4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
29 * DVA-based Adjustable Replacement Cache
31 * While much of the theory of operation used here is
32 * based on the self-tuning, low overhead replacement cache
33 * presented by Megiddo and Modha at FAST 2003, there are some
34 * significant differences:
36 * 1. The Megiddo and Modha model assumes any page is evictable.
37 * Pages in its cache cannot be "locked" into memory. This makes
38 * the eviction algorithm simple: evict the last page in the list.
39 * This also make the performance characteristics easy to reason
40 * about. Our cache is not so simple. At any given moment, some
41 * subset of the blocks in the cache are un-evictable because we
42 * have handed out a reference to them. Blocks are only evictable
43 * when there are no external references active. This makes
44 * eviction far more problematic: we choose to evict the evictable
45 * blocks that are the "lowest" in the list.
47 * There are times when it is not possible to evict the requested
48 * space. In these circumstances we are unable to adjust the cache
49 * size. To prevent the cache growing unbounded at these times we
50 * implement a "cache throttle" that slows the flow of new data
51 * into the cache until we can make space available.
53 * 2. The Megiddo and Modha model assumes a fixed cache size.
54 * Pages are evicted when the cache is full and there is a cache
55 * miss. Our model has a variable sized cache. It grows with
56 * high use, but also tries to react to memory pressure from the
57 * operating system: decreasing its size when system memory is
60 * 3. The Megiddo and Modha model assumes a fixed page size. All
61 * elements of the cache are therefore exactly the same size. So
62 * when adjusting the cache size following a cache miss, its simply
63 * a matter of choosing a single page to evict. In our model, we
64 * have variable sized cache blocks (rangeing from 512 bytes to
65 * 128K bytes). We therefore choose a set of blocks to evict to make
66 * space for a cache miss that approximates as closely as possible
67 * the space used by the new block.
69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
70 * by N. Megiddo & D. Modha, FAST 2003
76 * A new reference to a cache buffer can be obtained in two
77 * ways: 1) via a hash table lookup using the DVA as a key,
78 * or 2) via one of the ARC lists. The arc_read() interface
79 * uses method 1, while the internal arc algorithms for
80 * adjusting the cache use method 2. We therefore provide two
81 * types of locks: 1) the hash table lock array, and 2) the
84 * Buffers do not have their own mutexs, rather they rely on the
85 * hash table mutexs for the bulk of their protection (i.e. most
86 * fields in the arc_buf_hdr_t are protected by these mutexs).
88 * buf_hash_find() returns the appropriate mutex (held) when it
89 * locates the requested buffer in the hash table. It returns
90 * NULL for the mutex if the buffer was not in the table.
92 * buf_hash_remove() expects the appropriate hash mutex to be
93 * already held before it is invoked.
95 * Each arc state also has a mutex which is used to protect the
96 * buffer list associated with the state. When attempting to
97 * obtain a hash table lock while holding an arc list lock you
98 * must use: mutex_tryenter() to avoid deadlock. Also note that
99 * the active state mutex must be held before the ghost state mutex.
101 * Arc buffers may have an associated eviction callback function.
102 * This function will be invoked prior to removing the buffer (e.g.
103 * in arc_do_user_evicts()). Note however that the data associated
104 * with the buffer may be evicted prior to the callback. The callback
105 * must be made with *no locks held* (to prevent deadlock). Additionally,
106 * the users of callbacks must ensure that their private data is
107 * protected from simultaneous callbacks from arc_buf_evict()
108 * and arc_do_user_evicts().
110 * Note that the majority of the performance stats are manipulated
111 * with atomic operations.
113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
115 * - L2ARC buflist creation
116 * - L2ARC buflist eviction
117 * - L2ARC write completion, which walks L2ARC buflists
118 * - ARC header destruction, as it removes from L2ARC buflists
119 * - ARC header release, as it removes from L2ARC buflists
124 #include <sys/zio_compress.h>
125 #include <sys/zfs_context.h>
127 #include <sys/refcount.h>
128 #include <sys/vdev.h>
129 #include <sys/vdev_impl.h>
130 #include <sys/dsl_pool.h>
132 #include <sys/dnlc.h>
134 #include <sys/callb.h>
135 #include <sys/kstat.h>
136 #include <sys/trim_map.h>
137 #include <zfs_fletcher.h>
140 #include <vm/vm_pageout.h>
144 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
145 boolean_t arc_watch = B_FALSE;
150 static kmutex_t arc_reclaim_thr_lock;
151 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
152 static uint8_t arc_thread_exit;
154 #define ARC_REDUCE_DNLC_PERCENT 3
155 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
157 typedef enum arc_reclaim_strategy {
158 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
159 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
160 } arc_reclaim_strategy_t;
163 * The number of iterations through arc_evict_*() before we
164 * drop & reacquire the lock.
166 int arc_evict_iterations = 100;
168 /* number of seconds before growing cache again */
169 static int arc_grow_retry = 60;
171 /* shift of arc_c for calculating both min and max arc_p */
172 static int arc_p_min_shift = 4;
174 /* log2(fraction of arc to reclaim) */
175 static int arc_shrink_shift = 5;
178 * minimum lifespan of a prefetch block in clock ticks
179 * (initialized in arc_init())
181 static int arc_min_prefetch_lifespan;
184 * If this percent of memory is free, don't throttle.
186 int arc_lotsfree_percent = 10;
189 extern int zfs_prefetch_disable;
192 * The arc has filled available memory and has now warmed up.
194 static boolean_t arc_warm;
197 * These tunables are for performance analysis.
199 uint64_t zfs_arc_max;
200 uint64_t zfs_arc_min;
201 uint64_t zfs_arc_meta_limit = 0;
202 int zfs_arc_grow_retry = 0;
203 int zfs_arc_shrink_shift = 0;
204 int zfs_arc_p_min_shift = 0;
205 int zfs_disable_dup_eviction = 0;
207 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
208 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
209 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
210 SYSCTL_DECL(_vfs_zfs);
211 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
213 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
217 * Note that buffers can be in one of 6 states:
218 * ARC_anon - anonymous (discussed below)
219 * ARC_mru - recently used, currently cached
220 * ARC_mru_ghost - recentely used, no longer in cache
221 * ARC_mfu - frequently used, currently cached
222 * ARC_mfu_ghost - frequently used, no longer in cache
223 * ARC_l2c_only - exists in L2ARC but not other states
224 * When there are no active references to the buffer, they are
225 * are linked onto a list in one of these arc states. These are
226 * the only buffers that can be evicted or deleted. Within each
227 * state there are multiple lists, one for meta-data and one for
228 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
229 * etc.) is tracked separately so that it can be managed more
230 * explicitly: favored over data, limited explicitly.
232 * Anonymous buffers are buffers that are not associated with
233 * a DVA. These are buffers that hold dirty block copies
234 * before they are written to stable storage. By definition,
235 * they are "ref'd" and are considered part of arc_mru
236 * that cannot be freed. Generally, they will aquire a DVA
237 * as they are written and migrate onto the arc_mru list.
239 * The ARC_l2c_only state is for buffers that are in the second
240 * level ARC but no longer in any of the ARC_m* lists. The second
241 * level ARC itself may also contain buffers that are in any of
242 * the ARC_m* states - meaning that a buffer can exist in two
243 * places. The reason for the ARC_l2c_only state is to keep the
244 * buffer header in the hash table, so that reads that hit the
245 * second level ARC benefit from these fast lookups.
248 #define ARCS_LOCK_PAD CACHE_LINE_SIZE
252 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))];
257 * must be power of two for mask use to work
260 #define ARC_BUFC_NUMDATALISTS 16
261 #define ARC_BUFC_NUMMETADATALISTS 16
262 #define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS)
264 typedef struct arc_state {
265 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
266 uint64_t arcs_size; /* total amount of data in this state */
267 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */
268 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE);
271 #define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock))
274 static arc_state_t ARC_anon;
275 static arc_state_t ARC_mru;
276 static arc_state_t ARC_mru_ghost;
277 static arc_state_t ARC_mfu;
278 static arc_state_t ARC_mfu_ghost;
279 static arc_state_t ARC_l2c_only;
281 typedef struct arc_stats {
282 kstat_named_t arcstat_hits;
283 kstat_named_t arcstat_misses;
284 kstat_named_t arcstat_demand_data_hits;
285 kstat_named_t arcstat_demand_data_misses;
286 kstat_named_t arcstat_demand_metadata_hits;
287 kstat_named_t arcstat_demand_metadata_misses;
288 kstat_named_t arcstat_prefetch_data_hits;
289 kstat_named_t arcstat_prefetch_data_misses;
290 kstat_named_t arcstat_prefetch_metadata_hits;
291 kstat_named_t arcstat_prefetch_metadata_misses;
292 kstat_named_t arcstat_mru_hits;
293 kstat_named_t arcstat_mru_ghost_hits;
294 kstat_named_t arcstat_mfu_hits;
295 kstat_named_t arcstat_mfu_ghost_hits;
296 kstat_named_t arcstat_allocated;
297 kstat_named_t arcstat_deleted;
298 kstat_named_t arcstat_stolen;
299 kstat_named_t arcstat_recycle_miss;
301 * Number of buffers that could not be evicted because the hash lock
302 * was held by another thread. The lock may not necessarily be held
303 * by something using the same buffer, since hash locks are shared
304 * by multiple buffers.
306 kstat_named_t arcstat_mutex_miss;
308 * Number of buffers skipped because they have I/O in progress, are
309 * indrect prefetch buffers that have not lived long enough, or are
310 * not from the spa we're trying to evict from.
312 kstat_named_t arcstat_evict_skip;
313 kstat_named_t arcstat_evict_l2_cached;
314 kstat_named_t arcstat_evict_l2_eligible;
315 kstat_named_t arcstat_evict_l2_ineligible;
316 kstat_named_t arcstat_hash_elements;
317 kstat_named_t arcstat_hash_elements_max;
318 kstat_named_t arcstat_hash_collisions;
319 kstat_named_t arcstat_hash_chains;
320 kstat_named_t arcstat_hash_chain_max;
321 kstat_named_t arcstat_p;
322 kstat_named_t arcstat_c;
323 kstat_named_t arcstat_c_min;
324 kstat_named_t arcstat_c_max;
325 kstat_named_t arcstat_size;
326 kstat_named_t arcstat_hdr_size;
327 kstat_named_t arcstat_data_size;
328 kstat_named_t arcstat_other_size;
329 kstat_named_t arcstat_l2_hits;
330 kstat_named_t arcstat_l2_misses;
331 kstat_named_t arcstat_l2_feeds;
332 kstat_named_t arcstat_l2_rw_clash;
333 kstat_named_t arcstat_l2_read_bytes;
334 kstat_named_t arcstat_l2_write_bytes;
335 kstat_named_t arcstat_l2_writes_sent;
336 kstat_named_t arcstat_l2_writes_done;
337 kstat_named_t arcstat_l2_writes_error;
338 kstat_named_t arcstat_l2_writes_hdr_miss;
339 kstat_named_t arcstat_l2_evict_lock_retry;
340 kstat_named_t arcstat_l2_evict_reading;
341 kstat_named_t arcstat_l2_free_on_write;
342 kstat_named_t arcstat_l2_abort_lowmem;
343 kstat_named_t arcstat_l2_cksum_bad;
344 kstat_named_t arcstat_l2_io_error;
345 kstat_named_t arcstat_l2_size;
346 kstat_named_t arcstat_l2_asize;
347 kstat_named_t arcstat_l2_hdr_size;
348 kstat_named_t arcstat_l2_compress_successes;
349 kstat_named_t arcstat_l2_compress_zeros;
350 kstat_named_t arcstat_l2_compress_failures;
351 kstat_named_t arcstat_l2_write_trylock_fail;
352 kstat_named_t arcstat_l2_write_passed_headroom;
353 kstat_named_t arcstat_l2_write_spa_mismatch;
354 kstat_named_t arcstat_l2_write_in_l2;
355 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
356 kstat_named_t arcstat_l2_write_not_cacheable;
357 kstat_named_t arcstat_l2_write_full;
358 kstat_named_t arcstat_l2_write_buffer_iter;
359 kstat_named_t arcstat_l2_write_pios;
360 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
361 kstat_named_t arcstat_l2_write_buffer_list_iter;
362 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
363 kstat_named_t arcstat_memory_throttle_count;
364 kstat_named_t arcstat_duplicate_buffers;
365 kstat_named_t arcstat_duplicate_buffers_size;
366 kstat_named_t arcstat_duplicate_reads;
369 static arc_stats_t arc_stats = {
370 { "hits", KSTAT_DATA_UINT64 },
371 { "misses", KSTAT_DATA_UINT64 },
372 { "demand_data_hits", KSTAT_DATA_UINT64 },
373 { "demand_data_misses", KSTAT_DATA_UINT64 },
374 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
375 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
376 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
377 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
378 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
379 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
380 { "mru_hits", KSTAT_DATA_UINT64 },
381 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
382 { "mfu_hits", KSTAT_DATA_UINT64 },
383 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
384 { "allocated", KSTAT_DATA_UINT64 },
385 { "deleted", KSTAT_DATA_UINT64 },
386 { "stolen", KSTAT_DATA_UINT64 },
387 { "recycle_miss", KSTAT_DATA_UINT64 },
388 { "mutex_miss", KSTAT_DATA_UINT64 },
389 { "evict_skip", KSTAT_DATA_UINT64 },
390 { "evict_l2_cached", KSTAT_DATA_UINT64 },
391 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
392 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
393 { "hash_elements", KSTAT_DATA_UINT64 },
394 { "hash_elements_max", KSTAT_DATA_UINT64 },
395 { "hash_collisions", KSTAT_DATA_UINT64 },
396 { "hash_chains", KSTAT_DATA_UINT64 },
397 { "hash_chain_max", KSTAT_DATA_UINT64 },
398 { "p", KSTAT_DATA_UINT64 },
399 { "c", KSTAT_DATA_UINT64 },
400 { "c_min", KSTAT_DATA_UINT64 },
401 { "c_max", KSTAT_DATA_UINT64 },
402 { "size", KSTAT_DATA_UINT64 },
403 { "hdr_size", KSTAT_DATA_UINT64 },
404 { "data_size", KSTAT_DATA_UINT64 },
405 { "other_size", KSTAT_DATA_UINT64 },
406 { "l2_hits", KSTAT_DATA_UINT64 },
407 { "l2_misses", KSTAT_DATA_UINT64 },
408 { "l2_feeds", KSTAT_DATA_UINT64 },
409 { "l2_rw_clash", KSTAT_DATA_UINT64 },
410 { "l2_read_bytes", KSTAT_DATA_UINT64 },
411 { "l2_write_bytes", KSTAT_DATA_UINT64 },
412 { "l2_writes_sent", KSTAT_DATA_UINT64 },
413 { "l2_writes_done", KSTAT_DATA_UINT64 },
414 { "l2_writes_error", KSTAT_DATA_UINT64 },
415 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
416 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
417 { "l2_evict_reading", KSTAT_DATA_UINT64 },
418 { "l2_free_on_write", KSTAT_DATA_UINT64 },
419 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
420 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
421 { "l2_io_error", KSTAT_DATA_UINT64 },
422 { "l2_size", KSTAT_DATA_UINT64 },
423 { "l2_asize", KSTAT_DATA_UINT64 },
424 { "l2_hdr_size", KSTAT_DATA_UINT64 },
425 { "l2_compress_successes", KSTAT_DATA_UINT64 },
426 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
427 { "l2_compress_failures", KSTAT_DATA_UINT64 },
428 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
429 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
430 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
431 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
432 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
433 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
434 { "l2_write_full", KSTAT_DATA_UINT64 },
435 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
436 { "l2_write_pios", KSTAT_DATA_UINT64 },
437 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
438 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
439 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
440 { "memory_throttle_count", KSTAT_DATA_UINT64 },
441 { "duplicate_buffers", KSTAT_DATA_UINT64 },
442 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
443 { "duplicate_reads", KSTAT_DATA_UINT64 }
446 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
448 #define ARCSTAT_INCR(stat, val) \
449 atomic_add_64(&arc_stats.stat.value.ui64, (val))
451 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
452 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
454 #define ARCSTAT_MAX(stat, val) { \
456 while ((val) > (m = arc_stats.stat.value.ui64) && \
457 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
461 #define ARCSTAT_MAXSTAT(stat) \
462 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
465 * We define a macro to allow ARC hits/misses to be easily broken down by
466 * two separate conditions, giving a total of four different subtypes for
467 * each of hits and misses (so eight statistics total).
469 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
472 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
474 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
478 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
480 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
485 static arc_state_t *arc_anon;
486 static arc_state_t *arc_mru;
487 static arc_state_t *arc_mru_ghost;
488 static arc_state_t *arc_mfu;
489 static arc_state_t *arc_mfu_ghost;
490 static arc_state_t *arc_l2c_only;
493 * There are several ARC variables that are critical to export as kstats --
494 * but we don't want to have to grovel around in the kstat whenever we wish to
495 * manipulate them. For these variables, we therefore define them to be in
496 * terms of the statistic variable. This assures that we are not introducing
497 * the possibility of inconsistency by having shadow copies of the variables,
498 * while still allowing the code to be readable.
