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;
194 int zfs_disable_dup_eviction = 0;
196 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
197 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
198 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
199 SYSCTL_DECL(_vfs_zfs);
200 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
202 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
206 * Note that buffers can be in one of 6 states:
207 * ARC_anon - anonymous (discussed below)
208 * ARC_mru - recently used, currently cached
209 * ARC_mru_ghost - recentely used, no longer in cache
210 * ARC_mfu - frequently used, currently cached
211 * ARC_mfu_ghost - frequently used, no longer in cache
212 * ARC_l2c_only - exists in L2ARC but not other states
213 * When there are no active references to the buffer, they are
214 * are linked onto a list in one of these arc states. These are
215 * the only buffers that can be evicted or deleted. Within each
216 * state there are multiple lists, one for meta-data and one for
217 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
218 * etc.) is tracked separately so that it can be managed more
219 * explicitly: favored over data, limited explicitly.
221 * Anonymous buffers are buffers that are not associated with
222 * a DVA. These are buffers that hold dirty block copies
223 * before they are written to stable storage. By definition,
224 * they are "ref'd" and are considered part of arc_mru
225 * that cannot be freed. Generally, they will aquire a DVA
226 * as they are written and migrate onto the arc_mru list.
228 * The ARC_l2c_only state is for buffers that are in the second
229 * level ARC but no longer in any of the ARC_m* lists. The second
230 * level ARC itself may also contain buffers that are in any of
231 * the ARC_m* states - meaning that a buffer can exist in two
232 * places. The reason for the ARC_l2c_only state is to keep the
233 * buffer header in the hash table, so that reads that hit the
234 * second level ARC benefit from these fast lookups.
237 #define ARCS_LOCK_PAD CACHE_LINE_SIZE
241 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))];
246 * must be power of two for mask use to work
249 #define ARC_BUFC_NUMDATALISTS 16
250 #define ARC_BUFC_NUMMETADATALISTS 16
251 #define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS)
253 typedef struct arc_state {
254 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
255 uint64_t arcs_size; /* total amount of data in this state */
256 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */
257 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE);
260 #define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock))
263 static arc_state_t ARC_anon;
264 static arc_state_t ARC_mru;
265 static arc_state_t ARC_mru_ghost;
266 static arc_state_t ARC_mfu;
267 static arc_state_t ARC_mfu_ghost;
268 static arc_state_t ARC_l2c_only;
270 typedef struct arc_stats {
271 kstat_named_t arcstat_hits;
272 kstat_named_t arcstat_misses;
273 kstat_named_t arcstat_demand_data_hits;
274 kstat_named_t arcstat_demand_data_misses;
275 kstat_named_t arcstat_demand_metadata_hits;
276 kstat_named_t arcstat_demand_metadata_misses;
277 kstat_named_t arcstat_prefetch_data_hits;
278 kstat_named_t arcstat_prefetch_data_misses;
279 kstat_named_t arcstat_prefetch_metadata_hits;
280 kstat_named_t arcstat_prefetch_metadata_misses;
281 kstat_named_t arcstat_mru_hits;
282 kstat_named_t arcstat_mru_ghost_hits;
283 kstat_named_t arcstat_mfu_hits;
284 kstat_named_t arcstat_mfu_ghost_hits;
285 kstat_named_t arcstat_allocated;
286 kstat_named_t arcstat_deleted;
287 kstat_named_t arcstat_stolen;
288 kstat_named_t arcstat_recycle_miss;
289 kstat_named_t arcstat_mutex_miss;
290 kstat_named_t arcstat_evict_skip;
291 kstat_named_t arcstat_evict_l2_cached;
292 kstat_named_t arcstat_evict_l2_eligible;
293 kstat_named_t arcstat_evict_l2_ineligible;
294 kstat_named_t arcstat_hash_elements;
295 kstat_named_t arcstat_hash_elements_max;
296 kstat_named_t arcstat_hash_collisions;
297 kstat_named_t arcstat_hash_chains;
298 kstat_named_t arcstat_hash_chain_max;
299 kstat_named_t arcstat_p;
300 kstat_named_t arcstat_c;
301 kstat_named_t arcstat_c_min;
302 kstat_named_t arcstat_c_max;
303 kstat_named_t arcstat_size;
304 kstat_named_t arcstat_hdr_size;
305 kstat_named_t arcstat_data_size;
306 kstat_named_t arcstat_other_size;
307 kstat_named_t arcstat_l2_hits;
308 kstat_named_t arcstat_l2_misses;
309 kstat_named_t arcstat_l2_feeds;
310 kstat_named_t arcstat_l2_rw_clash;
311 kstat_named_t arcstat_l2_read_bytes;
312 kstat_named_t arcstat_l2_write_bytes;
313 kstat_named_t arcstat_l2_writes_sent;
314 kstat_named_t arcstat_l2_writes_done;
315 kstat_named_t arcstat_l2_writes_error;
316 kstat_named_t arcstat_l2_writes_hdr_miss;
317 kstat_named_t arcstat_l2_evict_lock_retry;
318 kstat_named_t arcstat_l2_evict_reading;
319 kstat_named_t arcstat_l2_free_on_write;
320 kstat_named_t arcstat_l2_abort_lowmem;
321 kstat_named_t arcstat_l2_cksum_bad;
322 kstat_named_t arcstat_l2_io_error;
323 kstat_named_t arcstat_l2_size;
324 kstat_named_t arcstat_l2_hdr_size;
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;
337 kstat_named_t arcstat_memory_throttle_count;
338 kstat_named_t arcstat_duplicate_buffers;
339 kstat_named_t arcstat_duplicate_buffers_size;
340 kstat_named_t arcstat_duplicate_reads;
343 static arc_stats_t arc_stats = {
344 { "hits", KSTAT_DATA_UINT64 },
345 { "misses", KSTAT_DATA_UINT64 },
346 { "demand_data_hits", KSTAT_DATA_UINT64 },
347 { "demand_data_misses", KSTAT_DATA_UINT64 },
348 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
349 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
350 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
351 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
352 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
353 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
354 { "mru_hits", KSTAT_DATA_UINT64 },
355 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
356 { "mfu_hits", KSTAT_DATA_UINT64 },
357 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
358 { "allocated", KSTAT_DATA_UINT64 },
359 { "deleted", KSTAT_DATA_UINT64 },
360 { "stolen", KSTAT_DATA_UINT64 },
361 { "recycle_miss", KSTAT_DATA_UINT64 },
362 { "mutex_miss", KSTAT_DATA_UINT64 },
363 { "evict_skip", KSTAT_DATA_UINT64 },
364 { "evict_l2_cached", KSTAT_DATA_UINT64 },
365 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
366 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
367 { "hash_elements", KSTAT_DATA_UINT64 },
368 { "hash_elements_max", KSTAT_DATA_UINT64 },
369 { "hash_collisions", KSTAT_DATA_UINT64 },
370 { "hash_chains", KSTAT_DATA_UINT64 },
371 { "hash_chain_max", KSTAT_DATA_UINT64 },
372 { "p", KSTAT_DATA_UINT64 },
373 { "c", KSTAT_DATA_UINT64 },
374 { "c_min", KSTAT_DATA_UINT64 },
375 { "c_max", KSTAT_DATA_UINT64 },
376 { "size", KSTAT_DATA_UINT64 },
377 { "hdr_size", KSTAT_DATA_UINT64 },
378 { "data_size", KSTAT_DATA_UINT64 },
379 { "other_size", KSTAT_DATA_UINT64 },
380 { "l2_hits", KSTAT_DATA_UINT64 },
381 { "l2_misses", KSTAT_DATA_UINT64 },
382 { "l2_feeds", KSTAT_DATA_UINT64 },
383 { "l2_rw_clash", KSTAT_DATA_UINT64 },
384 { "l2_read_bytes", KSTAT_DATA_UINT64 },
385 { "l2_write_bytes", KSTAT_DATA_UINT64 },
386 { "l2_writes_sent", KSTAT_DATA_UINT64 },
387 { "l2_writes_done", KSTAT_DATA_UINT64 },
388 { "l2_writes_error", KSTAT_DATA_UINT64 },
389 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
390 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
391 { "l2_evict_reading", KSTAT_DATA_UINT64 },
392 { "l2_free_on_write", KSTAT_DATA_UINT64 },
393 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
394 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
395 { "l2_io_error", KSTAT_DATA_UINT64 },
396 { "l2_size", KSTAT_DATA_UINT64 },
397 { "l2_hdr_size", KSTAT_DATA_UINT64 },
398 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
399 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
400 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
401 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
402 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
403 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
404 { "l2_write_full", KSTAT_DATA_UINT64 },
405 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
406 { "l2_write_pios", KSTAT_DATA_UINT64 },
407 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
408 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
409 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
410 { "memory_throttle_count", KSTAT_DATA_UINT64 },
411 { "duplicate_buffers", KSTAT_DATA_UINT64 },
412 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
413 { "duplicate_reads", KSTAT_DATA_UINT64 }
416 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
418 #define ARCSTAT_INCR(stat, val) \
419 atomic_add_64(&arc_stats.stat.value.ui64, (val));
421 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
422 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
424 #define ARCSTAT_MAX(stat, val) { \
426 while ((val) > (m = arc_stats.stat.value.ui64) && \
427 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
431 #define ARCSTAT_MAXSTAT(stat) \
432 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
435 * We define a macro to allow ARC hits/misses to be easily broken down by
436 * two separate conditions, giving a total of four different subtypes for
437 * each of hits and misses (so eight statistics total).
439 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
442 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
444 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
448 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
450 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
455 static arc_state_t *arc_anon;
456 static arc_state_t *arc_mru;
457 static arc_state_t *arc_mru_ghost;
458 static arc_state_t *arc_mfu;
459 static arc_state_t *arc_mfu_ghost;
460 static arc_state_t *arc_l2c_only;
463 * There are several ARC variables that are critical to export as kstats --
464 * but we don't want to have to grovel around in the kstat whenever we wish to
465 * manipulate them. For these variables, we therefore define them to be in
466 * terms of the statistic variable. This assures that we are not introducing
467 * the possibility of inconsistency by having shadow copies of the variables,
468 * while still allowing the code to be readable.
470 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
471 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
472 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
473 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
474 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
476 static int arc_no_grow; /* Don't try to grow cache size */
477 static uint64_t arc_tempreserve;
478 static uint64_t arc_loaned_bytes;
479 static uint64_t arc_meta_used;
480 static uint64_t arc_meta_limit;
481 static uint64_t arc_meta_max = 0;
482 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0,
483 "ARC metadata used");
484 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0,
485 "ARC metadata limit");
487 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
489 typedef struct arc_callback arc_callback_t;
491 struct arc_callback {
493 arc_done_func_t *acb_done;
495 zio_t *acb_zio_dummy;
496 arc_callback_t *acb_next;
499 typedef struct arc_write_callback arc_write_callback_t;
501 struct arc_write_callback {
503 arc_done_func_t *awcb_ready;
504 arc_done_func_t *awcb_done;
509 /* protected by hash lock */
514 kmutex_t b_freeze_lock;
515 zio_cksum_t *b_freeze_cksum;
518 arc_buf_hdr_t *b_hash_next;
523 arc_callback_t *b_acb;
527 arc_buf_contents_t b_type;
531 /* protected by arc state mutex */
532 arc_state_t *b_state;
533 list_node_t b_arc_node;
535 /* updated atomically */
536 clock_t b_arc_access;
538 /* self protecting */
541 l2arc_buf_hdr_t *b_l2hdr;
542 list_node_t b_l2node;
545 static arc_buf_t *arc_eviction_list;
546 static kmutex_t arc_eviction_mtx;
547 static arc_buf_hdr_t arc_eviction_hdr;
548 static void arc_get_data_buf(arc_buf_t *buf);
549 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
550 static int arc_evict_needed(arc_buf_contents_t type);
551 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
553 static void arc_buf_watch(arc_buf_t *buf);
556 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
558 #define GHOST_STATE(state) \
559 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
560 (state) == arc_l2c_only)
563 * Private ARC flags. These flags are private ARC only flags that will show up
564 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
565 * be passed in as arc_flags in things like arc_read. However, these flags
566 * should never be passed and should only be set by ARC code. When adding new
567 * public flags, make sure not to smash the private ones.
570 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
571 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
572 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
573 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
574 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
575 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
576 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
577 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
578 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
579 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
581 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
582 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
583 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
584 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
585 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
586 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
587 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
588 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
589 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
590 (hdr)->b_l2hdr != NULL)
591 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
592 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
593 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
599 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
600 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
603 * Hash table routines
606 #define HT_LOCK_PAD CACHE_LINE_SIZE
611 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
615 #define BUF_LOCKS 256
616 typedef struct buf_hash_table {
618 arc_buf_hdr_t **ht_table;
619 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
622 static buf_hash_table_t buf_hash_table;
624 #define BUF_HASH_INDEX(spa, dva, birth) \
625 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
626 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
627 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
628 #define HDR_LOCK(hdr) \
629 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
631 uint64_t zfs_crc64_table[256];
637 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
638 #define L2ARC_HEADROOM 2 /* num of writes */
639 #define L2ARC_FEED_SECS 1 /* caching interval secs */
640 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
642 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
643 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
646 * L2ARC Performance Tunables
648 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
649 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
650 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
651 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
652 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
653 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
654 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
655 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
657 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
658 &l2arc_write_max, 0, "max write size");
659 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
660 &l2arc_write_boost, 0, "extra write during warmup");
661 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
662 &l2arc_headroom, 0, "number of dev writes");
663 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
664 &l2arc_feed_secs, 0, "interval seconds");
665 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
666 &l2arc_feed_min_ms, 0, "min interval milliseconds");
668 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
669 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
670 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
671 &l2arc_feed_again, 0, "turbo warmup");
672 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
673 &l2arc_norw, 0, "no reads during writes");
675 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
676 &ARC_anon.arcs_size, 0, "size of anonymous state");
677 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
678 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
679 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
680 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
682 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
683 &ARC_mru.arcs_size, 0, "size of mru state");
684 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
685 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
686 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
687 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
689 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
690 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state");
691 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
692 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
693 "size of metadata in mru ghost state");
694 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
695 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
696 "size of data in mru ghost state");
698 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
699 &ARC_mfu.arcs_size, 0, "size of mfu state");
700 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
701 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
702 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
703 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
705 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
706 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state");
707 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
708 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
709 "size of metadata in mfu ghost state");
710 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
711 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
712 "size of data in mfu ghost state");
714 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
715 &ARC_l2c_only.arcs_size, 0, "size of mru state");
720 typedef struct l2arc_dev {
721 vdev_t *l2ad_vdev; /* vdev */
722 spa_t *l2ad_spa; /* spa */
723 uint64_t l2ad_hand; /* next write location */
724 uint64_t l2ad_write; /* desired write size, bytes */
725 uint64_t l2ad_boost; /* warmup write boost, bytes */
726 uint64_t l2ad_start; /* first addr on device */
727 uint64_t l2ad_end; /* last addr on device */
728 uint64_t l2ad_evict; /* last addr eviction reached */
729 boolean_t l2ad_first; /* first sweep through */
730 boolean_t l2ad_writing; /* currently writing */
731 list_t *l2ad_buflist; /* buffer list */
732 list_node_t l2ad_node; /* device list node */
735 static list_t L2ARC_dev_list; /* device list */
736 static list_t *l2arc_dev_list; /* device list pointer */
737 static kmutex_t l2arc_dev_mtx; /* device list mutex */
738 static l2arc_dev_t *l2arc_dev_last; /* last device used */
739 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
740 static list_t L2ARC_free_on_write; /* free after write buf list */
741 static list_t *l2arc_free_on_write; /* free after write list ptr */
742 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
743 static uint64_t l2arc_ndev; /* number of devices */
745 typedef struct l2arc_read_callback {
746 arc_buf_t *l2rcb_buf; /* read buffer */
747 spa_t *l2rcb_spa; /* spa */
748 blkptr_t l2rcb_bp; /* original blkptr */
749 zbookmark_t l2rcb_zb; /* original bookmark */
750 int l2rcb_flags; /* original flags */
751 } l2arc_read_callback_t;
753 typedef struct l2arc_write_callback {
754 l2arc_dev_t *l2wcb_dev; /* device info */
755 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
756 } l2arc_write_callback_t;
758 struct l2arc_buf_hdr {
759 /* protected by arc_buf_hdr mutex */
760 l2arc_dev_t *b_dev; /* L2ARC device */
761 uint64_t b_daddr; /* disk address, offset byte */
764 typedef struct l2arc_data_free {
765 /* protected by l2arc_free_on_write_mtx */
768 void (*l2df_func)(void *, size_t);
769 list_node_t l2df_list_node;
772 static kmutex_t l2arc_feed_thr_lock;
773 static kcondvar_t l2arc_feed_thr_cv;
774 static uint8_t l2arc_thread_exit;
776 static void l2arc_read_done(zio_t *zio);
777 static void l2arc_hdr_stat_add(void);
778 static void l2arc_hdr_stat_remove(void);
781 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
783 uint8_t *vdva = (uint8_t *)dva;
784 uint64_t crc = -1ULL;
787 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
789 for (i = 0; i < sizeof (dva_t); i++)
790 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
792 crc ^= (spa>>8) ^ birth;
797 #define BUF_EMPTY(buf) \
798 ((buf)->b_dva.dva_word[0] == 0 && \
799 (buf)->b_dva.dva_word[1] == 0 && \
802 #define BUF_EQUAL(spa, dva, birth, buf) \
803 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
804 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
805 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
808 buf_discard_identity(arc_buf_hdr_t *hdr)
810 hdr->b_dva.dva_word[0] = 0;
811 hdr->b_dva.dva_word[1] = 0;
816 static arc_buf_hdr_t *
817 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
819 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
820 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
823 mutex_enter(hash_lock);
824 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
825 buf = buf->b_hash_next) {
826 if (BUF_EQUAL(spa, dva, birth, buf)) {
831 mutex_exit(hash_lock);
837 * Insert an entry into the hash table. If there is already an element
838 * equal to elem in the hash table, then the already existing element
839 * will be returned and the new element will not be inserted.
