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, 2014 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2014 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 mutexs, rather they rely on the
86 * hash table mutexs for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexs).
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>
133 #include <sys/dnlc.h>
135 #include <sys/callb.h>
136 #include <sys/kstat.h>
137 #include <sys/trim_map.h>
138 #include <zfs_fletcher.h>
141 #include <vm/vm_pageout.h>
142 #include <machine/vmparam.h>
146 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
147 boolean_t arc_watch = B_FALSE;
152 static kmutex_t arc_reclaim_thr_lock;
153 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
154 static uint8_t arc_thread_exit;
156 uint_t arc_reduce_dnlc_percent = 3;
159 * The number of iterations through arc_evict_*() before we
160 * drop & reacquire the lock.
162 int arc_evict_iterations = 100;
164 /* number of seconds before growing cache again */
165 static int arc_grow_retry = 60;
167 /* shift of arc_c for calculating both min and max arc_p */
168 static int arc_p_min_shift = 4;
170 /* log2(fraction of arc to reclaim) */
171 static int arc_shrink_shift = 7;
174 * log2(fraction of ARC which must be free to allow growing).
175 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
176 * when reading a new block into the ARC, we will evict an equal-sized block
179 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
180 * we will still not allow it to grow.
182 int arc_no_grow_shift = 5;
186 * minimum lifespan of a prefetch block in clock ticks
187 * (initialized in arc_init())
189 static int arc_min_prefetch_lifespan;
192 * If this percent of memory is free, don't throttle.
194 int arc_lotsfree_percent = 10;
197 extern int zfs_prefetch_disable;
200 * The arc has filled available memory and has now warmed up.
202 static boolean_t arc_warm;
204 uint64_t zfs_arc_max;
205 uint64_t zfs_arc_min;
206 uint64_t zfs_arc_meta_limit = 0;
207 uint64_t zfs_arc_meta_min = 0;
208 int zfs_arc_grow_retry = 0;
209 int zfs_arc_shrink_shift = 0;
210 int zfs_arc_p_min_shift = 0;
211 int zfs_disable_dup_eviction = 0;
212 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
213 u_int zfs_arc_free_target = 0;
215 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
216 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
220 arc_free_target_init(void *unused __unused)
223 zfs_arc_free_target = vm_pageout_wakeup_thresh;
225 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
226 arc_free_target_init, NULL);
228 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
229 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
230 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
231 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
232 TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize);
233 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
234 SYSCTL_DECL(_vfs_zfs);
235 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
237 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
239 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
240 &zfs_arc_average_blocksize, 0,
241 "ARC average blocksize");
242 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
243 &arc_shrink_shift, 0,
244 "log2(fraction of arc to reclaim)");
247 * We don't have a tunable for arc_free_target due to the dependency on
248 * pagedaemon initialisation.
250 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
251 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
252 sysctl_vfs_zfs_arc_free_target, "IU",
253 "Desired number of free pages below which ARC triggers reclaim");
256 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
261 val = zfs_arc_free_target;
262 err = sysctl_handle_int(oidp, &val, 0, req);
263 if (err != 0 || req->newptr == NULL)
268 if (val > cnt.v_page_count)
271 zfs_arc_free_target = val;
277 * Must be declared here, before the definition of corresponding kstat
278 * macro which uses the same names will confuse the compiler.
280 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
281 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
282 sysctl_vfs_zfs_arc_meta_limit, "QU",
283 "ARC metadata limit");
287 * Note that buffers can be in one of 6 states:
288 * ARC_anon - anonymous (discussed below)
289 * ARC_mru - recently used, currently cached
290 * ARC_mru_ghost - recentely used, no longer in cache
291 * ARC_mfu - frequently used, currently cached
292 * ARC_mfu_ghost - frequently used, no longer in cache
293 * ARC_l2c_only - exists in L2ARC but not other states
294 * When there are no active references to the buffer, they are
295 * are linked onto a list in one of these arc states. These are
296 * the only buffers that can be evicted or deleted. Within each
297 * state there are multiple lists, one for meta-data and one for
298 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
299 * etc.) is tracked separately so that it can be managed more
300 * explicitly: favored over data, limited explicitly.
302 * Anonymous buffers are buffers that are not associated with
303 * a DVA. These are buffers that hold dirty block copies
304 * before they are written to stable storage. By definition,
305 * they are "ref'd" and are considered part of arc_mru
306 * that cannot be freed. Generally, they will aquire a DVA
307 * as they are written and migrate onto the arc_mru list.
309 * The ARC_l2c_only state is for buffers that are in the second
310 * level ARC but no longer in any of the ARC_m* lists. The second
311 * level ARC itself may also contain buffers that are in any of
312 * the ARC_m* states - meaning that a buffer can exist in two
313 * places. The reason for the ARC_l2c_only state is to keep the
314 * buffer header in the hash table, so that reads that hit the
315 * second level ARC benefit from these fast lookups.
318 #define ARCS_LOCK_PAD CACHE_LINE_SIZE
322 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))];
327 * must be power of two for mask use to work
330 #define ARC_BUFC_NUMDATALISTS 16
331 #define ARC_BUFC_NUMMETADATALISTS 16
332 #define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS)
334 typedef struct arc_state {
335 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
336 uint64_t arcs_size; /* total amount of data in this state */
337 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */
338 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE);
341 #define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock))
344 static arc_state_t ARC_anon;
345 static arc_state_t ARC_mru;
346 static arc_state_t ARC_mru_ghost;
347 static arc_state_t ARC_mfu;
348 static arc_state_t ARC_mfu_ghost;
349 static arc_state_t ARC_l2c_only;
351 typedef struct arc_stats {
352 kstat_named_t arcstat_hits;
353 kstat_named_t arcstat_misses;
354 kstat_named_t arcstat_demand_data_hits;
355 kstat_named_t arcstat_demand_data_misses;
356 kstat_named_t arcstat_demand_metadata_hits;
357 kstat_named_t arcstat_demand_metadata_misses;
358 kstat_named_t arcstat_prefetch_data_hits;
359 kstat_named_t arcstat_prefetch_data_misses;
360 kstat_named_t arcstat_prefetch_metadata_hits;
361 kstat_named_t arcstat_prefetch_metadata_misses;
362 kstat_named_t arcstat_mru_hits;
363 kstat_named_t arcstat_mru_ghost_hits;
364 kstat_named_t arcstat_mfu_hits;
365 kstat_named_t arcstat_mfu_ghost_hits;
366 kstat_named_t arcstat_allocated;
367 kstat_named_t arcstat_deleted;
368 kstat_named_t arcstat_stolen;
369 kstat_named_t arcstat_recycle_miss;
371 * Number of buffers that could not be evicted because the hash lock
372 * was held by another thread. The lock may not necessarily be held
373 * by something using the same buffer, since hash locks are shared
374 * by multiple buffers.
376 kstat_named_t arcstat_mutex_miss;
378 * Number of buffers skipped because they have I/O in progress, are
379 * indrect prefetch buffers that have not lived long enough, or are
380 * not from the spa we're trying to evict from.
382 kstat_named_t arcstat_evict_skip;
383 kstat_named_t arcstat_evict_l2_cached;
384 kstat_named_t arcstat_evict_l2_eligible;
385 kstat_named_t arcstat_evict_l2_ineligible;
386 kstat_named_t arcstat_hash_elements;
387 kstat_named_t arcstat_hash_elements_max;
388 kstat_named_t arcstat_hash_collisions;
389 kstat_named_t arcstat_hash_chains;
390 kstat_named_t arcstat_hash_chain_max;
391 kstat_named_t arcstat_p;
392 kstat_named_t arcstat_c;
393 kstat_named_t arcstat_c_min;
394 kstat_named_t arcstat_c_max;
395 kstat_named_t arcstat_size;
397 * Number of bytes consumed by internal ARC structures necessary
398 * for tracking purposes; these structures are not actually
399 * backed by ARC buffers. This includes arc_buf_hdr_t structures
400 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
401 * caches), and arc_buf_t structures (allocated via arc_buf_t
404 kstat_named_t arcstat_hdr_size;
406 * Number of bytes consumed by ARC buffers of type equal to
407 * ARC_BUFC_DATA. This is generally consumed by buffers backing
408 * on disk user data (e.g. plain file contents).
410 kstat_named_t arcstat_data_size;
412 * Number of bytes consumed by ARC buffers of type equal to
413 * ARC_BUFC_METADATA. This is generally consumed by buffers
414 * backing on disk data that is used for internal ZFS
415 * structures (e.g. ZAP, dnode, indirect blocks, etc).
417 kstat_named_t arcstat_metadata_size;
419 * Number of bytes consumed by various buffers and structures
420 * not actually backed with ARC buffers. This includes bonus
421 * buffers (allocated directly via zio_buf_* functions),
422 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
423 * cache), and dnode_t structures (allocated via dnode_t cache).
425 kstat_named_t arcstat_other_size;
427 * Total number of bytes consumed by ARC buffers residing in the
428 * arc_anon state. This includes *all* buffers in the arc_anon
429 * state; e.g. data, metadata, evictable, and unevictable buffers
430 * are all included in this value.
432 kstat_named_t arcstat_anon_size;
434 * Number of bytes consumed by ARC buffers that meet the
435 * following criteria: backing buffers of type ARC_BUFC_DATA,
436 * residing in the arc_anon state, and are eligible for eviction
437 * (e.g. have no outstanding holds on the buffer).
439 kstat_named_t arcstat_anon_evictable_data;
441 * Number of bytes consumed by ARC buffers that meet the
442 * following criteria: backing buffers of type ARC_BUFC_METADATA,
443 * residing in the arc_anon state, and are eligible for eviction
444 * (e.g. have no outstanding holds on the buffer).
446 kstat_named_t arcstat_anon_evictable_metadata;
448 * Total number of bytes consumed by ARC buffers residing in the
449 * arc_mru state. This includes *all* buffers in the arc_mru
450 * state; e.g. data, metadata, evictable, and unevictable buffers
451 * are all included in this value.
453 kstat_named_t arcstat_mru_size;
455 * Number of bytes consumed by ARC buffers that meet the
456 * following criteria: backing buffers of type ARC_BUFC_DATA,
457 * residing in the arc_mru state, and are eligible for eviction
458 * (e.g. have no outstanding holds on the buffer).
460 kstat_named_t arcstat_mru_evictable_data;
462 * Number of bytes consumed by ARC buffers that meet the
463 * following criteria: backing buffers of type ARC_BUFC_METADATA,
464 * residing in the arc_mru state, and are eligible for eviction
465 * (e.g. have no outstanding holds on the buffer).
467 kstat_named_t arcstat_mru_evictable_metadata;
469 * Total number of bytes that *would have been* consumed by ARC
470 * buffers in the arc_mru_ghost state. The key thing to note
471 * here, is the fact that this size doesn't actually indicate
472 * RAM consumption. The ghost lists only consist of headers and
473 * don't actually have ARC buffers linked off of these headers.
474 * Thus, *if* the headers had associated ARC buffers, these
475 * buffers *would have* consumed this number of bytes.
477 kstat_named_t arcstat_mru_ghost_size;
479 * Number of bytes that *would have been* consumed by ARC
480 * buffers that are eligible for eviction, of type
481 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
483 kstat_named_t arcstat_mru_ghost_evictable_data;
485 * Number of bytes that *would have been* consumed by ARC
486 * buffers that are eligible for eviction, of type
487 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
489 kstat_named_t arcstat_mru_ghost_evictable_metadata;
491 * Total number of bytes consumed by ARC buffers residing in the
492 * arc_mfu state. This includes *all* buffers in the arc_mfu
493 * state; e.g. data, metadata, evictable, and unevictable buffers
494 * are all included in this value.
496 kstat_named_t arcstat_mfu_size;
498 * Number of bytes consumed by ARC buffers that are eligible for
499 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
502 kstat_named_t arcstat_mfu_evictable_data;
504 * Number of bytes consumed by ARC buffers that are eligible for
505 * eviction, of type ARC_BUFC_METADATA, and reside in the
508 kstat_named_t arcstat_mfu_evictable_metadata;
510 * Total number of bytes that *would have been* consumed by ARC
511 * buffers in the arc_mfu_ghost state. See the comment above
512 * arcstat_mru_ghost_size for more details.
514 kstat_named_t arcstat_mfu_ghost_size;
516 * Number of bytes that *would have been* consumed by ARC
517 * buffers that are eligible for eviction, of type
518 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
520 kstat_named_t arcstat_mfu_ghost_evictable_data;
522 * Number of bytes that *would have been* consumed by ARC
523 * buffers that are eligible for eviction, of type
524 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
526 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
527 kstat_named_t arcstat_l2_hits;
528 kstat_named_t arcstat_l2_misses;
529 kstat_named_t arcstat_l2_feeds;
530 kstat_named_t arcstat_l2_rw_clash;
531 kstat_named_t arcstat_l2_read_bytes;
532 kstat_named_t arcstat_l2_write_bytes;
533 kstat_named_t arcstat_l2_writes_sent;
534 kstat_named_t arcstat_l2_writes_done;
535 kstat_named_t arcstat_l2_writes_error;
536 kstat_named_t arcstat_l2_writes_hdr_miss;
537 kstat_named_t arcstat_l2_evict_lock_retry;
538 kstat_named_t arcstat_l2_evict_reading;
539 kstat_named_t arcstat_l2_evict_l1cached;
540 kstat_named_t arcstat_l2_free_on_write;
541 kstat_named_t arcstat_l2_cdata_free_on_write;
542 kstat_named_t arcstat_l2_abort_lowmem;
543 kstat_named_t arcstat_l2_cksum_bad;
544 kstat_named_t arcstat_l2_io_error;
545 kstat_named_t arcstat_l2_size;
546 kstat_named_t arcstat_l2_asize;
547 kstat_named_t arcstat_l2_hdr_size;
548 kstat_named_t arcstat_l2_compress_successes;
549 kstat_named_t arcstat_l2_compress_zeros;
550 kstat_named_t arcstat_l2_compress_failures;
551 kstat_named_t arcstat_l2_write_trylock_fail;
552 kstat_named_t arcstat_l2_write_passed_headroom;
553 kstat_named_t arcstat_l2_write_spa_mismatch;
554 kstat_named_t arcstat_l2_write_in_l2;
555 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
556 kstat_named_t arcstat_l2_write_not_cacheable;
557 kstat_named_t arcstat_l2_write_full;
558 kstat_named_t arcstat_l2_write_buffer_iter;
559 kstat_named_t arcstat_l2_write_pios;
560 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
561 kstat_named_t arcstat_l2_write_buffer_list_iter;
562 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
563 kstat_named_t arcstat_memory_throttle_count;
564 kstat_named_t arcstat_duplicate_buffers;
565 kstat_named_t arcstat_duplicate_buffers_size;
566 kstat_named_t arcstat_duplicate_reads;
567 kstat_named_t arcstat_meta_used;
568 kstat_named_t arcstat_meta_limit;
569 kstat_named_t arcstat_meta_max;
570 kstat_named_t arcstat_meta_min;
573 static arc_stats_t arc_stats = {
574 { "hits", KSTAT_DATA_UINT64 },
575 { "misses", KSTAT_DATA_UINT64 },
576 { "demand_data_hits", KSTAT_DATA_UINT64 },
577 { "demand_data_misses", KSTAT_DATA_UINT64 },
578 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
579 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
580 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
581 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
582 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
583 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
584 { "mru_hits", KSTAT_DATA_UINT64 },
585 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
586 { "mfu_hits", KSTAT_DATA_UINT64 },
587 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
588 { "allocated", KSTAT_DATA_UINT64 },
589 { "deleted", KSTAT_DATA_UINT64 },
590 { "stolen", KSTAT_DATA_UINT64 },
591 { "recycle_miss", KSTAT_DATA_UINT64 },
592 { "mutex_miss", KSTAT_DATA_UINT64 },
593 { "evict_skip", KSTAT_DATA_UINT64 },
594 { "evict_l2_cached", KSTAT_DATA_UINT64 },
595 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
596 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
597 { "hash_elements", KSTAT_DATA_UINT64 },
598 { "hash_elements_max", KSTAT_DATA_UINT64 },
599 { "hash_collisions", KSTAT_DATA_UINT64 },
600 { "hash_chains", KSTAT_DATA_UINT64 },
601 { "hash_chain_max", KSTAT_DATA_UINT64 },
602 { "p", KSTAT_DATA_UINT64 },
603 { "c", KSTAT_DATA_UINT64 },
604 { "c_min", KSTAT_DATA_UINT64 },
605 { "c_max", KSTAT_DATA_UINT64 },
606 { "size", KSTAT_DATA_UINT64 },
607 { "hdr_size", KSTAT_DATA_UINT64 },
608 { "data_size", KSTAT_DATA_UINT64 },
609 { "metadata_size", KSTAT_DATA_UINT64 },
610 { "other_size", KSTAT_DATA_UINT64 },
611 { "anon_size", KSTAT_DATA_UINT64 },
612 { "anon_evictable_data", KSTAT_DATA_UINT64 },
613 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
614 { "mru_size", KSTAT_DATA_UINT64 },
615 { "mru_evictable_data", KSTAT_DATA_UINT64 },
616 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
617 { "mru_ghost_size", KSTAT_DATA_UINT64 },
618 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
619 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
620 { "mfu_size", KSTAT_DATA_UINT64 },
621 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
622 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
623 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
624 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
625 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
626 { "l2_hits", KSTAT_DATA_UINT64 },
627 { "l2_misses", KSTAT_DATA_UINT64 },
628 { "l2_feeds", KSTAT_DATA_UINT64 },
629 { "l2_rw_clash", KSTAT_DATA_UINT64 },
630 { "l2_read_bytes", KSTAT_DATA_UINT64 },
631 { "l2_write_bytes", KSTAT_DATA_UINT64 },
632 { "l2_writes_sent", KSTAT_DATA_UINT64 },
633 { "l2_writes_done", KSTAT_DATA_UINT64 },
634 { "l2_writes_error", KSTAT_DATA_UINT64 },
635 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
636 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
637 { "l2_evict_reading", KSTAT_DATA_UINT64 },
638 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
639 { "l2_free_on_write", KSTAT_DATA_UINT64 },
640 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 },
641 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
642 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
643 { "l2_io_error", KSTAT_DATA_UINT64 },
644 { "l2_size", KSTAT_DATA_UINT64 },
645 { "l2_asize", KSTAT_DATA_UINT64 },
646 { "l2_hdr_size", KSTAT_DATA_UINT64 },
647 { "l2_compress_successes", KSTAT_DATA_UINT64 },
648 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
649 { "l2_compress_failures", KSTAT_DATA_UINT64 },
650 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
651 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
652 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
653 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
654 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
655 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
656 { "l2_write_full", KSTAT_DATA_UINT64 },
657 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
658 { "l2_write_pios", KSTAT_DATA_UINT64 },
659 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
660 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
661 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
662 { "memory_throttle_count", KSTAT_DATA_UINT64 },
663 { "duplicate_buffers", KSTAT_DATA_UINT64 },
664 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
665 { "duplicate_reads", KSTAT_DATA_UINT64 },
666 { "arc_meta_used", KSTAT_DATA_UINT64 },
667 { "arc_meta_limit", KSTAT_DATA_UINT64 },
668 { "arc_meta_max", KSTAT_DATA_UINT64 },
669 { "arc_meta_min", KSTAT_DATA_UINT64 }
672 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
674 #define ARCSTAT_INCR(stat, val) \
675 atomic_add_64(&arc_stats.stat.value.ui64, (val))
677 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
678 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
680 #define ARCSTAT_MAX(stat, val) { \
682 while ((val) > (m = arc_stats.stat.value.ui64) && \
683 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
687 #define ARCSTAT_MAXSTAT(stat) \
688 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
691 * We define a macro to allow ARC hits/misses to be easily broken down by
692 * two separate conditions, giving a total of four different subtypes for
693 * each of hits and misses (so eight statistics total).
695 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
698 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
700 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
704 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
706 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
711 static arc_state_t *arc_anon;
712 static arc_state_t *arc_mru;
713 static arc_state_t *arc_mru_ghost;
714 static arc_state_t *arc_mfu;
715 static arc_state_t *arc_mfu_ghost;
716 static arc_state_t *arc_l2c_only;
719 * There are several ARC variables that are critical to export as kstats --
720 * but we don't want to have to grovel around in the kstat whenever we wish to
721 * manipulate them. For these variables, we therefore define them to be in
722 * terms of the statistic variable. This assures that we are not introducing
723 * the possibility of inconsistency by having shadow copies of the variables,
724 * while still allowing the code to be readable.
