4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2014 Nexenta Systems, Inc. All rights reserved.
30 * DVA-based Adjustable Replacement Cache
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal arc algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
96 * Each arc state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an arc list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
102 * Arc buffers may have an associated eviction callback function.
103 * This function will be invoked prior to removing the buffer (e.g.
104 * in arc_do_user_evicts()). Note however that the data associated
105 * with the buffer may be evicted prior to the callback. The callback
106 * must be made with *no locks held* (to prevent deadlock). Additionally,
107 * the users of callbacks must ensure that their private data is
108 * protected from simultaneous callbacks from arc_clear_callback()
109 * and arc_do_user_evicts().
111 * Note that the majority of the performance stats are manipulated
112 * with atomic operations.
114 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
116 * - L2ARC buflist creation
117 * - L2ARC buflist eviction
118 * - L2ARC write completion, which walks L2ARC buflists
119 * - ARC header destruction, as it removes from L2ARC buflists
120 * - ARC header release, as it removes from L2ARC buflists
125 #include <sys/zio_compress.h>
126 #include <sys/zfs_context.h>
128 #include <sys/refcount.h>
129 #include <sys/vdev.h>
130 #include <sys/vdev_impl.h>
131 #include <sys/dsl_pool.h>
133 #include <sys/vmsystm.h>
135 #include <sys/fs/swapnode.h>
136 #include <sys/dnlc.h>
138 #include <sys/callb.h>
139 #include <sys/kstat.h>
140 #include <zfs_fletcher.h>
143 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
144 boolean_t arc_watch = B_FALSE;
148 static kmutex_t arc_reclaim_thr_lock;
149 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
150 static uint8_t arc_thread_exit;
152 #define ARC_REDUCE_DNLC_PERCENT 3
153 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
155 typedef enum arc_reclaim_strategy {
156 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
157 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
158 } arc_reclaim_strategy_t;
161 * The number of iterations through arc_evict_*() before we
162 * drop & reacquire the lock.
164 int arc_evict_iterations = 100;
166 /* number of seconds before growing cache again */
167 static int arc_grow_retry = 60;
169 /* shift of arc_c for calculating both min and max arc_p */
170 static int arc_p_min_shift = 4;
172 /* log2(fraction of arc to reclaim) */
173 static int arc_shrink_shift = 5;
176 * minimum lifespan of a prefetch block in clock ticks
177 * (initialized in arc_init())
179 static int arc_min_prefetch_lifespan;
182 * If this percent of memory is free, don't throttle.
184 int arc_lotsfree_percent = 10;
189 * The arc has filled available memory and has now warmed up.
191 static boolean_t arc_warm;
194 * These tunables are for performance analysis.
196 uint64_t zfs_arc_max;
197 uint64_t zfs_arc_min;
198 uint64_t zfs_arc_meta_limit = 0;
199 uint64_t zfs_arc_meta_min = 0;
200 int zfs_arc_grow_retry = 0;
201 int zfs_arc_shrink_shift = 0;
202 int zfs_arc_p_min_shift = 0;
203 int zfs_disable_dup_eviction = 0;
204 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
207 * Note that buffers can be in one of 6 states:
208 * ARC_anon - anonymous (discussed below)
209 * ARC_mru - recently used, currently cached
210 * ARC_mru_ghost - recentely used, no longer in cache
211 * ARC_mfu - frequently used, currently cached
212 * ARC_mfu_ghost - frequently used, no longer in cache
213 * ARC_l2c_only - exists in L2ARC but not other states
214 * When there are no active references to the buffer, they are
215 * are linked onto a list in one of these arc states. These are
216 * the only buffers that can be evicted or deleted. Within each
217 * state there are multiple lists, one for meta-data and one for
218 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
219 * etc.) is tracked separately so that it can be managed more
220 * explicitly: favored over data, limited explicitly.
222 * Anonymous buffers are buffers that are not associated with
223 * a DVA. These are buffers that hold dirty block copies
224 * before they are written to stable storage. By definition,
225 * they are "ref'd" and are considered part of arc_mru
226 * that cannot be freed. Generally, they will aquire a DVA
227 * as they are written and migrate onto the arc_mru list.
229 * The ARC_l2c_only state is for buffers that are in the second
230 * level ARC but no longer in any of the ARC_m* lists. The second
231 * level ARC itself may also contain buffers that are in any of
232 * the ARC_m* states - meaning that a buffer can exist in two
233 * places. The reason for the ARC_l2c_only state is to keep the
234 * buffer header in the hash table, so that reads that hit the
235 * second level ARC benefit from these fast lookups.
238 typedef struct arc_state {
239 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
240 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
241 uint64_t arcs_size; /* total amount of data in this state */
246 static arc_state_t ARC_anon;
247 static arc_state_t ARC_mru;
248 static arc_state_t ARC_mru_ghost;
249 static arc_state_t ARC_mfu;
250 static arc_state_t ARC_mfu_ghost;
251 static arc_state_t ARC_l2c_only;
253 typedef struct arc_stats {
254 kstat_named_t arcstat_hits;
255 kstat_named_t arcstat_misses;
256 kstat_named_t arcstat_demand_data_hits;
257 kstat_named_t arcstat_demand_data_misses;
258 kstat_named_t arcstat_demand_metadata_hits;
259 kstat_named_t arcstat_demand_metadata_misses;
260 kstat_named_t arcstat_prefetch_data_hits;
261 kstat_named_t arcstat_prefetch_data_misses;
262 kstat_named_t arcstat_prefetch_metadata_hits;
263 kstat_named_t arcstat_prefetch_metadata_misses;
264 kstat_named_t arcstat_mru_hits;
265 kstat_named_t arcstat_mru_ghost_hits;
266 kstat_named_t arcstat_mfu_hits;
267 kstat_named_t arcstat_mfu_ghost_hits;
268 kstat_named_t arcstat_deleted;
269 kstat_named_t arcstat_recycle_miss;
271 * Number of buffers that could not be evicted because the hash lock
272 * was held by another thread. The lock may not necessarily be held
273 * by something using the same buffer, since hash locks are shared
274 * by multiple buffers.
276 kstat_named_t arcstat_mutex_miss;
278 * Number of buffers skipped because they have I/O in progress, are
279 * indrect prefetch buffers that have not lived long enough, or are
280 * not from the spa we're trying to evict from.
282 kstat_named_t arcstat_evict_skip;
283 kstat_named_t arcstat_evict_l2_cached;
284 kstat_named_t arcstat_evict_l2_eligible;
285 kstat_named_t arcstat_evict_l2_ineligible;
286 kstat_named_t arcstat_hash_elements;
287 kstat_named_t arcstat_hash_elements_max;
288 kstat_named_t arcstat_hash_collisions;
289 kstat_named_t arcstat_hash_chains;
290 kstat_named_t arcstat_hash_chain_max;
291 kstat_named_t arcstat_p;
292 kstat_named_t arcstat_c;
293 kstat_named_t arcstat_c_min;
294 kstat_named_t arcstat_c_max;
295 kstat_named_t arcstat_size;
296 kstat_named_t arcstat_hdr_size;
297 kstat_named_t arcstat_data_size;
298 kstat_named_t arcstat_other_size;
299 kstat_named_t arcstat_l2_hits;
300 kstat_named_t arcstat_l2_misses;
301 kstat_named_t arcstat_l2_feeds;
302 kstat_named_t arcstat_l2_rw_clash;
303 kstat_named_t arcstat_l2_read_bytes;
304 kstat_named_t arcstat_l2_write_bytes;
305 kstat_named_t arcstat_l2_writes_sent;
306 kstat_named_t arcstat_l2_writes_done;
307 kstat_named_t arcstat_l2_writes_error;
308 kstat_named_t arcstat_l2_writes_hdr_miss;
309 kstat_named_t arcstat_l2_evict_lock_retry;
310 kstat_named_t arcstat_l2_evict_reading;
311 kstat_named_t arcstat_l2_free_on_write;
312 kstat_named_t arcstat_l2_abort_lowmem;
313 kstat_named_t arcstat_l2_cksum_bad;
314 kstat_named_t arcstat_l2_io_error;
315 kstat_named_t arcstat_l2_size;
316 kstat_named_t arcstat_l2_asize;
317 kstat_named_t arcstat_l2_hdr_size;
318 kstat_named_t arcstat_l2_compress_successes;
319 kstat_named_t arcstat_l2_compress_zeros;
320 kstat_named_t arcstat_l2_compress_failures;
321 kstat_named_t arcstat_memory_throttle_count;
322 kstat_named_t arcstat_duplicate_buffers;
323 kstat_named_t arcstat_duplicate_buffers_size;
324 kstat_named_t arcstat_duplicate_reads;
325 kstat_named_t arcstat_meta_used;
326 kstat_named_t arcstat_meta_limit;
327 kstat_named_t arcstat_meta_max;
328 kstat_named_t arcstat_meta_min;
331 static arc_stats_t arc_stats = {
332 { "hits", KSTAT_DATA_UINT64 },
333 { "misses", KSTAT_DATA_UINT64 },
334 { "demand_data_hits", KSTAT_DATA_UINT64 },
335 { "demand_data_misses", KSTAT_DATA_UINT64 },
336 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
337 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
338 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
339 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
340 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
341 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
342 { "mru_hits", KSTAT_DATA_UINT64 },
343 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
344 { "mfu_hits", KSTAT_DATA_UINT64 },
345 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
346 { "deleted", KSTAT_DATA_UINT64 },
347 { "recycle_miss", KSTAT_DATA_UINT64 },
348 { "mutex_miss", KSTAT_DATA_UINT64 },
349 { "evict_skip", KSTAT_DATA_UINT64 },
350 { "evict_l2_cached", KSTAT_DATA_UINT64 },
351 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
352 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
353 { "hash_elements", KSTAT_DATA_UINT64 },
354 { "hash_elements_max", KSTAT_DATA_UINT64 },
355 { "hash_collisions", KSTAT_DATA_UINT64 },
356 { "hash_chains", KSTAT_DATA_UINT64 },
357 { "hash_chain_max", KSTAT_DATA_UINT64 },
358 { "p", KSTAT_DATA_UINT64 },
359 { "c", KSTAT_DATA_UINT64 },
360 { "c_min", KSTAT_DATA_UINT64 },
361 { "c_max", KSTAT_DATA_UINT64 },
362 { "size", KSTAT_DATA_UINT64 },
363 { "hdr_size", KSTAT_DATA_UINT64 },
364 { "data_size", KSTAT_DATA_UINT64 },
365 { "other_size", KSTAT_DATA_UINT64 },
366 { "l2_hits", KSTAT_DATA_UINT64 },
367 { "l2_misses", KSTAT_DATA_UINT64 },
368 { "l2_feeds", KSTAT_DATA_UINT64 },
369 { "l2_rw_clash", KSTAT_DATA_UINT64 },
370 { "l2_read_bytes", KSTAT_DATA_UINT64 },
371 { "l2_write_bytes", KSTAT_DATA_UINT64 },
372 { "l2_writes_sent", KSTAT_DATA_UINT64 },
373 { "l2_writes_done", KSTAT_DATA_UINT64 },
374 { "l2_writes_error", KSTAT_DATA_UINT64 },
375 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
376 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
377 { "l2_evict_reading", KSTAT_DATA_UINT64 },
378 { "l2_free_on_write", KSTAT_DATA_UINT64 },
379 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
380 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
381 { "l2_io_error", KSTAT_DATA_UINT64 },
382 { "l2_size", KSTAT_DATA_UINT64 },
383 { "l2_asize", KSTAT_DATA_UINT64 },
384 { "l2_hdr_size", KSTAT_DATA_UINT64 },
385 { "l2_compress_successes", KSTAT_DATA_UINT64 },
386 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
387 { "l2_compress_failures", KSTAT_DATA_UINT64 },
388 { "memory_throttle_count", KSTAT_DATA_UINT64 },
389 { "duplicate_buffers", KSTAT_DATA_UINT64 },
390 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
391 { "duplicate_reads", KSTAT_DATA_UINT64 },
392 { "arc_meta_used", KSTAT_DATA_UINT64 },
393 { "arc_meta_limit", KSTAT_DATA_UINT64 },
394 { "arc_meta_max", KSTAT_DATA_UINT64 },
395 { "arc_meta_min", KSTAT_DATA_UINT64 }
398 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
400 #define ARCSTAT_INCR(stat, val) \
401 atomic_add_64(&arc_stats.stat.value.ui64, (val))
403 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
404 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
406 #define ARCSTAT_MAX(stat, val) { \
408 while ((val) > (m = arc_stats.stat.value.ui64) && \
409 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
413 #define ARCSTAT_MAXSTAT(stat) \
414 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
417 * We define a macro to allow ARC hits/misses to be easily broken down by
418 * two separate conditions, giving a total of four different subtypes for
419 * each of hits and misses (so eight statistics total).
421 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
424 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
426 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
430 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
432 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
437 static arc_state_t *arc_anon;
438 static arc_state_t *arc_mru;
439 static arc_state_t *arc_mru_ghost;
440 static arc_state_t *arc_mfu;
441 static arc_state_t *arc_mfu_ghost;
442 static arc_state_t *arc_l2c_only;
445 * There are several ARC variables that are critical to export as kstats --
446 * but we don't want to have to grovel around in the kstat whenever we wish to
447 * manipulate them. For these variables, we therefore define them to be in
448 * terms of the statistic variable. This assures that we are not introducing
449 * the possibility of inconsistency by having shadow copies of the variables,
450 * while still allowing the code to be readable.
452 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
453 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
454 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
455 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
456 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
457 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
458 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
459 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
460 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
462 #define L2ARC_IS_VALID_COMPRESS(_c_) \
463 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
465 static int arc_no_grow; /* Don't try to grow cache size */
466 static uint64_t arc_tempreserve;
467 static uint64_t arc_loaned_bytes;
469 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
471 typedef struct arc_callback arc_callback_t;
473 struct arc_callback {
475 arc_done_func_t *acb_done;
477 zio_t *acb_zio_dummy;
478 arc_callback_t *acb_next;
481 typedef struct arc_write_callback arc_write_callback_t;
483 struct arc_write_callback {
485 arc_done_func_t *awcb_ready;
486 arc_done_func_t *awcb_physdone;
487 arc_done_func_t *awcb_done;
492 /* protected by hash lock */
497 kmutex_t b_freeze_lock;
498 zio_cksum_t *b_freeze_cksum;
501 arc_buf_hdr_t *b_hash_next;
506 arc_callback_t *b_acb;
510 arc_buf_contents_t b_type;
514 /* protected by arc state mutex */
515 arc_state_t *b_state;
516 list_node_t b_arc_node;
518 /* updated atomically */
519 clock_t b_arc_access;
521 /* self protecting */
524 l2arc_buf_hdr_t *b_l2hdr;
525 list_node_t b_l2node;
528 static arc_buf_t *arc_eviction_list;
529 static kmutex_t arc_eviction_mtx;
530 static arc_buf_hdr_t arc_eviction_hdr;
532 #define GHOST_STATE(state) \
533 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
534 (state) == arc_l2c_only)
536 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
537 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
538 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
539 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
540 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
541 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
542 #define HDR_FREE_IN_PROGRESS(hdr) \
543 ((hdr)->b_flags & ARC_FLAG_FREE_IN_PROGRESS)
544 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
545 #define HDR_L2_READING(hdr) \
546 ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS && \
547 (hdr)->b_l2hdr != NULL)
548 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
549 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
550 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
556 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
557 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
560 * Hash table routines
563 #define HT_LOCK_PAD 64
568 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
572 #define BUF_LOCKS 256
573 typedef struct buf_hash_table {
575 arc_buf_hdr_t **ht_table;
576 struct ht_lock ht_locks[BUF_LOCKS];
579 static buf_hash_table_t buf_hash_table;
581 #define BUF_HASH_INDEX(spa, dva, birth) \
582 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
583 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
584 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
585 #define HDR_LOCK(hdr) \
586 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
588 uint64_t zfs_crc64_table[256];
594 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
595 #define L2ARC_HEADROOM 2 /* num of writes */
597 * If we discover during ARC scan any buffers to be compressed, we boost
598 * our headroom for the next scanning cycle by this percentage multiple.
