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 <vm/vm_pageout.h>
143 #include <machine/vmparam.h>
147 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
148 boolean_t arc_watch = B_FALSE;
153 static kmutex_t arc_reclaim_lock;
154 static kcondvar_t arc_reclaim_thread_cv;
155 static boolean_t arc_reclaim_thread_exit;
156 static kcondvar_t arc_reclaim_waiters_cv;
158 static kmutex_t arc_user_evicts_lock;
159 static kcondvar_t arc_user_evicts_cv;
160 static boolean_t arc_user_evicts_thread_exit;
162 uint_t arc_reduce_dnlc_percent = 3;
165 * The number of headers to evict in arc_evict_state_impl() before
166 * dropping the sublist lock and evicting from another sublist. A lower
167 * value means we're more likely to evict the "correct" header (i.e. the
168 * oldest header in the arc state), but comes with higher overhead
169 * (i.e. more invocations of arc_evict_state_impl()).
171 int zfs_arc_evict_batch_limit = 10;
174 * The number of sublists used for each of the arc state lists. If this
175 * is not set to a suitable value by the user, it will be configured to
176 * the number of CPUs on the system in arc_init().
178 int zfs_arc_num_sublists_per_state = 0;
180 /* number of seconds before growing cache again */
181 static int arc_grow_retry = 60;
183 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
184 int zfs_arc_overflow_shift = 8;
186 /* shift of arc_c for calculating both min and max arc_p */
187 static int arc_p_min_shift = 4;
189 /* log2(fraction of arc to reclaim) */
190 static int arc_shrink_shift = 7;
193 * log2(fraction of ARC which must be free to allow growing).
194 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
195 * when reading a new block into the ARC, we will evict an equal-sized block
198 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
199 * we will still not allow it to grow.
201 int arc_no_grow_shift = 5;
205 * minimum lifespan of a prefetch block in clock ticks
206 * (initialized in arc_init())
208 static int arc_min_prefetch_lifespan;
211 * If this percent of memory is free, don't throttle.
213 int arc_lotsfree_percent = 10;
216 extern boolean_t zfs_prefetch_disable;
219 * The arc has filled available memory and has now warmed up.
221 static boolean_t arc_warm;
224 * These tunables are for performance analysis.
226 uint64_t zfs_arc_max;
227 uint64_t zfs_arc_min;
228 uint64_t zfs_arc_meta_limit = 0;
229 uint64_t zfs_arc_meta_min = 0;
230 int zfs_arc_grow_retry = 0;
231 int zfs_arc_shrink_shift = 0;
232 int zfs_arc_p_min_shift = 0;
233 int zfs_disable_dup_eviction = 0;
234 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
235 u_int zfs_arc_free_target = 0;
237 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
238 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
242 arc_free_target_init(void *unused __unused)
245 zfs_arc_free_target = vm_pageout_wakeup_thresh;
247 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
248 arc_free_target_init, NULL);
250 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
251 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
252 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
253 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
254 TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize);
255 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
256 SYSCTL_DECL(_vfs_zfs);
257 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
259 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
261 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
262 &zfs_arc_average_blocksize, 0,
263 "ARC average blocksize");
264 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
265 &arc_shrink_shift, 0,
266 "log2(fraction of arc to reclaim)");
269 * We don't have a tunable for arc_free_target due to the dependency on
270 * pagedaemon initialisation.
272 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
273 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
274 sysctl_vfs_zfs_arc_free_target, "IU",
275 "Desired number of free pages below which ARC triggers reclaim");
278 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
283 val = zfs_arc_free_target;
284 err = sysctl_handle_int(oidp, &val, 0, req);
285 if (err != 0 || req->newptr == NULL)
290 if (val > cnt.v_page_count)
293 zfs_arc_free_target = val;
299 * Must be declared here, before the definition of corresponding kstat
300 * macro which uses the same names will confuse the compiler.
302 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
303 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
304 sysctl_vfs_zfs_arc_meta_limit, "QU",
305 "ARC metadata limit");
309 * Note that buffers can be in one of 6 states:
310 * ARC_anon - anonymous (discussed below)
311 * ARC_mru - recently used, currently cached
312 * ARC_mru_ghost - recentely used, no longer in cache
313 * ARC_mfu - frequently used, currently cached
314 * ARC_mfu_ghost - frequently used, no longer in cache
315 * ARC_l2c_only - exists in L2ARC but not other states
316 * When there are no active references to the buffer, they are
317 * are linked onto a list in one of these arc states. These are
318 * the only buffers that can be evicted or deleted. Within each
319 * state there are multiple lists, one for meta-data and one for
320 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
321 * etc.) is tracked separately so that it can be managed more
322 * explicitly: favored over data, limited explicitly.
324 * Anonymous buffers are buffers that are not associated with
325 * a DVA. These are buffers that hold dirty block copies
326 * before they are written to stable storage. By definition,
327 * they are "ref'd" and are considered part of arc_mru
328 * that cannot be freed. Generally, they will aquire a DVA
329 * as they are written and migrate onto the arc_mru list.
331 * The ARC_l2c_only state is for buffers that are in the second
332 * level ARC but no longer in any of the ARC_m* lists. The second
333 * level ARC itself may also contain buffers that are in any of
334 * the ARC_m* states - meaning that a buffer can exist in two
335 * places. The reason for the ARC_l2c_only state is to keep the
336 * buffer header in the hash table, so that reads that hit the
337 * second level ARC benefit from these fast lookups.
340 typedef struct arc_state {
342 * list of evictable buffers
344 multilist_t arcs_list[ARC_BUFC_NUMTYPES];
346 * total amount of evictable data in this state
348 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES];
350 * total amount of data in this state; this includes: evictable,
351 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
353 refcount_t arcs_size;
357 static arc_state_t ARC_anon;
358 static arc_state_t ARC_mru;
359 static arc_state_t ARC_mru_ghost;
360 static arc_state_t ARC_mfu;
361 static arc_state_t ARC_mfu_ghost;
362 static arc_state_t ARC_l2c_only;
364 typedef struct arc_stats {
365 kstat_named_t arcstat_hits;
366 kstat_named_t arcstat_misses;
367 kstat_named_t arcstat_demand_data_hits;
368 kstat_named_t arcstat_demand_data_misses;
369 kstat_named_t arcstat_demand_metadata_hits;
370 kstat_named_t arcstat_demand_metadata_misses;
371 kstat_named_t arcstat_prefetch_data_hits;
372 kstat_named_t arcstat_prefetch_data_misses;
373 kstat_named_t arcstat_prefetch_metadata_hits;
374 kstat_named_t arcstat_prefetch_metadata_misses;
375 kstat_named_t arcstat_mru_hits;
376 kstat_named_t arcstat_mru_ghost_hits;
377 kstat_named_t arcstat_mfu_hits;
378 kstat_named_t arcstat_mfu_ghost_hits;
379 kstat_named_t arcstat_allocated;
380 kstat_named_t arcstat_deleted;
382 * Number of buffers that could not be evicted because the hash lock
383 * was held by another thread. The lock may not necessarily be held
384 * by something using the same buffer, since hash locks are shared
385 * by multiple buffers.
387 kstat_named_t arcstat_mutex_miss;
389 * Number of buffers skipped because they have I/O in progress, are
390 * indrect prefetch buffers that have not lived long enough, or are
391 * not from the spa we're trying to evict from.
393 kstat_named_t arcstat_evict_skip;
395 * Number of times arc_evict_state() was unable to evict enough
396 * buffers to reach it's target amount.
398 kstat_named_t arcstat_evict_not_enough;
399 kstat_named_t arcstat_evict_l2_cached;
400 kstat_named_t arcstat_evict_l2_eligible;
401 kstat_named_t arcstat_evict_l2_ineligible;
402 kstat_named_t arcstat_evict_l2_skip;
403 kstat_named_t arcstat_hash_elements;
404 kstat_named_t arcstat_hash_elements_max;
405 kstat_named_t arcstat_hash_collisions;
406 kstat_named_t arcstat_hash_chains;
407 kstat_named_t arcstat_hash_chain_max;
408 kstat_named_t arcstat_p;
409 kstat_named_t arcstat_c;
410 kstat_named_t arcstat_c_min;
411 kstat_named_t arcstat_c_max;
412 kstat_named_t arcstat_size;
414 * Number of bytes consumed by internal ARC structures necessary
415 * for tracking purposes; these structures are not actually
416 * backed by ARC buffers. This includes arc_buf_hdr_t structures
417 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
418 * caches), and arc_buf_t structures (allocated via arc_buf_t
421 kstat_named_t arcstat_hdr_size;
423 * Number of bytes consumed by ARC buffers of type equal to
424 * ARC_BUFC_DATA. This is generally consumed by buffers backing
425 * on disk user data (e.g. plain file contents).
427 kstat_named_t arcstat_data_size;
429 * Number of bytes consumed by ARC buffers of type equal to
430 * ARC_BUFC_METADATA. This is generally consumed by buffers
431 * backing on disk data that is used for internal ZFS
432 * structures (e.g. ZAP, dnode, indirect blocks, etc).
434 kstat_named_t arcstat_metadata_size;
436 * Number of bytes consumed by various buffers and structures
437 * not actually backed with ARC buffers. This includes bonus
438 * buffers (allocated directly via zio_buf_* functions),
439 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
440 * cache), and dnode_t structures (allocated via dnode_t cache).
442 kstat_named_t arcstat_other_size;
444 * Total number of bytes consumed by ARC buffers residing in the
445 * arc_anon state. This includes *all* buffers in the arc_anon
446 * state; e.g. data, metadata, evictable, and unevictable buffers
447 * are all included in this value.
449 kstat_named_t arcstat_anon_size;
451 * Number of bytes consumed by ARC buffers that meet the
452 * following criteria: backing buffers of type ARC_BUFC_DATA,
453 * residing in the arc_anon state, and are eligible for eviction
454 * (e.g. have no outstanding holds on the buffer).
456 kstat_named_t arcstat_anon_evictable_data;
458 * Number of bytes consumed by ARC buffers that meet the
459 * following criteria: backing buffers of type ARC_BUFC_METADATA,
460 * residing in the arc_anon state, and are eligible for eviction
461 * (e.g. have no outstanding holds on the buffer).
463 kstat_named_t arcstat_anon_evictable_metadata;
465 * Total number of bytes consumed by ARC buffers residing in the
466 * arc_mru state. This includes *all* buffers in the arc_mru
467 * state; e.g. data, metadata, evictable, and unevictable buffers
468 * are all included in this value.
470 kstat_named_t arcstat_mru_size;
472 * Number of bytes consumed by ARC buffers that meet the
473 * following criteria: backing buffers of type ARC_BUFC_DATA,
474 * residing in the arc_mru state, and are eligible for eviction
475 * (e.g. have no outstanding holds on the buffer).
477 kstat_named_t arcstat_mru_evictable_data;
479 * Number of bytes consumed by ARC buffers that meet the
480 * following criteria: backing buffers of type ARC_BUFC_METADATA,
481 * residing in the arc_mru state, and are eligible for eviction
482 * (e.g. have no outstanding holds on the buffer).
484 kstat_named_t arcstat_mru_evictable_metadata;
486 * Total number of bytes that *would have been* consumed by ARC
487 * buffers in the arc_mru_ghost state. The key thing to note
488 * here, is the fact that this size doesn't actually indicate
489 * RAM consumption. The ghost lists only consist of headers and
490 * don't actually have ARC buffers linked off of these headers.
491 * Thus, *if* the headers had associated ARC buffers, these
492 * buffers *would have* consumed this number of bytes.
494 kstat_named_t arcstat_mru_ghost_size;
496 * Number of bytes that *would have been* consumed by ARC
497 * buffers that are eligible for eviction, of type
498 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
500 kstat_named_t arcstat_mru_ghost_evictable_data;
502 * Number of bytes that *would have been* consumed by ARC
503 * buffers that are eligible for eviction, of type
504 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
506 kstat_named_t arcstat_mru_ghost_evictable_metadata;
508 * Total number of bytes consumed by ARC buffers residing in the
509 * arc_mfu state. This includes *all* buffers in the arc_mfu
510 * state; e.g. data, metadata, evictable, and unevictable buffers
511 * are all included in this value.
513 kstat_named_t arcstat_mfu_size;
515 * Number of bytes consumed by ARC buffers that are eligible for
516 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
519 kstat_named_t arcstat_mfu_evictable_data;
521 * Number of bytes consumed by ARC buffers that are eligible for
522 * eviction, of type ARC_BUFC_METADATA, and reside in the
525 kstat_named_t arcstat_mfu_evictable_metadata;
527 * Total number of bytes that *would have been* consumed by ARC
528 * buffers in the arc_mfu_ghost state. See the comment above
529 * arcstat_mru_ghost_size for more details.
531 kstat_named_t arcstat_mfu_ghost_size;
533 * Number of bytes that *would have been* consumed by ARC
534 * buffers that are eligible for eviction, of type
535 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
537 kstat_named_t arcstat_mfu_ghost_evictable_data;
539 * Number of bytes that *would have been* consumed by ARC
540 * buffers that are eligible for eviction, of type
541 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
543 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
544 kstat_named_t arcstat_l2_hits;
545 kstat_named_t arcstat_l2_misses;
546 kstat_named_t arcstat_l2_feeds;
547 kstat_named_t arcstat_l2_rw_clash;
548 kstat_named_t arcstat_l2_read_bytes;
549 kstat_named_t arcstat_l2_write_bytes;
550 kstat_named_t arcstat_l2_writes_sent;
551 kstat_named_t arcstat_l2_writes_done;
552 kstat_named_t arcstat_l2_writes_error;
553 kstat_named_t arcstat_l2_writes_lock_retry;
554 kstat_named_t arcstat_l2_evict_lock_retry;
555 kstat_named_t arcstat_l2_evict_reading;
556 kstat_named_t arcstat_l2_evict_l1cached;
557 kstat_named_t arcstat_l2_free_on_write;
558 kstat_named_t arcstat_l2_cdata_free_on_write;
559 kstat_named_t arcstat_l2_abort_lowmem;
560 kstat_named_t arcstat_l2_cksum_bad;
561 kstat_named_t arcstat_l2_io_error;
562 kstat_named_t arcstat_l2_size;
563 kstat_named_t arcstat_l2_asize;
564 kstat_named_t arcstat_l2_hdr_size;
565 kstat_named_t arcstat_l2_compress_successes;
566 kstat_named_t arcstat_l2_compress_zeros;
567 kstat_named_t arcstat_l2_compress_failures;
568 kstat_named_t arcstat_l2_write_trylock_fail;
569 kstat_named_t arcstat_l2_write_passed_headroom;
570 kstat_named_t arcstat_l2_write_spa_mismatch;
571 kstat_named_t arcstat_l2_write_in_l2;
572 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
573 kstat_named_t arcstat_l2_write_not_cacheable;
574 kstat_named_t arcstat_l2_write_full;
575 kstat_named_t arcstat_l2_write_buffer_iter;
576 kstat_named_t arcstat_l2_write_pios;
577 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
578 kstat_named_t arcstat_l2_write_buffer_list_iter;
579 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
580 kstat_named_t arcstat_memory_throttle_count;
581 kstat_named_t arcstat_duplicate_buffers;
582 kstat_named_t arcstat_duplicate_buffers_size;
583 kstat_named_t arcstat_duplicate_reads;
584 kstat_named_t arcstat_meta_used;
585 kstat_named_t arcstat_meta_limit;
586 kstat_named_t arcstat_meta_max;
587 kstat_named_t arcstat_meta_min;
588 kstat_named_t arcstat_sync_wait_for_async;
589 kstat_named_t arcstat_demand_hit_predictive_prefetch;
592 static arc_stats_t arc_stats = {
593 { "hits", KSTAT_DATA_UINT64 },
594 { "misses", KSTAT_DATA_UINT64 },
595 { "demand_data_hits", KSTAT_DATA_UINT64 },
596 { "demand_data_misses", KSTAT_DATA_UINT64 },
597 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
598 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
599 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
600 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
601 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
602 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
603 { "mru_hits", KSTAT_DATA_UINT64 },
604 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
605 { "mfu_hits", KSTAT_DATA_UINT64 },
606 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
607 { "allocated", KSTAT_DATA_UINT64 },
608 { "deleted", KSTAT_DATA_UINT64 },
609 { "mutex_miss", KSTAT_DATA_UINT64 },
610 { "evict_skip", KSTAT_DATA_UINT64 },
611 { "evict_not_enough", KSTAT_DATA_UINT64 },
612 { "evict_l2_cached", KSTAT_DATA_UINT64 },
613 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
614 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
615 { "evict_l2_skip", KSTAT_DATA_UINT64 },
616 { "hash_elements", KSTAT_DATA_UINT64 },
617 { "hash_elements_max", KSTAT_DATA_UINT64 },
618 { "hash_collisions", KSTAT_DATA_UINT64 },
619 { "hash_chains", KSTAT_DATA_UINT64 },
620 { "hash_chain_max", KSTAT_DATA_UINT64 },
621 { "p", KSTAT_DATA_UINT64 },
622 { "c", KSTAT_DATA_UINT64 },
623 { "c_min", KSTAT_DATA_UINT64 },
624 { "c_max", KSTAT_DATA_UINT64 },
625 { "size", KSTAT_DATA_UINT64 },
626 { "hdr_size", KSTAT_DATA_UINT64 },
627 { "data_size", KSTAT_DATA_UINT64 },
628 { "metadata_size", KSTAT_DATA_UINT64 },
629 { "other_size", KSTAT_DATA_UINT64 },
630 { "anon_size", KSTAT_DATA_UINT64 },
631 { "anon_evictable_data", KSTAT_DATA_UINT64 },
632 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
633 { "mru_size", KSTAT_DATA_UINT64 },
634 { "mru_evictable_data", KSTAT_DATA_UINT64 },
635 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
636 { "mru_ghost_size", KSTAT_DATA_UINT64 },
637 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
638 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
639 { "mfu_size", KSTAT_DATA_UINT64 },
640 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
641 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
642 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
643 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
644 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
645 { "l2_hits", KSTAT_DATA_UINT64 },
646 { "l2_misses", KSTAT_DATA_UINT64 },
647 { "l2_feeds", KSTAT_DATA_UINT64 },
648 { "l2_rw_clash", KSTAT_DATA_UINT64 },
649 { "l2_read_bytes", KSTAT_DATA_UINT64 },
650 { "l2_write_bytes", KSTAT_DATA_UINT64 },
651 { "l2_writes_sent", KSTAT_DATA_UINT64 },
652 { "l2_writes_done", KSTAT_DATA_UINT64 },
653 { "l2_writes_error", KSTAT_DATA_UINT64 },
654 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
655 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
656 { "l2_evict_reading", KSTAT_DATA_UINT64 },
657 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
658 { "l2_free_on_write", KSTAT_DATA_UINT64 },
659 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 },
660 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
661 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
662 { "l2_io_error", KSTAT_DATA_UINT64 },
663 { "l2_size", KSTAT_DATA_UINT64 },
664 { "l2_asize", KSTAT_DATA_UINT64 },
665 { "l2_hdr_size", KSTAT_DATA_UINT64 },
666 { "l2_compress_successes", KSTAT_DATA_UINT64 },
667 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
668 { "l2_compress_failures", KSTAT_DATA_UINT64 },
669 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
670 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
671 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
672 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
673 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
674 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
675 { "l2_write_full", KSTAT_DATA_UINT64 },
676 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
677 { "l2_write_pios", KSTAT_DATA_UINT64 },
678 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
679 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
680 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
681 { "memory_throttle_count", KSTAT_DATA_UINT64 },
682 { "duplicate_buffers", KSTAT_DATA_UINT64 },
683 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
684 { "duplicate_reads", KSTAT_DATA_UINT64 },
685 { "arc_meta_used", KSTAT_DATA_UINT64 },
686 { "arc_meta_limit", KSTAT_DATA_UINT64 },
687 { "arc_meta_max", KSTAT_DATA_UINT64 },
688 { "arc_meta_min", KSTAT_DATA_UINT64 },
689 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
690 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
693 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
695 #define ARCSTAT_INCR(stat, val) \
696 atomic_add_64(&arc_stats.stat.value.ui64, (val))
698 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
699 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
701 #define ARCSTAT_MAX(stat, val) { \
703 while ((val) > (m = arc_stats.stat.value.ui64) && \
704 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
708 #define ARCSTAT_MAXSTAT(stat) \
709 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
712 * We define a macro to allow ARC hits/misses to be easily broken down by
713 * two separate conditions, giving a total of four different subtypes for
714 * each of hits and misses (so eight statistics total).
716 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
719 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
721 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
725 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
727 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
732 static arc_state_t *arc_anon;
733 static arc_state_t *arc_mru;
734 static arc_state_t *arc_mru_ghost;
735 static arc_state_t *arc_mfu;
736 static arc_state_t *arc_mfu_ghost;
737 static arc_state_t *arc_l2c_only;
740 * There are several ARC variables that are critical to export as kstats --
741 * but we don't want to have to grovel around in the kstat whenever we wish to
742 * manipulate them. For these variables, we therefore define them to be in
743 * terms of the statistic variable. This assures that we are not introducing
744 * the possibility of inconsistency by having shadow copies of the variables,
745 * while still allowing the code to be readable.
747 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
748 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
749 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
750 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
751 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
752 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
753 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
754 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
755 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
757 #define L2ARC_IS_VALID_COMPRESS(_c_) \
758 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
760 static int arc_no_grow; /* Don't try to grow cache size */
761 static uint64_t arc_tempreserve;
762 static uint64_t arc_loaned_bytes;
764 typedef struct arc_callback arc_callback_t;
766 struct arc_callback {
768 arc_done_func_t *acb_done;
770 zio_t *acb_zio_dummy;
771 arc_callback_t *acb_next;
774 typedef struct arc_write_callback arc_write_callback_t;
776 struct arc_write_callback {
778 arc_done_func_t *awcb_ready;
779 arc_done_func_t *awcb_physdone;
780 arc_done_func_t *awcb_done;
785 * ARC buffers are separated into multiple structs as a memory saving measure:
786 * - Common fields struct, always defined, and embedded within it:
787 * - L2-only fields, always allocated but undefined when not in L2ARC
788 * - L1-only fields, only allocated when in L1ARC
790 * Buffer in L1 Buffer only in L2
791 * +------------------------+ +------------------------+
792 * | arc_buf_hdr_t | | arc_buf_hdr_t |
796 * +------------------------+ +------------------------+
797 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
798 * | (undefined if L1-only) | | |
799 * +------------------------+ +------------------------+
800 * | l1arc_buf_hdr_t |
805 * +------------------------+
807 * Because it's possible for the L2ARC to become extremely large, we can wind
808 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
809 * is minimized by only allocating the fields necessary for an L1-cached buffer
810 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
811 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
812 * words in pointers. arc_hdr_realloc() is used to switch a header between
813 * these two allocation states.