500 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
501 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
502 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
503 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
504 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
506 #define L2ARC_IS_VALID_COMPRESS(_c_) \
507 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
509 static int arc_no_grow; /* Don't try to grow cache size */
510 static uint64_t arc_tempreserve;
511 static uint64_t arc_loaned_bytes;
512 static uint64_t arc_meta_used;
513 static uint64_t arc_meta_limit;
514 static uint64_t arc_meta_max = 0;
515 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0,
516 "ARC metadata used");
517 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0,
518 "ARC metadata limit");
520 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
522 typedef struct arc_callback arc_callback_t;
524 struct arc_callback {
526 arc_done_func_t *acb_done;
528 zio_t *acb_zio_dummy;
529 arc_callback_t *acb_next;
532 typedef struct arc_write_callback arc_write_callback_t;
534 struct arc_write_callback {
536 arc_done_func_t *awcb_ready;
537 arc_done_func_t *awcb_physdone;
538 arc_done_func_t *awcb_done;
543 /* protected by hash lock */
548 kmutex_t b_freeze_lock;
549 zio_cksum_t *b_freeze_cksum;
552 arc_buf_hdr_t *b_hash_next;
557 arc_callback_t *b_acb;
561 arc_buf_contents_t b_type;
565 /* protected by arc state mutex */
566 arc_state_t *b_state;
567 list_node_t b_arc_node;
569 /* updated atomically */
570 clock_t b_arc_access;
572 /* self protecting */
575 l2arc_buf_hdr_t *b_l2hdr;
576 list_node_t b_l2node;
579 static arc_buf_t *arc_eviction_list;
580 static kmutex_t arc_eviction_mtx;
581 static arc_buf_hdr_t arc_eviction_hdr;
582 static void arc_get_data_buf(arc_buf_t *buf);
583 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
584 static int arc_evict_needed(arc_buf_contents_t type);
585 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
587 static void arc_buf_watch(arc_buf_t *buf);
590 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
592 #define GHOST_STATE(state) \
593 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
594 (state) == arc_l2c_only)
597 * Private ARC flags. These flags are private ARC only flags that will show up
598 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
599 * be passed in as arc_flags in things like arc_read. However, these flags
600 * should never be passed and should only be set by ARC code. When adding new
601 * public flags, make sure not to smash the private ones.
604 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
605 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
606 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
607 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
608 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
609 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
610 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
611 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
612 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
613 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
615 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
616 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
617 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
618 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
619 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
620 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
621 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
622 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
623 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
624 (hdr)->b_l2hdr != NULL)
625 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
626 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
627 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
633 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
634 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
637 * Hash table routines
640 #define HT_LOCK_PAD CACHE_LINE_SIZE
645 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
649 #define BUF_LOCKS 256
650 typedef struct buf_hash_table {
652 arc_buf_hdr_t **ht_table;
653 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
656 static buf_hash_table_t buf_hash_table;
658 #define BUF_HASH_INDEX(spa, dva, birth) \
659 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
660 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
661 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
662 #define HDR_LOCK(hdr) \
663 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
665 uint64_t zfs_crc64_table[256];
671 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
672 #define L2ARC_HEADROOM 2 /* num of writes */
674 * If we discover during ARC scan any buffers to be compressed, we boost
675 * our headroom for the next scanning cycle by this percentage multiple.
677 #define L2ARC_HEADROOM_BOOST 200
678 #define L2ARC_FEED_SECS 1 /* caching interval secs */
679 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
681 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
682 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
684 /* L2ARC Performance Tunables */
685 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
686 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
687 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
688 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
689 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
690 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
691 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
692 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
693 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
695 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
696 &l2arc_write_max, 0, "max write size");
697 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
698 &l2arc_write_boost, 0, "extra write during warmup");
699 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
700 &l2arc_headroom, 0, "number of dev writes");
701 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
702 &l2arc_feed_secs, 0, "interval seconds");
703 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
704 &l2arc_feed_min_ms, 0, "min interval milliseconds");
706 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
707 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
708 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
709 &l2arc_feed_again, 0, "turbo warmup");
710 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
711 &l2arc_norw, 0, "no reads during writes");
713 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
714 &ARC_anon.arcs_size, 0, "size of anonymous state");
715 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
716 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
717 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
718 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
720 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
721 &ARC_mru.arcs_size, 0, "size of mru state");
722 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
723 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
724 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
725 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
727 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
728 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state");
729 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
730 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
731 "size of metadata in mru ghost state");
732 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
733 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
734 "size of data in mru ghost state");
736 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
737 &ARC_mfu.arcs_size, 0, "size of mfu state");
738 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
739 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
740 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
741 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
743 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
744 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state");
745 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
746 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
747 "size of metadata in mfu ghost state");
748 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
749 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
750 "size of data in mfu ghost state");
752 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
753 &ARC_l2c_only.arcs_size, 0, "size of mru state");
758 typedef struct l2arc_dev {
759 vdev_t *l2ad_vdev; /* vdev */
760 spa_t *l2ad_spa; /* spa */
761 uint64_t l2ad_hand; /* next write location */
762 uint64_t l2ad_start; /* first addr on device */
763 uint64_t l2ad_end; /* last addr on device */
764 uint64_t l2ad_evict; /* last addr eviction reached */
765 boolean_t l2ad_first; /* first sweep through */
766 boolean_t l2ad_writing; /* currently writing */
767 list_t *l2ad_buflist; /* buffer list */
768 list_node_t l2ad_node; /* device list node */
771 static list_t L2ARC_dev_list; /* device list */
772 static list_t *l2arc_dev_list; /* device list pointer */
773 static kmutex_t l2arc_dev_mtx; /* device list mutex */
774 static l2arc_dev_t *l2arc_dev_last; /* last device used */
775 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
776 static list_t L2ARC_free_on_write; /* free after write buf list */
777 static list_t *l2arc_free_on_write; /* free after write list ptr */
778 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
779 static uint64_t l2arc_ndev; /* number of devices */
781 typedef struct l2arc_read_callback {
782 arc_buf_t *l2rcb_buf; /* read buffer */
783 spa_t *l2rcb_spa; /* spa */
784 blkptr_t l2rcb_bp; /* original blkptr */
785 zbookmark_t l2rcb_zb; /* original bookmark */
786 int l2rcb_flags; /* original flags */
787 enum zio_compress l2rcb_compress; /* applied compress */
788 } l2arc_read_callback_t;
790 typedef struct l2arc_write_callback {
791 l2arc_dev_t *l2wcb_dev; /* device info */
792 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
793 } l2arc_write_callback_t;
795 struct l2arc_buf_hdr {
796 /* protected by arc_buf_hdr mutex */
797 l2arc_dev_t *b_dev; /* L2ARC device */
798 uint64_t b_daddr; /* disk address, offset byte */
799 /* compression applied to buffer data */
800 enum zio_compress b_compress;
801 /* real alloc'd buffer size depending on b_compress applied */
803 /* temporary buffer holder for in-flight compressed data */
807 typedef struct l2arc_data_free {
808 /* protected by l2arc_free_on_write_mtx */
811 void (*l2df_func)(void *, size_t);
812 list_node_t l2df_list_node;
815 static kmutex_t l2arc_feed_thr_lock;
816 static kcondvar_t l2arc_feed_thr_cv;
817 static uint8_t l2arc_thread_exit;
819 static void l2arc_read_done(zio_t *zio);
820 static void l2arc_hdr_stat_add(void);
821 static void l2arc_hdr_stat_remove(void);
823 static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr);
824 static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr,
825 enum zio_compress c);
826 static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab);
829 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
831 uint8_t *vdva = (uint8_t *)dva;
832 uint64_t crc = -1ULL;
835 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
837 for (i = 0; i < sizeof (dva_t); i++)
838 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
840 crc ^= (spa>>8) ^ birth;
845 #define BUF_EMPTY(buf) \
846 ((buf)->b_dva.dva_word[0] == 0 && \
847 (buf)->b_dva.dva_word[1] == 0 && \
848 (buf)->b_cksum0 == 0)
850 #define BUF_EQUAL(spa, dva, birth, buf) \
851 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
852 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
853 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
856 buf_discard_identity(arc_buf_hdr_t *hdr)
858 hdr->b_dva.dva_word[0] = 0;
859 hdr->b_dva.dva_word[1] = 0;
864 static arc_buf_hdr_t *
865 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
867 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
868 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
871 mutex_enter(hash_lock);
872 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
873 buf = buf->b_hash_next) {
874 if (BUF_EQUAL(spa, dva, birth, buf)) {
879 mutex_exit(hash_lock);
885 * Insert an entry into the hash table. If there is already an element
886 * equal to elem in the hash table, then the already existing element
887 * will be returned and the new element will not be inserted.
888 * Otherwise returns NULL.
890 static arc_buf_hdr_t *
891 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
893 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
894 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
898 ASSERT(!HDR_IN_HASH_TABLE(buf));
900 mutex_enter(hash_lock);
901 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
902 fbuf = fbuf->b_hash_next, i++) {
903 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
907 buf->b_hash_next = buf_hash_table.ht_table[idx];
908 buf_hash_table.ht_table[idx] = buf;
909 buf->b_flags |= ARC_IN_HASH_TABLE;
911 /* collect some hash table performance data */
913 ARCSTAT_BUMP(arcstat_hash_collisions);
915 ARCSTAT_BUMP(arcstat_hash_chains);
917 ARCSTAT_MAX(arcstat_hash_chain_max, i);
920 ARCSTAT_BUMP(arcstat_hash_elements);
921 ARCSTAT_MAXSTAT(arcstat_hash_elements);
927 buf_hash_remove(arc_buf_hdr_t *buf)
929 arc_buf_hdr_t *fbuf, **bufp;
930 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
932 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
933 ASSERT(HDR_IN_HASH_TABLE(buf));
935 bufp = &buf_hash_table.ht_table[idx];
936 while ((fbuf = *bufp) != buf) {
937 ASSERT(fbuf != NULL);
938 bufp = &fbuf->b_hash_next;
940 *bufp = buf->b_hash_next;
941 buf->b_hash_next = NULL;
942 buf->b_flags &= ~ARC_IN_HASH_TABLE;
944 /* collect some hash table performance data */
945 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
947 if (buf_hash_table.ht_table[idx] &&
948 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
949 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
953 * Global data structures and functions for the buf kmem cache.
955 static kmem_cache_t *hdr_cache;
956 static kmem_cache_t *buf_cache;
963 kmem_free(buf_hash_table.ht_table,
964 (buf_hash_table.ht_mask + 1) * sizeof (void *));
965 for (i = 0; i < BUF_LOCKS; i++)
966 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
967 kmem_cache_destroy(hdr_cache);
968 kmem_cache_destroy(buf_cache);
972 * Constructor callback - called when the cache is empty
973 * and a new buf is requested.
977 hdr_cons(void *vbuf, void *unused, int kmflag)
979 arc_buf_hdr_t *buf = vbuf;
981 bzero(buf, sizeof (arc_buf_hdr_t));
982 refcount_create(&buf->b_refcnt);
983 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
984 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
985 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
992 buf_cons(void *vbuf, void *unused, int kmflag)
994 arc_buf_t *buf = vbuf;
996 bzero(buf, sizeof (arc_buf_t));
997 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
998 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1004 * Destructor callback - called when a cached buf is
1005 * no longer required.
1009 hdr_dest(void *vbuf, void *unused)
1011 arc_buf_hdr_t *buf = vbuf;
1013 ASSERT(BUF_EMPTY(buf));
1014 refcount_destroy(&buf->b_refcnt);
1015 cv_destroy(&buf->b_cv);
1016 mutex_destroy(&buf->b_freeze_lock);
1017 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
1022 buf_dest(void *vbuf, void *unused)
1024 arc_buf_t *buf = vbuf;
1026 mutex_destroy(&buf->b_evict_lock);
1027 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1031 * Reclaim callback -- invoked when memory is low.
1035 hdr_recl(void *unused)
1037 dprintf("hdr_recl called\n");
1039 * umem calls the reclaim func when we destroy the buf cache,
1040 * which is after we do arc_fini().
1043 cv_signal(&arc_reclaim_thr_cv);
1050 uint64_t hsize = 1ULL << 12;
1054 * The hash table is big enough to fill all of physical memory
1055 * with an average 64K block size. The table will take up
1056 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
1058 while (hsize * 65536 < (uint64_t)physmem * PAGESIZE)
1061 buf_hash_table.ht_mask = hsize - 1;
1062 buf_hash_table.ht_table =
1063 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1064 if (buf_hash_table.ht_table == NULL) {
1065 ASSERT(hsize > (1ULL << 8));
1070 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
1071 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
1072 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1073 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1075 for (i = 0; i < 256; i++)
1076 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1077 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1079 for (i = 0; i < BUF_LOCKS; i++) {
1080 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1081 NULL, MUTEX_DEFAULT, NULL);
1085 #define ARC_MINTIME (hz>>4) /* 62 ms */
1088 arc_cksum_verify(arc_buf_t *buf)
1092 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1095 mutex_enter(&buf->b_hdr->b_freeze_lock);
1096 if (buf->b_hdr->b_freeze_cksum == NULL ||
1097 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
1098 mutex_exit(&buf->b_hdr->b_freeze_lock);
1101 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1102 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1103 panic("buffer modified while frozen!");
1104 mutex_exit(&buf->b_hdr->b_freeze_lock);
1108 arc_cksum_equal(arc_buf_t *buf)
1113 mutex_enter(&buf->b_hdr->b_freeze_lock);
1114 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1115 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1116 mutex_exit(&buf->b_hdr->b_freeze_lock);
1122 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1124 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1127 mutex_enter(&buf->b_hdr->b_freeze_lock);
1128 if (buf->b_hdr->b_freeze_cksum != NULL) {
1129 mutex_exit(&buf->b_hdr->b_freeze_lock);
1132 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1133 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1134 buf->b_hdr->b_freeze_cksum);
1135 mutex_exit(&buf->b_hdr->b_freeze_lock);
1138 #endif /* illumos */
1143 typedef struct procctl {
1151 arc_buf_unwatch(arc_buf_t *buf)
1158 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1159 ctl.prwatch.pr_size = 0;
1160 ctl.prwatch.pr_wflags = 0;
1161 result = write(arc_procfd, &ctl, sizeof (ctl));
1162 ASSERT3U(result, ==, sizeof (ctl));
1169 arc_buf_watch(arc_buf_t *buf)
1176 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1177 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1178 ctl.prwatch.pr_wflags = WA_WRITE;
1179 result = write(arc_procfd, &ctl, sizeof (ctl));
1180 ASSERT3U(result, ==, sizeof (ctl));
1184 #endif /* illumos */
1187 arc_buf_thaw(arc_buf_t *buf)
1189 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1190 if (buf->b_hdr->b_state != arc_anon)
1191 panic("modifying non-anon buffer!");
1192 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1193 panic("modifying buffer while i/o in progress!");
1194 arc_cksum_verify(buf);
1197 mutex_enter(&buf->b_hdr->b_freeze_lock);
1198 if (buf->b_hdr->b_freeze_cksum != NULL) {
1199 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1200 buf->b_hdr->b_freeze_cksum = NULL;
1203 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1204 if (buf->b_hdr->b_thawed)
1205 kmem_free(buf->b_hdr->b_thawed, 1);
1206 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1209 mutex_exit(&buf->b_hdr->b_freeze_lock);
1212 arc_buf_unwatch(buf);
1213 #endif /* illumos */
1217 arc_buf_freeze(arc_buf_t *buf)
1219 kmutex_t *hash_lock;
1221 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1224 hash_lock = HDR_LOCK(buf->b_hdr);
1225 mutex_enter(hash_lock);
1227 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1228 buf->b_hdr->b_state == arc_anon);
1229 arc_cksum_compute(buf, B_FALSE);
1230 mutex_exit(hash_lock);
1235 get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock)
1237 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth);
1239 if (ab->b_type == ARC_BUFC_METADATA)
1240 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1);
1242 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1);
1243 buf_hashid += ARC_BUFC_NUMMETADATALISTS;
1246 *list = &state->arcs_lists[buf_hashid];
1247 *lock = ARCS_LOCK(state, buf_hashid);
1252 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1254 ASSERT(MUTEX_HELD(hash_lock));
1256 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1257 (ab->b_state != arc_anon)) {
1258 uint64_t delta = ab->b_size * ab->b_datacnt;
1259 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1263 get_buf_info(ab, ab->b_state, &list, &lock);
1264 ASSERT(!MUTEX_HELD(lock));
1266 ASSERT(list_link_active(&ab->b_arc_node));
1267 list_remove(list, ab);
1268 if (GHOST_STATE(ab->b_state)) {
1269 ASSERT0(ab->b_datacnt);
1270 ASSERT3P(ab->b_buf, ==, NULL);
1274 ASSERT3U(*size, >=, delta);
1275 atomic_add_64(size, -delta);
1277 /* remove the prefetch flag if we get a reference */
1278 if (ab->b_flags & ARC_PREFETCH)
1279 ab->b_flags &= ~ARC_PREFETCH;
1284 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1287 arc_state_t *state = ab->b_state;
1289 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1290 ASSERT(!GHOST_STATE(state));
1292 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1293 (state != arc_anon)) {
1294 uint64_t *size = &state->arcs_lsize[ab->b_type];
1298 get_buf_info(ab, state, &list, &lock);
1299 ASSERT(!MUTEX_HELD(lock));
1301 ASSERT(!list_link_active(&ab->b_arc_node));
1302 list_insert_head(list, ab);
1303 ASSERT(ab->b_datacnt > 0);
1304 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1311 * Move the supplied buffer to the indicated state. The mutex
1312 * for the buffer must be held by the caller.
1315 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1317 arc_state_t *old_state = ab->b_state;
1318 int64_t refcnt = refcount_count(&ab->b_refcnt);
1319 uint64_t from_delta, to_delta;
1323 ASSERT(MUTEX_HELD(hash_lock));
1324 ASSERT3P(new_state, !=, old_state);
1325 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1326 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1327 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1329 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1332 * If this buffer is evictable, transfer it from the
1333 * old state list to the new state list.
1336 if (old_state != arc_anon) {
1338 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1340 get_buf_info(ab, old_state, &list, &lock);
1341 use_mutex = !MUTEX_HELD(lock);
1345 ASSERT(list_link_active(&ab->b_arc_node));
1346 list_remove(list, ab);
1349 * If prefetching out of the ghost cache,
1350 * we will have a non-zero datacnt.