726 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
727 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
728 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
729 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
730 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
731 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
732 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
733 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
734 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
736 #define L2ARC_IS_VALID_COMPRESS(_c_) \
737 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
739 static int arc_no_grow; /* Don't try to grow cache size */
740 static uint64_t arc_tempreserve;
741 static uint64_t arc_loaned_bytes;
743 typedef struct arc_callback arc_callback_t;
745 struct arc_callback {
747 arc_done_func_t *acb_done;
749 zio_t *acb_zio_dummy;
750 arc_callback_t *acb_next;
753 typedef struct arc_write_callback arc_write_callback_t;
755 struct arc_write_callback {
757 arc_done_func_t *awcb_ready;
758 arc_done_func_t *awcb_physdone;
759 arc_done_func_t *awcb_done;
764 * ARC buffers are separated into multiple structs as a memory saving measure:
765 * - Common fields struct, always defined, and embedded within it:
766 * - L2-only fields, always allocated but undefined when not in L2ARC
767 * - L1-only fields, only allocated when in L1ARC
769 * Buffer in L1 Buffer only in L2
770 * +------------------------+ +------------------------+
771 * | arc_buf_hdr_t | | arc_buf_hdr_t |
775 * +------------------------+ +------------------------+
776 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
777 * | (undefined if L1-only) | | |
778 * +------------------------+ +------------------------+
779 * | l1arc_buf_hdr_t |
784 * +------------------------+
786 * Because it's possible for the L2ARC to become extremely large, we can wind
787 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
788 * is minimized by only allocating the fields necessary for an L1-cached buffer
789 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
790 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
791 * words in pointers. arc_hdr_realloc() is used to switch a header between
792 * these two allocation states.
794 typedef struct l1arc_buf_hdr {
795 kmutex_t b_freeze_lock;
798 * used for debugging wtih kmem_flags - by allocating and freeing
799 * b_thawed when the buffer is thawed, we get a record of the stack
800 * trace that thawed it.
807 /* for waiting on writes to complete */
810 /* protected by arc state mutex */
811 arc_state_t *b_state;
812 list_node_t b_arc_node;
814 /* updated atomically */
815 clock_t b_arc_access;
817 /* self protecting */
820 arc_callback_t *b_acb;
821 /* temporary buffer holder for in-flight compressed data */
825 typedef struct l2arc_dev l2arc_dev_t;
827 typedef struct l2arc_buf_hdr {
828 /* protected by arc_buf_hdr mutex */
829 l2arc_dev_t *b_dev; /* L2ARC device */
830 uint64_t b_daddr; /* disk address, offset byte */
831 /* real alloc'd buffer size depending on b_compress applied */
834 list_node_t b_l2node;
838 /* protected by hash lock */
842 * Even though this checksum is only set/verified when a buffer is in
843 * the L1 cache, it needs to be in the set of common fields because it
844 * must be preserved from the time before a buffer is written out to
845 * L2ARC until after it is read back in.
847 zio_cksum_t *b_freeze_cksum;
849 arc_buf_hdr_t *b_hash_next;
856 /* L2ARC fields. Undefined when not in L2ARC. */
857 l2arc_buf_hdr_t b_l2hdr;
858 /* L1ARC fields. Undefined when in l2arc_only state */
859 l1arc_buf_hdr_t b_l1hdr;
864 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
869 val = arc_meta_limit;
870 err = sysctl_handle_64(oidp, &val, 0, req);
871 if (err != 0 || req->newptr == NULL)
874 if (val <= 0 || val > arc_c_max)
877 arc_meta_limit = val;
882 static arc_buf_t *arc_eviction_list;
883 static kmutex_t arc_eviction_mtx;
884 static arc_buf_hdr_t arc_eviction_hdr;
886 #define GHOST_STATE(state) \
887 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
888 (state) == arc_l2c_only)
890 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
891 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
892 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
893 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
894 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
895 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
897 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
898 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
899 #define HDR_L2_READING(hdr) \
900 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
901 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
902 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
903 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
904 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
906 #define HDR_ISTYPE_METADATA(hdr) \
907 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
908 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
910 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
911 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
913 /* For storing compression mode in b_flags */
914 #define HDR_COMPRESS_OFFSET 24
915 #define HDR_COMPRESS_NBITS 7
917 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \
918 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS))
919 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \
920 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp))
926 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
927 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
930 * Hash table routines
933 #define HT_LOCK_PAD CACHE_LINE_SIZE
938 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
942 #define BUF_LOCKS 256
943 typedef struct buf_hash_table {
945 arc_buf_hdr_t **ht_table;
946 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
949 static buf_hash_table_t buf_hash_table;
951 #define BUF_HASH_INDEX(spa, dva, birth) \
952 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
953 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
954 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
955 #define HDR_LOCK(hdr) \
956 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
958 uint64_t zfs_crc64_table[256];
964 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
965 #define L2ARC_HEADROOM 2 /* num of writes */
967 * If we discover during ARC scan any buffers to be compressed, we boost
968 * our headroom for the next scanning cycle by this percentage multiple.
970 #define L2ARC_HEADROOM_BOOST 200
971 #define L2ARC_FEED_SECS 1 /* caching interval secs */
972 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
975 * Used to distinguish headers that are being process by
976 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
977 * address. This can happen when the header is added to the l2arc's list
978 * of buffers to write in the first stage of l2arc_write_buffers(), but
979 * has not yet been written out which happens in the second stage of
980 * l2arc_write_buffers().
982 #define L2ARC_ADDR_UNSET ((uint64_t)(-1))
984 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
985 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
987 /* L2ARC Performance Tunables */
988 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
989 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
990 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
991 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
992 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
993 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
994 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
995 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
996 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
998 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
999 &l2arc_write_max, 0, "max write size");
1000 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1001 &l2arc_write_boost, 0, "extra write during warmup");
1002 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1003 &l2arc_headroom, 0, "number of dev writes");
1004 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1005 &l2arc_feed_secs, 0, "interval seconds");
1006 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1007 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1009 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1010 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1011 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1012 &l2arc_feed_again, 0, "turbo warmup");
1013 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1014 &l2arc_norw, 0, "no reads during writes");
1016 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1017 &ARC_anon.arcs_size, 0, "size of anonymous state");
1018 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
1019 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
1020 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
1021 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
1023 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1024 &ARC_mru.arcs_size, 0, "size of mru state");
1025 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
1026 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
1027 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
1028 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
1030 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1031 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state");
1032 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
1033 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
1034 "size of metadata in mru ghost state");
1035 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
1036 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
1037 "size of data in mru ghost state");
1039 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1040 &ARC_mfu.arcs_size, 0, "size of mfu state");
1041 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
1042 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
1043 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
1044 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
1046 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1047 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state");
1048 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
1049 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
1050 "size of metadata in mfu ghost state");
1051 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
1052 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
1053 "size of data in mfu ghost state");
1055 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1056 &ARC_l2c_only.arcs_size, 0, "size of mru state");
1062 vdev_t *l2ad_vdev; /* vdev */
1063 spa_t *l2ad_spa; /* spa */
1064 uint64_t l2ad_hand; /* next write location */
1065 uint64_t l2ad_start; /* first addr on device */
1066 uint64_t l2ad_end; /* last addr on device */
1067 boolean_t l2ad_first; /* first sweep through */
1068 boolean_t l2ad_writing; /* currently writing */
1069 kmutex_t l2ad_mtx; /* lock for buffer list */
1070 list_t l2ad_buflist; /* buffer list */
1071 list_node_t l2ad_node; /* device list node */
1072 refcount_t l2ad_alloc; /* allocated bytes */
1075 static list_t L2ARC_dev_list; /* device list */
1076 static list_t *l2arc_dev_list; /* device list pointer */
1077 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1078 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1079 static list_t L2ARC_free_on_write; /* free after write buf list */
1080 static list_t *l2arc_free_on_write; /* free after write list ptr */
1081 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1082 static uint64_t l2arc_ndev; /* number of devices */
1084 typedef struct l2arc_read_callback {
1085 arc_buf_t *l2rcb_buf; /* read buffer */
1086 spa_t *l2rcb_spa; /* spa */
1087 blkptr_t l2rcb_bp; /* original blkptr */
1088 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1089 int l2rcb_flags; /* original flags */
1090 enum zio_compress l2rcb_compress; /* applied compress */
1091 } l2arc_read_callback_t;
1093 typedef struct l2arc_write_callback {
1094 l2arc_dev_t *l2wcb_dev; /* device info */
1095 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1096 } l2arc_write_callback_t;
1098 typedef struct l2arc_data_free {
1099 /* protected by l2arc_free_on_write_mtx */
1102 void (*l2df_func)(void *, size_t);
1103 list_node_t l2df_list_node;
1104 } l2arc_data_free_t;
1106 static kmutex_t l2arc_feed_thr_lock;
1107 static kcondvar_t l2arc_feed_thr_cv;
1108 static uint8_t l2arc_thread_exit;
1110 static void arc_get_data_buf(arc_buf_t *);
1111 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1112 static int arc_evict_needed(arc_buf_contents_t);
1113 static void arc_evict_ghost(arc_state_t *, uint64_t, int64_t);
1114 static void arc_buf_watch(arc_buf_t *);
1116 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1117 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1119 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1120 static void l2arc_read_done(zio_t *);
1122 static boolean_t l2arc_compress_buf(arc_buf_hdr_t *);
1123 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress);
1124 static void l2arc_release_cdata_buf(arc_buf_hdr_t *);
1127 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1129 uint8_t *vdva = (uint8_t *)dva;
1130 uint64_t crc = -1ULL;
1133 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1135 for (i = 0; i < sizeof (dva_t); i++)
1136 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1138 crc ^= (spa>>8) ^ birth;
1143 #define BUF_EMPTY(buf) \
1144 ((buf)->b_dva.dva_word[0] == 0 && \
1145 (buf)->b_dva.dva_word[1] == 0)
1147 #define BUF_EQUAL(spa, dva, birth, buf) \
1148 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1149 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1150 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
1153 buf_discard_identity(arc_buf_hdr_t *hdr)
1155 hdr->b_dva.dva_word[0] = 0;
1156 hdr->b_dva.dva_word[1] = 0;
1160 static arc_buf_hdr_t *
1161 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1163 const dva_t *dva = BP_IDENTITY(bp);
1164 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1165 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1166 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1169 mutex_enter(hash_lock);
1170 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1171 hdr = hdr->b_hash_next) {
1172 if (BUF_EQUAL(spa, dva, birth, hdr)) {
1177 mutex_exit(hash_lock);
1183 * Insert an entry into the hash table. If there is already an element
1184 * equal to elem in the hash table, then the already existing element
1185 * will be returned and the new element will not be inserted.
1186 * Otherwise returns NULL.
1187 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1189 static arc_buf_hdr_t *
1190 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1192 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1193 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1194 arc_buf_hdr_t *fhdr;
1197 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1198 ASSERT(hdr->b_birth != 0);
1199 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1201 if (lockp != NULL) {
1203 mutex_enter(hash_lock);
1205 ASSERT(MUTEX_HELD(hash_lock));
1208 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1209 fhdr = fhdr->b_hash_next, i++) {
1210 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1214 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1215 buf_hash_table.ht_table[idx] = hdr;
1216 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
1218 /* collect some hash table performance data */
1220 ARCSTAT_BUMP(arcstat_hash_collisions);
1222 ARCSTAT_BUMP(arcstat_hash_chains);
1224 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1227 ARCSTAT_BUMP(arcstat_hash_elements);
1228 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1234 buf_hash_remove(arc_buf_hdr_t *hdr)
1236 arc_buf_hdr_t *fhdr, **hdrp;
1237 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1239 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1240 ASSERT(HDR_IN_HASH_TABLE(hdr));
1242 hdrp = &buf_hash_table.ht_table[idx];
1243 while ((fhdr = *hdrp) != hdr) {
1244 ASSERT(fhdr != NULL);
1245 hdrp = &fhdr->b_hash_next;
1247 *hdrp = hdr->b_hash_next;
1248 hdr->b_hash_next = NULL;
1249 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE;
1251 /* collect some hash table performance data */
1252 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1254 if (buf_hash_table.ht_table[idx] &&
1255 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1256 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1260 * Global data structures and functions for the buf kmem cache.
1262 static kmem_cache_t *hdr_full_cache;
1263 static kmem_cache_t *hdr_l2only_cache;
1264 static kmem_cache_t *buf_cache;
1271 kmem_free(buf_hash_table.ht_table,
1272 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1273 for (i = 0; i < BUF_LOCKS; i++)
1274 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1275 kmem_cache_destroy(hdr_full_cache);
1276 kmem_cache_destroy(hdr_l2only_cache);
1277 kmem_cache_destroy(buf_cache);
1281 * Constructor callback - called when the cache is empty
1282 * and a new buf is requested.
1286 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1288 arc_buf_hdr_t *hdr = vbuf;
1290 bzero(hdr, HDR_FULL_SIZE);
1291 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1292 refcount_create(&hdr->b_l1hdr.b_refcnt);
1293 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1294 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1301 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1303 arc_buf_hdr_t *hdr = vbuf;
1305 bzero(hdr, HDR_L2ONLY_SIZE);
1306 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1313 buf_cons(void *vbuf, void *unused, int kmflag)
1315 arc_buf_t *buf = vbuf;
1317 bzero(buf, sizeof (arc_buf_t));
1318 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1319 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1325 * Destructor callback - called when a cached buf is
1326 * no longer required.
1330 hdr_full_dest(void *vbuf, void *unused)
1332 arc_buf_hdr_t *hdr = vbuf;
1334 ASSERT(BUF_EMPTY(hdr));
1335 cv_destroy(&hdr->b_l1hdr.b_cv);
1336 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1337 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1338 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1343 hdr_l2only_dest(void *vbuf, void *unused)
1345 arc_buf_hdr_t *hdr = vbuf;
1347 ASSERT(BUF_EMPTY(hdr));
1348 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1353 buf_dest(void *vbuf, void *unused)
1355 arc_buf_t *buf = vbuf;
1357 mutex_destroy(&buf->b_evict_lock);
1358 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1362 * Reclaim callback -- invoked when memory is low.
1366 hdr_recl(void *unused)
1368 dprintf("hdr_recl called\n");
1370 * umem calls the reclaim func when we destroy the buf cache,
1371 * which is after we do arc_fini().
1374 cv_signal(&arc_reclaim_thr_cv);
1381 uint64_t hsize = 1ULL << 12;
1385 * The hash table is big enough to fill all of physical memory
1386 * with an average block size of zfs_arc_average_blocksize (default 8K).
1387 * By default, the table will take up
1388 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1390 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1393 buf_hash_table.ht_mask = hsize - 1;
1394 buf_hash_table.ht_table =
1395 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1396 if (buf_hash_table.ht_table == NULL) {
1397 ASSERT(hsize > (1ULL << 8));
1402 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1403 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1404 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1405 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1407 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1408 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1410 for (i = 0; i < 256; i++)
1411 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1412 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1414 for (i = 0; i < BUF_LOCKS; i++) {
1415 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1416 NULL, MUTEX_DEFAULT, NULL);
1421 * Transition between the two allocation states for the arc_buf_hdr struct.
1422 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1423 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1424 * version is used when a cache buffer is only in the L2ARC in order to reduce
1427 static arc_buf_hdr_t *
1428 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
1430 ASSERT(HDR_HAS_L2HDR(hdr));
1432 arc_buf_hdr_t *nhdr;
1433 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1435 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
1436 (old == hdr_l2only_cache && new == hdr_full_cache));
1438 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
1440 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
1441 buf_hash_remove(hdr);
1443 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
1445 if (new == hdr_full_cache) {
1446 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1448 * arc_access and arc_change_state need to be aware that a
1449 * header has just come out of L2ARC, so we set its state to
1450 * l2c_only even though it's about to change.
1452 nhdr->b_l1hdr.b_state = arc_l2c_only;
1454 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1455 ASSERT0(hdr->b_l1hdr.b_datacnt);
1456 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
1458 * We might be removing the L1hdr of a buffer which was just
1459 * written out to L2ARC. If such a buffer is compressed then we
1460 * need to free its b_tmp_cdata before destroying the header.
1462 if (hdr->b_l1hdr.b_tmp_cdata != NULL &&
1463 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
1464 l2arc_release_cdata_buf(hdr);
1465 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR;
1468 * The header has been reallocated so we need to re-insert it into any
1471 (void) buf_hash_insert(nhdr, NULL);
1473 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
1475 mutex_enter(&dev->l2ad_mtx);
1478 * We must place the realloc'ed header back into the list at
1479 * the same spot. Otherwise, if it's placed earlier in the list,
1480 * l2arc_write_buffers() could find it during the function's
1481 * write phase, and try to write it out to the l2arc.
1483 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
1484 list_remove(&dev->l2ad_buflist, hdr);
1486 mutex_exit(&dev->l2ad_mtx);
1489 * Since we're using the pointer address as the tag when
1490 * incrementing and decrementing the l2ad_alloc refcount, we
1491 * must remove the old pointer (that we're about to destroy) and
1492 * add the new pointer to the refcount. Otherwise we'd remove
1493 * the wrong pointer address when calling arc_hdr_destroy() later.
1496 (void) refcount_remove_many(&dev->l2ad_alloc,
1497 hdr->b_l2hdr.b_asize, hdr);
1499 (void) refcount_add_many(&dev->l2ad_alloc,
1500 nhdr->b_l2hdr.b_asize, nhdr);
1502 buf_discard_identity(hdr);
1503 hdr->b_freeze_cksum = NULL;
1504 kmem_cache_free(old, hdr);
1510 #define ARC_MINTIME (hz>>4) /* 62 ms */
1513 arc_cksum_verify(arc_buf_t *buf)
1517 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1520 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1521 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) {
1522 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1525 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1526 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1527 panic("buffer modified while frozen!");
1528 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1532 arc_cksum_equal(arc_buf_t *buf)
1537 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1538 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1539 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1540 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1546 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1548 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1551 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1552 if (buf->b_hdr->b_freeze_cksum != NULL) {
1553 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1556 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1557 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1558 buf->b_hdr->b_freeze_cksum);
1559 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1562 #endif /* illumos */
1567 typedef struct procctl {
1575 arc_buf_unwatch(arc_buf_t *buf)
1582 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1583 ctl.prwatch.pr_size = 0;
1584 ctl.prwatch.pr_wflags = 0;
1585 result = write(arc_procfd, &ctl, sizeof (ctl));
1586 ASSERT3U(result, ==, sizeof (ctl));
1593 arc_buf_watch(arc_buf_t *buf)
1600 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1601 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1602 ctl.prwatch.pr_wflags = WA_WRITE;
1603 result = write(arc_procfd, &ctl, sizeof (ctl));
1604 ASSERT3U(result, ==, sizeof (ctl));
1608 #endif /* illumos */
1610 static arc_buf_contents_t
1611 arc_buf_type(arc_buf_hdr_t *hdr)
1613 if (HDR_ISTYPE_METADATA(hdr)) {
1614 return (ARC_BUFC_METADATA);
1616 return (ARC_BUFC_DATA);
1621 arc_bufc_to_flags(arc_buf_contents_t type)
1625 /* metadata field is 0 if buffer contains normal data */
1627 case ARC_BUFC_METADATA:
1628 return (ARC_FLAG_BUFC_METADATA);
1632 panic("undefined ARC buffer type!");
1633 return ((uint32_t)-1);
1637 arc_buf_thaw(arc_buf_t *buf)
1639 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1640 if (buf->b_hdr->b_l1hdr.b_state != arc_anon)
1641 panic("modifying non-anon buffer!");
1642 if (HDR_IO_IN_PROGRESS(buf->b_hdr))
1643 panic("modifying buffer while i/o in progress!");
1644 arc_cksum_verify(buf);
1647 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1648 if (buf->b_hdr->b_freeze_cksum != NULL) {
1649 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1650 buf->b_hdr->b_freeze_cksum = NULL;
1654 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1655 if (buf->b_hdr->b_l1hdr.b_thawed != NULL)
1656 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1);
1657 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1661 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1664 arc_buf_unwatch(buf);
1665 #endif /* illumos */
1669 arc_buf_freeze(arc_buf_t *buf)
1671 kmutex_t *hash_lock;
1673 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1676 hash_lock = HDR_LOCK(buf->b_hdr);
1677 mutex_enter(hash_lock);
1679 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1680 buf->b_hdr->b_l1hdr.b_state == arc_anon);
1681 arc_cksum_compute(buf, B_FALSE);
1682 mutex_exit(hash_lock);
1687 get_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock)
1689 uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1691 if (arc_buf_type(hdr) == ARC_BUFC_METADATA)
1692 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1);
1694 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1);
1695 buf_hashid += ARC_BUFC_NUMMETADATALISTS;
1698 *list = &state->arcs_lists[buf_hashid];
1699 *lock = ARCS_LOCK(state, buf_hashid);
1704 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1706 ASSERT(HDR_HAS_L1HDR(hdr));
1707 ASSERT(MUTEX_HELD(hash_lock));
1708 arc_state_t *state = hdr->b_l1hdr.b_state;
1710 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
1711 (state != arc_anon)) {
1712 /* We don't use the L2-only state list. */
1713 if (state != arc_l2c_only) {
1714 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt;
1715 uint64_t *size = &state->arcs_lsize[arc_buf_type(hdr)];
1719 get_buf_info(hdr, state, &list, &lock);
1720 ASSERT(!MUTEX_HELD(lock));
1722 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node));
1723 list_remove(list, hdr);
1724 if (GHOST_STATE(state)) {
1725 ASSERT0(hdr->b_l1hdr.b_datacnt);
1726 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
1727 delta = hdr->b_size;
1730 ASSERT3U(*size, >=, delta);
1731 atomic_add_64(size, -delta);
1734 /* remove the prefetch flag if we get a reference */
1735 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
1740 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1743 arc_state_t *state = hdr->b_l1hdr.b_state;
1745 ASSERT(HDR_HAS_L1HDR(hdr));
1746 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1747 ASSERT(!GHOST_STATE(state));
1750 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1751 * check to prevent usage of the arc_l2c_only list.
