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 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2013 by Delphix. All rights reserved.
28 * DVA-based Adjustable Replacement Cache
30 * While much of the theory of operation used here is
31 * based on the self-tuning, low overhead replacement cache
32 * presented by Megiddo and Modha at FAST 2003, there are some
33 * significant differences:
35 * 1. The Megiddo and Modha model assumes any page is evictable.
36 * Pages in its cache cannot be "locked" into memory. This makes
37 * the eviction algorithm simple: evict the last page in the list.
38 * This also make the performance characteristics easy to reason
39 * about. Our cache is not so simple. At any given moment, some
40 * subset of the blocks in the cache are un-evictable because we
41 * have handed out a reference to them. Blocks are only evictable
42 * when there are no external references active. This makes
43 * eviction far more problematic: we choose to evict the evictable
44 * blocks that are the "lowest" in the list.
46 * There are times when it is not possible to evict the requested
47 * space. In these circumstances we are unable to adjust the cache
48 * size. To prevent the cache growing unbounded at these times we
49 * implement a "cache throttle" that slows the flow of new data
50 * into the cache until we can make space available.
52 * 2. The Megiddo and Modha model assumes a fixed cache size.
53 * Pages are evicted when the cache is full and there is a cache
54 * miss. Our model has a variable sized cache. It grows with
55 * high use, but also tries to react to memory pressure from the
56 * operating system: decreasing its size when system memory is
59 * 3. The Megiddo and Modha model assumes a fixed page size. All
60 * elements of the cache are therefor exactly the same size. So
61 * when adjusting the cache size following a cache miss, its simply
62 * a matter of choosing a single page to evict. In our model, we
63 * have variable sized cache blocks (rangeing from 512 bytes to
64 * 128K bytes). We therefor choose a set of blocks to evict to make
65 * space for a cache miss that approximates as closely as possible
66 * the space used by the new block.
68 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
69 * by N. Megiddo & D. Modha, FAST 2003
75 * A new reference to a cache buffer can be obtained in two
76 * ways: 1) via a hash table lookup using the DVA as a key,
77 * or 2) via one of the ARC lists. The arc_read() interface
78 * uses method 1, while the internal arc algorithms for
79 * adjusting the cache use method 2. We therefor provide two
80 * types of locks: 1) the hash table lock array, and 2) the
83 * Buffers do not have their own mutexs, rather they rely on the
84 * hash table mutexs for the bulk of their protection (i.e. most
85 * fields in the arc_buf_hdr_t are protected by these mutexs).
87 * buf_hash_find() returns the appropriate mutex (held) when it
88 * locates the requested buffer in the hash table. It returns
89 * NULL for the mutex if the buffer was not in the table.
91 * buf_hash_remove() expects the appropriate hash mutex to be
92 * already held before it is invoked.
94 * Each arc state also has a mutex which is used to protect the
95 * buffer list associated with the state. When attempting to
96 * obtain a hash table lock while holding an arc list lock you
97 * must use: mutex_tryenter() to avoid deadlock. Also note that
98 * the active state mutex must be held before the ghost state mutex.
100 * Arc buffers may have an associated eviction callback function.
101 * This function will be invoked prior to removing the buffer (e.g.
102 * in arc_do_user_evicts()). Note however that the data associated
103 * with the buffer may be evicted prior to the callback. The callback
104 * must be made with *no locks held* (to prevent deadlock). Additionally,
105 * the users of callbacks must ensure that their private data is
106 * protected from simultaneous callbacks from arc_buf_evict()
107 * and arc_do_user_evicts().
109 * Note that the majority of the performance stats are manipulated
110 * with atomic operations.
112 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
114 * - L2ARC buflist creation
115 * - L2ARC buflist eviction
116 * - L2ARC write completion, which walks L2ARC buflists
117 * - ARC header destruction, as it removes from L2ARC buflists
118 * - ARC header release, as it removes from L2ARC buflists
123 #include <sys/zfs_context.h>
125 #include <sys/refcount.h>
126 #include <sys/vdev.h>
127 #include <sys/vdev_impl.h>
129 #include <sys/dnlc.h>
131 #include <sys/callb.h>
132 #include <sys/kstat.h>
133 #include <sys/trim_map.h>
134 #include <zfs_fletcher.h>
137 #include <vm/vm_pageout.h>
141 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
142 boolean_t arc_watch = B_FALSE;
147 static kmutex_t arc_reclaim_thr_lock;
148 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
149 static uint8_t arc_thread_exit;
151 extern int zfs_write_limit_shift;
152 extern uint64_t zfs_write_limit_max;
153 extern kmutex_t zfs_write_limit_lock;
155 #define ARC_REDUCE_DNLC_PERCENT 3
156 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
158 typedef enum arc_reclaim_strategy {
159 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
160 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
161 } arc_reclaim_strategy_t;
163 /* number of seconds before growing cache again */
164 static int arc_grow_retry = 60;
166 /* shift of arc_c for calculating both min and max arc_p */
167 static int arc_p_min_shift = 4;
169 /* log2(fraction of arc to reclaim) */
170 static int arc_shrink_shift = 5;
173 * minimum lifespan of a prefetch block in clock ticks
174 * (initialized in arc_init())
176 static int arc_min_prefetch_lifespan;
179 extern int zfs_prefetch_disable;
182 * The arc has filled available memory and has now warmed up.
184 static boolean_t arc_warm;
187 * These tunables are for performance analysis.
189 uint64_t zfs_arc_max;
190 uint64_t zfs_arc_min;
191 uint64_t zfs_arc_meta_limit = 0;
192 int zfs_arc_grow_retry = 0;
193 int zfs_arc_shrink_shift = 0;
194 int zfs_arc_p_min_shift = 0;
195 int zfs_disable_dup_eviction = 0;
197 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
198 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
199 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
200 SYSCTL_DECL(_vfs_zfs);
201 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
203 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
207 * Note that buffers can be in one of 6 states:
208 * ARC_anon - anonymous (discussed below)
209 * ARC_mru - recently used, currently cached
210 * ARC_mru_ghost - recentely used, no longer in cache
211 * ARC_mfu - frequently used, currently cached
212 * ARC_mfu_ghost - frequently used, no longer in cache
213 * ARC_l2c_only - exists in L2ARC but not other states
214 * When there are no active references to the buffer, they are
215 * are linked onto a list in one of these arc states. These are
216 * the only buffers that can be evicted or deleted. Within each
217 * state there are multiple lists, one for meta-data and one for
218 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
219 * etc.) is tracked separately so that it can be managed more
220 * explicitly: favored over data, limited explicitly.
222 * Anonymous buffers are buffers that are not associated with
223 * a DVA. These are buffers that hold dirty block copies
224 * before they are written to stable storage. By definition,
225 * they are "ref'd" and are considered part of arc_mru
226 * that cannot be freed. Generally, they will aquire a DVA
227 * as they are written and migrate onto the arc_mru list.
229 * The ARC_l2c_only state is for buffers that are in the second
230 * level ARC but no longer in any of the ARC_m* lists. The second
231 * level ARC itself may also contain buffers that are in any of
232 * the ARC_m* states - meaning that a buffer can exist in two
233 * places. The reason for the ARC_l2c_only state is to keep the
234 * buffer header in the hash table, so that reads that hit the
235 * second level ARC benefit from these fast lookups.
238 #define ARCS_LOCK_PAD CACHE_LINE_SIZE
242 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))];
247 * must be power of two for mask use to work
250 #define ARC_BUFC_NUMDATALISTS 16
251 #define ARC_BUFC_NUMMETADATALISTS 16
252 #define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS)
254 typedef struct arc_state {
255 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
256 uint64_t arcs_size; /* total amount of data in this state */
257 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */
258 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE);
261 #define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock))
264 static arc_state_t ARC_anon;
265 static arc_state_t ARC_mru;
266 static arc_state_t ARC_mru_ghost;
267 static arc_state_t ARC_mfu;
268 static arc_state_t ARC_mfu_ghost;
269 static arc_state_t ARC_l2c_only;
271 typedef struct arc_stats {
272 kstat_named_t arcstat_hits;
273 kstat_named_t arcstat_misses;
274 kstat_named_t arcstat_demand_data_hits;
275 kstat_named_t arcstat_demand_data_misses;
276 kstat_named_t arcstat_demand_metadata_hits;
277 kstat_named_t arcstat_demand_metadata_misses;
278 kstat_named_t arcstat_prefetch_data_hits;
279 kstat_named_t arcstat_prefetch_data_misses;
280 kstat_named_t arcstat_prefetch_metadata_hits;
281 kstat_named_t arcstat_prefetch_metadata_misses;
282 kstat_named_t arcstat_mru_hits;
283 kstat_named_t arcstat_mru_ghost_hits;
284 kstat_named_t arcstat_mfu_hits;
285 kstat_named_t arcstat_mfu_ghost_hits;
286 kstat_named_t arcstat_allocated;
287 kstat_named_t arcstat_deleted;
288 kstat_named_t arcstat_stolen;
289 kstat_named_t arcstat_recycle_miss;
291 * Number of buffers that could not be evicted because the hash lock
292 * was held by another thread. The lock may not necessarily be held
293 * by something using the same buffer, since hash locks are shared
294 * by multiple buffers.
296 kstat_named_t arcstat_mutex_miss;
298 * Number of buffers skipped because they have I/O in progress, are
299 * indrect prefetch buffers that have not lived long enough, or are
300 * not from the spa we're trying to evict from.
302 kstat_named_t arcstat_evict_skip;
303 kstat_named_t arcstat_evict_l2_cached;
304 kstat_named_t arcstat_evict_l2_eligible;
305 kstat_named_t arcstat_evict_l2_ineligible;
306 kstat_named_t arcstat_hash_elements;
307 kstat_named_t arcstat_hash_elements_max;
308 kstat_named_t arcstat_hash_collisions;
309 kstat_named_t arcstat_hash_chains;
310 kstat_named_t arcstat_hash_chain_max;
311 kstat_named_t arcstat_p;
312 kstat_named_t arcstat_c;
313 kstat_named_t arcstat_c_min;
314 kstat_named_t arcstat_c_max;
315 kstat_named_t arcstat_size;
316 kstat_named_t arcstat_hdr_size;
317 kstat_named_t arcstat_data_size;
318 kstat_named_t arcstat_other_size;
319 kstat_named_t arcstat_l2_hits;
320 kstat_named_t arcstat_l2_misses;
321 kstat_named_t arcstat_l2_feeds;
322 kstat_named_t arcstat_l2_rw_clash;
323 kstat_named_t arcstat_l2_read_bytes;
324 kstat_named_t arcstat_l2_write_bytes;
325 kstat_named_t arcstat_l2_writes_sent;
326 kstat_named_t arcstat_l2_writes_done;
327 kstat_named_t arcstat_l2_writes_error;
328 kstat_named_t arcstat_l2_writes_hdr_miss;
329 kstat_named_t arcstat_l2_evict_lock_retry;
330 kstat_named_t arcstat_l2_evict_reading;
331 kstat_named_t arcstat_l2_free_on_write;
332 kstat_named_t arcstat_l2_abort_lowmem;
333 kstat_named_t arcstat_l2_cksum_bad;
334 kstat_named_t arcstat_l2_io_error;
335 kstat_named_t arcstat_l2_size;
336 kstat_named_t arcstat_l2_hdr_size;
337 kstat_named_t arcstat_l2_write_trylock_fail;
338 kstat_named_t arcstat_l2_write_passed_headroom;
339 kstat_named_t arcstat_l2_write_spa_mismatch;
340 kstat_named_t arcstat_l2_write_in_l2;
341 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
342 kstat_named_t arcstat_l2_write_not_cacheable;
343 kstat_named_t arcstat_l2_write_full;
344 kstat_named_t arcstat_l2_write_buffer_iter;
345 kstat_named_t arcstat_l2_write_pios;
346 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
347 kstat_named_t arcstat_l2_write_buffer_list_iter;
348 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
349 kstat_named_t arcstat_memory_throttle_count;
350 kstat_named_t arcstat_duplicate_buffers;
351 kstat_named_t arcstat_duplicate_buffers_size;
352 kstat_named_t arcstat_duplicate_reads;
355 static arc_stats_t arc_stats = {
356 { "hits", KSTAT_DATA_UINT64 },
357 { "misses", KSTAT_DATA_UINT64 },
358 { "demand_data_hits", KSTAT_DATA_UINT64 },
359 { "demand_data_misses", KSTAT_DATA_UINT64 },
360 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
361 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
362 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
363 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
364 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
365 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
366 { "mru_hits", KSTAT_DATA_UINT64 },
367 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
368 { "mfu_hits", KSTAT_DATA_UINT64 },
369 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
370 { "allocated", KSTAT_DATA_UINT64 },
371 { "deleted", KSTAT_DATA_UINT64 },
372 { "stolen", KSTAT_DATA_UINT64 },
373 { "recycle_miss", KSTAT_DATA_UINT64 },
374 { "mutex_miss", KSTAT_DATA_UINT64 },
375 { "evict_skip", KSTAT_DATA_UINT64 },
376 { "evict_l2_cached", KSTAT_DATA_UINT64 },
377 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
378 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
379 { "hash_elements", KSTAT_DATA_UINT64 },
380 { "hash_elements_max", KSTAT_DATA_UINT64 },
381 { "hash_collisions", KSTAT_DATA_UINT64 },
382 { "hash_chains", KSTAT_DATA_UINT64 },
383 { "hash_chain_max", KSTAT_DATA_UINT64 },
384 { "p", KSTAT_DATA_UINT64 },
385 { "c", KSTAT_DATA_UINT64 },
386 { "c_min", KSTAT_DATA_UINT64 },
387 { "c_max", KSTAT_DATA_UINT64 },
388 { "size", KSTAT_DATA_UINT64 },
389 { "hdr_size", KSTAT_DATA_UINT64 },
390 { "data_size", KSTAT_DATA_UINT64 },
391 { "other_size", KSTAT_DATA_UINT64 },
392 { "l2_hits", KSTAT_DATA_UINT64 },
393 { "l2_misses", KSTAT_DATA_UINT64 },
394 { "l2_feeds", KSTAT_DATA_UINT64 },
395 { "l2_rw_clash", KSTAT_DATA_UINT64 },
396 { "l2_read_bytes", KSTAT_DATA_UINT64 },
397 { "l2_write_bytes", KSTAT_DATA_UINT64 },
398 { "l2_writes_sent", KSTAT_DATA_UINT64 },
399 { "l2_writes_done", KSTAT_DATA_UINT64 },
400 { "l2_writes_error", KSTAT_DATA_UINT64 },
401 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
402 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
403 { "l2_evict_reading", KSTAT_DATA_UINT64 },
404 { "l2_free_on_write", KSTAT_DATA_UINT64 },
405 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
406 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
407 { "l2_io_error", KSTAT_DATA_UINT64 },
408 { "l2_size", KSTAT_DATA_UINT64 },
409 { "l2_hdr_size", KSTAT_DATA_UINT64 },
410 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
411 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
412 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
413 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
414 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
415 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
416 { "l2_write_full", KSTAT_DATA_UINT64 },
417 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
418 { "l2_write_pios", KSTAT_DATA_UINT64 },
419 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
420 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
421 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
422 { "memory_throttle_count", KSTAT_DATA_UINT64 },
423 { "duplicate_buffers", KSTAT_DATA_UINT64 },
424 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
425 { "duplicate_reads", KSTAT_DATA_UINT64 }
428 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
430 #define ARCSTAT_INCR(stat, val) \
431 atomic_add_64(&arc_stats.stat.value.ui64, (val));
433 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
434 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
436 #define ARCSTAT_MAX(stat, val) { \
438 while ((val) > (m = arc_stats.stat.value.ui64) && \
439 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
443 #define ARCSTAT_MAXSTAT(stat) \
444 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
447 * We define a macro to allow ARC hits/misses to be easily broken down by
448 * two separate conditions, giving a total of four different subtypes for
449 * each of hits and misses (so eight statistics total).
451 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
454 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
456 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
460 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
462 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
467 static arc_state_t *arc_anon;
468 static arc_state_t *arc_mru;
469 static arc_state_t *arc_mru_ghost;
470 static arc_state_t *arc_mfu;
471 static arc_state_t *arc_mfu_ghost;
472 static arc_state_t *arc_l2c_only;
475 * There are several ARC variables that are critical to export as kstats --
476 * but we don't want to have to grovel around in the kstat whenever we wish to
477 * manipulate them. For these variables, we therefore define them to be in
478 * terms of the statistic variable. This assures that we are not introducing
479 * the possibility of inconsistency by having shadow copies of the variables,
480 * while still allowing the code to be readable.
