4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
30 * DVA-based Adjustable Replacement Cache
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal arc algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
96 * Each arc state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an arc list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
102 * Arc buffers may have an associated eviction callback function.
103 * This function will be invoked prior to removing the buffer (e.g.
104 * in arc_do_user_evicts()). Note however that the data associated
105 * with the buffer may be evicted prior to the callback. The callback
106 * must be made with *no locks held* (to prevent deadlock). Additionally,
107 * the users of callbacks must ensure that their private data is
108 * protected from simultaneous callbacks from arc_clear_callback()
109 * and arc_do_user_evicts().
111 * Note that the majority of the performance stats are manipulated
112 * with atomic operations.
114 * The L2ARC uses the l2ad_mtx on each vdev for the following:
116 * - L2ARC buflist creation
117 * - L2ARC buflist eviction
118 * - L2ARC write completion, which walks L2ARC buflists
119 * - ARC header destruction, as it removes from L2ARC buflists
120 * - ARC header release, as it removes from L2ARC buflists
125 #include <sys/zio_compress.h>
126 #include <sys/zfs_context.h>
128 #include <sys/refcount.h>
129 #include <sys/vdev.h>
130 #include <sys/vdev_impl.h>
131 #include <sys/dsl_pool.h>
132 #include <sys/multilist.h>
134 #include <sys/vmsystm.h>
136 #include <sys/fs/swapnode.h>
137 #include <sys/dnlc.h>
139 #include <sys/callb.h>
140 #include <sys/kstat.h>
141 #include <zfs_fletcher.h>
144 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
145 boolean_t arc_watch = B_FALSE;
149 static kmutex_t arc_reclaim_lock;
150 static kcondvar_t arc_reclaim_thread_cv;
151 static boolean_t arc_reclaim_thread_exit;
152 static kcondvar_t arc_reclaim_waiters_cv;
154 static kmutex_t arc_user_evicts_lock;
155 static kcondvar_t arc_user_evicts_cv;
156 static boolean_t arc_user_evicts_thread_exit;
158 uint_t arc_reduce_dnlc_percent = 3;
161 * The number of headers to evict in arc_evict_state_impl() before
162 * dropping the sublist lock and evicting from another sublist. A lower
163 * value means we're more likely to evict the "correct" header (i.e. the
164 * oldest header in the arc state), but comes with higher overhead
165 * (i.e. more invocations of arc_evict_state_impl()).
167 int zfs_arc_evict_batch_limit = 10;
170 * The number of sublists used for each of the arc state lists. If this
171 * is not set to a suitable value by the user, it will be configured to
172 * the number of CPUs on the system in arc_init().
174 int zfs_arc_num_sublists_per_state = 0;
176 /* number of seconds before growing cache again */
177 static int arc_grow_retry = 60;
179 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
180 int zfs_arc_overflow_shift = 8;
182 /* shift of arc_c for calculating both min and max arc_p */
183 static int arc_p_min_shift = 4;
185 /* log2(fraction of arc to reclaim) */
186 static int arc_shrink_shift = 7;
189 * log2(fraction of ARC which must be free to allow growing).
190 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
191 * when reading a new block into the ARC, we will evict an equal-sized block
194 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
195 * we will still not allow it to grow.
197 int arc_no_grow_shift = 5;
201 * minimum lifespan of a prefetch block in clock ticks
202 * (initialized in arc_init())
204 static int arc_min_prefetch_lifespan;
207 * If this percent of memory is free, don't throttle.
209 int arc_lotsfree_percent = 10;
214 * The arc has filled available memory and has now warmed up.
216 static boolean_t arc_warm;
219 * These tunables are for performance analysis.
221 uint64_t zfs_arc_max;
222 uint64_t zfs_arc_min;
223 uint64_t zfs_arc_meta_limit = 0;
224 uint64_t zfs_arc_meta_min = 0;
225 int zfs_arc_grow_retry = 0;
226 int zfs_arc_shrink_shift = 0;
227 int zfs_arc_p_min_shift = 0;
228 int zfs_disable_dup_eviction = 0;
229 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
232 * Note that buffers can be in one of 6 states:
233 * ARC_anon - anonymous (discussed below)
234 * ARC_mru - recently used, currently cached
235 * ARC_mru_ghost - recentely used, no longer in cache
236 * ARC_mfu - frequently used, currently cached
237 * ARC_mfu_ghost - frequently used, no longer in cache
238 * ARC_l2c_only - exists in L2ARC but not other states
239 * When there are no active references to the buffer, they are
240 * are linked onto a list in one of these arc states. These are
241 * the only buffers that can be evicted or deleted. Within each
242 * state there are multiple lists, one for meta-data and one for
243 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
244 * etc.) is tracked separately so that it can be managed more
245 * explicitly: favored over data, limited explicitly.
247 * Anonymous buffers are buffers that are not associated with
248 * a DVA. These are buffers that hold dirty block copies
249 * before they are written to stable storage. By definition,
250 * they are "ref'd" and are considered part of arc_mru
251 * that cannot be freed. Generally, they will aquire a DVA
252 * as they are written and migrate onto the arc_mru list.
254 * The ARC_l2c_only state is for buffers that are in the second
255 * level ARC but no longer in any of the ARC_m* lists. The second
256 * level ARC itself may also contain buffers that are in any of
257 * the ARC_m* states - meaning that a buffer can exist in two
258 * places. The reason for the ARC_l2c_only state is to keep the
259 * buffer header in the hash table, so that reads that hit the
260 * second level ARC benefit from these fast lookups.
263 typedef struct arc_state {
265 * list of evictable buffers
267 multilist_t arcs_list[ARC_BUFC_NUMTYPES];
269 * total amount of evictable data in this state
271 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES];
273 * total amount of data in this state; this includes: evictable,
274 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
276 refcount_t arcs_size;
280 static arc_state_t ARC_anon;
281 static arc_state_t ARC_mru;
282 static arc_state_t ARC_mru_ghost;
283 static arc_state_t ARC_mfu;
284 static arc_state_t ARC_mfu_ghost;
285 static arc_state_t ARC_l2c_only;
287 typedef struct arc_stats {
288 kstat_named_t arcstat_hits;
289 kstat_named_t arcstat_misses;
290 kstat_named_t arcstat_demand_data_hits;
291 kstat_named_t arcstat_demand_data_misses;
292 kstat_named_t arcstat_demand_metadata_hits;
293 kstat_named_t arcstat_demand_metadata_misses;
294 kstat_named_t arcstat_prefetch_data_hits;
295 kstat_named_t arcstat_prefetch_data_misses;
296 kstat_named_t arcstat_prefetch_metadata_hits;
297 kstat_named_t arcstat_prefetch_metadata_misses;
298 kstat_named_t arcstat_mru_hits;
299 kstat_named_t arcstat_mru_ghost_hits;
300 kstat_named_t arcstat_mfu_hits;
301 kstat_named_t arcstat_mfu_ghost_hits;
302 kstat_named_t arcstat_deleted;
304 * Number of buffers that could not be evicted because the hash lock
305 * was held by another thread. The lock may not necessarily be held
306 * by something using the same buffer, since hash locks are shared
307 * by multiple buffers.
309 kstat_named_t arcstat_mutex_miss;
311 * Number of buffers skipped because they have I/O in progress, are
312 * indrect prefetch buffers that have not lived long enough, or are
313 * not from the spa we're trying to evict from.
315 kstat_named_t arcstat_evict_skip;
317 * Number of times arc_evict_state() was unable to evict enough
318 * buffers to reach it's target amount.
320 kstat_named_t arcstat_evict_not_enough;
321 kstat_named_t arcstat_evict_l2_cached;
322 kstat_named_t arcstat_evict_l2_eligible;
323 kstat_named_t arcstat_evict_l2_ineligible;
324 kstat_named_t arcstat_evict_l2_skip;
325 kstat_named_t arcstat_hash_elements;
326 kstat_named_t arcstat_hash_elements_max;
327 kstat_named_t arcstat_hash_collisions;
328 kstat_named_t arcstat_hash_chains;
329 kstat_named_t arcstat_hash_chain_max;
330 kstat_named_t arcstat_p;
331 kstat_named_t arcstat_c;
332 kstat_named_t arcstat_c_min;
333 kstat_named_t arcstat_c_max;
334 kstat_named_t arcstat_size;
336 * Number of bytes consumed by internal ARC structures necessary
337 * for tracking purposes; these structures are not actually
338 * backed by ARC buffers. This includes arc_buf_hdr_t structures
339 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
340 * caches), and arc_buf_t structures (allocated via arc_buf_t
343 kstat_named_t arcstat_hdr_size;
345 * Number of bytes consumed by ARC buffers of type equal to
346 * ARC_BUFC_DATA. This is generally consumed by buffers backing
347 * on disk user data (e.g. plain file contents).
349 kstat_named_t arcstat_data_size;
351 * Number of bytes consumed by ARC buffers of type equal to
352 * ARC_BUFC_METADATA. This is generally consumed by buffers
353 * backing on disk data that is used for internal ZFS
354 * structures (e.g. ZAP, dnode, indirect blocks, etc).
356 kstat_named_t arcstat_metadata_size;
358 * Number of bytes consumed by various buffers and structures
359 * not actually backed with ARC buffers. This includes bonus
360 * buffers (allocated directly via zio_buf_* functions),
361 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
362 * cache), and dnode_t structures (allocated via dnode_t cache).
364 kstat_named_t arcstat_other_size;
366 * Total number of bytes consumed by ARC buffers residing in the
367 * arc_anon state. This includes *all* buffers in the arc_anon
368 * state; e.g. data, metadata, evictable, and unevictable buffers
369 * are all included in this value.
371 kstat_named_t arcstat_anon_size;
373 * Number of bytes consumed by ARC buffers that meet the
374 * following criteria: backing buffers of type ARC_BUFC_DATA,
375 * residing in the arc_anon state, and are eligible for eviction
376 * (e.g. have no outstanding holds on the buffer).
378 kstat_named_t arcstat_anon_evictable_data;
380 * Number of bytes consumed by ARC buffers that meet the
381 * following criteria: backing buffers of type ARC_BUFC_METADATA,
382 * residing in the arc_anon state, and are eligible for eviction
383 * (e.g. have no outstanding holds on the buffer).
385 kstat_named_t arcstat_anon_evictable_metadata;
387 * Total number of bytes consumed by ARC buffers residing in the
388 * arc_mru state. This includes *all* buffers in the arc_mru
389 * state; e.g. data, metadata, evictable, and unevictable buffers
390 * are all included in this value.
392 kstat_named_t arcstat_mru_size;
394 * Number of bytes consumed by ARC buffers that meet the
395 * following criteria: backing buffers of type ARC_BUFC_DATA,
396 * residing in the arc_mru state, and are eligible for eviction
397 * (e.g. have no outstanding holds on the buffer).
399 kstat_named_t arcstat_mru_evictable_data;
401 * Number of bytes consumed by ARC buffers that meet the
402 * following criteria: backing buffers of type ARC_BUFC_METADATA,
403 * residing in the arc_mru state, and are eligible for eviction
404 * (e.g. have no outstanding holds on the buffer).
406 kstat_named_t arcstat_mru_evictable_metadata;
408 * Total number of bytes that *would have been* consumed by ARC
409 * buffers in the arc_mru_ghost state. The key thing to note
410 * here, is the fact that this size doesn't actually indicate
411 * RAM consumption. The ghost lists only consist of headers and
412 * don't actually have ARC buffers linked off of these headers.
413 * Thus, *if* the headers had associated ARC buffers, these
414 * buffers *would have* consumed this number of bytes.
416 kstat_named_t arcstat_mru_ghost_size;
418 * Number of bytes that *would have been* consumed by ARC
419 * buffers that are eligible for eviction, of type
420 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
422 kstat_named_t arcstat_mru_ghost_evictable_data;
424 * Number of bytes that *would have been* consumed by ARC
425 * buffers that are eligible for eviction, of type
426 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
428 kstat_named_t arcstat_mru_ghost_evictable_metadata;
430 * Total number of bytes consumed by ARC buffers residing in the
431 * arc_mfu state. This includes *all* buffers in the arc_mfu
432 * state; e.g. data, metadata, evictable, and unevictable buffers
433 * are all included in this value.
435 kstat_named_t arcstat_mfu_size;
437 * Number of bytes consumed by ARC buffers that are eligible for
438 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
441 kstat_named_t arcstat_mfu_evictable_data;
443 * Number of bytes consumed by ARC buffers that are eligible for
444 * eviction, of type ARC_BUFC_METADATA, and reside in the
447 kstat_named_t arcstat_mfu_evictable_metadata;
449 * Total number of bytes that *would have been* consumed by ARC
450 * buffers in the arc_mfu_ghost state. See the comment above
451 * arcstat_mru_ghost_size for more details.
453 kstat_named_t arcstat_mfu_ghost_size;
455 * Number of bytes that *would have been* consumed by ARC
456 * buffers that are eligible for eviction, of type
457 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
459 kstat_named_t arcstat_mfu_ghost_evictable_data;
461 * Number of bytes that *would have been* consumed by ARC
462 * buffers that are eligible for eviction, of type
463 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
465 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
466 kstat_named_t arcstat_l2_hits;
467 kstat_named_t arcstat_l2_misses;
468 kstat_named_t arcstat_l2_feeds;
469 kstat_named_t arcstat_l2_rw_clash;
470 kstat_named_t arcstat_l2_read_bytes;
471 kstat_named_t arcstat_l2_write_bytes;
472 kstat_named_t arcstat_l2_writes_sent;
473 kstat_named_t arcstat_l2_writes_done;
474 kstat_named_t arcstat_l2_writes_error;
475 kstat_named_t arcstat_l2_writes_lock_retry;
476 kstat_named_t arcstat_l2_evict_lock_retry;
477 kstat_named_t arcstat_l2_evict_reading;
478 kstat_named_t arcstat_l2_evict_l1cached;
479 kstat_named_t arcstat_l2_free_on_write;
480 kstat_named_t arcstat_l2_cdata_free_on_write;
481 kstat_named_t arcstat_l2_abort_lowmem;
482 kstat_named_t arcstat_l2_cksum_bad;
483 kstat_named_t arcstat_l2_io_error;
484 kstat_named_t arcstat_l2_size;
485 kstat_named_t arcstat_l2_asize;
486 kstat_named_t arcstat_l2_hdr_size;
487 kstat_named_t arcstat_l2_compress_successes;
488 kstat_named_t arcstat_l2_compress_zeros;
489 kstat_named_t arcstat_l2_compress_failures;
490 kstat_named_t arcstat_memory_throttle_count;
491 kstat_named_t arcstat_duplicate_buffers;
492 kstat_named_t arcstat_duplicate_buffers_size;
493 kstat_named_t arcstat_duplicate_reads;
494 kstat_named_t arcstat_meta_used;
495 kstat_named_t arcstat_meta_limit;
496 kstat_named_t arcstat_meta_max;
497 kstat_named_t arcstat_meta_min;
498 kstat_named_t arcstat_sync_wait_for_async;
499 kstat_named_t arcstat_demand_hit_predictive_prefetch;
502 static arc_stats_t arc_stats = {
503 { "hits", KSTAT_DATA_UINT64 },
504 { "misses", KSTAT_DATA_UINT64 },
505 { "demand_data_hits", KSTAT_DATA_UINT64 },
506 { "demand_data_misses", KSTAT_DATA_UINT64 },
507 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
508 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
509 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
510 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
511 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
512 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
513 { "mru_hits", KSTAT_DATA_UINT64 },
514 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
515 { "mfu_hits", KSTAT_DATA_UINT64 },
516 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
517 { "deleted", KSTAT_DATA_UINT64 },
518 { "mutex_miss", KSTAT_DATA_UINT64 },
519 { "evict_skip", KSTAT_DATA_UINT64 },
520 { "evict_not_enough", KSTAT_DATA_UINT64 },
521 { "evict_l2_cached", KSTAT_DATA_UINT64 },
522 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
523 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
524 { "evict_l2_skip", KSTAT_DATA_UINT64 },
525 { "hash_elements", KSTAT_DATA_UINT64 },
526 { "hash_elements_max", KSTAT_DATA_UINT64 },
527 { "hash_collisions", KSTAT_DATA_UINT64 },
528 { "hash_chains", KSTAT_DATA_UINT64 },
529 { "hash_chain_max", KSTAT_DATA_UINT64 },
530 { "p", KSTAT_DATA_UINT64 },
531 { "c", KSTAT_DATA_UINT64 },
532 { "c_min", KSTAT_DATA_UINT64 },
533 { "c_max", KSTAT_DATA_UINT64 },
534 { "size", KSTAT_DATA_UINT64 },
535 { "hdr_size", KSTAT_DATA_UINT64 },
536 { "data_size", KSTAT_DATA_UINT64 },
537 { "metadata_size", KSTAT_DATA_UINT64 },
538 { "other_size", KSTAT_DATA_UINT64 },
539 { "anon_size", KSTAT_DATA_UINT64 },
540 { "anon_evictable_data", KSTAT_DATA_UINT64 },
541 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
542 { "mru_size", KSTAT_DATA_UINT64 },
543 { "mru_evictable_data", KSTAT_DATA_UINT64 },
544 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
545 { "mru_ghost_size", KSTAT_DATA_UINT64 },
546 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
547 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
548 { "mfu_size", KSTAT_DATA_UINT64 },
549 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
550 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
551 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
552 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
553 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
554 { "l2_hits", KSTAT_DATA_UINT64 },
555 { "l2_misses", KSTAT_DATA_UINT64 },
556 { "l2_feeds", KSTAT_DATA_UINT64 },
557 { "l2_rw_clash", KSTAT_DATA_UINT64 },
558 { "l2_read_bytes", KSTAT_DATA_UINT64 },
559 { "l2_write_bytes", KSTAT_DATA_UINT64 },
560 { "l2_writes_sent", KSTAT_DATA_UINT64 },
561 { "l2_writes_done", KSTAT_DATA_UINT64 },
562 { "l2_writes_error", KSTAT_DATA_UINT64 },
563 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
564 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
565 { "l2_evict_reading", KSTAT_DATA_UINT64 },
566 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
567 { "l2_free_on_write", KSTAT_DATA_UINT64 },
568 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 },
569 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
570 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
571 { "l2_io_error", KSTAT_DATA_UINT64 },
572 { "l2_size", KSTAT_DATA_UINT64 },
573 { "l2_asize", KSTAT_DATA_UINT64 },
574 { "l2_hdr_size", KSTAT_DATA_UINT64 },
575 { "l2_compress_successes", KSTAT_DATA_UINT64 },
576 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
577 { "l2_compress_failures", KSTAT_DATA_UINT64 },
578 { "memory_throttle_count", KSTAT_DATA_UINT64 },
579 { "duplicate_buffers", KSTAT_DATA_UINT64 },
580 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
581 { "duplicate_reads", KSTAT_DATA_UINT64 },
582 { "arc_meta_used", KSTAT_DATA_UINT64 },
583 { "arc_meta_limit", KSTAT_DATA_UINT64 },
584 { "arc_meta_max", KSTAT_DATA_UINT64 },
585 { "arc_meta_min", KSTAT_DATA_UINT64 },
586 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
587 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
590 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
592 #define ARCSTAT_INCR(stat, val) \
593 atomic_add_64(&arc_stats.stat.value.ui64, (val))
595 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
596 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
598 #define ARCSTAT_MAX(stat, val) { \
600 while ((val) > (m = arc_stats.stat.value.ui64) && \
601 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
605 #define ARCSTAT_MAXSTAT(stat) \
606 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
609 * We define a macro to allow ARC hits/misses to be easily broken down by
610 * two separate conditions, giving a total of four different subtypes for
611 * each of hits and misses (so eight statistics total).
613 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
616 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
618 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
622 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
624 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
629 static arc_state_t *arc_anon;
630 static arc_state_t *arc_mru;
631 static arc_state_t *arc_mru_ghost;
632 static arc_state_t *arc_mfu;
633 static arc_state_t *arc_mfu_ghost;
634 static arc_state_t *arc_l2c_only;
637 * There are several ARC variables that are critical to export as kstats --
638 * but we don't want to have to grovel around in the kstat whenever we wish to
639 * manipulate them. For these variables, we therefore define them to be in
640 * terms of the statistic variable. This assures that we are not introducing
641 * the possibility of inconsistency by having shadow copies of the variables,
642 * while still allowing the code to be readable.
644 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
645 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
646 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
647 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
648 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
649 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
650 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
651 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
652 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
654 #define L2ARC_IS_VALID_COMPRESS(_c_) \
655 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
657 static int arc_no_grow; /* Don't try to grow cache size */
658 static uint64_t arc_tempreserve;
659 static uint64_t arc_loaned_bytes;
661 typedef struct arc_callback arc_callback_t;
663 struct arc_callback {
665 arc_done_func_t *acb_done;
667 zio_t *acb_zio_dummy;
668 arc_callback_t *acb_next;
671 typedef struct arc_write_callback arc_write_callback_t;
673 struct arc_write_callback {
675 arc_done_func_t *awcb_ready;
676 arc_done_func_t *awcb_children_ready;
677 arc_done_func_t *awcb_physdone;
678 arc_done_func_t *awcb_done;
683 * ARC buffers are separated into multiple structs as a memory saving measure:
684 * - Common fields struct, always defined, and embedded within it:
685 * - L2-only fields, always allocated but undefined when not in L2ARC
686 * - L1-only fields, only allocated when in L1ARC
688 * Buffer in L1 Buffer only in L2
689 * +------------------------+ +------------------------+
690 * | arc_buf_hdr_t | | arc_buf_hdr_t |
694 * +------------------------+ +------------------------+
695 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
696 * | (undefined if L1-only) | | |
697 * +------------------------+ +------------------------+
698 * | l1arc_buf_hdr_t |
703 * +------------------------+
705 * Because it's possible for the L2ARC to become extremely large, we can wind
706 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
707 * is minimized by only allocating the fields necessary for an L1-cached buffer
708 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
709 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
710 * words in pointers. arc_hdr_realloc() is used to switch a header between
711 * these two allocation states.