1352 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1353 /* ghost elements have a ghost size */
1354 ASSERT(ab->b_buf == NULL);
1355 from_delta = ab->b_size;
1357 ASSERT3U(*size, >=, from_delta);
1358 atomic_add_64(size, -from_delta);
1363 if (new_state != arc_anon) {
1365 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1367 get_buf_info(ab, new_state, &list, &lock);
1368 use_mutex = !MUTEX_HELD(lock);
1372 list_insert_head(list, ab);
1374 /* ghost elements have a ghost size */
1375 if (GHOST_STATE(new_state)) {
1376 ASSERT(ab->b_datacnt == 0);
1377 ASSERT(ab->b_buf == NULL);
1378 to_delta = ab->b_size;
1380 atomic_add_64(size, to_delta);
1387 ASSERT(!BUF_EMPTY(ab));
1388 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1389 buf_hash_remove(ab);
1391 /* adjust state sizes */
1393 atomic_add_64(&new_state->arcs_size, to_delta);
1395 ASSERT3U(old_state->arcs_size, >=, from_delta);
1396 atomic_add_64(&old_state->arcs_size, -from_delta);
1398 ab->b_state = new_state;
1400 /* adjust l2arc hdr stats */
1401 if (new_state == arc_l2c_only)
1402 l2arc_hdr_stat_add();
1403 else if (old_state == arc_l2c_only)
1404 l2arc_hdr_stat_remove();
1408 arc_space_consume(uint64_t space, arc_space_type_t type)
1410 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1413 case ARC_SPACE_DATA:
1414 ARCSTAT_INCR(arcstat_data_size, space);
1416 case ARC_SPACE_OTHER:
1417 ARCSTAT_INCR(arcstat_other_size, space);
1419 case ARC_SPACE_HDRS:
1420 ARCSTAT_INCR(arcstat_hdr_size, space);
1422 case ARC_SPACE_L2HDRS:
1423 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1427 atomic_add_64(&arc_meta_used, space);
1428 atomic_add_64(&arc_size, space);
1432 arc_space_return(uint64_t space, arc_space_type_t type)
1434 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1437 case ARC_SPACE_DATA:
1438 ARCSTAT_INCR(arcstat_data_size, -space);
1440 case ARC_SPACE_OTHER:
1441 ARCSTAT_INCR(arcstat_other_size, -space);
1443 case ARC_SPACE_HDRS:
1444 ARCSTAT_INCR(arcstat_hdr_size, -space);
1446 case ARC_SPACE_L2HDRS:
1447 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1451 ASSERT(arc_meta_used >= space);
1452 if (arc_meta_max < arc_meta_used)
1453 arc_meta_max = arc_meta_used;
1454 atomic_add_64(&arc_meta_used, -space);
1455 ASSERT(arc_size >= space);
1456 atomic_add_64(&arc_size, -space);
1460 arc_data_buf_alloc(uint64_t size)
1462 if (arc_evict_needed(ARC_BUFC_DATA))
1463 cv_signal(&arc_reclaim_thr_cv);
1464 atomic_add_64(&arc_size, size);
1465 return (zio_data_buf_alloc(size));
1469 arc_data_buf_free(void *buf, uint64_t size)
1471 zio_data_buf_free(buf, size);
1472 ASSERT(arc_size >= size);
1473 atomic_add_64(&arc_size, -size);
1477 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1482 ASSERT3U(size, >, 0);
1483 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1484 ASSERT(BUF_EMPTY(hdr));
1487 hdr->b_spa = spa_load_guid(spa);
1488 hdr->b_state = arc_anon;
1489 hdr->b_arc_access = 0;
1490 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1493 buf->b_efunc = NULL;
1494 buf->b_private = NULL;
1497 arc_get_data_buf(buf);
1500 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1501 (void) refcount_add(&hdr->b_refcnt, tag);
1506 static char *arc_onloan_tag = "onloan";
1509 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1510 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1511 * buffers must be returned to the arc before they can be used by the DMU or
1515 arc_loan_buf(spa_t *spa, int size)
1519 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1521 atomic_add_64(&arc_loaned_bytes, size);
1526 * Return a loaned arc buffer to the arc.
1529 arc_return_buf(arc_buf_t *buf, void *tag)
1531 arc_buf_hdr_t *hdr = buf->b_hdr;
1533 ASSERT(buf->b_data != NULL);
1534 (void) refcount_add(&hdr->b_refcnt, tag);
1535 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1537 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1540 /* Detach an arc_buf from a dbuf (tag) */
1542 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1546 ASSERT(buf->b_data != NULL);
1548 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1549 (void) refcount_remove(&hdr->b_refcnt, tag);
1550 buf->b_efunc = NULL;
1551 buf->b_private = NULL;
1553 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1557 arc_buf_clone(arc_buf_t *from)
1560 arc_buf_hdr_t *hdr = from->b_hdr;
1561 uint64_t size = hdr->b_size;
1563 ASSERT(hdr->b_state != arc_anon);
1565 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1568 buf->b_efunc = NULL;
1569 buf->b_private = NULL;
1570 buf->b_next = hdr->b_buf;
1572 arc_get_data_buf(buf);
1573 bcopy(from->b_data, buf->b_data, size);
1576 * This buffer already exists in the arc so create a duplicate
1577 * copy for the caller. If the buffer is associated with user data
1578 * then track the size and number of duplicates. These stats will be
1579 * updated as duplicate buffers are created and destroyed.
1581 if (hdr->b_type == ARC_BUFC_DATA) {
1582 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1583 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1585 hdr->b_datacnt += 1;
1590 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1593 kmutex_t *hash_lock;
1596 * Check to see if this buffer is evicted. Callers
1597 * must verify b_data != NULL to know if the add_ref
1600 mutex_enter(&buf->b_evict_lock);
1601 if (buf->b_data == NULL) {
1602 mutex_exit(&buf->b_evict_lock);
1605 hash_lock = HDR_LOCK(buf->b_hdr);
1606 mutex_enter(hash_lock);
1608 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1609 mutex_exit(&buf->b_evict_lock);
1611 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1612 add_reference(hdr, hash_lock, tag);
1613 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1614 arc_access(hdr, hash_lock);
1615 mutex_exit(hash_lock);
1616 ARCSTAT_BUMP(arcstat_hits);
1617 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1618 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1619 data, metadata, hits);
1623 * Free the arc data buffer. If it is an l2arc write in progress,
1624 * the buffer is placed on l2arc_free_on_write to be freed later.
1627 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1629 arc_buf_hdr_t *hdr = buf->b_hdr;
1631 if (HDR_L2_WRITING(hdr)) {
1632 l2arc_data_free_t *df;
1633 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1634 df->l2df_data = buf->b_data;
1635 df->l2df_size = hdr->b_size;
1636 df->l2df_func = free_func;
1637 mutex_enter(&l2arc_free_on_write_mtx);
1638 list_insert_head(l2arc_free_on_write, df);
1639 mutex_exit(&l2arc_free_on_write_mtx);
1640 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1642 free_func(buf->b_data, hdr->b_size);
1647 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1651 /* free up data associated with the buf */
1653 arc_state_t *state = buf->b_hdr->b_state;
1654 uint64_t size = buf->b_hdr->b_size;
1655 arc_buf_contents_t type = buf->b_hdr->b_type;
1657 arc_cksum_verify(buf);
1659 arc_buf_unwatch(buf);
1660 #endif /* illumos */
1663 if (type == ARC_BUFC_METADATA) {
1664 arc_buf_data_free(buf, zio_buf_free);
1665 arc_space_return(size, ARC_SPACE_DATA);
1667 ASSERT(type == ARC_BUFC_DATA);
1668 arc_buf_data_free(buf, zio_data_buf_free);
1669 ARCSTAT_INCR(arcstat_data_size, -size);
1670 atomic_add_64(&arc_size, -size);
1673 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1674 uint64_t *cnt = &state->arcs_lsize[type];
1676 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1677 ASSERT(state != arc_anon);
1679 ASSERT3U(*cnt, >=, size);
1680 atomic_add_64(cnt, -size);
1682 ASSERT3U(state->arcs_size, >=, size);
1683 atomic_add_64(&state->arcs_size, -size);
1687 * If we're destroying a duplicate buffer make sure
1688 * that the appropriate statistics are updated.
1690 if (buf->b_hdr->b_datacnt > 1 &&
1691 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1692 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1693 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1695 ASSERT(buf->b_hdr->b_datacnt > 0);
1696 buf->b_hdr->b_datacnt -= 1;
1699 /* only remove the buf if requested */
1703 /* remove the buf from the hdr list */
1704 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1706 *bufp = buf->b_next;
1709 ASSERT(buf->b_efunc == NULL);
1711 /* clean up the buf */
1713 kmem_cache_free(buf_cache, buf);
1717 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1719 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1720 ASSERT3P(hdr->b_state, ==, arc_anon);
1721 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1722 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1724 if (l2hdr != NULL) {
1725 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1727 * To prevent arc_free() and l2arc_evict() from
1728 * attempting to free the same buffer at the same time,
1729 * a FREE_IN_PROGRESS flag is given to arc_free() to
1730 * give it priority. l2arc_evict() can't destroy this
1731 * header while we are waiting on l2arc_buflist_mtx.
1733 * The hdr may be removed from l2ad_buflist before we
1734 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1736 if (!buflist_held) {
1737 mutex_enter(&l2arc_buflist_mtx);
1738 l2hdr = hdr->b_l2hdr;
1741 if (l2hdr != NULL) {
1742 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr,
1744 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1745 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1746 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
1747 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1748 if (hdr->b_state == arc_l2c_only)
1749 l2arc_hdr_stat_remove();
1750 hdr->b_l2hdr = NULL;
1754 mutex_exit(&l2arc_buflist_mtx);
1757 if (!BUF_EMPTY(hdr)) {
1758 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1759 buf_discard_identity(hdr);
1761 while (hdr->b_buf) {
1762 arc_buf_t *buf = hdr->b_buf;
1765 mutex_enter(&arc_eviction_mtx);
1766 mutex_enter(&buf->b_evict_lock);
1767 ASSERT(buf->b_hdr != NULL);
1768 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1769 hdr->b_buf = buf->b_next;
1770 buf->b_hdr = &arc_eviction_hdr;
1771 buf->b_next = arc_eviction_list;
1772 arc_eviction_list = buf;
1773 mutex_exit(&buf->b_evict_lock);
1774 mutex_exit(&arc_eviction_mtx);
1776 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1779 if (hdr->b_freeze_cksum != NULL) {
1780 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1781 hdr->b_freeze_cksum = NULL;
1783 if (hdr->b_thawed) {
1784 kmem_free(hdr->b_thawed, 1);
1785 hdr->b_thawed = NULL;
1788 ASSERT(!list_link_active(&hdr->b_arc_node));
1789 ASSERT3P(hdr->b_hash_next, ==, NULL);
1790 ASSERT3P(hdr->b_acb, ==, NULL);
1791 kmem_cache_free(hdr_cache, hdr);
1795 arc_buf_free(arc_buf_t *buf, void *tag)
1797 arc_buf_hdr_t *hdr = buf->b_hdr;
1798 int hashed = hdr->b_state != arc_anon;
1800 ASSERT(buf->b_efunc == NULL);
1801 ASSERT(buf->b_data != NULL);
1804 kmutex_t *hash_lock = HDR_LOCK(hdr);
1806 mutex_enter(hash_lock);
1808 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1810 (void) remove_reference(hdr, hash_lock, tag);
1811 if (hdr->b_datacnt > 1) {
1812 arc_buf_destroy(buf, FALSE, TRUE);
1814 ASSERT(buf == hdr->b_buf);
1815 ASSERT(buf->b_efunc == NULL);
1816 hdr->b_flags |= ARC_BUF_AVAILABLE;
1818 mutex_exit(hash_lock);
1819 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1822 * We are in the middle of an async write. Don't destroy
1823 * this buffer unless the write completes before we finish
1824 * decrementing the reference count.
1826 mutex_enter(&arc_eviction_mtx);
1827 (void) remove_reference(hdr, NULL, tag);
1828 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1829 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1830 mutex_exit(&arc_eviction_mtx);
1832 arc_hdr_destroy(hdr);
1834 if (remove_reference(hdr, NULL, tag) > 0)
1835 arc_buf_destroy(buf, FALSE, TRUE);
1837 arc_hdr_destroy(hdr);
1842 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1844 arc_buf_hdr_t *hdr = buf->b_hdr;
1845 kmutex_t *hash_lock = HDR_LOCK(hdr);
1846 boolean_t no_callback = (buf->b_efunc == NULL);
1848 if (hdr->b_state == arc_anon) {
1849 ASSERT(hdr->b_datacnt == 1);
1850 arc_buf_free(buf, tag);
1851 return (no_callback);
1854 mutex_enter(hash_lock);
1856 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1857 ASSERT(hdr->b_state != arc_anon);
1858 ASSERT(buf->b_data != NULL);
1860 (void) remove_reference(hdr, hash_lock, tag);
1861 if (hdr->b_datacnt > 1) {
1863 arc_buf_destroy(buf, FALSE, TRUE);
1864 } else if (no_callback) {
1865 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1866 ASSERT(buf->b_efunc == NULL);
1867 hdr->b_flags |= ARC_BUF_AVAILABLE;
1869 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1870 refcount_is_zero(&hdr->b_refcnt));
1871 mutex_exit(hash_lock);
1872 return (no_callback);
1876 arc_buf_size(arc_buf_t *buf)
1878 return (buf->b_hdr->b_size);
1882 * Called from the DMU to determine if the current buffer should be
1883 * evicted. In order to ensure proper locking, the eviction must be initiated
1884 * from the DMU. Return true if the buffer is associated with user data and
1885 * duplicate buffers still exist.
1888 arc_buf_eviction_needed(arc_buf_t *buf)
1891 boolean_t evict_needed = B_FALSE;
1893 if (zfs_disable_dup_eviction)
1896 mutex_enter(&buf->b_evict_lock);
1900 * We are in arc_do_user_evicts(); let that function
1901 * perform the eviction.
1903 ASSERT(buf->b_data == NULL);
1904 mutex_exit(&buf->b_evict_lock);
1906 } else if (buf->b_data == NULL) {
1908 * We have already been added to the arc eviction list;
1909 * recommend eviction.
1911 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1912 mutex_exit(&buf->b_evict_lock);
1916 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1917 evict_needed = B_TRUE;
1919 mutex_exit(&buf->b_evict_lock);
1920 return (evict_needed);
1924 * Evict buffers from list until we've removed the specified number of
1925 * bytes. Move the removed buffers to the appropriate evict state.
1926 * If the recycle flag is set, then attempt to "recycle" a buffer:
1927 * - look for a buffer to evict that is `bytes' long.
1928 * - return the data block from this buffer rather than freeing it.
1929 * This flag is used by callers that are trying to make space for a
1930 * new buffer in a full arc cache.
1932 * This function makes a "best effort". It skips over any buffers
1933 * it can't get a hash_lock on, and so may not catch all candidates.
1934 * It may also return without evicting as much space as requested.
1937 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1938 arc_buf_contents_t type)
1940 arc_state_t *evicted_state;
1941 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1942 int64_t bytes_remaining;
1943 arc_buf_hdr_t *ab, *ab_prev = NULL;
1944 list_t *evicted_list, *list, *evicted_list_start, *list_start;
1945 kmutex_t *lock, *evicted_lock;
1946 kmutex_t *hash_lock;
1947 boolean_t have_lock;
1948 void *stolen = NULL;
1949 arc_buf_hdr_t marker = { 0 };
1951 static int evict_metadata_offset, evict_data_offset;
1952 int i, idx, offset, list_count, lists;
1954 ASSERT(state == arc_mru || state == arc_mfu);
1956 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1958 if (type == ARC_BUFC_METADATA) {
1960 list_count = ARC_BUFC_NUMMETADATALISTS;
1961 list_start = &state->arcs_lists[0];
1962 evicted_list_start = &evicted_state->arcs_lists[0];
1963 idx = evict_metadata_offset;
1965 offset = ARC_BUFC_NUMMETADATALISTS;
1966 list_start = &state->arcs_lists[offset];
1967 evicted_list_start = &evicted_state->arcs_lists[offset];
1968 list_count = ARC_BUFC_NUMDATALISTS;
1969 idx = evict_data_offset;
1971 bytes_remaining = evicted_state->arcs_lsize[type];
1975 list = &list_start[idx];
1976 evicted_list = &evicted_list_start[idx];
1977 lock = ARCS_LOCK(state, (offset + idx));
1978 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx));
1981 mutex_enter(evicted_lock);
1983 for (ab = list_tail(list); ab; ab = ab_prev) {
1984 ab_prev = list_prev(list, ab);
1985 bytes_remaining -= (ab->b_size * ab->b_datacnt);
1986 /* prefetch buffers have a minimum lifespan */
1987 if (HDR_IO_IN_PROGRESS(ab) ||
1988 (spa && ab->b_spa != spa) ||
1989 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1990 ddi_get_lbolt() - ab->b_arc_access <
1991 arc_min_prefetch_lifespan)) {
1995 /* "lookahead" for better eviction candidate */
1996 if (recycle && ab->b_size != bytes &&
1997 ab_prev && ab_prev->b_size == bytes)
2000 /* ignore markers */
2005 * It may take a long time to evict all the bufs requested.
2006 * To avoid blocking all arc activity, periodically drop
2007 * the arcs_mtx and give other threads a chance to run
2008 * before reacquiring the lock.
2010 * If we are looking for a buffer to recycle, we are in
2011 * the hot code path, so don't sleep.