482 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
483 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
484 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
485 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
486 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
488 static int arc_no_grow; /* Don't try to grow cache size */
489 static uint64_t arc_tempreserve;
490 static uint64_t arc_loaned_bytes;
491 static uint64_t arc_meta_used;
492 static uint64_t arc_meta_limit;
493 static uint64_t arc_meta_max = 0;
494 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RDTUN,
495 &arc_meta_used, 0, "ARC metadata used");
496 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RDTUN,
497 &arc_meta_limit, 0, "ARC metadata limit");
499 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
501 typedef struct arc_callback arc_callback_t;
503 struct arc_callback {
505 arc_done_func_t *acb_done;
507 zio_t *acb_zio_dummy;
508 arc_callback_t *acb_next;
511 typedef struct arc_write_callback arc_write_callback_t;
513 struct arc_write_callback {
515 arc_done_func_t *awcb_ready;
516 arc_done_func_t *awcb_done;
521 /* protected by hash lock */
526 kmutex_t b_freeze_lock;
527 zio_cksum_t *b_freeze_cksum;
530 arc_buf_hdr_t *b_hash_next;
535 arc_callback_t *b_acb;
539 arc_buf_contents_t b_type;
543 /* protected by arc state mutex */
544 arc_state_t *b_state;
545 list_node_t b_arc_node;
547 /* updated atomically */
548 clock_t b_arc_access;
550 /* self protecting */
553 l2arc_buf_hdr_t *b_l2hdr;
554 list_node_t b_l2node;
557 static arc_buf_t *arc_eviction_list;
558 static kmutex_t arc_eviction_mtx;
559 static arc_buf_hdr_t arc_eviction_hdr;
560 static void arc_get_data_buf(arc_buf_t *buf);
561 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
562 static int arc_evict_needed(arc_buf_contents_t type);
563 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
565 static void arc_buf_watch(arc_buf_t *buf);
568 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
570 #define GHOST_STATE(state) \
571 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
572 (state) == arc_l2c_only)
575 * Private ARC flags. These flags are private ARC only flags that will show up
576 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
577 * be passed in as arc_flags in things like arc_read. However, these flags
578 * should never be passed and should only be set by ARC code. When adding new
579 * public flags, make sure not to smash the private ones.
582 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
583 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
584 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
585 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
586 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
587 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
588 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
589 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
590 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
591 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
593 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
594 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
595 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
596 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
597 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
598 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
599 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
600 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
601 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
602 (hdr)->b_l2hdr != NULL)
603 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
604 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
605 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
611 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
612 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
615 * Hash table routines
618 #define HT_LOCK_PAD CACHE_LINE_SIZE
623 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
627 #define BUF_LOCKS 256
628 typedef struct buf_hash_table {
630 arc_buf_hdr_t **ht_table;
631 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
634 static buf_hash_table_t buf_hash_table;
636 #define BUF_HASH_INDEX(spa, dva, birth) \
637 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
638 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
639 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
640 #define HDR_LOCK(hdr) \
641 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
643 uint64_t zfs_crc64_table[256];
649 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
650 #define L2ARC_HEADROOM 2 /* num of writes */
651 #define L2ARC_FEED_SECS 1 /* caching interval secs */
652 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
654 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
655 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
658 * L2ARC Performance Tunables
660 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
661 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
662 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
663 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
664 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
665 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
666 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
667 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
669 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
670 &l2arc_write_max, 0, "max write size");
671 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
672 &l2arc_write_boost, 0, "extra write during warmup");
673 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
674 &l2arc_headroom, 0, "number of dev writes");
675 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
676 &l2arc_feed_secs, 0, "interval seconds");
677 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
678 &l2arc_feed_min_ms, 0, "min interval milliseconds");
680 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
681 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
682 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
683 &l2arc_feed_again, 0, "turbo warmup");
684 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
685 &l2arc_norw, 0, "no reads during writes");
687 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
688 &ARC_anon.arcs_size, 0, "size of anonymous state");
689 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
690 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
691 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
692 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
694 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
695 &ARC_mru.arcs_size, 0, "size of mru state");
696 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
697 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
698 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
699 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
701 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
702 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state");
703 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
704 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
705 "size of metadata in mru ghost state");
706 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
707 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
708 "size of data in mru ghost state");
710 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
711 &ARC_mfu.arcs_size, 0, "size of mfu state");
712 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
713 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
714 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
715 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
717 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
718 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state");
719 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
720 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
721 "size of metadata in mfu ghost state");
722 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
723 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
724 "size of data in mfu ghost state");
726 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
727 &ARC_l2c_only.arcs_size, 0, "size of mru state");
732 typedef struct l2arc_dev {
733 vdev_t *l2ad_vdev; /* vdev */
734 spa_t *l2ad_spa; /* spa */
735 uint64_t l2ad_hand; /* next write location */
736 uint64_t l2ad_write; /* desired write size, bytes */
737 uint64_t l2ad_boost; /* warmup write boost, bytes */
738 uint64_t l2ad_start; /* first addr on device */
739 uint64_t l2ad_end; /* last addr on device */
740 uint64_t l2ad_evict; /* last addr eviction reached */
741 boolean_t l2ad_first; /* first sweep through */
742 boolean_t l2ad_writing; /* currently writing */
743 list_t *l2ad_buflist; /* buffer list */
744 list_node_t l2ad_node; /* device list node */
747 static list_t L2ARC_dev_list; /* device list */
748 static list_t *l2arc_dev_list; /* device list pointer */
749 static kmutex_t l2arc_dev_mtx; /* device list mutex */
750 static l2arc_dev_t *l2arc_dev_last; /* last device used */
751 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
752 static list_t L2ARC_free_on_write; /* free after write buf list */
753 static list_t *l2arc_free_on_write; /* free after write list ptr */
754 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
755 static uint64_t l2arc_ndev; /* number of devices */
757 typedef struct l2arc_read_callback {
758 arc_buf_t *l2rcb_buf; /* read buffer */
759 spa_t *l2rcb_spa; /* spa */
760 blkptr_t l2rcb_bp; /* original blkptr */
761 zbookmark_t l2rcb_zb; /* original bookmark */
762 int l2rcb_flags; /* original flags */
763 } l2arc_read_callback_t;
765 typedef struct l2arc_write_callback {
766 l2arc_dev_t *l2wcb_dev; /* device info */
767 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
768 } l2arc_write_callback_t;
770 struct l2arc_buf_hdr {
771 /* protected by arc_buf_hdr mutex */
772 l2arc_dev_t *b_dev; /* L2ARC device */
773 uint64_t b_daddr; /* disk address, offset byte */
776 typedef struct l2arc_data_free {
777 /* protected by l2arc_free_on_write_mtx */
780 void (*l2df_func)(void *, size_t);
781 list_node_t l2df_list_node;
784 static kmutex_t l2arc_feed_thr_lock;
785 static kcondvar_t l2arc_feed_thr_cv;
786 static uint8_t l2arc_thread_exit;
788 static void l2arc_read_done(zio_t *zio);
789 static void l2arc_hdr_stat_add(void);
790 static void l2arc_hdr_stat_remove(void);
793 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
795 uint8_t *vdva = (uint8_t *)dva;
796 uint64_t crc = -1ULL;
799 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
801 for (i = 0; i < sizeof (dva_t); i++)
802 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
804 crc ^= (spa>>8) ^ birth;
809 #define BUF_EMPTY(buf) \
810 ((buf)->b_dva.dva_word[0] == 0 && \
811 (buf)->b_dva.dva_word[1] == 0 && \
814 #define BUF_EQUAL(spa, dva, birth, buf) \
815 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
816 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
817 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
820 buf_discard_identity(arc_buf_hdr_t *hdr)
822 hdr->b_dva.dva_word[0] = 0;
823 hdr->b_dva.dva_word[1] = 0;
828 static arc_buf_hdr_t *
829 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
831 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
832 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
835 mutex_enter(hash_lock);
836 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
837 buf = buf->b_hash_next) {
838 if (BUF_EQUAL(spa, dva, birth, buf)) {
843 mutex_exit(hash_lock);
849 * Insert an entry into the hash table. If there is already an element
850 * equal to elem in the hash table, then the already existing element
851 * will be returned and the new element will not be inserted.
852 * Otherwise returns NULL.
854 static arc_buf_hdr_t *
855 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
857 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
858 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
862 ASSERT(!HDR_IN_HASH_TABLE(buf));
864 mutex_enter(hash_lock);
865 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
866 fbuf = fbuf->b_hash_next, i++) {
867 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
871 buf->b_hash_next = buf_hash_table.ht_table[idx];
872 buf_hash_table.ht_table[idx] = buf;
873 buf->b_flags |= ARC_IN_HASH_TABLE;
875 /* collect some hash table performance data */
877 ARCSTAT_BUMP(arcstat_hash_collisions);
879 ARCSTAT_BUMP(arcstat_hash_chains);
881 ARCSTAT_MAX(arcstat_hash_chain_max, i);
884 ARCSTAT_BUMP(arcstat_hash_elements);
885 ARCSTAT_MAXSTAT(arcstat_hash_elements);
891 buf_hash_remove(arc_buf_hdr_t *buf)
893 arc_buf_hdr_t *fbuf, **bufp;
894 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
896 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
897 ASSERT(HDR_IN_HASH_TABLE(buf));
899 bufp = &buf_hash_table.ht_table[idx];
900 while ((fbuf = *bufp) != buf) {
901 ASSERT(fbuf != NULL);
902 bufp = &fbuf->b_hash_next;
904 *bufp = buf->b_hash_next;
905 buf->b_hash_next = NULL;
906 buf->b_flags &= ~ARC_IN_HASH_TABLE;
908 /* collect some hash table performance data */
909 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
911 if (buf_hash_table.ht_table[idx] &&
912 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
913 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
917 * Global data structures and functions for the buf kmem cache.
919 static kmem_cache_t *hdr_cache;
920 static kmem_cache_t *buf_cache;
927 kmem_free(buf_hash_table.ht_table,
928 (buf_hash_table.ht_mask + 1) * sizeof (void *));
929 for (i = 0; i < BUF_LOCKS; i++)
930 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
931 kmem_cache_destroy(hdr_cache);
932 kmem_cache_destroy(buf_cache);
936 * Constructor callback - called when the cache is empty
937 * and a new buf is requested.
941 hdr_cons(void *vbuf, void *unused, int kmflag)
943 arc_buf_hdr_t *buf = vbuf;
945 bzero(buf, sizeof (arc_buf_hdr_t));
946 refcount_create(&buf->b_refcnt);
947 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
948 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
949 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
956 buf_cons(void *vbuf, void *unused, int kmflag)
958 arc_buf_t *buf = vbuf;
960 bzero(buf, sizeof (arc_buf_t));
961 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
962 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
968 * Destructor callback - called when a cached buf is
969 * no longer required.
973 hdr_dest(void *vbuf, void *unused)
975 arc_buf_hdr_t *buf = vbuf;
977 ASSERT(BUF_EMPTY(buf));
978 refcount_destroy(&buf->b_refcnt);
979 cv_destroy(&buf->b_cv);
980 mutex_destroy(&buf->b_freeze_lock);
981 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
986 buf_dest(void *vbuf, void *unused)
988 arc_buf_t *buf = vbuf;
990 mutex_destroy(&buf->b_evict_lock);
991 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
995 * Reclaim callback -- invoked when memory is low.
999 hdr_recl(void *unused)
1001 dprintf("hdr_recl called\n");
1003 * umem calls the reclaim func when we destroy the buf cache,
1004 * which is after we do arc_fini().
1007 cv_signal(&arc_reclaim_thr_cv);
1014 uint64_t hsize = 1ULL << 12;
1018 * The hash table is big enough to fill all of physical memory
1019 * with an average 64K block size. The table will take up
1020 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
1022 while (hsize * 65536 < (uint64_t)physmem * PAGESIZE)
1025 buf_hash_table.ht_mask = hsize - 1;
1026 buf_hash_table.ht_table =
1027 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1028 if (buf_hash_table.ht_table == NULL) {
1029 ASSERT(hsize > (1ULL << 8));
1034 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
1035 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
1036 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1037 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1039 for (i = 0; i < 256; i++)
1040 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1041 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1043 for (i = 0; i < BUF_LOCKS; i++) {
1044 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1045 NULL, MUTEX_DEFAULT, NULL);
1049 #define ARC_MINTIME (hz>>4) /* 62 ms */
1052 arc_cksum_verify(arc_buf_t *buf)
1056 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1059 mutex_enter(&buf->b_hdr->b_freeze_lock);
1060 if (buf->b_hdr->b_freeze_cksum == NULL ||
1061 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
1062 mutex_exit(&buf->b_hdr->b_freeze_lock);
1065 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1066 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1067 panic("buffer modified while frozen!");
1068 mutex_exit(&buf->b_hdr->b_freeze_lock);
1072 arc_cksum_equal(arc_buf_t *buf)
1077 mutex_enter(&buf->b_hdr->b_freeze_lock);
1078 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1079 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1080 mutex_exit(&buf->b_hdr->b_freeze_lock);
1086 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1088 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1091 mutex_enter(&buf->b_hdr->b_freeze_lock);
1092 if (buf->b_hdr->b_freeze_cksum != NULL) {
1093 mutex_exit(&buf->b_hdr->b_freeze_lock);
1096 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1097 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1098 buf->b_hdr->b_freeze_cksum);
1099 mutex_exit(&buf->b_hdr->b_freeze_lock);
1102 #endif /* illumos */
1107 typedef struct procctl {
1115 arc_buf_unwatch(arc_buf_t *buf)
1122 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1123 ctl.prwatch.pr_size = 0;
1124 ctl.prwatch.pr_wflags = 0;
1125 result = write(arc_procfd, &ctl, sizeof (ctl));
1126 ASSERT3U(result, ==, sizeof (ctl));
1133 arc_buf_watch(arc_buf_t *buf)
1140 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1141 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1142 ctl.prwatch.pr_wflags = WA_WRITE;
1143 result = write(arc_procfd, &ctl, sizeof (ctl));
1144 ASSERT3U(result, ==, sizeof (ctl));
1148 #endif /* illumos */
1151 arc_buf_thaw(arc_buf_t *buf)
1153 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1154 if (buf->b_hdr->b_state != arc_anon)
1155 panic("modifying non-anon buffer!");
1156 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1157 panic("modifying buffer while i/o in progress!");
1158 arc_cksum_verify(buf);
1161 mutex_enter(&buf->b_hdr->b_freeze_lock);
1162 if (buf->b_hdr->b_freeze_cksum != NULL) {
1163 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1164 buf->b_hdr->b_freeze_cksum = NULL;
1167 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1168 if (buf->b_hdr->b_thawed)
1169 kmem_free(buf->b_hdr->b_thawed, 1);
1170 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1173 mutex_exit(&buf->b_hdr->b_freeze_lock);
1176 arc_buf_unwatch(buf);
1177 #endif /* illumos */
1181 arc_buf_freeze(arc_buf_t *buf)
1183 kmutex_t *hash_lock;
1185 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1188 hash_lock = HDR_LOCK(buf->b_hdr);
1189 mutex_enter(hash_lock);
1191 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1192 buf->b_hdr->b_state == arc_anon);
1193 arc_cksum_compute(buf, B_FALSE);
1194 mutex_exit(hash_lock);
1199 get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock)
1201 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth);
1203 if (ab->b_type == ARC_BUFC_METADATA)
1204 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1);
1206 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1);
1207 buf_hashid += ARC_BUFC_NUMMETADATALISTS;
1210 *list = &state->arcs_lists[buf_hashid];
1211 *lock = ARCS_LOCK(state, buf_hashid);
1216 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1218 ASSERT(MUTEX_HELD(hash_lock));
1220 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1221 (ab->b_state != arc_anon)) {
1222 uint64_t delta = ab->b_size * ab->b_datacnt;
1223 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1227 get_buf_info(ab, ab->b_state, &list, &lock);
1228 ASSERT(!MUTEX_HELD(lock));
1230 ASSERT(list_link_active(&ab->b_arc_node));
1231 list_remove(list, ab);
1232 if (GHOST_STATE(ab->b_state)) {
1233 ASSERT0(ab->b_datacnt);
1234 ASSERT3P(ab->b_buf, ==, NULL);
1238 ASSERT3U(*size, >=, delta);
1239 atomic_add_64(size, -delta);
1241 /* remove the prefetch flag if we get a reference */
1242 if (ab->b_flags & ARC_PREFETCH)
1243 ab->b_flags &= ~ARC_PREFETCH;
1248 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1251 arc_state_t *state = ab->b_state;
1253 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1254 ASSERT(!GHOST_STATE(state));
1256 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1257 (state != arc_anon)) {
1258 uint64_t *size = &state->arcs_lsize[ab->b_type];
1262 get_buf_info(ab, state, &list, &lock);
1263 ASSERT(!MUTEX_HELD(lock));
1265 ASSERT(!list_link_active(&ab->b_arc_node));
1266 list_insert_head(list, ab);
1267 ASSERT(ab->b_datacnt > 0);
1268 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1275 * Move the supplied buffer to the indicated state. The mutex
1276 * for the buffer must be held by the caller.