600 #define L2ARC_HEADROOM_BOOST 200
601 #define L2ARC_FEED_SECS 1 /* caching interval secs */
602 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
604 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
605 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
607 /* L2ARC Performance Tunables */
608 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
609 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
610 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
611 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
612 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
613 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
614 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
615 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
616 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
621 typedef struct l2arc_dev {
622 vdev_t *l2ad_vdev; /* vdev */
623 spa_t *l2ad_spa; /* spa */
624 uint64_t l2ad_hand; /* next write location */
625 uint64_t l2ad_start; /* first addr on device */
626 uint64_t l2ad_end; /* last addr on device */
627 uint64_t l2ad_evict; /* last addr eviction reached */
628 boolean_t l2ad_first; /* first sweep through */
629 boolean_t l2ad_writing; /* currently writing */
630 list_t *l2ad_buflist; /* buffer list */
631 list_node_t l2ad_node; /* device list node */
634 static list_t L2ARC_dev_list; /* device list */
635 static list_t *l2arc_dev_list; /* device list pointer */
636 static kmutex_t l2arc_dev_mtx; /* device list mutex */
637 static l2arc_dev_t *l2arc_dev_last; /* last device used */
638 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
639 static list_t L2ARC_free_on_write; /* free after write buf list */
640 static list_t *l2arc_free_on_write; /* free after write list ptr */
641 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
642 static uint64_t l2arc_ndev; /* number of devices */
644 typedef struct l2arc_read_callback {
645 arc_buf_t *l2rcb_buf; /* read buffer */
646 spa_t *l2rcb_spa; /* spa */
647 blkptr_t l2rcb_bp; /* original blkptr */
648 zbookmark_phys_t l2rcb_zb; /* original bookmark */
649 int l2rcb_flags; /* original flags */
650 enum zio_compress l2rcb_compress; /* applied compress */
651 } l2arc_read_callback_t;
653 typedef struct l2arc_write_callback {
654 l2arc_dev_t *l2wcb_dev; /* device info */
655 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
656 } l2arc_write_callback_t;
658 struct l2arc_buf_hdr {
659 /* protected by arc_buf_hdr mutex */
660 l2arc_dev_t *b_dev; /* L2ARC device */
661 uint64_t b_daddr; /* disk address, offset byte */
662 /* compression applied to buffer data */
663 enum zio_compress b_compress;
664 /* real alloc'd buffer size depending on b_compress applied */
666 /* temporary buffer holder for in-flight compressed data */
670 typedef struct l2arc_data_free {
671 /* protected by l2arc_free_on_write_mtx */
674 void (*l2df_func)(void *, size_t);
675 list_node_t l2df_list_node;
678 static kmutex_t l2arc_feed_thr_lock;
679 static kcondvar_t l2arc_feed_thr_cv;
680 static uint8_t l2arc_thread_exit;
682 static void arc_get_data_buf(arc_buf_t *);
683 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
684 static int arc_evict_needed(arc_buf_contents_t);
685 static void arc_evict_ghost(arc_state_t *, uint64_t, int64_t);
686 static void arc_buf_watch(arc_buf_t *);
688 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
689 static void l2arc_read_done(zio_t *);
690 static void l2arc_hdr_stat_add(void);
691 static void l2arc_hdr_stat_remove(void);
693 static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *);
694 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress);
695 static void l2arc_release_cdata_buf(arc_buf_hdr_t *);
698 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
700 uint8_t *vdva = (uint8_t *)dva;
701 uint64_t crc = -1ULL;
704 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
706 for (i = 0; i < sizeof (dva_t); i++)
707 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
709 crc ^= (spa>>8) ^ birth;
714 #define BUF_EMPTY(buf) \
715 ((buf)->b_dva.dva_word[0] == 0 && \
716 (buf)->b_dva.dva_word[1] == 0 && \
717 (buf)->b_cksum0 == 0)
719 #define BUF_EQUAL(spa, dva, birth, buf) \
720 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
721 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
722 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
725 buf_discard_identity(arc_buf_hdr_t *hdr)
727 hdr->b_dva.dva_word[0] = 0;
728 hdr->b_dva.dva_word[1] = 0;
733 static arc_buf_hdr_t *
734 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
736 const dva_t *dva = BP_IDENTITY(bp);
737 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
738 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
739 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
742 mutex_enter(hash_lock);
743 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
744 hdr = hdr->b_hash_next) {
745 if (BUF_EQUAL(spa, dva, birth, hdr)) {
750 mutex_exit(hash_lock);
756 * Insert an entry into the hash table. If there is already an element
757 * equal to elem in the hash table, then the already existing element
758 * will be returned and the new element will not be inserted.
759 * Otherwise returns NULL.
761 static arc_buf_hdr_t *
762 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
764 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
765 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
769 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
770 ASSERT(hdr->b_birth != 0);
771 ASSERT(!HDR_IN_HASH_TABLE(hdr));
773 mutex_enter(hash_lock);
774 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
775 fhdr = fhdr->b_hash_next, i++) {
776 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
780 hdr->b_hash_next = buf_hash_table.ht_table[idx];
781 buf_hash_table.ht_table[idx] = hdr;
782 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
784 /* collect some hash table performance data */
786 ARCSTAT_BUMP(arcstat_hash_collisions);
788 ARCSTAT_BUMP(arcstat_hash_chains);
790 ARCSTAT_MAX(arcstat_hash_chain_max, i);
793 ARCSTAT_BUMP(arcstat_hash_elements);
794 ARCSTAT_MAXSTAT(arcstat_hash_elements);
800 buf_hash_remove(arc_buf_hdr_t *hdr)
802 arc_buf_hdr_t *fhdr, **hdrp;
803 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
805 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
806 ASSERT(HDR_IN_HASH_TABLE(hdr));
808 hdrp = &buf_hash_table.ht_table[idx];
809 while ((fhdr = *hdrp) != hdr) {
810 ASSERT(fhdr != NULL);
811 hdrp = &fhdr->b_hash_next;
813 *hdrp = hdr->b_hash_next;
814 hdr->b_hash_next = NULL;
815 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE;
817 /* collect some hash table performance data */
818 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
820 if (buf_hash_table.ht_table[idx] &&
821 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
822 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
826 * Global data structures and functions for the buf kmem cache.
828 static kmem_cache_t *hdr_cache;
829 static kmem_cache_t *buf_cache;
836 kmem_free(buf_hash_table.ht_table,
837 (buf_hash_table.ht_mask + 1) * sizeof (void *));
838 for (i = 0; i < BUF_LOCKS; i++)
839 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
840 kmem_cache_destroy(hdr_cache);
841 kmem_cache_destroy(buf_cache);
845 * Constructor callback - called when the cache is empty
846 * and a new buf is requested.
850 hdr_cons(void *vbuf, void *unused, int kmflag)
852 arc_buf_hdr_t *hdr = vbuf;
854 bzero(hdr, sizeof (arc_buf_hdr_t));
855 refcount_create(&hdr->b_refcnt);
856 cv_init(&hdr->b_cv, NULL, CV_DEFAULT, NULL);
857 mutex_init(&hdr->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
858 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
865 buf_cons(void *vbuf, void *unused, int kmflag)
867 arc_buf_t *buf = vbuf;
869 bzero(buf, sizeof (arc_buf_t));
870 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
871 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
877 * Destructor callback - called when a cached buf is
878 * no longer required.
882 hdr_dest(void *vbuf, void *unused)
884 arc_buf_hdr_t *hdr = vbuf;
886 ASSERT(BUF_EMPTY(hdr));
887 refcount_destroy(&hdr->b_refcnt);
888 cv_destroy(&hdr->b_cv);
889 mutex_destroy(&hdr->b_freeze_lock);
890 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
895 buf_dest(void *vbuf, void *unused)
897 arc_buf_t *buf = vbuf;
899 mutex_destroy(&buf->b_evict_lock);
900 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
904 * Reclaim callback -- invoked when memory is low.
908 hdr_recl(void *unused)
910 dprintf("hdr_recl called\n");
912 * umem calls the reclaim func when we destroy the buf cache,
913 * which is after we do arc_fini().
916 cv_signal(&arc_reclaim_thr_cv);
923 uint64_t hsize = 1ULL << 12;
927 * The hash table is big enough to fill all of physical memory
928 * with an average block size of zfs_arc_average_blocksize (default 8K).
929 * By default, the table will take up
930 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
932 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
935 buf_hash_table.ht_mask = hsize - 1;
936 buf_hash_table.ht_table =
937 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
938 if (buf_hash_table.ht_table == NULL) {
939 ASSERT(hsize > (1ULL << 8));
944 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
945 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
946 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
947 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
949 for (i = 0; i < 256; i++)
950 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
951 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
953 for (i = 0; i < BUF_LOCKS; i++) {
954 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
955 NULL, MUTEX_DEFAULT, NULL);
959 #define ARC_MINTIME (hz>>4) /* 62 ms */
962 arc_cksum_verify(arc_buf_t *buf)
966 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
969 mutex_enter(&buf->b_hdr->b_freeze_lock);
970 if (buf->b_hdr->b_freeze_cksum == NULL ||
971 (buf->b_hdr->b_flags & ARC_FLAG_IO_ERROR)) {
972 mutex_exit(&buf->b_hdr->b_freeze_lock);
975 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
976 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
977 panic("buffer modified while frozen!");
978 mutex_exit(&buf->b_hdr->b_freeze_lock);
982 arc_cksum_equal(arc_buf_t *buf)
987 mutex_enter(&buf->b_hdr->b_freeze_lock);
988 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
989 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
990 mutex_exit(&buf->b_hdr->b_freeze_lock);
996 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
998 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1001 mutex_enter(&buf->b_hdr->b_freeze_lock);
1002 if (buf->b_hdr->b_freeze_cksum != NULL) {
1003 mutex_exit(&buf->b_hdr->b_freeze_lock);
1006 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1007 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1008 buf->b_hdr->b_freeze_cksum);
1009 mutex_exit(&buf->b_hdr->b_freeze_lock);
1014 typedef struct procctl {
1022 arc_buf_unwatch(arc_buf_t *buf)
1029 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1030 ctl.prwatch.pr_size = 0;
1031 ctl.prwatch.pr_wflags = 0;
1032 result = write(arc_procfd, &ctl, sizeof (ctl));
1033 ASSERT3U(result, ==, sizeof (ctl));
1040 arc_buf_watch(arc_buf_t *buf)
1047 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1048 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1049 ctl.prwatch.pr_wflags = WA_WRITE;
1050 result = write(arc_procfd, &ctl, sizeof (ctl));
1051 ASSERT3U(result, ==, sizeof (ctl));
1057 arc_buf_thaw(arc_buf_t *buf)
1059 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1060 if (buf->b_hdr->b_state != arc_anon)
1061 panic("modifying non-anon buffer!");
1062 if (buf->b_hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1063 panic("modifying buffer while i/o in progress!");
1064 arc_cksum_verify(buf);
1067 mutex_enter(&buf->b_hdr->b_freeze_lock);
1068 if (buf->b_hdr->b_freeze_cksum != NULL) {
1069 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1070 buf->b_hdr->b_freeze_cksum = NULL;
1073 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1074 if (buf->b_hdr->b_thawed)
1075 kmem_free(buf->b_hdr->b_thawed, 1);
1076 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1079 mutex_exit(&buf->b_hdr->b_freeze_lock);
1081 arc_buf_unwatch(buf);
1085 arc_buf_freeze(arc_buf_t *buf)
1087 kmutex_t *hash_lock;
1089 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1092 hash_lock = HDR_LOCK(buf->b_hdr);
1093 mutex_enter(hash_lock);
1095 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1096 buf->b_hdr->b_state == arc_anon);
1097 arc_cksum_compute(buf, B_FALSE);
1098 mutex_exit(hash_lock);
1103 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1105 ASSERT(MUTEX_HELD(hash_lock));
1107 if ((refcount_add(&hdr->b_refcnt, tag) == 1) &&
1108 (hdr->b_state != arc_anon)) {
1109 uint64_t delta = hdr->b_size * hdr->b_datacnt;
1110 list_t *list = &hdr->b_state->arcs_list[hdr->b_type];
1111 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
1113 ASSERT(!MUTEX_HELD(&hdr->b_state->arcs_mtx));
1114 mutex_enter(&hdr->b_state->arcs_mtx);
1115 ASSERT(list_link_active(&hdr->b_arc_node));
1116 list_remove(list, hdr);
1117 if (GHOST_STATE(hdr->b_state)) {
1118 ASSERT0(hdr->b_datacnt);
1119 ASSERT3P(hdr->b_buf, ==, NULL);
1120 delta = hdr->b_size;
1123 ASSERT3U(*size, >=, delta);
1124 atomic_add_64(size, -delta);
1125 mutex_exit(&hdr->b_state->arcs_mtx);
1126 /* remove the prefetch flag if we get a reference */
1127 if (hdr->b_flags & ARC_FLAG_PREFETCH)
1128 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
1133 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1136 arc_state_t *state = hdr->b_state;
1138 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1139 ASSERT(!GHOST_STATE(state));
1141 if (((cnt = refcount_remove(&hdr->b_refcnt, tag)) == 0) &&
1142 (state != arc_anon)) {
1143 uint64_t *size = &state->arcs_lsize[hdr->b_type];
1145 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1146 mutex_enter(&state->arcs_mtx);
1147 ASSERT(!list_link_active(&hdr->b_arc_node));
1148 list_insert_head(&state->arcs_list[hdr->b_type], hdr);
1149 ASSERT(hdr->b_datacnt > 0);
1150 atomic_add_64(size, hdr->b_size * hdr->b_datacnt);
1151 mutex_exit(&state->arcs_mtx);
1157 * Move the supplied buffer to the indicated state. The mutex
1158 * for the buffer must be held by the caller.
1161 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
1162 kmutex_t *hash_lock)
1164 arc_state_t *old_state = hdr->b_state;
1165 int64_t refcnt = refcount_count(&hdr->b_refcnt);
1166 uint64_t from_delta, to_delta;
1168 ASSERT(MUTEX_HELD(hash_lock));
1169 ASSERT3P(new_state, !=, old_state);
1170 ASSERT(refcnt == 0 || hdr->b_datacnt > 0);
1171 ASSERT(hdr->b_datacnt == 0 || !GHOST_STATE(new_state));
1172 ASSERT(hdr->b_datacnt <= 1 || old_state != arc_anon);
1174 from_delta = to_delta = hdr->b_datacnt * hdr->b_size;
1177 * If this buffer is evictable, transfer it from the
1178 * old state list to the new state list.
1181 if (old_state != arc_anon) {
1182 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1183 uint64_t *size = &old_state->arcs_lsize[hdr->b_type];
1186 mutex_enter(&old_state->arcs_mtx);
1188 ASSERT(list_link_active(&hdr->b_arc_node));
1189 list_remove(&old_state->arcs_list[hdr->b_type], hdr);
1192 * If prefetching out of the ghost cache,
1193 * we will have a non-zero datacnt.
1195 if (GHOST_STATE(old_state) && hdr->b_datacnt == 0) {
1196 /* ghost elements have a ghost size */
1197 ASSERT(hdr->b_buf == NULL);
1198 from_delta = hdr->b_size;
1200 ASSERT3U(*size, >=, from_delta);
1201 atomic_add_64(size, -from_delta);
1204 mutex_exit(&old_state->arcs_mtx);
1206 if (new_state != arc_anon) {
1207 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1208 uint64_t *size = &new_state->arcs_lsize[hdr->b_type];
1211 mutex_enter(&new_state->arcs_mtx);
1213 list_insert_head(&new_state->arcs_list[hdr->b_type],
1216 /* ghost elements have a ghost size */
1217 if (GHOST_STATE(new_state)) {
1218 ASSERT(hdr->b_datacnt == 0);
1219 ASSERT(hdr->b_buf == NULL);
1220 to_delta = hdr->b_size;
1222 atomic_add_64(size, to_delta);
1225 mutex_exit(&new_state->arcs_mtx);
1229 ASSERT(!BUF_EMPTY(hdr));
1230 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
1231 buf_hash_remove(hdr);
1233 /* adjust state sizes */
1235 atomic_add_64(&new_state->arcs_size, to_delta);
1237 ASSERT3U(old_state->arcs_size, >=, from_delta);
1238 atomic_add_64(&old_state->arcs_size, -from_delta);
1240 hdr->b_state = new_state;
1242 /* adjust l2arc hdr stats */
1243 if (new_state == arc_l2c_only)
1244 l2arc_hdr_stat_add();
1245 else if (old_state == arc_l2c_only)
1246 l2arc_hdr_stat_remove();
1250 arc_space_consume(uint64_t space, arc_space_type_t type)
1252 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1255 case ARC_SPACE_DATA:
1256 ARCSTAT_INCR(arcstat_data_size, space);
1258 case ARC_SPACE_OTHER:
1259 ARCSTAT_INCR(arcstat_other_size, space);
1261 case ARC_SPACE_HDRS:
1262 ARCSTAT_INCR(arcstat_hdr_size, space);
1264 case ARC_SPACE_L2HDRS:
1265 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1269 ARCSTAT_INCR(arcstat_meta_used, space);
1270 atomic_add_64(&arc_size, space);
1274 arc_space_return(uint64_t space, arc_space_type_t type)
1276 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1279 case ARC_SPACE_DATA:
1280 ARCSTAT_INCR(arcstat_data_size, -space);
1282 case ARC_SPACE_OTHER:
1283 ARCSTAT_INCR(arcstat_other_size, -space);
1285 case ARC_SPACE_HDRS:
1286 ARCSTAT_INCR(arcstat_hdr_size, -space);
1288 case ARC_SPACE_L2HDRS:
1289 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1293 ASSERT(arc_meta_used >= space);
1294 if (arc_meta_max < arc_meta_used)
1295 arc_meta_max = arc_meta_used;
1296 ARCSTAT_INCR(arcstat_meta_used, -space);
1297 ASSERT(arc_size >= space);
1298 atomic_add_64(&arc_size, -space);
1302 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1307 ASSERT3U(size, >, 0);
1308 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1309 ASSERT(BUF_EMPTY(hdr));
1312 hdr->b_spa = spa_load_guid(spa);
1313 hdr->b_state = arc_anon;
1314 hdr->b_arc_access = 0;
1315 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1318 buf->b_efunc = NULL;
1319 buf->b_private = NULL;
1322 arc_get_data_buf(buf);
1325 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1326 (void) refcount_add(&hdr->b_refcnt, tag);
1331 static char *arc_onloan_tag = "onloan";
1334 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1335 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1336 * buffers must be returned to the arc before they can be used by the DMU or
1340 arc_loan_buf(spa_t *spa, int size)
1344 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1346 atomic_add_64(&arc_loaned_bytes, size);
1351 * Return a loaned arc buffer to the arc.
