4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
25 * Copyright 2014 Nexenta Systems, Inc. All rights reserved.
29 * DVA-based Adjustable Replacement Cache
31 * While much of the theory of operation used here is
32 * based on the self-tuning, low overhead replacement cache
33 * presented by Megiddo and Modha at FAST 2003, there are some
34 * significant differences:
36 * 1. The Megiddo and Modha model assumes any page is evictable.
37 * Pages in its cache cannot be "locked" into memory. This makes
38 * the eviction algorithm simple: evict the last page in the list.
39 * This also make the performance characteristics easy to reason
40 * about. Our cache is not so simple. At any given moment, some
41 * subset of the blocks in the cache are un-evictable because we
42 * have handed out a reference to them. Blocks are only evictable
43 * when there are no external references active. This makes
44 * eviction far more problematic: we choose to evict the evictable
45 * blocks that are the "lowest" in the list.
47 * There are times when it is not possible to evict the requested
48 * space. In these circumstances we are unable to adjust the cache
49 * size. To prevent the cache growing unbounded at these times we
50 * implement a "cache throttle" that slows the flow of new data
51 * into the cache until we can make space available.
53 * 2. The Megiddo and Modha model assumes a fixed cache size.
54 * Pages are evicted when the cache is full and there is a cache
55 * miss. Our model has a variable sized cache. It grows with
56 * high use, but also tries to react to memory pressure from the
57 * operating system: decreasing its size when system memory is
60 * 3. The Megiddo and Modha model assumes a fixed page size. All
61 * elements of the cache are therefore exactly the same size. So
62 * when adjusting the cache size following a cache miss, its simply
63 * a matter of choosing a single page to evict. In our model, we
64 * have variable sized cache blocks (rangeing from 512 bytes to
65 * 128K bytes). We therefore choose a set of blocks to evict to make
66 * space for a cache miss that approximates as closely as possible
67 * the space used by the new block.
69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
70 * by N. Megiddo & D. Modha, FAST 2003
76 * A new reference to a cache buffer can be obtained in two
77 * ways: 1) via a hash table lookup using the DVA as a key,
78 * or 2) via one of the ARC lists. The arc_read() interface
79 * uses method 1, while the internal arc algorithms for
80 * adjusting the cache use method 2. We therefore provide two
81 * types of locks: 1) the hash table lock array, and 2) the
84 * Buffers do not have their own mutexs, rather they rely on the
85 * hash table mutexs for the bulk of their protection (i.e. most
86 * fields in the arc_buf_hdr_t are protected by these mutexs).
88 * buf_hash_find() returns the appropriate mutex (held) when it
89 * locates the requested buffer in the hash table. It returns
90 * NULL for the mutex if the buffer was not in the table.
92 * buf_hash_remove() expects the appropriate hash mutex to be
93 * already held before it is invoked.
95 * Each arc state also has a mutex which is used to protect the
96 * buffer list associated with the state. When attempting to
97 * obtain a hash table lock while holding an arc list lock you
98 * must use: mutex_tryenter() to avoid deadlock. Also note that
99 * the active state mutex must be held before the ghost state mutex.
101 * Arc buffers may have an associated eviction callback function.
102 * This function will be invoked prior to removing the buffer (e.g.
103 * in arc_do_user_evicts()). Note however that the data associated
104 * with the buffer may be evicted prior to the callback. The callback
105 * must be made with *no locks held* (to prevent deadlock). Additionally,
106 * the users of callbacks must ensure that their private data is
107 * protected from simultaneous callbacks from arc_buf_evict()
108 * and arc_do_user_evicts().
110 * Note that the majority of the performance stats are manipulated
111 * with atomic operations.
113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
115 * - L2ARC buflist creation
116 * - L2ARC buflist eviction
117 * - L2ARC write completion, which walks L2ARC buflists
118 * - ARC header destruction, as it removes from L2ARC buflists
119 * - ARC header release, as it removes from L2ARC buflists
124 #include <sys/zio_compress.h>
125 #include <sys/zfs_context.h>
127 #include <sys/refcount.h>
128 #include <sys/vdev.h>
129 #include <sys/vdev_impl.h>
130 #include <sys/dsl_pool.h>
132 #include <sys/dnlc.h>
134 #include <sys/callb.h>
135 #include <sys/kstat.h>
136 #include <sys/trim_map.h>
137 #include <zfs_fletcher.h>
140 #include <vm/vm_pageout.h>
144 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
145 boolean_t arc_watch = B_FALSE;
150 static kmutex_t arc_reclaim_thr_lock;
151 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
152 static uint8_t arc_thread_exit;
154 #define ARC_REDUCE_DNLC_PERCENT 3
155 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
157 typedef enum arc_reclaim_strategy {
158 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
159 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
160 } arc_reclaim_strategy_t;
163 * The number of iterations through arc_evict_*() before we
164 * drop & reacquire the lock.
166 int arc_evict_iterations = 100;
168 /* number of seconds before growing cache again */
169 static int arc_grow_retry = 60;
171 /* shift of arc_c for calculating both min and max arc_p */
172 static int arc_p_min_shift = 4;
174 /* log2(fraction of arc to reclaim) */
175 static int arc_shrink_shift = 5;
178 * minimum lifespan of a prefetch block in clock ticks
179 * (initialized in arc_init())
181 static int arc_min_prefetch_lifespan;
184 * If this percent of memory is free, don't throttle.
186 int arc_lotsfree_percent = 10;
189 extern int zfs_prefetch_disable;
192 * The arc has filled available memory and has now warmed up.
194 static boolean_t arc_warm;
197 * These tunables are for performance analysis.
199 uint64_t zfs_arc_max;
200 uint64_t zfs_arc_min;
201 uint64_t zfs_arc_meta_limit = 0;
202 int zfs_arc_grow_retry = 0;
203 int zfs_arc_shrink_shift = 0;
204 int zfs_arc_p_min_shift = 0;
205 int zfs_disable_dup_eviction = 0;
207 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
208 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
209 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
210 SYSCTL_DECL(_vfs_zfs);
211 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
213 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
217 * Note that buffers can be in one of 6 states:
218 * ARC_anon - anonymous (discussed below)
219 * ARC_mru - recently used, currently cached
220 * ARC_mru_ghost - recentely used, no longer in cache
221 * ARC_mfu - frequently used, currently cached
222 * ARC_mfu_ghost - frequently used, no longer in cache
223 * ARC_l2c_only - exists in L2ARC but not other states
224 * When there are no active references to the buffer, they are
225 * are linked onto a list in one of these arc states. These are
226 * the only buffers that can be evicted or deleted. Within each
227 * state there are multiple lists, one for meta-data and one for
228 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
229 * etc.) is tracked separately so that it can be managed more
230 * explicitly: favored over data, limited explicitly.
232 * Anonymous buffers are buffers that are not associated with
233 * a DVA. These are buffers that hold dirty block copies
234 * before they are written to stable storage. By definition,
235 * they are "ref'd" and are considered part of arc_mru
236 * that cannot be freed. Generally, they will aquire a DVA
237 * as they are written and migrate onto the arc_mru list.
239 * The ARC_l2c_only state is for buffers that are in the second
240 * level ARC but no longer in any of the ARC_m* lists. The second
241 * level ARC itself may also contain buffers that are in any of
242 * the ARC_m* states - meaning that a buffer can exist in two
243 * places. The reason for the ARC_l2c_only state is to keep the
244 * buffer header in the hash table, so that reads that hit the
245 * second level ARC benefit from these fast lookups.
248 #define ARCS_LOCK_PAD CACHE_LINE_SIZE
252 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))];
257 * must be power of two for mask use to work
260 #define ARC_BUFC_NUMDATALISTS 16
261 #define ARC_BUFC_NUMMETADATALISTS 16
262 #define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS)
264 typedef struct arc_state {
265 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
266 uint64_t arcs_size; /* total amount of data in this state */
267 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */
268 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE);
271 #define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock))
274 static arc_state_t ARC_anon;
275 static arc_state_t ARC_mru;
276 static arc_state_t ARC_mru_ghost;
277 static arc_state_t ARC_mfu;
278 static arc_state_t ARC_mfu_ghost;
279 static arc_state_t ARC_l2c_only;
281 typedef struct arc_stats {
282 kstat_named_t arcstat_hits;
283 kstat_named_t arcstat_misses;
284 kstat_named_t arcstat_demand_data_hits;
285 kstat_named_t arcstat_demand_data_misses;
286 kstat_named_t arcstat_demand_metadata_hits;
287 kstat_named_t arcstat_demand_metadata_misses;
288 kstat_named_t arcstat_prefetch_data_hits;
289 kstat_named_t arcstat_prefetch_data_misses;
290 kstat_named_t arcstat_prefetch_metadata_hits;
291 kstat_named_t arcstat_prefetch_metadata_misses;
292 kstat_named_t arcstat_mru_hits;
293 kstat_named_t arcstat_mru_ghost_hits;
294 kstat_named_t arcstat_mfu_hits;
295 kstat_named_t arcstat_mfu_ghost_hits;
296 kstat_named_t arcstat_allocated;
297 kstat_named_t arcstat_deleted;
298 kstat_named_t arcstat_stolen;
299 kstat_named_t arcstat_recycle_miss;
301 * Number of buffers that could not be evicted because the hash lock
302 * was held by another thread. The lock may not necessarily be held
303 * by something using the same buffer, since hash locks are shared
304 * by multiple buffers.
306 kstat_named_t arcstat_mutex_miss;
308 * Number of buffers skipped because they have I/O in progress, are
309 * indrect prefetch buffers that have not lived long enough, or are
310 * not from the spa we're trying to evict from.
312 kstat_named_t arcstat_evict_skip;
313 kstat_named_t arcstat_evict_l2_cached;
314 kstat_named_t arcstat_evict_l2_eligible;
315 kstat_named_t arcstat_evict_l2_ineligible;
316 kstat_named_t arcstat_hash_elements;
317 kstat_named_t arcstat_hash_elements_max;
318 kstat_named_t arcstat_hash_collisions;
319 kstat_named_t arcstat_hash_chains;
320 kstat_named_t arcstat_hash_chain_max;
321 kstat_named_t arcstat_p;
322 kstat_named_t arcstat_c;
323 kstat_named_t arcstat_c_min;
324 kstat_named_t arcstat_c_max;
325 kstat_named_t arcstat_size;
326 kstat_named_t arcstat_hdr_size;
327 kstat_named_t arcstat_data_size;
328 kstat_named_t arcstat_other_size;
329 kstat_named_t arcstat_l2_hits;
330 kstat_named_t arcstat_l2_misses;
331 kstat_named_t arcstat_l2_feeds;
332 kstat_named_t arcstat_l2_rw_clash;
333 kstat_named_t arcstat_l2_read_bytes;
334 kstat_named_t arcstat_l2_write_bytes;
335 kstat_named_t arcstat_l2_writes_sent;
336 kstat_named_t arcstat_l2_writes_done;
337 kstat_named_t arcstat_l2_writes_error;
338 kstat_named_t arcstat_l2_writes_hdr_miss;
339 kstat_named_t arcstat_l2_evict_lock_retry;
340 kstat_named_t arcstat_l2_evict_reading;
341 kstat_named_t arcstat_l2_free_on_write;
342 kstat_named_t arcstat_l2_abort_lowmem;
343 kstat_named_t arcstat_l2_cksum_bad;
344 kstat_named_t arcstat_l2_io_error;
345 kstat_named_t arcstat_l2_size;
346 kstat_named_t arcstat_l2_asize;
347 kstat_named_t arcstat_l2_hdr_size;
348 kstat_named_t arcstat_l2_compress_successes;
349 kstat_named_t arcstat_l2_compress_zeros;
350 kstat_named_t arcstat_l2_compress_failures;
351 kstat_named_t arcstat_l2_write_trylock_fail;
352 kstat_named_t arcstat_l2_write_passed_headroom;
353 kstat_named_t arcstat_l2_write_spa_mismatch;
354 kstat_named_t arcstat_l2_write_in_l2;
355 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
356 kstat_named_t arcstat_l2_write_not_cacheable;
357 kstat_named_t arcstat_l2_write_full;
358 kstat_named_t arcstat_l2_write_buffer_iter;
359 kstat_named_t arcstat_l2_write_pios;
360 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
361 kstat_named_t arcstat_l2_write_buffer_list_iter;
362 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
363 kstat_named_t arcstat_memory_throttle_count;
364 kstat_named_t arcstat_duplicate_buffers;
365 kstat_named_t arcstat_duplicate_buffers_size;
366 kstat_named_t arcstat_duplicate_reads;
369 static arc_stats_t arc_stats = {
370 { "hits", KSTAT_DATA_UINT64 },
371 { "misses", KSTAT_DATA_UINT64 },
372 { "demand_data_hits", KSTAT_DATA_UINT64 },
373 { "demand_data_misses", KSTAT_DATA_UINT64 },
374 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
375 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
376 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
377 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
378 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
379 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
380 { "mru_hits", KSTAT_DATA_UINT64 },
381 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
382 { "mfu_hits", KSTAT_DATA_UINT64 },
383 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
384 { "allocated", KSTAT_DATA_UINT64 },
385 { "deleted", KSTAT_DATA_UINT64 },
386 { "stolen", KSTAT_DATA_UINT64 },
387 { "recycle_miss", KSTAT_DATA_UINT64 },
388 { "mutex_miss", KSTAT_DATA_UINT64 },
389 { "evict_skip", KSTAT_DATA_UINT64 },
390 { "evict_l2_cached", KSTAT_DATA_UINT64 },
391 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
392 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
393 { "hash_elements", KSTAT_DATA_UINT64 },
394 { "hash_elements_max", KSTAT_DATA_UINT64 },
395 { "hash_collisions", KSTAT_DATA_UINT64 },
396 { "hash_chains", KSTAT_DATA_UINT64 },
397 { "hash_chain_max", KSTAT_DATA_UINT64 },
398 { "p", KSTAT_DATA_UINT64 },
399 { "c", KSTAT_DATA_UINT64 },
400 { "c_min", KSTAT_DATA_UINT64 },
401 { "c_max", KSTAT_DATA_UINT64 },
402 { "size", KSTAT_DATA_UINT64 },
403 { "hdr_size", KSTAT_DATA_UINT64 },
404 { "data_size", KSTAT_DATA_UINT64 },
405 { "other_size", KSTAT_DATA_UINT64 },
406 { "l2_hits", KSTAT_DATA_UINT64 },
407 { "l2_misses", KSTAT_DATA_UINT64 },
408 { "l2_feeds", KSTAT_DATA_UINT64 },
409 { "l2_rw_clash", KSTAT_DATA_UINT64 },
410 { "l2_read_bytes", KSTAT_DATA_UINT64 },
411 { "l2_write_bytes", KSTAT_DATA_UINT64 },
412 { "l2_writes_sent", KSTAT_DATA_UINT64 },
413 { "l2_writes_done", KSTAT_DATA_UINT64 },
414 { "l2_writes_error", KSTAT_DATA_UINT64 },
415 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
416 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
417 { "l2_evict_reading", KSTAT_DATA_UINT64 },
418 { "l2_free_on_write", KSTAT_DATA_UINT64 },
419 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
420 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
421 { "l2_io_error", KSTAT_DATA_UINT64 },
422 { "l2_size", KSTAT_DATA_UINT64 },
423 { "l2_asize", KSTAT_DATA_UINT64 },
424 { "l2_hdr_size", KSTAT_DATA_UINT64 },
425 { "l2_compress_successes", KSTAT_DATA_UINT64 },
426 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
427 { "l2_compress_failures", KSTAT_DATA_UINT64 },
428 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
429 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
430 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
431 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
432 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
433 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
434 { "l2_write_full", KSTAT_DATA_UINT64 },
435 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
436 { "l2_write_pios", KSTAT_DATA_UINT64 },
437 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
438 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
439 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
440 { "memory_throttle_count", KSTAT_DATA_UINT64 },
441 { "duplicate_buffers", KSTAT_DATA_UINT64 },
442 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
443 { "duplicate_reads", KSTAT_DATA_UINT64 }
446 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
448 #define ARCSTAT_INCR(stat, val) \
449 atomic_add_64(&arc_stats.stat.value.ui64, (val))
451 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
452 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
454 #define ARCSTAT_MAX(stat, val) { \
456 while ((val) > (m = arc_stats.stat.value.ui64) && \
457 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
461 #define ARCSTAT_MAXSTAT(stat) \
462 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
465 * We define a macro to allow ARC hits/misses to be easily broken down by
466 * two separate conditions, giving a total of four different subtypes for
467 * each of hits and misses (so eight statistics total).
469 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
472 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
474 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
478 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
480 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
485 static arc_state_t *arc_anon;
486 static arc_state_t *arc_mru;
487 static arc_state_t *arc_mru_ghost;
488 static arc_state_t *arc_mfu;
489 static arc_state_t *arc_mfu_ghost;
490 static arc_state_t *arc_l2c_only;
493 * There are several ARC variables that are critical to export as kstats --
494 * but we don't want to have to grovel around in the kstat whenever we wish to
495 * manipulate them. For these variables, we therefore define them to be in
496 * terms of the statistic variable. This assures that we are not introducing
497 * the possibility of inconsistency by having shadow copies of the variables,
498 * while still allowing the code to be readable.
