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/dnlc.h>
135 #include <sys/racct.h>
137 #include <sys/callb.h>
138 #include <sys/kstat.h>
139 #include <sys/trim_map.h>
140 #include <zfs_fletcher.h>
143 #include <machine/vmparam.h>
147 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
148 boolean_t arc_watch = B_FALSE;
153 static kmutex_t arc_reclaim_lock;
154 static kcondvar_t arc_reclaim_thread_cv;
155 static boolean_t arc_reclaim_thread_exit;
156 static kcondvar_t arc_reclaim_waiters_cv;
158 static kmutex_t arc_user_evicts_lock;
159 static kcondvar_t arc_user_evicts_cv;
160 static boolean_t arc_user_evicts_thread_exit;
162 static kmutex_t arc_dnlc_evicts_lock;
163 static kcondvar_t arc_dnlc_evicts_cv;
164 static boolean_t arc_dnlc_evicts_thread_exit;
166 uint_t arc_reduce_dnlc_percent = 3;
169 * The number of headers to evict in arc_evict_state_impl() before
170 * dropping the sublist lock and evicting from another sublist. A lower
171 * value means we're more likely to evict the "correct" header (i.e. the
172 * oldest header in the arc state), but comes with higher overhead
173 * (i.e. more invocations of arc_evict_state_impl()).
175 int zfs_arc_evict_batch_limit = 10;
178 * The number of sublists used for each of the arc state lists. If this
179 * is not set to a suitable value by the user, it will be configured to
180 * the number of CPUs on the system in arc_init().
182 int zfs_arc_num_sublists_per_state = 0;
184 /* number of seconds before growing cache again */
185 static int arc_grow_retry = 60;
187 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
188 int zfs_arc_overflow_shift = 8;
190 /* shift of arc_c for calculating both min and max arc_p */
191 static int arc_p_min_shift = 4;
193 /* log2(fraction of arc to reclaim) */
194 static int arc_shrink_shift = 7;
197 * log2(fraction of ARC which must be free to allow growing).
198 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
199 * when reading a new block into the ARC, we will evict an equal-sized block
202 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
203 * we will still not allow it to grow.
205 int arc_no_grow_shift = 5;
209 * minimum lifespan of a prefetch block in clock ticks
210 * (initialized in arc_init())
212 static int arc_min_prefetch_lifespan;
215 * If this percent of memory is free, don't throttle.
217 int arc_lotsfree_percent = 10;
220 extern boolean_t zfs_prefetch_disable;
223 * The arc has filled available memory and has now warmed up.
225 static boolean_t arc_warm;
228 * These tunables are for performance analysis.
230 uint64_t zfs_arc_max;
231 uint64_t zfs_arc_min;
232 uint64_t zfs_arc_meta_limit = 0;
233 uint64_t zfs_arc_meta_min = 0;
234 int zfs_arc_grow_retry = 0;
235 int zfs_arc_shrink_shift = 0;
236 int zfs_arc_p_min_shift = 0;
237 int zfs_disable_dup_eviction = 0;
238 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
239 u_int zfs_arc_free_target = 0;
241 /* Absolute min for arc min / max is 16MB. */
242 static uint64_t arc_abs_min = 16 << 20;
244 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
245 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
246 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
247 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
249 #if defined(__FreeBSD__) && defined(_KERNEL)
251 arc_free_target_init(void *unused __unused)
254 zfs_arc_free_target = vm_pageout_wakeup_thresh;
256 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
257 arc_free_target_init, NULL);
259 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
260 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
261 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
262 SYSCTL_DECL(_vfs_zfs);
263 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
264 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
265 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
266 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
267 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
268 &zfs_arc_average_blocksize, 0,
269 "ARC average blocksize");
270 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
271 &arc_shrink_shift, 0,
272 "log2(fraction of arc to reclaim)");
275 * We don't have a tunable for arc_free_target due to the dependency on
276 * pagedaemon initialisation.
278 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
279 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
280 sysctl_vfs_zfs_arc_free_target, "IU",
281 "Desired number of free pages below which ARC triggers reclaim");
284 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
289 val = zfs_arc_free_target;
290 err = sysctl_handle_int(oidp, &val, 0, req);
291 if (err != 0 || req->newptr == NULL)
296 if (val > vm_cnt.v_page_count)
299 zfs_arc_free_target = val;
305 * Must be declared here, before the definition of corresponding kstat
306 * macro which uses the same names will confuse the compiler.
308 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
309 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
310 sysctl_vfs_zfs_arc_meta_limit, "QU",
311 "ARC metadata limit");
315 * Note that buffers can be in one of 6 states:
316 * ARC_anon - anonymous (discussed below)
317 * ARC_mru - recently used, currently cached
318 * ARC_mru_ghost - recentely used, no longer in cache
319 * ARC_mfu - frequently used, currently cached
320 * ARC_mfu_ghost - frequently used, no longer in cache
321 * ARC_l2c_only - exists in L2ARC but not other states
322 * When there are no active references to the buffer, they are
323 * are linked onto a list in one of these arc states. These are
324 * the only buffers that can be evicted or deleted. Within each
325 * state there are multiple lists, one for meta-data and one for
326 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
327 * etc.) is tracked separately so that it can be managed more
328 * explicitly: favored over data, limited explicitly.
330 * Anonymous buffers are buffers that are not associated with
331 * a DVA. These are buffers that hold dirty block copies
332 * before they are written to stable storage. By definition,
333 * they are "ref'd" and are considered part of arc_mru
334 * that cannot be freed. Generally, they will aquire a DVA
335 * as they are written and migrate onto the arc_mru list.
337 * The ARC_l2c_only state is for buffers that are in the second
338 * level ARC but no longer in any of the ARC_m* lists. The second
339 * level ARC itself may also contain buffers that are in any of
340 * the ARC_m* states - meaning that a buffer can exist in two
341 * places. The reason for the ARC_l2c_only state is to keep the
342 * buffer header in the hash table, so that reads that hit the
343 * second level ARC benefit from these fast lookups.
346 typedef struct arc_state {
348 * list of evictable buffers
350 multilist_t arcs_list[ARC_BUFC_NUMTYPES];
352 * total amount of evictable data in this state
354 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES];
356 * total amount of data in this state; this includes: evictable,
357 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
359 refcount_t arcs_size;
363 static arc_state_t ARC_anon;
364 static arc_state_t ARC_mru;
365 static arc_state_t ARC_mru_ghost;
366 static arc_state_t ARC_mfu;
367 static arc_state_t ARC_mfu_ghost;
368 static arc_state_t ARC_l2c_only;
370 typedef struct arc_stats {
371 kstat_named_t arcstat_hits;
372 kstat_named_t arcstat_misses;
373 kstat_named_t arcstat_demand_data_hits;
374 kstat_named_t arcstat_demand_data_misses;
375 kstat_named_t arcstat_demand_metadata_hits;
376 kstat_named_t arcstat_demand_metadata_misses;
377 kstat_named_t arcstat_prefetch_data_hits;
378 kstat_named_t arcstat_prefetch_data_misses;
379 kstat_named_t arcstat_prefetch_metadata_hits;
380 kstat_named_t arcstat_prefetch_metadata_misses;
381 kstat_named_t arcstat_mru_hits;
382 kstat_named_t arcstat_mru_ghost_hits;
383 kstat_named_t arcstat_mfu_hits;
384 kstat_named_t arcstat_mfu_ghost_hits;
385 kstat_named_t arcstat_allocated;
386 kstat_named_t arcstat_deleted;
388 * Number of buffers that could not be evicted because the hash lock
389 * was held by another thread. The lock may not necessarily be held
390 * by something using the same buffer, since hash locks are shared
391 * by multiple buffers.
393 kstat_named_t arcstat_mutex_miss;
395 * Number of buffers skipped because they have I/O in progress, are
396 * indrect prefetch buffers that have not lived long enough, or are
397 * not from the spa we're trying to evict from.
399 kstat_named_t arcstat_evict_skip;
401 * Number of times arc_evict_state() was unable to evict enough
402 * buffers to reach it's target amount.
404 kstat_named_t arcstat_evict_not_enough;
405 kstat_named_t arcstat_evict_l2_cached;
406 kstat_named_t arcstat_evict_l2_eligible;
407 kstat_named_t arcstat_evict_l2_ineligible;
408 kstat_named_t arcstat_evict_l2_skip;
409 kstat_named_t arcstat_hash_elements;
410 kstat_named_t arcstat_hash_elements_max;
411 kstat_named_t arcstat_hash_collisions;
412 kstat_named_t arcstat_hash_chains;
413 kstat_named_t arcstat_hash_chain_max;
414 kstat_named_t arcstat_p;
415 kstat_named_t arcstat_c;
416 kstat_named_t arcstat_c_min;
417 kstat_named_t arcstat_c_max;
418 kstat_named_t arcstat_size;
420 * Number of bytes consumed by internal ARC structures necessary
421 * for tracking purposes; these structures are not actually
422 * backed by ARC buffers. This includes arc_buf_hdr_t structures
423 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
424 * caches), and arc_buf_t structures (allocated via arc_buf_t
427 kstat_named_t arcstat_hdr_size;
429 * Number of bytes consumed by ARC buffers of type equal to
430 * ARC_BUFC_DATA. This is generally consumed by buffers backing
431 * on disk user data (e.g. plain file contents).
433 kstat_named_t arcstat_data_size;
435 * Number of bytes consumed by ARC buffers of type equal to
436 * ARC_BUFC_METADATA. This is generally consumed by buffers
437 * backing on disk data that is used for internal ZFS
438 * structures (e.g. ZAP, dnode, indirect blocks, etc).
440 kstat_named_t arcstat_metadata_size;
442 * Number of bytes consumed by various buffers and structures
443 * not actually backed with ARC buffers. This includes bonus
444 * buffers (allocated directly via zio_buf_* functions),
445 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
446 * cache), and dnode_t structures (allocated via dnode_t cache).
448 kstat_named_t arcstat_other_size;
450 * Total number of bytes consumed by ARC buffers residing in the
451 * arc_anon state. This includes *all* buffers in the arc_anon
452 * state; e.g. data, metadata, evictable, and unevictable buffers
453 * are all included in this value.
455 kstat_named_t arcstat_anon_size;
457 * Number of bytes consumed by ARC buffers that meet the
458 * following criteria: backing buffers of type ARC_BUFC_DATA,
459 * residing in the arc_anon state, and are eligible for eviction
460 * (e.g. have no outstanding holds on the buffer).
462 kstat_named_t arcstat_anon_evictable_data;
464 * Number of bytes consumed by ARC buffers that meet the
465 * following criteria: backing buffers of type ARC_BUFC_METADATA,
466 * residing in the arc_anon state, and are eligible for eviction
467 * (e.g. have no outstanding holds on the buffer).
469 kstat_named_t arcstat_anon_evictable_metadata;
471 * Total number of bytes consumed by ARC buffers residing in the
472 * arc_mru state. This includes *all* buffers in the arc_mru
473 * state; e.g. data, metadata, evictable, and unevictable buffers
474 * are all included in this value.
476 kstat_named_t arcstat_mru_size;
478 * Number of bytes consumed by ARC buffers that meet the
479 * following criteria: backing buffers of type ARC_BUFC_DATA,
480 * residing in the arc_mru state, and are eligible for eviction
481 * (e.g. have no outstanding holds on the buffer).
483 kstat_named_t arcstat_mru_evictable_data;
485 * Number of bytes consumed by ARC buffers that meet the
486 * following criteria: backing buffers of type ARC_BUFC_METADATA,
487 * residing in the arc_mru state, and are eligible for eviction
488 * (e.g. have no outstanding holds on the buffer).
490 kstat_named_t arcstat_mru_evictable_metadata;
492 * Total number of bytes that *would have been* consumed by ARC
493 * buffers in the arc_mru_ghost state. The key thing to note
494 * here, is the fact that this size doesn't actually indicate
495 * RAM consumption. The ghost lists only consist of headers and
496 * don't actually have ARC buffers linked off of these headers.
497 * Thus, *if* the headers had associated ARC buffers, these
498 * buffers *would have* consumed this number of bytes.
500 kstat_named_t arcstat_mru_ghost_size;
502 * Number of bytes that *would have been* consumed by ARC
503 * buffers that are eligible for eviction, of type
504 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
506 kstat_named_t arcstat_mru_ghost_evictable_data;
508 * Number of bytes that *would have been* consumed by ARC
509 * buffers that are eligible for eviction, of type
510 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
512 kstat_named_t arcstat_mru_ghost_evictable_metadata;
514 * Total number of bytes consumed by ARC buffers residing in the
515 * arc_mfu state. This includes *all* buffers in the arc_mfu
516 * state; e.g. data, metadata, evictable, and unevictable buffers
517 * are all included in this value.
519 kstat_named_t arcstat_mfu_size;
521 * Number of bytes consumed by ARC buffers that are eligible for
522 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
525 kstat_named_t arcstat_mfu_evictable_data;
527 * Number of bytes consumed by ARC buffers that are eligible for
528 * eviction, of type ARC_BUFC_METADATA, and reside in the
531 kstat_named_t arcstat_mfu_evictable_metadata;
533 * Total number of bytes that *would have been* consumed by ARC
534 * buffers in the arc_mfu_ghost state. See the comment above
535 * arcstat_mru_ghost_size for more details.
537 kstat_named_t arcstat_mfu_ghost_size;
539 * Number of bytes that *would have been* consumed by ARC
540 * buffers that are eligible for eviction, of type
541 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
543 kstat_named_t arcstat_mfu_ghost_evictable_data;
545 * Number of bytes that *would have been* consumed by ARC
546 * buffers that are eligible for eviction, of type
547 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
549 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
550 kstat_named_t arcstat_l2_hits;
551 kstat_named_t arcstat_l2_misses;
552 kstat_named_t arcstat_l2_feeds;
553 kstat_named_t arcstat_l2_rw_clash;
554 kstat_named_t arcstat_l2_read_bytes;
555 kstat_named_t arcstat_l2_write_bytes;
556 kstat_named_t arcstat_l2_writes_sent;
557 kstat_named_t arcstat_l2_writes_done;
558 kstat_named_t arcstat_l2_writes_error;
559 kstat_named_t arcstat_l2_writes_lock_retry;
560 kstat_named_t arcstat_l2_evict_lock_retry;
561 kstat_named_t arcstat_l2_evict_reading;
562 kstat_named_t arcstat_l2_evict_l1cached;
563 kstat_named_t arcstat_l2_free_on_write;
564 kstat_named_t arcstat_l2_cdata_free_on_write;
565 kstat_named_t arcstat_l2_abort_lowmem;
566 kstat_named_t arcstat_l2_cksum_bad;
567 kstat_named_t arcstat_l2_io_error;
568 kstat_named_t arcstat_l2_size;
569 kstat_named_t arcstat_l2_asize;
570 kstat_named_t arcstat_l2_hdr_size;
571 kstat_named_t arcstat_l2_compress_successes;
572 kstat_named_t arcstat_l2_compress_zeros;
573 kstat_named_t arcstat_l2_compress_failures;
574 kstat_named_t arcstat_l2_padding_needed;
575 kstat_named_t arcstat_l2_write_trylock_fail;
576 kstat_named_t arcstat_l2_write_passed_headroom;
577 kstat_named_t arcstat_l2_write_spa_mismatch;
578 kstat_named_t arcstat_l2_write_in_l2;
579 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
580 kstat_named_t arcstat_l2_write_not_cacheable;
581 kstat_named_t arcstat_l2_write_full;
582 kstat_named_t arcstat_l2_write_buffer_iter;
583 kstat_named_t arcstat_l2_write_pios;
584 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
585 kstat_named_t arcstat_l2_write_buffer_list_iter;
586 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
587 kstat_named_t arcstat_memory_throttle_count;
588 kstat_named_t arcstat_duplicate_buffers;
589 kstat_named_t arcstat_duplicate_buffers_size;
590 kstat_named_t arcstat_duplicate_reads;
591 kstat_named_t arcstat_meta_used;
592 kstat_named_t arcstat_meta_limit;
593 kstat_named_t arcstat_meta_max;
594 kstat_named_t arcstat_meta_min;
595 kstat_named_t arcstat_sync_wait_for_async;
596 kstat_named_t arcstat_demand_hit_predictive_prefetch;
599 static arc_stats_t arc_stats = {
600 { "hits", KSTAT_DATA_UINT64 },
601 { "misses", KSTAT_DATA_UINT64 },
602 { "demand_data_hits", KSTAT_DATA_UINT64 },
603 { "demand_data_misses", KSTAT_DATA_UINT64 },
604 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
605 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
606 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
607 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
608 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
609 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
610 { "mru_hits", KSTAT_DATA_UINT64 },
611 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
612 { "mfu_hits", KSTAT_DATA_UINT64 },
613 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
614 { "allocated", KSTAT_DATA_UINT64 },
615 { "deleted", KSTAT_DATA_UINT64 },
616 { "mutex_miss", KSTAT_DATA_UINT64 },
617 { "evict_skip", KSTAT_DATA_UINT64 },
618 { "evict_not_enough", KSTAT_DATA_UINT64 },
619 { "evict_l2_cached", KSTAT_DATA_UINT64 },
620 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
621 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
622 { "evict_l2_skip", KSTAT_DATA_UINT64 },
623 { "hash_elements", KSTAT_DATA_UINT64 },
624 { "hash_elements_max", KSTAT_DATA_UINT64 },
625 { "hash_collisions", KSTAT_DATA_UINT64 },
626 { "hash_chains", KSTAT_DATA_UINT64 },
627 { "hash_chain_max", KSTAT_DATA_UINT64 },
628 { "p", KSTAT_DATA_UINT64 },
629 { "c", KSTAT_DATA_UINT64 },
630 { "c_min", KSTAT_DATA_UINT64 },
631 { "c_max", KSTAT_DATA_UINT64 },
632 { "size", KSTAT_DATA_UINT64 },
633 { "hdr_size", KSTAT_DATA_UINT64 },
634 { "data_size", KSTAT_DATA_UINT64 },
635 { "metadata_size", KSTAT_DATA_UINT64 },
636 { "other_size", KSTAT_DATA_UINT64 },
637 { "anon_size", KSTAT_DATA_UINT64 },
638 { "anon_evictable_data", KSTAT_DATA_UINT64 },
639 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
640 { "mru_size", KSTAT_DATA_UINT64 },
641 { "mru_evictable_data", KSTAT_DATA_UINT64 },
642 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
643 { "mru_ghost_size", KSTAT_DATA_UINT64 },
644 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
645 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
646 { "mfu_size", KSTAT_DATA_UINT64 },
647 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
648 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
649 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
650 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
651 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
652 { "l2_hits", KSTAT_DATA_UINT64 },
653 { "l2_misses", KSTAT_DATA_UINT64 },
654 { "l2_feeds", KSTAT_DATA_UINT64 },
655 { "l2_rw_clash", KSTAT_DATA_UINT64 },
656 { "l2_read_bytes", KSTAT_DATA_UINT64 },
657 { "l2_write_bytes", KSTAT_DATA_UINT64 },
658 { "l2_writes_sent", KSTAT_DATA_UINT64 },
659 { "l2_writes_done", KSTAT_DATA_UINT64 },
660 { "l2_writes_error", KSTAT_DATA_UINT64 },
661 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
662 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
663 { "l2_evict_reading", KSTAT_DATA_UINT64 },
664 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
665 { "l2_free_on_write", KSTAT_DATA_UINT64 },
666 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 },
667 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
668 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
669 { "l2_io_error", KSTAT_DATA_UINT64 },
670 { "l2_size", KSTAT_DATA_UINT64 },
671 { "l2_asize", KSTAT_DATA_UINT64 },
672 { "l2_hdr_size", KSTAT_DATA_UINT64 },
673 { "l2_compress_successes", KSTAT_DATA_UINT64 },
674 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
675 { "l2_compress_failures", KSTAT_DATA_UINT64 },
676 { "l2_padding_needed", KSTAT_DATA_UINT64 },
677 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
678 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
679 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
680 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
681 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
682 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
683 { "l2_write_full", KSTAT_DATA_UINT64 },
684 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
685 { "l2_write_pios", KSTAT_DATA_UINT64 },
686 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
687 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
688 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
689 { "memory_throttle_count", KSTAT_DATA_UINT64 },
690 { "duplicate_buffers", KSTAT_DATA_UINT64 },
691 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
692 { "duplicate_reads", KSTAT_DATA_UINT64 },
693 { "arc_meta_used", KSTAT_DATA_UINT64 },
694 { "arc_meta_limit", KSTAT_DATA_UINT64 },
695 { "arc_meta_max", KSTAT_DATA_UINT64 },
696 { "arc_meta_min", KSTAT_DATA_UINT64 },
697 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
698 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
701 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
703 #define ARCSTAT_INCR(stat, val) \
704 atomic_add_64(&arc_stats.stat.value.ui64, (val))
706 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
707 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
709 #define ARCSTAT_MAX(stat, val) { \
711 while ((val) > (m = arc_stats.stat.value.ui64) && \
712 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
716 #define ARCSTAT_MAXSTAT(stat) \
717 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
720 * We define a macro to allow ARC hits/misses to be easily broken down by
721 * two separate conditions, giving a total of four different subtypes for
722 * each of hits and misses (so eight statistics total).
724 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
727 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
729 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
733 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
735 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
740 static arc_state_t *arc_anon;
741 static arc_state_t *arc_mru;
742 static arc_state_t *arc_mru_ghost;
743 static arc_state_t *arc_mfu;
744 static arc_state_t *arc_mfu_ghost;
745 static arc_state_t *arc_l2c_only;
748 * There are several ARC variables that are critical to export as kstats --
749 * but we don't want to have to grovel around in the kstat whenever we wish to
750 * manipulate them. For these variables, we therefore define them to be in
751 * terms of the statistic variable. This assures that we are not introducing
752 * the possibility of inconsistency by having shadow copies of the variables,
753 * while still allowing the code to be readable.
755 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
756 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
757 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
758 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
759 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
760 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
761 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
762 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
763 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
765 #define L2ARC_IS_VALID_COMPRESS(_c_) \
766 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
768 static int arc_no_grow; /* Don't try to grow cache size */
769 static uint64_t arc_tempreserve;
770 static uint64_t arc_loaned_bytes;
772 typedef struct arc_callback arc_callback_t;
774 struct arc_callback {
776 arc_done_func_t *acb_done;
778 zio_t *acb_zio_dummy;
779 arc_callback_t *acb_next;
782 typedef struct arc_write_callback arc_write_callback_t;
784 struct arc_write_callback {
786 arc_done_func_t *awcb_ready;
787 arc_done_func_t *awcb_children_ready;
788 arc_done_func_t *awcb_physdone;
789 arc_done_func_t *awcb_done;
794 * ARC buffers are separated into multiple structs as a memory saving measure:
795 * - Common fields struct, always defined, and embedded within it:
796 * - L2-only fields, always allocated but undefined when not in L2ARC
797 * - L1-only fields, only allocated when in L1ARC
799 * Buffer in L1 Buffer only in L2
800 * +------------------------+ +------------------------+
801 * | arc_buf_hdr_t | | arc_buf_hdr_t |
805 * +------------------------+ +------------------------+
806 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
807 * | (undefined if L1-only) | | |
808 * +------------------------+ +------------------------+
809 * | l1arc_buf_hdr_t |
814 * +------------------------+
816 * Because it's possible for the L2ARC to become extremely large, we can wind
817 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
818 * is minimized by only allocating the fields necessary for an L1-cached buffer
819 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
820 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
821 * words in pointers. arc_hdr_realloc() is used to switch a header between
822 * these two allocation states.