2013 if (!recycle && count++ > arc_evict_iterations) {
2014 list_insert_after(list, ab, &marker);
2015 mutex_exit(evicted_lock);
2017 kpreempt(KPREEMPT_SYNC);
2019 mutex_enter(evicted_lock);
2020 ab_prev = list_prev(list, &marker);
2021 list_remove(list, &marker);
2026 hash_lock = HDR_LOCK(ab);
2027 have_lock = MUTEX_HELD(hash_lock);
2028 if (have_lock || mutex_tryenter(hash_lock)) {
2029 ASSERT0(refcount_count(&ab->b_refcnt));
2030 ASSERT(ab->b_datacnt > 0);
2032 arc_buf_t *buf = ab->b_buf;
2033 if (!mutex_tryenter(&buf->b_evict_lock)) {
2038 bytes_evicted += ab->b_size;
2039 if (recycle && ab->b_type == type &&
2040 ab->b_size == bytes &&
2041 !HDR_L2_WRITING(ab)) {
2042 stolen = buf->b_data;
2047 mutex_enter(&arc_eviction_mtx);
2048 arc_buf_destroy(buf,
2049 buf->b_data == stolen, FALSE);
2050 ab->b_buf = buf->b_next;
2051 buf->b_hdr = &arc_eviction_hdr;
2052 buf->b_next = arc_eviction_list;
2053 arc_eviction_list = buf;
2054 mutex_exit(&arc_eviction_mtx);
2055 mutex_exit(&buf->b_evict_lock);
2057 mutex_exit(&buf->b_evict_lock);
2058 arc_buf_destroy(buf,
2059 buf->b_data == stolen, TRUE);
2064 ARCSTAT_INCR(arcstat_evict_l2_cached,
2067 if (l2arc_write_eligible(ab->b_spa, ab)) {
2068 ARCSTAT_INCR(arcstat_evict_l2_eligible,
2072 arcstat_evict_l2_ineligible,
2077 if (ab->b_datacnt == 0) {
2078 arc_change_state(evicted_state, ab, hash_lock);
2079 ASSERT(HDR_IN_HASH_TABLE(ab));
2080 ab->b_flags |= ARC_IN_HASH_TABLE;
2081 ab->b_flags &= ~ARC_BUF_AVAILABLE;
2082 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
2085 mutex_exit(hash_lock);
2086 if (bytes >= 0 && bytes_evicted >= bytes)
2088 if (bytes_remaining > 0) {
2089 mutex_exit(evicted_lock);
2091 idx = ((idx + 1) & (list_count - 1));
2100 mutex_exit(evicted_lock);
2103 idx = ((idx + 1) & (list_count - 1));
2106 if (bytes_evicted < bytes) {
2107 if (lists < list_count)
2110 dprintf("only evicted %lld bytes from %x",
2111 (longlong_t)bytes_evicted, state);
2113 if (type == ARC_BUFC_METADATA)
2114 evict_metadata_offset = idx;
2116 evict_data_offset = idx;
2119 ARCSTAT_INCR(arcstat_evict_skip, skipped);
2122 ARCSTAT_INCR(arcstat_mutex_miss, missed);
2125 * Note: we have just evicted some data into the ghost state,
2126 * potentially putting the ghost size over the desired size. Rather
2127 * that evicting from the ghost list in this hot code path, leave
2128 * this chore to the arc_reclaim_thread().
2132 ARCSTAT_BUMP(arcstat_stolen);
2137 * Remove buffers from list until we've removed the specified number of
2138 * bytes. Destroy the buffers that are removed.
2141 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
2143 arc_buf_hdr_t *ab, *ab_prev;
2144 arc_buf_hdr_t marker = { 0 };
2145 list_t *list, *list_start;
2146 kmutex_t *hash_lock, *lock;
2147 uint64_t bytes_deleted = 0;
2148 uint64_t bufs_skipped = 0;
2150 static int evict_offset;
2151 int list_count, idx = evict_offset;
2152 int offset, lists = 0;
2154 ASSERT(GHOST_STATE(state));
2157 * data lists come after metadata lists
2159 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS];
2160 list_count = ARC_BUFC_NUMDATALISTS;
2161 offset = ARC_BUFC_NUMMETADATALISTS;
2164 list = &list_start[idx];
2165 lock = ARCS_LOCK(state, idx + offset);
2168 for (ab = list_tail(list); ab; ab = ab_prev) {
2169 ab_prev = list_prev(list, ab);
2170 if (ab->b_type > ARC_BUFC_NUMTYPES)
2171 panic("invalid ab=%p", (void *)ab);
2172 if (spa && ab->b_spa != spa)
2175 /* ignore markers */
2179 hash_lock = HDR_LOCK(ab);
2180 /* caller may be trying to modify this buffer, skip it */
2181 if (MUTEX_HELD(hash_lock))
2185 * It may take a long time to evict all the bufs requested.
2186 * To avoid blocking all arc activity, periodically drop
2187 * the arcs_mtx and give other threads a chance to run
2188 * before reacquiring the lock.
2190 if (count++ > arc_evict_iterations) {
2191 list_insert_after(list, ab, &marker);
2193 kpreempt(KPREEMPT_SYNC);
2195 ab_prev = list_prev(list, &marker);
2196 list_remove(list, &marker);
2200 if (mutex_tryenter(hash_lock)) {
2201 ASSERT(!HDR_IO_IN_PROGRESS(ab));
2202 ASSERT(ab->b_buf == NULL);
2203 ARCSTAT_BUMP(arcstat_deleted);
2204 bytes_deleted += ab->b_size;
2206 if (ab->b_l2hdr != NULL) {
2208 * This buffer is cached on the 2nd Level ARC;
2209 * don't destroy the header.
2211 arc_change_state(arc_l2c_only, ab, hash_lock);
2212 mutex_exit(hash_lock);
2214 arc_change_state(arc_anon, ab, hash_lock);
2215 mutex_exit(hash_lock);
2216 arc_hdr_destroy(ab);
2219 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
2220 if (bytes >= 0 && bytes_deleted >= bytes)
2222 } else if (bytes < 0) {
2224 * Insert a list marker and then wait for the
2225 * hash lock to become available. Once its
2226 * available, restart from where we left off.
2228 list_insert_after(list, ab, &marker);
2230 mutex_enter(hash_lock);
2231 mutex_exit(hash_lock);
2233 ab_prev = list_prev(list, &marker);
2234 list_remove(list, &marker);
2241 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1));
2244 if (lists < list_count)
2248 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] &&
2249 (bytes < 0 || bytes_deleted < bytes)) {
2250 list_start = &state->arcs_lists[0];
2251 list_count = ARC_BUFC_NUMMETADATALISTS;
2257 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2261 if (bytes_deleted < bytes)
2262 dprintf("only deleted %lld bytes from %p",
2263 (longlong_t)bytes_deleted, state);
2269 int64_t adjustment, delta;
2275 adjustment = MIN((int64_t)(arc_size - arc_c),
2276 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2279 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2280 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2281 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
2282 adjustment -= delta;
2285 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2286 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2287 (void) arc_evict(arc_mru, 0, delta, FALSE,
2295 adjustment = arc_size - arc_c;
2297 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2298 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2299 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2300 adjustment -= delta;
2303 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2304 int64_t delta = MIN(adjustment,
2305 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2306 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2311 * Adjust ghost lists
2314 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2316 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2317 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2318 arc_evict_ghost(arc_mru_ghost, 0, delta);
2322 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2324 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2325 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2326 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2331 arc_do_user_evicts(void)
2333 static arc_buf_t *tmp_arc_eviction_list;
2336 * Move list over to avoid LOR
2339 mutex_enter(&arc_eviction_mtx);
2340 tmp_arc_eviction_list = arc_eviction_list;
2341 arc_eviction_list = NULL;
2342 mutex_exit(&arc_eviction_mtx);
2344 while (tmp_arc_eviction_list != NULL) {
2345 arc_buf_t *buf = tmp_arc_eviction_list;
2346 tmp_arc_eviction_list = buf->b_next;
2347 mutex_enter(&buf->b_evict_lock);
2349 mutex_exit(&buf->b_evict_lock);
2351 if (buf->b_efunc != NULL)
2352 VERIFY(buf->b_efunc(buf) == 0);
2354 buf->b_efunc = NULL;
2355 buf->b_private = NULL;
2356 kmem_cache_free(buf_cache, buf);
2359 if (arc_eviction_list != NULL)
2364 * Flush all *evictable* data from the cache for the given spa.
2365 * NOTE: this will not touch "active" (i.e. referenced) data.
2368 arc_flush(spa_t *spa)
2373 guid = spa_load_guid(spa);
2375 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) {
2376 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2380 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) {
2381 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2385 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) {
2386 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2390 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) {
2391 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2396 arc_evict_ghost(arc_mru_ghost, guid, -1);
2397 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2399 mutex_enter(&arc_reclaim_thr_lock);
2400 arc_do_user_evicts();
2401 mutex_exit(&arc_reclaim_thr_lock);
2402 ASSERT(spa || arc_eviction_list == NULL);
2408 if (arc_c > arc_c_min) {
2412 to_free = arc_c >> arc_shrink_shift;
2414 to_free = arc_c >> arc_shrink_shift;
2416 if (arc_c > arc_c_min + to_free)
2417 atomic_add_64(&arc_c, -to_free);
2421 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2422 if (arc_c > arc_size)
2423 arc_c = MAX(arc_size, arc_c_min);
2425 arc_p = (arc_c >> 1);
2426 ASSERT(arc_c >= arc_c_min);
2427 ASSERT((int64_t)arc_p >= 0);
2430 if (arc_size > arc_c)
2434 static int needfree = 0;
2437 arc_reclaim_needed(void)
2446 * Cooperate with pagedaemon when it's time for it to scan
2447 * and reclaim some pages.
2449 if (vm_paging_needed())
2454 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2459 * check that we're out of range of the pageout scanner. It starts to
2460 * schedule paging if freemem is less than lotsfree and needfree.
2461 * lotsfree is the high-water mark for pageout, and needfree is the
2462 * number of needed free pages. We add extra pages here to make sure
2463 * the scanner doesn't start up while we're freeing memory.
2465 if (freemem < lotsfree + needfree + extra)
2469 * check to make sure that swapfs has enough space so that anon
2470 * reservations can still succeed. anon_resvmem() checks that the
2471 * availrmem is greater than swapfs_minfree, and the number of reserved
2472 * swap pages. We also add a bit of extra here just to prevent
2473 * circumstances from getting really dire.
2475 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2480 * If we're on an i386 platform, it's possible that we'll exhaust the
2481 * kernel heap space before we ever run out of available physical
2482 * memory. Most checks of the size of the heap_area compare against
2483 * tune.t_minarmem, which is the minimum available real memory that we
2484 * can have in the system. However, this is generally fixed at 25 pages
2485 * which is so low that it's useless. In this comparison, we seek to
2486 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2487 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2490 if (btop(vmem_size(heap_arena, VMEM_FREE)) <
2491 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
2495 if (kmem_used() > (kmem_size() * 3) / 4)
2500 if (spa_get_random(100) == 0)
2506 extern kmem_cache_t *zio_buf_cache[];
2507 extern kmem_cache_t *zio_data_buf_cache[];
2510 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2513 kmem_cache_t *prev_cache = NULL;
2514 kmem_cache_t *prev_data_cache = NULL;
2517 if (arc_meta_used >= arc_meta_limit) {
2519 * We are exceeding our meta-data cache limit.
2520 * Purge some DNLC entries to release holds on meta-data.
2522 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2526 * Reclaim unused memory from all kmem caches.
2533 * An aggressive reclamation will shrink the cache size as well as
2534 * reap free buffers from the arc kmem caches.
2536 if (strat == ARC_RECLAIM_AGGR)
2539 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2540 if (zio_buf_cache[i] != prev_cache) {
2541 prev_cache = zio_buf_cache[i];
2542 kmem_cache_reap_now(zio_buf_cache[i]);
2544 if (zio_data_buf_cache[i] != prev_data_cache) {
2545 prev_data_cache = zio_data_buf_cache[i];
2546 kmem_cache_reap_now(zio_data_buf_cache[i]);
2549 kmem_cache_reap_now(buf_cache);
2550 kmem_cache_reap_now(hdr_cache);
2554 arc_reclaim_thread(void *dummy __unused)
2556 clock_t growtime = 0;
2557 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2560 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2562 mutex_enter(&arc_reclaim_thr_lock);
2563 while (arc_thread_exit == 0) {
2564 if (arc_reclaim_needed()) {
2567 if (last_reclaim == ARC_RECLAIM_CONS) {
2568 last_reclaim = ARC_RECLAIM_AGGR;
2570 last_reclaim = ARC_RECLAIM_CONS;
2574 last_reclaim = ARC_RECLAIM_AGGR;
2578 /* reset the growth delay for every reclaim */
2579 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2581 if (needfree && last_reclaim == ARC_RECLAIM_CONS) {
2583 * If needfree is TRUE our vm_lowmem hook
2584 * was called and in that case we must free some
2585 * memory, so switch to aggressive mode.
2588 last_reclaim = ARC_RECLAIM_AGGR;
2590 arc_kmem_reap_now(last_reclaim);
2593 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2594 arc_no_grow = FALSE;
2599 if (arc_eviction_list != NULL)
2600 arc_do_user_evicts();
2609 /* block until needed, or one second, whichever is shorter */
2610 CALLB_CPR_SAFE_BEGIN(&cpr);
2611 (void) cv_timedwait(&arc_reclaim_thr_cv,
2612 &arc_reclaim_thr_lock, hz);
2613 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2616 arc_thread_exit = 0;
2617 cv_broadcast(&arc_reclaim_thr_cv);
2618 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2623 * Adapt arc info given the number of bytes we are trying to add and
2624 * the state that we are comming from. This function is only called
2625 * when we are adding new content to the cache.
2628 arc_adapt(int bytes, arc_state_t *state)
2631 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2633 if (state == arc_l2c_only)
2638 * Adapt the target size of the MRU list:
2639 * - if we just hit in the MRU ghost list, then increase
2640 * the target size of the MRU list.
2641 * - if we just hit in the MFU ghost list, then increase
2642 * the target size of the MFU list by decreasing the
2643 * target size of the MRU list.
2645 if (state == arc_mru_ghost) {
2646 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2647 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2648 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2650 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2651 } else if (state == arc_mfu_ghost) {
2654 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2655 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2656 mult = MIN(mult, 10);
2658 delta = MIN(bytes * mult, arc_p);
2659 arc_p = MAX(arc_p_min, arc_p - delta);
2661 ASSERT((int64_t)arc_p >= 0);
2663 if (arc_reclaim_needed()) {
2664 cv_signal(&arc_reclaim_thr_cv);
2671 if (arc_c >= arc_c_max)
2675 * If we're within (2 * maxblocksize) bytes of the target
2676 * cache size, increment the target cache size
2678 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2679 atomic_add_64(&arc_c, (int64_t)bytes);
2680 if (arc_c > arc_c_max)
2682 else if (state == arc_anon)
2683 atomic_add_64(&arc_p, (int64_t)bytes);
2687 ASSERT((int64_t)arc_p >= 0);
2691 * Check if the cache has reached its limits and eviction is required
2695 arc_evict_needed(arc_buf_contents_t type)
2697 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2703 * If zio data pages are being allocated out of a separate heap segment,
2704 * then enforce that the size of available vmem for this area remains
2705 * above about 1/32nd free.
2707 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2708 vmem_size(zio_arena, VMEM_FREE) <
2709 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2714 if (arc_reclaim_needed())
2717 return (arc_size > arc_c);
2721 * The buffer, supplied as the first argument, needs a data block.
2722 * So, if we are at cache max, determine which cache should be victimized.
2723 * We have the following cases:
2725 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2726 * In this situation if we're out of space, but the resident size of the MFU is
2727 * under the limit, victimize the MFU cache to satisfy this insertion request.
2729 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2730 * Here, we've used up all of the available space for the MRU, so we need to
2731 * evict from our own cache instead. Evict from the set of resident MRU
2734 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2735 * c minus p represents the MFU space in the cache, since p is the size of the
2736 * cache that is dedicated to the MRU. In this situation there's still space on
2737 * the MFU side, so the MRU side needs to be victimized.
2739 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2740 * MFU's resident set is consuming more space than it has been allotted. In
2741 * this situation, we must victimize our own cache, the MFU, for this insertion.
2744 arc_get_data_buf(arc_buf_t *buf)
2746 arc_state_t *state = buf->b_hdr->b_state;
2747 uint64_t size = buf->b_hdr->b_size;
2748 arc_buf_contents_t type = buf->b_hdr->b_type;
2750 arc_adapt(size, state);
2753 * We have not yet reached cache maximum size,
2754 * just allocate a new buffer.
2756 if (!arc_evict_needed(type)) {
2757 if (type == ARC_BUFC_METADATA) {
2758 buf->b_data = zio_buf_alloc(size);
2759 arc_space_consume(size, ARC_SPACE_DATA);
2761 ASSERT(type == ARC_BUFC_DATA);
2762 buf->b_data = zio_data_buf_alloc(size);
2763 ARCSTAT_INCR(arcstat_data_size, size);
2764 atomic_add_64(&arc_size, size);
2770 * If we are prefetching from the mfu ghost list, this buffer
2771 * will end up on the mru list; so steal space from there.
2773 if (state == arc_mfu_ghost)
2774 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2775 else if (state == arc_mru_ghost)
2778 if (state == arc_mru || state == arc_anon) {
2779 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2780 state = (arc_mfu->arcs_lsize[type] >= size &&
2781 arc_p > mru_used) ? arc_mfu : arc_mru;
2784 uint64_t mfu_space = arc_c - arc_p;
2785 state = (arc_mru->arcs_lsize[type] >= size &&
2786 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2788 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2789 if (type == ARC_BUFC_METADATA) {
2790 buf->b_data = zio_buf_alloc(size);
2791 arc_space_consume(size, ARC_SPACE_DATA);
2793 ASSERT(type == ARC_BUFC_DATA);
2794 buf->b_data = zio_data_buf_alloc(size);
2795 ARCSTAT_INCR(arcstat_data_size, size);
2796 atomic_add_64(&arc_size, size);
2798 ARCSTAT_BUMP(arcstat_recycle_miss);
2800 ASSERT(buf->b_data != NULL);
2803 * Update the state size. Note that ghost states have a
2804 * "ghost size" and so don't need to be updated.
2806 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2807 arc_buf_hdr_t *hdr = buf->b_hdr;
2809 atomic_add_64(&hdr->b_state->arcs_size, size);
2810 if (list_link_active(&hdr->b_arc_node)) {
2811 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2812 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2815 * If we are growing the cache, and we are adding anonymous
2816 * data, and we have outgrown arc_p, update arc_p
2818 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2819 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2820 arc_p = MIN(arc_c, arc_p + size);
2822 ARCSTAT_BUMP(arcstat_allocated);
2826 * This routine is called whenever a buffer is accessed.