500 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
501 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
502 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
503 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
504 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
506 #define L2ARC_IS_VALID_COMPRESS(_c_) \
507 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
509 static int arc_no_grow; /* Don't try to grow cache size */
510 static uint64_t arc_tempreserve;
511 static uint64_t arc_loaned_bytes;
512 static uint64_t arc_meta_used;
513 static uint64_t arc_meta_limit;
514 static uint64_t arc_meta_max = 0;
515 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0,
516 "ARC metadata used");
517 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0,
518 "ARC metadata limit");
520 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
522 typedef struct arc_callback arc_callback_t;
524 struct arc_callback {
526 arc_done_func_t *acb_done;
528 zio_t *acb_zio_dummy;
529 arc_callback_t *acb_next;
532 typedef struct arc_write_callback arc_write_callback_t;
534 struct arc_write_callback {
536 arc_done_func_t *awcb_ready;
537 arc_done_func_t *awcb_physdone;
538 arc_done_func_t *awcb_done;
543 /* protected by hash lock */
548 kmutex_t b_freeze_lock;
549 zio_cksum_t *b_freeze_cksum;
552 arc_buf_hdr_t *b_hash_next;
557 arc_callback_t *b_acb;
561 arc_buf_contents_t b_type;
565 /* protected by arc state mutex */
566 arc_state_t *b_state;
567 list_node_t b_arc_node;
569 /* updated atomically */
570 clock_t b_arc_access;
572 /* self protecting */
575 l2arc_buf_hdr_t *b_l2hdr;
576 list_node_t b_l2node;
579 static arc_buf_t *arc_eviction_list;
580 static kmutex_t arc_eviction_mtx;
581 static arc_buf_hdr_t arc_eviction_hdr;
582 static void arc_get_data_buf(arc_buf_t *buf);
583 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
584 static int arc_evict_needed(arc_buf_contents_t type);
585 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
587 static void arc_buf_watch(arc_buf_t *buf);
590 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
592 #define GHOST_STATE(state) \
593 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
594 (state) == arc_l2c_only)
597 * Private ARC flags. These flags are private ARC only flags that will show up
598 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
599 * be passed in as arc_flags in things like arc_read. However, these flags
600 * should never be passed and should only be set by ARC code. When adding new
601 * public flags, make sure not to smash the private ones.
604 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
605 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
606 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
607 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
608 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
609 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
610 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
611 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
612 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
613 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
615 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
616 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
617 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
618 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
619 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
620 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
621 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
622 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
623 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
624 (hdr)->b_l2hdr != NULL)
625 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
626 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
627 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
633 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
634 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
637 * Hash table routines
640 #define HT_LOCK_PAD CACHE_LINE_SIZE
645 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
649 #define BUF_LOCKS 256
650 typedef struct buf_hash_table {
652 arc_buf_hdr_t **ht_table;
653 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
656 static buf_hash_table_t buf_hash_table;
658 #define BUF_HASH_INDEX(spa, dva, birth) \
659 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
660 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
661 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
662 #define HDR_LOCK(hdr) \
663 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
665 uint64_t zfs_crc64_table[256];
671 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
672 #define L2ARC_HEADROOM 2 /* num of writes */
674 * If we discover during ARC scan any buffers to be compressed, we boost
675 * our headroom for the next scanning cycle by this percentage multiple.
677 #define L2ARC_HEADROOM_BOOST 200
678 #define L2ARC_FEED_SECS 1 /* caching interval secs */
679 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
681 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
682 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
684 /* L2ARC Performance Tunables */
685 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
686 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
687 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
688 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
689 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
690 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
691 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
692 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
693 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
695 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
696 &l2arc_write_max, 0, "max write size");
697 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
698 &l2arc_write_boost, 0, "extra write during warmup");
699 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
700 &l2arc_headroom, 0, "number of dev writes");
701 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
702 &l2arc_feed_secs, 0, "interval seconds");
703 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
704 &l2arc_feed_min_ms, 0, "min interval milliseconds");
706 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
707 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
708 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
709 &l2arc_feed_again, 0, "turbo warmup");
710 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
711 &l2arc_norw, 0, "no reads during writes");
713 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
714 &ARC_anon.arcs_size, 0, "size of anonymous state");
715 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
716 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
717 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
718 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
720 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
721 &ARC_mru.arcs_size, 0, "size of mru state");
722 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
723 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
724 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
725 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
727 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
728 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state");
729 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
730 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
731 "size of metadata in mru ghost state");
732 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
733 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
734 "size of data in mru ghost state");
736 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
737 &ARC_mfu.arcs_size, 0, "size of mfu state");
738 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
739 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
740 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
741 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
743 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
744 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state");
745 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
746 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
747 "size of metadata in mfu ghost state");
748 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
749 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
750 "size of data in mfu ghost state");
752 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
753 &ARC_l2c_only.arcs_size, 0, "size of mru state");
758 typedef struct l2arc_dev {
759 vdev_t *l2ad_vdev; /* vdev */
760 spa_t *l2ad_spa; /* spa */
761 uint64_t l2ad_hand; /* next write location */
762 uint64_t l2ad_start; /* first addr on device */
763 uint64_t l2ad_end; /* last addr on device */
764 uint64_t l2ad_evict; /* last addr eviction reached */
765 boolean_t l2ad_first; /* first sweep through */
766 boolean_t l2ad_writing; /* currently writing */
767 list_t *l2ad_buflist; /* buffer list */
768 list_node_t l2ad_node; /* device list node */
771 static list_t L2ARC_dev_list; /* device list */
772 static list_t *l2arc_dev_list; /* device list pointer */
773 static kmutex_t l2arc_dev_mtx; /* device list mutex */
774 static l2arc_dev_t *l2arc_dev_last; /* last device used */
775 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
776 static list_t L2ARC_free_on_write; /* free after write buf list */
777 static list_t *l2arc_free_on_write; /* free after write list ptr */
778 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
779 static uint64_t l2arc_ndev; /* number of devices */
781 typedef struct l2arc_read_callback {
782 arc_buf_t *l2rcb_buf; /* read buffer */
783 spa_t *l2rcb_spa; /* spa */
784 blkptr_t l2rcb_bp; /* original blkptr */
785 zbookmark_phys_t l2rcb_zb; /* original bookmark */
786 int l2rcb_flags; /* original flags */
787 enum zio_compress l2rcb_compress; /* applied compress */
788 } l2arc_read_callback_t;
790 typedef struct l2arc_write_callback {
791 l2arc_dev_t *l2wcb_dev; /* device info */
792 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
793 } l2arc_write_callback_t;
795 struct l2arc_buf_hdr {
796 /* protected by arc_buf_hdr mutex */
797 l2arc_dev_t *b_dev; /* L2ARC device */
798 uint64_t b_daddr; /* disk address, offset byte */
799 /* compression applied to buffer data */
800 enum zio_compress b_compress;
801 /* real alloc'd buffer size depending on b_compress applied */
803 /* temporary buffer holder for in-flight compressed data */
807 typedef struct l2arc_data_free {
808 /* protected by l2arc_free_on_write_mtx */
811 void (*l2df_func)(void *, size_t);
812 list_node_t l2df_list_node;
815 static kmutex_t l2arc_feed_thr_lock;
816 static kcondvar_t l2arc_feed_thr_cv;
817 static uint8_t l2arc_thread_exit;
819 static void l2arc_read_done(zio_t *zio);
820 static void l2arc_hdr_stat_add(void);
821 static void l2arc_hdr_stat_remove(void);
823 static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr);
824 static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr,
825 enum zio_compress c);
826 static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab);
829 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
831 uint8_t *vdva = (uint8_t *)dva;
832 uint64_t crc = -1ULL;
835 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
837 for (i = 0; i < sizeof (dva_t); i++)
838 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
840 crc ^= (spa>>8) ^ birth;
845 #define BUF_EMPTY(buf) \
846 ((buf)->b_dva.dva_word[0] == 0 && \
847 (buf)->b_dva.dva_word[1] == 0 && \
848 (buf)->b_cksum0 == 0)
850 #define BUF_EQUAL(spa, dva, birth, buf) \
851 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
852 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
853 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
856 buf_discard_identity(arc_buf_hdr_t *hdr)
858 hdr->b_dva.dva_word[0] = 0;
859 hdr->b_dva.dva_word[1] = 0;
864 static arc_buf_hdr_t *
865 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
867 const dva_t *dva = BP_IDENTITY(bp);
868 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
869 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
870 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
873 mutex_enter(hash_lock);
874 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
875 buf = buf->b_hash_next) {
876 if (BUF_EQUAL(spa, dva, birth, buf)) {
881 mutex_exit(hash_lock);
887 * Insert an entry into the hash table. If there is already an element
888 * equal to elem in the hash table, then the already existing element
889 * will be returned and the new element will not be inserted.
890 * Otherwise returns NULL.
892 static arc_buf_hdr_t *
893 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
895 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
896 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
900 ASSERT(!DVA_IS_EMPTY(&buf->b_dva));
901 ASSERT(buf->b_birth != 0);
902 ASSERT(!HDR_IN_HASH_TABLE(buf));
904 mutex_enter(hash_lock);
905 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
906 fbuf = fbuf->b_hash_next, i++) {
907 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
911 buf->b_hash_next = buf_hash_table.ht_table[idx];
912 buf_hash_table.ht_table[idx] = buf;
913 buf->b_flags |= ARC_IN_HASH_TABLE;
915 /* collect some hash table performance data */
917 ARCSTAT_BUMP(arcstat_hash_collisions);
919 ARCSTAT_BUMP(arcstat_hash_chains);
921 ARCSTAT_MAX(arcstat_hash_chain_max, i);
924 ARCSTAT_BUMP(arcstat_hash_elements);
925 ARCSTAT_MAXSTAT(arcstat_hash_elements);
931 buf_hash_remove(arc_buf_hdr_t *buf)
933 arc_buf_hdr_t *fbuf, **bufp;
934 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
936 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
937 ASSERT(HDR_IN_HASH_TABLE(buf));
939 bufp = &buf_hash_table.ht_table[idx];
940 while ((fbuf = *bufp) != buf) {
941 ASSERT(fbuf != NULL);
942 bufp = &fbuf->b_hash_next;
944 *bufp = buf->b_hash_next;
945 buf->b_hash_next = NULL;
946 buf->b_flags &= ~ARC_IN_HASH_TABLE;
948 /* collect some hash table performance data */
949 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
951 if (buf_hash_table.ht_table[idx] &&
952 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
953 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
957 * Global data structures and functions for the buf kmem cache.
959 static kmem_cache_t *hdr_cache;
960 static kmem_cache_t *buf_cache;
967 kmem_free(buf_hash_table.ht_table,
968 (buf_hash_table.ht_mask + 1) * sizeof (void *));
969 for (i = 0; i < BUF_LOCKS; i++)
970 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
971 kmem_cache_destroy(hdr_cache);
972 kmem_cache_destroy(buf_cache);
976 * Constructor callback - called when the cache is empty
977 * and a new buf is requested.
981 hdr_cons(void *vbuf, void *unused, int kmflag)
983 arc_buf_hdr_t *buf = vbuf;
985 bzero(buf, sizeof (arc_buf_hdr_t));
986 refcount_create(&buf->b_refcnt);
987 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
988 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
989 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
996 buf_cons(void *vbuf, void *unused, int kmflag)
998 arc_buf_t *buf = vbuf;
1000 bzero(buf, sizeof (arc_buf_t));
1001 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1002 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1008 * Destructor callback - called when a cached buf is
1009 * no longer required.
1013 hdr_dest(void *vbuf, void *unused)
1015 arc_buf_hdr_t *buf = vbuf;
1017 ASSERT(BUF_EMPTY(buf));
1018 refcount_destroy(&buf->b_refcnt);
1019 cv_destroy(&buf->b_cv);
1020 mutex_destroy(&buf->b_freeze_lock);
1021 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
1026 buf_dest(void *vbuf, void *unused)
1028 arc_buf_t *buf = vbuf;
1030 mutex_destroy(&buf->b_evict_lock);
1031 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1035 * Reclaim callback -- invoked when memory is low.
1039 hdr_recl(void *unused)
1041 dprintf("hdr_recl called\n");
1043 * umem calls the reclaim func when we destroy the buf cache,
1044 * which is after we do arc_fini().
1047 cv_signal(&arc_reclaim_thr_cv);
1054 uint64_t hsize = 1ULL << 12;
1058 * The hash table is big enough to fill all of physical memory
1059 * with an average 64K block size. The table will take up
1060 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
1062 while (hsize * 65536 < (uint64_t)physmem * PAGESIZE)
1065 buf_hash_table.ht_mask = hsize - 1;
1066 buf_hash_table.ht_table =
1067 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1068 if (buf_hash_table.ht_table == NULL) {
1069 ASSERT(hsize > (1ULL << 8));
1074 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
1075 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
1076 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1077 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1079 for (i = 0; i < 256; i++)
1080 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1081 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1083 for (i = 0; i < BUF_LOCKS; i++) {
1084 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1085 NULL, MUTEX_DEFAULT, NULL);
1089 #define ARC_MINTIME (hz>>4) /* 62 ms */
1092 arc_cksum_verify(arc_buf_t *buf)
1096 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1099 mutex_enter(&buf->b_hdr->b_freeze_lock);
1100 if (buf->b_hdr->b_freeze_cksum == NULL ||
1101 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
1102 mutex_exit(&buf->b_hdr->b_freeze_lock);
1105 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1106 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1107 panic("buffer modified while frozen!");
1108 mutex_exit(&buf->b_hdr->b_freeze_lock);
1112 arc_cksum_equal(arc_buf_t *buf)
1117 mutex_enter(&buf->b_hdr->b_freeze_lock);
1118 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1119 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1120 mutex_exit(&buf->b_hdr->b_freeze_lock);
1126 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1128 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1131 mutex_enter(&buf->b_hdr->b_freeze_lock);
1132 if (buf->b_hdr->b_freeze_cksum != NULL) {
1133 mutex_exit(&buf->b_hdr->b_freeze_lock);
1136 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1137 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1138 buf->b_hdr->b_freeze_cksum);
1139 mutex_exit(&buf->b_hdr->b_freeze_lock);
1142 #endif /* illumos */
1147 typedef struct procctl {
1155 arc_buf_unwatch(arc_buf_t *buf)
1162 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1163 ctl.prwatch.pr_size = 0;
1164 ctl.prwatch.pr_wflags = 0;
1165 result = write(arc_procfd, &ctl, sizeof (ctl));
1166 ASSERT3U(result, ==, sizeof (ctl));
1173 arc_buf_watch(arc_buf_t *buf)
1180 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1181 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1182 ctl.prwatch.pr_wflags = WA_WRITE;
1183 result = write(arc_procfd, &ctl, sizeof (ctl));
1184 ASSERT3U(result, ==, sizeof (ctl));
1188 #endif /* illumos */
1191 arc_buf_thaw(arc_buf_t *buf)
1193 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1194 if (buf->b_hdr->b_state != arc_anon)
1195 panic("modifying non-anon buffer!");
1196 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1197 panic("modifying buffer while i/o in progress!");
1198 arc_cksum_verify(buf);
1201 mutex_enter(&buf->b_hdr->b_freeze_lock);
1202 if (buf->b_hdr->b_freeze_cksum != NULL) {
1203 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1204 buf->b_hdr->b_freeze_cksum = NULL;
1207 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1208 if (buf->b_hdr->b_thawed)
1209 kmem_free(buf->b_hdr->b_thawed, 1);
1210 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1213 mutex_exit(&buf->b_hdr->b_freeze_lock);
1216 arc_buf_unwatch(buf);
1217 #endif /* illumos */
1221 arc_buf_freeze(arc_buf_t *buf)
1223 kmutex_t *hash_lock;
1225 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1228 hash_lock = HDR_LOCK(buf->b_hdr);
1229 mutex_enter(hash_lock);
1231 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1232 buf->b_hdr->b_state == arc_anon);
1233 arc_cksum_compute(buf, B_FALSE);
1234 mutex_exit(hash_lock);
1239 get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock)
1241 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth);
1243 if (ab->b_type == ARC_BUFC_METADATA)
1244 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1);
1246 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1);
1247 buf_hashid += ARC_BUFC_NUMMETADATALISTS;
1250 *list = &state->arcs_lists[buf_hashid];
1251 *lock = ARCS_LOCK(state, buf_hashid);
1256 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1258 ASSERT(MUTEX_HELD(hash_lock));
1260 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1261 (ab->b_state != arc_anon)) {
1262 uint64_t delta = ab->b_size * ab->b_datacnt;
1263 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1267 get_buf_info(ab, ab->b_state, &list, &lock);
1268 ASSERT(!MUTEX_HELD(lock));
1270 ASSERT(list_link_active(&ab->b_arc_node));
1271 list_remove(list, ab);
1272 if (GHOST_STATE(ab->b_state)) {
1273 ASSERT0(ab->b_datacnt);
1274 ASSERT3P(ab->b_buf, ==, NULL);
1278 ASSERT3U(*size, >=, delta);
1279 atomic_add_64(size, -delta);
1281 /* remove the prefetch flag if we get a reference */
1282 if (ab->b_flags & ARC_PREFETCH)
1283 ab->b_flags &= ~ARC_PREFETCH;
1288 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1291 arc_state_t *state = ab->b_state;
1293 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1294 ASSERT(!GHOST_STATE(state));
1296 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1297 (state != arc_anon)) {
1298 uint64_t *size = &state->arcs_lsize[ab->b_type];
1302 get_buf_info(ab, state, &list, &lock);
1303 ASSERT(!MUTEX_HELD(lock));
1305 ASSERT(!list_link_active(&ab->b_arc_node));
1306 list_insert_head(list, ab);
1307 ASSERT(ab->b_datacnt > 0);
1308 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1315 * Move the supplied buffer to the indicated state. The mutex
1316 * for the buffer must be held by the caller.
1319 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1321 arc_state_t *old_state = ab->b_state;
1322 int64_t refcnt = refcount_count(&ab->b_refcnt);
1323 uint64_t from_delta, to_delta;
1327 ASSERT(MUTEX_HELD(hash_lock));
1328 ASSERT3P(new_state, !=, old_state);
1329 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1330 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1331 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1333 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1336 * If this buffer is evictable, transfer it from the
1337 * old state list to the new state list.
1340 if (old_state != arc_anon) {
1342 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1344 get_buf_info(ab, old_state, &list, &lock);
1345 use_mutex = !MUTEX_HELD(lock);
1349 ASSERT(list_link_active(&ab->b_arc_node));
1350 list_remove(list, ab);
1353 * If prefetching out of the ghost cache,
1354 * we will have a non-zero datacnt.
