4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
30 * DVA-based Adjustable Replacement Cache
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal arc algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
96 * Each arc state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an arc list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
102 * Arc buffers may have an associated eviction callback function.
103 * This function will be invoked prior to removing the buffer (e.g.
104 * in arc_do_user_evicts()). Note however that the data associated
105 * with the buffer may be evicted prior to the callback. The callback
106 * must be made with *no locks held* (to prevent deadlock). Additionally,
107 * the users of callbacks must ensure that their private data is
108 * protected from simultaneous callbacks from arc_clear_callback()
109 * and arc_do_user_evicts().
111 * Note that the majority of the performance stats are manipulated
112 * with atomic operations.
114 * The L2ARC uses the l2ad_mtx on each vdev for the following:
116 * - L2ARC buflist creation
117 * - L2ARC buflist eviction
118 * - L2ARC write completion, which walks L2ARC buflists
119 * - ARC header destruction, as it removes from L2ARC buflists
120 * - ARC header release, as it removes from L2ARC buflists
125 #include <sys/zio_compress.h>
126 #include <sys/zfs_context.h>
128 #include <sys/refcount.h>
129 #include <sys/vdev.h>
130 #include <sys/vdev_impl.h>
131 #include <sys/dsl_pool.h>
132 #include <sys/multilist.h>
134 #include <sys/dnlc.h>
136 #include <sys/callb.h>
137 #include <sys/kstat.h>
138 #include <sys/trim_map.h>
139 #include <zfs_fletcher.h>
142 #include <vm/vm_pageout.h>
143 #include <machine/vmparam.h>
147 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
148 boolean_t arc_watch = B_FALSE;
153 static kmutex_t arc_reclaim_lock;
154 static kcondvar_t arc_reclaim_thread_cv;
155 static boolean_t arc_reclaim_thread_exit;
156 static kcondvar_t arc_reclaim_waiters_cv;
158 static kmutex_t arc_user_evicts_lock;
159 static kcondvar_t arc_user_evicts_cv;
160 static boolean_t arc_user_evicts_thread_exit;
162 uint_t arc_reduce_dnlc_percent = 3;
165 * The number of headers to evict in arc_evict_state_impl() before
166 * dropping the sublist lock and evicting from another sublist. A lower
167 * value means we're more likely to evict the "correct" header (i.e. the
168 * oldest header in the arc state), but comes with higher overhead
169 * (i.e. more invocations of arc_evict_state_impl()).
171 int zfs_arc_evict_batch_limit = 10;
174 * The number of sublists used for each of the arc state lists. If this
175 * is not set to a suitable value by the user, it will be configured to
176 * the number of CPUs on the system in arc_init().
178 int zfs_arc_num_sublists_per_state = 0;
180 /* number of seconds before growing cache again */
181 static int arc_grow_retry = 60;
183 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
184 int zfs_arc_overflow_shift = 8;
186 /* shift of arc_c for calculating both min and max arc_p */
187 static int arc_p_min_shift = 4;
189 /* log2(fraction of arc to reclaim) */
190 static int arc_shrink_shift = 7;
193 * log2(fraction of ARC which must be free to allow growing).
194 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
195 * when reading a new block into the ARC, we will evict an equal-sized block
198 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
199 * we will still not allow it to grow.
201 int arc_no_grow_shift = 5;
205 * minimum lifespan of a prefetch block in clock ticks
206 * (initialized in arc_init())
208 static int arc_min_prefetch_lifespan;
211 * If this percent of memory is free, don't throttle.
213 int arc_lotsfree_percent = 10;
216 extern int zfs_prefetch_disable;
219 * The arc has filled available memory and has now warmed up.
221 static boolean_t arc_warm;
224 * These tunables are for performance analysis.
226 uint64_t zfs_arc_max;
227 uint64_t zfs_arc_min;
228 uint64_t zfs_arc_meta_limit = 0;
229 uint64_t zfs_arc_meta_min = 0;
230 int zfs_arc_grow_retry = 0;
231 int zfs_arc_shrink_shift = 0;
232 int zfs_arc_p_min_shift = 0;
233 int zfs_disable_dup_eviction = 0;
234 uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
235 u_int zfs_arc_free_target = 0;
237 static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS);
238 static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS);
242 arc_free_target_init(void *unused __unused)
245 zfs_arc_free_target = vm_pageout_wakeup_thresh;
247 SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY,
248 arc_free_target_init, NULL);
250 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
251 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
252 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
253 TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min);
254 TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize);
255 TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift);
256 SYSCTL_DECL(_vfs_zfs);
257 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
259 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
261 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN,
262 &zfs_arc_average_blocksize, 0,
263 "ARC average blocksize");
264 SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW,
265 &arc_shrink_shift, 0,
266 "log2(fraction of arc to reclaim)");
269 * We don't have a tunable for arc_free_target due to the dependency on
270 * pagedaemon initialisation.
272 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target,
273 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int),
274 sysctl_vfs_zfs_arc_free_target, "IU",
275 "Desired number of free pages below which ARC triggers reclaim");
278 sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS)
283 val = zfs_arc_free_target;
284 err = sysctl_handle_int(oidp, &val, 0, req);
285 if (err != 0 || req->newptr == NULL)
290 if (val > cnt.v_page_count)
293 zfs_arc_free_target = val;
299 * Must be declared here, before the definition of corresponding kstat
300 * macro which uses the same names will confuse the compiler.
302 SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit,
303 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t),
304 sysctl_vfs_zfs_arc_meta_limit, "QU",
305 "ARC metadata limit");
309 * Note that buffers can be in one of 6 states:
310 * ARC_anon - anonymous (discussed below)
311 * ARC_mru - recently used, currently cached
312 * ARC_mru_ghost - recentely used, no longer in cache
313 * ARC_mfu - frequently used, currently cached
314 * ARC_mfu_ghost - frequently used, no longer in cache
315 * ARC_l2c_only - exists in L2ARC but not other states
316 * When there are no active references to the buffer, they are
317 * are linked onto a list in one of these arc states. These are
318 * the only buffers that can be evicted or deleted. Within each
319 * state there are multiple lists, one for meta-data and one for
320 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
321 * etc.) is tracked separately so that it can be managed more
322 * explicitly: favored over data, limited explicitly.
324 * Anonymous buffers are buffers that are not associated with
325 * a DVA. These are buffers that hold dirty block copies
326 * before they are written to stable storage. By definition,
327 * they are "ref'd" and are considered part of arc_mru
328 * that cannot be freed. Generally, they will aquire a DVA
329 * as they are written and migrate onto the arc_mru list.
331 * The ARC_l2c_only state is for buffers that are in the second
332 * level ARC but no longer in any of the ARC_m* lists. The second
333 * level ARC itself may also contain buffers that are in any of
334 * the ARC_m* states - meaning that a buffer can exist in two
335 * places. The reason for the ARC_l2c_only state is to keep the
336 * buffer header in the hash table, so that reads that hit the
337 * second level ARC benefit from these fast lookups.
340 typedef struct arc_state {
342 * list of evictable buffers
344 multilist_t arcs_list[ARC_BUFC_NUMTYPES];
346 * total amount of evictable data in this state
348 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES];
350 * total amount of data in this state; this includes: evictable,
351 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
353 refcount_t arcs_size;
357 static arc_state_t ARC_anon;
358 static arc_state_t ARC_mru;
359 static arc_state_t ARC_mru_ghost;
360 static arc_state_t ARC_mfu;
361 static arc_state_t ARC_mfu_ghost;
362 static arc_state_t ARC_l2c_only;
364 typedef struct arc_stats {
365 kstat_named_t arcstat_hits;
366 kstat_named_t arcstat_misses;
367 kstat_named_t arcstat_demand_data_hits;
368 kstat_named_t arcstat_demand_data_misses;
369 kstat_named_t arcstat_demand_metadata_hits;
370 kstat_named_t arcstat_demand_metadata_misses;
371 kstat_named_t arcstat_prefetch_data_hits;
372 kstat_named_t arcstat_prefetch_data_misses;
373 kstat_named_t arcstat_prefetch_metadata_hits;
374 kstat_named_t arcstat_prefetch_metadata_misses;
375 kstat_named_t arcstat_mru_hits;
376 kstat_named_t arcstat_mru_ghost_hits;
377 kstat_named_t arcstat_mfu_hits;
378 kstat_named_t arcstat_mfu_ghost_hits;
379 kstat_named_t arcstat_allocated;
380 kstat_named_t arcstat_deleted;
382 * Number of buffers that could not be evicted because the hash lock
383 * was held by another thread. The lock may not necessarily be held
384 * by something using the same buffer, since hash locks are shared
385 * by multiple buffers.
387 kstat_named_t arcstat_mutex_miss;
389 * Number of buffers skipped because they have I/O in progress, are
390 * indrect prefetch buffers that have not lived long enough, or are
391 * not from the spa we're trying to evict from.
393 kstat_named_t arcstat_evict_skip;
395 * Number of times arc_evict_state() was unable to evict enough
396 * buffers to reach it's target amount.
398 kstat_named_t arcstat_evict_not_enough;
399 kstat_named_t arcstat_evict_l2_cached;
400 kstat_named_t arcstat_evict_l2_eligible;
401 kstat_named_t arcstat_evict_l2_ineligible;
402 kstat_named_t arcstat_evict_l2_skip;
403 kstat_named_t arcstat_hash_elements;
404 kstat_named_t arcstat_hash_elements_max;
405 kstat_named_t arcstat_hash_collisions;
406 kstat_named_t arcstat_hash_chains;
407 kstat_named_t arcstat_hash_chain_max;
408 kstat_named_t arcstat_p;
409 kstat_named_t arcstat_c;
410 kstat_named_t arcstat_c_min;
411 kstat_named_t arcstat_c_max;
412 kstat_named_t arcstat_size;
414 * Number of bytes consumed by internal ARC structures necessary
415 * for tracking purposes; these structures are not actually
416 * backed by ARC buffers. This includes arc_buf_hdr_t structures
417 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
418 * caches), and arc_buf_t structures (allocated via arc_buf_t
421 kstat_named_t arcstat_hdr_size;
423 * Number of bytes consumed by ARC buffers of type equal to
424 * ARC_BUFC_DATA. This is generally consumed by buffers backing
425 * on disk user data (e.g. plain file contents).
427 kstat_named_t arcstat_data_size;
429 * Number of bytes consumed by ARC buffers of type equal to
430 * ARC_BUFC_METADATA. This is generally consumed by buffers
431 * backing on disk data that is used for internal ZFS
432 * structures (e.g. ZAP, dnode, indirect blocks, etc).
434 kstat_named_t arcstat_metadata_size;
436 * Number of bytes consumed by various buffers and structures
437 * not actually backed with ARC buffers. This includes bonus
438 * buffers (allocated directly via zio_buf_* functions),
439 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
440 * cache), and dnode_t structures (allocated via dnode_t cache).
442 kstat_named_t arcstat_other_size;
444 * Total number of bytes consumed by ARC buffers residing in the
445 * arc_anon state. This includes *all* buffers in the arc_anon
446 * state; e.g. data, metadata, evictable, and unevictable buffers
447 * are all included in this value.
449 kstat_named_t arcstat_anon_size;
451 * Number of bytes consumed by ARC buffers that meet the
452 * following criteria: backing buffers of type ARC_BUFC_DATA,
453 * residing in the arc_anon state, and are eligible for eviction
454 * (e.g. have no outstanding holds on the buffer).
456 kstat_named_t arcstat_anon_evictable_data;
458 * Number of bytes consumed by ARC buffers that meet the
459 * following criteria: backing buffers of type ARC_BUFC_METADATA,
460 * residing in the arc_anon state, and are eligible for eviction
461 * (e.g. have no outstanding holds on the buffer).
463 kstat_named_t arcstat_anon_evictable_metadata;
465 * Total number of bytes consumed by ARC buffers residing in the
466 * arc_mru state. This includes *all* buffers in the arc_mru
467 * state; e.g. data, metadata, evictable, and unevictable buffers
468 * are all included in this value.
470 kstat_named_t arcstat_mru_size;
472 * Number of bytes consumed by ARC buffers that meet the
473 * following criteria: backing buffers of type ARC_BUFC_DATA,
474 * residing in the arc_mru state, and are eligible for eviction
475 * (e.g. have no outstanding holds on the buffer).
477 kstat_named_t arcstat_mru_evictable_data;
479 * Number of bytes consumed by ARC buffers that meet the
480 * following criteria: backing buffers of type ARC_BUFC_METADATA,
481 * residing in the arc_mru state, and are eligible for eviction
482 * (e.g. have no outstanding holds on the buffer).
484 kstat_named_t arcstat_mru_evictable_metadata;
486 * Total number of bytes that *would have been* consumed by ARC
487 * buffers in the arc_mru_ghost state. The key thing to note
488 * here, is the fact that this size doesn't actually indicate
489 * RAM consumption. The ghost lists only consist of headers and
490 * don't actually have ARC buffers linked off of these headers.
491 * Thus, *if* the headers had associated ARC buffers, these
492 * buffers *would have* consumed this number of bytes.
494 kstat_named_t arcstat_mru_ghost_size;
496 * Number of bytes that *would have been* consumed by ARC
497 * buffers that are eligible for eviction, of type
498 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
500 kstat_named_t arcstat_mru_ghost_evictable_data;
502 * Number of bytes that *would have been* consumed by ARC
503 * buffers that are eligible for eviction, of type
504 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
506 kstat_named_t arcstat_mru_ghost_evictable_metadata;
508 * Total number of bytes consumed by ARC buffers residing in the
509 * arc_mfu state. This includes *all* buffers in the arc_mfu
510 * state; e.g. data, metadata, evictable, and unevictable buffers
511 * are all included in this value.
513 kstat_named_t arcstat_mfu_size;
515 * Number of bytes consumed by ARC buffers that are eligible for
516 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
519 kstat_named_t arcstat_mfu_evictable_data;
521 * Number of bytes consumed by ARC buffers that are eligible for
522 * eviction, of type ARC_BUFC_METADATA, and reside in the
525 kstat_named_t arcstat_mfu_evictable_metadata;
527 * Total number of bytes that *would have been* consumed by ARC
528 * buffers in the arc_mfu_ghost state. See the comment above
529 * arcstat_mru_ghost_size for more details.
531 kstat_named_t arcstat_mfu_ghost_size;
533 * Number of bytes that *would have been* consumed by ARC
534 * buffers that are eligible for eviction, of type
535 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
537 kstat_named_t arcstat_mfu_ghost_evictable_data;
539 * Number of bytes that *would have been* consumed by ARC
540 * buffers that are eligible for eviction, of type
541 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
543 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
544 kstat_named_t arcstat_l2_hits;
545 kstat_named_t arcstat_l2_misses;
546 kstat_named_t arcstat_l2_feeds;
547 kstat_named_t arcstat_l2_rw_clash;
548 kstat_named_t arcstat_l2_read_bytes;
549 kstat_named_t arcstat_l2_write_bytes;
550 kstat_named_t arcstat_l2_writes_sent;
551 kstat_named_t arcstat_l2_writes_done;
552 kstat_named_t arcstat_l2_writes_error;
553 kstat_named_t arcstat_l2_writes_lock_retry;
554 kstat_named_t arcstat_l2_evict_lock_retry;
555 kstat_named_t arcstat_l2_evict_reading;
556 kstat_named_t arcstat_l2_evict_l1cached;
557 kstat_named_t arcstat_l2_free_on_write;
558 kstat_named_t arcstat_l2_cdata_free_on_write;
559 kstat_named_t arcstat_l2_abort_lowmem;
560 kstat_named_t arcstat_l2_cksum_bad;
561 kstat_named_t arcstat_l2_io_error;
562 kstat_named_t arcstat_l2_size;
563 kstat_named_t arcstat_l2_asize;
564 kstat_named_t arcstat_l2_hdr_size;
565 kstat_named_t arcstat_l2_compress_successes;
566 kstat_named_t arcstat_l2_compress_zeros;
567 kstat_named_t arcstat_l2_compress_failures;
568 kstat_named_t arcstat_l2_write_trylock_fail;
569 kstat_named_t arcstat_l2_write_passed_headroom;
570 kstat_named_t arcstat_l2_write_spa_mismatch;
571 kstat_named_t arcstat_l2_write_in_l2;
572 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
573 kstat_named_t arcstat_l2_write_not_cacheable;
574 kstat_named_t arcstat_l2_write_full;
575 kstat_named_t arcstat_l2_write_buffer_iter;
576 kstat_named_t arcstat_l2_write_pios;
577 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
578 kstat_named_t arcstat_l2_write_buffer_list_iter;
579 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
580 kstat_named_t arcstat_memory_throttle_count;
581 kstat_named_t arcstat_duplicate_buffers;
582 kstat_named_t arcstat_duplicate_buffers_size;
583 kstat_named_t arcstat_duplicate_reads;
584 kstat_named_t arcstat_meta_used;
585 kstat_named_t arcstat_meta_limit;
586 kstat_named_t arcstat_meta_max;
587 kstat_named_t arcstat_meta_min;
590 static arc_stats_t arc_stats = {
591 { "hits", KSTAT_DATA_UINT64 },
592 { "misses", KSTAT_DATA_UINT64 },
593 { "demand_data_hits", KSTAT_DATA_UINT64 },
594 { "demand_data_misses", KSTAT_DATA_UINT64 },
595 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
596 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
597 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
598 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
599 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
600 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
601 { "mru_hits", KSTAT_DATA_UINT64 },
602 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
603 { "mfu_hits", KSTAT_DATA_UINT64 },
604 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
605 { "allocated", KSTAT_DATA_UINT64 },
606 { "deleted", KSTAT_DATA_UINT64 },
607 { "mutex_miss", KSTAT_DATA_UINT64 },
608 { "evict_skip", KSTAT_DATA_UINT64 },
609 { "evict_not_enough", KSTAT_DATA_UINT64 },
610 { "evict_l2_cached", KSTAT_DATA_UINT64 },
611 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
612 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
613 { "evict_l2_skip", KSTAT_DATA_UINT64 },
614 { "hash_elements", KSTAT_DATA_UINT64 },
615 { "hash_elements_max", KSTAT_DATA_UINT64 },
616 { "hash_collisions", KSTAT_DATA_UINT64 },
617 { "hash_chains", KSTAT_DATA_UINT64 },
618 { "hash_chain_max", KSTAT_DATA_UINT64 },
619 { "p", KSTAT_DATA_UINT64 },
620 { "c", KSTAT_DATA_UINT64 },
621 { "c_min", KSTAT_DATA_UINT64 },
622 { "c_max", KSTAT_DATA_UINT64 },
623 { "size", KSTAT_DATA_UINT64 },
624 { "hdr_size", KSTAT_DATA_UINT64 },
625 { "data_size", KSTAT_DATA_UINT64 },
626 { "metadata_size", KSTAT_DATA_UINT64 },
627 { "other_size", KSTAT_DATA_UINT64 },
628 { "anon_size", KSTAT_DATA_UINT64 },
629 { "anon_evictable_data", KSTAT_DATA_UINT64 },
630 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
631 { "mru_size", KSTAT_DATA_UINT64 },
632 { "mru_evictable_data", KSTAT_DATA_UINT64 },
633 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
634 { "mru_ghost_size", KSTAT_DATA_UINT64 },
635 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
636 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
637 { "mfu_size", KSTAT_DATA_UINT64 },
638 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
639 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
640 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
641 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
642 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
643 { "l2_hits", KSTAT_DATA_UINT64 },
644 { "l2_misses", KSTAT_DATA_UINT64 },
645 { "l2_feeds", KSTAT_DATA_UINT64 },
646 { "l2_rw_clash", KSTAT_DATA_UINT64 },
647 { "l2_read_bytes", KSTAT_DATA_UINT64 },
648 { "l2_write_bytes", KSTAT_DATA_UINT64 },
649 { "l2_writes_sent", KSTAT_DATA_UINT64 },
650 { "l2_writes_done", KSTAT_DATA_UINT64 },
651 { "l2_writes_error", KSTAT_DATA_UINT64 },
652 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
653 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
654 { "l2_evict_reading", KSTAT_DATA_UINT64 },
655 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
656 { "l2_free_on_write", KSTAT_DATA_UINT64 },
657 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 },
658 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
659 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
660 { "l2_io_error", KSTAT_DATA_UINT64 },
661 { "l2_size", KSTAT_DATA_UINT64 },
662 { "l2_asize", KSTAT_DATA_UINT64 },
663 { "l2_hdr_size", KSTAT_DATA_UINT64 },
664 { "l2_compress_successes", KSTAT_DATA_UINT64 },
665 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
666 { "l2_compress_failures", KSTAT_DATA_UINT64 },
667 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
668 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
669 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
670 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
671 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
672 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
673 { "l2_write_full", KSTAT_DATA_UINT64 },
674 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
675 { "l2_write_pios", KSTAT_DATA_UINT64 },
676 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
677 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
678 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
679 { "memory_throttle_count", KSTAT_DATA_UINT64 },
680 { "duplicate_buffers", KSTAT_DATA_UINT64 },
681 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
682 { "duplicate_reads", KSTAT_DATA_UINT64 },
683 { "arc_meta_used", KSTAT_DATA_UINT64 },
684 { "arc_meta_limit", KSTAT_DATA_UINT64 },
685 { "arc_meta_max", KSTAT_DATA_UINT64 },
686 { "arc_meta_min", KSTAT_DATA_UINT64 }
689 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
691 #define ARCSTAT_INCR(stat, val) \
692 atomic_add_64(&arc_stats.stat.value.ui64, (val))
694 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
695 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
697 #define ARCSTAT_MAX(stat, val) { \
699 while ((val) > (m = arc_stats.stat.value.ui64) && \
700 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
704 #define ARCSTAT_MAXSTAT(stat) \
705 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
708 * We define a macro to allow ARC hits/misses to be easily broken down by
709 * two separate conditions, giving a total of four different subtypes for
710 * each of hits and misses (so eight statistics total).
712 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
715 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
717 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
721 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
723 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
728 static arc_state_t *arc_anon;
729 static arc_state_t *arc_mru;
730 static arc_state_t *arc_mru_ghost;
731 static arc_state_t *arc_mfu;
732 static arc_state_t *arc_mfu_ghost;
733 static arc_state_t *arc_l2c_only;
736 * There are several ARC variables that are critical to export as kstats --
737 * but we don't want to have to grovel around in the kstat whenever we wish to
738 * manipulate them. For these variables, we therefore define them to be in
739 * terms of the statistic variable. This assures that we are not introducing
740 * the possibility of inconsistency by having shadow copies of the variables,
741 * while still allowing the code to be readable.