713 typedef struct l1arc_buf_hdr {
714 kmutex_t b_freeze_lock;
717 * used for debugging wtih kmem_flags - by allocating and freeing
718 * b_thawed when the buffer is thawed, we get a record of the stack
719 * trace that thawed it.
726 /* for waiting on writes to complete */
729 /* protected by arc state mutex */
730 arc_state_t *b_state;
731 multilist_node_t b_arc_node;
733 /* updated atomically */
734 clock_t b_arc_access;
736 /* self protecting */
739 arc_callback_t *b_acb;
740 /* temporary buffer holder for in-flight compressed data */
744 typedef struct l2arc_dev l2arc_dev_t;
746 typedef struct l2arc_buf_hdr {
747 /* protected by arc_buf_hdr mutex */
748 l2arc_dev_t *b_dev; /* L2ARC device */
749 uint64_t b_daddr; /* disk address, offset byte */
750 /* real alloc'd buffer size depending on b_compress applied */
754 list_node_t b_l2node;
758 /* protected by hash lock */
762 * Even though this checksum is only set/verified when a buffer is in
763 * the L1 cache, it needs to be in the set of common fields because it
764 * must be preserved from the time before a buffer is written out to
765 * L2ARC until after it is read back in.
767 zio_cksum_t *b_freeze_cksum;
769 arc_buf_hdr_t *b_hash_next;
776 /* L2ARC fields. Undefined when not in L2ARC. */
777 l2arc_buf_hdr_t b_l2hdr;
778 /* L1ARC fields. Undefined when in l2arc_only state */
779 l1arc_buf_hdr_t b_l1hdr;
782 static arc_buf_t *arc_eviction_list;
783 static arc_buf_hdr_t arc_eviction_hdr;
785 #define GHOST_STATE(state) \
786 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
787 (state) == arc_l2c_only)
789 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
790 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
791 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
792 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
793 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
794 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
796 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
797 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
798 #define HDR_L2_READING(hdr) \
799 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
800 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
801 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
802 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
803 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
805 #define HDR_ISTYPE_METADATA(hdr) \
806 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
807 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
809 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
810 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
816 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
817 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
820 * Hash table routines
823 #define HT_LOCK_PAD 64
828 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
832 #define BUF_LOCKS 256
833 typedef struct buf_hash_table {
835 arc_buf_hdr_t **ht_table;
836 struct ht_lock ht_locks[BUF_LOCKS];
839 static buf_hash_table_t buf_hash_table;
841 #define BUF_HASH_INDEX(spa, dva, birth) \
842 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
843 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
844 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
845 #define HDR_LOCK(hdr) \
846 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
848 uint64_t zfs_crc64_table[256];
854 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
855 #define L2ARC_HEADROOM 2 /* num of writes */
857 * If we discover during ARC scan any buffers to be compressed, we boost
858 * our headroom for the next scanning cycle by this percentage multiple.
860 #define L2ARC_HEADROOM_BOOST 200
861 #define L2ARC_FEED_SECS 1 /* caching interval secs */
862 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
865 * Used to distinguish headers that are being process by
866 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
867 * address. This can happen when the header is added to the l2arc's list
868 * of buffers to write in the first stage of l2arc_write_buffers(), but
869 * has not yet been written out which happens in the second stage of
870 * l2arc_write_buffers().
872 #define L2ARC_ADDR_UNSET ((uint64_t)(-1))
874 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
875 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
877 /* L2ARC Performance Tunables */
878 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
879 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
880 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
881 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
882 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
883 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
884 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
885 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
886 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
892 vdev_t *l2ad_vdev; /* vdev */
893 spa_t *l2ad_spa; /* spa */
894 uint64_t l2ad_hand; /* next write location */
895 uint64_t l2ad_start; /* first addr on device */
896 uint64_t l2ad_end; /* last addr on device */
897 boolean_t l2ad_first; /* first sweep through */
898 boolean_t l2ad_writing; /* currently writing */
899 kmutex_t l2ad_mtx; /* lock for buffer list */
900 list_t l2ad_buflist; /* buffer list */
901 list_node_t l2ad_node; /* device list node */
902 refcount_t l2ad_alloc; /* allocated bytes */
905 static list_t L2ARC_dev_list; /* device list */
906 static list_t *l2arc_dev_list; /* device list pointer */
907 static kmutex_t l2arc_dev_mtx; /* device list mutex */
908 static l2arc_dev_t *l2arc_dev_last; /* last device used */
909 static list_t L2ARC_free_on_write; /* free after write buf list */
910 static list_t *l2arc_free_on_write; /* free after write list ptr */
911 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
912 static uint64_t l2arc_ndev; /* number of devices */
914 typedef struct l2arc_read_callback {
915 arc_buf_t *l2rcb_buf; /* read buffer */
916 spa_t *l2rcb_spa; /* spa */
917 blkptr_t l2rcb_bp; /* original blkptr */
918 zbookmark_phys_t l2rcb_zb; /* original bookmark */
919 int l2rcb_flags; /* original flags */
920 enum zio_compress l2rcb_compress; /* applied compress */
921 } l2arc_read_callback_t;
923 typedef struct l2arc_write_callback {
924 l2arc_dev_t *l2wcb_dev; /* device info */
925 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
926 } l2arc_write_callback_t;
928 typedef struct l2arc_data_free {
929 /* protected by l2arc_free_on_write_mtx */
932 void (*l2df_func)(void *, size_t);
933 list_node_t l2df_list_node;
936 static kmutex_t l2arc_feed_thr_lock;
937 static kcondvar_t l2arc_feed_thr_cv;
938 static uint8_t l2arc_thread_exit;
940 static void arc_get_data_buf(arc_buf_t *);
941 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
942 static boolean_t arc_is_overflowing();
943 static void arc_buf_watch(arc_buf_t *);
945 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
946 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
948 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
949 static void l2arc_read_done(zio_t *);
951 static boolean_t l2arc_compress_buf(arc_buf_hdr_t *);
952 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress);
953 static void l2arc_release_cdata_buf(arc_buf_hdr_t *);
956 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
958 uint8_t *vdva = (uint8_t *)dva;
959 uint64_t crc = -1ULL;
962 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
964 for (i = 0; i < sizeof (dva_t); i++)
965 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
967 crc ^= (spa>>8) ^ birth;
972 #define BUF_EMPTY(buf) \
973 ((buf)->b_dva.dva_word[0] == 0 && \
974 (buf)->b_dva.dva_word[1] == 0)
976 #define BUF_EQUAL(spa, dva, birth, buf) \
977 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
978 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
979 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
982 buf_discard_identity(arc_buf_hdr_t *hdr)
984 hdr->b_dva.dva_word[0] = 0;
985 hdr->b_dva.dva_word[1] = 0;
989 static arc_buf_hdr_t *
990 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
992 const dva_t *dva = BP_IDENTITY(bp);
993 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
994 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
995 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
998 mutex_enter(hash_lock);
999 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1000 hdr = hdr->b_hash_next) {
1001 if (BUF_EQUAL(spa, dva, birth, hdr)) {
1006 mutex_exit(hash_lock);
1012 * Insert an entry into the hash table. If there is already an element
1013 * equal to elem in the hash table, then the already existing element
1014 * will be returned and the new element will not be inserted.
1015 * Otherwise returns NULL.
1016 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1018 static arc_buf_hdr_t *
1019 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1021 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1022 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1023 arc_buf_hdr_t *fhdr;
1026 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1027 ASSERT(hdr->b_birth != 0);
1028 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1030 if (lockp != NULL) {
1032 mutex_enter(hash_lock);
1034 ASSERT(MUTEX_HELD(hash_lock));
1037 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1038 fhdr = fhdr->b_hash_next, i++) {
1039 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1043 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1044 buf_hash_table.ht_table[idx] = hdr;
1045 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
1047 /* collect some hash table performance data */
1049 ARCSTAT_BUMP(arcstat_hash_collisions);
1051 ARCSTAT_BUMP(arcstat_hash_chains);
1053 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1056 ARCSTAT_BUMP(arcstat_hash_elements);
1057 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1063 buf_hash_remove(arc_buf_hdr_t *hdr)
1065 arc_buf_hdr_t *fhdr, **hdrp;
1066 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1068 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1069 ASSERT(HDR_IN_HASH_TABLE(hdr));
1071 hdrp = &buf_hash_table.ht_table[idx];
1072 while ((fhdr = *hdrp) != hdr) {
1073 ASSERT(fhdr != NULL);
1074 hdrp = &fhdr->b_hash_next;
1076 *hdrp = hdr->b_hash_next;
1077 hdr->b_hash_next = NULL;
1078 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE;
1080 /* collect some hash table performance data */
1081 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1083 if (buf_hash_table.ht_table[idx] &&
1084 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1085 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1089 * Global data structures and functions for the buf kmem cache.
1091 static kmem_cache_t *hdr_full_cache;
1092 static kmem_cache_t *hdr_l2only_cache;
1093 static kmem_cache_t *buf_cache;
1100 kmem_free(buf_hash_table.ht_table,
1101 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1102 for (i = 0; i < BUF_LOCKS; i++)
1103 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1104 kmem_cache_destroy(hdr_full_cache);
1105 kmem_cache_destroy(hdr_l2only_cache);
1106 kmem_cache_destroy(buf_cache);
1110 * Constructor callback - called when the cache is empty
1111 * and a new buf is requested.
1115 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1117 arc_buf_hdr_t *hdr = vbuf;
1119 bzero(hdr, HDR_FULL_SIZE);
1120 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1121 refcount_create(&hdr->b_l1hdr.b_refcnt);
1122 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1123 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1124 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1131 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1133 arc_buf_hdr_t *hdr = vbuf;
1135 bzero(hdr, HDR_L2ONLY_SIZE);
1136 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1143 buf_cons(void *vbuf, void *unused, int kmflag)
1145 arc_buf_t *buf = vbuf;
1147 bzero(buf, sizeof (arc_buf_t));
1148 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1149 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1155 * Destructor callback - called when a cached buf is
1156 * no longer required.
1160 hdr_full_dest(void *vbuf, void *unused)
1162 arc_buf_hdr_t *hdr = vbuf;
1164 ASSERT(BUF_EMPTY(hdr));
1165 cv_destroy(&hdr->b_l1hdr.b_cv);
1166 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1167 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1168 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1169 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1174 hdr_l2only_dest(void *vbuf, void *unused)
1176 arc_buf_hdr_t *hdr = vbuf;
1178 ASSERT(BUF_EMPTY(hdr));
1179 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1184 buf_dest(void *vbuf, void *unused)
1186 arc_buf_t *buf = vbuf;
1188 mutex_destroy(&buf->b_evict_lock);
1189 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1193 * Reclaim callback -- invoked when memory is low.
1197 hdr_recl(void *unused)
1199 dprintf("hdr_recl called\n");
1201 * umem calls the reclaim func when we destroy the buf cache,
1202 * which is after we do arc_fini().
1205 cv_signal(&arc_reclaim_thread_cv);
1212 uint64_t hsize = 1ULL << 12;
1216 * The hash table is big enough to fill all of physical memory
1217 * with an average block size of zfs_arc_average_blocksize (default 8K).
1218 * By default, the table will take up
1219 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1221 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1224 buf_hash_table.ht_mask = hsize - 1;
1225 buf_hash_table.ht_table =
1226 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1227 if (buf_hash_table.ht_table == NULL) {
1228 ASSERT(hsize > (1ULL << 8));
1233 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1234 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1235 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1236 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1238 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1239 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1241 for (i = 0; i < 256; i++)
1242 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1243 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1245 for (i = 0; i < BUF_LOCKS; i++) {
1246 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1247 NULL, MUTEX_DEFAULT, NULL);
1252 * Transition between the two allocation states for the arc_buf_hdr struct.
1253 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1254 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1255 * version is used when a cache buffer is only in the L2ARC in order to reduce
1258 static arc_buf_hdr_t *
1259 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
1261 ASSERT(HDR_HAS_L2HDR(hdr));
1263 arc_buf_hdr_t *nhdr;
1264 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1266 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
1267 (old == hdr_l2only_cache && new == hdr_full_cache));
1269 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
1271 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
1272 buf_hash_remove(hdr);
1274 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
1276 if (new == hdr_full_cache) {
1277 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1279 * arc_access and arc_change_state need to be aware that a
1280 * header has just come out of L2ARC, so we set its state to
1281 * l2c_only even though it's about to change.
1283 nhdr->b_l1hdr.b_state = arc_l2c_only;
1285 /* Verify previous threads set to NULL before freeing */
1286 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1288 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1289 ASSERT0(hdr->b_l1hdr.b_datacnt);
1292 * If we've reached here, We must have been called from
1293 * arc_evict_hdr(), as such we should have already been
1294 * removed from any ghost list we were previously on
1295 * (which protects us from racing with arc_evict_state),
1296 * thus no locking is needed during this check.
1298 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1301 * A buffer must not be moved into the arc_l2c_only
1302 * state if it's not finished being written out to the
1303 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1304 * might try to be accessed, even though it was removed.
1306 VERIFY(!HDR_L2_WRITING(hdr));
1307 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1310 if (hdr->b_l1hdr.b_thawed != NULL) {
1311 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1312 hdr->b_l1hdr.b_thawed = NULL;
1316 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR;
1319 * The header has been reallocated so we need to re-insert it into any
1322 (void) buf_hash_insert(nhdr, NULL);
1324 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
1326 mutex_enter(&dev->l2ad_mtx);
1329 * We must place the realloc'ed header back into the list at
1330 * the same spot. Otherwise, if it's placed earlier in the list,
1331 * l2arc_write_buffers() could find it during the function's
1332 * write phase, and try to write it out to the l2arc.
1334 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
1335 list_remove(&dev->l2ad_buflist, hdr);
1337 mutex_exit(&dev->l2ad_mtx);
1340 * Since we're using the pointer address as the tag when
1341 * incrementing and decrementing the l2ad_alloc refcount, we
1342 * must remove the old pointer (that we're about to destroy) and
1343 * add the new pointer to the refcount. Otherwise we'd remove
1344 * the wrong pointer address when calling arc_hdr_destroy() later.
1347 (void) refcount_remove_many(&dev->l2ad_alloc,
1348 hdr->b_l2hdr.b_asize, hdr);
1350 (void) refcount_add_many(&dev->l2ad_alloc,
1351 nhdr->b_l2hdr.b_asize, nhdr);
1353 buf_discard_identity(hdr);
1354 hdr->b_freeze_cksum = NULL;
1355 kmem_cache_free(old, hdr);
1361 #define ARC_MINTIME (hz>>4) /* 62 ms */
1364 arc_cksum_verify(arc_buf_t *buf)
1368 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1371 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1372 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) {
1373 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1376 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc);
1377 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1378 panic("buffer modified while frozen!");
1379 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1383 arc_cksum_equal(arc_buf_t *buf)
1388 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1389 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc);
1390 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1391 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1397 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1399 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1402 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1403 if (buf->b_hdr->b_freeze_cksum != NULL) {
1404 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1407 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1408 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1409 NULL, buf->b_hdr->b_freeze_cksum);
1410 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1415 typedef struct procctl {
1423 arc_buf_unwatch(arc_buf_t *buf)
1430 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1431 ctl.prwatch.pr_size = 0;
1432 ctl.prwatch.pr_wflags = 0;
1433 result = write(arc_procfd, &ctl, sizeof (ctl));
1434 ASSERT3U(result, ==, sizeof (ctl));
1441 arc_buf_watch(arc_buf_t *buf)
1448 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1449 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1450 ctl.prwatch.pr_wflags = WA_WRITE;
1451 result = write(arc_procfd, &ctl, sizeof (ctl));
1452 ASSERT3U(result, ==, sizeof (ctl));
1457 static arc_buf_contents_t
1458 arc_buf_type(arc_buf_hdr_t *hdr)
1460 if (HDR_ISTYPE_METADATA(hdr)) {
1461 return (ARC_BUFC_METADATA);
1463 return (ARC_BUFC_DATA);
1468 arc_bufc_to_flags(arc_buf_contents_t type)
1472 /* metadata field is 0 if buffer contains normal data */
1474 case ARC_BUFC_METADATA:
1475 return (ARC_FLAG_BUFC_METADATA);
1479 panic("undefined ARC buffer type!");
1480 return ((uint32_t)-1);
1484 arc_buf_thaw(arc_buf_t *buf)
1486 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1487 if (buf->b_hdr->b_l1hdr.b_state != arc_anon)
1488 panic("modifying non-anon buffer!");
1489 if (HDR_IO_IN_PROGRESS(buf->b_hdr))
1490 panic("modifying buffer while i/o in progress!");
1491 arc_cksum_verify(buf);
1494 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1495 if (buf->b_hdr->b_freeze_cksum != NULL) {
1496 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1497 buf->b_hdr->b_freeze_cksum = NULL;
1501 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1502 if (buf->b_hdr->b_l1hdr.b_thawed != NULL)
1503 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1);
1504 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1508 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1510 arc_buf_unwatch(buf);
1514 arc_buf_freeze(arc_buf_t *buf)
1516 kmutex_t *hash_lock;
1518 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1521 hash_lock = HDR_LOCK(buf->b_hdr);
1522 mutex_enter(hash_lock);
1524 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1525 buf->b_hdr->b_l1hdr.b_state == arc_anon);
1526 arc_cksum_compute(buf, B_FALSE);
1527 mutex_exit(hash_lock);
1532 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1534 ASSERT(HDR_HAS_L1HDR(hdr));
1535 ASSERT(MUTEX_HELD(hash_lock));
1536 arc_state_t *state = hdr->b_l1hdr.b_state;
1538 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
1539 (state != arc_anon)) {
1540 /* We don't use the L2-only state list. */
1541 if (state != arc_l2c_only) {
1542 arc_buf_contents_t type = arc_buf_type(hdr);
1543 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt;
1544 multilist_t *list = &state->arcs_list[type];
1545 uint64_t *size = &state->arcs_lsize[type];
1547 multilist_remove(list, hdr);
1549 if (GHOST_STATE(state)) {
1550 ASSERT0(hdr->b_l1hdr.b_datacnt);
1551 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
1552 delta = hdr->b_size;
1555 ASSERT3U(*size, >=, delta);
1556 atomic_add_64(size, -delta);
1558 /* remove the prefetch flag if we get a reference */
1559 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
1564 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1567 arc_state_t *state = hdr->b_l1hdr.b_state;
1569 ASSERT(HDR_HAS_L1HDR(hdr));
1570 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1571 ASSERT(!GHOST_STATE(state));
1574 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1575 * check to prevent usage of the arc_l2c_only list.
1577 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
1578 (state != arc_anon)) {
1579 arc_buf_contents_t type = arc_buf_type(hdr);
1580 multilist_t *list = &state->arcs_list[type];
1581 uint64_t *size = &state->arcs_lsize[type];
1583 multilist_insert(list, hdr);
1585 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
1586 atomic_add_64(size, hdr->b_size *
1587 hdr->b_l1hdr.b_datacnt);
1593 * Move the supplied buffer to the indicated state. The hash lock
1594 * for the buffer must be held by the caller.
1597 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
1598 kmutex_t *hash_lock)
1600 arc_state_t *old_state;
1603 uint64_t from_delta, to_delta;
1604 arc_buf_contents_t buftype = arc_buf_type(hdr);
1607 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1608 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1609 * L1 hdr doesn't always exist when we change state to arc_anon before
1610 * destroying a header, in which case reallocating to add the L1 hdr is
1613 if (HDR_HAS_L1HDR(hdr)) {
1614 old_state = hdr->b_l1hdr.b_state;
1615 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
1616 datacnt = hdr->b_l1hdr.b_datacnt;
1618 old_state = arc_l2c_only;
1623 ASSERT(MUTEX_HELD(hash_lock));
1624 ASSERT3P(new_state, !=, old_state);
1625 ASSERT(refcnt == 0 || datacnt > 0);
1626 ASSERT(!GHOST_STATE(new_state) || datacnt == 0);
1627 ASSERT(old_state != arc_anon || datacnt <= 1);
1629 from_delta = to_delta = datacnt * hdr->b_size;
1632 * If this buffer is evictable, transfer it from the
1633 * old state list to the new state list.
1636 if (old_state != arc_anon && old_state != arc_l2c_only) {
1637 uint64_t *size = &old_state->arcs_lsize[buftype];
1639 ASSERT(HDR_HAS_L1HDR(hdr));
1640 multilist_remove(&old_state->arcs_list[buftype], hdr);
1643 * If prefetching out of the ghost cache,
1644 * we will have a non-zero datacnt.
1646 if (GHOST_STATE(old_state) && datacnt == 0) {
1647 /* ghost elements have a ghost size */
1648 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1649 from_delta = hdr->b_size;
1651 ASSERT3U(*size, >=, from_delta);
1652 atomic_add_64(size, -from_delta);
1654 if (new_state != arc_anon && new_state != arc_l2c_only) {
1655 uint64_t *size = &new_state->arcs_lsize[buftype];
1658 * An L1 header always exists here, since if we're
1659 * moving to some L1-cached state (i.e. not l2c_only or
1660 * anonymous), we realloc the header to add an L1hdr
1663 ASSERT(HDR_HAS_L1HDR(hdr));
1664 multilist_insert(&new_state->arcs_list[buftype], hdr);
1666 /* ghost elements have a ghost size */
1667 if (GHOST_STATE(new_state)) {
1669 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1670 to_delta = hdr->b_size;
1672 atomic_add_64(size, to_delta);
1676 ASSERT(!BUF_EMPTY(hdr));
1677 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
1678 buf_hash_remove(hdr);
1680 /* adjust state sizes (ignore arc_l2c_only) */
1682 if (to_delta && new_state != arc_l2c_only) {
1683 ASSERT(HDR_HAS_L1HDR(hdr));
1684 if (GHOST_STATE(new_state)) {
1688 * We moving a header to a ghost state, we first
1689 * remove all arc buffers. Thus, we'll have a
1690 * datacnt of zero, and no arc buffer to use for
1691 * the reference. As a result, we use the arc
1692 * header pointer for the reference.