1356 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1357 /* ghost elements have a ghost size */
1358 ASSERT(ab->b_buf == NULL);
1359 from_delta = ab->b_size;
1361 ASSERT3U(*size, >=, from_delta);
1362 atomic_add_64(size, -from_delta);
1367 if (new_state != arc_anon) {
1369 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1371 get_buf_info(ab, new_state, &list, &lock);
1372 use_mutex = !MUTEX_HELD(lock);
1376 list_insert_head(list, ab);
1378 /* ghost elements have a ghost size */
1379 if (GHOST_STATE(new_state)) {
1380 ASSERT(ab->b_datacnt == 0);
1381 ASSERT(ab->b_buf == NULL);
1382 to_delta = ab->b_size;
1384 atomic_add_64(size, to_delta);
1391 ASSERT(!BUF_EMPTY(ab));
1392 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1393 buf_hash_remove(ab);
1395 /* adjust state sizes */
1397 atomic_add_64(&new_state->arcs_size, to_delta);
1399 ASSERT3U(old_state->arcs_size, >=, from_delta);
1400 atomic_add_64(&old_state->arcs_size, -from_delta);
1402 ab->b_state = new_state;
1404 /* adjust l2arc hdr stats */
1405 if (new_state == arc_l2c_only)
1406 l2arc_hdr_stat_add();
1407 else if (old_state == arc_l2c_only)
1408 l2arc_hdr_stat_remove();
1412 arc_space_consume(uint64_t space, arc_space_type_t type)
1414 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1417 case ARC_SPACE_DATA:
1418 ARCSTAT_INCR(arcstat_data_size, space);
1420 case ARC_SPACE_OTHER:
1421 ARCSTAT_INCR(arcstat_other_size, space);
1423 case ARC_SPACE_HDRS:
1424 ARCSTAT_INCR(arcstat_hdr_size, space);
1426 case ARC_SPACE_L2HDRS:
1427 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1431 atomic_add_64(&arc_meta_used, space);
1432 atomic_add_64(&arc_size, space);
1436 arc_space_return(uint64_t space, arc_space_type_t type)
1438 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1441 case ARC_SPACE_DATA:
1442 ARCSTAT_INCR(arcstat_data_size, -space);
1444 case ARC_SPACE_OTHER:
1445 ARCSTAT_INCR(arcstat_other_size, -space);
1447 case ARC_SPACE_HDRS:
1448 ARCSTAT_INCR(arcstat_hdr_size, -space);
1450 case ARC_SPACE_L2HDRS:
1451 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1455 ASSERT(arc_meta_used >= space);
1456 if (arc_meta_max < arc_meta_used)
1457 arc_meta_max = arc_meta_used;
1458 atomic_add_64(&arc_meta_used, -space);
1459 ASSERT(arc_size >= space);
1460 atomic_add_64(&arc_size, -space);
1464 arc_data_buf_alloc(uint64_t size)
1466 if (arc_evict_needed(ARC_BUFC_DATA))
1467 cv_signal(&arc_reclaim_thr_cv);
1468 atomic_add_64(&arc_size, size);
1469 return (zio_data_buf_alloc(size));
1473 arc_data_buf_free(void *buf, uint64_t size)
1475 zio_data_buf_free(buf, size);
1476 ASSERT(arc_size >= size);
1477 atomic_add_64(&arc_size, -size);
1481 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1486 ASSERT3U(size, >, 0);
1487 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1488 ASSERT(BUF_EMPTY(hdr));
1491 hdr->b_spa = spa_load_guid(spa);
1492 hdr->b_state = arc_anon;
1493 hdr->b_arc_access = 0;
1494 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1497 buf->b_efunc = NULL;
1498 buf->b_private = NULL;
1501 arc_get_data_buf(buf);
1504 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1505 (void) refcount_add(&hdr->b_refcnt, tag);
1510 static char *arc_onloan_tag = "onloan";
1513 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1514 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1515 * buffers must be returned to the arc before they can be used by the DMU or
1519 arc_loan_buf(spa_t *spa, int size)
1523 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1525 atomic_add_64(&arc_loaned_bytes, size);
1530 * Return a loaned arc buffer to the arc.
1533 arc_return_buf(arc_buf_t *buf, void *tag)
1535 arc_buf_hdr_t *hdr = buf->b_hdr;
1537 ASSERT(buf->b_data != NULL);
1538 (void) refcount_add(&hdr->b_refcnt, tag);
1539 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1541 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1544 /* Detach an arc_buf from a dbuf (tag) */
1546 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1550 ASSERT(buf->b_data != NULL);
1552 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1553 (void) refcount_remove(&hdr->b_refcnt, tag);
1554 buf->b_efunc = NULL;
1555 buf->b_private = NULL;
1557 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1561 arc_buf_clone(arc_buf_t *from)
1564 arc_buf_hdr_t *hdr = from->b_hdr;
1565 uint64_t size = hdr->b_size;
1567 ASSERT(hdr->b_state != arc_anon);
1569 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1572 buf->b_efunc = NULL;
1573 buf->b_private = NULL;
1574 buf->b_next = hdr->b_buf;
1576 arc_get_data_buf(buf);
1577 bcopy(from->b_data, buf->b_data, size);
1580 * This buffer already exists in the arc so create a duplicate
1581 * copy for the caller. If the buffer is associated with user data
1582 * then track the size and number of duplicates. These stats will be
1583 * updated as duplicate buffers are created and destroyed.
1585 if (hdr->b_type == ARC_BUFC_DATA) {
1586 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1587 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1589 hdr->b_datacnt += 1;
1594 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1597 kmutex_t *hash_lock;
1600 * Check to see if this buffer is evicted. Callers
1601 * must verify b_data != NULL to know if the add_ref
1604 mutex_enter(&buf->b_evict_lock);
1605 if (buf->b_data == NULL) {
1606 mutex_exit(&buf->b_evict_lock);
1609 hash_lock = HDR_LOCK(buf->b_hdr);
1610 mutex_enter(hash_lock);
1612 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1613 mutex_exit(&buf->b_evict_lock);
1615 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1616 add_reference(hdr, hash_lock, tag);
1617 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1618 arc_access(hdr, hash_lock);
1619 mutex_exit(hash_lock);
1620 ARCSTAT_BUMP(arcstat_hits);
1621 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1622 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1623 data, metadata, hits);
1627 * Free the arc data buffer. If it is an l2arc write in progress,
1628 * the buffer is placed on l2arc_free_on_write to be freed later.
1631 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1633 arc_buf_hdr_t *hdr = buf->b_hdr;
1635 if (HDR_L2_WRITING(hdr)) {
1636 l2arc_data_free_t *df;
1637 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1638 df->l2df_data = buf->b_data;
1639 df->l2df_size = hdr->b_size;
1640 df->l2df_func = free_func;
1641 mutex_enter(&l2arc_free_on_write_mtx);
1642 list_insert_head(l2arc_free_on_write, df);
1643 mutex_exit(&l2arc_free_on_write_mtx);
1644 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1646 free_func(buf->b_data, hdr->b_size);
1651 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1655 /* free up data associated with the buf */
1657 arc_state_t *state = buf->b_hdr->b_state;
1658 uint64_t size = buf->b_hdr->b_size;
1659 arc_buf_contents_t type = buf->b_hdr->b_type;
1661 arc_cksum_verify(buf);
1663 arc_buf_unwatch(buf);
1664 #endif /* illumos */
1667 if (type == ARC_BUFC_METADATA) {
1668 arc_buf_data_free(buf, zio_buf_free);
1669 arc_space_return(size, ARC_SPACE_DATA);
1671 ASSERT(type == ARC_BUFC_DATA);
1672 arc_buf_data_free(buf, zio_data_buf_free);
1673 ARCSTAT_INCR(arcstat_data_size, -size);
1674 atomic_add_64(&arc_size, -size);
1677 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1678 uint64_t *cnt = &state->arcs_lsize[type];
1680 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1681 ASSERT(state != arc_anon);
1683 ASSERT3U(*cnt, >=, size);
1684 atomic_add_64(cnt, -size);
1686 ASSERT3U(state->arcs_size, >=, size);
1687 atomic_add_64(&state->arcs_size, -size);
1691 * If we're destroying a duplicate buffer make sure
1692 * that the appropriate statistics are updated.
1694 if (buf->b_hdr->b_datacnt > 1 &&
1695 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1696 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1697 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1699 ASSERT(buf->b_hdr->b_datacnt > 0);
1700 buf->b_hdr->b_datacnt -= 1;
1703 /* only remove the buf if requested */
1707 /* remove the buf from the hdr list */
1708 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1710 *bufp = buf->b_next;
1713 ASSERT(buf->b_efunc == NULL);
1715 /* clean up the buf */
1717 kmem_cache_free(buf_cache, buf);
1721 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1723 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1724 ASSERT3P(hdr->b_state, ==, arc_anon);
1725 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1726 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1728 if (l2hdr != NULL) {
1729 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1731 * To prevent arc_free() and l2arc_evict() from
1732 * attempting to free the same buffer at the same time,
1733 * a FREE_IN_PROGRESS flag is given to arc_free() to
1734 * give it priority. l2arc_evict() can't destroy this
1735 * header while we are waiting on l2arc_buflist_mtx.
1737 * The hdr may be removed from l2ad_buflist before we
1738 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1740 if (!buflist_held) {
1741 mutex_enter(&l2arc_buflist_mtx);
1742 l2hdr = hdr->b_l2hdr;
1745 if (l2hdr != NULL) {
1746 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr,
1748 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1749 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1750 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
1751 vdev_space_update(l2hdr->b_dev->l2ad_vdev,
1752 -l2hdr->b_asize, 0, 0);
1753 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1754 if (hdr->b_state == arc_l2c_only)
1755 l2arc_hdr_stat_remove();
1756 hdr->b_l2hdr = NULL;
1760 mutex_exit(&l2arc_buflist_mtx);
1763 if (!BUF_EMPTY(hdr)) {
1764 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1765 buf_discard_identity(hdr);
1767 while (hdr->b_buf) {
1768 arc_buf_t *buf = hdr->b_buf;
1771 mutex_enter(&arc_eviction_mtx);
1772 mutex_enter(&buf->b_evict_lock);
1773 ASSERT(buf->b_hdr != NULL);
1774 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1775 hdr->b_buf = buf->b_next;
1776 buf->b_hdr = &arc_eviction_hdr;
1777 buf->b_next = arc_eviction_list;
1778 arc_eviction_list = buf;
1779 mutex_exit(&buf->b_evict_lock);
1780 mutex_exit(&arc_eviction_mtx);
1782 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1785 if (hdr->b_freeze_cksum != NULL) {
1786 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1787 hdr->b_freeze_cksum = NULL;
1789 if (hdr->b_thawed) {
1790 kmem_free(hdr->b_thawed, 1);
1791 hdr->b_thawed = NULL;
1794 ASSERT(!list_link_active(&hdr->b_arc_node));
1795 ASSERT3P(hdr->b_hash_next, ==, NULL);
1796 ASSERT3P(hdr->b_acb, ==, NULL);
1797 kmem_cache_free(hdr_cache, hdr);
1801 arc_buf_free(arc_buf_t *buf, void *tag)
1803 arc_buf_hdr_t *hdr = buf->b_hdr;
1804 int hashed = hdr->b_state != arc_anon;
1806 ASSERT(buf->b_efunc == NULL);
1807 ASSERT(buf->b_data != NULL);
1810 kmutex_t *hash_lock = HDR_LOCK(hdr);
1812 mutex_enter(hash_lock);
1814 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1816 (void) remove_reference(hdr, hash_lock, tag);
1817 if (hdr->b_datacnt > 1) {
1818 arc_buf_destroy(buf, FALSE, TRUE);
1820 ASSERT(buf == hdr->b_buf);
1821 ASSERT(buf->b_efunc == NULL);
1822 hdr->b_flags |= ARC_BUF_AVAILABLE;
1824 mutex_exit(hash_lock);
1825 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1828 * We are in the middle of an async write. Don't destroy
1829 * this buffer unless the write completes before we finish
1830 * decrementing the reference count.
1832 mutex_enter(&arc_eviction_mtx);
1833 (void) remove_reference(hdr, NULL, tag);
1834 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1835 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1836 mutex_exit(&arc_eviction_mtx);
1838 arc_hdr_destroy(hdr);
1840 if (remove_reference(hdr, NULL, tag) > 0)
1841 arc_buf_destroy(buf, FALSE, TRUE);
1843 arc_hdr_destroy(hdr);
1848 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1850 arc_buf_hdr_t *hdr = buf->b_hdr;
1851 kmutex_t *hash_lock = HDR_LOCK(hdr);
1852 boolean_t no_callback = (buf->b_efunc == NULL);
1854 if (hdr->b_state == arc_anon) {
1855 ASSERT(hdr->b_datacnt == 1);
1856 arc_buf_free(buf, tag);
1857 return (no_callback);
1860 mutex_enter(hash_lock);
1862 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1863 ASSERT(hdr->b_state != arc_anon);
1864 ASSERT(buf->b_data != NULL);
1866 (void) remove_reference(hdr, hash_lock, tag);
1867 if (hdr->b_datacnt > 1) {
1869 arc_buf_destroy(buf, FALSE, TRUE);
1870 } else if (no_callback) {
1871 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1872 ASSERT(buf->b_efunc == NULL);
1873 hdr->b_flags |= ARC_BUF_AVAILABLE;
1875 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1876 refcount_is_zero(&hdr->b_refcnt));
1877 mutex_exit(hash_lock);
1878 return (no_callback);
1882 arc_buf_size(arc_buf_t *buf)
1884 return (buf->b_hdr->b_size);
1888 * Called from the DMU to determine if the current buffer should be
1889 * evicted. In order to ensure proper locking, the eviction must be initiated
1890 * from the DMU. Return true if the buffer is associated with user data and
1891 * duplicate buffers still exist.
1894 arc_buf_eviction_needed(arc_buf_t *buf)
1897 boolean_t evict_needed = B_FALSE;
1899 if (zfs_disable_dup_eviction)
1902 mutex_enter(&buf->b_evict_lock);
1906 * We are in arc_do_user_evicts(); let that function
1907 * perform the eviction.
1909 ASSERT(buf->b_data == NULL);
1910 mutex_exit(&buf->b_evict_lock);
1912 } else if (buf->b_data == NULL) {
1914 * We have already been added to the arc eviction list;
1915 * recommend eviction.
1917 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1918 mutex_exit(&buf->b_evict_lock);
1922 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1923 evict_needed = B_TRUE;
1925 mutex_exit(&buf->b_evict_lock);
1926 return (evict_needed);
1930 * Evict buffers from list until we've removed the specified number of
1931 * bytes. Move the removed buffers to the appropriate evict state.
1932 * If the recycle flag is set, then attempt to "recycle" a buffer:
1933 * - look for a buffer to evict that is `bytes' long.
1934 * - return the data block from this buffer rather than freeing it.
1935 * This flag is used by callers that are trying to make space for a
1936 * new buffer in a full arc cache.
1938 * This function makes a "best effort". It skips over any buffers
1939 * it can't get a hash_lock on, and so may not catch all candidates.
1940 * It may also return without evicting as much space as requested.
1943 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1944 arc_buf_contents_t type)
1946 arc_state_t *evicted_state;
1947 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1948 int64_t bytes_remaining;
1949 arc_buf_hdr_t *ab, *ab_prev = NULL;
1950 list_t *evicted_list, *list, *evicted_list_start, *list_start;
1951 kmutex_t *lock, *evicted_lock;
1952 kmutex_t *hash_lock;
1953 boolean_t have_lock;
1954 void *stolen = NULL;
1955 arc_buf_hdr_t marker = { 0 };
1957 static int evict_metadata_offset, evict_data_offset;
1958 int i, idx, offset, list_count, lists;
1960 ASSERT(state == arc_mru || state == arc_mfu);
1962 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1964 if (type == ARC_BUFC_METADATA) {
1966 list_count = ARC_BUFC_NUMMETADATALISTS;
1967 list_start = &state->arcs_lists[0];
1968 evicted_list_start = &evicted_state->arcs_lists[0];
1969 idx = evict_metadata_offset;
1971 offset = ARC_BUFC_NUMMETADATALISTS;
1972 list_start = &state->arcs_lists[offset];
1973 evicted_list_start = &evicted_state->arcs_lists[offset];
1974 list_count = ARC_BUFC_NUMDATALISTS;
1975 idx = evict_data_offset;
1977 bytes_remaining = evicted_state->arcs_lsize[type];
1981 list = &list_start[idx];
1982 evicted_list = &evicted_list_start[idx];
1983 lock = ARCS_LOCK(state, (offset + idx));
1984 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx));
1987 mutex_enter(evicted_lock);
1989 for (ab = list_tail(list); ab; ab = ab_prev) {
1990 ab_prev = list_prev(list, ab);
1991 bytes_remaining -= (ab->b_size * ab->b_datacnt);
1992 /* prefetch buffers have a minimum lifespan */
1993 if (HDR_IO_IN_PROGRESS(ab) ||
1994 (spa && ab->b_spa != spa) ||
1995 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1996 ddi_get_lbolt() - ab->b_arc_access <
1997 arc_min_prefetch_lifespan)) {
2001 /* "lookahead" for better eviction candidate */
2002 if (recycle && ab->b_size != bytes &&
2003 ab_prev && ab_prev->b_size == bytes)
2006 /* ignore markers */
2011 * It may take a long time to evict all the bufs requested.
2012 * To avoid blocking all arc activity, periodically drop
2013 * the arcs_mtx and give other threads a chance to run
2014 * before reacquiring the lock.
2016 * If we are looking for a buffer to recycle, we are in
2017 * the hot code path, so don't sleep.