743 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
744 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
745 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
746 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
747 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
748 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
749 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
750 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
751 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
753 #define L2ARC_IS_VALID_COMPRESS(_c_) \
754 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
756 static int arc_no_grow; /* Don't try to grow cache size */
757 static uint64_t arc_tempreserve;
758 static uint64_t arc_loaned_bytes;
760 typedef struct arc_callback arc_callback_t;
762 struct arc_callback {
764 arc_done_func_t *acb_done;
766 zio_t *acb_zio_dummy;
767 arc_callback_t *acb_next;
770 typedef struct arc_write_callback arc_write_callback_t;
772 struct arc_write_callback {
774 arc_done_func_t *awcb_ready;
775 arc_done_func_t *awcb_physdone;
776 arc_done_func_t *awcb_done;
781 * ARC buffers are separated into multiple structs as a memory saving measure:
782 * - Common fields struct, always defined, and embedded within it:
783 * - L2-only fields, always allocated but undefined when not in L2ARC
784 * - L1-only fields, only allocated when in L1ARC
786 * Buffer in L1 Buffer only in L2
787 * +------------------------+ +------------------------+
788 * | arc_buf_hdr_t | | arc_buf_hdr_t |
792 * +------------------------+ +------------------------+
793 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
794 * | (undefined if L1-only) | | |
795 * +------------------------+ +------------------------+
796 * | l1arc_buf_hdr_t |
801 * +------------------------+
803 * Because it's possible for the L2ARC to become extremely large, we can wind
804 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
805 * is minimized by only allocating the fields necessary for an L1-cached buffer
806 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
807 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
808 * words in pointers. arc_hdr_realloc() is used to switch a header between
809 * these two allocation states.
811 typedef struct l1arc_buf_hdr {
812 kmutex_t b_freeze_lock;
815 * used for debugging wtih kmem_flags - by allocating and freeing
816 * b_thawed when the buffer is thawed, we get a record of the stack
817 * trace that thawed it.
824 /* for waiting on writes to complete */
827 /* protected by arc state mutex */
828 arc_state_t *b_state;
829 multilist_node_t b_arc_node;
831 /* updated atomically */
832 clock_t b_arc_access;
834 /* self protecting */
837 arc_callback_t *b_acb;
838 /* temporary buffer holder for in-flight compressed data */
842 typedef struct l2arc_dev l2arc_dev_t;
844 typedef struct l2arc_buf_hdr {
845 /* protected by arc_buf_hdr mutex */
846 l2arc_dev_t *b_dev; /* L2ARC device */
847 uint64_t b_daddr; /* disk address, offset byte */
848 /* real alloc'd buffer size depending on b_compress applied */
851 list_node_t b_l2node;
855 /* protected by hash lock */
859 * Even though this checksum is only set/verified when a buffer is in
860 * the L1 cache, it needs to be in the set of common fields because it
861 * must be preserved from the time before a buffer is written out to
862 * L2ARC until after it is read back in.
864 zio_cksum_t *b_freeze_cksum;
866 arc_buf_hdr_t *b_hash_next;
873 /* L2ARC fields. Undefined when not in L2ARC. */
874 l2arc_buf_hdr_t b_l2hdr;
875 /* L1ARC fields. Undefined when in l2arc_only state */
876 l1arc_buf_hdr_t b_l1hdr;
881 sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS)
886 val = arc_meta_limit;
887 err = sysctl_handle_64(oidp, &val, 0, req);
888 if (err != 0 || req->newptr == NULL)
891 if (val <= 0 || val > arc_c_max)
894 arc_meta_limit = val;
899 static arc_buf_t *arc_eviction_list;
900 static arc_buf_hdr_t arc_eviction_hdr;
902 #define GHOST_STATE(state) \
903 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
904 (state) == arc_l2c_only)
906 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
907 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
908 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
909 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
910 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
911 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
913 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
914 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
915 #define HDR_L2_READING(hdr) \
916 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
917 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
918 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
919 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
920 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
922 #define HDR_ISTYPE_METADATA(hdr) \
923 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
924 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
926 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
927 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
929 /* For storing compression mode in b_flags */
930 #define HDR_COMPRESS_OFFSET 24
931 #define HDR_COMPRESS_NBITS 7
933 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \
934 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS))
935 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \
936 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp))
942 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
943 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
946 * Hash table routines
949 #define HT_LOCK_PAD CACHE_LINE_SIZE
954 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
958 #define BUF_LOCKS 256
959 typedef struct buf_hash_table {
961 arc_buf_hdr_t **ht_table;
962 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
965 static buf_hash_table_t buf_hash_table;
967 #define BUF_HASH_INDEX(spa, dva, birth) \
968 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
969 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
970 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
971 #define HDR_LOCK(hdr) \
972 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
974 uint64_t zfs_crc64_table[256];
980 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
981 #define L2ARC_HEADROOM 2 /* num of writes */
983 * If we discover during ARC scan any buffers to be compressed, we boost
984 * our headroom for the next scanning cycle by this percentage multiple.
986 #define L2ARC_HEADROOM_BOOST 200
987 #define L2ARC_FEED_SECS 1 /* caching interval secs */
988 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
991 * Used to distinguish headers that are being process by
992 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
993 * address. This can happen when the header is added to the l2arc's list
994 * of buffers to write in the first stage of l2arc_write_buffers(), but
995 * has not yet been written out which happens in the second stage of
996 * l2arc_write_buffers().
998 #define L2ARC_ADDR_UNSET ((uint64_t)(-1))
1000 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1001 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1003 /* L2ARC Performance Tunables */
1004 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1005 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1006 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1007 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1008 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1009 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1010 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1011 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1012 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1014 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
1015 &l2arc_write_max, 0, "max write size");
1016 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
1017 &l2arc_write_boost, 0, "extra write during warmup");
1018 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
1019 &l2arc_headroom, 0, "number of dev writes");
1020 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
1021 &l2arc_feed_secs, 0, "interval seconds");
1022 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
1023 &l2arc_feed_min_ms, 0, "min interval milliseconds");
1025 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
1026 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
1027 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
1028 &l2arc_feed_again, 0, "turbo warmup");
1029 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
1030 &l2arc_norw, 0, "no reads during writes");
1032 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
1033 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state");
1034 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
1035 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
1036 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
1037 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
1039 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
1040 &ARC_mru.arcs_size.rc_count, 0, "size of mru state");
1041 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
1042 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
1043 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
1044 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
1046 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
1047 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state");
1048 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
1049 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
1050 "size of metadata in mru ghost state");
1051 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
1052 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
1053 "size of data in mru ghost state");
1055 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
1056 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state");
1057 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
1058 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
1059 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
1060 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
1062 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
1063 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state");
1064 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
1065 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
1066 "size of metadata in mfu ghost state");
1067 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
1068 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
1069 "size of data in mfu ghost state");
1071 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
1072 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state");
1078 vdev_t *l2ad_vdev; /* vdev */
1079 spa_t *l2ad_spa; /* spa */
1080 uint64_t l2ad_hand; /* next write location */
1081 uint64_t l2ad_start; /* first addr on device */
1082 uint64_t l2ad_end; /* last addr on device */
1083 boolean_t l2ad_first; /* first sweep through */
1084 boolean_t l2ad_writing; /* currently writing */
1085 kmutex_t l2ad_mtx; /* lock for buffer list */
1086 list_t l2ad_buflist; /* buffer list */
1087 list_node_t l2ad_node; /* device list node */
1088 refcount_t l2ad_alloc; /* allocated bytes */
1091 static list_t L2ARC_dev_list; /* device list */
1092 static list_t *l2arc_dev_list; /* device list pointer */
1093 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1094 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1095 static list_t L2ARC_free_on_write; /* free after write buf list */
1096 static list_t *l2arc_free_on_write; /* free after write list ptr */
1097 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1098 static uint64_t l2arc_ndev; /* number of devices */
1100 typedef struct l2arc_read_callback {
1101 arc_buf_t *l2rcb_buf; /* read buffer */
1102 spa_t *l2rcb_spa; /* spa */
1103 blkptr_t l2rcb_bp; /* original blkptr */
1104 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1105 int l2rcb_flags; /* original flags */
1106 enum zio_compress l2rcb_compress; /* applied compress */
1107 } l2arc_read_callback_t;
1109 typedef struct l2arc_write_callback {
1110 l2arc_dev_t *l2wcb_dev; /* device info */
1111 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1112 } l2arc_write_callback_t;
1114 typedef struct l2arc_data_free {
1115 /* protected by l2arc_free_on_write_mtx */
1118 void (*l2df_func)(void *, size_t);
1119 list_node_t l2df_list_node;
1120 } l2arc_data_free_t;
1122 static kmutex_t l2arc_feed_thr_lock;
1123 static kcondvar_t l2arc_feed_thr_cv;
1124 static uint8_t l2arc_thread_exit;
1126 static void arc_get_data_buf(arc_buf_t *);
1127 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1128 static boolean_t arc_is_overflowing();
1129 static void arc_buf_watch(arc_buf_t *);
1131 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1132 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1134 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1135 static void l2arc_read_done(zio_t *);
1137 static boolean_t l2arc_compress_buf(arc_buf_hdr_t *);
1138 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress);
1139 static void l2arc_release_cdata_buf(arc_buf_hdr_t *);
1142 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1144 uint8_t *vdva = (uint8_t *)dva;
1145 uint64_t crc = -1ULL;
1148 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1150 for (i = 0; i < sizeof (dva_t); i++)
1151 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1153 crc ^= (spa>>8) ^ birth;
1158 #define BUF_EMPTY(buf) \
1159 ((buf)->b_dva.dva_word[0] == 0 && \
1160 (buf)->b_dva.dva_word[1] == 0)
1162 #define BUF_EQUAL(spa, dva, birth, buf) \
1163 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1164 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1165 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
1168 buf_discard_identity(arc_buf_hdr_t *hdr)
1170 hdr->b_dva.dva_word[0] = 0;
1171 hdr->b_dva.dva_word[1] = 0;
1175 static arc_buf_hdr_t *
1176 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1178 const dva_t *dva = BP_IDENTITY(bp);
1179 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1180 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1181 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1184 mutex_enter(hash_lock);
1185 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1186 hdr = hdr->b_hash_next) {
1187 if (BUF_EQUAL(spa, dva, birth, hdr)) {
1192 mutex_exit(hash_lock);
1198 * Insert an entry into the hash table. If there is already an element
1199 * equal to elem in the hash table, then the already existing element
1200 * will be returned and the new element will not be inserted.
1201 * Otherwise returns NULL.
1202 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1204 static arc_buf_hdr_t *
1205 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1207 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1208 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1209 arc_buf_hdr_t *fhdr;
1212 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1213 ASSERT(hdr->b_birth != 0);
1214 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1216 if (lockp != NULL) {
1218 mutex_enter(hash_lock);
1220 ASSERT(MUTEX_HELD(hash_lock));
1223 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1224 fhdr = fhdr->b_hash_next, i++) {
1225 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1229 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1230 buf_hash_table.ht_table[idx] = hdr;
1231 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
1233 /* collect some hash table performance data */
1235 ARCSTAT_BUMP(arcstat_hash_collisions);
1237 ARCSTAT_BUMP(arcstat_hash_chains);
1239 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1242 ARCSTAT_BUMP(arcstat_hash_elements);
1243 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1249 buf_hash_remove(arc_buf_hdr_t *hdr)
1251 arc_buf_hdr_t *fhdr, **hdrp;
1252 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1254 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1255 ASSERT(HDR_IN_HASH_TABLE(hdr));
1257 hdrp = &buf_hash_table.ht_table[idx];
1258 while ((fhdr = *hdrp) != hdr) {
1259 ASSERT(fhdr != NULL);
1260 hdrp = &fhdr->b_hash_next;
1262 *hdrp = hdr->b_hash_next;
1263 hdr->b_hash_next = NULL;
1264 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE;
1266 /* collect some hash table performance data */
1267 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1269 if (buf_hash_table.ht_table[idx] &&
1270 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1271 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1275 * Global data structures and functions for the buf kmem cache.
1277 static kmem_cache_t *hdr_full_cache;
1278 static kmem_cache_t *hdr_l2only_cache;
1279 static kmem_cache_t *buf_cache;
1286 kmem_free(buf_hash_table.ht_table,
1287 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1288 for (i = 0; i < BUF_LOCKS; i++)
1289 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1290 kmem_cache_destroy(hdr_full_cache);
1291 kmem_cache_destroy(hdr_l2only_cache);
1292 kmem_cache_destroy(buf_cache);
1296 * Constructor callback - called when the cache is empty
1297 * and a new buf is requested.
1301 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1303 arc_buf_hdr_t *hdr = vbuf;
1305 bzero(hdr, HDR_FULL_SIZE);
1306 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1307 refcount_create(&hdr->b_l1hdr.b_refcnt);
1308 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1309 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1310 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1317 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1319 arc_buf_hdr_t *hdr = vbuf;
1321 bzero(hdr, HDR_L2ONLY_SIZE);
1322 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1329 buf_cons(void *vbuf, void *unused, int kmflag)
1331 arc_buf_t *buf = vbuf;
1333 bzero(buf, sizeof (arc_buf_t));
1334 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1335 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1341 * Destructor callback - called when a cached buf is
1342 * no longer required.
1346 hdr_full_dest(void *vbuf, void *unused)
1348 arc_buf_hdr_t *hdr = vbuf;
1350 ASSERT(BUF_EMPTY(hdr));
1351 cv_destroy(&hdr->b_l1hdr.b_cv);
1352 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1353 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1354 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1355 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1360 hdr_l2only_dest(void *vbuf, void *unused)
1362 arc_buf_hdr_t *hdr = vbuf;
1364 ASSERT(BUF_EMPTY(hdr));
1365 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1370 buf_dest(void *vbuf, void *unused)
1372 arc_buf_t *buf = vbuf;
1374 mutex_destroy(&buf->b_evict_lock);
1375 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1379 * Reclaim callback -- invoked when memory is low.
1383 hdr_recl(void *unused)
1385 dprintf("hdr_recl called\n");
1387 * umem calls the reclaim func when we destroy the buf cache,
1388 * which is after we do arc_fini().
1391 cv_signal(&arc_reclaim_thread_cv);
1398 uint64_t hsize = 1ULL << 12;
1402 * The hash table is big enough to fill all of physical memory
1403 * with an average block size of zfs_arc_average_blocksize (default 8K).
1404 * By default, the table will take up
1405 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1407 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE)
1410 buf_hash_table.ht_mask = hsize - 1;
1411 buf_hash_table.ht_table =
1412 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1413 if (buf_hash_table.ht_table == NULL) {
1414 ASSERT(hsize > (1ULL << 8));
1419 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1420 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1421 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1422 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1424 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1425 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1427 for (i = 0; i < 256; i++)
1428 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1429 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1431 for (i = 0; i < BUF_LOCKS; i++) {
1432 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1433 NULL, MUTEX_DEFAULT, NULL);
1438 * Transition between the two allocation states for the arc_buf_hdr struct.
1439 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1440 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1441 * version is used when a cache buffer is only in the L2ARC in order to reduce
1444 static arc_buf_hdr_t *
1445 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
1447 ASSERT(HDR_HAS_L2HDR(hdr));
1449 arc_buf_hdr_t *nhdr;
1450 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1452 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
1453 (old == hdr_l2only_cache && new == hdr_full_cache));
1455 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
1457 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
1458 buf_hash_remove(hdr);
1460 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
1462 if (new == hdr_full_cache) {
1463 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1465 * arc_access and arc_change_state need to be aware that a
1466 * header has just come out of L2ARC, so we set its state to
1467 * l2c_only even though it's about to change.
1469 nhdr->b_l1hdr.b_state = arc_l2c_only;
1471 /* Verify previous threads set to NULL before freeing */
1472 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1474 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1475 ASSERT0(hdr->b_l1hdr.b_datacnt);
1478 * If we've reached here, We must have been called from
1479 * arc_evict_hdr(), as such we should have already been
1480 * removed from any ghost list we were previously on
1481 * (which protects us from racing with arc_evict_state),
1482 * thus no locking is needed during this check.
1484 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1487 * A buffer must not be moved into the arc_l2c_only
1488 * state if it's not finished being written out to the
1489 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1490 * might try to be accessed, even though it was removed.
1492 VERIFY(!HDR_L2_WRITING(hdr));
1493 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1495 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR;
1498 * The header has been reallocated so we need to re-insert it into any
1501 (void) buf_hash_insert(nhdr, NULL);
1503 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
1505 mutex_enter(&dev->l2ad_mtx);
1508 * We must place the realloc'ed header back into the list at
1509 * the same spot. Otherwise, if it's placed earlier in the list,
1510 * l2arc_write_buffers() could find it during the function's
1511 * write phase, and try to write it out to the l2arc.
1513 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
1514 list_remove(&dev->l2ad_buflist, hdr);
1516 mutex_exit(&dev->l2ad_mtx);
1519 * Since we're using the pointer address as the tag when
1520 * incrementing and decrementing the l2ad_alloc refcount, we
1521 * must remove the old pointer (that we're about to destroy) and
1522 * add the new pointer to the refcount. Otherwise we'd remove
1523 * the wrong pointer address when calling arc_hdr_destroy() later.
1526 (void) refcount_remove_many(&dev->l2ad_alloc,
1527 hdr->b_l2hdr.b_asize, hdr);
1529 (void) refcount_add_many(&dev->l2ad_alloc,
1530 nhdr->b_l2hdr.b_asize, nhdr);
1532 buf_discard_identity(hdr);
1533 hdr->b_freeze_cksum = NULL;
1534 kmem_cache_free(old, hdr);
1540 #define ARC_MINTIME (hz>>4) /* 62 ms */
1543 arc_cksum_verify(arc_buf_t *buf)
1547 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1550 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1551 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) {
1552 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1555 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1556 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1557 panic("buffer modified while frozen!");
1558 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1562 arc_cksum_equal(arc_buf_t *buf)
1567 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1568 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1569 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1570 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1576 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1578 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1581 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1582 if (buf->b_hdr->b_freeze_cksum != NULL) {
1583 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1586 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1587 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1588 buf->b_hdr->b_freeze_cksum);
1589 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1592 #endif /* illumos */
1597 typedef struct procctl {
1605 arc_buf_unwatch(arc_buf_t *buf)
1612 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1613 ctl.prwatch.pr_size = 0;
1614 ctl.prwatch.pr_wflags = 0;
1615 result = write(arc_procfd, &ctl, sizeof (ctl));
1616 ASSERT3U(result, ==, sizeof (ctl));
1623 arc_buf_watch(arc_buf_t *buf)
1630 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1631 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1632 ctl.prwatch.pr_wflags = WA_WRITE;
1633 result = write(arc_procfd, &ctl, sizeof (ctl));
1634 ASSERT3U(result, ==, sizeof (ctl));
1638 #endif /* illumos */
1640 static arc_buf_contents_t
1641 arc_buf_type(arc_buf_hdr_t *hdr)
1643 if (HDR_ISTYPE_METADATA(hdr)) {
1644 return (ARC_BUFC_METADATA);
1646 return (ARC_BUFC_DATA);
1651 arc_bufc_to_flags(arc_buf_contents_t type)
1655 /* metadata field is 0 if buffer contains normal data */
1657 case ARC_BUFC_METADATA:
1658 return (ARC_FLAG_BUFC_METADATA);
1662 panic("undefined ARC buffer type!");
1663 return ((uint32_t)-1);
1667 arc_buf_thaw(arc_buf_t *buf)
1669 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1670 if (buf->b_hdr->b_l1hdr.b_state != arc_anon)
1671 panic("modifying non-anon buffer!");
1672 if (HDR_IO_IN_PROGRESS(buf->b_hdr))
1673 panic("modifying buffer while i/o in progress!");
1674 arc_cksum_verify(buf);
1677 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1678 if (buf->b_hdr->b_freeze_cksum != NULL) {
1679 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1680 buf->b_hdr->b_freeze_cksum = NULL;
1684 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1685 if (buf->b_hdr->b_l1hdr.b_thawed != NULL)
1686 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1);
1687 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1691 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1694 arc_buf_unwatch(buf);
1695 #endif /* illumos */
1699 arc_buf_freeze(arc_buf_t *buf)
1701 kmutex_t *hash_lock;
1703 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1706 hash_lock = HDR_LOCK(buf->b_hdr);
1707 mutex_enter(hash_lock);
1709 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1710 buf->b_hdr->b_l1hdr.b_state == arc_anon);
1711 arc_cksum_compute(buf, B_FALSE);
1712 mutex_exit(hash_lock);
1717 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1719 ASSERT(HDR_HAS_L1HDR(hdr));
1720 ASSERT(MUTEX_HELD(hash_lock));
1721 arc_state_t *state = hdr->b_l1hdr.b_state;
1723 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
1724 (state != arc_anon)) {
1725 /* We don't use the L2-only state list. */
1726 if (state != arc_l2c_only) {
1727 arc_buf_contents_t type = arc_buf_type(hdr);
1728 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt;
1729 multilist_t *list = &state->arcs_list[type];
1730 uint64_t *size = &state->arcs_lsize[type];
1732 multilist_remove(list, hdr);
1734 if (GHOST_STATE(state)) {
1735 ASSERT0(hdr->b_l1hdr.b_datacnt);
1736 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
1737 delta = hdr->b_size;
1740 ASSERT3U(*size, >=, delta);
1741 atomic_add_64(size, -delta);
1743 /* remove the prefetch flag if we get a reference */
1744 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
1749 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1752 arc_state_t *state = hdr->b_l1hdr.b_state;
1754 ASSERT(HDR_HAS_L1HDR(hdr));
1755 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1756 ASSERT(!GHOST_STATE(state));
1759 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1760 * check to prevent usage of the arc_l2c_only list.
1762 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
1763 (state != arc_anon)) {
1764 arc_buf_contents_t type = arc_buf_type(hdr);
1765 multilist_t *list = &state->arcs_list[type];
1766 uint64_t *size = &state->arcs_lsize[type];
1768 multilist_insert(list, hdr);
1770 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
1771 atomic_add_64(size, hdr->b_size *
1772 hdr->b_l1hdr.b_datacnt);
1778 * Move the supplied buffer to the indicated state. The hash lock
1779 * for the buffer must be held by the caller.
1782 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
1783 kmutex_t *hash_lock)
1785 arc_state_t *old_state;
1788 uint64_t from_delta, to_delta;
1789 arc_buf_contents_t buftype = arc_buf_type(hdr);
1792 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1793 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1794 * L1 hdr doesn't always exist when we change state to arc_anon before
1795 * destroying a header, in which case reallocating to add the L1 hdr is
1798 if (HDR_HAS_L1HDR(hdr)) {
1799 old_state = hdr->b_l1hdr.b_state;
1800 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
1801 datacnt = hdr->b_l1hdr.b_datacnt;
1803 old_state = arc_l2c_only;
1808 ASSERT(MUTEX_HELD(hash_lock));
1809 ASSERT3P(new_state, !=, old_state);
1810 ASSERT(refcnt == 0 || datacnt > 0);
1811 ASSERT(!GHOST_STATE(new_state) || datacnt == 0);
1812 ASSERT(old_state != arc_anon || datacnt <= 1);
1814 from_delta = to_delta = datacnt * hdr->b_size;
1817 * If this buffer is evictable, transfer it from the
1818 * old state list to the new state list.