1279 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1281 arc_state_t *old_state = ab->b_state;
1282 int64_t refcnt = refcount_count(&ab->b_refcnt);
1283 uint64_t from_delta, to_delta;
1287 ASSERT(MUTEX_HELD(hash_lock));
1288 ASSERT(new_state != old_state);
1289 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1290 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1291 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1293 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1296 * If this buffer is evictable, transfer it from the
1297 * old state list to the new state list.
1300 if (old_state != arc_anon) {
1302 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1304 get_buf_info(ab, old_state, &list, &lock);
1305 use_mutex = !MUTEX_HELD(lock);
1309 ASSERT(list_link_active(&ab->b_arc_node));
1310 list_remove(list, ab);
1313 * If prefetching out of the ghost cache,
1314 * we will have a non-zero datacnt.
1316 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1317 /* ghost elements have a ghost size */
1318 ASSERT(ab->b_buf == NULL);
1319 from_delta = ab->b_size;
1321 ASSERT3U(*size, >=, from_delta);
1322 atomic_add_64(size, -from_delta);
1327 if (new_state != arc_anon) {
1329 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1331 get_buf_info(ab, new_state, &list, &lock);
1332 use_mutex = !MUTEX_HELD(lock);
1336 list_insert_head(list, ab);
1338 /* ghost elements have a ghost size */
1339 if (GHOST_STATE(new_state)) {
1340 ASSERT(ab->b_datacnt == 0);
1341 ASSERT(ab->b_buf == NULL);
1342 to_delta = ab->b_size;
1344 atomic_add_64(size, to_delta);
1351 ASSERT(!BUF_EMPTY(ab));
1352 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1353 buf_hash_remove(ab);
1355 /* adjust state sizes */
1357 atomic_add_64(&new_state->arcs_size, to_delta);
1359 ASSERT3U(old_state->arcs_size, >=, from_delta);
1360 atomic_add_64(&old_state->arcs_size, -from_delta);
1362 ab->b_state = new_state;
1364 /* adjust l2arc hdr stats */
1365 if (new_state == arc_l2c_only)
1366 l2arc_hdr_stat_add();
1367 else if (old_state == arc_l2c_only)
1368 l2arc_hdr_stat_remove();
1372 arc_space_consume(uint64_t space, arc_space_type_t type)
1374 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1377 case ARC_SPACE_DATA:
1378 ARCSTAT_INCR(arcstat_data_size, space);
1380 case ARC_SPACE_OTHER:
1381 ARCSTAT_INCR(arcstat_other_size, space);
1383 case ARC_SPACE_HDRS:
1384 ARCSTAT_INCR(arcstat_hdr_size, space);
1386 case ARC_SPACE_L2HDRS:
1387 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1391 atomic_add_64(&arc_meta_used, space);
1392 atomic_add_64(&arc_size, space);
1396 arc_space_return(uint64_t space, arc_space_type_t type)
1398 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1401 case ARC_SPACE_DATA:
1402 ARCSTAT_INCR(arcstat_data_size, -space);
1404 case ARC_SPACE_OTHER:
1405 ARCSTAT_INCR(arcstat_other_size, -space);
1407 case ARC_SPACE_HDRS:
1408 ARCSTAT_INCR(arcstat_hdr_size, -space);
1410 case ARC_SPACE_L2HDRS:
1411 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1415 ASSERT(arc_meta_used >= space);
1416 if (arc_meta_max < arc_meta_used)
1417 arc_meta_max = arc_meta_used;
1418 atomic_add_64(&arc_meta_used, -space);
1419 ASSERT(arc_size >= space);
1420 atomic_add_64(&arc_size, -space);
1424 arc_data_buf_alloc(uint64_t size)
1426 if (arc_evict_needed(ARC_BUFC_DATA))
1427 cv_signal(&arc_reclaim_thr_cv);
1428 atomic_add_64(&arc_size, size);
1429 return (zio_data_buf_alloc(size));
1433 arc_data_buf_free(void *buf, uint64_t size)
1435 zio_data_buf_free(buf, size);
1436 ASSERT(arc_size >= size);
1437 atomic_add_64(&arc_size, -size);
1441 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1446 ASSERT3U(size, >, 0);
1447 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1448 ASSERT(BUF_EMPTY(hdr));
1451 hdr->b_spa = spa_load_guid(spa);
1452 hdr->b_state = arc_anon;
1453 hdr->b_arc_access = 0;
1454 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1457 buf->b_efunc = NULL;
1458 buf->b_private = NULL;
1461 arc_get_data_buf(buf);
1464 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1465 (void) refcount_add(&hdr->b_refcnt, tag);
1470 static char *arc_onloan_tag = "onloan";
1473 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1474 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1475 * buffers must be returned to the arc before they can be used by the DMU or
1479 arc_loan_buf(spa_t *spa, int size)
1483 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1485 atomic_add_64(&arc_loaned_bytes, size);
1490 * Return a loaned arc buffer to the arc.
1493 arc_return_buf(arc_buf_t *buf, void *tag)
1495 arc_buf_hdr_t *hdr = buf->b_hdr;
1497 ASSERT(buf->b_data != NULL);
1498 (void) refcount_add(&hdr->b_refcnt, tag);
1499 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1501 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1504 /* Detach an arc_buf from a dbuf (tag) */
1506 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1510 ASSERT(buf->b_data != NULL);
1512 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1513 (void) refcount_remove(&hdr->b_refcnt, tag);
1514 buf->b_efunc = NULL;
1515 buf->b_private = NULL;
1517 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1521 arc_buf_clone(arc_buf_t *from)
1524 arc_buf_hdr_t *hdr = from->b_hdr;
1525 uint64_t size = hdr->b_size;
1527 ASSERT(hdr->b_state != arc_anon);
1529 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1532 buf->b_efunc = NULL;
1533 buf->b_private = NULL;
1534 buf->b_next = hdr->b_buf;
1536 arc_get_data_buf(buf);
1537 bcopy(from->b_data, buf->b_data, size);
1540 * This buffer already exists in the arc so create a duplicate
1541 * copy for the caller. If the buffer is associated with user data
1542 * then track the size and number of duplicates. These stats will be
1543 * updated as duplicate buffers are created and destroyed.
1545 if (hdr->b_type == ARC_BUFC_DATA) {
1546 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1547 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1549 hdr->b_datacnt += 1;
1554 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1557 kmutex_t *hash_lock;
1560 * Check to see if this buffer is evicted. Callers
1561 * must verify b_data != NULL to know if the add_ref
1564 mutex_enter(&buf->b_evict_lock);
1565 if (buf->b_data == NULL) {
1566 mutex_exit(&buf->b_evict_lock);
1569 hash_lock = HDR_LOCK(buf->b_hdr);
1570 mutex_enter(hash_lock);
1572 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1573 mutex_exit(&buf->b_evict_lock);
1575 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1576 add_reference(hdr, hash_lock, tag);
1577 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1578 arc_access(hdr, hash_lock);
1579 mutex_exit(hash_lock);
1580 ARCSTAT_BUMP(arcstat_hits);
1581 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1582 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1583 data, metadata, hits);
1587 * Free the arc data buffer. If it is an l2arc write in progress,
1588 * the buffer is placed on l2arc_free_on_write to be freed later.
1591 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1593 arc_buf_hdr_t *hdr = buf->b_hdr;
1595 if (HDR_L2_WRITING(hdr)) {
1596 l2arc_data_free_t *df;
1597 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1598 df->l2df_data = buf->b_data;
1599 df->l2df_size = hdr->b_size;
1600 df->l2df_func = free_func;
1601 mutex_enter(&l2arc_free_on_write_mtx);
1602 list_insert_head(l2arc_free_on_write, df);
1603 mutex_exit(&l2arc_free_on_write_mtx);
1604 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1606 free_func(buf->b_data, hdr->b_size);
1611 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1615 /* free up data associated with the buf */
1617 arc_state_t *state = buf->b_hdr->b_state;
1618 uint64_t size = buf->b_hdr->b_size;
1619 arc_buf_contents_t type = buf->b_hdr->b_type;
1621 arc_cksum_verify(buf);
1623 arc_buf_unwatch(buf);
1624 #endif /* illumos */
1627 if (type == ARC_BUFC_METADATA) {
1628 arc_buf_data_free(buf, zio_buf_free);
1629 arc_space_return(size, ARC_SPACE_DATA);
1631 ASSERT(type == ARC_BUFC_DATA);
1632 arc_buf_data_free(buf, zio_data_buf_free);
1633 ARCSTAT_INCR(arcstat_data_size, -size);
1634 atomic_add_64(&arc_size, -size);
1637 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1638 uint64_t *cnt = &state->arcs_lsize[type];
1640 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1641 ASSERT(state != arc_anon);
1643 ASSERT3U(*cnt, >=, size);
1644 atomic_add_64(cnt, -size);
1646 ASSERT3U(state->arcs_size, >=, size);
1647 atomic_add_64(&state->arcs_size, -size);
1651 * If we're destroying a duplicate buffer make sure
1652 * that the appropriate statistics are updated.
1654 if (buf->b_hdr->b_datacnt > 1 &&
1655 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1656 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1657 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1659 ASSERT(buf->b_hdr->b_datacnt > 0);
1660 buf->b_hdr->b_datacnt -= 1;
1663 /* only remove the buf if requested */
1667 /* remove the buf from the hdr list */
1668 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1670 *bufp = buf->b_next;
1673 ASSERT(buf->b_efunc == NULL);
1675 /* clean up the buf */
1677 kmem_cache_free(buf_cache, buf);
1681 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1683 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1684 ASSERT3P(hdr->b_state, ==, arc_anon);
1685 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1686 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1688 if (l2hdr != NULL) {
1689 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1691 * To prevent arc_free() and l2arc_evict() from
1692 * attempting to free the same buffer at the same time,
1693 * a FREE_IN_PROGRESS flag is given to arc_free() to
1694 * give it priority. l2arc_evict() can't destroy this
1695 * header while we are waiting on l2arc_buflist_mtx.
1697 * The hdr may be removed from l2ad_buflist before we
1698 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1700 if (!buflist_held) {
1701 mutex_enter(&l2arc_buflist_mtx);
1702 l2hdr = hdr->b_l2hdr;
1705 if (l2hdr != NULL) {
1706 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr,
1708 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1709 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1710 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1711 if (hdr->b_state == arc_l2c_only)
1712 l2arc_hdr_stat_remove();
1713 hdr->b_l2hdr = NULL;
1717 mutex_exit(&l2arc_buflist_mtx);
1720 if (!BUF_EMPTY(hdr)) {
1721 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1722 buf_discard_identity(hdr);
1724 while (hdr->b_buf) {
1725 arc_buf_t *buf = hdr->b_buf;
1728 mutex_enter(&arc_eviction_mtx);
1729 mutex_enter(&buf->b_evict_lock);
1730 ASSERT(buf->b_hdr != NULL);
1731 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1732 hdr->b_buf = buf->b_next;
1733 buf->b_hdr = &arc_eviction_hdr;
1734 buf->b_next = arc_eviction_list;
1735 arc_eviction_list = buf;
1736 mutex_exit(&buf->b_evict_lock);
1737 mutex_exit(&arc_eviction_mtx);
1739 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1742 if (hdr->b_freeze_cksum != NULL) {
1743 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1744 hdr->b_freeze_cksum = NULL;
1746 if (hdr->b_thawed) {
1747 kmem_free(hdr->b_thawed, 1);
1748 hdr->b_thawed = NULL;
1751 ASSERT(!list_link_active(&hdr->b_arc_node));
1752 ASSERT3P(hdr->b_hash_next, ==, NULL);
1753 ASSERT3P(hdr->b_acb, ==, NULL);
1754 kmem_cache_free(hdr_cache, hdr);
1758 arc_buf_free(arc_buf_t *buf, void *tag)
1760 arc_buf_hdr_t *hdr = buf->b_hdr;
1761 int hashed = hdr->b_state != arc_anon;
1763 ASSERT(buf->b_efunc == NULL);
1764 ASSERT(buf->b_data != NULL);
1767 kmutex_t *hash_lock = HDR_LOCK(hdr);
1769 mutex_enter(hash_lock);
1771 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1773 (void) remove_reference(hdr, hash_lock, tag);
1774 if (hdr->b_datacnt > 1) {
1775 arc_buf_destroy(buf, FALSE, TRUE);
1777 ASSERT(buf == hdr->b_buf);
1778 ASSERT(buf->b_efunc == NULL);
1779 hdr->b_flags |= ARC_BUF_AVAILABLE;
1781 mutex_exit(hash_lock);
1782 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1785 * We are in the middle of an async write. Don't destroy
1786 * this buffer unless the write completes before we finish
1787 * decrementing the reference count.
1789 mutex_enter(&arc_eviction_mtx);
1790 (void) remove_reference(hdr, NULL, tag);
1791 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1792 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1793 mutex_exit(&arc_eviction_mtx);
1795 arc_hdr_destroy(hdr);
1797 if (remove_reference(hdr, NULL, tag) > 0)
1798 arc_buf_destroy(buf, FALSE, TRUE);
1800 arc_hdr_destroy(hdr);
1805 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1807 arc_buf_hdr_t *hdr = buf->b_hdr;
1808 kmutex_t *hash_lock = HDR_LOCK(hdr);
1809 boolean_t no_callback = (buf->b_efunc == NULL);
1811 if (hdr->b_state == arc_anon) {
1812 ASSERT(hdr->b_datacnt == 1);
1813 arc_buf_free(buf, tag);
1814 return (no_callback);
1817 mutex_enter(hash_lock);
1819 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1820 ASSERT(hdr->b_state != arc_anon);
1821 ASSERT(buf->b_data != NULL);
1823 (void) remove_reference(hdr, hash_lock, tag);
1824 if (hdr->b_datacnt > 1) {
1826 arc_buf_destroy(buf, FALSE, TRUE);
1827 } else if (no_callback) {
1828 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1829 ASSERT(buf->b_efunc == NULL);
1830 hdr->b_flags |= ARC_BUF_AVAILABLE;
1832 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1833 refcount_is_zero(&hdr->b_refcnt));
1834 mutex_exit(hash_lock);
1835 return (no_callback);
1839 arc_buf_size(arc_buf_t *buf)
1841 return (buf->b_hdr->b_size);
1845 * Called from the DMU to determine if the current buffer should be
1846 * evicted. In order to ensure proper locking, the eviction must be initiated
1847 * from the DMU. Return true if the buffer is associated with user data and
1848 * duplicate buffers still exist.
1851 arc_buf_eviction_needed(arc_buf_t *buf)
1854 boolean_t evict_needed = B_FALSE;
1856 if (zfs_disable_dup_eviction)
1859 mutex_enter(&buf->b_evict_lock);
1863 * We are in arc_do_user_evicts(); let that function
1864 * perform the eviction.
1866 ASSERT(buf->b_data == NULL);
1867 mutex_exit(&buf->b_evict_lock);
1869 } else if (buf->b_data == NULL) {
1871 * We have already been added to the arc eviction list;
1872 * recommend eviction.
1874 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1875 mutex_exit(&buf->b_evict_lock);
1879 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1880 evict_needed = B_TRUE;
1882 mutex_exit(&buf->b_evict_lock);
1883 return (evict_needed);
1887 * Evict buffers from list until we've removed the specified number of
1888 * bytes. Move the removed buffers to the appropriate evict state.
1889 * If the recycle flag is set, then attempt to "recycle" a buffer:
1890 * - look for a buffer to evict that is `bytes' long.
1891 * - return the data block from this buffer rather than freeing it.
1892 * This flag is used by callers that are trying to make space for a
1893 * new buffer in a full arc cache.
1895 * This function makes a "best effort". It skips over any buffers
1896 * it can't get a hash_lock on, and so may not catch all candidates.