824 typedef struct l1arc_buf_hdr {
825 kmutex_t b_freeze_lock;
828 * used for debugging wtih kmem_flags - by allocating and freeing
829 * b_thawed when the buffer is thawed, we get a record of the stack
830 * trace that thawed it.
837 /* for waiting on writes to complete */
840 /* protected by arc state mutex */
841 arc_state_t *b_state;
842 multilist_node_t b_arc_node;
844 /* updated atomically */
845 clock_t b_arc_access;
847 /* self protecting */
850 arc_callback_t *b_acb;
851 /* temporary buffer holder for in-flight compressed or padded data */
855 typedef struct l2arc_dev l2arc_dev_t;
857 typedef struct l2arc_buf_hdr {
858 /* protected by arc_buf_hdr mutex */
859 l2arc_dev_t *b_dev; /* L2ARC device */
860 uint64_t b_daddr; /* disk address, offset byte */
861 /* real alloc'd buffer size depending on b_compress applied */
865 list_node_t b_l2node;
869 /* protected by hash lock */
873 * Even though this checksum is only set/verified when a buffer is in
874 * the L1 cache, it needs to be in the set of common fields because it
875 * must be preserved from the time before a buffer is written out to
876 * L2ARC until after it is read back in.
878 zio_cksum_t *b_freeze_cksum;
880 arc_buf_hdr_t *b_hash_next;
887 /* L2ARC fields. Undefined when not in L2ARC. */
888 l2arc_buf_hdr_t b_l2hdr;
889 /* L1ARC fields. Undefined when in l2arc_only state */
890 l1arc_buf_hdr_t b_l1hdr;
893 #if defined(__FreeBSD__) && defined(_KERNEL)
895 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
900 val = arc_meta_limit;
901 err = sysctl_handle_64(oidp, &val, 0, req);
902 if (err != 0 || req->newptr == NULL)
905 if (val <= 0 || val > arc_c_max)
908 arc_meta_limit = val;
913 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
919 err = sysctl_handle_64(oidp, &val, 0, req);
920 if (err != 0 || req->newptr == NULL)
923 if (zfs_arc_max == 0) {
924 /* Loader tunable so blindly set */
929 if (val < arc_abs_min || val > kmem_size())
933 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
939 arc_p = (arc_c >> 1);
941 if (zfs_arc_meta_limit == 0) {
942 /* limit meta-data to 1/4 of the arc capacity */
943 arc_meta_limit = arc_c_max / 4;
946 /* if kmem_flags are set, lets try to use less memory */
947 if (kmem_debugging())
956 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
962 err = sysctl_handle_64(oidp, &val, 0, req);
963 if (err != 0 || req->newptr == NULL)
966 if (zfs_arc_min == 0) {
967 /* Loader tunable so blindly set */
972 if (val < arc_abs_min || val > arc_c_max)
977 if (zfs_arc_meta_min == 0)
978 arc_meta_min = arc_c_min / 2;
980 if (arc_c < arc_c_min)
983 zfs_arc_min = arc_c_min;
989 static arc_buf_t *arc_eviction_list;
990 static arc_buf_hdr_t arc_eviction_hdr;
992 #define GHOST_STATE(state) \
993 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
994 (state) == arc_l2c_only)
996 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
997 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
998 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
999 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1000 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
1001 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
1003 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1004 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
1005 #define HDR_L2_READING(hdr) \
1006 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1007 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1008 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1009 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1010 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1012 #define HDR_ISTYPE_METADATA(hdr) \
1013 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1014 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1016 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1017 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1023 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1024 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1027 * Hash table routines
1030 #define HT_LOCK_PAD CACHE_LINE_SIZE
1035 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1039 #define BUF_LOCKS 256
1040 typedef struct buf_hash_table {
1042 arc_buf_hdr_t **ht_table;
1043 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1046 static buf_hash_table_t buf_hash_table;
1048 #define BUF_HASH_INDEX(spa, dva, birth) \
1049 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1050 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1051 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1052 #define HDR_LOCK(hdr) \
1053 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1055 uint64_t zfs_crc64_table[256];
1061 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1062 #define L2ARC_HEADROOM 2 /* num of writes */
1064 * If we discover during ARC scan any buffers to be compressed, we boost
1065 * our headroom for the next scanning cycle by this percentage multiple.
1067 #define L2ARC_HEADROOM_BOOST 200
1068 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1069 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1072 * Used to distinguish headers that are being process by
1073 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
1074 * address. This can happen when the header is added to the l2arc's list
1075 * of buffers to write in the first stage of l2arc_write_buffers(), but
1076 * has not yet been written out which happens in the second stage of
1077 * l2arc_write_buffers().
1079 #define L2ARC_ADDR_UNSET ((uint64_t)(-1))
1081 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1082 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1084 /* L2ARC Performance Tunables */
1085 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1086 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1087 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1088 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1089 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1090 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1091 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1092 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1093 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1095 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1096 &l2arc_write_max, 0, "max write size");
1097 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1098 &l2arc_write_boost, 0, "extra write during warmup");
1099 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1100 &l2arc_headroom, 0, "number of dev writes");
1101 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1102 &l2arc_feed_secs, 0, "interval seconds");
1103 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1104 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1106 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1107 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1108 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1109 &l2arc_feed_again, 0, "turbo warmup");
1110 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1111 &l2arc_norw, 0, "no reads during writes");
1113 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1114 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1115 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
1116 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
1117 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
1118 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
1120 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1121 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1122 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
1123 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
1124 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
1125 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
1127 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1128 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1129 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
1130 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
1131 "size of metadata in mru ghost state");
1132 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
1133 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
1134 "size of data in mru ghost state");
1136 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1137 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1138 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
1139 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
1140 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
1141 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
1143 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1144 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1145 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
1146 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
1147 "size of metadata in mfu ghost state");
1148 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
1149 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
1150 "size of data in mfu ghost state");
1152 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1153 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1159 vdev_t *l2ad_vdev; /* vdev */
1160 spa_t *l2ad_spa; /* spa */
1161 uint64_t l2ad_hand; /* next write location */
1162 uint64_t l2ad_start; /* first addr on device */
1163 uint64_t l2ad_end; /* last addr on device */
1164 boolean_t l2ad_first; /* first sweep through */
1165 boolean_t l2ad_writing; /* currently writing */
1166 kmutex_t l2ad_mtx; /* lock for buffer list */
1167 list_t l2ad_buflist; /* buffer list */
1168 list_node_t l2ad_node; /* device list node */
1169 refcount_t l2ad_alloc; /* allocated bytes */
1172 static list_t L2ARC_dev_list; /* device list */
1173 static list_t *l2arc_dev_list; /* device list pointer */
1174 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1175 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1176 static list_t L2ARC_free_on_write; /* free after write buf list */
1177 static list_t *l2arc_free_on_write; /* free after write list ptr */
1178 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1179 static uint64_t l2arc_ndev; /* number of devices */
1181 typedef struct l2arc_read_callback {
1182 arc_buf_t *l2rcb_buf; /* read buffer */
1183 spa_t *l2rcb_spa; /* spa */
1184 blkptr_t l2rcb_bp; /* original blkptr */
1185 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1186 int l2rcb_flags; /* original flags */
1187 enum zio_compress l2rcb_compress; /* applied compress */
1188 void *l2rcb_data; /* temporary buffer */
1189 } l2arc_read_callback_t;
1191 typedef struct l2arc_write_callback {
1192 l2arc_dev_t *l2wcb_dev; /* device info */
1193 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1194 } l2arc_write_callback_t;
1196 typedef struct l2arc_data_free {
1197 /* protected by l2arc_free_on_write_mtx */
1200 void (*l2df_func)(void *, size_t);
1201 list_node_t l2df_list_node;
1202 } l2arc_data_free_t;
1204 static kmutex_t l2arc_feed_thr_lock;
1205 static kcondvar_t l2arc_feed_thr_cv;
1206 static uint8_t l2arc_thread_exit;
1208 static void arc_get_data_buf(arc_buf_t *);
1209 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1210 static boolean_t arc_is_overflowing();
1211 static void arc_buf_watch(arc_buf_t *);
1213 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1214 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1216 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1217 static void l2arc_read_done(zio_t *);
1219 static boolean_t l2arc_transform_buf(arc_buf_hdr_t *, boolean_t);
1220 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress);
1221 static void l2arc_release_cdata_buf(arc_buf_hdr_t *);
1224 l2arc_trim(const arc_buf_hdr_t *hdr)
1226 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1228 ASSERT(HDR_HAS_L2HDR(hdr));
1229 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1231 if (hdr->b_l2hdr.b_daddr == L2ARC_ADDR_UNSET)
1233 if (hdr->b_l2hdr.b_asize != 0) {
1234 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1235 hdr->b_l2hdr.b_asize, 0);
1237 ASSERT3U(hdr->b_l2hdr.b_compress, ==, ZIO_COMPRESS_EMPTY);
1242 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1244 uint8_t *vdva = (uint8_t *)dva;
1245 uint64_t crc = -1ULL;
1248 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1250 for (i = 0; i < sizeof (dva_t); i++)
1251 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1253 crc ^= (spa>>8) ^ birth;
1258 #define BUF_EMPTY(buf) \
1259 ((buf)->b_dva.dva_word[0] == 0 && \
1260 (buf)->b_dva.dva_word[1] == 0)
1262 #define BUF_EQUAL(spa, dva, birth, buf) \
1263 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1264 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1265 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
1268 buf_discard_identity(arc_buf_hdr_t *hdr)
1270 hdr->b_dva.dva_word[0] = 0;
1271 hdr->b_dva.dva_word[1] = 0;
1275 static arc_buf_hdr_t *
1276 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1278 const dva_t *dva = BP_IDENTITY(bp);
1279 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1280 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1281 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1284 mutex_enter(hash_lock);
1285 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1286 hdr = hdr->b_hash_next) {
1287 if (BUF_EQUAL(spa, dva, birth, hdr)) {
1292 mutex_exit(hash_lock);
1298 * Insert an entry into the hash table. If there is already an element
1299 * equal to elem in the hash table, then the already existing element
1300 * will be returned and the new element will not be inserted.
1301 * Otherwise returns NULL.
1302 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1304 static arc_buf_hdr_t *
1305 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1307 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1308 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1309 arc_buf_hdr_t *fhdr;
1312 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1313 ASSERT(hdr->b_birth != 0);
1314 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1316 if (lockp != NULL) {
1318 mutex_enter(hash_lock);
1320 ASSERT(MUTEX_HELD(hash_lock));
1323 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1324 fhdr = fhdr->b_hash_next, i++) {
1325 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1329 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1330 buf_hash_table.ht_table[idx] = hdr;
1331 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
1333 /* collect some hash table performance data */
1335 ARCSTAT_BUMP(arcstat_hash_collisions);
1337 ARCSTAT_BUMP(arcstat_hash_chains);
1339 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1342 ARCSTAT_BUMP(arcstat_hash_elements);
1343 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1349 buf_hash_remove(arc_buf_hdr_t *hdr)
1351 arc_buf_hdr_t *fhdr, **hdrp;
1352 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1354 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1355 ASSERT(HDR_IN_HASH_TABLE(hdr));
1357 hdrp = &buf_hash_table.ht_table[idx];
1358 while ((fhdr = *hdrp) != hdr) {
1359 ASSERT(fhdr != NULL);
1360 hdrp = &fhdr->b_hash_next;
1362 *hdrp = hdr->b_hash_next;
1363 hdr->b_hash_next = NULL;
1364 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE;
1366 /* collect some hash table performance data */
1367 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1369 if (buf_hash_table.ht_table[idx] &&
1370 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1371 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1375 * Global data structures and functions for the buf kmem cache.
1377 static kmem_cache_t *hdr_full_cache;
1378 static kmem_cache_t *hdr_l2only_cache;
1379 static kmem_cache_t *buf_cache;
1386 kmem_free(buf_hash_table.ht_table,
1387 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1388 for (i = 0; i < BUF_LOCKS; i++)
1389 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1390 kmem_cache_destroy(hdr_full_cache);
1391 kmem_cache_destroy(hdr_l2only_cache);
1392 kmem_cache_destroy(buf_cache);
1396 * Constructor callback - called when the cache is empty
1397 * and a new buf is requested.
1401 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1403 arc_buf_hdr_t *hdr = vbuf;
1405 bzero(hdr, HDR_FULL_SIZE);
1406 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1407 refcount_create(&hdr->b_l1hdr.b_refcnt);
1408 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1409 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1410 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1417 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1419 arc_buf_hdr_t *hdr = vbuf;
1421 bzero(hdr, HDR_L2ONLY_SIZE);
1422 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1429 buf_cons(void *vbuf, void *unused, int kmflag)
1431 arc_buf_t *buf = vbuf;
1433 bzero(buf, sizeof (arc_buf_t));
1434 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1435 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1441 * Destructor callback - called when a cached buf is
1442 * no longer required.
1446 hdr_full_dest(void *vbuf, void *unused)
1448 arc_buf_hdr_t *hdr = vbuf;
1450 ASSERT(BUF_EMPTY(hdr));
1451 cv_destroy(&hdr->b_l1hdr.b_cv);
1452 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1453 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1454 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1455 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1460 hdr_l2only_dest(void *vbuf, void *unused)
1462 arc_buf_hdr_t *hdr = vbuf;
1464 ASSERT(BUF_EMPTY(hdr));
1465 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1470 buf_dest(void *vbuf, void *unused)
1472 arc_buf_t *buf = vbuf;
1474 mutex_destroy(&buf->b_evict_lock);
1475 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1479 * Reclaim callback -- invoked when memory is low.
1483 hdr_recl(void *unused)
1485 dprintf("hdr_recl called\n");
1487 * umem calls the reclaim func when we destroy the buf cache,
1488 * which is after we do arc_fini().
1491 cv_signal(&arc_reclaim_thread_cv);
1498 uint64_t hsize = 1ULL << 12;
1502 * The hash table is big enough to fill all of physical memory
1503 * with an average block size of zfs_arc_average_blocksize (default 8K).
1504 * By default, the table will take up
1505 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1507 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1510 buf_hash_table.ht_mask = hsize - 1;
1511 buf_hash_table.ht_table =
1512 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1513 if (buf_hash_table.ht_table == NULL) {
1514 ASSERT(hsize > (1ULL << 8));
1519 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1520 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1521 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1522 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1524 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1525 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1527 for (i = 0; i < 256; i++)
1528 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1529 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1531 for (i = 0; i < BUF_LOCKS; i++) {
1532 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1533 NULL, MUTEX_DEFAULT, NULL);
1538 * Transition between the two allocation states for the arc_buf_hdr struct.
1539 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1540 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1541 * version is used when a cache buffer is only in the L2ARC in order to reduce
1544 static arc_buf_hdr_t *
1545 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
1547 ASSERT(HDR_HAS_L2HDR(hdr));
1549 arc_buf_hdr_t *nhdr;
1550 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1552 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
1553 (old == hdr_l2only_cache && new == hdr_full_cache));
1555 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
1557 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
1558 buf_hash_remove(hdr);
1560 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
1562 if (new == hdr_full_cache) {
1563 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1565 * arc_access and arc_change_state need to be aware that a
1566 * header has just come out of L2ARC, so we set its state to
1567 * l2c_only even though it's about to change.
1569 nhdr->b_l1hdr.b_state = arc_l2c_only;
1571 /* Verify previous threads set to NULL before freeing */
1572 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1574 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1575 ASSERT0(hdr->b_l1hdr.b_datacnt);
1578 * If we've reached here, We must have been called from
1579 * arc_evict_hdr(), as such we should have already been
1580 * removed from any ghost list we were previously on
1581 * (which protects us from racing with arc_evict_state),
1582 * thus no locking is needed during this check.
1584 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1587 * A buffer must not be moved into the arc_l2c_only
1588 * state if it's not finished being written out to the
1589 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1590 * might try to be accessed, even though it was removed.
1592 VERIFY(!HDR_L2_WRITING(hdr));
1593 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1596 if (hdr->b_l1hdr.b_thawed != NULL) {
1597 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1598 hdr->b_l1hdr.b_thawed = NULL;
1602 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR;
1605 * The header has been reallocated so we need to re-insert it into any
1608 (void) buf_hash_insert(nhdr, NULL);
1610 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
1612 mutex_enter(&dev->l2ad_mtx);
1615 * We must place the realloc'ed header back into the list at
1616 * the same spot. Otherwise, if it's placed earlier in the list,
1617 * l2arc_write_buffers() could find it during the function's
1618 * write phase, and try to write it out to the l2arc.
1620 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
1621 list_remove(&dev->l2ad_buflist, hdr);
1623 mutex_exit(&dev->l2ad_mtx);
1626 * Since we're using the pointer address as the tag when
1627 * incrementing and decrementing the l2ad_alloc refcount, we
1628 * must remove the old pointer (that we're about to destroy) and
1629 * add the new pointer to the refcount. Otherwise we'd remove
1630 * the wrong pointer address when calling arc_hdr_destroy() later.
1633 (void) refcount_remove_many(&dev->l2ad_alloc,
1634 hdr->b_l2hdr.b_asize, hdr);
1636 (void) refcount_add_many(&dev->l2ad_alloc,
1637 nhdr->b_l2hdr.b_asize, nhdr);
1639 buf_discard_identity(hdr);
1640 hdr->b_freeze_cksum = NULL;
1641 kmem_cache_free(old, hdr);
1647 #define ARC_MINTIME (hz>>4) /* 62 ms */
1650 arc_cksum_verify(arc_buf_t *buf)
1654 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1657 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1658 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) {
1659 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1662 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc);
1663 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1664 panic("buffer modified while frozen!");
1665 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1669 arc_cksum_equal(arc_buf_t *buf)
1674 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1675 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc);
1676 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1677 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1683 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1685 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1688 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1689 if (buf->b_hdr->b_freeze_cksum != NULL) {
1690 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1693 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1694 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1695 NULL, buf->b_hdr->b_freeze_cksum);
1696 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1704 typedef struct procctl {
1712 arc_buf_unwatch(arc_buf_t *buf)
1719 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1720 ctl.prwatch.pr_size = 0;
1721 ctl.prwatch.pr_wflags = 0;
1722 result = write(arc_procfd, &ctl, sizeof (ctl));
1723 ASSERT3U(result, ==, sizeof (ctl));
1730 arc_buf_watch(arc_buf_t *buf)
1737 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1738 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1739 ctl.prwatch.pr_wflags = WA_WRITE;
1740 result = write(arc_procfd, &ctl, sizeof (ctl));
1741 ASSERT3U(result, ==, sizeof (ctl));
1745 #endif /* illumos */
1747 static arc_buf_contents_t
1748 arc_buf_type(arc_buf_hdr_t *hdr)
1750 if (HDR_ISTYPE_METADATA(hdr)) {
1751 return (ARC_BUFC_METADATA);
1753 return (ARC_BUFC_DATA);
1758 arc_bufc_to_flags(arc_buf_contents_t type)
1762 /* metadata field is 0 if buffer contains normal data */
1764 case ARC_BUFC_METADATA:
1765 return (ARC_FLAG_BUFC_METADATA);
1769 panic("undefined ARC buffer type!");
1770 return ((uint32_t)-1);
1774 arc_buf_thaw(arc_buf_t *buf)
1776 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1777 if (buf->b_hdr->b_l1hdr.b_state != arc_anon)
1778 panic("modifying non-anon buffer!");
1779 if (HDR_IO_IN_PROGRESS(buf->b_hdr))
1780 panic("modifying buffer while i/o in progress!");
1781 arc_cksum_verify(buf);
1784 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1785 if (buf->b_hdr->b_freeze_cksum != NULL) {
1786 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1787 buf->b_hdr->b_freeze_cksum = NULL;
1791 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1792 if (buf->b_hdr->b_l1hdr.b_thawed != NULL)
1793 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1);
1794 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1798 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1801 arc_buf_unwatch(buf);
1806 arc_buf_freeze(arc_buf_t *buf)
1808 kmutex_t *hash_lock;
1810 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1813 hash_lock = HDR_LOCK(buf->b_hdr);
1814 mutex_enter(hash_lock);
1816 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1817 buf->b_hdr->b_l1hdr.b_state == arc_anon);
1818 arc_cksum_compute(buf, B_FALSE);
1819 mutex_exit(hash_lock);
1824 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1826 ASSERT(HDR_HAS_L1HDR(hdr));
1827 ASSERT(MUTEX_HELD(hash_lock));
1828 arc_state_t *state = hdr->b_l1hdr.b_state;
1830 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
1831 (state != arc_anon)) {
1832 /* We don't use the L2-only state list. */
1833 if (state != arc_l2c_only) {
1834 arc_buf_contents_t type = arc_buf_type(hdr);
1835 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt;
1836 multilist_t *list = &state->arcs_list[type];
1837 uint64_t *size = &state->arcs_lsize[type];
1839 multilist_remove(list, hdr);
1841 if (GHOST_STATE(state)) {
1842 ASSERT0(hdr->b_l1hdr.b_datacnt);
1843 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
1844 delta = hdr->b_size;
1847 ASSERT3U(*size, >=, delta);
1848 atomic_add_64(size, -delta);
1850 /* remove the prefetch flag if we get a reference */
1851 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
1856 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1859 arc_state_t *state = hdr->b_l1hdr.b_state;
1861 ASSERT(HDR_HAS_L1HDR(hdr));
1862 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1863 ASSERT(!GHOST_STATE(state));
1866 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1867 * check to prevent usage of the arc_l2c_only list.
1869 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
1870 (state != arc_anon)) {
1871 arc_buf_contents_t type = arc_buf_type(hdr);
1872 multilist_t *list = &state->arcs_list[type];
1873 uint64_t *size = &state->arcs_lsize[type];
1875 multilist_insert(list, hdr);
1877 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
1878 atomic_add_64(size, hdr->b_size *
1879 hdr->b_l1hdr.b_datacnt);
1885 * Move the supplied buffer to the indicated state. The hash lock
1886 * for the buffer must be held by the caller.
1889 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
1890 kmutex_t *hash_lock)
1892 arc_state_t *old_state;
1895 uint64_t from_delta, to_delta;
1896 arc_buf_contents_t buftype = arc_buf_type(hdr);
1899 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1900 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1901 * L1 hdr doesn't always exist when we change state to arc_anon before
1902 * destroying a header, in which case reallocating to add the L1 hdr is
1905 if (HDR_HAS_L1HDR(hdr)) {
1906 old_state = hdr->b_l1hdr.b_state;
1907 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
1908 datacnt = hdr->b_l1hdr.b_datacnt;
1910 old_state = arc_l2c_only;
1915 ASSERT(MUTEX_HELD(hash_lock));
1916 ASSERT3P(new_state, !=, old_state);
1917 ASSERT(refcnt == 0 || datacnt > 0);
1918 ASSERT(!GHOST_STATE(new_state) || datacnt == 0);
1919 ASSERT(old_state != arc_anon || datacnt <= 1);
1921 from_delta = to_delta = datacnt * hdr->b_size;
1924 * If this buffer is evictable, transfer it from the
1925 * old state list to the new state list.
1928 if (old_state != arc_anon && old_state != arc_l2c_only) {
1929 uint64_t *size = &old_state->arcs_lsize[buftype];
1931 ASSERT(HDR_HAS_L1HDR(hdr));
1932 multilist_remove(&old_state->arcs_list[buftype], hdr);
1935 * If prefetching out of the ghost cache,
1936 * we will have a non-zero datacnt.