1821 if (old_state != arc_anon && old_state != arc_l2c_only) {
1822 uint64_t *size = &old_state->arcs_lsize[buftype];
1824 ASSERT(HDR_HAS_L1HDR(hdr));
1825 multilist_remove(&old_state->arcs_list[buftype], hdr);
1828 * If prefetching out of the ghost cache,
1829 * we will have a non-zero datacnt.
1831 if (GHOST_STATE(old_state) && datacnt == 0) {
1832 /* ghost elements have a ghost size */
1833 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1834 from_delta = hdr->b_size;
1836 ASSERT3U(*size, >=, from_delta);
1837 atomic_add_64(size, -from_delta);
1839 if (new_state != arc_anon && new_state != arc_l2c_only) {
1840 uint64_t *size = &new_state->arcs_lsize[buftype];
1843 * An L1 header always exists here, since if we're
1844 * moving to some L1-cached state (i.e. not l2c_only or
1845 * anonymous), we realloc the header to add an L1hdr
1848 ASSERT(HDR_HAS_L1HDR(hdr));
1849 multilist_insert(&new_state->arcs_list[buftype], hdr);
1851 /* ghost elements have a ghost size */
1852 if (GHOST_STATE(new_state)) {
1854 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1855 to_delta = hdr->b_size;
1857 atomic_add_64(size, to_delta);
1861 ASSERT(!BUF_EMPTY(hdr));
1862 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
1863 buf_hash_remove(hdr);
1865 /* adjust state sizes (ignore arc_l2c_only) */
1867 if (to_delta && new_state != arc_l2c_only) {
1868 ASSERT(HDR_HAS_L1HDR(hdr));
1869 if (GHOST_STATE(new_state)) {
1873 * We moving a header to a ghost state, we first
1874 * remove all arc buffers. Thus, we'll have a
1875 * datacnt of zero, and no arc buffer to use for
1876 * the reference. As a result, we use the arc
1877 * header pointer for the reference.
1879 (void) refcount_add_many(&new_state->arcs_size,
1882 ASSERT3U(datacnt, !=, 0);
1885 * Each individual buffer holds a unique reference,
1886 * thus we must remove each of these references one
1889 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1890 buf = buf->b_next) {
1891 (void) refcount_add_many(&new_state->arcs_size,
1897 if (from_delta && old_state != arc_l2c_only) {
1898 ASSERT(HDR_HAS_L1HDR(hdr));
1899 if (GHOST_STATE(old_state)) {
1901 * When moving a header off of a ghost state,
1902 * there's the possibility for datacnt to be
1903 * non-zero. This is because we first add the
1904 * arc buffer to the header prior to changing
1905 * the header's state. Since we used the header
1906 * for the reference when putting the header on
1907 * the ghost state, we must balance that and use
1908 * the header when removing off the ghost state
1909 * (even though datacnt is non zero).
1912 IMPLY(datacnt == 0, new_state == arc_anon ||
1913 new_state == arc_l2c_only);
1915 (void) refcount_remove_many(&old_state->arcs_size,
1918 ASSERT3P(datacnt, !=, 0);
1921 * Each individual buffer holds a unique reference,
1922 * thus we must remove each of these references one
1925 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1926 buf = buf->b_next) {
1927 (void) refcount_remove_many(
1928 &old_state->arcs_size, hdr->b_size, buf);
1933 if (HDR_HAS_L1HDR(hdr))
1934 hdr->b_l1hdr.b_state = new_state;
1937 * L2 headers should never be on the L2 state list since they don't
1938 * have L1 headers allocated.
1940 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
1941 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
1945 arc_space_consume(uint64_t space, arc_space_type_t type)
1947 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1950 case ARC_SPACE_DATA:
1951 ARCSTAT_INCR(arcstat_data_size, space);
1953 case ARC_SPACE_META:
1954 ARCSTAT_INCR(arcstat_metadata_size, space);
1956 case ARC_SPACE_OTHER:
1957 ARCSTAT_INCR(arcstat_other_size, space);
1959 case ARC_SPACE_HDRS:
1960 ARCSTAT_INCR(arcstat_hdr_size, space);
1962 case ARC_SPACE_L2HDRS:
1963 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1967 if (type != ARC_SPACE_DATA)
1968 ARCSTAT_INCR(arcstat_meta_used, space);
1970 atomic_add_64(&arc_size, space);
1974 arc_space_return(uint64_t space, arc_space_type_t type)
1976 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1979 case ARC_SPACE_DATA:
1980 ARCSTAT_INCR(arcstat_data_size, -space);
1982 case ARC_SPACE_META:
1983 ARCSTAT_INCR(arcstat_metadata_size, -space);
1985 case ARC_SPACE_OTHER:
1986 ARCSTAT_INCR(arcstat_other_size, -space);
1988 case ARC_SPACE_HDRS:
1989 ARCSTAT_INCR(arcstat_hdr_size, -space);
1991 case ARC_SPACE_L2HDRS:
1992 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1996 if (type != ARC_SPACE_DATA) {
1997 ASSERT(arc_meta_used >= space);
1998 if (arc_meta_max < arc_meta_used)
1999 arc_meta_max = arc_meta_used;
2000 ARCSTAT_INCR(arcstat_meta_used, -space);
2003 ASSERT(arc_size >= space);
2004 atomic_add_64(&arc_size, -space);
2008 arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
2013 ASSERT3U(size, >, 0);
2014 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
2015 ASSERT(BUF_EMPTY(hdr));
2016 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
2018 hdr->b_spa = spa_load_guid(spa);
2020 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2023 buf->b_efunc = NULL;
2024 buf->b_private = NULL;
2027 hdr->b_flags = arc_bufc_to_flags(type);
2028 hdr->b_flags |= ARC_FLAG_HAS_L1HDR;
2030 hdr->b_l1hdr.b_buf = buf;
2031 hdr->b_l1hdr.b_state = arc_anon;
2032 hdr->b_l1hdr.b_arc_access = 0;
2033 hdr->b_l1hdr.b_datacnt = 1;
2034 hdr->b_l1hdr.b_tmp_cdata = NULL;
2036 arc_get_data_buf(buf);
2037 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2038 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2043 static char *arc_onloan_tag = "onloan";
2046 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2047 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2048 * buffers must be returned to the arc before they can be used by the DMU or
2052 arc_loan_buf(spa_t *spa, int size)
2056 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
2058 atomic_add_64(&arc_loaned_bytes, size);
2063 * Return a loaned arc buffer to the arc.
2066 arc_return_buf(arc_buf_t *buf, void *tag)
2068 arc_buf_hdr_t *hdr = buf->b_hdr;
2070 ASSERT(buf->b_data != NULL);
2071 ASSERT(HDR_HAS_L1HDR(hdr));
2072 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2073 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2075 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
2078 /* Detach an arc_buf from a dbuf (tag) */
2080 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2082 arc_buf_hdr_t *hdr = buf->b_hdr;
2084 ASSERT(buf->b_data != NULL);
2085 ASSERT(HDR_HAS_L1HDR(hdr));
2086 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2087 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2088 buf->b_efunc = NULL;
2089 buf->b_private = NULL;
2091 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
2095 arc_buf_clone(arc_buf_t *from)
2098 arc_buf_hdr_t *hdr = from->b_hdr;
2099 uint64_t size = hdr->b_size;
2101 ASSERT(HDR_HAS_L1HDR(hdr));
2102 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2104 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2107 buf->b_efunc = NULL;
2108 buf->b_private = NULL;
2109 buf->b_next = hdr->b_l1hdr.b_buf;
2110 hdr->b_l1hdr.b_buf = buf;
2111 arc_get_data_buf(buf);
2112 bcopy(from->b_data, buf->b_data, size);
2115 * This buffer already exists in the arc so create a duplicate
2116 * copy for the caller. If the buffer is associated with user data
2117 * then track the size and number of duplicates. These stats will be
2118 * updated as duplicate buffers are created and destroyed.
2120 if (HDR_ISTYPE_DATA(hdr)) {
2121 ARCSTAT_BUMP(arcstat_duplicate_buffers);
2122 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
2124 hdr->b_l1hdr.b_datacnt += 1;
2129 arc_buf_add_ref(arc_buf_t *buf, void* tag)
2132 kmutex_t *hash_lock;
2135 * Check to see if this buffer is evicted. Callers
2136 * must verify b_data != NULL to know if the add_ref
2139 mutex_enter(&buf->b_evict_lock);
2140 if (buf->b_data == NULL) {
2141 mutex_exit(&buf->b_evict_lock);
2144 hash_lock = HDR_LOCK(buf->b_hdr);
2145 mutex_enter(hash_lock);
2147 ASSERT(HDR_HAS_L1HDR(hdr));
2148 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2149 mutex_exit(&buf->b_evict_lock);
2151 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
2152 hdr->b_l1hdr.b_state == arc_mfu);
2154 add_reference(hdr, hash_lock, tag);
2155 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2156 arc_access(hdr, hash_lock);
2157 mutex_exit(hash_lock);
2158 ARCSTAT_BUMP(arcstat_hits);
2159 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
2160 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
2161 data, metadata, hits);
2165 arc_buf_free_on_write(void *data, size_t size,
2166 void (*free_func)(void *, size_t))
2168 l2arc_data_free_t *df;
2170 df = kmem_alloc(sizeof (*df), KM_SLEEP);
2171 df->l2df_data = data;
2172 df->l2df_size = size;
2173 df->l2df_func = free_func;
2174 mutex_enter(&l2arc_free_on_write_mtx);
2175 list_insert_head(l2arc_free_on_write, df);
2176 mutex_exit(&l2arc_free_on_write_mtx);
2180 * Free the arc data buffer. If it is an l2arc write in progress,
2181 * the buffer is placed on l2arc_free_on_write to be freed later.
2184 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
2186 arc_buf_hdr_t *hdr = buf->b_hdr;
2188 if (HDR_L2_WRITING(hdr)) {
2189 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
2190 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2192 free_func(buf->b_data, hdr->b_size);
2197 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
2199 ASSERT(HDR_HAS_L2HDR(hdr));
2200 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx));
2203 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
2204 * that doesn't exist, the header is in the arc_l2c_only state,
2205 * and there isn't anything to free (it's already been freed).
2207 if (!HDR_HAS_L1HDR(hdr))
2211 * The header isn't being written to the l2arc device, thus it
2212 * shouldn't have a b_tmp_cdata to free.
2214 if (!HDR_L2_WRITING(hdr)) {
2215 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2220 * The header does not have compression enabled. This can be due
2221 * to the buffer not being compressible, or because we're
2222 * freeing the buffer before the second phase of
2223 * l2arc_write_buffer() has started (which does the compression
2224 * step). In either case, b_tmp_cdata does not point to a
2225 * separately compressed buffer, so there's nothing to free (it
2226 * points to the same buffer as the arc_buf_t's b_data field).
2228 if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) {
2229 hdr->b_l1hdr.b_tmp_cdata = NULL;
2234 * There's nothing to free since the buffer was all zero's and
2235 * compressed to a zero length buffer.
2237 if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_EMPTY) {
2238 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2242 ASSERT(L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr)));
2244 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata,
2245 hdr->b_size, zio_data_buf_free);
2247 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
2248 hdr->b_l1hdr.b_tmp_cdata = NULL;
2252 * Free up buf->b_data and if 'remove' is set, then pull the
2253 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2256 arc_buf_destroy(arc_buf_t *buf, boolean_t remove)
2260 /* free up data associated with the buf */
2261 if (buf->b_data != NULL) {
2262 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
2263 uint64_t size = buf->b_hdr->b_size;
2264 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
2266 arc_cksum_verify(buf);
2268 arc_buf_unwatch(buf);
2269 #endif /* illumos */
2271 if (type == ARC_BUFC_METADATA) {
2272 arc_buf_data_free(buf, zio_buf_free);
2273 arc_space_return(size, ARC_SPACE_META);
2275 ASSERT(type == ARC_BUFC_DATA);
2276 arc_buf_data_free(buf, zio_data_buf_free);
2277 arc_space_return(size, ARC_SPACE_DATA);
2280 /* protected by hash lock, if in the hash table */
2281 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) {
2282 uint64_t *cnt = &state->arcs_lsize[type];
2284 ASSERT(refcount_is_zero(
2285 &buf->b_hdr->b_l1hdr.b_refcnt));
2286 ASSERT(state != arc_anon && state != arc_l2c_only);
2288 ASSERT3U(*cnt, >=, size);
2289 atomic_add_64(cnt, -size);
2292 (void) refcount_remove_many(&state->arcs_size, size, buf);
2296 * If we're destroying a duplicate buffer make sure
2297 * that the appropriate statistics are updated.
2299 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 &&
2300 HDR_ISTYPE_DATA(buf->b_hdr)) {
2301 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
2302 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
2304 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0);
2305 buf->b_hdr->b_l1hdr.b_datacnt -= 1;
2308 /* only remove the buf if requested */
2312 /* remove the buf from the hdr list */
2313 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf;
2314 bufp = &(*bufp)->b_next)
2316 *bufp = buf->b_next;
2319 ASSERT(buf->b_efunc == NULL);
2321 /* clean up the buf */
2323 kmem_cache_free(buf_cache, buf);
2327 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2329 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2330 l2arc_dev_t *dev = l2hdr->b_dev;
2332 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2333 ASSERT(HDR_HAS_L2HDR(hdr));
2335 list_remove(&dev->l2ad_buflist, hdr);
2338 * We don't want to leak the b_tmp_cdata buffer that was
2339 * allocated in l2arc_write_buffers()
2341 arc_buf_l2_cdata_free(hdr);
2344 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2345 * this header is being processed by l2arc_write_buffers() (i.e.
2346 * it's in the first stage of l2arc_write_buffers()).
2347 * Re-affirming that truth here, just to serve as a reminder. If
2348 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2349 * may not have its HDR_L2_WRITING flag set. (the write may have
2350 * completed, in which case HDR_L2_WRITING will be false and the
2351 * b_daddr field will point to the address of the buffer on disk).
2353 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr));
2356 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2357 * l2arc_write_buffers(). Since we've just removed this header
2358 * from the l2arc buffer list, this header will never reach the
2359 * second stage of l2arc_write_buffers(), which increments the
2360 * accounting stats for this header. Thus, we must be careful
2361 * not to decrement them for this header either.
2363 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) {
2364 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
2365 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
2367 vdev_space_update(dev->l2ad_vdev,
2368 -l2hdr->b_asize, 0, 0);
2370 (void) refcount_remove_many(&dev->l2ad_alloc,
2371 l2hdr->b_asize, hdr);
2374 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
2378 arc_hdr_destroy(arc_buf_hdr_t *hdr)
2380 if (HDR_HAS_L1HDR(hdr)) {
2381 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
2382 hdr->b_l1hdr.b_datacnt > 0);
2383 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2384 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2386 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2387 ASSERT(!HDR_IN_HASH_TABLE(hdr));
2389 if (HDR_HAS_L2HDR(hdr)) {
2390 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2391 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
2394 mutex_enter(&dev->l2ad_mtx);
2397 * Even though we checked this conditional above, we
2398 * need to check this again now that we have the
2399 * l2ad_mtx. This is because we could be racing with
2400 * another thread calling l2arc_evict() which might have
2401 * destroyed this header's L2 portion as we were waiting
2402 * to acquire the l2ad_mtx. If that happens, we don't
2403 * want to re-destroy the header's L2 portion.
2405 if (HDR_HAS_L2HDR(hdr)) {
2406 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET)
2407 trim_map_free(dev->l2ad_vdev,
2408 hdr->b_l2hdr.b_daddr,
2409 hdr->b_l2hdr.b_asize, 0);
2410 arc_hdr_l2hdr_destroy(hdr);
2414 mutex_exit(&dev->l2ad_mtx);
2417 if (!BUF_EMPTY(hdr))
2418 buf_discard_identity(hdr);
2420 if (hdr->b_freeze_cksum != NULL) {
2421 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
2422 hdr->b_freeze_cksum = NULL;
2425 if (HDR_HAS_L1HDR(hdr)) {
2426 while (hdr->b_l1hdr.b_buf) {
2427 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2429 if (buf->b_efunc != NULL) {
2430 mutex_enter(&arc_user_evicts_lock);
2431 mutex_enter(&buf->b_evict_lock);
2432 ASSERT(buf->b_hdr != NULL);
2433 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE);
2434 hdr->b_l1hdr.b_buf = buf->b_next;
2435 buf->b_hdr = &arc_eviction_hdr;
2436 buf->b_next = arc_eviction_list;
2437 arc_eviction_list = buf;
2438 mutex_exit(&buf->b_evict_lock);
2439 cv_signal(&arc_user_evicts_cv);
2440 mutex_exit(&arc_user_evicts_lock);
2442 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE);
2446 if (hdr->b_l1hdr.b_thawed != NULL) {
2447 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2448 hdr->b_l1hdr.b_thawed = NULL;
2453 ASSERT3P(hdr->b_hash_next, ==, NULL);
2454 if (HDR_HAS_L1HDR(hdr)) {
2455 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2456 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
2457 kmem_cache_free(hdr_full_cache, hdr);
2459 kmem_cache_free(hdr_l2only_cache, hdr);
2464 arc_buf_free(arc_buf_t *buf, void *tag)
2466 arc_buf_hdr_t *hdr = buf->b_hdr;
2467 int hashed = hdr->b_l1hdr.b_state != arc_anon;
2469 ASSERT(buf->b_efunc == NULL);
2470 ASSERT(buf->b_data != NULL);
2473 kmutex_t *hash_lock = HDR_LOCK(hdr);
2475 mutex_enter(hash_lock);
2477 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2479 (void) remove_reference(hdr, hash_lock, tag);
2480 if (hdr->b_l1hdr.b_datacnt > 1) {
2481 arc_buf_destroy(buf, TRUE);
2483 ASSERT(buf == hdr->b_l1hdr.b_buf);
2484 ASSERT(buf->b_efunc == NULL);
2485 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2487 mutex_exit(hash_lock);
2488 } else if (HDR_IO_IN_PROGRESS(hdr)) {
2491 * We are in the middle of an async write. Don't destroy
2492 * this buffer unless the write completes before we finish
2493 * decrementing the reference count.
2495 mutex_enter(&arc_user_evicts_lock);
2496 (void) remove_reference(hdr, NULL, tag);
2497 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2498 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
2499 mutex_exit(&arc_user_evicts_lock);
2501 arc_hdr_destroy(hdr);
2503 if (remove_reference(hdr, NULL, tag) > 0)
2504 arc_buf_destroy(buf, TRUE);
2506 arc_hdr_destroy(hdr);
2511 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
2513 arc_buf_hdr_t *hdr = buf->b_hdr;
2514 kmutex_t *hash_lock = HDR_LOCK(hdr);
2515 boolean_t no_callback = (buf->b_efunc == NULL);
2517 if (hdr->b_l1hdr.b_state == arc_anon) {
2518 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
2519 arc_buf_free(buf, tag);
2520 return (no_callback);
2523 mutex_enter(hash_lock);
2525 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
2526 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2527 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2528 ASSERT(buf->b_data != NULL);
2530 (void) remove_reference(hdr, hash_lock, tag);
2531 if (hdr->b_l1hdr.b_datacnt > 1) {
2533 arc_buf_destroy(buf, TRUE);
2534 } else if (no_callback) {
2535 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL);
2536 ASSERT(buf->b_efunc == NULL);
2537 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2539 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 ||
2540 refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2541 mutex_exit(hash_lock);
2542 return (no_callback);
2546 arc_buf_size(arc_buf_t *buf)
2548 return (buf->b_hdr->b_size);
2552 * Called from the DMU to determine if the current buffer should be
2553 * evicted. In order to ensure proper locking, the eviction must be initiated
2554 * from the DMU. Return true if the buffer is associated with user data and
2555 * duplicate buffers still exist.
2558 arc_buf_eviction_needed(arc_buf_t *buf)
2561 boolean_t evict_needed = B_FALSE;
2563 if (zfs_disable_dup_eviction)
2566 mutex_enter(&buf->b_evict_lock);
2570 * We are in arc_do_user_evicts(); let that function
2571 * perform the eviction.
2573 ASSERT(buf->b_data == NULL);
2574 mutex_exit(&buf->b_evict_lock);
2576 } else if (buf->b_data == NULL) {
2578 * We have already been added to the arc eviction list;
2579 * recommend eviction.
2581 ASSERT3P(hdr, ==, &arc_eviction_hdr);
2582 mutex_exit(&buf->b_evict_lock);
2586 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr))
2587 evict_needed = B_TRUE;
2589 mutex_exit(&buf->b_evict_lock);
2590 return (evict_needed);
2594 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2595 * state of the header is dependent on it's state prior to entering this
2596 * function. The following transitions are possible:
2598 * - arc_mru -> arc_mru_ghost
2599 * - arc_mfu -> arc_mfu_ghost
2600 * - arc_mru_ghost -> arc_l2c_only
2601 * - arc_mru_ghost -> deleted
2602 * - arc_mfu_ghost -> arc_l2c_only
2603 * - arc_mfu_ghost -> deleted
2606 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
2608 arc_state_t *evicted_state, *state;
2609 int64_t bytes_evicted = 0;
2611 ASSERT(MUTEX_HELD(hash_lock));
2612 ASSERT(HDR_HAS_L1HDR(hdr));
2614 state = hdr->b_l1hdr.b_state;
2615 if (GHOST_STATE(state)) {
2616 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2617 ASSERT(hdr->b_l1hdr.b_buf == NULL);
2620 * l2arc_write_buffers() relies on a header's L1 portion
2621 * (i.e. it's b_tmp_cdata field) during it's write phase.
2622 * Thus, we cannot push a header onto the arc_l2c_only
2623 * state (removing it's L1 piece) until the header is
2624 * done being written to the l2arc.
2626 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
2627 ARCSTAT_BUMP(arcstat_evict_l2_skip);
2628 return (bytes_evicted);
2631 ARCSTAT_BUMP(arcstat_deleted);
2632 bytes_evicted += hdr->b_size;
2634 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
2636 if (HDR_HAS_L2HDR(hdr)) {
2638 * This buffer is cached on the 2nd Level ARC;
2639 * don't destroy the header.
