU then MRU, followed by
4477 * the data lists. This function returns a locked list, and also returns
4481 l2arc_list_locked(int list_num, kmutex_t **lock)
4486 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS);
4488 if (list_num < ARC_BUFC_NUMMETADATALISTS) {
4490 list = &arc_mfu->arcs_lists[idx];
4491 *lock = ARCS_LOCK(arc_mfu, idx);
4492 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) {
4493 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4494 list = &arc_mru->arcs_lists[idx];
4495 *lock = ARCS_LOCK(arc_mru, idx);
4496 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 +
4497 ARC_BUFC_NUMDATALISTS)) {
4498 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4499 list = &arc_mfu->arcs_lists[idx];
4500 *lock = ARCS_LOCK(arc_mfu, idx);
4502 idx = list_num - ARC_BUFC_NUMLISTS;
4503 list = &arc_mru->arcs_lists[idx];
4504 *lock = ARCS_LOCK(arc_mru, idx);
4507 ASSERT(!(MUTEX_HELD(*lock)));
4513 * Evict buffers from the device write hand to the distance specified in
4514 * bytes. This distance may span populated buffers, it may span nothing.
4515 * This is clearing a region on the L2ARC device ready for writing.
4516 * If the 'all' boolean is set, every buffer is evicted.
4519 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4522 l2arc_buf_hdr_t *abl2;
4523 arc_buf_hdr_t *ab, *ab_prev;
4524 kmutex_t *hash_lock;
4527 buflist = dev->l2ad_buflist;
4529 if (buflist == NULL)
4532 if (!all && dev->l2ad_first) {
4534 * This is the first sweep through the device. There is
4540 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4542 * When nearing the end of the device, evict to the end
4543 * before the device write hand jumps to the start.
4545 taddr = dev->l2ad_end;
4547 taddr = dev->l2ad_hand + distance;
4549 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4550 uint64_t, taddr, boolean_t, all);
4553 mutex_enter(&l2arc_buflist_mtx);
4554 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4555 ab_prev = list_prev(buflist, ab);
4557 hash_lock = HDR_LOCK(ab);
4558 if (!mutex_tryenter(hash_lock)) {
4560 * Missed the hash lock. Retry.
4562 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4563 mutex_exit(&l2arc_buflist_mtx);
4564 mutex_enter(hash_lock);
4565 mutex_exit(hash_lock);
4569 if (HDR_L2_WRITE_HEAD(ab)) {
4571 * We hit a write head node. Leave it for
4572 * l2arc_write_done().
4574 list_remove(buflist, ab);
4575 mutex_exit(hash_lock);
4579 if (!all && ab->b_l2hdr != NULL &&
4580 (ab->b_l2hdr->b_daddr > taddr ||
4581 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4583 * We've evicted to the target address,
4584 * or the end of the device.
4586 mutex_exit(hash_lock);
4590 if (HDR_FREE_IN_PROGRESS(ab)) {
4592 * Already on the path to destruction.
4594 mutex_exit(hash_lock);
4598 if (ab->b_state == arc_l2c_only) {
4599 ASSERT(!HDR_L2_READING(ab));
4601 * This doesn't exist in the ARC. Destroy.
4602 * arc_hdr_destroy() will call list_remove()
4603 * and decrement arcstat_l2_size.
4605 arc_change_state(arc_anon, ab, hash_lock);
4606 arc_hdr_destroy(ab);
4609 * Invalidate issued or about to be issued
4610 * reads, since we may be about to write
4611 * over this location.
4613 if (HDR_L2_READING(ab)) {
4614 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4615 ab->b_flags |= ARC_L2_EVICTED;
4619 * Tell ARC this no longer exists in L2ARC.
4621 if (ab->b_l2hdr != NULL) {
4624 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4625 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4627 list_remove(buflist, ab);
4630 * This may have been leftover after a
4633 ab->b_flags &= ~ARC_L2_WRITING;
4635 mutex_exit(hash_lock);
4637 mutex_exit(&l2arc_buflist_mtx);
4639 spa_l2cache_space_update(dev->l2ad_vdev, 0, -(taddr - dev->l2ad_evict));
4640 dev->l2ad_evict = taddr;
4644 * Find and write ARC buffers to the L2ARC device.
4646 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4647 * for reading until they have completed writing.
4650 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4652 arc_buf_hdr_t *ab, *ab_prev, *head;
4653 l2arc_buf_hdr_t *hdrl2;
4655 uint64_t passed_sz, write_sz, buf_sz, headroom;
4657 kmutex_t *hash_lock, *list_lock;
4658 boolean_t have_lock, full;
4659 l2arc_write_callback_t *cb;
4661 uint64_t guid = spa_guid(spa);
4664 ASSERT(dev->l2ad_vdev != NULL);
4669 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4670 head->b_flags |= ARC_L2_WRITE_HEAD;
4672 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
4674 * Copy buffers for L2ARC writing.
4676 mutex_enter(&l2arc_buflist_mtx);
4677 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) {
4678 list = l2arc_list_locked(try, &list_lock);
4680 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
4683 * L2ARC fast warmup.
