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) 2011 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 <zfs_fletcher.h>
136 #include <vm/vm_pageout.h>
140 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
141 boolean_t arc_watch = B_FALSE;
146 static kmutex_t arc_reclaim_thr_lock;
147 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
148 static uint8_t arc_thread_exit;
150 extern int zfs_write_limit_shift;
151 extern uint64_t zfs_write_limit_max;
152 extern kmutex_t zfs_write_limit_lock;
154 #define ARC_REDUCE_DNLC_PERCENT 3
155 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
157 typedef enum arc_reclaim_strategy {
158 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
159 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
160 } arc_reclaim_strategy_t;
162 /* number of seconds before growing cache again */
163 static int arc_grow_retry = 60;
165 /* shift of arc_c for calculating both min and max arc_p */
166 static int arc_p_min_shift = 4;
168 /* log2(fraction of arc to reclaim) */
169 static int arc_shrink_shift = 5;
172 * minimum lifespan of a prefetch block in clock ticks
173 * (initialized in arc_init())
175 static int arc_min_prefetch_lifespan;
178 extern int zfs_prefetch_disable;
181 * The arc has filled available memory and has now warmed up.
183 static boolean_t arc_warm;
186 * These tunables are for performance analysis.
188 uint64_t zfs_arc_max;
189 uint64_t zfs_arc_min;
190 uint64_t zfs_arc_meta_limit = 0;
191 int zfs_arc_grow_retry = 0;
192 int zfs_arc_shrink_shift = 0;
193 int zfs_arc_p_min_shift = 0;
195 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
196 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
197 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
198 SYSCTL_DECL(_vfs_zfs);
199 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
201 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
205 * Note that buffers can be in one of 6 states:
206 * ARC_anon - anonymous (discussed below)
207 * ARC_mru - recently used, currently cached
208 * ARC_mru_ghost - recentely used, no longer in cache
209 * ARC_mfu - frequently used, currently cached
210 * ARC_mfu_ghost - frequently used, no longer in cache
211 * ARC_l2c_only - exists in L2ARC but not other states
212 * When there are no active references to the buffer, they are
213 * are linked onto a list in one of these arc states. These are
214 * the only buffers that can be evicted or deleted. Within each
215 * state there are multiple lists, one for meta-data and one for
216 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
217 * etc.) is tracked separately so that it can be managed more
218 * explicitly: favored over data, limited explicitly.
220 * Anonymous buffers are buffers that are not associated with
221 * a DVA. These are buffers that hold dirty block copies
222 * before they are written to stable storage. By definition,
223 * they are "ref'd" and are considered part of arc_mru
224 * that cannot be freed. Generally, they will aquire a DVA
225 * as they are written and migrate onto the arc_mru list.
227 * The ARC_l2c_only state is for buffers that are in the second
228 * level ARC but no longer in any of the ARC_m* lists. The second
229 * level ARC itself may also contain buffers that are in any of
230 * the ARC_m* states - meaning that a buffer can exist in two
231 * places. The reason for the ARC_l2c_only state is to keep the
232 * buffer header in the hash table, so that reads that hit the
233 * second level ARC benefit from these fast lookups.
236 #define ARCS_LOCK_PAD CACHE_LINE_SIZE
240 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))];
245 * must be power of two for mask use to work
248 #define ARC_BUFC_NUMDATALISTS 16
249 #define ARC_BUFC_NUMMETADATALISTS 16
250 #define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS)
252 typedef struct arc_state {
253 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
254 uint64_t arcs_size; /* total amount of data in this state */
255 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */
256 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE);
259 #define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock))
262 static arc_state_t ARC_anon;
263 static arc_state_t ARC_mru;
264 static arc_state_t ARC_mru_ghost;
265 static arc_state_t ARC_mfu;
266 static arc_state_t ARC_mfu_ghost;
267 static arc_state_t ARC_l2c_only;
269 typedef struct arc_stats {
270 kstat_named_t arcstat_hits;
271 kstat_named_t arcstat_misses;
272 kstat_named_t arcstat_demand_data_hits;
273 kstat_named_t arcstat_demand_data_misses;
274 kstat_named_t arcstat_demand_metadata_hits;
275 kstat_named_t arcstat_demand_metadata_misses;
276 kstat_named_t arcstat_prefetch_data_hits;
277 kstat_named_t arcstat_prefetch_data_misses;
278 kstat_named_t arcstat_prefetch_metadata_hits;
279 kstat_named_t arcstat_prefetch_metadata_misses;
280 kstat_named_t arcstat_mru_hits;
281 kstat_named_t arcstat_mru_ghost_hits;
282 kstat_named_t arcstat_mfu_hits;
283 kstat_named_t arcstat_mfu_ghost_hits;
284 kstat_named_t arcstat_allocated;
285 kstat_named_t arcstat_deleted;
286 kstat_named_t arcstat_stolen;
287 kstat_named_t arcstat_recycle_miss;
288 kstat_named_t arcstat_mutex_miss;
289 kstat_named_t arcstat_evict_skip;
290 kstat_named_t arcstat_evict_l2_cached;
291 kstat_named_t arcstat_evict_l2_eligible;
292 kstat_named_t arcstat_evict_l2_ineligible;
293 kstat_named_t arcstat_hash_elements;
294 kstat_named_t arcstat_hash_elements_max;
295 kstat_named_t arcstat_hash_collisions;
296 kstat_named_t arcstat_hash_chains;
297 kstat_named_t arcstat_hash_chain_max;
298 kstat_named_t arcstat_p;
299 kstat_named_t arcstat_c;
300 kstat_named_t arcstat_c_min;
301 kstat_named_t arcstat_c_max;
302 kstat_named_t arcstat_size;
303 kstat_named_t arcstat_hdr_size;
304 kstat_named_t arcstat_data_size;
305 kstat_named_t arcstat_other_size;
306 kstat_named_t arcstat_l2_hits;
307 kstat_named_t arcstat_l2_misses;
308 kstat_named_t arcstat_l2_feeds;
309 kstat_named_t arcstat_l2_rw_clash;
310 kstat_named_t arcstat_l2_read_bytes;
311 kstat_named_t arcstat_l2_write_bytes;
312 kstat_named_t arcstat_l2_writes_sent;
313 kstat_named_t arcstat_l2_writes_done;
314 kstat_named_t arcstat_l2_writes_error;
315 kstat_named_t arcstat_l2_writes_hdr_miss;
316 kstat_named_t arcstat_l2_evict_lock_retry;
317 kstat_named_t arcstat_l2_evict_reading;
318 kstat_named_t arcstat_l2_free_on_write;
319 kstat_named_t arcstat_l2_abort_lowmem;
320 kstat_named_t arcstat_l2_cksum_bad;
321 kstat_named_t arcstat_l2_io_error;
322 kstat_named_t arcstat_l2_size;
323 kstat_named_t arcstat_l2_hdr_size;
324 kstat_named_t arcstat_memory_throttle_count;
325 kstat_named_t arcstat_l2_write_trylock_fail;
326 kstat_named_t arcstat_l2_write_passed_headroom;
327 kstat_named_t arcstat_l2_write_spa_mismatch;
328 kstat_named_t arcstat_l2_write_in_l2;
329 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
330 kstat_named_t arcstat_l2_write_not_cacheable;
331 kstat_named_t arcstat_l2_write_full;
332 kstat_named_t arcstat_l2_write_buffer_iter;
333 kstat_named_t arcstat_l2_write_pios;
334 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
335 kstat_named_t arcstat_l2_write_buffer_list_iter;
336 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
339 static arc_stats_t arc_stats = {
340 { "hits", KSTAT_DATA_UINT64 },
341 { "misses", KSTAT_DATA_UINT64 },
342 { "demand_data_hits", KSTAT_DATA_UINT64 },
343 { "demand_data_misses", KSTAT_DATA_UINT64 },
344 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
345 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
346 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
347 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
348 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
349 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
350 { "mru_hits", KSTAT_DATA_UINT64 },
351 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
352 { "mfu_hits", KSTAT_DATA_UINT64 },
353 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
354 { "allocated", KSTAT_DATA_UINT64 },
355 { "deleted", KSTAT_DATA_UINT64 },
356 { "stolen", KSTAT_DATA_UINT64 },
357 { "recycle_miss", KSTAT_DATA_UINT64 },
358 { "mutex_miss", KSTAT_DATA_UINT64 },
359 { "evict_skip", KSTAT_DATA_UINT64 },
360 { "evict_l2_cached", KSTAT_DATA_UINT64 },
361 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
362 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
363 { "hash_elements", KSTAT_DATA_UINT64 },
364 { "hash_elements_max", KSTAT_DATA_UINT64 },
365 { "hash_collisions", KSTAT_DATA_UINT64 },
366 { "hash_chains", KSTAT_DATA_UINT64 },
367 { "hash_chain_max", KSTAT_DATA_UINT64 },
368 { "p", KSTAT_DATA_UINT64 },
369 { "c", KSTAT_DATA_UINT64 },
370 { "c_min", KSTAT_DATA_UINT64 },
371 { "c_max", KSTAT_DATA_UINT64 },
372 { "size", KSTAT_DATA_UINT64 },
373 { "hdr_size", KSTAT_DATA_UINT64 },
374 { "data_size", KSTAT_DATA_UINT64 },
375 { "other_size", KSTAT_DATA_UINT64 },
376 { "l2_hits", KSTAT_DATA_UINT64 },
377 { "l2_misses", KSTAT_DATA_UINT64 },
378 { "l2_feeds", KSTAT_DATA_UINT64 },
379 { "l2_rw_clash", KSTAT_DATA_UINT64 },
380 { "l2_read_bytes", KSTAT_DATA_UINT64 },
381 { "l2_write_bytes", KSTAT_DATA_UINT64 },
382 { "l2_writes_sent", KSTAT_DATA_UINT64 },
383 { "l2_writes_done", KSTAT_DATA_UINT64 },
384 { "l2_writes_error", KSTAT_DATA_UINT64 },
385 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
386 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
387 { "l2_evict_reading", KSTAT_DATA_UINT64 },
388 { "l2_free_on_write", KSTAT_DATA_UINT64 },
389 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
390 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
391 { "l2_io_error", KSTAT_DATA_UINT64 },
392 { "l2_size", KSTAT_DATA_UINT64 },
393 { "l2_hdr_size", KSTAT_DATA_UINT64 },
394 { "memory_throttle_count", KSTAT_DATA_UINT64 },
395 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
396 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
397 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
398 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
399 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
400 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
401 { "l2_write_full", KSTAT_DATA_UINT64 },
402 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
403 { "l2_write_pios", KSTAT_DATA_UINT64 },
404 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
405 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
406 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }
409 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
411 #define ARCSTAT_INCR(stat, val) \
412 atomic_add_64(&arc_stats.stat.value.ui64, (val));
414 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
415 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
417 #define ARCSTAT_MAX(stat, val) { \
419 while ((val) > (m = arc_stats.stat.value.ui64) && \
420 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
424 #define ARCSTAT_MAXSTAT(stat) \
425 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
428 * We define a macro to allow ARC hits/misses to be easily broken down by
429 * two separate conditions, giving a total of four different subtypes for
430 * each of hits and misses (so eight statistics total).
432 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
435 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
437 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
441 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
443 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
448 static arc_state_t *arc_anon;
449 static arc_state_t *arc_mru;
450 static arc_state_t *arc_mru_ghost;
451 static arc_state_t *arc_mfu;
452 static arc_state_t *arc_mfu_ghost;
453 static arc_state_t *arc_l2c_only;
456 * There are several ARC variables that are critical to export as kstats --
457 * but we don't want to have to grovel around in the kstat whenever we wish to
458 * manipulate them. For these variables, we therefore define them to be in
459 * terms of the statistic variable. This assures that we are not introducing
460 * the possibility of inconsistency by having shadow copies of the variables,
461 * while still allowing the code to be readable.
463 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
464 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
465 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
466 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
467 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
469 static int arc_no_grow; /* Don't try to grow cache size */
470 static uint64_t arc_tempreserve;
471 static uint64_t arc_loaned_bytes;
472 static uint64_t arc_meta_used;
473 static uint64_t arc_meta_limit;
474 static uint64_t arc_meta_max = 0;
475 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RDTUN,
476 &arc_meta_used, 0, "ARC metadata used");
477 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RDTUN,
478 &arc_meta_limit, 0, "ARC metadata limit");
480 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
482 typedef struct arc_callback arc_callback_t;
484 struct arc_callback {
486 arc_done_func_t *acb_done;
488 zio_t *acb_zio_dummy;
489 arc_callback_t *acb_next;
492 typedef struct arc_write_callback arc_write_callback_t;
494 struct arc_write_callback {
496 arc_done_func_t *awcb_ready;
497 arc_done_func_t *awcb_done;
502 /* protected by hash lock */
507 kmutex_t b_freeze_lock;
508 zio_cksum_t *b_freeze_cksum;
511 arc_buf_hdr_t *b_hash_next;
516 arc_callback_t *b_acb;
520 arc_buf_contents_t b_type;
524 /* protected by arc state mutex */
525 arc_state_t *b_state;
526 list_node_t b_arc_node;
528 /* updated atomically */
529 clock_t b_arc_access;
531 /* self protecting */
534 l2arc_buf_hdr_t *b_l2hdr;
535 list_node_t b_l2node;
538 static arc_buf_t *arc_eviction_list;
539 static kmutex_t arc_eviction_mtx;
540 static arc_buf_hdr_t arc_eviction_hdr;
541 static void arc_get_data_buf(arc_buf_t *buf);
542 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
543 static int arc_evict_needed(arc_buf_contents_t type);
544 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
546 static void arc_buf_watch(arc_buf_t *buf);
549 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
551 #define GHOST_STATE(state) \
552 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
553 (state) == arc_l2c_only)
556 * Private ARC flags. These flags are private ARC only flags that will show up
557 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
558 * be passed in as arc_flags in things like arc_read. However, these flags
559 * should never be passed and should only be set by ARC code. When adding new
560 * public flags, make sure not to smash the private ones.