1694 (void) refcount_add_many(&new_state->arcs_size,
1697 ASSERT3U(datacnt, !=, 0);
1700 * Each individual buffer holds a unique reference,
1701 * thus we must remove each of these references one
1704 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1705 buf = buf->b_next) {
1706 (void) refcount_add_many(&new_state->arcs_size,
1712 if (from_delta && old_state != arc_l2c_only) {
1713 ASSERT(HDR_HAS_L1HDR(hdr));
1714 if (GHOST_STATE(old_state)) {
1716 * When moving a header off of a ghost state,
1717 * there's the possibility for datacnt to be
1718 * non-zero. This is because we first add the
1719 * arc buffer to the header prior to changing
1720 * the header's state. Since we used the header
1721 * for the reference when putting the header on
1722 * the ghost state, we must balance that and use
1723 * the header when removing off the ghost state
1724 * (even though datacnt is non zero).
1727 IMPLY(datacnt == 0, new_state == arc_anon ||
1728 new_state == arc_l2c_only);
1730 (void) refcount_remove_many(&old_state->arcs_size,
1733 ASSERT3P(datacnt, !=, 0);
1736 * Each individual buffer holds a unique reference,
1737 * thus we must remove each of these references one
1740 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1741 buf = buf->b_next) {
1742 (void) refcount_remove_many(
1743 &old_state->arcs_size, hdr->b_size, buf);
1748 if (HDR_HAS_L1HDR(hdr))
1749 hdr->b_l1hdr.b_state = new_state;
1752 * L2 headers should never be on the L2 state list since they don't
1753 * have L1 headers allocated.
1755 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
1756 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
1760 arc_space_consume(uint64_t space, arc_space_type_t type)
1762 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1765 case ARC_SPACE_DATA:
1766 ARCSTAT_INCR(arcstat_data_size, space);
1768 case ARC_SPACE_META:
1769 ARCSTAT_INCR(arcstat_metadata_size, space);
1771 case ARC_SPACE_OTHER:
1772 ARCSTAT_INCR(arcstat_other_size, space);
1774 case ARC_SPACE_HDRS:
1775 ARCSTAT_INCR(arcstat_hdr_size, space);
1777 case ARC_SPACE_L2HDRS:
1778 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1782 if (type != ARC_SPACE_DATA)
1783 ARCSTAT_INCR(arcstat_meta_used, space);
1785 atomic_add_64(&arc_size, space);
1789 arc_space_return(uint64_t space, arc_space_type_t type)
1791 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1794 case ARC_SPACE_DATA:
1795 ARCSTAT_INCR(arcstat_data_size, -space);
1797 case ARC_SPACE_META:
1798 ARCSTAT_INCR(arcstat_metadata_size, -space);
1800 case ARC_SPACE_OTHER:
1801 ARCSTAT_INCR(arcstat_other_size, -space);
1803 case ARC_SPACE_HDRS:
1804 ARCSTAT_INCR(arcstat_hdr_size, -space);
1806 case ARC_SPACE_L2HDRS:
1807 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1811 if (type != ARC_SPACE_DATA) {
1812 ASSERT(arc_meta_used >= space);
1813 if (arc_meta_max < arc_meta_used)
1814 arc_meta_max = arc_meta_used;
1815 ARCSTAT_INCR(arcstat_meta_used, -space);
1818 ASSERT(arc_size >= space);
1819 atomic_add_64(&arc_size, -space);
1823 arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
1828 ASSERT3U(size, >, 0);
1829 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
1830 ASSERT(BUF_EMPTY(hdr));
1831 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
1833 hdr->b_spa = spa_load_guid(spa);
1835 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1838 buf->b_efunc = NULL;
1839 buf->b_private = NULL;
1842 hdr->b_flags = arc_bufc_to_flags(type);
1843 hdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1845 hdr->b_l1hdr.b_buf = buf;
1846 hdr->b_l1hdr.b_state = arc_anon;
1847 hdr->b_l1hdr.b_arc_access = 0;
1848 hdr->b_l1hdr.b_datacnt = 1;
1849 hdr->b_l1hdr.b_tmp_cdata = NULL;
1851 arc_get_data_buf(buf);
1852 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
1853 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
1858 static char *arc_onloan_tag = "onloan";
1861 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1862 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1863 * buffers must be returned to the arc before they can be used by the DMU or
1867 arc_loan_buf(spa_t *spa, int size)
1871 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1873 atomic_add_64(&arc_loaned_bytes, size);
1878 * Return a loaned arc buffer to the arc.
1881 arc_return_buf(arc_buf_t *buf, void *tag)
1883 arc_buf_hdr_t *hdr = buf->b_hdr;
1885 ASSERT(buf->b_data != NULL);
1886 ASSERT(HDR_HAS_L1HDR(hdr));
1887 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
1888 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
1890 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1893 /* Detach an arc_buf from a dbuf (tag) */
1895 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1897 arc_buf_hdr_t *hdr = buf->b_hdr;
1899 ASSERT(buf->b_data != NULL);
1900 ASSERT(HDR_HAS_L1HDR(hdr));
1901 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
1902 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
1903 buf->b_efunc = NULL;
1904 buf->b_private = NULL;
1906 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1910 arc_buf_clone(arc_buf_t *from)
1913 arc_buf_hdr_t *hdr = from->b_hdr;
1914 uint64_t size = hdr->b_size;
1916 ASSERT(HDR_HAS_L1HDR(hdr));
1917 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
1919 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1922 buf->b_efunc = NULL;
1923 buf->b_private = NULL;
1924 buf->b_next = hdr->b_l1hdr.b_buf;
1925 hdr->b_l1hdr.b_buf = buf;
1926 arc_get_data_buf(buf);
1927 bcopy(from->b_data, buf->b_data, size);
1930 * This buffer already exists in the arc so create a duplicate
1931 * copy for the caller. If the buffer is associated with user data
1932 * then track the size and number of duplicates. These stats will be
1933 * updated as duplicate buffers are created and destroyed.
1935 if (HDR_ISTYPE_DATA(hdr)) {
1936 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1937 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1939 hdr->b_l1hdr.b_datacnt += 1;
1944 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1947 kmutex_t *hash_lock;
1950 * Check to see if this buffer is evicted. Callers
1951 * must verify b_data != NULL to know if the add_ref
1954 mutex_enter(&buf->b_evict_lock);
1955 if (buf->b_data == NULL) {
1956 mutex_exit(&buf->b_evict_lock);
1959 hash_lock = HDR_LOCK(buf->b_hdr);
1960 mutex_enter(hash_lock);
1962 ASSERT(HDR_HAS_L1HDR(hdr));
1963 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1964 mutex_exit(&buf->b_evict_lock);
1966 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
1967 hdr->b_l1hdr.b_state == arc_mfu);
1969 add_reference(hdr, hash_lock, tag);
1970 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1971 arc_access(hdr, hash_lock);
1972 mutex_exit(hash_lock);
1973 ARCSTAT_BUMP(arcstat_hits);
1974 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
1975 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
1976 data, metadata, hits);
1980 arc_buf_free_on_write(void *data, size_t size,
1981 void (*free_func)(void *, size_t))
1983 l2arc_data_free_t *df;
1985 df = kmem_alloc(sizeof (*df), KM_SLEEP);
1986 df->l2df_data = data;
1987 df->l2df_size = size;
1988 df->l2df_func = free_func;
1989 mutex_enter(&l2arc_free_on_write_mtx);
1990 list_insert_head(l2arc_free_on_write, df);
1991 mutex_exit(&l2arc_free_on_write_mtx);
1995 * Free the arc data buffer. If it is an l2arc write in progress,
1996 * the buffer is placed on l2arc_free_on_write to be freed later.
1999 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
2001 arc_buf_hdr_t *hdr = buf->b_hdr;
2003 if (HDR_L2_WRITING(hdr)) {
2004 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
2005 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2007 free_func(buf->b_data, hdr->b_size);
2012 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
2014 ASSERT(HDR_HAS_L2HDR(hdr));
2015 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx));
2018 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
2019 * that doesn't exist, the header is in the arc_l2c_only state,
2020 * and there isn't anything to free (it's already been freed).
2022 if (!HDR_HAS_L1HDR(hdr))
2026 * The header isn't being written to the l2arc device, thus it
2027 * shouldn't have a b_tmp_cdata to free.
2029 if (!HDR_L2_WRITING(hdr)) {
2030 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2035 * The header does not have compression enabled. This can be due
2036 * to the buffer not being compressible, or because we're
2037 * freeing the buffer before the second phase of
2038 * l2arc_write_buffer() has started (which does the compression
2039 * step). In either case, b_tmp_cdata does not point to a
2040 * separately compressed buffer, so there's nothing to free (it
2041 * points to the same buffer as the arc_buf_t's b_data field).
2043 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_OFF) {
2044 hdr->b_l1hdr.b_tmp_cdata = NULL;
2049 * There's nothing to free since the buffer was all zero's and
2050 * compressed to a zero length buffer.
2052 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) {
2053 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2057 ASSERT(L2ARC_IS_VALID_COMPRESS(hdr->b_l2hdr.b_compress));
2059 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata,
2060 hdr->b_size, zio_data_buf_free);
2062 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
2063 hdr->b_l1hdr.b_tmp_cdata = NULL;
2067 * Free up buf->b_data and if 'remove' is set, then pull the
2068 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2071 arc_buf_destroy(arc_buf_t *buf, boolean_t remove)
2075 /* free up data associated with the buf */
2076 if (buf->b_data != NULL) {
2077 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
2078 uint64_t size = buf->b_hdr->b_size;
2079 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
2081 arc_cksum_verify(buf);
2082 arc_buf_unwatch(buf);
2084 if (type == ARC_BUFC_METADATA) {
2085 arc_buf_data_free(buf, zio_buf_free);
2086 arc_space_return(size, ARC_SPACE_META);
2088 ASSERT(type == ARC_BUFC_DATA);
2089 arc_buf_data_free(buf, zio_data_buf_free);
2090 arc_space_return(size, ARC_SPACE_DATA);
2093 /* protected by hash lock, if in the hash table */
2094 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) {
2095 uint64_t *cnt = &state->arcs_lsize[type];
2097 ASSERT(refcount_is_zero(
2098 &buf->b_hdr->b_l1hdr.b_refcnt));
2099 ASSERT(state != arc_anon && state != arc_l2c_only);
2101 ASSERT3U(*cnt, >=, size);
2102 atomic_add_64(cnt, -size);
2105 (void) refcount_remove_many(&state->arcs_size, size, buf);
2109 * If we're destroying a duplicate buffer make sure
2110 * that the appropriate statistics are updated.
2112 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 &&
2113 HDR_ISTYPE_DATA(buf->b_hdr)) {
2114 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
2115 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
2117 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0);
2118 buf->b_hdr->b_l1hdr.b_datacnt -= 1;
2121 /* only remove the buf if requested */
2125 /* remove the buf from the hdr list */
2126 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf;
2127 bufp = &(*bufp)->b_next)
2129 *bufp = buf->b_next;
2132 ASSERT(buf->b_efunc == NULL);
2134 /* clean up the buf */
2136 kmem_cache_free(buf_cache, buf);
2140 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2142 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2143 l2arc_dev_t *dev = l2hdr->b_dev;
2145 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2146 ASSERT(HDR_HAS_L2HDR(hdr));
2148 list_remove(&dev->l2ad_buflist, hdr);
2151 * We don't want to leak the b_tmp_cdata buffer that was
2152 * allocated in l2arc_write_buffers()
2154 arc_buf_l2_cdata_free(hdr);
2157 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2158 * this header is being processed by l2arc_write_buffers() (i.e.
2159 * it's in the first stage of l2arc_write_buffers()).
2160 * Re-affirming that truth here, just to serve as a reminder. If
2161 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2162 * may not have its HDR_L2_WRITING flag set. (the write may have
2163 * completed, in which case HDR_L2_WRITING will be false and the
2164 * b_daddr field will point to the address of the buffer on disk).
2166 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr));
2169 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2170 * l2arc_write_buffers(). Since we've just removed this header
2171 * from the l2arc buffer list, this header will never reach the
2172 * second stage of l2arc_write_buffers(), which increments the
2173 * accounting stats for this header. Thus, we must be careful
2174 * not to decrement them for this header either.
2176 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) {
2177 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
2178 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
2180 vdev_space_update(dev->l2ad_vdev,
2181 -l2hdr->b_asize, 0, 0);
2183 (void) refcount_remove_many(&dev->l2ad_alloc,
2184 l2hdr->b_asize, hdr);
2187 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
2191 arc_hdr_destroy(arc_buf_hdr_t *hdr)
2193 if (HDR_HAS_L1HDR(hdr)) {
2194 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
2195 hdr->b_l1hdr.b_datacnt > 0);
2196 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2197 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2199 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2200 ASSERT(!HDR_IN_HASH_TABLE(hdr));
2202 if (HDR_HAS_L2HDR(hdr)) {
2203 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2204 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
2207 mutex_enter(&dev->l2ad_mtx);
2210 * Even though we checked this conditional above, we
2211 * need to check this again now that we have the
2212 * l2ad_mtx. This is because we could be racing with
2213 * another thread calling l2arc_evict() which might have
2214 * destroyed this header's L2 portion as we were waiting
2215 * to acquire the l2ad_mtx. If that happens, we don't
2216 * want to re-destroy the header's L2 portion.
2218 if (HDR_HAS_L2HDR(hdr))
2219 arc_hdr_l2hdr_destroy(hdr);
2222 mutex_exit(&dev->l2ad_mtx);
2225 if (!BUF_EMPTY(hdr))
2226 buf_discard_identity(hdr);
2228 if (hdr->b_freeze_cksum != NULL) {
2229 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
2230 hdr->b_freeze_cksum = NULL;
2233 if (HDR_HAS_L1HDR(hdr)) {
2234 while (hdr->b_l1hdr.b_buf) {
2235 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2237 if (buf->b_efunc != NULL) {
2238 mutex_enter(&arc_user_evicts_lock);
2239 mutex_enter(&buf->b_evict_lock);
2240 ASSERT(buf->b_hdr != NULL);
2241 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE);
2242 hdr->b_l1hdr.b_buf = buf->b_next;
2243 buf->b_hdr = &arc_eviction_hdr;
2244 buf->b_next = arc_eviction_list;
2245 arc_eviction_list = buf;
2246 mutex_exit(&buf->b_evict_lock);
2247 cv_signal(&arc_user_evicts_cv);
2248 mutex_exit(&arc_user_evicts_lock);
2250 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE);
2254 if (hdr->b_l1hdr.b_thawed != NULL) {
2255 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2256 hdr->b_l1hdr.b_thawed = NULL;
2261 ASSERT3P(hdr->b_hash_next, ==, NULL);
2262 if (HDR_HAS_L1HDR(hdr)) {
2263 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2264 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
2265 kmem_cache_free(hdr_full_cache, hdr);
2267 kmem_cache_free(hdr_l2only_cache, hdr);
2272 arc_buf_free(arc_buf_t *buf, void *tag)
2274 arc_buf_hdr_t *hdr = buf->b_hdr;
2275 int hashed = hdr->b_l1hdr.b_state != arc_anon;
2277 ASSERT(buf->b_efunc == NULL);
2278 ASSERT(buf->b_data != NULL);
2281 kmutex_t *hash_lock = HDR_LOCK(hdr);
2283 mutex_enter(hash_lock);
2285 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2287 (void) remove_reference(hdr, hash_lock, tag);
2288 if (hdr->b_l1hdr.b_datacnt > 1) {
2289 arc_buf_destroy(buf, TRUE);
2291 ASSERT(buf == hdr->b_l1hdr.b_buf);
2292 ASSERT(buf->b_efunc == NULL);
2293 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2295 mutex_exit(hash_lock);
2296 } else if (HDR_IO_IN_PROGRESS(hdr)) {
2299 * We are in the middle of an async write. Don't destroy
2300 * this buffer unless the write completes before we finish
2301 * decrementing the reference count.
2303 mutex_enter(&arc_user_evicts_lock);
2304 (void) remove_reference(hdr, NULL, tag);
2305 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2306 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
2307 mutex_exit(&arc_user_evicts_lock);
2309 arc_hdr_destroy(hdr);
2311 if (remove_reference(hdr, NULL, tag) > 0)
2312 arc_buf_destroy(buf, TRUE);
2314 arc_hdr_destroy(hdr);
2319 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
2321 arc_buf_hdr_t *hdr = buf->b_hdr;
2322 kmutex_t *hash_lock = HDR_LOCK(hdr);
2323 boolean_t no_callback = (buf->b_efunc == NULL);
2325 if (hdr->b_l1hdr.b_state == arc_anon) {
2326 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
2327 arc_buf_free(buf, tag);
2328 return (no_callback);
2331 mutex_enter(hash_lock);
2333 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
2334 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2335 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2336 ASSERT(buf->b_data != NULL);
2338 (void) remove_reference(hdr, hash_lock, tag);
2339 if (hdr->b_l1hdr.b_datacnt > 1) {
2341 arc_buf_destroy(buf, TRUE);
2342 } else if (no_callback) {
2343 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL);
2344 ASSERT(buf->b_efunc == NULL);
2345 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2347 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 ||
2348 refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2349 mutex_exit(hash_lock);
2350 return (no_callback);
2354 arc_buf_size(arc_buf_t *buf)
2356 return (buf->b_hdr->b_size);
2360 * Called from the DMU to determine if the current buffer should be
2361 * evicted. In order to ensure proper locking, the eviction must be initiated
2362 * from the DMU. Return true if the buffer is associated with user data and
2363 * duplicate buffers still exist.
2366 arc_buf_eviction_needed(arc_buf_t *buf)
2369 boolean_t evict_needed = B_FALSE;
2371 if (zfs_disable_dup_eviction)
2374 mutex_enter(&buf->b_evict_lock);
2378 * We are in arc_do_user_evicts(); let that function
2379 * perform the eviction.
2381 ASSERT(buf->b_data == NULL);
2382 mutex_exit(&buf->b_evict_lock);
2384 } else if (buf->b_data == NULL) {
2386 * We have already been added to the arc eviction list;
2387 * recommend eviction.
2389 ASSERT3P(hdr, ==, &arc_eviction_hdr);
2390 mutex_exit(&buf->b_evict_lock);
2394 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr))
2395 evict_needed = B_TRUE;
2397 mutex_exit(&buf->b_evict_lock);
2398 return (evict_needed);
2402 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2403 * state of the header is dependent on it's state prior to entering this
2404 * function. The following transitions are possible:
2406 * - arc_mru -> arc_mru_ghost
2407 * - arc_mfu -> arc_mfu_ghost
2408 * - arc_mru_ghost -> arc_l2c_only
2409 * - arc_mru_ghost -> deleted
2410 * - arc_mfu_ghost -> arc_l2c_only
2411 * - arc_mfu_ghost -> deleted
2414 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
2416 arc_state_t *evicted_state, *state;
2417 int64_t bytes_evicted = 0;
2419 ASSERT(MUTEX_HELD(hash_lock));
2420 ASSERT(HDR_HAS_L1HDR(hdr));
2422 state = hdr->b_l1hdr.b_state;
2423 if (GHOST_STATE(state)) {
2424 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2425 ASSERT(hdr->b_l1hdr.b_buf == NULL);
2428 * l2arc_write_buffers() relies on a header's L1 portion
2429 * (i.e. it's b_tmp_cdata field) during it's write phase.
2430 * Thus, we cannot push a header onto the arc_l2c_only
2431 * state (removing it's L1 piece) until the header is
2432 * done being written to the l2arc.
2434 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
2435 ARCSTAT_BUMP(arcstat_evict_l2_skip);
2436 return (bytes_evicted);
2439 ARCSTAT_BUMP(arcstat_deleted);
2440 bytes_evicted += hdr->b_size;
2442 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
2444 if (HDR_HAS_L2HDR(hdr)) {
2446 * This buffer is cached on the 2nd Level ARC;
2447 * don't destroy the header.
2449 arc_change_state(arc_l2c_only, hdr, hash_lock);
2451 * dropping from L1+L2 cached to L2-only,
2452 * realloc to remove the L1 header.
2454 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
2457 arc_change_state(arc_anon, hdr, hash_lock);
2458 arc_hdr_destroy(hdr);
2460 return (bytes_evicted);
2463 ASSERT(state == arc_mru || state == arc_mfu);
2464 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2466 /* prefetch buffers have a minimum lifespan */
2467 if (HDR_IO_IN_PROGRESS(hdr) ||
2468 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
2469 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
2470 arc_min_prefetch_lifespan)) {
2471 ARCSTAT_BUMP(arcstat_evict_skip);
2472 return (bytes_evicted);
2475 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
2476 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0);
2477 while (hdr->b_l1hdr.b_buf) {
2478 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2479 if (!mutex_tryenter(&buf->b_evict_lock)) {
2480 ARCSTAT_BUMP(arcstat_mutex_miss);
2483 if (buf->b_data != NULL)
2484 bytes_evicted += hdr->b_size;
2485 if (buf->b_efunc != NULL) {
2486 mutex_enter(&arc_user_evicts_lock);
2487 arc_buf_destroy(buf, FALSE);
2488 hdr->b_l1hdr.b_buf = buf->b_next;
2489 buf->b_hdr = &arc_eviction_hdr;
2490 buf->b_next = arc_eviction_list;
2491 arc_eviction_list = buf;
2492 cv_signal(&arc_user_evicts_cv);
2493 mutex_exit(&arc_user_evicts_lock);
2494 mutex_exit(&buf->b_evict_lock);
2496 mutex_exit(&buf->b_evict_lock);
2497 arc_buf_destroy(buf, TRUE);
2501 if (HDR_HAS_L2HDR(hdr)) {
2502 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size);
2504 if (l2arc_write_eligible(hdr->b_spa, hdr))
2505 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size);
2507 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size);
2510 if (hdr->b_l1hdr.b_datacnt == 0) {
2511 arc_change_state(evicted_state, hdr, hash_lock);
2512 ASSERT(HDR_IN_HASH_TABLE(hdr));
2513 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
2514 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
2515 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
2518 return (bytes_evicted);
2522 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
2523 uint64_t spa, int64_t bytes)
2525 multilist_sublist_t *mls;
2526 uint64_t bytes_evicted = 0;
2528 kmutex_t *hash_lock;
2529 int evict_count = 0;
2531 ASSERT3P(marker, !=, NULL);
2532 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2534 mls = multilist_sublist_lock(ml, idx);
2536 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
2537 hdr = multilist_sublist_prev(mls, marker)) {
2538 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
2539 (evict_count >= zfs_arc_evict_batch_limit))
2543 * To keep our iteration location, move the marker
2544 * forward. Since we're not holding hdr's hash lock, we
2545 * must be very careful and not remove 'hdr' from the
2546 * sublist. Otherwise, other consumers might mistake the
2547 * 'hdr' as not being on a sublist when they call the
2548 * multilist_link_active() function (they all rely on
2549 * the hash lock protecting concurrent insertions and
2550 * removals). multilist_sublist_move_forward() was
2551 * specifically implemented to ensure this is the case
2552 * (only 'marker' will be removed and re-inserted).