4685 * Until the ARC is warm and starts to evict, read from the
4686 * head of the ARC lists rather than the tail.
4688 headroom = target_sz * l2arc_headroom;
4689 if (arc_warm == B_FALSE)
4690 ab = list_head(list);
4692 ab = list_tail(list);
4694 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
4696 for (; ab; ab = ab_prev) {
4697 if (arc_warm == B_FALSE)
4698 ab_prev = list_next(list, ab);
4700 ab_prev = list_prev(list, ab);
4701 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size);
4703 hash_lock = HDR_LOCK(ab);
4704 have_lock = MUTEX_HELD(hash_lock);
4705 if (!have_lock && !mutex_tryenter(hash_lock)) {
4706 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
4708 * Skip this buffer rather than waiting.
4713 passed_sz += ab->b_size;
4714 if (passed_sz > headroom) {
4718 mutex_exit(hash_lock);
4719 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
4723 if (!l2arc_write_eligible(guid, ab)) {
4724 mutex_exit(hash_lock);
4728 if ((write_sz + ab->b_size) > target_sz) {
4730 mutex_exit(hash_lock);
4731 ARCSTAT_BUMP(arcstat_l2_write_full);
4737 * Insert a dummy header on the buflist so
4738 * l2arc_write_done() can find where the
4739 * write buffers begin without searching.
4741 list_insert_head(dev->l2ad_buflist, head);
4744 sizeof (l2arc_write_callback_t), KM_SLEEP);
4745 cb->l2wcb_dev = dev;
4746 cb->l2wcb_head = head;
4747 pio = zio_root(spa, l2arc_write_done, cb,
4749 ARCSTAT_BUMP(arcstat_l2_write_pios);
4753 * Create and add a new L2ARC header.
4755 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4757 hdrl2->b_daddr = dev->l2ad_hand;
4759 ab->b_flags |= ARC_L2_WRITING;
4760 ab->b_l2hdr = hdrl2;
4761 list_insert_head(dev->l2ad_buflist, ab);
4762 buf_data = ab->b_buf->b_data;
4763 buf_sz = ab->b_size;
4766 * Compute and store the buffer cksum before
4767 * writing. On debug the cksum is verified first.
4769 arc_cksum_verify(ab->b_buf);
4770 arc_cksum_compute(ab->b_buf, B_TRUE);
4772 mutex_exit(hash_lock);
4774 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4775 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4776 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4777 ZIO_FLAG_CANFAIL, B_FALSE);
4779 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4781 (void) zio_nowait(wzio);
4784 * Keep the clock hand suitably device-aligned.
4786 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4789 dev->l2ad_hand += buf_sz;
4792 mutex_exit(list_lock);
4797 mutex_exit(&l2arc_buflist_mtx);
4800 ASSERT3U(write_sz, ==, 0);
4801 kmem_cache_free(hdr_cache, head);
4805 ASSERT3U(write_sz, <=, target_sz);
4806 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4807 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4808 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4809 spa_l2cache_space_update(dev->l2ad_vdev, 0, write_sz);
4812 * Bump device hand to the device start if it is approaching the end.
4813 * l2arc_evict() will already have evicted ahead for this case.
4815 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4816 spa_l2cache_space_update(dev->l2ad_vdev, 0,
4817 dev->l2ad_end - dev->l2ad_hand);
4818 dev->l2ad_hand = dev->l2ad_start;
4819 dev->l2ad_evict = dev->l2ad_start;
4820 dev->l2ad_first = B_FALSE;
4823 dev->l2ad_writing = B_TRUE;
4824 (void) zio_wait(pio);
4825 dev->l2ad_writing = B_FALSE;
4831 * This thread feeds the L2ARC at regular intervals. This is the beating
4832 * heart of the L2ARC.
4835 l2arc_feed_thread(void *dummy __unused)
4840 uint64_t size, wrote;
4841 clock_t begin, next = LBOLT;
4843 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4845 mutex_enter(&l2arc_feed_thr_lock);
4847 while (l2arc_thread_exit == 0) {
4848 CALLB_CPR_SAFE_BEGIN(&cpr);
4849 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4851 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4855 * Quick check for L2ARC devices.
4857 mutex_enter(&l2arc_dev_mtx);
4858 if (l2arc_ndev == 0) {
4859 mutex_exit(&l2arc_dev_mtx);
4862 mutex_exit(&l2arc_dev_mtx);
4866 * This selects the next l2arc device to write to, and in
4867 * doing so the next spa to feed from: dev->l2ad_spa. This
4868 * will return NULL if there are now no l2arc devices or if
4869 * they are all faulted.
4871 * If a device is returned, its spa's config lock is also
4872 * held to prevent device removal. l2arc_dev_get_next()
4873 * will grab and release l2arc_dev_mtx.
4875 if ((dev = l2arc_dev_get_next()) == NULL)
4878 spa = dev->l2ad_spa;
4879 ASSERT(spa != NULL);
4882 * Avoid contributing to memory pressure.