2827 * NOTE: the hash lock is dropped in this function.
2830 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2834 ASSERT(MUTEX_HELD(hash_lock));
2836 if (buf->b_state == arc_anon) {
2838 * This buffer is not in the cache, and does not
2839 * appear in our "ghost" list. Add the new buffer
2843 ASSERT(buf->b_arc_access == 0);
2844 buf->b_arc_access = ddi_get_lbolt();
2845 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2846 arc_change_state(arc_mru, buf, hash_lock);
2848 } else if (buf->b_state == arc_mru) {
2849 now = ddi_get_lbolt();
2852 * If this buffer is here because of a prefetch, then either:
2853 * - clear the flag if this is a "referencing" read
2854 * (any subsequent access will bump this into the MFU state).
2856 * - move the buffer to the head of the list if this is
2857 * another prefetch (to make it less likely to be evicted).
2859 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2860 if (refcount_count(&buf->b_refcnt) == 0) {
2861 ASSERT(list_link_active(&buf->b_arc_node));
2863 buf->b_flags &= ~ARC_PREFETCH;
2864 ARCSTAT_BUMP(arcstat_mru_hits);
2866 buf->b_arc_access = now;
2871 * This buffer has been "accessed" only once so far,
2872 * but it is still in the cache. Move it to the MFU
2875 if (now > buf->b_arc_access + ARC_MINTIME) {
2877 * More than 125ms have passed since we
2878 * instantiated this buffer. Move it to the
2879 * most frequently used state.
2881 buf->b_arc_access = now;
2882 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2883 arc_change_state(arc_mfu, buf, hash_lock);
2885 ARCSTAT_BUMP(arcstat_mru_hits);
2886 } else if (buf->b_state == arc_mru_ghost) {
2887 arc_state_t *new_state;
2889 * This buffer has been "accessed" recently, but
2890 * was evicted from the cache. Move it to the
2894 if (buf->b_flags & ARC_PREFETCH) {
2895 new_state = arc_mru;
2896 if (refcount_count(&buf->b_refcnt) > 0)
2897 buf->b_flags &= ~ARC_PREFETCH;
2898 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2900 new_state = arc_mfu;
2901 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2904 buf->b_arc_access = ddi_get_lbolt();
2905 arc_change_state(new_state, buf, hash_lock);
2907 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2908 } else if (buf->b_state == arc_mfu) {
2910 * This buffer has been accessed more than once and is
2911 * still in the cache. Keep it in the MFU state.
2913 * NOTE: an add_reference() that occurred when we did
2914 * the arc_read() will have kicked this off the list.
2915 * If it was a prefetch, we will explicitly move it to
2916 * the head of the list now.
2918 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2919 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2920 ASSERT(list_link_active(&buf->b_arc_node));
2922 ARCSTAT_BUMP(arcstat_mfu_hits);
2923 buf->b_arc_access = ddi_get_lbolt();
2924 } else if (buf->b_state == arc_mfu_ghost) {
2925 arc_state_t *new_state = arc_mfu;
2927 * This buffer has been accessed more than once but has
2928 * been evicted from the cache. Move it back to the
2932 if (buf->b_flags & ARC_PREFETCH) {
2934 * This is a prefetch access...
2935 * move this block back to the MRU state.
2937 ASSERT0(refcount_count(&buf->b_refcnt));
2938 new_state = arc_mru;
2941 buf->b_arc_access = ddi_get_lbolt();
2942 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2943 arc_change_state(new_state, buf, hash_lock);
2945 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2946 } else if (buf->b_state == arc_l2c_only) {
2948 * This buffer is on the 2nd Level ARC.
2951 buf->b_arc_access = ddi_get_lbolt();
2952 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2953 arc_change_state(arc_mfu, buf, hash_lock);
2955 ASSERT(!"invalid arc state");
2959 /* a generic arc_done_func_t which you can use */
2962 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2964 if (zio == NULL || zio->io_error == 0)
2965 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2966 VERIFY(arc_buf_remove_ref(buf, arg));
2969 /* a generic arc_done_func_t */
2971 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2973 arc_buf_t **bufp = arg;
2974 if (zio && zio->io_error) {
2975 VERIFY(arc_buf_remove_ref(buf, arg));
2979 ASSERT(buf->b_data);
2984 arc_read_done(zio_t *zio)
2986 arc_buf_hdr_t *hdr, *found;
2988 arc_buf_t *abuf; /* buffer we're assigning to callback */
2989 kmutex_t *hash_lock;
2990 arc_callback_t *callback_list, *acb;
2991 int freeable = FALSE;
2993 buf = zio->io_private;
2997 * The hdr was inserted into hash-table and removed from lists
2998 * prior to starting I/O. We should find this header, since
2999 * it's in the hash table, and it should be legit since it's
3000 * not possible to evict it during the I/O. The only possible
3001 * reason for it not to be found is if we were freed during the
3004 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
3007 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
3008 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
3009 (found == hdr && HDR_L2_READING(hdr)));
3011 hdr->b_flags &= ~ARC_L2_EVICTED;
3012 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
3013 hdr->b_flags &= ~ARC_L2CACHE;
3015 /* byteswap if necessary */
3016 callback_list = hdr->b_acb;
3017 ASSERT(callback_list != NULL);
3018 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
3019 dmu_object_byteswap_t bswap =
3020 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
3021 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
3022 byteswap_uint64_array :
3023 dmu_ot_byteswap[bswap].ob_func;
3024 func(buf->b_data, hdr->b_size);
3027 arc_cksum_compute(buf, B_FALSE);
3030 #endif /* illumos */
3032 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
3034 * Only call arc_access on anonymous buffers. This is because
3035 * if we've issued an I/O for an evicted buffer, we've already
3036 * called arc_access (to prevent any simultaneous readers from
3037 * getting confused).
3039 arc_access(hdr, hash_lock);
3042 /* create copies of the data buffer for the callers */
3044 for (acb = callback_list; acb; acb = acb->acb_next) {
3045 if (acb->acb_done) {
3047 ARCSTAT_BUMP(arcstat_duplicate_reads);
3048 abuf = arc_buf_clone(buf);
3050 acb->acb_buf = abuf;
3055 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3056 ASSERT(!HDR_BUF_AVAILABLE(hdr));
3058 ASSERT(buf->b_efunc == NULL);
3059 ASSERT(hdr->b_datacnt == 1);
3060 hdr->b_flags |= ARC_BUF_AVAILABLE;
3063 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
3065 if (zio->io_error != 0) {
3066 hdr->b_flags |= ARC_IO_ERROR;
3067 if (hdr->b_state != arc_anon)
3068 arc_change_state(arc_anon, hdr, hash_lock);
3069 if (HDR_IN_HASH_TABLE(hdr))
3070 buf_hash_remove(hdr);
3071 freeable = refcount_is_zero(&hdr->b_refcnt);
3075 * Broadcast before we drop the hash_lock to avoid the possibility
3076 * that the hdr (and hence the cv) might be freed before we get to
3077 * the cv_broadcast().
3079 cv_broadcast(&hdr->b_cv);
3082 mutex_exit(hash_lock);
3085 * This block was freed while we waited for the read to
3086 * complete. It has been removed from the hash table and
3087 * moved to the anonymous state (so that it won't show up
3090 ASSERT3P(hdr->b_state, ==, arc_anon);
3091 freeable = refcount_is_zero(&hdr->b_refcnt);
3094 /* execute each callback and free its structure */
3095 while ((acb = callback_list) != NULL) {
3097 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3099 if (acb->acb_zio_dummy != NULL) {
3100 acb->acb_zio_dummy->io_error = zio->io_error;
3101 zio_nowait(acb->acb_zio_dummy);
3104 callback_list = acb->acb_next;
3105 kmem_free(acb, sizeof (arc_callback_t));
3109 arc_hdr_destroy(hdr);
3113 * "Read" the block block at the specified DVA (in bp) via the
3114 * cache. If the block is found in the cache, invoke the provided
3115 * callback immediately and return. Note that the `zio' parameter
3116 * in the callback will be NULL in this case, since no IO was
3117 * required. If the block is not in the cache pass the read request
3118 * on to the spa with a substitute callback function, so that the
3119 * requested block will be added to the cache.
3121 * If a read request arrives for a block that has a read in-progress,
3122 * either wait for the in-progress read to complete (and return the
3123 * results); or, if this is a read with a "done" func, add a record
3124 * to the read to invoke the "done" func when the read completes,
3125 * and return; or just return.
3127 * arc_read_done() will invoke all the requested "done" functions
3128 * for readers of this block.
3131 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3132 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags,
3133 const zbookmark_t *zb)
3136 arc_buf_t *buf = NULL;
3137 kmutex_t *hash_lock;
3139 uint64_t guid = spa_load_guid(spa);
3142 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3144 if (hdr && hdr->b_datacnt > 0) {
3146 *arc_flags |= ARC_CACHED;
3148 if (HDR_IO_IN_PROGRESS(hdr)) {
3150 if (*arc_flags & ARC_WAIT) {
3151 cv_wait(&hdr->b_cv, hash_lock);
3152 mutex_exit(hash_lock);
3155 ASSERT(*arc_flags & ARC_NOWAIT);
3158 arc_callback_t *acb = NULL;
3160 acb = kmem_zalloc(sizeof (arc_callback_t),
3162 acb->acb_done = done;
3163 acb->acb_private = private;
3165 acb->acb_zio_dummy = zio_null(pio,
3166 spa, NULL, NULL, NULL, zio_flags);
3168 ASSERT(acb->acb_done != NULL);
3169 acb->acb_next = hdr->b_acb;
3171 add_reference(hdr, hash_lock, private);
3172 mutex_exit(hash_lock);
3175 mutex_exit(hash_lock);
3179 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3182 add_reference(hdr, hash_lock, private);
3184 * If this block is already in use, create a new
3185 * copy of the data so that we will be guaranteed
3186 * that arc_release() will always succeed.
3190 ASSERT(buf->b_data);
3191 if (HDR_BUF_AVAILABLE(hdr)) {
3192 ASSERT(buf->b_efunc == NULL);
3193 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3195 buf = arc_buf_clone(buf);
3198 } else if (*arc_flags & ARC_PREFETCH &&
3199 refcount_count(&hdr->b_refcnt) == 0) {
3200 hdr->b_flags |= ARC_PREFETCH;
3202 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
3203 arc_access(hdr, hash_lock);
3204 if (*arc_flags & ARC_L2CACHE)
3205 hdr->b_flags |= ARC_L2CACHE;
3206 if (*arc_flags & ARC_L2COMPRESS)
3207 hdr->b_flags |= ARC_L2COMPRESS;
3208 mutex_exit(hash_lock);
3209 ARCSTAT_BUMP(arcstat_hits);
3210 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3211 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3212 data, metadata, hits);
3215 done(NULL, buf, private);
3217 uint64_t size = BP_GET_LSIZE(bp);
3218 arc_callback_t *acb;
3221 boolean_t devw = B_FALSE;
3224 /* this block is not in the cache */
3225 arc_buf_hdr_t *exists;
3226 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
3227 buf = arc_buf_alloc(spa, size, private, type);
3229 hdr->b_dva = *BP_IDENTITY(bp);
3230 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3231 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3232 exists = buf_hash_insert(hdr, &hash_lock);
3234 /* somebody beat us to the hash insert */
3235 mutex_exit(hash_lock);
3236 buf_discard_identity(hdr);
3237 (void) arc_buf_remove_ref(buf, private);
3238 goto top; /* restart the IO request */
3240 /* if this is a prefetch, we don't have a reference */
3241 if (*arc_flags & ARC_PREFETCH) {
3242 (void) remove_reference(hdr, hash_lock,
3244 hdr->b_flags |= ARC_PREFETCH;
3246 if (*arc_flags & ARC_L2CACHE)
3247 hdr->b_flags |= ARC_L2CACHE;
3248 if (*arc_flags & ARC_L2COMPRESS)
3249 hdr->b_flags |= ARC_L2COMPRESS;
3250 if (BP_GET_LEVEL(bp) > 0)
3251 hdr->b_flags |= ARC_INDIRECT;
3253 /* this block is in the ghost cache */
3254 ASSERT(GHOST_STATE(hdr->b_state));
3255 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3256 ASSERT0(refcount_count(&hdr->b_refcnt));
3257 ASSERT(hdr->b_buf == NULL);
3259 /* if this is a prefetch, we don't have a reference */
3260 if (*arc_flags & ARC_PREFETCH)
3261 hdr->b_flags |= ARC_PREFETCH;
3263 add_reference(hdr, hash_lock, private);
3264 if (*arc_flags & ARC_L2CACHE)
3265 hdr->b_flags |= ARC_L2CACHE;
3266 if (*arc_flags & ARC_L2COMPRESS)
3267 hdr->b_flags |= ARC_L2COMPRESS;
3268 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3271 buf->b_efunc = NULL;
3272 buf->b_private = NULL;
3275 ASSERT(hdr->b_datacnt == 0);
3277 arc_get_data_buf(buf);
3278 arc_access(hdr, hash_lock);
3281 ASSERT(!GHOST_STATE(hdr->b_state));
3283 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
3284 acb->acb_done = done;
3285 acb->acb_private = private;
3287 ASSERT(hdr->b_acb == NULL);
3289 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3291 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3292 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3293 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3294 addr = hdr->b_l2hdr->b_daddr;
3296 * Lock out device removal.
3298 if (vdev_is_dead(vd) ||
3299 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3303 mutex_exit(hash_lock);
3306 * At this point, we have a level 1 cache miss. Try again in
3307 * L2ARC if possible.
3309 ASSERT3U(hdr->b_size, ==, size);
3310 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3311 uint64_t, size, zbookmark_t *, zb);
3312 ARCSTAT_BUMP(arcstat_misses);
3313 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3314 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3315 data, metadata, misses);
3317 curthread->td_ru.ru_inblock++;
3320 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3322 * Read from the L2ARC if the following are true:
3323 * 1. The L2ARC vdev was previously cached.
3324 * 2. This buffer still has L2ARC metadata.
3325 * 3. This buffer isn't currently writing to the L2ARC.
3326 * 4. The L2ARC entry wasn't evicted, which may
3327 * also have invalidated the vdev.
3328 * 5. This isn't prefetch and l2arc_noprefetch is set.
3330 if (hdr->b_l2hdr != NULL &&
3331 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3332 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3333 l2arc_read_callback_t *cb;
3335 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3336 ARCSTAT_BUMP(arcstat_l2_hits);
3338 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3340 cb->l2rcb_buf = buf;
3341 cb->l2rcb_spa = spa;
3344 cb->l2rcb_flags = zio_flags;
3345 cb->l2rcb_compress = hdr->b_l2hdr->b_compress;
3347 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
3348 addr + size < vd->vdev_psize -
3349 VDEV_LABEL_END_SIZE);
3352 * l2arc read. The SCL_L2ARC lock will be
3353 * released by l2arc_read_done().
3354 * Issue a null zio if the underlying buffer
3355 * was squashed to zero size by compression.
3357 if (hdr->b_l2hdr->b_compress ==
3358 ZIO_COMPRESS_EMPTY) {
3359 rzio = zio_null(pio, spa, vd,
3360 l2arc_read_done, cb,
3361 zio_flags | ZIO_FLAG_DONT_CACHE |
3363 ZIO_FLAG_DONT_PROPAGATE |
3364 ZIO_FLAG_DONT_RETRY);
3366 rzio = zio_read_phys(pio, vd, addr,
3367 hdr->b_l2hdr->b_asize,
3368 buf->b_data, ZIO_CHECKSUM_OFF,
3369 l2arc_read_done, cb, priority,
3370 zio_flags | ZIO_FLAG_DONT_CACHE |
3372 ZIO_FLAG_DONT_PROPAGATE |
3373 ZIO_FLAG_DONT_RETRY, B_FALSE);
3375 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3377 ARCSTAT_INCR(arcstat_l2_read_bytes,
3378 hdr->b_l2hdr->b_asize);
3380 if (*arc_flags & ARC_NOWAIT) {
3385 ASSERT(*arc_flags & ARC_WAIT);
3386 if (zio_wait(rzio) == 0)
3389 /* l2arc read error; goto zio_read() */
3391 DTRACE_PROBE1(l2arc__miss,
3392 arc_buf_hdr_t *, hdr);
3393 ARCSTAT_BUMP(arcstat_l2_misses);
3394 if (HDR_L2_WRITING(hdr))
3395 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3396 spa_config_exit(spa, SCL_L2ARC, vd);
3400 spa_config_exit(spa, SCL_L2ARC, vd);
3401 if (l2arc_ndev != 0) {
3402 DTRACE_PROBE1(l2arc__miss,
3403 arc_buf_hdr_t *, hdr);
3404 ARCSTAT_BUMP(arcstat_l2_misses);
3408 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3409 arc_read_done, buf, priority, zio_flags, zb);
3411 if (*arc_flags & ARC_WAIT)
3412 return (zio_wait(rzio));
3414 ASSERT(*arc_flags & ARC_NOWAIT);
3421 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3423 ASSERT(buf->b_hdr != NULL);
3424 ASSERT(buf->b_hdr->b_state != arc_anon);
3425 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3426 ASSERT(buf->b_efunc == NULL);
3427 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3429 buf->b_efunc = func;
3430 buf->b_private = private;
3434 * This is used by the DMU to let the ARC know that a buffer is
3435 * being evicted, so the ARC should clean up. If this arc buf
3436 * is not yet in the evicted state, it will be put there.
3439 arc_buf_evict(arc_buf_t *buf)
3442 kmutex_t *hash_lock;
3444 list_t *list, *evicted_list;
3445 kmutex_t *lock, *evicted_lock;
3447 mutex_enter(&buf->b_evict_lock);
3451 * We are in arc_do_user_evicts().