2554 multilist_sublist_move_forward(mls, marker);
2557 * The only case where the b_spa field should ever be
2558 * zero, is the marker headers inserted by
2559 * arc_evict_state(). It's possible for multiple threads
2560 * to be calling arc_evict_state() concurrently (e.g.
2561 * dsl_pool_close() and zio_inject_fault()), so we must
2562 * skip any markers we see from these other threads.
2564 if (hdr->b_spa == 0)
2567 /* we're only interested in evicting buffers of a certain spa */
2568 if (spa != 0 && hdr->b_spa != spa) {
2569 ARCSTAT_BUMP(arcstat_evict_skip);
2573 hash_lock = HDR_LOCK(hdr);
2576 * We aren't calling this function from any code path
2577 * that would already be holding a hash lock, so we're
2578 * asserting on this assumption to be defensive in case
2579 * this ever changes. Without this check, it would be
2580 * possible to incorrectly increment arcstat_mutex_miss
2581 * below (e.g. if the code changed such that we called
2582 * this function with a hash lock held).
2584 ASSERT(!MUTEX_HELD(hash_lock));
2586 if (mutex_tryenter(hash_lock)) {
2587 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
2588 mutex_exit(hash_lock);
2590 bytes_evicted += evicted;
2593 * If evicted is zero, arc_evict_hdr() must have
2594 * decided to skip this header, don't increment
2595 * evict_count in this case.
2601 * If arc_size isn't overflowing, signal any
2602 * threads that might happen to be waiting.
2604 * For each header evicted, we wake up a single
2605 * thread. If we used cv_broadcast, we could
2606 * wake up "too many" threads causing arc_size
2607 * to significantly overflow arc_c; since
2608 * arc_get_data_buf() doesn't check for overflow
2609 * when it's woken up (it doesn't because it's
2610 * possible for the ARC to be overflowing while
2611 * full of un-evictable buffers, and the
2612 * function should proceed in this case).
2614 * If threads are left sleeping, due to not
2615 * using cv_broadcast, they will be woken up
2616 * just before arc_reclaim_thread() sleeps.
2618 mutex_enter(&arc_reclaim_lock);
2619 if (!arc_is_overflowing())
2620 cv_signal(&arc_reclaim_waiters_cv);
2621 mutex_exit(&arc_reclaim_lock);
2623 ARCSTAT_BUMP(arcstat_mutex_miss);
2627 multilist_sublist_unlock(mls);
2629 return (bytes_evicted);
2633 * Evict buffers from the given arc state, until we've removed the
2634 * specified number of bytes. Move the removed buffers to the
2635 * appropriate evict state.
2637 * This function makes a "best effort". It skips over any buffers
2638 * it can't get a hash_lock on, and so, may not catch all candidates.
2639 * It may also return without evicting as much space as requested.
2641 * If bytes is specified using the special value ARC_EVICT_ALL, this
2642 * will evict all available (i.e. unlocked and evictable) buffers from
2643 * the given arc state; which is used by arc_flush().
2646 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
2647 arc_buf_contents_t type)
2649 uint64_t total_evicted = 0;
2650 multilist_t *ml = &state->arcs_list[type];
2652 arc_buf_hdr_t **markers;
2654 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2656 num_sublists = multilist_get_num_sublists(ml);
2659 * If we've tried to evict from each sublist, made some
2660 * progress, but still have not hit the target number of bytes
2661 * to evict, we want to keep trying. The markers allow us to
2662 * pick up where we left off for each individual sublist, rather
2663 * than starting from the tail each time.
2665 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
2666 for (int i = 0; i < num_sublists; i++) {
2667 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
2670 * A b_spa of 0 is used to indicate that this header is
2671 * a marker. This fact is used in arc_adjust_type() and
2672 * arc_evict_state_impl().
2674 markers[i]->b_spa = 0;
2676 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2677 multilist_sublist_insert_tail(mls, markers[i]);
2678 multilist_sublist_unlock(mls);
2682 * While we haven't hit our target number of bytes to evict, or
2683 * we're evicting all available buffers.
2685 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
2687 * Start eviction using a randomly selected sublist,
2688 * this is to try and evenly balance eviction across all
2689 * sublists. Always starting at the same sublist
2690 * (e.g. index 0) would cause evictions to favor certain
2691 * sublists over others.
2693 int sublist_idx = multilist_get_random_index(ml);
2694 uint64_t scan_evicted = 0;
2696 for (int i = 0; i < num_sublists; i++) {
2697 uint64_t bytes_remaining;
2698 uint64_t bytes_evicted;
2700 if (bytes == ARC_EVICT_ALL)
2701 bytes_remaining = ARC_EVICT_ALL;
2702 else if (total_evicted < bytes)
2703 bytes_remaining = bytes - total_evicted;
2707 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
2708 markers[sublist_idx], spa, bytes_remaining);
2710 scan_evicted += bytes_evicted;
2711 total_evicted += bytes_evicted;
2713 /* we've reached the end, wrap to the beginning */
2714 if (++sublist_idx >= num_sublists)
2719 * If we didn't evict anything during this scan, we have
2720 * no reason to believe we'll evict more during another
2721 * scan, so break the loop.
2723 if (scan_evicted == 0) {
2724 /* This isn't possible, let's make that obvious */
2725 ASSERT3S(bytes, !=, 0);
2728 * When bytes is ARC_EVICT_ALL, the only way to
2729 * break the loop is when scan_evicted is zero.
2730 * In that case, we actually have evicted enough,
2731 * so we don't want to increment the kstat.
2733 if (bytes != ARC_EVICT_ALL) {
2734 ASSERT3S(total_evicted, <, bytes);
2735 ARCSTAT_BUMP(arcstat_evict_not_enough);
2742 for (int i = 0; i < num_sublists; i++) {
2743 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2744 multilist_sublist_remove(mls, markers[i]);
2745 multilist_sublist_unlock(mls);
2747 kmem_cache_free(hdr_full_cache, markers[i]);
2749 kmem_free(markers, sizeof (*markers) * num_sublists);
2751 return (total_evicted);
2755 * Flush all "evictable" data of the given type from the arc state
2756 * specified. This will not evict any "active" buffers (i.e. referenced).
2758 * When 'retry' is set to FALSE, the function will make a single pass
2759 * over the state and evict any buffers that it can. Since it doesn't
2760 * continually retry the eviction, it might end up leaving some buffers
2761 * in the ARC due to lock misses.
2763 * When 'retry' is set to TRUE, the function will continually retry the
2764 * eviction until *all* evictable buffers have been removed from the
2765 * state. As a result, if concurrent insertions into the state are
2766 * allowed (e.g. if the ARC isn't shutting down), this function might
2767 * wind up in an infinite loop, continually trying to evict buffers.
2770 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
2773 uint64_t evicted = 0;
2775 while (state->arcs_lsize[type] != 0) {
2776 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
2786 * Evict the specified number of bytes from the state specified,
2787 * restricting eviction to the spa and type given. This function
2788 * prevents us from trying to evict more from a state's list than
2789 * is "evictable", and to skip evicting altogether when passed a
2790 * negative value for "bytes". In contrast, arc_evict_state() will
2791 * evict everything it can, when passed a negative value for "bytes".
2794 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
2795 arc_buf_contents_t type)
2799 if (bytes > 0 && state->arcs_lsize[type] > 0) {
2800 delta = MIN(state->arcs_lsize[type], bytes);
2801 return (arc_evict_state(state, spa, delta, type));
2808 * Evict metadata buffers from the cache, such that arc_meta_used is
2809 * capped by the arc_meta_limit tunable.
2812 arc_adjust_meta(void)
2814 uint64_t total_evicted = 0;
2818 * If we're over the meta limit, we want to evict enough
2819 * metadata to get back under the meta limit. We don't want to
2820 * evict so much that we drop the MRU below arc_p, though. If
2821 * we're over the meta limit more than we're over arc_p, we
2822 * evict some from the MRU here, and some from the MFU below.
2824 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
2825 (int64_t)(refcount_count(&arc_anon->arcs_size) +
2826 refcount_count(&arc_mru->arcs_size) - arc_p));
2828 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2831 * Similar to the above, we want to evict enough bytes to get us
2832 * below the meta limit, but not so much as to drop us below the
2833 * space alloted to the MFU (which is defined as arc_c - arc_p).
2835 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
2836 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
2838 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
2840 return (total_evicted);
2844 * Return the type of the oldest buffer in the given arc state
2846 * This function will select a random sublist of type ARC_BUFC_DATA and
2847 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
2848 * is compared, and the type which contains the "older" buffer will be
2851 static arc_buf_contents_t
2852 arc_adjust_type(arc_state_t *state)
2854 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
2855 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
2856 int data_idx = multilist_get_random_index(data_ml);
2857 int meta_idx = multilist_get_random_index(meta_ml);
2858 multilist_sublist_t *data_mls;
2859 multilist_sublist_t *meta_mls;
2860 arc_buf_contents_t type;
2861 arc_buf_hdr_t *data_hdr;
2862 arc_buf_hdr_t *meta_hdr;
2865 * We keep the sublist lock until we're finished, to prevent
2866 * the headers from being destroyed via arc_evict_state().
2868 data_mls = multilist_sublist_lock(data_ml, data_idx);
2869 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
2872 * These two loops are to ensure we skip any markers that
2873 * might be at the tail of the lists due to arc_evict_state().
2876 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
2877 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
2878 if (data_hdr->b_spa != 0)
2882 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
2883 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
2884 if (meta_hdr->b_spa != 0)
2888 if (data_hdr == NULL && meta_hdr == NULL) {
2889 type = ARC_BUFC_DATA;
2890 } else if (data_hdr == NULL) {
2891 ASSERT3P(meta_hdr, !=, NULL);
2892 type = ARC_BUFC_METADATA;
2893 } else if (meta_hdr == NULL) {
2894 ASSERT3P(data_hdr, !=, NULL);
2895 type = ARC_BUFC_DATA;
2897 ASSERT3P(data_hdr, !=, NULL);
2898 ASSERT3P(meta_hdr, !=, NULL);
2900 /* The headers can't be on the sublist without an L1 header */
2901 ASSERT(HDR_HAS_L1HDR(data_hdr));
2902 ASSERT(HDR_HAS_L1HDR(meta_hdr));
2904 if (data_hdr->b_l1hdr.b_arc_access <
2905 meta_hdr->b_l1hdr.b_arc_access) {
2906 type = ARC_BUFC_DATA;
2908 type = ARC_BUFC_METADATA;
2912 multilist_sublist_unlock(meta_mls);
2913 multilist_sublist_unlock(data_mls);
2919 * Evict buffers from the cache, such that arc_size is capped by arc_c.
2924 uint64_t total_evicted = 0;
2929 * If we're over arc_meta_limit, we want to correct that before
2930 * potentially evicting data buffers below.
2932 total_evicted += arc_adjust_meta();
2937 * If we're over the target cache size, we want to evict enough
2938 * from the list to get back to our target size. We don't want
2939 * to evict too much from the MRU, such that it drops below
2940 * arc_p. So, if we're over our target cache size more than
2941 * the MRU is over arc_p, we'll evict enough to get back to
2942 * arc_p here, and then evict more from the MFU below.
2944 target = MIN((int64_t)(arc_size - arc_c),
2945 (int64_t)(refcount_count(&arc_anon->arcs_size) +
2946 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
2949 * If we're below arc_meta_min, always prefer to evict data.
2950 * Otherwise, try to satisfy the requested number of bytes to
2951 * evict from the type which contains older buffers; in an
2952 * effort to keep newer buffers in the cache regardless of their
2953 * type. If we cannot satisfy the number of bytes from this
2954 * type, spill over into the next type.
2956 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
2957 arc_meta_used > arc_meta_min) {
2958 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2959 total_evicted += bytes;
2962 * If we couldn't evict our target number of bytes from
2963 * metadata, we try to get the rest from data.
2968 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
2970 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
2971 total_evicted += bytes;
2974 * If we couldn't evict our target number of bytes from
2975 * data, we try to get the rest from metadata.
2980 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2986 * Now that we've tried to evict enough from the MRU to get its
2987 * size back to arc_p, if we're still above the target cache
2988 * size, we evict the rest from the MFU.
2990 target = arc_size - arc_c;
2992 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
2993 arc_meta_used > arc_meta_min) {
2994 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
2995 total_evicted += bytes;
2998 * If we couldn't evict our target number of bytes from
2999 * metadata, we try to get the rest from data.
3004 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3006 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3007 total_evicted += bytes;
3010 * If we couldn't evict our target number of bytes from
3011 * data, we try to get the rest from data.
3016 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3020 * Adjust ghost lists
3022 * In addition to the above, the ARC also defines target values
3023 * for the ghost lists. The sum of the mru list and mru ghost
3024 * list should never exceed the target size of the cache, and
3025 * the sum of the mru list, mfu list, mru ghost list, and mfu
3026 * ghost list should never exceed twice the target size of the
3027 * cache. The following logic enforces these limits on the ghost
3028 * caches, and evicts from them as needed.
3030 target = refcount_count(&arc_mru->arcs_size) +
3031 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3033 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3034 total_evicted += bytes;
3039 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3042 * We assume the sum of the mru list and mfu list is less than
3043 * or equal to arc_c (we enforced this above), which means we
3044 * can use the simpler of the two equations below:
3046 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3047 * mru ghost + mfu ghost <= arc_c
3049 target = refcount_count(&arc_mru_ghost->arcs_size) +
3050 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3052 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3053 total_evicted += bytes;
3058 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3060 return (total_evicted);
3064 arc_do_user_evicts(void)
3066 mutex_enter(&arc_user_evicts_lock);
3067 while (arc_eviction_list != NULL) {
3068 arc_buf_t *buf = arc_eviction_list;
3069 arc_eviction_list = buf->b_next;
3070 mutex_enter(&buf->b_evict_lock);
3072 mutex_exit(&buf->b_evict_lock);
3073 mutex_exit(&arc_user_evicts_lock);
3075 if (buf->b_efunc != NULL)
3076 VERIFY0(buf->b_efunc(buf->b_private));
3078 buf->b_efunc = NULL;
3079 buf->b_private = NULL;
3080 kmem_cache_free(buf_cache, buf);
3081 mutex_enter(&arc_user_evicts_lock);
3083 mutex_exit(&arc_user_evicts_lock);
3087 arc_flush(spa_t *spa, boolean_t retry)
3092 * If retry is TRUE, a spa must not be specified since we have
3093 * no good way to determine if all of a spa's buffers have been
3094 * evicted from an arc state.
3096 ASSERT(!retry || spa == 0);
3099 guid = spa_load_guid(spa);
3101 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3102 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3104 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3105 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3107 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3108 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3110 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3111 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3113 arc_do_user_evicts();
3114 ASSERT(spa || arc_eviction_list == NULL);
3118 arc_shrink(int64_t to_free)
3120 if (arc_c > arc_c_min) {
3122 if (arc_c > arc_c_min + to_free)
3123 atomic_add_64(&arc_c, -to_free);
3127 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3128 if (arc_c > arc_size)
3129 arc_c = MAX(arc_size, arc_c_min);
3131 arc_p = (arc_c >> 1);
3132 ASSERT(arc_c >= arc_c_min);
3133 ASSERT((int64_t)arc_p >= 0);
3136 if (arc_size > arc_c)
3137 (void) arc_adjust();
3140 typedef enum free_memory_reason_t {
3145 FMR_PAGES_PP_MAXIMUM,
3148 } free_memory_reason_t;
3150 int64_t last_free_memory;
3151 free_memory_reason_t last_free_reason;
3154 * Additional reserve of pages for pp_reserve.
3156 int64_t arc_pages_pp_reserve = 64;
3159 * Additional reserve of pages for swapfs.
3161 int64_t arc_swapfs_reserve = 64;
3164 * Return the amount of memory that can be consumed before reclaim will be
3165 * needed. Positive if there is sufficient free memory, negative indicates
3166 * the amount of memory that needs to be freed up.
3169 arc_available_memory(void)
3171 int64_t lowest = INT64_MAX;
3173 free_memory_reason_t r = FMR_UNKNOWN;
3177 n = PAGESIZE * (-needfree);
3185 * check that we're out of range of the pageout scanner. It starts to
3186 * schedule paging if freemem is less than lotsfree and needfree.
3187 * lotsfree is the high-water mark for pageout, and needfree is the
3188 * number of needed free pages. We add extra pages here to make sure
3189 * the scanner doesn't start up while we're freeing memory.
3191 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3198 * check to make sure that swapfs has enough space so that anon
3199 * reservations can still succeed. anon_resvmem() checks that the
3200 * availrmem is greater than swapfs_minfree, and the number of reserved
3201 * swap pages. We also add a bit of extra here just to prevent
3202 * circumstances from getting really dire.
3204 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3205 desfree - arc_swapfs_reserve);
3208 r = FMR_SWAPFS_MINFREE;
3213 * Check that we have enough availrmem that memory locking (e.g., via
3214 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3215 * stores the number of pages that cannot be locked; when availrmem
3216 * drops below pages_pp_maximum, page locking mechanisms such as
3217 * page_pp_lock() will fail.)
3219 n = PAGESIZE * (availrmem - pages_pp_maximum -
3220 arc_pages_pp_reserve);
3223 r = FMR_PAGES_PP_MAXIMUM;
3228 * If we're on an i386 platform, it's possible that we'll exhaust the
3229 * kernel heap space before we ever run out of available physical
3230 * memory. Most checks of the size of the heap_area compare against
3231 * tune.t_minarmem, which is the minimum available real memory that we
3232 * can have in the system. However, this is generally fixed at 25 pages
3233 * which is so low that it's useless. In this comparison, we seek to
3234 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3235 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3238 n = vmem_size(heap_arena, VMEM_FREE) -
3239 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3247 * If zio data pages are being allocated out of a separate heap segment,
3248 * then enforce that the size of available vmem for this arena remains
3249 * above about 1/16th free.
3251 * Note: The 1/16th arena free requirement was put in place
3252 * to aggressively evict memory from the arc in order to avoid
3253 * memory fragmentation issues.
3255 if (zio_arena != NULL) {
3256 n = vmem_size(zio_arena, VMEM_FREE) -
3257 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3264 /* Every 100 calls, free a small amount */
3265 if (spa_get_random(100) == 0)
3269 last_free_memory = lowest;
3270 last_free_reason = r;
3277 * Determine if the system is under memory pressure and is asking
3278 * to reclaim memory. A return value of TRUE indicates that the system
3279 * is under memory pressure and that the arc should adjust accordingly.
3282 arc_reclaim_needed(void)
3284 return (arc_available_memory() < 0);
3288 arc_kmem_reap_now(void)
3291 kmem_cache_t *prev_cache = NULL;
3292 kmem_cache_t *prev_data_cache = NULL;
3293 extern kmem_cache_t *zio_buf_cache[];
3294 extern kmem_cache_t *zio_data_buf_cache[];
3295 extern kmem_cache_t *range_seg_cache;
3298 if (arc_meta_used >= arc_meta_limit) {
3300 * We are exceeding our meta-data cache limit.
3301 * Purge some DNLC entries to release holds on meta-data.
3303 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
3307 * Reclaim unused memory from all kmem caches.
3313 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
3314 if (zio_buf_cache[i] != prev_cache) {
3315 prev_cache = zio_buf_cache[i];
3316 kmem_cache_reap_now(zio_buf_cache[i]);
3318 if (zio_data_buf_cache[i] != prev_data_cache) {
3319 prev_data_cache = zio_data_buf_cache[i];
3320 kmem_cache_reap_now(zio_data_buf_cache[i]);
3323 kmem_cache_reap_now(buf_cache);
3324 kmem_cache_reap_now(hdr_full_cache);
3325 kmem_cache_reap_now(hdr_l2only_cache);
3326 kmem_cache_reap_now(range_seg_cache);
3328 if (zio_arena != NULL) {
3330 * Ask the vmem arena to reclaim unused memory from its
3333 vmem_qcache_reap(zio_arena);
3338 * Threads can block in arc_get_data_buf() waiting for this thread to evict
3339 * enough data and signal them to proceed. When this happens, the threads in
3340 * arc_get_data_buf() are sleeping while holding the hash lock for their
3341 * particular arc header. Thus, we must be careful to never sleep on a
3342 * hash lock in this thread. This is to prevent the following deadlock:
3344 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3345 * waiting for the reclaim thread to signal it.
3347 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3348 * fails, and goes to sleep forever.
3350 * This possible deadlock is avoided by always acquiring a hash lock
3351 * using mutex_tryenter() from arc_reclaim_thread().
3354 arc_reclaim_thread(void)
3356 hrtime_t growtime = 0;
3359 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
3361 mutex_enter(&arc_reclaim_lock);
3362 while (!arc_reclaim_thread_exit) {
3363 int64_t free_memory = arc_available_memory();
3364 uint64_t evicted = 0;
3366 mutex_exit(&arc_reclaim_lock);
3368 if (free_memory < 0) {
3370 arc_no_grow = B_TRUE;
3374 * Wait at least zfs_grow_retry (default 60) seconds
3375 * before considering growing.
3377 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
3379 arc_kmem_reap_now();
3382 * If we are still low on memory, shrink the ARC
3383 * so that we have arc_shrink_min free space.