1753 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
1754 (state != arc_anon)) {
1755 uint64_t *size = &state->arcs_lsize[arc_buf_type(hdr)];
1759 get_buf_info(hdr, state, &list, &lock);
1760 ASSERT(!MUTEX_HELD(lock));
1762 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
1763 list_insert_head(list, hdr);
1764 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
1765 atomic_add_64(size, hdr->b_size *
1766 hdr->b_l1hdr.b_datacnt);
1773 * Move the supplied buffer to the indicated state. The mutex
1774 * for the buffer must be held by the caller.
1777 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
1778 kmutex_t *hash_lock)
1780 arc_state_t *old_state;
1783 uint64_t from_delta, to_delta;
1784 arc_buf_contents_t buftype = arc_buf_type(hdr);
1789 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1790 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1791 * L1 hdr doesn't always exist when we change state to arc_anon before
1792 * destroying a header, in which case reallocating to add the L1 hdr is
1795 if (HDR_HAS_L1HDR(hdr)) {
1796 old_state = hdr->b_l1hdr.b_state;
1797 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
1798 datacnt = hdr->b_l1hdr.b_datacnt;
1800 old_state = arc_l2c_only;
1805 ASSERT(MUTEX_HELD(hash_lock));
1806 ASSERT3P(new_state, !=, old_state);
1807 ASSERT(refcnt == 0 || datacnt > 0);
1808 ASSERT(!GHOST_STATE(new_state) || datacnt == 0);
1809 ASSERT(old_state != arc_anon || datacnt <= 1);
1811 from_delta = to_delta = datacnt * hdr->b_size;
1814 * If this buffer is evictable, transfer it from the
1815 * old state list to the new state list.
1818 if (old_state != arc_anon && old_state != arc_l2c_only) {
1820 uint64_t *size = &old_state->arcs_lsize[buftype];
1822 get_buf_info(hdr, old_state, &list, &lock);
1823 use_mutex = !MUTEX_HELD(lock);
1827 ASSERT(HDR_HAS_L1HDR(hdr));
1828 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node));
1829 list_remove(list, hdr);
1832 * If prefetching out of the ghost cache,
1833 * we will have a non-zero datacnt.
1835 if (GHOST_STATE(old_state) && datacnt == 0) {
1836 /* ghost elements have a ghost size */
1837 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1838 from_delta = hdr->b_size;
1840 ASSERT3U(*size, >=, from_delta);
1841 atomic_add_64(size, -from_delta);
1846 if (new_state != arc_anon && new_state != arc_l2c_only) {
1848 uint64_t *size = &new_state->arcs_lsize[buftype];
1851 * An L1 header always exists here, since if we're
1852 * moving to some L1-cached state (i.e. not l2c_only or
1853 * anonymous), we realloc the header to add an L1hdr
1856 ASSERT(HDR_HAS_L1HDR(hdr));
1857 get_buf_info(hdr, new_state, &list, &lock);
1858 use_mutex = !MUTEX_HELD(lock);
1862 list_insert_head(list, hdr);
1864 /* ghost elements have a ghost size */
1865 if (GHOST_STATE(new_state)) {
1866 ASSERT(datacnt == 0);
1867 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1868 to_delta = hdr->b_size;
1870 atomic_add_64(size, to_delta);
1877 ASSERT(!BUF_EMPTY(hdr));
1878 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
1879 buf_hash_remove(hdr);
1881 /* adjust state sizes (ignore arc_l2c_only) */
1882 if (to_delta && new_state != arc_l2c_only)
1883 atomic_add_64(&new_state->arcs_size, to_delta);
1884 if (from_delta && old_state != arc_l2c_only) {
1885 ASSERT3U(old_state->arcs_size, >=, from_delta);
1886 atomic_add_64(&old_state->arcs_size, -from_delta);
1888 if (HDR_HAS_L1HDR(hdr))
1889 hdr->b_l1hdr.b_state = new_state;
1892 * L2 headers should never be on the L2 state list since they don't
1893 * have L1 headers allocated.
1896 ASSERT(list_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
1897 list_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
1902 arc_space_consume(uint64_t space, arc_space_type_t type)
1904 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1907 case ARC_SPACE_DATA:
1908 ARCSTAT_INCR(arcstat_data_size, space);
1910 case ARC_SPACE_META:
1911 ARCSTAT_INCR(arcstat_metadata_size, space);
1913 case ARC_SPACE_OTHER:
1914 ARCSTAT_INCR(arcstat_other_size, space);
1916 case ARC_SPACE_HDRS:
1917 ARCSTAT_INCR(arcstat_hdr_size, space);
1919 case ARC_SPACE_L2HDRS:
1920 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1924 if (type != ARC_SPACE_DATA)
1925 ARCSTAT_INCR(arcstat_meta_used, space);
1927 atomic_add_64(&arc_size, space);
1931 arc_space_return(uint64_t space, arc_space_type_t type)
1933 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1936 case ARC_SPACE_DATA:
1937 ARCSTAT_INCR(arcstat_data_size, -space);
1939 case ARC_SPACE_META:
1940 ARCSTAT_INCR(arcstat_metadata_size, -space);
1942 case ARC_SPACE_OTHER:
1943 ARCSTAT_INCR(arcstat_other_size, -space);
1945 case ARC_SPACE_HDRS:
1946 ARCSTAT_INCR(arcstat_hdr_size, -space);
1948 case ARC_SPACE_L2HDRS:
1949 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1953 if (type != ARC_SPACE_DATA) {
1954 ASSERT(arc_meta_used >= space);
1955 if (arc_meta_max < arc_meta_used)
1956 arc_meta_max = arc_meta_used;
1957 ARCSTAT_INCR(arcstat_meta_used, -space);
1960 ASSERT(arc_size >= space);
1961 atomic_add_64(&arc_size, -space);
1965 arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
1970 ASSERT3U(size, >, 0);
1971 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
1972 ASSERT(BUF_EMPTY(hdr));
1973 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
1975 hdr->b_spa = spa_load_guid(spa);
1977 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1980 buf->b_efunc = NULL;
1981 buf->b_private = NULL;
1984 hdr->b_flags = arc_bufc_to_flags(type);
1985 hdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1987 hdr->b_l1hdr.b_buf = buf;
1988 hdr->b_l1hdr.b_state = arc_anon;
1989 hdr->b_l1hdr.b_arc_access = 0;
1990 hdr->b_l1hdr.b_datacnt = 1;
1992 arc_get_data_buf(buf);
1993 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
1994 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
1999 static char *arc_onloan_tag = "onloan";
2002 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2003 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2004 * buffers must be returned to the arc before they can be used by the DMU or
2008 arc_loan_buf(spa_t *spa, int size)
2012 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
2014 atomic_add_64(&arc_loaned_bytes, size);
2019 * Return a loaned arc buffer to the arc.
2022 arc_return_buf(arc_buf_t *buf, void *tag)
2024 arc_buf_hdr_t *hdr = buf->b_hdr;
2026 ASSERT(buf->b_data != NULL);
2027 ASSERT(HDR_HAS_L1HDR(hdr));
2028 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2029 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2031 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
2034 /* Detach an arc_buf from a dbuf (tag) */
2036 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2038 arc_buf_hdr_t *hdr = buf->b_hdr;
2040 ASSERT(buf->b_data != NULL);
2041 ASSERT(HDR_HAS_L1HDR(hdr));
2042 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2043 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2044 buf->b_efunc = NULL;
2045 buf->b_private = NULL;
2047 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
2051 arc_buf_clone(arc_buf_t *from)
2054 arc_buf_hdr_t *hdr = from->b_hdr;
2055 uint64_t size = hdr->b_size;
2057 ASSERT(HDR_HAS_L1HDR(hdr));
2058 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2060 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2063 buf->b_efunc = NULL;
2064 buf->b_private = NULL;
2065 buf->b_next = hdr->b_l1hdr.b_buf;
2066 hdr->b_l1hdr.b_buf = buf;
2067 arc_get_data_buf(buf);
2068 bcopy(from->b_data, buf->b_data, size);
2071 * This buffer already exists in the arc so create a duplicate
2072 * copy for the caller. If the buffer is associated with user data
2073 * then track the size and number of duplicates. These stats will be
2074 * updated as duplicate buffers are created and destroyed.
2076 if (HDR_ISTYPE_DATA(hdr)) {
2077 ARCSTAT_BUMP(arcstat_duplicate_buffers);
2078 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
2080 hdr->b_l1hdr.b_datacnt += 1;
2085 arc_buf_add_ref(arc_buf_t *buf, void* tag)
2088 kmutex_t *hash_lock;
2091 * Check to see if this buffer is evicted. Callers
2092 * must verify b_data != NULL to know if the add_ref
2095 mutex_enter(&buf->b_evict_lock);
2096 if (buf->b_data == NULL) {
2097 mutex_exit(&buf->b_evict_lock);
2100 hash_lock = HDR_LOCK(buf->b_hdr);
2101 mutex_enter(hash_lock);
2103 ASSERT(HDR_HAS_L1HDR(hdr));
2104 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2105 mutex_exit(&buf->b_evict_lock);
2107 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
2108 hdr->b_l1hdr.b_state == arc_mfu);
2110 add_reference(hdr, hash_lock, tag);
2111 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2112 arc_access(hdr, hash_lock);
2113 mutex_exit(hash_lock);
2114 ARCSTAT_BUMP(arcstat_hits);
2115 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
2116 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
2117 data, metadata, hits);
2121 arc_buf_free_on_write(void *data, size_t size,
2122 void (*free_func)(void *, size_t))
2124 l2arc_data_free_t *df;
2126 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
2127 df->l2df_data = data;
2128 df->l2df_size = size;
2129 df->l2df_func = free_func;
2130 mutex_enter(&l2arc_free_on_write_mtx);
2131 list_insert_head(l2arc_free_on_write, df);
2132 mutex_exit(&l2arc_free_on_write_mtx);
2136 * Free the arc data buffer. If it is an l2arc write in progress,
2137 * the buffer is placed on l2arc_free_on_write to be freed later.
2140 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
2142 arc_buf_hdr_t *hdr = buf->b_hdr;
2144 if (HDR_L2_WRITING(hdr)) {
2145 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
2146 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2148 free_func(buf->b_data, hdr->b_size);
2153 * Free up buf->b_data and if 'remove' is set, then pull the
2154 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2157 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
2159 ASSERT(HDR_HAS_L2HDR(hdr));
2160 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx));
2163 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
2164 * that doesn't exist, the header is in the arc_l2c_only state,
2165 * and there isn't anything to free (it's already been freed).
2167 if (!HDR_HAS_L1HDR(hdr))
2170 if (hdr->b_l1hdr.b_tmp_cdata == NULL)
2173 ASSERT(HDR_L2_WRITING(hdr));
2174 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, hdr->b_size,
2177 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
2178 hdr->b_l1hdr.b_tmp_cdata = NULL;
2182 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove)
2186 /* free up data associated with the buf */
2187 if (buf->b_data != NULL) {
2188 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
2189 uint64_t size = buf->b_hdr->b_size;
2190 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
2192 arc_cksum_verify(buf);
2194 arc_buf_unwatch(buf);
2195 #endif /* illumos */
2198 if (type == ARC_BUFC_METADATA) {
2199 arc_buf_data_free(buf, zio_buf_free);
2200 arc_space_return(size, ARC_SPACE_META);
2202 ASSERT(type == ARC_BUFC_DATA);
2203 arc_buf_data_free(buf, zio_data_buf_free);
2204 arc_space_return(size, ARC_SPACE_DATA);
2207 if (list_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) {
2208 uint64_t *cnt = &state->arcs_lsize[type];
2210 ASSERT(refcount_is_zero(
2211 &buf->b_hdr->b_l1hdr.b_refcnt));
2212 ASSERT(state != arc_anon && state != arc_l2c_only);
2214 ASSERT3U(*cnt, >=, size);
2215 atomic_add_64(cnt, -size);
2217 ASSERT3U(state->arcs_size, >=, size);
2218 atomic_add_64(&state->arcs_size, -size);
2222 * If we're destroying a duplicate buffer make sure
2223 * that the appropriate statistics are updated.
2225 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 &&
2226 HDR_ISTYPE_DATA(buf->b_hdr)) {
2227 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
2228 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
2230 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0);
2231 buf->b_hdr->b_l1hdr.b_datacnt -= 1;
2234 /* only remove the buf if requested */
2238 /* remove the buf from the hdr list */
2239 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf;
2240 bufp = &(*bufp)->b_next)
2242 *bufp = buf->b_next;
2245 ASSERT(buf->b_efunc == NULL);
2247 /* clean up the buf */
2249 kmem_cache_free(buf_cache, buf);
2253 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2255 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2256 l2arc_dev_t *dev = l2hdr->b_dev;
2258 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2259 ASSERT(HDR_HAS_L2HDR(hdr));
2261 list_remove(&dev->l2ad_buflist, hdr);
2264 * We don't want to leak the b_tmp_cdata buffer that was
2265 * allocated in l2arc_write_buffers()
2267 arc_buf_l2_cdata_free(hdr);
2270 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2271 * this header is being processed by l2arc_write_buffers() (i.e.
2272 * it's in the first stage of l2arc_write_buffers()).
2273 * Re-affirming that truth here, just to serve as a reminder. If
2274 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2275 * may not have its HDR_L2_WRITING flag set. (the write may have
2276 * completed, in which case HDR_L2_WRITING will be false and the
2277 * b_daddr field will point to the address of the buffer on disk).
2279 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr));
2282 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2283 * l2arc_write_buffers(). Since we've just removed this header
2284 * from the l2arc buffer list, this header will never reach the
2285 * second stage of l2arc_write_buffers(), which increments the
2286 * accounting stats for this header. Thus, we must be careful
2287 * not to decrement them for this header either.
2289 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) {
2290 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
2291 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
2293 vdev_space_update(dev->l2ad_vdev,
2294 -l2hdr->b_asize, 0, 0);
2296 (void) refcount_remove_many(&dev->l2ad_alloc,
2297 l2hdr->b_asize, hdr);
2300 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
2304 arc_hdr_destroy(arc_buf_hdr_t *hdr)
2306 if (HDR_HAS_L1HDR(hdr)) {
2307 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
2308 hdr->b_l1hdr.b_datacnt > 0);
2309 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2310 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2312 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2313 ASSERT(!HDR_IN_HASH_TABLE(hdr));
2315 if (HDR_HAS_L2HDR(hdr)) {
2316 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2317 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
2320 mutex_enter(&dev->l2ad_mtx);
2323 * Even though we checked this conditional above, we
2324 * need to check this again now that we have the
2325 * l2ad_mtx. This is because we could be racing with
2326 * another thread calling l2arc_evict() which might have
2327 * destroyed this header's L2 portion as we were waiting
2328 * to acquire the l2ad_mtx. If that happens, we don't
2329 * want to re-destroy the header's L2 portion.
2331 if (HDR_HAS_L2HDR(hdr)) {
2332 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr,
2333 hdr->b_l2hdr.b_asize, 0);
2334 arc_hdr_l2hdr_destroy(hdr);
2338 mutex_exit(&dev->l2ad_mtx);
2341 if (!BUF_EMPTY(hdr))
2342 buf_discard_identity(hdr);
2343 if (hdr->b_freeze_cksum != NULL) {
2344 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
2345 hdr->b_freeze_cksum = NULL;
2348 if (HDR_HAS_L1HDR(hdr)) {
2349 while (hdr->b_l1hdr.b_buf) {
2350 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2352 if (buf->b_efunc != NULL) {
2353 mutex_enter(&arc_eviction_mtx);
2354 mutex_enter(&buf->b_evict_lock);
2355 ASSERT(buf->b_hdr != NULL);
2356 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE,
2358 hdr->b_l1hdr.b_buf = buf->b_next;
2359 buf->b_hdr = &arc_eviction_hdr;
2360 buf->b_next = arc_eviction_list;
2361 arc_eviction_list = buf;
2362 mutex_exit(&buf->b_evict_lock);
2363 mutex_exit(&arc_eviction_mtx);
2365 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE,
2370 if (hdr->b_l1hdr.b_thawed != NULL) {
2371 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2372 hdr->b_l1hdr.b_thawed = NULL;
2377 ASSERT3P(hdr->b_hash_next, ==, NULL);
2378 if (HDR_HAS_L1HDR(hdr)) {
2379 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
2380 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
2381 kmem_cache_free(hdr_full_cache, hdr);
2383 kmem_cache_free(hdr_l2only_cache, hdr);
2388 arc_buf_free(arc_buf_t *buf, void *tag)
2390 arc_buf_hdr_t *hdr = buf->b_hdr;
2391 int hashed = hdr->b_l1hdr.b_state != arc_anon;
2393 ASSERT(buf->b_efunc == NULL);
2394 ASSERT(buf->b_data != NULL);
2397 kmutex_t *hash_lock = HDR_LOCK(hdr);
2399 mutex_enter(hash_lock);
2401 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2403 (void) remove_reference(hdr, hash_lock, tag);
2404 if (hdr->b_l1hdr.b_datacnt > 1) {
2405 arc_buf_destroy(buf, FALSE, TRUE);
2407 ASSERT(buf == hdr->b_l1hdr.b_buf);
2408 ASSERT(buf->b_efunc == NULL);
2409 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2411 mutex_exit(hash_lock);
2412 } else if (HDR_IO_IN_PROGRESS(hdr)) {
2415 * We are in the middle of an async write. Don't destroy
2416 * this buffer unless the write completes before we finish
2417 * decrementing the reference count.
2419 mutex_enter(&arc_eviction_mtx);
2420 (void) remove_reference(hdr, NULL, tag);
2421 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2422 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
2423 mutex_exit(&arc_eviction_mtx);
2425 arc_hdr_destroy(hdr);
2427 if (remove_reference(hdr, NULL, tag) > 0)
2428 arc_buf_destroy(buf, FALSE, TRUE);
2430 arc_hdr_destroy(hdr);
2435 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
2437 arc_buf_hdr_t *hdr = buf->b_hdr;
2438 kmutex_t *hash_lock = HDR_LOCK(hdr);
2439 boolean_t no_callback = (buf->b_efunc == NULL);
2441 if (hdr->b_l1hdr.b_state == arc_anon) {
2442 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
2443 arc_buf_free(buf, tag);
2444 return (no_callback);
2447 mutex_enter(hash_lock);
2449 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
2450 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2451 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2452 ASSERT(buf->b_data != NULL);
2454 (void) remove_reference(hdr, hash_lock, tag);
2455 if (hdr->b_l1hdr.b_datacnt > 1) {
2457 arc_buf_destroy(buf, FALSE, TRUE);
2458 } else if (no_callback) {
2459 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL);
2460 ASSERT(buf->b_efunc == NULL);
2461 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2463 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 ||
2464 refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2465 mutex_exit(hash_lock);
2466 return (no_callback);
2470 arc_buf_size(arc_buf_t *buf)
2472 return (buf->b_hdr->b_size);
2476 * Called from the DMU to determine if the current buffer should be
2477 * evicted. In order to ensure proper locking, the eviction must be initiated
2478 * from the DMU. Return true if the buffer is associated with user data and
2479 * duplicate buffers still exist.
2482 arc_buf_eviction_needed(arc_buf_t *buf)
2485 boolean_t evict_needed = B_FALSE;
2487 if (zfs_disable_dup_eviction)
2490 mutex_enter(&buf->b_evict_lock);
2494 * We are in arc_do_user_evicts(); let that function
2495 * perform the eviction.
2497 ASSERT(buf->b_data == NULL);
2498 mutex_exit(&buf->b_evict_lock);
2500 } else if (buf->b_data == NULL) {
2502 * We have already been added to the arc eviction list;
2503 * recommend eviction.
2505 ASSERT3P(hdr, ==, &arc_eviction_hdr);
2506 mutex_exit(&buf->b_evict_lock);
2510 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr))
2511 evict_needed = B_TRUE;
2513 mutex_exit(&buf->b_evict_lock);
2514 return (evict_needed);
2518 * Evict buffers from list until we've removed the specified number of
2519 * bytes. Move the removed buffers to the appropriate evict state.
