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, 2015 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>
136 #include <sys/callb.h>
137 #include <sys/kstat.h>
138 #include <sys/trim_map.h>
139 #include <zfs_fletcher.h>
142 #include <machine/vmparam.h>
146 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
147 boolean_t arc_watch = B_FALSE;
152 static kmutex_t arc_reclaim_lock;
153 static kcondvar_t arc_reclaim_thread_cv;
154 static boolean_t arc_reclaim_thread_exit;
155 static kcondvar_t arc_reclaim_waiters_cv;
157 static kmutex_t arc_user_evicts_lock;
158 static kcondvar_t arc_user_evicts_cv;
159 static boolean_t arc_user_evicts_thread_exit;
161 uint_t arc_reduce_dnlc_percent = 3;
164 * The number of headers to evict in arc_evict_state_impl() before
165 * dropping the sublist lock and evicting from another sublist. A lower
166 * value means we're more likely to evict the "correct" header (i.e. the
167 * oldest header in the arc state), but comes with higher overhead
168 * (i.e. more invocations of arc_evict_state_impl()).
170 int zfs_arc_evict_batch_limit = 10;
173 * The number of sublists used for each of the arc state lists. If this
174 * is not set to a suitable value by the user, it will be configured to
175 * the number of CPUs on the system in arc_init().
177 int zfs_arc_num_sublists_per_state = 0;
179 /* number of seconds before growing cache again */
180 static int arc_grow_retry = 60;
182 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
183 int zfs_arc_overflow_shift = 8;
185 /* shift of arc_c for calculating both min and max arc_p */
186 static int arc_p_min_shift = 4;
188 /* log2(fraction of arc to reclaim) */
189 static int arc_shrink_shift = 7;
192 * log2(fraction of ARC which must be free to allow growing).
193 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
194 * when reading a new block into the ARC, we will evict an equal-sized block
197 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
198 * we will still not allow it to grow.
200 int arc_no_grow_shift = 5;
204 * minimum lifespan of a prefetch block in clock ticks
205 * (initialized in arc_init())
207 static int arc_min_prefetch_lifespan;
210 * If this percent of memory is free, don't throttle.
212 int arc_lotsfree_percent = 10;
215 extern boolean_t zfs_prefetch_disable;
218 * The arc has filled available memory and has now warmed up.
220 static boolean_t arc_warm;
223 * These tunables are for performance analysis.
225 uint64_t zfs_arc_max;
226 uint64_t zfs_arc_min;
227 uint64_t zfs_arc_meta_limit = 0;
228 uint64_t zfs_arc_meta_min = 0;
229 int zfs_arc_grow_retry = 0;
230 int zfs_arc_shrink_shift = 0;
231 int zfs_arc_p_min_shift = 0;
232 int zfs_disable_dup_eviction = 0;
233 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
234 u_int zfs_arc_free_target = 0;
236 /* Absolute min for arc min / max is 16MB. */
237 static uint64_t arc_abs_min = 16 << 20;
239 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
240 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
241 static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS);
242 static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS);
244 #if defined(__FreeBSD__) && defined(_KERNEL)
246 arc_free_target_init(void *unused __unused)
249 zfs_arc_free_target = vm_pageout_wakeup_thresh;
251 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
252 arc_free_target_init, NULL);
254 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
255 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
256 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
257 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
258 TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize);
259 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
260 SYSCTL_DECL(_vfs_zfs);
261 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN,
262 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size");
263 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN,
264 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size");
265 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
266 &zfs_arc_average_blocksize, 0,
267 "ARC average blocksize");
268 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
269 &arc_shrink_shift, 0,
270 "log2(fraction of arc to reclaim)");
273 * We don't have a tunable for arc_free_target due to the dependency on
274 * pagedaemon initialisation.
276 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
277 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
278 sysctl_vfs_zfs_arc_free_target, "IU",
279 "Desired number of free pages below which ARC triggers reclaim");
282 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
287 val = zfs_arc_free_target;
288 err = sysctl_handle_int(oidp, &val, 0, req);
289 if (err != 0 || req->newptr == NULL)
294 if (val > cnt.v_page_count)
297 zfs_arc_free_target = val;
303 * Must be declared here, before the definition of corresponding kstat
304 * macro which uses the same names will confuse the compiler.
306 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
307 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
308 sysctl_vfs_zfs_arc_meta_limit, "QU",
309 "ARC metadata limit");
313 * Note that buffers can be in one of 6 states:
314 * ARC_anon - anonymous (discussed below)
315 * ARC_mru - recently used, currently cached
316 * ARC_mru_ghost - recentely used, no longer in cache
317 * ARC_mfu - frequently used, currently cached
318 * ARC_mfu_ghost - frequently used, no longer in cache
319 * ARC_l2c_only - exists in L2ARC but not other states
320 * When there are no active references to the buffer, they are
321 * are linked onto a list in one of these arc states. These are
322 * the only buffers that can be evicted or deleted. Within each
323 * state there are multiple lists, one for meta-data and one for
324 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
325 * etc.) is tracked separately so that it can be managed more
326 * explicitly: favored over data, limited explicitly.
328 * Anonymous buffers are buffers that are not associated with
329 * a DVA. These are buffers that hold dirty block copies
330 * before they are written to stable storage. By definition,
331 * they are "ref'd" and are considered part of arc_mru
332 * that cannot be freed. Generally, they will aquire a DVA
333 * as they are written and migrate onto the arc_mru list.
335 * The ARC_l2c_only state is for buffers that are in the second
336 * level ARC but no longer in any of the ARC_m* lists. The second
337 * level ARC itself may also contain buffers that are in any of
338 * the ARC_m* states - meaning that a buffer can exist in two
339 * places. The reason for the ARC_l2c_only state is to keep the
340 * buffer header in the hash table, so that reads that hit the
341 * second level ARC benefit from these fast lookups.
344 typedef struct arc_state {
346 * list of evictable buffers
348 multilist_t arcs_list[ARC_BUFC_NUMTYPES];
350 * total amount of evictable data in this state
352 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES];
354 * total amount of data in this state; this includes: evictable,
355 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
357 refcount_t arcs_size;
361 static arc_state_t ARC_anon;
362 static arc_state_t ARC_mru;
363 static arc_state_t ARC_mru_ghost;
364 static arc_state_t ARC_mfu;
365 static arc_state_t ARC_mfu_ghost;
366 static arc_state_t ARC_l2c_only;
368 typedef struct arc_stats {
369 kstat_named_t arcstat_hits;
370 kstat_named_t arcstat_misses;
371 kstat_named_t arcstat_demand_data_hits;
372 kstat_named_t arcstat_demand_data_misses;
373 kstat_named_t arcstat_demand_metadata_hits;
374 kstat_named_t arcstat_demand_metadata_misses;
375 kstat_named_t arcstat_prefetch_data_hits;
376 kstat_named_t arcstat_prefetch_data_misses;
377 kstat_named_t arcstat_prefetch_metadata_hits;
378 kstat_named_t arcstat_prefetch_metadata_misses;
379 kstat_named_t arcstat_mru_hits;
380 kstat_named_t arcstat_mru_ghost_hits;
381 kstat_named_t arcstat_mfu_hits;
382 kstat_named_t arcstat_mfu_ghost_hits;
383 kstat_named_t arcstat_allocated;
384 kstat_named_t arcstat_deleted;
386 * Number of buffers that could not be evicted because the hash lock
387 * was held by another thread. The lock may not necessarily be held
388 * by something using the same buffer, since hash locks are shared
389 * by multiple buffers.
391 kstat_named_t arcstat_mutex_miss;
393 * Number of buffers skipped because they have I/O in progress, are
394 * indrect prefetch buffers that have not lived long enough, or are
395 * not from the spa we're trying to evict from.
397 kstat_named_t arcstat_evict_skip;
399 * Number of times arc_evict_state() was unable to evict enough
400 * buffers to reach it's target amount.
402 kstat_named_t arcstat_evict_not_enough;
403 kstat_named_t arcstat_evict_l2_cached;
404 kstat_named_t arcstat_evict_l2_eligible;
405 kstat_named_t arcstat_evict_l2_ineligible;
406 kstat_named_t arcstat_evict_l2_skip;
407 kstat_named_t arcstat_hash_elements;
408 kstat_named_t arcstat_hash_elements_max;
409 kstat_named_t arcstat_hash_collisions;
410 kstat_named_t arcstat_hash_chains;
411 kstat_named_t arcstat_hash_chain_max;
412 kstat_named_t arcstat_p;
413 kstat_named_t arcstat_c;
414 kstat_named_t arcstat_c_min;
415 kstat_named_t arcstat_c_max;
416 kstat_named_t arcstat_size;
418 * Number of bytes consumed by internal ARC structures necessary
419 * for tracking purposes; these structures are not actually
420 * backed by ARC buffers. This includes arc_buf_hdr_t structures
421 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
422 * caches), and arc_buf_t structures (allocated via arc_buf_t
425 kstat_named_t arcstat_hdr_size;
427 * Number of bytes consumed by ARC buffers of type equal to
428 * ARC_BUFC_DATA. This is generally consumed by buffers backing
429 * on disk user data (e.g. plain file contents).
431 kstat_named_t arcstat_data_size;
433 * Number of bytes consumed by ARC buffers of type equal to
434 * ARC_BUFC_METADATA. This is generally consumed by buffers
435 * backing on disk data that is used for internal ZFS
436 * structures (e.g. ZAP, dnode, indirect blocks, etc).
438 kstat_named_t arcstat_metadata_size;
440 * Number of bytes consumed by various buffers and structures
441 * not actually backed with ARC buffers. This includes bonus
442 * buffers (allocated directly via zio_buf_* functions),
443 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
444 * cache), and dnode_t structures (allocated via dnode_t cache).
446 kstat_named_t arcstat_other_size;
448 * Total number of bytes consumed by ARC buffers residing in the
449 * arc_anon state. This includes *all* buffers in the arc_anon
450 * state; e.g. data, metadata, evictable, and unevictable buffers
451 * are all included in this value.
453 kstat_named_t arcstat_anon_size;
455 * Number of bytes consumed by ARC buffers that meet the
456 * following criteria: backing buffers of type ARC_BUFC_DATA,
457 * residing in the arc_anon state, and are eligible for eviction
458 * (e.g. have no outstanding holds on the buffer).
460 kstat_named_t arcstat_anon_evictable_data;
462 * Number of bytes consumed by ARC buffers that meet the
463 * following criteria: backing buffers of type ARC_BUFC_METADATA,
464 * residing in the arc_anon state, and are eligible for eviction
465 * (e.g. have no outstanding holds on the buffer).
467 kstat_named_t arcstat_anon_evictable_metadata;
469 * Total number of bytes consumed by ARC buffers residing in the
470 * arc_mru state. This includes *all* buffers in the arc_mru
471 * state; e.g. data, metadata, evictable, and unevictable buffers
472 * are all included in this value.
474 kstat_named_t arcstat_mru_size;
476 * Number of bytes consumed by ARC buffers that meet the
477 * following criteria: backing buffers of type ARC_BUFC_DATA,
478 * residing in the arc_mru state, and are eligible for eviction
479 * (e.g. have no outstanding holds on the buffer).
481 kstat_named_t arcstat_mru_evictable_data;
483 * Number of bytes consumed by ARC buffers that meet the
484 * following criteria: backing buffers of type ARC_BUFC_METADATA,
485 * residing in the arc_mru state, and are eligible for eviction
486 * (e.g. have no outstanding holds on the buffer).
488 kstat_named_t arcstat_mru_evictable_metadata;
490 * Total number of bytes that *would have been* consumed by ARC
491 * buffers in the arc_mru_ghost state. The key thing to note
492 * here, is the fact that this size doesn't actually indicate
493 * RAM consumption. The ghost lists only consist of headers and
494 * don't actually have ARC buffers linked off of these headers.
495 * Thus, *if* the headers had associated ARC buffers, these
496 * buffers *would have* consumed this number of bytes.
498 kstat_named_t arcstat_mru_ghost_size;
500 * Number of bytes that *would have been* consumed by ARC
501 * buffers that are eligible for eviction, of type
502 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
504 kstat_named_t arcstat_mru_ghost_evictable_data;
506 * Number of bytes that *would have been* consumed by ARC
507 * buffers that are eligible for eviction, of type
508 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
510 kstat_named_t arcstat_mru_ghost_evictable_metadata;
512 * Total number of bytes consumed by ARC buffers residing in the
513 * arc_mfu state. This includes *all* buffers in the arc_mfu
514 * state; e.g. data, metadata, evictable, and unevictable buffers
515 * are all included in this value.
517 kstat_named_t arcstat_mfu_size;
519 * Number of bytes consumed by ARC buffers that are eligible for
520 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
523 kstat_named_t arcstat_mfu_evictable_data;
525 * Number of bytes consumed by ARC buffers that are eligible for
526 * eviction, of type ARC_BUFC_METADATA, and reside in the
529 kstat_named_t arcstat_mfu_evictable_metadata;
531 * Total number of bytes that *would have been* consumed by ARC
532 * buffers in the arc_mfu_ghost state. See the comment above
533 * arcstat_mru_ghost_size for more details.
535 kstat_named_t arcstat_mfu_ghost_size;
537 * Number of bytes that *would have been* consumed by ARC
538 * buffers that are eligible for eviction, of type
539 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
541 kstat_named_t arcstat_mfu_ghost_evictable_data;
543 * Number of bytes that *would have been* consumed by ARC
544 * buffers that are eligible for eviction, of type
545 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
547 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
548 kstat_named_t arcstat_l2_hits;
549 kstat_named_t arcstat_l2_misses;
550 kstat_named_t arcstat_l2_feeds;
551 kstat_named_t arcstat_l2_rw_clash;
552 kstat_named_t arcstat_l2_read_bytes;
553 kstat_named_t arcstat_l2_write_bytes;
554 kstat_named_t arcstat_l2_writes_sent;
555 kstat_named_t arcstat_l2_writes_done;
556 kstat_named_t arcstat_l2_writes_error;
557 kstat_named_t arcstat_l2_writes_lock_retry;
558 kstat_named_t arcstat_l2_evict_lock_retry;
559 kstat_named_t arcstat_l2_evict_reading;
560 kstat_named_t arcstat_l2_evict_l1cached;
561 kstat_named_t arcstat_l2_free_on_write;
562 kstat_named_t arcstat_l2_cdata_free_on_write;
563 kstat_named_t arcstat_l2_abort_lowmem;
564 kstat_named_t arcstat_l2_cksum_bad;
565 kstat_named_t arcstat_l2_io_error;
566 kstat_named_t arcstat_l2_size;
567 kstat_named_t arcstat_l2_asize;
568 kstat_named_t arcstat_l2_hdr_size;
569 kstat_named_t arcstat_l2_compress_successes;
570 kstat_named_t arcstat_l2_compress_zeros;
571 kstat_named_t arcstat_l2_compress_failures;
572 kstat_named_t arcstat_l2_padding_needed;
573 kstat_named_t arcstat_l2_write_trylock_fail;
574 kstat_named_t arcstat_l2_write_passed_headroom;
575 kstat_named_t arcstat_l2_write_spa_mismatch;
576 kstat_named_t arcstat_l2_write_in_l2;
577 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
578 kstat_named_t arcstat_l2_write_not_cacheable;
579 kstat_named_t arcstat_l2_write_full;
580 kstat_named_t arcstat_l2_write_buffer_iter;
581 kstat_named_t arcstat_l2_write_pios;
582 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
583 kstat_named_t arcstat_l2_write_buffer_list_iter;
584 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
585 kstat_named_t arcstat_memory_throttle_count;
586 kstat_named_t arcstat_duplicate_buffers;
587 kstat_named_t arcstat_duplicate_buffers_size;
588 kstat_named_t arcstat_duplicate_reads;
589 kstat_named_t arcstat_meta_used;
590 kstat_named_t arcstat_meta_limit;
591 kstat_named_t arcstat_meta_max;
592 kstat_named_t arcstat_meta_min;
593 kstat_named_t arcstat_sync_wait_for_async;
594 kstat_named_t arcstat_demand_hit_predictive_prefetch;
597 static arc_stats_t arc_stats = {
598 { "hits", KSTAT_DATA_UINT64 },
599 { "misses", KSTAT_DATA_UINT64 },
600 { "demand_data_hits", KSTAT_DATA_UINT64 },
601 { "demand_data_misses", KSTAT_DATA_UINT64 },
602 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
603 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
604 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
605 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
606 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
607 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
608 { "mru_hits", KSTAT_DATA_UINT64 },
609 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
610 { "mfu_hits", KSTAT_DATA_UINT64 },
611 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
612 { "allocated", KSTAT_DATA_UINT64 },
613 { "deleted", KSTAT_DATA_UINT64 },
614 { "mutex_miss", KSTAT_DATA_UINT64 },
615 { "evict_skip", KSTAT_DATA_UINT64 },
616 { "evict_not_enough", KSTAT_DATA_UINT64 },
617 { "evict_l2_cached", KSTAT_DATA_UINT64 },
618 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
619 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
620 { "evict_l2_skip", KSTAT_DATA_UINT64 },
621 { "hash_elements", KSTAT_DATA_UINT64 },
622 { "hash_elements_max", KSTAT_DATA_UINT64 },
623 { "hash_collisions", KSTAT_DATA_UINT64 },
624 { "hash_chains", KSTAT_DATA_UINT64 },
625 { "hash_chain_max", KSTAT_DATA_UINT64 },
626 { "p", KSTAT_DATA_UINT64 },
627 { "c", KSTAT_DATA_UINT64 },
628 { "c_min", KSTAT_DATA_UINT64 },
629 { "c_max", KSTAT_DATA_UINT64 },
630 { "size", KSTAT_DATA_UINT64 },
631 { "hdr_size", KSTAT_DATA_UINT64 },
632 { "data_size", KSTAT_DATA_UINT64 },
633 { "metadata_size", KSTAT_DATA_UINT64 },
634 { "other_size", KSTAT_DATA_UINT64 },
635 { "anon_size", KSTAT_DATA_UINT64 },
636 { "anon_evictable_data", KSTAT_DATA_UINT64 },
637 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
638 { "mru_size", KSTAT_DATA_UINT64 },
639 { "mru_evictable_data", KSTAT_DATA_UINT64 },
640 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
641 { "mru_ghost_size", KSTAT_DATA_UINT64 },
642 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
643 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
644 { "mfu_size", KSTAT_DATA_UINT64 },
645 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
646 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
647 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
648 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
649 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
650 { "l2_hits", KSTAT_DATA_UINT64 },
651 { "l2_misses", KSTAT_DATA_UINT64 },
652 { "l2_feeds", KSTAT_DATA_UINT64 },
653 { "l2_rw_clash", KSTAT_DATA_UINT64 },
654 { "l2_read_bytes", KSTAT_DATA_UINT64 },
655 { "l2_write_bytes", KSTAT_DATA_UINT64 },
656 { "l2_writes_sent", KSTAT_DATA_UINT64 },
657 { "l2_writes_done", KSTAT_DATA_UINT64 },
658 { "l2_writes_error", KSTAT_DATA_UINT64 },
659 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
660 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
661 { "l2_evict_reading", KSTAT_DATA_UINT64 },
662 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
663 { "l2_free_on_write", KSTAT_DATA_UINT64 },
664 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 },
665 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
666 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
667 { "l2_io_error", KSTAT_DATA_UINT64 },
668 { "l2_size", KSTAT_DATA_UINT64 },
669 { "l2_asize", KSTAT_DATA_UINT64 },
670 { "l2_hdr_size", KSTAT_DATA_UINT64 },
671 { "l2_compress_successes", KSTAT_DATA_UINT64 },
672 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
673 { "l2_compress_failures", KSTAT_DATA_UINT64 },
674 { "l2_padding_needed", KSTAT_DATA_UINT64 },
675 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
676 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
677 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
678 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
679 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
680 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
681 { "l2_write_full", KSTAT_DATA_UINT64 },
682 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
683 { "l2_write_pios", KSTAT_DATA_UINT64 },
684 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
685 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
686 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
687 { "memory_throttle_count", KSTAT_DATA_UINT64 },
688 { "duplicate_buffers", KSTAT_DATA_UINT64 },
689 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
690 { "duplicate_reads", KSTAT_DATA_UINT64 },
691 { "arc_meta_used", KSTAT_DATA_UINT64 },
692 { "arc_meta_limit", KSTAT_DATA_UINT64 },
693 { "arc_meta_max", KSTAT_DATA_UINT64 },
694 { "arc_meta_min", KSTAT_DATA_UINT64 },
695 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
696 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
699 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
701 #define ARCSTAT_INCR(stat, val) \
702 atomic_add_64(&arc_stats.stat.value.ui64, (val))
704 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
705 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
707 #define ARCSTAT_MAX(stat, val) { \
709 while ((val) > (m = arc_stats.stat.value.ui64) && \
710 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
714 #define ARCSTAT_MAXSTAT(stat) \
715 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
718 * We define a macro to allow ARC hits/misses to be easily broken down by
719 * two separate conditions, giving a total of four different subtypes for
720 * each of hits and misses (so eight statistics total).
722 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
725 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
727 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
731 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
733 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
738 static arc_state_t *arc_anon;
739 static arc_state_t *arc_mru;
740 static arc_state_t *arc_mru_ghost;
741 static arc_state_t *arc_mfu;
742 static arc_state_t *arc_mfu_ghost;
743 static arc_state_t *arc_l2c_only;
746 * There are several ARC variables that are critical to export as kstats --
747 * but we don't want to have to grovel around in the kstat whenever we wish to
748 * manipulate them. For these variables, we therefore define them to be in
749 * terms of the statistic variable. This assures that we are not introducing
750 * the possibility of inconsistency by having shadow copies of the variables,
751 * while still allowing the code to be readable.
753 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
754 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
755 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
756 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
757 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
758 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
759 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
760 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
761 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
763 #define L2ARC_IS_VALID_COMPRESS(_c_) \
764 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
766 static int arc_no_grow; /* Don't try to grow cache size */
767 static uint64_t arc_tempreserve;
768 static uint64_t arc_loaned_bytes;
770 typedef struct arc_callback arc_callback_t;
772 struct arc_callback {
774 arc_done_func_t *acb_done;
776 zio_t *acb_zio_dummy;
777 arc_callback_t *acb_next;
780 typedef struct arc_write_callback arc_write_callback_t;
782 struct arc_write_callback {
784 arc_done_func_t *awcb_ready;
785 arc_done_func_t *awcb_physdone;
786 arc_done_func_t *awcb_done;
791 * ARC buffers are separated into multiple structs as a memory saving measure:
792 * - Common fields struct, always defined, and embedded within it:
793 * - L2-only fields, always allocated but undefined when not in L2ARC
794 * - L1-only fields, only allocated when in L1ARC
796 * Buffer in L1 Buffer only in L2
797 * +------------------------+ +------------------------+
798 * | arc_buf_hdr_t | | arc_buf_hdr_t |
802 * +------------------------+ +------------------------+
803 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
804 * | (undefined if L1-only) | | |
805 * +------------------------+ +------------------------+
806 * | l1arc_buf_hdr_t |
811 * +------------------------+
813 * Because it's possible for the L2ARC to become extremely large, we can wind
814 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
815 * is minimized by only allocating the fields necessary for an L1-cached buffer
816 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
817 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
818 * words in pointers. arc_hdr_realloc() is used to switch a header between
819 * these two allocation states.