815 typedef struct l1arc_buf_hdr {
816 kmutex_t b_freeze_lock;
819 * used for debugging wtih kmem_flags - by allocating and freeing
820 * b_thawed when the buffer is thawed, we get a record of the stack
821 * trace that thawed it.
828 /* for waiting on writes to complete */
831 /* protected by arc state mutex */
832 arc_state_t *b_state;
833 multilist_node_t b_arc_node;
835 /* updated atomically */
836 clock_t b_arc_access;
838 /* self protecting */
841 arc_callback_t *b_acb;
842 /* temporary buffer holder for in-flight compressed data */
846 typedef struct l2arc_dev l2arc_dev_t;
848 typedef struct l2arc_buf_hdr {
849 /* protected by arc_buf_hdr mutex */
850 l2arc_dev_t *b_dev; /* L2ARC device */
851 uint64_t b_daddr; /* disk address, offset byte */
852 /* real alloc'd buffer size depending on b_compress applied */
856 list_node_t b_l2node;
860 /* protected by hash lock */
864 * Even though this checksum is only set/verified when a buffer is in
865 * the L1 cache, it needs to be in the set of common fields because it
866 * must be preserved from the time before a buffer is written out to
867 * L2ARC until after it is read back in.
869 zio_cksum_t *b_freeze_cksum;
871 arc_buf_hdr_t *b_hash_next;
878 /* L2ARC fields. Undefined when not in L2ARC. */
879 l2arc_buf_hdr_t b_l2hdr;
880 /* L1ARC fields. Undefined when in l2arc_only state */
881 l1arc_buf_hdr_t b_l1hdr;
886 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
891 val = arc_meta_limit;
892 err = sysctl_handle_64(oidp, &val, 0, req);
893 if (err != 0 || req->newptr == NULL)
896 if (val <= 0 || val > arc_c_max)
899 arc_meta_limit = val;
904 static arc_buf_t *arc_eviction_list;
905 static arc_buf_hdr_t arc_eviction_hdr;
907 #define GHOST_STATE(state) \
908 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
909 (state) == arc_l2c_only)
911 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
912 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
913 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
914 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
915 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
916 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
918 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
919 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
920 #define HDR_L2_READING(hdr) \
921 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
922 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
923 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
924 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
925 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
927 #define HDR_ISTYPE_METADATA(hdr) \
928 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
929 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
931 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
932 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
938 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
939 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
942 * Hash table routines
945 #define HT_LOCK_PAD CACHE_LINE_SIZE
950 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
954 #define BUF_LOCKS 256
955 typedef struct buf_hash_table {
957 arc_buf_hdr_t **ht_table;
958 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
961 static buf_hash_table_t buf_hash_table;
963 #define BUF_HASH_INDEX(spa, dva, birth) \
964 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
965 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
966 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
967 #define HDR_LOCK(hdr) \
968 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
970 uint64_t zfs_crc64_table[256];
976 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
977 #define L2ARC_HEADROOM 2 /* num of writes */
979 * If we discover during ARC scan any buffers to be compressed, we boost
980 * our headroom for the next scanning cycle by this percentage multiple.
982 #define L2ARC_HEADROOM_BOOST 200
983 #define L2ARC_FEED_SECS 1 /* caching interval secs */
984 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
987 * Used to distinguish headers that are being process by
988 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
989 * address. This can happen when the header is added to the l2arc's list
990 * of buffers to write in the first stage of l2arc_write_buffers(), but
991 * has not yet been written out which happens in the second stage of
992 * l2arc_write_buffers().
994 #define L2ARC_ADDR_UNSET ((uint64_t)(-1))
996 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
997 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
999 /* L2ARC Performance Tunables */
1000 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1001 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1002 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1003 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1004 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1005 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1006 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1007 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1008 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1010 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1011 &l2arc_write_max, 0, "max write size");
1012 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1013 &l2arc_write_boost, 0, "extra write during warmup");
1014 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1015 &l2arc_headroom, 0, "number of dev writes");
1016 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1017 &l2arc_feed_secs, 0, "interval seconds");
1018 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1019 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1021 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1022 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1023 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1024 &l2arc_feed_again, 0, "turbo warmup");
1025 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1026 &l2arc_norw, 0, "no reads during writes");
1028 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1029 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1030 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
1031 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
1032 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
1033 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
1035 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1036 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1037 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
1038 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
1039 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
1040 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
1042 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1043 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1044 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
1045 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
1046 "size of metadata in mru ghost state");
1047 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
1048 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
1049 "size of data in mru ghost state");
1051 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1052 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1053 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
1054 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
1055 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
1056 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
1058 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1059 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1060 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
1061 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
1062 "size of metadata in mfu ghost state");
1063 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
1064 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
1065 "size of data in mfu ghost state");
1067 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1068 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1074 vdev_t *l2ad_vdev; /* vdev */
1075 spa_t *l2ad_spa; /* spa */
1076 uint64_t l2ad_hand; /* next write location */
1077 uint64_t l2ad_start; /* first addr on device */
1078 uint64_t l2ad_end; /* last addr on device */
1079 boolean_t l2ad_first; /* first sweep through */
1080 boolean_t l2ad_writing; /* currently writing */
1081 kmutex_t l2ad_mtx; /* lock for buffer list */
1082 list_t l2ad_buflist; /* buffer list */
1083 list_node_t l2ad_node; /* device list node */
1084 refcount_t l2ad_alloc; /* allocated bytes */
1087 static list_t L2ARC_dev_list; /* device list */
1088 static list_t *l2arc_dev_list; /* device list pointer */
1089 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1090 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1091 static list_t L2ARC_free_on_write; /* free after write buf list */
1092 static list_t *l2arc_free_on_write; /* free after write list ptr */
1093 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1094 static uint64_t l2arc_ndev; /* number of devices */
1096 typedef struct l2arc_read_callback {
1097 arc_buf_t *l2rcb_buf; /* read buffer */
1098 spa_t *l2rcb_spa; /* spa */
1099 blkptr_t l2rcb_bp; /* original blkptr */
1100 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1101 int l2rcb_flags; /* original flags */
1102 enum zio_compress l2rcb_compress; /* applied compress */
1103 } l2arc_read_callback_t;
1105 typedef struct l2arc_write_callback {
1106 l2arc_dev_t *l2wcb_dev; /* device info */
1107 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1108 } l2arc_write_callback_t;
1110 typedef struct l2arc_data_free {
1111 /* protected by l2arc_free_on_write_mtx */
1114 void (*l2df_func)(void *, size_t);
1115 list_node_t l2df_list_node;
1116 } l2arc_data_free_t;
1118 static kmutex_t l2arc_feed_thr_lock;
1119 static kcondvar_t l2arc_feed_thr_cv;
1120 static uint8_t l2arc_thread_exit;
1122 static void arc_get_data_buf(arc_buf_t *);
1123 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1124 static boolean_t arc_is_overflowing();
1125 static void arc_buf_watch(arc_buf_t *);
1127 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1128 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1130 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1131 static void l2arc_read_done(zio_t *);
1133 static boolean_t l2arc_compress_buf(arc_buf_hdr_t *);
1134 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress);
1135 static void l2arc_release_cdata_buf(arc_buf_hdr_t *);
1138 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1140 uint8_t *vdva = (uint8_t *)dva;
1141 uint64_t crc = -1ULL;
1144 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1146 for (i = 0; i < sizeof (dva_t); i++)
1147 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1149 crc ^= (spa>>8) ^ birth;
1154 #define BUF_EMPTY(buf) \
1155 ((buf)->b_dva.dva_word[0] == 0 && \
1156 (buf)->b_dva.dva_word[1] == 0)
1158 #define BUF_EQUAL(spa, dva, birth, buf) \
1159 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1160 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1161 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
1164 buf_discard_identity(arc_buf_hdr_t *hdr)
1166 hdr->b_dva.dva_word[0] = 0;
1167 hdr->b_dva.dva_word[1] = 0;
1171 static arc_buf_hdr_t *
1172 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1174 const dva_t *dva = BP_IDENTITY(bp);
1175 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1176 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1177 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1180 mutex_enter(hash_lock);
1181 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1182 hdr = hdr->b_hash_next) {
1183 if (BUF_EQUAL(spa, dva, birth, hdr)) {
1188 mutex_exit(hash_lock);
1194 * Insert an entry into the hash table. If there is already an element
1195 * equal to elem in the hash table, then the already existing element
1196 * will be returned and the new element will not be inserted.
1197 * Otherwise returns NULL.
1198 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1200 static arc_buf_hdr_t *
1201 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1203 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1204 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1205 arc_buf_hdr_t *fhdr;
1208 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1209 ASSERT(hdr->b_birth != 0);
1210 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1212 if (lockp != NULL) {
1214 mutex_enter(hash_lock);
1216 ASSERT(MUTEX_HELD(hash_lock));
1219 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1220 fhdr = fhdr->b_hash_next, i++) {
1221 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1225 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1226 buf_hash_table.ht_table[idx] = hdr;
1227 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
1229 /* collect some hash table performance data */
1231 ARCSTAT_BUMP(arcstat_hash_collisions);
1233 ARCSTAT_BUMP(arcstat_hash_chains);
1235 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1238 ARCSTAT_BUMP(arcstat_hash_elements);
1239 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1245 buf_hash_remove(arc_buf_hdr_t *hdr)
1247 arc_buf_hdr_t *fhdr, **hdrp;
1248 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1250 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1251 ASSERT(HDR_IN_HASH_TABLE(hdr));
1253 hdrp = &buf_hash_table.ht_table[idx];
1254 while ((fhdr = *hdrp) != hdr) {
1255 ASSERT(fhdr != NULL);
1256 hdrp = &fhdr->b_hash_next;
1258 *hdrp = hdr->b_hash_next;
1259 hdr->b_hash_next = NULL;
1260 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE;
1262 /* collect some hash table performance data */
1263 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1265 if (buf_hash_table.ht_table[idx] &&
1266 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1267 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1271 * Global data structures and functions for the buf kmem cache.
1273 static kmem_cache_t *hdr_full_cache;
1274 static kmem_cache_t *hdr_l2only_cache;
1275 static kmem_cache_t *buf_cache;
1282 kmem_free(buf_hash_table.ht_table,
1283 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1284 for (i = 0; i < BUF_LOCKS; i++)
1285 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1286 kmem_cache_destroy(hdr_full_cache);
1287 kmem_cache_destroy(hdr_l2only_cache);
1288 kmem_cache_destroy(buf_cache);
1292 * Constructor callback - called when the cache is empty
1293 * and a new buf is requested.
1297 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1299 arc_buf_hdr_t *hdr = vbuf;
1301 bzero(hdr, HDR_FULL_SIZE);
1302 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1303 refcount_create(&hdr->b_l1hdr.b_refcnt);
1304 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1305 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1306 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1313 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1315 arc_buf_hdr_t *hdr = vbuf;
1317 bzero(hdr, HDR_L2ONLY_SIZE);
1318 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1325 buf_cons(void *vbuf, void *unused, int kmflag)
1327 arc_buf_t *buf = vbuf;
1329 bzero(buf, sizeof (arc_buf_t));
1330 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1331 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1337 * Destructor callback - called when a cached buf is
1338 * no longer required.
1342 hdr_full_dest(void *vbuf, void *unused)
1344 arc_buf_hdr_t *hdr = vbuf;
1346 ASSERT(BUF_EMPTY(hdr));
1347 cv_destroy(&hdr->b_l1hdr.b_cv);
1348 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1349 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1350 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1351 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1356 hdr_l2only_dest(void *vbuf, void *unused)
1358 arc_buf_hdr_t *hdr = vbuf;
1360 ASSERT(BUF_EMPTY(hdr));
1361 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1366 buf_dest(void *vbuf, void *unused)
1368 arc_buf_t *buf = vbuf;
1370 mutex_destroy(&buf->b_evict_lock);
1371 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1375 * Reclaim callback -- invoked when memory is low.
1379 hdr_recl(void *unused)
1381 dprintf("hdr_recl called\n");
1383 * umem calls the reclaim func when we destroy the buf cache,
1384 * which is after we do arc_fini().
1387 cv_signal(&arc_reclaim_thread_cv);
1394 uint64_t hsize = 1ULL << 12;
1398 * The hash table is big enough to fill all of physical memory
1399 * with an average block size of zfs_arc_average_blocksize (default 8K).
1400 * By default, the table will take up
1401 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1403 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1406 buf_hash_table.ht_mask = hsize - 1;
1407 buf_hash_table.ht_table =
1408 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1409 if (buf_hash_table.ht_table == NULL) {
1410 ASSERT(hsize > (1ULL << 8));
1415 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1416 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1417 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1418 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1420 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1421 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1423 for (i = 0; i < 256; i++)
1424 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1425 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1427 for (i = 0; i < BUF_LOCKS; i++) {
1428 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1429 NULL, MUTEX_DEFAULT, NULL);
1434 * Transition between the two allocation states for the arc_buf_hdr struct.
1435 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1436 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1437 * version is used when a cache buffer is only in the L2ARC in order to reduce
1440 static arc_buf_hdr_t *
1441 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
1443 ASSERT(HDR_HAS_L2HDR(hdr));
1445 arc_buf_hdr_t *nhdr;
1446 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1448 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
1449 (old == hdr_l2only_cache && new == hdr_full_cache));
1451 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
1453 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
1454 buf_hash_remove(hdr);
1456 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
1458 if (new == hdr_full_cache) {
1459 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1461 * arc_access and arc_change_state need to be aware that a
1462 * header has just come out of L2ARC, so we set its state to
1463 * l2c_only even though it's about to change.
1465 nhdr->b_l1hdr.b_state = arc_l2c_only;
1467 /* Verify previous threads set to NULL before freeing */
1468 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1470 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1471 ASSERT0(hdr->b_l1hdr.b_datacnt);
1474 * If we've reached here, We must have been called from
1475 * arc_evict_hdr(), as such we should have already been
1476 * removed from any ghost list we were previously on
1477 * (which protects us from racing with arc_evict_state),
1478 * thus no locking is needed during this check.
1480 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1483 * A buffer must not be moved into the arc_l2c_only
1484 * state if it's not finished being written out to the
1485 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1486 * might try to be accessed, even though it was removed.
1488 VERIFY(!HDR_L2_WRITING(hdr));
1489 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1492 if (hdr->b_l1hdr.b_thawed != NULL) {
1493 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1494 hdr->b_l1hdr.b_thawed = NULL;
1498 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR;
1501 * The header has been reallocated so we need to re-insert it into any
1504 (void) buf_hash_insert(nhdr, NULL);
1506 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
1508 mutex_enter(&dev->l2ad_mtx);
1511 * We must place the realloc'ed header back into the list at
1512 * the same spot. Otherwise, if it's placed earlier in the list,
1513 * l2arc_write_buffers() could find it during the function's
1514 * write phase, and try to write it out to the l2arc.
1516 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
1517 list_remove(&dev->l2ad_buflist, hdr);
1519 mutex_exit(&dev->l2ad_mtx);
1522 * Since we're using the pointer address as the tag when
1523 * incrementing and decrementing the l2ad_alloc refcount, we
1524 * must remove the old pointer (that we're about to destroy) and
1525 * add the new pointer to the refcount. Otherwise we'd remove
1526 * the wrong pointer address when calling arc_hdr_destroy() later.
1529 (void) refcount_remove_many(&dev->l2ad_alloc,
1530 hdr->b_l2hdr.b_asize, hdr);
1532 (void) refcount_add_many(&dev->l2ad_alloc,
1533 nhdr->b_l2hdr.b_asize, nhdr);
1535 buf_discard_identity(hdr);
1536 hdr->b_freeze_cksum = NULL;
1537 kmem_cache_free(old, hdr);
1543 #define ARC_MINTIME (hz>>4) /* 62 ms */
1546 arc_cksum_verify(arc_buf_t *buf)
1550 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1553 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1554 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) {
1555 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1558 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc);
1559 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1560 panic("buffer modified while frozen!");
1561 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1565 arc_cksum_equal(arc_buf_t *buf)
1570 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1571 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc);
1572 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1573 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1579 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1581 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1584 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1585 if (buf->b_hdr->b_freeze_cksum != NULL) {
1586 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1589 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1590 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1591 NULL, buf->b_hdr->b_freeze_cksum);
1592 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1595 #endif /* illumos */
1600 typedef struct procctl {
1608 arc_buf_unwatch(arc_buf_t *buf)
1615 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1616 ctl.prwatch.pr_size = 0;
1617 ctl.prwatch.pr_wflags = 0;
1618 result = write(arc_procfd, &ctl, sizeof (ctl));
1619 ASSERT3U(result, ==, sizeof (ctl));
1626 arc_buf_watch(arc_buf_t *buf)
1633 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1634 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1635 ctl.prwatch.pr_wflags = WA_WRITE;
1636 result = write(arc_procfd, &ctl, sizeof (ctl));
1637 ASSERT3U(result, ==, sizeof (ctl));
1641 #endif /* illumos */
1643 static arc_buf_contents_t
1644 arc_buf_type(arc_buf_hdr_t *hdr)
1646 if (HDR_ISTYPE_METADATA(hdr)) {
1647 return (ARC_BUFC_METADATA);
1649 return (ARC_BUFC_DATA);
1654 arc_bufc_to_flags(arc_buf_contents_t type)
1658 /* metadata field is 0 if buffer contains normal data */
1660 case ARC_BUFC_METADATA:
1661 return (ARC_FLAG_BUFC_METADATA);
1665 panic("undefined ARC buffer type!");
1666 return ((uint32_t)-1);
1670 arc_buf_thaw(arc_buf_t *buf)
1672 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1673 if (buf->b_hdr->b_l1hdr.b_state != arc_anon)
1674 panic("modifying non-anon buffer!");
1675 if (HDR_IO_IN_PROGRESS(buf->b_hdr))
1676 panic("modifying buffer while i/o in progress!");
1677 arc_cksum_verify(buf);
1680 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1681 if (buf->b_hdr->b_freeze_cksum != NULL) {
1682 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1683 buf->b_hdr->b_freeze_cksum = NULL;
1687 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1688 if (buf->b_hdr->b_l1hdr.b_thawed != NULL)
1689 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1);
1690 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1694 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1697 arc_buf_unwatch(buf);
1698 #endif /* illumos */
1702 arc_buf_freeze(arc_buf_t *buf)
1704 kmutex_t *hash_lock;
1706 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1709 hash_lock = HDR_LOCK(buf->b_hdr);
1710 mutex_enter(hash_lock);
1712 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1713 buf->b_hdr->b_l1hdr.b_state == arc_anon);
1714 arc_cksum_compute(buf, B_FALSE);
1715 mutex_exit(hash_lock);
1720 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1722 ASSERT(HDR_HAS_L1HDR(hdr));
1723 ASSERT(MUTEX_HELD(hash_lock));
1724 arc_state_t *state = hdr->b_l1hdr.b_state;
1726 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
1727 (state != arc_anon)) {
1728 /* We don't use the L2-only state list. */
1729 if (state != arc_l2c_only) {
1730 arc_buf_contents_t type = arc_buf_type(hdr);
1731 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt;
1732 multilist_t *list = &state->arcs_list[type];
1733 uint64_t *size = &state->arcs_lsize[type];
1735 multilist_remove(list, hdr);
1737 if (GHOST_STATE(state)) {
1738 ASSERT0(hdr->b_l1hdr.b_datacnt);
1739 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
1740 delta = hdr->b_size;
1743 ASSERT3U(*size, >=, delta);
1744 atomic_add_64(size, -delta);
1746 /* remove the prefetch flag if we get a reference */
1747 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
1752 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1755 arc_state_t *state = hdr->b_l1hdr.b_state;
1757 ASSERT(HDR_HAS_L1HDR(hdr));
1758 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1759 ASSERT(!GHOST_STATE(state));
1762 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1763 * check to prevent usage of the arc_l2c_only list.
1765 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
1766 (state != arc_anon)) {
1767 arc_buf_contents_t type = arc_buf_type(hdr);
1768 multilist_t *list = &state->arcs_list[type];
1769 uint64_t *size = &state->arcs_lsize[type];
1771 multilist_insert(list, hdr);
1773 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
1774 atomic_add_64(size, hdr->b_size *
1775 hdr->b_l1hdr.b_datacnt);
1781 * Move the supplied buffer to the indicated state. The hash lock
1782 * for the buffer must be held by the caller.
1785 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
1786 kmutex_t *hash_lock)
1788 arc_state_t *old_state;
1791 uint64_t from_delta, to_delta;
1792 arc_buf_contents_t buftype = arc_buf_type(hdr);
1795 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1796 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1797 * L1 hdr doesn't always exist when we change state to arc_anon before
1798 * destroying a header, in which case reallocating to add the L1 hdr is
1801 if (HDR_HAS_L1HDR(hdr)) {
1802 old_state = hdr->b_l1hdr.b_state;
1803 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
1804 datacnt = hdr->b_l1hdr.b_datacnt;
1806 old_state = arc_l2c_only;
1811 ASSERT(MUTEX_HELD(hash_lock));
1812 ASSERT3P(new_state, !=, old_state);
1813 ASSERT(refcnt == 0 || datacnt > 0);
1814 ASSERT(!GHOST_STATE(new_state) || datacnt == 0);
1815 ASSERT(old_state != arc_anon || datacnt <= 1);
1817 from_delta = to_delta = datacnt * hdr->b_size;
1820 * If this buffer is evictable, transfer it from the
1821 * old state list to the new state list.
1824 if (old_state != arc_anon && old_state != arc_l2c_only) {
1825 uint64_t *size = &old_state->arcs_lsize[buftype];
1827 ASSERT(HDR_HAS_L1HDR(hdr));
1828 multilist_remove(&old_state->arcs_list[buftype], hdr);
1831 * If prefetching out of the ghost cache,
1832 * we will have a non-zero datacnt.
1834 if (GHOST_STATE(old_state) && datacnt == 0) {
1835 /* ghost elements have a ghost size */
1836 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1837 from_delta = hdr->b_size;
1839 ASSERT3U(*size, >=, from_delta);
1840 atomic_add_64(size, -from_delta);
1842 if (new_state != arc_anon && new_state != arc_l2c_only) {
1843 uint64_t *size = &new_state->arcs_lsize[buftype];
1846 * An L1 header always exists here, since if we're
1847 * moving to some L1-cached state (i.e. not l2c_only or
1848 * anonymous), we realloc the header to add an L1hdr
1851 ASSERT(HDR_HAS_L1HDR(hdr));
1852 multilist_insert(&new_state->arcs_list[buftype], hdr);
1854 /* ghost elements have a ghost size */
1855 if (GHOST_STATE(new_state)) {
1857 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1858 to_delta = hdr->b_size;
1860 atomic_add_64(size, to_delta);
1864 ASSERT(!BUF_EMPTY(hdr));
1865 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
1866 buf_hash_remove(hdr);
1868 /* adjust state sizes (ignore arc_l2c_only) */
1870 if (to_delta && new_state != arc_l2c_only) {
1871 ASSERT(HDR_HAS_L1HDR(hdr));
1872 if (GHOST_STATE(new_state)) {
1876 * We moving a header to a ghost state, we first
1877 * remove all arc buffers. Thus, we'll have a
1878 * datacnt of zero, and no arc buffer to use for
1879 * the reference. As a result, we use the arc
1880 * header pointer for the reference.