1897 * It may also return without evicting as much space as requested.
1900 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1901 arc_buf_contents_t type)
1903 arc_state_t *evicted_state;
1904 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1905 int64_t bytes_remaining;
1906 arc_buf_hdr_t *ab, *ab_prev = NULL;
1907 list_t *evicted_list, *list, *evicted_list_start, *list_start;
1908 kmutex_t *lock, *evicted_lock;
1909 kmutex_t *hash_lock;
1910 boolean_t have_lock;
1911 void *stolen = NULL;
1912 static int evict_metadata_offset, evict_data_offset;
1913 int i, idx, offset, list_count, count;
1915 ASSERT(state == arc_mru || state == arc_mfu);
1917 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1919 if (type == ARC_BUFC_METADATA) {
1921 list_count = ARC_BUFC_NUMMETADATALISTS;
1922 list_start = &state->arcs_lists[0];
1923 evicted_list_start = &evicted_state->arcs_lists[0];
1924 idx = evict_metadata_offset;
1926 offset = ARC_BUFC_NUMMETADATALISTS;
1927 list_start = &state->arcs_lists[offset];
1928 evicted_list_start = &evicted_state->arcs_lists[offset];
1929 list_count = ARC_BUFC_NUMDATALISTS;
1930 idx = evict_data_offset;
1932 bytes_remaining = evicted_state->arcs_lsize[type];
1936 list = &list_start[idx];
1937 evicted_list = &evicted_list_start[idx];
1938 lock = ARCS_LOCK(state, (offset + idx));
1939 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx));
1942 mutex_enter(evicted_lock);
1944 for (ab = list_tail(list); ab; ab = ab_prev) {
1945 ab_prev = list_prev(list, ab);
1946 bytes_remaining -= (ab->b_size * ab->b_datacnt);
1947 /* prefetch buffers have a minimum lifespan */
1948 if (HDR_IO_IN_PROGRESS(ab) ||
1949 (spa && ab->b_spa != spa) ||
1950 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1951 ddi_get_lbolt() - ab->b_arc_access <
1952 arc_min_prefetch_lifespan)) {
1956 /* "lookahead" for better eviction candidate */
1957 if (recycle && ab->b_size != bytes &&
1958 ab_prev && ab_prev->b_size == bytes)
1960 hash_lock = HDR_LOCK(ab);
1961 have_lock = MUTEX_HELD(hash_lock);
1962 if (have_lock || mutex_tryenter(hash_lock)) {
1963 ASSERT0(refcount_count(&ab->b_refcnt));
1964 ASSERT(ab->b_datacnt > 0);
1966 arc_buf_t *buf = ab->b_buf;
1967 if (!mutex_tryenter(&buf->b_evict_lock)) {
1972 bytes_evicted += ab->b_size;
1973 if (recycle && ab->b_type == type &&
1974 ab->b_size == bytes &&
1975 !HDR_L2_WRITING(ab)) {
1976 stolen = buf->b_data;
1981 mutex_enter(&arc_eviction_mtx);
1982 arc_buf_destroy(buf,
1983 buf->b_data == stolen, FALSE);
1984 ab->b_buf = buf->b_next;
1985 buf->b_hdr = &arc_eviction_hdr;
1986 buf->b_next = arc_eviction_list;
1987 arc_eviction_list = buf;
1988 mutex_exit(&arc_eviction_mtx);
1989 mutex_exit(&buf->b_evict_lock);
1991 mutex_exit(&buf->b_evict_lock);
1992 arc_buf_destroy(buf,
1993 buf->b_data == stolen, TRUE);
1998 ARCSTAT_INCR(arcstat_evict_l2_cached,
2001 if (l2arc_write_eligible(ab->b_spa, ab)) {
2002 ARCSTAT_INCR(arcstat_evict_l2_eligible,
2006 arcstat_evict_l2_ineligible,
2011 if (ab->b_datacnt == 0) {
2012 arc_change_state(evicted_state, ab, hash_lock);
2013 ASSERT(HDR_IN_HASH_TABLE(ab));
2014 ab->b_flags |= ARC_IN_HASH_TABLE;
2015 ab->b_flags &= ~ARC_BUF_AVAILABLE;
2016 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
2019 mutex_exit(hash_lock);
2020 if (bytes >= 0 && bytes_evicted >= bytes)
2022 if (bytes_remaining > 0) {
2023 mutex_exit(evicted_lock);
2025 idx = ((idx + 1) & (list_count - 1));
2034 mutex_exit(evicted_lock);
2037 idx = ((idx + 1) & (list_count - 1));
2040 if (bytes_evicted < bytes) {
2041 if (count < list_count)
2044 dprintf("only evicted %lld bytes from %x",
2045 (longlong_t)bytes_evicted, state);
2047 if (type == ARC_BUFC_METADATA)
2048 evict_metadata_offset = idx;
2050 evict_data_offset = idx;
2053 ARCSTAT_INCR(arcstat_evict_skip, skipped);
2056 ARCSTAT_INCR(arcstat_mutex_miss, missed);
2059 * We have just evicted some data into the ghost state, make
2060 * sure we also adjust the ghost state size if necessary.
2063 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
2064 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
2065 arc_mru_ghost->arcs_size - arc_c;
2067 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
2069 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
2070 arc_evict_ghost(arc_mru_ghost, 0, todelete);
2071 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
2072 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
2073 arc_mru_ghost->arcs_size +
2074 arc_mfu_ghost->arcs_size - arc_c);
2075 arc_evict_ghost(arc_mfu_ghost, 0, todelete);
2079 ARCSTAT_BUMP(arcstat_stolen);
2085 * Remove buffers from list until we've removed the specified number of
2086 * bytes. Destroy the buffers that are removed.
2089 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
2091 arc_buf_hdr_t *ab, *ab_prev;
2092 arc_buf_hdr_t marker = { 0 };
2093 list_t *list, *list_start;
2094 kmutex_t *hash_lock, *lock;
2095 uint64_t bytes_deleted = 0;
2096 uint64_t bufs_skipped = 0;
2097 static int evict_offset;
2098 int list_count, idx = evict_offset;
2099 int offset, count = 0;
2101 ASSERT(GHOST_STATE(state));
2104 * data lists come after metadata lists
2106 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS];
2107 list_count = ARC_BUFC_NUMDATALISTS;
2108 offset = ARC_BUFC_NUMMETADATALISTS;
2111 list = &list_start[idx];
2112 lock = ARCS_LOCK(state, idx + offset);
2115 for (ab = list_tail(list); ab; ab = ab_prev) {
2116 ab_prev = list_prev(list, ab);
2117 if (spa && ab->b_spa != spa)
2120 /* ignore markers */
2124 hash_lock = HDR_LOCK(ab);
2125 /* caller may be trying to modify this buffer, skip it */
2126 if (MUTEX_HELD(hash_lock))
2128 if (mutex_tryenter(hash_lock)) {
2129 ASSERT(!HDR_IO_IN_PROGRESS(ab));
2130 ASSERT(ab->b_buf == NULL);
2131 ARCSTAT_BUMP(arcstat_deleted);
2132 bytes_deleted += ab->b_size;
2134 if (ab->b_l2hdr != NULL) {
2136 * This buffer is cached on the 2nd Level ARC;
2137 * don't destroy the header.
2139 arc_change_state(arc_l2c_only, ab, hash_lock);
2140 mutex_exit(hash_lock);
2142 arc_change_state(arc_anon, ab, hash_lock);
2143 mutex_exit(hash_lock);
2144 arc_hdr_destroy(ab);
2147 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
2148 if (bytes >= 0 && bytes_deleted >= bytes)
2150 } else if (bytes < 0) {
2152 * Insert a list marker and then wait for the
2153 * hash lock to become available. Once its
2154 * available, restart from where we left off.
2156 list_insert_after(list, ab, &marker);
2158 mutex_enter(hash_lock);
2159 mutex_exit(hash_lock);
2161 ab_prev = list_prev(list, &marker);
2162 list_remove(list, &marker);
2167 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1));
2170 if (count < list_count)
2174 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] &&
2175 (bytes < 0 || bytes_deleted < bytes)) {
2176 list_start = &state->arcs_lists[0];
2177 list_count = ARC_BUFC_NUMMETADATALISTS;
2183 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2187 if (bytes_deleted < bytes)
2188 dprintf("only deleted %lld bytes from %p",
2189 (longlong_t)bytes_deleted, state);
2195 int64_t adjustment, delta;
2201 adjustment = MIN((int64_t)(arc_size - arc_c),
2202 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2205 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2206 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2207 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
2208 adjustment -= delta;
2211 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2212 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2213 (void) arc_evict(arc_mru, 0, delta, FALSE,
2221 adjustment = arc_size - arc_c;
2223 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2224 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2225 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2226 adjustment -= delta;
2229 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2230 int64_t delta = MIN(adjustment,
2231 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2232 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2237 * Adjust ghost lists
2240 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2242 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2243 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2244 arc_evict_ghost(arc_mru_ghost, 0, delta);
2248 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2250 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2251 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2252 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2257 arc_do_user_evicts(void)
2259 static arc_buf_t *tmp_arc_eviction_list;
2262 * Move list over to avoid LOR
2265 mutex_enter(&arc_eviction_mtx);
2266 tmp_arc_eviction_list = arc_eviction_list;
2267 arc_eviction_list = NULL;
2268 mutex_exit(&arc_eviction_mtx);
2270 while (tmp_arc_eviction_list != NULL) {
2271 arc_buf_t *buf = tmp_arc_eviction_list;
2272 tmp_arc_eviction_list = buf->b_next;
2273 mutex_enter(&buf->b_evict_lock);
2275 mutex_exit(&buf->b_evict_lock);
2277 if (buf->b_efunc != NULL)
2278 VERIFY(buf->b_efunc(buf) == 0);
2280 buf->b_efunc = NULL;
2281 buf->b_private = NULL;
2282 kmem_cache_free(buf_cache, buf);
2285 if (arc_eviction_list != NULL)
2290 * Flush all *evictable* data from the cache for the given spa.
2291 * NOTE: this will not touch "active" (i.e. referenced) data.
2294 arc_flush(spa_t *spa)
2299 guid = spa_load_guid(spa);
2301 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) {
2302 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2306 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) {
2307 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2311 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) {
2312 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2316 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) {
2317 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2322 arc_evict_ghost(arc_mru_ghost, guid, -1);
2323 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2325 mutex_enter(&arc_reclaim_thr_lock);
2326 arc_do_user_evicts();
2327 mutex_exit(&arc_reclaim_thr_lock);
2328 ASSERT(spa || arc_eviction_list == NULL);
2334 if (arc_c > arc_c_min) {
2338 to_free = arc_c >> arc_shrink_shift;
2340 to_free = arc_c >> arc_shrink_shift;
2342 if (arc_c > arc_c_min + to_free)
2343 atomic_add_64(&arc_c, -to_free);
2347 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2348 if (arc_c > arc_size)
2349 arc_c = MAX(arc_size, arc_c_min);
2351 arc_p = (arc_c >> 1);
2352 ASSERT(arc_c >= arc_c_min);
2353 ASSERT((int64_t)arc_p >= 0);
2356 if (arc_size > arc_c)
2360 static int needfree = 0;
2363 arc_reclaim_needed(void)
2372 * Cooperate with pagedaemon when it's time for it to scan
2373 * and reclaim some pages.
2375 if (vm_paging_needed())
2380 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2385 * check that we're out of range of the pageout scanner. It starts to
2386 * schedule paging if freemem is less than lotsfree and needfree.
2387 * lotsfree is the high-water mark for pageout, and needfree is the
2388 * number of needed free pages. We add extra pages here to make sure
2389 * the scanner doesn't start up while we're freeing memory.
2391 if (freemem < lotsfree + needfree + extra)
2395 * check to make sure that swapfs has enough space so that anon
2396 * reservations can still succeed. anon_resvmem() checks that the
2397 * availrmem is greater than swapfs_minfree, and the number of reserved
2398 * swap pages. We also add a bit of extra here just to prevent
2399 * circumstances from getting really dire.
2401 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2406 * If we're on an i386 platform, it's possible that we'll exhaust the
2407 * kernel heap space before we ever run out of available physical
2408 * memory. Most checks of the size of the heap_area compare against
2409 * tune.t_minarmem, which is the minimum available real memory that we
2410 * can have in the system. However, this is generally fixed at 25 pages
2411 * which is so low that it's useless. In this comparison, we seek to
2412 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2413 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2416 if (btop(vmem_size(heap_arena, VMEM_FREE)) <
2417 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
2421 if (kmem_used() > (kmem_size() * 3) / 4)
2426 if (spa_get_random(100) == 0)
2432 extern kmem_cache_t *zio_buf_cache[];
2433 extern kmem_cache_t *zio_data_buf_cache[];
2436 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2439 kmem_cache_t *prev_cache = NULL;
2440 kmem_cache_t *prev_data_cache = NULL;
2443 if (arc_meta_used >= arc_meta_limit) {
2445 * We are exceeding our meta-data cache limit.
2446 * Purge some DNLC entries to release holds on meta-data.
2448 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2452 * Reclaim unused memory from all kmem caches.
2459 * An aggressive reclamation will shrink the cache size as well as
2460 * reap free buffers from the arc kmem caches.
2462 if (strat == ARC_RECLAIM_AGGR)
2465 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2466 if (zio_buf_cache[i] != prev_cache) {
2467 prev_cache = zio_buf_cache[i];
2468 kmem_cache_reap_now(zio_buf_cache[i]);
2470 if (zio_data_buf_cache[i] != prev_data_cache) {
2471 prev_data_cache = zio_data_buf_cache[i];
2472 kmem_cache_reap_now(zio_data_buf_cache[i]);
2475 kmem_cache_reap_now(buf_cache);
2476 kmem_cache_reap_now(hdr_cache);
2480 arc_reclaim_thread(void *dummy __unused)
2482 clock_t growtime = 0;
2483 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2486 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2488 mutex_enter(&arc_reclaim_thr_lock);
2489 while (arc_thread_exit == 0) {
2490 if (arc_reclaim_needed()) {
2493 if (last_reclaim == ARC_RECLAIM_CONS) {
2494 last_reclaim = ARC_RECLAIM_AGGR;
2496 last_reclaim = ARC_RECLAIM_CONS;
2500 last_reclaim = ARC_RECLAIM_AGGR;
2504 /* reset the growth delay for every reclaim */
2505 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2507 if (needfree && last_reclaim == ARC_RECLAIM_CONS) {
2509 * If needfree is TRUE our vm_lowmem hook
2510 * was called and in that case we must free some
2511 * memory, so switch to aggressive mode.
2514 last_reclaim = ARC_RECLAIM_AGGR;
2516 arc_kmem_reap_now(last_reclaim);
2519 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2520 arc_no_grow = FALSE;
2525 if (arc_eviction_list != NULL)
2526 arc_do_user_evicts();
2535 /* block until needed, or one second, whichever is shorter */
2536 CALLB_CPR_SAFE_BEGIN(&cpr);
2537 (void) cv_timedwait(&arc_reclaim_thr_cv,
2538 &arc_reclaim_thr_lock, hz);
2539 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2542 arc_thread_exit = 0;
2543 cv_broadcast(&arc_reclaim_thr_cv);
2544 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2549 * Adapt arc info given the number of bytes we are trying to add and
2550 * the state that we are comming from. This function is only called
2551 * when we are adding new content to the cache.
2554 arc_adapt(int bytes, arc_state_t *state)
2557 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2559 if (state == arc_l2c_only)
2564 * Adapt the target size of the MRU list:
2565 * - if we just hit in the MRU ghost list, then increase
2566 * the target size of the MRU list.
2567 * - if we just hit in the MFU ghost list, then increase
2568 * the target size of the MFU list by decreasing the
2569 * target size of the MRU list.
2571 if (state == arc_mru_ghost) {
2572 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2573 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2574 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2576 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2577 } else if (state == arc_mfu_ghost) {
2580 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2581 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2582 mult = MIN(mult, 10);
2584 delta = MIN(bytes * mult, arc_p);
2585 arc_p = MAX(arc_p_min, arc_p - delta);
2587 ASSERT((int64_t)arc_p >= 0);
2589 if (arc_reclaim_needed()) {
2590 cv_signal(&arc_reclaim_thr_cv);
2597 if (arc_c >= arc_c_max)
2601 * If we're within (2 * maxblocksize) bytes of the target
2602 * cache size, increment the target cache size
2604 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2605 atomic_add_64(&arc_c, (int64_t)bytes);
2606 if (arc_c > arc_c_max)
2608 else if (state == arc_anon)
2609 atomic_add_64(&arc_p, (int64_t)bytes);
2613 ASSERT((int64_t)arc_p >= 0);
2617 * Check if the cache has reached its limits and eviction is required
2621 arc_evict_needed(arc_buf_contents_t type)
2623 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2629 * If zio data pages are being allocated out of a separate heap segment,
2630 * then enforce that the size of available vmem for this area remains
2631 * above about 1/32nd free.