821 typedef struct l1arc_buf_hdr {
822 kmutex_t b_freeze_lock;
825 * used for debugging wtih kmem_flags - by allocating and freeing
826 * b_thawed when the buffer is thawed, we get a record of the stack
827 * trace that thawed it.
834 /* for waiting on writes to complete */
837 /* protected by arc state mutex */
838 arc_state_t *b_state;
839 multilist_node_t b_arc_node;
841 /* updated atomically */
842 clock_t b_arc_access;
844 /* self protecting */
847 arc_callback_t *b_acb;
848 /* temporary buffer holder for in-flight compressed or padded data */
852 typedef struct l2arc_dev l2arc_dev_t;
854 typedef struct l2arc_buf_hdr {
855 /* protected by arc_buf_hdr mutex */
856 l2arc_dev_t *b_dev; /* L2ARC device */
857 uint64_t b_daddr; /* disk address, offset byte */
858 /* real alloc'd buffer size depending on b_compress applied */
862 list_node_t b_l2node;
866 /* protected by hash lock */
870 * Even though this checksum is only set/verified when a buffer is in
871 * the L1 cache, it needs to be in the set of common fields because it
872 * must be preserved from the time before a buffer is written out to
873 * L2ARC until after it is read back in.
875 zio_cksum_t *b_freeze_cksum;
877 arc_buf_hdr_t *b_hash_next;
884 /* L2ARC fields. Undefined when not in L2ARC. */
885 l2arc_buf_hdr_t b_l2hdr;
886 /* L1ARC fields. Undefined when in l2arc_only state */
887 l1arc_buf_hdr_t b_l1hdr;
890 #if defined(__FreeBSD__) && defined(_KERNEL)
892 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
897 val = arc_meta_limit;
898 err = sysctl_handle_64(oidp, &val, 0, req);
899 if (err != 0 || req->newptr == NULL)
902 if (val <= 0 || val > arc_c_max)
905 arc_meta_limit = val;
910 sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS)
916 err = sysctl_handle_64(oidp, &val, 0, req);
917 if (err != 0 || req->newptr == NULL)
920 if (zfs_arc_max == 0) {
921 /* Loader tunable so blindly set */
926 if (val < arc_abs_min || val > kmem_size())
930 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit)
936 arc_p = (arc_c >> 1);
938 if (zfs_arc_meta_limit == 0) {
939 /* limit meta-data to 1/4 of the arc capacity */
940 arc_meta_limit = arc_c_max / 4;
943 /* if kmem_flags are set, lets try to use less memory */
944 if (kmem_debugging())
953 sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS)
959 err = sysctl_handle_64(oidp, &val, 0, req);
960 if (err != 0 || req->newptr == NULL)
963 if (zfs_arc_min == 0) {
964 /* Loader tunable so blindly set */
969 if (val < arc_abs_min || val > arc_c_max)
974 if (zfs_arc_meta_min == 0)
975 arc_meta_min = arc_c_min / 2;
977 if (arc_c < arc_c_min)
980 zfs_arc_min = arc_c_min;
986 static arc_buf_t *arc_eviction_list;
987 static arc_buf_hdr_t arc_eviction_hdr;
989 #define GHOST_STATE(state) \
990 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
991 (state) == arc_l2c_only)
993 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
994 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
995 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
996 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
997 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
998 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
1000 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1001 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
1002 #define HDR_L2_READING(hdr) \
1003 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1004 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1005 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1006 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1007 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1009 #define HDR_ISTYPE_METADATA(hdr) \
1010 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1011 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1013 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1014 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1020 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1021 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1024 * Hash table routines
1027 #define HT_LOCK_PAD CACHE_LINE_SIZE
1032 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1036 #define BUF_LOCKS 256
1037 typedef struct buf_hash_table {
1039 arc_buf_hdr_t **ht_table;
1040 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
1043 static buf_hash_table_t buf_hash_table;
1045 #define BUF_HASH_INDEX(spa, dva, birth) \
1046 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1047 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1048 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1049 #define HDR_LOCK(hdr) \
1050 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1052 uint64_t zfs_crc64_table[256];
1058 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1059 #define L2ARC_HEADROOM 2 /* num of writes */
1061 * If we discover during ARC scan any buffers to be compressed, we boost
1062 * our headroom for the next scanning cycle by this percentage multiple.
1064 #define L2ARC_HEADROOM_BOOST 200
1065 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1066 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1069 * Used to distinguish headers that are being process by
1070 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
1071 * address. This can happen when the header is added to the l2arc's list
1072 * of buffers to write in the first stage of l2arc_write_buffers(), but
1073 * has not yet been written out which happens in the second stage of
1074 * l2arc_write_buffers().
1076 #define L2ARC_ADDR_UNSET ((uint64_t)(-1))
1078 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1079 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1081 /* L2ARC Performance Tunables */
1082 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1083 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1084 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1085 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1086 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1087 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1088 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1089 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1090 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1092 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1093 &l2arc_write_max, 0, "max write size");
1094 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1095 &l2arc_write_boost, 0, "extra write during warmup");
1096 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1097 &l2arc_headroom, 0, "number of dev writes");
1098 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1099 &l2arc_feed_secs, 0, "interval seconds");
1100 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1101 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1103 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1104 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1105 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1106 &l2arc_feed_again, 0, "turbo warmup");
1107 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1108 &l2arc_norw, 0, "no reads during writes");
1110 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1111 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1112 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
1113 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
1114 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
1115 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
1117 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1118 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1119 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
1120 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
1121 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
1122 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
1124 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1125 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1126 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
1127 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
1128 "size of metadata in mru ghost state");
1129 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
1130 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
1131 "size of data in mru ghost state");
1133 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1134 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1135 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
1136 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
1137 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
1138 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
1140 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1141 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1142 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
1143 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
1144 "size of metadata in mfu ghost state");
1145 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
1146 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
1147 "size of data in mfu ghost state");
1149 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1150 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1156 vdev_t *l2ad_vdev; /* vdev */
1157 spa_t *l2ad_spa; /* spa */
1158 uint64_t l2ad_hand; /* next write location */
1159 uint64_t l2ad_start; /* first addr on device */
1160 uint64_t l2ad_end; /* last addr on device */
1161 boolean_t l2ad_first; /* first sweep through */
1162 boolean_t l2ad_writing; /* currently writing */
1163 kmutex_t l2ad_mtx; /* lock for buffer list */
1164 list_t l2ad_buflist; /* buffer list */
1165 list_node_t l2ad_node; /* device list node */
1166 refcount_t l2ad_alloc; /* allocated bytes */
1169 static list_t L2ARC_dev_list; /* device list */
1170 static list_t *l2arc_dev_list; /* device list pointer */
1171 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1172 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1173 static list_t L2ARC_free_on_write; /* free after write buf list */
1174 static list_t *l2arc_free_on_write; /* free after write list ptr */
1175 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1176 static uint64_t l2arc_ndev; /* number of devices */
1178 typedef struct l2arc_read_callback {
1179 arc_buf_t *l2rcb_buf; /* read buffer */
1180 spa_t *l2rcb_spa; /* spa */
1181 blkptr_t l2rcb_bp; /* original blkptr */
1182 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1183 int l2rcb_flags; /* original flags */
1184 enum zio_compress l2rcb_compress; /* applied compress */
1185 void *l2rcb_data; /* temporary buffer */
1186 } l2arc_read_callback_t;
1188 typedef struct l2arc_write_callback {
1189 l2arc_dev_t *l2wcb_dev; /* device info */
1190 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1191 } l2arc_write_callback_t;
1193 typedef struct l2arc_data_free {
1194 /* protected by l2arc_free_on_write_mtx */
1197 void (*l2df_func)(void *, size_t);
1198 list_node_t l2df_list_node;
1199 } l2arc_data_free_t;
1201 static kmutex_t l2arc_feed_thr_lock;
1202 static kcondvar_t l2arc_feed_thr_cv;
1203 static uint8_t l2arc_thread_exit;
1205 static void arc_get_data_buf(arc_buf_t *);
1206 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1207 static boolean_t arc_is_overflowing();
1208 static void arc_buf_watch(arc_buf_t *);
1210 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1211 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1213 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1214 static void l2arc_read_done(zio_t *);
1216 static boolean_t l2arc_transform_buf(arc_buf_hdr_t *, boolean_t);
1217 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress);
1218 static void l2arc_release_cdata_buf(arc_buf_hdr_t *);
1221 l2arc_trim(const arc_buf_hdr_t *hdr)
1223 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1225 ASSERT(HDR_HAS_L2HDR(hdr));
1226 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
1228 if (hdr->b_l2hdr.b_daddr == L2ARC_ADDR_UNSET)
1230 if (hdr->b_l2hdr.b_asize != 0) {
1231 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
1232 hdr->b_l2hdr.b_asize, 0);
1234 ASSERT3U(hdr->b_l2hdr.b_compress, ==, ZIO_COMPRESS_EMPTY);
1239 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1241 uint8_t *vdva = (uint8_t *)dva;
1242 uint64_t crc = -1ULL;
1245 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1247 for (i = 0; i < sizeof (dva_t); i++)
1248 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1250 crc ^= (spa>>8) ^ birth;
1255 #define BUF_EMPTY(buf) \
1256 ((buf)->b_dva.dva_word[0] == 0 && \
1257 (buf)->b_dva.dva_word[1] == 0)
1259 #define BUF_EQUAL(spa, dva, birth, buf) \
1260 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1261 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1262 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
1265 buf_discard_identity(arc_buf_hdr_t *hdr)
1267 hdr->b_dva.dva_word[0] = 0;
1268 hdr->b_dva.dva_word[1] = 0;
1272 static arc_buf_hdr_t *
1273 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1275 const dva_t *dva = BP_IDENTITY(bp);
1276 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1277 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1278 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1281 mutex_enter(hash_lock);
1282 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1283 hdr = hdr->b_hash_next) {
1284 if (BUF_EQUAL(spa, dva, birth, hdr)) {
1289 mutex_exit(hash_lock);
1295 * Insert an entry into the hash table. If there is already an element
1296 * equal to elem in the hash table, then the already existing element
1297 * will be returned and the new element will not be inserted.
1298 * Otherwise returns NULL.
1299 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1301 static arc_buf_hdr_t *
1302 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1304 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1305 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1306 arc_buf_hdr_t *fhdr;
1309 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1310 ASSERT(hdr->b_birth != 0);
1311 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1313 if (lockp != NULL) {
1315 mutex_enter(hash_lock);
1317 ASSERT(MUTEX_HELD(hash_lock));
1320 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1321 fhdr = fhdr->b_hash_next, i++) {
1322 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1326 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1327 buf_hash_table.ht_table[idx] = hdr;
1328 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
1330 /* collect some hash table performance data */
1332 ARCSTAT_BUMP(arcstat_hash_collisions);
1334 ARCSTAT_BUMP(arcstat_hash_chains);
1336 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1339 ARCSTAT_BUMP(arcstat_hash_elements);
1340 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1346 buf_hash_remove(arc_buf_hdr_t *hdr)
1348 arc_buf_hdr_t *fhdr, **hdrp;
1349 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1351 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1352 ASSERT(HDR_IN_HASH_TABLE(hdr));
1354 hdrp = &buf_hash_table.ht_table[idx];
1355 while ((fhdr = *hdrp) != hdr) {
1356 ASSERT(fhdr != NULL);
1357 hdrp = &fhdr->b_hash_next;
1359 *hdrp = hdr->b_hash_next;
1360 hdr->b_hash_next = NULL;
1361 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE;
1363 /* collect some hash table performance data */
1364 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1366 if (buf_hash_table.ht_table[idx] &&
1367 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1368 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1372 * Global data structures and functions for the buf kmem cache.
1374 static kmem_cache_t *hdr_full_cache;
1375 static kmem_cache_t *hdr_l2only_cache;
1376 static kmem_cache_t *buf_cache;
1383 kmem_free(buf_hash_table.ht_table,
1384 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1385 for (i = 0; i < BUF_LOCKS; i++)
1386 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1387 kmem_cache_destroy(hdr_full_cache);
1388 kmem_cache_destroy(hdr_l2only_cache);
1389 kmem_cache_destroy(buf_cache);
1393 * Constructor callback - called when the cache is empty
1394 * and a new buf is requested.
1398 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1400 arc_buf_hdr_t *hdr = vbuf;
1402 bzero(hdr, HDR_FULL_SIZE);
1403 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1404 refcount_create(&hdr->b_l1hdr.b_refcnt);
1405 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1406 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1407 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1414 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1416 arc_buf_hdr_t *hdr = vbuf;
1418 bzero(hdr, HDR_L2ONLY_SIZE);
1419 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1426 buf_cons(void *vbuf, void *unused, int kmflag)
1428 arc_buf_t *buf = vbuf;
1430 bzero(buf, sizeof (arc_buf_t));
1431 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1432 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1438 * Destructor callback - called when a cached buf is
1439 * no longer required.
1443 hdr_full_dest(void *vbuf, void *unused)
1445 arc_buf_hdr_t *hdr = vbuf;
1447 ASSERT(BUF_EMPTY(hdr));
1448 cv_destroy(&hdr->b_l1hdr.b_cv);
1449 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1450 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1451 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1452 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1457 hdr_l2only_dest(void *vbuf, void *unused)
1459 arc_buf_hdr_t *hdr = vbuf;
1461 ASSERT(BUF_EMPTY(hdr));
1462 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1467 buf_dest(void *vbuf, void *unused)
1469 arc_buf_t *buf = vbuf;
1471 mutex_destroy(&buf->b_evict_lock);
1472 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1476 * Reclaim callback -- invoked when memory is low.
1480 hdr_recl(void *unused)
1482 dprintf("hdr_recl called\n");
1484 * umem calls the reclaim func when we destroy the buf cache,
1485 * which is after we do arc_fini().
1488 cv_signal(&arc_reclaim_thread_cv);
1495 uint64_t hsize = 1ULL << 12;
1499 * The hash table is big enough to fill all of physical memory
1500 * with an average block size of zfs_arc_average_blocksize (default 8K).
1501 * By default, the table will take up
1502 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1504 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1507 buf_hash_table.ht_mask = hsize - 1;
1508 buf_hash_table.ht_table =
1509 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1510 if (buf_hash_table.ht_table == NULL) {
1511 ASSERT(hsize > (1ULL << 8));
1516 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1517 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1518 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1519 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1521 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1522 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1524 for (i = 0; i < 256; i++)
1525 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1526 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1528 for (i = 0; i < BUF_LOCKS; i++) {
1529 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1530 NULL, MUTEX_DEFAULT, NULL);
1535 * Transition between the two allocation states for the arc_buf_hdr struct.
1536 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1537 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1538 * version is used when a cache buffer is only in the L2ARC in order to reduce
1541 static arc_buf_hdr_t *
1542 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
1544 ASSERT(HDR_HAS_L2HDR(hdr));
1546 arc_buf_hdr_t *nhdr;
1547 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1549 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
1550 (old == hdr_l2only_cache && new == hdr_full_cache));
1552 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
1554 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
1555 buf_hash_remove(hdr);
1557 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
1559 if (new == hdr_full_cache) {
1560 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1562 * arc_access and arc_change_state need to be aware that a
1563 * header has just come out of L2ARC, so we set its state to
1564 * l2c_only even though it's about to change.
1566 nhdr->b_l1hdr.b_state = arc_l2c_only;
1568 /* Verify previous threads set to NULL before freeing */
1569 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1571 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1572 ASSERT0(hdr->b_l1hdr.b_datacnt);
1575 * If we've reached here, We must have been called from
1576 * arc_evict_hdr(), as such we should have already been
1577 * removed from any ghost list we were previously on
1578 * (which protects us from racing with arc_evict_state),
1579 * thus no locking is needed during this check.
1581 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1584 * A buffer must not be moved into the arc_l2c_only
1585 * state if it's not finished being written out to the
1586 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1587 * might try to be accessed, even though it was removed.
1589 VERIFY(!HDR_L2_WRITING(hdr));
1590 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1593 if (hdr->b_l1hdr.b_thawed != NULL) {
1594 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1595 hdr->b_l1hdr.b_thawed = NULL;
1599 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR;
1602 * The header has been reallocated so we need to re-insert it into any
1605 (void) buf_hash_insert(nhdr, NULL);
1607 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
1609 mutex_enter(&dev->l2ad_mtx);
1612 * We must place the realloc'ed header back into the list at
1613 * the same spot. Otherwise, if it's placed earlier in the list,
1614 * l2arc_write_buffers() could find it during the function's
1615 * write phase, and try to write it out to the l2arc.
1617 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
1618 list_remove(&dev->l2ad_buflist, hdr);
1620 mutex_exit(&dev->l2ad_mtx);
1623 * Since we're using the pointer address as the tag when
1624 * incrementing and decrementing the l2ad_alloc refcount, we
1625 * must remove the old pointer (that we're about to destroy) and
1626 * add the new pointer to the refcount. Otherwise we'd remove
1627 * the wrong pointer address when calling arc_hdr_destroy() later.
1630 (void) refcount_remove_many(&dev->l2ad_alloc,
1631 hdr->b_l2hdr.b_asize, hdr);
1633 (void) refcount_add_many(&dev->l2ad_alloc,
1634 nhdr->b_l2hdr.b_asize, nhdr);
1636 buf_discard_identity(hdr);
1637 hdr->b_freeze_cksum = NULL;
1638 kmem_cache_free(old, hdr);
1644 #define ARC_MINTIME (hz>>4) /* 62 ms */
1647 arc_cksum_verify(arc_buf_t *buf)
1651 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1654 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1655 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) {
1656 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1659 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc);
1660 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1661 panic("buffer modified while frozen!");
1662 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1666 arc_cksum_equal(arc_buf_t *buf)
1671 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1672 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc);
1673 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1674 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1680 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1682 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1685 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1686 if (buf->b_hdr->b_freeze_cksum != NULL) {
1687 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1690 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1691 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1692 NULL, buf->b_hdr->b_freeze_cksum);
1693 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1701 typedef struct procctl {
1709 arc_buf_unwatch(arc_buf_t *buf)
1716 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1717 ctl.prwatch.pr_size = 0;
1718 ctl.prwatch.pr_wflags = 0;
1719 result = write(arc_procfd, &ctl, sizeof (ctl));
1720 ASSERT3U(result, ==, sizeof (ctl));
1727 arc_buf_watch(arc_buf_t *buf)
1734 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1735 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1736 ctl.prwatch.pr_wflags = WA_WRITE;
1737 result = write(arc_procfd, &ctl, sizeof (ctl));
1738 ASSERT3U(result, ==, sizeof (ctl));
1742 #endif /* illumos */
1744 static arc_buf_contents_t
1745 arc_buf_type(arc_buf_hdr_t *hdr)
1747 if (HDR_ISTYPE_METADATA(hdr)) {
1748 return (ARC_BUFC_METADATA);
1750 return (ARC_BUFC_DATA);
1755 arc_bufc_to_flags(arc_buf_contents_t type)
1759 /* metadata field is 0 if buffer contains normal data */
1761 case ARC_BUFC_METADATA:
1762 return (ARC_FLAG_BUFC_METADATA);
1766 panic("undefined ARC buffer type!");
1767 return ((uint32_t)-1);
1771 arc_buf_thaw(arc_buf_t *buf)
1773 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1774 if (buf->b_hdr->b_l1hdr.b_state != arc_anon)
1775 panic("modifying non-anon buffer!");
1776 if (HDR_IO_IN_PROGRESS(buf->b_hdr))
1777 panic("modifying buffer while i/o in progress!");
1778 arc_cksum_verify(buf);
1781 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1782 if (buf->b_hdr->b_freeze_cksum != NULL) {
1783 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1784 buf->b_hdr->b_freeze_cksum = NULL;
1788 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1789 if (buf->b_hdr->b_l1hdr.b_thawed != NULL)
1790 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1);
1791 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1795 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1798 arc_buf_unwatch(buf);
1803 arc_buf_freeze(arc_buf_t *buf)
1805 kmutex_t *hash_lock;
1807 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1810 hash_lock = HDR_LOCK(buf->b_hdr);
1811 mutex_enter(hash_lock);
1813 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1814 buf->b_hdr->b_l1hdr.b_state == arc_anon);
1815 arc_cksum_compute(buf, B_FALSE);
1816 mutex_exit(hash_lock);
1821 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1823 ASSERT(HDR_HAS_L1HDR(hdr));
1824 ASSERT(MUTEX_HELD(hash_lock));
1825 arc_state_t *state = hdr->b_l1hdr.b_state;
1827 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
1828 (state != arc_anon)) {
1829 /* We don't use the L2-only state list. */
1830 if (state != arc_l2c_only) {
1831 arc_buf_contents_t type = arc_buf_type(hdr);
1832 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt;
1833 multilist_t *list = &state->arcs_list[type];
1834 uint64_t *size = &state->arcs_lsize[type];
1836 multilist_remove(list, hdr);
1838 if (GHOST_STATE(state)) {
1839 ASSERT0(hdr->b_l1hdr.b_datacnt);
1840 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
1841 delta = hdr->b_size;
1844 ASSERT3U(*size, >=, delta);
1845 atomic_add_64(size, -delta);
1847 /* remove the prefetch flag if we get a reference */
1848 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
1853 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1856 arc_state_t *state = hdr->b_l1hdr.b_state;
1858 ASSERT(HDR_HAS_L1HDR(hdr));
1859 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1860 ASSERT(!GHOST_STATE(state));
1863 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1864 * check to prevent usage of the arc_l2c_only list.
1866 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
1867 (state != arc_anon)) {
1868 arc_buf_contents_t type = arc_buf_type(hdr);
1869 multilist_t *list = &state->arcs_list[type];
1870 uint64_t *size = &state->arcs_lsize[type];
1872 multilist_insert(list, hdr);
1874 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
1875 atomic_add_64(size, hdr->b_size *
1876 hdr->b_l1hdr.b_datacnt);
1882 * Move the supplied buffer to the indicated state. The hash lock
1883 * for the buffer must be held by the caller.
1886 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
1887 kmutex_t *hash_lock)
1889 arc_state_t *old_state;
1892 uint64_t from_delta, to_delta;
1893 arc_buf_contents_t buftype = arc_buf_type(hdr);
1896 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1897 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1898 * L1 hdr doesn't always exist when we change state to arc_anon before
1899 * destroying a header, in which case reallocating to add the L1 hdr is
1902 if (HDR_HAS_L1HDR(hdr)) {
1903 old_state = hdr->b_l1hdr.b_state;
1904 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
1905 datacnt = hdr->b_l1hdr.b_datacnt;
1907 old_state = arc_l2c_only;
1912 ASSERT(MUTEX_HELD(hash_lock));
1913 ASSERT3P(new_state, !=, old_state);
1914 ASSERT(refcnt == 0 || datacnt > 0);
1915 ASSERT(!GHOST_STATE(new_state) || datacnt == 0);
1916 ASSERT(old_state != arc_anon || datacnt <= 1);
1918 from_delta = to_delta = datacnt * hdr->b_size;
1921 * If this buffer is evictable, transfer it from the
1922 * old state list to the new state list.