3385 free_memory = arc_available_memory();
3388 (arc_c >> arc_shrink_shift) - free_memory;
3391 to_free = MAX(to_free, ptob(needfree));
3393 arc_shrink(to_free);
3395 } else if (free_memory < arc_c >> arc_no_grow_shift) {
3396 arc_no_grow = B_TRUE;
3397 } else if (gethrtime() >= growtime) {
3398 arc_no_grow = B_FALSE;
3401 evicted = arc_adjust();
3403 mutex_enter(&arc_reclaim_lock);
3406 * If evicted is zero, we couldn't evict anything via
3407 * arc_adjust(). This could be due to hash lock
3408 * collisions, but more likely due to the majority of
3409 * arc buffers being unevictable. Therefore, even if
3410 * arc_size is above arc_c, another pass is unlikely to
3411 * be helpful and could potentially cause us to enter an
3414 if (arc_size <= arc_c || evicted == 0) {
3416 * We're either no longer overflowing, or we
3417 * can't evict anything more, so we should wake
3418 * up any threads before we go to sleep.
3420 cv_broadcast(&arc_reclaim_waiters_cv);
3423 * Block until signaled, or after one second (we
3424 * might need to perform arc_kmem_reap_now()
3425 * even if we aren't being signalled)
3427 CALLB_CPR_SAFE_BEGIN(&cpr);
3428 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
3429 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
3430 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
3434 arc_reclaim_thread_exit = FALSE;
3435 cv_broadcast(&arc_reclaim_thread_cv);
3436 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
3441 arc_user_evicts_thread(void)
3445 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG);
3447 mutex_enter(&arc_user_evicts_lock);
3448 while (!arc_user_evicts_thread_exit) {
3449 mutex_exit(&arc_user_evicts_lock);
3451 arc_do_user_evicts();
3454 * This is necessary in order for the mdb ::arc dcmd to
3455 * show up to date information. Since the ::arc command
3456 * does not call the kstat's update function, without
3457 * this call, the command may show stale stats for the
3458 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3459 * with this change, the data might be up to 1 second
3460 * out of date; but that should suffice. The arc_state_t
3461 * structures can be queried directly if more accurate
3462 * information is needed.
3464 if (arc_ksp != NULL)
3465 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
3467 mutex_enter(&arc_user_evicts_lock);
3470 * Block until signaled, or after one second (we need to
3471 * call the arc's kstat update function regularly).
3473 CALLB_CPR_SAFE_BEGIN(&cpr);
3474 (void) cv_timedwait(&arc_user_evicts_cv,
3475 &arc_user_evicts_lock, ddi_get_lbolt() + hz);
3476 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock);
3479 arc_user_evicts_thread_exit = FALSE;
3480 cv_broadcast(&arc_user_evicts_cv);
3481 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */
3486 * Adapt arc info given the number of bytes we are trying to add and
3487 * the state that we are comming from. This function is only called
3488 * when we are adding new content to the cache.
3491 arc_adapt(int bytes, arc_state_t *state)
3494 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
3495 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
3496 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
3498 if (state == arc_l2c_only)
3503 * Adapt the target size of the MRU list:
3504 * - if we just hit in the MRU ghost list, then increase
3505 * the target size of the MRU list.
3506 * - if we just hit in the MFU ghost list, then increase
3507 * the target size of the MFU list by decreasing the
3508 * target size of the MRU list.
3510 if (state == arc_mru_ghost) {
3511 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
3512 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
3514 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
3515 } else if (state == arc_mfu_ghost) {
3518 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
3519 mult = MIN(mult, 10);
3521 delta = MIN(bytes * mult, arc_p);
3522 arc_p = MAX(arc_p_min, arc_p - delta);
3524 ASSERT((int64_t)arc_p >= 0);
3526 if (arc_reclaim_needed()) {
3527 cv_signal(&arc_reclaim_thread_cv);
3534 if (arc_c >= arc_c_max)
3538 * If we're within (2 * maxblocksize) bytes of the target
3539 * cache size, increment the target cache size
3541 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
3542 atomic_add_64(&arc_c, (int64_t)bytes);
3543 if (arc_c > arc_c_max)
3545 else if (state == arc_anon)
3546 atomic_add_64(&arc_p, (int64_t)bytes);
3550 ASSERT((int64_t)arc_p >= 0);
3554 * Check if arc_size has grown past our upper threshold, determined by
3555 * zfs_arc_overflow_shift.
3558 arc_is_overflowing(void)
3560 /* Always allow at least one block of overflow */
3561 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
3562 arc_c >> zfs_arc_overflow_shift);
3564 return (arc_size >= arc_c + overflow);
3568 * The buffer, supplied as the first argument, needs a data block. If we
3569 * are hitting the hard limit for the cache size, we must sleep, waiting
3570 * for the eviction thread to catch up. If we're past the target size
3571 * but below the hard limit, we'll only signal the reclaim thread and
3575 arc_get_data_buf(arc_buf_t *buf)
3577 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
3578 uint64_t size = buf->b_hdr->b_size;
3579 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
3581 arc_adapt(size, state);
3584 * If arc_size is currently overflowing, and has grown past our
3585 * upper limit, we must be adding data faster than the evict
3586 * thread can evict. Thus, to ensure we don't compound the
3587 * problem by adding more data and forcing arc_size to grow even
3588 * further past it's target size, we halt and wait for the
3589 * eviction thread to catch up.
3591 * It's also possible that the reclaim thread is unable to evict
3592 * enough buffers to get arc_size below the overflow limit (e.g.
3593 * due to buffers being un-evictable, or hash lock collisions).
3594 * In this case, we want to proceed regardless if we're
3595 * overflowing; thus we don't use a while loop here.
3597 if (arc_is_overflowing()) {
3598 mutex_enter(&arc_reclaim_lock);
3601 * Now that we've acquired the lock, we may no longer be
3602 * over the overflow limit, lets check.
3604 * We're ignoring the case of spurious wake ups. If that
3605 * were to happen, it'd let this thread consume an ARC
3606 * buffer before it should have (i.e. before we're under
3607 * the overflow limit and were signalled by the reclaim
3608 * thread). As long as that is a rare occurrence, it
3609 * shouldn't cause any harm.
3611 if (arc_is_overflowing()) {
3612 cv_signal(&arc_reclaim_thread_cv);
3613 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
3616 mutex_exit(&arc_reclaim_lock);
3619 if (type == ARC_BUFC_METADATA) {
3620 buf->b_data = zio_buf_alloc(size);
3621 arc_space_consume(size, ARC_SPACE_META);
3623 ASSERT(type == ARC_BUFC_DATA);
3624 buf->b_data = zio_data_buf_alloc(size);
3625 arc_space_consume(size, ARC_SPACE_DATA);
3629 * Update the state size. Note that ghost states have a
3630 * "ghost size" and so don't need to be updated.
3632 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) {
3633 arc_buf_hdr_t *hdr = buf->b_hdr;
3634 arc_state_t *state = hdr->b_l1hdr.b_state;
3636 (void) refcount_add_many(&state->arcs_size, size, buf);
3639 * If this is reached via arc_read, the link is
3640 * protected by the hash lock. If reached via
3641 * arc_buf_alloc, the header should not be accessed by
3642 * any other thread. And, if reached via arc_read_done,
3643 * the hash lock will protect it if it's found in the
3644 * hash table; otherwise no other thread should be
3645 * trying to [add|remove]_reference it.
3647 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3648 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3649 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type],
3653 * If we are growing the cache, and we are adding anonymous
3654 * data, and we have outgrown arc_p, update arc_p
3656 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
3657 (refcount_count(&arc_anon->arcs_size) +
3658 refcount_count(&arc_mru->arcs_size) > arc_p))
3659 arc_p = MIN(arc_c, arc_p + size);
3664 * This routine is called whenever a buffer is accessed.
3665 * NOTE: the hash lock is dropped in this function.
3668 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3672 ASSERT(MUTEX_HELD(hash_lock));
3673 ASSERT(HDR_HAS_L1HDR(hdr));
3675 if (hdr->b_l1hdr.b_state == arc_anon) {
3677 * This buffer is not in the cache, and does not
3678 * appear in our "ghost" list. Add the new buffer
3682 ASSERT0(hdr->b_l1hdr.b_arc_access);
3683 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3684 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3685 arc_change_state(arc_mru, hdr, hash_lock);
3687 } else if (hdr->b_l1hdr.b_state == arc_mru) {
3688 now = ddi_get_lbolt();
3691 * If this buffer is here because of a prefetch, then either:
3692 * - clear the flag if this is a "referencing" read
3693 * (any subsequent access will bump this into the MFU state).
3695 * - move the buffer to the head of the list if this is
3696 * another prefetch (to make it less likely to be evicted).
3698 if (HDR_PREFETCH(hdr)) {
3699 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
3700 /* link protected by hash lock */
3701 ASSERT(multilist_link_active(
3702 &hdr->b_l1hdr.b_arc_node));
3704 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3705 ARCSTAT_BUMP(arcstat_mru_hits);
3707 hdr->b_l1hdr.b_arc_access = now;
3712 * This buffer has been "accessed" only once so far,
3713 * but it is still in the cache. Move it to the MFU
3716 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
3718 * More than 125ms have passed since we
3719 * instantiated this buffer. Move it to the
3720 * most frequently used state.
3722 hdr->b_l1hdr.b_arc_access = now;
3723 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3724 arc_change_state(arc_mfu, hdr, hash_lock);
3726 ARCSTAT_BUMP(arcstat_mru_hits);
3727 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
3728 arc_state_t *new_state;
3730 * This buffer has been "accessed" recently, but
3731 * was evicted from the cache. Move it to the
3735 if (HDR_PREFETCH(hdr)) {
3736 new_state = arc_mru;
3737 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
3738 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3739 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3741 new_state = arc_mfu;
3742 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3745 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3746 arc_change_state(new_state, hdr, hash_lock);
3748 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
3749 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
3751 * This buffer has been accessed more than once and is
3752 * still in the cache. Keep it in the MFU state.
3754 * NOTE: an add_reference() that occurred when we did
3755 * the arc_read() will have kicked this off the list.
3756 * If it was a prefetch, we will explicitly move it to
3757 * the head of the list now.
3759 if ((HDR_PREFETCH(hdr)) != 0) {
3760 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3761 /* link protected by hash_lock */
3762 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3764 ARCSTAT_BUMP(arcstat_mfu_hits);
3765 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3766 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
3767 arc_state_t *new_state = arc_mfu;
3769 * This buffer has been accessed more than once but has
3770 * been evicted from the cache. Move it back to the
3774 if (HDR_PREFETCH(hdr)) {
3776 * This is a prefetch access...
3777 * move this block back to the MRU state.
3779 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3780 new_state = arc_mru;
3783 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3784 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3785 arc_change_state(new_state, hdr, hash_lock);
3787 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
3788 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
3790 * This buffer is on the 2nd Level ARC.
3793 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3794 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3795 arc_change_state(arc_mfu, hdr, hash_lock);
3797 ASSERT(!"invalid arc state");
3801 /* a generic arc_done_func_t which you can use */
3804 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
3806 if (zio == NULL || zio->io_error == 0)
3807 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
3808 VERIFY(arc_buf_remove_ref(buf, arg));
3811 /* a generic arc_done_func_t */
3813 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
3815 arc_buf_t **bufp = arg;
3816 if (zio && zio->io_error) {
3817 VERIFY(arc_buf_remove_ref(buf, arg));
3821 ASSERT(buf->b_data);
3826 arc_read_done(zio_t *zio)
3830 arc_buf_t *abuf; /* buffer we're assigning to callback */
3831 kmutex_t *hash_lock = NULL;
3832 arc_callback_t *callback_list, *acb;
3833 int freeable = FALSE;
3835 buf = zio->io_private;
3839 * The hdr was inserted into hash-table and removed from lists
3840 * prior to starting I/O. We should find this header, since
3841 * it's in the hash table, and it should be legit since it's
3842 * not possible to evict it during the I/O. The only possible
3843 * reason for it not to be found is if we were freed during the
3846 if (HDR_IN_HASH_TABLE(hdr)) {
3847 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
3848 ASSERT3U(hdr->b_dva.dva_word[0], ==,
3849 BP_IDENTITY(zio->io_bp)->dva_word[0]);
3850 ASSERT3U(hdr->b_dva.dva_word[1], ==,
3851 BP_IDENTITY(zio->io_bp)->dva_word[1]);
3853 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
3856 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
3857 hash_lock == NULL) ||
3859 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
3860 (found == hdr && HDR_L2_READING(hdr)));
3863 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED;
3864 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
3865 hdr->b_flags &= ~ARC_FLAG_L2CACHE;
3867 /* byteswap if necessary */
3868 callback_list = hdr->b_l1hdr.b_acb;
3869 ASSERT(callback_list != NULL);
3870 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
3871 dmu_object_byteswap_t bswap =
3872 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
3873 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
3874 byteswap_uint64_array :
3875 dmu_ot_byteswap[bswap].ob_func;
3876 func(buf->b_data, hdr->b_size);
3879 arc_cksum_compute(buf, B_FALSE);
3882 if (hash_lock && zio->io_error == 0 &&
3883 hdr->b_l1hdr.b_state == arc_anon) {
3885 * Only call arc_access on anonymous buffers. This is because
3886 * if we've issued an I/O for an evicted buffer, we've already
3887 * called arc_access (to prevent any simultaneous readers from
3888 * getting confused).
3890 arc_access(hdr, hash_lock);
3893 /* create copies of the data buffer for the callers */
3895 for (acb = callback_list; acb; acb = acb->acb_next) {
3896 if (acb->acb_done) {
3898 ARCSTAT_BUMP(arcstat_duplicate_reads);
3899 abuf = arc_buf_clone(buf);
3901 acb->acb_buf = abuf;
3905 hdr->b_l1hdr.b_acb = NULL;
3906 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
3907 ASSERT(!HDR_BUF_AVAILABLE(hdr));
3909 ASSERT(buf->b_efunc == NULL);
3910 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
3911 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
3914 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
3915 callback_list != NULL);
3917 if (zio->io_error != 0) {
3918 hdr->b_flags |= ARC_FLAG_IO_ERROR;
3919 if (hdr->b_l1hdr.b_state != arc_anon)
3920 arc_change_state(arc_anon, hdr, hash_lock);
3921 if (HDR_IN_HASH_TABLE(hdr))
3922 buf_hash_remove(hdr);
3923 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3927 * Broadcast before we drop the hash_lock to avoid the possibility
3928 * that the hdr (and hence the cv) might be freed before we get to
3929 * the cv_broadcast().
3931 cv_broadcast(&hdr->b_l1hdr.b_cv);
3933 if (hash_lock != NULL) {
3934 mutex_exit(hash_lock);
3937 * This block was freed while we waited for the read to
3938 * complete. It has been removed from the hash table and
3939 * moved to the anonymous state (so that it won't show up
3942 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3943 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3946 /* execute each callback and free its structure */
3947 while ((acb = callback_list) != NULL) {
3949 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3951 if (acb->acb_zio_dummy != NULL) {
3952 acb->acb_zio_dummy->io_error = zio->io_error;
3953 zio_nowait(acb->acb_zio_dummy);
3956 callback_list = acb->acb_next;
3957 kmem_free(acb, sizeof (arc_callback_t));
3961 arc_hdr_destroy(hdr);
3965 * "Read" the block at the specified DVA (in bp) via the
3966 * cache. If the block is found in the cache, invoke the provided
3967 * callback immediately and return. Note that the `zio' parameter
3968 * in the callback will be NULL in this case, since no IO was
3969 * required. If the block is not in the cache pass the read request
3970 * on to the spa with a substitute callback function, so that the
3971 * requested block will be added to the cache.
3973 * If a read request arrives for a block that has a read in-progress,
3974 * either wait for the in-progress read to complete (and return the
3975 * results); or, if this is a read with a "done" func, add a record
3976 * to the read to invoke the "done" func when the read completes,
3977 * and return; or just return.
3979 * arc_read_done() will invoke all the requested "done" functions
3980 * for readers of this block.
3983 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3984 void *private, zio_priority_t priority, int zio_flags,
3985 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
3987 arc_buf_hdr_t *hdr = NULL;
3988 arc_buf_t *buf = NULL;
3989 kmutex_t *hash_lock = NULL;
3991 uint64_t guid = spa_load_guid(spa);
3993 ASSERT(!BP_IS_EMBEDDED(bp) ||
3994 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
3997 if (!BP_IS_EMBEDDED(bp)) {
3999 * Embedded BP's have no DVA and require no I/O to "read".
4000 * Create an anonymous arc buf to back it.
4002 hdr = buf_hash_find(guid, bp, &hash_lock);
4005 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) {
4007 *arc_flags |= ARC_FLAG_CACHED;
4009 if (HDR_IO_IN_PROGRESS(hdr)) {
4011 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
4012 priority == ZIO_PRIORITY_SYNC_READ) {
4014 * This sync read must wait for an
4015 * in-progress async read (e.g. a predictive
4016 * prefetch). Async reads are queued
4017 * separately at the vdev_queue layer, so
4018 * this is a form of priority inversion.
4019 * Ideally, we would "inherit" the demand
4020 * i/o's priority by moving the i/o from
4021 * the async queue to the synchronous queue,
4022 * but there is currently no mechanism to do
4023 * so. Track this so that we can evaluate
4024 * the magnitude of this potential performance
4027 * Note that if the prefetch i/o is already
4028 * active (has been issued to the device),
4029 * the prefetch improved performance, because
4030 * we issued it sooner than we would have
4031 * without the prefetch.
4033 DTRACE_PROBE1(arc__sync__wait__for__async,
4034 arc_buf_hdr_t *, hdr);
4035 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
4037 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4038 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH;
4041 if (*arc_flags & ARC_FLAG_WAIT) {
4042 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4043 mutex_exit(hash_lock);
4046 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4049 arc_callback_t *acb = NULL;
4051 acb = kmem_zalloc(sizeof (arc_callback_t),
4053 acb->acb_done = done;
4054 acb->acb_private = private;
4056 acb->acb_zio_dummy = zio_null(pio,
4057 spa, NULL, NULL, NULL, zio_flags);
4059 ASSERT(acb->acb_done != NULL);
4060 acb->acb_next = hdr->b_l1hdr.b_acb;
4061 hdr->b_l1hdr.b_acb = acb;
4062 add_reference(hdr, hash_lock, private);
4063 mutex_exit(hash_lock);
4066 mutex_exit(hash_lock);
4070 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4071 hdr->b_l1hdr.b_state == arc_mfu);
4074 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4076 * This is a demand read which does not have to
4077 * wait for i/o because we did a predictive
4078 * prefetch i/o for it, which has completed.
4081 arc__demand__hit__predictive__prefetch,
4082 arc_buf_hdr_t *, hdr);
4084 arcstat_demand_hit_predictive_prefetch);
4085 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH;
4087 add_reference(hdr, hash_lock, private);
4089 * If this block is already in use, create a new
4090 * copy of the data so that we will be guaranteed
4091 * that arc_release() will always succeed.
4093 buf = hdr->b_l1hdr.b_buf;
4095 ASSERT(buf->b_data);
4096 if (HDR_BUF_AVAILABLE(hdr)) {
4097 ASSERT(buf->b_efunc == NULL);
4098 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4100 buf = arc_buf_clone(buf);
4103 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4104 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4105 hdr->b_flags |= ARC_FLAG_PREFETCH;
4107 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4108 arc_access(hdr, hash_lock);
4109 if (*arc_flags & ARC_FLAG_L2CACHE)
4110 hdr->b_flags |= ARC_FLAG_L2CACHE;
4111 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4112 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4113 mutex_exit(hash_lock);
4114 ARCSTAT_BUMP(arcstat_hits);
4115 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4116 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4117 data, metadata, hits);
4120 done(NULL, buf, private);
4122 uint64_t size = BP_GET_LSIZE(bp);
4123 arc_callback_t *acb;
4126 boolean_t devw = B_FALSE;
4127 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
4128 int32_t b_asize = 0;
4131 /* this block is not in the cache */
4132 arc_buf_hdr_t *exists = NULL;
4133 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4134 buf = arc_buf_alloc(spa, size, private, type);
4136 if (!BP_IS_EMBEDDED(bp)) {
4137 hdr->b_dva = *BP_IDENTITY(bp);
4138 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4139 exists = buf_hash_insert(hdr, &hash_lock);
4141 if (exists != NULL) {
4142 /* somebody beat us to the hash insert */
4143 mutex_exit(hash_lock);
4144 buf_discard_identity(hdr);
4145 (void) arc_buf_remove_ref(buf, private);
4146 goto top; /* restart the IO request */
4150 * If there is a callback, we pass our reference to
4151 * it; otherwise we remove our reference.
4154 (void) remove_reference(hdr, hash_lock,
4157 if (*arc_flags & ARC_FLAG_PREFETCH)
4158 hdr->b_flags |= ARC_FLAG_PREFETCH;
4159 if (*arc_flags & ARC_FLAG_L2CACHE)
4160 hdr->b_flags |= ARC_FLAG_L2CACHE;
4161 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4162 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4163 if (BP_GET_LEVEL(bp) > 0)
4164 hdr->b_flags |= ARC_FLAG_INDIRECT;
4167 * This block is in the ghost cache. If it was L2-only
4168 * (and thus didn't have an L1 hdr), we realloc the
4169 * header to add an L1 hdr.
4171 if (!HDR_HAS_L1HDR(hdr)) {
4172 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4176 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4177 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4178 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4179 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4182 * If there is a callback, we pass a reference to it.
4185 add_reference(hdr, hash_lock, private);
4186 if (*arc_flags & ARC_FLAG_PREFETCH)
4187 hdr->b_flags |= ARC_FLAG_PREFETCH;
4188 if (*arc_flags & ARC_FLAG_L2CACHE)
4189 hdr->b_flags |= ARC_FLAG_L2CACHE;
4190 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4191 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4192 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
4195 buf->b_efunc = NULL;
4196 buf->b_private = NULL;
4198 hdr->b_l1hdr.b_buf = buf;
4199 ASSERT0(hdr->b_l1hdr.b_datacnt);
4200 hdr->b_l1hdr.b_datacnt = 1;
4201 arc_get_data_buf(buf);
4202 arc_access(hdr, hash_lock);
4205 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
4206 hdr->b_flags |= ARC_FLAG_PREDICTIVE_PREFETCH;
4207 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4209 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4210 acb->acb_done = done;
4211 acb->acb_private = private;
4213 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4214 hdr->b_l1hdr.b_acb = acb;
4215 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4217 if (HDR_HAS_L2HDR(hdr) &&
4218 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
4219 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
4220 addr = hdr->b_l2hdr.b_daddr;
4221 b_compress = hdr->b_l2hdr.b_compress;
4222 b_asize = hdr->b_l2hdr.b_asize;
4224 * Lock out device removal.