1882 (void) refcount_add_many(&new_state->arcs_size,
1885 ASSERT3U(datacnt, !=, 0);
1888 * Each individual buffer holds a unique reference,
1889 * thus we must remove each of these references one
1892 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1893 buf = buf->b_next) {
1894 (void) refcount_add_many(&new_state->arcs_size,
1900 if (from_delta && old_state != arc_l2c_only) {
1901 ASSERT(HDR_HAS_L1HDR(hdr));
1902 if (GHOST_STATE(old_state)) {
1904 * When moving a header off of a ghost state,
1905 * there's the possibility for datacnt to be
1906 * non-zero. This is because we first add the
1907 * arc buffer to the header prior to changing
1908 * the header's state. Since we used the header
1909 * for the reference when putting the header on
1910 * the ghost state, we must balance that and use
1911 * the header when removing off the ghost state
1912 * (even though datacnt is non zero).
1915 IMPLY(datacnt == 0, new_state == arc_anon ||
1916 new_state == arc_l2c_only);
1918 (void) refcount_remove_many(&old_state->arcs_size,
1921 ASSERT3P(datacnt, !=, 0);
1924 * Each individual buffer holds a unique reference,
1925 * thus we must remove each of these references one
1928 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1929 buf = buf->b_next) {
1930 (void) refcount_remove_many(
1931 &old_state->arcs_size, hdr->b_size, buf);
1936 if (HDR_HAS_L1HDR(hdr))
1937 hdr->b_l1hdr.b_state = new_state;
1940 * L2 headers should never be on the L2 state list since they don't
1941 * have L1 headers allocated.
1943 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
1944 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
1948 arc_space_consume(uint64_t space, arc_space_type_t type)
1950 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1953 case ARC_SPACE_DATA:
1954 ARCSTAT_INCR(arcstat_data_size, space);
1956 case ARC_SPACE_META:
1957 ARCSTAT_INCR(arcstat_metadata_size, space);
1959 case ARC_SPACE_OTHER:
1960 ARCSTAT_INCR(arcstat_other_size, space);
1962 case ARC_SPACE_HDRS:
1963 ARCSTAT_INCR(arcstat_hdr_size, space);
1965 case ARC_SPACE_L2HDRS:
1966 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1970 if (type != ARC_SPACE_DATA)
1971 ARCSTAT_INCR(arcstat_meta_used, space);
1973 atomic_add_64(&arc_size, space);
1977 arc_space_return(uint64_t space, arc_space_type_t type)
1979 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1982 case ARC_SPACE_DATA:
1983 ARCSTAT_INCR(arcstat_data_size, -space);
1985 case ARC_SPACE_META:
1986 ARCSTAT_INCR(arcstat_metadata_size, -space);
1988 case ARC_SPACE_OTHER:
1989 ARCSTAT_INCR(arcstat_other_size, -space);
1991 case ARC_SPACE_HDRS:
1992 ARCSTAT_INCR(arcstat_hdr_size, -space);
1994 case ARC_SPACE_L2HDRS:
1995 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1999 if (type != ARC_SPACE_DATA) {
2000 ASSERT(arc_meta_used >= space);
2001 if (arc_meta_max < arc_meta_used)
2002 arc_meta_max = arc_meta_used;
2003 ARCSTAT_INCR(arcstat_meta_used, -space);
2006 ASSERT(arc_size >= space);
2007 atomic_add_64(&arc_size, -space);
2011 arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
2016 ASSERT3U(size, >, 0);
2017 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
2018 ASSERT(BUF_EMPTY(hdr));
2019 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
2021 hdr->b_spa = spa_load_guid(spa);
2023 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2026 buf->b_efunc = NULL;
2027 buf->b_private = NULL;
2030 hdr->b_flags = arc_bufc_to_flags(type);
2031 hdr->b_flags |= ARC_FLAG_HAS_L1HDR;
2033 hdr->b_l1hdr.b_buf = buf;
2034 hdr->b_l1hdr.b_state = arc_anon;
2035 hdr->b_l1hdr.b_arc_access = 0;
2036 hdr->b_l1hdr.b_datacnt = 1;
2037 hdr->b_l1hdr.b_tmp_cdata = NULL;
2039 arc_get_data_buf(buf);
2040 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2041 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2046 static char *arc_onloan_tag = "onloan";
2049 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2050 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2051 * buffers must be returned to the arc before they can be used by the DMU or
2055 arc_loan_buf(spa_t *spa, int size)
2059 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
2061 atomic_add_64(&arc_loaned_bytes, size);
2066 * Return a loaned arc buffer to the arc.
2069 arc_return_buf(arc_buf_t *buf, void *tag)
2071 arc_buf_hdr_t *hdr = buf->b_hdr;
2073 ASSERT(buf->b_data != NULL);
2074 ASSERT(HDR_HAS_L1HDR(hdr));
2075 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2076 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2078 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
2081 /* Detach an arc_buf from a dbuf (tag) */
2083 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2085 arc_buf_hdr_t *hdr = buf->b_hdr;
2087 ASSERT(buf->b_data != NULL);
2088 ASSERT(HDR_HAS_L1HDR(hdr));
2089 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2090 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2091 buf->b_efunc = NULL;
2092 buf->b_private = NULL;
2094 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
2098 arc_buf_clone(arc_buf_t *from)
2101 arc_buf_hdr_t *hdr = from->b_hdr;
2102 uint64_t size = hdr->b_size;
2104 ASSERT(HDR_HAS_L1HDR(hdr));
2105 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2107 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2110 buf->b_efunc = NULL;
2111 buf->b_private = NULL;
2112 buf->b_next = hdr->b_l1hdr.b_buf;
2113 hdr->b_l1hdr.b_buf = buf;
2114 arc_get_data_buf(buf);
2115 bcopy(from->b_data, buf->b_data, size);
2118 * This buffer already exists in the arc so create a duplicate
2119 * copy for the caller. If the buffer is associated with user data
2120 * then track the size and number of duplicates. These stats will be
2121 * updated as duplicate buffers are created and destroyed.
2123 if (HDR_ISTYPE_DATA(hdr)) {
2124 ARCSTAT_BUMP(arcstat_duplicate_buffers);
2125 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
2127 hdr->b_l1hdr.b_datacnt += 1;
2132 arc_buf_add_ref(arc_buf_t *buf, void* tag)
2135 kmutex_t *hash_lock;
2138 * Check to see if this buffer is evicted. Callers
2139 * must verify b_data != NULL to know if the add_ref
2142 mutex_enter(&buf->b_evict_lock);
2143 if (buf->b_data == NULL) {
2144 mutex_exit(&buf->b_evict_lock);
2147 hash_lock = HDR_LOCK(buf->b_hdr);
2148 mutex_enter(hash_lock);
2150 ASSERT(HDR_HAS_L1HDR(hdr));
2151 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2152 mutex_exit(&buf->b_evict_lock);
2154 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
2155 hdr->b_l1hdr.b_state == arc_mfu);
2157 add_reference(hdr, hash_lock, tag);
2158 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2159 arc_access(hdr, hash_lock);
2160 mutex_exit(hash_lock);
2161 ARCSTAT_BUMP(arcstat_hits);
2162 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
2163 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
2164 data, metadata, hits);
2168 arc_buf_free_on_write(void *data, size_t size,
2169 void (*free_func)(void *, size_t))
2171 l2arc_data_free_t *df;
2173 df = kmem_alloc(sizeof (*df), KM_SLEEP);
2174 df->l2df_data = data;
2175 df->l2df_size = size;
2176 df->l2df_func = free_func;
2177 mutex_enter(&l2arc_free_on_write_mtx);
2178 list_insert_head(l2arc_free_on_write, df);
2179 mutex_exit(&l2arc_free_on_write_mtx);
2183 * Free the arc data buffer. If it is an l2arc write in progress,
2184 * the buffer is placed on l2arc_free_on_write to be freed later.
2187 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
2189 arc_buf_hdr_t *hdr = buf->b_hdr;
2191 if (HDR_L2_WRITING(hdr)) {
2192 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
2193 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2195 free_func(buf->b_data, hdr->b_size);
2200 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
2202 ASSERT(HDR_HAS_L2HDR(hdr));
2203 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx));
2206 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
2207 * that doesn't exist, the header is in the arc_l2c_only state,
2208 * and there isn't anything to free (it's already been freed).
2210 if (!HDR_HAS_L1HDR(hdr))
2214 * The header isn't being written to the l2arc device, thus it
2215 * shouldn't have a b_tmp_cdata to free.
2217 if (!HDR_L2_WRITING(hdr)) {
2218 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2223 * The header does not have compression enabled. This can be due
2224 * to the buffer not being compressible, or because we're
2225 * freeing the buffer before the second phase of
2226 * l2arc_write_buffer() has started (which does the compression
2227 * step). In either case, b_tmp_cdata does not point to a
2228 * separately compressed buffer, so there's nothing to free (it
2229 * points to the same buffer as the arc_buf_t's b_data field).
2231 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_OFF) {
2232 hdr->b_l1hdr.b_tmp_cdata = NULL;
2237 * There's nothing to free since the buffer was all zero's and
2238 * compressed to a zero length buffer.
2240 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) {
2241 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2245 ASSERT(L2ARC_IS_VALID_COMPRESS(hdr->b_l2hdr.b_compress));
2247 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata,
2248 hdr->b_size, zio_data_buf_free);
2250 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
2251 hdr->b_l1hdr.b_tmp_cdata = NULL;
2255 * Free up buf->b_data and if 'remove' is set, then pull the
2256 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2259 arc_buf_destroy(arc_buf_t *buf, boolean_t remove)
2263 /* free up data associated with the buf */
2264 if (buf->b_data != NULL) {
2265 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
2266 uint64_t size = buf->b_hdr->b_size;
2267 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
2269 arc_cksum_verify(buf);
2271 arc_buf_unwatch(buf);
2272 #endif /* illumos */
2274 if (type == ARC_BUFC_METADATA) {
2275 arc_buf_data_free(buf, zio_buf_free);
2276 arc_space_return(size, ARC_SPACE_META);
2278 ASSERT(type == ARC_BUFC_DATA);
2279 arc_buf_data_free(buf, zio_data_buf_free);
2280 arc_space_return(size, ARC_SPACE_DATA);
2283 /* protected by hash lock, if in the hash table */
2284 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) {
2285 uint64_t *cnt = &state->arcs_lsize[type];
2287 ASSERT(refcount_is_zero(
2288 &buf->b_hdr->b_l1hdr.b_refcnt));
2289 ASSERT(state != arc_anon && state != arc_l2c_only);
2291 ASSERT3U(*cnt, >=, size);
2292 atomic_add_64(cnt, -size);
2295 (void) refcount_remove_many(&state->arcs_size, size, buf);
2299 * If we're destroying a duplicate buffer make sure
2300 * that the appropriate statistics are updated.
2302 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 &&
2303 HDR_ISTYPE_DATA(buf->b_hdr)) {
2304 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
2305 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
2307 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0);
2308 buf->b_hdr->b_l1hdr.b_datacnt -= 1;
2311 /* only remove the buf if requested */
2315 /* remove the buf from the hdr list */
2316 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf;
2317 bufp = &(*bufp)->b_next)
2319 *bufp = buf->b_next;
2322 ASSERT(buf->b_efunc == NULL);
2324 /* clean up the buf */
2326 kmem_cache_free(buf_cache, buf);
2330 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2332 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2333 l2arc_dev_t *dev = l2hdr->b_dev;
2335 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2336 ASSERT(HDR_HAS_L2HDR(hdr));
2338 list_remove(&dev->l2ad_buflist, hdr);
2341 * We don't want to leak the b_tmp_cdata buffer that was
2342 * allocated in l2arc_write_buffers()
2344 arc_buf_l2_cdata_free(hdr);
2347 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2348 * this header is being processed by l2arc_write_buffers() (i.e.
2349 * it's in the first stage of l2arc_write_buffers()).
2350 * Re-affirming that truth here, just to serve as a reminder. If
2351 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2352 * may not have its HDR_L2_WRITING flag set. (the write may have
2353 * completed, in which case HDR_L2_WRITING will be false and the
2354 * b_daddr field will point to the address of the buffer on disk).
2356 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr));
2359 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2360 * l2arc_write_buffers(). Since we've just removed this header
2361 * from the l2arc buffer list, this header will never reach the
2362 * second stage of l2arc_write_buffers(), which increments the
2363 * accounting stats for this header. Thus, we must be careful
2364 * not to decrement them for this header either.
2366 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) {
2367 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
2368 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
2370 vdev_space_update(dev->l2ad_vdev,
2371 -l2hdr->b_asize, 0, 0);
2373 (void) refcount_remove_many(&dev->l2ad_alloc,
2374 l2hdr->b_asize, hdr);
2377 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
2381 arc_hdr_destroy(arc_buf_hdr_t *hdr)
2383 if (HDR_HAS_L1HDR(hdr)) {
2384 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
2385 hdr->b_l1hdr.b_datacnt > 0);
2386 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2387 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2389 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2390 ASSERT(!HDR_IN_HASH_TABLE(hdr));
2392 if (HDR_HAS_L2HDR(hdr)) {
2393 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2394 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
2397 mutex_enter(&dev->l2ad_mtx);
2400 * Even though we checked this conditional above, we
2401 * need to check this again now that we have the
2402 * l2ad_mtx. This is because we could be racing with
2403 * another thread calling l2arc_evict() which might have
2404 * destroyed this header's L2 portion as we were waiting
2405 * to acquire the l2ad_mtx. If that happens, we don't
2406 * want to re-destroy the header's L2 portion.
2408 if (HDR_HAS_L2HDR(hdr)) {
2409 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET)
2410 trim_map_free(dev->l2ad_vdev,
2411 hdr->b_l2hdr.b_daddr,
2412 hdr->b_l2hdr.b_asize, 0);
2413 arc_hdr_l2hdr_destroy(hdr);
2417 mutex_exit(&dev->l2ad_mtx);
2420 if (!BUF_EMPTY(hdr))
2421 buf_discard_identity(hdr);
2423 if (hdr->b_freeze_cksum != NULL) {
2424 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
2425 hdr->b_freeze_cksum = NULL;
2428 if (HDR_HAS_L1HDR(hdr)) {
2429 while (hdr->b_l1hdr.b_buf) {
2430 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2432 if (buf->b_efunc != NULL) {
2433 mutex_enter(&arc_user_evicts_lock);
2434 mutex_enter(&buf->b_evict_lock);
2435 ASSERT(buf->b_hdr != NULL);
2436 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE);
2437 hdr->b_l1hdr.b_buf = buf->b_next;
2438 buf->b_hdr = &arc_eviction_hdr;
2439 buf->b_next = arc_eviction_list;
2440 arc_eviction_list = buf;
2441 mutex_exit(&buf->b_evict_lock);
2442 cv_signal(&arc_user_evicts_cv);
2443 mutex_exit(&arc_user_evicts_lock);
2445 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE);
2449 if (hdr->b_l1hdr.b_thawed != NULL) {
2450 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2451 hdr->b_l1hdr.b_thawed = NULL;
2456 ASSERT3P(hdr->b_hash_next, ==, NULL);
2457 if (HDR_HAS_L1HDR(hdr)) {
2458 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2459 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
2460 kmem_cache_free(hdr_full_cache, hdr);
2462 kmem_cache_free(hdr_l2only_cache, hdr);
2467 arc_buf_free(arc_buf_t *buf, void *tag)
2469 arc_buf_hdr_t *hdr = buf->b_hdr;
2470 int hashed = hdr->b_l1hdr.b_state != arc_anon;
2472 ASSERT(buf->b_efunc == NULL);
2473 ASSERT(buf->b_data != NULL);
2476 kmutex_t *hash_lock = HDR_LOCK(hdr);
2478 mutex_enter(hash_lock);
2480 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2482 (void) remove_reference(hdr, hash_lock, tag);
2483 if (hdr->b_l1hdr.b_datacnt > 1) {
2484 arc_buf_destroy(buf, TRUE);
2486 ASSERT(buf == hdr->b_l1hdr.b_buf);
2487 ASSERT(buf->b_efunc == NULL);
2488 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2490 mutex_exit(hash_lock);
2491 } else if (HDR_IO_IN_PROGRESS(hdr)) {
2494 * We are in the middle of an async write. Don't destroy
2495 * this buffer unless the write completes before we finish
2496 * decrementing the reference count.
2498 mutex_enter(&arc_user_evicts_lock);
2499 (void) remove_reference(hdr, NULL, tag);
2500 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2501 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
2502 mutex_exit(&arc_user_evicts_lock);
2504 arc_hdr_destroy(hdr);
2506 if (remove_reference(hdr, NULL, tag) > 0)
2507 arc_buf_destroy(buf, TRUE);
2509 arc_hdr_destroy(hdr);
2514 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
2516 arc_buf_hdr_t *hdr = buf->b_hdr;
2517 kmutex_t *hash_lock = HDR_LOCK(hdr);
2518 boolean_t no_callback = (buf->b_efunc == NULL);
2520 if (hdr->b_l1hdr.b_state == arc_anon) {
2521 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
2522 arc_buf_free(buf, tag);
2523 return (no_callback);
2526 mutex_enter(hash_lock);
2528 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
2529 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2530 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2531 ASSERT(buf->b_data != NULL);
2533 (void) remove_reference(hdr, hash_lock, tag);
2534 if (hdr->b_l1hdr.b_datacnt > 1) {
2536 arc_buf_destroy(buf, TRUE);
2537 } else if (no_callback) {
2538 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL);
2539 ASSERT(buf->b_efunc == NULL);
2540 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2542 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 ||
2543 refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2544 mutex_exit(hash_lock);
2545 return (no_callback);
2549 arc_buf_size(arc_buf_t *buf)
2551 return (buf->b_hdr->b_size);
2555 * Called from the DMU to determine if the current buffer should be
2556 * evicted. In order to ensure proper locking, the eviction must be initiated
2557 * from the DMU. Return true if the buffer is associated with user data and
2558 * duplicate buffers still exist.
2561 arc_buf_eviction_needed(arc_buf_t *buf)
2564 boolean_t evict_needed = B_FALSE;
2566 if (zfs_disable_dup_eviction)
2569 mutex_enter(&buf->b_evict_lock);
2573 * We are in arc_do_user_evicts(); let that function
2574 * perform the eviction.
2576 ASSERT(buf->b_data == NULL);
2577 mutex_exit(&buf->b_evict_lock);
2579 } else if (buf->b_data == NULL) {
2581 * We have already been added to the arc eviction list;
2582 * recommend eviction.
2584 ASSERT3P(hdr, ==, &arc_eviction_hdr);
2585 mutex_exit(&buf->b_evict_lock);
2589 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr))
2590 evict_needed = B_TRUE;
2592 mutex_exit(&buf->b_evict_lock);
2593 return (evict_needed);
2597 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2598 * state of the header is dependent on it's state prior to entering this
2599 * function. The following transitions are possible:
2601 * - arc_mru -> arc_mru_ghost
2602 * - arc_mfu -> arc_mfu_ghost
2603 * - arc_mru_ghost -> arc_l2c_only
2604 * - arc_mru_ghost -> deleted
2605 * - arc_mfu_ghost -> arc_l2c_only
2606 * - arc_mfu_ghost -> deleted
2609 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
2611 arc_state_t *evicted_state, *state;
2612 int64_t bytes_evicted = 0;
2614 ASSERT(MUTEX_HELD(hash_lock));
2615 ASSERT(HDR_HAS_L1HDR(hdr));
2617 state = hdr->b_l1hdr.b_state;
2618 if (GHOST_STATE(state)) {
2619 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2620 ASSERT(hdr->b_l1hdr.b_buf == NULL);
2623 * l2arc_write_buffers() relies on a header's L1 portion
2624 * (i.e. it's b_tmp_cdata field) during it's write phase.
2625 * Thus, we cannot push a header onto the arc_l2c_only
2626 * state (removing it's L1 piece) until the header is
2627 * done being written to the l2arc.
2629 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
2630 ARCSTAT_BUMP(arcstat_evict_l2_skip);
2631 return (bytes_evicted);
2634 ARCSTAT_BUMP(arcstat_deleted);
2635 bytes_evicted += hdr->b_size;
2637 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
2639 if (HDR_HAS_L2HDR(hdr)) {
2641 * This buffer is cached on the 2nd Level ARC;
2642 * don't destroy the header.
2644 arc_change_state(arc_l2c_only, hdr, hash_lock);
2646 * dropping from L1+L2 cached to L2-only,
2647 * realloc to remove the L1 header.
2649 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
2652 arc_change_state(arc_anon, hdr, hash_lock);
2653 arc_hdr_destroy(hdr);
2655 return (bytes_evicted);
2658 ASSERT(state == arc_mru || state == arc_mfu);
2659 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2661 /* prefetch buffers have a minimum lifespan */
2662 if (HDR_IO_IN_PROGRESS(hdr) ||
2663 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
2664 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
2665 arc_min_prefetch_lifespan)) {
2666 ARCSTAT_BUMP(arcstat_evict_skip);
2667 return (bytes_evicted);
2670 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
2671 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0);
2672 while (hdr->b_l1hdr.b_buf) {
2673 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2674 if (!mutex_tryenter(&buf->b_evict_lock)) {
2675 ARCSTAT_BUMP(arcstat_mutex_miss);
2678 if (buf->b_data != NULL)
2679 bytes_evicted += hdr->b_size;
2680 if (buf->b_efunc != NULL) {
2681 mutex_enter(&arc_user_evicts_lock);
2682 arc_buf_destroy(buf, FALSE);
2683 hdr->b_l1hdr.b_buf = buf->b_next;
2684 buf->b_hdr = &arc_eviction_hdr;
2685 buf->b_next = arc_eviction_list;
2686 arc_eviction_list = buf;
2687 cv_signal(&arc_user_evicts_cv);
2688 mutex_exit(&arc_user_evicts_lock);
2689 mutex_exit(&buf->b_evict_lock);
2691 mutex_exit(&buf->b_evict_lock);
2692 arc_buf_destroy(buf, TRUE);
2696 if (HDR_HAS_L2HDR(hdr)) {
2697 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size);
2699 if (l2arc_write_eligible(hdr->b_spa, hdr))
2700 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size);
2702 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size);
2705 if (hdr->b_l1hdr.b_datacnt == 0) {
2706 arc_change_state(evicted_state, hdr, hash_lock);
2707 ASSERT(HDR_IN_HASH_TABLE(hdr));
2708 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
2709 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
2710 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
2713 return (bytes_evicted);
2717 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
2718 uint64_t spa, int64_t bytes)
2720 multilist_sublist_t *mls;
2721 uint64_t bytes_evicted = 0;
2723 kmutex_t *hash_lock;
2724 int evict_count = 0;
2726 ASSERT3P(marker, !=, NULL);
2727 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2729 mls = multilist_sublist_lock(ml, idx);
2731 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
2732 hdr = multilist_sublist_prev(mls, marker)) {
2733 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
2734 (evict_count >= zfs_arc_evict_batch_limit))
2738 * To keep our iteration location, move the marker
2739 * forward. Since we're not holding hdr's hash lock, we
2740 * must be very careful and not remove 'hdr' from the
2741 * sublist. Otherwise, other consumers might mistake the
2742 * 'hdr' as not being on a sublist when they call the
2743 * multilist_link_active() function (they all rely on
2744 * the hash lock protecting concurrent insertions and
2745 * removals). multilist_sublist_move_forward() was
2746 * specifically implemented to ensure this is the case
2747 * (only 'marker' will be removed and re-inserted).