2641 arc_change_state(arc_l2c_only, hdr, hash_lock);
2643 * dropping from L1+L2 cached to L2-only,
2644 * realloc to remove the L1 header.
2646 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
2649 arc_change_state(arc_anon, hdr, hash_lock);
2650 arc_hdr_destroy(hdr);
2652 return (bytes_evicted);
2655 ASSERT(state == arc_mru || state == arc_mfu);
2656 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2658 /* prefetch buffers have a minimum lifespan */
2659 if (HDR_IO_IN_PROGRESS(hdr) ||
2660 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
2661 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
2662 arc_min_prefetch_lifespan)) {
2663 ARCSTAT_BUMP(arcstat_evict_skip);
2664 return (bytes_evicted);
2667 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
2668 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0);
2669 while (hdr->b_l1hdr.b_buf) {
2670 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2671 if (!mutex_tryenter(&buf->b_evict_lock)) {
2672 ARCSTAT_BUMP(arcstat_mutex_miss);
2675 if (buf->b_data != NULL)
2676 bytes_evicted += hdr->b_size;
2677 if (buf->b_efunc != NULL) {
2678 mutex_enter(&arc_user_evicts_lock);
2679 arc_buf_destroy(buf, FALSE);
2680 hdr->b_l1hdr.b_buf = buf->b_next;
2681 buf->b_hdr = &arc_eviction_hdr;
2682 buf->b_next = arc_eviction_list;
2683 arc_eviction_list = buf;
2684 cv_signal(&arc_user_evicts_cv);
2685 mutex_exit(&arc_user_evicts_lock);
2686 mutex_exit(&buf->b_evict_lock);
2688 mutex_exit(&buf->b_evict_lock);
2689 arc_buf_destroy(buf, TRUE);
2693 if (HDR_HAS_L2HDR(hdr)) {
2694 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size);
2696 if (l2arc_write_eligible(hdr->b_spa, hdr))
2697 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size);
2699 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size);
2702 if (hdr->b_l1hdr.b_datacnt == 0) {
2703 arc_change_state(evicted_state, hdr, hash_lock);
2704 ASSERT(HDR_IN_HASH_TABLE(hdr));
2705 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
2706 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
2707 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
2710 return (bytes_evicted);
2714 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
2715 uint64_t spa, int64_t bytes)
2717 multilist_sublist_t *mls;
2718 uint64_t bytes_evicted = 0;
2720 kmutex_t *hash_lock;
2721 int evict_count = 0;
2723 ASSERT3P(marker, !=, NULL);
2724 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2726 mls = multilist_sublist_lock(ml, idx);
2728 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
2729 hdr = multilist_sublist_prev(mls, marker)) {
2730 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
2731 (evict_count >= zfs_arc_evict_batch_limit))
2735 * To keep our iteration location, move the marker
2736 * forward. Since we're not holding hdr's hash lock, we
2737 * must be very careful and not remove 'hdr' from the
2738 * sublist. Otherwise, other consumers might mistake the
2739 * 'hdr' as not being on a sublist when they call the
2740 * multilist_link_active() function (they all rely on
2741 * the hash lock protecting concurrent insertions and
2742 * removals). multilist_sublist_move_forward() was
2743 * specifically implemented to ensure this is the case
2744 * (only 'marker' will be removed and re-inserted).
2746 multilist_sublist_move_forward(mls, marker);
2749 * The only case where the b_spa field should ever be
2750 * zero, is the marker headers inserted by
2751 * arc_evict_state(). It's possible for multiple threads
2752 * to be calling arc_evict_state() concurrently (e.g.
2753 * dsl_pool_close() and zio_inject_fault()), so we must
2754 * skip any markers we see from these other threads.
2756 if (hdr->b_spa == 0)
2759 /* we're only interested in evicting buffers of a certain spa */
2760 if (spa != 0 && hdr->b_spa != spa) {
2761 ARCSTAT_BUMP(arcstat_evict_skip);
2765 hash_lock = HDR_LOCK(hdr);
2768 * We aren't calling this function from any code path
2769 * that would already be holding a hash lock, so we're
2770 * asserting on this assumption to be defensive in case
2771 * this ever changes. Without this check, it would be
2772 * possible to incorrectly increment arcstat_mutex_miss
2773 * below (e.g. if the code changed such that we called
2774 * this function with a hash lock held).
2776 ASSERT(!MUTEX_HELD(hash_lock));
2778 if (mutex_tryenter(hash_lock)) {
2779 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
2780 mutex_exit(hash_lock);
2782 bytes_evicted += evicted;
2785 * If evicted is zero, arc_evict_hdr() must have
2786 * decided to skip this header, don't increment
2787 * evict_count in this case.
2793 * If arc_size isn't overflowing, signal any
2794 * threads that might happen to be waiting.
2796 * For each header evicted, we wake up a single
2797 * thread. If we used cv_broadcast, we could
2798 * wake up "too many" threads causing arc_size
2799 * to significantly overflow arc_c; since
2800 * arc_get_data_buf() doesn't check for overflow
2801 * when it's woken up (it doesn't because it's
2802 * possible for the ARC to be overflowing while
2803 * full of un-evictable buffers, and the
2804 * function should proceed in this case).
2806 * If threads are left sleeping, due to not
2807 * using cv_broadcast, they will be woken up
2808 * just before arc_reclaim_thread() sleeps.
2810 mutex_enter(&arc_reclaim_lock);
2811 if (!arc_is_overflowing())
2812 cv_signal(&arc_reclaim_waiters_cv);
2813 mutex_exit(&arc_reclaim_lock);
2815 ARCSTAT_BUMP(arcstat_mutex_miss);
2819 multilist_sublist_unlock(mls);
2821 return (bytes_evicted);
2825 * Evict buffers from the given arc state, until we've removed the
2826 * specified number of bytes. Move the removed buffers to the
2827 * appropriate evict state.
2829 * This function makes a "best effort". It skips over any buffers
2830 * it can't get a hash_lock on, and so, may not catch all candidates.
2831 * It may also return without evicting as much space as requested.
2833 * If bytes is specified using the special value ARC_EVICT_ALL, this
2834 * will evict all available (i.e. unlocked and evictable) buffers from
2835 * the given arc state; which is used by arc_flush().
2838 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
2839 arc_buf_contents_t type)
2841 uint64_t total_evicted = 0;
2842 multilist_t *ml = &state->arcs_list[type];
2844 arc_buf_hdr_t **markers;
2846 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2848 num_sublists = multilist_get_num_sublists(ml);
2851 * If we've tried to evict from each sublist, made some
2852 * progress, but still have not hit the target number of bytes
2853 * to evict, we want to keep trying. The markers allow us to
2854 * pick up where we left off for each individual sublist, rather
2855 * than starting from the tail each time.
2857 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
2858 for (int i = 0; i < num_sublists; i++) {
2859 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
2862 * A b_spa of 0 is used to indicate that this header is
2863 * a marker. This fact is used in arc_adjust_type() and
2864 * arc_evict_state_impl().
2866 markers[i]->b_spa = 0;
2868 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2869 multilist_sublist_insert_tail(mls, markers[i]);
2870 multilist_sublist_unlock(mls);
2874 * While we haven't hit our target number of bytes to evict, or
2875 * we're evicting all available buffers.
2877 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
2879 * Start eviction using a randomly selected sublist,
2880 * this is to try and evenly balance eviction across all
2881 * sublists. Always starting at the same sublist
2882 * (e.g. index 0) would cause evictions to favor certain
2883 * sublists over others.
2885 int sublist_idx = multilist_get_random_index(ml);
2886 uint64_t scan_evicted = 0;
2888 for (int i = 0; i < num_sublists; i++) {
2889 uint64_t bytes_remaining;
2890 uint64_t bytes_evicted;
2892 if (bytes == ARC_EVICT_ALL)
2893 bytes_remaining = ARC_EVICT_ALL;
2894 else if (total_evicted < bytes)
2895 bytes_remaining = bytes - total_evicted;
2899 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
2900 markers[sublist_idx], spa, bytes_remaining);
2902 scan_evicted += bytes_evicted;
2903 total_evicted += bytes_evicted;
2905 /* we've reached the end, wrap to the beginning */
2906 if (++sublist_idx >= num_sublists)
2911 * If we didn't evict anything during this scan, we have
2912 * no reason to believe we'll evict more during another
2913 * scan, so break the loop.
2915 if (scan_evicted == 0) {
2916 /* This isn't possible, let's make that obvious */
2917 ASSERT3S(bytes, !=, 0);
2920 * When bytes is ARC_EVICT_ALL, the only way to
2921 * break the loop is when scan_evicted is zero.
2922 * In that case, we actually have evicted enough,
2923 * so we don't want to increment the kstat.
2925 if (bytes != ARC_EVICT_ALL) {
2926 ASSERT3S(total_evicted, <, bytes);
2927 ARCSTAT_BUMP(arcstat_evict_not_enough);
2934 for (int i = 0; i < num_sublists; i++) {
2935 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2936 multilist_sublist_remove(mls, markers[i]);
2937 multilist_sublist_unlock(mls);
2939 kmem_cache_free(hdr_full_cache, markers[i]);
2941 kmem_free(markers, sizeof (*markers) * num_sublists);
2943 return (total_evicted);
2947 * Flush all "evictable" data of the given type from the arc state
2948 * specified. This will not evict any "active" buffers (i.e. referenced).
2950 * When 'retry' is set to FALSE, the function will make a single pass
2951 * over the state and evict any buffers that it can. Since it doesn't
2952 * continually retry the eviction, it might end up leaving some buffers
2953 * in the ARC due to lock misses.
2955 * When 'retry' is set to TRUE, the function will continually retry the
2956 * eviction until *all* evictable buffers have been removed from the
2957 * state. As a result, if concurrent insertions into the state are
2958 * allowed (e.g. if the ARC isn't shutting down), this function might
2959 * wind up in an infinite loop, continually trying to evict buffers.
2962 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
2965 uint64_t evicted = 0;
2967 while (state->arcs_lsize[type] != 0) {
2968 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
2978 * Evict the specified number of bytes from the state specified,
2979 * restricting eviction to the spa and type given. This function
2980 * prevents us from trying to evict more from a state's list than
2981 * is "evictable", and to skip evicting altogether when passed a
2982 * negative value for "bytes". In contrast, arc_evict_state() will
2983 * evict everything it can, when passed a negative value for "bytes".
2986 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
2987 arc_buf_contents_t type)
2991 if (bytes > 0 && state->arcs_lsize[type] > 0) {
2992 delta = MIN(state->arcs_lsize[type], bytes);
2993 return (arc_evict_state(state, spa, delta, type));
3000 * Evict metadata buffers from the cache, such that arc_meta_used is
3001 * capped by the arc_meta_limit tunable.
3004 arc_adjust_meta(void)
3006 uint64_t total_evicted = 0;
3010 * If we're over the meta limit, we want to evict enough
3011 * metadata to get back under the meta limit. We don't want to
3012 * evict so much that we drop the MRU below arc_p, though. If
3013 * we're over the meta limit more than we're over arc_p, we
3014 * evict some from the MRU here, and some from the MFU below.
3016 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3017 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3018 refcount_count(&arc_mru->arcs_size) - arc_p));
3020 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3023 * Similar to the above, we want to evict enough bytes to get us
3024 * below the meta limit, but not so much as to drop us below the
3025 * space alloted to the MFU (which is defined as arc_c - arc_p).
3027 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3028 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
3030 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3032 return (total_evicted);
3036 * Return the type of the oldest buffer in the given arc state
3038 * This function will select a random sublist of type ARC_BUFC_DATA and
3039 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3040 * is compared, and the type which contains the "older" buffer will be
3043 static arc_buf_contents_t
3044 arc_adjust_type(arc_state_t *state)
3046 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
3047 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
3048 int data_idx = multilist_get_random_index(data_ml);
3049 int meta_idx = multilist_get_random_index(meta_ml);
3050 multilist_sublist_t *data_mls;
3051 multilist_sublist_t *meta_mls;
3052 arc_buf_contents_t type;
3053 arc_buf_hdr_t *data_hdr;
3054 arc_buf_hdr_t *meta_hdr;
3057 * We keep the sublist lock until we're finished, to prevent
3058 * the headers from being destroyed via arc_evict_state().
3060 data_mls = multilist_sublist_lock(data_ml, data_idx);
3061 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3064 * These two loops are to ensure we skip any markers that
3065 * might be at the tail of the lists due to arc_evict_state().
3068 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3069 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3070 if (data_hdr->b_spa != 0)
3074 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3075 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3076 if (meta_hdr->b_spa != 0)
3080 if (data_hdr == NULL && meta_hdr == NULL) {
3081 type = ARC_BUFC_DATA;
3082 } else if (data_hdr == NULL) {
3083 ASSERT3P(meta_hdr, !=, NULL);
3084 type = ARC_BUFC_METADATA;
3085 } else if (meta_hdr == NULL) {
3086 ASSERT3P(data_hdr, !=, NULL);
3087 type = ARC_BUFC_DATA;
3089 ASSERT3P(data_hdr, !=, NULL);
3090 ASSERT3P(meta_hdr, !=, NULL);
3092 /* The headers can't be on the sublist without an L1 header */
3093 ASSERT(HDR_HAS_L1HDR(data_hdr));
3094 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3096 if (data_hdr->b_l1hdr.b_arc_access <
3097 meta_hdr->b_l1hdr.b_arc_access) {
3098 type = ARC_BUFC_DATA;
3100 type = ARC_BUFC_METADATA;
3104 multilist_sublist_unlock(meta_mls);
3105 multilist_sublist_unlock(data_mls);
3111 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3116 uint64_t total_evicted = 0;
3121 * If we're over arc_meta_limit, we want to correct that before
3122 * potentially evicting data buffers below.
3124 total_evicted += arc_adjust_meta();
3129 * If we're over the target cache size, we want to evict enough
3130 * from the list to get back to our target size. We don't want
3131 * to evict too much from the MRU, such that it drops below
3132 * arc_p. So, if we're over our target cache size more than
3133 * the MRU is over arc_p, we'll evict enough to get back to
3134 * arc_p here, and then evict more from the MFU below.
3136 target = MIN((int64_t)(arc_size - arc_c),
3137 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3138 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
3141 * If we're below arc_meta_min, always prefer to evict data.
3142 * Otherwise, try to satisfy the requested number of bytes to
3143 * evict from the type which contains older buffers; in an
3144 * effort to keep newer buffers in the cache regardless of their
3145 * type. If we cannot satisfy the number of bytes from this
3146 * type, spill over into the next type.
3148 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
3149 arc_meta_used > arc_meta_min) {
3150 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3151 total_evicted += bytes;
3154 * If we couldn't evict our target number of bytes from
3155 * metadata, we try to get the rest from data.
3160 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3162 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3163 total_evicted += bytes;
3166 * If we couldn't evict our target number of bytes from
3167 * data, we try to get the rest from metadata.
3172 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3178 * Now that we've tried to evict enough from the MRU to get its
3179 * size back to arc_p, if we're still above the target cache
3180 * size, we evict the rest from the MFU.
3182 target = arc_size - arc_c;
3184 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
3185 arc_meta_used > arc_meta_min) {
3186 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3187 total_evicted += bytes;
3190 * If we couldn't evict our target number of bytes from
3191 * metadata, we try to get the rest from data.
3196 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3198 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3199 total_evicted += bytes;
3202 * If we couldn't evict our target number of bytes from
3203 * data, we try to get the rest from data.
3208 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3212 * Adjust ghost lists
3214 * In addition to the above, the ARC also defines target values
3215 * for the ghost lists. The sum of the mru list and mru ghost
3216 * list should never exceed the target size of the cache, and
3217 * the sum of the mru list, mfu list, mru ghost list, and mfu
3218 * ghost list should never exceed twice the target size of the
3219 * cache. The following logic enforces these limits on the ghost
3220 * caches, and evicts from them as needed.
3222 target = refcount_count(&arc_mru->arcs_size) +
3223 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3225 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3226 total_evicted += bytes;
3231 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3234 * We assume the sum of the mru list and mfu list is less than
3235 * or equal to arc_c (we enforced this above), which means we
3236 * can use the simpler of the two equations below:
3238 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3239 * mru ghost + mfu ghost <= arc_c
3241 target = refcount_count(&arc_mru_ghost->arcs_size) +
3242 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3244 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3245 total_evicted += bytes;
3250 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3252 return (total_evicted);
3256 arc_do_user_evicts(void)
3258 mutex_enter(&arc_user_evicts_lock);
3259 while (arc_eviction_list != NULL) {
3260 arc_buf_t *buf = arc_eviction_list;
3261 arc_eviction_list = buf->b_next;
3262 mutex_enter(&buf->b_evict_lock);
3264 mutex_exit(&buf->b_evict_lock);
3265 mutex_exit(&arc_user_evicts_lock);
3267 if (buf->b_efunc != NULL)
3268 VERIFY0(buf->b_efunc(buf->b_private));
3270 buf->b_efunc = NULL;
3271 buf->b_private = NULL;
3272 kmem_cache_free(buf_cache, buf);
3273 mutex_enter(&arc_user_evicts_lock);
3275 mutex_exit(&arc_user_evicts_lock);
3279 arc_flush(spa_t *spa, boolean_t retry)
3284 * If retry is TRUE, a spa must not be specified since we have
3285 * no good way to determine if all of a spa's buffers have been
3286 * evicted from an arc state.
3288 ASSERT(!retry || spa == 0);
3291 guid = spa_load_guid(spa);
3293 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3294 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3296 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3297 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3299 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3300 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3302 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3303 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3305 arc_do_user_evicts();
3306 ASSERT(spa || arc_eviction_list == NULL);
3310 arc_shrink(int64_t to_free)
3312 if (arc_c > arc_c_min) {
3313 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t,
3314 arc_c_min, uint64_t, arc_p, uint64_t, to_free);
3315 if (arc_c > arc_c_min + to_free)
3316 atomic_add_64(&arc_c, -to_free);
3320 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3321 if (arc_c > arc_size)
3322 arc_c = MAX(arc_size, arc_c_min);
3324 arc_p = (arc_c >> 1);
3326 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t,
3329 ASSERT(arc_c >= arc_c_min);
3330 ASSERT((int64_t)arc_p >= 0);
3333 if (arc_size > arc_c) {
3334 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size,
3336 (void) arc_adjust();
3340 static long needfree = 0;
3342 typedef enum free_memory_reason_t {
3347 FMR_PAGES_PP_MAXIMUM,
3351 } free_memory_reason_t;
3353 int64_t last_free_memory;
3354 free_memory_reason_t last_free_reason;
3357 * Additional reserve of pages for pp_reserve.
3359 int64_t arc_pages_pp_reserve = 64;
3362 * Additional reserve of pages for swapfs.
3364 int64_t arc_swapfs_reserve = 64;
3367 * Return the amount of memory that can be consumed before reclaim will be
3368 * needed. Positive if there is sufficient free memory, negative indicates
3369 * the amount of memory that needs to be freed up.
3372 arc_available_memory(void)
3374 int64_t lowest = INT64_MAX;
3376 free_memory_reason_t r = FMR_UNKNOWN;
3380 n = PAGESIZE * (-needfree);
3388 * Cooperate with pagedaemon when it's time for it to scan
3389 * and reclaim some pages.
3391 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target);
3399 * check that we're out of range of the pageout scanner. It starts to
3400 * schedule paging if freemem is less than lotsfree and needfree.
3401 * lotsfree is the high-water mark for pageout, and needfree is the
3402 * number of needed free pages. We add extra pages here to make sure
3403 * the scanner doesn't start up while we're freeing memory.
3405 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3412 * check to make sure that swapfs has enough space so that anon
3413 * reservations can still succeed. anon_resvmem() checks that the
3414 * availrmem is greater than swapfs_minfree, and the number of reserved
3415 * swap pages. We also add a bit of extra here just to prevent
3416 * circumstances from getting really dire.
3418 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3419 desfree - arc_swapfs_reserve);
3422 r = FMR_SWAPFS_MINFREE;
3427 * Check that we have enough availrmem that memory locking (e.g., via
3428 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3429 * stores the number of pages that cannot be locked; when availrmem
3430 * drops below pages_pp_maximum, page locking mechanisms such as
3431 * page_pp_lock() will fail.)
3433 n = PAGESIZE * (availrmem - pages_pp_maximum -
3434 arc_pages_pp_reserve);
3437 r = FMR_PAGES_PP_MAXIMUM;
3441 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
3443 * If we're on an i386 platform, it's possible that we'll exhaust the
3444 * kernel heap space before we ever run out of available physical
3445 * memory. Most checks of the size of the heap_area compare against
3446 * tune.t_minarmem, which is the minimum available real memory that we
3447 * can have in the system. However, this is generally fixed at 25 pages
3448 * which is so low that it's useless. In this comparison, we seek to
3449 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3450 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3453 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
3454 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3459 #define zio_arena NULL
3461 #define zio_arena heap_arena
3465 * If zio data pages are being allocated out of a separate heap segment,
3466 * then enforce that the size of available vmem for this arena remains
3467 * above about 1/16th free.
3469 * Note: The 1/16th arena free requirement was put in place
3470 * to aggressively evict memory from the arc in order to avoid
3471 * memory fragmentation issues.
3473 if (zio_arena != NULL) {
3474 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
3475 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3483 * Above limits know nothing about real level of KVA fragmentation.
3484 * Start aggressive reclamation if too little sequential KVA left.
3487 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ?
3488 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) :
3497 /* Every 100 calls, free a small amount */
3498 if (spa_get_random(100) == 0)
3500 #endif /* _KERNEL */
3502 last_free_memory = lowest;
3503 last_free_reason = r;
3504 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r);
3510 * Determine if the system is under memory pressure and is asking
3511 * to reclaim memory. A return value of TRUE indicates that the system
3512 * is under memory pressure and that the arc should adjust accordingly.
3515 arc_reclaim_needed(void)
3517 return (arc_available_memory() < 0);
3520 extern kmem_cache_t *zio_buf_cache[];
3521 extern kmem_cache_t *zio_data_buf_cache[];
3522 extern kmem_cache_t *range_seg_cache;
3524 static __noinline void
3525 arc_kmem_reap_now(void)
3528 kmem_cache_t *prev_cache = NULL;
3529 kmem_cache_t *prev_data_cache = NULL;
3531 DTRACE_PROBE(arc__kmem_reap_start);
3533 if (arc_meta_used >= arc_meta_limit) {
3535 * We are exceeding our meta-data cache limit.
3536 * Purge some DNLC entries to release holds on meta-data.
3538 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
3542 * Reclaim unused memory from all kmem caches.