1354 arc_return_buf(arc_buf_t *buf, void *tag)
1356 arc_buf_hdr_t *hdr = buf->b_hdr;
1358 ASSERT(buf->b_data != NULL);
1359 (void) refcount_add(&hdr->b_refcnt, tag);
1360 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1362 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1365 /* Detach an arc_buf from a dbuf (tag) */
1367 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1371 ASSERT(buf->b_data != NULL);
1373 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1374 (void) refcount_remove(&hdr->b_refcnt, tag);
1375 buf->b_efunc = NULL;
1376 buf->b_private = NULL;
1378 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1382 arc_buf_clone(arc_buf_t *from)
1385 arc_buf_hdr_t *hdr = from->b_hdr;
1386 uint64_t size = hdr->b_size;
1388 ASSERT(hdr->b_state != arc_anon);
1390 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1393 buf->b_efunc = NULL;
1394 buf->b_private = NULL;
1395 buf->b_next = hdr->b_buf;
1397 arc_get_data_buf(buf);
1398 bcopy(from->b_data, buf->b_data, size);
1401 * This buffer already exists in the arc so create a duplicate
1402 * copy for the caller. If the buffer is associated with user data
1403 * then track the size and number of duplicates. These stats will be
1404 * updated as duplicate buffers are created and destroyed.
1406 if (hdr->b_type == ARC_BUFC_DATA) {
1407 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1408 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1410 hdr->b_datacnt += 1;
1415 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1418 kmutex_t *hash_lock;
1421 * Check to see if this buffer is evicted. Callers
1422 * must verify b_data != NULL to know if the add_ref
1425 mutex_enter(&buf->b_evict_lock);
1426 if (buf->b_data == NULL) {
1427 mutex_exit(&buf->b_evict_lock);
1430 hash_lock = HDR_LOCK(buf->b_hdr);
1431 mutex_enter(hash_lock);
1433 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1434 mutex_exit(&buf->b_evict_lock);
1436 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1437 add_reference(hdr, hash_lock, tag);
1438 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1439 arc_access(hdr, hash_lock);
1440 mutex_exit(hash_lock);
1441 ARCSTAT_BUMP(arcstat_hits);
1442 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH),
1443 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1444 data, metadata, hits);
1448 * Free the arc data buffer. If it is an l2arc write in progress,
1449 * the buffer is placed on l2arc_free_on_write to be freed later.
1452 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1454 arc_buf_hdr_t *hdr = buf->b_hdr;
1456 if (HDR_L2_WRITING(hdr)) {
1457 l2arc_data_free_t *df;
1458 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1459 df->l2df_data = buf->b_data;
1460 df->l2df_size = hdr->b_size;
1461 df->l2df_func = free_func;
1462 mutex_enter(&l2arc_free_on_write_mtx);
1463 list_insert_head(l2arc_free_on_write, df);
1464 mutex_exit(&l2arc_free_on_write_mtx);
1465 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1467 free_func(buf->b_data, hdr->b_size);
1472 * Free up buf->b_data and if 'remove' is set, then pull the
1473 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
1476 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove)
1480 /* free up data associated with the buf */
1482 arc_state_t *state = buf->b_hdr->b_state;
1483 uint64_t size = buf->b_hdr->b_size;
1484 arc_buf_contents_t type = buf->b_hdr->b_type;
1486 arc_cksum_verify(buf);
1487 arc_buf_unwatch(buf);
1490 if (type == ARC_BUFC_METADATA) {
1491 arc_buf_data_free(buf, zio_buf_free);
1492 arc_space_return(size, ARC_SPACE_DATA);
1494 ASSERT(type == ARC_BUFC_DATA);
1495 arc_buf_data_free(buf, zio_data_buf_free);
1496 ARCSTAT_INCR(arcstat_data_size, -size);
1497 atomic_add_64(&arc_size, -size);
1500 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1501 uint64_t *cnt = &state->arcs_lsize[type];
1503 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1504 ASSERT(state != arc_anon);
1506 ASSERT3U(*cnt, >=, size);
1507 atomic_add_64(cnt, -size);
1509 ASSERT3U(state->arcs_size, >=, size);
1510 atomic_add_64(&state->arcs_size, -size);
1514 * If we're destroying a duplicate buffer make sure
1515 * that the appropriate statistics are updated.
1517 if (buf->b_hdr->b_datacnt > 1 &&
1518 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1519 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1520 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1522 ASSERT(buf->b_hdr->b_datacnt > 0);
1523 buf->b_hdr->b_datacnt -= 1;
1526 /* only remove the buf if requested */
1530 /* remove the buf from the hdr list */
1531 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1533 *bufp = buf->b_next;
1536 ASSERT(buf->b_efunc == NULL);
1538 /* clean up the buf */
1540 kmem_cache_free(buf_cache, buf);
1544 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1546 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1547 ASSERT3P(hdr->b_state, ==, arc_anon);
1548 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1549 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1551 if (l2hdr != NULL) {
1552 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1554 * To prevent arc_free() and l2arc_evict() from
1555 * attempting to free the same buffer at the same time,
1556 * a FREE_IN_PROGRESS flag is given to arc_free() to
1557 * give it priority. l2arc_evict() can't destroy this
1558 * header while we are waiting on l2arc_buflist_mtx.
1560 * The hdr may be removed from l2ad_buflist before we
1561 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1563 if (!buflist_held) {
1564 mutex_enter(&l2arc_buflist_mtx);
1565 l2hdr = hdr->b_l2hdr;
1568 if (l2hdr != NULL) {
1569 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1570 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1571 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
1572 vdev_space_update(l2hdr->b_dev->l2ad_vdev,
1573 -l2hdr->b_asize, 0, 0);
1574 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1575 if (hdr->b_state == arc_l2c_only)
1576 l2arc_hdr_stat_remove();
1577 hdr->b_l2hdr = NULL;
1581 mutex_exit(&l2arc_buflist_mtx);
1584 if (!BUF_EMPTY(hdr)) {
1585 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1586 buf_discard_identity(hdr);
1588 while (hdr->b_buf) {
1589 arc_buf_t *buf = hdr->b_buf;
1592 mutex_enter(&arc_eviction_mtx);
1593 mutex_enter(&buf->b_evict_lock);
1594 ASSERT(buf->b_hdr != NULL);
1595 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1596 hdr->b_buf = buf->b_next;
1597 buf->b_hdr = &arc_eviction_hdr;
1598 buf->b_next = arc_eviction_list;
1599 arc_eviction_list = buf;
1600 mutex_exit(&buf->b_evict_lock);
1601 mutex_exit(&arc_eviction_mtx);
1603 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1606 if (hdr->b_freeze_cksum != NULL) {
1607 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1608 hdr->b_freeze_cksum = NULL;
1610 if (hdr->b_thawed) {
1611 kmem_free(hdr->b_thawed, 1);
1612 hdr->b_thawed = NULL;
1615 ASSERT(!list_link_active(&hdr->b_arc_node));
1616 ASSERT3P(hdr->b_hash_next, ==, NULL);
1617 ASSERT3P(hdr->b_acb, ==, NULL);
1618 kmem_cache_free(hdr_cache, hdr);
1622 arc_buf_free(arc_buf_t *buf, void *tag)
1624 arc_buf_hdr_t *hdr = buf->b_hdr;
1625 int hashed = hdr->b_state != arc_anon;
1627 ASSERT(buf->b_efunc == NULL);
1628 ASSERT(buf->b_data != NULL);
1631 kmutex_t *hash_lock = HDR_LOCK(hdr);
1633 mutex_enter(hash_lock);
1635 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1637 (void) remove_reference(hdr, hash_lock, tag);
1638 if (hdr->b_datacnt > 1) {
1639 arc_buf_destroy(buf, FALSE, TRUE);
1641 ASSERT(buf == hdr->b_buf);
1642 ASSERT(buf->b_efunc == NULL);
1643 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
1645 mutex_exit(hash_lock);
1646 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1649 * We are in the middle of an async write. Don't destroy
1650 * this buffer unless the write completes before we finish
1651 * decrementing the reference count.
1653 mutex_enter(&arc_eviction_mtx);
1654 (void) remove_reference(hdr, NULL, tag);
1655 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1656 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1657 mutex_exit(&arc_eviction_mtx);
1659 arc_hdr_destroy(hdr);
1661 if (remove_reference(hdr, NULL, tag) > 0)
1662 arc_buf_destroy(buf, FALSE, TRUE);
1664 arc_hdr_destroy(hdr);
1669 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1671 arc_buf_hdr_t *hdr = buf->b_hdr;
1672 kmutex_t *hash_lock = HDR_LOCK(hdr);
1673 boolean_t no_callback = (buf->b_efunc == NULL);
1675 if (hdr->b_state == arc_anon) {
1676 ASSERT(hdr->b_datacnt == 1);
1677 arc_buf_free(buf, tag);
1678 return (no_callback);
1681 mutex_enter(hash_lock);
1683 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1684 ASSERT(hdr->b_state != arc_anon);
1685 ASSERT(buf->b_data != NULL);
1687 (void) remove_reference(hdr, hash_lock, tag);
1688 if (hdr->b_datacnt > 1) {
1690 arc_buf_destroy(buf, FALSE, TRUE);
1691 } else if (no_callback) {
1692 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1693 ASSERT(buf->b_efunc == NULL);
1694 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
1696 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1697 refcount_is_zero(&hdr->b_refcnt));
1698 mutex_exit(hash_lock);
1699 return (no_callback);
1703 arc_buf_size(arc_buf_t *buf)
1705 return (buf->b_hdr->b_size);
1709 * Called from the DMU to determine if the current buffer should be
1710 * evicted. In order to ensure proper locking, the eviction must be initiated
1711 * from the DMU. Return true if the buffer is associated with user data and
1712 * duplicate buffers still exist.
1715 arc_buf_eviction_needed(arc_buf_t *buf)
1718 boolean_t evict_needed = B_FALSE;
1720 if (zfs_disable_dup_eviction)
1723 mutex_enter(&buf->b_evict_lock);
1727 * We are in arc_do_user_evicts(); let that function
1728 * perform the eviction.
1730 ASSERT(buf->b_data == NULL);
1731 mutex_exit(&buf->b_evict_lock);
1733 } else if (buf->b_data == NULL) {
1735 * We have already been added to the arc eviction list;
1736 * recommend eviction.
1738 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1739 mutex_exit(&buf->b_evict_lock);
1743 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1744 evict_needed = B_TRUE;
1746 mutex_exit(&buf->b_evict_lock);
1747 return (evict_needed);
1751 * Evict buffers from list until we've removed the specified number of
1752 * bytes. Move the removed buffers to the appropriate evict state.
1753 * If the recycle flag is set, then attempt to "recycle" a buffer:
1754 * - look for a buffer to evict that is `bytes' long.
1755 * - return the data block from this buffer rather than freeing it.
1756 * This flag is used by callers that are trying to make space for a
1757 * new buffer in a full arc cache.
1759 * This function makes a "best effort". It skips over any buffers
1760 * it can't get a hash_lock on, and so may not catch all candidates.
1761 * It may also return without evicting as much space as requested.
1764 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1765 arc_buf_contents_t type)
1767 arc_state_t *evicted_state;
1768 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1769 arc_buf_hdr_t *hdr, *hdr_prev = NULL;
1770 kmutex_t *hash_lock;
1771 boolean_t have_lock;
1772 void *stolen = NULL;
1773 arc_buf_hdr_t marker = { 0 };
1776 ASSERT(state == arc_mru || state == arc_mfu);
1778 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1780 mutex_enter(&state->arcs_mtx);
1781 mutex_enter(&evicted_state->arcs_mtx);
1784 * Decide which "type" (data vs metadata) to recycle from.
1786 * If we are over the metadata limit, recycle from metadata.
1787 * If we are under the metadata minimum, recycle from data.
1788 * Otherwise, recycle from whichever type has the oldest (least
1789 * recently accessed) header.
1792 arc_buf_hdr_t *data_hdr =
1793 list_tail(&state->arcs_list[ARC_BUFC_DATA]);
1794 arc_buf_hdr_t *metadata_hdr =
1795 list_tail(&state->arcs_list[ARC_BUFC_METADATA]);
1796 arc_buf_contents_t realtype;
1797 if (data_hdr == NULL) {
1798 realtype = ARC_BUFC_METADATA;
1799 } else if (metadata_hdr == NULL) {
1800 realtype = ARC_BUFC_DATA;
1801 } else if (arc_meta_used >= arc_meta_limit) {
1802 realtype = ARC_BUFC_METADATA;
1803 } else if (arc_meta_used <= arc_meta_min) {
1804 realtype = ARC_BUFC_DATA;
1806 if (data_hdr->b_arc_access <
1807 metadata_hdr->b_arc_access) {
1808 realtype = ARC_BUFC_DATA;
1810 realtype = ARC_BUFC_METADATA;
1813 if (realtype != type) {
1815 * If we want to evict from a different list,
1816 * we can not recycle, because DATA vs METADATA
1817 * buffers are segregated into different kmem
1818 * caches (and vmem arenas).
1825 list_t *list = &state->arcs_list[type];
1827 for (hdr = list_tail(list); hdr; hdr = hdr_prev) {
1828 hdr_prev = list_prev(list, hdr);
1829 /* prefetch buffers have a minimum lifespan */
1830 if (HDR_IO_IN_PROGRESS(hdr) ||
1831 (spa && hdr->b_spa != spa) ||
1832 (hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT) &&
1833 ddi_get_lbolt() - hdr->b_arc_access <
1834 arc_min_prefetch_lifespan)) {
1838 /* "lookahead" for better eviction candidate */
1839 if (recycle && hdr->b_size != bytes &&
1840 hdr_prev && hdr_prev->b_size == bytes)
1843 /* ignore markers */
1844 if (hdr->b_spa == 0)
1848 * It may take a long time to evict all the bufs requested.
1849 * To avoid blocking all arc activity, periodically drop
1850 * the arcs_mtx and give other threads a chance to run
1851 * before reacquiring the lock.
1853 * If we are looking for a buffer to recycle, we are in
1854 * the hot code path, so don't sleep.
1856 if (!recycle && count++ > arc_evict_iterations) {
1857 list_insert_after(list, hdr, &marker);
1858 mutex_exit(&evicted_state->arcs_mtx);
1859 mutex_exit(&state->arcs_mtx);
1860 kpreempt(KPREEMPT_SYNC);
1861 mutex_enter(&state->arcs_mtx);
1862 mutex_enter(&evicted_state->arcs_mtx);
1863 hdr_prev = list_prev(list, &marker);
1864 list_remove(list, &marker);
1869 hash_lock = HDR_LOCK(hdr);
1870 have_lock = MUTEX_HELD(hash_lock);
1871 if (have_lock || mutex_tryenter(hash_lock)) {
1872 ASSERT0(refcount_count(&hdr->b_refcnt));
1873 ASSERT(hdr->b_datacnt > 0);
1874 while (hdr->b_buf) {
1875 arc_buf_t *buf = hdr->b_buf;
1876 if (!mutex_tryenter(&buf->b_evict_lock)) {
1881 bytes_evicted += hdr->b_size;
1882 if (recycle && hdr->b_type == type &&
1883 hdr->b_size == bytes &&
1884 !HDR_L2_WRITING(hdr)) {
1885 stolen = buf->b_data;
1890 mutex_enter(&arc_eviction_mtx);
1891 arc_buf_destroy(buf,
1892 buf->b_data == stolen, FALSE);
1893 hdr->b_buf = buf->b_next;
1894 buf->b_hdr = &arc_eviction_hdr;
1895 buf->b_next = arc_eviction_list;
1896 arc_eviction_list = buf;
1897 mutex_exit(&arc_eviction_mtx);
1898 mutex_exit(&buf->b_evict_lock);
1900 mutex_exit(&buf->b_evict_lock);
1901 arc_buf_destroy(buf,
1902 buf->b_data == stolen, TRUE);
1907 ARCSTAT_INCR(arcstat_evict_l2_cached,
1910 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
1911 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1915 arcstat_evict_l2_ineligible,
1920 if (hdr->b_datacnt == 0) {
1921 arc_change_state(evicted_state, hdr, hash_lock);
1922 ASSERT(HDR_IN_HASH_TABLE(hdr));
1923 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
1924 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
1925 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
1928 mutex_exit(hash_lock);
1929 if (bytes >= 0 && bytes_evicted >= bytes)
1936 mutex_exit(&evicted_state->arcs_mtx);
1937 mutex_exit(&state->arcs_mtx);
1939 if (bytes_evicted < bytes)
1940 dprintf("only evicted %lld bytes from %x",
1941 (longlong_t)bytes_evicted, state);
1944 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1947 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1950 * Note: we have just evicted some data into the ghost state,
1951 * potentially putting the ghost size over the desired size. Rather
1952 * that evicting from the ghost list in this hot code path, leave
1953 * this chore to the arc_reclaim_thread().