2520 * If the recycle flag is set, then attempt to "recycle" a buffer:
2521 * - look for a buffer to evict that is `bytes' long.
2522 * - return the data block from this buffer rather than freeing it.
2523 * This flag is used by callers that are trying to make space for a
2524 * new buffer in a full arc cache.
2526 * This function makes a "best effort". It skips over any buffers
2527 * it can't get a hash_lock on, and so may not catch all candidates.
2528 * It may also return without evicting as much space as requested.
2531 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
2532 arc_buf_contents_t type)
2534 arc_state_t *evicted_state;
2535 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
2536 int64_t bytes_remaining;
2537 arc_buf_hdr_t *hdr, *hdr_prev = NULL;
2538 list_t *evicted_list, *list, *evicted_list_start, *list_start;
2539 kmutex_t *lock, *evicted_lock;
2540 kmutex_t *hash_lock;
2541 boolean_t have_lock;
2542 void *stolen = NULL;
2543 arc_buf_hdr_t marker = { 0 };
2545 static int evict_metadata_offset, evict_data_offset;
2546 int i, idx, offset, list_count, lists;
2548 ASSERT(state == arc_mru || state == arc_mfu);
2550 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2553 * Decide which "type" (data vs metadata) to recycle from.
2555 * If we are over the metadata limit, recycle from metadata.
2556 * If we are under the metadata minimum, recycle from data.
2557 * Otherwise, recycle from whichever type has the oldest (least
2558 * recently accessed) header. This is not yet implemented.
2561 arc_buf_contents_t realtype;
2562 if (state->arcs_lsize[ARC_BUFC_DATA] == 0) {
2563 realtype = ARC_BUFC_METADATA;
2564 } else if (state->arcs_lsize[ARC_BUFC_METADATA] == 0) {
2565 realtype = ARC_BUFC_DATA;
2566 } else if (arc_meta_used >= arc_meta_limit) {
2567 realtype = ARC_BUFC_METADATA;
2568 } else if (arc_meta_used <= arc_meta_min) {
2569 realtype = ARC_BUFC_DATA;
2571 } else if (HDR_HAS_L1HDR(data_hdr) &&
2572 HDR_HAS_L1HDR(metadata_hdr) &&
2573 data_hdr->b_l1hdr.b_arc_access <
2574 metadata_hdr->b_l1hdr.b_arc_access) {
2575 realtype = ARC_BUFC_DATA;
2577 realtype = ARC_BUFC_METADATA;
2584 if (realtype != type) {
2586 * If we want to evict from a different list,
2587 * we can not recycle, because DATA vs METADATA
2588 * buffers are segregated into different kmem
2589 * caches (and vmem arenas).
2596 if (type == ARC_BUFC_METADATA) {
2598 list_count = ARC_BUFC_NUMMETADATALISTS;
2599 list_start = &state->arcs_lists[0];
2600 evicted_list_start = &evicted_state->arcs_lists[0];
2601 idx = evict_metadata_offset;
2603 offset = ARC_BUFC_NUMMETADATALISTS;
2604 list_start = &state->arcs_lists[offset];
2605 evicted_list_start = &evicted_state->arcs_lists[offset];
2606 list_count = ARC_BUFC_NUMDATALISTS;
2607 idx = evict_data_offset;
2609 bytes_remaining = evicted_state->arcs_lsize[type];
2613 list = &list_start[idx];
2614 evicted_list = &evicted_list_start[idx];
2615 lock = ARCS_LOCK(state, (offset + idx));
2616 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx));
2619 * The ghost list lock must be acquired first in order to prevent
2620 * a 3 party deadlock:
2622 * - arc_evict_ghost acquires arc_*_ghost->arcs_mtx, followed by
2623 * l2ad_mtx in arc_hdr_realloc
2624 * - l2arc_write_buffers acquires l2ad_mtx, followed by arc_*->arcs_mtx
2625 * - arc_evict acquires arc_*_ghost->arcs_mtx, followed by
2626 * arc_*_ghost->arcs_mtx and forms a deadlock cycle.
2628 * This situation is avoided by acquiring the ghost list lock first.
2630 mutex_enter(evicted_lock);
2633 for (hdr = list_tail(list); hdr; hdr = hdr_prev) {
2634 hdr_prev = list_prev(list, hdr);
2635 if (HDR_HAS_L1HDR(hdr)) {
2637 (hdr->b_size * hdr->b_l1hdr.b_datacnt);
2639 /* prefetch buffers have a minimum lifespan */
2640 if (HDR_IO_IN_PROGRESS(hdr) ||
2641 (spa && hdr->b_spa != spa) ||
2642 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
2643 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
2644 arc_min_prefetch_lifespan)) {
2648 /* "lookahead" for better eviction candidate */
2649 if (recycle && hdr->b_size != bytes &&
2650 hdr_prev && hdr_prev->b_size == bytes)
2653 /* ignore markers */
2654 if (hdr->b_spa == 0)
2658 * It may take a long time to evict all the bufs requested.
2659 * To avoid blocking all arc activity, periodically drop
2660 * the arcs_mtx and give other threads a chance to run
2661 * before reacquiring the lock.
2663 * If we are looking for a buffer to recycle, we are in
2664 * the hot code path, so don't sleep.
2666 if (!recycle && count++ > arc_evict_iterations) {
2667 list_insert_after(list, hdr, &marker);
2669 mutex_exit(evicted_lock);
2670 kpreempt(KPREEMPT_SYNC);
2671 mutex_enter(evicted_lock);
2673 hdr_prev = list_prev(list, &marker);
2674 list_remove(list, &marker);
2679 hash_lock = HDR_LOCK(hdr);
2680 have_lock = MUTEX_HELD(hash_lock);
2681 if (have_lock || mutex_tryenter(hash_lock)) {
2682 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
2683 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0);
2684 while (hdr->b_l1hdr.b_buf) {
2685 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2686 if (!mutex_tryenter(&buf->b_evict_lock)) {
2690 if (buf->b_data != NULL) {
2691 bytes_evicted += hdr->b_size;
2693 arc_buf_type(hdr) == type &&
2694 hdr->b_size == bytes &&
2695 !HDR_L2_WRITING(hdr)) {
2696 stolen = buf->b_data;
2700 if (buf->b_efunc != NULL) {
2701 mutex_enter(&arc_eviction_mtx);
2702 arc_buf_destroy(buf,
2703 buf->b_data == stolen, FALSE);
2704 hdr->b_l1hdr.b_buf = buf->b_next;
2705 buf->b_hdr = &arc_eviction_hdr;
2706 buf->b_next = arc_eviction_list;
2707 arc_eviction_list = buf;
2708 mutex_exit(&arc_eviction_mtx);
2709 mutex_exit(&buf->b_evict_lock);
2711 mutex_exit(&buf->b_evict_lock);
2712 arc_buf_destroy(buf,
2713 buf->b_data == stolen, TRUE);
2717 if (HDR_HAS_L2HDR(hdr)) {
2718 ARCSTAT_INCR(arcstat_evict_l2_cached,
2721 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
2722 ARCSTAT_INCR(arcstat_evict_l2_eligible,
2726 arcstat_evict_l2_ineligible,
2731 if (hdr->b_l1hdr.b_datacnt == 0) {
2732 arc_change_state(evicted_state, hdr, hash_lock);
2733 ASSERT(HDR_IN_HASH_TABLE(hdr));
2734 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
2735 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
2736 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
2739 mutex_exit(hash_lock);
2740 if (bytes >= 0 && bytes_evicted >= bytes)
2742 if (bytes_remaining > 0) {
2743 mutex_exit(evicted_lock);
2745 idx = ((idx + 1) & (list_count - 1));
2755 mutex_exit(evicted_lock);
2757 idx = ((idx + 1) & (list_count - 1));
2760 if (bytes_evicted < bytes) {
2761 if (lists < list_count)
2764 dprintf("only evicted %lld bytes from %x",
2765 (longlong_t)bytes_evicted, state);
2767 if (type == ARC_BUFC_METADATA)
2768 evict_metadata_offset = idx;
2770 evict_data_offset = idx;
2773 ARCSTAT_INCR(arcstat_evict_skip, skipped);
2776 ARCSTAT_INCR(arcstat_mutex_miss, missed);
2779 * Note: we have just evicted some data into the ghost state,
2780 * potentially putting the ghost size over the desired size. Rather
2781 * that evicting from the ghost list in this hot code path, leave
2782 * this chore to the arc_reclaim_thread().
2786 ARCSTAT_BUMP(arcstat_stolen);
2791 * Remove buffers from list until we've removed the specified number of
2792 * bytes. Destroy the buffers that are removed.
2795 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
2797 arc_buf_hdr_t *hdr, *hdr_prev;
2798 arc_buf_hdr_t marker = { 0 };
2799 list_t *list, *list_start;
2800 kmutex_t *hash_lock, *lock;
2801 uint64_t bytes_deleted = 0;
2802 uint64_t bufs_skipped = 0;
2804 static int evict_offset;
2805 int list_count, idx = evict_offset;
2806 int offset, lists = 0;
2808 ASSERT(GHOST_STATE(state));
2811 * data lists come after metadata lists
2813 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS];
2814 list_count = ARC_BUFC_NUMDATALISTS;
2815 offset = ARC_BUFC_NUMMETADATALISTS;
2818 list = &list_start[idx];
2819 lock = ARCS_LOCK(state, idx + offset);
2822 for (hdr = list_tail(list); hdr; hdr = hdr_prev) {
2823 hdr_prev = list_prev(list, hdr);
2824 if (arc_buf_type(hdr) >= ARC_BUFC_NUMTYPES)
2825 panic("invalid hdr=%p", (void *)hdr);
2826 if (spa && hdr->b_spa != spa)
2829 /* ignore markers */
2830 if (hdr->b_spa == 0)
2833 hash_lock = HDR_LOCK(hdr);
2834 /* caller may be trying to modify this buffer, skip it */
2835 if (MUTEX_HELD(hash_lock))
2839 * It may take a long time to evict all the bufs requested.
2840 * To avoid blocking all arc activity, periodically drop
2841 * the arcs_mtx and give other threads a chance to run
2842 * before reacquiring the lock.
2844 if (count++ > arc_evict_iterations) {
2845 list_insert_after(list, hdr, &marker);
2847 kpreempt(KPREEMPT_SYNC);
2849 hdr_prev = list_prev(list, &marker);
2850 list_remove(list, &marker);
2854 if (mutex_tryenter(hash_lock)) {
2855 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2856 ASSERT(!HDR_HAS_L1HDR(hdr) ||
2857 hdr->b_l1hdr.b_buf == NULL);
2858 ARCSTAT_BUMP(arcstat_deleted);
2859 bytes_deleted += hdr->b_size;
2861 if (HDR_HAS_L2HDR(hdr)) {
2863 * This buffer is cached on the 2nd Level ARC;
2864 * don't destroy the header.
2866 arc_change_state(arc_l2c_only, hdr, hash_lock);
2868 * dropping from L1+L2 cached to L2-only,
2869 * realloc to remove the L1 header.
2871 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
2873 mutex_exit(hash_lock);
2875 arc_change_state(arc_anon, hdr, hash_lock);
2876 mutex_exit(hash_lock);
2877 arc_hdr_destroy(hdr);
2880 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
2881 if (bytes >= 0 && bytes_deleted >= bytes)
2883 } else if (bytes < 0) {
2885 * Insert a list marker and then wait for the
2886 * hash lock to become available. Once its
2887 * available, restart from where we left off.
2889 list_insert_after(list, hdr, &marker);
2891 mutex_enter(hash_lock);
2892 mutex_exit(hash_lock);
2894 hdr_prev = list_prev(list, &marker);
2895 list_remove(list, &marker);
2902 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1));
2905 if (lists < list_count)
2909 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] &&
2910 (bytes < 0 || bytes_deleted < bytes)) {
2911 list_start = &state->arcs_lists[0];
2912 list_count = ARC_BUFC_NUMMETADATALISTS;
2918 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2922 if (bytes_deleted < bytes)
2923 dprintf("only deleted %lld bytes from %p",
2924 (longlong_t)bytes_deleted, state);
2930 int64_t adjustment, delta;
2936 adjustment = MIN((int64_t)(arc_size - arc_c),
2937 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2940 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2941 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2942 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
2943 adjustment -= delta;
2946 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2947 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2948 (void) arc_evict(arc_mru, 0, delta, FALSE,
2956 adjustment = arc_size - arc_c;
2958 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2959 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2960 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2961 adjustment -= delta;
2964 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2965 int64_t delta = MIN(adjustment,
2966 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2967 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2972 * Adjust ghost lists
2975 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2977 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2978 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2979 arc_evict_ghost(arc_mru_ghost, 0, delta);
2983 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2985 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2986 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2987 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2992 arc_do_user_evicts(void)
2994 static arc_buf_t *tmp_arc_eviction_list;
2997 * Move list over to avoid LOR
3000 mutex_enter(&arc_eviction_mtx);
3001 tmp_arc_eviction_list = arc_eviction_list;
3002 arc_eviction_list = NULL;
3003 mutex_exit(&arc_eviction_mtx);
3005 while (tmp_arc_eviction_list != NULL) {
3006 arc_buf_t *buf = tmp_arc_eviction_list;
3007 tmp_arc_eviction_list = buf->b_next;
3008 mutex_enter(&buf->b_evict_lock);
3010 mutex_exit(&buf->b_evict_lock);
3012 if (buf->b_efunc != NULL)
3013 VERIFY0(buf->b_efunc(buf->b_private));
3015 buf->b_efunc = NULL;
3016 buf->b_private = NULL;
3017 kmem_cache_free(buf_cache, buf);
3020 if (arc_eviction_list != NULL)
3025 * Flush all *evictable* data from the cache for the given spa.
3026 * NOTE: this will not touch "active" (i.e. referenced) data.
3029 arc_flush(spa_t *spa)
3034 guid = spa_load_guid(spa);
3036 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) {
3037 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
3041 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) {
3042 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
3046 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) {
3047 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
3051 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) {
3052 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
3057 arc_evict_ghost(arc_mru_ghost, guid, -1);
3058 arc_evict_ghost(arc_mfu_ghost, guid, -1);
3060 mutex_enter(&arc_reclaim_thr_lock);
3061 arc_do_user_evicts();
3062 mutex_exit(&arc_reclaim_thr_lock);
3063 ASSERT(spa || arc_eviction_list == NULL);
3067 arc_shrink(int64_t to_free)
3070 if (arc_c > arc_c_min) {
3071 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
3072 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
3073 if (arc_c > arc_c_min + to_free)
3074 atomic_add_64(&arc_c, -to_free);
3078 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3079 if (arc_c > arc_size)
3080 arc_c = MAX(arc_size, arc_c_min);
3082 arc_p = (arc_c >> 1);
3084 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
3087 ASSERT(arc_c >= arc_c_min);
3088 ASSERT((int64_t)arc_p >= 0);
3091 if (arc_size > arc_c) {
3092 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size,
3098 static long needfree = 0;
3100 typedef enum free_memory_reason_t {
3105 FMR_PAGES_PP_MAXIMUM,
3109 } free_memory_reason_t;
3111 int64_t last_free_memory;
3112 free_memory_reason_t last_free_reason;
3115 * Additional reserve of pages for pp_reserve.
3117 int64_t arc_pages_pp_reserve = 64;
3120 * Additional reserve of pages for swapfs.
3122 int64_t arc_swapfs_reserve = 64;
3125 * Return the amount of memory that can be consumed before reclaim will be
3126 * needed. Positive if there is sufficient free memory, negative indicates
3127 * the amount of memory that needs to be freed up.
3130 arc_available_memory(void)
3132 int64_t lowest = INT64_MAX;
3134 free_memory_reason_t r = FMR_UNKNOWN;
3138 n = PAGESIZE * (-needfree);
3146 * Cooperate with pagedaemon when it's time for it to scan
3147 * and reclaim some pages.
3149 n = PAGESIZE * (int64_t)(freemem - zfs_arc_free_target);
3157 * check that we're out of range of the pageout scanner. It starts to
3158 * schedule paging if freemem is less than lotsfree and needfree.
3159 * lotsfree is the high-water mark for pageout, and needfree is the
3160 * number of needed free pages. We add extra pages here to make sure
3161 * the scanner doesn't start up while we're freeing memory.
3163 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3170 * check to make sure that swapfs has enough space so that anon
3171 * reservations can still succeed. anon_resvmem() checks that the
3172 * availrmem is greater than swapfs_minfree, and the number of reserved
3173 * swap pages. We also add a bit of extra here just to prevent
3174 * circumstances from getting really dire.
3176 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3177 desfree - arc_swapfs_reserve);
3180 r = FMR_SWAPFS_MINFREE;
3185 * Check that we have enough availrmem that memory locking (e.g., via
3186 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3187 * stores the number of pages that cannot be locked; when availrmem
3188 * drops below pages_pp_maximum, page locking mechanisms such as
3189 * page_pp_lock() will fail.)
3191 n = PAGESIZE * (availrmem - pages_pp_maximum -
3192 arc_pages_pp_reserve);
3195 r = FMR_PAGES_PP_MAXIMUM;
3199 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
3201 * If we're on an i386 platform, it's possible that we'll exhaust the
3202 * kernel heap space before we ever run out of available physical
3203 * memory. Most checks of the size of the heap_area compare against
3204 * tune.t_minarmem, which is the minimum available real memory that we
3205 * can have in the system. However, this is generally fixed at 25 pages
3206 * which is so low that it's useless. In this comparison, we seek to
3207 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3208 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3211 n = vmem_size(heap_arena, VMEM_FREE) -
3212 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)
3217 #define zio_arena NULL
3219 #define zio_arena heap_arena
3223 * If zio data pages are being allocated out of a separate heap segment,
3224 * then enforce that the size of available vmem for this arena remains
3225 * above about 1/16th free.
3227 * Note: The 1/16th arena free requirement was put in place
3228 * to aggressively evict memory from the arc in order to avoid
3229 * memory fragmentation issues.
3231 if (zio_arena != NULL) {
3232 n = vmem_size(zio_arena, VMEM_FREE) -
3233 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3241 * Above limits know nothing about real level of KVA fragmentation.
3242 * Start aggressive reclamation if too little sequential KVA left.
3245 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ?
3246 -(vmem_size(heap_arena, VMEM_ALLOC) >> 4) : INT64_MAX;
3254 /* Every 100 calls, free a small amount */
3255 if (spa_get_random(100) == 0)
3257 #endif /* _KERNEL */
3259 last_free_memory = lowest;
3260 last_free_reason = r;
3261 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
3267 * Determine if the system is under memory pressure and is asking
3268 * to reclaim memory. A return value of TRUE indicates that the system
3269 * is under memory pressure and that the arc should adjust accordingly.
3272 arc_reclaim_needed(void)
3274 return (arc_available_memory() < 0);
3277 extern kmem_cache_t *zio_buf_cache[];
3278 extern kmem_cache_t *zio_data_buf_cache[];
3279 extern kmem_cache_t *range_seg_cache;
3281 static __noinline void
3282 arc_kmem_reap_now(void)
3285 kmem_cache_t *prev_cache = NULL;
3286 kmem_cache_t *prev_data_cache = NULL;
3288 DTRACE_PROBE(arc__kmem_reap_start);
3290 if (arc_meta_used >= arc_meta_limit) {
3292 * We are exceeding our meta-data cache limit.
3293 * Purge some DNLC entries to release holds on meta-data.
3295 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
3299 * Reclaim unused memory from all kmem caches.
3305 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
3306 if (zio_buf_cache[i] != prev_cache) {
3307 prev_cache = zio_buf_cache[i];
3308 kmem_cache_reap_now(zio_buf_cache[i]);
3310 if (zio_data_buf_cache[i] != prev_data_cache) {
3311 prev_data_cache = zio_data_buf_cache[i];
3312 kmem_cache_reap_now(zio_data_buf_cache[i]);
3315 kmem_cache_reap_now(buf_cache);
3316 kmem_cache_reap_now(hdr_full_cache);
3317 kmem_cache_reap_now(hdr_l2only_cache);
3318 kmem_cache_reap_now(range_seg_cache);
3321 if (zio_arena != NULL) {
3323 * Ask the vmem arena to reclaim unused memory from its
3326 vmem_qcache_reap(zio_arena);
3329 DTRACE_PROBE(arc__kmem_reap_end);
3333 arc_reclaim_thread(void *dummy __unused)
3335 clock_t growtime = 0;
3338 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
3340 mutex_enter(&arc_reclaim_thr_lock);
3341 while (arc_thread_exit == 0) {
3342 int64_t free_memory = arc_available_memory();
3343 if (free_memory < 0) {
3345 arc_no_grow = B_TRUE;
3349 * Wait at least zfs_grow_retry (default 60) seconds
3350 * before considering growing.
3352 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
3354 arc_kmem_reap_now();
3357 * If we are still low on memory, shrink the ARC
3358 * so that we have arc_shrink_min free space.