3548 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
3549 if (zio_buf_cache[i] != prev_cache) {
3550 prev_cache = zio_buf_cache[i];
3551 kmem_cache_reap_now(zio_buf_cache[i]);
3553 if (zio_data_buf_cache[i] != prev_data_cache) {
3554 prev_data_cache = zio_data_buf_cache[i];
3555 kmem_cache_reap_now(zio_data_buf_cache[i]);
3558 kmem_cache_reap_now(buf_cache);
3559 kmem_cache_reap_now(hdr_full_cache);
3560 kmem_cache_reap_now(hdr_l2only_cache);
3561 kmem_cache_reap_now(range_seg_cache);
3564 if (zio_arena != NULL) {
3566 * Ask the vmem arena to reclaim unused memory from its
3569 vmem_qcache_reap(zio_arena);
3572 DTRACE_PROBE(arc__kmem_reap_end);
3576 * Threads can block in arc_get_data_buf() waiting for this thread to evict
3577 * enough data and signal them to proceed. When this happens, the threads in
3578 * arc_get_data_buf() are sleeping while holding the hash lock for their
3579 * particular arc header. Thus, we must be careful to never sleep on a
3580 * hash lock in this thread. This is to prevent the following deadlock:
3582 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3583 * waiting for the reclaim thread to signal it.
3585 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3586 * fails, and goes to sleep forever.
3588 * This possible deadlock is avoided by always acquiring a hash lock
3589 * using mutex_tryenter() from arc_reclaim_thread().
3592 arc_reclaim_thread(void *dummy __unused)
3594 clock_t growtime = 0;
3597 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
3599 mutex_enter(&arc_reclaim_lock);
3600 while (!arc_reclaim_thread_exit) {
3601 int64_t free_memory = arc_available_memory();
3602 uint64_t evicted = 0;
3604 mutex_exit(&arc_reclaim_lock);
3606 if (free_memory < 0) {
3608 arc_no_grow = B_TRUE;
3612 * Wait at least zfs_grow_retry (default 60) seconds
3613 * before considering growing.
3615 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
3617 arc_kmem_reap_now();
3620 * If we are still low on memory, shrink the ARC
3621 * so that we have arc_shrink_min free space.
3623 free_memory = arc_available_memory();
3626 (arc_c >> arc_shrink_shift) - free_memory;
3629 to_free = MAX(to_free, ptob(needfree));
3631 arc_shrink(to_free);
3633 } else if (free_memory < arc_c >> arc_no_grow_shift) {
3634 arc_no_grow = B_TRUE;
3635 } else if (ddi_get_lbolt() >= growtime) {
3636 arc_no_grow = B_FALSE;
3639 evicted = arc_adjust();
3641 mutex_enter(&arc_reclaim_lock);
3644 * If evicted is zero, we couldn't evict anything via
3645 * arc_adjust(). This could be due to hash lock
3646 * collisions, but more likely due to the majority of
3647 * arc buffers being unevictable. Therefore, even if
3648 * arc_size is above arc_c, another pass is unlikely to
3649 * be helpful and could potentially cause us to enter an
3652 if (arc_size <= arc_c || evicted == 0) {
3657 * We're either no longer overflowing, or we
3658 * can't evict anything more, so we should wake
3659 * up any threads before we go to sleep.
3661 cv_broadcast(&arc_reclaim_waiters_cv);
3664 * Block until signaled, or after one second (we
3665 * might need to perform arc_kmem_reap_now()
3666 * even if we aren't being signalled)
3668 CALLB_CPR_SAFE_BEGIN(&cpr);
3669 (void) cv_timedwait(&arc_reclaim_thread_cv,
3670 &arc_reclaim_lock, hz);
3671 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
3675 arc_reclaim_thread_exit = FALSE;
3676 cv_broadcast(&arc_reclaim_thread_cv);
3677 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
3682 arc_user_evicts_thread(void *dummy __unused)
3686 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG);
3688 mutex_enter(&arc_user_evicts_lock);
3689 while (!arc_user_evicts_thread_exit) {
3690 mutex_exit(&arc_user_evicts_lock);
3692 arc_do_user_evicts();
3695 * This is necessary in order for the mdb ::arc dcmd to
3696 * show up to date information. Since the ::arc command
3697 * does not call the kstat's update function, without
3698 * this call, the command may show stale stats for the
3699 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3700 * with this change, the data might be up to 1 second
3701 * out of date; but that should suffice. The arc_state_t
3702 * structures can be queried directly if more accurate
3703 * information is needed.
3705 if (arc_ksp != NULL)
3706 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
3708 mutex_enter(&arc_user_evicts_lock);
3711 * Block until signaled, or after one second (we need to
3712 * call the arc's kstat update function regularly).
3714 CALLB_CPR_SAFE_BEGIN(&cpr);
3715 (void) cv_timedwait(&arc_user_evicts_cv,
3716 &arc_user_evicts_lock, hz);
3717 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock);
3720 arc_user_evicts_thread_exit = FALSE;
3721 cv_broadcast(&arc_user_evicts_cv);
3722 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */
3727 * Adapt arc info given the number of bytes we are trying to add and
3728 * the state that we are comming from. This function is only called
3729 * when we are adding new content to the cache.
3732 arc_adapt(int bytes, arc_state_t *state)
3735 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
3736 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
3737 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
3739 if (state == arc_l2c_only)
3744 * Adapt the target size of the MRU list:
3745 * - if we just hit in the MRU ghost list, then increase
3746 * the target size of the MRU list.
3747 * - if we just hit in the MFU ghost list, then increase
3748 * the target size of the MFU list by decreasing the
3749 * target size of the MRU list.
3751 if (state == arc_mru_ghost) {
3752 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
3753 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
3755 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
3756 } else if (state == arc_mfu_ghost) {
3759 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
3760 mult = MIN(mult, 10);
3762 delta = MIN(bytes * mult, arc_p);
3763 arc_p = MAX(arc_p_min, arc_p - delta);
3765 ASSERT((int64_t)arc_p >= 0);
3767 if (arc_reclaim_needed()) {
3768 cv_signal(&arc_reclaim_thread_cv);
3775 if (arc_c >= arc_c_max)
3779 * If we're within (2 * maxblocksize) bytes of the target
3780 * cache size, increment the target cache size
3782 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
3783 DTRACE_PROBE1(arc__inc_adapt, int, bytes);
3784 atomic_add_64(&arc_c, (int64_t)bytes);
3785 if (arc_c > arc_c_max)
3787 else if (state == arc_anon)
3788 atomic_add_64(&arc_p, (int64_t)bytes);
3792 ASSERT((int64_t)arc_p >= 0);
3796 * Check if arc_size has grown past our upper threshold, determined by
3797 * zfs_arc_overflow_shift.
3800 arc_is_overflowing(void)
3802 /* Always allow at least one block of overflow */
3803 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
3804 arc_c >> zfs_arc_overflow_shift);
3806 return (arc_size >= arc_c + overflow);
3810 * The buffer, supplied as the first argument, needs a data block. If we
3811 * are hitting the hard limit for the cache size, we must sleep, waiting
3812 * for the eviction thread to catch up. If we're past the target size
3813 * but below the hard limit, we'll only signal the reclaim thread and
3817 arc_get_data_buf(arc_buf_t *buf)
3819 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
3820 uint64_t size = buf->b_hdr->b_size;
3821 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
3823 arc_adapt(size, state);
3826 * If arc_size is currently overflowing, and has grown past our
3827 * upper limit, we must be adding data faster than the evict
3828 * thread can evict. Thus, to ensure we don't compound the
3829 * problem by adding more data and forcing arc_size to grow even
3830 * further past it's target size, we halt and wait for the
3831 * eviction thread to catch up.
3833 * It's also possible that the reclaim thread is unable to evict
3834 * enough buffers to get arc_size below the overflow limit (e.g.
3835 * due to buffers being un-evictable, or hash lock collisions).
3836 * In this case, we want to proceed regardless if we're
3837 * overflowing; thus we don't use a while loop here.
3839 if (arc_is_overflowing()) {
3840 mutex_enter(&arc_reclaim_lock);
3843 * Now that we've acquired the lock, we may no longer be
3844 * over the overflow limit, lets check.
3846 * We're ignoring the case of spurious wake ups. If that
3847 * were to happen, it'd let this thread consume an ARC
3848 * buffer before it should have (i.e. before we're under
3849 * the overflow limit and were signalled by the reclaim
3850 * thread). As long as that is a rare occurrence, it
3851 * shouldn't cause any harm.
3853 if (arc_is_overflowing()) {
3854 cv_signal(&arc_reclaim_thread_cv);
3855 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
3858 mutex_exit(&arc_reclaim_lock);
3861 if (type == ARC_BUFC_METADATA) {
3862 buf->b_data = zio_buf_alloc(size);
3863 arc_space_consume(size, ARC_SPACE_META);
3865 ASSERT(type == ARC_BUFC_DATA);
3866 buf->b_data = zio_data_buf_alloc(size);
3867 arc_space_consume(size, ARC_SPACE_DATA);
3871 * Update the state size. Note that ghost states have a
3872 * "ghost size" and so don't need to be updated.
3874 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) {
3875 arc_buf_hdr_t *hdr = buf->b_hdr;
3876 arc_state_t *state = hdr->b_l1hdr.b_state;
3878 (void) refcount_add_many(&state->arcs_size, size, buf);
3881 * If this is reached via arc_read, the link is
3882 * protected by the hash lock. If reached via
3883 * arc_buf_alloc, the header should not be accessed by
3884 * any other thread. And, if reached via arc_read_done,
3885 * the hash lock will protect it if it's found in the
3886 * hash table; otherwise no other thread should be
3887 * trying to [add|remove]_reference it.
3889 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3890 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3891 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type],
3895 * If we are growing the cache, and we are adding anonymous
3896 * data, and we have outgrown arc_p, update arc_p
3898 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
3899 (refcount_count(&arc_anon->arcs_size) +
3900 refcount_count(&arc_mru->arcs_size) > arc_p))
3901 arc_p = MIN(arc_c, arc_p + size);
3903 ARCSTAT_BUMP(arcstat_allocated);
3907 * This routine is called whenever a buffer is accessed.
3908 * NOTE: the hash lock is dropped in this function.
3911 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3915 ASSERT(MUTEX_HELD(hash_lock));
3916 ASSERT(HDR_HAS_L1HDR(hdr));
3918 if (hdr->b_l1hdr.b_state == arc_anon) {
3920 * This buffer is not in the cache, and does not
3921 * appear in our "ghost" list. Add the new buffer
3925 ASSERT0(hdr->b_l1hdr.b_arc_access);
3926 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3927 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3928 arc_change_state(arc_mru, hdr, hash_lock);
3930 } else if (hdr->b_l1hdr.b_state == arc_mru) {
3931 now = ddi_get_lbolt();
3934 * If this buffer is here because of a prefetch, then either:
3935 * - clear the flag if this is a "referencing" read
3936 * (any subsequent access will bump this into the MFU state).
3938 * - move the buffer to the head of the list if this is
3939 * another prefetch (to make it less likely to be evicted).
3941 if (HDR_PREFETCH(hdr)) {
3942 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
3943 /* link protected by hash lock */
3944 ASSERT(multilist_link_active(
3945 &hdr->b_l1hdr.b_arc_node));
3947 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3948 ARCSTAT_BUMP(arcstat_mru_hits);
3950 hdr->b_l1hdr.b_arc_access = now;
3955 * This buffer has been "accessed" only once so far,
3956 * but it is still in the cache. Move it to the MFU
3959 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
3961 * More than 125ms have passed since we
3962 * instantiated this buffer. Move it to the
3963 * most frequently used state.
3965 hdr->b_l1hdr.b_arc_access = now;
3966 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3967 arc_change_state(arc_mfu, hdr, hash_lock);
3969 ARCSTAT_BUMP(arcstat_mru_hits);
3970 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
3971 arc_state_t *new_state;
3973 * This buffer has been "accessed" recently, but
3974 * was evicted from the cache. Move it to the
3978 if (HDR_PREFETCH(hdr)) {
3979 new_state = arc_mru;
3980 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
3981 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3982 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3984 new_state = arc_mfu;
3985 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3988 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3989 arc_change_state(new_state, hdr, hash_lock);
3991 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
3992 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
3994 * This buffer has been accessed more than once and is
3995 * still in the cache. Keep it in the MFU state.
3997 * NOTE: an add_reference() that occurred when we did
3998 * the arc_read() will have kicked this off the list.
3999 * If it was a prefetch, we will explicitly move it to
4000 * the head of the list now.
4002 if ((HDR_PREFETCH(hdr)) != 0) {
4003 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4004 /* link protected by hash_lock */
4005 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4007 ARCSTAT_BUMP(arcstat_mfu_hits);
4008 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4009 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4010 arc_state_t *new_state = arc_mfu;
4012 * This buffer has been accessed more than once but has
4013 * been evicted from the cache. Move it back to the
4017 if (HDR_PREFETCH(hdr)) {
4019 * This is a prefetch access...
4020 * move this block back to the MRU state.
4022 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4023 new_state = arc_mru;
4026 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4027 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4028 arc_change_state(new_state, hdr, hash_lock);
4030 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4031 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4033 * This buffer is on the 2nd Level ARC.
4036 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4037 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4038 arc_change_state(arc_mfu, hdr, hash_lock);
4040 ASSERT(!"invalid arc state");
4044 /* a generic arc_done_func_t which you can use */
4047 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4049 if (zio == NULL || zio->io_error == 0)
4050 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
4051 VERIFY(arc_buf_remove_ref(buf, arg));
4054 /* a generic arc_done_func_t */
4056 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4058 arc_buf_t **bufp = arg;
4059 if (zio && zio->io_error) {
4060 VERIFY(arc_buf_remove_ref(buf, arg));
4064 ASSERT(buf->b_data);
4069 arc_read_done(zio_t *zio)
4073 arc_buf_t *abuf; /* buffer we're assigning to callback */
4074 kmutex_t *hash_lock = NULL;
4075 arc_callback_t *callback_list, *acb;
4076 int freeable = FALSE;
4078 buf = zio->io_private;
4082 * The hdr was inserted into hash-table and removed from lists
4083 * prior to starting I/O. We should find this header, since
4084 * it's in the hash table, and it should be legit since it's
4085 * not possible to evict it during the I/O. The only possible
4086 * reason for it not to be found is if we were freed during the
4089 if (HDR_IN_HASH_TABLE(hdr)) {
4090 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
4091 ASSERT3U(hdr->b_dva.dva_word[0], ==,
4092 BP_IDENTITY(zio->io_bp)->dva_word[0]);
4093 ASSERT3U(hdr->b_dva.dva_word[1], ==,
4094 BP_IDENTITY(zio->io_bp)->dva_word[1]);
4096 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
4099 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
4100 hash_lock == NULL) ||
4102 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
4103 (found == hdr && HDR_L2_READING(hdr)));
4106 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED;
4107 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
4108 hdr->b_flags &= ~ARC_FLAG_L2CACHE;
4110 /* byteswap if necessary */
4111 callback_list = hdr->b_l1hdr.b_acb;
4112 ASSERT(callback_list != NULL);
4113 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
4114 dmu_object_byteswap_t bswap =
4115 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
4116 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
4117 byteswap_uint64_array :
4118 dmu_ot_byteswap[bswap].ob_func;
4119 func(buf->b_data, hdr->b_size);
4122 arc_cksum_compute(buf, B_FALSE);
4125 #endif /* illumos */
4127 if (hash_lock && zio->io_error == 0 &&
4128 hdr->b_l1hdr.b_state == arc_anon) {
4130 * Only call arc_access on anonymous buffers. This is because
4131 * if we've issued an I/O for an evicted buffer, we've already
4132 * called arc_access (to prevent any simultaneous readers from
4133 * getting confused).
4135 arc_access(hdr, hash_lock);
4138 /* create copies of the data buffer for the callers */
4140 for (acb = callback_list; acb; acb = acb->acb_next) {
4141 if (acb->acb_done) {
4143 ARCSTAT_BUMP(arcstat_duplicate_reads);
4144 abuf = arc_buf_clone(buf);
4146 acb->acb_buf = abuf;
4150 hdr->b_l1hdr.b_acb = NULL;
4151 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4152 ASSERT(!HDR_BUF_AVAILABLE(hdr));
4154 ASSERT(buf->b_efunc == NULL);
4155 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
4156 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4159 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
4160 callback_list != NULL);
4162 if (zio->io_error != 0) {
4163 hdr->b_flags |= ARC_FLAG_IO_ERROR;
4164 if (hdr->b_l1hdr.b_state != arc_anon)
4165 arc_change_state(arc_anon, hdr, hash_lock);
4166 if (HDR_IN_HASH_TABLE(hdr))
4167 buf_hash_remove(hdr);
4168 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4172 * Broadcast before we drop the hash_lock to avoid the possibility
4173 * that the hdr (and hence the cv) might be freed before we get to
4174 * the cv_broadcast().
4176 cv_broadcast(&hdr->b_l1hdr.b_cv);
4178 if (hash_lock != NULL) {
4179 mutex_exit(hash_lock);
4182 * This block was freed while we waited for the read to
4183 * complete. It has been removed from the hash table and
4184 * moved to the anonymous state (so that it won't show up
4187 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
4188 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4191 /* execute each callback and free its structure */
4192 while ((acb = callback_list) != NULL) {
4194 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
4196 if (acb->acb_zio_dummy != NULL) {
4197 acb->acb_zio_dummy->io_error = zio->io_error;
4198 zio_nowait(acb->acb_zio_dummy);
4201 callback_list = acb->acb_next;
4202 kmem_free(acb, sizeof (arc_callback_t));
4206 arc_hdr_destroy(hdr);
4210 * "Read" the block at the specified DVA (in bp) via the
4211 * cache. If the block is found in the cache, invoke the provided
4212 * callback immediately and return. Note that the `zio' parameter
4213 * in the callback will be NULL in this case, since no IO was
4214 * required. If the block is not in the cache pass the read request
4215 * on to the spa with a substitute callback function, so that the
4216 * requested block will be added to the cache.
4218 * If a read request arrives for a block that has a read in-progress,
4219 * either wait for the in-progress read to complete (and return the
4220 * results); or, if this is a read with a "done" func, add a record
4221 * to the read to invoke the "done" func when the read completes,
4222 * and return; or just return.
4224 * arc_read_done() will invoke all the requested "done" functions
4225 * for readers of this block.
4228 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
4229 void *private, zio_priority_t priority, int zio_flags,
4230 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
4232 arc_buf_hdr_t *hdr = NULL;
4233 arc_buf_t *buf = NULL;
4234 kmutex_t *hash_lock = NULL;
4236 uint64_t guid = spa_load_guid(spa);
4238 ASSERT(!BP_IS_EMBEDDED(bp) ||
4239 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
4242 if (!BP_IS_EMBEDDED(bp)) {
4244 * Embedded BP's have no DVA and require no I/O to "read".
4245 * Create an anonymous arc buf to back it.
4247 hdr = buf_hash_find(guid, bp, &hash_lock);
4250 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) {
4252 *arc_flags |= ARC_FLAG_CACHED;
4254 if (HDR_IO_IN_PROGRESS(hdr)) {
4256 if (*arc_flags & ARC_FLAG_WAIT) {
4257 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4258 mutex_exit(hash_lock);
4261 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4264 arc_callback_t *acb = NULL;
4266 acb = kmem_zalloc(sizeof (arc_callback_t),
4268 acb->acb_done = done;
4269 acb->acb_private = private;
4271 acb->acb_zio_dummy = zio_null(pio,
4272 spa, NULL, NULL, NULL, zio_flags);
4274 ASSERT(acb->acb_done != NULL);
4275 acb->acb_next = hdr->b_l1hdr.b_acb;
4276 hdr->b_l1hdr.b_acb = acb;
4277 add_reference(hdr, hash_lock, private);
4278 mutex_exit(hash_lock);
4281 mutex_exit(hash_lock);
4285 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4286 hdr->b_l1hdr.b_state == arc_mfu);
4289 add_reference(hdr, hash_lock, private);
4291 * If this block is already in use, create a new
4292 * copy of the data so that we will be guaranteed
4293 * that arc_release() will always succeed.
4295 buf = hdr->b_l1hdr.b_buf;
4297 ASSERT(buf->b_data);
4298 if (HDR_BUF_AVAILABLE(hdr)) {
4299 ASSERT(buf->b_efunc == NULL);
4300 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4302 buf = arc_buf_clone(buf);
4305 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4306 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4307 hdr->b_flags |= ARC_FLAG_PREFETCH;
4309 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4310 arc_access(hdr, hash_lock);
4311 if (*arc_flags & ARC_FLAG_L2CACHE)
4312 hdr->b_flags |= ARC_FLAG_L2CACHE;
4313 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4314 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4315 mutex_exit(hash_lock);
4316 ARCSTAT_BUMP(arcstat_hits);
4317 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4318 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4319 data, metadata, hits);
4322 done(NULL, buf, private);
4324 uint64_t size = BP_GET_LSIZE(bp);
4325 arc_callback_t *acb;
4328 boolean_t devw = B_FALSE;
4329 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
4330 int32_t b_asize = 0;
4333 /* this block is not in the cache */
4334 arc_buf_hdr_t *exists = NULL;
4335 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4336 buf = arc_buf_alloc(spa, size, private, type);
4338 if (!BP_IS_EMBEDDED(bp)) {
4339 hdr->b_dva = *BP_IDENTITY(bp);
4340 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4341 exists = buf_hash_insert(hdr, &hash_lock);
4343 if (exists != NULL) {
4344 /* somebody beat us to the hash insert */
4345 mutex_exit(hash_lock);
4346 buf_discard_identity(hdr);
4347 (void) arc_buf_remove_ref(buf, private);
4348 goto top; /* restart the IO request */
4351 /* if this is a prefetch, we don't have a reference */
4352 if (*arc_flags & ARC_FLAG_PREFETCH) {
4353 (void) remove_reference(hdr, hash_lock,
4355 hdr->b_flags |= ARC_FLAG_PREFETCH;
4357 if (*arc_flags & ARC_FLAG_L2CACHE)
4358 hdr->b_flags |= ARC_FLAG_L2CACHE;
4359 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4360 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4361 if (BP_GET_LEVEL(bp) > 0)
4362 hdr->b_flags |= ARC_FLAG_INDIRECT;
4365 * This block is in the ghost cache. If it was L2-only
4366 * (and thus didn't have an L1 hdr), we realloc the
4367 * header to add an L1 hdr.