4226 if (vdev_is_dead(vd) ||
4227 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
4231 if (hash_lock != NULL)
4232 mutex_exit(hash_lock);
4235 * At this point, we have a level 1 cache miss. Try again in
4236 * L2ARC if possible.
4238 ASSERT3U(hdr->b_size, ==, size);
4239 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
4240 uint64_t, size, zbookmark_phys_t *, zb);
4241 ARCSTAT_BUMP(arcstat_misses);
4242 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4243 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4244 data, metadata, misses);
4246 if (priority == ZIO_PRIORITY_ASYNC_READ)
4247 hdr->b_flags |= ARC_FLAG_PRIO_ASYNC_READ;
4249 hdr->b_flags &= ~ARC_FLAG_PRIO_ASYNC_READ;
4251 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
4253 * Read from the L2ARC if the following are true:
4254 * 1. The L2ARC vdev was previously cached.
4255 * 2. This buffer still has L2ARC metadata.
4256 * 3. This buffer isn't currently writing to the L2ARC.
4257 * 4. The L2ARC entry wasn't evicted, which may
4258 * also have invalidated the vdev.
4259 * 5. This isn't prefetch and l2arc_noprefetch is set.
4261 if (HDR_HAS_L2HDR(hdr) &&
4262 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
4263 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
4264 l2arc_read_callback_t *cb;
4266 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
4267 ARCSTAT_BUMP(arcstat_l2_hits);
4269 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
4271 cb->l2rcb_buf = buf;
4272 cb->l2rcb_spa = spa;
4275 cb->l2rcb_flags = zio_flags;
4276 cb->l2rcb_compress = b_compress;
4278 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
4279 addr + size < vd->vdev_psize -
4280 VDEV_LABEL_END_SIZE);
4283 * l2arc read. The SCL_L2ARC lock will be
4284 * released by l2arc_read_done().
4285 * Issue a null zio if the underlying buffer
4286 * was squashed to zero size by compression.
4288 if (b_compress == ZIO_COMPRESS_EMPTY) {
4289 rzio = zio_null(pio, spa, vd,
4290 l2arc_read_done, cb,
4291 zio_flags | ZIO_FLAG_DONT_CACHE |
4293 ZIO_FLAG_DONT_PROPAGATE |
4294 ZIO_FLAG_DONT_RETRY);
4296 rzio = zio_read_phys(pio, vd, addr,
4297 b_asize, buf->b_data,
4299 l2arc_read_done, cb, priority,
4300 zio_flags | ZIO_FLAG_DONT_CACHE |
4302 ZIO_FLAG_DONT_PROPAGATE |
4303 ZIO_FLAG_DONT_RETRY, B_FALSE);
4305 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
4307 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
4309 if (*arc_flags & ARC_FLAG_NOWAIT) {
4314 ASSERT(*arc_flags & ARC_FLAG_WAIT);
4315 if (zio_wait(rzio) == 0)
4318 /* l2arc read error; goto zio_read() */
4320 DTRACE_PROBE1(l2arc__miss,
4321 arc_buf_hdr_t *, hdr);
4322 ARCSTAT_BUMP(arcstat_l2_misses);
4323 if (HDR_L2_WRITING(hdr))
4324 ARCSTAT_BUMP(arcstat_l2_rw_clash);
4325 spa_config_exit(spa, SCL_L2ARC, vd);
4329 spa_config_exit(spa, SCL_L2ARC, vd);
4330 if (l2arc_ndev != 0) {
4331 DTRACE_PROBE1(l2arc__miss,
4332 arc_buf_hdr_t *, hdr);
4333 ARCSTAT_BUMP(arcstat_l2_misses);
4337 rzio = zio_read(pio, spa, bp, buf->b_data, size,
4338 arc_read_done, buf, priority, zio_flags, zb);
4340 if (*arc_flags & ARC_FLAG_WAIT)
4341 return (zio_wait(rzio));
4343 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4350 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
4352 ASSERT(buf->b_hdr != NULL);
4353 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon);
4354 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) ||
4356 ASSERT(buf->b_efunc == NULL);
4357 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
4359 buf->b_efunc = func;
4360 buf->b_private = private;
4364 * Notify the arc that a block was freed, and thus will never be used again.
4367 arc_freed(spa_t *spa, const blkptr_t *bp)
4370 kmutex_t *hash_lock;
4371 uint64_t guid = spa_load_guid(spa);
4373 ASSERT(!BP_IS_EMBEDDED(bp));
4375 hdr = buf_hash_find(guid, bp, &hash_lock);
4378 if (HDR_BUF_AVAILABLE(hdr)) {
4379 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
4380 add_reference(hdr, hash_lock, FTAG);
4381 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4382 mutex_exit(hash_lock);
4384 arc_release(buf, FTAG);
4385 (void) arc_buf_remove_ref(buf, FTAG);
4387 mutex_exit(hash_lock);
4393 * Clear the user eviction callback set by arc_set_callback(), first calling
4394 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4395 * clearing the callback may result in the arc_buf being destroyed. However,
4396 * it will not result in the *last* arc_buf being destroyed, hence the data
4397 * will remain cached in the ARC. We make a copy of the arc buffer here so
4398 * that we can process the callback without holding any locks.
4400 * It's possible that the callback is already in the process of being cleared
4401 * by another thread. In this case we can not clear the callback.
4403 * Returns B_TRUE if the callback was successfully called and cleared.
4406 arc_clear_callback(arc_buf_t *buf)
4409 kmutex_t *hash_lock;
4410 arc_evict_func_t *efunc = buf->b_efunc;
4411 void *private = buf->b_private;
4413 mutex_enter(&buf->b_evict_lock);
4417 * We are in arc_do_user_evicts().
4419 ASSERT(buf->b_data == NULL);
4420 mutex_exit(&buf->b_evict_lock);
4422 } else if (buf->b_data == NULL) {
4424 * We are on the eviction list; process this buffer now
4425 * but let arc_do_user_evicts() do the reaping.
4427 buf->b_efunc = NULL;
4428 mutex_exit(&buf->b_evict_lock);
4429 VERIFY0(efunc(private));
4432 hash_lock = HDR_LOCK(hdr);
4433 mutex_enter(hash_lock);
4435 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4437 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <,
4438 hdr->b_l1hdr.b_datacnt);
4439 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4440 hdr->b_l1hdr.b_state == arc_mfu);
4442 buf->b_efunc = NULL;
4443 buf->b_private = NULL;
4445 if (hdr->b_l1hdr.b_datacnt > 1) {
4446 mutex_exit(&buf->b_evict_lock);
4447 arc_buf_destroy(buf, TRUE);
4449 ASSERT(buf == hdr->b_l1hdr.b_buf);
4450 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4451 mutex_exit(&buf->b_evict_lock);
4454 mutex_exit(hash_lock);
4455 VERIFY0(efunc(private));
4460 * Release this buffer from the cache, making it an anonymous buffer. This
4461 * must be done after a read and prior to modifying the buffer contents.
4462 * If the buffer has more than one reference, we must make
4463 * a new hdr for the buffer.
4466 arc_release(arc_buf_t *buf, void *tag)
4468 arc_buf_hdr_t *hdr = buf->b_hdr;
4471 * It would be nice to assert that if it's DMU metadata (level >
4472 * 0 || it's the dnode file), then it must be syncing context.
4473 * But we don't know that information at this level.
4476 mutex_enter(&buf->b_evict_lock);
4478 ASSERT(HDR_HAS_L1HDR(hdr));
4481 * We don't grab the hash lock prior to this check, because if
4482 * the buffer's header is in the arc_anon state, it won't be
4483 * linked into the hash table.
4485 if (hdr->b_l1hdr.b_state == arc_anon) {
4486 mutex_exit(&buf->b_evict_lock);
4487 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4488 ASSERT(!HDR_IN_HASH_TABLE(hdr));
4489 ASSERT(!HDR_HAS_L2HDR(hdr));
4490 ASSERT(BUF_EMPTY(hdr));
4492 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1);
4493 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
4494 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
4496 ASSERT3P(buf->b_efunc, ==, NULL);
4497 ASSERT3P(buf->b_private, ==, NULL);
4499 hdr->b_l1hdr.b_arc_access = 0;
4505 kmutex_t *hash_lock = HDR_LOCK(hdr);
4506 mutex_enter(hash_lock);
4509 * This assignment is only valid as long as the hash_lock is
4510 * held, we must be careful not to reference state or the
4511 * b_state field after dropping the lock.
4513 arc_state_t *state = hdr->b_l1hdr.b_state;
4514 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4515 ASSERT3P(state, !=, arc_anon);
4517 /* this buffer is not on any list */
4518 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
4520 if (HDR_HAS_L2HDR(hdr)) {
4521 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4524 * We have to recheck this conditional again now that
4525 * we're holding the l2ad_mtx to prevent a race with
4526 * another thread which might be concurrently calling
4527 * l2arc_evict(). In that case, l2arc_evict() might have
4528 * destroyed the header's L2 portion as we were waiting
4529 * to acquire the l2ad_mtx.
4531 if (HDR_HAS_L2HDR(hdr))
4532 arc_hdr_l2hdr_destroy(hdr);
4534 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4538 * Do we have more than one buf?
4540 if (hdr->b_l1hdr.b_datacnt > 1) {
4541 arc_buf_hdr_t *nhdr;
4543 uint64_t blksz = hdr->b_size;
4544 uint64_t spa = hdr->b_spa;
4545 arc_buf_contents_t type = arc_buf_type(hdr);
4546 uint32_t flags = hdr->b_flags;
4548 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
4550 * Pull the data off of this hdr and attach it to
4551 * a new anonymous hdr.
4553 (void) remove_reference(hdr, hash_lock, tag);
4554 bufp = &hdr->b_l1hdr.b_buf;
4555 while (*bufp != buf)
4556 bufp = &(*bufp)->b_next;
4557 *bufp = buf->b_next;
4560 ASSERT3P(state, !=, arc_l2c_only);
4562 (void) refcount_remove_many(
4563 &state->arcs_size, hdr->b_size, buf);
4565 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
4566 ASSERT3P(state, !=, arc_l2c_only);
4567 uint64_t *size = &state->arcs_lsize[type];
4568 ASSERT3U(*size, >=, hdr->b_size);
4569 atomic_add_64(size, -hdr->b_size);
4573 * We're releasing a duplicate user data buffer, update
4574 * our statistics accordingly.
4576 if (HDR_ISTYPE_DATA(hdr)) {
4577 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
4578 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
4581 hdr->b_l1hdr.b_datacnt -= 1;
4582 arc_cksum_verify(buf);
4583 arc_buf_unwatch(buf);
4585 mutex_exit(hash_lock);
4587 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
4588 nhdr->b_size = blksz;
4591 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING;
4592 nhdr->b_flags |= arc_bufc_to_flags(type);
4593 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
4595 nhdr->b_l1hdr.b_buf = buf;
4596 nhdr->b_l1hdr.b_datacnt = 1;
4597 nhdr->b_l1hdr.b_state = arc_anon;
4598 nhdr->b_l1hdr.b_arc_access = 0;
4599 nhdr->b_l1hdr.b_tmp_cdata = NULL;
4600 nhdr->b_freeze_cksum = NULL;
4602 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
4604 mutex_exit(&buf->b_evict_lock);
4605 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf);
4607 mutex_exit(&buf->b_evict_lock);
4608 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
4609 /* protected by hash lock, or hdr is on arc_anon */
4610 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4611 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4612 arc_change_state(arc_anon, hdr, hash_lock);
4613 hdr->b_l1hdr.b_arc_access = 0;
4614 mutex_exit(hash_lock);
4616 buf_discard_identity(hdr);
4619 buf->b_efunc = NULL;
4620 buf->b_private = NULL;
4624 arc_released(arc_buf_t *buf)
4628 mutex_enter(&buf->b_evict_lock);
4629 released = (buf->b_data != NULL &&
4630 buf->b_hdr->b_l1hdr.b_state == arc_anon);
4631 mutex_exit(&buf->b_evict_lock);
4637 arc_referenced(arc_buf_t *buf)
4641 mutex_enter(&buf->b_evict_lock);
4642 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
4643 mutex_exit(&buf->b_evict_lock);
4644 return (referenced);
4649 arc_write_ready(zio_t *zio)
4651 arc_write_callback_t *callback = zio->io_private;
4652 arc_buf_t *buf = callback->awcb_buf;
4653 arc_buf_hdr_t *hdr = buf->b_hdr;
4655 ASSERT(HDR_HAS_L1HDR(hdr));
4656 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
4657 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4658 callback->awcb_ready(zio, buf, callback->awcb_private);
4661 * If the IO is already in progress, then this is a re-write
4662 * attempt, so we need to thaw and re-compute the cksum.
4663 * It is the responsibility of the callback to handle the
4664 * accounting for any re-write attempt.
4666 if (HDR_IO_IN_PROGRESS(hdr)) {
4667 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
4668 if (hdr->b_freeze_cksum != NULL) {
4669 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
4670 hdr->b_freeze_cksum = NULL;
4672 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
4674 arc_cksum_compute(buf, B_FALSE);
4675 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4679 arc_write_children_ready(zio_t *zio)
4681 arc_write_callback_t *callback = zio->io_private;
4682 arc_buf_t *buf = callback->awcb_buf;
4684 callback->awcb_children_ready(zio, buf, callback->awcb_private);
4688 * The SPA calls this callback for each physical write that happens on behalf
4689 * of a logical write. See the comment in dbuf_write_physdone() for details.
4692 arc_write_physdone(zio_t *zio)
4694 arc_write_callback_t *cb = zio->io_private;
4695 if (cb->awcb_physdone != NULL)
4696 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
4700 arc_write_done(zio_t *zio)
4702 arc_write_callback_t *callback = zio->io_private;
4703 arc_buf_t *buf = callback->awcb_buf;
4704 arc_buf_hdr_t *hdr = buf->b_hdr;
4706 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4708 if (zio->io_error == 0) {
4709 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
4710 buf_discard_identity(hdr);
4712 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
4713 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
4716 ASSERT(BUF_EMPTY(hdr));
4720 * If the block to be written was all-zero or compressed enough to be
4721 * embedded in the BP, no write was performed so there will be no
4722 * dva/birth/checksum. The buffer must therefore remain anonymous
4725 if (!BUF_EMPTY(hdr)) {
4726 arc_buf_hdr_t *exists;
4727 kmutex_t *hash_lock;
4729 ASSERT(zio->io_error == 0);
4731 arc_cksum_verify(buf);
4733 exists = buf_hash_insert(hdr, &hash_lock);
4734 if (exists != NULL) {
4736 * This can only happen if we overwrite for
4737 * sync-to-convergence, because we remove
4738 * buffers from the hash table when we arc_free().
4740 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
4741 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4742 panic("bad overwrite, hdr=%p exists=%p",
4743 (void *)hdr, (void *)exists);
4744 ASSERT(refcount_is_zero(
4745 &exists->b_l1hdr.b_refcnt));
4746 arc_change_state(arc_anon, exists, hash_lock);
4747 mutex_exit(hash_lock);
4748 arc_hdr_destroy(exists);
4749 exists = buf_hash_insert(hdr, &hash_lock);
4750 ASSERT3P(exists, ==, NULL);
4751 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
4753 ASSERT(zio->io_prop.zp_nopwrite);
4754 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4755 panic("bad nopwrite, hdr=%p exists=%p",
4756 (void *)hdr, (void *)exists);
4759 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
4760 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
4761 ASSERT(BP_GET_DEDUP(zio->io_bp));
4762 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
4765 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4766 /* if it's not anon, we are doing a scrub */
4767 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
4768 arc_access(hdr, hash_lock);
4769 mutex_exit(hash_lock);
4771 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4774 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4775 callback->awcb_done(zio, buf, callback->awcb_private);
4777 kmem_free(callback, sizeof (arc_write_callback_t));
4781 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
4782 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
4783 const zio_prop_t *zp, arc_done_func_t *ready,
4784 arc_done_func_t *children_ready, arc_done_func_t *physdone,
4785 arc_done_func_t *done, void *private, zio_priority_t priority,
4786 int zio_flags, const zbookmark_phys_t *zb)
4788 arc_buf_hdr_t *hdr = buf->b_hdr;
4789 arc_write_callback_t *callback;
4792 ASSERT(ready != NULL);
4793 ASSERT(done != NULL);
4794 ASSERT(!HDR_IO_ERROR(hdr));
4795 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4796 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4797 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4799 hdr->b_flags |= ARC_FLAG_L2CACHE;
4801 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4802 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
4803 callback->awcb_ready = ready;
4804 callback->awcb_children_ready = children_ready;
4805 callback->awcb_physdone = physdone;
4806 callback->awcb_done = done;
4807 callback->awcb_private = private;
4808 callback->awcb_buf = buf;
4810 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
4812 (children_ready != NULL) ? arc_write_children_ready : NULL,
4813 arc_write_physdone, arc_write_done, callback,
4814 priority, zio_flags, zb);
4820 arc_memory_throttle(uint64_t reserve, uint64_t txg)
4823 uint64_t available_memory = ptob(freemem);
4824 static uint64_t page_load = 0;
4825 static uint64_t last_txg = 0;
4829 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
4832 if (freemem > physmem * arc_lotsfree_percent / 100)
4835 if (txg > last_txg) {
4840 * If we are in pageout, we know that memory is already tight,
4841 * the arc is already going to be evicting, so we just want to
4842 * continue to let page writes occur as quickly as possible.
4844 if (curproc == proc_pageout) {
4845 if (page_load > MAX(ptob(minfree), available_memory) / 4)
4846 return (SET_ERROR(ERESTART));
4847 /* Note: reserve is inflated, so we deflate */
4848 page_load += reserve / 8;
4850 } else if (page_load > 0 && arc_reclaim_needed()) {
4851 /* memory is low, delay before restarting */
4852 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
4853 return (SET_ERROR(EAGAIN));
4861 arc_tempreserve_clear(uint64_t reserve)
4863 atomic_add_64(&arc_tempreserve, -reserve);
4864 ASSERT((int64_t)arc_tempreserve >= 0);
4868 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
4873 if (reserve > arc_c/4 && !arc_no_grow)
4874 arc_c = MIN(arc_c_max, reserve * 4);
4875 if (reserve > arc_c)
4876 return (SET_ERROR(ENOMEM));
4879 * Don't count loaned bufs as in flight dirty data to prevent long
4880 * network delays from blocking transactions that are ready to be
4881 * assigned to a txg.
4883 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
4884 arc_loaned_bytes), 0);
4887 * Writes will, almost always, require additional memory allocations
4888 * in order to compress/encrypt/etc the data. We therefore need to
4889 * make sure that there is sufficient available memory for this.
4891 error = arc_memory_throttle(reserve, txg);
4896 * Throttle writes when the amount of dirty data in the cache
4897 * gets too large. We try to keep the cache less than half full
4898 * of dirty blocks so that our sync times don't grow too large.
4899 * Note: if two requests come in concurrently, we might let them
4900 * both succeed, when one of them should fail. Not a huge deal.
4903 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
4904 anon_size > arc_c / 4) {
4905 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
4906 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
4907 arc_tempreserve>>10,
4908 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
4909 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
4910 reserve>>10, arc_c>>10);
4911 return (SET_ERROR(ERESTART));
4913 atomic_add_64(&arc_tempreserve, reserve);
4918 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
4919 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
4921 size->value.ui64 = refcount_count(&state->arcs_size);
4922 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
4923 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
4927 arc_kstat_update(kstat_t *ksp, int rw)
4929 arc_stats_t *as = ksp->ks_data;
4931 if (rw == KSTAT_WRITE) {
4934 arc_kstat_update_state(arc_anon,
4935 &as->arcstat_anon_size,
4936 &as->arcstat_anon_evictable_data,
4937 &as->arcstat_anon_evictable_metadata);
4938 arc_kstat_update_state(arc_mru,
4939 &as->arcstat_mru_size,
4940 &as->arcstat_mru_evictable_data,
4941 &as->arcstat_mru_evictable_metadata);
4942 arc_kstat_update_state(arc_mru_ghost,
4943 &as->arcstat_mru_ghost_size,
4944 &as->arcstat_mru_ghost_evictable_data,
4945 &as->arcstat_mru_ghost_evictable_metadata);
4946 arc_kstat_update_state(arc_mfu,
4947 &as->arcstat_mfu_size,
4948 &as->arcstat_mfu_evictable_data,
4949 &as->arcstat_mfu_evictable_metadata);
4950 arc_kstat_update_state(arc_mfu_ghost,
4951 &as->arcstat_mfu_ghost_size,
4952 &as->arcstat_mfu_ghost_evictable_data,
4953 &as->arcstat_mfu_ghost_evictable_metadata);
4960 * This function *must* return indices evenly distributed between all
4961 * sublists of the multilist. This is needed due to how the ARC eviction
4962 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
4963 * distributed between all sublists and uses this assumption when
4964 * deciding which sublist to evict from and how much to evict from it.
4967 arc_state_multilist_index_func(multilist_t *ml, void *obj)
4969 arc_buf_hdr_t *hdr = obj;
4972 * We rely on b_dva to generate evenly distributed index
4973 * numbers using buf_hash below. So, as an added precaution,
4974 * let's make sure we never add empty buffers to the arc lists.
4976 ASSERT(!BUF_EMPTY(hdr));
4979 * The assumption here, is the hash value for a given
4980 * arc_buf_hdr_t will remain constant throughout it's lifetime
4981 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
4982 * Thus, we don't need to store the header's sublist index
4983 * on insertion, as this index can be recalculated on removal.