2749 multilist_sublist_move_forward(mls, marker);
2752 * The only case where the b_spa field should ever be
2753 * zero, is the marker headers inserted by
2754 * arc_evict_state(). It's possible for multiple threads
2755 * to be calling arc_evict_state() concurrently (e.g.
2756 * dsl_pool_close() and zio_inject_fault()), so we must
2757 * skip any markers we see from these other threads.
2759 if (hdr->b_spa == 0)
2762 /* we're only interested in evicting buffers of a certain spa */
2763 if (spa != 0 && hdr->b_spa != spa) {
2764 ARCSTAT_BUMP(arcstat_evict_skip);
2768 hash_lock = HDR_LOCK(hdr);
2771 * We aren't calling this function from any code path
2772 * that would already be holding a hash lock, so we're
2773 * asserting on this assumption to be defensive in case
2774 * this ever changes. Without this check, it would be
2775 * possible to incorrectly increment arcstat_mutex_miss
2776 * below (e.g. if the code changed such that we called
2777 * this function with a hash lock held).
2779 ASSERT(!MUTEX_HELD(hash_lock));
2781 if (mutex_tryenter(hash_lock)) {
2782 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
2783 mutex_exit(hash_lock);
2785 bytes_evicted += evicted;
2788 * If evicted is zero, arc_evict_hdr() must have
2789 * decided to skip this header, don't increment
2790 * evict_count in this case.
2796 * If arc_size isn't overflowing, signal any
2797 * threads that might happen to be waiting.
2799 * For each header evicted, we wake up a single
2800 * thread. If we used cv_broadcast, we could
2801 * wake up "too many" threads causing arc_size
2802 * to significantly overflow arc_c; since
2803 * arc_get_data_buf() doesn't check for overflow
2804 * when it's woken up (it doesn't because it's
2805 * possible for the ARC to be overflowing while
2806 * full of un-evictable buffers, and the
2807 * function should proceed in this case).
2809 * If threads are left sleeping, due to not
2810 * using cv_broadcast, they will be woken up
2811 * just before arc_reclaim_thread() sleeps.
2813 mutex_enter(&arc_reclaim_lock);
2814 if (!arc_is_overflowing())
2815 cv_signal(&arc_reclaim_waiters_cv);
2816 mutex_exit(&arc_reclaim_lock);
2818 ARCSTAT_BUMP(arcstat_mutex_miss);
2822 multilist_sublist_unlock(mls);
2824 return (bytes_evicted);
2828 * Evict buffers from the given arc state, until we've removed the
2829 * specified number of bytes. Move the removed buffers to the
2830 * appropriate evict state.
2832 * This function makes a "best effort". It skips over any buffers
2833 * it can't get a hash_lock on, and so, may not catch all candidates.
2834 * It may also return without evicting as much space as requested.
2836 * If bytes is specified using the special value ARC_EVICT_ALL, this
2837 * will evict all available (i.e. unlocked and evictable) buffers from
2838 * the given arc state; which is used by arc_flush().
2841 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
2842 arc_buf_contents_t type)
2844 uint64_t total_evicted = 0;
2845 multilist_t *ml = &state->arcs_list[type];
2847 arc_buf_hdr_t **markers;
2849 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2851 num_sublists = multilist_get_num_sublists(ml);
2854 * If we've tried to evict from each sublist, made some
2855 * progress, but still have not hit the target number of bytes
2856 * to evict, we want to keep trying. The markers allow us to
2857 * pick up where we left off for each individual sublist, rather
2858 * than starting from the tail each time.
2860 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
2861 for (int i = 0; i < num_sublists; i++) {
2862 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
2865 * A b_spa of 0 is used to indicate that this header is
2866 * a marker. This fact is used in arc_adjust_type() and
2867 * arc_evict_state_impl().
2869 markers[i]->b_spa = 0;
2871 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2872 multilist_sublist_insert_tail(mls, markers[i]);
2873 multilist_sublist_unlock(mls);
2877 * While we haven't hit our target number of bytes to evict, or
2878 * we're evicting all available buffers.
2880 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
2882 * Start eviction using a randomly selected sublist,
2883 * this is to try and evenly balance eviction across all
2884 * sublists. Always starting at the same sublist
2885 * (e.g. index 0) would cause evictions to favor certain
2886 * sublists over others.
2888 int sublist_idx = multilist_get_random_index(ml);
2889 uint64_t scan_evicted = 0;
2891 for (int i = 0; i < num_sublists; i++) {
2892 uint64_t bytes_remaining;
2893 uint64_t bytes_evicted;
2895 if (bytes == ARC_EVICT_ALL)
2896 bytes_remaining = ARC_EVICT_ALL;
2897 else if (total_evicted < bytes)
2898 bytes_remaining = bytes - total_evicted;
2902 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
2903 markers[sublist_idx], spa, bytes_remaining);
2905 scan_evicted += bytes_evicted;
2906 total_evicted += bytes_evicted;
2908 /* we've reached the end, wrap to the beginning */
2909 if (++sublist_idx >= num_sublists)
2914 * If we didn't evict anything during this scan, we have
2915 * no reason to believe we'll evict more during another
2916 * scan, so break the loop.
2918 if (scan_evicted == 0) {
2919 /* This isn't possible, let's make that obvious */
2920 ASSERT3S(bytes, !=, 0);
2923 * When bytes is ARC_EVICT_ALL, the only way to
2924 * break the loop is when scan_evicted is zero.
2925 * In that case, we actually have evicted enough,
2926 * so we don't want to increment the kstat.
2928 if (bytes != ARC_EVICT_ALL) {
2929 ASSERT3S(total_evicted, <, bytes);
2930 ARCSTAT_BUMP(arcstat_evict_not_enough);
2937 for (int i = 0; i < num_sublists; i++) {
2938 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2939 multilist_sublist_remove(mls, markers[i]);
2940 multilist_sublist_unlock(mls);
2942 kmem_cache_free(hdr_full_cache, markers[i]);
2944 kmem_free(markers, sizeof (*markers) * num_sublists);
2946 return (total_evicted);
2950 * Flush all "evictable" data of the given type from the arc state
2951 * specified. This will not evict any "active" buffers (i.e. referenced).
2953 * When 'retry' is set to FALSE, the function will make a single pass
2954 * over the state and evict any buffers that it can. Since it doesn't
2955 * continually retry the eviction, it might end up leaving some buffers
2956 * in the ARC due to lock misses.
2958 * When 'retry' is set to TRUE, the function will continually retry the
2959 * eviction until *all* evictable buffers have been removed from the
2960 * state. As a result, if concurrent insertions into the state are
2961 * allowed (e.g. if the ARC isn't shutting down), this function might
2962 * wind up in an infinite loop, continually trying to evict buffers.
2965 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
2968 uint64_t evicted = 0;
2970 while (state->arcs_lsize[type] != 0) {
2971 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
2981 * Evict the specified number of bytes from the state specified,
2982 * restricting eviction to the spa and type given. This function
2983 * prevents us from trying to evict more from a state's list than
2984 * is "evictable", and to skip evicting altogether when passed a
2985 * negative value for "bytes". In contrast, arc_evict_state() will
2986 * evict everything it can, when passed a negative value for "bytes".
2989 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
2990 arc_buf_contents_t type)
2994 if (bytes > 0 && state->arcs_lsize[type] > 0) {
2995 delta = MIN(state->arcs_lsize[type], bytes);
2996 return (arc_evict_state(state, spa, delta, type));
3003 * Evict metadata buffers from the cache, such that arc_meta_used is
3004 * capped by the arc_meta_limit tunable.
3007 arc_adjust_meta(void)
3009 uint64_t total_evicted = 0;
3013 * If we're over the meta limit, we want to evict enough
3014 * metadata to get back under the meta limit. We don't want to
3015 * evict so much that we drop the MRU below arc_p, though. If
3016 * we're over the meta limit more than we're over arc_p, we
3017 * evict some from the MRU here, and some from the MFU below.
3019 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3020 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3021 refcount_count(&arc_mru->arcs_size) - arc_p));
3023 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3026 * Similar to the above, we want to evict enough bytes to get us
3027 * below the meta limit, but not so much as to drop us below the
3028 * space alloted to the MFU (which is defined as arc_c - arc_p).
3030 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3031 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
3033 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3035 return (total_evicted);
3039 * Return the type of the oldest buffer in the given arc state
3041 * This function will select a random sublist of type ARC_BUFC_DATA and
3042 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3043 * is compared, and the type which contains the "older" buffer will be
3046 static arc_buf_contents_t
3047 arc_adjust_type(arc_state_t *state)
3049 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
3050 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
3051 int data_idx = multilist_get_random_index(data_ml);
3052 int meta_idx = multilist_get_random_index(meta_ml);
3053 multilist_sublist_t *data_mls;
3054 multilist_sublist_t *meta_mls;
3055 arc_buf_contents_t type;
3056 arc_buf_hdr_t *data_hdr;
3057 arc_buf_hdr_t *meta_hdr;
3060 * We keep the sublist lock until we're finished, to prevent
3061 * the headers from being destroyed via arc_evict_state().
3063 data_mls = multilist_sublist_lock(data_ml, data_idx);
3064 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3067 * These two loops are to ensure we skip any markers that
3068 * might be at the tail of the lists due to arc_evict_state().
3071 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3072 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3073 if (data_hdr->b_spa != 0)
3077 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3078 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3079 if (meta_hdr->b_spa != 0)
3083 if (data_hdr == NULL && meta_hdr == NULL) {
3084 type = ARC_BUFC_DATA;
3085 } else if (data_hdr == NULL) {
3086 ASSERT3P(meta_hdr, !=, NULL);
3087 type = ARC_BUFC_METADATA;
3088 } else if (meta_hdr == NULL) {
3089 ASSERT3P(data_hdr, !=, NULL);
3090 type = ARC_BUFC_DATA;
3092 ASSERT3P(data_hdr, !=, NULL);
3093 ASSERT3P(meta_hdr, !=, NULL);
3095 /* The headers can't be on the sublist without an L1 header */
3096 ASSERT(HDR_HAS_L1HDR(data_hdr));
3097 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3099 if (data_hdr->b_l1hdr.b_arc_access <
3100 meta_hdr->b_l1hdr.b_arc_access) {
3101 type = ARC_BUFC_DATA;
3103 type = ARC_BUFC_METADATA;
3107 multilist_sublist_unlock(meta_mls);
3108 multilist_sublist_unlock(data_mls);
3114 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3119 uint64_t total_evicted = 0;
3124 * If we're over arc_meta_limit, we want to correct that before
3125 * potentially evicting data buffers below.
3127 total_evicted += arc_adjust_meta();
3132 * If we're over the target cache size, we want to evict enough
3133 * from the list to get back to our target size. We don't want
3134 * to evict too much from the MRU, such that it drops below
3135 * arc_p. So, if we're over our target cache size more than
3136 * the MRU is over arc_p, we'll evict enough to get back to
3137 * arc_p here, and then evict more from the MFU below.
3139 target = MIN((int64_t)(arc_size - arc_c),
3140 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3141 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
3144 * If we're below arc_meta_min, always prefer to evict data.
3145 * Otherwise, try to satisfy the requested number of bytes to
3146 * evict from the type which contains older buffers; in an
3147 * effort to keep newer buffers in the cache regardless of their
3148 * type. If we cannot satisfy the number of bytes from this
3149 * type, spill over into the next type.
3151 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
3152 arc_meta_used > arc_meta_min) {
3153 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3154 total_evicted += bytes;
3157 * If we couldn't evict our target number of bytes from
3158 * metadata, we try to get the rest from data.
3163 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3165 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3166 total_evicted += bytes;
3169 * If we couldn't evict our target number of bytes from
3170 * data, we try to get the rest from metadata.
3175 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3181 * Now that we've tried to evict enough from the MRU to get its
3182 * size back to arc_p, if we're still above the target cache
3183 * size, we evict the rest from the MFU.
3185 target = arc_size - arc_c;
3187 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
3188 arc_meta_used > arc_meta_min) {
3189 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3190 total_evicted += bytes;
3193 * If we couldn't evict our target number of bytes from
3194 * metadata, we try to get the rest from data.
3199 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3201 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3202 total_evicted += bytes;
3205 * If we couldn't evict our target number of bytes from
3206 * data, we try to get the rest from data.
3211 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3215 * Adjust ghost lists
3217 * In addition to the above, the ARC also defines target values
3218 * for the ghost lists. The sum of the mru list and mru ghost
3219 * list should never exceed the target size of the cache, and
3220 * the sum of the mru list, mfu list, mru ghost list, and mfu
3221 * ghost list should never exceed twice the target size of the
3222 * cache. The following logic enforces these limits on the ghost
3223 * caches, and evicts from them as needed.
3225 target = refcount_count(&arc_mru->arcs_size) +
3226 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3228 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3229 total_evicted += bytes;
3234 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3237 * We assume the sum of the mru list and mfu list is less than
3238 * or equal to arc_c (we enforced this above), which means we
3239 * can use the simpler of the two equations below:
3241 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3242 * mru ghost + mfu ghost <= arc_c
3244 target = refcount_count(&arc_mru_ghost->arcs_size) +
3245 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3247 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3248 total_evicted += bytes;
3253 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3255 return (total_evicted);
3259 arc_do_user_evicts(void)
3261 mutex_enter(&arc_user_evicts_lock);
3262 while (arc_eviction_list != NULL) {
3263 arc_buf_t *buf = arc_eviction_list;
3264 arc_eviction_list = buf->b_next;
3265 mutex_enter(&buf->b_evict_lock);
3267 mutex_exit(&buf->b_evict_lock);
3268 mutex_exit(&arc_user_evicts_lock);
3270 if (buf->b_efunc != NULL)
3271 VERIFY0(buf->b_efunc(buf->b_private));
3273 buf->b_efunc = NULL;
3274 buf->b_private = NULL;
3275 kmem_cache_free(buf_cache, buf);
3276 mutex_enter(&arc_user_evicts_lock);
3278 mutex_exit(&arc_user_evicts_lock);
3282 arc_flush(spa_t *spa, boolean_t retry)
3287 * If retry is TRUE, a spa must not be specified since we have
3288 * no good way to determine if all of a spa's buffers have been
3289 * evicted from an arc state.
3291 ASSERT(!retry || spa == 0);
3294 guid = spa_load_guid(spa);
3296 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3297 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3299 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3300 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3302 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3303 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3305 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3306 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3308 arc_do_user_evicts();
3309 ASSERT(spa || arc_eviction_list == NULL);
3313 arc_shrink(int64_t to_free)
3315 if (arc_c > arc_c_min) {
3316 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
3317 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
3318 if (arc_c > arc_c_min + to_free)
3319 atomic_add_64(&arc_c, -to_free);
3323 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3324 if (arc_c > arc_size)
3325 arc_c = MAX(arc_size, arc_c_min);
3327 arc_p = (arc_c >> 1);
3329 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
3332 ASSERT(arc_c >= arc_c_min);
3333 ASSERT((int64_t)arc_p >= 0);
3336 if (arc_size > arc_c) {
3337 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size,
3339 (void) arc_adjust();
3343 static long needfree = 0;
3345 typedef enum free_memory_reason_t {
3350 FMR_PAGES_PP_MAXIMUM,
3354 } free_memory_reason_t;
3356 int64_t last_free_memory;
3357 free_memory_reason_t last_free_reason;
3360 * Additional reserve of pages for pp_reserve.
3362 int64_t arc_pages_pp_reserve = 64;
3365 * Additional reserve of pages for swapfs.
3367 int64_t arc_swapfs_reserve = 64;
3370 * Return the amount of memory that can be consumed before reclaim will be
3371 * needed. Positive if there is sufficient free memory, negative indicates
3372 * the amount of memory that needs to be freed up.
3375 arc_available_memory(void)
3377 int64_t lowest = INT64_MAX;
3379 free_memory_reason_t r = FMR_UNKNOWN;
3383 n = PAGESIZE * (-needfree);
3391 * Cooperate with pagedaemon when it's time for it to scan
3392 * and reclaim some pages.
3394 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
3402 * check that we're out of range of the pageout scanner. It starts to
3403 * schedule paging if freemem is less than lotsfree and needfree.
3404 * lotsfree is the high-water mark for pageout, and needfree is the
3405 * number of needed free pages. We add extra pages here to make sure
3406 * the scanner doesn't start up while we're freeing memory.
3408 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3415 * check to make sure that swapfs has enough space so that anon
3416 * reservations can still succeed. anon_resvmem() checks that the
3417 * availrmem is greater than swapfs_minfree, and the number of reserved
3418 * swap pages. We also add a bit of extra here just to prevent
3419 * circumstances from getting really dire.
3421 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3422 desfree - arc_swapfs_reserve);
3425 r = FMR_SWAPFS_MINFREE;
3430 * Check that we have enough availrmem that memory locking (e.g., via
3431 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3432 * stores the number of pages that cannot be locked; when availrmem
3433 * drops below pages_pp_maximum, page locking mechanisms such as
3434 * page_pp_lock() will fail.)
3436 n = PAGESIZE * (availrmem - pages_pp_maximum -
3437 arc_pages_pp_reserve);
3440 r = FMR_PAGES_PP_MAXIMUM;
3444 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
3446 * If we're on an i386 platform, it's possible that we'll exhaust the
3447 * kernel heap space before we ever run out of available physical
3448 * memory. Most checks of the size of the heap_area compare against
3449 * tune.t_minarmem, which is the minimum available real memory that we
3450 * can have in the system. However, this is generally fixed at 25 pages
3451 * which is so low that it's useless. In this comparison, we seek to
3452 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3453 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3456 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
3457 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3462 #define zio_arena NULL
3464 #define zio_arena heap_arena
3468 * If zio data pages are being allocated out of a separate heap segment,
3469 * then enforce that the size of available vmem for this arena remains
3470 * above about 1/16th free.
3472 * Note: The 1/16th arena free requirement was put in place
3473 * to aggressively evict memory from the arc in order to avoid
3474 * memory fragmentation issues.
3476 if (zio_arena != NULL) {
3477 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
3478 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3486 * Above limits know nothing about real level of KVA fragmentation.
3487 * Start aggressive reclamation if too little sequential KVA left.
3490 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ?
3491 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) :
3500 /* Every 100 calls, free a small amount */
3501 if (spa_get_random(100) == 0)
3503 #endif /* _KERNEL */
3505 last_free_memory = lowest;
3506 last_free_reason = r;
3507 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
3513 * Determine if the system is under memory pressure and is asking
3514 * to reclaim memory. A return value of TRUE indicates that the system
3515 * is under memory pressure and that the arc should adjust accordingly.
3518 arc_reclaim_needed(void)
3520 return (arc_available_memory() < 0);
3523 extern kmem_cache_t *zio_buf_cache[];
3524 extern kmem_cache_t *zio_data_buf_cache[];
3525 extern kmem_cache_t *range_seg_cache;
3527 static __noinline void
3528 arc_kmem_reap_now(void)
3531 kmem_cache_t *prev_cache = NULL;
3532 kmem_cache_t *prev_data_cache = NULL;
3534 DTRACE_PROBE(arc__kmem_reap_start);
3536 if (arc_meta_used >= arc_meta_limit) {
3538 * We are exceeding our meta-data cache limit.
3539 * Purge some DNLC entries to release holds on meta-data.
3541 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
3545 * Reclaim unused memory from all kmem caches.
3551 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
3552 if (zio_buf_cache[i] != prev_cache) {
3553 prev_cache = zio_buf_cache[i];
3554 kmem_cache_reap_now(zio_buf_cache[i]);
3556 if (zio_data_buf_cache[i] != prev_data_cache) {
3557 prev_data_cache = zio_data_buf_cache[i];
3558 kmem_cache_reap_now(zio_data_buf_cache[i]);
3561 kmem_cache_reap_now(buf_cache);
3562 kmem_cache_reap_now(hdr_full_cache);
3563 kmem_cache_reap_now(hdr_l2only_cache);
3564 kmem_cache_reap_now(range_seg_cache);
3567 if (zio_arena != NULL) {
3569 * Ask the vmem arena to reclaim unused memory from its
3572 vmem_qcache_reap(zio_arena);
3575 DTRACE_PROBE(arc__kmem_reap_end);
3579 * Threads can block in arc_get_data_buf() waiting for this thread to evict
3580 * enough data and signal them to proceed. When this happens, the threads in
3581 * arc_get_data_buf() are sleeping while holding the hash lock for their
3582 * particular arc header. Thus, we must be careful to never sleep on a
3583 * hash lock in this thread. This is to prevent the following deadlock:
3585 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3586 * waiting for the reclaim thread to signal it.
3588 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3589 * fails, and goes to sleep forever.
3591 * This possible deadlock is avoided by always acquiring a hash lock
3592 * using mutex_tryenter() from arc_reclaim_thread().
3595 arc_reclaim_thread(void *dummy __unused)
3597 clock_t growtime = 0;
3600 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
3602 mutex_enter(&arc_reclaim_lock);
3603 while (!arc_reclaim_thread_exit) {
3604 int64_t free_memory = arc_available_memory();
3605 uint64_t evicted = 0;
3607 mutex_exit(&arc_reclaim_lock);
3609 if (free_memory < 0) {
3611 arc_no_grow = B_TRUE;
3615 * Wait at least zfs_grow_retry (default 60) seconds
3616 * before considering growing.