3453 ASSERT(buf->b_data == NULL);
3454 mutex_exit(&buf->b_evict_lock);
3456 } else if (buf->b_data == NULL) {
3457 arc_buf_t copy = *buf; /* structure assignment */
3459 * We are on the eviction list; process this buffer now
3460 * but let arc_do_user_evicts() do the reaping.
3462 buf->b_efunc = NULL;
3463 mutex_exit(&buf->b_evict_lock);
3464 VERIFY(copy.b_efunc(©) == 0);
3467 hash_lock = HDR_LOCK(hdr);
3468 mutex_enter(hash_lock);
3470 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3472 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3473 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3476 * Pull this buffer off of the hdr
3479 while (*bufp != buf)
3480 bufp = &(*bufp)->b_next;
3481 *bufp = buf->b_next;
3483 ASSERT(buf->b_data != NULL);
3484 arc_buf_destroy(buf, FALSE, FALSE);
3486 if (hdr->b_datacnt == 0) {
3487 arc_state_t *old_state = hdr->b_state;
3488 arc_state_t *evicted_state;
3490 ASSERT(hdr->b_buf == NULL);
3491 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3494 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3496 get_buf_info(hdr, old_state, &list, &lock);
3497 get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock);
3499 mutex_enter(evicted_lock);
3501 arc_change_state(evicted_state, hdr, hash_lock);
3502 ASSERT(HDR_IN_HASH_TABLE(hdr));
3503 hdr->b_flags |= ARC_IN_HASH_TABLE;
3504 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3506 mutex_exit(evicted_lock);
3509 mutex_exit(hash_lock);
3510 mutex_exit(&buf->b_evict_lock);
3512 VERIFY(buf->b_efunc(buf) == 0);
3513 buf->b_efunc = NULL;
3514 buf->b_private = NULL;
3517 kmem_cache_free(buf_cache, buf);
3522 * Release this buffer from the cache, making it an anonymous buffer. This
3523 * must be done after a read and prior to modifying the buffer contents.
3524 * If the buffer has more than one reference, we must make
3525 * a new hdr for the buffer.
3528 arc_release(arc_buf_t *buf, void *tag)
3531 kmutex_t *hash_lock = NULL;
3532 l2arc_buf_hdr_t *l2hdr;
3536 * It would be nice to assert that if it's DMU metadata (level >
3537 * 0 || it's the dnode file), then it must be syncing context.
3538 * But we don't know that information at this level.
3541 mutex_enter(&buf->b_evict_lock);
3544 /* this buffer is not on any list */
3545 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3547 if (hdr->b_state == arc_anon) {
3548 /* this buffer is already released */
3549 ASSERT(buf->b_efunc == NULL);
3551 hash_lock = HDR_LOCK(hdr);
3552 mutex_enter(hash_lock);
3554 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3557 l2hdr = hdr->b_l2hdr;
3559 mutex_enter(&l2arc_buflist_mtx);
3560 hdr->b_l2hdr = NULL;
3561 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3563 buf_size = hdr->b_size;
3566 * Do we have more than one buf?
3568 if (hdr->b_datacnt > 1) {
3569 arc_buf_hdr_t *nhdr;
3571 uint64_t blksz = hdr->b_size;
3572 uint64_t spa = hdr->b_spa;
3573 arc_buf_contents_t type = hdr->b_type;
3574 uint32_t flags = hdr->b_flags;
3576 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3578 * Pull the data off of this hdr and attach it to
3579 * a new anonymous hdr.
3581 (void) remove_reference(hdr, hash_lock, tag);
3583 while (*bufp != buf)
3584 bufp = &(*bufp)->b_next;
3585 *bufp = buf->b_next;
3588 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3589 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3590 if (refcount_is_zero(&hdr->b_refcnt)) {
3591 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3592 ASSERT3U(*size, >=, hdr->b_size);
3593 atomic_add_64(size, -hdr->b_size);
3597 * We're releasing a duplicate user data buffer, update
3598 * our statistics accordingly.
3600 if (hdr->b_type == ARC_BUFC_DATA) {
3601 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3602 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3605 hdr->b_datacnt -= 1;
3606 arc_cksum_verify(buf);
3608 arc_buf_unwatch(buf);
3609 #endif /* illumos */
3611 mutex_exit(hash_lock);
3613 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3614 nhdr->b_size = blksz;
3616 nhdr->b_type = type;
3618 nhdr->b_state = arc_anon;
3619 nhdr->b_arc_access = 0;
3620 nhdr->b_flags = flags & ARC_L2_WRITING;
3621 nhdr->b_l2hdr = NULL;
3622 nhdr->b_datacnt = 1;
3623 nhdr->b_freeze_cksum = NULL;
3624 (void) refcount_add(&nhdr->b_refcnt, tag);
3626 mutex_exit(&buf->b_evict_lock);
3627 atomic_add_64(&arc_anon->arcs_size, blksz);
3629 mutex_exit(&buf->b_evict_lock);
3630 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3631 ASSERT(!list_link_active(&hdr->b_arc_node));
3632 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3633 if (hdr->b_state != arc_anon)
3634 arc_change_state(arc_anon, hdr, hash_lock);
3635 hdr->b_arc_access = 0;
3637 mutex_exit(hash_lock);
3639 buf_discard_identity(hdr);
3642 buf->b_efunc = NULL;
3643 buf->b_private = NULL;
3646 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
3647 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr,
3649 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3650 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3651 mutex_exit(&l2arc_buflist_mtx);
3656 arc_released(arc_buf_t *buf)
3660 mutex_enter(&buf->b_evict_lock);
3661 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3662 mutex_exit(&buf->b_evict_lock);
3667 arc_has_callback(arc_buf_t *buf)
3671 mutex_enter(&buf->b_evict_lock);
3672 callback = (buf->b_efunc != NULL);
3673 mutex_exit(&buf->b_evict_lock);
3679 arc_referenced(arc_buf_t *buf)
3683 mutex_enter(&buf->b_evict_lock);
3684 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3685 mutex_exit(&buf->b_evict_lock);
3686 return (referenced);
3691 arc_write_ready(zio_t *zio)
3693 arc_write_callback_t *callback = zio->io_private;
3694 arc_buf_t *buf = callback->awcb_buf;
3695 arc_buf_hdr_t *hdr = buf->b_hdr;
3697 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3698 callback->awcb_ready(zio, buf, callback->awcb_private);
3701 * If the IO is already in progress, then this is a re-write
3702 * attempt, so we need to thaw and re-compute the cksum.
3703 * It is the responsibility of the callback to handle the
3704 * accounting for any re-write attempt.
3706 if (HDR_IO_IN_PROGRESS(hdr)) {
3707 mutex_enter(&hdr->b_freeze_lock);
3708 if (hdr->b_freeze_cksum != NULL) {
3709 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3710 hdr->b_freeze_cksum = NULL;
3712 mutex_exit(&hdr->b_freeze_lock);
3714 arc_cksum_compute(buf, B_FALSE);
3715 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3719 * The SPA calls this callback for each physical write that happens on behalf
3720 * of a logical write. See the comment in dbuf_write_physdone() for details.
3723 arc_write_physdone(zio_t *zio)
3725 arc_write_callback_t *cb = zio->io_private;
3726 if (cb->awcb_physdone != NULL)
3727 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
3731 arc_write_done(zio_t *zio)
3733 arc_write_callback_t *callback = zio->io_private;
3734 arc_buf_t *buf = callback->awcb_buf;
3735 arc_buf_hdr_t *hdr = buf->b_hdr;
3737 ASSERT(hdr->b_acb == NULL);
3739 if (zio->io_error == 0) {
3740 if (BP_IS_HOLE(zio->io_bp)) {
3741 buf_discard_identity(hdr);
3743 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3744 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3745 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3748 ASSERT(BUF_EMPTY(hdr));
3752 * If the block to be written was all-zero, we may have
3753 * compressed it away. In this case no write was performed
3754 * so there will be no dva/birth/checksum. The buffer must
3755 * therefore remain anonymous (and uncached).
3757 if (!BUF_EMPTY(hdr)) {
3758 arc_buf_hdr_t *exists;
3759 kmutex_t *hash_lock;
3761 ASSERT(zio->io_error == 0);
3763 arc_cksum_verify(buf);
3765 exists = buf_hash_insert(hdr, &hash_lock);
3768 * This can only happen if we overwrite for
3769 * sync-to-convergence, because we remove
3770 * buffers from the hash table when we arc_free().
3772 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3773 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3774 panic("bad overwrite, hdr=%p exists=%p",
3775 (void *)hdr, (void *)exists);
3776 ASSERT(refcount_is_zero(&exists->b_refcnt));
3777 arc_change_state(arc_anon, exists, hash_lock);
3778 mutex_exit(hash_lock);
3779 arc_hdr_destroy(exists);
3780 exists = buf_hash_insert(hdr, &hash_lock);
3781 ASSERT3P(exists, ==, NULL);
3782 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3784 ASSERT(zio->io_prop.zp_nopwrite);
3785 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3786 panic("bad nopwrite, hdr=%p exists=%p",
3787 (void *)hdr, (void *)exists);
3790 ASSERT(hdr->b_datacnt == 1);
3791 ASSERT(hdr->b_state == arc_anon);
3792 ASSERT(BP_GET_DEDUP(zio->io_bp));
3793 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3796 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3797 /* if it's not anon, we are doing a scrub */
3798 if (!exists && hdr->b_state == arc_anon)
3799 arc_access(hdr, hash_lock);
3800 mutex_exit(hash_lock);
3802 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3805 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3806 callback->awcb_done(zio, buf, callback->awcb_private);
3808 kmem_free(callback, sizeof (arc_write_callback_t));
3812 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3813 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
3814 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
3815 arc_done_func_t *done, void *private, zio_priority_t priority,
3816 int zio_flags, const zbookmark_t *zb)
3818 arc_buf_hdr_t *hdr = buf->b_hdr;
3819 arc_write_callback_t *callback;
3822 ASSERT(ready != NULL);
3823 ASSERT(done != NULL);
3824 ASSERT(!HDR_IO_ERROR(hdr));
3825 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3826 ASSERT(hdr->b_acb == NULL);
3828 hdr->b_flags |= ARC_L2CACHE;
3830 hdr->b_flags |= ARC_L2COMPRESS;
3831 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3832 callback->awcb_ready = ready;
3833 callback->awcb_physdone = physdone;
3834 callback->awcb_done = done;
3835 callback->awcb_private = private;
3836 callback->awcb_buf = buf;
3838 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3839 arc_write_ready, arc_write_physdone, arc_write_done, callback,
3840 priority, zio_flags, zb);
3846 arc_memory_throttle(uint64_t reserve, uint64_t txg)
3849 uint64_t available_memory =
3850 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count);
3851 static uint64_t page_load = 0;
3852 static uint64_t last_txg = 0;
3857 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3861 if (cnt.v_free_count + cnt.v_cache_count >
3862 (uint64_t)physmem * arc_lotsfree_percent / 100)
3865 if (txg > last_txg) {
3870 * If we are in pageout, we know that memory is already tight,
3871 * the arc is already going to be evicting, so we just want to
3872 * continue to let page writes occur as quickly as possible.
3874 if (curproc == pageproc) {
3875 if (page_load > available_memory / 4)
3876 return (SET_ERROR(ERESTART));
3877 /* Note: reserve is inflated, so we deflate */
3878 page_load += reserve / 8;
3880 } else if (page_load > 0 && arc_reclaim_needed()) {
3881 /* memory is low, delay before restarting */
3882 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3883 return (SET_ERROR(EAGAIN));
3891 arc_tempreserve_clear(uint64_t reserve)
3893 atomic_add_64(&arc_tempreserve, -reserve);
3894 ASSERT((int64_t)arc_tempreserve >= 0);
3898 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3903 if (reserve > arc_c/4 && !arc_no_grow)
3904 arc_c = MIN(arc_c_max, reserve * 4);
3905 if (reserve > arc_c)
3906 return (SET_ERROR(ENOMEM));
3909 * Don't count loaned bufs as in flight dirty data to prevent long
3910 * network delays from blocking transactions that are ready to be
3911 * assigned to a txg.
3913 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3916 * Writes will, almost always, require additional memory allocations
3917 * in order to compress/encrypt/etc the data. We therefore need to
3918 * make sure that there is sufficient available memory for this.
3920 error = arc_memory_throttle(reserve, txg);
3925 * Throttle writes when the amount of dirty data in the cache
3926 * gets too large. We try to keep the cache less than half full
3927 * of dirty blocks so that our sync times don't grow too large.
3928 * Note: if two requests come in concurrently, we might let them
3929 * both succeed, when one of them should fail. Not a huge deal.
3932 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3933 anon_size > arc_c / 4) {
3934 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3935 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3936 arc_tempreserve>>10,
3937 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3938 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3939 reserve>>10, arc_c>>10);
3940 return (SET_ERROR(ERESTART));
3942 atomic_add_64(&arc_tempreserve, reserve);
3946 static kmutex_t arc_lowmem_lock;
3948 static eventhandler_tag arc_event_lowmem = NULL;
3951 arc_lowmem(void *arg __unused, int howto __unused)
3954 /* Serialize access via arc_lowmem_lock. */
3955 mutex_enter(&arc_lowmem_lock);
3956 mutex_enter(&arc_reclaim_thr_lock);
3958 cv_signal(&arc_reclaim_thr_cv);
3961 * It is unsafe to block here in arbitrary threads, because we can come
3962 * here from ARC itself and may hold ARC locks and thus risk a deadlock
3963 * with ARC reclaim thread.
3965 if (curproc == pageproc) {
3967 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0);
3969 mutex_exit(&arc_reclaim_thr_lock);
3970 mutex_exit(&arc_lowmem_lock);
3977 int i, prefetch_tunable_set = 0;
3979 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3980 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3981 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL);
3983 /* Convert seconds to clock ticks */
3984 arc_min_prefetch_lifespan = 1 * hz;
3986 /* Start out with 1/8 of all memory */
3987 arc_c = kmem_size() / 8;
3992 * On architectures where the physical memory can be larger
3993 * than the addressable space (intel in 32-bit mode), we may
3994 * need to limit the cache to 1/8 of VM size.
3996 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3999 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
4000 arc_c_min = MAX(arc_c / 4, 64<<18);
4001 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
4002 if (arc_c * 8 >= 1<<30)
4003 arc_c_max = (arc_c * 8) - (1<<30);
4005 arc_c_max = arc_c_min;
4006 arc_c_max = MAX(arc_c * 5, arc_c_max);
4010 * Allow the tunables to override our calculations if they are
4011 * reasonable (ie. over 16MB)
4013 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size())
4014 arc_c_max = zfs_arc_max;
4015 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max)
4016 arc_c_min = zfs_arc_min;
4020 arc_p = (arc_c >> 1);
4022 /* limit meta-data to 1/4 of the arc capacity */
4023 arc_meta_limit = arc_c_max / 4;
4025 /* Allow the tunable to override if it is reasonable */
4026 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
4027 arc_meta_limit = zfs_arc_meta_limit;
4029 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
4030 arc_c_min = arc_meta_limit / 2;
4032 if (zfs_arc_grow_retry > 0)
4033 arc_grow_retry = zfs_arc_grow_retry;
4035 if (zfs_arc_shrink_shift > 0)
4036 arc_shrink_shift = zfs_arc_shrink_shift;
4038 if (zfs_arc_p_min_shift > 0)
4039 arc_p_min_shift = zfs_arc_p_min_shift;
4041 /* if kmem_flags are set, lets try to use less memory */
4042 if (kmem_debugging())
4044 if (arc_c < arc_c_min)
4047 zfs_arc_min = arc_c_min;
4048 zfs_arc_max = arc_c_max;
4050 arc_anon = &ARC_anon;
4052 arc_mru_ghost = &ARC_mru_ghost;
4054 arc_mfu_ghost = &ARC_mfu_ghost;
4055 arc_l2c_only = &ARC_l2c_only;
4058 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
4059 mutex_init(&arc_anon->arcs_locks[i].arcs_lock,
4060 NULL, MUTEX_DEFAULT, NULL);
4061 mutex_init(&arc_mru->arcs_locks[i].arcs_lock,
4062 NULL, MUTEX_DEFAULT, NULL);
4063 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock,
4064 NULL, MUTEX_DEFAULT, NULL);
4065 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock,
4066 NULL, MUTEX_DEFAULT, NULL);
4067 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock,
4068 NULL, MUTEX_DEFAULT, NULL);
4069 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock,
4070 NULL, MUTEX_DEFAULT, NULL);
4072 list_create(&arc_mru->arcs_lists[i],
4073 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4074 list_create(&arc_mru_ghost->arcs_lists[i],
4075 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4076 list_create(&arc_mfu->arcs_lists[i],
4077 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4078 list_create(&arc_mfu_ghost->arcs_lists[i],
4079 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4080 list_create(&arc_mfu_ghost->arcs_lists[i],
4081 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4082 list_create(&arc_l2c_only->arcs_lists[i],
4083 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
4088 arc_thread_exit = 0;
4089 arc_eviction_list = NULL;
4090 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
4091 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
4093 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
4094 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
4096 if (arc_ksp != NULL) {
4097 arc_ksp->ks_data = &arc_stats;
4098 kstat_install(arc_ksp);
4101 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
4102 TS_RUN, minclsyspri);
4105 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
4106 EVENTHANDLER_PRI_FIRST);
4113 * Calculate maximum amount of dirty data per pool.
4115 * If it has been set by /etc/system, take that.
4116 * Otherwise, use a percentage of physical memory defined by
4117 * zfs_dirty_data_max_percent (default 10%) with a cap at
4118 * zfs_dirty_data_max_max (default 4GB).