1960 * Remove buffers from list until we've removed the specified number of
1961 * bytes. Destroy the buffers that are removed.
1964 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1966 arc_buf_hdr_t *hdr, *hdr_prev;
1967 arc_buf_hdr_t marker = { 0 };
1968 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1969 kmutex_t *hash_lock;
1970 uint64_t bytes_deleted = 0;
1971 uint64_t bufs_skipped = 0;
1974 ASSERT(GHOST_STATE(state));
1976 mutex_enter(&state->arcs_mtx);
1977 for (hdr = list_tail(list); hdr; hdr = hdr_prev) {
1978 hdr_prev = list_prev(list, hdr);
1979 if (hdr->b_type > ARC_BUFC_NUMTYPES)
1980 panic("invalid hdr=%p", (void *)hdr);
1981 if (spa && hdr->b_spa != spa)
1984 /* ignore markers */
1985 if (hdr->b_spa == 0)
1988 hash_lock = HDR_LOCK(hdr);
1989 /* caller may be trying to modify this buffer, skip it */
1990 if (MUTEX_HELD(hash_lock))
1994 * It may take a long time to evict all the bufs requested.
1995 * To avoid blocking all arc activity, periodically drop
1996 * the arcs_mtx and give other threads a chance to run
1997 * before reacquiring the lock.
1999 if (count++ > arc_evict_iterations) {
2000 list_insert_after(list, hdr, &marker);
2001 mutex_exit(&state->arcs_mtx);
2002 kpreempt(KPREEMPT_SYNC);
2003 mutex_enter(&state->arcs_mtx);
2004 hdr_prev = list_prev(list, &marker);
2005 list_remove(list, &marker);
2009 if (mutex_tryenter(hash_lock)) {
2010 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2011 ASSERT(hdr->b_buf == NULL);
2012 ARCSTAT_BUMP(arcstat_deleted);
2013 bytes_deleted += hdr->b_size;
2015 if (hdr->b_l2hdr != NULL) {
2017 * This buffer is cached on the 2nd Level ARC;
2018 * don't destroy the header.
2020 arc_change_state(arc_l2c_only, hdr, hash_lock);
2021 mutex_exit(hash_lock);
2023 arc_change_state(arc_anon, hdr, hash_lock);
2024 mutex_exit(hash_lock);
2025 arc_hdr_destroy(hdr);
2028 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
2029 if (bytes >= 0 && bytes_deleted >= bytes)
2031 } else if (bytes < 0) {
2033 * Insert a list marker and then wait for the
2034 * hash lock to become available. Once its
2035 * available, restart from where we left off.
2037 list_insert_after(list, hdr, &marker);
2038 mutex_exit(&state->arcs_mtx);
2039 mutex_enter(hash_lock);
2040 mutex_exit(hash_lock);
2041 mutex_enter(&state->arcs_mtx);
2042 hdr_prev = list_prev(list, &marker);
2043 list_remove(list, &marker);
2049 mutex_exit(&state->arcs_mtx);
2051 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
2052 (bytes < 0 || bytes_deleted < bytes)) {
2053 list = &state->arcs_list[ARC_BUFC_METADATA];
2058 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2062 if (bytes_deleted < bytes)
2063 dprintf("only deleted %lld bytes from %p",
2064 (longlong_t)bytes_deleted, state);
2070 int64_t adjustment, delta;
2076 adjustment = MIN((int64_t)(arc_size - arc_c),
2077 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2080 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2081 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2082 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA);
2083 adjustment -= delta;
2086 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2087 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2088 (void) arc_evict(arc_mru, NULL, delta, FALSE,
2096 adjustment = arc_size - arc_c;
2098 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2099 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2100 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA);
2101 adjustment -= delta;
2104 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2105 int64_t delta = MIN(adjustment,
2106 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2107 (void) arc_evict(arc_mfu, NULL, delta, FALSE,
2112 * Adjust ghost lists
2115 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2117 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2118 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2119 arc_evict_ghost(arc_mru_ghost, NULL, delta);
2123 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2125 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2126 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2127 arc_evict_ghost(arc_mfu_ghost, NULL, delta);
2132 arc_do_user_evicts(void)
2134 mutex_enter(&arc_eviction_mtx);
2135 while (arc_eviction_list != NULL) {
2136 arc_buf_t *buf = arc_eviction_list;
2137 arc_eviction_list = buf->b_next;
2138 mutex_enter(&buf->b_evict_lock);
2140 mutex_exit(&buf->b_evict_lock);
2141 mutex_exit(&arc_eviction_mtx);
2143 if (buf->b_efunc != NULL)
2144 VERIFY0(buf->b_efunc(buf->b_private));
2146 buf->b_efunc = NULL;
2147 buf->b_private = NULL;
2148 kmem_cache_free(buf_cache, buf);
2149 mutex_enter(&arc_eviction_mtx);
2151 mutex_exit(&arc_eviction_mtx);
2155 * Flush all *evictable* data from the cache for the given spa.
2156 * NOTE: this will not touch "active" (i.e. referenced) data.
2159 arc_flush(spa_t *spa)
2164 guid = spa_load_guid(spa);
2166 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2167 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2171 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2172 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2176 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2177 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2181 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2182 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2187 arc_evict_ghost(arc_mru_ghost, guid, -1);
2188 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2190 mutex_enter(&arc_reclaim_thr_lock);
2191 arc_do_user_evicts();
2192 mutex_exit(&arc_reclaim_thr_lock);
2193 ASSERT(spa || arc_eviction_list == NULL);
2199 if (arc_c > arc_c_min) {
2203 to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
2205 to_free = arc_c >> arc_shrink_shift;
2207 if (arc_c > arc_c_min + to_free)
2208 atomic_add_64(&arc_c, -to_free);
2212 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2213 if (arc_c > arc_size)
2214 arc_c = MAX(arc_size, arc_c_min);
2216 arc_p = (arc_c >> 1);
2217 ASSERT(arc_c >= arc_c_min);
2218 ASSERT((int64_t)arc_p >= 0);
2221 if (arc_size > arc_c)
2226 * Determine if the system is under memory pressure and is asking
2227 * to reclaim memory. A return value of 1 indicates that the system
2228 * is under memory pressure and that the arc should adjust accordingly.
2231 arc_reclaim_needed(void)
2241 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2246 * check that we're out of range of the pageout scanner. It starts to
2247 * schedule paging if freemem is less than lotsfree and needfree.
2248 * lotsfree is the high-water mark for pageout, and needfree is the
2249 * number of needed free pages. We add extra pages here to make sure
2250 * the scanner doesn't start up while we're freeing memory.
2252 if (freemem < lotsfree + needfree + extra)
2256 * check to make sure that swapfs has enough space so that anon
2257 * reservations can still succeed. anon_resvmem() checks that the
2258 * availrmem is greater than swapfs_minfree, and the number of reserved
2259 * swap pages. We also add a bit of extra here just to prevent
2260 * circumstances from getting really dire.
2262 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2266 * Check that we have enough availrmem that memory locking (e.g., via
2267 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
2268 * stores the number of pages that cannot be locked; when availrmem
2269 * drops below pages_pp_maximum, page locking mechanisms such as
2270 * page_pp_lock() will fail.)
2272 if (availrmem <= pages_pp_maximum)
2277 * If we're on an i386 platform, it's possible that we'll exhaust the
2278 * kernel heap space before we ever run out of available physical
2279 * memory. Most checks of the size of the heap_area compare against
2280 * tune.t_minarmem, which is the minimum available real memory that we
2281 * can have in the system. However, this is generally fixed at 25 pages
2282 * which is so low that it's useless. In this comparison, we seek to
2283 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2284 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2287 if (vmem_size(heap_arena, VMEM_FREE) <
2288 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2))
2293 * If zio data pages are being allocated out of a separate heap segment,
2294 * then enforce that the size of available vmem for this arena remains
2295 * above about 1/16th free.
2297 * Note: The 1/16th arena free requirement was put in place
2298 * to aggressively evict memory from the arc in order to avoid
2299 * memory fragmentation issues.
2301 if (zio_arena != NULL &&
2302 vmem_size(zio_arena, VMEM_FREE) <
2303 (vmem_size(zio_arena, VMEM_ALLOC) >> 4))
2306 if (spa_get_random(100) == 0)
2313 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2316 kmem_cache_t *prev_cache = NULL;
2317 kmem_cache_t *prev_data_cache = NULL;
2318 extern kmem_cache_t *zio_buf_cache[];
2319 extern kmem_cache_t *zio_data_buf_cache[];
2320 extern kmem_cache_t *range_seg_cache;
2323 if (arc_meta_used >= arc_meta_limit) {
2325 * We are exceeding our meta-data cache limit.
2326 * Purge some DNLC entries to release holds on meta-data.
2328 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2332 * Reclaim unused memory from all kmem caches.
2339 * An aggressive reclamation will shrink the cache size as well as
2340 * reap free buffers from the arc kmem caches.
2342 if (strat == ARC_RECLAIM_AGGR)
2345 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2346 if (zio_buf_cache[i] != prev_cache) {
2347 prev_cache = zio_buf_cache[i];
2348 kmem_cache_reap_now(zio_buf_cache[i]);
2350 if (zio_data_buf_cache[i] != prev_data_cache) {
2351 prev_data_cache = zio_data_buf_cache[i];
2352 kmem_cache_reap_now(zio_data_buf_cache[i]);
2355 kmem_cache_reap_now(buf_cache);
2356 kmem_cache_reap_now(hdr_cache);
2357 kmem_cache_reap_now(range_seg_cache);
2360 * Ask the vmem areana to reclaim unused memory from its
2363 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR)
2364 vmem_qcache_reap(zio_arena);
2368 arc_reclaim_thread(void)
2370 clock_t growtime = 0;
2371 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2374 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2376 mutex_enter(&arc_reclaim_thr_lock);
2377 while (arc_thread_exit == 0) {
2378 if (arc_reclaim_needed()) {
2381 if (last_reclaim == ARC_RECLAIM_CONS) {
2382 last_reclaim = ARC_RECLAIM_AGGR;
2384 last_reclaim = ARC_RECLAIM_CONS;
2388 last_reclaim = ARC_RECLAIM_AGGR;
2392 /* reset the growth delay for every reclaim */
2393 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2395 arc_kmem_reap_now(last_reclaim);
2398 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2399 arc_no_grow = FALSE;
2404 if (arc_eviction_list != NULL)
2405 arc_do_user_evicts();
2407 /* block until needed, or one second, whichever is shorter */
2408 CALLB_CPR_SAFE_BEGIN(&cpr);
2409 (void) cv_timedwait(&arc_reclaim_thr_cv,
2410 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2411 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2414 arc_thread_exit = 0;
2415 cv_broadcast(&arc_reclaim_thr_cv);
2416 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2421 * Adapt arc info given the number of bytes we are trying to add and
2422 * the state that we are comming from. This function is only called
2423 * when we are adding new content to the cache.
2426 arc_adapt(int bytes, arc_state_t *state)
2429 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2431 if (state == arc_l2c_only)
2436 * Adapt the target size of the MRU list:
2437 * - if we just hit in the MRU ghost list, then increase
2438 * the target size of the MRU list.
2439 * - if we just hit in the MFU ghost list, then increase
2440 * the target size of the MFU list by decreasing the
2441 * target size of the MRU list.
2443 if (state == arc_mru_ghost) {
2444 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2445 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2446 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2448 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2449 } else if (state == arc_mfu_ghost) {
2452 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2453 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2454 mult = MIN(mult, 10);
2456 delta = MIN(bytes * mult, arc_p);
2457 arc_p = MAX(arc_p_min, arc_p - delta);
2459 ASSERT((int64_t)arc_p >= 0);
2461 if (arc_reclaim_needed()) {
2462 cv_signal(&arc_reclaim_thr_cv);
2469 if (arc_c >= arc_c_max)
2473 * If we're within (2 * maxblocksize) bytes of the target
2474 * cache size, increment the target cache size
2476 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2477 atomic_add_64(&arc_c, (int64_t)bytes);
2478 if (arc_c > arc_c_max)
2480 else if (state == arc_anon)
2481 atomic_add_64(&arc_p, (int64_t)bytes);
2485 ASSERT((int64_t)arc_p >= 0);
2489 * Check if the cache has reached its limits and eviction is required
2493 arc_evict_needed(arc_buf_contents_t type)
2495 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2498 if (arc_reclaim_needed())
2501 return (arc_size > arc_c);
2505 * The buffer, supplied as the first argument, needs a data block.
2506 * So, if we are at cache max, determine which cache should be victimized.
2507 * We have the following cases:
2509 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2510 * In this situation if we're out of space, but the resident size of the MFU is
2511 * under the limit, victimize the MFU cache to satisfy this insertion request.
2513 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2514 * Here, we've used up all of the available space for the MRU, so we need to
2515 * evict from our own cache instead. Evict from the set of resident MRU
2518 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2519 * c minus p represents the MFU space in the cache, since p is the size of the
2520 * cache that is dedicated to the MRU. In this situation there's still space on
2521 * the MFU side, so the MRU side needs to be victimized.
2523 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2524 * MFU's resident set is consuming more space than it has been allotted. In
2525 * this situation, we must victimize our own cache, the MFU, for this insertion.
2528 arc_get_data_buf(arc_buf_t *buf)
2530 arc_state_t *state = buf->b_hdr->b_state;
2531 uint64_t size = buf->b_hdr->b_size;
2532 arc_buf_contents_t type = buf->b_hdr->b_type;
2534 arc_adapt(size, state);
2537 * We have not yet reached cache maximum size,
2538 * just allocate a new buffer.
2540 if (!arc_evict_needed(type)) {
2541 if (type == ARC_BUFC_METADATA) {
2542 buf->b_data = zio_buf_alloc(size);
2543 arc_space_consume(size, ARC_SPACE_DATA);
2545 ASSERT(type == ARC_BUFC_DATA);
2546 buf->b_data = zio_data_buf_alloc(size);
2547 ARCSTAT_INCR(arcstat_data_size, size);
2548 atomic_add_64(&arc_size, size);
2554 * If we are prefetching from the mfu ghost list, this buffer
2555 * will end up on the mru list; so steal space from there.
2557 if (state == arc_mfu_ghost)
2558 state = buf->b_hdr->b_flags & ARC_FLAG_PREFETCH ?
2560 else if (state == arc_mru_ghost)
2563 if (state == arc_mru || state == arc_anon) {
2564 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2565 state = (arc_mfu->arcs_lsize[type] >= size &&
2566 arc_p > mru_used) ? arc_mfu : arc_mru;
2569 uint64_t mfu_space = arc_c - arc_p;
2570 state = (arc_mru->arcs_lsize[type] >= size &&
2571 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2573 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) {
2574 if (type == ARC_BUFC_METADATA) {
2575 buf->b_data = zio_buf_alloc(size);
2576 arc_space_consume(size, ARC_SPACE_DATA);
2578 ASSERT(type == ARC_BUFC_DATA);
2579 buf->b_data = zio_data_buf_alloc(size);
2580 ARCSTAT_INCR(arcstat_data_size, size);
2581 atomic_add_64(&arc_size, size);
2583 ARCSTAT_BUMP(arcstat_recycle_miss);
2585 ASSERT(buf->b_data != NULL);
2588 * Update the state size. Note that ghost states have a
2589 * "ghost size" and so don't need to be updated.
2591 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2592 arc_buf_hdr_t *hdr = buf->b_hdr;
2594 atomic_add_64(&hdr->b_state->arcs_size, size);
2595 if (list_link_active(&hdr->b_arc_node)) {
2596 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2597 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2600 * If we are growing the cache, and we are adding anonymous
2601 * data, and we have outgrown arc_p, update arc_p
2603 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2604 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2605 arc_p = MIN(arc_c, arc_p + size);
2610 * This routine is called whenever a buffer is accessed.
2611 * NOTE: the hash lock is dropped in this function.
2614 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
2618 ASSERT(MUTEX_HELD(hash_lock));
2620 if (hdr->b_state == arc_anon) {
2622 * This buffer is not in the cache, and does not
2623 * appear in our "ghost" list. Add the new buffer
2627 ASSERT(hdr->b_arc_access == 0);
2628 hdr->b_arc_access = ddi_get_lbolt();
2629 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
2630 arc_change_state(arc_mru, hdr, hash_lock);
2632 } else if (hdr->b_state == arc_mru) {
2633 now = ddi_get_lbolt();
2636 * If this buffer is here because of a prefetch, then either:
2637 * - clear the flag if this is a "referencing" read
2638 * (any subsequent access will bump this into the MFU state).