840 * Otherwise returns NULL.
842 static arc_buf_hdr_t *
843 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
845 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
846 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
850 ASSERT(!HDR_IN_HASH_TABLE(buf));
852 mutex_enter(hash_lock);
853 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
854 fbuf = fbuf->b_hash_next, i++) {
855 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
859 buf->b_hash_next = buf_hash_table.ht_table[idx];
860 buf_hash_table.ht_table[idx] = buf;
861 buf->b_flags |= ARC_IN_HASH_TABLE;
863 /* collect some hash table performance data */
865 ARCSTAT_BUMP(arcstat_hash_collisions);
867 ARCSTAT_BUMP(arcstat_hash_chains);
869 ARCSTAT_MAX(arcstat_hash_chain_max, i);
872 ARCSTAT_BUMP(arcstat_hash_elements);
873 ARCSTAT_MAXSTAT(arcstat_hash_elements);
879 buf_hash_remove(arc_buf_hdr_t *buf)
881 arc_buf_hdr_t *fbuf, **bufp;
882 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
884 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
885 ASSERT(HDR_IN_HASH_TABLE(buf));
887 bufp = &buf_hash_table.ht_table[idx];
888 while ((fbuf = *bufp) != buf) {
889 ASSERT(fbuf != NULL);
890 bufp = &fbuf->b_hash_next;
892 *bufp = buf->b_hash_next;
893 buf->b_hash_next = NULL;
894 buf->b_flags &= ~ARC_IN_HASH_TABLE;
896 /* collect some hash table performance data */
897 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
899 if (buf_hash_table.ht_table[idx] &&
900 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
901 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
905 * Global data structures and functions for the buf kmem cache.
907 static kmem_cache_t *hdr_cache;
908 static kmem_cache_t *buf_cache;
915 kmem_free(buf_hash_table.ht_table,
916 (buf_hash_table.ht_mask + 1) * sizeof (void *));
917 for (i = 0; i < BUF_LOCKS; i++)
918 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
919 kmem_cache_destroy(hdr_cache);
920 kmem_cache_destroy(buf_cache);
924 * Constructor callback - called when the cache is empty
925 * and a new buf is requested.
929 hdr_cons(void *vbuf, void *unused, int kmflag)
931 arc_buf_hdr_t *buf = vbuf;
933 bzero(buf, sizeof (arc_buf_hdr_t));
934 refcount_create(&buf->b_refcnt);
935 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
936 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
937 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
944 buf_cons(void *vbuf, void *unused, int kmflag)
946 arc_buf_t *buf = vbuf;
948 bzero(buf, sizeof (arc_buf_t));
949 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
950 rw_init(&buf->b_data_lock, NULL, RW_DEFAULT, NULL);
951 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
957 * Destructor callback - called when a cached buf is
958 * no longer required.
962 hdr_dest(void *vbuf, void *unused)
964 arc_buf_hdr_t *buf = vbuf;
966 ASSERT(BUF_EMPTY(buf));
967 refcount_destroy(&buf->b_refcnt);
968 cv_destroy(&buf->b_cv);
969 mutex_destroy(&buf->b_freeze_lock);
970 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
975 buf_dest(void *vbuf, void *unused)
977 arc_buf_t *buf = vbuf;
979 mutex_destroy(&buf->b_evict_lock);
980 rw_destroy(&buf->b_data_lock);
981 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
985 * Reclaim callback -- invoked when memory is low.
989 hdr_recl(void *unused)
991 dprintf("hdr_recl called\n");
993 * umem calls the reclaim func when we destroy the buf cache,
994 * which is after we do arc_fini().
997 cv_signal(&arc_reclaim_thr_cv);
1004 uint64_t hsize = 1ULL << 12;
1008 * The hash table is big enough to fill all of physical memory
1009 * with an average 64K block size. The table will take up
1010 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
1012 while (hsize * 65536 < (uint64_t)physmem * PAGESIZE)
1015 buf_hash_table.ht_mask = hsize - 1;
1016 buf_hash_table.ht_table =
1017 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1018 if (buf_hash_table.ht_table == NULL) {
1019 ASSERT(hsize > (1ULL << 8));
1024 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
1025 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
1026 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1027 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1029 for (i = 0; i < 256; i++)
1030 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1031 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1033 for (i = 0; i < BUF_LOCKS; i++) {
1034 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1035 NULL, MUTEX_DEFAULT, NULL);
1039 #define ARC_MINTIME (hz>>4) /* 62 ms */
1042 arc_cksum_verify(arc_buf_t *buf)
1046 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1049 mutex_enter(&buf->b_hdr->b_freeze_lock);
1050 if (buf->b_hdr->b_freeze_cksum == NULL ||
1051 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
1052 mutex_exit(&buf->b_hdr->b_freeze_lock);
1055 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1056 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1057 panic("buffer modified while frozen!");
1058 mutex_exit(&buf->b_hdr->b_freeze_lock);
1062 arc_cksum_equal(arc_buf_t *buf)
1067 mutex_enter(&buf->b_hdr->b_freeze_lock);
1068 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1069 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1070 mutex_exit(&buf->b_hdr->b_freeze_lock);
1076 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1078 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1081 mutex_enter(&buf->b_hdr->b_freeze_lock);
1082 if (buf->b_hdr->b_freeze_cksum != NULL) {
1083 mutex_exit(&buf->b_hdr->b_freeze_lock);
1086 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1087 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1088 buf->b_hdr->b_freeze_cksum);
1089 mutex_exit(&buf->b_hdr->b_freeze_lock);
1092 #endif /* illumos */
1097 typedef struct procctl {
1105 arc_buf_unwatch(arc_buf_t *buf)
1112 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1113 ctl.prwatch.pr_size = 0;
1114 ctl.prwatch.pr_wflags = 0;
1115 result = write(arc_procfd, &ctl, sizeof (ctl));
1116 ASSERT3U(result, ==, sizeof (ctl));
1123 arc_buf_watch(arc_buf_t *buf)
1130 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1131 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1132 ctl.prwatch.pr_wflags = WA_WRITE;
1133 result = write(arc_procfd, &ctl, sizeof (ctl));
1134 ASSERT3U(result, ==, sizeof (ctl));
1138 #endif /* illumos */
1141 arc_buf_thaw(arc_buf_t *buf)
1143 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1144 if (buf->b_hdr->b_state != arc_anon)
1145 panic("modifying non-anon buffer!");
1146 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1147 panic("modifying buffer while i/o in progress!");
1148 arc_cksum_verify(buf);
1151 mutex_enter(&buf->b_hdr->b_freeze_lock);
1152 if (buf->b_hdr->b_freeze_cksum != NULL) {
1153 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1154 buf->b_hdr->b_freeze_cksum = NULL;
1157 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1158 if (buf->b_hdr->b_thawed)
1159 kmem_free(buf->b_hdr->b_thawed, 1);
1160 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1163 mutex_exit(&buf->b_hdr->b_freeze_lock);
1166 arc_buf_unwatch(buf);
1167 #endif /* illumos */
1171 arc_buf_freeze(arc_buf_t *buf)
1173 kmutex_t *hash_lock;
1175 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1178 hash_lock = HDR_LOCK(buf->b_hdr);
1179 mutex_enter(hash_lock);
1181 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1182 buf->b_hdr->b_state == arc_anon);
1183 arc_cksum_compute(buf, B_FALSE);
1184 mutex_exit(hash_lock);
1189 get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock)
1191 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth);
1193 if (ab->b_type == ARC_BUFC_METADATA)
1194 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1);
1196 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1);
1197 buf_hashid += ARC_BUFC_NUMMETADATALISTS;
1200 *list = &state->arcs_lists[buf_hashid];
1201 *lock = ARCS_LOCK(state, buf_hashid);
1206 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1208 ASSERT(MUTEX_HELD(hash_lock));
1210 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1211 (ab->b_state != arc_anon)) {
1212 uint64_t delta = ab->b_size * ab->b_datacnt;
1213 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1217 get_buf_info(ab, ab->b_state, &list, &lock);
1218 ASSERT(!MUTEX_HELD(lock));
1220 ASSERT(list_link_active(&ab->b_arc_node));
1221 list_remove(list, ab);
1222 if (GHOST_STATE(ab->b_state)) {
1223 ASSERT0(ab->b_datacnt);
1224 ASSERT3P(ab->b_buf, ==, NULL);
1228 ASSERT3U(*size, >=, delta);
1229 atomic_add_64(size, -delta);
1231 /* remove the prefetch flag if we get a reference */
1232 if (ab->b_flags & ARC_PREFETCH)
1233 ab->b_flags &= ~ARC_PREFETCH;
1238 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1241 arc_state_t *state = ab->b_state;
1243 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1244 ASSERT(!GHOST_STATE(state));
1246 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1247 (state != arc_anon)) {
1248 uint64_t *size = &state->arcs_lsize[ab->b_type];
1252 get_buf_info(ab, state, &list, &lock);
1253 ASSERT(!MUTEX_HELD(lock));
1255 ASSERT(!list_link_active(&ab->b_arc_node));
1256 list_insert_head(list, ab);
1257 ASSERT(ab->b_datacnt > 0);
1258 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1265 * Move the supplied buffer to the indicated state. The mutex
1266 * for the buffer must be held by the caller.
1269 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1271 arc_state_t *old_state = ab->b_state;
1272 int64_t refcnt = refcount_count(&ab->b_refcnt);
1273 uint64_t from_delta, to_delta;
1277 ASSERT(MUTEX_HELD(hash_lock));
1278 ASSERT(new_state != old_state);
1279 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1280 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1281 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1283 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1286 * If this buffer is evictable, transfer it from the
1287 * old state list to the new state list.
1290 if (old_state != arc_anon) {
1292 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1294 get_buf_info(ab, old_state, &list, &lock);
1295 use_mutex = !MUTEX_HELD(lock);
1299 ASSERT(list_link_active(&ab->b_arc_node));
1300 list_remove(list, ab);
1303 * If prefetching out of the ghost cache,
1304 * we will have a non-zero datacnt.
1306 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1307 /* ghost elements have a ghost size */
1308 ASSERT(ab->b_buf == NULL);
1309 from_delta = ab->b_size;
1311 ASSERT3U(*size, >=, from_delta);
1312 atomic_add_64(size, -from_delta);
1317 if (new_state != arc_anon) {
1319 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1321 get_buf_info(ab, new_state, &list, &lock);
1322 use_mutex = !MUTEX_HELD(lock);
1326 list_insert_head(list, ab);
1328 /* ghost elements have a ghost size */
1329 if (GHOST_STATE(new_state)) {
1330 ASSERT(ab->b_datacnt == 0);
1331 ASSERT(ab->b_buf == NULL);
1332 to_delta = ab->b_size;
1334 atomic_add_64(size, to_delta);
1341 ASSERT(!BUF_EMPTY(ab));
1342 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1343 buf_hash_remove(ab);
1345 /* adjust state sizes */
1347 atomic_add_64(&new_state->arcs_size, to_delta);
1349 ASSERT3U(old_state->arcs_size, >=, from_delta);
1350 atomic_add_64(&old_state->arcs_size, -from_delta);
1352 ab->b_state = new_state;
1354 /* adjust l2arc hdr stats */
1355 if (new_state == arc_l2c_only)
1356 l2arc_hdr_stat_add();
1357 else if (old_state == arc_l2c_only)
1358 l2arc_hdr_stat_remove();
1362 arc_space_consume(uint64_t space, arc_space_type_t type)
1364 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1367 case ARC_SPACE_DATA:
1368 ARCSTAT_INCR(arcstat_data_size, space);
1370 case ARC_SPACE_OTHER:
1371 ARCSTAT_INCR(arcstat_other_size, space);
1373 case ARC_SPACE_HDRS:
1374 ARCSTAT_INCR(arcstat_hdr_size, space);
1376 case ARC_SPACE_L2HDRS:
1377 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1381 atomic_add_64(&arc_meta_used, space);
1382 atomic_add_64(&arc_size, space);
1386 arc_space_return(uint64_t space, arc_space_type_t type)
1388 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1391 case ARC_SPACE_DATA:
1392 ARCSTAT_INCR(arcstat_data_size, -space);
1394 case ARC_SPACE_OTHER:
1395 ARCSTAT_INCR(arcstat_other_size, -space);
1397 case ARC_SPACE_HDRS:
1398 ARCSTAT_INCR(arcstat_hdr_size, -space);
1400 case ARC_SPACE_L2HDRS:
1401 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1405 ASSERT(arc_meta_used >= space);
1406 if (arc_meta_max < arc_meta_used)
1407 arc_meta_max = arc_meta_used;
1408 atomic_add_64(&arc_meta_used, -space);
1409 ASSERT(arc_size >= space);
1410 atomic_add_64(&arc_size, -space);
1414 arc_data_buf_alloc(uint64_t size)
1416 if (arc_evict_needed(ARC_BUFC_DATA))
1417 cv_signal(&arc_reclaim_thr_cv);
1418 atomic_add_64(&arc_size, size);
1419 return (zio_data_buf_alloc(size));
1423 arc_data_buf_free(void *buf, uint64_t size)
1425 zio_data_buf_free(buf, size);
1426 ASSERT(arc_size >= size);
1427 atomic_add_64(&arc_size, -size);
1431 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1436 ASSERT3U(size, >, 0);
1437 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1438 ASSERT(BUF_EMPTY(hdr));
1441 hdr->b_spa = spa_load_guid(spa);
1442 hdr->b_state = arc_anon;
1443 hdr->b_arc_access = 0;
1444 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1447 buf->b_efunc = NULL;
1448 buf->b_private = NULL;
1451 arc_get_data_buf(buf);
1454 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1455 (void) refcount_add(&hdr->b_refcnt, tag);
1460 static char *arc_onloan_tag = "onloan";
1463 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1464 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1465 * buffers must be returned to the arc before they can be used by the DMU or
1469 arc_loan_buf(spa_t *spa, int size)
1473 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1475 atomic_add_64(&arc_loaned_bytes, size);
1480 * Return a loaned arc buffer to the arc.