2019 if (!recycle && count++ > arc_evict_iterations) {
2020 list_insert_after(list, ab, &marker);
2021 mutex_exit(evicted_lock);
2023 kpreempt(KPREEMPT_SYNC);
2025 mutex_enter(evicted_lock);
2026 ab_prev = list_prev(list, &marker);
2027 list_remove(list, &marker);
2032 hash_lock = HDR_LOCK(ab);
2033 have_lock = MUTEX_HELD(hash_lock);
2034 if (have_lock || mutex_tryenter(hash_lock)) {
2035 ASSERT0(refcount_count(&ab->b_refcnt));
2036 ASSERT(ab->b_datacnt > 0);
2038 arc_buf_t *buf = ab->b_buf;
2039 if (!mutex_tryenter(&buf->b_evict_lock)) {
2044 bytes_evicted += ab->b_size;
2045 if (recycle && ab->b_type == type &&
2046 ab->b_size == bytes &&
2047 !HDR_L2_WRITING(ab)) {
2048 stolen = buf->b_data;
2053 mutex_enter(&arc_eviction_mtx);
2054 arc_buf_destroy(buf,
2055 buf->b_data == stolen, FALSE);
2056 ab->b_buf = buf->b_next;
2057 buf->b_hdr = &arc_eviction_hdr;
2058 buf->b_next = arc_eviction_list;
2059 arc_eviction_list = buf;
2060 mutex_exit(&arc_eviction_mtx);
2061 mutex_exit(&buf->b_evict_lock);
2063 mutex_exit(&buf->b_evict_lock);
2064 arc_buf_destroy(buf,
2065 buf->b_data == stolen, TRUE);
2070 ARCSTAT_INCR(arcstat_evict_l2_cached,
2073 if (l2arc_write_eligible(ab->b_spa, ab)) {
2074 ARCSTAT_INCR(arcstat_evict_l2_eligible,
2078 arcstat_evict_l2_ineligible,
2083 if (ab->b_datacnt == 0) {
2084 arc_change_state(evicted_state, ab, hash_lock);
2085 ASSERT(HDR_IN_HASH_TABLE(ab));
2086 ab->b_flags |= ARC_IN_HASH_TABLE;
2087 ab->b_flags &= ~ARC_BUF_AVAILABLE;
2088 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
2091 mutex_exit(hash_lock);
2092 if (bytes >= 0 && bytes_evicted >= bytes)
2094 if (bytes_remaining > 0) {
2095 mutex_exit(evicted_lock);
2097 idx = ((idx + 1) & (list_count - 1));
2106 mutex_exit(evicted_lock);
2109 idx = ((idx + 1) & (list_count - 1));
2112 if (bytes_evicted < bytes) {
2113 if (lists < list_count)
2116 dprintf("only evicted %lld bytes from %x",
2117 (longlong_t)bytes_evicted, state);
2119 if (type == ARC_BUFC_METADATA)
2120 evict_metadata_offset = idx;
2122 evict_data_offset = idx;
2125 ARCSTAT_INCR(arcstat_evict_skip, skipped);
2128 ARCSTAT_INCR(arcstat_mutex_miss, missed);
2131 * Note: we have just evicted some data into the ghost state,
2132 * potentially putting the ghost size over the desired size. Rather
2133 * that evicting from the ghost list in this hot code path, leave
2134 * this chore to the arc_reclaim_thread().
2138 ARCSTAT_BUMP(arcstat_stolen);
2143 * Remove buffers from list until we've removed the specified number of
2144 * bytes. Destroy the buffers that are removed.
2147 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
2149 arc_buf_hdr_t *ab, *ab_prev;
2150 arc_buf_hdr_t marker = { 0 };
2151 list_t *list, *list_start;
2152 kmutex_t *hash_lock, *lock;
2153 uint64_t bytes_deleted = 0;
2154 uint64_t bufs_skipped = 0;
2156 static int evict_offset;
2157 int list_count, idx = evict_offset;
2158 int offset, lists = 0;
2160 ASSERT(GHOST_STATE(state));
2163 * data lists come after metadata lists
2165 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS];
2166 list_count = ARC_BUFC_NUMDATALISTS;
2167 offset = ARC_BUFC_NUMMETADATALISTS;
2170 list = &list_start[idx];
2171 lock = ARCS_LOCK(state, idx + offset);
2174 for (ab = list_tail(list); ab; ab = ab_prev) {
2175 ab_prev = list_prev(list, ab);
2176 if (ab->b_type > ARC_BUFC_NUMTYPES)
2177 panic("invalid ab=%p", (void *)ab);
2178 if (spa && ab->b_spa != spa)
2181 /* ignore markers */
2185 hash_lock = HDR_LOCK(ab);
2186 /* caller may be trying to modify this buffer, skip it */
2187 if (MUTEX_HELD(hash_lock))
2191 * It may take a long time to evict all the bufs requested.
2192 * To avoid blocking all arc activity, periodically drop
2193 * the arcs_mtx and give other threads a chance to run
2194 * before reacquiring the lock.
2196 if (count++ > arc_evict_iterations) {
2197 list_insert_after(list, ab, &marker);
2199 kpreempt(KPREEMPT_SYNC);
2201 ab_prev = list_prev(list, &marker);
2202 list_remove(list, &marker);
2206 if (mutex_tryenter(hash_lock)) {
2207 ASSERT(!HDR_IO_IN_PROGRESS(ab));
2208 ASSERT(ab->b_buf == NULL);
2209 ARCSTAT_BUMP(arcstat_deleted);
2210 bytes_deleted += ab->b_size;
2212 if (ab->b_l2hdr != NULL) {
2214 * This buffer is cached on the 2nd Level ARC;
2215 * don't destroy the header.
2217 arc_change_state(arc_l2c_only, ab, hash_lock);
2218 mutex_exit(hash_lock);
2220 arc_change_state(arc_anon, ab, hash_lock);
2221 mutex_exit(hash_lock);
2222 arc_hdr_destroy(ab);
2225 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
2226 if (bytes >= 0 && bytes_deleted >= bytes)
2228 } else if (bytes < 0) {
2230 * Insert a list marker and then wait for the
2231 * hash lock to become available. Once its
2232 * available, restart from where we left off.
2234 list_insert_after(list, ab, &marker);
2236 mutex_enter(hash_lock);
2237 mutex_exit(hash_lock);
2239 ab_prev = list_prev(list, &marker);
2240 list_remove(list, &marker);
2247 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1));
2250 if (lists < list_count)
2254 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] &&
2255 (bytes < 0 || bytes_deleted < bytes)) {
2256 list_start = &state->arcs_lists[0];
2257 list_count = ARC_BUFC_NUMMETADATALISTS;
2263 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2267 if (bytes_deleted < bytes)
2268 dprintf("only deleted %lld bytes from %p",
2269 (longlong_t)bytes_deleted, state);
2275 int64_t adjustment, delta;
2281 adjustment = MIN((int64_t)(arc_size - arc_c),
2282 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2285 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2286 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2287 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
2288 adjustment -= delta;
2291 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2292 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2293 (void) arc_evict(arc_mru, 0, delta, FALSE,
2301 adjustment = arc_size - arc_c;
2303 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2304 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2305 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2306 adjustment -= delta;
2309 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2310 int64_t delta = MIN(adjustment,
2311 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2312 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2317 * Adjust ghost lists
2320 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2322 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2323 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2324 arc_evict_ghost(arc_mru_ghost, 0, delta);
2328 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2330 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2331 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2332 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2337 arc_do_user_evicts(void)
2339 static arc_buf_t *tmp_arc_eviction_list;
2342 * Move list over to avoid LOR
2345 mutex_enter(&arc_eviction_mtx);
2346 tmp_arc_eviction_list = arc_eviction_list;
2347 arc_eviction_list = NULL;
2348 mutex_exit(&arc_eviction_mtx);
2350 while (tmp_arc_eviction_list != NULL) {
2351 arc_buf_t *buf = tmp_arc_eviction_list;
2352 tmp_arc_eviction_list = buf->b_next;
2353 mutex_enter(&buf->b_evict_lock);
2355 mutex_exit(&buf->b_evict_lock);
2357 if (buf->b_efunc != NULL)
2358 VERIFY(buf->b_efunc(buf) == 0);
2360 buf->b_efunc = NULL;
2361 buf->b_private = NULL;
2362 kmem_cache_free(buf_cache, buf);
2365 if (arc_eviction_list != NULL)
2370 * Flush all *evictable* data from the cache for the given spa.
2371 * NOTE: this will not touch "active" (i.e. referenced) data.
2374 arc_flush(spa_t *spa)
2379 guid = spa_load_guid(spa);
2381 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) {
2382 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2386 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) {
2387 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2391 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) {
2392 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2396 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) {
2397 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2402 arc_evict_ghost(arc_mru_ghost, guid, -1);
2403 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2405 mutex_enter(&arc_reclaim_thr_lock);
2406 arc_do_user_evicts();
2407 mutex_exit(&arc_reclaim_thr_lock);
2408 ASSERT(spa || arc_eviction_list == NULL);
2414 if (arc_c > arc_c_min) {
2418 to_free = arc_c >> arc_shrink_shift;
2420 to_free = arc_c >> arc_shrink_shift;
2422 if (arc_c > arc_c_min + to_free)
2423 atomic_add_64(&arc_c, -to_free);
2427 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2428 if (arc_c > arc_size)
2429 arc_c = MAX(arc_size, arc_c_min);
2431 arc_p = (arc_c >> 1);
2432 ASSERT(arc_c >= arc_c_min);
2433 ASSERT((int64_t)arc_p >= 0);
2436 if (arc_size > arc_c)
2440 static int needfree = 0;
2443 arc_reclaim_needed(void)
2452 * Cooperate with pagedaemon when it's time for it to scan
2453 * and reclaim some pages.
2455 if (vm_paging_needed())
2460 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2465 * check that we're out of range of the pageout scanner. It starts to
2466 * schedule paging if freemem is less than lotsfree and needfree.
2467 * lotsfree is the high-water mark for pageout, and needfree is the
2468 * number of needed free pages. We add extra pages here to make sure
2469 * the scanner doesn't start up while we're freeing memory.
2471 if (freemem < lotsfree + needfree + extra)
2475 * check to make sure that swapfs has enough space so that anon
2476 * reservations can still succeed. anon_resvmem() checks that the
2477 * availrmem is greater than swapfs_minfree, and the number of reserved
2478 * swap pages. We also add a bit of extra here just to prevent
2479 * circumstances from getting really dire.
2481 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2486 * If we're on an i386 platform, it's possible that we'll exhaust the
2487 * kernel heap space before we ever run out of available physical
2488 * memory. Most checks of the size of the heap_area compare against
2489 * tune.t_minarmem, which is the minimum available real memory that we
2490 * can have in the system. However, this is generally fixed at 25 pages
2491 * which is so low that it's useless. In this comparison, we seek to
2492 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2493 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2496 if (btop(vmem_size(heap_arena, VMEM_FREE)) <
2497 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
2501 if (kmem_used() > (kmem_size() * 3) / 4)
2506 if (spa_get_random(100) == 0)
2512 extern kmem_cache_t *zio_buf_cache[];
2513 extern kmem_cache_t *zio_data_buf_cache[];
2516 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2519 kmem_cache_t *prev_cache = NULL;
2520 kmem_cache_t *prev_data_cache = NULL;
2523 if (arc_meta_used >= arc_meta_limit) {
2525 * We are exceeding our meta-data cache limit.
2526 * Purge some DNLC entries to release holds on meta-data.
2528 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2532 * Reclaim unused memory from all kmem caches.
2539 * An aggressive reclamation will shrink the cache size as well as
2540 * reap free buffers from the arc kmem caches.
2542 if (strat == ARC_RECLAIM_AGGR)
2545 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2546 if (zio_buf_cache[i] != prev_cache) {
2547 prev_cache = zio_buf_cache[i];
2548 kmem_cache_reap_now(zio_buf_cache[i]);
2550 if (zio_data_buf_cache[i] != prev_data_cache) {
2551 prev_data_cache = zio_data_buf_cache[i];
2552 kmem_cache_reap_now(zio_data_buf_cache[i]);
2555 kmem_cache_reap_now(buf_cache);
2556 kmem_cache_reap_now(hdr_cache);
2560 arc_reclaim_thread(void *dummy __unused)
2562 clock_t growtime = 0;
2563 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2566 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2568 mutex_enter(&arc_reclaim_thr_lock);
2569 while (arc_thread_exit == 0) {
2570 if (arc_reclaim_needed()) {
2573 if (last_reclaim == ARC_RECLAIM_CONS) {
2574 last_reclaim = ARC_RECLAIM_AGGR;
2576 last_reclaim = ARC_RECLAIM_CONS;
2580 last_reclaim = ARC_RECLAIM_AGGR;
2584 /* reset the growth delay for every reclaim */
2585 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2587 if (needfree && last_reclaim == ARC_RECLAIM_CONS) {
2589 * If needfree is TRUE our vm_lowmem hook
2590 * was called and in that case we must free some
2591 * memory, so switch to aggressive mode.
2594 last_reclaim = ARC_RECLAIM_AGGR;
2596 arc_kmem_reap_now(last_reclaim);
2599 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2600 arc_no_grow = FALSE;
2605 if (arc_eviction_list != NULL)
2606 arc_do_user_evicts();
2615 /* block until needed, or one second, whichever is shorter */
2616 CALLB_CPR_SAFE_BEGIN(&cpr);
2617 (void) cv_timedwait(&arc_reclaim_thr_cv,
2618 &arc_reclaim_thr_lock, hz);
2619 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2622 arc_thread_exit = 0;
2623 cv_broadcast(&arc_reclaim_thr_cv);
2624 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2629 * Adapt arc info given the number of bytes we are trying to add and
2630 * the state that we are comming from. This function is only called
2631 * when we are adding new content to the cache.
2634 arc_adapt(int bytes, arc_state_t *state)
2637 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2639 if (state == arc_l2c_only)
2644 * Adapt the target size of the MRU list:
2645 * - if we just hit in the MRU ghost list, then increase
2646 * the target size of the MRU list.
2647 * - if we just hit in the MFU ghost list, then increase
2648 * the target size of the MFU list by decreasing the
2649 * target size of the MRU list.
2651 if (state == arc_mru_ghost) {
2652 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2653 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2654 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2656 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2657 } else if (state == arc_mfu_ghost) {
2660 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2661 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2662 mult = MIN(mult, 10);
2664 delta = MIN(bytes * mult, arc_p);
2665 arc_p = MAX(arc_p_min, arc_p - delta);
2667 ASSERT((int64_t)arc_p >= 0);
2669 if (arc_reclaim_needed()) {
2670 cv_signal(&arc_reclaim_thr_cv);
2677 if (arc_c >= arc_c_max)
2681 * If we're within (2 * maxblocksize) bytes of the target
2682 * cache size, increment the target cache size
2684 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2685 atomic_add_64(&arc_c, (int64_t)bytes);
2686 if (arc_c > arc_c_max)
2688 else if (state == arc_anon)
2689 atomic_add_64(&arc_p, (int64_t)bytes);
2693 ASSERT((int64_t)arc_p >= 0);
2697 * Check if the cache has reached its limits and eviction is required
2701 arc_evict_needed(arc_buf_contents_t type)
2703 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2709 * If zio data pages are being allocated out of a separate heap segment,
2710 * then enforce that the size of available vmem for this area remains
2711 * above about 1/32nd free.
2713 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2714 vmem_size(zio_arena, VMEM_FREE) <
2715 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2720 if (arc_reclaim_needed())
2723 return (arc_size > arc_c);
2727 * The buffer, supplied as the first argument, needs a data block.
2728 * So, if we are at cache max, determine which cache should be victimized.
2729 * We have the following cases:
2731 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2732 * In this situation if we're out of space, but the resident size of the MFU is
2733 * under the limit, victimize the MFU cache to satisfy this insertion request.
2735 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2736 * Here, we've used up all of the available space for the MRU, so we need to
2737 * evict from our own cache instead. Evict from the set of resident MRU
2740 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2741 * c minus p represents the MFU space in the cache, since p is the size of the
2742 * cache that is dedicated to the MRU. In this situation there's still space on
2743 * the MFU side, so the MRU side needs to be victimized.
2745 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2746 * MFU's resident set is consuming more space than it has been allotted. In
2747 * this situation, we must victimize our own cache, the MFU, for this insertion.
2750 arc_get_data_buf(arc_buf_t *buf)
2752 arc_state_t *state = buf->b_hdr->b_state;
2753 uint64_t size = buf->b_hdr->b_size;
2754 arc_buf_contents_t type = buf->b_hdr->b_type;
2756 arc_adapt(size, state);
2759 * We have not yet reached cache maximum size,
2760 * just allocate a new buffer.
2762 if (!arc_evict_needed(type)) {
2763 if (type == ARC_BUFC_METADATA) {
2764 buf->b_data = zio_buf_alloc(size);
2765 arc_space_consume(size, ARC_SPACE_DATA);
2767 ASSERT(type == ARC_BUFC_DATA);
2768 buf->b_data = zio_data_buf_alloc(size);
2769 ARCSTAT_INCR(arcstat_data_size, size);
2770 atomic_add_64(&arc_size, size);
2776 * If we are prefetching from the mfu ghost list, this buffer
2777 * will end up on the mru list; so steal space from there.
2779 if (state == arc_mfu_ghost)
2780 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2781 else if (state == arc_mru_ghost)
2784 if (state == arc_mru || state == arc_anon) {
2785 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2786 state = (arc_mfu->arcs_lsize[type] >= size &&
2787 arc_p > mru_used) ? arc_mfu : arc_mru;
2790 uint64_t mfu_space = arc_c - arc_p;
2791 state = (arc_mru->arcs_lsize[type] >= size &&
2792 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2794 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2795 if (type == ARC_BUFC_METADATA) {
2796 buf->b_data = zio_buf_alloc(size);
2797 arc_space_consume(size, ARC_SPACE_DATA);
2799 ASSERT(type == ARC_BUFC_DATA);
2800 buf->b_data = zio_data_buf_alloc(size);
2801 ARCSTAT_INCR(arcstat_data_size, size);
2802 atomic_add_64(&arc_size, size);
2804 ARCSTAT_BUMP(arcstat_recycle_miss);
2806 ASSERT(buf->b_data != NULL);
2809 * Update the state size. Note that ghost states have a
2810 * "ghost size" and so don't need to be updated.
2812 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2813 arc_buf_hdr_t *hdr = buf->b_hdr;
2815 atomic_add_64(&hdr->b_state->arcs_size, size);
2816 if (list_link_active(&hdr->b_arc_node)) {
2817 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2818 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2821 * If we are growing the cache, and we are adding anonymous
2822 * data, and we have outgrown arc_p, update arc_p
2824 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2825 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2826 arc_p = MIN(arc_c, arc_p + size);
2828 ARCSTAT_BUMP(arcstat_allocated);
2832 * This routine is called whenever a buffer is accessed.
2833 * NOTE: the hash lock is dropped in this function.
2836 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2840 ASSERT(MUTEX_HELD(hash_lock));
2842 if (buf->b_state == arc_anon) {
2844 * This buffer is not in the cache, and does not
2845 * appear in our "ghost" list. Add the new buffer
2849 ASSERT(buf->b_arc_access == 0);
2850 buf->b_arc_access = ddi_get_lbolt();
2851 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2852 arc_change_state(arc_mru, buf, hash_lock);
2854 } else if (buf->b_state == arc_mru) {
2855 now = ddi_get_lbolt();
2858 * If this buffer is here because of a prefetch, then either:
2859 * - clear the flag if this is a "referencing" read
2860 * (any subsequent access will bump this into the MFU state).