1938 if (GHOST_STATE(old_state) && datacnt == 0) {
1939 /* ghost elements have a ghost size */
1940 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1941 from_delta = hdr->b_size;
1943 ASSERT3U(*size, >=, from_delta);
1944 atomic_add_64(size, -from_delta);
1946 if (new_state != arc_anon && new_state != arc_l2c_only) {
1947 uint64_t *size = &new_state->arcs_lsize[buftype];
1950 * An L1 header always exists here, since if we're
1951 * moving to some L1-cached state (i.e. not l2c_only or
1952 * anonymous), we realloc the header to add an L1hdr
1955 ASSERT(HDR_HAS_L1HDR(hdr));
1956 multilist_insert(&new_state->arcs_list[buftype], hdr);
1958 /* ghost elements have a ghost size */
1959 if (GHOST_STATE(new_state)) {
1961 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1962 to_delta = hdr->b_size;
1964 atomic_add_64(size, to_delta);
1968 ASSERT(!BUF_EMPTY(hdr));
1969 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
1970 buf_hash_remove(hdr);
1972 /* adjust state sizes (ignore arc_l2c_only) */
1974 if (to_delta && new_state != arc_l2c_only) {
1975 ASSERT(HDR_HAS_L1HDR(hdr));
1976 if (GHOST_STATE(new_state)) {
1980 * We moving a header to a ghost state, we first
1981 * remove all arc buffers. Thus, we'll have a
1982 * datacnt of zero, and no arc buffer to use for
1983 * the reference. As a result, we use the arc
1984 * header pointer for the reference.
1986 (void) refcount_add_many(&new_state->arcs_size,
1989 ASSERT3U(datacnt, !=, 0);
1992 * Each individual buffer holds a unique reference,
1993 * thus we must remove each of these references one
1996 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1997 buf = buf->b_next) {
1998 (void) refcount_add_many(&new_state->arcs_size,
2004 if (from_delta && old_state != arc_l2c_only) {
2005 ASSERT(HDR_HAS_L1HDR(hdr));
2006 if (GHOST_STATE(old_state)) {
2008 * When moving a header off of a ghost state,
2009 * there's the possibility for datacnt to be
2010 * non-zero. This is because we first add the
2011 * arc buffer to the header prior to changing
2012 * the header's state. Since we used the header
2013 * for the reference when putting the header on
2014 * the ghost state, we must balance that and use
2015 * the header when removing off the ghost state
2016 * (even though datacnt is non zero).
2019 IMPLY(datacnt == 0, new_state == arc_anon ||
2020 new_state == arc_l2c_only);
2022 (void) refcount_remove_many(&old_state->arcs_size,
2025 ASSERT3P(datacnt, !=, 0);
2028 * Each individual buffer holds a unique reference,
2029 * thus we must remove each of these references one
2032 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2033 buf = buf->b_next) {
2034 (void) refcount_remove_many(
2035 &old_state->arcs_size, hdr->b_size, buf);
2040 if (HDR_HAS_L1HDR(hdr))
2041 hdr->b_l1hdr.b_state = new_state;
2044 * L2 headers should never be on the L2 state list since they don't
2045 * have L1 headers allocated.
2047 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2048 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2052 arc_space_consume(uint64_t space, arc_space_type_t type)
2054 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2057 case ARC_SPACE_DATA:
2058 ARCSTAT_INCR(arcstat_data_size, space);
2060 case ARC_SPACE_META:
2061 ARCSTAT_INCR(arcstat_metadata_size, space);
2063 case ARC_SPACE_OTHER:
2064 ARCSTAT_INCR(arcstat_other_size, space);
2066 case ARC_SPACE_HDRS:
2067 ARCSTAT_INCR(arcstat_hdr_size, space);
2069 case ARC_SPACE_L2HDRS:
2070 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
2074 if (type != ARC_SPACE_DATA)
2075 ARCSTAT_INCR(arcstat_meta_used, space);
2077 atomic_add_64(&arc_size, space);
2081 arc_space_return(uint64_t space, arc_space_type_t type)
2083 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2086 case ARC_SPACE_DATA:
2087 ARCSTAT_INCR(arcstat_data_size, -space);
2089 case ARC_SPACE_META:
2090 ARCSTAT_INCR(arcstat_metadata_size, -space);
2092 case ARC_SPACE_OTHER:
2093 ARCSTAT_INCR(arcstat_other_size, -space);
2095 case ARC_SPACE_HDRS:
2096 ARCSTAT_INCR(arcstat_hdr_size, -space);
2098 case ARC_SPACE_L2HDRS:
2099 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
2103 if (type != ARC_SPACE_DATA) {
2104 ASSERT(arc_meta_used >= space);
2105 if (arc_meta_max < arc_meta_used)
2106 arc_meta_max = arc_meta_used;
2107 ARCSTAT_INCR(arcstat_meta_used, -space);
2110 ASSERT(arc_size >= space);
2111 atomic_add_64(&arc_size, -space);
2115 arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
2120 ASSERT3U(size, >, 0);
2121 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
2122 ASSERT(BUF_EMPTY(hdr));
2123 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
2125 hdr->b_spa = spa_load_guid(spa);
2127 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2130 buf->b_efunc = NULL;
2131 buf->b_private = NULL;
2134 hdr->b_flags = arc_bufc_to_flags(type);
2135 hdr->b_flags |= ARC_FLAG_HAS_L1HDR;
2137 hdr->b_l1hdr.b_buf = buf;
2138 hdr->b_l1hdr.b_state = arc_anon;
2139 hdr->b_l1hdr.b_arc_access = 0;
2140 hdr->b_l1hdr.b_datacnt = 1;
2141 hdr->b_l1hdr.b_tmp_cdata = NULL;
2143 arc_get_data_buf(buf);
2144 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2145 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2150 static char *arc_onloan_tag = "onloan";
2153 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2154 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2155 * buffers must be returned to the arc before they can be used by the DMU or
2159 arc_loan_buf(spa_t *spa, int size)
2163 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
2165 atomic_add_64(&arc_loaned_bytes, size);
2170 * Return a loaned arc buffer to the arc.
2173 arc_return_buf(arc_buf_t *buf, void *tag)
2175 arc_buf_hdr_t *hdr = buf->b_hdr;
2177 ASSERT(buf->b_data != NULL);
2178 ASSERT(HDR_HAS_L1HDR(hdr));
2179 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2180 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2182 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
2185 /* Detach an arc_buf from a dbuf (tag) */
2187 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2189 arc_buf_hdr_t *hdr = buf->b_hdr;
2191 ASSERT(buf->b_data != NULL);
2192 ASSERT(HDR_HAS_L1HDR(hdr));
2193 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2194 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2195 buf->b_efunc = NULL;
2196 buf->b_private = NULL;
2198 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
2202 arc_buf_clone(arc_buf_t *from)
2205 arc_buf_hdr_t *hdr = from->b_hdr;
2206 uint64_t size = hdr->b_size;
2208 ASSERT(HDR_HAS_L1HDR(hdr));
2209 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2211 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2214 buf->b_efunc = NULL;
2215 buf->b_private = NULL;
2216 buf->b_next = hdr->b_l1hdr.b_buf;
2217 hdr->b_l1hdr.b_buf = buf;
2218 arc_get_data_buf(buf);
2219 bcopy(from->b_data, buf->b_data, size);
2222 * This buffer already exists in the arc so create a duplicate
2223 * copy for the caller. If the buffer is associated with user data
2224 * then track the size and number of duplicates. These stats will be
2225 * updated as duplicate buffers are created and destroyed.
2227 if (HDR_ISTYPE_DATA(hdr)) {
2228 ARCSTAT_BUMP(arcstat_duplicate_buffers);
2229 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
2231 hdr->b_l1hdr.b_datacnt += 1;
2236 arc_buf_add_ref(arc_buf_t *buf, void* tag)
2239 kmutex_t *hash_lock;
2242 * Check to see if this buffer is evicted. Callers
2243 * must verify b_data != NULL to know if the add_ref
2246 mutex_enter(&buf->b_evict_lock);
2247 if (buf->b_data == NULL) {
2248 mutex_exit(&buf->b_evict_lock);
2251 hash_lock = HDR_LOCK(buf->b_hdr);
2252 mutex_enter(hash_lock);
2254 ASSERT(HDR_HAS_L1HDR(hdr));
2255 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2256 mutex_exit(&buf->b_evict_lock);
2258 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
2259 hdr->b_l1hdr.b_state == arc_mfu);
2261 add_reference(hdr, hash_lock, tag);
2262 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2263 arc_access(hdr, hash_lock);
2264 mutex_exit(hash_lock);
2265 ARCSTAT_BUMP(arcstat_hits);
2266 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
2267 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
2268 data, metadata, hits);
2272 arc_buf_free_on_write(void *data, size_t size,
2273 void (*free_func)(void *, size_t))
2275 l2arc_data_free_t *df;
2277 df = kmem_alloc(sizeof (*df), KM_SLEEP);
2278 df->l2df_data = data;
2279 df->l2df_size = size;
2280 df->l2df_func = free_func;
2281 mutex_enter(&l2arc_free_on_write_mtx);
2282 list_insert_head(l2arc_free_on_write, df);
2283 mutex_exit(&l2arc_free_on_write_mtx);
2287 * Free the arc data buffer. If it is an l2arc write in progress,
2288 * the buffer is placed on l2arc_free_on_write to be freed later.
2291 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
2293 arc_buf_hdr_t *hdr = buf->b_hdr;
2295 if (HDR_L2_WRITING(hdr)) {
2296 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
2297 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2299 free_func(buf->b_data, hdr->b_size);
2304 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
2306 size_t align, asize, len;
2308 ASSERT(HDR_HAS_L2HDR(hdr));
2309 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx));
2312 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
2313 * that doesn't exist, the header is in the arc_l2c_only state,
2314 * and there isn't anything to free (it's already been freed).
2316 if (!HDR_HAS_L1HDR(hdr))
2320 * The header isn't being written to the l2arc device, thus it
2321 * shouldn't have a b_tmp_cdata to free.
2323 if (!HDR_L2_WRITING(hdr)) {
2324 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2329 * The bufer has been chosen for writing to L2ARC, but it's
2330 * not being written just yet. In other words,
2331 * b_tmp_cdata points to exactly the same buffer as b_data,
2332 * l2arc_transform_buf hasn't been called.
2334 if (hdr->b_l2hdr.b_daddr == L2ARC_ADDR_UNSET) {
2335 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==,
2336 hdr->b_l1hdr.b_buf->b_data);
2337 ASSERT3U(hdr->b_l2hdr.b_compress, ==, ZIO_COMPRESS_OFF);
2338 hdr->b_l1hdr.b_tmp_cdata = NULL;
2343 * There's nothing to free since the buffer was all zero's and
2344 * compressed to a zero length buffer.
2346 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) {
2347 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2352 * Nothing to do if the temporary buffer was not required.
2354 if (hdr->b_l1hdr.b_tmp_cdata == NULL)
2357 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
2359 align = (size_t)1 << hdr->b_l2hdr.b_dev->l2ad_vdev->vdev_ashift;
2360 asize = P2ROUNDUP(len, align);
2361 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, asize,
2363 hdr->b_l1hdr.b_tmp_cdata = NULL;
2367 * Free up buf->b_data and if 'remove' is set, then pull the
2368 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2371 arc_buf_destroy(arc_buf_t *buf, boolean_t remove)
2375 /* free up data associated with the buf */
2376 if (buf->b_data != NULL) {
2377 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
2378 uint64_t size = buf->b_hdr->b_size;
2379 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
2381 arc_cksum_verify(buf);
2383 arc_buf_unwatch(buf);
2386 if (type == ARC_BUFC_METADATA) {
2387 arc_buf_data_free(buf, zio_buf_free);
2388 arc_space_return(size, ARC_SPACE_META);
2390 ASSERT(type == ARC_BUFC_DATA);
2391 arc_buf_data_free(buf, zio_data_buf_free);
2392 arc_space_return(size, ARC_SPACE_DATA);
2395 /* protected by hash lock, if in the hash table */
2396 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) {
2397 uint64_t *cnt = &state->arcs_lsize[type];
2399 ASSERT(refcount_is_zero(
2400 &buf->b_hdr->b_l1hdr.b_refcnt));
2401 ASSERT(state != arc_anon && state != arc_l2c_only);
2403 ASSERT3U(*cnt, >=, size);
2404 atomic_add_64(cnt, -size);
2407 (void) refcount_remove_many(&state->arcs_size, size, buf);
2411 * If we're destroying a duplicate buffer make sure
2412 * that the appropriate statistics are updated.
2414 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 &&
2415 HDR_ISTYPE_DATA(buf->b_hdr)) {
2416 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
2417 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
2419 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0);
2420 buf->b_hdr->b_l1hdr.b_datacnt -= 1;
2423 /* only remove the buf if requested */
2427 /* remove the buf from the hdr list */
2428 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf;
2429 bufp = &(*bufp)->b_next)
2431 *bufp = buf->b_next;
2434 ASSERT(buf->b_efunc == NULL);
2436 /* clean up the buf */
2438 kmem_cache_free(buf_cache, buf);
2442 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2444 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2445 l2arc_dev_t *dev = l2hdr->b_dev;
2447 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2448 ASSERT(HDR_HAS_L2HDR(hdr));
2450 list_remove(&dev->l2ad_buflist, hdr);
2453 * We don't want to leak the b_tmp_cdata buffer that was
2454 * allocated in l2arc_write_buffers()
2456 arc_buf_l2_cdata_free(hdr);
2459 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2460 * this header is being processed by l2arc_write_buffers() (i.e.
2461 * it's in the first stage of l2arc_write_buffers()).
2462 * Re-affirming that truth here, just to serve as a reminder. If
2463 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2464 * may not have its HDR_L2_WRITING flag set. (the write may have
2465 * completed, in which case HDR_L2_WRITING will be false and the
2466 * b_daddr field will point to the address of the buffer on disk).
2468 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr));
2471 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2472 * l2arc_write_buffers(). Since we've just removed this header
2473 * from the l2arc buffer list, this header will never reach the
2474 * second stage of l2arc_write_buffers(), which increments the
2475 * accounting stats for this header. Thus, we must be careful
2476 * not to decrement them for this header either.
2478 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) {
2479 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
2480 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
2482 vdev_space_update(dev->l2ad_vdev,
2483 -l2hdr->b_asize, 0, 0);
2485 (void) refcount_remove_many(&dev->l2ad_alloc,
2486 l2hdr->b_asize, hdr);
2489 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
2493 arc_hdr_destroy(arc_buf_hdr_t *hdr)
2495 if (HDR_HAS_L1HDR(hdr)) {
2496 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
2497 hdr->b_l1hdr.b_datacnt > 0);
2498 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2499 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2501 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2502 ASSERT(!HDR_IN_HASH_TABLE(hdr));
2504 if (HDR_HAS_L2HDR(hdr)) {
2505 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2506 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
2509 mutex_enter(&dev->l2ad_mtx);
2512 * Even though we checked this conditional above, we
2513 * need to check this again now that we have the
2514 * l2ad_mtx. This is because we could be racing with
2515 * another thread calling l2arc_evict() which might have
2516 * destroyed this header's L2 portion as we were waiting
2517 * to acquire the l2ad_mtx. If that happens, we don't
2518 * want to re-destroy the header's L2 portion.
2520 if (HDR_HAS_L2HDR(hdr)) {
2522 arc_hdr_l2hdr_destroy(hdr);
2526 mutex_exit(&dev->l2ad_mtx);
2529 if (!BUF_EMPTY(hdr))
2530 buf_discard_identity(hdr);
2532 if (hdr->b_freeze_cksum != NULL) {
2533 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
2534 hdr->b_freeze_cksum = NULL;
2537 if (HDR_HAS_L1HDR(hdr)) {
2538 while (hdr->b_l1hdr.b_buf) {
2539 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2541 if (buf->b_efunc != NULL) {
2542 mutex_enter(&arc_user_evicts_lock);
2543 mutex_enter(&buf->b_evict_lock);
2544 ASSERT(buf->b_hdr != NULL);
2545 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE);
2546 hdr->b_l1hdr.b_buf = buf->b_next;
2547 buf->b_hdr = &arc_eviction_hdr;
2548 buf->b_next = arc_eviction_list;
2549 arc_eviction_list = buf;
2550 mutex_exit(&buf->b_evict_lock);
2551 cv_signal(&arc_user_evicts_cv);
2552 mutex_exit(&arc_user_evicts_lock);
2554 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE);
2558 if (hdr->b_l1hdr.b_thawed != NULL) {
2559 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2560 hdr->b_l1hdr.b_thawed = NULL;
2565 ASSERT3P(hdr->b_hash_next, ==, NULL);
2566 if (HDR_HAS_L1HDR(hdr)) {
2567 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2568 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
2569 kmem_cache_free(hdr_full_cache, hdr);
2571 kmem_cache_free(hdr_l2only_cache, hdr);
2576 arc_buf_free(arc_buf_t *buf, void *tag)
2578 arc_buf_hdr_t *hdr = buf->b_hdr;
2579 int hashed = hdr->b_l1hdr.b_state != arc_anon;
2581 ASSERT(buf->b_efunc == NULL);
2582 ASSERT(buf->b_data != NULL);
2585 kmutex_t *hash_lock = HDR_LOCK(hdr);
2587 mutex_enter(hash_lock);
2589 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2591 (void) remove_reference(hdr, hash_lock, tag);
2592 if (hdr->b_l1hdr.b_datacnt > 1) {
2593 arc_buf_destroy(buf, TRUE);
2595 ASSERT(buf == hdr->b_l1hdr.b_buf);
2596 ASSERT(buf->b_efunc == NULL);
2597 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2599 mutex_exit(hash_lock);
2600 } else if (HDR_IO_IN_PROGRESS(hdr)) {
2603 * We are in the middle of an async write. Don't destroy
2604 * this buffer unless the write completes before we finish
2605 * decrementing the reference count.
2607 mutex_enter(&arc_user_evicts_lock);
2608 (void) remove_reference(hdr, NULL, tag);
2609 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2610 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
2611 mutex_exit(&arc_user_evicts_lock);
2613 arc_hdr_destroy(hdr);
2615 if (remove_reference(hdr, NULL, tag) > 0)
2616 arc_buf_destroy(buf, TRUE);
2618 arc_hdr_destroy(hdr);
2623 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
2625 arc_buf_hdr_t *hdr = buf->b_hdr;
2626 kmutex_t *hash_lock = HDR_LOCK(hdr);
2627 boolean_t no_callback = (buf->b_efunc == NULL);
2629 if (hdr->b_l1hdr.b_state == arc_anon) {
2630 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
2631 arc_buf_free(buf, tag);
2632 return (no_callback);
2635 mutex_enter(hash_lock);
2637 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
2638 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2639 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2640 ASSERT(buf->b_data != NULL);
2642 (void) remove_reference(hdr, hash_lock, tag);
2643 if (hdr->b_l1hdr.b_datacnt > 1) {
2645 arc_buf_destroy(buf, TRUE);
2646 } else if (no_callback) {
2647 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL);
2648 ASSERT(buf->b_efunc == NULL);
2649 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2651 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 ||
2652 refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2653 mutex_exit(hash_lock);
2654 return (no_callback);
2658 arc_buf_size(arc_buf_t *buf)
2660 return (buf->b_hdr->b_size);
2664 * Called from the DMU to determine if the current buffer should be
2665 * evicted. In order to ensure proper locking, the eviction must be initiated
2666 * from the DMU. Return true if the buffer is associated with user data and
2667 * duplicate buffers still exist.
2670 arc_buf_eviction_needed(arc_buf_t *buf)
2673 boolean_t evict_needed = B_FALSE;
2675 if (zfs_disable_dup_eviction)
2678 mutex_enter(&buf->b_evict_lock);
2682 * We are in arc_do_user_evicts(); let that function
2683 * perform the eviction.
2685 ASSERT(buf->b_data == NULL);
2686 mutex_exit(&buf->b_evict_lock);
2688 } else if (buf->b_data == NULL) {
2690 * We have already been added to the arc eviction list;
2691 * recommend eviction.
2693 ASSERT3P(hdr, ==, &arc_eviction_hdr);
2694 mutex_exit(&buf->b_evict_lock);
2698 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr))
2699 evict_needed = B_TRUE;
2701 mutex_exit(&buf->b_evict_lock);
2702 return (evict_needed);
2706 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2707 * state of the header is dependent on it's state prior to entering this
2708 * function. The following transitions are possible:
2710 * - arc_mru -> arc_mru_ghost
2711 * - arc_mfu -> arc_mfu_ghost
2712 * - arc_mru_ghost -> arc_l2c_only
2713 * - arc_mru_ghost -> deleted
2714 * - arc_mfu_ghost -> arc_l2c_only
2715 * - arc_mfu_ghost -> deleted
2718 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
2720 arc_state_t *evicted_state, *state;
2721 int64_t bytes_evicted = 0;
2723 ASSERT(MUTEX_HELD(hash_lock));
2724 ASSERT(HDR_HAS_L1HDR(hdr));
2726 state = hdr->b_l1hdr.b_state;
2727 if (GHOST_STATE(state)) {
2728 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2729 ASSERT(hdr->b_l1hdr.b_buf == NULL);
2732 * l2arc_write_buffers() relies on a header's L1 portion
2733 * (i.e. it's b_tmp_cdata field) during it's write phase.
2734 * Thus, we cannot push a header onto the arc_l2c_only
2735 * state (removing it's L1 piece) until the header is
2736 * done being written to the l2arc.
2738 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
2739 ARCSTAT_BUMP(arcstat_evict_l2_skip);
2740 return (bytes_evicted);
2743 ARCSTAT_BUMP(arcstat_deleted);
2744 bytes_evicted += hdr->b_size;
2746 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
2748 if (HDR_HAS_L2HDR(hdr)) {
2750 * This buffer is cached on the 2nd Level ARC;
2751 * don't destroy the header.
2753 arc_change_state(arc_l2c_only, hdr, hash_lock);
2755 * dropping from L1+L2 cached to L2-only,
2756 * realloc to remove the L1 header.
2758 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
2761 arc_change_state(arc_anon, hdr, hash_lock);
2762 arc_hdr_destroy(hdr);
2764 return (bytes_evicted);
2767 ASSERT(state == arc_mru || state == arc_mfu);
2768 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2770 /* prefetch buffers have a minimum lifespan */
2771 if (HDR_IO_IN_PROGRESS(hdr) ||
2772 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
2773 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
2774 arc_min_prefetch_lifespan)) {
2775 ARCSTAT_BUMP(arcstat_evict_skip);
2776 return (bytes_evicted);
2779 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
2780 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0);
2781 while (hdr->b_l1hdr.b_buf) {
2782 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2783 if (!mutex_tryenter(&buf->b_evict_lock)) {
2784 ARCSTAT_BUMP(arcstat_mutex_miss);
2787 if (buf->b_data != NULL)
2788 bytes_evicted += hdr->b_size;
2789 if (buf->b_efunc != NULL) {
2790 mutex_enter(&arc_user_evicts_lock);
2791 arc_buf_destroy(buf, FALSE);
2792 hdr->b_l1hdr.b_buf = buf->b_next;
2793 buf->b_hdr = &arc_eviction_hdr;
2794 buf->b_next = arc_eviction_list;
2795 arc_eviction_list = buf;
2796 cv_signal(&arc_user_evicts_cv);
2797 mutex_exit(&arc_user_evicts_lock);
2798 mutex_exit(&buf->b_evict_lock);
2800 mutex_exit(&buf->b_evict_lock);
2801 arc_buf_destroy(buf, TRUE);
2805 if (HDR_HAS_L2HDR(hdr)) {
2806 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size);
2808 if (l2arc_write_eligible(hdr->b_spa, hdr))
2809 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size);
2811 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size);
2814 if (hdr->b_l1hdr.b_datacnt == 0) {
2815 arc_change_state(evicted_state, hdr, hash_lock);
2816 ASSERT(HDR_IN_HASH_TABLE(hdr));
2817 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
2818 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
2819 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
2822 return (bytes_evicted);
2826 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
2827 uint64_t spa, int64_t bytes)
2829 multilist_sublist_t *mls;
2830 uint64_t bytes_evicted = 0;
2832 kmutex_t *hash_lock;
2833 int evict_count = 0;
2835 ASSERT3P(marker, !=, NULL);
2836 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2838 mls = multilist_sublist_lock(ml, idx);
2840 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
2841 hdr = multilist_sublist_prev(mls, marker)) {
2842 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
2843 (evict_count >= zfs_arc_evict_batch_limit))
2847 * To keep our iteration location, move the marker
2848 * forward. Since we're not holding hdr's hash lock, we
2849 * must be very careful and not remove 'hdr' from the
2850 * sublist. Otherwise, other consumers might mistake the
2851 * 'hdr' as not being on a sublist when they call the
2852 * multilist_link_active() function (they all rely on
2853 * the hash lock protecting concurrent insertions and
2854 * removals). multilist_sublist_move_forward() was
2855 * specifically implemented to ensure this is the case
2856 * (only 'marker' will be removed and re-inserted).