2640 * - move the buffer to the head of the list if this is
2641 * another prefetch (to make it less likely to be evicted).
2643 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) {
2644 if (refcount_count(&hdr->b_refcnt) == 0) {
2645 ASSERT(list_link_active(&hdr->b_arc_node));
2647 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
2648 ARCSTAT_BUMP(arcstat_mru_hits);
2650 hdr->b_arc_access = now;
2655 * This buffer has been "accessed" only once so far,
2656 * but it is still in the cache. Move it to the MFU
2659 if (now > hdr->b_arc_access + ARC_MINTIME) {
2661 * More than 125ms have passed since we
2662 * instantiated this buffer. Move it to the
2663 * most frequently used state.
2665 hdr->b_arc_access = now;
2666 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
2667 arc_change_state(arc_mfu, hdr, hash_lock);
2669 ARCSTAT_BUMP(arcstat_mru_hits);
2670 } else if (hdr->b_state == arc_mru_ghost) {
2671 arc_state_t *new_state;
2673 * This buffer has been "accessed" recently, but
2674 * was evicted from the cache. Move it to the
2678 if (hdr->b_flags & ARC_FLAG_PREFETCH) {
2679 new_state = arc_mru;
2680 if (refcount_count(&hdr->b_refcnt) > 0)
2681 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
2682 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
2684 new_state = arc_mfu;
2685 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
2688 hdr->b_arc_access = ddi_get_lbolt();
2689 arc_change_state(new_state, hdr, hash_lock);
2691 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2692 } else if (hdr->b_state == arc_mfu) {
2694 * This buffer has been accessed more than once and is
2695 * still in the cache. Keep it in the MFU state.
2697 * NOTE: an add_reference() that occurred when we did
2698 * the arc_read() will have kicked this off the list.
2699 * If it was a prefetch, we will explicitly move it to
2700 * the head of the list now.
2702 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) {
2703 ASSERT(refcount_count(&hdr->b_refcnt) == 0);
2704 ASSERT(list_link_active(&hdr->b_arc_node));
2706 ARCSTAT_BUMP(arcstat_mfu_hits);
2707 hdr->b_arc_access = ddi_get_lbolt();
2708 } else if (hdr->b_state == arc_mfu_ghost) {
2709 arc_state_t *new_state = arc_mfu;
2711 * This buffer has been accessed more than once but has
2712 * been evicted from the cache. Move it back to the
2716 if (hdr->b_flags & ARC_FLAG_PREFETCH) {
2718 * This is a prefetch access...
2719 * move this block back to the MRU state.
2721 ASSERT0(refcount_count(&hdr->b_refcnt));
2722 new_state = arc_mru;
2725 hdr->b_arc_access = ddi_get_lbolt();
2726 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
2727 arc_change_state(new_state, hdr, hash_lock);
2729 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2730 } else if (hdr->b_state == arc_l2c_only) {
2732 * This buffer is on the 2nd Level ARC.
2735 hdr->b_arc_access = ddi_get_lbolt();
2736 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
2737 arc_change_state(arc_mfu, hdr, hash_lock);
2739 ASSERT(!"invalid arc state");
2743 /* a generic arc_done_func_t which you can use */
2746 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2748 if (zio == NULL || zio->io_error == 0)
2749 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2750 VERIFY(arc_buf_remove_ref(buf, arg));
2753 /* a generic arc_done_func_t */
2755 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2757 arc_buf_t **bufp = arg;
2758 if (zio && zio->io_error) {
2759 VERIFY(arc_buf_remove_ref(buf, arg));
2763 ASSERT(buf->b_data);
2768 arc_read_done(zio_t *zio)
2772 arc_buf_t *abuf; /* buffer we're assigning to callback */
2773 kmutex_t *hash_lock = NULL;
2774 arc_callback_t *callback_list, *acb;
2775 int freeable = FALSE;
2777 buf = zio->io_private;
2781 * The hdr was inserted into hash-table and removed from lists
2782 * prior to starting I/O. We should find this header, since
2783 * it's in the hash table, and it should be legit since it's
2784 * not possible to evict it during the I/O. The only possible
2785 * reason for it not to be found is if we were freed during the
2788 if (HDR_IN_HASH_TABLE(hdr)) {
2789 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
2790 ASSERT3U(hdr->b_dva.dva_word[0], ==,
2791 BP_IDENTITY(zio->io_bp)->dva_word[0]);
2792 ASSERT3U(hdr->b_dva.dva_word[1], ==,
2793 BP_IDENTITY(zio->io_bp)->dva_word[1]);
2795 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
2798 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
2799 hash_lock == NULL) ||
2801 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2802 (found == hdr && HDR_L2_READING(hdr)));
2805 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED;
2806 if (l2arc_noprefetch && (hdr->b_flags & ARC_FLAG_PREFETCH))
2807 hdr->b_flags &= ~ARC_FLAG_L2CACHE;
2809 /* byteswap if necessary */
2810 callback_list = hdr->b_acb;
2811 ASSERT(callback_list != NULL);
2812 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2813 dmu_object_byteswap_t bswap =
2814 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2815 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2816 byteswap_uint64_array :
2817 dmu_ot_byteswap[bswap].ob_func;
2818 func(buf->b_data, hdr->b_size);
2821 arc_cksum_compute(buf, B_FALSE);
2824 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2826 * Only call arc_access on anonymous buffers. This is because
2827 * if we've issued an I/O for an evicted buffer, we've already
2828 * called arc_access (to prevent any simultaneous readers from
2829 * getting confused).
2831 arc_access(hdr, hash_lock);
2834 /* create copies of the data buffer for the callers */
2836 for (acb = callback_list; acb; acb = acb->acb_next) {
2837 if (acb->acb_done) {
2839 ARCSTAT_BUMP(arcstat_duplicate_reads);
2840 abuf = arc_buf_clone(buf);
2842 acb->acb_buf = abuf;
2847 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
2848 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2850 ASSERT(buf->b_efunc == NULL);
2851 ASSERT(hdr->b_datacnt == 1);
2852 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2855 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2857 if (zio->io_error != 0) {
2858 hdr->b_flags |= ARC_FLAG_IO_ERROR;
2859 if (hdr->b_state != arc_anon)
2860 arc_change_state(arc_anon, hdr, hash_lock);
2861 if (HDR_IN_HASH_TABLE(hdr))
2862 buf_hash_remove(hdr);
2863 freeable = refcount_is_zero(&hdr->b_refcnt);
2867 * Broadcast before we drop the hash_lock to avoid the possibility
2868 * that the hdr (and hence the cv) might be freed before we get to
2869 * the cv_broadcast().
2871 cv_broadcast(&hdr->b_cv);
2874 mutex_exit(hash_lock);
2877 * This block was freed while we waited for the read to
2878 * complete. It has been removed from the hash table and
2879 * moved to the anonymous state (so that it won't show up
2882 ASSERT3P(hdr->b_state, ==, arc_anon);
2883 freeable = refcount_is_zero(&hdr->b_refcnt);
2886 /* execute each callback and free its structure */
2887 while ((acb = callback_list) != NULL) {
2889 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2891 if (acb->acb_zio_dummy != NULL) {
2892 acb->acb_zio_dummy->io_error = zio->io_error;
2893 zio_nowait(acb->acb_zio_dummy);
2896 callback_list = acb->acb_next;
2897 kmem_free(acb, sizeof (arc_callback_t));
2901 arc_hdr_destroy(hdr);
2905 * "Read" the block at the specified DVA (in bp) via the
2906 * cache. If the block is found in the cache, invoke the provided
2907 * callback immediately and return. Note that the `zio' parameter
2908 * in the callback will be NULL in this case, since no IO was
2909 * required. If the block is not in the cache pass the read request
2910 * on to the spa with a substitute callback function, so that the
2911 * requested block will be added to the cache.
2913 * If a read request arrives for a block that has a read in-progress,
2914 * either wait for the in-progress read to complete (and return the
2915 * results); or, if this is a read with a "done" func, add a record
2916 * to the read to invoke the "done" func when the read completes,
2917 * and return; or just return.
2919 * arc_read_done() will invoke all the requested "done" functions
2920 * for readers of this block.
2923 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
2924 void *private, zio_priority_t priority, int zio_flags,
2925 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
2927 arc_buf_hdr_t *hdr = NULL;
2928 arc_buf_t *buf = NULL;
2929 kmutex_t *hash_lock = NULL;
2931 uint64_t guid = spa_load_guid(spa);
2933 ASSERT(!BP_IS_EMBEDDED(bp) ||
2934 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
2937 if (!BP_IS_EMBEDDED(bp)) {
2939 * Embedded BP's have no DVA and require no I/O to "read".
2940 * Create an anonymous arc buf to back it.
2942 hdr = buf_hash_find(guid, bp, &hash_lock);
2945 if (hdr != NULL && hdr->b_datacnt > 0) {
2947 *arc_flags |= ARC_FLAG_CACHED;
2949 if (HDR_IO_IN_PROGRESS(hdr)) {
2951 if (*arc_flags & ARC_FLAG_WAIT) {
2952 cv_wait(&hdr->b_cv, hash_lock);
2953 mutex_exit(hash_lock);
2956 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
2959 arc_callback_t *acb = NULL;
2961 acb = kmem_zalloc(sizeof (arc_callback_t),
2963 acb->acb_done = done;
2964 acb->acb_private = private;
2966 acb->acb_zio_dummy = zio_null(pio,
2967 spa, NULL, NULL, NULL, zio_flags);
2969 ASSERT(acb->acb_done != NULL);
2970 acb->acb_next = hdr->b_acb;
2972 add_reference(hdr, hash_lock, private);
2973 mutex_exit(hash_lock);
2976 mutex_exit(hash_lock);
2980 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2983 add_reference(hdr, hash_lock, private);
2985 * If this block is already in use, create a new
2986 * copy of the data so that we will be guaranteed
2987 * that arc_release() will always succeed.
2991 ASSERT(buf->b_data);
2992 if (HDR_BUF_AVAILABLE(hdr)) {
2993 ASSERT(buf->b_efunc == NULL);
2994 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
2996 buf = arc_buf_clone(buf);
2999 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
3000 refcount_count(&hdr->b_refcnt) == 0) {
3001 hdr->b_flags |= ARC_FLAG_PREFETCH;
3003 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
3004 arc_access(hdr, hash_lock);
3005 if (*arc_flags & ARC_FLAG_L2CACHE)
3006 hdr->b_flags |= ARC_FLAG_L2CACHE;
3007 if (*arc_flags & ARC_FLAG_L2COMPRESS)
3008 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
3009 mutex_exit(hash_lock);
3010 ARCSTAT_BUMP(arcstat_hits);
3011 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH),
3012 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3013 data, metadata, hits);
3016 done(NULL, buf, private);
3018 uint64_t size = BP_GET_LSIZE(bp);
3019 arc_callback_t *acb;
3022 boolean_t devw = B_FALSE;
3023 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
3024 uint64_t b_asize = 0;
3027 /* this block is not in the cache */
3028 arc_buf_hdr_t *exists = NULL;
3029 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
3030 buf = arc_buf_alloc(spa, size, private, type);
3032 if (!BP_IS_EMBEDDED(bp)) {
3033 hdr->b_dva = *BP_IDENTITY(bp);
3034 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3035 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3036 exists = buf_hash_insert(hdr, &hash_lock);
3038 if (exists != NULL) {
3039 /* somebody beat us to the hash insert */
3040 mutex_exit(hash_lock);
3041 buf_discard_identity(hdr);
3042 (void) arc_buf_remove_ref(buf, private);
3043 goto top; /* restart the IO request */
3046 /* if this is a prefetch, we don't have a reference */
3047 if (*arc_flags & ARC_FLAG_PREFETCH) {
3048 (void) remove_reference(hdr, hash_lock,
3050 hdr->b_flags |= ARC_FLAG_PREFETCH;
3052 if (*arc_flags & ARC_FLAG_L2CACHE)
3053 hdr->b_flags |= ARC_FLAG_L2CACHE;
3054 if (*arc_flags & ARC_FLAG_L2COMPRESS)
3055 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
3056 if (BP_GET_LEVEL(bp) > 0)
3057 hdr->b_flags |= ARC_FLAG_INDIRECT;
3059 /* this block is in the ghost cache */
3060 ASSERT(GHOST_STATE(hdr->b_state));
3061 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3062 ASSERT0(refcount_count(&hdr->b_refcnt));
3063 ASSERT(hdr->b_buf == NULL);
3065 /* if this is a prefetch, we don't have a reference */
3066 if (*arc_flags & ARC_FLAG_PREFETCH)
3067 hdr->b_flags |= ARC_FLAG_PREFETCH;
3069 add_reference(hdr, hash_lock, private);
3070 if (*arc_flags & ARC_FLAG_L2CACHE)
3071 hdr->b_flags |= ARC_FLAG_L2CACHE;
3072 if (*arc_flags & ARC_FLAG_L2COMPRESS)
3073 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
3074 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3077 buf->b_efunc = NULL;
3078 buf->b_private = NULL;
3081 ASSERT(hdr->b_datacnt == 0);
3083 arc_get_data_buf(buf);
3084 arc_access(hdr, hash_lock);
3087 ASSERT(!GHOST_STATE(hdr->b_state));
3089 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
3090 acb->acb_done = done;
3091 acb->acb_private = private;
3093 ASSERT(hdr->b_acb == NULL);
3095 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
3097 if (hdr->b_l2hdr != NULL &&
3098 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3099 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3100 addr = hdr->b_l2hdr->b_daddr;
3101 b_compress = hdr->b_l2hdr->b_compress;
3102 b_asize = hdr->b_l2hdr->b_asize;
3104 * Lock out device removal.
3106 if (vdev_is_dead(vd) ||
3107 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3111 if (hash_lock != NULL)
3112 mutex_exit(hash_lock);
3115 * At this point, we have a level 1 cache miss. Try again in
3116 * L2ARC if possible.
3118 ASSERT3U(hdr->b_size, ==, size);
3119 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3120 uint64_t, size, zbookmark_phys_t *, zb);
3121 ARCSTAT_BUMP(arcstat_misses);
3122 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH),
3123 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3124 data, metadata, misses);
3126 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3128 * Read from the L2ARC if the following are true:
3129 * 1. The L2ARC vdev was previously cached.
3130 * 2. This buffer still has L2ARC metadata.
3131 * 3. This buffer isn't currently writing to the L2ARC.
3132 * 4. The L2ARC entry wasn't evicted, which may
3133 * also have invalidated the vdev.
3134 * 5. This isn't prefetch and l2arc_noprefetch is set.
3136 if (hdr->b_l2hdr != NULL &&
3137 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3138 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3139 l2arc_read_callback_t *cb;
3141 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3142 ARCSTAT_BUMP(arcstat_l2_hits);
3144 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3146 cb->l2rcb_buf = buf;
3147 cb->l2rcb_spa = spa;
3150 cb->l2rcb_flags = zio_flags;
3151 cb->l2rcb_compress = b_compress;
3153 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
3154 addr + size < vd->vdev_psize -
3155 VDEV_LABEL_END_SIZE);
3158 * l2arc read. The SCL_L2ARC lock will be
3159 * released by l2arc_read_done().
3160 * Issue a null zio if the underlying buffer
3161 * was squashed to zero size by compression.
3163 if (b_compress == ZIO_COMPRESS_EMPTY) {
3164 rzio = zio_null(pio, spa, vd,
3165 l2arc_read_done, cb,
3166 zio_flags | ZIO_FLAG_DONT_CACHE |
3168 ZIO_FLAG_DONT_PROPAGATE |
3169 ZIO_FLAG_DONT_RETRY);
3171 rzio = zio_read_phys(pio, vd, addr,
3172 b_asize, buf->b_data,
3174 l2arc_read_done, cb, priority,
3175 zio_flags | ZIO_FLAG_DONT_CACHE |
3177 ZIO_FLAG_DONT_PROPAGATE |
3178 ZIO_FLAG_DONT_RETRY, B_FALSE);
3180 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3182 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
3184 if (*arc_flags & ARC_FLAG_NOWAIT) {
3189 ASSERT(*arc_flags & ARC_FLAG_WAIT);
3190 if (zio_wait(rzio) == 0)
3193 /* l2arc read error; goto zio_read() */
3195 DTRACE_PROBE1(l2arc__miss,
3196 arc_buf_hdr_t *, hdr);
3197 ARCSTAT_BUMP(arcstat_l2_misses);
3198 if (HDR_L2_WRITING(hdr))
3199 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3200 spa_config_exit(spa, SCL_L2ARC, vd);
3204 spa_config_exit(spa, SCL_L2ARC, vd);
3205 if (l2arc_ndev != 0) {
3206 DTRACE_PROBE1(l2arc__miss,
3207 arc_buf_hdr_t *, hdr);
3208 ARCSTAT_BUMP(arcstat_l2_misses);
3212 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3213 arc_read_done, buf, priority, zio_flags, zb);
3215 if (*arc_flags & ARC_FLAG_WAIT)
3216 return (zio_wait(rzio));
3218 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
3225 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3227 ASSERT(buf->b_hdr != NULL);
3228 ASSERT(buf->b_hdr->b_state != arc_anon);
3229 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3230 ASSERT(buf->b_efunc == NULL);
3231 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3233 buf->b_efunc = func;
3234 buf->b_private = private;
3238 * Notify the arc that a block was freed, and thus will never be used again.