563 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
564 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
565 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
566 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
567 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
568 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
569 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
570 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
571 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
572 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
574 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
575 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
576 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
577 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
578 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
579 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
580 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
581 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
582 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
583 (hdr)->b_l2hdr != NULL)
584 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
585 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
586 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
592 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
593 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
596 * Hash table routines
599 #define HT_LOCK_PAD CACHE_LINE_SIZE
604 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
608 #define BUF_LOCKS 256
609 typedef struct buf_hash_table {
611 arc_buf_hdr_t **ht_table;
612 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
615 static buf_hash_table_t buf_hash_table;
617 #define BUF_HASH_INDEX(spa, dva, birth) \
618 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
619 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
620 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
621 #define HDR_LOCK(hdr) \
622 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
624 uint64_t zfs_crc64_table[256];
630 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
631 #define L2ARC_HEADROOM 2 /* num of writes */
632 #define L2ARC_FEED_SECS 1 /* caching interval secs */
633 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
635 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
636 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
639 * L2ARC Performance Tunables
641 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
642 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
643 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
644 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
645 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
646 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
647 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
648 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
650 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
651 &l2arc_write_max, 0, "max write size");
652 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
653 &l2arc_write_boost, 0, "extra write during warmup");
654 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
655 &l2arc_headroom, 0, "number of dev writes");
656 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
657 &l2arc_feed_secs, 0, "interval seconds");
658 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
659 &l2arc_feed_min_ms, 0, "min interval milliseconds");
661 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
662 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
663 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
664 &l2arc_feed_again, 0, "turbo warmup");
665 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
666 &l2arc_norw, 0, "no reads during writes");
668 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
669 &ARC_anon.arcs_size, 0, "size of anonymous state");
670 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
671 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
672 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
673 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
675 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
676 &ARC_mru.arcs_size, 0, "size of mru state");
677 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
678 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
679 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
680 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
682 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
683 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state");
684 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
685 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
686 "size of metadata in mru ghost state");
687 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
688 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
689 "size of data in mru ghost state");
691 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
692 &ARC_mfu.arcs_size, 0, "size of mfu state");
693 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
694 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
695 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
696 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
698 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
699 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state");
700 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
701 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
702 "size of metadata in mfu ghost state");
703 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
704 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
705 "size of data in mfu ghost state");
707 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
708 &ARC_l2c_only.arcs_size, 0, "size of mru state");
713 typedef struct l2arc_dev {
714 vdev_t *l2ad_vdev; /* vdev */
715 spa_t *l2ad_spa; /* spa */
716 uint64_t l2ad_hand; /* next write location */
717 uint64_t l2ad_write; /* desired write size, bytes */
718 uint64_t l2ad_boost; /* warmup write boost, bytes */
719 uint64_t l2ad_start; /* first addr on device */
720 uint64_t l2ad_end; /* last addr on device */
721 uint64_t l2ad_evict; /* last addr eviction reached */
722 boolean_t l2ad_first; /* first sweep through */
723 boolean_t l2ad_writing; /* currently writing */
724 list_t *l2ad_buflist; /* buffer list */
725 list_node_t l2ad_node; /* device list node */
728 static list_t L2ARC_dev_list; /* device list */
729 static list_t *l2arc_dev_list; /* device list pointer */
730 static kmutex_t l2arc_dev_mtx; /* device list mutex */
731 static l2arc_dev_t *l2arc_dev_last; /* last device used */
732 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
733 static list_t L2ARC_free_on_write; /* free after write buf list */
734 static list_t *l2arc_free_on_write; /* free after write list ptr */
735 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
736 static uint64_t l2arc_ndev; /* number of devices */
738 typedef struct l2arc_read_callback {
739 arc_buf_t *l2rcb_buf; /* read buffer */
740 spa_t *l2rcb_spa; /* spa */
741 blkptr_t l2rcb_bp; /* original blkptr */
742 zbookmark_t l2rcb_zb; /* original bookmark */
743 int l2rcb_flags; /* original flags */
744 } l2arc_read_callback_t;
746 typedef struct l2arc_write_callback {
747 l2arc_dev_t *l2wcb_dev; /* device info */
748 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
749 } l2arc_write_callback_t;
751 struct l2arc_buf_hdr {
752 /* protected by arc_buf_hdr mutex */
753 l2arc_dev_t *b_dev; /* L2ARC device */
754 uint64_t b_daddr; /* disk address, offset byte */
757 typedef struct l2arc_data_free {
758 /* protected by l2arc_free_on_write_mtx */
761 void (*l2df_func)(void *, size_t);
762 list_node_t l2df_list_node;
765 static kmutex_t l2arc_feed_thr_lock;
766 static kcondvar_t l2arc_feed_thr_cv;
767 static uint8_t l2arc_thread_exit;
769 static void l2arc_read_done(zio_t *zio);
770 static void l2arc_hdr_stat_add(void);
771 static void l2arc_hdr_stat_remove(void);
774 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
776 uint8_t *vdva = (uint8_t *)dva;
777 uint64_t crc = -1ULL;
780 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
782 for (i = 0; i < sizeof (dva_t); i++)
783 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
785 crc ^= (spa>>8) ^ birth;
790 #define BUF_EMPTY(buf) \
791 ((buf)->b_dva.dva_word[0] == 0 && \
792 (buf)->b_dva.dva_word[1] == 0 && \
795 #define BUF_EQUAL(spa, dva, birth, buf) \
796 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
797 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
798 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
801 buf_discard_identity(arc_buf_hdr_t *hdr)
803 hdr->b_dva.dva_word[0] = 0;
804 hdr->b_dva.dva_word[1] = 0;
809 static arc_buf_hdr_t *
810 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
812 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
813 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
816 mutex_enter(hash_lock);
817 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
818 buf = buf->b_hash_next) {
819 if (BUF_EQUAL(spa, dva, birth, buf)) {
824 mutex_exit(hash_lock);
830 * Insert an entry into the hash table. If there is already an element
831 * equal to elem in the hash table, then the already existing element
832 * will be returned and the new element will not be inserted.
833 * Otherwise returns NULL.
835 static arc_buf_hdr_t *
836 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
838 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
839 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
843 ASSERT(!HDR_IN_HASH_TABLE(buf));
845 mutex_enter(hash_lock);
846 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
847 fbuf = fbuf->b_hash_next, i++) {
848 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
852 buf->b_hash_next = buf_hash_table.ht_table[idx];
853 buf_hash_table.ht_table[idx] = buf;
854 buf->b_flags |= ARC_IN_HASH_TABLE;
856 /* collect some hash table performance data */
858 ARCSTAT_BUMP(arcstat_hash_collisions);
860 ARCSTAT_BUMP(arcstat_hash_chains);
862 ARCSTAT_MAX(arcstat_hash_chain_max, i);
865 ARCSTAT_BUMP(arcstat_hash_elements);
866 ARCSTAT_MAXSTAT(arcstat_hash_elements);
872 buf_hash_remove(arc_buf_hdr_t *buf)
874 arc_buf_hdr_t *fbuf, **bufp;
875 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
877 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
878 ASSERT(HDR_IN_HASH_TABLE(buf));
880 bufp = &buf_hash_table.ht_table[idx];
881 while ((fbuf = *bufp) != buf) {
882 ASSERT(fbuf != NULL);
883 bufp = &fbuf->b_hash_next;
885 *bufp = buf->b_hash_next;
886 buf->b_hash_next = NULL;
887 buf->b_flags &= ~ARC_IN_HASH_TABLE;
889 /* collect some hash table performance data */
890 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
892 if (buf_hash_table.ht_table[idx] &&
893 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
894 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
898 * Global data structures and functions for the buf kmem cache.
900 static kmem_cache_t *hdr_cache;
901 static kmem_cache_t *buf_cache;
908 kmem_free(buf_hash_table.ht_table,
909 (buf_hash_table.ht_mask + 1) * sizeof (void *));
910 for (i = 0; i < BUF_LOCKS; i++)
911 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
912 kmem_cache_destroy(hdr_cache);
913 kmem_cache_destroy(buf_cache);
917 * Constructor callback - called when the cache is empty
918 * and a new buf is requested.
922 hdr_cons(void *vbuf, void *unused, int kmflag)
924 arc_buf_hdr_t *buf = vbuf;
926 bzero(buf, sizeof (arc_buf_hdr_t));
927 refcount_create(&buf->b_refcnt);
928 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
929 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
930 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
937 buf_cons(void *vbuf, void *unused, int kmflag)
939 arc_buf_t *buf = vbuf;
941 bzero(buf, sizeof (arc_buf_t));
942 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
943 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
949 * Destructor callback - called when a cached buf is
950 * no longer required.
954 hdr_dest(void *vbuf, void *unused)
956 arc_buf_hdr_t *buf = vbuf;
958 ASSERT(BUF_EMPTY(buf));
959 refcount_destroy(&buf->b_refcnt);
960 cv_destroy(&buf->b_cv);
961 mutex_destroy(&buf->b_freeze_lock);
962 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
967 buf_dest(void *vbuf, void *unused)
969 arc_buf_t *buf = vbuf;
971 mutex_destroy(&buf->b_evict_lock);
972 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
976 * Reclaim callback -- invoked when memory is low.
980 hdr_recl(void *unused)
982 dprintf("hdr_recl called\n");
984 * umem calls the reclaim func when we destroy the buf cache,
985 * which is after we do arc_fini().
988 cv_signal(&arc_reclaim_thr_cv);
995 uint64_t hsize = 1ULL << 12;
999 * The hash table is big enough to fill all of physical memory
1000 * with an average 64K block size. The table will take up
1001 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
1003 while (hsize * 65536 < (uint64_t)physmem * PAGESIZE)
1006 buf_hash_table.ht_mask = hsize - 1;
1007 buf_hash_table.ht_table =
1008 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1009 if (buf_hash_table.ht_table == NULL) {
1010 ASSERT(hsize > (1ULL << 8));
1015 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
1016 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
1017 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1018 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1020 for (i = 0; i < 256; i++)
1021 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1022 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1024 for (i = 0; i < BUF_LOCKS; i++) {
1025 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1026 NULL, MUTEX_DEFAULT, NULL);
1030 #define ARC_MINTIME (hz>>4) /* 62 ms */
1033 arc_cksum_verify(arc_buf_t *buf)
1037 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1040 mutex_enter(&buf->b_hdr->b_freeze_lock);
1041 if (buf->b_hdr->b_freeze_cksum == NULL ||
1042 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
1043 mutex_exit(&buf->b_hdr->b_freeze_lock);
1046 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1047 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1048 panic("buffer modified while frozen!");
1049 mutex_exit(&buf->b_hdr->b_freeze_lock);
1053 arc_cksum_equal(arc_buf_t *buf)
1058 mutex_enter(&buf->b_hdr->b_freeze_lock);
1059 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1060 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1061 mutex_exit(&buf->b_hdr->b_freeze_lock);
1067 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1069 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1072 mutex_enter(&buf->b_hdr->b_freeze_lock);
1073 if (buf->b_hdr->b_freeze_cksum != NULL) {
1074 mutex_exit(&buf->b_hdr->b_freeze_lock);
1077 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1078 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1079 buf->b_hdr->b_freeze_cksum);
1080 mutex_exit(&buf->b_hdr->b_freeze_lock);
1083 #endif /* illumos */
1088 typedef struct procctl {
1096 arc_buf_unwatch(arc_buf_t *buf)
1103 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1104 ctl.prwatch.pr_size = 0;
1105 ctl.prwatch.pr_wflags = 0;
1106 result = write(arc_procfd, &ctl, sizeof (ctl));
1107 ASSERT3U(result, ==, sizeof (ctl));
1114 arc_buf_watch(arc_buf_t *buf)
1121 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1122 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1123 ctl.prwatch.pr_wflags = WA_WRITE;
1124 result = write(arc_procfd, &ctl, sizeof (ctl));
1125 ASSERT3U(result, ==, sizeof (ctl));
1129 #endif /* illumos */
1132 arc_buf_thaw(arc_buf_t *buf)
1134 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1135 if (buf->b_hdr->b_state != arc_anon)
1136 panic("modifying non-anon buffer!");
1137 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1138 panic("modifying buffer while i/o in progress!");
1139 arc_cksum_verify(buf);
1142 mutex_enter(&buf->b_hdr->b_freeze_lock);
1143 if (buf->b_hdr->b_freeze_cksum != NULL) {
1144 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1145 buf->b_hdr->b_freeze_cksum = NULL;
1148 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1149 if (buf->b_hdr->b_thawed)
1150 kmem_free(buf->b_hdr->b_thawed, 1);
1151 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1154 mutex_exit(&buf->b_hdr->b_freeze_lock);
1157 arc_buf_unwatch(buf);
1158 #endif /* illumos */
1162 arc_buf_freeze(arc_buf_t *buf)
1164 kmutex_t *hash_lock;
1166 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1169 hash_lock = HDR_LOCK(buf->b_hdr);
1170 mutex_enter(hash_lock);
1172 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1173 buf->b_hdr->b_state == arc_anon);
1174 arc_cksum_compute(buf, B_FALSE);
1175 mutex_exit(hash_lock);
1180 get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock)
1182 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth);
1184 if (ab->b_type == ARC_BUFC_METADATA)
1185 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1);
1187 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1);
1188 buf_hashid += ARC_BUFC_NUMMETADATALISTS;
1191 *list = &state->arcs_lists[buf_hashid];
1192 *lock = ARCS_LOCK(state, buf_hashid);
1197 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1199 ASSERT(MUTEX_HELD(hash_lock));
1201 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1202 (ab->b_state != arc_anon)) {
1203 uint64_t delta = ab->b_size * ab->b_datacnt;
1204 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1208 get_buf_info(ab, ab->b_state, &list, &lock);
1209 ASSERT(!MUTEX_HELD(lock));
1211 ASSERT(list_link_active(&ab->b_arc_node));
1212 list_remove(list, ab);
1213 if (GHOST_STATE(ab->b_state)) {
1214 ASSERT0(ab->b_datacnt);
1215 ASSERT3P(ab->b_buf, ==, NULL);
1219 ASSERT3U(*size, >=, delta);
1220 atomic_add_64(size, -delta);
1222 /* remove the prefetch flag if we get a reference */
1223 if (ab->b_flags & ARC_PREFETCH)
1224 ab->b_flags &= ~ARC_PREFETCH;
1229 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1232 arc_state_t *state = ab->b_state;
1234 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1235 ASSERT(!GHOST_STATE(state));
1237 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1238 (state != arc_anon)) {
1239 uint64_t *size = &state->arcs_lsize[ab->b_type];
1243 get_buf_info(ab, state, &list, &lock);
1244 ASSERT(!MUTEX_HELD(lock));
1246 ASSERT(!list_link_active(&ab->b_arc_node));
1247 list_insert_head(list, ab);
1248 ASSERT(ab->b_datacnt > 0);
1249 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1256 * Move the supplied buffer to the indicated state. The mutex
1257 * for the buffer must be held by the caller.
1260 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1262 arc_state_t *old_state = ab->b_state;
1263 int64_t refcnt = refcount_count(&ab->b_refcnt);
1264 uint64_t from_delta, to_delta;
1268 ASSERT(MUTEX_HELD(hash_lock));
1269 ASSERT(new_state != old_state);
1270 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1271 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1272 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1274 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1277 * If this buffer is evictable, transfer it from the
1278 * old state list to the new state list.
1281 if (old_state != arc_anon) {
1283 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1285 get_buf_info(ab, old_state, &list, &lock);
1286 use_mutex = !MUTEX_HELD(lock);
1290 ASSERT(list_link_active(&ab->b_arc_node));
1291 list_remove(list, ab);
1294 * If prefetching out of the ghost cache,
1295 * we will have a non-zero datacnt.