1925 if (old_state != arc_anon && old_state != arc_l2c_only) {
1926 uint64_t *size = &old_state->arcs_lsize[buftype];
1928 ASSERT(HDR_HAS_L1HDR(hdr));
1929 multilist_remove(&old_state->arcs_list[buftype], hdr);
1932 * If prefetching out of the ghost cache,
1933 * we will have a non-zero datacnt.
1935 if (GHOST_STATE(old_state) && datacnt == 0) {
1936 /* ghost elements have a ghost size */
1937 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1938 from_delta = hdr->b_size;
1940 ASSERT3U(*size, >=, from_delta);
1941 atomic_add_64(size, -from_delta);
1943 if (new_state != arc_anon && new_state != arc_l2c_only) {
1944 uint64_t *size = &new_state->arcs_lsize[buftype];
1947 * An L1 header always exists here, since if we're
1948 * moving to some L1-cached state (i.e. not l2c_only or
1949 * anonymous), we realloc the header to add an L1hdr
1952 ASSERT(HDR_HAS_L1HDR(hdr));
1953 multilist_insert(&new_state->arcs_list[buftype], hdr);
1955 /* ghost elements have a ghost size */
1956 if (GHOST_STATE(new_state)) {
1958 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1959 to_delta = hdr->b_size;
1961 atomic_add_64(size, to_delta);
1965 ASSERT(!BUF_EMPTY(hdr));
1966 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
1967 buf_hash_remove(hdr);
1969 /* adjust state sizes (ignore arc_l2c_only) */
1971 if (to_delta && new_state != arc_l2c_only) {
1972 ASSERT(HDR_HAS_L1HDR(hdr));
1973 if (GHOST_STATE(new_state)) {
1977 * We moving a header to a ghost state, we first
1978 * remove all arc buffers. Thus, we'll have a
1979 * datacnt of zero, and no arc buffer to use for
1980 * the reference. As a result, we use the arc
1981 * header pointer for the reference.
1983 (void) refcount_add_many(&new_state->arcs_size,
1986 ASSERT3U(datacnt, !=, 0);
1989 * Each individual buffer holds a unique reference,
1990 * thus we must remove each of these references one
1993 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1994 buf = buf->b_next) {
1995 (void) refcount_add_many(&new_state->arcs_size,
2001 if (from_delta && old_state != arc_l2c_only) {
2002 ASSERT(HDR_HAS_L1HDR(hdr));
2003 if (GHOST_STATE(old_state)) {
2005 * When moving a header off of a ghost state,
2006 * there's the possibility for datacnt to be
2007 * non-zero. This is because we first add the
2008 * arc buffer to the header prior to changing
2009 * the header's state. Since we used the header
2010 * for the reference when putting the header on
2011 * the ghost state, we must balance that and use
2012 * the header when removing off the ghost state
2013 * (even though datacnt is non zero).
2016 IMPLY(datacnt == 0, new_state == arc_anon ||
2017 new_state == arc_l2c_only);
2019 (void) refcount_remove_many(&old_state->arcs_size,
2022 ASSERT3P(datacnt, !=, 0);
2025 * Each individual buffer holds a unique reference,
2026 * thus we must remove each of these references one
2029 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2030 buf = buf->b_next) {
2031 (void) refcount_remove_many(
2032 &old_state->arcs_size, hdr->b_size, buf);
2037 if (HDR_HAS_L1HDR(hdr))
2038 hdr->b_l1hdr.b_state = new_state;
2041 * L2 headers should never be on the L2 state list since they don't
2042 * have L1 headers allocated.
2044 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2045 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2049 arc_space_consume(uint64_t space, arc_space_type_t type)
2051 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2054 case ARC_SPACE_DATA:
2055 ARCSTAT_INCR(arcstat_data_size, space);
2057 case ARC_SPACE_META:
2058 ARCSTAT_INCR(arcstat_metadata_size, space);
2060 case ARC_SPACE_OTHER:
2061 ARCSTAT_INCR(arcstat_other_size, space);
2063 case ARC_SPACE_HDRS:
2064 ARCSTAT_INCR(arcstat_hdr_size, space);
2066 case ARC_SPACE_L2HDRS:
2067 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
2071 if (type != ARC_SPACE_DATA)
2072 ARCSTAT_INCR(arcstat_meta_used, space);
2074 atomic_add_64(&arc_size, space);
2078 arc_space_return(uint64_t space, arc_space_type_t type)
2080 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2083 case ARC_SPACE_DATA:
2084 ARCSTAT_INCR(arcstat_data_size, -space);
2086 case ARC_SPACE_META:
2087 ARCSTAT_INCR(arcstat_metadata_size, -space);
2089 case ARC_SPACE_OTHER:
2090 ARCSTAT_INCR(arcstat_other_size, -space);
2092 case ARC_SPACE_HDRS:
2093 ARCSTAT_INCR(arcstat_hdr_size, -space);
2095 case ARC_SPACE_L2HDRS:
2096 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
2100 if (type != ARC_SPACE_DATA) {
2101 ASSERT(arc_meta_used >= space);
2102 if (arc_meta_max < arc_meta_used)
2103 arc_meta_max = arc_meta_used;
2104 ARCSTAT_INCR(arcstat_meta_used, -space);
2107 ASSERT(arc_size >= space);
2108 atomic_add_64(&arc_size, -space);
2112 arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
2117 ASSERT3U(size, >, 0);
2118 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
2119 ASSERT(BUF_EMPTY(hdr));
2120 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
2122 hdr->b_spa = spa_load_guid(spa);
2124 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2127 buf->b_efunc = NULL;
2128 buf->b_private = NULL;
2131 hdr->b_flags = arc_bufc_to_flags(type);
2132 hdr->b_flags |= ARC_FLAG_HAS_L1HDR;
2134 hdr->b_l1hdr.b_buf = buf;
2135 hdr->b_l1hdr.b_state = arc_anon;
2136 hdr->b_l1hdr.b_arc_access = 0;
2137 hdr->b_l1hdr.b_datacnt = 1;
2138 hdr->b_l1hdr.b_tmp_cdata = NULL;
2140 arc_get_data_buf(buf);
2141 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2142 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2147 static char *arc_onloan_tag = "onloan";
2150 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2151 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2152 * buffers must be returned to the arc before they can be used by the DMU or
2156 arc_loan_buf(spa_t *spa, int size)
2160 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
2162 atomic_add_64(&arc_loaned_bytes, size);
2167 * Return a loaned arc buffer to the arc.
2170 arc_return_buf(arc_buf_t *buf, void *tag)
2172 arc_buf_hdr_t *hdr = buf->b_hdr;
2174 ASSERT(buf->b_data != NULL);
2175 ASSERT(HDR_HAS_L1HDR(hdr));
2176 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2177 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2179 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
2182 /* Detach an arc_buf from a dbuf (tag) */
2184 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2186 arc_buf_hdr_t *hdr = buf->b_hdr;
2188 ASSERT(buf->b_data != NULL);
2189 ASSERT(HDR_HAS_L1HDR(hdr));
2190 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2191 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2192 buf->b_efunc = NULL;
2193 buf->b_private = NULL;
2195 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
2199 arc_buf_clone(arc_buf_t *from)
2202 arc_buf_hdr_t *hdr = from->b_hdr;
2203 uint64_t size = hdr->b_size;
2205 ASSERT(HDR_HAS_L1HDR(hdr));
2206 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2208 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2211 buf->b_efunc = NULL;
2212 buf->b_private = NULL;
2213 buf->b_next = hdr->b_l1hdr.b_buf;
2214 hdr->b_l1hdr.b_buf = buf;
2215 arc_get_data_buf(buf);
2216 bcopy(from->b_data, buf->b_data, size);
2219 * This buffer already exists in the arc so create a duplicate
2220 * copy for the caller. If the buffer is associated with user data
2221 * then track the size and number of duplicates. These stats will be
2222 * updated as duplicate buffers are created and destroyed.
2224 if (HDR_ISTYPE_DATA(hdr)) {
2225 ARCSTAT_BUMP(arcstat_duplicate_buffers);
2226 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
2228 hdr->b_l1hdr.b_datacnt += 1;
2233 arc_buf_add_ref(arc_buf_t *buf, void* tag)
2236 kmutex_t *hash_lock;
2239 * Check to see if this buffer is evicted. Callers
2240 * must verify b_data != NULL to know if the add_ref
2243 mutex_enter(&buf->b_evict_lock);
2244 if (buf->b_data == NULL) {
2245 mutex_exit(&buf->b_evict_lock);
2248 hash_lock = HDR_LOCK(buf->b_hdr);
2249 mutex_enter(hash_lock);
2251 ASSERT(HDR_HAS_L1HDR(hdr));
2252 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2253 mutex_exit(&buf->b_evict_lock);
2255 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
2256 hdr->b_l1hdr.b_state == arc_mfu);
2258 add_reference(hdr, hash_lock, tag);
2259 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2260 arc_access(hdr, hash_lock);
2261 mutex_exit(hash_lock);
2262 ARCSTAT_BUMP(arcstat_hits);
2263 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
2264 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
2265 data, metadata, hits);
2269 arc_buf_free_on_write(void *data, size_t size,
2270 void (*free_func)(void *, size_t))
2272 l2arc_data_free_t *df;
2274 df = kmem_alloc(sizeof (*df), KM_SLEEP);
2275 df->l2df_data = data;
2276 df->l2df_size = size;
2277 df->l2df_func = free_func;
2278 mutex_enter(&l2arc_free_on_write_mtx);
2279 list_insert_head(l2arc_free_on_write, df);
2280 mutex_exit(&l2arc_free_on_write_mtx);
2284 * Free the arc data buffer. If it is an l2arc write in progress,
2285 * the buffer is placed on l2arc_free_on_write to be freed later.
2288 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
2290 arc_buf_hdr_t *hdr = buf->b_hdr;
2292 if (HDR_L2_WRITING(hdr)) {
2293 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
2294 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2296 free_func(buf->b_data, hdr->b_size);
2301 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
2303 size_t align, asize, len;
2305 ASSERT(HDR_HAS_L2HDR(hdr));
2306 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx));
2309 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
2310 * that doesn't exist, the header is in the arc_l2c_only state,
2311 * and there isn't anything to free (it's already been freed).
2313 if (!HDR_HAS_L1HDR(hdr))
2317 * The header isn't being written to the l2arc device, thus it
2318 * shouldn't have a b_tmp_cdata to free.
2320 if (!HDR_L2_WRITING(hdr)) {
2321 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2326 * The bufer has been chosen for writing to L2ARC, but it's
2327 * not being written just yet. In other words,
2328 * b_tmp_cdata points to exactly the same buffer as b_data,
2329 * l2arc_transform_buf hasn't been called.
2331 if (hdr->b_l2hdr.b_daddr == L2ARC_ADDR_UNSET) {
2332 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==,
2333 hdr->b_l1hdr.b_buf->b_data);
2334 ASSERT3U(hdr->b_l2hdr.b_compress, ==, ZIO_COMPRESS_OFF);
2335 hdr->b_l1hdr.b_tmp_cdata = NULL;
2340 * There's nothing to free since the buffer was all zero's and
2341 * compressed to a zero length buffer.
2343 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) {
2344 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2349 * Nothing to do if the temporary buffer was not required.
2351 if (hdr->b_l1hdr.b_tmp_cdata == NULL)
2354 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
2356 align = (size_t)1 << hdr->b_l2hdr.b_dev->l2ad_vdev->vdev_ashift;
2357 asize = P2ROUNDUP(len, align);
2358 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, asize,
2360 hdr->b_l1hdr.b_tmp_cdata = NULL;
2364 * Free up buf->b_data and if 'remove' is set, then pull the
2365 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2368 arc_buf_destroy(arc_buf_t *buf, boolean_t remove)
2372 /* free up data associated with the buf */
2373 if (buf->b_data != NULL) {
2374 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
2375 uint64_t size = buf->b_hdr->b_size;
2376 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
2378 arc_cksum_verify(buf);
2380 arc_buf_unwatch(buf);
2383 if (type == ARC_BUFC_METADATA) {
2384 arc_buf_data_free(buf, zio_buf_free);
2385 arc_space_return(size, ARC_SPACE_META);
2387 ASSERT(type == ARC_BUFC_DATA);
2388 arc_buf_data_free(buf, zio_data_buf_free);
2389 arc_space_return(size, ARC_SPACE_DATA);
2392 /* protected by hash lock, if in the hash table */
2393 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) {
2394 uint64_t *cnt = &state->arcs_lsize[type];
2396 ASSERT(refcount_is_zero(
2397 &buf->b_hdr->b_l1hdr.b_refcnt));
2398 ASSERT(state != arc_anon && state != arc_l2c_only);
2400 ASSERT3U(*cnt, >=, size);
2401 atomic_add_64(cnt, -size);
2404 (void) refcount_remove_many(&state->arcs_size, size, buf);
2408 * If we're destroying a duplicate buffer make sure
2409 * that the appropriate statistics are updated.
2411 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 &&
2412 HDR_ISTYPE_DATA(buf->b_hdr)) {
2413 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
2414 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
2416 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0);
2417 buf->b_hdr->b_l1hdr.b_datacnt -= 1;
2420 /* only remove the buf if requested */
2424 /* remove the buf from the hdr list */
2425 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf;
2426 bufp = &(*bufp)->b_next)
2428 *bufp = buf->b_next;
2431 ASSERT(buf->b_efunc == NULL);
2433 /* clean up the buf */
2435 kmem_cache_free(buf_cache, buf);
2439 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2441 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2442 l2arc_dev_t *dev = l2hdr->b_dev;
2444 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2445 ASSERT(HDR_HAS_L2HDR(hdr));
2447 list_remove(&dev->l2ad_buflist, hdr);
2450 * We don't want to leak the b_tmp_cdata buffer that was
2451 * allocated in l2arc_write_buffers()
2453 arc_buf_l2_cdata_free(hdr);
2456 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2457 * this header is being processed by l2arc_write_buffers() (i.e.
2458 * it's in the first stage of l2arc_write_buffers()).
2459 * Re-affirming that truth here, just to serve as a reminder. If
2460 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2461 * may not have its HDR_L2_WRITING flag set. (the write may have
2462 * completed, in which case HDR_L2_WRITING will be false and the
2463 * b_daddr field will point to the address of the buffer on disk).
2465 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr));
2468 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2469 * l2arc_write_buffers(). Since we've just removed this header
2470 * from the l2arc buffer list, this header will never reach the
2471 * second stage of l2arc_write_buffers(), which increments the
2472 * accounting stats for this header. Thus, we must be careful
2473 * not to decrement them for this header either.
2475 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) {
2476 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
2477 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
2479 vdev_space_update(dev->l2ad_vdev,
2480 -l2hdr->b_asize, 0, 0);
2482 (void) refcount_remove_many(&dev->l2ad_alloc,
2483 l2hdr->b_asize, hdr);
2486 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
2490 arc_hdr_destroy(arc_buf_hdr_t *hdr)
2492 if (HDR_HAS_L1HDR(hdr)) {
2493 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
2494 hdr->b_l1hdr.b_datacnt > 0);
2495 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2496 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2498 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2499 ASSERT(!HDR_IN_HASH_TABLE(hdr));
2501 if (HDR_HAS_L2HDR(hdr)) {
2502 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2503 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
2506 mutex_enter(&dev->l2ad_mtx);
2509 * Even though we checked this conditional above, we
2510 * need to check this again now that we have the
2511 * l2ad_mtx. This is because we could be racing with
2512 * another thread calling l2arc_evict() which might have
2513 * destroyed this header's L2 portion as we were waiting
2514 * to acquire the l2ad_mtx. If that happens, we don't
2515 * want to re-destroy the header's L2 portion.
2517 if (HDR_HAS_L2HDR(hdr)) {
2519 arc_hdr_l2hdr_destroy(hdr);
2523 mutex_exit(&dev->l2ad_mtx);
2526 if (!BUF_EMPTY(hdr))
2527 buf_discard_identity(hdr);
2529 if (hdr->b_freeze_cksum != NULL) {
2530 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
2531 hdr->b_freeze_cksum = NULL;
2534 if (HDR_HAS_L1HDR(hdr)) {
2535 while (hdr->b_l1hdr.b_buf) {
2536 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2538 if (buf->b_efunc != NULL) {
2539 mutex_enter(&arc_user_evicts_lock);
2540 mutex_enter(&buf->b_evict_lock);
2541 ASSERT(buf->b_hdr != NULL);
2542 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE);
2543 hdr->b_l1hdr.b_buf = buf->b_next;
2544 buf->b_hdr = &arc_eviction_hdr;
2545 buf->b_next = arc_eviction_list;
2546 arc_eviction_list = buf;
2547 mutex_exit(&buf->b_evict_lock);
2548 cv_signal(&arc_user_evicts_cv);
2549 mutex_exit(&arc_user_evicts_lock);
2551 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE);
2555 if (hdr->b_l1hdr.b_thawed != NULL) {
2556 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2557 hdr->b_l1hdr.b_thawed = NULL;
2562 ASSERT3P(hdr->b_hash_next, ==, NULL);
2563 if (HDR_HAS_L1HDR(hdr)) {
2564 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2565 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
2566 kmem_cache_free(hdr_full_cache, hdr);
2568 kmem_cache_free(hdr_l2only_cache, hdr);
2573 arc_buf_free(arc_buf_t *buf, void *tag)
2575 arc_buf_hdr_t *hdr = buf->b_hdr;
2576 int hashed = hdr->b_l1hdr.b_state != arc_anon;
2578 ASSERT(buf->b_efunc == NULL);
2579 ASSERT(buf->b_data != NULL);
2582 kmutex_t *hash_lock = HDR_LOCK(hdr);
2584 mutex_enter(hash_lock);
2586 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2588 (void) remove_reference(hdr, hash_lock, tag);
2589 if (hdr->b_l1hdr.b_datacnt > 1) {
2590 arc_buf_destroy(buf, TRUE);
2592 ASSERT(buf == hdr->b_l1hdr.b_buf);
2593 ASSERT(buf->b_efunc == NULL);
2594 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2596 mutex_exit(hash_lock);
2597 } else if (HDR_IO_IN_PROGRESS(hdr)) {
2600 * We are in the middle of an async write. Don't destroy
2601 * this buffer unless the write completes before we finish
2602 * decrementing the reference count.
2604 mutex_enter(&arc_user_evicts_lock);
2605 (void) remove_reference(hdr, NULL, tag);
2606 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2607 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
2608 mutex_exit(&arc_user_evicts_lock);
2610 arc_hdr_destroy(hdr);
2612 if (remove_reference(hdr, NULL, tag) > 0)
2613 arc_buf_destroy(buf, TRUE);
2615 arc_hdr_destroy(hdr);
2620 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
2622 arc_buf_hdr_t *hdr = buf->b_hdr;
2623 kmutex_t *hash_lock = HDR_LOCK(hdr);
2624 boolean_t no_callback = (buf->b_efunc == NULL);
2626 if (hdr->b_l1hdr.b_state == arc_anon) {
2627 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
2628 arc_buf_free(buf, tag);
2629 return (no_callback);
2632 mutex_enter(hash_lock);
2634 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
2635 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2636 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2637 ASSERT(buf->b_data != NULL);
2639 (void) remove_reference(hdr, hash_lock, tag);
2640 if (hdr->b_l1hdr.b_datacnt > 1) {
2642 arc_buf_destroy(buf, TRUE);
2643 } else if (no_callback) {
2644 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL);
2645 ASSERT(buf->b_efunc == NULL);
2646 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2648 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 ||
2649 refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2650 mutex_exit(hash_lock);
2651 return (no_callback);
2655 arc_buf_size(arc_buf_t *buf)
2657 return (buf->b_hdr->b_size);
2661 * Called from the DMU to determine if the current buffer should be
2662 * evicted. In order to ensure proper locking, the eviction must be initiated
2663 * from the DMU. Return true if the buffer is associated with user data and
2664 * duplicate buffers still exist.
2667 arc_buf_eviction_needed(arc_buf_t *buf)
2670 boolean_t evict_needed = B_FALSE;
2672 if (zfs_disable_dup_eviction)
2675 mutex_enter(&buf->b_evict_lock);
2679 * We are in arc_do_user_evicts(); let that function
2680 * perform the eviction.
2682 ASSERT(buf->b_data == NULL);
2683 mutex_exit(&buf->b_evict_lock);
2685 } else if (buf->b_data == NULL) {
2687 * We have already been added to the arc eviction list;
2688 * recommend eviction.
2690 ASSERT3P(hdr, ==, &arc_eviction_hdr);
2691 mutex_exit(&buf->b_evict_lock);
2695 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr))
2696 evict_needed = B_TRUE;
2698 mutex_exit(&buf->b_evict_lock);
2699 return (evict_needed);
2703 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2704 * state of the header is dependent on it's state prior to entering this
2705 * function. The following transitions are possible:
2707 * - arc_mru -> arc_mru_ghost
2708 * - arc_mfu -> arc_mfu_ghost
2709 * - arc_mru_ghost -> arc_l2c_only
2710 * - arc_mru_ghost -> deleted
2711 * - arc_mfu_ghost -> arc_l2c_only
2712 * - arc_mfu_ghost -> deleted
2715 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
2717 arc_state_t *evicted_state, *state;
2718 int64_t bytes_evicted = 0;
2720 ASSERT(MUTEX_HELD(hash_lock));
2721 ASSERT(HDR_HAS_L1HDR(hdr));
2723 state = hdr->b_l1hdr.b_state;
2724 if (GHOST_STATE(state)) {
2725 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2726 ASSERT(hdr->b_l1hdr.b_buf == NULL);
2729 * l2arc_write_buffers() relies on a header's L1 portion
2730 * (i.e. it's b_tmp_cdata field) during it's write phase.
2731 * Thus, we cannot push a header onto the arc_l2c_only
2732 * state (removing it's L1 piece) until the header is
2733 * done being written to the l2arc.
2735 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
2736 ARCSTAT_BUMP(arcstat_evict_l2_skip);
2737 return (bytes_evicted);
2740 ARCSTAT_BUMP(arcstat_deleted);
2741 bytes_evicted += hdr->b_size;
2743 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
2745 if (HDR_HAS_L2HDR(hdr)) {
2747 * This buffer is cached on the 2nd Level ARC;
2748 * don't destroy the header.
2750 arc_change_state(arc_l2c_only, hdr, hash_lock);
2752 * dropping from L1+L2 cached to L2-only,
2753 * realloc to remove the L1 header.