4369 if (!HDR_HAS_L1HDR(hdr)) {
4370 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4374 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4375 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4376 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4377 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4379 /* if this is a prefetch, we don't have a reference */
4380 if (*arc_flags & ARC_FLAG_PREFETCH)
4381 hdr->b_flags |= ARC_FLAG_PREFETCH;
4383 add_reference(hdr, hash_lock, private);
4384 if (*arc_flags & ARC_FLAG_L2CACHE)
4385 hdr->b_flags |= ARC_FLAG_L2CACHE;
4386 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4387 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4388 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
4391 buf->b_efunc = NULL;
4392 buf->b_private = NULL;
4394 hdr->b_l1hdr.b_buf = buf;
4395 ASSERT0(hdr->b_l1hdr.b_datacnt);
4396 hdr->b_l1hdr.b_datacnt = 1;
4397 arc_get_data_buf(buf);
4398 arc_access(hdr, hash_lock);
4401 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4403 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4404 acb->acb_done = done;
4405 acb->acb_private = private;
4407 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4408 hdr->b_l1hdr.b_acb = acb;
4409 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4411 if (HDR_HAS_L2HDR(hdr) &&
4412 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
4413 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
4414 addr = hdr->b_l2hdr.b_daddr;
4415 b_compress = HDR_GET_COMPRESS(hdr);
4416 b_asize = hdr->b_l2hdr.b_asize;
4418 * Lock out device removal.
4420 if (vdev_is_dead(vd) ||
4421 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
4425 if (hash_lock != NULL)
4426 mutex_exit(hash_lock);
4429 * At this point, we have a level 1 cache miss. Try again in
4430 * L2ARC if possible.
4432 ASSERT3U(hdr->b_size, ==, size);
4433 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
4434 uint64_t, size, zbookmark_phys_t *, zb);
4435 ARCSTAT_BUMP(arcstat_misses);
4436 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4437 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4438 data, metadata, misses);
4440 curthread->td_ru.ru_inblock++;
4443 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
4445 * Read from the L2ARC if the following are true:
4446 * 1. The L2ARC vdev was previously cached.
4447 * 2. This buffer still has L2ARC metadata.
4448 * 3. This buffer isn't currently writing to the L2ARC.
4449 * 4. The L2ARC entry wasn't evicted, which may
4450 * also have invalidated the vdev.
4451 * 5. This isn't prefetch and l2arc_noprefetch is set.
4453 if (HDR_HAS_L2HDR(hdr) &&
4454 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
4455 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
4456 l2arc_read_callback_t *cb;
4458 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
4459 ARCSTAT_BUMP(arcstat_l2_hits);
4461 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
4463 cb->l2rcb_buf = buf;
4464 cb->l2rcb_spa = spa;
4467 cb->l2rcb_flags = zio_flags;
4468 cb->l2rcb_compress = b_compress;
4470 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
4471 addr + size < vd->vdev_psize -
4472 VDEV_LABEL_END_SIZE);
4475 * l2arc read. The SCL_L2ARC lock will be
4476 * released by l2arc_read_done().
4477 * Issue a null zio if the underlying buffer
4478 * was squashed to zero size by compression.
4480 if (b_compress == ZIO_COMPRESS_EMPTY) {
4481 rzio = zio_null(pio, spa, vd,
4482 l2arc_read_done, cb,
4483 zio_flags | ZIO_FLAG_DONT_CACHE |
4485 ZIO_FLAG_DONT_PROPAGATE |
4486 ZIO_FLAG_DONT_RETRY);
4488 rzio = zio_read_phys(pio, vd, addr,
4489 b_asize, buf->b_data,
4491 l2arc_read_done, cb, priority,
4492 zio_flags | ZIO_FLAG_DONT_CACHE |
4494 ZIO_FLAG_DONT_PROPAGATE |
4495 ZIO_FLAG_DONT_RETRY, B_FALSE);
4497 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
4499 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
4501 if (*arc_flags & ARC_FLAG_NOWAIT) {
4506 ASSERT(*arc_flags & ARC_FLAG_WAIT);
4507 if (zio_wait(rzio) == 0)
4510 /* l2arc read error; goto zio_read() */
4512 DTRACE_PROBE1(l2arc__miss,
4513 arc_buf_hdr_t *, hdr);
4514 ARCSTAT_BUMP(arcstat_l2_misses);
4515 if (HDR_L2_WRITING(hdr))
4516 ARCSTAT_BUMP(arcstat_l2_rw_clash);
4517 spa_config_exit(spa, SCL_L2ARC, vd);
4521 spa_config_exit(spa, SCL_L2ARC, vd);
4522 if (l2arc_ndev != 0) {
4523 DTRACE_PROBE1(l2arc__miss,
4524 arc_buf_hdr_t *, hdr);
4525 ARCSTAT_BUMP(arcstat_l2_misses);
4529 rzio = zio_read(pio, spa, bp, buf->b_data, size,
4530 arc_read_done, buf, priority, zio_flags, zb);
4532 if (*arc_flags & ARC_FLAG_WAIT)
4533 return (zio_wait(rzio));
4535 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4542 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
4544 ASSERT(buf->b_hdr != NULL);
4545 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon);
4546 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) ||
4548 ASSERT(buf->b_efunc == NULL);
4549 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
4551 buf->b_efunc = func;
4552 buf->b_private = private;
4556 * Notify the arc that a block was freed, and thus will never be used again.
4559 arc_freed(spa_t *spa, const blkptr_t *bp)
4562 kmutex_t *hash_lock;
4563 uint64_t guid = spa_load_guid(spa);
4565 ASSERT(!BP_IS_EMBEDDED(bp));
4567 hdr = buf_hash_find(guid, bp, &hash_lock);
4570 if (HDR_BUF_AVAILABLE(hdr)) {
4571 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
4572 add_reference(hdr, hash_lock, FTAG);
4573 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4574 mutex_exit(hash_lock);
4576 arc_release(buf, FTAG);
4577 (void) arc_buf_remove_ref(buf, FTAG);
4579 mutex_exit(hash_lock);
4585 * Clear the user eviction callback set by arc_set_callback(), first calling
4586 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4587 * clearing the callback may result in the arc_buf being destroyed. However,
4588 * it will not result in the *last* arc_buf being destroyed, hence the data
4589 * will remain cached in the ARC. We make a copy of the arc buffer here so
4590 * that we can process the callback without holding any locks.
4592 * It's possible that the callback is already in the process of being cleared
4593 * by another thread. In this case we can not clear the callback.
4595 * Returns B_TRUE if the callback was successfully called and cleared.
4598 arc_clear_callback(arc_buf_t *buf)
4601 kmutex_t *hash_lock;
4602 arc_evict_func_t *efunc = buf->b_efunc;
4603 void *private = buf->b_private;
4605 mutex_enter(&buf->b_evict_lock);
4609 * We are in arc_do_user_evicts().
4611 ASSERT(buf->b_data == NULL);
4612 mutex_exit(&buf->b_evict_lock);
4614 } else if (buf->b_data == NULL) {
4616 * We are on the eviction list; process this buffer now
4617 * but let arc_do_user_evicts() do the reaping.
4619 buf->b_efunc = NULL;
4620 mutex_exit(&buf->b_evict_lock);
4621 VERIFY0(efunc(private));
4624 hash_lock = HDR_LOCK(hdr);
4625 mutex_enter(hash_lock);
4627 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4629 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <,
4630 hdr->b_l1hdr.b_datacnt);
4631 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4632 hdr->b_l1hdr.b_state == arc_mfu);
4634 buf->b_efunc = NULL;
4635 buf->b_private = NULL;
4637 if (hdr->b_l1hdr.b_datacnt > 1) {
4638 mutex_exit(&buf->b_evict_lock);
4639 arc_buf_destroy(buf, TRUE);
4641 ASSERT(buf == hdr->b_l1hdr.b_buf);
4642 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4643 mutex_exit(&buf->b_evict_lock);
4646 mutex_exit(hash_lock);
4647 VERIFY0(efunc(private));
4652 * Release this buffer from the cache, making it an anonymous buffer. This
4653 * must be done after a read and prior to modifying the buffer contents.
4654 * If the buffer has more than one reference, we must make
4655 * a new hdr for the buffer.
4658 arc_release(arc_buf_t *buf, void *tag)
4660 arc_buf_hdr_t *hdr = buf->b_hdr;
4663 * It would be nice to assert that if it's DMU metadata (level >
4664 * 0 || it's the dnode file), then it must be syncing context.
4665 * But we don't know that information at this level.
4668 mutex_enter(&buf->b_evict_lock);
4670 ASSERT(HDR_HAS_L1HDR(hdr));
4673 * We don't grab the hash lock prior to this check, because if
4674 * the buffer's header is in the arc_anon state, it won't be
4675 * linked into the hash table.
4677 if (hdr->b_l1hdr.b_state == arc_anon) {
4678 mutex_exit(&buf->b_evict_lock);
4679 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4680 ASSERT(!HDR_IN_HASH_TABLE(hdr));
4681 ASSERT(!HDR_HAS_L2HDR(hdr));
4682 ASSERT(BUF_EMPTY(hdr));
4683 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1);
4684 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
4685 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
4687 ASSERT3P(buf->b_efunc, ==, NULL);
4688 ASSERT3P(buf->b_private, ==, NULL);
4690 hdr->b_l1hdr.b_arc_access = 0;
4696 kmutex_t *hash_lock = HDR_LOCK(hdr);
4697 mutex_enter(hash_lock);
4700 * This assignment is only valid as long as the hash_lock is
4701 * held, we must be careful not to reference state or the
4702 * b_state field after dropping the lock.
4704 arc_state_t *state = hdr->b_l1hdr.b_state;
4705 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4706 ASSERT3P(state, !=, arc_anon);
4708 /* this buffer is not on any list */
4709 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
4711 if (HDR_HAS_L2HDR(hdr)) {
4712 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4715 * We have to recheck this conditional again now that
4716 * we're holding the l2ad_mtx to prevent a race with
4717 * another thread which might be concurrently calling
4718 * l2arc_evict(). In that case, l2arc_evict() might have
4719 * destroyed the header's L2 portion as we were waiting
4720 * to acquire the l2ad_mtx.
4722 if (HDR_HAS_L2HDR(hdr)) {
4723 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET)
4724 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev,
4725 hdr->b_l2hdr.b_daddr,
4726 hdr->b_l2hdr.b_asize, 0);
4727 arc_hdr_l2hdr_destroy(hdr);
4730 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4734 * Do we have more than one buf?
4736 if (hdr->b_l1hdr.b_datacnt > 1) {
4737 arc_buf_hdr_t *nhdr;
4739 uint64_t blksz = hdr->b_size;
4740 uint64_t spa = hdr->b_spa;
4741 arc_buf_contents_t type = arc_buf_type(hdr);
4742 uint32_t flags = hdr->b_flags;
4744 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
4746 * Pull the data off of this hdr and attach it to
4747 * a new anonymous hdr.
4749 (void) remove_reference(hdr, hash_lock, tag);
4750 bufp = &hdr->b_l1hdr.b_buf;
4751 while (*bufp != buf)
4752 bufp = &(*bufp)->b_next;
4753 *bufp = buf->b_next;
4756 ASSERT3P(state, !=, arc_l2c_only);
4758 (void) refcount_remove_many(
4759 &state->arcs_size, hdr->b_size, buf);
4761 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
4762 ASSERT3P(state, !=, arc_l2c_only);
4763 uint64_t *size = &state->arcs_lsize[type];
4764 ASSERT3U(*size, >=, hdr->b_size);
4765 atomic_add_64(size, -hdr->b_size);
4769 * We're releasing a duplicate user data buffer, update
4770 * our statistics accordingly.
4772 if (HDR_ISTYPE_DATA(hdr)) {
4773 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
4774 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
4777 hdr->b_l1hdr.b_datacnt -= 1;
4778 arc_cksum_verify(buf);
4780 arc_buf_unwatch(buf);
4781 #endif /* illumos */
4783 mutex_exit(hash_lock);
4785 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
4786 nhdr->b_size = blksz;
4789 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING;
4790 nhdr->b_flags |= arc_bufc_to_flags(type);
4791 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
4793 nhdr->b_l1hdr.b_buf = buf;
4794 nhdr->b_l1hdr.b_datacnt = 1;
4795 nhdr->b_l1hdr.b_state = arc_anon;
4796 nhdr->b_l1hdr.b_arc_access = 0;
4797 nhdr->b_l1hdr.b_tmp_cdata = NULL;
4798 nhdr->b_freeze_cksum = NULL;
4800 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
4802 mutex_exit(&buf->b_evict_lock);
4803 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf);
4805 mutex_exit(&buf->b_evict_lock);
4806 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
4807 /* protected by hash lock, or hdr is on arc_anon */
4808 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4809 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4810 arc_change_state(arc_anon, hdr, hash_lock);
4811 hdr->b_l1hdr.b_arc_access = 0;
4812 mutex_exit(hash_lock);
4814 buf_discard_identity(hdr);
4817 buf->b_efunc = NULL;
4818 buf->b_private = NULL;
4822 arc_released(arc_buf_t *buf)
4826 mutex_enter(&buf->b_evict_lock);
4827 released = (buf->b_data != NULL &&
4828 buf->b_hdr->b_l1hdr.b_state == arc_anon);
4829 mutex_exit(&buf->b_evict_lock);
4835 arc_referenced(arc_buf_t *buf)
4839 mutex_enter(&buf->b_evict_lock);
4840 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
4841 mutex_exit(&buf->b_evict_lock);
4842 return (referenced);
4847 arc_write_ready(zio_t *zio)
4849 arc_write_callback_t *callback = zio->io_private;
4850 arc_buf_t *buf = callback->awcb_buf;
4851 arc_buf_hdr_t *hdr = buf->b_hdr;
4853 ASSERT(HDR_HAS_L1HDR(hdr));
4854 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
4855 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4856 callback->awcb_ready(zio, buf, callback->awcb_private);
4859 * If the IO is already in progress, then this is a re-write
4860 * attempt, so we need to thaw and re-compute the cksum.
4861 * It is the responsibility of the callback to handle the
4862 * accounting for any re-write attempt.
4864 if (HDR_IO_IN_PROGRESS(hdr)) {
4865 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
4866 if (hdr->b_freeze_cksum != NULL) {
4867 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
4868 hdr->b_freeze_cksum = NULL;
4870 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
4872 arc_cksum_compute(buf, B_FALSE);
4873 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4877 * The SPA calls this callback for each physical write that happens on behalf
4878 * of a logical write. See the comment in dbuf_write_physdone() for details.
4881 arc_write_physdone(zio_t *zio)
4883 arc_write_callback_t *cb = zio->io_private;
4884 if (cb->awcb_physdone != NULL)
4885 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
4889 arc_write_done(zio_t *zio)
4891 arc_write_callback_t *callback = zio->io_private;
4892 arc_buf_t *buf = callback->awcb_buf;
4893 arc_buf_hdr_t *hdr = buf->b_hdr;
4895 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4897 if (zio->io_error == 0) {
4898 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
4899 buf_discard_identity(hdr);
4901 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
4902 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
4905 ASSERT(BUF_EMPTY(hdr));
4909 * If the block to be written was all-zero or compressed enough to be
4910 * embedded in the BP, no write was performed so there will be no
4911 * dva/birth/checksum. The buffer must therefore remain anonymous
4914 if (!BUF_EMPTY(hdr)) {
4915 arc_buf_hdr_t *exists;
4916 kmutex_t *hash_lock;
4918 ASSERT(zio->io_error == 0);
4920 arc_cksum_verify(buf);
4922 exists = buf_hash_insert(hdr, &hash_lock);
4923 if (exists != NULL) {
4925 * This can only happen if we overwrite for
4926 * sync-to-convergence, because we remove
4927 * buffers from the hash table when we arc_free().
4929 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
4930 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4931 panic("bad overwrite, hdr=%p exists=%p",
4932 (void *)hdr, (void *)exists);
4933 ASSERT(refcount_is_zero(
4934 &exists->b_l1hdr.b_refcnt));
4935 arc_change_state(arc_anon, exists, hash_lock);
4936 mutex_exit(hash_lock);
4937 arc_hdr_destroy(exists);
4938 exists = buf_hash_insert(hdr, &hash_lock);
4939 ASSERT3P(exists, ==, NULL);
4940 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
4942 ASSERT(zio->io_prop.zp_nopwrite);
4943 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4944 panic("bad nopwrite, hdr=%p exists=%p",
4945 (void *)hdr, (void *)exists);
4948 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
4949 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
4950 ASSERT(BP_GET_DEDUP(zio->io_bp));
4951 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
4954 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4955 /* if it's not anon, we are doing a scrub */
4956 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
4957 arc_access(hdr, hash_lock);
4958 mutex_exit(hash_lock);
4960 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4963 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4964 callback->awcb_done(zio, buf, callback->awcb_private);
4966 kmem_free(callback, sizeof (arc_write_callback_t));
4970 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
4971 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
4972 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
4973 arc_done_func_t *done, void *private, zio_priority_t priority,
4974 int zio_flags, const zbookmark_phys_t *zb)
4976 arc_buf_hdr_t *hdr = buf->b_hdr;
4977 arc_write_callback_t *callback;
4980 ASSERT(ready != NULL);
4981 ASSERT(done != NULL);
4982 ASSERT(!HDR_IO_ERROR(hdr));
4983 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4984 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4985 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4987 hdr->b_flags |= ARC_FLAG_L2CACHE;
4989 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4990 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
4991 callback->awcb_ready = ready;
4992 callback->awcb_physdone = physdone;
4993 callback->awcb_done = done;
4994 callback->awcb_private = private;
4995 callback->awcb_buf = buf;
4997 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
4998 arc_write_ready, arc_write_physdone, arc_write_done, callback,
4999 priority, zio_flags, zb);
5005 arc_memory_throttle(uint64_t reserve, uint64_t txg)
5008 uint64_t available_memory = ptob(freemem);
5009 static uint64_t page_load = 0;
5010 static uint64_t last_txg = 0;
5012 #if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC)
5014 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE)));
5017 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100)
5020 if (txg > last_txg) {
5025 * If we are in pageout, we know that memory is already tight,
5026 * the arc is already going to be evicting, so we just want to
5027 * continue to let page writes occur as quickly as possible.
5029 if (curproc == pageproc) {
5030 if (page_load > MAX(ptob(minfree), available_memory) / 4)
5031 return (SET_ERROR(ERESTART));
5032 /* Note: reserve is inflated, so we deflate */
5033 page_load += reserve / 8;
5035 } else if (page_load > 0 && arc_reclaim_needed()) {
5036 /* memory is low, delay before restarting */
5037 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
5038 return (SET_ERROR(EAGAIN));
5046 arc_tempreserve_clear(uint64_t reserve)
5048 atomic_add_64(&arc_tempreserve, -reserve);
5049 ASSERT((int64_t)arc_tempreserve >= 0);
5053 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
5058 if (reserve > arc_c/4 && !arc_no_grow) {
5059 arc_c = MIN(arc_c_max, reserve * 4);
5060 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c);
5062 if (reserve > arc_c)
5063 return (SET_ERROR(ENOMEM));
5066 * Don't count loaned bufs as in flight dirty data to prevent long
5067 * network delays from blocking transactions that are ready to be
5068 * assigned to a txg.
5070 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
5071 arc_loaned_bytes), 0);
5074 * Writes will, almost always, require additional memory allocations
5075 * in order to compress/encrypt/etc the data. We therefore need to
5076 * make sure that there is sufficient available memory for this.
5078 error = arc_memory_throttle(reserve, txg);
5083 * Throttle writes when the amount of dirty data in the cache
5084 * gets too large. We try to keep the cache less than half full
5085 * of dirty blocks so that our sync times don't grow too large.
5086 * Note: if two requests come in concurrently, we might let them
5087 * both succeed, when one of them should fail. Not a huge deal.
5090 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
5091 anon_size > arc_c / 4) {
5092 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5093 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5094 arc_tempreserve>>10,
5095 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
5096 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
5097 reserve>>10, arc_c>>10);
5098 return (SET_ERROR(ERESTART));
5100 atomic_add_64(&arc_tempreserve, reserve);
5105 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
5106 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
5108 size->value.ui64 = refcount_count(&state->arcs_size);
5109 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
5110 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
5114 arc_kstat_update(kstat_t *ksp, int rw)
5116 arc_stats_t *as = ksp->ks_data;
5118 if (rw == KSTAT_WRITE) {
5121 arc_kstat_update_state(arc_anon,
5122 &as->arcstat_anon_size,
5123 &as->arcstat_anon_evictable_data,
5124 &as->arcstat_anon_evictable_metadata);
5125 arc_kstat_update_state(arc_mru,
5126 &as->arcstat_mru_size,
5127 &as->arcstat_mru_evictable_data,
5128 &as->arcstat_mru_evictable_metadata);
5129 arc_kstat_update_state(arc_mru_ghost,
5130 &as->arcstat_mru_ghost_size,
5131 &as->arcstat_mru_ghost_evictable_data,
5132 &as->arcstat_mru_ghost_evictable_metadata);
5133 arc_kstat_update_state(arc_mfu,
5134 &as->arcstat_mfu_size,
5135 &as->arcstat_mfu_evictable_data,
5136 &as->arcstat_mfu_evictable_metadata);
5137 arc_kstat_update_state(arc_mfu_ghost,
5138 &as->arcstat_mfu_ghost_size,
5139 &as->arcstat_mfu_ghost_evictable_data,
5140 &as->arcstat_mfu_ghost_evictable_metadata);
5147 * This function *must* return indices evenly distributed between all
5148 * sublists of the multilist. This is needed due to how the ARC eviction
5149 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5150 * distributed between all sublists and uses this assumption when
5151 * deciding which sublist to evict from and how much to evict from it.
5154 arc_state_multilist_index_func(multilist_t *ml, void *obj)
5156 arc_buf_hdr_t *hdr = obj;
5159 * We rely on b_dva to generate evenly distributed index
5160 * numbers using buf_hash below. So, as an added precaution,
5161 * let's make sure we never add empty buffers to the arc lists.
5163 ASSERT(!BUF_EMPTY(hdr));
5166 * The assumption here, is the hash value for a given
5167 * arc_buf_hdr_t will remain constant throughout it's lifetime
5168 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
5169 * Thus, we don't need to store the header's sublist index
5170 * on insertion, as this index can be recalculated on removal.
5172 * Also, the low order bits of the hash value are thought to be
5173 * distributed evenly. Otherwise, in the case that the multilist
5174 * has a power of two number of sublists, each sublists' usage
5175 * would not be evenly distributed.
5177 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
5178 multilist_get_num_sublists(ml));
5182 static eventhandler_tag arc_event_lowmem = NULL;
5185 arc_lowmem(void *arg __unused, int howto __unused)
5188 mutex_enter(&arc_reclaim_lock);
5189 /* XXX: Memory deficit should be passed as argument. */
5190 needfree = btoc(arc_c >> arc_shrink_shift);
5191 DTRACE_PROBE(arc__needfree);
5192 cv_signal(&arc_reclaim_thread_cv);
5195 * It is unsafe to block here in arbitrary threads, because we can come
5196 * here from ARC itself and may hold ARC locks and thus risk a deadlock
5197 * with ARC reclaim thread.