2633 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2634 vmem_size(zio_arena, VMEM_FREE) <
2635 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2640 if (arc_reclaim_needed())
2643 return (arc_size > arc_c);
2647 * The buffer, supplied as the first argument, needs a data block.
2648 * So, if we are at cache max, determine which cache should be victimized.
2649 * We have the following cases:
2651 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2652 * In this situation if we're out of space, but the resident size of the MFU is
2653 * under the limit, victimize the MFU cache to satisfy this insertion request.
2655 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2656 * Here, we've used up all of the available space for the MRU, so we need to
2657 * evict from our own cache instead. Evict from the set of resident MRU
2660 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2661 * c minus p represents the MFU space in the cache, since p is the size of the
2662 * cache that is dedicated to the MRU. In this situation there's still space on
2663 * the MFU side, so the MRU side needs to be victimized.
2665 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2666 * MFU's resident set is consuming more space than it has been allotted. In
2667 * this situation, we must victimize our own cache, the MFU, for this insertion.
2670 arc_get_data_buf(arc_buf_t *buf)
2672 arc_state_t *state = buf->b_hdr->b_state;
2673 uint64_t size = buf->b_hdr->b_size;
2674 arc_buf_contents_t type = buf->b_hdr->b_type;
2676 arc_adapt(size, state);
2679 * We have not yet reached cache maximum size,
2680 * just allocate a new buffer.
2682 if (!arc_evict_needed(type)) {
2683 if (type == ARC_BUFC_METADATA) {
2684 buf->b_data = zio_buf_alloc(size);
2685 arc_space_consume(size, ARC_SPACE_DATA);
2687 ASSERT(type == ARC_BUFC_DATA);
2688 buf->b_data = zio_data_buf_alloc(size);
2689 ARCSTAT_INCR(arcstat_data_size, size);
2690 atomic_add_64(&arc_size, size);
2696 * If we are prefetching from the mfu ghost list, this buffer
2697 * will end up on the mru list; so steal space from there.
2699 if (state == arc_mfu_ghost)
2700 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2701 else if (state == arc_mru_ghost)
2704 if (state == arc_mru || state == arc_anon) {
2705 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2706 state = (arc_mfu->arcs_lsize[type] >= size &&
2707 arc_p > mru_used) ? arc_mfu : arc_mru;
2710 uint64_t mfu_space = arc_c - arc_p;
2711 state = (arc_mru->arcs_lsize[type] >= size &&
2712 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2714 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2715 if (type == ARC_BUFC_METADATA) {
2716 buf->b_data = zio_buf_alloc(size);
2717 arc_space_consume(size, ARC_SPACE_DATA);
2719 ASSERT(type == ARC_BUFC_DATA);
2720 buf->b_data = zio_data_buf_alloc(size);
2721 ARCSTAT_INCR(arcstat_data_size, size);
2722 atomic_add_64(&arc_size, size);
2724 ARCSTAT_BUMP(arcstat_recycle_miss);
2726 ASSERT(buf->b_data != NULL);
2729 * Update the state size. Note that ghost states have a
2730 * "ghost size" and so don't need to be updated.
2732 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2733 arc_buf_hdr_t *hdr = buf->b_hdr;
2735 atomic_add_64(&hdr->b_state->arcs_size, size);
2736 if (list_link_active(&hdr->b_arc_node)) {
2737 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2738 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2741 * If we are growing the cache, and we are adding anonymous
2742 * data, and we have outgrown arc_p, update arc_p
2744 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2745 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2746 arc_p = MIN(arc_c, arc_p + size);
2748 ARCSTAT_BUMP(arcstat_allocated);
2752 * This routine is called whenever a buffer is accessed.
2753 * NOTE: the hash lock is dropped in this function.
2756 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2760 ASSERT(MUTEX_HELD(hash_lock));
2762 if (buf->b_state == arc_anon) {
2764 * This buffer is not in the cache, and does not
2765 * appear in our "ghost" list. Add the new buffer
2769 ASSERT(buf->b_arc_access == 0);
2770 buf->b_arc_access = ddi_get_lbolt();
2771 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2772 arc_change_state(arc_mru, buf, hash_lock);
2774 } else if (buf->b_state == arc_mru) {
2775 now = ddi_get_lbolt();
2778 * If this buffer is here because of a prefetch, then either:
2779 * - clear the flag if this is a "referencing" read
2780 * (any subsequent access will bump this into the MFU state).
2782 * - move the buffer to the head of the list if this is
2783 * another prefetch (to make it less likely to be evicted).
2785 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2786 if (refcount_count(&buf->b_refcnt) == 0) {
2787 ASSERT(list_link_active(&buf->b_arc_node));
2789 buf->b_flags &= ~ARC_PREFETCH;
2790 ARCSTAT_BUMP(arcstat_mru_hits);
2792 buf->b_arc_access = now;
2797 * This buffer has been "accessed" only once so far,
2798 * but it is still in the cache. Move it to the MFU
2801 if (now > buf->b_arc_access + ARC_MINTIME) {
2803 * More than 125ms have passed since we
2804 * instantiated this buffer. Move it to the
2805 * most frequently used state.
2807 buf->b_arc_access = now;
2808 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2809 arc_change_state(arc_mfu, buf, hash_lock);
2811 ARCSTAT_BUMP(arcstat_mru_hits);
2812 } else if (buf->b_state == arc_mru_ghost) {
2813 arc_state_t *new_state;
2815 * This buffer has been "accessed" recently, but
2816 * was evicted from the cache. Move it to the
2820 if (buf->b_flags & ARC_PREFETCH) {
2821 new_state = arc_mru;
2822 if (refcount_count(&buf->b_refcnt) > 0)
2823 buf->b_flags &= ~ARC_PREFETCH;
2824 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2826 new_state = arc_mfu;
2827 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2830 buf->b_arc_access = ddi_get_lbolt();
2831 arc_change_state(new_state, buf, hash_lock);
2833 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2834 } else if (buf->b_state == arc_mfu) {
2836 * This buffer has been accessed more than once and is
2837 * still in the cache. Keep it in the MFU state.
2839 * NOTE: an add_reference() that occurred when we did
2840 * the arc_read() will have kicked this off the list.
2841 * If it was a prefetch, we will explicitly move it to
2842 * the head of the list now.
2844 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2845 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2846 ASSERT(list_link_active(&buf->b_arc_node));
2848 ARCSTAT_BUMP(arcstat_mfu_hits);
2849 buf->b_arc_access = ddi_get_lbolt();
2850 } else if (buf->b_state == arc_mfu_ghost) {
2851 arc_state_t *new_state = arc_mfu;
2853 * This buffer has been accessed more than once but has
2854 * been evicted from the cache. Move it back to the
2858 if (buf->b_flags & ARC_PREFETCH) {
2860 * This is a prefetch access...
2861 * move this block back to the MRU state.
2863 ASSERT0(refcount_count(&buf->b_refcnt));
2864 new_state = arc_mru;
2867 buf->b_arc_access = ddi_get_lbolt();
2868 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2869 arc_change_state(new_state, buf, hash_lock);
2871 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2872 } else if (buf->b_state == arc_l2c_only) {
2874 * This buffer is on the 2nd Level ARC.
2877 buf->b_arc_access = ddi_get_lbolt();
2878 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2879 arc_change_state(arc_mfu, buf, hash_lock);
2881 ASSERT(!"invalid arc state");
2885 /* a generic arc_done_func_t which you can use */
2888 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2890 if (zio == NULL || zio->io_error == 0)
2891 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2892 VERIFY(arc_buf_remove_ref(buf, arg));
2895 /* a generic arc_done_func_t */
2897 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2899 arc_buf_t **bufp = arg;
2900 if (zio && zio->io_error) {
2901 VERIFY(arc_buf_remove_ref(buf, arg));
2905 ASSERT(buf->b_data);
2910 arc_read_done(zio_t *zio)
2912 arc_buf_hdr_t *hdr, *found;
2914 arc_buf_t *abuf; /* buffer we're assigning to callback */
2915 kmutex_t *hash_lock;
2916 arc_callback_t *callback_list, *acb;
2917 int freeable = FALSE;
2919 buf = zio->io_private;
2923 * The hdr was inserted into hash-table and removed from lists
2924 * prior to starting I/O. We should find this header, since
2925 * it's in the hash table, and it should be legit since it's
2926 * not possible to evict it during the I/O. The only possible
2927 * reason for it not to be found is if we were freed during the
2930 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2933 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2934 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2935 (found == hdr && HDR_L2_READING(hdr)));
2937 hdr->b_flags &= ~ARC_L2_EVICTED;
2938 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2939 hdr->b_flags &= ~ARC_L2CACHE;
2941 /* byteswap if necessary */
2942 callback_list = hdr->b_acb;
2943 ASSERT(callback_list != NULL);
2944 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2945 dmu_object_byteswap_t bswap =
2946 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2947 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2948 byteswap_uint64_array :
2949 dmu_ot_byteswap[bswap].ob_func;
2950 func(buf->b_data, hdr->b_size);
2953 arc_cksum_compute(buf, B_FALSE);
2956 #endif /* illumos */
2958 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2960 * Only call arc_access on anonymous buffers. This is because
2961 * if we've issued an I/O for an evicted buffer, we've already
2962 * called arc_access (to prevent any simultaneous readers from
2963 * getting confused).
2965 arc_access(hdr, hash_lock);
2968 /* create copies of the data buffer for the callers */
2970 for (acb = callback_list; acb; acb = acb->acb_next) {
2971 if (acb->acb_done) {
2973 ARCSTAT_BUMP(arcstat_duplicate_reads);
2974 abuf = arc_buf_clone(buf);
2976 acb->acb_buf = abuf;
2981 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2982 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2984 ASSERT(buf->b_efunc == NULL);
2985 ASSERT(hdr->b_datacnt == 1);
2986 hdr->b_flags |= ARC_BUF_AVAILABLE;
2989 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2991 if (zio->io_error != 0) {
2992 hdr->b_flags |= ARC_IO_ERROR;
2993 if (hdr->b_state != arc_anon)
2994 arc_change_state(arc_anon, hdr, hash_lock);
2995 if (HDR_IN_HASH_TABLE(hdr))
2996 buf_hash_remove(hdr);
2997 freeable = refcount_is_zero(&hdr->b_refcnt);
3001 * Broadcast before we drop the hash_lock to avoid the possibility
3002 * that the hdr (and hence the cv) might be freed before we get to
3003 * the cv_broadcast().
3005 cv_broadcast(&hdr->b_cv);
3008 mutex_exit(hash_lock);
3011 * This block was freed while we waited for the read to
3012 * complete. It has been removed from the hash table and
3013 * moved to the anonymous state (so that it won't show up
3016 ASSERT3P(hdr->b_state, ==, arc_anon);
3017 freeable = refcount_is_zero(&hdr->b_refcnt);
3020 /* execute each callback and free its structure */
3021 while ((acb = callback_list) != NULL) {
3023 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3025 if (acb->acb_zio_dummy != NULL) {
3026 acb->acb_zio_dummy->io_error = zio->io_error;
3027 zio_nowait(acb->acb_zio_dummy);
3030 callback_list = acb->acb_next;
3031 kmem_free(acb, sizeof (arc_callback_t));
3035 arc_hdr_destroy(hdr);
3039 * "Read" the block block at the specified DVA (in bp) via the
3040 * cache. If the block is found in the cache, invoke the provided
3041 * callback immediately and return. Note that the `zio' parameter
3042 * in the callback will be NULL in this case, since no IO was
3043 * required. If the block is not in the cache pass the read request
3044 * on to the spa with a substitute callback function, so that the
3045 * requested block will be added to the cache.
3047 * If a read request arrives for a block that has a read in-progress,
3048 * either wait for the in-progress read to complete (and return the
3049 * results); or, if this is a read with a "done" func, add a record
3050 * to the read to invoke the "done" func when the read completes,
3051 * and return; or just return.
3053 * arc_read_done() will invoke all the requested "done" functions
3054 * for readers of this block.
3057 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3058 void *private, int priority, int zio_flags, uint32_t *arc_flags,
3059 const zbookmark_t *zb)
3062 arc_buf_t *buf = NULL;
3063 kmutex_t *hash_lock;
3065 uint64_t guid = spa_load_guid(spa);
3068 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3070 if (hdr && hdr->b_datacnt > 0) {
3072 *arc_flags |= ARC_CACHED;
3074 if (HDR_IO_IN_PROGRESS(hdr)) {
3076 if (*arc_flags & ARC_WAIT) {
3077 cv_wait(&hdr->b_cv, hash_lock);
3078 mutex_exit(hash_lock);
3081 ASSERT(*arc_flags & ARC_NOWAIT);
3084 arc_callback_t *acb = NULL;
3086 acb = kmem_zalloc(sizeof (arc_callback_t),
3088 acb->acb_done = done;
3089 acb->acb_private = private;
3091 acb->acb_zio_dummy = zio_null(pio,
3092 spa, NULL, NULL, NULL, zio_flags);
3094 ASSERT(acb->acb_done != NULL);
3095 acb->acb_next = hdr->b_acb;
3097 add_reference(hdr, hash_lock, private);
3098 mutex_exit(hash_lock);
3101 mutex_exit(hash_lock);
3105 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3108 add_reference(hdr, hash_lock, private);
3110 * If this block is already in use, create a new
3111 * copy of the data so that we will be guaranteed
3112 * that arc_release() will always succeed.
3116 ASSERT(buf->b_data);
3117 if (HDR_BUF_AVAILABLE(hdr)) {
3118 ASSERT(buf->b_efunc == NULL);
3119 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3121 buf = arc_buf_clone(buf);
3124 } else if (*arc_flags & ARC_PREFETCH &&
3125 refcount_count(&hdr->b_refcnt) == 0) {
3126 hdr->b_flags |= ARC_PREFETCH;
3128 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
3129 arc_access(hdr, hash_lock);
3130 if (*arc_flags & ARC_L2CACHE)
3131 hdr->b_flags |= ARC_L2CACHE;
3132 mutex_exit(hash_lock);
3133 ARCSTAT_BUMP(arcstat_hits);
3134 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3135 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3136 data, metadata, hits);
3139 done(NULL, buf, private);
3141 uint64_t size = BP_GET_LSIZE(bp);
3142 arc_callback_t *acb;
3145 boolean_t devw = B_FALSE;
3148 /* this block is not in the cache */
3149 arc_buf_hdr_t *exists;
3150 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
3151 buf = arc_buf_alloc(spa, size, private, type);
3153 hdr->b_dva = *BP_IDENTITY(bp);
3154 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3155 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3156 exists = buf_hash_insert(hdr, &hash_lock);
3158 /* somebody beat us to the hash insert */
3159 mutex_exit(hash_lock);
3160 buf_discard_identity(hdr);
3161 (void) arc_buf_remove_ref(buf, private);
3162 goto top; /* restart the IO request */
3164 /* if this is a prefetch, we don't have a reference */
3165 if (*arc_flags & ARC_PREFETCH) {
3166 (void) remove_reference(hdr, hash_lock,
3168 hdr->b_flags |= ARC_PREFETCH;
3170 if (*arc_flags & ARC_L2CACHE)
3171 hdr->b_flags |= ARC_L2CACHE;
3172 if (BP_GET_LEVEL(bp) > 0)
3173 hdr->b_flags |= ARC_INDIRECT;
3175 /* this block is in the ghost cache */
3176 ASSERT(GHOST_STATE(hdr->b_state));
3177 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3178 ASSERT0(refcount_count(&hdr->b_refcnt));
3179 ASSERT(hdr->b_buf == NULL);
3181 /* if this is a prefetch, we don't have a reference */
3182 if (*arc_flags & ARC_PREFETCH)
3183 hdr->b_flags |= ARC_PREFETCH;
3185 add_reference(hdr, hash_lock, private);
3186 if (*arc_flags & ARC_L2CACHE)
3187 hdr->b_flags |= ARC_L2CACHE;
3188 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3191 buf->b_efunc = NULL;
3192 buf->b_private = NULL;
3195 ASSERT(hdr->b_datacnt == 0);
3197 arc_get_data_buf(buf);
3198 arc_access(hdr, hash_lock);
3201 ASSERT(!GHOST_STATE(hdr->b_state));
3203 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
3204 acb->acb_done = done;
3205 acb->acb_private = private;
3207 ASSERT(hdr->b_acb == NULL);
3209 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3211 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3212 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3213 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3214 addr = hdr->b_l2hdr->b_daddr;
3216 * Lock out device removal.