1483 arc_return_buf(arc_buf_t *buf, void *tag)
1485 arc_buf_hdr_t *hdr = buf->b_hdr;
1487 ASSERT(buf->b_data != NULL);
1488 (void) refcount_add(&hdr->b_refcnt, tag);
1489 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1491 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1494 /* Detach an arc_buf from a dbuf (tag) */
1496 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1500 ASSERT(buf->b_data != NULL);
1502 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1503 (void) refcount_remove(&hdr->b_refcnt, tag);
1504 buf->b_efunc = NULL;
1505 buf->b_private = NULL;
1507 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1511 arc_buf_clone(arc_buf_t *from)
1514 arc_buf_hdr_t *hdr = from->b_hdr;
1515 uint64_t size = hdr->b_size;
1517 ASSERT(hdr->b_state != arc_anon);
1519 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1522 buf->b_efunc = NULL;
1523 buf->b_private = NULL;
1524 buf->b_next = hdr->b_buf;
1526 arc_get_data_buf(buf);
1527 bcopy(from->b_data, buf->b_data, size);
1530 * This buffer already exists in the arc so create a duplicate
1531 * copy for the caller. If the buffer is associated with user data
1532 * then track the size and number of duplicates. These stats will be
1533 * updated as duplicate buffers are created and destroyed.
1535 if (hdr->b_type == ARC_BUFC_DATA) {
1536 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1537 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1539 hdr->b_datacnt += 1;
1544 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1547 kmutex_t *hash_lock;
1550 * Check to see if this buffer is evicted. Callers
1551 * must verify b_data != NULL to know if the add_ref
1554 mutex_enter(&buf->b_evict_lock);
1555 if (buf->b_data == NULL) {
1556 mutex_exit(&buf->b_evict_lock);
1559 hash_lock = HDR_LOCK(buf->b_hdr);
1560 mutex_enter(hash_lock);
1562 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1563 mutex_exit(&buf->b_evict_lock);
1565 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1566 add_reference(hdr, hash_lock, tag);
1567 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1568 arc_access(hdr, hash_lock);
1569 mutex_exit(hash_lock);
1570 ARCSTAT_BUMP(arcstat_hits);
1571 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1572 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1573 data, metadata, hits);
1577 * Free the arc data buffer. If it is an l2arc write in progress,
1578 * the buffer is placed on l2arc_free_on_write to be freed later.
1581 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1583 arc_buf_hdr_t *hdr = buf->b_hdr;
1585 if (HDR_L2_WRITING(hdr)) {
1586 l2arc_data_free_t *df;
1587 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1588 df->l2df_data = buf->b_data;
1589 df->l2df_size = hdr->b_size;
1590 df->l2df_func = free_func;
1591 mutex_enter(&l2arc_free_on_write_mtx);
1592 list_insert_head(l2arc_free_on_write, df);
1593 mutex_exit(&l2arc_free_on_write_mtx);
1594 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1596 free_func(buf->b_data, hdr->b_size);
1601 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1605 /* free up data associated with the buf */
1607 arc_state_t *state = buf->b_hdr->b_state;
1608 uint64_t size = buf->b_hdr->b_size;
1609 arc_buf_contents_t type = buf->b_hdr->b_type;
1611 arc_cksum_verify(buf);
1613 arc_buf_unwatch(buf);
1614 #endif /* illumos */
1617 if (type == ARC_BUFC_METADATA) {
1618 arc_buf_data_free(buf, zio_buf_free);
1619 arc_space_return(size, ARC_SPACE_DATA);
1621 ASSERT(type == ARC_BUFC_DATA);
1622 arc_buf_data_free(buf, zio_data_buf_free);
1623 ARCSTAT_INCR(arcstat_data_size, -size);
1624 atomic_add_64(&arc_size, -size);
1627 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1628 uint64_t *cnt = &state->arcs_lsize[type];
1630 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1631 ASSERT(state != arc_anon);
1633 ASSERT3U(*cnt, >=, size);
1634 atomic_add_64(cnt, -size);
1636 ASSERT3U(state->arcs_size, >=, size);
1637 atomic_add_64(&state->arcs_size, -size);
1641 * If we're destroying a duplicate buffer make sure
1642 * that the appropriate statistics are updated.
1644 if (buf->b_hdr->b_datacnt > 1 &&
1645 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1646 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1647 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1649 ASSERT(buf->b_hdr->b_datacnt > 0);
1650 buf->b_hdr->b_datacnt -= 1;
1653 /* only remove the buf if requested */
1657 /* remove the buf from the hdr list */
1658 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1660 *bufp = buf->b_next;
1663 ASSERT(buf->b_efunc == NULL);
1665 /* clean up the buf */
1667 kmem_cache_free(buf_cache, buf);
1671 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1673 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1674 ASSERT3P(hdr->b_state, ==, arc_anon);
1675 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1676 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1678 if (l2hdr != NULL) {
1679 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1681 * To prevent arc_free() and l2arc_evict() from
1682 * attempting to free the same buffer at the same time,
1683 * a FREE_IN_PROGRESS flag is given to arc_free() to
1684 * give it priority. l2arc_evict() can't destroy this
1685 * header while we are waiting on l2arc_buflist_mtx.
1687 * The hdr may be removed from l2ad_buflist before we
1688 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1690 if (!buflist_held) {
1691 mutex_enter(&l2arc_buflist_mtx);
1692 l2hdr = hdr->b_l2hdr;
1695 if (l2hdr != NULL) {
1696 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1697 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1698 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1699 if (hdr->b_state == arc_l2c_only)
1700 l2arc_hdr_stat_remove();
1701 hdr->b_l2hdr = NULL;
1705 mutex_exit(&l2arc_buflist_mtx);
1708 if (!BUF_EMPTY(hdr)) {
1709 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1710 buf_discard_identity(hdr);
1712 while (hdr->b_buf) {
1713 arc_buf_t *buf = hdr->b_buf;
1716 mutex_enter(&arc_eviction_mtx);
1717 mutex_enter(&buf->b_evict_lock);
1718 ASSERT(buf->b_hdr != NULL);
1719 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1720 hdr->b_buf = buf->b_next;
1721 buf->b_hdr = &arc_eviction_hdr;
1722 buf->b_next = arc_eviction_list;
1723 arc_eviction_list = buf;
1724 mutex_exit(&buf->b_evict_lock);
1725 mutex_exit(&arc_eviction_mtx);
1727 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1730 if (hdr->b_freeze_cksum != NULL) {
1731 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1732 hdr->b_freeze_cksum = NULL;
1734 if (hdr->b_thawed) {
1735 kmem_free(hdr->b_thawed, 1);
1736 hdr->b_thawed = NULL;
1739 ASSERT(!list_link_active(&hdr->b_arc_node));
1740 ASSERT3P(hdr->b_hash_next, ==, NULL);
1741 ASSERT3P(hdr->b_acb, ==, NULL);
1742 kmem_cache_free(hdr_cache, hdr);
1746 arc_buf_free(arc_buf_t *buf, void *tag)
1748 arc_buf_hdr_t *hdr = buf->b_hdr;
1749 int hashed = hdr->b_state != arc_anon;
1751 ASSERT(buf->b_efunc == NULL);
1752 ASSERT(buf->b_data != NULL);
1755 kmutex_t *hash_lock = HDR_LOCK(hdr);
1757 mutex_enter(hash_lock);
1759 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1761 (void) remove_reference(hdr, hash_lock, tag);
1762 if (hdr->b_datacnt > 1) {
1763 arc_buf_destroy(buf, FALSE, TRUE);
1765 ASSERT(buf == hdr->b_buf);
1766 ASSERT(buf->b_efunc == NULL);
1767 hdr->b_flags |= ARC_BUF_AVAILABLE;
1769 mutex_exit(hash_lock);
1770 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1773 * We are in the middle of an async write. Don't destroy
1774 * this buffer unless the write completes before we finish
1775 * decrementing the reference count.
1777 mutex_enter(&arc_eviction_mtx);
1778 (void) remove_reference(hdr, NULL, tag);
1779 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1780 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1781 mutex_exit(&arc_eviction_mtx);
1783 arc_hdr_destroy(hdr);
1785 if (remove_reference(hdr, NULL, tag) > 0)
1786 arc_buf_destroy(buf, FALSE, TRUE);
1788 arc_hdr_destroy(hdr);
1793 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1795 arc_buf_hdr_t *hdr = buf->b_hdr;
1796 kmutex_t *hash_lock = HDR_LOCK(hdr);
1797 int no_callback = (buf->b_efunc == NULL);
1799 if (hdr->b_state == arc_anon) {
1800 ASSERT(hdr->b_datacnt == 1);
1801 arc_buf_free(buf, tag);
1802 return (no_callback);
1805 mutex_enter(hash_lock);
1807 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1808 ASSERT(hdr->b_state != arc_anon);
1809 ASSERT(buf->b_data != NULL);
1811 (void) remove_reference(hdr, hash_lock, tag);
1812 if (hdr->b_datacnt > 1) {
1814 arc_buf_destroy(buf, FALSE, TRUE);
1815 } else if (no_callback) {
1816 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1817 ASSERT(buf->b_efunc == NULL);
1818 hdr->b_flags |= ARC_BUF_AVAILABLE;
1820 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1821 refcount_is_zero(&hdr->b_refcnt));
1822 mutex_exit(hash_lock);
1823 return (no_callback);
1827 arc_buf_size(arc_buf_t *buf)
1829 return (buf->b_hdr->b_size);
1833 * Called from the DMU to determine if the current buffer should be
1834 * evicted. In order to ensure proper locking, the eviction must be initiated
1835 * from the DMU. Return true if the buffer is associated with user data and
1836 * duplicate buffers still exist.
1839 arc_buf_eviction_needed(arc_buf_t *buf)
1842 boolean_t evict_needed = B_FALSE;
1844 if (zfs_disable_dup_eviction)
1847 mutex_enter(&buf->b_evict_lock);
1851 * We are in arc_do_user_evicts(); let that function
1852 * perform the eviction.
1854 ASSERT(buf->b_data == NULL);
1855 mutex_exit(&buf->b_evict_lock);
1857 } else if (buf->b_data == NULL) {
1859 * We have already been added to the arc eviction list;
1860 * recommend eviction.
1862 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1863 mutex_exit(&buf->b_evict_lock);
1867 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1868 evict_needed = B_TRUE;
1870 mutex_exit(&buf->b_evict_lock);
1871 return (evict_needed);
1875 * Evict buffers from list until we've removed the specified number of
1876 * bytes. Move the removed buffers to the appropriate evict state.
1877 * If the recycle flag is set, then attempt to "recycle" a buffer:
1878 * - look for a buffer to evict that is `bytes' long.
1879 * - return the data block from this buffer rather than freeing it.
1880 * This flag is used by callers that are trying to make space for a
1881 * new buffer in a full arc cache.
1883 * This function makes a "best effort". It skips over any buffers
1884 * it can't get a hash_lock on, and so may not catch all candidates.
1885 * It may also return without evicting as much space as requested.
1888 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1889 arc_buf_contents_t type)
1891 arc_state_t *evicted_state;
1892 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1893 int64_t bytes_remaining;
1894 arc_buf_hdr_t *ab, *ab_prev = NULL;
1895 list_t *evicted_list, *list, *evicted_list_start, *list_start;
1896 kmutex_t *lock, *evicted_lock;
1897 kmutex_t *hash_lock;
1898 boolean_t have_lock;
1899 void *stolen = NULL;
1900 static int evict_metadata_offset, evict_data_offset;
1901 int i, idx, offset, list_count, count;
1903 ASSERT(state == arc_mru || state == arc_mfu);
1905 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1907 if (type == ARC_BUFC_METADATA) {
1909 list_count = ARC_BUFC_NUMMETADATALISTS;
1910 list_start = &state->arcs_lists[0];
1911 evicted_list_start = &evicted_state->arcs_lists[0];
1912 idx = evict_metadata_offset;
1914 offset = ARC_BUFC_NUMMETADATALISTS;
1915 list_start = &state->arcs_lists[offset];
1916 evicted_list_start = &evicted_state->arcs_lists[offset];
1917 list_count = ARC_BUFC_NUMDATALISTS;
1918 idx = evict_data_offset;
1920 bytes_remaining = evicted_state->arcs_lsize[type];
1924 list = &list_start[idx];
1925 evicted_list = &evicted_list_start[idx];
1926 lock = ARCS_LOCK(state, (offset + idx));
1927 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx));
1930 mutex_enter(evicted_lock);
1932 for (ab = list_tail(list); ab; ab = ab_prev) {
1933 ab_prev = list_prev(list, ab);
1934 bytes_remaining -= (ab->b_size * ab->b_datacnt);
1935 /* prefetch buffers have a minimum lifespan */
1936 if (HDR_IO_IN_PROGRESS(ab) ||
1937 (spa && ab->b_spa != spa) ||
1938 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1939 ddi_get_lbolt() - ab->b_arc_access <
1940 arc_min_prefetch_lifespan)) {
1944 /* "lookahead" for better eviction candidate */
1945 if (recycle && ab->b_size != bytes &&
1946 ab_prev && ab_prev->b_size == bytes)
1948 hash_lock = HDR_LOCK(ab);
1949 have_lock = MUTEX_HELD(hash_lock);
1950 if (have_lock || mutex_tryenter(hash_lock)) {
1951 ASSERT0(refcount_count(&ab->b_refcnt));
1952 ASSERT(ab->b_datacnt > 0);
1954 arc_buf_t *buf = ab->b_buf;
1955 if (!mutex_tryenter(&buf->b_evict_lock)) {
1960 bytes_evicted += ab->b_size;
1961 if (recycle && ab->b_type == type &&
1962 ab->b_size == bytes &&
1963 !HDR_L2_WRITING(ab)) {
1964 stolen = buf->b_data;
1969 mutex_enter(&arc_eviction_mtx);
1970 arc_buf_destroy(buf,
1971 buf->b_data == stolen, FALSE);
1972 ab->b_buf = buf->b_next;
1973 buf->b_hdr = &arc_eviction_hdr;
1974 buf->b_next = arc_eviction_list;
1975 arc_eviction_list = buf;
1976 mutex_exit(&arc_eviction_mtx);
1977 mutex_exit(&buf->b_evict_lock);
1979 mutex_exit(&buf->b_evict_lock);
1980 arc_buf_destroy(buf,
1981 buf->b_data == stolen, TRUE);
1986 ARCSTAT_INCR(arcstat_evict_l2_cached,
1989 if (l2arc_write_eligible(ab->b_spa, ab)) {
1990 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1994 arcstat_evict_l2_ineligible,
1999 if (ab->b_datacnt == 0) {
2000 arc_change_state(evicted_state, ab, hash_lock);
2001 ASSERT(HDR_IN_HASH_TABLE(ab));
2002 ab->b_flags |= ARC_IN_HASH_TABLE;
2003 ab->b_flags &= ~ARC_BUF_AVAILABLE;
2004 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
2007 mutex_exit(hash_lock);
2008 if (bytes >= 0 && bytes_evicted >= bytes)
2010 if (bytes_remaining > 0) {
2011 mutex_exit(evicted_lock);
2013 idx = ((idx + 1) & (list_count - 1));
2022 mutex_exit(evicted_lock);
2025 idx = ((idx + 1) & (list_count - 1));
2028 if (bytes_evicted < bytes) {
2029 if (count < list_count)
2032 dprintf("only evicted %lld bytes from %x",
2033 (longlong_t)bytes_evicted, state);
2035 if (type == ARC_BUFC_METADATA)
2036 evict_metadata_offset = idx;
2038 evict_data_offset = idx;
2041 ARCSTAT_INCR(arcstat_evict_skip, skipped);
2044 ARCSTAT_INCR(arcstat_mutex_miss, missed);
2047 * We have just evicted some date into the ghost state, make
2048 * sure we also adjust the ghost state size if necessary.