3618 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
3620 arc_kmem_reap_now();
3623 * If we are still low on memory, shrink the ARC
3624 * so that we have arc_shrink_min free space.
3626 free_memory = arc_available_memory();
3629 (arc_c >> arc_shrink_shift) - free_memory;
3632 to_free = MAX(to_free, ptob(needfree));
3634 arc_shrink(to_free);
3636 } else if (free_memory < arc_c >> arc_no_grow_shift) {
3637 arc_no_grow = B_TRUE;
3638 } else if (ddi_get_lbolt() >= growtime) {
3639 arc_no_grow = B_FALSE;
3642 evicted = arc_adjust();
3644 mutex_enter(&arc_reclaim_lock);
3647 * If evicted is zero, we couldn't evict anything via
3648 * arc_adjust(). This could be due to hash lock
3649 * collisions, but more likely due to the majority of
3650 * arc buffers being unevictable. Therefore, even if
3651 * arc_size is above arc_c, another pass is unlikely to
3652 * be helpful and could potentially cause us to enter an
3655 if (arc_size <= arc_c || evicted == 0) {
3660 * We're either no longer overflowing, or we
3661 * can't evict anything more, so we should wake
3662 * up any threads before we go to sleep.
3664 cv_broadcast(&arc_reclaim_waiters_cv);
3667 * Block until signaled, or after one second (we
3668 * might need to perform arc_kmem_reap_now()
3669 * even if we aren't being signalled)
3671 CALLB_CPR_SAFE_BEGIN(&cpr);
3672 (void) cv_timedwait(&arc_reclaim_thread_cv,
3673 &arc_reclaim_lock, hz);
3674 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
3678 arc_reclaim_thread_exit = FALSE;
3679 cv_broadcast(&arc_reclaim_thread_cv);
3680 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
3685 arc_user_evicts_thread(void *dummy __unused)
3689 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG);
3691 mutex_enter(&arc_user_evicts_lock);
3692 while (!arc_user_evicts_thread_exit) {
3693 mutex_exit(&arc_user_evicts_lock);
3695 arc_do_user_evicts();
3698 * This is necessary in order for the mdb ::arc dcmd to
3699 * show up to date information. Since the ::arc command
3700 * does not call the kstat's update function, without
3701 * this call, the command may show stale stats for the
3702 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3703 * with this change, the data might be up to 1 second
3704 * out of date; but that should suffice. The arc_state_t
3705 * structures can be queried directly if more accurate
3706 * information is needed.
3708 if (arc_ksp != NULL)
3709 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
3711 mutex_enter(&arc_user_evicts_lock);
3714 * Block until signaled, or after one second (we need to
3715 * call the arc's kstat update function regularly).
3717 CALLB_CPR_SAFE_BEGIN(&cpr);
3718 (void) cv_timedwait(&arc_user_evicts_cv,
3719 &arc_user_evicts_lock, hz);
3720 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock);
3723 arc_user_evicts_thread_exit = FALSE;
3724 cv_broadcast(&arc_user_evicts_cv);
3725 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */
3730 * Adapt arc info given the number of bytes we are trying to add and
3731 * the state that we are comming from. This function is only called
3732 * when we are adding new content to the cache.
3735 arc_adapt(int bytes, arc_state_t *state)
3738 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
3739 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
3740 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
3742 if (state == arc_l2c_only)
3747 * Adapt the target size of the MRU list:
3748 * - if we just hit in the MRU ghost list, then increase
3749 * the target size of the MRU list.
3750 * - if we just hit in the MFU ghost list, then increase
3751 * the target size of the MFU list by decreasing the
3752 * target size of the MRU list.
3754 if (state == arc_mru_ghost) {
3755 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
3756 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
3758 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
3759 } else if (state == arc_mfu_ghost) {
3762 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
3763 mult = MIN(mult, 10);
3765 delta = MIN(bytes * mult, arc_p);
3766 arc_p = MAX(arc_p_min, arc_p - delta);
3768 ASSERT((int64_t)arc_p >= 0);
3770 if (arc_reclaim_needed()) {
3771 cv_signal(&arc_reclaim_thread_cv);
3778 if (arc_c >= arc_c_max)
3782 * If we're within (2 * maxblocksize) bytes of the target
3783 * cache size, increment the target cache size
3785 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
3786 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
3787 atomic_add_64(&arc_c, (int64_t)bytes);
3788 if (arc_c > arc_c_max)
3790 else if (state == arc_anon)
3791 atomic_add_64(&arc_p, (int64_t)bytes);
3795 ASSERT((int64_t)arc_p >= 0);
3799 * Check if arc_size has grown past our upper threshold, determined by
3800 * zfs_arc_overflow_shift.
3803 arc_is_overflowing(void)
3805 /* Always allow at least one block of overflow */
3806 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
3807 arc_c >> zfs_arc_overflow_shift);
3809 return (arc_size >= arc_c + overflow);
3813 * The buffer, supplied as the first argument, needs a data block. If we
3814 * are hitting the hard limit for the cache size, we must sleep, waiting
3815 * for the eviction thread to catch up. If we're past the target size
3816 * but below the hard limit, we'll only signal the reclaim thread and
3820 arc_get_data_buf(arc_buf_t *buf)
3822 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
3823 uint64_t size = buf->b_hdr->b_size;
3824 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
3826 arc_adapt(size, state);
3829 * If arc_size is currently overflowing, and has grown past our
3830 * upper limit, we must be adding data faster than the evict
3831 * thread can evict. Thus, to ensure we don't compound the
3832 * problem by adding more data and forcing arc_size to grow even
3833 * further past it's target size, we halt and wait for the
3834 * eviction thread to catch up.
3836 * It's also possible that the reclaim thread is unable to evict
3837 * enough buffers to get arc_size below the overflow limit (e.g.
3838 * due to buffers being un-evictable, or hash lock collisions).
3839 * In this case, we want to proceed regardless if we're
3840 * overflowing; thus we don't use a while loop here.
3842 if (arc_is_overflowing()) {
3843 mutex_enter(&arc_reclaim_lock);
3846 * Now that we've acquired the lock, we may no longer be
3847 * over the overflow limit, lets check.
3849 * We're ignoring the case of spurious wake ups. If that
3850 * were to happen, it'd let this thread consume an ARC
3851 * buffer before it should have (i.e. before we're under
3852 * the overflow limit and were signalled by the reclaim
3853 * thread). As long as that is a rare occurrence, it
3854 * shouldn't cause any harm.
3856 if (arc_is_overflowing()) {
3857 cv_signal(&arc_reclaim_thread_cv);
3858 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
3861 mutex_exit(&arc_reclaim_lock);
3864 if (type == ARC_BUFC_METADATA) {
3865 buf->b_data = zio_buf_alloc(size);
3866 arc_space_consume(size, ARC_SPACE_META);
3868 ASSERT(type == ARC_BUFC_DATA);
3869 buf->b_data = zio_data_buf_alloc(size);
3870 arc_space_consume(size, ARC_SPACE_DATA);
3874 * Update the state size. Note that ghost states have a
3875 * "ghost size" and so don't need to be updated.
3877 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) {
3878 arc_buf_hdr_t *hdr = buf->b_hdr;
3879 arc_state_t *state = hdr->b_l1hdr.b_state;
3881 (void) refcount_add_many(&state->arcs_size, size, buf);
3884 * If this is reached via arc_read, the link is
3885 * protected by the hash lock. If reached via
3886 * arc_buf_alloc, the header should not be accessed by
3887 * any other thread. And, if reached via arc_read_done,
3888 * the hash lock will protect it if it's found in the
3889 * hash table; otherwise no other thread should be
3890 * trying to [add|remove]_reference it.
3892 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3893 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3894 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type],
3898 * If we are growing the cache, and we are adding anonymous
3899 * data, and we have outgrown arc_p, update arc_p
3901 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
3902 (refcount_count(&arc_anon->arcs_size) +
3903 refcount_count(&arc_mru->arcs_size) > arc_p))
3904 arc_p = MIN(arc_c, arc_p + size);
3906 ARCSTAT_BUMP(arcstat_allocated);
3910 * This routine is called whenever a buffer is accessed.
3911 * NOTE: the hash lock is dropped in this function.
3914 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3918 ASSERT(MUTEX_HELD(hash_lock));
3919 ASSERT(HDR_HAS_L1HDR(hdr));
3921 if (hdr->b_l1hdr.b_state == arc_anon) {
3923 * This buffer is not in the cache, and does not
3924 * appear in our "ghost" list. Add the new buffer
3928 ASSERT0(hdr->b_l1hdr.b_arc_access);
3929 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3930 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3931 arc_change_state(arc_mru, hdr, hash_lock);
3933 } else if (hdr->b_l1hdr.b_state == arc_mru) {
3934 now = ddi_get_lbolt();
3937 * If this buffer is here because of a prefetch, then either:
3938 * - clear the flag if this is a "referencing" read
3939 * (any subsequent access will bump this into the MFU state).
3941 * - move the buffer to the head of the list if this is
3942 * another prefetch (to make it less likely to be evicted).
3944 if (HDR_PREFETCH(hdr)) {
3945 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
3946 /* link protected by hash lock */
3947 ASSERT(multilist_link_active(
3948 &hdr->b_l1hdr.b_arc_node));
3950 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3951 ARCSTAT_BUMP(arcstat_mru_hits);
3953 hdr->b_l1hdr.b_arc_access = now;
3958 * This buffer has been "accessed" only once so far,
3959 * but it is still in the cache. Move it to the MFU
3962 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
3964 * More than 125ms have passed since we
3965 * instantiated this buffer. Move it to the
3966 * most frequently used state.
3968 hdr->b_l1hdr.b_arc_access = now;
3969 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3970 arc_change_state(arc_mfu, hdr, hash_lock);
3972 ARCSTAT_BUMP(arcstat_mru_hits);
3973 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
3974 arc_state_t *new_state;
3976 * This buffer has been "accessed" recently, but
3977 * was evicted from the cache. Move it to the
3981 if (HDR_PREFETCH(hdr)) {
3982 new_state = arc_mru;
3983 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
3984 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3985 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3987 new_state = arc_mfu;
3988 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3991 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3992 arc_change_state(new_state, hdr, hash_lock);
3994 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
3995 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
3997 * This buffer has been accessed more than once and is
3998 * still in the cache. Keep it in the MFU state.
4000 * NOTE: an add_reference() that occurred when we did
4001 * the arc_read() will have kicked this off the list.
4002 * If it was a prefetch, we will explicitly move it to
4003 * the head of the list now.
4005 if ((HDR_PREFETCH(hdr)) != 0) {
4006 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4007 /* link protected by hash_lock */
4008 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4010 ARCSTAT_BUMP(arcstat_mfu_hits);
4011 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4012 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4013 arc_state_t *new_state = arc_mfu;
4015 * This buffer has been accessed more than once but has
4016 * been evicted from the cache. Move it back to the
4020 if (HDR_PREFETCH(hdr)) {
4022 * This is a prefetch access...
4023 * move this block back to the MRU state.
4025 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4026 new_state = arc_mru;
4029 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4030 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4031 arc_change_state(new_state, hdr, hash_lock);
4033 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4034 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4036 * This buffer is on the 2nd Level ARC.
4039 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4040 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4041 arc_change_state(arc_mfu, hdr, hash_lock);
4043 ASSERT(!"invalid arc state");
4047 /* a generic arc_done_func_t which you can use */
4050 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4052 if (zio == NULL || zio->io_error == 0)
4053 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
4054 VERIFY(arc_buf_remove_ref(buf, arg));
4057 /* a generic arc_done_func_t */
4059 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4061 arc_buf_t **bufp = arg;
4062 if (zio && zio->io_error) {
4063 VERIFY(arc_buf_remove_ref(buf, arg));
4067 ASSERT(buf->b_data);
4072 arc_read_done(zio_t *zio)
4076 arc_buf_t *abuf; /* buffer we're assigning to callback */
4077 kmutex_t *hash_lock = NULL;
4078 arc_callback_t *callback_list, *acb;
4079 int freeable = FALSE;
4081 buf = zio->io_private;
4085 * The hdr was inserted into hash-table and removed from lists
4086 * prior to starting I/O. We should find this header, since
4087 * it's in the hash table, and it should be legit since it's
4088 * not possible to evict it during the I/O. The only possible
4089 * reason for it not to be found is if we were freed during the
4092 if (HDR_IN_HASH_TABLE(hdr)) {
4093 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
4094 ASSERT3U(hdr->b_dva.dva_word[0], ==,
4095 BP_IDENTITY(zio->io_bp)->dva_word[0]);
4096 ASSERT3U(hdr->b_dva.dva_word[1], ==,
4097 BP_IDENTITY(zio->io_bp)->dva_word[1]);
4099 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
4102 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
4103 hash_lock == NULL) ||
4105 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
4106 (found == hdr && HDR_L2_READING(hdr)));
4109 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED;
4110 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
4111 hdr->b_flags &= ~ARC_FLAG_L2CACHE;
4113 /* byteswap if necessary */
4114 callback_list = hdr->b_l1hdr.b_acb;
4115 ASSERT(callback_list != NULL);
4116 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
4117 dmu_object_byteswap_t bswap =
4118 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
4119 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
4120 byteswap_uint64_array :
4121 dmu_ot_byteswap[bswap].ob_func;
4122 func(buf->b_data, hdr->b_size);
4125 arc_cksum_compute(buf, B_FALSE);
4128 #endif /* illumos */
4130 if (hash_lock && zio->io_error == 0 &&
4131 hdr->b_l1hdr.b_state == arc_anon) {
4133 * Only call arc_access on anonymous buffers. This is because
4134 * if we've issued an I/O for an evicted buffer, we've already
4135 * called arc_access (to prevent any simultaneous readers from
4136 * getting confused).
4138 arc_access(hdr, hash_lock);
4141 /* create copies of the data buffer for the callers */
4143 for (acb = callback_list; acb; acb = acb->acb_next) {
4144 if (acb->acb_done) {
4146 ARCSTAT_BUMP(arcstat_duplicate_reads);
4147 abuf = arc_buf_clone(buf);
4149 acb->acb_buf = abuf;
4153 hdr->b_l1hdr.b_acb = NULL;
4154 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4155 ASSERT(!HDR_BUF_AVAILABLE(hdr));
4157 ASSERT(buf->b_efunc == NULL);
4158 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
4159 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4162 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
4163 callback_list != NULL);
4165 if (zio->io_error != 0) {
4166 hdr->b_flags |= ARC_FLAG_IO_ERROR;
4167 if (hdr->b_l1hdr.b_state != arc_anon)
4168 arc_change_state(arc_anon, hdr, hash_lock);
4169 if (HDR_IN_HASH_TABLE(hdr))
4170 buf_hash_remove(hdr);
4171 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4175 * Broadcast before we drop the hash_lock to avoid the possibility
4176 * that the hdr (and hence the cv) might be freed before we get to
4177 * the cv_broadcast().
4179 cv_broadcast(&hdr->b_l1hdr.b_cv);
4181 if (hash_lock != NULL) {
4182 mutex_exit(hash_lock);
4185 * This block was freed while we waited for the read to
4186 * complete. It has been removed from the hash table and
4187 * moved to the anonymous state (so that it won't show up
4190 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
4191 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4194 /* execute each callback and free its structure */
4195 while ((acb = callback_list) != NULL) {
4197 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
4199 if (acb->acb_zio_dummy != NULL) {
4200 acb->acb_zio_dummy->io_error = zio->io_error;
4201 zio_nowait(acb->acb_zio_dummy);
4204 callback_list = acb->acb_next;
4205 kmem_free(acb, sizeof (arc_callback_t));
4209 arc_hdr_destroy(hdr);
4213 * "Read" the block at the specified DVA (in bp) via the
4214 * cache. If the block is found in the cache, invoke the provided
4215 * callback immediately and return. Note that the `zio' parameter
4216 * in the callback will be NULL in this case, since no IO was
4217 * required. If the block is not in the cache pass the read request
4218 * on to the spa with a substitute callback function, so that the
4219 * requested block will be added to the cache.
4221 * If a read request arrives for a block that has a read in-progress,
4222 * either wait for the in-progress read to complete (and return the
4223 * results); or, if this is a read with a "done" func, add a record
4224 * to the read to invoke the "done" func when the read completes,
4225 * and return; or just return.
4227 * arc_read_done() will invoke all the requested "done" functions
4228 * for readers of this block.
4231 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
4232 void *private, zio_priority_t priority, int zio_flags,
4233 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
4235 arc_buf_hdr_t *hdr = NULL;
4236 arc_buf_t *buf = NULL;
4237 kmutex_t *hash_lock = NULL;
4239 uint64_t guid = spa_load_guid(spa);
4241 ASSERT(!BP_IS_EMBEDDED(bp) ||
4242 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
4245 if (!BP_IS_EMBEDDED(bp)) {
4247 * Embedded BP's have no DVA and require no I/O to "read".
4248 * Create an anonymous arc buf to back it.
4250 hdr = buf_hash_find(guid, bp, &hash_lock);
4253 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) {
4255 *arc_flags |= ARC_FLAG_CACHED;
4257 if (HDR_IO_IN_PROGRESS(hdr)) {
4259 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
4260 priority == ZIO_PRIORITY_SYNC_READ) {
4262 * This sync read must wait for an
4263 * in-progress async read (e.g. a predictive
4264 * prefetch). Async reads are queued
4265 * separately at the vdev_queue layer, so
4266 * this is a form of priority inversion.
4267 * Ideally, we would "inherit" the demand
4268 * i/o's priority by moving the i/o from
4269 * the async queue to the synchronous queue,
4270 * but there is currently no mechanism to do
4271 * so. Track this so that we can evaluate
4272 * the magnitude of this potential performance
4275 * Note that if the prefetch i/o is already
4276 * active (has been issued to the device),
4277 * the prefetch improved performance, because
4278 * we issued it sooner than we would have
4279 * without the prefetch.
4281 DTRACE_PROBE1(arc__sync__wait__for__async,
4282 arc_buf_hdr_t *, hdr);
4283 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
4285 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4286 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH;
4289 if (*arc_flags & ARC_FLAG_WAIT) {
4290 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4291 mutex_exit(hash_lock);
4294 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4297 arc_callback_t *acb = NULL;
4299 acb = kmem_zalloc(sizeof (arc_callback_t),
4301 acb->acb_done = done;
4302 acb->acb_private = private;
4304 acb->acb_zio_dummy = zio_null(pio,
4305 spa, NULL, NULL, NULL, zio_flags);
4307 ASSERT(acb->acb_done != NULL);
4308 acb->acb_next = hdr->b_l1hdr.b_acb;
4309 hdr->b_l1hdr.b_acb = acb;
4310 add_reference(hdr, hash_lock, private);
4311 mutex_exit(hash_lock);
4314 mutex_exit(hash_lock);
4318 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4319 hdr->b_l1hdr.b_state == arc_mfu);
4322 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4324 * This is a demand read which does not have to
4325 * wait for i/o because we did a predictive
4326 * prefetch i/o for it, which has completed.
4329 arc__demand__hit__predictive__prefetch,
4330 arc_buf_hdr_t *, hdr);
4332 arcstat_demand_hit_predictive_prefetch);
4333 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH;
4335 add_reference(hdr, hash_lock, private);
4337 * If this block is already in use, create a new
4338 * copy of the data so that we will be guaranteed
4339 * that arc_release() will always succeed.
4341 buf = hdr->b_l1hdr.b_buf;
4343 ASSERT(buf->b_data);
4344 if (HDR_BUF_AVAILABLE(hdr)) {
4345 ASSERT(buf->b_efunc == NULL);
4346 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4348 buf = arc_buf_clone(buf);
4351 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4352 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4353 hdr->b_flags |= ARC_FLAG_PREFETCH;
4355 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4356 arc_access(hdr, hash_lock);
4357 if (*arc_flags & ARC_FLAG_L2CACHE)
4358 hdr->b_flags |= ARC_FLAG_L2CACHE;
4359 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4360 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4361 mutex_exit(hash_lock);
4362 ARCSTAT_BUMP(arcstat_hits);
4363 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4364 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4365 data, metadata, hits);
4368 done(NULL, buf, private);
4370 uint64_t size = BP_GET_LSIZE(bp);
4371 arc_callback_t *acb;
4374 boolean_t devw = B_FALSE;
4375 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
4376 int32_t b_asize = 0;
4379 /* this block is not in the cache */
4380 arc_buf_hdr_t *exists = NULL;
4381 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4382 buf = arc_buf_alloc(spa, size, private, type);
4384 if (!BP_IS_EMBEDDED(bp)) {
4385 hdr->b_dva = *BP_IDENTITY(bp);
4386 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4387 exists = buf_hash_insert(hdr, &hash_lock);
4389 if (exists != NULL) {
4390 /* somebody beat us to the hash insert */
4391 mutex_exit(hash_lock);
4392 buf_discard_identity(hdr);
4393 (void) arc_buf_remove_ref(buf, private);
4394 goto top; /* restart the IO request */
4398 * If there is a callback, we pass our reference to
4399 * it; otherwise we remove our reference.
4402 (void) remove_reference(hdr, hash_lock,
4405 if (*arc_flags & ARC_FLAG_PREFETCH)
4406 hdr->b_flags |= ARC_FLAG_PREFETCH;
4407 if (*arc_flags & ARC_FLAG_L2CACHE)
4408 hdr->b_flags |= ARC_FLAG_L2CACHE;
4409 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4410 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4411 if (BP_GET_LEVEL(bp) > 0)
4412 hdr->b_flags |= ARC_FLAG_INDIRECT;
4415 * This block is in the ghost cache. If it was L2-only
4416 * (and thus didn't have an L1 hdr), we realloc the
4417 * header to add an L1 hdr.
4419 if (!HDR_HAS_L1HDR(hdr)) {
4420 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4424 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4425 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4426 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4427 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4430 * If there is a callback, we pass a reference to it.
4433 add_reference(hdr, hash_lock, private);
4434 if (*arc_flags & ARC_FLAG_PREFETCH)
4435 hdr->b_flags |= ARC_FLAG_PREFETCH;
4436 if (*arc_flags & ARC_FLAG_L2CACHE)
4437 hdr->b_flags |= ARC_FLAG_L2CACHE;
4438 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4439 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4440 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
4443 buf->b_efunc = NULL;
4444 buf->b_private = NULL;
4446 hdr->b_l1hdr.b_buf = buf;
4447 ASSERT0(hdr->b_l1hdr.b_datacnt);
4448 hdr->b_l1hdr.b_datacnt = 1;
4449 arc_get_data_buf(buf);
4450 arc_access(hdr, hash_lock);
4453 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
4454 hdr->b_flags |= ARC_FLAG_PREDICTIVE_PREFETCH;
4455 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4457 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4458 acb->acb_done = done;
4459 acb->acb_private = private;
4461 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4462 hdr->b_l1hdr.b_acb = acb;
4463 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4465 if (HDR_HAS_L2HDR(hdr) &&
4466 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
4467 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
4468 addr = hdr->b_l2hdr.b_daddr;
4469 b_compress = hdr->b_l2hdr.b_compress;
4470 b_asize = hdr->b_l2hdr.b_asize;
4472 * Lock out device removal.