2862 * - move the buffer to the head of the list if this is
2863 * another prefetch (to make it less likely to be evicted).
2865 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2866 if (refcount_count(&buf->b_refcnt) == 0) {
2867 ASSERT(list_link_active(&buf->b_arc_node));
2869 buf->b_flags &= ~ARC_PREFETCH;
2870 ARCSTAT_BUMP(arcstat_mru_hits);
2872 buf->b_arc_access = now;
2877 * This buffer has been "accessed" only once so far,
2878 * but it is still in the cache. Move it to the MFU
2881 if (now > buf->b_arc_access + ARC_MINTIME) {
2883 * More than 125ms have passed since we
2884 * instantiated this buffer. Move it to the
2885 * most frequently used state.
2887 buf->b_arc_access = now;
2888 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2889 arc_change_state(arc_mfu, buf, hash_lock);
2891 ARCSTAT_BUMP(arcstat_mru_hits);
2892 } else if (buf->b_state == arc_mru_ghost) {
2893 arc_state_t *new_state;
2895 * This buffer has been "accessed" recently, but
2896 * was evicted from the cache. Move it to the
2900 if (buf->b_flags & ARC_PREFETCH) {
2901 new_state = arc_mru;
2902 if (refcount_count(&buf->b_refcnt) > 0)
2903 buf->b_flags &= ~ARC_PREFETCH;
2904 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2906 new_state = arc_mfu;
2907 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2910 buf->b_arc_access = ddi_get_lbolt();
2911 arc_change_state(new_state, buf, hash_lock);
2913 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2914 } else if (buf->b_state == arc_mfu) {
2916 * This buffer has been accessed more than once and is
2917 * still in the cache. Keep it in the MFU state.
2919 * NOTE: an add_reference() that occurred when we did
2920 * the arc_read() will have kicked this off the list.
2921 * If it was a prefetch, we will explicitly move it to
2922 * the head of the list now.
2924 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2925 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2926 ASSERT(list_link_active(&buf->b_arc_node));
2928 ARCSTAT_BUMP(arcstat_mfu_hits);
2929 buf->b_arc_access = ddi_get_lbolt();
2930 } else if (buf->b_state == arc_mfu_ghost) {
2931 arc_state_t *new_state = arc_mfu;
2933 * This buffer has been accessed more than once but has
2934 * been evicted from the cache. Move it back to the
2938 if (buf->b_flags & ARC_PREFETCH) {
2940 * This is a prefetch access...
2941 * move this block back to the MRU state.
2943 ASSERT0(refcount_count(&buf->b_refcnt));
2944 new_state = arc_mru;
2947 buf->b_arc_access = ddi_get_lbolt();
2948 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2949 arc_change_state(new_state, buf, hash_lock);
2951 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2952 } else if (buf->b_state == arc_l2c_only) {
2954 * This buffer is on the 2nd Level ARC.
2957 buf->b_arc_access = ddi_get_lbolt();
2958 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2959 arc_change_state(arc_mfu, buf, hash_lock);
2961 ASSERT(!"invalid arc state");
2965 /* a generic arc_done_func_t which you can use */
2968 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2970 if (zio == NULL || zio->io_error == 0)
2971 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2972 VERIFY(arc_buf_remove_ref(buf, arg));
2975 /* a generic arc_done_func_t */
2977 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2979 arc_buf_t **bufp = arg;
2980 if (zio && zio->io_error) {
2981 VERIFY(arc_buf_remove_ref(buf, arg));
2985 ASSERT(buf->b_data);
2990 arc_read_done(zio_t *zio)
2994 arc_buf_t *abuf; /* buffer we're assigning to callback */
2995 kmutex_t *hash_lock = NULL;
2996 arc_callback_t *callback_list, *acb;
2997 int freeable = FALSE;
2999 buf = zio->io_private;
3003 * The hdr was inserted into hash-table and removed from lists
3004 * prior to starting I/O. We should find this header, since
3005 * it's in the hash table, and it should be legit since it's
3006 * not possible to evict it during the I/O. The only possible
3007 * reason for it not to be found is if we were freed during the
3010 if (HDR_IN_HASH_TABLE(hdr)) {
3011 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
3012 ASSERT3U(hdr->b_dva.dva_word[0], ==,
3013 BP_IDENTITY(zio->io_bp)->dva_word[0]);
3014 ASSERT3U(hdr->b_dva.dva_word[1], ==,
3015 BP_IDENTITY(zio->io_bp)->dva_word[1]);
3017 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
3020 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
3021 hash_lock == NULL) ||
3023 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
3024 (found == hdr && HDR_L2_READING(hdr)));
3027 hdr->b_flags &= ~ARC_L2_EVICTED;
3028 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
3029 hdr->b_flags &= ~ARC_L2CACHE;
3031 /* byteswap if necessary */
3032 callback_list = hdr->b_acb;
3033 ASSERT(callback_list != NULL);
3034 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
3035 dmu_object_byteswap_t bswap =
3036 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
3037 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
3038 byteswap_uint64_array :
3039 dmu_ot_byteswap[bswap].ob_func;
3040 func(buf->b_data, hdr->b_size);
3043 arc_cksum_compute(buf, B_FALSE);
3046 #endif /* illumos */
3048 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
3050 * Only call arc_access on anonymous buffers. This is because
3051 * if we've issued an I/O for an evicted buffer, we've already
3052 * called arc_access (to prevent any simultaneous readers from
3053 * getting confused).
3055 arc_access(hdr, hash_lock);
3058 /* create copies of the data buffer for the callers */
3060 for (acb = callback_list; acb; acb = acb->acb_next) {
3061 if (acb->acb_done) {
3063 ARCSTAT_BUMP(arcstat_duplicate_reads);
3064 abuf = arc_buf_clone(buf);
3066 acb->acb_buf = abuf;
3071 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3072 ASSERT(!HDR_BUF_AVAILABLE(hdr));
3074 ASSERT(buf->b_efunc == NULL);
3075 ASSERT(hdr->b_datacnt == 1);
3076 hdr->b_flags |= ARC_BUF_AVAILABLE;
3079 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
3081 if (zio->io_error != 0) {
3082 hdr->b_flags |= ARC_IO_ERROR;
3083 if (hdr->b_state != arc_anon)
3084 arc_change_state(arc_anon, hdr, hash_lock);
3085 if (HDR_IN_HASH_TABLE(hdr))
3086 buf_hash_remove(hdr);
3087 freeable = refcount_is_zero(&hdr->b_refcnt);
3091 * Broadcast before we drop the hash_lock to avoid the possibility
3092 * that the hdr (and hence the cv) might be freed before we get to
3093 * the cv_broadcast().
3095 cv_broadcast(&hdr->b_cv);
3098 mutex_exit(hash_lock);
3101 * This block was freed while we waited for the read to
3102 * complete. It has been removed from the hash table and
3103 * moved to the anonymous state (so that it won't show up
3106 ASSERT3P(hdr->b_state, ==, arc_anon);
3107 freeable = refcount_is_zero(&hdr->b_refcnt);
3110 /* execute each callback and free its structure */
3111 while ((acb = callback_list) != NULL) {
3113 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3115 if (acb->acb_zio_dummy != NULL) {
3116 acb->acb_zio_dummy->io_error = zio->io_error;
3117 zio_nowait(acb->acb_zio_dummy);
3120 callback_list = acb->acb_next;
3121 kmem_free(acb, sizeof (arc_callback_t));
3125 arc_hdr_destroy(hdr);
3129 * "Read" the block block at the specified DVA (in bp) via the
3130 * cache. If the block is found in the cache, invoke the provided
3131 * callback immediately and return. Note that the `zio' parameter
3132 * in the callback will be NULL in this case, since no IO was
3133 * required. If the block is not in the cache pass the read request
3134 * on to the spa with a substitute callback function, so that the
3135 * requested block will be added to the cache.
3137 * If a read request arrives for a block that has a read in-progress,
3138 * either wait for the in-progress read to complete (and return the
3139 * results); or, if this is a read with a "done" func, add a record
3140 * to the read to invoke the "done" func when the read completes,
3141 * and return; or just return.
3143 * arc_read_done() will invoke all the requested "done" functions
3144 * for readers of this block.
3147 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3148 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags,
3149 const zbookmark_phys_t *zb)
3151 arc_buf_hdr_t *hdr = NULL;
3152 arc_buf_t *buf = NULL;
3153 kmutex_t *hash_lock = NULL;
3155 uint64_t guid = spa_load_guid(spa);
3157 ASSERT(!BP_IS_EMBEDDED(bp) ||
3158 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
3161 if (!BP_IS_EMBEDDED(bp)) {
3163 * Embedded BP's have no DVA and require no I/O to "read".
3164 * Create an anonymous arc buf to back it.
3166 hdr = buf_hash_find(guid, bp, &hash_lock);
3169 if (hdr != NULL && hdr->b_datacnt > 0) {
3171 *arc_flags |= ARC_CACHED;
3173 if (HDR_IO_IN_PROGRESS(hdr)) {
3175 if (*arc_flags & ARC_WAIT) {
3176 cv_wait(&hdr->b_cv, hash_lock);
3177 mutex_exit(hash_lock);
3180 ASSERT(*arc_flags & ARC_NOWAIT);
3183 arc_callback_t *acb = NULL;
3185 acb = kmem_zalloc(sizeof (arc_callback_t),
3187 acb->acb_done = done;
3188 acb->acb_private = private;
3190 acb->acb_zio_dummy = zio_null(pio,
3191 spa, NULL, NULL, NULL, zio_flags);
3193 ASSERT(acb->acb_done != NULL);
3194 acb->acb_next = hdr->b_acb;
3196 add_reference(hdr, hash_lock, private);
3197 mutex_exit(hash_lock);
3200 mutex_exit(hash_lock);
3204 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3207 add_reference(hdr, hash_lock, private);
3209 * If this block is already in use, create a new
3210 * copy of the data so that we will be guaranteed
3211 * that arc_release() will always succeed.
3215 ASSERT(buf->b_data);
3216 if (HDR_BUF_AVAILABLE(hdr)) {
3217 ASSERT(buf->b_efunc == NULL);
3218 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3220 buf = arc_buf_clone(buf);
3223 } else if (*arc_flags & ARC_PREFETCH &&
3224 refcount_count(&hdr->b_refcnt) == 0) {
3225 hdr->b_flags |= ARC_PREFETCH;
3227 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
3228 arc_access(hdr, hash_lock);
3229 if (*arc_flags & ARC_L2CACHE)
3230 hdr->b_flags |= ARC_L2CACHE;
3231 if (*arc_flags & ARC_L2COMPRESS)
3232 hdr->b_flags |= ARC_L2COMPRESS;
3233 mutex_exit(hash_lock);
3234 ARCSTAT_BUMP(arcstat_hits);
3235 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3236 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3237 data, metadata, hits);
3240 done(NULL, buf, private);
3242 uint64_t size = BP_GET_LSIZE(bp);
3243 arc_callback_t *acb;
3246 boolean_t devw = B_FALSE;
3247 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
3248 uint64_t b_asize = 0;
3251 /* this block is not in the cache */
3252 arc_buf_hdr_t *exists = NULL;
3253 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
3254 buf = arc_buf_alloc(spa, size, private, type);
3256 if (!BP_IS_EMBEDDED(bp)) {
3257 hdr->b_dva = *BP_IDENTITY(bp);
3258 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3259 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3260 exists = buf_hash_insert(hdr, &hash_lock);
3262 if (exists != NULL) {
3263 /* somebody beat us to the hash insert */
3264 mutex_exit(hash_lock);
3265 buf_discard_identity(hdr);
3266 (void) arc_buf_remove_ref(buf, private);
3267 goto top; /* restart the IO request */
3269 /* if this is a prefetch, we don't have a reference */
3270 if (*arc_flags & ARC_PREFETCH) {
3271 (void) remove_reference(hdr, hash_lock,
3273 hdr->b_flags |= ARC_PREFETCH;
3275 if (*arc_flags & ARC_L2CACHE)
3276 hdr->b_flags |= ARC_L2CACHE;
3277 if (*arc_flags & ARC_L2COMPRESS)
3278 hdr->b_flags |= ARC_L2COMPRESS;
3279 if (BP_GET_LEVEL(bp) > 0)
3280 hdr->b_flags |= ARC_INDIRECT;
3282 /* this block is in the ghost cache */
3283 ASSERT(GHOST_STATE(hdr->b_state));
3284 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3285 ASSERT0(refcount_count(&hdr->b_refcnt));
3286 ASSERT(hdr->b_buf == NULL);
3288 /* if this is a prefetch, we don't have a reference */
3289 if (*arc_flags & ARC_PREFETCH)
3290 hdr->b_flags |= ARC_PREFETCH;
3292 add_reference(hdr, hash_lock, private);
3293 if (*arc_flags & ARC_L2CACHE)
3294 hdr->b_flags |= ARC_L2CACHE;
3295 if (*arc_flags & ARC_L2COMPRESS)
3296 hdr->b_flags |= ARC_L2COMPRESS;
3297 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3300 buf->b_efunc = NULL;
3301 buf->b_private = NULL;
3304 ASSERT(hdr->b_datacnt == 0);
3306 arc_get_data_buf(buf);
3307 arc_access(hdr, hash_lock);
3310 ASSERT(!GHOST_STATE(hdr->b_state));
3312 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
3313 acb->acb_done = done;
3314 acb->acb_private = private;
3316 ASSERT(hdr->b_acb == NULL);
3318 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3320 if (hdr->b_l2hdr != NULL &&
3321 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3322 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3323 addr = hdr->b_l2hdr->b_daddr;
3324 b_compress = hdr->b_l2hdr->b_compress;
3325 b_asize = hdr->b_l2hdr->b_asize;
3327 * Lock out device removal.
3329 if (vdev_is_dead(vd) ||
3330 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3334 if (hash_lock != NULL)
3335 mutex_exit(hash_lock);
3338 * At this point, we have a level 1 cache miss. Try again in
3339 * L2ARC if possible.
3341 ASSERT3U(hdr->b_size, ==, size);
3342 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3343 uint64_t, size, zbookmark_phys_t *, zb);
3344 ARCSTAT_BUMP(arcstat_misses);
3345 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3346 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3347 data, metadata, misses);
3349 curthread->td_ru.ru_inblock++;
3352 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3354 * Read from the L2ARC if the following are true:
3355 * 1. The L2ARC vdev was previously cached.
3356 * 2. This buffer still has L2ARC metadata.
3357 * 3. This buffer isn't currently writing to the L2ARC.
3358 * 4. The L2ARC entry wasn't evicted, which may
3359 * also have invalidated the vdev.
3360 * 5. This isn't prefetch and l2arc_noprefetch is set.
3362 if (hdr->b_l2hdr != NULL &&
3363 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3364 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3365 l2arc_read_callback_t *cb;
3367 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3368 ARCSTAT_BUMP(arcstat_l2_hits);
3370 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3372 cb->l2rcb_buf = buf;
3373 cb->l2rcb_spa = spa;
3376 cb->l2rcb_flags = zio_flags;
3377 cb->l2rcb_compress = b_compress;
3379 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
3380 addr + size < vd->vdev_psize -
3381 VDEV_LABEL_END_SIZE);
3384 * l2arc read. The SCL_L2ARC lock will be
3385 * released by l2arc_read_done().
3386 * Issue a null zio if the underlying buffer
3387 * was squashed to zero size by compression.
3389 if (b_compress == ZIO_COMPRESS_EMPTY) {
3390 rzio = zio_null(pio, spa, vd,
3391 l2arc_read_done, cb,
3392 zio_flags | ZIO_FLAG_DONT_CACHE |
3394 ZIO_FLAG_DONT_PROPAGATE |
3395 ZIO_FLAG_DONT_RETRY);
3397 rzio = zio_read_phys(pio, vd, addr,
3398 b_asize, buf->b_data,
3400 l2arc_read_done, cb, priority,
3401 zio_flags | ZIO_FLAG_DONT_CACHE |
3403 ZIO_FLAG_DONT_PROPAGATE |
3404 ZIO_FLAG_DONT_RETRY, B_FALSE);
3406 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3408 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
3410 if (*arc_flags & ARC_NOWAIT) {
3415 ASSERT(*arc_flags & ARC_WAIT);
3416 if (zio_wait(rzio) == 0)
3419 /* l2arc read error; goto zio_read() */
3421 DTRACE_PROBE1(l2arc__miss,
3422 arc_buf_hdr_t *, hdr);
3423 ARCSTAT_BUMP(arcstat_l2_misses);
3424 if (HDR_L2_WRITING(hdr))
3425 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3426 spa_config_exit(spa, SCL_L2ARC, vd);
3430 spa_config_exit(spa, SCL_L2ARC, vd);
3431 if (l2arc_ndev != 0) {
3432 DTRACE_PROBE1(l2arc__miss,
3433 arc_buf_hdr_t *, hdr);
3434 ARCSTAT_BUMP(arcstat_l2_misses);
3438 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3439 arc_read_done, buf, priority, zio_flags, zb);
3441 if (*arc_flags & ARC_WAIT)
3442 return (zio_wait(rzio));
3444 ASSERT(*arc_flags & ARC_NOWAIT);
3451 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3453 ASSERT(buf->b_hdr != NULL);
3454 ASSERT(buf->b_hdr->b_state != arc_anon);
3455 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3456 ASSERT(buf->b_efunc == NULL);
3457 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3459 buf->b_efunc = func;
3460 buf->b_private = private;
3464 * Notify the arc that a block was freed, and thus will never be used again.
3467 arc_freed(spa_t *spa, const blkptr_t *bp)
3470 kmutex_t *hash_lock;
3471 uint64_t guid = spa_load_guid(spa);
3473 ASSERT(!BP_IS_EMBEDDED(bp));
3475 hdr = buf_hash_find(guid, bp, &hash_lock);
3478 if (HDR_BUF_AVAILABLE(hdr)) {
3479 arc_buf_t *buf = hdr->b_buf;
3480 add_reference(hdr, hash_lock, FTAG);
3481 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3482 mutex_exit(hash_lock);
3484 arc_release(buf, FTAG);
3485 (void) arc_buf_remove_ref(buf, FTAG);
3487 mutex_exit(hash_lock);
3493 * This is used by the DMU to let the ARC know that a buffer is
3494 * being evicted, so the ARC should clean up. If this arc buf
3495 * is not yet in the evicted state, it will be put there.