4985 * Also, the low order bits of the hash value are thought to be
4986 * distributed evenly. Otherwise, in the case that the multilist
4987 * has a power of two number of sublists, each sublists' usage
4988 * would not be evenly distributed.
4990 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
4991 multilist_get_num_sublists(ml));
4998 * allmem is "all memory that we could possibly use".
5001 uint64_t allmem = ptob(physmem - swapfs_minfree);
5003 uint64_t allmem = (physmem * PAGESIZE) / 2;
5006 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
5007 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
5008 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
5010 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
5011 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL);
5013 /* Convert seconds to clock ticks */
5014 arc_min_prefetch_lifespan = 1 * hz;
5016 /* Start out with 1/8 of all memory */
5021 * On architectures where the physical memory can be larger
5022 * than the addressable space (intel in 32-bit mode), we may
5023 * need to limit the cache to 1/8 of VM size.
5025 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
5028 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
5029 arc_c_min = MAX(allmem / 32, 64 << 20);
5030 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
5031 if (allmem >= 1 << 30)
5032 arc_c_max = allmem - (1 << 30);
5034 arc_c_max = arc_c_min;
5035 arc_c_max = MAX(allmem * 3 / 4, arc_c_max);
5038 * In userland, there's only the memory pressure that we artificially
5039 * create (see arc_available_memory()). Don't let arc_c get too
5040 * small, because it can cause transactions to be larger than
5041 * arc_c, causing arc_tempreserve_space() to fail.
5044 arc_c_min = arc_c_max / 2;
5048 * Allow the tunables to override our calculations if they are
5049 * reasonable (ie. over 64MB)
5051 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem)
5052 arc_c_max = zfs_arc_max;
5053 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
5054 arc_c_min = zfs_arc_min;
5057 arc_p = (arc_c >> 1);
5059 /* limit meta-data to 1/4 of the arc capacity */
5060 arc_meta_limit = arc_c_max / 4;
5062 /* Allow the tunable to override if it is reasonable */
5063 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
5064 arc_meta_limit = zfs_arc_meta_limit;
5066 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
5067 arc_c_min = arc_meta_limit / 2;
5069 if (zfs_arc_meta_min > 0) {
5070 arc_meta_min = zfs_arc_meta_min;
5072 arc_meta_min = arc_c_min / 2;
5075 if (zfs_arc_grow_retry > 0)
5076 arc_grow_retry = zfs_arc_grow_retry;
5078 if (zfs_arc_shrink_shift > 0)
5079 arc_shrink_shift = zfs_arc_shrink_shift;
5082 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
5084 if (arc_no_grow_shift >= arc_shrink_shift)
5085 arc_no_grow_shift = arc_shrink_shift - 1;
5087 if (zfs_arc_p_min_shift > 0)
5088 arc_p_min_shift = zfs_arc_p_min_shift;
5090 if (zfs_arc_num_sublists_per_state < 1)
5091 zfs_arc_num_sublists_per_state = MAX(boot_ncpus, 1);
5093 /* if kmem_flags are set, lets try to use less memory */
5094 if (kmem_debugging())
5096 if (arc_c < arc_c_min)
5099 arc_anon = &ARC_anon;
5101 arc_mru_ghost = &ARC_mru_ghost;
5103 arc_mfu_ghost = &ARC_mfu_ghost;
5104 arc_l2c_only = &ARC_l2c_only;
5107 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5108 sizeof (arc_buf_hdr_t),
5109 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5110 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5111 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5112 sizeof (arc_buf_hdr_t),
5113 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5114 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5115 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5116 sizeof (arc_buf_hdr_t),
5117 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5118 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5119 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5120 sizeof (arc_buf_hdr_t),
5121 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5122 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5123 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5124 sizeof (arc_buf_hdr_t),
5125 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5126 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5127 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5128 sizeof (arc_buf_hdr_t),
5129 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5130 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5131 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5132 sizeof (arc_buf_hdr_t),
5133 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5134 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5135 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5136 sizeof (arc_buf_hdr_t),
5137 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5138 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5139 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5140 sizeof (arc_buf_hdr_t),
5141 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5142 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5143 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
5144 sizeof (arc_buf_hdr_t),
5145 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5146 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5148 refcount_create(&arc_anon->arcs_size);
5149 refcount_create(&arc_mru->arcs_size);
5150 refcount_create(&arc_mru_ghost->arcs_size);
5151 refcount_create(&arc_mfu->arcs_size);
5152 refcount_create(&arc_mfu_ghost->arcs_size);
5153 refcount_create(&arc_l2c_only->arcs_size);
5157 arc_reclaim_thread_exit = FALSE;
5158 arc_user_evicts_thread_exit = FALSE;
5159 arc_eviction_list = NULL;
5160 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
5162 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
5163 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
5165 if (arc_ksp != NULL) {
5166 arc_ksp->ks_data = &arc_stats;
5167 arc_ksp->ks_update = arc_kstat_update;
5168 kstat_install(arc_ksp);
5171 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
5172 TS_RUN, minclsyspri);
5174 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0,
5175 TS_RUN, minclsyspri);
5181 * Calculate maximum amount of dirty data per pool.
5183 * If it has been set by /etc/system, take that.
5184 * Otherwise, use a percentage of physical memory defined by
5185 * zfs_dirty_data_max_percent (default 10%) with a cap at
5186 * zfs_dirty_data_max_max (default 4GB).
5188 if (zfs_dirty_data_max == 0) {
5189 zfs_dirty_data_max = physmem * PAGESIZE *
5190 zfs_dirty_data_max_percent / 100;
5191 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
5192 zfs_dirty_data_max_max);
5199 mutex_enter(&arc_reclaim_lock);
5200 arc_reclaim_thread_exit = TRUE;
5202 * The reclaim thread will set arc_reclaim_thread_exit back to
5203 * FALSE when it is finished exiting; we're waiting for that.
5205 while (arc_reclaim_thread_exit) {
5206 cv_signal(&arc_reclaim_thread_cv);
5207 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
5209 mutex_exit(&arc_reclaim_lock);
5211 mutex_enter(&arc_user_evicts_lock);
5212 arc_user_evicts_thread_exit = TRUE;
5214 * The user evicts thread will set arc_user_evicts_thread_exit
5215 * to FALSE when it is finished exiting; we're waiting for that.
5217 while (arc_user_evicts_thread_exit) {
5218 cv_signal(&arc_user_evicts_cv);
5219 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock);
5221 mutex_exit(&arc_user_evicts_lock);
5223 /* Use TRUE to ensure *all* buffers are evicted */
5224 arc_flush(NULL, TRUE);
5228 if (arc_ksp != NULL) {
5229 kstat_delete(arc_ksp);
5233 mutex_destroy(&arc_reclaim_lock);
5234 cv_destroy(&arc_reclaim_thread_cv);
5235 cv_destroy(&arc_reclaim_waiters_cv);
5237 mutex_destroy(&arc_user_evicts_lock);
5238 cv_destroy(&arc_user_evicts_cv);
5240 refcount_destroy(&arc_anon->arcs_size);
5241 refcount_destroy(&arc_mru->arcs_size);
5242 refcount_destroy(&arc_mru_ghost->arcs_size);
5243 refcount_destroy(&arc_mfu->arcs_size);
5244 refcount_destroy(&arc_mfu_ghost->arcs_size);
5245 refcount_destroy(&arc_l2c_only->arcs_size);
5247 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
5248 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5249 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
5250 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
5251 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
5252 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
5253 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
5254 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
5255 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
5256 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
5260 ASSERT0(arc_loaned_bytes);
5266 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5267 * It uses dedicated storage devices to hold cached data, which are populated
5268 * using large infrequent writes. The main role of this cache is to boost
5269 * the performance of random read workloads. The intended L2ARC devices
5270 * include short-stroked disks, solid state disks, and other media with
5271 * substantially faster read latency than disk.
5273 * +-----------------------+
5275 * +-----------------------+
5278 * l2arc_feed_thread() arc_read()
5282 * +---------------+ |
5284 * +---------------+ |
5289 * +-------+ +-------+
5291 * | cache | | cache |
5292 * +-------+ +-------+
5293 * +=========+ .-----.
5294 * : L2ARC : |-_____-|
5295 * : devices : | Disks |
5296 * +=========+ `-_____-'
5298 * Read requests are satisfied from the following sources, in order:
5301 * 2) vdev cache of L2ARC devices
5303 * 4) vdev cache of disks
5306 * Some L2ARC device types exhibit extremely slow write performance.
5307 * To accommodate for this there are some significant differences between
5308 * the L2ARC and traditional cache design:
5310 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5311 * the ARC behave as usual, freeing buffers and placing headers on ghost
5312 * lists. The ARC does not send buffers to the L2ARC during eviction as
5313 * this would add inflated write latencies for all ARC memory pressure.
5315 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5316 * It does this by periodically scanning buffers from the eviction-end of
5317 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5318 * not already there. It scans until a headroom of buffers is satisfied,
5319 * which itself is a buffer for ARC eviction. If a compressible buffer is
5320 * found during scanning and selected for writing to an L2ARC device, we
5321 * temporarily boost scanning headroom during the next scan cycle to make
5322 * sure we adapt to compression effects (which might significantly reduce
5323 * the data volume we write to L2ARC). The thread that does this is
5324 * l2arc_feed_thread(), illustrated below; example sizes are included to
5325 * provide a better sense of ratio than this diagram:
5328 * +---------------------+----------+
5329 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5330 * +---------------------+----------+ | o L2ARC eligible
5331 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5332 * +---------------------+----------+ |
5333 * 15.9 Gbytes ^ 32 Mbytes |
5335 * l2arc_feed_thread()
5337 * l2arc write hand <--[oooo]--'
5341 * +==============================+
5342 * L2ARC dev |####|#|###|###| |####| ... |
5343 * +==============================+
5346 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5347 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5348 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5349 * safe to say that this is an uncommon case, since buffers at the end of
5350 * the ARC lists have moved there due to inactivity.
5352 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5353 * then the L2ARC simply misses copying some buffers. This serves as a
5354 * pressure valve to prevent heavy read workloads from both stalling the ARC
5355 * with waits and clogging the L2ARC with writes. This also helps prevent
5356 * the potential for the L2ARC to churn if it attempts to cache content too
5357 * quickly, such as during backups of the entire pool.
5359 * 5. After system boot and before the ARC has filled main memory, there are
5360 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5361 * lists can remain mostly static. Instead of searching from tail of these
5362 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5363 * for eligible buffers, greatly increasing its chance of finding them.
5365 * The L2ARC device write speed is also boosted during this time so that
5366 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5367 * there are no L2ARC reads, and no fear of degrading read performance
5368 * through increased writes.
5370 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5371 * the vdev queue can aggregate them into larger and fewer writes. Each
5372 * device is written to in a rotor fashion, sweeping writes through
5373 * available space then repeating.
5375 * 7. The L2ARC does not store dirty content. It never needs to flush
5376 * write buffers back to disk based storage.
5378 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5379 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5381 * The performance of the L2ARC can be tweaked by a number of tunables, which
5382 * may be necessary for different workloads:
5384 * l2arc_write_max max write bytes per interval
5385 * l2arc_write_boost extra write bytes during device warmup
5386 * l2arc_noprefetch skip caching prefetched buffers
5387 * l2arc_headroom number of max device writes to precache
5388 * l2arc_headroom_boost when we find compressed buffers during ARC
5389 * scanning, we multiply headroom by this
5390 * percentage factor for the next scan cycle,
5391 * since more compressed buffers are likely to
5393 * l2arc_feed_secs seconds between L2ARC writing
5395 * Tunables may be removed or added as future performance improvements are
5396 * integrated, and also may become zpool properties.
5398 * There are three key functions that control how the L2ARC warms up:
5400 * l2arc_write_eligible() check if a buffer is eligible to cache
5401 * l2arc_write_size() calculate how much to write
5402 * l2arc_write_interval() calculate sleep delay between writes
5404 * These three functions determine what to write, how much, and how quickly
5409 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
5412 * A buffer is *not* eligible for the L2ARC if it:
5413 * 1. belongs to a different spa.
5414 * 2. is already cached on the L2ARC.
5415 * 3. has an I/O in progress (it may be an incomplete read).
5416 * 4. is flagged not eligible (zfs property).
5418 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
5419 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
5426 l2arc_write_size(void)
5431 * Make sure our globals have meaningful values in case the user
5434 size = l2arc_write_max;
5436 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
5437 "be greater than zero, resetting it to the default (%d)",
5439 size = l2arc_write_max = L2ARC_WRITE_SIZE;
5442 if (arc_warm == B_FALSE)
5443 size += l2arc_write_boost;
5450 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
5452 clock_t interval, next, now;
5455 * If the ARC lists are busy, increase our write rate; if the
5456 * lists are stale, idle back. This is achieved by checking
5457 * how much we previously wrote - if it was more than half of
5458 * what we wanted, schedule the next write much sooner.
5460 if (l2arc_feed_again && wrote > (wanted / 2))
5461 interval = (hz * l2arc_feed_min_ms) / 1000;
5463 interval = hz * l2arc_feed_secs;
5465 now = ddi_get_lbolt();
5466 next = MAX(now, MIN(now + interval, began + interval));
5472 * Cycle through L2ARC devices. This is how L2ARC load balances.
5473 * If a device is returned, this also returns holding the spa config lock.
5475 static l2arc_dev_t *
5476 l2arc_dev_get_next(void)
5478 l2arc_dev_t *first, *next = NULL;
5481 * Lock out the removal of spas (spa_namespace_lock), then removal
5482 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
5483 * both locks will be dropped and a spa config lock held instead.
5485 mutex_enter(&spa_namespace_lock);
5486 mutex_enter(&l2arc_dev_mtx);
5488 /* if there are no vdevs, there is nothing to do */
5489 if (l2arc_ndev == 0)
5493 next = l2arc_dev_last;
5495 /* loop around the list looking for a non-faulted vdev */
5497 next = list_head(l2arc_dev_list);
5499 next = list_next(l2arc_dev_list, next);
5501 next = list_head(l2arc_dev_list);
5504 /* if we have come back to the start, bail out */
5507 else if (next == first)
5510 } while (vdev_is_dead(next->l2ad_vdev));
5512 /* if we were unable to find any usable vdevs, return NULL */
5513 if (vdev_is_dead(next->l2ad_vdev))
5516 l2arc_dev_last = next;
5519 mutex_exit(&l2arc_dev_mtx);
5522 * Grab the config lock to prevent the 'next' device from being
5523 * removed while we are writing to it.
5526 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
5527 mutex_exit(&spa_namespace_lock);
5533 * Free buffers that were tagged for destruction.
5536 l2arc_do_free_on_write()
5539 l2arc_data_free_t *df, *df_prev;
5541 mutex_enter(&l2arc_free_on_write_mtx);
5542 buflist = l2arc_free_on_write;
5544 for (df = list_tail(buflist); df; df = df_prev) {
5545 df_prev = list_prev(buflist, df);
5546 ASSERT(df->l2df_data != NULL);
5547 ASSERT(df->l2df_func != NULL);
5548 df->l2df_func(df->l2df_data, df->l2df_size);
5549 list_remove(buflist, df);
5550 kmem_free(df, sizeof (l2arc_data_free_t));
5553 mutex_exit(&l2arc_free_on_write_mtx);
5557 * A write to a cache device has completed. Update all headers to allow
5558 * reads from these buffers to begin.
5561 l2arc_write_done(zio_t *zio)
5563 l2arc_write_callback_t *cb;
5566 arc_buf_hdr_t *head, *hdr, *hdr_prev;
5567 kmutex_t *hash_lock;
5568 int64_t bytes_dropped = 0;
5570 cb = zio->io_private;
5572 dev = cb->l2wcb_dev;
5573 ASSERT(dev != NULL);
5574 head = cb->l2wcb_head;
5575 ASSERT(head != NULL);
5576 buflist = &dev->l2ad_buflist;
5577 ASSERT(buflist != NULL);
5578 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
5579 l2arc_write_callback_t *, cb);
5581 if (zio->io_error != 0)
5582 ARCSTAT_BUMP(arcstat_l2_writes_error);
5585 * All writes completed, or an error was hit.
5588 mutex_enter(&dev->l2ad_mtx);
5589 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
5590 hdr_prev = list_prev(buflist, hdr);
5592 hash_lock = HDR_LOCK(hdr);
5595 * We cannot use mutex_enter or else we can deadlock
5596 * with l2arc_write_buffers (due to swapping the order
5597 * the hash lock and l2ad_mtx are taken).
5599 if (!mutex_tryenter(hash_lock)) {
5601 * Missed the hash lock. We must retry so we
5602 * don't leave the ARC_FLAG_L2_WRITING bit set.
5604 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
5607 * We don't want to rescan the headers we've
5608 * already marked as having been written out, so
5609 * we reinsert the head node so we can pick up
5610 * where we left off.
5612 list_remove(buflist, head);
5613 list_insert_after(buflist, hdr, head);
5615 mutex_exit(&dev->l2ad_mtx);
5618 * We wait for the hash lock to become available
5619 * to try and prevent busy waiting, and increase
5620 * the chance we'll be able to acquire the lock
5621 * the next time around.
5623 mutex_enter(hash_lock);
5624 mutex_exit(hash_lock);
5629 * We could not have been moved into the arc_l2c_only
5630 * state while in-flight due to our ARC_FLAG_L2_WRITING
5631 * bit being set. Let's just ensure that's being enforced.
5633 ASSERT(HDR_HAS_L1HDR(hdr));
5636 * We may have allocated a buffer for L2ARC compression,
5637 * we must release it to avoid leaking this data.
5639 l2arc_release_cdata_buf(hdr);
5641 if (zio->io_error != 0) {
5643 * Error - drop L2ARC entry.
5645 list_remove(buflist, hdr);
5646 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
5648 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize);
5649 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
5651 bytes_dropped += hdr->b_l2hdr.b_asize;
5652 (void) refcount_remove_many(&dev->l2ad_alloc,
5653 hdr->b_l2hdr.b_asize, hdr);
5657 * Allow ARC to begin reads and ghost list evictions to
5660 hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
5662 mutex_exit(hash_lock);
5665 atomic_inc_64(&l2arc_writes_done);
5666 list_remove(buflist, head);
5667 ASSERT(!HDR_HAS_L1HDR(head));
5668 kmem_cache_free(hdr_l2only_cache, head);
5669 mutex_exit(&dev->l2ad_mtx);
5671 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
5673 l2arc_do_free_on_write();
5675 kmem_free(cb, sizeof (l2arc_write_callback_t));
5679 * A read to a cache device completed. Validate buffer contents before
5680 * handing over to the regular ARC routines.
5683 l2arc_read_done(zio_t *zio)
5685 l2arc_read_callback_t *cb;
5688 kmutex_t *hash_lock;
5691 ASSERT(zio->io_vd != NULL);
5692 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
5694 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
5696 cb = zio->io_private;
5698 buf = cb->l2rcb_buf;
5699 ASSERT(buf != NULL);
5701 hash_lock = HDR_LOCK(buf->b_hdr);
5702 mutex_enter(hash_lock);
5704 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5707 * If the buffer was compressed, decompress it first.
5709 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
5710 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
5711 ASSERT(zio->io_data != NULL);
5712 ASSERT3U(zio->io_size, ==, hdr->b_size);
5713 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size);
5716 * Check this survived the L2ARC journey.
5718 equal = arc_cksum_equal(buf);
5719 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
5720 mutex_exit(hash_lock);
5721 zio->io_private = buf;
5722 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
5723 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
5726 mutex_exit(hash_lock);
5728 * Buffer didn't survive caching. Increment stats and
5729 * reissue to the original storage device.
5731 if (zio->io_error != 0) {
5732 ARCSTAT_BUMP(arcstat_l2_io_error);
5734 zio->io_error = SET_ERROR(EIO);
5737 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
5740 * If there's no waiter, issue an async i/o to the primary
5741 * storage now. If there *is* a waiter, the caller must
5742 * issue the i/o in a context where it's OK to block.
5744 if (zio->io_waiter == NULL) {
5745 zio_t *pio = zio_unique_parent(zio);
5747 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
5749 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
5750 buf->b_data, hdr->b_size, arc_read_done, buf,
5751 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
5755 kmem_free(cb, sizeof (l2arc_read_callback_t));
5759 * This is the list priority from which the L2ARC will search for pages to
5760 * cache. This is used within loops (0..3) to cycle through lists in the
5761 * desired order. This order can have a significant effect on cache
5764 * Currently the metadata lists are hit first, MFU then MRU, followed by
5765 * the data lists. This function returns a locked list, and also returns
5768 static multilist_sublist_t *
5769 l2arc_sublist_lock(int list_num)
5771 multilist_t *ml = NULL;
5774 ASSERT(list_num >= 0 && list_num <= 3);
5778 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
5781 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
5784 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
5787 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
5792 * Return a randomly-selected sublist. This is acceptable
5793 * because the caller feeds only a little bit of data for each
5794 * call (8MB). Subsequent calls will result in different
5795 * sublists being selected.
5797 idx = multilist_get_random_index(ml);
5798 return (multilist_sublist_lock(ml, idx));
5802 * Evict buffers from the device write hand to the distance specified in
5803 * bytes. This distance may span populated buffers, it may span nothing.
5804 * This is clearing a region on the L2ARC device ready for writing.
5805 * If the 'all' boolean is set, every buffer is evicted.
5808 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
5811 arc_buf_hdr_t *hdr, *hdr_prev;
5812 kmutex_t *hash_lock;
5815 buflist = &dev->l2ad_buflist;
5817 if (!all && dev->l2ad_first) {
5819 * This is the first sweep through the device. There is
5825 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
5827 * When nearing the end of the device, evict to the end
5828 * before the device write hand jumps to the start.
5830 taddr = dev->l2ad_end;
5832 taddr = dev->l2ad_hand + distance;
5834 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
5835 uint64_t, taddr, boolean_t, all);
5838 mutex_enter(&dev->l2ad_mtx);
5839 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
5840 hdr_prev = list_prev(buflist, hdr);
5842 hash_lock = HDR_LOCK(hdr);
5845 * We cannot use mutex_enter or else we can deadlock
5846 * with l2arc_write_buffers (due to swapping the order
5847 * the hash lock and l2ad_mtx are taken).