5199 if (curproc == pageproc)
5200 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5201 mutex_exit(&arc_reclaim_lock);
5208 int i, prefetch_tunable_set = 0;
5210 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
5211 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
5212 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
5214 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
5215 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL);
5217 /* Convert seconds to clock ticks */
5218 arc_min_prefetch_lifespan = 1 * hz;
5220 /* Start out with 1/8 of all memory */
5221 arc_c = kmem_size() / 8;
5226 * On architectures where the physical memory can be larger
5227 * than the addressable space (intel in 32-bit mode), we may
5228 * need to limit the cache to 1/8 of VM size.
5230 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
5233 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
5234 arc_c_min = MAX(arc_c / 4, 16 << 20);
5235 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
5236 if (arc_c * 8 >= 1 << 30)
5237 arc_c_max = (arc_c * 8) - (1 << 30);
5239 arc_c_max = arc_c_min;
5240 arc_c_max = MAX(arc_c * 5, arc_c_max);
5244 * Allow the tunables to override our calculations if they are
5245 * reasonable (ie. over 16MB)
5247 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size())
5248 arc_c_max = zfs_arc_max;
5249 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max)
5250 arc_c_min = zfs_arc_min;
5254 arc_p = (arc_c >> 1);
5256 /* limit meta-data to 1/4 of the arc capacity */
5257 arc_meta_limit = arc_c_max / 4;
5259 /* Allow the tunable to override if it is reasonable */
5260 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
5261 arc_meta_limit = zfs_arc_meta_limit;
5263 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
5264 arc_c_min = arc_meta_limit / 2;
5266 if (zfs_arc_meta_min > 0) {
5267 arc_meta_min = zfs_arc_meta_min;
5269 arc_meta_min = arc_c_min / 2;
5272 if (zfs_arc_grow_retry > 0)
5273 arc_grow_retry = zfs_arc_grow_retry;
5275 if (zfs_arc_shrink_shift > 0)
5276 arc_shrink_shift = zfs_arc_shrink_shift;
5279 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
5281 if (arc_no_grow_shift >= arc_shrink_shift)
5282 arc_no_grow_shift = arc_shrink_shift - 1;
5284 if (zfs_arc_p_min_shift > 0)
5285 arc_p_min_shift = zfs_arc_p_min_shift;
5287 if (zfs_arc_num_sublists_per_state < 1)
5288 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1);
5290 /* if kmem_flags are set, lets try to use less memory */
5291 if (kmem_debugging())
5293 if (arc_c < arc_c_min)
5296 zfs_arc_min = arc_c_min;
5297 zfs_arc_max = arc_c_max;
5299 arc_anon = &ARC_anon;
5301 arc_mru_ghost = &ARC_mru_ghost;
5303 arc_mfu_ghost = &ARC_mfu_ghost;
5304 arc_l2c_only = &ARC_l2c_only;
5307 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5308 sizeof (arc_buf_hdr_t),
5309 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5310 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5311 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5312 sizeof (arc_buf_hdr_t),
5313 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5314 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5315 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5316 sizeof (arc_buf_hdr_t),
5317 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5318 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5319 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5320 sizeof (arc_buf_hdr_t),
5321 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5322 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5323 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5324 sizeof (arc_buf_hdr_t),
5325 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5326 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5327 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5328 sizeof (arc_buf_hdr_t),
5329 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5330 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5331 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5332 sizeof (arc_buf_hdr_t),
5333 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5334 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5335 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5336 sizeof (arc_buf_hdr_t),
5337 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5338 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5339 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5340 sizeof (arc_buf_hdr_t),
5341 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5342 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5343 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
5344 sizeof (arc_buf_hdr_t),
5345 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5346 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5348 refcount_create(&arc_anon->arcs_size);
5349 refcount_create(&arc_mru->arcs_size);
5350 refcount_create(&arc_mru_ghost->arcs_size);
5351 refcount_create(&arc_mfu->arcs_size);
5352 refcount_create(&arc_mfu_ghost->arcs_size);
5353 refcount_create(&arc_l2c_only->arcs_size);
5357 arc_reclaim_thread_exit = FALSE;
5358 arc_user_evicts_thread_exit = FALSE;
5359 arc_eviction_list = NULL;
5360 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
5362 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
5363 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
5365 if (arc_ksp != NULL) {
5366 arc_ksp->ks_data = &arc_stats;
5367 arc_ksp->ks_update = arc_kstat_update;
5368 kstat_install(arc_ksp);
5371 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
5372 TS_RUN, minclsyspri);
5375 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
5376 EVENTHANDLER_PRI_FIRST);
5379 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0,
5380 TS_RUN, minclsyspri);
5386 * Calculate maximum amount of dirty data per pool.
5388 * If it has been set by /etc/system, take that.
5389 * Otherwise, use a percentage of physical memory defined by
5390 * zfs_dirty_data_max_percent (default 10%) with a cap at
5391 * zfs_dirty_data_max_max (default 4GB).
5393 if (zfs_dirty_data_max == 0) {
5394 zfs_dirty_data_max = ptob(physmem) *
5395 zfs_dirty_data_max_percent / 100;
5396 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
5397 zfs_dirty_data_max_max);
5401 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
5402 prefetch_tunable_set = 1;
5405 if (prefetch_tunable_set == 0) {
5406 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
5408 printf(" add \"vfs.zfs.prefetch_disable=0\" "
5409 "to /boot/loader.conf.\n");
5410 zfs_prefetch_disable = 1;
5413 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
5414 prefetch_tunable_set == 0) {
5415 printf("ZFS NOTICE: Prefetch is disabled by default if less "
5416 "than 4GB of RAM is present;\n"
5417 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
5418 "to /boot/loader.conf.\n");
5419 zfs_prefetch_disable = 1;
5422 /* Warn about ZFS memory and address space requirements. */
5423 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
5424 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
5425 "expect unstable behavior.\n");
5427 if (kmem_size() < 512 * (1 << 20)) {
5428 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
5429 "expect unstable behavior.\n");
5430 printf(" Consider tuning vm.kmem_size and "
5431 "vm.kmem_size_max\n");
5432 printf(" in /boot/loader.conf.\n");
5440 mutex_enter(&arc_reclaim_lock);
5441 arc_reclaim_thread_exit = TRUE;
5443 * The reclaim thread will set arc_reclaim_thread_exit back to
5444 * FALSE when it is finished exiting; we're waiting for that.
5446 while (arc_reclaim_thread_exit) {
5447 cv_signal(&arc_reclaim_thread_cv);
5448 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
5450 mutex_exit(&arc_reclaim_lock);
5452 mutex_enter(&arc_user_evicts_lock);
5453 arc_user_evicts_thread_exit = TRUE;
5455 * The user evicts thread will set arc_user_evicts_thread_exit
5456 * to FALSE when it is finished exiting; we're waiting for that.
5458 while (arc_user_evicts_thread_exit) {
5459 cv_signal(&arc_user_evicts_cv);
5460 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock);
5462 mutex_exit(&arc_user_evicts_lock);
5464 /* Use TRUE to ensure *all* buffers are evicted */
5465 arc_flush(NULL, TRUE);
5469 if (arc_ksp != NULL) {
5470 kstat_delete(arc_ksp);
5474 mutex_destroy(&arc_reclaim_lock);
5475 cv_destroy(&arc_reclaim_thread_cv);
5476 cv_destroy(&arc_reclaim_waiters_cv);
5478 mutex_destroy(&arc_user_evicts_lock);
5479 cv_destroy(&arc_user_evicts_cv);
5481 refcount_destroy(&arc_anon->arcs_size);
5482 refcount_destroy(&arc_mru->arcs_size);
5483 refcount_destroy(&arc_mru_ghost->arcs_size);
5484 refcount_destroy(&arc_mfu->arcs_size);
5485 refcount_destroy(&arc_mfu_ghost->arcs_size);
5486 refcount_destroy(&arc_l2c_only->arcs_size);
5488 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
5489 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5490 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
5491 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
5492 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
5493 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
5494 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
5495 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
5499 ASSERT0(arc_loaned_bytes);
5502 if (arc_event_lowmem != NULL)
5503 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
5510 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5511 * It uses dedicated storage devices to hold cached data, which are populated
5512 * using large infrequent writes. The main role of this cache is to boost
5513 * the performance of random read workloads. The intended L2ARC devices
5514 * include short-stroked disks, solid state disks, and other media with
5515 * substantially faster read latency than disk.
5517 * +-----------------------+
5519 * +-----------------------+
5522 * l2arc_feed_thread() arc_read()
5526 * +---------------+ |
5528 * +---------------+ |
5533 * +-------+ +-------+
5535 * | cache | | cache |
5536 * +-------+ +-------+
5537 * +=========+ .-----.
5538 * : L2ARC : |-_____-|
5539 * : devices : | Disks |
5540 * +=========+ `-_____-'
5542 * Read requests are satisfied from the following sources, in order:
5545 * 2) vdev cache of L2ARC devices
5547 * 4) vdev cache of disks
5550 * Some L2ARC device types exhibit extremely slow write performance.
5551 * To accommodate for this there are some significant differences between
5552 * the L2ARC and traditional cache design:
5554 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5555 * the ARC behave as usual, freeing buffers and placing headers on ghost
5556 * lists. The ARC does not send buffers to the L2ARC during eviction as
5557 * this would add inflated write latencies for all ARC memory pressure.
5559 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5560 * It does this by periodically scanning buffers from the eviction-end of
5561 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5562 * not already there. It scans until a headroom of buffers is satisfied,
5563 * which itself is a buffer for ARC eviction. If a compressible buffer is
5564 * found during scanning and selected for writing to an L2ARC device, we
5565 * temporarily boost scanning headroom during the next scan cycle to make
5566 * sure we adapt to compression effects (which might significantly reduce
5567 * the data volume we write to L2ARC). The thread that does this is
5568 * l2arc_feed_thread(), illustrated below; example sizes are included to
5569 * provide a better sense of ratio than this diagram:
5572 * +---------------------+----------+
5573 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5574 * +---------------------+----------+ | o L2ARC eligible
5575 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5576 * +---------------------+----------+ |
5577 * 15.9 Gbytes ^ 32 Mbytes |
5579 * l2arc_feed_thread()
5581 * l2arc write hand <--[oooo]--'
5585 * +==============================+
5586 * L2ARC dev |####|#|###|###| |####| ... |
5587 * +==============================+
5590 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5591 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5592 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5593 * safe to say that this is an uncommon case, since buffers at the end of
5594 * the ARC lists have moved there due to inactivity.
5596 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5597 * then the L2ARC simply misses copying some buffers. This serves as a
5598 * pressure valve to prevent heavy read workloads from both stalling the ARC
5599 * with waits and clogging the L2ARC with writes. This also helps prevent
5600 * the potential for the L2ARC to churn if it attempts to cache content too
5601 * quickly, such as during backups of the entire pool.
5603 * 5. After system boot and before the ARC has filled main memory, there are
5604 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5605 * lists can remain mostly static. Instead of searching from tail of these
5606 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5607 * for eligible buffers, greatly increasing its chance of finding them.
5609 * The L2ARC device write speed is also boosted during this time so that
5610 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5611 * there are no L2ARC reads, and no fear of degrading read performance
5612 * through increased writes.
5614 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5615 * the vdev queue can aggregate them into larger and fewer writes. Each
5616 * device is written to in a rotor fashion, sweeping writes through
5617 * available space then repeating.
5619 * 7. The L2ARC does not store dirty content. It never needs to flush
5620 * write buffers back to disk based storage.
5622 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5623 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5625 * The performance of the L2ARC can be tweaked by a number of tunables, which
5626 * may be necessary for different workloads:
5628 * l2arc_write_max max write bytes per interval
5629 * l2arc_write_boost extra write bytes during device warmup
5630 * l2arc_noprefetch skip caching prefetched buffers
5631 * l2arc_headroom number of max device writes to precache
5632 * l2arc_headroom_boost when we find compressed buffers during ARC
5633 * scanning, we multiply headroom by this
5634 * percentage factor for the next scan cycle,
5635 * since more compressed buffers are likely to
5637 * l2arc_feed_secs seconds between L2ARC writing
5639 * Tunables may be removed or added as future performance improvements are
5640 * integrated, and also may become zpool properties.
5642 * There are three key functions that control how the L2ARC warms up:
5644 * l2arc_write_eligible() check if a buffer is eligible to cache
5645 * l2arc_write_size() calculate how much to write
5646 * l2arc_write_interval() calculate sleep delay between writes
5648 * These three functions determine what to write, how much, and how quickly
5653 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
5656 * A buffer is *not* eligible for the L2ARC if it:
5657 * 1. belongs to a different spa.
5658 * 2. is already cached on the L2ARC.
5659 * 3. has an I/O in progress (it may be an incomplete read).
5660 * 4. is flagged not eligible (zfs property).
5662 if (hdr->b_spa != spa_guid) {
5663 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
5666 if (HDR_HAS_L2HDR(hdr)) {
5667 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
5670 if (HDR_IO_IN_PROGRESS(hdr)) {
5671 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
5674 if (!HDR_L2CACHE(hdr)) {
5675 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
5683 l2arc_write_size(void)
5688 * Make sure our globals have meaningful values in case the user
5691 size = l2arc_write_max;
5693 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
5694 "be greater than zero, resetting it to the default (%d)",
5696 size = l2arc_write_max = L2ARC_WRITE_SIZE;
5699 if (arc_warm == B_FALSE)
5700 size += l2arc_write_boost;
5707 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
5709 clock_t interval, next, now;
5712 * If the ARC lists are busy, increase our write rate; if the
5713 * lists are stale, idle back. This is achieved by checking
5714 * how much we previously wrote - if it was more than half of
5715 * what we wanted, schedule the next write much sooner.
5717 if (l2arc_feed_again && wrote > (wanted / 2))
5718 interval = (hz * l2arc_feed_min_ms) / 1000;
5720 interval = hz * l2arc_feed_secs;
5722 now = ddi_get_lbolt();
5723 next = MAX(now, MIN(now + interval, began + interval));
5729 * Cycle through L2ARC devices. This is how L2ARC load balances.
5730 * If a device is returned, this also returns holding the spa config lock.
5732 static l2arc_dev_t *
5733 l2arc_dev_get_next(void)
5735 l2arc_dev_t *first, *next = NULL;
5738 * Lock out the removal of spas (spa_namespace_lock), then removal
5739 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
5740 * both locks will be dropped and a spa config lock held instead.
5742 mutex_enter(&spa_namespace_lock);
5743 mutex_enter(&l2arc_dev_mtx);
5745 /* if there are no vdevs, there is nothing to do */
5746 if (l2arc_ndev == 0)
5750 next = l2arc_dev_last;
5752 /* loop around the list looking for a non-faulted vdev */
5754 next = list_head(l2arc_dev_list);
5756 next = list_next(l2arc_dev_list, next);
5758 next = list_head(l2arc_dev_list);
5761 /* if we have come back to the start, bail out */
5764 else if (next == first)
5767 } while (vdev_is_dead(next->l2ad_vdev));
5769 /* if we were unable to find any usable vdevs, return NULL */
5770 if (vdev_is_dead(next->l2ad_vdev))
5773 l2arc_dev_last = next;
5776 mutex_exit(&l2arc_dev_mtx);
5779 * Grab the config lock to prevent the 'next' device from being
5780 * removed while we are writing to it.
5783 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
5784 mutex_exit(&spa_namespace_lock);
5790 * Free buffers that were tagged for destruction.
5793 l2arc_do_free_on_write()
5796 l2arc_data_free_t *df, *df_prev;
5798 mutex_enter(&l2arc_free_on_write_mtx);
5799 buflist = l2arc_free_on_write;
5801 for (df = list_tail(buflist); df; df = df_prev) {
5802 df_prev = list_prev(buflist, df);
5803 ASSERT(df->l2df_data != NULL);
5804 ASSERT(df->l2df_func != NULL);
5805 df->l2df_func(df->l2df_data, df->l2df_size);
5806 list_remove(buflist, df);
5807 kmem_free(df, sizeof (l2arc_data_free_t));
5810 mutex_exit(&l2arc_free_on_write_mtx);
5814 * A write to a cache device has completed. Update all headers to allow
5815 * reads from these buffers to begin.
5818 l2arc_write_done(zio_t *zio)
5820 l2arc_write_callback_t *cb;
5823 arc_buf_hdr_t *head, *hdr, *hdr_prev;
5824 kmutex_t *hash_lock;
5825 int64_t bytes_dropped = 0;
5827 cb = zio->io_private;
5829 dev = cb->l2wcb_dev;
5830 ASSERT(dev != NULL);
5831 head = cb->l2wcb_head;
5832 ASSERT(head != NULL);
5833 buflist = &dev->l2ad_buflist;
5834 ASSERT(buflist != NULL);
5835 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
5836 l2arc_write_callback_t *, cb);
5838 if (zio->io_error != 0)
5839 ARCSTAT_BUMP(arcstat_l2_writes_error);
5842 * All writes completed, or an error was hit.
5845 mutex_enter(&dev->l2ad_mtx);
5846 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
5847 hdr_prev = list_prev(buflist, hdr);
5849 hash_lock = HDR_LOCK(hdr);
5852 * We cannot use mutex_enter or else we can deadlock
5853 * with l2arc_write_buffers (due to swapping the order
5854 * the hash lock and l2ad_mtx are taken).
5856 if (!mutex_tryenter(hash_lock)) {
5858 * Missed the hash lock. We must retry so we
5859 * don't leave the ARC_FLAG_L2_WRITING bit set.
5861 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
5864 * We don't want to rescan the headers we've
5865 * already marked as having been written out, so
5866 * we reinsert the head node so we can pick up
5867 * where we left off.
5869 list_remove(buflist, head);
5870 list_insert_after(buflist, hdr, head);
5872 mutex_exit(&dev->l2ad_mtx);
5875 * We wait for the hash lock to become available
5876 * to try and prevent busy waiting, and increase
5877 * the chance we'll be able to acquire the lock
5878 * the next time around.
5880 mutex_enter(hash_lock);
5881 mutex_exit(hash_lock);
5886 * We could not have been moved into the arc_l2c_only
5887 * state while in-flight due to our ARC_FLAG_L2_WRITING
5888 * bit being set. Let's just ensure that's being enforced.
5890 ASSERT(HDR_HAS_L1HDR(hdr));
5893 * We may have allocated a buffer for L2ARC compression,
5894 * we must release it to avoid leaking this data.
5896 l2arc_release_cdata_buf(hdr);
5898 if (zio->io_error != 0) {
5900 * Error - drop L2ARC entry.
5902 list_remove(buflist, hdr);
5903 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev,
5904 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0);
5905 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
5907 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize);
5908 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
5910 bytes_dropped += hdr->b_l2hdr.b_asize;
5911 (void) refcount_remove_many(&dev->l2ad_alloc,
5912 hdr->b_l2hdr.b_asize, hdr);
5916 * Allow ARC to begin reads and ghost list evictions to
5919 hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
5921 mutex_exit(hash_lock);
5924 atomic_inc_64(&l2arc_writes_done);
5925 list_remove(buflist, head);
5926 ASSERT(!HDR_HAS_L1HDR(head));
5927 kmem_cache_free(hdr_l2only_cache, head);
5928 mutex_exit(&dev->l2ad_mtx);
5930 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
5932 l2arc_do_free_on_write();
5934 kmem_free(cb, sizeof (l2arc_write_callback_t));
5938 * A read to a cache device completed. Validate buffer contents before
5939 * handing over to the regular ARC routines.
5942 l2arc_read_done(zio_t *zio)
5944 l2arc_read_callback_t *cb;
5947 kmutex_t *hash_lock;
5950 ASSERT(zio->io_vd != NULL);
5951 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
5953 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
5955 cb = zio->io_private;
5957 buf = cb->l2rcb_buf;
5958 ASSERT(buf != NULL);
5960 hash_lock = HDR_LOCK(buf->b_hdr);
5961 mutex_enter(hash_lock);
5963 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5966 * If the buffer was compressed, decompress it first.
5968 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
5969 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
5970 ASSERT(zio->io_data != NULL);
5973 * Check this survived the L2ARC journey.
5975 equal = arc_cksum_equal(buf);
5976 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
5977 mutex_exit(hash_lock);
5978 zio->io_private = buf;
5979 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
5980 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
5983 mutex_exit(hash_lock);
5985 * Buffer didn't survive caching. Increment stats and
5986 * reissue to the original storage device.
5988 if (zio->io_error != 0) {
5989 ARCSTAT_BUMP(arcstat_l2_io_error);
5991 zio->io_error = SET_ERROR(EIO);
5994 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
5997 * If there's no waiter, issue an async i/o to the primary
5998 * storage now. If there *is* a waiter, the caller must
5999 * issue the i/o in a context where it's OK to block.
6001 if (zio->io_waiter == NULL) {
6002 zio_t *pio = zio_unique_parent(zio);
6004 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
6006 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
6007 buf->b_data, zio->io_size, arc_read_done, buf,
6008 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
6012 kmem_free(cb, sizeof (l2arc_read_callback_t));
6016 * This is the list priority from which the L2ARC will search for pages to
6017 * cache. This is used within loops (0..3) to cycle through lists in the
6018 * desired order. This order can have a significant effect on cache
6021 * Currently the metadata lists are hit first, MFU then MRU, followed by
6022 * the data lists. This function returns a locked list, and also returns
6025 static multilist_sublist_t *
6026 l2arc_sublist_lock(int list_num)
6028 multilist_t *ml = NULL;
6031 ASSERT(list_num >= 0 && list_num <= 3);
6035 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
6038 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
6041 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
6044 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
6049 * Return a randomly-selected sublist. This is acceptable
6050 * because the caller feeds only a little bit of data for each
6051 * call (8MB). Subsequent calls will result in different
6052 * sublists being selected.
6054 idx = multilist_get_random_index(ml);
6055 return (multilist_sublist_lock(ml, idx));
6059 * Evict buffers from the device write hand to the distance specified in
6060 * bytes. This distance may span populated buffers, it may span nothing.
6061 * This is clearing a region on the L2ARC device ready for writing.
6062 * If the 'all' boolean is set, every buffer is evicted.
6065 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
6068 arc_buf_hdr_t *hdr, *hdr_prev;
6069 kmutex_t *hash_lock;
6072 buflist = &dev->l2ad_buflist;
6074 if (!all && dev->l2ad_first) {
6076 * This is the first sweep through the device. There is
6082 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
6084 * When nearing the end of the device, evict to the end
6085 * before the device write hand jumps to the start.