2051 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
2052 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
2053 arc_mru_ghost->arcs_size - arc_c;
2055 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
2057 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
2058 arc_evict_ghost(arc_mru_ghost, 0, todelete);
2059 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
2060 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
2061 arc_mru_ghost->arcs_size +
2062 arc_mfu_ghost->arcs_size - arc_c);
2063 arc_evict_ghost(arc_mfu_ghost, 0, todelete);
2067 ARCSTAT_BUMP(arcstat_stolen);
2073 * Remove buffers from list until we've removed the specified number of
2074 * bytes. Destroy the buffers that are removed.
2077 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
2079 arc_buf_hdr_t *ab, *ab_prev;
2080 arc_buf_hdr_t marker = { 0 };
2081 list_t *list, *list_start;
2082 kmutex_t *hash_lock, *lock;
2083 uint64_t bytes_deleted = 0;
2084 uint64_t bufs_skipped = 0;
2085 static int evict_offset;
2086 int list_count, idx = evict_offset;
2087 int offset, count = 0;
2089 ASSERT(GHOST_STATE(state));
2092 * data lists come after metadata lists
2094 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS];
2095 list_count = ARC_BUFC_NUMDATALISTS;
2096 offset = ARC_BUFC_NUMMETADATALISTS;
2099 list = &list_start[idx];
2100 lock = ARCS_LOCK(state, idx + offset);
2103 for (ab = list_tail(list); ab; ab = ab_prev) {
2104 ab_prev = list_prev(list, ab);
2105 if (spa && ab->b_spa != spa)
2108 /* ignore markers */
2112 hash_lock = HDR_LOCK(ab);
2113 /* caller may be trying to modify this buffer, skip it */
2114 if (MUTEX_HELD(hash_lock))
2116 if (mutex_tryenter(hash_lock)) {
2117 ASSERT(!HDR_IO_IN_PROGRESS(ab));
2118 ASSERT(ab->b_buf == NULL);
2119 ARCSTAT_BUMP(arcstat_deleted);
2120 bytes_deleted += ab->b_size;
2122 if (ab->b_l2hdr != NULL) {
2124 * This buffer is cached on the 2nd Level ARC;
2125 * don't destroy the header.
2127 arc_change_state(arc_l2c_only, ab, hash_lock);
2128 mutex_exit(hash_lock);
2130 arc_change_state(arc_anon, ab, hash_lock);
2131 mutex_exit(hash_lock);
2132 arc_hdr_destroy(ab);
2135 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
2136 if (bytes >= 0 && bytes_deleted >= bytes)
2138 } else if (bytes < 0) {
2140 * Insert a list marker and then wait for the
2141 * hash lock to become available. Once its
2142 * available, restart from where we left off.
2144 list_insert_after(list, ab, &marker);
2146 mutex_enter(hash_lock);
2147 mutex_exit(hash_lock);
2149 ab_prev = list_prev(list, &marker);
2150 list_remove(list, &marker);
2155 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1));
2158 if (count < list_count)
2162 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] &&
2163 (bytes < 0 || bytes_deleted < bytes)) {
2164 list_start = &state->arcs_lists[0];
2165 list_count = ARC_BUFC_NUMMETADATALISTS;
2171 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2175 if (bytes_deleted < bytes)
2176 dprintf("only deleted %lld bytes from %p",
2177 (longlong_t)bytes_deleted, state);
2183 int64_t adjustment, delta;
2189 adjustment = MIN((int64_t)(arc_size - arc_c),
2190 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2193 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2194 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2195 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
2196 adjustment -= delta;
2199 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2200 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2201 (void) arc_evict(arc_mru, 0, delta, FALSE,
2209 adjustment = arc_size - arc_c;
2211 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2212 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2213 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2214 adjustment -= delta;
2217 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2218 int64_t delta = MIN(adjustment,
2219 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2220 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2225 * Adjust ghost lists
2228 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2230 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2231 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2232 arc_evict_ghost(arc_mru_ghost, 0, delta);
2236 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2238 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2239 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2240 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2245 arc_do_user_evicts(void)
2247 static arc_buf_t *tmp_arc_eviction_list;
2250 * Move list over to avoid LOR
2253 mutex_enter(&arc_eviction_mtx);
2254 tmp_arc_eviction_list = arc_eviction_list;
2255 arc_eviction_list = NULL;
2256 mutex_exit(&arc_eviction_mtx);
2258 while (tmp_arc_eviction_list != NULL) {
2259 arc_buf_t *buf = tmp_arc_eviction_list;
2260 tmp_arc_eviction_list = buf->b_next;
2261 mutex_enter(&buf->b_evict_lock);
2263 mutex_exit(&buf->b_evict_lock);
2265 if (buf->b_efunc != NULL)
2266 VERIFY(buf->b_efunc(buf) == 0);
2268 buf->b_efunc = NULL;
2269 buf->b_private = NULL;
2270 kmem_cache_free(buf_cache, buf);
2273 if (arc_eviction_list != NULL)
2278 * Flush all *evictable* data from the cache for the given spa.
2279 * NOTE: this will not touch "active" (i.e. referenced) data.
2282 arc_flush(spa_t *spa)
2287 guid = spa_load_guid(spa);
2289 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) {
2290 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2294 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) {
2295 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2299 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) {
2300 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2304 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) {
2305 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2310 arc_evict_ghost(arc_mru_ghost, guid, -1);
2311 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2313 mutex_enter(&arc_reclaim_thr_lock);
2314 arc_do_user_evicts();
2315 mutex_exit(&arc_reclaim_thr_lock);
2316 ASSERT(spa || arc_eviction_list == NULL);
2322 if (arc_c > arc_c_min) {
2326 to_free = arc_c >> arc_shrink_shift;
2328 to_free = arc_c >> arc_shrink_shift;
2330 if (arc_c > arc_c_min + to_free)
2331 atomic_add_64(&arc_c, -to_free);
2335 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2336 if (arc_c > arc_size)
2337 arc_c = MAX(arc_size, arc_c_min);
2339 arc_p = (arc_c >> 1);
2340 ASSERT(arc_c >= arc_c_min);
2341 ASSERT((int64_t)arc_p >= 0);
2344 if (arc_size > arc_c)
2348 static int needfree = 0;
2351 arc_reclaim_needed(void)
2360 * Cooperate with pagedaemon when it's time for it to scan
2361 * and reclaim some pages.
2363 if (vm_paging_needed())
2368 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2373 * check that we're out of range of the pageout scanner. It starts to
2374 * schedule paging if freemem is less than lotsfree and needfree.
2375 * lotsfree is the high-water mark for pageout, and needfree is the
2376 * number of needed free pages. We add extra pages here to make sure
2377 * the scanner doesn't start up while we're freeing memory.
2379 if (freemem < lotsfree + needfree + extra)
2383 * check to make sure that swapfs has enough space so that anon
2384 * reservations can still succeed. anon_resvmem() checks that the
2385 * availrmem is greater than swapfs_minfree, and the number of reserved
2386 * swap pages. We also add a bit of extra here just to prevent
2387 * circumstances from getting really dire.
2389 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2394 * If we're on an i386 platform, it's possible that we'll exhaust the
2395 * kernel heap space before we ever run out of available physical
2396 * memory. Most checks of the size of the heap_area compare against
2397 * tune.t_minarmem, which is the minimum available real memory that we
2398 * can have in the system. However, this is generally fixed at 25 pages
2399 * which is so low that it's useless. In this comparison, we seek to
2400 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2401 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2404 if (btop(vmem_size(heap_arena, VMEM_FREE)) <
2405 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
2409 if (kmem_used() > (kmem_size() * 3) / 4)
2414 if (spa_get_random(100) == 0)
2420 extern kmem_cache_t *zio_buf_cache[];
2421 extern kmem_cache_t *zio_data_buf_cache[];
2424 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2427 kmem_cache_t *prev_cache = NULL;
2428 kmem_cache_t *prev_data_cache = NULL;
2431 if (arc_meta_used >= arc_meta_limit) {
2433 * We are exceeding our meta-data cache limit.
2434 * Purge some DNLC entries to release holds on meta-data.
2436 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2440 * Reclaim unused memory from all kmem caches.
2447 * An aggressive reclamation will shrink the cache size as well as
2448 * reap free buffers from the arc kmem caches.
2450 if (strat == ARC_RECLAIM_AGGR)
2453 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2454 if (zio_buf_cache[i] != prev_cache) {
2455 prev_cache = zio_buf_cache[i];
2456 kmem_cache_reap_now(zio_buf_cache[i]);
2458 if (zio_data_buf_cache[i] != prev_data_cache) {
2459 prev_data_cache = zio_data_buf_cache[i];
2460 kmem_cache_reap_now(zio_data_buf_cache[i]);
2463 kmem_cache_reap_now(buf_cache);
2464 kmem_cache_reap_now(hdr_cache);
2468 arc_reclaim_thread(void *dummy __unused)
2470 clock_t growtime = 0;
2471 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2474 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2476 mutex_enter(&arc_reclaim_thr_lock);
2477 while (arc_thread_exit == 0) {
2478 if (arc_reclaim_needed()) {
2481 if (last_reclaim == ARC_RECLAIM_CONS) {
2482 last_reclaim = ARC_RECLAIM_AGGR;
2484 last_reclaim = ARC_RECLAIM_CONS;
2488 last_reclaim = ARC_RECLAIM_AGGR;
2492 /* reset the growth delay for every reclaim */
2493 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2495 if (needfree && last_reclaim == ARC_RECLAIM_CONS) {
2497 * If needfree is TRUE our vm_lowmem hook
2498 * was called and in that case we must free some
2499 * memory, so switch to aggressive mode.
2502 last_reclaim = ARC_RECLAIM_AGGR;
2504 arc_kmem_reap_now(last_reclaim);
2507 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2508 arc_no_grow = FALSE;
2513 if (arc_eviction_list != NULL)
2514 arc_do_user_evicts();
2523 /* block until needed, or one second, whichever is shorter */
2524 CALLB_CPR_SAFE_BEGIN(&cpr);
2525 (void) cv_timedwait(&arc_reclaim_thr_cv,
2526 &arc_reclaim_thr_lock, hz);
2527 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2530 arc_thread_exit = 0;
2531 cv_broadcast(&arc_reclaim_thr_cv);
2532 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2537 * Adapt arc info given the number of bytes we are trying to add and
2538 * the state that we are comming from. This function is only called
2539 * when we are adding new content to the cache.
2542 arc_adapt(int bytes, arc_state_t *state)
2545 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2547 if (state == arc_l2c_only)
2552 * Adapt the target size of the MRU list:
2553 * - if we just hit in the MRU ghost list, then increase
2554 * the target size of the MRU list.
2555 * - if we just hit in the MFU ghost list, then increase
2556 * the target size of the MFU list by decreasing the
2557 * target size of the MRU list.
2559 if (state == arc_mru_ghost) {
2560 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2561 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2562 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2564 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2565 } else if (state == arc_mfu_ghost) {
2568 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2569 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2570 mult = MIN(mult, 10);
2572 delta = MIN(bytes * mult, arc_p);
2573 arc_p = MAX(arc_p_min, arc_p - delta);
2575 ASSERT((int64_t)arc_p >= 0);
2577 if (arc_reclaim_needed()) {
2578 cv_signal(&arc_reclaim_thr_cv);
2585 if (arc_c >= arc_c_max)
2589 * If we're within (2 * maxblocksize) bytes of the target
2590 * cache size, increment the target cache size
2592 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2593 atomic_add_64(&arc_c, (int64_t)bytes);
2594 if (arc_c > arc_c_max)
2596 else if (state == arc_anon)
2597 atomic_add_64(&arc_p, (int64_t)bytes);
2601 ASSERT((int64_t)arc_p >= 0);
2605 * Check if the cache has reached its limits and eviction is required
2609 arc_evict_needed(arc_buf_contents_t type)
2611 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2617 * If zio data pages are being allocated out of a separate heap segment,
2618 * then enforce that the size of available vmem for this area remains
2619 * above about 1/32nd free.
2621 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2622 vmem_size(zio_arena, VMEM_FREE) <
2623 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2628 if (arc_reclaim_needed())
2631 return (arc_size > arc_c);
2635 * The buffer, supplied as the first argument, needs a data block.
2636 * So, if we are at cache max, determine which cache should be victimized.
2637 * We have the following cases:
2639 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2640 * In this situation if we're out of space, but the resident size of the MFU is
2641 * under the limit, victimize the MFU cache to satisfy this insertion request.
2643 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2644 * Here, we've used up all of the available space for the MRU, so we need to
2645 * evict from our own cache instead. Evict from the set of resident MRU
2648 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2649 * c minus p represents the MFU space in the cache, since p is the size of the
2650 * cache that is dedicated to the MRU. In this situation there's still space on
2651 * the MFU side, so the MRU side needs to be victimized.
2653 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2654 * MFU's resident set is consuming more space than it has been allotted. In
2655 * this situation, we must victimize our own cache, the MFU, for this insertion.
2658 arc_get_data_buf(arc_buf_t *buf)
2660 arc_state_t *state = buf->b_hdr->b_state;
2661 uint64_t size = buf->b_hdr->b_size;
2662 arc_buf_contents_t type = buf->b_hdr->b_type;
2664 arc_adapt(size, state);
2667 * We have not yet reached cache maximum size,
2668 * just allocate a new buffer.
2670 if (!arc_evict_needed(type)) {
2671 if (type == ARC_BUFC_METADATA) {
2672 buf->b_data = zio_buf_alloc(size);
2673 arc_space_consume(size, ARC_SPACE_DATA);
2675 ASSERT(type == ARC_BUFC_DATA);
2676 buf->b_data = zio_data_buf_alloc(size);
2677 ARCSTAT_INCR(arcstat_data_size, size);
2678 atomic_add_64(&arc_size, size);
2684 * If we are prefetching from the mfu ghost list, this buffer
2685 * will end up on the mru list; so steal space from there.
2687 if (state == arc_mfu_ghost)
2688 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2689 else if (state == arc_mru_ghost)
2692 if (state == arc_mru || state == arc_anon) {
2693 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2694 state = (arc_mfu->arcs_lsize[type] >= size &&
2695 arc_p > mru_used) ? arc_mfu : arc_mru;
2698 uint64_t mfu_space = arc_c - arc_p;
2699 state = (arc_mru->arcs_lsize[type] >= size &&
2700 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2702 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2703 if (type == ARC_BUFC_METADATA) {
2704 buf->b_data = zio_buf_alloc(size);
2705 arc_space_consume(size, ARC_SPACE_DATA);
2707 ASSERT(type == ARC_BUFC_DATA);
2708 buf->b_data = zio_data_buf_alloc(size);
2709 ARCSTAT_INCR(arcstat_data_size, size);
2710 atomic_add_64(&arc_size, size);
2712 ARCSTAT_BUMP(arcstat_recycle_miss);
2714 ASSERT(buf->b_data != NULL);
2717 * Update the state size. Note that ghost states have a
2718 * "ghost size" and so don't need to be updated.