3218 if (vdev_is_dead(vd) ||
3219 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3223 mutex_exit(hash_lock);
3226 * At this point, we have a level 1 cache miss. Try again in
3227 * L2ARC if possible.
3229 ASSERT3U(hdr->b_size, ==, size);
3230 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3231 uint64_t, size, zbookmark_t *, zb);
3232 ARCSTAT_BUMP(arcstat_misses);
3233 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3234 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3235 data, metadata, misses);
3237 curthread->td_ru.ru_inblock++;
3240 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3242 * Read from the L2ARC if the following are true:
3243 * 1. The L2ARC vdev was previously cached.
3244 * 2. This buffer still has L2ARC metadata.
3245 * 3. This buffer isn't currently writing to the L2ARC.
3246 * 4. The L2ARC entry wasn't evicted, which may
3247 * also have invalidated the vdev.
3248 * 5. This isn't prefetch and l2arc_noprefetch is set.
3250 if (hdr->b_l2hdr != NULL &&
3251 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3252 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3253 l2arc_read_callback_t *cb;
3255 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3256 ARCSTAT_BUMP(arcstat_l2_hits);
3258 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3260 cb->l2rcb_buf = buf;
3261 cb->l2rcb_spa = spa;
3264 cb->l2rcb_flags = zio_flags;
3266 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
3267 addr + size < vd->vdev_psize -
3268 VDEV_LABEL_END_SIZE);
3271 * l2arc read. The SCL_L2ARC lock will be
3272 * released by l2arc_read_done().
3274 rzio = zio_read_phys(pio, vd, addr, size,
3275 buf->b_data, ZIO_CHECKSUM_OFF,
3276 l2arc_read_done, cb, priority, zio_flags |
3277 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
3278 ZIO_FLAG_DONT_PROPAGATE |
3279 ZIO_FLAG_DONT_RETRY, B_FALSE);
3280 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3282 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
3284 if (*arc_flags & ARC_NOWAIT) {
3289 ASSERT(*arc_flags & ARC_WAIT);
3290 if (zio_wait(rzio) == 0)
3293 /* l2arc read error; goto zio_read() */
3295 DTRACE_PROBE1(l2arc__miss,
3296 arc_buf_hdr_t *, hdr);
3297 ARCSTAT_BUMP(arcstat_l2_misses);
3298 if (HDR_L2_WRITING(hdr))
3299 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3300 spa_config_exit(spa, SCL_L2ARC, vd);
3304 spa_config_exit(spa, SCL_L2ARC, vd);
3305 if (l2arc_ndev != 0) {
3306 DTRACE_PROBE1(l2arc__miss,
3307 arc_buf_hdr_t *, hdr);
3308 ARCSTAT_BUMP(arcstat_l2_misses);
3312 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3313 arc_read_done, buf, priority, zio_flags, zb);
3315 if (*arc_flags & ARC_WAIT)
3316 return (zio_wait(rzio));
3318 ASSERT(*arc_flags & ARC_NOWAIT);
3325 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3327 ASSERT(buf->b_hdr != NULL);
3328 ASSERT(buf->b_hdr->b_state != arc_anon);
3329 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3330 ASSERT(buf->b_efunc == NULL);
3331 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3333 buf->b_efunc = func;
3334 buf->b_private = private;
3338 * This is used by the DMU to let the ARC know that a buffer is
3339 * being evicted, so the ARC should clean up. If this arc buf
3340 * is not yet in the evicted state, it will be put there.
3343 arc_buf_evict(arc_buf_t *buf)
3346 kmutex_t *hash_lock;
3348 list_t *list, *evicted_list;
3349 kmutex_t *lock, *evicted_lock;
3351 mutex_enter(&buf->b_evict_lock);
3355 * We are in arc_do_user_evicts().
3357 ASSERT(buf->b_data == NULL);
3358 mutex_exit(&buf->b_evict_lock);
3360 } else if (buf->b_data == NULL) {
3361 arc_buf_t copy = *buf; /* structure assignment */
3363 * We are on the eviction list; process this buffer now
3364 * but let arc_do_user_evicts() do the reaping.
3366 buf->b_efunc = NULL;
3367 mutex_exit(&buf->b_evict_lock);
3368 VERIFY(copy.b_efunc(©) == 0);
3371 hash_lock = HDR_LOCK(hdr);
3372 mutex_enter(hash_lock);
3374 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3376 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3377 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3380 * Pull this buffer off of the hdr
3383 while (*bufp != buf)
3384 bufp = &(*bufp)->b_next;
3385 *bufp = buf->b_next;
3387 ASSERT(buf->b_data != NULL);
3388 arc_buf_destroy(buf, FALSE, FALSE);
3390 if (hdr->b_datacnt == 0) {
3391 arc_state_t *old_state = hdr->b_state;
3392 arc_state_t *evicted_state;
3394 ASSERT(hdr->b_buf == NULL);
3395 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3398 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3400 get_buf_info(hdr, old_state, &list, &lock);
3401 get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock);
3403 mutex_enter(evicted_lock);
3405 arc_change_state(evicted_state, hdr, hash_lock);
3406 ASSERT(HDR_IN_HASH_TABLE(hdr));
3407 hdr->b_flags |= ARC_IN_HASH_TABLE;
3408 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3410 mutex_exit(evicted_lock);
3413 mutex_exit(hash_lock);
3414 mutex_exit(&buf->b_evict_lock);
3416 VERIFY(buf->b_efunc(buf) == 0);
3417 buf->b_efunc = NULL;
3418 buf->b_private = NULL;
3421 kmem_cache_free(buf_cache, buf);
3426 * Release this buffer from the cache, making it an anonymous buffer. This
3427 * must be done after a read and prior to modifying the buffer contents.
3428 * If the buffer has more than one reference, we must make
3429 * a new hdr for the buffer.
3432 arc_release(arc_buf_t *buf, void *tag)
3435 kmutex_t *hash_lock = NULL;
3436 l2arc_buf_hdr_t *l2hdr;
3440 * It would be nice to assert that if it's DMU metadata (level >
3441 * 0 || it's the dnode file), then it must be syncing context.
3442 * But we don't know that information at this level.
3445 mutex_enter(&buf->b_evict_lock);
3448 /* this buffer is not on any list */
3449 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3451 if (hdr->b_state == arc_anon) {
3452 /* this buffer is already released */
3453 ASSERT(buf->b_efunc == NULL);
3455 hash_lock = HDR_LOCK(hdr);
3456 mutex_enter(hash_lock);
3458 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3461 l2hdr = hdr->b_l2hdr;
3463 mutex_enter(&l2arc_buflist_mtx);
3464 hdr->b_l2hdr = NULL;
3466 buf_size = hdr->b_size;
3469 * Do we have more than one buf?
3471 if (hdr->b_datacnt > 1) {
3472 arc_buf_hdr_t *nhdr;
3474 uint64_t blksz = hdr->b_size;
3475 uint64_t spa = hdr->b_spa;
3476 arc_buf_contents_t type = hdr->b_type;
3477 uint32_t flags = hdr->b_flags;
3479 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3481 * Pull the data off of this hdr and attach it to
3482 * a new anonymous hdr.
3484 (void) remove_reference(hdr, hash_lock, tag);
3486 while (*bufp != buf)
3487 bufp = &(*bufp)->b_next;
3488 *bufp = buf->b_next;
3491 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3492 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3493 if (refcount_is_zero(&hdr->b_refcnt)) {
3494 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3495 ASSERT3U(*size, >=, hdr->b_size);
3496 atomic_add_64(size, -hdr->b_size);
3500 * We're releasing a duplicate user data buffer, update
3501 * our statistics accordingly.
3503 if (hdr->b_type == ARC_BUFC_DATA) {
3504 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3505 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3508 hdr->b_datacnt -= 1;
3509 arc_cksum_verify(buf);
3511 arc_buf_unwatch(buf);
3512 #endif /* illumos */
3514 mutex_exit(hash_lock);
3516 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3517 nhdr->b_size = blksz;
3519 nhdr->b_type = type;
3521 nhdr->b_state = arc_anon;
3522 nhdr->b_arc_access = 0;
3523 nhdr->b_flags = flags & ARC_L2_WRITING;
3524 nhdr->b_l2hdr = NULL;
3525 nhdr->b_datacnt = 1;
3526 nhdr->b_freeze_cksum = NULL;
3527 (void) refcount_add(&nhdr->b_refcnt, tag);
3529 mutex_exit(&buf->b_evict_lock);
3530 atomic_add_64(&arc_anon->arcs_size, blksz);
3532 mutex_exit(&buf->b_evict_lock);
3533 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3534 ASSERT(!list_link_active(&hdr->b_arc_node));
3535 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3536 if (hdr->b_state != arc_anon)
3537 arc_change_state(arc_anon, hdr, hash_lock);
3538 hdr->b_arc_access = 0;
3540 mutex_exit(hash_lock);
3542 buf_discard_identity(hdr);
3545 buf->b_efunc = NULL;
3546 buf->b_private = NULL;
3549 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr,
3551 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3552 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3553 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3554 mutex_exit(&l2arc_buflist_mtx);
3559 arc_released(arc_buf_t *buf)
3563 mutex_enter(&buf->b_evict_lock);
3564 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3565 mutex_exit(&buf->b_evict_lock);
3570 arc_has_callback(arc_buf_t *buf)
3574 mutex_enter(&buf->b_evict_lock);
3575 callback = (buf->b_efunc != NULL);
3576 mutex_exit(&buf->b_evict_lock);
3582 arc_referenced(arc_buf_t *buf)
3586 mutex_enter(&buf->b_evict_lock);
3587 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3588 mutex_exit(&buf->b_evict_lock);
3589 return (referenced);
3594 arc_write_ready(zio_t *zio)
3596 arc_write_callback_t *callback = zio->io_private;
3597 arc_buf_t *buf = callback->awcb_buf;
3598 arc_buf_hdr_t *hdr = buf->b_hdr;
3600 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3601 callback->awcb_ready(zio, buf, callback->awcb_private);
3604 * If the IO is already in progress, then this is a re-write
3605 * attempt, so we need to thaw and re-compute the cksum.
3606 * It is the responsibility of the callback to handle the
3607 * accounting for any re-write attempt.
3609 if (HDR_IO_IN_PROGRESS(hdr)) {
3610 mutex_enter(&hdr->b_freeze_lock);
3611 if (hdr->b_freeze_cksum != NULL) {
3612 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3613 hdr->b_freeze_cksum = NULL;
3615 mutex_exit(&hdr->b_freeze_lock);
3617 arc_cksum_compute(buf, B_FALSE);
3618 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3622 arc_write_done(zio_t *zio)
3624 arc_write_callback_t *callback = zio->io_private;
3625 arc_buf_t *buf = callback->awcb_buf;
3626 arc_buf_hdr_t *hdr = buf->b_hdr;
3628 ASSERT(hdr->b_acb == NULL);
3630 if (zio->io_error == 0) {
3631 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3632 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3633 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3635 ASSERT(BUF_EMPTY(hdr));
3639 * If the block to be written was all-zero, we may have
3640 * compressed it away. In this case no write was performed
3641 * so there will be no dva/birth/checksum. The buffer must
3642 * therefore remain anonymous (and uncached).
3644 if (!BUF_EMPTY(hdr)) {
3645 arc_buf_hdr_t *exists;
3646 kmutex_t *hash_lock;
3648 ASSERT(zio->io_error == 0);
3650 arc_cksum_verify(buf);
3652 exists = buf_hash_insert(hdr, &hash_lock);
3655 * This can only happen if we overwrite for
3656 * sync-to-convergence, because we remove
3657 * buffers from the hash table when we arc_free().
3659 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3660 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3661 panic("bad overwrite, hdr=%p exists=%p",
3662 (void *)hdr, (void *)exists);
3663 ASSERT(refcount_is_zero(&exists->b_refcnt));
3664 arc_change_state(arc_anon, exists, hash_lock);
3665 mutex_exit(hash_lock);
3666 arc_hdr_destroy(exists);
3667 exists = buf_hash_insert(hdr, &hash_lock);
3668 ASSERT3P(exists, ==, NULL);
3669 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3671 ASSERT(zio->io_prop.zp_nopwrite);
3672 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3673 panic("bad nopwrite, hdr=%p exists=%p",
3674 (void *)hdr, (void *)exists);
3677 ASSERT(hdr->b_datacnt == 1);
3678 ASSERT(hdr->b_state == arc_anon);
3679 ASSERT(BP_GET_DEDUP(zio->io_bp));
3680 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3683 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3684 /* if it's not anon, we are doing a scrub */
3685 if (!exists && hdr->b_state == arc_anon)
3686 arc_access(hdr, hash_lock);
3687 mutex_exit(hash_lock);
3689 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3692 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3693 callback->awcb_done(zio, buf, callback->awcb_private);
3695 kmem_free(callback, sizeof (arc_write_callback_t));
3699 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3700 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3701 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3702 int priority, int zio_flags, const zbookmark_t *zb)
3704 arc_buf_hdr_t *hdr = buf->b_hdr;
3705 arc_write_callback_t *callback;
3708 ASSERT(ready != NULL);
3709 ASSERT(done != NULL);
3710 ASSERT(!HDR_IO_ERROR(hdr));
3711 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3712 ASSERT(hdr->b_acb == NULL);
3714 hdr->b_flags |= ARC_L2CACHE;
3715 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3716 callback->awcb_ready = ready;
3717 callback->awcb_done = done;
3718 callback->awcb_private = private;
3719 callback->awcb_buf = buf;
3721 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3722 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3728 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3731 uint64_t available_memory =
3732 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count);
3733 static uint64_t page_load = 0;
3734 static uint64_t last_txg = 0;
3739 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3742 if (available_memory >= zfs_write_limit_max)
3745 if (txg > last_txg) {
3750 * If we are in pageout, we know that memory is already tight,
3751 * the arc is already going to be evicting, so we just want to
3752 * continue to let page writes occur as quickly as possible.
3754 if (curproc == pageproc) {
3755 if (page_load > available_memory / 4)
3756 return (SET_ERROR(ERESTART));
3757 /* Note: reserve is inflated, so we deflate */
3758 page_load += reserve / 8;
3760 } else if (page_load > 0 && arc_reclaim_needed()) {
3761 /* memory is low, delay before restarting */
3762 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3763 return (SET_ERROR(EAGAIN));
3767 if (arc_size > arc_c_min) {
3768 uint64_t evictable_memory =
3769 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3770 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3771 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3772 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3773 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3776 if (inflight_data > available_memory / 4) {
3777 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3778 return (SET_ERROR(ERESTART));
3785 arc_tempreserve_clear(uint64_t reserve)
3787 atomic_add_64(&arc_tempreserve, -reserve);
3788 ASSERT((int64_t)arc_tempreserve >= 0);
3792 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3799 * Once in a while, fail for no reason. Everything should cope.
3801 if (spa_get_random(10000) == 0) {
3802 dprintf("forcing random failure\n");
3803 return (SET_ERROR(ERESTART));
3806 if (reserve > arc_c/4 && !arc_no_grow)
3807 arc_c = MIN(arc_c_max, reserve * 4);
3808 if (reserve > arc_c)
3809 return (SET_ERROR(ENOMEM));
3812 * Don't count loaned bufs as in flight dirty data to prevent long
3813 * network delays from blocking transactions that are ready to be
3814 * assigned to a txg.
3816 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3819 * Writes will, almost always, require additional memory allocations
3820 * in order to compress/encrypt/etc the data. We therefor need to
3821 * make sure that there is sufficient available memory for this.
3823 if (error = arc_memory_throttle(reserve, anon_size, txg))
3827 * Throttle writes when the amount of dirty data in the cache
3828 * gets too large. We try to keep the cache less than half full
3829 * of dirty blocks so that our sync times don't grow too large.
3830 * Note: if two requests come in concurrently, we might let them
3831 * both succeed, when one of them should fail. Not a huge deal.