6087 taddr = dev->l2ad_end;
6089 taddr = dev->l2ad_hand + distance;
6091 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
6092 uint64_t, taddr, boolean_t, all);
6095 mutex_enter(&dev->l2ad_mtx);
6096 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
6097 hdr_prev = list_prev(buflist, hdr);
6099 hash_lock = HDR_LOCK(hdr);
6102 * We cannot use mutex_enter or else we can deadlock
6103 * with l2arc_write_buffers (due to swapping the order
6104 * the hash lock and l2ad_mtx are taken).
6106 if (!mutex_tryenter(hash_lock)) {
6108 * Missed the hash lock. Retry.
6110 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
6111 mutex_exit(&dev->l2ad_mtx);
6112 mutex_enter(hash_lock);
6113 mutex_exit(hash_lock);
6117 if (HDR_L2_WRITE_HEAD(hdr)) {
6119 * We hit a write head node. Leave it for
6120 * l2arc_write_done().
6122 list_remove(buflist, hdr);
6123 mutex_exit(hash_lock);
6127 if (!all && HDR_HAS_L2HDR(hdr) &&
6128 (hdr->b_l2hdr.b_daddr > taddr ||
6129 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
6131 * We've evicted to the target address,
6132 * or the end of the device.
6134 mutex_exit(hash_lock);
6138 ASSERT(HDR_HAS_L2HDR(hdr));
6139 if (!HDR_HAS_L1HDR(hdr)) {
6140 ASSERT(!HDR_L2_READING(hdr));
6142 * This doesn't exist in the ARC. Destroy.
6143 * arc_hdr_destroy() will call list_remove()
6144 * and decrement arcstat_l2_size.
6146 arc_change_state(arc_anon, hdr, hash_lock);
6147 arc_hdr_destroy(hdr);
6149 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
6150 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
6152 * Invalidate issued or about to be issued
6153 * reads, since we may be about to write
6154 * over this location.
6156 if (HDR_L2_READING(hdr)) {
6157 ARCSTAT_BUMP(arcstat_l2_evict_reading);
6158 hdr->b_flags |= ARC_FLAG_L2_EVICTED;
6161 /* Ensure this header has finished being written */
6162 ASSERT(!HDR_L2_WRITING(hdr));
6163 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
6165 arc_hdr_l2hdr_destroy(hdr);
6167 mutex_exit(hash_lock);
6169 mutex_exit(&dev->l2ad_mtx);
6173 * Find and write ARC buffers to the L2ARC device.
6175 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6176 * for reading until they have completed writing.
6177 * The headroom_boost is an in-out parameter used to maintain headroom boost
6178 * state between calls to this function.
6180 * Returns the number of bytes actually written (which may be smaller than
6181 * the delta by which the device hand has changed due to alignment).
6184 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
6185 boolean_t *headroom_boost)
6187 arc_buf_hdr_t *hdr, *hdr_prev, *head;
6188 uint64_t write_asize, write_sz, headroom, buf_compress_minsz;
6191 l2arc_write_callback_t *cb;
6193 uint64_t guid = spa_load_guid(spa);
6194 const boolean_t do_headroom_boost = *headroom_boost;
6197 ASSERT(dev->l2ad_vdev != NULL);
6199 /* Lower the flag now, we might want to raise it again later. */
6200 *headroom_boost = B_FALSE;
6203 write_sz = write_asize = 0;
6205 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
6206 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD;
6207 head->b_flags |= ARC_FLAG_HAS_L2HDR;
6209 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
6211 * We will want to try to compress buffers that are at least 2x the
6212 * device sector size.
6214 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
6217 * Copy buffers for L2ARC writing.
6219 for (try = 0; try <= 3; try++) {
6220 multilist_sublist_t *mls = l2arc_sublist_lock(try);
6221 uint64_t passed_sz = 0;
6223 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
6226 * L2ARC fast warmup.
6228 * Until the ARC is warm and starts to evict, read from the
6229 * head of the ARC lists rather than the tail.
6231 if (arc_warm == B_FALSE)
6232 hdr = multilist_sublist_head(mls);
6234 hdr = multilist_sublist_tail(mls);
6236 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
6238 headroom = target_sz * l2arc_headroom;
6239 if (do_headroom_boost)
6240 headroom = (headroom * l2arc_headroom_boost) / 100;
6242 for (; hdr; hdr = hdr_prev) {
6243 kmutex_t *hash_lock;
6247 if (arc_warm == B_FALSE)
6248 hdr_prev = multilist_sublist_next(mls, hdr);
6250 hdr_prev = multilist_sublist_prev(mls, hdr);
6251 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size);
6253 hash_lock = HDR_LOCK(hdr);
6254 if (!mutex_tryenter(hash_lock)) {
6255 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
6257 * Skip this buffer rather than waiting.
6262 passed_sz += hdr->b_size;
6263 if (passed_sz > headroom) {
6267 mutex_exit(hash_lock);
6268 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
6272 if (!l2arc_write_eligible(guid, hdr)) {
6273 mutex_exit(hash_lock);
6278 * Assume that the buffer is not going to be compressed
6279 * and could take more space on disk because of a larger
6282 buf_sz = hdr->b_size;
6283 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
6285 if ((write_asize + buf_a_sz) > target_sz) {
6287 mutex_exit(hash_lock);
6288 ARCSTAT_BUMP(arcstat_l2_write_full);
6294 * Insert a dummy header on the buflist so
6295 * l2arc_write_done() can find where the
6296 * write buffers begin without searching.
6298 mutex_enter(&dev->l2ad_mtx);
6299 list_insert_head(&dev->l2ad_buflist, head);
6300 mutex_exit(&dev->l2ad_mtx);
6303 sizeof (l2arc_write_callback_t), KM_SLEEP);
6304 cb->l2wcb_dev = dev;
6305 cb->l2wcb_head = head;
6306 pio = zio_root(spa, l2arc_write_done, cb,
6308 ARCSTAT_BUMP(arcstat_l2_write_pios);
6312 * Create and add a new L2ARC header.
6314 hdr->b_l2hdr.b_dev = dev;
6315 hdr->b_flags |= ARC_FLAG_L2_WRITING;
6317 * Temporarily stash the data buffer in b_tmp_cdata.
6318 * The subsequent write step will pick it up from
6319 * there. This is because can't access b_l1hdr.b_buf
6320 * without holding the hash_lock, which we in turn
6321 * can't access without holding the ARC list locks
6322 * (which we want to avoid during compression/writing).
6324 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
6325 hdr->b_l2hdr.b_asize = hdr->b_size;
6326 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data;
6329 * Explicitly set the b_daddr field to a known
6330 * value which means "invalid address". This
6331 * enables us to differentiate which stage of
6332 * l2arc_write_buffers() the particular header
6333 * is in (e.g. this loop, or the one below).
6334 * ARC_FLAG_L2_WRITING is not enough to make
6335 * this distinction, and we need to know in
6336 * order to do proper l2arc vdev accounting in
6337 * arc_release() and arc_hdr_destroy().
6339 * Note, we can't use a new flag to distinguish
6340 * the two stages because we don't hold the
6341 * header's hash_lock below, in the second stage
6342 * of this function. Thus, we can't simply
6343 * change the b_flags field to denote that the
6344 * IO has been sent. We can change the b_daddr
6345 * field of the L2 portion, though, since we'll
6346 * be holding the l2ad_mtx; which is why we're
6347 * using it to denote the header's state change.
6349 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET;
6350 hdr->b_flags |= ARC_FLAG_HAS_L2HDR;
6352 mutex_enter(&dev->l2ad_mtx);
6353 list_insert_head(&dev->l2ad_buflist, hdr);
6354 mutex_exit(&dev->l2ad_mtx);
6357 * Compute and store the buffer cksum before
6358 * writing. On debug the cksum is verified first.
6360 arc_cksum_verify(hdr->b_l1hdr.b_buf);
6361 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE);
6363 mutex_exit(hash_lock);
6366 write_asize += buf_a_sz;
6369 multilist_sublist_unlock(mls);
6375 /* No buffers selected for writing? */
6378 ASSERT(!HDR_HAS_L1HDR(head));
6379 kmem_cache_free(hdr_l2only_cache, head);
6383 mutex_enter(&dev->l2ad_mtx);
6386 * Note that elsewhere in this file arcstat_l2_asize
6387 * and the used space on l2ad_vdev are updated using b_asize,
6388 * which is not necessarily rounded up to the device block size.
6389 * Too keep accounting consistent we do the same here as well:
6390 * stats_size accumulates the sum of b_asize of the written buffers,
6391 * while write_asize accumulates the sum of b_asize rounded up
6392 * to the device block size.
6393 * The latter sum is used only to validate the corectness of the code.
6395 uint64_t stats_size = 0;
6399 * Now start writing the buffers. We're starting at the write head
6400 * and work backwards, retracing the course of the buffer selector
6403 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr;
6404 hdr = list_prev(&dev->l2ad_buflist, hdr)) {
6408 * We rely on the L1 portion of the header below, so
6409 * it's invalid for this header to have been evicted out
6410 * of the ghost cache, prior to being written out. The
6411 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6413 ASSERT(HDR_HAS_L1HDR(hdr));
6416 * We shouldn't need to lock the buffer here, since we flagged
6417 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6418 * take care to only access its L2 cache parameters. In
6419 * particular, hdr->l1hdr.b_buf may be invalid by now due to
6422 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
6424 if ((HDR_L2COMPRESS(hdr)) &&
6425 hdr->b_l2hdr.b_asize >= buf_compress_minsz) {
6426 if (l2arc_compress_buf(hdr)) {
6428 * If compression succeeded, enable headroom
6429 * boost on the next scan cycle.
6431 *headroom_boost = B_TRUE;
6436 * Pick up the buffer data we had previously stashed away
6437 * (and now potentially also compressed).
6439 buf_data = hdr->b_l1hdr.b_tmp_cdata;
6440 buf_sz = hdr->b_l2hdr.b_asize;
6443 * We need to do this regardless if buf_sz is zero or
6444 * not, otherwise, when this l2hdr is evicted we'll
6445 * remove a reference that was never added.
6447 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr);
6449 /* Compression may have squashed the buffer to zero length. */
6453 wzio = zio_write_phys(pio, dev->l2ad_vdev,
6454 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
6455 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
6456 ZIO_FLAG_CANFAIL, B_FALSE);
6458 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
6460 (void) zio_nowait(wzio);
6462 stats_size += buf_sz;
6465 * Keep the clock hand suitably device-aligned.
6467 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
6468 write_asize += buf_a_sz;
6469 dev->l2ad_hand += buf_a_sz;
6473 mutex_exit(&dev->l2ad_mtx);
6475 ASSERT3U(write_asize, <=, target_sz);
6476 ARCSTAT_BUMP(arcstat_l2_writes_sent);
6477 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
6478 ARCSTAT_INCR(arcstat_l2_size, write_sz);
6479 ARCSTAT_INCR(arcstat_l2_asize, stats_size);
6480 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0);
6483 * Bump device hand to the device start if it is approaching the end.
6484 * l2arc_evict() will already have evicted ahead for this case.
6486 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
6487 dev->l2ad_hand = dev->l2ad_start;
6488 dev->l2ad_first = B_FALSE;
6491 dev->l2ad_writing = B_TRUE;
6492 (void) zio_wait(pio);
6493 dev->l2ad_writing = B_FALSE;
6495 return (write_asize);
6499 * Compresses an L2ARC buffer.
6500 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6501 * size in l2hdr->b_asize. This routine tries to compress the data and
6502 * depending on the compression result there are three possible outcomes:
6503 * *) The buffer was incompressible. The original l2hdr contents were left
6504 * untouched and are ready for writing to an L2 device.
6505 * *) The buffer was all-zeros, so there is no need to write it to an L2
6506 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6507 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6508 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6509 * data buffer which holds the compressed data to be written, and b_asize
6510 * tells us how much data there is. b_compress is set to the appropriate
6511 * compression algorithm. Once writing is done, invoke
6512 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6514 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6515 * buffer was incompressible).
6518 l2arc_compress_buf(arc_buf_hdr_t *hdr)
6521 size_t csize, len, rounded;
6522 ASSERT(HDR_HAS_L2HDR(hdr));
6523 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
6525 ASSERT(HDR_HAS_L1HDR(hdr));
6526 ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF);
6527 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6529 len = l2hdr->b_asize;
6530 cdata = zio_data_buf_alloc(len);
6531 ASSERT3P(cdata, !=, NULL);
6532 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata,
6533 cdata, l2hdr->b_asize);
6536 /* zero block, indicate that there's nothing to write */
6537 zio_data_buf_free(cdata, len);
6538 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY);
6540 hdr->b_l1hdr.b_tmp_cdata = NULL;
6541 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
6545 rounded = P2ROUNDUP(csize,
6546 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift);
6547 if (rounded < len) {
6549 * Compression succeeded, we'll keep the cdata around for
6550 * writing and release it afterwards.
6552 if (rounded > csize) {
6553 bzero((char *)cdata + csize, rounded - csize);
6556 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4);
6557 l2hdr->b_asize = csize;
6558 hdr->b_l1hdr.b_tmp_cdata = cdata;
6559 ARCSTAT_BUMP(arcstat_l2_compress_successes);
6563 * Compression failed, release the compressed buffer.
6564 * l2hdr will be left unmodified.
6566 zio_data_buf_free(cdata, len);
6567 ARCSTAT_BUMP(arcstat_l2_compress_failures);
6573 * Decompresses a zio read back from an l2arc device. On success, the
6574 * underlying zio's io_data buffer is overwritten by the uncompressed
6575 * version. On decompression error (corrupt compressed stream), the
6576 * zio->io_error value is set to signal an I/O error.
6578 * Please note that the compressed data stream is not checksummed, so
6579 * if the underlying device is experiencing data corruption, we may feed
6580 * corrupt data to the decompressor, so the decompressor needs to be
6581 * able to handle this situation (LZ4 does).
6584 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
6586 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
6588 if (zio->io_error != 0) {
6590 * An io error has occured, just restore the original io
6591 * size in preparation for a main pool read.
6593 zio->io_orig_size = zio->io_size = hdr->b_size;
6597 if (c == ZIO_COMPRESS_EMPTY) {
6599 * An empty buffer results in a null zio, which means we
6600 * need to fill its io_data after we're done restoring the
6601 * buffer's contents.
6603 ASSERT(hdr->b_l1hdr.b_buf != NULL);
6604 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size);
6605 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data;
6607 ASSERT(zio->io_data != NULL);
6609 * We copy the compressed data from the start of the arc buffer
6610 * (the zio_read will have pulled in only what we need, the
6611 * rest is garbage which we will overwrite at decompression)
6612 * and then decompress back to the ARC data buffer. This way we
6613 * can minimize copying by simply decompressing back over the
6614 * original compressed data (rather than decompressing to an
6615 * aux buffer and then copying back the uncompressed buffer,
6616 * which is likely to be much larger).
6621 csize = zio->io_size;
6622 cdata = zio_data_buf_alloc(csize);
6623 bcopy(zio->io_data, cdata, csize);
6624 if (zio_decompress_data(c, cdata, zio->io_data, csize,
6626 zio->io_error = EIO;
6627 zio_data_buf_free(cdata, csize);
6630 /* Restore the expected uncompressed IO size. */
6631 zio->io_orig_size = zio->io_size = hdr->b_size;
6635 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6636 * This buffer serves as a temporary holder of compressed data while
6637 * the buffer entry is being written to an l2arc device. Once that is
6638 * done, we can dispose of it.
6641 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr)
6643 enum zio_compress comp = HDR_GET_COMPRESS(hdr);
6645 ASSERT(HDR_HAS_L1HDR(hdr));
6646 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp));
6648 if (comp == ZIO_COMPRESS_OFF) {
6650 * In this case, b_tmp_cdata points to the same buffer
6651 * as the arc_buf_t's b_data field. We don't want to
6652 * free it, since the arc_buf_t will handle that.
6654 hdr->b_l1hdr.b_tmp_cdata = NULL;
6655 } else if (comp == ZIO_COMPRESS_EMPTY) {
6657 * In this case, b_tmp_cdata was compressed to an empty
6658 * buffer, thus there's nothing to free and b_tmp_cdata
6659 * should have been set to NULL in l2arc_write_buffers().
6661 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
6664 * If the data was compressed, then we've allocated a
6665 * temporary buffer for it, so now we need to release it.
6667 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6668 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata,
6670 hdr->b_l1hdr.b_tmp_cdata = NULL;
6676 * This thread feeds the L2ARC at regular intervals. This is the beating
6677 * heart of the L2ARC.
6680 l2arc_feed_thread(void *dummy __unused)
6685 uint64_t size, wrote;
6686 clock_t begin, next = ddi_get_lbolt();
6687 boolean_t headroom_boost = B_FALSE;
6689 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
6691 mutex_enter(&l2arc_feed_thr_lock);
6693 while (l2arc_thread_exit == 0) {
6694 CALLB_CPR_SAFE_BEGIN(&cpr);
6695 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
6696 next - ddi_get_lbolt());
6697 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
6698 next = ddi_get_lbolt() + hz;
6701 * Quick check for L2ARC devices.
6703 mutex_enter(&l2arc_dev_mtx);
6704 if (l2arc_ndev == 0) {
6705 mutex_exit(&l2arc_dev_mtx);
6708 mutex_exit(&l2arc_dev_mtx);
6709 begin = ddi_get_lbolt();
6712 * This selects the next l2arc device to write to, and in
6713 * doing so the next spa to feed from: dev->l2ad_spa. This
6714 * will return NULL if there are now no l2arc devices or if
6715 * they are all faulted.
6717 * If a device is returned, its spa's config lock is also
6718 * held to prevent device removal. l2arc_dev_get_next()
6719 * will grab and release l2arc_dev_mtx.
6721 if ((dev = l2arc_dev_get_next()) == NULL)
6724 spa = dev->l2ad_spa;
6725 ASSERT(spa != NULL);
6728 * If the pool is read-only then force the feed thread to
6729 * sleep a little longer.
6731 if (!spa_writeable(spa)) {
6732 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
6733 spa_config_exit(spa, SCL_L2ARC, dev);
6738 * Avoid contributing to memory pressure.
6740 if (arc_reclaim_needed()) {
6741 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
6742 spa_config_exit(spa, SCL_L2ARC, dev);
6746 ARCSTAT_BUMP(arcstat_l2_feeds);
6748 size = l2arc_write_size();
6751 * Evict L2ARC buffers that will be overwritten.
6753 l2arc_evict(dev, size, B_FALSE);
6756 * Write ARC buffers.
6758 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
6761 * Calculate interval between writes.
6763 next = l2arc_write_interval(begin, size, wrote);
6764 spa_config_exit(spa, SCL_L2ARC, dev);
6767 l2arc_thread_exit = 0;
6768 cv_broadcast(&l2arc_feed_thr_cv);
6769 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
6774 l2arc_vdev_present(vdev_t *vd)
6778 mutex_enter(&l2arc_dev_mtx);
6779 for (dev = list_head(l2arc_dev_list); dev != NULL;
6780 dev = list_next(l2arc_dev_list, dev)) {
6781 if (dev->l2ad_vdev == vd)
6784 mutex_exit(&l2arc_dev_mtx);
6786 return (dev != NULL);
6790 * Add a vdev for use by the L2ARC. By this point the spa has already
6791 * validated the vdev and opened it.
6794 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
6796 l2arc_dev_t *adddev;
6798 ASSERT(!l2arc_vdev_present(vd));
6800 vdev_ashift_optimize(vd);
6803 * Create a new l2arc device entry.
6805 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
6806 adddev->l2ad_spa = spa;
6807 adddev->l2ad_vdev = vd;
6808 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
6809 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
6810 adddev->l2ad_hand = adddev->l2ad_start;
6811 adddev->l2ad_first = B_TRUE;
6812 adddev->l2ad_writing = B_FALSE;
6814 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
6816 * This is a list of all ARC buffers that are still valid on the
6819 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
6820 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
6822 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
6823 refcount_create(&adddev->l2ad_alloc);
6826 * Add device to global list
6828 mutex_enter(&l2arc_dev_mtx);
6829 list_insert_head(l2arc_dev_list, adddev);
6830 atomic_inc_64(&l2arc_ndev);
6831 mutex_exit(&l2arc_dev_mtx);
6835 * Remove a vdev from the L2ARC.
6838 l2arc_remove_vdev(vdev_t *vd)
6840 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
6843 * Find the device by vdev
6845 mutex_enter(&l2arc_dev_mtx);
6846 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
6847 nextdev = list_next(l2arc_dev_list, dev);
6848 if (vd == dev->l2ad_vdev) {
6853 ASSERT(remdev != NULL);
6856 * Remove device from global list
6858 list_remove(l2arc_dev_list, remdev);
6859 l2arc_dev_last = NULL; /* may have been invalidated */
6860 atomic_dec_64(&l2arc_ndev);
6861 mutex_exit(&l2arc_dev_mtx);
6864 * Clear all buflists and ARC references. L2ARC device flush.
6866 l2arc_evict(remdev, 0, B_TRUE);
6867 list_destroy(&remdev->l2ad_buflist);
6868 mutex_destroy(&remdev->l2ad_mtx);
6869 refcount_destroy(&remdev->l2ad_alloc);
6870 kmem_free(remdev, sizeof (l2arc_dev_t));
6876 l2arc_thread_exit = 0;
6878 l2arc_writes_sent = 0;
6879 l2arc_writes_done = 0;
6881 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
6882 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
6883 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
6884 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
6886 l2arc_dev_list = &L2ARC_dev_list;
6887 l2arc_free_on_write = &L2ARC_free_on_write;
6888 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
6889 offsetof(l2arc_dev_t, l2ad_node));
6890 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
6891 offsetof(l2arc_data_free_t, l2df_list_node));
6898 * This is called from dmu_fini(), which is called from spa_fini();
6899 * Because of this, we can assume that all l2arc devices have
6900 * already been removed when the pools themselves were removed.
6903 l2arc_do_free_on_write();
6905 mutex_destroy(&l2arc_feed_thr_lock);
6906 cv_destroy(&l2arc_feed_thr_cv);
6907 mutex_destroy(&l2arc_dev_mtx);
6908 mutex_destroy(&l2arc_free_on_write_mtx);
6910 list_destroy(l2arc_dev_list);
6911 list_destroy(l2arc_free_on_write);
6917 if (!(spa_mode_global & FWRITE))
6920 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
6921 TS_RUN, minclsyspri);
6927 if (!(spa_mode_global & FWRITE))
6930 mutex_enter(&l2arc_feed_thr_lock);
6931 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
6932 l2arc_thread_exit = 1;
6933 while (l2arc_thread_exit != 0)
6934 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
6935 mutex_exit(&l2arc_feed_thr_lock);