3241 arc_freed(spa_t *spa, const blkptr_t *bp)
3244 kmutex_t *hash_lock;
3245 uint64_t guid = spa_load_guid(spa);
3247 ASSERT(!BP_IS_EMBEDDED(bp));
3249 hdr = buf_hash_find(guid, bp, &hash_lock);
3252 if (HDR_BUF_AVAILABLE(hdr)) {
3253 arc_buf_t *buf = hdr->b_buf;
3254 add_reference(hdr, hash_lock, FTAG);
3255 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
3256 mutex_exit(hash_lock);
3258 arc_release(buf, FTAG);
3259 (void) arc_buf_remove_ref(buf, FTAG);
3261 mutex_exit(hash_lock);
3267 * Clear the user eviction callback set by arc_set_callback(), first calling
3268 * it if it exists. Because the presence of a callback keeps an arc_buf cached
3269 * clearing the callback may result in the arc_buf being destroyed. However,
3270 * it will not result in the *last* arc_buf being destroyed, hence the data
3271 * will remain cached in the ARC. We make a copy of the arc buffer here so
3272 * that we can process the callback without holding any locks.
3274 * It's possible that the callback is already in the process of being cleared
3275 * by another thread. In this case we can not clear the callback.
3277 * Returns B_TRUE if the callback was successfully called and cleared.
3280 arc_clear_callback(arc_buf_t *buf)
3283 kmutex_t *hash_lock;
3284 arc_evict_func_t *efunc = buf->b_efunc;
3285 void *private = buf->b_private;
3287 mutex_enter(&buf->b_evict_lock);
3291 * We are in arc_do_user_evicts().
3293 ASSERT(buf->b_data == NULL);
3294 mutex_exit(&buf->b_evict_lock);
3296 } else if (buf->b_data == NULL) {
3298 * We are on the eviction list; process this buffer now
3299 * but let arc_do_user_evicts() do the reaping.
3301 buf->b_efunc = NULL;
3302 mutex_exit(&buf->b_evict_lock);
3303 VERIFY0(efunc(private));
3306 hash_lock = HDR_LOCK(hdr);
3307 mutex_enter(hash_lock);
3309 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3311 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3312 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3314 buf->b_efunc = NULL;
3315 buf->b_private = NULL;
3317 if (hdr->b_datacnt > 1) {
3318 mutex_exit(&buf->b_evict_lock);
3319 arc_buf_destroy(buf, FALSE, TRUE);
3321 ASSERT(buf == hdr->b_buf);
3322 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
3323 mutex_exit(&buf->b_evict_lock);
3326 mutex_exit(hash_lock);
3327 VERIFY0(efunc(private));
3332 * Release this buffer from the cache, making it an anonymous buffer. This
3333 * must be done after a read and prior to modifying the buffer contents.
3334 * If the buffer has more than one reference, we must make
3335 * a new hdr for the buffer.
3338 arc_release(arc_buf_t *buf, void *tag)
3341 kmutex_t *hash_lock = NULL;
3342 l2arc_buf_hdr_t *l2hdr;
3346 * It would be nice to assert that if it's DMU metadata (level >
3347 * 0 || it's the dnode file), then it must be syncing context.
3348 * But we don't know that information at this level.
3351 mutex_enter(&buf->b_evict_lock);
3354 /* this buffer is not on any list */
3355 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3357 if (hdr->b_state == arc_anon) {
3358 /* this buffer is already released */
3359 ASSERT(buf->b_efunc == NULL);
3361 hash_lock = HDR_LOCK(hdr);
3362 mutex_enter(hash_lock);
3364 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3367 l2hdr = hdr->b_l2hdr;
3369 mutex_enter(&l2arc_buflist_mtx);
3370 hdr->b_l2hdr = NULL;
3371 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3373 buf_size = hdr->b_size;
3376 * Do we have more than one buf?
3378 if (hdr->b_datacnt > 1) {
3379 arc_buf_hdr_t *nhdr;
3381 uint64_t blksz = hdr->b_size;
3382 uint64_t spa = hdr->b_spa;
3383 arc_buf_contents_t type = hdr->b_type;
3384 uint32_t flags = hdr->b_flags;
3386 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3388 * Pull the data off of this hdr and attach it to
3389 * a new anonymous hdr.
3391 (void) remove_reference(hdr, hash_lock, tag);
3393 while (*bufp != buf)
3394 bufp = &(*bufp)->b_next;
3395 *bufp = buf->b_next;
3398 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3399 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3400 if (refcount_is_zero(&hdr->b_refcnt)) {
3401 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3402 ASSERT3U(*size, >=, hdr->b_size);
3403 atomic_add_64(size, -hdr->b_size);
3407 * We're releasing a duplicate user data buffer, update
3408 * our statistics accordingly.
3410 if (hdr->b_type == ARC_BUFC_DATA) {
3411 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3412 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3415 hdr->b_datacnt -= 1;
3416 arc_cksum_verify(buf);
3417 arc_buf_unwatch(buf);
3419 mutex_exit(hash_lock);
3421 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3422 nhdr->b_size = blksz;
3424 nhdr->b_type = type;
3426 nhdr->b_state = arc_anon;
3427 nhdr->b_arc_access = 0;
3428 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING;
3429 nhdr->b_l2hdr = NULL;
3430 nhdr->b_datacnt = 1;
3431 nhdr->b_freeze_cksum = NULL;
3432 (void) refcount_add(&nhdr->b_refcnt, tag);
3434 mutex_exit(&buf->b_evict_lock);
3435 atomic_add_64(&arc_anon->arcs_size, blksz);
3437 mutex_exit(&buf->b_evict_lock);
3438 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3439 ASSERT(!list_link_active(&hdr->b_arc_node));
3440 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3441 if (hdr->b_state != arc_anon)
3442 arc_change_state(arc_anon, hdr, hash_lock);
3443 hdr->b_arc_access = 0;
3445 mutex_exit(hash_lock);
3447 buf_discard_identity(hdr);
3450 buf->b_efunc = NULL;
3451 buf->b_private = NULL;
3454 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
3455 vdev_space_update(l2hdr->b_dev->l2ad_vdev,
3456 -l2hdr->b_asize, 0, 0);
3457 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3458 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3459 mutex_exit(&l2arc_buflist_mtx);
3464 arc_released(arc_buf_t *buf)
3468 mutex_enter(&buf->b_evict_lock);
3469 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3470 mutex_exit(&buf->b_evict_lock);
3476 arc_referenced(arc_buf_t *buf)
3480 mutex_enter(&buf->b_evict_lock);
3481 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3482 mutex_exit(&buf->b_evict_lock);
3483 return (referenced);
3488 arc_write_ready(zio_t *zio)
3490 arc_write_callback_t *callback = zio->io_private;
3491 arc_buf_t *buf = callback->awcb_buf;
3492 arc_buf_hdr_t *hdr = buf->b_hdr;
3494 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3495 callback->awcb_ready(zio, buf, callback->awcb_private);
3498 * If the IO is already in progress, then this is a re-write
3499 * attempt, so we need to thaw and re-compute the cksum.
3500 * It is the responsibility of the callback to handle the
3501 * accounting for any re-write attempt.
3503 if (HDR_IO_IN_PROGRESS(hdr)) {
3504 mutex_enter(&hdr->b_freeze_lock);
3505 if (hdr->b_freeze_cksum != NULL) {
3506 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3507 hdr->b_freeze_cksum = NULL;
3509 mutex_exit(&hdr->b_freeze_lock);
3511 arc_cksum_compute(buf, B_FALSE);
3512 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
3516 * The SPA calls this callback for each physical write that happens on behalf
3517 * of a logical write. See the comment in dbuf_write_physdone() for details.
3520 arc_write_physdone(zio_t *zio)
3522 arc_write_callback_t *cb = zio->io_private;
3523 if (cb->awcb_physdone != NULL)
3524 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
3528 arc_write_done(zio_t *zio)
3530 arc_write_callback_t *callback = zio->io_private;
3531 arc_buf_t *buf = callback->awcb_buf;
3532 arc_buf_hdr_t *hdr = buf->b_hdr;
3534 ASSERT(hdr->b_acb == NULL);
3536 if (zio->io_error == 0) {
3537 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
3538 buf_discard_identity(hdr);
3540 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3541 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3542 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3545 ASSERT(BUF_EMPTY(hdr));
3549 * If the block to be written was all-zero or compressed enough to be
3550 * embedded in the BP, no write was performed so there will be no
3551 * dva/birth/checksum. The buffer must therefore remain anonymous
3554 if (!BUF_EMPTY(hdr)) {
3555 arc_buf_hdr_t *exists;
3556 kmutex_t *hash_lock;
3558 ASSERT(zio->io_error == 0);
3560 arc_cksum_verify(buf);
3562 exists = buf_hash_insert(hdr, &hash_lock);
3565 * This can only happen if we overwrite for
3566 * sync-to-convergence, because we remove
3567 * buffers from the hash table when we arc_free().
3569 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3570 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3571 panic("bad overwrite, hdr=%p exists=%p",
3572 (void *)hdr, (void *)exists);
3573 ASSERT(refcount_is_zero(&exists->b_refcnt));
3574 arc_change_state(arc_anon, exists, hash_lock);
3575 mutex_exit(hash_lock);
3576 arc_hdr_destroy(exists);
3577 exists = buf_hash_insert(hdr, &hash_lock);
3578 ASSERT3P(exists, ==, NULL);
3579 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3581 ASSERT(zio->io_prop.zp_nopwrite);
3582 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3583 panic("bad nopwrite, hdr=%p exists=%p",
3584 (void *)hdr, (void *)exists);
3587 ASSERT(hdr->b_datacnt == 1);
3588 ASSERT(hdr->b_state == arc_anon);
3589 ASSERT(BP_GET_DEDUP(zio->io_bp));
3590 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3593 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
3594 /* if it's not anon, we are doing a scrub */
3595 if (!exists && hdr->b_state == arc_anon)
3596 arc_access(hdr, hash_lock);
3597 mutex_exit(hash_lock);
3599 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
3602 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3603 callback->awcb_done(zio, buf, callback->awcb_private);
3605 kmem_free(callback, sizeof (arc_write_callback_t));
3609 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3610 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
3611 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
3612 arc_done_func_t *done, void *private, zio_priority_t priority,
3613 int zio_flags, const zbookmark_phys_t *zb)
3615 arc_buf_hdr_t *hdr = buf->b_hdr;
3616 arc_write_callback_t *callback;
3619 ASSERT(ready != NULL);
3620 ASSERT(done != NULL);
3621 ASSERT(!HDR_IO_ERROR(hdr));
3622 ASSERT((hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) == 0);
3623 ASSERT(hdr->b_acb == NULL);
3625 hdr->b_flags |= ARC_FLAG_L2CACHE;
3627 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
3628 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3629 callback->awcb_ready = ready;
3630 callback->awcb_physdone = physdone;
3631 callback->awcb_done = done;
3632 callback->awcb_private = private;
3633 callback->awcb_buf = buf;
3635 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3636 arc_write_ready, arc_write_physdone, arc_write_done, callback,
3637 priority, zio_flags, zb);
3643 arc_memory_throttle(uint64_t reserve, uint64_t txg)
3646 uint64_t available_memory = ptob(freemem);
3647 static uint64_t page_load = 0;
3648 static uint64_t last_txg = 0;
3652 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3655 if (freemem > physmem * arc_lotsfree_percent / 100)
3658 if (txg > last_txg) {
3663 * If we are in pageout, we know that memory is already tight,
3664 * the arc is already going to be evicting, so we just want to
3665 * continue to let page writes occur as quickly as possible.
3667 if (curproc == proc_pageout) {
3668 if (page_load > MAX(ptob(minfree), available_memory) / 4)
3669 return (SET_ERROR(ERESTART));
3670 /* Note: reserve is inflated, so we deflate */
3671 page_load += reserve / 8;
3673 } else if (page_load > 0 && arc_reclaim_needed()) {
3674 /* memory is low, delay before restarting */
3675 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3676 return (SET_ERROR(EAGAIN));
3684 arc_tempreserve_clear(uint64_t reserve)
3686 atomic_add_64(&arc_tempreserve, -reserve);
3687 ASSERT((int64_t)arc_tempreserve >= 0);
3691 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3696 if (reserve > arc_c/4 && !arc_no_grow)
3697 arc_c = MIN(arc_c_max, reserve * 4);
3698 if (reserve > arc_c)
3699 return (SET_ERROR(ENOMEM));
3702 * Don't count loaned bufs as in flight dirty data to prevent long
3703 * network delays from blocking transactions that are ready to be
3704 * assigned to a txg.
3706 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3709 * Writes will, almost always, require additional memory allocations
3710 * in order to compress/encrypt/etc the data. We therefore need to
3711 * make sure that there is sufficient available memory for this.
3713 error = arc_memory_throttle(reserve, txg);
3718 * Throttle writes when the amount of dirty data in the cache
3719 * gets too large. We try to keep the cache less than half full
3720 * of dirty blocks so that our sync times don't grow too large.
3721 * Note: if two requests come in concurrently, we might let them
3722 * both succeed, when one of them should fail. Not a huge deal.
3725 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3726 anon_size > arc_c / 4) {
3727 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3728 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3729 arc_tempreserve>>10,
3730 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3731 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3732 reserve>>10, arc_c>>10);
3733 return (SET_ERROR(ERESTART));
3735 atomic_add_64(&arc_tempreserve, reserve);
3742 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3743 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3745 /* Convert seconds to clock ticks */
3746 arc_min_prefetch_lifespan = 1 * hz;
3748 /* Start out with 1/8 of all memory */
3749 arc_c = physmem * PAGESIZE / 8;
3753 * On architectures where the physical memory can be larger
3754 * than the addressable space (intel in 32-bit mode), we may
3755 * need to limit the cache to 1/8 of VM size.
3757 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3760 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3761 arc_c_min = MAX(arc_c / 4, 64<<20);
3762 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3763 if (arc_c * 8 >= 1<<30)
3764 arc_c_max = (arc_c * 8) - (1<<30);
3766 arc_c_max = arc_c_min;
3767 arc_c_max = MAX(arc_c * 6, arc_c_max);
3770 * Allow the tunables to override our calculations if they are
3771 * reasonable (ie. over 64MB)
3773 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3774 arc_c_max = zfs_arc_max;
3775 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3776 arc_c_min = zfs_arc_min;
3779 arc_p = (arc_c >> 1);
3781 /* limit meta-data to 1/4 of the arc capacity */
3782 arc_meta_limit = arc_c_max / 4;
3784 /* Allow the tunable to override if it is reasonable */
3785 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3786 arc_meta_limit = zfs_arc_meta_limit;
3788 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3789 arc_c_min = arc_meta_limit / 2;
3791 if (zfs_arc_meta_min > 0) {
3792 arc_meta_min = zfs_arc_meta_min;
3794 arc_meta_min = arc_c_min / 2;
3797 if (zfs_arc_grow_retry > 0)
3798 arc_grow_retry = zfs_arc_grow_retry;
3800 if (zfs_arc_shrink_shift > 0)
3801 arc_shrink_shift = zfs_arc_shrink_shift;
3803 if (zfs_arc_p_min_shift > 0)
3804 arc_p_min_shift = zfs_arc_p_min_shift;
3806 /* if kmem_flags are set, lets try to use less memory */
3807 if (kmem_debugging())
3809 if (arc_c < arc_c_min)
3812 arc_anon = &ARC_anon;
3814 arc_mru_ghost = &ARC_mru_ghost;
3816 arc_mfu_ghost = &ARC_mfu_ghost;
3817 arc_l2c_only = &ARC_l2c_only;
3820 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3821 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3822 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3823 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3824 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3825 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3827 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3828 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3829 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3830 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3831 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3832 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3833 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3834 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3835 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3836 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3837 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3838 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3839 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3840 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3841 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3842 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3843 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3844 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3845 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3846 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3850 arc_thread_exit = 0;
3851 arc_eviction_list = NULL;
3852 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3853 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3855 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3856 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3858 if (arc_ksp != NULL) {
3859 arc_ksp->ks_data = &arc_stats;
3860 kstat_install(arc_ksp);
3863 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3864 TS_RUN, minclsyspri);
3870 * Calculate maximum amount of dirty data per pool.
3872 * If it has been set by /etc/system, take that.
3873 * Otherwise, use a percentage of physical memory defined by
3874 * zfs_dirty_data_max_percent (default 10%) with a cap at
3875 * zfs_dirty_data_max_max (default 4GB).