4120 if (zfs_dirty_data_max == 0) {
4121 zfs_dirty_data_max = ptob(physmem) *
4122 zfs_dirty_data_max_percent / 100;
4123 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
4124 zfs_dirty_data_max_max);
4128 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
4129 prefetch_tunable_set = 1;
4132 if (prefetch_tunable_set == 0) {
4133 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
4135 printf(" add \"vfs.zfs.prefetch_disable=0\" "
4136 "to /boot/loader.conf.\n");
4137 zfs_prefetch_disable = 1;
4140 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
4141 prefetch_tunable_set == 0) {
4142 printf("ZFS NOTICE: Prefetch is disabled by default if less "
4143 "than 4GB of RAM is present;\n"
4144 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
4145 "to /boot/loader.conf.\n");
4146 zfs_prefetch_disable = 1;
4149 /* Warn about ZFS memory and address space requirements. */
4150 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
4151 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
4152 "expect unstable behavior.\n");
4154 if (kmem_size() < 512 * (1 << 20)) {
4155 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
4156 "expect unstable behavior.\n");
4157 printf(" Consider tuning vm.kmem_size and "
4158 "vm.kmem_size_max\n");
4159 printf(" in /boot/loader.conf.\n");
4169 mutex_enter(&arc_reclaim_thr_lock);
4170 arc_thread_exit = 1;
4171 cv_signal(&arc_reclaim_thr_cv);
4172 while (arc_thread_exit != 0)
4173 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
4174 mutex_exit(&arc_reclaim_thr_lock);
4180 if (arc_ksp != NULL) {
4181 kstat_delete(arc_ksp);
4185 mutex_destroy(&arc_eviction_mtx);
4186 mutex_destroy(&arc_reclaim_thr_lock);
4187 cv_destroy(&arc_reclaim_thr_cv);
4189 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
4190 list_destroy(&arc_mru->arcs_lists[i]);
4191 list_destroy(&arc_mru_ghost->arcs_lists[i]);
4192 list_destroy(&arc_mfu->arcs_lists[i]);
4193 list_destroy(&arc_mfu_ghost->arcs_lists[i]);
4194 list_destroy(&arc_l2c_only->arcs_lists[i]);
4196 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock);
4197 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock);
4198 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock);
4199 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock);
4200 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock);
4201 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock);
4206 ASSERT(arc_loaned_bytes == 0);
4208 mutex_destroy(&arc_lowmem_lock);
4210 if (arc_event_lowmem != NULL)
4211 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
4218 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
4219 * It uses dedicated storage devices to hold cached data, which are populated
4220 * using large infrequent writes. The main role of this cache is to boost
4221 * the performance of random read workloads. The intended L2ARC devices
4222 * include short-stroked disks, solid state disks, and other media with
4223 * substantially faster read latency than disk.
4225 * +-----------------------+
4227 * +-----------------------+
4230 * l2arc_feed_thread() arc_read()
4234 * +---------------+ |
4236 * +---------------+ |
4241 * +-------+ +-------+
4243 * | cache | | cache |
4244 * +-------+ +-------+
4245 * +=========+ .-----.
4246 * : L2ARC : |-_____-|
4247 * : devices : | Disks |
4248 * +=========+ `-_____-'
4250 * Read requests are satisfied from the following sources, in order:
4253 * 2) vdev cache of L2ARC devices
4255 * 4) vdev cache of disks
4258 * Some L2ARC device types exhibit extremely slow write performance.
4259 * To accommodate for this there are some significant differences between
4260 * the L2ARC and traditional cache design:
4262 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4263 * the ARC behave as usual, freeing buffers and placing headers on ghost
4264 * lists. The ARC does not send buffers to the L2ARC during eviction as
4265 * this would add inflated write latencies for all ARC memory pressure.
4267 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4268 * It does this by periodically scanning buffers from the eviction-end of
4269 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4270 * not already there. It scans until a headroom of buffers is satisfied,
4271 * which itself is a buffer for ARC eviction. If a compressible buffer is
4272 * found during scanning and selected for writing to an L2ARC device, we
4273 * temporarily boost scanning headroom during the next scan cycle to make
4274 * sure we adapt to compression effects (which might significantly reduce
4275 * the data volume we write to L2ARC). The thread that does this is
4276 * l2arc_feed_thread(), illustrated below; example sizes are included to
4277 * provide a better sense of ratio than this diagram:
4280 * +---------------------+----------+
4281 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4282 * +---------------------+----------+ | o L2ARC eligible
4283 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4284 * +---------------------+----------+ |
4285 * 15.9 Gbytes ^ 32 Mbytes |
4287 * l2arc_feed_thread()
4289 * l2arc write hand <--[oooo]--'
4293 * +==============================+
4294 * L2ARC dev |####|#|###|###| |####| ... |
4295 * +==============================+
4298 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4299 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4300 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4301 * safe to say that this is an uncommon case, since buffers at the end of
4302 * the ARC lists have moved there due to inactivity.
4304 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4305 * then the L2ARC simply misses copying some buffers. This serves as a
4306 * pressure valve to prevent heavy read workloads from both stalling the ARC
4307 * with waits and clogging the L2ARC with writes. This also helps prevent
4308 * the potential for the L2ARC to churn if it attempts to cache content too
4309 * quickly, such as during backups of the entire pool.
4311 * 5. After system boot and before the ARC has filled main memory, there are
4312 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4313 * lists can remain mostly static. Instead of searching from tail of these
4314 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4315 * for eligible buffers, greatly increasing its chance of finding them.
4317 * The L2ARC device write speed is also boosted during this time so that
4318 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4319 * there are no L2ARC reads, and no fear of degrading read performance
4320 * through increased writes.
4322 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4323 * the vdev queue can aggregate them into larger and fewer writes. Each
4324 * device is written to in a rotor fashion, sweeping writes through
4325 * available space then repeating.
4327 * 7. The L2ARC does not store dirty content. It never needs to flush
4328 * write buffers back to disk based storage.
4330 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4331 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4333 * The performance of the L2ARC can be tweaked by a number of tunables, which
4334 * may be necessary for different workloads:
4336 * l2arc_write_max max write bytes per interval
4337 * l2arc_write_boost extra write bytes during device warmup
4338 * l2arc_noprefetch skip caching prefetched buffers
4339 * l2arc_headroom number of max device writes to precache
4340 * l2arc_headroom_boost when we find compressed buffers during ARC
4341 * scanning, we multiply headroom by this
4342 * percentage factor for the next scan cycle,
4343 * since more compressed buffers are likely to
4345 * l2arc_feed_secs seconds between L2ARC writing
4347 * Tunables may be removed or added as future performance improvements are
4348 * integrated, and also may become zpool properties.
4350 * There are three key functions that control how the L2ARC warms up:
4352 * l2arc_write_eligible() check if a buffer is eligible to cache
4353 * l2arc_write_size() calculate how much to write
4354 * l2arc_write_interval() calculate sleep delay between writes
4356 * These three functions determine what to write, how much, and how quickly
4361 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4364 * A buffer is *not* eligible for the L2ARC if it:
4365 * 1. belongs to a different spa.
4366 * 2. is already cached on the L2ARC.
4367 * 3. has an I/O in progress (it may be an incomplete read).
4368 * 4. is flagged not eligible (zfs property).
4370 if (ab->b_spa != spa_guid) {
4371 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
4374 if (ab->b_l2hdr != NULL) {
4375 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
4378 if (HDR_IO_IN_PROGRESS(ab)) {
4379 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
4382 if (!HDR_L2CACHE(ab)) {
4383 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
4391 l2arc_write_size(void)
4396 * Make sure our globals have meaningful values in case the user
4399 size = l2arc_write_max;
4401 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
4402 "be greater than zero, resetting it to the default (%d)",
4404 size = l2arc_write_max = L2ARC_WRITE_SIZE;
4407 if (arc_warm == B_FALSE)
4408 size += l2arc_write_boost;
4415 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4417 clock_t interval, next, now;
4420 * If the ARC lists are busy, increase our write rate; if the
4421 * lists are stale, idle back. This is achieved by checking
4422 * how much we previously wrote - if it was more than half of
4423 * what we wanted, schedule the next write much sooner.
4425 if (l2arc_feed_again && wrote > (wanted / 2))
4426 interval = (hz * l2arc_feed_min_ms) / 1000;
4428 interval = hz * l2arc_feed_secs;
4430 now = ddi_get_lbolt();
4431 next = MAX(now, MIN(now + interval, began + interval));
4437 l2arc_hdr_stat_add(void)
4439 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4440 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4444 l2arc_hdr_stat_remove(void)
4446 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4447 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4451 * Cycle through L2ARC devices. This is how L2ARC load balances.
4452 * If a device is returned, this also returns holding the spa config lock.
4454 static l2arc_dev_t *
4455 l2arc_dev_get_next(void)
4457 l2arc_dev_t *first, *next = NULL;
4460 * Lock out the removal of spas (spa_namespace_lock), then removal
4461 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4462 * both locks will be dropped and a spa config lock held instead.
4464 mutex_enter(&spa_namespace_lock);
4465 mutex_enter(&l2arc_dev_mtx);
4467 /* if there are no vdevs, there is nothing to do */
4468 if (l2arc_ndev == 0)
4472 next = l2arc_dev_last;
4474 /* loop around the list looking for a non-faulted vdev */
4476 next = list_head(l2arc_dev_list);
4478 next = list_next(l2arc_dev_list, next);
4480 next = list_head(l2arc_dev_list);
4483 /* if we have come back to the start, bail out */
4486 else if (next == first)
4489 } while (vdev_is_dead(next->l2ad_vdev));
4491 /* if we were unable to find any usable vdevs, return NULL */
4492 if (vdev_is_dead(next->l2ad_vdev))
4495 l2arc_dev_last = next;
4498 mutex_exit(&l2arc_dev_mtx);
4501 * Grab the config lock to prevent the 'next' device from being
4502 * removed while we are writing to it.
4505 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4506 mutex_exit(&spa_namespace_lock);
4512 * Free buffers that were tagged for destruction.
4515 l2arc_do_free_on_write()
4518 l2arc_data_free_t *df, *df_prev;
4520 mutex_enter(&l2arc_free_on_write_mtx);
4521 buflist = l2arc_free_on_write;
4523 for (df = list_tail(buflist); df; df = df_prev) {
4524 df_prev = list_prev(buflist, df);
4525 ASSERT(df->l2df_data != NULL);
4526 ASSERT(df->l2df_func != NULL);
4527 df->l2df_func(df->l2df_data, df->l2df_size);
4528 list_remove(buflist, df);
4529 kmem_free(df, sizeof (l2arc_data_free_t));
4532 mutex_exit(&l2arc_free_on_write_mtx);
4536 * A write to a cache device has completed. Update all headers to allow
4537 * reads from these buffers to begin.
4540 l2arc_write_done(zio_t *zio)
4542 l2arc_write_callback_t *cb;
4545 arc_buf_hdr_t *head, *ab, *ab_prev;
4546 l2arc_buf_hdr_t *abl2;
4547 kmutex_t *hash_lock;
4549 cb = zio->io_private;
4551 dev = cb->l2wcb_dev;
4552 ASSERT(dev != NULL);
4553 head = cb->l2wcb_head;
4554 ASSERT(head != NULL);
4555 buflist = dev->l2ad_buflist;
4556 ASSERT(buflist != NULL);
4557 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4558 l2arc_write_callback_t *, cb);
4560 if (zio->io_error != 0)
4561 ARCSTAT_BUMP(arcstat_l2_writes_error);
4563 mutex_enter(&l2arc_buflist_mtx);
4566 * All writes completed, or an error was hit.
4568 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4569 ab_prev = list_prev(buflist, ab);
4573 * Release the temporary compressed buffer as soon as possible.
4575 if (abl2->b_compress != ZIO_COMPRESS_OFF)
4576 l2arc_release_cdata_buf(ab);
4578 hash_lock = HDR_LOCK(ab);
4579 if (!mutex_tryenter(hash_lock)) {
4581 * This buffer misses out. It may be in a stage
4582 * of eviction. Its ARC_L2_WRITING flag will be
4583 * left set, denying reads to this buffer.
4585 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4589 if (zio->io_error != 0) {
4591 * Error - drop L2ARC entry.
4593 list_remove(buflist, ab);
4594 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4596 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr,
4598 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4599 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4603 * Allow ARC to begin reads to this L2ARC entry.
4605 ab->b_flags &= ~ARC_L2_WRITING;
4607 mutex_exit(hash_lock);
4610 atomic_inc_64(&l2arc_writes_done);
4611 list_remove(buflist, head);
4612 kmem_cache_free(hdr_cache, head);
4613 mutex_exit(&l2arc_buflist_mtx);
4615 l2arc_do_free_on_write();
4617 kmem_free(cb, sizeof (l2arc_write_callback_t));
4621 * A read to a cache device completed. Validate buffer contents before
4622 * handing over to the regular ARC routines.
4625 l2arc_read_done(zio_t *zio)
4627 l2arc_read_callback_t *cb;
4630 kmutex_t *hash_lock;
4633 ASSERT(zio->io_vd != NULL);
4634 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4636 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4638 cb = zio->io_private;
4640 buf = cb->l2rcb_buf;
4641 ASSERT(buf != NULL);
4643 hash_lock = HDR_LOCK(buf->b_hdr);
4644 mutex_enter(hash_lock);
4646 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4649 * If the buffer was compressed, decompress it first.
4651 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
4652 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
4653 ASSERT(zio->io_data != NULL);
4656 * Check this survived the L2ARC journey.
4658 equal = arc_cksum_equal(buf);
4659 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4660 mutex_exit(hash_lock);
4661 zio->io_private = buf;
4662 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4663 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4666 mutex_exit(hash_lock);
4668 * Buffer didn't survive caching. Increment stats and
4669 * reissue to the original storage device.
4671 if (zio->io_error != 0) {
4672 ARCSTAT_BUMP(arcstat_l2_io_error);
4674 zio->io_error = SET_ERROR(EIO);
4677 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4680 * If there's no waiter, issue an async i/o to the primary
4681 * storage now. If there *is* a waiter, the caller must
4682 * issue the i/o in a context where it's OK to block.
4684 if (zio->io_waiter == NULL) {
4685 zio_t *pio = zio_unique_parent(zio);
4687 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4689 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4690 buf->b_data, zio->io_size, arc_read_done, buf,
4691 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4695 kmem_free(cb, sizeof (l2arc_read_callback_t));
4699 * This is the list priority from which the L2ARC will search for pages to
4700 * cache. This is used within loops (0..3) to cycle through lists in the
4701 * desired order. This order can have a significant effect on cache
4704 * Currently the metadata lists are hit first, MFU then MRU, followed by
4705 * the data lists. This function returns a locked list, and also returns
4709 l2arc_list_locked(int list_num, kmutex_t **lock)
4711 list_t *list = NULL;
4714 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS);
4716 if (list_num < ARC_BUFC_NUMMETADATALISTS) {
4718 list = &arc_mfu->arcs_lists[idx];
4719 *lock = ARCS_LOCK(arc_mfu, idx);
4720 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) {
4721 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4722 list = &arc_mru->arcs_lists[idx];
4723 *lock = ARCS_LOCK(arc_mru, idx);
4724 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 +
4725 ARC_BUFC_NUMDATALISTS)) {
4726 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4727 list = &arc_mfu->arcs_lists[idx];
4728 *lock = ARCS_LOCK(arc_mfu, idx);
4730 idx = list_num - ARC_BUFC_NUMLISTS;
4731 list = &arc_mru->arcs_lists[idx];
4732 *lock = ARCS_LOCK(arc_mru, idx);
4735 ASSERT(!(MUTEX_HELD(*lock)));
4741 * Evict buffers from the device write hand to the distance specified in
4742 * bytes. This distance may span populated buffers, it may span nothing.
4743 * This is clearing a region on the L2ARC device ready for writing.
4744 * If the 'all' boolean is set, every buffer is evicted.
4747 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4750 l2arc_buf_hdr_t *abl2;
4751 arc_buf_hdr_t *ab, *ab_prev;
4752 kmutex_t *hash_lock;
4755 buflist = dev->l2ad_buflist;
4757 if (buflist == NULL)
4760 if (!all && dev->l2ad_first) {
4762 * This is the first sweep through the device. There is
4768 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4770 * When nearing the end of the device, evict to the end
4771 * before the device write hand jumps to the start.
4773 taddr = dev->l2ad_end;
4775 taddr = dev->l2ad_hand + distance;
4777 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4778 uint64_t, taddr, boolean_t, all);
4781 mutex_enter(&l2arc_buflist_mtx);
4782 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4783 ab_prev = list_prev(buflist, ab);
4785 hash_lock = HDR_LOCK(ab);
4786 if (!mutex_tryenter(hash_lock)) {
4788 * Missed the hash lock. Retry.
4790 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4791 mutex_exit(&l2arc_buflist_mtx);
4792 mutex_enter(hash_lock);
4793 mutex_exit(hash_lock);
4797 if (HDR_L2_WRITE_HEAD(ab)) {
4799 * We hit a write head node. Leave it for
4800 * l2arc_write_done().
4802 list_remove(buflist, ab);
4803 mutex_exit(hash_lock);
4807 if (!all && ab->b_l2hdr != NULL &&
4808 (ab->b_l2hdr->b_daddr > taddr ||
4809 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4811 * We've evicted to the target address,
4812 * or the end of the device.
4814 mutex_exit(hash_lock);
4818 if (HDR_FREE_IN_PROGRESS(ab)) {
4820 * Already on the path to destruction.
4822 mutex_exit(hash_lock);
4826 if (ab->b_state == arc_l2c_only) {
4827 ASSERT(!HDR_L2_READING(ab));
4829 * This doesn't exist in the ARC. Destroy.
4830 * arc_hdr_destroy() will call list_remove()
4831 * and decrement arcstat_l2_size.
4833 arc_change_state(arc_anon, ab, hash_lock);
4834 arc_hdr_destroy(ab);
4837 * Invalidate issued or about to be issued
4838 * reads, since we may be about to write
4839 * over this location.
4841 if (HDR_L2_READING(ab)) {
4842 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4843 ab->b_flags |= ARC_L2_EVICTED;
4847 * Tell ARC this no longer exists in L2ARC.