1297 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1298 /* ghost elements have a ghost size */
1299 ASSERT(ab->b_buf == NULL);
1300 from_delta = ab->b_size;
1302 ASSERT3U(*size, >=, from_delta);
1303 atomic_add_64(size, -from_delta);
1308 if (new_state != arc_anon) {
1310 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1312 get_buf_info(ab, new_state, &list, &lock);
1313 use_mutex = !MUTEX_HELD(lock);
1317 list_insert_head(list, ab);
1319 /* ghost elements have a ghost size */
1320 if (GHOST_STATE(new_state)) {
1321 ASSERT(ab->b_datacnt == 0);
1322 ASSERT(ab->b_buf == NULL);
1323 to_delta = ab->b_size;
1325 atomic_add_64(size, to_delta);
1332 ASSERT(!BUF_EMPTY(ab));
1333 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1334 buf_hash_remove(ab);
1336 /* adjust state sizes */
1338 atomic_add_64(&new_state->arcs_size, to_delta);
1340 ASSERT3U(old_state->arcs_size, >=, from_delta);
1341 atomic_add_64(&old_state->arcs_size, -from_delta);
1343 ab->b_state = new_state;
1345 /* adjust l2arc hdr stats */
1346 if (new_state == arc_l2c_only)
1347 l2arc_hdr_stat_add();
1348 else if (old_state == arc_l2c_only)
1349 l2arc_hdr_stat_remove();
1353 arc_space_consume(uint64_t space, arc_space_type_t type)
1355 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1358 case ARC_SPACE_DATA:
1359 ARCSTAT_INCR(arcstat_data_size, space);
1361 case ARC_SPACE_OTHER:
1362 ARCSTAT_INCR(arcstat_other_size, space);
1364 case ARC_SPACE_HDRS:
1365 ARCSTAT_INCR(arcstat_hdr_size, space);
1367 case ARC_SPACE_L2HDRS:
1368 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1372 atomic_add_64(&arc_meta_used, space);
1373 atomic_add_64(&arc_size, space);
1377 arc_space_return(uint64_t space, arc_space_type_t type)
1379 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1382 case ARC_SPACE_DATA:
1383 ARCSTAT_INCR(arcstat_data_size, -space);
1385 case ARC_SPACE_OTHER:
1386 ARCSTAT_INCR(arcstat_other_size, -space);
1388 case ARC_SPACE_HDRS:
1389 ARCSTAT_INCR(arcstat_hdr_size, -space);
1391 case ARC_SPACE_L2HDRS:
1392 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1396 ASSERT(arc_meta_used >= space);
1397 if (arc_meta_max < arc_meta_used)
1398 arc_meta_max = arc_meta_used;
1399 atomic_add_64(&arc_meta_used, -space);
1400 ASSERT(arc_size >= space);
1401 atomic_add_64(&arc_size, -space);
1405 arc_data_buf_alloc(uint64_t size)
1407 if (arc_evict_needed(ARC_BUFC_DATA))
1408 cv_signal(&arc_reclaim_thr_cv);
1409 atomic_add_64(&arc_size, size);
1410 return (zio_data_buf_alloc(size));
1414 arc_data_buf_free(void *buf, uint64_t size)
1416 zio_data_buf_free(buf, size);
1417 ASSERT(arc_size >= size);
1418 atomic_add_64(&arc_size, -size);
1422 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1427 ASSERT3U(size, >, 0);
1428 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1429 ASSERT(BUF_EMPTY(hdr));
1432 hdr->b_spa = spa_load_guid(spa);
1433 hdr->b_state = arc_anon;
1434 hdr->b_arc_access = 0;
1435 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1438 buf->b_efunc = NULL;
1439 buf->b_private = NULL;
1442 arc_get_data_buf(buf);
1445 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1446 (void) refcount_add(&hdr->b_refcnt, tag);
1451 static char *arc_onloan_tag = "onloan";
1454 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1455 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1456 * buffers must be returned to the arc before they can be used by the DMU or
1460 arc_loan_buf(spa_t *spa, int size)
1464 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1466 atomic_add_64(&arc_loaned_bytes, size);
1471 * Return a loaned arc buffer to the arc.
1474 arc_return_buf(arc_buf_t *buf, void *tag)
1476 arc_buf_hdr_t *hdr = buf->b_hdr;
1478 ASSERT(buf->b_data != NULL);
1479 (void) refcount_add(&hdr->b_refcnt, tag);
1480 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1482 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1485 /* Detach an arc_buf from a dbuf (tag) */
1487 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1491 ASSERT(buf->b_data != NULL);
1493 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1494 (void) refcount_remove(&hdr->b_refcnt, tag);
1495 buf->b_efunc = NULL;
1496 buf->b_private = NULL;
1498 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1502 arc_buf_clone(arc_buf_t *from)
1505 arc_buf_hdr_t *hdr = from->b_hdr;
1506 uint64_t size = hdr->b_size;
1508 ASSERT(hdr->b_state != arc_anon);
1510 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1513 buf->b_efunc = NULL;
1514 buf->b_private = NULL;
1515 buf->b_next = hdr->b_buf;
1517 arc_get_data_buf(buf);
1518 bcopy(from->b_data, buf->b_data, size);
1519 hdr->b_datacnt += 1;
1524 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1527 kmutex_t *hash_lock;
1530 * Check to see if this buffer is evicted. Callers
1531 * must verify b_data != NULL to know if the add_ref
1534 mutex_enter(&buf->b_evict_lock);
1535 if (buf->b_data == NULL) {
1536 mutex_exit(&buf->b_evict_lock);
1539 hash_lock = HDR_LOCK(buf->b_hdr);
1540 mutex_enter(hash_lock);
1542 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1543 mutex_exit(&buf->b_evict_lock);
1545 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1546 add_reference(hdr, hash_lock, tag);
1547 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1548 arc_access(hdr, hash_lock);
1549 mutex_exit(hash_lock);
1550 ARCSTAT_BUMP(arcstat_hits);
1551 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1552 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1553 data, metadata, hits);
1557 * Free the arc data buffer. If it is an l2arc write in progress,
1558 * the buffer is placed on l2arc_free_on_write to be freed later.
1561 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1563 arc_buf_hdr_t *hdr = buf->b_hdr;
1565 if (HDR_L2_WRITING(hdr)) {
1566 l2arc_data_free_t *df;
1567 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1568 df->l2df_data = buf->b_data;
1569 df->l2df_size = hdr->b_size;
1570 df->l2df_func = free_func;
1571 mutex_enter(&l2arc_free_on_write_mtx);
1572 list_insert_head(l2arc_free_on_write, df);
1573 mutex_exit(&l2arc_free_on_write_mtx);
1574 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1576 free_func(buf->b_data, hdr->b_size);
1581 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1585 /* free up data associated with the buf */
1587 arc_state_t *state = buf->b_hdr->b_state;
1588 uint64_t size = buf->b_hdr->b_size;
1589 arc_buf_contents_t type = buf->b_hdr->b_type;
1591 arc_cksum_verify(buf);
1593 arc_buf_unwatch(buf);
1594 #endif /* illumos */
1597 if (type == ARC_BUFC_METADATA) {
1598 arc_buf_data_free(buf, zio_buf_free);
1599 arc_space_return(size, ARC_SPACE_DATA);
1601 ASSERT(type == ARC_BUFC_DATA);
1602 arc_buf_data_free(buf, zio_data_buf_free);
1603 ARCSTAT_INCR(arcstat_data_size, -size);
1604 atomic_add_64(&arc_size, -size);
1607 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1608 uint64_t *cnt = &state->arcs_lsize[type];
1610 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1611 ASSERT(state != arc_anon);
1613 ASSERT3U(*cnt, >=, size);
1614 atomic_add_64(cnt, -size);
1616 ASSERT3U(state->arcs_size, >=, size);
1617 atomic_add_64(&state->arcs_size, -size);
1619 ASSERT(buf->b_hdr->b_datacnt > 0);
1620 buf->b_hdr->b_datacnt -= 1;
1623 /* only remove the buf if requested */
1627 /* remove the buf from the hdr list */
1628 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1630 *bufp = buf->b_next;
1633 ASSERT(buf->b_efunc == NULL);
1635 /* clean up the buf */
1637 kmem_cache_free(buf_cache, buf);
1641 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1643 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1644 ASSERT3P(hdr->b_state, ==, arc_anon);
1645 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1646 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1648 if (l2hdr != NULL) {
1649 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1651 * To prevent arc_free() and l2arc_evict() from
1652 * attempting to free the same buffer at the same time,
1653 * a FREE_IN_PROGRESS flag is given to arc_free() to
1654 * give it priority. l2arc_evict() can't destroy this
1655 * header while we are waiting on l2arc_buflist_mtx.
1657 * The hdr may be removed from l2ad_buflist before we
1658 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1660 if (!buflist_held) {
1661 mutex_enter(&l2arc_buflist_mtx);
1662 l2hdr = hdr->b_l2hdr;
1665 if (l2hdr != NULL) {
1666 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1667 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1668 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1669 if (hdr->b_state == arc_l2c_only)
1670 l2arc_hdr_stat_remove();
1671 hdr->b_l2hdr = NULL;
1675 mutex_exit(&l2arc_buflist_mtx);
1678 if (!BUF_EMPTY(hdr)) {
1679 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1680 buf_discard_identity(hdr);
1682 while (hdr->b_buf) {
1683 arc_buf_t *buf = hdr->b_buf;
1686 mutex_enter(&arc_eviction_mtx);
1687 mutex_enter(&buf->b_evict_lock);
1688 ASSERT(buf->b_hdr != NULL);
1689 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1690 hdr->b_buf = buf->b_next;
1691 buf->b_hdr = &arc_eviction_hdr;
1692 buf->b_next = arc_eviction_list;
1693 arc_eviction_list = buf;
1694 mutex_exit(&buf->b_evict_lock);
1695 mutex_exit(&arc_eviction_mtx);
1697 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1700 if (hdr->b_freeze_cksum != NULL) {
1701 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1702 hdr->b_freeze_cksum = NULL;
1704 if (hdr->b_thawed) {
1705 kmem_free(hdr->b_thawed, 1);
1706 hdr->b_thawed = NULL;
1709 ASSERT(!list_link_active(&hdr->b_arc_node));
1710 ASSERT3P(hdr->b_hash_next, ==, NULL);
1711 ASSERT3P(hdr->b_acb, ==, NULL);
1712 kmem_cache_free(hdr_cache, hdr);
1716 arc_buf_free(arc_buf_t *buf, void *tag)
1718 arc_buf_hdr_t *hdr = buf->b_hdr;
1719 int hashed = hdr->b_state != arc_anon;
1721 ASSERT(buf->b_efunc == NULL);
1722 ASSERT(buf->b_data != NULL);
1725 kmutex_t *hash_lock = HDR_LOCK(hdr);
1727 mutex_enter(hash_lock);
1729 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1731 (void) remove_reference(hdr, hash_lock, tag);
1732 if (hdr->b_datacnt > 1) {
1733 arc_buf_destroy(buf, FALSE, TRUE);
1735 ASSERT(buf == hdr->b_buf);
1736 ASSERT(buf->b_efunc == NULL);
1737 hdr->b_flags |= ARC_BUF_AVAILABLE;
1739 mutex_exit(hash_lock);
1740 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1743 * We are in the middle of an async write. Don't destroy
1744 * this buffer unless the write completes before we finish
1745 * decrementing the reference count.
1747 mutex_enter(&arc_eviction_mtx);
1748 (void) remove_reference(hdr, NULL, tag);
1749 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1750 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1751 mutex_exit(&arc_eviction_mtx);
1753 arc_hdr_destroy(hdr);
1755 if (remove_reference(hdr, NULL, tag) > 0)
1756 arc_buf_destroy(buf, FALSE, TRUE);
1758 arc_hdr_destroy(hdr);
1763 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1765 arc_buf_hdr_t *hdr = buf->b_hdr;
1766 kmutex_t *hash_lock = HDR_LOCK(hdr);
1767 int no_callback = (buf->b_efunc == NULL);
1769 if (hdr->b_state == arc_anon) {
1770 ASSERT(hdr->b_datacnt == 1);
1771 arc_buf_free(buf, tag);
1772 return (no_callback);
1775 mutex_enter(hash_lock);
1777 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1778 ASSERT(hdr->b_state != arc_anon);
1779 ASSERT(buf->b_data != NULL);
1781 (void) remove_reference(hdr, hash_lock, tag);
1782 if (hdr->b_datacnt > 1) {
1784 arc_buf_destroy(buf, FALSE, TRUE);
1785 } else if (no_callback) {
1786 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1787 ASSERT(buf->b_efunc == NULL);
1788 hdr->b_flags |= ARC_BUF_AVAILABLE;
1790 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1791 refcount_is_zero(&hdr->b_refcnt));
1792 mutex_exit(hash_lock);
1793 return (no_callback);
1797 arc_buf_size(arc_buf_t *buf)
1799 return (buf->b_hdr->b_size);
1803 * Evict buffers from list until we've removed the specified number of
1804 * bytes. Move the removed buffers to the appropriate evict state.
1805 * If the recycle flag is set, then attempt to "recycle" a buffer:
1806 * - look for a buffer to evict that is `bytes' long.
1807 * - return the data block from this buffer rather than freeing it.
1808 * This flag is used by callers that are trying to make space for a
1809 * new buffer in a full arc cache.
1811 * This function makes a "best effort". It skips over any buffers
1812 * it can't get a hash_lock on, and so may not catch all candidates.
1813 * It may also return without evicting as much space as requested.
1816 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1817 arc_buf_contents_t type)
1819 arc_state_t *evicted_state;
1820 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1821 int64_t bytes_remaining;
1822 arc_buf_hdr_t *ab, *ab_prev = NULL;
1823 list_t *evicted_list, *list, *evicted_list_start, *list_start;
1824 kmutex_t *lock, *evicted_lock;
1825 kmutex_t *hash_lock;
1826 boolean_t have_lock;
1827 void *stolen = NULL;
1828 static int evict_metadata_offset, evict_data_offset;
1829 int i, idx, offset, list_count, count;
1831 ASSERT(state == arc_mru || state == arc_mfu);
1833 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1835 if (type == ARC_BUFC_METADATA) {
1837 list_count = ARC_BUFC_NUMMETADATALISTS;
1838 list_start = &state->arcs_lists[0];
1839 evicted_list_start = &evicted_state->arcs_lists[0];
1840 idx = evict_metadata_offset;
1842 offset = ARC_BUFC_NUMMETADATALISTS;
1843 list_start = &state->arcs_lists[offset];
1844 evicted_list_start = &evicted_state->arcs_lists[offset];
1845 list_count = ARC_BUFC_NUMDATALISTS;
1846 idx = evict_data_offset;
1848 bytes_remaining = evicted_state->arcs_lsize[type];
1852 list = &list_start[idx];
1853 evicted_list = &evicted_list_start[idx];
1854 lock = ARCS_LOCK(state, (offset + idx));
1855 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx));
1858 mutex_enter(evicted_lock);
1860 for (ab = list_tail(list); ab; ab = ab_prev) {
1861 ab_prev = list_prev(list, ab);
1862 bytes_remaining -= (ab->b_size * ab->b_datacnt);
1863 /* prefetch buffers have a minimum lifespan */
1864 if (HDR_IO_IN_PROGRESS(ab) ||
1865 (spa && ab->b_spa != spa) ||
1866 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1867 ddi_get_lbolt() - ab->b_arc_access <
1868 arc_min_prefetch_lifespan)) {
1872 /* "lookahead" for better eviction candidate */
1873 if (recycle && ab->b_size != bytes &&
1874 ab_prev && ab_prev->b_size == bytes)
1876 hash_lock = HDR_LOCK(ab);
1877 have_lock = MUTEX_HELD(hash_lock);
1878 if (have_lock || mutex_tryenter(hash_lock)) {
1879 ASSERT0(refcount_count(&ab->b_refcnt));
1880 ASSERT(ab->b_datacnt > 0);
1882 arc_buf_t *buf = ab->b_buf;
1883 if (!mutex_tryenter(&buf->b_evict_lock)) {
1888 bytes_evicted += ab->b_size;
1889 if (recycle && ab->b_type == type &&
1890 ab->b_size == bytes &&
1891 !HDR_L2_WRITING(ab)) {
1892 stolen = buf->b_data;
1897 mutex_enter(&arc_eviction_mtx);
1898 arc_buf_destroy(buf,
1899 buf->b_data == stolen, FALSE);
1900 ab->b_buf = buf->b_next;
1901 buf->b_hdr = &arc_eviction_hdr;
1902 buf->b_next = arc_eviction_list;
1903 arc_eviction_list = buf;
1904 mutex_exit(&arc_eviction_mtx);
1905 mutex_exit(&buf->b_evict_lock);
1907 mutex_exit(&buf->b_evict_lock);
1908 arc_buf_destroy(buf,
1909 buf->b_data == stolen, TRUE);
1914 ARCSTAT_INCR(arcstat_evict_l2_cached,
1917 if (l2arc_write_eligible(ab->b_spa, ab)) {
1918 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1922 arcstat_evict_l2_ineligible,
1927 if (ab->b_datacnt == 0) {
1928 arc_change_state(evicted_state, ab, hash_lock);
1929 ASSERT(HDR_IN_HASH_TABLE(ab));
1930 ab->b_flags |= ARC_IN_HASH_TABLE;
1931 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1932 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1935 mutex_exit(hash_lock);
1936 if (bytes >= 0 && bytes_evicted >= bytes)
1938 if (bytes_remaining > 0) {
1939 mutex_exit(evicted_lock);
1941 idx = ((idx + 1) & (list_count - 1));
1950 mutex_exit(evicted_lock);
1953 idx = ((idx + 1) & (list_count - 1));
1956 if (bytes_evicted < bytes) {
1957 if (count < list_count)
1960 dprintf("only evicted %lld bytes from %x",
1961 (longlong_t)bytes_evicted, state);
1963 if (type == ARC_BUFC_METADATA)
1964 evict_metadata_offset = idx;
1966 evict_data_offset = idx;
1969 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1972 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1975 * We have just evicted some date into the ghost state, make
1976 * sure we also adjust the ghost state size if necessary.