4474 if (vdev_is_dead(vd) ||
4475 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
4479 if (hash_lock != NULL)
4480 mutex_exit(hash_lock);
4483 * At this point, we have a level 1 cache miss. Try again in
4484 * L2ARC if possible.
4486 ASSERT3U(hdr->b_size, ==, size);
4487 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
4488 uint64_t, size, zbookmark_phys_t *, zb);
4489 ARCSTAT_BUMP(arcstat_misses);
4490 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4491 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4492 data, metadata, misses);
4494 curthread->td_ru.ru_inblock++;
4497 if (priority == ZIO_PRIORITY_ASYNC_READ)
4498 hdr->b_flags |= ARC_FLAG_PRIO_ASYNC_READ;
4500 hdr->b_flags &= ~ARC_FLAG_PRIO_ASYNC_READ;
4502 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
4504 * Read from the L2ARC if the following are true:
4505 * 1. The L2ARC vdev was previously cached.
4506 * 2. This buffer still has L2ARC metadata.
4507 * 3. This buffer isn't currently writing to the L2ARC.
4508 * 4. The L2ARC entry wasn't evicted, which may
4509 * also have invalidated the vdev.
4510 * 5. This isn't prefetch and l2arc_noprefetch is set.
4512 if (HDR_HAS_L2HDR(hdr) &&
4513 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
4514 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
4515 l2arc_read_callback_t *cb;
4517 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
4518 ARCSTAT_BUMP(arcstat_l2_hits);
4520 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
4522 cb->l2rcb_buf = buf;
4523 cb->l2rcb_spa = spa;
4526 cb->l2rcb_flags = zio_flags;
4527 cb->l2rcb_compress = b_compress;
4529 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
4530 addr + size < vd->vdev_psize -
4531 VDEV_LABEL_END_SIZE);
4534 * l2arc read. The SCL_L2ARC lock will be
4535 * released by l2arc_read_done().
4536 * Issue a null zio if the underlying buffer
4537 * was squashed to zero size by compression.
4539 if (b_compress == ZIO_COMPRESS_EMPTY) {
4540 rzio = zio_null(pio, spa, vd,
4541 l2arc_read_done, cb,
4542 zio_flags | ZIO_FLAG_DONT_CACHE |
4544 ZIO_FLAG_DONT_PROPAGATE |
4545 ZIO_FLAG_DONT_RETRY);
4547 rzio = zio_read_phys(pio, vd, addr,
4548 b_asize, buf->b_data,
4550 l2arc_read_done, cb, priority,
4551 zio_flags | ZIO_FLAG_DONT_CACHE |
4553 ZIO_FLAG_DONT_PROPAGATE |
4554 ZIO_FLAG_DONT_RETRY, B_FALSE);
4556 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
4558 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
4560 if (*arc_flags & ARC_FLAG_NOWAIT) {
4565 ASSERT(*arc_flags & ARC_FLAG_WAIT);
4566 if (zio_wait(rzio) == 0)
4569 /* l2arc read error; goto zio_read() */
4571 DTRACE_PROBE1(l2arc__miss,
4572 arc_buf_hdr_t *, hdr);
4573 ARCSTAT_BUMP(arcstat_l2_misses);
4574 if (HDR_L2_WRITING(hdr))
4575 ARCSTAT_BUMP(arcstat_l2_rw_clash);
4576 spa_config_exit(spa, SCL_L2ARC, vd);
4580 spa_config_exit(spa, SCL_L2ARC, vd);
4581 if (l2arc_ndev != 0) {
4582 DTRACE_PROBE1(l2arc__miss,
4583 arc_buf_hdr_t *, hdr);
4584 ARCSTAT_BUMP(arcstat_l2_misses);
4588 rzio = zio_read(pio, spa, bp, buf->b_data, size,
4589 arc_read_done, buf, priority, zio_flags, zb);
4591 if (*arc_flags & ARC_FLAG_WAIT)
4592 return (zio_wait(rzio));
4594 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4601 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
4603 ASSERT(buf->b_hdr != NULL);
4604 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon);
4605 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) ||
4607 ASSERT(buf->b_efunc == NULL);
4608 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
4610 buf->b_efunc = func;
4611 buf->b_private = private;
4615 * Notify the arc that a block was freed, and thus will never be used again.
4618 arc_freed(spa_t *spa, const blkptr_t *bp)
4621 kmutex_t *hash_lock;
4622 uint64_t guid = spa_load_guid(spa);
4624 ASSERT(!BP_IS_EMBEDDED(bp));
4626 hdr = buf_hash_find(guid, bp, &hash_lock);
4629 if (HDR_BUF_AVAILABLE(hdr)) {
4630 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
4631 add_reference(hdr, hash_lock, FTAG);
4632 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4633 mutex_exit(hash_lock);
4635 arc_release(buf, FTAG);
4636 (void) arc_buf_remove_ref(buf, FTAG);
4638 mutex_exit(hash_lock);
4644 * Clear the user eviction callback set by arc_set_callback(), first calling
4645 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4646 * clearing the callback may result in the arc_buf being destroyed. However,
4647 * it will not result in the *last* arc_buf being destroyed, hence the data
4648 * will remain cached in the ARC. We make a copy of the arc buffer here so
4649 * that we can process the callback without holding any locks.
4651 * It's possible that the callback is already in the process of being cleared
4652 * by another thread. In this case we can not clear the callback.
4654 * Returns B_TRUE if the callback was successfully called and cleared.
4657 arc_clear_callback(arc_buf_t *buf)
4660 kmutex_t *hash_lock;
4661 arc_evict_func_t *efunc = buf->b_efunc;
4662 void *private = buf->b_private;
4664 mutex_enter(&buf->b_evict_lock);
4668 * We are in arc_do_user_evicts().
4670 ASSERT(buf->b_data == NULL);
4671 mutex_exit(&buf->b_evict_lock);
4673 } else if (buf->b_data == NULL) {
4675 * We are on the eviction list; process this buffer now
4676 * but let arc_do_user_evicts() do the reaping.
4678 buf->b_efunc = NULL;
4679 mutex_exit(&buf->b_evict_lock);
4680 VERIFY0(efunc(private));
4683 hash_lock = HDR_LOCK(hdr);
4684 mutex_enter(hash_lock);
4686 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4688 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <,
4689 hdr->b_l1hdr.b_datacnt);
4690 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4691 hdr->b_l1hdr.b_state == arc_mfu);
4693 buf->b_efunc = NULL;
4694 buf->b_private = NULL;
4696 if (hdr->b_l1hdr.b_datacnt > 1) {
4697 mutex_exit(&buf->b_evict_lock);
4698 arc_buf_destroy(buf, TRUE);
4700 ASSERT(buf == hdr->b_l1hdr.b_buf);
4701 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4702 mutex_exit(&buf->b_evict_lock);
4705 mutex_exit(hash_lock);
4706 VERIFY0(efunc(private));
4711 * Release this buffer from the cache, making it an anonymous buffer. This
4712 * must be done after a read and prior to modifying the buffer contents.
4713 * If the buffer has more than one reference, we must make
4714 * a new hdr for the buffer.
4717 arc_release(arc_buf_t *buf, void *tag)
4719 arc_buf_hdr_t *hdr = buf->b_hdr;
4722 * It would be nice to assert that if it's DMU metadata (level >
4723 * 0 || it's the dnode file), then it must be syncing context.
4724 * But we don't know that information at this level.
4727 mutex_enter(&buf->b_evict_lock);
4729 ASSERT(HDR_HAS_L1HDR(hdr));
4732 * We don't grab the hash lock prior to this check, because if
4733 * the buffer's header is in the arc_anon state, it won't be
4734 * linked into the hash table.
4736 if (hdr->b_l1hdr.b_state == arc_anon) {
4737 mutex_exit(&buf->b_evict_lock);
4738 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4739 ASSERT(!HDR_IN_HASH_TABLE(hdr));
4740 ASSERT(!HDR_HAS_L2HDR(hdr));
4741 ASSERT(BUF_EMPTY(hdr));
4742 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1);
4743 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
4744 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
4746 ASSERT3P(buf->b_efunc, ==, NULL);
4747 ASSERT3P(buf->b_private, ==, NULL);
4749 hdr->b_l1hdr.b_arc_access = 0;
4755 kmutex_t *hash_lock = HDR_LOCK(hdr);
4756 mutex_enter(hash_lock);
4759 * This assignment is only valid as long as the hash_lock is
4760 * held, we must be careful not to reference state or the
4761 * b_state field after dropping the lock.
4763 arc_state_t *state = hdr->b_l1hdr.b_state;
4764 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4765 ASSERT3P(state, !=, arc_anon);
4767 /* this buffer is not on any list */
4768 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
4770 if (HDR_HAS_L2HDR(hdr)) {
4771 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4774 * We have to recheck this conditional again now that
4775 * we're holding the l2ad_mtx to prevent a race with
4776 * another thread which might be concurrently calling
4777 * l2arc_evict(). In that case, l2arc_evict() might have
4778 * destroyed the header's L2 portion as we were waiting
4779 * to acquire the l2ad_mtx.
4781 if (HDR_HAS_L2HDR(hdr)) {
4782 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET)
4783 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev,
4784 hdr->b_l2hdr.b_daddr,
4785 hdr->b_l2hdr.b_asize, 0);
4786 arc_hdr_l2hdr_destroy(hdr);
4789 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4793 * Do we have more than one buf?
4795 if (hdr->b_l1hdr.b_datacnt > 1) {
4796 arc_buf_hdr_t *nhdr;
4798 uint64_t blksz = hdr->b_size;
4799 uint64_t spa = hdr->b_spa;
4800 arc_buf_contents_t type = arc_buf_type(hdr);
4801 uint32_t flags = hdr->b_flags;
4803 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
4805 * Pull the data off of this hdr and attach it to
4806 * a new anonymous hdr.
4808 (void) remove_reference(hdr, hash_lock, tag);
4809 bufp = &hdr->b_l1hdr.b_buf;
4810 while (*bufp != buf)
4811 bufp = &(*bufp)->b_next;
4812 *bufp = buf->b_next;
4815 ASSERT3P(state, !=, arc_l2c_only);
4817 (void) refcount_remove_many(
4818 &state->arcs_size, hdr->b_size, buf);
4820 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
4821 ASSERT3P(state, !=, arc_l2c_only);
4822 uint64_t *size = &state->arcs_lsize[type];
4823 ASSERT3U(*size, >=, hdr->b_size);
4824 atomic_add_64(size, -hdr->b_size);
4828 * We're releasing a duplicate user data buffer, update
4829 * our statistics accordingly.
4831 if (HDR_ISTYPE_DATA(hdr)) {
4832 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
4833 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
4836 hdr->b_l1hdr.b_datacnt -= 1;
4837 arc_cksum_verify(buf);
4839 arc_buf_unwatch(buf);
4840 #endif /* illumos */
4842 mutex_exit(hash_lock);
4844 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
4845 nhdr->b_size = blksz;
4848 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING;
4849 nhdr->b_flags |= arc_bufc_to_flags(type);
4850 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
4852 nhdr->b_l1hdr.b_buf = buf;
4853 nhdr->b_l1hdr.b_datacnt = 1;
4854 nhdr->b_l1hdr.b_state = arc_anon;
4855 nhdr->b_l1hdr.b_arc_access = 0;
4856 nhdr->b_l1hdr.b_tmp_cdata = NULL;
4857 nhdr->b_freeze_cksum = NULL;
4859 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
4861 mutex_exit(&buf->b_evict_lock);
4862 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf);
4864 mutex_exit(&buf->b_evict_lock);
4865 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
4866 /* protected by hash lock, or hdr is on arc_anon */
4867 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4868 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4869 arc_change_state(arc_anon, hdr, hash_lock);
4870 hdr->b_l1hdr.b_arc_access = 0;
4871 mutex_exit(hash_lock);
4873 buf_discard_identity(hdr);
4876 buf->b_efunc = NULL;
4877 buf->b_private = NULL;
4881 arc_released(arc_buf_t *buf)
4885 mutex_enter(&buf->b_evict_lock);
4886 released = (buf->b_data != NULL &&
4887 buf->b_hdr->b_l1hdr.b_state == arc_anon);
4888 mutex_exit(&buf->b_evict_lock);
4894 arc_referenced(arc_buf_t *buf)
4898 mutex_enter(&buf->b_evict_lock);
4899 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
4900 mutex_exit(&buf->b_evict_lock);
4901 return (referenced);
4906 arc_write_ready(zio_t *zio)
4908 arc_write_callback_t *callback = zio->io_private;
4909 arc_buf_t *buf = callback->awcb_buf;
4910 arc_buf_hdr_t *hdr = buf->b_hdr;
4912 ASSERT(HDR_HAS_L1HDR(hdr));
4913 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
4914 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4915 callback->awcb_ready(zio, buf, callback->awcb_private);
4918 * If the IO is already in progress, then this is a re-write
4919 * attempt, so we need to thaw and re-compute the cksum.
4920 * It is the responsibility of the callback to handle the
4921 * accounting for any re-write attempt.
4923 if (HDR_IO_IN_PROGRESS(hdr)) {
4924 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
4925 if (hdr->b_freeze_cksum != NULL) {
4926 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
4927 hdr->b_freeze_cksum = NULL;
4929 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
4931 arc_cksum_compute(buf, B_FALSE);
4932 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4936 * The SPA calls this callback for each physical write that happens on behalf
4937 * of a logical write. See the comment in dbuf_write_physdone() for details.
4940 arc_write_physdone(zio_t *zio)
4942 arc_write_callback_t *cb = zio->io_private;
4943 if (cb->awcb_physdone != NULL)
4944 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
4948 arc_write_done(zio_t *zio)
4950 arc_write_callback_t *callback = zio->io_private;
4951 arc_buf_t *buf = callback->awcb_buf;
4952 arc_buf_hdr_t *hdr = buf->b_hdr;
4954 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4956 if (zio->io_error == 0) {
4957 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
4958 buf_discard_identity(hdr);
4960 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
4961 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
4964 ASSERT(BUF_EMPTY(hdr));
4968 * If the block to be written was all-zero or compressed enough to be
4969 * embedded in the BP, no write was performed so there will be no
4970 * dva/birth/checksum. The buffer must therefore remain anonymous
4973 if (!BUF_EMPTY(hdr)) {
4974 arc_buf_hdr_t *exists;
4975 kmutex_t *hash_lock;
4977 ASSERT(zio->io_error == 0);
4979 arc_cksum_verify(buf);
4981 exists = buf_hash_insert(hdr, &hash_lock);
4982 if (exists != NULL) {
4984 * This can only happen if we overwrite for
4985 * sync-to-convergence, because we remove
4986 * buffers from the hash table when we arc_free().
4988 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
4989 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4990 panic("bad overwrite, hdr=%p exists=%p",
4991 (void *)hdr, (void *)exists);
4992 ASSERT(refcount_is_zero(
4993 &exists->b_l1hdr.b_refcnt));
4994 arc_change_state(arc_anon, exists, hash_lock);
4995 mutex_exit(hash_lock);
4996 arc_hdr_destroy(exists);
4997 exists = buf_hash_insert(hdr, &hash_lock);
4998 ASSERT3P(exists, ==, NULL);
4999 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
5001 ASSERT(zio->io_prop.zp_nopwrite);
5002 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5003 panic("bad nopwrite, hdr=%p exists=%p",
5004 (void *)hdr, (void *)exists);
5007 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
5008 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
5009 ASSERT(BP_GET_DEDUP(zio->io_bp));
5010 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
5013 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
5014 /* if it's not anon, we are doing a scrub */
5015 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
5016 arc_access(hdr, hash_lock);
5017 mutex_exit(hash_lock);
5019 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
5022 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5023 callback->awcb_done(zio, buf, callback->awcb_private);
5025 kmem_free(callback, sizeof (arc_write_callback_t));
5029 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
5030 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
5031 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
5032 arc_done_func_t *done, void *private, zio_priority_t priority,
5033 int zio_flags, const zbookmark_phys_t *zb)
5035 arc_buf_hdr_t *hdr = buf->b_hdr;
5036 arc_write_callback_t *callback;
5039 ASSERT(ready != NULL);
5040 ASSERT(done != NULL);
5041 ASSERT(!HDR_IO_ERROR(hdr));
5042 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5043 ASSERT(hdr->b_l1hdr.b_acb == NULL);
5044 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
5046 hdr->b_flags |= ARC_FLAG_L2CACHE;
5048 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
5049 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
5050 callback->awcb_ready = ready;
5051 callback->awcb_physdone = physdone;
5052 callback->awcb_done = done;
5053 callback->awcb_private = private;
5054 callback->awcb_buf = buf;
5056 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
5057 arc_write_ready, arc_write_physdone, arc_write_done, callback,
5058 priority, zio_flags, zb);
5064 arc_memory_throttle(uint64_t reserve, uint64_t txg)
5067 uint64_t available_memory = ptob(freemem);
5068 static uint64_t page_load = 0;
5069 static uint64_t last_txg = 0;
5071 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
5073 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE)));
5076 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
5079 if (txg > last_txg) {
5084 * If we are in pageout, we know that memory is already tight,
5085 * the arc is already going to be evicting, so we just want to
5086 * continue to let page writes occur as quickly as possible.
5088 if (curproc == pageproc) {
5089 if (page_load > MAX(ptob(minfree), available_memory) / 4)
5090 return (SET_ERROR(ERESTART));
5091 /* Note: reserve is inflated, so we deflate */
5092 page_load += reserve / 8;
5094 } else if (page_load > 0 && arc_reclaim_needed()) {
5095 /* memory is low, delay before restarting */
5096 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
5097 return (SET_ERROR(EAGAIN));
5105 arc_tempreserve_clear(uint64_t reserve)
5107 atomic_add_64(&arc_tempreserve, -reserve);
5108 ASSERT((int64_t)arc_tempreserve >= 0);
5112 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
5117 if (reserve > arc_c/4 && !arc_no_grow) {
5118 arc_c = MIN(arc_c_max, reserve * 4);
5119 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
5121 if (reserve > arc_c)
5122 return (SET_ERROR(ENOMEM));
5125 * Don't count loaned bufs as in flight dirty data to prevent long
5126 * network delays from blocking transactions that are ready to be
5127 * assigned to a txg.
5129 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
5130 arc_loaned_bytes), 0);
5133 * Writes will, almost always, require additional memory allocations
5134 * in order to compress/encrypt/etc the data. We therefore need to
5135 * make sure that there is sufficient available memory for this.
5137 error = arc_memory_throttle(reserve, txg);
5142 * Throttle writes when the amount of dirty data in the cache
5143 * gets too large. We try to keep the cache less than half full
5144 * of dirty blocks so that our sync times don't grow too large.
5145 * Note: if two requests come in concurrently, we might let them
5146 * both succeed, when one of them should fail. Not a huge deal.
5149 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
5150 anon_size > arc_c / 4) {
5151 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5152 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5153 arc_tempreserve>>10,
5154 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
5155 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
5156 reserve>>10, arc_c>>10);
5157 return (SET_ERROR(ERESTART));
5159 atomic_add_64(&arc_tempreserve, reserve);
5164 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
5165 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
5167 size->value.ui64 = refcount_count(&state->arcs_size);
5168 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
5169 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
5173 arc_kstat_update(kstat_t *ksp, int rw)
5175 arc_stats_t *as = ksp->ks_data;
5177 if (rw == KSTAT_WRITE) {
5180 arc_kstat_update_state(arc_anon,
5181 &as->arcstat_anon_size,
5182 &as->arcstat_anon_evictable_data,
5183 &as->arcstat_anon_evictable_metadata);
5184 arc_kstat_update_state(arc_mru,
5185 &as->arcstat_mru_size,
5186 &as->arcstat_mru_evictable_data,
5187 &as->arcstat_mru_evictable_metadata);
5188 arc_kstat_update_state(arc_mru_ghost,
5189 &as->arcstat_mru_ghost_size,
5190 &as->arcstat_mru_ghost_evictable_data,
5191 &as->arcstat_mru_ghost_evictable_metadata);
5192 arc_kstat_update_state(arc_mfu,
5193 &as->arcstat_mfu_size,
5194 &as->arcstat_mfu_evictable_data,
5195 &as->arcstat_mfu_evictable_metadata);
5196 arc_kstat_update_state(arc_mfu_ghost,
5197 &as->arcstat_mfu_ghost_size,
5198 &as->arcstat_mfu_ghost_evictable_data,
5199 &as->arcstat_mfu_ghost_evictable_metadata);
5206 * This function *must* return indices evenly distributed between all
5207 * sublists of the multilist. This is needed due to how the ARC eviction
5208 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5209 * distributed between all sublists and uses this assumption when
5210 * deciding which sublist to evict from and how much to evict from it.
5213 arc_state_multilist_index_func(multilist_t *ml, void *obj)
5215 arc_buf_hdr_t *hdr = obj;
5218 * We rely on b_dva to generate evenly distributed index
5219 * numbers using buf_hash below. So, as an added precaution,
5220 * let's make sure we never add empty buffers to the arc lists.
5222 ASSERT(!BUF_EMPTY(hdr));
5225 * The assumption here, is the hash value for a given
5226 * arc_buf_hdr_t will remain constant throughout it's lifetime
5227 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
5228 * Thus, we don't need to store the header's sublist index
5229 * on insertion, as this index can be recalculated on removal.
5231 * Also, the low order bits of the hash value are thought to be
5232 * distributed evenly. Otherwise, in the case that the multilist
5233 * has a power of two number of sublists, each sublists' usage
5234 * would not be evenly distributed.
5236 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
5237 multilist_get_num_sublists(ml));
5241 static eventhandler_tag arc_event_lowmem = NULL;
5244 arc_lowmem(void *arg __unused, int howto __unused)
5247 mutex_enter(&arc_reclaim_lock);
5248 /* XXX: Memory deficit should be passed as argument. */
5249 needfree = btoc(arc_c >> arc_shrink_shift);
5250 DTRACE_PROBE(arc__needfree);
5251 cv_signal(&arc_reclaim_thread_cv);
5254 * It is unsafe to block here in arbitrary threads, because we can come
5255 * here from ARC itself and may hold ARC locks and thus risk a deadlock
5256 * with ARC reclaim thread.