2858 multilist_sublist_move_forward(mls, marker);
2861 * The only case where the b_spa field should ever be
2862 * zero, is the marker headers inserted by
2863 * arc_evict_state(). It's possible for multiple threads
2864 * to be calling arc_evict_state() concurrently (e.g.
2865 * dsl_pool_close() and zio_inject_fault()), so we must
2866 * skip any markers we see from these other threads.
2868 if (hdr->b_spa == 0)
2871 /* we're only interested in evicting buffers of a certain spa */
2872 if (spa != 0 && hdr->b_spa != spa) {
2873 ARCSTAT_BUMP(arcstat_evict_skip);
2877 hash_lock = HDR_LOCK(hdr);
2880 * We aren't calling this function from any code path
2881 * that would already be holding a hash lock, so we're
2882 * asserting on this assumption to be defensive in case
2883 * this ever changes. Without this check, it would be
2884 * possible to incorrectly increment arcstat_mutex_miss
2885 * below (e.g. if the code changed such that we called
2886 * this function with a hash lock held).
2888 ASSERT(!MUTEX_HELD(hash_lock));
2890 if (mutex_tryenter(hash_lock)) {
2891 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
2892 mutex_exit(hash_lock);
2894 bytes_evicted += evicted;
2897 * If evicted is zero, arc_evict_hdr() must have
2898 * decided to skip this header, don't increment
2899 * evict_count in this case.
2905 * If arc_size isn't overflowing, signal any
2906 * threads that might happen to be waiting.
2908 * For each header evicted, we wake up a single
2909 * thread. If we used cv_broadcast, we could
2910 * wake up "too many" threads causing arc_size
2911 * to significantly overflow arc_c; since
2912 * arc_get_data_buf() doesn't check for overflow
2913 * when it's woken up (it doesn't because it's
2914 * possible for the ARC to be overflowing while
2915 * full of un-evictable buffers, and the
2916 * function should proceed in this case).
2918 * If threads are left sleeping, due to not
2919 * using cv_broadcast, they will be woken up
2920 * just before arc_reclaim_thread() sleeps.
2922 mutex_enter(&arc_reclaim_lock);
2923 if (!arc_is_overflowing())
2924 cv_signal(&arc_reclaim_waiters_cv);
2925 mutex_exit(&arc_reclaim_lock);
2927 ARCSTAT_BUMP(arcstat_mutex_miss);
2931 multilist_sublist_unlock(mls);
2933 return (bytes_evicted);
2937 * Evict buffers from the given arc state, until we've removed the
2938 * specified number of bytes. Move the removed buffers to the
2939 * appropriate evict state.
2941 * This function makes a "best effort". It skips over any buffers
2942 * it can't get a hash_lock on, and so, may not catch all candidates.
2943 * It may also return without evicting as much space as requested.
2945 * If bytes is specified using the special value ARC_EVICT_ALL, this
2946 * will evict all available (i.e. unlocked and evictable) buffers from
2947 * the given arc state; which is used by arc_flush().
2950 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
2951 arc_buf_contents_t type)
2953 uint64_t total_evicted = 0;
2954 multilist_t *ml = &state->arcs_list[type];
2956 arc_buf_hdr_t **markers;
2958 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2960 num_sublists = multilist_get_num_sublists(ml);
2963 * If we've tried to evict from each sublist, made some
2964 * progress, but still have not hit the target number of bytes
2965 * to evict, we want to keep trying. The markers allow us to
2966 * pick up where we left off for each individual sublist, rather
2967 * than starting from the tail each time.
2969 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
2970 for (int i = 0; i < num_sublists; i++) {
2971 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
2974 * A b_spa of 0 is used to indicate that this header is
2975 * a marker. This fact is used in arc_adjust_type() and
2976 * arc_evict_state_impl().
2978 markers[i]->b_spa = 0;
2980 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2981 multilist_sublist_insert_tail(mls, markers[i]);
2982 multilist_sublist_unlock(mls);
2986 * While we haven't hit our target number of bytes to evict, or
2987 * we're evicting all available buffers.
2989 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
2991 * Start eviction using a randomly selected sublist,
2992 * this is to try and evenly balance eviction across all
2993 * sublists. Always starting at the same sublist
2994 * (e.g. index 0) would cause evictions to favor certain
2995 * sublists over others.
2997 int sublist_idx = multilist_get_random_index(ml);
2998 uint64_t scan_evicted = 0;
3000 for (int i = 0; i < num_sublists; i++) {
3001 uint64_t bytes_remaining;
3002 uint64_t bytes_evicted;
3004 if (bytes == ARC_EVICT_ALL)
3005 bytes_remaining = ARC_EVICT_ALL;
3006 else if (total_evicted < bytes)
3007 bytes_remaining = bytes - total_evicted;
3011 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3012 markers[sublist_idx], spa, bytes_remaining);
3014 scan_evicted += bytes_evicted;
3015 total_evicted += bytes_evicted;
3017 /* we've reached the end, wrap to the beginning */
3018 if (++sublist_idx >= num_sublists)
3023 * If we didn't evict anything during this scan, we have
3024 * no reason to believe we'll evict more during another
3025 * scan, so break the loop.
3027 if (scan_evicted == 0) {
3028 /* This isn't possible, let's make that obvious */
3029 ASSERT3S(bytes, !=, 0);
3032 * When bytes is ARC_EVICT_ALL, the only way to
3033 * break the loop is when scan_evicted is zero.
3034 * In that case, we actually have evicted enough,
3035 * so we don't want to increment the kstat.
3037 if (bytes != ARC_EVICT_ALL) {
3038 ASSERT3S(total_evicted, <, bytes);
3039 ARCSTAT_BUMP(arcstat_evict_not_enough);
3046 for (int i = 0; i < num_sublists; i++) {
3047 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3048 multilist_sublist_remove(mls, markers[i]);
3049 multilist_sublist_unlock(mls);
3051 kmem_cache_free(hdr_full_cache, markers[i]);
3053 kmem_free(markers, sizeof (*markers) * num_sublists);
3055 return (total_evicted);
3059 * Flush all "evictable" data of the given type from the arc state
3060 * specified. This will not evict any "active" buffers (i.e. referenced).
3062 * When 'retry' is set to FALSE, the function will make a single pass
3063 * over the state and evict any buffers that it can. Since it doesn't
3064 * continually retry the eviction, it might end up leaving some buffers
3065 * in the ARC due to lock misses.
3067 * When 'retry' is set to TRUE, the function will continually retry the
3068 * eviction until *all* evictable buffers have been removed from the
3069 * state. As a result, if concurrent insertions into the state are
3070 * allowed (e.g. if the ARC isn't shutting down), this function might
3071 * wind up in an infinite loop, continually trying to evict buffers.
3074 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3077 uint64_t evicted = 0;
3079 while (state->arcs_lsize[type] != 0) {
3080 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3090 * Evict the specified number of bytes from the state specified,
3091 * restricting eviction to the spa and type given. This function
3092 * prevents us from trying to evict more from a state's list than
3093 * is "evictable", and to skip evicting altogether when passed a
3094 * negative value for "bytes". In contrast, arc_evict_state() will
3095 * evict everything it can, when passed a negative value for "bytes".
3098 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3099 arc_buf_contents_t type)
3103 if (bytes > 0 && state->arcs_lsize[type] > 0) {
3104 delta = MIN(state->arcs_lsize[type], bytes);
3105 return (arc_evict_state(state, spa, delta, type));
3112 * Evict metadata buffers from the cache, such that arc_meta_used is
3113 * capped by the arc_meta_limit tunable.
3116 arc_adjust_meta(void)
3118 uint64_t total_evicted = 0;
3122 * If we're over the meta limit, we want to evict enough
3123 * metadata to get back under the meta limit. We don't want to
3124 * evict so much that we drop the MRU below arc_p, though. If
3125 * we're over the meta limit more than we're over arc_p, we
3126 * evict some from the MRU here, and some from the MFU below.
3128 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3129 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3130 refcount_count(&arc_mru->arcs_size) - arc_p));
3132 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3135 * Similar to the above, we want to evict enough bytes to get us
3136 * below the meta limit, but not so much as to drop us below the
3137 * space alloted to the MFU (which is defined as arc_c - arc_p).
3139 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3140 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
3142 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3144 return (total_evicted);
3148 * Return the type of the oldest buffer in the given arc state
3150 * This function will select a random sublist of type ARC_BUFC_DATA and
3151 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3152 * is compared, and the type which contains the "older" buffer will be
3155 static arc_buf_contents_t
3156 arc_adjust_type(arc_state_t *state)
3158 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
3159 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
3160 int data_idx = multilist_get_random_index(data_ml);
3161 int meta_idx = multilist_get_random_index(meta_ml);
3162 multilist_sublist_t *data_mls;
3163 multilist_sublist_t *meta_mls;
3164 arc_buf_contents_t type;
3165 arc_buf_hdr_t *data_hdr;
3166 arc_buf_hdr_t *meta_hdr;
3169 * We keep the sublist lock until we're finished, to prevent
3170 * the headers from being destroyed via arc_evict_state().
3172 data_mls = multilist_sublist_lock(data_ml, data_idx);
3173 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3176 * These two loops are to ensure we skip any markers that
3177 * might be at the tail of the lists due to arc_evict_state().
3180 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3181 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3182 if (data_hdr->b_spa != 0)
3186 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3187 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3188 if (meta_hdr->b_spa != 0)
3192 if (data_hdr == NULL && meta_hdr == NULL) {
3193 type = ARC_BUFC_DATA;
3194 } else if (data_hdr == NULL) {
3195 ASSERT3P(meta_hdr, !=, NULL);
3196 type = ARC_BUFC_METADATA;
3197 } else if (meta_hdr == NULL) {
3198 ASSERT3P(data_hdr, !=, NULL);
3199 type = ARC_BUFC_DATA;
3201 ASSERT3P(data_hdr, !=, NULL);
3202 ASSERT3P(meta_hdr, !=, NULL);
3204 /* The headers can't be on the sublist without an L1 header */
3205 ASSERT(HDR_HAS_L1HDR(data_hdr));
3206 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3208 if (data_hdr->b_l1hdr.b_arc_access <
3209 meta_hdr->b_l1hdr.b_arc_access) {
3210 type = ARC_BUFC_DATA;
3212 type = ARC_BUFC_METADATA;
3216 multilist_sublist_unlock(meta_mls);
3217 multilist_sublist_unlock(data_mls);
3223 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3228 uint64_t total_evicted = 0;
3233 * If we're over arc_meta_limit, we want to correct that before
3234 * potentially evicting data buffers below.
3236 total_evicted += arc_adjust_meta();
3241 * If we're over the target cache size, we want to evict enough
3242 * from the list to get back to our target size. We don't want
3243 * to evict too much from the MRU, such that it drops below
3244 * arc_p. So, if we're over our target cache size more than
3245 * the MRU is over arc_p, we'll evict enough to get back to
3246 * arc_p here, and then evict more from the MFU below.
3248 target = MIN((int64_t)(arc_size - arc_c),
3249 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3250 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
3253 * If we're below arc_meta_min, always prefer to evict data.
3254 * Otherwise, try to satisfy the requested number of bytes to
3255 * evict from the type which contains older buffers; in an
3256 * effort to keep newer buffers in the cache regardless of their
3257 * type. If we cannot satisfy the number of bytes from this
3258 * type, spill over into the next type.
3260 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
3261 arc_meta_used > arc_meta_min) {
3262 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3263 total_evicted += bytes;
3266 * If we couldn't evict our target number of bytes from
3267 * metadata, we try to get the rest from data.
3272 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3274 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3275 total_evicted += bytes;
3278 * If we couldn't evict our target number of bytes from
3279 * data, we try to get the rest from metadata.
3284 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3290 * Now that we've tried to evict enough from the MRU to get its
3291 * size back to arc_p, if we're still above the target cache
3292 * size, we evict the rest from the MFU.
3294 target = arc_size - arc_c;
3296 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
3297 arc_meta_used > arc_meta_min) {
3298 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3299 total_evicted += bytes;
3302 * If we couldn't evict our target number of bytes from
3303 * metadata, we try to get the rest from data.
3308 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3310 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3311 total_evicted += bytes;
3314 * If we couldn't evict our target number of bytes from
3315 * data, we try to get the rest from data.
3320 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3324 * Adjust ghost lists
3326 * In addition to the above, the ARC also defines target values
3327 * for the ghost lists. The sum of the mru list and mru ghost
3328 * list should never exceed the target size of the cache, and
3329 * the sum of the mru list, mfu list, mru ghost list, and mfu
3330 * ghost list should never exceed twice the target size of the
3331 * cache. The following logic enforces these limits on the ghost
3332 * caches, and evicts from them as needed.
3334 target = refcount_count(&arc_mru->arcs_size) +
3335 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3337 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3338 total_evicted += bytes;
3343 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3346 * We assume the sum of the mru list and mfu list is less than
3347 * or equal to arc_c (we enforced this above), which means we
3348 * can use the simpler of the two equations below:
3350 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3351 * mru ghost + mfu ghost <= arc_c
3353 target = refcount_count(&arc_mru_ghost->arcs_size) +
3354 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3356 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3357 total_evicted += bytes;
3362 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3364 return (total_evicted);
3368 arc_do_user_evicts(void)
3370 mutex_enter(&arc_user_evicts_lock);
3371 while (arc_eviction_list != NULL) {
3372 arc_buf_t *buf = arc_eviction_list;
3373 arc_eviction_list = buf->b_next;
3374 mutex_enter(&buf->b_evict_lock);
3376 mutex_exit(&buf->b_evict_lock);
3377 mutex_exit(&arc_user_evicts_lock);
3379 if (buf->b_efunc != NULL)
3380 VERIFY0(buf->b_efunc(buf->b_private));
3382 buf->b_efunc = NULL;
3383 buf->b_private = NULL;
3384 kmem_cache_free(buf_cache, buf);
3385 mutex_enter(&arc_user_evicts_lock);
3387 mutex_exit(&arc_user_evicts_lock);
3391 arc_flush(spa_t *spa, boolean_t retry)
3396 * If retry is TRUE, a spa must not be specified since we have
3397 * no good way to determine if all of a spa's buffers have been
3398 * evicted from an arc state.
3400 ASSERT(!retry || spa == 0);
3403 guid = spa_load_guid(spa);
3405 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3406 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3408 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3409 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3411 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3412 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3414 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3415 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3417 arc_do_user_evicts();
3418 ASSERT(spa || arc_eviction_list == NULL);
3422 arc_shrink(int64_t to_free)
3424 if (arc_c > arc_c_min) {
3425 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
3426 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
3427 if (arc_c > arc_c_min + to_free)
3428 atomic_add_64(&arc_c, -to_free);
3432 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3433 if (arc_c > arc_size)
3434 arc_c = MAX(arc_size, arc_c_min);
3436 arc_p = (arc_c >> 1);
3438 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
3441 ASSERT(arc_c >= arc_c_min);
3442 ASSERT((int64_t)arc_p >= 0);
3445 if (arc_size > arc_c) {
3446 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size,
3448 (void) arc_adjust();
3452 static long needfree = 0;
3454 typedef enum free_memory_reason_t {
3459 FMR_PAGES_PP_MAXIMUM,
3463 } free_memory_reason_t;
3465 int64_t last_free_memory;
3466 free_memory_reason_t last_free_reason;
3469 * Additional reserve of pages for pp_reserve.
3471 int64_t arc_pages_pp_reserve = 64;
3474 * Additional reserve of pages for swapfs.
3476 int64_t arc_swapfs_reserve = 64;
3479 * Return the amount of memory that can be consumed before reclaim will be
3480 * needed. Positive if there is sufficient free memory, negative indicates
3481 * the amount of memory that needs to be freed up.
3484 arc_available_memory(void)
3486 int64_t lowest = INT64_MAX;
3488 free_memory_reason_t r = FMR_UNKNOWN;
3492 n = PAGESIZE * (-needfree);
3500 * Cooperate with pagedaemon when it's time for it to scan
3501 * and reclaim some pages.
3503 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
3511 * check that we're out of range of the pageout scanner. It starts to
3512 * schedule paging if freemem is less than lotsfree and needfree.
3513 * lotsfree is the high-water mark for pageout, and needfree is the
3514 * number of needed free pages. We add extra pages here to make sure
3515 * the scanner doesn't start up while we're freeing memory.
3517 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3524 * check to make sure that swapfs has enough space so that anon
3525 * reservations can still succeed. anon_resvmem() checks that the
3526 * availrmem is greater than swapfs_minfree, and the number of reserved
3527 * swap pages. We also add a bit of extra here just to prevent
3528 * circumstances from getting really dire.
3530 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3531 desfree - arc_swapfs_reserve);
3534 r = FMR_SWAPFS_MINFREE;
3539 * Check that we have enough availrmem that memory locking (e.g., via
3540 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3541 * stores the number of pages that cannot be locked; when availrmem
3542 * drops below pages_pp_maximum, page locking mechanisms such as
3543 * page_pp_lock() will fail.)
3545 n = PAGESIZE * (availrmem - pages_pp_maximum -
3546 arc_pages_pp_reserve);
3549 r = FMR_PAGES_PP_MAXIMUM;
3552 #endif /* illumos */
3553 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
3555 * If we're on an i386 platform, it's possible that we'll exhaust the
3556 * kernel heap space before we ever run out of available physical
3557 * memory. Most checks of the size of the heap_area compare against
3558 * tune.t_minarmem, which is the minimum available real memory that we
3559 * can have in the system. However, this is generally fixed at 25 pages
3560 * which is so low that it's useless. In this comparison, we seek to
3561 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3562 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3565 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
3566 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3571 #define zio_arena NULL
3573 #define zio_arena heap_arena
3577 * If zio data pages are being allocated out of a separate heap segment,
3578 * then enforce that the size of available vmem for this arena remains
3579 * above about 1/16th free.
3581 * Note: The 1/16th arena free requirement was put in place
3582 * to aggressively evict memory from the arc in order to avoid
3583 * memory fragmentation issues.
3585 if (zio_arena != NULL) {
3586 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
3587 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3595 * Above limits know nothing about real level of KVA fragmentation.
3596 * Start aggressive reclamation if too little sequential KVA left.
3599 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ?
3600 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) :
3609 /* Every 100 calls, free a small amount */
3610 if (spa_get_random(100) == 0)
3612 #endif /* _KERNEL */
3614 last_free_memory = lowest;
3615 last_free_reason = r;
3616 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
3622 * Determine if the system is under memory pressure and is asking
3623 * to reclaim memory. A return value of TRUE indicates that the system
3624 * is under memory pressure and that the arc should adjust accordingly.
3627 arc_reclaim_needed(void)
3629 return (arc_available_memory() < 0);
3632 extern kmem_cache_t *zio_buf_cache[];
3633 extern kmem_cache_t *zio_data_buf_cache[];
3634 extern kmem_cache_t *range_seg_cache;
3636 static __noinline void
3637 arc_kmem_reap_now(void)
3640 kmem_cache_t *prev_cache = NULL;
3641 kmem_cache_t *prev_data_cache = NULL;
3643 DTRACE_PROBE(arc__kmem_reap_start);
3645 if (arc_meta_used >= arc_meta_limit) {
3647 * We are exceeding our meta-data cache limit.
3648 * Purge some DNLC entries to release holds on meta-data.
3650 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
3654 * Reclaim unused memory from all kmem caches.
3660 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
3661 if (zio_buf_cache[i] != prev_cache) {
3662 prev_cache = zio_buf_cache[i];
3663 kmem_cache_reap_now(zio_buf_cache[i]);
3665 if (zio_data_buf_cache[i] != prev_data_cache) {
3666 prev_data_cache = zio_data_buf_cache[i];
3667 kmem_cache_reap_now(zio_data_buf_cache[i]);
3670 kmem_cache_reap_now(buf_cache);
3671 kmem_cache_reap_now(hdr_full_cache);
3672 kmem_cache_reap_now(hdr_l2only_cache);
3673 kmem_cache_reap_now(range_seg_cache);
3676 if (zio_arena != NULL) {
3678 * Ask the vmem arena to reclaim unused memory from its
3681 vmem_qcache_reap(zio_arena);
3684 DTRACE_PROBE(arc__kmem_reap_end);
3688 * Threads can block in arc_get_data_buf() waiting for this thread to evict
3689 * enough data and signal them to proceed. When this happens, the threads in
3690 * arc_get_data_buf() are sleeping while holding the hash lock for their
3691 * particular arc header. Thus, we must be careful to never sleep on a
3692 * hash lock in this thread. This is to prevent the following deadlock:
3694 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3695 * waiting for the reclaim thread to signal it.
3697 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3698 * fails, and goes to sleep forever.
3700 * This possible deadlock is avoided by always acquiring a hash lock
3701 * using mutex_tryenter() from arc_reclaim_thread().
3704 arc_reclaim_thread(void *dummy __unused)
3706 hrtime_t growtime = 0;
3709 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
3711 mutex_enter(&arc_reclaim_lock);
3712 while (!arc_reclaim_thread_exit) {
3713 int64_t free_memory = arc_available_memory();
3714 uint64_t evicted = 0;
3716 mutex_exit(&arc_reclaim_lock);
3718 if (free_memory < 0) {
3720 arc_no_grow = B_TRUE;
3724 * Wait at least zfs_grow_retry (default 60) seconds
3725 * before considering growing.
3727 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
3729 arc_kmem_reap_now();
3732 * If we are still low on memory, shrink the ARC
3733 * so that we have arc_shrink_min free space.
3735 free_memory = arc_available_memory();
3738 (arc_c >> arc_shrink_shift) - free_memory;
3741 to_free = MAX(to_free, ptob(needfree));
3743 arc_shrink(to_free);
3745 } else if (free_memory < arc_c >> arc_no_grow_shift) {
3746 arc_no_grow = B_TRUE;
3747 } else if (gethrtime() >= growtime) {
3748 arc_no_grow = B_FALSE;
3751 evicted = arc_adjust();
3753 mutex_enter(&arc_reclaim_lock);
3756 * If evicted is zero, we couldn't evict anything via
3757 * arc_adjust(). This could be due to hash lock
3758 * collisions, but more likely due to the majority of
3759 * arc buffers being unevictable. Therefore, even if
3760 * arc_size is above arc_c, another pass is unlikely to
3761 * be helpful and could potentially cause us to enter an
3764 if (arc_size <= arc_c || evicted == 0) {
3769 * We're either no longer overflowing, or we
3770 * can't evict anything more, so we should wake
3771 * up any threads before we go to sleep.