1979 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1980 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1981 arc_mru_ghost->arcs_size - arc_c;
1983 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1985 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1986 arc_evict_ghost(arc_mru_ghost, 0, todelete);
1987 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1988 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1989 arc_mru_ghost->arcs_size +
1990 arc_mfu_ghost->arcs_size - arc_c);
1991 arc_evict_ghost(arc_mfu_ghost, 0, todelete);
1995 ARCSTAT_BUMP(arcstat_stolen);
2001 * Remove buffers from list until we've removed the specified number of
2002 * bytes. Destroy the buffers that are removed.
2005 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
2007 arc_buf_hdr_t *ab, *ab_prev;
2008 arc_buf_hdr_t marker = { 0 };
2009 list_t *list, *list_start;
2010 kmutex_t *hash_lock, *lock;
2011 uint64_t bytes_deleted = 0;
2012 uint64_t bufs_skipped = 0;
2013 static int evict_offset;
2014 int list_count, idx = evict_offset;
2015 int offset, count = 0;
2017 ASSERT(GHOST_STATE(state));
2020 * data lists come after metadata lists
2022 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS];
2023 list_count = ARC_BUFC_NUMDATALISTS;
2024 offset = ARC_BUFC_NUMMETADATALISTS;
2027 list = &list_start[idx];
2028 lock = ARCS_LOCK(state, idx + offset);
2031 for (ab = list_tail(list); ab; ab = ab_prev) {
2032 ab_prev = list_prev(list, ab);
2033 if (spa && ab->b_spa != spa)
2036 /* ignore markers */
2040 hash_lock = HDR_LOCK(ab);
2041 /* caller may be trying to modify this buffer, skip it */
2042 if (MUTEX_HELD(hash_lock))
2044 if (mutex_tryenter(hash_lock)) {
2045 ASSERT(!HDR_IO_IN_PROGRESS(ab));
2046 ASSERT(ab->b_buf == NULL);
2047 ARCSTAT_BUMP(arcstat_deleted);
2048 bytes_deleted += ab->b_size;
2050 if (ab->b_l2hdr != NULL) {
2052 * This buffer is cached on the 2nd Level ARC;
2053 * don't destroy the header.
2055 arc_change_state(arc_l2c_only, ab, hash_lock);
2056 mutex_exit(hash_lock);
2058 arc_change_state(arc_anon, ab, hash_lock);
2059 mutex_exit(hash_lock);
2060 arc_hdr_destroy(ab);
2063 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
2064 if (bytes >= 0 && bytes_deleted >= bytes)
2066 } else if (bytes < 0) {
2068 * Insert a list marker and then wait for the
2069 * hash lock to become available. Once its
2070 * available, restart from where we left off.
2072 list_insert_after(list, ab, &marker);
2074 mutex_enter(hash_lock);
2075 mutex_exit(hash_lock);
2077 ab_prev = list_prev(list, &marker);
2078 list_remove(list, &marker);
2083 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1));
2086 if (count < list_count)
2090 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] &&
2091 (bytes < 0 || bytes_deleted < bytes)) {
2092 list_start = &state->arcs_lists[0];
2093 list_count = ARC_BUFC_NUMMETADATALISTS;
2099 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2103 if (bytes_deleted < bytes)
2104 dprintf("only deleted %lld bytes from %p",
2105 (longlong_t)bytes_deleted, state);
2111 int64_t adjustment, delta;
2117 adjustment = MIN((int64_t)(arc_size - arc_c),
2118 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2121 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2122 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2123 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
2124 adjustment -= delta;
2127 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2128 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2129 (void) arc_evict(arc_mru, 0, delta, FALSE,
2137 adjustment = arc_size - arc_c;
2139 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2140 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2141 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2142 adjustment -= delta;
2145 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2146 int64_t delta = MIN(adjustment,
2147 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2148 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2153 * Adjust ghost lists
2156 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2158 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2159 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2160 arc_evict_ghost(arc_mru_ghost, 0, delta);
2164 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2166 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2167 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2168 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2173 arc_do_user_evicts(void)
2175 static arc_buf_t *tmp_arc_eviction_list;
2178 * Move list over to avoid LOR
2181 mutex_enter(&arc_eviction_mtx);
2182 tmp_arc_eviction_list = arc_eviction_list;
2183 arc_eviction_list = NULL;
2184 mutex_exit(&arc_eviction_mtx);
2186 while (tmp_arc_eviction_list != NULL) {
2187 arc_buf_t *buf = tmp_arc_eviction_list;
2188 tmp_arc_eviction_list = buf->b_next;
2189 mutex_enter(&buf->b_evict_lock);
2191 mutex_exit(&buf->b_evict_lock);
2193 if (buf->b_efunc != NULL)
2194 VERIFY(buf->b_efunc(buf) == 0);
2196 buf->b_efunc = NULL;
2197 buf->b_private = NULL;
2198 kmem_cache_free(buf_cache, buf);
2201 if (arc_eviction_list != NULL)
2206 * Flush all *evictable* data from the cache for the given spa.
2207 * NOTE: this will not touch "active" (i.e. referenced) data.
2210 arc_flush(spa_t *spa)
2215 guid = spa_load_guid(spa);
2217 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) {
2218 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2222 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) {
2223 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2227 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) {
2228 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2232 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) {
2233 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2238 arc_evict_ghost(arc_mru_ghost, guid, -1);
2239 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2241 mutex_enter(&arc_reclaim_thr_lock);
2242 arc_do_user_evicts();
2243 mutex_exit(&arc_reclaim_thr_lock);
2244 ASSERT(spa || arc_eviction_list == NULL);
2250 if (arc_c > arc_c_min) {
2254 to_free = arc_c >> arc_shrink_shift;
2256 to_free = arc_c >> arc_shrink_shift;
2258 if (arc_c > arc_c_min + to_free)
2259 atomic_add_64(&arc_c, -to_free);
2263 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2264 if (arc_c > arc_size)
2265 arc_c = MAX(arc_size, arc_c_min);
2267 arc_p = (arc_c >> 1);
2268 ASSERT(arc_c >= arc_c_min);
2269 ASSERT((int64_t)arc_p >= 0);
2272 if (arc_size > arc_c)
2276 static int needfree = 0;
2279 arc_reclaim_needed(void)
2288 * Cooperate with pagedaemon when it's time for it to scan
2289 * and reclaim some pages.
2291 if (vm_paging_needed())
2296 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2301 * check that we're out of range of the pageout scanner. It starts to
2302 * schedule paging if freemem is less than lotsfree and needfree.
2303 * lotsfree is the high-water mark for pageout, and needfree is the
2304 * number of needed free pages. We add extra pages here to make sure
2305 * the scanner doesn't start up while we're freeing memory.
2307 if (freemem < lotsfree + needfree + extra)
2311 * check to make sure that swapfs has enough space so that anon
2312 * reservations can still succeed. anon_resvmem() checks that the
2313 * availrmem is greater than swapfs_minfree, and the number of reserved
2314 * swap pages. We also add a bit of extra here just to prevent
2315 * circumstances from getting really dire.
2317 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2322 * If we're on an i386 platform, it's possible that we'll exhaust the
2323 * kernel heap space before we ever run out of available physical
2324 * memory. Most checks of the size of the heap_area compare against
2325 * tune.t_minarmem, which is the minimum available real memory that we
2326 * can have in the system. However, this is generally fixed at 25 pages
2327 * which is so low that it's useless. In this comparison, we seek to
2328 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2329 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2332 if (btop(vmem_size(heap_arena, VMEM_FREE)) <
2333 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
2337 if (kmem_used() > (kmem_size() * 3) / 4)
2342 if (spa_get_random(100) == 0)
2348 extern kmem_cache_t *zio_buf_cache[];
2349 extern kmem_cache_t *zio_data_buf_cache[];
2352 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2355 kmem_cache_t *prev_cache = NULL;
2356 kmem_cache_t *prev_data_cache = NULL;
2359 if (arc_meta_used >= arc_meta_limit) {
2361 * We are exceeding our meta-data cache limit.
2362 * Purge some DNLC entries to release holds on meta-data.
2364 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2368 * Reclaim unused memory from all kmem caches.
2375 * An aggressive reclamation will shrink the cache size as well as
2376 * reap free buffers from the arc kmem caches.
2378 if (strat == ARC_RECLAIM_AGGR)
2381 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2382 if (zio_buf_cache[i] != prev_cache) {
2383 prev_cache = zio_buf_cache[i];
2384 kmem_cache_reap_now(zio_buf_cache[i]);
2386 if (zio_data_buf_cache[i] != prev_data_cache) {
2387 prev_data_cache = zio_data_buf_cache[i];
2388 kmem_cache_reap_now(zio_data_buf_cache[i]);
2391 kmem_cache_reap_now(buf_cache);
2392 kmem_cache_reap_now(hdr_cache);
2396 arc_reclaim_thread(void *dummy __unused)
2398 clock_t growtime = 0;
2399 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2402 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2404 mutex_enter(&arc_reclaim_thr_lock);
2405 while (arc_thread_exit == 0) {
2406 if (arc_reclaim_needed()) {
2409 if (last_reclaim == ARC_RECLAIM_CONS) {
2410 last_reclaim = ARC_RECLAIM_AGGR;
2412 last_reclaim = ARC_RECLAIM_CONS;
2416 last_reclaim = ARC_RECLAIM_AGGR;
2420 /* reset the growth delay for every reclaim */
2421 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2423 if (needfree && last_reclaim == ARC_RECLAIM_CONS) {
2425 * If needfree is TRUE our vm_lowmem hook
2426 * was called and in that case we must free some
2427 * memory, so switch to aggressive mode.
2430 last_reclaim = ARC_RECLAIM_AGGR;
2432 arc_kmem_reap_now(last_reclaim);
2435 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2436 arc_no_grow = FALSE;
2441 if (arc_eviction_list != NULL)
2442 arc_do_user_evicts();
2451 /* block until needed, or one second, whichever is shorter */
2452 CALLB_CPR_SAFE_BEGIN(&cpr);
2453 (void) cv_timedwait(&arc_reclaim_thr_cv,
2454 &arc_reclaim_thr_lock, hz);
2455 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2458 arc_thread_exit = 0;
2459 cv_broadcast(&arc_reclaim_thr_cv);
2460 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2465 * Adapt arc info given the number of bytes we are trying to add and
2466 * the state that we are comming from. This function is only called
2467 * when we are adding new content to the cache.
2470 arc_adapt(int bytes, arc_state_t *state)
2473 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2475 if (state == arc_l2c_only)
2480 * Adapt the target size of the MRU list:
2481 * - if we just hit in the MRU ghost list, then increase
2482 * the target size of the MRU list.
2483 * - if we just hit in the MFU ghost list, then increase
2484 * the target size of the MFU list by decreasing the
2485 * target size of the MRU list.
2487 if (state == arc_mru_ghost) {
2488 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2489 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2490 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2492 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2493 } else if (state == arc_mfu_ghost) {
2496 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2497 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2498 mult = MIN(mult, 10);
2500 delta = MIN(bytes * mult, arc_p);
2501 arc_p = MAX(arc_p_min, arc_p - delta);
2503 ASSERT((int64_t)arc_p >= 0);
2505 if (arc_reclaim_needed()) {
2506 cv_signal(&arc_reclaim_thr_cv);
2513 if (arc_c >= arc_c_max)
2517 * If we're within (2 * maxblocksize) bytes of the target
2518 * cache size, increment the target cache size
2520 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2521 atomic_add_64(&arc_c, (int64_t)bytes);
2522 if (arc_c > arc_c_max)
2524 else if (state == arc_anon)
2525 atomic_add_64(&arc_p, (int64_t)bytes);
2529 ASSERT((int64_t)arc_p >= 0);
2533 * Check if the cache has reached its limits and eviction is required
2537 arc_evict_needed(arc_buf_contents_t type)
2539 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2545 * If zio data pages are being allocated out of a separate heap segment,
2546 * then enforce that the size of available vmem for this area remains
2547 * above about 1/32nd free.
2549 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2550 vmem_size(zio_arena, VMEM_FREE) <
2551 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2556 if (arc_reclaim_needed())
2559 return (arc_size > arc_c);
2563 * The buffer, supplied as the first argument, needs a data block.
2564 * So, if we are at cache max, determine which cache should be victimized.
2565 * We have the following cases:
2567 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2568 * In this situation if we're out of space, but the resident size of the MFU is
2569 * under the limit, victimize the MFU cache to satisfy this insertion request.
2571 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2572 * Here, we've used up all of the available space for the MRU, so we need to
2573 * evict from our own cache instead. Evict from the set of resident MRU
2576 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2577 * c minus p represents the MFU space in the cache, since p is the size of the
2578 * cache that is dedicated to the MRU. In this situation there's still space on
2579 * the MFU side, so the MRU side needs to be victimized.
2581 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2582 * MFU's resident set is consuming more space than it has been allotted. In
2583 * this situation, we must victimize our own cache, the MFU, for this insertion.
2586 arc_get_data_buf(arc_buf_t *buf)
2588 arc_state_t *state = buf->b_hdr->b_state;
2589 uint64_t size = buf->b_hdr->b_size;
2590 arc_buf_contents_t type = buf->b_hdr->b_type;
2592 arc_adapt(size, state);
2595 * We have not yet reached cache maximum size,
2596 * just allocate a new buffer.
2598 if (!arc_evict_needed(type)) {
2599 if (type == ARC_BUFC_METADATA) {
2600 buf->b_data = zio_buf_alloc(size);
2601 arc_space_consume(size, ARC_SPACE_DATA);
2603 ASSERT(type == ARC_BUFC_DATA);
2604 buf->b_data = zio_data_buf_alloc(size);
2605 ARCSTAT_INCR(arcstat_data_size, size);
2606 atomic_add_64(&arc_size, size);
2612 * If we are prefetching from the mfu ghost list, this buffer
2613 * will end up on the mru list; so steal space from there.