3498 arc_buf_evict(arc_buf_t *buf)
3501 kmutex_t *hash_lock;
3503 list_t *list, *evicted_list;
3504 kmutex_t *lock, *evicted_lock;
3506 mutex_enter(&buf->b_evict_lock);
3510 * We are in arc_do_user_evicts().
3512 ASSERT(buf->b_data == NULL);
3513 mutex_exit(&buf->b_evict_lock);
3515 } else if (buf->b_data == NULL) {
3516 arc_buf_t copy = *buf; /* structure assignment */
3518 * We are on the eviction list; process this buffer now
3519 * but let arc_do_user_evicts() do the reaping.
3521 buf->b_efunc = NULL;
3522 mutex_exit(&buf->b_evict_lock);
3523 VERIFY(copy.b_efunc(©) == 0);
3526 hash_lock = HDR_LOCK(hdr);
3527 mutex_enter(hash_lock);
3529 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3531 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3532 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3535 * Pull this buffer off of the hdr
3538 while (*bufp != buf)
3539 bufp = &(*bufp)->b_next;
3540 *bufp = buf->b_next;
3542 ASSERT(buf->b_data != NULL);
3543 arc_buf_destroy(buf, FALSE, FALSE);
3545 if (hdr->b_datacnt == 0) {
3546 arc_state_t *old_state = hdr->b_state;
3547 arc_state_t *evicted_state;
3549 ASSERT(hdr->b_buf == NULL);
3550 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3553 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3555 get_buf_info(hdr, old_state, &list, &lock);
3556 get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock);
3558 mutex_enter(evicted_lock);
3560 arc_change_state(evicted_state, hdr, hash_lock);
3561 ASSERT(HDR_IN_HASH_TABLE(hdr));
3562 hdr->b_flags |= ARC_IN_HASH_TABLE;
3563 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3565 mutex_exit(evicted_lock);
3568 mutex_exit(hash_lock);
3569 mutex_exit(&buf->b_evict_lock);
3571 VERIFY(buf->b_efunc(buf) == 0);
3572 buf->b_efunc = NULL;
3573 buf->b_private = NULL;
3576 kmem_cache_free(buf_cache, buf);
3581 * Release this buffer from the cache, making it an anonymous buffer. This
3582 * must be done after a read and prior to modifying the buffer contents.
3583 * If the buffer has more than one reference, we must make
3584 * a new hdr for the buffer.
3587 arc_release(arc_buf_t *buf, void *tag)
3590 kmutex_t *hash_lock = NULL;
3591 l2arc_buf_hdr_t *l2hdr;
3595 * It would be nice to assert that if it's DMU metadata (level >
3596 * 0 || it's the dnode file), then it must be syncing context.
3597 * But we don't know that information at this level.
3600 mutex_enter(&buf->b_evict_lock);
3603 /* this buffer is not on any list */
3604 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3606 if (hdr->b_state == arc_anon) {
3607 /* this buffer is already released */
3608 ASSERT(buf->b_efunc == NULL);
3610 hash_lock = HDR_LOCK(hdr);
3611 mutex_enter(hash_lock);
3613 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3616 l2hdr = hdr->b_l2hdr;
3618 mutex_enter(&l2arc_buflist_mtx);
3619 hdr->b_l2hdr = NULL;
3620 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3622 buf_size = hdr->b_size;
3625 * Do we have more than one buf?
3627 if (hdr->b_datacnt > 1) {
3628 arc_buf_hdr_t *nhdr;
3630 uint64_t blksz = hdr->b_size;
3631 uint64_t spa = hdr->b_spa;
3632 arc_buf_contents_t type = hdr->b_type;
3633 uint32_t flags = hdr->b_flags;
3635 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3637 * Pull the data off of this hdr and attach it to
3638 * a new anonymous hdr.
3640 (void) remove_reference(hdr, hash_lock, tag);
3642 while (*bufp != buf)
3643 bufp = &(*bufp)->b_next;
3644 *bufp = buf->b_next;
3647 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3648 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3649 if (refcount_is_zero(&hdr->b_refcnt)) {
3650 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3651 ASSERT3U(*size, >=, hdr->b_size);
3652 atomic_add_64(size, -hdr->b_size);
3656 * We're releasing a duplicate user data buffer, update
3657 * our statistics accordingly.
3659 if (hdr->b_type == ARC_BUFC_DATA) {
3660 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3661 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3664 hdr->b_datacnt -= 1;
3665 arc_cksum_verify(buf);
3667 arc_buf_unwatch(buf);
3668 #endif /* illumos */
3670 mutex_exit(hash_lock);
3672 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3673 nhdr->b_size = blksz;
3675 nhdr->b_type = type;
3677 nhdr->b_state = arc_anon;
3678 nhdr->b_arc_access = 0;
3679 nhdr->b_flags = flags & ARC_L2_WRITING;
3680 nhdr->b_l2hdr = NULL;
3681 nhdr->b_datacnt = 1;
3682 nhdr->b_freeze_cksum = NULL;
3683 (void) refcount_add(&nhdr->b_refcnt, tag);
3685 mutex_exit(&buf->b_evict_lock);
3686 atomic_add_64(&arc_anon->arcs_size, blksz);
3688 mutex_exit(&buf->b_evict_lock);
3689 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3690 ASSERT(!list_link_active(&hdr->b_arc_node));
3691 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3692 if (hdr->b_state != arc_anon)
3693 arc_change_state(arc_anon, hdr, hash_lock);
3694 hdr->b_arc_access = 0;
3696 mutex_exit(hash_lock);
3698 buf_discard_identity(hdr);
3701 buf->b_efunc = NULL;
3702 buf->b_private = NULL;
3705 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
3706 vdev_space_update(l2hdr->b_dev->l2ad_vdev,
3707 -l2hdr->b_asize, 0, 0);
3708 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr,
3710 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3711 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3712 mutex_exit(&l2arc_buflist_mtx);
3717 arc_released(arc_buf_t *buf)
3721 mutex_enter(&buf->b_evict_lock);
3722 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3723 mutex_exit(&buf->b_evict_lock);
3728 arc_has_callback(arc_buf_t *buf)
3732 mutex_enter(&buf->b_evict_lock);
3733 callback = (buf->b_efunc != NULL);
3734 mutex_exit(&buf->b_evict_lock);
3740 arc_referenced(arc_buf_t *buf)
3744 mutex_enter(&buf->b_evict_lock);
3745 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3746 mutex_exit(&buf->b_evict_lock);
3747 return (referenced);
3752 arc_write_ready(zio_t *zio)
3754 arc_write_callback_t *callback = zio->io_private;
3755 arc_buf_t *buf = callback->awcb_buf;
3756 arc_buf_hdr_t *hdr = buf->b_hdr;
3758 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3759 callback->awcb_ready(zio, buf, callback->awcb_private);
3762 * If the IO is already in progress, then this is a re-write
3763 * attempt, so we need to thaw and re-compute the cksum.
3764 * It is the responsibility of the callback to handle the
3765 * accounting for any re-write attempt.
3767 if (HDR_IO_IN_PROGRESS(hdr)) {
3768 mutex_enter(&hdr->b_freeze_lock);
3769 if (hdr->b_freeze_cksum != NULL) {
3770 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3771 hdr->b_freeze_cksum = NULL;
3773 mutex_exit(&hdr->b_freeze_lock);
3775 arc_cksum_compute(buf, B_FALSE);
3776 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3780 * The SPA calls this callback for each physical write that happens on behalf
3781 * of a logical write. See the comment in dbuf_write_physdone() for details.
3784 arc_write_physdone(zio_t *zio)
3786 arc_write_callback_t *cb = zio->io_private;
3787 if (cb->awcb_physdone != NULL)
3788 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
3792 arc_write_done(zio_t *zio)
3794 arc_write_callback_t *callback = zio->io_private;
3795 arc_buf_t *buf = callback->awcb_buf;
3796 arc_buf_hdr_t *hdr = buf->b_hdr;
3798 ASSERT(hdr->b_acb == NULL);
3800 if (zio->io_error == 0) {
3801 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
3802 buf_discard_identity(hdr);
3804 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3805 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3806 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3809 ASSERT(BUF_EMPTY(hdr));
3813 * If the block to be written was all-zero or compressed enough to be
3814 * embedded in the BP, no write was performed so there will be no
3815 * dva/birth/checksum. The buffer must therefore remain anonymous
3818 if (!BUF_EMPTY(hdr)) {
3819 arc_buf_hdr_t *exists;
3820 kmutex_t *hash_lock;
3822 ASSERT(zio->io_error == 0);
3824 arc_cksum_verify(buf);
3826 exists = buf_hash_insert(hdr, &hash_lock);
3829 * This can only happen if we overwrite for
3830 * sync-to-convergence, because we remove
3831 * buffers from the hash table when we arc_free().
3833 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3834 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3835 panic("bad overwrite, hdr=%p exists=%p",
3836 (void *)hdr, (void *)exists);
3837 ASSERT(refcount_is_zero(&exists->b_refcnt));
3838 arc_change_state(arc_anon, exists, hash_lock);
3839 mutex_exit(hash_lock);
3840 arc_hdr_destroy(exists);
3841 exists = buf_hash_insert(hdr, &hash_lock);
3842 ASSERT3P(exists, ==, NULL);
3843 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3845 ASSERT(zio->io_prop.zp_nopwrite);
3846 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3847 panic("bad nopwrite, hdr=%p exists=%p",
3848 (void *)hdr, (void *)exists);
3851 ASSERT(hdr->b_datacnt == 1);
3852 ASSERT(hdr->b_state == arc_anon);
3853 ASSERT(BP_GET_DEDUP(zio->io_bp));
3854 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3857 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3858 /* if it's not anon, we are doing a scrub */
3859 if (!exists && hdr->b_state == arc_anon)
3860 arc_access(hdr, hash_lock);
3861 mutex_exit(hash_lock);
3863 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3866 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3867 callback->awcb_done(zio, buf, callback->awcb_private);
3869 kmem_free(callback, sizeof (arc_write_callback_t));
3873 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3874 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
3875 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
3876 arc_done_func_t *done, void *private, zio_priority_t priority,
3877 int zio_flags, const zbookmark_phys_t *zb)
3879 arc_buf_hdr_t *hdr = buf->b_hdr;
3880 arc_write_callback_t *callback;
3883 ASSERT(ready != NULL);
3884 ASSERT(done != NULL);
3885 ASSERT(!HDR_IO_ERROR(hdr));
3886 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3887 ASSERT(hdr->b_acb == NULL);
3889 hdr->b_flags |= ARC_L2CACHE;
3891 hdr->b_flags |= ARC_L2COMPRESS;
3892 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3893 callback->awcb_ready = ready;
3894 callback->awcb_physdone = physdone;
3895 callback->awcb_done = done;
3896 callback->awcb_private = private;
3897 callback->awcb_buf = buf;
3899 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3900 arc_write_ready, arc_write_physdone, arc_write_done, callback,
3901 priority, zio_flags, zb);
3907 arc_memory_throttle(uint64_t reserve, uint64_t txg)
3910 uint64_t available_memory =
3911 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count);
3912 static uint64_t page_load = 0;
3913 static uint64_t last_txg = 0;
3918 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3922 if (cnt.v_free_count + cnt.v_cache_count >
3923 (uint64_t)physmem * arc_lotsfree_percent / 100)
3926 if (txg > last_txg) {
3931 * If we are in pageout, we know that memory is already tight,
3932 * the arc is already going to be evicting, so we just want to
3933 * continue to let page writes occur as quickly as possible.
3935 if (curproc == pageproc) {
3936 if (page_load > available_memory / 4)
3937 return (SET_ERROR(ERESTART));
3938 /* Note: reserve is inflated, so we deflate */
3939 page_load += reserve / 8;
3941 } else if (page_load > 0 && arc_reclaim_needed()) {
3942 /* memory is low, delay before restarting */
3943 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3944 return (SET_ERROR(EAGAIN));
3952 arc_tempreserve_clear(uint64_t reserve)
3954 atomic_add_64(&arc_tempreserve, -reserve);
3955 ASSERT((int64_t)arc_tempreserve >= 0);
3959 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3964 if (reserve > arc_c/4 && !arc_no_grow)
3965 arc_c = MIN(arc_c_max, reserve * 4);
3966 if (reserve > arc_c)
3967 return (SET_ERROR(ENOMEM));
3970 * Don't count loaned bufs as in flight dirty data to prevent long
3971 * network delays from blocking transactions that are ready to be
3972 * assigned to a txg.
3974 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3977 * Writes will, almost always, require additional memory allocations
3978 * in order to compress/encrypt/etc the data. We therefore need to
3979 * make sure that there is sufficient available memory for this.
3981 error = arc_memory_throttle(reserve, txg);
3986 * Throttle writes when the amount of dirty data in the cache
3987 * gets too large. We try to keep the cache less than half full
3988 * of dirty blocks so that our sync times don't grow too large.
3989 * Note: if two requests come in concurrently, we might let them
3990 * both succeed, when one of them should fail. Not a huge deal.
3993 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3994 anon_size > arc_c / 4) {
3995 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3996 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3997 arc_tempreserve>>10,
3998 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3999 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
4000 reserve>>10, arc_c>>10);
4001 return (SET_ERROR(ERESTART));
4003 atomic_add_64(&arc_tempreserve, reserve);
4007 static kmutex_t arc_lowmem_lock;
4009 static eventhandler_tag arc_event_lowmem = NULL;
4012 arc_lowmem(void *arg __unused, int howto __unused)
4015 /* Serialize access via arc_lowmem_lock. */
4016 mutex_enter(&arc_lowmem_lock);
4017 mutex_enter(&arc_reclaim_thr_lock);
4019 cv_signal(&arc_reclaim_thr_cv);
4022 * It is unsafe to block here in arbitrary threads, because we can come
4023 * here from ARC itself and may hold ARC locks and thus risk a deadlock
4024 * with ARC reclaim thread.
4026 if (curproc == pageproc) {
4028 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0);
4030 mutex_exit(&arc_reclaim_thr_lock);
4031 mutex_exit(&arc_lowmem_lock);
4038 int i, prefetch_tunable_set = 0;
4040 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4041 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
4042 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL);
4044 /* Convert seconds to clock ticks */
4045 arc_min_prefetch_lifespan = 1 * hz;
4047 /* Start out with 1/8 of all memory */
4048 arc_c = kmem_size() / 8;
4053 * On architectures where the physical memory can be larger
4054 * than the addressable space (intel in 32-bit mode), we may
4055 * need to limit the cache to 1/8 of VM size.
4057 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
4060 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
4061 arc_c_min = MAX(arc_c / 4, 64<<18);
4062 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
4063 if (arc_c * 8 >= 1<<30)
4064 arc_c_max = (arc_c * 8) - (1<<30);
4066 arc_c_max = arc_c_min;
4067 arc_c_max = MAX(arc_c * 5, arc_c_max);
4071 * Allow the tunables to override our calculations if they are
4072 * reasonable (ie. over 16MB)
4074 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size())
4075 arc_c_max = zfs_arc_max;
4076 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max)
4077 arc_c_min = zfs_arc_min;
4081 arc_p = (arc_c >> 1);
4083 /* limit meta-data to 1/4 of the arc capacity */
4084 arc_meta_limit = arc_c_max / 4;
4086 /* Allow the tunable to override if it is reasonable */
4087 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
4088 arc_meta_limit = zfs_arc_meta_limit;
4090 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
4091 arc_c_min = arc_meta_limit / 2;
4093 if (zfs_arc_grow_retry > 0)
4094 arc_grow_retry = zfs_arc_grow_retry;
4096 if (zfs_arc_shrink_shift > 0)
4097 arc_shrink_shift = zfs_arc_shrink_shift;
4099 if (zfs_arc_p_min_shift > 0)
4100 arc_p_min_shift = zfs_arc_p_min_shift;
4102 /* if kmem_flags are set, lets try to use less memory */
4103 if (kmem_debugging())
4105 if (arc_c < arc_c_min)
4108 zfs_arc_min = arc_c_min;
4109 zfs_arc_max = arc_c_max;
4111 arc_anon = &ARC_anon;
4113 arc_mru_ghost = &ARC_mru_ghost;
4115 arc_mfu_ghost = &ARC_mfu_ghost;
4116 arc_l2c_only = &ARC_l2c_only;
4119 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
4120 mutex_init(&arc_anon->arcs_locks[i].arcs_lock,
4121 NULL, MUTEX_DEFAULT, NULL);
4122 mutex_init(&arc_mru->arcs_locks[i].arcs_lock,
4123 NULL, MUTEX_DEFAULT, NULL);
4124 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock,
4125 NULL, MUTEX_DEFAULT, NULL);
4126 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock,
4127 NULL, MUTEX_DEFAULT, NULL);
4128 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock,
4129 NULL, MUTEX_DEFAULT, NULL);
4130 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock,
4131 NULL, MUTEX_DEFAULT, NULL);
4133 list_create(&arc_mru->arcs_lists[i],
4134 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4135 list_create(&arc_mru_ghost->arcs_lists[i],
4136 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4137 list_create(&arc_mfu->arcs_lists[i],
4138 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4139 list_create(&arc_mfu_ghost->arcs_lists[i],
4140 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4141 list_create(&arc_mfu_ghost->arcs_lists[i],
4142 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4143 list_create(&arc_l2c_only->arcs_lists[i],
4144 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4149 arc_thread_exit = 0;
4150 arc_eviction_list = NULL;
4151 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
4152 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
4154 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
4155 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
4157 if (arc_ksp != NULL) {
4158 arc_ksp->ks_data = &arc_stats;
4159 kstat_install(arc_ksp);
4162 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
4163 TS_RUN, minclsyspri);
4166 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
4167 EVENTHANDLER_PRI_FIRST);
4174 * Calculate maximum amount of dirty data per pool.
4176 * If it has been set by /etc/system, take that.
4177 * Otherwise, use a percentage of physical memory defined by
4178 * zfs_dirty_data_max_percent (default 10%) with a cap at
4179 * zfs_dirty_data_max_max (default 4GB).