2755 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
2758 arc_change_state(arc_anon, hdr, hash_lock);
2759 arc_hdr_destroy(hdr);
2761 return (bytes_evicted);
2764 ASSERT(state == arc_mru || state == arc_mfu);
2765 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2767 /* prefetch buffers have a minimum lifespan */
2768 if (HDR_IO_IN_PROGRESS(hdr) ||
2769 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
2770 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
2771 arc_min_prefetch_lifespan)) {
2772 ARCSTAT_BUMP(arcstat_evict_skip);
2773 return (bytes_evicted);
2776 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
2777 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0);
2778 while (hdr->b_l1hdr.b_buf) {
2779 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2780 if (!mutex_tryenter(&buf->b_evict_lock)) {
2781 ARCSTAT_BUMP(arcstat_mutex_miss);
2784 if (buf->b_data != NULL)
2785 bytes_evicted += hdr->b_size;
2786 if (buf->b_efunc != NULL) {
2787 mutex_enter(&arc_user_evicts_lock);
2788 arc_buf_destroy(buf, FALSE);
2789 hdr->b_l1hdr.b_buf = buf->b_next;
2790 buf->b_hdr = &arc_eviction_hdr;
2791 buf->b_next = arc_eviction_list;
2792 arc_eviction_list = buf;
2793 cv_signal(&arc_user_evicts_cv);
2794 mutex_exit(&arc_user_evicts_lock);
2795 mutex_exit(&buf->b_evict_lock);
2797 mutex_exit(&buf->b_evict_lock);
2798 arc_buf_destroy(buf, TRUE);
2802 if (HDR_HAS_L2HDR(hdr)) {
2803 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size);
2805 if (l2arc_write_eligible(hdr->b_spa, hdr))
2806 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size);
2808 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size);
2811 if (hdr->b_l1hdr.b_datacnt == 0) {
2812 arc_change_state(evicted_state, hdr, hash_lock);
2813 ASSERT(HDR_IN_HASH_TABLE(hdr));
2814 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
2815 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
2816 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
2819 return (bytes_evicted);
2823 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
2824 uint64_t spa, int64_t bytes)
2826 multilist_sublist_t *mls;
2827 uint64_t bytes_evicted = 0;
2829 kmutex_t *hash_lock;
2830 int evict_count = 0;
2832 ASSERT3P(marker, !=, NULL);
2833 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2835 mls = multilist_sublist_lock(ml, idx);
2837 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
2838 hdr = multilist_sublist_prev(mls, marker)) {
2839 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
2840 (evict_count >= zfs_arc_evict_batch_limit))
2844 * To keep our iteration location, move the marker
2845 * forward. Since we're not holding hdr's hash lock, we
2846 * must be very careful and not remove 'hdr' from the
2847 * sublist. Otherwise, other consumers might mistake the
2848 * 'hdr' as not being on a sublist when they call the
2849 * multilist_link_active() function (they all rely on
2850 * the hash lock protecting concurrent insertions and
2851 * removals). multilist_sublist_move_forward() was
2852 * specifically implemented to ensure this is the case
2853 * (only 'marker' will be removed and re-inserted).
2855 multilist_sublist_move_forward(mls, marker);
2858 * The only case where the b_spa field should ever be
2859 * zero, is the marker headers inserted by
2860 * arc_evict_state(). It's possible for multiple threads
2861 * to be calling arc_evict_state() concurrently (e.g.
2862 * dsl_pool_close() and zio_inject_fault()), so we must
2863 * skip any markers we see from these other threads.
2865 if (hdr->b_spa == 0)
2868 /* we're only interested in evicting buffers of a certain spa */
2869 if (spa != 0 && hdr->b_spa != spa) {
2870 ARCSTAT_BUMP(arcstat_evict_skip);
2874 hash_lock = HDR_LOCK(hdr);
2877 * We aren't calling this function from any code path
2878 * that would already be holding a hash lock, so we're
2879 * asserting on this assumption to be defensive in case
2880 * this ever changes. Without this check, it would be
2881 * possible to incorrectly increment arcstat_mutex_miss
2882 * below (e.g. if the code changed such that we called
2883 * this function with a hash lock held).
2885 ASSERT(!MUTEX_HELD(hash_lock));
2887 if (mutex_tryenter(hash_lock)) {
2888 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
2889 mutex_exit(hash_lock);
2891 bytes_evicted += evicted;
2894 * If evicted is zero, arc_evict_hdr() must have
2895 * decided to skip this header, don't increment
2896 * evict_count in this case.
2902 * If arc_size isn't overflowing, signal any
2903 * threads that might happen to be waiting.
2905 * For each header evicted, we wake up a single
2906 * thread. If we used cv_broadcast, we could
2907 * wake up "too many" threads causing arc_size
2908 * to significantly overflow arc_c; since
2909 * arc_get_data_buf() doesn't check for overflow
2910 * when it's woken up (it doesn't because it's
2911 * possible for the ARC to be overflowing while
2912 * full of un-evictable buffers, and the
2913 * function should proceed in this case).
2915 * If threads are left sleeping, due to not
2916 * using cv_broadcast, they will be woken up
2917 * just before arc_reclaim_thread() sleeps.
2919 mutex_enter(&arc_reclaim_lock);
2920 if (!arc_is_overflowing())
2921 cv_signal(&arc_reclaim_waiters_cv);
2922 mutex_exit(&arc_reclaim_lock);
2924 ARCSTAT_BUMP(arcstat_mutex_miss);
2928 multilist_sublist_unlock(mls);
2930 return (bytes_evicted);
2934 * Evict buffers from the given arc state, until we've removed the
2935 * specified number of bytes. Move the removed buffers to the
2936 * appropriate evict state.
2938 * This function makes a "best effort". It skips over any buffers
2939 * it can't get a hash_lock on, and so, may not catch all candidates.
2940 * It may also return without evicting as much space as requested.
2942 * If bytes is specified using the special value ARC_EVICT_ALL, this
2943 * will evict all available (i.e. unlocked and evictable) buffers from
2944 * the given arc state; which is used by arc_flush().
2947 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
2948 arc_buf_contents_t type)
2950 uint64_t total_evicted = 0;
2951 multilist_t *ml = &state->arcs_list[type];
2953 arc_buf_hdr_t **markers;
2955 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2957 num_sublists = multilist_get_num_sublists(ml);
2960 * If we've tried to evict from each sublist, made some
2961 * progress, but still have not hit the target number of bytes
2962 * to evict, we want to keep trying. The markers allow us to
2963 * pick up where we left off for each individual sublist, rather
2964 * than starting from the tail each time.
2966 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
2967 for (int i = 0; i < num_sublists; i++) {
2968 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
2971 * A b_spa of 0 is used to indicate that this header is
2972 * a marker. This fact is used in arc_adjust_type() and
2973 * arc_evict_state_impl().
2975 markers[i]->b_spa = 0;
2977 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2978 multilist_sublist_insert_tail(mls, markers[i]);
2979 multilist_sublist_unlock(mls);
2983 * While we haven't hit our target number of bytes to evict, or
2984 * we're evicting all available buffers.
2986 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
2988 * Start eviction using a randomly selected sublist,
2989 * this is to try and evenly balance eviction across all
2990 * sublists. Always starting at the same sublist
2991 * (e.g. index 0) would cause evictions to favor certain
2992 * sublists over others.
2994 int sublist_idx = multilist_get_random_index(ml);
2995 uint64_t scan_evicted = 0;
2997 for (int i = 0; i < num_sublists; i++) {
2998 uint64_t bytes_remaining;
2999 uint64_t bytes_evicted;
3001 if (bytes == ARC_EVICT_ALL)
3002 bytes_remaining = ARC_EVICT_ALL;
3003 else if (total_evicted < bytes)
3004 bytes_remaining = bytes - total_evicted;
3008 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3009 markers[sublist_idx], spa, bytes_remaining);
3011 scan_evicted += bytes_evicted;
3012 total_evicted += bytes_evicted;
3014 /* we've reached the end, wrap to the beginning */
3015 if (++sublist_idx >= num_sublists)
3020 * If we didn't evict anything during this scan, we have
3021 * no reason to believe we'll evict more during another
3022 * scan, so break the loop.
3024 if (scan_evicted == 0) {
3025 /* This isn't possible, let's make that obvious */
3026 ASSERT3S(bytes, !=, 0);
3029 * When bytes is ARC_EVICT_ALL, the only way to
3030 * break the loop is when scan_evicted is zero.
3031 * In that case, we actually have evicted enough,
3032 * so we don't want to increment the kstat.
3034 if (bytes != ARC_EVICT_ALL) {
3035 ASSERT3S(total_evicted, <, bytes);
3036 ARCSTAT_BUMP(arcstat_evict_not_enough);
3043 for (int i = 0; i < num_sublists; i++) {
3044 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3045 multilist_sublist_remove(mls, markers[i]);
3046 multilist_sublist_unlock(mls);
3048 kmem_cache_free(hdr_full_cache, markers[i]);
3050 kmem_free(markers, sizeof (*markers) * num_sublists);
3052 return (total_evicted);
3056 * Flush all "evictable" data of the given type from the arc state
3057 * specified. This will not evict any "active" buffers (i.e. referenced).
3059 * When 'retry' is set to FALSE, the function will make a single pass
3060 * over the state and evict any buffers that it can. Since it doesn't
3061 * continually retry the eviction, it might end up leaving some buffers
3062 * in the ARC due to lock misses.
3064 * When 'retry' is set to TRUE, the function will continually retry the
3065 * eviction until *all* evictable buffers have been removed from the
3066 * state. As a result, if concurrent insertions into the state are
3067 * allowed (e.g. if the ARC isn't shutting down), this function might
3068 * wind up in an infinite loop, continually trying to evict buffers.
3071 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3074 uint64_t evicted = 0;
3076 while (state->arcs_lsize[type] != 0) {
3077 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3087 * Evict the specified number of bytes from the state specified,
3088 * restricting eviction to the spa and type given. This function
3089 * prevents us from trying to evict more from a state's list than
3090 * is "evictable", and to skip evicting altogether when passed a
3091 * negative value for "bytes". In contrast, arc_evict_state() will
3092 * evict everything it can, when passed a negative value for "bytes".
3095 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3096 arc_buf_contents_t type)
3100 if (bytes > 0 && state->arcs_lsize[type] > 0) {
3101 delta = MIN(state->arcs_lsize[type], bytes);
3102 return (arc_evict_state(state, spa, delta, type));
3109 * Evict metadata buffers from the cache, such that arc_meta_used is
3110 * capped by the arc_meta_limit tunable.
3113 arc_adjust_meta(void)
3115 uint64_t total_evicted = 0;
3119 * If we're over the meta limit, we want to evict enough
3120 * metadata to get back under the meta limit. We don't want to
3121 * evict so much that we drop the MRU below arc_p, though. If
3122 * we're over the meta limit more than we're over arc_p, we
3123 * evict some from the MRU here, and some from the MFU below.
3125 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3126 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3127 refcount_count(&arc_mru->arcs_size) - arc_p));
3129 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3132 * Similar to the above, we want to evict enough bytes to get us
3133 * below the meta limit, but not so much as to drop us below the
3134 * space alloted to the MFU (which is defined as arc_c - arc_p).
3136 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3137 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
3139 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3141 return (total_evicted);
3145 * Return the type of the oldest buffer in the given arc state
3147 * This function will select a random sublist of type ARC_BUFC_DATA and
3148 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3149 * is compared, and the type which contains the "older" buffer will be
3152 static arc_buf_contents_t
3153 arc_adjust_type(arc_state_t *state)
3155 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
3156 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
3157 int data_idx = multilist_get_random_index(data_ml);
3158 int meta_idx = multilist_get_random_index(meta_ml);
3159 multilist_sublist_t *data_mls;
3160 multilist_sublist_t *meta_mls;
3161 arc_buf_contents_t type;
3162 arc_buf_hdr_t *data_hdr;
3163 arc_buf_hdr_t *meta_hdr;
3166 * We keep the sublist lock until we're finished, to prevent
3167 * the headers from being destroyed via arc_evict_state().
3169 data_mls = multilist_sublist_lock(data_ml, data_idx);
3170 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3173 * These two loops are to ensure we skip any markers that
3174 * might be at the tail of the lists due to arc_evict_state().
3177 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3178 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3179 if (data_hdr->b_spa != 0)
3183 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3184 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3185 if (meta_hdr->b_spa != 0)
3189 if (data_hdr == NULL && meta_hdr == NULL) {
3190 type = ARC_BUFC_DATA;
3191 } else if (data_hdr == NULL) {
3192 ASSERT3P(meta_hdr, !=, NULL);
3193 type = ARC_BUFC_METADATA;
3194 } else if (meta_hdr == NULL) {
3195 ASSERT3P(data_hdr, !=, NULL);
3196 type = ARC_BUFC_DATA;
3198 ASSERT3P(data_hdr, !=, NULL);
3199 ASSERT3P(meta_hdr, !=, NULL);
3201 /* The headers can't be on the sublist without an L1 header */
3202 ASSERT(HDR_HAS_L1HDR(data_hdr));
3203 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3205 if (data_hdr->b_l1hdr.b_arc_access <
3206 meta_hdr->b_l1hdr.b_arc_access) {
3207 type = ARC_BUFC_DATA;
3209 type = ARC_BUFC_METADATA;
3213 multilist_sublist_unlock(meta_mls);
3214 multilist_sublist_unlock(data_mls);
3220 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3225 uint64_t total_evicted = 0;
3230 * If we're over arc_meta_limit, we want to correct that before
3231 * potentially evicting data buffers below.
3233 total_evicted += arc_adjust_meta();
3238 * If we're over the target cache size, we want to evict enough
3239 * from the list to get back to our target size. We don't want
3240 * to evict too much from the MRU, such that it drops below
3241 * arc_p. So, if we're over our target cache size more than
3242 * the MRU is over arc_p, we'll evict enough to get back to
3243 * arc_p here, and then evict more from the MFU below.
3245 target = MIN((int64_t)(arc_size - arc_c),
3246 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3247 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
3250 * If we're below arc_meta_min, always prefer to evict data.
3251 * Otherwise, try to satisfy the requested number of bytes to
3252 * evict from the type which contains older buffers; in an
3253 * effort to keep newer buffers in the cache regardless of their
3254 * type. If we cannot satisfy the number of bytes from this
3255 * type, spill over into the next type.
3257 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
3258 arc_meta_used > arc_meta_min) {
3259 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3260 total_evicted += bytes;
3263 * If we couldn't evict our target number of bytes from
3264 * metadata, we try to get the rest from data.
3269 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3271 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3272 total_evicted += bytes;
3275 * If we couldn't evict our target number of bytes from
3276 * data, we try to get the rest from metadata.
3281 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3287 * Now that we've tried to evict enough from the MRU to get its
3288 * size back to arc_p, if we're still above the target cache
3289 * size, we evict the rest from the MFU.
3291 target = arc_size - arc_c;
3293 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
3294 arc_meta_used > arc_meta_min) {
3295 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3296 total_evicted += bytes;
3299 * If we couldn't evict our target number of bytes from
3300 * metadata, we try to get the rest from data.
3305 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3307 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3308 total_evicted += bytes;
3311 * If we couldn't evict our target number of bytes from
3312 * data, we try to get the rest from data.
3317 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3321 * Adjust ghost lists
3323 * In addition to the above, the ARC also defines target values
3324 * for the ghost lists. The sum of the mru list and mru ghost
3325 * list should never exceed the target size of the cache, and
3326 * the sum of the mru list, mfu list, mru ghost list, and mfu
3327 * ghost list should never exceed twice the target size of the
3328 * cache. The following logic enforces these limits on the ghost
3329 * caches, and evicts from them as needed.
3331 target = refcount_count(&arc_mru->arcs_size) +
3332 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3334 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3335 total_evicted += bytes;
3340 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3343 * We assume the sum of the mru list and mfu list is less than
3344 * or equal to arc_c (we enforced this above), which means we
3345 * can use the simpler of the two equations below:
3347 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3348 * mru ghost + mfu ghost <= arc_c
3350 target = refcount_count(&arc_mru_ghost->arcs_size) +
3351 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3353 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3354 total_evicted += bytes;
3359 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3361 return (total_evicted);
3365 arc_do_user_evicts(void)
3367 mutex_enter(&arc_user_evicts_lock);
3368 while (arc_eviction_list != NULL) {
3369 arc_buf_t *buf = arc_eviction_list;
3370 arc_eviction_list = buf->b_next;
3371 mutex_enter(&buf->b_evict_lock);
3373 mutex_exit(&buf->b_evict_lock);
3374 mutex_exit(&arc_user_evicts_lock);
3376 if (buf->b_efunc != NULL)
3377 VERIFY0(buf->b_efunc(buf->b_private));
3379 buf->b_efunc = NULL;
3380 buf->b_private = NULL;
3381 kmem_cache_free(buf_cache, buf);
3382 mutex_enter(&arc_user_evicts_lock);
3384 mutex_exit(&arc_user_evicts_lock);
3388 arc_flush(spa_t *spa, boolean_t retry)
3393 * If retry is TRUE, a spa must not be specified since we have
3394 * no good way to determine if all of a spa's buffers have been
3395 * evicted from an arc state.
3397 ASSERT(!retry || spa == 0);
3400 guid = spa_load_guid(spa);
3402 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3403 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3405 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3406 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3408 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3409 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3411 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3412 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3414 arc_do_user_evicts();
3415 ASSERT(spa || arc_eviction_list == NULL);
3419 arc_shrink(int64_t to_free)
3421 if (arc_c > arc_c_min) {
3422 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
3423 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
3424 if (arc_c > arc_c_min + to_free)
3425 atomic_add_64(&arc_c, -to_free);
3429 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3430 if (arc_c > arc_size)
3431 arc_c = MAX(arc_size, arc_c_min);
3433 arc_p = (arc_c >> 1);
3435 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
3438 ASSERT(arc_c >= arc_c_min);
3439 ASSERT((int64_t)arc_p >= 0);
3442 if (arc_size > arc_c) {
3443 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size,
3445 (void) arc_adjust();
3449 static long needfree = 0;
3451 typedef enum free_memory_reason_t {
3456 FMR_PAGES_PP_MAXIMUM,
3460 } free_memory_reason_t;
3462 int64_t last_free_memory;
3463 free_memory_reason_t last_free_reason;
3466 * Additional reserve of pages for pp_reserve.
3468 int64_t arc_pages_pp_reserve = 64;
3471 * Additional reserve of pages for swapfs.
3473 int64_t arc_swapfs_reserve = 64;
3476 * Return the amount of memory that can be consumed before reclaim will be
3477 * needed. Positive if there is sufficient free memory, negative indicates
3478 * the amount of memory that needs to be freed up.
3481 arc_available_memory(void)
3483 int64_t lowest = INT64_MAX;
3485 free_memory_reason_t r = FMR_UNKNOWN;
3489 n = PAGESIZE * (-needfree);
3497 * Cooperate with pagedaemon when it's time for it to scan
3498 * and reclaim some pages.
3500 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
3508 * check that we're out of range of the pageout scanner. It starts to
3509 * schedule paging if freemem is less than lotsfree and needfree.
3510 * lotsfree is the high-water mark for pageout, and needfree is the
3511 * number of needed free pages. We add extra pages here to make sure
3512 * the scanner doesn't start up while we're freeing memory.
3514 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3521 * check to make sure that swapfs has enough space so that anon
3522 * reservations can still succeed. anon_resvmem() checks that the
3523 * availrmem is greater than swapfs_minfree, and the number of reserved
3524 * swap pages. We also add a bit of extra here just to prevent
3525 * circumstances from getting really dire.
3527 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3528 desfree - arc_swapfs_reserve);
3531 r = FMR_SWAPFS_MINFREE;
3536 * Check that we have enough availrmem that memory locking (e.g., via
3537 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3538 * stores the number of pages that cannot be locked; when availrmem
3539 * drops below pages_pp_maximum, page locking mechanisms such as
3540 * page_pp_lock() will fail.)
3542 n = PAGESIZE * (availrmem - pages_pp_maximum -
3543 arc_pages_pp_reserve);
3546 r = FMR_PAGES_PP_MAXIMUM;
3549 #endif /* illumos */
3550 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
3552 * If we're on an i386 platform, it's possible that we'll exhaust the
3553 * kernel heap space before we ever run out of available physical
3554 * memory. Most checks of the size of the heap_area compare against
3555 * tune.t_minarmem, which is the minimum available real memory that we
3556 * can have in the system. However, this is generally fixed at 25 pages
3557 * which is so low that it's useless. In this comparison, we seek to
3558 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3559 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3562 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
3563 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3568 #define zio_arena NULL
3570 #define zio_arena heap_arena
3574 * If zio data pages are being allocated out of a separate heap segment,
3575 * then enforce that the size of available vmem for this arena remains
3576 * above about 1/16th free.
3578 * Note: The 1/16th arena free requirement was put in place
3579 * to aggressively evict memory from the arc in order to avoid
3580 * memory fragmentation issues.
3582 if (zio_arena != NULL) {
3583 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
3584 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3592 * Above limits know nothing about real level of KVA fragmentation.
3593 * Start aggressive reclamation if too little sequential KVA left.
3596 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ?
3597 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) :
3606 /* Every 100 calls, free a small amount */
3607 if (spa_get_random(100) == 0)
3609 #endif /* _KERNEL */
3611 last_free_memory = lowest;
3612 last_free_reason = r;
3613 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
3619 * Determine if the system is under memory pressure and is asking
3620 * to reclaim memory. A return value of TRUE indicates that the system
3621 * is under memory pressure and that the arc should adjust accordingly.
3624 arc_reclaim_needed(void)
3626 return (arc_available_memory() < 0);
3629 extern kmem_cache_t *zio_buf_cache[];
3630 extern kmem_cache_t *zio_data_buf_cache[];
3631 extern kmem_cache_t *range_seg_cache;
3633 static __noinline void
3634 arc_kmem_reap_now(void)
3637 kmem_cache_t *prev_cache = NULL;
3638 kmem_cache_t *prev_data_cache = NULL;
3640 DTRACE_PROBE(arc__kmem_reap_start);
3642 if (arc_meta_used >= arc_meta_limit) {
3644 * We are exceeding our meta-data cache limit.
3645 * Purge some DNLC entries to release holds on meta-data.
3647 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
3651 * Reclaim unused memory from all kmem caches.
3657 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
3658 if (zio_buf_cache[i] != prev_cache) {
3659 prev_cache = zio_buf_cache[i];
3660 kmem_cache_reap_now(zio_buf_cache[i]);
3662 if (zio_data_buf_cache[i] != prev_data_cache) {
3663 prev_data_cache = zio_data_buf_cache[i];
3664 kmem_cache_reap_now(zio_data_buf_cache[i]);
3667 kmem_cache_reap_now(buf_cache);
3668 kmem_cache_reap_now(hdr_full_cache);
3669 kmem_cache_reap_now(hdr_l2only_cache);
3670 kmem_cache_reap_now(range_seg_cache);
3673 if (zio_arena != NULL) {
3675 * Ask the vmem arena to reclaim unused memory from its
3678 vmem_qcache_reap(zio_arena);
3681 DTRACE_PROBE(arc__kmem_reap_end);
3685 * Threads can block in arc_get_data_buf() waiting for this thread to evict
3686 * enough data and signal them to proceed. When this happens, the threads in
3687 * arc_get_data_buf() are sleeping while holding the hash lock for their
3688 * particular arc header. Thus, we must be careful to never sleep on a
3689 * hash lock in this thread. This is to prevent the following deadlock:
3691 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3692 * waiting for the reclaim thread to signal it.
3694 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3695 * fails, and goes to sleep forever.
3697 * This possible deadlock is avoided by always acquiring a hash lock
3698 * using mutex_tryenter() from arc_reclaim_thread().
3701 arc_reclaim_thread(void *dummy __unused)
3703 hrtime_t growtime = 0;
3706 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
3708 mutex_enter(&arc_reclaim_lock);
3709 while (!arc_reclaim_thread_exit) {
3710 int64_t free_memory = arc_available_memory();
3711 uint64_t evicted = 0;
3713 mutex_exit(&arc_reclaim_lock);
3715 if (free_memory < 0) {
3717 arc_no_grow = B_TRUE;
3721 * Wait at least zfs_grow_retry (default 60) seconds
3722 * before considering growing.