3877 if (zfs_dirty_data_max == 0) {
3878 zfs_dirty_data_max = physmem * PAGESIZE *
3879 zfs_dirty_data_max_percent / 100;
3880 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
3881 zfs_dirty_data_max_max);
3888 mutex_enter(&arc_reclaim_thr_lock);
3889 arc_thread_exit = 1;
3890 while (arc_thread_exit != 0)
3891 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3892 mutex_exit(&arc_reclaim_thr_lock);
3898 if (arc_ksp != NULL) {
3899 kstat_delete(arc_ksp);
3903 mutex_destroy(&arc_eviction_mtx);
3904 mutex_destroy(&arc_reclaim_thr_lock);
3905 cv_destroy(&arc_reclaim_thr_cv);
3907 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3908 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3909 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3910 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3911 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3912 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3913 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3914 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3916 mutex_destroy(&arc_anon->arcs_mtx);
3917 mutex_destroy(&arc_mru->arcs_mtx);
3918 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3919 mutex_destroy(&arc_mfu->arcs_mtx);
3920 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3921 mutex_destroy(&arc_l2c_only->arcs_mtx);
3925 ASSERT(arc_loaned_bytes == 0);
3931 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3932 * It uses dedicated storage devices to hold cached data, which are populated
3933 * using large infrequent writes. The main role of this cache is to boost
3934 * the performance of random read workloads. The intended L2ARC devices
3935 * include short-stroked disks, solid state disks, and other media with
3936 * substantially faster read latency than disk.
3938 * +-----------------------+
3940 * +-----------------------+
3943 * l2arc_feed_thread() arc_read()
3947 * +---------------+ |
3949 * +---------------+ |
3954 * +-------+ +-------+
3956 * | cache | | cache |
3957 * +-------+ +-------+
3958 * +=========+ .-----.
3959 * : L2ARC : |-_____-|
3960 * : devices : | Disks |
3961 * +=========+ `-_____-'
3963 * Read requests are satisfied from the following sources, in order:
3966 * 2) vdev cache of L2ARC devices
3968 * 4) vdev cache of disks
3971 * Some L2ARC device types exhibit extremely slow write performance.
3972 * To accommodate for this there are some significant differences between
3973 * the L2ARC and traditional cache design:
3975 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3976 * the ARC behave as usual, freeing buffers and placing headers on ghost
3977 * lists. The ARC does not send buffers to the L2ARC during eviction as
3978 * this would add inflated write latencies for all ARC memory pressure.
3980 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3981 * It does this by periodically scanning buffers from the eviction-end of
3982 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3983 * not already there. It scans until a headroom of buffers is satisfied,
3984 * which itself is a buffer for ARC eviction. If a compressible buffer is
3985 * found during scanning and selected for writing to an L2ARC device, we
3986 * temporarily boost scanning headroom during the next scan cycle to make
3987 * sure we adapt to compression effects (which might significantly reduce
3988 * the data volume we write to L2ARC). The thread that does this is
3989 * l2arc_feed_thread(), illustrated below; example sizes are included to
3990 * provide a better sense of ratio than this diagram:
3993 * +---------------------+----------+
3994 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3995 * +---------------------+----------+ | o L2ARC eligible
3996 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
3997 * +---------------------+----------+ |
3998 * 15.9 Gbytes ^ 32 Mbytes |
4000 * l2arc_feed_thread()
4002 * l2arc write hand <--[oooo]--'
4006 * +==============================+
4007 * L2ARC dev |####|#|###|###| |####| ... |
4008 * +==============================+
4011 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4012 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4013 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4014 * safe to say that this is an uncommon case, since buffers at the end of
4015 * the ARC lists have moved there due to inactivity.
4017 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4018 * then the L2ARC simply misses copying some buffers. This serves as a
4019 * pressure valve to prevent heavy read workloads from both stalling the ARC
4020 * with waits and clogging the L2ARC with writes. This also helps prevent
4021 * the potential for the L2ARC to churn if it attempts to cache content too
4022 * quickly, such as during backups of the entire pool.
4024 * 5. After system boot and before the ARC has filled main memory, there are
4025 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4026 * lists can remain mostly static. Instead of searching from tail of these
4027 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4028 * for eligible buffers, greatly increasing its chance of finding them.
4030 * The L2ARC device write speed is also boosted during this time so that
4031 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4032 * there are no L2ARC reads, and no fear of degrading read performance
4033 * through increased writes.
4035 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4036 * the vdev queue can aggregate them into larger and fewer writes. Each
4037 * device is written to in a rotor fashion, sweeping writes through
4038 * available space then repeating.
4040 * 7. The L2ARC does not store dirty content. It never needs to flush
4041 * write buffers back to disk based storage.
4043 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4044 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4046 * The performance of the L2ARC can be tweaked by a number of tunables, which
4047 * may be necessary for different workloads:
4049 * l2arc_write_max max write bytes per interval
4050 * l2arc_write_boost extra write bytes during device warmup
4051 * l2arc_noprefetch skip caching prefetched buffers
4052 * l2arc_headroom number of max device writes to precache
4053 * l2arc_headroom_boost when we find compressed buffers during ARC
4054 * scanning, we multiply headroom by this
4055 * percentage factor for the next scan cycle,
4056 * since more compressed buffers are likely to
4058 * l2arc_feed_secs seconds between L2ARC writing
4060 * Tunables may be removed or added as future performance improvements are
4061 * integrated, and also may become zpool properties.
4063 * There are three key functions that control how the L2ARC warms up:
4065 * l2arc_write_eligible() check if a buffer is eligible to cache
4066 * l2arc_write_size() calculate how much to write
4067 * l2arc_write_interval() calculate sleep delay between writes
4069 * These three functions determine what to write, how much, and how quickly
4074 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
4077 * A buffer is *not* eligible for the L2ARC if it:
4078 * 1. belongs to a different spa.
4079 * 2. is already cached on the L2ARC.
4080 * 3. has an I/O in progress (it may be an incomplete read).
4081 * 4. is flagged not eligible (zfs property).
4083 if (hdr->b_spa != spa_guid || hdr->b_l2hdr != NULL ||
4084 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
4091 l2arc_write_size(void)
4096 * Make sure our globals have meaningful values in case the user
4099 size = l2arc_write_max;
4101 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
4102 "be greater than zero, resetting it to the default (%d)",
4104 size = l2arc_write_max = L2ARC_WRITE_SIZE;
4107 if (arc_warm == B_FALSE)
4108 size += l2arc_write_boost;
4115 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4117 clock_t interval, next, now;
4120 * If the ARC lists are busy, increase our write rate; if the
4121 * lists are stale, idle back. This is achieved by checking
4122 * how much we previously wrote - if it was more than half of
4123 * what we wanted, schedule the next write much sooner.
4125 if (l2arc_feed_again && wrote > (wanted / 2))
4126 interval = (hz * l2arc_feed_min_ms) / 1000;
4128 interval = hz * l2arc_feed_secs;
4130 now = ddi_get_lbolt();
4131 next = MAX(now, MIN(now + interval, began + interval));
4137 l2arc_hdr_stat_add(void)
4139 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4140 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4144 l2arc_hdr_stat_remove(void)
4146 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4147 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4151 * Cycle through L2ARC devices. This is how L2ARC load balances.
4152 * If a device is returned, this also returns holding the spa config lock.
4154 static l2arc_dev_t *
4155 l2arc_dev_get_next(void)
4157 l2arc_dev_t *first, *next = NULL;
4160 * Lock out the removal of spas (spa_namespace_lock), then removal
4161 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4162 * both locks will be dropped and a spa config lock held instead.
4164 mutex_enter(&spa_namespace_lock);
4165 mutex_enter(&l2arc_dev_mtx);
4167 /* if there are no vdevs, there is nothing to do */
4168 if (l2arc_ndev == 0)
4172 next = l2arc_dev_last;
4174 /* loop around the list looking for a non-faulted vdev */
4176 next = list_head(l2arc_dev_list);
4178 next = list_next(l2arc_dev_list, next);
4180 next = list_head(l2arc_dev_list);
4183 /* if we have come back to the start, bail out */
4186 else if (next == first)
4189 } while (vdev_is_dead(next->l2ad_vdev));
4191 /* if we were unable to find any usable vdevs, return NULL */
4192 if (vdev_is_dead(next->l2ad_vdev))
4195 l2arc_dev_last = next;
4198 mutex_exit(&l2arc_dev_mtx);
4201 * Grab the config lock to prevent the 'next' device from being
4202 * removed while we are writing to it.
4205 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4206 mutex_exit(&spa_namespace_lock);
4212 * Free buffers that were tagged for destruction.
4215 l2arc_do_free_on_write()
4218 l2arc_data_free_t *df, *df_prev;
4220 mutex_enter(&l2arc_free_on_write_mtx);
4221 buflist = l2arc_free_on_write;
4223 for (df = list_tail(buflist); df; df = df_prev) {
4224 df_prev = list_prev(buflist, df);
4225 ASSERT(df->l2df_data != NULL);
4226 ASSERT(df->l2df_func != NULL);
4227 df->l2df_func(df->l2df_data, df->l2df_size);
4228 list_remove(buflist, df);
4229 kmem_free(df, sizeof (l2arc_data_free_t));
4232 mutex_exit(&l2arc_free_on_write_mtx);
4236 * A write to a cache device has completed. Update all headers to allow
4237 * reads from these buffers to begin.
4240 l2arc_write_done(zio_t *zio)
4242 l2arc_write_callback_t *cb;
4245 arc_buf_hdr_t *head, *hdr, *hdr_prev;
4246 l2arc_buf_hdr_t *abl2;
4247 kmutex_t *hash_lock;
4248 int64_t bytes_dropped = 0;
4250 cb = zio->io_private;
4252 dev = cb->l2wcb_dev;
4253 ASSERT(dev != NULL);
4254 head = cb->l2wcb_head;
4255 ASSERT(head != NULL);
4256 buflist = dev->l2ad_buflist;
4257 ASSERT(buflist != NULL);
4258 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4259 l2arc_write_callback_t *, cb);
4261 if (zio->io_error != 0)
4262 ARCSTAT_BUMP(arcstat_l2_writes_error);
4264 mutex_enter(&l2arc_buflist_mtx);
4267 * All writes completed, or an error was hit.
4269 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
4270 hdr_prev = list_prev(buflist, hdr);
4271 abl2 = hdr->b_l2hdr;
4274 * Release the temporary compressed buffer as soon as possible.
4276 if (abl2->b_compress != ZIO_COMPRESS_OFF)
4277 l2arc_release_cdata_buf(hdr);
4279 hash_lock = HDR_LOCK(hdr);
4280 if (!mutex_tryenter(hash_lock)) {
4282 * This buffer misses out. It may be in a stage
4283 * of eviction. Its ARC_L2_WRITING flag will be
4284 * left set, denying reads to this buffer.
4286 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4290 if (zio->io_error != 0) {
4292 * Error - drop L2ARC entry.
4294 list_remove(buflist, hdr);
4295 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4296 bytes_dropped += abl2->b_asize;
4297 hdr->b_l2hdr = NULL;
4298 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4299 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
4303 * Allow ARC to begin reads to this L2ARC entry.
4305 hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
4307 mutex_exit(hash_lock);
4310 atomic_inc_64(&l2arc_writes_done);
4311 list_remove(buflist, head);
4312 kmem_cache_free(hdr_cache, head);
4313 mutex_exit(&l2arc_buflist_mtx);
4315 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
4317 l2arc_do_free_on_write();
4319 kmem_free(cb, sizeof (l2arc_write_callback_t));
4323 * A read to a cache device completed. Validate buffer contents before
4324 * handing over to the regular ARC routines.
4327 l2arc_read_done(zio_t *zio)
4329 l2arc_read_callback_t *cb;
4332 kmutex_t *hash_lock;
4335 ASSERT(zio->io_vd != NULL);
4336 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4338 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4340 cb = zio->io_private;
4342 buf = cb->l2rcb_buf;
4343 ASSERT(buf != NULL);
4345 hash_lock = HDR_LOCK(buf->b_hdr);
4346 mutex_enter(hash_lock);
4348 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4351 * If the buffer was compressed, decompress it first.
4353 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
4354 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
4355 ASSERT(zio->io_data != NULL);
4358 * Check this survived the L2ARC journey.
4360 equal = arc_cksum_equal(buf);
4361 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4362 mutex_exit(hash_lock);
4363 zio->io_private = buf;
4364 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4365 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4368 mutex_exit(hash_lock);
4370 * Buffer didn't survive caching. Increment stats and
4371 * reissue to the original storage device.
4373 if (zio->io_error != 0) {
4374 ARCSTAT_BUMP(arcstat_l2_io_error);
4376 zio->io_error = SET_ERROR(EIO);
4379 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4382 * If there's no waiter, issue an async i/o to the primary
4383 * storage now. If there *is* a waiter, the caller must
4384 * issue the i/o in a context where it's OK to block.
4386 if (zio->io_waiter == NULL) {
4387 zio_t *pio = zio_unique_parent(zio);
4389 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4391 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4392 buf->b_data, zio->io_size, arc_read_done, buf,
4393 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4397 kmem_free(cb, sizeof (l2arc_read_callback_t));
4401 * This is the list priority from which the L2ARC will search for pages to
4402 * cache. This is used within loops (0..3) to cycle through lists in the
4403 * desired order. This order can have a significant effect on cache
4406 * Currently the metadata lists are hit first, MFU then MRU, followed by
4407 * the data lists. This function returns a locked list, and also returns
4411 l2arc_list_locked(int list_num, kmutex_t **lock)
4413 list_t *list = NULL;
4415 ASSERT(list_num >= 0 && list_num <= 3);
4419 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4420 *lock = &arc_mfu->arcs_mtx;
4423 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4424 *lock = &arc_mru->arcs_mtx;
4427 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4428 *lock = &arc_mfu->arcs_mtx;
4431 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4432 *lock = &arc_mru->arcs_mtx;
4436 ASSERT(!(MUTEX_HELD(*lock)));
4442 * Evict buffers from the device write hand to the distance specified in
4443 * bytes. This distance may span populated buffers, it may span nothing.
4444 * This is clearing a region on the L2ARC device ready for writing.
4445 * If the 'all' boolean is set, every buffer is evicted.
4448 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4451 l2arc_buf_hdr_t *abl2;
4452 arc_buf_hdr_t *hdr, *hdr_prev;
4453 kmutex_t *hash_lock;
4455 int64_t bytes_evicted = 0;
4457 buflist = dev->l2ad_buflist;
4459 if (buflist == NULL)
4462 if (!all && dev->l2ad_first) {
4464 * This is the first sweep through the device. There is
4470 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4472 * When nearing the end of the device, evict to the end
4473 * before the device write hand jumps to the start.
4475 taddr = dev->l2ad_end;
4477 taddr = dev->l2ad_hand + distance;
4479 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4480 uint64_t, taddr, boolean_t, all);
4483 mutex_enter(&l2arc_buflist_mtx);
4484 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
4485 hdr_prev = list_prev(buflist, hdr);
4487 hash_lock = HDR_LOCK(hdr);
4488 if (!mutex_tryenter(hash_lock)) {
4490 * Missed the hash lock. Retry.
4492 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4493 mutex_exit(&l2arc_buflist_mtx);
4494 mutex_enter(hash_lock);
4495 mutex_exit(hash_lock);
4499 if (HDR_L2_WRITE_HEAD(hdr)) {
4501 * We hit a write head node. Leave it for
4502 * l2arc_write_done().
4504 list_remove(buflist, hdr);
4505 mutex_exit(hash_lock);
4509 if (!all && hdr->b_l2hdr != NULL &&
4510 (hdr->b_l2hdr->b_daddr > taddr ||
4511 hdr->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4513 * We've evicted to the target address,
4514 * or the end of the device.
4516 mutex_exit(hash_lock);
4520 if (HDR_FREE_IN_PROGRESS(hdr)) {
4522 * Already on the path to destruction.
4524 mutex_exit(hash_lock);
4528 if (hdr->b_state == arc_l2c_only) {
4529 ASSERT(!HDR_L2_READING(hdr));
4531 * This doesn't exist in the ARC. Destroy.
4532 * arc_hdr_destroy() will call list_remove()
4533 * and decrement arcstat_l2_size.
4535 arc_change_state(arc_anon, hdr, hash_lock);
4536 arc_hdr_destroy(hdr);
4539 * Invalidate issued or about to be issued
4540 * reads, since we may be about to write
4541 * over this location.
4543 if (HDR_L2_READING(hdr)) {
4544 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4545 hdr->b_flags |= ARC_FLAG_L2_EVICTED;
4549 * Tell ARC this no longer exists in L2ARC.
4551 if (hdr->b_l2hdr != NULL) {
4552 abl2 = hdr->b_l2hdr;
4553 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4554 bytes_evicted += abl2->b_asize;
4555 hdr->b_l2hdr = NULL;
4556 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4557 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
4559 list_remove(buflist, hdr);
4562 * This may have been leftover after a
4565 hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
4567 mutex_exit(hash_lock);
4569 mutex_exit(&l2arc_buflist_mtx);
4571 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0);
4572 dev->l2ad_evict = taddr;
4576 * Find and write ARC buffers to the L2ARC device.