3773 cv_broadcast(&arc_reclaim_waiters_cv);
3776 * Block until signaled, or after one second (we
3777 * might need to perform arc_kmem_reap_now()
3778 * even if we aren't being signalled)
3780 CALLB_CPR_SAFE_BEGIN(&cpr);
3781 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
3782 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
3783 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
3787 arc_reclaim_thread_exit = FALSE;
3788 cv_broadcast(&arc_reclaim_thread_cv);
3789 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
3794 arc_user_evicts_thread(void *dummy __unused)
3798 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG);
3800 mutex_enter(&arc_user_evicts_lock);
3801 while (!arc_user_evicts_thread_exit) {
3802 mutex_exit(&arc_user_evicts_lock);
3804 arc_do_user_evicts();
3807 * This is necessary in order for the mdb ::arc dcmd to
3808 * show up to date information. Since the ::arc command
3809 * does not call the kstat's update function, without
3810 * this call, the command may show stale stats for the
3811 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3812 * with this change, the data might be up to 1 second
3813 * out of date; but that should suffice. The arc_state_t
3814 * structures can be queried directly if more accurate
3815 * information is needed.
3817 if (arc_ksp != NULL)
3818 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
3820 mutex_enter(&arc_user_evicts_lock);
3823 * Block until signaled, or after one second (we need to
3824 * call the arc's kstat update function regularly).
3826 CALLB_CPR_SAFE_BEGIN(&cpr);
3827 (void) cv_timedwait(&arc_user_evicts_cv,
3828 &arc_user_evicts_lock, hz);
3829 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock);
3832 arc_user_evicts_thread_exit = FALSE;
3833 cv_broadcast(&arc_user_evicts_cv);
3834 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */
3838 static u_int arc_dnlc_evicts_arg;
3839 extern struct vfsops zfs_vfsops;
3842 arc_dnlc_evicts_thread(void *dummy __unused)
3847 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG);
3849 mutex_enter(&arc_dnlc_evicts_lock);
3850 while (!arc_dnlc_evicts_thread_exit) {
3851 CALLB_CPR_SAFE_BEGIN(&cpr);
3852 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
3853 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock);
3854 if (arc_dnlc_evicts_arg != 0) {
3855 percent = arc_dnlc_evicts_arg;
3856 mutex_exit(&arc_dnlc_evicts_lock);
3858 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops);
3860 mutex_enter(&arc_dnlc_evicts_lock);
3862 * Clear our token only after vnlru_free()
3863 * pass is done, to avoid false queueing of
3866 arc_dnlc_evicts_arg = 0;
3869 arc_dnlc_evicts_thread_exit = FALSE;
3870 cv_broadcast(&arc_dnlc_evicts_cv);
3871 CALLB_CPR_EXIT(&cpr);
3876 dnlc_reduce_cache(void *arg)
3880 percent = (u_int)(uintptr_t)arg;
3881 mutex_enter(&arc_dnlc_evicts_lock);
3882 if (arc_dnlc_evicts_arg == 0) {
3883 arc_dnlc_evicts_arg = percent;
3884 cv_broadcast(&arc_dnlc_evicts_cv);
3886 mutex_exit(&arc_dnlc_evicts_lock);
3890 * Adapt arc info given the number of bytes we are trying to add and
3891 * the state that we are comming from. This function is only called
3892 * when we are adding new content to the cache.
3895 arc_adapt(int bytes, arc_state_t *state)
3898 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
3899 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
3900 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
3902 if (state == arc_l2c_only)
3907 * Adapt the target size of the MRU list:
3908 * - if we just hit in the MRU ghost list, then increase
3909 * the target size of the MRU list.
3910 * - if we just hit in the MFU ghost list, then increase
3911 * the target size of the MFU list by decreasing the
3912 * target size of the MRU list.
3914 if (state == arc_mru_ghost) {
3915 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
3916 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
3918 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
3919 } else if (state == arc_mfu_ghost) {
3922 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
3923 mult = MIN(mult, 10);
3925 delta = MIN(bytes * mult, arc_p);
3926 arc_p = MAX(arc_p_min, arc_p - delta);
3928 ASSERT((int64_t)arc_p >= 0);
3930 if (arc_reclaim_needed()) {
3931 cv_signal(&arc_reclaim_thread_cv);
3938 if (arc_c >= arc_c_max)
3942 * If we're within (2 * maxblocksize) bytes of the target
3943 * cache size, increment the target cache size
3945 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
3946 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
3947 atomic_add_64(&arc_c, (int64_t)bytes);
3948 if (arc_c > arc_c_max)
3950 else if (state == arc_anon)
3951 atomic_add_64(&arc_p, (int64_t)bytes);
3955 ASSERT((int64_t)arc_p >= 0);
3959 * Check if arc_size has grown past our upper threshold, determined by
3960 * zfs_arc_overflow_shift.
3963 arc_is_overflowing(void)
3965 /* Always allow at least one block of overflow */
3966 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
3967 arc_c >> zfs_arc_overflow_shift);
3969 return (arc_size >= arc_c + overflow);
3973 * The buffer, supplied as the first argument, needs a data block. If we
3974 * are hitting the hard limit for the cache size, we must sleep, waiting
3975 * for the eviction thread to catch up. If we're past the target size
3976 * but below the hard limit, we'll only signal the reclaim thread and
3980 arc_get_data_buf(arc_buf_t *buf)
3982 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
3983 uint64_t size = buf->b_hdr->b_size;
3984 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
3986 arc_adapt(size, state);
3989 * If arc_size is currently overflowing, and has grown past our
3990 * upper limit, we must be adding data faster than the evict
3991 * thread can evict. Thus, to ensure we don't compound the
3992 * problem by adding more data and forcing arc_size to grow even
3993 * further past it's target size, we halt and wait for the
3994 * eviction thread to catch up.
3996 * It's also possible that the reclaim thread is unable to evict
3997 * enough buffers to get arc_size below the overflow limit (e.g.
3998 * due to buffers being un-evictable, or hash lock collisions).
3999 * In this case, we want to proceed regardless if we're
4000 * overflowing; thus we don't use a while loop here.
4002 if (arc_is_overflowing()) {
4003 mutex_enter(&arc_reclaim_lock);
4006 * Now that we've acquired the lock, we may no longer be
4007 * over the overflow limit, lets check.
4009 * We're ignoring the case of spurious wake ups. If that
4010 * were to happen, it'd let this thread consume an ARC
4011 * buffer before it should have (i.e. before we're under
4012 * the overflow limit and were signalled by the reclaim
4013 * thread). As long as that is a rare occurrence, it
4014 * shouldn't cause any harm.
4016 if (arc_is_overflowing()) {
4017 cv_signal(&arc_reclaim_thread_cv);
4018 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4021 mutex_exit(&arc_reclaim_lock);
4024 if (type == ARC_BUFC_METADATA) {
4025 buf->b_data = zio_buf_alloc(size);
4026 arc_space_consume(size, ARC_SPACE_META);
4028 ASSERT(type == ARC_BUFC_DATA);
4029 buf->b_data = zio_data_buf_alloc(size);
4030 arc_space_consume(size, ARC_SPACE_DATA);
4034 * Update the state size. Note that ghost states have a
4035 * "ghost size" and so don't need to be updated.
4037 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) {
4038 arc_buf_hdr_t *hdr = buf->b_hdr;
4039 arc_state_t *state = hdr->b_l1hdr.b_state;
4041 (void) refcount_add_many(&state->arcs_size, size, buf);
4044 * If this is reached via arc_read, the link is
4045 * protected by the hash lock. If reached via
4046 * arc_buf_alloc, the header should not be accessed by
4047 * any other thread. And, if reached via arc_read_done,
4048 * the hash lock will protect it if it's found in the
4049 * hash table; otherwise no other thread should be
4050 * trying to [add|remove]_reference it.
4052 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4053 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4054 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type],
4058 * If we are growing the cache, and we are adding anonymous
4059 * data, and we have outgrown arc_p, update arc_p
4061 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
4062 (refcount_count(&arc_anon->arcs_size) +
4063 refcount_count(&arc_mru->arcs_size) > arc_p))
4064 arc_p = MIN(arc_c, arc_p + size);
4066 ARCSTAT_BUMP(arcstat_allocated);
4070 * This routine is called whenever a buffer is accessed.
4071 * NOTE: the hash lock is dropped in this function.
4074 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4078 ASSERT(MUTEX_HELD(hash_lock));
4079 ASSERT(HDR_HAS_L1HDR(hdr));
4081 if (hdr->b_l1hdr.b_state == arc_anon) {
4083 * This buffer is not in the cache, and does not
4084 * appear in our "ghost" list. Add the new buffer
4088 ASSERT0(hdr->b_l1hdr.b_arc_access);
4089 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4090 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4091 arc_change_state(arc_mru, hdr, hash_lock);
4093 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4094 now = ddi_get_lbolt();
4097 * If this buffer is here because of a prefetch, then either:
4098 * - clear the flag if this is a "referencing" read
4099 * (any subsequent access will bump this into the MFU state).
4101 * - move the buffer to the head of the list if this is
4102 * another prefetch (to make it less likely to be evicted).
4104 if (HDR_PREFETCH(hdr)) {
4105 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4106 /* link protected by hash lock */
4107 ASSERT(multilist_link_active(
4108 &hdr->b_l1hdr.b_arc_node));
4110 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
4111 ARCSTAT_BUMP(arcstat_mru_hits);
4113 hdr->b_l1hdr.b_arc_access = now;
4118 * This buffer has been "accessed" only once so far,
4119 * but it is still in the cache. Move it to the MFU
4122 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4124 * More than 125ms have passed since we
4125 * instantiated this buffer. Move it to the
4126 * most frequently used state.
4128 hdr->b_l1hdr.b_arc_access = now;
4129 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4130 arc_change_state(arc_mfu, hdr, hash_lock);
4132 ARCSTAT_BUMP(arcstat_mru_hits);
4133 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
4134 arc_state_t *new_state;
4136 * This buffer has been "accessed" recently, but
4137 * was evicted from the cache. Move it to the
4141 if (HDR_PREFETCH(hdr)) {
4142 new_state = arc_mru;
4143 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
4144 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
4145 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4147 new_state = arc_mfu;
4148 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4151 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4152 arc_change_state(new_state, hdr, hash_lock);
4154 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
4155 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
4157 * This buffer has been accessed more than once and is
4158 * still in the cache. Keep it in the MFU state.
4160 * NOTE: an add_reference() that occurred when we did
4161 * the arc_read() will have kicked this off the list.
4162 * If it was a prefetch, we will explicitly move it to
4163 * the head of the list now.
4165 if ((HDR_PREFETCH(hdr)) != 0) {
4166 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4167 /* link protected by hash_lock */
4168 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4170 ARCSTAT_BUMP(arcstat_mfu_hits);
4171 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4172 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4173 arc_state_t *new_state = arc_mfu;
4175 * This buffer has been accessed more than once but has
4176 * been evicted from the cache. Move it back to the
4180 if (HDR_PREFETCH(hdr)) {
4182 * This is a prefetch access...
4183 * move this block back to the MRU state.
4185 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4186 new_state = arc_mru;
4189 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4190 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4191 arc_change_state(new_state, hdr, hash_lock);
4193 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4194 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4196 * This buffer is on the 2nd Level ARC.
4199 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4200 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4201 arc_change_state(arc_mfu, hdr, hash_lock);
4203 ASSERT(!"invalid arc state");
4207 /* a generic arc_done_func_t which you can use */
4210 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4212 if (zio == NULL || zio->io_error == 0)
4213 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
4214 VERIFY(arc_buf_remove_ref(buf, arg));
4217 /* a generic arc_done_func_t */
4219 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4221 arc_buf_t **bufp = arg;
4222 if (zio && zio->io_error) {
4223 VERIFY(arc_buf_remove_ref(buf, arg));
4227 ASSERT(buf->b_data);
4232 arc_read_done(zio_t *zio)
4236 arc_buf_t *abuf; /* buffer we're assigning to callback */
4237 kmutex_t *hash_lock = NULL;
4238 arc_callback_t *callback_list, *acb;
4239 int freeable = FALSE;
4241 buf = zio->io_private;
4245 * The hdr was inserted into hash-table and removed from lists
4246 * prior to starting I/O. We should find this header, since
4247 * it's in the hash table, and it should be legit since it's
4248 * not possible to evict it during the I/O. The only possible
4249 * reason for it not to be found is if we were freed during the
4252 if (HDR_IN_HASH_TABLE(hdr)) {
4253 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
4254 ASSERT3U(hdr->b_dva.dva_word[0], ==,
4255 BP_IDENTITY(zio->io_bp)->dva_word[0]);
4256 ASSERT3U(hdr->b_dva.dva_word[1], ==,
4257 BP_IDENTITY(zio->io_bp)->dva_word[1]);
4259 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
4262 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
4263 hash_lock == NULL) ||
4265 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
4266 (found == hdr && HDR_L2_READING(hdr)));
4269 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED;
4270 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
4271 hdr->b_flags &= ~ARC_FLAG_L2CACHE;
4273 /* byteswap if necessary */
4274 callback_list = hdr->b_l1hdr.b_acb;
4275 ASSERT(callback_list != NULL);
4276 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
4277 dmu_object_byteswap_t bswap =
4278 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
4279 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
4280 byteswap_uint64_array :
4281 dmu_ot_byteswap[bswap].ob_func;
4282 func(buf->b_data, hdr->b_size);
4285 arc_cksum_compute(buf, B_FALSE);
4290 if (hash_lock && zio->io_error == 0 &&
4291 hdr->b_l1hdr.b_state == arc_anon) {
4293 * Only call arc_access on anonymous buffers. This is because
4294 * if we've issued an I/O for an evicted buffer, we've already
4295 * called arc_access (to prevent any simultaneous readers from
4296 * getting confused).
4298 arc_access(hdr, hash_lock);
4301 /* create copies of the data buffer for the callers */
4303 for (acb = callback_list; acb; acb = acb->acb_next) {
4304 if (acb->acb_done) {
4306 ARCSTAT_BUMP(arcstat_duplicate_reads);
4307 abuf = arc_buf_clone(buf);
4309 acb->acb_buf = abuf;
4313 hdr->b_l1hdr.b_acb = NULL;
4314 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4315 ASSERT(!HDR_BUF_AVAILABLE(hdr));
4317 ASSERT(buf->b_efunc == NULL);
4318 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
4319 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4322 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
4323 callback_list != NULL);
4325 if (zio->io_error != 0) {
4326 hdr->b_flags |= ARC_FLAG_IO_ERROR;
4327 if (hdr->b_l1hdr.b_state != arc_anon)
4328 arc_change_state(arc_anon, hdr, hash_lock);
4329 if (HDR_IN_HASH_TABLE(hdr))
4330 buf_hash_remove(hdr);
4331 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4335 * Broadcast before we drop the hash_lock to avoid the possibility
4336 * that the hdr (and hence the cv) might be freed before we get to
4337 * the cv_broadcast().
4339 cv_broadcast(&hdr->b_l1hdr.b_cv);
4341 if (hash_lock != NULL) {
4342 mutex_exit(hash_lock);
4345 * This block was freed while we waited for the read to
4346 * complete. It has been removed from the hash table and
4347 * moved to the anonymous state (so that it won't show up
4350 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
4351 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4354 /* execute each callback and free its structure */
4355 while ((acb = callback_list) != NULL) {
4357 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
4359 if (acb->acb_zio_dummy != NULL) {
4360 acb->acb_zio_dummy->io_error = zio->io_error;
4361 zio_nowait(acb->acb_zio_dummy);
4364 callback_list = acb->acb_next;
4365 kmem_free(acb, sizeof (arc_callback_t));
4369 arc_hdr_destroy(hdr);
4373 * "Read" the block at the specified DVA (in bp) via the
4374 * cache. If the block is found in the cache, invoke the provided
4375 * callback immediately and return. Note that the `zio' parameter
4376 * in the callback will be NULL in this case, since no IO was
4377 * required. If the block is not in the cache pass the read request
4378 * on to the spa with a substitute callback function, so that the
4379 * requested block will be added to the cache.
4381 * If a read request arrives for a block that has a read in-progress,
4382 * either wait for the in-progress read to complete (and return the
4383 * results); or, if this is a read with a "done" func, add a record
4384 * to the read to invoke the "done" func when the read completes,
4385 * and return; or just return.
4387 * arc_read_done() will invoke all the requested "done" functions
4388 * for readers of this block.
4391 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
4392 void *private, zio_priority_t priority, int zio_flags,
4393 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
4395 arc_buf_hdr_t *hdr = NULL;
4396 arc_buf_t *buf = NULL;
4397 kmutex_t *hash_lock = NULL;
4399 uint64_t guid = spa_load_guid(spa);
4401 ASSERT(!BP_IS_EMBEDDED(bp) ||
4402 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
4405 if (!BP_IS_EMBEDDED(bp)) {
4407 * Embedded BP's have no DVA and require no I/O to "read".
4408 * Create an anonymous arc buf to back it.
4410 hdr = buf_hash_find(guid, bp, &hash_lock);
4413 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) {
4415 *arc_flags |= ARC_FLAG_CACHED;
4417 if (HDR_IO_IN_PROGRESS(hdr)) {
4419 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
4420 priority == ZIO_PRIORITY_SYNC_READ) {
4422 * This sync read must wait for an
4423 * in-progress async read (e.g. a predictive
4424 * prefetch). Async reads are queued
4425 * separately at the vdev_queue layer, so
4426 * this is a form of priority inversion.
4427 * Ideally, we would "inherit" the demand
4428 * i/o's priority by moving the i/o from
4429 * the async queue to the synchronous queue,
4430 * but there is currently no mechanism to do
4431 * so. Track this so that we can evaluate
4432 * the magnitude of this potential performance
4435 * Note that if the prefetch i/o is already
4436 * active (has been issued to the device),
4437 * the prefetch improved performance, because
4438 * we issued it sooner than we would have
4439 * without the prefetch.
4441 DTRACE_PROBE1(arc__sync__wait__for__async,
4442 arc_buf_hdr_t *, hdr);
4443 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
4445 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4446 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH;
4449 if (*arc_flags & ARC_FLAG_WAIT) {
4450 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4451 mutex_exit(hash_lock);
4454 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4457 arc_callback_t *acb = NULL;
4459 acb = kmem_zalloc(sizeof (arc_callback_t),
4461 acb->acb_done = done;
4462 acb->acb_private = private;
4464 acb->acb_zio_dummy = zio_null(pio,
4465 spa, NULL, NULL, NULL, zio_flags);
4467 ASSERT(acb->acb_done != NULL);
4468 acb->acb_next = hdr->b_l1hdr.b_acb;
4469 hdr->b_l1hdr.b_acb = acb;
4470 add_reference(hdr, hash_lock, private);
4471 mutex_exit(hash_lock);
4474 mutex_exit(hash_lock);
4478 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4479 hdr->b_l1hdr.b_state == arc_mfu);
4482 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4484 * This is a demand read which does not have to
4485 * wait for i/o because we did a predictive
4486 * prefetch i/o for it, which has completed.
4489 arc__demand__hit__predictive__prefetch,
4490 arc_buf_hdr_t *, hdr);
4492 arcstat_demand_hit_predictive_prefetch);
4493 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH;
4495 add_reference(hdr, hash_lock, private);
4497 * If this block is already in use, create a new
4498 * copy of the data so that we will be guaranteed
4499 * that arc_release() will always succeed.
4501 buf = hdr->b_l1hdr.b_buf;
4503 ASSERT(buf->b_data);
4504 if (HDR_BUF_AVAILABLE(hdr)) {
4505 ASSERT(buf->b_efunc == NULL);
4506 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4508 buf = arc_buf_clone(buf);
4511 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4512 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4513 hdr->b_flags |= ARC_FLAG_PREFETCH;
4515 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4516 arc_access(hdr, hash_lock);
4517 if (*arc_flags & ARC_FLAG_L2CACHE)
4518 hdr->b_flags |= ARC_FLAG_L2CACHE;
4519 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4520 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4521 mutex_exit(hash_lock);
4522 ARCSTAT_BUMP(arcstat_hits);
4523 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4524 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4525 data, metadata, hits);
4528 done(NULL, buf, private);
4530 uint64_t size = BP_GET_LSIZE(bp);
4531 arc_callback_t *acb;
4534 boolean_t devw = B_FALSE;
4535 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
4536 int32_t b_asize = 0;
4539 /* this block is not in the cache */
4540 arc_buf_hdr_t *exists = NULL;
4541 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4542 buf = arc_buf_alloc(spa, size, private, type);
4544 if (!BP_IS_EMBEDDED(bp)) {
4545 hdr->b_dva = *BP_IDENTITY(bp);
4546 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4547 exists = buf_hash_insert(hdr, &hash_lock);
4549 if (exists != NULL) {
4550 /* somebody beat us to the hash insert */
4551 mutex_exit(hash_lock);
4552 buf_discard_identity(hdr);
4553 (void) arc_buf_remove_ref(buf, private);
4554 goto top; /* restart the IO request */
4558 * If there is a callback, we pass our reference to
4559 * it; otherwise we remove our reference.
4562 (void) remove_reference(hdr, hash_lock,
4565 if (*arc_flags & ARC_FLAG_PREFETCH)
4566 hdr->b_flags |= ARC_FLAG_PREFETCH;
4567 if (*arc_flags & ARC_FLAG_L2CACHE)
4568 hdr->b_flags |= ARC_FLAG_L2CACHE;
4569 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4570 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4571 if (BP_GET_LEVEL(bp) > 0)
4572 hdr->b_flags |= ARC_FLAG_INDIRECT;
4575 * This block is in the ghost cache. If it was L2-only
4576 * (and thus didn't have an L1 hdr), we realloc the
4577 * header to add an L1 hdr.
4579 if (!HDR_HAS_L1HDR(hdr)) {
4580 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4584 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4585 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4586 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4587 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4590 * If there is a callback, we pass a reference to it.
4593 add_reference(hdr, hash_lock, private);
4594 if (*arc_flags & ARC_FLAG_PREFETCH)
4595 hdr->b_flags |= ARC_FLAG_PREFETCH;
4596 if (*arc_flags & ARC_FLAG_L2CACHE)
4597 hdr->b_flags |= ARC_FLAG_L2CACHE;
4598 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4599 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4600 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
4603 buf->b_efunc = NULL;
4604 buf->b_private = NULL;
4606 hdr->b_l1hdr.b_buf = buf;
4607 ASSERT0(hdr->b_l1hdr.b_datacnt);
4608 hdr->b_l1hdr.b_datacnt = 1;
4609 arc_get_data_buf(buf);
4610 arc_access(hdr, hash_lock);
4613 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
4614 hdr->b_flags |= ARC_FLAG_PREDICTIVE_PREFETCH;
4615 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4617 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4618 acb->acb_done = done;
4619 acb->acb_private = private;
4621 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4622 hdr->b_l1hdr.b_acb = acb;
4623 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4625 if (HDR_HAS_L2HDR(hdr) &&
4626 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
4627 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
4628 addr = hdr->b_l2hdr.b_daddr;
4629 b_compress = hdr->b_l2hdr.b_compress;
4630 b_asize = hdr->b_l2hdr.b_asize;
4632 * Lock out device removal.
4634 if (vdev_is_dead(vd) ||
4635 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
4639 if (hash_lock != NULL)
4640 mutex_exit(hash_lock);
4643 * At this point, we have a level 1 cache miss. Try again in
4644 * L2ARC if possible.