2615 if (state == arc_mfu_ghost)
2616 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2617 else if (state == arc_mru_ghost)
2620 if (state == arc_mru || state == arc_anon) {
2621 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2622 state = (arc_mfu->arcs_lsize[type] >= size &&
2623 arc_p > mru_used) ? arc_mfu : arc_mru;
2626 uint64_t mfu_space = arc_c - arc_p;
2627 state = (arc_mru->arcs_lsize[type] >= size &&
2628 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2630 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2631 if (type == ARC_BUFC_METADATA) {
2632 buf->b_data = zio_buf_alloc(size);
2633 arc_space_consume(size, ARC_SPACE_DATA);
2635 ASSERT(type == ARC_BUFC_DATA);
2636 buf->b_data = zio_data_buf_alloc(size);
2637 ARCSTAT_INCR(arcstat_data_size, size);
2638 atomic_add_64(&arc_size, size);
2640 ARCSTAT_BUMP(arcstat_recycle_miss);
2642 ASSERT(buf->b_data != NULL);
2645 * Update the state size. Note that ghost states have a
2646 * "ghost size" and so don't need to be updated.
2648 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2649 arc_buf_hdr_t *hdr = buf->b_hdr;
2651 atomic_add_64(&hdr->b_state->arcs_size, size);
2652 if (list_link_active(&hdr->b_arc_node)) {
2653 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2654 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2657 * If we are growing the cache, and we are adding anonymous
2658 * data, and we have outgrown arc_p, update arc_p
2660 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2661 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2662 arc_p = MIN(arc_c, arc_p + size);
2664 ARCSTAT_BUMP(arcstat_allocated);
2668 * This routine is called whenever a buffer is accessed.
2669 * NOTE: the hash lock is dropped in this function.
2672 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2676 ASSERT(MUTEX_HELD(hash_lock));
2678 if (buf->b_state == arc_anon) {
2680 * This buffer is not in the cache, and does not
2681 * appear in our "ghost" list. Add the new buffer
2685 ASSERT(buf->b_arc_access == 0);
2686 buf->b_arc_access = ddi_get_lbolt();
2687 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2688 arc_change_state(arc_mru, buf, hash_lock);
2690 } else if (buf->b_state == arc_mru) {
2691 now = ddi_get_lbolt();
2694 * If this buffer is here because of a prefetch, then either:
2695 * - clear the flag if this is a "referencing" read
2696 * (any subsequent access will bump this into the MFU state).
2698 * - move the buffer to the head of the list if this is
2699 * another prefetch (to make it less likely to be evicted).
2701 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2702 if (refcount_count(&buf->b_refcnt) == 0) {
2703 ASSERT(list_link_active(&buf->b_arc_node));
2705 buf->b_flags &= ~ARC_PREFETCH;
2706 ARCSTAT_BUMP(arcstat_mru_hits);
2708 buf->b_arc_access = now;
2713 * This buffer has been "accessed" only once so far,
2714 * but it is still in the cache. Move it to the MFU
2717 if (now > buf->b_arc_access + ARC_MINTIME) {
2719 * More than 125ms have passed since we
2720 * instantiated this buffer. Move it to the
2721 * most frequently used state.
2723 buf->b_arc_access = now;
2724 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2725 arc_change_state(arc_mfu, buf, hash_lock);
2727 ARCSTAT_BUMP(arcstat_mru_hits);
2728 } else if (buf->b_state == arc_mru_ghost) {
2729 arc_state_t *new_state;
2731 * This buffer has been "accessed" recently, but
2732 * was evicted from the cache. Move it to the
2736 if (buf->b_flags & ARC_PREFETCH) {
2737 new_state = arc_mru;
2738 if (refcount_count(&buf->b_refcnt) > 0)
2739 buf->b_flags &= ~ARC_PREFETCH;
2740 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2742 new_state = arc_mfu;
2743 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2746 buf->b_arc_access = ddi_get_lbolt();
2747 arc_change_state(new_state, buf, hash_lock);
2749 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2750 } else if (buf->b_state == arc_mfu) {
2752 * This buffer has been accessed more than once and is
2753 * still in the cache. Keep it in the MFU state.
2755 * NOTE: an add_reference() that occurred when we did
2756 * the arc_read() will have kicked this off the list.
2757 * If it was a prefetch, we will explicitly move it to
2758 * the head of the list now.
2760 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2761 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2762 ASSERT(list_link_active(&buf->b_arc_node));
2764 ARCSTAT_BUMP(arcstat_mfu_hits);
2765 buf->b_arc_access = ddi_get_lbolt();
2766 } else if (buf->b_state == arc_mfu_ghost) {
2767 arc_state_t *new_state = arc_mfu;
2769 * This buffer has been accessed more than once but has
2770 * been evicted from the cache. Move it back to the
2774 if (buf->b_flags & ARC_PREFETCH) {
2776 * This is a prefetch access...
2777 * move this block back to the MRU state.
2779 ASSERT0(refcount_count(&buf->b_refcnt));
2780 new_state = arc_mru;
2783 buf->b_arc_access = ddi_get_lbolt();
2784 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2785 arc_change_state(new_state, buf, hash_lock);
2787 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2788 } else if (buf->b_state == arc_l2c_only) {
2790 * This buffer is on the 2nd Level ARC.
2793 buf->b_arc_access = ddi_get_lbolt();
2794 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2795 arc_change_state(arc_mfu, buf, hash_lock);
2797 ASSERT(!"invalid arc state");
2801 /* a generic arc_done_func_t which you can use */
2804 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2806 if (zio == NULL || zio->io_error == 0)
2807 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2808 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2811 /* a generic arc_done_func_t */
2813 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2815 arc_buf_t **bufp = arg;
2816 if (zio && zio->io_error) {
2817 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2821 ASSERT(buf->b_data);
2826 arc_read_done(zio_t *zio)
2828 arc_buf_hdr_t *hdr, *found;
2830 arc_buf_t *abuf; /* buffer we're assigning to callback */
2831 kmutex_t *hash_lock;
2832 arc_callback_t *callback_list, *acb;
2833 int freeable = FALSE;
2835 buf = zio->io_private;
2839 * The hdr was inserted into hash-table and removed from lists
2840 * prior to starting I/O. We should find this header, since
2841 * it's in the hash table, and it should be legit since it's
2842 * not possible to evict it during the I/O. The only possible
2843 * reason for it not to be found is if we were freed during the
2846 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2849 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2850 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2851 (found == hdr && HDR_L2_READING(hdr)));
2853 hdr->b_flags &= ~ARC_L2_EVICTED;
2854 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2855 hdr->b_flags &= ~ARC_L2CACHE;
2857 /* byteswap if necessary */
2858 callback_list = hdr->b_acb;
2859 ASSERT(callback_list != NULL);
2860 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2861 dmu_object_byteswap_t bswap =
2862 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2863 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2864 byteswap_uint64_array :
2865 dmu_ot_byteswap[bswap].ob_func;
2866 func(buf->b_data, hdr->b_size);
2869 arc_cksum_compute(buf, B_FALSE);
2872 #endif /* illumos */
2874 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2876 * Only call arc_access on anonymous buffers. This is because
2877 * if we've issued an I/O for an evicted buffer, we've already
2878 * called arc_access (to prevent any simultaneous readers from
2879 * getting confused).
2881 arc_access(hdr, hash_lock);
2884 /* create copies of the data buffer for the callers */
2886 for (acb = callback_list; acb; acb = acb->acb_next) {
2887 if (acb->acb_done) {
2889 abuf = arc_buf_clone(buf);
2890 acb->acb_buf = abuf;
2895 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2896 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2898 ASSERT(buf->b_efunc == NULL);
2899 ASSERT(hdr->b_datacnt == 1);
2900 hdr->b_flags |= ARC_BUF_AVAILABLE;
2903 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2905 if (zio->io_error != 0) {
2906 hdr->b_flags |= ARC_IO_ERROR;
2907 if (hdr->b_state != arc_anon)
2908 arc_change_state(arc_anon, hdr, hash_lock);
2909 if (HDR_IN_HASH_TABLE(hdr))
2910 buf_hash_remove(hdr);
2911 freeable = refcount_is_zero(&hdr->b_refcnt);
2915 * Broadcast before we drop the hash_lock to avoid the possibility
2916 * that the hdr (and hence the cv) might be freed before we get to
2917 * the cv_broadcast().
2919 cv_broadcast(&hdr->b_cv);
2922 mutex_exit(hash_lock);
2925 * This block was freed while we waited for the read to
2926 * complete. It has been removed from the hash table and
2927 * moved to the anonymous state (so that it won't show up
2930 ASSERT3P(hdr->b_state, ==, arc_anon);
2931 freeable = refcount_is_zero(&hdr->b_refcnt);
2934 /* execute each callback and free its structure */
2935 while ((acb = callback_list) != NULL) {
2937 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2939 if (acb->acb_zio_dummy != NULL) {
2940 acb->acb_zio_dummy->io_error = zio->io_error;
2941 zio_nowait(acb->acb_zio_dummy);
2944 callback_list = acb->acb_next;
2945 kmem_free(acb, sizeof (arc_callback_t));
2949 arc_hdr_destroy(hdr);
2953 * "Read" the block block at the specified DVA (in bp) via the
2954 * cache. If the block is found in the cache, invoke the provided
2955 * callback immediately and return. Note that the `zio' parameter
2956 * in the callback will be NULL in this case, since no IO was
2957 * required. If the block is not in the cache pass the read request
2958 * on to the spa with a substitute callback function, so that the
2959 * requested block will be added to the cache.
2961 * If a read request arrives for a block that has a read in-progress,
2962 * either wait for the in-progress read to complete (and return the
2963 * results); or, if this is a read with a "done" func, add a record
2964 * to the read to invoke the "done" func when the read completes,
2965 * and return; or just return.
2967 * arc_read_done() will invoke all the requested "done" functions
2968 * for readers of this block.
2971 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
2972 void *private, int priority, int zio_flags, uint32_t *arc_flags,
2973 const zbookmark_t *zb)
2977 kmutex_t *hash_lock;
2979 uint64_t guid = spa_load_guid(spa);
2982 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2984 if (hdr && hdr->b_datacnt > 0) {
2986 *arc_flags |= ARC_CACHED;
2988 if (HDR_IO_IN_PROGRESS(hdr)) {
2990 if (*arc_flags & ARC_WAIT) {
2991 cv_wait(&hdr->b_cv, hash_lock);
2992 mutex_exit(hash_lock);
2995 ASSERT(*arc_flags & ARC_NOWAIT);
2998 arc_callback_t *acb = NULL;
3000 acb = kmem_zalloc(sizeof (arc_callback_t),
3002 acb->acb_done = done;
3003 acb->acb_private = private;
3005 acb->acb_zio_dummy = zio_null(pio,
3006 spa, NULL, NULL, NULL, zio_flags);
3008 ASSERT(acb->acb_done != NULL);
3009 acb->acb_next = hdr->b_acb;
3011 add_reference(hdr, hash_lock, private);
3012 mutex_exit(hash_lock);
3015 mutex_exit(hash_lock);
3019 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3022 add_reference(hdr, hash_lock, private);
3024 * If this block is already in use, create a new
3025 * copy of the data so that we will be guaranteed
3026 * that arc_release() will always succeed.
3030 ASSERT(buf->b_data);
3031 if (HDR_BUF_AVAILABLE(hdr)) {
3032 ASSERT(buf->b_efunc == NULL);
3033 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3035 buf = arc_buf_clone(buf);
3038 } else if (*arc_flags & ARC_PREFETCH &&
3039 refcount_count(&hdr->b_refcnt) == 0) {
3040 hdr->b_flags |= ARC_PREFETCH;
3042 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
3043 arc_access(hdr, hash_lock);
3044 if (*arc_flags & ARC_L2CACHE)
3045 hdr->b_flags |= ARC_L2CACHE;
3046 mutex_exit(hash_lock);
3047 ARCSTAT_BUMP(arcstat_hits);
3048 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3049 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3050 data, metadata, hits);
3053 done(NULL, buf, private);
3055 uint64_t size = BP_GET_LSIZE(bp);
3056 arc_callback_t *acb;
3059 boolean_t devw = B_FALSE;
3062 /* this block is not in the cache */
3063 arc_buf_hdr_t *exists;
3064 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
3065 buf = arc_buf_alloc(spa, size, private, type);
3067 hdr->b_dva = *BP_IDENTITY(bp);
3068 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3069 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3070 exists = buf_hash_insert(hdr, &hash_lock);
3072 /* somebody beat us to the hash insert */
3073 mutex_exit(hash_lock);
3074 buf_discard_identity(hdr);
3075 (void) arc_buf_remove_ref(buf, private);
3076 goto top; /* restart the IO request */
3078 /* if this is a prefetch, we don't have a reference */
3079 if (*arc_flags & ARC_PREFETCH) {
3080 (void) remove_reference(hdr, hash_lock,
3082 hdr->b_flags |= ARC_PREFETCH;
3084 if (*arc_flags & ARC_L2CACHE)
3085 hdr->b_flags |= ARC_L2CACHE;
3086 if (BP_GET_LEVEL(bp) > 0)
3087 hdr->b_flags |= ARC_INDIRECT;
3089 /* this block is in the ghost cache */
3090 ASSERT(GHOST_STATE(hdr->b_state));
3091 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3092 ASSERT0(refcount_count(&hdr->b_refcnt));
3093 ASSERT(hdr->b_buf == NULL);
3095 /* if this is a prefetch, we don't have a reference */
3096 if (*arc_flags & ARC_PREFETCH)
3097 hdr->b_flags |= ARC_PREFETCH;
3099 add_reference(hdr, hash_lock, private);
3100 if (*arc_flags & ARC_L2CACHE)
3101 hdr->b_flags |= ARC_L2CACHE;
3102 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3105 buf->b_efunc = NULL;
3106 buf->b_private = NULL;
3109 ASSERT(hdr->b_datacnt == 0);
3111 arc_get_data_buf(buf);
3112 arc_access(hdr, hash_lock);
3115 ASSERT(!GHOST_STATE(hdr->b_state));
3117 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
3118 acb->acb_done = done;
3119 acb->acb_private = private;
3121 ASSERT(hdr->b_acb == NULL);
3123 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3125 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3126 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3127 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3128 addr = hdr->b_l2hdr->b_daddr;
3130 * Lock out device removal.
3132 if (vdev_is_dead(vd) ||
3133 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3137 mutex_exit(hash_lock);
3139 ASSERT3U(hdr->b_size, ==, size);
3140 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3141 uint64_t, size, zbookmark_t *, zb);
3142 ARCSTAT_BUMP(arcstat_misses);
3143 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3144 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3145 data, metadata, misses);
3147 curthread->td_ru.ru_inblock++;
3150 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3152 * Read from the L2ARC if the following are true:
3153 * 1. The L2ARC vdev was previously cached.
3154 * 2. This buffer still has L2ARC metadata.
3155 * 3. This buffer isn't currently writing to the L2ARC.
3156 * 4. The L2ARC entry wasn't evicted, which may
3157 * also have invalidated the vdev.
3158 * 5. This isn't prefetch and l2arc_noprefetch is set.
3160 if (hdr->b_l2hdr != NULL &&
3161 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3162 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3163 l2arc_read_callback_t *cb;
3165 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3166 ARCSTAT_BUMP(arcstat_l2_hits);
3168 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3170 cb->l2rcb_buf = buf;
3171 cb->l2rcb_spa = spa;
3174 cb->l2rcb_flags = zio_flags;
3177 * l2arc read. The SCL_L2ARC lock will be
3178 * released by l2arc_read_done().