3834 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3835 anon_size > arc_c / 4) {
3836 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3837 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3838 arc_tempreserve>>10,
3839 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3840 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3841 reserve>>10, arc_c>>10);
3842 return (SET_ERROR(ERESTART));
3844 atomic_add_64(&arc_tempreserve, reserve);
3848 static kmutex_t arc_lowmem_lock;
3850 static eventhandler_tag arc_event_lowmem = NULL;
3853 arc_lowmem(void *arg __unused, int howto __unused)
3856 /* Serialize access via arc_lowmem_lock. */
3857 mutex_enter(&arc_lowmem_lock);
3858 mutex_enter(&arc_reclaim_thr_lock);
3860 cv_signal(&arc_reclaim_thr_cv);
3863 * It is unsafe to block here in arbitrary threads, because we can come
3864 * here from ARC itself and may hold ARC locks and thus risk a deadlock
3865 * with ARC reclaim thread.
3867 if (curproc == pageproc) {
3869 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0);
3871 mutex_exit(&arc_reclaim_thr_lock);
3872 mutex_exit(&arc_lowmem_lock);
3879 int i, prefetch_tunable_set = 0;
3881 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3882 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3883 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL);
3885 /* Convert seconds to clock ticks */
3886 arc_min_prefetch_lifespan = 1 * hz;
3888 /* Start out with 1/8 of all memory */
3889 arc_c = kmem_size() / 8;
3894 * On architectures where the physical memory can be larger
3895 * than the addressable space (intel in 32-bit mode), we may
3896 * need to limit the cache to 1/8 of VM size.
3898 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3901 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
3902 arc_c_min = MAX(arc_c / 4, 64<<18);
3903 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
3904 if (arc_c * 8 >= 1<<30)
3905 arc_c_max = (arc_c * 8) - (1<<30);
3907 arc_c_max = arc_c_min;
3908 arc_c_max = MAX(arc_c * 5, arc_c_max);
3912 * Allow the tunables to override our calculations if they are
3913 * reasonable (ie. over 16MB)
3915 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size())
3916 arc_c_max = zfs_arc_max;
3917 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max)
3918 arc_c_min = zfs_arc_min;
3922 arc_p = (arc_c >> 1);
3924 /* limit meta-data to 1/4 of the arc capacity */
3925 arc_meta_limit = arc_c_max / 4;
3927 /* Allow the tunable to override if it is reasonable */
3928 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3929 arc_meta_limit = zfs_arc_meta_limit;
3931 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3932 arc_c_min = arc_meta_limit / 2;
3934 if (zfs_arc_grow_retry > 0)
3935 arc_grow_retry = zfs_arc_grow_retry;
3937 if (zfs_arc_shrink_shift > 0)
3938 arc_shrink_shift = zfs_arc_shrink_shift;
3940 if (zfs_arc_p_min_shift > 0)
3941 arc_p_min_shift = zfs_arc_p_min_shift;
3943 /* if kmem_flags are set, lets try to use less memory */
3944 if (kmem_debugging())
3946 if (arc_c < arc_c_min)
3949 zfs_arc_min = arc_c_min;
3950 zfs_arc_max = arc_c_max;
3952 arc_anon = &ARC_anon;
3954 arc_mru_ghost = &ARC_mru_ghost;
3956 arc_mfu_ghost = &ARC_mfu_ghost;
3957 arc_l2c_only = &ARC_l2c_only;
3960 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
3961 mutex_init(&arc_anon->arcs_locks[i].arcs_lock,
3962 NULL, MUTEX_DEFAULT, NULL);
3963 mutex_init(&arc_mru->arcs_locks[i].arcs_lock,
3964 NULL, MUTEX_DEFAULT, NULL);
3965 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock,
3966 NULL, MUTEX_DEFAULT, NULL);
3967 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock,
3968 NULL, MUTEX_DEFAULT, NULL);
3969 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock,
3970 NULL, MUTEX_DEFAULT, NULL);
3971 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock,
3972 NULL, MUTEX_DEFAULT, NULL);
3974 list_create(&arc_mru->arcs_lists[i],
3975 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3976 list_create(&arc_mru_ghost->arcs_lists[i],
3977 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3978 list_create(&arc_mfu->arcs_lists[i],
3979 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3980 list_create(&arc_mfu_ghost->arcs_lists[i],
3981 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3982 list_create(&arc_mfu_ghost->arcs_lists[i],
3983 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3984 list_create(&arc_l2c_only->arcs_lists[i],
3985 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3990 arc_thread_exit = 0;
3991 arc_eviction_list = NULL;
3992 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3993 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3995 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3996 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3998 if (arc_ksp != NULL) {
3999 arc_ksp->ks_data = &arc_stats;
4000 kstat_install(arc_ksp);
4003 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
4004 TS_RUN, minclsyspri);
4007 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
4008 EVENTHANDLER_PRI_FIRST);
4014 if (zfs_write_limit_max == 0)
4015 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
4017 zfs_write_limit_shift = 0;
4018 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
4021 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
4022 prefetch_tunable_set = 1;
4025 if (prefetch_tunable_set == 0) {
4026 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
4028 printf(" add \"vfs.zfs.prefetch_disable=0\" "
4029 "to /boot/loader.conf.\n");
4030 zfs_prefetch_disable = 1;
4033 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
4034 prefetch_tunable_set == 0) {
4035 printf("ZFS NOTICE: Prefetch is disabled by default if less "
4036 "than 4GB of RAM is present;\n"
4037 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
4038 "to /boot/loader.conf.\n");
4039 zfs_prefetch_disable = 1;
4042 /* Warn about ZFS memory and address space requirements. */
4043 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
4044 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
4045 "expect unstable behavior.\n");
4047 if (kmem_size() < 512 * (1 << 20)) {
4048 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
4049 "expect unstable behavior.\n");
4050 printf(" Consider tuning vm.kmem_size and "
4051 "vm.kmem_size_max\n");
4052 printf(" in /boot/loader.conf.\n");
4062 mutex_enter(&arc_reclaim_thr_lock);
4063 arc_thread_exit = 1;
4064 cv_signal(&arc_reclaim_thr_cv);
4065 while (arc_thread_exit != 0)
4066 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
4067 mutex_exit(&arc_reclaim_thr_lock);
4073 if (arc_ksp != NULL) {
4074 kstat_delete(arc_ksp);
4078 mutex_destroy(&arc_eviction_mtx);
4079 mutex_destroy(&arc_reclaim_thr_lock);
4080 cv_destroy(&arc_reclaim_thr_cv);
4082 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
4083 list_destroy(&arc_mru->arcs_lists[i]);
4084 list_destroy(&arc_mru_ghost->arcs_lists[i]);
4085 list_destroy(&arc_mfu->arcs_lists[i]);
4086 list_destroy(&arc_mfu_ghost->arcs_lists[i]);
4087 list_destroy(&arc_l2c_only->arcs_lists[i]);
4089 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock);
4090 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock);
4091 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock);
4092 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock);
4093 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock);
4094 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock);
4097 mutex_destroy(&zfs_write_limit_lock);
4101 ASSERT(arc_loaned_bytes == 0);
4103 mutex_destroy(&arc_lowmem_lock);
4105 if (arc_event_lowmem != NULL)
4106 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
4113 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
4114 * It uses dedicated storage devices to hold cached data, which are populated
4115 * using large infrequent writes. The main role of this cache is to boost
4116 * the performance of random read workloads. The intended L2ARC devices
4117 * include short-stroked disks, solid state disks, and other media with
4118 * substantially faster read latency than disk.
4120 * +-----------------------+
4122 * +-----------------------+
4125 * l2arc_feed_thread() arc_read()
4129 * +---------------+ |
4131 * +---------------+ |
4136 * +-------+ +-------+
4138 * | cache | | cache |
4139 * +-------+ +-------+
4140 * +=========+ .-----.
4141 * : L2ARC : |-_____-|
4142 * : devices : | Disks |
4143 * +=========+ `-_____-'
4145 * Read requests are satisfied from the following sources, in order:
4148 * 2) vdev cache of L2ARC devices
4150 * 4) vdev cache of disks
4153 * Some L2ARC device types exhibit extremely slow write performance.
4154 * To accommodate for this there are some significant differences between
4155 * the L2ARC and traditional cache design:
4157 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4158 * the ARC behave as usual, freeing buffers and placing headers on ghost
4159 * lists. The ARC does not send buffers to the L2ARC during eviction as
4160 * this would add inflated write latencies for all ARC memory pressure.
4162 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4163 * It does this by periodically scanning buffers from the eviction-end of
4164 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4165 * not already there. It scans until a headroom of buffers is satisfied,
4166 * which itself is a buffer for ARC eviction. The thread that does this is
4167 * l2arc_feed_thread(), illustrated below; example sizes are included to
4168 * provide a better sense of ratio than this diagram:
4171 * +---------------------+----------+
4172 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4173 * +---------------------+----------+ | o L2ARC eligible
4174 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4175 * +---------------------+----------+ |
4176 * 15.9 Gbytes ^ 32 Mbytes |
4178 * l2arc_feed_thread()
4180 * l2arc write hand <--[oooo]--'
4184 * +==============================+
4185 * L2ARC dev |####|#|###|###| |####| ... |
4186 * +==============================+
4189 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4190 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4191 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4192 * safe to say that this is an uncommon case, since buffers at the end of
4193 * the ARC lists have moved there due to inactivity.
4195 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4196 * then the L2ARC simply misses copying some buffers. This serves as a
4197 * pressure valve to prevent heavy read workloads from both stalling the ARC
4198 * with waits and clogging the L2ARC with writes. This also helps prevent
4199 * the potential for the L2ARC to churn if it attempts to cache content too
4200 * quickly, such as during backups of the entire pool.
4202 * 5. After system boot and before the ARC has filled main memory, there are
4203 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4204 * lists can remain mostly static. Instead of searching from tail of these
4205 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4206 * for eligible buffers, greatly increasing its chance of finding them.
4208 * The L2ARC device write speed is also boosted during this time so that
4209 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4210 * there are no L2ARC reads, and no fear of degrading read performance
4211 * through increased writes.
4213 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4214 * the vdev queue can aggregate them into larger and fewer writes. Each
4215 * device is written to in a rotor fashion, sweeping writes through
4216 * available space then repeating.
4218 * 7. The L2ARC does not store dirty content. It never needs to flush
4219 * write buffers back to disk based storage.
4221 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4222 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4224 * The performance of the L2ARC can be tweaked by a number of tunables, which
4225 * may be necessary for different workloads:
4227 * l2arc_write_max max write bytes per interval
4228 * l2arc_write_boost extra write bytes during device warmup
4229 * l2arc_noprefetch skip caching prefetched buffers
4230 * l2arc_headroom number of max device writes to precache
4231 * l2arc_feed_secs seconds between L2ARC writing
4233 * Tunables may be removed or added as future performance improvements are
4234 * integrated, and also may become zpool properties.
4236 * There are three key functions that control how the L2ARC warms up:
4238 * l2arc_write_eligible() check if a buffer is eligible to cache
4239 * l2arc_write_size() calculate how much to write
4240 * l2arc_write_interval() calculate sleep delay between writes
4242 * These three functions determine what to write, how much, and how quickly
4247 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4250 * A buffer is *not* eligible for the L2ARC if it:
4251 * 1. belongs to a different spa.
4252 * 2. is already cached on the L2ARC.
4253 * 3. has an I/O in progress (it may be an incomplete read).
4254 * 4. is flagged not eligible (zfs property).
4256 if (ab->b_spa != spa_guid) {
4257 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
4260 if (ab->b_l2hdr != NULL) {
4261 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
4264 if (HDR_IO_IN_PROGRESS(ab)) {
4265 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
4268 if (!HDR_L2CACHE(ab)) {
4269 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
4277 l2arc_write_size(l2arc_dev_t *dev)
4281 size = dev->l2ad_write;
4283 if (arc_warm == B_FALSE)
4284 size += dev->l2ad_boost;
4291 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4293 clock_t interval, next, now;
4296 * If the ARC lists are busy, increase our write rate; if the
4297 * lists are stale, idle back. This is achieved by checking
4298 * how much we previously wrote - if it was more than half of
4299 * what we wanted, schedule the next write much sooner.
4301 if (l2arc_feed_again && wrote > (wanted / 2))
4302 interval = (hz * l2arc_feed_min_ms) / 1000;
4304 interval = hz * l2arc_feed_secs;
4306 now = ddi_get_lbolt();
4307 next = MAX(now, MIN(now + interval, began + interval));
4313 l2arc_hdr_stat_add(void)
4315 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4316 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4320 l2arc_hdr_stat_remove(void)
4322 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4323 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4327 * Cycle through L2ARC devices. This is how L2ARC load balances.
4328 * If a device is returned, this also returns holding the spa config lock.
4330 static l2arc_dev_t *
4331 l2arc_dev_get_next(void)
4333 l2arc_dev_t *first, *next = NULL;
4336 * Lock out the removal of spas (spa_namespace_lock), then removal
4337 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4338 * both locks will be dropped and a spa config lock held instead.
4340 mutex_enter(&spa_namespace_lock);
4341 mutex_enter(&l2arc_dev_mtx);
4343 /* if there are no vdevs, there is nothing to do */
4344 if (l2arc_ndev == 0)
4348 next = l2arc_dev_last;
4350 /* loop around the list looking for a non-faulted vdev */
4352 next = list_head(l2arc_dev_list);
4354 next = list_next(l2arc_dev_list, next);
4356 next = list_head(l2arc_dev_list);
4359 /* if we have come back to the start, bail out */
4362 else if (next == first)
4365 } while (vdev_is_dead(next->l2ad_vdev));
4367 /* if we were unable to find any usable vdevs, return NULL */
4368 if (vdev_is_dead(next->l2ad_vdev))
4371 l2arc_dev_last = next;
4374 mutex_exit(&l2arc_dev_mtx);
4377 * Grab the config lock to prevent the 'next' device from being
4378 * removed while we are writing to it.
4381 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4382 mutex_exit(&spa_namespace_lock);
4388 * Free buffers that were tagged for destruction.
4391 l2arc_do_free_on_write()
4394 l2arc_data_free_t *df, *df_prev;
4396 mutex_enter(&l2arc_free_on_write_mtx);
4397 buflist = l2arc_free_on_write;
4399 for (df = list_tail(buflist); df; df = df_prev) {
4400 df_prev = list_prev(buflist, df);
4401 ASSERT(df->l2df_data != NULL);
4402 ASSERT(df->l2df_func != NULL);
4403 df->l2df_func(df->l2df_data, df->l2df_size);
4404 list_remove(buflist, df);
4405 kmem_free(df, sizeof (l2arc_data_free_t));
4408 mutex_exit(&l2arc_free_on_write_mtx);
4412 * A write to a cache device has completed. Update all headers to allow
4413 * reads from these buffers to begin.
4416 l2arc_write_done(zio_t *zio)
4418 l2arc_write_callback_t *cb;
4421 arc_buf_hdr_t *head, *ab, *ab_prev;
4422 l2arc_buf_hdr_t *abl2;
4423 kmutex_t *hash_lock;
4425 cb = zio->io_private;
4427 dev = cb->l2wcb_dev;
4428 ASSERT(dev != NULL);
4429 head = cb->l2wcb_head;
4430 ASSERT(head != NULL);
4431 buflist = dev->l2ad_buflist;
4432 ASSERT(buflist != NULL);
4433 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4434 l2arc_write_callback_t *, cb);
4436 if (zio->io_error != 0)
4437 ARCSTAT_BUMP(arcstat_l2_writes_error);
4439 mutex_enter(&l2arc_buflist_mtx);
4442 * All writes completed, or an error was hit.
4444 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4445 ab_prev = list_prev(buflist, ab);
4447 hash_lock = HDR_LOCK(ab);
4448 if (!mutex_tryenter(hash_lock)) {
4450 * This buffer misses out. It may be in a stage
4451 * of eviction. Its ARC_L2_WRITING flag will be
4452 * left set, denying reads to this buffer.
4454 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4458 if (zio->io_error != 0) {
4460 * Error - drop L2ARC entry.
4462 list_remove(buflist, ab);
4465 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr,
4467 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4468 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4472 * Allow ARC to begin reads to this L2ARC entry.
4474 ab->b_flags &= ~ARC_L2_WRITING;
4476 mutex_exit(hash_lock);
4479 atomic_inc_64(&l2arc_writes_done);
4480 list_remove(buflist, head);
4481 kmem_cache_free(hdr_cache, head);
4482 mutex_exit(&l2arc_buflist_mtx);
4484 l2arc_do_free_on_write();
4486 kmem_free(cb, sizeof (l2arc_write_callback_t));
4490 * A read to a cache device completed. Validate buffer contents before
4491 * handing over to the regular ARC routines.