4646 ASSERT3U(hdr->b_size, ==, size);
4647 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
4648 uint64_t, size, zbookmark_phys_t *, zb);
4649 ARCSTAT_BUMP(arcstat_misses);
4650 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4651 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4652 data, metadata, misses);
4657 racct_add_force(curproc, RACCT_READBPS, size);
4658 racct_add_force(curproc, RACCT_READIOPS, 1);
4659 PROC_UNLOCK(curproc);
4662 curthread->td_ru.ru_inblock++;
4665 if (priority == ZIO_PRIORITY_ASYNC_READ)
4666 hdr->b_flags |= ARC_FLAG_PRIO_ASYNC_READ;
4668 hdr->b_flags &= ~ARC_FLAG_PRIO_ASYNC_READ;
4670 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
4672 * Read from the L2ARC if the following are true:
4673 * 1. The L2ARC vdev was previously cached.
4674 * 2. This buffer still has L2ARC metadata.
4675 * 3. This buffer isn't currently writing to the L2ARC.
4676 * 4. The L2ARC entry wasn't evicted, which may
4677 * also have invalidated the vdev.
4678 * 5. This isn't prefetch and l2arc_noprefetch is set.
4680 if (HDR_HAS_L2HDR(hdr) &&
4681 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
4682 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
4683 l2arc_read_callback_t *cb;
4686 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
4687 ARCSTAT_BUMP(arcstat_l2_hits);
4689 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
4691 cb->l2rcb_buf = buf;
4692 cb->l2rcb_spa = spa;
4695 cb->l2rcb_flags = zio_flags;
4696 cb->l2rcb_compress = b_compress;
4697 if (b_asize > hdr->b_size) {
4698 ASSERT3U(b_compress, ==,
4700 b_data = zio_data_buf_alloc(b_asize);
4701 cb->l2rcb_data = b_data;
4703 b_data = buf->b_data;
4706 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
4707 addr + size < vd->vdev_psize -
4708 VDEV_LABEL_END_SIZE);
4711 * l2arc read. The SCL_L2ARC lock will be
4712 * released by l2arc_read_done().
4713 * Issue a null zio if the underlying buffer
4714 * was squashed to zero size by compression.
4716 if (b_compress == ZIO_COMPRESS_EMPTY) {
4717 ASSERT3U(b_asize, ==, 0);
4718 rzio = zio_null(pio, spa, vd,
4719 l2arc_read_done, cb,
4720 zio_flags | ZIO_FLAG_DONT_CACHE |
4722 ZIO_FLAG_DONT_PROPAGATE |
4723 ZIO_FLAG_DONT_RETRY);
4725 rzio = zio_read_phys(pio, vd, addr,
4728 l2arc_read_done, cb, priority,
4729 zio_flags | ZIO_FLAG_DONT_CACHE |
4731 ZIO_FLAG_DONT_PROPAGATE |
4732 ZIO_FLAG_DONT_RETRY, B_FALSE);
4734 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
4736 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
4738 if (*arc_flags & ARC_FLAG_NOWAIT) {
4743 ASSERT(*arc_flags & ARC_FLAG_WAIT);
4744 if (zio_wait(rzio) == 0)
4747 /* l2arc read error; goto zio_read() */
4749 DTRACE_PROBE1(l2arc__miss,
4750 arc_buf_hdr_t *, hdr);
4751 ARCSTAT_BUMP(arcstat_l2_misses);
4752 if (HDR_L2_WRITING(hdr))
4753 ARCSTAT_BUMP(arcstat_l2_rw_clash);
4754 spa_config_exit(spa, SCL_L2ARC, vd);
4758 spa_config_exit(spa, SCL_L2ARC, vd);
4759 if (l2arc_ndev != 0) {
4760 DTRACE_PROBE1(l2arc__miss,
4761 arc_buf_hdr_t *, hdr);
4762 ARCSTAT_BUMP(arcstat_l2_misses);
4766 rzio = zio_read(pio, spa, bp, buf->b_data, size,
4767 arc_read_done, buf, priority, zio_flags, zb);
4769 if (*arc_flags & ARC_FLAG_WAIT)
4770 return (zio_wait(rzio));
4772 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4779 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
4781 ASSERT(buf->b_hdr != NULL);
4782 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon);
4783 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) ||
4785 ASSERT(buf->b_efunc == NULL);
4786 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
4788 buf->b_efunc = func;
4789 buf->b_private = private;
4793 * Notify the arc that a block was freed, and thus will never be used again.
4796 arc_freed(spa_t *spa, const blkptr_t *bp)
4799 kmutex_t *hash_lock;
4800 uint64_t guid = spa_load_guid(spa);
4802 ASSERT(!BP_IS_EMBEDDED(bp));
4804 hdr = buf_hash_find(guid, bp, &hash_lock);
4807 if (HDR_BUF_AVAILABLE(hdr)) {
4808 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
4809 add_reference(hdr, hash_lock, FTAG);
4810 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4811 mutex_exit(hash_lock);
4813 arc_release(buf, FTAG);
4814 (void) arc_buf_remove_ref(buf, FTAG);
4816 mutex_exit(hash_lock);
4822 * Clear the user eviction callback set by arc_set_callback(), first calling
4823 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4824 * clearing the callback may result in the arc_buf being destroyed. However,
4825 * it will not result in the *last* arc_buf being destroyed, hence the data
4826 * will remain cached in the ARC. We make a copy of the arc buffer here so
4827 * that we can process the callback without holding any locks.
4829 * It's possible that the callback is already in the process of being cleared
4830 * by another thread. In this case we can not clear the callback.
4832 * Returns B_TRUE if the callback was successfully called and cleared.
4835 arc_clear_callback(arc_buf_t *buf)
4838 kmutex_t *hash_lock;
4839 arc_evict_func_t *efunc = buf->b_efunc;
4840 void *private = buf->b_private;
4842 mutex_enter(&buf->b_evict_lock);
4846 * We are in arc_do_user_evicts().
4848 ASSERT(buf->b_data == NULL);
4849 mutex_exit(&buf->b_evict_lock);
4851 } else if (buf->b_data == NULL) {
4853 * We are on the eviction list; process this buffer now
4854 * but let arc_do_user_evicts() do the reaping.
4856 buf->b_efunc = NULL;
4857 mutex_exit(&buf->b_evict_lock);
4858 VERIFY0(efunc(private));
4861 hash_lock = HDR_LOCK(hdr);
4862 mutex_enter(hash_lock);
4864 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4866 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <,
4867 hdr->b_l1hdr.b_datacnt);
4868 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4869 hdr->b_l1hdr.b_state == arc_mfu);
4871 buf->b_efunc = NULL;
4872 buf->b_private = NULL;
4874 if (hdr->b_l1hdr.b_datacnt > 1) {
4875 mutex_exit(&buf->b_evict_lock);
4876 arc_buf_destroy(buf, TRUE);
4878 ASSERT(buf == hdr->b_l1hdr.b_buf);
4879 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4880 mutex_exit(&buf->b_evict_lock);
4883 mutex_exit(hash_lock);
4884 VERIFY0(efunc(private));
4889 * Release this buffer from the cache, making it an anonymous buffer. This
4890 * must be done after a read and prior to modifying the buffer contents.
4891 * If the buffer has more than one reference, we must make
4892 * a new hdr for the buffer.
4895 arc_release(arc_buf_t *buf, void *tag)
4897 arc_buf_hdr_t *hdr = buf->b_hdr;
4900 * It would be nice to assert that if it's DMU metadata (level >
4901 * 0 || it's the dnode file), then it must be syncing context.
4902 * But we don't know that information at this level.
4905 mutex_enter(&buf->b_evict_lock);
4907 ASSERT(HDR_HAS_L1HDR(hdr));
4910 * We don't grab the hash lock prior to this check, because if
4911 * the buffer's header is in the arc_anon state, it won't be
4912 * linked into the hash table.
4914 if (hdr->b_l1hdr.b_state == arc_anon) {
4915 mutex_exit(&buf->b_evict_lock);
4916 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4917 ASSERT(!HDR_IN_HASH_TABLE(hdr));
4918 ASSERT(!HDR_HAS_L2HDR(hdr));
4919 ASSERT(BUF_EMPTY(hdr));
4920 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1);
4921 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
4922 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
4924 ASSERT3P(buf->b_efunc, ==, NULL);
4925 ASSERT3P(buf->b_private, ==, NULL);
4927 hdr->b_l1hdr.b_arc_access = 0;
4933 kmutex_t *hash_lock = HDR_LOCK(hdr);
4934 mutex_enter(hash_lock);
4937 * This assignment is only valid as long as the hash_lock is
4938 * held, we must be careful not to reference state or the
4939 * b_state field after dropping the lock.
4941 arc_state_t *state = hdr->b_l1hdr.b_state;
4942 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4943 ASSERT3P(state, !=, arc_anon);
4945 /* this buffer is not on any list */
4946 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
4948 if (HDR_HAS_L2HDR(hdr)) {
4949 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4952 * We have to recheck this conditional again now that
4953 * we're holding the l2ad_mtx to prevent a race with
4954 * another thread which might be concurrently calling
4955 * l2arc_evict(). In that case, l2arc_evict() might have
4956 * destroyed the header's L2 portion as we were waiting
4957 * to acquire the l2ad_mtx.
4959 if (HDR_HAS_L2HDR(hdr)) {
4961 arc_hdr_l2hdr_destroy(hdr);
4964 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4968 * Do we have more than one buf?
4970 if (hdr->b_l1hdr.b_datacnt > 1) {
4971 arc_buf_hdr_t *nhdr;
4973 uint64_t blksz = hdr->b_size;
4974 uint64_t spa = hdr->b_spa;
4975 arc_buf_contents_t type = arc_buf_type(hdr);
4976 uint32_t flags = hdr->b_flags;
4978 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
4980 * Pull the data off of this hdr and attach it to
4981 * a new anonymous hdr.
4983 (void) remove_reference(hdr, hash_lock, tag);
4984 bufp = &hdr->b_l1hdr.b_buf;
4985 while (*bufp != buf)
4986 bufp = &(*bufp)->b_next;
4987 *bufp = buf->b_next;
4990 ASSERT3P(state, !=, arc_l2c_only);
4992 (void) refcount_remove_many(
4993 &state->arcs_size, hdr->b_size, buf);
4995 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
4996 ASSERT3P(state, !=, arc_l2c_only);
4997 uint64_t *size = &state->arcs_lsize[type];
4998 ASSERT3U(*size, >=, hdr->b_size);
4999 atomic_add_64(size, -hdr->b_size);
5003 * We're releasing a duplicate user data buffer, update
5004 * our statistics accordingly.
5006 if (HDR_ISTYPE_DATA(hdr)) {
5007 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
5008 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
5011 hdr->b_l1hdr.b_datacnt -= 1;
5012 arc_cksum_verify(buf);
5014 arc_buf_unwatch(buf);
5017 mutex_exit(hash_lock);
5019 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
5020 nhdr->b_size = blksz;
5023 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING;
5024 nhdr->b_flags |= arc_bufc_to_flags(type);
5025 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
5027 nhdr->b_l1hdr.b_buf = buf;
5028 nhdr->b_l1hdr.b_datacnt = 1;
5029 nhdr->b_l1hdr.b_state = arc_anon;
5030 nhdr->b_l1hdr.b_arc_access = 0;
5031 nhdr->b_l1hdr.b_tmp_cdata = NULL;
5032 nhdr->b_freeze_cksum = NULL;
5034 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5036 mutex_exit(&buf->b_evict_lock);
5037 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf);
5039 mutex_exit(&buf->b_evict_lock);
5040 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5041 /* protected by hash lock, or hdr is on arc_anon */
5042 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5043 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5044 arc_change_state(arc_anon, hdr, hash_lock);
5045 hdr->b_l1hdr.b_arc_access = 0;
5046 mutex_exit(hash_lock);
5048 buf_discard_identity(hdr);
5051 buf->b_efunc = NULL;
5052 buf->b_private = NULL;
5056 arc_released(arc_buf_t *buf)
5060 mutex_enter(&buf->b_evict_lock);
5061 released = (buf->b_data != NULL &&
5062 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5063 mutex_exit(&buf->b_evict_lock);
5069 arc_referenced(arc_buf_t *buf)
5073 mutex_enter(&buf->b_evict_lock);
5074 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5075 mutex_exit(&buf->b_evict_lock);
5076 return (referenced);
5081 arc_write_ready(zio_t *zio)
5083 arc_write_callback_t *callback = zio->io_private;
5084 arc_buf_t *buf = callback->awcb_buf;
5085 arc_buf_hdr_t *hdr = buf->b_hdr;
5087 ASSERT(HDR_HAS_L1HDR(hdr));
5088 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5089 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
5090 callback->awcb_ready(zio, buf, callback->awcb_private);
5093 * If the IO is already in progress, then this is a re-write
5094 * attempt, so we need to thaw and re-compute the cksum.
5095 * It is the responsibility of the callback to handle the
5096 * accounting for any re-write attempt.
5098 if (HDR_IO_IN_PROGRESS(hdr)) {
5099 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
5100 if (hdr->b_freeze_cksum != NULL) {
5101 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
5102 hdr->b_freeze_cksum = NULL;
5104 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
5106 arc_cksum_compute(buf, B_FALSE);
5107 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
5111 arc_write_children_ready(zio_t *zio)
5113 arc_write_callback_t *callback = zio->io_private;
5114 arc_buf_t *buf = callback->awcb_buf;
5116 callback->awcb_children_ready(zio, buf, callback->awcb_private);
5120 * The SPA calls this callback for each physical write that happens on behalf
5121 * of a logical write. See the comment in dbuf_write_physdone() for details.
5124 arc_write_physdone(zio_t *zio)
5126 arc_write_callback_t *cb = zio->io_private;
5127 if (cb->awcb_physdone != NULL)
5128 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
5132 arc_write_done(zio_t *zio)
5134 arc_write_callback_t *callback = zio->io_private;
5135 arc_buf_t *buf = callback->awcb_buf;
5136 arc_buf_hdr_t *hdr = buf->b_hdr;
5138 ASSERT(hdr->b_l1hdr.b_acb == NULL);
5140 if (zio->io_error == 0) {
5141 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5142 buf_discard_identity(hdr);
5144 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
5145 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
5148 ASSERT(BUF_EMPTY(hdr));
5152 * If the block to be written was all-zero or compressed enough to be
5153 * embedded in the BP, no write was performed so there will be no
5154 * dva/birth/checksum. The buffer must therefore remain anonymous
5157 if (!BUF_EMPTY(hdr)) {
5158 arc_buf_hdr_t *exists;
5159 kmutex_t *hash_lock;
5161 ASSERT(zio->io_error == 0);
5163 arc_cksum_verify(buf);
5165 exists = buf_hash_insert(hdr, &hash_lock);
5166 if (exists != NULL) {
5168 * This can only happen if we overwrite for
5169 * sync-to-convergence, because we remove
5170 * buffers from the hash table when we arc_free().
5172 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5173 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5174 panic("bad overwrite, hdr=%p exists=%p",
5175 (void *)hdr, (void *)exists);
5176 ASSERT(refcount_is_zero(
5177 &exists->b_l1hdr.b_refcnt));
5178 arc_change_state(arc_anon, exists, hash_lock);
5179 mutex_exit(hash_lock);
5180 arc_hdr_destroy(exists);
5181 exists = buf_hash_insert(hdr, &hash_lock);
5182 ASSERT3P(exists, ==, NULL);
5183 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
5185 ASSERT(zio->io_prop.zp_nopwrite);
5186 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5187 panic("bad nopwrite, hdr=%p exists=%p",
5188 (void *)hdr, (void *)exists);
5191 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
5192 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
5193 ASSERT(BP_GET_DEDUP(zio->io_bp));
5194 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
5197 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
5198 /* if it's not anon, we are doing a scrub */
5199 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
5200 arc_access(hdr, hash_lock);
5201 mutex_exit(hash_lock);
5203 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
5206 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5207 callback->awcb_done(zio, buf, callback->awcb_private);
5209 kmem_free(callback, sizeof (arc_write_callback_t));
5213 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
5214 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
5215 const zio_prop_t *zp, arc_done_func_t *ready,
5216 arc_done_func_t *children_ready, arc_done_func_t *physdone,
5217 arc_done_func_t *done, void *private, zio_priority_t priority,
5218 int zio_flags, const zbookmark_phys_t *zb)
5220 arc_buf_hdr_t *hdr = buf->b_hdr;
5221 arc_write_callback_t *callback;
5224 ASSERT(ready != NULL);
5225 ASSERT(done != NULL);
5226 ASSERT(!HDR_IO_ERROR(hdr));
5227 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5228 ASSERT(hdr->b_l1hdr.b_acb == NULL);
5229 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
5231 hdr->b_flags |= ARC_FLAG_L2CACHE;
5233 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
5234 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
5235 callback->awcb_ready = ready;
5236 callback->awcb_children_ready = children_ready;
5237 callback->awcb_physdone = physdone;
5238 callback->awcb_done = done;
5239 callback->awcb_private = private;
5240 callback->awcb_buf = buf;
5242 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
5244 (children_ready != NULL) ? arc_write_children_ready : NULL,
5245 arc_write_physdone, arc_write_done, callback,
5246 priority, zio_flags, zb);
5252 arc_memory_throttle(uint64_t reserve, uint64_t txg)
5255 uint64_t available_memory = ptob(freemem);
5256 static uint64_t page_load = 0;
5257 static uint64_t last_txg = 0;
5259 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
5261 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE)));
5264 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
5267 if (txg > last_txg) {
5272 * If we are in pageout, we know that memory is already tight,
5273 * the arc is already going to be evicting, so we just want to
5274 * continue to let page writes occur as quickly as possible.
5276 if (curproc == pageproc) {
5277 if (page_load > MAX(ptob(minfree), available_memory) / 4)
5278 return (SET_ERROR(ERESTART));
5279 /* Note: reserve is inflated, so we deflate */
5280 page_load += reserve / 8;
5282 } else if (page_load > 0 && arc_reclaim_needed()) {
5283 /* memory is low, delay before restarting */
5284 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
5285 return (SET_ERROR(EAGAIN));
5293 arc_tempreserve_clear(uint64_t reserve)
5295 atomic_add_64(&arc_tempreserve, -reserve);
5296 ASSERT((int64_t)arc_tempreserve >= 0);
5300 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
5305 if (reserve > arc_c/4 && !arc_no_grow) {
5306 arc_c = MIN(arc_c_max, reserve * 4);
5307 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
5309 if (reserve > arc_c)
5310 return (SET_ERROR(ENOMEM));
5313 * Don't count loaned bufs as in flight dirty data to prevent long
5314 * network delays from blocking transactions that are ready to be
5315 * assigned to a txg.
5317 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
5318 arc_loaned_bytes), 0);
5321 * Writes will, almost always, require additional memory allocations
5322 * in order to compress/encrypt/etc the data. We therefore need to
5323 * make sure that there is sufficient available memory for this.
5325 error = arc_memory_throttle(reserve, txg);
5330 * Throttle writes when the amount of dirty data in the cache
5331 * gets too large. We try to keep the cache less than half full
5332 * of dirty blocks so that our sync times don't grow too large.
5333 * Note: if two requests come in concurrently, we might let them
5334 * both succeed, when one of them should fail. Not a huge deal.
5337 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
5338 anon_size > arc_c / 4) {
5339 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5340 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5341 arc_tempreserve>>10,
5342 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
5343 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
5344 reserve>>10, arc_c>>10);
5345 return (SET_ERROR(ERESTART));
5347 atomic_add_64(&arc_tempreserve, reserve);
5352 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
5353 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
5355 size->value.ui64 = refcount_count(&state->arcs_size);
5356 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
5357 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
5361 arc_kstat_update(kstat_t *ksp, int rw)
5363 arc_stats_t *as = ksp->ks_data;
5365 if (rw == KSTAT_WRITE) {
5368 arc_kstat_update_state(arc_anon,
5369 &as->arcstat_anon_size,
5370 &as->arcstat_anon_evictable_data,
5371 &as->arcstat_anon_evictable_metadata);
5372 arc_kstat_update_state(arc_mru,
5373 &as->arcstat_mru_size,
5374 &as->arcstat_mru_evictable_data,
5375 &as->arcstat_mru_evictable_metadata);
5376 arc_kstat_update_state(arc_mru_ghost,
5377 &as->arcstat_mru_ghost_size,
5378 &as->arcstat_mru_ghost_evictable_data,
5379 &as->arcstat_mru_ghost_evictable_metadata);
5380 arc_kstat_update_state(arc_mfu,
5381 &as->arcstat_mfu_size,
5382 &as->arcstat_mfu_evictable_data,
5383 &as->arcstat_mfu_evictable_metadata);
5384 arc_kstat_update_state(arc_mfu_ghost,
5385 &as->arcstat_mfu_ghost_size,
5386 &as->arcstat_mfu_ghost_evictable_data,
5387 &as->arcstat_mfu_ghost_evictable_metadata);
5394 * This function *must* return indices evenly distributed between all
5395 * sublists of the multilist. This is needed due to how the ARC eviction
5396 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5397 * distributed between all sublists and uses this assumption when
5398 * deciding which sublist to evict from and how much to evict from it.
5401 arc_state_multilist_index_func(multilist_t *ml, void *obj)
5403 arc_buf_hdr_t *hdr = obj;
5406 * We rely on b_dva to generate evenly distributed index
5407 * numbers using buf_hash below. So, as an added precaution,
5408 * let's make sure we never add empty buffers to the arc lists.
5410 ASSERT(!BUF_EMPTY(hdr));
5413 * The assumption here, is the hash value for a given
5414 * arc_buf_hdr_t will remain constant throughout it's lifetime
5415 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
5416 * Thus, we don't need to store the header's sublist index
5417 * on insertion, as this index can be recalculated on removal.
5419 * Also, the low order bits of the hash value are thought to be
5420 * distributed evenly. Otherwise, in the case that the multilist
5421 * has a power of two number of sublists, each sublists' usage
5422 * would not be evenly distributed.
5424 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
5425 multilist_get_num_sublists(ml));
5429 static eventhandler_tag arc_event_lowmem = NULL;
5432 arc_lowmem(void *arg __unused, int howto __unused)
5435 mutex_enter(&arc_reclaim_lock);
5436 /* XXX: Memory deficit should be passed as argument. */
5437 needfree = btoc(arc_c >> arc_shrink_shift);
5438 DTRACE_PROBE(arc__needfree);
5439 cv_signal(&arc_reclaim_thread_cv);
5442 * It is unsafe to block here in arbitrary threads, because we can come
5443 * here from ARC itself and may hold ARC locks and thus risk a deadlock
5444 * with ARC reclaim thread.
5446 if (curproc == pageproc)
5447 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5448 mutex_exit(&arc_reclaim_lock);
5455 int i, prefetch_tunable_set = 0;
5457 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
5458 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
5459 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
5461 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
5462 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL);
5464 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
5465 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL);
5467 /* Convert seconds to clock ticks */
5468 arc_min_prefetch_lifespan = 1 * hz;
5470 /* Start out with 1/8 of all memory */
5471 arc_c = kmem_size() / 8;
5476 * On architectures where the physical memory can be larger
5477 * than the addressable space (intel in 32-bit mode), we may
5478 * need to limit the cache to 1/8 of VM size.
5480 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
5482 #endif /* illumos */
5483 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
5484 arc_c_min = MAX(arc_c / 4, arc_abs_min);
5485 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
5486 if (arc_c * 8 >= 1 << 30)
5487 arc_c_max = (arc_c * 8) - (1 << 30);
5489 arc_c_max = arc_c_min;
5490 arc_c_max = MAX(arc_c * 5, arc_c_max);
5493 * In userland, there's only the memory pressure that we artificially
5494 * create (see arc_available_memory()). Don't let arc_c get too
5495 * small, because it can cause transactions to be larger than
5496 * arc_c, causing arc_tempreserve_space() to fail.