2720 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2721 arc_buf_hdr_t *hdr = buf->b_hdr;
2723 atomic_add_64(&hdr->b_state->arcs_size, size);
2724 if (list_link_active(&hdr->b_arc_node)) {
2725 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2726 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2729 * If we are growing the cache, and we are adding anonymous
2730 * data, and we have outgrown arc_p, update arc_p
2732 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2733 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2734 arc_p = MIN(arc_c, arc_p + size);
2736 ARCSTAT_BUMP(arcstat_allocated);
2740 * This routine is called whenever a buffer is accessed.
2741 * NOTE: the hash lock is dropped in this function.
2744 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2748 ASSERT(MUTEX_HELD(hash_lock));
2750 if (buf->b_state == arc_anon) {
2752 * This buffer is not in the cache, and does not
2753 * appear in our "ghost" list. Add the new buffer
2757 ASSERT(buf->b_arc_access == 0);
2758 buf->b_arc_access = ddi_get_lbolt();
2759 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2760 arc_change_state(arc_mru, buf, hash_lock);
2762 } else if (buf->b_state == arc_mru) {
2763 now = ddi_get_lbolt();
2766 * If this buffer is here because of a prefetch, then either:
2767 * - clear the flag if this is a "referencing" read
2768 * (any subsequent access will bump this into the MFU state).
2770 * - move the buffer to the head of the list if this is
2771 * another prefetch (to make it less likely to be evicted).
2773 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2774 if (refcount_count(&buf->b_refcnt) == 0) {
2775 ASSERT(list_link_active(&buf->b_arc_node));
2777 buf->b_flags &= ~ARC_PREFETCH;
2778 ARCSTAT_BUMP(arcstat_mru_hits);
2780 buf->b_arc_access = now;
2785 * This buffer has been "accessed" only once so far,
2786 * but it is still in the cache. Move it to the MFU
2789 if (now > buf->b_arc_access + ARC_MINTIME) {
2791 * More than 125ms have passed since we
2792 * instantiated this buffer. Move it to the
2793 * most frequently used state.
2795 buf->b_arc_access = now;
2796 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2797 arc_change_state(arc_mfu, buf, hash_lock);
2799 ARCSTAT_BUMP(arcstat_mru_hits);
2800 } else if (buf->b_state == arc_mru_ghost) {
2801 arc_state_t *new_state;
2803 * This buffer has been "accessed" recently, but
2804 * was evicted from the cache. Move it to the
2808 if (buf->b_flags & ARC_PREFETCH) {
2809 new_state = arc_mru;
2810 if (refcount_count(&buf->b_refcnt) > 0)
2811 buf->b_flags &= ~ARC_PREFETCH;
2812 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2814 new_state = arc_mfu;
2815 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2818 buf->b_arc_access = ddi_get_lbolt();
2819 arc_change_state(new_state, buf, hash_lock);
2821 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2822 } else if (buf->b_state == arc_mfu) {
2824 * This buffer has been accessed more than once and is
2825 * still in the cache. Keep it in the MFU state.
2827 * NOTE: an add_reference() that occurred when we did
2828 * the arc_read() will have kicked this off the list.
2829 * If it was a prefetch, we will explicitly move it to
2830 * the head of the list now.
2832 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2833 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2834 ASSERT(list_link_active(&buf->b_arc_node));
2836 ARCSTAT_BUMP(arcstat_mfu_hits);
2837 buf->b_arc_access = ddi_get_lbolt();
2838 } else if (buf->b_state == arc_mfu_ghost) {
2839 arc_state_t *new_state = arc_mfu;
2841 * This buffer has been accessed more than once but has
2842 * been evicted from the cache. Move it back to the
2846 if (buf->b_flags & ARC_PREFETCH) {
2848 * This is a prefetch access...
2849 * move this block back to the MRU state.
2851 ASSERT0(refcount_count(&buf->b_refcnt));
2852 new_state = arc_mru;
2855 buf->b_arc_access = ddi_get_lbolt();
2856 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2857 arc_change_state(new_state, buf, hash_lock);
2859 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2860 } else if (buf->b_state == arc_l2c_only) {
2862 * This buffer is on the 2nd Level ARC.
2865 buf->b_arc_access = ddi_get_lbolt();
2866 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2867 arc_change_state(arc_mfu, buf, hash_lock);
2869 ASSERT(!"invalid arc state");
2873 /* a generic arc_done_func_t which you can use */
2876 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2878 if (zio == NULL || zio->io_error == 0)
2879 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2880 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2883 /* a generic arc_done_func_t */
2885 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2887 arc_buf_t **bufp = arg;
2888 if (zio && zio->io_error) {
2889 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2893 ASSERT(buf->b_data);
2898 arc_read_done(zio_t *zio)
2900 arc_buf_hdr_t *hdr, *found;
2902 arc_buf_t *abuf; /* buffer we're assigning to callback */
2903 kmutex_t *hash_lock;
2904 arc_callback_t *callback_list, *acb;
2905 int freeable = FALSE;
2907 buf = zio->io_private;
2911 * The hdr was inserted into hash-table and removed from lists
2912 * prior to starting I/O. We should find this header, since
2913 * it's in the hash table, and it should be legit since it's
2914 * not possible to evict it during the I/O. The only possible
2915 * reason for it not to be found is if we were freed during the
2918 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2921 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2922 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2923 (found == hdr && HDR_L2_READING(hdr)));
2925 hdr->b_flags &= ~ARC_L2_EVICTED;
2926 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2927 hdr->b_flags &= ~ARC_L2CACHE;
2929 /* byteswap if necessary */
2930 callback_list = hdr->b_acb;
2931 ASSERT(callback_list != NULL);
2932 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2933 dmu_object_byteswap_t bswap =
2934 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2935 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2936 byteswap_uint64_array :
2937 dmu_ot_byteswap[bswap].ob_func;
2938 func(buf->b_data, hdr->b_size);
2941 arc_cksum_compute(buf, B_FALSE);
2944 #endif /* illumos */
2946 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2948 * Only call arc_access on anonymous buffers. This is because
2949 * if we've issued an I/O for an evicted buffer, we've already
2950 * called arc_access (to prevent any simultaneous readers from
2951 * getting confused).
2953 arc_access(hdr, hash_lock);
2956 /* create copies of the data buffer for the callers */
2958 for (acb = callback_list; acb; acb = acb->acb_next) {
2959 if (acb->acb_done) {
2961 ARCSTAT_BUMP(arcstat_duplicate_reads);
2962 abuf = arc_buf_clone(buf);
2964 acb->acb_buf = abuf;
2969 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2970 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2972 ASSERT(buf->b_efunc == NULL);
2973 ASSERT(hdr->b_datacnt == 1);
2974 hdr->b_flags |= ARC_BUF_AVAILABLE;
2977 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2979 if (zio->io_error != 0) {
2980 hdr->b_flags |= ARC_IO_ERROR;
2981 if (hdr->b_state != arc_anon)
2982 arc_change_state(arc_anon, hdr, hash_lock);
2983 if (HDR_IN_HASH_TABLE(hdr))
2984 buf_hash_remove(hdr);
2985 freeable = refcount_is_zero(&hdr->b_refcnt);
2989 * Broadcast before we drop the hash_lock to avoid the possibility
2990 * that the hdr (and hence the cv) might be freed before we get to
2991 * the cv_broadcast().
2993 cv_broadcast(&hdr->b_cv);
2996 mutex_exit(hash_lock);
2999 * This block was freed while we waited for the read to
3000 * complete. It has been removed from the hash table and
3001 * moved to the anonymous state (so that it won't show up
3004 ASSERT3P(hdr->b_state, ==, arc_anon);
3005 freeable = refcount_is_zero(&hdr->b_refcnt);
3008 /* execute each callback and free its structure */
3009 while ((acb = callback_list) != NULL) {
3011 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3013 if (acb->acb_zio_dummy != NULL) {
3014 acb->acb_zio_dummy->io_error = zio->io_error;
3015 zio_nowait(acb->acb_zio_dummy);
3018 callback_list = acb->acb_next;
3019 kmem_free(acb, sizeof (arc_callback_t));
3023 arc_hdr_destroy(hdr);
3027 * "Read" the block block at the specified DVA (in bp) via the
3028 * cache. If the block is found in the cache, invoke the provided
3029 * callback immediately and return. Note that the `zio' parameter
3030 * in the callback will be NULL in this case, since no IO was
3031 * required. If the block is not in the cache pass the read request
3032 * on to the spa with a substitute callback function, so that the
3033 * requested block will be added to the cache.
3035 * If a read request arrives for a block that has a read in-progress,
3036 * either wait for the in-progress read to complete (and return the
3037 * results); or, if this is a read with a "done" func, add a record
3038 * to the read to invoke the "done" func when the read completes,
3039 * and return; or just return.
3041 * arc_read_done() will invoke all the requested "done" functions
3042 * for readers of this block.
3044 * Normal callers should use arc_read and pass the arc buffer and offset
3045 * for the bp. But if you know you don't need locking, you can use
3049 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_buf_t *pbuf,
3050 arc_done_func_t *done, void *private, int priority, int zio_flags,
3051 uint32_t *arc_flags, const zbookmark_t *zb)
3057 * XXX This happens from traverse callback funcs, for
3058 * the objset_phys_t block.
3060 return (arc_read_nolock(pio, spa, bp, done, private, priority,
3061 zio_flags, arc_flags, zb));
3064 ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt));
3065 ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size);
3066 rw_enter(&pbuf->b_data_lock, RW_READER);
3068 err = arc_read_nolock(pio, spa, bp, done, private, priority,
3069 zio_flags, arc_flags, zb);
3070 rw_exit(&pbuf->b_data_lock);
3076 arc_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bp,
3077 arc_done_func_t *done, void *private, int priority, int zio_flags,
3078 uint32_t *arc_flags, const zbookmark_t *zb)
3082 kmutex_t *hash_lock;
3084 uint64_t guid = spa_load_guid(spa);
3087 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3089 if (hdr && hdr->b_datacnt > 0) {
3091 *arc_flags |= ARC_CACHED;
3093 if (HDR_IO_IN_PROGRESS(hdr)) {
3095 if (*arc_flags & ARC_WAIT) {
3096 cv_wait(&hdr->b_cv, hash_lock);
3097 mutex_exit(hash_lock);
3100 ASSERT(*arc_flags & ARC_NOWAIT);
3103 arc_callback_t *acb = NULL;
3105 acb = kmem_zalloc(sizeof (arc_callback_t),
3107 acb->acb_done = done;
3108 acb->acb_private = private;
3110 acb->acb_zio_dummy = zio_null(pio,
3111 spa, NULL, NULL, NULL, zio_flags);
3113 ASSERT(acb->acb_done != NULL);
3114 acb->acb_next = hdr->b_acb;
3116 add_reference(hdr, hash_lock, private);
3117 mutex_exit(hash_lock);
3120 mutex_exit(hash_lock);
3124 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3127 add_reference(hdr, hash_lock, private);
3129 * If this block is already in use, create a new
3130 * copy of the data so that we will be guaranteed
3131 * that arc_release() will always succeed.
3135 ASSERT(buf->b_data);
3136 if (HDR_BUF_AVAILABLE(hdr)) {
3137 ASSERT(buf->b_efunc == NULL);
3138 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3140 buf = arc_buf_clone(buf);
3143 } else if (*arc_flags & ARC_PREFETCH &&
3144 refcount_count(&hdr->b_refcnt) == 0) {
3145 hdr->b_flags |= ARC_PREFETCH;
3147 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
3148 arc_access(hdr, hash_lock);
3149 if (*arc_flags & ARC_L2CACHE)
3150 hdr->b_flags |= ARC_L2CACHE;
3151 mutex_exit(hash_lock);
3152 ARCSTAT_BUMP(arcstat_hits);
3153 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3154 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3155 data, metadata, hits);
3158 done(NULL, buf, private);
3160 uint64_t size = BP_GET_LSIZE(bp);
3161 arc_callback_t *acb;
3164 boolean_t devw = B_FALSE;
3167 /* this block is not in the cache */
3168 arc_buf_hdr_t *exists;
3169 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
3170 buf = arc_buf_alloc(spa, size, private, type);
3172 hdr->b_dva = *BP_IDENTITY(bp);
3173 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3174 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3175 exists = buf_hash_insert(hdr, &hash_lock);
3177 /* somebody beat us to the hash insert */
3178 mutex_exit(hash_lock);
3179 buf_discard_identity(hdr);
3180 (void) arc_buf_remove_ref(buf, private);
3181 goto top; /* restart the IO request */
3183 /* if this is a prefetch, we don't have a reference */
3184 if (*arc_flags & ARC_PREFETCH) {
3185 (void) remove_reference(hdr, hash_lock,
3187 hdr->b_flags |= ARC_PREFETCH;
3189 if (*arc_flags & ARC_L2CACHE)
3190 hdr->b_flags |= ARC_L2CACHE;
3191 if (BP_GET_LEVEL(bp) > 0)
3192 hdr->b_flags |= ARC_INDIRECT;
3194 /* this block is in the ghost cache */
3195 ASSERT(GHOST_STATE(hdr->b_state));
3196 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3197 ASSERT0(refcount_count(&hdr->b_refcnt));
3198 ASSERT(hdr->b_buf == NULL);
3200 /* if this is a prefetch, we don't have a reference */
3201 if (*arc_flags & ARC_PREFETCH)
3202 hdr->b_flags |= ARC_PREFETCH;
3204 add_reference(hdr, hash_lock, private);
3205 if (*arc_flags & ARC_L2CACHE)
3206 hdr->b_flags |= ARC_L2CACHE;
3207 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3210 buf->b_efunc = NULL;
3211 buf->b_private = NULL;
3214 ASSERT(hdr->b_datacnt == 0);
3216 arc_get_data_buf(buf);
3217 arc_access(hdr, hash_lock);
3220 ASSERT(!GHOST_STATE(hdr->b_state));
3222 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
3223 acb->acb_done = done;
3224 acb->acb_private = private;
3226 ASSERT(hdr->b_acb == NULL);
3228 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3230 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3231 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3232 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3233 addr = hdr->b_l2hdr->b_daddr;
3235 * Lock out device removal.
3237 if (vdev_is_dead(vd) ||
3238 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3242 mutex_exit(hash_lock);
3244 ASSERT3U(hdr->b_size, ==, size);
3245 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3246 uint64_t, size, zbookmark_t *, zb);
3247 ARCSTAT_BUMP(arcstat_misses);
3248 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3249 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3250 data, metadata, misses);
3252 curthread->td_ru.ru_inblock++;
3255 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3257 * Read from the L2ARC if the following are true:
3258 * 1. The L2ARC vdev was previously cached.
3259 * 2. This buffer still has L2ARC metadata.
3260 * 3. This buffer isn't currently writing to the L2ARC.
3261 * 4. The L2ARC entry wasn't evicted, which may
3262 * also have invalidated the vdev.
3263 * 5. This isn't prefetch and l2arc_noprefetch is set.
3265 if (hdr->b_l2hdr != NULL &&
3266 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3267 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3268 l2arc_read_callback_t *cb;
3270 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3271 ARCSTAT_BUMP(arcstat_l2_hits);
3273 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3275 cb->l2rcb_buf = buf;
3276 cb->l2rcb_spa = spa;
3279 cb->l2rcb_flags = zio_flags;
3282 * l2arc read. The SCL_L2ARC lock will be
3283 * released by l2arc_read_done().