4494 l2arc_read_done(zio_t *zio)
4496 l2arc_read_callback_t *cb;
4499 kmutex_t *hash_lock;
4502 ASSERT(zio->io_vd != NULL);
4503 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4505 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4507 cb = zio->io_private;
4509 buf = cb->l2rcb_buf;
4510 ASSERT(buf != NULL);
4512 hash_lock = HDR_LOCK(buf->b_hdr);
4513 mutex_enter(hash_lock);
4515 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4518 * Check this survived the L2ARC journey.
4520 equal = arc_cksum_equal(buf);
4521 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4522 mutex_exit(hash_lock);
4523 zio->io_private = buf;
4524 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4525 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4528 mutex_exit(hash_lock);
4530 * Buffer didn't survive caching. Increment stats and
4531 * reissue to the original storage device.
4533 if (zio->io_error != 0) {
4534 ARCSTAT_BUMP(arcstat_l2_io_error);
4536 zio->io_error = SET_ERROR(EIO);
4539 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4542 * If there's no waiter, issue an async i/o to the primary
4543 * storage now. If there *is* a waiter, the caller must
4544 * issue the i/o in a context where it's OK to block.
4546 if (zio->io_waiter == NULL) {
4547 zio_t *pio = zio_unique_parent(zio);
4549 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4551 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4552 buf->b_data, zio->io_size, arc_read_done, buf,
4553 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4557 kmem_free(cb, sizeof (l2arc_read_callback_t));
4561 * This is the list priority from which the L2ARC will search for pages to
4562 * cache. This is used within loops (0..3) to cycle through lists in the
4563 * desired order. This order can have a significant effect on cache
4566 * Currently the metadata lists are hit first, MFU then MRU, followed by
4567 * the data lists. This function returns a locked list, and also returns
4571 l2arc_list_locked(int list_num, kmutex_t **lock)
4573 list_t *list = NULL;
4576 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS);
4578 if (list_num < ARC_BUFC_NUMMETADATALISTS) {
4580 list = &arc_mfu->arcs_lists[idx];
4581 *lock = ARCS_LOCK(arc_mfu, idx);
4582 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) {
4583 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4584 list = &arc_mru->arcs_lists[idx];
4585 *lock = ARCS_LOCK(arc_mru, idx);
4586 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 +
4587 ARC_BUFC_NUMDATALISTS)) {
4588 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4589 list = &arc_mfu->arcs_lists[idx];
4590 *lock = ARCS_LOCK(arc_mfu, idx);
4592 idx = list_num - ARC_BUFC_NUMLISTS;
4593 list = &arc_mru->arcs_lists[idx];
4594 *lock = ARCS_LOCK(arc_mru, idx);
4597 ASSERT(!(MUTEX_HELD(*lock)));
4603 * Evict buffers from the device write hand to the distance specified in
4604 * bytes. This distance may span populated buffers, it may span nothing.
4605 * This is clearing a region on the L2ARC device ready for writing.
4606 * If the 'all' boolean is set, every buffer is evicted.
4609 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4612 l2arc_buf_hdr_t *abl2;
4613 arc_buf_hdr_t *ab, *ab_prev;
4614 kmutex_t *hash_lock;
4617 buflist = dev->l2ad_buflist;
4619 if (buflist == NULL)
4622 if (!all && dev->l2ad_first) {
4624 * This is the first sweep through the device. There is
4630 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4632 * When nearing the end of the device, evict to the end
4633 * before the device write hand jumps to the start.
4635 taddr = dev->l2ad_end;
4637 taddr = dev->l2ad_hand + distance;
4639 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4640 uint64_t, taddr, boolean_t, all);
4643 mutex_enter(&l2arc_buflist_mtx);
4644 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4645 ab_prev = list_prev(buflist, ab);
4647 hash_lock = HDR_LOCK(ab);
4648 if (!mutex_tryenter(hash_lock)) {
4650 * Missed the hash lock. Retry.
4652 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4653 mutex_exit(&l2arc_buflist_mtx);
4654 mutex_enter(hash_lock);
4655 mutex_exit(hash_lock);
4659 if (HDR_L2_WRITE_HEAD(ab)) {
4661 * We hit a write head node. Leave it for
4662 * l2arc_write_done().
4664 list_remove(buflist, ab);
4665 mutex_exit(hash_lock);
4669 if (!all && ab->b_l2hdr != NULL &&
4670 (ab->b_l2hdr->b_daddr > taddr ||
4671 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4673 * We've evicted to the target address,
4674 * or the end of the device.
4676 mutex_exit(hash_lock);
4680 if (HDR_FREE_IN_PROGRESS(ab)) {
4682 * Already on the path to destruction.
4684 mutex_exit(hash_lock);
4688 if (ab->b_state == arc_l2c_only) {
4689 ASSERT(!HDR_L2_READING(ab));
4691 * This doesn't exist in the ARC. Destroy.
4692 * arc_hdr_destroy() will call list_remove()
4693 * and decrement arcstat_l2_size.
4695 arc_change_state(arc_anon, ab, hash_lock);
4696 arc_hdr_destroy(ab);
4699 * Invalidate issued or about to be issued
4700 * reads, since we may be about to write
4701 * over this location.
4703 if (HDR_L2_READING(ab)) {
4704 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4705 ab->b_flags |= ARC_L2_EVICTED;
4709 * Tell ARC this no longer exists in L2ARC.
4711 if (ab->b_l2hdr != NULL) {
4714 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4715 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4717 list_remove(buflist, ab);
4720 * This may have been leftover after a
4723 ab->b_flags &= ~ARC_L2_WRITING;
4725 mutex_exit(hash_lock);
4727 mutex_exit(&l2arc_buflist_mtx);
4729 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4730 dev->l2ad_evict = taddr;
4734 * Find and write ARC buffers to the L2ARC device.
4736 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4737 * for reading until they have completed writing.
4740 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4742 arc_buf_hdr_t *ab, *ab_prev, *head;
4743 l2arc_buf_hdr_t *hdrl2;
4745 uint64_t passed_sz, write_sz, buf_sz, headroom;
4747 kmutex_t *hash_lock, *list_lock;
4748 boolean_t have_lock, full;
4749 l2arc_write_callback_t *cb;
4751 uint64_t guid = spa_load_guid(spa);
4754 ASSERT(dev->l2ad_vdev != NULL);
4759 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4760 head->b_flags |= ARC_L2_WRITE_HEAD;
4762 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
4764 * Copy buffers for L2ARC writing.
4766 mutex_enter(&l2arc_buflist_mtx);
4767 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) {
4768 list = l2arc_list_locked(try, &list_lock);
4770 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
4773 * L2ARC fast warmup.
4775 * Until the ARC is warm and starts to evict, read from the
4776 * head of the ARC lists rather than the tail.
4778 headroom = target_sz * l2arc_headroom;
4779 if (arc_warm == B_FALSE)
4780 ab = list_head(list);
4782 ab = list_tail(list);
4784 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
4786 for (; ab; ab = ab_prev) {
4787 if (arc_warm == B_FALSE)
4788 ab_prev = list_next(list, ab);
4790 ab_prev = list_prev(list, ab);
4791 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size);
4793 hash_lock = HDR_LOCK(ab);
4794 have_lock = MUTEX_HELD(hash_lock);
4795 if (!have_lock && !mutex_tryenter(hash_lock)) {
4796 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
4798 * Skip this buffer rather than waiting.
4803 passed_sz += ab->b_size;
4804 if (passed_sz > headroom) {
4808 mutex_exit(hash_lock);
4809 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
4813 if (!l2arc_write_eligible(guid, ab)) {
4814 mutex_exit(hash_lock);
4818 if ((write_sz + ab->b_size) > target_sz) {
4820 mutex_exit(hash_lock);
4821 ARCSTAT_BUMP(arcstat_l2_write_full);
4827 * Insert a dummy header on the buflist so
4828 * l2arc_write_done() can find where the
4829 * write buffers begin without searching.
4831 list_insert_head(dev->l2ad_buflist, head);
4834 sizeof (l2arc_write_callback_t), KM_SLEEP);
4835 cb->l2wcb_dev = dev;
4836 cb->l2wcb_head = head;
4837 pio = zio_root(spa, l2arc_write_done, cb,
4839 ARCSTAT_BUMP(arcstat_l2_write_pios);
4843 * Create and add a new L2ARC header.
4845 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4847 hdrl2->b_daddr = dev->l2ad_hand;
4849 ab->b_flags |= ARC_L2_WRITING;
4850 ab->b_l2hdr = hdrl2;
4851 list_insert_head(dev->l2ad_buflist, ab);
4852 buf_data = ab->b_buf->b_data;
4853 buf_sz = ab->b_size;
4856 * Compute and store the buffer cksum before
4857 * writing. On debug the cksum is verified first.
4859 arc_cksum_verify(ab->b_buf);
4860 arc_cksum_compute(ab->b_buf, B_TRUE);
4862 mutex_exit(hash_lock);
4864 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4865 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4866 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4867 ZIO_FLAG_CANFAIL, B_FALSE);
4869 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4871 (void) zio_nowait(wzio);
4874 * Keep the clock hand suitably device-aligned.
4876 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4879 dev->l2ad_hand += buf_sz;
4882 mutex_exit(list_lock);
4887 mutex_exit(&l2arc_buflist_mtx);
4891 kmem_cache_free(hdr_cache, head);
4895 ASSERT3U(write_sz, <=, target_sz);
4896 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4897 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4898 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4899 vdev_space_update(dev->l2ad_vdev, write_sz, 0, 0);
4902 * Bump device hand to the device start if it is approaching the end.
4903 * l2arc_evict() will already have evicted ahead for this case.
4905 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4906 vdev_space_update(dev->l2ad_vdev,
4907 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4908 dev->l2ad_hand = dev->l2ad_start;
4909 dev->l2ad_evict = dev->l2ad_start;
4910 dev->l2ad_first = B_FALSE;
4913 dev->l2ad_writing = B_TRUE;
4914 (void) zio_wait(pio);
4915 dev->l2ad_writing = B_FALSE;
4921 * This thread feeds the L2ARC at regular intervals. This is the beating
4922 * heart of the L2ARC.
4925 l2arc_feed_thread(void *dummy __unused)
4930 uint64_t size, wrote;
4931 clock_t begin, next = ddi_get_lbolt();
4933 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4935 mutex_enter(&l2arc_feed_thr_lock);
4937 while (l2arc_thread_exit == 0) {
4938 CALLB_CPR_SAFE_BEGIN(&cpr);
4939 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4940 next - ddi_get_lbolt());
4941 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4942 next = ddi_get_lbolt() + hz;
4945 * Quick check for L2ARC devices.
4947 mutex_enter(&l2arc_dev_mtx);
4948 if (l2arc_ndev == 0) {
4949 mutex_exit(&l2arc_dev_mtx);
4952 mutex_exit(&l2arc_dev_mtx);
4953 begin = ddi_get_lbolt();
4956 * This selects the next l2arc device to write to, and in
4957 * doing so the next spa to feed from: dev->l2ad_spa. This
4958 * will return NULL if there are now no l2arc devices or if
4959 * they are all faulted.
4961 * If a device is returned, its spa's config lock is also
4962 * held to prevent device removal. l2arc_dev_get_next()
4963 * will grab and release l2arc_dev_mtx.
4965 if ((dev = l2arc_dev_get_next()) == NULL)
4968 spa = dev->l2ad_spa;
4969 ASSERT(spa != NULL);
4972 * If the pool is read-only then force the feed thread to
4973 * sleep a little longer.
4975 if (!spa_writeable(spa)) {
4976 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4977 spa_config_exit(spa, SCL_L2ARC, dev);
4982 * Avoid contributing to memory pressure.
4984 if (arc_reclaim_needed()) {
4985 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4986 spa_config_exit(spa, SCL_L2ARC, dev);
4990 ARCSTAT_BUMP(arcstat_l2_feeds);
4992 size = l2arc_write_size(dev);
4995 * Evict L2ARC buffers that will be overwritten.
4997 l2arc_evict(dev, size, B_FALSE);
5000 * Write ARC buffers.
5002 wrote = l2arc_write_buffers(spa, dev, size);
5005 * Calculate interval between writes.
5007 next = l2arc_write_interval(begin, size, wrote);
5008 spa_config_exit(spa, SCL_L2ARC, dev);
5011 l2arc_thread_exit = 0;
5012 cv_broadcast(&l2arc_feed_thr_cv);
5013 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
5018 l2arc_vdev_present(vdev_t *vd)
5022 mutex_enter(&l2arc_dev_mtx);
5023 for (dev = list_head(l2arc_dev_list); dev != NULL;
5024 dev = list_next(l2arc_dev_list, dev)) {
5025 if (dev->l2ad_vdev == vd)
5028 mutex_exit(&l2arc_dev_mtx);
5030 return (dev != NULL);
5034 * Add a vdev for use by the L2ARC. By this point the spa has already
5035 * validated the vdev and opened it.
5038 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5040 l2arc_dev_t *adddev;
5042 ASSERT(!l2arc_vdev_present(vd));
5045 * Create a new l2arc device entry.
5047 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5048 adddev->l2ad_spa = spa;
5049 adddev->l2ad_vdev = vd;
5050 adddev->l2ad_write = l2arc_write_max;
5051 adddev->l2ad_boost = l2arc_write_boost;
5052 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5053 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5054 adddev->l2ad_hand = adddev->l2ad_start;
5055 adddev->l2ad_evict = adddev->l2ad_start;
5056 adddev->l2ad_first = B_TRUE;
5057 adddev->l2ad_writing = B_FALSE;
5058 ASSERT3U(adddev->l2ad_write, >, 0);
5061 * This is a list of all ARC buffers that are still valid on the
5064 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5065 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5066 offsetof(arc_buf_hdr_t, b_l2node));
5068 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5071 * Add device to global list
5073 mutex_enter(&l2arc_dev_mtx);
5074 list_insert_head(l2arc_dev_list, adddev);
5075 atomic_inc_64(&l2arc_ndev);
5076 mutex_exit(&l2arc_dev_mtx);
5080 * Remove a vdev from the L2ARC.
5083 l2arc_remove_vdev(vdev_t *vd)
5085 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5088 * Find the device by vdev
5090 mutex_enter(&l2arc_dev_mtx);
5091 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5092 nextdev = list_next(l2arc_dev_list, dev);
5093 if (vd == dev->l2ad_vdev) {
5098 ASSERT(remdev != NULL);
5101 * Remove device from global list
5103 list_remove(l2arc_dev_list, remdev);
5104 l2arc_dev_last = NULL; /* may have been invalidated */
5105 atomic_dec_64(&l2arc_ndev);
5106 mutex_exit(&l2arc_dev_mtx);
5109 * Clear all buflists and ARC references. L2ARC device flush.
5111 l2arc_evict(remdev, 0, B_TRUE);
5112 list_destroy(remdev->l2ad_buflist);
5113 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5114 kmem_free(remdev, sizeof (l2arc_dev_t));
5120 l2arc_thread_exit = 0;
5122 l2arc_writes_sent = 0;
5123 l2arc_writes_done = 0;
5125 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5126 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5127 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5128 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5129 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5131 l2arc_dev_list = &L2ARC_dev_list;
5132 l2arc_free_on_write = &L2ARC_free_on_write;
5133 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5134 offsetof(l2arc_dev_t, l2ad_node));
5135 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5136 offsetof(l2arc_data_free_t, l2df_list_node));
5143 * This is called from dmu_fini(), which is called from spa_fini();
5144 * Because of this, we can assume that all l2arc devices have
5145 * already been removed when the pools themselves were removed.
5148 l2arc_do_free_on_write();
5150 mutex_destroy(&l2arc_feed_thr_lock);
5151 cv_destroy(&l2arc_feed_thr_cv);
5152 mutex_destroy(&l2arc_dev_mtx);
5153 mutex_destroy(&l2arc_buflist_mtx);
5154 mutex_destroy(&l2arc_free_on_write_mtx);
5156 list_destroy(l2arc_dev_list);
5157 list_destroy(l2arc_free_on_write);
5163 if (!(spa_mode_global & FWRITE))
5166 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5167 TS_RUN, minclsyspri);
5173 if (!(spa_mode_global & FWRITE))
5176 mutex_enter(&l2arc_feed_thr_lock);
5177 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5178 l2arc_thread_exit = 1;
5179 while (l2arc_thread_exit != 0)
5180 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5181 mutex_exit(&l2arc_feed_thr_lock);