3360 free_memory = arc_available_memory();
3363 (arc_c >> arc_shrink_shift) - free_memory;
3366 to_free = MAX(to_free, ptob(needfree));
3368 arc_shrink(to_free);
3370 } else if (free_memory < arc_c >> arc_no_grow_shift) {
3371 arc_no_grow = B_TRUE;
3372 } else if (ddi_get_lbolt() >= growtime) {
3373 arc_no_grow = B_FALSE;
3378 if (arc_eviction_list != NULL)
3379 arc_do_user_evicts();
3389 * This is necessary in order for the mdb ::arc dcmd to
3390 * show up to date information. Since the ::arc command
3391 * does not call the kstat's update function, without
3392 * this call, the command may show stale stats for the
3393 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3394 * with this change, the data might be up to 1 second
3395 * out of date; but that should suffice. The arc_state_t
3396 * structures can be queried directly if more accurate
3397 * information is needed.
3399 if (arc_ksp != NULL)
3400 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
3402 /* block until needed, or one second, whichever is shorter */
3403 CALLB_CPR_SAFE_BEGIN(&cpr);
3404 (void) cv_timedwait(&arc_reclaim_thr_cv,
3405 &arc_reclaim_thr_lock, hz);
3406 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
3409 arc_thread_exit = 0;
3410 cv_broadcast(&arc_reclaim_thr_cv);
3411 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
3416 * Adapt arc info given the number of bytes we are trying to add and
3417 * the state that we are comming from. This function is only called
3418 * when we are adding new content to the cache.
3421 arc_adapt(int bytes, arc_state_t *state)
3424 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
3426 if (state == arc_l2c_only)
3431 * Adapt the target size of the MRU list:
3432 * - if we just hit in the MRU ghost list, then increase
3433 * the target size of the MRU list.
3434 * - if we just hit in the MFU ghost list, then increase
3435 * the target size of the MFU list by decreasing the
3436 * target size of the MRU list.
3438 if (state == arc_mru_ghost) {
3439 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
3440 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
3441 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
3443 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
3444 } else if (state == arc_mfu_ghost) {
3447 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
3448 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
3449 mult = MIN(mult, 10);
3451 delta = MIN(bytes * mult, arc_p);
3452 arc_p = MAX(arc_p_min, arc_p - delta);
3454 ASSERT((int64_t)arc_p >= 0);
3456 if (arc_reclaim_needed()) {
3457 cv_signal(&arc_reclaim_thr_cv);
3464 if (arc_c >= arc_c_max)
3468 * If we're within (2 * maxblocksize) bytes of the target
3469 * cache size, increment the target cache size
3471 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
3472 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
3473 atomic_add_64(&arc_c, (int64_t)bytes);
3474 if (arc_c > arc_c_max)
3476 else if (state == arc_anon)
3477 atomic_add_64(&arc_p, (int64_t)bytes);
3481 ASSERT((int64_t)arc_p >= 0);
3485 * Check if the cache has reached its limits and eviction is required
3489 arc_evict_needed(arc_buf_contents_t type)
3491 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
3494 if (arc_reclaim_needed())
3497 return (arc_size > arc_c);
3501 * The buffer, supplied as the first argument, needs a data block.
3502 * So, if we are at cache max, determine which cache should be victimized.
3503 * We have the following cases:
3505 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
3506 * In this situation if we're out of space, but the resident size of the MFU is
3507 * under the limit, victimize the MFU cache to satisfy this insertion request.
3509 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
3510 * Here, we've used up all of the available space for the MRU, so we need to
3511 * evict from our own cache instead. Evict from the set of resident MRU
3514 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
3515 * c minus p represents the MFU space in the cache, since p is the size of the
3516 * cache that is dedicated to the MRU. In this situation there's still space on
3517 * the MFU side, so the MRU side needs to be victimized.
3519 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
3520 * MFU's resident set is consuming more space than it has been allotted. In
3521 * this situation, we must victimize our own cache, the MFU, for this insertion.
3524 arc_get_data_buf(arc_buf_t *buf)
3526 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
3527 uint64_t size = buf->b_hdr->b_size;
3528 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
3530 arc_adapt(size, state);
3533 * We have not yet reached cache maximum size,
3534 * just allocate a new buffer.
3536 if (!arc_evict_needed(type)) {
3537 if (type == ARC_BUFC_METADATA) {
3538 buf->b_data = zio_buf_alloc(size);
3539 arc_space_consume(size, ARC_SPACE_META);
3541 ASSERT(type == ARC_BUFC_DATA);
3542 buf->b_data = zio_data_buf_alloc(size);
3543 arc_space_consume(size, ARC_SPACE_DATA);
3549 * If we are prefetching from the mfu ghost list, this buffer
3550 * will end up on the mru list; so steal space from there.
3552 if (state == arc_mfu_ghost)
3553 state = HDR_PREFETCH(buf->b_hdr) ? arc_mru : arc_mfu;
3554 else if (state == arc_mru_ghost)
3557 if (state == arc_mru || state == arc_anon) {
3558 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
3559 state = (arc_mfu->arcs_lsize[type] >= size &&
3560 arc_p > mru_used) ? arc_mfu : arc_mru;
3563 uint64_t mfu_space = arc_c - arc_p;
3564 state = (arc_mru->arcs_lsize[type] >= size &&
3565 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
3567 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
3568 if (type == ARC_BUFC_METADATA) {
3569 buf->b_data = zio_buf_alloc(size);
3570 arc_space_consume(size, ARC_SPACE_META);
3572 ASSERT(type == ARC_BUFC_DATA);
3573 buf->b_data = zio_data_buf_alloc(size);
3574 arc_space_consume(size, ARC_SPACE_DATA);
3576 ARCSTAT_BUMP(arcstat_recycle_miss);
3578 ASSERT(buf->b_data != NULL);
3581 * Update the state size. Note that ghost states have a
3582 * "ghost size" and so don't need to be updated.
3584 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) {
3585 arc_buf_hdr_t *hdr = buf->b_hdr;
3587 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_size, size);
3588 if (list_link_active(&hdr->b_l1hdr.b_arc_node)) {
3589 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3590 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type],
3594 * If we are growing the cache, and we are adding anonymous
3595 * data, and we have outgrown arc_p, update arc_p
3597 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
3598 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
3599 arc_p = MIN(arc_c, arc_p + size);
3601 ARCSTAT_BUMP(arcstat_allocated);
3605 * This routine is called whenever a buffer is accessed.
3606 * NOTE: the hash lock is dropped in this function.
3609 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3613 ASSERT(MUTEX_HELD(hash_lock));
3614 ASSERT(HDR_HAS_L1HDR(hdr));
3616 if (hdr->b_l1hdr.b_state == arc_anon) {
3618 * This buffer is not in the cache, and does not
3619 * appear in our "ghost" list. Add the new buffer
3623 ASSERT0(hdr->b_l1hdr.b_arc_access);
3624 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3625 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3626 arc_change_state(arc_mru, hdr, hash_lock);
3628 } else if (hdr->b_l1hdr.b_state == arc_mru) {
3629 now = ddi_get_lbolt();
3632 * If this buffer is here because of a prefetch, then either:
3633 * - clear the flag if this is a "referencing" read
3634 * (any subsequent access will bump this into the MFU state).
3636 * - move the buffer to the head of the list if this is
3637 * another prefetch (to make it less likely to be evicted).
3639 if (HDR_PREFETCH(hdr)) {
3640 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
3641 ASSERT(list_link_active(
3642 &hdr->b_l1hdr.b_arc_node));
3644 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3645 ARCSTAT_BUMP(arcstat_mru_hits);
3647 hdr->b_l1hdr.b_arc_access = now;
3652 * This buffer has been "accessed" only once so far,
3653 * but it is still in the cache. Move it to the MFU
3656 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
3658 * More than 125ms have passed since we
3659 * instantiated this buffer. Move it to the
3660 * most frequently used state.
3662 hdr->b_l1hdr.b_arc_access = now;
3663 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3664 arc_change_state(arc_mfu, hdr, hash_lock);
3666 ARCSTAT_BUMP(arcstat_mru_hits);
3667 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
3668 arc_state_t *new_state;
3670 * This buffer has been "accessed" recently, but
3671 * was evicted from the cache. Move it to the
3675 if (HDR_PREFETCH(hdr)) {
3676 new_state = arc_mru;
3677 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
3678 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3679 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3681 new_state = arc_mfu;
3682 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3685 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3686 arc_change_state(new_state, hdr, hash_lock);
3688 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
3689 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
3691 * This buffer has been accessed more than once and is
3692 * still in the cache. Keep it in the MFU state.
3694 * NOTE: an add_reference() that occurred when we did
3695 * the arc_read() will have kicked this off the list.
3696 * If it was a prefetch, we will explicitly move it to
3697 * the head of the list now.
3699 if ((HDR_PREFETCH(hdr)) != 0) {
3700 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3701 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node));
3703 ARCSTAT_BUMP(arcstat_mfu_hits);
3704 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3705 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
3706 arc_state_t *new_state = arc_mfu;
3708 * This buffer has been accessed more than once but has
3709 * been evicted from the cache. Move it back to the
3713 if (HDR_PREFETCH(hdr)) {
3715 * This is a prefetch access...
3716 * move this block back to the MRU state.
3718 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3719 new_state = arc_mru;
3722 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3723 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3724 arc_change_state(new_state, hdr, hash_lock);
3726 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
3727 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
3729 * This buffer is on the 2nd Level ARC.
3732 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3733 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3734 arc_change_state(arc_mfu, hdr, hash_lock);
3736 ASSERT(!"invalid arc state");
3740 /* a generic arc_done_func_t which you can use */
3743 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
3745 if (zio == NULL || zio->io_error == 0)
3746 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
3747 VERIFY(arc_buf_remove_ref(buf, arg));
3750 /* a generic arc_done_func_t */
3752 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
3754 arc_buf_t **bufp = arg;
3755 if (zio && zio->io_error) {
3756 VERIFY(arc_buf_remove_ref(buf, arg));
3760 ASSERT(buf->b_data);
3765 arc_read_done(zio_t *zio)
3769 arc_buf_t *abuf; /* buffer we're assigning to callback */
3770 kmutex_t *hash_lock = NULL;
3771 arc_callback_t *callback_list, *acb;
3772 int freeable = FALSE;
3774 buf = zio->io_private;
3778 * The hdr was inserted into hash-table and removed from lists
3779 * prior to starting I/O. We should find this header, since
3780 * it's in the hash table, and it should be legit since it's
3781 * not possible to evict it during the I/O. The only possible
3782 * reason for it not to be found is if we were freed during the
3785 if (HDR_IN_HASH_TABLE(hdr)) {
3786 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
3787 ASSERT3U(hdr->b_dva.dva_word[0], ==,
3788 BP_IDENTITY(zio->io_bp)->dva_word[0]);
3789 ASSERT3U(hdr->b_dva.dva_word[1], ==,
3790 BP_IDENTITY(zio->io_bp)->dva_word[1]);
3792 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
3795 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
3796 hash_lock == NULL) ||
3798 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
3799 (found == hdr && HDR_L2_READING(hdr)));
3802 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED;
3803 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
3804 hdr->b_flags &= ~ARC_FLAG_L2CACHE;
3806 /* byteswap if necessary */
3807 callback_list = hdr->b_l1hdr.b_acb;
3808 ASSERT(callback_list != NULL);
3809 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
3810 dmu_object_byteswap_t bswap =
3811 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
3812 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
3813 byteswap_uint64_array :
3814 dmu_ot_byteswap[bswap].ob_func;
3815 func(buf->b_data, hdr->b_size);
3818 arc_cksum_compute(buf, B_FALSE);
3821 #endif /* illumos */
3823 if (hash_lock && zio->io_error == 0 &&
3824 hdr->b_l1hdr.b_state == arc_anon) {
3826 * Only call arc_access on anonymous buffers. This is because
3827 * if we've issued an I/O for an evicted buffer, we've already
3828 * called arc_access (to prevent any simultaneous readers from
3829 * getting confused).
3831 arc_access(hdr, hash_lock);
3834 /* create copies of the data buffer for the callers */
3836 for (acb = callback_list; acb; acb = acb->acb_next) {
3837 if (acb->acb_done) {
3839 ARCSTAT_BUMP(arcstat_duplicate_reads);
3840 abuf = arc_buf_clone(buf);
3842 acb->acb_buf = abuf;
3846 hdr->b_l1hdr.b_acb = NULL;
3847 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
3848 ASSERT(!HDR_BUF_AVAILABLE(hdr));
3850 ASSERT(buf->b_efunc == NULL);
3851 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
3852 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
3855 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
3856 callback_list != NULL);
3858 if (zio->io_error != 0) {
3859 hdr->b_flags |= ARC_FLAG_IO_ERROR;
3860 if (hdr->b_l1hdr.b_state != arc_anon)
3861 arc_change_state(arc_anon, hdr, hash_lock);
3862 if (HDR_IN_HASH_TABLE(hdr))
3863 buf_hash_remove(hdr);
3864 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3868 * Broadcast before we drop the hash_lock to avoid the possibility
3869 * that the hdr (and hence the cv) might be freed before we get to
3870 * the cv_broadcast().
3872 cv_broadcast(&hdr->b_l1hdr.b_cv);
3874 if (hash_lock != NULL) {
3875 mutex_exit(hash_lock);
3878 * This block was freed while we waited for the read to
3879 * complete. It has been removed from the hash table and
3880 * moved to the anonymous state (so that it won't show up
3883 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3884 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3887 /* execute each callback and free its structure */
3888 while ((acb = callback_list) != NULL) {
3890 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3892 if (acb->acb_zio_dummy != NULL) {
3893 acb->acb_zio_dummy->io_error = zio->io_error;
3894 zio_nowait(acb->acb_zio_dummy);
3897 callback_list = acb->acb_next;
3898 kmem_free(acb, sizeof (arc_callback_t));
3902 arc_hdr_destroy(hdr);
3906 * "Read" the block block at the specified DVA (in bp) via the
3907 * cache. If the block is found in the cache, invoke the provided
3908 * callback immediately and return. Note that the `zio' parameter
3909 * in the callback will be NULL in this case, since no IO was
3910 * required. If the block is not in the cache pass the read request
3911 * on to the spa with a substitute callback function, so that the
3912 * requested block will be added to the cache.
3914 * If a read request arrives for a block that has a read in-progress,
3915 * either wait for the in-progress read to complete (and return the
3916 * results); or, if this is a read with a "done" func, add a record
3917 * to the read to invoke the "done" func when the read completes,
3918 * and return; or just return.
3920 * arc_read_done() will invoke all the requested "done" functions
3921 * for readers of this block.
3924 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3925 void *private, zio_priority_t priority, int zio_flags,
3926 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
3928 arc_buf_hdr_t *hdr = NULL;
3929 arc_buf_t *buf = NULL;
3930 kmutex_t *hash_lock = NULL;
3932 uint64_t guid = spa_load_guid(spa);
3934 ASSERT(!BP_IS_EMBEDDED(bp) ||
3935 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
3938 if (!BP_IS_EMBEDDED(bp)) {
3940 * Embedded BP's have no DVA and require no I/O to "read".
3941 * Create an anonymous arc buf to back it.
3943 hdr = buf_hash_find(guid, bp, &hash_lock);
3946 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) {
3948 *arc_flags |= ARC_FLAG_CACHED;
3950 if (HDR_IO_IN_PROGRESS(hdr)) {
3952 if (*arc_flags & ARC_FLAG_WAIT) {
3953 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
3954 mutex_exit(hash_lock);
3957 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
3960 arc_callback_t *acb = NULL;
3962 acb = kmem_zalloc(sizeof (arc_callback_t),
3964 acb->acb_done = done;
3965 acb->acb_private = private;
3967 acb->acb_zio_dummy = zio_null(pio,
3968 spa, NULL, NULL, NULL, zio_flags);
3970 ASSERT(acb->acb_done != NULL);
3971 acb->acb_next = hdr->b_l1hdr.b_acb;
3972 hdr->b_l1hdr.b_acb = acb;
3973 add_reference(hdr, hash_lock, private);
3974 mutex_exit(hash_lock);
3977 mutex_exit(hash_lock);
3981 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
3982 hdr->b_l1hdr.b_state == arc_mfu);
3985 add_reference(hdr, hash_lock, private);
3987 * If this block is already in use, create a new
3988 * copy of the data so that we will be guaranteed
3989 * that arc_release() will always succeed.
3991 buf = hdr->b_l1hdr.b_buf;
3993 ASSERT(buf->b_data);
3994 if (HDR_BUF_AVAILABLE(hdr)) {
3995 ASSERT(buf->b_efunc == NULL);
3996 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
3998 buf = arc_buf_clone(buf);
4001 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4002 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4003 hdr->b_flags |= ARC_FLAG_PREFETCH;
4005 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4006 arc_access(hdr, hash_lock);
4007 if (*arc_flags & ARC_FLAG_L2CACHE)
4008 hdr->b_flags |= ARC_FLAG_L2CACHE;
4009 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4010 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4011 mutex_exit(hash_lock);
4012 ARCSTAT_BUMP(arcstat_hits);
4013 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4014 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4015 data, metadata, hits);
4018 done(NULL, buf, private);
4020 uint64_t size = BP_GET_LSIZE(bp);
4021 arc_callback_t *acb;
4024 boolean_t devw = B_FALSE;
4025 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
4026 int32_t b_asize = 0;
4029 /* this block is not in the cache */
4030 arc_buf_hdr_t *exists = NULL;
4031 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4032 buf = arc_buf_alloc(spa, size, private, type);
4034 if (!BP_IS_EMBEDDED(bp)) {
4035 hdr->b_dva = *BP_IDENTITY(bp);
4036 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4037 exists = buf_hash_insert(hdr, &hash_lock);
4039 if (exists != NULL) {
4040 /* somebody beat us to the hash insert */
4041 mutex_exit(hash_lock);
4042 buf_discard_identity(hdr);
4043 (void) arc_buf_remove_ref(buf, private);
4044 goto top; /* restart the IO request */
4047 /* if this is a prefetch, we don't have a reference */
4048 if (*arc_flags & ARC_FLAG_PREFETCH) {
4049 (void) remove_reference(hdr, hash_lock,
4051 hdr->b_flags |= ARC_FLAG_PREFETCH;
4053 if (*arc_flags & ARC_FLAG_L2CACHE)
4054 hdr->b_flags |= ARC_FLAG_L2CACHE;
4055 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4056 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4057 if (BP_GET_LEVEL(bp) > 0)
4058 hdr->b_flags |= ARC_FLAG_INDIRECT;
4061 * This block is in the ghost cache. If it was L2-only
4062 * (and thus didn't have an L1 hdr), we realloc the
4063 * header to add an L1 hdr.
4065 if (!HDR_HAS_L1HDR(hdr)) {
4066 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4070 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4071 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4072 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4073 ASSERT(hdr->b_l1hdr.b_buf == NULL);
4075 /* if this is a prefetch, we don't have a reference */
4076 if (*arc_flags & ARC_FLAG_PREFETCH)
4077 hdr->b_flags |= ARC_FLAG_PREFETCH;
4079 add_reference(hdr, hash_lock, private);
4080 if (*arc_flags & ARC_FLAG_L2CACHE)
4081 hdr->b_flags |= ARC_FLAG_L2CACHE;
4082 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4083 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4084 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
4087 buf->b_efunc = NULL;
4088 buf->b_private = NULL;
4090 hdr->b_l1hdr.b_buf = buf;
4091 ASSERT0(hdr->b_l1hdr.b_datacnt);
4092 hdr->b_l1hdr.b_datacnt = 1;
4093 arc_get_data_buf(buf);
4094 arc_access(hdr, hash_lock);
4097 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4099 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4100 acb->acb_done = done;
4101 acb->acb_private = private;
4103 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4104 hdr->b_l1hdr.b_acb = acb;
4105 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4107 if (HDR_HAS_L2HDR(hdr) &&
4108 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
4109 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
4110 addr = hdr->b_l2hdr.b_daddr;
4111 b_compress = HDR_GET_COMPRESS(hdr);
4112 b_asize = hdr->b_l2hdr.b_asize;
4114 * Lock out device removal.
4116 if (vdev_is_dead(vd) ||
4117 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
4121 if (hash_lock != NULL)
4122 mutex_exit(hash_lock);
4125 * At this point, we have a level 1 cache miss. Try again in
4126 * L2ARC if possible.
4128 ASSERT3U(hdr->b_size, ==, size);
4129 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
4130 uint64_t, size, zbookmark_phys_t *, zb);
4131 ARCSTAT_BUMP(arcstat_misses);
4132 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4133 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4134 data, metadata, misses);
4136 curthread->td_ru.ru_inblock++;
4139 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
4141 * Read from the L2ARC if the following are true:
4142 * 1. The L2ARC vdev was previously cached.
4143 * 2. This buffer still has L2ARC metadata.
4144 * 3. This buffer isn't currently writing to the L2ARC.
4145 * 4. The L2ARC entry wasn't evicted, which may
4146 * also have invalidated the vdev.