3724 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
3726 arc_kmem_reap_now();
3729 * If we are still low on memory, shrink the ARC
3730 * so that we have arc_shrink_min free space.
3732 free_memory = arc_available_memory();
3735 (arc_c >> arc_shrink_shift) - free_memory;
3738 to_free = MAX(to_free, ptob(needfree));
3740 arc_shrink(to_free);
3742 } else if (free_memory < arc_c >> arc_no_grow_shift) {
3743 arc_no_grow = B_TRUE;
3744 } else if (gethrtime() >= growtime) {
3745 arc_no_grow = B_FALSE;
3748 evicted = arc_adjust();
3750 mutex_enter(&arc_reclaim_lock);
3753 * If evicted is zero, we couldn't evict anything via
3754 * arc_adjust(). This could be due to hash lock
3755 * collisions, but more likely due to the majority of
3756 * arc buffers being unevictable. Therefore, even if
3757 * arc_size is above arc_c, another pass is unlikely to
3758 * be helpful and could potentially cause us to enter an
3761 if (arc_size <= arc_c || evicted == 0) {
3766 * We're either no longer overflowing, or we
3767 * can't evict anything more, so we should wake
3768 * up any threads before we go to sleep.
3770 cv_broadcast(&arc_reclaim_waiters_cv);
3773 * Block until signaled, or after one second (we
3774 * might need to perform arc_kmem_reap_now()
3775 * even if we aren't being signalled)
3777 CALLB_CPR_SAFE_BEGIN(&cpr);
3778 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
3779 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
3780 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
3784 arc_reclaim_thread_exit = FALSE;
3785 cv_broadcast(&arc_reclaim_thread_cv);
3786 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
3791 arc_user_evicts_thread(void *dummy __unused)
3795 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG);
3797 mutex_enter(&arc_user_evicts_lock);
3798 while (!arc_user_evicts_thread_exit) {
3799 mutex_exit(&arc_user_evicts_lock);
3801 arc_do_user_evicts();
3804 * This is necessary in order for the mdb ::arc dcmd to
3805 * show up to date information. Since the ::arc command
3806 * does not call the kstat's update function, without
3807 * this call, the command may show stale stats for the
3808 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3809 * with this change, the data might be up to 1 second
3810 * out of date; but that should suffice. The arc_state_t
3811 * structures can be queried directly if more accurate
3812 * information is needed.
3814 if (arc_ksp != NULL)
3815 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
3817 mutex_enter(&arc_user_evicts_lock);
3820 * Block until signaled, or after one second (we need to
3821 * call the arc's kstat update function regularly).
3823 CALLB_CPR_SAFE_BEGIN(&cpr);
3824 (void) cv_timedwait(&arc_user_evicts_cv,
3825 &arc_user_evicts_lock, hz);
3826 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock);
3829 arc_user_evicts_thread_exit = FALSE;
3830 cv_broadcast(&arc_user_evicts_cv);
3831 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */
3836 * Adapt arc info given the number of bytes we are trying to add and
3837 * the state that we are comming from. This function is only called
3838 * when we are adding new content to the cache.
3841 arc_adapt(int bytes, arc_state_t *state)
3844 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
3845 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
3846 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
3848 if (state == arc_l2c_only)
3853 * Adapt the target size of the MRU list:
3854 * - if we just hit in the MRU ghost list, then increase
3855 * the target size of the MRU list.
3856 * - if we just hit in the MFU ghost list, then increase
3857 * the target size of the MFU list by decreasing the
3858 * target size of the MRU list.
3860 if (state == arc_mru_ghost) {
3861 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
3862 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
3864 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
3865 } else if (state == arc_mfu_ghost) {
3868 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
3869 mult = MIN(mult, 10);
3871 delta = MIN(bytes * mult, arc_p);
3872 arc_p = MAX(arc_p_min, arc_p - delta);
3874 ASSERT((int64_t)arc_p >= 0);
3876 if (arc_reclaim_needed()) {
3877 cv_signal(&arc_reclaim_thread_cv);
3884 if (arc_c >= arc_c_max)
3888 * If we're within (2 * maxblocksize) bytes of the target
3889 * cache size, increment the target cache size
3891 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
3892 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
3893 atomic_add_64(&arc_c, (int64_t)bytes);
3894 if (arc_c > arc_c_max)
3896 else if (state == arc_anon)
3897 atomic_add_64(&arc_p, (int64_t)bytes);
3901 ASSERT((int64_t)arc_p >= 0);
3905 * Check if arc_size has grown past our upper threshold, determined by
3906 * zfs_arc_overflow_shift.
3909 arc_is_overflowing(void)
3911 /* Always allow at least one block of overflow */
3912 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
3913 arc_c >> zfs_arc_overflow_shift);
3915 return (arc_size >= arc_c + overflow);
3919 * The buffer, supplied as the first argument, needs a data block. If we
3920 * are hitting the hard limit for the cache size, we must sleep, waiting
3921 * for the eviction thread to catch up. If we're past the target size
3922 * but below the hard limit, we'll only signal the reclaim thread and
3926 arc_get_data_buf(arc_buf_t *buf)
3928 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
3929 uint64_t size = buf->b_hdr->b_size;
3930 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
3932 arc_adapt(size, state);
3935 * If arc_size is currently overflowing, and has grown past our
3936 * upper limit, we must be adding data faster than the evict
3937 * thread can evict. Thus, to ensure we don't compound the
3938 * problem by adding more data and forcing arc_size to grow even
3939 * further past it's target size, we halt and wait for the
3940 * eviction thread to catch up.
3942 * It's also possible that the reclaim thread is unable to evict
3943 * enough buffers to get arc_size below the overflow limit (e.g.
3944 * due to buffers being un-evictable, or hash lock collisions).
3945 * In this case, we want to proceed regardless if we're
3946 * overflowing; thus we don't use a while loop here.
3948 if (arc_is_overflowing()) {
3949 mutex_enter(&arc_reclaim_lock);
3952 * Now that we've acquired the lock, we may no longer be
3953 * over the overflow limit, lets check.
3955 * We're ignoring the case of spurious wake ups. If that
3956 * were to happen, it'd let this thread consume an ARC
3957 * buffer before it should have (i.e. before we're under
3958 * the overflow limit and were signalled by the reclaim
3959 * thread). As long as that is a rare occurrence, it
3960 * shouldn't cause any harm.
3962 if (arc_is_overflowing()) {
3963 cv_signal(&arc_reclaim_thread_cv);
3964 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
3967 mutex_exit(&arc_reclaim_lock);
3970 if (type == ARC_BUFC_METADATA) {
3971 buf->b_data = zio_buf_alloc(size);
3972 arc_space_consume(size, ARC_SPACE_META);
3974 ASSERT(type == ARC_BUFC_DATA);
3975 buf->b_data = zio_data_buf_alloc(size);
3976 arc_space_consume(size, ARC_SPACE_DATA);
3980 * Update the state size. Note that ghost states have a
3981 * "ghost size" and so don't need to be updated.
3983 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) {
3984 arc_buf_hdr_t *hdr = buf->b_hdr;
3985 arc_state_t *state = hdr->b_l1hdr.b_state;
3987 (void) refcount_add_many(&state->arcs_size, size, buf);
3990 * If this is reached via arc_read, the link is
3991 * protected by the hash lock. If reached via
3992 * arc_buf_alloc, the header should not be accessed by
3993 * any other thread. And, if reached via arc_read_done,
3994 * the hash lock will protect it if it's found in the
3995 * hash table; otherwise no other thread should be
3996 * trying to [add|remove]_reference it.
3998 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3999 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4000 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type],
4004 * If we are growing the cache, and we are adding anonymous
4005 * data, and we have outgrown arc_p, update arc_p
4007 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
4008 (refcount_count(&arc_anon->arcs_size) +
4009 refcount_count(&arc_mru->arcs_size) > arc_p))
4010 arc_p = MIN(arc_c, arc_p + size);
4012 ARCSTAT_BUMP(arcstat_allocated);
4016 * This routine is called whenever a buffer is accessed.
4017 * NOTE: the hash lock is dropped in this function.
4020 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4024 ASSERT(MUTEX_HELD(hash_lock));
4025 ASSERT(HDR_HAS_L1HDR(hdr));
4027 if (hdr->b_l1hdr.b_state == arc_anon) {
4029 * This buffer is not in the cache, and does not
4030 * appear in our "ghost" list. Add the new buffer
4034 ASSERT0(hdr->b_l1hdr.b_arc_access);
4035 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4036 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4037 arc_change_state(arc_mru, hdr, hash_lock);
4039 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4040 now = ddi_get_lbolt();
4043 * If this buffer is here because of a prefetch, then either:
4044 * - clear the flag if this is a "referencing" read
4045 * (any subsequent access will bump this into the MFU state).
4047 * - move the buffer to the head of the list if this is
4048 * another prefetch (to make it less likely to be evicted).
4050 if (HDR_PREFETCH(hdr)) {
4051 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4052 /* link protected by hash lock */
4053 ASSERT(multilist_link_active(
4054 &hdr->b_l1hdr.b_arc_node));
4056 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
4057 ARCSTAT_BUMP(arcstat_mru_hits);
4059 hdr->b_l1hdr.b_arc_access = now;
4064 * This buffer has been "accessed" only once so far,
4065 * but it is still in the cache. Move it to the MFU
4068 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4070 * More than 125ms have passed since we
4071 * instantiated this buffer. Move it to the
4072 * most frequently used state.
4074 hdr->b_l1hdr.b_arc_access = now;
4075 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4076 arc_change_state(arc_mfu, hdr, hash_lock);
4078 ARCSTAT_BUMP(arcstat_mru_hits);
4079 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
4080 arc_state_t *new_state;
4082 * This buffer has been "accessed" recently, but
4083 * was evicted from the cache. Move it to the
4087 if (HDR_PREFETCH(hdr)) {
4088 new_state = arc_mru;
4089 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
4090 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
4091 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4093 new_state = arc_mfu;
4094 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4097 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4098 arc_change_state(new_state, hdr, hash_lock);
4100 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
4101 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
4103 * This buffer has been accessed more than once and is
4104 * still in the cache. Keep it in the MFU state.
4106 * NOTE: an add_reference() that occurred when we did
4107 * the arc_read() will have kicked this off the list.
4108 * If it was a prefetch, we will explicitly move it to
4109 * the head of the list now.
4111 if ((HDR_PREFETCH(hdr)) != 0) {
4112 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4113 /* link protected by hash_lock */
4114 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4116 ARCSTAT_BUMP(arcstat_mfu_hits);
4117 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4118 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4119 arc_state_t *new_state = arc_mfu;
4121 * This buffer has been accessed more than once but has
4122 * been evicted from the cache. Move it back to the
4126 if (HDR_PREFETCH(hdr)) {
4128 * This is a prefetch access...
4129 * move this block back to the MRU state.
4131 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4132 new_state = arc_mru;
4135 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4136 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4137 arc_change_state(new_state, hdr, hash_lock);
4139 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4140 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4142 * This buffer is on the 2nd Level ARC.
4145 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4146 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4147 arc_change_state(arc_mfu, hdr, hash_lock);
4149 ASSERT(!"invalid arc state");
4153 /* a generic arc_done_func_t which you can use */
4156 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4158 if (zio == NULL || zio->io_error == 0)
4159 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
4160 VERIFY(arc_buf_remove_ref(buf, arg));
4163 /* a generic arc_done_func_t */
4165 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4167 arc_buf_t **bufp = arg;
4168 if (zio && zio->io_error) {
4169 VERIFY(arc_buf_remove_ref(buf, arg));
4173 ASSERT(buf->b_data);
4178 arc_read_done(zio_t *zio)
4182 arc_buf_t *abuf; /* buffer we're assigning to callback */
4183 kmutex_t *hash_lock = NULL;
4184 arc_callback_t *callback_list, *acb;
4185 int freeable = FALSE;
4187 buf = zio->io_private;
4191 * The hdr was inserted into hash-table and removed from lists
4192 * prior to starting I/O. We should find this header, since
4193 * it's in the hash table, and it should be legit since it's
4194 * not possible to evict it during the I/O. The only possible
4195 * reason for it not to be found is if we were freed during the
4198 if (HDR_IN_HASH_TABLE(hdr)) {
4199 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
4200 ASSERT3U(hdr->b_dva.dva_word[0], ==,
4201 BP_IDENTITY(zio->io_bp)->dva_word[0]);
4202 ASSERT3U(hdr->b_dva.dva_word[1], ==,
4203 BP_IDENTITY(zio->io_bp)->dva_word[1]);
4205 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
4208 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
4209 hash_lock == NULL) ||
4211 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
4212 (found == hdr && HDR_L2_READING(hdr)));
4215 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED;
4216 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
4217 hdr->b_flags &= ~ARC_FLAG_L2CACHE;
4219 /* byteswap if necessary */
4220 callback_list = hdr->b_l1hdr.b_acb;
4221 ASSERT(callback_list != NULL);
4222 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
4223 dmu_object_byteswap_t bswap =
4224 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
4225 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
4226 byteswap_uint64_array :
4227 dmu_ot_byteswap[bswap].ob_func;
4228 func(buf->b_data, hdr->b_size);
4231 arc_cksum_compute(buf, B_FALSE);
4236 if (hash_lock && zio->io_error == 0 &&
4237 hdr->b_l1hdr.b_state == arc_anon) {
4239 * Only call arc_access on anonymous buffers. This is because
4240 * if we've issued an I/O for an evicted buffer, we've already
4241 * called arc_access (to prevent any simultaneous readers from
4242 * getting confused).
4244 arc_access(hdr, hash_lock);
4247 /* create copies of the data buffer for the callers */
4249 for (acb = callback_list; acb; acb = acb->acb_next) {
4250 if (acb->acb_done) {
4252 ARCSTAT_BUMP(arcstat_duplicate_reads);
4253 abuf = arc_buf_clone(buf);
4255 acb->acb_buf = abuf;
4259 hdr->b_l1hdr.b_acb = NULL;
4260 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4261 ASSERT(!HDR_BUF_AVAILABLE(hdr));
4263 ASSERT(buf->b_efunc == NULL);
4264 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
4265 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4268 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
4269 callback_list != NULL);
4271 if (zio->io_error != 0) {
4272 hdr->b_flags |= ARC_FLAG_IO_ERROR;
4273 if (hdr->b_l1hdr.b_state != arc_anon)
4274 arc_change_state(arc_anon, hdr, hash_lock);
4275 if (HDR_IN_HASH_TABLE(hdr))
4276 buf_hash_remove(hdr);
4277 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4281 * Broadcast before we drop the hash_lock to avoid the possibility
4282 * that the hdr (and hence the cv) might be freed before we get to
4283 * the cv_broadcast().
4285 cv_broadcast(&hdr->b_l1hdr.b_cv);
4287 if (hash_lock != NULL) {
4288 mutex_exit(hash_lock);
4291 * This block was freed while we waited for the read to
4292 * complete. It has been removed from the hash table and
4293 * moved to the anonymous state (so that it won't show up
4296 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
4297 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4300 /* execute each callback and free its structure */
4301 while ((acb = callback_list) != NULL) {
4303 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
4305 if (acb->acb_zio_dummy != NULL) {
4306 acb->acb_zio_dummy->io_error = zio->io_error;
4307 zio_nowait(acb->acb_zio_dummy);
4310 callback_list = acb->acb_next;
4311 kmem_free(acb, sizeof (arc_callback_t));
4315 arc_hdr_destroy(hdr);
4319 * "Read" the block at the specified DVA (in bp) via the
4320 * cache. If the block is found in the cache, invoke the provided
4321 * callback immediately and return. Note that the `zio' parameter
4322 * in the callback will be NULL in this case, since no IO was
4323 * required. If the block is not in the cache pass the read request
4324 * on to the spa with a substitute callback function, so that the
4325 * requested block will be added to the cache.
4327 * If a read request arrives for a block that has a read in-progress,
4328 * either wait for the in-progress read to complete (and return the
4329 * results); or, if this is a read with a "done" func, add a record
4330 * to the read to invoke the "done" func when the read completes,
4331 * and return; or just return.
4333 * arc_read_done() will invoke all the requested "done" functions
4334 * for readers of this block.
4337 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
4338 void *private, zio_priority_t priority, int zio_flags,
4339 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
4341 arc_buf_hdr_t *hdr = NULL;
4342 arc_buf_t *buf = NULL;
4343 kmutex_t *hash_lock = NULL;
4345 uint64_t guid = spa_load_guid(spa);
4347 ASSERT(!BP_IS_EMBEDDED(bp) ||
4348 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
4351 if (!BP_IS_EMBEDDED(bp)) {
4353 * Embedded BP's have no DVA and require no I/O to "read".
4354 * Create an anonymous arc buf to back it.
4356 hdr = buf_hash_find(guid, bp, &hash_lock);
4359 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) {
4361 *arc_flags |= ARC_FLAG_CACHED;
4363 if (HDR_IO_IN_PROGRESS(hdr)) {
4365 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
4366 priority == ZIO_PRIORITY_SYNC_READ) {
4368 * This sync read must wait for an
4369 * in-progress async read (e.g. a predictive
4370 * prefetch). Async reads are queued
4371 * separately at the vdev_queue layer, so
4372 * this is a form of priority inversion.
4373 * Ideally, we would "inherit" the demand
4374 * i/o's priority by moving the i/o from
4375 * the async queue to the synchronous queue,
4376 * but there is currently no mechanism to do
4377 * so. Track this so that we can evaluate
4378 * the magnitude of this potential performance
4381 * Note that if the prefetch i/o is already
4382 * active (has been issued to the device),
4383 * the prefetch improved performance, because
4384 * we issued it sooner than we would have
4385 * without the prefetch.
4387 DTRACE_PROBE1(arc__sync__wait__for__async,
4388 arc_buf_hdr_t *, hdr);
4389 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
4391 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4392 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH;
4395 if (*arc_flags & ARC_FLAG_WAIT) {
4396 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4397 mutex_exit(hash_lock);
4400 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4403 arc_callback_t *acb = NULL;
4405 acb = kmem_zalloc(sizeof (arc_callback_t),
4407 acb->acb_done = done;
4408 acb->acb_private = private;
4410 acb->acb_zio_dummy = zio_null(pio,
4411 spa, NULL, NULL, NULL, zio_flags);
4413 ASSERT(acb->acb_done != NULL);
4414 acb->acb_next = hdr->b_l1hdr.b_acb;
4415 hdr->b_l1hdr.b_acb = acb;
4416 add_reference(hdr, hash_lock, private);
4417 mutex_exit(hash_lock);
4420 mutex_exit(hash_lock);
4424 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4425 hdr->b_l1hdr.b_state == arc_mfu);
4428 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4430 * This is a demand read which does not have to
4431 * wait for i/o because we did a predictive
4432 * prefetch i/o for it, which has completed.
4435 arc__demand__hit__predictive__prefetch,
4436 arc_buf_hdr_t *, hdr);
4438 arcstat_demand_hit_predictive_prefetch);
4439 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH;
4441 add_reference(hdr, hash_lock, private);
4443 * If this block is already in use, create a new
4444 * copy of the data so that we will be guaranteed
4445 * that arc_release() will always succeed.
4447 buf = hdr->b_l1hdr.b_buf;
4449 ASSERT(buf->b_data);
4450 if (HDR_BUF_AVAILABLE(hdr)) {
4451 ASSERT(buf->b_efunc == NULL);
4452 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4454 buf = arc_buf_clone(buf);
4457 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4458 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4459 hdr->b_flags |= ARC_FLAG_PREFETCH;
4461 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4462 arc_access(hdr, hash_lock);
4463 if (*arc_flags & ARC_FLAG_L2CACHE)
4464 hdr->b_flags |= ARC_FLAG_L2CACHE;
4465 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4466 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4467 mutex_exit(hash_lock);
4468 ARCSTAT_BUMP(arcstat_hits);
4469 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4470 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4471 data, metadata, hits);
4474 done(NULL, buf, private);
4476 uint64_t size = BP_GET_LSIZE(bp);
4477 arc_callback_t *acb;
4480 boolean_t devw = B_FALSE;
4481 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
4482 int32_t b_asize = 0;
4485 /* this block is not in the cache */
4486 arc_buf_hdr_t *exists = NULL;
4487 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4488 buf = arc_buf_alloc(spa, size, private, type);
4490 if (!BP_IS_EMBEDDED(bp)) {
4491 hdr->b_dva = *BP_IDENTITY(bp);
4492 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4493 exists = buf_hash_insert(hdr, &hash_lock);
4495 if (exists != NULL) {
4496 /* somebody beat us to the hash insert */
4497 mutex_exit(hash_lock);
4498 buf_discard_identity(hdr);
4499 (void) arc_buf_remove_ref(buf, private);
4500 goto top; /* restart the IO request */
4504 * If there is a callback, we pass our reference to
4505 * it; otherwise we remove our reference.
4508 (void) remove_reference(hdr, hash_lock,
4511 if (*arc_flags & ARC_FLAG_PREFETCH)
4512 hdr->b_flags |= ARC_FLAG_PREFETCH;
4513 if (*arc_flags & ARC_FLAG_L2CACHE)
4514 hdr->b_flags |= ARC_FLAG_L2CACHE;
4515 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4516 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4517 if (BP_GET_LEVEL(bp) > 0)
4518 hdr->b_flags |= ARC_FLAG_INDIRECT;
4521 * This block is in the ghost cache. If it was L2-only
4522 * (and thus didn't have an L1 hdr), we realloc the
4523 * header to add an L1 hdr.
4525 if (!HDR_HAS_L1HDR(hdr)) {
4526 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4530 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4531 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4532 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4533 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4536 * If there is a callback, we pass a reference to it.
4539 add_reference(hdr, hash_lock, private);
4540 if (*arc_flags & ARC_FLAG_PREFETCH)
4541 hdr->b_flags |= ARC_FLAG_PREFETCH;
4542 if (*arc_flags & ARC_FLAG_L2CACHE)
4543 hdr->b_flags |= ARC_FLAG_L2CACHE;
4544 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4545 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4546 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
4549 buf->b_efunc = NULL;
4550 buf->b_private = NULL;
4552 hdr->b_l1hdr.b_buf = buf;
4553 ASSERT0(hdr->b_l1hdr.b_datacnt);
4554 hdr->b_l1hdr.b_datacnt = 1;
4555 arc_get_data_buf(buf);
4556 arc_access(hdr, hash_lock);
4559 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
4560 hdr->b_flags |= ARC_FLAG_PREDICTIVE_PREFETCH;
4561 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4563 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4564 acb->acb_done = done;
4565 acb->acb_private = private;
4567 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4568 hdr->b_l1hdr.b_acb = acb;
4569 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4571 if (HDR_HAS_L2HDR(hdr) &&
4572 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
4573 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
4574 addr = hdr->b_l2hdr.b_daddr;
4575 b_compress = hdr->b_l2hdr.b_compress;
4576 b_asize = hdr->b_l2hdr.b_asize;
4578 * Lock out device removal.