3285 rzio = zio_read_phys(pio, vd, addr, size,
3286 buf->b_data, ZIO_CHECKSUM_OFF,
3287 l2arc_read_done, cb, priority, zio_flags |
3288 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
3289 ZIO_FLAG_DONT_PROPAGATE |
3290 ZIO_FLAG_DONT_RETRY, B_FALSE);
3291 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3293 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
3295 if (*arc_flags & ARC_NOWAIT) {
3300 ASSERT(*arc_flags & ARC_WAIT);
3301 if (zio_wait(rzio) == 0)
3304 /* l2arc read error; goto zio_read() */
3306 DTRACE_PROBE1(l2arc__miss,
3307 arc_buf_hdr_t *, hdr);
3308 ARCSTAT_BUMP(arcstat_l2_misses);
3309 if (HDR_L2_WRITING(hdr))
3310 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3311 spa_config_exit(spa, SCL_L2ARC, vd);
3315 spa_config_exit(spa, SCL_L2ARC, vd);
3316 if (l2arc_ndev != 0) {
3317 DTRACE_PROBE1(l2arc__miss,
3318 arc_buf_hdr_t *, hdr);
3319 ARCSTAT_BUMP(arcstat_l2_misses);
3323 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3324 arc_read_done, buf, priority, zio_flags, zb);
3326 if (*arc_flags & ARC_WAIT)
3327 return (zio_wait(rzio));
3329 ASSERT(*arc_flags & ARC_NOWAIT);
3336 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3338 ASSERT(buf->b_hdr != NULL);
3339 ASSERT(buf->b_hdr->b_state != arc_anon);
3340 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3341 ASSERT(buf->b_efunc == NULL);
3342 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3344 buf->b_efunc = func;
3345 buf->b_private = private;
3349 * This is used by the DMU to let the ARC know that a buffer is
3350 * being evicted, so the ARC should clean up. If this arc buf
3351 * is not yet in the evicted state, it will be put there.
3354 arc_buf_evict(arc_buf_t *buf)
3357 kmutex_t *hash_lock;
3359 list_t *list, *evicted_list;
3360 kmutex_t *lock, *evicted_lock;
3362 mutex_enter(&buf->b_evict_lock);
3366 * We are in arc_do_user_evicts().
3368 ASSERT(buf->b_data == NULL);
3369 mutex_exit(&buf->b_evict_lock);
3371 } else if (buf->b_data == NULL) {
3372 arc_buf_t copy = *buf; /* structure assignment */
3374 * We are on the eviction list; process this buffer now
3375 * but let arc_do_user_evicts() do the reaping.
3377 buf->b_efunc = NULL;
3378 mutex_exit(&buf->b_evict_lock);
3379 VERIFY(copy.b_efunc(©) == 0);
3382 hash_lock = HDR_LOCK(hdr);
3383 mutex_enter(hash_lock);
3385 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3387 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3388 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3391 * Pull this buffer off of the hdr
3394 while (*bufp != buf)
3395 bufp = &(*bufp)->b_next;
3396 *bufp = buf->b_next;
3398 ASSERT(buf->b_data != NULL);
3399 arc_buf_destroy(buf, FALSE, FALSE);
3401 if (hdr->b_datacnt == 0) {
3402 arc_state_t *old_state = hdr->b_state;
3403 arc_state_t *evicted_state;
3405 ASSERT(hdr->b_buf == NULL);
3406 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3409 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3411 get_buf_info(hdr, old_state, &list, &lock);
3412 get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock);
3414 mutex_enter(evicted_lock);
3416 arc_change_state(evicted_state, hdr, hash_lock);
3417 ASSERT(HDR_IN_HASH_TABLE(hdr));
3418 hdr->b_flags |= ARC_IN_HASH_TABLE;
3419 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3421 mutex_exit(evicted_lock);
3424 mutex_exit(hash_lock);
3425 mutex_exit(&buf->b_evict_lock);
3427 VERIFY(buf->b_efunc(buf) == 0);
3428 buf->b_efunc = NULL;
3429 buf->b_private = NULL;
3432 kmem_cache_free(buf_cache, buf);
3437 * Release this buffer from the cache. This must be done
3438 * after a read and prior to modifying the buffer contents.
3439 * If the buffer has more than one reference, we must make
3440 * a new hdr for the buffer.
3443 arc_release(arc_buf_t *buf, void *tag)
3446 kmutex_t *hash_lock = NULL;
3447 l2arc_buf_hdr_t *l2hdr;
3451 * It would be nice to assert that if it's DMU metadata (level >
3452 * 0 || it's the dnode file), then it must be syncing context.
3453 * But we don't know that information at this level.
3456 mutex_enter(&buf->b_evict_lock);
3459 /* this buffer is not on any list */
3460 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3462 if (hdr->b_state == arc_anon) {
3463 /* this buffer is already released */
3464 ASSERT(buf->b_efunc == NULL);
3466 hash_lock = HDR_LOCK(hdr);
3467 mutex_enter(hash_lock);
3469 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3472 l2hdr = hdr->b_l2hdr;
3474 mutex_enter(&l2arc_buflist_mtx);
3475 hdr->b_l2hdr = NULL;
3476 buf_size = hdr->b_size;
3480 * Do we have more than one buf?
3482 if (hdr->b_datacnt > 1) {
3483 arc_buf_hdr_t *nhdr;
3485 uint64_t blksz = hdr->b_size;
3486 uint64_t spa = hdr->b_spa;
3487 arc_buf_contents_t type = hdr->b_type;
3488 uint32_t flags = hdr->b_flags;
3490 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3492 * Pull the data off of this hdr and attach it to
3493 * a new anonymous hdr.
3495 (void) remove_reference(hdr, hash_lock, tag);
3497 while (*bufp != buf)
3498 bufp = &(*bufp)->b_next;
3499 *bufp = buf->b_next;
3502 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3503 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3504 if (refcount_is_zero(&hdr->b_refcnt)) {
3505 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3506 ASSERT3U(*size, >=, hdr->b_size);
3507 atomic_add_64(size, -hdr->b_size);
3511 * We're releasing a duplicate user data buffer, update
3512 * our statistics accordingly.
3514 if (hdr->b_type == ARC_BUFC_DATA) {
3515 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3516 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3519 hdr->b_datacnt -= 1;
3520 arc_cksum_verify(buf);
3522 arc_buf_unwatch(buf);
3523 #endif /* illumos */
3525 mutex_exit(hash_lock);
3527 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3528 nhdr->b_size = blksz;
3530 nhdr->b_type = type;
3532 nhdr->b_state = arc_anon;
3533 nhdr->b_arc_access = 0;
3534 nhdr->b_flags = flags & ARC_L2_WRITING;
3535 nhdr->b_l2hdr = NULL;
3536 nhdr->b_datacnt = 1;
3537 nhdr->b_freeze_cksum = NULL;
3538 (void) refcount_add(&nhdr->b_refcnt, tag);
3540 mutex_exit(&buf->b_evict_lock);
3541 atomic_add_64(&arc_anon->arcs_size, blksz);
3543 mutex_exit(&buf->b_evict_lock);
3544 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3545 ASSERT(!list_link_active(&hdr->b_arc_node));
3546 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3547 if (hdr->b_state != arc_anon)
3548 arc_change_state(arc_anon, hdr, hash_lock);
3549 hdr->b_arc_access = 0;
3551 mutex_exit(hash_lock);
3553 buf_discard_identity(hdr);
3556 buf->b_efunc = NULL;
3557 buf->b_private = NULL;
3560 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3561 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3562 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3563 mutex_exit(&l2arc_buflist_mtx);
3568 * Release this buffer. If it does not match the provided BP, fill it
3569 * with that block's contents.
3573 arc_release_bp(arc_buf_t *buf, void *tag, blkptr_t *bp, spa_t *spa,
3576 arc_release(buf, tag);
3581 arc_released(arc_buf_t *buf)
3585 mutex_enter(&buf->b_evict_lock);
3586 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3587 mutex_exit(&buf->b_evict_lock);
3592 arc_has_callback(arc_buf_t *buf)
3596 mutex_enter(&buf->b_evict_lock);
3597 callback = (buf->b_efunc != NULL);
3598 mutex_exit(&buf->b_evict_lock);
3604 arc_referenced(arc_buf_t *buf)
3608 mutex_enter(&buf->b_evict_lock);
3609 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3610 mutex_exit(&buf->b_evict_lock);
3611 return (referenced);
3616 arc_write_ready(zio_t *zio)
3618 arc_write_callback_t *callback = zio->io_private;
3619 arc_buf_t *buf = callback->awcb_buf;
3620 arc_buf_hdr_t *hdr = buf->b_hdr;
3622 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3623 callback->awcb_ready(zio, buf, callback->awcb_private);
3626 * If the IO is already in progress, then this is a re-write
3627 * attempt, so we need to thaw and re-compute the cksum.
3628 * It is the responsibility of the callback to handle the
3629 * accounting for any re-write attempt.
3631 if (HDR_IO_IN_PROGRESS(hdr)) {
3632 mutex_enter(&hdr->b_freeze_lock);
3633 if (hdr->b_freeze_cksum != NULL) {
3634 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3635 hdr->b_freeze_cksum = NULL;
3637 mutex_exit(&hdr->b_freeze_lock);
3639 arc_cksum_compute(buf, B_FALSE);
3640 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3644 arc_write_done(zio_t *zio)
3646 arc_write_callback_t *callback = zio->io_private;
3647 arc_buf_t *buf = callback->awcb_buf;
3648 arc_buf_hdr_t *hdr = buf->b_hdr;
3650 ASSERT(hdr->b_acb == NULL);
3652 if (zio->io_error == 0) {
3653 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3654 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3655 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3657 ASSERT(BUF_EMPTY(hdr));
3661 * If the block to be written was all-zero, we may have
3662 * compressed it away. In this case no write was performed
3663 * so there will be no dva/birth/checksum. The buffer must
3664 * therefore remain anonymous (and uncached).
3666 if (!BUF_EMPTY(hdr)) {
3667 arc_buf_hdr_t *exists;
3668 kmutex_t *hash_lock;
3670 ASSERT(zio->io_error == 0);
3672 arc_cksum_verify(buf);
3674 exists = buf_hash_insert(hdr, &hash_lock);
3677 * This can only happen if we overwrite for
3678 * sync-to-convergence, because we remove
3679 * buffers from the hash table when we arc_free().
3681 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3682 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3683 panic("bad overwrite, hdr=%p exists=%p",
3684 (void *)hdr, (void *)exists);
3685 ASSERT(refcount_is_zero(&exists->b_refcnt));
3686 arc_change_state(arc_anon, exists, hash_lock);
3687 mutex_exit(hash_lock);
3688 arc_hdr_destroy(exists);
3689 exists = buf_hash_insert(hdr, &hash_lock);
3690 ASSERT3P(exists, ==, NULL);
3691 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3693 ASSERT(zio->io_prop.zp_nopwrite);
3694 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3695 panic("bad nopwrite, hdr=%p exists=%p",
3696 (void *)hdr, (void *)exists);
3699 ASSERT(hdr->b_datacnt == 1);
3700 ASSERT(hdr->b_state == arc_anon);
3701 ASSERT(BP_GET_DEDUP(zio->io_bp));
3702 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3705 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3706 /* if it's not anon, we are doing a scrub */
3707 if (!exists && hdr->b_state == arc_anon)
3708 arc_access(hdr, hash_lock);
3709 mutex_exit(hash_lock);
3711 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3714 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3715 callback->awcb_done(zio, buf, callback->awcb_private);
3717 kmem_free(callback, sizeof (arc_write_callback_t));
3721 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3722 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3723 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3724 int priority, int zio_flags, const zbookmark_t *zb)
3726 arc_buf_hdr_t *hdr = buf->b_hdr;
3727 arc_write_callback_t *callback;
3730 ASSERT(ready != NULL);
3731 ASSERT(done != NULL);
3732 ASSERT(!HDR_IO_ERROR(hdr));
3733 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3734 ASSERT(hdr->b_acb == NULL);
3736 hdr->b_flags |= ARC_L2CACHE;
3737 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3738 callback->awcb_ready = ready;
3739 callback->awcb_done = done;
3740 callback->awcb_private = private;
3741 callback->awcb_buf = buf;
3743 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3744 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3750 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3753 uint64_t available_memory =
3754 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count);
3755 static uint64_t page_load = 0;
3756 static uint64_t last_txg = 0;
3761 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3764 if (available_memory >= zfs_write_limit_max)
3767 if (txg > last_txg) {
3772 * If we are in pageout, we know that memory is already tight,
3773 * the arc is already going to be evicting, so we just want to
3774 * continue to let page writes occur as quickly as possible.
3776 if (curproc == pageproc) {
3777 if (page_load > available_memory / 4)
3779 /* Note: reserve is inflated, so we deflate */
3780 page_load += reserve / 8;
3782 } else if (page_load > 0 && arc_reclaim_needed()) {
3783 /* memory is low, delay before restarting */
3784 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3789 if (arc_size > arc_c_min) {
3790 uint64_t evictable_memory =
3791 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3792 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3793 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3794 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3795 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3798 if (inflight_data > available_memory / 4) {
3799 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3807 arc_tempreserve_clear(uint64_t reserve)
3809 atomic_add_64(&arc_tempreserve, -reserve);
3810 ASSERT((int64_t)arc_tempreserve >= 0);
3814 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3821 * Once in a while, fail for no reason. Everything should cope.
3823 if (spa_get_random(10000) == 0) {
3824 dprintf("forcing random failure\n");
3828 if (reserve > arc_c/4 && !arc_no_grow)
3829 arc_c = MIN(arc_c_max, reserve * 4);
3830 if (reserve > arc_c)
3834 * Don't count loaned bufs as in flight dirty data to prevent long
3835 * network delays from blocking transactions that are ready to be
3836 * assigned to a txg.
3838 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3841 * Writes will, almost always, require additional memory allocations
3842 * in order to compress/encrypt/etc the data. We therefor need to
3843 * make sure that there is sufficient available memory for this.
3845 if (error = arc_memory_throttle(reserve, anon_size, txg))
3849 * Throttle writes when the amount of dirty data in the cache
3850 * gets too large. We try to keep the cache less than half full
3851 * of dirty blocks so that our sync times don't grow too large.
3852 * Note: if two requests come in concurrently, we might let them
3853 * both succeed, when one of them should fail. Not a huge deal.
3856 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3857 anon_size > arc_c / 4) {
3858 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3859 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3860 arc_tempreserve>>10,
3861 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3862 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3863 reserve>>10, arc_c>>10);
3866 atomic_add_64(&arc_tempreserve, reserve);
3870 static kmutex_t arc_lowmem_lock;
3872 static eventhandler_tag arc_event_lowmem = NULL;
3875 arc_lowmem(void *arg __unused, int howto __unused)
3878 /* Serialize access via arc_lowmem_lock. */
3879 mutex_enter(&arc_lowmem_lock);
3880 mutex_enter(&arc_reclaim_thr_lock);
3882 cv_signal(&arc_reclaim_thr_cv);
3885 * It is unsafe to block here in arbitrary threads, because we can come
3886 * here from ARC itself and may hold ARC locks and thus risk a deadlock
3887 * with ARC reclaim thread.
3889 if (curproc == pageproc) {
3891 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0);
3893 mutex_exit(&arc_reclaim_thr_lock);
3894 mutex_exit(&arc_lowmem_lock);
3901 int i, prefetch_tunable_set = 0;
3903 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3904 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3905 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL);
3907 /* Convert seconds to clock ticks */
3908 arc_min_prefetch_lifespan = 1 * hz;
3910 /* Start out with 1/8 of all memory */
3911 arc_c = kmem_size() / 8;
3916 * On architectures where the physical memory can be larger
3917 * than the addressable space (intel in 32-bit mode), we may
3918 * need to limit the cache to 1/8 of VM size.