3180 rzio = zio_read_phys(pio, vd, addr, size,
3181 buf->b_data, ZIO_CHECKSUM_OFF,
3182 l2arc_read_done, cb, priority, zio_flags |
3183 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
3184 ZIO_FLAG_DONT_PROPAGATE |
3185 ZIO_FLAG_DONT_RETRY, B_FALSE);
3186 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3188 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
3190 if (*arc_flags & ARC_NOWAIT) {
3195 ASSERT(*arc_flags & ARC_WAIT);
3196 if (zio_wait(rzio) == 0)
3199 /* l2arc read error; goto zio_read() */
3201 DTRACE_PROBE1(l2arc__miss,
3202 arc_buf_hdr_t *, hdr);
3203 ARCSTAT_BUMP(arcstat_l2_misses);
3204 if (HDR_L2_WRITING(hdr))
3205 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3206 spa_config_exit(spa, SCL_L2ARC, vd);
3210 spa_config_exit(spa, SCL_L2ARC, vd);
3211 if (l2arc_ndev != 0) {
3212 DTRACE_PROBE1(l2arc__miss,
3213 arc_buf_hdr_t *, hdr);
3214 ARCSTAT_BUMP(arcstat_l2_misses);
3218 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3219 arc_read_done, buf, priority, zio_flags, zb);
3221 if (*arc_flags & ARC_WAIT)
3222 return (zio_wait(rzio));
3224 ASSERT(*arc_flags & ARC_NOWAIT);
3231 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3233 ASSERT(buf->b_hdr != NULL);
3234 ASSERT(buf->b_hdr->b_state != arc_anon);
3235 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3236 ASSERT(buf->b_efunc == NULL);
3237 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3239 buf->b_efunc = func;
3240 buf->b_private = private;
3244 * This is used by the DMU to let the ARC know that a buffer is
3245 * being evicted, so the ARC should clean up. If this arc buf
3246 * is not yet in the evicted state, it will be put there.
3249 arc_buf_evict(arc_buf_t *buf)
3252 kmutex_t *hash_lock;
3254 list_t *list, *evicted_list;
3255 kmutex_t *lock, *evicted_lock;
3257 mutex_enter(&buf->b_evict_lock);
3261 * We are in arc_do_user_evicts().
3263 ASSERT(buf->b_data == NULL);
3264 mutex_exit(&buf->b_evict_lock);
3266 } else if (buf->b_data == NULL) {
3267 arc_buf_t copy = *buf; /* structure assignment */
3269 * We are on the eviction list; process this buffer now
3270 * but let arc_do_user_evicts() do the reaping.
3272 buf->b_efunc = NULL;
3273 mutex_exit(&buf->b_evict_lock);
3274 VERIFY(copy.b_efunc(©) == 0);
3277 hash_lock = HDR_LOCK(hdr);
3278 mutex_enter(hash_lock);
3280 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3282 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3283 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3286 * Pull this buffer off of the hdr
3289 while (*bufp != buf)
3290 bufp = &(*bufp)->b_next;
3291 *bufp = buf->b_next;
3293 ASSERT(buf->b_data != NULL);
3294 arc_buf_destroy(buf, FALSE, FALSE);
3296 if (hdr->b_datacnt == 0) {
3297 arc_state_t *old_state = hdr->b_state;
3298 arc_state_t *evicted_state;
3300 ASSERT(hdr->b_buf == NULL);
3301 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3304 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3306 get_buf_info(hdr, old_state, &list, &lock);
3307 get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock);
3309 mutex_enter(evicted_lock);
3311 arc_change_state(evicted_state, hdr, hash_lock);
3312 ASSERT(HDR_IN_HASH_TABLE(hdr));
3313 hdr->b_flags |= ARC_IN_HASH_TABLE;
3314 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3316 mutex_exit(evicted_lock);
3319 mutex_exit(hash_lock);
3320 mutex_exit(&buf->b_evict_lock);
3322 VERIFY(buf->b_efunc(buf) == 0);
3323 buf->b_efunc = NULL;
3324 buf->b_private = NULL;
3327 kmem_cache_free(buf_cache, buf);
3332 * Release this buffer from the cache. This must be done
3333 * after a read and prior to modifying the buffer contents.
3334 * If the buffer has more than one reference, we must make
3335 * a new hdr for the buffer.
3338 arc_release(arc_buf_t *buf, void *tag)
3341 kmutex_t *hash_lock = NULL;
3342 l2arc_buf_hdr_t *l2hdr;
3346 * It would be nice to assert that if it's DMU metadata (level >
3347 * 0 || it's the dnode file), then it must be syncing context.
3348 * But we don't know that information at this level.
3351 mutex_enter(&buf->b_evict_lock);
3354 /* this buffer is not on any list */
3355 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3357 if (hdr->b_state == arc_anon) {
3358 /* this buffer is already released */
3359 ASSERT(buf->b_efunc == NULL);
3361 hash_lock = HDR_LOCK(hdr);
3362 mutex_enter(hash_lock);
3364 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3367 l2hdr = hdr->b_l2hdr;
3369 mutex_enter(&l2arc_buflist_mtx);
3370 hdr->b_l2hdr = NULL;
3371 buf_size = hdr->b_size;
3375 * Do we have more than one buf?
3377 if (hdr->b_datacnt > 1) {
3378 arc_buf_hdr_t *nhdr;
3380 uint64_t blksz = hdr->b_size;
3381 uint64_t spa = hdr->b_spa;
3382 arc_buf_contents_t type = hdr->b_type;
3383 uint32_t flags = hdr->b_flags;
3385 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3387 * Pull the data off of this hdr and attach it to
3388 * a new anonymous hdr.
3390 (void) remove_reference(hdr, hash_lock, tag);
3392 while (*bufp != buf)
3393 bufp = &(*bufp)->b_next;
3394 *bufp = buf->b_next;
3397 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3398 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3399 if (refcount_is_zero(&hdr->b_refcnt)) {
3400 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3401 ASSERT3U(*size, >=, hdr->b_size);
3402 atomic_add_64(size, -hdr->b_size);
3404 hdr->b_datacnt -= 1;
3405 arc_cksum_verify(buf);
3407 arc_buf_unwatch(buf);
3408 #endif /* illumos */
3410 mutex_exit(hash_lock);
3412 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3413 nhdr->b_size = blksz;
3415 nhdr->b_type = type;
3417 nhdr->b_state = arc_anon;
3418 nhdr->b_arc_access = 0;
3419 nhdr->b_flags = flags & ARC_L2_WRITING;
3420 nhdr->b_l2hdr = NULL;
3421 nhdr->b_datacnt = 1;
3422 nhdr->b_freeze_cksum = NULL;
3423 (void) refcount_add(&nhdr->b_refcnt, tag);
3425 mutex_exit(&buf->b_evict_lock);
3426 atomic_add_64(&arc_anon->arcs_size, blksz);
3428 mutex_exit(&buf->b_evict_lock);
3429 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3430 ASSERT(!list_link_active(&hdr->b_arc_node));
3431 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3432 if (hdr->b_state != arc_anon)
3433 arc_change_state(arc_anon, hdr, hash_lock);
3434 hdr->b_arc_access = 0;
3436 mutex_exit(hash_lock);
3438 buf_discard_identity(hdr);
3441 buf->b_efunc = NULL;
3442 buf->b_private = NULL;
3445 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3446 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3447 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3448 mutex_exit(&l2arc_buflist_mtx);
3453 arc_released(arc_buf_t *buf)
3457 mutex_enter(&buf->b_evict_lock);
3458 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3459 mutex_exit(&buf->b_evict_lock);
3464 arc_has_callback(arc_buf_t *buf)
3468 mutex_enter(&buf->b_evict_lock);
3469 callback = (buf->b_efunc != NULL);
3470 mutex_exit(&buf->b_evict_lock);
3476 arc_referenced(arc_buf_t *buf)
3480 mutex_enter(&buf->b_evict_lock);
3481 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3482 mutex_exit(&buf->b_evict_lock);
3483 return (referenced);
3488 arc_write_ready(zio_t *zio)
3490 arc_write_callback_t *callback = zio->io_private;
3491 arc_buf_t *buf = callback->awcb_buf;
3492 arc_buf_hdr_t *hdr = buf->b_hdr;
3494 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3495 callback->awcb_ready(zio, buf, callback->awcb_private);
3498 * If the IO is already in progress, then this is a re-write
3499 * attempt, so we need to thaw and re-compute the cksum.
3500 * It is the responsibility of the callback to handle the
3501 * accounting for any re-write attempt.
3503 if (HDR_IO_IN_PROGRESS(hdr)) {
3504 mutex_enter(&hdr->b_freeze_lock);
3505 if (hdr->b_freeze_cksum != NULL) {
3506 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3507 hdr->b_freeze_cksum = NULL;
3509 mutex_exit(&hdr->b_freeze_lock);
3511 arc_cksum_compute(buf, B_FALSE);
3512 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3516 arc_write_done(zio_t *zio)
3518 arc_write_callback_t *callback = zio->io_private;
3519 arc_buf_t *buf = callback->awcb_buf;
3520 arc_buf_hdr_t *hdr = buf->b_hdr;
3522 ASSERT(hdr->b_acb == NULL);
3524 if (zio->io_error == 0) {
3525 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3526 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3527 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3529 ASSERT(BUF_EMPTY(hdr));
3533 * If the block to be written was all-zero, we may have
3534 * compressed it away. In this case no write was performed
3535 * so there will be no dva/birth/checksum. The buffer must
3536 * therefore remain anonymous (and uncached).
3538 if (!BUF_EMPTY(hdr)) {
3539 arc_buf_hdr_t *exists;
3540 kmutex_t *hash_lock;
3542 ASSERT(zio->io_error == 0);
3544 arc_cksum_verify(buf);
3546 exists = buf_hash_insert(hdr, &hash_lock);
3549 * This can only happen if we overwrite for
3550 * sync-to-convergence, because we remove
3551 * buffers from the hash table when we arc_free().
3553 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3554 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3555 panic("bad overwrite, hdr=%p exists=%p",
3556 (void *)hdr, (void *)exists);
3557 ASSERT(refcount_is_zero(&exists->b_refcnt));
3558 arc_change_state(arc_anon, exists, hash_lock);
3559 mutex_exit(hash_lock);
3560 arc_hdr_destroy(exists);
3561 exists = buf_hash_insert(hdr, &hash_lock);
3562 ASSERT3P(exists, ==, NULL);
3563 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3565 ASSERT(zio->io_prop.zp_nopwrite);
3566 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3567 panic("bad nopwrite, hdr=%p exists=%p",
3568 (void *)hdr, (void *)exists);
3571 ASSERT(hdr->b_datacnt == 1);
3572 ASSERT(hdr->b_state == arc_anon);
3573 ASSERT(BP_GET_DEDUP(zio->io_bp));
3574 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3577 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3578 /* if it's not anon, we are doing a scrub */
3579 if (!exists && hdr->b_state == arc_anon)
3580 arc_access(hdr, hash_lock);
3581 mutex_exit(hash_lock);
3583 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3586 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3587 callback->awcb_done(zio, buf, callback->awcb_private);
3589 kmem_free(callback, sizeof (arc_write_callback_t));
3593 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3594 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3595 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3596 int priority, int zio_flags, const zbookmark_t *zb)
3598 arc_buf_hdr_t *hdr = buf->b_hdr;
3599 arc_write_callback_t *callback;
3602 ASSERT(ready != NULL);
3603 ASSERT(done != NULL);
3604 ASSERT(!HDR_IO_ERROR(hdr));
3605 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3606 ASSERT(hdr->b_acb == NULL);
3608 hdr->b_flags |= ARC_L2CACHE;
3609 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3610 callback->awcb_ready = ready;
3611 callback->awcb_done = done;
3612 callback->awcb_private = private;
3613 callback->awcb_buf = buf;
3615 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3616 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3622 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3625 uint64_t available_memory =
3626 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count);
3627 static uint64_t page_load = 0;
3628 static uint64_t last_txg = 0;
3633 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3636 if (available_memory >= zfs_write_limit_max)
3639 if (txg > last_txg) {
3644 * If we are in pageout, we know that memory is already tight,
3645 * the arc is already going to be evicting, so we just want to
3646 * continue to let page writes occur as quickly as possible.
3648 if (curproc == pageproc) {
3649 if (page_load > available_memory / 4)
3651 /* Note: reserve is inflated, so we deflate */
3652 page_load += reserve / 8;
3654 } else if (page_load > 0 && arc_reclaim_needed()) {
3655 /* memory is low, delay before restarting */
3656 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3661 if (arc_size > arc_c_min) {
3662 uint64_t evictable_memory =
3663 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3664 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3665 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3666 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3667 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3670 if (inflight_data > available_memory / 4) {
3671 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3679 arc_tempreserve_clear(uint64_t reserve)
3681 atomic_add_64(&arc_tempreserve, -reserve);
3682 ASSERT((int64_t)arc_tempreserve >= 0);
3686 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3693 * Once in a while, fail for no reason. Everything should cope.
3695 if (spa_get_random(10000) == 0) {
3696 dprintf("forcing random failure\n");
3700 if (reserve > arc_c/4 && !arc_no_grow)
3701 arc_c = MIN(arc_c_max, reserve * 4);
3702 if (reserve > arc_c)
3706 * Don't count loaned bufs as in flight dirty data to prevent long
3707 * network delays from blocking transactions that are ready to be
3708 * assigned to a txg.
3710 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3713 * Writes will, almost always, require additional memory allocations
3714 * in order to compress/encrypt/etc the data. We therefor need to
3715 * make sure that there is sufficient available memory for this.
3717 if (error = arc_memory_throttle(reserve, anon_size, txg))
3721 * Throttle writes when the amount of dirty data in the cache
3722 * gets too large. We try to keep the cache less than half full
3723 * of dirty blocks so that our sync times don't grow too large.
3724 * Note: if two requests come in concurrently, we might let them
3725 * both succeed, when one of them should fail. Not a huge deal.
3728 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3729 anon_size > arc_c / 4) {
3730 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3731 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3732 arc_tempreserve>>10,
3733 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3734 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3735 reserve>>10, arc_c>>10);
3738 atomic_add_64(&arc_tempreserve, reserve);
3742 static kmutex_t arc_lowmem_lock;
3744 static eventhandler_tag arc_event_lowmem = NULL;
3747 arc_lowmem(void *arg __unused, int howto __unused)
3750 /* Serialize access via arc_lowmem_lock. */
3751 mutex_enter(&arc_lowmem_lock);
3752 mutex_enter(&arc_reclaim_thr_lock);
3754 cv_signal(&arc_reclaim_thr_cv);
3757 * It is unsafe to block here in arbitrary threads, because we can come
3758 * here from ARC itself and may hold ARC locks and thus risk a deadlock
3759 * with ARC reclaim thread.
3761 if (curproc == pageproc) {
3763 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0);
3765 mutex_exit(&arc_reclaim_thr_lock);
3766 mutex_exit(&arc_lowmem_lock);
3773 int i, prefetch_tunable_set = 0;
3775 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3776 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3777 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL);
3779 /* Convert seconds to clock ticks */
3780 arc_min_prefetch_lifespan = 1 * hz;
3782 /* Start out with 1/8 of all memory */
3783 arc_c = kmem_size() / 8;
3788 * On architectures where the physical memory can be larger
3789 * than the addressable space (intel in 32-bit mode), we may
3790 * need to limit the cache to 1/8 of VM size.