4580 if (vdev_is_dead(vd) ||
4581 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
4585 if (hash_lock != NULL)
4586 mutex_exit(hash_lock);
4589 * At this point, we have a level 1 cache miss. Try again in
4590 * L2ARC if possible.
4592 ASSERT3U(hdr->b_size, ==, size);
4593 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
4594 uint64_t, size, zbookmark_phys_t *, zb);
4595 ARCSTAT_BUMP(arcstat_misses);
4596 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4597 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4598 data, metadata, misses);
4600 curthread->td_ru.ru_inblock++;
4603 if (priority == ZIO_PRIORITY_ASYNC_READ)
4604 hdr->b_flags |= ARC_FLAG_PRIO_ASYNC_READ;
4606 hdr->b_flags &= ~ARC_FLAG_PRIO_ASYNC_READ;
4608 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
4610 * Read from the L2ARC if the following are true:
4611 * 1. The L2ARC vdev was previously cached.
4612 * 2. This buffer still has L2ARC metadata.
4613 * 3. This buffer isn't currently writing to the L2ARC.
4614 * 4. The L2ARC entry wasn't evicted, which may
4615 * also have invalidated the vdev.
4616 * 5. This isn't prefetch and l2arc_noprefetch is set.
4618 if (HDR_HAS_L2HDR(hdr) &&
4619 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
4620 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
4621 l2arc_read_callback_t *cb;
4624 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
4625 ARCSTAT_BUMP(arcstat_l2_hits);
4627 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
4629 cb->l2rcb_buf = buf;
4630 cb->l2rcb_spa = spa;
4633 cb->l2rcb_flags = zio_flags;
4634 cb->l2rcb_compress = b_compress;
4635 if (b_asize > hdr->b_size) {
4636 ASSERT3U(b_compress, ==,
4638 b_data = zio_data_buf_alloc(b_asize);
4639 cb->l2rcb_data = b_data;
4641 b_data = buf->b_data;
4644 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
4645 addr + size < vd->vdev_psize -
4646 VDEV_LABEL_END_SIZE);
4649 * l2arc read. The SCL_L2ARC lock will be
4650 * released by l2arc_read_done().
4651 * Issue a null zio if the underlying buffer
4652 * was squashed to zero size by compression.
4654 if (b_compress == ZIO_COMPRESS_EMPTY) {
4655 ASSERT3U(b_asize, ==, 0);
4656 rzio = zio_null(pio, spa, vd,
4657 l2arc_read_done, cb,
4658 zio_flags | ZIO_FLAG_DONT_CACHE |
4660 ZIO_FLAG_DONT_PROPAGATE |
4661 ZIO_FLAG_DONT_RETRY);
4663 rzio = zio_read_phys(pio, vd, addr,
4666 l2arc_read_done, cb, priority,
4667 zio_flags | ZIO_FLAG_DONT_CACHE |
4669 ZIO_FLAG_DONT_PROPAGATE |
4670 ZIO_FLAG_DONT_RETRY, B_FALSE);
4672 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
4674 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
4676 if (*arc_flags & ARC_FLAG_NOWAIT) {
4681 ASSERT(*arc_flags & ARC_FLAG_WAIT);
4682 if (zio_wait(rzio) == 0)
4685 /* l2arc read error; goto zio_read() */
4687 DTRACE_PROBE1(l2arc__miss,
4688 arc_buf_hdr_t *, hdr);
4689 ARCSTAT_BUMP(arcstat_l2_misses);
4690 if (HDR_L2_WRITING(hdr))
4691 ARCSTAT_BUMP(arcstat_l2_rw_clash);
4692 spa_config_exit(spa, SCL_L2ARC, vd);
4696 spa_config_exit(spa, SCL_L2ARC, vd);
4697 if (l2arc_ndev != 0) {
4698 DTRACE_PROBE1(l2arc__miss,
4699 arc_buf_hdr_t *, hdr);
4700 ARCSTAT_BUMP(arcstat_l2_misses);
4704 rzio = zio_read(pio, spa, bp, buf->b_data, size,
4705 arc_read_done, buf, priority, zio_flags, zb);
4707 if (*arc_flags & ARC_FLAG_WAIT)
4708 return (zio_wait(rzio));
4710 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4717 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
4719 ASSERT(buf->b_hdr != NULL);
4720 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon);
4721 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) ||
4723 ASSERT(buf->b_efunc == NULL);
4724 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
4726 buf->b_efunc = func;
4727 buf->b_private = private;
4731 * Notify the arc that a block was freed, and thus will never be used again.
4734 arc_freed(spa_t *spa, const blkptr_t *bp)
4737 kmutex_t *hash_lock;
4738 uint64_t guid = spa_load_guid(spa);
4740 ASSERT(!BP_IS_EMBEDDED(bp));
4742 hdr = buf_hash_find(guid, bp, &hash_lock);
4745 if (HDR_BUF_AVAILABLE(hdr)) {
4746 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
4747 add_reference(hdr, hash_lock, FTAG);
4748 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4749 mutex_exit(hash_lock);
4751 arc_release(buf, FTAG);
4752 (void) arc_buf_remove_ref(buf, FTAG);
4754 mutex_exit(hash_lock);
4760 * Clear the user eviction callback set by arc_set_callback(), first calling
4761 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4762 * clearing the callback may result in the arc_buf being destroyed. However,
4763 * it will not result in the *last* arc_buf being destroyed, hence the data
4764 * will remain cached in the ARC. We make a copy of the arc buffer here so
4765 * that we can process the callback without holding any locks.
4767 * It's possible that the callback is already in the process of being cleared
4768 * by another thread. In this case we can not clear the callback.
4770 * Returns B_TRUE if the callback was successfully called and cleared.
4773 arc_clear_callback(arc_buf_t *buf)
4776 kmutex_t *hash_lock;
4777 arc_evict_func_t *efunc = buf->b_efunc;
4778 void *private = buf->b_private;
4780 mutex_enter(&buf->b_evict_lock);
4784 * We are in arc_do_user_evicts().
4786 ASSERT(buf->b_data == NULL);
4787 mutex_exit(&buf->b_evict_lock);
4789 } else if (buf->b_data == NULL) {
4791 * We are on the eviction list; process this buffer now
4792 * but let arc_do_user_evicts() do the reaping.
4794 buf->b_efunc = NULL;
4795 mutex_exit(&buf->b_evict_lock);
4796 VERIFY0(efunc(private));
4799 hash_lock = HDR_LOCK(hdr);
4800 mutex_enter(hash_lock);
4802 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4804 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <,
4805 hdr->b_l1hdr.b_datacnt);
4806 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4807 hdr->b_l1hdr.b_state == arc_mfu);
4809 buf->b_efunc = NULL;
4810 buf->b_private = NULL;
4812 if (hdr->b_l1hdr.b_datacnt > 1) {
4813 mutex_exit(&buf->b_evict_lock);
4814 arc_buf_destroy(buf, TRUE);
4816 ASSERT(buf == hdr->b_l1hdr.b_buf);
4817 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4818 mutex_exit(&buf->b_evict_lock);
4821 mutex_exit(hash_lock);
4822 VERIFY0(efunc(private));
4827 * Release this buffer from the cache, making it an anonymous buffer. This
4828 * must be done after a read and prior to modifying the buffer contents.
4829 * If the buffer has more than one reference, we must make
4830 * a new hdr for the buffer.
4833 arc_release(arc_buf_t *buf, void *tag)
4835 arc_buf_hdr_t *hdr = buf->b_hdr;
4838 * It would be nice to assert that if it's DMU metadata (level >
4839 * 0 || it's the dnode file), then it must be syncing context.
4840 * But we don't know that information at this level.
4843 mutex_enter(&buf->b_evict_lock);
4845 ASSERT(HDR_HAS_L1HDR(hdr));
4848 * We don't grab the hash lock prior to this check, because if
4849 * the buffer's header is in the arc_anon state, it won't be
4850 * linked into the hash table.
4852 if (hdr->b_l1hdr.b_state == arc_anon) {
4853 mutex_exit(&buf->b_evict_lock);
4854 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4855 ASSERT(!HDR_IN_HASH_TABLE(hdr));
4856 ASSERT(!HDR_HAS_L2HDR(hdr));
4857 ASSERT(BUF_EMPTY(hdr));
4858 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1);
4859 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
4860 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
4862 ASSERT3P(buf->b_efunc, ==, NULL);
4863 ASSERT3P(buf->b_private, ==, NULL);
4865 hdr->b_l1hdr.b_arc_access = 0;
4871 kmutex_t *hash_lock = HDR_LOCK(hdr);
4872 mutex_enter(hash_lock);
4875 * This assignment is only valid as long as the hash_lock is
4876 * held, we must be careful not to reference state or the
4877 * b_state field after dropping the lock.
4879 arc_state_t *state = hdr->b_l1hdr.b_state;
4880 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4881 ASSERT3P(state, !=, arc_anon);
4883 /* this buffer is not on any list */
4884 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
4886 if (HDR_HAS_L2HDR(hdr)) {
4887 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4890 * We have to recheck this conditional again now that
4891 * we're holding the l2ad_mtx to prevent a race with
4892 * another thread which might be concurrently calling
4893 * l2arc_evict(). In that case, l2arc_evict() might have
4894 * destroyed the header's L2 portion as we were waiting
4895 * to acquire the l2ad_mtx.
4897 if (HDR_HAS_L2HDR(hdr)) {
4899 arc_hdr_l2hdr_destroy(hdr);
4902 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4906 * Do we have more than one buf?
4908 if (hdr->b_l1hdr.b_datacnt > 1) {
4909 arc_buf_hdr_t *nhdr;
4911 uint64_t blksz = hdr->b_size;
4912 uint64_t spa = hdr->b_spa;
4913 arc_buf_contents_t type = arc_buf_type(hdr);
4914 uint32_t flags = hdr->b_flags;
4916 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
4918 * Pull the data off of this hdr and attach it to
4919 * a new anonymous hdr.
4921 (void) remove_reference(hdr, hash_lock, tag);
4922 bufp = &hdr->b_l1hdr.b_buf;
4923 while (*bufp != buf)
4924 bufp = &(*bufp)->b_next;
4925 *bufp = buf->b_next;
4928 ASSERT3P(state, !=, arc_l2c_only);
4930 (void) refcount_remove_many(
4931 &state->arcs_size, hdr->b_size, buf);
4933 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
4934 ASSERT3P(state, !=, arc_l2c_only);
4935 uint64_t *size = &state->arcs_lsize[type];
4936 ASSERT3U(*size, >=, hdr->b_size);
4937 atomic_add_64(size, -hdr->b_size);
4941 * We're releasing a duplicate user data buffer, update
4942 * our statistics accordingly.
4944 if (HDR_ISTYPE_DATA(hdr)) {
4945 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
4946 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
4949 hdr->b_l1hdr.b_datacnt -= 1;
4950 arc_cksum_verify(buf);
4952 arc_buf_unwatch(buf);
4955 mutex_exit(hash_lock);
4957 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
4958 nhdr->b_size = blksz;
4961 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING;
4962 nhdr->b_flags |= arc_bufc_to_flags(type);
4963 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
4965 nhdr->b_l1hdr.b_buf = buf;
4966 nhdr->b_l1hdr.b_datacnt = 1;
4967 nhdr->b_l1hdr.b_state = arc_anon;
4968 nhdr->b_l1hdr.b_arc_access = 0;
4969 nhdr->b_l1hdr.b_tmp_cdata = NULL;
4970 nhdr->b_freeze_cksum = NULL;
4972 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
4974 mutex_exit(&buf->b_evict_lock);
4975 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf);
4977 mutex_exit(&buf->b_evict_lock);
4978 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
4979 /* protected by hash lock, or hdr is on arc_anon */
4980 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4981 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4982 arc_change_state(arc_anon, hdr, hash_lock);
4983 hdr->b_l1hdr.b_arc_access = 0;
4984 mutex_exit(hash_lock);
4986 buf_discard_identity(hdr);
4989 buf->b_efunc = NULL;
4990 buf->b_private = NULL;
4994 arc_released(arc_buf_t *buf)
4998 mutex_enter(&buf->b_evict_lock);
4999 released = (buf->b_data != NULL &&
5000 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5001 mutex_exit(&buf->b_evict_lock);
5007 arc_referenced(arc_buf_t *buf)
5011 mutex_enter(&buf->b_evict_lock);
5012 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5013 mutex_exit(&buf->b_evict_lock);
5014 return (referenced);
5019 arc_write_ready(zio_t *zio)
5021 arc_write_callback_t *callback = zio->io_private;
5022 arc_buf_t *buf = callback->awcb_buf;
5023 arc_buf_hdr_t *hdr = buf->b_hdr;
5025 ASSERT(HDR_HAS_L1HDR(hdr));
5026 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5027 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
5028 callback->awcb_ready(zio, buf, callback->awcb_private);
5031 * If the IO is already in progress, then this is a re-write
5032 * attempt, so we need to thaw and re-compute the cksum.
5033 * It is the responsibility of the callback to handle the
5034 * accounting for any re-write attempt.
5036 if (HDR_IO_IN_PROGRESS(hdr)) {
5037 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
5038 if (hdr->b_freeze_cksum != NULL) {
5039 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
5040 hdr->b_freeze_cksum = NULL;
5042 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
5044 arc_cksum_compute(buf, B_FALSE);
5045 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
5049 * The SPA calls this callback for each physical write that happens on behalf
5050 * of a logical write. See the comment in dbuf_write_physdone() for details.
5053 arc_write_physdone(zio_t *zio)
5055 arc_write_callback_t *cb = zio->io_private;
5056 if (cb->awcb_physdone != NULL)
5057 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
5061 arc_write_done(zio_t *zio)
5063 arc_write_callback_t *callback = zio->io_private;
5064 arc_buf_t *buf = callback->awcb_buf;
5065 arc_buf_hdr_t *hdr = buf->b_hdr;
5067 ASSERT(hdr->b_l1hdr.b_acb == NULL);
5069 if (zio->io_error == 0) {
5070 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5071 buf_discard_identity(hdr);
5073 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
5074 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
5077 ASSERT(BUF_EMPTY(hdr));
5081 * If the block to be written was all-zero or compressed enough to be
5082 * embedded in the BP, no write was performed so there will be no
5083 * dva/birth/checksum. The buffer must therefore remain anonymous
5086 if (!BUF_EMPTY(hdr)) {
5087 arc_buf_hdr_t *exists;
5088 kmutex_t *hash_lock;
5090 ASSERT(zio->io_error == 0);
5092 arc_cksum_verify(buf);
5094 exists = buf_hash_insert(hdr, &hash_lock);
5095 if (exists != NULL) {
5097 * This can only happen if we overwrite for
5098 * sync-to-convergence, because we remove
5099 * buffers from the hash table when we arc_free().
5101 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5102 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5103 panic("bad overwrite, hdr=%p exists=%p",
5104 (void *)hdr, (void *)exists);
5105 ASSERT(refcount_is_zero(
5106 &exists->b_l1hdr.b_refcnt));
5107 arc_change_state(arc_anon, exists, hash_lock);
5108 mutex_exit(hash_lock);
5109 arc_hdr_destroy(exists);
5110 exists = buf_hash_insert(hdr, &hash_lock);
5111 ASSERT3P(exists, ==, NULL);
5112 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
5114 ASSERT(zio->io_prop.zp_nopwrite);
5115 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5116 panic("bad nopwrite, hdr=%p exists=%p",
5117 (void *)hdr, (void *)exists);
5120 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
5121 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
5122 ASSERT(BP_GET_DEDUP(zio->io_bp));
5123 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
5126 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
5127 /* if it's not anon, we are doing a scrub */
5128 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
5129 arc_access(hdr, hash_lock);
5130 mutex_exit(hash_lock);
5132 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
5135 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5136 callback->awcb_done(zio, buf, callback->awcb_private);
5138 kmem_free(callback, sizeof (arc_write_callback_t));
5142 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
5143 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
5144 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
5145 arc_done_func_t *done, void *private, zio_priority_t priority,
5146 int zio_flags, const zbookmark_phys_t *zb)
5148 arc_buf_hdr_t *hdr = buf->b_hdr;
5149 arc_write_callback_t *callback;
5152 ASSERT(ready != NULL);
5153 ASSERT(done != NULL);
5154 ASSERT(!HDR_IO_ERROR(hdr));
5155 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5156 ASSERT(hdr->b_l1hdr.b_acb == NULL);
5157 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
5159 hdr->b_flags |= ARC_FLAG_L2CACHE;
5161 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
5162 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
5163 callback->awcb_ready = ready;
5164 callback->awcb_physdone = physdone;
5165 callback->awcb_done = done;
5166 callback->awcb_private = private;
5167 callback->awcb_buf = buf;
5169 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
5170 arc_write_ready, arc_write_physdone, arc_write_done, callback,
5171 priority, zio_flags, zb);
5177 arc_memory_throttle(uint64_t reserve, uint64_t txg)
5180 uint64_t available_memory = ptob(freemem);
5181 static uint64_t page_load = 0;
5182 static uint64_t last_txg = 0;
5184 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
5186 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE)));
5189 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
5192 if (txg > last_txg) {
5197 * If we are in pageout, we know that memory is already tight,
5198 * the arc is already going to be evicting, so we just want to
5199 * continue to let page writes occur as quickly as possible.
5201 if (curproc == pageproc) {
5202 if (page_load > MAX(ptob(minfree), available_memory) / 4)
5203 return (SET_ERROR(ERESTART));
5204 /* Note: reserve is inflated, so we deflate */
5205 page_load += reserve / 8;
5207 } else if (page_load > 0 && arc_reclaim_needed()) {
5208 /* memory is low, delay before restarting */
5209 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
5210 return (SET_ERROR(EAGAIN));
5218 arc_tempreserve_clear(uint64_t reserve)
5220 atomic_add_64(&arc_tempreserve, -reserve);
5221 ASSERT((int64_t)arc_tempreserve >= 0);
5225 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
5230 if (reserve > arc_c/4 && !arc_no_grow) {
5231 arc_c = MIN(arc_c_max, reserve * 4);
5232 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
5234 if (reserve > arc_c)
5235 return (SET_ERROR(ENOMEM));
5238 * Don't count loaned bufs as in flight dirty data to prevent long
5239 * network delays from blocking transactions that are ready to be
5240 * assigned to a txg.
5242 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
5243 arc_loaned_bytes), 0);
5246 * Writes will, almost always, require additional memory allocations
5247 * in order to compress/encrypt/etc the data. We therefore need to
5248 * make sure that there is sufficient available memory for this.
5250 error = arc_memory_throttle(reserve, txg);
5255 * Throttle writes when the amount of dirty data in the cache
5256 * gets too large. We try to keep the cache less than half full
5257 * of dirty blocks so that our sync times don't grow too large.
5258 * Note: if two requests come in concurrently, we might let them
5259 * both succeed, when one of them should fail. Not a huge deal.
5262 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
5263 anon_size > arc_c / 4) {
5264 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5265 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5266 arc_tempreserve>>10,
5267 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
5268 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
5269 reserve>>10, arc_c>>10);
5270 return (SET_ERROR(ERESTART));
5272 atomic_add_64(&arc_tempreserve, reserve);
5277 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
5278 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
5280 size->value.ui64 = refcount_count(&state->arcs_size);
5281 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
5282 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
5286 arc_kstat_update(kstat_t *ksp, int rw)
5288 arc_stats_t *as = ksp->ks_data;
5290 if (rw == KSTAT_WRITE) {
5293 arc_kstat_update_state(arc_anon,
5294 &as->arcstat_anon_size,
5295 &as->arcstat_anon_evictable_data,
5296 &as->arcstat_anon_evictable_metadata);
5297 arc_kstat_update_state(arc_mru,
5298 &as->arcstat_mru_size,
5299 &as->arcstat_mru_evictable_data,
5300 &as->arcstat_mru_evictable_metadata);
5301 arc_kstat_update_state(arc_mru_ghost,
5302 &as->arcstat_mru_ghost_size,
5303 &as->arcstat_mru_ghost_evictable_data,
5304 &as->arcstat_mru_ghost_evictable_metadata);
5305 arc_kstat_update_state(arc_mfu,
5306 &as->arcstat_mfu_size,
5307 &as->arcstat_mfu_evictable_data,
5308 &as->arcstat_mfu_evictable_metadata);
5309 arc_kstat_update_state(arc_mfu_ghost,
5310 &as->arcstat_mfu_ghost_size,
5311 &as->arcstat_mfu_ghost_evictable_data,
5312 &as->arcstat_mfu_ghost_evictable_metadata);
5319 * This function *must* return indices evenly distributed between all
5320 * sublists of the multilist. This is needed due to how the ARC eviction
5321 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5322 * distributed between all sublists and uses this assumption when
5323 * deciding which sublist to evict from and how much to evict from it.
5326 arc_state_multilist_index_func(multilist_t *ml, void *obj)
5328 arc_buf_hdr_t *hdr = obj;
5331 * We rely on b_dva to generate evenly distributed index
5332 * numbers using buf_hash below. So, as an added precaution,
5333 * let's make sure we never add empty buffers to the arc lists.
5335 ASSERT(!BUF_EMPTY(hdr));
5338 * The assumption here, is the hash value for a given
5339 * arc_buf_hdr_t will remain constant throughout it's lifetime
5340 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
5341 * Thus, we don't need to store the header's sublist index
5342 * on insertion, as this index can be recalculated on removal.
5344 * Also, the low order bits of the hash value are thought to be
5345 * distributed evenly. Otherwise, in the case that the multilist
5346 * has a power of two number of sublists, each sublists' usage
5347 * would not be evenly distributed.
5349 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
5350 multilist_get_num_sublists(ml));
5354 static eventhandler_tag arc_event_lowmem = NULL;
5357 arc_lowmem(void *arg __unused, int howto __unused)
5360 mutex_enter(&arc_reclaim_lock);
5361 /* XXX: Memory deficit should be passed as argument. */
5362 needfree = btoc(arc_c >> arc_shrink_shift);
5363 DTRACE_PROBE(arc__needfree);
5364 cv_signal(&arc_reclaim_thread_cv);
5367 * It is unsafe to block here in arbitrary threads, because we can come
5368 * here from ARC itself and may hold ARC locks and thus risk a deadlock
5369 * with ARC reclaim thread.
5371 if (curproc == pageproc)
5372 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5373 mutex_exit(&arc_reclaim_lock);
5380 int i, prefetch_tunable_set = 0;
5382 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
5383 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
5384 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
5386 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
5387 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL);
5389 /* Convert seconds to clock ticks */
5390 arc_min_prefetch_lifespan = 1 * hz;
5392 /* Start out with 1/8 of all memory */
5393 arc_c = kmem_size() / 8;
5398 * On architectures where the physical memory can be larger
5399 * than the addressable space (intel in 32-bit mode), we may
5400 * need to limit the cache to 1/8 of VM size.
5402 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
5404 #endif /* illumos */
5405 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */
5406 arc_c_min = MAX(arc_c / 4, arc_abs_min);
5407 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
5408 if (arc_c * 8 >= 1 << 30)
5409 arc_c_max = (arc_c * 8) - (1 << 30);
5411 arc_c_max = arc_c_min;
5412 arc_c_max = MAX(arc_c * 5, arc_c_max);
5415 * In userland, there's only the memory pressure that we artificially
5416 * create (see arc_available_memory()). Don't let arc_c get too
5417 * small, because it can cause transactions to be larger than
5418 * arc_c, causing arc_tempreserve_space() to fail.