3920 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3923 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
3924 arc_c_min = MAX(arc_c / 4, 64<<18);
3925 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
3926 if (arc_c * 8 >= 1<<30)
3927 arc_c_max = (arc_c * 8) - (1<<30);
3929 arc_c_max = arc_c_min;
3930 arc_c_max = MAX(arc_c * 5, arc_c_max);
3934 * Allow the tunables to override our calculations if they are
3935 * reasonable (ie. over 16MB)
3937 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size())
3938 arc_c_max = zfs_arc_max;
3939 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max)
3940 arc_c_min = zfs_arc_min;
3944 arc_p = (arc_c >> 1);
3946 /* limit meta-data to 1/4 of the arc capacity */
3947 arc_meta_limit = arc_c_max / 4;
3949 /* Allow the tunable to override if it is reasonable */
3950 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3951 arc_meta_limit = zfs_arc_meta_limit;
3953 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3954 arc_c_min = arc_meta_limit / 2;
3956 if (zfs_arc_grow_retry > 0)
3957 arc_grow_retry = zfs_arc_grow_retry;
3959 if (zfs_arc_shrink_shift > 0)
3960 arc_shrink_shift = zfs_arc_shrink_shift;
3962 if (zfs_arc_p_min_shift > 0)
3963 arc_p_min_shift = zfs_arc_p_min_shift;
3965 /* if kmem_flags are set, lets try to use less memory */
3966 if (kmem_debugging())
3968 if (arc_c < arc_c_min)
3971 zfs_arc_min = arc_c_min;
3972 zfs_arc_max = arc_c_max;
3974 arc_anon = &ARC_anon;
3976 arc_mru_ghost = &ARC_mru_ghost;
3978 arc_mfu_ghost = &ARC_mfu_ghost;
3979 arc_l2c_only = &ARC_l2c_only;
3982 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
3983 mutex_init(&arc_anon->arcs_locks[i].arcs_lock,
3984 NULL, MUTEX_DEFAULT, NULL);
3985 mutex_init(&arc_mru->arcs_locks[i].arcs_lock,
3986 NULL, MUTEX_DEFAULT, NULL);
3987 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock,
3988 NULL, MUTEX_DEFAULT, NULL);
3989 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock,
3990 NULL, MUTEX_DEFAULT, NULL);
3991 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock,
3992 NULL, MUTEX_DEFAULT, NULL);
3993 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock,
3994 NULL, MUTEX_DEFAULT, NULL);
3996 list_create(&arc_mru->arcs_lists[i],
3997 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3998 list_create(&arc_mru_ghost->arcs_lists[i],
3999 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4000 list_create(&arc_mfu->arcs_lists[i],
4001 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4002 list_create(&arc_mfu_ghost->arcs_lists[i],
4003 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4004 list_create(&arc_mfu_ghost->arcs_lists[i],
4005 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4006 list_create(&arc_l2c_only->arcs_lists[i],
4007 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4012 arc_thread_exit = 0;
4013 arc_eviction_list = NULL;
4014 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
4015 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
4017 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
4018 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
4020 if (arc_ksp != NULL) {
4021 arc_ksp->ks_data = &arc_stats;
4022 kstat_install(arc_ksp);
4025 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
4026 TS_RUN, minclsyspri);
4029 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
4030 EVENTHANDLER_PRI_FIRST);
4036 if (zfs_write_limit_max == 0)
4037 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
4039 zfs_write_limit_shift = 0;
4040 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
4043 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
4044 prefetch_tunable_set = 1;
4047 if (prefetch_tunable_set == 0) {
4048 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
4050 printf(" add \"vfs.zfs.prefetch_disable=0\" "
4051 "to /boot/loader.conf.\n");
4052 zfs_prefetch_disable = 1;
4055 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
4056 prefetch_tunable_set == 0) {
4057 printf("ZFS NOTICE: Prefetch is disabled by default if less "
4058 "than 4GB of RAM is present;\n"
4059 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
4060 "to /boot/loader.conf.\n");
4061 zfs_prefetch_disable = 1;
4064 /* Warn about ZFS memory and address space requirements. */
4065 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
4066 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
4067 "expect unstable behavior.\n");
4069 if (kmem_size() < 512 * (1 << 20)) {
4070 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
4071 "expect unstable behavior.\n");
4072 printf(" Consider tuning vm.kmem_size and "
4073 "vm.kmem_size_max\n");
4074 printf(" in /boot/loader.conf.\n");
4084 mutex_enter(&arc_reclaim_thr_lock);
4085 arc_thread_exit = 1;
4086 cv_signal(&arc_reclaim_thr_cv);
4087 while (arc_thread_exit != 0)
4088 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
4089 mutex_exit(&arc_reclaim_thr_lock);
4095 if (arc_ksp != NULL) {
4096 kstat_delete(arc_ksp);
4100 mutex_destroy(&arc_eviction_mtx);
4101 mutex_destroy(&arc_reclaim_thr_lock);
4102 cv_destroy(&arc_reclaim_thr_cv);
4104 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
4105 list_destroy(&arc_mru->arcs_lists[i]);
4106 list_destroy(&arc_mru_ghost->arcs_lists[i]);
4107 list_destroy(&arc_mfu->arcs_lists[i]);
4108 list_destroy(&arc_mfu_ghost->arcs_lists[i]);
4109 list_destroy(&arc_l2c_only->arcs_lists[i]);
4111 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock);
4112 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock);
4113 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock);
4114 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock);
4115 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock);
4116 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock);
4119 mutex_destroy(&zfs_write_limit_lock);
4123 ASSERT(arc_loaned_bytes == 0);
4125 mutex_destroy(&arc_lowmem_lock);
4127 if (arc_event_lowmem != NULL)
4128 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
4135 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
4136 * It uses dedicated storage devices to hold cached data, which are populated
4137 * using large infrequent writes. The main role of this cache is to boost
4138 * the performance of random read workloads. The intended L2ARC devices
4139 * include short-stroked disks, solid state disks, and other media with
4140 * substantially faster read latency than disk.
4142 * +-----------------------+
4144 * +-----------------------+
4147 * l2arc_feed_thread() arc_read()
4151 * +---------------+ |
4153 * +---------------+ |
4158 * +-------+ +-------+
4160 * | cache | | cache |
4161 * +-------+ +-------+
4162 * +=========+ .-----.
4163 * : L2ARC : |-_____-|
4164 * : devices : | Disks |
4165 * +=========+ `-_____-'
4167 * Read requests are satisfied from the following sources, in order:
4170 * 2) vdev cache of L2ARC devices
4172 * 4) vdev cache of disks
4175 * Some L2ARC device types exhibit extremely slow write performance.
4176 * To accommodate for this there are some significant differences between
4177 * the L2ARC and traditional cache design:
4179 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4180 * the ARC behave as usual, freeing buffers and placing headers on ghost
4181 * lists. The ARC does not send buffers to the L2ARC during eviction as
4182 * this would add inflated write latencies for all ARC memory pressure.
4184 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4185 * It does this by periodically scanning buffers from the eviction-end of
4186 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4187 * not already there. It scans until a headroom of buffers is satisfied,
4188 * which itself is a buffer for ARC eviction. The thread that does this is
4189 * l2arc_feed_thread(), illustrated below; example sizes are included to
4190 * provide a better sense of ratio than this diagram:
4193 * +---------------------+----------+
4194 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4195 * +---------------------+----------+ | o L2ARC eligible
4196 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4197 * +---------------------+----------+ |
4198 * 15.9 Gbytes ^ 32 Mbytes |
4200 * l2arc_feed_thread()
4202 * l2arc write hand <--[oooo]--'
4206 * +==============================+
4207 * L2ARC dev |####|#|###|###| |####| ... |
4208 * +==============================+
4211 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4212 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4213 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4214 * safe to say that this is an uncommon case, since buffers at the end of
4215 * the ARC lists have moved there due to inactivity.
4217 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4218 * then the L2ARC simply misses copying some buffers. This serves as a
4219 * pressure valve to prevent heavy read workloads from both stalling the ARC
4220 * with waits and clogging the L2ARC with writes. This also helps prevent
4221 * the potential for the L2ARC to churn if it attempts to cache content too
4222 * quickly, such as during backups of the entire pool.
4224 * 5. After system boot and before the ARC has filled main memory, there are
4225 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4226 * lists can remain mostly static. Instead of searching from tail of these
4227 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4228 * for eligible buffers, greatly increasing its chance of finding them.
4230 * The L2ARC device write speed is also boosted during this time so that
4231 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4232 * there are no L2ARC reads, and no fear of degrading read performance
4233 * through increased writes.
4235 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4236 * the vdev queue can aggregate them into larger and fewer writes. Each
4237 * device is written to in a rotor fashion, sweeping writes through
4238 * available space then repeating.
4240 * 7. The L2ARC does not store dirty content. It never needs to flush
4241 * write buffers back to disk based storage.
4243 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4244 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4246 * The performance of the L2ARC can be tweaked by a number of tunables, which
4247 * may be necessary for different workloads:
4249 * l2arc_write_max max write bytes per interval
4250 * l2arc_write_boost extra write bytes during device warmup
4251 * l2arc_noprefetch skip caching prefetched buffers
4252 * l2arc_headroom number of max device writes to precache
4253 * l2arc_feed_secs seconds between L2ARC writing
4255 * Tunables may be removed or added as future performance improvements are
4256 * integrated, and also may become zpool properties.
4258 * There are three key functions that control how the L2ARC warms up:
4260 * l2arc_write_eligible() check if a buffer is eligible to cache
4261 * l2arc_write_size() calculate how much to write
4262 * l2arc_write_interval() calculate sleep delay between writes
4264 * These three functions determine what to write, how much, and how quickly
4269 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4272 * A buffer is *not* eligible for the L2ARC if it:
4273 * 1. belongs to a different spa.
4274 * 2. is already cached on the L2ARC.
4275 * 3. has an I/O in progress (it may be an incomplete read).
4276 * 4. is flagged not eligible (zfs property).
4278 if (ab->b_spa != spa_guid) {
4279 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
4282 if (ab->b_l2hdr != NULL) {
4283 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
4286 if (HDR_IO_IN_PROGRESS(ab)) {
4287 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
4290 if (!HDR_L2CACHE(ab)) {
4291 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
4299 l2arc_write_size(l2arc_dev_t *dev)
4303 size = dev->l2ad_write;
4305 if (arc_warm == B_FALSE)
4306 size += dev->l2ad_boost;
4313 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4315 clock_t interval, next, now;
4318 * If the ARC lists are busy, increase our write rate; if the
4319 * lists are stale, idle back. This is achieved by checking
4320 * how much we previously wrote - if it was more than half of
4321 * what we wanted, schedule the next write much sooner.
4323 if (l2arc_feed_again && wrote > (wanted / 2))
4324 interval = (hz * l2arc_feed_min_ms) / 1000;
4326 interval = hz * l2arc_feed_secs;
4328 now = ddi_get_lbolt();
4329 next = MAX(now, MIN(now + interval, began + interval));
4335 l2arc_hdr_stat_add(void)
4337 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4338 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4342 l2arc_hdr_stat_remove(void)
4344 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4345 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4349 * Cycle through L2ARC devices. This is how L2ARC load balances.
4350 * If a device is returned, this also returns holding the spa config lock.
4352 static l2arc_dev_t *
4353 l2arc_dev_get_next(void)
4355 l2arc_dev_t *first, *next = NULL;
4358 * Lock out the removal of spas (spa_namespace_lock), then removal
4359 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4360 * both locks will be dropped and a spa config lock held instead.
4362 mutex_enter(&spa_namespace_lock);
4363 mutex_enter(&l2arc_dev_mtx);
4365 /* if there are no vdevs, there is nothing to do */
4366 if (l2arc_ndev == 0)
4370 next = l2arc_dev_last;
4372 /* loop around the list looking for a non-faulted vdev */
4374 next = list_head(l2arc_dev_list);
4376 next = list_next(l2arc_dev_list, next);
4378 next = list_head(l2arc_dev_list);
4381 /* if we have come back to the start, bail out */
4384 else if (next == first)
4387 } while (vdev_is_dead(next->l2ad_vdev));
4389 /* if we were unable to find any usable vdevs, return NULL */
4390 if (vdev_is_dead(next->l2ad_vdev))
4393 l2arc_dev_last = next;
4396 mutex_exit(&l2arc_dev_mtx);
4399 * Grab the config lock to prevent the 'next' device from being
4400 * removed while we are writing to it.
4403 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4404 mutex_exit(&spa_namespace_lock);
4410 * Free buffers that were tagged for destruction.
4413 l2arc_do_free_on_write()
4416 l2arc_data_free_t *df, *df_prev;
4418 mutex_enter(&l2arc_free_on_write_mtx);
4419 buflist = l2arc_free_on_write;
4421 for (df = list_tail(buflist); df; df = df_prev) {
4422 df_prev = list_prev(buflist, df);
4423 ASSERT(df->l2df_data != NULL);
4424 ASSERT(df->l2df_func != NULL);
4425 df->l2df_func(df->l2df_data, df->l2df_size);
4426 list_remove(buflist, df);
4427 kmem_free(df, sizeof (l2arc_data_free_t));
4430 mutex_exit(&l2arc_free_on_write_mtx);
4434 * A write to a cache device has completed. Update all headers to allow
4435 * reads from these buffers to begin.
4438 l2arc_write_done(zio_t *zio)
4440 l2arc_write_callback_t *cb;
4443 arc_buf_hdr_t *head, *ab, *ab_prev;
4444 l2arc_buf_hdr_t *abl2;
4445 kmutex_t *hash_lock;
4447 cb = zio->io_private;
4449 dev = cb->l2wcb_dev;
4450 ASSERT(dev != NULL);
4451 head = cb->l2wcb_head;
4452 ASSERT(head != NULL);
4453 buflist = dev->l2ad_buflist;
4454 ASSERT(buflist != NULL);
4455 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4456 l2arc_write_callback_t *, cb);
4458 if (zio->io_error != 0)
4459 ARCSTAT_BUMP(arcstat_l2_writes_error);
4461 mutex_enter(&l2arc_buflist_mtx);
4464 * All writes completed, or an error was hit.
4466 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4467 ab_prev = list_prev(buflist, ab);
4469 hash_lock = HDR_LOCK(ab);
4470 if (!mutex_tryenter(hash_lock)) {
4472 * This buffer misses out. It may be in a stage
4473 * of eviction. Its ARC_L2_WRITING flag will be
4474 * left set, denying reads to this buffer.
4476 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4480 if (zio->io_error != 0) {
4482 * Error - drop L2ARC entry.
4484 list_remove(buflist, ab);
4487 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4488 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4492 * Allow ARC to begin reads to this L2ARC entry.
4494 ab->b_flags &= ~ARC_L2_WRITING;
4496 mutex_exit(hash_lock);
4499 atomic_inc_64(&l2arc_writes_done);
4500 list_remove(buflist, head);
4501 kmem_cache_free(hdr_cache, head);
4502 mutex_exit(&l2arc_buflist_mtx);
4504 l2arc_do_free_on_write();
4506 kmem_free(cb, sizeof (l2arc_write_callback_t));
4510 * A read to a cache device completed. Validate buffer contents before
4511 * handing over to the regular ARC routines.
4514 l2arc_read_done(zio_t *zio)
4516 l2arc_read_callback_t *cb;
4519 kmutex_t *hash_lock;
4522 ASSERT(zio->io_vd != NULL);
4523 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4525 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4527 cb = zio->io_private;
4529 buf = cb->l2rcb_buf;
4530 ASSERT(buf != NULL);
4532 hash_lock = HDR_LOCK(buf->b_hdr);
4533 mutex_enter(hash_lock);
4535 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4538 * Check this survived the L2ARC journey.
4540 equal = arc_cksum_equal(buf);
4541 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4542 mutex_exit(hash_lock);
4543 zio->io_private = buf;
4544 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4545 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4548 mutex_exit(hash_lock);
4550 * Buffer didn't survive caching. Increment stats and
4551 * reissue to the original storage device.