5258 if (curproc == pageproc)
5259 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5260 mutex_exit(&arc_reclaim_lock);
5267 int i, prefetch_tunable_set = 0;
5269 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
5270 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
5271 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
5273 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
5274 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL);
5276 /* Convert seconds to clock ticks */
5277 arc_min_prefetch_lifespan = 1 * hz;
5279 /* Start out with 1/8 of all memory */
5280 arc_c = kmem_size() / 8;
5285 * On architectures where the physical memory can be larger
5286 * than the addressable space (intel in 32-bit mode), we may
5287 * need to limit the cache to 1/8 of VM size.
5289 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
5292 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
5293 arc_c_min = MAX(arc_c / 4, 16 << 20);
5294 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
5295 if (arc_c * 8 >= 1 << 30)
5296 arc_c_max = (arc_c * 8) - (1 << 30);
5298 arc_c_max = arc_c_min;
5299 arc_c_max = MAX(arc_c * 5, arc_c_max);
5302 * In userland, there's only the memory pressure that we artificially
5303 * create (see arc_available_memory()). Don't let arc_c get too
5304 * small, because it can cause transactions to be larger than
5305 * arc_c, causing arc_tempreserve_space() to fail.
5308 arc_c_min = arc_c_max / 2;
5313 * Allow the tunables to override our calculations if they are
5314 * reasonable (ie. over 16MB)
5316 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size())
5317 arc_c_max = zfs_arc_max;
5318 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max)
5319 arc_c_min = zfs_arc_min;
5323 arc_p = (arc_c >> 1);
5325 /* limit meta-data to 1/4 of the arc capacity */
5326 arc_meta_limit = arc_c_max / 4;
5328 /* Allow the tunable to override if it is reasonable */
5329 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
5330 arc_meta_limit = zfs_arc_meta_limit;
5332 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
5333 arc_c_min = arc_meta_limit / 2;
5335 if (zfs_arc_meta_min > 0) {
5336 arc_meta_min = zfs_arc_meta_min;
5338 arc_meta_min = arc_c_min / 2;
5341 if (zfs_arc_grow_retry > 0)
5342 arc_grow_retry = zfs_arc_grow_retry;
5344 if (zfs_arc_shrink_shift > 0)
5345 arc_shrink_shift = zfs_arc_shrink_shift;
5348 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
5350 if (arc_no_grow_shift >= arc_shrink_shift)
5351 arc_no_grow_shift = arc_shrink_shift - 1;
5353 if (zfs_arc_p_min_shift > 0)
5354 arc_p_min_shift = zfs_arc_p_min_shift;
5356 if (zfs_arc_num_sublists_per_state < 1)
5357 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1);
5359 /* if kmem_flags are set, lets try to use less memory */
5360 if (kmem_debugging())
5362 if (arc_c < arc_c_min)
5365 zfs_arc_min = arc_c_min;
5366 zfs_arc_max = arc_c_max;
5368 arc_anon = &ARC_anon;
5370 arc_mru_ghost = &ARC_mru_ghost;
5372 arc_mfu_ghost = &ARC_mfu_ghost;
5373 arc_l2c_only = &ARC_l2c_only;
5376 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5377 sizeof (arc_buf_hdr_t),
5378 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5379 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5380 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5381 sizeof (arc_buf_hdr_t),
5382 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5383 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5384 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5385 sizeof (arc_buf_hdr_t),
5386 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5387 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5388 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5389 sizeof (arc_buf_hdr_t),
5390 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5391 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5392 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5393 sizeof (arc_buf_hdr_t),
5394 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5395 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5396 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5397 sizeof (arc_buf_hdr_t),
5398 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5399 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5400 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5401 sizeof (arc_buf_hdr_t),
5402 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5403 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5404 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5405 sizeof (arc_buf_hdr_t),
5406 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5407 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5408 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5409 sizeof (arc_buf_hdr_t),
5410 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5411 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5412 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
5413 sizeof (arc_buf_hdr_t),
5414 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5415 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5417 refcount_create(&arc_anon->arcs_size);
5418 refcount_create(&arc_mru->arcs_size);
5419 refcount_create(&arc_mru_ghost->arcs_size);
5420 refcount_create(&arc_mfu->arcs_size);
5421 refcount_create(&arc_mfu_ghost->arcs_size);
5422 refcount_create(&arc_l2c_only->arcs_size);
5426 arc_reclaim_thread_exit = FALSE;
5427 arc_user_evicts_thread_exit = FALSE;
5428 arc_eviction_list = NULL;
5429 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
5431 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
5432 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
5434 if (arc_ksp != NULL) {
5435 arc_ksp->ks_data = &arc_stats;
5436 arc_ksp->ks_update = arc_kstat_update;
5437 kstat_install(arc_ksp);
5440 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
5441 TS_RUN, minclsyspri);
5444 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
5445 EVENTHANDLER_PRI_FIRST);
5448 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0,
5449 TS_RUN, minclsyspri);
5455 * Calculate maximum amount of dirty data per pool.
5457 * If it has been set by /etc/system, take that.
5458 * Otherwise, use a percentage of physical memory defined by
5459 * zfs_dirty_data_max_percent (default 10%) with a cap at
5460 * zfs_dirty_data_max_max (default 4GB).
5462 if (zfs_dirty_data_max == 0) {
5463 zfs_dirty_data_max = ptob(physmem) *
5464 zfs_dirty_data_max_percent / 100;
5465 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
5466 zfs_dirty_data_max_max);
5470 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
5471 prefetch_tunable_set = 1;
5474 if (prefetch_tunable_set == 0) {
5475 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
5477 printf(" add \"vfs.zfs.prefetch_disable=0\" "
5478 "to /boot/loader.conf.\n");
5479 zfs_prefetch_disable = 1;
5482 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
5483 prefetch_tunable_set == 0) {
5484 printf("ZFS NOTICE: Prefetch is disabled by default if less "
5485 "than 4GB of RAM is present;\n"
5486 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
5487 "to /boot/loader.conf.\n");
5488 zfs_prefetch_disable = 1;
5491 /* Warn about ZFS memory and address space requirements. */
5492 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
5493 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
5494 "expect unstable behavior.\n");
5496 if (kmem_size() < 512 * (1 << 20)) {
5497 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
5498 "expect unstable behavior.\n");
5499 printf(" Consider tuning vm.kmem_size and "
5500 "vm.kmem_size_max\n");
5501 printf(" in /boot/loader.conf.\n");
5509 mutex_enter(&arc_reclaim_lock);
5510 arc_reclaim_thread_exit = TRUE;
5512 * The reclaim thread will set arc_reclaim_thread_exit back to
5513 * FALSE when it is finished exiting; we're waiting for that.
5515 while (arc_reclaim_thread_exit) {
5516 cv_signal(&arc_reclaim_thread_cv);
5517 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
5519 mutex_exit(&arc_reclaim_lock);
5521 mutex_enter(&arc_user_evicts_lock);
5522 arc_user_evicts_thread_exit = TRUE;
5524 * The user evicts thread will set arc_user_evicts_thread_exit
5525 * to FALSE when it is finished exiting; we're waiting for that.
5527 while (arc_user_evicts_thread_exit) {
5528 cv_signal(&arc_user_evicts_cv);
5529 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock);
5531 mutex_exit(&arc_user_evicts_lock);
5533 /* Use TRUE to ensure *all* buffers are evicted */
5534 arc_flush(NULL, TRUE);
5538 if (arc_ksp != NULL) {
5539 kstat_delete(arc_ksp);
5543 mutex_destroy(&arc_reclaim_lock);
5544 cv_destroy(&arc_reclaim_thread_cv);
5545 cv_destroy(&arc_reclaim_waiters_cv);
5547 mutex_destroy(&arc_user_evicts_lock);
5548 cv_destroy(&arc_user_evicts_cv);
5550 refcount_destroy(&arc_anon->arcs_size);
5551 refcount_destroy(&arc_mru->arcs_size);
5552 refcount_destroy(&arc_mru_ghost->arcs_size);
5553 refcount_destroy(&arc_mfu->arcs_size);
5554 refcount_destroy(&arc_mfu_ghost->arcs_size);
5555 refcount_destroy(&arc_l2c_only->arcs_size);
5557 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
5558 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5559 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
5560 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
5561 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
5562 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
5563 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
5564 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
5568 ASSERT0(arc_loaned_bytes);
5571 if (arc_event_lowmem != NULL)
5572 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
5579 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5580 * It uses dedicated storage devices to hold cached data, which are populated
5581 * using large infrequent writes. The main role of this cache is to boost
5582 * the performance of random read workloads. The intended L2ARC devices
5583 * include short-stroked disks, solid state disks, and other media with
5584 * substantially faster read latency than disk.
5586 * +-----------------------+
5588 * +-----------------------+
5591 * l2arc_feed_thread() arc_read()
5595 * +---------------+ |
5597 * +---------------+ |
5602 * +-------+ +-------+
5604 * | cache | | cache |
5605 * +-------+ +-------+
5606 * +=========+ .-----.
5607 * : L2ARC : |-_____-|
5608 * : devices : | Disks |
5609 * +=========+ `-_____-'
5611 * Read requests are satisfied from the following sources, in order:
5614 * 2) vdev cache of L2ARC devices
5616 * 4) vdev cache of disks
5619 * Some L2ARC device types exhibit extremely slow write performance.
5620 * To accommodate for this there are some significant differences between
5621 * the L2ARC and traditional cache design:
5623 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5624 * the ARC behave as usual, freeing buffers and placing headers on ghost
5625 * lists. The ARC does not send buffers to the L2ARC during eviction as
5626 * this would add inflated write latencies for all ARC memory pressure.
5628 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5629 * It does this by periodically scanning buffers from the eviction-end of
5630 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5631 * not already there. It scans until a headroom of buffers is satisfied,
5632 * which itself is a buffer for ARC eviction. If a compressible buffer is
5633 * found during scanning and selected for writing to an L2ARC device, we
5634 * temporarily boost scanning headroom during the next scan cycle to make
5635 * sure we adapt to compression effects (which might significantly reduce
5636 * the data volume we write to L2ARC). The thread that does this is
5637 * l2arc_feed_thread(), illustrated below; example sizes are included to
5638 * provide a better sense of ratio than this diagram:
5641 * +---------------------+----------+
5642 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5643 * +---------------------+----------+ | o L2ARC eligible
5644 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5645 * +---------------------+----------+ |
5646 * 15.9 Gbytes ^ 32 Mbytes |
5648 * l2arc_feed_thread()
5650 * l2arc write hand <--[oooo]--'
5654 * +==============================+
5655 * L2ARC dev |####|#|###|###| |####| ... |
5656 * +==============================+
5659 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5660 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5661 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5662 * safe to say that this is an uncommon case, since buffers at the end of
5663 * the ARC lists have moved there due to inactivity.
5665 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5666 * then the L2ARC simply misses copying some buffers. This serves as a
5667 * pressure valve to prevent heavy read workloads from both stalling the ARC
5668 * with waits and clogging the L2ARC with writes. This also helps prevent
5669 * the potential for the L2ARC to churn if it attempts to cache content too
5670 * quickly, such as during backups of the entire pool.
5672 * 5. After system boot and before the ARC has filled main memory, there are
5673 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5674 * lists can remain mostly static. Instead of searching from tail of these
5675 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5676 * for eligible buffers, greatly increasing its chance of finding them.
5678 * The L2ARC device write speed is also boosted during this time so that
5679 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5680 * there are no L2ARC reads, and no fear of degrading read performance
5681 * through increased writes.
5683 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5684 * the vdev queue can aggregate them into larger and fewer writes. Each
5685 * device is written to in a rotor fashion, sweeping writes through
5686 * available space then repeating.
5688 * 7. The L2ARC does not store dirty content. It never needs to flush
5689 * write buffers back to disk based storage.
5691 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5692 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5694 * The performance of the L2ARC can be tweaked by a number of tunables, which
5695 * may be necessary for different workloads:
5697 * l2arc_write_max max write bytes per interval
5698 * l2arc_write_boost extra write bytes during device warmup
5699 * l2arc_noprefetch skip caching prefetched buffers
5700 * l2arc_headroom number of max device writes to precache
5701 * l2arc_headroom_boost when we find compressed buffers during ARC
5702 * scanning, we multiply headroom by this
5703 * percentage factor for the next scan cycle,
5704 * since more compressed buffers are likely to
5706 * l2arc_feed_secs seconds between L2ARC writing
5708 * Tunables may be removed or added as future performance improvements are
5709 * integrated, and also may become zpool properties.
5711 * There are three key functions that control how the L2ARC warms up:
5713 * l2arc_write_eligible() check if a buffer is eligible to cache
5714 * l2arc_write_size() calculate how much to write
5715 * l2arc_write_interval() calculate sleep delay between writes
5717 * These three functions determine what to write, how much, and how quickly
5722 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
5725 * A buffer is *not* eligible for the L2ARC if it:
5726 * 1. belongs to a different spa.
5727 * 2. is already cached on the L2ARC.
5728 * 3. has an I/O in progress (it may be an incomplete read).
5729 * 4. is flagged not eligible (zfs property).
5731 if (hdr->b_spa != spa_guid) {
5732 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
5735 if (HDR_HAS_L2HDR(hdr)) {
5736 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
5739 if (HDR_IO_IN_PROGRESS(hdr)) {
5740 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
5743 if (!HDR_L2CACHE(hdr)) {
5744 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
5752 l2arc_write_size(void)
5757 * Make sure our globals have meaningful values in case the user
5760 size = l2arc_write_max;
5762 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
5763 "be greater than zero, resetting it to the default (%d)",
5765 size = l2arc_write_max = L2ARC_WRITE_SIZE;
5768 if (arc_warm == B_FALSE)
5769 size += l2arc_write_boost;
5776 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
5778 clock_t interval, next, now;
5781 * If the ARC lists are busy, increase our write rate; if the
5782 * lists are stale, idle back. This is achieved by checking
5783 * how much we previously wrote - if it was more than half of
5784 * what we wanted, schedule the next write much sooner.
5786 if (l2arc_feed_again && wrote > (wanted / 2))
5787 interval = (hz * l2arc_feed_min_ms) / 1000;
5789 interval = hz * l2arc_feed_secs;
5791 now = ddi_get_lbolt();
5792 next = MAX(now, MIN(now + interval, began + interval));
5798 * Cycle through L2ARC devices. This is how L2ARC load balances.
5799 * If a device is returned, this also returns holding the spa config lock.
5801 static l2arc_dev_t *
5802 l2arc_dev_get_next(void)
5804 l2arc_dev_t *first, *next = NULL;
5807 * Lock out the removal of spas (spa_namespace_lock), then removal
5808 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
5809 * both locks will be dropped and a spa config lock held instead.
5811 mutex_enter(&spa_namespace_lock);
5812 mutex_enter(&l2arc_dev_mtx);
5814 /* if there are no vdevs, there is nothing to do */
5815 if (l2arc_ndev == 0)
5819 next = l2arc_dev_last;
5821 /* loop around the list looking for a non-faulted vdev */
5823 next = list_head(l2arc_dev_list);
5825 next = list_next(l2arc_dev_list, next);
5827 next = list_head(l2arc_dev_list);
5830 /* if we have come back to the start, bail out */
5833 else if (next == first)
5836 } while (vdev_is_dead(next->l2ad_vdev));
5838 /* if we were unable to find any usable vdevs, return NULL */
5839 if (vdev_is_dead(next->l2ad_vdev))
5842 l2arc_dev_last = next;
5845 mutex_exit(&l2arc_dev_mtx);
5848 * Grab the config lock to prevent the 'next' device from being
5849 * removed while we are writing to it.
5852 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
5853 mutex_exit(&spa_namespace_lock);
5859 * Free buffers that were tagged for destruction.
5862 l2arc_do_free_on_write()
5865 l2arc_data_free_t *df, *df_prev;
5867 mutex_enter(&l2arc_free_on_write_mtx);
5868 buflist = l2arc_free_on_write;
5870 for (df = list_tail(buflist); df; df = df_prev) {
5871 df_prev = list_prev(buflist, df);
5872 ASSERT(df->l2df_data != NULL);
5873 ASSERT(df->l2df_func != NULL);
5874 df->l2df_func(df->l2df_data, df->l2df_size);
5875 list_remove(buflist, df);
5876 kmem_free(df, sizeof (l2arc_data_free_t));
5879 mutex_exit(&l2arc_free_on_write_mtx);
5883 * A write to a cache device has completed. Update all headers to allow
5884 * reads from these buffers to begin.
5887 l2arc_write_done(zio_t *zio)
5889 l2arc_write_callback_t *cb;
5892 arc_buf_hdr_t *head, *hdr, *hdr_prev;
5893 kmutex_t *hash_lock;
5894 int64_t bytes_dropped = 0;
5896 cb = zio->io_private;
5898 dev = cb->l2wcb_dev;
5899 ASSERT(dev != NULL);
5900 head = cb->l2wcb_head;
5901 ASSERT(head != NULL);
5902 buflist = &dev->l2ad_buflist;
5903 ASSERT(buflist != NULL);
5904 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
5905 l2arc_write_callback_t *, cb);
5907 if (zio->io_error != 0)
5908 ARCSTAT_BUMP(arcstat_l2_writes_error);
5911 * All writes completed, or an error was hit.
5914 mutex_enter(&dev->l2ad_mtx);
5915 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
5916 hdr_prev = list_prev(buflist, hdr);
5918 hash_lock = HDR_LOCK(hdr);
5921 * We cannot use mutex_enter or else we can deadlock
5922 * with l2arc_write_buffers (due to swapping the order
5923 * the hash lock and l2ad_mtx are taken).
5925 if (!mutex_tryenter(hash_lock)) {
5927 * Missed the hash lock. We must retry so we
5928 * don't leave the ARC_FLAG_L2_WRITING bit set.
5930 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
5933 * We don't want to rescan the headers we've
5934 * already marked as having been written out, so
5935 * we reinsert the head node so we can pick up
5936 * where we left off.
5938 list_remove(buflist, head);
5939 list_insert_after(buflist, hdr, head);
5941 mutex_exit(&dev->l2ad_mtx);
5944 * We wait for the hash lock to become available
5945 * to try and prevent busy waiting, and increase
5946 * the chance we'll be able to acquire the lock
5947 * the next time around.
5949 mutex_enter(hash_lock);
5950 mutex_exit(hash_lock);
5955 * We could not have been moved into the arc_l2c_only
5956 * state while in-flight due to our ARC_FLAG_L2_WRITING
5957 * bit being set. Let's just ensure that's being enforced.
5959 ASSERT(HDR_HAS_L1HDR(hdr));
5962 * We may have allocated a buffer for L2ARC compression,
5963 * we must release it to avoid leaking this data.
5965 l2arc_release_cdata_buf(hdr);
5967 if (zio->io_error != 0) {
5969 * Error - drop L2ARC entry.
5971 list_remove(buflist, hdr);
5972 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev,
5973 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0);
5974 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
5976 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize);
5977 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
5979 bytes_dropped += hdr->b_l2hdr.b_asize;
5980 (void) refcount_remove_many(&dev->l2ad_alloc,
5981 hdr->b_l2hdr.b_asize, hdr);
5985 * Allow ARC to begin reads and ghost list evictions to
5988 hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
5990 mutex_exit(hash_lock);
5993 atomic_inc_64(&l2arc_writes_done);
5994 list_remove(buflist, head);
5995 ASSERT(!HDR_HAS_L1HDR(head));
5996 kmem_cache_free(hdr_l2only_cache, head);
5997 mutex_exit(&dev->l2ad_mtx);
5999 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
6001 l2arc_do_free_on_write();
6003 kmem_free(cb, sizeof (l2arc_write_callback_t));
6007 * A read to a cache device completed. Validate buffer contents before
6008 * handing over to the regular ARC routines.
6011 l2arc_read_done(zio_t *zio)
6013 l2arc_read_callback_t *cb;
6016 kmutex_t *hash_lock;
6019 ASSERT(zio->io_vd != NULL);
6020 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
6022 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
6024 cb = zio->io_private;
6026 buf = cb->l2rcb_buf;
6027 ASSERT(buf != NULL);
6029 hash_lock = HDR_LOCK(buf->b_hdr);
6030 mutex_enter(hash_lock);
6032 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6035 * If the buffer was compressed, decompress it first.
6037 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
6038 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
6039 ASSERT(zio->io_data != NULL);
6040 ASSERT3U(zio->io_size, ==, hdr->b_size);
6041 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size);
6044 * Check this survived the L2ARC journey.
6046 equal = arc_cksum_equal(buf);
6047 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
6048 mutex_exit(hash_lock);
6049 zio->io_private = buf;
6050 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
6051 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
6054 mutex_exit(hash_lock);
6056 * Buffer didn't survive caching. Increment stats and
6057 * reissue to the original storage device.
6059 if (zio->io_error != 0) {
6060 ARCSTAT_BUMP(arcstat_l2_io_error);
6062 zio->io_error = SET_ERROR(EIO);
6065 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
6068 * If there's no waiter, issue an async i/o to the primary
6069 * storage now. If there *is* a waiter, the caller must
6070 * issue the i/o in a context where it's OK to block.
6072 if (zio->io_waiter == NULL) {
6073 zio_t *pio = zio_unique_parent(zio);
6075 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
6077 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
6078 buf->b_data, hdr->b_size, arc_read_done, buf,
6079 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
6083 kmem_free(cb, sizeof (l2arc_read_callback_t));
6087 * This is the list priority from which the L2ARC will search for pages to
6088 * cache. This is used within loops (0..3) to cycle through lists in the
6089 * desired order. This order can have a significant effect on cache
6092 * Currently the metadata lists are hit first, MFU then MRU, followed by
6093 * the data lists. This function returns a locked list, and also returns
6096 static multilist_sublist_t *
6097 l2arc_sublist_lock(int list_num)
6099 multilist_t *ml = NULL;
6102 ASSERT(list_num >= 0 && list_num <= 3);
6106 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
6109 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
6112 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
6115 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
6120 * Return a randomly-selected sublist. This is acceptable
6121 * because the caller feeds only a little bit of data for each
6122 * call (8MB). Subsequent calls will result in different
6123 * sublists being selected.
6125 idx = multilist_get_random_index(ml);
6126 return (multilist_sublist_lock(ml, idx));
6130 * Evict buffers from the device write hand to the distance specified in
6131 * bytes. This distance may span populated buffers, it may span nothing.
6132 * This is clearing a region on the L2ARC device ready for writing.
6133 * If the 'all' boolean is set, every buffer is evicted.
6136 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
6139 arc_buf_hdr_t *hdr, *hdr_prev;
6140 kmutex_t *hash_lock;
6143 buflist = &dev->l2ad_buflist;
6145 if (!all && dev->l2ad_first) {
6147 * This is the first sweep through the device. There is
6153 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
6155 * When nearing the end of the device, evict to the end
6156 * before the device write hand jumps to the start.