5421 arc_c_min = arc_c_max / 2;
5426 * Allow the tunables to override our calculations if they are
5429 if (zfs_arc_max > arc_abs_min && zfs_arc_max < kmem_size())
5430 arc_c_max = zfs_arc_max;
5431 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max)
5432 arc_c_min = zfs_arc_min;
5436 arc_p = (arc_c >> 1);
5438 /* limit meta-data to 1/4 of the arc capacity */
5439 arc_meta_limit = arc_c_max / 4;
5441 /* Allow the tunable to override if it is reasonable */
5442 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
5443 arc_meta_limit = zfs_arc_meta_limit;
5445 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
5446 arc_c_min = arc_meta_limit / 2;
5448 if (zfs_arc_meta_min > 0) {
5449 arc_meta_min = zfs_arc_meta_min;
5451 arc_meta_min = arc_c_min / 2;
5454 if (zfs_arc_grow_retry > 0)
5455 arc_grow_retry = zfs_arc_grow_retry;
5457 if (zfs_arc_shrink_shift > 0)
5458 arc_shrink_shift = zfs_arc_shrink_shift;
5461 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
5463 if (arc_no_grow_shift >= arc_shrink_shift)
5464 arc_no_grow_shift = arc_shrink_shift - 1;
5466 if (zfs_arc_p_min_shift > 0)
5467 arc_p_min_shift = zfs_arc_p_min_shift;
5469 if (zfs_arc_num_sublists_per_state < 1)
5470 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1);
5472 /* if kmem_flags are set, lets try to use less memory */
5473 if (kmem_debugging())
5475 if (arc_c < arc_c_min)
5478 zfs_arc_min = arc_c_min;
5479 zfs_arc_max = arc_c_max;
5481 arc_anon = &ARC_anon;
5483 arc_mru_ghost = &ARC_mru_ghost;
5485 arc_mfu_ghost = &ARC_mfu_ghost;
5486 arc_l2c_only = &ARC_l2c_only;
5489 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5490 sizeof (arc_buf_hdr_t),
5491 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5492 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5493 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5494 sizeof (arc_buf_hdr_t),
5495 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5496 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5497 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5498 sizeof (arc_buf_hdr_t),
5499 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5500 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5501 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5502 sizeof (arc_buf_hdr_t),
5503 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5504 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5505 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5506 sizeof (arc_buf_hdr_t),
5507 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5508 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5509 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5510 sizeof (arc_buf_hdr_t),
5511 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5512 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5513 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5514 sizeof (arc_buf_hdr_t),
5515 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5516 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5517 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5518 sizeof (arc_buf_hdr_t),
5519 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5520 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5521 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5522 sizeof (arc_buf_hdr_t),
5523 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5524 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5525 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
5526 sizeof (arc_buf_hdr_t),
5527 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5528 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5530 refcount_create(&arc_anon->arcs_size);
5531 refcount_create(&arc_mru->arcs_size);
5532 refcount_create(&arc_mru_ghost->arcs_size);
5533 refcount_create(&arc_mfu->arcs_size);
5534 refcount_create(&arc_mfu_ghost->arcs_size);
5535 refcount_create(&arc_l2c_only->arcs_size);
5539 arc_reclaim_thread_exit = FALSE;
5540 arc_user_evicts_thread_exit = FALSE;
5541 arc_eviction_list = NULL;
5542 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
5544 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
5545 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
5547 if (arc_ksp != NULL) {
5548 arc_ksp->ks_data = &arc_stats;
5549 arc_ksp->ks_update = arc_kstat_update;
5550 kstat_install(arc_ksp);
5553 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
5554 TS_RUN, minclsyspri);
5557 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
5558 EVENTHANDLER_PRI_FIRST);
5561 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0,
5562 TS_RUN, minclsyspri);
5568 * Calculate maximum amount of dirty data per pool.
5570 * If it has been set by /etc/system, take that.
5571 * Otherwise, use a percentage of physical memory defined by
5572 * zfs_dirty_data_max_percent (default 10%) with a cap at
5573 * zfs_dirty_data_max_max (default 4GB).
5575 if (zfs_dirty_data_max == 0) {
5576 zfs_dirty_data_max = ptob(physmem) *
5577 zfs_dirty_data_max_percent / 100;
5578 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
5579 zfs_dirty_data_max_max);
5583 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
5584 prefetch_tunable_set = 1;
5587 if (prefetch_tunable_set == 0) {
5588 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
5590 printf(" add \"vfs.zfs.prefetch_disable=0\" "
5591 "to /boot/loader.conf.\n");
5592 zfs_prefetch_disable = 1;
5595 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
5596 prefetch_tunable_set == 0) {
5597 printf("ZFS NOTICE: Prefetch is disabled by default if less "
5598 "than 4GB of RAM is present;\n"
5599 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
5600 "to /boot/loader.conf.\n");
5601 zfs_prefetch_disable = 1;
5604 /* Warn about ZFS memory and address space requirements. */
5605 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
5606 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
5607 "expect unstable behavior.\n");
5609 if (kmem_size() < 512 * (1 << 20)) {
5610 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
5611 "expect unstable behavior.\n");
5612 printf(" Consider tuning vm.kmem_size and "
5613 "vm.kmem_size_max\n");
5614 printf(" in /boot/loader.conf.\n");
5622 mutex_enter(&arc_reclaim_lock);
5623 arc_reclaim_thread_exit = TRUE;
5625 * The reclaim thread will set arc_reclaim_thread_exit back to
5626 * FALSE when it is finished exiting; we're waiting for that.
5628 while (arc_reclaim_thread_exit) {
5629 cv_signal(&arc_reclaim_thread_cv);
5630 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
5632 mutex_exit(&arc_reclaim_lock);
5634 mutex_enter(&arc_user_evicts_lock);
5635 arc_user_evicts_thread_exit = TRUE;
5637 * The user evicts thread will set arc_user_evicts_thread_exit
5638 * to FALSE when it is finished exiting; we're waiting for that.
5640 while (arc_user_evicts_thread_exit) {
5641 cv_signal(&arc_user_evicts_cv);
5642 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock);
5644 mutex_exit(&arc_user_evicts_lock);
5646 /* Use TRUE to ensure *all* buffers are evicted */
5647 arc_flush(NULL, TRUE);
5651 if (arc_ksp != NULL) {
5652 kstat_delete(arc_ksp);
5656 mutex_destroy(&arc_reclaim_lock);
5657 cv_destroy(&arc_reclaim_thread_cv);
5658 cv_destroy(&arc_reclaim_waiters_cv);
5660 mutex_destroy(&arc_user_evicts_lock);
5661 cv_destroy(&arc_user_evicts_cv);
5663 refcount_destroy(&arc_anon->arcs_size);
5664 refcount_destroy(&arc_mru->arcs_size);
5665 refcount_destroy(&arc_mru_ghost->arcs_size);
5666 refcount_destroy(&arc_mfu->arcs_size);
5667 refcount_destroy(&arc_mfu_ghost->arcs_size);
5668 refcount_destroy(&arc_l2c_only->arcs_size);
5670 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
5671 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5672 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
5673 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
5674 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
5675 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
5676 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
5677 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
5678 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
5679 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
5683 ASSERT0(arc_loaned_bytes);
5686 if (arc_event_lowmem != NULL)
5687 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
5694 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5695 * It uses dedicated storage devices to hold cached data, which are populated
5696 * using large infrequent writes. The main role of this cache is to boost
5697 * the performance of random read workloads. The intended L2ARC devices
5698 * include short-stroked disks, solid state disks, and other media with
5699 * substantially faster read latency than disk.
5701 * +-----------------------+
5703 * +-----------------------+
5706 * l2arc_feed_thread() arc_read()
5710 * +---------------+ |
5712 * +---------------+ |
5717 * +-------+ +-------+
5719 * | cache | | cache |
5720 * +-------+ +-------+
5721 * +=========+ .-----.
5722 * : L2ARC : |-_____-|
5723 * : devices : | Disks |
5724 * +=========+ `-_____-'
5726 * Read requests are satisfied from the following sources, in order:
5729 * 2) vdev cache of L2ARC devices
5731 * 4) vdev cache of disks
5734 * Some L2ARC device types exhibit extremely slow write performance.
5735 * To accommodate for this there are some significant differences between
5736 * the L2ARC and traditional cache design:
5738 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5739 * the ARC behave as usual, freeing buffers and placing headers on ghost
5740 * lists. The ARC does not send buffers to the L2ARC during eviction as
5741 * this would add inflated write latencies for all ARC memory pressure.
5743 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5744 * It does this by periodically scanning buffers from the eviction-end of
5745 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5746 * not already there. It scans until a headroom of buffers is satisfied,
5747 * which itself is a buffer for ARC eviction. If a compressible buffer is
5748 * found during scanning and selected for writing to an L2ARC device, we
5749 * temporarily boost scanning headroom during the next scan cycle to make
5750 * sure we adapt to compression effects (which might significantly reduce
5751 * the data volume we write to L2ARC). The thread that does this is
5752 * l2arc_feed_thread(), illustrated below; example sizes are included to
5753 * provide a better sense of ratio than this diagram:
5756 * +---------------------+----------+
5757 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5758 * +---------------------+----------+ | o L2ARC eligible
5759 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5760 * +---------------------+----------+ |
5761 * 15.9 Gbytes ^ 32 Mbytes |
5763 * l2arc_feed_thread()
5765 * l2arc write hand <--[oooo]--'
5769 * +==============================+
5770 * L2ARC dev |####|#|###|###| |####| ... |
5771 * +==============================+
5774 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5775 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5776 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5777 * safe to say that this is an uncommon case, since buffers at the end of
5778 * the ARC lists have moved there due to inactivity.
5780 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5781 * then the L2ARC simply misses copying some buffers. This serves as a
5782 * pressure valve to prevent heavy read workloads from both stalling the ARC
5783 * with waits and clogging the L2ARC with writes. This also helps prevent
5784 * the potential for the L2ARC to churn if it attempts to cache content too
5785 * quickly, such as during backups of the entire pool.
5787 * 5. After system boot and before the ARC has filled main memory, there are
5788 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5789 * lists can remain mostly static. Instead of searching from tail of these
5790 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5791 * for eligible buffers, greatly increasing its chance of finding them.
5793 * The L2ARC device write speed is also boosted during this time so that
5794 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5795 * there are no L2ARC reads, and no fear of degrading read performance
5796 * through increased writes.
5798 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5799 * the vdev queue can aggregate them into larger and fewer writes. Each
5800 * device is written to in a rotor fashion, sweeping writes through
5801 * available space then repeating.
5803 * 7. The L2ARC does not store dirty content. It never needs to flush
5804 * write buffers back to disk based storage.
5806 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5807 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5809 * The performance of the L2ARC can be tweaked by a number of tunables, which
5810 * may be necessary for different workloads:
5812 * l2arc_write_max max write bytes per interval
5813 * l2arc_write_boost extra write bytes during device warmup
5814 * l2arc_noprefetch skip caching prefetched buffers
5815 * l2arc_headroom number of max device writes to precache
5816 * l2arc_headroom_boost when we find compressed buffers during ARC
5817 * scanning, we multiply headroom by this
5818 * percentage factor for the next scan cycle,
5819 * since more compressed buffers are likely to
5821 * l2arc_feed_secs seconds between L2ARC writing
5823 * Tunables may be removed or added as future performance improvements are
5824 * integrated, and also may become zpool properties.
5826 * There are three key functions that control how the L2ARC warms up:
5828 * l2arc_write_eligible() check if a buffer is eligible to cache
5829 * l2arc_write_size() calculate how much to write
5830 * l2arc_write_interval() calculate sleep delay between writes
5832 * These three functions determine what to write, how much, and how quickly
5837 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
5840 * A buffer is *not* eligible for the L2ARC if it:
5841 * 1. belongs to a different spa.
5842 * 2. is already cached on the L2ARC.
5843 * 3. has an I/O in progress (it may be an incomplete read).
5844 * 4. is flagged not eligible (zfs property).
5846 if (hdr->b_spa != spa_guid) {
5847 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
5850 if (HDR_HAS_L2HDR(hdr)) {
5851 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
5854 if (HDR_IO_IN_PROGRESS(hdr)) {
5855 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
5858 if (!HDR_L2CACHE(hdr)) {
5859 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
5867 l2arc_write_size(void)
5872 * Make sure our globals have meaningful values in case the user
5875 size = l2arc_write_max;
5877 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
5878 "be greater than zero, resetting it to the default (%d)",
5880 size = l2arc_write_max = L2ARC_WRITE_SIZE;
5883 if (arc_warm == B_FALSE)
5884 size += l2arc_write_boost;
5891 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
5893 clock_t interval, next, now;
5896 * If the ARC lists are busy, increase our write rate; if the
5897 * lists are stale, idle back. This is achieved by checking
5898 * how much we previously wrote - if it was more than half of
5899 * what we wanted, schedule the next write much sooner.
5901 if (l2arc_feed_again && wrote > (wanted / 2))
5902 interval = (hz * l2arc_feed_min_ms) / 1000;
5904 interval = hz * l2arc_feed_secs;
5906 now = ddi_get_lbolt();
5907 next = MAX(now, MIN(now + interval, began + interval));
5913 * Cycle through L2ARC devices. This is how L2ARC load balances.
5914 * If a device is returned, this also returns holding the spa config lock.
5916 static l2arc_dev_t *
5917 l2arc_dev_get_next(void)
5919 l2arc_dev_t *first, *next = NULL;
5922 * Lock out the removal of spas (spa_namespace_lock), then removal
5923 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
5924 * both locks will be dropped and a spa config lock held instead.
5926 mutex_enter(&spa_namespace_lock);
5927 mutex_enter(&l2arc_dev_mtx);
5929 /* if there are no vdevs, there is nothing to do */
5930 if (l2arc_ndev == 0)
5934 next = l2arc_dev_last;
5936 /* loop around the list looking for a non-faulted vdev */
5938 next = list_head(l2arc_dev_list);
5940 next = list_next(l2arc_dev_list, next);
5942 next = list_head(l2arc_dev_list);
5945 /* if we have come back to the start, bail out */
5948 else if (next == first)
5951 } while (vdev_is_dead(next->l2ad_vdev));
5953 /* if we were unable to find any usable vdevs, return NULL */
5954 if (vdev_is_dead(next->l2ad_vdev))
5957 l2arc_dev_last = next;
5960 mutex_exit(&l2arc_dev_mtx);
5963 * Grab the config lock to prevent the 'next' device from being
5964 * removed while we are writing to it.
5967 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
5968 mutex_exit(&spa_namespace_lock);
5974 * Free buffers that were tagged for destruction.
5977 l2arc_do_free_on_write()
5980 l2arc_data_free_t *df, *df_prev;
5982 mutex_enter(&l2arc_free_on_write_mtx);
5983 buflist = l2arc_free_on_write;
5985 for (df = list_tail(buflist); df; df = df_prev) {
5986 df_prev = list_prev(buflist, df);
5987 ASSERT(df->l2df_data != NULL);
5988 ASSERT(df->l2df_func != NULL);
5989 df->l2df_func(df->l2df_data, df->l2df_size);
5990 list_remove(buflist, df);
5991 kmem_free(df, sizeof (l2arc_data_free_t));
5994 mutex_exit(&l2arc_free_on_write_mtx);
5998 * A write to a cache device has completed. Update all headers to allow
5999 * reads from these buffers to begin.
6002 l2arc_write_done(zio_t *zio)
6004 l2arc_write_callback_t *cb;
6007 arc_buf_hdr_t *head, *hdr, *hdr_prev;
6008 kmutex_t *hash_lock;
6009 int64_t bytes_dropped = 0;
6011 cb = zio->io_private;
6013 dev = cb->l2wcb_dev;
6014 ASSERT(dev != NULL);
6015 head = cb->l2wcb_head;
6016 ASSERT(head != NULL);
6017 buflist = &dev->l2ad_buflist;
6018 ASSERT(buflist != NULL);
6019 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
6020 l2arc_write_callback_t *, cb);
6022 if (zio->io_error != 0)
6023 ARCSTAT_BUMP(arcstat_l2_writes_error);
6026 * All writes completed, or an error was hit.
6029 mutex_enter(&dev->l2ad_mtx);
6030 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
6031 hdr_prev = list_prev(buflist, hdr);
6033 hash_lock = HDR_LOCK(hdr);
6036 * We cannot use mutex_enter or else we can deadlock
6037 * with l2arc_write_buffers (due to swapping the order
6038 * the hash lock and l2ad_mtx are taken).
6040 if (!mutex_tryenter(hash_lock)) {
6042 * Missed the hash lock. We must retry so we
6043 * don't leave the ARC_FLAG_L2_WRITING bit set.
6045 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
6048 * We don't want to rescan the headers we've
6049 * already marked as having been written out, so
6050 * we reinsert the head node so we can pick up
6051 * where we left off.
6053 list_remove(buflist, head);
6054 list_insert_after(buflist, hdr, head);
6056 mutex_exit(&dev->l2ad_mtx);
6059 * We wait for the hash lock to become available
6060 * to try and prevent busy waiting, and increase
6061 * the chance we'll be able to acquire the lock
6062 * the next time around.
6064 mutex_enter(hash_lock);
6065 mutex_exit(hash_lock);
6070 * We could not have been moved into the arc_l2c_only
6071 * state while in-flight due to our ARC_FLAG_L2_WRITING
6072 * bit being set. Let's just ensure that's being enforced.
6074 ASSERT(HDR_HAS_L1HDR(hdr));
6077 * We may have allocated a buffer for L2ARC compression,
6078 * we must release it to avoid leaking this data.
6080 l2arc_release_cdata_buf(hdr);
6082 if (zio->io_error != 0) {
6084 * Error - drop L2ARC entry.
6086 list_remove(buflist, hdr);
6088 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
6090 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize);
6091 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
6093 bytes_dropped += hdr->b_l2hdr.b_asize;
6094 (void) refcount_remove_many(&dev->l2ad_alloc,
6095 hdr->b_l2hdr.b_asize, hdr);
6099 * Allow ARC to begin reads and ghost list evictions to
6102 hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
6104 mutex_exit(hash_lock);
6107 atomic_inc_64(&l2arc_writes_done);
6108 list_remove(buflist, head);
6109 ASSERT(!HDR_HAS_L1HDR(head));
6110 kmem_cache_free(hdr_l2only_cache, head);
6111 mutex_exit(&dev->l2ad_mtx);
6113 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
6115 l2arc_do_free_on_write();
6117 kmem_free(cb, sizeof (l2arc_write_callback_t));
6121 * A read to a cache device completed. Validate buffer contents before
6122 * handing over to the regular ARC routines.
6125 l2arc_read_done(zio_t *zio)
6127 l2arc_read_callback_t *cb;
6130 kmutex_t *hash_lock;
6133 ASSERT(zio->io_vd != NULL);
6134 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
6136 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
6138 cb = zio->io_private;
6140 buf = cb->l2rcb_buf;
6141 ASSERT(buf != NULL);
6143 hash_lock = HDR_LOCK(buf->b_hdr);
6144 mutex_enter(hash_lock);
6146 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6149 * If the data was read into a temporary buffer,
6150 * move it and free the buffer.
6152 if (cb->l2rcb_data != NULL) {
6153 ASSERT3U(hdr->b_size, <, zio->io_size);
6154 ASSERT3U(cb->l2rcb_compress, ==, ZIO_COMPRESS_OFF);
6155 if (zio->io_error == 0)
6156 bcopy(cb->l2rcb_data, buf->b_data, hdr->b_size);
6159 * The following must be done regardless of whether
6160 * there was an error:
6161 * - free the temporary buffer
6162 * - point zio to the real ARC buffer
6163 * - set zio size accordingly
6164 * These are required because zio is either re-used for
6165 * an I/O of the block in the case of the error
6166 * or the zio is passed to arc_read_done() and it
6169 zio_data_buf_free(cb->l2rcb_data, zio->io_size);
6170 zio->io_size = zio->io_orig_size = hdr->b_size;
6171 zio->io_data = zio->io_orig_data = buf->b_data;
6175 * If the buffer was compressed, decompress it first.
6177 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
6178 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
6179 ASSERT(zio->io_data != NULL);
6180 ASSERT3U(zio->io_size, ==, hdr->b_size);
6181 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size);
6184 * Check this survived the L2ARC journey.
6186 equal = arc_cksum_equal(buf);
6187 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
6188 mutex_exit(hash_lock);
6189 zio->io_private = buf;
6190 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
6191 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
6194 mutex_exit(hash_lock);
6196 * Buffer didn't survive caching. Increment stats and
6197 * reissue to the original storage device.
6199 if (zio->io_error != 0) {
6200 ARCSTAT_BUMP(arcstat_l2_io_error);
6202 zio->io_error = SET_ERROR(EIO);
6205 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
6208 * If there's no waiter, issue an async i/o to the primary
6209 * storage now. If there *is* a waiter, the caller must
6210 * issue the i/o in a context where it's OK to block.
6212 if (zio->io_waiter == NULL) {
6213 zio_t *pio = zio_unique_parent(zio);
6215 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
6217 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
6218 buf->b_data, hdr->b_size, arc_read_done, buf,
6219 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
6223 kmem_free(cb, sizeof (l2arc_read_callback_t));
6227 * This is the list priority from which the L2ARC will search for pages to
6228 * cache. This is used within loops (0..3) to cycle through lists in the
6229 * desired order. This order can have a significant effect on cache
6232 * Currently the metadata lists are hit first, MFU then MRU, followed by
6233 * the data lists. This function returns a locked list, and also returns
6236 static multilist_sublist_t *
6237 l2arc_sublist_lock(int list_num)
6239 multilist_t *ml = NULL;
6242 ASSERT(list_num >= 0 && list_num <= 3);
6246 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
6249 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
6252 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
6255 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
6260 * Return a randomly-selected sublist. This is acceptable
6261 * because the caller feeds only a little bit of data for each
6262 * call (8MB). Subsequent calls will result in different
6263 * sublists being selected.
6265 idx = multilist_get_random_index(ml);
6266 return (multilist_sublist_lock(ml, idx));
6270 * Evict buffers from the device write hand to the distance specified in
6271 * bytes. This distance may span populated buffers, it may span nothing.
6272 * This is clearing a region on the L2ARC device ready for writing.
6273 * If the 'all' boolean is set, every buffer is evicted.
6276 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
6279 arc_buf_hdr_t *hdr, *hdr_prev;
6280 kmutex_t *hash_lock;
6283 buflist = &dev->l2ad_buflist;
6285 if (!all && dev->l2ad_first) {
6287 * This is the first sweep through the device. There is
6293 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
6295 * When nearing the end of the device, evict to the end
6296 * before the device write hand jumps to the start.
6298 taddr = dev->l2ad_end;
6300 taddr = dev->l2ad_hand + distance;
6302 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
6303 uint64_t, taddr, boolean_t, all);
6306 mutex_enter(&dev->l2ad_mtx);
6307 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
6308 hdr_prev = list_prev(buflist, hdr);
6310 hash_lock = HDR_LOCK(hdr);
6313 * We cannot use mutex_enter or else we can deadlock
6314 * with l2arc_write_buffers (due to swapping the order
6315 * the hash lock and l2ad_mtx are taken).