4553 if (zio->io_error != 0) {
4554 ARCSTAT_BUMP(arcstat_l2_io_error);
4556 zio->io_error = EIO;
4559 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4562 * If there's no waiter, issue an async i/o to the primary
4563 * storage now. If there *is* a waiter, the caller must
4564 * issue the i/o in a context where it's OK to block.
4566 if (zio->io_waiter == NULL) {
4567 zio_t *pio = zio_unique_parent(zio);
4569 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4571 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4572 buf->b_data, zio->io_size, arc_read_done, buf,
4573 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4577 kmem_free(cb, sizeof (l2arc_read_callback_t));
4581 * This is the list priority from which the L2ARC will search for pages to
4582 * cache. This is used within loops (0..3) to cycle through lists in the
4583 * desired order. This order can have a significant effect on cache
4586 * Currently the metadata lists are hit first, MFU then MRU, followed by
4587 * the data lists. This function returns a locked list, and also returns
4591 l2arc_list_locked(int list_num, kmutex_t **lock)
4596 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS);
4598 if (list_num < ARC_BUFC_NUMMETADATALISTS) {
4600 list = &arc_mfu->arcs_lists[idx];
4601 *lock = ARCS_LOCK(arc_mfu, idx);
4602 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) {
4603 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4604 list = &arc_mru->arcs_lists[idx];
4605 *lock = ARCS_LOCK(arc_mru, idx);
4606 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 +
4607 ARC_BUFC_NUMDATALISTS)) {
4608 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4609 list = &arc_mfu->arcs_lists[idx];
4610 *lock = ARCS_LOCK(arc_mfu, idx);
4612 idx = list_num - ARC_BUFC_NUMLISTS;
4613 list = &arc_mru->arcs_lists[idx];
4614 *lock = ARCS_LOCK(arc_mru, idx);
4617 ASSERT(!(MUTEX_HELD(*lock)));
4623 * Evict buffers from the device write hand to the distance specified in
4624 * bytes. This distance may span populated buffers, it may span nothing.
4625 * This is clearing a region on the L2ARC device ready for writing.
4626 * If the 'all' boolean is set, every buffer is evicted.
4629 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4632 l2arc_buf_hdr_t *abl2;
4633 arc_buf_hdr_t *ab, *ab_prev;
4634 kmutex_t *hash_lock;
4637 buflist = dev->l2ad_buflist;
4639 if (buflist == NULL)
4642 if (!all && dev->l2ad_first) {
4644 * This is the first sweep through the device. There is
4650 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4652 * When nearing the end of the device, evict to the end
4653 * before the device write hand jumps to the start.
4655 taddr = dev->l2ad_end;
4657 taddr = dev->l2ad_hand + distance;
4659 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4660 uint64_t, taddr, boolean_t, all);
4663 mutex_enter(&l2arc_buflist_mtx);
4664 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4665 ab_prev = list_prev(buflist, ab);
4667 hash_lock = HDR_LOCK(ab);
4668 if (!mutex_tryenter(hash_lock)) {
4670 * Missed the hash lock. Retry.
4672 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4673 mutex_exit(&l2arc_buflist_mtx);
4674 mutex_enter(hash_lock);
4675 mutex_exit(hash_lock);
4679 if (HDR_L2_WRITE_HEAD(ab)) {
4681 * We hit a write head node. Leave it for
4682 * l2arc_write_done().
4684 list_remove(buflist, ab);
4685 mutex_exit(hash_lock);
4689 if (!all && ab->b_l2hdr != NULL &&
4690 (ab->b_l2hdr->b_daddr > taddr ||
4691 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4693 * We've evicted to the target address,
4694 * or the end of the device.
4696 mutex_exit(hash_lock);
4700 if (HDR_FREE_IN_PROGRESS(ab)) {
4702 * Already on the path to destruction.
4704 mutex_exit(hash_lock);
4708 if (ab->b_state == arc_l2c_only) {
4709 ASSERT(!HDR_L2_READING(ab));
4711 * This doesn't exist in the ARC. Destroy.
4712 * arc_hdr_destroy() will call list_remove()
4713 * and decrement arcstat_l2_size.
4715 arc_change_state(arc_anon, ab, hash_lock);
4716 arc_hdr_destroy(ab);
4719 * Invalidate issued or about to be issued
4720 * reads, since we may be about to write
4721 * over this location.
4723 if (HDR_L2_READING(ab)) {
4724 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4725 ab->b_flags |= ARC_L2_EVICTED;
4729 * Tell ARC this no longer exists in L2ARC.
4731 if (ab->b_l2hdr != NULL) {
4734 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4735 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4737 list_remove(buflist, ab);
4740 * This may have been leftover after a
4743 ab->b_flags &= ~ARC_L2_WRITING;
4745 mutex_exit(hash_lock);
4747 mutex_exit(&l2arc_buflist_mtx);
4749 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4750 dev->l2ad_evict = taddr;
4754 * Find and write ARC buffers to the L2ARC device.
4756 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4757 * for reading until they have completed writing.
4760 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4762 arc_buf_hdr_t *ab, *ab_prev, *head;
4763 l2arc_buf_hdr_t *hdrl2;
4765 uint64_t passed_sz, write_sz, buf_sz, headroom;
4767 kmutex_t *hash_lock, *list_lock;
4768 boolean_t have_lock, full;
4769 l2arc_write_callback_t *cb;
4771 uint64_t guid = spa_load_guid(spa);
4774 ASSERT(dev->l2ad_vdev != NULL);
4779 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4780 head->b_flags |= ARC_L2_WRITE_HEAD;
4782 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
4784 * Copy buffers for L2ARC writing.
4786 mutex_enter(&l2arc_buflist_mtx);
4787 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) {
4788 list = l2arc_list_locked(try, &list_lock);
4790 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
4793 * L2ARC fast warmup.
4795 * Until the ARC is warm and starts to evict, read from the
4796 * head of the ARC lists rather than the tail.
4798 headroom = target_sz * l2arc_headroom;
4799 if (arc_warm == B_FALSE)
4800 ab = list_head(list);
4802 ab = list_tail(list);
4804 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
4806 for (; ab; ab = ab_prev) {
4807 if (arc_warm == B_FALSE)
4808 ab_prev = list_next(list, ab);
4810 ab_prev = list_prev(list, ab);
4811 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size);
4813 hash_lock = HDR_LOCK(ab);
4814 have_lock = MUTEX_HELD(hash_lock);
4815 if (!have_lock && !mutex_tryenter(hash_lock)) {
4816 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
4818 * Skip this buffer rather than waiting.
4823 passed_sz += ab->b_size;
4824 if (passed_sz > headroom) {
4828 mutex_exit(hash_lock);
4829 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
4833 if (!l2arc_write_eligible(guid, ab)) {
4834 mutex_exit(hash_lock);
4838 if ((write_sz + ab->b_size) > target_sz) {
4840 mutex_exit(hash_lock);
4841 ARCSTAT_BUMP(arcstat_l2_write_full);
4847 * Insert a dummy header on the buflist so
4848 * l2arc_write_done() can find where the
4849 * write buffers begin without searching.
4851 list_insert_head(dev->l2ad_buflist, head);
4854 sizeof (l2arc_write_callback_t), KM_SLEEP);
4855 cb->l2wcb_dev = dev;
4856 cb->l2wcb_head = head;
4857 pio = zio_root(spa, l2arc_write_done, cb,
4859 ARCSTAT_BUMP(arcstat_l2_write_pios);
4863 * Create and add a new L2ARC header.
4865 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4867 hdrl2->b_daddr = dev->l2ad_hand;
4869 ab->b_flags |= ARC_L2_WRITING;
4870 ab->b_l2hdr = hdrl2;
4871 list_insert_head(dev->l2ad_buflist, ab);
4872 buf_data = ab->b_buf->b_data;
4873 buf_sz = ab->b_size;
4876 * Compute and store the buffer cksum before
4877 * writing. On debug the cksum is verified first.
4879 arc_cksum_verify(ab->b_buf);
4880 arc_cksum_compute(ab->b_buf, B_TRUE);
4882 mutex_exit(hash_lock);
4884 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4885 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4886 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4887 ZIO_FLAG_CANFAIL, B_FALSE);
4889 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4891 (void) zio_nowait(wzio);
4894 * Keep the clock hand suitably device-aligned.
4896 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4899 dev->l2ad_hand += buf_sz;
4902 mutex_exit(list_lock);
4907 mutex_exit(&l2arc_buflist_mtx);
4911 kmem_cache_free(hdr_cache, head);
4915 ASSERT3U(write_sz, <=, target_sz);
4916 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4917 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4918 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4919 vdev_space_update(dev->l2ad_vdev, write_sz, 0, 0);
4922 * Bump device hand to the device start if it is approaching the end.
4923 * l2arc_evict() will already have evicted ahead for this case.
4925 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4926 vdev_space_update(dev->l2ad_vdev,
4927 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4928 dev->l2ad_hand = dev->l2ad_start;
4929 dev->l2ad_evict = dev->l2ad_start;
4930 dev->l2ad_first = B_FALSE;
4933 dev->l2ad_writing = B_TRUE;
4934 (void) zio_wait(pio);
4935 dev->l2ad_writing = B_FALSE;
4941 * This thread feeds the L2ARC at regular intervals. This is the beating
4942 * heart of the L2ARC.
4945 l2arc_feed_thread(void *dummy __unused)
4950 uint64_t size, wrote;
4951 clock_t begin, next = ddi_get_lbolt();
4953 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4955 mutex_enter(&l2arc_feed_thr_lock);
4957 while (l2arc_thread_exit == 0) {
4958 CALLB_CPR_SAFE_BEGIN(&cpr);
4959 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4960 next - ddi_get_lbolt());
4961 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4962 next = ddi_get_lbolt() + hz;
4965 * Quick check for L2ARC devices.
4967 mutex_enter(&l2arc_dev_mtx);
4968 if (l2arc_ndev == 0) {
4969 mutex_exit(&l2arc_dev_mtx);
4972 mutex_exit(&l2arc_dev_mtx);
4973 begin = ddi_get_lbolt();
4976 * This selects the next l2arc device to write to, and in
4977 * doing so the next spa to feed from: dev->l2ad_spa. This
4978 * will return NULL if there are now no l2arc devices or if
4979 * they are all faulted.
4981 * If a device is returned, its spa's config lock is also
4982 * held to prevent device removal. l2arc_dev_get_next()
4983 * will grab and release l2arc_dev_mtx.
4985 if ((dev = l2arc_dev_get_next()) == NULL)
4988 spa = dev->l2ad_spa;
4989 ASSERT(spa != NULL);
4992 * If the pool is read-only then force the feed thread to
4993 * sleep a little longer.
4995 if (!spa_writeable(spa)) {
4996 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4997 spa_config_exit(spa, SCL_L2ARC, dev);
5002 * Avoid contributing to memory pressure.
5004 if (arc_reclaim_needed()) {
5005 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
5006 spa_config_exit(spa, SCL_L2ARC, dev);
5010 ARCSTAT_BUMP(arcstat_l2_feeds);
5012 size = l2arc_write_size(dev);
5015 * Evict L2ARC buffers that will be overwritten.
5017 l2arc_evict(dev, size, B_FALSE);
5020 * Write ARC buffers.
5022 wrote = l2arc_write_buffers(spa, dev, size);
5025 * Calculate interval between writes.
5027 next = l2arc_write_interval(begin, size, wrote);
5028 spa_config_exit(spa, SCL_L2ARC, dev);
5031 l2arc_thread_exit = 0;
5032 cv_broadcast(&l2arc_feed_thr_cv);
5033 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
5038 l2arc_vdev_present(vdev_t *vd)
5042 mutex_enter(&l2arc_dev_mtx);
5043 for (dev = list_head(l2arc_dev_list); dev != NULL;
5044 dev = list_next(l2arc_dev_list, dev)) {
5045 if (dev->l2ad_vdev == vd)
5048 mutex_exit(&l2arc_dev_mtx);
5050 return (dev != NULL);
5054 * Add a vdev for use by the L2ARC. By this point the spa has already
5055 * validated the vdev and opened it.
5058 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5060 l2arc_dev_t *adddev;
5062 ASSERT(!l2arc_vdev_present(vd));
5065 * Create a new l2arc device entry.
5067 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5068 adddev->l2ad_spa = spa;
5069 adddev->l2ad_vdev = vd;
5070 adddev->l2ad_write = l2arc_write_max;
5071 adddev->l2ad_boost = l2arc_write_boost;
5072 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5073 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5074 adddev->l2ad_hand = adddev->l2ad_start;
5075 adddev->l2ad_evict = adddev->l2ad_start;
5076 adddev->l2ad_first = B_TRUE;
5077 adddev->l2ad_writing = B_FALSE;
5078 ASSERT3U(adddev->l2ad_write, >, 0);
5081 * This is a list of all ARC buffers that are still valid on the
5084 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5085 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5086 offsetof(arc_buf_hdr_t, b_l2node));
5088 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5091 * Add device to global list
5093 mutex_enter(&l2arc_dev_mtx);
5094 list_insert_head(l2arc_dev_list, adddev);
5095 atomic_inc_64(&l2arc_ndev);
5096 mutex_exit(&l2arc_dev_mtx);
5100 * Remove a vdev from the L2ARC.
5103 l2arc_remove_vdev(vdev_t *vd)
5105 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5108 * Find the device by vdev
5110 mutex_enter(&l2arc_dev_mtx);
5111 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5112 nextdev = list_next(l2arc_dev_list, dev);
5113 if (vd == dev->l2ad_vdev) {
5118 ASSERT(remdev != NULL);
5121 * Remove device from global list
5123 list_remove(l2arc_dev_list, remdev);
5124 l2arc_dev_last = NULL; /* may have been invalidated */
5125 atomic_dec_64(&l2arc_ndev);
5126 mutex_exit(&l2arc_dev_mtx);
5129 * Clear all buflists and ARC references. L2ARC device flush.
5131 l2arc_evict(remdev, 0, B_TRUE);
5132 list_destroy(remdev->l2ad_buflist);
5133 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5134 kmem_free(remdev, sizeof (l2arc_dev_t));
5140 l2arc_thread_exit = 0;
5142 l2arc_writes_sent = 0;
5143 l2arc_writes_done = 0;
5145 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5146 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5147 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5148 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5149 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5151 l2arc_dev_list = &L2ARC_dev_list;
5152 l2arc_free_on_write = &L2ARC_free_on_write;
5153 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5154 offsetof(l2arc_dev_t, l2ad_node));
5155 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5156 offsetof(l2arc_data_free_t, l2df_list_node));
5163 * This is called from dmu_fini(), which is called from spa_fini();
5164 * Because of this, we can assume that all l2arc devices have
5165 * already been removed when the pools themselves were removed.
5168 l2arc_do_free_on_write();
5170 mutex_destroy(&l2arc_feed_thr_lock);
5171 cv_destroy(&l2arc_feed_thr_cv);
5172 mutex_destroy(&l2arc_dev_mtx);
5173 mutex_destroy(&l2arc_buflist_mtx);
5174 mutex_destroy(&l2arc_free_on_write_mtx);
5176 list_destroy(l2arc_dev_list);
5177 list_destroy(l2arc_free_on_write);
5183 if (!(spa_mode_global & FWRITE))
5186 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5187 TS_RUN, minclsyspri);
5193 if (!(spa_mode_global & FWRITE))
5196 mutex_enter(&l2arc_feed_thr_lock);
5197 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5198 l2arc_thread_exit = 1;
5199 while (l2arc_thread_exit != 0)
5200 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5201 mutex_exit(&l2arc_feed_thr_lock);