4181 if (zfs_dirty_data_max == 0) {
4182 zfs_dirty_data_max = ptob(physmem) *
4183 zfs_dirty_data_max_percent / 100;
4184 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
4185 zfs_dirty_data_max_max);
4189 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
4190 prefetch_tunable_set = 1;
4193 if (prefetch_tunable_set == 0) {
4194 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
4196 printf(" add \"vfs.zfs.prefetch_disable=0\" "
4197 "to /boot/loader.conf.\n");
4198 zfs_prefetch_disable = 1;
4201 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
4202 prefetch_tunable_set == 0) {
4203 printf("ZFS NOTICE: Prefetch is disabled by default if less "
4204 "than 4GB of RAM is present;\n"
4205 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
4206 "to /boot/loader.conf.\n");
4207 zfs_prefetch_disable = 1;
4210 /* Warn about ZFS memory and address space requirements. */
4211 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
4212 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
4213 "expect unstable behavior.\n");
4215 if (kmem_size() < 512 * (1 << 20)) {
4216 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
4217 "expect unstable behavior.\n");
4218 printf(" Consider tuning vm.kmem_size and "
4219 "vm.kmem_size_max\n");
4220 printf(" in /boot/loader.conf.\n");
4230 mutex_enter(&arc_reclaim_thr_lock);
4231 arc_thread_exit = 1;
4232 cv_signal(&arc_reclaim_thr_cv);
4233 while (arc_thread_exit != 0)
4234 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
4235 mutex_exit(&arc_reclaim_thr_lock);
4241 if (arc_ksp != NULL) {
4242 kstat_delete(arc_ksp);
4246 mutex_destroy(&arc_eviction_mtx);
4247 mutex_destroy(&arc_reclaim_thr_lock);
4248 cv_destroy(&arc_reclaim_thr_cv);
4250 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
4251 list_destroy(&arc_mru->arcs_lists[i]);
4252 list_destroy(&arc_mru_ghost->arcs_lists[i]);
4253 list_destroy(&arc_mfu->arcs_lists[i]);
4254 list_destroy(&arc_mfu_ghost->arcs_lists[i]);
4255 list_destroy(&arc_l2c_only->arcs_lists[i]);
4257 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock);
4258 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock);
4259 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock);
4260 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock);
4261 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock);
4262 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock);
4267 ASSERT(arc_loaned_bytes == 0);
4269 mutex_destroy(&arc_lowmem_lock);
4271 if (arc_event_lowmem != NULL)
4272 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
4279 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
4280 * It uses dedicated storage devices to hold cached data, which are populated
4281 * using large infrequent writes. The main role of this cache is to boost
4282 * the performance of random read workloads. The intended L2ARC devices
4283 * include short-stroked disks, solid state disks, and other media with
4284 * substantially faster read latency than disk.
4286 * +-----------------------+
4288 * +-----------------------+
4291 * l2arc_feed_thread() arc_read()
4295 * +---------------+ |
4297 * +---------------+ |
4302 * +-------+ +-------+
4304 * | cache | | cache |
4305 * +-------+ +-------+
4306 * +=========+ .-----.
4307 * : L2ARC : |-_____-|
4308 * : devices : | Disks |
4309 * +=========+ `-_____-'
4311 * Read requests are satisfied from the following sources, in order:
4314 * 2) vdev cache of L2ARC devices
4316 * 4) vdev cache of disks
4319 * Some L2ARC device types exhibit extremely slow write performance.
4320 * To accommodate for this there are some significant differences between
4321 * the L2ARC and traditional cache design:
4323 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4324 * the ARC behave as usual, freeing buffers and placing headers on ghost
4325 * lists. The ARC does not send buffers to the L2ARC during eviction as
4326 * this would add inflated write latencies for all ARC memory pressure.
4328 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4329 * It does this by periodically scanning buffers from the eviction-end of
4330 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4331 * not already there. It scans until a headroom of buffers is satisfied,
4332 * which itself is a buffer for ARC eviction. If a compressible buffer is
4333 * found during scanning and selected for writing to an L2ARC device, we
4334 * temporarily boost scanning headroom during the next scan cycle to make
4335 * sure we adapt to compression effects (which might significantly reduce
4336 * the data volume we write to L2ARC). The thread that does this is
4337 * l2arc_feed_thread(), illustrated below; example sizes are included to
4338 * provide a better sense of ratio than this diagram:
4341 * +---------------------+----------+
4342 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4343 * +---------------------+----------+ | o L2ARC eligible
4344 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4345 * +---------------------+----------+ |
4346 * 15.9 Gbytes ^ 32 Mbytes |
4348 * l2arc_feed_thread()
4350 * l2arc write hand <--[oooo]--'
4354 * +==============================+
4355 * L2ARC dev |####|#|###|###| |####| ... |
4356 * +==============================+
4359 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4360 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4361 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4362 * safe to say that this is an uncommon case, since buffers at the end of
4363 * the ARC lists have moved there due to inactivity.
4365 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4366 * then the L2ARC simply misses copying some buffers. This serves as a
4367 * pressure valve to prevent heavy read workloads from both stalling the ARC
4368 * with waits and clogging the L2ARC with writes. This also helps prevent
4369 * the potential for the L2ARC to churn if it attempts to cache content too
4370 * quickly, such as during backups of the entire pool.
4372 * 5. After system boot and before the ARC has filled main memory, there are
4373 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4374 * lists can remain mostly static. Instead of searching from tail of these
4375 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4376 * for eligible buffers, greatly increasing its chance of finding them.
4378 * The L2ARC device write speed is also boosted during this time so that
4379 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4380 * there are no L2ARC reads, and no fear of degrading read performance
4381 * through increased writes.
4383 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4384 * the vdev queue can aggregate them into larger and fewer writes. Each
4385 * device is written to in a rotor fashion, sweeping writes through
4386 * available space then repeating.
4388 * 7. The L2ARC does not store dirty content. It never needs to flush
4389 * write buffers back to disk based storage.
4391 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4392 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4394 * The performance of the L2ARC can be tweaked by a number of tunables, which
4395 * may be necessary for different workloads:
4397 * l2arc_write_max max write bytes per interval
4398 * l2arc_write_boost extra write bytes during device warmup
4399 * l2arc_noprefetch skip caching prefetched buffers
4400 * l2arc_headroom number of max device writes to precache
4401 * l2arc_headroom_boost when we find compressed buffers during ARC
4402 * scanning, we multiply headroom by this
4403 * percentage factor for the next scan cycle,
4404 * since more compressed buffers are likely to
4406 * l2arc_feed_secs seconds between L2ARC writing
4408 * Tunables may be removed or added as future performance improvements are
4409 * integrated, and also may become zpool properties.
4411 * There are three key functions that control how the L2ARC warms up:
4413 * l2arc_write_eligible() check if a buffer is eligible to cache
4414 * l2arc_write_size() calculate how much to write
4415 * l2arc_write_interval() calculate sleep delay between writes
4417 * These three functions determine what to write, how much, and how quickly
4422 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4425 * A buffer is *not* eligible for the L2ARC if it:
4426 * 1. belongs to a different spa.
4427 * 2. is already cached on the L2ARC.
4428 * 3. has an I/O in progress (it may be an incomplete read).
4429 * 4. is flagged not eligible (zfs property).
4431 if (ab->b_spa != spa_guid) {
4432 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
4435 if (ab->b_l2hdr != NULL) {
4436 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
4439 if (HDR_IO_IN_PROGRESS(ab)) {
4440 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
4443 if (!HDR_L2CACHE(ab)) {
4444 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
4452 l2arc_write_size(void)
4457 * Make sure our globals have meaningful values in case the user
4460 size = l2arc_write_max;
4462 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
4463 "be greater than zero, resetting it to the default (%d)",
4465 size = l2arc_write_max = L2ARC_WRITE_SIZE;
4468 if (arc_warm == B_FALSE)
4469 size += l2arc_write_boost;
4476 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4478 clock_t interval, next, now;
4481 * If the ARC lists are busy, increase our write rate; if the
4482 * lists are stale, idle back. This is achieved by checking
4483 * how much we previously wrote - if it was more than half of
4484 * what we wanted, schedule the next write much sooner.
4486 if (l2arc_feed_again && wrote > (wanted / 2))
4487 interval = (hz * l2arc_feed_min_ms) / 1000;
4489 interval = hz * l2arc_feed_secs;
4491 now = ddi_get_lbolt();
4492 next = MAX(now, MIN(now + interval, began + interval));
4498 l2arc_hdr_stat_add(void)
4500 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4501 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4505 l2arc_hdr_stat_remove(void)
4507 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4508 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4512 * Cycle through L2ARC devices. This is how L2ARC load balances.
4513 * If a device is returned, this also returns holding the spa config lock.
4515 static l2arc_dev_t *
4516 l2arc_dev_get_next(void)
4518 l2arc_dev_t *first, *next = NULL;
4521 * Lock out the removal of spas (spa_namespace_lock), then removal
4522 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4523 * both locks will be dropped and a spa config lock held instead.
4525 mutex_enter(&spa_namespace_lock);
4526 mutex_enter(&l2arc_dev_mtx);
4528 /* if there are no vdevs, there is nothing to do */
4529 if (l2arc_ndev == 0)
4533 next = l2arc_dev_last;
4535 /* loop around the list looking for a non-faulted vdev */
4537 next = list_head(l2arc_dev_list);
4539 next = list_next(l2arc_dev_list, next);
4541 next = list_head(l2arc_dev_list);
4544 /* if we have come back to the start, bail out */
4547 else if (next == first)
4550 } while (vdev_is_dead(next->l2ad_vdev));
4552 /* if we were unable to find any usable vdevs, return NULL */
4553 if (vdev_is_dead(next->l2ad_vdev))
4556 l2arc_dev_last = next;
4559 mutex_exit(&l2arc_dev_mtx);
4562 * Grab the config lock to prevent the 'next' device from being
4563 * removed while we are writing to it.
4566 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4567 mutex_exit(&spa_namespace_lock);
4573 * Free buffers that were tagged for destruction.
4576 l2arc_do_free_on_write()
4579 l2arc_data_free_t *df, *df_prev;
4581 mutex_enter(&l2arc_free_on_write_mtx);
4582 buflist = l2arc_free_on_write;
4584 for (df = list_tail(buflist); df; df = df_prev) {
4585 df_prev = list_prev(buflist, df);
4586 ASSERT(df->l2df_data != NULL);
4587 ASSERT(df->l2df_func != NULL);
4588 df->l2df_func(df->l2df_data, df->l2df_size);
4589 list_remove(buflist, df);
4590 kmem_free(df, sizeof (l2arc_data_free_t));
4593 mutex_exit(&l2arc_free_on_write_mtx);
4597 * A write to a cache device has completed. Update all headers to allow
4598 * reads from these buffers to begin.
4601 l2arc_write_done(zio_t *zio)
4603 l2arc_write_callback_t *cb;
4606 arc_buf_hdr_t *head, *ab, *ab_prev;
4607 l2arc_buf_hdr_t *abl2;
4608 kmutex_t *hash_lock;
4609 int64_t bytes_dropped = 0;
4611 cb = zio->io_private;
4613 dev = cb->l2wcb_dev;
4614 ASSERT(dev != NULL);
4615 head = cb->l2wcb_head;
4616 ASSERT(head != NULL);
4617 buflist = dev->l2ad_buflist;
4618 ASSERT(buflist != NULL);
4619 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4620 l2arc_write_callback_t *, cb);
4622 if (zio->io_error != 0)
4623 ARCSTAT_BUMP(arcstat_l2_writes_error);
4625 mutex_enter(&l2arc_buflist_mtx);
4628 * All writes completed, or an error was hit.
4630 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4631 ab_prev = list_prev(buflist, ab);
4635 * Release the temporary compressed buffer as soon as possible.
4637 if (abl2->b_compress != ZIO_COMPRESS_OFF)
4638 l2arc_release_cdata_buf(ab);
4640 hash_lock = HDR_LOCK(ab);
4641 if (!mutex_tryenter(hash_lock)) {
4643 * This buffer misses out. It may be in a stage
4644 * of eviction. Its ARC_L2_WRITING flag will be
4645 * left set, denying reads to this buffer.
4647 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4651 if (zio->io_error != 0) {
4653 * Error - drop L2ARC entry.
4655 list_remove(buflist, ab);
4656 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4657 bytes_dropped += abl2->b_asize;
4659 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr,
4661 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4662 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4666 * Allow ARC to begin reads to this L2ARC entry.
4668 ab->b_flags &= ~ARC_L2_WRITING;
4670 mutex_exit(hash_lock);
4673 atomic_inc_64(&l2arc_writes_done);
4674 list_remove(buflist, head);
4675 kmem_cache_free(hdr_cache, head);
4676 mutex_exit(&l2arc_buflist_mtx);
4678 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
4680 l2arc_do_free_on_write();
4682 kmem_free(cb, sizeof (l2arc_write_callback_t));
4686 * A read to a cache device completed. Validate buffer contents before
4687 * handing over to the regular ARC routines.
4690 l2arc_read_done(zio_t *zio)
4692 l2arc_read_callback_t *cb;
4695 kmutex_t *hash_lock;
4698 ASSERT(zio->io_vd != NULL);
4699 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4701 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4703 cb = zio->io_private;
4705 buf = cb->l2rcb_buf;
4706 ASSERT(buf != NULL);
4708 hash_lock = HDR_LOCK(buf->b_hdr);
4709 mutex_enter(hash_lock);
4711 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4714 * If the buffer was compressed, decompress it first.
4716 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
4717 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
4718 ASSERT(zio->io_data != NULL);
4721 * Check this survived the L2ARC journey.
4723 equal = arc_cksum_equal(buf);
4724 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4725 mutex_exit(hash_lock);
4726 zio->io_private = buf;
4727 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4728 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4731 mutex_exit(hash_lock);
4733 * Buffer didn't survive caching. Increment stats and
4734 * reissue to the original storage device.
4736 if (zio->io_error != 0) {
4737 ARCSTAT_BUMP(arcstat_l2_io_error);
4739 zio->io_error = SET_ERROR(EIO);
4742 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4745 * If there's no waiter, issue an async i/o to the primary
4746 * storage now. If there *is* a waiter, the caller must
4747 * issue the i/o in a context where it's OK to block.
4749 if (zio->io_waiter == NULL) {
4750 zio_t *pio = zio_unique_parent(zio);
4752 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4754 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4755 buf->b_data, zio->io_size, arc_read_done, buf,
4756 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4760 kmem_free(cb, sizeof (l2arc_read_callback_t));
4764 * This is the list priority from which the L2ARC will search for pages to
4765 * cache. This is used within loops (0..3) to cycle through lists in the
4766 * desired order. This order can have a significant effect on cache
4769 * Currently the metadata lists are hit first, MFU then MRU, followed by
4770 * the data lists. This function returns a locked list, and also returns
4774 l2arc_list_locked(int list_num, kmutex_t **lock)
4776 list_t *list = NULL;
4779 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS);
4781 if (list_num < ARC_BUFC_NUMMETADATALISTS) {
4783 list = &arc_mfu->arcs_lists[idx];
4784 *lock = ARCS_LOCK(arc_mfu, idx);
4785 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) {
4786 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4787 list = &arc_mru->arcs_lists[idx];
4788 *lock = ARCS_LOCK(arc_mru, idx);
4789 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 +
4790 ARC_BUFC_NUMDATALISTS)) {
4791 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4792 list = &arc_mfu->arcs_lists[idx];
4793 *lock = ARCS_LOCK(arc_mfu, idx);
4795 idx = list_num - ARC_BUFC_NUMLISTS;
4796 list = &arc_mru->arcs_lists[idx];
4797 *lock = ARCS_LOCK(arc_mru, idx);
4800 ASSERT(!(MUTEX_HELD(*lock)));
4806 * Evict buffers from the device write hand to the distance specified in
4807 * bytes. This distance may span populated buffers, it may span nothing.
4808 * This is clearing a region on the L2ARC device ready for writing.
4809 * If the 'all' boolean is set, every buffer is evicted.
4812 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4815 l2arc_buf_hdr_t *abl2;
4816 arc_buf_hdr_t *ab, *ab_prev;
4817 kmutex_t *hash_lock;
4819 int64_t bytes_evicted = 0;
4821 buflist = dev->l2ad_buflist;
4823 if (buflist == NULL)
4826 if (!all && dev->l2ad_first) {
4828 * This is the first sweep through the device. There is
4834 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4836 * When nearing the end of the device, evict to the end
4837 * before the device write hand jumps to the start.
4839 taddr = dev->l2ad_end;
4841 taddr = dev->l2ad_hand + distance;
4843 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4844 uint64_t, taddr, boolean_t, all);
4847 mutex_enter(&l2arc_buflist_mtx);
4848 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4849 ab_prev = list_prev(buflist, ab);
4851 hash_lock = HDR_LOCK(ab);
4852 if (!mutex_tryenter(hash_lock)) {
4854 * Missed the hash lock. Retry.
4856 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4857 mutex_exit(&l2arc_buflist_mtx);
4858 mutex_enter(hash_lock);
4859 mutex_exit(hash_lock);
4863 if (HDR_L2_WRITE_HEAD(ab)) {
4865 * We hit a write head node. Leave it for
4866 * l2arc_write_done().
4868 list_remove(buflist, ab);
4869 mutex_exit(hash_lock);
4873 if (!all && ab->b_l2hdr != NULL &&
4874 (ab->b_l2hdr->b_daddr > taddr ||
4875 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4877 * We've evicted to the target address,
4878 * or the end of the device.
4880 mutex_exit(hash_lock);
4884 if (HDR_FREE_IN_PROGRESS(ab)) {
4886 * Already on the path to destruction.
4888 mutex_exit(hash_lock);
4892 if (ab->b_state == arc_l2c_only) {
4893 ASSERT(!HDR_L2_READING(ab));
4895 * This doesn't exist in the ARC. Destroy.
4896 * arc_hdr_destroy() will call list_remove()
4897 * and decrement arcstat_l2_size.
4899 arc_change_state(arc_anon, ab, hash_lock);
4900 arc_hdr_destroy(ab);
4903 * Invalidate issued or about to be issued
4904 * reads, since we may be about to write
4905 * over this location.