5849 if (!mutex_tryenter(hash_lock)) {
5851 * Missed the hash lock. Retry.
5853 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
5854 mutex_exit(&dev->l2ad_mtx);
5855 mutex_enter(hash_lock);
5856 mutex_exit(hash_lock);
5860 if (HDR_L2_WRITE_HEAD(hdr)) {
5862 * We hit a write head node. Leave it for
5863 * l2arc_write_done().
5865 list_remove(buflist, hdr);
5866 mutex_exit(hash_lock);
5870 if (!all && HDR_HAS_L2HDR(hdr) &&
5871 (hdr->b_l2hdr.b_daddr > taddr ||
5872 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
5874 * We've evicted to the target address,
5875 * or the end of the device.
5877 mutex_exit(hash_lock);
5881 ASSERT(HDR_HAS_L2HDR(hdr));
5882 if (!HDR_HAS_L1HDR(hdr)) {
5883 ASSERT(!HDR_L2_READING(hdr));
5885 * This doesn't exist in the ARC. Destroy.
5886 * arc_hdr_destroy() will call list_remove()
5887 * and decrement arcstat_l2_size.
5889 arc_change_state(arc_anon, hdr, hash_lock);
5890 arc_hdr_destroy(hdr);
5892 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
5893 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
5895 * Invalidate issued or about to be issued
5896 * reads, since we may be about to write
5897 * over this location.
5899 if (HDR_L2_READING(hdr)) {
5900 ARCSTAT_BUMP(arcstat_l2_evict_reading);
5901 hdr->b_flags |= ARC_FLAG_L2_EVICTED;
5904 /* Ensure this header has finished being written */
5905 ASSERT(!HDR_L2_WRITING(hdr));
5906 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
5908 arc_hdr_l2hdr_destroy(hdr);
5910 mutex_exit(hash_lock);
5912 mutex_exit(&dev->l2ad_mtx);
5916 * Find and write ARC buffers to the L2ARC device.
5918 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
5919 * for reading until they have completed writing.
5920 * The headroom_boost is an in-out parameter used to maintain headroom boost
5921 * state between calls to this function.
5923 * Returns the number of bytes actually written (which may be smaller than
5924 * the delta by which the device hand has changed due to alignment).
5927 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
5928 boolean_t *headroom_boost)
5930 arc_buf_hdr_t *hdr, *hdr_prev, *head;
5931 uint64_t write_asize, write_sz, headroom,
5935 l2arc_write_callback_t *cb;
5937 uint64_t guid = spa_load_guid(spa);
5938 const boolean_t do_headroom_boost = *headroom_boost;
5940 ASSERT(dev->l2ad_vdev != NULL);
5942 /* Lower the flag now, we might want to raise it again later. */
5943 *headroom_boost = B_FALSE;
5946 write_sz = write_asize = 0;
5948 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
5949 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD;
5950 head->b_flags |= ARC_FLAG_HAS_L2HDR;
5953 * We will want to try to compress buffers that are at least 2x the
5954 * device sector size.
5956 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
5959 * Copy buffers for L2ARC writing.
5961 for (int try = 0; try <= 3; try++) {
5962 multilist_sublist_t *mls = l2arc_sublist_lock(try);
5963 uint64_t passed_sz = 0;
5966 * L2ARC fast warmup.
5968 * Until the ARC is warm and starts to evict, read from the
5969 * head of the ARC lists rather than the tail.
5971 if (arc_warm == B_FALSE)
5972 hdr = multilist_sublist_head(mls);
5974 hdr = multilist_sublist_tail(mls);
5976 headroom = target_sz * l2arc_headroom;
5977 if (do_headroom_boost)
5978 headroom = (headroom * l2arc_headroom_boost) / 100;
5980 for (; hdr; hdr = hdr_prev) {
5981 kmutex_t *hash_lock;
5985 if (arc_warm == B_FALSE)
5986 hdr_prev = multilist_sublist_next(mls, hdr);
5988 hdr_prev = multilist_sublist_prev(mls, hdr);
5990 hash_lock = HDR_LOCK(hdr);
5991 if (!mutex_tryenter(hash_lock)) {
5993 * Skip this buffer rather than waiting.
5998 passed_sz += hdr->b_size;
5999 if (passed_sz > headroom) {
6003 mutex_exit(hash_lock);
6007 if (!l2arc_write_eligible(guid, hdr)) {
6008 mutex_exit(hash_lock);
6013 * Assume that the buffer is not going to be compressed
6014 * and could take more space on disk because of a larger
6017 buf_sz = hdr->b_size;
6018 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
6020 if ((write_asize + buf_a_sz) > target_sz) {
6022 mutex_exit(hash_lock);
6028 * Insert a dummy header on the buflist so
6029 * l2arc_write_done() can find where the
6030 * write buffers begin without searching.
6032 mutex_enter(&dev->l2ad_mtx);
6033 list_insert_head(&dev->l2ad_buflist, head);
6034 mutex_exit(&dev->l2ad_mtx);
6037 sizeof (l2arc_write_callback_t), KM_SLEEP);
6038 cb->l2wcb_dev = dev;
6039 cb->l2wcb_head = head;
6040 pio = zio_root(spa, l2arc_write_done, cb,
6045 * Create and add a new L2ARC header.
6047 hdr->b_l2hdr.b_dev = dev;
6048 hdr->b_flags |= ARC_FLAG_L2_WRITING;
6050 * Temporarily stash the data buffer in b_tmp_cdata.
6051 * The subsequent write step will pick it up from
6052 * there. This is because can't access b_l1hdr.b_buf
6053 * without holding the hash_lock, which we in turn
6054 * can't access without holding the ARC list locks
6055 * (which we want to avoid during compression/writing).
6057 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF;
6058 hdr->b_l2hdr.b_asize = hdr->b_size;
6059 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data;
6062 * Explicitly set the b_daddr field to a known
6063 * value which means "invalid address". This
6064 * enables us to differentiate which stage of
6065 * l2arc_write_buffers() the particular header
6066 * is in (e.g. this loop, or the one below).
6067 * ARC_FLAG_L2_WRITING is not enough to make
6068 * this distinction, and we need to know in
6069 * order to do proper l2arc vdev accounting in
6070 * arc_release() and arc_hdr_destroy().
6072 * Note, we can't use a new flag to distinguish
6073 * the two stages because we don't hold the
6074 * header's hash_lock below, in the second stage
6075 * of this function. Thus, we can't simply
6076 * change the b_flags field to denote that the
6077 * IO has been sent. We can change the b_daddr
6078 * field of the L2 portion, though, since we'll
6079 * be holding the l2ad_mtx; which is why we're
6080 * using it to denote the header's state change.
6082 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET;
6084 hdr->b_flags |= ARC_FLAG_HAS_L2HDR;
6086 mutex_enter(&dev->l2ad_mtx);
6087 list_insert_head(&dev->l2ad_buflist, hdr);
6088 mutex_exit(&dev->l2ad_mtx);
6091 * Compute and store the buffer cksum before
6092 * writing. On debug the cksum is verified first.
6094 arc_cksum_verify(hdr->b_l1hdr.b_buf);
6095 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE);
6097 mutex_exit(hash_lock);
6100 write_asize += buf_a_sz;
6103 multilist_sublist_unlock(mls);
6109 /* No buffers selected for writing? */
6112 ASSERT(!HDR_HAS_L1HDR(head));
6113 kmem_cache_free(hdr_l2only_cache, head);
6117 mutex_enter(&dev->l2ad_mtx);
6120 * Note that elsewhere in this file arcstat_l2_asize
6121 * and the used space on l2ad_vdev are updated using b_asize,
6122 * which is not necessarily rounded up to the device block size.
6123 * Too keep accounting consistent we do the same here as well:
6124 * stats_size accumulates the sum of b_asize of the written buffers,
6125 * while write_asize accumulates the sum of b_asize rounded up
6126 * to the device block size.
6127 * The latter sum is used only to validate the corectness of the code.
6129 uint64_t stats_size = 0;
6133 * Now start writing the buffers. We're starting at the write head
6134 * and work backwards, retracing the course of the buffer selector
6137 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr;
6138 hdr = list_prev(&dev->l2ad_buflist, hdr)) {
6142 * We rely on the L1 portion of the header below, so
6143 * it's invalid for this header to have been evicted out
6144 * of the ghost cache, prior to being written out. The
6145 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6147 ASSERT(HDR_HAS_L1HDR(hdr));
6150 * We shouldn't need to lock the buffer here, since we flagged
6151 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6152 * take care to only access its L2 cache parameters. In
6153 * particular, hdr->l1hdr.b_buf may be invalid by now due to
6156 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
6158 if ((HDR_L2COMPRESS(hdr)) &&
6159 hdr->b_l2hdr.b_asize >= buf_compress_minsz) {
6160 if (l2arc_compress_buf(hdr)) {
6162 * If compression succeeded, enable headroom
6163 * boost on the next scan cycle.
6165 *headroom_boost = B_TRUE;
6170 * Pick up the buffer data we had previously stashed away
6171 * (and now potentially also compressed).
6173 buf_data = hdr->b_l1hdr.b_tmp_cdata;
6174 buf_sz = hdr->b_l2hdr.b_asize;
6177 * We need to do this regardless if buf_sz is zero or
6178 * not, otherwise, when this l2hdr is evicted we'll
6179 * remove a reference that was never added.
6181 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr);
6183 /* Compression may have squashed the buffer to zero length. */
6187 wzio = zio_write_phys(pio, dev->l2ad_vdev,
6188 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
6189 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
6190 ZIO_FLAG_CANFAIL, B_FALSE);
6192 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
6194 (void) zio_nowait(wzio);
6196 stats_size += buf_sz;
6199 * Keep the clock hand suitably device-aligned.
6201 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
6202 write_asize += buf_a_sz;
6203 dev->l2ad_hand += buf_a_sz;
6207 mutex_exit(&dev->l2ad_mtx);
6209 ASSERT3U(write_asize, <=, target_sz);
6210 ARCSTAT_BUMP(arcstat_l2_writes_sent);
6211 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
6212 ARCSTAT_INCR(arcstat_l2_size, write_sz);
6213 ARCSTAT_INCR(arcstat_l2_asize, stats_size);
6214 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0);
6217 * Bump device hand to the device start if it is approaching the end.
6218 * l2arc_evict() will already have evicted ahead for this case.
6220 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
6221 dev->l2ad_hand = dev->l2ad_start;
6222 dev->l2ad_first = B_FALSE;
6225 dev->l2ad_writing = B_TRUE;
6226 (void) zio_wait(pio);
6227 dev->l2ad_writing = B_FALSE;
6229 return (write_asize);
6233 * Compresses an L2ARC buffer.
6234 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6235 * size in l2hdr->b_asize. This routine tries to compress the data and
6236 * depending on the compression result there are three possible outcomes:
6237 * *) The buffer was incompressible. The original l2hdr contents were left
6238 * untouched and are ready for writing to an L2 device.
6239 * *) The buffer was all-zeros, so there is no need to write it to an L2
6240 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6241 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6242 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6243 * data buffer which holds the compressed data to be written, and b_asize
6244 * tells us how much data there is. b_compress is set to the appropriate
6245 * compression algorithm. Once writing is done, invoke
6246 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6248 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6249 * buffer was incompressible).
6252 l2arc_compress_buf(arc_buf_hdr_t *hdr)
6255 size_t csize, len, rounded;
6256 ASSERT(HDR_HAS_L2HDR(hdr));
6257 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
6259 ASSERT(HDR_HAS_L1HDR(hdr));
6260 ASSERT3S(l2hdr->b_compress, ==, ZIO_COMPRESS_OFF);
6261 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6263 len = l2hdr->b_asize;
6264 cdata = zio_data_buf_alloc(len);
6265 ASSERT3P(cdata, !=, NULL);
6266 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata,
6267 cdata, l2hdr->b_asize);
6269 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE);
6270 if (rounded > csize) {
6271 bzero((char *)cdata + csize, rounded - csize);
6276 /* zero block, indicate that there's nothing to write */
6277 zio_data_buf_free(cdata, len);
6278 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
6280 hdr->b_l1hdr.b_tmp_cdata = NULL;
6281 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
6283 } else if (csize > 0 && csize < len) {
6285 * Compression succeeded, we'll keep the cdata around for
6286 * writing and release it afterwards.
6288 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
6289 l2hdr->b_asize = csize;
6290 hdr->b_l1hdr.b_tmp_cdata = cdata;
6291 ARCSTAT_BUMP(arcstat_l2_compress_successes);
6295 * Compression failed, release the compressed buffer.
6296 * l2hdr will be left unmodified.
6298 zio_data_buf_free(cdata, len);
6299 ARCSTAT_BUMP(arcstat_l2_compress_failures);
6305 * Decompresses a zio read back from an l2arc device. On success, the
6306 * underlying zio's io_data buffer is overwritten by the uncompressed
6307 * version. On decompression error (corrupt compressed stream), the
6308 * zio->io_error value is set to signal an I/O error.
6310 * Please note that the compressed data stream is not checksummed, so
6311 * if the underlying device is experiencing data corruption, we may feed
6312 * corrupt data to the decompressor, so the decompressor needs to be
6313 * able to handle this situation (LZ4 does).
6316 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
6318 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
6320 if (zio->io_error != 0) {
6322 * An io error has occured, just restore the original io
6323 * size in preparation for a main pool read.
6325 zio->io_orig_size = zio->io_size = hdr->b_size;
6329 if (c == ZIO_COMPRESS_EMPTY) {
6331 * An empty buffer results in a null zio, which means we
6332 * need to fill its io_data after we're done restoring the
6333 * buffer's contents.
6335 ASSERT(hdr->b_l1hdr.b_buf != NULL);
6336 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size);
6337 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data;
6339 ASSERT(zio->io_data != NULL);
6341 * We copy the compressed data from the start of the arc buffer
6342 * (the zio_read will have pulled in only what we need, the
6343 * rest is garbage which we will overwrite at decompression)
6344 * and then decompress back to the ARC data buffer. This way we
6345 * can minimize copying by simply decompressing back over the
6346 * original compressed data (rather than decompressing to an
6347 * aux buffer and then copying back the uncompressed buffer,
6348 * which is likely to be much larger).
6353 csize = zio->io_size;
6354 cdata = zio_data_buf_alloc(csize);
6355 bcopy(zio->io_data, cdata, csize);
6356 if (zio_decompress_data(c, cdata, zio->io_data, csize,
6358 zio->io_error = EIO;
6359 zio_data_buf_free(cdata, csize);
6362 /* Restore the expected uncompressed IO size. */
6363 zio->io_orig_size = zio->io_size = hdr->b_size;
6367 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6368 * This buffer serves as a temporary holder of compressed data while
6369 * the buffer entry is being written to an l2arc device. Once that is
6370 * done, we can dispose of it.
6373 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr)
6375 ASSERT(HDR_HAS_L2HDR(hdr));
6376 enum zio_compress comp = hdr->b_l2hdr.b_compress;
6378 ASSERT(HDR_HAS_L1HDR(hdr));
6379 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp));
6381 if (comp == ZIO_COMPRESS_OFF) {
6383 * In this case, b_tmp_cdata points to the same buffer
6384 * as the arc_buf_t's b_data field. We don't want to
6385 * free it, since the arc_buf_t will handle that.
6387 hdr->b_l1hdr.b_tmp_cdata = NULL;
6388 } else if (comp == ZIO_COMPRESS_EMPTY) {
6390 * In this case, b_tmp_cdata was compressed to an empty
6391 * buffer, thus there's nothing to free and b_tmp_cdata
6392 * should have been set to NULL in l2arc_write_buffers().
6394 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
6397 * If the data was compressed, then we've allocated a
6398 * temporary buffer for it, so now we need to release it.
6400 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6401 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata,
6403 hdr->b_l1hdr.b_tmp_cdata = NULL;
6409 * This thread feeds the L2ARC at regular intervals. This is the beating
6410 * heart of the L2ARC.
6413 l2arc_feed_thread(void)
6418 uint64_t size, wrote;
6419 clock_t begin, next = ddi_get_lbolt();
6420 boolean_t headroom_boost = B_FALSE;
6422 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
6424 mutex_enter(&l2arc_feed_thr_lock);
6426 while (l2arc_thread_exit == 0) {
6427 CALLB_CPR_SAFE_BEGIN(&cpr);
6428 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
6430 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
6431 next = ddi_get_lbolt() + hz;
6434 * Quick check for L2ARC devices.
6436 mutex_enter(&l2arc_dev_mtx);
6437 if (l2arc_ndev == 0) {
6438 mutex_exit(&l2arc_dev_mtx);
6441 mutex_exit(&l2arc_dev_mtx);
6442 begin = ddi_get_lbolt();
6445 * This selects the next l2arc device to write to, and in
6446 * doing so the next spa to feed from: dev->l2ad_spa. This
6447 * will return NULL if there are now no l2arc devices or if
6448 * they are all faulted.
6450 * If a device is returned, its spa's config lock is also
6451 * held to prevent device removal. l2arc_dev_get_next()
6452 * will grab and release l2arc_dev_mtx.
6454 if ((dev = l2arc_dev_get_next()) == NULL)
6457 spa = dev->l2ad_spa;
6458 ASSERT(spa != NULL);
6461 * If the pool is read-only then force the feed thread to
6462 * sleep a little longer.
6464 if (!spa_writeable(spa)) {
6465 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
6466 spa_config_exit(spa, SCL_L2ARC, dev);
6471 * Avoid contributing to memory pressure.
6473 if (arc_reclaim_needed()) {
6474 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
6475 spa_config_exit(spa, SCL_L2ARC, dev);
6479 ARCSTAT_BUMP(arcstat_l2_feeds);
6481 size = l2arc_write_size();
6484 * Evict L2ARC buffers that will be overwritten.
6486 l2arc_evict(dev, size, B_FALSE);
6489 * Write ARC buffers.
6491 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
6494 * Calculate interval between writes.
6496 next = l2arc_write_interval(begin, size, wrote);
6497 spa_config_exit(spa, SCL_L2ARC, dev);
6500 l2arc_thread_exit = 0;
6501 cv_broadcast(&l2arc_feed_thr_cv);
6502 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
6507 l2arc_vdev_present(vdev_t *vd)
6511 mutex_enter(&l2arc_dev_mtx);
6512 for (dev = list_head(l2arc_dev_list); dev != NULL;
6513 dev = list_next(l2arc_dev_list, dev)) {
6514 if (dev->l2ad_vdev == vd)
6517 mutex_exit(&l2arc_dev_mtx);
6519 return (dev != NULL);
6523 * Add a vdev for use by the L2ARC. By this point the spa has already
6524 * validated the vdev and opened it.
6527 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
6529 l2arc_dev_t *adddev;
6531 ASSERT(!l2arc_vdev_present(vd));
6534 * Create a new l2arc device entry.
6536 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
6537 adddev->l2ad_spa = spa;
6538 adddev->l2ad_vdev = vd;
6539 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
6540 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
6541 adddev->l2ad_hand = adddev->l2ad_start;
6542 adddev->l2ad_first = B_TRUE;
6543 adddev->l2ad_writing = B_FALSE;
6545 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
6547 * This is a list of all ARC buffers that are still valid on the
6550 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
6551 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
6553 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
6554 refcount_create(&adddev->l2ad_alloc);
6557 * Add device to global list
6559 mutex_enter(&l2arc_dev_mtx);
6560 list_insert_head(l2arc_dev_list, adddev);
6561 atomic_inc_64(&l2arc_ndev);
6562 mutex_exit(&l2arc_dev_mtx);
6566 * Remove a vdev from the L2ARC.
6569 l2arc_remove_vdev(vdev_t *vd)
6571 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
6574 * Find the device by vdev
6576 mutex_enter(&l2arc_dev_mtx);
6577 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
6578 nextdev = list_next(l2arc_dev_list, dev);
6579 if (vd == dev->l2ad_vdev) {
6584 ASSERT(remdev != NULL);
6587 * Remove device from global list
6589 list_remove(l2arc_dev_list, remdev);
6590 l2arc_dev_last = NULL; /* may have been invalidated */
6591 atomic_dec_64(&l2arc_ndev);
6592 mutex_exit(&l2arc_dev_mtx);
6595 * Clear all buflists and ARC references. L2ARC device flush.
6597 l2arc_evict(remdev, 0, B_TRUE);
6598 list_destroy(&remdev->l2ad_buflist);
6599 mutex_destroy(&remdev->l2ad_mtx);
6600 refcount_destroy(&remdev->l2ad_alloc);
6601 kmem_free(remdev, sizeof (l2arc_dev_t));
6607 l2arc_thread_exit = 0;
6609 l2arc_writes_sent = 0;
6610 l2arc_writes_done = 0;
6612 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
6613 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
6614 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
6615 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
6617 l2arc_dev_list = &L2ARC_dev_list;
6618 l2arc_free_on_write = &L2ARC_free_on_write;
6619 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
6620 offsetof(l2arc_dev_t, l2ad_node));
6621 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
6622 offsetof(l2arc_data_free_t, l2df_list_node));
6629 * This is called from dmu_fini(), which is called from spa_fini();
6630 * Because of this, we can assume that all l2arc devices have
6631 * already been removed when the pools themselves were removed.
6634 l2arc_do_free_on_write();
6636 mutex_destroy(&l2arc_feed_thr_lock);
6637 cv_destroy(&l2arc_feed_thr_cv);
6638 mutex_destroy(&l2arc_dev_mtx);
6639 mutex_destroy(&l2arc_free_on_write_mtx);
6641 list_destroy(l2arc_dev_list);
6642 list_destroy(l2arc_free_on_write);
6648 if (!(spa_mode_global & FWRITE))
6651 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
6652 TS_RUN, minclsyspri);
6658 if (!(spa_mode_global & FWRITE))
6661 mutex_enter(&l2arc_feed_thr_lock);
6662 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
6663 l2arc_thread_exit = 1;
6664 while (l2arc_thread_exit != 0)
6665 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
6666 mutex_exit(&l2arc_feed_thr_lock);