4884 if (arc_reclaim_needed()) {
4885 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4886 spa_config_exit(spa, SCL_L2ARC, dev);
4890 ARCSTAT_BUMP(arcstat_l2_feeds);
4892 size = l2arc_write_size(dev);
4895 * Evict L2ARC buffers that will be overwritten.
4897 l2arc_evict(dev, size, B_FALSE);
4900 * Write ARC buffers.
4902 wrote = l2arc_write_buffers(spa, dev, size);
4905 * Calculate interval between writes.
4907 next = l2arc_write_interval(begin, size, wrote);
4908 spa_config_exit(spa, SCL_L2ARC, dev);
4911 l2arc_thread_exit = 0;
4912 cv_broadcast(&l2arc_feed_thr_cv);
4913 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4918 l2arc_vdev_present(vdev_t *vd)
4922 mutex_enter(&l2arc_dev_mtx);
4923 for (dev = list_head(l2arc_dev_list); dev != NULL;
4924 dev = list_next(l2arc_dev_list, dev)) {
4925 if (dev->l2ad_vdev == vd)
4928 mutex_exit(&l2arc_dev_mtx);
4930 return (dev != NULL);
4934 * Add a vdev for use by the L2ARC. By this point the spa has already
4935 * validated the vdev and opened it.
4938 l2arc_add_vdev(spa_t *spa, vdev_t *vd, uint64_t start, uint64_t end)
4940 l2arc_dev_t *adddev;
4942 ASSERT(!l2arc_vdev_present(vd));
4945 * Create a new l2arc device entry.
4947 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4948 adddev->l2ad_spa = spa;
4949 adddev->l2ad_vdev = vd;
4950 adddev->l2ad_write = l2arc_write_max;
4951 adddev->l2ad_boost = l2arc_write_boost;
4952 adddev->l2ad_start = start;
4953 adddev->l2ad_end = end;
4954 adddev->l2ad_hand = adddev->l2ad_start;
4955 adddev->l2ad_evict = adddev->l2ad_start;
4956 adddev->l2ad_first = B_TRUE;
4957 adddev->l2ad_writing = B_FALSE;
4958 ASSERT3U(adddev->l2ad_write, >, 0);
4961 * This is a list of all ARC buffers that are still valid on the
4964 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4965 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4966 offsetof(arc_buf_hdr_t, b_l2node));
4968 spa_l2cache_space_update(vd, adddev->l2ad_end - adddev->l2ad_hand, 0);
4971 * Add device to global list
4973 mutex_enter(&l2arc_dev_mtx);
4974 list_insert_head(l2arc_dev_list, adddev);
4975 atomic_inc_64(&l2arc_ndev);
4976 mutex_exit(&l2arc_dev_mtx);
4980 * Remove a vdev from the L2ARC.
4983 l2arc_remove_vdev(vdev_t *vd)
4985 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4988 * Find the device by vdev
4990 mutex_enter(&l2arc_dev_mtx);
4991 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4992 nextdev = list_next(l2arc_dev_list, dev);
4993 if (vd == dev->l2ad_vdev) {
4998 ASSERT(remdev != NULL);
5001 * Remove device from global list
5003 list_remove(l2arc_dev_list, remdev);
5004 l2arc_dev_last = NULL; /* may have been invalidated */
5005 atomic_dec_64(&l2arc_ndev);
5006 mutex_exit(&l2arc_dev_mtx);
5009 * Clear all buflists and ARC references. L2ARC device flush.
5011 l2arc_evict(remdev, 0, B_TRUE);
5012 list_destroy(remdev->l2ad_buflist);
5013 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5014 kmem_free(remdev, sizeof (l2arc_dev_t));
5020 l2arc_thread_exit = 0;
5022 l2arc_writes_sent = 0;
5023 l2arc_writes_done = 0;
5025 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5026 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5027 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5028 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5029 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5031 l2arc_dev_list = &L2ARC_dev_list;
5032 l2arc_free_on_write = &L2ARC_free_on_write;
5033 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5034 offsetof(l2arc_dev_t, l2ad_node));
5035 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5036 offsetof(l2arc_data_free_t, l2df_list_node));
5043 * This is called from dmu_fini(), which is called from spa_fini();
5044 * Because of this, we can assume that all l2arc devices have
5045 * already been removed when the pools themselves were removed.
5048 l2arc_do_free_on_write();
5050 mutex_destroy(&l2arc_feed_thr_lock);
5051 cv_destroy(&l2arc_feed_thr_cv);
5052 mutex_destroy(&l2arc_dev_mtx);
5053 mutex_destroy(&l2arc_buflist_mtx);
5054 mutex_destroy(&l2arc_free_on_write_mtx);
5056 list_destroy(l2arc_dev_list);
5057 list_destroy(l2arc_free_on_write);
5063 if (!(spa_mode_global & FWRITE))
5066 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5067 TS_RUN, minclsyspri);
5073 if (!(spa_mode_global & FWRITE))
5076 mutex_enter(&l2arc_feed_thr_lock);
5077 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5078 l2arc_thread_exit = 1;
5079 while (l2arc_thread_exit != 0)
5080 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5081 mutex_exit(&l2arc_feed_thr_lock);