4578 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
4579 * for reading until they have completed writing.
4580 * The headroom_boost is an in-out parameter used to maintain headroom boost
4581 * state between calls to this function.
4583 * Returns the number of bytes actually written (which may be smaller than
4584 * the delta by which the device hand has changed due to alignment).
4587 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
4588 boolean_t *headroom_boost)
4590 arc_buf_hdr_t *hdr, *hdr_prev, *head;
4592 uint64_t write_asize, write_psize, write_sz, headroom,
4595 kmutex_t *list_lock;
4597 l2arc_write_callback_t *cb;
4599 uint64_t guid = spa_load_guid(spa);
4600 const boolean_t do_headroom_boost = *headroom_boost;
4602 ASSERT(dev->l2ad_vdev != NULL);
4604 /* Lower the flag now, we might want to raise it again later. */
4605 *headroom_boost = B_FALSE;
4608 write_sz = write_asize = write_psize = 0;
4610 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4611 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD;
4614 * We will want to try to compress buffers that are at least 2x the
4615 * device sector size.
4617 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
4620 * Copy buffers for L2ARC writing.
4622 mutex_enter(&l2arc_buflist_mtx);
4623 for (int try = 0; try <= 3; try++) {
4624 uint64_t passed_sz = 0;
4626 list = l2arc_list_locked(try, &list_lock);
4629 * L2ARC fast warmup.
4631 * Until the ARC is warm and starts to evict, read from the
4632 * head of the ARC lists rather than the tail.
4634 if (arc_warm == B_FALSE)
4635 hdr = list_head(list);
4637 hdr = list_tail(list);
4639 headroom = target_sz * l2arc_headroom;
4640 if (do_headroom_boost)
4641 headroom = (headroom * l2arc_headroom_boost) / 100;
4643 for (; hdr; hdr = hdr_prev) {
4644 l2arc_buf_hdr_t *l2hdr;
4645 kmutex_t *hash_lock;
4648 if (arc_warm == B_FALSE)
4649 hdr_prev = list_next(list, hdr);
4651 hdr_prev = list_prev(list, hdr);
4653 hash_lock = HDR_LOCK(hdr);
4654 if (!mutex_tryenter(hash_lock)) {
4656 * Skip this buffer rather than waiting.
4661 passed_sz += hdr->b_size;
4662 if (passed_sz > headroom) {
4666 mutex_exit(hash_lock);
4670 if (!l2arc_write_eligible(guid, hdr)) {
4671 mutex_exit(hash_lock);
4675 if ((write_sz + hdr->b_size) > target_sz) {
4677 mutex_exit(hash_lock);
4683 * Insert a dummy header on the buflist so
4684 * l2arc_write_done() can find where the
4685 * write buffers begin without searching.
4687 list_insert_head(dev->l2ad_buflist, head);
4690 sizeof (l2arc_write_callback_t), KM_SLEEP);
4691 cb->l2wcb_dev = dev;
4692 cb->l2wcb_head = head;
4693 pio = zio_root(spa, l2arc_write_done, cb,
4698 * Create and add a new L2ARC header.
4700 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4702 hdr->b_flags |= ARC_FLAG_L2_WRITING;
4705 * Temporarily stash the data buffer in b_tmp_cdata.
4706 * The subsequent write step will pick it up from
4707 * there. This is because can't access hdr->b_buf
4708 * without holding the hash_lock, which we in turn
4709 * can't access without holding the ARC list locks
4710 * (which we want to avoid during compression/writing).
4712 l2hdr->b_compress = ZIO_COMPRESS_OFF;
4713 l2hdr->b_asize = hdr->b_size;
4714 l2hdr->b_tmp_cdata = hdr->b_buf->b_data;
4716 buf_sz = hdr->b_size;
4717 hdr->b_l2hdr = l2hdr;
4719 list_insert_head(dev->l2ad_buflist, hdr);
4722 * Compute and store the buffer cksum before
4723 * writing. On debug the cksum is verified first.
4725 arc_cksum_verify(hdr->b_buf);
4726 arc_cksum_compute(hdr->b_buf, B_TRUE);
4728 mutex_exit(hash_lock);
4733 mutex_exit(list_lock);
4739 /* No buffers selected for writing? */
4742 mutex_exit(&l2arc_buflist_mtx);
4743 kmem_cache_free(hdr_cache, head);
4748 * Now start writing the buffers. We're starting at the write head
4749 * and work backwards, retracing the course of the buffer selector
4752 for (hdr = list_prev(dev->l2ad_buflist, head); hdr;
4753 hdr = list_prev(dev->l2ad_buflist, hdr)) {
4754 l2arc_buf_hdr_t *l2hdr;
4758 * We shouldn't need to lock the buffer here, since we flagged
4759 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
4760 * take care to only access its L2 cache parameters. In
4761 * particular, hdr->b_buf may be invalid by now due to
4764 l2hdr = hdr->b_l2hdr;
4765 l2hdr->b_daddr = dev->l2ad_hand;
4767 if ((hdr->b_flags & ARC_FLAG_L2COMPRESS) &&
4768 l2hdr->b_asize >= buf_compress_minsz) {
4769 if (l2arc_compress_buf(l2hdr)) {
4771 * If compression succeeded, enable headroom
4772 * boost on the next scan cycle.
4774 *headroom_boost = B_TRUE;
4779 * Pick up the buffer data we had previously stashed away
4780 * (and now potentially also compressed).
4782 buf_data = l2hdr->b_tmp_cdata;
4783 buf_sz = l2hdr->b_asize;
4785 /* Compression may have squashed the buffer to zero length. */
4789 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4790 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4791 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4792 ZIO_FLAG_CANFAIL, B_FALSE);
4794 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4796 (void) zio_nowait(wzio);
4798 write_asize += buf_sz;
4800 * Keep the clock hand suitably device-aligned.
4802 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4803 write_psize += buf_p_sz;
4804 dev->l2ad_hand += buf_p_sz;
4808 mutex_exit(&l2arc_buflist_mtx);
4810 ASSERT3U(write_asize, <=, target_sz);
4811 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4812 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
4813 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4814 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
4815 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
4818 * Bump device hand to the device start if it is approaching the end.
4819 * l2arc_evict() will already have evicted ahead for this case.
4821 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4822 dev->l2ad_hand = dev->l2ad_start;
4823 dev->l2ad_evict = dev->l2ad_start;
4824 dev->l2ad_first = B_FALSE;
4827 dev->l2ad_writing = B_TRUE;
4828 (void) zio_wait(pio);
4829 dev->l2ad_writing = B_FALSE;
4831 return (write_asize);
4835 * Compresses an L2ARC buffer.
4836 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its
4837 * size in l2hdr->b_asize. This routine tries to compress the data and
4838 * depending on the compression result there are three possible outcomes:
4839 * *) The buffer was incompressible. The original l2hdr contents were left
4840 * untouched and are ready for writing to an L2 device.
4841 * *) The buffer was all-zeros, so there is no need to write it to an L2
4842 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
4843 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
4844 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
4845 * data buffer which holds the compressed data to be written, and b_asize
4846 * tells us how much data there is. b_compress is set to the appropriate
4847 * compression algorithm. Once writing is done, invoke
4848 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
4850 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
4851 * buffer was incompressible).
4854 l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr)
4857 size_t csize, len, rounded;
4859 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
4860 ASSERT(l2hdr->b_tmp_cdata != NULL);
4862 len = l2hdr->b_asize;
4863 cdata = zio_data_buf_alloc(len);
4864 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata,
4865 cdata, l2hdr->b_asize);
4867 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE);
4868 if (rounded > csize) {
4869 bzero((char *)cdata + csize, rounded - csize);
4874 /* zero block, indicate that there's nothing to write */
4875 zio_data_buf_free(cdata, len);
4876 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
4878 l2hdr->b_tmp_cdata = NULL;
4879 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
4881 } else if (csize > 0 && csize < len) {
4883 * Compression succeeded, we'll keep the cdata around for
4884 * writing and release it afterwards.
4886 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
4887 l2hdr->b_asize = csize;
4888 l2hdr->b_tmp_cdata = cdata;
4889 ARCSTAT_BUMP(arcstat_l2_compress_successes);
4893 * Compression failed, release the compressed buffer.
4894 * l2hdr will be left unmodified.
4896 zio_data_buf_free(cdata, len);
4897 ARCSTAT_BUMP(arcstat_l2_compress_failures);
4903 * Decompresses a zio read back from an l2arc device. On success, the
4904 * underlying zio's io_data buffer is overwritten by the uncompressed
4905 * version. On decompression error (corrupt compressed stream), the
4906 * zio->io_error value is set to signal an I/O error.
4908 * Please note that the compressed data stream is not checksummed, so
4909 * if the underlying device is experiencing data corruption, we may feed
4910 * corrupt data to the decompressor, so the decompressor needs to be
4911 * able to handle this situation (LZ4 does).
4914 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
4916 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
4918 if (zio->io_error != 0) {
4920 * An io error has occured, just restore the original io
4921 * size in preparation for a main pool read.
4923 zio->io_orig_size = zio->io_size = hdr->b_size;
4927 if (c == ZIO_COMPRESS_EMPTY) {
4929 * An empty buffer results in a null zio, which means we
4930 * need to fill its io_data after we're done restoring the
4931 * buffer's contents.
4933 ASSERT(hdr->b_buf != NULL);
4934 bzero(hdr->b_buf->b_data, hdr->b_size);
4935 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data;
4937 ASSERT(zio->io_data != NULL);
4939 * We copy the compressed data from the start of the arc buffer
4940 * (the zio_read will have pulled in only what we need, the
4941 * rest is garbage which we will overwrite at decompression)
4942 * and then decompress back to the ARC data buffer. This way we
4943 * can minimize copying by simply decompressing back over the
4944 * original compressed data (rather than decompressing to an
4945 * aux buffer and then copying back the uncompressed buffer,
4946 * which is likely to be much larger).
4951 csize = zio->io_size;
4952 cdata = zio_data_buf_alloc(csize);
4953 bcopy(zio->io_data, cdata, csize);
4954 if (zio_decompress_data(c, cdata, zio->io_data, csize,
4956 zio->io_error = EIO;
4957 zio_data_buf_free(cdata, csize);
4960 /* Restore the expected uncompressed IO size. */
4961 zio->io_orig_size = zio->io_size = hdr->b_size;
4965 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
4966 * This buffer serves as a temporary holder of compressed data while
4967 * the buffer entry is being written to an l2arc device. Once that is
4968 * done, we can dispose of it.
4971 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr)
4973 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
4975 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) {
4977 * If the data was compressed, then we've allocated a
4978 * temporary buffer for it, so now we need to release it.
4980 ASSERT(l2hdr->b_tmp_cdata != NULL);
4981 zio_data_buf_free(l2hdr->b_tmp_cdata, hdr->b_size);
4983 l2hdr->b_tmp_cdata = NULL;
4987 * This thread feeds the L2ARC at regular intervals. This is the beating
4988 * heart of the L2ARC.
4991 l2arc_feed_thread(void)
4996 uint64_t size, wrote;
4997 clock_t begin, next = ddi_get_lbolt();
4998 boolean_t headroom_boost = B_FALSE;
5000 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
5002 mutex_enter(&l2arc_feed_thr_lock);
5004 while (l2arc_thread_exit == 0) {
5005 CALLB_CPR_SAFE_BEGIN(&cpr);
5006 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
5008 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
5009 next = ddi_get_lbolt() + hz;
5012 * Quick check for L2ARC devices.
5014 mutex_enter(&l2arc_dev_mtx);
5015 if (l2arc_ndev == 0) {
5016 mutex_exit(&l2arc_dev_mtx);
5019 mutex_exit(&l2arc_dev_mtx);
5020 begin = ddi_get_lbolt();
5023 * This selects the next l2arc device to write to, and in
5024 * doing so the next spa to feed from: dev->l2ad_spa. This
5025 * will return NULL if there are now no l2arc devices or if
5026 * they are all faulted.
5028 * If a device is returned, its spa's config lock is also
5029 * held to prevent device removal. l2arc_dev_get_next()
5030 * will grab and release l2arc_dev_mtx.
5032 if ((dev = l2arc_dev_get_next()) == NULL)
5035 spa = dev->l2ad_spa;
5036 ASSERT(spa != NULL);
5039 * If the pool is read-only then force the feed thread to
5040 * sleep a little longer.
5042 if (!spa_writeable(spa)) {
5043 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
5044 spa_config_exit(spa, SCL_L2ARC, dev);
5049 * Avoid contributing to memory pressure.
5051 if (arc_reclaim_needed()) {
5052 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
5053 spa_config_exit(spa, SCL_L2ARC, dev);
5057 ARCSTAT_BUMP(arcstat_l2_feeds);
5059 size = l2arc_write_size();
5062 * Evict L2ARC buffers that will be overwritten.
5064 l2arc_evict(dev, size, B_FALSE);
5067 * Write ARC buffers.
5069 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
5072 * Calculate interval between writes.
5074 next = l2arc_write_interval(begin, size, wrote);
5075 spa_config_exit(spa, SCL_L2ARC, dev);
5078 l2arc_thread_exit = 0;
5079 cv_broadcast(&l2arc_feed_thr_cv);
5080 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
5085 l2arc_vdev_present(vdev_t *vd)
5089 mutex_enter(&l2arc_dev_mtx);
5090 for (dev = list_head(l2arc_dev_list); dev != NULL;
5091 dev = list_next(l2arc_dev_list, dev)) {
5092 if (dev->l2ad_vdev == vd)
5095 mutex_exit(&l2arc_dev_mtx);
5097 return (dev != NULL);
5101 * Add a vdev for use by the L2ARC. By this point the spa has already
5102 * validated the vdev and opened it.
5105 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5107 l2arc_dev_t *adddev;
5109 ASSERT(!l2arc_vdev_present(vd));
5112 * Create a new l2arc device entry.
5114 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5115 adddev->l2ad_spa = spa;
5116 adddev->l2ad_vdev = vd;
5117 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5118 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5119 adddev->l2ad_hand = adddev->l2ad_start;
5120 adddev->l2ad_evict = adddev->l2ad_start;
5121 adddev->l2ad_first = B_TRUE;
5122 adddev->l2ad_writing = B_FALSE;
5125 * This is a list of all ARC buffers that are still valid on the
5128 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5129 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5130 offsetof(arc_buf_hdr_t, b_l2node));
5132 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5135 * Add device to global list
5137 mutex_enter(&l2arc_dev_mtx);
5138 list_insert_head(l2arc_dev_list, adddev);
5139 atomic_inc_64(&l2arc_ndev);
5140 mutex_exit(&l2arc_dev_mtx);
5144 * Remove a vdev from the L2ARC.
5147 l2arc_remove_vdev(vdev_t *vd)
5149 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5152 * Find the device by vdev
5154 mutex_enter(&l2arc_dev_mtx);
5155 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5156 nextdev = list_next(l2arc_dev_list, dev);
5157 if (vd == dev->l2ad_vdev) {
5162 ASSERT(remdev != NULL);
5165 * Remove device from global list
5167 list_remove(l2arc_dev_list, remdev);
5168 l2arc_dev_last = NULL; /* may have been invalidated */
5169 atomic_dec_64(&l2arc_ndev);
5170 mutex_exit(&l2arc_dev_mtx);
5173 * Clear all buflists and ARC references. L2ARC device flush.
5175 l2arc_evict(remdev, 0, B_TRUE);
5176 list_destroy(remdev->l2ad_buflist);
5177 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5178 kmem_free(remdev, sizeof (l2arc_dev_t));
5184 l2arc_thread_exit = 0;
5186 l2arc_writes_sent = 0;
5187 l2arc_writes_done = 0;
5189 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5190 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5191 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5192 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5193 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5195 l2arc_dev_list = &L2ARC_dev_list;
5196 l2arc_free_on_write = &L2ARC_free_on_write;
5197 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5198 offsetof(l2arc_dev_t, l2ad_node));
5199 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5200 offsetof(l2arc_data_free_t, l2df_list_node));
5207 * This is called from dmu_fini(), which is called from spa_fini();
5208 * Because of this, we can assume that all l2arc devices have
5209 * already been removed when the pools themselves were removed.
5212 l2arc_do_free_on_write();
5214 mutex_destroy(&l2arc_feed_thr_lock);
5215 cv_destroy(&l2arc_feed_thr_cv);
5216 mutex_destroy(&l2arc_dev_mtx);
5217 mutex_destroy(&l2arc_buflist_mtx);
5218 mutex_destroy(&l2arc_free_on_write_mtx);
5220 list_destroy(l2arc_dev_list);
5221 list_destroy(l2arc_free_on_write);
5227 if (!(spa_mode_global & FWRITE))
5230 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5231 TS_RUN, minclsyspri);
5237 if (!(spa_mode_global & FWRITE))
5240 mutex_enter(&l2arc_feed_thr_lock);
5241 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5242 l2arc_thread_exit = 1;
5243 while (l2arc_thread_exit != 0)
5244 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5245 mutex_exit(&l2arc_feed_thr_lock);