3792 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3795 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
3796 arc_c_min = MAX(arc_c / 4, 64<<18);
3797 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
3798 if (arc_c * 8 >= 1<<30)
3799 arc_c_max = (arc_c * 8) - (1<<30);
3801 arc_c_max = arc_c_min;
3802 arc_c_max = MAX(arc_c * 5, arc_c_max);
3806 * Allow the tunables to override our calculations if they are
3807 * reasonable (ie. over 16MB)
3809 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size())
3810 arc_c_max = zfs_arc_max;
3811 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max)
3812 arc_c_min = zfs_arc_min;
3816 arc_p = (arc_c >> 1);
3818 /* limit meta-data to 1/4 of the arc capacity */
3819 arc_meta_limit = arc_c_max / 4;
3821 /* Allow the tunable to override if it is reasonable */
3822 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3823 arc_meta_limit = zfs_arc_meta_limit;
3825 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3826 arc_c_min = arc_meta_limit / 2;
3828 if (zfs_arc_grow_retry > 0)
3829 arc_grow_retry = zfs_arc_grow_retry;
3831 if (zfs_arc_shrink_shift > 0)
3832 arc_shrink_shift = zfs_arc_shrink_shift;
3834 if (zfs_arc_p_min_shift > 0)
3835 arc_p_min_shift = zfs_arc_p_min_shift;
3837 /* if kmem_flags are set, lets try to use less memory */
3838 if (kmem_debugging())
3840 if (arc_c < arc_c_min)
3843 zfs_arc_min = arc_c_min;
3844 zfs_arc_max = arc_c_max;
3846 arc_anon = &ARC_anon;
3848 arc_mru_ghost = &ARC_mru_ghost;
3850 arc_mfu_ghost = &ARC_mfu_ghost;
3851 arc_l2c_only = &ARC_l2c_only;
3854 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
3855 mutex_init(&arc_anon->arcs_locks[i].arcs_lock,
3856 NULL, MUTEX_DEFAULT, NULL);
3857 mutex_init(&arc_mru->arcs_locks[i].arcs_lock,
3858 NULL, MUTEX_DEFAULT, NULL);
3859 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock,
3860 NULL, MUTEX_DEFAULT, NULL);
3861 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock,
3862 NULL, MUTEX_DEFAULT, NULL);
3863 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock,
3864 NULL, MUTEX_DEFAULT, NULL);
3865 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock,
3866 NULL, MUTEX_DEFAULT, NULL);
3868 list_create(&arc_mru->arcs_lists[i],
3869 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3870 list_create(&arc_mru_ghost->arcs_lists[i],
3871 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3872 list_create(&arc_mfu->arcs_lists[i],
3873 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3874 list_create(&arc_mfu_ghost->arcs_lists[i],
3875 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3876 list_create(&arc_mfu_ghost->arcs_lists[i],
3877 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3878 list_create(&arc_l2c_only->arcs_lists[i],
3879 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3884 arc_thread_exit = 0;
3885 arc_eviction_list = NULL;
3886 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3887 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3889 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3890 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3892 if (arc_ksp != NULL) {
3893 arc_ksp->ks_data = &arc_stats;
3894 kstat_install(arc_ksp);
3897 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3898 TS_RUN, minclsyspri);
3901 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
3902 EVENTHANDLER_PRI_FIRST);
3908 if (zfs_write_limit_max == 0)
3909 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3911 zfs_write_limit_shift = 0;
3912 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3915 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
3916 prefetch_tunable_set = 1;
3919 if (prefetch_tunable_set == 0) {
3920 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
3922 printf(" add \"vfs.zfs.prefetch_disable=0\" "
3923 "to /boot/loader.conf.\n");
3924 zfs_prefetch_disable = 1;
3927 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
3928 prefetch_tunable_set == 0) {
3929 printf("ZFS NOTICE: Prefetch is disabled by default if less "
3930 "than 4GB of RAM is present;\n"
3931 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
3932 "to /boot/loader.conf.\n");
3933 zfs_prefetch_disable = 1;
3936 /* Warn about ZFS memory and address space requirements. */
3937 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
3938 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
3939 "expect unstable behavior.\n");
3941 if (kmem_size() < 512 * (1 << 20)) {
3942 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
3943 "expect unstable behavior.\n");
3944 printf(" Consider tuning vm.kmem_size and "
3945 "vm.kmem_size_max\n");
3946 printf(" in /boot/loader.conf.\n");
3956 mutex_enter(&arc_reclaim_thr_lock);
3957 arc_thread_exit = 1;
3958 cv_signal(&arc_reclaim_thr_cv);
3959 while (arc_thread_exit != 0)
3960 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3961 mutex_exit(&arc_reclaim_thr_lock);
3967 if (arc_ksp != NULL) {
3968 kstat_delete(arc_ksp);
3972 mutex_destroy(&arc_eviction_mtx);
3973 mutex_destroy(&arc_reclaim_thr_lock);
3974 cv_destroy(&arc_reclaim_thr_cv);
3976 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
3977 list_destroy(&arc_mru->arcs_lists[i]);
3978 list_destroy(&arc_mru_ghost->arcs_lists[i]);
3979 list_destroy(&arc_mfu->arcs_lists[i]);
3980 list_destroy(&arc_mfu_ghost->arcs_lists[i]);
3981 list_destroy(&arc_l2c_only->arcs_lists[i]);
3983 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock);
3984 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock);
3985 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock);
3986 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock);
3987 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock);
3988 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock);
3991 mutex_destroy(&zfs_write_limit_lock);
3995 ASSERT(arc_loaned_bytes == 0);
3997 mutex_destroy(&arc_lowmem_lock);
3999 if (arc_event_lowmem != NULL)
4000 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
4007 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
4008 * It uses dedicated storage devices to hold cached data, which are populated
4009 * using large infrequent writes. The main role of this cache is to boost
4010 * the performance of random read workloads. The intended L2ARC devices
4011 * include short-stroked disks, solid state disks, and other media with
4012 * substantially faster read latency than disk.
4014 * +-----------------------+
4016 * +-----------------------+
4019 * l2arc_feed_thread() arc_read()
4023 * +---------------+ |
4025 * +---------------+ |
4030 * +-------+ +-------+
4032 * | cache | | cache |
4033 * +-------+ +-------+
4034 * +=========+ .-----.
4035 * : L2ARC : |-_____-|
4036 * : devices : | Disks |
4037 * +=========+ `-_____-'
4039 * Read requests are satisfied from the following sources, in order:
4042 * 2) vdev cache of L2ARC devices
4044 * 4) vdev cache of disks
4047 * Some L2ARC device types exhibit extremely slow write performance.
4048 * To accommodate for this there are some significant differences between
4049 * the L2ARC and traditional cache design:
4051 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4052 * the ARC behave as usual, freeing buffers and placing headers on ghost
4053 * lists. The ARC does not send buffers to the L2ARC during eviction as
4054 * this would add inflated write latencies for all ARC memory pressure.
4056 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4057 * It does this by periodically scanning buffers from the eviction-end of
4058 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4059 * not already there. It scans until a headroom of buffers is satisfied,
4060 * which itself is a buffer for ARC eviction. The thread that does this is
4061 * l2arc_feed_thread(), illustrated below; example sizes are included to
4062 * provide a better sense of ratio than this diagram:
4065 * +---------------------+----------+
4066 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4067 * +---------------------+----------+ | o L2ARC eligible
4068 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4069 * +---------------------+----------+ |
4070 * 15.9 Gbytes ^ 32 Mbytes |
4072 * l2arc_feed_thread()
4074 * l2arc write hand <--[oooo]--'
4078 * +==============================+
4079 * L2ARC dev |####|#|###|###| |####| ... |
4080 * +==============================+
4083 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4084 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4085 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4086 * safe to say that this is an uncommon case, since buffers at the end of
4087 * the ARC lists have moved there due to inactivity.
4089 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4090 * then the L2ARC simply misses copying some buffers. This serves as a
4091 * pressure valve to prevent heavy read workloads from both stalling the ARC
4092 * with waits and clogging the L2ARC with writes. This also helps prevent
4093 * the potential for the L2ARC to churn if it attempts to cache content too
4094 * quickly, such as during backups of the entire pool.
4096 * 5. After system boot and before the ARC has filled main memory, there are
4097 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4098 * lists can remain mostly static. Instead of searching from tail of these
4099 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4100 * for eligible buffers, greatly increasing its chance of finding them.
4102 * The L2ARC device write speed is also boosted during this time so that
4103 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4104 * there are no L2ARC reads, and no fear of degrading read performance
4105 * through increased writes.
4107 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4108 * the vdev queue can aggregate them into larger and fewer writes. Each
4109 * device is written to in a rotor fashion, sweeping writes through
4110 * available space then repeating.
4112 * 7. The L2ARC does not store dirty content. It never needs to flush
4113 * write buffers back to disk based storage.
4115 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4116 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4118 * The performance of the L2ARC can be tweaked by a number of tunables, which
4119 * may be necessary for different workloads:
4121 * l2arc_write_max max write bytes per interval
4122 * l2arc_write_boost extra write bytes during device warmup
4123 * l2arc_noprefetch skip caching prefetched buffers
4124 * l2arc_headroom number of max device writes to precache
4125 * l2arc_feed_secs seconds between L2ARC writing
4127 * Tunables may be removed or added as future performance improvements are
4128 * integrated, and also may become zpool properties.
4130 * There are three key functions that control how the L2ARC warms up:
4132 * l2arc_write_eligible() check if a buffer is eligible to cache
4133 * l2arc_write_size() calculate how much to write
4134 * l2arc_write_interval() calculate sleep delay between writes
4136 * These three functions determine what to write, how much, and how quickly
4141 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4144 * A buffer is *not* eligible for the L2ARC if it:
4145 * 1. belongs to a different spa.
4146 * 2. is already cached on the L2ARC.
4147 * 3. has an I/O in progress (it may be an incomplete read).
4148 * 4. is flagged not eligible (zfs property).
4150 if (ab->b_spa != spa_guid) {
4151 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
4154 if (ab->b_l2hdr != NULL) {
4155 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
4158 if (HDR_IO_IN_PROGRESS(ab)) {
4159 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
4162 if (!HDR_L2CACHE(ab)) {
4163 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
4171 l2arc_write_size(l2arc_dev_t *dev)
4175 size = dev->l2ad_write;
4177 if (arc_warm == B_FALSE)
4178 size += dev->l2ad_boost;
4185 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4187 clock_t interval, next, now;
4190 * If the ARC lists are busy, increase our write rate; if the
4191 * lists are stale, idle back. This is achieved by checking
4192 * how much we previously wrote - if it was more than half of
4193 * what we wanted, schedule the next write much sooner.
4195 if (l2arc_feed_again && wrote > (wanted / 2))
4196 interval = (hz * l2arc_feed_min_ms) / 1000;
4198 interval = hz * l2arc_feed_secs;
4200 now = ddi_get_lbolt();
4201 next = MAX(now, MIN(now + interval, began + interval));
4207 l2arc_hdr_stat_add(void)
4209 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4210 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4214 l2arc_hdr_stat_remove(void)
4216 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4217 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4221 * Cycle through L2ARC devices. This is how L2ARC load balances.
4222 * If a device is returned, this also returns holding the spa config lock.
4224 static l2arc_dev_t *
4225 l2arc_dev_get_next(void)
4227 l2arc_dev_t *first, *next = NULL;
4230 * Lock out the removal of spas (spa_namespace_lock), then removal
4231 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4232 * both locks will be dropped and a spa config lock held instead.
4234 mutex_enter(&spa_namespace_lock);
4235 mutex_enter(&l2arc_dev_mtx);
4237 /* if there are no vdevs, there is nothing to do */
4238 if (l2arc_ndev == 0)
4242 next = l2arc_dev_last;
4244 /* loop around the list looking for a non-faulted vdev */
4246 next = list_head(l2arc_dev_list);
4248 next = list_next(l2arc_dev_list, next);
4250 next = list_head(l2arc_dev_list);
4253 /* if we have come back to the start, bail out */
4256 else if (next == first)
4259 } while (vdev_is_dead(next->l2ad_vdev));
4261 /* if we were unable to find any usable vdevs, return NULL */
4262 if (vdev_is_dead(next->l2ad_vdev))
4265 l2arc_dev_last = next;
4268 mutex_exit(&l2arc_dev_mtx);
4271 * Grab the config lock to prevent the 'next' device from being
4272 * removed while we are writing to it.
4275 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4276 mutex_exit(&spa_namespace_lock);
4282 * Free buffers that were tagged for destruction.
4285 l2arc_do_free_on_write()
4288 l2arc_data_free_t *df, *df_prev;
4290 mutex_enter(&l2arc_free_on_write_mtx);
4291 buflist = l2arc_free_on_write;
4293 for (df = list_tail(buflist); df; df = df_prev) {
4294 df_prev = list_prev(buflist, df);
4295 ASSERT(df->l2df_data != NULL);
4296 ASSERT(df->l2df_func != NULL);
4297 df->l2df_func(df->l2df_data, df->l2df_size);
4298 list_remove(buflist, df);
4299 kmem_free(df, sizeof (l2arc_data_free_t));
4302 mutex_exit(&l2arc_free_on_write_mtx);
4306 * A write to a cache device has completed. Update all headers to allow
4307 * reads from these buffers to begin.
4310 l2arc_write_done(zio_t *zio)
4312 l2arc_write_callback_t *cb;
4315 arc_buf_hdr_t *head, *ab, *ab_prev;
4316 l2arc_buf_hdr_t *abl2;
4317 kmutex_t *hash_lock;
4319 cb = zio->io_private;
4321 dev = cb->l2wcb_dev;
4322 ASSERT(dev != NULL);
4323 head = cb->l2wcb_head;
4324 ASSERT(head != NULL);
4325 buflist = dev->l2ad_buflist;
4326 ASSERT(buflist != NULL);
4327 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4328 l2arc_write_callback_t *, cb);
4330 if (zio->io_error != 0)
4331 ARCSTAT_BUMP(arcstat_l2_writes_error);
4333 mutex_enter(&l2arc_buflist_mtx);
4336 * All writes completed, or an error was hit.
4338 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4339 ab_prev = list_prev(buflist, ab);
4341 hash_lock = HDR_LOCK(ab);
4342 if (!mutex_tryenter(hash_lock)) {
4344 * This buffer misses out. It may be in a stage
4345 * of eviction. Its ARC_L2_WRITING flag will be
4346 * left set, denying reads to this buffer.
4348 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4352 if (zio->io_error != 0) {
4354 * Error - drop L2ARC entry.
4356 list_remove(buflist, ab);
4359 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4360 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4364 * Allow ARC to begin reads to this L2ARC entry.
4366 ab->b_flags &= ~ARC_L2_WRITING;
4368 mutex_exit(hash_lock);
4371 atomic_inc_64(&l2arc_writes_done);
4372 list_remove(buflist, head);
4373 kmem_cache_free(hdr_cache, head);
4374 mutex_exit(&l2arc_buflist_mtx);
4376 l2arc_do_free_on_write();
4378 kmem_free(cb, sizeof (l2arc_write_callback_t));
4382 * A read to a cache device completed. Validate buffer contents before
4383 * handing over to the regular ARC routines.
4386 l2arc_read_done(zio_t *zio)
4388 l2arc_read_callback_t *cb;
4391 kmutex_t *hash_lock;
4394 ASSERT(zio->io_vd != NULL);
4395 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4397 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4399 cb = zio->io_private;
4401 buf = cb->l2rcb_buf;
4402 ASSERT(buf != NULL);
4404 hash_lock = HDR_LOCK(buf->b_hdr);
4405 mutex_enter(hash_lock);
4407 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4410 * Check this survived the L2ARC journey.
4412 equal = arc_cksum_equal(buf);
4413 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4414 mutex_exit(hash_lock);
4415 zio->io_private = buf;
4416 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4417 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4420 mutex_exit(hash_lock);
4422 * Buffer didn't survive caching. Increment stats and
4423 * reissue to the original storage device.