4147 * 5. This isn't prefetch and l2arc_noprefetch is set.
4149 if (HDR_HAS_L2HDR(hdr) &&
4150 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
4151 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
4152 l2arc_read_callback_t *cb;
4154 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
4155 ARCSTAT_BUMP(arcstat_l2_hits);
4157 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
4159 cb->l2rcb_buf = buf;
4160 cb->l2rcb_spa = spa;
4163 cb->l2rcb_flags = zio_flags;
4164 cb->l2rcb_compress = b_compress;
4166 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
4167 addr + size < vd->vdev_psize -
4168 VDEV_LABEL_END_SIZE);
4171 * l2arc read. The SCL_L2ARC lock will be
4172 * released by l2arc_read_done().
4173 * Issue a null zio if the underlying buffer
4174 * was squashed to zero size by compression.
4176 if (b_compress == ZIO_COMPRESS_EMPTY) {
4177 rzio = zio_null(pio, spa, vd,
4178 l2arc_read_done, cb,
4179 zio_flags | ZIO_FLAG_DONT_CACHE |
4181 ZIO_FLAG_DONT_PROPAGATE |
4182 ZIO_FLAG_DONT_RETRY);
4184 rzio = zio_read_phys(pio, vd, addr,
4185 b_asize, buf->b_data,
4187 l2arc_read_done, cb, priority,
4188 zio_flags | ZIO_FLAG_DONT_CACHE |
4190 ZIO_FLAG_DONT_PROPAGATE |
4191 ZIO_FLAG_DONT_RETRY, B_FALSE);
4193 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
4195 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
4197 if (*arc_flags & ARC_FLAG_NOWAIT) {
4202 ASSERT(*arc_flags & ARC_FLAG_WAIT);
4203 if (zio_wait(rzio) == 0)
4206 /* l2arc read error; goto zio_read() */
4208 DTRACE_PROBE1(l2arc__miss,
4209 arc_buf_hdr_t *, hdr);
4210 ARCSTAT_BUMP(arcstat_l2_misses);
4211 if (HDR_L2_WRITING(hdr))
4212 ARCSTAT_BUMP(arcstat_l2_rw_clash);
4213 spa_config_exit(spa, SCL_L2ARC, vd);
4217 spa_config_exit(spa, SCL_L2ARC, vd);
4218 if (l2arc_ndev != 0) {
4219 DTRACE_PROBE1(l2arc__miss,
4220 arc_buf_hdr_t *, hdr);
4221 ARCSTAT_BUMP(arcstat_l2_misses);
4225 rzio = zio_read(pio, spa, bp, buf->b_data, size,
4226 arc_read_done, buf, priority, zio_flags, zb);
4228 if (*arc_flags & ARC_FLAG_WAIT)
4229 return (zio_wait(rzio));
4231 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4238 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
4240 ASSERT(buf->b_hdr != NULL);
4241 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon);
4242 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) ||
4244 ASSERT(buf->b_efunc == NULL);
4245 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
4247 buf->b_efunc = func;
4248 buf->b_private = private;
4252 * Notify the arc that a block was freed, and thus will never be used again.
4255 arc_freed(spa_t *spa, const blkptr_t *bp)
4258 kmutex_t *hash_lock;
4259 uint64_t guid = spa_load_guid(spa);
4261 ASSERT(!BP_IS_EMBEDDED(bp));
4263 hdr = buf_hash_find(guid, bp, &hash_lock);
4266 if (HDR_BUF_AVAILABLE(hdr)) {
4267 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
4268 add_reference(hdr, hash_lock, FTAG);
4269 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4270 mutex_exit(hash_lock);
4272 arc_release(buf, FTAG);
4273 (void) arc_buf_remove_ref(buf, FTAG);
4275 mutex_exit(hash_lock);
4281 * Clear the user eviction callback set by arc_set_callback(), first calling
4282 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4283 * clearing the callback may result in the arc_buf being destroyed. However,
4284 * it will not result in the *last* arc_buf being destroyed, hence the data
4285 * will remain cached in the ARC. We make a copy of the arc buffer here so
4286 * that we can process the callback without holding any locks.
4288 * It's possible that the callback is already in the process of being cleared
4289 * by another thread. In this case we can not clear the callback.
4291 * Returns B_TRUE if the callback was successfully called and cleared.
4294 arc_clear_callback(arc_buf_t *buf)
4297 kmutex_t *hash_lock;
4298 arc_evict_func_t *efunc = buf->b_efunc;
4299 void *private = buf->b_private;
4300 list_t *list, *evicted_list;
4301 kmutex_t *lock, *evicted_lock;
4303 mutex_enter(&buf->b_evict_lock);
4307 * We are in arc_do_user_evicts().
4309 ASSERT(buf->b_data == NULL);
4310 mutex_exit(&buf->b_evict_lock);
4312 } else if (buf->b_data == NULL) {
4314 * We are on the eviction list; process this buffer now
4315 * but let arc_do_user_evicts() do the reaping.
4317 buf->b_efunc = NULL;
4318 mutex_exit(&buf->b_evict_lock);
4319 VERIFY0(efunc(private));
4322 hash_lock = HDR_LOCK(hdr);
4323 mutex_enter(hash_lock);
4325 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4327 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <,
4328 hdr->b_l1hdr.b_datacnt);
4329 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4330 hdr->b_l1hdr.b_state == arc_mfu);
4332 buf->b_efunc = NULL;
4333 buf->b_private = NULL;
4335 if (hdr->b_l1hdr.b_datacnt > 1) {
4336 mutex_exit(&buf->b_evict_lock);
4337 arc_buf_destroy(buf, FALSE, TRUE);
4339 ASSERT(buf == hdr->b_l1hdr.b_buf);
4340 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4341 mutex_exit(&buf->b_evict_lock);
4344 mutex_exit(hash_lock);
4345 VERIFY0(efunc(private));
4350 * Release this buffer from the cache, making it an anonymous buffer. This
4351 * must be done after a read and prior to modifying the buffer contents.
4352 * If the buffer has more than one reference, we must make
4353 * a new hdr for the buffer.
4356 arc_release(arc_buf_t *buf, void *tag)
4358 arc_buf_hdr_t *hdr = buf->b_hdr;
4361 * It would be nice to assert that if it's DMU metadata (level >
4362 * 0 || it's the dnode file), then it must be syncing context.
4363 * But we don't know that information at this level.
4366 mutex_enter(&buf->b_evict_lock);
4368 * We don't grab the hash lock prior to this check, because if
4369 * the buffer's header is in the arc_anon state, it won't be
4370 * linked into the hash table.
4372 if (hdr->b_l1hdr.b_state == arc_anon) {
4373 mutex_exit(&buf->b_evict_lock);
4374 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4375 ASSERT(!HDR_IN_HASH_TABLE(hdr));
4376 ASSERT(!HDR_HAS_L2HDR(hdr));
4377 ASSERT(BUF_EMPTY(hdr));
4378 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1);
4379 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
4380 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
4382 ASSERT3P(buf->b_efunc, ==, NULL);
4383 ASSERT3P(buf->b_private, ==, NULL);
4385 hdr->b_l1hdr.b_arc_access = 0;
4391 kmutex_t *hash_lock = HDR_LOCK(hdr);
4392 mutex_enter(hash_lock);
4395 * This assignment is only valid as long as the hash_lock is
4396 * held, we must be careful not to reference state or the
4397 * b_state field after dropping the lock.
4399 arc_state_t *state = hdr->b_l1hdr.b_state;
4400 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4401 ASSERT3P(state, !=, arc_anon);
4403 /* this buffer is not on any list */
4404 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
4406 if (HDR_HAS_L2HDR(hdr)) {
4407 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4410 * We have to recheck this conditional again now that
4411 * we're holding the l2ad_mtx to prevent a race with
4412 * another thread which might be concurrently calling
4413 * l2arc_evict(). In that case, l2arc_evict() might have
4414 * destroyed the header's L2 portion as we were waiting
4415 * to acquire the l2ad_mtx.
4417 if (HDR_HAS_L2HDR(hdr)) {
4418 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev,
4419 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0);
4420 arc_hdr_l2hdr_destroy(hdr);
4423 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4427 * Do we have more than one buf?
4429 if (hdr->b_l1hdr.b_datacnt > 1) {
4430 arc_buf_hdr_t *nhdr;
4432 uint64_t blksz = hdr->b_size;
4433 uint64_t spa = hdr->b_spa;
4434 arc_buf_contents_t type = arc_buf_type(hdr);
4435 uint32_t flags = hdr->b_flags;
4437 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
4439 * Pull the data off of this hdr and attach it to
4440 * a new anonymous hdr.
4442 (void) remove_reference(hdr, hash_lock, tag);
4443 bufp = &hdr->b_l1hdr.b_buf;
4444 while (*bufp != buf)
4445 bufp = &(*bufp)->b_next;
4446 *bufp = buf->b_next;
4449 ASSERT3P(state, !=, arc_l2c_only);
4450 ASSERT3U(state->arcs_size, >=, hdr->b_size);
4451 atomic_add_64(&state->arcs_size, -hdr->b_size);
4452 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
4453 ASSERT3P(state, !=, arc_l2c_only);
4454 uint64_t *size = &state->arcs_lsize[type];
4455 ASSERT3U(*size, >=, hdr->b_size);
4456 atomic_add_64(size, -hdr->b_size);
4460 * We're releasing a duplicate user data buffer, update
4461 * our statistics accordingly.
4463 if (HDR_ISTYPE_DATA(hdr)) {
4464 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
4465 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
4468 hdr->b_l1hdr.b_datacnt -= 1;
4469 arc_cksum_verify(buf);
4471 arc_buf_unwatch(buf);
4472 #endif /* illumos */
4474 mutex_exit(hash_lock);
4476 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
4477 nhdr->b_size = blksz;
4480 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING;
4481 nhdr->b_flags |= arc_bufc_to_flags(type);
4482 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
4484 nhdr->b_l1hdr.b_buf = buf;
4485 nhdr->b_l1hdr.b_datacnt = 1;
4486 nhdr->b_l1hdr.b_state = arc_anon;
4487 nhdr->b_l1hdr.b_arc_access = 0;
4488 nhdr->b_freeze_cksum = NULL;
4490 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
4492 mutex_exit(&buf->b_evict_lock);
4493 atomic_add_64(&arc_anon->arcs_size, blksz);
4495 mutex_exit(&buf->b_evict_lock);
4496 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
4497 /* protected by hash lock */
4498 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
4499 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4500 arc_change_state(arc_anon, hdr, hash_lock);
4501 hdr->b_l1hdr.b_arc_access = 0;
4502 mutex_exit(hash_lock);
4504 buf_discard_identity(hdr);
4507 buf->b_efunc = NULL;
4508 buf->b_private = NULL;
4512 arc_released(arc_buf_t *buf)
4516 mutex_enter(&buf->b_evict_lock);
4517 released = (buf->b_data != NULL &&
4518 buf->b_hdr->b_l1hdr.b_state == arc_anon);
4519 mutex_exit(&buf->b_evict_lock);
4525 arc_referenced(arc_buf_t *buf)
4529 mutex_enter(&buf->b_evict_lock);
4530 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
4531 mutex_exit(&buf->b_evict_lock);
4532 return (referenced);
4537 arc_write_ready(zio_t *zio)
4539 arc_write_callback_t *callback = zio->io_private;
4540 arc_buf_t *buf = callback->awcb_buf;
4541 arc_buf_hdr_t *hdr = buf->b_hdr;
4543 ASSERT(HDR_HAS_L1HDR(hdr));
4544 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
4545 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4546 callback->awcb_ready(zio, buf, callback->awcb_private);
4549 * If the IO is already in progress, then this is a re-write
4550 * attempt, so we need to thaw and re-compute the cksum.
4551 * It is the responsibility of the callback to handle the
4552 * accounting for any re-write attempt.
4554 if (HDR_IO_IN_PROGRESS(hdr)) {
4555 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
4556 if (hdr->b_freeze_cksum != NULL) {
4557 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
4558 hdr->b_freeze_cksum = NULL;
4560 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
4562 arc_cksum_compute(buf, B_FALSE);
4563 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4567 * The SPA calls this callback for each physical write that happens on behalf
4568 * of a logical write. See the comment in dbuf_write_physdone() for details.
4571 arc_write_physdone(zio_t *zio)
4573 arc_write_callback_t *cb = zio->io_private;
4574 if (cb->awcb_physdone != NULL)
4575 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
4579 arc_write_done(zio_t *zio)
4581 arc_write_callback_t *callback = zio->io_private;
4582 arc_buf_t *buf = callback->awcb_buf;
4583 arc_buf_hdr_t *hdr = buf->b_hdr;
4585 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4587 if (zio->io_error == 0) {
4588 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
4589 buf_discard_identity(hdr);
4591 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
4592 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
4595 ASSERT(BUF_EMPTY(hdr));
4599 * If the block to be written was all-zero or compressed enough to be
4600 * embedded in the BP, no write was performed so there will be no
4601 * dva/birth/checksum. The buffer must therefore remain anonymous
4604 if (!BUF_EMPTY(hdr)) {
4605 arc_buf_hdr_t *exists;
4606 kmutex_t *hash_lock;
4608 ASSERT(zio->io_error == 0);
4610 arc_cksum_verify(buf);
4612 exists = buf_hash_insert(hdr, &hash_lock);
4613 if (exists != NULL) {
4615 * This can only happen if we overwrite for
4616 * sync-to-convergence, because we remove
4617 * buffers from the hash table when we arc_free().
4619 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
4620 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4621 panic("bad overwrite, hdr=%p exists=%p",
4622 (void *)hdr, (void *)exists);
4623 ASSERT(refcount_is_zero(
4624 &exists->b_l1hdr.b_refcnt));
4625 arc_change_state(arc_anon, exists, hash_lock);
4626 mutex_exit(hash_lock);
4627 arc_hdr_destroy(exists);
4628 exists = buf_hash_insert(hdr, &hash_lock);
4629 ASSERT3P(exists, ==, NULL);
4630 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
4632 ASSERT(zio->io_prop.zp_nopwrite);
4633 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4634 panic("bad nopwrite, hdr=%p exists=%p",
4635 (void *)hdr, (void *)exists);
4638 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
4639 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
4640 ASSERT(BP_GET_DEDUP(zio->io_bp));
4641 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
4644 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4645 /* if it's not anon, we are doing a scrub */
4646 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
4647 arc_access(hdr, hash_lock);
4648 mutex_exit(hash_lock);
4650 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4653 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4654 callback->awcb_done(zio, buf, callback->awcb_private);
4656 kmem_free(callback, sizeof (arc_write_callback_t));
4660 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
4661 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
4662 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
4663 arc_done_func_t *done, void *private, zio_priority_t priority,
4664 int zio_flags, const zbookmark_phys_t *zb)
4666 arc_buf_hdr_t *hdr = buf->b_hdr;
4667 arc_write_callback_t *callback;
4670 ASSERT(ready != NULL);
4671 ASSERT(done != NULL);
4672 ASSERT(!HDR_IO_ERROR(hdr));
4673 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4674 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4675 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4677 hdr->b_flags |= ARC_FLAG_L2CACHE;
4679 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4680 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
4681 callback->awcb_ready = ready;
4682 callback->awcb_physdone = physdone;
4683 callback->awcb_done = done;
4684 callback->awcb_private = private;
4685 callback->awcb_buf = buf;
4687 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
4688 arc_write_ready, arc_write_physdone, arc_write_done, callback,
4689 priority, zio_flags, zb);
4695 arc_memory_throttle(uint64_t reserve, uint64_t txg)
4698 uint64_t available_memory = ptob(freemem);
4699 static uint64_t page_load = 0;
4700 static uint64_t last_txg = 0;
4702 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
4704 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE)));
4707 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
4710 if (txg > last_txg) {
4715 * If we are in pageout, we know that memory is already tight,
4716 * the arc is already going to be evicting, so we just want to
4717 * continue to let page writes occur as quickly as possible.
4719 if (curproc == pageproc) {
4720 if (page_load > MAX(ptob(minfree), available_memory) / 4)
4721 return (SET_ERROR(ERESTART));
4722 /* Note: reserve is inflated, so we deflate */
4723 page_load += reserve / 8;
4725 } else if (page_load > 0 && arc_reclaim_needed()) {
4726 /* memory is low, delay before restarting */
4727 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
4728 return (SET_ERROR(EAGAIN));
4736 arc_tempreserve_clear(uint64_t reserve)
4738 atomic_add_64(&arc_tempreserve, -reserve);
4739 ASSERT((int64_t)arc_tempreserve >= 0);
4743 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
4748 if (reserve > arc_c/4 && !arc_no_grow) {
4749 arc_c = MIN(arc_c_max, reserve * 4);
4750 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
4752 if (reserve > arc_c)
4753 return (SET_ERROR(ENOMEM));
4756 * Don't count loaned bufs as in flight dirty data to prevent long
4757 * network delays from blocking transactions that are ready to be
4758 * assigned to a txg.
4760 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
4763 * Writes will, almost always, require additional memory allocations
4764 * in order to compress/encrypt/etc the data. We therefore need to
4765 * make sure that there is sufficient available memory for this.
4767 error = arc_memory_throttle(reserve, txg);
4772 * Throttle writes when the amount of dirty data in the cache
4773 * gets too large. We try to keep the cache less than half full
4774 * of dirty blocks so that our sync times don't grow too large.
4775 * Note: if two requests come in concurrently, we might let them
4776 * both succeed, when one of them should fail. Not a huge deal.
4779 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
4780 anon_size > arc_c / 4) {
4781 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
4782 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
4783 arc_tempreserve>>10,
4784 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
4785 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
4786 reserve>>10, arc_c>>10);
4787 return (SET_ERROR(ERESTART));
4789 atomic_add_64(&arc_tempreserve, reserve);
4794 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
4795 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
4797 size->value.ui64 = state->arcs_size;
4798 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
4799 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
4803 arc_kstat_update(kstat_t *ksp, int rw)
4805 arc_stats_t *as = ksp->ks_data;
4807 if (rw == KSTAT_WRITE) {
4810 arc_kstat_update_state(arc_anon,
4811 &as->arcstat_anon_size,
4812 &as->arcstat_anon_evictable_data,
4813 &as->arcstat_anon_evictable_metadata);
4814 arc_kstat_update_state(arc_mru,
4815 &as->arcstat_mru_size,
4816 &as->arcstat_mru_evictable_data,
4817 &as->arcstat_mru_evictable_metadata);
4818 arc_kstat_update_state(arc_mru_ghost,
4819 &as->arcstat_mru_ghost_size,
4820 &as->arcstat_mru_ghost_evictable_data,
4821 &as->arcstat_mru_ghost_evictable_metadata);
4822 arc_kstat_update_state(arc_mfu,
4823 &as->arcstat_mfu_size,
4824 &as->arcstat_mfu_evictable_data,
4825 &as->arcstat_mfu_evictable_metadata);
4826 arc_kstat_update_state(arc_mfu_ghost,
4827 &as->arcstat_mfu_ghost_size,
4828 &as->arcstat_mfu_ghost_evictable_data,
4829 &as->arcstat_mfu_ghost_evictable_metadata);
4836 static eventhandler_tag arc_event_lowmem = NULL;
4839 arc_lowmem(void *arg __unused, int howto __unused)
4842 mutex_enter(&arc_reclaim_thr_lock);
4843 /* XXX: Memory deficit should be passed as argument. */
4844 needfree = btoc(arc_c >> arc_shrink_shift);
4845 DTRACE_PROBE(arc__needfree);
4846 cv_signal(&arc_reclaim_thr_cv);
4849 * It is unsafe to block here in arbitrary threads, because we can come
4850 * here from ARC itself and may hold ARC locks and thus risk a deadlock
4851 * with ARC reclaim thread.
4853 if (curproc == pageproc)
4854 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0);
4855 mutex_exit(&arc_reclaim_thr_lock);
4862 int i, prefetch_tunable_set = 0;
4864 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4865 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
4867 /* Convert seconds to clock ticks */
4868 arc_min_prefetch_lifespan = 1 * hz;
4870 /* Start out with 1/8 of all memory */
4871 arc_c = kmem_size() / 8;
4876 * On architectures where the physical memory can be larger
4877 * than the addressable space (intel in 32-bit mode), we may
4878 * need to limit the cache to 1/8 of VM size.