4907 if (HDR_L2_READING(ab)) {
4908 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4909 ab->b_flags |= ARC_L2_EVICTED;
4913 * Tell ARC this no longer exists in L2ARC.
4915 if (ab->b_l2hdr != NULL) {
4917 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4918 bytes_evicted += abl2->b_asize;
4920 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4921 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4923 list_remove(buflist, ab);
4926 * This may have been leftover after a
4929 ab->b_flags &= ~ARC_L2_WRITING;
4931 mutex_exit(hash_lock);
4933 mutex_exit(&l2arc_buflist_mtx);
4935 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0);
4936 dev->l2ad_evict = taddr;
4940 * Find and write ARC buffers to the L2ARC device.
4942 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4943 * for reading until they have completed writing.
4944 * The headroom_boost is an in-out parameter used to maintain headroom boost
4945 * state between calls to this function.
4947 * Returns the number of bytes actually written (which may be smaller than
4948 * the delta by which the device hand has changed due to alignment).
4951 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
4952 boolean_t *headroom_boost)
4954 arc_buf_hdr_t *ab, *ab_prev, *head;
4956 uint64_t write_asize, write_psize, write_sz, headroom,
4959 kmutex_t *list_lock;
4961 l2arc_write_callback_t *cb;
4963 uint64_t guid = spa_load_guid(spa);
4964 const boolean_t do_headroom_boost = *headroom_boost;
4967 ASSERT(dev->l2ad_vdev != NULL);
4969 /* Lower the flag now, we might want to raise it again later. */
4970 *headroom_boost = B_FALSE;
4973 write_sz = write_asize = write_psize = 0;
4975 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4976 head->b_flags |= ARC_L2_WRITE_HEAD;
4978 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
4980 * We will want to try to compress buffers that are at least 2x the
4981 * device sector size.
4983 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
4986 * Copy buffers for L2ARC writing.
4988 mutex_enter(&l2arc_buflist_mtx);
4989 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) {
4990 uint64_t passed_sz = 0;
4992 list = l2arc_list_locked(try, &list_lock);
4993 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
4996 * L2ARC fast warmup.
4998 * Until the ARC is warm and starts to evict, read from the
4999 * head of the ARC lists rather than the tail.
5001 if (arc_warm == B_FALSE)
5002 ab = list_head(list);
5004 ab = list_tail(list);
5006 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
5008 headroom = target_sz * l2arc_headroom;
5009 if (do_headroom_boost)
5010 headroom = (headroom * l2arc_headroom_boost) / 100;
5012 for (; ab; ab = ab_prev) {
5013 l2arc_buf_hdr_t *l2hdr;
5014 kmutex_t *hash_lock;
5017 if (arc_warm == B_FALSE)
5018 ab_prev = list_next(list, ab);
5020 ab_prev = list_prev(list, ab);
5021 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size);
5023 hash_lock = HDR_LOCK(ab);
5024 if (!mutex_tryenter(hash_lock)) {
5025 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
5027 * Skip this buffer rather than waiting.
5032 passed_sz += ab->b_size;
5033 if (passed_sz > headroom) {
5037 mutex_exit(hash_lock);
5038 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
5042 if (!l2arc_write_eligible(guid, ab)) {
5043 mutex_exit(hash_lock);
5047 if ((write_sz + ab->b_size) > target_sz) {
5049 mutex_exit(hash_lock);
5050 ARCSTAT_BUMP(arcstat_l2_write_full);
5056 * Insert a dummy header on the buflist so
5057 * l2arc_write_done() can find where the
5058 * write buffers begin without searching.
5060 list_insert_head(dev->l2ad_buflist, head);
5063 sizeof (l2arc_write_callback_t), KM_SLEEP);
5064 cb->l2wcb_dev = dev;
5065 cb->l2wcb_head = head;
5066 pio = zio_root(spa, l2arc_write_done, cb,
5068 ARCSTAT_BUMP(arcstat_l2_write_pios);
5072 * Create and add a new L2ARC header.
5074 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
5076 ab->b_flags |= ARC_L2_WRITING;
5079 * Temporarily stash the data buffer in b_tmp_cdata.
5080 * The subsequent write step will pick it up from
5081 * there. This is because can't access ab->b_buf
5082 * without holding the hash_lock, which we in turn
5083 * can't access without holding the ARC list locks
5084 * (which we want to avoid during compression/writing).
5086 l2hdr->b_compress = ZIO_COMPRESS_OFF;
5087 l2hdr->b_asize = ab->b_size;
5088 l2hdr->b_tmp_cdata = ab->b_buf->b_data;
5090 buf_sz = ab->b_size;
5091 ab->b_l2hdr = l2hdr;
5093 list_insert_head(dev->l2ad_buflist, ab);
5096 * Compute and store the buffer cksum before
5097 * writing. On debug the cksum is verified first.
5099 arc_cksum_verify(ab->b_buf);
5100 arc_cksum_compute(ab->b_buf, B_TRUE);
5102 mutex_exit(hash_lock);
5107 mutex_exit(list_lock);
5113 /* No buffers selected for writing? */
5116 mutex_exit(&l2arc_buflist_mtx);
5117 kmem_cache_free(hdr_cache, head);
5122 * Now start writing the buffers. We're starting at the write head
5123 * and work backwards, retracing the course of the buffer selector
5126 for (ab = list_prev(dev->l2ad_buflist, head); ab;
5127 ab = list_prev(dev->l2ad_buflist, ab)) {
5128 l2arc_buf_hdr_t *l2hdr;
5132 * We shouldn't need to lock the buffer here, since we flagged
5133 * it as ARC_L2_WRITING in the previous step, but we must take
5134 * care to only access its L2 cache parameters. In particular,
5135 * ab->b_buf may be invalid by now due to ARC eviction.
5137 l2hdr = ab->b_l2hdr;
5138 l2hdr->b_daddr = dev->l2ad_hand;
5140 if ((ab->b_flags & ARC_L2COMPRESS) &&
5141 l2hdr->b_asize >= buf_compress_minsz) {
5142 if (l2arc_compress_buf(l2hdr)) {
5144 * If compression succeeded, enable headroom
5145 * boost on the next scan cycle.
5147 *headroom_boost = B_TRUE;
5152 * Pick up the buffer data we had previously stashed away
5153 * (and now potentially also compressed).
5155 buf_data = l2hdr->b_tmp_cdata;
5156 buf_sz = l2hdr->b_asize;
5158 /* Compression may have squashed the buffer to zero length. */
5162 wzio = zio_write_phys(pio, dev->l2ad_vdev,
5163 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
5164 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
5165 ZIO_FLAG_CANFAIL, B_FALSE);
5167 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
5169 (void) zio_nowait(wzio);
5171 write_asize += buf_sz;
5173 * Keep the clock hand suitably device-aligned.
5175 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
5176 write_psize += buf_p_sz;
5177 dev->l2ad_hand += buf_p_sz;
5181 mutex_exit(&l2arc_buflist_mtx);
5183 ASSERT3U(write_asize, <=, target_sz);
5184 ARCSTAT_BUMP(arcstat_l2_writes_sent);
5185 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
5186 ARCSTAT_INCR(arcstat_l2_size, write_sz);
5187 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
5188 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
5191 * Bump device hand to the device start if it is approaching the end.
5192 * l2arc_evict() will already have evicted ahead for this case.
5194 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
5195 dev->l2ad_hand = dev->l2ad_start;
5196 dev->l2ad_evict = dev->l2ad_start;
5197 dev->l2ad_first = B_FALSE;
5200 dev->l2ad_writing = B_TRUE;
5201 (void) zio_wait(pio);
5202 dev->l2ad_writing = B_FALSE;
5204 return (write_asize);
5208 * Compresses an L2ARC buffer.
5209 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its
5210 * size in l2hdr->b_asize. This routine tries to compress the data and
5211 * depending on the compression result there are three possible outcomes:
5212 * *) The buffer was incompressible. The original l2hdr contents were left
5213 * untouched and are ready for writing to an L2 device.
5214 * *) The buffer was all-zeros, so there is no need to write it to an L2
5215 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
5216 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
5217 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
5218 * data buffer which holds the compressed data to be written, and b_asize
5219 * tells us how much data there is. b_compress is set to the appropriate
5220 * compression algorithm. Once writing is done, invoke
5221 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
5223 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
5224 * buffer was incompressible).
5227 l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr)
5230 size_t csize, len, rounded;
5232 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
5233 ASSERT(l2hdr->b_tmp_cdata != NULL);
5235 len = l2hdr->b_asize;
5236 cdata = zio_data_buf_alloc(len);
5237 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata,
5238 cdata, l2hdr->b_asize, (size_t)(1ULL << l2hdr->b_dev->l2ad_vdev->vdev_ashift));
5240 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE);
5241 if (rounded > csize) {
5242 bzero((char *)cdata + csize, rounded - csize);
5247 /* zero block, indicate that there's nothing to write */
5248 zio_data_buf_free(cdata, len);
5249 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
5251 l2hdr->b_tmp_cdata = NULL;
5252 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
5254 } else if (csize > 0 && csize < len) {
5256 * Compression succeeded, we'll keep the cdata around for
5257 * writing and release it afterwards.
5259 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
5260 l2hdr->b_asize = csize;
5261 l2hdr->b_tmp_cdata = cdata;
5262 ARCSTAT_BUMP(arcstat_l2_compress_successes);
5266 * Compression failed, release the compressed buffer.
5267 * l2hdr will be left unmodified.
5269 zio_data_buf_free(cdata, len);
5270 ARCSTAT_BUMP(arcstat_l2_compress_failures);
5276 * Decompresses a zio read back from an l2arc device. On success, the
5277 * underlying zio's io_data buffer is overwritten by the uncompressed
5278 * version. On decompression error (corrupt compressed stream), the
5279 * zio->io_error value is set to signal an I/O error.
5281 * Please note that the compressed data stream is not checksummed, so
5282 * if the underlying device is experiencing data corruption, we may feed
5283 * corrupt data to the decompressor, so the decompressor needs to be
5284 * able to handle this situation (LZ4 does).
5287 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
5289 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
5291 if (zio->io_error != 0) {
5293 * An io error has occured, just restore the original io
5294 * size in preparation for a main pool read.
5296 zio->io_orig_size = zio->io_size = hdr->b_size;
5300 if (c == ZIO_COMPRESS_EMPTY) {
5302 * An empty buffer results in a null zio, which means we
5303 * need to fill its io_data after we're done restoring the
5304 * buffer's contents.
5306 ASSERT(hdr->b_buf != NULL);
5307 bzero(hdr->b_buf->b_data, hdr->b_size);
5308 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data;
5310 ASSERT(zio->io_data != NULL);
5312 * We copy the compressed data from the start of the arc buffer
5313 * (the zio_read will have pulled in only what we need, the
5314 * rest is garbage which we will overwrite at decompression)
5315 * and then decompress back to the ARC data buffer. This way we
5316 * can minimize copying by simply decompressing back over the
5317 * original compressed data (rather than decompressing to an
5318 * aux buffer and then copying back the uncompressed buffer,
5319 * which is likely to be much larger).
5324 csize = zio->io_size;
5325 cdata = zio_data_buf_alloc(csize);
5326 bcopy(zio->io_data, cdata, csize);
5327 if (zio_decompress_data(c, cdata, zio->io_data, csize,
5329 zio->io_error = EIO;
5330 zio_data_buf_free(cdata, csize);
5333 /* Restore the expected uncompressed IO size. */
5334 zio->io_orig_size = zio->io_size = hdr->b_size;
5338 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
5339 * This buffer serves as a temporary holder of compressed data while
5340 * the buffer entry is being written to an l2arc device. Once that is
5341 * done, we can dispose of it.
5344 l2arc_release_cdata_buf(arc_buf_hdr_t *ab)
5346 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr;
5348 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) {
5350 * If the data was compressed, then we've allocated a
5351 * temporary buffer for it, so now we need to release it.
5353 ASSERT(l2hdr->b_tmp_cdata != NULL);
5354 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size);
5356 l2hdr->b_tmp_cdata = NULL;
5360 * This thread feeds the L2ARC at regular intervals. This is the beating
5361 * heart of the L2ARC.
5364 l2arc_feed_thread(void *dummy __unused)
5369 uint64_t size, wrote;
5370 clock_t begin, next = ddi_get_lbolt();
5371 boolean_t headroom_boost = B_FALSE;
5373 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
5375 mutex_enter(&l2arc_feed_thr_lock);
5377 while (l2arc_thread_exit == 0) {
5378 CALLB_CPR_SAFE_BEGIN(&cpr);
5379 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
5380 next - ddi_get_lbolt());
5381 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
5382 next = ddi_get_lbolt() + hz;
5385 * Quick check for L2ARC devices.
5387 mutex_enter(&l2arc_dev_mtx);
5388 if (l2arc_ndev == 0) {
5389 mutex_exit(&l2arc_dev_mtx);
5392 mutex_exit(&l2arc_dev_mtx);
5393 begin = ddi_get_lbolt();
5396 * This selects the next l2arc device to write to, and in
5397 * doing so the next spa to feed from: dev->l2ad_spa. This
5398 * will return NULL if there are now no l2arc devices or if
5399 * they are all faulted.
5401 * If a device is returned, its spa's config lock is also
5402 * held to prevent device removal. l2arc_dev_get_next()
5403 * will grab and release l2arc_dev_mtx.
5405 if ((dev = l2arc_dev_get_next()) == NULL)
5408 spa = dev->l2ad_spa;
5409 ASSERT(spa != NULL);
5412 * If the pool is read-only then force the feed thread to
5413 * sleep a little longer.
5415 if (!spa_writeable(spa)) {
5416 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
5417 spa_config_exit(spa, SCL_L2ARC, dev);
5422 * Avoid contributing to memory pressure.
5424 if (arc_reclaim_needed()) {
5425 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
5426 spa_config_exit(spa, SCL_L2ARC, dev);
5430 ARCSTAT_BUMP(arcstat_l2_feeds);
5432 size = l2arc_write_size();
5435 * Evict L2ARC buffers that will be overwritten.
5437 l2arc_evict(dev, size, B_FALSE);
5440 * Write ARC buffers.
5442 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
5445 * Calculate interval between writes.
5447 next = l2arc_write_interval(begin, size, wrote);
5448 spa_config_exit(spa, SCL_L2ARC, dev);
5451 l2arc_thread_exit = 0;
5452 cv_broadcast(&l2arc_feed_thr_cv);
5453 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
5458 l2arc_vdev_present(vdev_t *vd)
5462 mutex_enter(&l2arc_dev_mtx);
5463 for (dev = list_head(l2arc_dev_list); dev != NULL;
5464 dev = list_next(l2arc_dev_list, dev)) {
5465 if (dev->l2ad_vdev == vd)
5468 mutex_exit(&l2arc_dev_mtx);
5470 return (dev != NULL);
5474 * Add a vdev for use by the L2ARC. By this point the spa has already
5475 * validated the vdev and opened it.
5478 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5480 l2arc_dev_t *adddev;
5482 ASSERT(!l2arc_vdev_present(vd));
5484 vdev_ashift_optimize(vd);
5487 * Create a new l2arc device entry.
5489 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5490 adddev->l2ad_spa = spa;
5491 adddev->l2ad_vdev = vd;
5492 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5493 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5494 adddev->l2ad_hand = adddev->l2ad_start;
5495 adddev->l2ad_evict = adddev->l2ad_start;
5496 adddev->l2ad_first = B_TRUE;
5497 adddev->l2ad_writing = B_FALSE;
5500 * This is a list of all ARC buffers that are still valid on the
5503 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5504 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5505 offsetof(arc_buf_hdr_t, b_l2node));
5507 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5510 * Add device to global list
5512 mutex_enter(&l2arc_dev_mtx);
5513 list_insert_head(l2arc_dev_list, adddev);
5514 atomic_inc_64(&l2arc_ndev);
5515 mutex_exit(&l2arc_dev_mtx);
5519 * Remove a vdev from the L2ARC.
5522 l2arc_remove_vdev(vdev_t *vd)
5524 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5527 * Find the device by vdev
5529 mutex_enter(&l2arc_dev_mtx);
5530 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5531 nextdev = list_next(l2arc_dev_list, dev);
5532 if (vd == dev->l2ad_vdev) {
5537 ASSERT(remdev != NULL);
5540 * Remove device from global list
5542 list_remove(l2arc_dev_list, remdev);
5543 l2arc_dev_last = NULL; /* may have been invalidated */
5544 atomic_dec_64(&l2arc_ndev);
5545 mutex_exit(&l2arc_dev_mtx);
5548 * Clear all buflists and ARC references. L2ARC device flush.
5550 l2arc_evict(remdev, 0, B_TRUE);
5551 list_destroy(remdev->l2ad_buflist);
5552 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5553 kmem_free(remdev, sizeof (l2arc_dev_t));
5559 l2arc_thread_exit = 0;
5561 l2arc_writes_sent = 0;
5562 l2arc_writes_done = 0;
5564 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5565 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5566 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5567 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5568 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5570 l2arc_dev_list = &L2ARC_dev_list;
5571 l2arc_free_on_write = &L2ARC_free_on_write;
5572 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5573 offsetof(l2arc_dev_t, l2ad_node));
5574 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5575 offsetof(l2arc_data_free_t, l2df_list_node));
5582 * This is called from dmu_fini(), which is called from spa_fini();
5583 * Because of this, we can assume that all l2arc devices have
5584 * already been removed when the pools themselves were removed.
5587 l2arc_do_free_on_write();
5589 mutex_destroy(&l2arc_feed_thr_lock);
5590 cv_destroy(&l2arc_feed_thr_cv);
5591 mutex_destroy(&l2arc_dev_mtx);
5592 mutex_destroy(&l2arc_buflist_mtx);
5593 mutex_destroy(&l2arc_free_on_write_mtx);
5595 list_destroy(l2arc_dev_list);
5596 list_destroy(l2arc_free_on_write);
5602 if (!(spa_mode_global & FWRITE))
5605 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5606 TS_RUN, minclsyspri);
5612 if (!(spa_mode_global & FWRITE))
5615 mutex_enter(&l2arc_feed_thr_lock);
5616 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5617 l2arc_thread_exit = 1;
5618 while (l2arc_thread_exit != 0)
5619 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5620 mutex_exit(&l2arc_feed_thr_lock);