5499 arc_c_min = arc_c_max / 2;
5504 * Allow the tunables to override our calculations if they are
5507 if (zfs_arc_max > arc_abs_min && zfs_arc_max < kmem_size())
5508 arc_c_max = zfs_arc_max;
5509 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
5510 arc_c_min = zfs_arc_min;
5514 arc_p = (arc_c >> 1);
5516 /* limit meta-data to 1/4 of the arc capacity */
5517 arc_meta_limit = arc_c_max / 4;
5519 /* Allow the tunable to override if it is reasonable */
5520 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
5521 arc_meta_limit = zfs_arc_meta_limit;
5523 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
5524 arc_c_min = arc_meta_limit / 2;
5526 if (zfs_arc_meta_min > 0) {
5527 arc_meta_min = zfs_arc_meta_min;
5529 arc_meta_min = arc_c_min / 2;
5532 if (zfs_arc_grow_retry > 0)
5533 arc_grow_retry = zfs_arc_grow_retry;
5535 if (zfs_arc_shrink_shift > 0)
5536 arc_shrink_shift = zfs_arc_shrink_shift;
5539 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
5541 if (arc_no_grow_shift >= arc_shrink_shift)
5542 arc_no_grow_shift = arc_shrink_shift - 1;
5544 if (zfs_arc_p_min_shift > 0)
5545 arc_p_min_shift = zfs_arc_p_min_shift;
5547 if (zfs_arc_num_sublists_per_state < 1)
5548 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1);
5550 /* if kmem_flags are set, lets try to use less memory */
5551 if (kmem_debugging())
5553 if (arc_c < arc_c_min)
5556 zfs_arc_min = arc_c_min;
5557 zfs_arc_max = arc_c_max;
5559 arc_anon = &ARC_anon;
5561 arc_mru_ghost = &ARC_mru_ghost;
5563 arc_mfu_ghost = &ARC_mfu_ghost;
5564 arc_l2c_only = &ARC_l2c_only;
5567 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5568 sizeof (arc_buf_hdr_t),
5569 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5570 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5571 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5572 sizeof (arc_buf_hdr_t),
5573 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5574 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5575 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5576 sizeof (arc_buf_hdr_t),
5577 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5578 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5579 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5580 sizeof (arc_buf_hdr_t),
5581 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5582 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5583 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5584 sizeof (arc_buf_hdr_t),
5585 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5586 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5587 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5588 sizeof (arc_buf_hdr_t),
5589 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5590 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5591 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5592 sizeof (arc_buf_hdr_t),
5593 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5594 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5595 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5596 sizeof (arc_buf_hdr_t),
5597 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5598 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5599 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5600 sizeof (arc_buf_hdr_t),
5601 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5602 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5603 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
5604 sizeof (arc_buf_hdr_t),
5605 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5606 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5608 refcount_create(&arc_anon->arcs_size);
5609 refcount_create(&arc_mru->arcs_size);
5610 refcount_create(&arc_mru_ghost->arcs_size);
5611 refcount_create(&arc_mfu->arcs_size);
5612 refcount_create(&arc_mfu_ghost->arcs_size);
5613 refcount_create(&arc_l2c_only->arcs_size);
5617 arc_reclaim_thread_exit = FALSE;
5618 arc_user_evicts_thread_exit = FALSE;
5619 arc_dnlc_evicts_thread_exit = FALSE;
5620 arc_eviction_list = NULL;
5621 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
5623 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
5624 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
5626 if (arc_ksp != NULL) {
5627 arc_ksp->ks_data = &arc_stats;
5628 arc_ksp->ks_update = arc_kstat_update;
5629 kstat_install(arc_ksp);
5632 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
5633 TS_RUN, minclsyspri);
5636 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
5637 EVENTHANDLER_PRI_FIRST);
5640 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0,
5641 TS_RUN, minclsyspri);
5643 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0,
5644 TS_RUN, minclsyspri);
5650 * Calculate maximum amount of dirty data per pool.
5652 * If it has been set by /etc/system, take that.
5653 * Otherwise, use a percentage of physical memory defined by
5654 * zfs_dirty_data_max_percent (default 10%) with a cap at
5655 * zfs_dirty_data_max_max (default 4GB).
5657 if (zfs_dirty_data_max == 0) {
5658 zfs_dirty_data_max = ptob(physmem) *
5659 zfs_dirty_data_max_percent / 100;
5660 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
5661 zfs_dirty_data_max_max);
5665 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
5666 prefetch_tunable_set = 1;
5669 if (prefetch_tunable_set == 0) {
5670 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
5672 printf(" add \"vfs.zfs.prefetch_disable=0\" "
5673 "to /boot/loader.conf.\n");
5674 zfs_prefetch_disable = 1;
5677 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
5678 prefetch_tunable_set == 0) {
5679 printf("ZFS NOTICE: Prefetch is disabled by default if less "
5680 "than 4GB of RAM is present;\n"
5681 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
5682 "to /boot/loader.conf.\n");
5683 zfs_prefetch_disable = 1;
5686 /* Warn about ZFS memory and address space requirements. */
5687 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
5688 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
5689 "expect unstable behavior.\n");
5691 if (kmem_size() < 512 * (1 << 20)) {
5692 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
5693 "expect unstable behavior.\n");
5694 printf(" Consider tuning vm.kmem_size and "
5695 "vm.kmem_size_max\n");
5696 printf(" in /boot/loader.conf.\n");
5704 mutex_enter(&arc_reclaim_lock);
5705 arc_reclaim_thread_exit = TRUE;
5707 * The reclaim thread will set arc_reclaim_thread_exit back to
5708 * FALSE when it is finished exiting; we're waiting for that.
5710 while (arc_reclaim_thread_exit) {
5711 cv_signal(&arc_reclaim_thread_cv);
5712 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
5714 mutex_exit(&arc_reclaim_lock);
5716 mutex_enter(&arc_user_evicts_lock);
5717 arc_user_evicts_thread_exit = TRUE;
5719 * The user evicts thread will set arc_user_evicts_thread_exit
5720 * to FALSE when it is finished exiting; we're waiting for that.
5722 while (arc_user_evicts_thread_exit) {
5723 cv_signal(&arc_user_evicts_cv);
5724 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock);
5726 mutex_exit(&arc_user_evicts_lock);
5728 mutex_enter(&arc_dnlc_evicts_lock);
5729 arc_dnlc_evicts_thread_exit = TRUE;
5731 * The user evicts thread will set arc_user_evicts_thread_exit
5732 * to FALSE when it is finished exiting; we're waiting for that.
5734 while (arc_dnlc_evicts_thread_exit) {
5735 cv_signal(&arc_dnlc_evicts_cv);
5736 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock);
5738 mutex_exit(&arc_dnlc_evicts_lock);
5740 /* Use TRUE to ensure *all* buffers are evicted */
5741 arc_flush(NULL, TRUE);
5745 if (arc_ksp != NULL) {
5746 kstat_delete(arc_ksp);
5750 mutex_destroy(&arc_reclaim_lock);
5751 cv_destroy(&arc_reclaim_thread_cv);
5752 cv_destroy(&arc_reclaim_waiters_cv);
5754 mutex_destroy(&arc_user_evicts_lock);
5755 cv_destroy(&arc_user_evicts_cv);
5757 mutex_destroy(&arc_dnlc_evicts_lock);
5758 cv_destroy(&arc_dnlc_evicts_cv);
5760 refcount_destroy(&arc_anon->arcs_size);
5761 refcount_destroy(&arc_mru->arcs_size);
5762 refcount_destroy(&arc_mru_ghost->arcs_size);
5763 refcount_destroy(&arc_mfu->arcs_size);
5764 refcount_destroy(&arc_mfu_ghost->arcs_size);
5765 refcount_destroy(&arc_l2c_only->arcs_size);
5767 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
5768 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5769 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
5770 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
5771 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
5772 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
5773 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
5774 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
5775 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
5776 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
5780 ASSERT0(arc_loaned_bytes);
5783 if (arc_event_lowmem != NULL)
5784 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
5791 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5792 * It uses dedicated storage devices to hold cached data, which are populated
5793 * using large infrequent writes. The main role of this cache is to boost
5794 * the performance of random read workloads. The intended L2ARC devices
5795 * include short-stroked disks, solid state disks, and other media with
5796 * substantially faster read latency than disk.
5798 * +-----------------------+
5800 * +-----------------------+
5803 * l2arc_feed_thread() arc_read()
5807 * +---------------+ |
5809 * +---------------+ |
5814 * +-------+ +-------+
5816 * | cache | | cache |
5817 * +-------+ +-------+
5818 * +=========+ .-----.
5819 * : L2ARC : |-_____-|
5820 * : devices : | Disks |
5821 * +=========+ `-_____-'
5823 * Read requests are satisfied from the following sources, in order:
5826 * 2) vdev cache of L2ARC devices
5828 * 4) vdev cache of disks
5831 * Some L2ARC device types exhibit extremely slow write performance.
5832 * To accommodate for this there are some significant differences between
5833 * the L2ARC and traditional cache design:
5835 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5836 * the ARC behave as usual, freeing buffers and placing headers on ghost
5837 * lists. The ARC does not send buffers to the L2ARC during eviction as
5838 * this would add inflated write latencies for all ARC memory pressure.
5840 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5841 * It does this by periodically scanning buffers from the eviction-end of
5842 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5843 * not already there. It scans until a headroom of buffers is satisfied,
5844 * which itself is a buffer for ARC eviction. If a compressible buffer is
5845 * found during scanning and selected for writing to an L2ARC device, we
5846 * temporarily boost scanning headroom during the next scan cycle to make
5847 * sure we adapt to compression effects (which might significantly reduce
5848 * the data volume we write to L2ARC). The thread that does this is
5849 * l2arc_feed_thread(), illustrated below; example sizes are included to
5850 * provide a better sense of ratio than this diagram:
5853 * +---------------------+----------+
5854 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5855 * +---------------------+----------+ | o L2ARC eligible
5856 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5857 * +---------------------+----------+ |
5858 * 15.9 Gbytes ^ 32 Mbytes |
5860 * l2arc_feed_thread()
5862 * l2arc write hand <--[oooo]--'
5866 * +==============================+
5867 * L2ARC dev |####|#|###|###| |####| ... |
5868 * +==============================+
5871 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5872 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5873 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5874 * safe to say that this is an uncommon case, since buffers at the end of
5875 * the ARC lists have moved there due to inactivity.
5877 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5878 * then the L2ARC simply misses copying some buffers. This serves as a
5879 * pressure valve to prevent heavy read workloads from both stalling the ARC
5880 * with waits and clogging the L2ARC with writes. This also helps prevent
5881 * the potential for the L2ARC to churn if it attempts to cache content too
5882 * quickly, such as during backups of the entire pool.
5884 * 5. After system boot and before the ARC has filled main memory, there are
5885 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5886 * lists can remain mostly static. Instead of searching from tail of these
5887 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5888 * for eligible buffers, greatly increasing its chance of finding them.
5890 * The L2ARC device write speed is also boosted during this time so that
5891 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5892 * there are no L2ARC reads, and no fear of degrading read performance
5893 * through increased writes.
5895 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5896 * the vdev queue can aggregate them into larger and fewer writes. Each
5897 * device is written to in a rotor fashion, sweeping writes through
5898 * available space then repeating.
5900 * 7. The L2ARC does not store dirty content. It never needs to flush
5901 * write buffers back to disk based storage.
5903 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5904 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5906 * The performance of the L2ARC can be tweaked by a number of tunables, which
5907 * may be necessary for different workloads:
5909 * l2arc_write_max max write bytes per interval
5910 * l2arc_write_boost extra write bytes during device warmup
5911 * l2arc_noprefetch skip caching prefetched buffers
5912 * l2arc_headroom number of max device writes to precache
5913 * l2arc_headroom_boost when we find compressed buffers during ARC
5914 * scanning, we multiply headroom by this
5915 * percentage factor for the next scan cycle,
5916 * since more compressed buffers are likely to
5918 * l2arc_feed_secs seconds between L2ARC writing
5920 * Tunables may be removed or added as future performance improvements are
5921 * integrated, and also may become zpool properties.
5923 * There are three key functions that control how the L2ARC warms up:
5925 * l2arc_write_eligible() check if a buffer is eligible to cache
5926 * l2arc_write_size() calculate how much to write
5927 * l2arc_write_interval() calculate sleep delay between writes
5929 * These three functions determine what to write, how much, and how quickly
5934 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
5937 * A buffer is *not* eligible for the L2ARC if it:
5938 * 1. belongs to a different spa.
5939 * 2. is already cached on the L2ARC.
5940 * 3. has an I/O in progress (it may be an incomplete read).
5941 * 4. is flagged not eligible (zfs property).
5943 if (hdr->b_spa != spa_guid) {
5944 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
5947 if (HDR_HAS_L2HDR(hdr)) {
5948 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
5951 if (HDR_IO_IN_PROGRESS(hdr)) {
5952 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
5955 if (!HDR_L2CACHE(hdr)) {
5956 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
5964 l2arc_write_size(void)
5969 * Make sure our globals have meaningful values in case the user
5972 size = l2arc_write_max;
5974 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
5975 "be greater than zero, resetting it to the default (%d)",
5977 size = l2arc_write_max = L2ARC_WRITE_SIZE;
5980 if (arc_warm == B_FALSE)
5981 size += l2arc_write_boost;
5988 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
5990 clock_t interval, next, now;
5993 * If the ARC lists are busy, increase our write rate; if the
5994 * lists are stale, idle back. This is achieved by checking
5995 * how much we previously wrote - if it was more than half of
5996 * what we wanted, schedule the next write much sooner.
5998 if (l2arc_feed_again && wrote > (wanted / 2))
5999 interval = (hz * l2arc_feed_min_ms) / 1000;
6001 interval = hz * l2arc_feed_secs;
6003 now = ddi_get_lbolt();
6004 next = MAX(now, MIN(now + interval, began + interval));
6010 * Cycle through L2ARC devices. This is how L2ARC load balances.
6011 * If a device is returned, this also returns holding the spa config lock.
6013 static l2arc_dev_t *
6014 l2arc_dev_get_next(void)
6016 l2arc_dev_t *first, *next = NULL;
6019 * Lock out the removal of spas (spa_namespace_lock), then removal
6020 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6021 * both locks will be dropped and a spa config lock held instead.
6023 mutex_enter(&spa_namespace_lock);
6024 mutex_enter(&l2arc_dev_mtx);
6026 /* if there are no vdevs, there is nothing to do */
6027 if (l2arc_ndev == 0)
6031 next = l2arc_dev_last;
6033 /* loop around the list looking for a non-faulted vdev */
6035 next = list_head(l2arc_dev_list);
6037 next = list_next(l2arc_dev_list, next);
6039 next = list_head(l2arc_dev_list);
6042 /* if we have come back to the start, bail out */
6045 else if (next == first)
6048 } while (vdev_is_dead(next->l2ad_vdev));
6050 /* if we were unable to find any usable vdevs, return NULL */
6051 if (vdev_is_dead(next->l2ad_vdev))
6054 l2arc_dev_last = next;
6057 mutex_exit(&l2arc_dev_mtx);
6060 * Grab the config lock to prevent the 'next' device from being
6061 * removed while we are writing to it.
6064 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
6065 mutex_exit(&spa_namespace_lock);
6071 * Free buffers that were tagged for destruction.
6074 l2arc_do_free_on_write()
6077 l2arc_data_free_t *df, *df_prev;
6079 mutex_enter(&l2arc_free_on_write_mtx);
6080 buflist = l2arc_free_on_write;
6082 for (df = list_tail(buflist); df; df = df_prev) {
6083 df_prev = list_prev(buflist, df);
6084 ASSERT(df->l2df_data != NULL);
6085 ASSERT(df->l2df_func != NULL);
6086 df->l2df_func(df->l2df_data, df->l2df_size);
6087 list_remove(buflist, df);
6088 kmem_free(df, sizeof (l2arc_data_free_t));
6091 mutex_exit(&l2arc_free_on_write_mtx);
6095 * A write to a cache device has completed. Update all headers to allow
6096 * reads from these buffers to begin.
6099 l2arc_write_done(zio_t *zio)
6101 l2arc_write_callback_t *cb;
6104 arc_buf_hdr_t *head, *hdr, *hdr_prev;
6105 kmutex_t *hash_lock;
6106 int64_t bytes_dropped = 0;
6108 cb = zio->io_private;
6110 dev = cb->l2wcb_dev;
6111 ASSERT(dev != NULL);
6112 head = cb->l2wcb_head;
6113 ASSERT(head != NULL);
6114 buflist = &dev->l2ad_buflist;
6115 ASSERT(buflist != NULL);
6116 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
6117 l2arc_write_callback_t *, cb);
6119 if (zio->io_error != 0)
6120 ARCSTAT_BUMP(arcstat_l2_writes_error);
6123 * All writes completed, or an error was hit.
6126 mutex_enter(&dev->l2ad_mtx);
6127 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
6128 hdr_prev = list_prev(buflist, hdr);
6130 hash_lock = HDR_LOCK(hdr);
6133 * We cannot use mutex_enter or else we can deadlock
6134 * with l2arc_write_buffers (due to swapping the order
6135 * the hash lock and l2ad_mtx are taken).
6137 if (!mutex_tryenter(hash_lock)) {
6139 * Missed the hash lock. We must retry so we
6140 * don't leave the ARC_FLAG_L2_WRITING bit set.
6142 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
6145 * We don't want to rescan the headers we've
6146 * already marked as having been written out, so
6147 * we reinsert the head node so we can pick up
6148 * where we left off.
6150 list_remove(buflist, head);
6151 list_insert_after(buflist, hdr, head);
6153 mutex_exit(&dev->l2ad_mtx);
6156 * We wait for the hash lock to become available
6157 * to try and prevent busy waiting, and increase
6158 * the chance we'll be able to acquire the lock
6159 * the next time around.
6161 mutex_enter(hash_lock);
6162 mutex_exit(hash_lock);
6167 * We could not have been moved into the arc_l2c_only
6168 * state while in-flight due to our ARC_FLAG_L2_WRITING
6169 * bit being set. Let's just ensure that's being enforced.
6171 ASSERT(HDR_HAS_L1HDR(hdr));
6174 * We may have allocated a buffer for L2ARC compression,
6175 * we must release it to avoid leaking this data.
6177 l2arc_release_cdata_buf(hdr);
6179 if (zio->io_error != 0) {
6181 * Error - drop L2ARC entry.
6183 list_remove(buflist, hdr);
6185 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
6187 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize);
6188 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
6190 bytes_dropped += hdr->b_l2hdr.b_asize;
6191 (void) refcount_remove_many(&dev->l2ad_alloc,
6192 hdr->b_l2hdr.b_asize, hdr);
6196 * Allow ARC to begin reads and ghost list evictions to
6199 hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
6201 mutex_exit(hash_lock);
6204 atomic_inc_64(&l2arc_writes_done);
6205 list_remove(buflist, head);
6206 ASSERT(!HDR_HAS_L1HDR(head));
6207 kmem_cache_free(hdr_l2only_cache, head);
6208 mutex_exit(&dev->l2ad_mtx);
6210 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
6212 l2arc_do_free_on_write();
6214 kmem_free(cb, sizeof (l2arc_write_callback_t));
6218 * A read to a cache device completed. Validate buffer contents before
6219 * handing over to the regular ARC routines.
6222 l2arc_read_done(zio_t *zio)
6224 l2arc_read_callback_t *cb;
6227 kmutex_t *hash_lock;
6230 ASSERT(zio->io_vd != NULL);
6231 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
6233 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
6235 cb = zio->io_private;
6237 buf = cb->l2rcb_buf;
6238 ASSERT(buf != NULL);
6240 hash_lock = HDR_LOCK(buf->b_hdr);
6241 mutex_enter(hash_lock);
6243 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6246 * If the data was read into a temporary buffer,
6247 * move it and free the buffer.
6249 if (cb->l2rcb_data != NULL) {
6250 ASSERT3U(hdr->b_size, <, zio->io_size);
6251 ASSERT3U(cb->l2rcb_compress, ==, ZIO_COMPRESS_OFF);
6252 if (zio->io_error == 0)
6253 bcopy(cb->l2rcb_data, buf->b_data, hdr->b_size);
6256 * The following must be done regardless of whether
6257 * there was an error:
6258 * - free the temporary buffer
6259 * - point zio to the real ARC buffer
6260 * - set zio size accordingly
6261 * These are required because zio is either re-used for
6262 * an I/O of the block in the case of the error
6263 * or the zio is passed to arc_read_done() and it
6266 zio_data_buf_free(cb->l2rcb_data, zio->io_size);
6267 zio->io_size = zio->io_orig_size = hdr->b_size;
6268 zio->io_data = zio->io_orig_data = buf->b_data;
6272 * If the buffer was compressed, decompress it first.
6274 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
6275 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
6276 ASSERT(zio->io_data != NULL);
6277 ASSERT3U(zio->io_size, ==, hdr->b_size);
6278 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size);
6281 * Check this survived the L2ARC journey.
6283 equal = arc_cksum_equal(buf);
6284 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
6285 mutex_exit(hash_lock);
6286 zio->io_private = buf;
6287 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
6288 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
6291 mutex_exit(hash_lock);
6293 * Buffer didn't survive caching. Increment stats and
6294 * reissue to the original storage device.
6296 if (zio->io_error != 0) {
6297 ARCSTAT_BUMP(arcstat_l2_io_error);
6299 zio->io_error = SET_ERROR(EIO);
6302 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
6305 * If there's no waiter, issue an async i/o to the primary
6306 * storage now. If there *is* a waiter, the caller must
6307 * issue the i/o in a context where it's OK to block.
6309 if (zio->io_waiter == NULL) {
6310 zio_t *pio = zio_unique_parent(zio);
6312 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
6314 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
6315 buf->b_data, hdr->b_size, arc_read_done, buf,
6316 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
6320 kmem_free(cb, sizeof (l2arc_read_callback_t));
6324 * This is the list priority from which the L2ARC will search for pages to
6325 * cache. This is used within loops (0..3) to cycle through lists in the
6326 * desired order. This order can have a significant effect on cache
6329 * Currently the metadata lists are hit first, MFU then MRU, followed by
6330 * the data lists. This function returns a locked list, and also returns
6333 static multilist_sublist_t *
6334 l2arc_sublist_lock(int list_num)
6336 multilist_t *ml = NULL;
6339 ASSERT(list_num >= 0 && list_num <= 3);
6343 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
6346 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
6349 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
6352 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
6357 * Return a randomly-selected sublist. This is acceptable
6358 * because the caller feeds only a little bit of data for each
6359 * call (8MB). Subsequent calls will result in different
6360 * sublists being selected.
6362 idx = multilist_get_random_index(ml);
6363 return (multilist_sublist_lock(ml, idx));
6367 * Evict buffers from the device write hand to the distance specified in
6368 * bytes. This distance may span populated buffers, it may span nothing.
6369 * This is clearing a region on the L2ARC device ready for writing.
6370 * If the 'all' boolean is set, every buffer is evicted.
6373 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
6376 arc_buf_hdr_t *hdr, *hdr_prev;
6377 kmutex_t *hash_lock;
6380 buflist = &dev->l2ad_buflist;
6382 if (!all && dev->l2ad_first) {
6384 * This is the first sweep through the device. There is
6390 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
6392 * When nearing the end of the device, evict to the end
6393 * before the device write hand jumps to the start.