4880 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
4883 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
4884 arc_c_min = MAX(arc_c / 4, 16 << 20);
4885 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
4886 if (arc_c * 8 >= 1 << 30)
4887 arc_c_max = (arc_c * 8) - (1 << 30);
4889 arc_c_max = arc_c_min;
4890 arc_c_max = MAX(arc_c * 5, arc_c_max);
4894 * Allow the tunables to override our calculations if they are
4895 * reasonable (ie. over 16MB)
4897 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size())
4898 arc_c_max = zfs_arc_max;
4899 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max)
4900 arc_c_min = zfs_arc_min;
4904 arc_p = (arc_c >> 1);
4906 /* limit meta-data to 1/4 of the arc capacity */
4907 arc_meta_limit = arc_c_max / 4;
4909 /* Allow the tunable to override if it is reasonable */
4910 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
4911 arc_meta_limit = zfs_arc_meta_limit;
4913 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
4914 arc_c_min = arc_meta_limit / 2;
4916 if (zfs_arc_meta_min > 0) {
4917 arc_meta_min = zfs_arc_meta_min;
4919 arc_meta_min = arc_c_min / 2;
4922 if (zfs_arc_grow_retry > 0)
4923 arc_grow_retry = zfs_arc_grow_retry;
4925 if (zfs_arc_shrink_shift > 0)
4926 arc_shrink_shift = zfs_arc_shrink_shift;
4929 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
4931 if (arc_no_grow_shift >= arc_shrink_shift)
4932 arc_no_grow_shift = arc_shrink_shift - 1;
4934 if (zfs_arc_p_min_shift > 0)
4935 arc_p_min_shift = zfs_arc_p_min_shift;
4937 /* if kmem_flags are set, lets try to use less memory */
4938 if (kmem_debugging())
4940 if (arc_c < arc_c_min)
4943 zfs_arc_min = arc_c_min;
4944 zfs_arc_max = arc_c_max;
4946 arc_anon = &ARC_anon;
4948 arc_mru_ghost = &ARC_mru_ghost;
4950 arc_mfu_ghost = &ARC_mfu_ghost;
4951 arc_l2c_only = &ARC_l2c_only;
4954 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
4955 mutex_init(&arc_anon->arcs_locks[i].arcs_lock,
4956 NULL, MUTEX_DEFAULT, NULL);
4957 mutex_init(&arc_mru->arcs_locks[i].arcs_lock,
4958 NULL, MUTEX_DEFAULT, NULL);
4959 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock,
4960 NULL, MUTEX_DEFAULT, NULL);
4961 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock,
4962 NULL, MUTEX_DEFAULT, NULL);
4963 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock,
4964 NULL, MUTEX_DEFAULT, NULL);
4965 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock,
4966 NULL, MUTEX_DEFAULT, NULL);
4968 list_create(&arc_mru->arcs_lists[i],
4969 sizeof (arc_buf_hdr_t),
4970 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node));
4971 list_create(&arc_mru_ghost->arcs_lists[i],
4972 sizeof (arc_buf_hdr_t),
4973 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node));
4974 list_create(&arc_mfu->arcs_lists[i],
4975 sizeof (arc_buf_hdr_t),
4976 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node));
4977 list_create(&arc_mfu_ghost->arcs_lists[i],
4978 sizeof (arc_buf_hdr_t),
4979 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node));
4980 list_create(&arc_mfu_ghost->arcs_lists[i],
4981 sizeof (arc_buf_hdr_t),
4982 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node));
4983 list_create(&arc_l2c_only->arcs_lists[i],
4984 sizeof (arc_buf_hdr_t),
4985 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node));
4990 arc_thread_exit = 0;
4991 arc_eviction_list = NULL;
4992 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
4993 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
4995 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
4996 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
4998 if (arc_ksp != NULL) {
4999 arc_ksp->ks_data = &arc_stats;
5000 arc_ksp->ks_update = arc_kstat_update;
5001 kstat_install(arc_ksp);
5004 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
5005 TS_RUN, minclsyspri);
5008 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
5009 EVENTHANDLER_PRI_FIRST);
5016 * Calculate maximum amount of dirty data per pool.
5018 * If it has been set by /etc/system, take that.
5019 * Otherwise, use a percentage of physical memory defined by
5020 * zfs_dirty_data_max_percent (default 10%) with a cap at
5021 * zfs_dirty_data_max_max (default 4GB).
5023 if (zfs_dirty_data_max == 0) {
5024 zfs_dirty_data_max = ptob(physmem) *
5025 zfs_dirty_data_max_percent / 100;
5026 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
5027 zfs_dirty_data_max_max);
5031 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
5032 prefetch_tunable_set = 1;
5035 if (prefetch_tunable_set == 0) {
5036 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
5038 printf(" add \"vfs.zfs.prefetch_disable=0\" "
5039 "to /boot/loader.conf.\n");
5040 zfs_prefetch_disable = 1;
5043 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
5044 prefetch_tunable_set == 0) {
5045 printf("ZFS NOTICE: Prefetch is disabled by default if less "
5046 "than 4GB of RAM is present;\n"
5047 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
5048 "to /boot/loader.conf.\n");
5049 zfs_prefetch_disable = 1;
5052 /* Warn about ZFS memory and address space requirements. */
5053 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
5054 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
5055 "expect unstable behavior.\n");
5057 if (kmem_size() < 512 * (1 << 20)) {
5058 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
5059 "expect unstable behavior.\n");
5060 printf(" Consider tuning vm.kmem_size and "
5061 "vm.kmem_size_max\n");
5062 printf(" in /boot/loader.conf.\n");
5072 mutex_enter(&arc_reclaim_thr_lock);
5073 arc_thread_exit = 1;
5074 cv_signal(&arc_reclaim_thr_cv);
5075 while (arc_thread_exit != 0)
5076 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
5077 mutex_exit(&arc_reclaim_thr_lock);
5083 if (arc_ksp != NULL) {
5084 kstat_delete(arc_ksp);
5088 mutex_destroy(&arc_eviction_mtx);
5089 mutex_destroy(&arc_reclaim_thr_lock);
5090 cv_destroy(&arc_reclaim_thr_cv);
5092 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
5093 list_destroy(&arc_mru->arcs_lists[i]);
5094 list_destroy(&arc_mru_ghost->arcs_lists[i]);
5095 list_destroy(&arc_mfu->arcs_lists[i]);
5096 list_destroy(&arc_mfu_ghost->arcs_lists[i]);
5097 list_destroy(&arc_l2c_only->arcs_lists[i]);
5099 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock);
5100 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock);
5101 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock);
5102 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock);
5103 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock);
5104 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock);
5109 ASSERT0(arc_loaned_bytes);
5112 if (arc_event_lowmem != NULL)
5113 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
5120 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5121 * It uses dedicated storage devices to hold cached data, which are populated
5122 * using large infrequent writes. The main role of this cache is to boost
5123 * the performance of random read workloads. The intended L2ARC devices
5124 * include short-stroked disks, solid state disks, and other media with
5125 * substantially faster read latency than disk.
5127 * +-----------------------+
5129 * +-----------------------+
5132 * l2arc_feed_thread() arc_read()
5136 * +---------------+ |
5138 * +---------------+ |
5143 * +-------+ +-------+
5145 * | cache | | cache |
5146 * +-------+ +-------+
5147 * +=========+ .-----.
5148 * : L2ARC : |-_____-|
5149 * : devices : | Disks |
5150 * +=========+ `-_____-'
5152 * Read requests are satisfied from the following sources, in order:
5155 * 2) vdev cache of L2ARC devices
5157 * 4) vdev cache of disks
5160 * Some L2ARC device types exhibit extremely slow write performance.
5161 * To accommodate for this there are some significant differences between
5162 * the L2ARC and traditional cache design:
5164 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5165 * the ARC behave as usual, freeing buffers and placing headers on ghost
5166 * lists. The ARC does not send buffers to the L2ARC during eviction as
5167 * this would add inflated write latencies for all ARC memory pressure.
5169 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5170 * It does this by periodically scanning buffers from the eviction-end of
5171 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5172 * not already there. It scans until a headroom of buffers is satisfied,
5173 * which itself is a buffer for ARC eviction. If a compressible buffer is
5174 * found during scanning and selected for writing to an L2ARC device, we
5175 * temporarily boost scanning headroom during the next scan cycle to make
5176 * sure we adapt to compression effects (which might significantly reduce
5177 * the data volume we write to L2ARC). The thread that does this is
5178 * l2arc_feed_thread(), illustrated below; example sizes are included to
5179 * provide a better sense of ratio than this diagram:
5182 * +---------------------+----------+
5183 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5184 * +---------------------+----------+ | o L2ARC eligible
5185 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5186 * +---------------------+----------+ |
5187 * 15.9 Gbytes ^ 32 Mbytes |
5189 * l2arc_feed_thread()
5191 * l2arc write hand <--[oooo]--'
5195 * +==============================+
5196 * L2ARC dev |####|#|###|###| |####| ... |
5197 * +==============================+
5200 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5201 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5202 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5203 * safe to say that this is an uncommon case, since buffers at the end of
5204 * the ARC lists have moved there due to inactivity.
5206 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5207 * then the L2ARC simply misses copying some buffers. This serves as a
5208 * pressure valve to prevent heavy read workloads from both stalling the ARC
5209 * with waits and clogging the L2ARC with writes. This also helps prevent
5210 * the potential for the L2ARC to churn if it attempts to cache content too
5211 * quickly, such as during backups of the entire pool.
5213 * 5. After system boot and before the ARC has filled main memory, there are
5214 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5215 * lists can remain mostly static. Instead of searching from tail of these
5216 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5217 * for eligible buffers, greatly increasing its chance of finding them.
5219 * The L2ARC device write speed is also boosted during this time so that
5220 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5221 * there are no L2ARC reads, and no fear of degrading read performance
5222 * through increased writes.
5224 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5225 * the vdev queue can aggregate them into larger and fewer writes. Each
5226 * device is written to in a rotor fashion, sweeping writes through
5227 * available space then repeating.
5229 * 7. The L2ARC does not store dirty content. It never needs to flush
5230 * write buffers back to disk based storage.
5232 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5233 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5235 * The performance of the L2ARC can be tweaked by a number of tunables, which
5236 * may be necessary for different workloads:
5238 * l2arc_write_max max write bytes per interval
5239 * l2arc_write_boost extra write bytes during device warmup
5240 * l2arc_noprefetch skip caching prefetched buffers
5241 * l2arc_headroom number of max device writes to precache
5242 * l2arc_headroom_boost when we find compressed buffers during ARC
5243 * scanning, we multiply headroom by this
5244 * percentage factor for the next scan cycle,
5245 * since more compressed buffers are likely to
5247 * l2arc_feed_secs seconds between L2ARC writing
5249 * Tunables may be removed or added as future performance improvements are
5250 * integrated, and also may become zpool properties.
5252 * There are three key functions that control how the L2ARC warms up:
5254 * l2arc_write_eligible() check if a buffer is eligible to cache
5255 * l2arc_write_size() calculate how much to write
5256 * l2arc_write_interval() calculate sleep delay between writes
5258 * These three functions determine what to write, how much, and how quickly
5263 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
5266 * A buffer is *not* eligible for the L2ARC if it:
5267 * 1. belongs to a different spa.
5268 * 2. is already cached on the L2ARC.
5269 * 3. has an I/O in progress (it may be an incomplete read).
5270 * 4. is flagged not eligible (zfs property).
5272 if (hdr->b_spa != spa_guid) {
5273 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
5276 if (HDR_HAS_L2HDR(hdr)) {
5277 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
5280 if (HDR_IO_IN_PROGRESS(hdr)) {
5281 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
5284 if (!HDR_L2CACHE(hdr)) {
5285 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
5293 l2arc_write_size(void)
5298 * Make sure our globals have meaningful values in case the user
5301 size = l2arc_write_max;
5303 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
5304 "be greater than zero, resetting it to the default (%d)",
5306 size = l2arc_write_max = L2ARC_WRITE_SIZE;
5309 if (arc_warm == B_FALSE)
5310 size += l2arc_write_boost;
5317 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
5319 clock_t interval, next, now;
5322 * If the ARC lists are busy, increase our write rate; if the
5323 * lists are stale, idle back. This is achieved by checking
5324 * how much we previously wrote - if it was more than half of
5325 * what we wanted, schedule the next write much sooner.
5327 if (l2arc_feed_again && wrote > (wanted / 2))
5328 interval = (hz * l2arc_feed_min_ms) / 1000;
5330 interval = hz * l2arc_feed_secs;
5332 now = ddi_get_lbolt();
5333 next = MAX(now, MIN(now + interval, began + interval));
5339 * Cycle through L2ARC devices. This is how L2ARC load balances.
5340 * If a device is returned, this also returns holding the spa config lock.
5342 static l2arc_dev_t *
5343 l2arc_dev_get_next(void)
5345 l2arc_dev_t *first, *next = NULL;
5348 * Lock out the removal of spas (spa_namespace_lock), then removal
5349 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
5350 * both locks will be dropped and a spa config lock held instead.
5352 mutex_enter(&spa_namespace_lock);
5353 mutex_enter(&l2arc_dev_mtx);
5355 /* if there are no vdevs, there is nothing to do */
5356 if (l2arc_ndev == 0)
5360 next = l2arc_dev_last;
5362 /* loop around the list looking for a non-faulted vdev */
5364 next = list_head(l2arc_dev_list);
5366 next = list_next(l2arc_dev_list, next);
5368 next = list_head(l2arc_dev_list);
5371 /* if we have come back to the start, bail out */
5374 else if (next == first)
5377 } while (vdev_is_dead(next->l2ad_vdev));
5379 /* if we were unable to find any usable vdevs, return NULL */
5380 if (vdev_is_dead(next->l2ad_vdev))
5383 l2arc_dev_last = next;
5386 mutex_exit(&l2arc_dev_mtx);
5389 * Grab the config lock to prevent the 'next' device from being
5390 * removed while we are writing to it.
5393 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
5394 mutex_exit(&spa_namespace_lock);
5400 * Free buffers that were tagged for destruction.
5403 l2arc_do_free_on_write()
5406 l2arc_data_free_t *df, *df_prev;
5408 mutex_enter(&l2arc_free_on_write_mtx);
5409 buflist = l2arc_free_on_write;
5411 for (df = list_tail(buflist); df; df = df_prev) {
5412 df_prev = list_prev(buflist, df);
5413 ASSERT(df->l2df_data != NULL);
5414 ASSERT(df->l2df_func != NULL);
5415 df->l2df_func(df->l2df_data, df->l2df_size);
5416 list_remove(buflist, df);
5417 kmem_free(df, sizeof (l2arc_data_free_t));
5420 mutex_exit(&l2arc_free_on_write_mtx);
5424 * A write to a cache device has completed. Update all headers to allow
5425 * reads from these buffers to begin.
5428 l2arc_write_done(zio_t *zio)
5430 l2arc_write_callback_t *cb;
5433 arc_buf_hdr_t *head, *hdr, *hdr_prev;
5434 kmutex_t *hash_lock;
5435 int64_t bytes_dropped = 0;
5437 cb = zio->io_private;
5439 dev = cb->l2wcb_dev;
5440 ASSERT(dev != NULL);
5441 head = cb->l2wcb_head;
5442 ASSERT(head != NULL);
5443 buflist = &dev->l2ad_buflist;
5444 ASSERT(buflist != NULL);
5445 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
5446 l2arc_write_callback_t *, cb);
5448 if (zio->io_error != 0)
5449 ARCSTAT_BUMP(arcstat_l2_writes_error);
5451 mutex_enter(&dev->l2ad_mtx);
5454 * All writes completed, or an error was hit.
5456 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
5457 hdr_prev = list_prev(buflist, hdr);
5459 hash_lock = HDR_LOCK(hdr);
5460 if (!mutex_tryenter(hash_lock)) {
5462 * This buffer misses out. It may be in a stage
5463 * of eviction. Its ARC_FLAG_L2_WRITING flag will be
5464 * left set, denying reads to this buffer.
5466 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
5471 * It's possible that this buffer got evicted from the L1 cache
5472 * before we grabbed the vdev + hash locks, in which case
5473 * arc_hdr_realloc freed b_tmp_cdata for us if it was allocated.
5474 * Only free the buffer if we still have an L1 hdr.
5476 if (HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_tmp_cdata != NULL &&
5477 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
5478 l2arc_release_cdata_buf(hdr);
5480 if (zio->io_error != 0) {
5482 * Error - drop L2ARC entry.
5484 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev,
5485 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0);
5486 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
5488 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize);
5489 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
5491 bytes_dropped += hdr->b_l2hdr.b_asize;
5492 (void) refcount_remove_many(&dev->l2ad_alloc,
5493 hdr->b_l2hdr.b_asize, hdr);
5497 * Allow ARC to begin reads to this L2ARC entry.
5499 hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
5501 mutex_exit(hash_lock);
5504 atomic_inc_64(&l2arc_writes_done);
5505 list_remove(buflist, head);
5506 ASSERT(!HDR_HAS_L1HDR(head));
5507 kmem_cache_free(hdr_l2only_cache, head);
5508 mutex_exit(&dev->l2ad_mtx);
5510 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
5512 l2arc_do_free_on_write();
5514 kmem_free(cb, sizeof (l2arc_write_callback_t));
5518 * A read to a cache device completed. Validate buffer contents before
5519 * handing over to the regular ARC routines.
5522 l2arc_read_done(zio_t *zio)
5524 l2arc_read_callback_t *cb;
5527 kmutex_t *hash_lock;
5530 ASSERT(zio->io_vd != NULL);
5531 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
5533 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
5535 cb = zio->io_private;
5537 buf = cb->l2rcb_buf;
5538 ASSERT(buf != NULL);
5540 hash_lock = HDR_LOCK(buf->b_hdr);
5541 mutex_enter(hash_lock);
5543 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5546 * If the buffer was compressed, decompress it first.
5548 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
5549 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
5550 ASSERT(zio->io_data != NULL);
5553 * Check this survived the L2ARC journey.
5555 equal = arc_cksum_equal(buf);
5556 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
5557 mutex_exit(hash_lock);
5558 zio->io_private = buf;
5559 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
5560 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
5563 mutex_exit(hash_lock);
5565 * Buffer didn't survive caching. Increment stats and
5566 * reissue to the original storage device.
5568 if (zio->io_error != 0) {
5569 ARCSTAT_BUMP(arcstat_l2_io_error);
5571 zio->io_error = SET_ERROR(EIO);
5574 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
5577 * If there's no waiter, issue an async i/o to the primary
5578 * storage now. If there *is* a waiter, the caller must
5579 * issue the i/o in a context where it's OK to block.
5581 if (zio->io_waiter == NULL) {
5582 zio_t *pio = zio_unique_parent(zio);
5584 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
5586 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
5587 buf->b_data, zio->io_size, arc_read_done, buf,
5588 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
5592 kmem_free(cb, sizeof (l2arc_read_callback_t));
5596 * This is the list priority from which the L2ARC will search for pages to
5597 * cache. This is used within loops (0..3) to cycle through lists in the
5598 * desired order. This order can have a significant effect on cache
5601 * Currently the metadata lists are hit first, MFU then MRU, followed by
5602 * the data lists. This function returns a locked list, and also returns
5606 l2arc_list_locked(int list_num, kmutex_t **lock)
5608 list_t *list = NULL;
5611 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS);
5613 if (list_num < ARC_BUFC_NUMMETADATALISTS) {
5615 list = &arc_mfu->arcs_lists[idx];
5616 *lock = ARCS_LOCK(arc_mfu, idx);
5617 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) {
5618 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
5619 list = &arc_mru->arcs_lists[idx];
5620 *lock = ARCS_LOCK(arc_mru, idx);
5621 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 +
5622 ARC_BUFC_NUMDATALISTS)) {
5623 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
5624 list = &arc_mfu->arcs_lists[idx];
5625 *lock = ARCS_LOCK(arc_mfu, idx);
5627 idx = list_num - ARC_BUFC_NUMLISTS;
5628 list = &arc_mru->arcs_lists[idx];
5629 *lock = ARCS_LOCK(arc_mru, idx);
5632 ASSERT(!(MUTEX_HELD(*lock)));
5638 * Evict buffers from the device write hand to the distance specified in
5639 * bytes. This distance may span populated buffers, it may span nothing.
5640 * This is clearing a region on the L2ARC device ready for writing.
5641 * If the 'all' boolean is set, every buffer is evicted.
5644 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
5647 arc_buf_hdr_t *hdr, *hdr_prev;
5648 kmutex_t *hash_lock;
5651 buflist = &dev->l2ad_buflist;
5653 if (!all && dev->l2ad_first) {
5655 * This is the first sweep through the device. There is
5661 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
5663 * When nearing the end of the device, evict to the end
5664 * before the device write hand jumps to the start.
5666 taddr = dev->l2ad_end;
5668 taddr = dev->l2ad_hand + distance;
5670 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
5671 uint64_t, taddr, boolean_t, all);
5674 mutex_enter(&dev->l2ad_mtx);
5675 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
5676 hdr_prev = list_prev(buflist, hdr);
5678 hash_lock = HDR_LOCK(hdr);
5679 if (!mutex_tryenter(hash_lock)) {
5681 * Missed the hash lock. Retry.
5683 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
5684 mutex_exit(&dev->l2ad_mtx);
5685 mutex_enter(hash_lock);
5686 mutex_exit(hash_lock);
5690 if (HDR_L2_WRITE_HEAD(hdr)) {
5692 * We hit a write head node. Leave it for
5693 * l2arc_write_done().