6317 if (!mutex_tryenter(hash_lock)) {
6319 * Missed the hash lock. Retry.
6321 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
6322 mutex_exit(&dev->l2ad_mtx);
6323 mutex_enter(hash_lock);
6324 mutex_exit(hash_lock);
6328 if (HDR_L2_WRITE_HEAD(hdr)) {
6330 * We hit a write head node. Leave it for
6331 * l2arc_write_done().
6333 list_remove(buflist, hdr);
6334 mutex_exit(hash_lock);
6338 if (!all && HDR_HAS_L2HDR(hdr) &&
6339 (hdr->b_l2hdr.b_daddr > taddr ||
6340 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
6342 * We've evicted to the target address,
6343 * or the end of the device.
6345 mutex_exit(hash_lock);
6349 ASSERT(HDR_HAS_L2HDR(hdr));
6350 if (!HDR_HAS_L1HDR(hdr)) {
6351 ASSERT(!HDR_L2_READING(hdr));
6353 * This doesn't exist in the ARC. Destroy.
6354 * arc_hdr_destroy() will call list_remove()
6355 * and decrement arcstat_l2_size.
6357 arc_change_state(arc_anon, hdr, hash_lock);
6358 arc_hdr_destroy(hdr);
6360 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
6361 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
6363 * Invalidate issued or about to be issued
6364 * reads, since we may be about to write
6365 * over this location.
6367 if (HDR_L2_READING(hdr)) {
6368 ARCSTAT_BUMP(arcstat_l2_evict_reading);
6369 hdr->b_flags |= ARC_FLAG_L2_EVICTED;
6372 /* Ensure this header has finished being written */
6373 ASSERT(!HDR_L2_WRITING(hdr));
6374 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
6376 arc_hdr_l2hdr_destroy(hdr);
6378 mutex_exit(hash_lock);
6380 mutex_exit(&dev->l2ad_mtx);
6384 * Find and write ARC buffers to the L2ARC device.
6386 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6387 * for reading until they have completed writing.
6388 * The headroom_boost is an in-out parameter used to maintain headroom boost
6389 * state between calls to this function.
6391 * Returns the number of bytes actually written (which may be smaller than
6392 * the delta by which the device hand has changed due to alignment).
6395 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
6396 boolean_t *headroom_boost)
6398 arc_buf_hdr_t *hdr, *hdr_prev, *head;
6399 uint64_t write_asize, write_sz, headroom,
6403 l2arc_write_callback_t *cb;
6405 uint64_t guid = spa_load_guid(spa);
6406 const boolean_t do_headroom_boost = *headroom_boost;
6409 ASSERT(dev->l2ad_vdev != NULL);
6411 /* Lower the flag now, we might want to raise it again later. */
6412 *headroom_boost = B_FALSE;
6415 write_sz = write_asize = 0;
6417 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
6418 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD;
6419 head->b_flags |= ARC_FLAG_HAS_L2HDR;
6421 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
6423 * We will want to try to compress buffers that are at least 2x the
6424 * device sector size.
6426 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
6429 * Copy buffers for L2ARC writing.
6431 for (try = 0; try <= 3; try++) {
6432 multilist_sublist_t *mls = l2arc_sublist_lock(try);
6433 uint64_t passed_sz = 0;
6435 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
6438 * L2ARC fast warmup.
6440 * Until the ARC is warm and starts to evict, read from the
6441 * head of the ARC lists rather than the tail.
6443 if (arc_warm == B_FALSE)
6444 hdr = multilist_sublist_head(mls);
6446 hdr = multilist_sublist_tail(mls);
6448 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
6450 headroom = target_sz * l2arc_headroom;
6451 if (do_headroom_boost)
6452 headroom = (headroom * l2arc_headroom_boost) / 100;
6454 for (; hdr; hdr = hdr_prev) {
6455 kmutex_t *hash_lock;
6460 if (arc_warm == B_FALSE)
6461 hdr_prev = multilist_sublist_next(mls, hdr);
6463 hdr_prev = multilist_sublist_prev(mls, hdr);
6464 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size);
6466 hash_lock = HDR_LOCK(hdr);
6467 if (!mutex_tryenter(hash_lock)) {
6468 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
6470 * Skip this buffer rather than waiting.
6475 passed_sz += hdr->b_size;
6476 if (passed_sz > headroom) {
6480 mutex_exit(hash_lock);
6481 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
6485 if (!l2arc_write_eligible(guid, hdr)) {
6486 mutex_exit(hash_lock);
6491 * Assume that the buffer is not going to be compressed
6492 * and could take more space on disk because of a larger
6495 buf_sz = hdr->b_size;
6496 align = (size_t)1 << dev->l2ad_vdev->vdev_ashift;
6497 buf_a_sz = P2ROUNDUP(buf_sz, align);
6499 if ((write_asize + buf_a_sz) > target_sz) {
6501 mutex_exit(hash_lock);
6502 ARCSTAT_BUMP(arcstat_l2_write_full);
6508 * Insert a dummy header on the buflist so
6509 * l2arc_write_done() can find where the
6510 * write buffers begin without searching.
6512 mutex_enter(&dev->l2ad_mtx);
6513 list_insert_head(&dev->l2ad_buflist, head);
6514 mutex_exit(&dev->l2ad_mtx);
6517 sizeof (l2arc_write_callback_t), KM_SLEEP);
6518 cb->l2wcb_dev = dev;
6519 cb->l2wcb_head = head;
6520 pio = zio_root(spa, l2arc_write_done, cb,
6522 ARCSTAT_BUMP(arcstat_l2_write_pios);
6526 * Create and add a new L2ARC header.
6528 hdr->b_l2hdr.b_dev = dev;
6529 hdr->b_flags |= ARC_FLAG_L2_WRITING;
6531 * Temporarily stash the data buffer in b_tmp_cdata.
6532 * The subsequent write step will pick it up from
6533 * there. This is because can't access b_l1hdr.b_buf
6534 * without holding the hash_lock, which we in turn
6535 * can't access without holding the ARC list locks
6536 * (which we want to avoid during compression/writing).
6538 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF;
6539 hdr->b_l2hdr.b_asize = hdr->b_size;
6540 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data;
6543 * Explicitly set the b_daddr field to a known
6544 * value which means "invalid address". This
6545 * enables us to differentiate which stage of
6546 * l2arc_write_buffers() the particular header
6547 * is in (e.g. this loop, or the one below).
6548 * ARC_FLAG_L2_WRITING is not enough to make
6549 * this distinction, and we need to know in
6550 * order to do proper l2arc vdev accounting in
6551 * arc_release() and arc_hdr_destroy().
6553 * Note, we can't use a new flag to distinguish
6554 * the two stages because we don't hold the
6555 * header's hash_lock below, in the second stage
6556 * of this function. Thus, we can't simply
6557 * change the b_flags field to denote that the
6558 * IO has been sent. We can change the b_daddr
6559 * field of the L2 portion, though, since we'll
6560 * be holding the l2ad_mtx; which is why we're
6561 * using it to denote the header's state change.
6563 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET;
6565 hdr->b_flags |= ARC_FLAG_HAS_L2HDR;
6567 mutex_enter(&dev->l2ad_mtx);
6568 list_insert_head(&dev->l2ad_buflist, hdr);
6569 mutex_exit(&dev->l2ad_mtx);
6572 * Compute and store the buffer cksum before
6573 * writing. On debug the cksum is verified first.
6575 arc_cksum_verify(hdr->b_l1hdr.b_buf);
6576 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE);
6578 mutex_exit(hash_lock);
6581 write_asize += buf_a_sz;
6584 multilist_sublist_unlock(mls);
6590 /* No buffers selected for writing? */
6593 ASSERT(!HDR_HAS_L1HDR(head));
6594 kmem_cache_free(hdr_l2only_cache, head);
6598 mutex_enter(&dev->l2ad_mtx);
6601 * Now start writing the buffers. We're starting at the write head
6602 * and work backwards, retracing the course of the buffer selector
6606 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr;
6607 hdr = list_prev(&dev->l2ad_buflist, hdr)) {
6612 * We rely on the L1 portion of the header below, so
6613 * it's invalid for this header to have been evicted out
6614 * of the ghost cache, prior to being written out. The
6615 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6617 ASSERT(HDR_HAS_L1HDR(hdr));
6620 * We shouldn't need to lock the buffer here, since we flagged
6621 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6622 * take care to only access its L2 cache parameters. In
6623 * particular, hdr->l1hdr.b_buf may be invalid by now due to
6626 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
6629 * Save a pointer to the original buffer data we had previously
6632 buf_data = hdr->b_l1hdr.b_tmp_cdata;
6634 compress = HDR_L2COMPRESS(hdr) &&
6635 hdr->b_l2hdr.b_asize >= buf_compress_minsz;
6636 if (l2arc_transform_buf(hdr, compress)) {
6638 * If compression succeeded, enable headroom
6639 * boost on the next scan cycle.
6641 *headroom_boost = B_TRUE;
6645 * Get the new buffer size that accounts for compression
6648 buf_sz = hdr->b_l2hdr.b_asize;
6651 * We need to do this regardless if buf_sz is zero or
6652 * not, otherwise, when this l2hdr is evicted we'll
6653 * remove a reference that was never added.
6655 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr);
6657 /* Compression may have squashed the buffer to zero length. */
6660 * If the data was padded or compressed, then it
6661 * it is in a new buffer.
6663 if (hdr->b_l1hdr.b_tmp_cdata != NULL)
6664 buf_data = hdr->b_l1hdr.b_tmp_cdata;
6665 wzio = zio_write_phys(pio, dev->l2ad_vdev,
6666 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
6667 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
6668 ZIO_FLAG_CANFAIL, B_FALSE);
6670 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
6672 (void) zio_nowait(wzio);
6674 write_asize += buf_sz;
6675 dev->l2ad_hand += buf_sz;
6679 mutex_exit(&dev->l2ad_mtx);
6681 ASSERT3U(write_asize, <=, target_sz);
6682 ARCSTAT_BUMP(arcstat_l2_writes_sent);
6683 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
6684 ARCSTAT_INCR(arcstat_l2_size, write_sz);
6685 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
6686 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
6689 * Bump device hand to the device start if it is approaching the end.
6690 * l2arc_evict() will already have evicted ahead for this case.
6692 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
6693 dev->l2ad_hand = dev->l2ad_start;
6694 dev->l2ad_first = B_FALSE;
6697 dev->l2ad_writing = B_TRUE;
6698 (void) zio_wait(pio);
6699 dev->l2ad_writing = B_FALSE;
6701 return (write_asize);
6705 * Transforms, possibly compresses and pads, an L2ARC buffer.
6706 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6707 * size in l2hdr->b_asize. This routine tries to compress the data and
6708 * depending on the compression result there are three possible outcomes:
6709 * *) The buffer was incompressible. The buffer size was already ashift aligned.
6710 * The original hdr contents were left untouched except for b_tmp_cdata,
6711 * which is reset to NULL. The caller must keep a pointer to the original
6713 * *) The buffer was incompressible. The buffer size was not ashift aligned.
6714 * b_tmp_cdata was replaced with a temporary data buffer which holds a padded
6715 * (aligned) copy of the data. Once writing is done, invoke
6716 * l2arc_release_cdata_buf on this hdr to free the temporary buffer.
6717 * *) The buffer was all-zeros, so there is no need to write it to an L2
6718 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6719 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6720 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6721 * data buffer which holds the compressed data to be written, and b_asize
6722 * tells us how much data there is. b_compress is set to the appropriate
6723 * compression algorithm. Once writing is done, invoke
6724 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6726 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6727 * buffer was incompressible).
6730 l2arc_transform_buf(arc_buf_hdr_t *hdr, boolean_t compress)
6733 size_t align, asize, csize, len, rounded;
6735 ASSERT(HDR_HAS_L2HDR(hdr));
6736 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
6738 ASSERT(HDR_HAS_L1HDR(hdr));
6739 ASSERT3S(l2hdr->b_compress, ==, ZIO_COMPRESS_OFF);
6740 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6742 len = l2hdr->b_asize;
6743 align = (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift;
6744 asize = P2ROUNDUP(len, align);
6745 cdata = zio_data_buf_alloc(asize);
6746 ASSERT3P(cdata, !=, NULL);
6748 csize = zio_compress_data(ZIO_COMPRESS_LZ4,
6749 hdr->b_l1hdr.b_tmp_cdata, cdata, len);
6754 /* zero block, indicate that there's nothing to write */
6755 zio_data_buf_free(cdata, asize);
6756 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
6758 hdr->b_l1hdr.b_tmp_cdata = NULL;
6759 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
6763 rounded = P2ROUNDUP(csize, align);
6764 ASSERT3U(rounded, <=, asize);
6765 if (rounded < len) {
6767 * Compression succeeded, we'll keep the cdata around for
6768 * writing and release it afterwards.
6770 if (rounded > csize) {
6771 bzero((char *)cdata + csize, rounded - csize);
6774 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
6775 l2hdr->b_asize = csize;
6776 hdr->b_l1hdr.b_tmp_cdata = cdata;
6777 ARCSTAT_BUMP(arcstat_l2_compress_successes);
6781 * Compression did not save space.
6783 if (P2PHASE(len, align) != 0) {
6785 * Use compression buffer for a copy of data padded to
6786 * the proper size. Compression algorithm remains set
6787 * to ZIO_COMPRESS_OFF.
6789 ASSERT3U(len, <, asize);
6790 bcopy(hdr->b_l1hdr.b_tmp_cdata, cdata, len);
6791 bzero((char *)cdata + len, asize - len);
6792 l2hdr->b_asize = asize;
6793 hdr->b_l1hdr.b_tmp_cdata = cdata;
6794 ARCSTAT_BUMP(arcstat_l2_padding_needed);
6796 ASSERT3U(len, ==, asize);
6798 * The original buffer is good as is,
6799 * release the compressed buffer.
6800 * l2hdr will be left unmodified except for b_tmp_cdata.
6802 zio_data_buf_free(cdata, asize);
6803 hdr->b_l1hdr.b_tmp_cdata = NULL;
6806 ARCSTAT_BUMP(arcstat_l2_compress_failures);
6812 * Decompresses a zio read back from an l2arc device. On success, the
6813 * underlying zio's io_data buffer is overwritten by the uncompressed
6814 * version. On decompression error (corrupt compressed stream), the
6815 * zio->io_error value is set to signal an I/O error.
6817 * Please note that the compressed data stream is not checksummed, so
6818 * if the underlying device is experiencing data corruption, we may feed
6819 * corrupt data to the decompressor, so the decompressor needs to be
6820 * able to handle this situation (LZ4 does).
6823 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
6825 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
6827 if (zio->io_error != 0) {
6829 * An io error has occured, just restore the original io
6830 * size in preparation for a main pool read.
6832 zio->io_orig_size = zio->io_size = hdr->b_size;
6836 if (c == ZIO_COMPRESS_EMPTY) {
6838 * An empty buffer results in a null zio, which means we
6839 * need to fill its io_data after we're done restoring the
6840 * buffer's contents.
6842 ASSERT(hdr->b_l1hdr.b_buf != NULL);
6843 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size);
6844 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data;
6846 ASSERT(zio->io_data != NULL);
6848 * We copy the compressed data from the start of the arc buffer
6849 * (the zio_read will have pulled in only what we need, the
6850 * rest is garbage which we will overwrite at decompression)
6851 * and then decompress back to the ARC data buffer. This way we
6852 * can minimize copying by simply decompressing back over the
6853 * original compressed data (rather than decompressing to an
6854 * aux buffer and then copying back the uncompressed buffer,
6855 * which is likely to be much larger).
6860 csize = zio->io_size;
6861 cdata = zio_data_buf_alloc(csize);
6862 bcopy(zio->io_data, cdata, csize);
6863 if (zio_decompress_data(c, cdata, zio->io_data, csize,
6865 zio->io_error = EIO;
6866 zio_data_buf_free(cdata, csize);
6869 /* Restore the expected uncompressed IO size. */
6870 zio->io_orig_size = zio->io_size = hdr->b_size;
6874 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6875 * This buffer serves as a temporary holder of compressed or padded data while
6876 * the buffer entry is being written to an l2arc device. Once that is
6877 * done, we can dispose of it.
6880 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr)
6882 size_t align, asize, len;
6883 enum zio_compress comp = hdr->b_l2hdr.b_compress;
6885 ASSERT(HDR_HAS_L2HDR(hdr));
6886 ASSERT(HDR_HAS_L1HDR(hdr));
6887 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp));
6889 if (hdr->b_l1hdr.b_tmp_cdata != NULL) {
6890 ASSERT(comp != ZIO_COMPRESS_EMPTY);
6892 align = (size_t)1 << hdr->b_l2hdr.b_dev->l2ad_vdev->vdev_ashift;
6893 asize = P2ROUNDUP(len, align);
6894 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, asize);
6895 hdr->b_l1hdr.b_tmp_cdata = NULL;
6897 ASSERT(comp == ZIO_COMPRESS_OFF || comp == ZIO_COMPRESS_EMPTY);
6902 * This thread feeds the L2ARC at regular intervals. This is the beating
6903 * heart of the L2ARC.
6906 l2arc_feed_thread(void *dummy __unused)
6911 uint64_t size, wrote;
6912 clock_t begin, next = ddi_get_lbolt();
6913 boolean_t headroom_boost = B_FALSE;
6915 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
6917 mutex_enter(&l2arc_feed_thr_lock);
6919 while (l2arc_thread_exit == 0) {
6920 CALLB_CPR_SAFE_BEGIN(&cpr);
6921 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
6922 next - ddi_get_lbolt());
6923 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
6924 next = ddi_get_lbolt() + hz;
6927 * Quick check for L2ARC devices.
6929 mutex_enter(&l2arc_dev_mtx);
6930 if (l2arc_ndev == 0) {
6931 mutex_exit(&l2arc_dev_mtx);
6934 mutex_exit(&l2arc_dev_mtx);
6935 begin = ddi_get_lbolt();
6938 * This selects the next l2arc device to write to, and in
6939 * doing so the next spa to feed from: dev->l2ad_spa. This
6940 * will return NULL if there are now no l2arc devices or if
6941 * they are all faulted.
6943 * If a device is returned, its spa's config lock is also
6944 * held to prevent device removal. l2arc_dev_get_next()
6945 * will grab and release l2arc_dev_mtx.
6947 if ((dev = l2arc_dev_get_next()) == NULL)
6950 spa = dev->l2ad_spa;
6951 ASSERT(spa != NULL);
6954 * If the pool is read-only then force the feed thread to
6955 * sleep a little longer.
6957 if (!spa_writeable(spa)) {
6958 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
6959 spa_config_exit(spa, SCL_L2ARC, dev);
6964 * Avoid contributing to memory pressure.
6966 if (arc_reclaim_needed()) {
6967 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
6968 spa_config_exit(spa, SCL_L2ARC, dev);
6972 ARCSTAT_BUMP(arcstat_l2_feeds);
6974 size = l2arc_write_size();
6977 * Evict L2ARC buffers that will be overwritten.
6979 l2arc_evict(dev, size, B_FALSE);
6982 * Write ARC buffers.
6984 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
6987 * Calculate interval between writes.
6989 next = l2arc_write_interval(begin, size, wrote);
6990 spa_config_exit(spa, SCL_L2ARC, dev);
6993 l2arc_thread_exit = 0;
6994 cv_broadcast(&l2arc_feed_thr_cv);
6995 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7000 l2arc_vdev_present(vdev_t *vd)
7004 mutex_enter(&l2arc_dev_mtx);
7005 for (dev = list_head(l2arc_dev_list); dev != NULL;
7006 dev = list_next(l2arc_dev_list, dev)) {
7007 if (dev->l2ad_vdev == vd)
7010 mutex_exit(&l2arc_dev_mtx);
7012 return (dev != NULL);
7016 * Add a vdev for use by the L2ARC. By this point the spa has already
7017 * validated the vdev and opened it.
7020 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7022 l2arc_dev_t *adddev;
7024 ASSERT(!l2arc_vdev_present(vd));
7026 vdev_ashift_optimize(vd);
7029 * Create a new l2arc device entry.
7031 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7032 adddev->l2ad_spa = spa;
7033 adddev->l2ad_vdev = vd;
7034 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7035 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7036 adddev->l2ad_hand = adddev->l2ad_start;
7037 adddev->l2ad_first = B_TRUE;
7038 adddev->l2ad_writing = B_FALSE;
7040 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7042 * This is a list of all ARC buffers that are still valid on the
7045 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7046 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7048 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7049 refcount_create(&adddev->l2ad_alloc);
7052 * Add device to global list
7054 mutex_enter(&l2arc_dev_mtx);
7055 list_insert_head(l2arc_dev_list, adddev);
7056 atomic_inc_64(&l2arc_ndev);
7057 mutex_exit(&l2arc_dev_mtx);
7061 * Remove a vdev from the L2ARC.
7064 l2arc_remove_vdev(vdev_t *vd)
7066 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7069 * Find the device by vdev
7071 mutex_enter(&l2arc_dev_mtx);
7072 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7073 nextdev = list_next(l2arc_dev_list, dev);
7074 if (vd == dev->l2ad_vdev) {
7079 ASSERT(remdev != NULL);
7082 * Remove device from global list
7084 list_remove(l2arc_dev_list, remdev);
7085 l2arc_dev_last = NULL; /* may have been invalidated */
7086 atomic_dec_64(&l2arc_ndev);
7087 mutex_exit(&l2arc_dev_mtx);
7090 * Clear all buflists and ARC references. L2ARC device flush.
7092 l2arc_evict(remdev, 0, B_TRUE);
7093 list_destroy(&remdev->l2ad_buflist);
7094 mutex_destroy(&remdev->l2ad_mtx);
7095 refcount_destroy(&remdev->l2ad_alloc);
7096 kmem_free(remdev, sizeof (l2arc_dev_t));
7102 l2arc_thread_exit = 0;
7104 l2arc_writes_sent = 0;
7105 l2arc_writes_done = 0;
7107 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7108 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7109 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7110 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7112 l2arc_dev_list = &L2ARC_dev_list;
7113 l2arc_free_on_write = &L2ARC_free_on_write;
7114 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7115 offsetof(l2arc_dev_t, l2ad_node));
7116 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7117 offsetof(l2arc_data_free_t, l2df_list_node));
7124 * This is called from dmu_fini(), which is called from spa_fini();
7125 * Because of this, we can assume that all l2arc devices have
7126 * already been removed when the pools themselves were removed.
7129 l2arc_do_free_on_write();
7131 mutex_destroy(&l2arc_feed_thr_lock);
7132 cv_destroy(&l2arc_feed_thr_cv);
7133 mutex_destroy(&l2arc_dev_mtx);
7134 mutex_destroy(&l2arc_free_on_write_mtx);
7136 list_destroy(l2arc_dev_list);
7137 list_destroy(l2arc_free_on_write);
7143 if (!(spa_mode_global & FWRITE))
7146 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7147 TS_RUN, minclsyspri);
7153 if (!(spa_mode_global & FWRITE))
7156 mutex_enter(&l2arc_feed_thr_lock);
7157 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7158 l2arc_thread_exit = 1;
7159 while (l2arc_thread_exit != 0)
7160 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7161 mutex_exit(&l2arc_feed_thr_lock);