6158 taddr = dev->l2ad_end;
6160 taddr = dev->l2ad_hand + distance;
6162 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
6163 uint64_t, taddr, boolean_t, all);
6166 mutex_enter(&dev->l2ad_mtx);
6167 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
6168 hdr_prev = list_prev(buflist, hdr);
6170 hash_lock = HDR_LOCK(hdr);
6173 * We cannot use mutex_enter or else we can deadlock
6174 * with l2arc_write_buffers (due to swapping the order
6175 * the hash lock and l2ad_mtx are taken).
6177 if (!mutex_tryenter(hash_lock)) {
6179 * Missed the hash lock. Retry.
6181 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
6182 mutex_exit(&dev->l2ad_mtx);
6183 mutex_enter(hash_lock);
6184 mutex_exit(hash_lock);
6188 if (HDR_L2_WRITE_HEAD(hdr)) {
6190 * We hit a write head node. Leave it for
6191 * l2arc_write_done().
6193 list_remove(buflist, hdr);
6194 mutex_exit(hash_lock);
6198 if (!all && HDR_HAS_L2HDR(hdr) &&
6199 (hdr->b_l2hdr.b_daddr > taddr ||
6200 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
6202 * We've evicted to the target address,
6203 * or the end of the device.
6205 mutex_exit(hash_lock);
6209 ASSERT(HDR_HAS_L2HDR(hdr));
6210 if (!HDR_HAS_L1HDR(hdr)) {
6211 ASSERT(!HDR_L2_READING(hdr));
6213 * This doesn't exist in the ARC. Destroy.
6214 * arc_hdr_destroy() will call list_remove()
6215 * and decrement arcstat_l2_size.
6217 arc_change_state(arc_anon, hdr, hash_lock);
6218 arc_hdr_destroy(hdr);
6220 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
6221 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
6223 * Invalidate issued or about to be issued
6224 * reads, since we may be about to write
6225 * over this location.
6227 if (HDR_L2_READING(hdr)) {
6228 ARCSTAT_BUMP(arcstat_l2_evict_reading);
6229 hdr->b_flags |= ARC_FLAG_L2_EVICTED;
6232 /* Ensure this header has finished being written */
6233 ASSERT(!HDR_L2_WRITING(hdr));
6234 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
6236 arc_hdr_l2hdr_destroy(hdr);
6238 mutex_exit(hash_lock);
6240 mutex_exit(&dev->l2ad_mtx);
6244 * Find and write ARC buffers to the L2ARC device.
6246 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6247 * for reading until they have completed writing.
6248 * The headroom_boost is an in-out parameter used to maintain headroom boost
6249 * state between calls to this function.
6251 * Returns the number of bytes actually written (which may be smaller than
6252 * the delta by which the device hand has changed due to alignment).
6255 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
6256 boolean_t *headroom_boost)
6258 arc_buf_hdr_t *hdr, *hdr_prev, *head;
6259 uint64_t write_asize, write_sz, headroom,
6263 l2arc_write_callback_t *cb;
6265 uint64_t guid = spa_load_guid(spa);
6266 const boolean_t do_headroom_boost = *headroom_boost;
6269 ASSERT(dev->l2ad_vdev != NULL);
6271 /* Lower the flag now, we might want to raise it again later. */
6272 *headroom_boost = B_FALSE;
6275 write_sz = write_asize = 0;
6277 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
6278 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD;
6279 head->b_flags |= ARC_FLAG_HAS_L2HDR;
6281 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
6283 * We will want to try to compress buffers that are at least 2x the
6284 * device sector size.
6286 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
6289 * Copy buffers for L2ARC writing.
6291 for (try = 0; try <= 3; try++) {
6292 multilist_sublist_t *mls = l2arc_sublist_lock(try);
6293 uint64_t passed_sz = 0;
6295 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
6298 * L2ARC fast warmup.
6300 * Until the ARC is warm and starts to evict, read from the
6301 * head of the ARC lists rather than the tail.
6303 if (arc_warm == B_FALSE)
6304 hdr = multilist_sublist_head(mls);
6306 hdr = multilist_sublist_tail(mls);
6308 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
6310 headroom = target_sz * l2arc_headroom;
6311 if (do_headroom_boost)
6312 headroom = (headroom * l2arc_headroom_boost) / 100;
6314 for (; hdr; hdr = hdr_prev) {
6315 kmutex_t *hash_lock;
6319 if (arc_warm == B_FALSE)
6320 hdr_prev = multilist_sublist_next(mls, hdr);
6322 hdr_prev = multilist_sublist_prev(mls, hdr);
6323 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size);
6325 hash_lock = HDR_LOCK(hdr);
6326 if (!mutex_tryenter(hash_lock)) {
6327 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
6329 * Skip this buffer rather than waiting.
6334 passed_sz += hdr->b_size;
6335 if (passed_sz > headroom) {
6339 mutex_exit(hash_lock);
6340 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
6344 if (!l2arc_write_eligible(guid, hdr)) {
6345 mutex_exit(hash_lock);
6350 * Assume that the buffer is not going to be compressed
6351 * and could take more space on disk because of a larger
6354 buf_sz = hdr->b_size;
6355 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
6357 if ((write_asize + buf_a_sz) > target_sz) {
6359 mutex_exit(hash_lock);
6360 ARCSTAT_BUMP(arcstat_l2_write_full);
6366 * Insert a dummy header on the buflist so
6367 * l2arc_write_done() can find where the
6368 * write buffers begin without searching.
6370 mutex_enter(&dev->l2ad_mtx);
6371 list_insert_head(&dev->l2ad_buflist, head);
6372 mutex_exit(&dev->l2ad_mtx);
6375 sizeof (l2arc_write_callback_t), KM_SLEEP);
6376 cb->l2wcb_dev = dev;
6377 cb->l2wcb_head = head;
6378 pio = zio_root(spa, l2arc_write_done, cb,
6380 ARCSTAT_BUMP(arcstat_l2_write_pios);
6384 * Create and add a new L2ARC header.
6386 hdr->b_l2hdr.b_dev = dev;
6387 hdr->b_flags |= ARC_FLAG_L2_WRITING;
6389 * Temporarily stash the data buffer in b_tmp_cdata.
6390 * The subsequent write step will pick it up from
6391 * there. This is because can't access b_l1hdr.b_buf
6392 * without holding the hash_lock, which we in turn
6393 * can't access without holding the ARC list locks
6394 * (which we want to avoid during compression/writing).
6396 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF;
6397 hdr->b_l2hdr.b_asize = hdr->b_size;
6398 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data;
6401 * Explicitly set the b_daddr field to a known
6402 * value which means "invalid address". This
6403 * enables us to differentiate which stage of
6404 * l2arc_write_buffers() the particular header
6405 * is in (e.g. this loop, or the one below).
6406 * ARC_FLAG_L2_WRITING is not enough to make
6407 * this distinction, and we need to know in
6408 * order to do proper l2arc vdev accounting in
6409 * arc_release() and arc_hdr_destroy().
6411 * Note, we can't use a new flag to distinguish
6412 * the two stages because we don't hold the
6413 * header's hash_lock below, in the second stage
6414 * of this function. Thus, we can't simply
6415 * change the b_flags field to denote that the
6416 * IO has been sent. We can change the b_daddr
6417 * field of the L2 portion, though, since we'll
6418 * be holding the l2ad_mtx; which is why we're
6419 * using it to denote the header's state change.
6421 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET;
6423 hdr->b_flags |= ARC_FLAG_HAS_L2HDR;
6425 mutex_enter(&dev->l2ad_mtx);
6426 list_insert_head(&dev->l2ad_buflist, hdr);
6427 mutex_exit(&dev->l2ad_mtx);
6430 * Compute and store the buffer cksum before
6431 * writing. On debug the cksum is verified first.
6433 arc_cksum_verify(hdr->b_l1hdr.b_buf);
6434 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE);
6436 mutex_exit(hash_lock);
6439 write_asize += buf_a_sz;
6442 multilist_sublist_unlock(mls);
6448 /* No buffers selected for writing? */
6451 ASSERT(!HDR_HAS_L1HDR(head));
6452 kmem_cache_free(hdr_l2only_cache, head);
6456 mutex_enter(&dev->l2ad_mtx);
6459 * Note that elsewhere in this file arcstat_l2_asize
6460 * and the used space on l2ad_vdev are updated using b_asize,
6461 * which is not necessarily rounded up to the device block size.
6462 * Too keep accounting consistent we do the same here as well:
6463 * stats_size accumulates the sum of b_asize of the written buffers,
6464 * while write_asize accumulates the sum of b_asize rounded up
6465 * to the device block size.
6466 * The latter sum is used only to validate the corectness of the code.
6468 uint64_t stats_size = 0;
6472 * Now start writing the buffers. We're starting at the write head
6473 * and work backwards, retracing the course of the buffer selector
6476 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr;
6477 hdr = list_prev(&dev->l2ad_buflist, hdr)) {
6481 * We rely on the L1 portion of the header below, so
6482 * it's invalid for this header to have been evicted out
6483 * of the ghost cache, prior to being written out. The
6484 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6486 ASSERT(HDR_HAS_L1HDR(hdr));
6489 * We shouldn't need to lock the buffer here, since we flagged
6490 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6491 * take care to only access its L2 cache parameters. In
6492 * particular, hdr->l1hdr.b_buf may be invalid by now due to
6495 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
6497 if ((HDR_L2COMPRESS(hdr)) &&
6498 hdr->b_l2hdr.b_asize >= buf_compress_minsz) {
6499 if (l2arc_compress_buf(hdr)) {
6501 * If compression succeeded, enable headroom
6502 * boost on the next scan cycle.
6504 *headroom_boost = B_TRUE;
6509 * Pick up the buffer data we had previously stashed away
6510 * (and now potentially also compressed).
6512 buf_data = hdr->b_l1hdr.b_tmp_cdata;
6513 buf_sz = hdr->b_l2hdr.b_asize;
6516 * We need to do this regardless if buf_sz is zero or
6517 * not, otherwise, when this l2hdr is evicted we'll
6518 * remove a reference that was never added.
6520 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr);
6522 /* Compression may have squashed the buffer to zero length. */
6526 wzio = zio_write_phys(pio, dev->l2ad_vdev,
6527 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
6528 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
6529 ZIO_FLAG_CANFAIL, B_FALSE);
6531 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
6533 (void) zio_nowait(wzio);
6535 stats_size += buf_sz;
6538 * Keep the clock hand suitably device-aligned.
6540 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
6541 write_asize += buf_a_sz;
6542 dev->l2ad_hand += buf_a_sz;
6546 mutex_exit(&dev->l2ad_mtx);
6548 ASSERT3U(write_asize, <=, target_sz);
6549 ARCSTAT_BUMP(arcstat_l2_writes_sent);
6550 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
6551 ARCSTAT_INCR(arcstat_l2_size, write_sz);
6552 ARCSTAT_INCR(arcstat_l2_asize, stats_size);
6553 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0);
6556 * Bump device hand to the device start if it is approaching the end.
6557 * l2arc_evict() will already have evicted ahead for this case.
6559 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
6560 dev->l2ad_hand = dev->l2ad_start;
6561 dev->l2ad_first = B_FALSE;
6564 dev->l2ad_writing = B_TRUE;
6565 (void) zio_wait(pio);
6566 dev->l2ad_writing = B_FALSE;
6568 return (write_asize);
6572 * Compresses an L2ARC buffer.
6573 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6574 * size in l2hdr->b_asize. This routine tries to compress the data and
6575 * depending on the compression result there are three possible outcomes:
6576 * *) The buffer was incompressible. The original l2hdr contents were left
6577 * untouched and are ready for writing to an L2 device.
6578 * *) The buffer was all-zeros, so there is no need to write it to an L2
6579 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6580 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6581 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6582 * data buffer which holds the compressed data to be written, and b_asize
6583 * tells us how much data there is. b_compress is set to the appropriate
6584 * compression algorithm. Once writing is done, invoke
6585 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6587 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6588 * buffer was incompressible).
6591 l2arc_compress_buf(arc_buf_hdr_t *hdr)
6594 size_t csize, len, rounded;
6595 ASSERT(HDR_HAS_L2HDR(hdr));
6596 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
6598 ASSERT(HDR_HAS_L1HDR(hdr));
6599 ASSERT3S(l2hdr->b_compress, ==, ZIO_COMPRESS_OFF);
6600 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6602 len = l2hdr->b_asize;
6603 cdata = zio_data_buf_alloc(len);
6604 ASSERT3P(cdata, !=, NULL);
6605 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata,
6606 cdata, l2hdr->b_asize);
6609 /* zero block, indicate that there's nothing to write */
6610 zio_data_buf_free(cdata, len);
6611 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
6613 hdr->b_l1hdr.b_tmp_cdata = NULL;
6614 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
6618 rounded = P2ROUNDUP(csize,
6619 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift);
6620 if (rounded < len) {
6622 * Compression succeeded, we'll keep the cdata around for
6623 * writing and release it afterwards.
6625 if (rounded > csize) {
6626 bzero((char *)cdata + csize, rounded - csize);
6629 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
6630 l2hdr->b_asize = csize;
6631 hdr->b_l1hdr.b_tmp_cdata = cdata;
6632 ARCSTAT_BUMP(arcstat_l2_compress_successes);
6636 * Compression failed, release the compressed buffer.
6637 * l2hdr will be left unmodified.
6639 zio_data_buf_free(cdata, len);
6640 ARCSTAT_BUMP(arcstat_l2_compress_failures);
6646 * Decompresses a zio read back from an l2arc device. On success, the
6647 * underlying zio's io_data buffer is overwritten by the uncompressed
6648 * version. On decompression error (corrupt compressed stream), the
6649 * zio->io_error value is set to signal an I/O error.
6651 * Please note that the compressed data stream is not checksummed, so
6652 * if the underlying device is experiencing data corruption, we may feed
6653 * corrupt data to the decompressor, so the decompressor needs to be
6654 * able to handle this situation (LZ4 does).
6657 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
6659 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
6661 if (zio->io_error != 0) {
6663 * An io error has occured, just restore the original io
6664 * size in preparation for a main pool read.
6666 zio->io_orig_size = zio->io_size = hdr->b_size;
6670 if (c == ZIO_COMPRESS_EMPTY) {
6672 * An empty buffer results in a null zio, which means we
6673 * need to fill its io_data after we're done restoring the
6674 * buffer's contents.
6676 ASSERT(hdr->b_l1hdr.b_buf != NULL);
6677 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size);
6678 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data;
6680 ASSERT(zio->io_data != NULL);
6682 * We copy the compressed data from the start of the arc buffer
6683 * (the zio_read will have pulled in only what we need, the
6684 * rest is garbage which we will overwrite at decompression)
6685 * and then decompress back to the ARC data buffer. This way we
6686 * can minimize copying by simply decompressing back over the
6687 * original compressed data (rather than decompressing to an
6688 * aux buffer and then copying back the uncompressed buffer,
6689 * which is likely to be much larger).
6694 csize = zio->io_size;
6695 cdata = zio_data_buf_alloc(csize);
6696 bcopy(zio->io_data, cdata, csize);
6697 if (zio_decompress_data(c, cdata, zio->io_data, csize,
6699 zio->io_error = EIO;
6700 zio_data_buf_free(cdata, csize);
6703 /* Restore the expected uncompressed IO size. */
6704 zio->io_orig_size = zio->io_size = hdr->b_size;
6708 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6709 * This buffer serves as a temporary holder of compressed data while
6710 * the buffer entry is being written to an l2arc device. Once that is
6711 * done, we can dispose of it.
6714 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr)
6716 ASSERT(HDR_HAS_L2HDR(hdr));
6717 enum zio_compress comp = hdr->b_l2hdr.b_compress;
6719 ASSERT(HDR_HAS_L1HDR(hdr));
6720 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp));
6722 if (comp == ZIO_COMPRESS_OFF) {
6724 * In this case, b_tmp_cdata points to the same buffer
6725 * as the arc_buf_t's b_data field. We don't want to
6726 * free it, since the arc_buf_t will handle that.
6728 hdr->b_l1hdr.b_tmp_cdata = NULL;
6729 } else if (comp == ZIO_COMPRESS_EMPTY) {
6731 * In this case, b_tmp_cdata was compressed to an empty
6732 * buffer, thus there's nothing to free and b_tmp_cdata
6733 * should have been set to NULL in l2arc_write_buffers().
6735 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
6738 * If the data was compressed, then we've allocated a
6739 * temporary buffer for it, so now we need to release it.
6741 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6742 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata,
6744 hdr->b_l1hdr.b_tmp_cdata = NULL;
6750 * This thread feeds the L2ARC at regular intervals. This is the beating
6751 * heart of the L2ARC.
6754 l2arc_feed_thread(void *dummy __unused)
6759 uint64_t size, wrote;
6760 clock_t begin, next = ddi_get_lbolt();
6761 boolean_t headroom_boost = B_FALSE;
6763 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
6765 mutex_enter(&l2arc_feed_thr_lock);
6767 while (l2arc_thread_exit == 0) {
6768 CALLB_CPR_SAFE_BEGIN(&cpr);
6769 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
6770 next - ddi_get_lbolt());
6771 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
6772 next = ddi_get_lbolt() + hz;
6775 * Quick check for L2ARC devices.
6777 mutex_enter(&l2arc_dev_mtx);
6778 if (l2arc_ndev == 0) {
6779 mutex_exit(&l2arc_dev_mtx);
6782 mutex_exit(&l2arc_dev_mtx);
6783 begin = ddi_get_lbolt();
6786 * This selects the next l2arc device to write to, and in
6787 * doing so the next spa to feed from: dev->l2ad_spa. This
6788 * will return NULL if there are now no l2arc devices or if
6789 * they are all faulted.
6791 * If a device is returned, its spa's config lock is also
6792 * held to prevent device removal. l2arc_dev_get_next()
6793 * will grab and release l2arc_dev_mtx.
6795 if ((dev = l2arc_dev_get_next()) == NULL)
6798 spa = dev->l2ad_spa;
6799 ASSERT(spa != NULL);
6802 * If the pool is read-only then force the feed thread to
6803 * sleep a little longer.
6805 if (!spa_writeable(spa)) {
6806 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
6807 spa_config_exit(spa, SCL_L2ARC, dev);
6812 * Avoid contributing to memory pressure.
6814 if (arc_reclaim_needed()) {
6815 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
6816 spa_config_exit(spa, SCL_L2ARC, dev);
6820 ARCSTAT_BUMP(arcstat_l2_feeds);
6822 size = l2arc_write_size();
6825 * Evict L2ARC buffers that will be overwritten.
6827 l2arc_evict(dev, size, B_FALSE);
6830 * Write ARC buffers.
6832 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
6835 * Calculate interval between writes.
6837 next = l2arc_write_interval(begin, size, wrote);
6838 spa_config_exit(spa, SCL_L2ARC, dev);
6841 l2arc_thread_exit = 0;
6842 cv_broadcast(&l2arc_feed_thr_cv);
6843 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
6848 l2arc_vdev_present(vdev_t *vd)
6852 mutex_enter(&l2arc_dev_mtx);
6853 for (dev = list_head(l2arc_dev_list); dev != NULL;
6854 dev = list_next(l2arc_dev_list, dev)) {
6855 if (dev->l2ad_vdev == vd)
6858 mutex_exit(&l2arc_dev_mtx);
6860 return (dev != NULL);
6864 * Add a vdev for use by the L2ARC. By this point the spa has already
6865 * validated the vdev and opened it.
6868 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
6870 l2arc_dev_t *adddev;
6872 ASSERT(!l2arc_vdev_present(vd));
6874 vdev_ashift_optimize(vd);
6877 * Create a new l2arc device entry.
6879 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
6880 adddev->l2ad_spa = spa;
6881 adddev->l2ad_vdev = vd;
6882 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
6883 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
6884 adddev->l2ad_hand = adddev->l2ad_start;
6885 adddev->l2ad_first = B_TRUE;
6886 adddev->l2ad_writing = B_FALSE;
6888 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
6890 * This is a list of all ARC buffers that are still valid on the
6893 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
6894 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
6896 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
6897 refcount_create(&adddev->l2ad_alloc);
6900 * Add device to global list
6902 mutex_enter(&l2arc_dev_mtx);
6903 list_insert_head(l2arc_dev_list, adddev);
6904 atomic_inc_64(&l2arc_ndev);
6905 mutex_exit(&l2arc_dev_mtx);
6909 * Remove a vdev from the L2ARC.
6912 l2arc_remove_vdev(vdev_t *vd)
6914 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
6917 * Find the device by vdev
6919 mutex_enter(&l2arc_dev_mtx);
6920 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
6921 nextdev = list_next(l2arc_dev_list, dev);
6922 if (vd == dev->l2ad_vdev) {
6927 ASSERT(remdev != NULL);
6930 * Remove device from global list
6932 list_remove(l2arc_dev_list, remdev);
6933 l2arc_dev_last = NULL; /* may have been invalidated */
6934 atomic_dec_64(&l2arc_ndev);
6935 mutex_exit(&l2arc_dev_mtx);
6938 * Clear all buflists and ARC references. L2ARC device flush.
6940 l2arc_evict(remdev, 0, B_TRUE);
6941 list_destroy(&remdev->l2ad_buflist);
6942 mutex_destroy(&remdev->l2ad_mtx);
6943 refcount_destroy(&remdev->l2ad_alloc);
6944 kmem_free(remdev, sizeof (l2arc_dev_t));
6950 l2arc_thread_exit = 0;
6952 l2arc_writes_sent = 0;
6953 l2arc_writes_done = 0;
6955 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
6956 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
6957 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
6958 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
6960 l2arc_dev_list = &L2ARC_dev_list;
6961 l2arc_free_on_write = &L2ARC_free_on_write;
6962 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
6963 offsetof(l2arc_dev_t, l2ad_node));
6964 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
6965 offsetof(l2arc_data_free_t, l2df_list_node));
6972 * This is called from dmu_fini(), which is called from spa_fini();
6973 * Because of this, we can assume that all l2arc devices have
6974 * already been removed when the pools themselves were removed.
6977 l2arc_do_free_on_write();
6979 mutex_destroy(&l2arc_feed_thr_lock);
6980 cv_destroy(&l2arc_feed_thr_cv);
6981 mutex_destroy(&l2arc_dev_mtx);
6982 mutex_destroy(&l2arc_free_on_write_mtx);
6984 list_destroy(l2arc_dev_list);
6985 list_destroy(l2arc_free_on_write);
6991 if (!(spa_mode_global & FWRITE))
6994 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
6995 TS_RUN, minclsyspri);
7001 if (!(spa_mode_global & FWRITE))
7004 mutex_enter(&l2arc_feed_thr_lock);
7005 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7006 l2arc_thread_exit = 1;
7007 while (l2arc_thread_exit != 0)
7008 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7009 mutex_exit(&l2arc_feed_thr_lock);