5695 list_remove(buflist, hdr);
5696 mutex_exit(hash_lock);
5700 if (!all && HDR_HAS_L2HDR(hdr) &&
5701 (hdr->b_l2hdr.b_daddr > taddr ||
5702 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
5704 * We've evicted to the target address,
5705 * or the end of the device.
5707 mutex_exit(hash_lock);
5711 ASSERT(HDR_HAS_L2HDR(hdr));
5712 if (!HDR_HAS_L1HDR(hdr)) {
5713 ASSERT(!HDR_L2_READING(hdr));
5715 * This doesn't exist in the ARC. Destroy.
5716 * arc_hdr_destroy() will call list_remove()
5717 * and decrement arcstat_l2_size.
5719 arc_change_state(arc_anon, hdr, hash_lock);
5720 arc_hdr_destroy(hdr);
5722 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
5723 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
5725 * Invalidate issued or about to be issued
5726 * reads, since we may be about to write
5727 * over this location.
5729 if (HDR_L2_READING(hdr)) {
5730 ARCSTAT_BUMP(arcstat_l2_evict_reading);
5731 hdr->b_flags |= ARC_FLAG_L2_EVICTED;
5734 arc_hdr_l2hdr_destroy(hdr);
5736 mutex_exit(hash_lock);
5738 mutex_exit(&dev->l2ad_mtx);
5742 * Find and write ARC buffers to the L2ARC device.
5744 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
5745 * for reading until they have completed writing.
5746 * The headroom_boost is an in-out parameter used to maintain headroom boost
5747 * state between calls to this function.
5749 * Returns the number of bytes actually written (which may be smaller than
5750 * the delta by which the device hand has changed due to alignment).
5753 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
5754 boolean_t *headroom_boost)
5756 arc_buf_hdr_t *hdr, *hdr_prev, *head;
5758 uint64_t write_asize, write_sz, headroom, buf_compress_minsz;
5760 kmutex_t *list_lock;
5762 l2arc_write_callback_t *cb;
5764 uint64_t guid = spa_load_guid(spa);
5765 const boolean_t do_headroom_boost = *headroom_boost;
5768 ASSERT(dev->l2ad_vdev != NULL);
5770 /* Lower the flag now, we might want to raise it again later. */
5771 *headroom_boost = B_FALSE;
5774 write_sz = write_asize = 0;
5776 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
5777 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD;
5778 head->b_flags |= ARC_FLAG_HAS_L2HDR;
5780 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
5782 * We will want to try to compress buffers that are at least 2x the
5783 * device sector size.
5785 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
5788 * Copy buffers for L2ARC writing.
5790 mutex_enter(&dev->l2ad_mtx);
5791 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) {
5792 uint64_t passed_sz = 0;
5794 list = l2arc_list_locked(try, &list_lock);
5795 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
5798 * L2ARC fast warmup.
5800 * Until the ARC is warm and starts to evict, read from the
5801 * head of the ARC lists rather than the tail.
5803 if (arc_warm == B_FALSE)
5804 hdr = list_head(list);
5806 hdr = list_tail(list);
5808 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
5810 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS;
5811 if (do_headroom_boost)
5812 headroom = (headroom * l2arc_headroom_boost) / 100;
5814 for (; hdr; hdr = hdr_prev) {
5815 kmutex_t *hash_lock;
5819 if (arc_warm == B_FALSE)
5820 hdr_prev = list_next(list, hdr);
5822 hdr_prev = list_prev(list, hdr);
5823 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size);
5825 hash_lock = HDR_LOCK(hdr);
5826 if (!mutex_tryenter(hash_lock)) {
5827 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
5829 * Skip this buffer rather than waiting.
5834 passed_sz += hdr->b_size;
5835 if (passed_sz > headroom) {
5839 mutex_exit(hash_lock);
5840 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
5844 if (!l2arc_write_eligible(guid, hdr)) {
5845 mutex_exit(hash_lock);
5850 * Assume that the buffer is not going to be compressed
5851 * and could take more space on disk because of a larger
5854 buf_sz = hdr->b_size;
5855 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
5857 if ((write_asize + buf_a_sz) > target_sz) {
5859 mutex_exit(hash_lock);
5860 ARCSTAT_BUMP(arcstat_l2_write_full);
5866 * Insert a dummy header on the buflist so
5867 * l2arc_write_done() can find where the
5868 * write buffers begin without searching.
5870 list_insert_head(&dev->l2ad_buflist, head);
5873 sizeof (l2arc_write_callback_t), KM_SLEEP);
5874 cb->l2wcb_dev = dev;
5875 cb->l2wcb_head = head;
5876 pio = zio_root(spa, l2arc_write_done, cb,
5878 ARCSTAT_BUMP(arcstat_l2_write_pios);
5882 * Create and add a new L2ARC header.
5884 hdr->b_l2hdr.b_dev = dev;
5885 hdr->b_flags |= ARC_FLAG_L2_WRITING;
5887 * Temporarily stash the data buffer in b_tmp_cdata.
5888 * The subsequent write step will pick it up from
5889 * there. This is because can't access b_l1hdr.b_buf
5890 * without holding the hash_lock, which we in turn
5891 * can't access without holding the ARC list locks
5892 * (which we want to avoid during compression/writing).
5894 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
5895 hdr->b_l2hdr.b_asize = hdr->b_size;
5896 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data;
5899 * Explicitly set the b_daddr field to a known
5900 * value which means "invalid address". This
5901 * enables us to differentiate which stage of
5902 * l2arc_write_buffers() the particular header
5903 * is in (e.g. this loop, or the one below).
5904 * ARC_FLAG_L2_WRITING is not enough to make
5905 * this distinction, and we need to know in
5906 * order to do proper l2arc vdev accounting in
5907 * arc_release() and arc_hdr_destroy().
5909 * Note, we can't use a new flag to distinguish
5910 * the two stages because we don't hold the
5911 * header's hash_lock below, in the second stage
5912 * of this function. Thus, we can't simply
5913 * change the b_flags field to denote that the
5914 * IO has been sent. We can change the b_daddr
5915 * field of the L2 portion, though, since we'll
5916 * be holding the l2ad_mtx; which is why we're
5917 * using it to denote the header's state change.
5919 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET;
5920 hdr->b_flags |= ARC_FLAG_HAS_L2HDR;
5922 list_insert_head(&dev->l2ad_buflist, hdr);
5925 * Compute and store the buffer cksum before
5926 * writing. On debug the cksum is verified first.
5928 arc_cksum_verify(hdr->b_l1hdr.b_buf);
5929 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE);
5931 mutex_exit(hash_lock);
5934 write_asize += buf_a_sz;
5937 mutex_exit(list_lock);
5943 /* No buffers selected for writing? */
5946 mutex_exit(&dev->l2ad_mtx);
5947 ASSERT(!HDR_HAS_L1HDR(head));
5948 kmem_cache_free(hdr_l2only_cache, head);
5953 * Note that elsewhere in this file arcstat_l2_asize
5954 * and the used space on l2ad_vdev are updated using b_asize,
5955 * which is not necessarily rounded up to the device block size.
5956 * Too keep accounting consistent we do the same here as well:
5957 * stats_size accumulates the sum of b_asize of the written buffers,
5958 * while write_asize accumulates the sum of b_asize rounded up
5959 * to the device block size.
5960 * The latter sum is used only to validate the corectness of the code.
5962 uint64_t stats_size = 0;
5966 * Now start writing the buffers. We're starting at the write head
5967 * and work backwards, retracing the course of the buffer selector
5970 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr;
5971 hdr = list_prev(&dev->l2ad_buflist, hdr)) {
5975 * We shouldn't need to lock the buffer here, since we flagged
5976 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
5977 * take care to only access its L2 cache parameters. In
5978 * particular, hdr->l1hdr.b_buf may be invalid by now due to
5981 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
5983 if ((HDR_L2COMPRESS(hdr)) &&
5984 hdr->b_l2hdr.b_asize >= buf_compress_minsz) {
5985 if (l2arc_compress_buf(hdr)) {
5987 * If compression succeeded, enable headroom
5988 * boost on the next scan cycle.
5990 *headroom_boost = B_TRUE;
5995 * Pick up the buffer data we had previously stashed away
5996 * (and now potentially also compressed).
5998 buf_data = hdr->b_l1hdr.b_tmp_cdata;
5999 buf_sz = hdr->b_l2hdr.b_asize;
6002 * If the data has not been compressed, then clear b_tmp_cdata
6003 * to make sure that it points only to a temporary compression
6006 if (!L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr)))
6007 hdr->b_l1hdr.b_tmp_cdata = NULL;
6010 * We need to do this regardless if buf_sz is zero or
6011 * not, otherwise, when this l2hdr is evicted we'll
6012 * remove a reference that was never added.
6014 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr);
6016 /* Compression may have squashed the buffer to zero length. */
6020 wzio = zio_write_phys(pio, dev->l2ad_vdev,
6021 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
6022 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
6023 ZIO_FLAG_CANFAIL, B_FALSE);
6025 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
6027 (void) zio_nowait(wzio);
6029 stats_size += buf_sz;
6032 * Keep the clock hand suitably device-aligned.
6034 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
6035 write_asize += buf_a_sz;
6036 dev->l2ad_hand += buf_a_sz;
6040 mutex_exit(&dev->l2ad_mtx);
6042 ASSERT3U(write_asize, <=, target_sz);
6043 ARCSTAT_BUMP(arcstat_l2_writes_sent);
6044 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
6045 ARCSTAT_INCR(arcstat_l2_size, write_sz);
6046 ARCSTAT_INCR(arcstat_l2_asize, stats_size);
6047 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0);
6050 * Bump device hand to the device start if it is approaching the end.
6051 * l2arc_evict() will already have evicted ahead for this case.
6053 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
6054 dev->l2ad_hand = dev->l2ad_start;
6055 dev->l2ad_first = B_FALSE;
6058 dev->l2ad_writing = B_TRUE;
6059 (void) zio_wait(pio);
6060 dev->l2ad_writing = B_FALSE;
6062 return (write_asize);
6066 * Compresses an L2ARC buffer.
6067 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6068 * size in l2hdr->b_asize. This routine tries to compress the data and
6069 * depending on the compression result there are three possible outcomes:
6070 * *) The buffer was incompressible. The original l2hdr contents were left
6071 * untouched and are ready for writing to an L2 device.
6072 * *) The buffer was all-zeros, so there is no need to write it to an L2
6073 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6074 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6075 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6076 * data buffer which holds the compressed data to be written, and b_asize
6077 * tells us how much data there is. b_compress is set to the appropriate
6078 * compression algorithm. Once writing is done, invoke
6079 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6081 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6082 * buffer was incompressible).
6085 l2arc_compress_buf(arc_buf_hdr_t *hdr)
6088 size_t csize, len, rounded;
6089 ASSERT(HDR_HAS_L2HDR(hdr));
6090 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
6092 ASSERT(HDR_HAS_L1HDR(hdr));
6093 ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF);
6094 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6096 len = l2hdr->b_asize;
6097 cdata = zio_data_buf_alloc(len);
6098 ASSERT3P(cdata, !=, NULL);
6099 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata,
6100 cdata, l2hdr->b_asize);
6103 /* zero block, indicate that there's nothing to write */
6104 zio_data_buf_free(cdata, len);
6105 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY);
6107 hdr->b_l1hdr.b_tmp_cdata = NULL;
6108 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
6112 rounded = P2ROUNDUP(csize,
6113 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift);
6114 if (rounded < len) {
6116 * Compression succeeded, we'll keep the cdata around for
6117 * writing and release it afterwards.
6119 if (rounded > csize) {
6120 bzero((char *)cdata + csize, rounded - csize);
6123 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4);
6124 l2hdr->b_asize = csize;
6125 hdr->b_l1hdr.b_tmp_cdata = cdata;
6126 ARCSTAT_BUMP(arcstat_l2_compress_successes);
6130 * Compression failed, release the compressed buffer.
6131 * l2hdr will be left unmodified.
6133 zio_data_buf_free(cdata, len);
6134 ARCSTAT_BUMP(arcstat_l2_compress_failures);
6140 * Decompresses a zio read back from an l2arc device. On success, the
6141 * underlying zio's io_data buffer is overwritten by the uncompressed
6142 * version. On decompression error (corrupt compressed stream), the
6143 * zio->io_error value is set to signal an I/O error.
6145 * Please note that the compressed data stream is not checksummed, so
6146 * if the underlying device is experiencing data corruption, we may feed
6147 * corrupt data to the decompressor, so the decompressor needs to be
6148 * able to handle this situation (LZ4 does).
6151 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
6153 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
6155 if (zio->io_error != 0) {
6157 * An io error has occured, just restore the original io
6158 * size in preparation for a main pool read.
6160 zio->io_orig_size = zio->io_size = hdr->b_size;
6164 if (c == ZIO_COMPRESS_EMPTY) {
6166 * An empty buffer results in a null zio, which means we
6167 * need to fill its io_data after we're done restoring the
6168 * buffer's contents.
6170 ASSERT(hdr->b_l1hdr.b_buf != NULL);
6171 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size);
6172 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data;
6174 ASSERT(zio->io_data != NULL);
6176 * We copy the compressed data from the start of the arc buffer
6177 * (the zio_read will have pulled in only what we need, the
6178 * rest is garbage which we will overwrite at decompression)
6179 * and then decompress back to the ARC data buffer. This way we
6180 * can minimize copying by simply decompressing back over the
6181 * original compressed data (rather than decompressing to an
6182 * aux buffer and then copying back the uncompressed buffer,
6183 * which is likely to be much larger).
6188 csize = zio->io_size;
6189 cdata = zio_data_buf_alloc(csize);
6190 bcopy(zio->io_data, cdata, csize);
6191 if (zio_decompress_data(c, cdata, zio->io_data, csize,
6193 zio->io_error = EIO;
6194 zio_data_buf_free(cdata, csize);
6197 /* Restore the expected uncompressed IO size. */
6198 zio->io_orig_size = zio->io_size = hdr->b_size;
6202 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6203 * This buffer serves as a temporary holder of compressed data while
6204 * the buffer entry is being written to an l2arc device. Once that is
6205 * done, we can dispose of it.
6208 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr)
6210 ASSERT(HDR_HAS_L1HDR(hdr));
6211 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_EMPTY) {
6213 * If the data was compressed, then we've allocated a
6214 * temporary buffer for it, so now we need to release it.
6216 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6217 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata,
6219 hdr->b_l1hdr.b_tmp_cdata = NULL;
6221 ASSERT(hdr->b_l1hdr.b_tmp_cdata == NULL);
6226 * This thread feeds the L2ARC at regular intervals. This is the beating
6227 * heart of the L2ARC.
6230 l2arc_feed_thread(void *dummy __unused)
6235 uint64_t size, wrote;
6236 clock_t begin, next = ddi_get_lbolt();
6237 boolean_t headroom_boost = B_FALSE;
6239 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
6241 mutex_enter(&l2arc_feed_thr_lock);
6243 while (l2arc_thread_exit == 0) {
6244 CALLB_CPR_SAFE_BEGIN(&cpr);
6245 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
6246 next - ddi_get_lbolt());
6247 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
6248 next = ddi_get_lbolt() + hz;
6251 * Quick check for L2ARC devices.
6253 mutex_enter(&l2arc_dev_mtx);
6254 if (l2arc_ndev == 0) {
6255 mutex_exit(&l2arc_dev_mtx);
6258 mutex_exit(&l2arc_dev_mtx);
6259 begin = ddi_get_lbolt();
6262 * This selects the next l2arc device to write to, and in
6263 * doing so the next spa to feed from: dev->l2ad_spa. This
6264 * will return NULL if there are now no l2arc devices or if
6265 * they are all faulted.
6267 * If a device is returned, its spa's config lock is also
6268 * held to prevent device removal. l2arc_dev_get_next()
6269 * will grab and release l2arc_dev_mtx.
6271 if ((dev = l2arc_dev_get_next()) == NULL)
6274 spa = dev->l2ad_spa;
6275 ASSERT(spa != NULL);
6278 * If the pool is read-only then force the feed thread to
6279 * sleep a little longer.
6281 if (!spa_writeable(spa)) {
6282 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
6283 spa_config_exit(spa, SCL_L2ARC, dev);
6288 * Avoid contributing to memory pressure.
6290 if (arc_reclaim_needed()) {
6291 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
6292 spa_config_exit(spa, SCL_L2ARC, dev);
6296 ARCSTAT_BUMP(arcstat_l2_feeds);
6298 size = l2arc_write_size();
6301 * Evict L2ARC buffers that will be overwritten.
6303 l2arc_evict(dev, size, B_FALSE);
6306 * Write ARC buffers.
6308 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
6311 * Calculate interval between writes.
6313 next = l2arc_write_interval(begin, size, wrote);
6314 spa_config_exit(spa, SCL_L2ARC, dev);
6317 l2arc_thread_exit = 0;
6318 cv_broadcast(&l2arc_feed_thr_cv);
6319 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
6324 l2arc_vdev_present(vdev_t *vd)
6328 mutex_enter(&l2arc_dev_mtx);
6329 for (dev = list_head(l2arc_dev_list); dev != NULL;
6330 dev = list_next(l2arc_dev_list, dev)) {
6331 if (dev->l2ad_vdev == vd)
6334 mutex_exit(&l2arc_dev_mtx);
6336 return (dev != NULL);
6340 * Add a vdev for use by the L2ARC. By this point the spa has already
6341 * validated the vdev and opened it.
6344 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
6346 l2arc_dev_t *adddev;
6348 ASSERT(!l2arc_vdev_present(vd));
6350 vdev_ashift_optimize(vd);
6353 * Create a new l2arc device entry.
6355 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
6356 adddev->l2ad_spa = spa;
6357 adddev->l2ad_vdev = vd;
6358 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
6359 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
6360 adddev->l2ad_hand = adddev->l2ad_start;
6361 adddev->l2ad_first = B_TRUE;
6362 adddev->l2ad_writing = B_FALSE;
6364 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
6366 * This is a list of all ARC buffers that are still valid on the
6369 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
6370 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
6372 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
6373 refcount_create(&adddev->l2ad_alloc);
6376 * Add device to global list
6378 mutex_enter(&l2arc_dev_mtx);
6379 list_insert_head(l2arc_dev_list, adddev);
6380 atomic_inc_64(&l2arc_ndev);
6381 mutex_exit(&l2arc_dev_mtx);
6385 * Remove a vdev from the L2ARC.
6388 l2arc_remove_vdev(vdev_t *vd)
6390 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
6393 * Find the device by vdev
6395 mutex_enter(&l2arc_dev_mtx);
6396 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
6397 nextdev = list_next(l2arc_dev_list, dev);
6398 if (vd == dev->l2ad_vdev) {
6403 ASSERT(remdev != NULL);
6406 * Remove device from global list
6408 list_remove(l2arc_dev_list, remdev);
6409 l2arc_dev_last = NULL; /* may have been invalidated */
6410 atomic_dec_64(&l2arc_ndev);
6411 mutex_exit(&l2arc_dev_mtx);
6414 * Clear all buflists and ARC references. L2ARC device flush.
6416 l2arc_evict(remdev, 0, B_TRUE);
6417 list_destroy(&remdev->l2ad_buflist);
6418 mutex_destroy(&remdev->l2ad_mtx);
6419 refcount_destroy(&remdev->l2ad_alloc);
6420 kmem_free(remdev, sizeof (l2arc_dev_t));
6426 l2arc_thread_exit = 0;
6428 l2arc_writes_sent = 0;
6429 l2arc_writes_done = 0;
6431 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
6432 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
6433 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
6434 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
6436 l2arc_dev_list = &L2ARC_dev_list;
6437 l2arc_free_on_write = &L2ARC_free_on_write;
6438 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
6439 offsetof(l2arc_dev_t, l2ad_node));
6440 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
6441 offsetof(l2arc_data_free_t, l2df_list_node));
6448 * This is called from dmu_fini(), which is called from spa_fini();
6449 * Because of this, we can assume that all l2arc devices have
6450 * already been removed when the pools themselves were removed.
6453 l2arc_do_free_on_write();
6455 mutex_destroy(&l2arc_feed_thr_lock);
6456 cv_destroy(&l2arc_feed_thr_cv);
6457 mutex_destroy(&l2arc_dev_mtx);
6458 mutex_destroy(&l2arc_free_on_write_mtx);
6460 list_destroy(l2arc_dev_list);
6461 list_destroy(l2arc_free_on_write);
6467 if (!(spa_mode_global & FWRITE))
6470 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
6471 TS_RUN, minclsyspri);
6477 if (!(spa_mode_global & FWRITE))
6480 mutex_enter(&l2arc_feed_thr_lock);
6481 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
6482 l2arc_thread_exit = 1;
6483 while (l2arc_thread_exit != 0)
6484 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
6485 mutex_exit(&l2arc_feed_thr_lock);