4849 if (ab->b_l2hdr != NULL) {
4851 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4853 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4854 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4856 list_remove(buflist, ab);
4859 * This may have been leftover after a
4862 ab->b_flags &= ~ARC_L2_WRITING;
4864 mutex_exit(hash_lock);
4866 mutex_exit(&l2arc_buflist_mtx);
4868 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4869 dev->l2ad_evict = taddr;
4873 * Find and write ARC buffers to the L2ARC device.
4875 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4876 * for reading until they have completed writing.
4877 * The headroom_boost is an in-out parameter used to maintain headroom boost
4878 * state between calls to this function.
4880 * Returns the number of bytes actually written (which may be smaller than
4881 * the delta by which the device hand has changed due to alignment).
4884 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
4885 boolean_t *headroom_boost)
4887 arc_buf_hdr_t *ab, *ab_prev, *head;
4889 uint64_t write_asize, write_psize, write_sz, headroom,
4892 kmutex_t *list_lock;
4894 l2arc_write_callback_t *cb;
4896 uint64_t guid = spa_load_guid(spa);
4897 const boolean_t do_headroom_boost = *headroom_boost;
4900 ASSERT(dev->l2ad_vdev != NULL);
4902 /* Lower the flag now, we might want to raise it again later. */
4903 *headroom_boost = B_FALSE;
4906 write_sz = write_asize = write_psize = 0;
4908 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4909 head->b_flags |= ARC_L2_WRITE_HEAD;
4911 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
4913 * We will want to try to compress buffers that are at least 2x the
4914 * device sector size.
4916 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
4919 * Copy buffers for L2ARC writing.
4921 mutex_enter(&l2arc_buflist_mtx);
4922 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) {
4923 uint64_t passed_sz = 0;
4925 list = l2arc_list_locked(try, &list_lock);
4926 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
4929 * L2ARC fast warmup.
4931 * Until the ARC is warm and starts to evict, read from the
4932 * head of the ARC lists rather than the tail.
4934 if (arc_warm == B_FALSE)
4935 ab = list_head(list);
4937 ab = list_tail(list);
4939 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
4941 headroom = target_sz * l2arc_headroom;
4942 if (do_headroom_boost)
4943 headroom = (headroom * l2arc_headroom_boost) / 100;
4945 for (; ab; ab = ab_prev) {
4946 l2arc_buf_hdr_t *l2hdr;
4947 kmutex_t *hash_lock;
4950 if (arc_warm == B_FALSE)
4951 ab_prev = list_next(list, ab);
4953 ab_prev = list_prev(list, ab);
4954 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size);
4956 hash_lock = HDR_LOCK(ab);
4957 if (!mutex_tryenter(hash_lock)) {
4958 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
4960 * Skip this buffer rather than waiting.
4965 passed_sz += ab->b_size;
4966 if (passed_sz > headroom) {
4970 mutex_exit(hash_lock);
4971 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
4975 if (!l2arc_write_eligible(guid, ab)) {
4976 mutex_exit(hash_lock);
4980 if ((write_sz + ab->b_size) > target_sz) {
4982 mutex_exit(hash_lock);
4983 ARCSTAT_BUMP(arcstat_l2_write_full);
4989 * Insert a dummy header on the buflist so
4990 * l2arc_write_done() can find where the
4991 * write buffers begin without searching.
4993 list_insert_head(dev->l2ad_buflist, head);
4996 sizeof (l2arc_write_callback_t), KM_SLEEP);
4997 cb->l2wcb_dev = dev;
4998 cb->l2wcb_head = head;
4999 pio = zio_root(spa, l2arc_write_done, cb,
5001 ARCSTAT_BUMP(arcstat_l2_write_pios);
5005 * Create and add a new L2ARC header.
5007 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
5009 ab->b_flags |= ARC_L2_WRITING;
5012 * Temporarily stash the data buffer in b_tmp_cdata.
5013 * The subsequent write step will pick it up from
5014 * there. This is because can't access ab->b_buf
5015 * without holding the hash_lock, which we in turn
5016 * can't access without holding the ARC list locks
5017 * (which we want to avoid during compression/writing).
5019 l2hdr->b_compress = ZIO_COMPRESS_OFF;
5020 l2hdr->b_asize = ab->b_size;
5021 l2hdr->b_tmp_cdata = ab->b_buf->b_data;
5023 buf_sz = ab->b_size;
5024 ab->b_l2hdr = l2hdr;
5026 list_insert_head(dev->l2ad_buflist, ab);
5029 * Compute and store the buffer cksum before
5030 * writing. On debug the cksum is verified first.
5032 arc_cksum_verify(ab->b_buf);
5033 arc_cksum_compute(ab->b_buf, B_TRUE);
5035 mutex_exit(hash_lock);
5040 mutex_exit(list_lock);
5046 /* No buffers selected for writing? */
5049 mutex_exit(&l2arc_buflist_mtx);
5050 kmem_cache_free(hdr_cache, head);
5055 * Now start writing the buffers. We're starting at the write head
5056 * and work backwards, retracing the course of the buffer selector
5059 for (ab = list_prev(dev->l2ad_buflist, head); ab;
5060 ab = list_prev(dev->l2ad_buflist, ab)) {
5061 l2arc_buf_hdr_t *l2hdr;
5065 * We shouldn't need to lock the buffer here, since we flagged
5066 * it as ARC_L2_WRITING in the previous step, but we must take
5067 * care to only access its L2 cache parameters. In particular,
5068 * ab->b_buf may be invalid by now due to ARC eviction.
5070 l2hdr = ab->b_l2hdr;
5071 l2hdr->b_daddr = dev->l2ad_hand;
5073 if ((ab->b_flags & ARC_L2COMPRESS) &&
5074 l2hdr->b_asize >= buf_compress_minsz) {
5075 if (l2arc_compress_buf(l2hdr)) {
5077 * If compression succeeded, enable headroom
5078 * boost on the next scan cycle.
5080 *headroom_boost = B_TRUE;
5085 * Pick up the buffer data we had previously stashed away
5086 * (and now potentially also compressed).
5088 buf_data = l2hdr->b_tmp_cdata;
5089 buf_sz = l2hdr->b_asize;
5091 /* Compression may have squashed the buffer to zero length. */
5095 wzio = zio_write_phys(pio, dev->l2ad_vdev,
5096 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
5097 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
5098 ZIO_FLAG_CANFAIL, B_FALSE);
5100 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
5102 (void) zio_nowait(wzio);
5104 write_asize += buf_sz;
5106 * Keep the clock hand suitably device-aligned.
5108 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
5109 write_psize += buf_p_sz;
5110 dev->l2ad_hand += buf_p_sz;
5114 mutex_exit(&l2arc_buflist_mtx);
5116 ASSERT3U(write_asize, <=, target_sz);
5117 ARCSTAT_BUMP(arcstat_l2_writes_sent);
5118 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
5119 ARCSTAT_INCR(arcstat_l2_size, write_sz);
5120 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
5121 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
5124 * Bump device hand to the device start if it is approaching the end.
5125 * l2arc_evict() will already have evicted ahead for this case.
5127 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
5128 vdev_space_update(dev->l2ad_vdev,
5129 dev->l2ad_end - dev->l2ad_hand, 0, 0);
5130 dev->l2ad_hand = dev->l2ad_start;
5131 dev->l2ad_evict = dev->l2ad_start;
5132 dev->l2ad_first = B_FALSE;
5135 dev->l2ad_writing = B_TRUE;
5136 (void) zio_wait(pio);
5137 dev->l2ad_writing = B_FALSE;
5139 return (write_asize);
5143 * Compresses an L2ARC buffer.
5144 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its
5145 * size in l2hdr->b_asize. This routine tries to compress the data and
5146 * depending on the compression result there are three possible outcomes:
5147 * *) The buffer was incompressible. The original l2hdr contents were left
5148 * untouched and are ready for writing to an L2 device.
5149 * *) The buffer was all-zeros, so there is no need to write it to an L2
5150 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
5151 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
5152 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
5153 * data buffer which holds the compressed data to be written, and b_asize
5154 * tells us how much data there is. b_compress is set to the appropriate
5155 * compression algorithm. Once writing is done, invoke
5156 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
5158 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
5159 * buffer was incompressible).
5162 l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr)
5167 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
5168 ASSERT(l2hdr->b_tmp_cdata != NULL);
5170 len = l2hdr->b_asize;
5171 cdata = zio_data_buf_alloc(len);
5172 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata,
5173 cdata, l2hdr->b_asize, (size_t)(1ULL << l2hdr->b_dev->l2ad_vdev->vdev_ashift));
5176 /* zero block, indicate that there's nothing to write */
5177 zio_data_buf_free(cdata, len);
5178 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
5180 l2hdr->b_tmp_cdata = NULL;
5181 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
5183 } else if (csize > 0 && csize < len) {
5185 * Compression succeeded, we'll keep the cdata around for
5186 * writing and release it afterwards.
5188 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
5189 l2hdr->b_asize = csize;
5190 l2hdr->b_tmp_cdata = cdata;
5191 ARCSTAT_BUMP(arcstat_l2_compress_successes);
5195 * Compression failed, release the compressed buffer.
5196 * l2hdr will be left unmodified.
5198 zio_data_buf_free(cdata, len);
5199 ARCSTAT_BUMP(arcstat_l2_compress_failures);
5205 * Decompresses a zio read back from an l2arc device. On success, the
5206 * underlying zio's io_data buffer is overwritten by the uncompressed
5207 * version. On decompression error (corrupt compressed stream), the
5208 * zio->io_error value is set to signal an I/O error.
5210 * Please note that the compressed data stream is not checksummed, so
5211 * if the underlying device is experiencing data corruption, we may feed
5212 * corrupt data to the decompressor, so the decompressor needs to be
5213 * able to handle this situation (LZ4 does).
5216 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
5218 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
5220 if (zio->io_error != 0) {
5222 * An io error has occured, just restore the original io
5223 * size in preparation for a main pool read.
5225 zio->io_orig_size = zio->io_size = hdr->b_size;
5229 if (c == ZIO_COMPRESS_EMPTY) {
5231 * An empty buffer results in a null zio, which means we
5232 * need to fill its io_data after we're done restoring the
5233 * buffer's contents.
5235 ASSERT(hdr->b_buf != NULL);
5236 bzero(hdr->b_buf->b_data, hdr->b_size);
5237 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data;
5239 ASSERT(zio->io_data != NULL);
5241 * We copy the compressed data from the start of the arc buffer
5242 * (the zio_read will have pulled in only what we need, the
5243 * rest is garbage which we will overwrite at decompression)
5244 * and then decompress back to the ARC data buffer. This way we
5245 * can minimize copying by simply decompressing back over the
5246 * original compressed data (rather than decompressing to an
5247 * aux buffer and then copying back the uncompressed buffer,
5248 * which is likely to be much larger).
5253 csize = zio->io_size;
5254 cdata = zio_data_buf_alloc(csize);
5255 bcopy(zio->io_data, cdata, csize);
5256 if (zio_decompress_data(c, cdata, zio->io_data, csize,
5258 zio->io_error = EIO;
5259 zio_data_buf_free(cdata, csize);
5262 /* Restore the expected uncompressed IO size. */
5263 zio->io_orig_size = zio->io_size = hdr->b_size;
5267 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
5268 * This buffer serves as a temporary holder of compressed data while
5269 * the buffer entry is being written to an l2arc device. Once that is
5270 * done, we can dispose of it.
5273 l2arc_release_cdata_buf(arc_buf_hdr_t *ab)
5275 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr;
5277 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) {
5279 * If the data was compressed, then we've allocated a
5280 * temporary buffer for it, so now we need to release it.
5282 ASSERT(l2hdr->b_tmp_cdata != NULL);
5283 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size);
5285 l2hdr->b_tmp_cdata = NULL;
5289 * This thread feeds the L2ARC at regular intervals. This is the beating
5290 * heart of the L2ARC.
5293 l2arc_feed_thread(void *dummy __unused)
5298 uint64_t size, wrote;
5299 clock_t begin, next = ddi_get_lbolt();
5300 boolean_t headroom_boost = B_FALSE;
5302 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
5304 mutex_enter(&l2arc_feed_thr_lock);
5306 while (l2arc_thread_exit == 0) {
5307 CALLB_CPR_SAFE_BEGIN(&cpr);
5308 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
5309 next - ddi_get_lbolt());
5310 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
5311 next = ddi_get_lbolt() + hz;
5314 * Quick check for L2ARC devices.
5316 mutex_enter(&l2arc_dev_mtx);
5317 if (l2arc_ndev == 0) {
5318 mutex_exit(&l2arc_dev_mtx);
5321 mutex_exit(&l2arc_dev_mtx);
5322 begin = ddi_get_lbolt();
5325 * This selects the next l2arc device to write to, and in
5326 * doing so the next spa to feed from: dev->l2ad_spa. This
5327 * will return NULL if there are now no l2arc devices or if
5328 * they are all faulted.
5330 * If a device is returned, its spa's config lock is also
5331 * held to prevent device removal. l2arc_dev_get_next()
5332 * will grab and release l2arc_dev_mtx.
5334 if ((dev = l2arc_dev_get_next()) == NULL)
5337 spa = dev->l2ad_spa;
5338 ASSERT(spa != NULL);
5341 * If the pool is read-only then force the feed thread to
5342 * sleep a little longer.
5344 if (!spa_writeable(spa)) {
5345 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
5346 spa_config_exit(spa, SCL_L2ARC, dev);
5351 * Avoid contributing to memory pressure.
5353 if (arc_reclaim_needed()) {
5354 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
5355 spa_config_exit(spa, SCL_L2ARC, dev);
5359 ARCSTAT_BUMP(arcstat_l2_feeds);
5361 size = l2arc_write_size();
5364 * Evict L2ARC buffers that will be overwritten.
5366 l2arc_evict(dev, size, B_FALSE);
5369 * Write ARC buffers.
5371 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
5374 * Calculate interval between writes.
5376 next = l2arc_write_interval(begin, size, wrote);
5377 spa_config_exit(spa, SCL_L2ARC, dev);
5380 l2arc_thread_exit = 0;
5381 cv_broadcast(&l2arc_feed_thr_cv);
5382 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
5387 l2arc_vdev_present(vdev_t *vd)
5391 mutex_enter(&l2arc_dev_mtx);
5392 for (dev = list_head(l2arc_dev_list); dev != NULL;
5393 dev = list_next(l2arc_dev_list, dev)) {
5394 if (dev->l2ad_vdev == vd)
5397 mutex_exit(&l2arc_dev_mtx);
5399 return (dev != NULL);
5403 * Add a vdev for use by the L2ARC. By this point the spa has already
5404 * validated the vdev and opened it.
5407 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5409 l2arc_dev_t *adddev;
5411 ASSERT(!l2arc_vdev_present(vd));
5413 vdev_ashift_optimize(vd);
5416 * Create a new l2arc device entry.
5418 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5419 adddev->l2ad_spa = spa;
5420 adddev->l2ad_vdev = vd;
5421 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5422 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5423 adddev->l2ad_hand = adddev->l2ad_start;
5424 adddev->l2ad_evict = adddev->l2ad_start;
5425 adddev->l2ad_first = B_TRUE;
5426 adddev->l2ad_writing = B_FALSE;
5429 * This is a list of all ARC buffers that are still valid on the
5432 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5433 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5434 offsetof(arc_buf_hdr_t, b_l2node));
5436 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5439 * Add device to global list
5441 mutex_enter(&l2arc_dev_mtx);
5442 list_insert_head(l2arc_dev_list, adddev);
5443 atomic_inc_64(&l2arc_ndev);
5444 mutex_exit(&l2arc_dev_mtx);
5448 * Remove a vdev from the L2ARC.
5451 l2arc_remove_vdev(vdev_t *vd)
5453 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5456 * Find the device by vdev
5458 mutex_enter(&l2arc_dev_mtx);
5459 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5460 nextdev = list_next(l2arc_dev_list, dev);
5461 if (vd == dev->l2ad_vdev) {
5466 ASSERT(remdev != NULL);
5469 * Remove device from global list
5471 list_remove(l2arc_dev_list, remdev);
5472 l2arc_dev_last = NULL; /* may have been invalidated */
5473 atomic_dec_64(&l2arc_ndev);
5474 mutex_exit(&l2arc_dev_mtx);
5477 * Clear all buflists and ARC references. L2ARC device flush.
5479 l2arc_evict(remdev, 0, B_TRUE);
5480 list_destroy(remdev->l2ad_buflist);
5481 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5482 kmem_free(remdev, sizeof (l2arc_dev_t));
5488 l2arc_thread_exit = 0;
5490 l2arc_writes_sent = 0;
5491 l2arc_writes_done = 0;
5493 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5494 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5495 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5496 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5497 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5499 l2arc_dev_list = &L2ARC_dev_list;
5500 l2arc_free_on_write = &L2ARC_free_on_write;
5501 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5502 offsetof(l2arc_dev_t, l2ad_node));
5503 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5504 offsetof(l2arc_data_free_t, l2df_list_node));
5511 * This is called from dmu_fini(), which is called from spa_fini();
5512 * Because of this, we can assume that all l2arc devices have
5513 * already been removed when the pools themselves were removed.
5516 l2arc_do_free_on_write();
5518 mutex_destroy(&l2arc_feed_thr_lock);
5519 cv_destroy(&l2arc_feed_thr_cv);
5520 mutex_destroy(&l2arc_dev_mtx);
5521 mutex_destroy(&l2arc_buflist_mtx);
5522 mutex_destroy(&l2arc_free_on_write_mtx);
5524 list_destroy(l2arc_dev_list);
5525 list_destroy(l2arc_free_on_write);
5531 if (!(spa_mode_global & FWRITE))
5534 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5535 TS_RUN, minclsyspri);
5541 if (!(spa_mode_global & FWRITE))
5544 mutex_enter(&l2arc_feed_thr_lock);
5545 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5546 l2arc_thread_exit = 1;
5547 while (l2arc_thread_exit != 0)
5548 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5549 mutex_exit(&l2arc_feed_thr_lock);