4425 if (zio->io_error != 0) {
4426 ARCSTAT_BUMP(arcstat_l2_io_error);
4428 zio->io_error = EIO;
4431 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4434 * If there's no waiter, issue an async i/o to the primary
4435 * storage now. If there *is* a waiter, the caller must
4436 * issue the i/o in a context where it's OK to block.
4438 if (zio->io_waiter == NULL) {
4439 zio_t *pio = zio_unique_parent(zio);
4441 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4443 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4444 buf->b_data, zio->io_size, arc_read_done, buf,
4445 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4449 kmem_free(cb, sizeof (l2arc_read_callback_t));
4453 * This is the list priority from which the L2ARC will search for pages to
4454 * cache. This is used within loops (0..3) to cycle through lists in the
4455 * desired order. This order can have a significant effect on cache
4458 * Currently the metadata lists are hit first, MFU then MRU, followed by
4459 * the data lists. This function returns a locked list, and also returns
4463 l2arc_list_locked(int list_num, kmutex_t **lock)
4468 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS);
4470 if (list_num < ARC_BUFC_NUMMETADATALISTS) {
4472 list = &arc_mfu->arcs_lists[idx];
4473 *lock = ARCS_LOCK(arc_mfu, idx);
4474 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) {
4475 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4476 list = &arc_mru->arcs_lists[idx];
4477 *lock = ARCS_LOCK(arc_mru, idx);
4478 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 +
4479 ARC_BUFC_NUMDATALISTS)) {
4480 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4481 list = &arc_mfu->arcs_lists[idx];
4482 *lock = ARCS_LOCK(arc_mfu, idx);
4484 idx = list_num - ARC_BUFC_NUMLISTS;
4485 list = &arc_mru->arcs_lists[idx];
4486 *lock = ARCS_LOCK(arc_mru, idx);
4489 ASSERT(!(MUTEX_HELD(*lock)));
4495 * Evict buffers from the device write hand to the distance specified in
4496 * bytes. This distance may span populated buffers, it may span nothing.
4497 * This is clearing a region on the L2ARC device ready for writing.
4498 * If the 'all' boolean is set, every buffer is evicted.
4501 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4504 l2arc_buf_hdr_t *abl2;
4505 arc_buf_hdr_t *ab, *ab_prev;
4506 kmutex_t *hash_lock;
4509 buflist = dev->l2ad_buflist;
4511 if (buflist == NULL)
4514 if (!all && dev->l2ad_first) {
4516 * This is the first sweep through the device. There is
4522 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4524 * When nearing the end of the device, evict to the end
4525 * before the device write hand jumps to the start.
4527 taddr = dev->l2ad_end;
4529 taddr = dev->l2ad_hand + distance;
4531 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4532 uint64_t, taddr, boolean_t, all);
4535 mutex_enter(&l2arc_buflist_mtx);
4536 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4537 ab_prev = list_prev(buflist, ab);
4539 hash_lock = HDR_LOCK(ab);
4540 if (!mutex_tryenter(hash_lock)) {
4542 * Missed the hash lock. Retry.
4544 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4545 mutex_exit(&l2arc_buflist_mtx);
4546 mutex_enter(hash_lock);
4547 mutex_exit(hash_lock);
4551 if (HDR_L2_WRITE_HEAD(ab)) {
4553 * We hit a write head node. Leave it for
4554 * l2arc_write_done().
4556 list_remove(buflist, ab);
4557 mutex_exit(hash_lock);
4561 if (!all && ab->b_l2hdr != NULL &&
4562 (ab->b_l2hdr->b_daddr > taddr ||
4563 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4565 * We've evicted to the target address,
4566 * or the end of the device.
4568 mutex_exit(hash_lock);
4572 if (HDR_FREE_IN_PROGRESS(ab)) {
4574 * Already on the path to destruction.
4576 mutex_exit(hash_lock);
4580 if (ab->b_state == arc_l2c_only) {
4581 ASSERT(!HDR_L2_READING(ab));
4583 * This doesn't exist in the ARC. Destroy.
4584 * arc_hdr_destroy() will call list_remove()
4585 * and decrement arcstat_l2_size.
4587 arc_change_state(arc_anon, ab, hash_lock);
4588 arc_hdr_destroy(ab);
4591 * Invalidate issued or about to be issued
4592 * reads, since we may be about to write
4593 * over this location.
4595 if (HDR_L2_READING(ab)) {
4596 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4597 ab->b_flags |= ARC_L2_EVICTED;
4601 * Tell ARC this no longer exists in L2ARC.
4603 if (ab->b_l2hdr != NULL) {
4606 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4607 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4609 list_remove(buflist, ab);
4612 * This may have been leftover after a
4615 ab->b_flags &= ~ARC_L2_WRITING;
4617 mutex_exit(hash_lock);
4619 mutex_exit(&l2arc_buflist_mtx);
4621 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4622 dev->l2ad_evict = taddr;
4626 * Find and write ARC buffers to the L2ARC device.
4628 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4629 * for reading until they have completed writing.
4632 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4634 arc_buf_hdr_t *ab, *ab_prev, *head;
4635 l2arc_buf_hdr_t *hdrl2;
4637 uint64_t passed_sz, write_sz, buf_sz, headroom;
4639 kmutex_t *hash_lock, *list_lock;
4640 boolean_t have_lock, full;
4641 l2arc_write_callback_t *cb;
4643 uint64_t guid = spa_load_guid(spa);
4646 ASSERT(dev->l2ad_vdev != NULL);
4651 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4652 head->b_flags |= ARC_L2_WRITE_HEAD;
4654 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
4656 * Copy buffers for L2ARC writing.
4658 mutex_enter(&l2arc_buflist_mtx);
4659 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) {
4660 list = l2arc_list_locked(try, &list_lock);
4662 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
4665 * L2ARC fast warmup.
4667 * Until the ARC is warm and starts to evict, read from the
4668 * head of the ARC lists rather than the tail.
4670 headroom = target_sz * l2arc_headroom;
4671 if (arc_warm == B_FALSE)
4672 ab = list_head(list);
4674 ab = list_tail(list);
4676 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
4678 for (; ab; ab = ab_prev) {
4679 if (arc_warm == B_FALSE)
4680 ab_prev = list_next(list, ab);
4682 ab_prev = list_prev(list, ab);
4683 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size);
4685 hash_lock = HDR_LOCK(ab);
4686 have_lock = MUTEX_HELD(hash_lock);
4687 if (!have_lock && !mutex_tryenter(hash_lock)) {
4688 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
4690 * Skip this buffer rather than waiting.
4695 passed_sz += ab->b_size;
4696 if (passed_sz > headroom) {
4700 mutex_exit(hash_lock);
4701 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
4705 if (!l2arc_write_eligible(guid, ab)) {
4706 mutex_exit(hash_lock);
4710 if ((write_sz + ab->b_size) > target_sz) {
4712 mutex_exit(hash_lock);
4713 ARCSTAT_BUMP(arcstat_l2_write_full);
4719 * Insert a dummy header on the buflist so
4720 * l2arc_write_done() can find where the
4721 * write buffers begin without searching.
4723 list_insert_head(dev->l2ad_buflist, head);
4726 sizeof (l2arc_write_callback_t), KM_SLEEP);
4727 cb->l2wcb_dev = dev;
4728 cb->l2wcb_head = head;
4729 pio = zio_root(spa, l2arc_write_done, cb,
4731 ARCSTAT_BUMP(arcstat_l2_write_pios);
4735 * Create and add a new L2ARC header.
4737 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4739 hdrl2->b_daddr = dev->l2ad_hand;
4741 ab->b_flags |= ARC_L2_WRITING;
4742 ab->b_l2hdr = hdrl2;
4743 list_insert_head(dev->l2ad_buflist, ab);
4744 buf_data = ab->b_buf->b_data;
4745 buf_sz = ab->b_size;
4748 * Compute and store the buffer cksum before
4749 * writing. On debug the cksum is verified first.
4751 arc_cksum_verify(ab->b_buf);
4752 arc_cksum_compute(ab->b_buf, B_TRUE);
4754 mutex_exit(hash_lock);
4756 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4757 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4758 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4759 ZIO_FLAG_CANFAIL, B_FALSE);
4761 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4763 (void) zio_nowait(wzio);
4766 * Keep the clock hand suitably device-aligned.
4768 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4771 dev->l2ad_hand += buf_sz;
4774 mutex_exit(list_lock);
4779 mutex_exit(&l2arc_buflist_mtx);
4783 kmem_cache_free(hdr_cache, head);
4787 ASSERT3U(write_sz, <=, target_sz);
4788 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4789 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4790 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4791 vdev_space_update(dev->l2ad_vdev, write_sz, 0, 0);
4794 * Bump device hand to the device start if it is approaching the end.
4795 * l2arc_evict() will already have evicted ahead for this case.
4797 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4798 vdev_space_update(dev->l2ad_vdev,
4799 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4800 dev->l2ad_hand = dev->l2ad_start;
4801 dev->l2ad_evict = dev->l2ad_start;
4802 dev->l2ad_first = B_FALSE;
4805 dev->l2ad_writing = B_TRUE;
4806 (void) zio_wait(pio);
4807 dev->l2ad_writing = B_FALSE;
4813 * This thread feeds the L2ARC at regular intervals. This is the beating
4814 * heart of the L2ARC.
4817 l2arc_feed_thread(void *dummy __unused)
4822 uint64_t size, wrote;
4823 clock_t begin, next = ddi_get_lbolt();
4825 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4827 mutex_enter(&l2arc_feed_thr_lock);
4829 while (l2arc_thread_exit == 0) {
4830 CALLB_CPR_SAFE_BEGIN(&cpr);
4831 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4832 next - ddi_get_lbolt());
4833 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4834 next = ddi_get_lbolt() + hz;
4837 * Quick check for L2ARC devices.
4839 mutex_enter(&l2arc_dev_mtx);
4840 if (l2arc_ndev == 0) {
4841 mutex_exit(&l2arc_dev_mtx);
4844 mutex_exit(&l2arc_dev_mtx);
4845 begin = ddi_get_lbolt();
4848 * This selects the next l2arc device to write to, and in
4849 * doing so the next spa to feed from: dev->l2ad_spa. This
4850 * will return NULL if there are now no l2arc devices or if
4851 * they are all faulted.
4853 * If a device is returned, its spa's config lock is also
4854 * held to prevent device removal. l2arc_dev_get_next()
4855 * will grab and release l2arc_dev_mtx.
4857 if ((dev = l2arc_dev_get_next()) == NULL)
4860 spa = dev->l2ad_spa;
4861 ASSERT(spa != NULL);
4864 * If the pool is read-only then force the feed thread to
4865 * sleep a little longer.
4867 if (!spa_writeable(spa)) {
4868 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4869 spa_config_exit(spa, SCL_L2ARC, dev);
4874 * Avoid contributing to memory pressure.
4876 if (arc_reclaim_needed()) {
4877 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4878 spa_config_exit(spa, SCL_L2ARC, dev);
4882 ARCSTAT_BUMP(arcstat_l2_feeds);
4884 size = l2arc_write_size(dev);
4887 * Evict L2ARC buffers that will be overwritten.
4889 l2arc_evict(dev, size, B_FALSE);
4892 * Write ARC buffers.
4894 wrote = l2arc_write_buffers(spa, dev, size);
4897 * Calculate interval between writes.
4899 next = l2arc_write_interval(begin, size, wrote);
4900 spa_config_exit(spa, SCL_L2ARC, dev);
4903 l2arc_thread_exit = 0;
4904 cv_broadcast(&l2arc_feed_thr_cv);
4905 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4910 l2arc_vdev_present(vdev_t *vd)
4914 mutex_enter(&l2arc_dev_mtx);
4915 for (dev = list_head(l2arc_dev_list); dev != NULL;
4916 dev = list_next(l2arc_dev_list, dev)) {
4917 if (dev->l2ad_vdev == vd)
4920 mutex_exit(&l2arc_dev_mtx);
4922 return (dev != NULL);
4926 * Add a vdev for use by the L2ARC. By this point the spa has already
4927 * validated the vdev and opened it.
4930 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
4932 l2arc_dev_t *adddev;
4934 ASSERT(!l2arc_vdev_present(vd));
4937 * Create a new l2arc device entry.
4939 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4940 adddev->l2ad_spa = spa;
4941 adddev->l2ad_vdev = vd;
4942 adddev->l2ad_write = l2arc_write_max;
4943 adddev->l2ad_boost = l2arc_write_boost;
4944 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
4945 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
4946 adddev->l2ad_hand = adddev->l2ad_start;
4947 adddev->l2ad_evict = adddev->l2ad_start;
4948 adddev->l2ad_first = B_TRUE;
4949 adddev->l2ad_writing = B_FALSE;
4950 ASSERT3U(adddev->l2ad_write, >, 0);
4953 * This is a list of all ARC buffers that are still valid on the
4956 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4957 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4958 offsetof(arc_buf_hdr_t, b_l2node));
4960 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
4963 * Add device to global list
4965 mutex_enter(&l2arc_dev_mtx);
4966 list_insert_head(l2arc_dev_list, adddev);
4967 atomic_inc_64(&l2arc_ndev);
4968 mutex_exit(&l2arc_dev_mtx);
4972 * Remove a vdev from the L2ARC.
4975 l2arc_remove_vdev(vdev_t *vd)
4977 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4980 * Find the device by vdev
4982 mutex_enter(&l2arc_dev_mtx);
4983 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4984 nextdev = list_next(l2arc_dev_list, dev);
4985 if (vd == dev->l2ad_vdev) {
4990 ASSERT(remdev != NULL);
4993 * Remove device from global list
4995 list_remove(l2arc_dev_list, remdev);
4996 l2arc_dev_last = NULL; /* may have been invalidated */
4997 atomic_dec_64(&l2arc_ndev);
4998 mutex_exit(&l2arc_dev_mtx);
5001 * Clear all buflists and ARC references. L2ARC device flush.
5003 l2arc_evict(remdev, 0, B_TRUE);
5004 list_destroy(remdev->l2ad_buflist);
5005 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5006 kmem_free(remdev, sizeof (l2arc_dev_t));
5012 l2arc_thread_exit = 0;
5014 l2arc_writes_sent = 0;
5015 l2arc_writes_done = 0;
5017 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5018 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5019 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5020 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5021 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5023 l2arc_dev_list = &L2ARC_dev_list;
5024 l2arc_free_on_write = &L2ARC_free_on_write;
5025 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5026 offsetof(l2arc_dev_t, l2ad_node));
5027 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5028 offsetof(l2arc_data_free_t, l2df_list_node));
5035 * This is called from dmu_fini(), which is called from spa_fini();
5036 * Because of this, we can assume that all l2arc devices have
5037 * already been removed when the pools themselves were removed.
5040 l2arc_do_free_on_write();
5042 mutex_destroy(&l2arc_feed_thr_lock);
5043 cv_destroy(&l2arc_feed_thr_cv);
5044 mutex_destroy(&l2arc_dev_mtx);
5045 mutex_destroy(&l2arc_buflist_mtx);
5046 mutex_destroy(&l2arc_free_on_write_mtx);
5048 list_destroy(l2arc_dev_list);
5049 list_destroy(l2arc_free_on_write);
5055 if (!(spa_mode_global & FWRITE))
5058 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5059 TS_RUN, minclsyspri);
5065 if (!(spa_mode_global & FWRITE))
5068 mutex_enter(&l2arc_feed_thr_lock);
5069 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5070 l2arc_thread_exit = 1;
5071 while (l2arc_thread_exit != 0)
5072 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5073 mutex_exit(&l2arc_feed_thr_lock);