6395 taddr = dev->l2ad_end;
6397 taddr = dev->l2ad_hand + distance;
6399 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
6400 uint64_t, taddr, boolean_t, all);
6403 mutex_enter(&dev->l2ad_mtx);
6404 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
6405 hdr_prev = list_prev(buflist, hdr);
6407 hash_lock = HDR_LOCK(hdr);
6410 * We cannot use mutex_enter or else we can deadlock
6411 * with l2arc_write_buffers (due to swapping the order
6412 * the hash lock and l2ad_mtx are taken).
6414 if (!mutex_tryenter(hash_lock)) {
6416 * Missed the hash lock. Retry.
6418 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
6419 mutex_exit(&dev->l2ad_mtx);
6420 mutex_enter(hash_lock);
6421 mutex_exit(hash_lock);
6425 if (HDR_L2_WRITE_HEAD(hdr)) {
6427 * We hit a write head node. Leave it for
6428 * l2arc_write_done().
6430 list_remove(buflist, hdr);
6431 mutex_exit(hash_lock);
6435 if (!all && HDR_HAS_L2HDR(hdr) &&
6436 (hdr->b_l2hdr.b_daddr > taddr ||
6437 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
6439 * We've evicted to the target address,
6440 * or the end of the device.
6442 mutex_exit(hash_lock);
6446 ASSERT(HDR_HAS_L2HDR(hdr));
6447 if (!HDR_HAS_L1HDR(hdr)) {
6448 ASSERT(!HDR_L2_READING(hdr));
6450 * This doesn't exist in the ARC. Destroy.
6451 * arc_hdr_destroy() will call list_remove()
6452 * and decrement arcstat_l2_size.
6454 arc_change_state(arc_anon, hdr, hash_lock);
6455 arc_hdr_destroy(hdr);
6457 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
6458 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
6460 * Invalidate issued or about to be issued
6461 * reads, since we may be about to write
6462 * over this location.
6464 if (HDR_L2_READING(hdr)) {
6465 ARCSTAT_BUMP(arcstat_l2_evict_reading);
6466 hdr->b_flags |= ARC_FLAG_L2_EVICTED;
6469 /* Ensure this header has finished being written */
6470 ASSERT(!HDR_L2_WRITING(hdr));
6471 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
6473 arc_hdr_l2hdr_destroy(hdr);
6475 mutex_exit(hash_lock);
6477 mutex_exit(&dev->l2ad_mtx);
6481 * Find and write ARC buffers to the L2ARC device.
6483 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6484 * for reading until they have completed writing.
6485 * The headroom_boost is an in-out parameter used to maintain headroom boost
6486 * state between calls to this function.
6488 * Returns the number of bytes actually written (which may be smaller than
6489 * the delta by which the device hand has changed due to alignment).
6492 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
6493 boolean_t *headroom_boost)
6495 arc_buf_hdr_t *hdr, *hdr_prev, *head;
6496 uint64_t write_asize, write_sz, headroom,
6500 l2arc_write_callback_t *cb;
6502 uint64_t guid = spa_load_guid(spa);
6503 const boolean_t do_headroom_boost = *headroom_boost;
6506 ASSERT(dev->l2ad_vdev != NULL);
6508 /* Lower the flag now, we might want to raise it again later. */
6509 *headroom_boost = B_FALSE;
6512 write_sz = write_asize = 0;
6514 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
6515 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD;
6516 head->b_flags |= ARC_FLAG_HAS_L2HDR;
6518 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
6520 * We will want to try to compress buffers that are at least 2x the
6521 * device sector size.
6523 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
6526 * Copy buffers for L2ARC writing.
6528 for (try = 0; try <= 3; try++) {
6529 multilist_sublist_t *mls = l2arc_sublist_lock(try);
6530 uint64_t passed_sz = 0;
6532 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
6535 * L2ARC fast warmup.
6537 * Until the ARC is warm and starts to evict, read from the
6538 * head of the ARC lists rather than the tail.
6540 if (arc_warm == B_FALSE)
6541 hdr = multilist_sublist_head(mls);
6543 hdr = multilist_sublist_tail(mls);
6545 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
6547 headroom = target_sz * l2arc_headroom;
6548 if (do_headroom_boost)
6549 headroom = (headroom * l2arc_headroom_boost) / 100;
6551 for (; hdr; hdr = hdr_prev) {
6552 kmutex_t *hash_lock;
6557 if (arc_warm == B_FALSE)
6558 hdr_prev = multilist_sublist_next(mls, hdr);
6560 hdr_prev = multilist_sublist_prev(mls, hdr);
6561 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size);
6563 hash_lock = HDR_LOCK(hdr);
6564 if (!mutex_tryenter(hash_lock)) {
6565 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
6567 * Skip this buffer rather than waiting.
6572 passed_sz += hdr->b_size;
6573 if (passed_sz > headroom) {
6577 mutex_exit(hash_lock);
6578 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
6582 if (!l2arc_write_eligible(guid, hdr)) {
6583 mutex_exit(hash_lock);
6588 * Assume that the buffer is not going to be compressed
6589 * and could take more space on disk because of a larger
6592 buf_sz = hdr->b_size;
6593 align = (size_t)1 << dev->l2ad_vdev->vdev_ashift;
6594 buf_a_sz = P2ROUNDUP(buf_sz, align);
6596 if ((write_asize + buf_a_sz) > target_sz) {
6598 mutex_exit(hash_lock);
6599 ARCSTAT_BUMP(arcstat_l2_write_full);
6605 * Insert a dummy header on the buflist so
6606 * l2arc_write_done() can find where the
6607 * write buffers begin without searching.
6609 mutex_enter(&dev->l2ad_mtx);
6610 list_insert_head(&dev->l2ad_buflist, head);
6611 mutex_exit(&dev->l2ad_mtx);
6614 sizeof (l2arc_write_callback_t), KM_SLEEP);
6615 cb->l2wcb_dev = dev;
6616 cb->l2wcb_head = head;
6617 pio = zio_root(spa, l2arc_write_done, cb,
6619 ARCSTAT_BUMP(arcstat_l2_write_pios);
6623 * Create and add a new L2ARC header.
6625 hdr->b_l2hdr.b_dev = dev;
6626 hdr->b_flags |= ARC_FLAG_L2_WRITING;
6628 * Temporarily stash the data buffer in b_tmp_cdata.
6629 * The subsequent write step will pick it up from
6630 * there. This is because can't access b_l1hdr.b_buf
6631 * without holding the hash_lock, which we in turn
6632 * can't access without holding the ARC list locks
6633 * (which we want to avoid during compression/writing).
6635 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF;
6636 hdr->b_l2hdr.b_asize = hdr->b_size;
6637 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data;
6640 * Explicitly set the b_daddr field to a known
6641 * value which means "invalid address". This
6642 * enables us to differentiate which stage of
6643 * l2arc_write_buffers() the particular header
6644 * is in (e.g. this loop, or the one below).
6645 * ARC_FLAG_L2_WRITING is not enough to make
6646 * this distinction, and we need to know in
6647 * order to do proper l2arc vdev accounting in
6648 * arc_release() and arc_hdr_destroy().
6650 * Note, we can't use a new flag to distinguish
6651 * the two stages because we don't hold the
6652 * header's hash_lock below, in the second stage
6653 * of this function. Thus, we can't simply
6654 * change the b_flags field to denote that the
6655 * IO has been sent. We can change the b_daddr
6656 * field of the L2 portion, though, since we'll
6657 * be holding the l2ad_mtx; which is why we're
6658 * using it to denote the header's state change.
6660 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET;
6662 hdr->b_flags |= ARC_FLAG_HAS_L2HDR;
6664 mutex_enter(&dev->l2ad_mtx);
6665 list_insert_head(&dev->l2ad_buflist, hdr);
6666 mutex_exit(&dev->l2ad_mtx);
6669 * Compute and store the buffer cksum before
6670 * writing. On debug the cksum is verified first.
6672 arc_cksum_verify(hdr->b_l1hdr.b_buf);
6673 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE);
6675 mutex_exit(hash_lock);
6678 write_asize += buf_a_sz;
6681 multilist_sublist_unlock(mls);
6687 /* No buffers selected for writing? */
6690 ASSERT(!HDR_HAS_L1HDR(head));
6691 kmem_cache_free(hdr_l2only_cache, head);
6695 mutex_enter(&dev->l2ad_mtx);
6698 * Now start writing the buffers. We're starting at the write head
6699 * and work backwards, retracing the course of the buffer selector
6703 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr;
6704 hdr = list_prev(&dev->l2ad_buflist, hdr)) {
6709 * We rely on the L1 portion of the header below, so
6710 * it's invalid for this header to have been evicted out
6711 * of the ghost cache, prior to being written out. The
6712 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6714 ASSERT(HDR_HAS_L1HDR(hdr));
6717 * We shouldn't need to lock the buffer here, since we flagged
6718 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6719 * take care to only access its L2 cache parameters. In
6720 * particular, hdr->l1hdr.b_buf may be invalid by now due to
6723 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
6726 * Save a pointer to the original buffer data we had previously
6729 buf_data = hdr->b_l1hdr.b_tmp_cdata;
6731 compress = HDR_L2COMPRESS(hdr) &&
6732 hdr->b_l2hdr.b_asize >= buf_compress_minsz;
6733 if (l2arc_transform_buf(hdr, compress)) {
6735 * If compression succeeded, enable headroom
6736 * boost on the next scan cycle.
6738 *headroom_boost = B_TRUE;
6742 * Get the new buffer size that accounts for compression
6745 buf_sz = hdr->b_l2hdr.b_asize;
6748 * We need to do this regardless if buf_sz is zero or
6749 * not, otherwise, when this l2hdr is evicted we'll
6750 * remove a reference that was never added.
6752 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr);
6754 /* Compression may have squashed the buffer to zero length. */
6757 * If the data was padded or compressed, then it
6758 * it is in a new buffer.
6760 if (hdr->b_l1hdr.b_tmp_cdata != NULL)
6761 buf_data = hdr->b_l1hdr.b_tmp_cdata;
6762 wzio = zio_write_phys(pio, dev->l2ad_vdev,
6763 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
6764 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
6765 ZIO_FLAG_CANFAIL, B_FALSE);
6767 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
6769 (void) zio_nowait(wzio);
6771 write_asize += buf_sz;
6772 dev->l2ad_hand += buf_sz;
6776 mutex_exit(&dev->l2ad_mtx);
6778 ASSERT3U(write_asize, <=, target_sz);
6779 ARCSTAT_BUMP(arcstat_l2_writes_sent);
6780 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
6781 ARCSTAT_INCR(arcstat_l2_size, write_sz);
6782 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
6783 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
6786 * Bump device hand to the device start if it is approaching the end.
6787 * l2arc_evict() will already have evicted ahead for this case.
6789 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
6790 dev->l2ad_hand = dev->l2ad_start;
6791 dev->l2ad_first = B_FALSE;
6794 dev->l2ad_writing = B_TRUE;
6795 (void) zio_wait(pio);
6796 dev->l2ad_writing = B_FALSE;
6798 return (write_asize);
6802 * Transforms, possibly compresses and pads, an L2ARC buffer.
6803 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6804 * size in l2hdr->b_asize. This routine tries to compress the data and
6805 * depending on the compression result there are three possible outcomes:
6806 * *) The buffer was incompressible. The buffer size was already ashift aligned.
6807 * The original hdr contents were left untouched except for b_tmp_cdata,
6808 * which is reset to NULL. The caller must keep a pointer to the original
6810 * *) The buffer was incompressible. The buffer size was not ashift aligned.
6811 * b_tmp_cdata was replaced with a temporary data buffer which holds a padded
6812 * (aligned) copy of the data. Once writing is done, invoke
6813 * l2arc_release_cdata_buf on this hdr to free the temporary buffer.
6814 * *) The buffer was all-zeros, so there is no need to write it to an L2
6815 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6816 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6817 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6818 * data buffer which holds the compressed data to be written, and b_asize
6819 * tells us how much data there is. b_compress is set to the appropriate
6820 * compression algorithm. Once writing is done, invoke
6821 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6823 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6824 * buffer was incompressible).
6827 l2arc_transform_buf(arc_buf_hdr_t *hdr, boolean_t compress)
6830 size_t align, asize, csize, len, rounded;
6832 ASSERT(HDR_HAS_L2HDR(hdr));
6833 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
6835 ASSERT(HDR_HAS_L1HDR(hdr));
6836 ASSERT3S(l2hdr->b_compress, ==, ZIO_COMPRESS_OFF);
6837 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6839 len = l2hdr->b_asize;
6840 align = (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift;
6841 asize = P2ROUNDUP(len, align);
6842 cdata = zio_data_buf_alloc(asize);
6843 ASSERT3P(cdata, !=, NULL);
6845 csize = zio_compress_data(ZIO_COMPRESS_LZ4,
6846 hdr->b_l1hdr.b_tmp_cdata, cdata, len);
6851 /* zero block, indicate that there's nothing to write */
6852 zio_data_buf_free(cdata, asize);
6853 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
6855 hdr->b_l1hdr.b_tmp_cdata = NULL;
6856 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
6860 rounded = P2ROUNDUP(csize, align);
6861 ASSERT3U(rounded, <=, asize);
6862 if (rounded < len) {
6864 * Compression succeeded, we'll keep the cdata around for
6865 * writing and release it afterwards.
6867 if (rounded > csize) {
6868 bzero((char *)cdata + csize, rounded - csize);
6871 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
6872 l2hdr->b_asize = csize;
6873 hdr->b_l1hdr.b_tmp_cdata = cdata;
6874 ARCSTAT_BUMP(arcstat_l2_compress_successes);
6878 * Compression did not save space.
6880 if (P2PHASE(len, align) != 0) {
6882 * Use compression buffer for a copy of data padded to
6883 * the proper size. Compression algorithm remains set
6884 * to ZIO_COMPRESS_OFF.
6886 ASSERT3U(len, <, asize);
6887 bcopy(hdr->b_l1hdr.b_tmp_cdata, cdata, len);
6888 bzero((char *)cdata + len, asize - len);
6889 l2hdr->b_asize = asize;
6890 hdr->b_l1hdr.b_tmp_cdata = cdata;
6891 ARCSTAT_BUMP(arcstat_l2_padding_needed);
6893 ASSERT3U(len, ==, asize);
6895 * The original buffer is good as is,
6896 * release the compressed buffer.
6897 * l2hdr will be left unmodified except for b_tmp_cdata.
6899 zio_data_buf_free(cdata, asize);
6900 hdr->b_l1hdr.b_tmp_cdata = NULL;
6903 ARCSTAT_BUMP(arcstat_l2_compress_failures);
6909 * Decompresses a zio read back from an l2arc device. On success, the
6910 * underlying zio's io_data buffer is overwritten by the uncompressed
6911 * version. On decompression error (corrupt compressed stream), the
6912 * zio->io_error value is set to signal an I/O error.
6914 * Please note that the compressed data stream is not checksummed, so
6915 * if the underlying device is experiencing data corruption, we may feed
6916 * corrupt data to the decompressor, so the decompressor needs to be
6917 * able to handle this situation (LZ4 does).
6920 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
6922 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
6924 if (zio->io_error != 0) {
6926 * An io error has occured, just restore the original io
6927 * size in preparation for a main pool read.
6929 zio->io_orig_size = zio->io_size = hdr->b_size;
6933 if (c == ZIO_COMPRESS_EMPTY) {
6935 * An empty buffer results in a null zio, which means we
6936 * need to fill its io_data after we're done restoring the
6937 * buffer's contents.
6939 ASSERT(hdr->b_l1hdr.b_buf != NULL);
6940 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size);
6941 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data;
6943 ASSERT(zio->io_data != NULL);
6945 * We copy the compressed data from the start of the arc buffer
6946 * (the zio_read will have pulled in only what we need, the
6947 * rest is garbage which we will overwrite at decompression)
6948 * and then decompress back to the ARC data buffer. This way we
6949 * can minimize copying by simply decompressing back over the
6950 * original compressed data (rather than decompressing to an
6951 * aux buffer and then copying back the uncompressed buffer,
6952 * which is likely to be much larger).
6957 csize = zio->io_size;
6958 cdata = zio_data_buf_alloc(csize);
6959 bcopy(zio->io_data, cdata, csize);
6960 if (zio_decompress_data(c, cdata, zio->io_data, csize,
6962 zio->io_error = EIO;
6963 zio_data_buf_free(cdata, csize);
6966 /* Restore the expected uncompressed IO size. */
6967 zio->io_orig_size = zio->io_size = hdr->b_size;
6971 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6972 * This buffer serves as a temporary holder of compressed or padded data while
6973 * the buffer entry is being written to an l2arc device. Once that is
6974 * done, we can dispose of it.
6977 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr)
6979 size_t align, asize, len;
6980 enum zio_compress comp = hdr->b_l2hdr.b_compress;
6982 ASSERT(HDR_HAS_L2HDR(hdr));
6983 ASSERT(HDR_HAS_L1HDR(hdr));
6984 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp));
6986 if (hdr->b_l1hdr.b_tmp_cdata != NULL) {
6987 ASSERT(comp != ZIO_COMPRESS_EMPTY);
6989 align = (size_t)1 << hdr->b_l2hdr.b_dev->l2ad_vdev->vdev_ashift;
6990 asize = P2ROUNDUP(len, align);
6991 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, asize);
6992 hdr->b_l1hdr.b_tmp_cdata = NULL;
6994 ASSERT(comp == ZIO_COMPRESS_OFF || comp == ZIO_COMPRESS_EMPTY);
6999 * This thread feeds the L2ARC at regular intervals. This is the beating
7000 * heart of the L2ARC.
7003 l2arc_feed_thread(void *dummy __unused)
7008 uint64_t size, wrote;
7009 clock_t begin, next = ddi_get_lbolt();
7010 boolean_t headroom_boost = B_FALSE;
7012 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7014 mutex_enter(&l2arc_feed_thr_lock);
7016 while (l2arc_thread_exit == 0) {
7017 CALLB_CPR_SAFE_BEGIN(&cpr);
7018 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7019 next - ddi_get_lbolt());
7020 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7021 next = ddi_get_lbolt() + hz;
7024 * Quick check for L2ARC devices.
7026 mutex_enter(&l2arc_dev_mtx);
7027 if (l2arc_ndev == 0) {
7028 mutex_exit(&l2arc_dev_mtx);
7031 mutex_exit(&l2arc_dev_mtx);
7032 begin = ddi_get_lbolt();
7035 * This selects the next l2arc device to write to, and in
7036 * doing so the next spa to feed from: dev->l2ad_spa. This
7037 * will return NULL if there are now no l2arc devices or if
7038 * they are all faulted.
7040 * If a device is returned, its spa's config lock is also
7041 * held to prevent device removal. l2arc_dev_get_next()
7042 * will grab and release l2arc_dev_mtx.
7044 if ((dev = l2arc_dev_get_next()) == NULL)
7047 spa = dev->l2ad_spa;
7048 ASSERT(spa != NULL);
7051 * If the pool is read-only then force the feed thread to
7052 * sleep a little longer.
7054 if (!spa_writeable(spa)) {
7055 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7056 spa_config_exit(spa, SCL_L2ARC, dev);
7061 * Avoid contributing to memory pressure.
7063 if (arc_reclaim_needed()) {
7064 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7065 spa_config_exit(spa, SCL_L2ARC, dev);
7069 ARCSTAT_BUMP(arcstat_l2_feeds);
7071 size = l2arc_write_size();
7074 * Evict L2ARC buffers that will be overwritten.
7076 l2arc_evict(dev, size, B_FALSE);
7079 * Write ARC buffers.
7081 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
7084 * Calculate interval between writes.
7086 next = l2arc_write_interval(begin, size, wrote);
7087 spa_config_exit(spa, SCL_L2ARC, dev);
7090 l2arc_thread_exit = 0;
7091 cv_broadcast(&l2arc_feed_thr_cv);
7092 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7097 l2arc_vdev_present(vdev_t *vd)
7101 mutex_enter(&l2arc_dev_mtx);
7102 for (dev = list_head(l2arc_dev_list); dev != NULL;
7103 dev = list_next(l2arc_dev_list, dev)) {
7104 if (dev->l2ad_vdev == vd)
7107 mutex_exit(&l2arc_dev_mtx);
7109 return (dev != NULL);
7113 * Add a vdev for use by the L2ARC. By this point the spa has already
7114 * validated the vdev and opened it.
7117 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7119 l2arc_dev_t *adddev;
7121 ASSERT(!l2arc_vdev_present(vd));
7123 vdev_ashift_optimize(vd);
7126 * Create a new l2arc device entry.
7128 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7129 adddev->l2ad_spa = spa;
7130 adddev->l2ad_vdev = vd;
7131 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7132 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7133 adddev->l2ad_hand = adddev->l2ad_start;
7134 adddev->l2ad_first = B_TRUE;
7135 adddev->l2ad_writing = B_FALSE;
7137 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7139 * This is a list of all ARC buffers that are still valid on the
7142 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7143 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7145 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7146 refcount_create(&adddev->l2ad_alloc);
7149 * Add device to global list
7151 mutex_enter(&l2arc_dev_mtx);
7152 list_insert_head(l2arc_dev_list, adddev);
7153 atomic_inc_64(&l2arc_ndev);
7154 mutex_exit(&l2arc_dev_mtx);
7158 * Remove a vdev from the L2ARC.
7161 l2arc_remove_vdev(vdev_t *vd)
7163 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7166 * Find the device by vdev
7168 mutex_enter(&l2arc_dev_mtx);
7169 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7170 nextdev = list_next(l2arc_dev_list, dev);
7171 if (vd == dev->l2ad_vdev) {
7176 ASSERT(remdev != NULL);
7179 * Remove device from global list
7181 list_remove(l2arc_dev_list, remdev);
7182 l2arc_dev_last = NULL; /* may have been invalidated */
7183 atomic_dec_64(&l2arc_ndev);
7184 mutex_exit(&l2arc_dev_mtx);
7187 * Clear all buflists and ARC references. L2ARC device flush.
7189 l2arc_evict(remdev, 0, B_TRUE);
7190 list_destroy(&remdev->l2ad_buflist);
7191 mutex_destroy(&remdev->l2ad_mtx);
7192 refcount_destroy(&remdev->l2ad_alloc);
7193 kmem_free(remdev, sizeof (l2arc_dev_t));
7199 l2arc_thread_exit = 0;
7201 l2arc_writes_sent = 0;
7202 l2arc_writes_done = 0;
7204 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7205 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7206 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7207 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7209 l2arc_dev_list = &L2ARC_dev_list;
7210 l2arc_free_on_write = &L2ARC_free_on_write;
7211 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7212 offsetof(l2arc_dev_t, l2ad_node));
7213 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7214 offsetof(l2arc_data_free_t, l2df_list_node));
7221 * This is called from dmu_fini(), which is called from spa_fini();
7222 * Because of this, we can assume that all l2arc devices have
7223 * already been removed when the pools themselves were removed.
7226 l2arc_do_free_on_write();
7228 mutex_destroy(&l2arc_feed_thr_lock);
7229 cv_destroy(&l2arc_feed_thr_cv);
7230 mutex_destroy(&l2arc_dev_mtx);
7231 mutex_destroy(&l2arc_free_on_write_mtx);
7233 list_destroy(l2arc_dev_list);
7234 list_destroy(l2arc_free_on_write);
7240 if (!(spa_mode_global & FWRITE))
7243 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7244 TS_RUN, minclsyspri);
7250 if (!(spa_mode_global & FWRITE))
7253 mutex_enter(&l2arc_feed_thr_lock);
7254 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7255 l2arc_thread_exit = 1;
7256 while (l2arc_thread_exit != 0)
7257 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7258 mutex_exit(&l2arc_feed_thr_lock);